Loading [MathJax]/jax/element/mml/optable/GeneralPunctuation.js
Special Issues

Pullback attractors for stochastic recurrent neural networks with discrete and distributed delays

  • In this paper, we investigate a class of stochastic recurrent neural networks with discrete and distributed delays for both biological and mathematical interests. We do not assume any Lipschitz condition on the nonlinear term, just a continuity assumption together with growth conditions so that the uniqueness of the Cauchy problem fails to be true. Moreover, the existence of pullback attractors with or without periodicity is presented for the multi-valued noncompact random dynamical system. In particular, a new method for checking the asymptotical compactness of solutions to the class of nonautonomous stochastic lattice systems with infinite delay is used.

    Citation: Meiyu Sui, Yejuan Wang, Peter E. Kloeden. Pullback attractors for stochastic recurrent neural networks with discrete and distributed delays[J]. Electronic Research Archive, 2021, 29(2): 2187-2221. doi: 10.3934/era.2020112

    Related Papers:

    [1] Meiyu Sui, Yejuan Wang, Peter E. Kloeden . Pullback attractors for stochastic recurrent neural networks with discrete and distributed delays. Electronic Research Archive, 2021, 29(2): 2187-2221. doi: 10.3934/era.2020112
    [2] Wenlong Sun . The boundedness and upper semicontinuity of the pullback attractors for a 2D micropolar fluid flows with delay. Electronic Research Archive, 2020, 28(3): 1343-1356. doi: 10.3934/era.2020071
    [3] Xin-Guang Yang, Lu Li, Xingjie Yan, Ling Ding . The structure and stability of pullback attractors for 3D Brinkman-Forchheimer equation with delay. Electronic Research Archive, 2020, 28(4): 1395-1418. doi: 10.3934/era.2020074
    [4] Shu Wang, Mengmeng Si, Rong Yang . Dynamics of stochastic 3D Brinkman-Forchheimer equations on unbounded domains. Electronic Research Archive, 2023, 31(2): 904-927. doi: 10.3934/era.2023045
    [5] Guowei Liu, Hao Xu, Caidi Zhao . Upper semi-continuity of pullback attractors for bipolar fluids with delay. Electronic Research Archive, 2023, 31(10): 5996-6011. doi: 10.3934/era.2023305
    [6] Yangrong Li, Shuang Yang, Qiangheng Zhang . Odd random attractors for stochastic non-autonomous Kuramoto-Sivashinsky equations without dissipation. Electronic Research Archive, 2020, 28(4): 1529-1544. doi: 10.3934/era.2020080
    [7] Pan Zhang, Lan Huang, Rui Lu, Xin-Guang Yang . Pullback dynamics of a 3D modified Navier-Stokes equations with double delays. Electronic Research Archive, 2021, 29(6): 4137-4157. doi: 10.3934/era.2021076
    [8] Lingrui Zhang, Xue-zhi Li, Keqin Su . Dynamical behavior of Benjamin-Bona-Mahony system with finite distributed delay in 3D. Electronic Research Archive, 2023, 31(11): 6881-6897. doi: 10.3934/era.2023348
    [9] Lianbing She, Nan Liu, Xin Li, Renhai Wang . Three types of weak pullback attractors for lattice pseudo-parabolic equations driven by locally Lipschitz noise. Electronic Research Archive, 2021, 29(5): 3097-3119. doi: 10.3934/era.2021028
    [10] Yadan Shi, Yongqin Xie, Ke Li, Zhipiao Tang . Attractors for the nonclassical diffusion equations with the driving delay term in time-dependent spaces. Electronic Research Archive, 2024, 32(12): 6847-6868. doi: 10.3934/era.2024320
  • In this paper, we investigate a class of stochastic recurrent neural networks with discrete and distributed delays for both biological and mathematical interests. We do not assume any Lipschitz condition on the nonlinear term, just a continuity assumption together with growth conditions so that the uniqueness of the Cauchy problem fails to be true. Moreover, the existence of pullback attractors with or without periodicity is presented for the multi-valued noncompact random dynamical system. In particular, a new method for checking the asymptotical compactness of solutions to the class of nonautonomous stochastic lattice systems with infinite delay is used.



    Recurrent Neural Networks arise in a wide range of applications such as classification, combinatorial optimization, parallel computing, signal processing and pattern recognition, (see, e.g. [7,9,14,16,22,23,28]). Due to the finite switching speed of neurons and amplifiers, time delays commonly occured in neural networks. Since time delays will affect the stability of the neural system and may lead to some complex dynamic behavior, it is critical to study delayed recurrent neural networks. In particular, signal propagation is not instantaneous and may not be suitably modeled with discrete delay, so it is more appropriate to incorporate continuously distributed delays in neural network models.

    Random effects arise naturally in neural network models to take into account the uncertainty. Given τR and t>τ, in this paper, we will consider the following general class of stochastic neural networks with discrete and distributed delays:

    ˙xi(t)=fi(xi(t))+i+Nj=iNaij(t)g1j(θtω,xj(t))+i+Nj=iNbij(t)g2j(θtω,xj(tˆh(t)))+i+Nj=iNcij(t)0g3j(θtω,r,xj(t+r))dr+Ji(t),iZ, (1.1)

    with the initial condition

    xi(t)=ϕi(tτ),t(,τ],iZ, (1.2)

    where Z denotes the integer set; xi(t) represents the state variable of the potential for the i-th neuron at time t; fi denotes the behaved function; gkj(k=1,2,3) are activation functions of the neuron; aij(t), bij(t) and cij(t) denote the connection weight, discretely delayed connection weight and distributively delayed connection weight, respectively, between the j-th and i-th neurons, aij(t), bij(t) and cij(t) belong to C(R;R+); ˆh(t) stands for discrete time varying delay and ˆh(t) belongs to C(R;[0,h]) with constant h>0; Ji(t) represents the external force.

    Robust analysis for stochastic neural networks with time-varying delay can be found in [20,35]. Exponential stability of stochastic neural networks with constant or time-varying delays has been studied in [8,15,16,19,21,30]. Exponential stability of stochastic recurrent neural networks with time-varying delays was investigated in [25]. Asymptotic stability of stochastic neural networks with discrete and distributed delays has been developed, e.g., Markovian jumping parameters [26,27,29], Brownian motion [12], impulsive effects [23], and infinite delay [2,18]. There has, however, been little mention of pullback attractors for stochastic neural networks.

