RecKAN: Kolmogorov-Arnold Networks with a Learnable Recursive Polynomial Basis
cs.LG · cs.AI
Quick verdict
RecKAN learns a 5-param recurrence generating the KAN basis itself; novel and interpretable, but benchmark margins are small and statistically untested.
Summary
RecKAN replaces the fixed basis of Kolmogorov-Arnold Networks with a basis generated by a learnable second-order polynomial recurrence, R(n+1)(x) = (ax²+bx+c)R_n(x) + (dx+e)R(n−1)(x), whose five coefficients are trained jointly with the network and shared across all layers. The paper shows the recurrence recovers Chebyshev (both kinds), Fibonacci, Pell, Jacobsthal, and Fermat polynomials as special cases, and proves degree bounds: at most one degree per recursion step when a=0, at most two when a≠0. Under a parameter-matched protocol, RecKAN posts the best mean accuracy among three KAN baselines (ChebyKAN, JacobiKAN, SplineKAN) on MNIST (97.39%), CIFAR-10 (54.39%), AG News (88.40%), and ECG5000 (93.89%), and the lowest MSE on ETTh1 (0.011172); RecKAN classifier heads also beat MLP heads on three CNN experiments by 0.08–0.52 points with 40 extra parameters. On two synthetic fits it outperforms larger MLPs, and learned coefficients move away from the Chebyshev-like initialization on high-frequency tasks. Limitations: several margins fall within the reported standard deviations; no significance tests are run and the number of runs is unreported; synthetic comparisons use training MSE at fixed checkpoints against an apparently under-tuned MLP; ETTh1 protocol details are absent; and no code or hyperparameters are provided, limiting reproducibility.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A five-coefficient second-order polynomial recurrence can serve as a learnable, jointly trained basis generator for KANs, subsuming Chebyshev (both kinds), Fibonacci, Pell, Jacobsthal, and Fermat polynomial families (Table 1).
- Polynomial degree grows by at most n−1 when a=0 and 2(n−1) when a≠0 (Proposition 1), so the learned basis can exceed any fixed classical family's degree at matched basis order.
- Under parameter matching, RecKAN achieves the best mean accuracy on all four classification benchmarks and the lowest ETTh1 MSE among ChebyKAN, JacobiKAN, and SplineKAN (Table 5).
- Replacing MLP classifier heads with RecKAN heads improves CNN accuracy on Fashion-MNIST, CIFAR-10, and SVHN with only 40 extra parameters (Table 7).
- Learned coefficients adapt to task structure: high-frequency/local tasks drive a away from zero, activating double degree growth (Tables 4 and 8).
Strengths
Genuinely new point in the KAN design space
Learning the basis-generating recurrence (5 shared coefficients, Section 3.2) rather than a fixed family is a real departure from prior variants; Table 1 maps exact coefficient settings to six classical families (Chebyshev 1st/2nd kind, Fibonacci, Pell, Jacobsthal, Fermat), making the subsumption claim concretely checkable.
Correct, honest theory linked to practice
Proposition 1's degree bounds (n−1 vs 2(n−1)) and Proposition 2's Chebyshev identification are simple but sound (Appendix B); Section 4.3 connects a>0 activation to high-frequency structure and explicitly labels the interpretation as descriptive rather than causal.
Disciplined parameter matching
Breadth of evaluation with coefficient introspection
Weaknesses
Margins within noise, no inferential statistics
ETTh1 claim statistically fragile
Weak synthetic comparisons
Reproducibility gaps
No MLP baseline in the main benchmark
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RecKAN (arXiv:2609.01729) targets a specific gap in the KAN design space: every existing variant (B-spline [5], Chebyshev [7], wavelet [8], fractional-Jacobi [9]) fixes the basis family before training and learns only combination weights. The paper instead generates the basis from a second-order recurrence R(n+1) = (ax²+bx+c)R_n + (dx+e)R(n−1) whose five coefficients are learned and shared network-wide (Section 3.2, Definition 1). Central claims: (1) the recurrence subsumes Chebyshev (both kinds), Fibonacci, Pell, Jacobsthal, and Fermat polynomials (Table 1); (2) degree grows by at most 1 per step when a=0 and 2 when a≠0 (Proposition 1), so the learned basis can exceed any classical family at matched order K; (3) parameter-matched benchmarks favor RecKAN on all four classification datasets and ETTh1; (4) RecKAN heads beat MLP heads on three CNN experiments; (5) learned coefficients move off the Chebyshev-like initialization on high-frequency tasks (Section 4.3, Tables 4 and 8).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is RecKAN and how does it differ from other KAN variants?
RecKAN is a Kolmogorov-Arnold Network whose edge basis functions are not drawn from a fixed family (splines, Chebyshev, wavelets, Jacobi). Instead, the basis is generated by a second-order polynomial recurrence R(n+1)(x)=(ax²+bx+c)R_n(x)+(dx+e)R(n−1)(x) whose five coefficients are learned jointly with the network and shared across all layers, adding only five parameters.
Does RecKAN actually beat Chebyshev KAN and other baselines?
In the paper's parameter-matched benchmarks it posts the best mean accuracy on MNIST, CIFAR-10, AG News, and ECG5000, and the lowest MSE on ETTh1 forecasting. However, several margins are small relative to the reported standard deviations (e.g., ECG5000 +0.083 points), no significance tests are reported, and the number of runs is unstated, so the gains should be treated as provisional.
Is code available for RecKAN?
The paper text contains no code or repository link. The method is fully specified by Algorithm 1 and public benchmarks are used, so reimplementation is possible, but exact training configurations (run counts, learning rates, epochs, ETTh1 protocol) are not disclosed.
Which classical polynomials can the learned recurrence reproduce?
With the paper's initial conditions R0=0, R1=1 and suitable coefficients: Chebyshev polynomials of the second kind, Fibonacci, Pell, Jacobsthal, and Fermat polynomials; Chebyshev polynomials of the first kind appear with a different initialization. All of these lie in the a=0 slice with one degree of growth per recursion step.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3–7 are raster images; only their captions were extractable, so visual fit quality (e.g., Figure 3) is assessed from captions and the reported numbers, not from the plotted curves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.