Abstract

A class of Cohen-Grossberg-type BAM neural networks with distributed delays and impulses are investigated in this paper. Sufficient conditions to guarantee the uniqueness and global exponential stability of the periodic solutions of such networks are established by using suitable Lyapunov function, the properties of 𝑀-matrix, and some suitable mathematical transformation. The results in this paper improve the earlier publications.

1. Introduction

The research of neural networks with delays involves not only the dynamic analysis of equilibrium point but also that of periodic oscillatory solution. The dynamic behavior of periodic oscillatory solution is very important in learning theory due to the fact that learning usually requires repetition [1, 2].

Cohen and Grossberg proposed the Cohen-Grossberg neural networks (CGNNs) in 1983 [3]. Kosko proposed bi directional associative memory neural networks (BAMNNs) in 1988 [4]. Some important results for periodic solutions of delayed CGNNs have been obtained in [510]. Xiang and Cao proposed a class of Cohen-Grossberg BAM neural networks (CGBAMNNs) with distributed delays in 2007 [11]; in addition, many evolutionary processes are characterized by abrupt changes at certain time; these changes are called to be impulsive phenomena, which are included in many neural networks such as Hopfield neural networks, BAM neural networks, CGNNs, and CGBAMNNs and can affect dynamical behaviors of the systems just as time delays. The results for periodic solutions of CGBAMNNs with or without impulses are obtained in [1115].

The objective of this paper is to study the existence and global exponential stability of periodic solutions of CGBAMNNs with distributed delays by using suitable Lyapunov function, the properties of 𝑀-matrix, and some suitable mathematical transformation. Comparing with the results in [13, 14], improved results are successively obtained, the conditions for the existence and globally exponential stability of the periodic solution of such system without impulses have nothing to do with inputs of the neurons and amplification functions; and we also point that CGBAMNNs model is a special case of CGNNs model, many results of CGBAMNNs can be directly obtained from the results of CGNNs.

The rest of this paper is organized as follows. Preliminaries are given in Section 2. Sufficient conditions which guarantee the uniqueness and global exponential stability of periodic solutions for CGBAMNNs with distributed delays and impulses are given in Section 3. Two examples are given in Section 4 to demonstrate the main results.

2. Preliminaries

Consider the following periodic CGNNs model with distributed delays and impulses:̇𝑥𝑖(𝑡)=𝑎𝑖𝑥𝑖𝑏(𝑡)𝑖𝑡,𝑥𝑖(𝑡)𝑛𝑗=1𝑝𝑖𝑗(𝑡)𝑓𝑗𝜌𝑗𝑥𝑗(𝑡)𝑛𝑗=1𝑢𝑖𝑗(𝑡)0+𝑘𝑖𝑗(𝑠)𝑓𝑗𝜌𝑗x𝑗(𝑡𝑠)𝑑𝑠𝐼𝑖(𝑡),𝑡>0,𝑡𝑡𝑘,Δ𝑥𝑖𝑡𝑘=𝛾𝑖𝑘𝑥𝑖𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,(2.1) where 1𝑖𝑛, 𝑡>0, and 𝑍+={1,2,}. 𝑥𝑖(𝑡) denotes the state variable of the 𝑖th neuron, 𝑓𝑗() denotes the signal function of the 𝑗th neuron at time 𝑡; 𝐼𝑖 denotes input of the 𝑖th neuron at time 𝑡; 𝑎𝑖() represents amplification function; 𝑏𝑖(𝑡,) is appropriately behaved function; 𝑝𝑖𝑗(𝑡) and 𝑢𝑖𝑗(𝑡) are connection weights of the neural networks at time 𝑡; respectively, 𝜌𝑗 is positive constant, which corresponds to the neuronal gain associated with the neuronal activations and 𝑘𝑖𝑗 corresponds to the delay kernel function; 𝑝𝑖𝑗(𝑡) and 𝑢𝑖𝑗(𝑡) are continuously periodic functions on [0,+) with common period 𝑇>0.

Δ𝑥𝑖(𝑡𝑘)=𝑥𝑖(𝑡+𝑘)𝑥𝑖(𝑡𝑘); 𝑡𝑘 is called impulsive moment and satisfies 0<𝑡1<𝑡2<,lim𝑘+𝑡𝑘=+; 𝑥𝑖(𝑡) and 𝑥𝑖(𝑡+) denote the left-hand and right-hand limits at 𝑡𝑘; respectively, we always assume 𝑥𝑖(𝑡𝑘)=𝑥𝑖(𝑡𝑘) and 𝑥𝑖(𝑡𝑘)=𝑥𝑖(𝑡𝑘),𝑘𝑍+.

For system (2.1), we assume the following.(H1)The amplification function 𝑎𝑖() is continuous, and there exist constants 𝑎𝑖,𝑎𝑖 such that 0<𝑎𝑖𝑎𝑖(𝑥𝑖(𝑡))𝑎𝑖 for 1𝑖𝑛.(H2)The behaved function 𝑏𝑖(𝑡,) is 𝑇-periodic about the first argument; there exists continuous 𝑇-periodic function 𝛼𝑖(𝑡) such that𝑏𝑖(𝑡,𝑥)𝑏𝑖(𝑡,𝑦)𝑥𝑦𝛼𝑖(𝑡)>0,(2.2)for all 𝑥𝑦,1𝑗𝑛.(H3)For activation function 𝑓𝑗(), there exists positive constant 𝐿𝑗 such that𝐿𝑗=sup𝑥y||||𝑓𝑗(𝑥)𝑓𝑗(𝑥)||||𝑥𝑦,(2.3)for all 𝑥𝑦, 1𝑗𝑛.(H4) The kernel function 𝑘𝑖𝑗(𝑠) is nonnegative continuous function on [0,+) and satisfies0+𝑠𝑒𝜆𝑠𝑘𝑖𝑗𝐾(𝑠)𝑑𝑠<+,𝑖𝑗(𝜆)=0+𝑒𝜆𝑠𝑘𝑖𝑗(𝑠)𝑑𝑠(2.4)is differentiable function for 𝜆[0,𝑟𝑖𝑗),0<𝑟𝑖𝑗<+, 𝐾𝑖𝑗(0)=1 and lim𝜆𝑟𝑖𝑗𝐾𝑖𝑗(𝜆)=+.(H5)There exists positive integer 𝑘0 such that 𝑡𝑘+𝑘0=𝑡𝑘+𝑇 and 𝛾𝑖(𝑘+𝑘0)=𝛾𝑖𝑘 hold.

