{"id":"0dd149ab-f011-4b99-a77c-c030e0949849","arxiv_id":"2411.18762","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Kernelized velocity-form models yield an offset-free data-driven predictive controller for nonlinear systems, learned by least squares and supported by feasibility and stability conditions.","lead":"This paper combines kernel-based learning with an incremental velocity model to build a data-driven predictive controller for nonlinear systems that aims to track references without steady-state offset. It learns the needed nonlinear terms with least squares, keeps the controller solvable by sequential quadratic programming, and demonstrates the idea on a nonlinear pendulum benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3's proof uses absolute inputs where the velocity model requires increments and omits the reference in the terminal-invariance step, so recursive feasibility and stability are not established as written.","rationale":"I focused on the proof of Theorem 4.3 because it is the load-bearing component for the paper's central claim of a framework for recursive feasibility and stability. The exactness assumption and terminal-LMI verification are important and are acknowledged in the paper as conditions, but the proof error is an internal flaw: it uses the wrong input variable (absolute u instead of Δu) and omits the reference in the terminal invariant-set argument. Both defects are concrete and can be checked by re-derivation or simulation. If the proof cannot be repaired, the theoretical contribution reduces to a heuristic scheme with an offset-free simulation, but no guarantees. The reader already flagged an 'algebraic inconsistency in the terminal step'; my concern is more specific and indicates that the displayed terminal equality is false unless r=0. This does not necessarily doom the approach, as the theorem may be fixable, so I recommend keeping the conditional verdict pending the re-derivation.","tokens_in":11885,"tokens_out":17812,"duration_ms":141274,"concrete_test":"Independently re-derive the shifted terminal state for the velocity model with explicit sums over Δu: starting from z_{1|k} with increments Δu_{1|k},…,Δu_{N−1|k},K(z_{N|k}−r), compute z_{N|k+1} and verify whether z_{N|k+1}−r equals A_cl(ρ̃)(z_{N|k}−r) with A_cl as in Assumption 4.2. Then simulate the proposed shifted sequence on the exact analytic velocity model from [2] for the pendulum benchmark with a nonzero reference (e.g., r=0.5 as in Fig. 3) and check whether the terminal constraint (z_{N|k+1}−r)∈Z_T is satisfied. If the algebraic re-derivation differs from the proof's final line, or the simulation violates the terminal constraint, the proof's construction is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central stability framework rests on Theorems 4.3–4.4, but the proof of Theorem 4.3 as printed is not a valid proof for the vKDPC problem. In Problem 4.1 the prediction model is z_{k+1} = A(ρ_k)z_k + B(ρ_k)Δu_k, so the decision variables are increments Δu, and the constraints in (15d) are on (z_k, Δu_k). The proof, however, constructs the shifted input sequence in absolute control inputs ũ_{k+1} = {u∗_{1|k},…,u∗_{N−1|k}, u∗_{N−1|k}+K(z∗_{N|k}−r)} and then applies the convolution formula of (13)–(14) as if the inputs were absolute. The correct shifted increment sequence is {Δu∗_{1|k},…,Δu∗_{N−1|k}, K(z∗_{N|k}−r)}. Moreover, the terminal state update under the feedback Δu = K(z−r) is A z + B K(z−r), which equals (A+BK)z only when r=0; the proof writes A_cl z∗_{N|k} = A(ρ)z∗_{N|k} + B(ρ)K(z∗_{N|k}−r) and then claims this lies in Z_T. But Assumption 4.2 defines Z_T in the error space (z−r) and requires invariance of A_cl Z_T ⊆ Z_T, i.e., the proof must show z̃_{N|k+1}−r ∈ Z_T. The displayed line omits the reference subtraction and therefore does not establish the terminal constraint. Since Theorem 4.3 underpins both recursive feasibility and the Lyapunov argument in Theorem 4.4, the theoretical guarantee is not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven predictive control scheme for nonlinear systems based on a kernelized velocity-form state-space model. The unknown gradients in the velocity model are parameterized with kernel functions and learned by least squares, and the resulting controller is formulated as a nonlinear program solved by sequential quadratic programming. The authors state terminal cost and set conditions intended to guarantee recursive feasibility and Lyapunov stability, and