Loading [MathJax]/jax/element/mml/optable/GeneralPunctuation.js
  数学杂志  2023, Vol. 43 Issue (2): 135-150   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
HONG Yu
SONG Xing-chuan
TIAN Yan-ling
ASYMPTOTIC BEHAVIOR OF NONLINEAR DELAYED DISCRETE SYSTEMS
HONG Yu, SONG Xing-chuan, TIAN Yan-ling    
School of Mathematics, South China Normal University, Guangdong 510631, China
Abstract: In this paper, we consider a two-dimensional vector nonlinear difference system with time delay. By using stability theory and optimal control theory, sufficient conditions for the global asymptotic stability of a unique positive equilibrium are obtained. For the associated optimal control problem of maximizing the consumption functional, the existence of optimal solutions is established as well as their stability. Finally, numerical simulation is carried out to illustrate the validity of our results. Our results generalize the corresponding results on a one-dimensional vector nonlinear difference system.
Keywords: delayed difference system     global asymptotic stability     optimal control    
一类非线性时滞离散系统的渐近行为
洪宇, 宋兴传, 田艳玲    
华南师范大学数学科学学院, 广东 广州 510631
摘要:本文研究了一类二维时滞非线性差分系统.利用稳定性理论和最优控制理论等方法,本文首先获得了保证正平衡点的存在唯一以及全局渐近稳定的充分条件;然后对消费函数的最大化的最优控制问题进行讨论,获得了最优解的存在性与稳定性条件;最后,通过数值模拟验证了结果的有效性.本文推广了一维时滞非线性差分系统中给出的相关结论.
关键词时滞差分系统    全局渐近稳定    最优控制方法    
1 Introduction

In many fields, including mathematical biology, medicine, homeland security and manufacturing industry, economics etc, much of the data are discrete rather than continuous. As well known, difference equation is an effective mathematical tool to study the variation law of discrete variables. A lot of papers and monographs have been written on this subject, see [1-4] and further references therein. Among them, delayed nonlinear equations were investigated by a large number of researchers, see in [1, 5-7]. For such equations, the asymptotic behavior of the solution to difference equations is an important topic arising from a large number of practical problems as well as the optimal control problem, see [8-11] and textbooks [2, 3, 12].

To describe an economic problem, Ivanov considered the following equation in [13]:

xn+1=axn+unf(xnK) (1.1)

where 0<a<1, un[0,1] is the control coefficient. xn is the economic output at any time n, the output of the next time interval n+1 is composed of two parts, one is axn, a certain proportion of the output at time n, the other is the output f(xnK) which results from the production output K steps back, caused by the delay factors due to some economic results. They dealt with two problems in [13]. One is the global asymptotic stability of its unique positive equilibrium for equation (1.1) without control, i.e un1 for any nN+. The other is an optimal control problem, which is associated with an important economic problem, maximization of consumption functional. They found a control sequence {un}n=1 to control the consumption achieve to maximal value, where the value of un is in [0,1] but not identically equal to 1.

But one variable or one discrete equation can not describe more complex situation sometimes. In our current paper, we consider a more complex situation where a firm has two factories. The same economic problem leads to the following difference equations with delay

{xn+1=a1xn+un(f1(xnK)+g1(ynK)),yn+1=a2yn+vn(f2(xnK)+g2(ynK)), (1.2)

where 0<ai<1, functions fi,gi:R+R+ are continuous on R+ for i=1,2 and K1 is the integer delay. xn,yn are the economic output of the two factories respectively, and can be measured at discrete time intervals (daily, monthly, or yearly). For example, the output of one factory at any time n denoted by xn, the output of the next time interval n+1 is composed of two parts, one is a certain proportion of the output of the company at time n, a1xn, and the other is the output f1(xnK)+g1(ynK). The economic significance of the two parts are similar to [13]. Obviously, our model is an generalization of equation (1.1). Moreover, it is interesting to discuss how to assign resources to the two factories to control the consumption achieve to maximal value. But there are many difficulties due to this general setting. First, some more complicated conditions will be proposed to guarantee the existence and the uniqueness of the positive equilibrium by using inverse function method. Second, we need to develop additional techniques to establish the stability of the positive equilibrium. Finally, the most difficult thing is to study the control problem. Not only are there four functions fi, gi(i=1,2) in the system, but also the control sequence {un, vn}n=1 are more complicated than those in [13]. Therefore, we will propose suitable assumptions on those functions and develop additional techniques.

Our paper is organized as follows. We establish the existence and the uniqueness of the positive equilibrium of the system considered in the current paper with un=vn1 and study the global asymptotic stability of the equilibrium in section 2. Section 3 is concerned with the associated optimal control problem. The properties of the solution to the control equation are obtained. For any one of the optimal equilibria, an associated control sequence is found to make the solution converge to it. Finally, numerical simulation is carried out in section 4 to illustrate the validity of our assumptions and results.

2 Preliminaries and Global Asymptotic Stability

In this section, we put forward two hypotheses, introduce basic definitions and state stability results concerning the difference system (1.2) with un=vn=1, that is the following system:

{xn+1=a1xn+(f1(xnK)+g1(ynK)),yn+1=a2yn+(f2(xnK)+g2(ynK)). (2.1)
2.1 Positive Equilibrium Analysis of Equation (2.1)

It is difficult to guarantee the existence and the uniqueness of the equilibrium of the system (2.1) since there are four functions fi, gi(i=1,2) in the right of the system. To overcome this difficulty, we introduce inverse function method. Some functions are given first. Define

H(1)1(r):=(1a1)rf1(r), H(2)1(r):=(1a2)rg2(r).

Their inverse functions of H(1)1(r) and H(2)1(r) are denoted by (H(1)1)1(r), (H(2)1)1(r) respectively. Define

H(1)2(r):=(1a2)rg2(r)f2((H(1)1)1(g1(r))),H(2)2(r):=(1a1)rf1(r)g1((H(2)1)1(f2(r))).

