Complex self-sustained oscillation patterns in modular excitable networks

Jason Danison∗ and Miguel Perez

BMCC, The City University of New York, 199 Chambers St, New York, New York 10007-1047
(Dated: July 25, 2018)
We study the relationship between the modularity of scale-free excitable networks and their ability
to support self-sustained oscillation patterns. In the process, we introduce a new method that can
be used to analyze the propagation of waves in complex excitable networks at different levels of
detail. We find that the probability for a network of given degree distribution exponent to be able
to support self-sustained oscillations is strongly affected by its modularity. In addition, both high
and low modularity networks are more likely to exhibit long-period oscillatory patterns than those
arXiv:1807.08849v1 [physics.soc-ph] 23 Jul 2018

with intermediate modularity, but the levels of complexity of the patterns in these two cases are
different. The long period oscillations cannot be explained by the minimum length Winfree loop,
but instead arise from the interplay between two or more weakly connected loops.

PACS numbers: 05.65.+b, 89.75.Fb, 89.75.Kd, 89.75.Hc

Keywords: Suggested keywords

I. INTRODUCTION the network[32]. Finally, no attempts have been made so

far to study the statistics of ensembles of networks.
The propagation of waves in continuous excitable The main obstacle to the establishment of a self-
media[1–6] as well as in discrete lattice structures[7] has sustained pattern of oscillations on small-world and even
been an important research topic for many years[8]. The Erdős-Rényi networks is the rapid spreading of excita-
propagation of spiral and target waves in both of these tion through the high degree nodes, which places the
situations is well understood. In recent years, a number bulk of the network in a refractory state. Intuitively,
of authors have studied the sustained activity patterns it is to be expected that a modular structure will be
in various complex network models[9–13], including sim- able to mitigate this effect by limiting the propagation of
ple small-world networks[14–19] as well as Erdős-Rényi the excitations, and that it might even lead to situations
networks[20–22]. where wave patterns that are only weakly coupled propa-
The interest in these network models is driven primar- gate across separate sets of communities. The results we
ily by neuroscience[23–30]. In particular, it has been present for ensembles of networks and for different val-
shown that wave propagation in complex neuronal net- ues of the coupling coefficient confirm and quantify these
works is involved in many brain functions such as vi- considerations.
sual perception[28], cognitive processes[23, 29] and sleep-
arousal patterns[30]. In addition, the presence of high-
degree neurons is crucial for the ability of the cortex to II. METHOD
perform its information processing functions[31]. Long-
period rhythmic synchronous firing has been observed in The concept of community structure arises from the
Barabási-Albert-like scale-free networks[32] and has been fact that many networks can be naturally divided into
hypothesized to be the basis for the memorization of long subsets of nodes such that the density of connections
temporal intervals. within these subsets is higher than between them. Net-
However, all studies published so far are based on sim- works that exhibit a clear structure of this kind are called
ple network models which, with the exception of Ref. [9], modular. The simplest and most straightforward way
do not allow any independent control over important as- to quantify the modularity of an undirected unipartite
pects of network topology like clustering, degree-degree network is by means of the modularity function Q in-
correlations or community structure. The latter in par- troduced by Newman and Girvan in Refs. [35–37]. This
ticular has been shown to have a significant bearing on a function is defined as
network’s ability to perform its functions as it has gen- K
erally been observed that subsets of a network whose X X  di dj

Q= Aij − , (1)
nodes are more densely connected than in a random “null 2m
k=1 i,j∈Ck
model” are likely to perform some function together[33–
36]. In addition, the role played by the strength of the where A is the adjacency matrix of the network, {Ck }
coupling between the neurons has not yet been system- with k = 1, K is the set of communities, di is the degree of
atically studied, even though it has beed demonstrated PN
node i and 2m = i=1 di . If the community structre of a
that this parameter can strongly affect the behavior of
network is unknown, the maximization of the modularity
function provides a way to identify it[36].
We considered ensembles of random scale-free networks
∗ [email protected]; Previously known as Bogdan Danila. with tunable modularity generated using the algorithm
 2

1 (a)
u
0.8 u - (v + b) / a 43
45

0.6 31

0.4 67
5
97 37
0.2

0
26 8
-0.2 99 33

-0.4 91 46

-0.6
560 562 564 566 568 570 572
time

FIG. 1. (Color online) Amplitude u(t) (black continuous) and (b)

97 45
the factor w(t) (red dashed) for a node where DPAD fails. The
horizontal gray line above zero represents the threshold b/a. 46

43 88
99 39 8
described in Ref. [38]. These undirected networks have 44
a built-in community structure that is provided on out- 5
33
put. The properties of the ensemble are controlled by
a number of parameters which include the network size 67
31
N , the average degree hdi, the maximum degree dmax 87
and the mixing parameter µ, which represents the aver- 91
37 100

age fraction of links running between different modules.

