ARTICLE

https://doi.org/10.1038/s41467-022-31733-2 OPEN

Functional control of oscillator networks

Tommaso Menara 1, Giacomo Baggio2, Dani Bassett 3,4 & Fabio Pasqualetti 5✉

Oscillatory activity is ubiquitous in natural and engineered network systems. The interaction
scheme underlying interdependent oscillatory components governs the emergence of
network-wide patterns of synchrony that regulate and enable complex functions. Yet,
1234567890():,;

understanding, and ultimately harnessing, the structure-function relationship in oscillator

networks remains an outstanding challenge of modern science. Here, we address this chal-
lenge by presenting a principled method to prescribe exact and robust functional config-
urations from local network interactions through optimal tuning of the oscillators’
parameters. To quantify the behavioral synchrony between coupled oscillators, we introduce
the notion of functional pattern, which encodes the pairwise relationships between the
oscillators’ phases. Our procedure is computationally efficient and provably correct, accounts
for constrained interaction types, and allows to concurrently assign multiple desired func-
tional patterns. Further, we derive algebraic and graph-theoretic conditions to guarantee the
feasibility and stability of target functional patterns. These conditions provide an interpretable
mapping between the structural constraints and their functional implications in oscillator
networks. As a proof of concept, we apply the proposed method to replicate empirically
recorded functional relationships from cortical oscillations in a human brain, and to redis-
tribute the active power flow in different models of electrical grids.

1 Department of Mechanical and Aerospace Engineering, University of California, San Diego, La Jolla, CA 92093, USA. 2 Department of Information Engineering,

University of Padova, Padova 35131, Italy. 3 Department of Physics & Astronomy, Department of Bioengineering, Department of Electrical & Systems Engineering,
Department of Neurology, Department of Psychiatry, University of Pennsylvania, Philadelphia, PA 19104, USA. 4 The Santa Fe Institute, Santa Fe, NM 87506, USA.
5 Department of Mechanical Engineering, University of California, Riverside, Riverside, CA 92521, USA. ✉email: [email protected]

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 1

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

T
he complex coordinated behavior of oscillatory compo- networks (where interconnections denote the number of vehicles
nents is linked to the function of many natural and tech- on a road with respect to its maximum capacity).
nological network systems1–3. For instance, distinctive To quantify the pairwise functional relations between oscilla-
network-wide patterns of synchrony4–6 determine the coordi- tory units, and inspired by the work in ref. 20, we define a local
nated motion of orbiting particle systems7, promote successful order parameter that, compared to the classical Pearson corre-
mating in populations of fireflies8, regulate the active power flow lation coefficient, does not depend on sampling time and is more
in electrical grids9, predict global climate change phenomena10, convenient when dealing with periodic phase signals (see Sup-
dictate the structural development of mother-of-pearl in plementary Information). Given a pair of phase oscillators i and j
mollusks11, and enable numerous cognitive functions in the with phase trajectories θi(t) and θj(t), we define the correlation
brain12,13. Since this rich repertoire of patterns emerges from the coefficient
properties of the underlying interaction network14, controlling
the collective configuration of interdependent units holds tre- ρij ¼ hcosðθj ðtÞ  θi ðtÞÞit ; ð2Þ
mendous potential across science and engineering15. Despite its where 〈 ⋅ 〉t denotes the average over time. A functional pattern
practical significance, a comprehensive method to enforce is formally defined as the symmetric matrix R whose i, jth entry
network-wide patterns of synchrony by intervening in the net- equals ρij. Importantly, a functional pattern explicitly encodes the
work’s structural parameters does not yet exist. pairwise, local, correlations across all of the oscillators, which are
In this work, we develop a rigorous framework that allows us to more informative than a global observable (e.g., the order
optimally control the spatial organization of the network com- parameter16,21). It is easy to see that, if two oscillators i and j
ponents and their oscillation frequencies to achieve desired pat- synchronize after a certain initial transient, ρij converges to 1 as
terns of synchrony. We abstract the rhythmic activity of a system time increases. If two oscillators i and j become phase-locked (i.e.,
as the output of a network of diffusively coupled oscillators16,17 their phase difference remains constant over time), then their
with Kuramoto dynamics. This modeling choice is motivated by correlation coefficient converges to some constant value with a
the rich dynamical repertoire and wide adoption of Kuramoto magnitude smaller than 1. If the phases of two oscillators i and j
oscillators18. Specifically, we consider an undirected network G ¼ evolve independently, then their correlation value remains small
fO; Eg of n oscillators with dynamics over time. A few questions arise naturally, which will be answered
n in this paper. Are all functional patterns achievable? Which
θ_ i ¼ ωi þ ∑ Aij sinðθj  θi Þ; ð1Þ network configurations allow for the emergence of multiple target
j¼1
functional patterns? And, if a certain functional pattern can be
where ωi 2 R and θi 2 S1 are the frequency and phase of the ith achieved, is it robust to perturbations? Surprisingly, we reveal that
oscillator, respectively, A = [Aij] is the weighted adjacency matrix controlling functional patterns can be cast as a convex optimi-
of G, and O ¼ f1; ¼ ; ng and E  O ´ O denote the oscillator zation problem, whose solution can be characterized explicitly.
and interconnection sets, respectively. In this work, we consider Figure 1 shows our framework and an example of control of
the case where the network G admits both cooperative (i.e., functional patterns for a network with 7 oscillators. In the paper,
Aij > 0) and competitive (i.e., Aij < 0)19 interactions among the we will validate our methods by replicating functional patterns
oscillators, as well as the more constrained case of purely coop- from brain recordings in an empirically reconstructed neuronal
erative interactions that arises in several real-word systems. For network, and by controlling the active power distribution in
instance, negative interactions are not physically meaningful in multiple models of the power grid.
networks of excitatory neurons, in power distribution networks While synchronization phenomena in oscillator networks have
(where the interconnection weight denotes conductance and been studied extensively (e.g., see refs. 22–26), the development of
susceptance of a transmission line), and in urban transportation control methods to impose desired synchronous behaviors has

Fig. 1 Network control to enforce a desired functional pattern from an abnormal or undesired one. The left panel contains a network of n = 7 oscillators
(top left panel, line thickness is proportional to the coupling strength), whose vector of natural frequencies ω has zero mean. The phase differences with


respect to θ1 (i.e., θi−θ1) converge to π8 π8 π6 π6 π3 2π
3 , as also illustrated in the phases' evolution from random initial conditions (bottom left panel, color
coded). The center left panel depicts the functional pattern R corresponding to such phase differences over time. The right panel illustrates the same
oscillator network after a selection of coupling strengths

π π π π and
π
 natural frequencies have been tuned (in red, the structural parameters A and ω are adjusted
to Ac and ωc) to obtain the phase differences 2π 3 3 6 6 8 8 , which encode the desired functional pattern in the center right. In this example, we have
computed the closest set (in the ℓ1-norm sense) of coupling strengths and natural frequencies to the original ones that enforce the emergence of the target
pattern. Importantly, only a subset of the original parameters has been modified.

