For thirty years, Benoît Mandelbrot studied cotton prices. The data went back a century. He was looking for patterns, looking for whether price movements could be modeled using standard statistics.

What he found was this: cotton prices didn't fit the bell curve that economists assumed. Instead, they showed something remarkable — self-similarity at different time scales.

When you looked at daily prices, weekly prices, monthly prices, and yearly prices, the pattern was the same. The distribution of moves looked identical whether you were zoomed in on a single day or zoomed out to a year.

This is the defining property of a power-law distribution: it looks the same at every scale. The distribution is self-similar. Fractal.

The finance industry ignored Mandelbrot for three decades, preferring the tractable Gaussian. Every major crash since has confirmed his view.

Today, power laws describe not just cotton prices but most quantities where one observation can dominate the total. Wealth, book sales, earthquake magnitudes, city populations, company sizes, war casualties. The pattern is universal.

And it changes everything about how you should think about risk.

What a Power Law Is

A power law is a relationship where one variable is proportional to another variable raised to a constant power.

In concrete terms: if you double the size of an event, the probability of an event that large is cut by a fixed ratio (say, by four). Double again, cut by four again.

For earthquake magnitudes (the Gutenberg-Richter Law): each unit increase in magnitude reduces frequency by a factor of ten. Magnitude-4 earthquakes happen ten times more often than magnitude-5. Magnitude-5 happens ten times more often than magnitude-6. The pattern continues down to the smallest tremors and up to the rare catastrophic quakes.

For wealth (Pareto's original observation): if you look at the wealthiest 20% of people, they hold 80% of wealth. Now look within that richest 20%. Again, the top 20% of the wealthy hold 80% of the wealth in that subgroup. The pattern repeats recursively at every scale.

For book sales: some books sell millions, most sell hundreds, a very few sell dozens or fewer. The distribution isn't peaked (like height) and it doesn't have a typical value. It has every scale, from tiny to massive, and the frequency follows a power law.

The Consequence: Infinite Variance

This is where power laws diverge radically from Gaussian distributions in practical terms.

For a Gaussian distribution, you can calculate the variance — the average squared distance from the mean. This number is finite. It tells you something about how spread out the distribution is. Standard deviation is the square root of variance. It's a meaningful summary.

For a power-law distribution, the variance is infinite.

What does that mean practically? It means: as you accumulate more data, the average deviation from the mean doesn't converge. It keeps growing. Adding one more observation — particularly an observation at the tail — can move the mean and variance dramatically.

Standard deviation and confidence intervals in the Gaussian sense do not exist for power-law quantities.

You can calculate the numbers. But the numbers are meaningless — they will shift dramatically each time a new tail observation arrives.

If you try to use standard statistical tools (mean, standard deviation, confidence intervals) on a power-law quantity, you're making a category error. The tools assume the distribution has certain properties that power-law distributions don't have.

Mandelbrot's Cotton Prices

Mandelbrot examined 100 years of cotton prices. The prices did not fit a Gaussian. The tails were far too fat.

Instead, he proposed they followed a Levy stable distribution — a distribution with infinite variance. Self-similar at different scales. Fractal.

The finance industry's response: ignore him. Gaussians were simpler, more tractable mathematically, easier to teach. Mandelbrot was correct but inconvenient.

Three decades of crashes, volatility spikes, and model failures later, traders and academics reluctantly acknowledged: the tails are fat. The distribution looks more like Mandelbrot described.

Even today, the field uses Gaussian models as the default, then patches them with "stress testing" to account for the fat tails that the models don't capture. It's intellectual dishonesty dressed up as risk management.

Mandelbrot's insight — that financial prices are fractal, self-similar across scales — is now widely accepted. But the consequence (abandoning Gaussian models for quantities with fat tails) has not been fully adopted. The frameworks persist because they're embedded in decades of textbooks, regulations, and institutional practice.

The Gutenberg-Richter Law: Power Laws in Earthquakes

Seismologists long ago discovered that earthquake magnitudes follow a power law.

For every unit increase in magnitude, the frequency decreases by a factor of 10. So:

The pattern holds because of the underlying physics of seismic energy. The magnitude scale is logarithmic; each step up represents roughly 32 times the energy release.

