Geometric unfolding of synchronization

dynamics on networks
Cite as: Chaos 31, 061105 (2021); https://doi.org/10.1063/5.0053837
Submitted: 12 April 2021 . Accepted: 17 May 2021 . Published Online: 07 June 2021

Lluís Arola-Fernández, Per Sebastian Skardal, and Alex Arenas

Chaos 31, 061105 (2021); https://doi.org/10.1063/5.0053837 31, 061105

© 2021 Author(s).
 Chaos ARTICLE scitation.org/journal/cha

Geometric unfolding of synchronization dynamics

on networks
Cite as: Chaos 31, 061105 (2021); doi: 10.1063/5.0053837
Submitted: 12 April 2021 · Accepted: 17 May 2021 ·
Published Online: 7 June 2021 View Online Export Citation CrossMark

Lluís Arola-Fernández,1 Per Sebastian Skardal,2 and Alex Arenas1,a)

AFFILIATIONS
1
Departament d’Enginyeria Informàtica i Matemàtiques, Universitat Rovira i Virgili, 43007 Tarragona, Spain
2
Department of Mathematics, Trinity College, Hartford, Connecticut 06106, USA

a)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
We study the synchronized state in a population of network-coupled, heterogeneous oscillators. In particular, we show that the steady-state
solution of the linearized dynamics may be written as a geometric series whose subsequent terms represent different spatial scales of the
network. Specifically, each additional term incorporates contributions from wider network neighborhoods. We prove that this geometric
expansion converges for arbitrary frequency distributions and for both undirected and directed networks provided that the adjacency matrix
is primitive. We also show that the error in the truncated series grows geometrically with the second largest eigenvalue of the normalized
adjacency matrix, analogously to the rate of convergence to the stationary distribution of a random walk. Last, we derive a local approximation
for the synchronized state by truncating the spatial series, at the first neighborhood term, to illustrate the practical advantages of our approach.
Published under an exclusive license by AIP Publishing. https://doi.org/10.1063/5.0053837

Synchronization in ensembles of network-coupled heterogeneous I. INTRODUCTION

oscillators plays a vital role in a number of natural and engi- The study of synchronization dynamics on networks of hetero-
neered phenomena, ranging from cell cycles to robust power geneous oscillators has shed light on a wide range of natural and
systems.1,2 In a wide range of practical applications of such sys- engineered phenomena and has provided significant insights into
tems, including synchronization optimization,3–5 control,6,7 and structural properties of complex networks.10–12 A particularly useful
global entrainment,8 the particular structure of a fully synchro- model for these purposes uses phase oscillators that evolve according
nized state9 or a target synchronized state is critical for achieving to the following equations of motion:
the task at hand. In this paper, we study the linearized synchro-
N
nized state in networks of heterogeneous oscillators and show X
that it can be written as a particular geometric series where sub- θ̇i = ωi + σ aij H(θj − θi ), (1)
j=1
sequent series terms describe contributions from larger network
neighborhoods. This spatial description of the state provides a where θi is the phase of oscillator i, i = 1, . . . , N, ωi is its natu-
contrast to, e.g., spectral methods that rely on global information. ral frequency, σ is the global coupling strength, and H(θj − θi ) is
In particular, truncation of the geometric series at a fixed order is a 2π -periodic non-linear coupling function of the state differences
equivalent to building the synchronized state using only local, i.e., of pairs of connected oscillators. The network connectivity and the
decentralized, information. We provide a rigorous proof of our relative intensity of the interactions are captured by the entries aij
main result, show that error in the truncated series decays geo- of the N × N, real-valued adjacency matrix A that, in general, can
metrically with the second largest eigenvalue of the normalized be undirected or directed, binary or weighted. While the particular
adjacency matrix, and illustrate the practical advantages of this choice of H(θ ) depends on the application at hand, notable choices
approach. H(θ ) = sin(θ ) and H(θ ) = sin(θ − α), where α ∈ [−π/2, π/2],

