Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Set-based value functions on compact sets exactly mark the domain of stabilization for input-constrained discrete-time systems, and physics-informed networks learn them without control-infima.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-07-13 15:12 UTC pith:MAGT5BBP

load-bearing objection Clean set-based Zubov extension to input-constrained discrete systems with a usable PINN pipeline; Assumption 1 is load-bearing but the rest holds. the 4 major comments →

arxiv 2604.00305 v1 pith:MAGT5BBP submitted 2026-03-31 eess.SY cs.NEmath.DSmath.OC

Set-Based Value Function Characterization and Neural Approximation of Stabilization Domains for Input-Constrained Discrete-Time Systems

Mohamed Serry , S. Sivaranjani , Jun Liu This is my paper
classification eess.SY cs.NEmath.DSmath.OC
keywords domain of stabilizationcontrolled invariant setset-based value functionBellman equationphysics-informed neural networkdiscrete-time nonlinear systemsinput constraintscontrol Lyapunov function
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

For discrete-time nonlinear systems with input constraints, knowing which initial states can be driven to a target controlled invariant set (the domain of stabilization) is hard: classical value functions force expensive dynamic programming, while most neural Lyapunov methods give only local or non-maximal estimates. This paper introduces value functions V and W that take compact sets, not points, as arguments. Their finite (respectively sub-1) sublevel sets on singletons recover exactly the domain of stabilization whenever the target is locally stabilizable. Because the functions live on sets, the Bellman-type equations they satisfy contain no explicit minimization over controls; the reachable-set operator already encodes the possible inputs. That structure lets the equations be written as residual losses and trained with physics-informed neural networks. On two low-dimensional examples the learned networks produce visibly larger certified domains than quadratic Lyapunov ellipsoids and yield simple feedback laws that keep trajectories inside the estimated domain while respecting the input bounds. The result matters because it turns an abstract set-theoretic characterization into a practical, data-and-equation-driven estimator that can be larger than classical certificates yet still tied to a rigorous optimality principle.

Core claim

The domain of stabilization of a locally ℓ_p-stabilizable controlled invariant set A equals the set of singletons on which the newly defined set-based value functions V and W remain finite (respectively strictly less than 1). These functions satisfy Bellman-type functional equations free of control-infima, which can be embedded directly into a physics-informed neural loss to produce accurate domain estimates and stabilizing controllers.

What carries the argument

Set-based value functions V(X) = sum_k Ψ(R(X,k)) and W = 1-exp(-V), defined on the metric space of compact subsets via the reachable-set map R; they turn the domain of stabilization into ordinary sublevel sets and yield infimum-free Bellman equations usable as training residuals.

Load-bearing premise

The target set must be locally stabilizable with a summable decay envelope; without that local guarantee the infinite sum that defines the value function can diverge even for states that actually belong to the true domain of stabilization.

What would settle it

On either numerical example, compute or tightly over-approximate the true domain of stabilization by exhaustive gridding or formal reachability; if the neural 0.97-sublevel set of the learned W is either substantially larger than that true domain or fails to admit a stabilizing feedback for some interior point, the claim that the learned network recovers the domain is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Domain-of-stabilization estimates larger than classical quadratic Lyapunov ellipsoids become available for the same systems.
  • Stabilizing feedback can be synthesized by simple grid search on the learned value function without solving a separate optimal-control problem.
  • The same residual-loss construction applies to any discrete-time system whose reachable sets admit a finite-dimensional embedding (hyper-rectangles, zonotopes, etc.).
  • Future formal verification of the learned networks would convert the estimates into certified maximal domains of stabilization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same set-based construction could be applied to continuous-time systems by replacing the reachable-set operator with a flow-pipe operator, provided a suitable local stabilizability assumption is retained.
  • Because the Bellman residual never evaluates an explicit min over controls, the method may scale better to high-dimensional input sets than classical dynamic-programming approaches.
  • If the finite-horizon trajectory sampling used for data generation is replaced by rigorous set-propagation tools, the training targets themselves become certified, closing the loop between learning and verification.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces set-based value functions V and W defined on the metric space of compact subsets of R^n that characterize the domain of stabilization (DOS) of a locally ℓ_p-stabilizable controlled invariant set A for input-constrained discrete-time nonlinear systems. Under Assumption 1, Theorem 2 proves V_∞ = W_1 = D_A. Continuity, monotonicity, positive-definiteness, and uniqueness of solutions to associated Bellman/Zubov-type equations (without explicit control infima) are established (Theorems 7–13). A physics-informed neural network is trained by embedding the residual of equation (9) together with finite-horizon trajectory-sampled approximations of W, and two low-dimensional examples illustrate larger DOS estimates than quadratic Lyapunov ellipsoids together with a hybrid stabilizing controller extracted from the learned network.