    The long-time behavior of multi-valued non-autonomous and random dynamical systems has been extensively developed over the last one and a half decades; see, e.g. [3,4,10,11,13,17,24] etc. The theory of pullback attractors for single-valued noncompact random dynamical systems has been established in [B.X. Wang]. The existence of pullback attractors has been studied in [33] for reaction-diffusion equations on an unbounded domain with non-autonomous deterministic as well as stochastic forcing terms for which the uniqueness of solutions need not hold (see also [34] for unbounded delay case). Based on the previous work, our main goal in this paper is to develop new theory of multi-valued noncompact random dynamical systems in a biological context to analyze the dynamics of a class of stochastic recurrent neural networks with discrete and distributed delays. It is worthy mentioning that we do not assume any Lipschitz condition on the nonlinear term, just a continuity assumption together with growth conditions.

    The paper is organized as follows. Section 2 gives some preliminary definitions and results regarding pullback attractors of multi-valued noncompact random dynamical systems, while in Section 3 the existence of solutions for the multi-valued noncompact random dynamical systems is considered. Sections 4-6 are devoted to the existence of pullback attractors and periodic attractors for stochastic recurrent neural networks with discrete and distributed delays.

    We now recall some basic definitions for multi-valued noncompact random dynamical systems and some results ensuring the existence of a pullback attractor for these systems.

    Let Q be a nonempty set, (Ω,F,P) be a probability space, and (X,d) be a Polish space with Borel σ-algebra B(X). Denote by P(X) and C(X) the sets of all nonempty and nonempty closed subsets of X, respectively. Let also denote by dist(A,B) the Hausdorff semidistance, i.e., for given subsets A and B of X we have

    dist(A,B)=sup{d(x,B):xA},

    where d(x,B)=inf{d(x,y):yB}. Finally, denote by Nr(A) the open r-neighborhood {yX:d(y,A)<r} of radius r>0 of a subset A of X.

    Assume that there are two groups {σt}tR and {θt}tR acting on Q and Ω, respectively. Specifically, σ:R×QQ is a mapping such that σ0 is the identity on Q, σt+τ=σtστ for all t,τR. Similarly, θ:R×ΩΩ is a (B(R)×F,F)-measurable mapping such that θ0 is the identity on Ω, θt+τ=θtθτ for all t,τR and θtP=P for all tR. In the sequel, we will call both (Q,{σt}tR) and (Ω,F,P,{θt}tR) parametric dynamical systems.

    Definition 2.1. Let (Q,{σt}tR) and (Ω,F,P,{θt}tR) be parametric dynamical systems. A multi-valued mapping Φ:R+×Q×Ω×XP(X) is called a multi-valued cocycle on X over (Q,{σt}tR) and (Ω,F,P,{θt}tR) if for all qQ, ωΩ and t,τR+, the following conditions are satisfied:

    (1) Φ(0,q,ω,) is the identity on X;

    (2) Φ(t+τ,q,ω,)=Φ(t,στq,θτω,Φ(τ,q,ω,)).

    For the above composition of multi-valued mappings, we use that for any nonempty set VX, Φ(t,q,ω,V) is defined by

    Φ(t,q,ω,V)=x0VΦ(t,q,ω,x0).

    Definition 2.2. (See [3,31,33].) A set-valued mapping K:Q×ΩP(X) is called measurable with respect to F in Ω if the mapping ωΩd(x,K(q,ω)) is (F,B(R))-measurable for every fixed xX and qQ.

    In what follows denote by D be a collection of some families of nonempty subsets of X parametrized by qQ and ωΩ.

    Definition 2.3. Let D be a collection of some families of nonempty subsets of X parametrized by qQ and ωΩ. D is said to be neighborhood closed if for each D={D(q,ω):qQ,ωΩ}D, there exists a positive number ε depending on D such that the family

    {B(q,ω):B(q,ω) is a nonempty subset of Nε(D(q,ω)),qQ,ωΩ} (2.1)

    also belongs to D.

    Note that the neighborhood closedness of D implies for each DD,

    {˜D(q,ω):˜D(q,ω) is a nonempty subset of D(q,ω),qQ,ωΩ}D. (2.2)

    A collection D satisfying (2.2) is said to be inclusion-closed in the literature, see, e.g., [11].

    Definition 2.4. (See [3,31,33].) Let D be a collection of some families of nonempty subsets of X and K={K(q,ω):qQ,ωΩ}D. Then K is called a D-pullback absorbing set for Φ if for all qQ, ωΩ and for every B={B(q,ω):qQ,ωΩ}D, there exists T=T(B,q,ω)>0 such that

    Φ(t,σtq,θtω,B(σtq,θtω))K(q,ω), for all tT.

    In addition, if K is measurable with respect to the P-completion of F, then K is said to be a measurable D-pullback absorbing set for Φ.

    Definition 2.5. Let D be a collection of some families of nonempty subsets of X. Then Φ is said to be D-pullback asymptotically upper semicompact in X if for all qQ and ωΩ, any sequence ynΦ(tn,σtnq,θtnω,xn) has a convergent subsequence in X whenever tn+(n), xnB(σtnq,θtnω) with {B(q,ω):qQ,ωΩ}D.

    Definition 2.6. Let D be a collection of some families of nonempty subsets of X and A={A(q,ω):qQ,ωΩ}D. Then A is said to be a D-pullback attractor for Φ if it satisfies:

    (1) A(q,ω) is compact for all qQ and ωΩ.

    (2) A is invariant, that is, for every qQ and ωΩ,

    Φ(t,q,ω,A(q,ω))=A(σtq,θtω),t0.

    (3) A attracts every member of D, that is, for every B={B(q,ω):qQ,ωΩ}D and for every qQ and ωΩ,

    limt+dist(Φ(t,σtq,θtω,B(σtq,θtω)),A(q,ω))=0.

    The following result shows a sufficient and necessary criterion for the existence and uniqueness of pullback attractors associated to multi-valued cocycles [33], see also [31] for the single-valued case.