The other parameters were kept at their default values, 26
95

including the exponent of the power-law degree distribu- 25

tion γ = −2. It is important to note that the networks

that were not fully connected were rejected.
FIG. 2. (Color online) Different degrees of simplification of
Network dynamics was defined by a variant of the Bär- the oscillating core for a network with N = 100 nodes and
Eiswirth model[4], hdi = 4: (a) with maximum individual thresholds Dth,i =
  Di,max and (b) with maximum collective threshold Dth =
dui 1 vi + b X
= ui (1 − ui ) ui − +c (uj − ui )(2) min{Di,max }.
dt ε a
j∈Ni
dvi
= f (ui ) − vi , (3) This method builds a simplified subnetwork by retain-
dt
ing only the links between each node i and the neigh-
where ui and vi are analogous to the concentrations of bor which provides the strongest driving at the moment
activator and inhibitor or to the membrane potential and when ui (t) crosses the threshold value b/a while increas-
recovery current, c is the strength of the coupling between ing. Following extensive testing, we identified many situ-
neighboring nodes, Ni is the set of neighbors of node i ations where this is not the best choice, since it can lead
and the function f (u) is defined by to a false identification of the dominant driving node.

1
Consider the example in Fig. 1, which shows the behav-
0 for u < 3
 ior of a node from a network generated with parameters
f (u) = 1 − 6.75u(1 − u)2 for 13 ≤ u ≤ 1 . (4) N = 100, hdi = 4, dmax = 15 and µ = 0.3 for which
the coupling strength was set to c = 0.5. This is a net-

1 for u > 1
work with a relatively high modularity Q = 0.564930. A
The system of differential equations was integrated us- few nodes exhibit two above-threshold maxima of u in the
ing a fourth-order Runge-Kutta routine with integration course of a period (continuous line), but only one of them
step h = 1/128. Following previous work[2, 4, 15, 17, represents true firing. The lower maximum occurs while
20, 32], we set ε = 0.04, a = 0.84 and b = 0.07 but we the node is strongly driven by a different set of neighbors,
explored a wide range of values for the coupling constant towards the end of its refractory period, but the concen-
c between 0.1 and 0.7. tration of inhibitor v is still too high and u drops as soon
A number of previous studies have used the dominant as the driving subsides. A simple way to avoid this prob-
phase-advanced driving (DPAD) method[17] to identify lem is to define the dominant phase-advanced node of i
the source of the sustained oscillations on the network. as its strongest driver at the moment when the quantity
 3

6000 80
(a) (b)
5000
60
4000

3000 40

2000
20
1000

0 0
0 0.1 0.2 0.3 0.4 0.5 700 750 800 850 900 950
25
(c) (d)
800
20
600
15

400
10

200 5

0 0
0 0.1 0.2 0.3 0.4 0.5 700 750 800 850 900
frequency time

FIG. 3. Discrete Fourier transforms ((a) and (c)) and plots of utot vs. t ((b) and (d)) for long-period oscillation patterns of two
networks with N = 100 and hdi = 4 but different modularities.

wi = ui − via+b (dashed line in Fig. 1) crosses 0 while

2. At the end of the simulation, compute the driving
increasing. This provides essentially the same timing as coefficients Dij = huj − ui i. If i and j are not
the older version in the case of true firing but avoids the connected set Dij = 0.
futile firing attempts during the refractory period.
In addition, there are situations when two or more 3. For every node i, identify the largest driving coef-
nodes provide roughly equal driving to a common neigh- ficient Di,max .
bor and, moreover, that neighbor would not be able to 4. Choose a threshold value 0 ≤ Dth ≤ min{Di,max }
fire at the right time to continue the propagation of the and if Dij < Dth set Dij = 0. Alternatively, one
wave without all major contributions. Therefore, it is can use separate thresholds 0 ≤ Dth,i ≤ Di,max for
important to be able to generate subnetworks exhibiting every node.
a variable degree of simplification. Such a subnetwork
would include the links between every node and all its 5. For every node i that does not contribute to the
major drivers. One way to achieve this is to build a “driv- driving of a neighbor set Di: = 0. Repeat this step
ing matrix” whose elements Dij are the averages of the until no such nodes are found.
differences (uj − ui ) recorded at times when wi (t) = 0
while increasing and then setting all elements below a 6. Symmetrize the matrix D by setting Dij =
certain threshold equal to zero. In addition, as a way to max(Dij , Dji ) and, if desired, set all non-zero coef-
focus on the nodes that are essential to the propagation ficients equal to 1.
of the self-sustained oscillations, one can recursively re-
move all the “dead-end” nodes which do not contribute 7. Treat the rows and columns of matrix D that con-
to the driving of any neighbor. tain non-zero elements as the adjacency matrix of
The detailed procedure is as follows: the “oscillating core” subnetwork.

