Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Characterization of Safety in Stochastic Difference Inclusions using Barrier Functions

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Concave barriers plus convex set-valued dynamics give a delta/Delta safety bound by replacing random inputs with their mean.

desk verdict A promising verification trick undermined by an overclaimed equality and examples that violate the paper's own convexity assumption; the main theorem may hold but needs repair. read the letter →

arxiv 2508.20204 v1 pith:7Z2V7ELV submitted 2025-08-27 math.OC cs.SYeess.SY

classification math.OCcs.SYeess.SY MSC 93E0393C55
keywords stochasticdifferenceinclusionsbarrierfunctionssafetyverificationconcaveconvexset-valuedmapssupermartingaleprobabilisticreachabilitymean-field
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

This paper studies discrete-time systems whose next state is chosen adversarially from a set G(x, v) that depends on the current state and a random input v. Its central claim is that when G has a convex graph and the barrier function B is concave, the difficult worst-case expectation over adversarial choices and random inputs can be replaced by a single check at the mean input: sup_{x+ in G(x, E[v])} B(x+) <= B(x). If B is at most delta on the initial set and at least Delta on the unsafe set, then the probability of ever reaching the unsafe set is at most delta/Delta, over any finite or infinite horizon. The value is that safety verification becomes a mean-field, convex-optimization problem rather than a dynamic game over distributions and adversarial picks.

What carries the argument

The load-bearing object is the pair (B, G): a concave, nonnegative, upper semicontinuous barrier function B together with a set-valued map G whose graph is convex. This combination guarantees that each worst-case barrier value lambda_k(x, v0, ..., v_{k-1}) = sup over G-compositions of B is concave in the random arguments. Concavity then allows Jensen's inequality to replace every random input v by its expectation E[v] in the supermartingale check, reducing the adversarial stochastic evolution to a deterministic mean-input system. The resulting condition sup_{x+ in G(x,E[v])} B(x+) <= B(x) is a convex verification problem and yields the closed-form probability bound delta/Delta.

What would settle it

Take the system of Example 37 with v uniformly distributed on [-0.2, 0.2] and B(X) = tr(X). Directly computing lambda_1(X, v) = tr(X) + 2 max(v, 0) shows it is convex in v, giving E[lambda_1(X, v)] = tr(X) + 0.1 > tr(X) = lambda_1(X, E[v]). Thus B satisfies the mean-input condition (25) but fails the true supermartingale inequality, contradicting Lemma 30 and Theorem 35 for this instance; simulating many trajectories and counting hits of the unsafe set would reveal whether the claimed delta/Delta = 0.25 bound actually holds or is violated.

Watch

Extended reading notes

Core claim

The paper establishes that, under Assumptions 4, 5, and 21 (upper semicontinuity, measurability, and convexity of the set-valued map's graph), a concave nonnegative barrier function B that satisfies the one-step mean-input condition sup_{x+ in G(x,E[v])} B(x+) <= B(x) on the safe set is automatically a supermartingale barrier. Consequently, if B <= delta on the initial set X0 and B >= Delta on the unsafe set Xu, the probability of reaching Xu from X0, regardless of adversarial choices and over any time horizon, is at most delta/Delta (Theorem 35). The key step is that concavity of B and convexity of G make the marginal value lambda_k concave in the random input, so Jensen's inequality permit

Load-bearing premise

The whole simplification depends on the set-valued map G having a convex graph; if that convexity fails, the worst-case barrier value is not guaranteed to be concave in the random input, so replacing the input by its mean is not justified and the supermartingale argument collapses.

Editorial extensions

If this is right

  • Safety verification for a broad class of stochastic set-valued systems reduces to checking one convex condition at the mean disturbance, instead of solving a dynamic game or computing high-dimensional expectations.
  • The delta/Delta bound holds for infinite time horizons whenever the mean-input decrease condition holds, so a single local check certifies safety forever.
  • The framework naturally covers distributional robustness: families of distributions parameterized by a set can be encoded in G, and the mean-substitution result applies to every member of the family simultaneously.
  • For linear or affine dynamics, the supermartingale condition becomes a set of linear inequalities, making the certificate computable by standard convex optimization.
  • The result unifies the deterministic barrier-certificate bound with the stochastic setting: the deterministic bound delta/Delta appears as the special case with no random input.