Remark 2.1. A typical example of kernel function is given by 𝑘𝑖𝑗(𝑠)=(𝑠𝑟/𝑟!)𝑟𝑟+1𝑖𝑗𝑒𝑟𝑖𝑗𝑠 for 𝑠[0,+), where 𝑟𝑖𝑗(0,+),𝑟{0,1,,𝑛}. These kernel functions are called as the gamma memory filter [16] and satisfy condition (H4).
For any continuous function 𝑆(𝑡) on [0,𝑇], 𝑆 and 𝑆 denote min𝑡[0,𝑇]{|𝑆(𝑡)|} and max𝑡[0,𝑇]{|𝑆(𝑡)|}, respectively.
For any 𝑥(𝑡)=(𝑥𝑖(𝑡),𝑥2(𝑡),,𝑥𝑘(𝑡))𝑇𝑅𝑘, 𝑡>0, define 𝑥(𝑡)=𝑘𝑖=1|𝑥𝑖(𝑡)|, and for any 𝜑(𝑠)=(𝜑1(𝑠),𝜑2(𝑠),,𝜑𝑘(𝑠))𝑇𝑅𝑘, 𝑠(,0], define 𝜑(𝑠)=sup𝑠(,0]𝑘𝑖=1|𝜑𝑖(𝑠)|.
Denote ]PC(,0,𝑅𝑘[]={𝜓,0𝑅𝑘]𝜓𝑠𝜓(𝑠)isboundedandcontinuousforallbutatmostanitenumberofpoints𝑠(,0,andatthesepoints𝑠,+,𝜓(𝑠)existand𝜓(𝑠)=𝜓(𝑠)}.(2.5) Then PC((,0],𝑅𝑘) is a Banach space with respect to .
The initial conditions of system (2.1) are given by 𝑥𝑖(𝑠)=𝜑𝑖(𝑠),𝑠0,1𝑖𝑛,(2.6) where 𝜑(𝑠)=(𝜑1(𝑠),𝜑2(𝑠),,𝜑𝑛(𝑠))PC([,0],𝑅𝑛).
Let 𝑥(𝑡,𝜑)=(𝑥1(𝑡,𝜑),𝑥2(𝑡,𝜑),,𝑥𝑛(𝑡,𝜑)𝑇) denote any solution of the system (2.1) with initial value 𝜑PC((,0],𝑅𝑛).

Definition 2.2. A solution 𝑥(𝑡,𝜑) of system (2.1) is said to be globally exponentially stable, if there exist two constants 𝜆>0,𝑀>0 such that 𝑥(𝑡,𝜓)𝑥(𝑡,𝜑)𝑀𝜓𝜑𝑒𝜆𝑡,𝑡>0,(2.7) for any solutions 𝑥(𝑡,𝜓) of system (2.1).

Definition 2.3. A real matrix 𝐴=(𝑎𝑖𝑗)𝑛×𝑛is said to be a nonsingular 𝑀-matrix if 𝑎𝑖𝑗0(𝑖,𝑗=1,2,,𝑛,𝑖𝑗), and all successive principle minors of 𝐴 are positive.

Lemma 2.4 (see [17]). A matrix with nonpositive off-diagonal elements 𝐴=(𝑎𝑖𝑗)𝑛×𝑛 is a nonsingular 𝑀-matrix if and if only there exists a vector 𝑝=(𝑝𝑖)1×𝑛>0 such that 𝑝𝑇𝐴>0 or 𝐴𝑝 holds.

Lemma 2.5. Under assumptions (H1)–(H5), system (2.1) has a 𝑇-periodic solution which is globally exponentially stable, if the following conditions hold.(H6)=𝐴𝐶 is a nonsingular 𝑀-matrix, where𝛼𝐴=diag1,𝛼2,,𝛼𝑛𝑐,𝐶=𝑖𝑗𝑛×𝑛,𝑐𝑖𝑗=𝑝𝑖𝑗+𝑢𝑖𝑗𝜌𝑗𝐿𝑗.(2.8)(H7)𝑎𝑖((1𝛾𝑖𝑘)𝑠)|1𝛾𝑖𝑘|𝑎𝑖(𝑠), for all 𝑠R,𝑖=1,2,,𝑛.

