Why everything becomes a bell curve

Here's something a bit odd once you notice it. The same curve keeps showing up in completely unrelated places: the heights of a thousand people, the errors in a batch of measurements, the marks on an exam, the noise on a sensor line, the spread of a dart-throw. Different causes, different units, and yet the histogram is always that same humped, symmetric bell. That can't be a coincidence, and it isn't. It's about the closest thing statistics has to a law of nature, and I think it's genuinely beautiful once you see why it has to be true.

The short version: when you add up a lot of independent random things, the sum stops caring what the individual things looked like and settles onto one fixed shape. That's the central limit theorem (CLT), and the rest of this post is us working out why the bell, and not some other curve, is the one everything falls into.

Watch it happen

Let's start by building one. A Galton board is a triangle of pegs; you drop a ball in the top and at every peg it bounces left or right, roughly fifty-fifty. By the time it reaches the bottom it's made a string of independent little decisions, and where it lands records how many went right. Drop one ball and it's anyone's guess. Drop a few thousand and a shape appears.

Each ball's final bin is a sum of $n$ independent $\pm 1$ steps, so the bins follow the binomial distribution: bin $k$ collects a fraction $\binom{n}{k}/2^n$ of the balls. Crank the rows up and the binomial gets smoother and smoother until it's indistinguishable from the bell curve drawn over the top. We didn't put the bell there. It emerged from nothing more than "add up a lot of coin flips".

A sum is a walk

There's a second way to picture the same fact, and it connects to a thing I find I reach for constantly: a sum of random steps is just a random walk. Each step nudges you left or right; the position after $n$ steps is the running total. Watch a few walks and you'll see them fan out from the origin, and the cloud of where-they-end-up is (of course) a bell.

The theorem, stated plainly

Take independent, identically distributed random variables $X_1, \dots, X_n$, each with mean $\mu$ and a finite variance $\sigma^2$. Their variances add:

$$\operatorname{Var}\!\left(\sum_{i=1}^{n} X_i\right) = n\sigma^2$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

Now centre the sum (subtract its mean) and rescale it by that $\sqrt{n}$ width. The claim of the CLT is that this standardised sum stops depending on the details of $X$ at all and converges to a single distribution:

$$Z_n = \frac{\sum_{i=1}^{n} X_i - n\mu}{\sigma\sqrt{n}} \;\xrightarrow{\;d\;}\; \mathcal{N}(0,1)$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

and that limit is the bell curve itself, whose density is

$$\varphi(z) = \frac{1}{\sqrt{2\pi}}\, e^{-z^2/2}$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

What I love here is how little the theorem asks for. It doesn't care whether $X$ is a coin flip or a die or a lopsided spinner. It only needs the pieces to be roughly independent and to have a finite variance. Hand it that, and out comes the same curve every single time.

When it doesn't work

A theorem is only as trustworthy as its assumptions, and the CLT's assumptions are quietly load-bearing. Push on them and the bell goes away.

The big one is finite variance. The proof leans on $\sigma^2$ being a real, finite number to rescale by. Some distributions don't have that. The classic troublemaker is the Cauchy: average a thousand Cauchy samples and you get... another Cauchy, exactly as wild as a single sample. No sharpening, no bell, no convergence. Its tails are too heavy; rare enormous values dominate the sum no matter how many you take.

Some food for thought, and I don't have a clean answer: the bell curve is so good at describing the well-behaved middle of things that it trains our intuition to expect the middle everywhere. A lot of expensive surprises (financial blow-ups, "unprecedented" failures) are really just someone's bell-curve intuition meeting a fat-tailed world. Worth keeping a little suspicion handy whenever a histogram looks reassuringly normal.

So why is everything a bell curve?

Because so much of what we measure is secretly a sum. A person's height is a sum of many small genetic and environmental nudges. A measurement error is a sum of many tiny independent errors in the apparatus. Sensor noise is a sum of countless little physical disturbances. Each of those is a Galton board you can't see, dropping its ball through a thousand invisible pegs, and the CLT guarantees where the pile ends up. The bell isn't a fact about heights or errors or noise. It's a fact about adding, and adding is everywhere.

I want to come back to the Gaussian from the other direction in a later post, because it's also the maximum-entropy distribution for a given variance, which is a completely different reason for it to be everywhere and yet lands on the same curve. Same shape, two unrelated stories. That kind of coincidence-that-isn't is my favourite thing in maths. Watch this space.

Reading further

• de Moivre (1733) and Laplace: the bell curve was first found exactly here, as the limit of the binomial — the Galton board, two centuries early. A nice account is in any history-of-statistics text; Stigler's The History of Statistics is the standard one.
• Galton, Natural Inheritance (1889): where the quincunx (the board above) comes from, and a lovely period read on the idea of regression to the mean. galton.org
• Strogatz, The Joy of x: the gentlest correct explanation of why the bell is special, if you want intuition over proof.
• Taleb, The Black Swan: the case against borrowing the bell for fat-tailed things, argued at length and with feeling. Read it as the warning callout above, expanded to a book.