    Theorem 2.7. Let D be a neighborhood closed collection of some families of nonempty subsets of X, and let Φ be a multi-valued cocycle on X over (Q,{σt}tR) and (Ω,F,P,{θt}tR) possessing the norm-to-weak upper semicontinuity on X, i.e., if xnx in X, then for any ynΦ(t,q,ω,xn), there exist a subsequence ynk and a yΦ(t,q,ω,x) such that ynky (weak convergence). Then Φ has a D-pullback attractor A in D if and only if Φ is D-pullback asymptotically upper semicompact in X and Φ has a closed D-pullback absorbing set K in D. The D-pullback attractor A is unique and is given by, for each qQ and ωΩ,

    A(q,ω)=Θ(K,q,ω)=BDΘ(B,q,ω), (2.3)

    where the family {Θ(B,q,ω):qQ,ωΩ} is called the Θ-limit set of B defined by

    Θ(B,q,ω)=τ0¯tτΦ(t,σtq,θtω,B(σtq,θtω)).

    By the similar arguments of Theorem 2.25 in [31], we have a sufficient and necessary criterion for the periodicity of pullback attractors of multi-valued cocycles.

    Theorem 2.8. Let D be a neighborhood closed collection of some families of nonempty subsets of X, and let Φ be a norm-to-weak upper semicontinuous periodic multi-valued cocycle with period T>0 on X over (Q,{σt}tR) and (Ω,F,P,{θt}tR), i.e., for every t0, qQ and ωΩ, there holds

    Φ(t,σTq,ω,)=Φ(t,q,ω,).

    Suppose Φ has a D-pullback attractor AD. Then A is periodic with period T, i.e., A(σTq,ω)=A(q,ω) for all qQ and ωΩ if and only if Φ has a closed D-pullback absorbing set KD with K being periodic with period T.

    Let (Ω,F,P) be a probability space. On this probability space we consider a measurable non-autonomous group θ:

    θ:(R×Ω,B(R)F)(Ω,F).

    In addition, we assume that P is ergodic with respect to θ, which means that every θt-invariant set has measure zero or one, tR. Therefore P is invariant with respect to θt. Then (Ω,F,P,{θt}tR), which is the model for a noise, is a parametric dynamical system. Suppose Q=R. Define a family {σt}tR of shift operators by σt(s)=s+t for all t,sR.

    We also recall the following well-known ergodic theorem.

    Theorem 3.1. Suppose Y is a real random variable in L1. Then

    limt±1tt0Y(θsω)ds=EY

    on a {θt}tR-invariant set of measure one.

    Outside this set of measure one we will replace the values of Y by EY so that this version of Y has the above limit for all ωΩ.

    Let

    l2={x=(xi)iZ,xiR:iZx2i<+},

    and equip it with the inner product and norm as

    (x,y)=iZxiyi,x2=(x,x),x=(xi)iZ,y=(yi)iZl2.

    We denote by Cγ,l2 the space

    Cγ,l2={φC((,0];l2):limsφ(s)eγs exists},

    where the parameter γ>0 will be determined later on. If we define

    φCγ,l2=(iZsups(,0]e2γs|φi(s)|2)12,φCγ,l2,

    then (Cγ,l2,Cγ,l2) is a Banach space. Given τR, T>τ and a function x:(,T]l2, for each t[τ,T) we denote by xt the function defined on (,0] by the relation xt(s)=x(t+s), s(,0]. In the following sections, C denotes an arbitrary positive constant, which may be different from line to line and even in the same line.

    We consider the following conditions:

    (C1) The mappings ωgkj(ω,x) are F-measurable for any fixed xR, and the mappings (t,x)gkj(θtω,x) are continuous from R×R into R for any fixed ωR, where k=1,2 and jZ. Similarly, the mappings ωg3j(ω,r,x) are F-measurable for any fixed (r,x)R×R, and the mappings (t,r,x)g3j(θtω,r,x) are continuous from R×R×R into R for any fixed ωΩ, where jZ.

    (C2) For each iZ, fi is continous, and there exist a positive constant h1>0 and h2=(h2i)iZl2 such that

    fi(x)xh1x2+h22i,iZ,xR.

    Besides, there exist a positive constant l1>0 and l2=(l2i)iZl2 such that

    |fi(x)|l1|x|+l2i,iZ,xR.

    (C3) For k=1,2 and jZ, there exist nonnegative functions pkj, qkj:ΩR, which are measurable with respect to F, such that

    |gkj(ω,x)|2p2kj(ω)|x|2+q2kj(ω),ωΩ,xR,

    where the mappings tpkj(θtω) and tqkj(θtω) are continuous from R into R for any fixed ωΩ.

    Besides, there exist nonnegative functions ˆp3j, ˆq3j:Ω×RR such that

    |g3j(ω,r,x)|ˆp3j(ω,r)|x|+ˆq3j(ω,r),ωΩ,rR,xR,

    where the mappings ωˆp3j(ω,r) and ωˆq3j(ω,r) are F-measurable for any fixed rR, and the mappings rˆp3j(ω,r) and rˆq3j(ω,r) are continuous from R into R for any fixed ωΩ.

    Also, for any ωΩ and tR, we define

    ˜p3j(θtω):=0eγrˆp3j(θtω,r)dr,˜q3j(θtω):=0ˆq3j(θtω,r)dr,

    where the mappings tˆp3j(θtω,r) and tˆq3j(θtω,r) are continuous from R into R for any fixed ωΩ and rR. In addition, there exists a measurable mapping Λ:Ω×RR such that the mapping tΛ(θtω) is continuous from R into R for any fixed ωΩ, and

    pkj(ω)Λ(ω),qkj(ω)Λ(ω),˜p3jΛ(ω),˜q3jΛ(ω),jZ,k=1,2,ωΩ.

    (C4) For i,jZ, aij, bij and cij belong to C(R;R+), and

    aij(t)˜aij,bij(t)˜bij,cij(t)˜cij,tR.

    Moreover,

    iZi+Nj=iN(˜a2ij+˜b2ij+˜c2ij)<.

    (C5) The external force (Ji)iZ belongs to C(R;l2), and

    τiZeh13r|Ji(r)|2dr<,τR,

    which implies that

    limk+τ|i|keh13r|Ji(r)|2dr=0,τR,

    where the constant h1 is the same as that of Assumption (C2).

    (C6) The mappings tp21j(θtω), tp22j(θtω) and t˜p23j(θtω) are sub-exponentially growing as t± for any fixed ωΩ, where jZ. In other words, for ε>0 and ωΩ, there exists a t0(ε,ω)>0 such that for |t|t0(ε,ω) it holds that

    p21j(θtω)eε|t|,p22j(θtω)eε|t|,˜p23j(θtω)eε|t|.