Significance. If the characterization and learning pipeline hold as claimed, the work offers a principled route to maximal DOS estimation for constrained discrete-time systems that avoids the explicit infimum over controls in the Bellman residual—an obstacle that has limited physics-informed learning for controlled systems. The set-based formulation, continuity/uniqueness analysis, and residual form free of argmin operators are genuine technical contributions relative to classical Zubov methods and recent neural Lyapunov work. The numerical examples give concrete evidence that the learned sublevel sets can substantially enlarge certified regions obtained from linear feedback plus quadratic Lyapunov analysis, and that a practical (if only partially certified) controller can be read off the network. The contribution is therefore of clear interest to the nonlinear control and learning-for-control communities, provided the approximation and certification gaps are addressed.

major comments (4)
  1. [Section V-A] Section V-A and the training loss: the data term uses ˜W_Ns built from finite trajectory samples (N_traj = 5000) of reachable sets. No error bound relating ˜R({x},k) to R({x},k), nor any analysis of how this bias propagates into the learned ω_nn or the claimed DOS estimate, is given. Because the central practical claim is accurate estimation of D_A via the learned value function, the uncontrolled approximation error is load-bearing; at minimum a quantitative discussion (or a conservative outer approximation of reachable sets) is needed.
  2. [Section VI-A] Section VI-A: the set D_nn and controller Π are obtained by grid search for thresholds ω_1, ω_2 and for decrease of ω_nn (or ν) on a finite grid. Only the inner ellipsoid E_c1 carries a Lyapunov certificate; outside it the decrease condition is numerical and non-formal. The abstract and conclusion state that the method “synthesizes stabilizing controllers,” which overstates what is rigorously guaranteed. The manuscript should clearly separate certified inner regions from heuristic outer enlargement, or supply formal verification of the decrease condition on D_nn.
  3. [Assumption 1 / Theorem 2] Assumption 1 (local ℓ_p-stabilizability) is invoked throughout Theorems 2, 7, 8 and the uniqueness arguments (Theorems 12–13). Without a summable decay envelope near A the series defining V may diverge inside the true DOS, so the exact characterization V_∞ = D_A fails. The paper treats the assumption as given and verifies it only by construction in the two examples (linear feedback + quadratic region). A sharper discussion of when the assumption holds for general nonlinear systems, and of the consequences of its violation, is required for the claimed generality.
  4. [Section IV-B] Theorems 9, 10, 12 and 13 are reduced to corresponding statements in the authors’ prior arXiv [23] with only brief sketches. While the reduction is explicit, a journal version should either make the controlled-system arguments self-contained or isolate precisely which steps are new (the set-valued F, the embedding T, and the residual free of an explicit infimum) versus inherited, so that the incremental contribution can be assessed independently of [23].
minor comments (5)
  1. [Section VI-C / Figures 1–2] Figures 1–2 are described in the text but the rendered plots in the manuscript source are largely unreadable (placeholder glyphs). Ensure high-resolution, labeled axes, and a clear legend distinguishing the NN estimate, the ellipsoidal estimate, and sample trajectories.
  2. [Section I] The organization paragraph in the Introduction refers to Section VI for numerical examples and Section VII for the conclusion, which matches the body; however the intermediate section numbering (value functions → DOS estimation → examples) could be stated more cleanly for the reader.
  3. [Section II] Notation for the asymmetric/symmetric Hausdorff distances d_a_H and d_s_H is introduced by reference to [23], [26] without a self-contained definition; a one-line definition would improve readability.
  4. [Section VI] Related neural Lyapunov / DOS estimation works [11]–[14] are cited but never used as numerical baselines. Even a brief qualitative comparison on the same two examples would strengthen the experimental section.
  5. [Section VI] Typographical inconsistencies appear (e.g., “th CIS”, “N step” vs N_s, mixed use of W_r and D_nn). A careful copy-edit pass is needed.

