{"id":"7b2cf7f4-1b31-44a7-8e44-5c0ee718fcd9","arxiv_id":"2411.14370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Rigorous convergence and stability proofs for the extended infinite-horizon MPC and its zone-control variant, covering general non-square gain matrices and any input horizon.","lead":"This paper proves, with mathematical rigor, that two popular model predictive control (MPC) algorithms used in refineries converge to the desired operating point and are stable, for a broader class of systems than previously proven. The result matters because industry relies on these controllers, and gaps in their theoretical guarantees were unresolved for twenty years.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2.8 asserts ϕ>0 for rectangles without proof; this unproved geometric bound is load-bearing for Theorem 2.3 and must be supplied or replaced.","rationale":"The reader and I identify the same weakest point: the unproved ϕ > 0 assertion in Proposition 2.8. I do not follow the reader's stronger claim that ϕ necessarily fails for unbounded rectangles; the identity |cos θ_x| = ‖Pr x − x‖/‖Π_r x − x‖ suggests the assertion may be true for fixed rectangles via a polyhedral error bound, but that is precisely what needs proof. The concern is load-bearing because Theorem 2.3's convergence proof relies on Proposition 2.8, and the Appendix makes the dependence on ϕ^{-2} explicit. The rest of the argument is checkable, and Section 3 does not use this geometric claim, so the paper is not fatally flawed. A conditional verdict is appropriate: the authors must add the missing geometric lemma or state and use a boundedness assumption. Since this matches the reader's conditional conclusion, I recommend no change to the verdict.","tokens_in":21508,"tokens_out":32483,"duration_ms":316654,"concrete_test":"Prove or disprove the missing inequality in Proposition 2.8: there exists a constant C = C(U, D0) such that for every x ∈ U, ‖Π_r x − Pr x‖ ≤ C ‖Pr x − x‖. Equivalently, use the Hoffman error bound for the polyhedral pair (Ur, U) to derive such a C, or construct a sequence x_n ∈ U with ‖Π_r x_n − Pr x_n‖ / ‖Pr x_n − x_n‖ → ∞. If the bound holds, add it as a lemma to Proposition 2.8; if the sequence exists for an unbounded rectangle, the proof must add a boundedness assumption or replace the ϕ^{-1} bound with a different argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Proposition 2.8, the proof needs the bound ‖Δũ(0|k)‖ ≤ αϕ^{-1}(c+ε), where ϕ := inf_{x∈U\\ker D0} |cos θ_x| and θ_x is the angle between Pr x − x and Π_r x − x. The sentence 'Since U is a rectangle we have that ϕ > 0' is asserted without proof. This is not a routine consequence of the word 'rectangle': writing α = Pr x − x and β = Π_r x − x, one has α ⟂ (β−α), so |cos θ_x| = ‖Pr x − x‖ / ‖Π_r x − x‖. Thus the claim is equivalent to a polyhedral error-bound/transversality property: the normal distance from x to Ur must control the tangential gap ‖Π_r x − Pr x‖ uniformly over U. That is a genuine geometric lemma, not an immediate observation. The proof of Theorem 2.3 uses Proposition 2.8 to conclude ‖δ_k^*‖_S → 0, and the Appendix defines C3 with a ϕ^{-2} factor, so the whole contradiction argument collapses if ϕ can be zero or cannot be proved positive. The paper never states whether U is bounded; if boundedness is silently intended, that assumption should be explicit and the compactness/transversality argument should be provided. As written, the one-line justification is insufficient and the central convergence theorem rests on an unproved geometric claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the extended infinite-horizon MPC of Odloak (2004) and the zone-control MPC of González and Odloak (2009). It claims convergence of the optimal cost to zero and uniform stability of the tracking error under a non-regular static gain D0 and for any input horizon m. The proofs use a Lyapunov-type decrease argument, a constructed slack weight S (or Su), and a contradiction argument based on an improved feasible