Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

December 23, 2013 18:44 WSPC/S0218-1274 1330041

International Journal of Bifurcation and Chaos, Vol. 23, No. 12 (2013) 1330041 (20 pages)
c World Scientific Publishing Company
DOI: 10.1142/S0218127413300413

BURSTING TYPES AND STABLE DOMAINS


OF RULKOV NEURON NETWORK WITH
MEAN FIELD COUPLING
HONGJUN CAO∗
Department of Mathematics, School of Science,
Beijing Jiaotong University, Shang Yuan Cun 3,
Haidian District, Beijing 100044, P. R. China
hjcao@bjtu.edu.cn
YANGUO WU
School of Mathematics and Statistics,
Nanyang Normal University,
Nanyang 473061, P. R. China
09122156@bjtu.edu.cn

Received June 19, 2013; Revised July 26, 2013

Based on the detailed bifurcation analysis and the master stability function, bursting types and
stable domains of the parameter space of the Rulkov map-based neuron network coupled by
the mean field are taken into account. One of our main findings is that besides the square-wave
bursting, there at least exist two kinds of triangle burstings after the mean field coupling, which
can be determined by the crisis bifurcation, the flip bifurcation, and the saddle-node bifurcation.
Under certain coupling conditions, there exists two kinds of striking transitions from the square-
wave bursting (the spiking) to the triangle bursting (the square-wave bursting). Stable domains
of fixed points, periodic solutions, quasiperiodic solutions and their corresponding firing regimes
in the parameter space are presented in a rigorous mathematical way. In particular, as a function
of the intrinsic control parameters of each single neuron and the external coupling strength, a
stable coefficient of the Neimark–Sacker bifurcation is derived in a parameter plane. These
results show that there exist complex dynamics and rich firing regimes in such a simple but
thought-provoking neuron network.

Keywords: Rulkov neuron model; bursting; bifurcation analysis; master stable function; mean
field coupling.

1. Introduction et al., 2006; Ibarz et al., 2007a; Ibarz et al., 2007b;


As a simplification of classical neuron models Ibarz et al., 2011], especially in the numerical
described by ordinary differential equations like simulation of large-scale neuron networks [Rulkov
Hodgkin–Huxley equation [Hodgkin & Huxley, et al., 2004; Izhikevich & Edelman, 2008]. Over the
1952], the chaotic Rulkov map-based neuron model past decade, the chaotic Rulkov map-based neu-
has been widely used in the area of computational ron model has played an important role in explor-
neuroscience [Rulkov, 2001; de Vries, 2001; Tanaka ing the information transmission and processing of


Author for correspondence

1330041-1
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

biological neurons and collective behaviors of made by Rulkov [2001] and de Vries [2001], a
large-scale coupled map-based neuron networks more comprehensive bifurcation analysis is pre-
[Yang & Lu, 2005; Tanaka et al., 2006; Wang et al., sented from a single neuron to N identical neurons
2008; Cao & Sanjuán, 2009; Izhikevich & Edelman, coupled by the mean field. We obtain the analyti-
2008; Cao & Ibarz, 2010]. cal expressions of the flip bifurcation and the cri-
Usually, as a basic neuron model, the chaotic sis bifurcation of the coupled neurons, by which it
Rulkov map-based neuron model receives much is useful to examine different types of bursting of
attention from two aspects: one is the classifica- the coupled neurons in a rigorous way rather than
tion of different kinds of bursting, and the other is only by numerical simulations. Different from other
whether the bursting can be recovered or destroyed works, much attention is paid to possible bifurca-
in a population of such nonbursting or bursting tion conditions leading to other types of bursting
neurons when they are coupled via various net- not only the square-wave bursting.
work connections. Among them, it is well known We find there are at least two kinds of trian-
that the single chaotic Rulkov map-based neuron gle bursting types in the coupled Rulkov map-based
model can produce the square-wave bursting due to neuron model. The first kind of triangle bursting
the bistability and the suitable position of a slow can be determined by a crisis bifurcation, a supflip
nullcline [Rulkov, 2001; de Vries, 2001]. According bifurcation and a saddle-node bifurcation, and the
to the classification of bursting mappings [Izhike- second one can be determined by two saddle-node
vich & Hoppensteadt, 2004], this kind of square- bifurcations, a subflip bifurcation, and a supflip
wave bursting belongs to the fold/homoclinic type, bifurcation. The duration between the active phase
namely, the square-wave bursting can be predicted and the silent phase of the first triangle burst-
by a fold (saddle-node) bifurcation and a crisis ing is the longest. Interestingly, after the mean
(homoclinic) bifurcation [Rulkov, 2001; de Vries, field coupling, we find that there are at least two
2001]. Besides that, few have considered whether kinds of striking transitions: the first transition is
there exist other kinds of bursting. Although de from the square-wave bursting to the triangle burst-
Vries mentioned that other types of bursting like the ing, and the second one is from the spiking of
triangle bursting (tapered type) could exist even in the single neuron to the square-wave bursting after
a single Rulkov map-based neuron model [de Vries, couplings.
2001], there is still a lack of adequate arguments to Secondly, we carry out a comprehensive qualita-
support this possibility. As far as the second aspect tive analysis and a codimension-2 bifurcation anal-
is concerned, Rulkov reported how the synchroniza- ysis for the coupled neurons. The goal is to clarify
tion among chaotically bursting neurons can lead the stable domains in the parameter space among
to the onset of regular bursting by establishing a fixed points, periodic solutions, quasiperiodic solu-
neuron network model with a mean field coupling tions and their corresponding firing regimes of
[Rulkov, 2001]. de Vries revealed that the bursting the coupled neuron network in a two-dimensional
can be recovered in a population of such nonburst- parameter plane. In particular, we make a detailed
ing neurons when they are coupled via the mean computation of a coefficient of the Neimark–Sacker
field, which is called the emergent phenomenon, and bifurcation [Kuznetsov, 1999] in a parameter plane.
this kind of emergent bursting in the network is due To attain this goal, the normal form restricted on
to coupling alone and is very robust to changes in the center manifold is calculated. As a function of
the coupling strength by using the geometric bifur- the intrinsic control parameters of each single neu-
cation analysis [de Vries, 2001]. While, it is still ron and the external coupling strength, the stability
not fully understood in the transition mechanism coefficient demonstrates that there exist different
why the bursting can be recovered or destroyed in a parameter regions where complex dynamics and
population of such nonbursting or bursting neurons rich behaviors in biological neurons arise in the sim-
after couplings. ple chaotic Rulkov neuron network.
Thus, we intend to characterize mathemati- These results show that there exist complex
cally how the transition takes place among differ- dynamics and rich firing regimes in such a sim-
ent kinds of firing regimes with or without the ple but thought-provoking neuron network. On the
mean field coupling. In this paper, first of all, based one hand, these dynamics and firing regimes are
on the previous geometrical bifurcation analysis very useful for simulating collective behaviors in