2 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

only recently attracted the attention of the research satisfy the following equation:
community27–30. Perhaps the work that is closest to our approach
is ref. 28, where the authors tailor interconnection weights and BDðxÞδ ¼ ω; ð4Þ
natural frequencies to achieve a specified level of phase cohe- where we note that the vector ω has zero mean and that, with a
siveness in a network of Kuramoto oscillators. Our work slight abuse of notation, D(x) denotes the jEj-dimensional
improves considerably upon this latter study, whose goal is lim- diagonal matrix of the sine of the phase differences uniquely
ited to prescribing an upper bound to the phase differences, by defined by the (n−1)-dimensional vector xdesired.
enabling the prescription of pairwise differences and by investi- We begin by studying the problem of attaining a desired
gating the stability properties of different functional patterns. functional pattern using only nonnegative weights. With the
Taken together, existing results highlight the importance of above notation, for a desired functional pattern corresponding to
controlling distinct configurations of synchrony, but remain the phase differences x, this problem reads as
mainly focused on the control of “macroscopic” observables of
synchrony (e.g., the average synchronization level of all the find δ ð5Þ
oscillators). In contrast, our control method prescribes desired
pairwise levels of correlation across all of the oscillators, thus subject to BDðxÞδ ¼ ω; ð5aÞ
enabling a precise “microscopic” description of functional
interactions. and δ ≥ 0: ð5bÞ
It should be noticed that the feasibility of the optimization
Results
problem (5) depends on the sign of the entries of the diagonal
Feasible functional patterns in positive networks. A functional
matrix D(x), but is independent of their magnitude. To see this,
pattern is an n × n matrix whose entries are the time-averaged
notice that
cosine of the differences in the oscillator phases (see Eq. (2)).
When the oscillators reach an equilibrium, the differences of the DðxÞ ¼ signðDðxÞÞjDðxÞj;
oscillator phases become constant, and the network evolves into a
phase-locked configuration. In this case, the functional pattern of where the sign( ⋅ ) and absolute value ∣ ⋅ ∣ operators are applied
the network also becomes constant and is uniquely determined by element-wise. Then, Problem (5) is feasible if and only if there
the phase differences at the equilibrium configuration. In this exists a nonnegative solution to
work, we study functional patterns for the special case of phase- BsignðDðxÞÞ jDðxÞjδ ¼ ω:
locked oscillators and, since a functional pattern can be specified |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} |fflfflffl{zfflfflffl}
using a set of phase differences at equilibrium, convert the pro- 
B 
δ
blem of generating a functional pattern into the problem of
ensuring a desired phase-locked equilibrium. We recall that, while The feasibility of the latter equation, in turn, depends on the
projections of the natural frequencies ω on the columns of B:  a
convenient for the analysis, phase-locked configurations play a
crucial role in the functioning of many natural and man-made nonnegative solution exists if ω belongs to the cone generated by
the columns of B.  This also implies that, if a network admits a
networks31–33.
For the undirected network G ¼ ðO; EÞ, let xij = θj−θi be the desired functional pattern x then, by tuning its weights, the same
network can generate any other functional pattern xnew such that
difference of the phases of the oscillators i and j, and let x 2 RjEj
sign(D(xnew)) = sign(D(x)). Thus, by properly tuning its weights,
be the vector of all phase differences with ði; jÞ 2 E and i < j. The
a network can generally generate a continuum of functional
network dynamics (1) can be conveniently rewritten in vector
patterns determined uniquely by the signs of its incidence matrix
form as (see the “Methods” section)
and the oscillators natural frequencies. This property is illustrated
in Fig. 2 for the case of a line network.
_
BDðxÞδ ¼ ½ω1    ωn T  θ; ð3Þ A sufficient condition for the feasibility of Problem (5) is as
follows:
There exists δ ≥ 0 such that BD(x)δ = ω if there exists a set S
where B 2 Rn ´ jEj is the (oriented) incidence matrix of the satisfying:
network G, DðxÞ 2 RjEj ´ jEj is a diagonal matrix of the sine (i.a) Dii ðxÞDjj ðxÞBT:;i B:;j ≤ 0 for all i; j 2 Swith i ≠ j and
functions in Eq. (1), and δ 2 RjEj is a vector collecting all the Dii, Djj ≠ 0;
weights Aij with i < j. Because we focus on phase-locked (i.b) ωTB:,iDii(x) > 0 for all i 2 S;
trajectories, all oscillators evolve with the same frequency (i.c) ω 2 ImðB:;S Þ.
 and
the vector θ_ satisfies θ_ ¼ ωmean 1, where ωmean ¼ n1 ∑ni¼1 ωi is the Equivalently, the above conditions ensure that ω is contained
within the cone generated by the columns of B  :;S (see Fig. 3a for a
average of the natural frequencies of the oscillators. Further, since
G contains only n oscillators, any phase difference xij can always self-contained example). To see this, rewrite the pattern assign-
be written as a function of n−1 independent differences; for ment problem BD(x)δ = ω as
instance, {x12, x23, …, xn−1,n}. For instance, for any pair of BDðxÞδ ¼ B:;S DS;S ðxÞδS þ B:;S~ DS;
j1 ~S ~ ðxÞδS
~ ¼ ω; ð6Þ
oscillators i and j with i < j, it holds xij ¼ ∑k¼i xk;kþ1 . This
implies that the vector of all phase differences in equation (3), and where the subscripts S and S ~ denote the entries corresponding to
in fact any n × n functional pattern, has only n − 1 degrees of the set S and the remaining ones, respectively. If the vectors B:,i,
freedom and can be uniquely specified with a set of n−1 i 2 S, are linearly independent, condition (i.a) implies that
independent differences xdesired (see the “Methods” section). DS;S BT:;S B:;S DS;S is an M-matrix; that is, a matrix which has
Following this reasoning and to avoid cluttered notation, let nonpositive off-diagonal elements and positive principal
ω ¼ ½ω1  ωmean    ωn  ωmean T , and notice that the problem of minors34. Otherwise, the argument holds verbatim by replacing
enforcing a desired functional pattern simplifies to (i) converting S with any subset S ind  S such that the vectors B:,i, i 2 S ind , are
the desired functional pattern to the corresponding phase linearly independent. Condition (i.c) guarantees the existence of a
differences xdesired, and (ii) computing the network weights δ to solution to B:;S DS;S δS ¼ ω. A particular solution to the latter

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 3

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

Fig. 2 Mapping between desired phase differences and interconnection weights. a A line network of n = 4 nodes and its parameters. The desired phase
differences are shown in red. b Left panel: space of the phase differences; right panel: space of the interconnection weights. The pattern x is illustrated in
red in the left panel, and the network weights that achieve such a pattern are represented in red in the right panel. For fixed natural frequencies ω, the green
parallelepiped on the left represents the continuum of functional patterns within 0.2 radians from x which can be generated by the positive interconnection
weights in the green parallelepiped on the right.

Fig. 3 Algebraic and graph-theoretic conditions for the existence of positive weights that attain a desired functional pattern. a The left side illustrates a
simple network of 3 oscillators with adjacency matrix B  and vector of natural frequencies ω. The right side illustrates the cone generated by the columns of
 In this example, S ¼ f1; 2g satisfies the conditions for the existence of δ ≥ 0 in Eq. (5), as ω is contained within the cone generated by the columns B
B.  .
:;S

b The (directed) Hamiltonian path described by the columns of B:;H , with H ¼ f1; 2g, in the network of panel (a). c The existence of such an Hamiltonian
path, together with a positive projection of ω onto B  , also ensure that there exists a strictly positive δ > 0 solution to BD(x)δ = ω. In particular, for any
:;H
choice of x12, x23 ∈ (0, π), Eq. (4) reveals that if 0 < A13 < 0:5= sinðx12 þ x23 Þ, then there exist strictly positive weights A12 > 0 and A23 > 0 such that δ > 0.

equation is interconnection weights solves such an equation (see the

1
“Methods” section):
δS ¼ ðB:;S DS;S ðxÞÞy ω ¼ ðDS;S ðxÞBT:;S B:;S DS;S ðxÞÞ DS;S ðxÞBTS;S ω > 0
δH ¼ ðB:;H DH;H ðxÞÞy ðω  B:;H~ DH;
~ H~ ðxÞδH
~ Þ:
where (⋅)†denotes the Moore–Penrose pseudo-inverse of a
Because B  :;H ¼ B:;H DH;H defines an Hamiltonian path and
matrix. The positivity of δS follows from condition (i.b) and the
fact that the inverse of an M-matrix is element-wise because of (ii.b), the vector ðB:;H DH;H ðxÞÞy ω contains only strictly
nonnegative34. We conclude that a solution to Eq. (6) is given positive entries. Thus, for any sufficiently small and positive
by δS ¼ ðB:;S DS;S ðxÞÞy ω > 0 and δS~ ¼ 0. vector δH~ , the weights δH are also strictly positive, ultimately
To avoid disconnecting edges or to maintain a fixed network proving the existence of a strictly positive solution to the pattern
topology, a functional pattern should be realized in Problem (5) assignment problem (see the “Methods” section). Figure 3c
using a strictly positive weight vector (that is, δ > 0 rather than illustrates a self-contained example.
δ ≥ 0). While, in general, this is a considerably harder problem, a Taken together, the results presented in this section reveal that
sufficient condition for the existence of a strictly positive solution the interplay between the network structure and the oscillators’
δ > 0 is that the network with incidence matrix B  contains an natural frequencies dictates whether a desired functional pattern
Hamiltonian path, that is, a directed path that visits all the is achievable under the constraint of nonnegative (or even strictly
oscillators exactly once (Fig. 3b shows a network containing an positive) interconnections. First, dense positive networks with a
Hamiltonian path). Namely, large number of edges are more likely to generate a desired
There exists a strictly positive solution δ > 0 to BD(x)δ = ω if functional pattern, since their incidence matrix features a larger
(ii.a) the network with incidence matrix B  contains a directed number of candidate vectors to satisfy conditions (i.a)–(i.c).
Hamiltonian path H; Second, densely connected networks are also more likely to
(ii.b) ωTB:,iDii(x) > 0 for all i 2 H; contain an Hamiltonian path, thus promoting also strictly
The incidence matrix B  :;H of a directed Hamiltonian path H positive network designs. Third, after an appropriate relabeling
has two key properties. First, it comprises n−1 linearly of the oscillators such that any interconnection from i to j in the
independent columns, since the path covers all the vertices and Hamiltonian path satisfies i < j, condition (ii.b) is equivalent to
contains no cycles. This guarantees that ω 2 ImðB:;H Þ. Second, the requiring that ωi < ωj. That is, the feasibility of a functional
 T:;i B
 :;H satisfy B  :;i ¼ 2 and pattern is guaranteed when the natural frequencies increase
columns of the incidence matrix B
monotonically with the ordering identified by the Hamiltonian
 T:;i B
B  :;j 2 f0; 1g for all i; j 2 H, i ≠ j. Then, letting the set S in the path. This also implies, for instance, that sparsely connected
result above identify the columns of the Hamiltonian path, positive networks, and not only dense ones, can attain a large
conditions (ii.a) and (ii.b) imply (i.a)–(i.c), thus ensuring the variety of functional patterns. An example is a connected line
existence of a nonnegative set of weights δ that solves the pattern network with increasing natural frequencies, which can generate,
assignment problem BD(x)δ = ω. Furthermore, by rewriting the among others, any functional pattern defined by phase differences
pattern assignment problem as in Eq. (6), the following vector of that are smaller than π2 (trivially, when the phase differences are

