TOPIC 22 · PHASE D

Scale-Free Networks and Preferential Attachment

Rich get richer — but being rich is not yet being scale-free

2026-08-08 · Networks

Your phone holds a couple of hundred numbers, and nobody you know holds twenty thousand. Online, one account has twenty million followers and the next has twenty, and that is not an anomaly — it is the normal condition. Same quantity, "how many connections"; on one side it has a typical value, on the other it has none. The difference is not in the people. It is in the wiring rule.

Last issue's world had no celebrities: everyone held roughly the same number of ties, a few of them crossed between circles, and that was enough to make distances short. Real networks contain something extra — giants. One airport carries hundreds of routes, one page is pointed at by millions, one protein takes part in hundreds of reactions. These nodes are not statistical outliers; they are the standard furniture of such networks.

Here is the counterintuitive turn. Explaining giants requires no assumption that they are better, stronger, or saw something earlier. Two rules — dull to the point of being neutral — are enough. And the same two rules predict a second thing: giants will certainly appear, and which node becomes one is essentially unpredictable. Those two sentences have to be said separately. Said together, they become a success story.

The division of labour with the next issue, up front: this one is about how hubs grow and how strong the evidence for the claim actually is; Topic 23 is about what follows from having hubs (near-immunity to random failure, instant collapse under targeted attack). And the last section here is a rare instance of this site dismantling its own material — "scale-free" is the most heavily marketed word in complexity science, and its evidence is much weaker than the popular books suggest.

01Why Random Wiring Grows No Giants

To see how strange something is you first need a control. Around 1959 the mathematicians Erdős and Rényi studied the laziest network imaginable: take N nodes, flip a coin for every pair, and draw an edge on heads. This is the random graph.

The number of edges a node holds is its degree. In a random graph the degrees follow what is called a Poisson distribution, and the meaning of that term is plain: almost every node sits close to the mean, twice the mean is already rare, and three times the mean effectively does not occur. In everyday language — it has a typical value.

Height is like that. The mean is around 1.7 m, you have seen two metres, you will never see five. Wealth is not like that, and neither is follower count. So the question is concrete: what kind of wiring produces the picture on the right?

Same 40 nodes, same 40 edges — only the wiring rule differs(1) Wire at random(2) Wire by existing degreethis one nodemax degree 7 — 3.5x the meanmax degree 14 — 7.0x the meanNo node stands out on the left; on the right one node owns 18% of all edge-endsmean degree is 2.0 on both sides
Same 40 nodes, same 40 edges, same mean degree — only the rule for allocating new edges differs. The largest node on the left is 3.5× the mean; on the right it is 7×, and that single node holds 18% of all edge-ends in the network.

Note that both sides carry the same number of edges. That matters: giants are not the product of more resource. Hand out the same resource under a different allocation rule and no giant appears.

🎯 DECISION LINE

To judge whether a system will grow giants, do not look at how concentrated it is now — look at how the new connections are allocated. If new links land uniformly at random, no giants. If they land in proportion to what each node already holds, giants are guaranteed. This is directly measurable: take the connections created recently (new users, new citations, new dependencies, new orders), sort them by the recipient's existing stock, and see whether the slope is flat or positive.

🌀 Engineering history · Erlang's trunk formula In 1917 the Danish engineer A. K. Erlang gave telephone exchanges a formula for how many trunk lines are enough, and it governed capacity planning in telecoms for a century. Its premise is exactly this section's: calls arrive independently at random, so traffic has a typical value and provisioning to the mean plus a margin suffices. Which means every "size it to average load" engineering habit carries a random-graph assumption inside it; the moment arrivals start triggering one another (a national broadcast, an earthquake, a ticket sale), the typical value is gone, and capacity computed from the mean fails precisely when it is needed. Capacity planning should not ask "how much on average" but "can one arrival cause the next".

02Two Rules Are Enough

In 1999 Barabási and Albert pointed out two things the random graph leaves out — and adding just those two is enough.