1330041-2
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

large-scale neural networks. On the other hand, Suppose that R(xn ) = xn+1 , then the fixed
these analyses and results can be easily extended points of Eq. (2) satisfy the following equation
to a lot of neuron networks.
The layout of this paper is as follows. In Sec. 2, R(xn ) = f (xn ) + γ + εxn = xn . (4)
we give a description of the chaotic Rulkov neuron
map with the mean field coupling. Bifurcation of the 3. Bifurcation of the Single
single parameter is presented in Sec. 3. Sections 4 Parameter
and 5 discuss firing regimes of the single neuron,
and the coupled neurons with the mean field cou- 3.1. Saddle-node bifurcation
pling, respectively. Stability analysis of N identi- Assume that there exists a bifurcation parameter
cal Rulkov neurons with the mean field coupling γ0 and a fixed point x0 satisfying the following four
includes the Neimark–Sacker bifurcation that is dis- conditions:
cussed in Sec. 6. Finally, we sum up our results with
some comments in Sec. 7. (i) R(x0 ) = x0 ;
(ii) ∂R
∂x (x0 , γ0 ) = 1;
2. Model Description ∂R
(iii) ∂γ (x0 , γ0 ) = 0;
We consider the following neuron network with the ∂2R
(iv) ∂x2 (x0 , γ0 ) = 0,
mean field coupling [Rulkov, 2001; de Vries, 2001]:
 then a saddle-node bifurcation will occur [Wiggins,

 N
xn+1,i = f (xn,i ) + yn,i + ε

xn,j ,
1990].
N (1) From Eq. (4), it is easy to know that ∂R∂γ = 1
 j=1

 for any√x and γ. There exist two zero points at
y 2 −1)
x = ± 33 such that ∂∂xR2 = 2α(3x
2
n+1,i = yn,i − η(xn,i − σi ), = 0, while
(1+x2 )3
where f (xn,i ) = α/(1 + x2n,i ), the additional sub- the two zero points are not roots of the equation
∂R −2αx
script i indicates the ith neuron, xn,i and yn,i repre- ∂x = (1+x2 )2 + ε = 1.
∂R
sent the transmembrane voltage and the slow gating It is noted that if x < 0, then ∂x > 0. Other-
2
process of the ith neuron, respectively, n is the dis- wise, if x > 0, then ∂R
= 2α(3x
< 0. From −1)
∂2R
2 3 ,
∂x
√ (1+x )∂x2
crete time (n = 1, 2, . . .). N is the total number
we have the following conclusion: if x > 3/3, then
of neurons, α, σ, and η are control parameters of
R√(x) > 0, and√R (x) increases monotonously; if
the single neuron, in which α and σ are O(1), while
− 3/3 < x < 3/3, then R (x) < 0,√and R (x)
0 < η  1. ε is the strength of the coupling. Due
decreases monotonously; when x < − 3/3, then
to 0 < η  1, xn,i is called the fast variable, while
R (x) √
> 0 and R (x) increases monotonously; if
yn,i is referred to as the slow variable.
x√= − 3/3, then R (x) takes
√ the maximum value
A synchronized manifold of Eqs. (1) is defined
3 3α/8 + ε, while if√x = 3/3, then R (x) take the
by Π = {(x, y) | x = x1 = x2 = · · · = xN = σ, y =
minimum value −3 3α/8 + ε. Therefore, we have
y1 = y2 = · · · = yN = σ − f (σ) − εσ}. Then,
the following conclusion.
the dynamics of Eqs. (1) restricted on the synchro-

nized manifold Π can be described by the follow- Proposition 1. If α ≥ (1 − ε)8 3/9, then there
ing two-dimensional system [Rulkov, 2001; de Vries, exists a saddle-node bifurcation.
2001]:


α √ R (x) is a continuous function of
In addition,
 x, so if α > 8 3/9
xn+1 =
1 + x2n
+ yn + εxn ,

√ + ε, then the maximum value
(2) Max{R (x)} = 3 3α/8 + ε > 1. There exist two


yn+1 = yn − η(xn − σ). pairs
√ of points
√ (x01 , γ) and (x02 , γ), which x01 <
− 3/3, − 3/3 < x02 < 0, such that R (x) = 1,
By using the fast–slow decomposition technique then we have the following stability conditions.
[Rinzel, 1985, 1987], Eqs. (2) can be further reduced
into the following one-dimensional fast subsystem if Proposition 2
we suppose that yn = γ:
(i) If x < x01 , then 0 < R (x) < 1, which means
xn+1 = f (xn ) + γ + εxn . (3) that the fixed points of Eq. (2 ) are stable;

1330041-3
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu


(ii) If x01 < x < x02 , then R (x) > 1, which means  ∂R −2αx


that the fixed points of Eq. (2 ) are unstable;  ∂x = (1 + x2 )2 + ε = 1,
(iii) If x02 < x < 0, then 0 < R (x) < 1, and the

 α
fixed points of Eq. (2 ) are stable. 
x = + γ + εx,
1 + x2
By solving the first two equations given in the
conditions of the saddle-node bifurcation the analytical expression of the saddle-node bifur-
cation is obtained by Rulkov [2001] as follows:

−18γ(1 − ε)2 − 2γ 3 ± [2γ 2 − 6(1 − ε)2 ] γ 2 − 3(1 − ε)2
α= . (5)
27(1 − ε)2

3.2. Flip bifurcation


From
If there exists a bifurcation parameter γ0 and a fixed 
point x0 satisfying the following four conditions  ∂R −2αx


 ∂x = (1 + x2 )2 + ε = −1,
(i) R(x0 ) = x0 ;

 α
(ii) ∂R
∂x (x0 , γ0 ) = −1; 
x = + γ + εx,
∂R 1 + x2
(iii) ∂γ (x0 , γ0 ) = 0;
∂2R we at first obtain the following two analytical
(iv) ∂x2
(x0 , γ0 ) = 0,
expressions:
then a flip bifurcation
√ exists [Wiggins, 1990]. 
When√ x = 3/3, R (x) takes the minimum γ ± γ 2 + (1 − 3ε)(1 + ε)
x= ,
value −3 3α/8 + ε, then we have the following (1 − 3ε) (6)
conclusion:
√ α = [(1 − ε)x − γ](1 + x2 ).
Proposition 3. If α ≥ (1 + ε)8 3/9, then there
exists a flip bifurcation. Then the analytical expression of the flip bifurca-
tion is obtained as follows:

−18γε3 + 18γε2 + 2γε − 2γ 3 ε − 2γ 3 − 2γ ± (2γ 2 ε + 14ε2 − 10ε + 2 + 2γ 2 − 6ε3 ) γ 2 − (ε + 1)(3ε − 1)
α= .
(3ε − 1)3
(7)
In particular, when ε = 0, the analytical expression
of flip bifurcation is obtained by Rulkov [2001] as From Eqs. (4) and (9), we let F (x, α) =
follows: 2 − ε + αx + (2 − ε)x2 , and it is noted that
 limx→±∞ F (x, α) = +∞ if 2 − ε > 0. There exists
α = 2(γ 2 + 1)(γ ± γ 2 + 1). (8)
only a solution x = −α/(4 − 2ε) of ∂F/∂x =
α + 2(2 − ε)x = 0. Because F (x, α) is a continuous
3.3. Crisis bifurcation function of x, F (x, α) is able to take the minimum
As we have known from the above analysis that value 2 − ε − α2 /4(2 − ε) at x = −α/(4 − 2ε), then
if x < 0, then R(x) increases monotonously, while we have the following conclusion.
if x > 0, then R(x) decreases monotonously. R(x)
Proposition 4. If α ≥ 4 − 2ε, then there exists the
can take the maximum value R(0) = α + γ. If the
maximum value is iterated onto a stable fixed point, crisis bifurcation.
then the active phase of the single neuron will be
From Eq. (9), we obtain the analytical expres-
changed into the silent phase. Namely, the crisis
sion of the crisis bifurcation as
bifurcation may take place if the following condi-

tion is satisfied: −γ ± γ 2 − 4[(1 − ε)2 + 1]
α= − γ. (10)
R(x) = R(R(0)). (9) 2(1 − ε)2

1330041-4
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

In particular, when ε = 0, the analytical expres- the flip bifurcation (the dashed line), respectively.
sion of the crisis bifurcation is obtained by Rulkov The range of α is divided into the following four
[2001] as: intervals by four different values of α1,2,3,4 .

