Explosive percolation via control of the largest cluster

N. A. M. Araújo1, ∗ and H. J. Herrmann1, 2, †

1
Computational Physics for Engineering Materials, IfB,
ETH Zurich, Schafmattstr. 6, 8093 Zurich, Switzerland
2
Departamento de Fı́sica, Universidade Federal do Ceará,
Campus do Pici, 60451-970 Fortaleza, Ceará, Brazil
We show that only considering the largest cluster suffices to obtain a first-order percolation
transition. As opposed to previous realizations of explosive percolation our novel models obtain
Gaussian cluster distributions and compact clusters as one would expect at first-order transitions.
We also discover that, surprisingly, the cluster perimeters are fractal at the transition point yielding
a fractal dimension of 1.23 ± 0.03 intriguingly close to that of watersheds.

PACS numbers: 64.60.ah, 89.75.Da, 64.60.al

Percolation, the paradigm for random connectivity, has

since Hammersley [1] been one of the most often applied
statistical models [2, 3]. Its phase transition being re-
lated to magnetic models [4] is in all dimensions one of
the most robust second-order transitions known. This ex-
plains the enormous excitement generated by the recent
work by Achlioptas, D’Souza, and Spencer [5] describ-
ing a stochastic rule apparently yielding a discontinuous
percolation transition on a fully connected graph. Subse-
quent work applied the process on other networks [6–11].
However, reported results of finite-size studies and size classical product rule
distributions are not consistent with a first-order transi-
tion. Since then various rules have been devised [12–14]
and even a Hamiltonian formalism was proposed [15],
all attempting a discontinuous transition towards an in-
finite cluster. In all proposed models one tries to keep
the clusters of similar size and some authors additionally
suppress the internal bonds of clusters [5, 15]. Could one
obtain a clear and consistent first-order percolation tran-
sition? It is the objective of the present Letter to answer
this question. One criterion is the cluster size distribution
at the percolation threshold. Radicchi and Fortunato [9]
largest cluster Gaussian
as well as Ziff [7] found a power-law distribution with
an exponent close to two. Although, different from the FIG. 1: (Color online) Snapshots of the system, obtained on
exponent of classical percolation the sole fact of finding a square lattice with 10242 sites, at pc , for four different bond
a power law is untypical for first-order transitions. Also percolation models, namely, classical [2], Achlioptas product
unusual for a first-order transition is that the clusters are rule [6], largest cluster model (α = 1), and Gaussian model
fractal, as we found happens for the Achlioptas rule, from (α = 1). The largest cluster and Gaussian models are intro-
the behavior of the order parameter with the system size duced in this Letter.
[7, 9]. It is a purpose of the present Letter to present a
model in which a Gaussian cluster size distribution and
compact clusters can be achieved in a systematic way, parameter, i.e., a second-order transition. On a fully con-
characterized by a fractal perimeter yielding a fractal di- nected graph, Achlioptas et al. [5], used the best-of-two
mension similar to the one of watersheds and random product rule studied in detail by Friedman and Lands-
polymers in strongly disordered media. berg [12]. Ziff reported simulations on a regular square
Usual bond percolation can be implemented on a lattice [6, 7], while Radicchi and Fortunato [8, 9] and Cho
square lattice by randomly occupying bonds between et al. [10] on scale-free networks.
neighboring sites, reaching its threshold at a certain frac- More recently, other approaches have been introduced
tion when opposite borders are first connected through to obtain explosive percolation. Instead of a best-of-two
one large cluster [2, 16, 17]. This percolation threshold, rule Manna et al. [13], Cho et al. [11], and Moreira et
is characterized by the continuous vanishing of the order al. [15] proposed a weighted rule where bonds are occu-
 2