4 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

smaller than π2 and the natural frequencies are increasing, a line parameterize infinite equilibria with a scalar γ 2 S1 in a cycle of
network contains an Hamiltonian path and the vector of natural n = 4 oscillators. Finally, as complete graphs are equivalent to a
frequencies has positive projections onto the columns of the composition of cycles, they also admit infinite compatible
incidence matrix). Figure 2a contains an example of such a patterns that can be parameterized akin to what occurs in a
network. simple cycle (see Supplementary Fig. 2).
We now turn our attention to finding the interconnection
Compatibility of multiple functional patterns. A single choice weights that simultaneously enable a collection of ℓ ≥ 1 desired
‘
of the interconnections weights can allow for multiple desired functional patterns fxðiÞ gi¼1 . We first notice that Eq. (7) reveals
functional patterns, as long as they are compatible with the net- that to achieve a desired functional pattern x(i) with components
work dynamics in Eq. (1). In this section, we provide a char- not equal to kπ, k 2 Z, the network weights δ must belong to an
acterization of compatible functional patterns in a given network, jEj-dimensional affine subspace of RjEj :
and derive conditions for the existence of a set of interconnection
1  
weights that achieve multiple desired functional patterns. Being δ 2 DðxðiÞ Þ By ω þ kerðBÞ ; 8i ¼ 1; ¼ ; ‘: ð9Þ
able to concurrently assign multiple functional patterns is crucial, 1
for instance, to the investigation and design of memory systems35, Let Γi ¼ DðxðiÞ Þ ðBy ω þ kerðBÞÞ. Then, to concurrently realize a
‘
where different patterns of activity correspond to distinct mem- collection of patterns fxðiÞ gi¼1 , a solution to Eq. (9) exists if and
ories. Furthermore, our results complement previous work on the T‘
only if i¼1 Γi ≠ ;. It is worth noting that, whenever the latter
search for equilibria in oscillator networks36. intersection coincides with a singleton, then there exists a single
To find a set of functional patterns that exist concurrently in a ‘
given network with fixed interconnection weights δ, we exploit choice of network weights that realizes fxðiÞ gi¼1 . However, if
T‘
the algebraic core of Eq. (4) and show that the kernel of the i¼1 Γi corresponds to a subspace, then infinite networks can
incidence matrix B uniquely determines the equilibria of the realize the desired collection of functional patterns. We conclude
network. In fact, for a given network (i.e., δ with nonzero by emphasizing that a positive δ that achieves the desired patterns
T
components) all compatible equilibria x(i), i ∈ {1, …, ℓ}, must exists if and only if ð ‘i¼1 Γi Þ \ RjEj≥ 0 ≠;. That is, if the network
satisfy weights belong to the nonempty intersection of the ℓ affine
DðxðiÞ Þδ ¼ By ω þ kerðBÞ: ð7Þ subspaces with the positive orthant.

From Eq. (7), we can see that the sine vector of all compatible Stability of functional patterns. A functional pattern is stable
equilibria must belong to a specific affine subspace of RjEj : when small deviations of the oscillators phases from the desired
  configuration lead to vanishing functional perturbations. Stability
sinðxðiÞ Þ 2 diagðδÞ1 By ω þ kerðBÞ : ð8Þ
is a desired property since it guarantees that the desired func-
Rewriting Eq. (4) in the above form connects the existence of tional pattern is robust against perturbations to the oscillators
distinct functional patterns with kerðBÞ, the latter featuring a dynamics. To study the stability of a functional pattern, we
number of well-known properties. For instance, it holds that analyze the Jacobian of the Kuramoto dynamics at the desired
dimðkerðBÞÞ ¼ jEj  n þ c, where c is the number of connected functional configuration, which reads as17
components in a network, and that kerðBÞ coincides with the ∂
subspace spanned by the signed path vectors of all undirected J ¼ θ_ ¼  Bdiag ðfAij cosðxij Þgði;jÞ2E ÞBT ;
cycles in the network37. Notice also that, after a suitable ∂θ |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} ð10Þ
reordering of the phase differences in x, we can write L ðxÞ
T
sinðxÞ ¼ ½sinðxTdesired Þ sinðxTdep Þ , where xdep denotes the phase where LðxÞ denotes the Laplacian matrix of the network with
differences dependent on n−1 desired phase differences xdesired. weights scaled by the cosines of the phase differences (the weight
Thus, all the xdesired for which sinðxdep Þ intersects the affine space between nodes i and j is Aij cosðθj  θi Þ). The functional pattern x
described by diagðδÞ1 ðBy ω þ kerðBÞÞ identify compatible func- is stable when the eigenvalues of the above Jacobian matrix have
tional patterns. negative real parts. For instance, if all phase differences are strictly
To showcase how the intimate relationship between the smaller than π2 (that is, the infinity-norm of x satisfies k x k1 < π2 ),
network structure and the kernel of its incidence matrix enables then the Jacobian in Eq. (10) is known to be stable17. In the case
the characterization of which (and how many) compatible that both cooperative and competitive interactions are allowed,
patterns coexist, we consider three essential network topologies: we can ensure stability of a desired pattern by specifying the
trees, cycles, and complete graphs. For the sake of simplicity, we network weights in δ such that Aij > 0 if jxij j< π2 and Aij < 0
let δ = 1 and ω = 0, so that Eq. (8) holds whenever otherwise, so that the matrix L becomes the Laplacian of a
sinðxðiÞ Þ 2 kerðBÞ. In networks with tree topologies it holds positive network (see Methods). Furthermore, we observe that in
kerðBÞ ¼ 0, and sinðxðiÞ Þ ¼ 0 is satisfied by 2n−1 patterns of the the particular case where some differences jxij j ¼ π2, the network
ðiÞ
form xjk ¼ 0; π, for all ðj; kÞ 2 E. Consider now cycle networks, may become disconnected since cosðxij Þ ¼ 0. Because the Lapla-
where kerðBÞ ¼ span 1. For any cycle of n ≥ 3 oscillators, two cian of a disconnected network has multiple eigenvalues at zero,
families of patterns are straightforward to derive. First, there are marginal stability may occur, and phase trajectories may not
ðiÞ converge to the desired pattern.
2n−1 patterns of the form xk;kþ1 ¼ 0; π, with k = 1, …, n−1, and
ðiÞ n1 ðiÞ When some phase differences are larger than π2 and the network
xn1 ¼  ∑k¼1 xk;kþ1 . Second, there are n−1 splay states17, where allows only for nonnegative weights, then stability of a functional
the oscillators’ phases evenly span the unitary circle, with pattern is more difficult to assess because the Jacobian matrix
ðiÞ
xjk ¼ 2πm
n , m = 1, …, n−1, ðj; kÞ 2 E. Figure 4 illustrates the
becomes a signed Laplacian38. The off-diagonal entries of a signed
compatible functional patterns satisfying Eq. (8) in a positive Laplacian satisfy Lij > 0 whenever jxij j> π2, thus possibly changing
network of three fully synchronizing oscillators. In general, the sign of its diagonal entries Lii ¼ ∑j Aij cosðxij Þ and violating
however, cycle networks of identical oscillators admit infinite the conditions for the use of classic results from algebraic graph
coexisting patterns. For instance, Fig. 5 shows how we can theory for the stability of Laplacian matrices. To derive a

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 5

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

Fig. 4 The intersection of an affine space with sinðxdep Þ determines the compatible functional patterns of 3 identical oscillators. Consider a fully
connected network of n = 3 identical oscillators with zero natural frequency and δ = 1. It is well known that x(1) = [0 0]T, x(2) = [π 0]T, x(3) = [0 π]T,
x(4) = [π π]T are phase difference equilibria. Furthermore, because sinðθÞ ¼ sinðπ  θÞ, this figure illustrates sinðx13 Þ as a function of x12 and x23 in four
different panels: sinðx12 þ x23 Þ (top left), sinðπ  x12 þ x23 Þ (top right), sinðx12 þ π  x23 Þ (bottom left), and sinðx12  x23 Þ (bottom right). The fourth panel
reveals that the two functional patterns compatible with x(j), j ∈ {1, …, 4}, correspond to x(5) = [2π/3 2π/3]T and x(6) = [−2π/3 −2π/3]T (in red).

