A curve and its mirror image
Take for . Its inverse undoes it: if then , so . The graph of the inverse is the graph of f reflected in the line y = x, because reflecting in that line swaps the two coordinates of every point. The point (1.5, 1.125) is on , and its mirror image (1.125, 1.5) is on .
The two curves cross on the mirror line itself, at (0, 0) and (2, 2), where swapping x and y changes nothing.
Rise and run swap
At x = 1.5 the gradient of is f'(1.5) = 1.5. On its tangent, a run of 0.5 gives a rise of 0.75.
Reflect that little triangle in y = x. What went across now goes up, and what went up now goes across: a run of 0.75 and a rise of 0.5. So the tangent to the inverse at (1.125, 1.5) has gradient , which is 1 ÷ 1.5.
Check it on the inverse directly. From x = 1.125 to x = 1.126, rises from 1.5 to , and the chord gradient with h = 0.001 is 0.666519, just under .
The curve , its inverse , and the mirror line y = x between them. At (1.5, 1.125) the tangent to runs 0.5 and rises 0.75, a gradient of 1.5. At the mirror point (1.125, 1.5) the tangent to the inverse runs 0.75 and rises 0.5, a gradient of .
The rule
Name the matching points. If f(a) = b, then (a, b) is on the graph of f and (b, a) is on the graph of . The tangents at those two points are mirror images, so their gradients are reciprocals: .
The derivative of f is read at a, the input, and the inverse's gradient belongs to b, the output. Reading f' at b gives the gradient of f somewhere else on the curve. For , f'(1.125) is 1.125, and 1 ÷ 1.125, about 0.889, is the gradient of the inverse at , not at x = 1.125.
The rule needs . Where the tangent to f is level, its mirror image is vertical, and a vertical line has no gradient. The rule also needs f to have an inverse at all near a, which a curve that keeps rising (or keeps falling) there does.
Why: the chain rule
Doing and then f returns the input: for every x the inverse accepts. Differentiate both sides. The right-hand side gives 1. The left-hand side is f applied to the inside function , so the chain rule gives .
So , and dividing gives . At x = b, , and this is the rule again: . The two rates multiply to 1, the way a rise over a run and a run over a rise do.
From two numbers
The rule needs only two facts about f. Given f(3) = 7 and f'(3) = 5, the input that gives 7 is 3, so . There is no need to find a formula for the inverse.
One function with those two facts is : f(3) = 9 − 3 + 1 = 7, and f'(x) = 2x − 1 gives f'(3) = 5. For its inverse can be written out: solving for x gives , and . Its derivative is , which at x = 7 is . The chord with h = 0.001 gives 0.199992.
Cube roots and logarithms
Take , with f(2) = 8 and . Its inverse is the cube root, so the cube root has gradient at 8. Directly, differentiates to , and at 8 that is . The chord from 8 with h = 0.001 gives 0.08333.
The same step gives the derivative of ln x from that of . is its own derivative, so at x = 2 its gradient is . Its inverse, ln x, has gradient at the matching point , about 7.389, and is there. The same holds at every point, which is why ln x differentiates to . A centered chord at gives 0.135335, the value of to six places.
At 0 the cube has a level tangent: f'(0) = 0. The rule then has nothing to divide by, and the cube root stands vertical at the origin. Its chord from 0 to 0.001 already has gradient 100, and from 0 to 0.000001 it has gradient 10000.
the inverse is the reflection in y = x, so its tangent is the reflected tangent and the slopes are reciprocals: f′(a) × (f⁻¹)′(f(a)) = 1
Drag the point and try to make the product of the two slopes anything but 1
The curve and its inverse , mirror images in y = x. At a = 1.4 the point on the cube is (1.4, 2.744), with gradient ; at the mirror point (2.744, 1.4) the cube root has gradient 1 ÷ 5.88, about 0.170. Drag the point along the cube: the two gradients always multiply to 1.
The usual mistakes
Keeping f' as it is. With f(2) = 9 and f'(2) = 4, the inverse has gradient at 9, not 4.
Turning over the point instead of the gradient. With f(2) = 9 and f'(2) = 4, the answer is not : the reciprocal is taken of f', not of the place it is read.
Reading f' at the output. With f(4) = 2 and f'(4) = 6, is . The derivative of f at 2 belongs to a different point of the curve.
Treating as . The inverse undoes f; it does not divide by it. The rule comes from , a composition, and not from a product.