In 1980s physics, criticality was a rare guest: you had to tune the temperature to a precise value before a system showed scale-free fluctuations. Yet scale-free things are everywhere in nature — earthquake magnitudes, river branching, price jumps, the 1/f noise in electronic devices.
Bak and his colleagues asked: is there a class of system that reaches criticality by itself, with nobody tuning anything? The sandpile is their minimal answer — the dumbest possible device that still demonstrates the point.
02The rules themselves
An N×N grid; each cell holds an integer z — read it as the number of grains stacked there.
Pick a cell at random and add 1 to its z.
Any cell with z > 3 topples: subtract 4 from itself, add 1 to each of its four neighbours.
A toppling may push a neighbour above 3, which topples in turn, until no cell exceeds 3. That whole chain is one avalanche.
Grains pushed past the boundary leave the system (otherwise it would fill up forever).
Only once the avalanche has fully stopped is the next grain added.
That last rule is easy to skip past, and it is what makes the model work: driving must be far slower than relaxation. This is separation of timescales — slow drive, fast release. Where a real system fails this condition, the model’s conclusions do not transfer.
Toppling conserves grains: all four leaving the centre land on its neighbours. Sand exits only at the boundary.
03What you see when it runs
The pile starts empty and added grains do essentially nothing. As cells fill up, avalanches appear and grow. Eventually the system settles into a statistically steady state: average density stops changing, yet avalanches keep occurring, ranging from a single cell to sweeps across the whole grid.
Plot avalanche size against time and you get an irregular spike train — no period, no precursor, no sign that a big one is coming.
No period, no precursor, no "typical size". Histogram these heights and you get a power law.
There is one more elegant property: in 1990 Deepak Dhar proved the model is abelian — when several cells are over threshold at once, the final configuration is identical no matter what order you topple them in. The process may be reordered freely; the outcome is unique.
04What it explains in the real world
The sandpile is not meant to predict any single event. It explains why an entire class of systems has no typical event size. Anything with slow accumulation, threshold release, and release propagating to neighbours can take it as a first approximation:
crustal stress and earthquakes (the Gutenberg-Richter law is the firmest real-world counterpart), forest fuel and wildfire, load redistribution and blackout size, extinction cascades in ecosystems. It makes exactly one core prediction: the size distribution should be a power law, not a bell curve. That prediction can be refuted by data — which is precisely what makes it a scientific model.
What it cannot explain
It does not explain real sandpiles. Awkward, but it has to be said: experiments with actual sand find periodic large collapses beyond a certain pile size, not a power law. The 1996 Nature rice-pile experiment sharpened the condition — elongated grains obey, near-spherical ones do not, and the difference is whether grains interlock and pass stress along. The model captures a class of mechanism, not sand.
It predicts no individual avalanche. It gives a distribution, not a schedule. Asking it when the next big one arrives, or how big this one will get, demands from a purely statistical model exactly what it declines to offer.
It does not apply without separation of timescales. Rule 6 requires the avalanche to finish before the next grain lands. If driving is as fast as relaxation — continuous loading while collapse is under way — the system enters a different regime and a power law need not appear.
It is not a substitute for judgments of responsibility. "Avalanche size is independent of who triggered it" is a claim about statistics. Every toppling has a real causal chain you can trace cell by cell. Reading it as "so nobody can be blamed" swaps a statistical claim for an ethical one.
Even the exponents deserve caution. Avalanche exponents for the 2D BTW model remain disputed (multifractal scaling is involved) and differ between definitions. Before quoting a number, say which quantity and which definition you mean.