Two subnetworks representing different degrees of sim-

1. For every node i and every j ∈ Ni , record the values plification of the network from which the example in
of (uj − ui ) at every moment when wi (t) = 0 and Fig. 1 was taken are shown in Figs. 2 (a) and (b), corre-
dwi
dt > 0. If the network’s phase space trajectory sponding to setting individual Dth,i = Di,max and respec-
settles on an attractor, one may wish to consider tively a global Dth = min{Di,max }. The arrows in these
only values recorded after this has happened. figures show the direction of propagation of the wave,
 4

1 100
c = 0.10
c = 0.11
0.8 c = 0.12
c = 0.13 10
c = 0.14
c = 0.15
0.6

<n>
fsust

1
0.4 c = 0.11
c = 0.12
0.1 c = 0.13
0.2 c = 0.14
c = 0.15
c = 0.20

0 0.01
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
1
100
0.8

0.6 c = 0.20 10 c = 0.30

c = 0.30 c = 0.40

<n>
fsust

c = 0.40 c = 0.50
c = 0.50 c = 0.60
0.4 c = 0.60 c = 0.70
c = 0.70
1
0.2

0 0.1
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
µ µ

FIG. 4. (Color online) The fraction fsust of networks of size FIG. 5. (Color online) The average number of successful phase
N = 100 and hdi = 4 that exhibit self-sustained oscillation space realizations hni for networks of size N = 100 and hdi = 4
patterns as a function of the mixing parameter µ, for different as a function of the mixing parameter µ, for different values
values of the coupling parameter c. of the coupling parameter c.

derived from the coefficients of the “driving matrix” Dij modularity and different coupling strength parameters c.
before symmetrization (Step 6). In Refs. [17, 20] the au- While detailed statistical results are presented in the fol-
thors describe only situations where the activity on the lowing section, here we focus on the presence of complex,
entire network can be traced to waves propagating along long-period wave patterns. In Refs. [22, 32], the length of
a single Winfree loop[17, 39]. We see from Fig. 2 (a) the period on both Erdős-Rényi and some simple models
that this description is incomplete, since multiple loops of small-world networks was connected to the length of
may be present within a certain wave pattern. Note the minimum Winfree loop on the network. Our results
that the two loops do not have the same length, which indicate that many complex excitable networks exhibit
means that the network must include a mechanism for periodicity that cannot be explained within this frame-
their synchronization. Fig. 2 (b) shows a still simple work, but results from the interplay between two or more
but much more complete picture of the network’s work- loops. This is especially true in the case of networks with
ings. Loop {46, 33, 37, 31, 5, 8} is doubled by another loop high modularity. On some of these networks the source
exhibiting weaker driving {46, 44/88, 100, 95, 25, 87, 39} of the oscillations can, in fact, be traced to a single short
with the propagation along the latter (longer) loop being loop, but the response from the rest of the network to
sped up by early activation of node 25 by node 37. At the driving produced by this loop involves other loops
the same time, these loops contribute to the driving of and spans a large number of its periods. Detailed anal-
loop {45, 97, 99, 91, 26, 67, 43} through the links {46, 45} ysis has frequently shown reversals of the direction of
and {25, 26}. It is important to note that some of the propagation along some of the loops that are part of the
additional links shown in Fig. 2 (b) are critical for the “oscillating core”. It is important to note, however, that
persistance of the wave pattern on the network. If any a given network may exhibit a large number of stable
one of links {46, 45}, {37, 25} or {39, 46} is removed, no oscillatory patterns with different periods.
self-sustained pattern can be established on this network. Qualitatively different types of long-period behav-
On the other hand, removing only {25, 26} still allows a ior have been observed in the case of low- and high-
slightly different self-sustained pattern. modularity networks. Representative results are shown
PN
This analysis method was applied systematically, com- in Fig. 3. The Fourier transform of utot = i=1 ui and
bined with discrete Fourier analysis of the oscillation a plot of utot vs. time are shown in Figs. 3 (a) and re-
pattern, to study ensembles of networks with different spectively (b) for a scale-free network of N = 100 nodes,
 5