Circularity Check

2 steps flagged

Modest load-bearing self-citations to concurrent arXiv [23] for Bellman equations, continuity lemmas, and set embeddings; core DOS characterization (Thm 2) is independently proved under Assumption 1.

specific steps
  1. self citation load bearing [Section IV-B, Theorems 9 and 10]
    "Theorem 9: V satisfies the equation (w.r.t. to the function v) v(X)=Ψ(X)+v(F(X)), X∈K(R^n). (8) Proof: See the proof of Theorem 18 in [23]. Theorem 10: For X∈K(R^n), W satisfies the equations ... Proof: See the proofs of Theorems 19 and 20 in [23]."

    The paper claims to 'derive the associated Bellman-type (Zubov-type) functional equations' that are then embedded as the physics-informed residual J_pi in the NN loss. The actual derivations are wholly deferred to the authors' concurrent arXiv [23] (uncontrolled systems). These equations are load-bearing for the learning method and for the uniqueness arguments that follow; without them the PINN pipeline has no governing residual.

  2. self citation load bearing [Lemma 6; Assumption 2 / Section V-B]
    "Lemma 6: ... whose proof is a trivial extension of the proof of Lemma 15 in [23]. ... Assume there exists a mapping T:S→R^L ... as discussed thoroughly in Remark 6 of [23]."

    Continuity of the set-valued infimum functional Ψ (used for continuity of V) and the injective finite-dimensional embedding T of singleton and one-step reachable sets (required to feed F({x}) into a standard NN) are imported from the same authors' prior work without independent derivation. Both are prerequisites for the claimed physics-informed training procedure.

full rationale

The paper's central characterization (Theorem 2: V_∞ = W_1 = D_A) is fully proved from the definitions of V/W, reachable-set properties (Lemma 1), and local ℓ_p-stabilizability (Assumption 1), without circular reduction. Positive-definiteness, monotonicity, continuity of V on K_{D_A}, blow-up outside D_A, and uniqueness of solutions to the functional equations (Theorems 7–8, 12–13, Lemma 11) likewise contain self-contained arguments. However, the Bellman/Zubov equations that are embedded into the physics-informed loss (Theorems 9–10), the continuity of the infimum map (Lemma 6), and the finite-dimensional embedding T of F({x}) (Assumption 2) are deferred entirely to the authors' concurrent arXiv [23] on the uncontrolled case. These imported pieces are load-bearing for the NN training pipeline and controller extraction, producing modest residual circularity of the self-citation type. No self-definitional loop, no fitted parameter re-labeled as prediction, and the numerical examples merely illustrate the learned approximation rather than redefine the DOS. Score 3 reflects that the main theoretical claim stands independently while the learning architecture leans on unverified-in-this-paper self-citations.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

The central characterization rests on one domain-specific stabilizability assumption, standard metric-space facts, and two invented functional objects (V and W) whose properties are derived rather than postulated. Free parameters appear only in the numerical learning stage and do not affect the theoretical claim.