Fig. 5 A homogeneous cycle network admits infinite compatible functional patterns. Since kerðBÞ ¼ span 1, the cycle network admits infinite compatible
equilibria, which can be parameterized by γ 2 S1 as xðiÞ ðγÞ ¼ ½π2  γ; π2 þ γ; π2  γ; π2 þ γT . Any arbitrarily small variation of γ yields sinðx ðiÞ ðγÞÞ 2 kerðBÞ. The
right panel illustrates the patterns associated with x(i)(γ), i = 1, …, 5 for increments of γ of 0.2 radians.

condition for the instability of the Jacobian in Eq. (10), we exploit is unstable whenever jxij j > π2 for any i, j. Instead, for a cycle
the notion of structural balance. We say that the cosine-scaled network with positive weights, the pattern x can be stable only if
network with Laplacian matrix L is structurally balanced if and it contains at most one phase difference π2 < jxij j<γ, where
only if its oscillators can be partitioned into two sets, O1 and O2 , γ ≈ 1.789776 solves γ  tanðγÞ ¼ 2π (see Supplementary Infor-
such that all ði; jÞ 2 E with Aij cosðxij Þ < 0 connect oscillators in mation). In the next section, we propose a heuristic procedure to
O1 to oscillators in O2 , and all ði; jÞ 2 E with Aij cosðxij Þ > 0 correct the interconnection weights in positive networks to
connect oscillators within Oi , i ∈ {1, 2}. If a network is promote stability of a functional pattern.
structurally balanced, then its Laplacian has mixed eigenvalues38.
Therefore, we conclude the following: Optimal interventions for desired functional patterns. Armed
If the functional pattern x yields a structurally balanced cosine- with conditions to guarantee the existence of positive inter-
scaled network, then x is unstable. connections that enable a desired functional pattern, we now
The above condition allows us to immediately assess the show that the problem of adjusting the network weights to gen-
instability of functional patterns for the special cases of line and erate a desired functional pattern can be cast as a convex opti-
cycle networks. In fact, for a line network with positive weights, x mization problem. Formally, for a desired functional pattern x

6 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

Fig. 6 Optimal interventions for desired functional patterns. a For the line network in Fig. 2a, we solve Problem (11) to assign the desired pattern
π π T π π 4π T T
xdesired ¼ ½4π
5 3 10 . The starting pattern x original ¼ ½10 3 5  is associated with interconnection weights δ = [3.4026 3.4641 6.4721] . Applying the optimal
correction α yields positive interconnection weights δ + α* = [6.4721 3.4026 3.4641]T that achieve the desired functional patterns xdesired. b Joint
*

allocation of two compatible equilibria for the phase difference dynamics. By taking θ1 as a reference, we choose two points for the phase differences

π π π π π π T
π π π π π πT
x1i = θi − θ1, i ∈ {2, …, 7}, to be imposed as equilibria in a network of n = 7 oscillators: xð1Þ
desired ¼ 6 6 4 4 6 4 and xð2Þ
desired ¼ 8 3 4 4 6 4 . In this example, we
find a set of interconnection weights (δ + α ) that solves the minimization problem (11) with constraint (12). The trajectories start at the (unstable)
*

equilibrium point xð1Þ ð2Þ 7

desired at time t = 0, and converge to the point x desired after a small perturbation p 2 T , with π ∈ [0 0.05], is applied to the phase
difference trajectories at time t = 50.

and network weights δ, we seek to solve δ + α*. To promote stability of the pattern x, we use a heuristic
procedure based on the classic Gerschgorin’s theorem39. Recall
min kαk2 ð11Þ
α that the stability of x is guaranteed when the associated Jacobian
matrix has a Laplacian structure, with negative diagonal entries
subject to BDðxÞðδ þ αÞ ¼ ω; ð11aÞ and nonnegative off-diagonal entries. Further, instability of x
depends primarily on the negative off-diagonal entries Aij cosðxij Þ
and ðδ þ αÞ ≥ 0; ð11bÞ of the Jacobian (these entries are negative because the sign of the
jEj network weight Aij is different from the sign of the cosine of the
where α 2 R are the controllable modifications of the network desired phase difference xij). Therefore, reducing the magnitude
weights, and ∥ ⋅ ∥2 denotes the ℓ2-norm. Figure 6a illustrates the of such entries Aij heuristically moves the eigenvalues of the
control of a functional pattern in a line network of n = 4 Jacobian towards the stable half of the complex plane (this
oscillators. phenomenon can be captured using the Gerschgorin circles, as we
The minimization problem (11) is convex, thus efficiently show in Fig. 7 for a network with 7 nodes). To formalize this
solvable even for large networks, and may admit multiple procedure, let δN and αN denote the entries of the weights δ and
minimizers, thus showing that different networks may exhibit tuning vector α, respectively, that are associated to negative
the same functional pattern. Moreover, in light of our results interconnections Aij cosðxij Þ < 0 in the cosine-scaled network.
above, Problem (11) can be easily adapted to assign a collection of
‘ Then, the optimization problem that enacts the proposed strategy
desired patterns fxðiÞ gi¼1 . To do so, we simply replace the becomes:
constraint (11a) with  
BDðxðiÞ Þðδ þ αÞ ¼ ω; 8i ¼ 1; ¼ ; ‘: ð12Þ min δN þ αN 2
α
subject to BDðxÞðδ þ αÞ ¼ ω; ð13Þ
Figure 6b illustrates an example where we jointly allocate two
functional patterns for a complete graph with n = 7 oscillators and ðδ þ αÞ ≥ 0:
(see Supplementary Information for more details on this
example).
Note that the minimization problem (11) does not guarantee Carefully reducing the weights δN þ αN promotes stability of
that the functional pattern x is stable for the network with weights the target pattern. Figure 7 illustrates the shift of the Jacobian’s

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 7

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

Fig. 7 Mechanism underlying the heuristic procedure to promote stability of functional patterns containing negative correlations.
π π π π πTFor the 7-oscillator
network in Supplementary Text 1.5, we apply the procedure in equation (13) to achieve stability of the pattern x desired ¼ 21π 32 6 6 8 8 3 , where
x12 ¼ θ2  θ1 > π2. The left plot illustrates the Gerschgorin disks (in blue) and the Jacobian's eigenvalues locations for the original network (as dark dots). The
complex axis is highlighted in purple. It can be observed in the zoomed-in panel that one eigenvalue is unstable (λ2 = 0.0565, in red). The optimal
correction α* is gradually applied to the existing interconnections from the left-most panel to the right-most one at 31 increments. The right zoomed-in panel
shows that, as a result of our procedure, n − 1 eigenvalues ultimately lie in the left-hand side of the complex plane (λ1 = 0 due to rotational symmetry and
λ2 = − 0.0178, in green).

eigenvalues while the optimal tuning vector α* is gradually applicable to directed networks. The latter extension can be
applied to a 7-oscillator network to achieve stability of a attained by replacing the constraints (11a) and (14a) with a
functional pattern containing negative correlations (the network suitable rewriting of the matrix form (4). We refer the interested
parameters can be found in the Supplementary Information). reader to the Supplementary Information for a comprehensive
Finally, we remark that the procedure in Eq. (13) can
 be further
 treatment of this extension and an example.
refined by introducing scaling constants to penalize δN þ αN 2
differently from the modification of other interconnection Applications to complex networks
weights (see Supplementary Information for further details and In the remainder of this paper, we apply our methods to an
an example). empirically reconstructed brain network and to the IEEE 39 New
The minimization problems (11) and (13) do not allow us to England power distribution network. In the former case, we use
tune the oscillators’ natural frequencies, and are constrained to the Kuramoto model to map structure to function and find that
networks with positive weights. When any parameter of the local metabolic changes underlie the emergence of functional
network is unconstrained and can be adjusted to enforce a desired patterns of recorded neural activity. In the latter case, we use our
functional pattern, the network optimization problem can be methods to restore the nominal network power flow after a fault.
generalized as


min  αT βT  Local metabolic changes govern the emergence of distinct
α;β 2 ð14Þ
functional patterns in the brain. The brain can be studied as a
network system in which Kuramoto oscillators approximate the
subject to BDðxÞðδ þ αÞ ¼ ½ω1    ωn T þ β; rhythmic activity of different brain regions12,30,40,41. Over short
ð14aÞ time frames, the brain is capable of exhibiting a rich repertoire of
functional patterns while the network structure and the inter-
where β denotes the correction to the natural frequencies. connection weights remain unaltered. Functional patterns of
Problem (14) always admits a solution because β can be chosen to brain activity not only underlie multiple cognitive processes but
satisfy the constraint (14a) for any choice of the network can also be used as biomarkers for different psychiatric and
parameters δ + α. Further, the (unique) solution to the neurological disorders42.
minimization problem (14) can also be computed in closed form: To shed light on the structure–function relationship of the
 
α
y   human brain, we utilize Kuramoto oscillators evolving on an
¼ BDðxÞ I n ½ω1    ωn T  BDðxÞδ empirically reconstructed brain network. We hypothesize that the
β
intermittent emergence of diverse patterns stems from changes in
where In denotes the n × n identity matrix. the oscillators’ natural frequencies—which can be thought of as
We conclude this section by noting that the minimization endogenous changes in metabolic regional activity regulated by
problems (11)–(14) can be readily extended to include other glial cells43 or exogenous interventions to modify undesired
vector norms besides the ℓ2-norm in the cost function (e.g., the synchronization patterns30. First, we show that phase-locked
ℓ1-norm to promote sparsity of the corrections), and to be trajectories of the Kuramoto model in Eq. (1) can be accurately

