Cluster aggregation model for discontinuous percolation transitions.

Abstract

The evolution of the Erdos-Rényi (ER) network by adding edges is a basis model for irreversible kinetic aggregation phenomena. Such ER processes can be described by a rate equation for the evolution of the cluster-size distribution with the connection kernel Kij approximately ij , where ij is the product of the sizes of two merging clusters. Here we study that when the giant cluster is discouraged to develop by a sublinear kernel Kij approximately (ij)omega with 0<or=omega<1/2 , the percolation transition (PT) is discontinuous. Such discontinuous PT can occur even when the ER dynamics evolves from proper initial conditions. The obtained evolutionary properties of the simple model sheds light on the origin of the discontinuous PT in other nonequilibrium kinetic systems.

Full text

Page 1

arXiv:0911.4001v2 [cond-mat.stat-mech] 26 May 2010
Cluster aggregation model for discontinuous percolation transition
Y.S. Cho, B. Kahng and D. Kim
Department of Physics and Astronomy, Seoul National University, Seoul 151-747, Korea
(Dated: May 27, 2010)
The evolution of the Erd˝ os-R´ enyi (ER) network by adding edges is a basis model for irreversible
kinetic aggregation phenomena. Such ER processes can be described by a rate equation for the
evolution of the cluster-size distribution with the connection kernel Kij ∼ ij, where ij is the product
of the sizes of two merging clusters. Here we study that when the giant cluster is discouraged to
develop by a sub-linear kernel Kij ∼ (ij)ωwith 0 ≤ ω < 1/2, the percolation transition (PT) is
discontinuous. Such discontinuous PT can occur even when the ER dynamics evolves from proper
initial conditions. The obtained evolutionary properties of the simple model sheds light on the origin
of the discontinuous PT in other non-equilibrium kinetic systems.
PACS numbers: 64.60.ah,02.50.Ey,89.75.Hc
Irreversible cluster aggregations are widespread phe-
nomena occurring in a diverse range of fields, including
dust and colloid formation, aerosol growth, droplet nucle-
ation and growth, gelation transition, etc [1]. The Smolu-
chowski coagulation equation [2–4] can successfully de-
scribe such cluster aggregation processes. In linear poly-
merization, molecules with two reactive ends can react
to form long chains. In this case, the reaction kernel is
given as Kij= 1, where i and j are the masses of the two
reactants. For the aggregation of branched polymers, the
reaction kernel has the form Kij= (ai+b)(aj+b), where
a and b are constants. When clusters have a compact
shape, the reaction kernel has the form Kij∼ (ij)1−1/d,
where d is the spatial dimension. Intensive studies have
been carried out using the Smoluchowski coagulation
equation with such different kernel types [1, 5–8], and
it is known that sol-gel transitions can occur at either
finite or infinite transition points. They are continuous
transitions.
During the past decade, the evolution of complex net-
works has been of much interest to the science commu-
nities in multidisciplinary fields. To study percolation
transition (PT) during network evolution, the branch-
ing process approach [9, 10] and the Potts model formal-
ism [11] have been used. Such complex network evolution
can also be viewed as a cluster aggregation phenomenon,
and can be studied by the rate-equation approach [6].
For example, in the evolution of the classical random
network, called the Erd˝ os-R´ enyi (ER) model, an edge
is added at each time step, thereby either connecting
two separate clusters (inter-cluster edge) or increasing
the edge number in one cluster without changing clus-
ter numbers (intra-cluster edge). Fig. 1 shows that the
frequency of inter-cluster connections is dominant until
the percolation threshold. Thus, the cluster aggregation
picture of the ER network evolution comes in naturally.
In this paper, we extend the cluster aggregation dynam-
ics in networks to more general cases. Specifically, the
model is as follows: In a system composed of N vertices,
we perform the following tasks at each time step.
• Two clusters of sizes i and j are chosen with proba-
bilities qiand qj, respectively. The two clusters can
be the same. Probability qiis given as ki/?
where ki and ni are the weight and density of an
i-sized cluster, respectively.
sksns,
• Two vertices are selected randomly one each from
the selected clusters. If they are not yet connected,
then they are connected by an edge. If they are
already connected, we choose another pair of ver-
tices in the same manner until a link can be added.
Self-loop cases are excluded.
We repeat these simple steps until a given time t ≡ L/N,
where L, the number of edges added to the system, is
tuned. This model is called the cluster aggregation net-
work model hereafter. In this model, the two selected
clusters can be the same, and thus, the evolution can
proceed even after one giant cluster remains. The ER
network corresponds to the case ki= i. Here, we show
that when the weight is sub-linear, as ki = iωwith
0 ≤ ω < 1/2, a discontinuous PT occurs at a finite
transition point. Moreover, under certain initial con-
ditions, the ER dynamics also exhibits a discontinuous
PT. This observation is remarkable, because a discontin-
uous PT has rarely been discovered in irreversible kinetic
systems, except for recent observations in the ER [13]
and other networks [14–19] under the so-called Achliop-
tas process [13]. On the other hand, it is noteworthy
that the cluster aggregation network model evolves by
single-edge dynamics, as compared with the ER network
under the Achlioptas process, which involves a pair of
edges at each time step. Thus, this cluster aggregation
network model allows us to study the underlying mecha-
nism of the discontinuous PT analytically for some cases,
which is shown later.
The cluster aggregation processes in the model are de-
scribed via a rate equation for the cluster density, which
takes the following form in the thermodynamic limit:
dns(t)
dt
=
?
i+j=s
kini
c(t)
kjnj
c(t)− 2ksns
c(t),(1)
where c(t) =?
kikj/c2. The first term on the right hand side repre-
sents the aggregation of two clusters of sizes i and j with
sksns(t). The connection kernel Kij ≡

