Percolation Theory

Percolation theory studies the emergence of long-range connectivity in random media. On a lattice, each edge (bond) or site is independently open with probability

p

; there exists a sharp critical threshold

p_{c}

above which an infinite connected cluster appears almost surely, and below which all clusters are finite. It is the canonical model of a geometric phase transition, governing flow in porous rock, epidemic spreading, network robustness, and conductivity of random composites. Origin: Broadbent & Hammersley (1957).

Percolation is the simplest probabilistic model exhibiting a continuous (second-order) phase transition with critical exponents, universality, and scaling. Given the infinite lattice $Z^{d}$ , one declares each edge open with probability $p$ and closed otherwise, independently. The central object is the percolation probability $θ (p) = P_{p} (0 \leftrightarrow \infty)$ , the chance the origin lies in an infinite open cluster. The surprising fact is that $θ$ is identically zero for $p < p_{c}$ and strictly positive for $p > p_{c}$ : connectivity switches on at a single sharp point, even though edges are decided by entirely local, independent coin flips. Near $p_{c}$ the model becomes scale-invariant, and its large-scale behavior depends only on dimension and broad structural class — not on lattice details.

The Core Insight

A purely local random rule (each edge open with the same independent probability $p$ ) produces global structure with a sharp threshold. Because clusters grow by a branching-like competition between the number of ways to extend and the per-edge survival probability $p$ , there is a critical balance: below it, finiteness wins exponentially; above it, an infinite cluster condenses out. The transition is sharp (no intermediate "slowly growing" regime), and near it the system has no characteristic length scale, giving rise to power laws and universality.

Formal Statement

Bond Percolation on a Lattice

Let $L = (V, E)$ be a connected, infinite, locally finite graph (canonically $Z^{d}$ with nearest-neighbor edges). Fix $p \in [0, 1]$ . To each edge $e \in E$ assign an independent random variable $ω (e) \in {0, 1}$ with $P (ω (e) = 1) = p$ . The configuration $ω \in Ω = {0, 1}^{E}$ is drawn from the product measure $P_{p}$ . An edge is open if $ω (e) = 1$ . The open cluster $C (x)$ of vertex $x$ is the set of vertices reachable from $x$ via open edges. Define the percolation probability:

θ (p) = P_{p} (∣ C (0) ∣ = \infty) = P_{p} (0 \leftrightarrow \infty),

where $0$ is a distinguished origin and $0 \leftrightarrow \infty$ means $0$ belongs to an infinite open cluster.

⟳

θ (p) = lim_{n \to \infty} P_{p} (0 \leftrightarrow \partial B_{n})

by monotone continuity, since

{0 \leftrightarrow \partial B_{n}}

is a decreasing sequence of events with intersection

{0 \leftrightarrow \infty}

Critical Probability

The critical probability (percolation threshold) is

p_{c} (L) = sup {p \in [0, 1] : θ (p) = 0} = in f {p : θ (p) > 0} .

Since $θ$ is nondecreasing in $p$ (by monotone coupling), this is well defined. For $Z^{d}$ with $d \geq 2$ one has $0 < p_{c} < 1$ , ensuring a genuine nontrivial transition.

⟳monotonicity of

θ

follows from the standard coupling: realize all

ω (e)

1 [U_{e} \leq p]

with

U_{e} \sim Unif [0, 1]

i.i.d.; increasing

p

can only open more edges, so

{0 \leftrightarrow \infty}

is preserved.

Existence of a Nontrivial Threshold on $\mathbb{Z}^d$, $d\ge 2$

For nearest-neighbor bond percolation on $Z^{d}$ with $d \geq 2$ , $0 < p_{c} (Z^{d}) < 1$ . More precisely, with $μ$ the connective constant (growth rate of self-avoiding walks),

\frac{1}{μ} \leq p_{c} (Z^{d}) \leq 1 - \frac{1}{μ ^{'}},

where $μ^{'}$ is the connective constant of the matching/dual lattice. Furthermore $θ (p)$ is continuous on $[0, 1] ∖ {p_{c}}$ , and $θ (p_{c}) = 0$ is known for $d = 2$ and $d \geq 11$ (and conjectured for all $d$ ).