pied according to a certain probability. However, despite

108
being rejection-free schemes, they are limited to small-
1
system sizes and/or reduced number of samples. Here,
we suggest an acceptance method where new bonds are P∞ 0.99
selected randomly and occupied according to a certain 10 6
weight yielding, for the first time, a clear first-order tran- 0.98
2.0
0 2500 5000
sition. The considered scheme allows to consider system χ L
10-1
sizes 64 times larger than before [13], specifically, we con-
sider systems of 40962 sites and averages over 104 sam- 104 χ∞/N
ples.
In our simplest rule (“largest cluster model”), as for 10-2
101 102 103 104
classical bond percolation, a link is randomly selected L
2
among the empty ones. If its occupation would not lead 10
101 102 103 104
to the formation or growth of the largest cluster, it is al-
L
ways occupied, otherwise, it is occupied with probability
FIG. 2: (Color online) Size dependence, for the largest cluster
( "  2 #) model, of the susceptibility (χ), fraction of sites in the largest
s − s̄ cluster (P∞ ), and its standard deviation per site (χ∞ /N ) at
min 1, exp −α , (1)
s̄ the percolation threshold, on a square lattice of linear size
(L) ranging from 32 to 4096. All bonds are occupied with
where s is the size of the cluster that would be formed the same probability except the ones that lead to the for-
mation/growth of the largest cluster, to which an occupation
by occupying this bond and s̄ the average cluster size af-
probability q is assigned, Eq. (1), with α = 1. Results have
ter occupying the bond. The parameter α controls the been averaged over 104 samples.
allowed size dispersion. Note that, for α ≤ 0, since the
size of the largest cluster is always greater (or equal)
than the average cluster size, all new bonds are occupied To estimate the percolation threshold we consider the av-
reducing to classical bond percolation, characterized by erage value of p (fraction of occupied bonds) at which a
a continuous transition at the percolation threshold [2]. connected path linking opposite boundaries of the sys-
For α > 0, the probability of Eq. (1) suppresses the for- tem is obtained. Considering different system sizes, for
mation of a cluster significantly larger than the average, α = 1, we obtain for the percolation threshold pc =
inducing a homogenization of cluster sizes. The Gaus- 0.632 ± 0.002. To identify the order of the transition,
sian function has been considered because this is what in the largest cluster model, Fig. 2 presents a finite-size
we expect for the cluster size distribution at a first-order study for P∞ , χ, and χ∞ /N , averaged over 104 samples
transition. However, to observe a discontinuous transi- of square lattices with linear sizes ranging from 32 to
tion any other function could be chosen, as far as it con- 4096. As we can see in the top inset of Fig. 2, above a
strains the largest cluster differing significantly, in size, certain system size, the order parameter, at the percola-
from the average cluster. tion threshold, does not show any finite-size dependence,
For nonequilibrium problems, where a free energy can- staying at a constant value in the thermodynamic limit
not be defined, transitions can still be classified based on (L → ∞). The second moment of the cluster size dis-
the behavior of the order parameter [18]. A first-order tribution (χ) scales with Ld (d = 2) which is a sign of a
transition, is characterized by a jump in the order param- first-order transition [19, 20]. The standard deviation of
eter, otherwise, a transition is denoted as continuous. For the largest cluster (smax ) per lattice site, which was also
percolation, we define as order parameter the fraction of considered in Refs. [6] and [7], converges, for larger sys-
sites in the largest cluster (P∞ ) [2]. Here we also consider tem sizes, to a constant value, corroborating the presence
two other quantities: the second moment of the cluster of a discontinuous transition.
size distribution (χ), defined as To explicitly control the cluster size distribution we
also implemented the following model. A new bond is
X chosen from the list of empty ones and occupied with
χ= s2i , (2) probability given by Eq. (1). For internal connections we
i
consider s as twice the cluster size. Since equation (1)
√is a
where the sum runs over all clusters i, and the standard Gaussian with average size s̄ and size dispersion s̄/ 2α,
deviation (χ∞ ) of the largest cluster size (smax ) over dif- we denote this model as Gaussian model. Note that here
ferent samples, the occupation probability is assigned to all new bonds
even when they are not related to the largest cluster.
p This not only guarantees the control over clusters greater
χ∞ = hs2max i − hsmax i2 . (3) than the average, as in the previous model, but also over
 3

108 104
Largest
1
Gaussian
P∞ 0.9 103
1.26±0.04
106 0.8
0 2500 5000 2.0 102
χ L Nsticks 1.23±0.03
100 1
10
104 χ∞/N 10-1
100
10-2
101 102 103 104
2 L 10-1
10
101 102 103 104 10-4 10-3 10-2 10-1 100
L 1/length

