What you're seeing
A random network of nodes, each either off (dim) or on (hot amber). We nudge a single node on and let the rule run. Sometimes the activation spreads to one neighbour, then a few more, and a front sweeps across almost the entire network — a global cascade. Far more often it tips no one and dies on the spot. Strikingly, a single dial — how densely the network is wired — flips which of these you get, and it does so at both ends: too sparse fails, and so does too dense.
The rule
Every node has the same threshold φ, a fraction between 0 and 1. A node flips on the moment the share of its neighbours that are already on reaches φ:
- Count a node's neighbours, and how many of them are on.
- If (on ÷ total neighbours) ≥ φ, the node switches on — permanently.
- Repeat in waves until no one new flips. The cascade size is the fraction left on.
The crucial word is fraction. One active neighbour is not enough unless it makes up a large enough share of your neighbours. This is complex contagion — and it is a different beast from simple contagion, where a single exposure can pass something on (a spark to a tree, a germ to a person). Complex contagion is the rule for behaviours, fads, and decisions: you tend to wait until enough of the people around you have committed before you do.
Why it matters
This is Duncan Watts's model of global cascades (2002), built on Mark Granovetter's earlier threshold models of collective behavior (1978). It explains a puzzle: why do almost all shocks fizzle, while a rare few — the same kind of shock — go global? The answer is the cascade window. Single-seed cascades happen only inside a band of connectivity:
- Too sparse (low average degree z): the network is fragmented, with no connected paths for the activation to travel along, so the shock stays local.
- In the window: one tip flips a neighbour, which flips others, and the activation sweeps the network's giant connected component.
- Too dense (high z): now every node has so many neighbours that one or two early movers are below its threshold fraction. The network becomes robust — no spark is a big enough share of anyone to tip them, so it never catches.
That last point is the counter-intuitive heart of the model: more connection can mean more stability, not less. A densely wired system can be globally safer from cascades precisely because each node is anchored by its many connections. The same logic speaks to viral memes, the spread of innovations and fads, bank runs and financial contagion, and cascading power-grid blackouts. Contrast this Atlas's percolation (which is about static connectivity — whether a path exists at all) and its forest-fire model (which is simple contagion — one burning neighbour is enough to ignite a tree). Cascades sit a step beyond both: the threshold makes ignition depend on a fraction.
In the wild
The model is a lens on a family of real phenomena:
- Viral spread of behaviours and products — adoption that needs social proof, not just exposure: joining a platform, a protest, a strike, a run on a bank.
- Financial contagion and systemic risk — a failing institution stresses its counterparties; whether the failure stays contained or cascades depends on the web of exposures.
- Cascading power-grid failures — an overloaded line sheds its load onto neighbours, which can trip in turn, sweeping a blackout across a continent.
- Simple vs. complex contagion — Centola and Macy (2007) argue many social spreads are complex: they need reinforcement from several sources, which is why the "weak ties" that speed simple contagion can actually slow a behaviour that needs a clustered push.
Be honest about what this is. It is a stylised model. Real networks are not Erdős–Rényi random; thresholds vary from person to person and are not fixed and known; and which real-world cascades are "complex" versus "simple" is itself debated. The seed here is a single node, but cascade behaviour depends strongly on seed size (Gleeson and Cahalane, 2007). The cascade window — with both a lower and an upper boundary in connectivity — is a robust qualitative insight, not a quantitative forecast of any particular grid, market, or meme.
Try this
- Fix the threshold and sweep the average degree. At low z the spark dies; raise it into the window and a single seed paints almost the whole network; push it to high z and the cascade dies again because every node is now too stable. (Each link retunes the sim and scrolls you up to watch — then press seed & run.)
- Watch the inset curve: it rises off zero at the window's lower edge and falls back to zero at its upper edge. The white line marks where your current z sits on it.
- Lower the threshold and the window widens — easier to tip means cascades survive across a broader range of connectivity. Raise φ and the window narrows toward nothing.
- Press seed & run several times at a fixed setting: in the window, most seeds fizzle and only some go global — the same rule, wildly different outcomes, just from where the seed lands.
References
- Watts, D. J. (2002). "A simple model of global cascades on random networks." Proceedings of the National Academy of Sciences 99(9), 5766–5771.
- Granovetter, M. (1978). "Threshold models of collective behavior." American Journal of Sociology 83(6), 1420–1443.
- Centola, D., & Macy, M. (2007). "Complex contagions and the weakness of long ties." American Journal of Sociology 113(3), 702–734.
- Gleeson, J. P., & Cahalane, D. J. (2007). "Seed size strongly affects cascades on random networks." Physical Review E 75, 056103.