The Engine: Preferential Attachment

In lesson 1 you learned to recognise a power law — the scale-free shape, the log–log line that goes straight, the menagerie of cities and words and wealth that all wear it. That was diagnosis: spotting the fingerprint. This lesson is the crime. Where does the shape come from? What machine, running in the background, keeps stamping the same lopsided pattern onto systems as different as scientific citations, Twitter followers, and the size of firms?

There is one dominant answer, and it’s the beating heart of this whole course: a reinforcing feedback loop in which having more of something makes you more likely to get even more of it. Turn that loop on, let it run, and a power law falls out the other end almost every time. Turn it off, and the same players produce a boring, even spread. The distribution isn’t decided by how good the players are. It’s decided by whether the loop is running.

One loop, three famous names

The mechanism has been discovered independently so many times that it collects names like a runaway pile collects tokens. Three of them matter, and they are the same loop viewed from three doorways.

The Matthew effect. Sociologist Robert Merton coined this in 1968, borrowing a line from the Gospel of Matthew: “for unto every one that hath shall be given… but from him that hath not shall be taken away.” Merton was studying science, and he noticed something unfair: when a famous scientist and an unknown one co-author a paper, the famous one soaks up nearly all the credit. Give a talk about the work, and the audience remembers the big name. Fame, in other words, is self-reinforcing — being known gets you known more. Merton generalised it: accumulated recognition attracts further recognition, so the eminent get more eminent and the obscure stay obscure, even for identical work.

Preferential attachment. The network scientists gave the loop a mechanical description. In the Barabási–Albert model of how networks grow, you add nodes one at a time, and each new node’s links attach to existing nodes with probability proportional to how many links each already has. A webpage with a thousand inbound links is a thousand times likelier to catch the next link than a page with one. Popular nodes get more popular by the simple arithmetic of being easy to find. The same idea appears decades earlier in the Yule–Simon process, built to explain why a few genera contain most of the species and a few words most of the usage — new arrivals attach preferentially to whatever is already abundant.

Cumulative advantage. The plain-language name, and the one Merton’s student Derek de Solla Price used for the citation version: an advantage, once gained, feeds back and compounds. More citations make a paper more visible, which earns more citations. More wealth earns more return, which is more wealth. More followers get recommended more, which wins more followers.

The mechanism, step by step

Strip away the vocabulary and the loop is embarrassingly simple. It has exactly three moving parts.

1. A quantity you can accumulate. Wealth, followers, citations, inbound links, plays, customers, reputation — anything that piles up and that you can have more or less of.
2. A feedback rule: more of it raises the RATE at which you get more. This is the whole engine. Your current stock isn’t just a score sitting there; it’s an input to how fast you gain. Ten thousand followers don’t merely mean you had reach — they mean the recommendation algorithm shows you to more people right now, so you gain followers faster than the account with a hundred.
3. Repetition over time. Run the rule again and again. Each pass, the leaders’ faster rate widens the gap, which raises their rate further, which widens the gap further.

That is a reinforcing feedback loop in the exact sense you met in the feedback-loops course — the same “output loops back and amplifies itself” structure as compound interest or a viral post — except here the thing being amplified is your share of a limited prize. And a reinforcing loop acting on a quantity, running long enough, does not produce a tidy bell curve. It produces a fat-tailed power law: the runaway leader out at the extreme, the long thin tail of also-rans. (That’s the shape from the fat-tails course, and cumulative advantage is one of the main machines that builds it.)

Why proportional growth gives you a power law

Here’s the intuition for why this loop, specifically, produces a power law and not some other lopsided shape — kept deliberately light, because the simulation below will do the real convincing.

The engine is proportional (or multiplicative) growth: each pile grows roughly in proportion to how big it already is. A pile with 200 tokens catches new tokens about twice as fast as a pile with 100 — not a fixed bonus, a percentage. When everyone grows by a percentage of their current size, the gaps between them don’t stay fixed; they multiply. That’s the same reason compound interest curves upward: proportional growth turns small differences into large ratios.

Add the second ingredient — new nodes keep entering (new accounts join, new papers get published, new pages appear) — and something clean happens to the overall distribution. Old, established nodes have had the longest time to compound their lead, so they sit at the giant end; nodes that entered recently haven’t had time, so they crowd the small end. Do the maths (Yule did, in 1925) and the resulting spread of sizes is scale-free: no characteristic “typical” size, and the probability of finding a node with k links falls off as a power of k. In prose you’ll see this written $P(k) \propto k^{-\gamma}$ — the fraction of nodes with size k drops in proportion to k raised to some negative power γ. Don’t memorise it; just hold the takeaway: proportional growth plus the steady arrival of newcomers is a power-law factory.

