{"id":"fde50970-742b-4dc7-9aa1-b619d4d2d4dc","arxiv_id":"2508.20288","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural Spline Operators train a physics-informed neural operator with B-spline output bases to predict long-term safety probabilities for stochastic systems with varying functional dynamics.","lead":"Neural Spline Operators (NeSO) learn a mapping from a stochastic system's dynamics function and safe-set parameters to the probability of staying safe over time, using B-spline basis functions inside a physics-informed neural operator. A generalist reader might care because online risk estimates that adapt to changing dynamics are a bottleneck in safety-critical control, and this method claims large speed-ups over simulation and PDE solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is not proven as stated: the proof approximates an infinite-dimensional coefficient map with a finite-dimensional MLP and quantifies uniformly over noncompact α∈R^ξ, both beyond the cited universal approximation results.","rationale":"The reader's CONDITIONAL verdict is appropriate, and my concern is closely related but more specific. The weakest assumption identified by the reader is the continuity of the map (f,α)→F; my concern is that even granting that continuity, the proof of Theorem 1 does not establish the claimed uniform approximation because it misapplies finite-dimensional MLP universal approximation to an infinite-dimensional coefficient map and quantifies over a noncompact parameter set. This directly affects Eq. (19), the central theoretical claim. I do not recommend REJECT because both defects appear repairable: restricting α to a compact range matches the experiments, and the coefficient approximation could likely be justified with an explicit grid-evaluation and density argument, though that argument is absent. The empirical results, especially the two case studies and the speed comparison, provide independent support for the method's practical value, and the modal decomposition in the multi-agent case is imported from a stated theorem. Therefore, the paper should remain CONDITIONAL: the theoretical guarantee needs to be rederived with correct domain restrictions and a valid approximation argument before the claim 'unconditional for any α∈R^ξ' is accepted.","tokens_in":15088,"tokens_out":17596,"duration_ms":188448,"concrete_test":"Rewrite the coefficient map explicitly: let E_M(g)∈R^M be the point values of the neural operator output function on a fixed grid, and let L_θ: R^M→R^N be the linear layer. Check whether sup_{g∈K3} ∥C*(g)-L_θ(E_M(g))∥₂ can be driven below ε as M grows and θ varies, with K3 the compact family of safety probabilities appearing in the case studies. If yes, state and prove the missing density argument; if no, there is no valid route in the proof from the output function g_h to the optimal coefficients. Analytically, perform this check on the one-dimensional family parameterized by α∈[1,2] using exact projection coefficients computed from ground-truth PDE solutions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix A, the proof of Theorem 1 (Eq. 19) has a category error that is load-bearing. After defining the optimal B-spline coefficients C*(g_h) = B_φ^{-1} b(g_h), where g_h = W(h) is a function in the Sobolev space Y=H^k(Ω2;R), the proof invokes the MLP universal approximation theorem [47] to assert that C*(·): Y→R^N can be approximated by a neural network. But [47] is a statement about continuous functions on finite-dimensional Euclidean spaces; C* is a bounded linear map on an infinite-dimensional Hilbert space, and no argument is given that the proposed architecture (a neural operator layer followed by a linear layer) can approximate it from the input h rather than from g_h. Separately, Theorem 1 quantifies uniformly over all α∈R^ξ, while the FNO universal approximation theorem [43] used in the same proof applies on compact input sets; K1×K2 must be compact for that step, and R^ξ is not compact. Both defects sit precisely on Eq. (19), the central theoretical guarantee. They may be repairable by restricting α to a compact range and by replacing the MLP step with a point-evaluation and density argument for the coefficient map, but as written the theorem's conclusion exceeds its assumptions and the proof does not establish the advertised universal approximation guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural Spline Operators (NeSO), a physics-informed neural operator framework for estimating long-term safety probabilities of stochastic systems with varying functional dynamics. The safety probability is characterized as the solution of a convection-diffusion PDE, and NeSO maps the dynamics function f and safe-set parameter α to B-spline control points through a coefficient neural functional; the control points are multiplied by B-spline bases to produce the safety probability field. A universal approximation theorem (Theorem 1, Eq. (19)) is claimed. Two case studies are presented: a one-dimensional recovery-probability problem with randomly parameterized sinusoidal dynamics, and a 14-dimensional multi-agent mass-spring-damper system whose safety probability is reconstructed as a product of per-mode probabilities. The reported results show that NeSO improves accuracy and training time over FNO in the first case and achieves substantial online speedups over Monte Carlo and PDE solvers in the second.","tokens_in":15364,"tokens_out":7559,"duration_ms":72438,"significance":"If the universal approximation theorem were established as stated, the paper would be a meaningful step: it would be the first physics-informed operator