Last updated: 31 October 2026.
Self-organized criticality (SOC) is the idea that some systems, when driven slowly and allowed to dissipate energy or load in sudden bursts, tune themselves to a critical state without anyone setting a parameter. In that state, small disturbances usually fade, but any of them can trigger an avalanche of any size, and the sizes of the avalanches follow a heavy-tailed pattern. Per Bak, Chao Tang and Kurt Wiesenfeld proposed it in 1987 using a simple sandpile model, and it became one of the most discussed ideas in complexity science. It is also one of the most debated: it explains something real about how thresholds and slow driving produce bursts, and it has been over-claimed as a theory of earthquakes, extinctions and markets. This article runs the original model, shows what it does and does not demonstrate, and draws out the lessons for the design of infrastructure and organisations.
Key takeaways
- SOC needs three ingredients: slow driving, a threshold that triggers local relaxation, and dissipation at the boundary. The system then settles near a critical state by itself.
- In my simulation of a 50 by 50 sandpile, 57% of added grains caused no avalanche at all, while the largest event involved 8,085 topplings from a single grain.
- The trigger is always tiny and identical. The size of an avalanche is determined by the state of the pile, not by the size of the cause.
- Real systems rarely show ideal SOC. Treat it as a mechanism that can generate heavy tails, and test it against alternatives before applying it.
The sandpile model
The Bak-Tang-Wiesenfeld (BTW) model lives on a grid. Each cell holds a number of grains. The rules are minimal:
- Add one grain to a randomly chosen cell.
- If a cell holds four or more grains, it topples: it sends one grain to each of its four neighbours and loses four.
- A toppling can push a neighbour over the threshold, which topples in turn. The chain continues until every cell is below the threshold. That chain is an avalanche.
- Grains that fall off the edge of the grid are lost.
- Repeat, adding grains slowly, one avalanche at a time.
Nothing in these rules mentions a target state. Even so, a pile started empty fills up, and once it has accumulated enough that edge losses balance additions, it stops changing statistically. At that point the average height in my run was about 2.10 grains per cell, close to the value often reported for this model, and the system is poised so that one extra grain may do nothing or may reorganise a large area.
What my simulation shows
I ran the model on a 50 by 50 grid with open boundaries. After a burn-in of 60,000 grains, I added a further 300,000 grains one at a time and recorded the number of topplings each caused.
- About 57% of grains caused no toppling at all.
- Among the avalanches that did occur (about 130,000), the median was in the tens of topplings, about 32% involved 100 or more topplings, and about 6% involved 1,000 or more.
- The largest avalanche involved 8,085 topplings, roughly three topplings for every cell on the grid, caused by a single added grain.
Figure 1. Share of avalanches at least as large as a given size, on logarithmic axes. Sizes span more than three orders of magnitude before the finite grid imposes a cutoff. The dashed line is a reference slope, not a fit.
Two cautions on reading the chart. First, the curve is bent, not a perfect straight line: on a small grid, the finite size curtails large events, and the measure used here (counting topplings) differs from other definitions of size used in the literature. Second, a straight line on log-log axes is not by itself proof of a power law; several other processes produce similar-looking curves. What the simulation does demonstrate is the key qualitative result: no characteristic scale. A system that produces events from one toppling to thousands, with the same trigger every time, cannot be described by an average avalanche.
Why the trigger does not explain the size
Every event in the run began with exactly one grain. The consequence ranged from nothing to a pile-wide cascade. What decided the outcome was the configuration: how many cells sat just below the threshold, and how they were connected. In a sandpile with many cells at three grains, one added grain starts a chain. In a freshly relaxed pile, the same grain does nothing.
This reframes root-cause analysis. Looking for the “cause” of a large avalanche in the last grain is like blaming the final straw. The relevant explanation is structural: the system has accumulated a connected network of near-threshold elements. For cascading events in technical systems, the same logic applies, as discussed in cascading failures.
Why the system tunes itself
The mechanism is a feedback structure, in the terms of reinforcing and balancing loops:
- Loading (slow). Grains arrive at a steady rate, raising the stock of stored grains and making toppling more likely.
- Relaxation (fast, threshold-triggered). When stress crosses the threshold, it is released, lowering the stock and sending load to neighbours.
- Dissipation at the boundary. Grains leave at the edge, so the stock cannot grow forever.
If the pile is too empty, grains accumulate and the system heads toward the critical state. If it is too full, large avalanches drain it. The critical state is an attractor, not a value anyone set. A separation of timescales is essential: loading is slow compared with relaxation, which is why avalanches finish before the next grain arrives. This resembles the delay structures covered in delays in systems thinking, with the added feature that relaxation is nonlinear.
