Eur. Phys. J.

Special Topics 223, 2649–2664 (2014)

© EDP Sciences, Springer-Verlag 2014 THE EUROPEAN
DOI: 10.1140/epjst/e2014-02284-9 PHYSICAL JOURNAL
SPECIAL TOPICS
Review

Adaptive pinning control: A review of the fully

decentralized strategy and its extensions
L.F.R. Turci1,a , P. De Lellis2 , E.E.N. Macau3 , M. Di Bernardo2 ,
and M.M.R. Simões1
1
Science and Technology Institute, Federal University of Alfenas, Poços de Caldas, Brazil
2
Department of Electrical Engineering and Information Technology, University of Naples
Federico II, Naples, Italy
3
Laboratory of Computing and Applied Mathematics, Instituto National de Pesquisas
Espaciais (INPE), São José dos Campos, Brazil

Received 28 July 2014 / Received in final form 17 October 2014

Published online 10 December 2014

Abstract. In this work, we review recent developments related to the

problem of guiding a complex network of agents toward a synchronized
state. Specifically, we focus on adaptive pinning control strategies, ex-
pounding those developed by the authors in the context of the exist-
ing literature, in which only a small fraction of the network nodes is
directly controlled. The methodologies described herein are adaptive
in the sense that the control and coupling gains are updated on the
basis of the local mismatch with the desired trajectory and between
coupled nodes, respectively. A selection of adaptive strategies recently
proposed in the literature is reviewed, and the main stability results
are expounded. As a numerical validation, the selected approaches are
applied to control an ensemble of coupled mobile agents moving in a
formation.

1 Introduction
In the last decades, much research attention has been focused on synchronization [1–9]
and consensus [10–17] in large ensembles of interacting dynamical agents. The emer-
gence of such collective behaviors can only be explained by analyzing the delicate
interplay between the individual dynamics of the agents and the topology of their in-
terconnections. From these analysis, the concept of complex networks arises. Namely,
each node of the network is a (possibly nonlinear) dynamical agent, while the edges
represent the interconnections among the agents. From a control perspective, a chal-
lenging goal is to steer the dynamics of the network towards a desired synchronous
solution defined a priori. In many real scenarios, directly controlling every node in
the network may be non feasible. This was the main motivation for the development
of the so-called pinning control strategies [17–25], in which only a small subset of
the network nodes is directly controlled. Specifically, an external node, the pinner, is
added to the original network and describes the desired trajectory, say s(t). This node
a
e-mail: [email protected]
2650 The European Physical Journal Special Topics

is connected through unidirectional edges to a (small) fraction of directly controlled

nodes, denoted pinned nodes, thus exerting the control action on a subset of the
network nodes. In classical pinning control schemes, the pinned nodes are controlled
through a proportional feedback action with constant gains [25–27]. The resulting
network equation reads

ẋi (t) = f (xi , t) − c Γ(xi (t) − xj (t)) − δi qΓ(xi (t) − s(t)), (1)
j∈Ni

for all i = 1, . . . , N , with


1, for i = 1, . . . , Ns ,
δi = (2)
0, for i = Ns + 1, . . . , N,

where xi (t) ∈ Rn is the state vector of node i, f : Rn × R+ → Rn the nonlinear

continuous vector field describing the node dynamics, c ≥ 0 a unique global coupling
gain among nodes assumed to be constant and time-invariant, Γ ∈ Rn×n the inner
coupling matrix describing the information exchanged among neighboring nodes, and
Ni the set of neighbors of node i, that is, the set of nodes connected to node i;
Ns is the number of pinned nodes, q the constant control gain and s(t) the desired
synchronous solution to be achieved, which is a solution of the equation ṡ(t) = f (s, t)
describing the pinner’s dynamics.
A key issue here is how to tune the control gain q (and the global coupling gain
c) to make s(t) locally (or globally) asymptotically stable solution for each network
node. In classical pinning control, these gains are constant and set up offline. As
discussed in [28] and [29], there is a critical interdependence between c and q which
is only partially uncovered in the existing literature.
An alternative approach is to adaptively adjust the coupling and control gains thus
avoiding the necessity of an offline tuning of the gains. Adaptive mechanisms can be
found in many natural settings, such as in schools of fish, flocks of birds, and herds
of ungulates which regulate their coordinated movement to avoid obstacles, escape
from predators, or seek for feeding sources [30–39]. In this work, we present a review
of the decentralized strategies for gain adaptation in pinning control. Specifically,
in the recent literature, preliminary attempts of embedding adaptive mechanisms in
pinning control are reported in [40–42], where a unique centralized adaptive law was
designed to adjust the control gain q = q(t), while the suitable coupling gain c still
needed to be computed offline. Subsequently, inspired by the adaptive laws for the
coupling gains developed in [43,44] for synchronization problems, a fully decentralized
approach was proposed in [45]. Differently from previous approaches, i) a control gain
qi , i = 1, . . . , N , is associated to each of the pinned nodes, ii) instead of a unique global
coupling gain c, a coupling gain σij is associated to each pair of network edges. iii)
both the coupling and control gains are adapted in a fully decentralized way. In this
framework, the network equation becomes
N

