Synchronization in Networks of Coupled Oscillators with Large Parameter Mismatches

Amirhossein Nazerian, Shirin Panahi, and Francesco Sorrentino∗

Department of Mechanical Engineering, University of New Mexico, Albuquerque, NM 87131, USA

Systems that synchronize in nature are intrinsically different from one another, with possibly large
differences from system to system. While a vast part of the literature has investigated the emergence
of network synchronization for the case of small parametric mismatches, we consider the general
case that parameter mismatches may be large. We present a unified stability analysis that predicts
why the range of stability either increases or decreases with parameter heterogeneity for a given
network. We introduce a parametric approach that allows us to estimate the effect of mismatches
on the stability of the synchronous solution in terms of contributions of pairs of eigenvalues of the
Laplacian. For cases in which synchronization occurs in a bounded interval of a parameter, we
study the effects of parameter heterogeneity on both transitions (asynchronous to synchronous and
synchronous to asynchronous.)

I. INTRODUCTION turbations when mismatches are absent. However, it is

Synchronization in networks of coupled dynamical os-
cillators continues to be the subject of intensive investi-
gation [1–7]. The emergence of synchronization in homo-
geneous networks consisting of coupled identical oscilla-
tors has been analyzed using the general framework of
the master stability function (MSF) [8, 9] and the con-
nection graph stability method [10]. One of the ongoing
challenges is to study the emergence of collective behav-
ior in heterogeneous networks [11–16], with a large part
of the literature on heterogeneous networks focusing on
phase oscillators [17–20]. Typical instances of synchro-
nization in the real world include situations in which the
individual parameters of the systems are heterogeneous
and are not fine-tuned [21, 22]. Hence, it is important to
study how synchronization may emerge in ensembles of
heterogeneous systems with large parameter variability.
Synchronization among non-phase oscillators with pa-
rameter mismatches has been studied in a number of pa-
pers; however, most of these papers focus on the case of
small mismatches. In particular, (i) Refs. [23–25] general-
ized the master stability approach of Pecora and Carroll
[8] to the case of slightly non-identical oscillators. As-
suming stability and small parametric mismatches, these
papers have found that deviations from the synchronous
state grow linearly with the mismatches; (ii) Acharyya
and Amriktar [26, 27] generalized the approach of (i) by
performing a higher-order expansion. The main differ-
ence between (i) and (ii) is that (ii) found that the sta-
bility is affected by parameter mismatches. Finally, a
recent paper (iii) [28] has focused on the case of Chua
oscillators and shown that, both in simulation and in ex-
periments, the stability of the synchronous solution is
enhanced by mismatches, which contrasts with previous
observations. Reference [28] has suggested that this syn-
chronization enhancement in the presence of mismatches
is a general phenomenon that is due to the ‘mixing’ of
the modes that describe the time evolution of the per-
∗
fsorrent@unm.edu
not clear why such mixing should always result in an
improvement for synchronizability.
In this paper, we provide a unified approach that can
be implemented to study the synchronization of a broad
class of dynamical networks, for which the stability of
the synchronous solution can be written in the form of a
linear time-invariant system [29]. This class includes net-
works of saddle focus oscillators [30, 31] and networks of
piecewise linear oscillators/maps. Similar to the classical
master stability function approach of Ref. [8], our ap-
proach decouples the effects of the dynamics from those
of the network topology. However, compared to the mas-
ter stability function approach, we perform a higher order
expansion in the heterogeneous parameters and consider
the effects of pairs of eigenvalues of the Laplacian (instead
of a single eigenvalue) on the stability of the synchronous
solution. Through this approach, we will be able to ex-
plain why parameter mismatches sometimes hinder and
sometimes favor network synchronization. We emphasize
that the previously introduced mode-mixing hypothesis
of Ref. [28] did not elucidate why parameter mismatches
sometimes enhance and sometimes hinder synchroniza-
tion.
The rest of the paper is organized as follows. The
general case of mismatches in the parameters of the indi-
vidual oscillators is discussed in Sec. II. Sec. III discusses
a case of particular interest in which mismatches affect
the frequencies of the individual oscillators. The case of
optoelectronic maps is discussed in Sec. IV. Finally, the
conclusions are provided in Sec. V.
II. PARAMETRIC MISMATCHES IN THE
LOCAL DYNAMICS
We consider a general equation for the time evolution
of N coupled dynamical systems, with parametric mis-
matches in the individual dynamics of each node,
X
ẋ xi (t), ri ) − σ
xi (t) = F (x xj (t)),
Lij H (x (1)
j

