Nonlinear Dyn (2019) 98:1877–1890

https://doi.org/10.1007/s11071-019-05293-y

ORIGINAL PAPER

Synchronization control of Markov jump neural networks

with mixed time-varying delay and parameter uncertain
based on sample point controller
Nuo Xu · Liankun Sun

Received: 11 June 2019 / Accepted: 7 October 2019 / Published online: 4 November 2019
© Springer Nature B.V. 2019

Abstract This paper put forward an improved syn- 1 Introduction

chronization problem for neural networks with Markov
jump parameters. The traditional Markov jump neu- The enlightenment stage of the neural network (NN) is
ral network (MJNN) only considers the basic external in the middle and late 1980s. After decades of develop-
time-varying delays, ignoring both the distributed and ment, it gradually develops to a mature stage and has
leakage delays in the internal transmission of the neu- been extended to all areas of real life. This special non-
ral network and the small time-varying errors in the linear network which imitates the structure of human
mode switching of Markov probability transition rates. brain and the method of processing information have
In this paper, we focus on the synchronization of MJNN made amazing achievements in many aspects and can
with mixed time-varying delay. And an improved solve many problems that are difficult to solve by dig-
Lyapunov–Krasovskii functional is constructed. The ital computers. In the past few decades, different kinds
convergence of inequalities is solved by using affine of NN have attracted attention. However, the simula-
Bessel–Legendre inequalities and Wirtinger double tion of human brain structure by artificial NN is still
integral inequalities. At the same time, a new method a low degree of research. Scholars have been looking
is used to optimize the mathematical geometric area for more accurate theories to imitate brain intelligence.
of the time-varying delay and reduce the conservative- Through continuous practice and theoretical research,
ness of the system. Finally, a sample point controller is people have found that chaos and time delay have been
constructed to synchronize the driving system and the found in the nervous system, whether micro-neurons
corresponding system. or macro-brain waves. Therefore, researchers focus
on chaotic NN with time-varying delay [1,10,12,33].
Keywords Markov jump parameters · Mixed time- Chaotic NN is to apply the advantages of NN system
varying delay · Time-varying transition rates · Neural to chaotic system, make up for each other’s shortcom-
network ings, and there is also a very large possibility of intel-
ligent information processing. However, in traditional
NN, only external time-varying delay is considered,
and the distributed delay and leakage delay of informa-
tion transmission within neurons are neglected. There-
fore, the first problem that this paper focuses on is the
N. Xu · L. Sun (B)
School of Computer Science and Technology, Tianjin
chaotic perturbation of NN with mixed time-varying
Polytechnic University, Tianjin 300387, China delay.
e-mail: [email protected]

123
1878 N. Xu, L. Sun

In recent years, the stability of NN with Markov vex inequality [34] and Bessel–Legendre inequali-
jump has become a research hot spot. This model allows ties [21]. These methods have effectively improved
NN to have multiple modes, and different modes can be the convergence accuracy, but there is still room
switched under the drive of a Markov chain. Therefore, for improvement. Wirtinger double integral inequali-
the study of the stability of Markov jump model has ties and affine Bessel–Legendre inequalities improve
more potential application value [13,14,18,23,25,26, Wirtinger integral inequality and Bessel–Legendre
29,30]. In [25,29], by constructing suitable LKF and inequality, respectively. Therefore, the fourth problem
using linear matrix inequality (LMI), the mean-square that this paper focuses on is how to use Wirtinger dou-
global exponential stability of a class of reaction- ble integral inequalities and affine Bessel–Legendre
diffusion Hopfield MJNN and the global robust expo- inequalities to improve the convergence accuracy.
nential stability of a class of time-varying delay MJNN In addition, when discussing the interval range of
are studied, respectively. However, the traditional prob- time-varying delays, the defaults are h 1 ≤ h ≤ h 2
abilistic transfer matrix of Markov jump parameters and d1 ≤ ḣ ≤ d2 , which are conservative and can
often neglects the small time-varying errors in proba- be optimized in two-dimensional space. Therefore, the
bility transition rates, which may make the switching fifth problem that this paper focuses on is to discuss the
process unstable and cause the system to collapse in optimization of time-varying delay intervals based on
severe cases. Therefore, the second problem that this two-dimensional level.
paper focuses on is the time-varying probabilistic trans- In summary, the contributions of this paper and the
fer parameters in MJNN. difficulties to be solved are as follows: Firstly, how to
Synchronization, as a nonlinear phenomenon, has unify the mixed time-varying delay and time-varying
appeared in many practical problems, such as physics, probability transfer under one MJNN. Secondly, how to
ecology and physiology. Therefore, the application of apply Wirtinger double integral inequalities and affine
synchronization theory has been widely studied in dif- Bessel–Legendre inequalities to Lyapunov functional
ferent scientific fields. In particular since 1990s, Pec- processing. Thirdly, how to synchronize MJNN-DS
ora and Carroll have paid attention to the importance of and MJNN-RS through the control of sample point con-
control and synchronization of chaotic systems. They troller. Fourthly, how to optimize the two-dimensional
put forward the concept of drive-response to achieve geometric area of time delay. In addition, these meth-
synchronization of chaotic systems. This method con- ods have the following advantages: the affine Bessel–
trols the response system by driving the external input Legendre inequalities improves the traditional Bessel–
of the system to achieve synchronization. So the the- Legendre inequality, and with the increase in N , the
ory of chaos synchronization and chaos control has optimization effect will be better. Compared with the
been widely studied. In order to achieve synchroniza- traditional state feedback controller, the sample point
tion, many control systems have been proposed, such controller can better transmit the effective information
as: synchronization method of driving-response [19]; of the system and achieve better control effect. The tra-
active–passive synchronization method [20]; synchro- ditional two-dimensional geometric area of time delay
nization method based on mutual coupling [36]; adap- is a rectangle. We reduce the conservativeness of the
tive synchronization method [9]; feedback control syn- system by reducing the area to a parallelogram.
chronization method [15]; projection synchronization Next, this paper will be based on the following
control [11]; and impulse control [7]. Therefore, the four parts. The first part introduces MJNN-DS and
third problem that this paper focuses on is how to con- MJNN-RS, sample point controller, and relevant use-
struct a suitable sample point controller to synchronize ful lemmas. In the second part, the synchronous anal-
MJNN drive system (MJNN-DS) and MJNN response ysis of MJNN mixed-time-varying-delayed error sys-
system (MJNN-RS). tem is carried out, and the convergence accuracy of
On the other hand, the synchronous analysis of LKF is improved by using Wirtinger double integral
MJNN usually constructs a suitable LKF and then inequalities and affine Bessel–Legendre inequalities.
converges the inequality. In recent years, scholars In the third part, the range of time-varying delay in
have proposed many useful inequality methods, such two-dimensional space is discussed, and the conserva-
as: Jensen inequality [37], Wirtinger integral inequal- tiveness of the system is reduced by reducing the two-
ity [22], free matrix inequality [32], interactive con- dimensional geometric area. In the fourth part, a numer-

