{"id":"606ab836-6329-43a9-a8b0-a648f65a509e","arxiv_id":"1908.09337","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fully distributed stochastic MPC algorithm for linear networks with multiplicative uncertainty, using ADMM and distributed terminal design to achieve recursive feasibility, mean square stability, and chance constraint satisfaction.","lead":"This paper presents a distributed stochastic model predictive control scheme for networks of linear systems with multiplicative noise, using ADMM so each subsystem optimizes locally and communicates only with neighbors. It removes the need for a central coordinator, which matters for large-scale networked control where centralized optimization is intractable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 19's proof covers only the S2 shifted branch; Algorithm 2's S1 feedback branch is never shown feasible or cost-decreasing, even under exact ADMM.","rationale":"I read the paper in good faith. Its contribution is a distributed SMPC formulation, and the strongest claim is Theorem 19: recursive feasibility, mean-square convergence, and chance-constraint satisfaction under initial feasibility. The reader's weakest assumption was that exact ADMM convergence (Assumption 18) is not met by the finite-tolerance implementation. That is a legitimate concern, but it is an explicit idealization common in optimization-based control. Looking for the least secure point in the theorem's own proof, I found a more direct internal gap: Appendix C's only feasible candidate at time k+1 uses S2 initial conditions, while Algorithm 2 solves with S1 whenever S1 is feasible. No argument shows that the S1 problem is feasible or that its optimal cost satisfies the Lyapunov decrease. This gap exists even with εc = 0, so it is more load-bearing than the ADMM tolerance issue. The concern is not that the theorem is necessarily false; it may be repairable by proving S1 feasibility, by always using S2 after initialization, or by qualifying Theorem 19 to the S2 branch. For that reason I keep the verdict CONDITIONAL rather than REJECT: the approach remains plausible, but the central claim as stated is not supported by the proof for the implemented algorithm. I partially agree with the reader because they also identify the S2/chance-constraint proof as not fully rigorous, but their primary stated weakness is the ADMM exactness assumption, which is not the most fundamental problem.","tokens_in":13880,"tokens_out":9262,"duration_ms":94898,"concrete_test":"Re-derive the recursive-feasibility and cost-decrease steps of Appendix C for the S1 branch: at time k+1, instantiate Problem 16 with (z0, Σ0) = (x(k+1), 0) and attempt to construct a feasible shifted solution from the previous optimal trajectories. If this construction is impossible, Theorem 19 is unproven for Algorithm 2. A complementary numerical probe: run Algorithm 2 with εc → 0 on the paper's 3-subsystem example and on several scalar multiplicative-noise systems; at every time where S1 is selected, compare the optimal cost under S1 with the previous shifted S2 cost J_m(1|k) + J_v(1|k). A single time with J*_S1(k+1) > J_m(1|k) + J_v(1|k) would show the proof's key inequality fails for the implemented feedback branch.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is Theorem 19, whose proof in Appendix C builds a candidate for time k+1 from the shifted optimal solution (z*(1|k), Σ*(1|k), shifted v*). This is precisely strategy S2 in the initialization constraint (23). However, Algorithm 2 (lines 4–8) solves Problem 16 with strategy S1 whenever every subsystem passes the S1 feasibility check, i.e. with (z0, Σ0) = (x(k+1), 0). The shifted candidate is not feasible for the S1 problem unless x(k+1) = z*(1|k) and Σ*(1|k) = 0, which is not generally true. The proof supplies no alternative feasible candidate for S1, so recursive feasibility of the closed loop generated by Algorithm 2 is not established. The same gap invalidates the cost-decrease inequality J*(k+1) ≤ J_m(1|k) + J_v(1|k): if S1 is selected, the feasible set is the one with (z0, Σ0) = (x(k+1), 0), and the shifted S2 solution does not lie in it. Consequently the mean-square convergence argument applies only to a hypothetical run that always uses S2, not to the implemented algorithm that prefers S1. This concern is independent of the ADMM tolerance issue raised by the reader: even under Assumption 18 with exact optimization (εc = 0), Theorem 19 is unproven for the S1 branch. The numerical example reports aggregate costs and violations over Monte-Carlo runs and does not test the Lyapunov inequality on S1 steps, so it cannot resolve the gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed stochastic model predictive control (DSMPC) scheme for networks of linear systems with multiplicative, unbounded zero-mean noise and local chance constraints. Chance