On the numerical integration of the multidimensional Kuramoto

model

Marcus A. M. de Aguiar
Instituto de Fı́sica ‘Gleb Wataghin’, Universidade Estadual de Campinas,
Unicamp 13083-970, Campinas, SP, Brazil
arXiv:2306.05939v4 [nlin.AO] 10 Aug 2023

The Kuramoto model, describing the synchronization dynamics of coupled oscil-

lators, has been generalized in many ways over the past years. One recent extension
of the model replaces the oscillators, originally characterized by a single phase, by
particles with D −1 internal phases, represented by a point on the surface of the unit
D-sphere. Particles are then more easily represented by D-dimensional unit vectors
than by D−1 spherical angles. However, numerical integration of the state equations
should ensure that the propagated vectors remain unit and that particles rotate on
the sphere as predicted by the dynamical equations. As discussed in [1] integration of
the three-dimensional Kuramoto model using Euler’s method with time step ∆t not
only changes the norm of the vectors but produces a small rotation of the particles
around the wrong axis. Importantly, the error in the axis’ direction does not vanish
in the limit ∆t → 0. Therefore, instead of displacing the unit vectors in the direction
of the velocity one should performed a sequence of direct small rotations, as dictated
by the equations of motion. This keeps the particles on the sphere at all times, en-
suring exact norm preservation, and rotates the particles around the proper axis for
small ∆t [1]. Here I propose an alternative way to do such integration by rotations
in 3D that can be generalized to more dimensions using Cayley-Hamilton’s theorem.
Explicit formulas are provided for 2, 3 and 4 dimensions. I also compare the results
with the forth order Runge-Kutta method, which seems to provide accurate results
even requiring renormalization of the vectors after each integration step.

I. INTRODUCTION

The Kuramoto model became a paradigm in the study of synchronization dynamics [2, 3].
It has been used to describe a variety of systems, such as circadian rhythms [4, 5], power
 2

grids [6–9], neuronal networks [10–13] and coupled metronomes [14]. The model has also
been extended in many ways, with the introduction of frustration [15–18], different types
of coupling functions [19–21], networks [22, 23], distributions of the oscillator’s natural fre-
quencies [24, 25], inertial terms [26–28], external periodic driving forces [29–31] and coupling
with particle swarms [32–34].
The original model describes the dynamics of N oscillators, represented by their phases
θi , and coupled according to the equations
N
k X
θ̇i = ωi + sin (θj − θi ) (1)
N j=1

where ωi are their natural frequencies, selected from a symmetric distribution g(ω), k is
the coupling strength and i = 1, ..., N . Kuramoto showed that, for k is sufficiently large,
the oscillators synchronize their phases, behaving as a single particle. The transition to
synchronization can be described by the complex order parameter
N
1 X iθi
z = peiψ ≡ e (2)
N i=1

with disordered motion resulting in p ≈ 0 and coherent motion in p ≈ 1. In the limit

N → ∞, the onset of synchronization can be described as a continuous phase transition,
p
where p = 0 for k < kc = 2/πg(0) and increases as p = 1 − kc /k for k > kc [22, 26].
Recently, Chandra et al [35] have shown that Kuramoto oscillators could also be repre-
sented by unit vectors σ⃗i = (cos θi , sin θi ) rotating on the unit circle. According to Eq.(1),
the dynamics of ⃗σi is given by

d⃗
σi k X
= Wi σ⃗i + [σ⃗j − (⃗
σi · σ⃗j )⃗
σi ] (3)
dt N j

where Wi is the anti-symmetric matrix

 
0 −ωi
Wi =  . (4)
ωi 0

The complex order parameter z, Eq.(2), can be written in terms of the real vector

1 X
p⃗ = σ⃗i = (p cos ψ, p sin ψ) (5)
N i

describing the center of mass of the system.

 3

II. MULTIDIMENSIONAL KURAMOTO MODEL

Eq.(3) can be extended to higher dimensions by simply considering unit vectors ⃗σi in
D-dimensions, rotating on the surface of the corresponding (D-1) unit sphere. The matrices
Wi become D × D anti-symmetric matrices containing the D(D − 1)/2 natural frequencies
of each oscillator. Finally, the D-dimensional model is further extended by replacing the
coupling constant k by a coupling matrix K [17, 18, 36]:

d⃗
σi 1
P
= Wi σ⃗i + N j [Kσ
⃗j − (⃗
σi · Kσ⃗j )⃗
σi ]
dt
(6)

= Wi σ⃗i + [K⃗p − (⃗
σi · K⃗p)⃗
σi ].

