Pith. sign in

REVIEW 2 major objections 5 minor 51 references

Certifying Plans under Model Mismatch: A Trilemma for Reachability from Scarce Data

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Sparse target data create a safety-certificate trilemma: soundness, finite width, and unrestricted model error cannot hold together once a plan leaves observed regions.

desk verdict A genuinely new trilemma for certifying fixed plans from sparse one-step data, wrapped in a sound method whose practical guarantee rests entirely on a Lipschitz declaration the data can refute but not confirm. read the letter →

arxiv 2608.02453 v1 pith:BKAK6OKW submitted 2026-08-03 cs.RO

Yanliang Huang , Zhen Zhang , Ahmad Hafez , Wenyuan Wu , Peng Xie , Zhuoqi Zeng , Amr Alanwar This is my paper · ORCID
classification cs.RO
keywords reachabilityanalysispre-executioncertificationmodelmismatchLipschitzboundset-membershipidentificationzonotopessim-to-realtransferscarcedata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a robot can certify a fixed control sequence as safe before executing it, when the only target-system data are a handful of isolated one-step transitions. It proves a trilemma: no deterministic certifier can simultaneously be sound for every target system consistent with those observations, return a reachable tube of finite projected width, and leave the model error unrestricted away from the observed locations. Bounding the model error to a smooth class does not dissolve the tension; it only makes the forced tube width finite, with a lower bound that depends on the plan and on how local model-error differences are amplified along the trajectory. The paper then gives a certifier, conditional on a supplied componentwise Lipschitz bound on the model error, that constructs a set-membership envelope and a zonotopic tube, and abstains when the tube exits the certification domain or cannot prove obstacle clearance.

What carries the argument

The load-bearing construction is a pair of data-indistinguishable residuals: a smooth bump perturbation localized at an unobserved plan query, added or subtracted to a baseline residual. Because both residuals agree at every sampled location and share the same noise realization, no certifier can tell them apart, yet their one-step state separation is 2a e_k. The terminal effect of that separation is propagated by the product of Jacobians G_{T,τ} = A_{T-1}⋯A_{τ+1}, giving the lower bound ρ_P(X_T) ≥ a∥P G_{T,τ} e_k∥₂ − C a². The practical side uses the standard Lipschitz-extension bound and set-membership envelopes to bound the residual over each queried zonotope, then propagates the nominal m

What would settle it

Find one admissible deterministic certifier that, on the paper's two-system construction, returns a tube with projected half-width smaller than the lower bound of Theorem 2 while staying sound over the bounded smooth class — that would refute the trilemma. Alternatively, run ForeReach with L certified by an independent method and observe a single target rollout that leaves the returned tube; that would falsify the containment theorem as an engineering guarantee.

Watch

Extended reading notes

Core claim

The central claim is that, given only N isolated one-step transitions with bounded observation noise, a pre-execution certificate for a fixed plan is informationally limited: once the plan rolls through a state-input region with no nearby observation, two smooth target systems can share the exact same dataset yet separate along the plan, so any certifier that is uniformly sound for all data-consistent systems must either decline to certify or report a reachable tube whose projected width is forced from below. For unrestricted model error the forced width is unbounded; for bounded smooth classes the paper derives a finite plan-dependent lower bound involving the product of Jacobians along the

Load-bearing premise

The guarantee collapses if the supplied componentwise Lipschitz bound L is false anywhere the plan propagates; observed transitions can refute L only when a violating pair happens to be sampled, not prove it in unobserved regions.

