The 2024 KAN paper's own refs [9]-[16], cited for prior neural Kolmogorov-Arnold work, span 1993-2023 — only one falls in the 1980s-90s window

kolmogorovarnoldkanneural-networkshistory-of-sciencecitation-analysistooling-bottleneck

Liu et al.'s 2024 KAN paper backs its claim that "the possibility of using Kolmogorov-Arnold representation theorem to build neural networks has been studied" with eight citations, [9]-[16]. Read against the paper's own bibliography, these are: Sprecher & Draghici 2002, Köppen (ICANN) 2002, Lin & Unbehauen 1993, Lai & Shen (arXiv) 2021, Leni, Fougerolle & Truchetet 2013, Fakhoury, Fakhoury & Speleers 2022, Montanelli & Yang 2020, and He (arXiv) 2023. Only Lin & Unbehauen (1993) falls within the 1980s-90s window; five of the eight date from 2013 or later (four strictly postdate it — Leni et al. is 2013 itself). The paper singles out Lai & Shen [12] by name as the one prior work given real approximation-theoretic treatment. The paper's companion claim — that prior attempts stalled because "many did not have the chance to leverage more modern techniques (e.g., back propagation)" — is textually true of the paper's own account, but the citation cluster it actually points to for "has been studied" is mostly recent, tooling-rich work, not the tooling-poor 1980s-90s era the stalling narrative implies. The well-known 1980s critique that a neural Kolmogorov network doesn't work — Girosi and Poggio's 1989 "Kolmogorov's theorem is irrelevant" — is cited by the KAN paper as ref [20] in Section 2.1, in the passage conceding the theorem's one-variable functions "can be non-smooth and even fractal, so they may not be learnable in practice [19, 20]" — a passage directly engaging the 1989 critique, though still not grouped among the [9]-[16] prior-work citations. (The paper's Related Works section attributes the pathology point to a different reference, Poggio 2022 [66], again not to [20].)