−3γ ± γ 2 − 8 Proposition 6
α= . (11)
2 √
(i) When 8 3/9 < α < 2.58, there coexist two
To sum up the above bifurcation analyses, we have saddle-node bifurcation values γsn1,sn2 , and
the following conclusion: two flip bifurcation values γf p1,f p2 , and their
Proposition 5
order is γsn1 < γsn2 < γf p1 < γf p2 ;
√ (ii) When 2.58 < α < 4, there coexist two saddle-
(i) If (1 − ε)8 3/9 < α < 4 − 2ε, then there node bifurcation values γsn1,sn2 , and two flip
coexist the saddle-node bifurcation and the flip bifurcation values γf p1,f p2 , their order becomes
bifurcation; √ γsn1 < γf p1 < γsn2 < √ γf p2 ;
(ii) If 4 − 2ε < α < (1 + ε)8 3/3, then there (iii) When 4 < α < 8 3/3, there coexist two
coexist the saddle-node bifurcation, the flip saddle-node bifurcation values γsn1,sn2 , two flip
bifurcation, and the
√ crisis bifurcation; bifurcation values γf p1,f p2 , and two crisis bifur-
(iii) If α = (1 + ε)8 3/3, then the saddle-node cation values γcs1,2 , and their order is γsn1 <
bifurcation is equal to the crisis bifurcation. γf p1 < γcs1 < γcs2 < γsn2 < γf p2 .

Corresponding to the four intervals of γ, four


4. Firing Regimes of the Single bifurcation diagrams are presented in Figs. 2(a)–
Neuron 2(d), respectively. Here and throughout this paper,
When ε = 0, based on the above mentioned bifur- the S-shaped curve denotes the branch of fixed
cation analysis of the single parameter, we discuss points of Eq. (1), the rectangle stands for the
different firing regimes of the single neuron in this saddle-node bifurcation, the ellipse denotes the flip
section. bifurcation, the diamond denotes the crisis bifurca-
In Fig. 1, three bifurcation curves entitled αsn , tion, xmin and xmax stand for the minimum values
αcs , and αf p are presented in the parameter plane and maximum values of xn . The slow nullcline is
(γ, α) when ε = 0, their corresponding bifurca- always kept fixed at xn = σ = −1.
tion values about γ are γsn , γcs , and γf p , which
denote the saddle-node bifurcation (the solid line),
4.1. Silent state
the crisis bifurcation (the dashed-dotted line), and √
When 8 3/9 < α < 2.58, the corresponding bifur-
cation diagram is presented in Fig. 2(a), in which
7 there are six bifurcation values located on the S-
αsn shaped curve of the fixed points: two saddle-node
α
6 αcs fp
bifurcation values γsn1,sn2 , and two flip bifurcation
√ values γf p1,f p2 in (γ, xn ) plane, and their relative
α4 = 8 3/3
5 position is γsn1 < γsn2 < γf p1 < γf p2 . The corre-
α3=4 sponding time evolution and phase portrait are pre-
4
α

sented in Figs. 3(a) and 3(b), respectively, in which


there exists a limit cycle in the phase plane (yn , xn ),
3 α2=2.58 which corresponds to the silent state of the single
neuron in (n, xn ) plane.
2

α1 = 8 3/9
1 4.2. Triangle bursting
−4 −3 −2
γ
−1 0 1 √
When 8 3/9 < α < 4, the fast subsystem (3)
Fig. 1. The two-parameter bifurcation diagram for the sin- exhibits the saddle-node bifurcations and the flip
gle neuron of the fast subsystem of Eq. (3) with ε = 0, bifurcations except for the crisis bifurcations. From
σ = −1, and η = 0.001 [Rulkov, 2001; de Vries, 2001]. Eqs. (6) and (8), it is easy to know that the

1330041-5
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

2 5

1.5 4
xmax
3
1
γfp2 xmax
2
0.5
1 γfp1 γfp2
γfp1
xn

xn
0
xmin 0
γsn1
−0.5
−1
xmin
−1
−2
γ γ
sn2 sn2
−1.5 −3
γsn1
−2 −4
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 −5 −4 −3 −2 −1 0 1
γ γ

(a) (b)

5 5

4 4
xmax
x
max
3 3

2 2

γ γfp1
xn

xn

1 γfp1 fp2 1 γfp2


0 0
γ γ
sn1 sn1
−1 γcs1 xmin
−1 γcs1
γsn2 xmin
−2 −2 γ
γ sn2
cs2
−3 −3
−5 −4 −3 −2 −1 0 1 −5 −4 −3 −2 −1 0 1
γ γ

(c) (d)
Fig. 2. Bifurcation diagrams of Eqs. (2) with ε = 0, σ = −1, η = 0.001, and (a) α = 2, (b) α = 3.9, (c) α = 4.15 and
(d) α = 4.7.

saddle-node bifurcation γsn2  and the√ flip bifurca- In this interval of α, since there is no crisis bifurca-
tion γf p1 intersect at α = 87 13 16 2
≈ 2.58. By tion, xn moves between the maximum values and
7 + 7
Proposition 6, when 2.58 < α < 4, there exist four the minimum values of xn . Therefore, the single
bifurcation values, and their relative positions are neuron gives rise to the triangle bursting. Once xn
γsn1 < γf p1 < γsn2 < γf p2 . Seen from Fig. 2(b), dur- moves toward left along the branch of fixed points
ing the silent phase, if iterates of the map are on or and falls to the lower stable branch of fixed points
near the stable fixed points on the bottom branch of at γsn1 , the single neuron becomes the silent state
the S-shaped curve, and all of them are below the again. To sum up, when 2.58 < α < 4, the single
y-nullcline xn = σ = −1, then yn increases slowly. neuron produces the triangle bursting between γsn1
When yn moves to γsn2 , namely, when yn > γsn2 , and γsn2 . This is a so-called fold-flip type bursting
then the single neuron gives rise to the active phase. [Izhikevich & Hoppensteadt, 2004]. The correspond-
During the active phase, because iterates lie above ing time evolution and the phase portrait of the
xn = σ = −1 on average, yn decreases slowly. tapered bursting are shown in Figs. 3(c) and 3(d).

1330041-6
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

4 −0.95

xn

xn
0 −1
−4 −1.05
0 5000 10000 −2.004 −2 −1.996
n yn

(a) (b)

4 5
xn

xn
0 0
−4 −5
0 5000 10000 −5 −4 −3 −2
n yn

(c) (d)

4 5
xn

xn
0 0
−4 −5
0 5000 10000 −3 −2.8 −2.6
n yn

(e) (f)

4 5
xn

xn

0 0
−4 −5
0 5000 10000 −3 −2.9 −2.8
n yn

(g) (h)
Fig. 3. Time evolutions and phase portraits of Eqs. (2) with ε = 0, σ = −1, η = 0.001, and (a, e) α = 2, (b, f) α = 3.9,
(c, g) α = 4.15 and (d, h) α = 4.7.