8 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

Fig. 8 Replication of empirically recorded functional connectivity in the brain through tuning of the natural frequencies of Kuramoto oscillators. The
anatomical brain organization provides the network backbone over which the oscillators evolve. The filtered fMRI time series provide the relative phase
differences between co-fluctuating brain regions, and thus define the desired phase differences x, which is used to calculate the metabolic change encoded
in the oscillators' natural frequencies. In this figure, we select the 40-s time window from t0 = 498 s to tf = 538 s for subject 18 in the second scanning
session. We obtain ∥R−F∥2 = 0.2879. Additionally, we verify that the analysis of the Jacobian spectrum (see Eq. (10)) accurately predicts the stability of
the phase-locked trajectories. Supplementary Fig. 7a illustrates the basin of attraction of R, which we numerically estimate to be half of the torus.

extracted from noisy measurements of neural activity and are a to x. We show in Fig. 8 that the assignment of natural frequencies
relatively accurate approximation of empirical data. according to the extracted phase differences promotes sponta-
We employ structural (i.e., interconnections between brain neous synchronization to accurately replicate the empirical
regions) and functional (i.e., time series of recorded neural functional connectivity F.
activity) data from ref. 40. Structural connectivity consists of a These results corroborate the postulate that structural connec-
sparse weighted matrix A whose entries represent the strength of tions in the brain support the intermittent activation of specific
the physical interconnection between two brain regions. Func- functional patterns during rest through regional metabolic
tional data comprise time series of neural activity recorded changes. Furthermore, we show that the Kuramoto model
through functional magnetic resonance imaging (fMRI) of represents an accurate and interpretable mapping between the
healthy subjects at rest. Because the phases of the measured brain anatomical organization and the functional patterns of
activity have been shown to carry most of the information frequency-synchronized neural co-fluctuations. Once a good
contained in the slow oscillations recorded through fMRI time mapping is inferred, it can be used to define a generative brain
series, we follow the steps in ref. 40 to obtain such phases from the model to replicate in silico how the brain efficiently enacts large-
data by filtering the time series in the [0.04, 0.07] Hz frequency scale integration of information, and to develop personalized
range (Supplementary Information). Next, since frequency intervention schemes for neurological disorders related to
synchronization is thought to be a crucial enabler of information abnormal synchronization phenomena48,49.
exchange between different brain regions and homeostasis of
brain dynamics44,45, we focus on functional patterns that arise Power flow control in power networks for fault recovery and
from phase-locked trajectories, as compatible with our analysis. prevention. Efficient and robust power delivery in electrical grids
For simplicity, we restrict our study to the cingulo-opercular is crucial for the correct functioning of this critical infrastructure.
cognitive system, which, at the spatial scale of our data, comprises Modern, reconfigurable power networks are expected to be resi-
n = 12 interacting brain regions46. lient to distributed faults and malicious cyber-physical attacks50
We identify time windows in the filtered fMRI time series while being able to rapidly adapt to varying load demands. In
where the signals are phase-locked, and compute two matrices for addition, climate change is straining service reliability, as
each time window: a matrix F 2 R12 ´ 12 of Pearson correlation underscored by recent events such as the Texas grid collapse after
coefficients (also known as functional connectivity), and a Winter Storm Uri in February 202151, and the New Orleans
functional pattern R 2 R12 ´ 12 (as in equation (2)) from the blackout after Hurricane Ida in August 202152. Therefore, there
phases extracted by solving the nonconvex phase synchronization exists a dire need to design control methods to efficiently operate
problem47. Strikingly, we find that ∥F−R∥2 ≪ 1 consistently these networks and react to unforeseen disruptive events.
(see Supplementary Information and Supplementary Fig. 7b), The Kuramoto model in Eq. (1) has been shown to be
thus demonstrating that our definition of the functional pattern is particularly relevant in the modeling of large distribution
a good replacement for the classical Pearson correlation networks and microgrids9. Preliminary work on the control of
arrangements in networks with oscillating states. frequency synchronization in electrical grids modeled through
Building upon the above finding, we test whether the Kuramoto oscillators has been developed in ref. 53. Here, we
oscillators’ natural frequencies embody the parameter that links present a method that leverages our findings to guarantee not
the brain network structure to its function (i.e., structural and only frequency synchronization but also a target pattern of active
functional matrices). We set ω = BD(x)δ, where x are phase power flow. Our method can be used for power (re)distribution
differences obtained from the previous step, and integrate the with respect to specific pricing strategies, fault prevention (e.g.,
Kuramoto model in Eq. (1) with random initial conditions close when a line overheats), and recovery (e.g., after the disconnection

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 9

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

Fig. 9 Fault recovery in the IEEE 39 New England power distribution network through minimal and local intervention. a New England power distribution
network. The generator terminal buses illustrate the type of generator (coal, nuclear, hydroelectric). We simulate a fault by disconnecting the transmission
line 25 (between loads 13 and 14). b The fault causes the voltage phases θ to depart from normal operating conditions, which could cause overheating of
some transmission lines (due to violation of the nominal thermal constraint limits) or abnormal power delivery. To recover the pre-fault active power flow
and promote a local (sparse) intervention, we solve the optimization in Eq. (11) by minimizing the 1-norm of the structural parameter modification δ with no
scaling parameters in the cost functional. The network returns to the initial operative conditions with a localized modification of the neighboring
transmission lines' impedances.

of a branch). Furthermore, thanks to the formal guarantees that remaining parameters (i.e., admittances) so that the original
our method prescribes, we are able to prevent Braess’ Paradox in functional pattern can be recovered. Figure 9b illustrates the
power networks54, which is a phenomenon where the addition of effectiveness of our procedure at recovering the nominal
interconnections to a network may impede its synchronization. pattern of active power flow by means of minimal and localized
It has been shown in ref. 9, Lemma 1 that the load dynamics interventions (see also Supplementary Fig. 8).
(nodes 1–29 in Fig. 9a) in a structure-preserving power grid The above application is based on a classical lossless structure-
model have the same stable synchronization manifold of Eq. (1). preserving power network model9. However, in the power
In this model, ωi ¼ p‘i =Di is the active power load at node i, systems literature, more complex dynamics that relax some of
where Di is the damping coefficient, and Aij ¼ jvi jjvj jIðY ij Þ=Di , the modeling assumptions have been proposed. For instance,
with vi denoting the nodal voltage magnitude and IðY ij Þ being ref. 58 uses a third-order model (or one-axis model) that takes
into account transient voltage dynamics. Instead, ref. 59 studies
the imaginary part of the admittance Yij (see Supplementary
the case of interconnections with power losses, which lead to a
Information for details about this model). In this example, we
network of phase-lagged Kuramoto–Sakaguchi oscillators60. We
choose a highly damped scenario where Di = 1, which is possibly
show in the Supplementary Information that our procedure to
due to local excitation controllers. Notice that, when the phase
recover a target functional pattern can still be applied successfully
angles θ are phase-locked and Aij is fixed, the active power flow is
to a wide range of situations involving these more realistic
given by Aij sinðθj  θi Þ.
models.
We posit that solving the problem in Eq. (11) to design a local
reconfiguration of the network parameters can recover the power
distribution before a line fault or provide the smallest parameter Discussion
changes to steer the load powers to desired values. In practice, Distinct configurations of synchrony govern the functioning of
control devices such as flexible alternating current transmission oscillatory network systems. This work presents a simple and
systems (FACTs) allow operators and engineers to change the mathematically grounded mapping between the structural para-
network parameters55,56. We demonstrate the effectiveness of our meters of arbitrary oscillator networks and their components’
approach by recovering a desired power distribution in the IEEE functional interactions. The tantalizing idea of prescribing pat-
39 New England power distribution network after a fault. During terns in networks of oscillators has been investigated before, yet
a regime of normal operation, we simulate a fault by disconnect- only partial results had been reported in the literature. Here, we
ing the line between two loads and solve the problem in Eq. (11) demonstrate that the control of patterns of synchrony can be cast
to find the minimum modification of the couplings aij that as optimal (convex) design and tuning problems. We also
recovers the nominal power distribution. investigate the feasibility of such optimizations in the cases of
We first utilize MATPOWER57 to solve the power flow networks that admit negative coupling weights and networks that
problem. Then we use the active powers pℓ and voltages v at the are constrained to positive couplings. Our control framework also
buses to obtain the natural frequencies ω and the adjacency allows us to prescribe multiple desired equilibria in Kuramoto
matrix A of the oscillators, respectively, while the voltage phase networks, a problem that is relevant in practice and had not been
angles are used as initial conditions θ(0) for the Kuramoto investigated before.
model in equation (1). We integrate the Kuramoto dynamics As stability of a functional pattern may be a compelling
and let the voltage phases θ(t) reach a frequency-synchronized property in many applications, we explore conditions to test and
steady state, which corresponds to a normal operating enforce the stability of functional patterns. We show that such
condition. The phase differences also represent a functional conditions are rather straightforward in the case of networks that
pattern across the loads. Next, to simulate a line fault, we admit both cooperative and competitive interactions. However,
disconnect one line. By solving the problem in Eq. (11) with stability of target functional patterns in networks that are con-
xdesired corresponding to the pre-fault steady-state voltage strained to only cooperative interactions is a more challenging
phase differences, we compute the smallest variation of the task, for which we demonstrate that any pattern associated with a