they illustrate the method on a discretized pendulum example, comparing it with an analytic velocity-form NMPC.","tokens_in":12281,"tokens_out":13270,"duration_ms":116065,"significance":"The main idea, preserving the velocity-model structure so that a fixed-dimensional kernelized predictor can be learned by a closed-form least-squares problem, is attractive, and the simulation results are encouraging. The least-squares derivation in Lemma 3.3 is clean for noiseless data, and the authors are explicit about the exactness assumption on the learned model. If the theoretical gaps in Section IV were repaired, the paper would be a useful contribution to data-driven nonlinear MPC. As it stands, however, the recursive feasibility and stability claims are not established as written, and the numerical example does not verify the assumptions under which those claims are proven.","major_comments":[{"comment":"The proof of Theorem 4.3 defines the shifted sequence tilde u_{k+1} in absolute control inputs, but the vKDPC problem in (15) optimizes over incremental inputs and the prediction matrices (13)-(14) act on Delta u. The displayed derivation then applies the convolution formula to absolute inputs, which is inconsistent with the model. The correct candidate is the incremental sequence {Delta u*_{1|k}, ..., Delta u*_{N-1|k}, K(z*_{N|k}-r)}. Consequently, recursive feasibility of Problem 4.1 at time k+1 is not established by the proof as written.","section":"Section IV, Theorem 4.3 proof"},{"comment":"Assumption 4.2 and the terminal step of the proof of Theorem 4.3 are not stated in consistent coordinates. The terminal constraint is (z_{N|k}-r) in Z_T, so Z_T is a set of deviations, while the proof concludes that tilde z_{N|k+1} = A_cl(rho) z*_{N|k} lies in Z_T. Since A(rho)r = r for the velocity model (2), one must instead show that A_cl(rho)(z*_{N|k}-r) lies in Z_T. The proof omits the subtraction of r and therefore does not establish the terminal constraint. Relatedly, the inclusion Z_T subset Z in Assumption 4.2 is not dimensionally consistent if Z_T is a set of deviations and Z is a set of absolute states; the condition should be {r+e : e in Z_T} subset Z, or Z should be redefined in error coordinates.","section":"Section IV, Assumption 4.2 and Theorem 4.3"},{"comment":"Before Assumption 4.2, the paper assumes that the kernelized velocity dynamics hat A, hat B, hat C are an exact approximation of the true velocity dynamics. For a finite sample of noiseless data from an unknown system, this exactness assumption is generally false. Theorems 4.3 and 4.4 are proved only for this exact representation, and Remark 4.5 merely defers the inexact case to robust methods in [18]. The simulation in Section V does not verify exactness or the terminal LMI over the computed terminal set, so the theoretical guarantees do not apply to the controller actually implemented in the example. The paper should either provide a quantified approximation bound and adapt the stability analysis, or clearly present the numerical results as purely empirical.","section":"Section IV, exactness assumption"},{"comment":"In Remark 4.6 and Section V, the terminal ingredients P, K, and Z_T are computed from the linearized dynamics at the reference, i.e., (hat A(x_r,u_r)+hat B(x_r,u_r)K). Assumption 4.2, however, requires the nonlinear, state-dependent LMI and set invariance to hold for all e in Z_T. No argument is given that the computed terminal set satisfies the nonlinear conditions, for example via Lipschitz bounds or a grid verification. This gap is load-bearing because the recursive feasibility and stability theorems rely on Assumption 4.2.","section":"Section V and Remark 4.6"}],"minor_comments":[{"comment":"The full rank hypothesis is ill-posed: the matrices [Ky_1 ... Ky_{s-1}] and [Kx_1 ... Kx_{s-1}] have dimensions (n s) x (s-1) and ((n+m)s) x (s-1), so they cannot have full row rank for s>1; if full column rank is meant, the least-squares solution is not unique and the pseudoinverse formulas give only the minimum-norm minimizer. Please clarify.","section":"Lemma 3.3"},{"comment":"In the proof of Theorem 4.4, the expression (z*_{N|k}-r)(Q+K^T R K)(z*_{N|k}-r) should be a quadratic form with transposes: (z*_{N|k}-r)^T (Q+K^T R K) (z*_{N|k}-r).","section":"Theorem 4.4 proof"},{"comment":"The claim of equal computational complexity should be qualified: the optimization variables have the same dimensions as in the analytic velocity model, but evaluating the kernel matrices has cost O(s) per entry.","section":"Remark 3.2"},{"comment":"There are several typos, e.g., 'dyanmics' in Remark 4.6 and 'a offset-free' in the abstract; also, Fig. 2 would benefit from labeled axes and a quantitative validation metric for the kernel model.