First, the network grows. The random graph assumes every node is present from the start and merely unconnected. Real networks are not built that way: pages are created one at a time, papers written one at a time, people register one at a time. Second, arrivals do not choose uniformly. A newcomer is more likely to attach to nodes that already hold many edges. That second rule is preferential attachment — "preferential" describes a bias in probability, not a preference in anyone's mind.

That deserves a sentence more, because it is the part most often misread as psychology. There is no intention in the rule: when you reach for a page to cite, you are more likely to land on one that many others have already cited simply because it is easier for you to come across. Preference is the name of the outcome, not the name of its cause. → ref · the preferential attachment model

Growth: every new node brings 2 edges and picks by existing degree(1) 6 nodesmax k = 4(2) 12 nodesmax k = 6(3) 24 nodesmax k = 9the earliest arrivals stay ahead in degreeP(new node picks an old node) is proportional to that node's current degree
One rule, three stages. Each new node brings 2 edges and picks targets with probability proportional to their current degree. The earliest arrivals pull further and further ahead — and no step anywhere used "who is better".

The degree distribution that comes out is a power law: the number of nodes holding k edges is proportional to k to the minus three. That 3 is derived from the rules, not fitted to data — which will matter later.

The exact meaning of scale-free here is this: the distribution has no typical value, so there is no "usual number of connections" that represents the network. It is not the mystical claim that there is no scale; it is the concrete statistical property of having no characteristic scale.

Both rules are needed, and this gets skipped. Growth with uniform attachment gives an exponential distribution and no hubs. Preferential attachment on a fixed set of nodes — no growth, just edges piling on — never settles: the distribution is not stationary, edges keep accumulating, and the network ends up connected to everything, with the power law only a phase it passes through. What does the work is the combination: a growing system allocating its increments in proportion to stock, not either piece alone. The same skeleton has an older name in probability theory → ref · the Pólya urn.

🌀 Biology · gene duplication makes the preference a by-product Nobody in a protein interaction network is chasing star proteins, and yet it grows a heavy tail. The mechanism: genes are occasionally duplicated wholesale, and the copy arrives wired to the same partners as the original — so the more partners a protein has, the more likely it is to be a neighbour of whichever gene got duplicated, and to gain a link for free, with a probability exactly proportional to its existing degree. That is the duplication–divergence model of Solé, Vázquez and others. It imposes a hard limit: observing the statistical signature of preferential attachment does not license inferring that anyone prefers anything, because "allocate by stock" can be entirely a by-product of copying.

03What Arriving Early Buys You

This model has one great virtue: it can be solved. The i-th node to arrive holds, by the time the network has grown to t nodes, a degree of kᵢ(t) = m·√(t/tᵢ) — where m is the number of edges each newcomer brings and tᵢ is when that node entered.

Read the formula again: degree depends only on when you arrived. Every node in the model is identical; arrival order is the only difference, and time amplifies it as a square root.

In this model the only difference between nodes is when they arrivednodes added so far (time) →degree karrived 4th → k = 20arrived 40th → k = 6arrived 160th → k = 3a fitter latecomer(Bianconi–Barabási)ki(t) = m·√(t/ti) — solid curves are the exact BA solution; the dashed one is schematic
The solid curves are the model's exact solution. They never cross — in this model a late node cannot catch an early one by any means. The dashed curve is the patch discussed below: give nodes a "fitness" and a latecomer can overtake.

This is what first-mover advantage means under this mechanism, and it is both stronger and weaker than the usual claim. Stronger, because the early node need not do anything right — pure arrival order suffices. Weaker, because it holds only under the model's assumptions, and those break quickly.

The obvious counterexample is search. Google entered in 1998, long after the link structure of the web had formed, and by the formula above should never have caught up. Bianconi and Barabási supplied the missing piece in 2001: give every node a fitness η, and allocate new edges in proportion to η times degree. Now a young but fit node really can overtake; and under some fitness distributions the system enters a winner-takes-all phase whose equations are those of Bose–Einstein condensation in physics, with the fittest node taking a finite fraction of all edges in the network.

