In 1777, Georges-Louis Leclerc, Comte de Buffon, posed a question about a parlor game: if a needle is dropped at random onto a floor of parallel boards, what is the probability that it comes to rest crossing one of the cracks? The answer turned out to contain \(\pi\), which means the game can be run in reverse. Drop many needles, count how often they cross a line, and the count estimates \(\pi\) itself, extracted from randomness by nothing more than arithmetic.

The result is worth knowing for its own sake, and it is also a perfect miniature of an idea at the center of modern statistics: that simulation can compute quantities which have nothing visibly random about them.

The setup and the answer

Rule a floor with parallel lines a distance \(d\) apart, and drop a needle of length \(\ell\), with \(\ell \le d\), so that its position and angle are entirely random. Buffon showed that

\[P(\text{needle crosses a line}) = \frac{2\ell}{\pi d}\]

The appearance of \(\pi\) is not mystical. A dropped needle has a random angle, angles live on a circle, and averaging over all angles drags the circle’s constant into the answer. The derivation is a short exercise in a college probability course, and for BC students it is a pleasant surprise: the average of \(\sin\theta\) over all angles is computed with an integral, so the formula is calculus and probability shaking hands.

If the needle’s length is exactly half the spacing, so \(\ell = \tfrac{d}{2}\), the formula collapses to \(P = \tfrac{1}{\pi}\). A needle crosses a line about 31.8% of the time, and the estimate becomes

\[\pi \approx \frac{\text{number of drops}}{\text{number of crossings}}\]

The experiment

The floor below uses that half-spacing needle. Drop them by the hundred and watch the estimate close in.

Needles that cross a line are drawn dark; needles that land clear are faint. The readout divides drops by crossings to estimate pi. The first hundred needles usually land within a few tenths of the truth; ten thousand may get two decimal places, on a lucky day. Notice how slowly the precision improves. Cutting the error in half requires roughly four times the needles, the same square-root law that governs margins of error in polling, and a preview of why real surveys cannot buy accuracy cheaply.

Why a statistics student should care

The needle is the ancestor of the Monte Carlo method, the technique of computing a fixed, deterministic quantity by designing a random experiment whose long-run behavior encodes it. The Law of Large Numbers guarantees the convergence: the proportion of crossings must settle toward \(\tfrac{1}{\pi}\), so the reciprocal must settle toward \(\pi\). Every simulation in the AP course, from randomization distributions to simulated p-values, runs on exactly this engine, and the investigative-task questions that ask students to interpret a simulation are asking them to reason the way Buffon’s floor does.

The slow convergence carries its own lesson. The estimate’s error shrinks in proportion to \(\tfrac{1}{\sqrt{n}}\), which is the same \(\sqrt{n}\) that sits in the denominator of every standard error formula in the course. A political poll of 1,000 people and a floor of 1,000 needles are bound by the same arithmetic, and neither can be made ten times more precise without one hundred times the effort.

A historical footnote with a caution attached: in 1901 the Italian mathematician Mario Lazzarini reported tossing a needle 3,408 times and obtaining pi to six decimal places, a result far too accurate for the number of trials. His needle-to-spacing ratio was chosen so that a particular lucky count would reproduce the famous fraction 355/113, and the trial count appears to have been chosen to stop at the right moment. The episode is now a standard example in the study of suspicious data, which makes it, fittingly, a statistics lesson twice over.