{"id":"e0019204-7385-4909-9dfe-efcd7db2bfca","arxiv_id":"2412.12469","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A neural adaptive spectral operator maps optimal control instances directly to control functions in one forward pass, with approximation error bounds and large inference speedups over classical solvers.","lead":"The authors propose NASM, a neural operator that maps optimal control problem instances directly to control functions in a single forward pass, avoiding iterative optimization at test time. It offers a reusable, fast alternative to classical optimal-control solvers, particularly when dynamics are unknown or expensive to model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The approximation guarantee is conditioned on unverified Sobolev regularity of the control operator and its encoded coefficient map; constrained/nonsmooth OCPs, which the paper itself says are only approximately handled, can violate that premise.","rationale":"The paper's strongest claim is an approximation guarantee with explicit network-size bounds. I looked for the least secure condition for that guarantee. Theorem 3's proof is omitted, but standard Fourier theory makes it credible once the H^s premise is granted; the real risk is that the premise itself is neither implied by the OCP formulation nor verified experimentally. The paper's own Appendix I narrows the framework by stating that boundary conditions and inequalities hold only approximately, and in constrained optimal control nonsmooth or bang-bang solutions are common. Theorem 4 adds a second, less visible premise: the map from encoded instance to Fourier coefficients must itself be Sobolev-smooth in the parameter space; time-regularity of controls does not give this. The reported benchmarks are smooth, unconstrained or softly constrained instances, so the empirical MAPE may hold while the theorem's assumptions remain untested. This does not undermine the empirical speedup or the merit of the architecture; those are plausible and consistent with the results. But the arbitrary-tolerance-with-bounded-size claim is exactly as strong as these unverified regularity hypotheses, making them the load-bearing point. A released-code audit of Fourier tail decay and coefficient-map finite differences would settle whether the theory covers the benchmarks. If it does not, the paper should be revised to state the guarantee as conditional on checked regularity and to qualify the constrained-OCP claim; this is an addressable issue rather than a fatal one, so the existing CONDITIONAL verdict stands.","tokens_in":26835,"tokens_out":10742,"duration_ms":106663,"concrete_test":"Audit both Sobolev premises on the released Quadrotor benchmark: (a) for the 100 DM-generated optimal controls, fit log||u - P_p u||_{L2} versus log p to estimate s in Theorem 3; (b) finite-difference the Fourier coefficient vector c(x_goal) with respect to x_goal over the ID distribution to estimate the H^s(E#mu) regularity of P∘G∘E^{-1} needed by Theorem 4. If either fitted exponent is not positive, that theorem's premise is false for the reported benchmark, and the approximation bound cannot explain the empirical MAPE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim (Theorem 1) rests on Theorem 2's decomposition and on two regularity premises: (i) Theorem 3 requires G(i) in H^s(T), s>0, with finite integrated squared H^s norm; (ii) Theorem 4 additionally requires the encoded Fourier-coefficient map P∘G∘E^{-1} to lie in H^s(E#mu). Neither premise is derived from the OCP structure nor checked on the benchmark controls. For active control/state constraints, optimal controls are often bang-bang or discontinuous, and the Fourier reconstruction error then decays too slowly for the bound to be useful; Appendix I concedes constraints are only 'satisfied approximately', so the learned operator is not guaranteed feasible. Condition (ii) is even more hidden: smoothness of u(t) in time does not imply Sobolev regularity of the coefficients as a function of the encoded instance e, and the theorem is stated for a coefficient vector A(e), not the actual time-dependent c_j(t,i) in Eq. (4). Theorem 3's proof is also omitted. Thus the claim that NASM approximates the control operator to arbitrary tolerance with bounded size and depth is not established for the general OCP setting advertised.