    Similarly, the mappings tq21j(θtω), tq22j(θtω) and t˜q23j(θtω) are sub-exponentially growing as t± for any fixed ωΩ, where jZ, which means that for ε>0 and ωΩ, there exists a t0(ε,ω)>0 such that for |t|t0(ε,ω) it holds that

    q21j(θtω)eε|t|,q22j(θtω)eε|t|,˜q23j(θtω)eε|t|.

    (C7) We suppose that EΛ2<, and also that

    limt±1tt0Λ2(θrω)dr=ˉΛ.

    By the ergodicity assumption and Theorem 3.1 we obtain that

    limt±1tt0Λ2(θrω)dr=EΛ2=:ˉΛ,

    on a {θt}tR-invariant set of full measure. Let us replace outside this set (which has measure zero) the values of Λ2(ω) by ˉΛ.

    Remark 1. Let us define

    αij(θtω):=˜a2ijp21j(θtω)+˜b2ijp22j(θtω)+˜c2ij˜p23j(θtω)

    and

    βij(θtω):=˜a2ijq21j(θtω)+˜b2ijq22j(θtω)+˜c2ij˜q23j(θtω),

    where i,jZ. Then from Assumption (C6) we obtain that αij(θtω) and βij(θtω) are sub-exponentially growing as t± for any fixed ωΩ.

    Lemma 3.2. Let (C1)-(C5) hold. Then for any fixed τR, ωΩ and each M>0, there exists T(M,ω)>0 such that if ϕCγ,l2 and ϕCγ,l2M, then problem (1.1) and (1.2) admits at least a solution x(t)=x(t;τ,ω,ϕ) defined on [τ,τ+T(M,ω)], and x belongs to the space C1([τ,τ+T(M,ω)];l2).

    The proof of Lemma 3.2 is given in the Appendix.

    By slightly modifying the proof of Lemmas 4.1 and 4.2, we see that every solution can be globally defined. Hence we now define a multi-valued mapping Φ:R+×R×Ω×Cγ,l2P(Cγ,l2) by

    Φ(t,τ,ω,ϕ)={xt+τ(,τ,θτω,ϕ)|x()isasolutionofEqs.(1.1)(1.2)withϕCγ,l2}.

    Lemma 3.3. The mapping Φ is a multi-valued cocycle on Cγ,l2 over (R,{σt}tR) and (Ω,F,P, {θt}tR).

    Proof. We only need to check condition (2) in Definition 2.1, since condition (1) follows immediately. Let zΦ(t+s,τ,ω,ϕ). Then there exists a solution x of (1.1)-(1.2) such that z=xt+τ+s(,τ,θτω,ϕ). We define a function u by ut+τ=xt+τ+s for t0 and τR. It is clear that uτ=xτ+s, and the function u solves (1.1) with ω, aij(t), bij(t), cij(t), Ji(t), ˆh(t) replaced by θsω, aij(t+s), bij(t+s), cij(t+s), Ji(t+s) and ˆh(t+s), respectively, and ϕ=xτ+s. Indeed, for r[t,0], we obtain

    ut+τ(r)=xt+τ+s(r)=ϕ(0)+t+τ+s+rτˆf(θrω,r,xr)dr=ϕ(0)+τ+sτˆf(θrω,r,xr)dr+t+τ+s+rτ+sˆf(θrω,r,xr)dr=xτ+s(0)+t+τ+s+rτ+sˆf(θrω,r,xr)dr=uτ(0)+t+τ+rτˆf(θr+sω,r+s,ur)dr,

    where ˆf(θrω,r,xr)=(ˆfi(θrω,r,xir))iZ and ˆfi(θrω,r,xir)=fi(xi(r))+i+Nj=iNaij(r)g1j(θrω,xj(r))+i+Nj=iNbij(r)g2j(θrω,xj(rˆh(r)))+i+Nj=iNcij(r)0g3j(θrω,r,xj(r+r))dr+Ji(r).

    Note that ut+τ(r)=xs(t+τ+r) for r<t. Therefore

    zΦ(t,τ+s,θsω,xτ+s)Φ(t,τ+s,θsω,Φ(s,τ,ω,ϕ)).

    Since z is arbitrary, we have Φ(t+s,τ,ω,ϕ)Φ(t,τ+s,θsω,Φ(s,τ,ω,ϕ)).

    On the other hand, let zΦ(t,τ+s,θsω,Φ(s,τ,ω,ϕ)). Then there exist x solving (1.1) and y solving (1.1)(with ω, aij(t), bij(t), cij(t), Ji(t) and ˆh(t) replaced by θsω, aij(t+s), bij(t+s), cij(t+s), Ji(t+s) and ˆh(t+s), respectively) and such that yτ=xs+τ and yt+τ=z. Define the function

    wt={xt,ifτts+τ,yts,ifs+τt,

    which is a solution to (1.1). Indeed, for ts+τ the equality wt=xt ensures that w() is a solution. If ts+τ, then for r[τ+st,0] we have

    wt(r)=yts(r)=yτ(0)+ts+rτˆf(θr+sω,r+s,yr)dr=xs+τ(0)+t+rτ+sˆf(θrω,r,wr)dr=xτ(0)+t+rτˆf(θrω,r,wr)dr.

    Also, for r[τt,τ+st] we find that

    wt(r)=yts(r)=xs(t+rs)=xτ(0)+t+rτˆf(θrω,r,xr)dr=xτ(0)+t+rτˆf(θrω,r,wr)dr.

    Finally, for r<τt it is easy to see that

    wt(r)=xs(t+rs)=ϕ(t+rτ).

    Therefore, z=yt+τ=wt+τ+sΦ(t+s,τ,ω,ϕ). Since z is arbitrary, we obtain that

    Φ(t,τ+s,θsω,Φ(s,τ,ω,ϕ))Φ(t+s,τ,ω,ϕ).

    In this section, we establish uniform estimates of solutions of problem (1.1)-(1.2) which are needed for proving the existence of pullback absorbing sets of the system.

    Let B be a bounded nonempty subset of Cγ,l2, and denote by BCγ,l2=supφBφCγ,l2. Assume D={D(τ,ω):τR,ωΩ} is a family of bounded nonempty subsets of Cγ,l2 satisfying, for every τR and ωΩ,

    limreh12rD(τ+r,θrω)2Cγ,l2=0, (4.1)

    where the constant h1 is the same as that of Assumption (C2). Denote by D the collection of all families of bounded nonempty subsets of Cγ,l2 which fulfill condition (4.1), i.e.,

    D={D={D(τ,ω):τR,ωΩ}:Dsatisfies (4.1)}.

