{"id":"0f9b2b4a-3f6a-4bbb-9b7b-74f5bb79fa60","arxiv_id":"2506.07958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The first NTK analysis of cPIKANs finds their kernel spectra stay stable during training, correlating with large accuracy gains over PINNs, especially when time is split into subdomains.","lead":"The authors study how Chebyshev-based physics-informed neural networks learn to solve partial differential equations by tracking a mathematical quantity called the neural tangent kernel. They report that these networks show more stable and focused learning dynamics than standard PINNs, and that splitting time into smaller domains improves convergence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central causal claim that NTK spectral behavior explains cPIKAN convergence rests on an unverified NTK-regime assumption, and the paper's own spectral-evolution plots indicate the kernel is far from constant during training.","rationale":"The reader's weakest assumption identified exactly the NTK-regime validity for finite-width, multi-layer cPIKANs trained with Adam/L-BFGS/hybrid optimizers. My stress-test agrees and sharpens it: the paper's displayed spectra evolve during training, which is direct evidence against kernel constancy, and the only theorem that would justify treating the initialization kernel as predictive has unverified assumptions and a proof gap in the bound. The paper does provide useful descriptive evidence: the derived expected NTK formula in Theorem 1 under Gaussian initialization is a coherent analytical contribution, and the observed spectral stabilization correlating with lower error is a real empirical pattern. But the abstract's language that spectral trends 'directly link kernel behavior to convergence rates' and 'predict' empirical performance goes beyond what the evidence supports. The condition for acceptance is clear: either verify that NTK drift is small during the relevant training phase and that Eq. (46) predicts the actual residual evolution, or rewrite the claims as correlational. Since the reader already assigned CONDITIONAL with these conditions, my recommendation is UNCHANGED rather than a new verdict.","tokens_in":27345,"tokens_out":2804,"duration_ms":35559,"concrete_test":"Re-run Experiment 4.1 and one decomposed case from Experiments 4.3 or 4.4 with parameter checkpoints, and compute at each training step: (i) relative NTK drift ||K(tau)-K(0)||_F / ||K(0)||_F, (ii) parameter drift ||theta(tau)-theta(0)||, and (iii) the NTK-predicted residual trajectory obtained by integrating dv/dtau = -K(tau)(v-G) against the actual residual trajectory. If the relative NTK drift exceeds roughly 10% during the phase where accuracy is set, or the NTK-predicted trajectory diverges from the actual residuals by more than the model's own error, then the spectral metrics are correlational rather than causal and the central claim must be tempered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the empirical NTK computed along the actual training trajectory remains a valid linearized descriptor of cPIKAN training dynamics, so that eigenvalue spectra and spectral entropy can be causally linked to convergence. Theorem 2 is the only theoretical support: it bounds NTK drift by C * ||theta(tau) - theta(0)|| under gradient flow, uniformly bounded gradients/Hessians, and vanishing parameter drift. None of these conditions is checked in Experiments 4.1-4.4, and the trained models are finite-width networks (widths 8-47, depths 2-4) optimized with Adam, L-BFGS, or hybrid updates rather than gradient flow. More directly, the paper's own figures show NTK eigenvalue spectra changing substantially over training (Figs. 2, 4, 6, 8-9, 11), which is evidence that the kernel is not near-constant. If the kernel evolves significantly, spectral statistics describe a moving object, not a fixed kernel whose eigenstructure predicts convergence. The paper's Section 5 concedes that these metrics 'do not fully explain how the NTK influences model accuracy or generalization.' There is also a technical issue in Appendix B: the proof replaces the integral of ||theta_dot(s)|| ds with ||theta(tau) - theta(0)||, which is generally false since path length can exceed displacement; as written, even the stated drift bound is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript analyzes the Neural Tangent Kernel (NTK) of Chebyshev-based Kolmogorov-Arnold Networks in both supervised (cKAN) and physics-informed (cPIKAN) settings. It derives an expected NTK formula for a one-hidden-layer cKAN, states a drift bound for finite-width cKANs, and gives a generic block-form NTK for physics-informed losses. The main