Did Bryson & Ho's Applied Optimal Control (1969) serve as a curriculum transmission vector for adjoint/gradient methods, and what does the argmin blog cite it as evidence of?
Topic question (harvested from): 2026-06-30-what-role-did-kelley-1960-and-bryson-1961-optimal-control-work-play-in-the-backprop-lineage
Related notes: claim-kelley-bryson-optimal-control-precursor · 20260706-1505-did-kelley-1960-derive · adjoint-method · backpropagation · optimal-control · Lagrange-multiplier · argmin-blog
This capture has two parts, kept separate because they turned out to have different evidentiary bases: (1) what the argmin blog itself says about the book, and (2) whether the book independently functioned as a curriculum-transmission vector. The two are related but not the same claim, and it matters for provenance to keep them apart.
Part 1 — What the argmin blog cites the book as evidence of
Claim type: Historical / technical-mechanism claim (what a specific named source argues) → floor Tier 1–2 required. Source: Ben Recht's own blog post, his own venue. Tier 1–2. ✓ clears floor.
The relevant post is Ben Recht's "Mates of Costate," published on the arg min blog on 2016-05-18 (archived at https://archives.argmin.net/2016/05/18/mates-of-costate/ — this is the older WordPress-era archive of the blog that now continues at argmin.net / Substack). This is the only post located, across a targeted search of Recht's blog, that names Bryson & Ho's Applied Optimal Control specifically.
Exact quoted passage:
"As I mentioned already, the method of adjoints originates in the study of controls. According to Dreyfus, this was first proposed by Bryson in a paper called 'A Gradient Method for Optimizing Multi-Stage Allocation Processes' that appeared in the Proceedings of the Harvard University Symposium on Digital Computers and Their Applications in 1961. I was unable to find this proceedings in our Engineering Library, but the Lagrangian derivation plays a prominent role in Bryson and Ho's 1968 book Applied Optimal Control. Note that Bryson's paper appeared only a year after Kalman's absurdly influential 'A New Approach to Linear Filtering and Prediction Problems'. This use of duality was very much at the birth of modern control theory."
What this passage actually cites the book as evidence of: Recht uses Applied Optimal Control as a substitute textual record for the argument he wants to make, because he could not get hold of Bryson's original 1961 conference paper. The book stands in as documentation that "the Lagrangian derivation" of the adjoint/gradient method "plays a prominent role" in the literature — i.e., it is cited as evidence that the mathematical technique (duality between a system and its adjoint, used to compute gradients) was worked out and written down in a citable, accessible engineering-control-theory text, and as part of a lineage that Recht traces back to "the birth of modern control theory" alongside Kalman's 1960 filtering paper.
Note the discrepancy: Recht dates the book "1968" in this passage; the book's actual publication (Blaisdell Publishing, Waltham, MA) is standardly dated 1969, matching the topic-question's date and Dreyfus's 1990 citation (below). This is Recht's own wording as fetched; the date discrepancy is preserved here rather than silently corrected, since it belongs to the primary quote. [minor discrepancy — not resolved, flagged]
What the passage does NOT say: Recht does not use the words "curriculum," "taught," "engineering education," or any equivalent framing in this post. His claim, in his own words, is about mathematical/derivational lineage — the technique's presence in a citable secondary/textbook source, and its place in a chronology bookended by Kalman (1960) and Bryson's own 1961 paper — not an explicit argument that the book served as a vector carrying the method into engineering classrooms. Two other Recht/argmin posts that mention Bryson's adjoint method in passing ("Fact, Fiction, and Forecast," https://www.argmin.net/p/fact-fiction-and-forecast, and "Reticulating Splines," https://www.argmin.net/p/reticulating-splines) do not mention Ho or the 1969 book at all — they cite only "Bryson's method of adjoints... (1962)" as equivalent to backpropagation's gradient computation, without textbook or curriculum framing.
Verdict on this part of the core question: The specific "curriculum transmission vector" framing named in the topic question does not appear to originate from the argmin blog itself. [unverified — the argmin blog's actual citation of the book is narrower (mathematical/derivational lineage), not a pedagogical-transmission argument; no post found making the curriculum-vector claim explicitly, after a targeted search of the most likely candidate posts]. If a Recht post making that specific argument exists elsewhere (a later Substack piece, a talk transcript, etc.), it was not located in this research pass.
Part 2 — Did the book independently function as a curriculum transmission vector?
Claim type: Historical claim about the book's origin and use in teaching. Uncontested, and the strongest possible source is available: the book's own co-author's first-person retrospective. Tier 1. ✓ clears floor (and exceeds it — this is a case where a surprising/load-bearing claim gets escalated to Tier 1, which is met).
Independent of what argmin says, Arthur Bryson's own 1980 "This Week's Citation Classic" commentary (Current Contents, Issue #8, 1980-02-25, https://garfield.library.upenn.edu/classics1980/A1980JE96000001.pdf) gives a direct, first-person account of the book's teaching origins:
"Having had some success in solving realistic problems, Yu-Chi Ho and I offered a two-week course on dynamic optimization at Harvard in the summer of 1963. This was attended by 100 engineers, most of whom were from industry. In 1967, we offered an intensive two-day continuing-education course sponsored by AIAA which featured simultaneous use of two viewgraphs. This course was repeated several times and is still available from AIAA in the form of audio-cassettes and a slide-book. We began offering a graduate course at Harvard in 1963, which was also given at MIT in 1966. As a result of the fine reception given to these courses, we were encouraged to extend our class notes and put them into book form."