    Obviously, D is neighborhood closed.

    Lemma 4.1. Suppose (C1)-(C5) hold and assume that

    58h1<γ. (4.2)

    Then for every τR and ωΩ, any solution x of Eqs. (1.1)-(1.2) with ω replaced by θτω satisfies for all t0,

    xτ(,τt,θτω,ϕ)2Cγ,l2Ceh1t+8h1(2N+1)2e2γh0tiZi+Nj=iNαij(θsω)dsϕ2Cγ,l2+C0t(iZi+Nj=iNβij(θsω)+J(τ+s)2+h22)×eh1s+0s8h1(2N+1)2e2γhiZi+Nj=iNαij(θsω)dsds,

    where αij(θsω) and βij(θsω) are the same as those of Remark 1.

    The proof of Lemma 4.1 is given in the Appendix.

    Lemma 4.2. Let (C1)-(C7) and (4.2) hold. Also, assume that

    8(2N+1)2e2γhiZi+Nj=iN(˜a2ij+˜b2ij+˜c2ij)EΛ2<12h21. (4.3)

    Then the closed ball K(τ,ω) in Cγ,l2 with center zero and random radius R(τ,ω) where

    (R(τ,ω))2=C0(iZi+Nj=iNβij(θsω)+J(τ+s)2+h22)×eh1s+0s8h1(2N+1)2e2γhiZi+Nj=iNαij(θsω)dsds

    is contained in D, and K={K(τ,ω):τR,ωΩ} is a measurable D-pullback absorbing set for Φ.

    Proof. It follows from Remark 1 that for any fixed ωΩ, the mappings tβij(θtω) are sub-exponentially growing for t±, where iZ and j=iN,,i+N. Hence for 0<ε<112h1 and ωΩ, there exists a t0(ε,ω) such that for |t|t0(ε,ω),

    βij(θtω)(˜a2ij+˜b2ij+˜c2ij)eε|t|,

    where iZ and j=iN,,i+N.

    Thanks to Assumptions (C4)-(C7), in view of Remark 1, we deduce that

    Ceh1r20iZi+Nj=iNβij(θs+rω)eh1s+0s8h1(2N+1)2e2γhiZi+Nj=iNαij(θs+rω)dsdsCeh1r20eε|s+r|eh1s×e8h1(2N+1)2e2γhiZi+Nj=iNMij((0s+r0r)(Λ2(θsω)ˉΛ)dsˉΛs)dsCeh1r20e3ε|s+r|eh1s8h1(2N+1)2e2γhiZi+Nj=iNMijˉΛsdsCe(12h13ε)r0e(12h13ε)sdsCe(12h13ε)r0 (4.4)

    as r, where we have used the notations ˉΛ=EΛ2 and

    Mij:=˜a2ij+˜b2ij+˜c2ij,

    and in the similar way, we have

    Ceh1r20(J(τ+r+s)2+h22)×eh1s+0s8h1(2N+1)2e2γhiZi+Nj=iNαij(θs+rω)dsdsCeh1r20(J(τ+r+s)2+h22)×eh1s+8h1(2N+1)2e2γhiZi+Nj=iNMij((0s+r0r)(Λ2(θsω)ˉΛ)dsˉΛs)dsCeh1r20(J(τ+r+s)2+h22)×eh1s+2ε|s+r|8h1(2N+1)2e2γhiZi+Nj=iNMijˉΛsdsCe(12h12ε)r0(J(τ+r+s)2+h22)e(12h12ε)sdsCe(12h12ε)ττ+re(12h12ε)s(J(s)2+h22)ds0 (4.5)

    as r. Therefore,

    eh1r2(R(τ+r,θrω))2=Ceh1r20(iZi+Nj=iNβij(θs+rω)+J(τ+r+s)2+h22)×eh1s+0s8h1(2N+1)2e2γhiZi+Nj=iNαij(θs+rω)dsds0asr.

    This implies that

    limreh1r2K(τ+r,θrω)2Cγ,l2=0, (4.6)

    and thus K={K(τ,ω):τR,ωΩ} belongs to D. By (C7), (4.3) and the ergodic Theorem 3.1, we obtain that

    eh1t+8h1(2N+1)2e2γh0tiZi+Nj=iNαij(θsω)dsB(τt,θtω)2Cγ,l2e12h1tB(τt,θtω)2Cγ,l20 (4.7)

    as t+, where ϕB(τt,θtω) and BD. Note that for each τR, (R(τ,ω))2:ΩR is (F,B(R))-measurable. Then it follows from Lemma 4.1, (4.6) and (4.7) that K={K(τ,ω):τR,ωΩ} is a closed measurable D-pullback absorbing set in D for Φ. This completes the proof of the lemma.

    In order to prove the asymptotically upper semicompactness for the multi-valued cocycle Φ, we need the following lemma.

    Lemma 5.1. Suppose (C1)-(C7), (4.2) and (4.3) hold. Let τR, ωΩ and B={B(τ,ω):τR,ωΩ}D. Then for every ε>0, there exist T=T(τ,ω,B,ε)>0 and ˉN=ˉN(τ,ω,B,ε)>0 such that any solution x() of Eqs. (1.1)-(1.2), given by x with xτ(,τt,θτω,ϕ)Φ(t,τt,θτω,ϕ) and ϕB(τt,θtω), satisfies

    |i|ˉNsups(,0]e2γs|xiτ(s,τt,θτω,ϕ)|2ε,for alltT. (5.1)

    Proof. Choose a smooth function ρ such that 0ρ(r)1 for rR+, and

    ρ(r)=0for0r1,ρ(r)=1forr2.

    Then there exists a constant C0 such that |ρ(r)|C0 for rR+. Define ρM(|i|) := ρ(|i|M). Multiplying (1.1) by ρM(|i|)xi we have that

    12ddtρM(|i|)|xi(t)|2=ρM(|i|)fi(xi(t))xi(t)+i+Nj=iNρM(|i|)aij(t)g1j(θtω,xj(t))xi(t)+i+Nj=iNρM(|i|)bij(t)g2j(θtω,xj(tˆh(t)))xi(t)+i+Nj=iNρM(|i|)cij(t)xi(t)0g3j(θtω,r,xj(t+r))dr+ρM(|i|)Ji(t)xi(t).