ẋi (t) = f (xi , t) − c(t) ij (t)Γxj (t) − δi qi (t)Γ(xi (t) − s(t)), (3)
j=1

for i = 1, . . . , N , where qi (t) is the time-varying control gain associated to node i,

ij (t) is the ij-th element of the Laplacian matrix L describing the network topology
and local strength of the coupling. Namely,
⎧
⎨ −σij (t), i = j, i ∈ E,
ij (t) = 0, i = j, i ∈
/ E, (4)
⎩ N
k=1 σij (t) i = j,
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2651

with σij = σji (t) ≥ 0 for all (i, j) ∈ E, where E is the set of network edges. The
idea here is to use a decentralized local strategy in which neighboring systems in the
network interact with each other to adapt their mutual coupling strength σij (t) until
eventually all the agents in the network synchronize. At the same time, the pinner
adapts its control gains qi (t) on the basis of the local output mismatch with each of
the pinned nodes to steer the dynamics of the whole network towards the desired ref-
erence trajectory. Inspired by the work in [45] three hybrid adaptive approaches were
proposed: non-identical reference hybrid pinning strategy [46], network chaos control
hybrid pinning strategy [46], and node-to-node fully adaptive decentralized pinning
strategy [47].
The outline of the paper is the following. In Sect. 2, after giving some mathematical
preliminaries, the control problem tackled in this paper is formulated. Then, in Sect. 3,
a review of the main smooth adaptive strategies for gain adaptation is given, together
with the sufficient conditions guaranteeing the asymptotic convergence of the network
towards the reference trajectory. In Sect. 4, hybrid adaptive derivations of the previous
strategies are illustrated. Then, in Sect. 5, these strategies are numerically validated
on a network of mobile agents that need to synchronize common processes. Finally,
discussion and comments are provided in Sect. 6.

2 Problem formulation
Here, we first give some definition and introduce the notation used in this paper, and
then describe the control problem studied in the paper.

2.1 Mathematical preliminaries

The notation used here is standard in the literature on complex dynamical networks.
Ir denotes the r-dimensional identity matrix, while · denotes the Euclidean norm
and its induced matrix norm. Moreover, given a Hermitian matrix M ∈ Rp×p , we
sort its eigenvalues as λmin (M ) = λ1 (M ) ≤ λ2 (M ) ≤ . . . ≤ λp−1 (M ) ≤ λp (M ) =
λmax (M ). If M is not symmetric, we define its symmetric part as Msym = (M +
M T )/2.
Similarly to what stated in [48, 49], we define now two related classes of vector
fields.
Definition 1. A vector field f : Rn × R+ → Rn is QUAD(H, Γ) with Γ ∈ Rn×n , if
there exists a positive scalar ε and a diagonal matrix H = diag{h1 , . . . , hn } such that

(x − y)T [f (x, t) − f (y, t)] − (x − y)T HΓ(x − y) ≤ −ε(x − y)T (x − y),

∀x, y ∈ Rn , ∀t ∈ R+ . (5)

Definition 2. A vector field f : Rn × R+ → Rn is weakly QUAD(H, Γ) with Γ ∈

Rn×n , if there exists a positive scalar ε and a diagonal matrix H = diag{h1 , . . . , hn }
such that

(x − s(t))T [f (x, t) − f (s, t)] − (x − s(t))T HΓ(x − s(t)) ≤ −ε(x − s(t))T (x − s(t)),
∀x ∈ Rn , ∀t ∈ R+ . (6)

Remark 1. Notice that all the QUAD systems are weakly QUAD, while the vicev-
ersa is not always true. In fact, s(t) is the trajectory of the pinner that is typically
constrained to a bounded subset of Rn .
2652 The European Physical Journal Special Topics

Definition 3. A vector field h(x, t) : Ω ⊆ Rn × R+ → Rn is Lipschitz continuous

if there exists a scalar ς, denoted Lipschitz constant, such that h(x, t) − h(y, t)2 ≤
ς x − y2 for all x, y ∈ Ω.
Lemma 1. [40] Let M be an m-dimensional irreducible matrix whose element
m (i, j) is
denoted by mij . Assume that Rank(M ) = m − 1, mij = mji ≤ 0, and j=1 mij = 0.
Let us consider now a diagonal matrix V = diag(v1 , . . . , vm ) with vi ≥ 0 and
 m
i=1 vi > 0. Then, the eigenvalues of the matrix Z = M + V are all positive.

Lemma 2. ([50], pp. 279-288) The Laplacian matrix L in a connected undirected

network is positive semi-definite. Moreover, it has a simple eigenvalue at 0 and all
the other eigenvalues are positive
Finally, we define M(ψ) the matrix

M(ψ) = L + ψΔ, (7)

where Δ = diag{δ1 , . . . , δN }, and ψ is an arbitrary positive scalar.

2.2 Problem statement

Let G = (N , E) be the graph representing the controlled network, where N is the set
of nodes and E is the set of edges. The control problem that we want to tackle is the
design of the adaptive laws for the coupling gains c(t) and σij (t), (i, j) ∈ E, and the
control gains qi (t), for i = 1, . . . , N , so that network (3) is controlled to the desired
trajectory s(t) described by the pinner. The pinner’s trajectory s(t) ∈ Rn , rooted in
s(0) ∈ Rn , is the reference solution, given by

ṡ(t) = f (s, t). (8)

Note that s(t) can describe a trajectory that converges towards an equilibrium point,
a periodic orbit, or a chaotic attractor.
Definition 4. According to [18, 19, 28, 29], we say that the network (3) is asymptot-
ically controlled to the desired trajectory (8) if

lim xi (t) − s(t) = 0, i = 1, . . . , N.

t→∞

In the next section, we describe the main adaptive strategies that have been proposed
to tackle this control problem.

3 Adaptive pinning strategies

In the following, we review the pinning adaptive strategies presented in [40–42, 45]
and illustrate the main global stability results.