learning method for risk quantification, with a B-spline output representation that enforces initial and boundary conditions and enables analytical derivatives. The experimental comparison is useful, the code is provided, and the online speedup claim is concrete. However, the theorem as written is not proven: the proof contains a category error in the approximation of the coefficient map, an unjustified uniformity over noncompact α∈R^ξ, and a false statement about orthogonality of B-splines. These are load-bearing because they support Eq. (19), the paper's central theoretical guarantee. The experimental claims are not deeply affected by these proof gaps, but the advertised guarantee currently exceeds what the proof establishes.","major_comments":[{"comment":"The proof invokes the MLP universal approximation theorem [47] to assert that the coefficient map C*(·): Y→R^N can be approximated by a neural network, but [47] applies to continuous functions on finite-dimensional Euclidean spaces. Here Y=H^k(Ω2;R) is infinite-dimensional, and the proposed coefficient neural functional receives h=(f,α), not the function gh=W(h) on which C* is evaluated. No argument shows that the composition h↦C*(W(h)) is approximable by a neural operator layer followed by a linear layer. Since this step is used to bound the term ∥C*(gh)-Ĉ*(ĝh)∥ in Eq. (62), the proof does not establish the advertised universal approximation in Eq. (19).","section":"Appendix A, proof of Theorem 1, Eq. (61)"},{"comment":"The theorem quantifies uniformly over all α∈R^ξ and over f in a class that is not made precise, whereas the FNO universal approximation theorem [43] used in Eq. (41) requires the input set to be compact and the operator W to be continuous on that set. The manuscript only assumes that K1 and K2 are subsets of Sobolev spaces and states in footnote 2 that continuity 'holds with regularity conditions,' without stating those conditions. To make Eq. (19) follow from [43], the statement must restrict (f,α) to a compact product set K1×K2 and state the regularity conditions under which (f,α)↦F is continuous.","section":"Section V, Theorem 1 and Appendix A, Eq. (41)"},{"comment":"The footnote claims that B-spline basis functions are orthogonal to each other; this is false. B-spline bases are locally supported and linearly independent but not orthogonal in L2, so the asserted invertibility of B_φ 'since φ_j are orthogonal' is not justified. The argument can be repaired by using the positive definiteness of the Gram matrix of a B-spline basis, but as written the proof contains a false mathematical claim at a point used to define the optimal coefficients C*(gh).","section":"Appendix A, footnote 4 and Eqs. (42)-(47)"}],"minor_comments":[{"comment":"The word 'inital' in the introduction should be corrected to 'initial.'","section":"Section I"},{"comment":"The statement 'for any function f∈F' uses an undefined symbol F; it should be f∈K1 or another explicitly defined class, and the notation should be made consistent with the earlier assumption f∈K1⊂C(Ω1;R^n).","section":"Section V, Theorem 1"},{"comment":"The PDE residual uses d/dx and d^2/dx^2 notation for a vector state x; this should be written with ∇_x and trace(σσ^T ∇_x^2 F) so that the residual is well-defined for n>1 as well as for the scalar case.","section":"Section IV, Eq. (14)"},{"comment":"The dynamics function in Eq. (21) depends only on t, while the problem formulation in Eq. (1) allows x-dependence; the paper should state explicitly that this case study uses time-dependent dynamics, so that S(τ) in Eq. (25) is independent of the state trajectory.","section":"Section VI-A, Eqs. (21) and (25)"},{"comment":"The reported NeSO online time of 56.47 s should be separated into the network evaluation time and any additional reconstruction overhead, and the 3999 s training time should be clearly labeled as offline, since the advertised speedup depends on this offline/online split.","section":"Section VI-B, Table II"}],"recommendation":"major_revision","confidential_remarks":"The contribution is largely an architectural integration: the PDE characterization is taken from prior work [3], the B-spline universal approximation from [5], and the modal decomposition from [46]. The universal approximation theorem is advertised as a main contribution and is currently not proven as stated, so I recommend major revision rather than rejection because the proof gaps appear repairable by restricting the parameter domain and replacing the flawed coefficient-map approximation step with a valid operator approximation argument. The experiments appear honest and the code availability is a positive feature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest read: the proposed architecture is genuinely new and the empirical results are decent, but the universal approximation theorem as stated is not proven. The stress-test note is correct: the proof approximates the coefficient map C* on the infinite-dimensional space Y using a finite-dimensional MLP theorem, and quantifies uniformly over all α∈R^ξ even though the FNO approximation theorem applies on compact input sets. Those are load-bearing gaps, not cosmetic.\n\nWhat actually works: putting B-spline output layers on top of a neural operator backbone is a sensible way to get analytic derivatives and hard enforcement of initial/boundary conditions in physics-informed operator learning. The first case study shows a real gain over plain FNO on both error and training time, with standard deviations reported over 10 test functions. The second case study is a plausible demonstration on a 14-dimensional multi-agent system, and the claimed speedup relative to Monte Carlo and PDE solving is credible for an amortized surrogate. Code is available, which is good.