{"id":"9c3986c3-331c-4b54-80df-3f477f604f2e","arxiv_id":"2502.06469","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A stochastic MPC scheme using online-optimized affine disturbance feedback achieves recursive feasibility and closed-loop chance constraint satisfaction for linear Gaussian systems.","lead":"This paper presents a new model predictive control method that re-optimizes its feedback policy at every time step for systems with random disturbances, while still guaranteeing that safety constraints are met with the required probability. It solves a gap in stochastic control where online-tuned policies usually lose their formal safety guarantees.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Offline terminal set construction lacks a termination guarantee and is demonstrated outside Assumption 1, but the central theorem is sound under its stated assumptions.","rationale":"I agree with the reader's identification of the terminal set construction as the weakest point. The paper is transparent about the limitation, and the theoretical contribution is sound conditional on Assumption 1. The numerical example's violation of the boundedness assumption is acknowledged in Remark 5. No circular reasoning or false claims were found. The central theorem is well-supported; the concern is about practical implementability and the gap between assumptions and the demonstration, not about a mathematical error in the proof.","tokens_in":26234,"tokens_out":32753,"duration_ms":258018,"concrete_test":"For the numerical example, construct a dense grid of (z, vec(Phi^x_N)) samples within the computed S_59 and directly verify the infinite tail inequalities (14) for i = 0, ..., 1000 using the exact definitions; also run Algorithm 1 with the S-procedure on a small system satisfying Assumption 1 to see if it terminates within a reasonable time. If any sampled point violates a tail constraint, or if the S-procedure fails to terminate, then the practical offline construction is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee (Theorem 2) is conditional on the existence of the terminal set S_mu from Algorithm 1. Theorem 1 and Proposition 1 guarantee termination of Algorithm 1 under Assumption 1, but the implementable version uses S-procedure sufficient conditions for the set inclusions (Appendices E-F), and Remark 4 explicitly states that these may not terminate. Moreover, the numerical example in Section VII uses a one-sided constraint (44), which makes L = {y in R | y <= 0.5} unbounded, violating the boundedness condition in Assumption 1; the example relies on Algorithm 1 converging anyway (Remark 5). Consequently, the closed-loop recursive feasibility and chance constraint guarantees are not actually proven for the demonstrated scenario, and there is no implementable procedure with a finite-time guarantee for computing the terminal set. If the set inclusion checks are inconclusive or the algorithm does not terminate, the entire proposed SMPC scheme cannot be constructed, and the guarantees in Theorem 2 are void. This is the most load-bearing concern because every online guarantee hinges on the offline terminal set.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a stochastic model predictive control (SMPC) scheme for linear time-invariant systems with additive i.i.d. Gaussian disturbances, in which the disturbance-feedback matrices are optimized online via the system level parameterization. The central mechanism is a reconditioning framework that updates the predicted probabilistic constraints on the current disturbance history, combined with a finitely determined maximal admissible terminal set for the infinite-horizon tail under a fixed terminal controller. The authors prove that the receding-horizon implementation is recursively feasible, satisfies the original chance constraints in closed loop, and yields a bounded asymptotic average cost (Theorem 2). The method is demonstrated on a building temperature control example, with numerical comparisons against several existing SMPC schemes.","tokens_in":26336,"tokens_out":17382,"duration_ms":155115,"significance":"If the results hold, the paper fills a genuine gap: existing SMPC methods with online-optimized feedback either lack closed-loop chance-constraint guarantees or rely on bounded disturbances. The SOC reformulation of the probabilistic constraints, the finite-determination argument for the probabilistic terminal set, and the recursive-feasibility proof via shifted candidate policies are the main technical strengths. The paper provides detailed proofs, a convex formulation, and released code, which are all positive features. The principal caveat is the offline terminal-set construction: the implementable S-procedure checks are only sufficient and may not terminate, and the numerical example lies outside the assumptions of the main theorem. These issues are fixable but currently leave the practical scheme without a finite-time certificate in the demonstrated