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-1

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

yield the network Kuramoto and Sakaguchi–Kuramoto models, failures20 and to rank specific nodes and edges according to its
respectively.13 dynamical stability.2,21 Furthermore, the object L† provides, by its
A particularly suitable framework useful for analyzing the own, a lot of insight on the network properties in diffusive pro-
interplay between structure and dynamics, described here by the cesses in general.22 In noisy dynamics, the coupling of L† with the
network A and the natural frequencies ω, respectively, emerges covariance matrix of the input noise determines the noise in the out-
when a strongly synchronized, phase-locked state becomes attain- put signal such that the network can be optimized to perform as a
able. Following Ref. 3, we assume that |H(0)|  1 so that for noise-canceling filter.23 It plays a central role in the theory of electric
a strongly synchronized state where |θj − θi |  1, we have that circuits, where it is used to obtain the resistance matrix that provides
H(θj − θi ) ≈ H(0) + H0 (0)(θj − θi ) [we note here that for the classi- a notion of diffusion distance between nodes and to determine the
cal Kuramoto model, we have H(0) = 0 and H0 (0) = 1]. By defining best spreader of the network,22 as well as other node or edge cen-
the effective frequency ω̂i = ωi + σ H(0)kin i and entering a suitable trality measures, such as the random-walk betweenness.24 From a
rotating frame, the synchronized state is then described by the fixed mathematical point of view, L† corresponds to the discrete Green’s
point satisfying θ̇ = 0. After linearizing, the synchronized state then function of the network25 since it quantifies how a current injected
satisfies the following equation: in a node propagates to any other node of the network.26
All the previous results are expressed in the literature in terms
ω̂ = σ H0 (0)Lθ , (2) of the spectra of L. The exact computation of L† through the spec-
where L = D − A is the Laplacian of the network and D is a diag- tral or singular value decomposition can be computationally costly,
onal matrix with the in-degrees or in-strengths of the nodes, i.e., and the optimization and analysis of the interplay between structure
Dii = ki = Nj=1 aij . Since the matrix L has zero row sum, it has
P and dynamics rely on numerical schemes that are often treated as
black-boxes. Therefore, the networks studied in most of these cases
a trivial eigenvalue λ1 = 0 with a constant associated eigenvector
are rather small, and the interesting properties that emerge in the
v1 ∝ 1, and therefore, it is singular and not invertible.3,14 Moreover,
optimal coupling of topology and dynamics can only be observed
this spectral property reveals an important physical characteristic of
a posteriori, from the outcome of numerical methods. Also, the
the system, namely, that the dynamics are invariant to a constant
computations and interpretations of the results require global infor-
shift to the phases, i.e., translation along the synchronization man-
mation of the network that is not available at the level of the nodes,
ifold defined as the span of the trivial eigenvector v1 . Thus, while
which usually operate in a decentralized manner.11
solutions to Eq. (2) are not unique, the one that minimizes the norm
Motivated by the previous limitations, we derive a geometric
||θ || is likely the most useful and is given by
expansion of Eq. (3) in terms of increasingly further neighborhoods
L† ω̂ of the nodes. This approach allows for a decentralized, fast, and
θ∗ = , (3) accurate computation of L† ω, up to a desired degree of accuracy
σ H0 (0)
(or amount of available information) and provides new analyti-
where L† is the Moore–Penrose pseudo-inverse of the Laplacian cal insights that allow for a deeper understanding of the interplay
matrix.15 Importantly, to write down the exact pseudo-inverse, we between topology and dynamics in network synchronization.
require a full spectral decomposition of the Laplacian matrix (i.e., The remainder of this paper is organized as follows. In Sec. II,
global information of the network). For a possibly directed net- we present our main theoretical results, namely, geometric series
work, the pseudo-inverse can be computed via the singular value expansions for computing Eq. (3), along with convergence proofs.
decomposition of L as4 In Sec. III, we use local approximations, generated from our theoret-
ical results to identify link removals that improve synchronization of
N
X v n uT n networks. In Sec. IV, we conclude with a discussion of our results.
L† = , (4)
n=2
µn
II. MAIN RESULTS
where 0 < µ2 ≤ · · · ≤ µN are the N − 1 singular values of L and
{vn } and {un } are the set of right and left singular vectors. Note that We begin by noting that rescaling the natural frequencies in
for the particular case of undirected networks, the singular values Eq. (2) enables us to set σ and H0 (0) both to one. To further simplify
are given by the eigenvalues of L and the left and right singular vec- notation, we drop the hat-notation, in which case Eq. (2) reduces
tors are given by the eigenvectors of L; therefore, L† is defined by its to3,6
eigenvalue decomposition. ω = Lθ . (5)
Equation (3) gives the linearized synchronized state and, there- To solve Eq. (5) without the Moore–Penrose pseudo-inverse, we
fore, is critical to understanding the interplay between structure and focus our attention more directly on L. First, using L = D − A, we
synchronization dynamics in tasks involving the fully synchronized write
state, and for this reason, there is a myriad variety of phenomena
L = D(I − D−1 A). (6)
and applications involving it. Equation (3) emerges in the derivation
of objective functions used to optimize the degree of synchroniza- While D is invertible (assuming that the network is connected and
tion in the system3–5,16 and in problems related to the control of thus each node has some positive degree), (I − D−1 A) is not. This
desired synchronized states,6 critical points, and in network infer- can be seen by noting that D−1 A is a stochastic matrix and, therefore,
ence techniques.17–19 Also, it is used to estimate analytically the local has a leading eigenvalue λ1 = 1. Then, I − D−1 A has a zero eigen-
and global susceptibility of a power-grid system against structural value, making it singular. However, replacing D−1 A with matrix X

