{"id":"ceab90b4-68e9-44fb-b0ca-829a4730a647","arxiv_id":"1908.00483","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Under an ISS Lyapunov condition and a mild assumption on disturbance probabilities, the state of a stochastic nonlinear system converges with probability 1 to the minimal robust positively invariant set.","lead":"This paper proves that many stochastic nonlinear control systems will almost surely settle into the smallest disturbance-invariant region around the origin, and never leave it. The result gives tight convergence and average-performance guarantees for three existing stochastic model predictive control algorithms that previously had only conservative or incomplete stability analyses.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 7 overreaches: the minimal RPI set need not satisfy Theorem 5's interior condition, and a rank-deficient D gives a concrete counterexample to indicator convergence.","rationale":"After checking the proof of Theorem 5, the Borel-Cantelli/block argument appears sound: ISS gives finite-time entry into any RPI set with nonempty interior when a block of small disturbances occurs, and Assumption 1 gives the blocks positive probability. The main issue is not Assumption 1, which is mild and explicitly scoped, but the step from 'any RPI set with interior' to 'the minimal RPI set'. Definition 2 does not impose interiority, and intersections of RPI sets can collapse to a lower-dimensional or singleton set. In the linear case D=0, the minimal RPI set is {0}; trajectories converge to it but do not enter it in finite time, so the indicator-function notion of convergence used in (7) fails exactly at the claimed tight limit. This is a load-bearing gap because the paper's headline and MPC applications promise convergence to the minimal RPI set, and the MPC corollaries inherit the same flaw unless the underlying closed-loop disturbance map has full-rank D or X∞ has nonempty interior. The reader already noted this issue in the rationale, so I do not change the CONDITIONAL verdict; I would make the fix a condition: either prove interiority of X∞ under the stated assumptions or weaken the conclusion to distance convergence or eventual entry into arbitrarily tight RPI neighborhoods.","tokens_in":14792,"tokens_out":13328,"duration_ms":150318,"concrete_test":"Run the scalar linear check: x_{k+1}=0.5 x_k + 0·w_k, W=[-1,1], w_k i.i.d. truncated Gaussian (or uniform), x0=1. Verify Assumptions 1-4 with Φ=0.5, D=0, Γ=[-1,1]. Compute X∞={0} from (9). If Corollary 7 or Corollary 13 is applied in this case, it asserts P{lim 1_{X∞}(x_k)=1}=1, but x_k=2^{-k}>0 for all k, so the asserted probability is 0. This single computation isolates the missing interiority condition; if the authors instead prove distance convergence or assume X∞ has nonempty interior, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5 is proved only for RPI sets Ω satisfying Assumption 2, which requires Ω to contain the origin in its interior (Section III.A). Corollary 7 applies this to the minimal RPI set X∞ of Definition 2, but X∞ need not have nonempty interior: X∞ is the intersection of all RPI sets containing the origin, and this intersection can be a singleton. Example: with n=1, x_{k+1}=0.5x_k+0·w_k, W=[-1,1], the minimal RPI set is X∞={0}, which has empty interior. For x0=1 the trajectory is x_k=2^{-k} and never equals 0, so P{lim 1_{X∞}(x_k)=1}=0 even though every RPI interval satisfying Assumption 2 satisfies (7). Hence the claimed tight ultimate bound in Remark 6 and the MPC corollaries (Corollaries 10, 13 and Section IV.C) rely on an unproved interiority assumption on X∞. The correct conclusion from Theorem 5 is convergence in distance to X∞ (or eventual entry into any interior RPI neighborhood), not necessarily eventual entry into X∞ itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies almost sure convergence of discrete-time stochastic nonlinear systems with additive i.i.d. disturbances. Under an input-to-state stability (ISS) assumption and a condition that arbitrarily small disturbances have positive probability, Theorem 5 proves that the state converges with probability 1 to any robust positively invariant (RPI) set containing the origin in its interior. The