setting.","major_comments":[{"comment":"Theorem 2's closed-loop guarantees are conditional on the existence of the terminal set S_μ, and Proposition 1 proves termination of Algorithm 1 only under exact set-inclusion checks. The actual implementation proposed in Appendices E–F uses lossy S-procedure sufficient conditions, and Remark 4 explicitly concedes that these may not terminate. Thus Algorithm 2 has no finite-time certificate for constructing S_μ in the form that is implemented. This is load-bearing because without S_μ the online problem (19) and all subsequent guarantees are void. Please either provide conditions under which the S-procedure checks are exact or terminate, or restate the main theorem with an explicit computability assumption and discuss the resulting gap.","section":"Algorithm 1, Remarks 4–5, Appendices E–F"},{"comment":"The numerical example uses the one-sided chance constraint [1,0,0]x_k ≥ −0.5, which makes the set L unbounded and hence violates the boundedness part of Assumption 1. Consequently, Theorem 2 does not apply to the demonstrated scenario; the paper relies on Remark 5 and on Algorithm 1 having converged in that specific instance. Please modify the example so that Assumption 1 is satisfied (for instance, by adding an upper bound on the room temperature so that L is bounded), or provide a separate proof that the one-sided case still yields a finitely determined S_μ under the other conditions of Assumption 1.","section":"Section VII, constraint (44), Remark 5"}],"minor_comments":[{"comment":"The notation in Remark 1 contains a corrupted symbol where 'mathbfΣ' appears; it should read Σ_w^0.","section":"Remark 1"},{"comment":"The notation χ2(·) for the inverse cumulative distribution function of the chi-squared distribution with one degree of freedom is nonstandard and could be confused with the chi-squared random variable; a more explicit notation such as Φ^{-1}_{χ^2_1}(·) would improve clarity.","section":"Section II-A"},{"comment":"The proof of recursive feasibility states that the candidate satisfies the constraints 'with equality', but this is not literally true in the degenerate case (33)–(34) and in the α<0 case (38), where the candidate satisfies the relevant set definition by construction rather than by equality. A sentence clarifying the case distinction would make the argument easier to verify.","section":"Appendix H, proof of Theorem 2 Part I"},{"comment":"The IF method's closed-loop cost is reported in parentheses with a satisfaction level of 0.0%; it would help to state explicitly that this cost is not comparable because the method violates the chance constraints.","section":"Section VII, Table I"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical result appears sound, and I agree with the reader's positive assessment of the proofs. My main concern is that the practical algorithm relies on a terminal-set computation that is not guaranteed to terminate in its implemented form, and the numerical example intentionally violates the boundedness assumption of the main theorem. These issues are fixable and do not invalidate the theorem, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper. It does what it says: for linear Gaussian systems with unbounded disturbances, it gives an SMPC scheme that optimizes affine disturbance feedback online and still guarantees recursive feasibility and closed-loop chance constraint satisfaction. The two ingredients—reconditioning the chance constraints on realized disturbances, and a finitely determined maximal admissible set for probabilistic constraints—are both built from existing ideas, but the synthesis is new and the proofs are solid. I checked the appendices; the argument for Theorem 2 really does go through with the candidate solution (30)–(32), and the cost decrease argument is standard but correct.\n\nThe paper is also honest about its limits. Remark 4 states that the S-procedure sufficient conditions in Algorithm 1 may not terminate, so the implementable version lacks a finite-time guarantee. Remark 5 notes that the numerical example violates the boundedness assumption in Assumption 1, so the guarantees of Theorem 2 are not formally proven for that example. These are real soft spots, but they are disclosed and they do not undermine the central theorem under its stated assumptions. If I have a complaint, it's that the terminal constraint (40) is very conservative: the RC method essentially reproduces the open-loop policy (23), as Figure 1 shows. RC-mod does better, but it's an approximation without the same guarantees. That limits the practical upside of the re-optimization part, though the closed-loop guarantee is still the point.\n\nThe numerical study is fine as a demonstration, not a validation. The comparison with CLP [17] and IF [13]+[21] makes the right qualitative points. The paper cites prior work properly; the self-citations are to genuinely related results.