i = 1, . . . , N , x i ∈ Rm is the dynamical state of the
node i, F : Rm 7→ Rm and H : Rm 7→ Rm are the
local dynamics of each node and the coupling functions
between the nodes, and σ is a positive scalar measuring
the strength of the coupling. The symmetric Laplacian
matrix L = [Lij ] is constructed as L = G − A, where A =
[Aij ] is the adjacency matrix that describes the network
topology, and G is a diagonal matrix where the diagonal
entries are the sums of the rows of A. Each parameter
PN
is set equal to ri = r̄ + ϵδi , where r̄ = N −1 i=1 ri is a
nominal averagePvalue and ϵδi is a parametric mismatch,
2 bians on the average solution, see [29]. This assumption
P
j δj = 0, and j δj = 1, and the scalar ϵ is the tunable
magnitude of the mismatches.
Our work applies to different classes of the master sta-
bility function [32], i.e., to Class II, for which stability of
the synchronous solution is achieved in an infinite range
of the coupling gain σ and to Class III, for which stability
of the synchronous solution is achieved in a finite range of
the coupling gain σ. For Class II, as we increase σ from
zero, there is only one transition from asynchrony to syn-
chrony (A → S). For Class III, as we increase σ from zero,
first there is a transition from asynchrony to synchrony
(A → S) followed by another transition from synchrony
to asynchrony (S → A.) We develop a theory that allows
us to study the effects of parameter mismatches on each
individual transition and applies to either the case of A
→ S transitions or S → A transitions.
We assume possibly large mismatches and small per-
turbations about the average solution. The particular
assumption of small perturbation is discussed in detail
in Sec. SI P of the SupplementaryPMaterial. By defining
x(t) = N1 i x i (t) and r̄ = N1 i ri , we can write an
x̄
equation for the average solution and the time evolution
of small perturbations w i (t) = x i (t) − x̄x(t),
1 X
x˙ =
x̄ x, rj ) + DF
F (x̄ F (x̄
x, rj )w
wj
N j
X It is very important to emphasize that while Eq. (2b)
w i =DF
ẇ F (x̄ wi − σ
x, r̄)w H (x̄
Lij DH x)ww j + F (x̄
x, rj )
j part, we are mostly concerened with the homogeneous
1 X 1 X
− F (x̄
DF wj −
x, rj )w x, rj )
F (x̄
N j N j
where D is the differential operator. We emphasize that
the derivations we present only apply to the case that
the perturbations in the states of the systems are small.
At the same time, we do not introduce an assumption of
small mismatches.
In the following, for each transition, either A → S or
S → A, we investigate how parameter mismatches af-
fect the network ‘synchronizability’ [33], i.e., the range
of the parameter σ over which the synchronous solution
is stable. When we say that mismatches improve (hin-
der) synchronization, we mean that for that particular
transition this range is extended (narrowed.) In particu-
lar we study separately, each one of the transitions that
may arise, either (A → S) or (S → A) and see how the
2
critical σ that characterizes those transitions is affected
by parameter heterogeneity. We remark that this anal-
ysis needs to be repeated for each individual transition
and in fact, in what follows we report the example of a
particular system which has two transitions, A → S and
S → A, and for which parameter heteorogeneity increases
the synchronizability in the case of the A → S transition
and reduced it in the case of the S → A transition.
We proceed under the assumption that the particular
choice of the local dynamics F (·) and of the coupling
function H (·) results in a weak dependence of the Jaco-
has been previously validated for saddle-focus oscillators
(eg the Rössler system) in [29] and is validated here for
the Chua circuit and Bernoulli maps, see SM Sec. SII for
a discussion on the validity of this assumption. A more
general theory for which the assumption is removed is
presented in Sec. IV of this paper. We can then write
DF x(t), ri ) ≈ DF
F (x̄ F (ri ), where DF
F (ri ) = F +ϵδi B, where
F = DF F (r̄) and B ∈ Rm×m are constant matrices.
Analogously, we can also write DH x(t)) ≈ H, where
H (x̄
H ∈ Rm×m is a constant matrix. We comment now on
this assumption. While the assumption only applies to
certain choices of the functions F and H , the theory we
obtain still applies to a variety of continuous time and
discrete systems, such as Rössler systems, Chua circuits,
Bernoulli maps, and opto-electronic maps, as we show
in the rest of this paper. More importantly, it allows us
to extend the theory to the case of large parameter mis-
matches and to analyze whether parameter mismatches
either enhance or hinder synchronization.
We now define W = [w w⊤ ⊤ ⊤ ⊤
1 w 2 . . . w N ] , and rewrite
the homogeneous part of Eq. (2b) in compact form,
W (t) = MW
Ẇ W (t), (3)
(2a) where M = M0 + ϵM1 , M0 = IN ⊗ F − σ(L ⊗ H) with IN
indicating the identity matrix of size N , and M1 = ∆⊗B,
∆ = diag(δ1 , ..., δN ).
has both an homogeneous part and a non-homogeneous
(2b)
part of the equation. The homogeneous part of Eq. (2b)
describes stability; in case the homogeneous part is sta-
ble, the non-homogeneous part quantifies the deviation
from the synchronous state. The rest of this paper is de-
voted to analyze how parametric mismatches affect the
homogeneous part of the equation. This is in contrast
with Refs. [23–25] that focused on the non-homogeneous
part of the equation and did not study how paramet-
ric mismatches affect the synchronizability but only the
level of synchronization achieved in the network in case
the synchronous solution is stable.
In order to capture the dynamics of the transverse per-
turbation about the average solution, we introduce the
N × m-dimensional matrix Ṽ , with m = (N − 1), that
has for columns all the eigenvectors of the matrix L ex-
cept for the eigenvector [1, 1, ..., 1], Ṽ ⊤ Ṽ = Im . We have
that Ṽ ⊤ LṼ is equal to the m dimensional matrix Γ̃, with
 3

∗
entries on the main diagonal, −γm ≤ −γm−1 ≤ · · · ≤ W 0 = ⊕m i m i i i
i=1 W0 and V0 = ⊕i=1 V0 where W0 V0 = I, and
−γ1 < 0. By multiplying Eq. (3) on the left by (Ṽ ⊤ ⊗ I) the superscript ∗ indicates the conjugate transpose. We
and on the right by (Ṽ ⊗ I) and obtain M̃ = M̃0 + ϵM̃1 , define:
M̃ , M̃0 , M̃1 ∈ Rnm×nm , where
Λ = ⊕m
i=1 Λi , Λi = diag(λ1i , . . . , λm
i )∈C
m×m
⊤
M̃0 = (Ṽ ⊗ Ṽ )M0 (V ⊗ I) = Im ⊗ F − σ Γ̃ ⊗ H, (4a) Λ † = ⊕m Λ†,i = diag(λ1†,i , . . . , λm m×m
i=1 Λ†,i , †,i ) ∈ C

˜ ⊗ B,
M̃1 = (Ṽ ⊤ ⊗ I)M1 (Ṽ ⊗ I) = ∆ (4b) where Λ† refers to any of the Λ0 , Λ1 , or Λ2 matrices. The
w j0,i ) and the right (vv j0,i ) eigenvectors corresponding
left (w
and we have used the property that V ⊤ ∆2 V = V ⊤ DV = to the eigenvalue λj0,i are the column j of W0i and V0i ,
0. The transformed matrix ∆ ˜ = V ⊤ ∆1 V ∈ Rm×m is
respectively. We obtain,
symmetric and has a zero trace.
Next, we analyze stability of the transverse perturba-  
tions about the average solution by studying the sign of ˜ ⊗ B)V0
Λ1 = diag(W0∗ M̃1 V0 ) = diag W0∗ (∆ (5a)
the eigenvalues of M̃ as a function of the magnitude of

Λ2 = −diag Q(Q ◦ Π) (5b)
the mismatches, ϵ. We use matrix perturbation theory
to estimate the variation of the eigenvalues of M̃ as a
function of ϵ. This will result in a parametric approach, where ◦ indicates the Hadamard product, Q = W0∗ (∆ ˜⊗
similar to that of the master stability function [8], to an- B)V0 . The matrix Q has block structure Q = [[Qik ]]
alyze the effect of mismatches on the synchronizability of where each block [Qik ] = ∆˜ ik [Rik ] and the block [Rik ] =
networks of oscillators with mismatches. i∗ k n×n
W0 BV0 ∈ C . The matrix Π has block structure in
We call Λ the matrix that has the eigenvalues of M̃ on the blocks [Πik ] ∈ Cn×n , with entries
its diagonal, and by using second-order matrix perturba-
tion theory [34], we approximate Λ(ϵ) = Λ0 + ϵΛ1 + ϵ2 Λ2 , 
0 p=q &i=k
where Λ0 is a diagonal matrix with the eigenvalues of
[Πik ]pq = 1 (6)
M̃0 on the main diagonal, and Λ1 and Λ2 are diagonal  p otherwise
matrices to be calculated as a function of M̃0 and M̃1 . λ0,i − λq0,k
Given the block diagonal structure of M̃0 , its left
(W0 ) and right (V0 ) eigenvectors are written in the form Equation (5b) is rewritten as in Eq. (7).