\n\nThe soft spots beyond the theorem: footnote 4's claim that B-spline bases are orthogonal is false, though the argument only needs a positive-definite Gram matrix, so this is easily fixed. The product-form decomposition F=∏F_k comes from a prior theorem; the paper gives no evidence about how sensitive the 14-dimensional reconstruction is to inexact decoupling. The second case study has no repeated-run variability, so the reported 0.0074 error is a single draw. There are also no precise training details for the PINO losses beyond weights.\n\nWho gets value from this: people working on operator learning surrogates for safe control, and to a lesser extent PINO method developers. I'd send it to peer review — the idea deserves referee time and the proof issues are repairable. But the theorem should not be accepted in its present form: restrict α to a compact set, and give a legitimate density argument for the coefficient map rather than invoking a finite-dimensional MLP theorem on an infinite-dimensional space.","headline":"A genuinely new architecture with decent empirical results, but the universal approximation theorem as written is not proven.","tokens_in":15890,"tokens_out":3088,"would_cite":true,"duration_ms":28143,"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":"Neural Spline Operators learn direct maps from dynamics functions to safety probabilities, with a universal approximation guarantee.","keywords":["physics-informed neural operators","risk quantification","safety probability","B-splines","stochastic dynamical systems","universal approximation","operator learning","multi-agent safety"],"falsifier":"Run a numerical experiment along a family of smooth dynamics functions that converge uniformly to a limit with a sharp drift change; compute the safety probability for each member with a high-accuracy PDE or sampling solver and check whether the probabilities converge. If they do not converge despite uniform convergence of the dynamics, the continuity premise behind the theorem fails, and NeSO would lose its guarantee on that family.","tokens_in":14884,"feed_emoji":"🎯","tokens_out":10136,"duration_ms":86971,"temperature":0.7,"pith_summary":"The paper introduces Neural Spline Operators (NeSO), which it positions as the first physics-informed operator learning framework for risk quantification. The central claim is that one trained operator can take the system's dynamics function $f$ and safe-set parameters $\\alpha$ as input and output the long-term safety probability $F(x,t)$ for any state and time horizon. This matters because sampling-based and PDE-based estimators become expensive or fragile when dynamics change in functional form, while existing physics-informed networks only accept fixed finite-dimensional parameter vectors. The paper proves a universal approximation theorem for NeSO and demonstrates the method on a system with randomly varying functional dynamics and on a 14-dimensional seven-agent system, reporting millisecond-scale online evaluation.","feed_headline":"One trained operator maps any dynamics to safety probabilities","feed_subtitle":"B-spline neural operators cut long-term risk evaluation to milliseconds, replacing sampling and PDE solves.","key_machinery":"The load-bearing object is the B-spline basis paired with learned control points. The safety probability is represented as $\\hat F(x,t)=C\\cdot B_{\\ell,d}(x,t)$, where $C$ is a control-point tensor produced by a coefficient neural functional $G_\\theta(f,\\alpha)$ and $B_{\\ell,d}$ is a tensor product of univariate B-spline bases in each state coordinate and in time. Two structural facts do the work: B-splines have closed-form derivatives, so the convection-diffusion residual can be formed without automatic differentiation, and the boundary control points of the tensor directly encode the initial condition and Dirichlet boundary conditions, so those constraints are enforced exactly rather than learned. The proof combines universal approximation of the neural operator layer with the fact that orthogonal projection onto the B-spline subspace is a bounded linear map that a feedforward network can reproduce.","core_discovery":"On its own terms, the paper's discovery is that the risk-quantification problem can be reorganized as an operator learning problem: instead of re-solving the convection-diffusion equation for each dynamics function, a network learns the mapping $(f,\\alpha)\\mapsto F$. The network output is a tensor of B-spline control points; multiplying this tensor by the B-spline basis functions produces the predicted safety probability surface. B-spline derivatives are available in closed form, so the physics-based residual is cheap to evaluate, and the endpoint control points can be set by hand to satisfy initial and Dirichlet boundary conditions exactly. Theorem 1 asserts that for any accuracy $\\varepsilon>0$ there exist B-spline counts and orders and network parameters $\\theta$ such that $|F(x,t)-G_\\theta(f,\\alpha)(x,t)|\\le\\varepsilon$ uniformly over $f$, $\\alpha$, $x$, and $t$, assuming the map from dynamics and safe sets to safety probabilities is continuous.","pith_inferences":["A stress test would push NeSO toward dynamics families that approach a discontinuous limit; the error behavior there would reveal whether the continuity assumption, not the architecture, is the operative bottleneck.","If the B-spline control-point trick is as generic as the proof suggests, the same construction could be applied to other PDE-constrained operators, fixing