acceptodds
Withdrawn from ICLR 2027

Learning dynamical systems with regional stability certificates

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Regionally stable systems produce bounded outputs only on part of their state and input space, outside it, their outputs grow unbounded even for bounded inputs. Unconstrained models can reproduce such behavior but provide no guarantees, whereas globally stable models provide guarantees but cannot predict unbounded outputs. We propose a dynamic recurrent neural network in Lur'e form with deadzone nonlinearity. Exploiting the deadzone, we derive sufficient conditions on the learnable parameters that yield a regional stability certificate, which consists of an invariant state set, a state-dependent input map, and an output set that contains the model predictions whenever the initial state and inputs are admissible. During training, we enforce these conditions with logarithmic barrier functions and enlarge the admissible input map with a semidefinite program. On two synthetic systems, the learned model recovers the regional stability properties of the data-generating system and reproduces its unbounded response. On datasets from a nonlinear system identification benchmark that do not contain unbounded outputs, they match the accuracy of globally stable models, while unconstrained models are more accurate but uncertified.

Primary area:
learning on time series and dynamical systems
Keywords:
system identification, stable recurrent neural networks, learning dynamical systems

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qVPaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the challenge of learning a regionally stable nonlinear system, i.e., a nonlinear system with local input-to-state stability guarantees. The approach is based on a parametrization of a Lur'e system and an LMI condition ensuring regional stability within a polyhedral domain. This LMI condition is incorporated into the training procedure of identifying the model parameters to learn systems with regional stability guarantees. The approach is demonstrated in several numerical benchmarks on small- to medium-scale nonlinear systems.

Critical strengths
  • The new LMI condition enforcing regional stability appears novel to me. It provides a rigorous way of establishing regional stability with concrete estimates on both the subset of state space and input space that ensure local input to state stability

  • The paper is written clearly and is comprehensible for someone with a control theory background. Thank you

Critical weaknesses
  • The numerical experiments are limited to small-scale systems where the proposed approach has on the order of 10s to 100s of trainable parameters. This does not seem comprehensive enough to address the scalability of the approach or how it handles systems with more states. Is this approach applicable to a control environment like those in Gym / RL training environments?
Questions for authors
  • Why not address regional stability in the following manner: first identify the region in which local ISS holds, then fit an existing model with global guarantees in this region?

  • The paper studies dynamical systems affected by disturbances. Are there versions of these results that extend to control systems where the user designs the control input?

  • The matrix XX in equation (13) is not defined

  • Any training approach that relies on solving SDPs at each iteration is fundamentally limited in the scale of problems that it can handle. Is there any way to circumvent the LMI constraints by directly constraining the conditions to hold via a parametrization? I can see the challenge in training in Table 3 where even a 1D nonlinear system takes 2 hours to train using the proposed approach.

1 comment

We thank the reviewer for the comments.

Regarding the Gym/RL environments:

The open-loop dynamics of the classical-control environments in Gym RL are not regionally stable. Our certificate requires a regionally exponentially stable equilibrium, so it does not apply to these systems directly. It does apply to the closed loop of such an environment with a stabilizing policy, which is a typical regionally stable system: a balancing policy keeps the cart-pole upright only for a bounded range of states and disturbances.

Below, we address the questions raised by the reviewer:

1.) The main contribution of our approach is that

(a) we derive conditions for regional stability that provide an invariant state set and output set and a set of admissible inputs for which the model is input-to-state stable and

(b) we develop a training algorithm that discovers the regional stability certificates while minimizing the model's prediction error.

From the description of the reviewer’s proposed approach, it is not clear how global ISS can be achieved while keeping it only in a specified region for an existing model. Moreover, discovering the stability region is precisely what our architecture provides.

2.) The closest related work that operates in a closed-loop setup is the reference La Bella et al. (2025).

3.) Thanks, we will add the missing definition.

4.) The semidefinite program does not have to be solved in every optimization step. It is solved only when

(i) the new parameters are not feasible or

(ii) the input trajectory is not in the input set.

Embedding the constraints into the model parameters, enabling unconstrained optimization, is not trivial and left for future work.

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

The paper proposes a Lur’e-form recurrent neural network with dead-zone nonlinearities that provides regional stability guarantees while retaining the ability to model unbounded outputs outside the certified region. Experiments show that it captures regional stability and unbounded behavior while achieving accuracy comparable to globally stable models on standard benchmarks.

Critical strengths
  1. The theory part seems to be solid.
Critical weaknesses
  1. My main concern is that this work appears to fall primarily within the scope of classical control theory, and its relevance to the broader machine learning community is not sufficiently clear. Most of the related work also comes from the control literature. The authors should better motivate the ML relevance and clarify what new insights or capabilities this work offers from a machine learning perspective.

  2. The motivation for regional stability could be strengthened. Why is regional stability particularly important for the applications considered here? More specifically, why is it desirable or practically relevant for a learned model to capture unstable or unbounded behavior outside a certain region?

