acceptodds
Under review as a conference paper at ICLR 2027

GAIN-Cert: Gap-Aware Certification of Disturbance-to-State Gain Curves

Abstract

Formal stability at one disturbance budget gives limited guidance for a learned system exposed to disturbances of varying magnitude. In inverter-rich power systems and other cyber-physical settings, the relevant question is which disturbance magnitudes keep the eventual state deviation within a prescribed operating limit. Repeating Lyapunov verification at selected budgets yields isolated certificates: interpolating them lacks a formal guarantee, whereas endpoint constructions can be needlessly conservative. We study certified disturbance-to-state gain-curve synthesis for nonlinear and learned ordinary differential equations. At each disturbance magnitude in a prescribed range, the curve bounds the ultimate state deviation for every admissible disturbance signal of up to that magnitude. For a fixed Lyapunov certificate and regional verifier, we define a disturbance-dependent threshold profile and separate conservatism caused by the certificate, state-norm comparison, and verifier from that introduced by representing and verifying a curve. Targeting the latter, we train a monotone integral network against a pointwise branch-and-bound reference and allocate knots adaptively where the bound changes rapidly. A piecewise envelope is checked jointly over state and disturbance regions, repaired where verification fails, and tightened after certification. We prove soundness of the resulting curve under explicit operating-region assumptions and measure the contributions of neural approximation, verification repair, and tightening to its gap. We evaluate on six classical ODEs and two learning-enabled systems against pointwise branch-and-bound and staircase, affine-level, and neural curve baselines. The comparisons show smaller normalized bound area at competitive end-to-end cost.

Primary area:
learning theory
Keywords:
robustness certification, learned dynamical systems, Lyapunov stability, neural networks, formal verification, input-to-state stability, neural ODEs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tPNqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks for a certified curve γ(ωˉ)\gamma(\bar\omega) that bounds the ultimate state deviation for every disturbance level in [0,ωˉmax][0,\bar\omega_{max}], given a fixed Lyapunov function VV and operating region. It defines a threshold H(ωˉ)H(\bar\omega) for feasible Lyapunov levels and builds the curve in four steps: pointwise branch-and-bound (BaB) at sampled budgets; a monotone integral neural network fitted to those points, with adaptive knots; conversion to a piecewise-affine level curve that is verified jointly over state and disturbance on each interval; and local upward repair followed by tightening. Theorem 2 proves soundness, and an identity splits the gap to BaB into proposal, repair, and tightening terms. Experiments on six low-dimensional ODEs and two learned systems report that the staircase, affine-level, and uniform-knot neural baselines have 12–18% larger normalized area than GAIN-Cert, and that an SOS sweep has a larger area on the four polynomial systems.

Critical strengths
  1. The question is a sensible one. Certifying the bound over a whole interval of disturbance levels, rather than at a few budgets, is useful, and the paper explains clearly why interpolating pointwise certificates is not sound (both the disturbance set and the shell V≥cV \ge c change with ωˉ\bar\omega). Theorem 1, Lemma 1, and Theorem 2 are correct as far as I checked. They are standard ultimate-boundedness arguments, but they are stated carefully.

  2. The accounting identity in Eq. (21), which separates proposal, repair, and tightening, is a reasonable diagnostic. The paper is also careful not to treat the pointwise BaB result as optimal.