3.1 Constant coupling gains

In [41] and [42], a unique centralized adaptive law was designed to adjust just the
control gain q = q(t), while the suitable coupling gain still needed to be computed
offline. So, constant coupling gains are considered, that is, c(t) = c, and σij (t) = σij
for all t, while an adaptive pinning gain qi (t) is associated to each pinned node.
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2653

Specifically, in [41] each pinning gain is proportional to the unique coupling gain, and
its adaptive law is given by

qi (t) = cki (t), (9)

k˙i (t) = βi eTi (t)Γei (t), (10)

with βi ≥ 0 and ki (0) ≥ 0, for i = 1, . . . , Ns . In [41], the authors gave global stability
results when Γ = I, which can be summarized in the following theorem.
Theorem 1. If Γ = I, the vector field f is QUAD(I, W ), and λmin (M(ψ))I +W > 0
for some ψ > 0, then network (3) is asymptotically controlled to the desired trajectory
(8) under the adaptive law (9)–(10).
The authors of [42] independently derived a similar adaptive law, which can be
viewed as a particular case of (10), with c = 1 and Γ = I. Namely
q̇i (t) = βi eTi (t)ei (t). (11)
The stability analysis also encompassed the case of directed networks (σij = σji ), and
focused on the case in which the vector field can be decomposed as
f (x, t) = Gx(t) + h(x, t), (12)
where G ∈ Rn×n is a norm-bounded matrix, that is, G2 ≤ ν, and h : Ω ⊆ Rn ×
R+ → Rn is a Lipschitz continuous vector field. After defining γ := Γ2 > 0, we can
give the stability results presented in [42].
Theorem 2. ([42], p. 998) If i) the vector field f can be decomposed as in (12),
with h being Lipschitz with constant ς, ii) there exist a natural number 1 ≤ Ns ≤ N
satisfying λNs +1 (cLsym + ψΔ/2) < −(ν + ς)/γ), for some ψ > 0, then network (3) is
asymptotically controlled to the desired trajectory (8) under the adaptive law (11).
We observe that this approach allows to deal with directed network but, while
qi (t) is adaptively selected, the coupling gains c and σij , and the minimum number
of pinned nodes Ns need to be estimated off-line.

3.2 Centralized gain adaptations

A totally adaptive strategy has been proposed in [40]. The the authors adapt the
coupling and control gains based on a centralized measure of the error vector e,
whose i-th component is defined as
ei (t) = xi (t) − s(t). (13)
Namely, the global coupling strength c is adapted as
N
ρ T
ċ(t) = e (t)P ei (t), (14)
2 i=1 i

with ρ > 0, c(0) ≥ 0, and the control gain is selected accordingly as

qi (t) = q(t) = κc(t), ∀i = 1, . . . , Ns , (15)
where κ > 0.
Before giving the main result presented in [40], we define the matrix A = −L and
its left eigenvalues ξ1 , . . . , ξN . Also, we define Ξ = diag{ξ1 , ..., ξN }.
2654 The European Physical Journal Special Topics

Theorem 3. ([40], p. 1322) If i) A is irreducible with rank(A) = N −1, ii) the vector
field f is QUAD(H, I), then network (3) is asymptotically controlled to the desired
trajectory (8) under the adaptive law (14) and (15) with P = H, for any Ns ≥ 1.

Here, we remark that this approach allows to steering the network nodes towards
the desired trajectory by directly controlling only one node in the network, that is,
for Ns = 1. However, the adaptation of the unique coupling and control gains c(t)
and q(t), respectively, requires the knowledge of all the nodes’ states, thus making
the approach centralized.

3.3 Fully decentralized adaptive pinning

Differently from the approaches illustrated above, in [45] the authors proposed a
completely decentralized adaptive approach, in which all the coupling gains σij (t),
(i, j) ∈ E, and the control gains qi (t), i = 1, . . . , Ns , are adapted based on local
information exchanged among neighboring nodes.
Inspired by the edge-based strategy [43], the local coupling strength σij (t) between
linked nodes is adaptively evolved according to the following strategy:

σ̇ij (t) = αij (xi (t) − xj (t))T Γ(xi (t) − xj (t)), (16)

for (i, j) ∈ E, where σij (0) = σji (0) ≥ 0, and αij is a positive scalar, thus making the
network topology undirected for all t. Similarly, the control gain of the i-th pinned
node is adapted on the basis of the local mismatch with the pinner’s trajectory as
in [41]. Namely,
q̇i (t) = di ei (t)T Γei (t), (17)
for i = 1, . . . , Ns , where di is a positive scalar. The combined adaptation of the
coupling and control gains allows the achievement of the desired trajectory. Specifi-
cally, the edge-based adaptive coupling strategy [43, 44] guarantees synchronization of
the network nodes, while the pinning control action steers the synchronous solution
towards the desired reference trajectory s(t).
Here, we further discuss the stability properties of this fully decentralized ap-
proach, and give sufficient conditions for asymptotic controllability of network (3),
extending the preliminary stability results illustrated in [45].
Theorem 4. If the vector field f is weakly QUAD(H, Γ), under the fully adaptive
strategy (16)–(17), network (3) is asymptotically controlled to the desired trajectory
(8), for any Ns ≥ 1. Moreover, the coupling strengths σij (t), (i, j) ∈ E, and the control
gains qi (t), i = 1, . . . , Ns , asymptotically converge to finite steady-state values.

Proof. In the following, we omit the explicit time dependence for the sake of brevity.
The network Eq. (3) can be recast as follows
N

ėi = f (xi , t) − f (s, t) − c(t) ij (t)Γej (t) − δi qi (t)Γei (t) (18)
j=1

for i = 1, . . . , N .
Consider the following candidate Lyapunov function
N Ns
1 T  γij  ηi
V (e, σij , qi ) = ei ei + (σij − σ̄ij )2 + (qi − q̄i )2 . (19)
2 i=1 (i,j)∈E
2 i=1
2
j>i
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2655

The derivative of V along the trajectory of the error system (18) is

⎡ ⎤
N
 N

V̇ = ⎣eTi (f (xi ) − f (s)) − ceTi ij Γej − δi qi eTi Γei ⎦
i=1 j=1

 Ns

+ (σij − σ̄ij )γij αij (ei − ej )T Γ(ei − ej ) + ηi βi (qi − q̄i )eTi Γei .
(i,j)∈E i=1
j>i

