{"id":"bf966c55-a1d3-4238-b8f2-cb8771818b3f","arxiv_id":"2501.17500","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"By restructuring the product-kernel Gram matrix, the authors derive a computationally efficient nonlinear data-enabled predictive controller that runs much faster than stacked-kernel baselines.","lead":"This paper introduces a kernel-based method for nonlinear data-enabled predictive control that models system trajectories as operators in a product reproducing kernel Hilbert space. The key advance is a reformulation that exploits the Kronecker product structure to cut computation time, allowing larger datasets and better control performance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Problem 4.4 is not an equivalent reduction of KerODeePC: Lemma 4.1's left-multiplication is non-invertible, and the projection \\hat g=Y†\\tilde g changes the prediction model unless Lemma 4.3's condition holds, which is unverified and generically false.","rationale":"The reader's weakest assumption concerns the universality of the product RKHS inherited from [18]. That is a real concern about asymptotic expressiveness, but it is not the most load-bearing issue for the central computational claim: even if the universality theorem were accepted, the efficient formulation still solves a different optimization problem than Problem 3.2. The false equivalence in Lemma 4.1 and the unverified condition in Lemma 4.3 are concrete, internal, and testable. If Lemma 4.3's condition fails, the speed comparison is between a constrained operator-learning problem and a relaxed output-fitting problem, so the 100× speedup does not establish the claimed advantage for the same method. This does not require rejecting the paper; the authors could verify the equivalence condition, report λ, or explicitly reframe Problem 4.4 as a relaxation with an error bound. Since the reader's verdict is already CONDITIONAL, my concern supports that verdict rather than changing it, but for a different reason than the one stated in the reader's weakest_assumption.","tokens_in":11036,"tokens_out":17011,"duration_ms":174899,"concrete_test":"Reproduce the T=400 Van der Pol setup. (1) For 100 random x in X, compute a basis of N(Ω(x)) and report max_{g∈basis} ‖Yg‖; if it exceeds numerical tolerance, Lemma 4.3's equivalence condition fails. (2) Solve Problem 4.4 with the exact λ used in the paper (report λ and solver tolerances) on a validation set of 50 input/state pairs, and compare the predicted y to y⊗=YK⊗^{-1}k⊗(u,x); if the median absolute error exceeds 1e-6, the efficient controller is not implementing the operator predictor. (3) Re-run the control comparison with λ→∞ and with λ=0; if tracking changes significantly, the finite-λ behavior is doing the work, not the learned operator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central speed claim compares Problem 4.4 against a stacked formulation of Problem 3.2. For that comparison to be meaningful, Problem 4.4 should be an equivalent, or explicitly controlled, approximation of the same operator model. It is neither. Starting from (7a), K⊗g = ku(u)⊗kx(x), left-multiplying by I_Tu⊗kx(x)^⊤/‖kx(x)‖² gives (11a), but the converse only yields (I_Tu⊗P_{kx})K⊗g = ku(u)⊗kx(x), not K⊗g = ku(u)⊗kx(x); hence (11) admits extra solutions. Lemma 4.3 would restore output equivalence if Y\\hat g=0 for all \\hat g∈N(Ω(x)), but with full-rank Ku and Kx, dim N(Ω(x)) = Tu(Tx−1); for the T=400 experiment this is 380 dimensions in R^400, and the 10-row matrix Y would have to annihilate that entire subspace. This condition is generically false and is never checked. Problem 4.4 then further replaces \\hat g by Y†\\tilde g, which is exact only when \\hat g lies in row(Y); no argument shows the relevant solutions lie there, nor that ΩY†\\tilde g=0 preserves enough freedom. The only formal connection to the original operator prediction is the λ→∞ limit, while the finite λ used in the experiments is not reported. Thus the reported speedup may be for a different, relaxed controller, not for the kernelized operator DeePC problem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a kernelized operator approach to nonlinear data-enabled predictive control. It models the input-output behavior of a nonlinear system as a product reproducing kernel Hilbert space operator G(u)(x) = Θ*k⊗(u,x), building on a result from a companion preprint. The KerODeePC problem (Problem 3.2) enforces K⊗g = k⊗(u,x) and Yg = y. To reduce computation, Lemma 4.1 rewrites the first constraint into a lower-dimensional equality Ω(x)g = ku(u), and Lemma 4.3 gives a condition under which the reduced system is