Critical weaknesses
  1. Some reported certificates appear challenging under the setup as written. Appendix A gives a single setting for all benchmarks: cmax=1.0c_{max}=1.0, κ=0.1\kappa=0.1, the state box [−2,2]n[-2,2]^n (after the equilibrium shift), and ωˉmax=0.5\bar\omega_{max}=0.5, followed by "Every method uses the same values for a given row", which leaves open whether the values differ by system. If these values apply and the box is centered on the shifted equilibrium:
  • Bistable Duffing: the stable well at (1,0)(1,0) disappears for ∣w∣>2/(33)≈0.385|w|>2/(3\sqrt3)\approx0.385. In simulation, a constant w=−0.36w=-0.36 moves the state to the other well (q≈−1.15q\approx-1.15), and a bounded disturbance ∣w∣≤0.19|w|\le0.19 pushing along the velocity reaches ∣q−1∣≈2.4|q-1|\approx2.4. Both leave the box.
  • FitzHugh–Nagumo: the rest point loses stability at w≈0.041w\approx0.041 (Hopf). A constant w=0.05w=0.05 gives a sustained limit cycle with ∣v−v∗∣≈2.8|v-v^*|\approx2.8, outside the box.

A certificate at budget ωˉ\bar\omega (Theorem 1) implies that AcmaxA_{c_{max}}, which contains the equilibrium, is forward invariant inside the box. Valid certificates therefore cannot exist beyond roughly 0.38 ωˉmax0.38\,\bar\omega_{max} for Duffing or 0.1 ωˉmax0.1\,\bar\omega_{max} for FitzHugh–Nagumo, yet Table 1 and Fig. 4 report smooth full-range curves for both. The learned-FHN and coupled-FHN experiments may face the same issue, but their models are not specified. Either the appendix does not describe what was run, or the verifier is unsound. VV is also never specified (its form, how it was obtained, α1\alpha_1), so none of the results can be checked.

  1. The numbers are unusually regular, and no variability is reported. The gap between uniform neural and GAIN-Cert is 0.044–0.045 on all six ODEs and on both learned systems. The direct affine-level minus uniform neural gap is 0.004–0.005 everywhere, and GAIN-Cert/staircase is 0.84–0.85 throughout. This holds across systems from a 2D linear ODE to a 4D cart-pole with very different nonlinearities. The curves in Fig. 4 look nearly identical across systems, with no feature where Duffing loses its well or FitzHugh–Nagumo starts oscillating. There are no seeds, standard deviations, or raw results. Table 2 is also unclear: "uniform neural" as described in Sec. 4 (same anchors, capacity, knot count, and verifier, uniform knots) reads like "Full pipeline, uniform knots", yet they get 0.401 and 0.373; and "uniform parametric level curve" (0.413) is presented as one of the earlier baselines but matches none of the Table 1 values. Together with point 1, I cannot treat the experimental evidence as reliable.
Questions for authors
  1. What are the actual ωˉmax\bar\omega_{max}, cmaxc_{max}, κ\kappa, and state box for each system, and is the box centered on the shifted equilibrium? If ωˉmax=0.5\bar\omega_{max}=0.5 for Duffing and FitzHugh–Nagumo, how do the certificates hold given the disturbances described above? Please release the certified VV and the final level knots.
  2. What is VV for each system, how was it obtained, and what α1\alpha_1 is used in the normalization?
  3. Which verifier is used, and how does it bound the cubic and trigonometric terms over state × disturbance boxes?
  4. Please report per-run variability (seeds) and explain why the gap to uniform neural is 0.044–0.045 on all eight settings.
  5. How do "uniform neural" and "full pipeline, uniform knots" differ, and what exactly is "uniform parametric level curve"?
  6. The final certified object is a piecewise-affine level curve on the same knots as the direct affine-level baseline. If both are tightened to convergence, why does the neural initializer give about 12% less area?
Reviewer_NQxzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GAIN-Cert, a framework for certifying disturbance-to-state gain curves for nonlinear and learning-enabled dynamical systems. Unlike conventional approaches that consider individual disturbance budgets, the proposed method aims to establish continuous gain bounds over an entire disturbance interval. It integrates pointwise branch-and-bound verification, neural curve approximation, and joint state–disturbance certification, followed by certificate refinement. The paper also introduces a gap decomposition to analyze different sources of conservatism. Overall, the experimental results are promising.

