Progress of Theoretical Physics Supplement No.

79, 1984 223

Cooperative Dynamics of Oscillator Community

- - A Study Based on Lattice of Rings - -

Y oshiki KURAMOTO

Research Institute for Fundamental Physics, Kyoto University, Kyoto 606

(Received November 8, 1984)

A phase description of systems of many limit cycle oscillators is established. As a result, the

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

systems are reduced to populations of simple elements called rings, their mutual coupling depending
only on the phase difference between the interacting pair. A few typical lattice models of rings are
proposed. Besides pattern formation and turbulencelike behavior, our ring models, as statistical
cooperative systems, can exhibit curious phase transitions with the order parameter essentially of
dynamic origin. Finally, such fields of rings are generalized in order to simulate excitable media,
and also to incorporate equilibrium cooperative fields of some known universality classes.

§ 1. Introduction
A simplest picture of a limit cycle oscillator may be provided by a ring with a
circulating representative point bound to it. This kind of one-dimensional representation
or phase description of limit cycle motion in fact turns out extremely usefull when we wish
to study large-scale dynamics, deterministic or statistical, of a naany-oscillator
community. Arthur Winfree was the first to propose and formulate such a view/>
although for a rather limited purpose of understanding physiological clocks in terms of
individual cellular oscillators. More recently, the present author tried to develop the
same idea in some directions. 2 >- 4 > The present paper provides a general survey of the
latter works, and includes also some new analytical and numerical results.
To be concrete, one may imagine, e.g., a regular lattice, each point of which being
occupied by such a ring. Analogously to a lattice of spins, the rings may be assumed to
couple to their neighbors. This seems to represent quite a new type of cooperative
system, particularly because of the absence of such things as free energy functional or
hamiltonian to govern the dynamics and/ or statistics. One may also say that the element
local systems are dynamically active (i.e., functioning under off-equilibriumt conditions) in
the present system, while they are passive in ordinary cooperative fields.
In general, when similar active elements come into contact with each other to form an
extended field, they are expected to produce a wealth of wave patterns and turbulencelike
phenomena, and this has been demonstrated typically in reaction-diffusion systems. In
the past, however, our main concern has been restricted to deterministic aspects of
homogeneous active fields. There seems to be another important class of problems,
namely, those associated with statistical dynamics of random or stochastic active fields.
The ring dynamics developed below seems to meet the purpose of extending our scope to
include statistical dynamics, insofar as the active components are of oscillatory nature.
We will show in this paper that our ring models not only recover most of the essential
features of oscillatory reaction-diffusion systems, but also exhibit, under the presence of
224 Y. Kuramoto

randomness, a new type of ordered phases. Moreover, the limitation of the phase
dynamics, which is often spoken of in connection with the appearance of topological
defects or phase singularities, actually does not exist in our case, because what we work
with are lattices and not a continuous space.
In § 2 a phase description of one- and two-oscillator systems will be formulated, where
two types of randomness will be incorporated in the theory. Some discussion will also be
given on the nature of ring-ring interaction. In § 3 our two-ring model will be generalized
to the case of N rings. From such an N -ring model a few specific subclass models are
generated according to the type of randomness. Some deterministic and statistical
aspects of N -ring models, i.e., waves, turbulence and phase-transitionlike phenomena, will
be· discussed. Because our purpose there is limited to demonstrating the richness in

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

behavior of our ring models, we will not go into full details of each topic. In § 4 it will
be argued that the ring models can be modified and extended in various ways. One
important modification is the inclusion of excitability. It will also be shown that a
suitably modified ring model with stochasticity includes as special cases equilibrium
cooperative fields such as Ising and XY spin systems.

§ 2. One- and two-oscillator ring dynamics

Let us consider a set of s coupled ordinary differential equations in vector form

dX =F(X) (2·1)
dt
which is assumed to admit a stable limit cycle solution Xo(wt) with period T 2Jr/w.
This solution may trivially be rewritten as Xo( ¢) where

(2·2)

The ¢ specifies the position of the representative point on the limit cycle orbit .£, and is
called phase; ¢ and ¢+2Jr represent an identical point on .£. We now ask how the
equation for ¢ deviates from (2 · 2) when the oscillator is perturbed either by external
forces or by other oscillators coupled to it. Since such perturbations, however weak it
may be, will no longer permit the representative point to stay strictly on .£, such a
question makes sense only if the definition of ¢ is extended to some region outside of .£.
We now define a scalar field ¢ (X) for some region containing .£ inside. It is most
convenient to choose ¢(X) in such a way that, as far as the perturbation is absent,
Eq. ( 2 · 2) may hold identically even if the representative point is outside of .£. This
definition immediately leads to the identity

~~ =gradx¢·F(X)=w. (2·3)

The phase ¢ (X) defined in this way is called asymptotic phase, s> and the manifolds of
constant ¢ are called isochrons. l),G>, 7 > Let the perturbed motion be described by
dX (2·4)
dt= F(X)+ sp(X, t ).
 Cooperative Dynamics of Oscillator Community 225

Here E is used only to indicate the smallness of p and, whenever no confusion is expected,
we suppress it by equating to 1. The perturbed equation for ¢ now becomes

~~ =gradx¢· {F(X)+Ep(X,t )} .

