{"id":"7e53f75c-8fd1-405e-84b3-56f1b930d707","arxiv_id":"2603.17843","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A certainty-equivalent MPC combined with least-mean-square parameter adaptation yields linear-in-noise cumulative tracking error bounds for uncertain nonlinear systems under input and soft state constraints.","lead":"This paper combines a certainty-equivalent model-predictive controller with online least-mean-square parameter estimation for nonlinear systems with unknown parameters. The main guarantee is a cumulative bound on tracking error and constraint violations that grows linearly with noise, disturbances, and parameter drift.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4 as stated is false: with h(x,u,θ)=θu and f independent of θ, LMS cannot estimate θ and a constant tracking error accumulates, so inequality (7) fails despite all of Assumptions 1–5.","rationale":"The reader identified Assumption 3c (feasible target for every θ) as the weakest assumption, but this is not where the central claim actually fails. The counterexample above satisfies Assumption 3c: y_rd,θ=0.5 for every θ∈[0.5,2]. It also satisfies Assumptions 1, 2, 4, and 5. The failure is that the output map h is allowed to depend on u and θ, while the proof of Theorem 4 bounds the output error as if h were a function of x alone. The LMS estimator only corrects parameters that affect the one-step state prediction through G(x,u,w); parameters appearing solely in the output map are not identifiable from the state measurements used by the controller. As a result, the closed loop can converge to the wrong setpoint in output space, producing a nonzero constant tracking error and violating the claimed linear-in-noise bound. This is not a matter of restrictive assumptions or missing numerical verification; the central theoretical result, Theorem 4, is false as stated. A repair would require either restricting h to be independent of the unknown parameters (or at least ensuring the output error is controlled by the state/input error via the stage cost), or feeding the output measurement into the adaptation law. Without such a change, the paper's main claim should be rejected. The paper may still contain useful ideas, but the headline semi-global adaptive tracking guarantee is not supported.","tokens_in":32569,"tokens_out":18309,"duration_ms":184364,"concrete_test":"Run the 1-D counterexample with the paper's code framework: f=0.5x+u, h=θu, U=[-1,1], Θ=[0.5,2], y_d=0.5, true θ=2, θhat0=1, x0=1, W=V={0}. Record the cumulative output error Σ||y_k-0.5||^2 over a long horizon. If, as expected, it grows linearly while the RHS of (7) is constant, Theorem 4 is falsified.","verdict_should_be":"REJECT","load_bearing_attack":"Theorem 4's proof bounds the output tracking error using only Lipschitz continuity of y=h(x) (proof of Thm. 4, Part I, around Eq. (D.11)), but the paper defines y_k=h(x_k,u_k,θ_k) in Eq. (4) and later allows h=Cθx+Dθu+fθ in Sec. 6.2. This is a proof gap, and it is fatal for the stated claim. Counterexample satisfying Assumptions 1–5: n_x=n_u=n_y=n_θ=1, U=[-1,1], Θ=[0.5,2], W=V={0}, f(x,u,θ,w)=0.5x+u+w (so G≡0 in Assumption 2), h(x,u,θ)=θu, y_d=0.5. Then S(θ)={(2u,u,θu)}, y_rd,θ=0.5, so Assumption 3 holds; Assumption 5 holds with ρ=0.5. Let true θ=2, θhat0=1, x0=1. The LMS regressor \\hatΦ_k=G(·)=0, so θhat_k=1 forever. The certainty-equivalent MPC (with θhat=1) steers the state to (x_s,u_s,y_s)=(1,0.5,0.5). The true output is y_k=2·0.5=1 for all k, so ||y_k-y_rd,θ||^2=0.25 each step. The RHS of (7) is constant (no w, v, or Δθ), so (7) fails as K→∞. The missing term is the dependence of h on u and θ; the LMS update (10) only uses state-prediction error and cannot estimate parameters that enter only the output map.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a certainty-equivalent adaptive MPC scheme for discrete-time nonlinear systems with linearly parameterized dynamics. It combines a tracking MPC with artificial references, finite-horizon rollout terminal costs, soft state constraints, and a projected least-mean-square (LMS) parameter adaptation. The central claim is Theorem 4: for open-loop exponentially stable systems satisfying Assumptions 1–5, the closed loop satisfies the cumulative bound (7), in which tracking error and state-constraint violation are bounded linearly by disturbance/noise energy and parameter-variation path length, plus an initial-condition term. A regional variant for stabilizable systems is stated in Theorem 7. The paper also contains nominal stability and inherent-robustness theorems, a discussion of the special linear case, and two numerical studies with open-source code.","tokens_in":33042,"tokens_out":10032,"duration_ms":98202,"significance":"If Theorem 4 were correct, the paper would be a substantial contribution: it would give an implementable adaptive MPC for nonlinear systems without a control Lyapunov function, with a strong worst-case cumulative performance guarantee. The manuscript is well organized, contains detailed appendices, and the simulation code is a useful asset. However, the main theorem is false under the stated assumptions. The proof of Theorem 4 relates the output error to the state error