Proof. Let 𝑥(𝑡,𝜓1) and 𝑥(𝑡,𝜓2) be two solutions of system (2.1) with initial value 𝜓1=(𝜑1,𝜑2,,𝜑𝑛) and 𝜓2=(𝜁1,𝜁2,,𝜁𝑛)PC((,0],𝑅𝑛), respectively.
Let𝐹𝑖(𝜃)=𝜇𝑖𝛼𝑖𝜃𝑎𝑖𝑛𝑗=1𝜇𝑗𝑝𝑗𝑖+𝑢𝑗𝑖𝐾𝑗𝑖𝜌(𝜃)𝑖𝐿𝑖,𝑖=1,2,,𝑛.(2.9) Since is a nonsingular 𝑀-matrix according to condition (H6), 𝑇 is also a nonsingular 𝑀-matrix; we know from Lemma 2.4 that there exists a vector 𝑝=(𝜇1,𝜇2,,𝜇𝑛)𝑇 such that 𝑇𝑝>0; that is, 𝜇𝑖𝛼𝑖𝑛𝑗=1𝜇𝑗𝑝𝑗𝑖+𝑢𝑗𝑖𝜌𝑖𝐿𝑖>0,(2.10) for 1𝑖𝑛, which indicates that 𝐹𝑖(0)>0. Since 𝐹𝑖(𝜃) are continuous and differential on [0,𝑟𝑗𝑖) and lim𝜃𝑟𝑗𝑖𝐹𝑖(𝜃)= according to condition (H4), 𝐹𝑖(𝜃)<0 for 𝜃[0,𝑢𝑗𝑖). There exist constants 𝜃𝑖 such that 𝐹𝑖(𝜃𝑖)=0 for 𝑖=1,2,,𝑛. So we can choose 𝜃0<𝜆min1,𝜃2,,𝜃𝑛,(2.11) such that 𝐹𝑖(𝜆)0.(2.12) Define 𝑋𝑖||𝑥(𝑡)=𝑖𝑡,𝜓2𝑥𝑖𝑡,𝜓1||.(2.13)
Now we define a Lyapunov function 𝑉(𝑡) by 𝑉(𝑡)=𝑛𝑖=1𝜇𝑖𝑉𝑖(𝑡)+𝑛𝑗=1𝑢𝑖𝑗𝐿𝑗𝜌𝑗0+𝑘𝑖𝑗(𝑠)𝑡𝑡𝑠𝑋𝑗(𝜇)𝑒𝜆(𝑠+𝜇),𝑑𝜇𝑑𝑠(2.14) in which 𝑉𝑖(𝑡)=𝑒𝜆𝑡𝑥sign𝑖𝑡,𝜓2𝑥𝑖𝑡,𝜓1𝑥𝑖(𝑡,𝜓2)𝑥𝑖𝑡,𝜓11𝑎𝑖(𝑠)𝑑𝑠.(2.15) for 𝑖=1,2,,𝑛.
When 𝑡𝑡𝑘,𝑘𝑍+, calculating the upper right derivative of 𝑉(𝑡) along solution of (2.1), similar to proof of Theorem  3.1 in [10], corresponding to case in which 𝑟1,𝑣𝑖𝑗𝑙(𝑡)=0 in [10], we obtain from (2.12)–(2.15) that𝐷+𝑉(𝑡)(2.1)𝑒𝜆𝑡𝑛𝑖=1𝜇𝑖𝛼𝑖𝜆𝑎𝑖𝑋𝑖(𝑡)𝑛𝑛𝑖=1𝑗=1𝜇𝑖𝑝𝑖𝑗+𝑢𝑖𝑗𝐾𝑖𝑗𝜌(𝜆)𝑗𝐿𝑗𝑋𝑗(𝑡)=𝑒𝜆𝑡𝑛𝑖=1𝜇𝑖𝛼𝑖𝜆𝑎𝑖𝑋𝑖(𝑡)𝑛𝑛𝑖=1𝑗=1𝜇𝑖𝑝𝑗𝑖+𝑢𝑗𝑖𝐾𝑗𝑖𝜌(𝜆)𝑖𝐿𝑖𝑋𝑖(𝑡)=𝑒𝑛𝜆𝑡𝑖=1𝐹𝑖(𝜆)𝑋𝑖(𝑡)0.(2.16) When 𝑡=𝑡𝑘,𝑘𝑍+, we have 𝑉𝑖𝑡+𝑘=𝑒𝜆𝑡+𝑥sign𝑖𝑡+𝑘,𝜓2𝑥𝑖𝑡+𝑘,𝜓1𝑥𝑖(𝑡+𝑘,𝜓2)𝑥𝑖(𝑡+𝑘,𝜓1)1𝑎𝑖(𝑠)𝑑𝑠=𝑒𝜆𝑡sign1𝛾𝑖𝑘𝑥𝑖𝑡𝑘,𝜓2𝑥𝑖𝑡𝑘,𝜓1(1𝛾𝑖𝑘)𝑥𝑖(𝑡𝑘,𝜓2)(1𝛾𝑖𝑘)𝑥𝑖(𝑡𝑘,𝜓1)1𝑎𝑖(𝑠)𝑑𝑠,=𝑒𝜆𝑡𝑥sign𝑖𝑡𝑘,𝜓2𝑥𝑖𝑡𝑘,𝜓1𝑥𝑖(𝑡𝑘,𝜓2)𝑥𝑖(𝑡𝑘,𝜓1)||1𝛾𝑖𝑘||𝑎𝑖1𝛾𝑖𝑘𝑠𝑑𝑠,(2.17) which, together with (H7), leads to 𝑉𝑖𝑡𝑘𝑉𝑖𝑡+𝑘=𝑒𝜆𝑡𝑥sign𝑖𝑡𝑘,𝜓2𝑥𝑖𝑡𝑘,𝜓1𝑥𝑖(𝑡𝑘,𝜓2)𝑥𝑖(𝑡𝑘,𝜓1)1𝑎𝑖||(𝑠)1𝛾𝑖𝑘||𝑎𝑖1𝛾𝑖𝑘𝑠𝑑𝑠,𝑒𝜆𝑡𝑥sign𝑖𝑡𝑘,𝜓2𝑥𝑖𝑡𝑘,𝜓1𝑥𝑖(𝑡𝑘,𝜓2)𝑥𝑖(𝑡𝑘,𝜓1)𝑎𝑖1𝛾𝑖𝑘𝑠||1𝛾𝑖𝑘||𝑎𝑖(𝑠)𝑎𝑖(𝑠)𝑎𝑖1𝛾𝑖𝑘𝑠𝑑𝑠0,(2.18) that is, 𝑉𝑖𝑡+𝑘𝑉𝑖𝑡𝑘.(2.19) It follows that 𝑉𝑡+𝑘=𝑛𝑖=1𝜇𝑖𝑉𝑖𝑡++𝑛𝑗=1𝑢𝑖𝑗𝐿𝑗𝜌𝑗0+𝑘𝑖𝑗(𝑠)𝑡+𝑡+𝑠𝑋𝑗(𝜇)𝑒𝜆(𝑠+𝜇)𝑑𝜇𝑑𝑠𝑛𝑖=1𝜇𝑖𝑉𝑖(𝑡)+𝑛𝑗=1𝑢𝑖𝑗𝐿𝑗𝜌𝑗0+𝑘𝑖𝑗(𝑠)𝑡𝑡𝑠𝑋𝑗(𝜇)𝑒𝜆(𝑠+𝜇)𝑡𝑑𝜇𝑑𝑠=𝑉𝑘.(2.20) Then we have 𝑉(𝑡)𝑉(0).(2.21) On the other hand, from (2.14), we have 𝑉(𝑡)𝑚0𝑒𝑛𝜆𝑡𝑖=1||𝑥𝑖𝑡,𝜓2𝑥𝑖𝑡,𝜓1||,𝑉(0)𝑀0sup𝑛𝑙(,0]𝑖=1||𝜑𝑖(𝑙)𝜁𝑖||,(𝑙)(2.22) in which 𝑚0=min1𝑖𝑛𝜇𝑖𝑎𝑖,𝑀0𝑀=max1,𝑀2,𝑀1=max1𝑖𝑛𝜇𝑖𝑎𝑖,𝑀2=𝑛𝑗=1𝜇𝑗max1𝑖𝑛𝑢𝑗𝑖𝜌𝑖𝐿𝑖0+𝑠𝑒𝜆𝑠max1𝑖𝑛𝑘𝑗𝑖(𝑠)𝑑𝑠.(2.23) Hence, from (2.21) and (2.22), we know that the following inequality holds for 𝑡>0: 𝑥𝑡,𝜓2𝑥𝑡,𝜓1𝜓𝑀2𝜓1𝑒𝜆𝑡,(2.24) in which 𝑀=𝑀0/𝑚0.
We can always choose a positive integer 𝑁 such that 𝑒𝜆0𝑁𝑇𝑀1/2 and define a Poincaré mapping 𝑃𝐶𝐶 by 𝑃(𝜉)=𝑥𝑇(𝜉); we have 𝑃𝑁𝜓2𝑃𝑁𝜓112𝜓2𝜓1,(2.25) which implies that 𝑃𝑁 is a contraction mapping. Similar to [10], using contraction mapping principle, we know that system (2.1) has a 𝑇-periodic solution which is globally exponentially stable. This completes the proof.

