Percolation Theory

Dr. Kim Christensen

Blackett Laboratory
Imperial College London
Prince Consort Road
SW7 2BW London
United Kingdom

October 9, 2002
Aim

The aim of the percolation theory course is to provide a challenging and stimulating introduction
to a selection of topics within modern theoretical condensed matter physics.
Percolation theory is the simplest model displaying a phase transition. The analytic solutions
to 1d and mean-field percolation are presented. While percolation cannot be solved exactly for
intermediate dimensions, the model enables the reader to become familiar with important concepts
such as fractals, scaling, and renormalisation group theory in a very intuitive way.
The text is accompanied by exercises with solutions and visual interactive simulations for the
percolation theory model to allow the readers to experience the behaviour, in the spirit ”seeing is be-
lieving”. The animations can be downloaded via the URL http://www.cmth.ph.ic.ac.uk/kim/cmth/
I greatly appriciate the suggestions and comments provided by Nicholas Moloney and Ole Peters
without whom, the text would have been incomprehensible and flooded with mistakes. However,
if you still are able to find any misprints, misspellings and mistakes in the notes, I would be very
grateful if you would report those to [email protected].

1
Contents

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Percolation in 1d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Percolation in the Bethe Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Cluster Number Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Cluster Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.6.1 Cluster Radius and Fractal Dimension . . . . . . . . . . . . . . . . . . . . . . 22
1.6.2 Finite Boxing of Percolating Clusters . . . . . . . . . . . . . . . . . . . . . . . 24
1.6.3 Mass of the Percolating Cluster . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.7 Fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.8 Finite-size scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.9 Real space renormalisation in percolation theory . . . . . . . . . . . . . . . . . . . . 31
1.9.1 Renormalisation group transformation in 1d. . . . . . . . . . . . . . . . . . . 34
1.9.2 Renormalisation group transformation on 2d triangular lattice. . . . . . . . . 35
1.9.3 Renormalisation group transformation on 2d square lattice of bond percolation. 36
1.9.4 Why is the renormalisation group transformation not exact? . . . . . . . . . 37

2
1.1 Introduction
Percolation theory is the simplest not exactly solved model displaying a phase transition. Often,
the insight into the percolation theory problem facilitates the understanding of many other physical
systems. Moreover, the concept of fractals, which is intimately related to the percolation theory
problem, is of general interest as it pops up more or less everywhere in Nature. The knowledge of
percolation, fractals, and scaling are of immense importance theoretically in such diverse fields as
biology, physics, and geophysics and also of practical importance in e.g. oil recovery. We will begin
gently by developing a basic understanding of percolation theory, providing a natural introduction
to the concept of scaling and renormalisation group theory.

1.2 Preliminaries
Let P (A) denote the probability for an event A and P (A 1 ∩ A2 ) the joint probability for event A1
and A2 .
Definition 1 Two events A1 and A2 are independent ⇔ P (A1 ∩ A2 ) = P (A1 )P (A2 ).

Definition 2 More generally, we define n ≥ 3 events A 1 , A2 , . . . , An to be mutually independent

if P (A1 ∩ A2 ∩ · · · ∩ An ) = P (A1 )P (A2 ) · · · P (An ) and if any subcollection containing at least two
but fewer than n events are mutually independent.

Let each site in a lattice be occupied at random with probability p, that is, each site is occupied
(with probability p) or empty (with probability 1−p) independent of the status (empty or occupied)
of any of the other sites in the lattice. We call p the occupation probability or the concentration.
Definition 3 A cluster is a group of nearest neighbouring occupied sites.
Percolation theory deals with the numbers and properties of the clusters formed when sites are
occupied with probability p, see Fig. (1.1).

Figure 1.1: Percolation in 2d square lattice of linear size L = 5. Sites are occupied with probability
p. In the lattice above, we have one cluster of size 7, a cluster of size 3 and two clusters of size 1
(isolated sites).

Definition 4 The cluster number ns (p) denotes the number of s-clusters per lattice site.
The (average) number of clusters of size s in a hypercubic lattice of linear size L is L d ns (p), d being
the dimensionality of the lattice. Defining the cluster number per lattice site as opposed to the
total number of s-clusters in the lattice ensures that the quantity will be independent of the lattice
size L.

3
 For finite lattices L < ∞, it is intuitively clear, that if the occupation probability p is small,
there is only a very tiny chance of having a cluster percolating between two opposite boundaries
(i.e., in 2d, from top-to-bottom or from left-to-right). For p close to 1, we almost certainly will
have a cluster percolating through the system. In Fig. 1.2, sites in 2d square lattices are occupied
at random with increasing occupation probability p. The occupied sites are shown in gray while
the sites belonging to the largest cluster are shown in black. Unoccupied sites are white. Note that
for p ≈ 0.59, a percolating cluster appears for the first time.

Figure 1.2: Percolation in 2d square lattices with system size L × L = 150 × 150. Occupation prob-
ability p = 0.45, 0.55, 0.59, 0.65, and 0.75, respectively. Notice, that the largest cluster percolates
through the lattice from top to bottom in this example when p ≥ 0.59.

Definition 5 The percolation threshold p c is the concentration (occupation probability) p at which

an infinite cluster appears for the first time in an infinite lattice.

Note, that pc is defined with respect to an infinite lattice, that is, in the limit of L → ∞. Table
(1.1) lists the percolation threshold in various lattices and dimensions.

Exercise 1 Why is pc not well defined in a finite lattice?

1.3 Percolation in 1d
We will consider the percolation problem in 1d where it can be solved analytically. Many of the
characteristic features encountered in higher dimensions are present in 1d as well, if we know

4
 Lattice # nn Site percolation Bond percolation
1d 2 1 1
2d Honeycomb 3 0.6962 1 − 2 sin(π/18) ≈ 0.65271
2d Square 4 0.592746 1/2
2d Triangular 6 1/2 2 sin(π/18) ≈ 0.34729
3d Diamond 4 0.43 0.388
3d Simple cubic 6 0.3116 0.2488
3d BCC 8 0.246 0.1803
3d FCC 12 0.198 0.119
4d Hypercubic 8 0.197 0.1601
5d Hypercubic 10 0.141 0.1182
6d Hypercubic 12 0.107 0.0942
7d Hypercubic 14 0.089 0.0787
Bethe lattice z 1/(z-1) 1/(z-1)

Table 1.1: The percolation threshold for the site percolation problem is given in column 3 for various
lattices in various dimensions. Column 2 lists the number of nearest-neighbours (nn), also known
as the coordination number. Within a given dimension, the percolation threshold decrease with
increasing number of nearest-neighbours. The site percolation problem has a counterpart called
the bond percolation problem: In a lattice, each bond between neighbouring lattice sites can be
occupied (open) with probability p and empty (closed) with probability (1−p). A cluster is a group
of connected occupied (open) bonds. NB: In all cases, a cluster is defined as a group of nearest
neighbouring occupied sites (bonds). Note that the percolation threshold for the site-percolation on
high-dimensional hypercubic lattices, where loops become irrelevant, approaches that of the Bethe
lattice 1/(z − 1), if we substitute the coordination number z with 2d.

where and how to look. Thus the 1d case serves as a transparent window into the world of phase
transitions, scaling, scaling relations, and renormalisation group theory.
Imagine a 1d lattice with an infinite number of sites of equal spacing arranged in a line. Each
site has a probability p of being occupied, and consequently 1 − p of being empty (not occupied).
These are the only two states possible, see Fig. 1.3.

Figure 1.3: Percolation in a 1d lattice. Sites are occupied with probability p. The crosses are empty
sites, the solid circles are occupied sites. In the part of the infinite 1d lattice shown above, there is
one cluster of size 5, one cluster of size 2, and three clusters of size 1.

What is of interest to us now and in future discussions in higher dimensions is the occupation
probability at which an infinite cluster is obtained for the first time. A percolating cluster in 1d
spans from −∞ to +∞. Clearly, in 1d this can only be achieved if all sites are occupied, that is,
the percolation threshold pc = 1, as a single empty site would prevent a cluster to percolate.
A precise “mathematical” derivation of p c = 1 in 1d goes as follows.

Definition 6 Let Π(p, L) denote the probability that a lattice of linear size L percolates at concen-
tration p.

5
Combining the two definitions (5) and (6), we have,
(
0 for p < pc
lim Π(p, L) =
L→∞ 1 for p ≥ pc .
Consider a finite 1d lattice of size L where each site is occupied with probability p. As the
events of occupying sites are independent, all sites are occupied with probability Π(p, L) = p L , see
Fig. (1.4), and (
0 for p < 1
lim Π(p, L) = lim pL =
L→∞ L→∞ 1 for p = 1,
implying pc = 1.

1.0
L=1
L=2
L=3
L=4
L=5
L = 10
L = 20
Π(p,L)

L = 50
0.5 L = 100

0.0
0.0 0.5 1.0
Occupation probability p

Figure 1.4: The probability of a 1d lattice of linear size L to percolation at occupation probability
p. In the limit L → ∞, Π(p, L) converges to a discontinous step function.

Let us consider the clusters formed in a 1d lattice. A cluster of size s is formed when s sites
are occupied next to one another bounded by two empty sites, see Fig. 1.3. When L → ∞, we
can ignore the effects of the boundary sites of the lattice and the probability of an arbitrary site
(occupied or not) being, say, the left hand side (LHS) of an s-cluster is
ns (p) = (1 − p)ps (1 − p) = (1 − p)2 ps . (1.1)
This expression is obtained from the assumption that the occupancy of each site is independent
of the state of any other site. If this was not the case then it would be much more complicated.
Note that, since all sites have equal probability of being occupied (or empty), the probability that
an arbitrary site is part of an s-cluster is s times the probability of it being the LHS of the cluster.
We can re-write the cluster number Eq.(1.1) for 1d percolation as
ns (p) = (1 − p)2 ps
= (1 − p)2 exp(ln(ps ))
= (1 − p)2 exp(s ln(p))
s
= (pc − p)2 exp(− ) (1.2)
sξ
with the definition of a cutoff cluster size or characteristic cluster size
−1 −1 1
sξ = = → = (pc − p)−1 for p → pc , (1.3)
ln(p) ln(pc − (pc − p)) pc − p

6
where we, to obtain the limit, have used p c = 1 and the Taylor expansion
1 1
ln(1 − x) = −x − x2 − x3 − · · · ≈ −x
2 3
where the last approximation is valid for x → 0, see Fig. 1.5.

