TOPIC 21 · PHASE D

Small World

What you lack is not more contacts, but a few shortcuts

2026-08-07 · Networks

Three sentences into a conversation with a stranger you discover a mutual friend, and you say "small world". Half of that is right: short paths really are everywhere — but the reason they exist has nothing to do with everyone being within reach, and knowing a path exists is not the same as being able to find it.

Last time we drew a system as a graph: nodes, edges, degree, clustering. This time we chase one oddity — real-world graphs have two properties that ought to rule each other out.

One is clustering: your friends mostly know each other. The other is closeness: pick any two people on Earth and the number of people in between is absurdly small. Clustering means ties are packed into small circles, which should make global distances huge. Closeness means ties are spread thin, which should destroy clustering. Reality has both. This issue asks how it gets away with that, what it costs, and what the claim has been misread as.

The division of labour with the next issue, stated up front: this one is about shortcuts — how a tiny number of cross-circle links crushes the distance across a whole graph. Topic 22 is about hubs — why some nodes carry thousands of edges. They are different things: in this issue's model everyone has roughly the same number of ties, there is not a single hub, and distances are short anyway.

01Six Degrees: A Misremembered Experiment

In 1967 the social psychologist Stanley Milgram did something wonderfully low-tech: he handed folders to randomly chosen people in Nebraska and Boston and asked them to get the folder to a stockbroker in Massachusetts whom they had never met. One rule — you may only mail it to someone you know on a first-name basis, who then mails it onward.

In the study Travers and Milgram published in 1969, 296 people started and 64 folders arrived. Those completed chains passed through 5.2 intermediaries on average. That "5.2" became "six degrees of separation".

Two things need saying. First, "six degrees of separation" is not Milgram's phrase — it comes from John Guare's 1990 play of that name. A literary phrase became, in the public mind, a scientific finding. Second, and this is the one that matters: 232 folders never arrived.

That is not a footnote about data quality. It changes what the number means. 5.2 is the average length of the chains that completed; chains that broke have no length at all and never enter the denominator. And the longer a chain, the more chances someone in the middle had to shrug and drop the folder in a drawer — failure and length are not independent, and failure is biased toward removing the long ones. From which a directional conclusion follows: the true average distance can only be larger than 5.2, never smaller.

The earliest run is more awkward still. In 2002 the education scholar Judith Kleinfeld went through Milgram's archived papers and found his first pilot, run in Kansas: 60 folders, 3 arrived — 5%. That study was never formally published; it surfaced in one popular-magazine article, and the line that got quoted was that one folder took only four days.

Travers–Milgram, 1969: who actually entered that average29621764296 started217 forwarded64 arrived−79−15379 never moved153 broke en routeaveraged over these 64 only5.2intermediaries (≈ 6.2 hops)the 232 broken chains have no lengthand longer chains break more often→ the true value can only be largerfailure correlates with the quantity being measured, so the direction of the bias is known
Two readings of the same experiment. 5.2 is the average over those 64 chains, not over 296.

So is "the world is small" wrong? No. Its evidence just arrived forty years late. In 2011 Facebook computed the shortest paths across its entire graph of 721 million active users: the average was 4.74 edges, i.e. 3.74 intermediaries. Recomputed on 1.59 billion users in 2016: 4.57 edges, 3.57 intermediaries.

Watch the units — Milgram's 5.2 counts intermediaries, which is 6.2 hops; Facebook's 4.74 counts hops. The two numbers only mean anything on the same ruler, and this is exactly where "six degrees" tends to get fudged.

🌀 Military statistics · Wald's bullet holes Asked in the Second World War where to armour bombers given the pattern of hits on returning aircraft, the statistician Abraham Wald pointed out that the places to armour were precisely the ones with no holes — planes hit there did not come back. Milgram's 5.2 runs on the same mechanism: only completed chains enter the denominator, and the probability of failure rises with the very quantity being measured. That yields something much harder than "the data is messy": the direction of the bias is known, so such an observation can be used directly as a lower bound rather than waiting for better data.

02Clustered and Close at the Same Time

Two rulers first, both from last issue. Average path length (L): pick any two nodes, count the hops along the shortest route, average over all pairs. Clustering coefficient (C): pick a node, look at what fraction of its friends are actually connected to each other — the quantified version of "do my friends know each other?"