123
Synchronization control of Markov jump neural networks 1879

ical example is constructed. The parameters of the sam- Remark 1 The first item on the right side of the equa-
ple point controller, the chaotic curve of MJNN system, tion is the stable negative feedback of the system, which
Markov jump response curve, synchronization analy- is often referred to as the “leakage” item. Since the self-
sis response curve and error analysis response curve attenuation process of neurons is not instantaneous,
are obtained through actual simulation. when the neurons are cut off from the neural network
In this paper, “0” represents zero matrix of suitable and external inputs, it takes time to reset to the isolated
dimension. Rn and Rn×m represent n-dimensional and static state. In order to describe this phenomenon, it is
n ×m-dimensional Euclidean spaces, respectively. “T” necessary to introduce a “leakage” delay. In this paper,
represents the matrix transposition. {Ω, , P} repre- σ is called leakage delay.
sents the probability space.
Consider the following MJNN-RS with mixed time-
varying delay:
2 Preliminaries
ẏ(t) = −C(r (t))y(t − σ ) + A(r (t)) f (x(t))
Consider the following MJNN-DS with mixed time- +B(r (t)) f (y(t − d1 (t)))
 t
varying delay:
+D(r (t)) f (y(s))ds + u(t) + J (2)
ẋ(t) = −C(r (t))x(t − σ ) + A(r (t)) f (x(t)) t−d2 (t)

+B(r (t)) f (x(t − d1 (t))) where y(t) = (y1 (t), y2 (t), · · · , yn (t))T ∈ Rn is the
 t
neuron state vector. The meanings of other symbols are
+D(r (t)) f (x(s))ds + J (1)
t−d2 (t) equivalent to MJNN driving system (1). u(t) represents
the sample point controller, which is defined as follows:
where x(t) = (x1 (t), x2 (t), · · · , xn (t))T ∈ Rn is the
neuron state vector. A(·), B(·), C(·) and D(·) are matri- u(t) = K (r (tk ))e(tk ), tk ≤ t < tk+1
ces of suitable dimensions with uncertainties, which are
expressed as follows: where K (·) is the feedback gain matrix of the sample
point controller, e(tk ) represents the discrete control
A(·) = Ā(·) + ΔA B(·) = B̄(·) + ΔB function, and tk is the sample point and satisfies:
C(·) = C̄(·) + ΔC D(·) = D̄(·) + ΔD
0 = t0 < t1 < · · · < tk < · · · < lim tk = +∞
where ΔA, ΔB, ΔC and ΔD are uncertain parameter k→+∞

terms, such as: Assuming that the period of sample points is

[ΔA, ΔB, ΔC, ΔD] = G F(t)[E 1 , E 2 , E 3 , E 4 ] bounded, for any k ≥ 0, there exists a normal quan-
tity d3 satisfying tk+1 − tk ≤ d3 .
where G and E i (i = 1, 2, 3) are real matrices of suit-
able dimensions, F(t) satisfies: F T (t)F(t) ≤ I . Remark 2 Obviously, due to the introduction of the dis-
f (·) is the neuron excitation function. J denotes crete term e(tk ), the synchronization analysis of the
external disturbances. r (t) represents a Markov jump system becomes more difficult. In this paper, the input
subset on a finite state space S = {1, · · ·, M}. Markov delay method is used to deal with the discrete term.
chain is defined in space {Ω, , P}. The transfer rate Define a smooth function:
matrix Π (t) = (μi j ) N ×N is defined as follows:
d3 (t) = t − tk , tk ≤ t ≤ tk+1
P{r (t + Δ) = j|r (t) = i}

μi j Δ + o(Δ); i f j = i, Easy to get: 0 ≤ d3 (t) ≤ d3 . In summary, the sample
=
1 + μii Δ + o(Δ); i f j = i point controller is converted as follows:

where μi j ≥ 0, if j = i, μii = − Nj=1, j=i μi j . σ , u(t)=K (r (tk ))e(tk ) ⇒ u(t)=K (r (tk ))e(t − d3 (t))
d1 (t) and d2 (t) represent the leakage delay, the external
time-varying delay and the distributed delay, respec- Let e(t) = y(t)−x(t), g(e(·)) = f (y(·))− f (x(·)).
tively, and the time-varying delay ranges are as follows: The MJNN error system with mixed time-varying delay
0 ≤ d1 (t) ≤ d1 , h 1 ≤ ḋ1 (t) ≤ h 2 , 0 ≤ d2 (t) ≤ d2 . is defined as follows:

123
1880 N. Xu, L. Sun

 n  n
ė(t) = −C(r (t))e(t − σ ) + A(r (t))g(e(t)) (n − m)2 x T (u)Hx(u)dudθ
+ B(r (t))g(e(t − d1 (t))) m θ
 t  2Θd1
T
HΘd1 + 4Θd2
T
H Θd2
+ D(r (t)) g(e(s))ds + u(t) (3)
t−d2 (t) where
Some lemmas are given below, which play a key role ⎧ n n
in the calculation of this paper. ⎨ Θd1 = m θ x(u)dudθ
n n
Θd2 = − m θ x(u)duds
⎩ nnn
Lemma 1 (Affine Bessel–Legendre inequalities)[8] If + s−r m θ u x(u)dvduds
3

the function x(·) satisfies x(·) : [a, b] → Rn and N ∈

N, given any positive definite matrix R = R T , there
Remark 5 Lemma 2 adds a multiple integral on the
exists a matrix X such that the following relation holds
 b basis of Wirtinger integral inequality [22]. At the same
− ẋ T (s)R ẋ(s)ds ≤ −ϑ NT (t)Ω(X )ϑ N (t) time, Θd1 and Θd2 on the right side of the inequal-
a ity contain more sub-terms. Therefore, Lemma 2 can
where express the internal information of the system more
Ω(X ) = X L N + L TN X T − (b − a)X R̄ X T completely in the derivative deformation of Lyapunov
L N = [Γ NT (0) Γ NT (1) . . . Γ NT (N )]T functional, so it has lower conservativeness.
 
1 1
R̄ = diag R −1 , R −1 , . . . , R −1 Lemma 3 [3]When has the M − Π (t) transfer ratio
3 2N + 1
⎧ matrix is located has the border area D apex, territory
⎨ x (b) x (a) i f N = 0
T T T
ϑN = T D1 by the following expression is composed:
⎩ x T (b) x T (a) 1 Φ T . . . 1 Φ T if N > 0
b−a 0 b−a N −1
 D1 = {Π (r (t))|Π (r (t))
[I − I ] i f N = 0
T

Γ N (k) = T 
M 
M
I (−1)k+1 I γ N0 k I . . . γ NNk−1 I i f N > 0 (l)
= rl (t)Π , rl (t) = 1, rl (t) ≥ 0 (4)
 b
l=1 l=1
Φk = L k (s)x(s)ds
a

−(2i + 1)(1 − (−1)k+i ) i f i ≤ k where Π (l) (l = 1, 2, · · ·, M) are vertices, r (t) is the
γ Ni k =
0 if i ≥ k +1 parameter vector, it is assumed that the changes are
k 
   
 u−a l
known. As a result, ṙ (t) is as follows:
k k +l
L k (u) = (−1)k (−1)l  
D2 = −vl ≤ṙl (t)≤vl , vl ≥0, l = 1, 2, · · ·, M − 1
l l b−a
l=0