0 0
10 10
p = 0.7
p = 0.7 p = 0.9
p = 0.9 p = 0.95
−2 p = 0.95
10 p = 0.99
p = 0.99
−2
10
−4
10

s ns(p)
ns(p)

2
−6
10
−4
10
−8
10

−10 −6
10 0 1 2 3 10
10 10 10 10 0 0 1 10 100
Cluster size s s(pc − p) = s/sξ

4
500 10
−1/ln(p) −1/ln(p)
1/(pc − p) 1/(pc − p)
3
400 10

2
300 10
sξ

sξ

1
200 10

0
100 10

−1
0 10 −4 −3 −2 −1 0
0.0 0.2 0.4 0.6 0.8 1.0 10 10 10 10 10
Occupation probability p pc − p
Figure 1.5: Percolation in 1d. (a) The cluster number distribution n s (p) = (pc − p)2 exp(− ssξ ) for
various values of p approaching pc = 1. The vertical lines indicate the cutoff cluster size s ξ (p). (b)
By plotting s2 ns (p) versus s/sξ ≈ s(pc − p), all the data collapses onto a function x 2 exp(−x). (c)
1
The characteristic cluster size sξ = − ln(p) diverges when p → pc = 1. (d) In the limit p → pc = 1,
−1
sξ ∝ (pc − p) .

Exercise 2 Verify the Taylor expansion of ln(1 − x) around x = 0 given above.

Thus the cutoff cluster size sξ diverges for p → pc as a power law in the distance from the critical
occupation probability pc , see Fig. 1.5. The divergence of the cutoff cluster size when p → p c is also
seen in higher dimensions where, however, another numerical exponent will describe the divergence.
Thus it is natural to introduce a symbol for the exponent.
Definition 7 The critical exponent σ is defined by
1
sξ ∝ |pc − p|− σ for p → pc . (1.4)

7
In 1d percolation theory, σ = 1 and log s ξ = constant − σ1 log |pc − p|, see Fig. 1.5.
Let us continue our journey into 1d percolation theory. For p < p c we can state that the
probability that an arbitrary site belongs to any (finite) cluster is simply the probability p of it
being occupied. Since the probability that an arbitrary sites belongs to an s-cluster is given by
sns (p), we arrive at
∞
X
sns (p) = p for p < pc . (1.5)
s=1

Using the formula for summing a geometric series, we can satisfy those who prefer a rigorous
mathematical proof:
∞
X ∞
X
sns (p) = s(1 − p)2 ps
s=1 s=1
∞
X
2 d(ps )
= (1 − p) p
s=1
dp
∞
!
d X
2 s
= (1 − p) p p
dp s=1
 
d p
= (1 − p)2 p
dp 1−p
= p.

How large on average is a cluster or, equivalently, how large is a cluster on average to which
an occupied site belongs? The probability that a site is occupied is p. The probability that an
arbitrary site belongs to an s-cluster is sn s (p). Thus the probability ws that the cluster to which
an occupied sites belongs contains s sites is

sns (p) sns (p)

ws = = P∞ .
p s=1 sns (p)

Thus, the mean cluster size or average cluster size S(p) is given by
∞
X
S(p) = sws
s=1
X∞
s2 ns (p)
= P∞
s=1 s=1 ns (p)s
X∞
1
= (1 − p)2 s2 ps
p s=1
 2 ∞
!
1 d X
= (1 − p)2 p p s
,
p dp s=1

where the operator

 2   
d d d d2
p = p p 6= p2 .
dp dp dp dp2
d
Using the formula for summing a geometric series and the operator p dp twice, we finally arrive at

1+p pc + p
S(p) = = (1.6)
1−p pc − p
where the last equality follows because in 1d the critical occupation probability p c = 1.

8
Exercise 3 Derive the result in Eq.(1.6) for the mean cluster size in 1d.
We thus see that the mean cluster size diverges for p → p −
c , where the minus sign signifies that we
are approaching pc from below, which is what we intuitively expect if considering an infinite lattice.
It is not possible to approach pc from above in 1d as pc = 1. This is actually the main difference
between 1d and higher dimensions, where, p c < 1 and we can approach pc from both below and
above.
In order to investigate in detail how the mean cluster size diverges, when taking the limit
p → p− c , we note that the numerator in Eq.(1.6) approaches 2p c , so

pc + p 2pc
S(p) = → ∝ (pc − p)−1 for p → p−
c . (1.7)
pc − p pc − p
Thus in 1d, the mean cluster diverges like a power law in the quantity (p c − p) when p → pc , see
Fig. 1.6. The same phenomenon will be encountered in higher dimensions.
Definition 8 The critical exponent γ is defined by

S(p) ∝ |pc − p|−γ for p → pc . (1.8)

In 1d percolation theory, γ = 1.

4 4
10 10
(1 + p)/(1 − p) = (pc + p)/(pc − p) (1 + p)/(1 − p) = (pc + p)/(pc − p)

3 3
10 10
Average cluster size S

Average cluster size S

2 2
10 10

1 1
10 10

0 0
10 10

−1 −1
10 −4 −3 −2 −1 0 10 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10
Occupation probability p pc − p
Figure 1.6: Percolation in 1d. The average cluster size S(p) = (1 + p)/(1 − p) = (p c + p)/(pc − p)
diverges when p → pc = 1. In the limit p → pc = 1, S(p) → 2/(pc − p).

Definition 9 The correlation function or pair connectivity g(r) is the probability that a site at
position r from an occupied site belongs to the same finite cluster.
Note this definition excludes the contribution from the infinite cluster. That need not worry us in
1d, where all clusters are finite if p < p c = 1. Let r = |r|. Clearly, g(r = 0) = 1, since the site is
occupied by definition. In 1d, for a site at position r to be occupied and belong to the same (finite)
cluster, this site and the (r − 1) intermediate sites must be occupied, leaving

g(r) = pr ,

for all p, which can also be expressed in the form

 
r r
g(r) = exp(ln(p )) = exp(r ln(p)) = exp − , (1.9)
ξ

9
where
1 −1 1
ξ=− = → = (pc − p)−1 for p → pc = 1, (1.10)
ln(p) ln(pc − (pc − p)) (pc − p)

where we use the expansion ln(1 − x) ≈ −x for small x. The quantity ξ is called the correlation
length which diverges for p → pc . The same phenomenon will be encountered in higher dimensions.
Note that in 1d we have sξ = ξ which is why we haven’t bothered displaying a figure for ξ, as it
would be identical to Fig. 1.5. However, this identity will not be true in higher dimensions, where
we will find sξ ∝ ξ D , where D is the fractal dimension, but more about this later.

Definition 10 The critical exponent ν is defined by

ξ ∝ |pc − p|−ν for p → pc . (1.11)

In 1d percolation theory, ν = 1.
By summing over all possible lattice sites r of the correlation function, the mean cluster size
can be shown to be X
g(r) = S(p). (1.12)
r

In 1d, this sum is straight forward identifying r with r = 0, ±1, ±2, . . ., see Problem 2, where you
will also discover, that the sum rule is valid in all dimensions d.
The general pattern for the exact solutions of the 1d percolation problem is that certain quan-
tities, such as the cutoff cluster size s ξ , the mean cluster size S(p), and the correlation length ξ
diverge at the percolation threshold. The divergence can be described by simple power laws of
the distance from the critical occupation probability |∆p| = |p c − p|, e.g. ξ ∝ (pc − p)−1 , at least
asymptotically close to pc where ∆p is small. The same phenomena will be encountered in higher
dimensions even though we cannot obtain exact analytic solutions.

1.4 Percolation in the Bethe Lattice

The percolation problem can be solved analytically in d = 1 and d = ∞. The infinitely dimensional
case is synonymous with the Bethe lattice, a special lattice where each site has z neighbouring
sites, such that each branch gives rise to z − 1 other branches, see Fig. 1.7.

Figure 1.7: The Bethe lattice with z = 3. Each site has three neighbours. Each branch contains
z − 1 = 2 subbranches.

10
 The 1d case is effectively a Bethe lattice with z = 2. Why does the Bethe lattice correspond to
the spatial dimension d = ∞ you might rightly ask! Well, in a hypercubic lattice, (a) the number
of surface sites relative to the total number of sites approaches a constant when d → ∞ and (b)
there are no closed loops when d → ∞. The Bethe lattice has both these properties.
(a) Let g denote the generation, that is, the distance from a “centre site”. Note, however, that in
an infinite Bethe lattice, all sites are equivalent, so the notion of a “centre site” is not to be
taken literally. In the figure above, the first “ring” of three sites belong to generation g = 1,
the second “ring” of six sites belongs to the second generation and so on. The total number
of sites in a Bethe lattice consisting of g generations is

Total no. sites = 1 + 3 · (1 + 2 + · · · + 2g−1 )

1 − 2g
= 1+3·
1−2
= 3 · 2g − 2,

while the number of surface sites is 3 · 2 g−1 . Thus

No. of surface sites 3 · 2g−1 1
= g
→ for g → ∞.
Total no. of sites 3·2 −2 2
and the surface/volume tends to a constant.

