{"id":"e5d5cc7e-9396-4a80-bd39-dee78f9c3764","arxiv_id":"2607.02851","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Kernel multi-step predictors enable data-driven MPC for nonlinear systems with recursive feasibility and practical stability when the horizon is long and representation error is small.","lead":"This paper builds a predictive controller for unknown nonlinear systems that uses only input-output data and kernel methods, then proves it stays feasible and practically stable. It matters because many real plants lack reliable models, yet still need constraint-aware control with formal guarantees.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The stability margin of Theorem 1 is controlled by a generally uncertifiable uniform kernel-error bound d-bar that enters every candidate-cost estimate.","rationale":"The reader correctly isolates Assumption 2/3 (and the resulting d-bar) as the weakest link. My reading confirms that this quantity is not merely a technical convenience: it is the sole parameter that makes the candidate-cost excess αY(d-bar) small enough for the Lyapunov decrease of Theorem 1 to hold. Because the paper supplies no data-driven certificate for d-bar (only the admission in Remark 2 that an approximate order of magnitude “is sufficient”), the existence statement remains non-constructive for any unknown plant. The numerical examples succeed with hand-tuned values, but that does not close the certification gap. The concern therefore reinforces, rather than overturns, the reader’s CONDITIONAL verdict; no stronger rejection is warranted because the proofs themselves appear free of circularity once the assumptions are granted.","tokens_in":29511,"tokens_out":638,"duration_ms":7001,"concrete_test":"On the 1-D vehicle example of §VI-A, replace the hand-tuned d-bar=10^{-3} by the data-driven residual max_a |y_ker[a]-y[a]| evaluated on a held-out validation trajectory of length 200; recompute the closed-loop trajectories of Fig. 1 with this residual (and with 2\times and 5\times that residual). If any of the three runs loses recursive feasibility or the practical Lyapunov decrease (22b) fails for the L=10 used in the paper, the claim that “sufficiently small d-bar” can be verified in practice is falsified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 1 asserts recursive feasibility and practical stability of the nominal DDKPC whenever L>L_Y-bar and d-bar≤d0. The proof constructs a candidate solution whose cost excess is bounded by αY(d-bar) (see (24), (28) and the definition of αY after (28)). That excess is absorbed into the Lyapunov decrease only when d-bar is smaller than an a-priori threshold that itself depends on the unknown constants Γ (Assumption 2) and the uniform bound of Assumption 3. Remark 2 candidly admits that both quantities are “generally intractable” for black-box plants and that only “approximate order of magnitude” is needed for tuning. Consequently the existence claim of Theorem 1 is non-constructive: one cannot verify, from data alone, that the hypothesis d-bar≤d0 holds, nor compute a concrete horizon L_Y-bar that guarantees the claimed decrease. The same uncertifiable bound reappears (as w-bar or d-hat) in the robust and online theorems, so the entire closed-loop guarantee rests on a quantity that the data-driven setting cannot certify.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a data-driven kernel-based predictive control (DDKPC) scheme for unknown nonlinear input-output systems. A multi-step predictor is constructed from offline (and later online) trajectories via the representer theorem; the resulting nonconvex program is analyzed without terminal ingredients. Theorem 1 establishes recursive feasibility and practical closed-loop stability for the noise-free case when the horizon is long enough and the uniform kernel-error bound is small enough. The same Lyapunov argument is extended to measurement noise via a unified uncertainty bound (Theorem 2) and to slowly time-varying plants via an ALD-managed fixed-budget dictionary with periodic predictor updates (Theorem 3). A penalty-relaxation formulation is proposed for real-time solution, and two numerical examples (1-D vehicle, payload-carrying quadruped) illustrate tracking performance.","tokens_in":29819,"tokens_out":1153,"duration_ms":9451,"significance":"If the claims hold, the paper supplies one of the first closed-loop recursive-feasibility and practical-stability certificates for a kernelized multi-step DDPC scheme that does not rely on terminal ingredients or full-state measurements. The online ALD extension and the unified-noise robustification are natural and useful increments over existing kernelized DDPC and GP-MPC literature. The candidate-solution construction in the proof of Theorem 1 is explicit and follows standard practical-MPC reasoning; the numerical examples (especially the contact-rich quadruped residual controller) give concrete evidence that the formulation can be implemented. The main limitation is that the stability margins remain non-constructive with respect to the kernel-error bound, which is standard but should be stated more carefully for a data-driven claim.","major_comments":[{"comment":"Theorem 