˜ 11 [R11 ] · · · ˜ 1n [R1n ] ˜ 11 [R11 ] ◦ [Π11 ] · · · ˜ 1n [R1n ] ◦ [Π1n ]

   
∆ ∆ ∆ ∆
Λ2 = diag − 
  .. .. ..   .. .. .. 
(7)
. . .  . . . 
˜
∆n1 [Rn1 ] · · · ˜ ˜
∆nn [Rnn ] ∆n1 [Rn1 ] ◦ [Πn1 ] · · · ˜
∆nn [Rnn ] ◦ [Πnn ]

The real part of the blocks on the main diagonal of Λ2 blocks F − σγi H, F − σγk H and B; and with knowledge
are of F , H and B, it can be parametrized in the pair (ζi ,
m
! ζk ). This approach can be applied to study variations
of all the eigenvalues of the matrix M̃0 . However, in
X
Λ̄2,i = diag ˜ 2
∆ik [Uik ] , (8)
k=1
what follows, we focus on the eigenvalue(s) of M̃0 with
the largest real part (either a single eigenvalue or a pair
where the bar notation indicates the real part, and of complex conjugate eigenvalues), which we call critical.
We first consider a transition A → S and label i = 1 the
∗
 ∗  
[Uik ] = −real W0i BV0k (W0k BV0i ) ◦ [Πki ] . (9) block to which the critical eigenvalue(s) belongs. We call
the real part of the critical eigenvalue(s) c0 , i.e., c0 =
The entries ∆˜ ik reflect information about the mismatches maxj λ̄j0,1 , and s = argmaxj λ̄j0,1 . Then Eq. (8) needs to
(and the eigenvectors of the Laplacian); the blocks [Uik ] be evaluated for i = 1 and for a fixed value of ζ1 = a,
reflect information about the matrices F , H, the eigen- where a is the lower MSF bound for the synchronization
couplings ζi = σγi and ζk = σγk , and the matrix B which in the case of the identical oscillators. We approximate,
determines which of the parameters inside the local dy- c(ϵ) = c0 + ϵc1 + ϵ2 c2 . (10)
namics is the mismatched parameter.
From Eq. (8), we are interested in the sign of the entries where c(ϵ) is the real part of the critical eigenvalue of M̃
on the main diagonal of the block [Uik ], i.e., [Uik ]ss . We and the expansion coefficients c1 and c2 are characterized
note that each block [Uik ] can be parametrized in the by F , H, B and the eigencouplings σγi . Both c1 and c2
 4

are relevant to describe how c varies with ϵ, however in #104

0
what follows we focus on the sign of c2 , which determines

the curvature of c(ϵ). If c2 > 0 (c2 < 0), we expect c(ϵ)
to increase (decrease) with ϵ, which results in a decrease
(increase) of the synchronizability with parameter het-
erogeneity. If we find that [Uik ]ss < 0 ([Uik ]ss > 0) for all
ζk ≥ a, then c2 < 0 (c2 > 0.) We remark that our main
result is that we can characterize the improvement (hin-
drance) of the synchronizability in terms of the [Uik ]ss
from Eq. (9). A general conclusion is that the larger (the
smaller) the values of the function [Uik ]ss the more one
can expect synchronizability to improve (worsen) as pa-
rameter heterogeneity increases.
For certain choices of the functions F and H , stability
of the synchronous solution is found for σ in a bounded
interval σmin ≤ σ ≤ σmax . For σ ≥ σmin there is an
A → S transition and for σ ≥ σmax there is an S → A
transition. Then, the same approach outlined above can
be used to determine the effects of mismatches on σmax in
correspondence of the S → A transition. In this case, we
set i = n and ζn = b, where b is the upper MSF bound
for synchronization in the case of identical oscillators.
Then s = argmaxj λ̄j0,n . Equation (10) still holds with
c1 = λs1,n and c2 = λs2,n . If for all a ≤ ζk ≤ b, [Unk ]ss
is positive (negative), then c2 > 0 (c2 < 0), indicating
that mismatches enhance (reduce) the synchronizability
independent of the topology and of the particular values
of the mismatches.
A. Example
As an example, we choose the individual oscillators to
be Chua systems,
 

β(y − x − g(x)) β = 10

x) =  x − αy + z  , κ = −17.85
F (x
κy 
α = 1
(11)
(
b−a a = −1.44
g(x) = bx+ (|x−1|−|x+1|),
2 b = −0.72
where x = [x y z]⊤ and we set the coupling function
x) = x. We study the effect of mismatches in the
H (x
parameter κ with the nominal value shown in Eq. (11).
Hence, the constant matrices F, B and H that define this
mismatched problem are
   