strategy.","tokens_in":21789,"tokens_out":29773,"duration_ms":272030,"significance":"If correct, the results provide the first rigorous convergence and stability proof for these industrially used MPC formulations without requiring D0 to be square or regular, and without restricting the input horizon to m=1. The paper gives explicit constructive conditions on S and Su, and many of the algebraic steps, including the decrease lemmas and the use of stability of F, are checkable and sound. However, the central convergence theorem rests on an unproved geometric claim, and the feasibility arguments for the increment constraints are incomplete.","major_comments":[{"comment":"The proof asserts without proof that for a rectangle U, ϕ := inf_{x∈U\\kerD0} |cos θ_x| > 0, where θ_x is the angle between Pr x − x and Π_r x − x. This is load-bearing: it is used to bound ‖Δũ(0|k)‖ ≤ αϕ^{-1}(c+ε), and the Appendix defines C3 with a ϕ^{-2} factor. The claim is not a routine consequence of U being a rectangle, because kerD0 is an arbitrary subspace, not necessarily coordinate-aligned. Moreover, U is not stated to be bounded, so the infimum can be approached at infinity. I am not claiming the assertion is false; but as written it is an unproved global transversality/error-bound property. Please supply a proof (for example via recession cones and the positive angle between kerD0 and the relevant faces of U) or add boundedness of U with a compactness-based argument. Without this, the contradiction in Proposition 2.8 and hence Theorem 2.3 is incomplete.","section":"§2.2, Proposition 2.8"},{"comment":"Feasibility of the constructed increment needs more care. In Proposition 2.8 the strategy sets Δũ(0|k)=α(Π_r u^*(0|k−1)−u^*(0|k−1)) and the text says 'Since U is convex, we can choose α small enough such that this strategy is feasible.' Feasibility also requires Δũ(0|k)∈ΔU, and ΔU is only assumed to be a rectangle containing the origin. If 0 is not in the interior of ΔU, no positive scaling of that vector may be admissible. Since α is later fixed as 3ε/(c+ε), the choice of ε must be coordinated with the size of ΔU. The same issue appears in Theorem 3.3, where (12) uses Δu^*(m−1|k)∈int ΔU and the perturbed increment Δu^*(m−1|k)−(1−α)δ^*_{u,k} is required to lie in ΔU. Please state explicitly that ΔU (and, where needed, U) have nonempty interior containing the origin, or otherwise justify the existence of the required α.","section":"§2.2, Proposition 2.8 and §3.2, Theorem 3.3"}],"minor_comments":[{"comment":"The convention for x_s(0|k) should be stated explicitly. The equations in Lemmas 2.5 and 2.7 suggest that x_s(0|k) is the state after the first move Δu^*(0|k) has been applied, rather than the measured state before the optimization. With the standard MPC convention, identities such as e_s^*(0|k)−δ_k^* = −D_0∑_{j=1}^{m−1}Δu^*(j|k) would appear inconsistent with (3). Stating the convention would remove ambiguity.","section":"§2"},{"comment":"There are typos in the acknowledgements: 'finantial' should be 'financial'.","section":"Acknowledgements"},{"comment":"The notation Γ_Z^2 ∨ (2Γ_{\\bar Q}Γ_{Z−R}) should define the symbol ∨, or use max, for readers.","section":"Appendix"},{"comment":"The phrase 'op erate' in the abstract appears to be a broken word from typesetting; please correct in the final version.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The main uncertainty is the geometric lemma in Proposition 2.8. If the authors can supply a rigorous proof, or an explicit boundedness and transversality assumption, the paper's central claims are likely to be correct. I also recommend that the editors ask for explicit assumptions on ΔU, since both feasibility constructions rely on the ability to take small nonzero increments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper extends Odloak 2004 and Gonzalez-Odloak 2009 convergence/stability proofs from regular D0 and m=1 to any D0 and any input horizon. The generalization matters because non-square gain matrices are the norm in real refineries. The main proof skeleton is sound: Lemma 2.5's cost decrease is real, Lemma 2.6's use of the decay of F is legitimate, and the LQ-factorization construction of S is a genuinely new device that makes the slack-weight choice explicit.