Power Laws & Fat Tails

In 1999, Albert-László Barabási and Réka Albert published Emergence of Scaling in Random Networks, showing that many real-world networks — the World Wide Web, scientific citation graphs, biological networks, the internet itself — share a remarkable property: their connectivity distributions follow power laws. A small number of hub nodes have enormously more connections than the median; the median is far below the average. The same shape recurs everywhere it is looked for. City sizes (Zipf's law). Wealth distributions in their upper tail (Pareto, 1896). Earthquake magnitudes (Gutenberg-Richter). Word frequencies in any natural-language text (Zipf again). Solar flares. Citation counts. Internet outage durations. The bell curve is the wrong intuition for most of these.

A power-law distribution has the form P(x) ∝ x^(−α) — the probability of an outcome of size x falls off polynomially. The defining consequences: no characteristic scale (zoom in or out and the distribution looks the same), fat tails (extreme outcomes are much more likely than Gaussian intuition predicts), and mean-dominating events (the average can be controlled by a single rare observation, which means the average itself is unstable). Standard statistical tools assuming normality fail in this regime. The 2008 financial crisis was, in part, a fat-tail failure: standard risk models priced extreme events as if returns were Gaussian; they were not. Important caveat from Clauset, Shalizi, and Newman (2009): when many claimed power-law distributions in real data are properly tested, most are better fit by log-normal or stretched-exponential distributions than by pure power laws. The qualitative fat-tailed claim usually survives; the strict power-law claim usually doesn't. The polymath-accurate vocabulary is fat-tailed rather than power-law. Mechanisms that generate fat tails include preferential attachment (Barabási's mechanism: new nodes connect to popular nodes), multiplicative growth (compounding produces log-normal distributions), and self-organized criticality (Bak's sandpile, applied to earthquakes, neural avalanches, market crashes).