A bell curve assembled out of coin flips

A ball is dropped onto a peg. Nothing about the apparatus is subtle, which is exactly why Galton built it as an argument rather than an experiment. It bounces left or right — genuinely unpredictably, with no bias either way. Below that peg are two more, and the ball hits one of them and bounces again. Twelve rows of this, and it falls into a bin at the bottom.

Where does it land? There is no answer, in the strong sense: this is not ignorance about a determined outcome but genuine indeterminacy, and no amount of information about the apparatus would fix it. Contrast the primes, which look random and are completely determined. That is the point: the path is a sequence of twelve coin flips and the outcome is not determined by anything.

Now drop six hundred balls.

The individual outcome is unpredictable and the aggregate is not. This is not a small effect that needs statistics to detect. It is a shape, visible immediately, and it turns up every time.

Why the middle wins #

The mechanism is elementary counting, and it is worth doing because it demystifies the whole thing.

To reach the far-left bin after twelve rows, a ball must bounce left twelve times in a row. There is exactly one route that does this. Its probability is $(1/2)^{12}$, about one in four thousand.

To reach the middle bin, a ball must bounce left six times and right six times, in any order. The number of ways to arrange six lefts among twelve bounces is $\binom{12}{6} = 924$.

So the middle bin is 924 times more likely than the far-left one, not because the middle is attractive but because there are 924 times as many ways to get there.

This is the single most useful idea in elementary probability and the most reliably misunderstood. The bins are not equally likely, and yet nothing is biased. The asymmetry is entirely in the counting — in how many distinct histories lead to the same destination — and that is a fact about combinatorics rather than about physics. It is the same move as counting one thing two ways, used here to compare rather than to equate. Every individual path is exactly as likely as every other individual path — the sequence LLLLLLLLLLLL has precisely the same probability as LRLRRLLRLRRL. What differs is how many paths end in the same place.

The bin counts are therefore the entries of row 12 of Pascal’s triangle, which is the same object counting the same thing: the number of paths from the apex to each position. The Galton board is Pascal’s triangle with gravity applied.

It also explains why the effect strengthens with more rows rather than washing out. Adding a row multiplies the number of routes to the middle far faster than the number to the edges, so the contrast sharpens. Twelve rows already give a ratio of 924 to 1; twenty give 184,756 to 1. The bell does not emerge gradually from noise — it is there from the third row and merely becomes harder to mistake for anything else.

The counts make the picture checkable rather than merely suggestive, and the six-row board is small enough to check by hand. There are $2^6 = 64$ routes, distributed $1, 6, 15, 20, 15, 6, 1$, so four hundred balls should fall as $6.25,\ 37.5,\ 93.75,\ 125,\ 93.75,\ 37.5,\ 6.25$. They fall as $9,\ 38,\ 97,\ 124,\ 90,\ 35,\ 7$. Every bin is within a few of a number computed before any ball was dropped, and the computation used nothing but counting.

The twelve-row board makes the sharper prediction. Its outermost bins should receive $1/4096$ of the balls each — out of six hundred, that is $0.15$ of a ball, so the expectation is that the extreme bins are empty or nearly so. They hold one ball and none. A shape that merely looked bell-like would not have earned that; a claim that a particular bin will be essentially empty is one that could have failed and did not.

The curve #

The predicted shape is the binomial distribution, and for a large number of rows it is very close to the normal distribution — the bell curve, with formula

$$\frac{1}{\sigma\sqrt{2\pi}}e^{-(x-\mu)^2/2\sigma^2}$$

which contains, for no reason visible in the setup, both $\pi$ and $e$. There are no circles in a peg board and no exponential growth. Both constants arrive out of the analysis rather than the apparatus, and $\pi$ in particular is turning up uninvited yet again — this time from the integral that normalises the curve.

The approximation of binomial by normal is de Moivre’s, from 1733, and it is the earliest form of what is now called the central limit theorem.

What the theorem actually says #

The central limit theorem is one of the most widely invoked and most frequently overstated results in mathematics, so it is worth stating carefully.

