{"id":"a2b70c0e-020d-4d72-a909-57359a0b3c97","arxiv_id":"2411.10622","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of Kolmogorov-Arnold Networks that restates existing theory and claims, without new experiments or derived results.","lead":"This paper is a survey of Kolmogorov-Arnold Networks, explaining the Kolmogorov-Arnold representation theorem, B-splines, and KAN architecture. It presents KANs as escaping the curse of dimensionality, but adds no new analysis and leaves its own open question unanswered.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dimension-free error claim is not established: Theorem 1's constant C and required grid size G can grow exponentially with input dimension even for functions with exact smooth KAN representations.","rationale":"The reader's weakest assumption identifies the representation/smoothness requirement and the Section 3 polynomial-structure assumption; that is a valid concern. My stress-test sharpens it: even when a smooth KAN representation exists, Theorem 1 as cited does not imply dimension-independent complexity, because the constant C and the grid size G can depend on n through the range and derivatives of the univariate components. The exponential example f_n(x)=exp(Σx_i) is exactly representable, yet the grid size needed for fixed accuracy grows exponentially with n. This directly undercuts the abstract's 'true freedom from the curse of dimensionality' and the Section 6 generalization to real-world tasks. The paper itself flags unresolved limitations in Section 6 (unrelated inputs) and Section 7 (polynomial structure cannot be proven, only not-disproven), but the conclusion does not carry these caveats. Since these are unverified and underspecified assumptions behind the main claim, and since the review provides no experiments or independent derivation, the REJECT verdict stands. The reader and I agree on the need to reject; I differ only in emphasizing that the hidden n-dependence in C and G is the more fundamental failure, not merely the smooth-representation hypothesis.","tokens_in":10850,"tokens_out":5533,"duration_ms":62262,"concrete_test":"Analytically instantiate Theorem 1 for the exactly representable family f_n(x)=exp(x_1+...+x_n), n=2,4,...,64. Using m=0 and the proof's constant C_n = ||(d^{k+1}/dt^{k+1})e^t||_{C^0([0,n])} = e^n, compute the minimal grid size G(n) satisfying C_n G^{-(k+1)} ≤ 1e-3 for k=3. If log G(n) grows linearly in n, the dimension-free error bound is vacuous for this family. As a complementary check, fit a KAN with grid sizes G_fixed=10 and G(n) to this target and report the C^0 error scaling; the fixed-grid error should grow with n despite the exact representability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that KANs achieve 'true freedom from the curse of dimensionality' (Abstract, Section 6) rests on Theorem 1, which states ||f - (Φ_{L-1}∘...∘Φ_0)x||_{C^m} ≤ C G^{-k-1+m}. The paper reads this as 'error not reliant on input dimensions,' but the theorem only bounds error in terms of G for a fixed f that admits a representation with (k+1)-times continuously differentiable components. The constant C depends on f and its representation, and nothing in the theorem controls how C or G scale with n. A concrete family shows the problem: f_n(x)=exp(x_1+...+x_n) is exactly representable as a KAN with Φ(t)=e^t on t∈[0,n]. For B-splines of degree k on a grid of size G, the approximation error is roughly C (n/G)^{k+1} with C ~ ||(e^t)^{(k+1)}||_{C^0([0,n])} = e^n. Holding error ≤ ε requires G ≳ n e^{n/(k+1)}, i.e., exponential in n. Thus a smooth, exactly representable target already exhibits an exponential parameter/data requirement, so the dimension-free error bound does not imply freedom from the curse. The paper also relies on the unverified Section 3 assumption that the process 'has a higher-order polynomial expression,' and Section 6 explicitly poses the unrelated-input-output question without resolving it, yet the conclusion repeats the strong claim. These are load-bearing: without controlling C and G uniformly in n, and without verifying the representation condition, the 'true freedom' claim collapses to a conditional statement about a narrow function class.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a broad survey of Kolmogorov-Arnold Networks (KANs). It reviews interpolation, B-splines, perceptrons, the universal approximation theorem, the Kolmogorov-Arnold representation theorem, and the KAN architecture, and it argues from Theorem 1 of Liu et al. that KANs achieve 'true freedom from the curse of dimensionality.' It also claims that KANs require far fewer parameters and exhibit better scaling than MLPs. The manuscript contains no new derivations or experiments; its conclusions are drawn from cited work and from the authors' interpretation of the error bound.","tokens_in":11144,"tokens_out":5191,"duration_ms":47009,"significance":"If the advertised claim were valid, it would be a major result for high-dimensional learning, and the paper provides a reasonably accessible introduction to the mathematical ingredients of KANs. The pedagogical sections on splines and interpolation are clear, and the paper explicitly identifies the polynomial-structure assumption in Section 3, which is a useful caution. The paper also cites a broad set of recent KAN application papers. However, the central quantitative claim is not established: the quoted error bound does not control the dependence of the constant and grid size on the input dimension, and the paper leaves the unrelated-input question in Section 6 unanswered before repeating the strong conclusion in Section 7. The paper therefore overstates what the cited theorem supports.","major_comments":[{"comment":"The claim that Theorem 1 shows 'true freedom from the curse of dimensionality' is not supported by the stated bound. The bound ||f − (Φ_{L−1}∘...