Noting that
N
 N
 
eTi ij Γej = σij (ei − ej )T Γ(ei − ej ),
i=1 j=1 (i,j)∈E
j>i

and as f is QUAD(Γ), we have

N N  Ns
V̇ ≤ i=1 eTi Zei − ε i=1 eTi ei − c (i,j)∈E σij (ei − ej )T Γ(ei − ej ) − i=1 qi eTi Γei +
 j>i
T
Ns T
(i,j)∈E (σij − σ̄ij )γij αij (ei − ej ) Γ(ei − ej ) +
j>i
i=1 ηi βi (qi − q̄i )ei Γei ,

where Z = HΓ.
1 c
If we choose ηi = βi and γij = αij , we get

N
 N
  Ns

V̇ ≤ eTi Zei − ε eTi ei − c σ̄ij (ei − ej )T Γ(ei − ej ) − q̄i eTi Γei . (20)
i=1 i=1 (i,j)∈E i=1
j>i

If we define the stack error vector as e = [eT1 , . . . , eTN ]T , we observe that


σ̄ij (ei − ej )T Γ(ei − ej ) = eT (L̄ ⊗ Γ)e, (21)
(i,j)∈E
j>i

where L̄ is a Laplacian-like matrix, whose ij th element ¯ij is defined as

⎧
⎪
⎪ σ̄ij , i = j, i ∈ E,
⎪
⎨ 0, i = j, i ∈/ E,
¯ij =  N
⎪
⎪−
⎪
⎩ σ̄ik i = j.
k=1

Combining (20) and (21), we have

V̇ ≤ eT (IN ⊗ Z)e − εeT e − eT (Q̄ ⊗ Γ)e − ceT (L̄ ⊗ Γ)e ≤


eT (IN ⊗ Z)e − εeT e − eT (Q̄ + cL̄) ⊗ Γ e,

where Q̄ = diag{q̄1 , . . . , q̄Ns , 0, . . . , 0} is an N -dimensional diagonal matrix.

From matrix algebra, we have
V̇ ≤ [λmax (Zsym ) − λmin (Q̄ + cL̄)]eT (IN ⊗ Γ)e − εeT e,

where Zsym = (Z +Z T )/2. From Lemma 2, for any Ns ≥ 1, we can make λmin (Q̄+cL̄)
arbitrarily large by appropriately choosing the scalars q̄i for i = 1, . . . , Ns and σ̄ij for
2656 The European Physical Journal Special Topics


(i, j) ∈ E. Therefore, λmax (Zsym ) − λmin (Q̄ + cL̄) can be made negative, and we
finally have
V̇ ≤ −εeT e := −W(e(t)) ≤ 0.
Hence, applying LaSalle-Yoshizawa Theorem (see [51], page 492), we conclude that the
error dynamics, the coupling strengths and the control gains are globally uniformly
bounded. Moreover, limt→∞ W(e(t)) = 0, which implies that:
limt→∞ e(t) = 0.
As we proved that the coupling strengths and control gains are bounded, and being
monotone increasing by definition (see Eqs. (16) and (17)), we can conclude that they
converge to finite steady-state values.
Here, we wish to remark that the above stability result is quite general. Most of
the results on global stability obtained in the study of adaptive pinning control of
complex networks can be seen as particular cases of Theorem 4. To emphasize this
point, we give the following corollaries.
Corollary 1. If the vector field f is QUAD(Γ), with Γ = In and H − εIn ≤ 0,
under the adaptive pinning control strategy given by (16) and (17), network (3) is
asymptotically controlled to the desired trajectory (8), for any Ns ≥ 1. Moreover,
the coupling strengths σij (t), (i, j) ∈ E, and the control gains qi (t), i = 1, . . . , Ns ,
asymptotically converge to finite steady-state values.
Proof. The thesis trivially follows from Theorem 4.
Corollary 2. If the vector field f is globally Lipschitz continuous and the inner cou-
pling matrix Γ is positive definite, then, under the adaptive pinning control strategy
given by (16) and (17), network (3) is asymptotically controlled to the desired tra-
jectory (8), for any Ns ≥ 1. Moreover, the coupling strengths σij (t), (i, j) ∈ E, and
the control gains qi (t), i = 1, . . . , Ns , asymptotically converge to finite steady-state
values.
Proof. Notice that, as illustrated in the Appendix, if f is globally Lipschitz con-
tinuous, then it is also QUAD(Γ) for any Γ > 0. Hence, the thesis follows from
Theorem 4.
Here, it is interesting to remark that the combined adaptation of the coupling
strengths and control gains allows to control Lipschitz systems to the desired trajec-
tory: when just the control gains are adapted, the network is asymptotically controlled
only when some inequalities on the eigenvalues of the Laplacian matrix are satisfied,
see for instance the results in [42].
We also point out that, in its general form, Theorem 4 does not require that the
oscillators are coupled through all the state variables. Moreover, differently from the
existing literature (see for instance [42]), the adaptation of the gain is based only on
the available system outputs, that is, Γxi , which do not necessarily correspond to
all the state vector. This is very relevant in the case in which the state cannot be
estimated through the output vector. Furthermore, we also emphasize that through
the combined adaptation of the coupling strengths and control gains it is possible to
control the network by pinning one node, while multiple pinning allows reducing the
intensity of the coupling, as illustrated numerically in Sect. 5.

4 Extensions of fully decentralized adaptive pinning

In this section we present three different strategies derived from Fully Decentralized
Adaptive Pinning: adaptive node-to-node pinning strategy, non-identical reference
hybrid pinning strategy, and network chaos control hybrid pinning strategy.
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2657

4.1 Adaptive node-to-node pinning

The problem of selecting the type of nodes that should be pinned to optimize the syn-
chronization performance [52] by minimizing the synchronization time, and driving
the system to a desired state is a very hard problem. However, a node-to-node pinning
strategy has been proposed by Porfiri et al. [53] to increase synchronization perfor-
mance, i.e., reduce synchronization time. In node-to-node pinning control just one
node is pinned, and at each instant of time multiple of the switching period T , a new
node is randomly chosen to be pinned.
In [47], the authors combined the node-to-node selection of the pinned nodes with
the modulation of the coupling and control gains illustrated in Sect. 3.3. Differently
from the fully decentralized adaptive pinning, the adaptive node-to-node pinning
control considers that Mpin nodes of the network are “pinnable”. Among this set
of nodes, Ns < Mpin are randomly chosen and effectively pinned at every period
[kT (k + 1)T ], k = 0, 1, . . .. Namely, in the network model (3), δi is zero for the non-
pinnable nodes, while, if i ∈ Mpin , it is described by the following binomial probability
function