And, on why the book was so widely cited:
"Speculating on why the book has been frequently cited, we believe it may be because it is oriented toward solving realistic problems and contains examples and exercises that are closely related to real engineering problems."
The same source states the book had been "cited over 520 times since 1969" as of the 1980 writing [unverified-quant — needs primary: this figure comes from Science Citation Index tallies referenced in Bryson's own retrospective; the retrospective itself is Tier 1, but the underlying SCI count was not independently re-verified against the index in this pass].
What this establishes: The book was not merely a reference text written in isolation — by its own co-author's account, it was compiled directly out of class notes from a 1963 two-week Harvard short course (100 engineers, mostly from industry), a Harvard graduate course running from 1963, a parallel MIT course from 1966, and a repeated 1967 AIAA continuing-education course distributed afterward as audio-cassettes and a slide-book. This is direct, primary-source evidence of a teaching-to-textbook pipeline that predates and feeds the 1969 publication — i.e., evidence that the book functioned as (or emerged from and then extended) a curriculum-transmission vector carrying adjoint/gradient optimal-control methods into engineering audiences, both academic (Harvard, MIT graduate courses) and industrial (the AIAA continuing-education course, the industry-heavy 1963 short course).
Corroborating technical-mechanism claim (Tier 1–2): Stuart Dreyfus, in his own 1990 peer-reviewed historical/technical account ("Artificial Neural Networks, Back Propagation, and the Kelley-Bryson Gradient Procedure," J. Guidance, Control, and Dynamics, 13(5):926-928, https://gwern.net/doc/ai/nn/1990-dreyfus.pdf), confirms the book's specific technical content:
"Bryson did, however, present a gradient solution of a multistage problem in Ref. 5 and Bryson and Ho explicitly gave in their 1969 book gradient formulas for exactly the multistage, free-terminal-state problem we are considering. Kelley used adjoint equations and Green's theorem in his derivation, and Bryson used Lagrange multipliers."
Dreyfus dates the book 1969, consistent with the topic question and standard bibliographic record, and independently confirms that the adjoint/gradient formulas for the relevant class of optimal-control problem are worked out explicitly in the book (not merely alluded to) — corroborating, from a different primary actor, the same substance Recht points to in "Mates of Costate."
Caution on a search-summary error: One search result encountered during this research misattributed the sentence "Bryson and Ho explicitly gave in their 1969 book gradient formulas for exactly the multistage, free-terminal-state problem" to "the argmin.net blog." Direct fetch of the actual Dreyfus 1990 PDF confirms this sentence is Dreyfus's own wording in his 1990 paper, not Recht's. This is noted here explicitly so the misattribution is not repeated at promotion.
Weaker corroboration (Tier 4, definitional/uncontested-historical use only): Wikipedia's "Backpropagation" article, History section, states: "Hecht-Nielsen credits the Robbins–Monro algorithm (1951) and Arthur Bryson and Yu-Chi Ho's Applied Optimal Control (1969) as presages of backpropagation." This attributes the "presage" framing to Robert Hecht-Nielsen's 1990 book Neurocomputing — a claim about who first drew this connection in the AI/ML literature, distinct from Recht's or Dreyfus's claims. Used here only to note that the book's role as a backprop-precursor was already being flagged in the neural-network literature by 1990, not as evidence for the curriculum-vector claim itself, which rests on the Tier 1 Bryson retrospective above.
Jürgen Schmidhuber's "Who Invented Backpropagation?" page (https://people.idsia.ch/~juergen/who-invented-backpropagation.html) lists "Bryson and Ho (1969)" in a chronological citation chain (Kelley 1960; Bryson 1961; Pontryagin et al. 1961; Dreyfus 1962; ...; Bryson and Ho 1969) with no elaboration specific to curriculum or textbook role — included here only to confirm the book's standard place in the priority literature, not as an independent source for either half of the core question.
Overall verdict on the core question
Two-part answer, evidenced at different tiers:
-
"What does the argmin blog cite the book as evidence of?" — Confirmed, narrowly: Ben Recht's 2016 "Mates of Costate" post cites Applied Optimal Control as the accessible written record of the Lagrangian/adjoint gradient derivation (standing in for an unobtainable 1961 conference paper), situating it in a lineage running from Kalman (1960) through Bryson (1961) to the book (which Recht here dates 1968) — a claim about mathematical/derivational lineage, not an explicit curriculum-transmission argument.
[the specific "curriculum transmission vector" framing in the topic question does not appear in the argmin text found; flagged as such rather than forced] -
"Did the book serve as a curriculum transmission vector into engineering education?" — Well supported, but by a source outside argmin: Arthur Bryson's own 1980 first-person retrospective describes a direct pipeline from a 1963 Harvard industry short course and a 1967 AIAA continuing-education course, through graduate courses at Harvard (1963) and MIT (1966), into the 1969 book. This is Tier 1 primary evidence for the curriculum-vector claim, but it is Bryson's own account, not something drawn from or attributed to the argmin blog.
The topic question links these two things as though they were one and the same evidentiary chain; the research here suggests they are two separate, independently well-sourced claims that happen to share a common object (the book), and that conflating them at promotion would overstate what argmin itself argues.