Exercise 4 Show, that for a general Bethe lattice with coordination number z
No. of surface sites z−2
→ for g → ∞.
Total no. of sites z−1

In a hypercubic lattice of linear size L, the surface is proportional to the volume only when
d → ∞: the surface in d dimensions is proportional to L d−1 while the volume is proportional
to Ld leaving
d−1 1
Surface ∝ Volume d = Volume1− d ,
that is, the surface is proportional to the volume if d → ∞.
(b) There are no closed loops in a Bethe lattice. Starting from the “centre site” going outwards,
one will never return to the starting point. In a hypercubic lattice, the chance (probability)
of having a loop approaches zero as the dimension d → ∞: As an example, let us place four
particles in a chain in a hypercubic lattice with dimension d. When the first particle has
been placed, there are 2d nearest neighbour sites, where the second particle can be placed.
However, for the third and fourth particle, there are only 2d − 1 possible sites, implying a
total no. of different chains 2d · (2d − 1) 2 . Calculating the number of ways to place four
particles in a loop, we arrive at 2d · (2d − 2) · 1, that is, the probability of having a four loop
No. loops 2d · (2d − 2) · 1 (2d − 2)
= 2
= →0 for d → ∞.
Total no. chains 2d · (2d − 1) (2d − 1)2
For d = ∞, there is no chance of having a closed loop, which is intuitively clear, isn’t it?
What is the critical occupation probability in a Bethe lattice, that is, at which occupation
probability does an infinite cluster (path) appear for the first time in an infinite Bethe lattice?
Starting from a “centre site” and going outwards, we encounter (z − 1) new neighbours. Thus,
on average, we have p(z − 1) new occupied sites on which we can continue the path. The critical
occupation probability is determined by the equation
1
pc (z − 1) = 1 ⇔ pc = .
z−1

11
Notice, that for z = 2, the 1d percolation problem, we recover p c = 1. For z = 3, e.g., pc = 12 . Note
that in a hypercubic lattice with d → ∞ one would expect p c → 1/(2d − 1) in order to be able to
continue walking along a path of occupied sites.

Definition 11 The strength of the infinite cluster P (p) is the probability that an arbitrary site
belongs to the infinite cluster.

Exercise 5 Discuss why, obviously, P (p) = 0 for p < p c in an infinite lattice.

In order to calculate P (p) for p > pc in the Bethe lattice, we introduce the quantity Q as the
probability that an arbitrary site is NOT connected to infinity through a fixed branch originating
at this site. Restricting ourselves to the lattice with z = 3 and using basic probability theory, we
have that the strength P (p) of the infinite network, that is, the probability that an arbitraryly
selected site is connected to infinity by occupied sites is

P (p) = (Prob. site is occupied) × (Prob. at least ONE branch lead to infinity)
= p(1 − Q3 ).

Next, we need to determine Q, the probability that a branch originating at that site does not
lead to infinity. This depends on whether or not the neighbouring site is occupied. We find

Q = (Prob. site is empty) + (Prob. site is occupied) × (Prob. no subbranch leads to infinity)
= (1 − p) + pQ2 .

We have relied on the fact that all sites in a Bethe lattice are equivalent, so Q is also the probability
that a subbranch is not connected to infinity. The quadratic equation can easily be solved:
p (
1± (2p − 1)2 1
Q= = 1−p
2p p .

Exercise 6 Show that, in a general Bethe lattice where each site has z neighbours (Problem 4),
(
1
Q= 2p(z−1)−2
1− p(z−1)(z−2) .

When p < pc , there are no infinite clusters, by definition, so with probability 1 there is no connection
to infinity. Thus Q = 1 is the trivial solution associated with P (p) = 0 for p < p c . The non-trivial
solution Q = 1−p
p is associated with p > pc , leaving

 0 for p < pc
P (p) =   
1−p 3
 p 1− p for p ≥ pc .

Using a Taylor expansion for P (p) around p = p c = 12 , it can be shown that (see Problem 4)

P (p) ∝ (p − pc ) for p → p+
c .

Definition 12 The critical exponent β is defined by

P (p) ∝ (p − pc )β for p → p+
c . (1.13)

12
 0
1.0 10
Analytic P(p)
p − pc = 0.0025
P(p) = 6 (p − pc)
0.8 −1
10

Strength P
Strength P

0.6
−2
10
0.4

−3
10
0.2

−4
0.0 10 −5 −4 −3 −2 −1 0
0.0 0.2 0.4 0.6 0.8 1.0 10 10 10 10 10 10
Occupation probability p p − pc
Figure 1.8: The strength P (p) of the infinite cluster in a Bethe lattice with z = 3. When p → p +
c ,
the strength P (p) → 6(p − pc ) implying β = 1 in a Bethe lattice.

In the Bethe lattice, β = 1, see Fig. 1.8.

The strength P (p) measures how large a fraction of the sites in the lattice belongs to the infinite
cluster (in an infinite lattice) and is known as the order parameter. The phenomenon that the order
parameter (strength) becomes non-zero for p > p c is known as a phase transition and p = p c is
often called a critical point or critical occupation probability.
Exercise 7 What is the strength P (p) in the 1d percolation problem?
What is the mean cluster size S(p) to which an occupied site (e.g. the “centre site”) belongs? Let
T be the mean cluster size in one branch.
S(p) = average cluster size to which the origin belongs
= (contribution from origin) + (contributions from the 3 branches)
= 1 + 3T.
Thus, we have to determine T . Again, as all sites are equivalent in an infinite Bethe lattice, T will
also denote the mean cluster size in the subbranch, so we can argue as follows: If the neighbour
to the “centre site” is empty, probability (1 − p), there is no contribution to T from this branch. If
the neighbour to the “centre site” is occupied (probability p), it contributes its own mass (unity)
to the cluster, and adds the mass T for each of its z − 1 = 2 subbranches. Thus,
p
T = (1 − p) · 0 + p · (1 + 2T ) ⇔ T = for p < pc .
1 − 2p
Therefore, the mean cluster size S(p) is given by
1+p 1+p 1+p
S(p) = 1 + 3T = = 1 = = Γ2 (pc − p)−1 , (1.14)
(1 − 2p) 2( 2 − p) 2(pc − p)

where Γ2 = 1+p 2 is the amplitude, and the critical exponent γ = 1 in the Bethe lattice. This is the
exact result for a mean cluster of size S(p) and we notice that it diverges for p → p c .
An occupied site either belongs to the infinite cluster or a finite cluster. Thus, the generalisation
of Eq.(1.5) is valid for all p
∞
X
P (p) + sns (p) = p ∀p. (1.15)
s=1

13
The sum runs over all finite s and excludes the infinite cluster. Of course, Eq.(1.15) reduces to
Eq.(1.5) for p < pc since the strength P (p) is identical zero in this range.
P∞
Exercise 8 Reflect upon the identity Eq.(1.15) for p = p c . What is the sum s=1 sns (pc )? Is your
answer also valid in d = 1?
Critical phenomena occur in a variety of systems. The Bethe lattice approximation for percolation
theory is somewhat analogous to the mean-field approximation for magnetic systems dislaying a
phase transition, or the Van der Waals equation for a fluid. In all cases, the order parameter
becomes non-zero above (or below) the critical point. In a magnetic system, the order parameter
is the spontaneous magnetisation per spin m, which is zero above the critical temperature (Curie
temperature) Tc and non-zero for T < Tc as the system enters an ordered state. We will see later,
that m ∝ (Tc − T )β . In the case of a fluid, the order parameter is identified as the the difference
between liquid and vapour density which becomes nonzero as T < T c . The analogy between mean-
field thermal critical phenomena and percolation is not complete since the critical exponent β for
the order parameter is 21 for thermal phase transitions and unity for percolation, but nevertheless,
the qualitative similarities are clear.
In both the thermal phase transitions mentioned above and percolation, the order parameter
goes to zero continuously as one approaches the critical point. Such phase transitions are called
continuous phase transitions or second-order phase transitions. If instead the order parameter
jumps to zero, one has a first order phase transition. Such transitions can occur in more complicated
percolation theory situations, like “bootstrap percolation” where on a square lattice a site remains
occupied only if three or four of its neighbours are still occupied [2].

1.5 Cluster Number Distribution

Unfortunately, it is not possible to obtain an exact form for the cluster number n s (p) in d > 1 or the
Bethe lattice since, unlike in 1d, there are a very large number of different ways in which clusters
can arrange themselves. Even for relatively small cluster numbers in 2d we run into difficulties.
Exact enumeration of small cluster “animals” has been tabulated. In a 2d square lattice a computer
has counted a total of 68, 557, 762, 666, 345, 165, 410, 168, 738 different cluster configurations with
s = 46 [3]. Nevertheless, we are still able to account for the general scaling behaviours of certain
quantities that characterise the clusters.
Definition 13 The perimeter or external boundary t of a cluster is the number of empty nearest
neighbours of the cluster.
Using the concept of a perimeter, we are able write down the general expression for the cluster
number ns (p) as X
ns (p) = gs,t ps (1 − p)t , (1.16)
t
where we have introduced gs,t , the number of different lattice configurations with size s and perime-
ter t.
In 1d, all clusters have two perimeter sites and thus
(
1 for t = 2
gs,t =
0 otherwise

implying ns (p) = ps (1 − p)2 .

Exercise 9 Consider clusters of size s = 3 in 2d on a square lattice. Verify, that g s,t is only
non-zero for t = 7 or t = 8 and that gs=3,t=7 = 4 and gs=3,t=8 = 2, implying ns=3 (p) = 4p3 (1 −
p)7 + 2p3 (1 − p)8 .