The textbook graphs each take one extreme. A regular lattice: imagine everyone standing in one big ring, each linked only to two neighbours on either side. C is high (your neighbours are neighbours of each other) but L is enormous — a message has to walk half the ring to reach the far side. A random graph: all edges thrown at random. L is tiny, but C is nearly zero (the chance your two friends happen to know each other is the same as for two strangers).

The real world is neither. In 1998 Duncan Watts and Steven Strogatz did something economical: instead of inventing a new model, they put a knob between the two extremes.

The rule fits in one sentence: lay out the regular lattice, then walk through every edge and, with probability p, detach one end and reattach it to a random node. p = 0 is the lattice, p = 1 is the random graph, and nobody had looked carefully at the middle. → ref · the small-world model

Look at what happens in that middle and the whole thing becomes clear: L falls far faster than C. By the time p is one per cent, path length has already dropped close to random-graph territory while clustering has barely moved. That middle stretch — spanning two or three orders of magnitude in p — is the "small world".

One knob p: from regular lattice to random graphp = 0 (regular lattice)large L · large Cp ≈ 0.05 (small world)L already small · C barely movedp = 1 (random graph)small L · C gone toothe two copper chords are the rewired shortcuts — 2 edges out of 40small-world regime10⁻⁴10⁻³10⁻²10⁻¹1rewiring probability p (log) →L / L(0) path lengthC / C(0) clusteringratio to the p = 0 valueL falls far faster than C — the two-to-three-decade gap in between is where real networks sit
The curves follow the shape of the original Watts–Strogatz result (N ≈ 1000, 10 edges per node). What matters is not the exact values but that the two lines fall out of step.

Why out of step? Because the two things a shortcut does differ in magnitude. Its benefit to distance is regional: it welds together two stretches of the ring that were far apart, and every pair of nodes that needed to travel from one stretch to the other gets a shorter route. Its damage to clustering is pointwise: it destroys exactly one triangle. A whole region's worth of gain against one node's worth of loss — which is why the first few shortcuts are almost free.

The flip side is that returns fall off just as fast. The first shortcut turns "walk half the ring" into "one jump"; the hundredth shaves half a hop off a route that was already short. This is the most usable line of the issue: the marginal return on shortcuts collapses.

Watts and Strogatz checked three real networks: the film-actor collaboration graph, the western US power grid, and the nervous system of the nematode C. elegans. All three sit in that middle stretch — the actor graph's clustering coefficient is more than a hundred times that of a size-matched random graph, while its average path length is about the same as the random graph's.

🎯 DECISION LINE

To shorten a system's "distance" (information, approvals, incident recovery), buy one shortcut and then measure immediately. Do not add ten at once. Concretely: before adding the link, write down the length of the longest critical path; write it down again afterwards. Two additions with no change means stop — the rest of the edges are pure cost.

🌀 Engineering history · the first transatlantic cable Before the transatlantic telegraph cable worked reliably in 1866, a message from London to New York travelled by ship and took a week and a half; afterwards, minutes. Nothing laid in the following century and a half ever produced a change of that magnitude again — later cables added bandwidth, not diameter. That is the historical version of collapsing marginal returns on shortcuts, and it implies an unflattering periodisation: if you date eras by how much the world's diameter shrank, the information revolution begins the day the first line opened, not during the decades it spread.

03Bridges Must Be Weak Ties

Shortcuts are valuable, then. Which raises the next question: where do shortcuts grow? Are they choosy about location?

In 1973 the sociologist Mark Granovetter gave an answer, and derived it from pure logic. One definition first: if removing an edge forces the two people at its ends to take a long detour to reconnect, that edge is a bridge — which is what a shortcut is called in a social network.

Then the key observation, which he called the forbidden triad: suppose A–B is a strong tie and A–C is a strong tie. Strong means frequent contact, substantial time invested, mutual importance. Then B and C are almost bound to run into each other, and very likely to connect. So the triangle "A–B strong, A–C strong, B and C total strangers" is vanishingly rare in reality — that is what "forbidden" means.

Push it one step and the conclusion arrives: no strong tie can ever be a bridge. A strong tie is necessarily wrapped in closed triangles, so removing it still leaves a ring of mutual friends to detour through. What is left to serve as a bridge is only the weak ties.