200
Fig. 4 shows the fraction fsust of networks of size
c = 0.15
c = 0.20
N = 100, hdi = 4 and dmax = 15 that exhibit self-
sustained oscillation patterns as a function of the mixing
150 parameter µ for different values of the coupling param-
eter ranging from c = 0.1 to c = 0.7. Note that a high
value of µ means a low average modularity of the net-
work ensemble, with hQi decreasing from about 0.8 to
<n>

100
about 0.4 from left to right. These results prove that
modularity plays a critical role in a network’s ability to
50
support self-sustained oscillation patterns. The fraction
fsust exhibits a rapid overall increase with c between 0.1
and 0.15. Above c = 0.2, fsust begins to decrease with
increasing c in the case of low modularity networks due
0
0 0.1 0.2 0.3 0.4 0.5 to the fact that the hubs can now be excited by the simul-
µ
taneous firing of a smaller number of neighbors, leading
FIG. 6. (Color online) The average number of successful phase
to “epileptic” firing. However, a highly modular struc-
space realizations hni for networks of size N = 300 and hdi = 5 ture is able to mitigate this effect and fsust remains high
as a function of the mixing parameter µ, for different values for such networks until it finally starts to decrease above
of the coupling coefficient c. c = 0.7.
The dependence on c and µ is similar but much more
pronounced if one looks at the average number of success-
average degree hdi = 4 and relatively low modularity, ful phase space realizations hni (out of 1000), displayed
Q = 0.404074. The coupling parameter in this case is in Figs. 5 and 6. This count is averaged over the 100
c = 0.30. The oscillatory pattern consists of a series of realizations of the network ensemble. The error bars in
nonidentical bursts of synchronous firing interrupted by these figures represent the error on the mean. The results
longer periods of low activity, with a period PL = 64.8. in Fig. 5 are for N = 100, hdi = 4 and dmax = 15 while
This pattern is qualitatively typical for the low modu- those in Fig. 6 are for N = 300, hdi = 5 and dmax = 20.
larity networks that exhibit long periodicity. Note that Thus, the ratio dmax / hdi is essentially the same but the
the same network also supports an oscillatory pattern ratio N/ hdi is significantly larger in the second case. The
with a much shorter period, PS = 7.42, slightly different larger but sparser networks exhibit a new feature, namely
from the 7.2 period of the prominent nineth harmonic an optimum for values of the mixing parameter around
of the long period variant. The picture is quite differ- µ = 0.2 but hni still decreases rapidly with decreasing
ent in the case of high modularity networks. Results for modularity beyond that. We also counted the number
one such network, also with N = 100 and hdi = 4 but of distinct oscillation pattern variants, which is typically
a much higher modularity Q = 0.765741, are shown in much less than the number of successful phase space re-
Figs. 3 (c) and (d). A lower value of the coupling pa- alizations, but the results for that quantity are similar to
rameter c = 0.15 was used in this case. The oscillatory those for hni.
pattern now consists of three distinct, less synchronous, The last set of results concerns the relationship be-
bursts, each confined to a different part of the network. tween modularity and the average period hP i of the wave
This shows that a modular structure may indeed prevent pattern on the network. The period was calculated from
global synchronous firing, instead causing the excitation the lowest frequency peak (not necessarily the highest)
to cycle through the set of communities. The resulting in the discrete Fourier transform of utot = N
P
i=1 ui using
period in this case is PL = 82.6. The same network also the last 4096 recorded sets of values. To better correlate
exhibits two periodic oscillation patterns of periods 32.7 the average period with modularity, we considered the
and 34.3, as well as sustained non-periodic oscillations. union of all network ensembles with given N , hdi and
dmax but different values of µ and the resulting range for
Q was divided into 10 bins. The period was averaged over
III. RESULTS FOR STATISTICAL ENSEMBLES all “successful” phase space realizations of all networks
with modularity within a given bin.
In this section we present results concerning the re- Results for networks of size N = 100 are shown in
lationship between modularity and the likelihood for a Fig. 7, where the horizontal bars represent the extent of
network to exhibit self-sustained oscillation patterns, as each modularity bin and the vertical bars represent the
well as between modularity and the period of the wave error on the mean. While the dependence of the average
pattern. The statistical ensemble for each set of parame- period on modularity changes in complex ways when the
ters N , hdi, dmax and µ consisted of 100 networks. Each coupling coefficient is varied, there is a clear trend of
network was started 1000 times with random initial con- overall decrease with increasing c, again as a result of
ditions, the sets {ui } and {vi } being independently and the network hubs’ increased susceptibility. The curves
uniformly distributed between 0 and 1. for c > 0.4 are statistically indistinguishable from that
 6