free parameters (4)
  • N_traj (number of sampled input trajectories)
    Chosen by hand (5000) to approximate reachable sets; accuracy of Ψ and therefore of the data-driven loss depends on this choice.
  • N_s / N_step (truncation horizon)
    Finite-horizon cutoff (30) used to generate training targets; not derived from theory.
  • λ_d, λ_pi (loss weights)
    Hand-tuned (0.1 and 1) to balance data and physics residuals.
  • ω_1, ω_2 (sublevel thresholds for controller extraction)
    Grid-searched scalars that define the practical neural DOS; not uniquely determined by the value function.
axioms (3)
  • domain assumption Assumption 1: the controlled invariant set A is locally ℓ_p-stabilizable with continuous non-decreasing decay envelope λ whose p-power series converges.
    Invoked to prove that V remains finite exactly on D_A (Theorem 2) and for continuity/uniqueness arguments.
  • ad hoc to paper Assumption 2: there exists an injective finite-dimensional embedding T of singletons and one-step reachable sets into R^L.
    Required to turn the set-valued Bellman residual into a standard NN loss; satisfied for hyper-rectangles but not automatic for arbitrary dynamics.
  • standard math Hausdorff continuity of the reachable-set map F and of the infimum functional Ψ (Lemmas 1, 6).
    Standard facts from set-valued analysis used throughout continuity proofs.
invented entities (1)
  • Set-based value functions V and W on K(R^n) no independent evidence
    purpose: Provide an exact characterization of the DOS via sublevel sets and supply Bellman equations free of explicit control minimization.
    Defined by infinite sums of Ψ(R(X,k)); their properties are proved rather than assumed, but they are new objects introduced for this purpose.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Set-Based Value Function Characterization and Neural Approximation of Stabilization Domains for Input-Constrained Discrete-Time Systems." pith.science (2026). https://pith.science/paper/MAGT5BBP

@misc{pith2026260400305,
  author       = {Pith},
  title        = {Pith review of: Set-Based Value Function Characterization and Neural Approximation of Stabilization Domains for Input-Constrained Discrete-Time Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MAGT5BBP}},
  note         = {Machine review of arXiv:2604.00305}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Analyzing nonlinear systems with stabilizable controlled invariant sets (CISs) requires accurate estimation of their domains of stabilization (DOS) together with associated stabilizing controllers. Despite extensive research, estimating DOSs for general nonlinear systems remains challenging due to fundamental theoretical and computational limitations. In this paper, we propose a novel framework for estimating DOSs for controlled input-constrained discrete-time systems. The DOS is characterized via newly introduced value functions defined on metric spaces of compact sets. We establish the fundamental properties of these value functions and derive the associated Bellman-type (Zubov-type) functional equations. Building on this characterization, we develop a physics-informed neural network (NN) framework that learns the value functions by embedding the derived functional equations directly into the training process. The proposed methodology is demonstrated through two numerical examples, illustrating its ability to accurately estimate DOSs and synthesize stabilizing controllers from the learned value functions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

  1. [1]

    Computing control Lyapunov functions via a Zubov type algorithm,

    L. Grune and F. Wurth, “Computing control Lyapunov functions via a Zubov type algorithm,” inProc. of CDC, vol. 3, pp. 2129–2134, 2000

    2000

  2. [2]

    Formal verification of control Lyapunov-barrier func- tions for safe stabilization with bounded controls,

    J. Liu, “Formal verification of control Lyapunov-barrier func- tions for safe stabilization with bounded controls,”arXiv preprint arXiv:2511.10510, 2025

    arXiv 2025

  3. [3]

    Controllability, realization and stability of discrete-time systems,

    L. Weiss, “Controllability, realization and stability of discrete-time systems,”SIAM Journal on Control, vol. 10, no. 2, pp. 230–251, 1972

    1972

  4. [4]

    Discrete-time asymptotic controlla- bility implies smooth control-Lyapunov function,

    C. M. Kellett and A. R. Teel, “Discrete-time asymptotic controlla- bility implies smooth control-Lyapunov function,”Systems & Control Letters, vol. 52, no. 5, pp. 349–359, 2004

    2004

  5. [5]

    Control synthesis for discrete-time nonlinear control sys- tems under state and input constraints,