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-2

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

that yields (I − X) invertible, we have that Inserting Eq. (12) into Eq. (11) yields
[D(I − X)]−1 = (I − X)−1 D−1 . (7) 1
D 2 φ m = Bm x
(13)
Moreover, the matrix (I − X)−1 may be expanded in the geometric = α1 λm
1 m m
1 k + α2 λ2 v2 + · · · + αn λn vn .
2
series,
X∞ Note now that for terms j = 2, . . . , N, λm
j decays geometrically, while
(I − X)−1 = Xm . (8) λm1 = 1. However, we now show that the coefficient α1 must be
m=0 zero. To see this, recall that the natural frequency vector has mean
The issue now arises that X cannot be replaced by D−1 A, or more zero, or in other words, ω is orthogonal to the constant vector 1,
specifically, we have that i.e., h1, ωi = 0. This is equivalent to hD1/2 1, D−1/2 ωi = 0, or more
simply, hk1/2 , xi = 0, which is precisely α1 . Thus, we have that
∞
−1 m
X 1
[D(I − D−1 A)] 6= (D−1 A) D−1 , (9) D 2 φ m = α2 λm m
2 v2 + · · · + αn λn vn . (14)
m=0
The convergence of the right-hand-side of Eq. P (10) now follows
namely, on the left hand side, the inverse is ill-posed, and this is quite easily: since each of the finitely-many series ∞ m
m=0 αj λj vj con-
reflected by the fact that the series on the right-hand side diverges. verges to αj vj /(1 − λj ) for j = 2, . . . , N, we have that the full series
However, this does not rule out the possibility of the right-hand side converges to
converging when it is applied to a vector of a particular form. In fact, ∞
under the relatively mild conditions of the network having a primi- φ = D−1/2
X
D1/2 φm
tive adjacency matrix, when the series is applied to an appropriately m=0
shifted frequency vector ω, the right-hand-side does converge and
∞ X
N
yields a solution to Eq. (5), which leads to the formulation of our X
first main result for undirected networks. = D−1/2 αj λm
j vj (15)
m=0 j=2
Theorem 1 Convergence of the geometric series for undi-
rected networks. Consider an undirected network with primitive XN
αj vj
adjacency matrix A and a frequency vector ω with zero mean, i.e., = D−1/2 ,
hωi = 0. Then, the infinite series j=2
1 − λj
∞ which concludes the proof. 
m
X
φ= (D−1 A) D−1 ω (10) A subtle issue that arises with the result given in Theorem 1
m=0 is that, while the infinite series applied to the zero-mean vector ω
converges. converges, it does not necessarily converge to the minimum norm
Proof. We begin by denoting the symmetric normalized adja- solution given by Eq. (3), which turns out to the zero-mean solution
cency matrix as B = D−1/2 AD−1/2 . Since A is symmetric, so is B, recovered by the Moore–Penrose pseudo-inverse approach. This
and therefore, its normalized eigenvectors {vj }Nj=1 form an orthonor- can be seen by noting that at initial truncation of the geometric
mal basis for RN . Moreover, since A is primitive, that is, there exists series, we have that hφ 0 i = h ωk i, which is not zero. However, this
some integer M > 0 such that AM is strictly positive (note that this can be easily fixed by simply applying a constant shift to the result-
is equivalent to A being both irreducible and aperiodic), so is B. The ing vector to ensure that it is orthogonal to the constant eigenvector
Perron–Frobenius theorem27 then implies that B has a single largest v1 and the solution converges to the one with minimal norm. In fact,
eigenvalue λ1 that is real and larger in magnitude than all other this correction leads us to a generalization of our main result, which
eigenvalues, i.e., λ1 > |λj | for j = 2, . . . , N. Moreover, since B is nor- importantly applies to both undirected and directed networks.
malized, we have λ1 = 1 and |λj | < 1 for j = 2, . . . , N. Finally, the Theorem 2 Convergence of the geometric series: General case.
leading eigenvector associated with the leading eigenvalue λ1 = 1 Consider a network with primitive adjacency matrix A. Then, the
has entries that are proportional to the square root of the degrees of infinite series
the respective nodes, i.e., v1 ∝ k1/2 . X∞


Next, it is useful to define φ= φ m − hφ m i , (16)
m=0
1 1
φ m = D− 2 Bm D− 2 ω (11) where
so that the right-hand-side of Eq. (10) is given by ∞ m
P
m=0 φ m . Defining φ m = (D−1 A) D−1 ω, (17)
1
the vector x = D− 2 ω, we now expand x via the orthonormal basis of converges.
eigenvectors of B, namely, Proof. The complete proof is shown in the Appendix. 
1
x = α1 k 2 + α2 v2 + · · · + αN vN , (12) Before proceeding to other practical questions and examples,
we remark that Theorem 2 includes no zero-mean condition on the
where αi = hxi , vi i are the coefficients given by projections of X onto frequency vector nor is it restricted to undirected networks; thus,
the different eigenvector directions, and we assume that the eigen- its potential usage is wider than Theorem 1. In Fig. (1), we illus-
vector v1 = ck1/2 , where c is a constant to ensure normalization. trate the utility of the more general case given by Eqs. (16) and (17)