Critical strengths

-The paper addresses a meaningful problem by extending disturbance-to-state certification from isolated disturbance budgets to a continuous disturbance range, providing a more comprehensive characterization of system robustness.

-The proposed framework integrates neural approximation with formal verification, and the experimental evaluation includes comparisons with several baselines and component ablation studies that demonstrate the effectiveness of the proposed approach.

-Overall, the paper is well written, with a relatively clear structure and an accessible presentation. However, the connections between different sections could be strengthened, particularly through clearer transitions, better explanations of how each component builds upon the preceding one, and more concise summaries of the key ideas and findings.

Critical weaknesses

-The main technical novelty of the proposed framework is not sufficiently clear. Theorem 1 follows from standard Lyapunov ultimate-boundedness arguments, while Theorem 2 is essentially a direct consequence of Theorem 1 once the joint verification condition is established. Similarly, the proposed framework combines several existing techniques, including pointwise BaB verification, monotone neural approximation, piecewise-affine interpolation, and counterexample-guided repair. Although their integration is useful, it remains unclear which component introduces a fundamentally new verification technique. In particular, the proposed gap decomposition in Eqs. (2) and (21) is primarily an algebraic accounting identity. It is not clear whether this decomposition leads to new theoretical bounds or directly guides the optimization procedure, beyond providing a diagnostic interpretation of the results. The authors should better distinguish the methodological contributions from standard Lyapunov analysis and existing learning-guided verification approaches.

-The necessity of the monotone integral neural network is not sufficiently justified. The framework first obtains pointwise certified bounds using BaB, fits a neural network to these values, and then converts the neural proposal into a piecewise-affine Lyapunov-level curve for formal verification. Since the final certified object is a piecewise-affine curve rather than the neural network itself, the neural component appears to serve primarily as an initialization heuristic. The paper includes a direct affine-level optimization baseline, which is useful, and the reported results favor GAIN-Cert. However, it remains unclear whether these improvements arise specifically from the neural representation or from differences in initialization, optimization, and local refinement. A more detailed comparison with conventional monotone interpolation or adaptive piecewise-affine optimization would help establish the necessity of the neural component.

-Although the experimental results show consistent improvements in normalized bound area, the main evaluation focuses on six relatively low-dimensional ODE systems. The additional learning-enabled experiments and dimension-scaling tests are mainly presented in the appendix. Since joint verification introduces an additional disturbance-budget parameter and couples the state region with the disturbance constraints, its computational scalability is an important concern. Moreover, the introduction motivates the problem using inverter-rich power systems and grid disturbances, but the evaluation does not include a corresponding power-system benchmark. The certification results for learned dynamics also apply to the learned vector field unless the modeling error is formally bounded. Consequently, the experiments provide limited evidence for the practical applications emphasized in the introduction.

Minor comments:

-Section 3.4 states that adjacent knot levels are increased until a validated counterexample is excluded or the derivative inequality is satisfied. However, the procedure for selecting which knots to update, determining the update magnitude, and propagating changes to preserve monotonicity is not sufficiently specified.

-Although Definition 1 introduces a certificate-dependent threshold H(ωˉ)H(\bar\omega), the relationship between this construction and existing ISS gain estimation or Lyapunov sublevel optimization techniques could be discussed more explicitly.

-There are minor notation and proofreading issues, including the repeated reference "Eq. Eq. (2)" in Section 3.1. The notation distinguishing the exact verification objective from its computed upper bound in Eq. (19) should also be made clearer.

-The abstract could be more concise. In particular, the method description would benefit from a more compact summary of the overall pipeline, emphasizing the logical relationships between its components rather than detailing each individual step. The technical details can be elaborated in the methodology section. Similarly, the experimental findings could be summarized around the main takeaways instead of listing too many individual results.

-In addition, introducing formal equations in the Introduction may interrupt the narrative flow; the relevant definitions could be moved to Section 2. The logical transition around Lines 67–68 is also unclear. In particular, the motivation for introducing a parameterized form does not seem to follow naturally from the preceding discussion and should be better explained.

