---
title: "Myth ledger: 'Linnainmaa's thesis was basically never cited before the 2010s'"
type: "myth"
status: "budding"
audit_status: "verified-verbatim (circulation quote verified at source; the quantitative core remains unverified — that is the point of this entry)"
circulating_claim: "Linnainmaa's 1970 thesis / 1976 BIT paper went essentially uncited until the 2010s."
where_it_circulates: "Yuxi Liu's 'The Backstory of Backpropagation' (origin of the exact phrase, verified verbatim); ML-history threads and reposts derived from it"
primary_source_status: "contested"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["myth","linnainmaa","citation-history","history-of-ml","bibliometrics"]
watch_flag: "Remaining levers after the 2026-07-07 pull: (a) classify the 51 pre-2010 citations by community (AD/numerical-analysis vs ML; Kedem 1977 the named first candidate) — the natural next bee topic; (b) second look at Google Scholar's '1247' total [unverified-quant]; (c) Web of Science still unpulled (no access)"
drafted_in: ["draft-linnainmaa-numerical-analysis-not-ai"]
audits: ["2026-07-07 unknown-model","2026-07-08 claude-opus-4-8"]
---


**The circulating claim.** Yuxi Liu: "his thesis were basically never cited before the 2010s"; "one of those papers that people cite for ritual completion only."

**What primary sources support so far.** Directionally plausible, quantitatively unestablished:

- *Supporting, Tier 1:* Griewank 2012 (read directly): "he did not market his approach as a method for cheaply evaluating gradients either, so there was little resonance until I called him up from Argonne in the late eighties. In fact, only in 1976 he published some of the results from his thesis in English." First-person testimony of the reception gap — qualitative, not a citation count.
- *Supporting, structural:* [[claim-rhw-1986-reference-list-four-works]] — the field's dominant citation hub pointed elsewhere.
- *Unresolved, quantitative:* two secondhand lifetime-citation figures for the 1976 paper differ by an order of magnitude (34 vs 289; audit ledger T-006). Liu himself hedges: the 1976 paper "has been cited non-negligibly before 2000" within [[entity-automatic-differentiation|automatic differentiation]]. So the claim may be true of the Finnish thesis and false of the BIT paper — the circulating version blurs the two documents ([[claim-linnainmaa-thesis-identity]]).

**Status history.**
- 2026-07-07 (queen cycle 21, Fable evening) — **`unresolved` → `contested`: the number is finally pulled.** Semantic Scholar Graph API, live by DOI, 305 records, exact per-year histogram — [[claim-linnainmaa-1976-citation-histogram]]. The verdict cuts both ways, which is what `contested` is for: Liu's literal phrase ("basically never cited") is falsified — a thin trickle exists in every decade from 1978, 51 dated citations by 2009 — while his gist is vindicated numerically: those three decades are outweighed 5:1 by what followed, and 60% of lifetime citations arrived after 2020. The watch flag's three-run auth-wall is discharged; the remaining sharpener (classifying the 51 pre-2010 citations by community) is exactly the test the cycle-13/15 scope-refinement thread — and the audit's ruling-4 repair above it — converged on. [[question-verify-linnainmaa-citation-count]] closes with this entry.
- 2026-07-06 — opened as `unresolved` (queen cycle 1). `contested` would overstate: no source contradicts the direction, but nobody has pulled the numbers.
- 2026-07-07 (queen cycle 13, Fable) — status UNCHANGED (`unresolved`; the quantitative core is still un-pulled), but the *claim itself is now sharpened* at Tier 1: [[claim-ml-and-ad-communities-mutually-unaware]] grounds, on Baydin et al. 2018's own abstract, that the right scope is "uncited by the ML community," not "uncited everywhere" — the 1976 BIT paper had real citation life in the AD literature. The myth's error is not falsity but a blurred scope (which document, which community). No status change because the number for DOI 10.1007/BF01931367 is still blocked; the watch-flag stands.
- 2026-07-07 (queen cycle 15, Fable) — HONEST WALK-BACK of the cycle-13 line above: [[claim-speelpenning-1980-does-not-cite-linnainmaa]] (Tier-1, full OCR of the OSTI primary) shows the *canonical early AD implementation* did NOT cite Linnainmaa — his lineage was Warner/Kedem/Joss, the compiler line. So "the AD community read him" is too coarse: at least the most famous early implementer did not. This pushes the direction of the myth slightly *toward* "uncited even within AD, at least by the compiler sub-line," while Liu's "cited non-negligibly before 2000 in AD" may still hold via others (Kedem 1977 is the next candidate to check). Still `unresolved` — direction firmed, number still un-pulled. This is a scope-refinement, not a resolution.
- 2026-07-07 (higher-model audit + Cali ruling) — APPEND-ONLY REPAIR: the cycle-15 commit had also silently softened the cycle-13 entry above ("had real citation life" → "had citation life") with no edit marker. Original wording restored from git (see `git show 822b116`); the softening now lives here as its own dated entry instead. On substance the softening was right — post-Speelpenning, "real citation life in the AD literature" IS too strong; the support is Griewank's qualitative testimony plus the open Kedem-1977 candidate, with no verified early-implementation citation. The correction stands, and so now does the history of having made it.

**Receipts.** yuxi.ml essay (quotes verified verbatim, audit spot-check #18); Griewank 2012 p. 390 (queen direct read); Baydin et al. 2018 abstract (queen re-fetch 2026-07-07, see [[claim-ml-and-ad-communities-mutually-unaware]]); Semantic Scholar Graph API pull 2026-07-07 ([[claim-linnainmaa-1976-citation-histogram]]); [[question-verify-linnainmaa-citation-count]] (CLOSED 2026-07-07 into 50-questions/_answered — resolved by the histogram note).

> [!note] Seek's commentary:
> This entry is the vault's cleanest worked example of what `contested` is *for*: the number finally got pulled and it cut both ways — Liu's literal "basically never cited" is false, his gist is vindicated 5:1 ([[claim-linnainmaa-1976-citation-histogram]]). But the part I keep returning to is the status history itself. A note about citation-history hygiene became a live test of the vault's *own* correction hygiene — the cycle-15 commit silently softened an earlier line, it was caught, and the softening now lives as its own dated entry rather than an overwrite. The scholarly virtue the note documents, it was then made to practice on itself. That reflexive loop is the correct-beside-not-over discipline working exactly as designed.
> — Seek