Interactive guide

One nudge, resized twice.

Inner: u = g(x)

Outer: y = f(u)

 curve    tangent    nudged point (true)

0.80
0.60

Jump to a composition

dy/dx = f′(u) × g′(x) = = — =

True Δy
Predicted Δy

Shrink Δx all the way down with the second slider and watch "True Δy" and "Predicted Δy" converge. The chain rule's product, f′(u)×g′(x), is only ever exact in that limit, at any real (nonzero) step size it's an approximation, a very good one close in, a worse one far out.

A composition is a pipeline, not a blend

f(g(x)) means: first run x through g, get a number u, then run that number through f. Nothing about f ever sees x directly, it only ever sees u. So a nudge to x has to travel through g first, arriving at f already resized by however much g stretches or shrinks things right there.

The chain rule is just bookkeeping for that two-stage resizing: g′(x) is how much g resizes a nudge at x; f′(u) is how much f resizes a nudge at u. Do both resizings in sequence, multiply, done.

Reading the two panels

  1. 1

    Left panel: Δx becomes Δu

    The tangent line's slope, g′(x), is the resizing factor for this step.

  2. 2

    Right panel: that same Δu becomes Δy

    Same idea, different slope: f′(u). The input here is the output of the left panel, not x.

  3. 3

    Two resizings in a row multiply

    Resize by 3, then by 2, and the combined effect is resizing by 6, not 5. Rates compose by multiplying, not adding.

  4. 4

    Curved functions only agree with this locally

    The tangent-based prediction and the true curve-based result match closely for small Δx, and drift apart for large ones.

Both tangent lines here are the exact idea from the Derivatives & Integration page, just used twice in a row instead of once. And the shrinking Δx slider is the exact idea from the Taylor Series page's higher terms: a smaller step means less of the curve's bending gets to matter, which is the entire reason a first-order, straight-line approximation can work at all.

The jargons

Two functions, one after the other, and the arithmetic for how their rates combine.

Composition f(g(x))

Applying g first, then f to whatever g produced. Also written (f∘g)(x).

Inner function g

The one applied first, directly to x. Its output, u, is the only thing the outer function ever sees.

Outer function f

The one applied last, to g's output rather than to x directly.

Chain rule

d/dx[f(g(x))] = f′(g(x)) × g′(x): multiply the two local resizing factors together.

Nudge Δx

A small step away from the current x, used to see how much the output moves in response.

Linear approximation

Predicting a nudge's effect using only the tangent line's slope, exact in the limit, approximate everywhere else.

A teaching tool: "true Δy" is computed from the actual curves, f(g(x+Δx)) − f(g(x)); "predicted Δy" comes only from the chain rule's product of slopes, so the gap between them is the approximation error made visible.