capture rejected Tier 1 2026-07-07

Capture: Speelpenning's 1980 thesis (UIUCDCS-R-80-1002) — free full text on OSTI, and does its bibliography (leaf 74) cite Linnainmaa 1970 or 1976?

Claim: A free full-text PDF of Speelpenning's thesis is openly hosted on OSTI.gov (DOE's own repository)

Claim type: historical/bibliographic (existence and location of a document). Uncontested — Tier 3–4 floor applies, but source achieved is Tier 1 (the primary document itself, on the sponsoring agency's own repository).

The thesis — "Compiling Fast Partial Derivatives of Functions Given by Algorithms," Bert Speelpenning's PhD dissertation, Department of Computer Science, University of Illinois at Urbana-Champaign, January 1980 — is openly available as a full-text PDF via OSTI.gov (the U.S. Department of Energy's Office of Scientific and Technical Information), which is the funding agency's own repository. The record page is at https://www.osti.gov/biblio/5254402 and the direct PDF is at https://www.osti.gov/servlets/purl/5254402/. Both URLs were verified to resolve in this session; the PDF was successfully fetched and OCR'd (82 pages, method "ocr").

Bibliographic identifiers from the OSTI record and confirmed against the thesis's own title page and NTIS bibliographic-data sheet (the standard back-matter page every DOE report of this era carries): report numbers COO-2383-0063; UILU-ENG-80-1702; UIUCDCS-R-80-1002; OSTI ID 5254402; DOI 10.2172/5254402; sponsoring organization Department of Computer Science, University of Illinois at Urbana-Champaign, under U.S. DOE/ERDA contract (the thesis's own text gives it as "US ENERGY/EY-76-S-02-2383").

A second, independent copy (scan, not the OSTI OCR) is hosted on the Internet Archive at https://archive.org/details/compilingfastpar1002spee, also verified to resolve, listing the same title, author "Speelpenning, B", 1980, and "Report (University of Illinois at Urbana-Champaign. Dept. of Computer Science) no. 1002" — corroborating the OSTI identification is the correct, complete document (104 pages per Internet Archive's count, vs. 82 pages in OSTI's scanned/OCR'd PDF — the discrepancy is plausibly front-matter/blank-page handling between the two scans; it was not resolved further this session).

Provenance:

source_url: https://www.osti.gov/servlets/purl/5254402/ (PDF); metadata cross-checked at https://www.osti.gov/biblio/5254402 and https://archive.org/details/compilingfastpar1002spee
source_author: Speelpenning, Bert
source_date: 1980-01 (January 1980)
source_tier: 1
exact quote (OSTI record metadata, via WebFetch of the biblio page): "Title: 'Compiling fast partial derivatives of functions given by algorithms' ... Report Numbers: COO-2383-0063; UILU-ENG-80-1702; UIUCDCS-R-80-1002 ... DOI: https://doi.org/10.2172/5254402"
exact quote (thesis title page, OCR'd text): "COMPILING FAST PARTIAL DERIVATIVES / OF FUNCTIONS GIVEN BY ALGORITHMS / by / Bert Speelpenning / January 1980 / DEPARTMENT OF COMPUTER SCIENCE / UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN"

Claim: The thesis's References section (printed page 74, immediately followed by the reference list running onto page 75) lists exactly eleven citations, and neither Linnainmaa 1970 nor Linnainmaa 1976 is among them

Claim type: technical-mechanism / historical (what a specific primary document's own bibliography contains). Tier 1–2 required — achieved Tier 1, direct OCR read of the primary thesis text itself, not a secondhand description.

The thesis's table of contents lists "REFERENCES" beginning on page 74 ("REFERENCES...74" in the OCR'd table of contents). Reading the thesis's actual page 74 confirms this: the page footer "74" appears immediately before the "REFERENCES" heading, followed by eleven bracketed citation-key entries and then the corresponding full citations, running onto page 75 (footer "75" appears immediately after the last entry, followed by "VITA," the thesis's closing biographical section). The complete reference list, transcribed exactly as OCR'd:

"[BEL57] [COM78] [DIJ76] [JOH75] [JOS76] [KAM76] [KED77] [KUC78] [STA76] [STR69] [WAR75]

74

REFERENCES

Bellman, Richard E., "Dynamic Programming", Princeton Univ. Press, 1957.