  3. If the primary objective is to ensure that a compact set is forward invariant, this can seemingly be achieved using projection or other relatively simple constraint-enforcement mechanisms.

  4. The experimental evaluation appears somewhat limited, as the datasets considered are generated from relatively low-dimensional systems. Moreover, the baselines are quite weak. It should be compared with some recent methods from the ML literature, e.g., SSM.

  5. In addition, I do not understand the use of boldface numbers in Table 2. If lower NRMSE indicates better performance, what criterion is used to determine which entries are highlighted in bold? It seems that the proposed method underperform the baseline in many cases.

Questions for authors

see the weaknesses.

1 comment

We thank the reviewer for the comments and appreciate the suggestions.

The reviewer has expressed a strong opinion about our work. While we are convinced that our answers should satisfy the criticism, we also acknowledge that priorities about what is relevant differ between researchers. In case our answers do not respond satisfactorily to the reviewer’s critique, we would appreciate answers to the following questions:

a.) Why is learning stable dynamics from input-output measurements not considered relevant from an ML perspective?

b.) What would be a valuable application of learned dynamics that goes beyond simulation and controller design for the ML community?

Below we address the weaknesses the reviewer raised:

1.) Relevance to ML

Learning stable models for dynamical systems from input-output data is, in our understanding, a machine learning problem (e.g., Miller and Hardt (2019), Erichson et al. (2019)). One of our main machine learning insights is that a global form of model stability is an overly restrictive inductive bias for many dynamical systems, as it rules out all dynamical systems with diverging outputs.

2.) Why regional stability?

(i) When the learned model is used to simulate the response of the underlying system, its predictions are guaranteed to remain bounded whenever the input and initial state are admissible. In contrast, an unconstrained RNN can diverge unpredictably, for example, even for slight deviations from training data.

(ii) The input-to-state bound relates inputs to states and gives a maximum gain between the input and the internal state, which also holds for unseen data.

(iii) We can check whether a given input is covered by the certificate; this can be used for controller design or in model-based RL to steer the system response towards the certified region.

3.) Why not projection?

(i) Characterizing this set is what Theorem 1 provides. Without this characterization, it is not clear what to project on. For an input-driven model, invariance is only meaningful relative to a set of admissible inputs, since, in general, a large enough input drives the system state out of any bounded set.

(ii) Naively projecting the state onto a compact set makes the learned model bound everywhere, so it will not be able to predict unbounded outputs. In addition, invariance does not imply input-to-state stability; thus, a projected model may oscillate. Finally, the proposed approach does not enforce an a priori invariant set, rather the size and shape of this set is discovered during learning and is therefore a result of the training algorithm.

5.) Our accuracy is on par with the best models while discovering knowledge about stable input regions and the stability of the learned model.

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

This paper addresses regional stability guarantees and learning from input-output data for a Lur'e-form recurrent model consisting of a linear state-space block and a fixed deadzone nonlinearity. It derives matrix inequality conditions that guarantee forward invariance of the state set, an input-to-state stability bound, and bounded outputs, provided that the initial state lies in the specified state set and the input at each time step belongs to a state-dependent admissible set. Training updates the model parameters and auxiliary variables using gradient-based optimization of a loss comprising prediction error and logarithmic barrier terms for the stability constraints. In two numerical simulation examples, the method achieves predictive accuracy close to that of an unconstrained model while providing a stability certificate, and reproduces divergent behavior that models constrained to be globally stable cannot represent. On five system-identification benchmarks, its predictive accuracy is broadly comparable to that of the globally stable baseline.

Critical strengths
  • On the theoretical side, the paper provides sufficient matrix inequality conditions guaranteeing an input-to-state stability bound and an output bound for the considered Lur'e-form model. These conditions allow explicit evaluation of the ranges of initial states and inputs for which the guarantees hold.
  • In two numerical simulation examples, the method achieves predictive accuracy close to that of a model without stability constraints while retaining regional stability guarantees. It also reproduces divergent behavior that the globally stable baseline cannot capture, illustrating the usefulness of restricting the region over which stability is guaranteed.
Critical weaknesses
  1. Suitability for the venue. The main contribution concerns stability analysis of a system formed by incorporating a nonlinear function into part of a linear state-space representation, and system identification subject to the resulting conditions. New methodology or insights for machine learning are not sufficiently established, and the work is primarily a study in control theory. I consider a control-theory venue more appropriate than ICLR.

  2. Positioning as an RNN contribution. The model is a specific Lur'e system with a linear block and nonlinearity confined to a fixed deadzone function. The guarantee in Theorem 1 depends on this structure and the properties of the nonlinear function, and does not directly apply to general RNN architectures. The results strongly depend on the chosen structure, limiting their generality as a contribution to RNN learning.

  3. Theoretical novelty. The generalized sector conditions, quadratic Lyapunov function, and matrix inequality analysis of invariant sets build on existing control theory. Although Theorem 1 provides an admissible set for exogenous inputs and an input-to-state stability bound, its nontrivial theoretical advance over closely related prior work is insufficiently clear. Moreover, the theorem concerns stability of a given model and does not provide guarantees on identification accuracy or error when learning from data.

