What does Baydin et al. 2018 (arXiv:1502.05767) say about Linnainmaa's specific contribution to autodiff, and does it call his work the origin of modern AD?
Core question
Does Baydin, Pearlmutter, Radul & Siskind's survey "Automatic Differentiation in Machine Learning: a Survey" (arXiv:1502.05767) characterise Seppo Linnainmaa's work as the origin of modern automatic differentiation, or does it describe his contribution more narrowly?
Finding: the survey mentions Linnainmaa by name in its §3.3 "Origins of AD and Backpropagation," but does not crown him as sole origin. It places his thesis inside a longer, hedged lineage — control-theory antecedents before him, Werbos narrated just ahead of him in the prose despite a later date, and Speelpenning credited with the first automatic implementation after him — and explicitly defers the authoritative historical verdict to a different paper (Griewank, 2012) rather than rendering one itself. The full text of the arXiv v4 PDF (43 pages) was read directly via extract_pdf (sha256 6f72720b86da67c0471efb879677e3909cdf32ff4d73c006663b8c441a4cf552, TLS verified) — §3.3 is on p.13-14 of the PDF.
Note: an earlier capture in this vault, 10-inbox/raw/20260701-1639-does-baydin-et-al.md, concluded (on 2026-07-01, using a truncated fetch) that the word "Linnainmaa" did not appear anywhere in the retrievable text. That finding was already corrected on 2026-07-07 per the audit note folded into claim-ml-and-ad-communities-mutually-unaware's frontmatter. This capture's full-PDF read confirms the correction and extends it with the surrounding sentence structure, which the 2026-07-07 correction did not quote in full.
Claim: Baydin et al. hedge Linnainmaa's priority as a citation convention, not as their own historical verdict
The survey's §3.3 narrates reverse-mode AD's lineage in this sequence: the Pontryagin maximum principle (Rozonoer 1959; Boltyanskii et al. 1960) as "the essence of the reverse mode" in continuous-time form; the control-theory community (Bryson & Denham 1962; Bryson & Ho 1969); then Werbos (1974), credited with casting it "in more formal terms with discrete-time variables topologically sorted in terms of dependency"; and only after Werbos does the text introduce Linnainmaa — even though Linnainmaa's thesis (1970) predates Werbos's (1974) by four years. The attribution to Linnainmaa is explicitly hedged as a citation convention rather than asserted as the authors' own settled fact:
"This method was understood in the control theory community (Bryson and Denham, 1962; Bryson and Ho, 1969) and cast in more formal terms with discrete-time variables topologically sorted in terms of dependency by Werbos (1974). Prior to Werbos, the work by Linnainmaa (1970, 1976) is often cited as the first published description of the reverse mode."
source_url: https://arxiv.org/pdf/1502.05767 · source_sha: 6f72720b86da67c0471efb879677e3909cdf32ff4d73c006663b8c441a4cf552 · source_author: Baydin, Pearlmutter, Radul, Siskind · source_date: 2018-02-05 · source_title: "Automatic Differentiation in Machine Learning: a Survey" · source_venue: arXiv preprint v4 (JMLR 18(153):1-43, 2018) · source_tier: 1
(The same sentence is already quoted in claim-ml-and-ad-communities-mutually-unaware's 2026-07-07 correction note, in support of a different claim — ML/AD community unawareness. This capture's distinct angle is the "is often cited as" hedge itself and its position in the narrative sequence, which bears directly on whether the survey treats Linnainmaa as the origin.)
Claim: the survey credits Speelpenning, not Linnainmaa, with the first actual automatic implementation of reverse-mode AD
Immediately following the Linnainmaa sentence, the survey draws an explicit distinction between "first published description" (attributed to Linnainmaa, hedged) and "first implementation that was actually automatic" (attributed to Speelpenning, unhedged):
"Speelpenning (1980) subsequently introduced reverse mode AD as we know it, in the sense that he gave the first implementation that was actually automatic, accepting a specification of a computational process written in a general-purpose programming language and automatically performing the reverse mode transformation."