The coupling matrix breaks the rotational symmetry and plays the role of a generalized
frustration: it rotates ⃗σj , hindering its alignment with ⃗σi and inhibiting synchronization.
The angle of rotation depends on σj , generalizing the constant frustration angle of the
Sakaguchi model [15]. Norm conservation, |⃗
σi | = 1, is guaranteed, as can be seen by taking
the scalar product of Eqs.(6) with σ⃗i . Similar extensions of the Kuramoto model to higher
dimensions were also considered in refs. [37–39].

III. DYNAMICS ON THE SPHERE

Eq.(6) describes the dynamics of particles on the surface of a D-dimensional unit sphere
and the right hand side, expressing the velocity of the i-th oscillator, is tangent to the
sphere at all times. Euler’s method attempts to solve these equations by considering the
approximation where ⃗σi (t + ∆t) ≈ ⃗σi (t) + ⃗σ˙ i (t)∆t. This, however, moves the particle away
from the sphere and increases the norm of ⃗σi , requiring it to be manually renormalized as

⃗σi (t) + ⃗σ˙ i (t)∆t

⃗σi (t + ∆t) ≈ . (7)
|⃗σi (t) + ⃗σ˙ i (t)∆t|

Although the error in the norm is of order ∆t2 , it was shown in [1] that Euler’s procedure
introduces an error in the particle’s axis of rotation that does not vanish in the limit ∆t → 0.
In order to overcome this problem, a new integration method, relying on a sequence of
rotations using Euler’s angles, was proposed and shown to converge to correct result as
∆t → 0 [1].
 4

Here I will follow the ideas introduced in [1] and propose a slightly different integration
scheme that also preserves norm exactly and that can be extended to higher dimensions. In
D = 3 the method is equivalent to that proposed in [1], but easier to implement numerically.
I start by defining
p⃗k = K⃗p (8)

and noting that the second term on the RHS of Eq. (6) can be written as

K⃗p − (⃗
σi · K⃗p)⃗
σi = p⃗K − (⃗
σi · p⃗K )⃗
σi = Ui⃗σi (9)

where
Ui = p⃗k σ⃗i T − σ⃗i p⃗kT (10)

is an anti-symmetric matrix. The superscript T stands for transpose and the dyadic matrices
⃗B
are defined by (A ⃗ T )ij = Ai Bj and satisfy (A
⃗B⃗ T )C
⃗ = (B
⃗ · C)
⃗ A.
⃗ Equation (6) can now be

written as
d⃗
σi
= Vi σ⃗i (11)
dt
where
Vi = Wi + Ui = Wi + p⃗k σ⃗i T − σ⃗i p⃗kT (12)

is itself anti-symmetric.
If the matrix Vi were time-independent, the solution of Eq.(11) would be σi (t) =
eVi t σ⃗i (0). However, because Vi depends not only on the time but also on the full set of
particles’ positions σ⃗j , this is not true. Still, one can approximate

σ⃗i (t + ∆t) ≈ eVi (t)∆t σ⃗i (t) (13)

where eVi (t)∆t plays the role of an infinitesimal evolution operator. The error in this ap-
proximation comes only from the fact that Vi does not remain constant throughout the
time interval from t to t + ∆t. Norm conservation, though, is exact: since ViT = −Vi ,
T
|⃗σi (t + ∆t)|2 = ⃗σiT (t)eVi (t) eVi (t) σ⃗i (t) = |⃗
σi (t)|2 .
In three dimensions Eq.(11) is equivalent to

d⃗
σi ⃗
= Ωi × σ⃗i (14)
dt
 5

⃗ i are related to Vi by
where the components of Ω
 
 0 −Ω i3 Ω 2i

Vi =  Ωi3 0 −Ωi1  (15)
 
 
−Ωi2 Ωi1 0

with
Ωi1 = ωi1 + pk2 σi3 − pk3 σi2
Ωi2 = ωi2 + pk3 σi1 − pk1 σi3 (16)
Ωi3 = ωi3 + pk1 σi2 − pk2 σi1 .

Equation (14) explicitly represents a rotation on the sphere. The vector Ω̂i gives the
instantaneous axis of rotation, and depends on the natural frequencies of particle i and on
the position of all other oscillators through p⃗. This equation can be used directly to integrate
the dynamics, avoiding the artifacts introduced by changing the norm of the unit vectors
⃗σi and projecting them back to the sphere [1]. However, this form of vector product does
not hold in more dimensions. In the next sections I will propose an integration method,
based on Eq.(11), instead of Eq.(14), that is perhaps simpler in three dimensions and that
generalizes to higher dimensions as well. I shall give explicit formulas for D = 2, D = 3 and
D = 4.