(5)
Remark 3 Unlike the traditional Bessel–Legendre M
inequalities [21], the right side of the inequality of Remark 6 Easy to get l=1 rl (t) = 1 is equivalent to
 M−1
Lemma 1 is the affine of the length of the integral ṙ l (t) + ṙ M (t) = 0. So ṙ M (t) is expressed by
l=1  M−1
interval, so it can be easily dealt with by convexity. | ṙ M (t) |≤ l=1 vl .
In addition, Lemma 1 can be transformed into existing
inequalities in literature under special conditions, such Lemma 4 [4] If the vector function x satisfies x :
as affine Jensen inequality [2] and affine Wirtinger inte- [0, ] → Rn , given any positive definite matrix U and
gral inequality [6], which shows that the inequality of positive scalar , the following relation holds
Lemma 1 is more general.  T  
−1
x(s)ds U x(s)ds
Remark 4 Lemma 1 has an additional decision vari-
able of (N + 1)(N + 2)n 2 because of the addition of  0 0

additional matrix X to the traditional Bessel–Legendre ≤ x T (s)U x(s)ds

0
inequalities [21].
Lemma 5 [28] For any real matrices D, E, F and
Lemma 2 (Wirtinger Double Integral inequalities)[17] scalar ε > 0, when F T F ≤ I is satisfied, the following
If constants m and n satisfy m < n, for any positive def- inequalities hold
inite matrix H, and x ∈ [m, n] → Rn , the following
inequalities hold D F E + E T F T D T ≤ ε D D T + ε−1 E T E

123
Synchronization control of Markov jump neural networks 1881

⎡ 2  2 ⎤
Assumption (A1) The neuron excitation function 2 d32
T ⎣ d1
f (·) satisfies the following conditions: + e21 Z1 + Z 2 ⎦ e21
2 2
f i (u i − vi ) 
0< ≤ li (i = 1, 2 · · · , N ) d1
u i − vi − [d1 e1 − e16 ]T Z 1 [d1 e1 − e16 ] − 2 − e1 − e16
2
where u i and vi are arbitrary real numbers, and u = v. T  
li are known constants. 3 d1 3
+ e17 Z 1 − e1 − e16 + e17
d1 2 d1

d3
− [d3 e1 − e18 ] Z 2 [d3 e1 − e18 ] − 2 − e1
T
3 Main results 2
T  
3 d3 3
Theorem 1 Given scalars di > 0, i = 1, 2, 3 and − e18 + e19 Z 2 − e1 − e18 + e19
d3 2 d3
ḋ1 (t), and satisfy, di (t) ∈ [0, di ], i = 1, 2, 3, ḋ1 (t) ∈
[h 1 , h 2 ], for any delay d(t), MJNN-DS (1) and MJNN- + d2 e1T Se1 − δ1 e23
T
e23 − e1T L T Le1
RS (2) achieve complete synchronization, if there exist
(l) − δ2 e24
T
e24 − e2T L T Le2 − δ3 d2−1 e22
T
e22 + Φ3
symmetry matrices PP > 0 ∈ R7n , Q i > 0, i =
1, 2, 3, 4 ∈ Rn , Ri > 0, i = 1, 2 ∈ Rn , Z i > (ls)
Ω P (di (t), ḋ1 (t)) (8)
0, i = 1, 2 ∈ Rn , S > 0 ∈ Rn , any matrices
X i , i = 1, 2, 3, 4 ∈ R4n×3n , M1 , M2 and χi are matri-
ces of suitable dimensions, such that the following hold:

(ls) (sl)
Υ P (di (t), ḋ1 (t)) + Υ P (di (t), ḋ1 (t)) < 0 (6)
(ls)
Υ P (di (t), ḋ1 (t))
⎡ ⎤
Ψ P (di (t), ḋ1 (t)) Π1T X 1 Π2T X 2 Π3T X 3 Π4T X 4 Φ1 Φ2
⎢ ∗ Δ22 0 0 0 0 0 ⎥
⎢ ⎥
⎢ ∗ ∗ Δ 0 0 0 0 ⎥
⎢ 33 ⎥
=⎢
⎢ ∗ ∗ ∗ Δ44 0 0 0 ⎥ ⎥ (7)
⎢ ∗ ∗ ∗ ∗ Δ55 0 0 ⎥
⎢ ⎥
⎣ ∗ ∗ ∗ ∗ ∗ −ε I 0 ⎦
∗ ∗ ∗ ∗ ∗ ∗ − 1ε I

where = H e{T0T (ḋ1 (t))PP(l) T1 (d1 (t))} + T1T (d1 (t))

Δ22 = −d1 (t)R1N Δ33 = −(d1 − d1 (t))R1N 
N
(l) (s)
μi j P j T1 (d1 (t))
Δ44 = −d3 (t)R2N Δ55 = −(d3 − d3 (t))R2N
j=1
Ri N = diag{Ri , 3Ri , 5Ri } i = 1, 2

M−1
Ψ P (di (t), ḋ1 (t)) + T1T (d1 (t)) ±(PP(n) − PP(M) )T1 (d1 (t)) (9)
= Ω P(ls) (di (t), ḋ1 (t)) n=1

 4
Φ1 = col{M1 G 0 ·!"
· · 0# M2 G 0 0 0} (10)
+ e1T Q i e1 − (1 − ḋ1 (t))e2T Q 1 e2
19n
i=1
Φ2 = col{0 ·!"
· · 0# −E 3T 0 E 4T E 1T E 2T } (11)
− e20T
Q 2 e20 − e3T Q 3 e3
19n
− e11T
Q 4 e11 + e21 T
(d1 R1 + d3 R2 )e21 ⎡ ⎤
In −In 0n 0n
− Π1 (X 1 M + M X 1 )Π1 − Π2T (X 2 M
T T T
M = ⎣ In In −2In 0n ⎦ (12)
+ M T X 2T )Π2 − Π3T (X 3 M + M T X 3T )Π3 In −In 6In −12In
− Π4T (X 4 M + M T X 4T )Π4 Π1 = col{e1 e2 e6 e8 }

