Thought Toys · Chance & inference · Exhibit 31

Averages that never settle

Everyone knows the law of averages: keep averaging more numbers and the answer settles down. Here is a curve for which it never does. Its bell shape looks perfectly tame, yet a rare monster draw keeps flinging the running average clean off the chart — no matter how many numbers you pile on. Drag the tail heavier and watch the law of averages quietly break.

A chart of the running average as more and more random numbers are drawn. The horizontal axis is the number of draws so far on a logarithmic scale; the vertical axis is the running average, centred on zero. For a light tail the amber trace funnels into the centre line inside a narrowing grey band. For the Cauchy the trace repeatedly spikes to the top or bottom edge and never settles.

running average now — · biggest single draw so far —

your turn — drag tail heaviness right toward the bell curve and watch the average finally settle into the funnel

What you're seeing

Draw a random number, then another, and keep a running average of all of them. The amber line is that average as the pile of draws grows to the right (the axis counts draws on a log scale, so the whole life of the run fits on screen). The grey funnel is where the law of large numbers promises the average will be: a band that pinches toward the true centre like one over the square root of how many numbers you've drawn. Draw enough and you should be trapped in the throat of the funnel.

With a light tail — slide the dial right, toward the bell curve — that's exactly what happens. The trace wobbles early, then homes in and stays. Now drag the dial back to the left. The tail gets heavier: extreme draws that a bell curve would essentially never produce start showing up. Past ν = 2 the funnel vanishes entirely, because the spread — the variance, the average squared distance from the middle — is now infinite; the promise has no width to make. And at the far left, ν = 1, you meet the Cauchy: a curve that looks like a slightly sharper bell, but whose mean does not exist.

Watch what the Cauchy does. The average creeps toward zero, looks like it's settling — and then one draw lands at +400, or −1,900, and the average jumps, erasing everything the previous thousand draws had earned. It never settles because it can't: the average of a million Cauchy numbers is, provably, no more reliable than a single one. This is why "just take more data" is not always an answer. When the tails are heavy enough — as they can be for market crashes, city sizes, or wealth — the one freak event isn't noise around the average. It is the average.

The rule, exactly. The dial is the degrees of freedom ν of a Student-t distribution; the tail gets heavier as ν falls. The mean exists only for ν > 1 and the variance only for ν > 2. At ν = 1 it is the standard Cauchy, and the punchline is exact: average of n Cauchy draws ~ Cauchy  — the same distribution, for every n so averaging does nothing to sharpen it. Verified in node (improve/verify/31-cauchy.js): for ν = 8 the run-to-run scatter of the sample mean shrinks about 4× when the sample grows 16× (one over root n, the law of large numbers), while for the Cauchy it does not shrink at all — its sample mean still has the Cauchy's own interquartile range of 2 at every sample size. Counter-example: clip the very same Cauchy tail (throw out draws beyond ±50) and the average settles like one over root n again — proving it is precisely the heavy tail, not the peaked shape, that breaks the law of averages. (That clipping quietly makes a different distribution — a truncated Cauchy has finite variance by construction. It's a diagnosis of the cause, not a licence to delete the outliers you meet in the wild, where you can't know in advance which extremes are real.) (Each draw here is t = Z·√(ν/χ²ν), the same generator the proof uses.)

Also in Chance & inference: The birthday paradox →

All 16 in Chance & inference
  1. 05The Galton board
  2. 06The Monty Hall problem
  3. 09Bayes' theorem
  4. 13Buffon's needle
  5. 14The central limit theorem
  6. 28Simpson's paradox
  7. 30Markov chains
  8. 31Averages that never settle — you are here
  9. 35The birthday paradox
  10. 38The drunkard's walk
  11. 39Zipf's law
  12. 40Benford's law
  13. 42The coupon collector's problem
  14. 48The wisdom of crowds
  15. 52Genetic drift
  16. 63The gaps are chaos. The count is law.

← the cabinet · Thought Toys — a cabinet of explorable explanations. Exhibit 31.