\n\nWho benefits: researchers working on stochastic MPC with unbounded disturbances and closed-loop guarantees. This is a solid theory contribution that deserves a serious referee. The referee should push on the offline construction, but the paper already frames that as an open issue. Accept.","headline":"Closes the gap between online-optimized feedback and closed-loop chance constraint guarantees in SMPC, with honest caveats about the offline terminal set computation.","tokens_in":26953,"tokens_out":1884,"would_cite":true,"duration_ms":23473,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","93E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Stochastic MPC optimizes feedback online without losing closed-loop safety guarantees","keywords":["stochastic model predictive control","chance constraints","disturbance feedback","system level synthesis","maximal admissible set","reconditioning","recursive feasibility","Gaussian disturbances"],"falsifier":"Simulate Algorithm 2 on a system satisfying Assumption 1 over many (e.g. $10^5$) i.i.d. Gaussian disturbance sequences and record, at every time step, whether the optimization (41) is infeasible or the empirical frequency of constraint violation falls below the required $p_j$; any such event contradicts the theorem. A cheaper offline falsifier is to run Algorithm 1 with the S-procedure sufficient conditions on a system satisfying Assumption 1 and observe non-termination, which would show the proposed construction is not actually computable.","tokens_in":1554,"feed_emoji":"🎛️","tokens_out":2001,"duration_ms":48357,"temperature":0.7,"pith_summary":"This paper proposes a stochastic model predictive control (SMPC) method for linear time-invariant systems with additive Gaussian disturbances that optimizes affine disturbance-feedback policies online at every time step. The authors claim this is the first SMPC scheme to combine online-optimized feedback with recursive feasibility, closed-loop satisfaction of probabilistic (chance) constraints, and a bounded asymptotic average cost. The key enablers are a finitely determined maximal admissible set used as a terminal constraint, and a reconditioning step that updates the predicted chance constraints using the realized past disturbances each time the optimization is solved. A building temperature control example illustrates that the method operates close to the probabilistic constraint while meeting it in closed loop.","feed_headline":"Stochastic MPC optimizes feedback online without losing safety guarantees","feed_subtitle":"A finitely determined terminal set plus reconditioning keeps chance constraints satisfied for all time, despite unbounded Gaussian noise.","key_machinery":"The machinery is the system level parameterization (SLP), which expresses the controlled error trajectory as a convolution of past disturbances with the optimized feedback matrices and turns the chance constraints into second-order cone constraints. A finitely determined maximal admissible set $S_\\mu = S_\\infty$ for the probabilistic tail constraints serves as the terminal set at $k=0$; Algorithm 1 computes it using maximal admissible set theory. At $k>0$, a reconditioning update computes each required probability level $\\alpha^i_{j|k}$ from the previous optimal policy shifted and conditioned on the latest disturbance $w_{k-1}$, and the terminal set is replaced by the fixed shifted tail (40), which trivially satisfies the tail constraints.","core_discovery":"The central claim is Theorem 2: if Assumption 1 holds and the initial SOCP (19) is feasible, then the receding-horizon scheme Algorithm 2, which solves (19) at $k=0$ and (41) for $k>0$, keeps (41) recursively feasible for all $k>0$, satisfies the chance constraints (2) in closed loop, and bounds the asymptotic average cost by the performance of the fixed terminal feedback $K$. This closes a gap identified in the paper: prior methods with online-optimized disturbance feedback either assumed bounded disturbances, used recovery mechanisms, or forfeited closed-loop chance-constraint guarantees, while methods with such guarantees used fixed feedback policies. The proof works by reconditioning the probabilistic constraints at each step on the realized disturbance history, so the predicted distribution of the trajectory under the candidate shifted policy matches the distribution that the previous policy would have produced in hindsight.","pith_inferences":["The practical construction of the terminal set relies on S-procedure sufficient conditions that, as the paper's Remark 4 states, may not terminate; should Algorithm 1 fail for a given instance, the scheme cannot be built even if the system satisfies Assumption 1.","The paper's numerical example actually violates the boundedness part of Assumption 1 and still converges, suggesting finite determination may hold under weaker conditions than assumed; this is a testable conjecture, not a paper claim.","A natural extension, implied rather than proven here, would replace Gaussian disturbances with subgaussian or bounded-variance distributions and keep the same reconditioning