In addition, it is noted that the duration of the a periodic way. It is the so-called homoclinic/fold
tapered bursting is decided by the distance ∆γ1 = type of the square-wave bursting [Izhikevich & Hop-
|γsn2 − γsn1 |. Figure 4(a) shows that the duration of pensteadt, 2004]. The corresponding time evolution
the tapered bursting increases as α increases. and the phase portrait of the square-wave burst-
ing are shown in Figs. 3(e) and 3(f). The duration
4.3. Square-wave bursting of square-wave bursting is decided by the distance
√ ∆γ2 = γsn2 − γcs2 . Contrary to the triangle burst-
When 4 < α < 8 3/3, there coexist the saddle- ing, Fig. 4(b) demonstrates that the duration of the
node bifurcation, the flip bifurcation and the crisis square-wave bursting decreases with the increasing
bifurcation. The corresponding bifurcation diagram of α.
is shown in Fig. 2(c). During the silent phase, iter-
ates of Eq. (2) changes in a very similar way as
the triangle bursting. When yn > γsn1 , the single 4.4. Spiking

neuron changes to the active phase. While, dur- When α > 8 3/3, there exist still two saddle-
ing the active phase, because iterates of Eq. (2) lie node bifurcations γsn1,sn2 , and two flip bifurcations
above xn = σ on average, yn decreases slowly. At γf p1,f p2 , but there exists only a crisis bifurcation
the right crisis bifurcation γcs2 , the minimum val- γcs1 as shown in Fig. 2(d). In this situation, both
ues of xn intersect with the middle branch of the the fixed points on the bottom and middle branches
unstable fixed point, the single neuron backs to the of the S-shaped curve are located within the chaotic
silent phase. It leads to the square-wave bursting in attractor. When yn > γsn2 , the single neuron goes

1330041-7
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

1.4 0.07

1.3
0.06
1.2
0.05
1.1

1 0.04

∆γ
∆γ

0.9 0.03

0.8
0.02
0.7
0.01
0.6

0.5 0
2.5 3 3.5 4 4.2 4.3 4.4 4.5 4.6
α α

(a) (b)
Fig. 4. The duration of the bursting when α is varied, where ε = 0, σ = −1, and η = 0.001: (a) The duration of the triangle
bursting and (b) the duration of the square-wave bursting.

into the chaotic oscillation and moves right. When parameter of Eq. (2) corresponding to two pairs of
the average values of xn are below xn = σ = −1, parameter values ε = 0.2, α = 3.8, and ε = 0.4,
iterates of Eq. (2) increase slowly and move right. α = 4.15, respectively.
Once iterates of Eq. (2) lie above xn = σ on aver- Seen from Figs. 5(a)–5(d), there coexist two flip
age, then iterates of Eq. (2) decrease and the single bifurcation points located at the upper branch of
neuron moves left. So the single neuron produces the fixed point curve. When xn moves along the
the spiking regime. The corresponding time evolu- lower stable branch of fixed points, the coupled neu-
tion and phase portrait of spiking regime are shown ron is in the resting state, and yn increases slowly.
in Figs. 3(g) and 3(h). When it moves past the right saddle-node bifurca-
tion γsn1 , the fixed point becomes unstable, and xn
moves to the vicinity of the upper branch of fixed
5. Firing Regimes of the Coupled
points between γf p1 and γf p2 . Since there is no crisis
Neurons with Mean Field bifurcation, xn moves between the maximum value
Coupling and the minimum value of iterates of xn . In this
In this section, we further explore the mechanism process, the coupled neuron gives rise to the trian-
responsible for the emergent phenomenon [de Vries, gle bifurcation. Once xn moves along the curve of
2001], and see what effect will take place after the fixed point and falls to the lower stable branch of
mean field coupling. fixed points at γsn2 , the coupled neuron goes back
to the resting state. So this is the mechanism for
the triangle bursting.
5.1. Triangle bursting Figure 5(e) shows time evolutions from n =
From Proposition 6, when 2.58 < α < 4, a single 1, . . . , 20 000 when ε = 0 to n = 20 001, . . . , 30 000
neuron can give rise to the triangle bursting, which when ε = 0.2. In this case, the single neuron and the
is determined by the saddle-node bifurcation and coupled neurons are both in the triangle bursting
the flip bifurcation. So when the coupled neurons regime, but the size of the triangle bursting is obvi-
have the saddle-node bifurcation and the flip bifur- ously larger than that of the single neuron; while in
cation, can the coupled neurons also produce the Fig. 5(f), the transition takes place from the square-
triangle bursting? In Figs. 5(a)–5(d), we give bifur- wave bursting of the single neuron with ε = 0
cation diagrams of the two-parameter and the single and α = 4.15 for n = 1, . . . , 20 000 to the triangle

1330041-8
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

bursting of the coupled bursting with ε = 0.4 and square-wave bursting, which is determined by the
α = 4.15 for n = 20 001, . . . , 30 000. saddle-node bifurcation and the crisis bifurcation.
To sum up, there exists the transition from the So what effect will take place if the coupled neurons
square-wave bursting to the triangle bursting after have also the crisis bifurcation and the saddle-node
the mean field coupling. bifurcation?
It can be seen from Figs. 6(a)–6(d), there coex-
ist six bifurcations values γsn1,sn2 , γf p1,f p2 , and
5.2. Square-wave bursting γcs1,cs2 , with the increase of the parameter α,
√ previous bifurcation analysis, when 4 <
From the γcs2 → γsn2 . Based on these bifurcation analyses,
α < 8 3/3, a single neuron will give rise to the two arrays of time evolutions are presented in

10 3
9
γ 2
γ sn
8 cs
1 γ γfp2
7
γ γsn1 fp1
fp
6 0
5
xn
−1
α

4
3 −2
γ
2 sn2
−3
1
−4
0
−1 −5
−5 −4 −3 −2 −1 0 1 −5 −4 −3 −2 −1 0 1
γ γ

(a) (b)

10 3
γ
9 cs γ 2
sn
8
γfp 1 γfp1
γ
7 fp2
0
6 γsn1
−1
5
xn

−2
α

4
3
−3 γsn2
−4
2
1 −5

0 −6

−1 −7
−5 −4 −3 −2 −1 −5 −4 −3 −2 −1 0 1
γ γ

(c) (d)
Fig. 5. (a) The two-parameter bifurcation diagram for the coupled neurons with ε = 0.2, α = 3.8, (b) when α = 3.8, the
corresponding single parameter bifurcation diagram for the coupled neurons with ε = 0.2, (c) the two-parameter bifurcation
diagram for the coupled neurons with ε = 0.4, α = 4.15, (d) when α = 4.15, the corresponding single parameter bifurcation
diagram for the coupled neurons with ε = 0.4, (e) time evolutions from n = 1, . . . , 20 000 when ε = 0 to n = 20 001, . . . , 30 000
when ε = 0.2. In this case, the single neuron and the coupled neurons are both in the triangle bursting regimes and (f) the
transition from the square-wave bursting of the single neuron with ε = 0 and α = 4.15 for n = 1, . . . , 20 000 to the triangle
bursting of the coupled bursting with ε = 0.4 and α = 4.15 for n = 20 001, . . . , 30 000.