Editorial extensions

If this is right

  • Any certifier that is uniformly sound over all residuals consistent with the data must abstain once the plan enters an unobserved region, unless it is willing to return a tube of unbounded projected width.
  • With bounded smooth residuals, every sound tube must carry at least a plan-dependent projected half-width; for small a the width scales linearly with the amplitude of the unobservable perturbation.
  • A supplied Lipschitz bound on the residual is a necessary practical ingredient: it is the only side information that turns the impossibility into a computable containment guarantee.
  • Observations placed along the plan's own trajectory restore certification; observations elsewhere can leave the tube narrow but unsafe.
  • The method's guarantee is conditional: a false but unrefuted Lipschitz declaration can yield a false certificate, and the pairwise test may stay silent even when the declared bound is off by a large factor.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that closed-loop feedback, which makes future queries depend on the observed region, may weaken the force of the trilemma; the lower-bound argument relies on fixed open-loop queries.
  • The indistinguishability argument is not limited to deterministic certifiers: two data-identical systems induce identical output distributions, so the same two-system construction would force any randomized certifier to either abstain or be uninformative with high probability.
  • One testable extension is to vary the gap between a plan query and the nearest observation; Theorem 2 predicts the certified tube width grows with the gap through the class-membership and localization thresholds.
  • If the trilemma transfers to trajectory-level calibration methods, it would explain why such methods need distributional or structural assumptions beyond the data to certify worst-case safety.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies pre-execution certification of fixed control sequences under model mismatch when target data consist of only a few isolated one-step transitions. It proves an information-theoretic tension: if the nominal rollout reaches an unobserved state-input region, any deterministic certifier that is uniformly sound over all data-consistent residuals must either abstain or return a tube of arbitrarily large projected width (Theorem 1); imposing bounded smoothness on the residual turns this into a finite, plan-dependent lower bound on the terminal projected half-width (Theorem 2), summarized as a trilemma (Corollary 2). On the constructive side, ForeReach requires a supplied componentwise Lipschitz bound L on the residual, checks observed pairs for contradictions, builds set-membership envelopes over queried zonotopes, propagates them under the nominal model, and returns a certificate only if the tube stays in the certification domain and every safety projection avoids the unsafe set (Theorem 3). Experiments on point-mass and six-state bicycle systems, including action chunks from a learned policy, show that unsupported plans are declined while corridor-supported plans can be certified. Section 5.4 demonstrates that an incorrect L can pass the pairwise consistency test and yield a false certificate, and the paper is explicit that the method's guarantee is conditional on a valid supplied L.

Significance. If the results hold, the paper gives a clean boundary for certification from sparse one-step data: without additional structure beyond the observations, uniform soundness and finite informative width cannot coexist once a plan enters an unobserved region, and the required width is governed by how the plan propagates local model-error differences. The lower-bound construction is specific and parameter-free in the sense that the two adversarial residual systems are explicitly built from the data and class bounds; the affine case yields an exact width formula, and the supplement includes a numerical witness that reproduces the predicted separation. The paper is also honest about the conditional nature of ForeReach: Theorem 3 assumes a supplied Lipschitz bound L, and Proposition S3 states that finite one-step data cannot verify L outside the sampled locations. This makes the contribution solid but narrower than a data-only certificate: the practical guarantee is only as strong as the external declaration, and the paper's own Section 5.4 shows a false certificate under a wrong declaration.

major comments (2)
  1. [Assumption 1 and Section 5.4] The practical certification guarantee of ForeReach rests entirely on the external componentwise Lipschitz bound L. Proposition 1 and Proposition S3 show that the pairwise consistency test is one-sided: it can reject an exposed violation but cannot establish L away from the observed samples. Section 5.4 concretely shows that an L underestimating the residual regularity by a factor of 31.7 passes the noiseless pairwise test and leads to a false certificate. While the paper is transparent about this, the main-text claims of 'certified recall' and the 'sound-and-informative corner' in Figure 3 are only meaningful for users who can independently certify L. I request a prominent 'scope of guarantee' statement in Section 5 or the Conclusion, explicitly stating that the reported coverage and recall numbers are conditional on a valid supplied L and that no amount of one-step transition data can r
  2. [Theorems 1 and 2 / Corollary 2] The impossibility result is per-plan and applies only when the nominal rollout possesses a data-free neighborhood and, for Theorem 2, the single-transition localization and a nonzero projected gain. Corollary 2's wording 'No admissible deterministic certifier is at once uniformly sound over its residual class, ρ̄-informative, and unrestricted beyond the observed locations' can be misread as a global statement for all plans. Please make the quantifier explicit next to Corollary 2: for plans whose entire rollout remains covered by observations, a conditional certifier can be both sound and informative, as the corridor-supported experiments demonstrate. This distinction is implicit but should be stated directly to avoid overgeneralization.
minor comments (5)
  1. [Section 4, Eq. (11)] The envelope formula in Eq. (11) is a key ingredient; the text notes that it follows from McShane's extension and bounded-noise set membership. It would help to state explicitly that the envelope is exact over the query set given Assumption 1, not a conservative heuristic.
  2. [Theorem 2, Eq. (6)] The sentence 'where k maximizes ∥P G_{T,τ} e_k∥_2 and this gain is nonzero' should add that if all projected gains are zero, the lower bound is vacuous; otherwise the statement can be misread as always producing a positive bound.
  3. [Algorithm 1] The abstention labels ⊥F, ⊥D, and ⊥S are defined in the text but not in the algorithm or its caption. A one-line definition near the algorithm would improve readability.
  4. [Section 5, Figure 3] The comparison in Figure 3 and Table S5 gives ForeReach a side-information advantage: the baselines do not receive a certified Lipschitz bound L. This should be stated in the caption or text so the reader does not interpret the comparison as data-only.
  5. [Supplement, Proposition S3] The non-verification result is central to the practical caveat. Consider moving a short version of Proposition S3 into the main text near Assumption 1, since many readers may not consult the supplement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: trilemma bounds are derived from explicit class definitions, and ForeReach's guarantee is explicitly conditional on supplied Lipschitz side information.

