|Home | About | Journals | Submit | Contact Us | Français|
The present study considers the robust stability for impulsive complex-valued neural networks (CVNNs) with discrete time delays. By applying the homeomorphic mapping theorem and some inequalities in a complex domain, some sufficient conditions are obtained to prove the existence and uniqueness of the equilibrium for the CVNNs. By constructing appropriate Lyapunov-Krasovskii functionals and employing the complex-valued matrix inequality skills, the study finds the conditions to guarantee its global robust stability. A numerical simulation illustrates the correctness of the proposed theoretical results.
Robustness is the ability of maintaining the performance of the controlling system under certain parameter perturbations. When these inner structural disturbances result in the instability of the system, additional control mechanisms should be used to improve these flimsy properties. When the control method is introduced, the uncertainty and error brought by the control itself also become another disturbance factor of the system. In real life and engineering practices, perturbations of system characteristics or parameters are often unavoidable. Perturbations exist for two main reasons. One is that the actual measurement is not accurate and usually deviates from its designed value. The other is slow drift of characteristics or parameters, which is influenced by the environmental factors in the running process of the system. When these uncertainties or random disturbances exist, the questions how and in what range to control the quality of the system or maintain its characteristics are of great importance. Therefore, robustness has become an important research topic in control theory, and it is also a basic problem that must be considered in the design of almost all kinds of control systems, such as image and signal processing, combinatorial optimization problems, pattern recognition, etc. It has attracted great attention of the scholars that work with neural networks [1–6].
For many applications of neural networks, on the one hand, the states change rapidly at a fixed time, and the duration of these abrupt changes is often neglected, assuming that they are caused by jumps. Such processes are studied by impulsive differential equations (for the relative theorems, we refer to ), and there are numerical applications of such equations in science and technology, mass services, etc. [8–11]. On the other hand, due to the neural processing and signal transmission, a time delay often occurs, which may cause instability and a poor performance of the system . Generally, delays may be caused by the measuring process and therefore the effect of time delay is common. Also many efforts are being made as regards the delay-dependent stability analysis of neural networks [13–23].
In the application of neural networks, complex signals are usually preferable [24–29], so it is necessary to analyze complex-valued neural networks (CVNNs), which deal with the complex-valued date, weight and neuron activation functions. However, much work mainly focuses on the boundedness, μ-stability, power-stability, exponent-stability, etc. [3, 18, 30–32], and little work considers the robust stability of neural networks with time delay and impulse in the complex domain. In [2, 6], the authors investigated a kind of recurrent CVNNs with time delays, but the activation functions are separated by real and imaginary parts and the analysis methods are also similar to those in their real domain. Therefore, the stability criteria cannot be applied if the activation functions cannot be expressed by separating their real and imaginary parts or if they are unbounded. Moreover, compared to real-valued neural networks, the advantage of CVNNs is that they can directly deal with two-dimensional data, which can also be processed by real-valued neural networks but then require double neurons. Consequently, as a class of complex-valued systems, CVNNs have undergone a growing number of studies that explore the application of neural networks. Therefore, the present study considers the robust stability of CVNNs with discrete time delay and impulse, which is valid regardless of whether the activation functions are separated or not. The relative results are extension of those in reference [2, 6]. Therefore, the robust stability for impulsive CVNNs with time delays is considered in this paper. Furthermore, compared with  and , the robust stability criteria for the addressed neural networks are valid regardless of whether the activation functions can be separated or not.
The structure of this paper is arranged as follows. Section 2 gives some preliminaries, including some notations and important lemmas, introducing the complex-valued recurrent networks model. The existence and uniqueness of the equilibrium are proved by using the homeomorphism mapping principle in Section 3. In Section 4, the global robust stability of the neural networks is investigated by building the proper Lyapunov functions. An example is given to illustrate the correction of our results. In Section 5, we conclude the paper.
Some notations of this paper are presented here firstly. i denotes the imaginary unit, i.e., . ℂn, ℝm×n, and ℂm×n represent the set of n-dimensional complex vectors, m × n real matrices, and complex matrices, respectively. The subscripts T and denote matrix transposition and matrix conjugate transposition, respectively. For complex vector z ∈ ℂn, let |z| = (|z1|,|z2|,…,|zn|)T be the module of the vector z and be the norm of the vector z. For complex matrix A = (aij)n×n ∈ ℂn×n, let |A| = (|aij|)n×n denote the module of the matrix A and denote the norm of the matrix A. I denotes the identity matrix with appropriate dimensions. The notation X ≥ Y (or X > Y) means that X − Y is positive semi-definite (or positive definite). In addition, λmax(P) and λmin(P) are defined as the largest and the smallest eigenvalue of positive definite matrix P, respectively.