= w+Egradx¢ · p(X,t ). (2·5)

This may be approximated as

~~ =w+EZ(r/>)·p(Xo(r/>),t), (2·6)

where

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

Z (¢)=(grad xrf> h=xo<S~~>· (2·7)

On going from (2·5) to (2·6), X has been

replaced by Xo( ¢ ), or the representative
point of phase ¢, which lies close to _[, has
been replaced by the point on _[ of the same
phase (see Fig. 1 ). This approximation is
allowed to the lowest order of E, because the
last term in ( 2· 5) is already of the order of E.
Note that the equation of motion (2·6) no
longer contains degrees of freedom other
than ¢. Note also that its right-hand side is
2Jr-periodic in ¢.
We now proceed to two-oscillator phase
dynamics, and start with evolution equations
describing a pair of oscillators coupled Fig. 1. The representative point A is approximated
symmetrically. We also include some by point B lying on the closed orbit, where A and
external forces which may be state- B belong to the same surface of constant phase.
dependent. Both the interactions and This approximation only produces an error of
O(s 2 ) in the evolution equation for phase, c being
external forces are assumed to be weak.
the strength of perturbation.
Then,

df/ =F1(X1)+E{ V(X1,Xz)+ a1(X1,t )},

df/ =Fz(Xz)+E{ V(Xz,Xl)+ az(Xz,t )}. (2·8)

The factor E which commonly multiplies V and aa does not necessarily imply that the
typical amplitudes of V and aa are the same order; it says nothing more than that V and
aa are small quantities. This remark is relevant especially when aa represents stochastic
forces fluctuating very rapidly about the zero value. Then its typical amplitude must be
much larger than that of V in order that the net effects of V and of aa may be comparable.
In any case, we will take account of the nontrivial lowest order contribution from each
type of perturbation. The two oscillators are now supposed to be different, but only
slightly, or
226 Y. Kuramoto

Fa(Xa )= F(Xa )+coFa(Xa ), a=1,2. (2·9)

Equation (2·8) then takes the form similar to (2·4), or

dJ/ =F(Xa)+cPa(Xa,t), (2·10)

where
Pa(Xa,t )= V(Xa,XP)+ aa(Xa,t )+oFa(Xa). (2·11)

The explicit t-dependence of Pa should be interpreted as including both the implicit

t -dependence of V via Xp and the explicit t -dependence of a a. It is straightforward to

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

derive coupled equations obeyed by the phases of the two oscillators, denoted as ¢hand ¢z.
Approximating Xa( t) by Xo( if> a( t )) as we did before, we get

(2·12)

where V( rl>a,¢P) is the abbreviation of V(Xo( rl>a ),Xo( rf>p) ), which is 2Jr-periodic in each of
rl>a and rf>p, and
ga( rl>a, t )=Z( rl>a )· {aa(Xo( rl>a ), t )+ oFa(Xo( rl>a) )} (2·13)

which is 2Jr-periodic in r/>a. Equation (2·12) may further be simplified owing to the
coexistence of two time scales which are vastly different, the one corresponding to the
unperturbed frequency w and the other to the change in the instantaneous frequency due
to perturbations. However, specific reduction procedure depends on the nature of
perturbation, so that we treat below two typical cases separately.
Case I aa=O
Introduce phase disturbances <Pa by
rl>a= wt + <Pa, (2·14)
and express (2·12) as

(2·15)

Because of the common factor con the right-hand side of this equation, the time evolution
of <Pa is slow. Thus, to the lowest order in c, the right-hand side may be approximated
by its time-average over a period of oscillation, thereby the constancy of <Pa being
pretended. This leads to

(2·16)

where

(2·17a)

(2·17b)
 Cooperative Dynamics of Oscillator Community 227

In terms of the original phase variables, we have

(2·18)

where
Wa = W+OW a, (2·19)
and r(¢) is a 2Jr-periodic function of¢.
Case II oFa=O, 9a being Gaussian random forces
Specifically, we assume statistical properties

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

9a(c/Ja,t)=O, }
(2· 20)
9a( cPa, t )gp( cPP, t') = 2Ll (cPa )oapO ( t- t' ).
Let P( cP1,cPz,t) denote the probability that at time t the phases of the oscillators 1 and 2
take on values cP1 and ¢z, respectively. Equation (2·12) is then transformed to a Fokker-
Planck equation

aP = _ (_2h___+ aiz ) }
at a¢1 a¢z '
(2·21)
Ia=[ w+c:{Z(¢a)V(¢a,¢ 11 )+ ~ d~~~a) }]P-c: a;a {LJ(¢a)P}.
This is easily transformed to the equation for Q(</h,¢z,t ), the probability distribution for
phase disturbances </h and ¢z, and we get

_QQ= ~ ( a11 + aJz)

at c a¢1 a¢z '
fa= {Z(wt + <Pa) V(wt + </Ja,Wt + </Jp) (2·22)

+ ~ dL1(~~7¢a)}Q- a~a {Ll(wt+<Pa)Q}.