1330041-9
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

6 6

4 4

2 2
xn

xn
0 0

−2 −2

−4 −4

−6 −6

1 1.5 2 2.5 3 1 1.5 2 2.5 3


n 4
x 10 n 4
x 10

(e) (f)
Fig. 5. (Continued)

Figs. 6(e) and 6(f), respectively. Seen from Fig. 6(e), square-wave bursting of the coupled bursting with
the square-wave bursting of the single neuron with ε = 0.2 and α = 4.7.
ε = 0 and α = 4.4 can be continually maintained In addition, the distance between γcs2 and γsn2
after the coupling with ε = 0.2 and α = 4.4. While, is obtained by numerical simulations shown in Fig. 7
there exists another transition from the spiking of corresponding to α = 4.4, while ε is varied, in which
the single neuron with ε = 0 and α = 4.7 to the the crisis bifurcation γcs2 is always less than the

10
3
9
γ
8 sn
2
7 γcs
γfp 1 γfp2
6 γ
fp1
5 0
xn

γsn1
α

4
−1 γ
cs1
3
γcs2
2 −2
1 γ
−3 sn2
0
−1 −4
−5 −4 −3 −2 −1 0 1 −5 −4 −3 −2 −1 0 1
γ γ

(a) (b)
Fig. 6. (a) The two-parameter bifurcation diagram for the coupled neurons with ε = 0.2, α = 4.4, (b) when α = 4.4, the
corresponding single parameter bifurcation diagram for the coupled neurons with ε = 0.2, (c) the two-parameter bifurcation
diagram for the coupled neurons with ε = 0.2, α = 4.7, (d) when α = 4.4, the corresponding single parameter bifurcation
diagram for the coupled neurons with ε = 0.2, (e) the square-wave bursting of the single neuron with ε = 0 and α = 4.4 for
n = 1, . . . , 20 000 is kept after the coupling with ε = 0.2 and α = 4.4 for n = 20 001, . . . , 30 000 and (f) the transition from the
spiking of the single neuron with ε = 0 and α = 4.7 for n = 1, . . . , 20 000 to the square-wave bursting of the coupled bursting
with ε = 0.2 and α = 4.7 for n = 20 001, . . . , 30 000.

1330041-10
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

10 4
9
3
8
7 2
γfp2
6 1 γfp1
5

xn
0
α

4 γsn1
−1
γcs1
3
2 −2 γsn2
1
−3
0 γ
cs2
−1 −4
−5 −4 −3 −2 −1 0 1 −5 −4 −3 −2 −1 0 1
γ γ

(c) (d)

3 3

2 2

1 1
xn

xn

0 0

−1 −1

−2 −2

−3 −3
1 1.5 2 2.5 3 1 1.5 2 2.5 3
n x 10
4 n x 10
4

(e) (f)
Fig. 6. (Continued)

saddle-node bifurcation γsn2 . In particular, there


exists a critical point ε0 such that γcs2 = γsn2 .

5.3. Triangle bursting


As we have known from the previous bifurcation
analysis that the triangle bursting and the square-
wave bursting may take place when either the cor-
responding parameter γ is between γf p1 and γf p2 ,
or γsn2 < γf p2 . What will happen if the coupling
strength ε is increased such that γsn2 > γf p2 ?
To answer this question, two bifurcation diagrams
of Eq. (3) are presented in Figs. 8(a) and 8(b)
Fig. 7. The distance between γcs2 and γsn2 when α = 4.4, with regard to ε = 0.8, α = 4.15, and α = 4.7,
and ε is varied. respectively.

1330041-11
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

55
50 γfp1
γ
0 fp2
45
γsn1
40
γ
fp
35 −5 γ
sn2
30

n
α

x
25 −10
20 γ
γcs sn
15
−15
10
5
−20
0
−5 −4 −3 −2 −1 0 −5 −4 −3 −2 −1 0
γ γ

(a) (b)

2 γ (d)
0
fp1
0 γfp2 −5
γsn1
n
x

−2 −10
γsn2 −15
−4
−20
−6 0 1 2 3 4
n 4
xn

x 10
−8

−10 (e)
0
−12
−5
n

−14
x

−10
−16
−15
−5 −4 −3 −2 −1 0
γ 0 1 2 3 4
n 4
x 10
(c)
Fig. 8. (a) The two-parameter bifurcation diagram for the coupled neurons with ε = 0.8, the corresponding single parameter
bifurcation diagrams and time evolutions for the coupled neurons with (b) ε = 0.8, α = 4.7, (c) ε = 0.8, α = 4.15, (d) ε = 0.8,
α = 4.7 and (e) ε = 0.8, α = 4.15.

Seen from Figs. 8(a)–8(c), γsn2 > γf p2 . The that the duty cycle of the first type of the triangle
square-wave bursting of the single neuron may pro- bursting is the longest one among any type of burst-
duce the first kind of triangle bursting after cou- ing shown in paper.
pling shown in Fig. 8(d), in which the active phase
of bursting is determined by γf p1 and γf p2 , while
the silent phase of bursting is determined by γsn1 5.4. Spiking

and γsn2 . Interestingly, there exists another kind of From Fig. 2, when α > 8 3/3, a single neuron
triangle bursting shown in Fig. 8(e), where it means √ produce the spiking. In particular, when α =
will
that the spiking neuron can produce the second kind 8 3/3, the saddle-node bifurcation γsn2 intersects
of triangle bursting after coupling, whose character- √ crisis bifurcation γcs2 . Therefore, when
with the
istic is the active phase of bursting determined by α > 8 3/3, or when the saddle-node bifurcation
γcs2 and γf p2 , while the silent phase of bursting is γsn1 intersects with the crisis bifurcation γcs1 , the
determined by γcs2 and γsn2 . In addition, it is noted single neuron may give rise to the spiking. Under

1330041-12
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

5.5 3
γcs
5
2
4.5
1
4 γ
γ fp1
sn 0
3.5

xn
γsn1
α

3 −1 γcs1
γ
2.5 fp
−2
γsn2
2
−3
1.5

1 −4
−5 −4 −3 −2 −1 0 1 2 −5 −4 −3 −2 −1 0 1
γ γ

(a) (b)

3 (d)
2

2 xn
0

1
γfp1 −2

0 1 1.5 2 2.5 3
n
γ
4
xn

x 10
−1
sn1 γ
cs1
(e)
−2 2
γ
sn2
xn

−3 0

−4 −2
−5 −4 −3 −2 −1 0 1
γ 1 1.5 2 2.5 3
n 4
x 10
(c)

Fig. 9. (a) The two-parameter bifurcation diagram for the coupled neurons with ε = −0.2, α = 4.15, (b) the correspond-
ing single parameter bifurcation diagrams for the coupled neurons with α = 4.15, ε = −0.2, (c) α = 4.7, ε = −0.2, time
evolutions: (d) α = 4.15 and ε = 0 for n = 1, . . . , 20 000, and ε = −0.2 for n = 20 001, . . . , 30 000. In this case, the single
neuron without couplings and the coupled neurons are both in the square-wave bursting regimes and (e) α = 4.7 and ε = 0 for
n = 1, . . . , 20 000, and ε = −0.2 for n = 20 001, . . . , 30 000. In this case, the single neuron without couplings and the coupled
neurons are both in the spiking regime.

the effect of the mean field coupling, what will hap- single neuron without couplings can also become
pen when γcs2 ≈ γsn2 or when γcs2 disappears? the spiking after the negative coupling.
In Figs. 9(a)–9(e), we give the two-parameter
bifurcation diagram, the corresponding single
6. Stability Analysis of N Identical
parameter bifurcation diagrams, and time evolu-
tions with α = 4.15, ε = −0.2 and α = 4.7,
Rulkov Neurons with Mean Field
ε = −0.2, respectively. Seen from Fig. 9(d), the Coupling
square-wave bursting of the single neuron without In this section, we discuss the stability condition
couplings is still maintained after the negative cou- of Eq. (1), the corresponding variational equation
pling. While, seen from Fig. 9(e), the spiking of the of Eqs. (1) restricted to Π is δxn+1 = Jδxn . Here,

