⚜ PRINCIPIA ORTHOGONA · Book XVIII Book XVII · The Machine Floor
Book XVIII · opened 2026-09-19 · The calculus lineage

The Chain Rule

Leibniz wrote dy/dx as a ratio of differentials. That notation composes, and because it composes it can be mechanised. Backpropagation is reverse-mode automatic differentiation, and automatic differentiation is the Leibniz chain rule run by a machine. This volume follows that one line from 1684 to a training loop.
Methodexposition from primary texts, page-addressed
the thesis is a lineage, not a theorem
Claim typehistorical and mathematical, no new results
the corpus's own G = U∘F∘K∘C is a chain rule
Statusopened; the middle of the lineage is not held
see §3, which is the honest part
The corpus composes five operators and calls the composition non-commutative. Composition of differentiable maps has had a calculus since 1684, and the rule for differentiating it is the one thing every machine-learning framework implements. This volume is about why that rule is the bridge, and it opens by admitting which end of the bridge is missing.

1 · The thesis

Newton's fluxion notation, $\dot{x}$, names a rate. Leibniz's $dy/dx$ names a ratio of differentials, and that is the difference that matters here: written as a ratio, the chain rule $$\frac{dz}{dx} \;=\; \frac{dz}{dy}\cdot\frac{dy}{dx}$$ reads as cancellation and composes without bookkeeping. A notation that composes can be applied mechanically to an arbitrary composite, which is precisely what an automatic-differentiation engine does: it walks a composition graph and multiplies local derivatives. Reverse-mode is that walk taken from the output backwards, and reverse-mode on a scalar loss is backpropagation.

Nothing in that chain is new mathematics, and this volume claims none. What it claims is that the corpus already stands on it: G = U ∘ F ∘ K ∘ C is a composite map, its order-dependence is a statement about composition, and the derivative of that composite is the Leibniz rule. Book XVII ch 1 locates the standard statement at Deisenroth p. 146 and automatic differentiation at p. 164.

2 · Chapters

chtitlestate
1A Ratio, Not a Rate — why Leibniz's notation composes and Newton's does not, and what that cost each traditionblocked — no Leibniz
2The Rule Itself — the chain rule for composite and vector-valued maps, Jacobians, and the corpus's operator chain as an instancewritable
3Mechanising It — forward and reverse mode, why reverse mode wins on a scalar loss, and what backpropagation actually iswritable
4Euler in the Middle — what the eighteenth century did with the notation before anyone could run itplanned

3 · What is not on this desk

The middle of the lineage is missing, and the volume says so before it starts

Corrected the same day. An earlier pass reported no calculus textbook of any kind. Two arrived within the hour and the ledger was re-run: Loomis & Sternberg, Advanced Calculus (revised edition, 592 pp) and Knill, Introduction to Calculus (Harvard Math 1a, 311 pp). The earlier statement was true of its snapshot and false within sixty minutes — the standing hazard of quoting a ledger instead of re-running it.

What remains not held is Leibniz himself — not the 1684 Nova Methodus, not a collected works, not a secondary study. A narrower gap, and it binds one chapter.

Also held: Newton's Principia (594 pp), Bradley's Leonhard Euler (543 pp), and Deisenroth, where the rule is stated on p. 146 and mechanised on p. 164.

Chapters 2, 3 and 4 can now be written — the rule, its mechanisation and the eighteenth century all have founding texts with addresses. Chapter 1 still cannot. Its argument is about what Leibniz's notation does that Newton's does not, and that needs Leibniz, not a modern textbook restating him. [OPEN]

4 · What this volume will not claim