One parameter crosses a line, and the whole thing connects
2026-08-04 · Self-Organization & Criticality
A rumour breaks out of its bubble, an epidemic takes off, a grid trips, a niche hobby is suddenly everywhere — and afterwards we all go looking for the thing that lit the fire. In a whole class of systems the sparks were always there. What changed is whether the paths were open; and whether they are open flips, inside a very narrow band, from almost never to everywhere.
Take a sheet of graph paper and fill in squares at random. At 45% the page is a scatter of unrelated blotches, the biggest one pitifully small. Keep going to 70% and one blotch has swallowed nearly every filled square, running from the top edge clear down to the bottom.
You might expect the middle to be gradual — fill in more, get a slightly bigger blotch. It isn't. For a long stretch the largest blotch barely grows; then, inside a narrow band, it eats the page in one gulp. There is no half-connected stage in between.
Worse: if the only gauge you own reads "how big is the biggest blotch", you see nothing at all until the line is crossed. Not because you weren't watching carefully — because that quantity genuinely does not move beforehand. This is why events of this kind always arrive "without warning".
Five issues on this site deal with criticality, each from a different angle: Topic 10 is bifurcation and tipping (the system jumps to another stable state), Topic 16 is phase transitions and universality (why details stop mattering at the critical point), Topic 18 is self-organized criticality (the system climbs there by itself), Topic 36 is critical slowing down (how to monitor it). This issue takes the topological view: never mind how hot, how full or how steep the system is — is it connected or not?
In 1957 two Englishmen, Simon Broadbent and John Hammersley, were working on a thoroughly practical problem: the filter in a gas mask is a lump of porous carbon, and gas crossing it has to travel along pores that happen to link up. The pores are randomly placed. How dense do they have to be before gas gets from one side to the other?
They stripped it to the bone: a grid where each position is open with probability p and shut otherwise; adjacent open positions count as connected; now ask when a path appears from top to bottom. The mathematics has been called percolation ever since. → ref · the percolation model
The numbers come out clean. On an ordinary square lattice, opening cells at random (site percolation) has a threshold of about 0.5927; if instead every cell stays and you randomly open the links between them (bond percolation), the threshold is exactly 1/2 — an "exactly" that Harry Kesten proved rigorously in 1980. Move to a three-dimensional cubic lattice and roughly 0.3116 is enough: more room means more ways around, so less density is needed.
Notice how different these two quantities are. p is per-cell, local, continuously adjustable. "Connected or not" belongs to the entire picture, is collective, and has exactly two answers. The knob in your hand is continuous; the result it controls is binary. Everything strange in the rest of this issue grows out of that one sentence.
The whole business is clearest as a curve. Horizontal axis, the occupation probability p; vertical axis, the share of the lattice taken by the largest single piece — physicists call such a quantity an order parameter, meaning it exists to tell you which phase the system is in.
The left-hand shape deserves a long look. From 0.30 up to 0.55 the curve hugs the floor: with over half the cells filled, the biggest piece is still under 2% of the board. Then between 0.56 and 0.62 — six more percentage points — it goes from 5% to 52%.
The right-hand panel shows how that is done. Near the threshold the mid-sized pieces all become large, large enough to be almost touching; a little more occupation and they start joining one another wholesale, merging into a single piece. So the spanning cluster is not one lump that slowly grew. It is a great many mid-sized lumps being stitched together at once. Once the stitching is over only crumbs are left, which is why that curve then collapses.
One reading has to be resisted here. The left curve looks as if "nothing happens until 0.59", but the real mechanism is that the typical size of a cluster climbs steadily as the threshold approaches and diverges at it — that length scale is called the correlation length. Accumulation is happening the whole time. It is simply accumulating in a quantity you are not watching: you are staring at how big the biggest piece is, while what is changing is how far apart the pieces are.
Stop using a size metric as an early warning for a connectivity event. "How big is the biggest piece" barely moves before the line is crossed; it is the outcome, not the precursor. Measure two other things instead: how large the runners-up have become (they grow together ahead of the threshold), and how many edges still separate one piece from another. Concretely: stop tracking only the size of your top community or top account, and start tracking whether overlap among the second tier is rising.
A square lattice is a toy. Real connectivity happens on networks: people to people, machines to machines, junction to junction. On a network the question becomes — how many links per node, on average, before the whole thing joins up?
The answer is uncomfortably elegant: if links are placed at random, one link per node is enough. That is the Erdős–Rényi result of 1960. Below one, the network is scattered debris; cross one and a huge lump appears that swallows a substantial fraction of all nodes. Its formal name is the giant component.
"One link on average" sounds far too low, and the reason it isn't is that connectivity does not require everybody to know many people. It only requires that nobody be left out entirely — and when links fall at random, leaving a node out is harder than it feels.