constraints are converted, via Cantelli's inequality and a linearization, into constraints on the predicted mean and an upper bound on the predicted covariance. The terminal ingredients, including a structured terminal controller, terminal cost, relaxation functions, and terminal covariance bound, are synthesized through distributed LMIs and a distributed LP. The online algorithm solves the global MPC problem with consensus ADMM, using two initialization strategies: S1, which initializes the mean at the measured state and the covariance at zero, and S2, which uses the shifted optimal prediction from the previous step. The main result, Theorem 19, claims recursive feasibility, mean-square convergence of the state to zero, and satisfaction of the chance constraints at all times, assuming exact solution of the distributed optimization. A three-subsystem numerical example illustrates the closed-loop behavior and compares the distributed design with a centralized one.","tokens_in":14251,"tokens_out":11541,"duration_ms":130127,"significance":"If Theorem 19 were fully established, the paper would make a useful contribution: it provides a fully distributed, scalable stochastic MPC formulation for multiplicative uncertainty, avoiding a central node in both design and online computation. The chance-constraint bounds are derived from a distribution-independent inequality rather than tuned to the example, and the terminal design is genuinely distributed. The numerical study gives evidence of practical behavior over many Monte-Carlo runs. However, the theoretical guarantees as stated are not proven for the algorithm actually implemented: the proof covers only the shifted S2 branch, and the exact-ADMM assumption is not satisfied by the finite-tolerance stopping rule. These gaps are load-bearing for the central claims.","major_comments":[{"comment":"The proof of Theorem 19 constructs the feasible candidate at time k+1 by shifting the previous optimal solution, i.e. by setting the initial mean and covariance to (z*(1|k), Σ*(1|k)). This is exactly strategy S2 in constraint (23). However, Algorithm 2 (lines 4-8) solves Problem 16 with strategy S1, i.e. with (z0,Σ0)=(x(k+1),0), whenever the S1 feasibility check succeeds. The shifted candidate is not generally feasible for the S1 problem because it requires z0=z*(1|k) and Σ0=Σ*(1|k), not (x(k+1),0). No alternative feasible candidate is supplied for the S1 branch, so the cost-decrease inequality J*(k+1) ≤ J_m(1|k)+J_v(1|k) is not established for the branch that Algorithm 2 prefers. Consequently, the mean-square convergence argument applies only to a hypothetical run that always uses S2. This gap is independent of the ADMM tolerance issue.","section":"Theorem 19, Algorithm 2, Appendix C"},{"comment":"Theorem 19 assumes that Problem 16 is solved exactly, i.e. that Algorithm 1 converges with ε_c=0. The implemented Algorithm 2 uses the finite stopping criterion ||E_i ξ − y_i||_∞ ≤ ε_c, and ADMM convergence is asymptotic, so ε_c=0 is not reached in finitely many iterations. The manuscript does not quantify how a nonzero ADMM residue affects recursive feasibility, mean-square convergence, or chance constraint satisfaction. The theorem therefore guarantees properties of a limit that the online algorithm never attains. Please either state Theorem 19 explicitly for the idealized exact solver and treat the numerical scheme as an approximation, or provide a suboptimality bound that preserves the closed-loop properties.","section":"Assumption 18, Algorithm 1"},{"comment":"The covariance propagation via (8)-(9) is valid only if Σ̂_{N_i}(t|k) is an upper bound on the true covariance Σ_{N_i}(t|k), as stated in (7). However, Problem 16 lists only (4a), (6), and (9) as constraints and does not include (7). If Σ̂ is a decision variable without the explicit constraint Σ̂_{N_i}(t|k) ⬣ Σ_{N_i}(t|k), the solver could choose Σ̂ too small, making the predicted covariance Σ_i(t+1|k) an underestimate of the true covariance and invalidating the chance constraint guarantee. Please add (7) to the problem statement, specify whether Σ̂ is a decision variable or a precomputed bound, and show that the constraint is preserved in the ADMM local problems.","section":"Problem 16, Eq. (7)"}],"minor_comments":[{"comment":"In (5b) and (6b) the subscript r should be s, consistent with the input constraint index in (2b).","section":"Eqs. (5b), (6b)"},{"comment":"The text says the synthesis is presented 'such that Assumption 4 is satisfied', but the relevant condition is Assumption 9 (the LMI (10)); please correct the cross-reference.","section":"Section 3.3"},{"comment":"The claim that recursive feasibility and constraint satisfaction follow directly from [8, Lem. 8] and [8, Lem. 9] should be expanded, because [8] is a deterministic setting and the terminal covariance condition (17) has no counterpart