Reading between the lines

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

  • When the convexity assumption on G fails, the mean-substitution step can break: if lambda_1(x, v) is convex rather than concave in v, Jensen gives the reverse inequality, so a barrier can satisfy the mean-input check while failing the true supermartingale inequality.
  • A testable alternative to convex G is to verify concavity of lambda_k(x, v0, ..., v_{k-1}) in the random arguments directly; for systems with non-convex graphs but concave marginals, the same delta/Delta bound should still hold.
  • The delta/Delta bound is likely conservative; in practice, reach probabilities may be far smaller. A sharper bound might be obtained by tracking the actual distribution of the barrier process rather than only its expectation.
  • The mean-input reduction suggests a design principle for stochastic safety: to make a system safe, choose the set-valued dynamics so that its mean image contracts the level sets of a concave barrier, which is a convex-design problem.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies safety of discrete-time stochastic difference inclusions of the form x^+ ∈ G(x,v), where v is stochastic and an adversary chooses the next state from the set G(x,v). It develops barrier-function conditions for bounding the probability of reaching an unsafe set. Under an upper semicontinuity/measurability assumption on G and, crucially, Assumption 21 that G is a convex set-valued map, the paper argues that for a concave barrier B one may replace the random input v by its expectation in the one-step condition and obtain a supermartingale barrier certificate. The main result, Theorem 35, bounds the reach probability by δ/Δ under the deterministic condition sup_{x^+∈G(x,E[v])} B(x^+) ≤ B(x). Numerical examples on linear and matrix-valued dynamics are presented as illustrations.

Significance. If the main theorem is correct, the paper offers a clean and useful sufficient condition: for concave barriers and convex-graph set-valued maps, checking the barrier condition at the expected disturbance suffices for a supermartingale certificate. The proof strategy is simple and mostly self-contained, relying on Jensen's inequality and standard supermartingale arguments. However, the paper overstates the result: the expectation substitution is an inequality, not an equality, and the paper's two main numerical examples do not satisfy Assumption 21 and therefore do not validate the machinery. The central theorem is plausible and likely fixable, but the current presentation and examples are not reliable as stated.

major comments (3)
  1. [§4, Remark 24] The equality E[sup_{k∈[1,N]} λ_k(x,v_0,...,v_{k-1})] = λ_j(x,E[v_0],...,E[v_{j-1}]) is false in general. Concavity of λ_j gives only the Jensen upper bound E[λ_j] ≤ λ_j(E[v_0],...,E[v_{j-1}]); equality holds only if λ_j is affine or the random variables are degenerate. This is not a cosmetic issue: the simplification used in Remark 24, in condition (32), and in Example 37 relies on the equality. The correct sufficient condition should be an inequality, e.g., λ_j(x,E[v_0],...,E[v_{j-1}]) ≤ ρΔ, not an equality.
  2. [§6, Example 37] The map G(X,v)={UXU^T+M(γ)v | γ∈[0,1]} does not have a convex graph. Take X1=X2=I, v1=0.2, v2=-0.2, and θ=1/2. Then G(I,0)={I}, but (1/2)G(I,0.2)+(1/2)G(I,-0.2) contains I+0.1(M(γ1)-M(γ2)), which is not I for γ1=1, γ2=0. Hence Assumption 21 and Lemma 22 do not apply. In fact λ_1(X,v)=tr(X)+2max(v,0), which is convex in v, and for v∼U[-0.2,0.2] one has E[λ_1]=tr(X)+0.1>B(X), so B is not a supermartingale. Thus the claimed bound 1-δ/Δ and Figure 2 are not supported by Theorem 35; the true finite-horizon expectation is tr(X)+0.1N, not tr(X).
  3. [§6, Example 36] The assertion that the dynamics x^+∈{(γA1+(1-γ)A2)x+bv} are convex because they are linear in (x,v) is incorrect. A set-valued map that is affine in the parameter γ need not have a convex graph. For the matrices in the example, take C=A1-A2, x1=(1,0), x2=(0,1), v1=v2=1, γ1=1, γ2=0. Graph convexity would require some γ∈[0,1] with γ C(x1+x2)=γ1 C x1+γ2 C x2, but for the given C this equation has no solution. Therefore Assumption 21 is not verified, and the sufficient condition (29) at E[v] cannot be justified by Lemma 30. The numerical results in Section 6 therefore do not demonstrate the paper's main result.
minor comments (5)
  1. [§5, Theorem 35] The statement says the probability is 'less than δ/Δ', but the proof yields the non-strict inequality ≤ δ/Δ. The statement should be 'less than or equal to δ/Δ', or the proof should be adjusted.
  2. [§6, Example 36] There is a typo: the second matrix is labeled 'A1 =' but the context indicates it should be A2. Please correct.
  3. [§5, Lemma 33] The conditional expectation notation E[Z_k | Z_j] is imprecise; the condition should be with respect to the filtration generated by the random inputs up to time j, not merely the single random variable Z_j. This is a clarity issue in an otherwise standard argument.
  4. [§4, Example 27] The augmented variable ζ=(v,v^2) has a non-convex support. If the argument relies on extending G and λ to the convex hull of the support, this extension and the domain of the Jensen inequality should be stated explicitly.
  5. [§2, Definition 14] The definition uses a strict inequality '< ρ' while the derived bounds are non-strict. For consistency, either use '≤ ρ' in the definition or note that the strict version follows by taking ρ slightly larger.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained and does not reduce to its assumptions or to self-citations.

