Uncertainty propagation in complex

networks: From noisy links to critical

properties
Cite as: Chaos 30, 023129 (2020); https://doi.org/10.1063/1.5129630
Submitted: 01 October 2019 . Accepted: 01 February 2020 . Published Online: 19 February 2020

Lluís Arola-Fernández , Guillem Mosquera-Doñate, Benjamin Steinegger , and Alex Arenas

COLLECTIONS

Paper published as part of the special topic on Symmetry and Optimization in the Synchronization and Collective
Behavior of Complex Systems
Note: This paper is part of the Focus Issue on Symmetry and Optimization in the Synchronization and Collective
Behavior of Complex Systems.

ARTICLES YOU MAY BE INTERESTED IN

Message-passing theory for cooperative epidemics

Chaos: An Interdisciplinary Journal of Nonlinear Science 30, 023131 (2020); https://
doi.org/10.1063/1.5140813

Noise stability of synchronization and optimal network structures

Chaos: An Interdisciplinary Journal of Nonlinear Science 30, 013148 (2020); https://
doi.org/10.1063/1.5121341

Totally asymmetric simple exclusion process on multiplex networks

Chaos: An Interdisciplinary Journal of Nonlinear Science 30, 023103 (2020); https://
doi.org/10.1063/1.5135618

Chaos 30, 023129 (2020); https://doi.org/10.1063/1.5129630 30, 023129

© 2020 Author(s).
 Chaos ARTICLE scitation.org/journal/cha

Uncertainty propagation in complex networks:

From noisy links to critical properties
Cite as: Chaos 30, 023129 (2020); doi: 10.1063/1.5129630
Submitted: 1 October 2019 · Accepted: 1 February 2020 ·
Published Online: 19 February 2020 View Online Export Citation CrossMark

Lluís Arola-Fernández,1 Guillem Mosquera-Doñate,2,3 Benjamin Steinegger,1 and Alex Arenas1,a)

AFFILIATIONS
1
Departament d’Enginyeria Informàtica i Matemàtiques, Universitat Rovira i Virgili, 43007 Tarragona, Catalonia, Spain
2
Center for Complexity Science, University of Warwick, Coventry CV4 7AL, United Kingdom
3
The Alan Turing Institute, London NW1 2DB, United Kingdom

Note: This paper is part of the Focus Issue on Symmetry and Optimization in the Synchronization and Collective Behavior of
Complex Systems.
a)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
Many complex networks are built up from empirical data prone to experimental error. Thus, the determination of the specific weights of
the links is a noisy measure. Noise propagates to those macroscopic variables researchers are interested in, such as the critical threshold
for synchronization of coupled oscillators or for the spreading of a disease. Here, we apply error propagation to estimate the macroscopic
uncertainty in the critical threshold for some dynamical processes in networks with noisy links. We obtain closed form expressions for the
mean and standard deviation of the critical threshold depending on the properties of the noise and the moments of the degree distribution
of the network. The analysis provides confidence intervals for critical predictions when dealing with uncertain measurements or intrinsic
fluctuations in empirical networked systems. Furthermore, our results unveil a nonmonotonous behavior of the uncertainty of the critical
threshold that depends on the specific network structure.
Published under license by AIP Publishing. https://doi.org/10.1063/1.5129630

The critical point (or threshold) refers to the minimum value of a I. INTRODUCTION
tuning parameter that triggers a phase transition in a dynamical The study of critical phenomena has been, and still is, a
process. This value becomes an uncertain quantity when noise is fruitful area of research in network science.1 Critical phenom-
present in the underlying network of interactions. For the class of ena in networks include a wide set of aspects, from structural
processes where the threshold is determined by the inverse of the changes in networks, or percolation phenomena,2 to epidemic3 or
largest eigenvalue of the adjacency matrix, we apply error prop- synchronization4 thresholds and many other phase transitions in
agation to estimate the critical uncertainty induced by quenched dynamical processes defined on networks.1,5 The estimation of the
noise on the network weights. We obtain closed form expressions critical threshold is of upmost importance to predict the onset of the
for the mean and standard deviation of the threshold depending phase transition, and hence a major concern in several applications,
on the properties of the noise and the moments of the network such as the containment of an infectious disease6 or the control of
degree distribution. Our results can be used in practical situa- synchronization in the power grid.7,8 However, an accurate estima-
tions, to provide confidence intervals for the predictions of the tion of the threshold is often elusive and costly because it depends on
critical threshold when dealing with uncertain measurements or the particular details of the whole network structure, usually through
intrinsic fluctuations in a networked system. Furthermore, the its eigenvalues.
results unveil several noise-amplifying properties of the networks, As network science becomes more and more extended, its
including a nonmonotonous behavior of the uncertainty of the potential applications grow fueled by the necessity of analyzing data
threshold depending on the heterogeneity and density of the net- produced in diverse fields of research, such as sociology, biology,
work. Accordingly, our analysis predicts the existence of particu- experimental physics, etc. However, the data collected in any of
lar structures that maximize the uncertainty of the threshold only the former fields are not free from experimental error, induced,
due to small noise in the weights, without altering the underlying for example, by sampling biases, device accuracy, or mistakes in
structure of the links. data entry. Nevertheless, the literature on network science usually