123
1882 N. Xu, L. Sun

 0  t
Π2 = col{e2 e3 e7 e9 } 1 1
e(s)dsdθ
Π3 = col{e1 e10 e12 e14 } d1 (t) −d1 (t) t+θ d1 − d1 (t)
 −d1 (t)  t−d1 (t) 
Π4 = col{e10 e11 e13 e15 } (13)
e(s)dsdθ
−d1 t+θ
⎡ ⎤
0 ·!"
· · 0# χi 0 ·!"
· · 0# −ε1 M2 C̄ P −ε1 M2 ε1 M2 D̄ P ε1 M2 Ā P ε1 M2 B̄ P
⎢ 9n ⎥
⎢ 9n ⎥
⎢0 · · · 0 0 0 ·!"
· · 0# 0 0 0 0 0 ⎥
⎢ !" # ⎥
⎢ 9n ⎥
⎢ 9n ⎥
⎢ .. .. .. ⎥
⎢ . . . ··· ··· ··· ··· ··· ⎥
⎢ ⎥
⎢0 · · · 0 χi 0 ·!"
· · 0# −M2 C̄ P −M2 − M2T M2 D̄ P M2 Ā P M2 B̄ P ⎥
⎢ !" # ⎥
Φ3 = ⎢ ⎥ (14)
⎢ 9n 9n ⎥
⎢0 · · · 0 0 0 ·!"
· · 0# 0 0 0 0 0 ⎥
⎢ !" # ⎥
⎢ 9n ⎥
⎢ 9n ⎥
⎢0 · · · 0 0 0 ·!"
· · 0# 0 0 0 0 0 ⎥
⎢ !" # ⎥
⎢ 9n ⎥
⎢ 9n ⎥
⎣0 · · · 0 0 0 ·!"
· · 0# 0 0 0 0 0 ⎦
!" #
9n 9n

ei (i = 1, · · · , 24) ∈ Rn×24n are identity matrices. By deriving V (x(t), t, r (t)), meanwhile, Lemma 1
Sample point controller parameters can be obtained: and Lemma 2 are used for V3 (t) and V4 (t) respectively,
K i = M2−1 χi . we get the following results
V̇1 (t) = H e{T0T (ḋ1 (t))PP T1 (d1 (t))} + T1T (d1 (t))
Proof An improved LKFs are defined: V (x(t), t, r (t))

N
= V1 (t) + V2 (t) + V3 (t) + V4 (t) + V5 (t). where μi j (r (t))P j (r (t))T1 (d1 (t))
j=1
V1 (t) = η1T (t)P(r (t))η1 (t)  
 t  t d PP (r (t))
+ T1T (d1 (t)) T1 (d1 (t))
V2 (t) = e (s)Q 1 e(s)ds+
T
eT (s)Q 2 e(s)ds dt
t−d1 (t) t−σ 
4
 t  t V̇2 (t) = e1T Q i e1 − (1 − ḋ1 (t))e2T Q 1 e2
+ eT (s)Q 3 e(s)ds+ eT (s)Q 4 e(s)ds i=1
t−d1 t−d3
 0  t − e20
T
− e3T Q 3 e3 − e11
Q 2 e20 T
Q 4 e11
 t
V3 (t) = ėT (s)R1 ė(s)dsdθ V̇3 (t) = e21 (d1 R1 + d3 R2 )e21 −
T
ėT (s)R1 ė(s)ds
−d1 t+θ
 0  t  t
t−d1

+ ėT (s)R2 ė(s)dsdθ − ėT (s)R2 ė(s)ds

−d3 t+θ t−d3
  t t ≤ e21
T
(d1 R1 + d3 R2 )e21 − Π1T (X 1 M + M T X 1T
d12 t
V4 (t) = ėT (v)Z 1 ė(v)dvdudθ −1 T
− d1 (t)X 1 R1N X 1 )Π1
2 t−d1 θ u
  t t −1 T
− Π2T (X 2 M + M T X 2T − (d1 − d1 (t))X 2 R1N X 2 )Π2
d2 t
+ 3 ėT (v)Z 2 ė(v)dvdudθ −1 T
− Π3T (X 3 M + M T X 3T − d3 (t)X 3 R2N X 3 )Π3
2 t−d3 θ u
 0  t − Π4T (X 4 M + M T X 4T
V5 (t) = e (s)Se(s)dsdθ
T −1 T
− (d3 − d3 (t))X 4 R2N X 4 )Π4
−d2 t+θ ⎡   2 ⎤
2 2
T ⎣ d1 d32
where V̇4 (t) ≤ e21 Z1 + Z2⎦
2 2
η1 (t) = col { e(t) e(t − d1 (t)) e(t − d1 )
 t  t−d1 (t) e21 − [d1 e1 − e16 ]T Z 1 [d1 e1 − e16 ]

e(s)ds e(s)ds d1
− 2 − e1
t−d1 (t) t−d1 2

123
Synchronization control of Markov jump neural networks 1883

T    t T
3 d1 3
−e16 +
d1
e17 Z 1 − e1 − e16 + e17
2 d1 ≤ −δ3 d2−1 g(e(s))ds
t−d2 (t)
− [d3 e1 − e18 ] Z 2 [d3 e1 − e18 ]
T  t 
 T g(e(s))ds (18)
d3 3
− 2 − e1 − e18 + e19 Z 2 t−d2 (t)
2 d3
 
d3 3
− e1 − e18 + e19 Given any constant M1 and M2 , the following equa-
2 d3
 t tion holds
V̇5 (t) = d2 e1T Se1 − eT (s)Se(s)ds
t−d2 (t)
0 = 2[eT (t)M1 + ėT (t)M2 ]
where [−ė(t) − C(r (t))e(t − σ ) + A(r (t))g(e(t))
T0 (ḋ1 (t)) +B(r (t))g(e(t − d1 (t))) + D(r (t))
  t 
= col e21 (1 − ḋ1 (t))e4 e5 e1 − (1 − ḋ1 (t))e2 g(e(s))ds + K (r (t))e(t − d3 (t))
(1 − ḋ1 (t))e2 − e3 e1 − (1 − ḋ1 (t))e6 t−d2 (t)
 (19)
− ḋ1 (t)e8 (1 − ḋ1 (t))e2 − e7 + ḋ1 (t)e9
T1 (d1 (t)) Add (15)–(19) to V̇1 (t)-V̇5 (t), and then deal with the
= col{ e1 e2 e3 d1 (t)e6 (d1 − d1 (t))e7 d1 (t)e8 items. Separating the definite items from the uncertain
d1 − d1 (t))e9 } items in A(·), B(·), C(·) and D(·), the following results
can be obtained:
And the definitions of Φ1 , Φ2 , M, Π1 , Π2 , Π3 and
⎡ ⎤
Π4 are shown in (10)–(13). 0 ·!"
· · 0# −M1 ΔC 0 M1 ΔD M1 ΔA M1 ΔB
⎢ 19n ⎥
⎢ ⎥
The following inequalities are defined according to ⎢0 · · · 0 0 0 0 0 0 ⎥
⎢ !" # ⎥
Assumption (A1) ⎢ 19n ⎥
⎢ ⎥
g T (e(t))g(e(t)) − eT (t)L T Le(t) ≤ 0 ⎢ .. ⎥
⎢ . ··· ··· ··· ··· ··· ⎥
⎢ ⎥
g T (e(t − d1 (t)))g(e(t − d1 (t))) ⎢0 · · · 0 −M2 ΔC 0 M2 ΔD M2 ΔA M2 ΔB ⎥
⎢ ⎥
Φ̄3 = ⎢ !" # ⎥
−eT (t − d1 (t))L T Le(t − d1 (t)) ≤ 0 ⎢ 19n ⎥
 ⎢0 · · · 0 0 0 0 0 0 ⎥
t ⎢ !" # ⎥
g T (e(s))g(e(s))ds ⎢ 19n ⎥
⎢ ⎥
t−d2 (t) ⎢0 · · · 0 0 0 0 0 0 ⎥
 t ⎢ !" # ⎥
⎢ 19n ⎥
− e (s)L Le(s)ds ≤ 0
T T ⎢ ⎥
⎣0 · · · 0 0 0 0 0 0 ⎦
t−d2 (t) !" #
where L = diag{l1 , l2 , · · · , ln }. Meanwhile, given any 19n