10 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

structurally balanced cosine-weighted network cannot be stable. may require the derivation of ad-hoc principles based on Lya-
To overcome this issue, we propose a heuristic procedure that punov’s stability theory70.
adjusts the oscillators’ coupling strengths to violate the structu- The framework presented in this work has other limitations,
rally balanced property and promote the stability of functional which can be addressed in follow-up studies. First, despite its
patterns with negative correlations. While heuristic, our proce- capabilities in modeling numerous oscillatory network systems,
dure for stability has proven successful in all our numerical stu- the Kuramoto model cannot capture the amplitude of the oscil-
dies. Notice that, differently from methods that study an lations, making it most suitable for oscillator systems where most
“average" description of the system at a near-synchronous state of the information is conveyed by phase interaction as demon-
(see, e.g., ref. 61), here we assess the stability of exact target phase- strated in ref. 40 for resting brain activity. To model brain activity
locked trajectories where phase values can instead be arbitrarily during cognitively demanding tasks such as learning, higher-
spread over the torus. This method can also be extended to assess dimensional oscillators may be more suitable71. Second, the use of
the stability of equilibria of higher-dimensional oscillators, pro- phase-locked trajectories is instrumental to the control and design
vided that the considered equilibrium state is a fixed point. of functional patterns. Yet, it is not necessary. In fact, restricting
We emphasize that our results are also intimately related to the the control to phase-locked dynamics does not capture exotic
long-standing economic problem of enhancing network opera- dynamical regimes in which only a subset of the oscillators is
tions while optimizing wiring costs. In any complex system where frequency-synchronized. Third and finally, in some situations, the
synchrony between components ensures appropriate functions, it network parameters are not fully known. While still an active area
is beneficial to maximize synchronization while minimizing the of research, network identification of oscillator systems may be
physical variations of the interconnection weights62. Compatible employed in such scenarios72.
with this principle, neural systems are thought to have evolved to Directions of future research can be both of a theoretical and
maximize information processing by promoting synchronization practical nature. For instance, follow-up studies can focus on the
through optimal spatial organization63. Inspired by the efficiency derivation of a general condition for the stability of a feasible func-
observed in neurobiological circuitry, Eq.(11) could be utilized for tional pattern in positive networks. Further, a thorough investigation
the design of optimal interaction schemes in large-scale computer of which network structures allow for multiple prescribed equilibria
networks whose performance relies on synchronization-based may be particularly relevant in the context of memory systems,
tasks64. where different patterns are associated with different memory states.
An important consideration that highlights the general con- Specific practical applications may also require the inclusion of
tributions of the present study is that being able to specify pairwise sparsity constraints on the accessible structural parameters for the
functional relationships between the oscillators also solves pro- implementation of the proposed control and design framework.
blems such as phase-locking, full, and cluster synchronization. Yet,
the converse is not true. In fact, even in the general setting of cluster Methods
synchronization65—where distinct groups of oscillators behave Matrix form of Eq. (1) and phase-locked solutions. For a given network of
cohesively—one can only achieve a desired synchronization level oscillators, we let the entries of the (oriented) incidence matrix B be defined
component-wise after choosing the orientation of each interconnection (i, j). In
within the same cluster, but not across clusters, which instead is particular, i points to j if i < j, and Bkℓ = −1 if oscillator k is the source node of the
possible with our approach. Specifically, in cluster synchronization, interconnection ℓ, Bkℓ = 1 if oscillator k is the sink node of the interconnection ℓ,
oscillators belonging to the same cluster are forced to synchronize, and Bkℓ = 0 otherwise. The matrix form of Eq. (1) can be written as
thus implying that the associated diagonal blocks of the functional 2 3
..
pattern R display values close to 1. Seminal work in ref. 66 devel- 6 . 7

T 6 7
oped a nonlinear feedback control to change the coupling functions θ_ ¼ ω1    ωn  B6 sinðxij Þ 7δ
4 5
in Eq. (2) to engineer clusters of synchronized oscillators, whereas ..
the authors of ref. 67 propose the formation of clusters through .

T
T
selective addition of interconnections to the network. More ¼ ω1    ωn  B diag fsinðxij Þgði;jÞ2E δ ¼ ω1    ωn  BDðxÞδ;
recently, the control of partially synchronized states with applica-
where D(x) is the diagonal matrix of the sine functions in Eq. (1).
tions to brain networks is studied in ref. 30 by means of structural When the oscillators evolve in a phase-locked configuration, the oscillator
interventions, and in ref. 68 via exogenous stimulation. Ultimately, frequencies become equal to each other and constant. In particular, since 1TB = 0,
the results presented in this work not only complement but go well we have 1T θ_ ¼ 1T k1 ¼ 1T ½ω1    ωn T , thus showing that in any phase locked
beyond the control of macroscopic synchronization observables trajectory, the oscillator's frequency k needs to equal the mean natural frequency
1 n
and partial synchronization by allowing to specify the synchrony n ∑i¼1 ωi .
level of pairwise interactions.
While our contributions include the analysis of the stability Any feasible functional pattern has n−1 degrees of freedom. The values of a
functional pattern can be uniquely specified using a set of n−1 correlation values. To
properties of functional patterns, here we do not assess their
see this, let us define the incremental variables x = Mθ, where M 2 RjEj ´ n is the
basins of attraction. We emphasize that, in general, the estimation matrix whose kth row, associated to xij, is all zeros except for bki = −1 and bkj = 1.
of the basin of attraction of the attractors of nonlinear systems Consider the first n−1 rows of M, associated with x12, x13, …x1n, and notice that they
remains an outstanding problem, and even the most recent results are linearly independent. Moreover, the row associated to xij, i > 1, can be obtained by
rely on numerical approaches or heavy modeling assumptions69. subtracting the row associated to x1i to the row associated to x1j, implying that the rank
of M is n−1. Any collection of n−1 linearly independent rows of M defines a full row-
Further, in the case of coupled Kuramoto oscillators, existing rank matrix M min (e.g., any n−1 rows corresponding to the transpose incidence matrix
literature shows that the number of equilibria for the phase dif- of a spanning tree37). We let xmin ¼ M min θ, where xmin is a smallest set of phase
ferences increases significantly with the cardinality of the differences that can be used to quantify the synchronization angles among all oscil-
network36, making the study of basins of attraction extremely lators. Because kerðM min Þ ¼ 1, we can obtain the phases θ from xmin modulo rotation:
challenging. In the Supplementary Information, we extend pre- θ ¼ M ymin xmin  c1, where M ymin denotes the Moore–Penrose pseudo-inverse of M min
and c is some real number. Further, since kerðM min Þ ¼ kerðMÞ, we can reconstruct all
vious work on identical oscillators to networks with hetero- phase differences x from xmin :
geneous oscillators, and show that functional patterns can be at
most almost-globally stable in cluster-synchronized positive net- MM ymin xmin ¼ Mðθ þ c1Þ ¼ Mθ þ 0 ¼ x:
works. Yet, a precise estimation of the basin of attraction for any The above equation reveals that all the differences x are encoded in xmin .
arbitrary target pattern goes beyond the scope of this work and That is, any xij can be written as a linear combination of the elements in xmin .