14
 In a Bethe lattice, however, we can show that there is a unique perimeter for a given cluster
size s.
Exercise 10 Show that in a Bethe lattice t = 2 + s(z − 2).
Now we apply this formula to our general result for cluster numbers (1.16) to get:
X
ns (p) = gs,t ps (1 − p)t = gs,2+s(z−2) ps (1 − p)2+s(z−2) . (1.17)
t

For simplicity, from here on we shall work with the z = 3 case rather than the general Bethe
lattice. For z = 3 we have pc = 12 . As we have said, determining gs,2+s for large s is difficult, but
we can avoid having to calculate gs,2+s by considering the ratio
 2   s
ns (p) (1 − p) p (1 − p)
=
ns (pc ) (1 − pc ) pc (1 − pc )
 2   
(1 − p) p (1 − p)
= exp s ln .
(1 − pc ) pc (1 − pc )
(1.18)

In order to proceed, define the function

p (1 − p) p (1 − p) 2
f (p) =
pc (1 − pc )
= 1 1 = 4p − 4p for z = 3.
2 (1 − 2 )
1
Exercise 11 Make a Taylor expansion of the function f (p) = 4p − 4p 2 around p = 2 = pc and
verify f (p) = 1 − 4(p − pc )2 .
Substituting the expression for f (p) given in Exercise 11, we find
 2
ns (p) (1 − p) s
= exp(− ) (1.19)
ns (pc ) (1 − pc ) sξ
with
−1
sξ =   2 
1
ln 1 − 4 p − 2
 −2
1 1 1
∝ p− for p → 2
4 2
1
∝ |p − pc |− σ . (1.20)

In the Bethe lattice, σ = 12 .

We now have a very simple exponential decay for the ratio of cluster numbers. This is to say
that large clusters with s  sξ are very rare indeed. This explains why s ξ is known as the cutoff
cluster size. In order to investigate the cluster number distribution in more detail, we have to
consider the term ns (pc ) appearing in the denominator.
From the analysis on the average cluster size, we know that,
∞
X s2 ns (p)
S(p) = P∞ →∞ for p → pc . (1.21)
s=1 s=1 sns (p)

However, the denominator remains finite at p = p c (why?) so the numerator must diverge for
p = pc , that is,
∞
X
s2 ns (pc ) = ∞.
s=1

15
Generally, we have that (
∞
X
a convergent for a < −1
s =
s=1
divergent for a ≥ −1.
Thus a power-law decay
∞
X ∞
X
−τ 2
ns (pc ) ∝ s ⇒ s ns (pc ) = s2−τ
s=1 s=1

would imply a divergence of the average cluster size if the critical exponent τ ≤ 3.
P P∞ 1−τ
However, see Eq.(1.15) and problem (8), we also need p c = ∞ s=1 sns (pc ) ∝ s=1 s <∞
implying that τ > 2. Therefore, we arrive at a general form for the cluster number distribution in
the Bethe lattice:
s
ns (p) ∝ s−τ exp(− ), for s  1 (1.22)
sξ
with
1
sξ ∝ |pc − p|− σ for p → pc , (1.23)
where σ = 12 and 2 < τ ≤ 3.
Actually, we can determine τ as it is not independent of the critical exponents γ and σ deter-
mined previously. Let us evaluate the average cluster size S(p) assuming Eq.(1.22) is valid for all
s. Neglecting the denominator which for p → p c will approach pc and thus be a constant we find
∞
X
S(p) ∝ s2 ns (p)
s=1
X∞
s
∝ s2−τ exp(−)
s=1
s ξ
Z ∞
s
≈ s2−τ exp(− ) ds
1 sξ
Z ∞
= (zsξ )2−τ exp(−z)sξ dz with z = s/sξ
1/sξ
Z ∞
= s3−τ
ξ z 2−τ exp(−z)dz
1/sξ
Z ∞
= s3−τ
ξ z 2−τ exp(−z)dz for p → pc where sξ → ∞
0
= s3−τ
ξ Γ(3 − τ ) (1.24)

where Γ is the Gamma function

Z ∞
Γ(x) = z x−1 exp(−z) dz.
0

Exercise 12 Show that the Gamma function Γ defined by the integral above satisfies the recursion
relation Γ(x + 1) = xΓ(x) and deduce that Γ(x + 1) = x! for x = 0, 1, 2, . . . where 0! = 1.

Using sξ ∝ (p − pc )−2 , we see that

S(p) ∝ (p − pc )2τ −6 ∝ (p − pc )−1 , (1.25)

where we have used Eq.(1.14). Thus we must require that

5
2τ − 6 = −1 ⇔ τ = . (1.26)
2

16
In summary, for the Bethe lattice
5 s
ns (p) ∝ s− 2 exp(− ) s1
sξ
sξ ∝ |p − pc |−2 for p → pc .

The above procedure is a general way of deriving a relationship between critical exponents, as we
shall see shortly.

1
10
0
10 p = 0.25
−1
p = 0.425 10
p = 0.475
p = 0.49 −3
p = 0.4975 10

(1−pc) /(1−p) s ns(p)

−5 −5
10 10

2 τ
ns(p)

−7
10
−9

2
10 p = 0.25
−10 p = 0.425
10 p = 0.475
−11
10 p = 0.49
p = 0.4975
−13
10
−15 −15
10 0 2 4 6 10 −5 −4 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
2
Cluster size s s 4(pc − p) = s/sξ
h i2
(1−p)
Figure 1.9: (a) The cluster number distribution n s (p) = (1−p c)
s−τ exp(− ssξ ), see Eq.(1.19) for
different values of p approaching pc = 1/(z − 1) = 1/2. The vertical lines indicate the cutoff cluster
size sξ (p) for the various p values. The cluster number distribution is well approximated by a
h i
(1−pc ) 2 τ
power law for s  sξ while is decays rapidly for s  sξ . (b) By plotting (1−p) s ns (p) versus
s/sξ ≈ s · 4(pc − p)2 , all the data collapses onto the scaling function e −x .

Now, we postulate a scaling ansatz (form) valid for all p and large s

Ansatz 1
s
ns (p) ∝ s−τ exp(− ), s  1,
sξ
1
sξ ∝ |p − pc |− σ p → pc , (1.27)

or equivalently
1
ns (p) ∝ s−τ exp(−s|p − pc | σ ) s  1 and p → pc . (1.28)

Note the role of sξ as a cutoff cluster size, where ns (p) is characterised by

(
s−τ for s  sξ
ns (p) ∝ (1.29)
decays rapidly for s  sξ ,

representing the crossover from a behaviour of “critical clusters” (power-law distributed) to that
of non-critical clusters. Note also that n s (pc ) = s−τ as sξ = ∞ at p = pc .
We can, however, immediately identify two problems (a third problem will surface shortly) with
the scaling ansatz 1.

17
(1) The scaling ansatz 1 fails to describe the 1d case. We found previously that
s s
ns (p) = (1 − p)2 exp(− ) = (pc − p)2 exp(− ),
sξ sξ

which is not a special case of our ansatz 1 since, instead of a power of s, we have (p c − p)2
in front of the exponential. You might be very tempted to say that τ = 0, but that does
actually not solve the problem, as we shall see below.

(2) For fixed s

∞
X
ns (p) = gs,t ps (1 − p)t (1.30)
t=1

is a polynomial in p with a finite number of terms. Thus all derivatives of n s (p) with respect
to p will remain finite for all p. However, using the scaling ansatz 1 for p > p c ns (p) =
1
As−τ exp(−s(p − pc ) σ ) we find, keeping s constant

dns (p) 1 1 1
= As−τ exp(−s(p − pc ) σ ) (p − pc ) σ −1 (1.31)
dp σ
and by differentiating once more
"  2   #
d2 ns (p) 1 1 1 1 1 1 1

2
= As−τ exp(−s(p − pc ) σ ) (p − pc ) σ −1 + exp(−s(p − pc ) )
σ − 1 (p − pc ) σ −2 .
dp σ σ σ

In 2d, e.g., σ = 0.4 ⇔ σ1 − 2 = 0.5, so both the first and second derivative of n s (p) w.r.t. p
remain finite as p → pc . The third derivative of ns (p), however, will contain a term with the
1
factor (p − pc ) σ −3 which will diverge as p → pc contrary to the derivatives of Eq.(1.30).

We shall later resolve these two problems, so there is no need for you to worry too much at the
moment.
Let us return to the question of deriving scaling relations assuming the scaling ansatz above for
the cluster number distribution. Following the previous calculation of the average cluster size S(p)
and noting that the Gamma function is just a number, we find
τ −3
S(p) ∝ s3−τ
ξ ∝ |p − pc | σ , for p → pc , (1.32)

as
1
sξ ∝ |p − pc |− σ for p → pc . (1.33)
However, by definition
S(p) ∝ |p − pc |−γ for p → pc , (1.34)
implying the scaling relation
3−τ
γ= . (1.35)
σ
We notice again that to ensure the divergence of S(p) for p → p c we must have τ < 3 in order to
leave γ > 0. Also, we recover problem (1) mentioned above because in 1d percolation theory γ = 1
and σ = 1 implying τ = 2. However, there is no s −τ term (yet!) in our expression for the 1d cluster
number distribution ns (p) and obviously the suggested “solution” of assuming τ = 0 is not valid
either.

18
 Now we focus our attention on the percolation strength P (p), starting from Eq.(1.15):
∞
X
P (p) = p − sns (p)
s=1
X∞
= pc − sns (p) + (p − pc )
s=1
∞
X ∞
X
= sns (pc ) − sns (p) + O(p − pc ) (O(p − pc ) = term of order(p − pc ))
s=1 s=1
X∞ ∞
X s
= s1−τ − s1−τ exp(− ) + O(p − pc ) using the scaling ansatz 1
s=1 s=1
sξ
∞
( )
X s
1−τ
∝ s 1 − exp(− ) + O(p − pc ).
s=1
sξ
∞
( )
X s
∝ s1−τ 1 − exp(− ) ∆s, (1.36)
s=1
sξ

where we have dropped the term of order (p − p c ) and introduced ∆s = 1. When p → pc , the cutoff
cluster size sξ → ∞. Thus the term {1−exp(− ssξ )} ≈ 0 for small s and the main contribution to the
integral will come from large s of the order of s ξ . To solve (1.36) we need to replace the summation
by an integral. Going from a discrete sum to a continuous sum does not, of course, create a
mathematical identity, unless one takes certain limits. However, when the main contribution R
from
P
the sum is for large s, it is legitimate to approximate the sum by an integral ∆s ≈ ds, as we
are only interested in the the general scaling behaviour of the system. Thus
Z ( )
∞ s
1−τ
P (p) ∝ s 1 − exp(− ) ds. (1.37)
1 sξ

One might proceed by integration by parts. You will then discover that only one of the terms are
non-zero. However, we will take another approach. We differentiate Eq.(1.37) w.r.t. 1/s ξ to find
Z ∞
dP (p) s
∝ s2−τ exp(− ) ds
d( s1ξ ) 1 sξ
Z ∞
∝ s3−τ
ξ z 2−τ exp(−z) dz substituting z = s/sξ .
1/sξ

As we are ultimately interested in the limit p → p c where the cutoff cluster size diverges, we replace
the lower limit 1/sξ with zero and thus
!τ −3
1
P (p) ∝ (1.38)
sξ

so integrating we find
!τ −2
1
P (p) = c1 + c2 (1.39)
sξ
At p = pc , sξ = ∞. But, P (pc ) = 0 implies that c1 = 0. So,
!τ −2
1
P (p) ∝ c2 ,
sξ
τ −2
∝ |p − pc | σ . (1.40)

19
Hence, we have derived a new scaling relation
τ −2
β= . (1.41)
σ
By definition (
0 p ≤ pc
P (p) = (1.42)
(p − pc )β p → p+c .