β(−1 − b) β 0 0 0 0
F = 1 −α 1 , B = 0 0 0 , (12)
0 κ 0 0 1 0
and H is a 3 × 3 matrix with only nonzero entry H11 = 1.
We study the sign of c2 from Eq. (10). Since in the
case of identical oscillators, the MSF predicts that there
is only one transition A → S for ζ1 ≥ 6.00, we set i = 1.
In Fig. 1 we plot [U1k ]ss as we vary the eigencoupling
ζk ≥ 6.00, from which we see that for all choices of ζk ,
[U1k ]ss
-2
s=1
-4 s=2
6 10 25 40 55 70 85 100
1k
FIG. 1: The plot shows the real part of the first, and
the second entry on the main diagonal of [U1k ] versus
ζk . See the previous sections for our choice of the
matrices F , H, and B. Since [U1k ]ss < 0 ∀ζk ≥ 6.00,
then c2 < 0 in Eq. (10).
(a) (b)
3
1
2
FIG. 2: (a) shows the synchronization error Es for
coupled Chua oscillators with mismatches in the
parameter κ, see Eq. (11). The solid black curve
encloses the region of stability predicted from the
eigenvalues of M̃ from Eq. (4). The dashed red curve is
the approximation of the solid black curve derived from
matrix perturbation theory. (b) shows the topology of
the network.
[U1k ]ss < 0; therefore, c2 < 0 and the synchronizability
increases with the magnitude of ϵ independent of the par-
ticular choices of the mismatches (δi ) and of the network
topology (Laplacian L.)
The result from Fig. 1 is confirmed in Fig. 2 where
we plot the synchronization error (Fig. 2(a)) for a net-
work of three coupled Chua’s oscillators (Fig. 2(b)) with
randomly generated mismatches equal to δ1 = −0.6534,
δ2 = 0.7507 and δ3 = −0.0973. The synchronization
q
PN
1 2
error is defined as Es = N i=1 (x i (t) − x̄(t)) ,
PN
where x̄ = 1/N i=1 xi is the average of the x compo-
nents of all oscillators, and < · > denotes the average
over the time period t. Whenever nodes are approxi-
mately synchronized, Es attains a small value close to
zero (the dark blue region.) The solid black curve delim-
its the region of stability predicted from the eigenvalues
of M̃ from Eq. (4). The dashed red curve is the approx-
imation of the solid black curve obtained from second-
order matrix perturbation theory. We note that the solid
black contour from Fig. 2 well predicts the transition to
synchronization. Moreover, the solid black curve and its
approximation given by the dashed red curve (the lat-
ter for small |ϵ|) correctly replicate the curvature of the
 5
h i
critical σ versus the magnitude of the mismatches. W = (IN + ϵ∆1 + ϵ∆2 ) ⊗ DF
Ẇ F (x̄
x) W
h i  
− σ (L + ϵ∆1 L + ϵD) ⊗ DH x) W + ϵ δ ⊗ F (x̄
H (x̄ x)
III. PARAMETRIC MISMATCHES IN THE (15b)
FREQUENCIES OF THE INDIVIDUAL
OSCILLATORS where δ = [δ1 , . . . , δN ]⊤ and d = [d1 , . . . , dN ]⊤ , and the
matrices ∆1 = diag(δ1 , ..., δN ),
In this section, we consider a problem studied in [28] in    
which mismatches affect the frequencies of the individual δ1 δ2 · · · δN d1 d2 · · · dN
oscillators and show how our general theory presented in −1 δ1 δ2 · · · δN  d1 d2 · · · dN 
−1 
 
Sec. II can be specialized to this particular problem. We ∆2 =  .. .. . . ..  , D =  .. .. . . ..  .
N . . . .  N . . . . 
consider an equation for the time evolution of N coupled δ1 δ2 · · · δN d1 d2 · · · dN
dynamical systems, with frequency mismatches,
(16)
h X i As already noted before, Eq. (15b) is non-
ẋ xi ) − σ
xi = (1 + ϵδi ) F (x xj ) ,
Lij H (x (13) homogeneous, with the homogeneous part determin-
j ing stability about the average solution and the non-
homogeneous term determining the deviations of the in-
i = 1, ..., N, x i ∈ Rm is the dynamical state of the node dividual trajectories from this solution, see also [23–27].
i, F : Rm 7→ Rm and H : Rm 7→ Rm are the local dynam- We proceed similarly to what already done in Sec. II
ics of each node and the coupling functions between the and write the homogeneous part of Eq. (15b) in the form
nodes, and σ is a positive scalar measuring the strength W = MW
Ẇ W , where M = M0 + ϵM1 and
of the coupling. Similar to the previous section, the sym-
metric Laplacian matrix L = [Lij ] encodes the network M0 = IN ⊗ F − σL ⊗ H (17a)
connectivity. The term 1 + ϵδi is the particular frequency M1 = (∆1 + ∆2 ) ⊗ F − σ(∆1 L + D) ⊗ H (17b)
of node i, with 1 being the nominal frequency and ϵδi
representing the possibly large frequency mismatch. As Similar to Sec. II, we multiply Eq. (17) on the left by
before, ϵ represents the tunable
P magnitude of the mis- (Ṽ ⊤ ⊗ I) and on the right by (Ṽ ⊗ I) and obtain M̃ =
matches, and i δi = 0, and i δi2 = 1. In Sec. SIII of
P
M̃0 + ϵM̃1 , where
the SM, we take arbitrary δi ’s and show how the equa-
tions can be rewritten in the form of Eq. (13). M̃0 = (Ṽ ⊤ ⊗ Ṽ )M0 (Ṽ ⊗ I) = In ⊗ F − σ Γ̃ ⊗ H, (18a)
The average solution x̄ x(t) and the perturbations
about the average solution x(t) =
w i (t) are defined as x̄
M̃1 = (Ṽ ⊤ ⊗ I)M1 (Ṽ ⊗ I)
P
1/N i x i (t) and w i (t) = x i (t) − x̄ x(t), i = 1, . . . , N .
Next, we rewrite Eq. (13) in x̄x(t) and w i (t) by expanding (18b)
=∆˜ ⊗ F − σ∆ ˜ Γ̃ ⊗ H = (∆
˜ ⊗ I)M̃0 ,
to the first order about the average solution,
ϵ Xh i and we have used the property that V ⊤ ∆2 V = V ⊤ DV =
x˙ = F (x̄
x̄ x) + F (x̄
δj DF x)ww j − σdj DH
H (x̄
x)ww j (14a) 0, with M̃ , M̃0 , M̃1 ∈ Rnm×nm . The rest of the analysis
N j is done similarly to Sec. II. Next we provide our final
result analogous to Eq. (9) for the case of mismatches in
the frequencies. Similar to Eq. (8), the real part of the
1 X blocks on the main diagonal of Λ2 are obtained,
w i =(1 + ϵδi )DF
ẇ F (x̄
x)wwi − F (x̄
ϵδj DF x)wwj
N j n
!
X
X 1 Λ̄2,i = diag ˜ 2
∆ik [Uik ] , (19)
−σ H (x̄
(Lij + ϵδi Lij + ϵdj )DH x)ww j + ϵδiF (x̄
x) k=1
j
N
(14b) where each block
∗
 ∗  
where DF F , DHH ∈ Rm×m are the [Uik ] = −real W0i V0k Λk0 (W0k V0i Λi0 ) ◦ [Πki ] .
Ptime-varying Jacobian
matrices of and H , and dj = i δi Lij . Local stability
F (20)
about the average solution is described the first order Here, W0i , V0i and Λi0 denote the left and right eigenvec-
expansion (14) under the assumption that the deviations tor matrices and a diagonal matrix with the eigenvalues
w⊤
w i are small. By defining W = [w ⊤ ⊤
1 , . . . , wN ] ∈ R
Nm
, of the ith block of M̃0 , respectively. The block [Πki ] is
we rewrite Eq. (14), defined the same as in Eq. (6). The entries ∆ ˜ ik reflect
information about the mismatches (and the eigenvectors
1h ⊤ i
of the Laplacian), while the blocks [Uik ] reflect informa-
x˙ = F (x̄
x̄ x)+ ϵδδ ⊗ DF x) − σϵdd⊤ ⊗ DH
F (x̄ H (x̄
x) W
N tion about the functions F , H , and the eigencouplings
(15a) ζi = σγi and ζk = σγk .