FIG. 3: (Color online) Size dependence, for the Gaussian FIG. 4: (Color online) Number of sticks necessary to follow
model, with α = 1, of the susceptibility (χ), fraction of sites the perimeter of the infinite cluster as a function of the stick
in the largest cluster (P∞ ), and its standard deviation per site length, to obtain the fractal dimension of the perimeter with
(χ∞ /N ) at the percolation threshold, on a square lattice of the yardstick method. For both the largest cluster and Gaus-
linear size (L) ranging from 32 to 4096. All bonds are occu- sian models, with α = 1. For the Gaussian model data were
pied with a probability given by Eq. (1). Results have been vertically shifted by a factor of 0.1. Results have been aver-
averaged over 104 samples. aged over 104 samples of lattices with linear size 2048.

the smaller ones. For α = 0, all bonds have the same sion and number of clusters are observed for the largest
probability and, therefore, the model reduces to classical cluster model. According to Eq. (1), increasing the value
bond percolation. For negative α, the growth of larger of α decreases the size dispersion.
clusters is favored in two different ways: they differ more As clearly seen in the snapshots of Fig. 1, clusters ob-
from the average value and have more empty bonds than tained with our models are compact but we find that the
the smaller ones. Yet, for all negative α, the model recov- surface is fractal. For the Gaussian model, we calculate
ers the classical universality class of percolation [2, 18]. for the cluster perimeter a fractal dimension of 1.23±0.03,
As example, for positive α, we present, in Fig. 3, a obtained with the yardstick method [21] (Fig. 4). For the
size dependence study of the order parameter, second largest cluster model, it is also characterized by a fractal
moment of the cluster size distribution, and standard de- perimeter with a fractal dimension of 1.26±0.04 (Fig. 4).
viation per site of the largest cluster, for the Gaussian Compact clusters with fractal surface were also reported
model, with α = 1, at the percolation threshold, on a for irreversible aggregation growth in the limit of high
regular square lattice with linear size (L) ranging from concentration by Kolb et al. [22]. For the present models,
32 to 4096. Results were averaged over 104 samples. We the percolation thresholds are larger than the ones from
extrapolate, for the infinite system, a percolation thresh- previous models due to the compactness of the clusters.
old pc = 0.56244 ± 0.00006. As for the largest cluster In Fig. 5 we see the cluster size distribution, P (s, α),
model, the density of the infinite cluster does not change for different system sizes, obtained with the Gaussian
significantly with the system size, the second moment of model. Measurements have been performed at the per-
the cluster size distribution scales with Ld (d = 2), and colation threshold on a square lattice with 10242 , 20482 ,
the standard deviation per site of the largest cluster con- and 40962 sites, and averaged over 104 samples. Three
verges to a non-zero constant. As before these results characteristic peaks are observed. In fact, the third peak
imply a first-order transition. (around 0.7) is only due to the largest cluster and only
Figure 1 shows snapshots for four different models of appears due to the small number of clusters at the perco-
bond percolation: classical, product rule, largest cluster lation threshold, being finite-size effect. This peak is not
model, and Gaussian model. All figures have been ob- observed when we compute the same distribution without
tained at their respective percolation thresholds (pc ). For considering the largest cluster. In the thermodynamic
classical percolation and for the product rule, clusters of limit, since an infinite number of clusters exists, the con-
very different sizes are obtained. In fact, the cluster size tribution of a single cluster to the distribution vanishes.
distribution is characterized by a power law [7, 9]. How- The presence of two main peaks is characteristic for a
ever, for the largest cluster and the Gaussian model, a first-order transition showing, for a finite system, at the
characteristic cluster size is observed. Both models lead percolation threshold, coexistence of the percolative and
to a localized cluster size distribution. Small size disper- non-percolative states [23].
 4