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-3

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

i.e., each node must be reachable from each node in a manner that
is eventually well-mixed, hence the need for the network adjacency
matrix to be irreducible and aperiodic, i.e., primitive.30
Turning now to a more computationally practical question,
we investigate the rate of convergence of this expansion. Assum-
ing for simplicity, the case of an undirected network (so we may
use Theorem P 1), we−1consider the M-order approximation defined
m −1
as θ (M) = M m=0 (D A) D ω. Using Eq. (13), the error is given by
∞
1 X
D 2 (θ ∗ − θ (M) ) = (α2 λm m
2 v2 + · · · + αn λn vn )
m=M+1
α2 v2 M+1 αN vN M+1
= λ + ··· + λ . (18)
1 − λ2 2 1 − λN N
As the order M of the approximation increases, the dominant term
in Eq. (18) is that which corresponds to the second largest (in mag-
FIG. 1. Scatter-plot of the exact equation (3) vs the truncated equation (16)
approximation of the stationary phases, for different neighborhood orders (M), nitude) eigenvalue λ2 of the normalized adjacency matrix B. (Recall
in a fixed Erdös–Rényi network of size N = 500, mean degree hki = 20 with a that the largest of the eigenvalue of B is λ1 = 1.) Save for the unlikely
normal distribution of frequencies N(0, 1) for three different truncation orders. scenario where x = D−1/2 ω is exactly orthogonal to v2 , in which
case α2 vanishes, then for large enough M, the mean square error
will scale geometrically with the magnitude of this second largest
eigenvalue, i.e.,
when applied to a random network with random allocated frequen-
cies. We truncate the expansion at different neighborhood orders M ||θ ∗ − θ (M) || ∼ |λ2 |M+1 . (19)
and compare the approximated solution against the exact one given
by Eq. (3), showing that, for this particular random configuration,
the approximation is accurate even for small M.
Investigating our results described in Eq. (10) and Eqs. (16)
and (17) more deeply, we see that the geometric expansion expresses
the solution of the linear system as a sum of contributions of terms
∼ ω/k coming from increasingly further neighborhoods of the
nodes. That is, the mth order term consists of terms ωj /kj cor-
responding to oscillators located precisely m links removed from
a given oscillator. Thus, the series is simply the Taylor expansion
of the solution expanded about each oscillator with higher orders
corresponding to larger network neighborhoods. It constructs the
exact solution by adding infinitely many incremental pieces of local
information in a polynomial basis that is not necessarily orthogo-
nal. Interestingly, this differs from the spectral decomposition used
for L† in Eq. (3). In the latter, the solution is constructed by adding
N − 1 pieces of global information [i.e., rank-one matrices for each
non-zero eigenvalue and its associated eigenvector in Eq. (4)], anal-
ogously to a Fourier expansion that expresses the solution on the
basis of orthogonal eigenmodes with its associated eigenfrequencies
(which carry global information on the original function).28,29
Furthermore, there is a clear connection between the applica-
bility of this method and Markov chains and random walks.30 In
particular, for a Markov chain to have a unique, globally attract-
FIG. 2. Error of the approximation depending on the truncation order for several
ing stationary state, it is necessary and sufficient to have a primitive interpolating network models, with fixed size N = 1000. We study the effect of
transition matrix (or, equivalently, irreducible and aperiodic). For a density (varying the average degree), and fixing the mean degree at hki = 10,
network with adjacency matrix A, the transition matrix for a random we also study the effect of modularity (interpolating from a random network to
walk is given by D−1 A (or AD−1 , depending on the definition of the a 2-module configuration), and clustering (using the algorithm31 to generate SF
random walk). Here, D−1 A appears in the geometric series because networks with controllable clustering and the small-world model to interpolate
the solution is expressed in terms of “in-neighborhoods.” Since this between a ring and a random network).10 In all the scenarios, we observe that,
for sufficiently high truncation order M, the error is dominated by λ2 , the sec-
“spatial” expansion unfolds in terms of in-neighborhoods of radius ond largest eigenvalue of the normalized adjacency matrix B. λ2 increases with
zero, one, two, three, etc., convergence of the expansion implies that sparsity, modularity, and clustering as expected.
eventually, these neighborhoods must include the whole network,