⟳lower bound from first-moment (Peierls-type) count of open self-avoiding paths from origin; upper bound from a contour/dual argument showing closed dual circuits cannot surround the origin when

p

is large.

Uniqueness of the Infinite Cluster — Burton–Keane

For $p > p_{c}$ on $Z^{d}$ , there is $P_{p}$ -almost surely a unique infinite open cluster.

⟳the number

N

of infinite clusters is a translation-invariant random variable, hence a.s. constant by ergodicity; rule out

N = k \geq 2

via finite energy, and rule out

N = \infty

by the trifurcation/counting argument that infinitely many "trifurcation points" would overflow the boundary of a box.

Why It Holds / Derivation

Sharpness of the Phase Transition (sketch of Aizenman–Barsky / Menshikov / Duminil-Copin–Tassion)

We outline why the transition at $p_{c}$ is sharp: subcritical clusters decay exponentially, not just have finite mean size. Define the susceptibility $χ (p) = E_{p} ∣ C (0) ∣$ .

Finite-volume order parameter. For a box $Λ_{n} = [- n, n]^{d}$ , set $φ_{p} (Λ_{n}) = p \sum_{x \in Λ_{n}, y \in / Λ_{n}} P_{p} (0 \leftrightarrow x in Λ_{n})$ , measuring the expected number of open boundary crossings emanating from clusters of the origin.

Differential inequality (Duminil-Copin–Tassion). One establishes

θ^{'} (p) \geq \frac{θ ( p )}{p ( 1 - p )} \cdot \frac{1 - p ~ _{n}}{p ~ _{n}}, \frac{d}{d p} P_{p} (0 \leftrightarrow \partial Λ_{n}) \geq \frac{1}{p} φ_{p} (Λ_{n}) (1 - P_{p} (0 \leftrightarrow \partial Λ_{n})),

relating the derivative of connection probabilities to the boundary functional

φ_{p}

Dichotomy. Define $\tilde{p}_{c} = sup {p : in f_{n} φ_{p} (Λ_{n}) < 1}$ . If $p < \tilde{p}_{c}$ , then for some $n$ , $φ_{p} (Λ_{n}) < 1$ , and iterating across nested boxes yields exponential decay:

P_{p} (0 \leftrightarrow \partial Λ_{n}) \leq e^{- c (p) n}, c (p) > 0 for p < p_{c} .

Coincidence of thresholds. If $p > \tilde{p}_{c}$ , integrating [4] forces $θ (p) \geq \frac{p - p ~ _{c}}{p ( 1 - p )} \cdot (const) > 0$ . Hence $\tilde{p}_{c} = p_{c}$ , so for all $p < p_{c}$ the cluster radius has exponential tails — the transition is sharp with no intermediate phase. $□$

⟳the heart is the OSSS / randomized-algorithm or the Duminil-Copin–Tassion differential inequality; the key is that

φ_{p} (Λ_{n})

acts as a finite-volume criterion whose crossing of the value

1

pinpoints

p_{c}

Mean Cluster Size Diverges at $p_c$

χ (p) = E_{p} ∣ C (0) ∣ = x \in Z^{d} \sum P_{p} (0 \leftrightarrow x) < \infty for p < p_{c}, χ (p) \to \infty as p ↑ p_{c} .

Near

p_{c}

(subcritical side) one expects

χ (p) \sim C (p_{c} - p)^{- γ}

⟳

χ (p) = \sum_{x} τ_{p} (0, x)

with two-point function

τ_{p}

; exponential decay of

τ_{p}

for

p < p_{c}

(from sharpness) gives finiteness, and the divergence is the susceptibility blow-up at the critical point.

John Michael Hammersley | 1920–2004

Hammersley, working at Oxford, co-introduced percolation with Simon Broadbent in 1957 to model the spread of fluid through a random porous medium (motivated by gas-mask filter design). He distinguished percolation processes (random medium, deterministic fluid) from diffusion processes (deterministic medium, random fluid), a conceptual split that defined the field. Hammersley also pioneered the rigorous probabilistic treatment of self-avoiding walks and subadditivity (the subadditive ergodic theorem), tools that remain central to lattice models. His insistence on rigorous bounds — rather than physical heuristics alone — set the mathematical tone for the subject.