P (δi = 1) = Ns /Mpin
P (δi = 0) = 1 − Ns /Mpin

for i ∈ Mpin , with the constraint


δi = Ns .
i∈Mpin

Due to the discontinuity induced by the periodic switching of the pinned nodes, the
stability of the error dynamics was proven under stricter assumption on the nodes’ dy-
namics. Specifically, in [47] the authors assumed that the vector field f is QUAD(H, Γ)
with H − I < 0 and Γ = I, while the stability of the classical fully decentralized
adaptive strategy only requires f to be weakly QUAD. However, the statistical analy-
sis presented in [47] has shown that, when the same number of nodes are pinned,
the adaptive node-to-node pinning strategy performs better than fully decentralized
adaptive pinning strategy in terms of convergence time.

4.2 Application to chaos control

The solution s(t) of (8) may describe a trajectory converging towards an equilibrium
point, a periodic orbit, or a chaotic attractor. Notice that, when the pinner dynamics
are chaotic, Unstable Periodic Orbits (UPOs) exist. In principle, an UPO may be
used as the reference trajectory s(t). However, as such periodic orbit is unstable,
the presence of noise and the finite precision effect imply that synchronization is not
stable as the reference node is not able to keep its trajectory on the unstable orbit
without a control action. In the literature, the stabilization of the periodic orbit can
be accomplished through control of chaos [54] approach.
In [46] the authors considered the problem of controlling the network towards
an UPO. Specifically, they combined the fully decentralized pinning control strategy
presented in Sect. 3.3 with the a chaos control strategy applied on the pinner. As
a result, the network dynamics are steered towards the stabilized desired periodic
trajectory.
Also, we emphasize that, the number of UPOs embedded in a chaotic orbit is
infinite although countable. This is a desirable property in applications in which
2658 The European Physical Journal Special Topics

different synchronous trajectories are applied. For example, in multitasking (multiple

agents) area exploration or a surveillance screening tasks, chaotic trajectories imply
in very efficient reference trajectories. However, as soon as one or more targets are
identified, a convenient approach strategy is to change agents trajectory to make them
follow periodic orbits around the pinpoint positions, allowing a better discrimination.

4.3 Nonidentical hybrid pinning

Notice that, without loss of generality, Eq. (8) can be rewritten as

ṡ(t) = Gs(t) + h(s, t).

In fact, the nonlinear vector field f (s, t) can always be written as the sum of a linear
and a nonlinear component. In [46], the authors also tackled the problem of pinning
control when the reference trajectory described by the pinner is not a solution of (8).
In other words, the authors extended the fully adaptive decentralized adaptive pinning
approach of Sect. 3.3 to the case in which the pinner dynamics is different from the
dynamics of the nodes of the controlled network. Specifically, the pinner trajectory
was described by
ṡ(t) = f (s, t) = Gs(t) + kh(s, t), (22)
k ∈ R, s(0) = s0 ∈ Rn . When k = 1, the network nodes will not converge towards
the desired trajectory, and a finite pinning error would remain, thus implying the
divergence of the control gains, see Eq. (17).
To avoid the divergence of qi (t), a saturation is introduced: its adaptation law is
modified as
q̇i (t) = di π(ei )ei (t)T Γei (t) with  ∈ R+ , (23)
as [46] where π(ei , t) is defined as

1, if ei (t) > ,
π(ei , t) = (24)
0, otherwise.

Before stating the main stability result, we give the following definition.
Definition 5. We say that the network (3) is asymptotically bounded controlled to
the desired trajectory (8) if

lim xi (t) − s(t) ≤ , 1 i = 1, . . . , N. (25)

t→∞

Theorem 5. ([42], p. 998) If the vector field f can be decomposed as in (12), with
h being Lipschitz, and Γ = I, then network (3) is asymptotically bounded controlled
to the desired trajectory (8) under the adaptive laws (16) and (23). Moreover, the
coupling strengths σij (t), (i, j) ∈ E, and the control gains qi (t), i = 1, . . . , Ns , asymp-
totically converge to finite steady-state values.

Proof. The thesis is proved using the same Lyapunov function given in (19), see [46],
pages 4–6, for details.
1
Here, the use of the symbol lim is not intended in the classical sense of limit. By (25),
we mean that for all ν > 0 there exists a tν > 0 such that for all t > tν we have that
x(t) − s(t)2 ≤  + ν. We remark that this does not imply the existence of the limit in a
classical sense.
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2659

We also observe that, while the pinning error remains bounded, but does not
converge to zero, the state of all the controlled nodes will asymptotically converge.
Namely, in [46] it is also shown that

lim xi (t) − xj (t) = 0,

t→∞

for all (i, j) ∈ E.

5 Application
Communication security using chaos has been extensively addressed in the litera-
ture [55–57]. Researchers have shown that communication efficiency for specific ap-
plications can be assured by a chaos-based secure communication scheme in which
signal is encrypted by the addition of a chaotic signal generated in the transmit-
ter [58, 59]. The encrypted signal is communicated through a communication channel
from the transmitter side to the receiver side. At the receiver, demodulation is per-
formed and the obtained signal is the decrypted signal. The chaotic signal generated
by the transmitter and the chaotic signal generated by the receiver must be equal,
i.e., signals must be synchronous.
Now, consider as a reference application an ensemble of N + 1 autonomous planar
moving platforms design to accomplish an exploration task in which each platform
has to securely communicate with a common receiver. In order to use chaos-based se-
cure communication scheme it must be guaranteed the synchronization of the chaotic
signals generated in each platform (transmitter) and the common receiver. The syn-
chronization can be achieved by the use of pinning control strategies, particularly,
fully decentralized adaptive pinning strategy.
Consider a Chua’s chaotic circuit is installed on each platform. The receiver, iden-
tified with a subscript s according to the notation used in this paper, defines the
reference trajectory s(t), which is described by the Chua’s circuit installed on it.
The vector field describing each Chua’s circuit is given by:
⎛ ⎞⎛ ⎞ ⎛ ⎞
−c1 c1 0 xi1 (t) −c1 g(xi1 )
f (xi ) = ⎝ 1 −1 1 ⎠ ⎝ xi2 (t) ⎠ + ⎝ 0 ⎠, (26)
0 −c2 −c3 xi3 (t) 0

