3342 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO.

5, MAY 2022

Synchronization of Complex Networks With

Nondifferentiable Time-Varying Delay
Shuaibing Zhu , Student Member, IEEE, Jin Zhou , Xinghuo Yu , Fellow, IEEE, and Jun-An Lu

Abstract—In this article, we investigate the synchronization of complex in real situations. However, the basic problem was neglected in most
networks with general time-varying delay, especially with nondifferen- existing studies, which is to be considered in this article.
tiable delay. In the literature, the time-varying delay is usually assumed During the communications among the interconnected systems of
to be differentiable. This assumption is strict and not easy to verify in
engineering. Until now, the synchronization of networks with nondifferen- a network, the coupling delay is hardly negligible due to the finite
tiable delay through adaptive control remains a challenging problem. By transmission speed of the signals. The synchronization problem of
analyzing the boundedness of the adaptive control gain and extending networks with differentiable time-varying delay has been extensively
the well-known Halanay inequality, we solve this problem and estab- studied [19]–[28]. In [19]–[26], it is assumed that the time-varying
lish several synchronization criteria for networks under the centralized
delay τ (t) is differentiable and satisfies
adaptive control and networks under the decentralized adaptive control.
Particularly, the boundedness of the centralized adaptive control gain is τ̇ (t) ≤ μ1 < 1
theoretically proved. Numerical simulations are provided to verify the
theoretical results. which is one of the most widelyused assumptions. In [27] and [28],
Index Terms—Adaptive control, complex network, nondifferentiable this assumption is relaxed to τ̇ (t) ≤ μ2 . In reality, the delay may
time-varying delay, synchronization. change in a random way that is unpredictable, so it is hardly possible
to know the analytical expression of the delay, making it diffi-
cult to verify whether the delay is differentiable. When the control
is linear, the synchronization problem of networks with nondiffer-
I. I NTRODUCTION entiable delay is generally solved in virtue of delay differential
Complex networks, generally consisting of numerous inequalities [29], [30]. However, dealing with the adaptive synchro-
interconnected systems, have drawn remarkable attention in nization of networks with nondifferentiable delay is still a challenging
the past two decades [1]–[10]. In network science, one of the most problem. In 2008, the adaptive robust convergence of neural networks
important problems is the synchronization among the interconnected with nondifferentiable delays was addressed in [31], whereas the
systems of a network. control therein is actually not adaptive. In 2013, the adaptive synchro-
A great deal of networks cannot realize synchronization by them- nization of networks with nondifferentiable delay was investigated
selves, so external control is generally inevitable. In the literature, in [32]. We find that the control therein is complicated and the
two kinds of control are widely used for realizing synchronization, LaSalle invariance principle used therein is, in fact, inapplicable as
which are the linear feedback control [11]–[13] and the adaptive the network error system is nonautonomous. To this end, the main
feedback control [14]–[16]. Despite easy implementation, the linear aim of this article is to solve this challenging problem.
control has two major disadvantages. First, its control gain should Motivated by the above discussions, this article investigates the
be designed according to the network parameters. Therefore, the lin- adaptive synchronization of complex networks with general time-
ear control is incapable when the network parameters are unknown. varying delay, especially with nondifferentiable delay. The main
Second, the linear control is not reusable, that is, the control gain has contributions are as follows.
to be redesigned when the network parameters change. Therefore, the 1) The synchronization criteria for networks under centralized
adaptive control is usually a preferred choice as its control gain can adaptive control are established by analyzing the bounded-
be adapted by itself according to certain updating laws, especially ness of the adaptive control gain. Compared with conventional
for networks with unknown topology [16]–[18]. A basic problem for techniques, the gain analysis technique here provides new per-
the adaptive control is to determine the boundedness of the control spectives on addressing the adaptive control and can well deal
gain because controllers cannot provide infinitely large control gain with the nondifferentiable delay.
2) The boundedness of the centralized adaptive control gain is
Manuscript received December 22, 2019; revised June 3, 2020; accepted theoretically proved so as to ensure the practicality of adaptive
September 7, 2020. Date of publication October 7, 2020; date of current
controllers.
version May 19, 2022. This work was supported in part by the National
Natural Science Foundation of China under Grant 61773294, Grant 61374173, 3) The synchronization criteria for networks under decentralized
and Grant 61773175; and in part by the Australian Research Council through adaptive control are established by analyzing the control gain
Discovery Scheme under Grant DP170102303. This article was recommended and extending the well-known Halanay inequality.
by Associate Editor J. Cao. (Corresponding author: Jin Zhou.) The remainder of this article is organized as follows. Section II
Shuaibing Zhu and Jun-An Lu are with the School of Mathematics and
Statistics, Wuhan University, Wuhan 430072, China (e-mail: zhushuaibing@
presents some preliminaries. Section III establishes the main results
whu.edu.cn; [email protected]). on the adaptive synchronization of networks with nondifferentiable
Jin Zhou is with the School of Mathematics and Statistics, Wuhan delay. Section IV verifies the theoretical results via numerical simu-
University, Wuhan 430072, China, and also with the Hubei Key Laboratory lations. Section V draws the conclusion.
of Computational Science, Wuhan University, Wuhan 430072, China (e-mail:
[email protected]).
Xinghuo Yu is with the School of Engineering, Royal Melbourne Institute II. P RELIMINARIES
of Technology, Melbourne, VIC 3001, Australia (e-mail: [email protected]). A. Notations and Network Model
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCYB.2020.3022976. First, some notations are introduced. Rn and Rn×m denote
Digital Object Identifier 10.1109/TCYB.2020.3022976 the n-dimensional Euclidean space and the set of all the
2168-2267 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022 3343