1330041-13
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

the Jacobian matrix J of Eqs. (1) at a fixed point and ∆ = (t + 1 − ε)2 − 4η, then Eq. (15) becomes
Φ(x, y) on Π for 1 ≤ i ≤ N can be written 
1 − t + ε ± (t + 1 − ε)2 − 4η
J = IN ⊗ F (x) + εG ⊗ H, (12) λN ± = . (16)
2
where IN is a N × N identity matrix, We will take into account the conditions such
  that |λN ± | < 1 according to the sign of ∆,
f  (x) 1 1 0 respectively.
F (x) = , H= ,
−η 1 0 0

and G is the adjacency matrix of the graphs of the 6.1. ∆ > 0



mean field connection given by When ∆ > 0, that is, t > 2 η − 1 + ε or
  √
t < −2 η − 1 + ε, then there exist two different
1 1 ··· 1
  eigenvalues of Eqs. (2) as λN ± = [1 − t +
real 
1 

1 1 · · · 1
. ε ± (t + 1 − ε)2 − 4η]/2. We have the following
G= (13)
N  ··· ··· 
 conclusion.
1 1 ··· 1 Proposition 7
It is noted that the matrix F (x) is the same for √
(i) If 2 η − 1 + ε < t < 1 + 0.5η + ε, then |λN ± | <
all neurons because they are identical on Π, thus 1 and the fixed point Φ is a stable sink;
the diagonalization of the Jacobian J can be done √
(ii) If t < −2 η − 1 + ε, then |λN ± | > 1 and the
by simply diagonalizing the connectivity matrix G. fixed point
We obtain a block diagonal matrix, in which each √ Φ is a unstable source;
(iii) If t > 1 + 2η + ε, then |λN + | < 1 and
block 2 × 2 is given as follows: λN − < −1, namely, the fixed point Φ is a
saddle.
Mk = F (x) + εsk H, (k = 1, . . . , N ), (14)

where sk = 0 (k = 1, . . . , N − 1), sN = 1 are the From Proposition 7, four bifurcation curves can
eigenvalues of G. Then, the eigenvalues of the com- be obtained by the following mathematical analyt-
plete system of Eqs. (1) are now obtained from each ical expressions.
block at xi = σ in Eqs. (1) as follows: √ (1 + σ 2 )2
 α1,2 = (±2 η − 1 + ε) ,

f  (σ) + 1 + εsk ± [f  (σ) − 1 + εsk ]2 − 4η
λk± = . (1 + σ 2 )2
2 α3 = (1 + 0.5η + ε) , (17)
(15) 2σ
 (1 + σ 2 )2
The stability of the fixed point on Π requires α4 = ( 1 + 2η + ε) .

that |λk± | < 1. Thus, as a function of the maximum
absolute value of the eigenvalues of Mk , neuron It is noted that αi (i = 1, . . . , 4) are odd func-
parameters α, σ, η, the coupling strength ε, and the tions about σ in the significant domain of σ, and
eigenvalues sk of the coupling matrix G, the mas- their geometrical presentations of αi are symmetri-
ter stability function [Pecora & Carroll, 1998] deter- cal about the origin O(0, 0) in the parametric plane
mine the stability of the neuron network around the (σ, α). Therefore, in the following discussion, we
fixed point. only consider the left-half plane of the parameter
Due to the fact that s1 = s2 = · · · = sN −1 = 0, plane (σ, α).
and sN = 1, therefore, the stability of the fixed The parameter space (σ, α) is divided into three
point on Π is mainly determined by |λN ± | < 1. domains by α1 , α2 , σ-axis, and α-axis as shown in
Therefore, in the following part of this section, we Fig. 10. The three domains in the parameter plane
will focus on the dynamics of the fixed points of (σ, α) are entitled “Sink”, “Source”, and “Saddle”
Eqs. (2). as shown in Figs. 10(a) (ε = 0) and 10(b) (ε = 0.4),
For the sake of simplicity, in the following part respectively, which visualize stable sinks, unstable
of this section, we suppose that t = 2ασ/(1 + σ 2 )2 , sources, and unstable saddles when the parameter

1330041-14
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

5 5

4 Source 4
Source
3 α 3
1
2 α2 2

1 α3 1 α1
α4 Sink α2
0 0
α

α
α3
−1 −1 Sink
α4
−2 −2

−3 Saddle −3

−4 −4 Saddle

−5 −5
−2 −1.5 −1 −0.5 0 −2 −1.5 −1 −0.5 0
σ σ

(a) (b)
Fig. 10. When ∆ > 0, fixed points and their types in the parameter plane (σ, α): (a) ε = 0 and (b) ε = 0.4.


values of (σ, α) are located at the three different Eqs. (2) as λ1,2 = [1 − t + ε ± i 4η − (t + 1 − ε)2 ]/
subregions, respectively. It is noted that ∆ < 0 2, then we have the following conclusion.
when the pair of parameter values (σ, α) is located
on the region between α1 and α2 , in which some Proposition 9
very rich nonlinear dynamical phenomena such as

the saddle-node bifurcation, the Neimark–Sacker (i) If η − 1 + ε < t < 2 η − 1 + ε, then the fixed
bifurcation may occur. Apart from this region where point Φ is a stable focus;

∆ < 0, most parameter values in the (σ, α) plane (ii) If −2 η − 1 + ε < t < η − 1 + ε, then the fixed
correspond to ∆ ≥ 0. In particular, when the pair point Φ is an unstable focus.
of parameter values (σ, α) is chosen on the bound-
aries of α1,2 , then ∆ = 0. In addition, it is hard From Proposition 9, the parameter region
to distinguish α4 from α3 due to α3 ≈ α4 when where ∆ < 0 is divided into two different subre-
0 < η  1. gions by α9 , where α5 is in fact the Neimark–Sacker
bifurcation line defined by
6.2. ∆ = 0
(1 + σ 2 )2
When ∆ = 0, i.e. 4η = (t + 1 − ε)2 ,
then the pair α5 = (η − 1 + ε) . (18)

of parameter values (σ, α) is just located on α1,2 ,
on which there exist two identical real eigenvalues As shown in Fig. 11(a), when the pair of para-
of Eqs. (2) as λN ± = (1 − t + ε)/2. We obtain the metric values of (σ, α) is chosen in the light gray
following conclusion: blue region enclosed by α1 and α5 , then Φ is a sta-
ble focus, while in the dark gray region enclosed by
Proposition 8 α5 and α1 , Φ is a unstable focus. Their correspond-
(i) Φ is a stable sink if the pair of parameter values ing time evolutions and phase portraits are given in
(σ, α) is located on α1 ; Figs. 11(b)–11(e).
(ii) Φ is a unstable source when the pair of param-
eter values (σ, α) is just located on α2 .
6.4. Neimark–Sacker bifurcation
When ∆ < 0, namely, 4η > (1 + t − ε)2 , then there
6.3. ∆ < 0 exists a pair of conjugate complex eigenvalues λN ±
Under this condition, i.e. 4η > (1 + t − ε)2 , there of Eqs. (2). The Neimark–Sacker (NS) bifurcation
exists a pair of conjugate complex eigenvalues of will occur [Wiggins, 1990; Kuznetsov, 1999] if the

1330041-15
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

From |λN ± |2 = 1, we have η = t + 1 − ε, there-


fore the points satisfying the first condition of the
NS bifurcation are located on the dashed line pre-
sented by α9 as shown in Fig. 11(a). When the pair
of parameter values of (σ, α) is chosen near the NS
bifurcation line, their types and stabilities of fixed
points are as presented by Proposition 9. While,
when the pair of parameter values of (σ, α) is chosen
on the NS bifurcation line α9 , the fixed point can be
transformed into a limit cycle with the quasiperi-
odic motion. The stability of this limit cycle will
be determined by a coefficient in the corresponding
normal form. In the remaining part of this section,
we will discuss the stability and how the fixed points
are transformed into quasiperiodic states or limit
cycles in the parameter plane. To attain this goal,
the coefficient of the normal form will be derived
(a) for the NS bifurcation in the parameter space.
The Jacobian matrix J  of Eqs. (2) can be
rewritten as
1
1