IV. THREE DIMENSIONS

Using Cayley-Hamilton’s theorem we write

D−1
X
eVi ∆t = βk Vik (17)
k=0

where D is the dimension of Vi . This expression also holds for the eigenvalues λ of Vi ,
D−1
X
λ∆t
e = βk λk . (18)
k=0

Writing Eq.(18) for the D eigenvalues of Vi gives a linear system that can be solved for the
coefficients βk . For D = 3 the eigenvalues are ±iΩi and 0, which results in

sin(Ωi ∆t) 1 − cos(Ωi ∆t) 2

eVi ∆t = 1 + Vi + Vi . (19)
Ωi Ω2i
 6

Computing Vi2 I find

⃗ iΩ
Vi2 = Ω ⃗ T − Ω2 1 (20)
i i

where
Ω2i = Ω2i1 + Ω2i2 + Ω2i3 , (21)

and obtain
sin(Ωi ∆t) 1 − cos(Ωi ∆t) ⃗ ⃗ T
eVi ∆t = cos(Ωi ∆t)1 + Vi + Ωi Ωi . (22)
Ωi Ω2i
Finally, Eq.(13) becomes

σ⃗i (t + ∆t) = cos(Ωi ∆t)⃗σi + sin(Ωi ∆t)Ω̂i × ⃗σi + (1 − cos(Ωi ∆t))(Ω̂i · ⃗σi )Ω̂i (23)

⃗ i /Ωi and all quantities on the RHS are computed at time t. Eq.(23) corresponds
where Ω̂i = Ω
to a rotation of ⃗σi around Ω̂i by the angle Ωi ∆t [40]. Once all vectors ⃗σ have been updated,
⃗ i ’s are recalculated and the process iterated. Eq. (23) contains all orders of the
p⃗k and Ω
exponential and would be exact if Ω ⃗ i were constant. The error of the integration process

can be estimated by calculating the how fast the matrix elements of Ui , responsible for the
interactions, are changing over time. This allows us to adjust the time step ∆t according its
rate of change, increasing it when the rate is smaller than a specified threshold and reducing
it otherwise. This procedure can reduce the integration time considerably.

V. FOUR DIMENSIONS

In D = 4 equations (10) and (11) still work but the exponential of Vi = Wi + Ui is

more complicated for three reasons: first, Wi has six independent components; second, the
eigenvalues of Vi are not immediately available and; third, the exponential of Vi involves
powers Vi2 and Vi3 , which need to be calculated. Using Cayley-Hamilton’s formula we write

eVi ∆t = β0 1 + β1 Vi + β2 Vi2 + β3 Vi3 (24)

In the case of identical oscillators, Wi = 0, Eq.(10) can be used to write down the powers
of Vi in terms of dyadic matrices as

Vi2 = (⃗pk · ⃗σi )[⃗pk ⃗σiT + ⃗σi p⃗kT ] − p⃗k p⃗kT − p2k ⃗σi ⃗σiT (25)

and
Vi3 = [(⃗pk · ⃗σi )2 − p2k ][⃗pk ⃗σiT − ⃗σi p⃗kT ]. (26)
 7

In the general case, however, there is no simple form for Vi2 or Vi3 like in the 3D case (see
Eq. (20)) and they have to be computed numerically as shown below.
To compute the coefficients βk in Eq.(24) we need the eigenvalues of Vi . For each eigen-
value λ it holds that

eλ∆t = β0 + β1 λ + β2 λ2 + β3 λ3 . (27)

Writing this equation for the four eigenvalues allows us to find all β’s. Because Vi is anti-
symmetric, its eigenvalues come in pairs of complex conjugates, ±iα1 and ±iα2 . The result
is
β0 = [α12 cos(α2 ∆t) − α22 cos(α1 ∆t)]/(α12 − α22 )
β1 = [α13 sin(α2 ∆t) − α23 sin(α1 ∆t)]/[(α1 α2 )(α12 − α22 )]
(28)
β2 = [cos(α2 ∆t) − cos(α1 ∆t)]/(α12 − α22 )
β3 = [α1 sin(α2 ∆t) − α2 sin(α1 ∆t)]/[(α1 α2 )(α12 − α22 )].

