Upload
jie-zhou
View
214
Download
0
Embed Size (px)
Citation preview
ARTICLE IN PRESS
Neurocomputing 72 (2009) 1688–1693
Contents lists available at ScienceDirect
Neurocomputing
0925-23
doi:10.1
� Corr
E-m
journal homepage: www.elsevier.com/locate/neucom
Global exponential stability of impulsive BAM neural networks withdistributed delays
Jie Zhou a,b,�, Shuyong Li b
a College of Science, Sichuan University of Science and Engineering, Zigong, Sichuan 643000, PR Chinab College of Mathematics and Software Science, Sichuan Normal University, Chengdu 610068, PR China
a r t i c l e i n f o
Article history:
Received 25 February 2008
Received in revised form
6 June 2008
Accepted 11 August 2008
Communicated by T. Heskesand differential inequality technique to obtain our results. It is shown that, in some case, the stability
Available online 1 October 2008
Keywords:
Global exponential stability
BAM
Distributed delays
Impulsive
12/$ - see front matter & 2008 Elsevier B.V. A
016/j.neucom.2008.08.008
esponding author.
ail address: [email protected] (J. Zhou
a b s t r a c t
In this paper, the existence and uniqueness, and the global exponential stability (GES) of equilibrium
point of a impulsive BAM neural networks with distributed delays are considered. Here we point out,
different from previous methods, we do not construct Lyapunov functional or use the properties of
nonsingular M-matrix. We use some basic analytical technique, such as contraction mapping principle
criteria can be easily checked. Several remarks are worked out to demonstrate the advantage of our
results.
& 2008 Elsevier B.V. All rights reserved.
1. Introduction
The bidirectional associative memory (BAM) neural networksmodel, known as an extension of the unidirectional autoassociatorof Hopfield [4,14,16], was first introduced by Kosko [6,7]. This classof networks possesses good application prospects in some fieldssuch as pattern recognition, signal and image process, artificialintelligence. Such applications heavily depend on the dynamicalbehaviors of the neural networks. Thus, the analysis of thedynamical behaviors is a necessary step for practical design ofneural networks.
One of the most investigated problems in dynamical behaviorsof BAM neural networks is the existence, uniqueness and globalstability of the equilibrium point. The property of the globalstability, which means that the domain of attraction of theequilibrium point is the whole space, is of importance fromthe theoretical as well as application of view in several field. Thereexist some results of stability for the constant delays anddistributed delays BAM [17,8,2,1,15,10].
Recently, there has been a somewhat new category of neuralnetworks, which is neither purely continuous-time nor purelydiscrete-time ones; these are called impulsive neural networks.This third category of neural networks displays a combination ofcharacteristics of both the continuous-time and discrete-timesystems [12,13,3,9]. Therefore, it is necessary to consider both
ll rights reserved.
).
impulsive effect and delay effect on the neural networks. In thispaper, inspired by Refs. [17,8,2,1,15,10,12,13,3,9,11], we shall studythe existence, uniqueness and global exponential stability (GES) ofthe equilibrium point of impulsive BAM neural networks withdistributed delays. Here we point out, different from previousmethods, we do not construct Lyapunov functional or use theproperties of nonsingular M-matrix. We use some basic analyticaltechnique, such as contraction mapping principle and differentialinequality technique to obtain our results, so our results aredifferent from the previous.