full rationale

The lower-bound derivation is self-contained. Theorems 1 and 2 construct two data-consistent residual systems r(±)=r0 ± aφek from the consistency class C(D_N,σ) and the bounded-smooth class Csm, and the projected-width bound follows directly from the definitions of uniform soundness and projected half-width; no fitted constant or observed value is renamed as a prediction. ForeReach's containment guarantee is conditional on Assumption 1 (the supplied componentwise Lipschitz bound L) and the observation-noise bound σ. The set-membership envelope in Eq. (11) is the McShane/Milanese–Novara Lipschitz extension of that supplied pair, and the paper explicitly proves that the pairwise consistency test is one-sided: Proposition 1 states non-violation leaves the assumption unresolved away from the data, Proposition S3 establishes finite-data non-verification, and Section 5.4 demonstrates a false certificate under an underestimated declaration. Thus L is treated as side information/axiom, not as a parameter fitted to the target data and then presented as a derived prediction. Self-citations appear only in related-work positioning and baseline comparisons and are not load-bearing for the trilemma or the containment theorem. No circular step is present.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The central theoretical result depends only on explicit consistency-class definitions and derivative bounds, not on fitted parameters. The method's practical guarantee, however, rests on the user-supplied Lipschitz declaration L, which is an unverifiable input. No new physical entities are introduced.

free parameters (1)
  • Componentwise Lipschitz bound L
    User-supplied side information (Assumption 1). All ForeReach guarantees are conditional on L holding over Zcert. The data can refute it but cannot establish it outside sampled locations, so it functions as an unverified free input rather than a fitted parameter.
assumptions (6)
  • domain assumption Assumption 1: a componentwise Lipschitz bound L holds for the residual over the certification domain Zcert.
    The entire ForeReach envelope and containment proof are conditional on this bound. Section 4 and Proposition 1 state that data can only reject, never confirm, this declaration.
  • domain assumption Observation noise is deterministic and componentwise bounded by sigma; sample locations z_i are exact (Eq. 3 in Section 2.2).
    The pairwise consistency test and the set-membership envelope both rely on this noise model. If noise is unbounded or sample locations are noisy, the guarantees fail.
  • domain assumption Zcert is compact and convex, contains all observed locations, and f0 is twice continuously differentiable on the interval hull of each queried zonotope.
    The nominal-map enclosure (Proposition S1) requires finite interval Hessian bounds on Box(Zt), and the domain check in Algorithm 1 ensures queries stay inside Zcert.
  • domain assumption The target dynamics are deterministic and the plan is a fixed open-loop control sequence; no feedback law changes the applied inputs.
    The certification task in Section 2.2 is defined for a fixed sequence pi. Feedback policies or online replanning are outside the stated setup.
  • domain assumption For the bounded-smooth lower bound, the residual class Csm with amplitude bound Rmax and derivative bounds H1, H2, plus class slacks mu_j > 0, are assumed (Definition S9 and Eq. S33).
    Theorem 2 and Theorem S1 only apply inside this explicitly stated smooth class. The lower bound is finite only because these derivative bounds are imposed.
  • ad hoc to paper A data-free rollout neighborhood and the single-transition localization condition hold for the baseline residual r0 (Definitions S8 and S10).
    The adversarial bump construction requires an open ball around one plan query with no sample locations and no other baseline query. This is a theorem premise that may not hold for every dataset or plan.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Certifying Plans under Model Mismatch: A Trilemma for Reachability from Scarce Data." pith.science (2026). https://pith.science/paper/BKAK6OKW

@misc{pith2026260802453,
  author       = {Pith},
  title        = {Pith review of: Certifying Plans under Model Mismatch: A Trilemma for Reachability from Scarce Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKAK6OKW}},
  note         = {Machine review of arXiv:2608.02453}
}
read the original abstract