structure, though the SOC reformulations would need new tail bounds.","The terminal constraint (40) at $k>0$ is deliberately conservative, tying the tail to the previous optimal solution; relaxing it with a re-computed finite index, as the modified variant RC-mod does, seems to recover performance at extra online cost."],"forward_implications":["Recursive feasibility is maintained even though the Gaussian disturbance has unbounded support, without softening constraints or adding a recovery mechanism.","The controller can operate close to the probabilistic constraint, achieving lower expected cost than fixed-feedback SMPC or robust constraint-tightening approaches.","Closed-loop chance constraint satisfaction is guaranteed despite the feedback matrices being random variables, because the constraints are reconditioned on the realized disturbance history at each step.","The method extends the reconditioning idea from mission-wide joint chance constraints to a receding-horizon setting with causal disturbance-feedback policies.","The asymptotic average cost is no worse than that of the static terminal feedback $K$, providing a performance ceiling."],"supporting_citations":[{"why":"Supplies the system level synthesis parameterization used to jointly optimize state trajectories and disturbance feedback matrices in a convex manner.","marker":"[21]"},{"why":"Provides the indirect-feedback SMPC paradigm and the exact second-order cone reformulation of scalar Gaussian chance constraints that the paper extends to optimized feedback.","marker":"[13]"},{"why":"Supplies the shrinking-horizon reconditioning idea that the paper generalizes to a receding-horizon setting with causal policies.","marker":"[27]"},{"why":"Inspires the finitely determined maximal admissible set construction for probabilistic constraints used as the terminal set.","marker":"[28]"},{"why":"Provides the classical maximal admissible set theory, including observability and finite-determination arguments that Theorem 1 builds on.","marker":"[29]"},{"why":"Establishes that optimization over state feedback policies can be written as a convex problem, the foundation for online-optimized disturbance feedback.","marker":"[20]"}],"fun_headline_variants":["Online stochastic MPC with recursive feasibility and chance constraints","Guaranteed chance constraints for online-optimized stochastic MPC","Closing the gap: online stochastic MPC with safety guarantees","Reconditioning enables full closed-loop guarantees in stochastic MPC"],"cache_read_input_tokens":29056,"weakest_assumption_plain":"The scheme is only guaranteed to exist if the offline Algorithm 1 terminates and returns a non-empty terminal set $S_\\mu$, which its guaranteed convergence requires Assumption 1 (bounded constraint set, observability of $(C_K,A_K)$, and the margin condition (17)); with the practical S-procedure checks, termination is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Online stochastic MPC with recursive feasibility and chance constraints","Guaranteed chance constraints for online-optimized stochastic MPC","Closing the gap: online stochastic MPC with safety guarantees","Reconditioning enables full closed-loop guarantees in stochastic MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2651,"prompt_tokens":816,"completion_tokens":1835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1770}},"tokens_in":432,"tokens_out":1835,"duration_ms":12777,"temperature":1.0,"reasoning_tokens":1770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:20:59.438618+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate Algorithm 2 on a system satisfying Assumption 1 over many (e.g. $10^5$) i.i.d. Gaussian disturbance sequences and record, at every time step, whether the optimization (41) is infeasible or the empirical frequency of constraint violation falls below the required $p_j$; any such event contradicts the theorem. A cheaper offline falsifier is to run Algorithm 1 with the S-procedure sufficient conditions on a system satisfying Assumption 1 and observe non-termination, which would show the proposed construction is not actually computable.","supporting_citations":[{"cited_title":"Recursive feasibility of stochastic model predic- tive control with mission-wide probabilistic constraints,","cited_arxiv_id":null,"evidence_quote":"Supplies the shrinking-horizon reconditioning idea that the paper generalizes to a receding-horizon setting with causal policies."},{"cited_title":"Chance-constrained controller state and reference governor,","cited_arxiv_id":null,"evidence_quote":"Inspires the finitely determined maximal admissible set construction for probabilistic constraints used as the terminal set."},{"cited_title":"Optimization over state feedback policies for robust control with constraints,","cited_arxiv_id":null,"evidence_quote":"Establishes that optimization over state feedback policies can be written as a convex problem, the foundation for online-optimized disturbance feedback."}],"review_version":1}