2. Preliminaries
In the following, we will consider the BAM neural networkswith distributed delays and subjected to impulsive state displace-ments at fixed instants, which can be described by a set of integro-differential equations:
dxiðtÞ
dt¼ �aixiðtÞ þ
Xn
j¼1
pijf jðyjðtÞÞ þXn
j¼1
mij
Z 10
kijðsÞf jðyjðt � sÞÞdsþ Ii;
t40; tatk; i ¼ 1;2; . . . ;n;
DxiðtkÞ ¼ IkðxiðtkÞÞ; i ¼ 1;2; . . .n; k ¼ 1;2; . . . ;
dyjðtÞ
dt¼ �bjyjðtÞ þ
Xn
i¼1
qjigiðxiðtÞÞ þXn
i¼1
nji
Z 10
rjiðsÞgiðxiðt � sÞÞdsþ Jj;
t40; tatk; j ¼ 1;2; . . . ;n;
DyjðtkÞ ¼ JkðyjðtkÞÞ; j ¼ 1;2; . . . ;n; k ¼ 1;2; . . . ;
8>>>>>>>>>>>>>>><>>>>>>>>>>>>>>>:
(2.1)
ARTICLE IN PRESS
J. Zhou, S. Li / Neurocomputing 72 (2009) 1688–1693 1689
where DxiðtkÞ ¼ xiðtþ
k Þ � xiðt�k Þ, DyjðtkÞ ¼ yjðt
þ
k Þ � yjðt�k Þ are the
impulses at moments tk and 0ot1ot2o � � � is a strictly increasingsequence such that limk!1tk ¼ þ1; xiðtÞ; yjðtÞ are the state ofneurons, ai40; bj40 represent the passive delay rates; pij; qji;mij;
nji are the synaptic connection strengths, I ¼ ðI1; I2; . . . ; InÞT, J ¼
ðJ1; J2; . . . ; JnÞT are the constant input vectors, f j; gi present the
activation functions of the neuron; the delayed feedbackkijðsÞ; rjiðsÞ are real value nonnegative continuous functions definedon ½0;1Þwith
R10 kijðsÞpkij, and
R10 rjiðsÞprji, kij; rij are nonnegative
constants.And the initial conditions associated with the system (2.1) are
of the form
xiðsÞ ¼ ciðsÞ; s 2 ð�1;0Þ; i ¼ 1;2; . . . ;n,
yjðsÞ ¼ fjðsÞ; s 2 ð�1;0Þ; j ¼ 1;2; . . . ;n,
where ciðsÞ;fjðsÞ are bounded continuous functions on ð�1;0�.As usual in the theory of impulsive differential equations, at
the points of discontinuity tk of the solution t/ðx1ðtÞ;
x2ðtÞ; . . . ; xnðtÞ; y1ðtÞ; y2ðtÞ; . . . ; ynðtÞÞT, denoted by ðxðtÞ; yðtÞÞ. We
assume that ðx1ðtkÞ; x2ðtkÞ; . . . ; xnðtkÞ; y1ðtkÞ; y2ðtkÞ; . . . ; ynðtkÞÞT�
ðx1ðtk � 0Þ, x2ðtk � 0Þ; . . . ; xnðtk � 0Þ; y1ðtk � 0Þ; y2ðtk � 0Þ; . . . ; yn
ðtk � 0ÞÞT. It is clear that, in general, the derivatives x0iðtkÞ do notexist. On the other hand, according to the equalities of (2.1) thereexist the limits x0iðtk � 0Þ. For the above convention, we assumex0iðtkÞ � x0iðtk � 0Þ; y0iðtkÞ � y0iðtk � 0Þ.
For convenience, we can rewrite (2.1) in the form
dxðtÞ
dt¼ �AxðtÞ þ Pf ðyðtÞÞ þM
Z 10
KðsÞf ðyðt � sÞÞdsþ I;
t40; tatk;
DxðtkÞ ¼ IkðxðtkÞÞ; k ¼ 1;2; . . . ;
dyðtÞ
dt¼ �ByðtÞ þ QgðxðtÞÞ þ N
Z 10
RðsÞgðxðt � sÞÞdsþ J;
t40; tatk;
DyðtkÞ ¼ JkðyðtkÞÞ; k ¼ 1;2; . . . ;
8>>>>>>>>>>>>><>>>>>>>>>>>>>:
(2.2)
in which xðtÞ ¼ colfxiðtÞg, yðtÞ ¼ colfyjðtÞg; A ¼ diagfaig;B ¼
diagfbjg; P ¼ ðpijÞn�n, Q ¼ ðqjiÞn�n, M ¼ ðmijÞn�n, N ¼ ðnjiÞn�n;f ¼ colff jg, g ¼ colfgig; KðsÞ ¼ ðkijðsÞÞn�n, RðsÞ ¼ ðrjiðsÞÞn�n; I ¼
colfIig, J ¼ colfJjg;DxðtkÞ ¼ colfDxiðtkÞg, DyðtkÞ ¼ colfDyjðtkÞg;cðsÞ ¼ colfciðsÞg, fðsÞ ¼ colffjðsÞg:
Definition 1. A constant vector ðx�; y�Þ ¼ ðx�1; . . . ; x�n; y�1; . . . ; y
�nÞ
T
is said to be the equilibrium point of the system (2.2), if itsatisfies
Ax� ¼ Pf ðy�Þ þM
Z 10
KðsÞf ðy�Þdsþ I;
By� ¼ Qgðx�Þ þ N
Z 10
RðsÞgðx�Þdsþ J;
8>>><>>>:
(2.3)
when the impulsive jumps Ikð�Þ; Jkð�Þ satisfy Ikðx�Þ ¼ 0; Jkðy
�Þ ¼ 0,where x� ¼ ðx�1; . . . ; x
�nÞ
T; y� ¼ ðy�1; . . . ; y�nÞ
T.
Definition 2. The equilibrium point ðx�; y�Þ of the model (2.2) issaid to be globally exponentially stable, if there exists e40 andEX1 such that
kxðtÞ � x�k2 þ kyðtÞ � y�k2pEðkc� x�k þ kf� y�kÞe�et for all t40,
where ðxðtÞ; yðtÞÞ is any solution of the system (2.2) with initialvalue ðcðsÞ;fðsÞÞ.