Now we find a third problem with our ansatz 1.

(3) In the derivation above we did not assume p > p c . Thus the ansatz 1 also predicts an infinite
cluster for p < pc , which is of course not correct.

Reflecting upon the three problems, we realise that they can all be resolved, as we shall see
1
shortly, by avoiding the argument s|p − p c | σ in the scaling ansatz. Thus let us assume the following
general scaling form

Ansatz 2
ns (p) = q0 s−τ f [q1 (p − pc )sσ ] s  1 and p → pc . (1.43)

The scaling ansatz 2 has the following properties:

• q0 and q1 are proportionality factors and pc the critical occupation probability. They all
depend on the lattice details. They are non-universal.

• τ , σ are the critical scaling indices which are independent of p, p c and the lattice structure.
They are universal.

• f , the scaling function, is independent of p, p c and detailed lattice structure. It is univer-

sal. The precise form of the scaling function f = f (z) has to be determined by computer
simulations or other numerical methods since the function f is not predicted by the scaling
assumption. Typically, however
( 1
constant |z|  1 ⇔ |p − pc |sσ  1 ⇔ s  |p − pc |− σ ∝ sξ
f (z) = (1.44)
decays fast |z|  1 ⇔ s  sξ .

that is, the scaling function f (z) reaches a constant value (except in 1d percolation theory)
for |z|  1 ⇔ s  sξ , and decays rather fast for |z|  1 ⇔ s  s ξ . Thus, the role of sξ as a
cutoff and crossover size is maintained. Later, we shall look more carefully at the crossover
properties.

• τ , σ and f depend on dimensionality.

• By plotting q10 sτ ns (p) versus (p−pc )q1 sσ , all the data will fall on the universal curve outlining
the scaling function f , see Fig. 1.9. This phenomenon is known as data collapse.

Problems 7 and 8 highlights some of the points above.

Let us check whether the general scaling ansatz 2 really resolves the three problems associated
with the former scaling ansatz 1.

20
(1) The 1d cluster number distribution
s
ns (p) = (pc − p)2 exp(− )
sξ
≈ (pc − p)2 exp(−s(pc − p)) for p → pc
= s−2 (s(pc − p))2 exp(−s(pc − p))
= s−2 f [sσ (pc − p)]
(1.45)

which is of the form of the general scaling ansatz 2 with τ = 2 and σ = 1 and f (x) =
x2 exp(−x). Thus the scaling relation Eq.(1.35) is indeed fulfilled. Note, however, that unlike
in higher dimensions, the scaling function increases like x 2 for x  1 but that it indeed decays
rapidly for x  1, see Fig. (1.5).
(2) Keeping s constant we find

∂ k ns (p)
= q0 s−τ f (k) [q1 (p − pc )sσ ]q1k sσk
∂pk
which is not divergent for p → pc assuming that f is analytic, that is, all derivatives of f (z)
with respect to z are finite everywhere, also at z = 0.
(3) Let us calculate the strength of the infinite cluster.
∞
X
−P (p) = sns (p) − p
s=1
X∞
= sns (p) − pc + (pc − p)
s=1
X∞
P
= [sns (p) − sns (pc )] + O(p − pc ) using sns (pc ) = pc
s=1
Z ∞
≈ q0 s1−τ (f [q1 (p − pc )sσ ] − f [0]) ds
1

Now substitute z = q1 (p − pc )sσ . Notice that if p > pc the argument z is positive while
for p < pc the argument is negative. The cluster size s, however, is always positive so
1 1 1 1
s = |z| σ (q1 |p − pc |)− σ implying that ds = σ1 |z| σ −1 (q1 |p − pc |)− σ dz with dz positive. Thus we
find
Z
1−τ τ −1 1 1 1
−P (p) = q0 |z| σ (q1 |p − pc |) [f (z) − f (0)] |z| σ −1 (q1 |p − pc |)− σ dz
σ

Z
σ
τ −2 2−τ 1
= q0 (q1 |p − pc |) σ |z| σ −1 [f (z) − f (0)] dz. (1.46)
σ
Now it is important to realise that the limits of the integral will depend on the value of p
since z = q1 (p − pc )sσ . If p > pc , the upper limit will be ∞ (lower limit p − p c approaches 0 as
p → p+ −
c ) while for p < pc , the lower limit will be −∞ (upper limit approaching 0 as p → p c ).
Thus if Z (
2−τ
−1 1 constant 6= 0 for p > pc
|z| σ [f (z) − f (0)] dz = (1.47)
σ 0 for p < pc
we have solved the third problem as well, implying
(
(p − pc )β for p → p+c
P (p) ∝
0 for p < pc

21
 with
τ −2
β= .
σ
Exercise 13 Discuss the qualitative behaviour of the scaling function f (z) as a function of z =
q1 (p − pc )sσ in order to satisfy the constraint given in Eq.(1.47).

Problem 8 will ask you to calculate the average cluster size S(p) assuming the general scaling
ansatz 2.

1.6 Cluster Structure

Till now we have only considered the distribution of cluster sizes and have come up with some useful
scaling laws and scaling relations. Now we turn our attention to the geometry of the clusters. We
shall do so by mainly studying the fractal geometry, which contains information about the density
of the clusters on different length scales.

1.6.1 Cluster Radius and Fractal Dimension

Let ri denote the position of the ith occupied site in a cluster of size s.
1 Ps
Definition 14 The centre of mass of an s-cluster r cm = s i=1 ri .

Definition 15 The radius of gyration R s of a given s-cluster is defined by

Rs2 = average square distance to the centre of mass

= h|ri − rcm |2 i
s
1X
= |ri − rcm |2
s i=1

We want to prove that

1 1 1 X
Rs2 = average square distance between two cluster sites = |ri − rj |2 . (1.48)
2 2 s2 ij

By definition, the lefthand side of Eq.(1.48),

s
1X
Rs2 = |ri − rcm |2
s i=1
s
1X
= (ri · ri + rcm · rcm − 2 ri · rcm )
s i=1
s s s
1X 1X 2X
= ri · r i + rcm · rcm − ri · rcm
s i=1 s i=1 s i=1
s s
1X 1X
= ri · ri + rcm · rcm − 2 ( ri ) · rcm
s i=1 s i=1
s
1X
= ri · ri − rcm · rcm .
s i=1

22
 Let us rewrite the righthand side of Eq.(1.48)
1 1 X
Rs2 = |ri − rj |2
2 s2 ij
s X s
1 1 X
= (ri · ri + rj · rj − 2 ri · rj )
2 s2 i=1 j=1
 !  
s
X s
X s
X s
X
1 1 
= s ri · ri + s rj · rj − 2 ri ·  r j 
2 s2 i=1 j=1 i=1 j=1
s
1X 1
= ri · r i − (s rcm ) · (s rcm )
s i=1
s2
s
1X
= ri · ri − rcm · rcm .
s i=1

The radius of gyration is a more useful quantity to work with since in many situations (like in
Polymer Science) we have to deal with more complicated structures than straight lines, squares or
spheres.
We also remember that the correlation function g(r) is defined as the probability that a site at
a position r from an occupied site belongs to the same finite cluster. We have the identity
X
g(r) = S(p) (1.49)
r

because the sum is the average number of sites to which an occupied site is connected, see Problem
2.

Definition 16 The correlation length ξ is defined as

P 2
r g(r)
ξ = Pr
2
. (1.50)
r g(r)

The correlation length represents some average distance of two sites belonging to the same cluster.
For a given s cluster, 2Rs2 is the average squared distance between two cluster sites. The
probability of a site being part of an s cluster is sn s(p), and it will be connected to s sites (if we
include self-connection). The corresponding average over 2R s2 gives the squared correlation length
P 2 2
s 2Rs s ns (p)
2
ξ = P 2
. (1.51)
ss ns (p)
The correlation length ξ is the upper cutoff of the radius of those clusters which contribute to
the mean cluster size near the percolation threshold. It is expected that ξ diverges as p → p c , like
ξ ∝ |p − pc |−ν . (1.52)
We have introduced a new critical exponent, ν, and we wish to determine how it is related to τ
and σ. For 2d percolation, plausible but not rigorous arguments give ν = 34 , in excellent agreement
with numerical results. In 3d, ν is somewhat smaller than 0.9, whilst for Bethe lattices one has
ν = 12 , analogous to the mean-field theories for thermal phase transitions.
As we have already seen, many quantities diverge at the percolation threshold. Most of these
involve sums over all cluster sizes s; their main contribution comes from s of the order s ξ ∝
1
|p − pc |− σ . Now we see that the correlation length ξ, which is also one of these quantities, is simply
the radius of those clusters which contribute the most to divergence. This effect is the foundation of
scaling theory. There is one and only one length ξ dominating the critical behaviour in an infinite
lattice.