Building Intuition

Picture a vast orchard where each pair of neighboring trees may or may not be connected by a passable footpath; each path exists with probability $p$ . For small $p$ , paths are rare, you can wander only within a small local cluster, and the orchard is fragmented. As $p$ grows, clusters merge. At a precise critical $p_{c}$ , the largest cluster suddenly spans the entire (infinite) orchard — you can walk forever. The miracle is the sharpness: there is no "almost-spanning, slowly-growing" regime. The competition is between entropy (the number of possible paths of length $n$ grows like $μ^{n}$ ) and energy (each path survives with probability $p^{n}$ ); when $p μ > 1$ paths proliferate, when $p μ < 1$ they die out exponentially, locating $p_{c} \approx 1/ μ$ . Right at criticality, clusters look the same at every magnification — a self-similar fractal — which is why critical exponents and conformal symmetry appear.

Percolation probability $\theta(p)$ vs. occupation probability $p$, showing the sharp transition at $p_c$.

<svg viewBox="0 0 520 360" xmlns="http://www.w3.org/2000/svg">
<rect width="520" height="360" fill="#0a0e17"/>
<!-- axes -->
<line x1="60" y1="300" x2="480" y2="300" stroke="#94a3b8" stroke-width="2"/>
<line x1="60" y1="300" x2="60" y2="40" stroke="#94a3b8" stroke-width="2"/>
<text x="470" y="325" fill="#e2e8f0" font-size="16">p</text>
<text x="20" y="55" fill="#e2e8f0" font-size="16">θ(p)</text>
<text x="50" y="320" fill="#e2e8f0" font-size="13">0</text>
<text x="455" y="320" fill="#e2e8f0" font-size="13">1</text>
<text x="40" y="55" fill="#e2e8f0" font-size="13">1</text>
<!-- 1 gridline -->
<line x1="60" y1="60" x2="470" y2="60" stroke="#334155" stroke-width="1" stroke-dasharray="4 4"/>
<!-- pc marker -->
<line x1="230" y1="300" x2="230" y2="40" stroke="#475569" stroke-width="1" stroke-dasharray="3 3"/>
<text x="212" y="320" fill="#d4a853" font-size="14">p_c</text>
<!-- theta = 0 segment -->
<line x1="60" y1="300" x2="230" y2="300" stroke="#d4a853" stroke-width="3"/>
<!-- theta rising curve for p>pc, going to 1 at p=1 -->
<path d="M230,300 C 270,200 360,90 470,60" fill="none" stroke="#d4a853" stroke-width="3"/>
<!-- annotations -->
<text x="110" y="285" fill="#e2e8f0" font-size="13">θ = 0  (subcritical)</text>
<text x="300" y="150" fill="#e2e8f0" font-size="13">θ ~ (p − p_c)^β</text>
<circle cx="230" cy="300" r="4" fill="#d4a853"/>
<text x="240" y="80" fill="#e2e8f0" font-size="13">supercritical:</text>
<text x="240" y="98" fill="#e2e8f0" font-size="13">infinite cluster</text>
</svg>

Worked Examples

Bethe Lattice (Cayley Tree) — Exact Threshold and Order Parameter

Given: An infinite regular tree where each vertex has coordination number

z

(so

z - 1

"children" per vertex away from the root). Bond percolation with parameter

p

. Step 1 — Branching criterion. Following an open edge into the tree, the expected number of further open children is

(z - 1) p

. The cluster is a Galton–Watson branching process with mean offspring

m = (z - 1) p

. Survival (infinite cluster) requires

m > 1

p_{c} = \frac{1}{z - 1} .