Remark 2.6. The result above also holds for (2.1) without impulses, and the existence and globally exponential stability of the periodic solution for (2.1) have nothing to do with amplification functions and inputs of the neuron. The results in [5] have more restrictions than Lemma 2.5 in this paper because conditions for the ones in [5] are relevant to amplification functions.

3. Periodic Solutions of CGBAMNNs with Distributed Delays and Impulses

Consider the following periodic CGBAMNNs model with distributed delays:̇𝑥𝑖(𝑡)=𝑎𝑖𝑥𝑖𝑏(𝑡)𝑖𝑡,𝑥𝑖(𝑡)𝑚𝑗=1𝑝𝑖𝑗(𝑡)𝑓𝑗𝜌𝑗𝑦𝑗(𝑡)𝑚𝑗=1𝑢𝑖𝑗(𝑡)0+𝑘𝑖𝑗(𝑠)𝑓𝑗𝜌𝑗𝑦𝑗(𝑡𝑠)𝑑𝑠𝐼𝑖(𝑡),𝑡>0,𝑡𝑡𝑘,Δ𝑥𝑖𝑡𝑘=𝛾𝑖𝑘𝑥𝑖𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,̇𝑦𝑗(𝑡)=𝑐𝑗𝑦𝑗𝑑(𝑡)𝑗𝑡,𝑦𝑗(𝑡)𝑛𝑖=1𝑞𝑗𝑖(𝑡)𝑔𝑖̃𝜌𝑖𝑥𝑖(𝑡)𝑛𝑖=1𝑣𝑗𝑖(𝑡)0+̃𝑘𝑗𝑖(𝑠)𝑔𝑖̃𝜌𝑖𝑥𝑖(𝑡𝑠)𝑑𝑠𝐽𝑗(𝑡),𝑡>0,𝑡𝑡𝑘,Δ𝑦𝑗𝑡𝑘=𝛿𝑗𝑘𝑦𝑗𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+(3.1) for 1𝑖𝑛,1𝑗𝑚, and 𝑍+={1,2,}; 𝑥𝑖(𝑡) and 𝑦𝑗(𝑡) denote the state variable of the 𝑖th neuron from the neural field 𝐹𝑋 and the 𝑗th neuron from the neural field 𝐹𝑌 at time 𝑡;  𝑓𝑗() and 𝑔𝑖() denote the signal functions of the 𝑗th neuron from the neural field 𝐹𝑌 and the 𝑖th neuron from the neural field 𝐹𝑋 at time t; respectively, 𝐼𝑖 and 𝐽𝑗 denote inputs of the 𝑖th neuron from the neural field 𝐹𝑋 and the 𝑗th neuron from the neural field 𝐹𝑌 at time 𝑡; respectively, 𝑎𝑖() and 𝑐𝑗() represent amplification functions; 𝑏𝑖(𝑡,) and 𝑑𝑗(𝑡,) are appropriately behaved functions; 𝑝𝑖𝑗(𝑡),𝑞𝑗𝑖(𝑡),𝑢𝑖𝑗(𝑡), and 𝑣𝑗𝑖(𝑡) are the connection weights; 𝜌𝑗,̃𝜌𝑖 are positive constants, which correspond to the neuronal gains associated with the neuronal activations; 𝑘𝑖𝑗 and ̃𝑘𝑗𝑖 correspond to the delay kernel functions; 𝑢𝑖𝑗(𝑡),𝑣𝑗𝑖(𝑡),𝑝𝑖𝑗(𝑡),𝑞𝑗𝑖(𝑡), 𝐼𝑖(𝑡), and 𝐽𝑗(𝑡) are all continuously periodic functions on [0,+) with common period 𝑇>0.