positive constant: δ1 , δ2 and δ3 , the following inequal-

ities can be obtained where Φ̄3 is a matrix consisting of uncertain terms and
Φ3 is a matrix consisting of deterministic terms, as
− δ1 [g T (e(t))g(e(t)) − eT (t)L T Le(t)] ≥ 0 (15) shown in (14). Next, lemma 5 is used for matrix Φ̄3 ,
− δ2 [g (e(t − d1 (t)))g(e(t − d1 (t)))
T
which can be obtained as follows:
−eT (t − d1 (t))L T Le(t − d1 (t))] ≥ 0 (16) ⎡ ⎤ ⎡ ⎤T ⎡ ⎤ ⎡ ⎤T
 t M1 G 0 0 M1 G
− δ3 [ g T (e(s))g(e(s))ds ⎢ 0 ⎥ ⎢ .. ⎥ ⎢ .. ⎥ ⎢ 0 ⎥
t−d2 (t) ⎢ ⎥ ⎢ . ⎥ ⎢ . ⎥ ⎢ ⎥
 ⎢ .. ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ .. ⎥
t ⎢ . ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ . ⎥
− eT (s)L T Le(s)ds] ≥ 0 ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
(17) ⎢ 0 ⎥ ⎢ T⎥ ⎢ T⎥ ⎢ ⎥
t−d2 (t) ⎢ ⎥ F(t) ⎢−E 3 ⎥ + ⎢−E 3 ⎥ F T (t) ⎢ 0 ⎥
⎢ M2 G ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ M2 G ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
From Lemma 4, the following can be obtained ⎢ 0 ⎥ ⎢ ET ⎥ ⎢ ET ⎥ ⎢ 0 ⎥
 t ⎢ ⎥ ⎢ 4 ⎥ ⎢ 4 ⎥ ⎢ ⎥
⎣ 0 ⎦ ⎣ ET ⎦ ⎣ ET ⎦ ⎣ 0 ⎦
−δ3 g T (e(s))g(e(s))ds 1 1
t−d2 (t) 0 E 2T E 2T 0

123
1884 N. Xu, L. Sun

⎡ ⎤⎡ ⎤T ⎡ ⎤⎡ ⎤T 
M1 G M1 G 0 0 3
+ e19 + d2 e1T Se1
⎢ 0 ⎥⎢ 0 ⎥ ⎢ .. ⎥ ⎢ .. ⎥ d3
⎢ ⎥⎢ ⎥ ⎢ . ⎥⎢ . ⎥
⎢ .. ⎥ ⎢ .. ⎥ ⎢ ⎥⎢ ⎥ −δ1 [e23
T
e23 − e1T L T Le1 ] − δ2 [e24
T
e24 − e2T L T Le2 ]
⎢ . ⎥⎢ . ⎥ ⎢ 0 ⎥⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ T⎥ ⎢ T⎥
≤ ε−1 ⎢ 0 ⎥ ⎢ 0 ⎥ + ε ⎢−E 3 ⎥ ⎢−E 3 ⎥ −δ3 d2−1 e22
T
e22 + Φ3 (23)
⎢ M2 G ⎥ ⎢ M2 G ⎥ ⎢ 0 ⎥⎢ 0 ⎥ Ω(di (t), ḋ1 (t))
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ 0 ⎥⎢ 0 ⎥ ⎢ ET ⎥ ⎢ ET ⎥
⎢ ⎥⎢ ⎥ ⎢ 4 ⎥⎢ 4 ⎥ = H e{T0T (ḋ1 (t))PP (r (t))T1 (d1 (t))}
⎣ 0 ⎦⎣ 0 ⎦ ⎣ ET ⎦ ⎣ ET ⎦
1 1
0 0 E 2T E 2T 
N
+T1T (d1 (t)) μi j P j (r (t))T1 (d1 (t))
(20) j=1
 
For convenience, let M1 = ε1 M2 and χi = M2 K i , d PP (r (t))
+T1T (d1 (t)) T1 (d1 (t)) (24)
ε1 is an arbitrary real number. To sum up, combined dt
with (20), we can get:
ξ(t)
V̇ (x(t), t, r (t)) ≤ ξ T Υ (di (t), ḋ1 (t))ξ(t) (21)
= col { e(t) e(t − d1 (t)) e(t − d1 )
where ė(t − d1 (t)) ė(t − d1 )
Υ (di (t), ḋ1 (t))
 t
1 1
⎡ ⎤ e(s)ds
Ψ (di (t), ḋ1 (t)) Π1T X 1 Π2T X 2 Π3T X 3 Π4T X 4 Φ1 Φ2 d1 (t) t−d1 (t) d1 − d1 (t)
⎢ ∗ Δ22 0 0 0 0 0 ⎥  t−d1 (t)  0  t
⎢ ⎥
⎢ ∗ ∗ Δ 0 0 0 0 ⎥ 1
⎢ 33 ⎥ e(s)ds e(s)dsdθ
=⎢ ⎢ ∗ ∗ ∗ Δ 44 0 0 0 ⎥
⎥ d12 (t) −d1 (t) t+θ
⎢ ⎥ t−d1
⎢ ∗ ∗ ∗ ∗ Δ 55 0 0 ⎥  −d1 (t)
⎣ ∗ ∗ ∗ ∗ ∗ −ε I 0 ⎦ 1
∗ ∗ ∗ ∗ ∗ ∗ −ε I
1
(d1 − d1 (t))2 −d1
(22)  t−d1 (t)
e(s)dsdθ e(t − d3 (t)) e(t − d3 )
Ψ (di (t), ḋ1 (t)) t+θ
 t  t−d3 (t)
1 1

4
e(s)ds e(s)ds
= Ω(di (t), ḋ1 (t)) + e1T Q i e1 d3 (t) t−d3 (t) d3 − d3 (t) t−d3
i=1 1
−(1 − ḋ1 (t))e2T Q 1 e2 d3 (t)
2
 0  t
−e20
T
Q 2 e20 − e3T Q 3 e3
e(s)dsdθ
−e11
T
Q 4 e11 + e21
T
(d1 R1 + d3 R2 )e21 −d3 (t) t+θ
 −d3 (t)  t−d3 (t)
−Π1T (X 1 M + M T X 1T )Π1 − Π2T (X 2 M 1
e(s)dsdθ
+M T X 2T )Π2 − Π3T (X 3 M + M T X 3T )Π3 (d3 − d3 (t))2 −d3 t+θ
 t  t  t
−Π4T (X 4 M + M T X 4T )Π4
⎡   2 ⎤ e(θ )dθ e(s)dsdθ
2 2 d32
t−d1 t−d1 θ
T ⎣ d  
+e21 1
Z1 + Z 2 ⎦ e21 t t
2 2 e(θ )dθ
t−d3 t−d3

−[d1 e1 − e16 ]T Z 1 [d1 e1 − e16 ] t
 T e(s)dsdθ e(t − σ ) ė(t)
d1 3 θ
−2 − e1 − e16 + e17  t
2 d1
  g(e(s))ds g(e(t))
d1 3 t−d2 (t)
Z 1 − e1 − e16 + e17
2 d1 g(e(t − d1 (t)))} (25)
−[d3 e1 − e18 ] Z 2 [d3 e1 − e18 ]
T
Therefore, as long as satisfying (22) is negative
 T 
d3 3 d3 definite, then V̇ (x(t), t, r (t)) is strictly negative def-
−2 − e1 − e18 + e19 Z 2 − e1 − e18 inite in the interval d1 (t) ∈ [0, d1 ], ḋ1 (t) ∈ [h 1 , h 2 ].
2 d3 2

123
Synchronization control of Markov jump neural networks 1885

According to Lyapunov stability theory, under the con-

trol of the sample point controller, the MJNN-DS (1)
and the MJNN-RS (2) are completely synchronized.
Sample point controller parameters can be obtained:
K i = M2−1 χi .