there; a short argument showing that the stochastic terminal constraints are invariant under the α_i update would make the proof self-contained.","section":"Appendix C"},{"comment":"The simulations do not report how often the S1 branch versus the S2 branch is selected, nor do they test the Lyapunov decrease on S1 steps; such a report would help assess whether the missing S1 analysis is a purely theoretical artifact or a practical issue.","section":"Algorithm 2, numerical example"}],"recommendation":"major_revision","confidential_remarks":"The main concerns are the unproven S1 branch in Theorem 19 and the exact-ADMM assumption; both are fixable in principle by either restricting the theorem to S2 or adding a dedicated analysis of S1 and a suboptimality bound. The missing constraint (7) in Problem 16 is also a serious technical omission that should be corrected. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Christoph,\n\nI read the Mark-Liu paper. The headline: it does something new relative to [9] — a fully distributed DSMPC for multiplicative uncertainty with distributed synthesis of terminal ingredients and an ADMM-based online solver. The LMI framework in Section 3 is a real extension of Conte et al. and looks carefully done. The numerical example is honest, with Monte-Carlo violations and cost comparisons at different tolerances. That part is the paper's strength.\n\nThe problem is Theorem 19. The proof in Appendix C constructs a feasible candidate for time k+1 by shifting the previous optimal solution, i.e. using the S2 initialization (prediction). But Algorithm 2, when every subsystem passes the S1 feasibility check, solves Problem 16 with S1 (feedback, zero covariance). The shifted candidate is not feasible for that problem unless the measured state happens to equal the predicted mean and the predicted covariance is zero, which is generally false. The proof never supplies a feasible candidate or a cost-decrease argument for the S1 branch. So the recursive feasibility claim is actually fine because the S1 check guarantees feasibility by construction, but the mean-square convergence argument only works for a hypothetical run that always uses S2. Since the algorithm prefers S1 whenever possible, the stated guarantee does not follow for the implemented controller. This is not a minor glitch; it's a missing half of the theorem.\n\nThere's a second, more conventional gap: Assumption 18 assumes exact ADMM convergence, while Algorithm 1 uses a finite tolerance εc. The paper doesn't quantify how suboptimality affects feasibility or the Lyapunov decrease. For MPC papers this is often waved away, but here it compounds the first issue: even with εc=0, Theorem 19 is not established for S1.\n\nI'd send it to review if I were the editor — the core idea deserves a chance to be fixed. The authors need to either prove the S1 branch enjoys a similar descent property, or restrict the algorithm to S2 and accept the loss of feedback, or add a safeguard that guarantees the cost decrease. The distributed synthesis part is solid on its own and could be published separately. For now, cite it for the synthesis method, not for the closed-loop guarantees.\n\nBest.","headline":"A promising distributed stochastic MPC scheme whose main theorem does not cover the actual preferred feedback branch (S1), leaving the convergence guarantee unproven for the implemented algorithm.","tokens_in":14710,"tokens_out":6020,"would_cite":true,"duration_ms":59449,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","93A14","93E20","90C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"A feasible start guarantees recursive feasibility, mean-square convergence, and chance constraint satisfaction in distributed stochastic MPC.","keywords":["distributed model predictive control","stochastic MPC","multiplicative uncertainty","chance constraints","Cantelli inequality","alternating direction method of multipliers","mean square stability","recursive feasibility"],"falsifier":"A decisive test is to run the Section 5 example with a boundary-feasible initial condition and decreasing stopping tolerances $\\epsilon_c=10^{-2},10^{-4},10^{-6}$; if any closed-loop trajectory under finite $\\epsilon_c$ becomes infeasible at the next step, or violates the chance constraint at the required $p_{i,x} \\ge 0.7$, then the theorem's guarantees do not extend to the implemented algorithm.","tokens_in":13651,"feed_emoji":"🕸️","tokens_out":9630,"duration_ms":89881,"temperature":0.7,"pith_summary":"This paper proposes a fully distributed stochastic model predictive control (MPC) scheme for networks of linear systems whose dynamics are corrupted by multiplicative, unbounded zero-mean noise. The authors replace probabilistic state and input constraints with mean-and-covariance constraints via Cantelli's inequality, then solve the resulting optimization with