100 70 70

c = 0.12
60 c = 0.14
60 c = 0.15
80
50 50
60 40 40

40 30 30
20 20
20
10 10
0 0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.3 0.4 0.5 0.6 0.7 0.8 0.3 0.4 0.5 0.6 0.7 0.8
40 40 40
35 c = 0.20
35 c = 0.30
35 c = 0.40
30 30 30
25 25 25
20 20 20
15 15 15
10 10 10
5 5 5
0 0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.3 0.4 0.5 0.6 0.7 0.8 0.3 0.4 0.5 0.6 0.7 0.8
Q

FIG. 7. (Color online) The average period hP i for different modularity ranges and different values of the coupling coefficient c
for networks of size N = 100 and average degree hdi = 4. The vertical bars represent the error on the mean while the horizontal
bars show the extent of each modularity bin.

60 200
c = 0.20 c = 0.15
50 c = 0.30 c = 0.20
c = 0.40
150
40
<PL>

<PL>

30 100

20
50
10

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
fmod fmod

FIG. 8. (Color online) The average period hP i for different ranges of fmod and different values of the coupling coefficient c for
networks of size N = 100 and hdi = 4 (a) and N = 300 and hdi = 5 (b). The vertical bars represent the error on the mean
while the horizontal bars show the extent of each fmod bin.

for c = 0.4. The most important feature is the presence tion that arises is whether there is a correlation between
of a minimum of the average period around Q = 0.6, period and some other quantity characterizing either the
which suggests different mechanisms for the generation of topology of the network or the oscillation pattern. Tests
long period oscillations at the two ends of the modularity failed to reveal any correlation between period and the
range. The results for N = 300 are qualitatively similar average degree of the network within a given modular-
but the modularity values are higher and the minimum ity range or for a given value of the mixing parameter
is around Q = 0.75. µ. Likewise, there is no correlation between period and
the size of the smallest oscillating core defined using in-
It is important to mention that, while the average pe-
dividual link thresholds Dth,i = Di,max or the size of the
riod varies as shown in Fig. 7, the actual values of the pe-
larger core defined using a global Dth = min{Di,max }.
riods in each Q range are distributed over wide intervals,
However, we found a positive correlation between period
from less than 10 up to hundreds in some bins. The ques-
 7

and the fraction fmod of modules that have at least one networks of different sizes and found that higly modu-
node in the larger core component, as shown in Figs. 8 lar networks are much more likely to be able to support
(a) and (b), where the average period is plotted against self-sustained oscillation patterns compared to low mod-
fmod . Interestingly enough, the correlation with the frac- ularity networks of the same size and average degree. In
tion of modules represented in the smaller core is much addition, the measure of the subset of points in phase
weaker. These results seem to hold regardless of network space from which a self-sustained oscillation pattern can
size, the value of the coupling parameter, or, more im- be initiated increases quickly with increasing modularity.
portantly, the value of the modularity and show that long The same is true about the number of distinct oscillation
periods are associated with propagation patterns where pattern variants.
most of the modules are involved in at least one loop, not We found that both low- and high-modularity networks
necessarily the ones that are the primary source of the can support long-period oscillations, but these oscilla-
oscillation. tions are qualitatively different, with series of synchro-
nized network-wide bursts in the case of low-modularity
networks and series of complex, localized bursts in the
case of networks with high modularity. Regardless of
modularity, such long period oscillations cannot be ex-
IV. CONCLUSIONS
plained by the length of any simple loop on the network,
but by interactions between excitations driven along dif-
We introduced a new method for analyzing the self- ferent loops. This proves that the memorization of com-
sustained oscillation patterns on complex excitable net- plex, long duration patterns does not necessarily require
works, which can be used to analyze the propagation pat- long minimum Winfree loops, as it has been shown to be
tern at different levels of detail. We studied the relation- the case with Erdős-Rényi and certain simple small-world
ship between modularity and the oscillatory patterns for networks.