where xi = [xi1 xi2 xi3 ] is the state of the i-th Chua’s circuit, i = 1, . . . , N , and
g(xi1 ) = m0 xi1 + 12 (m1 − m0 )(|xi1 + 1| − |xi1 − 1|). The parameters of the circuit are
set as c1 = 10, c2 = 15, c3 =, m0 = 0.68, m1 = 1.27. When these values are selected,
the system exhibits chaotic behavior [60].
In order to proceed the exploration task each moving platform follows a random
walk trajectory given by:

yij (t + Δ) = yij (t) + vij (t)Δt
(27)
θij (t) = φij (t + Δt)

where a subscript i identifies the i-th platform, i = 1, . . . , N, s, while a subscript

j identifies the j-th planar coordinate, j = 1, 2; moreover, Δt is a time variation
sufficiently small to guarantee network synchronization under fast topology switch-
ing [61–63], vij (t) is the velocity with modulus v and phase θij (t), while φij (t) is a
random variable uniformly distributed in [−π, π] [64].
In this specific application, platforms ensemble form an active network in the
sense that platforms can communicate with each other. The network topology is time
2660 The European Physical Journal Special Topics

Agents Motion
1.5

0.5

y (t)
2 −0.5

−1

−1.5

−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
y (t)
1

Fig. 1. Mobile Platforms Random Walk yi (t): figure shows mobile platforms trajectories
in the coordinate plane under random walk motion behavior, according to Eq. (27). Each
different color represents one different mobile platform, and the circle represent its initial
positions.

varying, as the presence of an edge connecting two platforms (nodes) i and j depends
on proximity rules: at any time instant t, there is an edge if and only if the Euclidean
distance dij (t) = yi (t) − yj (t) between the two moving platforms is below a given
threshold r. Accordingly, the Laplacian matrix L describing the network topology
and local strength of the coupling becomes
⎧
⎪
⎪ −σij (t), i = j, dij (t) < r,
⎨
ij (t) = 0, i = j, dij (t) ≥ r, (28)
⎪
⎪
⎩ N
k=1 σij (t) i = j.

where σij (t) is adaptively adjusted according (16).

The control aim is to synchronize all the chaotic circuits towards s(t) applying
the fully decentralized adaptive pinning strategy presented in Sect. 3.3. Under the
synchronability conditions for time-varying topology networks [61–63], by applying
the fully decentralized adaptive pinning strategy, all the circuits’ dynamics can be
synchronized to a desired dynamics (reference trajectory) while each mobile platform
follows its random trajectory.
In the selected numerical scenario, we consider a network of 10 + 1 moving plat-
forms. The total number of pinned nodes, Ns , is set to 1 (one), and the pinned node
is randomly selected among the platforms. The control gain q(t) that defines the in-
tensity of the control action under the pinned node is adaptively adjusted according
Eq. (17).
The initial conditions for the 10 controlled circuits are randomly selected from a
normal distribution with mean 0 and standard deviation 1, so the initial conditions
for the pinning node process (reference platform). The value of αij is set to 1 any
(i, j) ∈ E, βi = 1 any i ∈ (1, N ), qi (0) is set to zero for any i ∈ (1, Ns ), and σij (0)
are set to zero any (i, j) ∈ E. As for the motion of the platforms, the parameters of
(27) are set as r = 2, v = 1 and Δt = 0.1, while the initial coordinates are randomly
selected from an uniform distribution in the interval [−3, 3].
The platforms random walk trajectories in the coordinate plane are illustrated in
Fig. 1. Each different color represents one different platform. One can observe that
distances among nodes are continuously changing and consequently so will be the
network topology according to (28).
As illustrated in Fig. 2 all the circuit’s state synchronize asymptotically converg-
ing to the reference trajectory, that is, all the signals generated in the platforms
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2661

Circuits’ State Variables

5

x (t)
0

1
−5
0 2 4 6 8 10
2

x (t)
0
2
−2
0 2 4 6 8 10
5
x (t)

0
3

−5
0 2 4 6 8 10
Units of time

Fig. 2. Processes Synchronous State: evolution of each state variable xj (t) of the chaotic
process of each platform i (network node). Observe states synchronize within each other
and with the reference chaotic trajectory. In the figure, each different color represents one
different platform process (Chua circuit), and the black curve is the reference trajectory.

Coupling Gains
2

1.5
σij(t)

0.5

0
0 2 4 6 8 10
Units of time
Fig. 3. Coupling Strength σij (t): evolution of each coupling strength σij (t) of the network
edges (links) formed among the platforms as they interact. Observe each coupling strength
converges to a bounded value as time evolves. In the figure, each different color represents
one different coupling strength σij (t) between two platforms.

Chua’s circuit synchronize with the signal generated at the receiver. In the figure,
each different color represents one different platform circuit.
Moreover, both the pinning and control gains asymptotically converge towards
finite steady-state values, as depicted in Figs. 3 and 4, guaranteeing limited energy
control action. In the Fig. 3 each different color represents one different coupling
strength σij (t) between two platforms.