consensus ADMM so no central coordinator is needed. The central claim is Theorem 19: if the distributed optimization is feasible at time zero and solved exactly, the closed loop stays feasible, the expected quadratic state norm tends to zero, and the original chance constraints hold at every time. The practical value is that networked control designers get scalability without sacrificing stability or safety certificates.","feed_headline":"Distributed MPC keeps chance constraints under multiplicative noise","feed_subtitle":"A fully decentralized design with no central node guarantees recursive feasibility and mean-square convergence.","key_machinery":"The central machinery is a distributed terminal design: a block-diagonal terminal cost $V_{f,i}(z_i)=z_i^\\top P_i z_i$, a structured local terminal controller $K_{N_i}$, time-varying local terminal sets $Z_{f,i}(\\alpha_i(k))$ whose sizes are updated by relaxation functions $\\Gamma_{N_i}$, and a block-diagonal terminal covariance bound $\\hat{\\Sigma}_{f,i}$. These are computed by solving distributed LMIs and a distributed linear program, and they make the terminal region invariant while satisfying tightened mean and covariance constraints. Online, the global optimization problem is rewritten in consensus form and solved in parallel by ADMM; each subsystem optimizes over its own and its neighbors' predicted trajectories and exchanges copies to reach agreement. Chance constraints are converted into deterministic constraints on means, inputs, and covariances via Cantelli's inequality, $f(p)=\\sqrt{p/(1-p)}$, and then linearized with a parameter $\\epsilon$ that trades conservatism against feasibility.","core_discovery":"The paper establishes that every stabilizing ingredient of stochastic MPC—terminal cost, terminal controller, terminal set scaling, and covariance bound—can be synthesized distributedly for multiplicative uncertainties, and the online distributed problem then inherits the recursive feasibility, mean-square stability, and chance constraint satisfaction of the centralized design. Specifically, Theorem 19 states: if Problem 16 is feasible at $k=0$ and Algorithm 1 converges exactly (Assumption 18), then the closed loop is recursively feasible, $\\mathbb{E}(\\|x(k)\\|_Q^2) \\to 0$ as $k \\to \\infty$, and the chance constraints (2) are satisfied for all $k \\ge 0$. This extends centralized stochastic MPC for multiplicative unbounded uncertainty to networks and removes the need for a central node in both offline design and online operation.","pith_inferences":["Because Theorem 19 assumes exact convergence of the ADMM solver ($\\epsilon_c=0$) whereas Algorithm 2 stops at a finite tolerance, the guarantees for the implemented controller are unquantified; a natural extension is to derive a bound on how suboptimality degrades recursive feasibility, mean-square convergence, and chance constraint satisfaction.","The numerical table suggests a trade-off: looser ADMM tolerances act as extra constraint tightening and reduce observed violations, while tighter tolerances operate closer to the probabilistic limit and show more violations; using the tolerance as an intentional tuning knob is a testable idea the paper leaves open.","As the paper notes in Remark 5, replacing the distribution-free Cantelli bound with the exact quantile of the disturbance should reduce conservatism; a quantitative comparison of feasible volume and closed-loop cost is a straightforward extension.","The block-diagonal upper bound on each neighbor covariance is a deliberate source of conservatism; relaxing it would be a step toward the coupling chance constraints the conclusion names as future work."],"forward_implications":["Under the theorem's conditions, closed-loop trajectories of every subsystem satisfy their chance constraints at all times $k \\ge 0$, even though the multiplicative noise has unbounded support.","The terminal ingredients can be computed without a central node, so the method scales to larger networks where centralized synthesis or centralized initialization is intractable.","With exact ADMM convergence, the closed loop is mean-square stable: $\\mathbb{E}(\\|x(k)\\|_Q^2) \\to 0$ as $k \\to \\infty$.","The linearization parameter $\\epsilon \\in (0,1]$ gives the designer a dial for trading conservatism of the approximated chance constraints against feasibility and closed-loop cost.","The two initialization strategies (feedback and prediction) let the controller maintain recursive feasibility even when resetting the mean and covariance to the measured state would be infeasible."],"supporting_citations":[{"why":"Supplies the distributed synthesis method for terminal cost, terminal controller, relaxation functions, and time-varying terminal sets, and the invariance lemmas used in the proof of Theorem 19.","marker":"[8]"},{"why":"Provides the centralized stochastic