Skip to the content
Chaos, Drawn

Chapter 4 · Tool oneThe law of large numbers

You cannot call one die. You can call a thousand of them cold. Chaos hides in the single throw; certainty is waiting in the pile. This is the first tool for living with chaos, and it is older than the word.

One die, three thousand times

A fair die is close enough to unpredictable for anyone's purpose — you will not call the next face. But keep a running average of the faces as you roll, and something firm happens.

150010001500200025003000123456number of rolls Naverage so farOne fair die, rolled 3,000 timesthe true average, 3.5±2·σ/√N
Fig 8. The running average staggers around early — a couple of high rolls and it lurches up — then it is reeled in, tighter and tighter, to 3.5. After 3,000 rolls this run sat at 3.499. The shaded funnel is how far off you should expect to be, and it closes like 1 over the square root of the number of rolls. · computed here · Nan · hongdam.net · CC BY 4.0
average of N rolls → μ,   with a spread of about σ / √N

Reading it: as the count N climbs, the average closes on the true mean μ (3.5 for a die), and the wobble around it shrinks as one over the square root of N. Four times the rolls, half the wobble.

Jacob Bernoulli proved this in a book published in 1713, and was proud enough to call it his golden theorem.16 It is the ground under every poll, every insurance premium, every casino floor. None of them can call the one. All of them can call the many, and they bet the building on it.

The bridge back to chaos

Here is where it gets good. Take the most chaotic rule on this whole site — the rabbit map with its knob turned all the way to 4, the one whose path never repeats and cannot be forecast a dozen steps out. You cannot say what its next value will be. But drop three hundred thousand of its values into bins and count them.

00.20.40.60.8100.250.50.7511.251.5value xhow often300,000 values of a map that never repeatsthe curve 1 / (π √(x(1−x)))
Fig 9. The single values are unforecastable. The pile they make is a fixed, exact curve — more time spent near the ends, less in the middle — and it is the same curve every run, 1 over π times the square root of x(1−x). · computed here · Nan · hongdam.net · CC BY 4.0

The individual step: chaos, no forecast. The long-run shape: a law, down to the decimals, known in closed form since von Neumann used this very map to make random numbers in 1947.15 That is the tool in one picture. When you cannot predict the case, predict the distribution. Stop asking which way this one falls and ask how the whole heap settles — the heap holds still even when every grain in it is jumping.

How to pick it up

Do not forecast the customer; forecast the month. Do not time the one trade; hold the many. Do not ask whether it rains Tuesday; ask how many wet days the season brings — that number barely moves. The law of large numbers turns a wall of chaos into a fact you can plan around, as long as you are willing to zoom out from the one to the many.