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-4

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

Therefore, the smaller |λ2 | is, the quicker the approximation will k are the strengths (or degrees of the nodes). The inverse second
converge. On the other hand, convergence will be slower for sparse moment is larger for heterogeneous networks; thus, homogeneity
networks with either strong modularity or clustering as well as a promotes synchronization in the linearized regime.35 This is oppo-
strong bipartite structure, in which cases λ2 tends to be close to 1 site to what occurs in the critical threshold, where Kc ∼ hki/hk2 i,
and −1, respectively,14,30 and a larger number of terms are needed and homogeneity delays the critical threshold.11 The interpretabil-
to obtain a desired level of accuracy. In Fig. 2, we plot the error ity of these effects emerges naturally from the local description of
of the truncated approximation depending on the truncation order synchrony, and although the particularities of each phenomenon
for different values of the network density, modularity, clustering may require further analysis, the geometric unfolding allows us to
coefficient, and small-worldness and compare against the theoreti- unveil the underlying mechanistic rules that control the interplay of
cal scaling predicted by Eq. (19) observing an excellent agreement structure and dynamics in synchronization.
for sufficiently large order M in all the cases. Furthermore, the local approximation in Eq. (21) is useful
in estimating the impact that removing or adding a link has on
III. LOCAL APPROXIMATION OF SYNCHRONIZATION the degree of synchronization. In Ref. 5, an approximation for the
change in the order parameter r is given for undirected networks
Next we generate a local approximation from the geomet- using a perturbation analysis to approximate changes in the Lapla-
ric expansion presented above to describe the degree of synchro- cian eigenvalues and in turn the change to the order parameter.
nization of P a system given by the Kuramoto order parameter,32 Using the geometric expansion, this problem can be addressed using
reiψ = N−1 j eiθj . First, after linearizing about a strongly synchro- only local information and directed networks may also be treated.
nized state where all |θi |  1 (note that an appropriate shift in initial First, we denote the change in the order parameter r as a result of
conditions allows us to set the mean phase ψ = 0), one obtains3 removing or adding a directed link coming from q to p as 1rpq .
r ≈ 1 − 2σ12 N ||θ ∗ ||2 . Note that the degree of synchronization Assuming that degrees are sufficiently large and no frequency or
increases, i.e., tends toward one, as the dispersion in the phases is degree correlations exist, it is straightforward to obtain from Eq. (21)
reduced. We approximate the solution of the phases by truncating that
Eq. (16) at the first-order term, which will be accurate for networks 1
 
ωp ωp ωq

with |λj |  1 for j = 2, . . . , N and no strong dynamical correla- 1rpq ≈ ∓ − , (22)
Nkp kp kp kq
tions among nodes. We also neglect the contribution of the shift to
the mean in Eq. (16), assuming that the distribution of ω/k is well where the ∓ sign accounts for removal/addition of links. First, the
centered around zero. Under these assumptions, we have that absolute value |1rpq | will be larger if node p, i.e., the node with an
incoming link from node q, has a large value of the ratio ωp /kp (i.e.,
ωi D ω E(1)
θi ≈ + , (20) large frequency and/or low degree) and the difference of phases (or
ki k i ratios) between nodes p and q is large. Thus, the most important
PN ωj links in terms of their contribution to the degree of synchronization
with h ωk i(1)
i = ki
1
j=1 aij kj being the average contribution of the
are those that connect nodes with large rations |ω/k|. Subsequently,
first neighbors arriving at node i. Now, we can then directly write
connecting very similar nodes and those that point toward nodes
the order parameter as
with small |ω/k| contribute less to the degree of synchronization.
N 
1 X ωi D ω E(1) 2
 Going one step further, using Eq. (22), we can predict which
r≈1− 2 + . (21) directed links will produce the counter-intuitive effect of increas-
2σ N i=1 ki k i
ing (decreasing) synchrony after its removal (addition). This effect,
The local unfolding of synchronization dynamics from the known as the Braess paradox in the context of road traffic,14 has been
geometric expansion that we use to write Eq. (21) allows us to studied for oscillatory networks,2,21,36 although the identification of
gain an analytical insight into the interplay between topology these particular links relied on numerical schemes or expressions in
and dynamics that improves synchronization as well as under- terms of the spectral decomposition of L. Also, while these works
stand several features that to date have only been investigated study if the perturbations break the stability of the current state,9
numerically,2–5,21,33 for instance, (i) linear degree–frequency corre- here we assume that the perturbations in the links drive the system
lations reduce the absolute value of the phases at a given trunca- toward a new steady-state. In our formalism, we can directly impose
tion order, which tend to reduce the overall dispersion (increasing 1rpq > 0 for a link removal to obtain the condition
synchronization),4,33,34 (ii) negative frequency–frequency correla- 
ωp ωp ωq