output-equivalent. Problem 4.4 further projects the decision variable onto the output space, resulting in an efficient formulation. Numerical experiments on a Van der Pol oscillator show order-of-magnitude speedups relative to a stacked kernel DeePC and comparable performance to NMPC when more data are used.","tokens_in":11437,"tokens_out":10516,"duration_ms":91600,"significance":"The structural idea of factorizing the kernel into input and state parts to reduce the equality constraints from TxTu to Tu is original and potentially valuable. The numerical results are impressive: a reduction from 33.87s to 0.33s per control action at T=400 and the ability to handle T=10000 online. The paper also contributes a systematic data-generation procedure. However, the value of the contribution depends critically on whether Problem 4.4 is a faithful and controllable approximation of Problem 3.2; the current manuscript does not establish this. If the equivalence issues are resolved, the method could be a practical tool for large-scale kernel-based predictive control.","major_comments":[{"comment":"Lemma 4.1's claim that (11) is 'equivalently re-written' from (7) is too strong. The left-multiplication by the pseudo-inverse (I_Tu ⊗ kx(x))† is not invertible: from (11a) one can only conclude that P_{range(I⊗kx)} K⊗ g = (I⊗kx)ku(u), not that K⊗g = k⊗(u,x). The reduced system (11) therefore admits extra solutions, namely all g ∈ K⊗^{-1}(range(I⊗kx))^⊥. Lemma 4.3 correctly identifies the necessary and sufficient condition for output equivalence (Y annihilates N(Ω(x))), but this condition is not verified anywhere in the manuscript and is generically false for the reported data dimensions. For the T=400 example with Tx=20 and Tu=20, dim N(Ω(x)) = Tu(Tx-1) = 380 while Y ∈ R^{10×400}; the requirement that a 10-row matrix annihilate a 380-dimensional subspace is a restrictive condition that should be checked numerically. The paper neither checks it nor discusses its plausibility, so the claim that Problem 4.4 preserves the original operator model is not supported.","section":"Section IV, Lemma 4.1 and Lemma 4.3"},{"comment":"Problem 4.4 is not an equivalent reformulation of Problem 3.2. It replaces the decision variable ĝ by Y†˜g, which restricts the search to row(Y), and it changes the regularization term from ||g||² to ||˜g||². Moreover, the constraint (15b) is only a relaxation of the requirement that ĝ ∈ N(Ω(x)); it enforces that Ω(x)Y†˜g=0, i.e., that the output-space variable lies in Y N(Ω(x)). The only formal connection to Problem 3.2 established in the text is the λ→∞ limit, where ˜g→0 and y tends to the unregularized operator predictor. Since the finite value of λ used in the experiments is not reported, the speedups in Table I cannot be attributed to solving the original KerODeePC problem; they are for a modified controller. The manuscript should either provide a bound on the suboptimality/approximation error for finite λ, or explicitly present Problem 4.4 as a relaxation and study its closed-loop behavior as a function of λ.","section":"Section IV, Problem 4.4 and the paragraph after (15)"},{"comment":"The paper's central modeling assumption is that the product RKHS H(k⊗, X×U) is universal in the space of nonlinear system operators and that the minimizer of (9) is a universal approximator. This theorem is imported from the companion preprint [18] and is neither stated with its assumptions nor proven in the present manuscript. The paper also does not verify the conditions of that theorem for the Van der Pol oscillator and the data generated by Algorithms 1 and 2. If the universality result does not apply, the prediction model (10) is biased and the claimed nonlinear generalization is unsupported. At minimum, the authors should state the theorem precisely (including the required assumptions on the system, the kernel, and the data) and indicate how their data-generation procedure meets those assumptions. Relying on an unpublished self-cited preprint for such a load-bearing step is risky for the reader.","section":"Section II, eq. (5)-(7); Section III-A"}],"minor_comments":[{"comment":"The definition of 1n is garbled: 'A square n × n matrix where all elements of 1/n is denoted by 1n' should read 'A square n × n matrix whose elements are all equal to 1/n is denoted by (1/n)1_n1_n^T' or similar, and the vector 1_n should be defined separately.","section":"Section I, Notation"},{"comment":"Equations (3) and (4) use k(z1, z1) on the left-hand side; they should be k(z1, z2).","section":"Section II, eq. (3) and (4)"},{"comment":"In Definition 2.2, item (i) says 'k(z, ·) ∈ H(k, X)' but the space should be