Capture: Does Table 5 in Griewank 2014 independently support a Tier-1 claim about Linnainmaa's backward pass?

Claim 1: Table 5 formally specifies the adjoint recursion — the mathematical structure of the backward sweep

Claim type: technical-mechanism Sourcing floor: Tier 1-2 required Source tier achieved: Tier 1 (direct primary read, p. 629)

Table 5, captioned "Adjoint recursion for general evaluation procedure," reads verbatim:

"v̄ℓ−i ≡ ȳm−i    i = 0 . . . m − 1 ūi += v̄i · φ'i(ui)    i = ℓ . . . 1 x̄j ≡ v̄j−n    j = n . . . 1"

This is the formal recursion for propagating adjoint ("bar") quantities backward through a general evaluation procedure — i.e., the mathematical object underlying what is colloquially called the "backward pass" in automatic differentiation and backpropagation.

Claim 2: Griewank 2014 gives an explicit, formal statement of the "backward" (reverse-order) structural property, in his own words, immediately below Table 5

Claim type: technical-mechanism Sourcing floor: Tier 1-2 required Source tier achieved: Tier 1 (direct primary read, p. 629)

Immediately after Table 5, the paper states:

"As one can see, adjoint statements += v̄i · φ'i(ui) are executed in the opposite order in which the underlying statements vi = φi(ui) occurred in the original program. The same is true for the initialization of the ȳl−i, which are input variables, and the deinitialization of the x̄j, which are output variables."

This is the formal structural description of what makes the sweep "backward": the adjoint statements execute in the exact reverse of the order the forward (primal) statements executed in. This is a self-contained, Tier-1, primary-source statement of the mechanism — it does not require or cite the 2012 paper for this specific claim.

Note on vocabulary: throughout the 2014 paper, Griewank's own terms for this are "reverse procedure," "reverse sweep," and "reverse mode" (e.g. p. 630: "Here we account for executing the forward sweep Table 2 and the reverse sweep Table 5 after one another"). The term "backward pass" (or "backward sweep") as a literal phrase does not appear in the paper's own vocabulary near Table 5 — this confirms and sharpens the finding of the 2026-06-28 capture on this same topic. Any note that attributes the literal phrase "backward pass" to Griewank should flag it as ML-community shorthand for his "reverse sweep/procedure," not his own wording.

Claim 3: Griewank 2014 directly attributes authorship of this reverse procedure to Linnainmaa, citing Linnainmaa's own 1970 thesis — not the 2012 paper — as the source of the attribution

Claim type: historical/biographical (load-bearing — escalated to Tier 1-2 per the rubric, since this is the central disputed point) Sourcing floor: Tier 1-2 (escalated; this is the load-bearing point of the whole line of inquiry) Source tier achieved: Tier 1 (direct primary read, p. 629)

The same paragraph continues, in Griewank's own authorial voice:

"Apparently the first author to write down this reverse procedure was Seppo Linnainmaa, who listed it in Fortran at the end of his Master Thesis [23], which is otherwise written in Finnish. He interpreted and used the quantities v̄i to estimate the propagation of errors in complicated programs. For more information on the history of the reverse mode see [16]."

Reference [23] in the paper's bibliography (p. 645) resolves to:

"LINNAINMAA S. 1970. The representation of the cumulative rounding error of an algorithm as a Taylor expansion of the local rounding error, Master's thesis (in Finnish), Department of Computer Science, University of Helsinki."

Reference [16] (p. 644) resolves to the paper this capture was trying to route around:

"GRIEWANK A. 2012. Who invented the Reverse Mode of Differentiation? In: MARTIN GRÖTSCHEL, (Ed.). Optimization Stories, Documenta Mathematica, Extra Volume ISMP (2012): 389–400."

The critical structural point: [16] (the 2012 paper) is cited only as a pointer "for more information on the history of the reverse mode" — a further-reading reference, not the citation attached to the attribution claim itself. The attribution claim ("the first author to write down this reverse procedure was Seppo Linnainmaa... who listed it in Fortran at the end of his Master Thesis") is cited to [23], Linnainmaa's own 1970 thesis, and is stated directly in Griewank's own words in the 2014 paper. This means the attribution is independently supported by a Tier-1 primary source (Griewank 2014 itself, in propria persona) without needing to read or rely on the content of the 2012 paper.

Related notes: claim-linnainmaa-priority-not-paternity, claim-linnainmaa-thesis-identity, claim-linnainmaa-rounding-error-problem, claim-linnainmaa-field-numerical-analysis, claim-linnainmaa-reverse-mode-single-pass, backpropagation-gap

Claim 4: A worked small-scale example of the reverse sweep (Table 6) confirms the mechanism is fully general and executable, not merely schematic

Claim type: technical-mechanism Sourcing floor: Tier 1-2 required Source tier achieved: Tier 1 (direct primary read, p. 629)

Griewank 2014 follows Table 5 with a worked example, "Table 6 – Reverse mode on small example," applied to the toy function from Table 1 (F(x) = [log(x1)·(x2+x3)/sin(x1), √x3 − exp(x2+x3)]). The text notes:

"Starting from their initial value v̄i = 0 for i ≤ l − m the adjoint quantities may be incremented repeatedly until they reach their final value and then occur once on the right-hand side as pre-factor of φ'i. This can be seen in the adjoint procedure Table 6... Here we see that v̄i for i = 3, 0 − 2 are each incremented twice before they reach their final value and can then occur on the right hand side."

This is a secondary, corroborating confirmation that Table 5's general recursion is not just notation — it is demonstrated as a concrete, steppable procedure, reinforcing that the "reverse procedure" language used for Linnainmaa in Claim 3 refers to this same fully-specified mechanism.

Further leads

• Griewank 2012 primary text — still unreached as of this run (prior captures record ECONNREFUSED/ECONNRESET across EMIS mirrors and CiteSeerX). This capture shows that reaching it is no longer necessary for the specific Linnainmaa/backward-pass attribution claim, since Griewank 2014 makes that claim independently and in his own words. It would still be valuable for the Copenhagen park anecdote (not repeated in the 2014 paper) and for any "priority but no paternity"-style framing, which was not found anywhere in the 2014 text.
• "Priority but no paternity" phrasing: not found anywhere in the Griewank 2014 full text read in this run (all 25 pages were at least skimmed via section headers; Sections 1-3 and 5 plus references were read in full detail). This increases confidence that the earlier flag — that this phrase is Yuxi Liu's own framing (Tier 3), not Griewank's — is correct, though this is not a direct disproof since the phrase could in principle appear in the untranscribed Section 4 pages (633-640), which were not read in this run.
• Ostrowski citation — reference [29] in the 2014 bibliography: "OSTROVSKII GM, VOLIN YU M & BORISOV WW. 1971. Über die Berechnung von Ableitungen. Wiss. Z. Tech. Hochschule für Chemie, 13: 382–384." This is a Tier-1-citable primary reference for the earlier-noted claim (previously Tier 4/unverified) that a chemical engineer used reverse-mode differentiation before Linnainmaa without publishing broadly; the 1971 date here is later than Linnainmaa's 1970 thesis, which is worth flagging as a discrepancy against the "five years before Linnainmaa" claim in the 2026-06-28 capture (sourced there only from a Tier 4 blog). This 1971 publication date does not by itself establish when Ostrowski's group first used the technique, so it neither confirms nor refutes the "used it earlier" claim — but it does mean any promoted note citing Ostrowski's priority should use this Tier-1 bibliographic entry rather than the Tier-4 blog account.