1 (and likewise Theorems 2–3) asserts recursive feasibility and practical stability once L > L_Ȳ and d̄ ≤ d0. The proof constructs a candidate whose cost excess is bounded by α_Y(d̄) (see (24), (28) and the definition of α_Y after (28)). That excess is absorbed only when d̄ is smaller than a threshold that itself depends on the unknown RKHS-norm bound Γ of Assumption 2 and the uniform bound of Assumption 3. Remark 2 candidly admits both quantities are “generally intractable” for black-box plants. Consequently the existence claim is non-constructive: from data alone one cannot verify d̄ ≤ d0 nor compute a concrete L_Ȳ that guarantees the claimed decrease. The same uncertifiable scale reappears as w̄ and d̂. The paper should either (i) supply a data-driven upper bound / validation procedure for d̄ (or an a-posteriori residual check that certifies the Lyapunov decrease online), or (","section":null},{"comment":"Section IV-A replaces the hard kernel constraint (19b) by a quadratic penalty and then assumes a numerical solver returns a point whose objective suboptimality δ_J and residual δ_K are “sufficiently small” to be absorbed into α_Y. No quantitative relation is given between (δ_J, δ_K) and the constants that appear in a_L and α_Y of Theorem 1. Because the program remains nonconvex, global optimality is not guaranteed; the closed-loop certificate therefore rests on an unstated numerical-accuracy hypothesis. A short corollary that makes the admissible (δ_J, δ_K) explicit in terms of the Lyapunov margin would close this gap.","section":null}],"minor_comments":[{"comment":"Assumption 1 (finite-memory input-output realization) is classical but non-trivial for general nonlinear systems; a short pointer to concrete classes (NARX, systems with well-defined relative degree) already present in Remark 1 could be moved into the assumption statement itself for clarity.","section":null},{"comment":"In the vehicle example the stacked input cost uses both ru and rΔu, while the theoretical stage cost is separable; a one-sentence remark that the analysis extends verbatim to any positive-definite quadratic form on the stacked input would avoid confusion.","section":null},{"comment":"Figure 3 (middle panel) shows that tracking error is non-monotonic in the update period T0; a brief discussion of how T0 should be chosen in practice would strengthen the online section.","section":null},{"comment":"Notation for the stacked windows (x[t1,t2]) and the multi-index a of the multi-step maps is dense; a short table of symbols would help the reader.","section":null},{"comment":"The conference precursor [27] is cited; it would be useful to state explicitly which theorems are new relative to that version (the robust and online results appear to be the main additions).","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical core is solid and the online ALD extension is a genuine contribution. The non-constructive character of the kernel-error hypothesis is the only load-bearing issue; once the authors either supply a data-driven residual certificate or clearly rephrase the theorems as pure existence statements, the paper should be acceptable. I do not see a fatal flaw that would justify rejection."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new piece is the closed-loop theory. They take a kernel multi-step predictor (representer theorem), put it inside a robust DDPC-style problem without terminal ingredients, and actually prove recursive feasibility plus practical stability for the nominal, noisy, and slowly time-varying cases. That package was missing from the kernel/GP-MPC and nonlinear DDPC literature they cite.\n\nWhat they do well is the candidate-solution construction. Theorem 1 is careful: shift the optimal sequence, append the local stabilizer once a state enters the local region, bound the cost excess by functions of d-bar, and extract an explicit horizon lower bound. The robust and online extensions reuse the same pattern with a unified uncertainty scale and an ALD-managed fixed-budget dictionary. The two numerical examples (vehicle + payload-carrying A1) are honest about solver quality and update period; they do not oversell.\n\nThe stress-test concern about d-bar is real but not fatal. Remark 2 already says Gamma and the uniform error bound are generally intractable for black-box plants and that only order-of-magnitude tuning is needed. That is the usual situation in learning-based MPC: the existence claim is non-constructive from data alone. It does not make the Lyapunov decrease circular; the kernel error is treated as an additive disturbance absorbed into the practical residual. Same story for the non-convex program: they require a sufficiently accurate local solution and show the residual can be absorbed. No code or data release is a minor practical minus.