H(k, Z); likewise item (ii) should refer to H(k, Z).","section":"Definition 2.2"},{"comment":"The sentence 'when λ → ∞, we have that ˜gk → 0 and thus ˆgk → 0 by Lemma 4.1' incorrectly invokes Lemma 4.1; the implication follows from the definition ˆg = Y†˜g, not from Lemma 4.1.","section":"End of Section IV"},{"comment":"Constraint (15c) contains an extra closing parenthesis after 'u[0,N−1](k)))'; it should read 'YΩ†(xk)ku(u[0,N−1](k)) + ˜gk = y[1,N](k)'.","section":"Problem 4.4, eq. (15c)"},{"comment":"The entry 'X' for the stacked Gram inversion at T=10000 could be clearer: state explicitly that the stacked formulation is not solvable online at that size, and give the corresponding product-kernel time in the same row.","section":"Section V, Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own preprint [18] for the universal approximation theorem. Given that the present manuscript's theory depends on it, the editor may wish to require that the theorem be either proved in an appendix or that [18] be formally published and accessible. Additionally, the experiments omit the value of λ, which is essential for interpreting the relationship between Problem 4.4 and Problem 3.2; a sensitivity study over λ would help. The computational comparison with the 'stacked kernel formulation' is fair in spirit, but because Problem 4.4 is a different optimization problem, the reported speedup conflates algorithmic reduction with problem relaxation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Start with the punchline: the paper has a genuinely useful computational observation. The product kernel structure means the Gram matrix is a Kronecker product, so inversion and construction cost polynomial in Tx and Tu separately rather than in the product TxTu. That idea is sound, and the k-means data-generation scheme is a reasonable practical addition. The problem is that the \"efficient formulation\" presented as the main contribution is not an equivalent reduction of the original KerODeePC problem, and the speed comparison is between two different optimization problems.\n\nThe algebra in Lemma 4.1 is correct as a manipulation, but the word \"equivalently\" is doing too much work. Multiplying K⊗g = ku(u)⊗kx(x) by the pseudo-inverse of (I⊗kx(x)) gives a necessary condition, not a sufficient one. The extra solutions are exactly the nullspace of Ω(x), and Lemma 4.3 states the right condition for model equivalence: Y must annihilate that nullspace. The paper never checks that condition. In the T=400 experiment, the nullspace has dimension 380 inside R^400 and Y has only 10 rows, so generically the condition is false. Then Problem 4.4 goes further and restricts ĝ to lie in the row space of Y via ĝ=Y†g̃, with no argument that the relevant solutions live there. The only formal bridge back to the original problem is the λ→∞ limit, and the paper does not report the finite λ used in the experiments. So the 100x speedup is real, but it is the speedup of a relaxed controller, not of solving Problem 3.2 faster.\n\nThe second soft spot is the foundation: the universality of the product RKHS is imported from the same group's preprint [18], and the approximation error is not quantified here. That is not fatal on its own, but it means the prediction model's bias is unquantified.\n\nOn the positive side, the writing is clear, the Kronecker trick is a legitimate contribution, and the empirical trends—more data helps, prediction error comparable to the stacked kernel baseline—look plausible. But the comparison is unfair: the stacked baseline solves a much larger problem. A revision should either verify the nullspace condition, or explicitly present Problem 4.4 as a heuristic relaxation and benchmark it against a same-problem baseline, not against a different formulation.\n\nMy recommendation: send it to peer review, but expect major revision. The core idea is worth keeping; the equivalence claims need to be scaled back and the experiments need to be redone or reframed.","headline":"A useful Kronecker-product computational trick buried under an overstated equivalence claim and an unfair speed comparison.","tokens_in":11938,"tokens_out":5648,"would_cite":false,"duration_ms":53685,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper turns nonlinear DeePC into a low-dimensional one-shot optimization by parameterizing the system with a product kernel, cutting per-action computation by about a factor of one hundred.","keywords":["data-enabled predictive control","DeePC","reproducing kernel Hilbert space","product kernel","kernelized operator","data-driven control","universal kernel","operator learning"],"falsifier":"Take