4 CH1-01-08-2.
1024
2048
4096
3
∗
Electronic address: [email protected]
†
P(s,α) 2 Electronic address: [email protected]
[1] S. R. Broadbent and J. M. Hammersley, Proceedings of
the Cambridge Philosophical Society 53, 629 (1957).
1
[2] D. Stauffer and A. Aharony, Introduction to Percolation
Theory (Taylor & Francis, London, 1994), 2nd ed.
[3] M. Sahimi, Applications of percolation theory (Taylor &
Francis, London, 1994).
0
0 0.2 0.4 0.6 0.8 1 [4] C. M. Fortuin and P. W. Kasteleyn, Physica 57, 536
s (1972).
[5] D. Achlioptas, R. M. D’Souza, and J. Spencer, Science
323, 1453 (2009).
FIG. 5: (Color online) Cluster size distribution for the Gaus- [6] R. M. Ziff, Phys. Rev. Lett. 103, 045701 (2009).
sian model for different system sizes (α = 1), at the percola- [7] R. M. Ziff (2009), arXiv:0912.1060.
tion threshold, on a square lattice, averaged over 104 samples. [8] F. Radicchi and S. Fortunato, Phys. Rev. Lett. 103,
Black-dashed lines are two Gaussian distributions fitting the 168701 (2009).
results from simulation. The black-solid line is the sum of [9] F. Radicchi and S. Fortunato, Phys. Rev. E 81, 036110
both curves. (2010).
[10] Y. S. Cho, J. S. Kim, J. Park, B. Kahng, and D. Kim,
Phys. Rev. Lett. 103, 135702 (2009).
In conclusion, the present work reveals that, to obtain [11] Y. S. Cho, B. Kahng, and D. Kim, Phys. Rev. E 81,
explosive percolation on a regular lattice it is sufficient 030103(R) (2010).
to control the formation and growth of the largest clus- [12] E. J. Friedman and A. S. Landsberg, Phys. Rev. Lett.
ter, instead of applying a rule to the overall set of empty 103, 255701 (2009).
[13] S. S. Manna and A. Chatterjee (2009), arXiv:0911.4674.
bonds. We propose the largest cluster model which sys- [14] R. M. D’Souza and M. Mitzenmacher, Phys. Rev. Lett.
tematically suppresses the formation of a largest clus- 104, 195702 (2010).
ter. We introduce as well, the Gaussian model, where [15] A. A. Moreira, E. A. Oliveira, S. D. S. Reis, H. J.
a weight is assigned to each selected bond, such that Herrmann, and J. S. Andrade Jr., Phys. Rev. E 81,
a Gaussian distribution of cluster sizes is obtained, re- 040101(R) (2010).
vealing the coexistence of two states at the percolation [16] D. Stauffer, Phys. Rep. 54, 1 (1979).
threshold. Our models, yielding clear first-order transi- [17] M. B. Isichenko, Rev. Mod. Phys. 64, 961 (1992).
[18] G. Ódor, Rev. Mod. Phys. 76, 663 (2004).
tions, show that explosive percolation can be obtained
[19] K. Binder, Z. Phys. B 43, 119 (1981).
under less stringent conditions that previously thought [20] K. Binder and D. P. Landau, Phys. Rev. B 30, 1477
shedding light on the minimum ingredients to trigger ex- (1984).
plosive percolation. In fact, we believe that our restric- [21] C. Tricot, J. F. Quiniou, D. Wehbi, C. Roques-Carmes,
tions on the formation of a largest cluster differing signifi- and B. Dubuc, Revue Phys. Appl. 23, 111 (1988).
cantly, in size, from the average, is the required necessary [22] M. Kolb and H. J. Herrmann, Phys. Rev. Lett. 59, 454
condition and hope that this statement can one day be (1987).
[23] K. Binder, K. Vollmayr, H.-P. Deutsch, J. D. Reger,
formally proven. The value of the novel fractal dimension
M. Scheucher, and D. P. Landau, Int. J. Mod. Phys. C
of percolation that we discovered in the cluster perime- 3, 1025 (1992).
ters is intriguingly close to the one found for watersheds [24] E. Fehr, J. S. Andrade Jr, S. D. da Cunha, L. R. da Silva,
[24] and random polymers in strongly disordered media H. J. Herrmann, D. Kadau, C. F. Moukarzel, and E. A.
[25]. Oliveira, J. Stat. Mech. P09007 (2009).
We acknowledge financial support from the ETH Com- [25] M. Porto, N. Schwartz, S. Havlin, and A. Bunde, Phys.
petence Center Coping with Crises in Complex Socio- Rev. E 60, R2448 (1999).
Economic Systems (CCSS) through ETH Research Grant