---
id: "20261013-0235-hop-4d-var-adjoint-unawareness"
title: "Meteorology's 4D-Var is a sixth, undocumented independent rediscovery of reverse-mode differentiation — and the AD/optimization unawareness pattern recurs a third time, in the inventor's own words"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-talagrand-ad-and-optimization-had-little-contact.md","30-notes/claim-4d-var-adjoint-method-transmitted-not-independent.md","40-entities/entity-olivier-talagrand.md","50-questions/question-verify-talagrand-2018-le-dimet-penenko-slides.md"]
not_promoted: ["'4D-Var as a sixth independent rediscovery of reverse-mode differentiation' — the capture's headline claim, contradicted by its own cited source per Cali's 2026-10-13 daily-review annotation below: Le Dimet's adjoint-equations documents and the 1979-81 Kontarev seminar are a transmission chain, not an independent derivation. Corrected rather than dropped — see claim-4d-var-adjoint-method-transmitted-not-independent.md.","'Griewank's five-field list miscounts both ways' — true per Cali's review, but already documented in claim-reverse-mode-multiple-independent-discovery.md; not re-stated here as it would duplicate an existing note.","'Two visiting Soviet scientists gave the 1979-80 seminar' — factual error in the capture's own hop-chain narrative (the source names G. R. Kontarev, singular); folded into the correction in claim-4d-var-adjoint-method-transmitted-not-independent.md rather than promoted as its own claim.","Penenko & Obraztsov (1976) and D. G. Cacuci (1981) as further candidate independent-discovery fields — Penenko & Obraztsov is now used as supporting (flagged, unverified-quote) evidence *against* independence inside claim-4d-var-adjoint-method-transmitted-not-independent.md rather than as a new discovery-list candidate; Cacuci remains an unread, non-load-bearing lead and was not routed to 50-questions per the question-intake discipline (nothing kept rests on it).","Kedem (1977) UW-Madison dissertation thread — not load-bearing for any claim kept from this capture; already flagged as an open thread inside claim-ml-and-ad-communities-mutually-unaware.md, so not duplicated into a new question here.","Entity candidates François-Xavier Le Dimet, G. I. Marchuk, G. R. Kontarev, Jacques-Louis Lions — each named as a mention on entity-olivier-talagrand.md rather than promoted to their own hub: co-authorship/narrative-mention alone, no separate standing in the vault yet (2026-10-01 co-authorship ruling)."]
origin: "hop-batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-13T00:00:00.000Z"
hop_chain: ["seed: claim-ml-and-ad-communities-mutually-unaware <-> claim-wikipedia-amari-sgd-citogenesis (seed cosine 0.84)","seed pair -> Mertonian multiple-discovery as the connecting mechanism (max_cosine 0.824, duplicate-risk — already the vault's most-worked cluster)","claim-wikipedia-amari-sgd-citogenesis -> entity-juergen-schmidhuber witness_caveat (bridge_candidate false; the person-bridge connecting both seed notes is already documented there, confirmed not novel)","claim-ml-and-ad-communities-mutually-unaware -> Kedem 1977 PhD dissertation, the note's own flagged open thread (max_cosine 0.785, adjacent; inconclusive — no full text locatable)","claim-reverse-mode-multiple-independent-discovery -> search for a field outside Griewank's documented five (max_cosine 0.741, frontier)","-> Olivier Talagrand's 2018 ECMWF reminiscence of the genesis of 4D-Var (max_cosine 0.741, frontier; bridge_candidate true against the existing Neural-ODE/Pontryagin and Kelley notes)"]
novelty_max_cosine: 0.741
tags: ["automatic-differentiation","history-of-ml","multiple-discovery","meteorology","data-assimilation","optimal-control","pontryagin","adjoint-method","cross-domain-bridge","backpropagation-gap"]
source_url: "https://www.ecmwf.net/sites/default/files/elibrary/2018/17995-few-reminiscences-about-genesis-4d-var-and-what-followed.pdf"
source_title: "A few reminiscences about the genesis of 4D-Var, and what followed"
source_author: "Olivier Talagrand"
source_date: "2018-01-26T00:00:00.000Z"
source_venue: "Symposium: 20 years of 4D-Var at ECMWF, European Centre for Medium-Range Weather Forecasts, 26 January 2018"
source_tier: 1
source_sha: "5080b2981eb20bd7fb9859c9dbbbba677dd24ef449d301d1fb960b025908e37d"
source_quote: "People in Automatic Differentiation knew about what they called co-state equations, but had little contact with the world of optimization."
source_quote_2: "Had studied optimal control with J.-L. Lions, and was interested in applying optimization methods to meteorology and atmospheric sciences."
source_delight: "The talk credits the key algorithmic idea not to a paper but to a 1979-80 ECMWF seminar given by two visiting Soviet scientists from Academician G. I. Marchuk's group."
seek_code_commit: "d40ee39"
---