1 − η 1
0 0 J = , (19)
−η
xn

1
xn

−1 −1
and the pair of conjugate complex roots can be
−2
−2
0 5000 10000 −1.4 −1.3 −1.2 rewritten as λN ± = (2 − η ± i 4η − η 2 )/2.
yn
n In a very straightforward way, it is easy to prove
(b) (c) that λnN ± = 1 (n = 1, 2, 3, 4). In the following, we
will determine the coefficient of the normal form
1 derived on the center manifold.
1
At first, the nonzero fixed point Φ(σ, σ − f (σ)−
0 0 εσ) on the synchronized manifold Π can be trans-
formed into the origin O(0, 0) by the following
xn
xn

−1 −1 transformation

−2 −2
−4 −2 0 un+1 = xn+1 − σ,
0 5000 10000 (20)
yn
n
vn+1 = yn+1 − σ + f (σ) + εσ.
(d) (e)
Fig. 11. (a) When ∆ < 0, the parameter domain of (σ, α) in Then Eqs. (2) becomes
the light gray region corresponds to stable foci, while in the 
dark gray region corresponds to unstable foci. Time evolu- 
 α α
tions and phase portraits of Eqs. (2) when ε = 0.4, σ = −1,

un+1 = εun + vn + 2
− ,
1 + (un + σ) 1 + σ2
η = 0.001: (b)–(c) in a limit cycle state with α = 1.3 and
(d)–(e) in a silent state with α = 1.15. 

vn+1 = vn − ηun .

following three conditions are satisfied (21)


(i) |λN ± | = 1; After Taylor’s expansion on the first equation
(ii) λiN ± = 1 for i = 1, . . . , 4; of the right-hand side of Eqs. (21), the equivalent
(iii) |λN ± | = 0 at a specific bifurcation point. map format of Eqs. (21) is as

1330041-16
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

  
 2ασ α(1 − 3σ 2 ) 2 4ασ(1 − σ 2 ) 3
u  ε− u+v− u + u + O(4)
→

(1 + σ 2 )2 (1 + σ 2 )3 (1 + σ 2 )4 .
 (22)
v
−ηu + v
Secondly, there exists a matrix M , such that
the Jacobian matrix J  can be transformed into the Thirdly, Eqs. (26) can be reduced into the following
following form: one-dimensional complex map after letting:
     
2 z 1 i x
 1− η −
4η − η  = , (28)
 2 2  z 1 −i y
 
M −1 J  M =   , (23)
  then we have
 4η − η 2 η 
1−
2 2 zn+1 = (Re λ + i Im λ)zn + m + ig. (29)
where Fourthly, expanding m + ig in a Taylor expansion
  with zn and zn , we obtain the following normal form
1 0
   zn+1 = λ(η)zn + ξ20 z 2n + ξ11 zn zn
M =
η 4η −

η 2 ,

2 2 + ξ02 z 2n + ξ21 z 2n zn + · · · , (30)
  (24)
where the bar of zn denotes the conjugate complex
1 0
  variable of zn , and
M −1 = 
 η 2 .

− ξ20 =
1 
[m − myy + 2g xy
4η − η 2 4η − η 2 8 xx
Letting + i(g xx − g yy − 2mxy )](0,0) ,
 
u x 1 
→M , (25) ξ11 = [m + myy + i(g xx + g yy )](0,0) ,
v y 4 xx
1  (31)
and substituting M and M −1 into Map. (22), then ξ02 = [m − myy − 2g xy
8 xx
we have
   + i(g xx − g yy + 2mxy )](0,0) ,
2
  1− η −
4η − η 

xn+1   xn ξ21 =
1
[m + m  
xyy + g xxy + g yyy
 2 2 
→   16 xxx
yn+1   yn
 4η − η 2 η  + i(g    
1− xxx + g xyy − mxxy − myyy )](0,0) ,
2 2
 where the superscript and the subscript denote the
m second or the third derivative of m or g about x,
+ , (26) and y, respectively.
g
We have
where −2α + 6ασ
  mxx (0, 0) = = a1 ,
α 2ασ (1 + σ 2 )3
m=− + − ε xn
1 + σ2 (1 + σ 2 )2 η
g xx (0, 0) =  a1 = b1 ,
α 4η − η 2
+ , (27)
1 + (xn + σ)2
mxy (0, 0) = gxy (0, 0) = myy (0, 0)
η
g= f.
4η − η 2 = gyy (0, 0) = 0,

1330041-17
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

24ασ(1 − σ 2 ) We have the following conclusion.


m
xxx (0, 0) = = a2 ,
(1 + σ 2 )4
Proposition 10
η
g 
xxx (0, 0) = a2 = b2 , (i) If a(0) < 0, then the fixed point O(0, 0) (i.e. Φ)
4η − η 2
is stable;
m  
xxy (0, 0) = g xxy (0, 0) = mxyy (0, 0) (ii) If a(0) > 0, then the fixed point (0, 0) (i.e. Φ)
is unstable.
= g 
xyy (0, 0) = myyy (0, 0)
Especially, if r 2 = − a1 and a(0) < 0, there exists
= g
yyy (0, 0) = 0. a stable quasiperiodic solution or limit cycle, and
(32) this is the so-called supercritical NS bifurcation.
Otherwise, the quasiperiodic solution or limit cycle
Then is unstable and the bifurcation is subcritical.
1 1
ξ20 = (a1 + ib1 ), ξ11 = (a1 + ib1 ), A contour plot of the coefficient a(0) as a func-
8 4
(33) tion of two parameters σ and α is useful to visualize
1 1
ξ02 = (a1 + ib1 ), ξ21 = (a2 + ib2 ). its sign of a(0) in the parametrical plane shown in
8 16 Fig. 12 for ε = 0.4, in which there exist some con-
Equations (30) becomes tours with negative values of a(0) in the range of
(σ, α) ∈ [−5, 5] × [−5, 5]. These negative contours
z → λz + cz 2 z + O(4), (34)
demonstrate that there exist stable quasiperiodic
where solutions or limit cycles. Figures 13(a) and 13(b)
further show the isoperiodic diagrams when ε = 0
λ(1 − 2λ) |ξ11 |2 2|ξ02 |2
c= ξ20 ξ11 + + + ξ21 . and ε = 0.4, respectively. For the sake of compari-
λ−1 1 − λ λ2 − λ son, two isoperiodic diagrams are in the same range
(35) of (σ, α) ∈ [−1.5, −0.1] × [−3, 3], in which the dis-
cretized parameter interval of σ and α is a mesh of
Finally, Eq. (34) can be transformed into the
800 × 1000 points. From Figs. 13(a) and 13(b) for
normal form in polar coordinates by z = reiθ and
ε = 0 and ε = 0.4, there exist three kinds of dif-
λ = eiθ0 , where r and θ are the radial and angular
ferent colors corresponding to periodic solutions (in
parts of complex variable z, respectively,
 and θ0 =
2
red), quasiperiodic solutions (in yellow), and chaotic
arctan[Im(λ)/Re(λ)] = arctan[ 4η − η /(2 − η)],
motions (in blue), respectively.
then the normal form is as
r → r[1 + a(0)r 2 ] + O(r 4 ), (36)
5
where
 
−iθ0 (1 − 2eiθ0 )e−2iθ0
a(0) = Re[ξ21 e ] − Re ξ20 ξ11
1 − eiθ0
1
− |ξ11 |2 − |ξ02 |2
2 0
α


3(a21+ b21 ) 1 θ0 θ0
=− + csc 2b2 cos
64 64 2 2
θ0 3θ0
− 2a2 sin − 2(2a1 b1 + b2 ) cos
2 2
−5
3θ0 5θ0 −5 0 5
+ 2(a21 b21
− + a2 ) sin + 2a1 b1 cos σ
2 2
 Fig. 12. A contour plot of the stable coefficient a(0) as a
5θ0
− (a21 − b21 ) sin . (37) function of σ and α in the range of (σ, α) ∈ [−5, 5] × [−5, 5]
2 for ε = 0.4.