Finally, writing V explicitly as

 
0 a b c
 
 −a 0 d e 
V= (29)
 

 −b −d 0 f 
 
−c −e −f 0

I obtain q
Ω2 1
√
α1 = 2
+ 2
Ω2 − 4Λ
(30)
q
Ω2 1
√
α2 = 2
− 2
Ω2 − 4Λ

where
Ω2 = a2 + b2 + c2 + d2 + e2 + f 2
(31)
Λ = a2 f 2 + e2 b2 + c2 d2 − 2af eb + 2af cd − 2abcd.

The algorithm for integrating the equations in D = 4 can then be summarized as follows:
for the initial distribution of oscillators, compute p⃗ and p⃗k = K⃗p. For each oscillator i
compute Vi = Wi + p⃗k σ⃗i T − σ⃗i p⃗kT and use Eqs. (30) and (31) to compute the angles α1 and
α2 . Compute the coefficients βk using Eq.(28) and update the oscillator’s position using

⃗σi (t + ∆t) = β0⃗σi (t) + β1 Vi⃗σi (t) + β2 Vi2⃗σi (t) + β3 Vi3⃗σi (t). (32)
 8

VI. HIGHER DIMENSIONS AND D=2

The formal evolution equations provided by Eqs.(23) and (32) extends naturally to any
number D of dimensions. In D = 5 one of the eigenvalues of Vi is always zero, implying
β0 = 1. The remaining four eigenvalues can be obtained analytically with expressions similar
to Eqs. (30) and (31). In 6 or more dimensions, however, solutions involve cubic or higher
order equations and need to be performed numerically.
The two dimensional case, corresponding to the Kuramoto model and its generalizations
in 2D [17], can be easily integrated using the same formalism. In this case
 
0 vi
Vi =   (33)
−vi 0

with vi = −ωi + pk1 σi2 − pk2 σi1 . The evolution operator becomes eVi ∆t = cos(vi ∆t) +
(sin(vi ∆t)/vi )Vi = R(vi ∆t) where R is the usual 2 × 2 rotation matrix
 
cos θ sin θ
R(θ) =   (34)
− sin θ cos θ

as expected. The components of ⃗σ are then easily updated according to ⃗σ (t + ∆t) =

R(vi ∆t) ⃗σ (t). In the next section I will show examples of numerical simulations in D = 3
and D = 4.

VII. SIMULATIONS

A. Three Dimensional systems

We first consider a single oscillator in 3D. In this case V is time independent and Eq.(23)
is exact for any ∆t. Choosing Ω ⃗ = (0, 0, ω) and ⃗σ (0) = (σ1 , σ2 , σ3 ) we can write

sin(ωt) 1 − cos(ωt)
⃗σ (t) = cos(ωt)⃗σ (0) + ⃗ × ⃗σ (0) +
ω (⃗ω · ⃗σ (0))⃗ω (35)
ω ω2

which results in the obvious rotation ⃗σ (t) = (σ1 cos(ωt)−σ2 sin(ωt), σ2 cos(ωt)+σ1 sin(ωt), σ3 ).
Although trivial, this is exactly one of the cases where Euler’s method fails [1]. Figure 1(a)
shows the third component of ⃗σ (t) as computed with Euler’s method using a time step of
0.0005 and compared with the exact constant result σ3 . AS discussed in [1], the reason σ3 (t)
 9

decreases to zero is that the module of ⃗σ (t + ∆t) is always slightly larger than 1 in Euler’s
method, as the velocity vector is tangent to the sphere. When ⃗σ (t + ∆t) gets renormalized,
by diving it by its module, the third component (that should stay constant) decreases a
little. Interestingly, integration with 4th order Runge-Kutta method (4RK) agrees with
the exact result. I hypothesize that the accuracy of 4RK comes from the fact that the
calculation with time step ∆t uses intermediate points at ∆t/4, ∆t/2 and 3∆t/4, adjusting
the velocity field much better around the sphere. Indeed one can verify that the module of
⃗σ (t + ∆t) is not always larger than 1 in RK4 simulations.

As a second example I consider N = 100 particles with coupling matrix K = 0.51 and
natural frequencies Wi given by vectors ω
⃗ i = (ωi1 , ωi2 , ωi3 + ω̄). The frequencies ωij are
Gaussian distributed with zero average and width δ = 0.1, but the particles have an overall
tendency to rotate around z with angular velocity ω̄ = 8. Initial conditions are chosen
randomly over the sphere. Fig.1(b) shows the same integration artifact observed for a single
particle.