    J.-L. Wu, “Control synthesis for discrete-time nonlinear control sys- tems under state and input constraints,”IEEE Transactions on Auto- matic Control, vol. 69, no. 12, pp. 8418–8432, 2024

    2024

  6. [6]

    Sliding mode control of discrete- time systems,

    A. J. Koshkouei and A. Zinober, “Sliding mode control of discrete- time systems,”J. Dyn. Sys., Meas., Control, vol. 122, no. 4, pp. 793– 802, 2000

    2000

  7. [7]

    Feedback linearization of discrete-time systems,

    B. Jakubczyk, “Feedback linearization of discrete-time systems,”Sys- tems & Control Letters, vol. 9, no. 5, pp. 411–416, 1987

    1987

  8. [8]

    Linearization of discrete-time systems,

    E. Aranda-Bricaire, ¨U. Kotta, and C. H. Moog, “Linearization of discrete-time systems,”SIAM Journal on Control and Optimization, vol. 34, no. 6, pp. 1999–2023, 1996

    1999

  9. [9]

    Nonlinear model predictive control,

    L. Gr ¨une and J. Pannek, “Nonlinear model predictive control,” in Nonlinear model predictive control: Theory and algorithms, pp. 45– 69, Springer, 2016. ���� ���� �������� �� ������ �� ���� ���� �������� �� ������ �� �� �� ���� ���� ���� ���� ���� �������� �� ������ ������ ������ ������ �� ������ �� �� ���� ���� ���� ���� ���� ���� �������� �� ������ Fig...

    2016

  10. [10]

    Approximately optimal nonlinear stabilization with preservation of the Lyapunov function property,

    L. Grune and O. Junge, “Approximately optimal nonlinear stabilization with preservation of the Lyapunov function property,” in2007 46th IEEE Conference on Decision and Control, pp. 702–707, IEEE, 2007

    2007

  11. [11]

    Lyapunov- stable neural-network control,

    H. Dai, B. Landry, L. Yang, M. Pavone, and R. Tedrake, “Lyapunov- stable neural-network control,”arXiv:2109.14152, 2021

    Pith/arXiv arXiv 2021

  12. [12]

    Neural Lyapunov control for discrete-time systems,

    J. Wu, A. Clark, Y . Kantaros, and Y . V orobeychik, “Neural Lyapunov control for discrete-time systems,”Advances in Neural Information Processing Systems, vol. 36, pp. 2939–2955, 2023

    2023

  13. [13]

    Neural Lyapunov control,

    Y .-C. Chang, N. Roohi, and S. Gao, “Neural Lyapunov control,” Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  14. [14]

    Neural Lyapunov control of unknown nonlinear systems with stability guarantees,

    R. Zhou, T. Quartz, H. De Sterck, and J. Liu, “Neural Lyapunov control of unknown nonlinear systems with stability guarantees,” Advances in Neural Information Processing Systems, 2022

    2022

  15. [15]

    Certified training with branch- and-bound: A case study on Lyapunov-stable neural control,

    Z. Shi, C.-J. Hsieh, and H. Zhang, “Certified training with branch- and-bound: A case study on Lyapunov-stable neural control,” arXiv:2411.18235, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    dreal: An smt solver for nonlinear theories over the reals,

    S. Gao, S. Kong, and E. M. Clarke, “dreal: An smt solver for nonlinear theories over the reals,” inProc. of CADE, pp. 208–214, Springer, 2013

    2013

  17. [17]

    Branch and bound for piecewise linear neural network verification,

    R. Bunel, J. Lu, I. Turkaslan, P. H. Torr, P. Kohli, and M. P. Kumar, “Branch and bound for piecewise linear neural network verification,” Journal of Machine Learning Research, vol. 21, no. 42, pp. 1–39, 2020

    2020

  18. [18]

    Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers,

    K. Xu, H. Zhang, S. Wang, Y . Wang, S. Jana, X. Lin, and C. J. Hsieh, “Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers,” inProc. of ICLR, 2021

    2021

  19. [19]