boundary conditions in the representation rather than through penalty terms.","The product-form reconstruction in the multi-agent study inherits an exact modal decoupling assumption; for non-symmetric or time-varying interaction graphs that factorization breaks, and NeSO would need a different decomposition to keep working.","The reported speed comparisons include training time for NeSO; a fair deployment comparison would amortize training across the number of query systems, which the paper's own cost formula starts to do."],"forward_implications":["A single trained NeSO can evaluate safety probabilities for new, unseen dynamics functions in milliseconds, replacing repeated sampling or PDE solves with one forward pass.","Because it accepts a function as input, NeSO covers dynamics families whose functional form changes, rather than only parameterized variants of a fixed form.","Assigning initial and boundary conditions through B-spline control points removes the need for additional penalty terms that enforce those constraints during training.","The universal approximation theorem transfers to other PDE-constrained operator learning problems whose solution depends continuously on an input function.","In the multi-agent case, factorizing the full safety probability into sub-system probabilities lets NeSO reconstruct a 14-dimensional safety landscape from low-dimensional solves."],"supporting_citations":[{"why":"Supplies the convection-diffusion equation and initial and boundary conditions whose solution is the long-term safety probability.","marker":"[3]"},{"why":"Shows that B-spline bases can serve as universal approximators with direct assignment of initial and boundary conditions.","marker":"[5]"},{"why":"Defines neural operators as maps between function spaces, the formalism NeSO extends to risk quantification.","marker":"[6]"},{"why":"Provides the spectral-convolution neural operator layer used as the coefficient backbone in the case studies.","marker":"[24]"},{"why":"Gives universal approximation and error bounds for spectral-convolution neural operators, used in the first step of the proof.","marker":"[43]"},{"why":"Cited as the regularity result under which the map from dynamics and safe sets to the safety probability is continuous.","marker":"[45]"},{"why":"Proves the modal decomposition that lets the full multi-agent safety probability factor as a product of sub-system probabilities.","marker":"[46]"},{"why":"Supplies the universal approximation theorem for feedforward networks used to approximate the projection map onto the B-spline subspace.","marker":"[47]"}],"fun_headline_variants":["B-spline operator turns risk evaluation into milliseconds","One neural operator maps any dynamics to safety risk","Universal operator for risk: learn once, apply to all dynamics","From PDEs to milliseconds: neural spline operator for risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that small changes in the dynamics function and safe-set parameters produce small changes in the safety probability; if that map has a jump, the universal approximation guarantee no longer applies.","fun_headline_variants_meta":{"raw":{"variants":["B-spline operator turns risk evaluation into milliseconds","One neural operator maps any dynamics to safety risk","Universal operator for risk: learn once, apply to all dynamics","From PDEs to milliseconds: neural spline operator for risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1552,"prompt_tokens":938,"completion_tokens":614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":554,"tokens_out":614,"duration_ms":5873,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:47:44.572765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a numerical experiment along a family of smooth dynamics functions that converge uniformly to a limit with a sharp drift change; compute the safety probability for each member with a high-accuracy PDE or sampling solver and check whether the probabilities converge. If they do not converge despite uniform convergence of the dynamics, the continuity premise behind the theorem fails, and NeSO would lose its guarantee on that family.","supporting_citations":[{"cited_title":"Safe control in the presence of stochastic uncertainties,","cited_arxiv_id":null,"evidence_quote":"Supplies the convection-diffusion equation and initial and boundary conditions whose solution is the long-term safety probability."},{"cited_title":"Neural operator: Learning maps be- tween function spaces with applications to pdes,","cited_arxiv_id":null,"evidence_quote":"Defines neural operators as maps between function spaces, the formalism NeSO extends to risk quantification."},{"cited_title":"On universal approximation and error bounds for fourier neural operators,","cited_arxiv_id":null,"evidence_quote":"Gives universal approximation and error bounds for spectral-convolution neural operators, used in the first step of the proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the regularity result under which the map from dynamics and safe sets to the safety probability is continuous."},{"cited_title":"Orthogonal modal representa- tion in long-term risk quantification for dynamic multi-agent systems,","cited_arxiv_id":null,"evidence_quote":"Proves the modal decomposition that lets the full multi-agent safety probability factor as a product of sub-system probabilities."},{"cited_title":"Multilayer feedforward networks are universal approximators,","cited_arxiv_id":null,"evidence_quote":"Supplies the universal approximation theorem for feedforward networks used to approximate the projection map onto the B-spline subspace."}],"review_version":2}