So does success actually breed success in real systems? In 2014 van de Rijt and colleagues ran a fine field experiment on Kickstarter, Wikipedia, Epinions and Change.org: pick recipients at random, give each a small initial success — a modest pledge, an award, an upvote, a signature — then do nothing further and watch. On all four platforms the randomly favoured group went on to succeed at significantly higher rates: on Kickstarter 70% of the projects given one random pledge attracted further funding, against 39% of the controls. But the effect showed diminishing returns rather than unbounded amplification. Reality, in other words, looks more like the version with fitness than like pure preferential attachment.

🎯 DECISION LINE

A small head start early really is worth betting heavily on — but only while two conditions hold together. (1) The system is still growing: stop the growth and preferential attachment stops with it, freezing shares where they are instead of widening them. (2) New connections really are allocated by stock — measure this, don't assume it. Concretely: replace "cumulative share" on the dashboard with "share of the most recent period's new connections". The moment those two curves diverge is the moment fitness has overtaken stock, and the cumulative curve will not show it for several periods.

🌀 Art and culture · eight parallel charts In 2006 Salganik, Dodds and Watts had 14,341 participants listen to 48 unknown songs. Some chose with nothing but song titles to go on; the rest were randomly assigned to eight sealed parallel worlds, each showing the download counts accumulated within that world only. Result: the worlds that showed social information were more unequal at the top — and each of the eight crowned a different winner, with one song, "Lockdown", ranking 1st in one world and 40th in another. That is this section's mechanism run as an experiment: cumulative advantage amplifies whichever clicks happened first. From which follows an unkind conclusion about post-mortems: in such markets "there will be a top" is reproducible while "who is at the top" is not — so analysing what the winner did right is systematically an exercise in explaining noise.

04How Scale-Free Got Oversold

Now the provenance of this idea, and the beating it later took. Neither appears often in the popular accounts.

Provenance first. "Rich get richer" had been independently discovered at least three times before 1999. The statistician Yule used it in 1925 to explain why some biological genera contain hundreds of species and others one. Simon wrote the same skeleton in general form in 1955, applying it to word frequency, income and city size. And Price moved it explicitly onto networks in 1976 — he was studying citations between papers, and he called the mechanism cumulative advantage.

The mechanism predates the name by seventy years1925Yulewhy some genera are huge1955Simonword frequency, income, city size1976Pricecitations: cumulative advantage1999Barabási–Albertnamed: scale-free networks2019Broido–Clauset / Voitalovsystematic tests, and the rebuttalBroido & Clauset 2019: one statistical pipeline applied to 927 real networksstrongest category4%at least the weakest category57%social networks with no evidence at all50%On the same data, log-normal fits most of these networks as well as a power law or better
Above: the mechanism was found seventy years before it was named. Below: the 2019 systematic test — one statistical pipeline run over 927 real networks.

This is not trivia. A mechanism rediscovered repeatedly is usually a basic one; but it also means what was genuinely new in 1999 was attaching it to networks as an object and giving it a marketable name — not "rich get richer" itself. And it was the name that went on to carry far more weight than the evidence could bear.

Now the beating. In 2019 Broido and Clauset did something dull and useful: they ran one strict statistical procedure over all 927 real networks they could assemble — social, biological, technological, transport, informational, no cherry-picking. Sorting them into five levels of evidence, they found that only 4% reached the strongest level; 57% reached at least one level; and half of all social networks failed to reach even the weakest. The most awkward finding: for most of these networks a log-normal distribution fits as well as a power law, or better.

The paper drew an immediate rebuttal. Voitalov and colleagues published "Scale-free networks well done" the same year, arguing the definition was too strict: if a power law is defined as statisticians standardly do — as a regularly varying distribution, constraining only the tail and letting the low-degree region deviate freely — then scale-free networks are not rare at all.

So the honest state of play is this. Almost nobody disputes that real degree distributions are heavy-tailed — the large nodes are far larger than a random graph permits. What is disputed is how "power law" should be defined and where the evidential bar sits. The popular line, "real networks are scale-free", merges two claims of very different strength into one.