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and addresses a relevant problem, but the stability analysis needs substantial revision. The proof errors in Theorem 4.3 are fixable, but the exactness assumption is a more fundamental limitation. I would encourage the editor to request a major revision with a corrected proof and either a verified terminal design or an explicit statement that the guarantees are only for exact velocity representations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper does something genuinely useful: it turns the kernelized velocity-form model from [13] into a practical offset-free DPC scheme by adding an explicit least-squares learning step and terminal ingredients for recursive feasibility and stability. The learning problem in Lemma 3.3 is clean, and the dimension-preserving structure means the online cost stays at the level of the analytic velocity model. The pendulum comparison against the analytic vNMPC is the right kind of sanity check, and the paper is honest about the exact-model assumption in the theoretical part.\n\nThe soft spot is real, though. Theorem 4.3's proof, as printed, is not a valid proof for the stated problem. The prediction model uses increments Δu, but the shifted candidate sequence is written in absolute inputs and plugged into the convolution formula as if it were built for u. The terminal step also fails to subtract the reference r before invoking invariance of Z_T, which is defined in error coordinates. Since Theorem 4.4 leans on Theorem 4.3, the formal stability guarantee isn't established. I agree with the stress-test note on this.\n\nThis is repairable: shift the increments, write the candidate tail with K(z_{N|k}-r), and show z̃_{N|k+1}-r ∈ Z_T. The proof structure is standard and I'd expect it to go through. But as it stands, the main theorem is not proven.\n\nTwo smaller issues. The 'exact approximation' assumption in Section IV is doing a lot of work, and the offset-free property is demonstrated in simulation but not proven for the stated disturbance model. The paper also cites [14] as very recent related work but doesn't benchmark against it; given how close that work is, a quantitative comparison would materially strengthen the paper. No code or data is shipped, which makes reproduction harder.\n\nOverall: the core idea is good, the learning is sound, and the gap is fixable. The paper deserves a serious referee, but the referee should insist on a corrected proof and a robustness discussion. I wouldn't cite it in its current form.","headline":"Useful data-driven MPC extension with a clean least-squares learning step, but the stability proof has a load-bearing gap that needs repair before the guarantees can be trusted.","tokens_in":12811,"tokens_out":3120,"would_cite":false,"duration_ms":34531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that a kernelized velocity-form data-driven predictive controller can achieve offset-free tracking for unknown nonlinear systems, with recursive feasibility and Lyapunov stability guarantees under an exactness assumption.","keywords":["data-driven predictive control","offset-free control","kernel methods","velocity form model","nonlinear systems","recursive feasibility","stability","sequential quadratic programming"],"falsifier":"For the pendulum example, recompute the terminal set $Z_T$ from the kernelized model and test whether $A_{cl}(\\rho) Z_T \\subseteq Z_T$, $K Z_T \\subseteq \\Delta U$, and the LMI of Assumption 4.2 hold at a dense grid of points inside $Z_T$; any violation would invalidate the recursive feasibility proof, since the paper does not verify Assumption 4.2 over the whole set.","tokens_in":11628,"feed_emoji":"🎛️","tokens_out":6395,"duration_ms":52112,"temperature":0.7,"pith_summary":"The paper tries to show that a data-driven predictive controller for nonlinear systems can be made offset-free without knowing the plant model, by learning a velocity (incremental) representation with kernel functions. Preserving the structure of an analytic velocity model reduces the learning step to a linear least squares problem, which is far simpler than the nonlinear optimization used in comparable Koopman-based approaches. The resulting