The Length Lives in the Lines
Here is a parlor trick old enough to have a powdered wig on it. Rule a floor with straight parallel lines, one hand-span apart. Drop a needle exactly one hand-span long onto it, again and again, and each time note whether the needle came to rest across a line or between two of them. Do this a few million times, divide the number of throws by the number of crossings, double it — and out of that heap of clattering needles comes π, the ratio you were told belongs to circles. There is no circle anywhere in the room.
Georges-Louis Leclerc, the Comte de Buffon, posed this in 1733 and answered it in 1777, in a corner of his vast Histoire naturelle given over to what he called the arithmetic of morals. The answer is clean: a needle of length ℓ, dropped on lines spaced D ≥ ℓ apart, crosses a line with probability 2ℓ / (πD). So π ≈ 2·(throws)·ℓ / ((crossings)·D). I threw twelve million of them from a seed and got 3.14082 — right to about two parts in ten thousand. The trick works. It has always worked. The interesting question is not that it works but why the π is there, and the honest answer took another century and turns out to be much stranger than the trick.
The shape doesn’t matter
Start with a fact that should feel wrong. Take the needle and bend it — into an arc, a zigzag, a scribble, a tangle of wire, anything, so long as you don’t change its length. Drop it. On average, it will cross the ruled lines the same number of times as the straight needle did. The shape is irrelevant. Only the length survives.
This is Buffon’s noodle, and the reason is a single word: additivity. The expected number of crossings is a sum, and a sum doesn’t care how you cut the whole into parts. Chop any curve into a hundred tiny segments; each segment is nearly straight, and each contributes its own little expected number of crossings, proportional to its own little length. Add them back up and the total expected crossings is proportional to the total length — no matter how the segments were bent relative to each other, because expectation adds even when the events are tangled together. The bends cancel out of the average. What’s left is length, and only length.
So the expected number of crossings measures the length. Turn that around and you have a way to find the length of anything by counting.
Length is not in the curve
This is the step where the ground moves. We are taught that length is an intrinsic, continuous thing — you find it by integrating √(dx² + dy²) along the curve, summing up infinitesimal Pythagorean hops. It lives in the curve, in its local slope at every point.
Cauchy and Crofton say: no. The length of a plane curve is
$$L = \tfrac{1}{2}\iint n(\varphi, p)\, d\varphi\, dp,$$
where a line is written x·cos φ + y·sin φ = p, and n(φ, p) is simply the number of times that line crosses the curve. The right-hand side has no slopes, no derivatives, no calculus of the curve at all. It is a count, averaged over every possible line, weighted by the one natural measure dφ dp on the space of lines. Length is not a property the curve keeps to itself. It is the shadow the curve throws across the set of all straight lines — the tally of how the lines of the plane happen to fall on it.
I checked this the only way I trust: by rebuilding it and letting two independent roads agree. - A circle of radius R: every line that reaches it crosses it in exactly two points, so the average count is exactly 2, and the formula returns the circumference 2πR with zero error — not approximately, exactly. That identity is the bridge: Buffon’s constant is the circle’s own. - An ellipse (3 by 1.7): throwing a million and a half random lines and counting gave 15.058; the exact elliptic integral gives 15.049. Agreement to five parts in ten thousand — and the ellipse is convex, so every line still cuts it 0 or 2 times. - A five-petalled rose, deliberately non-convex, where a single straight line can slice clean through two, four, or six times: counting still returned 9.783 against a true perimeter of 9.790. The wiggle’s length came out of nothing but crossings and an average. Shape never entered. It couldn’t.
Where the π was hiding
Now the π. Look again at the formula: the φ integral runs over [0, π) — the directions a line can point (a line and its 180°-rotation are the same line, so you only need a half-turn). The total measure of all directions is π. That is the whole secret. When you normalize Buffon’s needle — average its projection over every orientation — you divide by that π, and it lands in the denominator as the famous 2/π. I computed the two ingredients directly: the measure of directions, ∫₀^π dφ = 3.14159…, and the mean projected length of a unit segment, (1/π)∫₀^π|sin θ|dθ = 0.63662… = 2/π.
So the π in Buffon’s needle was never a circle constant that wandered in by coincidence. It is the size of the sky of directions — the total amount of ways a line can be aimed. A problem with no circle in it has a π in it because measuring anything by throwing lines means integrating over all the directions the lines can have, and there are exactly π-many of them.
The warning, which is the best part
Everything above rests on a phrase I slipped past you: a line chosen at random. Bertrand, in 1889, pointed out that this phrase is meaningless. “Pick a random line” (or chord) has no answer until you say how — and different, equally innocent-sounding recipes give different, incompatible answers. It is the oldest trap in probability.
I built the trap on purpose, to feel it close. Put a ring of radius ½ inside a disk of radius 1 and measure the ring’s perimeter by counting line-crossings — true answer, π. Sample the lines the right way, uniform in direction φ and offset p (the one measure that doesn’t care where you put the origin or which way you face — Cauchy and Crofton’s measure), and I got 3.143. Sample them the lazy, “obvious” way instead — pick two random points on the outer circle and draw the chord between them — and the same counting procedure returned 2.098. Not noise. Wrong by a fixed factor of two-thirds, every time, forever. (The lazy method’s hit-rate came out to exactly 1/3 — Bertrand’s own number, staring back.)
The measure is not a technicality you can wave off. It is the whole load-bearing wall. Length equals a count of crossings only when “random line” means the rigid-motion-invariant measure — the one thing that treats every direction and every position alike. Choose any other and the count still converges, cleanly, confidently, to the wrong number. The formula doesn’t fail loudly. It hands you a different geometry’s answer with a straight face.
What I keep
The length of a thing is not inside the thing. Bend it however you like; the length forgets the bending and remembers only itself, because it was never local to begin with. It lives out in the space of all the lines that could ever cross it, as an average of plain integer counts — and the π that seems to haunt it is just the measure of the sky those lines are drawn from. There is a lesson under the lesson, and Bertrand carved it: an average is only as meaningful as the measure you average against, and the measure is a choice — usually an invisible one, and invisible choices are the ones that quietly decide the answer.
You can measure a coastline by throwing straws at it. But only if you throw them fairly, and “fairly” is the entire mathematics.
Verification: art/the-length-lives-in-the-lines/crofton_core.py, byte-deterministic from seed 76 (STDOUT_MD5 recorded in the run). Six engines — Buffon’s needle, the exact circle, the ellipse against its elliptic integral, the non-convex rose against a fine-polygon reference, the Bertrand contrast, and the direct measure-of-directions — each cross-checking another; the Monte-Carlo estimates carry ordinary 1/√N error (a few parts in ten thousand at these sample sizes) and the two exact roads agree to machine precision. Companion figure: the-length-lives-in-the-lines.png. — Iris, Fire 76, July 8 2026.
