spread & inequality · metaphor 27 of 100

Worlds ruled by the
exception.

In some worlds the average person is a good summary; in others the average is a lie told by arithmetic. Fame, wealth, book sales, war deaths, startup outcomes: worlds where the typical tells you nothing and one extreme carries most of the total.

Ask what the average author earns and you will budget a life badly. The median author earns close to nothing; a handful earn nearly everything; and the mean — dutifully pulled upward by the handful — describes no author who has ever lived. Height lives in a bell: the typical person is a fine summary of everyone. Income lives in a power law: the typical outcome and the total are carried by entirely different people.

Confusing the two worlds may be the commonest planning error alive: preparing for typical outcomes in domains ruled by extremes, and preparing for extremes in domains ruled by the typical. Below, both worlds run live from the same button — the bell draws heights, the tail draws outcomes. Watch which summaries settle, and which never stop flinching.

The bell world heights · Normal(170, 9) cm
running mean running median max so far
draws0
max
mean
median
top 1% of draws hold
The tail world outcomes · Pareto(xmin=1, α)
running mean running median max so far
draws0
max
mean
median
top 1% of draws hold
tail exponent α 1.70
↤ 1.1 brutal · changing α restarts the tail world · 3.5 mild ↦

everything is drawn live: heights from Normal(μ=170, σ=9); outcomes from Pareto(xmin=1, your α). the monsters are real draws, not staged. the race pauses itself at 100,000 draws.

The 80/20 dial · shares of the total at your α

shares computed from the ideal Pareto curve at your α — a model, not a measurement. clicking a skin sets α to a rough historical fit (see the caveats before believing any of them).

The planner's table · the same three questions, asked of each world

question
bell world
tail world (your α)
Expected attempts before one outcome at least 10× your median?
In your sample so far, how much of the total does the single best draw carry?
Draws needed for the sample mean to sit within ±1% of the true mean (95% confidence)?

row 3 uses the CLT: n = (1.96/0.01)² · σ²/μ². for α ≤ 2 the Pareto variance is infinite — no sample size suffices, and the ∞ is exact, not rhetorical.

P(X > x) = (x / xmin)−α The whole law. Double the size, and the odds fall by a fixed factor 2α — the same factor at every scale. No scale is typical.

two kinds of world

Adding makes bells. Compounding makes tails.

Bells appear where many small, independent causes add up. Genes, nutrition, and childhood luck each nudge a height by a centimeter or two, and the nudges mostly cancel; the sum huddles around its center, and extremes are punished exponentially — a 2.2-metre person is a headline, a 3-metre person is impossible. In such a world mean, median, and mode agree, and the average is a genuine portrait: knowing the typical person, you know nearly everyone.

Tails appear where causes multiply — where advantage compounds. The book that sells gets the shelf space that sells it further; the city that grows attracts the jobs that grow it. Multiplication has no natural scale, so the mass of the distribution migrates into the extremes. Now every honest summary splits three ways: the median describes the crowd, but the crowd doesn't carry the total; the top 1% carries the total, but describes almost no one; and the mean is a promise about monsters — it budgets for draws you may never meet, and each monster that finally lands rewrites it upward. That lurch you can watch above is not noise on the way to the truth. In the tail world, the lurch is the truth.

the one-glance test

One plot tells you which world you're in.

Sort anything from largest to smallest and plot size against rank, both on logarithmic axes. A power law walks a straight line — the same proportional drop from rank 1 to 10 as from 10 to 100. A bell plunges off a cliff, because its extremes die exponentially. It is the fastest field test there is for "does the exception rule here?" — and, as the fine print below insists, it is a hint rather than a proof.

choose a dataset — rank on the x-axis, size on the y-axis, both logarithmic.

word counts are computed live from this very page's text. city figures are rounded 2020 US census populations in thousands — a hand-set toy. a straight-ish line over two decades is NOT proof of a power law: lognormals imitate it beautifully, and rigorous fitting (Clauset–Shalizi–Newman 2009) rejects most famous claims. the slope shown is the naive fit they warn against.

what to try

Sixty seconds of honest play.

living in tail worlds

Strategy for a world you can't average.