Remark 7 In (24), because of the existence of d PPdt(r (t)) ,

we cannot directly calculate the results by MATLAB,
so we use Lemma 3 to deform PP (r (t)). The results
are as follows:
D1 = {PP (r (t))|PP (r (t))


M
(l)

M
= rl (t)PP , rl (t) = 1, rl (t) ≥ 0 (26) Fig. 1 Area in general sense
l=1 l=1
(l)
where PP expresses respective polyhedron apex. The long as (6) is satisfied, (22) is strictly negative definite.
time-varying transition rates in PP (r (t)) deforms as This completes the proof.
follows:
Remark 8 When using Lemma 1 to deal with V̇ (t),

M
(l) d PP (r (t)) we set the Legendre parameter N = 2. If we take
PP (r (t)) = rl (t)PP ⇒
l=1
dt N = 1, we just need to replace η1 (t) with η̄1 (t) =
t
[eT (t) eT (t − d1 (t)) eT (t − d1 ) t−d1 (t) eT (s)ds

M
(l)

M−1
(n) (M)  t−d1 (t) T
= ṙl (t)PP = ṙn (t)(PP − PP ) t−d1 e (s)ds]T , and the rest of the processing is
l=1 n=1 basically the same as Theorem 1.
According to (22), we have
Remark 9 If we increase the Legendre parameter N ,
Υ (di (t), ḋ1 (t)) we can get a stricter bound for the integral term in

M V̇3 (t). In this case, Lyapunov functions, especially
(ll)
= rl2 (t)Ῡ P (di (t), ḋ1 (t)) η1 (t) in V1 , should be changed appropriately in the
l=1 order of increasing N to obtain a less conservative sta-
   b condition. 1N =
bility  b 1 band N = 2 correspond to
M−1 M
+ rl (t)rs (t)(Ῡ P(ls) (di (t), ḋ1 (t)) a e(s)ds and (b−a) a θ e(s)dsdθ in η1 (t), respec-
 b  b > 2,b corresponding to the following:
tively. When N
l=1 s=l+1
(sl) 1
· · · α N −1 e(α N )dα N · · · dα2 dα1 .
+Ῡ P (di (t), ḋ1 (t))) < 0 (b−a) N −1 a α1
(ls)
where Ῡ P (di (t), ḋ1 (t)) is equivalent to Υ (di (t), Remark 10 When considering the range of time-varying
(ls)
ḋ1 (t)) except Ω̄ P (di (t), ḋ1 (t)), it is expressed as fol- delays, it is generally set to: h 1 ≤ h ≤ h 2 , d1 ≤ ḣ ≤ d2 .
lows: For the convenience of the next discussion, we present
(ls) the delay and its derivatives in a two-dimensional plane
Ω̄ P (di (t), ḋ1 (t)) as shown in Fig. 1.
= H e{T0T (ḋ1 (t))PP(l) T1 (d1 (t))}
The plane presents a rectangle, and the coordinates

N
(l) (s) of its four vertices are: (h 1 , d1 ), (h 1 , d2 ), (h 2 , d1 ) and
+T1T (d1 (t)) μi j P j T1 (d1 (t)) (h 2 , d2 ). The area of rectangle is the range of time-
j=1
varying delays. It is improved by [21]. Change the four

M−1
(n) (M) vertices into the following: (0, 0), (h 1 , d2 ), (h 2 , 0) and
+T1T (d1 (t)) ṙn (t)(PP − PP )T1 (d1 (t)) (h 2 , d1 ). The shape is shown in Fig. 2. It can be seen
n=1
that in the same time-delay interval, the area of Fig. 2 is
(27)
smaller than that of Fig. 1, which indicates that it is less
We can get (6)–(9) by using the method of dealing conservative. Therefore, Theorem 1 can be optimized
 M−1 (n) (M)
with n=1 ṙn (t)(PP − PP ) in [3]. Therefore, as by this theory, and the results are as follows.

123
1886 N. Xu, L. Sun

Table 1 Different results to d1 for Example 1

μ22 = −1 μ22 = −2 μ22 = −3

By [5] 0.6898 1.1077 1.2455

By Theorem 0.6976 1.1384 1.5091
1 of [35]
By Theorem 0.9324 1.2508 1.7531
1 of [16]
By Theorem 1.0231 1.3027 1.8251
1 (N=1) of
this paper
By Theorem 1.0952 1.4125 1.8976
1 (N=2) of
Fig. 2 Area improved in this paper this paper

Theorem 2 Same as Theorem 1, MJNN-DS (1) and

MJNN-RS (2) achieve complete synchronization, if the in Table 1. From Table 1, we can see that with the
following hold: increase in N , the better the effect of affine Bessel–
(ls) (sl) Legendre inequalities is.
Υ P (0, 0) + Υ P (0, 0) < 0 (28)
Υ P(ls) (0, h 2 ) + Υ P(sl) (0, h 2 ) < 0 (29) Example 2 The following two modes and the matrix
(ls) (sl)
parameter s[16] are hold:
Υ P (di , h 1 ) + Υ P (di , h 1 ) < 0 (30)  
−3.4888 0.8057
Υ P(ls) (di , 0) + Υ P(sl) (di , 0) < 0 i = 1, 3 (31) A1 =
−0.6451 −3.2684
 
where Υ P(ls) (·) + Υ P(sl) (·) < 0 is defined in (6). A2 =
−2.4898 0.2895
1.3396 −0.0211
 
−0.8620 −1.2919
4 Numerical examples B1 =
−0.6841 −2.0729
 
−2.8306 0.4978
Firstly, Examples 1 and 2 illustrate the validity of B2 =
affine Bessel–Legendre inequalities and Wirtinger dou- −0.8436 −1.0115
ble integral inequalities. Secondly, Example 3 illus- with transition rates matrix
trates the effectiveness of optimizing two-dimensional  
−0.1 0.1
space of time delay. Finally, Example 4 shows that Π=
0.8 −0.8
under the control of sample point controller, MJNN-DS
(1) and MJNN-RS (2) can achieve synchronization. By setting different upper bounds of delay deriva-
tives h 2 , we obtain different upper bounds of delay as
Example 1 Consider the following two modes and the shown in Table 2. From Table 2, we can see that with
matrix parameter s[16]: the increase in N , the better the effect of affine Bessel–
    Legendre inequalities is.
0.5 −1 −5 1
A1 = A2 =
0 −3 1 0.2 Example 3 Consider MJNN-DS (1) and MJNN-RS (2)
   