  4. Contribution and convergence of the learning method. The learning method combines gradient descent, logarithmic barriers, and semidefinite programs to restore feasibility of the guarantee conditions and enlarge the certified range, but the methodological novelty of this combination is unclear. If a constraint is violated, the algorithm attempts a repair and discards the update if the repair fails; however, neither monotonic objective decrease nor convergence to a stationary point is established for the overall algorithm. The stability guarantee for state trajectories in Theorem 1 is distinct from a convergence guarantee for this learning procedure.

Questions for authors
  1. Beyond existing control-theoretic insights, what new methodology or insights for RNN learning follow from these results, which depend on a specific Lur'e-form model?
  2. Please explain the nontrivial theoretical additions in Theorem 1 relative to closely related prior work, such as La Bella et al. (2025).
  3. What is novel about the proposed learning procedure beyond combining existing methods? What guarantees or supporting evidence are available concerning convergence of the overall learning algorithm?
2 comments

We thank the reviewer for the comments and value the questions.

The reviewer has expressed a strong opinion about our work. While we are convinced that our answers should satisfy the criticism, we also acknowledge that priorities about what is relevant differ between researchers. In case our answers do not respond satisfactorily to the reviewer’s critique, we would appreciate answers to the following questions:

a) Why would a particular RNN architecture that is suitable for learning dynamic systems not be considered an ML contribution?

b) Why is knowledge discovered about a learned model (i.e., regional stability) not a valuable insight, though it may be exploited in various applications?

Below we address the weaknesses and questions the reviewer raised:

1.) (Question 1) Relevance to ML Learning stable models for dynamical systems from input-output data is, in our understanding, a machine learning problem (e.g., Miller and Hardt (2019), Erichson et al. (2019)). One of our main machine learning insights is that a global form of model stability is an overly restrictive inductive bias for many dynamical systems, as it rules out all dynamical systems with diverging outputs.

2.) Generality of the model class Our model class subsumes standard RNNs with saturating activation (hard tanh) and zero bias terms. The saturating activation function is a deliberate choice that enables exploiting generalized sector conditions for regional stability.

3.) (Question 2) Novelty of Theorem 1 relative to La Bella et al. (2025) (i) La Bella et al. (2025) consider autonomous systems, while the novelty of Theorem 1 lies in regional stability for systems with external inputs, which is the main enabler for exploiting Theorem 1 for the purpose of learning dynamical systems with inputs. We provide an input condition, which, once satisfied, certifies whether applying a given input to the learned model keeps its state and output in the targeted region. (ii) We consider a modeling problem for systems with external inputs and provide a regional input-to-state stability result for the learned model. The architecture in La Bella et al. (2025) is a closed-loop interconnection (without external inputs) that provides exponential stability in a region of the state space.

4.) (Question 3) Learning procedure Determining the region of allowed inputs is an explicit training objective, optimized jointly with the prediction error. This is the key innovation of our approach.

Reviewer_KHjrOpenReview ↗

Assessment Following the Authors’ Response

Thank you for your response. As noted in my original review, I recognize that global stability constitutes an inductive bias and acknowledge the benefit of combining regional stability guarantees with more flexible modeling.

Regarding Q1, my concern is the limited generality of the theoretical results and learning procedure. They depend strongly on a specific Lur’e structure and nonlinearity, restricting their applicability to that model class. The explanation that the model class includes some standard RNNs with hard tanh activations does not resolve this limitation. The experiments demonstrate predictive accuracy close to that of an unconstrained model while retaining guarantees on two synthetic systems. However, the real-data benchmarks contain no diverging trajectories and therefore do not directly test the distinctive benefits of regional stability. Given the generality of the results and the scope of the empirical evidence, concerns remain about the significance of the contribution to learning methodology and insights.

Regarding Q2, I recognize the differences from prior work in the scope and conclusions of the guarantees. My question concerned the theoretical challenges newly addressed in extending the existing results. The response explains the differences in the guarantees, but the theoretical ideas required for this extension and their significance remain insufficiently clear.

Regarding Q3, I understand the proposed formulation’s joint optimization of prediction error and the admissible input region. However, my question concerned how the procedure combining gradient-based optimization, logarithmic barriers, and semidefinite programming relates to existing research on constrained optimization, and what methodological additions it provides. The response explains the optimization objective, but the relationship to existing optimization methods and the advances beyond them remain unclear. The response also does not directly address the question about convergence guarantees or supporting evidence for the overall learning algorithm.

I understand the explanations in your response and acknowledge the reported results. Nevertheless, given the scope limitations and the unresolved theoretical and methodological issues above, my assessment remains that the paper has not demonstrated a contribution of sufficient significance to support acceptance. I therefore maintain my Clear Reject rating.

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

Related papers

Discussion (0)

Sign in to comment.