    Beta-crown: Efficient bound propagation with per-neuron split constraints for neural network robustness verification,

    S. Wang, H. Zhang, K. Xu, X. Lin, S. Jana, C.-J. Hsieh, and J. Z. Kolter, “Beta-crown: Efficient bound propagation with per-neuron split constraints for neural network robustness verification,”Advances in Neural Information Processing Systems, vol. 34, pp. 29909–29921, 2021

    2021

  20. [20]

    Complete verification via multi-neuron relaxation guided branch-and-bound,

    C. Ferrari, M. N. Mueller, N. Jovanovi ´c, and M. Vechev, “Complete verification via multi-neuron relaxation guided branch-and-bound,” in Proc. of ICLR, 2022

    2022

  21. [21]

    Scalable neural network verification with branch-and-bound inferred cutting planes,

    D. Zhou, C. Brix, G. A. Hanasusanto, and H. Zhang, “Scalable neural network verification with branch-and-bound inferred cutting planes,” inAdvances in Neural Information Processing Systems, 2024

    2024

  22. [22]

    Safe domains of attraction for discrete-time nonlinear systems: Characterization and verifiable neural network estimation,

    M. Serry, H. Li, R. Zhou, H. Zhang, and J. Liu, “Safe domains of attraction for discrete-time nonlinear systems: Characterization and verifiable neural network estimation,” inProc. of CDC, pp. 5774– 5781, IEEE, 2025

    2025

  23. [23]

    Safe and robust do- mains of attraction for discrete-time systems: A set-based charac- terization and certifiable neural network estimation,

    M. Serry, M. Fitzsimmons, and J. Liu, “Safe and robust do- mains of attraction for discrete-time systems: A set-based charac- terization and certifiable neural network estimation,”arXiv preprint arXiv:2603.03082, 2026

    arXiv 2026

  24. [24]

    Physics-informed neural network Lyapunov functions: PDE characterization, learning, and verification,

    J. Liu, Y . Meng, M. Fitzsimmons, and R. Zhou, “Physics-informed neural network Lyapunov functions: PDE characterization, learning, and verification,”Automatica, vol. 175, p. 112193, 2025

    2025

  25. [25]

    Formally verified physics-informed neural control Lyapunov functions,

    J. Liu, M. Fitzsimmons, R. Zhou, and Y . Meng, “Formally verified physics-informed neural control Lyapunov functions,” in2025 Amer- ican Control Conference (ACC), pp. 1347–1354, IEEE, 2025

    2025

  26. [26]

    Overapproximating reachable tubes of linear time-varying systems,

    M. Serry and G. Reissig, “Overapproximating reachable tubes of linear time-varying systems,”IEEE Transactions on Automatic Control, vol. 67, no. 1, pp. 443–450, 2021

    2021

  27. [27]

    Rigorously computed orbits of dynamical systems without the wrapping effect,

    W. K ¨uhn, “Rigorously computed orbits of dynamical systems without the wrapping effect,”Computing, vol. 61, no. 1, pp. 47–67, 1998

    1998

  28. [28]

    Guaranteed state estimation by zonotopes,

    T. Alamo, J. M. Bravo, and E. F. Camacho, “Guaranteed state estimation by zonotopes,”Automatica, vol. 41, no. 6, pp. 1035–1043, 2005

    2005

  29. [29]

    Accurate uncertainty propagation for discrete-time nonlinear systems using differential inequalities with model redundancy,

    X. Yang and J. K. Scott, “Accurate uncertainty propagation for discrete-time nonlinear systems using differential inequalities with model redundancy,”IEEE Transactions on Automatic Control, vol. 65, no. 12, pp. 5043–5057, 2020

    2020

  30. [30]

    An introduction to CORA 2015,

    M. Althoff, “An introduction to CORA 2015,” inProc. of the Workshop on Applied Verification for Continuous and Hybrid Systems, pp. 120– 151, 2015

    2015

This paper was first reviewed by grok-4.5 on July 13, 2026.