\n\nThe soft spot is Proposition 2.8's claim that φ := inf |cos θ_x| > 0 over the rectangle U. That infimum is exactly the ratio of the normal distance to Ur to the distance to U∩Ur. It is not a routine consequence of 'rectangle.' For bounded U it should follow from a compactness plus face-by-face transversality argument; for unbounded U you need to be careful about directions at infinity. The claim is probably true, but the paper simply asserts it and the constant C3 in the appendix carries a φ^{-2} factor, so the convergence theorem collapses without it. A referee should ask for a lemma: either assume U bounded and prove the infimum is positive, or state and prove the general polyhedral version. This is fixable, not fatal.\n\nMinor issues: several 'after some elementary computation' steps are actually lengthy substitutions, and a few inequalities in Theorem 2.3 are written as if they follow immediately from Lemma 2.6/2.7 when they need a subsequence argument. These are cosmetic once you fill them in. No code or numerics, but this is a proof paper and doesn't need them. Citations trace the Odloak lineage properly; prior work is credited and the new theorems are clearly marked as not in the earlier papers.\n\nBottom line: the central claim is plausible and valuable, the proof framework is transparent, and the one load-bearing gap is a missing geometric lemma rather than a hidden circularity. Send it to a serious referee, and tell them to make the authors prove φ > 0 properly. I'd bring it to a reading group, and I'd cite it if I worked on MPC stability theory.","headline":"A solid proof extension of the Odloak MPC stability results; the main claim is likely right, but the paper needs a real proof of the load-bearing φ>0 claim before it is complete.","tokens_in":22361,"tokens_out":16966,"would_cite":true,"duration_ms":152273,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C95","93D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that for the extended infinite-horizon MPC and its zone-control variant, the optimal cost converges to zero and the closed loop is uniformly stable for any input horizon and any (possibly singular) static gain matrix.","keywords":["model predictive control","infinite horizon","nominal stability","convergence","zone control","OPOM model","non-regular gain matrix"],"falsifier":"For a stable plant with singular $D_0$, choose a rectangle $U$ with a side parallel to the affine set $U_r = u_r + \\ker D_0$, and compute $\\phi := \\inf_{x\\in U\\setminus\\ker D_0}|\\cos\\theta_x|$. If $\\phi = 0$ and the closed-loop simulation from a nonzero steady state shows $\\limsup_{k\\to\\infty}\\|u^*(0|k)-P_r u^*(0|k)\\| > 0$ (equivalently, $V_k^*$ does not tend to 0) for some horizon $m\\ge 2$, the central claim would be false; at minimum such an example would pinpoint the missing hypothesis in Proposition 2.8.","tokens_in":21227,"feed_emoji":"🎛️","tokens_out":8335,"duration_ms":68762,"temperature":0.7,"pith_summary":"This paper supplies complete proofs of two facts that earlier analyses of these infinite-horizon MPC algorithms had only argued in restricted settings. For the extended nominally stable MPC with the output prediction-oriented (OPOM) model, under stability of the open-loop system and reachability of the set-point, the optimal cost $V_k^*$ can be driven to zero as $k\\to\\infty$ by choosing the slack-weighting matrix $S$ appropriately, and the tracking error is uniformly stable. The same conclusions hold for the zone-control variant when the target weight $S_u$ is chosen large enough. These results hold for any input horizon $m$ and for any static gain matrix $D_0$, which need not be square or full rank. This matters because industrial MPC implementations commonly have more outputs than inputs or linearly dependent gains, and earlier proofs assumed a regular $D_0$ and $m=1$.","feed_headline":"Infinite-horizon MPC: cost vanishes for any horizon, any gain","feed_subtitle":"New proofs cover non-square gain matrices and all input horizons m, not just m=1.","key_machinery":"The load-bearing object is the finite-horizon equivalent of the infinite-horizon cost together with a specially built slack penalty matrix $S$. Because the open-loop system is stable, the infinite sum in $V_k$ collapses to the finite sum (4) with terminal weight $\\bar Q$ solving the Lyapunov relation $\\bar Q - F^T\\bar QF = F^T\\Psi^T Q\\Psi F$; the terminal constraint $x_s(m-1|k) - \\delta_k - r = 0$ keeps the tail bounded. The matrix $\\hat S = K_1^T (L^{-1})^T L^{-1} K_1 + K_2^T K_2$, built from an LQ factorization of $D_0^\\perp$, is the device that converts output-side slack costs into input-side distances. The other mechanism is the monotonicity lemma (Lemma 2.5), which shows $V_{k+1}^* \\le \\tilde V_{k+1} \\le V_k^*$ by shifting the optimal move sequence forward, and from it the paper derives $\\Delta u^*(0|k)\\to 0$ and $x_d^*(0|k)\\to 0$, the decay facts needed for the contradiction.","core_discovery":"The central discovery is that a single geometric-algebraic construction removes the two standing restrictions on these controllers. Theorem 2.3 states that under the stability and set-point assumptions one can pick $S = \\beta\\hat S$ so that $\\lim_{k\\to\\infty} V_k^* = 0$, and Theorem 2.4 derives uniform stability of the output error. The construction of $\\hat S$ via the LQ factorization of the restriction $D_0^\\perp$ makes $\\|D_0 v\\|_{\\hat S} = \\|v\\|$ for every input direction $v\\in(\\ker D_0)^\\perp$, so that distances in output space become distances in input space. This reduction permits a contradiction argument in Proposition 2.8: if the implemented input sequence stays away from the reference affine set $U_r = u_r + \\ker D_0$, a feasible one-step move toward $U_r$ produces a cost strictly lower than the optimum. The zone-control analog, Theorems 3.3 and 3.4, extends the same mechanism to the formulation with output zones and input targets, with $S_u > H + I_{n_u}$ replacing the condition on $S$.","pith_inferences":["Beyond the paper's claims, the same projection-and-LQ construction may yield closed-form slack weights for robust or invariant-set MPC variants, since the only model-dependent objects are $D_0$, $\\Psi$, $F$, and the Lyapunov weight.","A testable extension not stated in the paper is to allow a set-point sequence $r_k$ that converges geometrically; the proof's estimates suggest the tracking error would inherit the same geometric decay rate.","The unresolved geometric assertion $\\phi>0$ is likely the first point to break in practice: for bounded rectangles it is true whenever the target affine subspace is not parallel to a face, but for unbounded $U$ the claimed uniform bound may fail, and the convergence conclusion would need separate hypotheses."],"forward_implications":["Recursive feasibility and a non-increasing optimal cost hold for any input horizon $m$, so a practitioner can choose longer horizons without losing the Lyapunov-based stability guarantee.","The explicit choice $S = \\beta\\hat S$ (with $\\beta$ a computable constant) and the zone-control choice $S_u > H + I_{n_u}$ provide concrete tuning rules for the slack weights.","Systems with non-square or singular static gain can be handled directly, covering plants with more outputs than inputs or with redundant actuator directions.","The proof templates in Sections 2 and 3 can be adapted to the derived MPC variants mentioned in the introduction (integrating systems, dead time, unstable systems, two-layer RTO-MPC), exactly as the authors claim."],"supporting_citations":[{"why":"Odloak 2004 supplies the extended infinite-horizon MPC cost, the terminal constraint (3), and the monotonicity argument that the paper generalizes to any $D_0$ and $m$.","marker":"[19]"},{"why":"Gonzalez and Odloak 2009 supply the zone-control MPC formulation with slack variables and terminal constraints (9)-(10) whose convergence and