Here we recall that the Euclidean vector norm is defined on Rn
by
kxk2 ¼Xn
i¼1
jxij2
!1=2
for x ¼ col fxig 2 Rn,
and the spectral norm [5] which is the matrix norm induced bythe Euclidean vector norm is defined by
kQk2 ¼ ðmaxfl; l is an eigenvalue of QTQgÞ1=2
where Q is an� n-matrix.
For c;f, we define kck ¼ sup�1oto0kcðtÞk2; kfk ¼ sup�1oto0
kfðtÞk2:Suppose that the system (2.2) has a unique equilibrium point
ðx�; y�Þ (which we will prove in the following section), and letuðtÞ ¼ xðtÞ � x�, vðtÞ ¼ yðtÞ � y�; FðvðtÞÞ ¼ f ðyðtÞÞ � f ðy�Þ, GðuðtÞÞ ¼
gðxðtÞÞ � gðx�Þ; CðtÞ ¼ cðtÞ � x�;FðtÞ ¼ fðtÞ � y�, then system (2.2)becomes
duðtÞ
dt¼ �AuðtÞ þ PFðvðtÞÞ þM
Z 10
KðsÞFðvðt � sÞÞds;
t40; tatk;
DuðtkÞ ¼ IkðuðtkÞÞ ¼ DxðtkÞ; k ¼ 1;2; . . . ;
dvðtÞ
dt¼ �BvðtÞ þ QGðuðtÞÞ þ N
Z 10
RðsÞGðuðt � sÞÞds;
t40; tatk;
DvðtkÞ ¼ JkðvðtkÞÞ ¼ DyðtkÞ; k ¼ 1;2; . . . :
8>>>>>>>>>>>>><>>>>>>>>>>>>>:
(2.4)
The initial conditions associated with the system (2.4) are
uðsÞ ¼ CðsÞ; s 2 ð�1;0Þ,
vðsÞ ¼ FðsÞ; s 2 ð�1;0Þ,
where CðsÞ;FðsÞ are bounded continuous functions on ð�1;0�:Throughout this paper, we make the following assumptions:
ðH1Þ
There exist positive constant numbers xj;Zi such thatjf jðy0Þ � f jðy
00Þjpxjjy0 � y00j; j ¼ 1;2; . . . ;n,
jgiðx0Þ � giðx
00ÞjpZijx0 � x00j; i ¼ 1;2; . . . ;n.
ðH2Þ
There exist positive constant number e such thatZ 10eeskijðsÞdso1;Z 1
0eesrjiðsÞdso1,
there must be constants l1 and l2 such thatZ 10
KðsÞds
��������
2
pZ 1
0eesKðsÞds
��������
2
¼ l1,Z 10
RðsÞds
��������
2
pZ 1
0eesRðsÞds
��������
2
¼ l2.
ðH3Þ
W ¼ maxfðkPk2 þ l1kMk2Þx=a�; ðkQk2 þ l2kNk2ÞZ=a�go1,where x ¼ maxjfxjg, Z ¼maxifZig; a0 ¼ minifaig, b0 ¼minjfbjg; a� ¼ minfa0; b0g.
ðH4Þ The impulsive operators IkðxðtkÞÞ; JkðyðtkÞÞ satisfyIkðxðtkÞÞ ¼ �akðxðtkÞ � x�Þ; 0oako2;
JkðyðtkÞÞ ¼ �bkðyðtkÞ � y�Þ; 0obko2;
((2.5)
where ak;bk are constants.
3. Existence and uniqueness of the equilibrium point
Theorem 1. If ðH1Þ, ðH3Þ and ðH4Þ hold, then for any C 2 C and
F 2 C, where C9Cðð�1;0�;RnÞ, there exists some constant D40,
such that
kuðtÞk2oD
2; kvðtÞk2o
D
2i:e:; kuðtÞk2 þ kvðtÞk2oD. (3.1)
That is, the solutions of system (2.4) are uniformly bounded.
Proof. For tatk and any C 2 C, F 2 C, there exists a large enoughnumber D40 such that kCkoD=2; kFkoD=2.
ARTICLE IN PRESS
J. Zhou, S. Li / Neurocomputing 72 (2009) 1688–16931690
If (3.1) is false, then there must be some t140 such that one of
the following holds:
(1)
kuðt1Þk2pD2; kvðt1Þk2o
D
2,
(2)
kuðt1Þk2oD2; kvðt1Þk2p
D
2, (3.2)
(3)
kuðt1Þk2pD2; kvðt1Þk2p
D
2i:e: kuðtÞk2 þ kvðtÞk2pD,
for�1otpt1; tatk. (3.3)
We just prove (3.3), the others are the same as (3.3).