Sim-to-real policies are designed under nominal dynamics, but target-system trials may yield only a few isolated one-step transitions. We study pre-execution certification of a fixed control sequence, such as an action chunk produced by a learned policy. If the sequence reaches an unobserved state-input region, the observations remain consistent with target systems whose trajectories separate along it by an arbitrarily large amount. Any deterministic certifier sound for all of them must then decline to certify or return a reachable tube with arbitrarily large projected width. For bounded smooth classes of the target-nominal model error, we derive a finite plan-dependent projected-width lower bound. These results expose a trilemma among uniform trajectory containment, finite projected width, and unrestricted model-error behavior beyond the observations. ForeReach requires a supplied componentwise Lipschitz bound on the model error. Observed transition pairs can refute this declaration but cannot establish it outside the observed locations. Conditional on a valid declaration, our method constructs a set-membership envelope for the model error, propagates a zonotopic reachable tube, and certifies only when propagation remains within the certification domain and every projected tube slice avoids the unsafe set. In two benchmark systems, calibration baselines may remain narrow after losing trajectory containment outside data support, whereas our method declines to certify unsupported sequences and recovers certification when relevant target data and sufficient obstacle clearance are available.

Figures

Figures reproduced from arXiv: 2608.02453 by the authors.

Figure 1
Figure 1. Finite observations admit data-consistent residuals that separate at an unobserved plan query. With a supplied Lipschitz [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Each point pairs complete-tube coverage on un [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 2
Figure 2. Certification geometry in position space. Panels (a) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith

  1. [1]

    Conformal Prediction Beyond Exchangeability , journal =

    Barber, Rina Foygel and Cand. Conformal Prediction Beyond Exchangeability , journal =. 2023 , doi =

    2023

  2. [2]

    and Barber, Rina Foygel and Cand

    Tibshirani, Ryan J. and Barber, Rina Foygel and Cand. Conformal Prediction Under Covariate Shift , booktitle =. 2019 , url =

    2019

  3. [3]

    Adaptive Conformal Inference Under Distribution Shift , booktitle =

    Gibbs, Isaac and Cand. Adaptive Conformal Inference Under Distribution Shift , booktitle =. 2021 , url =

    2021

  4. [4]

    Biometrika , volume =

    Guan, Leying , title =. Biometrika , volume =. 2023 , doi =

    2023

  5. [5]

    Electronic Journal of Statistics , volume =

    Bhattacharyya, Aabesh and Barber, Rina Foygel , title =. Electronic Journal of Statistics , volume =. 2026 , doi =

    2026

  6. [6]

    and Listgarten, Jennifer and Jordan, Michael I

    Fannjiang, Clara and Bates, Stephen and Angelopoulos, Anastasios N. and Listgarten, Jennifer and Jordan, Michael I. , title =. Proceedings of the National Academy of Sciences , volume =. 2022 , doi =

    2022

  7. [7]

    Data-Driven Reachability Analysis Using Matrix Zonotopes , booktitle =

    Alanwar, Amr and Koch, Anne and Allg. Data-Driven Reachability Analysis Using Matrix Zonotopes , booktitle =. 2021 , editor =

    2021

  8. [8]

    Data-Driven Reachability Analysis From Noisy Data , journal =

    Alanwar, Amr and Koch, Anne and Allg. Data-Driven Reachability Analysis From Noisy Data , journal =. 2023 , doi =

    2023

Show all 51 references
  1. [9]

    Reachset-Conformant System Identification , journal =

    L. Reachset-Conformant System Identification , journal =. 2026 , doi =

    2026

  2. [10]

    Nonlinear Analysis: Hybrid Systems , volume =

    Wang, Shuqi and Feng, Mingyang and Chen, Yu and Gao, Yue and Yin, Xiang , title =. Nonlinear Analysis: Hybrid Systems , volume =. 2026 , doi =

    2026

  3. [11]

    Proceedings of The 8th Annual Learning for Dynamics and Control Conference , series =

    Srinivasan, Anutam and Leeman, Antoine and Chou, Glen , title =. Proceedings of The 8th Annual Learning for Dynamics and Control Conference , series =. 2026 , publisher =

    2026

  4. [12]

    Proceedings of The 8th Annual Learning for Dynamics and Control Conference , series =

    Nath, Devesh and Yin, Haoran and Chou, Glen , title =. Proceedings of The 8th Annual Learning for Dynamics and Control Conference , series =. 2026 , publisher =