controller, called vKDPC, is a nonlinear program that can be solved with sequential quadratic programming, and the paper gives terminal conditions under which recursive feasibility and Lyapunov stability hold. On a pendulum benchmark, the data-driven controller tracks piecewise-constant references with zero offset after disturbances, matching an analytic velocity-form NMPC in performance. If true, this provides a practical route to offset-free predictive control for unknown nonlinear systems with stability guarantees.","feed_headline":"Kernelized velocity model yields offset-free data-driven MPC","feed_subtitle":"A least-squares kernel fit of incremental dynamics matches model-based NMPC on a pendulum, with stability guarantees.","key_machinery":"The central object is the velocity form (incremental) state-space model, where the state is $z_k = \\mathrm{col}(y_{k-1}, \\Delta x_k)$ and the dynamics matrices $A$, $B$, $C$ depend nonlinearly on scheduling variables. The paper's key move is to parameterize each unknown gradient matrix ($\\nabla_x f$, $\\nabla_u f$, $\\nabla_x h$) as a product of a constant coefficient matrix and a kernel basis matrix built from the data. Because these matrices enter linearly, the representer theorem applies and the coefficients are obtained by least squares. This kernelized velocity model has the same dimension as the analytic model, and the prediction matrices $\\Psi$ and $\\Gamma$ are formed by products of these kernel matrices, so the MPC problem can be solved with sequential QP by fixing the scheduling variables and iterating.","core_discovery":"The central claim is that the structure of the analytic velocity form can be preserved exactly when functions are replaced by kernel-based representations, because the unknown gradient matrices appear linearly. This turns model learning into two independent least-squares problems: one for the output-gradient coefficients and one for the state/input-gradient coefficients. The learned model has the same state dimension as the analytic velocity model, so the online predictive control problem remains a parameterized nonlinear program solvable by sequential QP. The paper then states Assumption 4.2, a terminal cost and terminal set condition involving a locally stabilizing feedback and an LMI, and proves that under exact kernel representation this yields recursive feasibility and asymptotic Lyapunov stability of the closed loop. Simulations on a discretized pendulum show offset-free tracking for piecewise-constant references under piecewise-constant disturbances, with performance essentially identical to the analytic velocity-form NMPC.","pith_inferences":["A natural next test is whether the least-squares kernel model remains accurate when the data are noisy; the paper assumes noiseless data, and the offset-free property in that setting is not established.","The same structural preservation could be applied to other basis functions, e.g., polynomials or neural features, where the linear-in-parameters property holds, potentially giving offset-free data-driven MPC beyond kernels.","The proof of recursive feasibility depends on the kernel model being exact; a practical user would need to verify Assumption 4.2's LMI over the entire terminal set, or rely on the robust variant, before deploying vKDPC safely.","Because the scheduling variables $\\rho$ are fixed during each QP step, the convergence of Algorithm 1 to a local optimum of Problem 4.1 is not guaranteed; a reader could test whether the 4\\text{--}5 average iterations reach the same cost as the analytic vNMPC on harder nonlinearities."],"forward_implications":["Any nonlinear system whose velocity-form gradient functions lie in the RKHS of the chosen kernel can be controlled offset-free from noiseless data alone, without solving a nonlinear identification problem.","The terminal cost and terminal set conditions (Assumption 4.2) apply to both analytic and data-driven velocity-form MPC, replacing the more conservative terminal equality constraint used in earlier velocity-form NMPC.","Because the kernelized model has the same dimension as the analytic model, the online computation cost of vKDPC is comparable to that of model-based velocity-form NMPC; in the pendulum example it averaged 0.0683 s per sample versus 0.0438 s.","The learning step reduces to solving two least-squares problems, and the paper reports 2.0 s to fit a 2000-sample dataset for the pendulum, so the approach scales linearly with data size in