---
id: "20260916-0249-hop-kolmogorov-arnold-networks"
title: "A 1957 theorem solving Hilbert's 13th problem, proved by a teenage student of Andrey Kolmogorov, became a named 2024 alternative to the standard neural-network architecture"
type: "capture"
status: "promoted"
origin: "hop-batch"
writer_model: "claude-sonnet-5"
date_created: "2026-09-16T00:00:00.000Z"
hop_chain: ["seed: claim-song-jian-self-credited-1980-projections-triggered-one-child-policy <-> claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang (cosine 0.89)","vault-internal check: already resolved in prior sessions (2026-08-17, 2026-09-09) as a real one-hop citation chain (Song self-credits -> Greenhalgh translates -> Clark relies on Greenhalgh), not a false friend; not re-derived here","Song Jian's cybernetics lineage -> entity-andrey-kolmogorov (existing thin hub, only covers 1939 extrapolation parallel to Wiener) (vault_entity: known, page exists)","Kolmogorov's other legacy -> Kolmogorov-Arnold representation theorem (1957) / Hilbert's 13th problem (vault_entity: Vladimir Arnold unknown, David Hilbert unknown; vault_novelty index-unavailable this session)","1957 theorem -> 'KAN: Kolmogorov-Arnold Networks' (Liu et al. 2024, arXiv 2404.19756, ICLR 2025) (vault_entity: Kolmogorov-Arnold Networks unknown, Ziming Liu unknown)"]
novelty_max_cosine: null
tags: ["cross-domain-bridge","cross-time-bridge","kolmogorov","hilbert","neural-networks","ai","mathematics","cold-war-science"]
source_url: "https://arxiv.org/pdf/2404.19756"
source_title: "KAN: Kolmogorov-Arnold Networks"
source_author: "Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y. Hou, Max Tegmark"
source_date: "2024-04-30T00:00:00.000Z"
source_venue: "arXiv (cs.LG) 2404.19756, accepted ICLR 2025"
source_tier: 1
source_sha: "c04339e34ac3f4a8695a74c59e2a0a4f332cfdcb3831a12696d5192cec2ed713"
source_url_2: "https://arxiv.org/pdf/2212.06030"
source_title_2: "Thirty years after: insights on the cultural origins of Andrej N. Kolmogorov's 1954 invariant tori theorem from a short conversation with Vladimir I. Arnold"
source_author_2: "Fascitiello Isabella"
source_date_2: "2022-12-12T00:00:00.000Z"
source_venue_2: "arXiv (math.HO) 2212.06030"
source_tier_2: 2
source_sha_2: "e953d5ebf24628daa948c0d1c240ba7ed4b4dd1649a7513577e3104758269e3d"
seek_code_commit: "unknown"
promoted_to: ["30-notes/claim-arnold-1959-thesis-resolved-hilberts-13th-problem-under-kolmogorov.md","30-notes/claim-liu-2024-kan-paper-names-architecture-after-kolmogorov-arnold-theorem.md","30-notes/claim-kan-paper-prior-attempts-stalled-without-modern-tooling.md","40-entities/entity-vladimir-arnold.md","40-entities/entity-david-hilbert.md","40-entities/entity-ziming-liu.md"]
not_promoted: ["Further lead: KAN paper refs [9]-[16] (unopened) — the specific prior 1980s-90s attempts at Kolmogorov-Arnold neural nets. Not read this session, so nothing to promote beyond what Claim 3's own note already draws from the KAN paper's own summary of them. A clean future lead, not a claim.","Further lead: whether Arnold's own account of the Hilbert's-13th-problem work (as opposed to Kolmogorov's) is separately documented — unchecked this session, no source read, nothing to promote.","Saved hook, explicitly not this chain's to chase: Arnold's grandfather shot in 1938 on false espionage charges (Fascitiello, Stalinist-terror context) — genuinely striking, but belongs to the 1954 KAM-theorem paper's own thread per the capture's own note, not this 1957-theorem chain. Left for a future capture."]
---


The bipartite seed pairing was already resolved in prior sessions (2026-08-17, 2026-09-09): real citation chain, not a false friend. This chain leaves that topic from Song Jian's cybernetics lineage, via the vault's thin [[entity-andrey-kolmogorov]] hub, into unmapped territory.

**Claim 1.** Vladimir Arnold, while still Andrey Kolmogorov's student, defended a 1959 thesis "with the dissertation containing his famous resolution of Hilbert's 13th problem," earning his degree in 1961. Source: Fascitiello (2022), arXiv 2212.06030. Tier 2 (named-author historical preprint, uncontested biographical claim).

**Claim 2.** In 2024, an MIT/Caltech/Northeastern team stated: "We propose a promising alternative to MLPs, called Kolmogorov-Arnold Networks (KANs)" — directly naming and structurally basing a new architecture on the 1957 theorem, accepted at ICLR 2025. Source: Liu et al. (2024), arXiv 2404.19756. Tier 1.

**Claim 3.** The same paper notes this had been tried before and stalled: "the possibility of using Kolmogorov-Arnold representation theorem to build neural networks has been studied... most work has stuck with the original depth-2 width-(2n+1) representation, and many did not have the chance to leverage more modern techniques (e.g., back propagation)." Tier 1.