Δ𝑥𝑖(𝑡𝑘)=𝑥𝑖(𝑡+𝑘)𝑥𝑖(𝑡𝑘), Δ𝑦𝑗(𝑡𝑘)=𝑦𝑗(𝑡+𝑘)𝑦𝑗(𝑡𝑘); 𝑡𝑘 is called impulsive moment and satisfies 0<𝑡1<𝑡2<,lim𝑘+𝑡𝑘=+; 𝑥𝑖(𝑡),𝑦𝑗(𝑡) and 𝑥𝑖(𝑡+),𝑦𝑗(𝑡+) denote the left-hand and right-hand limits at 𝑡𝑘; respectively, we always assume 𝑥i(𝑡𝑘)=𝑥𝑖(𝑡𝑘),𝑦𝑗(𝑡𝑘)=𝑦𝑗(𝑡𝑘),𝑥𝑖(𝑡𝑘)=𝑥𝑖(𝑡𝑘), and 𝑦𝑗(𝑡𝑘)=𝑦𝑗(𝑡𝑘),𝑘𝑍+.

For system (3.1), we assume the following.(H8) Amplification functions 𝑎𝑖() and 𝑐𝑗() are continuous and there exist constants 𝑎𝑖,𝑎𝑖 and 𝑐𝑗,𝑐𝑗 such that 0<𝑎𝑖𝑎𝑖(𝑥𝑖(𝑡))𝑎𝑖,0<𝑐𝑗𝑐𝑗(𝑦𝑗(𝑡))𝑐𝑗,1𝑖𝑛,1𝑗𝑚.(H9)𝑏𝑖(𝑡,𝑢),𝑑𝑗(𝑡,𝑢) are 𝑇-periodic about the first argument, there exist continuous, 𝑇-periodic functions 𝛼𝑖(𝑡) and 𝛽𝑗(𝑡) such that 𝑏𝑖(𝑡,𝑥)𝑏𝑖(𝑡,𝑦)𝑥𝑦𝛼𝑖𝑑(𝑡)>0,𝑗(𝑡,𝑥)𝑑𝑗(𝑡,𝑦)𝑥𝑦𝛽𝑗(𝑡)>0(3.2) for all 𝑥𝑦,1𝑖𝑛,1𝑗𝑚.(H10)For activation functions 𝑓𝑗() and 𝑔𝑖(), there exist constant 𝐿𝑗 and 𝐿𝑖 such that 𝐿𝑗=sup𝑥𝑦||||𝑓𝑗(𝑥)𝑓𝑗(𝑦)||||,𝐿𝑥𝑦𝑖=sup𝑥𝑦||||𝑔𝑖(𝑥)𝑔𝑖(𝑦)||||𝑥𝑦,𝑥𝑦𝑅,1𝑖𝑛,1𝑗𝑚.(3.3)(H11)The kernel functions 𝑘𝑖𝑗(𝑠) and ̃𝑘𝑗𝑖(𝑠) are nonnegative continuous functions on [0,+) and satisfy 0+𝑠𝑒𝜆𝑠𝑘𝑖𝑗(𝑠)𝑑𝑠<+,0+𝑠𝑒𝜆𝑠̃𝑘𝑗𝑖𝐾(𝑠)𝑑𝑠<+,𝑖𝑗(𝜆)=0+𝑒𝜆𝑠𝑘𝑖𝑗𝐾(𝑠)𝑑𝑠,𝑗𝑖(𝜆)=0+𝑒𝜆𝑠̃𝑘𝑗𝑖(𝑠)𝑑𝑠,(3.4) are differentiable functions for 𝜆[0,𝑟𝑖𝑗) and 𝜆[0,̃𝑟𝑗𝑖); respectively, 0<𝑟𝑖𝑗<+,0<̃𝑟𝑗𝑖<+,  𝐾𝑖𝑗𝐾(0)=1,𝑗𝑖(0)=1, lim𝜆𝑟𝑖𝑗𝐾𝑖𝑗(𝜆)=+ and lim𝜆̃𝑟𝑗𝑖𝐾𝑗𝑖(𝜆)=+.(H12)There exists positive integer 𝑘0 such that 𝑡𝑘+𝑘0=𝑡𝑘+𝑇 and 𝛾𝑖(𝑘+𝑘0)=𝛾𝑖𝑘,𝛿𝑗(𝑘+𝑘0)=𝛿𝑗𝑘 hold.

We assume that system (3.1) has the following initial conditions:𝑥𝑖(𝑠)=𝜑𝑖(𝑠),𝑦𝑗(𝑠)=𝜙𝑗(𝑠),𝑖=1,2,,𝑛,𝑗=1,2,,𝑚,𝑠0,(3.5) where 𝜓=(𝜑,𝜙)PC((,0],𝑅𝑛+𝑚), 𝜑(𝑠)=(𝜑1(𝑠),𝜑2(𝑠),,𝜑𝑛(𝑠)),𝜙(𝑠)=(𝜙1(𝑠),𝜙2(𝑠),,𝜙𝑚(𝑠)).

Let 𝑍(𝑡,𝜓)=(𝑥(𝑡,𝜓),𝑦(𝑡,𝜓)) denote any solution of the system (3.1) with initial value 𝜓=(𝜑,𝜙)PC, 𝑥(𝑡,𝜓)=(𝑥1(𝑡,𝜓),𝑥2(𝑡,𝜓),,𝑥(𝑡𝑛,𝜓)), 𝑦(𝑡,𝜓)=(𝑦1(𝑡,𝜓),𝑦2(𝑡,𝜓),,𝑦𝑚(𝑡,𝜓)).