    2026

  5. [13]

    and Isaacson, Seth and Zhang, Bohao and Ejakov, Matthew and Skinner, Katherine A

    Kwon, Yongseok and Michaux, Jonathan B. and Isaacson, Seth and Zhang, Bohao and Ejakov, Matthew and Skinner, Katherine A. and Vasudevan, Ram , title =. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2025 , doi =

    2025

  6. [14]

    Proceedings of Robotics: Science and Systems , address =

    Michaux, Jonathan and Li, Adam and Chen, Qingyi and Chen, Che and Vasudevan, Ram , title =. Proceedings of Robotics: Science and Systems , address =. 2024 , doi =

    2024

  7. [15]

    2026 , eprint =

    Huang, Yanliang and Zhang, Zhen and Xie, Peng and Zeng, Zhuoqi and Alanwar, Amr , title =. 2026 , eprint =

    2026

  8. [16]

    2026 , eprint =

    Huang, Yanliang and Xie, Peng and Wu, Wenyuan and Zeng, Zhuoqi and Alanwar, Amr , title =. 2026 , eprint =

    2026

  9. [17]

    2026 , eprint =

    Zhang, Zhen and Xie, Peng and Wu, Wenyuan and Huang, Yanliang and Alanwar, Amr , title =. 2026 , eprint =

    2026

  10. [18]

    2026 , eprint =

    Zhang, Zhen and Hafez, Ahmad and Xie, Peng and Huang, Yanliang and Wu, Wenyuan and Alanwar, Amr , title =. 2026 , eprint =

    2026

  11. [19]

    and Schwager, Mac , title =

    Sun, Jiankai and Jiang, Yiqi and Qiu, Jianing and Nobel, Parth and Kochenderfer, Mykel J. and Schwager, Mac , title =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  12. [20]

    , title =

    Lindemann, Lars and Cleaveland, Matthew and Shim, Gihyun and Pappas, George J. , title =. IEEE Robotics and Automation Letters , volume =. 2023 , doi =

    2023

  13. [21]

    ARCH14-15

    Althoff, Matthias , title =. ARCH14-15. 1st and 2nd International Workshop on Applied veRification for Continuous and Hybrid Systems , editor =. 2015 , publisher =. doi:10.29007/zbkv , url =

    2015 doi

  14. [22]

    IEEE Robotics and Automation Letters , volume =

    Knuth, Craig and Chou, Glen and Ozay, Necmiye and Berenson, Dmitry , title =. IEEE Robotics and Automation Letters , volume =. 2021 , doi =

    2021

  15. [23]

    IEEE Robotics and Automation Letters , volume =

    Knuth, Craig and Chou, Glen and Ozay, Necmiye and Berenson, Dmitry , title =. IEEE Robotics and Automation Letters , volume =. 2022 , doi =

    2022

  16. [24]

    2021 60th IEEE Conference on Decision and Control (CDC) , pages =

    Chou, Glen and Ozay, Necmiye and Berenson, Dmitry , title =. 2021 60th IEEE Conference on Decision and Control (CDC) , pages =. 2021 , publisher =. doi:10.1109/CDC45484.2021.9683354 , note =

    2021

  17. [25]

    , title =

    Knuth, Craig and Chou, Glen and Reese, Jamie and Moore, Joseph L. , title =. 2023 IEEE International Conference on Robotics and Automation , pages =. 2023 , doi =

    2023

  18. [26]

    2025 IEEE 64th Conference on Decision and Control (CDC) , pages =

    Dietrich, Elizabeth and Devonport, Rosalyn and Tu, Stephen and Arcak, Murat , title =. 2025 IEEE 64th Conference on Decision and Control (CDC) , pages =. 2025 , doi =

    2025

  19. [27]

    Tony and Low, Mark G

    Cai, T. Tony and Low, Mark G. , title =. The Annals of Statistics , volume =. 2004 , doi =

    2004

  20. [28]

    The Limits of Distribution-Free Conditional Predictive Inference , journal =

    Barber, Rina Foygel and Cand. The Limits of Distribution-Free Conditional Predictive Inference , journal =. 2021 , doi =

    2021

  21. [29]

    2020 American Control Conference , pages =

    Jin, Zeyuan and Khajenejad, Mohammad and Yong, Sze Zheng , title =. 2020 American Control Conference , pages =. 2020 , doi =. 2003.11662 , archivePrefix =

    2020 arXiv

  22. [30]

    IEEE Control Systems Letters , volume =

    Jin, Zeyuan and Khajenejad, Mohammad and Yong, Sze Zheng , title =. IEEE Control Systems Letters , volume =. 2022 , doi =

    2022

  23. [31]