full rationale

The paper's central derivation chain is: define λ_k^B recursively via suprema over G (Eq. 15); under Assumptions 4, 5, and 21 plus concavity of B, prove λ_k is concave (Lemma 22); use concavity to apply Jensen's inequality, obtaining the supermartingale condition E[λ_1(x,v)] ≤ λ_1(x,E[v]) (Lemma 30) and the conditional supermartingale bound (Lemma 33); then apply the supermartingale inequality to get the reachability probability bound B(x0)/Δ (Theorem 35). Every step uses standard, externally valid inequalities (Markov, Boole, Jensen, supermartingale) applied to the assumed barrier candidate and the assumed convex-valued map. No parameter is fitted to data and then renamed a prediction; no uniqueness theorem or load-bearing result is imported from the authors' prior work; citations to [13,14] are contextual only. The apparent technical problems in Remark 24 and Examples 36–37—where an equality substitution is used where only Jensen's inequality is justified, and where the set-valued map may fail Assumption 21—are correctness and validity concerns, not circularity. Even if those examples are flawed, the theorem's proof does not reduce to its inputs by construction; under the stated assumptions it is self-contained. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central results rest on three domain assumptions about the system (upper semicontinuity and measurability, i.i.d. inputs, graph convexity) plus standard inequalities from martingale theory. No invented entities or fitted data parameters appear; the barrier functions in the examples are hand-chosen design variables.

free parameters (1)
  • Barrier function coefficients (e.g., c=(0,1), c0=0 in Example 36; B(X)=tr(X) in Example 37) = Example 36: B(x)=x2, i.e., c=(0,1); Example 37: B(X)=tr(X)
    Chosen by hand so that condition (25) holds; the safety ratio delta/Delta depends on this choice and the paper provides no synthesis procedure. These are design choices, not data fits.
assumptions (5)
  • domain assumption Assumption 4: G is upper semicontinuous, locally bounded, with nonempty closed convex values, and v to graph(G(.,v)) is measurable
    Imposed on the system to make worst-case reach probabilities and marginal functions well-defined. Invoked throughout Section 2.3 and the appendix.
  • domain assumption Assumption 5: random inputs v_i are i.i.d. with distribution mu
    Needed for the supermartingale filtration and expectation substitutions (Section 2.3).
  • domain assumption Assumption 21: G is a convex set-valued map (its graph is convex)
    This is the load-bearing convexity used in Lemmas 20, 22, 30, 33, and 35. It is restrictive and not checked in Example 37.
  • standard math Jensen's inequality for concave functions and Doob's maximal inequality for nonnegative supermartingales
    Used in Lemma 30 and Theorem 35 to replace v by E[v] for upper bounds and to bound hitting probabilities.
  • standard math Measurable selection and marginal measurability theorems from Rockafellar and Wets (e.g., [22, Thm 14.13, Cor 14.14, Prop 14.47])
    Used in Appendix A to prove measurability of lambda and compositions of G.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Characterization of Safety in Stochastic Difference Inclusions using Barrier Functions." pith.science (2026). https://pith.science/paper/7Z2V7ELV

@misc{pith2026250820204,
  author       = {Pith},
  title        = {Pith review of: Characterization of Safety in Stochastic Difference Inclusions using Barrier Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7Z2V7ELV}},
  note         = {Machine review of arXiv:2508.20204}
}
read the original abstract