Again, we have a common factor con the right-hand side of (2·22), so that one may apply
the same reasoning as in Case I and time-average fa to obtain

(2·23)

where r( <Pa- </Jp) is given by (2·17a), and

1 (27!
D= 2JrJo Ll(A+</Ja)dA. (2· 24)

The Fokker-Planck equation in (2·22) with fa given by (2·23) is equivalent to the

Langevin equations

(2·25)

or
228 Y. Kuramoto

1/ =w+T(c/Ja-c/Jp)+ga(t), (2·26)

where
ga(t)=O,
ga(t )gp( t' )= 2Doapo( t- t' ). } (2·27)

We have obtained two kinds of ring models for a pair of coupled oscillators, each
given by (2·18) and (2·26). In the first case, no stochastic forces are present but the
oscillators have different natural frequencies. In the second case, the oscillators are
identical, but only statistically. Of course, there may be a number of possible variants

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

other than these two extreme cases. We now look into how the oscillators behave in Case
I and Case II, respectively.
Case I
Let ¢ denote the phase difference, i.e., ¢ = ¢1- ¢z, by which (2·18) is reexpressed as

d¢-
dt-o+B(¢), (2·28)

where

B(¢)=T(¢)-T(-¢). (2·29b)
Obviously, B( ¢) is an odd and 2Jr-periodic function of ¢. If we restrict consideration to
the domain 0 < ¢ < 2Jr, then
B(O)=B(Jr)=O. (2·30)

The oscillators are phase-locked or mutually entrained provided (2·28) has an equilibrium
solution. When the oscillators are identical, i.e., o= 0, then the relative phase ¢ reaches
equilibrium at 0 and 7r (and possibly other values, too), one of which being stable and the
other unstable. If dT/dc/Jio<O, then ¢=0 gives a stable equilibrium. In this case, the
coupling is called attractive; this is similar to ferromagnetic coupling between spins. On
the contrary, if dT/ dc/Jio >0, then ¢ = Jr gives a stable equilibrium, the coupling then being
called repulsive; this is similar to antiferromagnetic coupling. Next, we assume that the
frequency difference is present. As lo I becomes larger, the stable-unstable pair of
equilibrium points come closer to each other, and at some critical value, denoted by oc, the
equilibria vanish simultaneously, beyond which the oscillators will oscillate with
independent frequencies. Near threshold, the frequency difference behaves like Io- ocl 112 .
Suppose that lol is small enough to admit a stable locking state at ¢ = ¢. Then the
oscillators have a common frequency OJ, where
OJ= (J)l + r( ¢)=wz+ r(- ¢). (2·31)

Empirically, OJ is usually higher than the simple average of natural frequencies, or

r( ¢)+ rc- ¢)>0, (2·32)

 Cooperative Dynamics of Oscillator Community 229

where (J)a are assumed positive. The property in (2 · 32) implies that the interaction is not
of variational type. In other words, there exists no common potential function H ( </>1- </>2)
such that the interaction parts of the evolution equations are generated by its gradients in
the form

} (2·33)

In fact, (2·33) would imply T(¢)+T(-¢)=0, or T(¢) being an odd function of</>, in
contradiction to (2·32). Besides (2·32), let us assume below for simplicity T(O)=O, which

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

is actually the case for diffusion-type coupling, V(X1,X2)cxX1- X2.
As a model for r (¢ ), it would be interesting to take
T(¢ )=- K {sin(¢+a )-sina },

lal< ~· (2·34)

Then the interaction is variational if a=O, and non-variational if a=I=O. Vve also see that
the interaction is attractive if K>O, and repulsive if K<O. The property in (2·32) is
satisfied if a> 0 for attractive interaction, and a< 0 for repulsive interaction.
Case II
Random forces permit no perfect entrainment. In terms of the phase difference ¢, the
stochastic equation (2 • 26) may be expressed as
d¢-
dt-B(¢ )+g(t ), (2·35)

where
(2·36)
Clearly,

g(t)=O, }
(2·37)
· g(t)g(t')=4LJ·B(t-t').