[1] R. Kapral, Physica D 86, 149 (1995). [20] Y. Qian, Phys. Rev. E 90, 032807 (2014).
[2] E. Meron, Physics Reports 218, 1 (1992). [21] Y. Qian and Z. Zhang, Commun Nonlinear Sci Numer
[3] J. Chen, L. Peng, Y. Zhao, S. You, N. Wu, and H. Ying, Simulat 47, 127 (2017).
Commun. Nonlin. Sci. Numer. Simulat. 19, 60 (2014). [22] Y. Qian, X. Cui, and Z. Zheng, Scientific Reports 7,
[4] M. Bär and M. Eiswirth, Phys. Rev. E 48, R1635 (1993). 5746 (2017).
[5] O. Steinbock, J. Schütze, and S. C. Müller, Phys. Rev. [23] A. Avena-Koenigsberger, B. Misic, and O. Sporns, Na-
Lett. 68, 248 (1992). ture Reviews Neuroscience 19, 17 (2018).
[6] I. Aranson, L. Kramer, and A. Weber, Phys. Rev. Lett. [24] Z. Zheng and Y. Qian, Chin. Phys. B 27, 018901 (2018).
72, 2316 (1994). [25] M. Bazhenov, I. Timofeev, M. Steriade, and T. J. Se-
[7] Y. Qian and Z. Zhang, PLoS ONE 11, e0149842 (2016). jnowski, Nat. Neurosci. 2, 168 (1999).
[8] M. C. Cross and P. C. Hohenberg, Rev. Mod. Phys. 65, [26] N. F. Rulkov, I. Timofeev, and M. Bazhenov, J. Comput.
851 (1993). Neurosci. 17, 203 (2004).
[9] M. Müller-Linow, C. C. Hilgetag, and M.-T. Hütt, PLoS [27] G. Buzsáki and A. Draguhn, Science 304, 1926 (2004).
Comput. Biol. 4, e1000190 (2008). [28] W. M. Usrey and R. C. Reid, Annu. Rev. Physiol. 61,
[10] O. I. Kanakov and G. V. Osipov, Chaos 17, 015111 435 (1999).
(2007). [29] L. M. Ward, Trends Cognit. Sci. 7, 553 (2003).
[11] J. Ma, X. Song, J. Tang, and C. Wang, Neurocomputing [30] M. Steriade, D. A. McCormick, and T. J. Sejnowski,
167, 378 (2015). Science 262, 679 (1993).
[12] X. Liao, Q. Xia, Y. Qian, L. Zhang, G. Hu, and Y. Mi, [31] N. M. Timme, S. Ito, M. Myroshnychenko, S. Nigam,
Phys. Rev. E 83, 056204 (2011). M. Shimono, F.-C. Yeh, P. Hottowy, A. M. Litke, and
[13] F. M. M. Kakmeni, E. M. Inack, and E. M. Yamakou, J. M. Beggs, PLoS Comput Biol 12, e1004858 (2016).
Phys. Rev. E 89, 052919 (2014). [32] Y. Mi, X. Liao, X. Huang, L. Zhang, W. Gu, G. Hu, and
[14] H. Riecke, A. Roxin, S. Madruga, and S. A. Solla, Chaos S. Wu, PNAS , E4931 (2013).
17, 026110 (2007). [33] S. Fortunato, Physics Reports 486, 75 (2010).
[15] Y. Qian, Commun Nonlinear Sci Numer Simulat 27, 12 [34] M. E. J. Newman and M. Girvan, Phys. Rev. E 69,
(2015). 026113 (2004).
[16] S. Sinha, J. Saramäki, and K. Kaski, Phys. Rev. E 76, [35] M. E. J. Newman, Phys. Rev. E 74, 036104 (2006).
015101(R) (2007). [36] M. E. J. Newman, Proc. Natl. Acad. Sci. USA 103, 8577
[17] Y. Qian, X. Huang, G. Hu, and X. Liao, Phys. Rev. E (2006).
81, 036101 (2010). [37] M. E. J. Newman, Phys. Rev. E 69, 066133 (2004).
[18] Y. Kobayashi, H. Kitahata, and M. Nagayama, Phys. [38] A. Lancichinetti, S. Fortunato, and F. Radicchi, Phys.
Rev. E 96, 022213 (2017). Rev. E 78, 046110 (2008).
[19] A. Roxin, H. Riecke, and S. A. Solla, Phys. Rev. Lett. [39] W. Jahnke and A. T. Winfree, Int. J. Bifurcation Chaos
92, 198101 (2004). Appl. Sci. Eng. 1, 445 (1991).