We study stochastic systems characterized by difference inclusions. Such stochastic differential inclusions are defined by set-valued maps involving the current state and stochastic input. For such systems, we investigate the problem of proving bounds on the worst-case probability of violating safety properties. Our approach uses the well-known concept of barrier functions from the study of stochastic control systems. However, barrier functions are hard to prove in the presence of stochastic inputs and adversarial choices due to the set-valued nature of the dynamics. In this paper, we show that under some assumptions on the set-valued map including upper semi-continuity and convexity combined with a concave barrier function vastly simplifies the proof of barrier conditions, allowing us to effectively substitute each random input in terms of its expectation. We prove key results based on the theory of set-valued maps and provide some interesting numerical examples. The ideas proposed here will contribute to the growing interest in problems of robust control and verification of stochastic systems in the presence of uncertain distributions and unmodeled dynamics.

Figures

Figures reproduced from arXiv: 2508.20204 by the authors.

Figure 1
Figure 1. 200 trajectories from randomly generated initial [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. The figure shows the 200 trajectories of [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

  1. [1]

    Probabilistic reachability and safety for controlled discrete time stochastic hybrid systems

    Alessandro Abate, Maria Prandini, John Lygeros, and Shankar Sastry. Probabilistic reachability and safety for controlled discrete time stochastic hybrid systems. Automatica, 44(11):2724–2734, 2008

  2. [2]

    J. P. Aubin. Viability Theory . Birkhauser Boston Inc., Cambridge, MA, USA, 1991

  3. [3]

    J. P. Aubin and H. Frankowska. Set-valued Analysis. Springer Science & Business Media, 2009

  4. [4]

    Programmatic strategy synthesis: Resolving nondeterminism in probabilistic programs

    Kevin Batz, Tom Jannik Biskup, Joost-Pieter Katoen, and Tobias Winkler. Programmatic strategy synthesis: Resolving nondeterminism in probabilistic programs. Proc. ACM Program. Lang., 8(POPL), Jan 2024

  5. [5]

    Optimal inequalities in probability theory: A convex optimization approach.SIAM Journal on Optimization , 15(3):780–804, 2005

    Dimitris Bertsimas and Ioana Popescu. Optimal inequalities in probability theory: A convex optimization approach.SIAM Journal on Optimization , 15(3):780–804, 2005

  6. [6]

    Risk- sensitive safety analysis using conditional value-at-risk.IEEE Transactions on Automatic Control, 67(12):6521–6536, 2021

    Margaret P Chapman, Riccardo Bonalli, Kevin M Smith, Insoon Yang, Marco Pavone, and Claire J Tomlin. Risk- sensitive safety analysis using conditional value-at-risk.IEEE Transactions on Automatic Control, 67(12):6521–6536, 2021

  7. [7]

    Safe control in the presence of stochastic uncertainties

    Albert Chern, Xiang Wang, Abhiram Iyer, and Yorie Nakahira. Safe control in the presence of stochastic uncertainties. In 2021 60th IEEE Conference on Decision and Control (CDC) , pages 6640–6645. IEEE, 2021

  8. [8]

    Distributionally robust optimization under moment uncertainty with application to data-driven problems

    Erick Delage and Yinyu Ye. Distributionally robust optimization under moment uncertainty with application to data-driven problems. Operations Research, 58(3):595–612, 2010

Show all 29 references
  1. [9]

    Real analysis: modern techniques and their applications, volume 40

    Gerald B Folland. Real analysis: modern techniques and their applications, volume 40. John Wiley & Sons, 1999

  2. [10]

    Robust nonlinear control design: state-space and Lyapunov techniques

    Randy Freeman and Petar V Kokotovic. Robust nonlinear control design: state-space and Lyapunov techniques. Springer Science & Business Media, 2008

  3. [11]

    Johnson, Samuel Kotz, and Narayanaswamy Balakrishnan

    Norman L. Johnson, Samuel Kotz, and Narayanaswamy Balakrishnan. Continuous Univariate Distributions, Vol. 1 , volume 1. Wiley, New York, 2nd edition, 1994

  4. [12]

    Safety barrier certificates for stochastic hybrid systems

    Abolfazl Lavaei, Sadegh Soudjani, and Emilio Frazzoli. Safety barrier certificates for stochastic hybrid systems. In 2022 American Control Conference (ACC), pages 880–885. IEEE, 2022