(n × m)-dimensional real matrices, respectively, the superscript T Lemma 2 (Barbǎlat Lemma [34, Lemma 8.2]): Let h : R → R be a
represents the transpose of a vector or a matrix,  ·  denotes  +∞ continuous function on [0, +∞). Then, limt→+∞ h(t) = 0
uniformly
the 2-norm of a vector or a matrix; λmax (·) denotes the max- if 0 h(ω)dω converges.
imum eigenvalue of a symmetric matrix, and In is the identity Lemma 3: Let α > β ≥ 0 be constants, and let g : R → R+ be a
matrix of dimension n. The Dini derivative is defined as D+ V(t) := non-negative continuous function. If V is a non-negative continuous
lim suph→0+ [(V(t +h)−V(t))/h]. For a continuous function ϕ on R, function on [t0 − τ̂ , t0 ] and satisfies
define ϕt,h = maxs∈[t−h,t] ϕ(s), where h is a positive constant.
Consider a controlled complex network consisting of N identical V̇(t) ≤ −αV(t) + βVt,τ̂ + g(t), t ≥ t0
systems with time-varying coupling delay
then

N
ẋi (t) = f (xi (t), t) + c aij xj (t − τ (t)) + ui (t) (1) V(t) ≤ M0 exp(−σ (t − t0 )) + ĝ(t), t ≥ t0 (6)
j=1 t
where τ̂ > 0, M0 = Vt0 ,τ̂ , ĝ(t) = t0 g(θ) exp(−σ (t − θ ))dθ , and
where 1 ≤ i ≤ N, xi ∈ Rn represents the state vector of the ith σ is the unique positive solution of the equation σ = α − β exp(σ τ̂ ).
node, f : Rn × R → Rn is a continuous function, c > 0 is the Proof: Consider the following comparison system:
coupling strength, A = (aij )N×N denotes the weighted outer coupling 
matrix,  denotes the inner coupling matrix, τ (t) is the continous Ẇ(t) = −αW(t) + βWt,τ̂ + g1 (t), t ≥ t0
(7)
time-varying coupling delay, and ui (t) is the control input to the W(t) = V(t), t ≤ t0
ith node. If there is an edge from node i to node j(j  = i), then
where g1 (t) = g(t) + ε and ε is a positive constant. Let M be any
aij > 0; otherwise, aij = 0; the diagonal elements of A are defined
 constant larger than M0 . We shall prove that
by aii = − N j=1,j=i aij , where 1 ≤ i ≤ N.
The objective of this article is to synchronize the states of all the W(t) < M exp(−σ (t − t0 )) + ĝ1 (t) := h(t), t ≥ t0 (8)
nodes to a desired state s(t), which is a solution of the following
t
isolated node system: where ĝ1 (t) = t0 g1 (θ) exp(−σ (t−θ))dθ. Define ĝ1 (t) = 0 if t < t0 .
When t ∈ [t0 − τ̂ , t0 ], one has W(t) ≤ M0 < M ≤ h(t). If (8) is not
ṡ(t) = f (s(t), t). (2)
true, then
Defining the synchronization error as ei (t) = xi (t) − s(t), one has the
t1 = inf{t > t0 |W(t) ≥ h(t)}
following error system:

N is well defined. It follows that W(t1 ) = h(t1 ), t1 > t0 , and W(t) <
ėi (t) = f (xi (t), t) − f (s(t), t) + c aij ej (t − τ (t)) + ui (t) (3) h(t) for t ∈ [t0 − τ̂ , t1 ). From (7), one has
j=1
d   
W(t) exp(αt) = exp(αt) βWt,τ̂ + g1 (t) . (9)
where 1 ≤ i ≤ N. Obviously, network (1) realizes synchronization if dt
e(t) = [eT T T
1 (t), . . . , eN (t)] converges to 0. Integrating (9) from t0 to t1 gives

