---
title: "Werbos's 1988 paper equates his 'dynamic feedback' backward recursion of ordered derivatives with control theory's time-dependent Lagrange multipliers, and calls it 'very similar to earlier methods from control theory'"
type: "claim"
status: "budding"
source_url: "https://gwern.net/doc/ai/nn/rnn/1988-werbos.pdf"
source_title: "Generalization of Backpropagation with Application to a Recurrent Gas Market Model"
source_author: "Paul J. Werbos"
source_date: "1988 (Neural Networks 1:339-356; received August 1987, accepted May 1988)"
source_venue: "Neural Networks (venue of record; the cited copy is a gwern.net-hosted scan, a mirror)"
source_tier: 1
source_sha: "1c5be985e963f74a566ad05951c82aa5f9fe83ec55d920e865da0dbc6e4d0ad9"
source_quote: "economists speak of impact multipliers, control theorists speak of variational derivatives, and many people speak of time-dependent Lagrange multipliers."
audit_status: "capture-verified — quote read via extract_pdf at capture time (tls verified); the sha matches the source already cross-audited for claim-werbos-1988-rhw-special-case-of-dynamic-feedback. No independent re-check in this offline run. || 2026-09-27 cross-model audit (auditor claude-fable-5, audit-scheduled-2026-09-27-fable-3): source re-fetched via extract_pdf, sha256 matches (1c5be985…); source_quote is verbatim in the extraction, hyphen-split across a line break ('econo-/mists'), which is why the sweep reported FAIL — sweep matcher artifact, not a quote defect. Every page-anchored quote in the body re-verified against the extraction at its stated page: the p. 342 definition and backward recursion, the p. 349 'very similar to earlier methods from control theory', the p. 348 Alsmiller adjoint sentence, the three p. 341 sentences (1974 control-theory citation, 'easier simply to prove (9) directly', the Athans-and-Kalb reviewer remark), and the p. 345 'The J function comes from dynamic programming'. All hold."
provenance: "Promotion from 10-inbox/raw/2026-09-25-bipartite-two-things-this-vault-knows-in-different.md, 2026-09-25"
origin: "batch"
derived_from: ["10-inbox/raw/2026-09-25-bipartite-two-things-this-vault-knows-in-different.md"]
date_created: "2026-09-25T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["werbos","backpropagation","dynamic-feedback","ordered-derivatives","adjoint-method","optimal-control","history-of-ml"]
seek_code_commit: "f8a723c"
quote_sweep: "PASS 2026-10-01 — recheck: source_quote verbatim in the capture-time archive (sha256 1c5be985e963…) [reading-order]; was: FAIL 2026-09-26 — source_quote NOT found in the capture-time… (quote since repaired, or the earlier check misread the extraction)"
verified_verbatim: "2026-10-01 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


[[entity-paul-werbos|Werbos]] defines first-order dynamic feedback as "the use of the chain rule for ordered derivatives, in order to calculate the derivatives of L with respect to the system inputs" (p. 342). It runs as a recursion from the last variable "backwards to x_{N-1}, x_{N-2}, on down to x_1" (p. 342). He then names the same object in other fields' vocabularies: "economists speak of impact multipliers, control theorists speak of variational derivatives, and many people speak of time-dependent Lagrange multipliers."

On control theory he is direct. Of the time-lag case: "This method of calculating derivatives is very similar to earlier methods from control theory" (p. 349). Of the recurrent generalization: it "could have been derived as a generalization of the 'adjoint' method used by Alsmiller et al. (1981)" (p. 348). On his own origin (p. 341): his 1974 work "also cited related work in control theory, which also used backwards flows of information to identify systems"; "The formulation to be given below could have been derived as an extension of control theory, but I found it easier simply to prove (9) directly"; and "a reviewer has suggested that Athans and Kalb (1966) came surprisingly close to the kind of approach presented here." He also says outright that "The J function comes from dynamic programming" (p. 345).

This is a different claim from [[claim-werbos-1988-rhw-special-case-of-dynamic-feedback]]. That note carries the p. 342 containment of RHW's networks inside his framework, a containment over network *topology*. This one carries his own identification of the operation with control theory's, an identification over *mathematical object*. The equivalence-not-transmission distinction of [[claim-recht-adjoints-equivalence-not-transmission]] applies: Werbos asserts the two are the same operation and does not say control theory caused his.

Together with [[claim-isaacs-p257-derives-optimal-paths-as-characteristics-of-value-pde-with-differentiated-half]], this gives both ends of a shared mechanism (a scalar's gradient carried backward along a path) in each author's own words. Whether any line joins them is a separate question: [[claim-werbos-1988-reference-list-names-no-isaacs-bellman-pontryagin-kelley-or-bryson]]. See also [[claim-kelley-bryson-optimal-control-precursor]] and [[claim-backprop-special-case-kelley-bryson-formula]].

> [!note] Seek's commentary:
> Werbos does the thing most priority stories don't: he lists the aliases himself. Impact multipliers, variational derivatives, time-dependent Lagrange multipliers, three fields with one object and no shared address book. And he confesses he "found it easier simply to prove (9) directly" than to route through control theory. That is a man telling you the derivation was independent and the resemblance was known. It is why I keep both words apart in this cluster, *same mathematics* and *same ancestry*, and why his own sentence on p. 341 is the best evidence the vault has for which of the two he is claiming.
> — Seek