The probability distribution p(¢,t) obeys a simple Fokker-Planck equation

ap _ a a2 p
Tt- -Jf{B( ¢ )p} + 2LJ a¢2 (2·38)

whose stationary solution Po(¢) is given by

Po(¢ )cxexp{- U(¢ )},

U(¢ )=-
2~ 1()B(¢')d¢'. (2·39)

The potential U ( ¢ ) is even and 27l'- periodic in ¢, and has extrema at least at 0 and 7l', one
of which being maximum and the other minimum.
230 Y. Kuramoto

§ 3. N -oscillator ring dynamics

It is obvious that the two-oscillator ring dynamics derived in § 2 can immediately be

extended to the case of N oscillators. As a trivial generalization of each of (2·18) and
(2·26), we get

(Model I) 1/ = aJi + ljru( c/>i- c/>j ), (3·1)

(Model II) 1/ = w+ ljru( cPi- c/>j )+ gJ t ), (3·2)

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

where ru(¢) is 2Jr-periodic in¢, and the Gaussian noise gi(t) has properties

~=0,

g i ( t )g j ( t = 2D 0 ijO ( t - t
I ) I ) • (3·3)
As a special case of Model I or l\tiodel II, we have a system without randomness, or

(Model 0) (3·4)

In what follows, we survey various cooperative phenomena expected from each of the
three models derived above.

3.1. Deterministic dynamics based on Model 0

If the interactions ru are chosen suitably, Model 0 turns out to share some important
features in common with oscillatory reaction-diffusion systems. To show this, we
consider a one-dimensional lattice of rings with lattice spacing a, and assume that the
rings are coupled only to their nearest neighbors, or

(3·5)

As before, it is required that r(O)=O. We now go over to a continuous space by the

following replacements :
cPi~cf>(x),
2
cPi±l~cf>(x+a)=c/>(x)+acf>l(x)+ ~ c/>"(x)+···, (3·6)
2 3
r(cf>i-cPi±l)~r(+ac/>l(x)- ~ ¢"(x)+ ~ cf>"l(x)-···).

Equation (3·5) is then transformed formally into a partial differential equation. We will
now concentrate on long-scale wave phenomena, and treat the space derivative a; ax
formally as a small expansion parameter. To be more explicit, one may associate an
indicator E of smallness to each space derivative and replace, e.g., a2 I ax 2 by E a I ax and
2 2 2

a2 I ax 2 ( a; ax )2 by E4 a2 I ax 2 ( a; ax )2 and so on. up to the order E ' we obtain a nonlinear

2

diffusion equation
 Cooperative Dynamics of Oscillator Community 231

(3·7)

where
]) = - a2 r' (0 ), (3 • 8a)

f-L = a 2 r" (0). (3. 8b)

The same equation as (3 · 7) arises also from reaction-diffusion equations. 4>,s>

Previously, 8 > this type of equation (with extension to higher dimensions) was applied to
the discussion of target pattern formation such as observed in the Belousov-Zhabotinsky
reaction. It should be noted that the attractive coupling leads to positive v, i.e., normal

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

diffusion, whereas the repulsive coupling leads to negative diffusion implying the
breakdown of the underlying long-scale assumption. It is also interesting to observe that
the variational type of coupling leads to vanishing J-L, while the corresponding term J-L( a¢
I ax )2 gives rise to a frequency change due to phase gradient. It was argued previously 8 >
that the same term is crucial to the formation of shocklike structures on collision of two
phase waves and hence to the steady expansion of target patterns.
In oscillatory reaction-diffusion systems, the state of spatially uniform oscillations is
not always stable. As a result of weak instability of uniform oscillations with respect to
long wavelength phase modes, we have phase turbulence. 9 > A suitable equation
describing such turbulence turned out to be

(3. 9)

where v is a small negative. This equation was first derived in Ref. 10 ), and its turbulent
behavior was numerically confirmed in Ref. 11 ). Similar equations are known to arise in
connection with combustion, 12 > chemical wave fronts in general, 13 > phase-unstable periodic
structures with propagation, 14 > and possibly many other problems. We now show that our
chain of rings can also generate the same equation provided the ring-ring couplings are of
some competing type. Let us assume attractive type nearest-neighbor coupling, and
include also next-nearest-neighbor coupling which is of repulsive type. A simplest choice
would be
ri,j( ¢) = -- K1 {sin(¢+ a1)- sina1} if j= i+l,
= K2 {sin(¢+ a2)- sina2} if j=i+2,
= 0 otherwise, (3·10)

where K1,K2>0. For simplicity, we put a2=0. By making a formal space-derivative

expansion of a¢/ at, and retaining leading terms as we did previously, 9 > the above form of
couplings leads to the phase turbulence equation (3· 9) with coefficients given by
(3·1la)

(3·11b)

(3·11c)
232 Y. Kuramoto

For our treatment to be consistent, Ivi has to be small and A positive, which is actually
fulfilled by assuming
(3·12)

For vanishing a1, the interaction is of variational type, the corresponding interaction
potential H being given by
H= ~{- K1cos(¢i- ¢i+l )+ Kzcos(¢i- ¢i+z)}. (3·13)
i