Theorem 3.1. Under assumptions (H8)–(H12), there exists a 𝑇-periodic solution which is asymptotically stable, if the following conditions hold.(H13)The following is a nonsingular 𝑀-matrix, and 𝐶𝐴=𝐴𝐶,(3.6) in which 𝛼𝐴=diag1,𝛼2,,𝛼𝑛,𝛽𝐴=diag1,𝛽2,,𝛽𝑚,𝑒𝐶=𝑖𝑗𝑚×𝑛,𝐶=̃𝑒𝑖𝑗𝑛×𝑚,̃𝑒𝑖𝑗=𝑞𝑗𝑖+𝑣𝑗𝑖̃𝜌𝑖𝐿𝑖,𝑒𝑖𝑗=𝑝𝑖𝑗+𝑢𝑖𝑗𝜌𝑗𝐿𝑗.(3.7)(H14)𝑎𝑖((1𝛾𝑖𝑘)𝑠)|1𝛾𝑖𝑘|𝑎𝑖(𝑠),𝑐𝑗((1𝛿𝑗𝑘)𝑠)|1𝛿𝑗𝑘|𝑐𝑗(𝑠),𝑠𝑅,𝑖=1,2,,𝑛,𝑗=1,2,,𝑚.

Proof. Let 𝑥𝑛+𝑗(𝑡)=𝑦𝑗(𝑡),𝑎𝑛+𝑗𝑡,𝑥𝑛+𝑗(𝑡)=𝑐𝑗𝑡,𝑦𝑗(𝑡),𝑏𝑛+𝑗𝑡,𝑥𝑛+𝑗(𝑡)=𝑑𝑗𝑡,𝑦𝑗,𝑝(𝑡)𝑛+𝑗,𝑖(𝑡)=𝑞𝑗𝑖(𝑡),𝑝𝑖,𝑛+𝑗(𝑡)=𝑝𝑖𝑗(𝑡),𝑢𝑛+𝑗,𝑖(𝑡)=𝑣𝑗𝑖(𝑡),𝑢𝑖,𝑛+𝑗(𝑡)=𝑢𝑖𝑗(𝑆𝑡),𝑖𝑥𝑖(𝑡)=𝑔𝑖𝑥𝑖(𝑡),𝑆𝑛+𝑗𝑥𝑛+𝑗(𝑡)=𝑓𝑗𝑥𝑗(𝑡),𝜑𝑛+𝑗(𝑠)=𝜓𝑗𝐼(𝑠),𝑛+𝑗=𝐽𝑗(𝑡),𝑘𝑛+𝑗,𝑖̃𝑘(𝑠)=𝑗𝑖(𝑠),𝑘𝑖𝑗(𝑠)=𝑘𝑖,𝑛+𝑗𝛼(𝑠),𝑛+𝑗(𝑡)=𝛽𝑗𝐿(𝑡),𝑛+𝑗=𝐿𝑗,̃𝜌𝑛+𝑗=𝜌𝑗.(3.8) It follows that system (3.1) can be rewritten as ̇𝑥𝑖(𝑡)=𝑎𝑖𝑥𝑖𝑏(𝑡)𝑖𝑡,𝑥𝑖(𝑡)𝑚𝑗=1𝑝𝑖,𝑛+𝑗(𝑡)𝑆𝑛+𝑗̃𝜌𝑛+𝑗𝑥𝑛+𝑗(𝑡)𝑚𝑗=1𝑢𝑖,𝑛+𝑗(𝑡)0+𝑘𝑖,𝑛+𝑗(𝑠)𝑆𝑛+𝑗̃𝜌𝑛+𝑗𝑥𝑛+𝑗(𝑡𝑠)𝑑𝑠𝐼𝑖(𝑡),𝑡𝑡𝑘,Δ𝑥𝑖𝑡𝑘=𝛾𝑖𝑘𝑥𝑖𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,̇𝑥𝑛+𝑗(𝑡)=𝑎𝑛+𝑗𝑥𝑛+𝑗𝑏(𝑡)𝑛+𝑗𝑡,𝑥𝑛+𝑗(𝑡)𝑛𝑖=1𝑝𝑛+𝑗,𝑖(𝑡)𝑆𝑖̃𝜌𝑖𝑥𝑖(𝑡)𝑛𝑖=1𝑢𝑛+𝑗,𝑖(𝑡)0+𝑘𝑛+𝑗,𝑖(𝑠)𝑆𝑖̃𝜌𝑖𝑥𝑖(𝑡𝑠)𝑑𝑠𝐼𝑛+𝑗(𝑡),𝑡𝑡𝑘,Δ𝑥𝑛+𝑗𝑡𝑘=𝛾𝑛+𝑗,𝑘𝑥𝑛+𝑗𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,(3.9) for 1𝑖𝑛,1𝑗𝑚.
Initial conditions are given by 𝑥𝑖(𝑠)=𝜑𝑖](𝑠),𝑠(,0,𝑖=1,2,,(𝑛+𝑚).(3.10) Thus system (3.9) is a special case of system (2.1) in mathematical form, under conditions (H8)–(H14), we obtain from Lemma 2.5 that system (3.9) has a 𝑇-periodic solution which is globally exponentially stable if 𝑎𝑖((1𝛾𝑖𝑘)𝑠)|1𝛾𝑖𝑘|𝑎𝑖(𝑠) and the following matrix is a 𝑀-matrix, and =𝐴𝐶,(3.11) where 𝛼𝐴=diag1,𝛼2,,𝛼𝑛+𝑚,𝐷=00𝑤1,𝑛+1𝑤1,𝑛+𝑚00𝑤𝑛,𝑛+1𝑤𝑛,𝑛+𝑚𝑤𝑛+1,1𝑤𝑛+1,𝑛𝑤00𝑛+𝑚,1𝑤𝑛+𝑚,𝑛,00(3.12) in which 𝑤𝑖𝑗=(𝑝𝑖𝑗+𝑢𝑖𝑗)̃𝜌𝑗𝐿𝑗.
Then, we know from (3.8) and (3.11) that Theorem 3.1 holds.
If 𝑎𝑖(𝑥𝑖(𝑡))=𝑐𝑗(𝑦𝑗(𝑡))=1, 𝑏𝑖(𝑡,𝑥𝑖(𝑡))=𝑏𝑖(𝑡)𝑥𝑖(𝑡) and 𝑑𝑗(𝑡,𝑦𝑗(𝑡))=𝑑𝑗(𝑡)𝑦𝑗(𝑡), where 𝑏𝑖(𝑡) and 𝑑𝑗(𝑡) are positive continuous 𝑇-periodic functions for 𝑖=1,2,,𝑛,𝑗=1,2,,𝑚. System (3.1) reduces to the following Hopfield-type BAM neural networks model: ̇𝑥𝑖(𝑡)=𝑏𝑖(𝑡)𝑥𝑖(𝑡)+𝑚𝑗=1𝑝𝑖𝑗(𝑡)𝑓𝑗𝜌𝑗𝑦𝑗+(𝑡)𝑚𝑗=1𝑢𝑖𝑗(𝑡)0+𝑘𝑖𝑗(𝑠)𝑓𝑗𝜌𝑗𝑦𝑗(𝑡𝑠)𝑑𝑠+𝐼𝑖(𝑡),𝑡>0,𝑡𝑡𝑘,Δ𝑥𝑖𝑡𝑘=𝛾𝑖𝑘𝑥𝑖𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,̇𝑦𝑗(𝑡)=𝑑𝑗(𝑡)𝑦𝑗(𝑡)+𝑛𝑖=1𝑞𝑗𝑖(𝑡)𝑔𝑖̃𝜌𝑖𝑥𝑖+(𝑡)𝑛𝑖=1𝑣𝑗𝑖(𝑡)0+̃𝑘𝑗𝑖(𝑠)𝑔𝑖̃𝜌𝑖𝑥𝑖(𝑡𝑠)𝑑𝑠+𝐽𝑗(𝑡),𝑡>0,𝑡𝑡𝑘,Δ𝑦𝑗𝑡𝑘=𝛿𝑗𝑘𝑦𝑗𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+.(3.13)