Why shortcuts only grow on weak tiesthe forbidden triadABCB and C almost certainly know each other tooso removing A–B still leaves a detourstrong tietwo dense clustersthe only bridge is that thin lineremove it and the two sides really do spliteveryone inside knows everyonelogic on the left, consequence on the right: bridges must live where ties are sparse — geometry, not preference
"The strength of weak ties" is not a motivational slogan; it is a geometric consequence of triangle closure.

This is more than an elegant derivation. Granovetter interviewed 282 professionals who had recently changed jobs; among those who got the job through a personal contact, 16.7% saw that contact often, 55.6% occasionally, and 27.8% rarely. More than eight in ten opportunities came from people they were not in regular touch with.

The mechanism is plain: people you are with every day know largely what you know. The value of information lies in being non-redundant, and non-redundancy grows exactly where ties are sparse. Which is why "get to know a few more colleagues" usually does nothing — they share your information source.

But the line got flattened into "the weaker the better", and that is wrong; there is now hard evidence. In 2022 Rajkumar and colleagues published a rare causal experiment in Science: over five years LinkedIn ran repeated randomised trials on its "People You May Know" algorithm, covering more than 20 million people, some 2 billion new ties and 600,000 job changes. The result is an inverted U — weak beats strong, yes, but past a point the effect turns back down. The most useful ties are the moderately weak ones: people with a few connections in common who are not in your daily circle.

Not "weaker is better": the shape measured by a five-year randomised experiment on LinkedInmoderately weak: peak← very strong tiesvery weak ties →tie strength (measured by number of mutual connections)relative effect on landing a new jobtoo strong = redundant information; too weak = they neither know you nor have reason to pass it on (Rajkumar et al., 2022)
The curve's shape is taken from that experiment's finding (inverted U, peaking at moderately weak ties); the vertical axis is relative effect, not absolute probability.

The shape makes sense: too strong means redundant information; too weak means the other person neither knows you nor has any reason to take an extra step for you. A bridge has to bear load, and both ends must hold.

🎯 DECISION LINE

When you are looking for opportunities, do not deepen ties that are already close (no new information there) and do not chase total strangers (no load-bearing there). Concretely: go through the people you contacted once or twice in the past year, pick three, and send each a specific message — what you are working on, where you are stuck. The test is not reply rate; it is whether one of the three brings back a name or a fact you would not otherwise have had.

🌀 Sociology · Fei Xiaotong's "differential mode of association" From the Soil describes Chinese relational structure as ripples spreading outward from the self — tight near the centre, thinning outward. Read that picture through this section's mechanism and it yields a line Fei did not write: the inner rings, whose members all know one another, are heavily triangle-closed, so they contain no bridges at all. From which a concrete division follows — in such a structure "using contacts to get something done" works well (strong ties supply reliability and obligation) while "using contacts to hear about opportunities" works badly (the same circle knows the same things). Treating the two as one activity is the most common waste this structure produces.

04Existence Is Not Findability

Go back to Milgram's experiment and you will notice we have not yet mentioned the truly astonishing part.

That short paths exist is one thing. That ordinary people found them using only local information is another — and a far stronger claim. None of those 64 people had ever seen the global graph. Each knew only who they knew, then guessed which of those was closer to a stockbroker in Boston. Guessing all the way down, they got there.

In 2000 the computer scientist Jon Kleinberg proved that this is not a given at all. Most small-world networks cannot do it.

His setup is clean. Arrange nodes on a two-dimensional lattice (a chessboard where each square links to its four neighbours), then give every node one extra long-range edge to somewhere distant. The crucial part: how far away that edge lands follows a distribution in which probability drops with distance, and how fast it drops is set by an exponent α. Small α means shortcuts tend to be long; large α means they tend to be short.

The conclusion is one sentence and it is hard: only when α exactly equals the dimension of the lattice (here, 2) can greedy forwarding using only local information deliver in roughly (log N)² steps. For any other α the number of steps grows as a power of the network size N — meaning a slightly larger network is completely unusable.

Same number of shortcuts — only one arrangement is findableα = 0 (too small)all shortcuts are long jumpsnothing fine-grained nearbyα = 2 = lattice dimensionshortcuts at every scaleeach step cuts a fixed fractionα = 4 (too large)almost no long jumpsyou can only crawl cell by cellcopper dot is the source, coloured lines are its long-range links — identical edge countsonly α = 2 works01234exponent α (how fast shortcuts decay with distance)steps needed to deliver using local information (schematic)off the lattice dimension, delivery time grows as a power of network size — unusable at any real scale
All three panels have exactly the same number of shortcuts, and yet finding a route ranges from easy to impossible. The only difference is how shortcut lengths are distributed.