MPC formulation for multiplicative unbounded uncertainty and chance constraints that this paper distributes.","marker":"[12]"},{"why":"Supplies the Cantelli-inequality-based reformulation and linearization of chance constraints used in constraints (5) and (6).","marker":"[10]"},{"why":"The consensus ADMM algorithm that decomposes the global MPC problem into local problems and underlies Algorithm 1.","marker":"[5]"},{"why":"Earlier distributed stochastic MPC with additive uncertainty that motivates the initialization strategies S1/S2 and the recursive feasibility treatment.","marker":"[11]"},{"why":"The existing distributed stochastic MPC with parameter uncertainty that this paper contrasts as requiring a central node for initialization.","marker":"[9]"},{"why":"The authors' prior non-iterative DSMPC using probabilistic reachable sets, whose conservatism and central-node requirements motivate the new approach.","marker":"[17]"}],"fun_headline_variants":["Distributed stochastic MPC works without a central node","Chance constraints met via distributed MPC for multiplicative noise","Decentralized stochastic MPC for uncertain networks","Scalable distributed MPC with multiplicative uncertainty","Fully distributed stochastic MPC for multiplicative noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the distributed optimization problem is solved exactly at every step, while the implemented algorithm stops at a finite tolerance, and the paper gives no bound on how this gap affects recursive feasibility, mean-square convergence, or chance constraint satisfaction.","fun_headline_variants_meta":{"raw":{"variants":["Distributed stochastic MPC works without a central node","Chance constraints met via distributed MPC for multiplicative noise","Decentralized stochastic MPC for uncertain networks","Scalable distributed MPC with multiplicative uncertainty","Fully distributed stochastic MPC for multiplicative noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00052,"raw_usage":{"total_tokens":2445,"prompt_tokens":798,"completion_tokens":1647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":1579}},"tokens_in":414,"tokens_out":1647,"duration_ms":12068,"temperature":1.0,"reasoning_tokens":1579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:15:29.631569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to run the Section 5 example with a boundary-feasible initial condition and decreasing stopping tolerances $\\epsilon_c=10^{-2},10^{-4},10^{-6}$; if any closed-loop trajectory under finite $\\epsilon_c$ becomes infeasible at the next step, or violates the chance constraint at the required $p_{i,x} \\ge 0.7$, then the theorem's guarantees do not extend to the implemented algorithm.","supporting_citations":[{"cited_title":"Distributed synthesis and stability o f cooperative distributed model predictive control for line ar systems","cited_arxiv_id":null,"evidence_quote":"Supplies the distributed synthesis method for terminal cost, terminal controller, relaxation functions, and time-varying terminal sets, and the invariance lemmas used in the proof of Theorem 19."},{"cited_title":"Model predic tive control of linear systems with multiplicative unbounded uncertainty and chance constraints","cited_arxiv_id":null,"evidence_quote":"Provides the centralized stochastic MPC formulation for multiplicative unbounded uncertainty and chance constraints that this paper distributes."},{"cited_title":"A probabilistic approach to model predictive control","cited_arxiv_id":null,"evidence_quote":"Supplies the Cantelli-inequality-based reformulation and linearization of chance constraints used in constraints (5) and (6)."},{"cited_title":"Distributed optimization and statistical learning via the alternating direction method of multipliers","cited_arxiv_id":null,"evidence_quote":"The consensus ADMM algorithm that decomposes the global MPC problem into local problems and underlies Algorithm 1."},{"cited_title":"Distributed predictive control of stochastic linear syste ms with chance constraints","cited_arxiv_id":null,"evidence_quote":"Earlier distributed stochastic MPC with additive uncertainty that motivates the initialization strategies S1/S2 and the recursive feasibility treatment."},{"cited_title":"Distributed stochastic mpc of linear systems with paramete r uncertainty and disturbances","cited_arxiv_id":null,"evidence_quote":"The existing distributed stochastic MPC with parameter uncertainty that this paper contrasts as requiring a central node for initialization."},{"cited_title":"Distributed stochastic model predictive control for dynamically coupled linear systems using probabilistic reachable sets","cited_arxiv_id":null,"evidence_quote":"The authors' prior non-iterative DSMPC using probabilistic reachable sets, whose conservatism and central-node requirements motivate the new approach."}],"review_version":1}