Comer, D., "MOUSE4: An improved implementation of the RATFOR preprocessor", Software-Practice and Experience, Vol. 8, 1978.

Dijkstra, Edsger W., "A Discipline of Programming", Prentice-Hall, 1976.

Johnson, S.C., "YACC - Yet Another Compiler-Compiler", C.S. Tech. Report 32, Bell Laboratories, July 1975.

Joss, Johan, "Algorithmisches Differenzieren", Ph.D. Thesis, ETH, Zurich, Switzerland, 1976.

Kam, J.B., Ullman, J., "Global Data Flow Analysis and Iterative Algorithms", JACM, Vol. 23, No. 1, Jan. 1976.

Kedem, Gershon, "Automatic Differentiation of Computer Programs", Proc. 1977 Army Numerical Analysis and Computer Conf., Madison, Wisc. 1977.

Kuck, David J., "The Structure of Computers and Computations", Vol. 1, John Wiley & Sons, 1978.

Standish, T., Harriman, D., Kibler, D. and Neighbors, J., "The Irvine Program Transformation Catalogue", C.S. Dept., U.C. Irvine, Irvine, Cal., Jan. 1976.

Strassen, Volker, "Gaussian Elimination is not optimal", Numer. Math. 13, 1969.

Warner, D.D., "A Partial Derivative Generator", C.S. Tech. Report 28, Bell Laboratories, April 1975.

75"

No entry cites Linnainmaa (neither the 1970 Helsinki master's-thesis version nor the 1976 BIT paper "Taylor expansion of the accumulated rounding error"). The eleven cited works are: Bellman 1957, Comer 1978, Dijkstra 1976, Johnson 1975, Joss 1976, Kam & Ullman 1976, Kedem 1977, Kuck 1978, Standish et al. 1976, Strassen 1969, and Warner 1975.

This is corroborated by the body of the thesis itself: Chapter 2, "Symbolic Differentiation of Algorithms: Previous Work," is Speelpenning's own literature review of precursor work, and it discusses exactly three prior efforts by name — Warner 1975, Joss 1976, and Kedem 1977 — with Joss's 1976 ETH Zurich PhD thesis ("Algorithmisches Differenzieren," written in German) singled out as "an important breakthrough" and the primary point of departure for Speelpenning's own method. The chapter explicitly frames the history as: Warner (1975, straight-line code only) → Joss (1976, arbitrary control flow, first general feasibility proof) → Kedem (1977, independent rediscovery in English, unaware of Joss). Linnainmaa's name does not appear anywhere in this historical account either.

A full read of the entire OCR'd document (all 82 pages / 3,757 lines of extracted text, front matter through the closing NTIS bibliographic-data sheet) turned up zero occurrences of "Linnainmaa" in any spelling or OCR-garbled variant searched for. This is a stronger check than leaf 74 alone: the name is absent from the whole thesis, not merely from the formal References list.

Provenance:

source_url: https://www.osti.gov/servlets/purl/5254402/ (PDF, extracted via extract_pdf; OCR text read directly, not summarized by an intermediary tool)
source_author: Speelpenning, Bert
source_date: 1980-01
source_tier: 1
exact quote: see full blockquote above (thesis pages 74–75, References section) and "2.2. Joss, 1976 / An important breakthrough was the doctor's thesis of Johann Joss at ETH in Switzerland in 1976 [JOS76]."

Seek's commentary: A caveat worth keeping attached to this finding — this is OCR'd text, and the code-listing sections of the thesis (Chapter 6, worked FORTRAN examples) are badly garbled by the OCR pass in a way the prose and the References section are not. The References section itself reads cleanly and internally-consistently (eleven bracket-keyed citations in the pre-list matching eleven full entries exactly, in the same order), which is good evidence the OCR did not silently drop or merge an entry there. But "the OCR shows no Linnainmaa citation" is not quite the same epistemic weight as "a human who has held the physical microfiche/paper copy confirms no Linnainmaa citation" — for a claim this load-bearing to the backpropagation-priority literature, a second, independent look at the Internet Archive scan (a different digitization, not OSTI's) would be worth doing before treating this as fully closed. That said, both the isolated References section and the independent corroboration from Chapter 2's own historical narrative point the same direction, which raises confidence considerably.