0.5 −0.2 −0.3 0.5 with two modes and the matrix parameters:
B1 = B2 =
0.2 0.3 0.4 −0.5    
0 0.6 0.5 0.9
A1 = B1 =
with transition rates matrix 0 1 0 2
     
−7 7 10 0.9 0
Π= C1 = D1 =
3 −3 10 1 1.6
   
Let h 2 = 0, μ11 = −7. We compare the upper −4.5 2 −3.5 1.2
A2 = B2 =
bounds of time-varying delays. The results are shown −0.8 −4.3 1.0 −1.9

123
Synchronization control of Markov jump neural networks 1887

Table 2 Different results to d1 for Example 2

h 2 = 0.6 h 2 = 0.8 h 2 = 1.6

By Theorem 1 of [27] 0.4428 0.3795 0.3469

By Theorem 2 of [27] 0.4492 0.4341 0.4314
By [31] 0.4927 0.4261 0.3860
By Theorem 1 of [35] 0.5159 0.4814 0.4789
By Theorem 1 of [16] 0.5342 0.5147 0.5029
By Theorem 1 (N=1) of this paper 0.5398 0.5204 0.5162
By Theorem 1 of [24] 0.5642 0.5456 0.5279
By Theorem 1 (N=2) of this paper 0.5679 0.5501 0.5326

    Fig. 3 Chaotic curve

0.6 1.6 1.6 0.5
C2 = D2 =
1.8 −1.1 −0.7 1.0
   
0.1 0 0.2 0
G= Ei =
0 0.1 0 0.2
 
0.1 0
i = 1, 2, 3, 4 L =
0 0.1
Assume that the transition rate matrix is time-
varying in the following vertex polyhedron Π (t) =
sin 2 (t)Π (1) + cos 2 (t)Π (2) :
   
−0.8 0.8 −0.6 0.6
Π (1) = Π (2) =
0.6 −0.6 0.8 −0.8
Let δ1 = δ2 = δ3 = ε = ε1 = 0.1. When N = 2,
the time-varying delay range obtained from Theorem
1: 0 ≤ d1 (t) ≤ 1.63. Similarly, the time-varying delay
range obtained from Theorem 2: 0 ≤ d1 (t) ≤ 1.70. It Fig. 4 Time response of r (t)
can be shown that Theorem 2 is effective in optimizing
the two-dimensional geometric space of time delay.  
0.1 0
i = 1, 2, 3, 4 L =
Example 4 Consider MJNN-DS (1) and MJNN-RS (2) 0 0.1
with two modes and the matrix parameters: Assume that the transition rate matrix is time-
    varying in the following vertex polyhedron Π (t) =
0 0.6 0.5 0.9
A1 = B1 = sin 2 (t)Π (1) + cos 2 (t)Π (2) :
0 1 0 2  
    −0.8 0.8
C1 =
10
D1 =
0.9 0 Π (1) =
10 1 1.6 0.6 −0.6
     
−0.6 0.6
A2 =
−4.5 2
B2 =
−3.5 1.2 Π (2) =
−0.8 −4.3 1.0 −1.9 0.8 −0.8
    Let δ1 = δ2 = δ3 = ε = ε1 = 0.1, the range of
0.6 1.6 1.6 0.5
C2 = D2 = mixed time-varying delay is: 0 ≤ d1 (t) ≤ 1.7, 0 ≤
1.8 −1.1 −0.7 1.0
  d2 (t) ≤ 0.3, 0 ≤ d3 (t) ≤ 1.9, 0.1 ≤ ḋ1 (t) ≤ 0.6.
0.1 0 Substitute the above data into Theorem 2, we can get the
G=
0 0.1 parameters of the sample point controller as follows:
   
0.2 0 −6.35965433 2.82331641
Ei = K1 =
0 0.2 2.82331676 −9.48809923

123
1888 N. Xu, L. Sun

Fig. 5 Time response of x1 (t), y1 (t) Fig. 7 Time response of e1 (t)

Fig. 8 Time response of e2 (t)

Fig. 6 Time response of x2 (t), y2 (t)

  with mixed time-varying delay and parameter uncer-

−6.35965406 2.82331609
K2 = tainties.
2.82331651 −9.48809847
The neuron excitation function is f (x) = tanhx.
The initial condition is x0 (θ ) = [−0.3, 2, 1.2]. As 5 Conclusions
shown in Fig. 3, when MJNN (1) takes the above param-
eters, it shows obvious chaotic characteristics. Figure 4 In this paper, a sample point controller is used to syn-
is a Markov jump response curve with time-varying chronize the DS and RS of MJNN with mixed time-
probability transition perturbations. Figures 5 and 6 varying delay and uncertain parameters. When dealing
describe the time state curves of MJNN-DS (1) and with error systems, Wirtinger double integral inequali-
MJNN-RS (2). Figures 7 and 8 describe the conver- ties and affine Bessel–Legendre inequalities are intro-
gence behavior of errors between MJNN-DS (1) and duced into Lyapunov functional to reduce conservative-
MJNN-RS (2). The numerical simulation shows the ness. In addition, when discussing the two-dimensional
validity of the sample point controller for the complete geometric area with time-varying delays, the conser-
synchronization of MJNN-DS (1) and MJNN-RS (2) vativeness is reduced by changing the vertex of the