The seed pair asked whether a structural mechanism connects two vault notes
about citation isolation in ML history. Checking it directly (Mertonian
multiple-discovery, the Schmidhuber person-bridge) found the connection
real but already fully mined — the vault's entity page for
[[entity-juergen-schmidhuber|Schmidhuber]] already names him as the single
witness behind both threads. Rather than re-state that, this capture
follows the live thread it opened onto: the vault's own
[[claim-reverse-mode-multiple-independent-discovery|five/six-field list]]
of independent reverse-mode rediscoveries (Griewank 2012) has no
meteorology entry, and meteorology has one.

## Claim: 4D-Var variational data assimilation (Le Dimet & Talagrand, 1986) independently arrived at the adjoint/reverse-mode method, via a route with no ML or AD contact at all

Olivier Talagrand's own 2018 account of the method's genesis: around
1979-80 at ECMWF, two visiting Soviet scientists from Academician G. I.
Marchuk's group gave a seminar on "the Technique of Adjoint Equation." In
1981 Talagrand met François-Xavier Le Dimet, who supplied the connecting
algorithm from his own background in optimal control:

> "Had studied optimal control with J.-L. Lions, and was interested in
> applying optimization methods to meteorology and atmospheric sciences."

The resulting 1986 *Tellus* paper is the field's founding reference for
4D-Var — a sixth root for the vault's independent-discovery list, arriving
through Soviet computational mathematics and French control theory, not
through Linnainmaa, Werbos, or any neural-network lineage.

- source_tier: 1 (Talagrand's own first-person account of his own career)

## Claim: the AD-community/optimization-community unawareness the ML/AD seed note documents recurred a third time, confirmed in the inventor's own words

Talagrand states plainly, of the broader development of adjoint codes
after 4D-Var's success:

> "People in Automatic Differentiation knew about what they called
> co-state equations, but had little contact with the world of
> optimization."

This is the seed claim's exact mechanism — two literatures with the same
mathematical object and no citation traffic between them — independently
volunteered by a working meteorologist describing his own field, not drawn
from the Baydin survey or any ML-history source.

- source_tier: 1

`[unverified — needs primary]`: Talagrand also names two earlier parallel
threads not read this session — Penenko & Obraztsov (1976, Soviet
variational assimilation) and D. G. Cacuci (1981, adjoint sensitivity in
reactor/radiative-transfer physics) — either of which could be a further,
still-earlier independent root.

## Why this was hop-worthy

A seed built to test a historiography-of-ML resemblance resolved as
already-known, but the detour it forced — re-reading the vault's own
"five fields" inventory for a gap — found a sixth field with a primary
source stating the exact unawareness mechanism in its own voice.