using a Lipschitz condition on y=h(x), whereas the paper defines y_k=h(x_k,u_k,θ_k) and later explicitly allows h to depend on u and θ. A simple counterexample satisfying Assumptions 1–5 shows that the LMS estimator cannot identify parameters that appear only in the output map, producing a permanent tracking offset and violating (7). This flaw is load-bearing: it invalidates the paper's principal advertised contribution, and it also propagates to the regional result Theorem 7.","major_comments":[{"comment":"The proof bounds the output tracking error using \"Lipschitz continuity of y=h(x)\", i.e., it treats the output as state-dependent only. But Eq. (4) defines y_k=h(x_k,u_k,θ_k), and Sec. 6.2 explicitly allows h=C_θ x + D_θ u + f_θ. This gap is fatal. Counterexample: n_x=n_u=n_y=n_θ=1, U=[-1,1], Θ=[0.5,2], W=V={0}, f=0.5x+u (so G≡0 in Assumption 2), h=θu, y_d=0.5. Then S(θ)={(2u,u,θu)} and y_rd,θ=0.5, so Assumptions 1, 3, and 5 hold. For true θ=2, θhat_0=1, x_0=1, the LMS regressor is zero, so θhat_k=1 forever; the MPC with θhat=1 selects the setpoint (1,0.5,0.5) and keeps x_k=1. The true output is y_k=2·0.5=1, hence ||y_k-y_rd,θ||^2=0.25 for all k, while the RHS of (7) is bounded. Inequality (7) therefore fails as K→∞. The structural cause is that the LMS update (10) only uses state-prediction error; parameters appearing only in h cannot be estimated.","section":"Appendix D, Eq. (D.11); Theorem 4"},{"comment":"The regional result inherits the same output-map gap. Its proof concludes with \"The remainder of the proof is analogous to Theorem 4\", and the invalid Part I of the proof of Theorem 4 is exactly the step that relates y_k to the state error. Unless the output map is restricted (e.g., h independent of u and θ), or the adaptation is changed to use output prediction error, the regional claim is unsupported for the same reason. A revision cannot be limited to patching Eq. (D.11); it must either narrow the class of output maps or alter the performance bound to include an unavoidable output-identification offset.","section":"Appendix E / Theorem 7"}],"minor_comments":[{"comment":"The text says \"Assumption 3c) follows from linear independence constraint qualification and second order sufficient conditions ... Assumption 3d) requires a feasible target y_d.\" The labels appear swapped: Assumption 3c is the feasible-target condition, while Assumption 3d is the setpoint-regularity condition.","section":"Sec. 6.2"},{"comment":"The abstract claims the approach \"requires no system-specific offline design,\" but Section 5 relies on a locally stabilizing feedback κ (e.g., LQR), which is a system-specific design choice. The claim should be qualified to the open-loop stable setting of Section 4.","section":"Abstract / Sec. 5"},{"comment":"The paper states that the quadrotor example does not verify the sufficient conditions of Theorem 7. The example should be presented explicitly as an illustration of practical behavior, not as numerical validation of the regional theorem.","section":"Sec. 7.2"},{"comment":"The key nominal-stability condition (B.3) is imported from [25, Thm. 6–7] via the phrase \"adapts the worst-case linear programming analysis.\" Since this condition is load-bearing for the later robustness and adaptation results, a more self-contained statement of the imported LP result would improve verifiability.","section":"Appendix B, Theorem 2"}],"recommendation":"reject","confidential_remarks":"The manuscript is ambitious and generally careful, but the main theorem is contradicted by a simple example under its own assumptions. This is not a local proof gap: the proposed LMS estimator is structurally unable to identify parameters that appear only in the output map, so the advertised semi-global cumulative guarantee cannot hold for the general output model in Eq. (4). The paper would need a substantially different problem formulation (e.g., state-dependent output only, or output-error adaptation with sufficient excitation) before the central claims could be reconsidered. The heavy reliance on the author's prior work is acceptable but makes independent verification harder."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a real contribution to adaptive MPC: combining certainty-equivalent tracking MPC with finite-tail rollouts, artificial setpoints, soft constraints, and LMS adaptation, then proving a cumulative error bound linear in noise, disturbance, and parameter-drift energy, is genuinely new. The nominal stability and robustness analysis is detailed and mostly clear. Second, the central theorem is false as stated. The stress-test counterexample is valid, and it is not a technicality.\n\nThe proof of Theorem 4 bounds the output tracking error in (D.11) using the line “Given Lipschitz continuity of y=h(x)”, but the paper defines y_k = h(x_k,u_k,θ_k) in (4), and Section 6.2 explicitly allows Dθ u + fθ. Assumption 3c only says the chosen target y_d is feasible for every θ; it makes y_rd,θ = y_d, but it does not remove the dependence of h on u and θ. Take f(x,u,θ,w)=0.5x+u, h(x,u,θ)=θu, Θ=[0.5,2], y_d=0.5, true θ=2, θhat_0=1. All Assumptions 1–5 hold. The LMS regressor G is identically zero, so θhat stays 1 forever; the certainty-equivalent MPC drives (x,u) to (1,0.5); the true output is 1, not 0.5. The left side of Inequality (7) grows linearly in K while the right side is constant. Objective 1 is not achieved.