From Eq. (19), we are interested in the sign of the en-
tries on the main diagonal of the block [Uik ],i.e., [Uik ]ss .
We note that each block [Uik ] can be parametrized in the
blocks F − σγi H and F − σγk H; and with knowledge of
F and H, it can be parametrized in the pair (ζi , ζk ).
Next, similar to Sec. II, we focus on the eigenvalue(s) of
M̃0 with the largest real part (either a single eigenvalue
or a pair of complex conjugate eigenvalues), which we
call critical. We label i = 1 the block to which the criti-
cal eigenvalue(s) belongs and expand the real part of the
critical eigenvalue(s) to second order in ϵ, see Eq. (10).
For not too large ϵ, c(ϵ) determines the crossing value
of σmin from asynchrony to synchrony. In particular, we
focus on characterizing the curvature coefficient c2 .
With the choice of F and H and having set ζ1 = a,
we sweep over ζk and evaluate [U1k ]ss . If for all ζk ≥ a,
[U1k ]ss is positive (negative) then c2 > 0 (c2 < 0), so
the presence of mismatches reduces (increases) the syn-
chronizability for any topology and any value of the mis-
matches.
A. Example
Next, we consider two examples, one for which the
range of stability increases with ϵ and one for which it
decreases. The first example is that of Chua systems
coupled in the ‘x’ variable, considered in [28],
  (
η(y − x − g(x))
η = 10
x) =  x − y + z  ,
F (x
−y/k k = 0.056
( (21)
1 a = −1.44
g(x) = bx+ (b−a)(|x−1|−|x+1|),
2 b = −0.72.
The matrices in Eq. (17) are,
 
η(−1 − b) η 0
F = 1 −1 1 , (22)
0 −1/k 0
with the same parameters as in Eq. (21), and H = [Hij ] ∈
R3×3 where Hij = 0 except for H11 = 1. With no pa-
rameter mismatches, the MSF [8, 35] predicts that sta-
bility of the synchronous solution for ζ1 = σγ1 ≥ a = 6.
We thus fix ζ1 = a = 6 and study the contribution of
ζk ≥ ζ1 on [U1k ]. Figure 3 shows [U1k ]ss as a function
of ζk ≥ ζ1 = 6. For any real positive ζk , F − ζk H has
one real and two complex and conjugate eigenvalues. We
order these eigenvalues such that the first and the second
eigenvalues are complex and conjugate and the third one
is real. The complex eigenvalues are critical as these have
real part larger than the real eigenvalue, so s = 1, 2. We
characterize the second order coefficient of the expansion
c2 relative to these critical eigenvalues. Figure 3 shows
that for all ζk ≥ ζ1 = 6, [U1k ]ss is negative. Thus, based
on Eq. (19) we conclude that c2 < 0. From Fig. 3 we see
that c2 < 0 holds for any network topology, and so the
6
0
[U1k ]ss
-50
s=1
s=2
-100
610 25 40 55 70 85 100
1k
FIG. 3: Second order expansion. The [U1k ]ss ,
s = 1, 2, as ζk is varied. The plot shows the real part of
the first, and the second entry on the main diagonal of
[U1k ], respectively. See Eq. (22) for the matrices F and
H. For coupled identical oscillators, synchronization is
achieved for ζ1 ≥ 6.00 based on the MSF.
enhancement of synchronization with the mismatches is
a network-independent effect. Based on the results from
Fig. 3, we expect to see that increasing mismatches will
improve the synchronizability for the case of Chua oscil-
lators presented above. This is confirmed in Fig. 4 (a)
where we plot the synchronization error E as the cou-
pling strength σ and the magnitude of the mismatches ϵ
are varied for coupled Chua circuits with the undirected
network shown in panel (b). Whenever nodes are ap-
proximately synchronized, E attains a small value close
to zero (the blue region.) The black contour delimits the
region of stability predicted from the eigenvalues of M̃ .
We see that the linear approximation well describes the
(a) (b)
6 4
1 2
5 3
FIG. 4: Synchronization error predicted by
matrix perturbation theory. (a) Synchronization
error E for a network of coupled Chua circuits,
Eqs. (13) and (21), as a function of the coupling
strength σ and the magnitude of the mismatches ϵ. The
black contour encloses the region of stability predicted
from the eigenvalues of M̃ from Eq. (18). The white
contour is an approximation of the black contour based
on matrix perturbation theory, Eq. (10). (b) Topology
of the network, with mismatches randomly set equal to
δ = [−0.067, −0.518, −0.358, 0.712, 0.294, −0.062]⊤ .
emergence of synchronization in the nonlinear system,
Eq. (13), even for relatively large mismatches, |ϵ| ≃ 0.2.
The linear approximation correctly predicts that the sta-
bility range increases for increasing |ϵ|, as previously re-
ported in [28]. The white curve in Fig. 4(a) is obtained
by plotting c(ϵ) from Eq. (10), which provides a good ap-
proximation of the black curve, even for relatively large
 7

values of |ϵ|. IV. DISCRETE TIME OPTO-ELECTRONIC

SYSTEMS
The second example is that of Bernoulli maps with
mismatches, described by discrete time dynamics,
We now consider a case of particular interest for which
h X i the dynamics is in the general form of Eq. (13) and the
xk+1
i = (1 + ϵδi ) f (xki ) − σ Lij h(xkj ) , (23) functions F and H are taken in the particular form of
j discrete time opto-electronic systems [36–38],
"