body presents four numerical PDE experiments (diffusion, Helmholtz, Allen-Cahn, forced Euler-Bernoulli vibration) comparing cPIKAN with PINN and bPIKAN under Adam, L-BFGS, and hybrid optimization, with temporal domain decomposition in the last two experiments. The paper claims that NTK eigenvalue spectra and spectral entropy reveal and directly explain cPIKAN convergence advantages over PINNs.","tokens_in":1706,"tokens_out":1864,"duration_ms":81064,"significance":"If the central claim were established, this would be a useful contribution: it is the first systematic NTK study of cPIKANs, it covers four benchmark PDEs, and it proposes spectral diagnostics for when domain decomposition helps. The paper's strengths include the explicit parameter-free expectation in Theorem 1, the clear empirical comparisons against PINN and bPIKAN baselines, and the reproducible tables of architecture, error, and runtime. However, the load-bearing causal claim that spectral behavior 'directly links' to convergence is not supported by the current theory or experiments; the cPIKAN-specific NTK is not actually derived beyond a generic block form, and the only drift theorem has a flawed proof. The empirical results remain valuable if reframed as descriptive observations rather than a proven mechanistic explanation.","major_comments":[{"comment":"The proof of Theorem 2 replaces the integral of ||theta_dot(s)|| ds from 0 to tau with ||theta(tau) - theta(0)||. This is false in general because path length is at least the displacement, so Eq. (B.4) does not follow from Eq. (B.3). As written, the advertised bound (36) is not established. The theorem must either state the correct bound with the integrated gradient norm and prove that this integral is small under the stated assumptions, or add assumptions that make the replacement valid. The additional claim that ||theta(tau)-theta(0)|| vanishes as N -> infinity is also not proved and is independent of the previous boundedness assumptions.","section":"Appendix B, Eq. (B.4)"},{"comment":"The physics-informed NTK for cPIKAN is not actually derived. Lemma 1 reproduces the generic block form of Wang et al. [60] and states that the proof follows the same reasoning, but no computation of the residual derivatives d(N[u])/dtheta or of the blocks K_rr and K_ru in terms of the Chebyshev structure is provided. Since the novelty claim is a cPIKAN-specific NTK analysis, this missing derivation is load-bearing. Theorem 1 is limited to a supervised one-hidden-layer cKAN and does not cover the multi-layer cPIKANs used in the experiments.","section":"Section 3.2, Lemma 1 and Eq. (45)"},{"comment":"The central claim that spectral trends 'directly link kernel behavior to convergence rates' is not supported by the evidence. Section 5 concedes that the spectral metrics 'do not fully explain how the NTK influences model accuracy or generalization.' The experiments show correlations only: for example, Section 4.2 explicitly states that 'a well-conditioned NTK spectrum alone does not guarantee high solution accuracy.' To support the causal or predictive claim, the paper needs either a convergence theorem involving the NTK spectrum or a quantitative test, such as measuring the alignment between the NTK top eigenspace and the residual/target during training, and showing that this alignment predicts the observed error drops.","section":"Abstract and Section 5"},{"comment":"The NTK-regime assumption is unverified for the actual training setups. The experiments use finite-width networks (widths 8-47, depths 2-4) trained with Adam, L-BFGS, or hybrid updates, while Theorem 2 assumes gradient flow and uniformly bounded gradients and Hessians. None of these conditions is checked. Moreover, the paper's own figures (Figs. 2, 4, 6, 8-9, 11) show eigenvalue spectra changing substantially during training, which indicates that the empirical NTK is not near-constant. The paper should either report kernel-drift magnitudes and verify the assumptions, or explicitly reframe the spectral analysis as a descriptive study of a time-varying kernel rather than a fixed kernel governing convergence.","section":"Experiments 4.1-4.4 and Theorem 2"},{"comment":"The spectral entropy metric is introduced without any theoretical connection to convergence or generalization. It is claimed to 'capture the transient and convergence behavior of the NTK spectrum,' but no argument or experiment establishes that reduced spectral entropy causes or predicts faster convergence. Since this metric is used to support the main claim in the forced-vibration experiment, it needs at least a formal definition of what it measures (e.g., effective rank versus variance