    In a similar way as in Lemma 4.1, by Assumptions (C2)-(C4), Young's inequality, and (Nj=1uj)2NNj=1u2j, we obtain that

    ρM(|i|)fi(xi(t))xi(t)ρM(|i|)h1|xi(t)|2+ρM(|i|)h22i, (5.2)
    i+Nj=iNρM(|i|)aij(t)g1j(θtω,xj(t))xi(t)4h1(2N+1)i+Nj=iNρM(|i|)a2ij(t)g21j(θtω,xj(t))+116h1ρM(|i|)|xi(t)|24h1(2N+1)i+Nj=iNρM(|i|)˜a2ijp21j(θtω)sups(,0]e2γs|xjt(s)|2+116h1ρM(|i|)|xi(t)|2+4h1(2N+1)i+Nj=iNρM(|i|)˜a2ijq21j(θtω), (5.3)
    i+Nj=iNρM(|i|)bij(t)g2j(θtω,xj(tˆh(t)))xi(t)4h1(2N+1)i+Nj=iNρM(|i|)b2ij(t)g22j(θtω,xj(tˆh(t)))+116h1ρM(|i|)|xi(t)|24h1(2N+1)i+Nj=iNρM(|i|)˜b2ijp22j(θtω)e2γhsups(,0]e2γs|xjt(s)|2+116h1ρM(|i|)|xi(t)|2+4h1(2N+1)i+Nj=iNρM(|i|)˜b2ijq22j(θtω), (5.4)
    |i+Nj=iNρM(|i|)cij(t)xi(t)0g3j(θtω,r,xj(t+r))dr|i+Nj=iNρM(|i|)cij(t)|xi(t)|0(ˆp3j(θtω,r)|xj(t+r)|+ˆq3j(θtω,r))dri+Nj=iNρM(|i|)˜cij(˜p3j(θtω)sups(,0]eγs|xjt(s)|+˜q3j(θtω))|xi(t)|4h1(2N+1)i+Nj=iNρM(|i|)˜c2ij˜p23j(θtω)sups(,0]e2γs|xjt(s)|2+18h1ρM(|i|)|xi(t)|2+4h1(2N+1)i+Nj=iNρM(|i|)˜c2ij˜q23j(θtω), (5.5)

    and

    ρM(|i|)Ji(t)xi(t)116h1ρM(|i|)|xi(t)|2+4h1ρM(|i|)|Ji(t)|2. (5.6)

    Note that e2γh>1. Then, from (5.2)-(5.6) it follows that

    ddt(ρM(|i|)|xi(t)|2)118h1ρM(|i|)|xi(t)|2+8h1(2N+1)ρM(|i|)i+Nj=iNαij(θtω)e2γhsups(,0]e2γs|xjt(s)|2+8h1ρM(|i|)|Ji(t)|2+8h1(2N+1)ρM(|i|)i+Nj=iNβij(θtω)+2ρM(|i|)h22i, (5.7)

    where αij(θtω) and βij(θtω) are given in Remark 1. And consequently,

    ddt(e54h1tρM(|i|)|xi(t)|2)=54h1e54h1tρM(|i|)|xi(t)|2+e54h1tddt(ρM(|i|)|xi(t)|2)18h1e54h1tρM(|i|)|xi(t)|2+8h1(2N+1)ρM(|i|)e54h1ti+Nj=iNβij(θtω)+8h1ρM(|i|)e54h1t|Ji(t)|2+2ρM(|i|)e54h1th22i+8h1(2N+1)e2γhρM(|i|)i+Nj=iNαij(θtω)e54h1tsups(,0]e2γs|xjt(s)|2. (5.8)

    Integrating (5.8) over [τt,t] with t0 and tτ, we find that for every ωΩ,

    e54h1tρM(|i|)|xi(t,τt,ω,ϕ)|2e54h1(τt)ρM(|i|)|xi(τt,τt,ω,ϕ)|218h1tτte54h1rρM(|i|)|xi(r,τt,ω,ϕ)|2dr+8h1(2N+1)e2γhρM(|i|)tτti+Nj=iNαij(θrω)e54h1r×sups(,0]e2γs|xjr(s,τt,ω,ϕ)|2dr+8h1(2N+1)ρM(|i|)tτti+Nj=iNβij(θrω)e54h1rdr+ρM(|i|)tτte54h1r(8h1|Ji(r)|2+2h22i)dr. (5.9)

    Neglecting the second term on the right-hand side of (5.9). Note that 58h1<γ, so e(2γ54h1)s1 for s0. Setting t+s instead of t, multiplying (5.9) by e54h1(t+s)e2γs and replacing ω by θτω, we have that for all s[τtt,0],

    ρM(|i|)e2γs|xit(s,τt,θτω,ϕ)|2e54h1(t+tτ)ρM(|i|)|xi(τt,τt,θτω,ϕ)|2+8h1(2N+1)e2γhe54h1tρM(|i|)tτti+Nj=iNαij(θrτω)e54h1r×sups(,0]e2γs|xjr(s,τt,θτω,ϕ)|2dr+8h1(2N+1)e54h1tρM(|i|)tτti+Nj=iNβij(θrτω)e54h1rdr+e54h1tρM(|i|)tτte54h1r(8h1|Ji(r)|2+2h22i)dr. (5.10)

    Note that for all ,

    and

    (5.11)

    Let . Then it follows that for all ,

    (5.12)

    Now we estimate each term on the right-hand side of (5.12). For the first term, since and , we see that

    (5.13)

    For the third term, Assumption ensures that we can find large enough such that for all ,

    (5.14)

    Let be given arbitrarily. Then there is such that for all ,

    (5.15)

    where is given in Lemma 4.2. By Assumption and Remark 1, we see that for and , there exists a such that for ,

    (5.16)

    where and . Hence, for the second term, using Assumption and (5.16), we have that for all and ,

    (5.17)

    Now we estimate the last term in (5.12). Similar to (5.17), we find that for all and ,

    (5.18)

    Note that and , using Assumptions , and the ergodic Theorem 3.1, in view of (4.3), we deduce that

    (5.19)

    as . In a similar way as in (5.19), by Assumptions and , (4.3), (5.16) and Theorem 4.1, we have

    (5.20)

    where we have used

    for sufficiently large . This and (5.20) ensure that for all ,

    (5.21)

    thanks to . From (4.3) we see that . Then similar to (5.20), we obtain that

    for sufficiently large and . Hence it follows from Assumption that for all ,

    (5.22)

    Hence for the last term in (5.12), by (F.43), (5.18)-(5.19) and (5.21)-(5.22), we can choose and sufficiently large such that

    (5.23)

    Finally, if we take and sufficiently large, then we deduce from (5.12)-(5.13), (5.14), (5.17) and (5.23) that for all ,

    (5.24)

    Thus the proof of this lemma is complete.