k+1
i = 1, ..., N, where f (xk ) = 2xk mod 1, and h(xk ) = xk . xi = (1 + ϵδi ) βF (xki )
Fig. 5(a) is a plot of the synchronization error versus ϵ
and σ, from which we see that as |ϵ| grows the range of
#
X 
σ that supports synchronization (blue area) shrinks. For −σ Lij F (xkj ) + α mod 2π,
this case, F = 2 and H = 1. In Sec. SIV of the SM j
we apply the same analysis presented before to approxi- (24)
mate the eigenvalues of the matrix M̃ , and obtain a more
subtle result that most of the times c2 > 0, but not al- where F (x) = (1 − cos(x))/2, β is the self-feedback
ways. The region inside the black contour is where all strength, and the offset α is introduced to suppress the
the absolute values of the matrix M̃ are between −1 and trivial solution xi = 0. The case of the opto-electronic
1, which corresponds to the stability of the synchronous systems is of particular interest because our approach
solution. The white contour is the approximation of the described in Secs. II and III can be applied without in-
black curve based on the matrix perturbation theory. We troducing the assumption that the Jacobians are weakly
see that the lower (upper) bound for σ increases (de- dependent on the synchronous solution.
creases) with |ϵ|, causing an overall reduction of the syn- We assume possibly large mismatches and small per-
chronizability with mismatches, which we find to be inde- turbations about the average solution, x̄. Then, by fol-
pendent of the particular choice of the network topology lowing steps similar to the previous sections, we obtain
(panel b). the following equation that describes the time evolution
of small transverse perturbations about the average so-
lution,
(a) (b) " #
k+1 k
6 5 Y
Ỹ k ˜
= DF (x̄ ) (IN −1 + ∆)(βIN −1 − σ Γ̃) Ỹ
Y . (25)
7 4

k
8 3 Y ∈ RN −1 , and the matrices Γ̃ and ∆
where Ỹ ˜ are exactly
9 2 the same as they were defined in the previous sections.
1 ˜
We define M̃ = (IN −1 + ∆)(βI N −1 − σ Γ̃). By diagonal-
izing M̃ , Eq. (25) can be rewritten,

FIG. 5: The case of Bernoulli maps. (a) Z k+1 = ΦDF (x̄k )Z

Zk (26)
Synchronization error for a network of nine coupled
Bernoulli maps, Eq. (23), as a function of the coupling where Φ is a diagonal matrix that has the eigenvalues of
strength σ and the magnitude of the mismatches ϵ. The M̃ on its main diagonal. Equation (26) can be decoupled
region inside the black contours (eigenvalues) is where into N independent equations,
all the absolute values of the eigenvalues of M̃ , Eq. (18),
zik+1 = ϕi DF (x̄k )zik , i = 1, . . . , N. (27)
are between -1 and 1. The white contour is an
approximation of the black contour based on matrix By defining the free parameter s, we can then study the
perturbation theory. (b) Topology of the network. Here stability of the parametric equation
δ = [0.1568, −0.0869, −0.6469, −0.4689, 0.0152, 0.1642,
0.3971, 0.1033, 0.3661]⊤ was chosen randomly. z k+1 = sDF (x̄k )z k , (28)

as a function of the variable s = ϕi . For practical pur-

We note that our analysis presented here for Chua os-
cillators and Bernoulli maps can be similarly applied to
different classes of oscillators and couplings. The case of
Rössler systems coupled in the x or y variables [29] stud-
ied in the Sec. SV of the SM. The cases where DF F = DH H,
including opto-electronic maps presented in the next sec-
tion.
poses, following [24, 39, 40], the average solution x̄k can
be approximated by the dynamics of a single uncoupled
map. We have calculated the Maximum Lyapunov Ex-
ponent (MLE) Ψ(s) of Eq. (28) as a function of the real
variable s for the optoelectric map [37], from which we
have found that for β = 2π and α = 0.525, Ψ(s) < 0 is
negative in the range [s− = −3.47, s+ = 3.47].
 8

1
where
0
(
(1−σγi )(1−σγj )
*(s) σγj −σγi i ̸= j
Π̃ = M̃0 ΠM̃0 = . (33)
-1 0 i=j

-2 We see that the sign of Λ2,i from Eq. (32) is fully defined
-20 -10 0 10 20
s for a given pair of eigencouplings ζi = σγi and ζj = σγj .
As discussed earlier, the network of coupled optoelectric
FIG. 6: Maximum Lyapunov Exponent. The MLE maps with mismatches can synchronize if the eigenvalues
associated with Eq. (28) as the parameter s is varied. of M̃ are in the range [−3.47, 3.47]. Hence, the condition
lims→0 Ψ(s) = −∞. The zero-crossing points are for synchronization is that,
s+ = 3.47, and s− = −3.47. Here, β = 2π and
α = 0.525. −3.47 ≤ Λ0,i + ϵΛ1,i + ϵ2 Λ2,i ≤ 3.47 (34)

which is equivalent to
In Fig. 6 we plot the Maximum Lyapunov Exponent
N
(MLE) Ψ(s) of Eq. (28) as a function of the real variable ˜ ii (1 − ζi ) + ϵ2
X
˜ 2ij Π̃ji ≤ 3.47
s for the optoelectric map [37]. We see that Ψ(s) < 0 is −3.47 ≤ (1 − ζi ) + ϵ∆ −∆
negative in the range [s− , s+ ], and stability is achieved j=1

for ϕ− > s− and ϕ+ < s+ , where ϕ− = mini ϕi and (35)