Page 2

2
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6
t ≡ L/N
0.8 1 1.2 1.4
fraction
ER
FIG. 1: (Color online) The fraction of each type of attached
edges, inter-cluster (•) or intra-cluster (▽) edges for the ER
model. Arrow indicates percolation threshold at pc = 1/2.
i + j = s and the second term represents a cluster of
size s merging with another cluster of any size.
rate equation differs from the Smoluchowski coagulation
equation in two aspects. First, the connection kernel is
time-dependent through c(t) when ω ?= 1. Second, the
second term on the right hand side of Eq. (1) includes the
process of merging with an infinite-size cluster. Hence,
Eq. (1) with ω = 1 and c = 1 describes the ER pro-
cess, while the conventional Smoluchowski coagulation
equation with ω = 1 does not, because only sol-sol reac-
tions are taken into account. However, the case including
the infinite-size cluster in the Smoluchowski coagulation
equation was also considered in Ref [6], which was called
the F-model. Owing to the presence of c(t), a PT occurs
at a finite transition point even when a PT does not occur
in the Smoluchowski coagulation equation, for example,
when ω = 0. Here, we study the cases ki= 1 (ω = 0),
ki= iωwith 0 < ω < 1, and ki= i (ω = 1), separately.
The case ω = 0: In this case, c(t) =?
the total density of the clusters, which decreases linearly
with time. The generating function of ns(t) is defined
as f(z,t) =?
range 0 < z < 1. Then, one can obtain the differential
equation for f(z,t) from Eq. (1) and solve in a closed
form as f(z,t) = (1 − t)2z/(1 − zt) for t < 1 and 0 for
t > 1 in the thermodynamic limit. Expanding f(z,t) as
a series in z, we obtain
The
sns becomes
sns(t)zs, where z is the fugacity in the
ns(t) = (1 − t)2ts−1
(2)
for t < 1. This formula shows that the cluster size dis-
tribution decays exponentially as s becomes large. Par-
ticularly, when δ ≡ 1 − t is small, ns(δ) ≈ δ2e−s/s∗with
s∗≈ 1/δ. The characteristic size s∗diverges as δ → 0.
As shown in the inset of Fig. 2(a), ns(t) is almost flat at
δ = 10−4for N = 106, indicating that large-size clusters
are relatively abundant. The merging of these clusters
causes a sudden jump in the giant cluster size, leading to
a first-order transition.
We find the giant cluster size G(t) by using the relation,
G(t) = 1−f′(1,t) ≡ 1−?′
excludes an infinite-size cluster. We find that
ssns(t), where the summation
G(t) =
?0 if
1 if
0 < t < 1,
t > 1,
(3)
10-6
10-5
10-4
10-3
10-2
10-1
100
5 10 15 20 25 30 35 40
ns
(a)
ki = 1
10-7
10-6
10-5
0 20000 40000
10-12
10-10
10-8
10-6
10-4
10-2
100
100
101
102
103
104
105
ns
(b)
ki = i0.4
10-12
10-10
10-8
10-6
10-4
10-2
100
100
101
102
103
104
105
ns
(c)
ki = i0.8
10-12
10-10
10-8
10-6
10-4
10-2
100
100
101
102
103
104
105
ns
s
(d)
ki = i
FIG. 2: (Color online) The cluster-size distributions of the
cluster aggregation network models with (a) ki = 1, (b) ki =
i0.4, (c) ki = i0.8, and (d) ki = 1 (ER model) with mono-
disperse initial condition. (a) is drawn in semi-logarithmic
scale, while the others are in double-logarithmic scales. Data
points in (a) are at t/tc = 0.3(•), 0.5(▽) and 0.7(◦), and
in (b), (c), and (d), they are obtained at t/tc = 0.50 (•),
0.95 (▽), 0.998 (◦)), and 1.003 (⋄)). Data points in the inset
of (a) are at δ = 1−t = 10−4. In (b), there exists a hump for
data points (◦). The system size N is taken as 106for (a) and
105for (b), (c), and (d). Solid lines in (b) and (c) represent
analytic formulae Eq. (4)
in the thermodynamic limit (Fig. 3(a)). Thus, the PT
is first-order at tc= 1. This result differs from what we
obtain from the Smoluchowski coagulation equation, in
which the transition point tc= ∞.
The case 0 < ω < 1: For this case, while exact solu-
tion for ns(t) is not obtained, ns(tc) is done under cer-