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-1

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

dismisses these error sources and produces results that are only valid from random matrix theory15,16 and spectral graph theory.17,18 How-
if data is error free. Some authors have concentrated their attention ever, it becomes very challenging to study noisy sparse networks
on the inference of missing data in networks;9–12 however, to the best with arbitrary degree distributions in these frameworks. Here, we
of our knowledge, no similar attention has been paid to the propaga- use an alternative approach, based on applying error propagation to
tion of uncertainty from the structure to the properties of dynamical the mean-field approximation (MFA) of Eq. (1). This approxima-
processes running on it. tion obviously restricts the validity range of the analysis; however,
The lack of works devoted to the analysis of error propagation the results are found to be very accurate in some scenarios, and,
in networks is probably due to the fact that many studies consider more importantly, they provide clear analytical insight on how the
networks unweighted, where a link is a binary variable denoting uncertainty in the structure affects the determination of the critical
its existence or not. However, the vast majority of networks are threshold.
weighted, i.e., the existence or not is valued by its intensity. The Our derivation starts assuming a mean-field approach. For
accurate determination of the weight is unlikely, and, therefore, simplicity, we restrict to the case of undirected (symmetric) net-
the error in their numerical values will influence any particular works. Under the aforementioned conditions, the critical threshold
measurement of the network properties. in Eq. (1) can be approximated19–22 by
Here, we present a study of error propagation in networks
hsi
where links are subject to uncertainty in their weights and wonder Kc = , (2)
about the effect that this uncertainty will have in the determination hs2 i
of the critical threshold. In particular, we focus on those dynamical where hsn i is the n-moment of the strength distribution (the strength
processes in which the critical point is known to be inversely propor- of a node is the sum of in-coming/out-going weights). Equation (2)
tional to the largest eigenvalue of the connectivity matrix. In Sec. II, can also be obtained directly from the equations of motion of
we present the particularities of our analysis and derive our main the dynamical process (for instance, in the Kuramoto model20) by
results; in Sec. III, we study the range of uncertainty in the critical assuming that the local field in a node is proportional to the global
point for different network structures; and, finally, in Sec. IV, we field weighted by the in-strength of the node.4 For the rest of the
discuss the implications and limitations of the current study, paving article, we will refer to Eq. (2) as the mean-field approximation
the way for new analysis to come. (MFA).
First, we test the accuracy of the critical threshold in the MFA,
II. PROPAGATION OF UNCERTAINTY IN THE CRITICAL
Eq. (2), by comparing it to Eq. (1), in Erdös–Rényi networks with
THRESHOLD uncertainty in the weights. In Fig. 1, we plot the threshold distribu-
We consider a dynamical process running on top of a complex tion for two different values of the intensity of the uncertainty σ . We
network with N units. We restrict the study to the class of dynam- observe that the MFA accurately determines the distribution, and
ical models in which a phase transition occurs at a critical value of that the values of the expected critical threshold Kc and its variance
the coupling intensity (the threshold), and where this value is given are clearly dependent on σ . In general, we expect our results to be
in terms of the largest eigenvalue λmax of the network connectiv- accurate in the cases in which the approximation of Eq. (2) remains
ity matrix C whose values represent the weighted structure of the valid.
network.13 Explicitly, we assume that Using Eq. (2), we can express Kc in terms of the moments of
K0 the degree distribution and noise parameters. The detailed calcula-
Kc = , (1) tions are shown in Appendix A. We obtain
λmax (C)
PN
where K0 is a constant that depends on the specific details of µi ki
Kc = PN 2 2 i=1 PN , (3)
the particular process. Without loss of generality, we fix K0 = 1. i=1 µi (ki − ki ) + i=1 hw ii ki
2
Equation (1) estimates the threshold for a wide variety of dynam-
ical processes, including the synchronization of heterogeneous where µi is the average weight of node i, and hw2 ii the average
phase-oscillators,4 the onset of endemicity of a disease in epidemic second moment of the weight distribution for node i. In random
models,3,14 and the phase transition in the Ising model in networks, homogeneous networks, for sufficiently large degree (ki  1), we
to name a few.1,2,5 The aim of this work is to understand how small can approximate µi = µ, and hw2 ii = σ 2 + µ2 in Eq. (3). This
noise in the entries of C affects the statistical properties of the macro- approximation allows to write down a simple relation between the
scopic threshold given by Eq. (1), without looking into the details of mean of the critical threshold and the uncertainty of the network as
a specific dynamical model. For the sake of simplicity, we assume µhki
that the quenched noise in the entries is gaussian and uncorrelated, hKc i ≈ . (4)
µ2 hk2 i + σ 2 hki
where each weight is independently drawn from a normal distribu-
tion N (µ, σ 2 ), being µ > 0 the average weight and σ 2 its variance. Interestingly, the naïve approximation in Eq. (4) already informs
Nevertheless, the proposed analysis can be extended to other dis- that the critical threshold decreases as the noise intensity σ
tributions of noise, either theoretical or obtained through empirical increases. This can be understood because the noise increases the
measurements, if the covariance matrix of the noise is known (see structural heterogeneity of the network, and heterogeneity tends to
Appendix B for details). make the critical threshold to vanish.3,4,22 Note that for µ = 1 and
To study the exact statistics of Kc in Eq. (1) induced by the σ = 0, we recover the usual threshold for unweighted, undirected
presence of noise, one could use in principle the available tools networks21 and for σ  1, hKc i ≈ hki/µhk2 i.