tions between connected neighbors tend to make the first-order − < 0. (23)
kp kp kq
term of opposite sign (but smaller) to the local term, and this reduces
the dispersion (increasing synchronization),3,4 (iii) synchronization The condition in Eq. (23) describes two regions of the (ωp /kp , ωq /kq)-
is reduced when the network becomes sparser. Even if the loss of plane, namely, the wedges ωq /kq > ωp /kp for ωp > 0 and ωq /kq
links is compensated by an increase of the weights, a sparse network < ωp /kp for ωp < 0. In terms of their area, these wedges describe
will have more dispersion on the phases with respect to a denser a quarter of the (ωp /kp , ωq /kq ) space. In Fig. (3), we plot the result-
one because in the latter, the higher-order terms sum over more ing change 1rpq for each possible link removal in an Erdös–Rényi
neighbors and thus they vanish out more rapidly. This provides a network of size N = 500 with mean degree hki = 50, color-coding
clear mechanistic interpretation of weight localization phenomena,35 the change so that positive (negative) changes are shaded more red
and (iv) if frequencies are randomly allocated, r ∼ 1 − hk−2 i, where (blue). We note that the positive changes fit well within the wedges

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-5

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

dynamics on networks by deriving a local approximation for the

degree of synchronization in the system. Furthermore, this approx-
imation correctly identifies that directed links may be removed to
improve synchronization properties, a phenomenon known as the
Braess paradox, as it captures the local interplay between network
structure and dynamics.
Our results are of both theoretical and practical relevance for
the study of synchronization on networks. The theory unfolds the
stationary dynamics of the system in the real (geometric) domain of
the nodes, and it provides an alternative understanding of the sys-
tem from its local construction unlike all previous results found in
the literature that solve many specific problems in terms of global
spectral information. Also, we can use the theory to tackle important
practical problems such as network inference18 and optimization,3–5
critical threshold estimation,8 control of synchronized states by local
weight tuning or structure modifications,6,38 and to predict how fluc-
tuations in the weights or the frequencies propagate to the macro-
scopic synchronization,35,39 just to name a few. Furthermore, our
FIG. 3. Distribution of removed directed links (p, q) in the phase-space theory can be applied beyond the synchronization problem since it
(ωp /kp , ωq /kq ) for an Erdös–Rényi network with N = 500 and hki = 50 and finds utility in any forced diffusion system on networks that reduces
gaussian g(ω) with N(0, 1). The color indicates whether the removed link to solving the linear system of Eq. (5). The challenge to unfold the
decreases (blue) or increases (red) the degree of synchrony, calculated exactly fully non-linear dynamics and the stability of the solutions of Eq. (1)
from Eq. (3). Straight lines bound the region predicted by Eq. (23) in the local
approximation.
within this geometric framework remains open for further research.

ACKNOWLEDGMENTS
predicted by our local theory, which are plotted in dashed black L.A.-F. and A.A. acknowledge the Spanish MINECO (Grant
lines. In fact, approximately a quarter of directed links have the No. PGC2018-094754-B-C2). A.A. also acknowledges the Generali-
potential to increase synchronization after its removal as expected. tat de Catalunya (Grant No. 2017SGR-896), the Universitat Rovira i
While this phenomenon has been investigated in the context of Virgili (Grant No. 2017PFR-URV-B2-41), and the ICREA Academia
identical oscillators,11,37 here, the frequencies of the oscillators play and the James S. McDonnell Foundation (Grant No. 220020325).
a critical role in determining which directed links are harmful or
redundant, and we have shown that the local approximation is
sufficient to capture this phenomenon. APPENDIX: PROOF OF THEOREM 2
It is important to remark that, while we can gain much insight In this appendix, we now present the proof of Theorem 2.
from the truncation of the expansion at the first-order term, one can Proof. Rather than making use of the symmetric normalized
obtain more accurate approximations by going to higher orders. For adjacency matrix, as in the proof of Theorem 1, here, we use the
instance, the truncation at second order improves Eq. (20) by adding more classically stochastic matrix D−1 A. However, this matrix shares
PN aij P ωk
the term h ωk i(2)
i = ki
1
j=1 kj k ajk kk , i.e., the weighted average similar properties with its symmetric counterpart, namely, because
over ratios of frequency-degree two hops away from the nodes. A is primitive, then so is D−1 A, and the Perron–Frobenius theorem
Similarly, further approximations can be obtained up to a desired guarantees similar eigenvalue properties, namely, there is a single
degree of accuracy depending on the available computational time largest eigenvalue λ1 that is real and larger in magnitude than all
or amount of accessible information from the system. other eigenvalues, i.e., λ1 > |λj | for j = 2, . . . , N. Also, since D−1 A is
stochastic, λ1 = 1 and |λj | < 1 for j = 2, . . . , N. On the other hand,
the leading eigenvector is now given by the constant vector v1 ∝ 1.
IV. CONCLUSIONS
Importantly, since D−1 A is not symmetric (even if A is), the eigen-
In this paper, we introduced a geometric expansion that vectors are not orthogonal to one another and cannot be used to
expresses the exact solution of the linearized synchronization prob- form an orthonormal basis for RN . Nonetheless, we may use these
lem on a complex network as a Taylor series in contributions from eigenvalues as a (non-orthogonal) basis for RN and uniquely expand
increasingly further network neighborhoods. We proved that this the vector x = D−1 ω using this basis, specifically,
series converges under relatively mild conditions, namely, that the
network has a primitive adjacency matrix, and this method can N
X
be used for both directed and undirected networks. Moreover, the x= αj vj . (A1)
approximation error from using a truncated series scales geometri- j=1
cally with the second largest eigenvalue of the normalized adjacency
matrix. We then applied these theoretical results to gain analytical We then look at each eigenmode j = 1, . . . , N of the term φ m ,
m PN (j)
and mechanistic insights on the optimal interplay of structure and namely, φ (j) −1
m = αj (D A) vj so that φ m = j=1 φ M . In terms of the