By the assumption ðH1Þ, ðH3Þ, we have
kuðt1Þk2 þ kvðt1Þk2
pe�a0t1kCk þZ t1
0e�a0ðt1�yÞkPk2xkvðyÞk2dy
þ
Z t1
0e�a0ðt1�yÞkMk2x
Z 10
vðy� sÞKðsÞds
��������
2
dy
þ e�b0t1kFk þZ t1
0e�b0ðt1�yÞkQk2ZkuðyÞk2 dy
þ
Z t1
0e�b0ðt1�yÞkNk2Z
Z 10
uðy� sÞRðsÞds
��������
2
dy
pe�a0t1D
2þ
Z t1
0e�a0ðt1�yÞkPk2x
D
2dy
þ
Z t1
0e�a0ðt1�yÞ dykMk2x
D
2
Z 10
KðsÞds
��������
2
þ e�b0t1D
2þ
Z t1
0e�b0ðt1�yÞkQk2Z
D
2dy
þ
Z t1
0e�b0ðt1�yÞdykNk2Z
D
2
Z 10
RðsÞds
��������
2
pe�a0t1D
2þ
D
2kPk2xþ
D
2kMk2xl1
� �Z t1
0e�a0ðt1�yÞ dy
FðS; ZÞ ¼ ðF1ðS; ZÞ; F2ðS; ZÞÞ
F1ðS; ZÞ ¼ Pf ðB�1ZðtÞÞ þM
Z 10
KðyÞf ðB�1ZðtÞÞdyþ I
F2ðS; ZÞ ¼ QgðA�1SðtÞÞ þ N
Z 10
RðyÞgðA�1SðtÞÞdyþ J
���������
8>>><>>>:
9>>>=>>>;
.
þ e�b0t1D
2þ
D
2kQk2Zþ
D
2kNk2Zl2
� �Z t1
0e�b0ðt1�yÞ dy
pe�a0t1D
2þkPk2xþ kMk2xl1Þ
a0ð1� e�a0t1 Þ
D
2þ e�b0t1
D
2
þkQk2Zþ kNk2Zl2Þ
b0ð1� e�b0t1 Þ
D
2oe�a0t1
D
2
þ ð1� e�a0t1 ÞD
2þ e�b0t1
D
2þ ð1� e�b0t1 Þ
D
2¼ D,
which contradicts the equality (3.3), so (3.1) holds.
For t ¼ tk, by ðxTðtkÞ; yTðtkÞÞ ¼ ðx
Tðtk � 0Þ; yTðtk � 0ÞÞ and ðH4Þ;
kuðtk þ 0Þk2 þ kvðtk þ 0Þk2
¼ kxðtk þ 0Þ � x�k2 þ kyðtk þ 0Þ � y�k2
¼ kxðtkÞ þ IkðxðtkÞÞ � x�k2 þ kyðtkÞ þ JkðyðtkÞÞ � y�k2
¼ kð1� akÞðxðtkÞ � x�Þk2 þ kð1� bkÞðyðtkÞ � y�Þk2
okxðtkÞ � x�k2 þ kyðtkÞ � y�k2 ¼ kuðtkÞk2 þ kvðtkÞk2
oD.
The proof is completed. &
Theorem 2. Assume that ðH1Þ, ðH3Þ and ðH4Þ hold, then the system
(2.3) has an unique solution ðx�; y�Þ which is an equilibrium point of
the system (2.2).
Proof. From Theorem 1, all solutions of system (2.2) are bounded,i.e., for any initial value, there exists some constant D040, suchthat
kxðtÞk2oD0; kyðtÞk2oD0.
So
Pf ðyðtÞÞ þM
Z 10
KðsÞf ðyðtÞÞdsþ I
��������
2
pkPk2xkyðtÞk2 þ kMk2xZ 1
0KðsÞyðtÞds
��������
2
þ kIk2
okPk2xD0 þ kMk2xD0l1 þ kIk29p1, (3.4)
QgðxðtÞÞ þ N
Z 10
RðsÞgðxðtÞÞdsþ J
��������
2
pkQk2ZkxðtÞk2 þ kNk2ZZ 1
0RðsÞxðtÞds
��������
2
þ kJk2
okQk2ZD0 þ kNk2ZD0l2 þ kJk29p2. (3.5)
Define SðtÞ ¼ AxðtÞ; ZðtÞ ¼ ByðtÞ such that
SðtÞ ¼ Pf ðB�1ZðtÞÞ þM
Z 10
KðyÞf ðB�1ZðtÞÞdyþ I,
ZðtÞ ¼ QgðA�1SðtÞÞ þ N
Z 10
RðyÞgðA�1SðtÞÞdyþ J.