B. Mathematical Preliminaries W(t1 ) exp(αt1 ) − W(t0 ) exp(αt0 )

 t1
Here, we present several assumptions and lemmas.  
= exp(αt) βWt,τ̂ + g1 (t) dt. (10)
Assumption 1 (A1): Suppose that the time-varying delay τ (t) t0
satisfies
There exists t̂ ∈ [t − τ̂ , t] such that ĝ1 t,τ̂ = ĝ1 (t̂), yielding
0 ≤ τ (t) ≤ τmax (4)
 t̂
where τmax is a positive constant. ĝ1 t,τ̂ = g1 (θ) exp −σ t̂ − θ dθ
t
Remark 1: Compared with much of the literature, the assumption  0t
(A1) here is very general, since τ (t) can be constant, differentiable, ≤ g1 (θ) exp −σ t − τ̂ − θ dθ
or even nondifferentiable. In addition, the set of nondifferentiable t0
points of τ (t) may be uncountable. = exp σ τ̂ ĝ1 (t). (11)
Assumption 2 (A2): Suppose that there exists a non-negative
constant γ satisfying Then, one has Wt,τ̂ < ht,τ̂ ≤ M exp(σ τ̂ ) exp(−σ (t − t0 )) +
exp(σ τ̂ )ĝ1 (t) for t ∈ [t0 , t1 ).
f (x, t) − f (y, t) ≤ γ x − y (5) Recalling (10), one has
for any x, y ∈ Rn and t ≥ 0. W(t1 ) exp(αt1 ) < 1 + 2 + 3 (12)
Lemma 1 (Halanay Inequality [33]): Let α > β be positive con-
 t1
stants. If V(t) is a non-negative continuous function on [t0 − τ̂ , t0 ] where = exp(σ τ̂ ) exp(−σ (t − t0 ))dt +
1 t0 exp(αt)βM

and satisfies W(t0 ) exp(αt0 ), 2 = tt01 exp(αt)β exp(σ τ̂ )ĝ1 (t)dt, and 3 =
 t1
D+ V(t) ≤ −αV(t) + βVt,τ̂ , t ≥ t0 t0 exp(αt)g1 (t)dt.
Considering W(t0 ) ≤ M and σ = α − β exp(σ τ̂ ), one has
then  t1
V(t) ≤ M0 exp(−σ (t − t0 )), t ≥ t0 1 ≤ M(α − σ ) exp(αt) exp(−σ (t − t0 ))dt
t0
where τ̂ > 0, M0 = Vt0 ,τ̂ , and σ is the unique positive solution + M exp(αt0 )
of the equation σ = α − β exp(σ τ̂ ). = M exp(αt1 ) exp(−σ (t1 − t0 )). (13)

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 3344 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022

Since g is continuous, one has It then follows from (A2) that:

 t1  t
= β exp σ τ̂ exp(αt)dt g1 (θ ) exp(−σ (t − θ))dθ V̇(t) ≤ 2(γ − d(t))e(t)2 + 2ceT (t)(A ⊗ )e(t − τ (t))
2
 t1
t0 t0
 t1 ≤ 2(γ − d(t))e(t)2 + 2cA ⊗ e(t)e(t − τ (t))
= (α − σ ) g1 (θ ) exp(σ θ )dθ exp((α − σ )t)dt ≤ −(2d(t) − 2γ − cA)e(t)2
t0 θ + cAe(t − τ (t))2 .
 t1 (18)
 
= g1 (θ ) exp(σ θ ) exp((α − σ )t1 ) − exp(αθ) dθ
t0 The following proof is divided into three steps.
= exp(αt1 )ĝ1 (t1 ) − Step 1 [Boundedness of d(t)]: From (16), it is derived that d(t) is
3. (14)
nondecreasing. If
Combining (12)–(14), it can be deduced that W(t1 ) < h(t1 ), a
contradiction with W(t1 ) = h(t1 ). Therefore, W(t) < h(t) for t ≥ t0 . d(t) < d0 := max{γ + cA, d(0)} + 1 ∀t ≥ 0
Since ĝ1 (t) = ĝ(t) + (ε/σ )[1 − exp(−σ (t − t0 ))] < ĝ(t) + (ε/σ ),
then d(t) is bounded. Otherwise, there exists t0 > 0 such that d(t0 ) =
one obtains
d0 . Then
ε
V(t) < W(t) < h(t) < M exp(−σ (t − t0 )) + ĝ(t) +
σ d(t) ≥ γ + cA + 1, t ≥ t0 .
where t ≥ t0 . Letting ε → 0+ and M → M0+ leads to (6). Combining with (18), one obtains
Lemma 3 is an extension of Lemma 1, which is to deal with the
case when disturbance g(t) exists. Particularly, this lemma becomes V̇(t) ≤ −αV(t) + βVt,τmax , t ≥ t0
the Halanay inequality if g(t) ≡ 0.
where α = cA + 2 and β = cA. Applying Lemma 1
gives
III. M AIN R ESULTS
In this section, two kinds of adaptive control, including the cen- V(t) ≤ M exp(−σ (t − t0 )), t ≥ t0
tralized adaptive control and the decentralized adaptive control, are where M = Vt0 ,τmax and σ is the unique positive solution of the
used to realize the adaptive synchronization of network (1). equation β exp(σ τmax ) + σ = α. From (16), it follows that:
 t
kM
A. Centralized Adaptive Control d(t) = d(t0 ) + k V(s)ds ≤ d0 + .
t0 σ
Consider the following centralized adaptive control
Hence, d(t) is bounded.
ui (t) = −d(t)ei (t), 1 ≤ i ≤ N (15) Step 2 [Boundedness of e(t)]: Since d(t) is nondecreasing and
with updating laws bounded, d(+∞) exists. Then
 +∞
d(+∞) − d(0)
ḋ(t) = ke(t)2 (16) V(t)dt = (19)
0 k
where k is a positive constant and d(0) ≥ 0. where d(+∞) = limt→+∞ d(t). If e(t) is unbounded, then V(t) is
In (15), the adaptive control gain d(t) of each control ui is identical, unbounded. There exists a positive constant  and a sequence {ωm }
and the updating laws (16) needs to know the global information e(t) in strictly increasing order satisfying that limm→+∞ ωm = +∞ and
of network (1). V(ωm ) ≥ . Let
The following theorem establishes adaptive synchronization criteria
for network (1) under centralized adaptive control (15). In the mean- Mm = max V(t) ≥ .
[ωm , ωm +2τmax ]
time, the boundedness of the adaptive control gain d(t) is guaranteed
theoretically. Due to the unboundedness of V(t)
Theorem 1: Let assumptions (A1) and (A2) hold. Then, network (1)
realizes synchronization under centralized adaptive control (15) with ω̂m = inf{t ≥ ωm + 2τmax | V(t) ≥ Mm }
updating laws (16). In addition, the control gain of adaptive con-
is well defined. The continuity of V(t) gives V(ω̂m ) = Mm and V(t) ≤
trol (15) is necessarily bounded.
Mm for t ∈ [ωm , ω̂m ]. When t ∈ [ωm +τmax , ω̂m ], it follows from (18)
Proof: Consider the following Lyapunov function:
that:

N
V(t) = eT (t)e(t) = eT V̇(t) ≤ (2γ + cA)V(t) + cAV(t − τ (t))
i (t)ei (t). (17)
i=1 ≤ 2(γ + cA)Mm .
Differentiating V along the solution of (3) yields In the case of τ0 := (1/[2(γ + cA)]) ≥ τmax , one has

N
 
V̇(t) = 2 eT V(t) ≥ V ω̂m − 2(γ + cA)Mm ω̂m − t
i (t) f (xi (t), t) − f (s(t), t)  
i=1 = Mm 1 − 2(γ + cA) ω̂m − t (20)

N 
N
+ 2c aij eT where t ∈ [ω̂m − τmax , ω̂m ] ⊂ [ωm + τmax , ω̂m ]. Then
i (t)ej (t − τ (t))
 ω̂m
i=1 j=1  
V(t)dt ≥ Mm τmax 1 − (γ + cA)τmax
 N
ω̂m −τmax
− 2d(t) eT
i (t)ei (t). 2
≥ (γ + cA)τmax
i=1

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022 3345

which contradicts (19). In the case of τ0 < τmax , (20) holds for with updating laws
t ∈ [ω̂m − τ0 , ω̂m ]. Then
 ω̂m ḋi (t) = ki ei (t)2 (22)
Mm 
V(t)dt ≥ ≥
ω̂m −τ0 4(γ + cA) 4(γ + cA) where ki is a positive constant and di (0) ≥ 0.
Different from the centralized adaptive control (15), each control
which contradicts (19). Therefore, e(t) is bounded.
here has its own control gain di (t), and the updating laws (22) only
Step 3 [Convergence of e(t) to Zero]: Considering
depend on local information ei (t).
f (xi (t), t) − f (s(t), t) ≤ γ ei (t) Centralized adaptive control may not be practical for large-scale
networks. The adaptive synchronization criteria for network (1) under
one obtains the boundedness of f (xi (t), t) − f (s(t), t). Since d(t) is
decentralized adaptive control (21) is therefore established in the
bounded, it follows from (3) that ė(t) is bounded. Then, V̇(t) =
following theorem.
2eT (t)ė(t) is bounded, implying that V(t) is uniformly continuous on
Theorem 2: Let assumptions (A1) and (A2) hold. Then, network (1)
[0, +∞). By using Lemma 2, it follows from (19) that
realizes synchronization under decentralized adaptive control (21)
lim V(t) = 0 with updating laws (22).
t→+∞
Proof: Define  = {i| di (t) < d0 ∀t ≥ 0}, where
leading to limt→+∞ e(t) = 0, that is, network (1) realizes
synchronization. d0 = max{γ + cA, d1 (0), . . . , dN (0)} + 1.
Remark 2: Theorem 1 establishes synchronization criteria for
network (1) under centralized adaptive control. It is noteworthy that Denote l = ||, where || is the number of the elements in .
the criteria here are independent of the network topology, and thereby Without loss of generality, let  = {1, . . . , l}. 
are convenient to use. This benefits from using the adaptive control. If l = N, then all di (t) are bounded, so 0+∞ e(t)2 dt con-
Remark 3: When the delay τ (t) is nondifferentiable, especially verges. Following steps 2 and 3 of the proof of Theorem 1, one
when τ (t) has an uncountable set of nondifferentiable points, conven- has limt→+∞ e(t) = 0.
tional methods does not work for solving the adaptive synchronization It is next to deal with the case of 0 ≤ l < N. Denote
problem of networks. In the literature, the following Lyapunov T
function: μ1 (t) = eT T
1 (t), . . . , el (t)
 t
1 T
V(t) = eT (t)e(t) + eT (θ )e(θ )dθ μ2 (t) = eT T
l+1 (t), . . . , eN (t)
2(1 − μ) t−τ (t) 
A1
d(t) − d∗
2
A=
+ A2
k
:= V1 (t) + V2 (t) + V3 (t) where A1 ∈ Rl×N and A2 ∈ R(N−l)×N . Particularly, one has μ1 = 0,
is usually used for analyzing the adaptive synchronization of networks μ2 = e, and A2 = A if l = 0.
with time-varying delay, where τ̇ (t) ≤ μ < 1 and d∗ is a large enough Consider the following Lyapunov functions:
constant. When τ (t) is not differentiable, V̇2 (t) does not exist, and
V1 (t) = μT1 (t)μ1 (t), V2 (t) = μT2 (t)μ2 (t).
D+ V2 (t) may be infinite, making it difficult to analyze V2 (t). If the
Lyapunov function is chosen as Differentiating V1 along the solution of (3) yields
V(t) = V1 (t) + V3 (t)

l
 
the Hanalay inequality (delay differential inequality that deals with V̇1 (t) = 2 eT
i (t) f (xi (t), t) − f (s(t), t)
delay, especially nondifferentiable delay) becomes difficult to use due i=1
to V3 (t). If the control is linear rather than adaptive, the problem 
l 
N
would be simple, since the Hanalay inequality can be used by + 2c aij eT
i (t)ej (t − τ (t))
choosing V(t) = V1 (t). i=1 j=1
Remark 4: Instead of using the term V3 (defined in Remark 3) 
l
to deal with the adaptive control, we analyze the boundedness of −2 di (t)eT
i (t)ei (t).
the adaptive control gain d(t), which provides new perspectives on i=1
addressing the adaptive control. Based on the gain analysis, we further Considering (A2) and di (t) ≥ di (0) ≥ 0, one has
derive the synchronization criteria of network (1). The gain analysis
here also helps avoid using the information on the differentiability of V̇1 (t) ≤ 2γ V1 (t) + 2cμT
1 (t)(A1 ⊗ )e(t − τ (t)). (23)
the delay.
Remark 5: In much of the literature, the boundedness of the control Since A1 2 = λmax (AT T T T
1 A1 ) ≤ λmax (A1 A1 +A2 A2 ) = λmax (A A) =
gain d(t) is neglected. However, this is an essential problem for the 2
A , one has
adaptive control as controllers cannot provide infinitely large control
gain in engineering. 2cμT
1 (t)(A1 ⊗ )e(t − τ (t))
Remark 6: When the initial synchronization error e(0) is large, the ≤ 2cA1 μ1 (t)e(t − τ (t))
 
control gain d(t) may be large, causing an overkill control effect. This ≤ cA V1 (t) + V1 (t − τ (t)) + V2 (t − τ (t))
kind of effect can be greatly alleviated by choosing a small value of  
≤ cA 2V1 t,τmax + V2 t,τmax (24)
the parameter k, which will be verified by simulations.
which further gives
B. Decentralized Adaptive Control
Consider the following decentralized adaptive control: V̇1 (t) ≤ α1 V1 t,τmax + β1 V2 t,τmax (25)

ui (t) = −di (t)ei (t), 1 ≤ i ≤ N (21) where α1 = 2γ + 2cA and β1 = cA.

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 3346 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022

According to the definition of l, there exists t0 ≥ 0 such that

di (t) ≥ d0 ∀t ≥ t0 for i > l. Similar to (23), one obtains

V̇2 (t) ≤ 2(γ − d0 )V2 (t) + 2cμT

2 (t)(A2 ⊗ )e(t − τ (t)) (26)

for t ≥ t0 . Similar to (24), one obtains

2cμT
2 (t)(A2 ⊗ )e(t − τ (t)) 
≤ cA 2V2 t,τmax + V1 t,τmax

which further gives

V̇2 (t) ≤ −α2 V2 (t) + β2 V2 t,τmax + g(t) ∀t ≥ t0 (27)

where α2 = 2(d0 −γ ) = 2β1 +2, β2 = 2β1 , and g(t) = β1 V1 t,τmax .

The following proof is divided into two steps.
Step 1 [Boundedness of V1 (t)]: Suppose that V1 is unbounded. Fig. 1. Trajectories of the synchronization errors eij (t) of network (34),
According
 +∞ to the definition of l and the updating laws (22), where 1 ≤ i ≤ 4 and 1 ≤ j ≤ 3.
0 ei (t)2 dt converges for 1 ≤ i ≤ l, thus
 +∞
χ0 := V1 (t)dt (28) Let ε be any positive constant. There exists a constant T ≥ t0 such
0 that g(t) ≤ ε for t ≥ T. According to (27) and Lemma 3, one has
is finite.  t
Let M0 = maxs∈[t0 −τmax ,t0 +2τmax ] V1 (s), and let M be any positive V2 (t) ≤ M̂T exp(−σ (t − T)) + g(θ) exp(−σ (t − θ ))dθ
T
constant larger than M0 and (2χ0 /τmax ). Then ε
≤ M̂T exp(−σ (t − T)) + ∀t ≥ T (32)
t1 = inf{t ≥ t0 | V1 (t) ≥ M} σ
where M̂T = V2 T,τmax and σ is defined in (31). This gives
is well defined. It is obvious that t1 > t0 + 2τmax , V1 (t1 ) = M, and limt→+∞ V2 (t) ≤ (ε/σ ). Letting ε → 0+ leads to
 
V1 (t) ≤ M ∀t ∈ t0 − τmax , t1 . (29) 0 ≤ limt→+∞ V2 (t) ≤ 0
There exists t2 ∈ [t1 − δ, t1 ] such that V1 (t2 ) = m, where δ = that is, limt→+∞ V2 (t) = 0.
(2χ0 /M) < τmax and m = mins∈[t1 −δ,t1 ] V1 (s). Since By limt→+∞ V1 (t) = 0 and limt→+∞ V2 (t) = 0, it can
 t1 deduced that limt→+∞ e(t) = 0, that is, network (1) realizes
mδ ≤ V1 (t)dt ≤ χ0 synchronization.
t1 −δ Remark 7: Theorem 2 establishes synchronization criteria for
one has m ≤ (M/2). Then, there exists t3 ∈ [t2 , t1 ] such that network (1) under decentralized adaptive control. Compared with
centralized control, decentralized control is easier to implement in
V1 (t1 ) − V1 (t2 ) M−m M2 engineering, especially for large-scale networks. It should be noted
V̇1 (t3 ) = ≥ ≥ . (30)
t1 − t2 δ 4χ0 that for centralized control, it is easy to theoretically guarantee the
boundedness of the control gain. However, it is not easy to prove
When t ∈ [t0 , t1 ], it follows from (29) that g(t) ≤ β1 M. According
the boundedness of the control gain for decentralized control, which
to (27) and Lemma 3, one has
deserves further study.
 t
V2 (t) ≤ M̂0 exp(−σ (t − t0 )) + g(θ ) exp(−σ (t − θ))dθ
t0 IV. N UMERICAL S IMULATIONS
 t
≤ M̂0 + β1 M exp(−σ (t − θ ))dθ The well-known Lorenz system [35] is described by
t0
β   ṡ(t) = f (s(t)) = Bs(t) + W(s(t)) (33)
≤ M̂0 + 1 M ∀t ∈ t0 , t1 (31)
σ
where
where M̂0 = V2 t0 ,τmax and σ is the unique positive solution of the ⎡ ⎤ ⎡ ⎤
−c1 c1 0 0
equation σ = α2 − β2 exp(σ τmax ). Combining (25), (29), and (30), B = ⎣ c3 −1 0 ⎦, W(s(t)) = ⎣ −s1 (t)s3 (t) ⎦
one has 0 0 −c2 s1 (t)s2 (t)
 
M2 β12 with c1 = 10, c2 = 8/3, and c3 = 28.
≤ V̇1 (t3 ) ≤ β1 M̂0 + α1 + M
4χ0 σ Example 1: Consider a network consisting of four Lorenz systems
under centralized control
leading to a contradiction if letting M → +∞. Hence, V1 (t) is
bounded. 
4
Step 2 [Convergence of e(t) to Zero]: Since V1 (t) is bounded, ẋi (t) = f (xi (t)) + c aij xj (t − τ (t)) − d(t)ei (t) (34)
g(t) is also bounded. Similar to (31), it can be proved that V2 (t) is j=1
bounded. Then, it follows from (25) that V̇1 (t) is bounded, so V1 is with updating laws
uniformly continuous on [0, +∞). According to (28) and Lemma 2,
one has limt→+∞ V1 (t) = 0 and thus limt→+∞ g(t) = 0. ḋ(t) = ke(t)2

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022 3347

Fig. 2. Adaptive control gain d(t) of network (34). Fig. 3. Trajectories of the synchronization errors eij (t) of network (35),
where 1 ≤ i ≤ 4 and 1 ≤ j ≤ 3.
TABLE I
TSYNC AND d(+∞) FOR D IFFERENT VALUES OF k, W HERE TSYNC I S THE
F IRST I NSTANT A FTER W HICH e(t) < 10−4 AND d(+∞) I S
A PPROXIMATED BY d(TSYNC )

where c = 1,  = I3 , k = 1, and
⎡ ⎤
−4 1 2 1
⎢ 1 −3 2 0 ⎥
A=⎢ ⎣ 2
⎥
0 −5 3 ⎦ Fig. 4. Adaptive control gain di (t) of network (35), where 1 ≤ i ≤ 4.
0 1 2 −3
τ (t) = |2m + 1 − t|, if t ∈ [2m, 2m + 2)
Example 2: Consider another network of four Lorenz systems
with m = 0, 1, 2, . . . under decentralized control
Most existing results on adaptive synchronization cannot deal with 
4
network (34) as the time-varying delay τ (t) is nondifferentiable, while ẋi (t) = f (xi (t)) + c aij xj (t − τ (t)) − di (t)ei (t) (35)
ours can. It is verified that (A1) holds with τmax = 1 and (A2) holds j=1
due to the boundedness of the Lorenz system [36]. According to
with updating laws
Theorem 1, network (34) can realize synchronization.
Set d(0) = 1, s(0) = [0, −1, −5]T , and ḋi (t) = ki ei (t)2
where ki = 1 and the other parameters are as set in Example 1.
xi (t) = 30i + [1, −10, −15]T , t ∈ [ − 1, 0]
According to Theorem 2, network (35) can realize synchronization,
as (A1) and (A2) hold (verified in Example 1).
where 1 ≤ i ≤ 4. The synchronization errors ei (t) and the control
Set di (0) = i, s(0) = [0, −1, 2]T , and
gain d(t) are presented in Figs. 1 and 2, respectively. It is shown
that network (34) realizes synchronization and the control gain d(t) xi (t) = i + [1, −1, −2]T , t ∈ [ − 1, 0]
is bounded, which demonstrates the effectiveness of Theorem 1.
From Fig. 2, we find that the control gain d(t) is large when we where 1 ≤ i ≤ 4. The synchronization errors ei (t) and the control
set k = 1. We next show how to reduce the control gain by tuning gain di (t) are presented in Figs. 3 and 4, respectively. Fig. 3 shows
the value of k. For simplicity, the synchronization is considered to be that network (35) realizes synchronization, which demonstrates the
realized at T0 if the synchronization error satisfies e(t) < 10−4 for effectiveness of Theorem 2. Fig. 4 shows the control gain di (t) is
t ≥ T0 . The lower bound of T0 is denoted by Tsync , and the upper bounded, but the boundedness of di (t) has not been theoretically
bound d(+∞) of the control gain d(t) is approximated by d(Tsync ). proved.
The values of Tsync and d(+∞) for several different values of k
are given in Table I. Roughly speaking, d(+∞) decreases with the V. C ONCLUSION
decrease of k when k ≥ 0.003. Therefore, we can choose a small k In this article, the synchronization of complex networks with
to avoid the overkill control effect. On the other hand, k should not a time-varying delay has been investigated. Unlike most existing
be too small, because Tsync increases quickly when k becomes very studies, the delay here could be nondifferentiable. The synchroniza-
small. tion criteria for networks under the centralized adaptive control and

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.
 3348 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022

networks under the decentralized adaptive control have been estab- [18] X. Zhao, J. Zhou, S. Zhu, C. Ma, and J. Lu, “Topology iden-
lished by analyzing the control gain and extending the well-known tification of multiplex delayed networks,” IEEE Trans. Circuits
Syst. II, Exp. Briefs, vol. 67, no. 2, pp. 290–294, Feb. 2020,
Halanay inequality. For centralized control, the boundedness of the
doi: 10.1109/TCSII.2019.2903287.
control gain has been theoretically proved. For decentralized control, [19] P. He, S. H. Ma, and T. Fan, “Finite-time mixed outer synchronization
however, determining the boundedness of the control gain is still an of complex networks with coupling time-varying delay,” Chaos, vol. 22,
open problem, which deserves further study. no. 4, pp. 175–423, 2012.
[20] J. L. Wang, H. N. Wu, and T. Huang, “Passivity-based synchronization
of a class of complex dynamical networks with time-varying delay,”
R EFERENCES Automatica, vol. 56, pp. 105–112, Jun. 2015.
[1] M. E. J. Newman, “The structure and function of complex networks,” [21] M. Xu, J. L. Wang, Y. L. Huang, P. C. Wei, and S. X. Wang, “Pinning
SIAM Rev., vol. 45, no. 2, pp. 167–256, 2003. synchronization of complex dynamical networks with and without time-
[2] D. Meng, “Dynamic distributed control for networks with cooperative– varying delay,” Neurocomputing, vol. 266, pp. 263–273, Nov. 2017.
antagonistic interactions,” IEEE Trans. Autom. Control, vol. 63, no. 8, [22] S. Gong, S. Yang, Z. Guo, and T. Huang, “Global exponential synchro-
pp. 2311–2326, Aug. 2018. nization of memristive competitive neural networks with time-varying
[3] Q. Song, W. Yu, J. Cao, and F. Liu, “Reaching synchronization in net- delay via nonlinear control,” Neural Process. Lett., vol. 49, no. 1,
worked harmonic oscillators with outdated position data,” IEEE Trans. pp. 103–119, 2019.
Cybern., vol. 46, no. 7, pp. 1566–1578, Jul. 2016. [23] J. Cao and J. Lu, “Adaptive synchronization of neural networks
[4] J. Zhou, J. Chen, J. Lu, and J. Lü, “On applicability of auxiliary system with or without time-varying delay,” Chaos, vol. 16, no. 1, 2006,
approach to detect generalized synchronization in complex network,” Art. no. 037203.
IEEE Trans. Autom. Control, vol. 62, no. 7, pp. 3468–3473, Jul. 2017. [24] S. C. Jeong, D. H. Ji, H. P. Ju, and S. C. Won, “Adaptive synchroniza-
[5] W. Xiong, L. Xu, D. W. C. Ho, J. Cao, and T. Huang, “Synchronous tion for uncertain chaotic neural networks with mixed time delays using
and asynchronous iterative learning strategies of T–S fuzzy systems with fuzzy disturbance observer,” Appl. Math. Comput., vol. 219, no. 11,
measurable and unmeasurable state information,” IEEE Trans. Fuzzy pp. 5984–5995, 2013.
Syst., vol. 26, no. 5, pp. 3042–3053, Oct. 2018. [25] J. Zhou, X. Wu, W. Yu, M. Small, and J. A. Lu, “Pinning synchro-
[6] Y. Xin, Y. Li, X. Huang, and Z. Cheng, “Quasi-synchronization of nization of delayed neural networks,” Chaos, vol. 18, no. 4, 2008,
delayed chaotic memristive neural networks,” IEEE Trans. Cybern., Art. no. 043111.
vol. 49, no. 2, pp. 712–718, Feb. 2019. [26] Y. I. Liang, X. Wang, and J. Eustace, “Adaptive synchronization in com-
[7] S. H. Strogatz, “Exploring complex networks,” Nature, vol. 410, plex networks with non-delay and variable delay couplings via pinning
no. 6825, pp. 268–276, 2001. control,” Neurocomputing, vol. 123, no. 5, pp. 292–298, 2014.
[8] R. Albert and A. L. Barabási, “Statistical mechanics of complex [27] T. H. Lee, Z. G. Wu, and H. P. Ju, “Synchronization of a complex
networks,” Rev. Mod. Phys., vol. 74, no. 1, pp. 47–97, 2002. dynamical network with coupling time-varying delays via sampled-data
[9] S. Zhu, J. Zhou, G. Chen, and J. Lu, “Estimating the region of attraction control,” Appl. Math. Comput., vol. 219, no. 3, pp. 1354–1366, 2012.
on a complex dynamical network,” SIAM J. Control Optim., vol. 57, [28] M. S. Ali and J. Yogambigai, “Passivity-based synchronization of
no. 2, pp. 1189–1208, 2019. stochastic switched complex dynamical networks with additive time-
[10] W. Xiong, D. W. C. Ho, and L. Xu, “Multilayered sampled-data iterative varying delays via impulsive control,” Neurocomputing, vol. 273,
learning tracking for discrete systems with cooperative-antagonistic pp. 209–221, Jan. 2018.
interactions,” IEEE Trans. Cybern., early access, May 29, 2019, [29] H. Zhang, Z. Wang, and D. Liu, “Robust stability analysis for interval
doi: 10.1109/TCYB.2019.2915664. Cohen–Grossberg neural networks with unknown time-varying delays,”
[11] M. Porfiri and M. D. Bernardo, “Criteria for global pinning- IEEE Trans. Neural Netw., vol. 19, no. 11, pp. 1942–1955, Nov. 2008.
controllability of complex networks,” Automatica, vol. 44, no. 12, [30] B. Wang and Z. H. Guan, “Synchronization of complex dynamical
pp. 3100–3106, 2008. networks with time-varying delays via impulsive distributed control,”
[12] J. Zhou, X. Yu, and J. Lu, “Node importance in controlled complex IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 57, no. 8, pp. 2182–2195,
networks,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 66, no. 3, Aug. 2010.
pp. 437–441, Mar. 2019. [31] W. Xiong, L. Song, and J. Cao, “Adaptive robust convergence of neural
[13] W. Yu, G. Chen, J. Lü, and J. Kurths, “Synchronization via pinning networks with time-varying delays,” Nonlinear Anal. Real World Appl.,
control on general complex networks,” SIAM J. Control Optim., vol. 51, vol. 9, no. 4, pp. 1283–1291, 2008.
no. 2, pp. 1395–1416, 2013. [32] B. Liu, X. Wang, H. Su, H. Zhou, Y. Shi, and R. Li, “Adaptive
[14] J. Zhou, J. Lu, and J. Lü, “Pinning adaptive synchronization of a general synchronization of complex dynamical networks with time-varying
complex dynamical network,” Automatica, vol. 44, no. 4, pp. 996–1003, delays,” Circuits Syst. Signal Process., vol. 33, no. 4, pp. 1173–1188,
2008. 2014.
[15] M. Tan, P. Qi, and Z. Xuan, “Adaptive stabilization and synchronization [33] J. Cao and J. Wang, “Absolute exponential stability of recurrent neu-
of non-diffusively coupled complex networks with nonidentical nodes ral networks with Lipschitz-continuous activation functions and time
of different dimensions,” Nonlinear Dyn., vol. 85, no. 1, pp. 303–316, delays,” Neural Netw., vol. 17, no. 3, pp. 379–390, 2004.
2016. [34] H. K. Khalil, Nonlinear Systems (3rd ed.). Upper Saddle River, NJ, USA:
[16] S. Zhu, J. Zhou, G. Chen, and J. Lu, “A new method for topology Prentice-Hall, 2002.
identification of complex dynamical networks,” IEEE Trans. Cybern., [35] E. N. Lorenz, “Deterministic nonperiodic flow,” J. Atmosp. Sci., vol. 20,
early access, Feb. 8, 2019, doi: 10.1109/TCYB.2019.2894838. no. 2, pp. 130–141, 1963.
[17] H. Liu, J. Lu, J. Lü, and D. J. Hill, “Structure identification of uncer- [36] J. Zhou and J. Lu, “Topology identification of weighted complex
tain general complex dynamical networks with time delay,” Automatica, dynamical networks,” Physica A, Stat. Mech. Appl., vol. 386, no. 1,
vol. 45, no. 8, pp. 1799–1807, 2009. pp. 481–491, 2007.

Authorized licensed use limited to: Manav Rachna International Institute of Research and Studies. Downloaded on August 12,2024 at 16:58:09 UTC from IEEE Xplore. Restrictions apply.