Page 3

3
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
G
t
(a)
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t
(b)
FIG. 3: (Color online) The density of the giant cluster size G
versus time t showing (a) discontinuous transitions in ki = 1
(◦), ki = i0.2(△), ki = i0.4(•), and (b) continuous transitions
in the cluster aggregation network model ki = i0.6(▽), ki =
i0.8(△), and the ER network (◦). N = 105in both (a) and
(b).
tain assumptions. To proceed, we define the generat-
ing function gω(µ,t) ≡?
presume that ns(tc) ∼ s−τ. Next, we use the assump-
tion made in Ref. [5, 7] for the Smoluchowski coagulation
equation that ns(t) = ns(tc)/(1 + b(t − tc)) near t = t+
where b is an s-independent constant. Then, comparing
the most singular terms in the series of the generating
functions f(eµ,tc) and g2
ssωns(t)eµs/c(t) (µ < 0), and
c,
ω(µ,tc) in µ, we find that
τ =
?1 + 2ω
3/2 + ω if
if0 < ω < 1/2,
1/2 < ω < 1.
(4)
This result is confirmed numerically in Fig 2. When t <
tc, ns(t) follows a power-law function with an exponential
cutoff for 1/2 < ω < 1, but it exhibits a hump in a large-
size region for 0 < ω < 1/2 (Fig. 2(b)).
We examine G(t) as a function of time for various ω
cases. G(t) exhibits a transition at finite tc, which is
continuous for 1/2 < ω ≤ 1, discontinuous for 0 ≤ ω <
1/2 (Fig. 3), and marginal for ω = 1/2. The first-order
transition is tested in Fig. 4 using the scaling approach
introduced in Ref. [13]. We define ∆ ≡ t1− t0, where t0
and t1are chosen as the times at which the value of G(t)
reaches 1/√N and 0.8 for the first time, respectively. We
find numerically that for 0 ≤ ω < 0.5, ∆ decays as N →
∞, while for 0.5 < ω ≤ 1, ∆ converges to a finite value.
This result suggests that the transition is discontinuous
(continuous) for 0 ≤ ω < 0.5 (0.5 < ω ≤ 1).
The case ω = 1: This case is exactly solvable as the
case of the Smoluchowski coagulation equation [6]. We
consider an arbitrary initial condition of ns(0). In this
case, c(t) =?
first moment M1(t) =?′
ing the largest cluster, is not. The generating function
g1(µ,t) =?
˙ g1= 2(g1− 1)g′
ssns(t) is conserved as c(t) = 1, but the
ssns(t), with the sum exclud-
ssns(t)exp(µs) satisfies the relation
1, (5)
where the dot (prime) is the derivative with respect to
time t (µ). Then, g1is the solution of
g1(µ,t) = 1 − H(−µ − 2t(g1(µ,t) − 1)), (6)
10-2
10-1
100
10-7
10-6
10-5
1/N
10-4
10-3
∆
ω = 0.8
= 0.6
= 0.5
= 0.4
= 0.2
FIG. 4: (Color online) Test of the discontinuous or continuous
PT for the cluster aggregation network models with various
ω values. When 0 < ω < 0.5, ∆ decays with increasing N;
however, when 0.5 < ω < 1, it converges to a finite value.
The straight lines are guidelines for eyes.
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
G
t
(a)
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t
(b)
FIG. 5: (Color online) Plot of the giant cluster size G versus
time for the ER network with different initial conditions. In
(a), the cluster-size distribution ns(0) is flat with different
cutoff values sm = Nη, with η = 0, 0.1, 0.2, and 0.3 from right
to left. In (b), ns(0) decays according to a power law with
exponent τ = 3, 2.5, 2, and 1.5 from right to left. N = 107in
both (a) and (b).
where H(µ) ≡ 1 − g1(−µ,0) is fixed by the initial con-
ditions of ns(0). The giant cluster size G(t), defined as
G = 1 − g1(0−,t), can be solved by the self-consistent
equation G = H(2tG). The obtained G(t) has the form
near tcas
G(t) =2M2
2(0)
M3(0)
(2M2(0)t − 1),(7)
where Mn(0) =
ment. Also, one finds that the second moment M2(t) ≡
?′
M2(t) =
|1 − 2M2(0)t|
for t < tc= 1/(2M2(0)), and for t > tcas t → t+
These solutions for arbitrary initial conditions are used
to study the first-order transition in the ER network be-
low. For the cluster aggregation network model, the ini-
tial condition is ns(0) = δs,1. Then, M2(0) = M3(0) = 1,
?′
ssnns(0) is the initial n-th mo-
ss2ns(t), obtained from g′
1(0−,t), behaves as
M2(0)
(8)
c.

