You drive 60 miles and it takes you exactly one hour. Your average speed was 60 mph — but you certainly didn't hold 60 the whole way. You crawled through a town, you overtook a lorry at 72, you sat at a red light doing 0. Question: was there ever an instant at which your speedometer read exactly 60 mph?

Yes — provably, unavoidably yes. If you spent the whole hour below 60 you couldn't have covered 60 miles; always above 60 and you'd have covered more. Somewhere between the crawling and the overtaking, your instantaneous speed must have passed through the average. That everyday certainty, made rigorous, is the Mean Value Theorem (MVT) — and it is not a curiosity: police average-speed cameras convict drivers on the strength of it, and half of calculus quietly leans on it.

In the language of the derivative, the MVT converts a fact about f' — a local, instantaneous quantity — into a fact about f over a whole interval. It says the average rate of change is achieved exactly, at some interior instant:

Geometrically: draw the secant joining the endpoints (a, f(a)) and (b, f(b)). Somewhere strictly between them, the tangent to the curve is parallel to that secant. We will earn this in two moves: first a level special case (Rolle's theorem), then a tilt.

The statement — every word doing work

Analysis theorems are contracts, and the fine print matters. The MVT asks for exactly two things, and asks for them on different sets:

• Continuity on the closed interval [a, b] — endpoints included. This is what lets the Extreme Value Theorem run: no jumps, no missing values, so a maximum and minimum genuinely exist.
• Differentiability on the open interval (a, b) — endpoints excluded. The conclusion delivers a c strictly inside, so we never need a derivative at a or b. That generosity is deliberate: it lets the theorem apply to f(x) = \sqrt{1 - x^2} on [-1, 1], whose graph is a semicircle with vertical (hence non-differentiable) tangents at the very ends.

Rolle's theorem is the MVT with the ends at the same height: the secant is horizontal, so the promised parallel tangent is horizontal too — f'(c) = 0. In the driving picture, Rolle says: leave home and return home, and at some moment your velocity was exactly zero (of course — you turned around). The MVT is the same statement after tilting the axes.

Centrepiece, part 1: Rolle's theorem

Rolle. If f is continuous on [a, b], differentiable on (a, b), and the ends are level, f(a) = f(b), then f'(c) = 0 for some c \in (a, b). We derive it line by line.

Step 1 — get an extremum for free. f is continuous on the closed bounded interval [a, b], so by the Extreme Value Theorem it attains a maximum M and a minimum m on [a, b].

Step 2 — split into two cases. Either M = m, or not.

Step 3 — the flat case. If M = m, the function is constant, so f'(c) = 0 at every interior point — done.

Step 4 — the non-flat case forces an interior extremum. If M \ne m, at least one of them differs from the common endpoint value f(a) = f(b). So a max or a min is attained at some point c that is not an endpoint — i.e. c \in (a, b).

Step 5 — Fermat's stationary-point lemma kills the derivative. At an interior extremum a differentiable function has f'(c) = 0. Why: at an interior maximum the one-sided difference quotients have opposite signs,

so the limit (which exists, since f is differentiable) must be \ge 0 and \le 0 at once — hence f'(c) = 0. That proves Rolle.

Notice where each hypothesis was spent: continuity on the closed interval bought the extremum in Step 1; differentiability on the open interval was cashed in at Step 5, and only at the interior point c where it was needed.

Centrepiece, part 2: the MVT by tilting

The MVT is Rolle applied to the function with the secant subtracted off. The picture: Rolle handles level-ended functions; a general f has tilted ends, so we un-tilt it, apply Rolle, and translate back.

Step 1 — build the tilt. Let L be the secant line through the endpoints,

and define the gap between the curve and the secant,

Step 2 — check Rolle's hypotheses for g. g is continuous on [a, b] and differentiable on (a, b) (a difference of such functions). And it is level at the ends — because the secant touches the curve there:

Step 3 — apply Rolle to g. Some c \in (a, b) has g'(c) = 0.