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-2

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

FIG. 2. Numerics [Eq. (1)] vs theory [Eqs. (4) and (6)]: mean and standard devi-
ation of the threshold Kc depending on the noise intensity σ for an Erdös–Rényi
network with N = 200, p = 0.3, µ = 1, and 5000 independent realizations for
each value of the noise intensity σ .

In Fig. 2, we show the accuracy of the theoretical expres-

sions for an Erdös–Rényi network, confirming the validity of the
FIG. 1. Empirical distribution of the critical point Kc governed by Eq. (1) (boxes) approach, at least for small noise and homogeneous structures. Note
and MFA (solid lines) in an Erdös–Rényi network with N = 200, p = 0.3, K0 = 1, that the linear approximation used in Eq. (5) is valid as far as24
and µ = 1 for two different noise intensities (σ = 0.2 blue and σ = 0.5 red). The
distribution corresponds to 104 independent realizations of the noise. 1
JT0 VJ0  Tr[(H0 V)2 ], (7)
2
where H0 ∈ R2N×2N is the Hessian matrix of the system evaluated at
Now, we estimate confidence intervals for the uncertainty the mean values of the random variables. The detailed calculations of
of Kc , that is, the standard deviation named here δKc [or the H0 are shown in Appendix B. Both terms in Eq. (7) depend implicitly
variance (δKc )2 ]. For this purpose, we use the method of error on the value of the noise, so their scaling with σ will determine the
propagation,23,24 that quantifies how the error in the microscopic range of validity of Eq. (6). We numerically examine the goodness of
variables of a system (the 2N random variables in our nodal both the linear, Eq. (5) and the second-order approximation for the
description) propagates through a macroscopic quantity (the critical uncertainty δKc
threshold Kc ). In a first-order expansion, we have
1
(δKc )2 ≈ JT0 VJ0 , (5) (δKc )2 ≈ JT0 VJ0 + Tr[(H0 V)2 ] (8)
2
with J ∈ R2N the Jacobian of the system evaluated at the mean values against the numerical results obtained for the Erdös–Rényi network
of the random variables µE and hwE2 i and V ∈ R2N×2N the covariance analyzed so far, and also for a real world network with large size
matrix, which depends on the full connectivity matrix C. The details and heterogeneous connectivity patterns (the worldwide air trans-
of these calculations (for uncorrelated gaussian noise and fixing portation network). The air transportation network was constructed
K0 = 1) are shown in Appendix B. Finally, we obtain the following using data from the website openflights.org, which has information
closed form expression: about the traffic between airports updated to 2012; data are avail-
able from Ref. 6. This network accounts for the largest connected
(δKc )2 ≈ a[µ4 (2hkihk3 i component, with 3154 nodes and 18 592 edges.
− hk2 i2 ) − 2µ2 σ 2 (hkihk2 i − hki2 ) + σ 4 hki2 ], (6) Figure 3 shows that the first and second-order solutions are
practically indistinguishable for small noise, therefore validating the
4
with a = 2σ 2 hki/[N(µ2 hk2 i + σ 2 hki) ]. result in Eq. (6) in this regime. The deviation of the theory from the
Equation (6) shows that, beyond the nonlinear dependence on actual values in the empirical network (right plot in Fig. 3) points
the network and noise parameters, the uncertainty in the threshold toward another direction: the goodness of the MFA itself. Basically,
is a finite-size effect, and decays with N−1/2 . To compare networks of the theory is expected to be accurate for networks that deviate from
different sizes, we will scale the threshold by the size N in the current a random structure as long as the MFA in Eq. (2) holds. It is not
analysis. the main goal of this article to convey an exhaustive verification of