∘Φ_0)x||_{C^m} ≤ C G^{−k−1+m} only says that for a fixed f whose KAN representation has (k+1)-times continuously differentiable components, the error decreases as the spline grid G grows. It does not state that the constant C or the required G are independent of the input dimension n. A concrete example shows why this matters: f_n(x)=exp(x_1+...+x_n) is exactly representable by a KAN with Φ(t)=e^t on [0,n], but the B-spline error of approximating e^t on an interval of length n scales like e^n/G^{k+1} up to polynomial factors in n, so holding the error below ε forces G to grow exponentially in n. Thus the theorem, as quoted, is compatible with an exponential parameter/data requirement, and the Abstract's unconditional claim is not a consequence of it.","section":"Abstract; Section 6, Theorem 1"},{"comment":"Section 6 raises the exact objection that would limit the curse-of-dimensionality claim: 'what happens when a model is trained with unrelated input and output features?' and asks whether the bound ε ≤ CG^{−k−1+m} still holds. No answer is given, and Section 7 then reasserts that KANs are promising for high-dimensional real-world tasks. This is load-bearing because Theorem 1 is conditional on f admitting a smooth KAN representation; an arbitrary mapping from unrelated features to labels need not satisfy that condition, so the error bound cannot be invoked for such tasks. The authors must either answer the question with a rigorous statement of the class of functions for which the bound holds, or explicitly restrict the conclusion to that class.","section":"Sections 6 and 7"},{"comment":"The paper's generalization from the theorem to practice relies on the assumption, introduced in Section 3, that 'the process we are trying to model has a higher-order polynomial expression.' Section 7 acknowledges that this structure 'can be challenging to prove' and proposes that it is 'reasonable to consider' whenever a polynomial form cannot be ruled out. That is not a verification: for image, video, or arbitrary regression tasks, there is no evidence that a KAN representation with smooth B-spline components exists. Without such evidence, the error bound does not apply and the claimed scalability advantages are not established. The conditional nature of the claim should be stated in the Abstract and Conclusion.","section":"Sections 3 and 7"}],"minor_comments":[{"comment":"Equation (21) states f : [0,1] → R, but the theorem applies to multivariate f : [0,1]^n → R; this is a typo that should be corrected.","section":"Equation (21)"},{"comment":"There are several typos, including 'the the rectified linear unit' in Section 3.3 and inconsistent capitalization 'KANS' in Section 6.","section":"Section 3.3"},{"comment":"The caption of Figure 5 says the B-spline interpolant manages to interpolate 'without needing any data,' which is misleading; the interpolant is constructed from the same 50 interpolation points used by the other methods.","section":"Figure 5 caption"},{"comment":"The reference list relies heavily on Wikipedia entries for mathematical background (e.g., [16], [23]–[26], [29], [30]); for a formal review, primary sources would be more appropriate.","section":"References"},{"comment":"The description of the Tschirnhaus transformation as reducing to 'n − 4 variables when n ≥ 7' is imprecise and is not used later in the paper; it should be rephrased or removed.","section":"Section 2"}],"recommendation":"reject","confidential_remarks":"The manuscript is a survey with no new technical content. Its central advertised claim is a misreading of the cited theorem, and the unresolved question in Section 6 undercuts the conclusion. If the journal is open to expository papers, a substantially rewritten version that states the conditional nature of KAN error bounds and honestly discusses dimension-dependent constants could be considered; in its current form I cannot recommend publication. I also note that the citation list is dominated by arXiv preprints and Wikipedia, which is a concern for a formal journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou can skip the abstract and go straight to Section 6 if you want the paper's one real moment. It asks: what happens when input and output features are unrelated? That is the right question, and the paper does not answer it. The rest of the paper is a survey that restates KAT, B-splines, perceptrons, UAT, and Liu et al.'s KAN architecture. As a primer it is serviceable—the interpolation section is clear enough, the figures are fine, and the reference list covers the early KAN literature. If you want a quick map of the terminology, you could do worse.