Cumulative Advantage & Power Laws

Line up the world’s cities by population, its books by copies sold, its websites by inbound links, its people by wealth, and you keep meeting the same eerie shape: a tiny handful of colossal winners, then a long, thinning tail of also-rans that never quite ends. Not “a bit above average” winners — winners hundreds or thousands of times bigger than the median. Tokyo dwarfs the thousandth-largest city. The best-selling novel outsells the thousandth by a factor you can’t fit on a normal chart. The top 1% of any of these holds a slice that would be flatly impossible if the world ran on bell curves. It doesn’t. It runs, over and over, on power laws — and this course is about the engine that manufactures them.

That engine has a name older than statistics: “to those who have, more shall be given.” Sociologists call it the Matthew effect; network scientists call it preferential attachment; the general principle is cumulative advantage — a reinforcing loop where having more of something (money, citations, followers, links, attention) makes you more likely to get even more. Success is not just rewarded; it is compounded into further success. Run that loop and small, often random, early differences don’t stay small — they snowball into vast, self-perpetuating gaps. The winner isn’t necessarily a thousand times better. They just got a little ahead early, and the loop did the rest.

This is an expert course sitting where probability meets science & engineering, and it leans on two models you’ve already built. From fat tails you bring the crucial fact that in some worlds the extremes dominate — the average is a lie, one outlier can outweigh everyone else, and planning for the “typical” case misses the point entirely. From feedback loops you bring the reinforcing loop itself — the “more begets more” spiral that runs backward on your intuitions. Cumulative advantage is precisely a reinforcing feedback loop acting on a quantity, and its fingerprint on the world is a fat-tailed, power-law distribution. If either model feels shaky, shore it up first; everything here composes them.

From there we build the whole model, floor to ceiling. We start with what a power law actually is — a scale-free distribution that looks the same lopsided shape at every zoom level, why it goes straight on a log–log plot, and why it’s a different animal from the bell curve. Then the heart of the course: the generative mechanism — preferential attachment and the Matthew effect, the loop that manufactures the whole distribution from nothing but “more begets more.” We meet its friendly everyday face, the 80/20 Pareto principle, and — more usefully — how to wield it: find the vital few, ignore the trivial many. We turn to winner-take-all and superstar markets, where a razor-thin edge in quality, plus scalability and network effects, pays off a hundred to one. We stare into the tyranny of the tail: why in these worlds the average is meaningless, the sample mean never settles, and a single event can outweigh all of recorded history — and why that makes outcomes far more about luck compounded than a proportional-reward intuition can accept. And we stay ruthlessly honest about where the model lies: the eyeball test on a log–log plot proves almost nothing (half the “power laws” in the wild are really log-normals), the winner’s dazzling résumé is inflated by survivorship and the Matthew effect, and the tail’s exponent — estimated from a pitiful handful of extreme points — is far shakier than the confident straight line suggests.

Throughout, you’ll have a cumulative-advantage engine to play with: a row of equal piles competing for tokens dropped one at a time. Turn the attachment strength down and luck keeps everyone roughly even; turn it up and you watch a runaway winner emerge from a trivial early lead, watch the pile chart grow one tower over a field of scraps, and flip to the log–log view to see the rank–size curve straighten into the tell-tale diagonal of a power law. Drag it until the punchline is in your hands: the same simple loop, run long enough, turns near-equality into near-monopoly — and it does so whether or not the winner ever deserved it. By the end you’ll stop asking the naïve question — is this gap proof that the winner is that much better? — and start asking the sharp one: is this a bell-curve world where the middle is the story, or a power-law world where a reinforcing loop already decided the winner, and the average is a distraction?