Since the constant term w in (3 · 5) can be

eliminated if we work with <A( ¢i- wt ), the

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

entire dynamics then becomes variational
with the same potential H as given above.
As time t goes to infinity, the system is then
expected to reach a minimum of H. As far
as equilibrium states are concerned, such
systems are the same as one-dimensional
X Y spin systems without thermal
fluctuation, where the nearest-neighbor
coupling is ferromagnetic and next-nearest-
neighbor coupling antiferromagnetic.
Anisotropic Ising model of similar type of
Fig. 2. Discrete-space ring model allows for
competing interactions is known as ANNNI
vortexlike modes. Arrows show unit vectors
model (axial next-nearest-neighbor model) with components (cos¢i, sin¢i ).
which produces complicated equilibrium
configurations of spins. Thus, the present ring lattice producing phase turbulence may be
viewed as a non-equilibrium extension of some class of equilibrium cooperative systems
with frustration.
It was pointed oues) that the phase description in general is bound to break down for
systems with topological defects because of phase singularities. involved. For instance,
the nonlinear phase diffusion equation (3·7) was shown to lead to a serious contradiction
when applied to rotating spiral waves. s) Since such a contradiction arises from the
requirement for perfect spatial continuity of the phase function, it is avoided in our case
because we are working with rings distributed on a discrete space. The system can then
allow for vortexlike singularities (and hence rotating spiral waves) as schematically
shown in Fig. 2. In the case of vanishing non-variational effects, such vortices are the
same thing as we meet in XY spin lattices. In a continuous space, a vortexlike singular
point has vanishing amplitude, whereas in our discrete-space ring model, each ring retains
its full amplitude. In a spatially coarse-grained picture, however, the ring model
produces amplitude degree of freedom. In particular, the four vectors in the central part
of the lattice in Fig. 2 give vanishing sum or vanishing amplitude. In any case, one need
not worry so much about the absence of amplitude degree of freedom in the discrete-space
ring model, as far as the large-scale dynamics is of one's ultimate concern.
Although our lattice model of rings may not precisely represent oscillatory reaction-
diffusion systems, we saw that it still shares a number of important features in common
with the latter. Unfortunately, however, there is one important property which is missing
 Cooperative Dynamics of Oscillator Community 233

in Model 0. This is the relaxational character of oscillations or, as its extreme,

excitability/ 6 > which is quite common is real chemical, biochemical and biological
oscillations. However, this property, too, can be incorporated into our model by a simple
modification, as will be seen in § 4.
3.2. Statistical ring dynamics based on Model J 2 l. 4 >
We saw in § 2 that the system of two rings with frequency difference can be analyzed
quite easily. The system is either in a phase-locked state (i.e., periodic state) or in an
unlocked state (i.e., quasi-periodic state). As the number N of rings is increased,
however, the system will show increasingly complicated motions. For instance, we
expect the appearance of a strange attractor for N 24. As N increases further, e.g., to
10 10 , some simple properties, though of statistical nature, will become n1ore and more

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

apparent. In this subsection, we consider what properties of this sort are expected from
Model I in the limit of large N.
Spatially extended lattices of rings with random frequency distribution seem to serve
as an interesting model showing phase transitions quite unlike those iln equilibrium
systems. The interactions between rings, if they are of the attractive type, will favor
synchronized collective oscillations. On the contrary, random frequency distribution will
tend to disintegrate such organized motions. Suppose each ring on a lattice is coupled to
its neighbors, and that the coupling strength is increased continuously. For sufficiently
weak coupling, the individual rings will oscillate almost independently, and no
macroscopically coherent behavior will be expected. As the coupling becomes stronger,
some local order may develop, and the entire system may look as consisting of patches of
self-synchronized clusters. The size of such clusters will increase with increasing
coupling strength, and at some critical value of the latter, one of the clusters will develop
to a macroscopic scale, beyond which a finite portion of the rings will find themselves in
an identical frequency state. Such a scenario seems quite probable, but it should be
confirmed by computer simulation, which we have not attempted yet, however.
For such a phase transition, there seem to be at least two ways of defining degree of
macroscopic order. The first one is given by a complex order parameter w defined by

(3·14)

where n( ¢, t) denotes the number density of the rings of phase ¢ at time t. N onvanishing
a implies the presence of macroscopic oscillation, and vanishing a its absence. Another
measure of order, denoted as r, may be provided by
r=Ns/N, (3·15)
where Ns is the number of rings falling in an identical frequency state to form a
condensate. It see1ns that nonvanishing r does not always imply nonvanishing a. This
may be inferred simply from the observation of a well-developed target pattern in the
Belousov-Zhabotinsky reaction, where the entire system is entrained to a central
pacemaker, while the monotone gradient of ¢ in the radial direction seems to cancel the
oscillation of the system as a whole.
There is a special subclass of Model I for which analytical expressions for a, r and
some other quantities are available. This is the case where the coupling is variational
234 Y. Kuramoto

and all ring pairs have equal coupling strength.- A simplest choice would be
rij=- N- 1Ksin(¢i- ¢j). (3·16)
Then, Model I takes the form

1/ =uJi-KO"sin(¢i-8). (3·17)