\n\nThe soft spot is load-bearing. The abstract and Section 6 claim that KANs achieve \"true freedom from the curse of dimensionality,\" based on Theorem 1. But Theorem 1 only gives ||f - (Φ...)x|| ≤ C G^{-k-1+m} for a fixed f that admits a representation with smooth B-spline components. The constant C depends on f and its representation, and nothing in the theorem controls how C or G scale with input dimension. The stress-test example is telling: f_n = exp(x_1+...+x_n) has an exact KAN representation, but holding error fixed requires G to grow like n e^{n/(k+1)}—exponential in n. So the theorem does not imply dimension-free approximation. The paper also relies on an unverified assumption in Section 3 that the target process \"has a higher-order polynomial expression.\" Section 6 raises the unrelated-input objection and leaves it unresolved; Section 7 then asserts the strong claim anyway. That is not a minor hedge—it is the gap between a conditional statement about a narrow function class and the broad claim in the abstract.\n\nThere is no new theorem, method, experiment, or dataset. The citation pattern is mostly fine, though leaning on a Wikipedia entry for the theorem in a review is weak. The paper's honest core is the question in Section 6; the rest is exposition with an overclaim at the top.\n\nWho is this for? A reader who wants a shallow orientation to KAN vocabulary might get something. But as a scientific survey it is not reliable, and as a research contribution it is empty. I would not send it to a serious peer review venue; desk-reject it, or accept it only as a tutorial with the dimension-free claim explicitly corrected and the polynomial-structure assumption flagged. If the authors revise it that way, it could be a useful educational note.","headline":"A readable KAN primer whose central dimension-free claim does not follow from the cited theorem.","tokens_in":11680,"tokens_out":3396,"would_cite":false,"duration_ms":33164,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","41A15","41A30"],"pacs":[],"model":"deepseek-v4-flash","headline":"KANs claim to break the curse of dimensionality","keywords":["Kolmogorov-Arnold Networks","Kolmogorov-Arnold representation theorem","curse of dimensionality","B-splines","universal approximation theorem","error scaling","neural network architecture"],"falsifier":"Construct a continuous function on [0,1]^n that is known not to admit a smooth KAT representation—for example, a function whose KAT components are fractal or discontinuous, such as the maximum of n variables—and train a KAN on it for increasing n. If the empirical error grows with n in a way that violates C $G^{{-k-1+m}}$ (e.g., requires exponentially more grid points), the claimed dimension-free bound fails.","tokens_in":10634,"feed_emoji":"🧠","tokens_out":3304,"duration_ms":32261,"temperature":0.7,"pith_summary":"This review argues that Kolmogorov-Arnold Networks (KANs), a neural architecture introduced in 2024, can escape the curse of dimensionality that limits traditional multilayer perceptrons. The paper explains the underlying mathematics—interpolation, B-splines, the universal approximation theorem, and the Kolmogorov-Arnold representation theorem—and presents the error bound that supports the claim. If correct, KANs would train accurate models on high-dimensional data with far fewer parameters and less data than MLPs. The review also stresses that the guarantee depends on the target function possessing a polynomial-like structure with smooth univariate components, a condition it does not verify for real-world tasks.","feed_headline":"KANs claim to break the curse of dimensionality","feed_subtitle":"Review explains the error bound that would let networks learn high-dimensional functions without exponential data.","key_machinery":"The central object is the Kolmogorov-Arnold representation theorem (KAT), which states that any continuous multivariate function on [0,1]^n can be written as a sum of outer univariate functions composed with sums of inner univariate functions: f(x) = Σ_{q=1}^{2n+1} Φ_q(Σ_{p=1}^n φ_{q,p}(x_p)). KANs turn this into a neural architecture by replacing fixed activations with learnable univariate functions approximated by B-splines. The theorem from Liu et al. is the load-bearing error bound: with B-spline order k and grid size G, the C^m error decays as C $G^{{-k-1+m}}$, independent of dimension, which is the source of the claimed curse-of-dimensionality freedom.","core_discovery":"The paper's central claim is that KANs achieve true freedom from the curse of dimensionality. This rests on a theorem from Liu et al. stating that for any function admitting a KAN representation with (k+1)-times continuously differentiable B-spline components, the approximation error satisfies ||f - (Φ_{L-1}∘...