Claim: The thesis's own bibliographic-data-sheet abstract states Jake (Speelpenning's system) is faster than numerical differencing for n > 8

Claim type: quantitative (a specific "n >" threshold). Tier 1–2 required — achieved Tier 1, direct quote from the primary document's own abstract page.

"Measurements performed on one particular machine suggest that Jake is faster than numerical differencing for n > 8. Somewhat weaker results have been obtained for the problem of computing Jacobians of arbitrary shape."

This is included here as a secondary, incidental finding surfaced while reading the full document for the Linnainmaa question — flagged for any future note on Speelpenning's actual performance results, not the primary target of this capture.

Provenance:

source_url: https://www.osti.gov/servlets/purl/5254402/
source_author: Speelpenning, Bert
source_date: 1980-01
source_tier: 1
exact quote: "Measurements performed on one particular machine suggest that Jake is faster than numerical differencing for n > 8."

Central question: RESOLVED — leaf 74 (and the thesis as a whole) cites neither Linnainmaa 1970 nor Linnainmaa 1976

Two things were asked: (1) whether a free full-text PDF of the thesis is reachable via OSTI/DOE/Argonne open repositories, and (2) whether the bibliography cites Linnainmaa. Both are answered directly from primary-source material read in this session:

Yes, a free full-text PDF is on OSTI.gov itself (https://www.osti.gov/servlets/purl/5254402/), independently corroborated by an Internet Archive scan. No Argonne-specific repository was found or needed — OSTI.gov's own record sufficed.
No. Leaf/page 74 (the References heading and the start of the reference list) — and the thesis's entire bibliography, and the entire body text — cite neither Linnainmaa 1970 nor Linnainmaa 1976. The precursor Speelpenning credits by name throughout is Johann Joss's 1976 ETH Zurich PhD thesis, alongside Warner 1975 and Kedem 1977.

This bears on the broader backpropagation/automatic-differentiation priority question (the subject of several other captures in this vault, e.g. discussions of Linnainmaa's 1970/1976 priority for reverse-mode AD): Speelpenning's independently-arrived-at 1980 method, one of the more frequently cited early "fast gradient" results in the AD literature, appears to have been developed without reference to Linnainmaa's earlier work — his own stated intellectual lineage runs through Warner → Joss → Kedem instead. Whether this reflects a genuine independent (re)discovery, limited circulation of Linnainmaa's Finnish-language 1970 thesis and Finnish/Scandinavian-venue 1976 paper outside Nordic/Soviet numerical-analysis circles, or simply a gap in Speelpenning's own literature search, is a separate question this capture does not resolve — it only establishes the citation fact itself.

Further leads

Cross-check this "no Linnainmaa citation" finding against the Internet Archive scan (https://archive.org/details/compilingfastpar1002spee) directly, ideally with a second independent OCR pass, given the caveat above about OCR reliability on a single-source scan.
This finding is a useful data point for any synthesis note in this vault about Linnainmaa's citation history / obscurity before ~2010s revival (see existing captures tagged linnainmaa and griewank) — Speelpenning 1980 not citing Linnainmaa is consistent with, though not proof of, the broader claim that Linnainmaa's priority went largely unrecognized in the Western AD/optimization literature for decades.
Joss's 1976 ETH Zurich thesis ("Algorithmisches Differenzieren," in German) is repeatedly named by Speelpenning as the key precedent and is itself a candidate for a dedicated capture — it predates Speelpenning by four years and is described in the thesis as proving a general feasibility result for differentiating algorithms with arbitrary control flow, independent of Linnainmaa's earlier and differently-framed (computational-graph / rounding-error) treatment.

Source

Tier 1 Speelpenning, Bert 1980-01

https://www.osti.gov/servlets/purl/5254402/

· batch run 2026-07-07 — researched via web search, OSTI.gov, and direct PDF extraction of the primary thesis text · raw markdown