Page 4

4
TABLE I: When the number of cluster sizes at initial time obeys a power law ns(0) = As−τfor s = 1,...,sm, listed are the
amplitude A, the second moment at initial time M2(0), the critical point tc, and the critical behavior of the giant cluster size
G(t). Type of PT is specified for each case. The listed tc and G are the ones in the thermodynamic limit.
τAM2(0)
(2−τ)sm
3−τ
sm
ln sm
s3−τ
m
ζ(τ−1)(3−τ)
ln sm
ζ(2)
ζ(τ−2)
ζ(τ−1)
ζ(2)
ζ(3)
ζ(τ−2)
ζ(τ−1)
tc
0
G(t > tc)
1
type of PT
discontinuous
(i)0 ≤ τ < 2
2−τ
s2−τ
m
1
ln sm
1
ζ(τ−1)
1
ζ(τ−1)
1
ζ(τ−1)
1
ζ(τ−1)
1
ζ(τ−1)
(ii)τ = 2
01
τ−2
3−τ
discontinuous
(iii) 2 < τ < 3
0
∝ t continuous
(iv)τ = 3
0
∝
1
2te−ζ(2)/2t
∝ (t − tc)
continuous
(v)3 < τ < 4
ζ(τ−1)
2ζ(τ−2)
ζ(3)
2ζ(2)
ζ(τ−1)
2ζ(τ−2)
1
τ−3
continuous
(vi)τ = 4
∝
t−tc
ln(t−tc)
∝ t − tc
continuous
(vii) τ > 4
continuous
and consequently, tc= 1/2, which is the well-known ER
value. The giant cluster size exhibits a continuous tran-
sition at tcwith ns(tc) ∼ s−5/2.
It is often the case that starting from ns(0) = δs,1,
the cluster-size distribution ns(t) exhibits a power-law
behavior (or with hump) in s just before or at the tran-
sition point, even when the dynamics is different from
the ER (see Fig. 2). To see how such ns evolves un-
der the ER dynamics from then on, we consider here
two particular cases in which M2(0) and M3(0) depend
on N. First, we assume that ns(0) follows a flat distri-
bution, ns(0) = n0, in the range 0 < s < sm, where
sm, the size of the largest cluster at t = 0, depends
on N as sm = Nη.Then, n0 = 2N−2η, M2(0) ∝
Nη, and M3(0) ∝ N2η.
tc(N) = 1/2M2(0) ∝ N−η, and G(t) ∼ r(2M2(0)t − 1)
for t > tc(N) from Eq. (7), where r turns out to be in
O(1).
one can show that G(t′) is the solution of G =˜H(3t′G),
where˜H(x) = 2?∞
lar function qualitatively similar to H(x) = 1 − e−xof
the standard ER problem. Hence, G(t′) has a mean field
behavior similar to the original ER case. This scaling be-
havior implies that while δG(t) ≡ G(t1)−G(t0) increases
by O(1), ∆ ≡ t1− t0does so by ∼ O(N−η). Thus, we
have a first-order transition as N → ∞ (Fig. 5(a)).
Second, we suppose that the initial condition is given
as ns(0) = As−τin the range 0 < s < sm, where A is
the normalization constant determined by the condition
?
τ, together with the initial second moment for arbitrary
Then, a PT takes place at
Thus, if time t is scaled as t′= tM2(0), then
n=1(−1)n+1xn/(n!(n + 2)) is a regu-
ssns= 1. A is given in Table I for various ranges of
sm, the critical point, and the giant cluster size G(t),
obtained from Eq. (6)
In short, the PT occurs at t = 0 for τ ≤ 3, and at finite
tcfor τ > 3. This behavior is related to the divergence
of the second moment M2(0) since time is scaled in the
form t′= 2tM2(0). The transition is discontinuous when