It says: if a large number of independent random quantities are added together, and each contributes only a small amount to the total, then the distribution of the total approaches a normal distribution — regardless of the distribution of the individual quantities.

That last clause is the extraordinary part. The individual bounces here are coin flips, which are about as far from bell-shaped as a distribution can be: two outcomes, equally likely, nothing in the middle. Add twelve of them and the result is already visibly a bell. The shape is not inherited from the inputs. It is created by the addition.

Which is why the normal distribution shows up so relentlessly in measurement error, in biological variation, in noise of every kind. Any quantity that is the sum of many small independent contributions ends up normal, and it does not matter what the contributions look like individually.

It does not say that everything is normally distributed, and this is where the theorem gets abused. The conditions are real:

• Independent. Contributions that influence each other break it. If one ball’s bounce made the next ball more likely to bounce the same way, the pile would be a different shape.
• Small individually. If one contribution dominates the sum, the total inherits that one’s distribution instead. Wealth is not normally distributed, and one reason is that its contributions are not small and not independent.
• Finite variance. Distributions with heavy tails — where extreme values are not rare enough — do not obey the theorem at all. Sums of Cauchy-distributed variables are Cauchy-distributed, no matter how many are added, and never approach a bell curve.

Assuming normality where those conditions fail is a well-documented way to lose a great deal of money. Financial returns have heavier tails than the normal distribution, and models that assumed otherwise have failed expensively and repeatedly.

Galton’s actual board #

Francis Galton built the physical device in the 1870s — he called it a quincunx — as a demonstration for his work on heredity, and it was a piece of rhetoric as much as apparatus. He wanted to make visible how a trait could vary continuously and predictably across a population while being unpredictable in any individual.

The idea he was extending came from Adolphe Quetelet, who had taken the bell curve out of astronomy and pointed it at people. In astronomy the curve described errors: many measurements of one star’s position scatter around the true value, and the spread is the instrument’s failing rather than the star’s. Quetelet applied the same curve to the chest measurements of Scottish soldiers, found it fitted, and concluded that there was an homme moyen — an average man — of whom actual men were imperfect copies.

That is a conclusion about which the mathematics is completely silent. A bell curve says a quantity is the sum of many small independent contributions. It does not say that the mean is a target and the spread a defect, and the whole difference between describing variation and pathologising it sits in that unearned step. Galton inherited the framing along with the curve.

He is a genuinely uncomfortable figure. He coined the term eugenics and the demonstration was in service of that programme; his statistical contributions are real and permanent, and what he wanted them for was not. Both things are true and neither cancels the other.

The device itself is honest and has been rebuilt many times, usually in museums, usually with beads. It works every time, which is the point.

It is also worth noticing what kind of argument a working machine is. It cannot prove the theorem — no finite apparatus can, for the reasons above — but it does something a proof does not, which is to make the claim impossible to disbelieve. Michelson’s harmonic analyser did the same for the Gibbs overshoot, and there the machine was ahead of the mathematicians: it drew the spikes, and the spikes were real, and someone had to go and work out why.

The bell is one attractor, not the only one #

The three conditions above read as a list of ways the theorem can fail, which undersells what is going on. The failures are not an absence of structure. They are a different structure.

Ask the question the theorem answers, but without assuming the answer: which distributions can arise as the limit of a sum of many independent copies of one thing? A limiting shape has to be stable — the sum of two independent copies of it must be the same shape again, up to rescaling and shifting. Otherwise adding one more term would change the answer, and it would not be a limit.

That requirement alone almost settles the matter. The stable distributions form a family with one main parameter, an index $\alpha$ between 0 and 2, controlling how heavy the tails are. The normal distribution is the case $\alpha = 2$ — and it is the only member of the family with a finite variance. The Cauchy distribution is $\alpha = 1$. The rest fill in between and below.

So the honest statement is not “sums become normal, except when they do not”. It is: sums are drawn toward a stable law, and which one depends on the tail behaviour of what is being added. Finite variance is precisely the condition that selects $\alpha = 2$, and since almost everything encountered in practice has finite variance, almost everything lands on the normal — which is why the special case acquired the general reputation.