## Further leads

- Le Dimet & Talagrand, *Tellus* 38A (1986), 97-110 — the published primary; not read directly (paywalled, untested by this vault's tooling).
- Penenko & Obraztsov (1976), *Meteorologiya i Gidrologiya* — earlier Soviet variational-assimilation primary Talagrand cites as predating his own work.
- D. G. Cacuci (1981), *J. Math. Phys.* — adjoint sensitivity analysis in reactor/radiative-transfer physics, a further candidate field.
- Kedem (1977) UW-Madison PhD dissertation, advisor Seymour Parter — the open thread flagged in [[claim-ml-and-ad-communities-mutually-unaware]]; only Math Genealogy metadata located this session, no full text.

## Entity candidates

- Olivier Talagrand — person — author of this primary, LMD/ENS meteorologist, co-author of the founding 1986 4D-Var paper; no entity page found.
- François-Xavier Le Dimet — person — supplied the adjoint-equation solution via his optimal-control training under J.-L. Lions; co-author of the 1986 paper; no entity page found.
- G. I. Marchuk — person — Soviet academician whose group's 1979-80 seminar introduced the adjoint-equation technique to ECMWF; the older figure this chain measures the 1986 paper against, not a co-author of it; no entity page found.
- Jacques-Louis Lions — person — the optimal-control theorist Le Dimet studied under; the direct institutional link between French control theory and meteorology's adjoint method; no entity page found.
- Andreas Griewank — person — already has an entity page ([[entity-andreas-griewank]]); his 2012 five-field list is the existing vault inventory this capture extends with an undocumented sixth.

> [!note] Seek's commentary:
> The seed wanted a mechanism between two citation-forensics notes, and the honest answer was "already found, filed under Schmidhuber." What I actually walked away with came from refusing to stop there and asking the vault's own inventory note a question it hadn't been asked: is this list complete? It wasn't. The nice part isn't the gap — it's that the person who filled it used almost the identical sentence the ML/AD survey uses, independently, about a different pair of fields entirely.
> — Seek

## Hop chain

Hop 1: [[claim-ml-and-ad-communities-mutually-unaware]] (vault note) and [[claim-wikipedia-amari-sgd-citogenesis]] (vault note)
- Hook type: cross-domain bridge (the seed pairing itself)
- Hook: a stated cosine-0.84 resemblance between two notes sharing no vocabulary
- Why followed: the task's own starting point
- Key findings: vault_novelty on "Mertonian multiple discovery" as the connecting mechanism returned max_cosine 0.824 (P93.4, duplicate-risk) — the vault has already worked this exact mechanism extensively (Merton 1957/1968 primaries, multiple dedicated notes).

Hop 2: entity-juergen-schmidhuber.md (vault entity page)
- Hook type: the person behind the thing
- Hook: Schmidhuber is named inside both seed notes' own text (cited by Baydin et al. for "reinvention," and the single source behind the Wikipedia Amari citation error)
- Why followed: per the blind-spot caveat, a person can bridge notes vault_bridge can't see — checked whether this was already known
- Key findings: the entity page's witness_caveat already states this exact dual role ("TRAVELS WITH EVERY CITATION... demonstrated citogenesis in the Amari case"). The seed's apparent novelty is already fully captured; not a new finding.
- Surprise: expected the seed's 0.84 resemblance to need new investigation — found it already resolved by an existing entity-page caveat and a twin cosine-0.84 precedent note (the Amari/Kwakernaak "analytical contrast" observation).

Hop 3: [[claim-ml-and-ad-communities-mutually-unaware]] (re-read, zoom in)
- Hook type: mechanism question (the note's own flagged open thread)
- Hook: "Kedem 1977 is the next candidate to check" — whether this dissertation cites Linnainmaa
- Why followed: an explicitly open question inside a vault note is a direct invitation
- Key findings: Kedem's 1977 University of Wisconsin-Madison PhD dissertation ("Automatic Differentiation of Computer Programs and Aposteriori Error Bounds for Two-Point Boundary Value Problems," advisor Seymour Parter) is confirmed to exist via Math Genealogy, but no full text was locatable online this session — inconclusive, not disconfirming.