stability are here proven for general $m$.","marker":"[10]"},{"why":"Rodrigues and Odloak 2003 provide the OPOM state-space model (1)-(2) and the infinite-horizon tracking formulation that the paper's proofs rely on.","marker":"[25]"},{"why":"Rawlings and Muske 1993 establish the reduction of an infinite-horizon MPC cost to a finite horizon with a Lyapunov terminal weight, the device used to derive the finite expression (4).","marker":"[24]"},{"why":"Keerthi and Gilbert 1988 establish the terminal-state-constraint stability technique that motivates constraint (3) and the Lyapunov interpretation of the cost.","marker":"[12]"}],"fun_headline_variants":["MPC stability proof extended to all input horizons and gains","New proof: infinite-horizon MPC stable for any gain and horizon","Geometric proof removes MPC's horizon and gain restrictions","MPC convergence proven for non-square gains and any horizon","Infinite-horizon MPC: new stability proof for all horizons and gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's convergence argument rests on an unproved geometric assertion that the angle between the direction from any feasible input $x$ to its orthogonal projection onto the target affine set and the direction from $x$ to the closest feasible target point is uniformly bounded away from 90 degrees over the rectangle $U$; if this angle can approach 90 degrees, the bound on the corrective input move collapses.","fun_headline_variants_meta":{"raw":{"variants":["MPC stability proof extended to all input horizons and gains","New proof: infinite-horizon MPC stable for any gain and horizon","Geometric proof removes MPC's horizon and gain restrictions","MPC convergence proven for non-square gains and any horizon","Infinite-horizon MPC: new stability proof for all horizons and gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00062,"raw_usage":{"total_tokens":2951,"prompt_tokens":1095,"completion_tokens":1856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":1769}},"tokens_in":711,"tokens_out":1856,"duration_ms":10607,"temperature":1.0,"reasoning_tokens":1769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:17:08.186806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a stable plant with singular $D_0$, choose a rectangle $U$ with a side parallel to the affine set $U_r = u_r + \\ker D_0$, and compute $\\phi := \\inf_{x\\in U\\setminus\\ker D_0}|\\cos\\theta_x|$. If $\\phi = 0$ and the closed-loop simulation from a nonzero steady state shows $\\limsup_{k\\to\\infty}\\|u^*(0|k)-P_r u^*(0|k)\\| > 0$ (equivalently, $V_k^*$ does not tend to 0) for some horizon $m\\ge 2$, the central claim would be false; at minimum such an example would pinpoint the missing hypothesis in Proposition 2.8.","supporting_citations":[{"cited_title":"Odloak, Extended robust model predictive control, A IChE Journal 50 (8) (2004) 1824– 1836","cited_arxiv_id":null,"evidence_quote":"Odloak 2004 supplies the extended infinite-horizon MPC cost, the terminal constraint (3), and the monotonicity argument that the paper generalizes to any $D_0$ and $m$."},{"cited_title":"González, D","cited_arxiv_id":null,"evidence_quote":"Gonzalez and Odloak 2009 supply the zone-control MPC formulation with slack variables and terminal constraints (9)-(10) whose convergence and stability are here proven for general $m$."},{"cited_title":"Rodrigues, D","cited_arxiv_id":null,"evidence_quote":"Rodrigues and Odloak 2003 provide the OPOM state-space model (1)-(2) and the infinite-horizon tracking formulation that the paper's proofs rely on."},{"cited_title":"Rawlings, K.R","cited_arxiv_id":null,"evidence_quote":"Rawlings and Muske 1993 establish the reduction of an infinite-horizon MPC cost to a finite horizon with a Lyapunov terminal weight, the device used to derive the finite expression (4)."},{"cited_title":"Keerthi, E.G","cited_arxiv_id":null,"evidence_quote":"Keerthi and Gilbert 1988 establish the terminal-state-constraint stability technique that motivates constraint (3) and the Lyapunov interpretation of the cost."}],"review_version":1}