The Galton Board: How Randomness Builds the Bell Curve

The construction: a lattice of coin flips

Build a board with n rows of pegs arranged in a triangular (quincunx) lattice: row 1 has a single peg, row 2 has two, row r has r pegs, for a total of n(n+1)/2 pegs. Directly below the last row sit n + 1 collecting bins. A ball dropped at the apex strikes the top peg and, ideally, is deflected left or right with equal probability. It then falls to a peg in the next row and repeats, making exactly one binary choice per row.

Number the bins k = 0, 1, …, n from left to right, where k counts how many times the ball went right. Every trip down the board is therefore a string of n independent coin flips — RLLRRL… — and the number of R's alone decides the final bin (the order does not matter for where it lands). Let X be that number of rights. Because the flips are independent and each is 1 with probability ½ and 0 with probability ½, X is a sum of n independent Bernoulli(½) variables:

X = B₁ + B₂ + ⋯ + Bₙ,   Bᵢ ∈ {0, 1},   P(Bᵢ = 1) = ½.

That single sentence is the whole physics of the device reduced to arithmetic. Everything else — the bell, the width, the limit theorem — is a consequence of studying the distribution of this one sum.

Counting paths: Pascal's triangle and the binomial

How many distinct peg-to-peg routes reach bin k? A route is a sequence of n moves containing exactly k rights and n − k lefts, so the count is the number of ways to choose which k of the n rows go right:

(number of paths to bin k) = C(n, k) = n! / [k!(n − k)!].

These are exactly the entries of Pascal's triangle, and the board is a physical Pascal's triangle: the number of paths through any peg equals the sum of the counts of the two pegs feeding into it, which is Pascal's rule C(n,k) = C(n−1,k−1) + C(n−1,k). Since every one of the 2ⁿ equally likely paths is a distinct coin-flip string, the probability of landing in bin k is the fraction of paths that go there:

P(X = k) = C(n, k) · (½)ⁿ = C(n, k) / 2ⁿ.

This is precisely the binomial distribution B(n, ½). The middle bins are wide because there are astronomically many ways to balance rights and lefts, while the edge bins are narrow because "all rights" or "all lefts" is a single path each (C(n,0) = C(n,n) = 1). For n = 10, the center bin has C(10,5) = 252 paths versus just 1 for the extremes — a 252-to-1 ratio that is the seed of the bell shape.

The numbers: why the spread grows like √n

Because X is a sum of independent pieces, its mean and variance add up piece by piece. Each Bernoulli(½) has mean ½ and variance ½·½ = ¼, so:

• Mean (peak bin): E[X] = n · ½ = n/2.
• Variance: Var(X) = n · ¼ = n/4.
• Standard deviation: σ = √(n/4) = √n / 2.

The √n growth of σ is the single most important number in the whole story, and it explains a subtle double behavior. In absolute terms the pile gets wider as you add rows — σ climbs like √n, so a 100-row board (σ = 5 bins) spreads twice as far from center as a 25-row board (σ = 2.5 bins). But in relative terms the pile gets narrower: the fractional width σ/n = 1/(2√n) shrinks to zero. That relative concentration is the Law of Large Numbers — the fraction of rights, X/n, homes in on ½ — while the absolute √n spreading is what a random walk does as it diffuses. Both facts live in the same formula. A useful rule of thumb: about 68% of balls land within σ = √n/2 bins of center, and about 95% within 2σ = √n bins, exactly the normal-curve tail proportions the next section derives.

Why a bell? The de Moivre–Laplace theorem

The bell is not an accident of the picture; it is a theorem. In 1733 Abraham de Moivre proved (for p = ½) that the binomial probabilities, when there are many rows, are approximated by a smooth exponential-of-a-square — the shape Laplace later generalized to any p in 1812. In modern form, for k near the mean np,

P(X = k) ≈ 1 / √(2π·np(1−p)) · exp( −(k − np)² / (2·np(1−p)) ).

For the fair board p = ½, so np(1−p) = n/4 and this collapses to

P(X = k) ≈ √(2/(πn)) · exp( −2(k − n/2)² / n ).

Here is the reasoning in one breath. Take logs of P(X = k) = C(n,k)/2ⁿ and feed in Stirling's approximation m! ≈ √(2πm)·(m/e)ᵐ. Writing k = n/2 + x and Taylor-expanding the resulting log around the peak x = 0, the constant and linear terms in x vanish (the peak is a maximum, so the first derivative is zero there), and the leading survivor is a quadratic: log P ≈ (const) − 2x²/n. Exponentiate and you have a Gaussian in the displacement x, with the √(2/(πn)) prefactor coming straight from Stirling's √(2πm) factors. As a check, the peak value √(2/(πn)) at n = 100 gives ≈ 0.0798, matching the exact center probability C(100,50)/2¹⁰⁰ ≈ 0.0796.