If we assume that as t ~ oo the order parameter behaves like

a= constant,
8=Qt,
} (3·18)

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

then, one may solve (3·17) for each ¢i. Given the distribution of OJi, these solutions for
¢i(t), which of course include unknown parameters a and Q, enable us to calculate n(¢,
t) and hence a and Q via (3·14). In this way, we obtain a self-consistent equation for the
complex order parameter. For some simple distribution of OJi, such a self-consistent
equation can be treated analytically.
Let the distribution g ( w) of Wi be symmetric about some frequency Wo and convex at
w= wo. Then the system given by (3·17) shows the following features. Due to the
assumed symmetry of g and the assumed variational nature of the interaction, the
frequency Q of the organized motion turns out identical to the mean frequency wo. The
critical value of the coupling strength, denoted as Kc, for the onset of macroscopic
oscillation is given by
Kc = 2/ Jrg ( Wo ). (3·19)
We define parameter 11 by
11=(K- Kc)/Kc. (3·20)
Near and above criticality, i.e. 0<11<t1, the order parameters are given by

8g(wo)f.1
a= (3·21a)
K c 2 g " ( Wo ) '

r=2Kcag(wo)+ O(o- 3 ). (3·21b)

The frequency of each oscillator has been changed to iiJi, where

2 (3·22)
=wo+(wi-wo)/1-( Ka ) otherwise. }
Wi- Wo
Thus, for given distribution g, we obtain the distribution G( iiJ) of iiJi in the form
G ( iiJ ) = ro ( iiJ- Wo )

+ g(wo+/(iiJ- wo) + (Ka ) /(iiJ-I~J2~~/(a )2 ·

2 2
)
(3·23)

The analytical results summarized above represent nothing more than one particular
solution of our special model. Moreover, we do not know how to confirm analyticallyits
 Cooperative Dynamics of Oscillator Community 235

stability. The difficulty of proving stability is related to the fact that our solution is valid
only for infinite N. To be more specific, we assumed in obtaining this solution statistical
independence of cPi as N goes to infinity except that the rings may be under a common
internal field of 'collective oscillation. Thus the stability in question is not purely
mechanical, but also of some statistical nature.
In order to check the stability and confirm the nonexistence of other stable solutions,
we carried out a computer simulation. This numerical study strongly suggests that the
above analytic solution is the only stable solution. Let g be given by a Lorentzian, i.e.,

(3· 24) '

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

for which explicit analytic solutions are available in the entire parameter range.
Numerical calculations were made for a system of 1000 rings. Initially, all rings were set
in an identical phase. Comparisons between numerical and analytical results are shown
in Figs. 3'"'--'7. Since the mean frequency Wo is irrelevant (i.e., elirninable by a
transformation from cPi to <Pi), it was set to zero from the beginning. Figure 3 shows the
order parameter a whose analytical expression is given by

a=/1 r; if 7J < 1,
=0 otherwise, (3·25)
where
r; = 2r/K. (3·26)
Although numerically calculated a( t)
fluctuates in time rather strongly due to 1.0
finite size effects, its long-time average
shows a smooth behavior as a function of 0
r;, and agrees with the analytical results
fairly well except for r; near criticality.
0.5
Figure 4 shows the true frequencies was a
function of natural frequencies w. From
this figure up to Fig. 7 (except for Fig. 6 ),
r; has been chosen to be 0.8. Each dot in
Fig. 4 corresponds to each ring, and Wi 0.0
were calculated from a long-time average 0.5 1.0
of dc/>d dt. The data points lie almost
perfectly on the analytic curve given by Fig. 3. Order parameter <J as a function of r; ;
comparison between computer simulation (open
(3·22), though the latter is not shown in the circles) and analytic solution (solid curve).
figure. Above criticality, the dots of
course lie on the straight line w= w. In Fig. 5 the distribution of w is shown in a
histogram. The corresponding analytic expression· is given by

(3·27)

although the coherent part ro(w-wo) is not indicated in the figure. Numerically, we
236 Y. Kuramoto

defined r as coming from those oscillator whose frequencies satisfy lw- wol < 0.05. The
dependence of r on 7J is shown in Fig. 6. Analytically,

15

1.0
.·
10

.. ··

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

0.5

,: 5

0.0
0.5 1.0
w 0 0.5 1.0
Fig. 4. True frequencies w of the oscillators versus Fig. 5. Number G of oscillators having frequency w
their natural frequencies w. The dots show the at 7J = 0.8 ; comparison between computer
result of computer simulation at 77 =0.8. In the simulation (histogram) and analytic solution
disordered state these dots come to lie on the (solid curve). Coherent peak which should
straight line w= cv. appear at w=O is not shown in the figure.