    First, let us prove some properties of the multi-valued cocycle .

    Lemma 6.1. Suppose -, (4.2) and (4.3) hold. Let be a sequence converging to in and fix . Then for any , and , there exist and such that for any solution of problem (1.1) with replaced by and it follows

    (6.1)

    Moreover, there exist and a subsequence satisfying

    (6.2)

    Proof. For any , there exist and such that

    and

    Hence,

    (6.3)

    if and . On the other hand, by slightly modifying the proof of Lemma 4.1, in view of in and Assumptions -, there exists such that

    (6.4)

    Integrating (5.8) over with , by (6.3)-(6.4), the continuity of , we can choose and sufficiently large such that for all ,

    (6.5)

    thanks to Assumptions and . From (6.3) and (6.5), the conclusion (6.1) follows immediately.

    Now it only remains to prove (6.2). Fix now . Taking into account (6.4), passing to a subsequence, we can state that weakly in . This and (6.1) imply that for any , there exist and such that

    (6.6)

    if . Therefore, strongly in , and consequently, is precompact for any .

    On the other hand, in view of (6.4) and in , by Assumptions - we deduce that there exists such that for all and ,

    (6.7)

    thanks to (6.4), and the continuity of . This implies that the sequence is equi-continuous. Then the Arzelà-Ascoli theorem ensures the existence of a subsequence converging in to some function . It is easy to show that is a solution of (1.1). Also, it is clear that .

    From Lemma 6.1 we have the following two results.

    Corollary 1. Suppose -, (4.2) and (4.3) hold. Then for any , and , the map has compact values.

    Corollary 2. Suppose -, (4.2) and (4.3) hold. Then for any , and , the map is upper semi-continuous, i.e., if in , then for any , there exists a subsequence and a such that in .

    We are now ready to show the existence of pullback attractors for .

    Theorem 6.2. Suppose -, (4.2) and (4.3) hold. Then the multi-valued cocycle associated with problem (1.1)-(1.2) has a unique -pullback attractor in .

    Proof. Note that by Lemma 4.2, Corollary 2 and Theorem 2.7, it only remains to prove the asymptotically upper semicompactness for .

    In order to prove the asymptotically upper semicompactness for , arguing as in Theorem 2.5 in [34], we only need to show that for any fixed , , every and any , there exist , , a and a such that

    (1) for all , ,

    (2) for each fixed

    (3) for all , , , with ,

    (4) for all , ,

    We divide the proof into two steps.

    Step 1. For , by making use of (F.41) with replaced by , we deduce that for all and with ,

    and further by (F.43) with replaced by , we obtain

    (6.8)

    Note that for all ,

    (6.9)

    thanks to . Note that and , by Assumption , , and the ergodic Theorem 3.1, in view of (4.2), (4.3) and (5.16), we find that there exists a and then we can choose large enough such that for all ,

    (6.10)
    (6.11)
    (6.12)

    where , and in a similar way, we have

    (6.13)
    (6.14)
    (6.15)
    (6.16)

    Inserting (6.11)-(6.16) into (6.8), in view of (6.10), we deduce that for all and ,

    which implies that holds true.

    Step 2. Thanks to Lemmas 4.1-4.2 and 5.1, and follow immediately.

    For , without loss of generality, we assume that with , by - we have that for all ,

    (6.17)

    where and are given in Lemma 4.1. Using (F.43), (5.19), (5.21)-(5.22) and (6.17), we obtain that for sufficiently large,

    (6.18)

    thanks to the continuity of , and , and thus holds. The proof is complete.

    By a similar argument as in [31], the following result can be obtained immediately by using Theorem 2.8.

    Theorem 6.3. Suppose -, (4.2) and (4.3) hold. If there exists such that for all ,

    (6.19)
    (6.20)

    where and , then the multi-valued cocycle associated with problem (1.1)-(1.2) has a unique periodic -pullback attractor in .

    Proof. Let us fix some . We can rewrite Eq. (1.1) as

    where and

    We divide the proof into two steps.

    Step 1. is well defined and bounded.

    We note that and

    In view of the Assumption and the trivial bound , we can obtain that

    (F.21)

    By and using the fact that , we have

    (F.22)

    In a similar way as above, by we deduce that

    (F.23)

    and

    (F.24)

    Then, using (F.21)-(F.24) and the assumption on we obtain that

    (F.25)

    Since belongs to for any fixed , in view of , it follows from (F.25) and that maps the bounded sets of into the bounded set of .

    Step 2. is continuous.

    We consider and such that , and and such that . Let be given arbitrarily. Then there exists such that for all ,

    (F.26)
    (F.27)

    Due to the continuity of and , in view of the Assumption , for any and sufficiently large , we have

    (F.28)

    By , - and (F.27), in view of the continuity of , we find that for all sufficiently large,

    (F.29)

    Arguing in the similar way as above, we deduce from , -, (F.27) and the continuity of , , and that for all sufficiently large,

    (F.30)

    and

    (F.31)

    thanks to Assumption and Lebesgue's dominated convergence theorem.

    Note that , hence for sufficiently large , we can deduce that

    (F.32)

    Then it follows from (F.28)-(F.32) that for all sufficiently large,

    (F.33)

    This implies that is continuous. Thus, Theorem 4 in [5] ensures that for any and , there exists at least one solution .

    Proof. Multiplying (1.1) by we obtain

    Let , and be positive parameters to be fixed later on. Note that takes the value in . Then by making use of Young's inequality, Assumptions - and , we find that

    (F.34)
    (F.35)
    (F.36)
    (F.37)

    and

    Note that . Therefore,

    (F.38)

    This implies that

    (F.39)

    Integrating (F.39) over with and , we obtain that for every ,

    (F.40)

    Let . Then we can neglect the second term on the right-hand side of (F.40). Note that , so for . Setting now instead of , multiplying (F.40) by and replacing by , we find that for all ,

    (F.41)

    where we have used the notations

    Note that for all ,

    Then, it holds

    (F.42)

    We observe that

    Then (F.42) can be rewritten as

    Using Gronwall's lemma, we have

    (F.43)

    Let , then for all we have

    and thus the proof of this lemma is finished.