Rescale to standard units by defining Z = (X − n/2)/(√n/2) = (2X − n)/√n. Since one bin of X is 2/√n units of Z, converting the probability into a density divides by that spacing and the √n factors cancel perfectly, leaving the standard normal density 1/√(2π)·e^(−z²/2). In symbols, (2X − n)/√n → N(0, 1) as n → ∞. The bell curve on the board is literally this limit made out of beans.

Each ball is a random walk; the board is the Central Limit Theorem

Recode each peg decision as a step of ±1 (right = +1, left = −1). Then a single ball performs a simple symmetric random walk: after n steps its net displacement is S = (rights − lefts) = 2X − n, an integer that ranges over −n, …, +n and lands the ball in the bin at horizontal offset S/2 from center. The distribution of where one walker ends up after n steps is exactly the distribution the board displays. So the Galton board is a parallel random-walk experiment: thousands of independent walkers, each taking n steps, and the histogram of their endpoints.

Viewed this way, the emergence of the bell is a hands-on demonstration of the Central Limit Theorem. The CLT says that a sum of many independent, identically distributed random variables with finite variance — whatever their individual shape — has a distribution that approaches the normal as the number of terms grows. The board uses the simplest possible summand, a fair ±1 coin, but the theorem's reach is the point: it is why the bell curve is ubiquitous in nature. Measurement errors, particle diffusion, heights in a population, the sum of many small independent jolts — each is a sum of many little random contributions, and each therefore inherits the same bell. Galton's genius was to make that abstract convergence something you can watch pour out of a funnel, one bounce at a time. The de Moivre–Laplace theorem is simply the CLT for the special summand the board provides, and it was proven a full century before the general theorem was fully formalized.

Variants, caveats, and Galton's own uses

Change the peg rule and you change the limit, which is what makes the board a laboratory rather than a toy. Bias the pegs to send balls right with a fixed probability p ≠ ½ and the pile still becomes a bell — now centered at np with width √(np(1−p)) — the general CLT in action; the shape is normal even though the mechanism is lopsided. Build a multiplicative board, where each stage scales the ball's position by a random factor instead of adding a step, and because the log of a product is a sum of logs, the CLT applies to the logarithm: the pile becomes right-skewed log-normal, the distribution that governs incomes, particle sizes, and file lengths. Galton himself built a two-stage quincunx to illustrate regression toward the mean, a phenomenon he discovered while studying inherited stature.

Three honest caveats keep the demonstration rigorous:

• The ½ is idealized. Real pegs, ball spin, and elastic bounces make the left/right split only approximately fair and the steps not perfectly independent; a well-machined board keeps the bias tiny, but the physical bell is an approximation to the mathematical one.
• Convergence is in shape, not in every tail. The normal approximation is superb near the center (within a few σ), but the extreme bins have relatively larger error and are better handled by large-deviation bounds than by the Gaussian.
• Finite n is discrete. The exact answer is always the binomial C(n,k)/2ⁿ; the normal curve is the n → ∞ limit, excellent already by n ≈ 20–30 near the peak.

Sir Francis Galton (1822–1911) first described the device to the Royal Institution around 1873–74 and featured it in Natural Inheritance (1889); the modern nicknames "bean machine" and "quincunx" both stuck. Beyond its charm, it is a genuine teaching engine for the binomial distribution, Pascal's triangle, the random walk, and above all the Central Limit Theorem — the reason the bell curve is the most recognizable shape in all of statistics.

The standard board and three cousins — same machine, different peg rule, different limit

Board / mechanism	Rule at each peg	Limiting shape (large n)	What it demonstrates
Standard Galton board	Left or right, p = ½ (symmetric)	Symmetric normal, centered at n/2	de Moivre–Laplace: sums of fair coin flips → bell curve
Biased board (p ≠ ½)	Right with fixed probability p	Normal, shifted to np, width √(np(1−p))	The full CLT: any fixed bias still gives a bell, just off-center
Multiplicative (log) board	Multiply position by a random factor	Log-normal (right-skewed)	Sums of logs are normal → products are log-normal
Single ball, many rows	n independent ±1 steps	Simple symmetric random walk of length n	Endpoint spreads like √n — diffusion in one dimension