Finally I consider an example where the coupling matrix K is not proportional to the
identity. I set

 
k cos γ k1 sin γ 0
 1 
K =  −k1 sin γ k1 cos γ 0  (36)
 
 
0 0 k2

with k1 = 1, k2 = 0.5 and γ = 0.5. As demonstrated in [18] this would result in a synchro-
nized state where the order parameter rotates in the x-y plane if all the averages natural
frequencies were zero. Fig. 1(c) shows results for the same settings as in fig. 1(b), i.e., ωij
Gaussian distributed with zero average, width δ = 0.1 and ω¯3 = 8. We see that the third
component of p⃗ increases from near zero, but Euler’s method imprecision makes it go in the
wrong direction. Even correcting the sign of p3 , its module does not converge to the correct
value. Panel (d) shows a similar simulation with ω¯3 = 0. Now the error is Euler’s method
is much smaller, although it is still there and propagate for longer times.
 10

0 .0 1 .0
(a ) (b )
-0 .1 0 .8

-0 .2 0 .6
p p
3 -0 .3 3 0 .4

-0 .4 0 .2

-0 .5 0 .0
0 5 0 1 0 0 1 5 0 2 0 0 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0
t t

0 .4 0 .3
(c ) (d )
0 .2 0 .2

p 3
0 .0
p 3
0 .1

-0 .2 0 .0

-0 .4 -0 .1
0 2 0 4 0 6 0 8 0 1 0 0 0 5 0 1 0 0 1 5 0 2 0 0
t t

Figure 1. Third component of order parameter as a function of time according to Euler’s method
(red thin line), Eq.(23)(black thick line) and 4th order Runge-Kutta (blue thin black line). In all
panels the average value of the natural frequencies is zero in directions 1 and 2. The coupling
matrix is diagonal in panels (a)-(b) with coupling intensity 0.5. For panels (c)-(d) the coupling
matrix is given by Eq.(36): (a) single particle with ω3 = 8; (b) n = 100 particles with ⟨ω3 ⟩ = 8; (c)
n = 100 particles with ⟨ω3 ⟩ = 8 (purple line shows −p3 for Euler’s method) ; (d) n = 100 particles
with ⟨ω3 ⟩ = 0.
 11

B. Four Dimensional Systems

Again I consider a single particle with coupling matrix proportional to the identity as a
first example. In this case U = 0 I choose W as
 
0 −ω 0 0
 
ω 0 0 0
V=W= . (37)
 
0 0 0 0
 
0 0 0 0

We find α1 = ω and α2 = 0. In this case the coefficients βk simplify to

β0 = 1
β1 = ∆t
(38)
β2 = [1 − cos(ω∆t)]/ω 2
β3 = [ω∆t − sin(ω∆t)]/ω 3 .

The matrices V2 and V3 can be readily calculated and the results is ⃗σ (t) = (σ1 cos(ωt) −
σ2 sin(ωt), σ2 cos(ωt) + σ1 sin(ωt), σ3 , σ4 ). Fig.2 shows p3 and p2 , the third and second com-
ponents of order parameter p⃗, as a function of time. Once again Euler’s method causes p3 to
decrease, consequently increasing the amplitude of oscillations in p2 and p1 . Using Eq.(32)
with adjustable time step also gives excellent results, and so does RK4 (not shown).
Next I consider N = 100 particles interacting with coupling matrix
 
k1 cos α k1 sin α 0 0
 
 −k sin α k cos α 0 0 
2 2
K= , (39)
 

 0 0 k cos β k sin β 

0 0 −k sin β k cos β

with k1 = 2.6, k2 = 1, k = 1.5, α = 0.5 and β = 0.2. I consider the full matrix of natural
frequencies
 
0 −ω6i ω5i −ω4i
 
 ω
 6i 0 −ω3i ω 
2i 
Wi =   (40)
 −ω5i ω3i 0 −ω1i 
 
ω4i −ω2i ω1i 0
and choose all ωij Gaussian distributed with zero average and width δ = 0.2. Fig.3 shows
comparisons between simulations with Euler’s method and Eq.(32). The overall effect on
 12

0 .0 1 .0
(a ) (b )
-0 .2
0 .5

-0 .4
p 3 p 0 .0
-0 .6 2

-0 .5
-0 .8

-1 .0 -1 .0
0 5 0 1 0 0
t 1 5 0 2 0 0 0 5 1 0
t 1 5 2 0

Figure 2. Time evolution of the (a) third component and (b) second component of the order
parameter for a single particle with ω = 8. Red thin line shows the result according to Euler’s
method (red thin line). Evolution according to Eq.(32) is shown as blue thick line (using fixed
time step 0.01) and black thin line (with adjustable time step). The coupling matrix is diagonal
with coupling intensity 3.

the module of p⃗ is not so large, but the trajectory p⃗(t) changes considerably as small errors
propagate after t ≈ 10. Simulations using RK4 method also works well and, for this range
of time, falls on top of the curve generated with Eq.(32.