6 Conclusion
In this paper, we reviewed the main works on pinning control with adaptive coupling
and pinning gains. Specifically, we illustrated three main approaches.
The first approach is decentralized, but consider the case of constant coupling
gains that are offline estimated to guarantee network synchronization [41, 42]. The
control gains, instead, were adaptively evolved based on the local pinning error ei to
guarantee pinning synchronization. In the second approach, both the coupling and
2662 The European Physical Journal Special Topics

Control Gain
4

qi(t)
2

0
0 2 4 6 8 10
Units of time

Fig. 4. Pinning Control Gain q(t): evolution of the pinning control gain q(t) of the
pinned node. Observe the control gain q(t) converges to a bounded value as time evolves.

control gains are adapted, but in a centralized fashion [40]. In fact, their evolution
is guided by a measure of the network error, thus requiring the knowledge of all the
nodes’ states. For these first two approaches, global stability results were provided
for the case of nodes’ dynamics described by Lipschitz or QUAD vector fields.
In the third approach instead, named fully adaptive pinning control [45], both the
coupling and control gains are adapted in a fully decentralized fashion. We illustrated
how the network can be asymptotically driven towards the pinner trajectory under
milder assumptions on the nodes’ trajectory. Specifically, the nodes’ vector fields have
to fulfil the so-called weakly QUAD assumption. Compared with the other approaches,
we observed that the effectiveness of fully adaptive decentralized adaptive pinning
class of function is proved for a wider class of systems, as the weakly QUAD condition
implies the QUAD condition, as well as the Lipschitz condition. The stability analysis
also encompassed the case in which the inner coupling matrix Γ is nondiagonal, and
not all the nodes’ states can be used to adapt the coupling and control gains. Moreover,
we showed that the network can be controlled by pinning only one node.
To further illustrate the versatility of the fully adaptive decentralized pinning
strategy, we also reviewed three related adaptive strategies directly derived from this
approach. The first one is the non-identical reference hybrid pinning strategy [46], in
which the pinner dynamics is different from the dynamics of the nodes of the con-
trolled network. The second is the network chaos control hybrid pinning strategy [46],
in which we combine pinning synchronization control, particularly fully adaptive de-
centralized pinning synchronization control, with chaos control technique in order to
synchronize a complex network not only to an equilibrium point, stable periodic orbit
or chaotic orbit, but also to any unstable periodic orbit embedded in the chaotic at-
tractor. Finally, the adaptive node-to-node pinning strategy [47], in which the pinned
node is periodically randomly changed.
In addition we have shown an application of fully adaptive decentralized pinning
control to the synchronization of chaotic circuits mounted on a time-varying network
of moving platforms.

The works reviewed in this article were sponsored by FAPESP, FAPEMIG and CAPES.
EENM is partially supported by CNPq and FAPESP (2011/50151-0).

Appendix
Theorem 6. If f is globally Lipshitz continuous, with Lipschitz constant equal to α,
then f is QUAD(Δ, ω̄), with Δ − ω̄I ≥ αI.
Proof. Being f Lipschitz, we have:
f (x, t) − f (y, t) ≤ αx − y. (29)
 Advanced Computational and Experimental Techniques in Nonlinear Dynamics 2663

It is also true that,

x − yf (x, t) − f (y, t) ≤ αx − y2 , (30)

or equivalently,

x − yf (x, t) − f (y, t) ≤ (x − y)T αI(x − y). (31)

From basic algebra and using the Cauchy-Schwarz inequality, we can write:

(x − y)T [f (x, t) − f (y, t)] ≤ x − yf (x, t) − f (y, t). (32)

Hence, we get,
(x − y)T [f (x, t) − f (y, t)] ≤ (x − y)T αI(x − y). (33)
Thus, from (33) and (5), with αI = HΓ − εI, follows the thesis.

References
1. I. Belykh, I. Belykh, M. Hasler, Chaos 16, 015102 (2006)
2. I. Belykh, M. Hasler, M. Lauret, H. Nijmeijer, Int. J. Bif. Chaos 15, 3423 (2005)
3. H. Gao, X. Meng, T. Chen, J. Lam, SIAM J. Cont. Optimiz. 48, 3643 (2010)
4. H.J.C. Huijberts, H. Nijmeijer, R.M.A. Willens, Int. J. Robust Nonlinear Contr. 10, 363
(2000)
5. J. Lü, X. Yu, G. Chen, Physica A 51, 787 (2004)
6. H. Nijmeijer, Physica D 154, 219 (2001)
7. L.M. Pecora, T.L. Caroll, Phys. Rev. Lett. 64, 821 (1990)
8. C.W. Wu, IEEE Trans. Autom. Contr. 51, 1207 (2006)
9. C.W. Wu, L.O. Chua, IEEE Trans. Circ. Syst. I 42, 430 (1995)
10. R.M. Murray, R. Olfati-Saber, IEEE Trans. Autom. Contr. 49, 1520 (2004)
11. M. Cao, A.S. Morse, B.D.O. Anderson, SIAM J. Contr. Optimiz. 47, 575 (2008)
12. B. Kozma, A. Barrat, Phys. Rev. E 77, 016102 (2008)
13. B. Liu, D.J. Hill, SIAM J. Contr. Optimiz. 49, 315 (2011)
14. R. Olfati-Saber, J.A. Fax, R.M. Murray, Proc. IEEE 95, 215 (2007)
15. W. Wang, J.-J.E. Slotine, IEEE Trans. Autom. Control 51, 1156 (2006)
16. P. DeLellis, M. diBernardo, F. Garofalo, D. Liuzza, Appl. Math. Comput. 217, 988
(2010)
17. M. Huang, J.H. Manton, SIAM J. Contr. Optimiz. 48, 134 (2009)
18. X. Li, X. Wang, G. Chen, IEEE Trans. Circ. Syst. I 51, 2074 (2004)
19. W. Yu, G. Chen, J. Lü, Automatica 45, 429 (2009)
20. X. Wang, G. Chen, Physica A 310, 521 (2002)
21. A. Rahmani, M. Ji, M. Mesbahi, M. Egerstedt, SIAM J. Contr. Optimiz. 48, 162 (2009)
22. P. DeLellis, M. di Bernardo, M. Porfiri, Chaos 21, 033119 (2011)
23. P. DeLellis, M. di Bernardo, F. Garofalo, IEEE Trans. Circ. Syst. I 60, 3033 (2013)
24. V. Mwaffo, P. DeLellis, M. Porfiri, Chaos 24, 013101 (2014)
25. L. Xiang, J.J.H. Zhu, Nonlinear Dyn. 64, 339 (2011)
26. L. Xiang, Z. Chen, Z. Liu, F. Chen, Z. Yuan, J. Phys. A 40, 14369 (2007)
27. L. Xiang, Z.X. Liu, Z.Q. Chen, F. Chen, Z.Z. Yuan, Physica A 379, 298 (2007)
28. F. Sorrentino, M. di Bernardo, F. Garofalo, G. Chen, Phys. Rev. E 75, 046103 (2007)
29. M. Porfiri, M. di Bernardo, Automatica 44, 3100 (2008)
30. J. Cortes, S. Martinez, F. Bullo, IEEE Trans. Autom. Contr. 51, 1289 (2006)
31. M. Lindhe, P. Ogren, K.H. Johansson, Flocking with obstacle avoidance: A new dis-
tributed coordination algorithm based on Voronoi partitions. In Proceedings of the 2005
IEEE International Conference on Robotics and Automation (2005), p. 1785
32. R. Olfati-Saber, IEEE Trans. Autom. Contr. 51, 401 (2006)
2664 The European Physical Journal Special Topics

