Self-Organized Criticality

After every crash and every outage we go looking for the culprit. But there is a whole class of system whose largest disasters are set off by the most ordinary thing imaginable — and in those systems, the act of hunting for a culprit carries no information at all.

01The story of one grain

In 1987 three physicists — Per Bak, Chao Tang and Kurt Wiesenfeld — wrote down a set of rules simple enough to be faintly ridiculous → ref · sandpile model:

Picture a sheet of graph paper with a few grains of sand in each square. Pick a square at random and add one grain. If that square now holds more than three grains it topples: four grains leave, one going to each of its four neighbours. A neighbour that now holds more than three topples in turn. Grains that spill off the edge are gone.

That's the entire model. No physics, no friction coefficient, no gravity. Yet run it and this happens: the grain you just added may do nothing at all, or it may set off a chain of topplings that sweeps across half the sheet. And you cannot tell the two apart in advance, because the grain you added was identical both times.

The most important sentence of this issue is already buried in there: in systems like this there is no correspondence between the triggering event and the size of the event. So when you ask what caused this great avalanche, the answer is always an ordinary grain of sand — true, and completely unilluminating.

02The system walks to the cliff edge by itself

In the phase-transition picture, the critical point is something you tune to: cool the magnet to precisely the right temperature and it sits at criticality. Miss by a little and it doesn't. This is why physics long treated criticality as rare and delicately prepared.

What's strange about the sandpile is that nobody is tuning anything. You add grains one at a time. Past a certain point the system settles onto a particular slope — too shallow and grains pile up and steepen it, too steep and a large avalanche unloads it. Both sides push inward.

Criticality here isn't a target to aim at; it's an attractor. The system's own dynamics carry it there and then it stays. That is the whole content of the word "self-organized": nobody organized it into criticality, it went by itself.

This changes the character of the whole subject. If criticality had to be finely prepared, critical systems in the real world would be vanishingly rare. If systems climb there on their own, then anything built out of slow accumulation plus threshold release drifts automatically into a state where something big could happen at any moment — crustal stress, forest fuel load, leverage in a financial system, the technical debt nobody in the org wants to raise.

03There is no such thing as a typical size

Run the sandpile ten thousand times, record how many cells each avalanche touched, and plot the distribution. You get a power law: enormous numbers of tiny avalanches, very few huge ones, no gap in between — and no typical value.

Compare human height. Height has a typical value, around 1.7 metres, with almost everyone clustered nearby. You have never met a three-metre person and never will. That's a bell curve; it has a scale.

Avalanche size has nothing of the kind. The longer you measure, the larger the largest avalanche you've seen, without limit. Asking how big a typical avalanche is, is like asking how big a typical earthquake is — the question has no mathematical answer.

That straight line is everywhere in the real world. The Gutenberg-Richter law puts it most bluntly for earthquakes: raise the magnitude by one and the number of such quakes falls to roughly a tenth. The relation holds across several orders of magnitude and a century has not overturned it. Blackout sizes, forest fire areas, and market drawdowns all produce comparable lines.

So "big events need no big cause" acquires an exact meaning: large and small events are two ends of one process, not two kinds of thing. To go looking for a special cause commensurate with the largest one is to assume it belongs to a separate category — and on the distribution there is no such dividing line to be found.

04Where this stops being true

Now the other side. Self-organized criticality is among the most abused ideas in complexity science, and you should know its limits before you use it.

First, real sand doesn't really obey it. This is a little awkward — around 1990 people actually ran the experiment, building real piles and measuring avalanches, and found that beyond a certain size what appeared were periodic large collapses rather than a power law. The sandpile model is a fine model; it just isn't a model of sandpiles.

Later work in Chicago and Oslo gave a more honest answer. In the rice-pile experiment published in Nature in 1996, elongated grains produced a power law and near-spherical grains did not. What mattered was whether grains could interlock and pass stress along. That's a specific precondition, and not everything satisfies it.

Second, "looks like a power law" usually isn't one. On log axes the eye readily reads a gentle curve as a straight line, and log-normal or exponentially truncated distributions look much the same when plotted. In 2009 Clauset, Shalizi and Newman re-examined a set of widely cited power-law datasets with a uniform statistical test, and a substantial share failed. Looking straight is not evidence.

Third, and most important: none of this is an excuse. "Big events need no big cause" says the magnitude needs no matching cause. It does not say there was no cause, and it does not say nobody is responsible. Every toppling in the sandpile has an exact causal chain you can trace cell by cell. SOC claims one thing only: the length of those chains is distributed without a scale, so the identity of the chain's starting point does not explain the size of the outcome. Reading it as "so no one can be blamed" swaps a claim about statistics for a claim about responsibility.