ϕ+ = maxi ϕi . Mismatches affect both s− and s+ . We Figure 7 illustrates the usefulness of our proposed
say that they enhance synchronization when they either method through an example. Panel (a) shows the syn-
increase ϕ− or decrease ϕ+ . chronization error as a function of both σ and ϵ for the
Therefore, stability is achieved for ϕ− > s− and ϕ+ < N = 5 network of coupled optoelectric maps shown in
s , where ϕ− = mini ϕi and ϕ+ = maxi ϕi . Mismatches
+ (b). For ϵ = 0 we see that as σ increases, there are two
affect both ϕ− and ϕ+ , and we say that they enhance transitions: one around σ = 0.94, for which the identical
synchronization when they either increase ϕ− or decrease systems switch from asynchrony to synchrony, and an-
ϕ+ . other one around σ = 1.94 where the transition is from
Next, we investigate how mismatches affect the eigen- synchrony to asynchrony. Panels (c-f) of Fig. 7 show the
values of M̃ . Once again, we use matrix perturbation application of our theory to predict whether or not mis-
theory to study the behavior of the eigenvalues of M̃ as matches enhance the network synchronizability. Inter-
the magnitude of the mismatches ϵ is varied. We write estingly, in this case, we see that both σ − and σ + de-
M̃ = M̃0 + ϵM̃1 , where crease with |ϵ|. In particular, panel (c) shows how Λ+ ,
the largest eigenvalue of M̃ , varies with ϵ for σ = 0.936,
M̃0 = IN − σΓ, M̃1 = ∆ ˜ M̃0 , (29) which is slightly less than σ − = 0.94. We see that for ei-
and Γ is the diagonal matrix with diagonal entries equal ther large enough ϵ or small enough ϵ, Λ+ decreases below
to the eigenvalues of the Laplacian L. The matrix 3.47, which causes synchronization. Instead, panel (f)
˜ = V −1 ∆V where V is a matrix whose columns are
∆ shows how Λ− , the smallest eigenvalue of M̃ , varies with
the eigenvectors of L. We have that the eigenvalues Λϵ ϵ for σ = 1.95, which is slightly more than σ + = 1.945.
of the matrix M̃ can be approximated to second order, In this case, we see that as |ϵ| increases, Λ− decreases
Λϵ = Λ0 + ϵΛ1 + ϵ2 Λ2 , where the matrix Λ0 is a diag- further away from −3.47, which is the threshold for syn-
onal matrix that has the eigenvalues of A0 on its main chronization. Hence, we conclude that the presence
diagonal (since M̃0 is also diagonal, Λ0 = M̃0 and the of mismatches in the frequencies of the coupled opto-
matrices for the right and the left eigenvectors of M̃0 are electronic maps helps the synchronization in the transi-
the identity) and, tion from asynchrony to synchrony (critical lower bound
  of σ) and hinders synchronization in the transition from
Λ1 = diag(M̃1 ) = diag ∆Λ ˜ 0 (30a) synchrony to asynchrony (critical upper bound of σ.)


Λ2 = −diag Q(Q ◦ Π) . (30b)
V. CONCLUSIONS
˜ ij Λ0,j ] and
Here, Q = [Qij ] = [∆

0 i=j This paper studies the emergence of synchronization
Πij = 1 (31) in networks of coupled oscillators with possibly large pa-
otherwise. rameter mismatches. We propose a parametric approach
Λ0,i − Λ0,j

to analyze how the size of the mismatches affects the syn-
Each Λ2,i is equal to, chronizability, i.e., the range of the coupling strength σ
N
X N
X for which the synchronous solution is stable. We iden-
Λ2,i = ˜ ij Λ0,j (∆
−∆ ˜ ji Λ0,i ◦Πji ) = − ˜ 2ij Π̃ji . (32)
∆ tify a critical eigenvalue c that is responsible for stabil-
j=1 j=1 ity of the synchronous solution and expand c to second
 9

(a) (c) (e)

(b)
4 3
𝜎 = 0.936 𝜎 = 1.95
1 2
5

(d) (f)
FIG. 7: The case of coupled ‘opto-electronic maps’. (a) Synchronization error for a network of N = 5 coupled
optoelectric maps as a function of the magnitude of the mismatches ϵ and of the coupling strength σ. Here,
δ = [−0.5323, −0.5265, 0.1526, 0.5058, 0.4003]⊤ was randomly chosen. When the maps are identical (ϵ = 0),
synchronization is achieved for σ − = 0.94 ≤ σ ≤ 1.945 = σ + . We see that for σ slightly lower than σ − (σ slightly
greater than σ + ), the network of mismatched oscillators synchronizes (does not synchronize). (b) Topology of the
network. (c) shows Λ+ , the maximum eigenvalue of M̃ as a function of ϵ for σ = 0.936. When ϵ = 0, the network is
not synchronized since Λ+ is greater than the critical bound s+ = 3.47 shown as a black line. As ϵ is either
increased or decreased, Λ+ eventually shifts below s+ = 3.47 and the network synchronizes. (d) shows Λ− , the
minimum eigenvalue of M̃ remaining inside the critical bounds as ϵ varies for σ = 0.936. (e) shows Λ+ remaining
inside the critical bounds as ϵ varies for σ = 1.95. (f) shows Λ− as a function of ϵ for σ = 1.95. When ϵ = 0, Λ− is
slightly less than s− . As ϵ is varied, Λ− shifts further away from s− ; hence synchronization is not achieved.

order in the scalar ϵ which measures the size of the mis-
matches; we focus on the second order coefficient of the
expansion, i.e., the curvature coefficient c2 : if this coeffi-
cient is positive (negative), we expect the synchronizabil-
ity to decrease (increase) with the size of the mismatches.
We derive an expression for how each pair of eigencou-
plings (ζi , ζj ) (the eigenvalues of the Laplacian matrix)
contributes to the curvature coefficient c2 .
Similar to the parametric approach of the master sta-
bility function [8], it becomes possible, given knowledge
of the functions F and H , to pre-compute the effects of
any pair (ζi , ζj ), seen as parameters. With this knowl-
edge, one can then find for any network realization (any
set of eigenvalues) the total curvature attained. For ex-
ample, given knowledge of F and H , our numerical ap-
proach consists of numerically computing the function
shown in Fig. 1 and then from this, computing the over-
all effect on the curvature through Eq. (8). In cases in
which the function is always positive (always negative),
we can conclude that the curvature is positive (negative)
independent of the network topology. Our main contri-
bution is exactly that we can explain when parameter
heterogeneity either increases or decreases the synchro-
nizability.
Is the enhancement of synchronizability with param-
eter mismatches a general phenomenon as reported in
[28]? We conclude that the answer is no, as our theory in-
dicates that this is dependent on a number of factors and
we find several examples of systems that become less syn-
chronizable as we increase |ϵ|. Our work applies to both
continuous-time and discrete-time networks. The ex-
amples involving Chua’s oscillators showed that the pres-
ence of parametric mismatches improves synchronizabil-
ity while the example of Bernoulli maps showed the op-
posite. In the case of opto-electronic maps, it was shown
that one-sided improvement and one-sided hindrance to
synchronizability were present simultaneously. In the
transition from asynchrony to synchrony (synchrony to
asynchrony), the synchronizability improved (hindered)
by the presence of mismatches.
Our work differs substantially from Refs. [23–25] since
these papers are in the framework of small parameter
mismatches ↔ small state deviations. It also differs from
[28] as we provide rigorous conditions, based on matrix
perturbation theory, that allow us to explain why param-
eter mismatches sometimes favor synchronization and
sometimes hinder it. We depart from the ‘mode mixing’
hypothesis [28] and the implication that this should al-
ways result in an improvement in synchronizability. Such
mode mixing (due to the presence of two non-commuting
matrices [41–43]) had already been observed in [26, 27]
but no conclusions were drawn on its usefulness towards
synchronization. Furthermore, our approach allows us to
investigate the general case in which synchronization is
possible in a bounded range of a given parameter and
to consider the effects of parameter mismatches on both
the lower and upper end of the range, which was not
considered before.
ACKNOWLEDGEMENTS
This work is supported by NIH grant 1R21EB028489-
01A1. The authors thank Chad Nathe for insightful con-
versations and his help with performing simulations.