Two assumptions are given as follows.

(H1) Assume that functions H(1)1, H(2)1 have inverse functions for r>0 and (H(1)1)1(r)>0, (H(2)1)1(r)>0 if r>0.

(H2) Assume that there is Y>0 such that the function H(1)2 satisfies

H(1)2(r)<0, if  0<r<Y;  H(1)2(r)>0, if  r>Y. (2.2)

Let X=(H(1)1)1(g1(Y)). Assume that the function H(2)2 satisfies

H(2)2(r)<0, if  0<r<X;  H(2)2(r)>0, if  r>X. (2.3)

Assumptions (H1), (H2) and simple calculations lead to the following lemma.

Lemma 2.1 Assume that assumptions (H1) and (H2) hold. Then difference system (2.1) has a unique positive equilibrium (X, Y). Moreover,

(X,Y)=(f1(X)+g1(Y)1a1, f2(X)+g2(Y)1a2).
2.2 Asymptotic Behavior of the Solution to Equation (2.1)

Before our main results, we introduce some definitions and notations.

Define F=(F1,F2)T, G=(G1,G2)T, where Fi=11aifi, Gi=11aigi, i=1,2. Define T=(T1, T2)T, where Ti(x,y)=Fi(x)+Gi(y), i=1,2. It is obvious that (X,Y) is a fixed point of the map T. The fixed point of map T is called attracting if there exists its open (with respect to R+×R+) neighborhood U such that T(U)U and limnTn(x,y)=(X,Y) for each (x,y)U, where Tn:=TTn1=T(Tn1). The largest interval UR+×R+ with the above property is called the domain of immediate attraction of the fixed point (X,Y). An interval IR+×R+ is said to be invariant under T if (T1(x,y),T2(x,y))I for all (x,y)I.

Given an initial string P0={PK, PK+1, ...,P0}, then system (2.1) has a solution (xn,yn) for n>0 by consecutive iterations. Define Pi:=(xi,yi), i{K,K+1,...,0}N+. The segment {P(j1)K+j, , PjK+j} is called the jth string Pj. Give a set SR+×R+, we say that PjS if P(j1)K+j+iS for all i{0,1,,K}. Clearly, xn0, yn0 for all n1 if the initial data for P0 are all in R+×R+.

Difference system (2.1) is equivalent to the following one

{μ1xn=xn+1+F1(xnK)+G1(ynK),μ2yn=yn+1+F2(xnK)+G2(ynK), (2.4)

where μi=ai1ai, xn:=xn+1xn, yn:=yn+1yn. Let A:=(μ100μ2) and Pn=(xn,yn)T, then the vector form of (2.4) is

APn=Pn+1+F(xnK)+G(ynK)=Pn+1+T(xnK,ynK).

In this subsection we obtain the global asymptotic stability of the unique positive equilibrium.

Theorem 2.1 Suppose that (X,Y) is an attracting fixed point of map T and set J is the associated domain of immediate attraction. Then for every initial string P0J, the corresponding solution Pn=Pn(P0) has the property

limn(xn,yn)=(X,Y). (2.5)

Proof By the definition of attracting fixed point of map T, one has

T(X, Y)=(X, Y), limnTn(x,y)=(X,Y), (x,y)J(X,Y).

Let an initial string P0={(xK,yK),(xK+1,yK+1),,(x0,y0)} be given such that P0J, then one can find a closed bounded interval I0:=[α1, β1]×[α2, β2]J such that P0I0 and T(I0)I0. Since J is the domain of immediate attraction of (X,Y), one also has

I0T(I0)T2(I0)Ti(I0) and i0Ti(I0)=(X,Y), (2.6)

where T(I0):=[γ1, δ1]×[γ2, δ2].

Let (xn,yn) be the solution of (2.4) with associated initial data P0. Then the following two cases will occur.

Case I. There is N0 such that (xN,yN)T(I0). Then one can show that (xn,yn)T(I0) for all nN. Suppose not, let ˜N>N be the first time when the solution leaves the interval T(I0) for n>N. Without loss of generality, we may assume that (x˜N,y˜N)[α1,γ1]×[α2,β2] and (xn,yn)T(I0)I0 for all N<n<˜N. Then x˜N1=x˜Nx˜N1<0. On the other hand, since T1(x˜NK,y˜NK)T1(I0)=[γ1,δ1], then system (2.4) shows that x˜N1=1μ1[x˜N+F1(x˜NK)+G1(y˜NK)]>0, a contradiction. Then (2.5) is valid from (2.6).

Case II. (xn,yn)T(I0) for all n0. Then we will show that limn(xn,yn)=(X, Y). Let A:=([α1,γ1]×[α2,δ2])([γ1,δ1]×[α2,γ2]), then AI0T(I0). We claim that (xn,yn)A for all n>0 if (x0,y0)A, and (xn,yn)[γ1,δ1]×[γ2,δ2] for all n>0. Indeed, if on the contrary, P1=P1(P0)A, then x1>δ1 or y1>δ2. Consider a modified initial string ˜P0:={(xK,yK),,(x1,y1),(δ1,δ2)}, in view of (δ1,δ2)T(I0), one has Pn(˜P0)T(I0) for all n>0. So that P1(˜P0)=(˜x1,˜y1)(δ1,δ2). From system (2.4) we can see that

x1(P0)=a1x0+f1(xK)+g1(yK)<a1δ1+f1(xK)+g1(yK)=˜x1(˜P0),
y1(P0)=a2y0+f2(xK)+g2(yK)<a2δ2+f2(xK)+g2(yK)=˜y1(˜P0),

a contradiction with P1(P0)A. Therefore, we obtain the conclusion that (xn,yn)A for all n>0. That means the solution (xn,yn) is bounded from above. Note that Fi, Gi(i=1,2) are continuous over I0, then limn(xn,yn)=(¯x,¯y)A. Since if α1¯x<γ1, then 1μ1[¯x+F1(¯x)+G1(¯y)]>0 and if α2¯y<γ2, then 1μ2[¯y+F2(¯x)+G2(¯y)]>0, then ¯x=γ1, ¯y=γ2. On the other hand, since limn(xn,yn)=(0,0), system (2.4) implies that ¯x=F1(¯x)+G1(¯y), ¯y=F2(¯x)+G2(¯y), which implies (¯x, ¯y)=(X, Y). Thus, from (2.4), one can conclude that (2.5) is valid. The theorem is proved.