concentration) and a demonstration that it tracks the convergence behavior beyond visual inspection.","section":"Section 4.4, Eq. (55)"}],"minor_comments":[{"comment":"The definition of D(x,x') in Eq. (A.13) contains an extra factor psi(x') compared with Eq. (32), making the two definitions inconsistent; this appears to be a typo that should be corrected.","section":"Appendix A, Eq. (A.13)"},{"comment":"The text calls the quantity in Eq. (55) 'spectral entropy' but the figures and captions label it 'spectral bias'; the terminology should be made consistent and the relationship between the two concepts clarified.","section":"Section 4.4, Eq. (55) and Figures 11-12"},{"comment":"The sentence 'B-splines are a common choice []' has an empty citation; a reference should be supplied.","section":"Section 2.1, B-splines"},{"comment":"The sentence 'LBFGS and hybrid approaches demonstrate a more rapid and structured shift in the eigenvalues' is vague; it would be helpful to quantify the shift, for example by reporting the effective rank or the largest eigenvalue ratio at selected epochs.","section":"Section 4.2, Figure 6 discussion"},{"comment":"The error reduction from 7.90e-3 (2 subdomains) to 6.21e-3 (4 subdomains) is small relative to the reduction from the full domain, so the text should avoid implying that the 4-subdomain improvement is of the same magnitude as the 2-subdomain improvement.","section":"Section 4.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The novelty relative to the authors' prior work [54] should be clarified; [54] already contains an NTK-based analysis of cKAN scaling, and the present manuscript's main new contribution is the empirical spectral study of cPIKANs across PDEs, optimizers, and domain decomposition. That is publishable if the causal language is softened and the theoretical gaps are addressed. The flawed proof in Appendix B is fixable by restating the theorem with the integrated gradient norm, but the paper must then re-derive the advertised conclusions or explicitly limit them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is the first systematic NTK analysis of cPIKANs, and it earns its place as a reference point for future work on KAN training dynamics. Theorem 1 is genuinely new: an expected-kernel formula for a nested one-hidden-layer cKAN under Gaussian initialization, derived without fitting or circular normalization. The experimental sweep is also respectable, four PDEs, parameter-matched and time-matched PINN baselines, and a sensible temporal-decomposition study across Allen-Cahn and Euler-Bernoulli problems. Section 5 is honest, admitting the spectral metrics do not fully explain accuracy or generalization, and that kind of candor matters.\n\nThe soft spots are real, though. The central claim in the abstract and conclusions, that spectral trends 'directly link' to convergence, is not supported. The paper's own figures show the NTK spectra changing substantially over training, which is direct evidence that the kernel is not near-constant; without a justified NTK regime for finite-width networks trained with Adam/L-BFGS/hybrid optimizers, the spectral statistics describe a moving object rather than a fixed kernel with predictive power. Theorem 2 is supposed to supply the missing link, but the proof in Appendix B is wrong as written: it replaces the integral of ||theta_dot(s)|| ds with ||theta(tau) - theta(0)||, and that inequality is generally false because path length can exceed displacement. So even the stated drift bound is not established. The assumptions (uniform gradient/Hessian bounds, gradient flow, vanishing drift) are also never checked in the experiments.\n\nThe empirical correlations are still informative as qualitative heuristics, especially the spectral-entropy reduction that accompanies successful domain decomposition. But without code or data, the numbers cannot be independently reproduced, and the predictive value of the spectra is asserted rather than demonstrated.\n\nWho should read this? Researchers comparing KAN-based and MLP-based PINNs, or thinking about when temporal decomposition helps, will get useful intuition from the experiments and the kernel formula. The Theorem 1 derivation is a solid contribution; the causal interpretation needs major tempering.