Linnainmaa has priority over reverse-mode automatic differentiation but not paternity — his work had no traceable causal influence on backpropagation's adoption in AI

Seppo Linnainmaa's 1970 Finnish master's thesis is the earliest documented formulation of reverse-mode automatic differentiation — the mathematical procedure underlying backpropagation. On this basis, he has priority: his is the first published instance of the algorithm.

He does not have paternity. His work had no traceable causal influence on the AI community's independent rediscovery of the same algorithm. Yuxi Liu's historical essay on the backstory of backpropagation gives the concise formulation: "He has priority but no paternity."

The independent rediscoveries

At least three independent derivations of backpropagation followed Linnainmaa — each without apparent awareness of his work:

• Paul Werbos (1974): Harvard PhD dissertation, developing backpropagation to formalize systems-theoretic models of psychic energy (see backpropagation-gap for the Freudian context). First application to multilayer perceptrons. Published more widely in 1981.
• Rumelhart, Hinton, and Williams (1986): Nature paper that demonstrated backpropagation's empirical power for cognitive tasks and became the canonical citation for the AI community. Rumelhart was unaware of both Linnainmaa and Werbos at the time of writing.
• Other parallel derivations exist in the optimal control literature (Pontryagin adjoint state method, Dreyfus 1962 chain-rule formulation); see backpropagation-gap for the full multi-domain trajectory.

Why the gap existed

Three structural factors isolated Linnainmaa's contribution from the AI lineage:

1. Language: The 1970 thesis was in Finnish. A Finnish-language master's thesis in 1970 had effectively zero cross-disciplinary reach.
2. Venue: The English publication (1976) appeared in BIT Numerical Mathematics — a numerical analysis journal with no overlap with the neural network community. See claim-linnainmaa-field-numerical-analysis.
3. Framing: The thesis addressed floating-point rounding error, not learning or optimization. AI researchers solving the credit-assignment problem in the 1970s–1980s were not reading the floating-point error analysis literature.

The citation gap was acute enough that Griewank (Tier 1) found it necessary to write a dedicated historical paper in 2012 — titled "Who Invented the Reverse Mode of Differentiation?" — four decades after the fact. The very necessity of that paper is evidence that the lineage was not traced until recently.

Note: the specific quantitative claim that the thesis was "basically never cited before the 2010s" comes from Yuxi Liu's essay (Tier 4) and has not been verified against citation data. See question-verify-linnainmaa-citation-count.

Vault connection

This priority/paternity distinction is the precise structure that makes backpropagation-gap interesting. The gap is not merely a citation accident; it is evidence that the algorithm training every modern LLM was not developed by or for AI researchers. It was borrowed — or more precisely, independently re-derived — without any channel of transmission from its origin. Werbos 1974 counts as an independent discoverer for the same reason Linnainmaa does: neither influenced the other, and neither influenced Rumelhart 1986.

Sources

• Liu, Yuxi. "The Backstory of Backpropagation." https://yuxi.ml/essays/posts/backstory-of-backpropagation/ (accessed 2026-06-26). Tier 4 (personal essay). Exact quotes confirmed in research session: "He has priority but no paternity"; "his thesis were basically never cited before the 2010s"; "one of those papers that people cite for ritual completion only."
• Schmidhuber, Jürgen. "Who Invented Backpropagation?" https://people.idsia.ch/~juergen/who-invented-backpropagation.html (accessed 2026-06-26). Tier 2. Frames Linnainmaa as inventor but does not claim the AI community was aware of him.
• Griewank, Andreas. "Who Invented the Reverse Mode of Differentiation?" Documenta Mathematica, Extra Volume ISMP, 2012, pp. 389–400. Tier 1. Not directly accessed. Its existence as a 2012 history-of-mathematics paper is itself evidence that the lineage was traced only recently.
• Wikipedia. "Backpropagation — History." https://en.wikipedia.org/wiki/Backpropagation#History (accessed via backpropagation-gap session, 2026-05-31). Tier 3. Notes Rumelhart was unaware of Werbos and Linnainmaa.

Revisit 2026-07-06 (queen cycle 1) — the primary text complicates "priority"

Griewank 2012 was finally read directly (capture 2026-07-06, re-verified by queen pdftotext extraction). Its prologue argues multiple independent discovery, not sole priority: "there have been many incarnations of this reversal technique, which has been suggested by several people from various fields since the late 1960s, if not earlier" — with Ostrowski "some five years earlier" (~1965, chemical engineering) and Hachtel et al. (1960s, circuit theory) named before Linnainmaa's section begins. Griewank centers Linnainmaa for stated narrative-scope reasons, "since we do not wish to specifically address electronic circuits or chemical processes," not as a priority verdict.

What survives of this note unchanged: the paternity half (no causal influence on AI's rediscoveries) — Griewank's own "little resonance until I called him up from Argonne in the late eighties" strengthens it to Tier 1. The priority half should now be read as: earliest general, formal statement of reverse-mode accumulation (which is what Griewank specifically credits), not earliest use of the reversal idea. Yuxi Liu's "priority but no paternity" epigram, quoted in the frontmatter, inherits this same narrowing.

The original body above is retained as written (2026-06-26); this section is the update. See claim-linnainmaa-copenhagen-anecdote and myth-werbos-invented-backpropagation for the surrounding corrections.