The following corollary gives a sufficient condition to guarantee that the fixed point (X,Y) is an attracting fixed point of map T.

Corollary 2.1 Suppose that

max(s,r)[0,X]×[0,Y](F1(r)+G1(s))X, max(s,r)[0,X]×[0,Y](F2(r)+G2(s))Y,min(s,r)[X,+)×[Y,+)(F1(r)+G1(s))X, min(s,r)[X,+)×[Y,+)(F2(r)+G2(s))Y. (2.7)

If there are increasing functions ω±i(i=1,2) with

ω1(r)>r for 0<r<X, ω+1(r)<r for r>X,ω2(s)>s for 0<s<Y, ω+2(s)<s for s>Y, (2.8)

such that

F1(r)+mins[0,Y]G1(s)ω1(r), F1(r)+maxsYG1(s)ω+1(r),minr[0,X]F2(r)+G2(s)ω2(s), maxrXF2(r)+G2(s)ω+2(s), (2.9)

then

limn(xn,yn)=(X,Y)

provided P0I0:=([0,X]×[0,Y])([X,+)×[Y,+)).

Proof It is easy to see that (2.7) implies that I0 is the invariant domain of operator T. Moreover, (2.8) and (2.9) imply that i0Ti(I0)=(X,Y). Therefore, (X,Y) is the the attracting fixed point of map T and the associated domain of immediate attraction is I0. Hence, one can obtain (2.5) when the initial string P0I0. The corollary is proved.

3 Convergence of Solutions in Optimal Control Problems

Consider the following difference system

{xn+1=a1xn+un(f1(xnK)+g1(ynK)),yn+1=a2yn+vn(f2(xnK)+g2(ynK)), n=0,1,2,, (3.1)

where the parameters ai(i=1,2) and the functions fi,gi(i=1,2) are as same as those of system (2.1), and un,vn are control sequences with values in [0,1] which are motivated by practical applications.

Obviously, the general setting leads to additional difficulties and much more technicality. How to handle with the four functions and the six solution subsequences in (3.1)? To this end, the functions fi,gi,(i=1,2) are assumed to satisfy the following conditions.

(H3) Let 1a=min{(1a1),(1a2)}. For a given bounded region [0,M]×[0,M], there exist positive constants hi(0,1], i=1,2 with (1a1)h1+(1a2)h21a such that for (x,y)[0,M]×[0,M],

Σ2i=1(fi(x)+gi(y))Σ2i=1(fi(h1(x+y))+gi(h2(x+y))). (3.2)

(H4) Let ˜F(r)=f1(r)+f2(r)(1a1)r, ˜G(s)=g1(s)+g2(s)(1a2)s. Define A={ˉs|˜G(ˉs)=max0sY˜G(s)}, B={ˉr|˜F(ˉr)=max0rX˜F(r)}. Then

(1) ˜F(r)<0 for r>X and ˜G(s)<0 for s>Y;

(2) f1(r)+g1(s)(1a1)r if sA for 0<r<X and f2(r)+g2(s)(1a2)s if rB for 0<s<Y;

(3) there is ˜M>0 such that f1(r)+f2(r)<(1b)r, g1(r)+g2(r)<(1b)r for r>˜M, where b>a.

Remark The condition (H3) plays an important role in our next discussion and it is reasonable. For example, it is easy to prove that (H3) is satisfied when fi(x)=x+x, gi(y)=y+y, i=1,2 when a1=a2. Let a1a2, choose λ<12 such that λ(1a1)+λ(1a2)1a. If f1(x)+f2(x)=λh(x), g1(y)+g2(y)=(1λ)h(y), where h is convex and decreasing, then one has

Σ2i=1(fi(x)+gi(y))λh(λx+(1λ)y)+(1λ)h(λx+(1λ)y)λh(λ(x+y))+(1λ)h(λ(x+y))=Σ2i=1(fi(λ(x+y))+gi(λ(x+y))).

Thus, (H3) is satisfied.

Next, we give the notations about the consumption. Let P0 be any fixed initial string. For arbitrary control U=(un,vn)n=0 system (3.1) defines a unique solution P=P(P0,U)=(Pn)n=1=(xn,yn)n=1, which is found by consecutive iterations for all n>0. Similar to [13], the consumption is denoted by [(1un)(f1(xnK)+g1(ynK))+(1vn)(f2(xnK)+g2(ynK))]. The following consumption functional

C(P0,U)=lim infn[(1un)(f1(xnK)+g1(ynK))+(1vn)(f2(xnK)+g2(ynK))] (3.3)

and its maximization problem over the solutions to system (3.1)

Maximize:C(P0,U), subject to (3.1) (3.4)

are used to denote the minimal level of consumption for sufficiently large time periods. A positive constant c is used to describe the maximum steady consumption that could be achieved in problem (3.4) and defined as

c=max{L1(x,y)+L2(x,y), (x,y)R+×R+},

where L1(x,y)=f1(x)+g1(y)(1a1)x, L2(x,y)=f2(x)+g2(y)(1a2)y, (x,y)R+×R+. And set τ, τ are defined as follows:

τ={(x,y)[0,X]×[0,Y]:L1(x,y)+L2(x,y)=c, there are  ¯u1,¯v1[0,1] such thatx=a1x+¯u1(f1(x)+g1(y)), y=a2y+¯v1(f2(x)+g2(y))} (3.5)

and

τ={¯p:there are  ¯x, ¯y such that ¯x+¯y=¯p and (¯x,¯y)τ},

where set τ contains the steady state guaranteing this consumption c, the element of τ is the sum of each component of the element in set τ. First, set τ is non-empty evidently. Actually, by assumption (H4), there is (x,y)[0,X]×[0,Y] such that L1(x,y)+L2(x,y)=c, and (x,y)τ since

ˉu1=(1a1)xf1(x)+g1(y)1, ˉv1=(1a2)xf2(x)+g2(y)1.