\n\nRecommendation: send it to peer review, but require major revision. The reviewers should ask for code/data, a corrected drift argument or a removal of Theorem 2 as stated, and a rewrite that frames the spectral trends as correlated diagnostics rather than mechanisms.","headline":"A useful first NTK study of cPIKANs with a new kernel formula and broad experiments, but the causal claim that spectra explain convergence is overreached and the drift-bound proof has a real flaw.","tokens_in":28216,"tokens_out":1518,"would_cite":false,"duration_ms":21595,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Stable kernel spectra track cPIKAN convergence gains.","keywords":["neural tangent kernel","physics-informed neural networks","Kolmogorov-Arnold networks","Chebyshev polynomials","spectral bias","domain decomposition","PDE solvers","convergence analysis"],"falsifier":"Train a cPIKAN on the Allen–Cahn equation with 8 temporal subdomains while measuring $\\|\\theta(\\tau)-\\theta(0)\\|$ and the NTK spectral drift; if the kernel moves substantially during training while accuracy still improves, or if a deliberately diffuse initial spectrum still converges to low error, the spectral-stability explanation is falsified. A sharper test is to compute the alignment between the top NTK eigenvectors and the training residual $(K_{ntk}\\psi - G)$ and check whether that alignment, not just eigenvalue concentration, tracks convergence.","tokens_in":27040,"feed_emoji":"🧠","tokens_out":6124,"duration_ms":73095,"temperature":0.7,"pith_summary":"The paper analyzes Chebyshev-based physics-informed Kolmogorov–Arnold networks (cPIKANs) through neural tangent kernel (NTK) theory, aiming to show that their training dynamics are more tractable than those of standard physics-informed neural networks (PINNs). It argues that the cPIKAN kernel stays more structured and better conditioned during training, and that this spectral structure helps explain why cPIKANs converge faster and more accurately on PDE benchmarks. The empirical centerpiece is a forced vibration problem where cPIKAN with 8 temporal subdomains achieves about 98% lower maximum error than a PINN with the same decomposition. A sympathetic reader would take the paper to be establishing kernel spectral statistics as a practical diagnostic for when domain decomposition and optimizer choice improve training.","feed_headline":"Stable kernel spectra track cPIKAN convergence gains","feed_subtitle":"Chebyshev-based neural PDE solvers keep their learning kernel structured, and that structure tracks their error drops.","key_machinery":"The carrying object is the empirical neural tangent kernel matrix of the physics-informed network, together with two derived diagnostics. The kernel's blocks are inner products of parameter gradients: $K_{uu}$ for network outputs, $K_{rr}$ for PDE residuals, and $K_{ur}$ for the cross-terms, and under gradient flow the stacked outputs follow $d\\psi/d\\tau = -K_{ntk}(\\psi-G)$. The diagnostics are the eigenvalue spectrum $\\lambda(K_{ij})$ and the spectral entropy $S = -\\sum_i p_i \\log p_i$ with $p_i = |\\lambda_i|/\\sum_j |\\lambda_j|$, which measures whether learning directions are concentrated in a few dominant modes. This machinery translates architecture and training choices, such as Chebyshev degree, input scaling, optimizer, and subdomain count, into a kernel-conditioning picture in which stable, concentrated spectra indicate effective learning directions.","core_discovery":"The central claim is that the NTK of a cPIKAN is analytically tractable and empirically informative. At initialization the expected kernel has a closed two-term form, $E[K_{ntk}(x,x')] = N[\\sum_n C_n(x,x') + \\sum_i \\sum_n T_n(\\tilde x_i)T_n(\\tilde x'_i)D(x,x')]$, where $C_n$ captures Chebyshev-feature correlations after tanh activations and $D$ captures gradients through the activation function. Theorem 2 bounds kernel drift during gradient flow by $2B_1B_2\\|\\theta(\\tau)-\\theta(0)\\|$, so the kernel stays nearly constant when parameter drift is small. In the physics-informed setting the kernel is the block matrix with output, residual, and cross blocks, and the prediction error evolves as $d\\psi/d\\tau = -K_{ntk}(\\psi-G)$. Across the diffusion, Helmholtz, Allen–Cahn, and forced-vibration experiments, the paper reports that cPIKAN spectra remain broader and more persistent in residual blocks, that spectral entropy drops as temporal subdomains increase, and that these spectral trends accompany large error reductions, most notably about 98% lower maximum error than PINN in the 8-subdomain forced vibration case.","pith_inferences":["If spectral entropy at initialization predicts final error, one could choose subdomain counts and optimizers from a single cheap NTK computation before training, without solving the PDE.","The paper's own future-outlook section concedes that spectral metrics do not fully explain accuracy or generalization, so the spectral-stability story is likely one mechanism among several rather than a complete theory.","The