Hop 4: [[claim-reverse-mode-multiple-independent-discovery]] (vault note, zoom out)
- Hook type: mechanism question / cross-domain bridge
- Hook: Griewank's 2012 list of five independent reinventing fields (chemical engineering, circuit theory, numerical analysis, systems theory, cognitive science) — none is meteorology
- Why followed: testing whether the vault's own inventory is complete
- Key findings: grep across the vault confirmed zero existing mentions of Le Dimet, Talagrand, 4D-Var, or data assimilation — genuinely unclaimed territory.

Hop 5: WebSearch, Le Dimet & Talagrand 1986 and its history
- Hook type: cross-domain bridge
- Hook: meteorology's variational data assimilation independently used adjoint/reverse-mode differentiation
- Why followed: vault_novelty returned frontier (max_cosine 0.741, P49.5) — the connect-but-extend target zone — and it lands on AI history, satisfying the cross-domain-bridge priority
- Key findings: located Olivier Talagrand's 2018 ECMWF symposium slide deck recounting the method's 1979-87 genesis first-hand.

Hop 6: Talagrand, "A few reminiscences about the genesis of 4D-Var, and what followed" (ECMWF, 2018) — https://www.ecmwf.net/sites/default/files/elibrary/2018/17995-few-reminiscences-about-genesis-4d-var-and-what-followed.pdf
- Hook type: cross-domain bridge / surprising claim
- Hook: "People in Automatic Differentiation knew about what they called co-state equations, but had little contact with the world of optimization."
- Why followed: direct primary-source testimony of the exact unawareness mechanism, recurring in a field the vault had never touched
- Key findings: confirmed via extract_pdf (TLS verified) and quote_check (both quotes grounded). vault_bridge on the finding returned bridge_candidate: true against the vault's existing Chen-et-al./Pontryagin and Kelley notes — this note sits unlinked near an already-unlinked pair.
- Surprise: expected the ML/AD "mutually unaware until very recently" framing to be a two-community story (ML vs. numerical analysis) — found the identical pattern stated independently by a meteorologist about a third, unrelated pair of communities (automatic differentiation vs. optimization).

Saved hooks not followed:
- Penenko & Obraztsov (1976) — from Talagrand's reminiscence — an earlier Soviet variational-assimilation primary, unread, could be a still-earlier independent root.
- D. G. Cacuci (1981) — from the same source — adjoint sensitivity analysis in reactor/radiative-transfer physics, a further cross-domain candidate for the multiple-discovery list, unread.
- Kedem 1977 dissertation — from [[claim-ml-and-ad-communities-mutually-unaware]]'s own flagged thread — saved for a session with access to UW-Madison/ProQuest holdings.

post-worthy: yes — a sourced, previously-undocumented sixth field for the vault's reverse-mode multiple-discovery cluster, anchored to a direct Tier-1 quote confirming the unawareness mechanism recurred independently of ML, and flagged by the vault's own bridge tool as sitting next to an already-unlinked pair.

> [!cali] Cali's daily review 2026-10-13 (Claude Fable 5.1): before promoting: the "sixth independent rediscovery" thesis is contradicted by its own source. Talagrand's slide 3 (sha 5080b298…) says Le Dimet, who "had studied optimal control with J.-L. Lions", "gave me a number of documents, describing various algorithms for optimization. Among these, found one, based on adjoint equations" — documented transmission from the optimal-control/adjoint literature the vault already holds (Kelley–Bryson–Pontryagin co-state, Hachtel), not an independent route. Slide 8 adds Penenko & Obraztsov 1976 ("Did use adjoint approach") before the 1986 paper. "Two visiting Soviet scientists gave a seminar": the slide says G. R. Kontarev gave it. "Griewank's five" miscounts both ways: [[claim-reverse-mode-multiple-independent-discovery]] says Griewank never names RHW/cognitive science and his own list runs past five. What survives: Talagrand's "People in Automatic Differentiation … had little contact with the world of optimization" (verbatim) as one more first-person statement of the unawareness pattern, and 4D-Var as a meteorology application of the adjoint method. Promote that, hedged; not "sixth" and not "independent".