It follows that set τ is non-empty. Second, τ is a closed set and it may contain more than one point. One would like to find a control sequence (un,vn),n0 that maximizes the minimal level of consumption for sufficiently large time periods for any given initial string P0, that means the solution to (3.1) will be convergent to the optimal steady state in set τ under the control sequence (un, vn).

In what follows, we present three theorems to describe the properties of the solutions to equation (3.1) as well as the relation between C(P0,U) and c over the solutions to (3.1) for any given initial string P0. Theorem 3.1 shows C(P0,U)c for all P0 and U; Theorem 3.2 shows that there exists an optimal control UP0 to problem (3.4) such that functional (3.3) achieves its maximum possible value; that is, C(P0,U)=c; Theorem 3.3 shows how to construct a control sequence to make the solution to (3.1) converge to an optimal control equilibrium in set τ or set τ.

Theorem 3.1 The functional (3.3) is bounded above over the solutions to (3.1), that is, for all P0 and U, there exists the inequality

C(P0,U)c. (3.6)

Proof Step 1. We claim that the solution P to (3.1) is bounded; that is, there exists ˉM such that

lim supn(xn+yn)ˉM, P.

Consider solution P=P(P0,U)=(xn,yn)n=0 corresponding to the initial string P0 and control U=(un,vn)n=0. Let Sn=xn+yn, r=max{X, Y}, L=max{f1(x)+f2(x), x[0,˜M]}+max{g1(y)+g2(y), y[0,˜M]}, ˉM=max{Sk,...,S1,S0; 11aL, ˉM1}, where ˉM1Lba. Let m be the value such that SkˉM for km and Sm+1>ˉM.

(1) xmK<˜M, ymK<˜M.

0<Sm+1Sm(1a1)xm+f1(xmK)+g1(ymK)(1a2)ym+f2(xmK)+g2(ymK)(1a)Sm+L(1a)Sm+(1a)ˉM(1a)Sm+(1a)Sm+1.

(2) xmK˜M, ˜MymK<ˉM.

0<Sm+1Sm(1a1)xm+f1(xmK)+g1(ymK)(1a2)ym+f2(xmK)+g2(ymK)(1a)Sm+L+(1b)ˉM(1a)Sm+(1a)ˉM(1a)Sm+(1a)Sm+1.

(3) xmK>˜M, ymK>˜M.

0<Sm+1Sm(1a1)xm+f1(xmK)+g1(ymK)(1a2)ym+f2(xmK)+g2(ymK)(1a)Sm+(1a)SmK(1a)Sm+(1a)Sm+1.

Thus, 0<Sm+1Sm<(1a)(Sm+1Sm), a contradiction is obtained since 0<a<1.

Step 2. We claim that for every solution P=P(P0,U) to (3.1) the inequality (3.6) holds.

Recall that C(P0,U)=lim infncn, where

cn=(1un)(f1(xnK)+g1(ynK))+(1vn)(f2(xnK)+g2(ynK)). (3.7)

From (3.1) and (3.7) it follows that

cn=f1(xnK)+g1(ynK)xn+1+a1xn+f2(xnK)+g2(ynK)yn+1+a2yn. (3.8)

Denote p1:=lim supnxn<, then there is km such that xkm+1p1; let p2:=lim supkmykm+1<, then there is {km}{km} such that ykm+1p2 as km. Without loss of generality, we have

limkmxkmK=x, limkmxkm=x;limkmykmK=y, limkmykm=y.

From (3.8) we have

limkmckm=f1(x

By the definition of c^* we have

f_1(x'')+g_1(y'')-(1-a)x''+f_2(x'')+g_2(y'')-(1-a)y''\leq c^*.

Therefore, {C}(\mathbf{P}_0, \mathbf{U})=\liminf_{n\rightarrow \infty}c_n\leq \lim_{n\rightarrow \infty}c_{k^{\prime}_m} \leq c^*. The theorem is proved.

Theorem 3.2 For any given initial string \mathbf{P}_0 , there exists an optimal control \mathbf{U}_{\mathbf{P}_0} to problem (3.4) such that functional (3.3) achieves its maximum possible value, that is,

C(\mathbf{P}_0 , \mathbf{U})=c^*.

In addition, if the set of optimal equilibrium points defined by (3.5) has an empty interior int \tau^{\prime}=\emptyset , then

\begin{equation} \lim\limits_{n\rightarrow \infty}(x_{n}+y_{n})\in\tau' \end{equation} (3.9)

for all optimal solution \mathbf{P} .

Proof We first claim that if \mathbf{P}=\mathbf{P}(\mathbf{P}_0, \mathbf{U}) =(x_n, y_n)_{n=0}^\infty is a solution to (3.1) such that {C}(\mathbf{P}_0, \mathbf{U})= c^* , then

\begin{align*} (1-a_1)x_{n-K}+(1-a_2)y_{n-K}+a_1x_n+a_2y_n-(x_{n+1}+y_{n+1})\geq -\xi_n, \ \forall n, \end{align*}

where \xi_n>0 and \xi_n\rightarrow 0 as n\rightarrow \infty . Moreover,

\begin{equation} \limsup\limits_{n\rightarrow \infty}(x_{n}+y_{n})\in\tau'. \end{equation} (3.10)

Since \liminf_{n\rightarrow \infty}c_{n}=c^* , then there is a sequence of positive numbers \xi_n\rightarrow 0 such that c_n\geq c^*-\xi_n . Then from (3.8)

f_1(x_{n-K})+g_1(y_{n-K})-x_{n+1}+a_1x_n+f_2(x_{n-K})+g_2(y_{n-K})-y_{n+1}+a_2y_n\geq c^*-\xi_n, \ \forall n.