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-3

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

and for the same noise parameters µ and σ . After some algebra, the
condition for a given network to display higher uncertainty in Kc
than a random regular network reads
hk2 i2 hk2 i2
 
hk3 i > 1+ . (11)
2hki hki4
Now, we can use Eq. (11) to evaluate the role of heterogene-
ity. Let us consider a power-law distribution p(k) ≈ k−γ , where
the exponent γ controls the tail of the distribution. For the
FIG. 3. Numerics vs theory: standard deviation of the critical threshold δKc value γ = 3, one recovers the well-known scale-free network that
depending on the noise intensity σ with µ = 1 for a (left) fixed Erdös–Rényi net- emerges from preferential attachment.26 For lower (higher) values
work (N = 200, hki = 60, p = 0.3) and (right) the empirical network of airports of γ , the network becomes more (less) heterogeneous. For a finite
(N = 3154, hki ≈ 12) for 2000 independent realizations for each value of the power-law network, the moments of the degree distribution are
noise. Results have been rescaled by N.
given by
n−γ +1 n−γ +1
(−γ + 1)(kmax − kmin )
hkn i = γ +1 γ +1
. (12)
the theory for particular networks, and we refer the reader to the (n − γ + 1)(kmax − kmin )
literature19,21,22,25 for details on the validity of the MFA. Moreover, it
is important to remark that even if the MFA holds, the method of By fixing the value of kmin , we can explore the space of networks with
error propagation (at any order) can only be applied in our problem a given (γ , kmax ), thus revealing the effect of heterogeneity and size.
when the mean of the signal µ is sufficiently large compared to the To simplify the visualization, we define
noise. 
3

2hkihk i
q = log   2 2
. (13)
III. THE ROLE OF THE STRUCTURE IN THE ERROR hk2 i2 1 + hkhkii4
PROPAGATION
This way, when q = 0, the uncertainty of the critical threshold of a
The network structure plays an important role in the uncer- network is the same than that of the regular one, and for positive
tainty range of Kc . After the finding of Eq. (6), some interesting (negative) values of q, we are measuring an increase (decrease) of
questions arise: does the heterogeneity induce an increase of the δK with respect to the homogeneous network. In Fig. 4, we show
critical fluctuations with respect to a homogeneous network? Is the the theoretical results obtained for the q value of networks in the
behavior of (δKc ) monotonous with the moments of the degree dis- space (γ , kmax ). We note that the three horizontal lines correspond
tribution of the network? If not, is there any particular structure that to the cases where the network has an integer exponent of 2, 3,
maximizes the uncertainty of the critical point induced by noise in or 4. In these cases, the first, second, or third moments diverge.
the weights?
To answer these questions, we consider the regime where net-
works are sufficiently large and σ  µ. Then, we can approximate
Eq. (6) by its leading term, neglecting terms in σ larger than O(σ 2 )
2hkihk3 i − hk2 i2
(δKc )2 ≈ 2σ 2 hKc i4 . (9)
Nhki3
Note that δKc increases linearly with the noise intensity and scales
with hKc i2 . We know that hKc i2 is reduced by the heterogeneity
of the degree distribution,3,4,22 and, therefore, one would expect
δKc to follow the same trend. However, the nonlinear depen-
dence on the moments of the degree distribution could change this
intuition.
To understand this effect, we choose first as a reference the
most homogeneous network we can consider, a regular network, i.e.,
ki = k, ∀i. We compute Kc and δKc for a regular network, obtaining
1
hKc ireg ≈ ,
µk
(10)
2σ 2
(δKc )2reg ≈ .
Nµ4 k3 FIG. 4. Colormap showing the theoretical dependence of q on the exponent γ
2 and the maximum degree of the network kmax . The value of kmin is fixed to kmin = 5
The role of the heterogeneity will be detected by comparing (δKc ) and the resolution of the map is 100 × 100.
with (δKc )2reg for networks with the same size and average degree,

