Percolation

The intuition — a forest catching fire

Picture a square grid of cells. Independently fill each cell with probability p — call a filled cell a tree, an empty cell bare ground. Two trees that touch edge-to-edge are connected; a group of mutually-touching trees is a cluster. Now slowly turn the dial on p and ask one question: is there a cluster that reaches all the way from the left edge to the right edge?

At small p you get scattered specks — tiny clusters, none of them large. Crank p up and the clusters grow and start to merge. The surprising part is what happens near a special value. The largest cluster does not grow smoothly to span the grid. Instead, over a vanishingly narrow window of p, a spanning cluster snaps into existence — like a forest that was a patchwork of safe firebreaks one moment and a single continent of fuel the next. That tipping point is the percolation threshold p_c, and the abruptness is the whole point: percolation is a phase transition built out of nothing but geometry and randomness — no temperature, no energy, no forces.

How percolation works

There are two equivalent flavors, and the difference is where you put the randomness:

• Site percolation. Each lattice node is occupied with probability p, independently. Two occupied neighbors are connected.
• Bond percolation. Every edge between neighboring nodes is open with probability p, independently. Connectivity flows only along open bonds.

In both cases you build the connected components (clusters) and watch for one that touches opposite boundaries. The mathematics is governed by a single observable: the probability P_∞(p) that a randomly chosen site belongs to the infinite (spanning) cluster. In the infinite-lattice limit this is exactly zero below the threshold and positive above it:

P_∞(p) = 0            for p < p_c
P_∞(p) > 0            for p > p_c
P_∞(p) ∝ (p − p_c)^β  just above p_c

P_∞ is the order parameter — the same role magnetization plays for a magnet. The control parameter p plays the role of temperature. The fact that P_∞ turns on continuously from zero (rather than jumping) is what makes percolation a continuous, or second-order, transition.

The governing physics — critical exponents and scaling

Near p_c every interesting quantity follows a power law in the distance from the threshold, |p − p_c|. Three quantities and three exponents do most of the work:

Order parameter        P_∞      ∝ (p − p_c)^β        (p > p_c)
Mean finite cluster    S        ∝ |p − p_c|^(−γ)
Correlation length     ξ        ∝ |p − p_c|^(−ν)

The correlation length ξ is the typical linear size of the finite clusters. As p → p_c it diverges — clusters of every size appear, and the system loses any characteristic length scale. That scale-free state is exactly why a fractal cluster emerges at the threshold (see below).

At criticality itself the cluster-size distribution n_s (number of clusters of s sites) becomes a pure power law, with no preferred size:

n_s ∝ s^(−τ)         at p = p_c

The exponents are not independent — they obey scaling relations that hold in every dimension, for example:

2β + γ = d·ν              (hyperscaling)
τ = 2 + β / (β + γ)
d_f = d − β/ν             (fractal dimension of the cluster)

where d is the spatial dimension. The remarkable fact — the heart of universality — is that these exponents do not depend on the lattice type or on whether you do site or bond percolation. The threshold p_c is lattice-specific; the exponents are not. Every 2D percolation problem shares β = 5/36, ν = 4/3, γ = 43/18.

Critical exponents by dimension

The exponents are known exactly in 2D (from conformal field theory and Coulomb-gas methods) and at the upper critical dimension d = 6, where they take mean-field (Bethe-lattice) values and stay frozen for all higher d. In 3D they are only known numerically.

Exponent	Meaning	2D (exact)	3D (numerical)	Mean field (d ≥ 6)
β	Order parameter P∞ ∝ (p − p_c)β	5/36 ≈ 0.139	≈ 0.418	1
γ	Mean cluster size S ∝ |p − p_c|−γ	43/18 ≈ 2.389	≈ 1.793	1
ν	Correlation length ξ ∝ |p − p_c|−ν	4/3 ≈ 1.333	≈ 0.876	1/2
τ	Cluster-size law n_s ∝ s−τ at p_c	187/91 ≈ 2.055	≈ 2.18	5/2
d_f	Fractal dim. of cluster at p_c	91/48 ≈ 1.896	≈ 2.523	4

Because hyperscaling 2β + γ = dν only holds below d = 6, dimension six is the upper critical dimension: above it, fluctuations stop mattering and mean-field theory becomes exact.

Thresholds — every lattice has its own p_c

The threshold is a pure number that depends on the geometry: how many neighbors each node has (coordination number z) and how the lattice is wired. A few are known exactly; most come from large Monte-Carlo simulations.

Lattice	Type	Coordination z	pc	Known
2D honeycomb	Site	3	≈ 0.6970	numerical
2D square	Site	4	≈ 0.5927	numerical
2D triangular	Site	6	1/2	exact
2D square	Bond	4	1/2	exact
2D triangular	Bond	6	2·sin(π/18) ≈ 0.3473	exact
3D simple cubic	Site	6	≈ 0.3116	numerical
3D simple cubic	Bond	6	≈ 0.2488	numerical
Bethe lattice (z neighbors)	Bond	z	1/(z − 1)	exact

Two patterns are worth burning in. First, more neighbors means a lower threshold — more ways to connect, so a giant cluster forms with fewer occupied sites. Second, more dimensions means a lower threshold — extra routes around obstacles. The Bethe lattice (an infinite tree) is the one case you can solve with pencil and paper: a bond cluster grows on average by (z − 1)·p new sites per generation, so it survives forever exactly when (z − 1)·p ≥ 1, giving p_c = 1/(z − 1).