The Cauchy behaviour mentioned above now looks less like a pathology. Adding Cauchy variables gives a Cauchy variable because the Cauchy distribution is already a fixed point, in exactly the sense the normal distribution is. It is not failing to converge; it has converged, to a different attractor, and it did so at the first term.

This reframing is worth having because it changes what a modelling assumption is. Choosing the normal distribution is not choosing “the shape sums make”; it is choosing $\alpha = 2$, which is a substantive claim about how often extreme values occur. The financial models that failed did not neglect a caveat. They picked one member of a family and did not notice they were picking.

How fast, and why this board flatters the theorem #

That twelve rows is “already enough” was read off the pictures, and it can be measured instead. Compare the exact binomial against the normal approximation and take the largest disagreement between their running totals. At six rows it is $0.0050$; at twelve, $0.0024$; at a hundred, $0.00027$; at a thousand, $0.000027$.

Those are falling like $1/n$ — the error multiplied by the row count gives $0.0284$, $0.0273$, $0.0272$, flat to three figures. And that is faster than the theorem is entitled to. The general rate, the Berry–Esseen bound, is $1/\sqrt n$, and it is not a pessimistic bound: run the same measurement on a board whose pegs send a ball right with probability $0.1$ and the error is $0.045$ at twelve rows, $0.017$ at a hundred, $0.0056$ at a thousand — error times $\sqrt n$ flat at $0.177$. By a thousand rows the biased board is two hundred times further from its bell than the fair one at the same size.

The reason is the one thing this apparatus cannot vary. Berry–Esseen’s leading term is proportional to the skewness of the quantity being added, and a fair coin has none. The first correction vanishes identically, and what is left is the next one, of order $1/n$. A biased peg is not symmetric and pays the full rate.

So the board is a demonstration in the most flattering case available, converging at the square of the general rate — which is why twelve rows are enough to look convincing, and why rather less should be concluded from that than it invites. The theorem’s real promise is that the biased board gets there too. It simply needs a hundred times as many rows to look as though it has.

What the picture cannot show #

The board has twelve rows and the theorem is about a limit. What the pile of six hundred balls shows is that twelve is already enough for the shape to be recognisable — which is a fact about binomial distributions, not a demonstration of the central limit theorem.

The distinction is not pedantic, because the theorem’s content is precisely the part that cannot be drawn: that the limiting shape does not depend on the distribution of the individual contributions. Every ball here makes the same fair binary choice. A board whose pegs were biased, or whose steps had wildly different sizes, would still converge to a bell — and that is the theorem. One apparatus can only ever exhibit one case of it.

The failures are equally invisible. A Cauchy-distributed sum never approaches a bell no matter how many terms are added, and no peg board can be built to show this, because the apparatus has bounded steps by construction. The conditions in the theorem are exactly the things a physical demonstration cannot violate, which makes the demonstration persuasive in a slightly misleading way.

The ladder from here #

Rungs above: the binomial distribution in its own right, and its mean and variance. De Moivre’s approximation, derived. Stirling’s formula, and the appearance of $e$ and $\pi$ in a factorial. The law of large numbers, and how it differs from the central limit theorem. Standard error, and why it falls like $1/\sqrt{n}$ — the same rate that makes Monte Carlo slow. Confidence intervals as coverage. Heavy-tailed distributions and the theorem’s failure modes. Random walks, and the connection back to the board read sideways. And the Berry–Esseen theorem, which says how fast the convergence actually is.

The general shape of the observation #

There is a category of result that says: the details do not matter.

The central limit theorem is one. It does not care whether the underlying randomness is coins, dice, measurement jitter or thermal noise; it says the sum forgets. That forgetting is what makes statistics possible at all, because it means a method built on the normal distribution can be applied to a process whose internals are unknown.

The mirror image is estimating $\pi$ by dropping needles on a floor, where randomness is used deliberately as a computational tool rather than tolerated as noise. In one, order emerges from randomness by aggregation. In the other, randomness is harnessed to compute something exact.

Both are the same underlying fact: a large number of independent trials has properties that no individual trial has, and those properties can be as sharp as anything in deterministic mathematics.