\n\nWhat the paper does well: the LMS analysis in Theorem 1 is solid, the adaptive-MPC combination is clearly motivated, the decomposition into nominal stability and inherent robustness is a serious technical effort, and the simulations are honest and reproducible via linked code. There is no circularity in the main bound.\n\nThe soft spots beyond the central flaw are secondary: Theorem 2 imports the LP stability condition from [25] as a black box, Assumptions 3c and 5 are genuinely restrictive, and Theorem 7’s uncertainty bounds are only qualitative, with the quadrotor example explicitly not verifying them. But the output-map issue is load-bearing. It changes the class of systems for which the claim can be true: essentially, h must not depend on the estimated parameters or on u, or the output parameters have to be estimated separately. I do not see a reading that repairs the theorem.\n\nThe paper deserves a serious referee, not a desk reject, and the referee should be asked to check the output dependence in the proof of Theorem 4. In its present form the main theorem needs a major revision or a substantially narrower claim, and I would not cite it yet.","headline":"The adaptive MPC machinery is novel and mostly careful, but Theorem 4 is false as stated: the proof silently treats y_k as h(x_k) while the system defines y_k = h(x_k,u_k,θ_k), and a simple scalar counterexample satisfies every stated assumption.","tokens_in":33476,"tokens_out":5231,"would_cite":false,"duration_ms":53961,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B45","93C40","93D25","93D30","93C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"A certainty-equivalent adaptive MPC proves that cumulative tracking error and constraint violation scale linearly with noise, disturbance, and parameter-drift energy, for open-loop stable nonlinear systems with large parametric uncertainty.","keywords":["adaptive model predictive control","certainty equivalence","least-mean-square parameter adaptation","nonlinear uncertain systems","tracking MPC with artificial references","inherent robustness","constraint violation bounds","open-loop stable systems"],"falsifier":"Take a linear open-loop stable system satisfying all of the paper's assumptions, set noise and disturbances to zero, keep the parameter constant, but choose a target y_d that is infeasible for one parameter value in Θ. Run the controller: if the cumulative tracking error grows linearly with the horizon K rather than staying bounded as Corollary 2 predicts, then Assumption 3c is load-bearing. Alternatively, run the same controller on a marginally stable (non-exponentially stable) linear system with persistent noise and check whether the right-hand side of inequality (7) is violated, which would","tokens_in":32457,"feed_emoji":"🎯","tokens_out":4250,"duration_ms":46134,"temperature":0.7,"pith_summary":"This paper sets out to show that an adaptive model predictive controller can track uncertain nonlinear systems without an offline-designed Lyapunov function: it couples a certainty-equivalent tracking MPC—one that trusts the current parameter estimate—with a projected least-mean-square update. For open-loop exponentially stable systems, the result is a semi-global guarantee: cumulative tracking error and state-constraint violation grow at most linearly with noise energy, disturbance energy, and the path length of parameter variation, with initial parameter error entering only as a transient constant. For stabilizable but not necessarily open-loop stable systems, the same combination gives a regional guarantee inside a computable region of attraction, provided the uncertainty stays sufficiently small. The paper matters because classical nonlinear adaptive control typically requires system-specific control-Lyapunov or barrier functions, while robust MPC requires small uncertainty; this scheme promises large parametric uncertainty with only standard regularity assumptions and no system-specific offline design.","feed_headline":"Adaptive MPC error scales linearly with noise and drift","feed_subtitle":"A certainty-equivalent controller plus least-mean-square estimation yields a semi-global robustness bound for nonlinear systems.","key_machinery":"The load-bearing object is the optimal value function J_N^*(x,θ) of a certainty-equivalent tracking MPC. The MPC optimizes over an artificial steady state (x_s,u_s,y_s) in the feasible-setpoint set S(θ), adds a quadratic offset penalty toward the target y_d, uses a finite-horizon rollout of the steady-state input (or of a local stabilizing feedback in the regional case) as a terminal penalty, and softens state constraints with quadratic penalties. Because the target is assumed feasible for every parameter (Assumption 3c), the optimal offset is zero and J_N^* itself becomes a Lyapunov function; Theorem 3 shows it contracts robustly, with growth bounded by prediction error and parameter-estima","core_discovery":"The central claim is Theorem 4: given compact bound sets, an open-loop exponentially stable nonlinear system