on global exponential stability of delayed cellular neural networks with time-varying delays
TRANSCRIPT
Applied Mathematics and Computation 162 (2005) 679–686
www.elsevier.com/locate/amc
On global exponential stabilityof delayed cellular neural networks
with time-varying delays
Qiang Zhang a,b,*, Xiaopeng Wei a, Jin Xu a
a Advanced Design Technology Center, Dalian University, Dalian 116622, Chinab School of Mechanical Engineering, Dalian University of Technology, Dalian 116024, China
Abstract
A new sufficient condition has been presented ensuring the global exponential sta-
bility of cellular neural networks with time-varying delays by using an approach based
on delay differential inequality combining with Young inequality. The results estab-
lished here extend those earlier given in the literature. Compared with the method of
Lyapunov functionals as in most previous studies, our method is simpler and more
effective for stability analysis.
� 2004 Elsevier Inc. All rights reserved.
Keywords: Global exponential stability; Cellular neural networks; Time-varying delays; Delay
differential inequality; Lyapunov functionals
1. Introduction
Cellular neural networks (CNNs) were introduced by Chua and Yang in
1988 [1]. CNNs have been extensively discussed in the past decade. CNNs with
time delay (DCNNs) proposed in [2] have found applications in many areasincluding classification of patterns and processing of moving images. Such
applications rely on the existence of an equilibrium point or of a unique
* Corresponding author. Address: Advanced Design Technology Center, Dalian University,
Dalian 116622, China.
E-mail address: [email protected] (Q. Zhang).
0096-3003/$ - see front matter � 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2004.01.004
680 Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686
equilibrium point, and its stability. To the best of our knowledge, cellular
neural networks with time-varying delays are seldom considered [3–17].However, in practice, time delays are usually variable, and sometimes vary
violently with time. Therefore, the studies of neural networks with time-varying
delays are more important and actual than those with constant delays. Fur-
thermore, in designing a neural network, one is not only interested in the
stability of the network, but also in performance. In particular, it is often
desired that a neural network converges fast enough in order to achieve fast
response. The aim of this brief is to provide some new results on global
exponential stability for cellular neural networks with time-varying delays. Theresults in the earlier references emerge as special cases of the main results given
here.
2. Global exponential convergence analysis
The dynamic behavior of a continuous time DCNN can be described by the
following state equations:
x0iðtÞ ¼ �cixiðtÞ þXn
j¼1
aijfjðxjðtÞÞ þXn
j¼1
bijfjðxjðt � sjðtÞÞÞ þ Ji;
i ¼ 1; 2; . . . ; n ð1Þ
or equivalently
x0ðtÞ ¼ �CxðtÞ þ Af ðxðtÞÞ þ Bf ðxðt � sðtÞÞÞ þ J ; ð2Þ
where n corresponds to the number of units in a neural network;
xðtÞ ¼ ½x1ðtÞ; . . . ; xnðtÞ�T 2 Rn corresponds to the state vector at time t;f ðxðtÞÞ ¼ ½f1ðx1ðtÞÞ; . . . ; fnðxnðtÞÞ�T 2 Rn denotes the activation function of the
neurons; f ðxðt� sðtÞÞÞ ¼ ½f1ðx1ðt� s1ðtÞÞÞ; . . . ; fnðxnðt� snðtÞÞÞ�T 2 Rn;C;A;B; Jare constant matrices; C ¼ diagðci > 0Þ (a positive diagonal matrix) represents
the rate with which the ith unit will reset its potential to the resting state in
isolation when disconnected from the network and external inputs. A ¼ faijg isreferred to as the feedback matrix, B ¼ fbijg represents the delayed feedback
matrix, while J ¼ ½J1; . . . ; Jn�T is an external bias vector, sjðtÞ is the transmis-
sion delay along the axon of the jth unit and satisfies 06 sjðtÞ6 s (s is a
constant). The activation function fi (i ¼ 1; 2; . . . ; n) satisfies the following
condition
(H) Each fi is bounded continuous and satisfies
jfiðn1Þ � fiðn2Þj6 Lijn1 � n2j
for each n1; n2 2 R, n1 6¼ n2.
Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686 681
This type of activation functions is clearly more general than both the usual
sigmoid activation functions in Hopfield networks and the piecewise linearfunction (PWL): fiðxÞ ¼ 1
2ðjxþ 1j � jx� 1jÞ in standard cellular networks [1].