1.2
1.0

r
0.8

0.4

-Tt 0 Tt
'Y]
Fig. 6. Order parameter r as a function of 7J ; Fig. 7. Steady profile of number density distribution
comparison between computer simulation (open at 77 = 0.8 ; comparison between computer
triangles) and analytic solution (solid curve). simulation (histogram) and analytic solution
(solid curve).
 Cooperative Dynamics of Oscillator Community 237

r=j_tan- 12 ~
J[ r;
if r;<l,

=0 otherwise. (3·28)
Figure 7 shows a steady profile of the number density distribution n( ¢- wot ). This was
obtained again through a long-time average because n fluctuates too strongly.
Analytically,

(3·29)

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

The exact reflection symmetry of the histogram is due to the same symmetry in the initial
distribution adopted. In conclusion, our numerical study, although being only
preliminary, clearly shows the existence of a phase transition and the stability of our
analytic solutions.
3.3. Statistical ring dynamics based on Model 11 3 >. 4>
A .pair of interacting rings which are statistically identical by no means show
complete phase locking, as we saw in § 2. This is analogous to a pair of spins with
thermal fluctuation, for which no symmetry-breaking instability can occur. Our concern
below is therefore the limit of large number of rings, for which one may well expect the
occurrence of phase transitions. This is in fact demonstrated below analytically for
Model II. Again, we assume the interaction to be of infinite range, or
Tu ( ¢ ) = N - 1 r (¢ ),
although no functional form of r need be assumed now. The Langevin equation (3·2)
then takes the form

(3·30)

One may easily write down the corresponding Fokker-Planck equation for the probability
distribution P(¢,t) of the i-th oscillator. This is given by

ap a [{ w+ lzn: d¢ , r(¢-¢ , )n(¢ , ,t)}P J+D az

at=-&! p
a¢z. (3·31)
0

Because different rings are statistically independent in the present particular model, P(¢,
t) may be identified with n( ¢,t ). In this way, we obtain a nonlinear evolution equation
for n(¢,t) in a closed form

(3·32)

Clearly, the total number is conserved, and we impose the condition

2n:
1
0
n(c/;,t)d¢=1. (3·33)

Further, n(¢,t) is a 2Jr-periodic function,

238 Y. Kuramoto

n(r/>,t )= n(r/>+2Jr,t ). (3·34)

The solution of (3 · 32) as t ~ oo subject to periodic boundary condition (3 · 34) and

normalization condition (3 · 33) was discussed previously, 3 >. 4 >so that We will not repeat its
detail except for the following remarks. Equation (3·32) always admits a constant
solution

n(¢,t)=(2Jr)-\ (3·35)

which corresponds to the absence of collective oscillation. The problem is its stability.
For sufficiently large D, for which (3·32) reduces to a simple diffusion equation, the

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

constant solution is obviously stable. If the interaction is attractive, there exists a
critical value De below which the constant solution is destabilized and gives way to a
steadily propagating solution n( ¢>- Qf ). The propagation velocity Q is identified to w if
the interaction is variational,*> but generally Q depends on D. The onset of collective
oscillation of this kind is mathematically described by a Hopf bifurcation for system
(3·32). Small amplitude solutions near De is easily constructed through ordinary
perturb~tion methods, which we actually did previously. 4) Finally, we remark that there
exists another interesting approach, 17) based on the assumption of weak nonlinearity, i.e.,
soft orbits, to mutual synchronization in populations of limit cycle oscillators. Some
phase transitions similar to the present ones are also obtained there.

§ 4. Discussions and further extensions of the ring models

In the present paper, we first clarified the mathematical basis of some ring models, and
then discussed a number of deterministic and statistical aspects of those models, though
without going into full details in each case. In this section we point out one serious
limitation of our ring models, and discuss how to remove it. By serious limitation we
mean that our models cannot reflect features peculiar to relaxation oscillations. This
drawback becomes fatal when the increasing relaxational character ultimately leads to a
non-oscillating excitable state. It is true that as far as the oscillations persist, there is
always a way of defining ¢> such that the local systems behave like d¢>/ dt = w, which is the
fact we made full use of. This representation is not very natural, however, when the
oscillation is relaxational or highly distorted, and even leads to a breakdown when the
phase is self-locked. A consequence of unnatural definition of ¢> is that the approximation
for the ring-ring coupling by a function of the phase difference alone becomes very bad.
This approximation may in principle be valid- for sufficiently weak coupling, but the
problem is that the range of its validity is narrowed more and more with the loss of the
smoothness of oscillations.
An easiest way of taking account of non-smoothness of oscillations would be to
change the definition of ¢> suitably and work with modified local ring dynamics such as

(4 ·1)