23
1.6.2 Finite Boxing of Percolating Clusters
We now want to find out how the Rs varies with s at the percolation threshold p = p c where ξ = ∞.
Let M (L) denote the mass of the percolating cluster within linear distance L. If the percolating
cluster was a compact object, then it would be of the form M (L) ∝ L 2 . However, at p = pc ,
the percolating cluster is a fractal object. Figure 1.10 displays the size of the largest cluster S ∞

Figure 1.10: The size of the largest cluster S ∞ at p = pc as a function of the lattice size L.

(the percolating cluster) as a function of lattice size L in a double-logarithmic plot. The data are
consistent with the existence of a fractal dimension D such that

M (L) ∝ LD . (1.53)

For the 2d case D = 91

48 < 2 and in 1d, D = 1.
Having obtained the relationship in (1.53) it is natural to assume

s ∝ RsD for s  1 at p = pc . (1.54)

Thus,
1
Rs ∝ s D for p → pc . (1.55)
Inserting Eq.(1.55) into Eq.(1.51) we have
P 2
2+ D
2 s 2s ns (p)
ξ ∝ P 2
ss ns (p)
(2)
τ −3− Dσ
|p − pc |
∝ 2
|p − pc |τ −1− σ
2
∝ |p − pc |− Dσ
∝ |p − pc |−2ν , (1.56)

where we have used ∞

X τ −1−k
sk ns (p) ∝ |p − pc | σ .
s=1

24
Therefore we end up with the scaling relation
1 1
ν= ⇔D= . (1.57)
Dσ νσ
Note that the behaviour at p = pc and the behaviour for s  sξ , or Rs  ξ are indistinguishable.
Therefore, ξ is the crossover length between a critical and noncritical behaviour. On length scales
1
much less than ξ we have ns ∝ s−τ and Rs ∝ s D but for length scales much larger than ξ the
scaleless behaviour disappears. Note also, that
1
sξ ∝ ξ D ∝ |p − pc |−νD ∝ |p − pc |− σ

confirming once again the scaling relation above between D, ν and σ.

If, however, p  pc , the fractal dimension D is expected to be equal to the Eucledian dimension
d of the lattice. So,
D=d for p > pc . (1.58)
Clusters above pc are not fractals but ‘normal’ objects with D = d, provided s  s ξ or equivalently
Rs  ξ. Thus sξ or ξ sets the cluster size and linear scale where there is a crossover from fractal
to non-fractal behaviour.

Exponent 1d 2d 3d 4d 5d 6d Bethe
α 1 −2/3 -0.62 -0.72 -0.86 -1 -1
β 0 5/36 0.41 0.64 0.84 1 1
γ 1 43/18 1.80 1.44 1.18 1 1
ν 1 4/3 0.88 0.68 0.57 1/2 1/2
σ 1 36/91 0.45 0.48 0.49 1/2 1/2
τ 2 187/91 2.18 2.31 2.41 5/2 5/2
D(p = pc ) 1 91/48 2.53 3.06 3.54 4 4

Table 1.2: The critical exponents for the percolation theory problems in dimensions d = 1, 2, 3, 4, 5, 6
and in the Bethe lattice.

1.6.3 Mass of the Percolating Cluster

Now we wish to see how the scaling relation of the mass M (L) changes for the two different cases.

(1) For L  ξ: (always the case at p = pc , where ξ = ∞) The percolating cluster appears fractal,
implying M (L) ∝ LD .

(2) For L  ξ: The cluster appear homogeneous

M (L) = (No. of lattice sites) × (Prob. site belongs to percolating cluster)

= Ld P (p)
= Ld (p − pc )β
β
= Ld ξ − ν since ξ ∝ |p − pc |−ν . (1.59)

Now consider L ≈ ξ, that is, we match the observations above by substituting L with ξ:

M (L) ∝ LD ∝ ξ D , (1.60)

25
and β β
M (L) ∝ Ld ξ − ν ∝ ξ d− ν . (1.61)
Therefore, we can state that
β
D =d− , (1.62)
ν
which is know as a hyperscaling relation because the Euclidean dimension d enters in the scaling
relation.
Since the percolating cluster has a constant density for L  ξ, it is natural to divide the system
 d
L
into boxes of linear size ξ. In d dimensions, the total volume L d will be divided into ξ boxes.
Since the cluster inside each of these boxes of size ξd
has a mass of order ξD, the total mass of the
cluster is given by
 d
L
M (L, ξ) = ξ D = ξ D−d Ld , (1.63)
ξ
which is, of course, equivalent to P (p)L d .
In summary, we have 
 LD Lξ
M (L, ξ) ∝  d (1.64)
 ξD L Lξ
ξ

which can be written in an alternative form of

L
M (L, ξ) = LD m( ), (1.65)
ξ
where the scaling function ( L
constant x = ξ 1
m(x) = L (1.66)
xd−D x= ξ  1.
The argument x of the scaling function is a dimensionless number, namely the ratio of the two
length scales Lξ . No other length scales play a role. The scaling function describes a crossover
from fractal behaviour at length scales L much smaller than the correlation length ξ to uniform
behaviour at length scales L much larger than the correlation length ξ. In Nature, there are many
examples of crossover phenomena from one type of behaviour to another. Take as an example the
table in front of you. On large length scales, the surface of the table is smooth, but going to very
small length scales, the surface will become very rough and probably fractal. The mountain range
is yet another example. Far away, i.e., on large length scales, it looks pretty smooth, but, as you
are well aware, if you are in the middle of the mountain range, on small length scales, it is pretty
rough and rugged.

Exercise 14 Give one or two more examples of such crossover phenomena in Nature.

The density has the form

(
M (L, ξ) LD−d L  ξ L
ρ(L, ξ) = ∝ D−d = LD−d ρ̃( ), (1.67)
Ld ξ Lξ ξ

with a scaling function ( L

constant x = ξ 1
ρ̃(x) = L (1.68)
xd−D x= ξ  1.
To be aware that the density actually decreases with the length scale when L < ξ can be of
immense practical importance. Imagine that you have just been hired by an oil company and you
were given a sample of dimensions 0.1×0.1×0.1 m 3 of the porous medium where the oil resides and

26
Figure 1.11: The density of sites belonging to the percolating cluster within a region of linear size
L. In region I, the ratio L/ξ  1 and in region II, L/ξ  1. Region III is related to the fact, that
the lattices used are finite.

had to estimate how much oil the company could expect to recover in the oil field of dimension,
say, 100 × 100 × 100 km3 . The percolation model is used as a model for the distribution of oil
and gas inside porous rock in oil reservoirs. The empty sites model hard rock while the occupied
sites are pores filled with oil. The average concentration of oil is p. Now, if p < p c there are
only finite clusters, so let us assume p > p c . The strength P (p) will be the probability of drilling
into a percolating cluster, but how much oil can you recover? You would head for an immediate
dismissal if you argued as follows: Measure the density of oil in the sample ρ(L 1 = 0.1m). The
mass of oil to be recovered is M = ρ(L1 = 0.1m) · (100000m)3 . The argument is wrong simply
because the density of oil at length scales L 2 = 100 km is not given by the density of oil at length
scales L1 = 0.1 m, see Fig. (1.11). The proper way to argue would be the following: measure the
density of oil in the sample ρ(L1 = 0.1m) = CLD−d 1 . The density of oil at the reservoir length scale
ρ(L2 = 100km) = CLD−d 2 with the same constant C. Taking the ratio we find
 D−d
L2
ρ(L2 = 100km) = ρ(L1 = 0.1m)
L1
 D−d
= 106 ρ(L1 = 0.1m)
≈ 1.5 · 10−3 ρ(L1 = 0.1m),

since in 3d D − d ≈ 2.53 − 3 = −0.47, see table 1.2. The former estimate is wrong by a factor 1000.

1.7 Fractals
The percolating cluster at p = pc is an example of a random fractal. When p 6= p c large clusters
appear fractal on length scales up to the correlation length ξ ∝ |p − p c |−ν , roughly speaking. In
mathematics, we can also encounter deterministic fractals as for example the Sierpinski carpet and
the Sierpinski gasket.
For a geometrical fractal, the mass scales with the linear size raised to the fractal dimension
log M (L)
M (L) = LD ⇔ D = .
log L

27
Figure 1.12: Two examples of deterministic fractals. (a) The Sierpinski carpet. The first three levels
in the algorithm for constructing the Sierpinski carpet are shown. The linear scale is enlarged by a
factor 3 with the condition that an occupied square (white) is replaced with with 3 × 3 squares in
which the centre square is empty (grey) while an empty square (grey) is replaced with 3 × 3 empty
squares. (b) The Sierpinski gasket. A similar algorithm generates the Sierpinski gasket.

Let n denote the iteration number. By inspection of the Sierpinski carpet, M = 8 n and L = 3n ,
implying
log 8n log 8
D= n
= ≈ 1.893.
log 3 log 3
In the case of the Sierpinski gasket, M = 3 n and L = 2n , yielding
log 3
D= ≈ 1.585.
log 2
Note that if the length scale is rescaled by a factor b, then
 D
L L LD M (L)
M( ) = = =
b b bD bD

so that the mass is reduced by a factor b D . Using this observation as the characteristic of fractal
behaviour for random fractals we can easily find, say, the scaling of the mass of the percolating
cluster M (L, ξ) as a function of L and ξ. Let us rescale the length scales by a factor b, that is,
L → Lb and ξ → ξb (no other length scales are present). Thus

L ξ
M (L, ξ) = bD M ( , )
b b
D D L ξ
= b b M( 2 , 2 )
b b
..
.
L ξ
= bDl M ( l , l ).
b b
(1.69)

28
 ξ
If we are at the critical point p = pc , ξ ∝ |p − pc |−ν = ∞, but then bl
= ∞ ∀l. If we stop the above
renormalisation procedure when bl = L we find
 D L ξ
M (L, ξ = ∞) = bl
M ( l , l = ∞)
b b
D
= L M (1, ∞)
= LD .
(1.70)

This proves that self-similarity is mathematically expressed by power-law behaviour. Now, assume
p > pc with ξ ∝ |p − pc |−ν < ∞ with

L ξ
M (L, ξ) = bDl M ( l , l ).
b b
We now have to consider two cases depending on the ratio L/ξ.