🎯 DECISION LINE

Never hang a decision on the numeric value of the exponent — it is sensitive to the lower cutoff and to the observation window, and two analysts can differ by 0.5 on the same data. Report two quantities that are insensitive to the distributional assumption instead: the maximum degree as a multiple of the median degree, and the share of all connections held by the top 1% of nodes. Both map directly onto what you actually need to do — who to watch, whose loss would hurt — and anyone can recheck them.

🌀 History of economics · what Pareto inferred from a curve In 1896 Pareto found that income distributions in several countries shared the same heavy-tailed shape, and from that cross-national regularity he concluded that inequality is a law of nature and redistribution futile. This section's mechanism shows the step is far too fast: the same distribution can be produced by entirely different mobility structures — a society reshuffled every generation and a hereditary one can have identical income distributions, because a distribution records the shape at this instant and not how long anyone has occupied a position. So to judge whether a society has set, neither the Gini coefficient nor the exponent will tell you; you need the intergenerational transition matrix. Shape carries no mobility information — that is an intrinsic limit of distributions, not a data-quality problem.

05Where This Breaks Down

The previous section was about strength of evidence. This one is about the boundary of the mechanism itself — when preferential attachment gives you no power law at all, and when the heavy tail you observe is not a property of the system.

First, "allocate by stock" has to be strictly proportional, and near enough is not good enough. In 2000 Krapivsky, Redner and Leyvraz worked out the general case: if new edges are allocated in proportion to degree raised to the power α, only α = 1 — strict proportionality — yields a power law. Slightly below 1, the hubs are flattened and the distribution becomes a stretched exponential; above 1, links collapse onto a handful of nodes, and further up almost every edge ends at a single one. It is a knife-edge condition. And in 2003 Jeong, Néda and Barabási measured the actual allocation of new edges in four real networks: citations and the internet came out close to linear, while actor collaboration and scientific co-authorship were sublinear. So "preferential attachment is running here" and "a power law should appear here" are two different statements.

Second, your measurement may manufacture the heavy tail by itself. This one is brutal. Internet topology is among the most famous scale-free cases, and its early evidence came from traceroute — start at one source, record the shortest paths to many destinations, stitch the visible edges into a graph. In 2005 Achlioptas, Clauset, Kempe and Moore proved that this sampling method yields an observed power law even when every node in the underlying graph has exactly the same degree.

Every node below has exactly 4 edges — but not as seen from one sourceThe real graph: every degree is 4sourcegold = the 13 edges traceroute sees, out of 28One graph, two ways of counting degree615223141456truth: all 14 nodes have degree 4as sampled: a few high, most lowdegree kOne sampling pass turns "all identical" into "a few high, most low" — a heavy tail can be pure measurement
Every node in the graph below holds exactly 4 edges; there is no hub anywhere. But sample it as shortest paths from one source, see only 13 of the 28 edges, and the counted degrees turn into "a few high, most low". The real result comes from analysis on much larger graphs — this only draws the mechanism.

Third, the things the BA model explicitly does not predict. The networks it generates have almost no clustering — the chance that two of your neighbours know each other approaches zero; no community structure; nodes that never die; edges that never rewire. Real networks violate all four. So using this model for any question that turns on clustering, communities or exit is using the wrong tool, however well its degree distribution fits.

Fourth, inferring mechanism from shape means inverting a many-to-one map. That case was made last issue → ref · identifying and misidentifying power laws; here is the network-specific roster: preferential attachment, fitness, duplication–divergence, random proportional growth, and the pure sampling bias above all deliver the same tail. To claim which one is running, go and measure some other observable consequence of it — for instance measure the allocation of new edges directly, instead of only glancing at the degree distribution.

Fifth, and most worth keeping: scale-free is not a medal. It does not mean the system is more efficient, more advanced or more "organic", nor that someone designed it better. It is a statement about the shape of a degree distribution, and a fairly weak one. Much of the reason the word travelled so far is that it sounds like it is saying something profound.