  5. [13]

    A converse robust-safety theorem for differential inclusions

    Mohamed Maghenem and Masoumeh Ghanbarpour. A converse robust-safety theorem for differential inclusions. ESAIM: Control, Optimisation and Calculus of Variations , 31:48, 2025

  6. [14]

    Sufficient conditions for robust safety in differential inclusions using barrier functions

    Mohamed Adlene Maghenem, Masoumeh Ghanbarpour Mamaghani, and Adnane Saoud. Sufficient conditions for robust safety in differential inclusions using barrier functions. Automatica, 171:111938, 2025

  7. [15]

    Abstraction, Refinement and Proof for Probabilistic Systems

    Annabelle McIver and Carroll Morgan. Abstraction, Refinement and Proof for Probabilistic Systems . Springer, 2005

  8. [16]

    Stochastic differential equations: an introduction with applications

    Bernt Oksendal. Stochastic differential equations: an introduction with applications. Springer Science & Business Media, 2013

  9. [17]

    Pflug and D

    G. Pflug and D. Wozabal. Ambiguity in portfolio selection. Quantitative Finance, 7(4):435–442, 2007

  10. [18]

    A lyapunov theorem certifying global weak reachability for stochastic difference inclusions with random inputs

    Corrado Possieri and Andrew R Teel. A lyapunov theorem certifying global weak reachability for stochastic difference inclusions with random inputs. Systems & Control Letters , 109:37–42, 2017

  11. [19]

    Prajna and A

    S. Prajna and A. Jadbabaie. Safety verification of hybrid systems using barrier certificates. In International Workshop on Hybrid Systems: Computation and Control , pages 477–

  12. [20]

    Stochastic safety verification using barrier certificates

    Stephen Prajna, Ali Jadbabaie, and George J Pappas. Stochastic safety verification using barrier certificates. In 2004 43rd IEEE conference on decision and control (CDC)(IEEE Cat. No. 04CH37601) , volume 1, pages 929–

  13. [21]

    Frameworks and results in distributionally robust optimization

    Hamed Rahimian and Sanjay Mehrotra. Frameworks and results in distributionally robust optimization. Open Journal of Mathematical Optimization , 3:1–85, July 2022

  14. [22]

    R. T. Rockafellar and J. B. R Wets. Variational Analysis, volume 317. Springer Science & Business Media, 1997

  15. [23]

    Probabilistic reachability and invariance computation of stochastic systems using linear programming

    Niklas Schmid and John Lygeros. Probabilistic reachability and invariance computation of stochastic systems using linear programming. IF AC-PapersOnLine, 56(2):11229– 11234, 2023

  16. [24]

    Data- driven distributionally robust safety verification using barrier certificates and conditional mean embeddings

    Oliver Sch¨ on, Zhengang Zhong, and Sadegh Soudjani. Data- driven distributionally robust safety verification using barrier certificates and conditional mean embeddings. In 2024 American Control Conference (ACC) , pages 3417–3423, 2024

  17. [25]

    Finite-time regional verification of stochastic non-linear systems

    Jacob Steinhardt and Russ Tedrake. Finite-time regional verification of stochastic non-linear systems. The International Journal of Robotics Research , 31(7):901–923, 2012

  18. [26]

    Verification of discrete time stochastic hybrid systems: A stochastic reach-avoid decision problem

    Sean Summers and John Lygeros. Verification of discrete time stochastic hybrid systems: A stochastic reach-avoid decision problem. Automatica, 46(12):1951–1961, 2010

  19. [27]

    A distributionally robust approach to regret optimal control using the wasserstein distance

    Feras Al Taha, Shuhao Yan, and Eilyan Bitar. A distributionally robust approach to regret optimal control using the wasserstein distance. In 2023 62nd IEEE Conference on Decision and Control (CDC) , pages 2768– 2775, 2023

  20. [28]

    Safety of stochastic systems: An analytic and computational approach

    Rafal Wisniewski and Luminita-Manuela Bujorianu. Safety of stochastic systems: An analytic and computational approach. Automatica, 133:109839, 2021

  21. [29]

    A dynamic game approach to distributionally robust safety specifications for stochastic systems

    Insoon Yang. A dynamic game approach to distributionally robust safety specifications for stochastic systems. Automatica, 94:94–101, 2018. 13

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.