Set the strength and watch a winner get manufactured

Enough words — go make one. Below is the engine from lesson 0, and this time you’re the one turning the crank. Fourteen piles start exactly equal. Tokens drop one at a time, and each token picks a pile — but you set how strongly a pile’s current size boosts its odds of grabbing the next token. That slider is the feedback rule, live in your hands.

Two things to burn in, because they carry the rest of the course.

First, strength controls inequality, not the players. Nudge the slider up and the top pile’s share and the Gini coefficient climb in lockstep; slide it back down and inequality collapses toward zero — with the exact same piles. The spread of outcomes is generated by how strong the “more begets more” loop is, not by any pile being better than the others. There are no better piles; they’re identical. Sit with that.

Second — the deepest idea in the lesson — the winner is chosen early, and largely by luck.

The role of luck: the winner is picked early, almost at random

Reset the engine at high strength and run it a few times. A different pile wins nearly every time. Not the “best” pile — there is no best pile — but whichever one happened to grab a couple of early tokens before the field had spread out. That fleeting, coin-flip lead is all the loop needs. Once a pile is even slightly ahead, its higher rate does the rest, and the gap only widens. The loop is a lead amplifier: hand it any tiny head start, from any source including pure chance, and it will inflate that head start into a runaway victory.

This is the counterintuitive core of cumulative advantage, and it flips a deep intuition on its head. We look at the giant — the arena-filling musician, the billion-play song, the thousand-citation paper — and reason backwards: it’s enormous, so it must have started enormously good. But the loop manufactures giants out of near-equal starts. Two identical seeds, planted the same day, diverge wildly — and which one becomes the redwood was decided by an early gust of luck, then locked in and magnified by the loop. Rerun history and a different one wins.

Worked example: two accounts, one lucky nudge

Make it concrete. Two brand-new creators, Ada and Bela, post equally good videos on the same platform the same week. By any blind measure their content is a wash. On day one, Ada’s clip happens to get surfaced to a slightly larger test audience — a scheduling quirk, nothing about quality — and ends the day with a hair more engagement. The recommendation algorithm reads engagement as “show this to more people,” so it feeds Ada’s next clip to a bigger audience. That’s the loop: a lead in reach buys more reach.

Watch the gap widen round by round, each creator gaining roughly in proportion to their current reach:

Nothing about their talent changed across those rows. Ada didn’t get five times better; she got a lucky 10% nudge on day one, and the loop — reach raising the rate of reach — did the rest, stretching a 1.1× edge into a 5× gap. Give Bela the lucky day-one nudge instead and she runs away. The 5× winner was decided in round one, by chance, then manufactured by the loop.

The control condition: what happens with no loop

The cleanest way to prove it’s the loop doing the work — not hidden quality — is to switch the loop off and re-run. Imagine the same tokens dropped by equal random assignment: every token picks a pile uniformly at random, ignoring how big each pile already is. No proportional boost, no feedback, no Matthew effect. What do you get?

You get near-equality. With fourteen piles and thousands of tokens, each pile lands near the same size (piles-of-coins-flipped math: the counts cluster tightly around the average, the classic thin-tailed bell-curve behaviour). No runaway winner, no fat tail, no power law — a boring, egalitarian spread. Same players, same number of tokens, same randomness. The only thing removed was the “current size boosts your rate” rule. So the inequality was never coming from the piles. It was the loop, and only the loop.

Tie it to the latticework

You now have the generative half of the course locked in. Lesson 1 gave you the shape — the power law, the log–log line. This lesson gives you the engine that stamps it out: a reinforcing loop (from the feedback-loops course) acting on an accumulable quantity, producing a fat-tailed distribution (from the fat-tails course). The two ideas snap together — recognise the shape, and you should immediately go hunting for the “more begets more” loop that built it; find a strong reinforcing loop acting on a quantity, and you should expect a power law downstream.

Recap

Cumulative advantage, the Matthew effect, and preferential attachment are one reinforcing loop under three famous names: the more of an accumulable quantity you already have, the faster you get more of it. Its three parts are a quantity, a feedback rule where current level raises the growth rate, and repetition over time. Because the growth is proportional and newcomers keep arriving, the loop manufactures a fat-tailed power law — and because it amplifies whatever tiny lead appears first, the winner is chosen early and largely by luck, not merit. Switch the loop off (equal random assignment) and the same players produce a boring, near-equal bell curve: proof that the inequality is made by the loop, not the players.

Next up: lesson 3, The 80/20 Principle — the friendly, everyday face of the power law this engine builds, why “the vital few and the trivial many” keeps recurring, and how to actually put the rule to work (including where 80/20 sharpens into 90/10, or nests inside itself).