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-4

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

IV. DISCUSSION AND CONCLUSIONS

The results found in Sec. III are of theoretical and practical
relevance for the field of network science and they should be
investigated further in detail. We have shown that particular net-
work structures, such as power-law degree distribution networks
with exponent γ ≈ 3, maximize the uncertainty of the critical
threshold in the presence of noisy weights. This fact should be taken
into account in the prediction of the critical threshold in empiri-
cal networks (which are usually heterogeneous) because, as proven,
the accuracy in the estimation crucially depends on the underlying
FIG. 5. Relative value of the theoretical (left) and numerical (right) uncertainty structure of the network. Moreover, the results might have a strong
δKc for scale-free networks in the range γ ∈ [2, 6] for sizes N = 500, 1000, and impact in the context of network optimization and adaptation,28–30
2000, µ = 1, σ = 0.05, and minimum degree fixed at kmin = 5 compared to reg-
ular networks with the same average degree, and the same characteristics of the especially considering the ubiquity and theoretical relevance2,26 of
noise. The results are obtained with 200 realizations of the noise for each net- power-law networks with exponent γ ≈ 3 and the well-established
work and then averaging with 200 networks for each configuration of the modified hypothesis that many biological networks are operating near the
preferential attachment algorithm. The high variance at each point shows that the critical point.31,32 In particular, one could wonder to which extent
results are very sensitive to the particular structure of the network, although the the existence of power-law networks with an exponent close to 3,
general trend is captured. maximizing the range of critical values, has been evolutionary favor-
able. In this sense, the current results make a natural connection
with the previous work in Ref. 33 where it was shown that scale-free
It is also important to remark that below γ = 2, it is not feasible networks with exponent γ = 3 are able to achieve a larger variety
to generate networks with a pure power-law distribution.2 Besides of macrostates with respect to homogeneous networks (specifically
these considerations, we observe an interesting result. As expected, near the critical threshold) by deterministically tuning the weights
for large values of the exponent γ , the networks show similar of the links.
uncertainty to that of a regular network. However, for γ < 4, uncer- From the methodological side, the formalism introduced in
tainty significantly increases, reaching a maximum as the expo- Sec. II represents a first step in the use of error propagation methods
nent approaches γ = 3, before decreasing again. When approaching to the analysis of complex networks with dynamical processes on top
the value of γ = 3, the network maximizes the third moment of of them. The formalism is flexible and it can be applied to other net-
the degree distribution, while minimizing its second moment and, work properties and in other scenarios, being of special importance
therefore, emerges as the optimal uncorrelated structure amplifying the case of colored noise obtained directly from empirical measure-
the uncertainty in the threshold. Conversely, uncertainty is min- ments. We conjecture that this line of research will receive more
imal for maximally heterogeneous networks, corresponding to an attention in the future due to the increasing amount of data (not
exponent γ ≈ 2. Interestingly, the nonmonotonous dependence on free of errors), that is, being collected for a large variety of systems.
γ is amplified as we increase the size of the system (in terms of its We remark also that the current method is based on a MFA of the
maximum degree). largest eigenvalue of the connectivity matrix, and this approxima-
To validate the previous theoretical prediction, we gener- tion neglects strong correlations of the eigenvalues in the presence
ate synthetic power-law networks using the modified preferential of noise.34,35 While definitely more results are needed, the present
attachment algorithm with an attractiveness parameter that control formalism provides analytical insight into the studied phenomena
the exponent.27 Fixing the value of the minimum degree kmin , and and turns out to give very accurate quantitative predictions if a few
tuning the exponent and the size of the network, we detect a max- assumptions on the network hold.
imum in the uncertainty δKc for the exponent γ = 3, as shown in To summarize, in this work we have studied how noise in
Fig. 5, thus confirming the prediction of the theory. the weights of a complex network affects the critical threshold of a
We observe good qualitative agreement for the nonmono- dynamical process. We have restricted our study to the wide family
tonous dependency on the heterogeneity, and also that system of processes where the threshold depends on the largest eigenvalue
size reinforces this dependency. The slight mismatch may be orig- of the connectivity matrix. In this scenario, and using the well-
inated by the fact that, for scale-free networks with γ > known MFA, we have applied error propagation to derive analytical
√5/2,
the critical threshold is better approximated by Kc ≈ 1/ kmax expressions for the mean and standard deviation of the threshold
than by the ratio of moments.19–22 Furthermore, the small sizes depending on the noise parameters and the moments of the degree
that are considered in the numerical analysis produce signifi- distribution. We validated our results against numerical simulations,
cant degree correlations which may contribute to the mismatch. showing good agreement when the initial MFA holds. Moreover,
Despite these flaws, the mean-field approach captures well the the formalism allowed us to carefully examine the effect that the
dependence on the exponent and size of these networks. The network structure plays in the amplification of the noise at the
results point toward the difficulty to accurately determine the crit- critical point. Surprisingly, we found a nonmonotonous behavior
ical threshold of scale-free networks, with exponent γ ≈ 3, since of the critical uncertainty with respect to the heterogeneity of the
δKc is maximized in the presence of noisy weights for these net- underlying network. By considering the paradigmatic case of uncor-
works. related power-law networks, we found that networks with exponent