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an instance-solution operator perspective for optimal control problems (OCPs), in which a neural operator maps an encoded OCP instance directly to the optimal control function, thereby avoiding iterative optimization at test time. The authors instantiate this operator with a new architecture, the Neural Adaptive Spectral Method (NASM), which combines a coefficient network, time-dependent coefficients, and adaptive basis functions. The theoretical part decomposes the approximation error into encoder, approximator, and reconstructor errors (Theorem 2) and states bounds on the reconstructor and approximator errors under Sobolev regularity assumptions on the control operator and its encoded coefficient map (Theorems 3 and 4). The experimental part compares NASM with direct and indirect solvers and with several neural operators on five synthetic systems and a real planar-pushing dataset, reporting large speedups relative to the direct method and competitive or better MAPE on in- and out-of-distribution benchmarks, including two problems with analytical solutions.","tokens_in":27073,"tokens_out":4010,"duration_ms":42376,"significance":"If the central claim were fully established, the paper would make a useful practical contribution: a reusable neural operator that solves families of OCPs in one forward pass, without knowledge of the dynamics, is attractive for real-time and model-free settings. The error-decomposition proof in Theorem 2 is clean and correctly reduces the overall bound to reconstructor and approximator errors. The use of two benchmarks with analytical optimal solutions (Brachistochrone and Zermelo) is a genuine check against circularity, since the neural models are still trained on direct-method data but evaluated against ground truth. The public code link and systematic baselines are also strengths. However, the advertised approximation guarantee is currently not established at the advertised level of generality: the key reconstructor bound is stated with its proof omitted, and the regularity assumptions on the optimal-control operator are asserted rather than verified or derived from OCP structure. These are load-bearing gaps.","major_comments":[{"comment":"The Fourier reconstructor error bound bER ≤ C p^{-s} is the backbone of the claimed arbitrary-accuracy result, but the proof is explicitly omitted: the text says 'The proof (omitted here) is based on an observation that a smooth function... has small (exponentially decaying) coefficients.' For a central theorem this is not sufficient. Please provide a complete proof or a precise citation with the constant C spelled out as a function of s, M, and the time interval T, including the case of vector-valued controls u: T→R^{d_u}.","section":"Appendix C, Theorem 3"},{"comment":"There is a mismatch between the architecture analyzed and the approximator error bound. The NASM definition in Eq. (4) uses time-dependent coefficients c_j(t,i), but Theorem 4 bounds the approximation of a map e ↦ A(e) with A(e) ∈ R^p, with error measured in ℓ2(R^p). The proof of Theorem 2 then uses Lip(R) bEA with bEA defined on R^p only. Please formalize how the coefficient map produced by the network is identified with the finite-dimensional vector A(e), and state the corresponding norm; otherwise the size and depth bounds in Eq. (10) do not apply to the architecture actually used in the experiments.","section":"Section 2 and Appendix C, Theorem 4"},{"comment":"The regularity assumptions G: I → H^s(T) and P ∘ G ∘ E^{-1} ∈ H^s(E#μ) are load-bearing, but they are neither derived from the OCP structure nor verified on the benchmark problems. The claim that these assumptions are 'trivial in many real-world continuous OCPs' is unsupported. For OCPs with active control or state constraints, optimal controls are frequently bang-bang or discontinuous, so G(i) need not belong to H^s for s>0, as the paper itself concedes in Appendix I where constraints are only 'satisfied approximately.' Please either restrict the formal theorems to a class of unconstrained or sufficiently regular OCPs, or provide sufficient conditions on f, d, X, and U under which the optimal control operator satisfies the required Sobolev regularity, and check those conditions on the benchmark instances.","section":"Appendix C, 'Estimation of Decomposed Errors'"},{"comment":"Theorem 1 is presented only informally, and no formal counterpart appears in the appendix. The formal content is distributed among Theorems 2-5, but the appendix does not assemble these into a single statement with explicit constants, network size, depth, and the exact regularity assumptions under which 'for any ε there exists a NASM' holds. Please state and prove a formal version of Theorem 1 that makes the dependency on the regularity parameters (s, M, m, p) and on the encoder error explicit.","section":"Section 2, Theorem 1"}],"minor_comments":[{"comment":"In the size calculation for the block-diagonal p-dimensional MLP, each layer i≥2 has p blocks of size size(W^{i,j}), so the total size is p·size(N_j), not p^2·size(N_j); the displayed p^2 bound is a valid but loose upper bound. Please correct the accounting or state explicitly that the p^2 factor is an overestimate.","section":"Appendix D, proof of Lemma 7"},{"comment":"The adaptive Fourier basis restricts |θ_k| ≤ 0.5 'to avoid overlapping the adaptive range of basis functions,' but no precise condition is given that guarantees the parameterized basis remains a basis or that the reconstruction Lipschitz constant stays bounded. Please state the condition formally.","section":"Eq. (5)"},{"comment":"The theorem assumes fi > 0 and the finiteness of C = sup_i Lip(f_i)/(f_i∘G(i)), but these assumptions are only introduced in text before Eq. (2). Please include them in the theorem statement itself, since the bound is not meaningful when the optimal cost can vanish.","section":"Theorem 2 statement"},{"comment":"For the Pushing dataset, MAPE is defined as ||u - u*||/||u*||, whereas for the synthetic environments it is defined through costs. This difference is noted in the text but not discussed; a sentence explaining why the two metrics support the same conclusions would improve clarity.","section":"Section 3.2 and Table 7"},{"comment":"The description of the FNO variant used is brief: 'we slightly modify it by adding time indices to its input.' Since FNO is a central baseline and the modification affects the validity of the comparison, please describe the modification in enough detail to be reproducible.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the empirical contribution is solid, but the theoretical guarantee advertised in the abstract and Section 2 is not yet proven as stated. The omitted proof of Theorem 3 and the unverified Sobolev regularity of the control operator are the main risks; both are fixable within the manuscript's scope. I would also suggest the editor ensure that the final version contains either a verification of the regularity assumptions on the benchmarks or a more modest formal claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, honest paper that deserves a serious referee, but the theoretical headline is narrower than the abstract suggests. The architecture—adaptive basis functions, time-dependent coefficients, neural aggregation—is genuinely new relative to SNO and DeepONet, and the application to learning an instance-to-control operator is a useful step. Empirical work is extensive: multiple synthetic benchmarks, a real pushing dataset, ablations, OOD generalization, fine-tuning, and two problems with analytical solutions to check the ground truth.\n\nThe best part is Theorem 2's error decomposition, which is proved cleanly and honestly adapted from Lanthaler et al. The comparison showing NASM gets DeepONet's reconstructor bound with fewer parameters is fair and correct, because the basis is fixed rather than learned.\n\nSoft spots, in order of seriousness. First, Theorem 3's proof is omitted, and it is load-bearing: the O(p^{-s}) Fourier reconstructor error requires G(i) in H^s(T). The paper asserts this is 'trivial' for many real OCPs, but does not verify it for any benchmark, and for constrained problems—which Appendix I admits are only approximately handled—optimal controls are often bang-bang with limited regularity. Second, Theorem 4's assumption on the coefficient map P∘G∘E^{-1} is even more hidden: smoothness of u(t) in time does not imply Sobolev regularity of the coefficients as a function of the encoded instance. So the 'arbitrary tolerance with bounded size and depth' claim is not established for the general setting advertised. Third, the synthetic experiments lack error bars, though the pushing dataset does report them. Training and evaluation both use the direct-method solver as ground truth; the Brachistochrone and Zermelo experiments mitigate this by checking against analytical solutions, which I credit.\n\nNone of this is fatal. The architecture and empirical evidence stand on their own, and the theoretical section is an honest attempt at a bound with clearly stated (if unverified) hypotheses. The abstract overreaches, but the paper body is mostly careful.\n\nFor you: if you work on learned control or neural operators, it is worth a read. I would accept it for peer review. With a referee requesting the missing proof, empirical verification of regularity, and error bars for synthetic runs, it could be a solid publication.","headline":"NASM is a real architectural step for learning control operators, but the theoretical guarantee rests on unverified Sobolev regularity and an omitted proof, so the paper is solid-conditioned rather than solid-outright.","tokens_in":27605,"tokens_out":1712,"would_cite":true,"duration_ms":16050,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49J15","46E35","41A25","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims optimal control can be solved by learning the operator that maps instances to solutions, and that its NASM architecture approximates this operator with bounded error and order-of-magnitude speedups.","keywords":["optimal control","neural operator","spectral method","operator learning","adaptive basis","approximation error","out-of-distribution generalization","amortized solving"],"falsifier":"Estimate the Sobolev regularity s of the optimal controls generated by the direct method for the paper's benchmarks; if some benchmark controls have s ≤ 0, the claimed $p^{{-s}}$ reconstruction bound does not apply to them, and a stronger test would be to run a family of OCPs with provably discontinuous optimal controls and check whether NASM's error decays with the