Step 4 — read off g'. Differentiating g = f - L and using that L has constant slope,

Step 5 — set it to zero. At the c from Step 3,

The tangent at c is parallel to the secant. That is the Mean Value Theorem — one clever substitution away from Rolle, which was itself one lemma away from the Extreme Value Theorem. Analysis is a tower, and this is one of its cleanest storeys.

Watch the tangent track the secant

For f(x) = \tfrac14 x^3 - x, drag the endpoints a and b. The dashed secant connects the endpoints; the solid tangent is drawn at the MVT point c where their slopes match. They stay parallel no matter where you put the ends.

Two things to notice as you drag. First, c moves — it depends on the interval, and nothing in the theorem tells you where it will be. Second, try squeezing a and b together: the secant slope tends to the derivative at the squeeze point, and c, trapped between them, is dragged along. That trapping trick — "c is squeezed, so c \to a" — is exactly how the MVT gets used in limit arguments like L'Hôpital's rule below.

Worked examples: pinning down the promised c

The theorem never tells you where c is — but for concrete functions you can often solve for it. The recipe: compute the secant slope, set f'(c) equal to it, solve, and check the solution lands inside (a, b).

Example 1 — f(x) = x^2 on [1, 3]. The secant slope is

Set f'(c) = 2c = 4, so c = 2 \in (1, 3). Done — and notice c is the exact midpoint. That's no accident: for any parabola x^2 on [a, b], the equation 2c = \frac{b^2 - a^2}{b - a} = a + b gives c = \tfrac{a+b}{2}, the midpoint, every time.

Example 2 — f(x) = \sqrt{x} on [1, 4]. Secant slope: \frac{2 - 1}{4 - 1} = \tfrac13. Set f'(c) = \frac{1}{2\sqrt{c}} = \tfrac13, so \sqrt{c} = \tfrac32 and c = \tfrac94 = 2.25 \in (1, 4). Not the midpoint this time — c sits left of centre, where the steep early part of the square root does most of the climbing.

Example 3 — the speed camera. An average-speed check zone is 5 miles long. A car's number plate is photographed entering at 14:00:00 and leaving at 14:04:00 — four minutes, i.e. \tfrac{1}{15} of an hour. Let s(t) be the car's position: continuous (cars don't teleport) and differentiable (the speedometer always reads something). The average speed is

so by the MVT there was an instant c at which s'(c) = 75 mph — exactly. In a 70 mph zone, that instant is the offence. No camera ever saw the speedometer; a theorem did.

The payoff: the corollaries that run calculus

Here is why the MVT is a workhorse rather than a museum piece. Facts you have used a hundred times — "zero derivative means constant", "positive derivative means increasing" — are not obvious; they are two-line consequences of the MVT, and there is no honest route to them without it.

Corollary 1: f' = 0 throughout an interval \Rightarrow f is constant. Take any two points x_1 < x_2 in the interval. By the MVT on [x_1, x_2] there is a c with

so f(x_2) = f(x_1) — the value never changes. This is exactly why two antiderivatives of the same function differ only by a constant (their difference has zero derivative), which is the fact the "+ C" in every integral, and the Fundamental Theorem of Calculus itself, stand on.

Corollary 2: f' > 0 throughout an interval \Rightarrow f is strictly increasing. Same two lines, different sign: for x_1 < x_2, the MVT gives a c with

since both factors are positive. So f(x_2) > f(x_1): every increasing/decreasing argument in curve sketching, every first-derivative test, is the MVT wearing work clothes.

Corollary 3: a speed limit for functions. If |f'(x)| \le M on an interval, then for any two points in it, |f(x_2) - f(x_1)| = |f'(c)|\,|x_2 - x_1| \le M\,|x_2 - x_1|. A bound on the derivative becomes a bound on how far the function can travel — the germ of Lipschitz estimates, error bounds for Taylor polynomials, and uniqueness theorems for differential equations.

See it explained