Step 2 — Numbers. For

z = 3

(binary-ish tree,

z - 1 = 2

p_{c} = 1/2

. For

z = 4

p_{c} = 1/3

. Step 3 — Order parameter. Let

Q

be the probability that an edge does not lead to an infinite subtree. For each of the

z - 1

children:

Q = (1 - p) + p Q^{z - 1}

. The probability the root is in the infinite cluster is

θ = 1 - Q^{z}

(root has

z

edges). For

z = 3

Q = (1 - p) + p Q^{2}

, giving

Q = \frac{1 - p}{p}

(taking the root other than

Q = 1

) for

p > 1/2

. Then near

p_{c}

θ \approx 6 (p - p_{c})

to leading order. Result:

p_{c} = \frac{1}{z - 1}, θ (p) \sim const (p - p_{c})^{β}, β = 1 (mean-field) .

The Bethe lattice gives the mean-field exponents (

β = 1

γ = 1

δ = 2

), valid also on

Z^{d}

for

d \geq d_{c} = 6

2D Square Lattice — Bond Percolation Threshold by Duality

Given: Bond percolation on

Z^{2}

. We want

p_{c}

. Step 1 — Self-duality. The dual lattice of

Z^{2}

is again

Z^{2}

(shifted). Each open primal edge corresponds to a closed dual edge crossing it, so the dual has parameter

1 - p

. A primal infinite cluster fails to exist iff a dual infinite cluster exists. Step 2 — Threshold relation. Self-duality and the Burton–Keane uniqueness/sharpness arguments imply

p_{c} + p_{c}^{dual} = 1

with

p_{c}^{dual} = p_{c}

(same lattice), so:

p_{c} = 1 - p_{c} \Rightarrow p_{c} = \frac{1}{2} .

This was rigorously proven by Kesten (1980). Step 3 — Numerical check (triangular site percolation for contrast). On the triangular lattice site percolation,

p_{c} = 1/2

exactly (also by self-matching); on the square site lattice

p_{c} \approx 0.5927

(no exact form known). Result:

p_{c} (Z^{2}, bond) = \frac{1}{2} (exact, Kesten 1980).

p_{c}

the probability of a left-right open crossing of an

n \times n

box tends to a nontrivial constant (Russo–Seymour–Welsh theory), confirming criticality.

Common Misconceptions

" $p_{c}$ depends only on dimension." False. The threshold

p_{c}

is highly non-universal — it depends on the lattice (square bond

1/2

, triangular site

1/2

, square site

\approx 0.5927

Z^{3}

bond

\approx 0.2488

). What is universal (depends only on dimension and broad symmetry class) are the critical exponents

β, γ, ν, δ

and scaling functions. Do not conflate the location of the transition with its critical behavior. "At $p_{c}$ there is a marginal/half-infinite cluster." It is not generally obvious that

θ (p_{c}) = 0

. This is a theorem (Harris–Kesten in

d = 2

; Hara–Slade for

d \geq 11

) and remains open for intermediate dimensions

3 \leq d \leq 10

. Assuming

θ (p_{c}) = 0

in those dimensions is unjustified. Long-range percolation breaks the " $0 < p_{c} < 1$ " intuition. Consider one-dimensional percolation on

Z

where each pair

{x, y}

is open independently with probability

p_{x y} = 1 - exp (- β ∣ x - y ∣^{- s})

. For

s = 2

and large enough

β

, an infinite cluster appears even in $d = 1$ , with a discontinuous order parameter (Aizenman–Newman):

θ (p_{c}^{+}) > 0

. This contradicts the naive expectation that 1D never percolates and that transitions are always continuous — locality of interactions is essential to the standard nearest-neighbor picture.

Connections

Percolation sits at a crossroads of probability and statistical mechanics. It is the $q \to 1$ limit of the random-cluster (Fortuin–Kasteleyn) model, which interpolates between percolation, the Ising Model ( $q = 2$ ), and Potts models ( $q \geq 2$ ); FK representation transfers correlation inequalities (FKG, BK) between them. Cluster growth is governed by branching processes (exact on trees) and connects to Galton–Watson Processes. In two dimensions, the critical interfaces converge to Schramm–Loewner Evolution $SLE_{6}$ (Smirnov's theorem for triangular site percolation), tying percolation to Conformal Field Theory with central charge $c = 0$ . The mean-field regime $d > 6$ is analyzed by the lace expansion, the same technique used for the self-avoiding walk. Dynamic versions connect to Interacting Particle Systems and the contact process for epidemics.