Worked example — the Bethe lattice threshold

Take a Bethe lattice with coordination number z = 3 (each node has three neighbors). Start at any node and follow open bonds outward. Each node you reach has z − 1 = 2 bonds leading away from where you came. Each of those is open with probability p, so the expected number of new occupied nodes one step out is:

branching factor  =  (z − 1) · p  =  2p

If the branching factor is below 1, the cluster dies out (a subcritical branching process); above 1, it grows without bound with positive probability. The knife-edge is:

2p = 1   ⟹   p_c = 1/2

Just above it, the order parameter rises linearly — the mean-field result P_∞ ∝ (p − p_c)^β with β = 1. Now compare to the 2D square lattice, where loops let clusters reconnect and reinforce: the threshold drops differently and the exponent softens to β = 5/36, so the giant cluster appears far more abruptly than the Bethe-lattice straight line. The lattice geometry changes the numbers; the existence of a sharp threshold does not.

The fractal at the threshold

Exactly at p_c the incipient spanning cluster is not a solid blob — it is a fractal, riddled with holes on every length scale, because the correlation length has diverged and there is no scale to make it look smooth. Its mass inside a box of side L grows more slowly than the box volume:

M(L) ∝ L^(d_f),   d_f = 91/48 ≈ 1.896 in 2D  (vs. d = 2)
                  d_f ≈ 2.523       in 3D  (vs. d = 3)

Because d_f < d, the cluster fills a vanishing fraction of space as L → ∞ — consistent with P_∞ = 0 right at the threshold. Only above p_c, on scales larger than the (now finite) correlation length ξ, does the infinite cluster look compact and fill a finite fraction of the lattice. Inside that cluster, the subset that actually carries current — the backbone — is a thinner fractal still (d_B ≈ 1.64 in 2D), with dead-end "dangling ends" hanging off it that connect but conduct nothing.

Where percolation shows up — and what it costs

• Oil and gas recovery. Crude flows through a sandstone reservoir only if its pore network percolates. Typical reservoir sandstone porosity runs 10–30%; below a connectivity threshold the oil is trapped no matter how hard you pump. Percolation theory underpins how much of the oil-in-place (often only 30–50%) is actually recoverable.
• Epidemics. The SIR epidemic model maps exactly onto bond percolation with p equal to the disease transmissibility. The epidemic threshold p_c is the dividing line between a contained outbreak and a population-spanning epidemic; the giant cluster is the final outbreak size. Vaccinating a fraction of nodes is exactly site-diluting the lattice below threshold.
• Conductive composites. Mix conductive filler (carbon black, silver flakes, carbon nanotubes) into an insulating polymer and conductivity is essentially zero until the filler fraction crosses p_c — then it jumps by 10 or more orders of magnitude over a tiny composition range. High-aspect-ratio fillers like nanotubes percolate at well under 1% by volume, which is why a pinch of them turns plastic conductive.
• Forest fires and spread of damage. Whether a fire crosses a landscape, or a blackout cascades across a power grid, is a percolation question on the connectivity of fuel or of overloaded lines.
• Network robustness. Randomly removing nodes from the Internet, a power grid, or a food web is site percolation in reverse: the network stays globally connected until the surviving fraction drops below p_c, at which point the giant component shatters into disconnected pieces.
• Materials and geology. Gelation (when a polymer solution sets into a gel) is the chemical-bond percolation of cross-links; groundwater flow, fracture networks in rock, and the porosity of coffee grounds (yes, that is literally where the word comes from) are all percolation problems.

Common misconceptions and edge cases

• "The threshold is 50%." Only for special cases — bond percolation on the square lattice and site percolation on the triangular lattice are exactly 1/2. Square-lattice site percolation is ≈ 0.5927, and 3D thresholds are much lower (≈ 0.31 for the simple cubic site problem). There is no universal threshold value.
• "The transition is gradual." In an infinite system it is a true singularity — exactly zero below p_c, positive above. The smooth S-curve you measure is a finite-size artifact, smeared over a window of width ≈ L^(−1/ν) that vanishes as the system grows.
• "More occupied sites always means a bigger biggest-cluster jump." Standard percolation is continuous: the giant cluster appears with zero initial size. Only contrived "explosive percolation" rules (which suppress large merges) can sharpen it toward a discontinuity — and even those turn out to be continuous on close inspection.
• "Spanning means conducting fully." Above p_c the spanning cluster carries current only through its backbone; dangling ends and isolated loops belong to the cluster but conduct nothing. Conductivity rises with its own exponent t (≈ 1.30 in 2D, ≈ 2.0 in 3D), distinct from β.
• "p_c depends on system size." The true threshold is a fixed lattice constant. What shifts with size is the apparent crossing point in a finite sample, which converges to p_c as L → ∞ — the basis of finite-size scaling.
• "Percolation needs physics — energy, temperature, forces." It needs none of them. It is a pure probability-and-geometry model, which is exactly why it is the cleanest available illustration of a critical phase transition and universality.