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 11

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2

For example, if n = 3 and xmin ¼2½x12 x23 T , then

3 x13 is a linear combination of
Received: 14 June 2021; Accepted: 28 June 2022;
1 1 0  y
1 1 0
the differences in xmin , i.e., x ¼ 4 0 1 1 5 xmin , in which
0 1 1
1 0 1
x13 = x12 + x23. Thus, because n−1 incremental variables define all the
remaining ones, the entries of any functional pattern must have only n−1
degrees of freedom. References
1. Cross, M. C. & Hohenberg, P. C. Pattern formation outside of equilibrium.
Existence of a strictly positive solution to Problem 5. Rewrite the pattern Rev. Mod. Phys. 65, 851 (1993).
assignment problem BD(x)δ = ω as 2. Pikovsky, A., Rosenblum, M. & Kurths, J. Synchronization: A Universal
Concept in Nonlinear Sciences (Cambridge University Press, 2003).
BDðxÞδ ¼ B:;H DH;H ðxÞδH þ B:;H~ DH;
~ H~ ðxÞδ H
~ ¼ ω; 3. Cavagna, A. et al. Scale-free correlations in starling flocks. Proc. Natl Acad. Sci.
where the subscripts H and H ~ denote the entries corresponding to the Hamiltonian USA 107, 11865–11870 (2010).
path in conditions (ii.a) and (ii.b) and the remaining ones, respectively. Since 4. Stewart, I., Golubitsky, M. & Pivato, M. Symmetry groupoids and patterns of
 :;H Þ ¼ ImðB:;H D:;H ðxÞÞ ¼ spanð1Þ? , ImðB ~ D ~ ðxÞÞ  spanð1Þ? , and ω ∈ synchrony in coupled cell networks. SIAM J. Appl. Dyn. Syst. 2, 609–646
ImðB :;H :;H (2003).
span(1)⊥, for any vector δH~ , the following set of weights solves the above equation: 5. Golubitsky, M., Stewart, I. & Török, A. Patterns of synchrony in coupled cell
1 networks with multiple arrows. SIAM J. Appl. Dyn. Syst. 4, 78–100 (2005).
δH ¼ ðB:;H DH;H ðxÞÞy ðω  B:;H~ DH;
~ H~ ðxÞδH
T
~ Þ ¼ ðDH;H ðxÞB:;H B:;H DH;H ðxÞÞ
6. Sorrentino, F. & Ott, E. Network synchronization of groups. Phys. Rev. E 76,
DH;H ðxÞBT:;H ðω  B:;H~ DH;
~ H~ ðxÞδ H
~Þ 056114 (2007).
7. Zhang, F. & Leonard, N. E. Coordinated patterns of unit speed particles on a
Because the matrix DH;H ðxÞBT:;H B:;H DH;H ðxÞ is an M-matrix, its inverse has non-
closed curve. Syst. Control Lett. 56, 397–407 (2007).
negative entries. Further, by condition (ii.b), DH;H ðxÞBT:;H ω is strictly positive. 8. Buck, J. & Buck, E. Biology of synchronous flashing of fireflies. Nature 211,
1
Then, the vector ðDH;H ðxÞBT:;H B:;H DH;H ðxÞÞ DH;H ðxÞBT:;H ω is also strictly positive, 562–564 (1966).
and so is the solution vector δH for any sufficiently small and positive vector δH~ . 9. Dörfler, F., Chertkov, M. & Bullo, F. Synchronization in complex oscillator
networks and smart grids. Proc. Natl Acad. Sci. USA 110, 2005–2010 (2013).
10. Trenberth, K. E. Spatial and temporal variations of the Southern Oscillation.
Enforcing stability of functional patterns in networks with cooperative and
Q. J. R. Meteorol. Soc. 102, 639–653 (1976).
competitive interactions. To ensure that the Jacobian matrix in Eq. (10) is the
11. Beliaev, M., Zöllner, D., Pacureanu, A., Zaslansky, P. & Zlotnikov, I. Dynamics
Laplacian of a positive network and, thus, stable, we solve the problem in Eq. (5)
of topological defects and structural synchronization in a forming periodic
with a slight modification of the constraints. Specifically, we post-multiply the
tissue. Nat. Phys. 17, 410–415 (2021).
matrix B in Eq. (5) as BΔ, where Δ ¼ diagðfsignðcosðxij ÞÞgði;jÞ2E Þ is a matrix that
12. Cabral, J., Hugues, E., Sporns, O. & Deco, G. Role of local network oscillations
changes the sign of the columns of B associated to negative weights in the cosine- in resting-state functional connectivity. NeuroImage 57, 130–139 (2011).
scaled network: 13. Fell, J. & Axmacher, N. The role of phase synchronization in memory
BΔDðxÞδ ¼ ω processes. Nat. Rev. Neurosci. 12, 105–118 (2011).
14. Strogatz, S. H. Exploring complex networks. Nature 410, 268–276 (2001).
Solving for positive interconnection weights the problem in Eq. (5) under the 15. Strogatz, S. H. & Stewart, I. Coupled oscillators and biological
above-modified constraint yields a stable Jacobian in a network where the final synchronization. Sci. Am. 269, 102–109 (1993).
couplings are Δδ. 16. Arenas, A., Díaz-Guilera, A., Kurths, J., Moreno, Y. & Zhou, C.
Synchronization in complex networks. Phys. Rep. 469, 93–153 (2008).
Heuristic procedure to promote stability of functional patterns in positive 17. Dörfler, F. & Bullo, F. Synchronization in complex networks of phase
networks. We provide a heuristic procedure to promote the stability of functional oscillators: a survey. Automatica 50, 1539–1564 (2014).
patterns that include negative correlations in a network with nonnegative weights. 18. Acebrón, J. A., Bonilla, L. L., Pérez Vicente, C. J., Ritort, F. & Spigler, R. The
Our procedure relies on the definition of Gerschgorin disks and the Gerschgorin Kuramoto model: a simple paradigm for synchronization phenomena. Rev.
Theorem. Mod. Phys. 77, 137 (2005).
Definition of Gerschgorin disk. Let M 2 Cn ´ n be a complex matrix. The ith 19. Hong, H. & Strogatz, S. H. Kuramoto model of coupled oscillators with
Gerschgorin disk is Di ¼ ðM ii ; r i Þ, i = 1, …, n, where the radius is ri = ∑j≠i∣Mij∣ and positive and negative coupling parameters: an example of conformist and
the center is Mii. contrarian oscillators. Phys. Rev. Lett. 106, 054102 (2011).
20. Arenas, A., Díaz-Guilera, A. & Pérez-Vicente, C. J. Synchronization reveals
Theorem 2. (Gerschgorin39). The eigenvalues of the matrix M lie within the union
S topological scales in complex networks. Phys. Rev. Lett. 96, 114102 (2006).
n
i¼1 Di of its Gerschgorin disks. 21. Arola-Fernández, L., Díaz-Guilera, A. & Arenas, A. Synchronization
Whenever all target phase differences in x satisfy jxij j ≤ π2, the Gerschgorin disks invariance under network structural transformations. Phys. Rev. E 97, 060301
of the Jacobian in Eq. (10) all lie in the closed left half-plane. However, for patterns (2018).
x containing phase differences jxij j ≥ π2, the union of the Gerschgorin disks 22. Pecora, L. M. & Carroll, T. L. Master stability functions for synchronized
intersects the right half-plane. Reducing the magnitude of the entries satisfying coupled systems. Phys. Rev. Lett. 80, 2109–2112 (1998).
Aij cosðxij Þ < 0 effectively shrinks the radius of the Gerschgorin disks that overlap 23. Strogatz, S. H. SYNC: The Emerging Science of Spontaneous Order (Hyperion,
with the right half-plane and shift their centers towards the left-half plane due to 2003).
the structure of the Jacobian matrix. We remark that the procedure in Eq. (13) is a 24. Sorrentino, F., Di Bernardo, M. & Garofalo, F. Synchronizability and
heuristic, and it is probably effective only when all interconnections with synchronization dynamics of weighed and unweighed scale free networks with
Aij cosðxij Þ < 0 can be removed so that all the Gerschgorin disks lie completely in degree mixing. Int. J. Bifurc. Chaos 17, 2419–2434 (2007).
the left-half plane. 25. Nishikawa, T. & Motter, A. E. Network synchronization landscape reveals
compensatory structures, quantization, and the positive effect of negative
Data availability interactions. Proc. Natl Acad. Sci. USA 107, 10342–10347 (2010).
The brain data that support the findings of this study are publicly available in 26. Belykh, I., Jeter, R. & Belykh, V. Foot force models of crowd dynamics on a
the Supplementary Information of40, at https://doi.org/10.1371/journal.pcbi.1004100. wobbly bridge. Sci. Adv. 3, e1701512 (2017).
s006. The IEEE 39 New England data parameters and interconnection scheme analyzed 27. Papadopoulos, L., Kim, J. Z., Kurths, J. & Bassett, D. S. Development of
in this study for the structure-preserving power network model can be found in the structural correlations and synchronization from adaptive rewiring in
reference textbook73 (see also Supplementary Information for modeling assumptions). networks of Kuramoto oscillators. Chaos: Interdiscip. J. Nonlinear Sci. 27,
The parameters for the simulations on the network-reduced IEEE 39 New England test 073115 (2017).
case in the Supplementary Information have been obtained from refs. 74,75. All data used 28. Fazlyab, M., Dörfler, F. & Preciado, V. M. Optimal network design for
in this study are also available in the public GitHub repository: https://github.com/ synchronization of coupled oscillators. Automatica 84, 181–189 (2017).
tommasomenara/functional_control. 29. Forrow, A., Woodhouse, F. G. & Dunkel, J. Functional control of network
dynamics using designed Laplacian spectra. Phys. Rev. X 8, 041043 (2018).
30. Menara, T., Baggio, G., Bassett, D. S. & Pasqualetti, F. A framework to control
Code availability functional connectivity in the human brain. In IEEE Conf. on Decision and
Source code and documentation for the numerical simulations presented here are freely Control, Nice, France, December 2019, 4697–4704 (IEEE, 2019).
available in GitHub at: https://github.com/tommasomenara/functional_control with the 31. Kaka, S. et al. Mutual phase-locking of microwave spin torque nano-
identifier https://doi.org/10.5281/zenodo.4546413. oscillators. Nature 437, 389–392 (2005).