number of basis functions at all.","tokens_in":26619,"feed_emoji":"🎛️","tokens_out":9956,"duration_ms":76973,"temperature":0.7,"pith_summary":"This paper tries to establish that optimal control problems can be solved by learning one operator that maps a problem instance — the cost functional, dynamics, and initial condition — directly to the optimal control function, without iterative optimization or an explicit dynamics model. The authors argue this instance-solution operator perspective makes optimal-control solvers reusable and fast, and they instantiate it with NASM, a neural operator that generalizes classical spectral methods by making both coefficients and basis functions adaptive. They prove that, under Sobolev regularity conditions, NASM can approximate the operator to any desired error tolerance with bounded network size and depth. Empirically, NASM matches or beats other neural operators in accuracy on five synthetic systems and a real robot-pushing dataset while running over 6000 times faster than a direct collocation solver, and it generalizes out of distribution, especially after few-shot fine-tuning.","feed_headline":"A learned operator solves optimal control in one pass","feed_subtitle":"NASM maps problem instances to controls with error bounds, 6000x speedup, and out-of-distribution generalization.","key_machinery":"The central object is the Neural Adaptive Spectral Method (NASM), a generalization of classical spectral approximations in which the solution is built from basis functions whose coefficients and whose adaptation parameters are both produced by a network: NNASM(i)(t) = aggregation of c_j(t, i) b_j(t; θ(t,i)). The adaptive parameters θ scale and shift the basis (e.g., sin(π[(1+θ1)t+θ2])) while being bounded to avoid overlap, which lets the same Fourier family represent non-periodic controls. The argument is carried by an error decomposition stating that the relative cost error of the full operator is bounded by a Lipschitz constant times the reconstructor error plus the approximator error, and each of those terms is then bounded separately using Fourier approximation theory and Sobolev-space neural approximation bounds. This decomposition is what turns the abstract operator-learning claim into a concrete guarantee with explicit rates and parameter counts.","core_discovery":"On the paper's own terms, the central claim is that a well-posed optimal control problem can be treated as an infinite-dimensional operator G that takes an instance i = (f, d, xinit) and returns the optimal control function u*. This operator perspective is then realized by NASM, whose output at time t is an aggregation of coefficients c_j(t, i) times adaptive basis functions b_j(t; θ(t,i)), where the network infers both coefficients and basis parameters. The main theoretical result decomposes the total approximation error into encoder, approximator, and reconstructor errors and bounds each: the Fourier reconstructor error decays as $p^{{-s}}$ for controls in H^s(T), and the MLP/CNN approximator error can be driven below any positive tolerance with bounded size and depth. The consequence is that a single trained NASM can output near-optimal controls for unseen problem instances in one forward pass. Experiments confirm the claim in practice: on the Quadrotor benchmark, NASM reaches an in-distribution MAPE of 6.17e-6 and an out-of-distribution MAPE of 1.21e-4 while taking about 6.5e-5 seconds per instance versus 9.2e-2 seconds for the direct method, and fine-tuning on 20% OOD data restores accuracy under extreme distribution shifts.","pith_inferences":["Inference: if the control-operator perspective holds up, it suggests a general amortization strategy — expensive optimal-control computation is paid once during training and then replayed at essentially zero marginal cost, which could change real-time control practice; the paper does not test closed-loop deployment.","Inference: the H^s regularity premise is the natural place to probe the theory; for bang-bang or otherwise discontinuous optimal controls, one would expect the p^{-s} decay to fail, suggesting a testable extension is to add basis families that represent discontinuities (e.g., wavelets) or to validate the regularity empirically from solver data.","Inference: the error decomposition implies a principled architecture-selection rule — pick fixed bases for periodic smooth problems and adaptive bases for non-periodic ones — which could be automated per instance rather than chosen by hand.","Inference: because the operator is learned from labels produced by a reference solver, an iterative self-improvement loop (train on the model's own solutions, then with a better solver) could push accuracy beyond the reference, matching the paper's observation that some neural operators beat the direct method on out-of-distribution Brachistochrone benchmarks."],"forward_implications":["A trained NASM can replace iterative optimal-control solving at inference time, yielding the reported over 6000x speedup while