Motivated by , we consider the following impulsive CVNN model with time delays:
where n is the number of neurons, zi(t) ∈ ℂ denotes the state of neuron i at time t, fj(t) is the complex-valued activation function, τj (j = 1, 2, …, n) are constant time delays and satisfy 0 ≤ τj ≤ ρ, ci ∈ ℝ with ci > 0 is the self-feedback connection weight, aij ∈ ℂ and bij ∈ ℂ are the connection weights, and Ji ∈ ℂ is the external input. Here Iik is a linear map, is the jump of zi at moments tk, and 0 < t1 < t2 < ⋯ is a strictly increasing sequence such that limk→∞tk = + ∞.
We rewrite (1) in the equivalent matrix-vector form
where z(t) = (z1(t),z2(t),…,zn(t))T ∈ ℂn, C = diag(c1, c2, …, cn), A = (aij)n×n ∈ ℂn×n, B = (bij)n×n ∈ ℂn×n, f(z(t)) = (f1(z1(t)),f2(z2(t)),…,fn(zn(t)))T, f(z(t − τ)) = (f1(z1(t−τ1)),f2(z2(t−τ2)),…,fn(zn(t−τn)))T, J = (J1,J2,…,Jn)T ∈ ℂn, Δz(tk) = (Δz1(tk),Δz2(tk),…,Δzn(tk))T, and .
or in the equivalent vector form
where φi( ⋅ ) is a complex-valued continuous function defined on [−ρ, 0] and φ(s) = (φ1(s),φ2(s),…,φn(s))T ∈ C([−ρ, 0], ℂn) with the norm .
The following assumptions will be needed in the study:
(H1) The parameters C = diag(c1, c2, …, cn), A = (aij)n×n, B = (bij)n×n, and J = (J1,J2,…,Jn)T in neural system (1) are assumed to be norm-bounded and satisfy
where , , , , , , , , , , with , , , , , and .
(H2) For i = 1, 2, …, n, the neuron activation function fi is continuous and satisfies
for any z1, z2 ∈ ℂ, where γi is a real constant. Furthermore, define Γ = diag(γ1, γ2, …, γn).
The neural network defined by (1) with the parameter ranges defined by (H1) is globally asymptotically robust stable if the unique equilibrium point of the neural system (1) is globally asymptotically stable for all C ∈ CI, A ∈ AI, B ∈ BI, and J ∈ JI.
For any a, b ∈ ℂn, if P ∈ ℂn×n is a positive definite Hermitian matrix, then a∗b + b∗a ≤ a∗Pa + b∗P−1b.
A given Hermitian matrix
where , , and , is equivalent to any of the following conditions:
If H(z):ℂn → ℂn is a continuous map and satisfies the following conditions:
then H(z) is a homeomorphism of ℂn onto itself.
Suppose A ∈ AI. Let R and S be real positive diagonal matrices. The function fi (i = 1, 2, …, n) satisfies (H2). Then, for any , the following inequalities hold:
where , , and f(z) = (f1(z1),f2(z2),…,fn(zn))T.
It should be noted that since A ∈ AI. Then we calculate directly that
Hence inequality (5) holds.
Next we prove inequality (6). Let S = diag(s1, s2, …, sn) and . Then . It is obvious that |Rz| = R|z| since R is a real positive diagonal matrix. From A ∈ AI, it follows that and for all i, j = 1, 2, …, n. Then and , which means . Hence . Noting that , by inequality (5), we infer
Therefore, inequality (6) holds.
In this section, we will give the sufficient conditions to prove the existence and uniqueness of equilibrium for system (1). An equilibrium solution of (1) is a constant complex vector , which satisfies
and , k = 1, 2, …, when ž is the impulsive jump.
Hence, proving the existence and uniqueness of solution (8) is equivalent to proving the existence of a unique zero point of map ℋ:ℂn → ℂn, where
We have the following theorem.
For the CVNN defined by (1), assume that the network parameters and the activation function satisfy assumptions (H1) and (H2), respectively. Then the neural network (1) has a unique equilibrium point for every input vector J = (J1,J2,…,Jn)T ∈ ℂn, if there exist two real positive diagonal matrices U and V such that the following linear matrix inequality (LMI) holds:
where , , , and .
We will use the homeomorphism mapping theorem on the complex domain to prove the theorem, that is, to show the map ℋ(z) is a homeomorphism of ℂn onto itself.