\n\nThis is for people already working on data-driven or learning-based MPC who want the missing stability arguments for kernel multi-step predictors. The math and citation pattern look solid; the free parameters are the usual ones. I would send it to peer review. Engage if that is your lane; the proofs are worth reading even if you stay skeptical of the a-priori bounds.","headline":"Solid recursive-feasibility proofs for kernel multi-step DDPC; the uncertifiable d-bar is a real but standard soft spot, not a collapse of the argument.","tokens_in":30443,"tokens_out":475,"would_cite":true,"duration_ms":5351,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Kernel multi-step predictors built only from input-output data give recursive feasibility and practical closed-loop stability for nonlinear predictive control when the horizon is long enough and representation error is small enough.","keywords":["data-driven control","kernel methods","model predictive control","nonlinear systems","recursive feasibility","practical stability","online dictionary","approximate linear dependence"],"falsifier":"On a compact nonlinear plant for which a high-fidelity multi-step residual bound can be computed offline, run the controller with a horizon shorter than the derived threshold or with an intentionally larger residual; recursive feasibility or the practical Lyapunov decrease should then fail.","tokens_in":30392,"feed_emoji":"⚙️","tokens_out":590,"duration_ms":5407,"temperature":0.7,"pith_summary":"The paper develops a data-driven kernel-based predictive controller that never needs an explicit state-space model of an unknown nonlinear plant. Past input-output windows are turned into a multi-step predictor via the representer theorem, so every future output element is written as a finite kernel expansion whose residual is bounded. When that residual is small and the prediction horizon is long enough, the resulting receding-horizon scheme is shown to be recursively feasible and practically stable about a desired equilibrium, without terminal costs or terminal sets. The same argument is extended to measurement noise by folding noise and representation error into one uncertainty radius, and to slowly time-varying plants by refreshing a fixed-size kernel dictionary online under an approximate-linear-dependence rule. The result matters because it supplies the first closed-loop guarantees for kernelized nonlinear data-driven MPC that work from pure input-output trajectories and remain valid under mild non-stationarity.","feed_headline":"Kernel predictors give stable data-driven control of nonlinear plants","feed_subtitle":"Long enough horizons and small residual error yield recursive feasibility without terminal sets","key_machinery":"The implicit multi-step kernel predictor obtained from the representer theorem: each future output coordinate is written as a linear combination of kernel evaluations on the offline (or online) dictionary, with a deterministic residual bound that is treated as a single additive uncertainty inside a robust data-driven MPC problem.","core_discovery":"For noise-free data the DDKPC optimization is recursively feasible and the closed-loop system is practically stable whenever the prediction horizon exceeds an explicit length that depends only on the sublevel set of the cost and the kernel residual is smaller than a matching threshold; the same practical Lyapunov decrease continues to hold after the residual is enlarged to absorb bounded measurement noise or the bounded-rate drift of a slowly time-varying plant.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Long horizons plus tiny kernel residual lock in DDKPC recursive feasibility","Kernel residual threshold and horizon length certify practical closed-loop stability","Unified residual absorbs noise while preserving DDKPC Lyapunov decrease","ALD online dictionary keeps practical stability under slow plant drift","Noise-free DDKPC is recursively feasible without terminal ingredients"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Every multi-step input-output map must live in a known reproducing-kernel Hilbert space whose norm is bounded by a known constant, so that a uniform residual size can be certified before the controller is run.","fun_headline_variants_meta":{"raw":{"variants":["Long horizons plus tiny kernel residual lock in DDKPC recursive feasibility","Kernel residual threshold and horizon length certify practical closed-loop stability","Unified residual absorbs noise while preserving DDKPC Lyapunov decrease","ALD online dictionary keeps practical stability under slow plant drift","Noise-free DDKPC is recursively feasible without terminal ingredients"]},"model":"grok-4.5","effort":"low","cost_usd":0.00518,"raw_usage":{"total_tokens":1452,"prompt_tokens":787,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":51800000,"prompt_tokens_details":{"text_tokens":787,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":598,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":787,"tokens_out":67,"duration_ms":4647,"temperature":1.0,"reasoning_tokens":598,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:36:01.028590+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a compact nonlinear plant for which a high-fidelity multi-step residual bound can be computed offline, run the controller with a horizon shorter than the derived threshold or with an intentionally larger residual; recursive feasibility or the practical Lyapunov decrease should then fail.","supporting_citations":[],"review_version":1}