VIII. CONCLUSIONS

Lee and collaborators have recently shown that the application of Euler’s method to the
integration of the 3D Kuramoto model results in numerical errors that cannot be corrected
by decreasing the simulation time steps [1]. The sources of error are the change in norm of
the propagated vectors, that move away from the sphere and need to be manually projected
back onto its surface, and the intrinsic error in the direction of the rotation axis produced
by Euler’s procedure. They have shown that proper integration must consider rotation of
the particles over the sphere and they do that with a sequence of rotations using Euler’s
angles. Here I proposed a slightly different way to perform such rotations that might be
easier to implement numerically and that can be extended to higher dimensions. The method
consists in writing the dynamical equations in the form ⃗σ˙ i = eVi ⃗σi and use Cayley-Hamilton’s
PD−1
theorem to write eVi = k=0 βk Vik , a finite sum that contains all orders of the exponential.
 13

1 .0
(a ) (b )
1 .0
0 .5
0 .8
p
p 3 0 .0
0 .6

0 .4 -0 .5

0 .2 -1 .0
0 1 0 2 0
t
3 0 4 0 5 0 0 1 0 2 0
t 3 0 4 0 5 0

1 .0
(c ) (d ) 0 .6

0 .5 0 .3

p
p
3 0 .0
2 0 .0
-0 .3

-0 .5 -0 .6 1 .0
-1 .0
0 .5
-0 .5
0 .0
0 .0

-1 .0 p 0 .5 -0 .5
p 2
-1 .0 -0 .5 0 .0
p 0 .5 1 .0
1
1 .0 -1 .0

Figure 3. Evolution of order parameter for the coupling matrix (39). Module of the order paramter
(a) and p3 (b) as a function ot time. Panel (c) shows the p1 × p2 projection and (d) the p1 × p2 × p3
projection. In all plots the red thin line shows the results with Euler’s method and the thick black
line the results using Eq.(32).

I provided explicit formulas for the cases of D = 2, 3 and 4 and compared the results
with different integration methods. I have also shown, numerically, that the fourth order
Runge-Kutta method does work well in three and four dimensions. The errors generated by
projecting the vectors back to the sphere after each integration time step do not seem affect
the accuracy of results in these cases.
It is not clear if the form of Eq.(11) is useful for exploring other properties of the Kuramoto
dynamics or if it is just a mathematical trick to facilitate numerical integration as I have
done here. As a final remark I note that the same formalism can be used to integrate the
Kuramoto model in the presence of external forces. Adding a period force to each oscillator
 14

like [29]
N
k X
θ̇i = ωi + sin (θj − θi ) + F sin(θi − Ωt) (41)
N j=1

results in the vector equation

d⃗
σi
= Wi σ⃗i + [(k⃗p − F⃗ ) − (⃗σi · (k⃗p − F⃗ ))⃗
σi ] (42)
dt

where F⃗ = (F cos(Ωt), F sin(Ωt)) [36]. This amounts to replace k⃗p by k⃗p − F⃗ and can be
extended to multi-dimensional systems with matrix coupling as

d⃗
σi
= Wi σ⃗i + [(K⃗p − F⃗ ) − (⃗σi · (K⃗p − F⃗ ))⃗
σi ] (43)
dt

where F⃗ is the generalized D-dimensional force applied to the oscillators. Therefore, to take
into account external forces, the only change in the integration algorithm is to replace the
vector p⃗K (t) = K⃗p(t) in Eq.(8) by p⃗K (t) = K⃗p(t) − F⃗ (t).