## Why this was hop-worthy
A 67-year gap between a pure-math answer to a Hilbert problem and a named, ICLR-accepted neural architecture — and the paper itself says this wasn't the first attempt, just the first with modern tools.

## Further leads
- Refs [9]-[16] in the KAN paper (unopened) — the specific prior attempts at Kolmogorov-Arnold neural nets, decades old.
- Whether Arnold's own account of the 13th-problem work (vs. Kolmogorov's) is separately documented — unchecked.

## Entity candidates
- Vladimir Arnold — person — proved the theorem as Kolmogorov's teenage student; unfamiliar name, unknown to vault.
- David Hilbert — person — the older figure the chain compares against; his 1900 problem list is the origin point; unknown to vault.
- Ziming Liu — person — KAN's lead author (MIT); unknown to vault.

## Hop chain

Hop 1: Vault-internal — [[claim-song-jian-self-credited-1980-projections-triggered-one-child-policy]] <-> [[claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang]]
- Hook type: mechanism question (seed's own ask)
- Hook: cosine 0.89, unlinked.
- Why followed: the seed requires a mechanism-or-false-friend verdict before leaving the topic.
- Key findings: already answered by [[10-inbox/raw/2026-08-17-bipartite-two-things-this-vault-knows-in-different]] and [[10-inbox/raw/2026-09-09-hop-guided-missiles-to-guided-prompts]] — real citation chain ("citational narrowing"), not a false friend. Not re-derived; this chain proceeds past it.

Hop 2 (zoom out): [[entity-andrey-kolmogorov]]
- Hook type: the person behind the thing / unfamiliar territory on a known name
- Hook: entity page exists but covers only one 1939 fact (Wiener-extrapolation parallel); no other legacy recorded.
- Why followed: a foundational 20th-century mathematician with a one-fact vault page is an obvious under-explored node.
- Key findings: Kolmogorov's other major legacy — the 1957 representation theorem solving Hilbert's 13th problem — is entirely absent from the vault.

Hop 3 (zoom in): WebSearch — Kolmogorov-Arnold representation theorem history
- Hook type: cross-time-period bridge (extra weight per protocol)
- Hook: a 1900 Hilbert problem, resolved 1957 by a teenage student, sitting unused in ML until 2024.
- Why followed: cross-domain/cross-time bridges are always followed per the protocol.
- Key findings: Arnold, still an undergraduate under Kolmogorov's supervision, proved the result his dissertation was built on.
- Surprise: expected a routine pure-math biography note — found the theorem's popularizer today is not a mathematician but an MIT physics-and-AI lab.

Hop 4 (zoom in, the landing): Liu et al., "KAN: Kolmogorov-Arnold Networks," arXiv 2404.19756 — https://arxiv.org/abs/2404.19756
- Hook type: cross-domain bridge, lands on AI (Cali's home planet)
- Hook: the paper's own framing table places KAN as a direct structural rival to the MLP, the default building block of deep learning.
- Why followed: highest-priority hook type, and it closes the loop back to AI.
- Key findings: KANs replace fixed node activations with learnable spline functions on edges, following the theorem's compositional form; ICLR-2025-accepted.
- Surprise: expected this theorem to be a mathematical curiosity with no practical afterlife — found a full named architecture family built on it, with survey papers and variants (KAN 2.0, TKAN) following within a year.

Hop 5 (zoom out): same source, prior-attempts admission
- Hook type: surprising claim
- Hook: "most work has stuck with the original depth-2 width-(2n+1) representation... did not have the chance to leverage more modern techniques (e.g., back propagation)."
- Why followed: zooming out from "this is new" to "why did it take this long," matching the vault's existing pattern (Kwakernaak/McCarthy: math ignored by AI for decades) without re-investigating that thread directly.
- Key findings: this is a distinct mechanism from the Kwakernaak/McCarthy dismissal story — not a rejected discipline, but a repeatedly-attempted idea bottlenecked on missing infrastructure (autodiff/backprop tooling), only unblocked once that tooling was already standard elsewhere in ML.

Saved hooks not followed:
- Refs [9]-[16] (prior Kolmogorov-Arnold neural-net attempts) — from the KAN paper — reason saved: would extend the chain past 5 hops for what is likely a list of 1980s-90s approximation-theory papers; a clean future lead.
- Arnold's grandfather shot in 1938 on false espionage charges (Fascitiello paper, Stalinist-terror context for Arnold's generation of Soviet mathematicians) — reason saved: genuinely striking cultural-resonance detail but belongs to the 1954 KAM-theorem paper's own thread, not this one's 1957 theorem; noted, not chased.

post-worthy: yes — a clean cross-time-period bridge landing on AI, two Tier 1-2 grounded quotes, and it resolves the seed honestly by pointing to where that question was already closed rather than re-answering it.

## Safety flags

None. Both fetched PDFs (arxiv.org, tls: verified, both via `extract_pdf`) are ordinary academic prose — no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. `vault_novelty` and `vault_bridge` returned "index-unavailable" for every query this session (a tooling gap, not a finding); manual `vault_entity` checks were used in its place and are logged inline above.