It follows from (3.4) and (3.5) that the set O R2n defined by
O ¼ fðS; ZÞjkSk2pp1; kZk2pp2g�
Consider a mapping F : O! O defined by
If ðS; ZÞ, ðS̄; Z̄Þ are any two points of O, then
kFðS; ZÞ � FðS̄; Z̄Þk2
¼ Pðf ðB�1ZðtÞÞ � f ðB�1Z̄ðtÞÞÞ þM
Z 10
KðyÞf ðB�1ZðtÞÞdy�����
�
Z 10
KðyÞf ðB�1Z̄ðtÞÞdy�����
2
þ Q ðgðA�1SðtÞÞ � gðA�1S̄ðtÞÞÞ þ N
Z 10
RðyÞgðA�1SðtÞÞdy�����
�
Z 10
RðyÞgðA�1S̄ðtÞÞdy�����
2
pkPk2kf ðB�1ZðtÞÞ � f ðB�1Z̄ðtÞÞk2 þ kMk2
Z 10
KðyÞðf ðB�1ZðtÞÞ
����� f ðB�1Z̄ðtÞÞÞdy2 þ kQk2kgðA
�1SðtÞÞ � gðA�1S̄ðtÞÞk2
þ kNk2
Z 10
RðyÞðgðA�1SðtÞÞ � gðA�1S̄ðtÞÞÞdy����
����2
pkPk2xkB�1k2kZðtÞ � Z̄ðtÞk2 þ kMk2xkB
�1k2kZðtÞ � Z̄ðtÞk2l1
þ kQk2ZkA�1k2kSðtÞÞ � S̄ðtÞk2 þ kNk2ZkA
�1k2kSðtÞÞ � S̄ðtÞk2l2
¼ðkPk2 þ l1kMk2Þx
b0kZðtÞ � Z̄ðtÞk2 þ
ðkQk2 þ l2kNk2ÞZa0
kSðtÞ � S̄ðtÞk2
pgðkZðtÞ � Z̄ðtÞk2 þ kSðtÞ � S̄ðtÞk2Þ
¼ gkðS; ZÞ � ðS̄; Z̄Þk2,
ARTICLE IN PRESS
J. Zhou, S. Li / Neurocomputing 72 (2009) 1688–1693 1691
where g ¼maxfðkPk2 þ l1kMk2Þx=b0, ðkQk2 þ l2kNk2ÞZ=a0g, ob-viously, 0ogo1, then it follows by the well known contractionmapping principle that the mapping F has a unique point ðS�; Z�Þ.We can easily see that (2.3) satisfies, and this completes theproof. &
4. Stability of the equilibrium point of BAM neural networkswith impulses
Theorem 3. Suppose that the conditions ðH1Þ, ðH2Þ, ðH3Þ and ðH4Þ
hold, then the equilibrium point of system (2.4) is globally
exponentially stable.
Proof. We first choose the constant e in ðH2Þ with 0oeoa� suchthat
W 0¼ max
ðkPk2 þ l1kMk2Þxa� � e
;ðkQk2 þ l2kNk2ÞZ
a� � e
� �o1. (4.1)
For any C 2 C, F 2 C, we shall prove that
kuðtÞk2 þ kvðtÞk2pEðkCk þ kFkÞe�et for all t40. (4.2)
To prove (4.2), we first show: for tatk, C 2 C, F 2 C, and b41,the following inequality holds,
kuðtÞk2 þ kvðtÞk2obEðkCk þ kFkÞe�et for all t40; tatk. (4.3)
If (4.3) is false, then there must be some t240, such that
kuðt2Þk2 þ kvðt2Þk2 ¼ bEðkCk þ kFkÞe�et2 (4.4)
and
kuðtÞk2 þ kvðtÞk2pbEðkCk þ kFkÞe�et ; 0ptpt2; tatk. (4.5)
By ðH1Þ–ðH4Þ and (4.1), (4.4), (4.5), we have
kuðt2Þk2 þ kvðt2Þk2
pe�a0t2kCk þZ t2
0e�a0ðt2�yÞkPk2xkvðyÞk2dy
þ
Z t2
0e�a0ðt2�yÞ
Z 10kMk2xkvðy� sÞk2kKðsÞk2ds
� �dy
þ e�b0t2kFk þZ t2
0e�b0ðt2�yÞkQk2ZkuðyÞk2dy
þ
Z t2
0e�b0ðt2�yÞ
Z 10kNk2Zkuðy� sÞk2kRðsÞk2ds
� �dy
pe�a0t2kCk þZ t2
0e�a0ðt2�yÞkPk2xbEkFke�eydy
þ
Z t2
0e�a0ðt2�yÞ
Z 10kMk2xbEkFke�eðy�sÞkKðsÞk2ds
� �dy
þ e�b0t2kFk þZ t2
0e�b0ðt2�yÞkQk2ZbEkCke�eydy
þ
Z t2
0e�b0ðt2�yÞ
Z 10kNk2ZbEkCke�eðy�sÞkRðsÞk2ds
� �dy
pe�b0t2kFk þ kPk2xbEkFkZ t2
0e�a0ðt2�yÞe�eydy
