Buffon's Needle: Estimating π by Dropping Needles

The setup and the exact formula

Rule a floor with infinitely many parallel lines, evenly spaced a distance d apart. Take a thin needle of length ℓ with ℓ ≤ d (the short-needle case) and drop it so that its center lands at a uniformly random position and its orientation is a uniformly random angle. The event of interest is that the needle crosses (touches or intersects) one of the lines.

The exact result Buffon derived is:

P(cross) = 2ℓ / (πd).

Two random variables fully describe a drop. Let y ∈ [0, d/2] be the distance from the needle's center to the nearest line, taken uniform on [0, d/2] by symmetry, and let θ ∈ [0, π/2] be the acute angle between the needle and the lines, uniform on [0, π/2]. The needle reaches the nearest line exactly when its half-projection perpendicular to the lines exceeds y, i.e. when

y ≤ (ℓ/2)·sin θ.

The crossing probability is the fraction of the (y, θ) rectangle satisfying that inequality — which is what the next section computes.

Why it's true: an integral over all drops

Both variables are independent and uniform, so the joint density is constant: 1/(d/2) · 1/(π/2) = 4/(πd) on the rectangle [0, d/2] × [0, π/2]. The needle crosses iff y ≤ (ℓ/2) sin θ, so we integrate that density over the favorable region:

P = ∫₀^{π/2} ∫₀^{(ℓ/2) sin θ} (4/(πd)) dy dθ = (4/(πd)) ∫₀^{π/2} (ℓ/2) sin θ dθ.

The inner integral just returns the height (ℓ/2) sin θ. Now ∫₀^{π/2} sin θ dθ = [−cos θ]₀^{π/2} = 0 − (−1) = 1, so

P = (4/(πd)) · (ℓ/2) · 1 = 2ℓ/(πd). ∎

The intuition: a needle at angle θ has a vertical shadow of length ℓ·sin θ. It crosses a line precisely when a line falls inside that shadow, which happens with probability (ℓ sin θ)/d for any fixed θ (as long as that's ≤ 1). Averaging over all angles, you need the mean of sin θ. That average is exactly 2/π — the mean value of sin over a quarter turn is (1/(π/2))∫₀^{π/2} sin θ dθ = 2/π. Multiply the two pieces, ℓ/d and 2/π, and you get 2ℓ/(πd). The π appears because it is baked into how much a rotating segment sticks out on average.

A worked example: estimating π from 1000 drops

Set ℓ = d (needle length equals line spacing), the case the animation uses. Then the crossing probability is P = 2ℓ/(πd) = 2/π ≈ 0.63662. Rearranged, this gives the estimator π̂ = 2/P, and empirically we replace P by C/N (crossings over drops):

π̂ = 2N / C.

Suppose we drop N = 1000 needles and observe C = 631 crossings. Then:

• Observed crossing rate: C/N = 631/1000 = 0.631.
• Estimate: π̂ = 2·1000 / 631 = 2000/631 ≈ 3.1696.
• Error: |3.1696 − 3.14159| ≈ 0.028, about 0.9%.

Is 631 a plausible count? The expected number of crossings is N·P = 1000·0.63662 ≈ 636.6, with standard deviation √(N·P(1−P)) = √(1000·0.6366·0.3634) ≈ √231.3 ≈ 15.2. So 631 sits well within one standard deviation of the mean — a completely ordinary outcome. If instead you saw C = 637 (right on the mean), you'd get π̂ = 2000/637 ≈ 3.1397, off by only 0.002. The estimate wobbles because C is random; that wobble is the whole story of how fast this method converges.

How fast does it converge — and why it's slow

The count C is Binomial(N, P), so the estimator π̂ = 2N/C inherits its noise. By the delta method the standard error of π̂ scales like:

SE(π̂) ≈ π·√((1 − P)/(N·P)) = π·√((π − 2)/(2N)).