33. I.D. Couzin, J. Krause, N.R. Franks, S.A. Levin, Nature 433, 513 (2005)
34. P. DeLellis, M. Porfiri, E.M. Bollt, Phys. Rev. E 87, 022818 (2013)
35. B. Nabet, N.E. Leonard, I.D. Couzin, S.A. Levin, J. Nonlinear Sci. 19, 399 (2009)
36. D.A. Paley, N.E. Leonard, R. Sepulchre, D. Grunbaum, J.K. Parrish, IEEE Contr. Syst.
Mag. 27, 89 (2007)
37. S. Ghosh, G. Rangarajan, S. Sinha, EPL 92, 40012 (2010)
38. J. Yu, C. Hu, H. Jiang, Z. Teng, Neurocomputing 74, 1776 (2011)
39. P. DeLellis, G. Polverino, G. Ustuner, N. Abaid, S. Macrı̀, E.M. Bollt, M. Porfiri,
Scientific Reports 4, 3723 (2014)
40. T. Chen, X. Liu, W. Lu, IEEE Trans. Circ. Syst. I 54, 1317 (2007)
41. L. Wang, H.P. Dai, H. Dong, Y.Y. Cao, Y.X. Sun, Eur. Phys. J. B 61, 335 (2008)
42. J. Lu J. Zhou, J. Lü, Automatica 44, 996 (2008)
43. P. De Lellis, M. di Bernardo, F. Garofalo, Chaos 18, 037110 (2008)
44. P. DeLellis, M. diBernardo, F. Garofalo, Automatica 45, 1312 (2009)
45. P. DeLellis, M. di Bernardo, L.F.R. Turci, Fully adaptive pinning control of complex net-
works. In Proceedings of 2010 IEEE International Symposium on Circuits and Systems
(ISCAS) (2010), p. 685
46. L.F.R. Turci, E.E.N. Macau, Int. J. Bif. Chaos 22, 1 (2012)
47. L.F.R. Turci, E.E.N. Macau, Chaos 22, 033151 (2012)
48. T. Chen, X. Liu, Network synchronization with an adaptive coupling strength,
http://arxiv.org/abs/math/0610580v1 (2006)
49. P. DeLellis, M. diBernardo, G. Russo, IEEE Trans. Circ. Syst. I 58, 576 (2011)
50. C.D. Godsil, G. Royle, Algebraic Graph Theory (Springer, 2001)
51. M. Krstic, I. Kanellakopoulos, P. Kokotovic, Nonlinear and Adaptive Control Design
(John Wiley and Sons, 1995)
52. L.F.R. Turci, E.E.N. Macau, Phys. Rev. E 84, 011120 (2011)
53. M. Porfiri, F. Fiorilli, Chaos 19, 013122 (2009)
54. E. Ott, C. Grebogi, J.A. Yorke, Phys. Rev. Lett. 64, 1196 (1990)
55. K. Fallahi, H. Leung, Comm. Nonlinear Sci. Numer. Simul. 15, 368 (2010)
56. N. Singh, A. Sinha, Opt. Lasers Eng. 48, 398 (2010)
57. O.I. Moskalenko, A.A. Koronovskii, A.E. Hramov, Phys. Lett. A 374, 2925 (2010)
58. M. Eisencraft, R.D. Fanganiello, J.M.V. Grzybowski, D.C. Soriano, R. Attux, A.M.
Batista, E.E.N. Macau, L.H.A. Monteiro, J.M.T. Romano, R. Suyama, T. Yoneyama,
Comm. Nonlinear Sci. Numer. Simul. 17, 4707 (2012)
59. J.V. Grzybowski, E.E. Neher Macau, T. Yoneyama, J. Phys. A: Math. Theor. 44, 175103
2011
60. P. Bartissol, L.O. Chua, IEEE Trans. Circ. Syst. 35, 1512 (1988)
61. M. Frasca, A. Buscarino, A. Rizzo, L. Fortuna, S. Boccaletti, Phys. Rev. Lett. 100,
044102 (2008)
62. D.J. Stilwell, E.M. Bollt, D.G. Roberson, SIAM J. Appl. Dyn. Syst. 5, 140 (2006)
63. V.N. Belykh, I.V. Belykh, M. Hasler, Physica D: Nonlinear Phenom. 195, 159 (2004)
64. L. Wang, Y.-X. Sun, J. Stat. Mech.: Theory Exper. 2009, P11005 (2009)