þ kMk2xbEkFkZ t2
0e�a0ðt2�yÞe�eydy
Z 10keesKðsÞk2ds
þ e�a0t2kCk þ kQk2ZbEkCkZ t2
0e�b0ðt2�yÞe�eydy
þ kNk2ZbEkCkZ t2
0e�b0ðt2�yÞe�eydy
Z 10keesRðsÞk2ds
pe�b0t2kFk þ kPk2xbEkFkZ t2
0e�a0ðt2�yÞe�eydy
þ kMk2xbEkFkl1
Z t2
0e�a0ðt2�yÞe�eydy
þ e�a0t2kCk þ kQk2ZbEkCkZ t2
0e�b0ðt2�yÞe�eydy
þ kNk2ZbEkCkl2
Z t2
0e�b0ðt2�yÞe�eydy
¼ bEkFke�et2e�b0t2þet2
bEþ kPk2x
Z t2
0eet2 e�a0ðt2�yÞe�eydy
�
þkMk2xl1
Z t2
0eet2 e�a0ðt2�yÞe�eydy
�
þ bEkCke�et2e�a0t2þet2
bEþ kQk2Z
Z t2
0eet2 e�b0ðt2�yÞe�eydy
�
þkNk2Zl2
Z t2
0eet2 e�b0ðt2�yÞe�eydy
�
¼ bEkFke�et2e�ðb0�eÞt2
bEþ ðkPk2xþ l1xkMk2Þ
1� e�ða0�eÞt2
a0 � e
� �
þ bEkCke�et2e�ða0�eÞt2
bEþ ðkQk2Zþ l2ZkNk2Þ
1� e�ðb0�eÞt2
b0 � e
� �
pbEkFke�et2e�ða
��eÞt2
bEþ ðkPk2 þ l1kMk2Þx
1� e�ða��eÞt2
a� � e
� �
þ bEkCke�et2e�ða
��eÞt2
bEþ ðkQk2Zþ l2ZkNk2Þ
1� e�ða��eÞt2
a� � e
� �
¼ bEkFke�et21
bE�ðkPk2 þ l1kMk2Þx
a� � e
� �e�ða
��eÞt2
�
þðkPk2 þ l1kMk2Þx
a� � e
�
þ bEkCke�et21
bE�ðkQk2 þ l2kNk2ÞZ
a� � e
� �e�ða
��eÞt2
�
þðkQk2 þ l2kNk2ÞZ
a� � e
�obEkFke�et2 þ bEkCke�et2 ¼ bEðkFk þ kCkÞe�et2 ,
which contradicts the equality of (4.4), and so (4.3) holds. Letb! 1, then (4.2) holds. For t ¼ tk, by ðuðtkÞ; vðtkÞÞ ¼ ðuðtk �
0Þ; vðtk � 0Þ and ðH4Þ,
kuðtk þ 0Þk2 þ kvðtk þ 0Þk2okuðtkÞk2 þ kvðtkÞk2pEðkFk þ kCkÞe�etk .
This completes the proof. &
Remark 1. In [1], Arik and Tavsanoglu considered the followingmodel:
duiðtÞ
dt¼ �aiuiðtÞ þ
Xn
j¼1
wijgjðzjðt � tÞÞ þ Ii; i ¼ 1;2; . . . ;n;
dzjðtÞ
dt¼ �bjzjðtÞ þ
Xn
i¼1
vjigiðuiðt � tÞÞ þ Jj; j ¼ 1;2; . . . ;m:
8>>>>>><>>>>>>:
Obviously, this is a BAM neural networks with constant timedelays. This is a special case of (2.1), When DxðtkÞ ¼ 0;DyðtkÞ ¼ 0,pij ¼ qij ¼ 0, and f jðyjðt � sÞ ¼ gjðxjðt � sÞ ¼ f jðyjðt � tÞ ¼ gjðxjðt � tÞ.If we take A ¼ B ¼ ð1Þ;W ¼ V ¼ ð1Þ;a ¼ b ¼ 3=2 andkðyÞ ¼ rðyÞ ¼ 10�y; e ¼ 0:1, by simple calculation, O140;O240,which means the condition of Theorem 1 in [1] failed, but ðH3Þ
holds.
Furthermore, when proving the existence of the equilibrium
point of the model, Arik and Tavsanoglu required that the
activation functions be bounded. Here we do not need this
assumption. We obtain GES for the equilibrium point of our
model, while Arik and Tavsanoglu obtain global asymptotic
stability for the equilibrium point of their model.
ARTICLE IN PRESS
−2 0 2 4 6 8 10 12−1
−0.5
0
0.5
1
1.5
time t
stat
es x
(t), y
(t)
x(t)
y(t)
Fig. 1. Stability of impulsive BAM neural networks (5.1) with distributed delays
where I1 ¼ I2 ¼ 0; ð0;0ÞT is the equilibrium point.