123
Synchronization control of Markov jump neural networks 1889

polyhedron without changing the range of the delays. 12. Mohammadzadeh, A., Ghaemi, S.: Robust synchronization
Finally, it is verified by numerical simulation that of uncertain fractional-order chaotic systems with time-
varying delay. Nonlinear Dyn. 93(4), 1809–1821 (2018)
MJNN-DS and MJNN-RS are fully synchronized under 13. Nagamani, G., Joo, Y.H., Radhika, T.: Delay-dependent dis-
the control of sample point controller, and the parame- sipativity criteria for Markovian jump neural networks with
ters of the controller are obtained. random delays and incomplete transition probabilities. Non-
linear Dyn. 91(4), 2503–2522 (2018)
Acknowledgements Project supported by the National Natural 14. Nagamani, G., Joo, Y.H., Radhika, T.: Delay-dependent dis-
Science Foundation of China (Grant Nos. 61403278, 61503280). sipativity criteria for Markovian jump neural networks with
The authors are very indebted to the Editor and the anonymous random delays and incomplete transition probabilities. Non-
reviewers for their insightful comments and valuable suggestions linear Dyn. 91(56), 2503–2522 (2018)
that have helped improve the academic research. 15. Novienko, V., Ratas, I.: In-phase synchronization in complex
oscillator networks by adaptive delayed feedback control.
Compliance with ethical standards Phys. Rev. E 98(4), 042302 (2018)
16. Nuo, Xu, Sun, L.: An improved delay-dependent stability
Conflict of interest The authors declare that they have no con- analysis for Markovian jump systems with interval time-
flict of interest. varying-delays. IEEE Access 6, 33055–33061 (2018)
17. Park, M., Kwon, O., Park, J.H., Lee, S., Cha, E.: Stability
of time-delay systems via Wirtinger-based double integral
inequality. Automatica 55, 204–208 (2015)
18. Park, I.S., Kwon, N.K., Park, P.G.: Dynamic output-
References feedback control for singular Markovian jump systems with
partly unknown transition rates. Nonlinear Dyn. 95(4), 1–12
1. Ahn, C.K., Shi, P., Wu, L.: Receding horizon stabiliza- (2019)
tion and disturbance attenuation for neural networks with 19. Rong, Z., Yang, Y., Xu, Z., et al.: Function projective syn-
time-varying delay. IEEE Trans. Cybern. 45(12), 2680–2692 chronization in drive—response dynamical network. Phys.
(2017) Lett. A 374(30), 3025–3028 (2010)
2. Briat, C.: Convergence and equivalence results for 20. Schibli, T.R., Kim, J., Kuzucu, O., et al.: Attosecond active
the Jensen’s inequality—application to time-delay and synchronization of passively mode-locked lasers by bal-
sampled-data systems. IEEE Transa. Autom. Control 56(7), anced cross correlation. Opt. Lett. 28(11), 947–9 (2003)
1660–1665 (2012) 21. Seuret, A.: Frdric Gouaisbaut. Stability of linear systems
3. Ding, Y., Liu, H.: Stability analysis of continuous-time with time-varying delays using Bessel–Legendre inequali-
Markovian jump time-delay systems with time-varying tran- ties. IEEE Trans. Autom. Control 63(1), 225–232 (2017)
sition rates. J. Frankl. Inst. 353(11), 2418–2430 (2016) 22. Seuret, A., Gouaisbaut, F.: Wirtinger-based integral inequal-
4. Gu, K., Kharitonov, V.L., Chen, J.: Stability of Time-Delay ity: application to time-delay systems. Automatica 49(9),
Systems. Birkh-user, Boston (2003) 2860–2866 (2013)
5. Guan, H., Gao, L.: Delay-dependent robust stability and H∞ 23. Shu, Y., Liu, X.G., Qiu, S., et al.: Dissipativity analysis for
control for jump linear systems with interval time-varying generalized neural networks with Markovian jump param-
delay. In: Proceedings of the 26th Chinese Control Confer- eters and time-varying delay. Nonlinear Dyn. 89(3), 2125–
ence, pp. 609–614. IEEE, Zhangjiajie (2007) 2140 (2017)
6. Gyurkovics, Eva: A Note on Wirtinger-Type Integral 24. Sun, L., Nuo, X.: Stability analysis of Markovian jump
Inequalities for Time-Delay Systems. Pergamon Press, Inc, system with multi-time-varying disturbances based on
Oxford (2015) improved interactive convex inequality and positive definite
7. Kasemsuk, C., Oyama, G., Hattori, N.: Management of condition. IEEE Access 7, 54910–54917 (2019)
impulse control disorders with deep brain stimulation: a 25. Syed, A.M., Marudai, M.: Stochastic stability of discrete-
double-edged sword. J. Neurol. Sci. 374, 63–68 (2017) time uncertain recurrent neural networks with Markovian
8. Lee, W.I., Lee, S.Y., Park, P.G.: Affine Bessel–Legendre jumping and time-varying delays. Math. Comput. Model.
inequality: application to stability analysis for systems with 54(9–10), 1979–1988 (2011)
time-varying delays. Automatica 93, S0005109818301687 26. Tao, J., Wu, Z.G., Su, H., et al.: Asynchronous and resilient
(2018) filtering for Markovian jump neural networks subject to
9. Li, R.G., Wu, H.N.: Adaptive synchronization control based extended dissipativity. IEEE Trans. Cybern. 99, 1–10 (2018)
on QPSO algorithm with interval estimation for fractional- 27. Wang, J., Luo, Y.: Further improvement of delay-dependent
order chaotic systems and its application in secret commu- stability for Markov jump systems with time-varying delay.
nication. Nonlinear Dyn 92(3), 1–25 (2018) In Proceedings of the 7th World Congress on Interligent
10. Lin, F.F., Zeng, Z.Z.: Synchronization of uncertain Control and Automation, pp. 6319-6324. IEEE, Chongqing,
fractional-order chaotic systems with time delay based on (2008)
adaptive neural network control. Acta Phys. Sin. 66, 9 (2017) 28. Wang, Y., Xie, L., de Souza, C.E.: Robust control of a class of
11. Mayer, J., Schuster, H.G., Claussen, J.C., et al.: Corticotha- uncertain nonlinear system. Syst. Control Lett. 19(2), 139–
lamic projections control synchronization in locally coupled 149 (1992)
bistable thalamic oscillators. Phys. Rev. Lett. 99(6), 068102 29. Wang, Y.F., Lin, P., Wang, L.S.: Exponential stability
(2007) of reaction-diffusion high-order Markovian jump Hopfield

123
1890 N. Xu, L. Sun

neural networks with time-varying delays. Nonlinear Anal. 35. Zhao, X., Zeng, Q.: Delay-dependent stability analysis for
Real World Appl. 13(3), 1353–1361 (2012) Markovian jump systems with interval time-varying-delay.
30. Wang, J., Chen, X., Feng, J., et al.: Synchronization of net- Int. J. Autom. Comput. 7(2), 224–229 (2010)
worked harmonic oscillators subject to Markovian jumping 36. Zhi, Z., Liu, K., Wang, W.Q., et al.: Robust adaptive beam-
coupling strengths. Nonlinear Dyn. 91(1), 1–13 (2018) forming against mutual coupling based on mutual coupling
31. Xu, S., Lam, J., Mao, X.: Delay-dependent H∞ control and coefficients estimation. IEEE Trans. Veh. Technol. 99, 1–1
filtering for uncertain Markovian jump systems with time- (2017)
varying delays. IEEE Trans. Circuits Syst. I Regul. P. 54(9), 37. Zhu, X.L., Yang, G.H.: Jensen integral inequality approach
2070–2077 (2007) to stability analysis of continuous-time systems with time-
32. Zeng, H.B., He, Y., Wu, M., et al.: Free-matrix-based integral varying delay. IET Control Theory Appl. 2(6), 524–534
inequality for stablilty analysis of systems with time-varying (2008)
delay. IEEE Trans. Autumatic Control 60(10), 2768–2772
(2015)
33. Zhang, X., Lv, X., Li, X.: Sampled-data-based lag synchro- Publisher’s Note Springer Nature remains neutral with regard
nization of chaotic delayed neural networks with impulsive to jurisdictional claims in published maps and institutional affil-
control. Nonlinear Dyn. 90(3), 2199–2207 (2017) iations.
34. Zhang, X.M., Han, Q.L., Seuret, A., et al.: An
improved reciprocally convex inequality and an augmented
Lyapunov–Krasovskii functional for stability of linear sys-
tems with time-varying delay. Automatica 84, 221–226
(2017)

123