(a) If L/ξ  1 ⇔ L  ξ (which is allways the case at p = p c ) we stop the iteration process when
we reach the smallest length scale b l = L and we “recover”
ξ
M (L, ξ) = LD M (1, ) = LD M (1, ∞) ∝ LD .
L

(b) If L/ξ  1 ⇔ L  ξ, we stop the iteration process when b l = ξ and we find

 d
L L
M (L, ξ) = ξ M ( , 1) ∝ ξ D
D
= ξ D−d Ld ,
ξ ξ
 d
since for a uniform system M ( Lξ , 1) ∝ L
ξ .

Exercise 15 Discuss why the system looks uniform when ξ/b l = 1.

1.8 Finite-size scaling

Let us consider the percolation problem for p 6= p c where the correlation length is finite
1
ξ ∝ |p − pc |−ν ⇒ |p − pc | ∝ ξ − ν .

The strength of the infinite cluster for p > p c can then be expressed in terms of the finite correlation
length
β
P (p) ∝ (p − pc )β ∝ ξ − ν
and similarly for the average cluster size
γ
S(p) ∝ |p − pc |−γ ∝ ξ ν .

These are the results in infinite lattices with L = ∞. Now, what happens at finite lattice sizes
when L < ∞? Well, nothing happens as long as L  ξ. The only relevant length scale is the

29
correlation length ξ. However, when L  ξ, the length scale will be set by L. This we have already
encountered before, we just didn’t stress that point. Better late that never:

P (L, ξ) = ρ(L, ξ)
M (L, ξ)
= d
( L
LD−d L  ξ
=
ξ D−d L  ξ.
( β
L− ν L  ξ
= β
ξ − ν L  ξ.
β L
= ξ− ν f ( )
ξ
where the scaling function    β
 L −
L ν L
1
f( ) = ξ ξ
ξ  constant L
 1.
ξ

1.0 20
L = 25 L = 25
L = 50 L = 50
0.8 L = 100 L = 100
Average cluster size S

L = 200 L = 200
Indicates pc Indicates pc
Strength P

0.6
10
0.4

0.2

0.0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Occupation probability P Occupation probability p

Figure 1.13: (a) The strength P (p) measured in a 2d square lattice for finite system sizes as a
function of occupation probability p. When p is far away from p c , L  ξ and the strength is
β
determined by ξ − ν . However, with p closer to pc , the finite-size effect is seen and P (p) decays with
β
system size as L− ν . (b) The average cluster size S(p) measured in a 2d square lattice for finite
system sizes L as a function of occupation probability p. When p is far away from p c , L  ξ and
γ
the average cluster size is determined by the correlation length ξ ν . γHowever, with p closer to pc ,
the finite-size effect is seen and S(p) increases with system size as L ν .

Similarly for the average cluster size, we would expect

( γ
Lν Lξ
S(L, ξ) ∝ γ
ξν L  ξ.

Generally, if we have a quantity

χ
X ∝ |p − pc |−χ ∝ ξ ν for L  ξ

30
then we would expect
( χ
Lν Lξ
X(L, ξ) ∝ χ
ξν L  ξ.
χ L
= ξ ν x1 ( ) (1.71)
ξ
or, in terms of occupation probabilities
1
X(L, p) = |p − pc |−χ x2 (L ν (p − pc )).

The important message we get is that by studying finite-size scaling, that is, studying X as a
function of (finite) system sizes L at p = p c , we can extract the exponent χν .

1.9 Real space renormalisation in percolation theory

Imagine a percolation lattice with n sites. The total number of different microscopic configurations
is 2n as a site can be in one of two states – occupied with probability p or empty with probability
1 − p. Table (1.3) shows the number of microscopic configurations in a two dimensional square
lattice of linear size L.
L 1 2 3 4 5 6 7 8 9 10 17 100
n 1 4 9 16 25 36 49 64 81 100 289 10000
2n 2 16 512 65536 3 · 107 7 · 1010 6 · 1014 2 · 1019 2 · 1024 1030 1087 103000

Table 1.3: The number of microscopic configurations 2 n as a function the linear size L of a two-
dimensional square lattice where n = L 2 , the number of sites in the lattice. For lattice size L = 17,
the number exceeds the estimated number of atoms in the universe ≈ 10 81 .

In principle, if we wanted to calculate the average of a quantity hAi (say the average cluster
size), we should evaluate the quantity A i associated with each configuration i and then weight the
quantity Ai with the probability pi (which would be of the form pm (1 − p)(n−m) , where m is the
number of occupied sites in that particulate configuration) of the system to be in that microscopic
state, that is, P
X
i pi Ai
hAi = P = pi Ai . (1.72)
i pi i
This procedure, however, soon becomes an impossible task due to the numbers of different
states involved. With today’s computing power one might do, say, a 7 × 7 system, but an 8 × 8
system would take approximately 40 years on a tera flop (10 12 calculations per second) while the
time required to do a 10 × 10 system would exceed the age of the universe which is approximately
15 · 109 years.
Real space renormalisation is a coarse graining procedure where we systematically reduce the
number of degrees of freedom (sites) in order to break a large problem down into a sequence of
smaller and more manageable stages by eliminating fluctuations on scales less than a given length
scale b. The effect will be to reveal the large scale behaviour of the system.
Before we introduce the real space renormalisation method, let us make some observations
related to the phase transition in a percolation model.

(a) The empty state at occupation probability p = 0 is self-similar with an associated correlation
length ξ = 0.

31
(b) The fully occupied state at occupation probability p = 1 is self-similar. The associated corre-
lation length ξ = 0. The ordered state has an order parameter P (p = 1) = 1.
(c) The states at the critical occupation probability p = p c are also self-similar and the correlation
length ξ = ∞. The infinite cluster is fractal and looks alike on all length scales and we have
clusters of all sizes from s = 1 to s = ∞.

Loosely speaking, the correlation length ξ sets the scale of the largest fluctuation from the
“averaged” state. For p 6= pc , the correlation length is identified with the linear size of the cutoff
cluster size sξ .
When p < pc the upper cutoff of the linear size of the fluctuations away from the empty state
is given by ξ and when p approaches zero, ξ approaches zero. At p = 0, there are no clusters left
at all and ξ = 0.
When p > pc , the correlation length can also be identified with the upper cutoff in the linear
size of the largest holes in the percolating (spanning) cluster which indicates how far away you are
from the fully occupied state. When p approaches one, ξ approaches zero. At p = 1, there are no
holes left and ξ = 0.
At the phase transition point p = pc , the correlation length is infinite. Thus there are fluctua-
tions from the smallest length scale up to infinity.
The real space renormalisation technique is based on the so-called block site (spin) technique in-
troduced by Leo Kadanoff and later formalised into the renormalisation group method by Kenneth
Wilson, who received the Nobel price in 1982 for “his theory for critical phenomena in connection
with phase transitions”, see, for example, the website on physics laureates http://www.nobel.se/physics/laureates/1
The renormalisation group method has three basic steps:

1. Divide the lattice into blocks of linear size b (in terms of the lattice constant) with each block
containing a few sites (spin).
2. Next, the coarse graining procedure takes place. The sites in the blocks are averaged in some
way (to be specified more precisely shortly) and the entire block is replaced by a single su-
per site (spin) which is occupied with a probability according to the renormalisation group
transformation p0 = Rb (p).
In the combined procedure 1 and 2, one should keep the symmetry of the original lattice such
that we can repeat the coarse graining procedure again. The result of these two operations are
to create a new lattice whose fundamental spacing is b times as large as the original lattice.
3. Restore original lattice constant by rescaling the length scales by the factor b.

These 3 steps define a renormalisation group transformation R b , where the number of degrees
of freedom is reduced drastically. Note that it is called a group because of the property that when
applying the transformation to a configuration {s i },

Rb2 (Rb1 ({si })) = Rb1 b2 ({si }) (1.73)

but it is of course not a group in a strictly mathematical sense because no inverse transformation
exists since we are reducing the number of degree of freedom.
The effect of the coarse graining procedure in step 2 is to eliminate from the system all fluctu-
ations whose scale is smaller than the block size b. Any small scale fluctuations of the sites over
a range of less than b lattice units will be smeared out. This is somewhat similar to the process
taking place if you view the system through a lens out of focus – all the smaller features are blurred
but the large scale features are unaffected.
Let us clarify the 3 steps by looking at a simple example. Consider a 2d square lattice with
9 × 9 sites, that is, 81 degrees of freedom and a total of 2 81 = 2 · 1024 different configurations.

32
Figure 1.14: The three renormalisation group steps illustrated in a 2d square lattice with b = 3.
The majority rule defines the renormalisation group transformation.

1. Divide the square lattice up into small 3 × 3 blocks with a total of 9 sites (b = 3.)

2. We now have to specify the course graining procedure which will define the renormalisation
group transformation. One possibility would be to use the so-called majority rule which
states that if a majority of the sites in the block is occupied, then the super site should be
occupied, otherwise empty. The probability, that a majority of sites are occupied in the 3 × 3
block then defines the renormalisation group transformation

p0 = Rb (p) = p9 + K9,1 p8 (1 − p) + K9,2 p7 (1 − p)2 + K9,3 p6 (1 − p)3 + K9,4 p5 (1 − p)4 ,

where K9,k is the number of different ways to place k empty sites into the cells of 9 sites.
The super site should be occupied with probability p 0 = Rb (p) and empty with probability
1 − p0 = 1 − Rb (p).