    Automatica , volume =

    Milanese, Mario and Novara, Carlo , title =. Automatica , volume =. 2004 , doi =

    2004

  24. [32]

    and Goubault, Eric and Putot, Sylvie and Topcu, Ufuk , title =

    Djeumou, Franck and Vinod, Abraham P. and Goubault, Eric and Putot, Sylvie and Topcu, Ufuk , title =. IEEE Transactions on Automatic Control , volume =. 2023 , doi =

    2023

  25. [33]

    and Kumar, Vikash and Levine, Sergey and Finn, Chelsea , title =

    Zhao, Tony Z. and Kumar, Vikash and Levine, Sergey and Finn, Chelsea , title =. Proceedings of Robotics: Science and Systems , address =. 2023 , doi =

    2023

  26. [34]

    Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin C. M. and Song, Shuran , title =. Proceedings of Robotics: Science and Systems , address =. 2023 , doi =

    2023

  27. [35]

    Proceedings of Robotics: Science and Systems , address =

    Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and Jakubczak, Szymon and Jones, Tim and Ke, Liyiming and Levine, Sergey and Li-Bell, Adrian and...

    2025

  28. [36]

    International Journal of Robust and Nonlinear Control , volume =

    Canale, Massimo and Fagiano, Lorenzo and Signorile, Maria Carmela , title =. International Journal of Robust and Nonlinear Control , volume =. 2014 , doi =

    2014

  29. [37]

    Bulletin of the American Mathematical Society , volume =

    McShane, Edward James , title =. Bulletin of the American Mathematical Society , volume =. 1934 , doi =

    1934

  30. [38]

    and Taha, Ahmad F

    Nugroho, Sebastian A. and Taha, Ahmad F. and Hoang, Vu , title =. IEEE Transactions on Automatic Control , volume =. 2022 , doi =. 2004.12061 , archivePrefix =

    2022 arXiv

  31. [39]

    , title =

    Fazlyab, Mahyar and Robey, Alexander and Hassani, Hamed and Morari, Manfred and Pappas, George J. , title =. Advances in Neural Information Processing Systems , volume =. 2019 , publisher =

    2019

  32. [40]

    Componentwise

    Manzano, Jos. Componentwise. IEEE Transactions on Automatic Control , volume =. 2021 , doi =

    2021

  33. [41]

    Robust Learning-Based

    Manzano, Jos. Robust Learning-Based. Automatica , volume =. 2020 , doi =

    2020

  34. [42]

    , title =

    Tsybakov, Alexandre B. , title =. 2009 , isbn =

    2009

  35. [43]

    and Wasilkowski, Grzegorz W

    Traub, Joseph F. and Wasilkowski, Grzegorz W. and Wo. Information-Based Complexity , publisher =. 1988 , isbn =

    1988

  36. [44]

    IEEE Transactions on Automatic Control , volume =

    Milanese, Mario and Tempo, Roberto , title =. IEEE Transactions on Automatic Control , volume =. 1985 , doi =

    1985

  37. [45]

    Automatica , volume =

    Deterministic error bounds for kernel-based learning techniques under bounded noise , author =. Automatica , volume =. 2021 , doi =

    2021

  38. [46]

    IEEE Transactions on Automatic Control , volume =

    Model Validation: A Connection Between Robust Control and Identification , author =. IEEE Transactions on Automatic Control , volume =. 1992 , doi =

    1992

  39. [47]

    Automatica , volume =

    Lazily Adapted Constant Kinky Inference for nonparametric regression and model-reference adaptive control , author =. Automatica , volume =. 2020 , doi =

    2020

  40. [48]

    2000 , doi =

    Global Optimization with Non-Convex Constraints: Sequential and Parallel Algorithms , author =. 2000 , doi =

    2000

  41. [49]

    IEEE Transactions on Automatic Control , volume =

    A General Safety Framework for Learning-Based Control in Uncertain Robotic Systems , author =. IEEE Transactions on Automatic Control , volume =. 2019 , doi =

    2019

  42. [50]

    2018 IEEE Conference on Decision and Control (CDC) , pages =

    Learning-Based Model Predictive Control for Safe Exploration , author =. 2018 IEEE Conference on Decision and Control (CDC) , pages =. 2018 , doi =

    2018

  43. [51]

    Rigorously Computed Orbits of Dynamical Systems without the Wrapping Effect , journal =

    K. Rigorously Computed Orbits of Dynamical Systems without the Wrapping Effect , journal =. 1998 , doi =

    1998

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.