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-6

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

6
full expression for φ [given in Eq. (16)], we then have that P. S. Skardal and A. Arenas, “Control of coupled oscillator networks with
application to microgrid technologies,” Sci. Adv. 1, e1500339 (2015).
∞ X
X N 7
P. S. Skardal and A. Arenas, “On controlling networks of limit-cycle oscillators,”
φ (j) (j)


φ= m − hφ m i Chaos 26, 094812 (2016).
8
m=0 j=1 F. Dörfler, M. Chertkov, and F. Bullo, “Synchronization in complex oscillator
networks and smart grids,” Proc. Natl. Acad. Sci. U.S.A. 110, 2005–2010 (2013).
∞
N X 9
X L. M. Pecora and T. L. Carroll, “Master stability functions for synchronized
φ (j) (j)


= m − hφ m i . (A2) coupled systems,” Phys. Rev. Lett. 80, 2109–2112 (1998).
10
j=1 m=0 D. J. Watts and S. H. Strogatz, “Collective dynamics of ’small-world’ networks,”
Nature 393, 440–442 (1998).
We now treat the contribution of each eigenmode separately. We 11
A. Arenas, A. Díaz-Guilera, J. Kurths, Y. Moreno, and C. Zhou, “Synchroniza-
begin with the eigenmodes j ≥ 2 for which |λj | < 1. First, we have tion in complex networks,” Phys. Rep. 469, 93–153 (2008).
12
that A. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization: A Universal Con-
m
cept in Nonlinear Sciences (Cambridge University Press, 2003), Vol. 12.
−1 m
φ (j)
m = αj (D A) vj = αj λj vj (A3) 13
Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Courier Corpora-
tion, 2003).
and 14
M. Newman, Networks: An Introduction (Oxford University Press, Inc., New
m m York, 2010).
hφ (j)
m i = hαj λj vj i = αj λj hvj i; (A4) 15
A. Ben-Israel and T. Greville, Generalized Inverses: Theory and Applications
(Wiley, New York, 1974).
therefore, together, we have that 16
P. S. Skardal, D. Taylor, J. Sun, and A. Arenas, “Collective frequency varia-
∞
X ∞ tion in network synchronization and reverse pagerank,” Phys. Rev. E 93, 042314

