---
title: "Linnainmaa's 1970 thesis was a contribution to numerical analysis, not artificial intelligence"
type: "claim"
status: "budding"
audit_status: "verified-verbatim"
source_url: "https://people.idsia.ch/~juergen/who-invented-backpropagation.html"
source_title: "Who Invented Backpropagation?"
source_author: "Jürgen Schmidhuber"
source_date: "2026-06-26T00:00:00.000Z"
source_quote: "albeit without reference to NNs"
source_tier: 2
provenance: "Promotion from 10-inbox/raw/2026-06-26-seppo-linnainmaa-..., 2026-06-26"
origin: "batch"
date_created: "2026-06-26T00:00:00.000Z"
tags: ["linnainmaa","numerical-analysis","automatic-differentiation","history-of-computation","backpropagation","floating-point"]
drafted_in: ["draft-linnainmaa-numerical-analysis-not-ai"]
---


[[entity-seppo-linnainmaa|Seppo Linnainmaa]]'s 1970 master's thesis — and its 1976 English publication — belongs entirely to [[numerical analysis]], the branch of mathematics concerned with the behavior of numerical computation. It contains no reference to neural networks or artificial intelligence.

The evidence is structural as well as testimonial. The English publication appeared in *BIT Numerical Mathematics* (vol. 16, no. 2, pp. 146–160, 1976; DOI: 10.1007/BF01931367), a journal founded in 1961 for computational and numerical mathematics. A paper's venue is not proof of its field, but BIT has no history of publishing AI or machine learning research — placing the work in numerical mathematics by publication norms as well as by content. [[entity-juergen-schmidhuber|Jürgen Schmidhuber]], who has surveyed the history of [[entity-backpropagation|backpropagation]]'s origins in detail, notes that Linnainmaa's thesis introduced the relevant mathematical structure "albeit without reference to NNs."

The problem Linnainmaa addressed — how rounding errors in floating-point arithmetic accumulate through a chain of operations — is a classical numerical analysis concern. It has no natural formulation within the AI problems of the era (pattern recognition, knowledge representation, theorem proving). The thesis is invisible in the AI literature of the 1970s and 1980s not because it was overlooked but because it was not addressed to that community.

## Why the field matters for the backpropagation gap

The disciplinary isolation of Linnainmaa's contribution is load-bearing for understanding [[backpropagation-gap]]. The algorithm that underlies all modern deep learning was not developed within any AI research program. It was derived in a numerical mathematics context, published in a numerical mathematics journal, and then independently re-derived — multiple times — by AI researchers who had no knowledge of Linnainmaa's work. The gap was not primarily a gap in citation; it was a gap in disciplinary contact.

See [[claim-linnainmaa-priority-not-paternity]] for the consequences of this isolation, and [[claim-linnainmaa-rounding-error-problem]] for what the thesis was actually solving.

## Sources

- Schmidhuber, Jürgen. "Who Invented Backpropagation?" https://people.idsia.ch/~juergen/who-invented-backpropagation.html (accessed 2026-06-26). Tier 2 (named expert, IDSIA director; secondary source for history but domain-authoritative). Exact quote: "albeit without reference to NNs."
- Linnainmaa, S. "Taylor expansion of the accumulated rounding error." *BIT Numerical Mathematics*, vol. 16, no. 2, pp. 146–160, 1976. DOI: 10.1007/BF01931367. Tier 1 (peer-reviewed primary). Venue: a numerical mathematics journal, not an AI venue; full paper paywalled at Springer Nature.
- Wikipedia. "Automatic differentiation." https://en.wikipedia.org/wiki/Automatic_differentiation (accessed 2026-06-26). Tier 3. Cites the BIT paper as the locus of Linnainmaa's contribution to the numerical mathematics tradition.

> [!note] Seek's commentary:
> "The gap was not primarily a gap in citation; it was a gap in disciplinary contact" is the deepest diagnosis in the whole cluster, and it reframes what "transmission" even means. Ideas don't cross fields through citations — citations flow *within* fields. They cross only through the rare person standing in two at once: Griewank phoning Linnainmaa from Argonne, Werbos reading Freud *and* control theory. Absent such a bridge, a published, correct, prior result stays invisible for forty years — not because anyone rejected it, but because no one in the receiving field was positioned to see it. The bridging individual, not the reference list, is the actual unit of transmission, and its scarcity is why the vault keeps finding forty-year gaps.
> — Seek