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-5

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

γ ≈ 3 (γ ≈ 2) emerge as the structures that maximize (minimize) APPENDIX B: CALCULATION OF THE VARIANCE
the uncertainty of the threshold, due to an interplay between the The propagation of uncertainty of a nonlinear function of
second and third moment of the degree distribution. the random variables as Eq. (3) requires to use a truncated Tay-
lor expansion.24 Up to second-order, and in the notation used
ACKNOWLEDGMENTS in the main text, the approximate variance of the function is
L.A.-F. and A.A. acknowledge the Spanish MINECO (Grant given by
No. PGC2018-094754-B-C2). A.A. also acknowledges the Generali- 1
tat de Catalunya (Grant No. 2017SGR-896), the Universitat Rovira i (δKc )2 ≈ JT0 VJ0 + Tr[(H0 V)2 ], (B1)
2
Virgili (Grant No. 2017PFR-URV-B2-41), and the ICREA Academia
and the James S. McDonnell Foundation (Grant No. 220020325). where the Jacobian vector and the Hessian matrix are evaluated at
G.M.-D. acknowledges funding from the Engineering and Phys- the mean values of the random variables µ E and hwE2 i. The Jacobian
ical Sciences Research Council (EPSRC) and MRC Centre for of the system in Eq. (3) is
Doctoral Training in Mathematics for Real World Systems (No.  
EP/L015374/1) and The Alan Turing Institute under the EPSRC ∂Kc ∂Kc ∂Kc ∂Kc
J= ,..., , ,..., . (B2)
Grant (No. EP/N510129/1). B.S. acknowledges financial support ∂µ1 ∂µN ∂hw2 i1 ∂hw2 iN
from the European Union’s Horizon 2020 research and innovation
programme under the Marie Skodowska-Curie Grant Agreement First, we compute the partial derivatives in Eq. (B2) explicitly from
No. 713679 and from the Universitat Rovira i Virgili (URV). Eq. (3), obtaining

∂Kc 1 ki (µ2 hk2 i + σ 2 hki) − 2µ2 (k2i − ki )hki

APPENDIX A: CALCULATION OF THE MEAN ≈ ,
∂µi N (µ2 hk2 i + σ 2 hki)2
We can write the degrees and strengths in terms of the binary
connections (aij = 0 or 1) and weights (wij ∈ R) of the connectiv- ∂Kc 1 ki µhki
≈− , (B3)
ity matrix C, i.e ki = Nj=1 aij and si = Nj=1 aij wij . For the average
P P
∂hw2 ii N (µ2 hk2 i + σ 2 hki)2
strength hsi, we have
  where the sign ≈ stands for assuming, in good approximation, that
N
1 X X
N the input parameters µ and σ 2 are the actual mean values of the
hsi = aij wij  . (A1) random variables µ E and σE 2 = hw
E 2i − µ
E 2.
N i=1 j=1
The Hessian matrix, the square matrix of the second-order par-
tial derivatives of the function in Eq. (3), can be directly obtained
Note that we can write Eq. (A1) equivalently as hsi = (1/N) i µi ki ,
P
by taking derivates from Eq. (B3). After some algebra, and defining
where µi is the average weight of node i. For sufficiently large degree
Q = µ2 hk2 i + σ 2 hki, we obtain
(ki  1), one can approximate µi = µ, and therefore hsi = µhki.
However, in general, it is important to keep the contribution of each ∂ 2 Kc 1
node because each µi has a specific uncertainty depending on the ≈ 2 3 [Q(2µ(k2j − kj )ki − (2 + 2δij µ(k2i − ki )kj ))
∂µi ∂µj NQ
degree of node i, and this affects the overall uncertainty on Kc . For
the second moment hs2 i, we have − (ki − 8µ3 hki(k2i − ki )(k2j − kj )). (B4)
 2
N N
2 1 X X The Hessian matrix of our system is symmetric, such that
hs i = aij wij 
N i=1 j=1 ∂ 2 Kc /∂µi ∂hw2 ij = ∂ 2 Kc /∂hw2 ii ∂µj . We obtain