🌀 History of science · the canals of Mars In 1877 Schiaparelli reported faint linear markings on the Martian surface; Lowell went on to draw an entire canal network and infer a civilisation moving water. The eventual explanation: near the limit of resolution, vision joins a string of discrete specks into lines — the structure was generated by the method of observation, not discovered by it. That is the same thing as this section's second point, and it yields the same test: to decide whether a structure is real, don't ask how clearly you see it, ask whether it survives a different sampling method. The canals were dissolved by better telescopes; the internet's power law has to be checked against a different way of mapping topology.

🎒 Scenarios · BigCat

  1. Investing and position sizingEvery so often you check whether you are diversified enough, and the check is always a count of positions. This issue's mechanism says hubs are the necessary output of the rule, not an accident of one industry — so in a network grown by preferential attachment, most nodes hang off a handful of hubs. Ten holdings that look unrelated may well share one cloud provider, one distribution channel, one pool of passive money. The denominator of diversification is not the number of positions but the number of distinct hubs left after deduplication. What to change: for each holding write down the three external nodes it depends on (whoever going dark would hurt it), deduplicate, and count — that number is your actual diversification, and it is often frighteningly small. Stop doing: treating position count as a diversification metric.
  2. Leading a teamEvery team eventually grows one person who is "fastest to ask", through whom all information passes. That is usually not a difference in ability but preferential attachment: they know more → everyone asks them → they know more still, widening as a square root. Telling people to stop asking them will not work — persuasion pushes on character, while what is running here is an allocation rule. What can be changed is the rule: write down an owner for each area, make the entry point a document rather than a person, so new questions land by ownership rather than by who already knows most. A measurable test: the fraction of all questions that go directly to people past their 90th day. If it stays near zero, the problem is not their capability, it is that the rule has not changed. Stop doing: treating "who gets asked most" as performance evidence.
  3. Practice and mindThe usual way to stop a recurring thought is to watch for it — "notice when it comes up and cut it off". The trouble is that monitoring is itself an access, and every retrieval makes the next retrieval easier (memory research calls this retrieval practice). That is preferential attachment applied to attention: what is accessed most becomes easiest to access, and you are adding the edges yourself. What can be changed: replace suppression with adding edges elsewhere — anchor attention on something low-stimulus and sustainable (counting the breath is doing exactly this mechanically), so that the increment of attention goes somewhere else rather than into auditing the old target. A useful test: if a method requires you to keep checking whether a particular thought is present, it is adding edges to that thought.

🌀 Crossings

Going Deeper

If preferential attachment needs nobody's preference, what explanatory work is left for "brand" and "network effects"?

What is left is the part that can be measured separately. Preferential attachment predicts that the allocation of new links depends on stock alone, independent of any intrinsic property of the target. Brand and network effects claim there is a premium that does not vary with stock — which is precisely the η of the fitness model. To separate them, compare the new links won by two targets holding equal stock but differing in the property; if there is no systematic difference, the much-discussed "brand" may just be another name for stock.

When growth stops, does inequality freeze or keep widening?

Under pure preferential attachment it freezes: with no new nodes arriving there are no new edges to allocate, and shares stay where they are. That yields something counterintuitive — in a market that has stopped growing, the penalty for arriving late stops increasing. But mind the premise: in reality, once growth stops the existing edges usually start being reallocated (customers churn, citations age, dependencies get replaced), and that is a different dynamic on which this issue's model is entirely silent.

If a heavy tail can be manufactured by the measurement, how do you design a measurement that could falsify itself?

One workable answer: change the most suspicious step and see whether the conclusion survives. Traceroute's problem is "looking out from a handful of sources", so add sources, or read topology from routing tables instead. More generally: write down the sentence "if this tail is an artefact of sampling, then changing the sampling to X should make it disappear", then go and do X. If you cannot write that sentence, the measurement is not falsifiable.

If real networks are log-normal rather than power-law, how much of next issue's "robust yet fragile" survives?

More than you would expect. Those conclusions really depend on there being a few nodes with degrees far above the rest, not on the distribution obeying one particular formula — a log-normal tail is heavy enough to have hubs too. What does collapse are the quantitative results that depend on the exponent's value, such as the claim that the percolation threshold under random failure tends to zero, which holds only for γ < 3.

Further Reading