Then

\begin{align*} &f_1(h_1(x_{n-K}+y_{n-K}))+g_1(h_2(x_{n-K}+y_{n-K}))+f_2(h_1(x_{n-K}+y_{n-K}))\\ &+g_2(h_2(x_{n-K}+y_{n-K})) -(x_{n+1}+y_{n+1})+a_1x_n+a_2y_n\geq c^*-\xi_n, \ \forall n. \end{align*}

On the other hand, by definition of c^* it follows that

\begin{align*} &f_1(h_1(x_{n-K}+y_{n-K}))+g_1(h_2(x_{n-K}+y_{n-K}))+f_2(h_1(x_{n-K}+y_{n-K}))\\ &+g_2(h_2(x_{n-K}+y_{n-K}))-(1-a_1)x_{n-K}-(1-a_2)y_{n-K}\\ \leq& f_1(h_1(x_{n-K}+y_{n-K}))+g_1(h_2(x_{n-K}+y_{n-K}))+f_2(h_1(x_{n-K}+y_{n-K}))\\ &+g_2(h_2(x_{n-K}+y_{n-K})) -(1-a_1)h_1(x_{n-K}+y_{n-K})-(1-a_2)h_2(x_{n-K}+y_{n-K})\leq c^*. \end{align*}

Therefore, from the last two inequalities we have

(1-a_1)x_{n-K}+(1-a_2)y_{n-K}+a_1x_n+a_2y_n-(x_{n+1}+y_{n+1})\geq -\xi_n, \ \forall\ n.

Now we prove (3.10). Denote p:=\limsup_{n\rightarrow \infty}(x_n+y_n)<\infty . By the definition of p there is a subsequence k_m\rightarrow \infty satisfying the following conditions:

x_{k_m+1}+y_{k_m+1}\rightarrow p;\ x_{k_m}+y_{k_m}\rightarrow p'\leq p;\ x_{k_m-K}+y_{k_m-K}\rightarrow p''\leq p.

From (3.2) and (3.8), together with the fact a<1 and p''\leq p , we have

\begin{equation} \begin{aligned} c^*=&\liminf\limits_{n\rightarrow \infty}c_n \leq\lim\limits_{k_m\rightarrow \infty}c_{k_m}\\ &=\lim\limits_{k_m\rightarrow \infty}(f_1(x_{k_m-K})+g_1(y_{k_m-K})-x_{k_m+1}+a_1x_{k_m}+f_2(x_{k_m-K})+g_2(y_{k_m-K}) -y_{k_m+1}+a_2y_{k_m})\\ &\leq \lim\limits_{k_m\rightarrow \infty}[f_1(h_1(x_{k_m-K}+y_{k_m-K}))+g_1(h_2(x_{k_m-K}+y_{k_m-K}))+f_2(h_1(x_{k_m-K}+y_{k_m-K}))\\ &\ \ \ \ \ \ \ \ \ \ \ \ +g_2(h_2(x_{k_m-K}+y_{k_m-K}))-(x_{k_m+1}+y_{k_m+1})+a_1x_{k_m}+a_2y_{k_m}]\\ &\leq f_1(h_1p'')+g_1(h_2p'')+f_2(h_1p'')+g_2(h_2p'')-(1-a)p\\ &\leq f_1(h_1p'')+g_1(h_2p'')+f_2(h_1p'')+g_2(h_2p'')-(1-a_1)h_1p-(1-a_2)h_2p\\ &\leq f_1(h_1p'')+g_1(h_2p'')-(1-a_1)h_1p''+f_2(h_1p'')+g_2(h_2p'')-h_2(1-a_2)p''. \end{aligned} \end{equation} (3.11)

This means that (h_1p'', h_2p'')\in\tau , from (3.6). Therefore,

c^*= f_1(h_1p'')+g_1(h_2p'')-(1-a_1)h_1p''+f_2(h_1p'')+g_2(h_2p'')-(1-a_2)h_2p''.

Now, if p'' < p , then going back to (3.11) we obtain a contradiction in the form

\begin{equation} \begin{aligned} \begin{split} c^*&\leq f_1(h_1p'')+g_1(h_2p'')+f_2(h_1p'')+g_2(h_2p'')-(1-a_1)h_1p-(1-a_2)h_2p\\ &< f_1(h_1p'')+g_1(h_2p'')+f_2(h_1p'')+g_2(h_2p'')-(1-a_1)h_1p^{\prime\prime}-(1-a_2)h_2p^{\prime\prime}\\ &\leq f_1(h_1p'')+g_1(h_2p'')-(1-a_1)h_1p''+f_2(h_1p'')+g_2(h_2p'')-(1-a_2)h_2p'' =c^*. \end{split} \end{aligned} \end{equation} (3.12)

Therefore, p=p^{\prime\prime}=\limsup_{n\rightarrow \infty}(x_n+y_n)\in\tau' , which implies that (3.10) is true.