J. Zhou, S. Li / Neurocomputing 72 (2009) 1688–16931692
Remark 2. When DxðtkÞ ¼ 0;DyðtkÞ ¼ 0 and P ¼ Q ¼ 0, the model(2.2) becomes the same as Zhao [1], Zhao [8, Theorem 7] provedthe global asymptotic stability of the equilibrium point if ðH1Þwassatisfied and
rðWÞo1; W ¼ K�1jMjx0 þ K�1
jNjZ0, (4.6)
where K ¼ diagfkig, with ki ¼ minfai; big, jMj ¼ ðjmijjÞn�n,jNj ¼ ðjnijjÞn�n, x0 ¼ diagfxg, Z0 ¼ diagfZg, ðH3Þ is weaker than(4.6). In fact, if we take M ¼ N ¼ ð11
�11 Þ, x1 ¼ x2 ¼ Z1 ¼ Z2 ¼
12,
a1 ¼ a2 ¼ b1 ¼ b2 ¼ 2, rij ¼ kij ¼ e�y; e ¼ 0:1, by simple calcula-tion, ðH3Þ holds. But
W ¼
1
20
01
2
0BBB@
1CCCA
1 1
1 1
! 1
20
01
2
0BBB@
1CCCAþ
1
20
01
2
0BBB@
1CCCA
1 1
1 1
! 1
20
01
2
0BBB@
1CCCA
¼
1
2
1
2
1
2
1
2
0BBB@
1CCCA,
where rðWÞ ¼ 1, the condition (4.6) failed.
Remark 3. When discussing the GES of BAM, many authors usedthe properties of nonsingular M-matrix. Here we point out ourresults do not use these properties. So we can deal with the casewhich does not satisfy M-matrix properties.
Li and Yang [9] discussed the system (2.1) with pij ¼ qji ¼
0; i ¼ 1;2; . . . ;n; j ¼ 1;2; . . . ;n. Li and Yang [9, Theorem 4] proved
the GES of the unique equilibrium point if ðH1Þ, ðH2Þ, ðH4Þ hold
and
S ¼A �TjNj
�LjMj B
!is an M-matrix where T ¼ diagðZ1; . . . ;ZnÞ,
L ¼ diagðx1; . . . ; xnÞ. (4.7)
However, let n ¼ 1; a1 ¼ b1 ¼ 1, M ¼ N ¼ ð1Þ, kðyÞ ¼ rðyÞ ¼ 10�y,
x ¼ Z ¼ 32, e ¼ 0:1.
By simple calculation,
Z 10
e0:1y10�ydy ¼1
ln 10� 0:1¼ l1 ¼ l2
ðH3Þ holds. But S ¼1 �
3
2
�3
21
0BB@
1CCA is not an M-matrix.
Remark 4. When DxðtkÞ ¼ 0;DyðtkÞ ¼ 0, the model (2.2) becomesthe same as Song and Cao [10]. In [10], Song and Cao constructed anew Lyapunov functional, employed the homeomorphism theory,M-matrix theory and Hardy inequality to obtain a sufficientcondition to ensure the existence, uniqueness and GES of theequilibrium point for the model. By the same discussion withRemarks 3, Theorem 3 can deal with the case which does notsatisfy M-matrix properties.
5. Illustrative examples
Example 1. Let n ¼ 1, f jðxÞ ¼ giðxÞ ¼ jxj; kijðsÞ ¼ rjiðsÞ ¼ e�s.Clearly, the activation functions f jð�Þ; gið�Þ satisfy hypothesis ðH1Þ
with xi ¼ Zi ¼ 1 and the kernel kijð�Þ; rjið�Þ satisfy hypothesis ðH2Þ.Consider impulsive BAM neural networks with distributed delaysas follows:
dxðtÞ
dt¼ �AxðtÞ þ Pf ðyðtÞÞ þM
Z 10
KðsÞf ðyðt � sÞÞdsþ I;
t40; tatk;
DxðtkÞ ¼ IkðxðtkÞÞ; k ¼ 1;2; . . . ;
dyðtÞ
dt¼ �ByðtÞ þ QgðxðtÞÞ þ N
Z 10
RðsÞgðxðt � sÞÞdsþ J;
t40; tatk;
DyðtkÞ ¼ JkðyðtkÞÞ; k ¼ 1;2; . . . ;
8>>>>>>>>>>>>><>>>>>>>>>>>>>:
(5.1)
where A ¼ B ¼ ð2:5Þ, P ¼ Q ¼ M ¼ N ¼ ð1Þ, I ¼ 5, J ¼ 3,ak ¼ 1þ 1
2 sinð1þ kÞ, bk ¼ 1þ 23 cosð2k3
Þ;0ot1ot2o � � � is a strictlyincreasing sequence such that limt!1tk ¼ þ1; k 2 Zþ.