-The contributions are currently presented as separate titled items, and their descriptions read more like a list of technical components than a synthesis of the paper's methodological advances. The authors could improve this section by emphasizing the overarching methodological insights and explaining how the individual components collectively address the target problem, rather than primarily enumerating implementation-level techniques.

-The methodology section would benefit from an overview of the overall framework at the beginning. A workflow diagram accompanied by a concise explanation of the main stages and their dependencies would help readers understand how pointwise verification, neural curve fitting, adaptive knot placement, interval certification, and subsequent refinement fit together before proceeding to the technical details.

-The experimental section relies heavily on line plots, which provide useful information about curve shapes but may not fully convey the overall performance and trade-offs of the proposed method. Additional forms of visualization, such as more informative comparative summaries or plots of verification cost and scalability, could improve the presentation. Furthermore, the experimental evaluation could be strengthened with more extensive results, particularly on higher-dimensional systems and more challenging learning-enabled benchmarks, to provide stronger evidence of the method's effectiveness and scalability.

Questions for authors

i. Could the authors clarify the main technical novelty of GAIN-Cert compared with existing ISS gain estimation and parameterized Lyapunov verification methods? In particular, what new verification capability or theoretical insight is introduced beyond combining existing techniques?

ii. How does the proposed gap decomposition directly influence the synthesis algorithm? Is it used to guide optimization, knot placement, or verification, or is it primarily a post-hoc diagnostic tool?

iii. Given that the final certificate is represented by a piecewise-affine Lyapunov-level curve, what specific advantage does the monotone integral neural network provide over direct monotone interpolation or piecewise-affine optimization? Could the authors further clarify the differences between the neural and direct optimization baselines?

iv. How does the computational cost of joint state–disturbance verification scale with state dimension, disturbance dimension, and the number of knots? What are the main computational bottlenecks observed in the learning-enabled benchmarks?

v. The AI use statement indicates that generative AI was used to "brainstorm research ideas" and "draft portions of the manuscript." Could the authors clarify the extent of AI involvement in developing the core research ideas and methodology? In particular, did AI assistance extend beyond general brainstorming to the design of the proposed framework or its theoretical components?

Reviewer_BKmmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a new method for certifying robustness to disturbances. The main focus of this work is in the learning of functions that approximate sublevel sets of a given Lyapunov function (for a stable dynamical system) for different disturbance magnitudes. The key idea here is to use a neural network to interpolate between branch and bound verified points, and then use that to obtain a superior disturbance gain value. Numerical experiments are provided showing the use of this approach on some standard toy examples.

Critical strengths
  • The paper identifies and addresses an interesting problem in the use of neural lyapunov controllers, and the proposed approach (what I understood of it) might be promising.
Critical weaknesses

I feel the paper has numerous critical weaknesses.

  • The writing feels very disjointed and imprecise. As a result, it is very hard to follow along with the paper. Words like "knots" are not defined ever in the paper. In equation 3, ζ\zeta is never formally defined. The dynamics are never properly discussed- is the disturbance additive or multiplicative? The algorithm used is also never specified, adding to the lack of clarity.
  • Neither the Lyapunov function nor the verifier is ever described or stated formally anywhere in this work. This is a problem since the contributions explicitly state that the bounds derived are "verifier depenent".
  • The "repair" stage is imprecisely stated. When is repair sufficient? When is repair impossible?
  • I feel that ICLR is perhaps a poor fit for this work as well. The only learning aspect of this work is to fit a monotone neural network to 1D γ\gamma obtained using the thresholds computed with branch and bound approaches. Perhaps a dedicated controls conference would be a better fit for this work.
Questions for authors
  • Why use a monotone neural network for the interpolant, especially given that the data is 1 dimensional?
  • When does the repair procedure fail?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.