τ < 2, but continuous when 2 < τ < 4. This difference
originates from the fact that the ratio M2
finite for the former, while it vanishes for the latter. For
τ > 4, both M2(0) and M3(0) are finite, resulting in the
classical percolation behavior at a finite tc.
In summary, we have introduced a cluster aggregation
network model, in which discontinuous percolation
transitions occur when the connection kernel is sub-
linear as Kij ∼ (ij)ωwith 0 ≤ ω < 1/2. Even for the
ER network, a discontinuous PT can also be obtained
by using initial conditions where M2(0) diverges and
M2
2(0)/M3(0) remains finite. The simple model mani-
fests explicitly the role of the abundance of large-size
clusters just before a transition point as a mechanism of
the discontinuous PT [17]. We expect that the cluster
aggregation network model can be used to study under-
lying dynamics of the explosive percolation transition
of random ER network under the Achlioptas process [13].
2(0)/M3(0) is
This work was supported by a KOSEF grant Acceler-
ation Research (CNRC) (Grant No.R17-2007-073-01001-
0) and by the NAP of KRCF.
Note added.–After the submission of this paper, we
became aware of a work [20], which starts from the same
motivation as ours.
[1] Kinetics of aggregation and gelation, edited by F. Family
and D.P. Landau, (North-Holland, Amsterdam, 1984).
[2] M.V. Smoluchowski, Physik. Zeits. 17, 557 (1916).
[3] P.J. Flory, J. Am. Chem. Soc. 13, 3083 (1941).
[4] W.H. Stockmayer, J. Chem. Phys. 11, 45 (1943).
[5] R.M. Ziff, E.M. Hendriks, and M.H. Ernst, Phys. Rev.
Lett. 49, 593 (1982).
[6] R.M. Ziff, E.M. Hendriks, and M.H. Ernst, J. Phys. A:
Math. Gen., 16, 2293 (1983).
[7] F. Levyraz and H.R. Tschudi, J. Phys. A 14, 3389 (1981).
[8] F. Leyvraz, Phys. Rep. 383, 95 (2003).
[9] D.S. Callaway, M.E.J. Newman, S.H. Strogatz, and D.J.

Page 5

5
Watts, Phys. Rev. Lett. 85, 5468 (2000).
[10] R. Cohen, D. ben-Avraham, and S. Havlin, Phys. Rev.
E 66, 036113 (2002).
[11] D.S. Lee, K.-I. Goh, B. Kahng, and D. Kim, Nucl. Phys.
B696, 351 (2004).
[12] P. Erd˝ os, A. R´ enyi, Publ. Math. Hugar. Acad. Sci. 5, 17
(1960).
[13] D. Achlioptas, R. M. D’Souza, and J. Spencer, Science
323, 1453 (2009).
[14] R. M. Ziff, Phys. Rev. Lett. 103, 045701 (2009);
arXiv:0912.1060.
[15] Y.S. Cho, J.S. Kim, J. Park, B. Kahng, and D. Kim,
Phys. Rev. Lett. 103, 135702 (2009).
[16] F. Radicchi and S. Fortunato, Phys. Rev. Lett. 103,
168701 (2009); arXiv:0911.3549.
[17] E.J. Friedman and A.S. Landsberg, Phys. Rev. Lett. 103,
255701 (2009).
[18] A.A Moreira, E.A. Oliveira, S.D.S. Reis, H.J. Herrmann,
J.S. Andrade Jr. arXiv:0910.5918.
[19] Herm´ an D. Rozenfeld, L.K. Gallos, and H.A. Makse,
arXiv:0911.4082.
[20] S.S. Manna and A. Chatterjee, arXiv:0911.4674.