Corollary 3.2. Under assumptions (H9)–(H12), there exists a 𝑇-periodic solution which is globally asymptotically stable, if the following conditions hold. (H13) The following is a nonsingular 𝑀-matrix, and 𝐶𝐴=AC,(3.14) in which 𝑏𝐴=diag1,𝑏2,,𝑏𝑛,𝑑𝐴=diag1,𝑑2,,𝑑𝑚,𝑒𝐶=𝑖𝑗𝑚×𝑛,𝐶=̃𝑒𝑖𝑗𝑛×𝑚,𝑒𝑖𝑗=𝑝𝑖𝑗+𝑢𝑖𝑗𝜌𝑗𝐿𝑗,̃𝑒𝑖𝑗=𝑞𝑗𝑖+𝑣𝑗𝑖̃𝜌𝑖𝐿𝑖.(3.15)(H14)0𝛾𝑖𝑘2,0𝛿j𝑘2 for 𝑖=1,2,,𝑛,𝑗=1,2,,𝑚,𝑘𝑍+.

Proof. As 𝑏𝑖(𝑡,𝑥𝑖(𝑡))=𝑏𝑖(𝑡)𝑥𝑖(𝑡) and 𝑑𝑗(𝑡,𝑦𝑗(𝑡))=𝑑𝑗(𝑡)𝑦𝑗(𝑡), we obtain 𝛼𝑖(𝑡)=𝑏𝑖(𝑡) and 𝛽𝑗(𝑡)=𝑑𝑗(𝑡) in (H2), (H13) implies (H13) holds. Since 𝑎𝑖(𝑥𝑖(𝑡))=𝑐𝑗(𝑦𝑗(𝑡))1, then condition (H14) reduces to (H14). Corollary 3.2 Holds from Theorem 3.1.

Remark 3.3. The conditions for the existence and globally exponential stability of the periodic solution of (3.1) without impulses have nothing to do with inputs of the neuron and amplification functions. The results in [13, 14] have more restrictions than Theorem 3.1 in this paper because conditions for the ones in [13, 14] are relevant to amplification functions and inputs of neurons our results should be better. In addition, Corollary 3.2 is similar to Theorem  2.1 in [15]; our results generalize the results in [15].

Remark 3.4. In view of proof of Theorem 3.1, since CGBAMNNs model is a special case of CGNNs model in form as BAM neural networks model is a special case of Hopfield neural networks model, many results of CGBAMNNs can be directly obtained from the ones of CGNNs, needing no repetitive discussions. Since system (3.1) reduces to autonomous system, Theorem 3.1 still holds, which means that system (3.1) has a equilibrium which is globally asymptotically stable; we know that many results in [18] can be directly obtained from the results in [19].

4. Two Simple Examples

Example 4.1. Consider the following CGNNs model with distributed delays: ̇𝑥1𝑥(𝑡)=21𝑥(𝑡)0.2tanh1(𝑡)sin𝑡0+𝑒𝑠𝑥tanh2,(𝑡𝑠)𝑑𝑠̇𝑥2(𝑥𝑡)=2+cos2(𝑥𝑡)2(𝑡)0.30+𝑒𝑠𝑥tanh1(.𝑡𝑠)𝑑𝑠5(4.1) Obviously, system (4.1) satisfies (H1)–(H5).
Note that =11.20.31,(4.2) it is a nonsingular 𝑀-matrix and system (4.1) also satisfies condition (H6). According to Lemma 2.5, system (4.1) has a 2𝜋-periodic solution which is globally exponentially stable. Figure 1 shows the dynamic behaviors of system (4.1) with initial condition (0.1,0.2).
However, It is easy to check that system (4.1) does not satisfy Theorem 4.3 or  4.4 in [5], so theorems in [5] cannot are used to ascertain the existence and stability of periodic solutions of system (4.1).