12 NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31733-2 ARTICLE

32. van Wingerden, M., Vinck, M., Lankelma, J. & Pennartz, C. M. A. Theta-band 64. Korniss, G., Novotny, M. A., Guclu, H., Toroczkai, Z. & Rikvold, P. A.
phase locking of orbitofrontal neurons during reward expectancy. J. Neurosci. Suppressing roughness of virtual times in parallel discrete-event simulations.
30, 7078–7087 (2010). Science 299, 677–679 (2003).
33. Kirst, C., Timme, M. & Battaglia, D. Dynamic information routing in complex 65. Pecora, L. M., Sorrentino, F., Hagerstrom, A. M., Murphy, T. E. & Roy, R.
networks. Nat. Commun. 7, 11061 (2016). Cluster synchronization and isolated desynchronization in complex networks
34. Horn, R. A. & Johnson, C. R. Matrix Analysis (Cambridge University Press, with symmetries. Nat. Commun. 5, 1–8 (2014).
1985). 66. Kiss, I. Z., Rusin, C. G., Kori, H. & Hudson, J. L. Engineering complex
35. Skardal, P. S. & Arenas, A. Memory selection and information switching in dynamical structures: sequential patterns and desynchronization. Science 316,
oscillator networks with higher-order interactions. J. Phys.: Complex. 2, 1886–1889 (2007).
015003 (2020). 67. D’Souza, R. M. & Mitzenmacher, M. Local cluster aggregation models of
36. Mehta, D., Daleo, N. S., Dörfler, F. & Hauenstein, J. D. Algebraic explosive percolation. Phys. Rev. Lett. 104, 195702 (2010).
geometrization of the Kuramoto model: equilibria and stability analysis. 68. Bansal, K. et al. Cognitive chimera states in human brain networks. Sci. Adv. 5,
Chaos: Interdiscip. J. Nonlinear Sci. 25, 053103 (2015). eaau8535 (2019).
37. Godsil, C. & Royle, G. F. Algebraic Graph Theory. Graduate Texts in 69. Chen, S., Fazlyab, M., Morari, M., Pappas, G. J. & Preciado, V. M. Learning
Mathematics (Springer, New York, 2001). region of attraction for nonlinear systems. 60th IEEE Conference on Decision
38. Zelazo, D. & Bürger, M. On the robustness of uncertain consensus networks. and Control (CDC) 6477–6484 (IEEE, 2021).
IEEE Trans. Control Netw. Syst. 4, 170–178 (2017). 70. Zhu, L. & Hill, D. J. Synchronization of Kuramoto oscillators: a regional
39. Geršgorin, S. Über die abgrenzung der eigenwerte einer matrix. Bull. Acad. Sci. stability framework. IEEE Trans. Autom. Control 65, 5070–5082 (2020).
URSS Cl. Sci. Math. na, 749–754 (1931). 71. Qin, Y., Menara, T., Bassett, D. S. & Pasqualetti, F. Phase-amplitude coupling
40. Ponce-Alvarez, A. et al. Resting-state temporal synchronization networks in neuronal oscillator networks. Phys. Rev. Res. 3, 023218 (2021).
emerge from connectivity topology and heterogeneity. PLoS Comput. Biol. 11, 72. Panaggio, M. J., Ciocanel, M., Lazarus, L., Topaz, C. M. & Xu, B. Model
1–23 (2015). reconstruction from temporal data for coupled oscillator networks. Chaos:
41. Politi, A. & Rosenblum, M. Equivalence of phase-oscillator and integrate-and- Interdiscip. J. Nonlinear Sci. 29, 103116 (2019).
fire models. Phys. Rev. E 91, 042916 (2015). 73. Sauer, P. W. & Pai, M. A. Power System Dynamics and Stability (Prentice-Hall,
42. Schnitzler, A. & Gross, J. Normal and pathological oscillatory communication 1998).
in the brain. Nat. Rev. Neurosci. 6, 285 (2005). 74. Susuki, Y., Mezić, I. & Hikihara, T. Coherent swing instability of power grids.
43. Fields, D. R. et al. Glial biology in learning and cognition. Neuroscientist 20, J. Nonlinear Sci. 21, 403–439 (2011).
426–431 (2014). 75. Moeini, A., Kamwa, I., Brunelle, P., & Sybille, G. Open data IEEE test systems
44. Varela, F., Lachaux, J. P., Rodriguez, E. & Martinerie, J. The brainweb: phase implemented in SimPowerSystems for education and research in power grid
synchronization and large-scale integration. Nat. Rev. Neurosci. 2, 229–239 dynamics and control. In 2015 50th International Universities Power
(2001). Engineering Conference (UPEC), 1–6 (IEEE, 2015).
45. Noori, R. et al. Activity-dependent myelination: a glial mechanism of
oscillatory self-organization in large-scale brain networks. Proc. Natl Acad. Sci.
USA 117, 13227–13237 (2020). Acknowledgements
46. Power, J. D. et al. Functional network organization of the human brain. T.M. was supported in part by the National Science Foundation (Grant NSF-557 NCS-
Neuron 72, 665–678 (2011). FO-1926829) and the Army Research Office (Grant ARO-W911NF1910360).
47. Boumal, N. Nonconvex phase synchronization. SIAM J. Optim. 26, 2355–2377
(2016). Author contributions
48. Váša, F. et al. Effects of lesions on synchrony and metastability in cortical T.M., G.B., and F.P. designed research; T.M. and G.B. performed research; T.M. analyzed
networks. Neuroimage 118, 456–467 (2015). data and developed code; D.S.B. and F.P. supervised the research, and T.M., G.B., D.S.B.,
49. Menara, T., Lisi, G., Pasqualetti, F. & Cortese, A. Brain network dynamics and F.P. wrote the paper.
fingerprints are resilient to data heterogeneity. J. Neural Eng. 18, 026004
(2021).
50. Wang, J.-W. & Rong, L.-L. Cascade-based attack vulnerability on the US Competing interests
power grid. Saf. Sci. 47, 1332–1336 (2009). The authors declare no competing interests.
51. Busby, J. W. et al. Cascading risks: Understanding the 2021 winter blackout in
Texas. Energy Res. Soc. Sci. 77, 102106 (2021).
52. Santana, R. & Reeves, J. Hurricane Ida traps Louisianans, shatters the power
Additional information
Supplementary information The online version contains supplementary material
grid (Associated Press, 2021).
available at https://doi.org/10.1038/s41467-022-31733-2.
53. Skardal, P. S. & Arenas, A. Control of coupled oscillator networks with
application to microgrid technologies. Sci. Adv. 1, e1500339 (2015).
Correspondence and requests for materials should be addressed to Fabio Pasqualetti.
54. Witthaut, D. & Timme, M. Braess’s paradox in oscillator networks,
desynchronization and power outage. N. J. Phys. 14, 083036 (2012).
Peer review information Nature Communications thanks Rico Berner, and the other,
55. L’Abbate, A. et al. The role of FACTS and HVDC in the future pan-European
anonymous, reviewer(s) for their contribution to the peer review of this work. Peer
transmission system development. In Proc. 9th IET International Conference
reviewer reports are available.
on AC and DC Power Transmission (ACDC 2010), 1–8 (IET, 2010).
56. Motter, A. E., Myers, S. A., Anghel, M. & Nishikawa, T. Spontaneous
Reprints and permission information is available at http://www.nature.com/reprints
synchrony in power-grid networks. Nat. Phys. 9, 191–197 (2013).
57. Zimmerman, R. D., Murillo-Sánchez, C. E. & Thomas, R. J. MATPOWER: Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
steady-state operations, planning, and analysis tools for power systems published maps and institutional affiliations.
research and education. IEEE Trans. Power Syst. 26, 12–19 (2010).
58. Sharafutdinov, K., Gorjão, L. R., Matthiae, M., Faulwasser, T. & Witthaut, D.
Rotor-angle versus voltage instability in the third-order model for
synchronous generators. Chaos: Interdiscip. J. Nonlinear Sci. 28, 033117 Open Access This article is licensed under a Creative Commons
(2018). Attribution 4.0 International License, which permits use, sharing,
59. Hellmann, F. et al. Network-induced multistability through lossy coupling and adaptation, distribution and reproduction in any medium or format, as long as you give
exotic solitary states. Nat. Commun. 11, 592 (2020). appropriate credit to the original author(s) and the source, provide a link to the Creative
60. Sakaguchi, H. & Kuramoto, Y. A soluble active rotater model showing Commons license, and indicate if changes were made. The images or other third party
phase transitions via mutual entertainment. Prog. Theor. Phys. 76, 576–581 material in this article are included in the article’s Creative Commons license, unless
(1986). indicated otherwise in a credit line to the material. If material is not included in the
61. Sun, J., Bollt, E. M. & Nishikawa, T. Master stability functions for coupled article’s Creative Commons license and your intended use is not permitted by statutory
nearly identical dynamical systems. EPL (Europhys. Lett.) 85, 60011 (2009). regulation or exceeds the permitted use, you will need to obtain permission directly from
62. Nishikawa, T. & Motter, A. E. Maximum performance at minimum cost the copyright holder. To view a copy of this license, visit http://creativecommons.org/
in network synchronization. Phys. D: Nonlinear Phenom. 224, 77–89 licenses/by/4.0/.
(2006).
63. Achard, S. & Bullmore, E. Efficiency and cost of economical brain functional
networks. PLOS Comput. Biol. 3, e17 (2007). © The Author(s) 2022

NATURE COMMUNICATIONS | (2022)13:4721 | https://doi.org/10.1038/s41467-022-31733-2 | www.nature.com/naturecommunications 13