ACKNOWLEDGMENTS

It is a pleasure to thank Joao U.F. Lizarraga for helpful comments. This work was partly
supported by FAPESP, grant 2021/14335-0 (ICTP-SAIFR) and CNPq, grant 301082/2019-
7.
 15

[1] H. K. Lee, H. Hong, and J. Yeo, “Improved numerical scheme for the generalized kuramoto
model,” Journal of Statistical Mechanics: Theory and Experiment, vol. 2023, no. 4, p. 043403,
2023.
[2] Y. Kuramoto, “Self-entrainment of a population of coupled non-linear oscillators,” in In-
ternational Symposium on Mathematical Problems in Theoretical Physics, pp. 420–422,
Berlin/Heidelberg: Springer-Verlag, 1975.
[3] Y. Kuramoto, “Chemical Waves,” in Chemical Oscillations, Waves, and Turbulence, pp. 89–
110, Springer Berlin Heidelberg, 1984.
[4] S. Yamaguchi, H. Isejima, T. Matsuo, R. Okura, K. Yagita, M. Kobayashi, and H. Okamura,
“Synchronization of cellular clocks in the suprachiasmatic nucleus,” Science, vol. 302, no. 5649,
pp. 1408–1412, 2003.
[5] C. Bick, M. Goodfellow, C. R. Laing, and E. A. Martens, “Understanding the dynamics of
biological and neural oscillator networks through exact mean-field reductions: a review,” The
Journal of Mathematical Neuroscience, vol. 10, no. 1, pp. 1–43, 2020.
[6] G. Filatrella, A. H. Nielsen, and N. F. Pedersen, “Analysis of a power grid using a kuramoto-
like model,” The European Physical Journal B, vol. 61, no. 4, pp. 485–491, 2008.
[7] A. E. Motter, S. A. Myers, M. Anghel, and T. Nishikawa, “Spontaneous synchrony in power-
grid networks,” Nature Physics, vol. 9, no. 3, pp. 191–197, 2013.
[8] T. Nishikawa and A. E. Motter, “Comparative analysis of existing models for power-grid
synchronization,” New Journal of Physics, vol. 17, p. 015012, jan 2015.
[9] F. Molnar, T. Nishikawa, and A. E. Motter, “Asymmetry underlies stability in power grids,”
Nature communications, vol. 12, no. 1, p. 1457, 2021.
[10] D. Cumin and C. Unsworth, “Generalising the kuramoto model for the study of neuronal
synchronisation in the brain,” Physica D: Nonlinear Phenomena, vol. 226, no. 2, pp. 181–196,
2007.
[11] D. Bhowmik and M. Shanahan, “How well do oscillator models capture the behaviour of bio-
logical neurons?,” in The 2012 International Joint Conference on Neural Networks (IJCNN),
pp. 1–8, IEEE, 2012.
[12] F. A. Ferrari, R. L. Viana, S. R. Lopes, and R. Stoop, “Phase synchronization of coupled
 16

bursting neurons and the generalized kuramoto model,” Neural Networks, vol. 66, pp. 107–
118, 2015.
[13] A. S. Reis, K. C. Iarosz, F. A. Ferrari, I. L. Caldas, A. M. Batista, and R. L. Viana, “Bursting
synchronization in neuronal assemblies of scale-free networks,” Chaos, Solitons & Fractals,
vol. 142, p. 110395, 2021.
[14] J. Pantaleone, “Synchronization of metronomes,” American Journal of Physics, vol. 70, no. 10,
pp. 992–1000, 2002.
[15] H. Sakaguchi and Y. Kuramoto, “A soluble active rotater model showing phase transitions
via mutual entertainment,” Progress of Theoretical Physics, vol. 76, no. 3, pp. 576–581, 1986.
[16] W. Yue, L. D. Smith, and G. A. Gottwald, “Model reduction for the kuramoto-sakaguchi
model: The importance of nonentrained rogue oscillators,” Physical Review E, vol. 101, no. 6,
p. 062213, 2020.
[17] G. L. Buzanello, A. E. D. Barioni, and M. A. de Aguiar, “Matrix coupling and generalized frus-
tration in kuramoto oscillators,” Chaos: An Interdisciplinary Journal of Nonlinear Science,
vol. 32, no. 9, p. 093130, 2022.
[18] M. A. M. de Aguiar, “Generalized frustration in the multidimensional kuramoto model,” Phys.
Rev. E, vol. 107, p. 044205, Apr 2023.
[19] H. Hong and S. H. Strogatz, “Kuramoto model of coupled oscillators with positive and negative
coupling parameters: an example of conformist and contrarian oscillators,” Physical Review
Letters, vol. 106, no. 5, p. 054102, 2011.
[20] M. S. Yeung and S. H. Strogatz, “Time delay in the kuramoto model of coupled oscillators,”
Physical Review Letters, vol. 82, no. 3, p. 648, 1999.
[21] M. Breakspear, S. Heitmann, and A. Daffertshofer, “Generative models of cortical oscillations:
neurobiological implications of the kuramoto model,” Frontiers in human neuroscience, vol. 4,
p. 190, 2010.
[22] F. A. Rodrigues, T. K. D. M. Peron, P. Ji, and J. Kurths, “The Kuramoto model in complex
networks,” Physics Reports, vol. 610, pp. 1–98, 2016.
[23] J. S. Climaco and A. Saa, “Optimal global synchronization of partially forced kuramoto oscil-
lators,” Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 29, no. 7, p. 073115,
2019.
[24] J. Gomez-Gardenes, S. Gomez, A. Arenas, and Y. Moreno, “Explosive synchronization tran-
 17