Example 4.2. Consider the following CGBAMNNs model with distributed delays and impulses: ̇𝑥1𝑥(𝑡)=2+sin1(𝑡)2𝑥1(𝑡)sin𝑡0+𝑒𝑠||𝑦1||(𝑡𝑠)𝑑𝑠1,𝑡>0,𝑡𝑡𝑘,Δ𝑥1𝑡𝑘=𝛾1𝑘𝑥𝑖𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,̇𝑦1𝑦(𝑡)=3+cos1(𝑡)(3+cos𝑡)𝑦1(𝑡)sin𝑡0+𝑒𝑠||𝑥1||(𝑡𝑠)𝑑𝑠1,𝑡>0,𝑡𝑡𝑘,Δ𝑦1𝑡𝑘=𝛿1𝑘𝑦1𝑡𝑘,𝑡=𝑡𝑘,𝑘𝑍+,(4.3) where 𝑡𝑘=𝜋𝑘,𝑘𝑍+.
Obviously, system (4.3) satisfies (H8)–(H12).

Case 1. 𝛾1𝑘=0,𝛿1𝑘=0. Note that =2112,(4.4) it is a nonsingular 𝑀-matrix and system (4.3) also satisfies condition (H13). According to Theorem 3.1, system (4.3) without impulses has a 2𝜋-periodic solution which is globally exponentially stable. Figure 2 shows the dynamic behaviors of system (4.3) with initial condition (0.1,0.2).
However, it is easy to check that system (4.3) without impulses does not satisfy Theorem  1 in [13] and theorems in [14]; so theorems in [13, 14] cannot be used to ascertain the existence and stability of periodic solutions of system (4.3).

Case 2. 𝛾1𝑘=0.7,𝛿1𝑘=(10.5sin(𝑡𝑘+1)). Note that 𝑎1(𝑠)=2+sin𝑠,𝑐1(𝑠)=3+cos𝑠, and 𝑎1/𝑎1=0.5>|1𝛾1𝑘|=0.3 and |1𝛿1𝑘|<0.5<𝑐1/𝑐1=2/3, which means condition (H14) also holds for system (4.3). Hence, system (4.3) with impulses still has that there exists a 2𝜋-periodic solution which is globally asymptotically stable. Figure 3 shows the dynamic behaviors of system (4.3) with initial condition (0.1,0.2).
This example illustrates the feasibility and effectiveness of the main results obtained in this paper, and it also shows that the conditions for the existence and globally exponential stability of the periodic solutions of CGBAMNNs without impulses have nothing to do with inputs of the neurons and amplification functions. If impulsive perturbations exist, the periodic solutions still exist and they are globally exponentially stable when we give some restrictions on impulsive perturbations.

5. Conclusions

A class of CGBAMNNs with distributed delays and impulses are investigated by using suitable Lyapunov functional, the properties of 𝑀-matrix, and some suitable mathematical transformation in this paper. Sufficient conditions to guarantee the uniqueness and global exponential stability of the periodic solutions of such networks are established without assuming the boundedness of the activation functions. Lemma 2.5 improves the results in [5], and Theorem 3.1 improves the results in [13, 14] and generalize the results in [15]. In addition, we point that CGBAMNNs model is a special case of CGNNs model; many results of CGBAMNNs can be directly obtained from the ones of CGNNs, needing no repetitive discussions. Our results are new, and two examples have been provided to demonstrate the effectiveness of our results.

Appendix

The source program (MATLAB 7.0) of Figure 1 is given as follows [14].clearT=70;N=7000;h=T/N;m=40/h;for i=1:mU(:,i)=[0.1; 0.2];endfor i=(m+1):(N+m)r(i)=i*h-40;x(i)=r(i);I=2;J=2+cos(U(2,i-1));A=[-I,0; 0,-J];B=[0,sin(x(i))*I; 0.3*J,0];U(:,i)=h*A*[(U(1,i-1)-0.2*tanh(U(1,i-1))); U(2,i-1)]+U(:,i-1);P(:,1)=[0; 0];for k=1:mP(:,1)=P(:,1)+h*exp(-(40-(k-1)*h))*[(tanh(U(1,i-m+k-1)));(tanh(U(2,i-m+k-1)))];endU(:,i)=U(:,i)+B*h*[(P(1,1));(P(2,1))]+h*[0; 5*J];endy=U(1,:);z=U(2,:);hold onplot(r,y,’:’)hold onplot(r,z)hold onplot3(r,y,z)

The source program (MATLAB 7.0) of Figures 2 and 3 is given as follows [14].clearT=70;N=7000;h=T/N;m=40/h;for i=1:mU(:,i)=[0.1;0.2];endfor i=(m+1):(N+m)r(i)=i*h-40;x(i)=r(i);I=2+sin(U(1,i-1));J=3+cos(U(2,i-1));A=[-I,0;0,-J];B=[0,I*sin(x(i)); J*sin(x(i)),0];U(:,i)=h*A*[2*U(1,i-1);(3+cos(x(i)))*U(2,i-1)]+U(:,i-1);P(:,1)=[0;0];for k=1:mP(:,1)=P(:,1)+h*exp(-(40-(k-1)*h))*[(abs(U(1,i-m+k-1)));(abs(U(2,i-m+k-1)))];endU(:,i)=U(:,i)+B*h*[(P(1,1));(P(2,1))]+[I; J]*h;if mod(i-m,314)==0U(:,i)=[0.3,0; 0,1/2*(sin(x(i)+1))]*U(:,i);endendy=U(1,:);z=U(2,:)hold onplot(r,y,’:’)hold onplot(r,z)hold onplot3(r,y,z)

Acknowledgment

The authors would like to thank the editor and the reviewers for their valuable suggestions and comments which greatly improved the original paper. Projects supported by the National Natural Science Foundation of China (no. 11071254).