Finally, we prove (3.9), where is similar to [13]. Denote q:=\liminf_{n\rightarrow \infty}(x_n+y_n)<\infty, \ p:=\limsup_{n\rightarrow \infty}(x_n+y_n)<\infty and p\in\tau' . Take any positive number \eta\in(0, \overline{\eta}] , where \overline{\eta} is defined as \overline{\eta}:=\min\{p-\overline{q}, \ \dfrac{p-q}{2}\}>0. Here \overline{q}:=a^{K}q+(1-a^{K})p and K is the delay in system (3.1). Denote \widetilde{p}:=p-\dfrac{1}{2}\eta\ <p . Similar to the proof of Proposition 3.4 in [13], we can construct a sequence k_m\rightarrow \infty satisfying

\begin{equation} x_{k_m+1}+y_{k_m+1}\rightarrow p \geq\widetilde{p};\ x_{k_m}+y_{k_m}\rightarrow p'\leq \widetilde{p} ;\ x_{k_m-K}+y_{k_m-K}\rightarrow p''\leq \widetilde{p}. \end{equation} (3.13)

Therefore, similar to the calculation in (3.11), there is

\begin{align*} c^*\leq f_1(h_1p'')+g_1(h_2p'')-(1-a_1)h_1p^{\prime\prime}+f_2(h_1p'') +g_2(h_2p'')-(1-a_2)h_2p^{\prime\prime}. \end{align*}

In view of the definition of c^* and (3.13), similar to (3.12), we obtain p''=\widetilde{p}=p-\dfrac{1}{2}\eta\in\tau', \ \forall\eta\in(0, \overline{\eta}] follows. This contradicts the assumption that \tau has an empty interior and (3.9) is proved.

Theorem 3.3 Suppose that functions f_1, \ g_2, \ H_1^{(i)}(i=1, 2) are increasing, while functions f_2 and g_1 are monotone and possess the same monotonicity. For any given initial string \mathbf{P}_0 and any optimal equilibrium (\tilde{x}, \tilde{y})\in\tau, \ \tilde{p}=\tilde{x}+\tilde{y}\in\tau' , there is a control \mathbf{U} such that the corresponding solution \mathbf{P}=\mathbf{P}(\mathbf{P}_0, \mathbf{U}) =(x_n, y_n)_{n=0}^\infty to (3.1) converges to that equilibrium (\tilde{x}, \tilde{y}) ; that is

\lim\limits_{n\rightarrow \infty}(x_{n}, y_n)=(\tilde{x}, \tilde{y})\in \tau.

Proof Take any \tilde{p}\in\tau^{\prime} and (\tilde{x}, \tilde{y})\in\tau . From (2.2), (2.3) and (3.5), at least one of the two statements \tilde{x}<X^* and \tilde{y}<Y^* holds.

If at least one of \tilde{x} and \tilde{y} is vanish, then one can use the method as the same as that in [13] to find the control sequence, we omit it here.

Next, we consider the case \tilde{x}>0, \ \tilde{y}>0 . Without loss of generality, we assume that \tilde{y}<Y^* , then from assumption (H2), there holds

\begin{equation} \begin{aligned} &f_2((H_1^{(1)})^{-1}(g_1(y)))+g_2(y)>(1-a_2)y>0, \ \forall y\in(0, \tilde{y}]. \end{aligned} \end{equation} (3.14)

For any given any initial string \mathbf{P}_0 , since \tilde{x}>0, \ \tilde{y}>0 , then there is a number n_1 such that

x_{n_1-j}+y_{n_1-j}\in(0, \widetilde{p}], \ \ \forall j=0, 1, \cdots, K.

For n\geq n_1 we define the sequences \tilde{u}_n, \ \tilde{v}_n as follows:

\tilde{u}_n=\min\{1, \dfrac{\tilde{x}-a_1x_n}{f_1(x_{n-K})+g_1(y_{n-K})}\}, \ \tilde{v}_n=\min\{1, \dfrac{\tilde{y}-a_2y_n}{f_2(x_{n-K})+g_2(y_{n-K})}\}, n\geq n_1.

Clearly, as long as \tilde{x}>a_1x_n, \tilde{y}>a_2y_n , the values of \tilde{u}_n, \tilde{v}_n stay in the interval [0, 1] ; that is, \tilde{u}_n, \tilde{v}_n can be used as two control parameters. Consider a sequence (x_n, y_n) defined by

\begin{align} \begin{split} x_{n+1}=a_1x_n+\tilde{u}_n(f_1(x_{n-K})+g_1(y_{n-K})), \\ y_{n+1}=a_2y_n+\tilde{v}_n(f_2(x_{n-K})+g_2(y_{n-K})), \end{split} \end{align} (3.15)

then

\begin{align*} x_{n+1}+y_{n+1}=a_1x_n+\tilde{u}_n(f_1(x_{n-K})+g_1(y_{n-K}))+a_2y_n+\tilde{v}_n(f_2(x_{n-K})+g_2(y_{n-K})). \end{align*}

First we show that this sequence (x_y, y_n) is a solution to (3.1); that is, \tilde{u}_n, \tilde{v}_n\in [0, 1] for all n\geq n_1. For n=n_1 , if \tilde{x}-a_1x_{n_1}\leq f_1(x_{n_1-K})+g_1(y_{n_1-K}) , then \tilde{u}_{n_1}= \dfrac{\tilde{x}-a_1x_{n_1}}{f_1(x_{n_1-K})+g_1(y_{n_1-K})} , and x_{n_1+1}=\tilde{x} ; else if \tilde{x}-a_1x_{n_1}>f_1(x_{n_1-K})+g_1(y_{n_1-K}) , then \tilde{u}_{n_1}=1 and x_{n_1+1}<\tilde{x} , therefore x_{n_1+1}\leq\tilde{x} . Similarly, y_{n_1+1}\leq\tilde{y} . Therefore,

\begin{aligned} x_{n_1+1}\leq \tilde{x}, \ y_{n_1+1}\leq \tilde{y}, \ x_{n_1+1}+y_{n_1+1}\leq \tilde{p}. \end{aligned}

Moreover, one has

\begin{aligned} x_{n_1+1}+y_{n_1+1}&=a_1x_{n_1}+f_1(x_{n_1-K})+g_1(y_{n_1-K})+a_2y_{n_1}+f_2(x_{n_1-K})+g_2(y_{n_1-K})\geq\delta, \end{aligned}

where

\begin{align*} \delta:=\min\limits_{j=0, 1, \cdots, K}\{a_1x_{n_1-j}+a_2y_{n_1-j}, f_1(x_{n_1-j})+g_1(y_{n_1-j}), f_2(x_{n_1-j})+g_2(y_{n_1-j})\}>0. \end{align*}

Therefore, x_{n_1+1}+y_{n_1+1}\in[\delta, \ \tilde{p}] . This in particular means that \tilde{p}-a_1x_{n_1+1}-a_2y_{n_1+1}>0 or \tilde{u}_{n_1+1} \in(0, 1], \tilde{v}_{n_1+1}\in(0, 1] .