a nonlinear benchmark with a non-smooth or discontinuous map (for example a relay or dead-zone nonlinearity), generate data with Algorithm 2, and check whether the multi-step prediction error (18) keeps decreasing as $T_x$ and $T_u$ grow; if the error plateaus above the measurement noise, the universal-approximation premise is violated for that system.","tokens_in":10833,"feed_emoji":"⚙️","tokens_out":9018,"duration_ms":74860,"temperature":0.7,"pith_summary":"This paper proposes a kernelized operator formulation of data-enabled predictive control (DeePC) for nonlinear systems. Instead of using a kernel on the stacked space of initial states and future inputs, it learns the input-to-output map as an operator in a product reproducing kernel Hilbert space, with a product kernel that factors into a state kernel and an input kernel. The central claim is that this product structure shrinks the online optimization: the equality constraints drop from the product of the two dataset sizes to the number of input trajectories, and the decision variables are projected into the output space. In the paper's Van der Pol example, each control action takes about 0.33 seconds instead of 33.87 seconds for the same data, and datasets with 10,000 trajectories become usable online. If the claim holds, nonlinear DeePC can exploit much larger datasets and reach performance comparable to model-based nonlinear MPC without an explicit model.","feed_headline":"Nonlinear data-driven control runs 100x faster","feed_subtitle":"A product-kernel trick shrinks the DeePC optimization so datasets 25x larger become practical.","key_machinery":"The central object is the product kernel $k_\\otimes((u_1,x_1),(u_2,x_2))=k_u(u_1,u_2)k_x(x_1,x_2)$ on the product space of input trajectories and initial states, which generates a product reproducing kernel Hilbert space and a Gram matrix $K_\\otimes=K_u\\otimes K_x$ whose inverse factors as $K_u^{-1}\\otimes K_x^{-1}$. The load-bearing identity is the reduction $\\Omega(x)=\\big(I_{T_u}\\otimes k_x^\\top(x)/\\|k_x(x)\\|_2^2\\big)K_\\otimes$, which replaces the $T_xT_u$-dimensional kernel equation by a $T_u$-dimensional one, and the companion projection $\\hat g=Y^\\dagger\\tilde g$ that moves decision variables into the $pN$-dimensional output space. This machinery is what decouples the input and state data, cuts the optimization size, and makes the Gram inversion and construction much cheaper.","core_discovery":"The paper claims that a nonlinear discrete-time system's behavior can be parameterized as $G(u)(x)=\\Theta^* k_\\otimes(u,x)$ in the product reproducing kernel Hilbert space built from the kernel $k_\\otimes((u_1,x_1),(u_2,x_2))=k_u(u_1,u_2)k_x(x_1,x_2)$, with Gram matrix $K_\\otimes=K_u\\otimes K_x$. It then shows that the DeePC prediction equations $K_\\otimes g=k_\\otimes(u,x)$, $Yg=y$ are equivalent to the reduced system $\\Omega(x)g=k_u(u)$, $Yg=y$ with $\\Omega(x)=\\big(I_{T_u}\\otimes k_x^\\top(x)/\\|k_x(x)\\|_2^2\\big)K_\\otimes$, and that projecting $g$ through $g=Y^\\dagger\\tilde g$ into the $pN$-dimensional output space yields a one-shot efficient formulation (Problem 4.4). The equivalence is established by Lemma 4.3 under the condition that the output data annihilate the null space of $\\Omega(x)$. The numerical section reports per-control-action computation times of $0.3319\\,\\mathrm{s}$ versus $33.8668\\,\\mathrm{s}$ for the stacked-kernel formulation at $T=400$, and shows that $T=10000$ datasets are tractable with the product formulation while the stacked formulation is not.","pith_inferences":["Because the state and input data enter through separate kernels, the same library of input sequences can be reused with many different initial-condition sets; one could update $K_x$ and the corresponding part of $K_\\otimes$ without recomputing the whole pipeline when adding new operating regions.","The projection $g=Y^\\dagger\\tilde g$ buys speed at the cost of limiting the decision space to the row space of the output data, so systems with many outputs or poor output richness may need feature selection in the output space to keep enough flexibility.","The k-means choice of initial conditions suggests a testable extension: distributing initial conditions along unstable or strongly nonlinear regions could improve the operator approximation for systems that are not globally stable, since the density theorem does not say where data must be sampled."],"forward_implications":["For the same dataset of 400 trajectories, the efficient product-kernel DeePC computes a control action in 0.3319 seconds whereas the stacked-kernel formulation needs 33.8668 seconds, a