Why that particular value? Because when α equals the dimension, shortcuts are roughly equal in number at every scale: some jump a thousand cells, some a hundred, some ten, some three. So however far you still are from the target, there is always a shortcut of the right size to cut the remaining distance by a fixed fraction. With α too small, every shortcut is a huge jump, and once you land near the target there is no fine-grained shortcut left — you crawl. With α too large, there are no long jumps to begin with.

This step is worth pausing on, because it replaces a default assumption. We habitually treat usability as a property of the network. The conclusion here is: navigability is not a property of the network; it is a joint property of the network and the distance ruler in your hand. Same graph, different ruler, and it flips from usable to useless. Milgram's subjects used geography plus occupation as their ruler — that ruler is exactly what let them judge, at every step, whether they were getting closer.

🎯 DECISION LINE

For any process that depends on people passing things to people (referrals, internal introductions, finding someone in another department, incident escalation), the bottleneck is usually not how many relationships exist but whether each person can judge who is closer. So build the ruler before the network. Concretely: give such a process a locally comparable labelling dimension (which system module owns it, which industry the customer is in, what tier the problem is) instead of expanding the address book. The test: grab any intermediary and see whether they can say within three seconds "here is who around me is closer to this". If they cannot, adding people will not help.

🌀 Engineering history · postal addresses A postal worker knows none of the recipients, and the mail still arrives, because an address narrows the field at every level: country → city → street → number. That is Kleinberg's condition already implemented in a human system — a "closer than this" test available at every scale. Turn it around: replace addresses with random serial numbers and the same postal network, same carriers, same roads collapse instantly. Which gives an easily missed engineering conclusion: a lookup system's usability can be destroyed or rescued without touching a single connection — and a perfectly flat tag system, with no hierarchy and no notion of nearness, is that string of random numbers.

05Where This Breaks Down

"Small world" is probably the most freely applied term in complexity science. Here are four boundaries.

First, "six degrees" was never cleanly measured. Beyond the completion rate, Kleinfeld pointed out a problem with the target itself: a Boston stockbroker is high-status and occupationally conspicuous, about the easiest kind of person to guess your way toward, and the senders were not a random street sample either. Swap in a low-income person in a remote area and chains get much longer and break much more often. In popular culture six degrees is common knowledge; empirically it is an estimate that has never been fully verified.

Second, "this is a small-world network" carries almost no information. The test is "L close to a random graph while C is far above it", and sparse, clustered graphs are everywhere in reality — nearly all of them qualify. Saying a network "is small-world" is close to saying nothing. If you are going to say it, give numbers: how many times L and C are relative to a size-matched random graph. A small world without ratios is an adjective, not a finding.

Third, there are no hubs in this model. Rewiring does not change the distribution of how many edges each node has — everyone ends up with roughly the same number of ties (near-Poisson). So it cannot explain why some people are especially pivotal; that is a different mechanism, and it is next issue's subject. Using small worlds to explain star nodes is the wrong tool.

Fourth, and hardest: short paths do not mean things get through. In 2007 Damon Centola and Michael Macy distinguished two kinds of contagion. Simple contagion: one exposure suffices — a rumour, a link, a virus. Complex contagion: you only move after repeated exposure from several different people — adopting a new way of working, joining a risky action, using a technology nobody has validated.

For simple contagion a long bridge is an accelerator. For complex contagion it is often a dead end: it gives the person on the far side exactly one exposure, below threshold, so nothing happens — and that edge could have gone into thickening local connections instead. Centola and Macy's conclusion is that complex contagion feeds on the width of bridges (several parallel connections between two groups), not their length.

One graph, one shortcut — the only difference is how many exposures adoption needssimple contagion: one exposure is enoughseedlong bridgeafter two steps: the far side lights up toocomplex contagion: needs two sourcesseed (has to be a small cluster)long bridgethe far side gets 1 exposure → it stops heregreen = adopted; hollow = heard it but below threshold. A long bridge accelerates messages and kills behaviours
The same graph run twice; the only change is the adoption threshold. Topic 24 will take the two kinds of contagion apart in full.

This boundary changes tactics outright: "get a few more people from other circles to share it" works for messages and fails for behaviours — behaviours need several people inside one circle doing it at the same time.

🎯 DECISION LINE

Before you act, decide whether you are spreading information or behaviour. Information: buy long bridges across circles, the more scattered the better. Behaviour: do not scatter — lay three to five lines into a single small circle so each person sees the same thing from several different people. The test is one question: "how many people does the target need to hear this from before they actually move?" If the answer is greater than one, long bridges are not your tool.

🌀 History and social movements · Freedom Summer, 1964 Doug McAdam went through the application forms of the volunteers who signed up to register Black voters in Mississippi that year: 720 went, 241 applied and then withdrew. The two groups barely differed in stated conviction. They differed in ties — those who actually went listed two to three times as many activists and fellow volunteers among their strong ties as the withdrawals did. That is complex contagion running: the threshold for high-risk action is not "knowing this exists" but "several people I know are going". Which yields a concrete recruiting conclusion: low-risk things spread on breadth, high-risk things spread on density — and the resources should go in opposite directions.

🎒 In Context · BigCat

  1. Writing & this siteEvery finished issue goes, by habit, to the handful of people most likely to care. But those people mostly know each other — by the forbidden triad they form a closed triangle, and their feedback overlaps heavily. Three people flagging the same problem looks like three pieces of evidence and is one. The change: every issue, hand the first two paragraphs to one reader from outside the circle (someone who does none of this work) and ask a single question — "at this point, do you still want to keep reading?" The test: if the last five issues drew feedback from the same group, you did not receive five reviews; you received one, repeated five times.
  2. Engineering & systems designIn architecture reviews the discussion almost always lands on "is this interface designed right", and nobody asks "should this edge exist at all". But every new direct cross-module link is a permanent operational cost and a permanent fault-propagation path, and its return collapses fast — the first one turns a two-week wait into a day, the tenth saves a few minutes. The change: add a line to the review template — "cross-module links added: N; for each, which specific critical path it shortens, from what to what". Any line that cannot be filled in with numbers gets cut.
  3. ParentingActivities keep getting added, but nearly all of them sit in the same circle — same school, same neighbourhood, same set of parents. Those ties overlap, and what they can bring in runs dry quickly. By this issue's mechanism, what needs adding is not the number of activities but one circle with no overlap with the existing one — and one is enough, because the first shortcut carries the largest return. The test: ask one question — "when was the last time my child learned something new from someone nobody at school knows?" If nothing comes to mind, what is missing is not more hours of class.

🌀 Crossings · Where Else This Mechanism Runs

Going Deeper

If returns on shortcuts collapse so fast, why are cross-circle links still so scarce in real societies?

Because their costs do not collapse. Maintaining a cross-circle weak tie has no mutual friends to share the load — no triangle to remind, to nudge, to supply things to talk about — so it runs on unilateral effort, while a strong tie is held up by an entire triangle and renews itself almost automatically. Falling returns plus flat costs means a low stock. That also says where to push: to grow bridges, cut the maintenance cost (a fixed occasion, a fixed rhythm) rather than trying to raise motivation.

Is small-world structure good or bad for a system?

Depends what travels on it. Information, innovation, synchronisation: good. Viruses, financial contagion, cascading failure: the same shortcut accelerates those too. So "should we build this shortcut" has no general answer, only "for what runs on this channel, do you fear slow more than fast?" Topics 23 and 25 take this apart.

Do online connections count as Granovetter's weak ties?

Partly. They really are non-redundant and really are cheap. But Granovetter's weak ties are still reciprocal and carry a minimum of obligation — the other person is willing to spend a little effort on you. A one-way follow has neither reciprocity nor obligation, and calling it a weak tie overstates what it does. The right-hand half of that 2022 inverted U is very likely measuring exactly this.

If "this is a small world" carries almost no information, what is the concept still good for?

Its value is not classification but a comparable baseline. The useful question is not "is it a small world" but "what multiples of a size-matched random graph are its L and C, and which way has that position moved over the past two years?" Treated as a state variable rather than a label, it stays alive.

If navigability depends on the metric, can an organisation get faster by changing only its labelling system and not its structure?

Partly, and this is usually the cheapest first step. But be careful: the ruler has to correlate with the actual connections. Navigating by a taxonomy unrelated to real collaboration paths makes people confidently wrong at every step — which is worse than having no ruler at all.

Further Reading