the coefficient matrices.","When the kernel model is inexact, the guarantees degrade from asymptotic stability to input-to-state stability, and recursive feasibility can be recovered by leaving the initial extended state as a free optimization variable, per the robust framework referenced in the paper."],"supporting_citations":[{"why":"Supplies the analytic velocity-form state-space model whose structure is preserved by the kernelized representation.","marker":"[2]"},{"why":"Shows that incremental inputs in data-driven predictive control yield offset-free tracking, the effect the paper extends to nonlinear systems.","marker":"[5]"},{"why":"Provides the generalized representer theorem that justifies parameterizing the unknown gradient functions as kernel expansions with finite data.","marker":"[8]"},{"why":"Gives universal kernel functions, which the paper invokes to justify the full-rank condition needed for the least-squares solution.","marker":"[17]"},{"why":"Supplies the robust recursive feasibility and input-to-state stability framework used when the kernelized velocity model is inexact, and the terminal-set computation method used in the simulations.","marker":"[18]"},{"why":"Baseline Koopman-based data-driven nonlinear MPC that requires nonlinear least squares, contrasted with the paper's linear least squares approach.","marker":"[6]"},{"why":"Recent kernel-based multi-step velocity-form predictors for nonlinear systems, the closest related approach.","marker":"[14]"}],"fun_headline_variants":["Kernelized velocity model yields offset-free MPC","Data-driven MPC with kernel least squares: no offsets","Kernel method in MPC: offset-free tracking guaranteed","Nonlinear data-driven control via kernelized models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The kernelized velocity model exactly represents the true velocity dynamics; if the finite-sample kernel model is only approximate, the paper's recursive feasibility and asymptotic stability guarantees do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Kernelized velocity model yields offset-free MPC","Data-driven MPC with kernel least squares: no offsets","Kernel method in MPC: offset-free tracking guaranteed","Nonlinear data-driven control via kernelized models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000121,"raw_usage":{"total_tokens":1033,"prompt_tokens":826,"completion_tokens":207,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":146}},"tokens_in":442,"tokens_out":207,"duration_ms":3138,"temperature":1.0,"reasoning_tokens":146,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:55:50.325075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the pendulum example, recompute the terminal set $Z_T$ from the kernelized model and test whether $A_{cl}(\\rho) Z_T \\subseteq Z_T$, $K Z_T \\subseteq \\Delta U$, and the LMI of Assumption 4.2 hold at a dense grid of points inside $Z_T$; any violation would invalidate the recursive feasibility proof, since the paper does not verify Assumption 4.2 over the whole set.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the analytic velocity-form state-space model whose structure is preserved by the kernelized representation."},{"cited_title":"Offset–free data–driven predic tive control,","cited_arxiv_id":null,"evidence_quote":"Shows that incremental inputs in data-driven predictive control yield offset-free tracking, the effect the paper extends to nonlinear systems."},{"cited_title":"A generalized r epresenter theorem,","cited_arxiv_id":null,"evidence_quote":"Provides the generalized representer theorem that justifies parameterizing the unknown gradient functions as kernel expansions with finite data."},{"cited_title":"Universal kernels,","cited_arxiv_id":null,"evidence_quote":"Gives universal kernel functions, which the paper invokes to justify the full-rank condition needed for the least-squares solution."},{"cited_title":"Data-Driven quasi- LPV Model Predictive Control Using Koopman Operator Techni ques,","cited_arxiv_id":null,"evidence_quote":"Baseline Koopman-based data-driven nonlinear MPC that requires nonlinear least squares, contrasted with the paper's linear least squares approach."},{"cited_title":"Kernel-based multi-step predictors for data-driven analysis and control of nonlinear systems through the velocity form","cited_arxiv_id":"2408.00688","evidence_quote":"Recent kernel-based multi-step velocity-form predictors for nonlinear systems, the closest related approach."}],"review_version":1}