The consequence: there is no such thing as a "typical" earthquake.

You can't say "the average earthquake is magnitude 5 and most are within half a magnitude of that." That would be Gaussian thinking. In reality, a region might experience hundreds of magnitude-4 earthquakes and one catastrophic magnitude-8. Both are part of the same distribution.

If you were building earthquake policy and someone told you "average earthquakes are magnitude 5, so we're designing for that," you'd be planning for the wrong distribution. You'd be focusing on the frequent small events and ignoring the rare large events where most of the damage comes from.

Yet this is exactly what happens in other domains. People plan for the mean instead of acknowledging the power-law structure.

The Long Tail: Book Sales and Amazon

Before the internet, bookstores were constrained by shelf space. A typical brick-and-mortar bookstore carried maybe 10,000 titles. These were chosen for mass appeal — the books most people wanted.

Amazon and other online retailers operate with essentially unlimited shelf space. Amazon's catalog includes millions of titles.

Chris Anderson documented this in The Long Tail: while individual obscure books sell very few copies, the aggregate of millions of long-tail titles exceeds the sales of the top hits.

Book sales follow a power law: a few massive bestsellers, then a long tail of gradually declining sales. The distribution has no peak. It has every magnitude, from blockbusters to books that sell one or two copies.

The practical consequence: the old retail model (focus on the hits because shelf space is limited) was forced by constraint, not by the actual distribution. Online retail, freed from shelf-space constraint, can embrace the entire distribution and profit from the tail.

The same power-law structure appears in music (a few hits dominate streaming, but the long tail of niche music is enormous), movies (a few blockbusters, millions of niche films), and digital products of all kinds.

Zipf's Law: City Populations

What size is a "normal" city?

The question has no sensible answer. City sizes span from a few thousand people (small town) to 37 million (Tokyo). And the distribution of cities across this range follows a power law called Zipf's Law.

The pattern: the n-th largest city in a country has population roughly proportional to 1/n times the largest city.

So if the largest city has 10 million people: - The 2nd largest has roughly 5 million - The 3rd largest has roughly 3.3 million - The 10th largest has roughly 1 million - The 100th largest has roughly 100,000

This pattern holds not just in one country but globally, across different eras, and even for different categories (largest companies, biggest websites).

The practical consequence: there is no such thing as a "typical" city.

If you're designing urban policy and someone talks about "average cities" or "typical city infrastructure," they're applying Gaussian thinking to a power-law distribution. The policy will work well for certain scales and catastrophically for others, because the distribution has no central tendency.

The same applies to companies (no "typical" company size), websites (no "typical" site traffic), and social networks (no "typical" account size).

How Variance and Standard Deviation Become Useless

For a Gaussian distribution with mean 100 and standard deviation 10:

You can use the standard deviation to reason about how likely different values are.

For a power-law distribution with the same apparent mean and standard deviation (calculated from a sample):

The calculation of standard deviation is mathematically valid. But applying it to describe a power-law distribution is a category error. The standard deviation doesn't describe what it claims to describe.

The Practical Recognition

How do you know if you're dealing with a power-law distribution rather than Gaussian?

Ask: Can one observation dominate the total?

If one extreme observation can move the aggregate meaningfully, you're in a power-law domain. The usual statistical tools are inadequate. You need to think explicitly about the tail — what extreme values are possible, how often they arrive, and how they structure outcomes.

What to Do About It

If you're managing something in a power-law domain (financial portfolio, company, city, any large system):

First: Stop using means and standard deviations as your summary statistics.

They're technically computable but practically misleading. Instead, understand the distribution across its whole range.

Second: Expect one observation to dominate.

Design your systems with that expectation. If one tail event can ruin you, that's a fragility problem. If one tail event can fund everything else, that's a resilience opportunity.

Third: Recognize that history provides limited information about the tail.

The worst event in your dataset is not the worst event possible. When designing for safety (financial regulation, infrastructure, insurance), assume the tail is fatter than your data suggests.

The power law is not a detail. It's a fundamental property of how Extremistan is structured. Ignore it at your peril.