For ℓ = d that constant is π·√((π−2)/2) ≈ π·√0.5708 ≈ 2.37. So the typical error after N drops is about 2.37/√N. To get one extra correct decimal digit you must shrink the error tenfold, which means multiplying N by one hundred. Concretely:

• N = 100 → error ≈ 0.24
• N = 10⁴ → error ≈ 0.024 (≈ 2 digits)
• N = 10⁶ → error ≈ 0.0024 (≈ 3 digits)
• N = 10⁸ → error ≈ 0.00024 (≈ 4 digits)

This O(1/√N) rate is the signature of every Monte Carlo method, and it is brutally slow compared to a good series. Lazzarini's famous 1901 'experiment' reported π ≈ 3.1415929 from 3408 tosses — suspiciously accurate to 6 digits, which N = 3408 cannot honestly deliver (the honest error is ~0.04). He almost certainly stopped at a toss count engineered so that 355/113 (a superb rational approximation to π) popped out. It is a cautionary tale about stopping rules, not a triumph of the method.

The long-needle case and the Buffon–Laplace grid

When ℓ > d the needle can cross several lines at once, and the clean formula changes. Now define m = ℓ/d > 1. Integrating the same geometry (accounting for angles where the projection exceeds d) gives the general expected number of crossings:

E[crossings] = (2/π)·(ℓ/d) = 2m/π,

which is remarkably clean and holds for all ℓ: the expected number of line-crossings per drop is simply 2ℓ/(πd). (For ℓ ≤ d this expectation equals the crossing probability because you can only cross 0 or 1 line.) That means you can estimate π honestly even with long needles by counting total crossings, not just whether a crossing occurred — and this is the statistically efficient thing to do.

Buffon–Laplace variant: replace the parallel lines with a rectangular grid of spacing a × b. A short needle (ℓ ≤ min(a,b)) now has crossing probability P = [2ℓ(a + b) − ℓ²] / (πab). Two families of lines create more crossings per drop, so the estimator has lower variance for the same N — a genuinely better π-machine than single lines.

Why it matters and what to watch out for

Buffon's needle (posed 1733, published 1777) is the founding problem of geometric probability and a direct ancestor of integral geometry — the Cauchy–Crofton formula, which recovers the length of any curve by counting how often random lines cross it, is Buffon's argument generalized. It is also the historical seed of the Monte Carlo method: using random sampling to estimate a deterministic quantity you cannot compute directly.

Common pitfalls to avoid:

• Using the wrong formula for long needles. P = 2ℓ/(πd) exceeds 1 once ℓ > πd/2 ≈ 1.57d — a probability can't do that. Beyond ℓ = d you must use the expected-crossings form, not the naive one.
• Non-uniform dropping. The derivation demands θ uniform on angles and center uniform in position. A biased hand, a needle that bounces, or clustered throws breaks the assumption and biases π̂.
• Cherry-picking the stopping count (the Lazzarini trap). Deciding to stop when the estimate looks good manufactures accuracy that the sample size cannot support.
• Expecting speed. This is a beautiful demonstration, not a practical π-computer — even a laptop running 10⁹ virtual drops only reaches ~4 correct digits, while the Leibniz series or Chudnovsky formula crush it.

Buffon's Needle vs. other Monte Carlo ways to estimate π

Method	Setup	π estimate	Efficiency
Buffon's needle (ℓ = d)	Drop needles on ruled lines, count crossings C	π ≈ 2N/C	Very slow; ~10⁶ drops for ~2–3 digits
Buffon–Laplace grid	Drop on a rectangular grid, count crossings of either axis	Higher crossing rate → less variance	Better than single lines, still O(1/√N)
Circle-in-square darts	Throw points in [0,1]², count fraction inside quarter-circle	π ≈ 4·(inside/total)	O(1/√N), simple, no needle geometry
Series (e.g. Leibniz)	Sum 4(1 − 1/3 + 1/5 − …)	Deterministic partial sums	O(1/N) — beats all random methods