Suppose that the system (1) is supplemented with initial conditions of the
form
xiðsÞ ¼ /iðsÞ; s 2 ½�s; 0�; i ¼ 1; 2; . . . ; n
in which /iðsÞ is continuous for s 2 ½�s; 0�, and Eq. (1) has an equilibrium
point x� ¼ ðx�1; x�2; . . . ; x�nÞ. We denote
k/ � x�k ¼ sup�s6 s6 0
Xn
j¼1
j/jðsÞ"
� x�j jr
#1=r
:
We say that an equilibrium point x� ¼ ðx�1; x�2; . . . ; x�nÞ is globally exponen-
tially stable if there exist constants � > 0 and M P 1 such that
kxðtÞ � x�k6Mk/ � x�ke��t; tP 0:
In order to simplify the proofs and compare our results, we present some
lemmas as follows.
Lemma 1 (Young inequality [18]). Assume that a > 0, b > 0, p > 1, 1p þ 1
q ¼ 1,then the following inequality:
ab61
pap þ 1
qbq
holds.
Lemma 2 (Halanay inequality[19]). Let a and b be constants with 0 < b < a.Let xðtÞ be a continuous nonnegative function on tP t0 � s satisfying inequality(3) for tP t0.
x0ðtÞ6 � axðtÞ þ b�xðtÞ; ð3Þ
where �xðtÞ¼def supt�s6 s6 tfxðsÞg. Then
xðtÞ6�xðt0Þe�rðt�t0Þ; ð4Þ
where r is a bound on the exponential convergence rate and is the unique positivesolution of
r ¼ a � bers: ð5Þ
Theorem. Assume that there exist real constants fij, gij and positive constantsci > 0, pP 1, i; j ¼ 1; 2; . . . ; n such that
682 Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686
min16 i6 n
pci
(�Xn
j¼1
cjciLijajijpð1�fjiÞ
�þ ðp � 1ÞLjjaijj
pfijp�1 þ ðp � 1ÞLjjbijj
pgijp�1
�)
> max16 i6 n
Xn
j¼1
cjciLijbjijpð1�gjiÞ
( ); ð6Þ
then the equilibrium point x� of system (1) is globally exponentially stable.
Proof. Let yiðtÞ ¼ xiðtÞ � x�i , define a function
V ðtÞ ¼Xn
i¼1
cijyiðtÞjp:
Calculating and estimating the upper right derivative DþV of V along the
solution of (1) as follows:
DþV ðtÞ ¼Xn
i¼1
cipjyiðtÞjp�1
signðyiðtÞÞy0iðtÞ
¼Xn
i¼1
cipjyiðtÞjp�1
signðyiðtÞÞ"� ciyiðtÞ þ
Xn
j¼1
aij fjðyjðtÞ�
þ x�j Þ � fjðx�j Þ
þXn
j¼1
bij fjðyjðt�
� sjðtÞÞ þ x�j Þ � fjðx�j Þ#
6
Xn
i¼1
cip
"� cijyiðtÞjp þ
Xn
j¼1
jaijjLjjyiðtÞjp�1jyjðtÞj
þXn
j¼1
jbijjLjjyiðtÞjp�1jyjðt � sjðtÞÞj#
¼Xn
i¼1
cip
(� cijyiðtÞjp þ
Xn
j¼1
LjjyjðtÞkaijj1�fij jaijjfijp�1jyiðtÞj
� �p�1
þXn
j¼1
Ljjyjðt � sjðtÞÞkbijj1�gij jbijjgijp�1jyiðtÞj
� p�1): ð7Þ� �
Let a ¼ jyjðtÞkaijj1�fij , b ¼ jaijjfijp�1jyiðtÞj
p�1
, by Lemma 1, we have
jyjðtÞkaijj1�fij jaijjfijp�1jyiðtÞj
� �p�1
61
pjaijjpð1�fijÞjyjðtÞjp þ
p � 1
pjaijj
pfijp�1jyiðtÞjp: ð8Þ
Similarly, let a ¼ jyjðt � sjðtÞÞkbijj1�gij , b ¼ jbijjgijp�1jyiðtÞj
� p�1
, by Lemma 1, we
get
Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686 683
jyjðt � sjðtÞÞkbijj1�gij jbijjgijp�1jyiðtÞj
� p�1
61
pjbijjpð1�gijÞjyjðt � sjðtÞÞjp þ
p � 1
pjbijj
pgijp�1jyiðtÞjp: ð9Þ
Substituting Eqs. (8) and (9) into Eq. (7), we obtain
DþV ðtÞ6Xn
i¼1
cip
(� cijyiðtÞjp þ
Xn
j¼1
1
pLjjaijjpð1�fijÞjyjðtÞjp
þXn
j¼1
p � 1
pLjjaijj
pfijp�1jyiðtÞjp þ
Xn
j¼1
1
pLjjbijjpð1�gijÞjyjðt � sjðtÞÞjp
þXn
j¼1
p � 1
pLjjbijj
pgijp�1jyiðtÞjp
)
¼Xn
i¼1
ci
"� pci þ
Xn
j¼1
ðp � 1ÞLjjaijjpfijp�1 þ
Xn
j¼1
ðp � 1ÞLjjbijjpgijp�1
þXn
j¼1
cjciLijajijpð1�fjiÞ
#jyiðtÞjp þ
Xn
i¼1
ciXn
j¼1
cjciLijbjijpð1�gjiÞ
" #jyiðt � siðtÞÞjp
6 � min16 i6 n
pci
(�Xn
j¼1
cjciLijajijpð1�fjiÞ
�þ ðp � 1ÞLjjaijj
pfijp�1
þ ðp � 1ÞLjjbijjpgijp�1
�)V ðtÞ þ max
16 i6 n
Xn
j¼1
cjciLijbjijpð1�gjiÞ
( )V ðtÞ:
According to Lemma 2, we obtain
cminkxðtÞ � x�kp 6 V ðtÞ6 V ðt0Þ expð�rðt � t0ÞÞ:
It follows from above that
kxðtÞ � x�k6 c1=pmax
c1=pmin
k/ � x�k exp
�� rpt�:
Therefore, the proof is completed. h
Corollary 1. If there exist constants ci > 0 such that
min16 i6 n
ci
(�Xn
j¼1
cjciLijajij
� �)> max
16 i6 n
Xn
j¼1
cjciLijbjij
( );
then the equilibrium point x� of system (1) is globally exponentially stable.
Proof. Taking p ¼ 1 in theorem above, then we can easily obtain Corollary
1. h
684 Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686
Corollary 2. If there exist constants ci > 0 such that
min16 i6 n
2ci �Xn
j¼1
cjciLijajij þ Lj jaijj þ jbijj
�� �( )> max
16 i6 n
Xn
j¼1
cjciLijbjij
( );
then the equilibrium point x� of system (1) is globally exponentially stable.
Proof. It is easy to check that the inequality (6) is satisfied by taking p ¼ 2,
fij ¼ gij ¼ 0:5, and hence the theorem implies Corollary 2. h
Remark 1. If we let ci ¼ 1, i ¼ 1; 2; . . . ; n, then Corollarys 1 and 2 correspondto theorems 1 and 2 in [6], respectively. That is, our theorem includes the main
results in [6] as special cases.
Remark 2. For the pure-delay network model of (1) (i.e., aij ¼ 0; 8i; j), if
Li ¼ ci ¼ 1, it follows from Corollary 1 above that the inequalityPnj¼1
cjcijbjij < 1, i ¼ 1; 2; . . . ; n ensures global exponential stability of the
equilibrium point of (1). This coincides with the result in [7]. Furthermore,
according to the property of M-matrix [8], we can conclude that the conditionPnj¼1
cjcijbjij < 1 yields qðjBjÞ < 1, where qð Þ denotes spectral radius and jBj
denotes absolute-value matrix given by jBj ¼ ðjbijjÞn�n. This is consistent with
the main result in [9].
Remark 3. In [10–12], by constructing Lyapunov functional, some results onthe global asymptotic stability of (1) are presented. Different from our results,
all of those results require that the delay function sjðtÞ be differentiable. Thus,
conditions given in [10–12] are more restrictive and conservative.
Remark 4. Young inequality was first used by Cao [13]. However, the results in
[13] only hold for constant delays.
3. Examples
In this section, we will give some examples to show the validity of our re-
sults.
Example 1. Consider a delayed neural network with bounded time-varying
delays
x0ðtÞ ¼ �xðtÞ þ Af ðxðtÞÞ þ Bf ðxðt � sðtÞÞÞ; ð10Þ
Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686 685
where
C ¼ I ; A ¼ 0:1 �0:10:1 0:1
�; B ¼ 0:5 �0:1
0:1 0:5
�
and the activation function is PWL: fiðxÞ ¼ 12ðjxþ 1j � jx� 1jÞ. Clearly, fi
satisfies the assumption (H) with Li ¼ 1, i ¼ 1; 2. One can easily check that
min16 i6 n
ci
(�Xn
j¼1
cjciLijajij
� �)¼ 0:8 > max
16 i6 n
Xn
j¼1
cjciLijbjij
( )¼ 0:6:
Therefore, by Corollary 1, the equilibrium point of (10) is globally exponen-tially stable. By a simple computation, we can easily seen that the matrix
Aþ AT ¼ 0:2 0
0 0:2
�is not negative semidefinite. Thus the condition in [14]
does not hold. For this example, our result is less restrictive than that given in
[14].
Example 2. Consider the following delayed neural networks
x0ðtÞ ¼ �CxðtÞ þ Af ðxðtÞÞ þ Bf ðxðt � sðtÞÞÞ; ð11Þ
whereC ¼ 1:5 0
0 1:5
�; A ¼ 0:1 �0:1
0:2 0:1
�; B ¼ 1 0:2
�0:3 0:2
�
and the activation function is described by: fiðxÞ ¼ tanhðxÞ. Clearly, fi satisfiesthe assumption (H) with Li ¼ 1, i ¼ 1; 2. One can check that
min16 i6 n
ci
(�Xn
j¼1
cjciLijajij
� �)¼ 1:2 < max
16 i6 n
Xn
j¼1
cjciLijbjij
( )¼ 1:3;
min16 i6 n
2ci
(�Xn
j¼1
cjciLijajij
�þ Ljðjaijj þ jbijjÞ
�)
¼ max16 i6 n
Xn
j¼1
cjciLijbjij
( )¼ 1:3:
Therefore, the main results in [6] do not hold for this example. On the other
hand, if we let p ¼ 3, fij ¼ gij ¼ 0:5, c1 ¼ c2 ¼ 1 in theorem above, one can
verify that
min16 i6 n
3ci
(�Xn
j¼1
jajij
þ 2ðjaijj þ jbijjÞ�)
¼ 1:4 > max16 i6 n
Xn
j¼1
jbjij( )
¼ 1:3:
Hence, the equilibrium point of Eq. (11) is globally exponentially stable.
686 Q. Zhang et al. / Appl. Math. Comput. 162 (2005) 679–686
Acknowledgements
The project was supported by the National Natural Science Foundation of
China (Grant nos. 60174037, 50275013).
References
[1] L.O. Chua, Y. Ling, Cellular neural networks: theory, IEEE Trans. Circuits Syst. I 35 (1988)
1257–1272.
[2] T. Roska, L.O. Chua, Cellular neural networks with nonlinear and delay-type templates, Int. J.
Circuit Theory Appl. 20 (1992) 469–481.
[3] T. Chu, An exponential convergence estimate for analog neural networks with delay, Phys.
Lett. A 283 (2001) 113–118.
[4] J. Zhang, Globally exponential stability of neural networks with variable delays, IEEE Trans.
Circuits Syst. I 50 (2003) 288–291.
[5] C. Hou, J. Qian, Stability analysis for neural dynamics with time-varying delays, IEEE Trans.
Neural Networks 9 (1998) 221–223.
[6] D. Zhou, J. Cao, Globally exponential stability conditions for cellular neural networks with
time-varying delays, Appl. Math. Comput. 131 (2002) 487–496.
[7] J. Peng, H. Qiao, Z.B. Xu, A new approach to stability of neural networks with time-varying
delays, Neural Networks 15 (2002) 95–103.
[8] A. berman, R.J. Plemmons, Nonnegative Matrices in the Mathematical Science, Academic,
New York, 1979.
[9] D. Xu, H. Zhao, H. Zhu, Global dynamics of Hopfield neural networks involving variable
delays, Comput. Math. Appl. 42 (2001) 39–45.
[10] X. Liao, G. Chen, E.N. Sanchez, LMI-based approach for asymptotically stability analysis of
delayed neural networks, IEEE Trans. Circuits Syst. I 49 (2002) 1033–1039.
[11] M. Joy, On the global convergence of a class of functional differential equations with
applications in neural network theory, J. Math. Anal. Appl. 232 (1999) 61–81.
[12] J. Cao, J. Wang, Global asymptotic stability of a general class of recurrent neural networks
with time-varying delays, IEEE Trans. Circuits Syst. I 50 (2003) 34–44.
[13] J. Cao, New results concerning exponential stability and periodic solutions of delayed cellular
neural networks, Phys. Lett. A 307 (2003) 136–147.
[14] X.M. Li, L.H. Huang, H. Zhu, Global stability of cellular neural networks with constant and
variable delays, Nonlinear Anal. 53 (2003) 319–333.
[15] Q. Zhang, R. Ma, J. Xu, Stability of cellular neural networks with delay, Electron. Lett. 37
(2001) 575–576.
[16] Q. Zhang, R. Ma, C. Wang, J. Xu, On the global stability of delayed neural networks, IEEE
Trans. Automat. Control 48 (2003) 794–797.
[17] Y. Zhang, Global exponential stability and periodic solutions of delay Hopfield neural
networks, Int. J. Syst. Sci. 27 (1996) 895–901.
[18] G.H. Hardy, J.E. Littlewood, G. Polya, Inequalities, second ed., Cambridge University Press,
London, 1952.
[19] K. Gopalsamy, Stability and Oscillations in Delay Differential Equations of Population
Dynamics, Kluwer Academic Publishers, Dordrecht, 1992.