3. Rescale the lattice of super sites by a factor 3 to restore original lattice constant.
Let ξ denote the correlation length in the original lattice. Assume we are close to p c . Then

ξ = constant |p − pc |−ν (1.74)

We denote the correlation length in the new renormalised lattice by ξ 0 . Since the length scales
have been rescaled by a factor b (step 3)
ξ
ξ0 = ⇔ (1.75)
b
constant |p − pc |−ν
constant |Rb (p) − pc |−ν =
b
where Rb (p) is the renormalisation group transformation that determines the new occupation prob-
ability in the rescaled lattice associated with the new (smaller as b > 1) correlation length.
Simple rearranging implies that
 −ν
|Rb (p) − pc | 1
= (1.76)
|p − pc | b
so the critical exponent ν is given by
log b
ν=  . (1.77)
|Rb (p)−pc |
log |p−pc |

The basic idea is that at the critical point p c we have self-similarity, that is,

ξ 0 = ξ. (1.78)

33
 2000

Corraletion length ξ

0
0.0 0.2 0.4 0.6 0.8 1.0
Occupation probability p

Figure 1.15: When renormalising the length scales by a factor b, the correlation length decreases
unless it was 0 or ∞ initially. Associated with a decrease in the correlation length is a flow in
occupation probability space either towards p = 0, if initially p < p c or towards p = 1, if initially
p > pc . The fixed points of the renormalisation group transformation are associated with the trivial
self-similar states at p = 0 or p = 1 or the nontrivial self-similar states at p = p c .

This equation can only be consistent with Eq.(1.76) if

(
∞
ξ= (1.79)
0.

This is intuitively clear as ξ = 0 is associated with either p = 0 (empty lattice) or p = 1 (fully occu-
pied lattice) the two trivially self-similar cases and ξ = ∞ with the critical point p c . Furthermore,
we can also conclude that

ξ = constant |p − pc |−ν = ξ 0 = constant |Rb (p) − pc |−ν (1.80)

which implies that the phase transition ξ = ∞ is identified with (one of) the fixed points p ? of the
renormalisation group transformation, that is, the solutions of the equation

Rb (p? ) = p? . (1.81)

Thus
|Rb (p) − pc | |Rb (p) − Rb (pc )|
=
|p − pc | |p − pc |
dRb (pc )
= for p → pc (1.82)
dp

leaving, see Eq.(1.77)

log b
ν=   (1.83)
dRb (pc )
log dp

1.9.1 Renormalisation group transformation in 1d.

Consider a renormalisation group transformation in 1d. Take a 1d lattice where each site is occupied
with probability p. Divide the lattice into blocks with b sites. Let the renormalisation group

34
transformation be determined by the probability of having a spanning cluster. The probability of
having a spanning cluster in a block of b sites is

Rb (p) = pb .

The fixed point equation can easily be solved

(
? ?b ? ? 0
Rb (p ) = p =p ⇔p =
1.

If we start with an empty lattice p = 0, the renormalised lattice will also be empty, associated with
the fixed point p? = 0. If we start with a fully occupied lattice p = 1, the renormalised lattice
will also be fully occupied since all the blocks contain a percolating cluster, associated with the
fixed point p? = 1. However, if we start with a lattice containing some empty sites p < 1, the
renormalised lattice will contain even more empty sites because p 0 = Rb (p) = pb < p. Repeating
the renormalisation procedure will gradually bring the renormalised occupation probability towards
zero. Clearly, p? = 0 is associated with ξ = 0, while p? = 1 is associated with ξ = ∞ and must
thus be identified with the critical occupation probability p c .
Furthermore, in order to calculate the correlation length exponent ν we need to take the deriva-
tive of the renormalisation group transformation evaluated at p ? = 1.

dRb (p)
|p? =1 = bpb−1 |p? =1 = b,
dp
implying
log b
ν= dRb (p)
= 1,
log dp |p =1
?

so in 1d the renormalisation group transformation is exact.

1.9.2 Renormalisation group transformation on 2d triangular lattice.

Divide the lattice into triangular cells containing three sites each. Let the probability for having a
spanning cluster define the renormalisation group transformation. Thus

Rb (p) = p3 + 3p2 (1 − p) = 3p2 − 2p3 . (1.84)

The fixed point equation



 0
Rb (p? ) = 3p?2 − 2p?3 = p? ⇔ p? = 1

 1/2.

The two trivial fixed point is associated with the self-similar states of an empty lattice and fully
occupied lattice. The unstable fixed point p ? = 1/2 will be associated with the nontrivial self-similar
states at pc . The critical exponent ν is
√ √
log b log 3 log 3
ν= = = = 1.355.
log dRb (p) |p? =1/2 log(6p − 6p2 )|p? =1/2 log 3/2
dp

The exact values are pc = 1/2 and ν = 4/3 so the renormalisation group transformation does a
good job.

35
Figure 1.16: Renormalisation group method on a 2d square lattice according to the majority rule
for 3 × 3 blocks. Shown are part of the original lattice L = 729 and the renormalised lattices
L = 243 (part of), L = 81, 27, and 3. (a) Initial occupation probability p < p ? in the original
lattice and the flow is towards the empty lattice. (b) Initial occupation probability p > p ? in the
original lattice and the flow is towards the fully occupied lattice. (c) Initial occupation probability
p = p? in the original lattice. This is a fixed point for the renormalisation group transformation
and consequently there is no flow. It looks like itself on all length scales.

1.9.3 Renormalisation group transformation on 2d square lattice of bond per-

colation.
Consider the block defined in Fig. (1.17) which, after the coarse graining procedure, is replaced by
two superbonds. The renormalisation group transformation is

p0 = Rb (p)
= probability to have a spanning cluster in horizontal direction
= p5 + p4 (1 − p) + 4p4 (1 − p) + 2p3 (1 − p)2 + 2p3 (1 − p)2 + 4p3 (1 − p)2 + 2p2 (1 − p)3
= 2p5 − 5p4 + 2p3 + 2p2 (1.85)

with the fixed point equation



 0
Rb (p? ) = 2p? 5 − 5p? 4 + 2p? 3 + 2p? 2 = p? ⇔ p? = 1

 1/2.

36
Figure 1.17: Renormalisation group method of 2d square lattice with bond percolation according to
the spanning cluster rule. (a) and (b) After the coarse graining procedure, the block is replaced with
two super bonds AC and AG. (c) The various configurations with a horizontal spanning cluster.
The two dangling bonds DG and EH need not be considered as they do not affect the probability
of having a spanning cluster from left to right.

Again, the two trivial fixed points are associated with the self-similar states of an empty lattice
and fully occupied lattice. The unstable fixed point p ? = 1/2 will be associated with the nontrivial
self-similar states at pc . For bond percolation in 2d, pc = 1/2, see Table 1.1. The critical exponent
ν is
log b log 2
ν= dRb (p)
= = 1.428,
log dp |p? =1/2 log 13/8

which should be compared to the analytical result of ν = 4/3 in 2d. Thus the renormalisation
group transformation defined above for bond percoaltion in 2d results in an exact prediction of p c
and a good estimate of ν.

1.9.4 Why is the renormalisation group transformation not exact?

There are two significant sources of error which is the reason for not obtaining the exact results

(1) Two sites that were connected in the original lattice can be disconnected and vica versa.

(2) The renormalised system is not really a true percolation system because the probabilities of
having bonds between supersites are no longer independent. As we shall see later, this is
the equivalent of introducing next-nearest-neighbour couplings in the renormalisation group
procedure in the 2d Ising model. Thus we should really introduce new parameters but we
choose here to truncate the problem to only one parameter, the occupation probability p.

37
 Glossary
Taken from The New Physics ed. Paul Davies, Cambridge University Press.

Coarse-graining An operation implementing some form of spatial averaging which smoothes out
relatively small length scale configurational structure while preserving the larger length-scale struc-
tures.
Correlation length The correlation length ξ gives a measure of the typical length scale over which
fluctuations of one microscopic variable are correlated with the fluctuations of another. In percola-
tion theory, it is the typical cluster diameter of the clusters s ξ which give the main contribution to
the divergence of the second (and higher) moments of the cluster distribution. Close to a critical
point ξ ∝ |p − pc |−ν for p → pc .
Critical phenomena The phenomena which occur in the neighbourhood of a continuous (second
order) phase transition, characterised by very long correlation lengths.
Critical exponents (or indices) Near a ctitical point, physical quantities are often proportional
to a power of another quantity, such as the distance from the critical point |p − p c |. The power that
occurs known as a critical exponent (index).
Critical point A point in a phase diagram, where the correlation length associated with the physical
system is, in principle, infinite.
Fractal geometry Generalisation of Euclidean geometry suitable for describing irregular and frag-
mented patterns such as the percolating cluster at p = p c . A noninteger fractal dimension D can
frequently be associated with such patterns.
Order parameter A variable such as the strength of the infinite cluster in percolation theory (or
the magnetisation in an Ising model) used to describe the degree of order in a phase above (below) its
critical point. In a continuous phase transition (second order phase transition), the order parameter
goes continuously to zero as the critical point is approached from above (below).
Phase transition A change of state such as occurs in the boiling or freezing of a liquid, or in
the change between ferromagnetic and paramagnetic states of a magnetic solid. An abrupt change,
characterised by a jump in an order parameter is known as first order; a change in which the
order parameter evolves smoothly to or from zero is called continuous or second order.
Renormalisation group In statistical physics, the renormalisation group method systematically
implement some form of coarse-graining operation (e.g. rescaling length scales with a factor b:
L → Lb , ξ → ξb ) to expose the character of the large-scale phenomena, in systems where many length
scales are important.
Scale invariance A physical system is said to exhibit scale-invariance if it remains unchanged (in
a statistical sense) by a coarse-graining operation.
Universality The phenomena whereby many microscopically quite different physical systems ex-
hibits critical point behaviour, with quantitatively identical features, such as critical exponents. In
percolation theory e.g. the critical exponents do not depend on the microscopic details of the lattice
but only on the Euclidean dimension d.

38
Bibliography

[1] D. Stauffer and A. Aharony, Introduction to Percolation Theory, (Taylor & Francis, 1994).

[2] J. Adler, Physica A, 171, 453 (1991).

[3] Iwan Jensen, J. Stat. Phys., 102, 865 (2000).

39