∂ 2 Kc 1
 
N N N
1 X X X ≈ 2 3 [−Qki kj + 4µ2 hkikj (k2i − ki )], (B5)
= aij w2ij + aij aik wij wik  . (A2) ∂µi ∂hw ij
2 NQ
N i=1 j=1
j6=k
and for the last term, we have
aij aik = k2i − ki , we obtain
P
Noticing that j6=k
∂Kc 2µki kj hki
" N N
# ≈ . (B6)
2 1 X 2 2 X
2
∂hw2 ii ∂hw2 ij N2 Q3
hs i = µ (k − ki ) + hw ii ki , (A3)
N i=1 i i i=1 For the covariance matrix, we can obtain explicit expression for
where hw2 ii is the average second moment of the i-node. Plugging the entries (V)ij when the noise in the weights is assumed gaus-
Eqs. (A1) and (A3) into Eq. (2) in the main text, we obtain sian and uncorrelated. By assumption, the network is symmetric and
PN so it will be the covariance matrix, which can be written in block
µ i ki form as
Kc = PN 2 2 i=1 PN , (A4)
i=1 µi (ki − ki ) + i=1 hw ii ki
2
vµ 2
 
vµ,hw2 i
V= 2 ,
which correspond to Eq. (3) in the main text. vµ,hw2 i vhw2 i

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-6

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

where vµ 2 , vµ,hw2 i , and vhw2 i 2 are symmetric matrices in RN×N that 4

A. Arenas, A. Díaz-Guilera, J. Kurths, Y. Moreno, and C. Zhou, “Synchronization
capture each covariance term between the two random variables in complex networks,” Phys. Rep. 469, 93–153 (2008).
5
(µi , hw2 ii ) of all nodes. Explicitly, S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D. U. Hwang,
“Complex networks: Structure and dynamics,” Phys. Rep. 424, 175–308
σ2 (2006).
 
aij
(vµ 2 )ij = δij + , (B7) 6
J. Matamalas, A. Arenas, and S. Gómez, “Effective approach to epidemic con-
ki kj tainment using link equations in complex networks,” Sci. Adv. 4, eaau4212
2µσ 2

aij
 (2018).
7
(vµ,hw2 i )ij = δij + , (B8) I. Dobson, B. Carreras, V. Lynch, and D. Newman, “Complex systems analysis
ki kj of series of blackouts: Cascading failure, critical points, and self-organization,”
Chaos 17, 026103 (2007).
2σ 2 (2µ2 + σ 2 )
 