speedup of roughly two orders of magnitude.","The product formulation remains solvable online for 10,000 trajectories (about 1.35 seconds per action), while the stacked formulation cannot handle that dataset online, so users can train on 25 times more data.","With 10,000 trajectories, the product-kernel controller's mean tracking error drops from 0.0994 to 0.0835 and its mean prediction error from 0.0216 to 0.0157, and the paper shows tracking performance comparable to nonlinear MPC.","Under the conditions of Lemma 4.3, the reduced model predicts the same outputs as the original product-kernel model, so the computational savings do not change the predicted trajectories."],"supporting_citations":[{"why":"supplies the product RKHS operator parameterization $G(u)(x)=\\Theta^*k_\\otimes(u,x)$ and the universality result that makes the predictor dense for nonlinear operators.","marker":"[18]"},{"why":"supplies the fundamental lemma and persistency-of-excitation background that the kernelized equations (7) generalize.","marker":"[1]"},{"why":"provides the stacked kernelized DeePC formulation used as the baseline for computation-time and prediction-error comparisons.","marker":"[14]"},{"why":"gives kernelized fundamental lemmas for linear, Hammerstein, and flat nonlinear systems, which the product formulation recovers as the $T_x=1$ special case.","marker":"[16]"},{"why":"defines universal kernels, including the Gaussian and Hardy reverse multiquadratic kernels used in the numerical example.","marker":"[19]"},{"why":"supplies the equivalent predictor representation and model-equivalence notion used in Lemma 4.3 to justify the reduced formulation.","marker":"[4]"}],"fun_headline_variants":["Kernel operator method speeds nonlinear control 100x","Bigger datasets, faster control: product kernel DeePC","Product RKHS shrinks DeePC optimization 100x","Nonlinear predictive control scales to 10000 samples","Speedup 100x: kernelized DeePC for large data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the inherited premise that the product reproducing kernel Hilbert space is dense enough to represent the true system's input-to-output operator on the relevant domain; if the real system is not well approximated there, the predictor is biased no matter how the optimization is sped up.","fun_headline_variants_meta":{"raw":{"variants":["Kernel operator method speeds nonlinear control 100x","Bigger datasets, faster control: product kernel DeePC","Product RKHS shrinks DeePC optimization 100x","Nonlinear predictive control scales to 10000 samples","Speedup 100x: kernelized DeePC for large data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1515,"prompt_tokens":971,"completion_tokens":544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":463}},"tokens_in":587,"tokens_out":544,"duration_ms":5467,"temperature":1.0,"reasoning_tokens":463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:40:21.721614+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a nonlinear benchmark with a non-smooth or discontinuous map (for example a relay or dead-zone nonlinearity), generate data with Algorithm 2, and check whether the multi-step prediction error (18) keeps decreasing as $T_x$ and $T_u$ grow; if the error plateaus above the measurement noise, the universal-approximation premise is violated for that system.","supporting_citations":[{"cited_title":"A note on persistency of excitation,","cited_arxiv_id":null,"evidence_quote":"supplies the fundamental lemma and persistency-of-excitation background that the kernelized equations (7) generalize."},{"cited_title":"Robust and kernel ized data- enabled predictive control for nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"provides the stacked kernelized DeePC formulation used as the baseline for computation-time and prediction-error comparisons."},{"cited_title":"Exploring the links be tween the fundamental lemma and kernel regression,","cited_arxiv_id":null,"evidence_quote":"gives kernelized fundamental lemmas for linear, Hammerstein, and flat nonlinear systems, which the product formulation recovers as the $T_x=1$ special case."},{"cited_title":"Universal kernels,","cited_arxiv_id":null,"evidence_quote":"defines universal kernels, including the Gaussian and Hardy reverse multiquadratic kernels used in the numerical example."},{"cited_title":"Basis-functions nonlinear data-enabled pre dictive control: Consistent and computationally efficient formulations,","cited_arxiv_id":null,"evidence_quote":"supplies the equivalent predictor representation and model-equivalence notion used in Lemma 4.3 to justify the reduced formulation."}],"review_version":1}