maintaining solution quality on in-distribution and many out-of-distribution instances.","Because the dynamics are never explicitly modeled, the same operator framework applies to systems whose equations are unknown or too noisy to write down, as demonstrated on the planar-pushing dataset.","The error bounds mean the approximation improves predictably with the number of basis functions when optimal controls are sufficiently smooth, and the architecture needs no more parameters than DeepONet to reach the same reconstructor accuracy.","If a few out-of-distribution samples are available, fine-tuning the operator restores accuracy even for large distribution shifts, making the learned solver reusable beyond its training distribution.","The framework is solver-agnostic: any numerical solver that produces (instance, control) pairs can provide training data, so the operator can inherit and potentially improve on the reference solver's solutions."],"supporting_citations":[{"why":"Supplies the optimal-control formulation (cost, dynamics, initial condition) that the instance-solution operator is defined over.","marker":"(Kirk 2004)"},{"why":"Provides the error-decomposition technique (encoder/approximator/reconstructor) and the DeepONet error bound that NASM's analysis extends and compares against.","marker":"(Lanthaler, Mishra, and Karniadakis 2022)"},{"why":"Defines the DeepONet operator-learning baseline and the branch/trunk structure that NASM generalizes.","marker":"(Lu et al. 2021)"},{"why":"Introduces the Spectral Neural Operator, the fixed-basis predecessor that NASM makes adaptive.","marker":"(Fanaskov and Oseledets 2022)"},{"why":"Supplies the Pontryagin Differentiable Programming baseline and the synthetic control-system evaluation protocol.","marker":"(Jin et al. 2020)"},{"why":"Provides the ReLU-MLP Sobolev approximation bounds used in the approximator-error theorem.","marker":"(Gühring, Kutyniok, and Petersen 2020)"},{"why":"Bridges MLP and CNN approximation so the error bound extends to convolutional backbones.","marker":"(Petersen and Voigtlaender 2020)"},{"why":"Supplies the real-world planar-pushing dataset used to test the operator on unknown, noisy dynamics.","marker":"(Yu et al. 2016)"},{"why":"Provides IPOPT, the nonlinear programming backend of the direct method that generates training labels and ground-truth costs.","marker":"(Biegler and Zavala 2009)"}],"fun_headline_variants":["Neural operator solves optimal control in one shot","One-pass optimal control via learned spectral operator","NASM: One-shot control with proven error bounds","Neural operator for optimal control with 6000x speedup","Adaptive spectral method learns control operator directly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof relies on optimal controls being reasonably smooth as functions of time and on cost functionals being Lipschitz with bounded relative growth; if a problem family has discontinuous bang-bang controls or active state and input constraints, these conditions can fail and the stated error bounds do not cover it.","fun_headline_variants_meta":{"raw":{"variants":["Neural operator solves optimal control in one shot","One-pass optimal control via learned spectral operator","NASM: One-shot control with proven error bounds","Neural operator for optimal control with 6000x speedup","Adaptive spectral method learns control operator directly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1442,"prompt_tokens":951,"completion_tokens":491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":417}},"tokens_in":567,"tokens_out":491,"duration_ms":4506,"temperature":1.0,"reasoning_tokens":417,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:02:32.915762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the Sobolev regularity s of the optimal controls generated by the direct method for the paper's benchmarks; if some benchmark controls have s ≤ 0, the claimed $p^{{-s}}$ reconstruction bound does not apply to them, and a stronger test would be to run a family of OCPs with provably discontinuous optimal controls and check whether NASM's error decays with the number of basis functions at all.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the optimal-control formulation (cost, dynamics, initial condition) that the instance-solution operator is defined over."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the error-decomposition technique (encoder/approximator/reconstructor) and the DeepONet error bound that NASM's analysis extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Pontryagin Differentiable Programming baseline and the synthetic control-system evaluation protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Bridges MLP and CNN approximation so the error bound extends to convolutional backbones."},{"cited_title":"T.; and Zavala, V","cited_arxiv_id":null,"evidence_quote":"Provides IPOPT, the nonlinear programming backend of the direct method that generates training labels and ground-truth costs."}],"review_version":1}