authors then claim (Corollary 7) that this implies almost sure convergence to the minimal RPI set X∞, and, under an additional linear-dynamics-on-a-limit-set condition, they prove a bound on the asymptotic average quadratic cost (Theorem 8). These results are applied to three existing stochastic MPC formulations from [18], [19], and [12], yielding new convergence statements for those controllers. The core Borel-Cantelli / ISS argument is carefully executed and is correct for RPI sets with nonempty interior; however, the extension to the minimal RPI set is not justified and is false in degenerate cases.","tokens_in":14998,"tokens_out":9224,"duration_ms":96245,"significance":"If repaired, the paper would provide a useful and general framework for deriving almost sure convergence and tight long-run performance bounds for stochastic MPC, going beyond earlier analyses that only establish boundedness or Lyapunov-type inequalities. The proof technique based on ISS Lyapunov functions plus Borel-Cantelli is elegant and is correctly applied to interior RPI sets. The claimed stronger result of almost sure entry into the minimal RPI set, however, is not valid without additional assumptions, and this affects several corollaries and the nonlinear MPC application. The proposed fix—either adding an interiority assumption on X∞ or restating the result as convergence in distance to X∞—is local and preserves much of the paper's value. The work is not accompanied by code or machine-checked proofs, but the main mathematical arguments are presented in a self-contained and generally careful way.","major_comments":[{"comment":"Corollary 7 is false as stated. Theorem 5 requires Assumption 2, which demands that the RPI set Ω contains the origin in its interior, but the minimal RPI set X∞ of Definition 2 is only the intersection of RPI sets containing the origin and need not have nonempty interior. For example, take n=1, x_{k+1}=0.5x_k+0·w_k, W=[-1,1], and let w_k be i.i.d. uniform on W so that Assumption 1 holds. Then X∞={0}, which does not contain the origin in its interior, and for x0=1 the trajectory is x_k=2^{-k}; it never equals 0, so P{lim_{k→∞} 1_{X∞}(x_k)=1}=0, contradicting (8). The proof of Corollary 7 asserts that the minimal RPI set 'also satisfies this assumption' (Assumption 2) without proof, and that assertion is false. The correct conclusion from Theorem 5 is P{lim_{k→∞} d(x_k,X∞)=0}=1, or eventual entry into any RPI set that contains the origin in its interior. This correction propagates to Remark 6, Corollaries 10 and 13, and the closing statement of Section IV.C.","section":"Section III.A, Corollary 7 and Remark 6"},{"comment":"Proposition 14 does not verify Assumption 2 for Ω=X∞. The proof states 'Assumption 2 holds because X∞ is bounded due to X∞⊆Z', but Assumption 2 also requires that Ω contains the origin in its interior, and boundedness does not imply this. In the same degenerate linear example as in the previous comment, X∞={0} is RPI and bounded but has empty interior, so Proposition 14 fails and Corollary 7 cannot be invoked for the nonlinear MPC law of (26). Please either add an explicit interiority condition on X∞ (for the nonlinear case, or on the disturbance matrix D and W in the linear cases) or restate the result as convergence in distance to X∞.","section":"Section IV.C, Proposition 14"},{"comment":"The proof of Theorem 8 invokes the summability bound ∑_{j} P{xj∉Γ} ≤ Nf p^{-Nf} from Lemma 4, but Lemma 4 applies only to an RPI set Γ satisfying Assumption 2, i.e., containing the origin in its interior. Assumption 4 as stated only requires Γ to be an RPI set on which the dynamics are linear; it does not require 0∈int Γ. The paragraph preceding Assumption 4 says 'an RPI set containing the origin,' but the formal assumption should state this explicitly. Without that interiority condition, the derivation of (10) is incomplete because P{xj∉Γ} need not be summable for Γ with empty interior (e.g., Γ={0} in the scalar example above, where P{xj∉{0}}=1 for all j).","section":"Section III.B, Theorem 8"}],"minor_comments":[{"comment":"In the paragraph before equation (6), '⌊N/Nf⌋' should be '⌊k/Nf⌋'.","section":"Lemma 4 proof"},{"comment":"'Borell-Cantelli' should be 