sitions in scale-free networks,” Physical Review Letters, vol. 106, no. 12, pp. 1–4, 2011.
[25] P. Ji, T. K. D. Peron, P. J. Menck, F. A. Rodrigues, and J. Kurths, “Cluster explosive
synchronization in complex networks,” Physical Review Letters, vol. 110, no. 21, pp. 1–5,
2013.
[26] J. A. Acebrón, L. L. Bonilla, C. J. P. Vicente, F. Ritort, and R. Spigler, “The Kuramoto
model: A simple paradigm for synchronization phenomena,” Reviews of Modern Physics,
vol. 77, no. 1, pp. 137–185, 2005.
[27] F. Dörfler and F. Bullo, “On the critical coupling for kuramoto oscillators,” SIAM Journal
on Applied Dynamical Systems, vol. 10, no. 3, pp. 1070–1099, 2011.
[28] S. Olmi, A. Navas, S. Boccaletti, and A. Torcini, “Hysteretic transitions in the kuramoto
model with inertia,” Physical Review E, vol. 90, no. 4, p. 042905, 2014.
[29] L. M. Childs and S. H. Strogatz, “Stability diagram for the forced Kuramoto model,” Chaos,
vol. 18, no. 4, pp. 1–9, 2008.
[30] C. A. Moreira and M. A. de Aguiar, “Global synchronization of partially forced kuramoto os-
cillators on networks,” Physica A: Statistical Mechanics and its Applications, vol. 514, pp. 487–
496, 2019.
[31] C. A. Moreira and M. A. de Aguiar, “Modular structure in c. elegans neural network and its
response to external localized stimuli,” Physica A: Statistical Mechanics and its Applications,
vol. 533, p. 122051, 2019.
[32] K. P. O’Keeffe, H. Hong, and S. H. Strogatz, “Oscillators that sync and swarm,” Nature
communications, vol. 8, no. 1, pp. 1–13, 2017.
[33] K. O’Keeffe, S. Ceron, and K. Petersen, “Collective behavior of swarmalators on a ring,”
Physical Review E, vol. 105, no. 1, p. 014211, 2022.
[34] R. Supekar, B. Song, A. Hastewell, G. P. Choi, A. Mietke, and J. Dunkel, “Learning hydro-
dynamic equations for active matter from particle simulations and experiments,” Proceedings
of the National Academy of Sciences, vol. 120, no. 7, p. e2206994120, 2023.
[35] S. Chandra, M. Girvan, and E. Ott, “Continuous versus discontinuous transitions in the d-
dimensional generalized kuramoto model: Odd d is different,” Physical Review X, vol. 9, no. 1,
p. 011002, 2019.
[36] A. E. D. Barioni and M. A. de Aguiar, “Complexity reduction in the 3d kuramoto model,”
Chaos, Solitons & Fractals, vol. 149, p. 111090, 2021.
 18

[37] T. Tanaka, “Solvable model of the collective motion of heterogeneous particles interacting on
a sphere,” New Journal of Physics, vol. 16, 01 2014.
[38] M. Lipton, R. Mirollo, and S. H. Strogatz, “The kuramoto model on a sphere: Explaining its
low-dimensional dynamics with group theory and hyperbolic geometry,” Chaos: An Interdis-
ciplinary Journal of Nonlinear Science, vol. 31, no. 9, p. 093113, 2021.
[39] A. Crnkić, V. Jaćimović, and M. Marković, “On synchronization in kuramoto models on
spheres,” Analysis and Mathematical Physics, vol. 11, no. 3, pp. 1–13, 2021.
[40] H. Goldstein, C. Poole, and J. Safko, “Classical mechanics,” 2002, chapter 4. (American
Association of Physics Teachers).