Continuing this process we obtain sequences \tilde{u}_{n}, \tilde{v}_{n} and x_{n}, y_{n}, x_{n}+y_{n} such that

\tilde{u}_{n} \in(0, 1], \ \tilde{v}_{n}\in(0, 1], \ \mbox{and}\ 0<x_{n+1}\leq \tilde{x}, \ 0<y_{n+1}\leq \tilde{y}, \ x_{n+1}+y_{n+1}\in[\delta, \widetilde{p}], \ \ \forall n\geq n_1.

Moreover,

\begin{equation} \tilde{u}_{n}= 1\ \mbox{if}\ x_{n}<\tilde{x}, \ \ \tilde{v}_{n}=1\ \mbox{if}\ y_{n}<\tilde{y}\ \ . \end{equation} (3.16)

Thus, (3.15) defines a solution to (3.1).

Now we show that \lim_{n\rightarrow \infty}(x_{n}+y_{n})\in\tau'. We first claim that if there is \bar{x}>0 such that \lim_{n\rightarrow \infty}x_n=\bar{x} , then there is \bar{y}>0 such that \lim_{n\rightarrow \infty}y_n=\bar{y} and \bar{y}+\bar{x}=\bar{p}\in\tau^{\prime} . If the claim is not true, then let \liminf_{n\rightarrow \infty}y_n=y^- , \limsup_{n\rightarrow \infty}y_n=y^+ . Since u_n\in[0, 1] , by taking the limit of (3.1) we have

y^+ \leq a_2y^++f_2(\bar{x})+g_2(y^+).

On the other hand, for y^- < \tilde{y} , then v_n=1, take the limit of (3.1) we have

y^-\geq a_2y^-+f_2(\bar{x})+g_2(y^-).

Thus, we obtain

y^–a_2y^–g_2(y^-) \geq y^+-a_2y^+-g_2(y^+).

Note that H_1^{(2)}(r)=(1-a_2)r-g_2(r) is increasing and y^-\leq y^+ , we have y^-= y^+. That is, \{y_n\} is also convergent. Set y':=\lim_{n\rightarrow \infty}y_n , \bar{y}=\bar{p}-\bar{x} , \bar{p}\in\tau' . We claim that y'=\bar{y} . By the definition of \tau' ,

\bar{y}=a_2\bar{y}+\bar{v}_1(f_2(\bar{x})+g_2(\bar{y}))\leq a_2\bar{y}+f_2(\bar{x})+g_2(\bar{y}),

together with y' = a_2y'+f_2(\bar{x})+g_2(y') , we obtain

\bar{y}-a_2\bar{y}-g_2(\bar{y}))\leq f_2(\bar{x}) = y' -a_2y'-g_2(y'),

thus our claim is true. Similarly, if there is \bar{y}>0 such that \lim_{n\rightarrow \infty}y_n=\bar{y} , then there is \bar{x}>0 such that \lim_{n\rightarrow \infty}x_n=\bar{x} and \bar{x}+\bar{y}=\bar{p}\in\tau^{\prime} .

Denote q:=\liminf_{n\rightarrow \infty}(x_n+y_n) , q_x:=\limsup_{n\rightarrow \infty}x_n , q_y:=\liminf_{n\rightarrow \infty}y_n , on the contrary assume, that is q<\tilde{p} . From x_{n}+y_{n}\in(\delta, \tilde{p}], \ \forall n\geq n_1 , we know that q\geq\delta>0 . Consider sequences k_m\rightarrow \infty and k'_m\rightarrow \infty such that

\begin{align*} &x_{k_m+1}\rightarrow q_x;\ \ y_{k'_m+1}\rightarrow q_y;\\ &x_{k_m}\rightarrow q_x'\leq q_x;\ \ y_{k'_m}\rightarrow q'_y\geq q_y;\\ &x_{k_m-K}\rightarrow q_x''\leq q_x;\ \ y_{k^{\prime}_m-K}\rightarrow q'_y\geq q_y. \end{align*}

Since for sufficiently large k'_m the inequalities \ y_{k^{\prime}_m+1}<\ \tilde{y} hold, from (3.16) we have \tilde{u}_{k_m}= 1, \ \tilde{v}_{k'_m}=1 . Then

\begin{equation} \begin{aligned} x_{k_m+1}\leq a_1x_{k_m}+f_1(x_{k_m-K})+g_1(y_{k_m-K}), \\ y_{k'_m+1}=a_2y_{k'_m}+f_2(x_{k'_m-K})+g_2(y_{k'_m-K}). \end{aligned} \end{equation} (3.17)