same spectral-entropy diagnostic could be applied to other KAN basis choices, such as wavelets or radial basis functions, to predict which variant benefits most from domain decomposition.","Because Theorem 2's drift bound rests on gradient and Hessian bounds that the experiments do not verify, the observed spectral stability could partly reflect well-scaled targets rather than a true NTK regime; measuring parameter drift during training would separate these explanations."],"forward_implications":["Temporal domain decomposition is not just a parallelization trick: for cPIKANs it lowers spectral entropy of $K_{uu}$ and $K_{u_t u_t}$ and accompanies error drops from $5.09\\times10^{-1}$ to $6.21\\times10^{-3}$ on the Allen–Cahn equation.","Optimizer choice changes NTK evolution: L-BFGS and hybrid ADAM+LBFGS shift eigenvalues faster and reduce the Helmholtz relative $L^2$ error from $6.61\\times10^{-2}$ with ADAM alone to about $5\\times10^{-3}$.","cPIKAN retains higher-magnitude eigenvalues in the residual block $K_{rr}$ through training, which the paper interprets as better information flow from PDE residuals than PINNs achieve.","The spectral metrics suggest a practical rule: when the NTK spectrum is diffuse or slow to stabilize, adding temporal subdomains or switching to a second-order optimizer should improve convergence.","The NTK-based analysis extends to high-order PDEs, where additional derivative terms enlarge the kernel into a structured block matrix that still yields interpretable spectral-entropy trends."],"supporting_citations":[{"why":"Introduces the neural tangent kernel and the infinite-width constant-kernel dynamics that the paper adapts to cPIKANs.","marker":"[57]"},{"why":"Derives the physics-informed NTK block structure and spectral-bias analysis for PINNs that Lemma 1 extends.","marker":"[60]"},{"why":"Establishes the scaled cPIKAN formulation and prior NTK-based justification for spatial variable scaling.","marker":"[54]"},{"why":"Defines the Chebyshev polynomial basis and cKAN architecture whose kernel is derived in Section 3.","marker":"[55]"},{"why":"Documents when finite-width NTK theory fails, which motivates the drift bound and the paper's caution about regime assumptions.","marker":"[58]"}],"fun_headline_variants":["cPIKAN NTK stays tractable, predicts error drops","Kernel spectra reveal why cPIKANs converge faster","First NTK study of cPIKANs: stable spectra, better PDE solvers","cPIKANs: NTK analysis links kernel structure to 98% error cut","PIKANs vs PINNs: NTK shows why cPIKANs win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that finite-width cPIKANs trained with Adam, L-BFGS, or hybrids remain close enough to the NTK regime for kernel spectra and spectral entropy to describe the learning dynamics that produce the observed accuracy gains; the paper itself notes these metrics do not by themselves explain accuracy or generalization.","fun_headline_variants_meta":{"raw":{"variants":["cPIKAN NTK stays tractable, predicts error drops","Kernel spectra reveal why cPIKANs converge faster","First NTK study of cPIKANs: stable spectra, better PDE solvers","cPIKANs: NTK analysis links kernel structure to 98% error cut","PIKANs vs PINNs: NTK shows why cPIKANs win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1653,"prompt_tokens":1116,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":732,"tokens_out":537,"duration_ms":5560,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:21:56.342278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a cPIKAN on the Allen–Cahn equation with 8 temporal subdomains while measuring $\\|\\theta(\\tau)-\\theta(0)\\|$ and the NTK spectral drift; if the kernel moves substantially during training while accuracy still improves, or if a deliberately diffuse initial spectrum still converges to low error, the spectral-stability explanation is falsified. A sharper test is to compute the alignment between the top NTK eigenvectors and the training residual $(K_{ntk}\\psi - G)$ and check whether that alignment, not just eigenvalue concentration, tracks convergence.","supporting_citations":[{"cited_title":"Jacot, F","cited_arxiv_id":null,"evidence_quote":"Introduces the neural tangent kernel and the infinite-width constant-kernel dynamics that the paper adapts to cPIKANs."},{"cited_title":"Mostajeran, S","cited_arxiv_id":null,"evidence_quote":"Establishes the scaled cPIKAN formulation and prior NTK-based justification for spatial variable scaling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents when finite-width NTK theory fails, which motivates the drift bound and the paper's caution about regime assumptions."}],"review_version":1}