aij 8
P. S. Skardal and A. Arenas, “Control of coupled oscillator networks with
(vhw2 i 2 )ij = δij + . (B9)
ki kj application to microgrid technologies,” Sci. Adv. 1, e1500339 (2015).
9
A. Clauset, C. Moore, and M. E. Newman, “Hierarchical structure and the
The first term in the sums is the contribution of the diagonal entries. prediction of missing links in networks,” Nature 453, 98 (2008).
The gaussian variances [σ 2 and 2σ 2 (2µ2 + σ 2 )] and covariance 10
R. Guimerà and M. Sales-Pardo, “Missing and spurious interactions and the
(2µσ 2 ) of a single weight wij drawn from (µ, σ 2 ) are divided by the reconstruction of complex networks,” Proc. Natl. Acad. Sci. 106, 22073–22078
number of elements (the degree ki ) involved in computing the aver- (2009).
11
D. Hric, T. P. Peixoto, and S. Fortunato, “Network structure, metadata, and
ages µi and hw2 ii . The second term accounts for the nondiagonal the prediction of missing nodes and annotations,” Phys. Rev. X 6, 031038
entries. If two nodes (i, j) are neighbors, i.e., aij = 1, then we have (2016).
to add an additional correlation due to the presence of the shared 12
T. Hoffmann, L. Peel, R. Lambiotte, and N. S. Jones, “Community detection in
weight, which is divided by the product of their degrees (ki and kj ). networks without observing edges,” Sci. Adv. 6(4), eaav1478 (2020).
13
For the first-order expansion, we can compute explicitly (δKc )2 M. E. J. Newman, “Analysis of weighted networks,” Phys. Rev. E 70, 056131
(2004).
in terms of the noise parameters (µ, σ ) and the moments of the 14
S. Gómez, A. Arenas, J. Borge-Holthoefer, S. Meloni, and Y. Moreno, “Discrete-
degree distribution. We can write Eq. (5) as time Markov chain approach to contact-based disease spreading in complex
N X N    networks,” Eurphys. Lett. 89, 38009 (2010).
X ∂Kc ∂Kc 15
E. Wigner, “Characteristic vectors of bordered matrices with infinite dimen-
(δKc )2 ≈ (σµ 2 )ij , (B10)
∂µi ∂µ j sions,” Ann. Math. 62, 548–564 (1955).
i=1 j=1 16
M. L. Mehta, Random Matrices (Elsevier Academic Press, 2004).
17
   D. A. Spielman, “Spectral graph theory and its applications,” in Proceedings of
∂Kc ∂Kc the 48th Annual IEEE Symposium on Foundations of Computer Science (IEEE,
+ (σhw2 i 2 )ij , (B11) 2007), pp. 29–38.
∂hw2 ii ∂hw2 ij 18
P. V. Mieghem, Graph Spectra for Complex Networks (Cambridge University
    Press, 2011).
19
∂Kc ∂Kc F. Chung, L. Lu, and V. Vu, “Spectra of random graphs with given expected
+2 (σµ,hw2 i )ij , (B12) degrees,” Proc. Natl. Acad. Sci. 100, 6313–6318 (2003).
∂µi ∂hw2 ij 20
J. G. Restrepo, E. Ott, and B. R. Hunt, “Onset of synchronization in large
and after some algebra, we obtain networks of coupled oscillators,” Phys. Rev. E 71, 036151 (2005).
21
J. G. Restrepo, E. Ott, and B. R. Hunt, “Approximating the largest eigenvalue of
2σ 2 hki 2 network adjacency matrices,” Phys. Rev. E 76, 056119 (2007).
(δKc )2 ≈ [Q − 4µ2 hk2 iQ + 2µ2 (2µ2 + σ 2 )hki2 22
C. Castellano and R. Pastor-Satorras, “Thresholds for epidemic spreading in
NQ4
networks,” Phys. Rev. Lett. 105, 218701 (2010).
+ 2µ4 (hkihk3 i + hk2 i(hk2 i − 4hki) + 2hki2 ) 23
H. H. Ku, “Notes on the use of propagation of error formulas,” J. Res. Nat.
Bureau Stand. C 70C(4), 263–273 (1966).
+ 8µ4 hki(hk2 i − hki)], (B13) 24
G. Mana and F. Pennecchi, “Uncertainty propagation in non-linear measure-
ment equations,” Metrologia 44, 246 (2007).
2 2
where we have used that i j aij ki kj = Nhk i /hki. Simplifying
P P
25
J. P. Gleeson, S. Melnik, J. A. Ward, M. A. Porter, and P. J. Mucha, “Accuracy of
further, we get the resulting Eq. (6) in the main text. Explicitly, mean-field theory for dynamics on real-world networks,” Phys. Rev. E 85, 026106
(2012).
(δKc )2 ≈ a[µ4 (2hkihk3 i 26
A.-L. Barabasi and R. Albert, “Emergence of scaling in random networks,”
Science 286, 509–512 (1999).
− hk2 i2 ) − 2µ2 σ 2 (hkihk2 i − hki2 ) + σ 4 hki2 ], (B14) 27
S. N. Dorogovtsev, J. F. F. Mendes, and A. N. Samukhin, “Structure of
4
growing networks with preferential linking,” Phys. Rev. Lett. 85, 4633–4636
with a = 2σ 2 hki/[N(µ2 hk2 i + σ 2 hki) ]. (2000).
28
S. Valverde, R. Ferrer-i Cancho, and R. V. Sole, “Scale-free networks from
optimal design,” Eurphys. Lett. 60, 512–517 (2002).
REFERENCES 29
D. B. Larremore, W. L. Shew, and J. G. Restrepo, “Predicting criticality and
1
S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, “Critical phenomena in dynamic range in complex networks: Effects of topology,” Phys. Rev. Lett. 106,
complex networks,” Rev. Mod. Phys. 80, 1275–1335 (2008). 058101 (2011).
2 30
M. Newman, Networks: An Introduction (Oxford University Press, Inc., New B. Corominas-Murtra, J. Goñi, R. V. Solé, and C. Rodríguez-Caso, “On the ori-
York, NY, 2010). gins of hierarchy in complex networks,” Proc. Natl. Acad. Sci. 110, 13316–13321
3
R. Pastor-Satorras, C. Castellano, P. Van Mieghem, and A. Vespignani, (2013).
31
“Epidemic processes in complex networks,” Rev. Mod. Phys. 87, 925–979 O. Kinouchi and M. Copelli, “Optimal dynamical range of excitable networks
(2015). at criticality,” Nat. Phys. 2, 348–351 (2006).

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-7

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

32 34
D. R. Chialvo, “Emergent complex neural dynamics,” Nat. Phys. 6, 744–750 S. Jalan and J. Bandyopdhyay, “Random matrix analysis of complex networks,”
(2010). Phys. Rev. E 76, 046107 (2007).
33 35
L. Arola-Fernández, A. Díaz-Guilera, and A. Arenas, “Synchronization invari- C. Sarkar and S. Jalan, “Spectral properties of complex networks,” Chaos
ance under network structural transformations,” Phys. Rev. E 97, 060301 (2018). 28, 102101 (2018).

Chaos 30, 023129 (2020); doi: 10.1063/1.5129630 30, 023129-8

Published under license by AIP Publishing.