But that one holds only for randomly scattered links. The general criterion looks at the distribution of degree — a node's degree being its number of links. A giant component exists when ⟨k²⟩ / ⟨k⟩ > 2 (the Molloy–Reed criterion; ⟨·⟩ means an average over all nodes). There is a square in that expression, which means a handful of exceptionally well-connected nodes drag the threshold way down, since squaring is brutally sensitive to large values.
In 2000 Reuven Cohen and colleagues took this to its limit: in a scale-free network — one where a few nodes have vastly more links than the rest — the percolation threshold tends to zero as the network grows. You essentially cannot break such a network apart by removing nodes at random. The same fact explains why these networks are so fragile against targeted attack, which is the subject of Topic 23.
Compressed into one usable sentence: connectivity is a property of edges, not of nodes. To make something join up, adding links — especially links between clumps — beats adding members by a wide margin. To keep something apart, cut links and hubs rather than shaving a little participation off everyone.
Whenever the goal is "get this connected" or "keep this from connecting", first work out whether your lever adds nodes or adds edges — the two differ by orders of magnitude. To push something out: don't spend the budget recruiting more participants, spend it manufacturing the few cross-circle connections (people who belong to both circles, artefacts both sides can cite). To block something: find the few edges that stitch separate clumps together and cut those, instead of uniformly reducing everyone's involvement.
Percolation is among the most useful and most abused models on this site: it is far too easy to apply, and anything at all can be described as "hitting the threshold". Before using it, know where its premises give way.
First, real networks are not randomly arranged. "Average degree one and it connects" assumes links fall at random. In practice, human, organizational and code connections clump heavily: a friend's friend is still a friend. The same number of links, if all of them are spent inside small circles, can leave a network far above average degree one and still in pieces.
Second, plenty of things do not travel across a single edge. Percolation assumes one contact suffices, which is roughly true of a virus. Changing a behaviour usually is not: switching supplier, joining a protest, adopting a new tool — people typically need to see it independently from several different others before they move. This is complex contagion (Centola and Macy, 2007). The consequence is not merely "a higher threshold". In ordinary percolation the most valuable link is the long one that bridges circles; under complex contagion long ties are often useless — one outsider's recommendation won't start you — and dense local circles do the work instead. Getting the model wrong costs you the budget, spent precisely on the kind of connection that doesn't help.
Third, direction has been erased. Percolation edges run both ways. Influence, money and information often run one way: he can see you without you seeing him. On a directed graph, reachability splits into "how many I can reach" and "how many can reach me" — two different giant components with two different thresholds. Anyone computing connectivity by treating follows, citations or supply relations as undirected edges is systematically overstating how joined-up things are.
Fourth, and most important: you cannot measure p. The 0.5927 belongs to an infinite lattice; in a finite system that cliff gets smoothed out, and the smaller the system the more it smooths. More fundamentally, nothing in the world displays "current connection probability: 0.57" — occupation probability is a parameter inside a model, not a reading. So what percolation hands you is never an alarm level. It is a list of what to watch: the size of the runner-up clusters, the count of cross-cluster edges, the overlap between communities. Anyone who tells you "we're at 0.6 now" has made that number up.
Before invoking percolation, answer all four: are the links roughly randomly placed, or heavily clumped? Does one contact suffice, or are several independent ones needed? Do the edges have direction? And which quantity can you actually measure? Pass all four and you may use the numbers; pass only the first two and you may use it qualitatively — that is, use its direction (add edges not nodes, watch the runners-up not the leader) but not its scale.
What's left is changing the observable. The order parameter — the largest cluster — doesn't move before the line, but the size of the runners-up, the distance between clusters and the count of cross-cluster edges all do. Topic 36's critical slowing down is the same idea in another key: don't watch the state, watch how the system responds to a nudge. The common thread is that suspicious quiet usually means the wrong gauge, not that nothing is happening.
Because they meet different degree distributions. Connecting a uniform network just takes edges. Breaking a network with hubs is nearly immune to random node removal — the hubs drag the threshold towards zero, so you have to keep deleting until only crumbs remain. That asymmetry is the source of Topic 23's "robust yet fragile": the same network's response to random failure and to targeted attack differs not in degree but in direction.
When the things you don't want travel on the same network. Cross-circle edges raise the connectivity of good news and bad news alike — the links that break information silos in an organization are the links that turn a local failure into a global one. Topic 25's cascading failures is that ledger. So "add edges" is never unconditionally good; ask first what runs on this network.
The 1/2 comes from the square lattice's self-duality — the graph made of open bonds and its dual have the same shape, so the threshold can only sit at the symmetric point, and Kesten turned that intuition into a proof. Site percolation has no such symmetry, so its threshold is just a number fixed by the lattice's shape, with no closed form. Which is itself worth remembering: the threshold is a property of the lattice, not of percolation. Move to a triangular lattice and the site threshold is exactly 1/2 again.