    [1] J.-P. Aubin and H. Franskowska, Set-Valued Analysis, Systems & Control: Foundations & Applications, 2. Birkhäuser Boston, Inc., Boston, MA, 1990.
    [2] Global asymptotic stability of stochastic recurrent neural networks with multiple discrete delays and unbounded distributed delays. Appl. Math. Comput. (2008) 204: 680-686.
    [3] Non-autonomous and random attractors for delay random semilinear equations without uniqueness. Discrete Contin. Dyn. Syst. (2008) 21: 415-443.
    [4] T. Caraballo, F. Morillas and J. Valero, Pullback attractors of stochastic lattice dynamical systems with a multiplicative noise and non-Lipschitz nonlinearities, Differential and Difference Equations with Applications, Springer Proc. Math. Stat., Springer, New York, 47 (2013), 341-349. doi: 10.1007/978-1-4614-7333-6_27
    [5] On differential equations with delay in Banach spaces and attractors for retarded lattice dynamical systems. Discrete Contin. Dyn. Syst. (2014) 34: 51-77.
    [6] D. N. Cheban, Global Attractors of Non-Autonomous Dissipative Dynamical Systems: Interdisciplinary Mathematical Sciences, 1, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2004. doi: 10.1142/9789812563088
    [7] Global exponential stability of delayed Hopfield neural networks. Neural Netw. (2001) 14: 977-980.
    [8] Stability results for stochastic delayed recurrent neural networks with discrete and distributed delays. J. Differential Equations (2018) 264: 3864-3898.
    [9] A. Cichocki and R. Unbehauen, Neural Networks for Optimization and Signal Processing, Wiley, Chichester, 1993.
    [10] Attractors for random dynamical systems. Probab. Theory Related Fields (1994) 100: 365-393.
    [11] Random attractors for the D stochastic Navier-Stokes equation with multiplicative white noise. Stochast. Stochast. Rep. (1996) 59: 21-45.
    [12] Mean square global asymptotic stability of stochastic recurrent neural networks with distributed delays. Appl. Math. Comput. (2009) 215: 791-795.
    [13] Non-autonomous lattice systems with switching effects and delayed recovery. J. Differential Equations (2016) 261: 2986-3009.
    [14] S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice-Hall, Englewood Cliffs, 1994.
    [15] Exponential stability analysis of stochastic delayed cellular neural network. Chaos Solitons Fractals (2006) 27: 1006-1010.
    [16] Mean square exponential stability of stochastic recurrent neural networks with time-varying delays. Comput. Math. Appl. (2008) 56: 1773-1778.
    [17] Pullback attractors of nonautonomous semidynamical systems. Stoch. Dyn. (2003) 3: 101-112.
    [18] Stability analysis of stochastic functional differential equations with infinite delay and its application to recurrent neural networks. J. Comput. Appl. Math. (2010) 234: 407-417.
    [19] Exponential stability analysis for delayed semi-Markovian recurrent neural networks: A homogeneous polynomial approach. IEEE Trans. Neural Netw. Learn. Syst. (2018) 29: 6374-6384.
    [20] Robust dissipativity and passivity analysis for discrete-time stochastic neural networks with time-varying delay. Complexity (2015) 21: 47-58.
    [21] Exponential stability of hybrid stochastic recurrent neural networks with time-varying delays. Nonlinear Anal. Hybrid Syst. (2008) 2: 1198-1204.
    [22] Cellular neural networks with nonlinear and delay-type template elements and non-uniform grids. Int. J. Circuit Theory Appl. (1992) 20: 469-481.
    [23] Asymptotic stability of stochastic delayed recurrent neural networks with impulsive effects. J. Optim. Theory Appl. (2010) 147: 583-596.
    [24] B. Schmalfuss, Measure Attractors of the Stochastic Navier-Stokes Equation, Report 258, Universität Bremen, Fachbereiche Mathematik/Informatik, Elektrotechnik/Physik, Forschungsschwerpunkt Dynamische Systeme, Bremen, 1991.
    [25] th moment exponential stability of stochastic recurrent neural networks with time-varying delays. Nonlinear Anal. Real World Appl. (2007) 8: 1171-1185.
    [26] Stochastic stability of discrete-time uncertain recurrent neural networks with Markovian jumping and time-varying delays. Math. Comput. Modelling (2011) 54: 1979-1988.
    [27] Stochastic stability of uncertain recurrent neural networks with Markovian jumping parameters. Acta Math. Sci. Ser. B (2015) 35: 1122-1136.
    [28] Image compression by delayed CNNs. IEEE Trans. Circuits Syst. I (1998) 45: 205-215.
    [29] Global asymptotic stability of stochastic reaction-diffusion recurrent neural networks with Markovian jumping parameters and mixed delays. J. Anal. (2019) 27: 277-292.
    [30] Almost sure exponential stability of stochastic recurrent neural networks with time-varying delays. Internat. J. Bifur. Chaos Appl. Sci. Engrg. (2010) 20: 539-544.
    [31] B. Wang, Sufficient and necessary criteria for existence of pullback attractors for non-compact random dynamical systems, J. Differential Equations, 253 (2012), 1544-1583. doi: 10.1016/j.jde.2012.05.015
    [32] Finite lattice approximation of infinite lattice systems with delays and non-Lipschitz nonlinearities. Asymptot. Anal. (2018) 106: 169-203.
    [33] Pullback attractors for multi-valued non-compact random dynamical systems generated by reaction-diffusion equations on an unbounded domain. J. Differential Equations (2015) 259: 728-776.
    [34] J. Wang, Y. Wang and D. Zhao, Pullback attractors for multi-valued non-compact random dynamical systems generated by semi-linear degenerate parabolic equations with unbounded delays, Stoch. Dyn., 16 (2016), 1750001, 49 pp. doi: 10.1142/S0219493717500010
    [35] Robustness analysis for connection weight matrices of global exponential stability of stochastic delayed recurrent neural networks. Circuits Systems Signal Process (2014) 33: 2065-2083.
  • Reader Comments
  • © 2021 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2626) PDF downloads(286) Cited by(0)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog