Ideas are not viruses — don't copy the virus playbook
2026-08-10 · Networks
"How many people does one patient infect on average?" — not one term in that number belongs to the virus alone. It is the product of the pathogen, your behaviour, and the network you happen to sit in. Which is why the same disease can be two different diseases in two cities.
A rumour goes round the office, an app is suddenly on everyone's phone, a flu season, a clip that won't stop resurfacing. We call all of it "spreading", and we borrow the word viral for all of it. That word comes from epidemiology — and the borrowing leaves the most important premise behind.
The premise left behind is this: a virus only needs to touch you once. Anything that asks you to change what you do — switch tools, drop a habit, be the first to say the plan is wrong — usually needs several different people to do it in front of you first. That one-word difference flips the whole playbook: the structure that makes viruses fly (long-range shortcuts, wide weak ties) is exactly the structure that stalls behaviour.
The six network topics divide the ground: T20 is structure itself (degree, paths, clustering), T21 is shortcuts and small worlds, T22 is how hubs grow, T23 is what that structure resists and what breaks it, T25 is cascading failure through load redistribution. This one asks a single question: once something is running across the network, what decides whether it stops?
In 1927 two Scottish researchers, Kermack and McKendrick, took a shortcut that turned out to be the whole field: track nobody in particular, and sort the population into three piles instead. → ref · the SIR model
Pile one is the susceptible: never had it, will catch it on contact. Pile two is the infectious: currently passing it on. Pile three is the removed: recovered or dead, either way out of the game. Their initials give the model its name — SIR. There are two rules: a susceptible who meets an infectious becomes infectious with some probability, and an infectious eventually becomes removed.
Out of those two rules falls one number, the basic reproduction number, written R₀: drop a single infectious person into a population where everyone is still susceptible, and count how many they infect on average. Above 1 it grows, below 1 it dies out on its own. The line is almost suspiciously clean.
What deserves a pause is what R₀ is made of. It is a product of three things: how many people you contact per unit time × the chance a contact transmits × how long you stay infectious. Only the middle term belongs mostly to the pathogen; the other two are behaviour and environment. So "this disease has an R₀ of 3" is an incomplete sentence — it needs "where, when, and among whom".
One number gets used even more than R₀: the effective reproduction number, R_eff — how many one infectious person is passing it to right now. It equals R₀ times the fraction of people still susceptible. That sentence looks unremarkable, and it is the seed of everything below: spreading eats its own fuel. Every person infected is a person removed from the susceptible pile, and the same virus with the same behaviour starts running slower.
"Reduce transmission" is not one action, it is three separately priced ones. Split R₀ into contact rate, per-contact probability and infectious period, and ask of each: what does cutting this one by a third cost? Masks and encryption cut the middle term; capacity limits and isolation cut the first; fast detection and fast takedown cut the third. Get all three price lists on the table before choosing — "we must control the spread" has no units.
If R_eff = R₀ × (fraction still susceptible), then once that fraction drops below 1/R₀, R_eff falls under 1 and the outbreak contracts. That gate has a name — the herd immunity threshold, equal to 1 − 1/R₀. At R₀ = 3 it is 67%.
This is exactly what Kermack and McKendrick set out to explain: why does an epidemic recede while a large crowd of never-infected people is still standing there? The fashionable answer at the time was that the pathogen weakened. They showed you need no such assumption — burn part of the fuel and the fire shrinks by itself.
But there is a turn here that nearly everyone drops. The epidemic does not end when the threshold is crossed. Crossing it only means new cases start falling. A large number of people are still mid-infection, and they do not stop breathing because a statistical line was crossed. They keep transmitting, susceptibles keep falling, and the system sails well past the threshold before it truly halts. That extra stretch is overshoot.
The numbers are worse than the intuition: at R₀ = 3 the threshold is 67%, but letting it burn out unmitigated infects about 94% of the population. All 27 extra percentage points are overshoot. That is the arithmetic problem with "let it reach herd immunity naturally" — it treats the threshold as the finish line when it is only the turning point.
Overshoot also settles a very practical argument: when to let go is a real question. Lift an intervention exactly when R_eff has just crossed 1 and the susceptible share is still hugging the threshold — the moment it lifts, R_eff bounces back above 1 and the whole thing restarts. Any "stop at the target" plan needs headroom for overshoot, which means the exit condition should never be a date; it should be a stock level held below a line for a stated duration.
Both sections above smuggled in an assumption called homogeneous mixing: everyone is equally likely to meet everyone, like a well-stirred soup. It makes the maths beautiful, and it puts the conclusions some distance from reality.
People are not soup. Some meet two hundred people a day; some meet five a week. T22 covered why contact degree distributions are so often heavy-tailed → ref · spotting and misreading power laws. Once degrees are uneven, the quantity that decides whether something takes off changes: it is no longer the mean degree ⟨k⟩ but ⟨k²⟩ / ⟨k⟩ — the mean of the square. Squaring is brutally sensitive to the tail: someone with 200 contacts weighs four hundred times as much as someone with ten.
In 2001 Pastor-Satorras and Vespignani carried this to its conclusion, and it startled the field: if the degree distribution is a power law with exponent at or below 3, then ⟨k²⟩ diverges as the network grows and the epidemic threshold goes to zero. There is no safe band of "too weakly transmissible to take off". Any transmissibility at all can establish itself.
There is a layer beyond "some people have more contacts": the distribution of secondary infections itself is heavy-tailed. The term is overdispersion. SARS-CoV-2's dispersion parameter was estimated around 0.1, meaning roughly 10% of cases produced about 80% of transmission. So the person described by "three infections on average" may not exist at all — most infect nobody, a few infect dozens.
That changes the shape of intervention outright. Under homogeneous mixing, an intervention must be universal to work, because every person contributes equally. Under overdispersion, transmission concentrates in a few settings — enclosed, poorly ventilated, long duration, loud — so setting-targeted measures are absurdly cheap: a tiny footprint buys most of the mass in the tail.
Whenever a spreading metric is quoted "per person", ask about the shape of its distribution first. If secondary transmission is heavy-tailed, the mean carries almost no information and the thing to measure is the tail: what fraction of events produced what fraction of the consequences. The action changes to match — stop shaving everyone uniformly and instead enumerate the high-transmission settings and handle them one by one. Blunt test: sort past events by consequence; if the top 10% account for more than half the total, uniform intervention is buying you the wrong thing.
Here is the most useful distinction in this topic. A virus needs one contact: you meet it, you catch it with some probability, and how many carriers you met last week is irrelevant. That is simple contagion.
A great deal of social behaviour is not like that. Quitting to start something; abandoning a tool you've used for ten years; being the first in the room to say the plan is wrong — the bar for these is several different people around me already did it. One person doing it is noise; three unrelated people doing it is signal. That is complex contagion, named by Centola and Macy in 2007 and prefigured by Granovetter's 1978 threshold model: everyone carries a private number — "I'll do it once n others have" — and those numbers vary across a population.
The point is that this distinction inverts the value of network structure. T21 showed that a handful of long-range shortcuts collapses a network's diameter → ref · the small-world model, which is wonderful for simple contagion: one shortcut suffices to carry a spark across. But complex contagion is not short of distance, it is short of multiple independent signals landing on the same person. A shortcut delivers exactly one and then it is spent; a dense circle of mutual acquaintances delivers the same thing from three directions, which is precisely what a threshold requires. The structure that makes viruses fly is the structure that stalls behaviour.
Centola turned this into an experiment in 2010. Over 1,500 participants were randomly assigned to one of two artificial social networks: a clustered lattice (your neighbours are also each other's neighbours, so ties are highly redundant) or a random network (long ties, short paths). The behaviour to spread was registering for a health forum. The result ran against small-world intuition: the clustered network spread it faster and further — 54% adoption versus 38%, at roughly four times the speed.
Before funding any "rollout", settle one yes-or-no question: is one exposure enough? Yes (see it and you'll share it, install it and it works) → buy reach: shortcuts, cross-cutting ties, headcount covered. No (it demands a changed practice or a taken risk) → buy density: saturate one small circle, then move to the next, and never spread thin. The two spend in opposite directions; splitting the difference lands you short on both. The check is concrete: ask people who already adopted, "how many different people did you hear it from?" Mostly one — you have a simple contagion. Mostly two or three — you have a complex one, and every dollar priced per head reached is being wasted.
Contagion models are the most heavily borrowed toolkit in complexity science, and therefore the most frequently broken in transit. Here is where they come apart.
One: R₀ is not a constant you can look up. Measles has long been quoted at 12–18, nearly as common knowledge. In 2017 Guerra and colleagues ran a systematic review, gathering 58 estimates from 18 studies, and concluded that the actual spread of estimates is far wider than that familiar range, and strongly dependent on population structure, era and estimation method. No surprise — two of R₀'s three factors were behaviour and environment all along. Quoting an R₀ without saying which network and which behaviours produced it is quoting a number with no units.
Two: the compartmental assumptions are load-bearing, not decorative. The model requires homogeneous mixing (dismantled in section 3), discrete irreversible transitions (catch it once, immune for life), and a population whose composition doesn't shift over the period. Real pathogens have latent periods (add an E compartment), reinfection, mutation, seasonality. The model gives you a skeleton, not a body.
Three: applied to ideas, what is missing is a definition of infection. When exactly has someone "accepted" an idea? On hearing it? On agreeing? On changing behaviour? Without a decidable moment of transition, R₀ cannot be measured and the whole apparatus degrades into metaphor. Memetics ran aground precisely here — it treated ideas as self-replicating units, yet decades on, "replication" still has no observable counterpart; the Journal of Memetics ceased publication in 2005, and what it lacked was never data.
Four: complex contagion has its own boundary. Centola's elegant experiment ran on networks the researchers built: assignment was randomised, the behaviour was a single one (registering for a health forum), and participants were strangers. It demonstrated forcefully that social reinforcement is real; it did not demonstrate that all social spreading is complex contagion. Recent work also identifies a trade-off between reach and reinforcement: too clustered and reinforcement is ample but reach is slow; too random and reach is fast but reinforcement is thin. The optimum sits between them, which means "build denser small circles" is a prescription with a dose, not a direction.
Before using any spreading model, write down three things: ① which observable event counts as your infection (a download? a payment? two straight weeks of use?); ② who the edges of your network actually connect; ③ whether one exposure is enough. If any of the three won't fill in, don't report an R₀ — report something you can actually measure, such as the distribution of "how many distinct sources had this new user heard of us from". A crude number you can fill in beats an elegant parameter you cannot.
Most real things are mixed: hearing about it is simple contagion, adopting it is complex. The interesting part is that the two layers have opposite optimal structures — the awareness layer wants shortcuts, the adoption layer wants density. That suggests a division of labour: use reach to build a stock of "have heard of it", then use density to convert it locally. Worth pressing on: can saturating the awareness layer actually raise the adoption threshold — everyone has heard of it, nobody has gone first, and waiting itself becomes a signal?
Intuition says an intervention only delays. Overshoot gives the other answer: intervention lowers R_eff, and final size is a function of R_eff — push it from 3 to 1.5 and the final attack rate falls from about 94% to about 58%. Flattening genuinely changes the total, because it changes the momentum with which the system charges past the threshold. The catch is that it must be sustained until susceptibles are actually drawn down, or lifting it produces a rebound.
Both. Bad: the mean stops predicting, small outbreaks have enormous variance, and early data barely extrapolates. Good: since most transmission concentrates in a few settings, cheap local measures can remove most of the risk — an opportunity that simply does not exist under homogeneous mixing. Which is why "is this system overdispersed?" belongs at the top of intervention design rather than in a statistical footnote.
This is the hardest part of porting SIR to ideas. Someone who has had the disease is no longer susceptible, but someone who heard an argument and declined it can still accept it later — arguably more easily, since thresholds accumulate. That suggests the dynamics of ideas resemble a memoried SIS (repeat infection possible) more than SIR, and in a process with memory the notion of "threshold" needs redefining. Whoever gives that pile an observable definition moves this whole question forward.