∘Φ_0)x||_{C^m} ≤ C $G^{{-k-1+m}}$, where G is the grid size and C is a constant independent of the input dimension n. Because the error bound does not involve n, the review concludes that KANs can handle high-dimensional inputs without the exponential growth in data requirements that plagues standard neural networks. It further claims that KANs need fewer parameters and achieve better scaling exponents than MLPs, positioning them as a practical alternative for high-dimensional learning.","pith_inferences":["The curse-of-dimensionality freedom is conditional: for a given real-world dataset, one must first verify that the true mapping lies in the KAN function class. Without that verification, the error bound offers no practical guarantee, and empirical performance could degrade with dimension.","A testable extension would be to compare KANs and MLPs on high-dimensional benchmark functions where the KAN representation condition is known to fail (e.g., functions with discontinuous or fractal KAT components) to see whether the claimed advantage disappears.","The review's emphasis on polynomial structure suggests KANs may be especially useful in scientific computing and symbolic regression, where such structure is often assumed; the paper does not explore this connection explicitly."],"forward_implications":["KANs could train on high-dimensional inputs without the exponential data growth that MLPs require, because the approximation error does not scale with input dimension.","KANs may need far fewer parameters than MLPs to reach comparable accuracy, since each univariate function is directly parameterized by B-spline coefficients.","The error bound suggests that increasing the B-spline grid size G is a controlled way to improve accuracy, unlike the unbounded neuron count N(ϵ) in the universal approximation theorem.","KANs are best suited to problems where the target function is believed to have a smooth, polynomial-like structure; where that structure is absent, the guarantees do not apply."],"supporting_citations":[{"why":"Supplies the KAN architecture, the error bound in Theorem 1, and the scaling-law results that the review's central claim rests on.","marker":"[1]"},{"why":"Provides the original Kolmogorov representation theorem, the mathematical foundation for decomposing multivariate functions into univariate superpositions.","marker":"[18]"},{"why":"Arnold's extension of the theorem to general dimensions, which the review uses to justify the decomposition in higher-dimensional spaces.","marker":"[19]"},{"why":"Cybenko's universal approximation theorem serves as the baseline: it guarantees approximation but gives no bound on neuron count, highlighting the contrast with KANs.","marker":"[31]"},{"why":"Barron's approximation bounds for sigmoidal superpositions provide the MLP error-scaling comparison that KANs are claimed to improve upon.","marker":"[33]"}],"fun_headline_variants":["KANs break the curse of dimensionality","KANs escape the curse of dimensionality","KANs dodge the curse of dimensionality","KANs triumph over the curse of dimensionality","KANs make the curse of dimensionality vanish"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire curse-of-dimensionality argument depends on the target function admitting a KAN representation with smooth B-spline components, and the review also assumes the underlying process has a higher-order polynomial expression; if the data contain discontinuities or lack such structure, the error bound and the freedom-from-curse conclusion collapse.","fun_headline_variants_meta":{"raw":{"variants":["KANs break the curse of dimensionality","KANs escape the curse of dimensionality","KANs dodge the curse of dimensionality","KANs triumph over the curse of dimensionality","KANs make the curse of dimensionality vanish"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2370,"prompt_tokens":964,"completion_tokens":1406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1337}},"tokens_in":580,"tokens_out":1406,"duration_ms":13140,"temperature":1.0,"reasoning_tokens":1337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:29:17.833166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a continuous function on [0,1]^n that is known not to admit a smooth KAT representation—for example, a function whose KAT components are fractal or discontinuous, such as the maximum of n variables—and train a KAN on it for increasing n. If the empirical error grows with n in a way that violates C $G^{{-k-1+m}}$ (e.g., requires exponentially more grid points), the claimed dimension-free bound fails.","supporting_citations":[{"cited_title":"Kolmogorov","cited_arxiv_id":null,"evidence_quote":"Provides the original Kolmogorov representation theorem, the mathematical foundation for decomposing multivariate functions into univariate superpositions."},{"cited_title":"Kolmogorov–arnold representation theorem, 2024","cited_arxiv_id":null,"evidence_quote":"Arnold's extension of the theorem to general dimensions, which the review uses to justify the decomposition in higher-dimensional spaces."},{"cited_title":"Deep Learning","cited_arxiv_id":null,"evidence_quote":"Cybenko's universal approximation theorem serves as the baseline: it guarantees approximation but gives no bound on neuron count, highlighting the contrast with KANs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Barron's approximation bounds for sigmoidal superpositions provide the MLP error-scaling comparison that KANs are claimed to improve upon."}],"review_version":1}