Capture: Griewank 2012 — backward-pass description of Linnainmaa's algorithm and Copenhagen park anecdote

Claim 1: The Griewank 2012 PDF is currently unreachable — direct verification was not possible

Claim type: factual/access log
Source tier: N/A (this is the researcher's access record)

The canonical PDF of Griewank (2012) lives at two EMIS mirrors and a CiteSeerX cache. As of 2026-06-28, all three are unreachable (connection refused or SSL cert failure). The Semantic Scholar page (https://www.semanticscholar.org/paper/Who-Invented-the-Reverse-Mode-of-Differentiation-Griewank/8ff0c546aff84566635f0a9a2e01feb3d6588c1c) resolves but does not expose the full text. This entire capture rests on secondaries. Any promotion to atomic notes should be held pending a successful direct read.

The paper's bibliographic identity is uncontested across all sources:

• Griewank, A. (2012). "Who Invented the Reverse Mode of Differentiation?" In Grötschel, M. (Ed.), Optimization Stories, Documenta Mathematica, Extra Volume ISMP: 389–400.

Claim 2: The Copenhagen park anecdote is attributed to Griewank 2012, page ~391

Claim type: historical/biographical
Sourcing floor: Tier 3–4 acceptable for uncontested historical claims
Source tier achieved: Tier 4 (Toronto AI Meetup post) and Tier 3 (Psychology Today UK) — clears the floor for this claim type

The apparent near-verbatim text from Griewank 2012 (attributed to page 391 by the Toronto AI Meetup post, which reproduces Schmidhuber's account citing Griewank 2012):

"Seppo Linnainmaa (Lin76) of Helsinki says the idea came to him on a sunny afternoon in a Copenhagen park in 1970."

The (Lin76) parenthetical is the paper's in-text citation key for Linnainmaa's 1976 published paper — its presence in the attributed quote is consistent with it being extracted verbatim from an academic paper rather than paraphrased. Psychology Today UK (Cami Rosso, 2019) independently paraphrases the same event as "while in a Copenhagen park on a sunny afternoon," citing Griewank 2012 in its reference list.

Confidence note: The quote is attributed to Griewank 2012 p. 391 by two independent secondary sources. It is internally consistent (the citation key matches Griewank's bibliographic style). However, exact verbatim accuracy cannot be guaranteed without reading the paper. If this quote is used in a promoted note, add a caveat that it comes via a Tier 4 intermediary and requires direct confirmation.

Relevant existing notes: claim-linnainmaa-priority-not-paternity, claim-linnainmaa-rounding-error-problem, claim-linnainmaa-field-numerical-analysis

Claim 3: The backward-pass structural description of Linnainmaa's algorithm was NOT verified from Griewank 2012 directly

Claim type: technical mechanism
Sourcing floor: Tier 1–2 required
Source tier achieved for Griewank 2012 specifically: none (paper inaccessible)
Status: [unverified-mechanism -- needs primary] for Griewank 2012 specifically

Secondary sources characterize Griewank 2012 as describing Linnainmaa's contribution in terms of a "reverse procedure" that accumulates derivatives in the direction opposite to the forward evaluation sweep. However, the exact language Griewank 2012 uses for the structural description — whether he calls it a "backward pass," a "backward sweep," a "reverse sweep," or uses adjoint-propagation language — could not be confirmed.

What CAN be confirmed from a Tier 1 source (Griewank 2014, "On Automatic Differentiation and Algorithmic Linearization," Pesquisa Operacional 34(3), 2014, https://www.scielo.br/j/pope/a/kZ8M4LPYFryXQkdwHz7QXVx/?lang=en):

"The adjoint quantities v̄ⱼ and their combinations ūᵢ = (v̄ⱼ)ⱼ≺ᵢ ∈ ℝⁿⁱ can be propagated backwards by the reverse procedure listed in Table 5."

The 2014 paper's "Table 5" is captioned "Adjoint recursion for general evaluation procedure" and shows operations executed in reverse order compared to the forward sweep. This is the same Griewank writing four years later about the same algorithm; it establishes that his vocabulary for the backward structural step is "reverse procedure" / "adjoint recursion," not "backward pass." But attributing this to Griewank 2012 specifically is not confirmed.

Implication for existing notes: If claim-linnainmaa-reverse-mode-single-pass or backpropagation-gap assert that Griewank 2012 specifically uses "backward pass" language, that claim needs direct verification.

Relevant existing notes: claim-linnainmaa-reverse-mode-single-pass, backpropagation-gap, claim-linnainmaa-thesis-identity

Claim 4: Griewank 2012 records biographical context for Linnainmaa and his motivation — corroborated via secondaries

Claim type: historical/biographical
Sourcing floor: Tier 3–4 acceptable
Source tier achieved: Tier 3–4 via multiple consistent secondaries

Several secondary sources citing Griewank 2012 converge on these biographical details:

• Linnainmaa enrolled in the first computer science class at the University of Helsinki in 1967. [unverified-quant -- needs primary for the "1967" date specifically, sourced via search-engine snippet not traceable to a stable URL]
• Linnainmaa developed the reverse mode as a tool for estimating the propagation of rounding errors in complex programs — not for gradient-based learning. Multiple secondaries are consistent on this; it aligns with claim-linnainmaa-rounding-error-problem and claim-linnainmaa-field-numerical-analysis.
• The algorithm was written in Fortran at the end of his Master's thesis, which is otherwise in Finnish (this detail appears in Griewank 2014 p. 621 citing [23] = the 1970 thesis, not necessarily in Griewank 2012 — distinguish sources carefully).

Further leads

• Griewank 2014 (Tier 1, accessible) — https://www.scielo.br/j/pope/a/kZ8M4LPYFryXQkdwHz7QXVx/?lang=en — contains a Table 5 "Adjoint recursion for general evaluation procedure" showing the backward sweep formally. This could support a Tier-1 claim about what the backward structural description looks like without requiring the inaccessible 2012 paper.
• Schmidhuber "Who Invented Backpropagation?" page — https://people.idsia.ch/~juergen/who-invented-backpropagation.html — cites Griewank 2012 as [BP5] but does not reproduce the Copenhagen park anecdote in its current form (page may have been revised since 2019 Toronto AI Meetup repost).
• Linnainmaa's original 1976 paper ("Taylor expansion of the accumulation of floating point rounding errors," Annales Academiae Scientiarum Fennicae) — the "(Lin76)" entry cited in the Copenhagen park quote — would be Tier 1 if accessible.
• Try EMIS again when accessible — both mirrors have failed today; retry when servers recover to get the primary text directly.