This matters for the core question because it shows the survey splitting "origin of modern AD" credit across at least three figures (Werbos for formalization, Linnainmaa for first published description, Speelpenning for the first working automatic system) rather than assigning it to Linnainmaa alone. See claim-speelpenning-1980-does-not-cite-linnainmaa for the separate, already-vaulted finding that Speelpenning's own paper does not cite Linnainmaa — a different claim (about Speelpenning's reference list) from this one (about how Baydin et al. retrospectively rank the two men's contributions).
source_url: https://arxiv.org/pdf/1502.05767 · source_sha: 6f72720b86da67c0471efb879677e3909cdf32ff4d73c006663b8c441a4cf552 · source_author: Baydin, Pearlmutter, Radul, Siskind · source_date: 2018-02-05 · source_title: "Automatic Differentiation in Machine Learning: a Survey" · source_venue: arXiv preprint v4 (JMLR 18(153):1-43, 2018) · source_tier: 1
Claim: the survey explicitly declines to adjudicate the origin question itself, deferring to Griewank (2012)
Rather than rendering its own verdict on who originated the reverse mode, the survey closes its historical section by pointing readers to a different, dedicated paper for that investigation:
"For a thorough review of the development of AD, we advise readers to refer to Rall (2006). Interested readers are highly recommended to read Griewank (2012) for an investigation of the origins of the reverse mode and Schmidhuber (2015) for the same for backpropagation."
This directly answers the second half of the core question: Baydin et al. 2018 does not itself characterise Linnainmaa's work as "the origin of modern automatic differentiation." It names him as one hedged claimant ("is often cited as") within a multi-figure lineage (Pontryagin/control theory → Werbos → Linnainmaa → Speelpenning), and explicitly outsources the origin-determination to Griewank (2012) — the paper this vault already treats as the Tier 1 primary investigation of the same question (see claim-linnainmaa-priority-not-paternity and claim-linnainmaa-reverse-mode-single-pass).
source_url: https://arxiv.org/pdf/1502.05767 · source_sha: 6f72720b86da67c0471efb879677e3909cdf32ff4d73c006663b8c441a4cf552 · source_author: Baydin, Pearlmutter, Radul, Siskind · source_date: 2018-02-05 · source_title: "Automatic Differentiation in Machine Learning: a Survey" · source_venue: arXiv preprint v4 (JMLR 18(153):1-43, 2018) · source_tier: 1
Further leads
- The survey's bibliography carries two separate Linnainmaa entries read directly from the PDF (p.36 of the printed survey): "Seppo Linnainmaa. The representation of the cumulative rounding error of an algorithm as a taylor expansion of the local rounding errors. Master's thesis, University of Helsinki, 1970." and "Seppo Linnainmaa. Taylor expansion of the accumulated rounding error. BIT Numerical Mathematics, 16(2):146-160, 1976." — confirms the vault's existing title/venue metadata for both works via a second independent Tier-1 source. (Baydin et al. 2018, same PDF as above.)
- The survey attributes "forward mode AD as a general method for evaluating partial derivatives" to Wengert (1964) and names Griewank (1989) as the figure who revived AD research interest in the 1980s — both one-line historical attributions not otherwise explored in this capture. (Baydin et al. 2018, §3.3.)
- Hecht-Nielsen (1989) is cited by Baydin et al. as naming Bryson & Ho (1969) and Werbos (1974) as "the two earliest known instances of backpropagation" — a separate, narrower priority claim about backpropagation specifically (not reverse-mode AD in general) that does not mention Linnainmaa at all. Worth a dedicated capture on whether this is a genuine disagreement with the Linnainmaa-first framing or simply a narrower scope (backprop-as-applied-to-networks vs. reverse-mode-AD-as-general-technique). (Baydin et al. 2018, §3.3.)
Entity candidates
- Andreas Griewank — person — Baydin et al. explicitly defer the origin-of-reverse-mode question to his 2012 paper rather than adjudicating it themselves; he is the figure the whole priority claim is measured against, not a co-author of this survey.
- Seppo Linnainmaa — person — the subject of the core question; already has vault entries, but this capture adds the specific "is often cited as" hedge and his position in Baydin et al.'s narrative sequence.
- Paul Werbos — person — credited by the survey with casting reverse-mode AD "in more formal terms with discrete-time variables topologically sorted in terms of dependency," and narrated immediately before Linnainmaa despite a later date (1974 vs. 1970) — a framing choice worth its own note.
- Bert Speelpenning — person — credited by the survey, not Linnainmaa, with "reverse mode AD as we know it" (first automatic implementation); the survey's own distinction between "first published description" and "first automatic implementation" rests on separating these two men's contributions.
- Pontryagin maximum principle — concept — the survey identifies this (via Rozonoer 1959; Boltyanskii et al. 1960) as "the essence of the reverse mode," predating all the named AD/backprop claimants and complicating any single-origin story.
- Jürgen Schmidhuber — person — named alongside Griewank as the survey's deferred authority, specifically "for the same for backpropagation" (as opposed to reverse-mode AD generally).