Where SOC has been proposed, and where it holds
| Domain | Proposed analogue | Status |
|---|---|---|
| Earthquakes | Slow tectonic loading, sudden slip; Gutenberg-Richter law | Magnitude statistics are heavy-tailed; whether this is SOC is debated, and the model gives no useful short-term prediction |
| Neuronal avalanches | Cascades of neural activity | Beggs and Plenz (2003) reported avalanche statistics in cortical tissue; interpretation is still discussed |
| Power grid blackouts | Load growth, operator action and upgrades acting on a slowly evolving network | Dobson and co-authors argued that grid evolution can drive the system toward criticality; blackout size data are heavy-tailed |
| Forest fires | Tree growth, lightning, fire spread | Models of the Drossel-Schwabl type reproduce heavy tails; real fire regimes depend heavily on suppression and climate |
| Financial crashes, extinctions | Bursts in a slowly driven system | Largely speculative; competing explanations are strong |
The critical literature is worth knowing. Gunnar Pruessner’s 2012 book on the subject stresses that ideal SOC is a precise property of specific models and that many real systems differ in key ways, such as having dissipation in the bulk, nonstationary driving or external tuning. Heavy tails also arise from other mechanisms, including multiplicative growth, preferential attachment and optimisation under constraints. A good analysis tests the mechanism, not just the shape of a histogram.
Lessons for engineered and organisational systems
1. Efficiency pushes systems toward the edge
Operators who raise utilisation, remove slack and trim redundancy are, in effect, loading the pile. Each step makes sense locally, and together they build a network of near-threshold elements. The grid example is the clearest: as discussed in resilience versus efficiency, incentives that reward lean operation steadily move the system toward the point where a small fault can propagate. The same pressure appears in data-centre growth, covered in AI data centres and the power grid.
2. Preventing small failures can enlarge big ones
In the sandpile, suppressing small avalanches allows stress to accumulate, which raises the chance of a very large release later. The forest-fire version is familiar: decades of suppressing small fires can leave fuel build-up that sustains fires larger than any before. Organisations see it too when every small incident is hidden or smoothed over and the real stress surfaces as a crisis. This is not an argument for tolerating failures that harm people; it is an argument to relieve stress in controlled increments, such as prescribed burns, scheduled maintenance windows and honest incident reporting.
3. Averages mislead
When there is no characteristic size, the mean of the distribution is dominated by rare large events. Planning for the typical avalanche misses the cases that matter. Risk metrics should show the tail, and capacity should be justified against it. A related pattern is a stable state that collapses and then persists, as in metastable failures.
4. Structure matters more than the trigger
Interventions should target the connectivity and the stock of near-threshold elements: add buffers, break coupling, introduce firebreaks, and keep margins. Hunting for ever more precise predictions of the next trigger is usually less effective. This echoes the point from emergence: the large-scale behaviour is a property of the interaction network and cannot be read off the individual parts.
Common mistakes
- Calling every power law SOC. A heavy tail is evidence, not proof.
- Expecting prediction. SOC explains the statistics of events, not when the next big one will happen.
- Ignoring boundaries and driving. Without slow driving and dissipation, the self-tuning argument fails.
- Treating small events as harmless noise. They are the visible part of a distribution that includes much larger events.
Frequently asked questions
What does self-organized criticality mean?
It means a system drives itself to a critical state, where events of all sizes can occur, through its own dynamics of slow loading, threshold-triggered release and dissipation, without outside fine-tuning.
Can SOC predict earthquakes or crashes?
No. It describes the statistics of event sizes, not the timing or size of any specific event. The same trigger can lead to nothing or to a large cascade.
Is the sandpile model realistic?
Not as a literal model of sand. Experiments with real sand piles show behaviour that differs from the idealised model. Its value is as a minimal model that shows how a threshold network can generate scale-free bursts.
What is the difference between SOC and ordinary criticality?
In ordinary critical phenomena, such as a magnet at its Curie point, an external parameter such as temperature must be tuned to a precise value. In SOC the system reaches the critical state by itself.
Conclusion
The sandpile shows in a few lines of code how a simple system can generate disproportionate events from identical small causes. Its lessons are structural: stress accumulates where releases are suppressed, margins erode under efficiency pressure, and tail events dominate risk. Its limits matter as well, since the idea has been stretched beyond its evidence in several fields. Used carefully, it is a prompt to look at network structure, stored stress and buffers rather than the last trigger. For a model of how failures spread across connected parts, see cascading failures, and the glossary entry on emergence.
Sources and further reading
- Bak, P., Tang, C. and Wiesenfeld, K., “Self-organized criticality: An explanation of the 1/f noise,” Physical Review Letters 59, 381, 1987.
- Bak, P., How Nature Works: The Science of Self-Organized Criticality, Copernicus, 1996.
- Pruessner, G., Self-Organised Criticality: Theory, Models and Characterisation, Cambridge University Press, 2012.
- Beggs, J. M. and Plenz, D., “Neuronal avalanches in neocortical circuits”, Journal of Neuroscience 23(35), 2003.
- Dobson, I., Carreras, B. A., Lynch, V. E. and Newman, D. E., “Complex systems analysis of series of blackouts: Cascading failure, critical points, and self-organization,” Chaos 17, 026103, 2007.

Discussion