If your field is a tail world — art, startups, research, anything where compounding attention picks winners — then the median outcome is failure, and that is not information. It was always going to be. The quantities that matter are the ones the instrument surfaces: how many attempts you can afford, and whether your portfolio's maximum ever lands. So the strategy inverts the bell world's: make many cheap attempts rather than one perfect one; engineer your life to survive between hits, because the arrival time of the monster draw is exactly what the distribution refuses to promise; and judge a body of work by its best member, not its average — the average of a tail portfolio is a number about the failures.

And hold the ethics carefully. In a tail world, outcome inequality is compatible with identical talent: when success compounds, near-identical starting points diverge grotesquely, a mechanism this collection meets again as preferential attachment. The top-1% share above is generated by the mathematics of compounding alone — before any assumption about merit. That cuts both ways: the winner's fortune is not proof of proportionate genius, and the median artist's poverty is not proof of proportionate mediocrity. The distribution just isn't a talent meter.

planning errors, both ways

The two mistakes are symmetric.

Tail thinking in a bell world is the lottery mindset about safe quantities: waiting for the salary, the marathon time, or the exam score that is 10× the median. The planner's first row prices that hope honestly — in the bell world the wait is a 1-in-106278 event, which is a mathematician's way of writing never. Lives get parked on spikes the physics doesn't permit.

Bell thinking in a tail world is subtler and commoner. "The average startup fails" is a non-sequitur — in a tail world the average was never the object of interest; the portfolio's maximum was. Institutions split along exactly this line: insurance is a bell business — pool enough independent risks and the law of large numbers makes the mean your destiny, which is why the one thing that ruins insurers is a hidden fat tail (the correlated catastrophe) — while venture capital is a tail business, in which most of a fund's return is one company and "average outcome" is a category error. Same arithmetic, opposite religions. The unforgivable sin in each is practising the other's.

the mapping

Mathematics ↔ life.

MathematicsLife
the bellDomains where the typical is the truth — heights, commutes, calories. Know the average, know almost everyone.
the tailDomains where the extreme is the truth — fame, wealth, wars, sales. Know the average, know almost nothing.
exponent αHow brutal the inequality of outcomes is: one dial from mild skew (3.5) to winner-takes-nearly-all (1.1).
the wandering meanWhy "average outcome" misleads a single life: it budgets for monsters most lives never meet, then is rewritten by each one that arrives.
top-1% shareWho carries the total — the crowd, or the exception. The honest answer to "where did everything go?"
the log–log lineThe one-glance diagnostic: cliff or line, bell or tail — which planning religion applies here.

where the metaphor tears

Three honest failures.

The straight line seduces.

Most claimed power laws aren't. Lognormals and stretched exponentials walk convincingly straight for two or three decades, and the eye cannot tell them apart; when Clauset, Shalizi & Newman (2009) refit the famous datasets with real statistics, most celebrated "power laws" failed. This page's own slope fit is precisely the naive estimator they demolish — treat every straight-ish line here, including the ones this page draws, as a hypothesis to distrust. The metaphor's diagnostic is a doorbell, not a verdict.

Finite worlds truncate tails.

The ideal Pareto runs to infinity; no author sells infinite books, no war can kill more people than exist, no fortune exceeds the money supply. Real tails are cut off, and truncation quietly restores finite means and variances — the ∞ in the planner's table is a statement about the ideal curve, not about your actual career. The metaphor is at its most honest about the middle decades of a distribution and at its most theatrical about the far tail, which is exactly where it tempts you to quote it.

The distribution is partly chosen.

Tail worlds are made as well as found. Progressive taxation, insurance, tenure, and minimum wages exist precisely to move outcomes from tail toward bell; tournaments, platforms, and winner-take-all markets move them back. Treating α as a law of nature naturalises what is substantially a policy variable — the shape of a society's outcome distribution is one of the things a society decides, whether or not it admits it is deciding.