*) For variational interaction, Model II is essentially the same as XY spin systems, as is easily seen by working
with phase disturbances cf;i to eliminate win (3·2).
 Cooperative Dynamics of Oscillator Community 239

where /(¢) is some 2x-periodic function of ¢. Clearly, (4·1) shows excitability or

self-locking provided the maximum of 1/ w is slightly greater than 1. Regarding the
property of r, we will conveniently retain the previous property that it is a function of the
phase difference, although this could not be justified from a rigorous perturbation-
theoretical viewpoint. Previously, we emphasized the importance of non-variational
nature of r as characterizing cooperativity in far-from-equilibrium situations. Its
importance seems to be lessened, however, for nonvanishing /. This is partly because the
local systems are already non-variational,*> and even with variational type coupling this
can lead to some properties peculiar to off-equilibrium situations, e.g., that oscillators tend
to be entrained to higher frequencies rather than lower frequencies.
Apart from the problems associated with excitability and relaxational character of

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

oscillations, the local ring dynamics in the form of (4 ·1) seems to generate a number of
problems of statistical-mechanical interest. Here we point our one of such problems.
Recall that Model II with w = 0 and variational interactions is equivalent to
thermodynamic systems of XY spins. Let us replace the local dynamics d¢~d dt = 0 in that
case by (4·1), but at first with w=O. If, in addition, we assume, e.g.,

/( ¢ )=- bsin2¢, (4·2)

the system still has nothing different from equilibrium systems because local systems (and
hence the entire system) possess a time-independent potential. In fact, assuming/(¢) in
the above form amounts to introducing uniaxial anisotropy to XY spins. The critical
behavior will then be changed to Ising-like. For sufficiently weak stochastic forces, the
system will stay in an Ising-like ordered state. We now introduce w into/(¢), and then
one could no longer imagine any equivalent equilibrium systems. The w term acts as a
rotating force. Well below the critical value We defined by Max (//We)= 1, the individual
rings will persist their state of self-locking, and the macroscopic state will not differ much
from an Ising-like ordered state. Beyond We, each ring will start oscillation, although at
first the oscillation will be extremely non-smooth. Macroscopically, one may expect a
phase transition from Ising-like ordered state to the state of coherent oscillation, although
the corresponding critical value of w will be considerably lower than We due to
fluctuations. Since there is no such things as free energy functional, the phase transitions
of this sort may provide a challenging problem to theorists. A study based on a mean
field equation analogous to (3 · 32) is now in progress. 18 >

Acknowledgements

I would like to express my sincere gratitude to Professor Kazuhisa Tomita who was
the first to introduce me to the world of nonlinear dynamics. Besides being an
outstanding physicist, he has been such an excellent teacher of physics that I have always
felt free in presenting him with my still premature ideas and never missed finding valuable
advices from him. It is my great pleasure to contribute the present paper on nonlinear
dynamics to the issue dedicated to him on the occasion of his retirement from Kyoto
University.

*> No new phase variable ¢ exists such that (4 ·1) is generated from an autonomous periodic potential H ( ¢ ),
which means that the local system is non-variational.
240 Y. Kuramoto

References

1) A. T. Winfree, ]. Theor. Biol. 16 (1967), 15.

2) Y. Kuramoto, in International Symposium on Mathematical Problems in Theoretical Physics, ed. by H. Araki,
Lecture Notes in Phys., Vol. 39 (Springer, New York, 1975 ), 420.
3) Y. Kuramoto, Physica 106A (1981), 128.
4) Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Springer, Berlin, Heidelberg, New York, Tokyo,
1984 ).
5) E. A. Coddington and N. Levinson, Differential Equations (McGraw-Hill, New York, 1955).
6) A. T. Winfree, The Geometry of Biological Time, Biomath. Vol. 8 (Springer, New York, 1980).
7) ]. Guckenheimer, ]. Math. Biol. 1 (1975), 259.
8) Y. Kuramoto and T. Yamada, Prog. Theor. Phys. 56 (1976), 724.

Downloaded from http://ptps.oxfordjournals.org/ at Pennsylvania State University on March 15, 2015

9) Y. Kuramoto, Prog. Theor. Phys. Suppl. No. 64 (1978), 346.
10) Y. Kuramoto and T. Tsuzuki, Prog. Theor. Phys. 55 (1976), 356.
11) T. Yamada andY. Kuramoto, Prog. Theor. Phys. 56 (1976), 681.
12) G. I. Sivashinsky, Acta Astronautica 4 (1977), 1177; 6 (1979), 569.
13) Y. Kuramoto, Prog. Theor. Phys. 63 (1980), 1185.
14) Y. Kuramoto, Prog. Theor. Phys. 71 (1984 ), 1182.
15) A. T. Winfree, in Theoretical Chemistry, Vol. 4, ed. by H. Eyring and D. Henderson (Academic, New York,
1978), 1.
16) See,
I
e.g., ]. }. Tyson, The Belousov-Zhabotinskii Reaction, Lecture Notes Biomath. Vol. 10 (Springer, Berlin,
Heidelberg, New York, 1976 ).
17) Y. Yamaguchi, K. Kometani and H. Shimizu,]. Stat. Phys. 26 (1981), 719.
Y. Yamaguchi and H. Shimizu, Physica llD (1984 ), 212.
18) S. Shinomoto and Y. Kuramoto, in preparation.