First, we prove that ℋ(z) is an injective map on ℂn. Let with , such that . Then
Multiplying both sides of (11) by , we obtain
Then taking the conjugate transpose of (12) leads to
Since V is a positive diagonal matrix, from assumption (H2) we get
An application of the Cauchy-Schwarz inequality yields
When z ≠ 0, we have
In this section, we will investigate the global robust stability of the unique equilibrium point for system (1). Firstly, the following assumption for the impulsive operators is needed: (H3) For i = 1, 2, …, n and k = 1, 2, …, Iik( ⋅ ) is such that
where δik ∈ [0, 2] is a real constant, and is the ith component of the equilibrium point . Then we have the following global robust stability theorem.
Suppose the conditions of Theorem 1 and (H3) hold. The equilibrium point of system (1) is globally robust stable, if there exist two real positive diagonal matrices P = diag(p1, p2, …, pn) and Q = diag(q1, q2, …, qn), such that the following linear matrix inequalities hold:
where , , , and .
where and . Meanwhile, the initial condition (4) can be transformed into
Consider the following Lyapunov-Krasovskii functional candidate:
When t = tk, k = 1, 2, …, it should be noted that . Then we compute
On the other hand, by the definition of V(t), we have
from which it can be concluded that the origin of (22), or equivalently the equilibrium point of system (1), is globally asymptotically robust stable by the standard Lyapunov theorem. The proof is completed.
In [9, 13], some dynamic characteristics, such as exponential stability and exponential anti-synchronization, were investigated for real-valued neural networks. Compared to [9, 13], the neural networks model in this paper is complex-valued, which can be viewed as an extension of real-valued neural networks.
In [33, 34], the criteria for the stability of CVNNs are expressed in terms of complex-valued LMIs. As pointed out in , complex-valued LMIs cannot be solved by the MATLAB LMI Toolbox straightforwardly. A feasible approach is to convert complex-valued LMIs to real-valued ones but this could double the dimension of the LMIs. In this paper, we express the stability criteria for CVNNs directly in terms of real-valued LMIs, which can be solved by the MATLAB LMI Toolbox straightforwardly.
In , the authors investigated the problem of global robust stability of recurrent CVNNs with time delays and uncertainties. In Theorem 3.4 of , to check robust stability of CVNNs, the boundedness of activation function fi is required. However, in this paper, the boundedness condition is removed. In Example 1, in the next section, the activation function fi is unbounded.
The following example demonstrates the effectiveness and superiority of our results.
Assume that the network parameters of system (1) are given as follows:
and the impulsive operator I( ⋅ ) satisfies assumption (H3).
Using the above matrices , , , and , we have
Thus, the conditions of Theorems 1 and 2 are satisfied, and system (1) has a unique equilibrium point which is globally asymptotically robust stable. To simulate the results, let us choose C, A, and B from the proper intervals above, and obtain the following specific system:
where f1(u) = f2(u) = 0.2(eu − 1), δ1k = 1 + ½sin(1 + k), δ2k = 1 + ⅔cos(2k3), k = 1, 2, …, and t1 = 0.5, tk = tk−1 + 0.2k, k = 2, 3, ….
In Figures 1--4,4, we see that the equilibrium point of system (34) is asymptotically stable, whether the delay τ1 = τ2 = 0.5 or τ1 = τ2 = 8. It should be noted that the criteria (10), (17), and (18) in Theorems 1 and 2 are independent from the delays τ. Therefore, in system (34), the delays have no influence on the stability of the equilibrium point.
In this paper, we have investigated the existence and uniqueness of the equilibrium as well as its robust stability for an impulsive CVNN with discrete time delays, by applying the homeomorphic mapping theorem and some important inequalities in the complex domain. We have presented some sufficient conditions to guarantee the existence of a unique equilibrium point for the CVNN. In addition, by constructing appropriate Lyapunov-Krasovskii functionals and employing complex-valued matrix inequalities, we also have obtained sufficient conditions to guarantee the robust stability of the CVNN. Finally, a numerical simulation has illustrated the correctness of the proposed theoretical results. Moreover, the conditions in Theorems 1 and 2 are irrelevant to the parameter τ, which illustrates that the parameter τ has no effect on the uniqueness and existence, neither on the robust stability of system (1). The figures in the article confirm this result.
This work is supported by the National Natural Science Foundation of China under Grants (NSFCs:11631012, 11401060), and the Program of Chongqing Innovation Team Project in University under Grant (CXTDX201601022).
The authors declare that they have no competing interests.
All the authors conceived the study, participated in its design and coordination and read and approved the final manuscript.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Yuanshun Tan, Email: nc.ude.utjqc@synat.
Sanyi Tang, Email: nc.ude.unns@gnatys.