By simple calculation, we obtainZ 10
e0:1se�sds ¼1
0:9¼ l1 ¼ l2,
which implies ðH2Þ holds, and
W ¼ maxðkPk2 þ l1kMk2Þx
a�;ðkQk2 þ l2kNk2ÞZ
a�
� �o1,
which implies ðH3Þ holds: obviously, ðH4Þ holds with 0oako2,
0obko2, k 2 Zþ.
So, by Theorem 3, the equilibrium point of system (5.1) is
globally exponentially stable (Fig. 1).
6. Conclusion
In this paper, GES of impulsive BAM neural networks withdistributed delays have been studied. Some sufficient conditionsfor the existence and GES of the equilibrium point have beenestablished. These obtained results are new and they complementpreviously known results. Moreover, an example is given toillustrate the effectiveness of our results. Moreover, our methodsin the paper may be extended for more complex networks.
Acknowledgments
The authors would like to thank the Associate Editor and fourreviewers for their careful reading of this paper and helpful
ARTICLE IN PRESS
J. Zhou, S. Li / Neurocomputing 72 (2009) 1688–1693 1693
comments, which have been very useful for improving the qualityof this paper. This work is supported by National ScienceFoundation of PR China under Grant 10671133 and by ScientificResearch Fund of Sichuan Provincial Education Department.
References
[1] S. Arik, V. Tavsanoglu, Global asymptotic stability analysis of bidirectionalassociative memory neural networks with constant time delays, Neurocom-puting 68 (2005) 161–176.
[2] J. Cao, Global asymptotic stability of delayed bi-directional associativememory neural networks, Appl. Math. Comput. 142 (2003) 333–339.
[3] Z. Gui, J. Zhang, C. Xing, Global exponential stability of impulsive cellularneural network with delay, in: Third International Conference on NaturalComputation, 2007, pp. 455–459.
[4] J. Hopfield, Neurons with graded response have collective computationalproperties like those of two sate neurons, Proc. Natl. Acad. Sci. 81 (1984)3088–3092.
[5] R. Horn, C. Johnson, Matrix Analysis, Cambridge University Press, London,1990.
[6] B. Kosko, Adaptive bidirectional associative memories, Appl. Opt. 26 (1987)4947–4960.
[7] B. Kosko, Bidirectional associative memories, IEEE Trans. Syst. Man Cybern. 18(1988) 49–60.
[8] S.Y. Li, D.Y. Xu, Exponential attraction domain and exponential convergentrate of associative memory neural networks with delays, Control Theory Appl.19 (2002) 442–444.
[9] Y. Li, C. Yang, Global exponential stability analysis on impulsive BAM neuralnetworks with distributed delays, J. Math. Anal. Appl. 324 (2006) 1125–1139.
[10] Q. Song, J. Cao, Global exponential stability of bidirectional associativememory neural networks with distributed delays, J. Comput. Appl. Math. 202(2007) 266–279.
[11] Y. Wang, W. Xiong, Q. Zhou, B. Xiao, Y. Yu, Global exponential stability ofcellular neural networks with continuously distributed delays and impulses,Phys. Lett. A 350 (2006) 89–95.
[12] D. Xu, Z. Yang, Impulsive delay differential inequality and stability of neuralnetworks, J. Math. Anal. Appl. 305 (2005) 107–120.
[13] Z. Yang, D. Xu, Impulsive effects on stability of CohenCGrossberg neuralnetworks with variable delays, Appl. Math. Comput. 177 (2006) 63–78.
[14] Q. Zhang, Stability condition for impulsive Hopfield neural networks withtime delay, Dyn. Cont. Discrete Impulsive Syst. Ser. A Math. Anal., Part 1(Suppl. S 13) (2006) 35–38.
[15] Q. Zhang, X. Wei, J. Xu, Global asymptotic stability of cellular neural networkswith infinite delay, Neural Network World 15 (2005) 579–589.
[16] Q. Zhang, C. Zhou, Dynamics of Hopfield neural networks with continuouslydistributed delays, Dyn. Cont. Discrete Impulsive Syst. Ser. A Math. Anal., Part2 (Suppl. S 13) (2006) 541–544.
[17] H. Zhao, Global stability of bidirectional associative memory neural networkswith distributed delays, Phys. Lett. A 297 (2002) 182–190.
Jie Zhou received the Master degree from SichuanNormal University, Chengdu, China. He is with collegeof science ,Sichuan University of Science and Engineer-ing, Zigong, Sichuan 643000, PR China. He is currentlya teaching assistant of college of science in SichuanUniversity of Science and Engineering, Zigong, PRChina.
Shuyong Li received the Ph.D. degree from SichuanUniversity, Chengdu, China, and the Post-DoctoralFellow in School of Science, Xi’an Jiaotong University,Xi’an, China. He was with College of Mathematics andSoftware science at Sichuan Normal University,Cheng-du, China. He is currently a Professor of SichuanNormal university. He also is the author or coauthorof more than 40 journal papers. His research interestsinclude functional differential systems, partial func-tional differential systems and neural networks.