By taking the limit of the first equation of (3.17) and since f_1, g_2 are increasing and f_2, g_1 are decreasing, we have

\begin{aligned} q_x&\leq a_1q_x'+f_1(q_x'')+g_1(\lim\limits_{k_m\rightarrow \infty}y_{k_m-K})\\ &\leq a_1q_x+f_1(q_x)+g_1(q_y), \end{aligned}

that is, g_1(q_y) \geq H_1^{(1)}(q_x). Since H_1^{(1)} is continuous and increasing, then q_x \leq (H_1^{(1)})^{-1}(g_1(q_y)). On the other hand, take the limit of the second equation of (3.17) we also obtain

\begin{aligned} q_y&=a_2q_y'+f_2(\lim\limits_{k_m\rightarrow \infty}x_{k'_m-K})+g_2(q_y'')\\ &\geq a_2q_y+f_2(q_x)+g_2(q_y)\geq a_2q_y+f_2((H_1^{(1)})^{-1}(g_1(q_y)))+g_2(q_y), \end{aligned}

which is a contradiction with (3.14). If f_2, g_1 are increasing, let q_x:=\liminf_{n\rightarrow \infty}x_n , by taking the limit of the first equation of (3.17) we have

q_x=a_1q_x'+f_1(q_x'')+g_1(\lim\limits_{k_m\rightarrow \infty}y_{k_m-K}) \geq a_1q_x+f_1(q_x)+g_1(q_y),

that is, g_1(q_y) \leq H_1^{(1)}(q_x). Thus we also obtain the similar contradiction. Therefore, \lim_{n\rightarrow \infty}(x_{n}+y_{n})=\tilde{p}\in\tau' . Together with 0<x_{n+1}\leq \tilde{x}, \ 0<y_{n+1}\leq \tilde{y} , one has \lim_{n\rightarrow \infty}(x_{n}, y_n)=(\tilde{x}, \tilde{y})\in \tau . Therefore, \lim_{n\rightarrow \infty}(x_{n}+y_{n})=\tilde{p}\in\tau' . Together with 0<x_{n+1}\leq \tilde{x}, \ 0<y_{n+1}\leq \tilde{y} , one has \lim_{n\rightarrow \infty}(x_{n}, y_n)=(\tilde{x}, \tilde{y})\in \tau .

4 Numerical Results

In this section, we will give an example. Given that a_1 = 0.51, a_2 = 0.65 and the initial string

\begin{align*} P_0 &= \left\lbrace (x_1, y_1), (x_2, y_2), (x_3, y_3), (x_4, y_4), (x_5, y_5), (x_6, y_6)\right\rbrace \\ &= \left\lbrace (24.8, 22.8), (20.4, 22), (23.6, 25), (21.2, 23.2), (22.6, 21.7) \right\rbrace . \end{align*}

\qquad \textbf{F} = \begin{pmatrix} f_1 \\ f_2 \end{pmatrix} = \begin{pmatrix} x^{0.78} \\ x^{0.65} \end{pmatrix} , \qquad \qquad \textbf{G} = \begin{pmatrix} g_1 \\ g_2 \end{pmatrix} = \begin{pmatrix} 0.06*\ln x \\ 0.08*\ln x \end{pmatrix}.

Under this circumstance, we obtain the equilibrium (X^{*}, Y^{*}) of system (1) is (27.3499, 25.2828)

From Figure 1, we can see that the higher value of time delay K is, the lower output of factory is. Moreover, whatever the value of K is, the yield of x_n and y_n will approach to equilibrium (27.3499, 25.2828) when n is sufficiently large.

Figure 1 Numerical simulation of Equation (2.1) with different time delay K , where the solid line is x_n and the dotted line is y_n

By Theorem 3.3 with optimal equilibrium (\tilde{x}, \tilde{y}) = (23.852, 22.217) , we obtain the numerical results of equation (3.1) as Figure 2 shown. The convergence of control sequences u_n and v_n could be seen in Table 1. In this example, we find that {(u_n, v_n)} converges to the same result {(0.9694, 0.9591)} with different time delay and the rate of convergence is barely affected by value K .

Figure 2 Numerical simulation of Equation (2.1) with different time delay K , where the solid line is x_n and the dotted line is y_n

Table 1
The convergence of control function u_n and v_n with different time delay K
References
[1]
Cooke K, Ivanov A. On the discretization of a delay differential equation[J]. Journal of Difference Equations and Applications, 2000, 6(1): 105-119. DOI:10.1080/10236190008808216
[2]
Hirsch M. Optimization and cooperative control strategies: proceedings of the 8th international conference on cooperative control and optimization[M]. New York: Springer Science and Business Media, 2009.
[3]
Mamedov M A. Asymptotical stability of optimal paths in nonconvex problems[M]. New York: Springer, 2009: 95-134.
[4]
Richtmyer R D, Morton K W. Difference methods for initial-value problems[J]. Physics Today, 1994, 12(4).
[5]
Braverman E, Liz E. Global stabilization of periodic orbits using a proportional feedback control with pulses[J]. Nonlinear Dynamics, 2012, 67(4): 2467-2475. DOI:10.1007/s11071-011-0160-x
[6]
El-Morshedy H A, Liz E. Globally attracting fixed points in higher order discrete population models[J]. Journal of Mathematical Biology, 2006, 53(3): 365-384.
[7]
Kuang Yang. Delay differential equations: with applications in population dynamics[M]. New York: Academic press, 1993.
[8]
Ivanov A F, Trofimchuk S I. Periodic solutions of a discretized differential delay equation[J]. Journal of Difference Equations and Applications, 2010, 16(2-3): 157-171. DOI:10.1080/10236190902998008
[9]
Gwinner J. On optimality conditions for infinite programs[M]. Berlin: Springer, 1981: 21-27.
[10]
Mammadov M A. Turnpike theory: stability of optimal trajectories[M]. Boston: Springer, 2009: 3948-3955.
[11]
Zaslavski A. Turnpike properties in the calculus of variations and optimal control[M]. New York: Springer Science and Business Media, 2005.
[12]
Pardalos P M, Yatsenko V A. Optimization and control of bilinear systems: theory, algorithms, and applications[M]. New York: Springer Science and Business Media, 2010.
[13]
Ivanov A F, Mammadov M A, Trofimchuk S I. Global stabilization in nonlinear discrete systems with time-delay[J]. Journal of Global Optimization, 2013, 56(2): 251-263.