'Borel-Cantelli'.","section":"Section III.A, paragraph after Lemma 4"},{"comment":"The proof invokes [21, Lem. 22] to conclude ISS, but the displayed inequality only treats w=0; please state the lemma or explain how it supplies the K-function bound on the disturbance term needed in (3b).","section":"Theorem 11 proof"},{"comment":"The definition '0 ∈ X ⊆ X' uses the symbol X for both the state space and an element of the intersection, which is confusing; the intended meaning is probably that each set in the intersection contains 0 and is a subset of the state space.","section":"Definition 2"},{"comment":"The reference [22] lists '2009 European' but the conference appears to be the 2019 European Control Conference; please correct the year.","section":"Reference [22]"}],"recommendation":"major_revision","confidential_remarks":"The main proof of Theorem 5 is sound for interior RPI sets, but the overclaim about convergence to the minimal RPI set is a substantive error that will need to be corrected in all corollaries and in Section IV.C. The revision should either strengthen the assumptions (e.g., full-row-rank D and full-dimensional W, or an explicit interiority condition on X∞) or consistently restate the results in terms of convergence in distance. The paper is otherwise a solid contribution and the corrections are local, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper does something real. It shows that for a discrete-time nonlinear system with an ISS Lyapunov function and an i.i.d. disturbance with positive probability of arbitrarily small realizations, the state enters any RPI set containing the origin in its interior with probability 1, and does so after finitely many excursions. The Borel-Cantelli argument (Lemma 4, Theorem 5) is careful and correct. The extension of [20] to general ISS systems and the removal of linearity/controllability assumptions relative to [22] are genuine advances. The applications to three existing MPC formulations, giving first convergence results for [19] and [12], are useful.\n\nThe soft spot is Corollary 7. It asserts the minimal RPI set X_infty also satisfies Assumption 2, i.e., contains the origin in its interior. That is not generally true. The minimal RPI set is the intersection of all RPI sets containing the origin, and that intersection can be a singleton. A scalar example: x_{k+1} = 0.5 x_k (D=0), W=[-1,1]; X_infty = {0}, which has empty interior, and trajectories never hit 0, so the indicator convergence (8) fails. The proof of Corollary 7 simply asserts the interior property; it does not follow. What Theorem 5 actually gives is convergence to any interior RPI set, hence eventual entry into arbitrarily small neighborhoods of X_infty, but not necessarily into X_infty itself. This matters because Corollaries 10 and 13, and the tight-ultimate-bound claims in Section IV, lean on Corollary 7. They are true under an additional assumption that X_infty has nonempty interior (which holds for many practical problems, e.g., when D is full row rank and W has interior), but the paper does not state that condition.\n\nThe other assumptions are handled honestly. Assumption 1's 'arbitrarily small disturbances' restriction is stated and discussed. Assumption 5 in the striped MPC section is indeed an assumption, not a theorem, and the paper does not verify it for concrete problem data; that is a minor gap. The citation pattern is fair, including the self-citations, which are used for background properties rather than to assume the conclusion.\n\nBottom line: the main theorem is worth having, and the flaw is fixable by either adding an interiority assumption for X_infty or rephrasing the corollaries in terms of distance convergence to X_infty. I would send this to review; a referee should push on the minimal-RPI-set claims.","headline":"Solid almost-sure convergence machinery for stochastic MPC, but the claimed convergence to the minimal RPI set outruns the assumptions—the set may lack interior, so the tight-ultimate-bound corollaries need an extra condition.","tokens_in":15551,"tokens_out":3019,"would_cite":true,"duration_ms":29751,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-14T15:53:44.982715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}