whose dynamics are linear in the unknown parameters, and a certainty-equivalent tracking MPC with a finite-horizon rollout terminal cost, the closed loop driven by the LMS parameter update satisfies, for all horizons K, sum_{k=0}^{K-1} (||y_k - y_{rd,θ_k}||^2 + ||x_k||_X^2) ≤ C1 sum_{k=0}^{K-1} (||w_k||^2 + ||v_k||^2 + ||Δθ_k||) + C2(||θ_0 - θhat_0||^2 + ||x_0 - x_{rd,θ_0}||^2), with uniform constants. In words: every unit of noise, disturbance, or parameter drift enters the tracking and constraint-violation budget linearly, and the initial parameter error matters only through a const","pith_inferences":["Because the guarantee is cumulative rather than pointwise, the scheme establishes performance under adaptation but not safety: pointwise state-constraint satisfaction would still require robust tubes or tightened constraints, which the paper explicitly does not claim.","The linear dependence on Δθ suggests time-varying parameters are tracked at a cost proportional to their total variation; a stress test would be to increase the drift bound in Lemma 9 until the contraction factor λ+c_θ||Δθ|| crosses 1 and observe where the stated guarantee breaks.","The quadrotor demonstration is presented as a simulation, not as a verification of the regional theorem's sufficient conditions; checking inequality (35) numerically would tell whether the guarantee, rather than only the simulation, drives the reported success.","If the feasible-target assumption fails, the Lyapunov argument would need an extra offset term; a natural extension is an output-offset estimator (as in offset-free MPC) to recover a modified bound with a persistent-offset constant."],"forward_implications":["If Theorem 4 is correct, adaptive MPC can accommodate arbitrarily large parametric uncertainty, noise, and disturbances for open-loop stable systems, with no offline design of a control-Lyapunov or barrier function.","In the disturbance-free, noise-free, constant-parameter case, the controller converges to the optimal feasible setpoint, and the cumulative tracking and constraint-violation error stays bounded even with a large initial parameter error.","The same controller directly handles input constraints, soft state constraints, and piecewise-constant reference changes, with computational cost only moderately above a standard MPC implementation.","For stabilizable systems, the regional version gives the same linear cumulative bound inside a specified region of attraction when noise, disturbance, and parameter uncertainty are small enough, using a local feedback such as an LQR as the terminal ingredient.","For linear systems, the assumptions reduce to Schur stability or stabilizability plus affine-in-θ parametrization, giving a broad generic class of linear problems to which the guarantee applies.","If the paper is right, it closes a practical gap: nonlinear adaptive control under constraints can be designed by choosing horizons and weights computationally, rather than by solving a symbolic Lyapunov design."],"fun_headline_variants":["Linear error scaling in adaptive MPC for nonlinear systems","Certainty-equivalent MPC guarantees linear error growth","Adaptive MPC error grows only linearly with disturbances","Semi-global adaptive MPC with linear error bounds","MPC + LMS: tracking error linear in noise and drift"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The strongest result assumes the desired output y_d is exactly a feasible steady-state output for every parameter in the known set (Assumption 3c), and that the system is globally open-loop exponentially stable for every parameter (Assumption 5); if either fails, the semi-global linear bound is not proven, and the paper's Appendix F notes that under only local Lipschitz continuity the semi-global claim additionally needs parameter drift to be sufficiently small.","fun_headline_variants_meta":{"raw":{"variants":["Linear error scaling in adaptive MPC for nonlinear systems","Certainty-equivalent MPC guarantees linear error growth","Adaptive MPC error grows only linearly with disturbances","Semi-global adaptive MPC with linear error bounds","MPC + LMS: tracking error linear in noise and drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001056,"raw_usage":{"total_tokens":4288,"prompt_tokens":784,"completion_tokens":3504,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":3429}},"tokens_in":528,"tokens_out":3504,"duration_ms":22434,"temperature":1.0,"reasoning_tokens":3429,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:55:35.351262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a linear open-loop stable system satisfying all of the paper's assumptions, set noise and disturbances to zero, keep the parameter constant, but choose a target y_d that is infeasible for one parameter value in Θ. Run the controller: if the cumulative tracking error grows linearly with the horizon K rather than staying bounded as Corollary 2 predicts, then Assumption 3c is load-bearing. Alternatively, run the same controller on a marginally stable (non-exponentially stable) linear system with persistent noise and check whether the right-hand side of inequality (7) is violated, which would","supporting_citations":[],"review_version":1}