X (2016).
φ (j) (j)
αj λm


m − hφ m i = j vj − hvj i 17
A. Arenas, A. Díaz-Guilera, and C. J. Pérez-Vicente, “Synchronization reveals
m=0 m=0 (A5) topological scales in complex networks,” Phys. Rev. Lett. 96, 114102 (2006).
αj
18
M. Timme, “Revealing network connectivity from response dynamics,” Phys.
= vj − hvj i ;
1 − λj Rev. Lett. 98, 224101 (2007).
19
M. De Domenico, “Diffusion geometry unravels the emergence of functional
i.e., each component converges for j ≥ 2. clusters in collective phenomena,” Phys. Rev. Lett. 118, 168301 (2017).
20
To complete the proof, we now show that the j = 1 eigenmode, D. Manik, M. Rohden, X. Ronellenfitsch, H. Zhang, S. Hallerberg, D. Witthaut,
and M. Timme, “Network susceptibilities: Theory and applications,” Phys. Rev. E
for which λ1 = 1, converges. In fact, it turns out that this component
95, 012319 (2017).
has no contribution due to the shift of the mean. As in Eqs. (A3) 21
T. Coletta and P. Jacquod, “Linear stability and the Braess paradox in
and (A4), we have that coupled-oscillator networks and electric power grids,” Phys. Rev. E 93, 032222
−1 m m
(2016).
φ (1)
m = α1 (D A) v1 = α1 λ1 v1 = α1 v1 (A6) 22
P. Van Mieghem, K. Devriendt, and H. Cetinay, “Pseudoinverse of the Laplacian
and best spreader node in a network,” Phys. Rev. E 96, 032311 (2017).
and 23
H. Ronellenfitsch, J. Dunkel, and M. Wilczek, “Optimal noise-canceling net-
hφ (1) m m works,” Phys. Rev. Lett. 121, 208301 (2018).
m i = hα1 λ1 v1 i = α1 λ1 hv1 i = α1 hv1 i; (A7) 24
M. E. J. Newman, “A measure of betweenness centrality based on random
therefore, walks,” Soc. Netw. 27, 39–54 (2005).
25
F. Chung and S. T. Yau, “Discrete Green’s functions,” J. Comb. Theory Ser. A
φ (1) (1)
m − hφm i = α1 (v1 − hv1 i) , (A8) 91, 191–214 (2000).
26
E. Estrada and N. Hatano, “Communicability in complex networks,” Phys.
but since v1 ∝ 1, i.e., it is constant, we have that v1 = hv1 i and each Rev. E 77, 036111 (2008).
27
term φ1(m) − hφ1(m) i vanishes, which completes the proof.  C. R. MacCluer, “The many proofs and applications of Perron’s theorem,”
SIAM Rev. 42, 487–498 (2000).
28
P. N. McGraw and M. Menzinger, “Analysis of nonlinear synchronization
DATA AVAILABILITY dynamics of oscillator networks by Laplacian spectral methods,” Phys. Rev. E 75,
027104 (2007).
The data that support the findings of this study are available 29
D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The
from the corresponding author upon reasonable request. emerging field of signal processing on graphs: Extending high-dimensional data
analysis to networks and other irregular domains,” IEEE Signal Process. Mag. 30,
83–98 (2013).
REFERENCES 30
N. Masuda, M. A. Porter, and R. Lambiotte, “Random walks and diffusion on
1
A. Prindle, P. Samayoa, I. Razinkov, T. Danino, L. S. Tsimring, and J. Hasty, “A networks,” Phys. Rep. 716-717, 1–58 (2017).
31
sensing array of radically coupled genetic ‘biopixels’,” Nature 481, 39–44 (2012). P. Holme and B. Kim, “Growing scale-free networks with tunable clustering,”
2
D. Witthaut and M. Timme, “Braesss paradox in oscillator networks, desynchro- Phys. Rev. E: Stat., Nonlinear, Soft Matter Phys. 65, 026107 (2002).
32
nization and power outage,” New J. Phys. 14, 083036 (2012). Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Dover Publica-
3
P. S. Skardal, D. Taylor, and J. Sun, “Optimal synchronization of complex tions, 2003).
33
networks,” Phys. Rev. Lett. 113, 144101 (2014). P. S. Skardal, J. Sun, D. Taylor, and J. G. Restrepo, “Effects of degree-frequency
4
P. S. Skardal, D. Taylor, and J. Sun, “Optimal synchronization of directed correlations on network synchronization: Universality and full phase-locking,”
complex networks,” Chaos 26, 094807 (2016). Europhys. Lett. 101, 20001 (2013).
5 34
D. Taylor, P. S. Skardal, and J. Sun, “Synchronization of heterogeneous oscillators J. Gómez-Gardeñes, S. Gómez, A. Arenas, and Y. Moreno, “Explosive syn-
under network modifications: Perturbation and optimization of the synchrony chronization transitions in scale-free networks,” Phys. Rev. Lett. 106, 128701
alignment function,” SIAM J. Appl. Math. 76, 1984–2008 (2016). (2011).

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-7

Published under an exclusive license by AIP Publishing.
 Chaos ARTICLE scitation.org/journal/cha

35 38
P. Skardal, D. Taylor, and J. Sun, “Synchronization of network-coupled oscilla- L. Arola-Fernández, A. Díaz-Guilera, and A. Arenas, “Synchronization invari-
tors with uncertain dynamics,” SIAM J. Appl. Math. 79, 2409–2433 (2019). ance under network structural transformations,” Phys. Rev. E 97, 060301
36
A. E. Motter and M. Timme, “Antagonistic phenomena in network dynamics,” (2018).
39
Annu. Rev. Condens. Matter Phys. 9, 463–484 (2018). L. Arola-Fernández, G. Mosquera-Doñate, B. Steinegger, and A. Arenas,
37
T. Nishikawa and A. Motter, “Synchronization is optimal in nondiagonalizable “Uncertainty propagation in complex networks: From noisy links to critical
networks,” Phys. Rev. E 73, 065106 (2006). properties,” Chaos 30, 023129 (2020).

Chaos 31, 061105 (2021); doi: 10.1063/5.0053837 31, 061105-8

Published under an exclusive license by AIP Publishing.