1330041-18
December 23, 2013 18:44 WSPC/S0218-1274 1330041

Rulkov Neuron Network with Mean Field Coupling

(a) (b)
Fig. 13. Isoperiodic diagrams in the range of (σ, α) ∈ [−1.5, −0.1] × [−3, 3]. (a) ε = 0 and (b) ε = 0.4.

7. Conclusion stability coefficient can determine the stable region


in a two-parameter space, which is very important
In this paper, we have thoroughly investigated the
to know which domain in the parameter space cor-
bifurcation for N identical Rulkov neurons coupled
responds to what kinds of transition from a fixed
by the mean field. We obtain the analytical expres-
point to quasiperiodic solution or vice versa.
sions of the flip bifurcation and the crisis bifurca-
Our results show that there exist very rich
tion of the coupled neurons, by which we are able
bifurcation phenomena and different firing regimes
to examine the effect of the coupling on different
when the parameter values (σ, α) and the coupling
types of bursting of the coupled neuron in a more
strength ε are chosen in different domains in the
rigorous way rather than only by numerical simu-
parameter space. Due to these bifurcations and sta-
lations. We have found that there are at least two
ble domains of various solutions, on the one hand,
types of triangle burstings in the coupled Rulkov
it is clear to distinguish the kinds of bursting of
map-based neuron model. The first type of trian-
the Rulkov neuron model. On the other hand, it is
gle bursting can be determined by two saddle-node
easy to determine the model parameter regions in
bifurcations, a subflip bifurcation, and a supflip
order to mimic different desirable firing regimes of
bifurcation, and the second one can be determined
the Rulkov chaotic coupled neuron network.
by a crisis bifurcation, a supflip bifurcation and
We hope these results can lay a solid theo-
a saddle-node bifurcation. There are at least two
retical foundation to build up large-scale neuron
kinds of striking transition mechanisms: after differ-
networks, reflect real aspects of functional roles
ent couplings, one is from the square-wave bursting
in neural assemblies, and make a significant con-
to the triangle bursting, and the second one is from
tribution to establish large-scale neuron networks
the spiking to the square-wave bursting.
in dealing with the neurological diseases such as
By using the master stability function, the
epilepsy, Parkinson’s disease, Alzheimer’s disease,
dynamics of full neuron network can be restricted
schizophrenia, and autism in the future.
on the synchronized manifold, then the dynamics of
the full neuron network can be determined by a two-
dimensional map-based model. We have taken into
Acknowledgment
account the stability conditions of fixed points on
the synchronized manifold. In particular, in order to This work is supported by Natural Science Founda-
determine the stability of limit cycles or quasiperi- tion of China (NSFC) under Project No. 11171017.
odic solutions, as a function of the intrinsic control
parameters of each single neuron and the external
coupling strength, the normal form restricted on References
the center manifold is obtained. We have obtained Cao, H. J. & Sanjuán, M. A. F. [2009] “A mechanism
the stability coefficient of the normal form. This for elliptic bursting and synchronization of bursts in

1330041-19
December 23, 2013 18:44 WSPC/S0218-1274 1330041

H. Cao & Y. Wu

a map-based neuron network,” Cogn. Process. 10, Rinzel, J. [1985] “Bursting oscillation in an excitable
23–31. membrane model,” Ordinary and Partial Differen-
Cao, H. J. & Ibarz, B. [2010] “Hybrid discrete-time tial Equations, eds. Sleeman, B. D. & Jarvis, R. J.
neural networks,” Phil. Trans. Roy. Soc. A 368, 5071– (Springer-Verlag, Berlin), pp. 304–316.
5086. Rinzel, J. [1987] “A formal classification of bursting
de Vries, G. [2001] “Bursting as an emergent phe- mechanisms in excitable systems,” Mathematical Top-
nomenon in coupled chaotic maps,” Phys. Rev. E ics in Population Biology, Morphogenesis and Neuro-
64, 051914. sciences, Lecture Notes in Biomathematics, Vol. 71,
Hodgkin, A. L. & Huxley, A. F. [1952] “A qualitative eds. Teramoto, E. & Yamaguti, M. (Springer-Verlag,
description of membrane current and application to Berlin).
conduction and excitation in nerve,” J. Physiol. 117, Rulkov, N. F. [2001] “Regularization of synchronized
500–544. chaotic bursts,” Phys. Rev. Lett. 86, 183–186.
Ibarz, B., Casado, J. M. & Sanjuán, M. A. F. [2007a] Rulkov, N. F., Timofeev, I. & Bazhenov, M. [2004]
“Patterns in inhibitory networks of simple map neu- “Oscillations in large-scale cortical networks: Map-
rons,” Phys. Rev. E 75, 041911. based model,” J. Comput. Neurosci. 17, 203–223.
Ibarz, B., Tanaka, G., Sanjuán, M. A. F. & Aihara, Tanaka, G., Ibarz, B., Sanjuán, M. F. A. & Aihara, K.
K. [2007b] “Sensitivity versus resonance in two- [2006] “Synchronization and propagation of bursts in
dimensional spiking-bursting neuron models,” Phys. networks of coupled map neurons,” Chaos 16, 013113.
Rev. E 75, 041902. Wang, H. X., Lu, Q. S. & Wang, Q. Y. [2008] “Bursting
Ibarz, B., Casado, J. M. & Sanjuán, M. A. F. [2011] and synchronization transition in the coupled mod-
“Map-based models in neuronal dynamics,” Phys. ified ML neurons,” Commun. Nonlin. Sci. Numer.
Rep. 501, 1–74. Simulat. 13, 1668–1675.
Izhikevich, E. M. & Hoppensteadt, F. [2004] “Classifi- Wiggins, S. [1990] Introduction to Applied Nonlinear
cation of bursting mappings,” Int. J. Bifurcation and Dynamical Systems and Chaos (Springer-Verlag, NY).
Chaos 14, 3847–3854. Yang, Z. Q. & Lu, Q. S. [2005] “The integer multiple
Izhikevich, E. M. & Edelman, G. M. [2008] “Large- ‘fold/homoclinic’ bursting induced by noise in the
scale model of mammalian thalamocortical systems,” Chay neuronal model,” Int. J. Nonlin. Sci. Numer.
PNAS 105, 3593–3598. Simulat. 6, 1–6.
Kuznetsov, Y. A. [1999] Elements of Applied Bifurcation
Theory, 2nd edition (Springer-Verlag, NY).
Pecora, L. M. & Carroll, T. L. [1998] “Master stabil-
ity functions for synchronized coupled systems,” Phys.
Rev. Lett. 80, 2109–2112.

1330041-20

You might also like