[1] M. Shafiei, F. Parastesh, M. Jalili, S. Jafari, M. Perc,
and M. Slavinec, The European Physical Journal B 92,
1 (2019).
[2] J. Ding, I. Belykh, A. Marandi, and M.-A. Miri, Physical
Review Applied 12, 054039 (2019).
[3] S. Boccaletti, A. N. Pisarchik, C. I. Del Genio, and
A. Amann, Synchronization: from coupled systems to
complex networks (Cambridge University Press, 2018).
[4] S. S. Sajjadi, D. Baleanu, A. Jajarmi, and H. M. Pirouz,
Chaos, Solitons & Fractals 138, 109919 (2020).
[5] N. M. Awal and I. R. Epstein, Physical Review E 101,
042222 (2020).
[6] H. Fan, L.-W. Kong, X. Wang, A. Hastings, and Y.-C.
Lai, National science review 8, nwaa269 (2021).
[7] A. Palacios, Physical Review E 103, 022206 (2021).
[8] L. Pecora and T. Carroll, Phys. Rev. Lett. 80, 2109
(1998).
[9] S. Boccaletti, G. Bianconi, R. Criado, C. I. Del Ge-
nio, J. Gómez-Gardenes, M. Romance, I. Sendina-Nadal,
Z. Wang, and M. Zanin, Physics reports 544, 1 (2014).
[10] V. N. Belykh, I. V. Belykh, and M. Hasler, Physica D:
nonlinear phenomena 195, 159 (2004).
[11] J. Xiang and G. Chen, Automatica 43, 1049 (2007).
[12] F. Ricci, R. Tonelli, L. Huang, and Y.-C. Lai, Physical
Review E 86, 027201 (2012).
[13] W. Lu, B. Liu, and T. Chen, Chaos: An Interdisciplinary
Journal of Nonlinear Science 20, 013120 (2010).
[14] I. Belykh, V. Belykh, K. Nevidin, and M. Hasler, Chaos:
An Interdisciplinary Journal of Nonlinear Science 13, 165
(2003).
[15] A.-K. Malchow, I. Omelchenko, E. Schöll, and P. Hövel,
Physical Review E 98, 012217 (2018).
[16] B. Ramakrishnan, F. Parastesh, S. Jafari, K. Rajagopal,
G. Stamov, and I. Stamova, Fractal and Fractional 6,
169 (2022).
[17] Y. Kuramoto, Chemical Oscillations, Waves and Turbo-
lence (Springer, Berlin, 1984).
[18] M. G. Rosenblum, A. S. Pikovsky, and J. Kurths, Phys-
ical Review Letters 78, 4193 (1997).
[19] J. C. Lacerda, C. Freitas, E. E. Macau, and J. Kurths,
Chaos: An Interdisciplinary Journal of Nonlinear Science
31, 113134 (2021).
[20] Y. Moreno and A. F. Pacheco, EPL (Europhysics Letters)
68, 603 (2004).
[21] W.-X. Wang, J. Ren, Y.-C. Lai, and B. Li, Chaos: An
Interdisciplinary Journal of Nonlinear Science 22, 033131
10
(2012).
[22] M. Wiedermann, J. F. Donges, J. Heitzig, and J. Kurths,
EPL (Europhysics Letters) 102, 28007 (2013).
[23] J. G. Restrepo, E. Ott, and B. R. Hunt, Phys. Rev. E
69, 066215 (2004).
[24] J. Sun, E. M. Bollt, and T. Nishikawa, Europhys. Lett.
85, 60011 (2009).
[25] F. Sorrentino and M. Porfiri, Europhys. Lett. 93, 50002
(2011).
[26] S. Acharyya and R. Amritkar, EPL (Europhysics Letters)
99, 40005 (2012).
[27] S. Acharyya and R. Amritkar, Physical Review E 92,
052902 (2015).
[28] Y. Sugitani, Y. Zhang, and A. E. Motter, Physical Re-
view Letters 126, 164101 (2021).
[29] L. Tang, K. Smith, K. Daley, and I. Belykh, Physical
Review E 106, 024214 (2022).
[30] J. G. Barajas-Ramı́rez, A. Franco-López, and H. G.
González-Hernández, Applied Mathematics and Compu-
tation 395, 125877 (2021).
[31] O. Rössler, Physics Letters A 57, 397 (1976).
[32] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and
D. U. Hwang, Phys. Rep. 424, 175 (2006).
[33] M. Barahona and L. Pecora, Phys. Rev. Lett. 89, 054101
(2002).
[34] B. Bamieh, arXiv preprint arXiv:2002.05001 (2020).
[35] L. Huang, Q. Chen, Y.-C. Lai, and L. M. Pecora, Phys.
Rev. E 80, 036204 (2009).
[36] A. M. Hagerstrom, T. E. Murphy, R. Roy, P. Hövel,
I. Omelchenko, and E. Schöll, Nature Physics 8, 658
(2012).
[37] L. M. Pecora, F. Sorrentino, A. M. Hagerstrom, T. E.
Murphy, and R. Roy, Nature Communications 5 (2014).
[38] C. R. Williams, T. E. Murphy, R. Roy, F. Sorrentino,
T. Dahms, and E. Schöll, Physical review letters 110,
064104 (2013).
[39] F. Sorrentino and M. Porfiri, EPL (Europhysics Letters)
93, 50002 (2011).
[40] S. Panahi and F. Sorrentino, Physical Review E 104,
054314 (2021).
[41] S. Boccaletti, D.-U. Hwang, M. Chavez, A. Amann,
J. Kurths, and L. M. Pecora, Physical Review E 74,
016102 (2006).
[42] F. Sorrentino, New J. Phys. 14, 033035 (2012).
[43] D. Irving and F. Sorrentino, Phys. Rev. E 86, 056102
(2012).