Real-World Applications

Porous media / oil recovery. Bond percolation models fluid flow through rock; $p_{c}$ predicts when an oil reservoir becomes hydraulically connected enough to drain — the founding application (Broadbent–Hammersley).
Epidemiology. The SIR epidemic on a contact network maps exactly onto bond percolation: the basic reproduction threshold $R_{0} = 1$ corresponds to $p = p_{c}$ , and the giant outbreak corresponds to the infinite cluster. Used in vaccination/immunization strategy design.
Network robustness / Internet resilience. Random removal of nodes (site percolation) on scale-free networks: such networks have $p_{c} \to 0$ , explaining extreme robustness to random failure but fragility to targeted attack (Cohen, Havlin, ben-Avraham).
Conductivity of composites. A random mixture of conducting and insulating grains conducts macroscopically only above $p_{c}$ ; conductivity $σ \sim (p - p_{c})^{t}$ governs carbon-black-filled polymers and printed electronics.
Forest fires and spread of infestation. Fire spreads to neighboring trees with probability $p$ ; whether a whole forest burns is a supercritical-percolation question, informing firebreak spacing.
Gelation and polymerization. The sol–gel transition (formation of a system-spanning polymer network) is a percolation transition on the Bethe lattice in the Flory–Stockmayer theory.
Materials with cracks / fracture networks. Connectivity of fracture networks in geothermal and groundwater modeling uses continuum percolation thresholds.

Historical Development

1894 / 1941 (precursors): Flory and Stockmayer model polymer gelation on a tree — implicitly the Bethe-lattice percolation transition, decades before the lattice formulation.
1957: Broadbent & Hammersley publish "Percolation processes I", formally defining bond percolation and the threshold $p_{c}$ , motivated by fluid flow through random media.
1960: Harris proves $θ (1/2) = 0$ for bond percolation on $Z^{2}$ , hence $p_{c} \geq 1/2$ , using the FKG-type correlation inequality avant la lettre.
1971–1972: Fortuin & Kasteleyn introduce the random-cluster model, unifying percolation with Ising/Potts.
1980: Kesten proves $p_{c} (Z^{2}, bond) = 1/2$ exactly — a landmark, completing the Harris bound.
1987: Burton & Keane give a clean proof of uniqueness of the infinite cluster via translation-invariance and trifurcation counting.
1986–1987: Aizenman & Barsky, and independently Menshikov, prove sharpness of the phase transition (exponential subcritical decay).
1990–2000: Hara & Slade develop the lace expansion, establishing mean-field exponents and $θ (p_{c}) = 0$ for high dimensions ( $d \geq 19$ , later improved to $d \geq 11$ ).
2001: Smirnov proves conformal invariance of critical site percolation on the triangular lattice and convergence of interfaces to $SLE_{6}$ , confirming Cardy's crossing formula.
2016: Duminil-Copin & Tassion give a short, robust new proof of sharpness valid on general transitive graphs.

Subtleties & Limits

Continuity at criticality in dimensions $3 \leq d \leq 10$ . Whether

θ (p_{c}) = 0

for nearest-neighbor percolation on

Z^{d}

is open in these intermediate dimensions; it is proven only in

d = 2

and

d \geq 11

. Conformal invariance for $d = 2$ general lattices. Smirnov's theorem is established for the triangular site model; universality of the conformal scaling limit (independence from lattice) is conjectural beyond special cases. Upper critical dimension and exponents. The mean-field exponents are rigorous for

d > 6

via the lace expansion, but the predicted values at and below

d = 6

(logarithmic corrections, non-mean-field exponents in

d = 3, 4, 5

) lack rigorous derivation. Exact $p_{c}$ values for most three-dimensional lattices remain unknown analytically. Existence of the scaling limit in $d = 3$ and any conformal-type structure there is a major open frontier.