Proceedings of the 17th World Congress

The International Federation of Automatic Control

Seoul, Korea, July 6-11, 2008

A Delay Decomposition Approach to

Stability of Linear Neutral Systems 
Qing-Long Han ∗
∗
School of Computing Sciences, Central Queensland University
Rockhampton QLD 4702, Australia
(Tel: 61-7-4930-9729; e-mail: q.han@cqu.edu.au)

Abstract: This paper is concerned with stability of linear neutral systems. Firstly, a new
approach, a delay decomposition approach, is proposed to deal with the stability issue. The idea
of the approach is that the delay interval is uniformly divided into N segments with N a positive
integer, and a proper Lyapunov-Krasovskii functional is chosen with different weighted matrices
corresponding to different segments in the Lyapunov-Krasovskii functional. Secondly, based on
the delay decomposition approach, some new delay-dependent stability criteria for linear neutral
systems are derived. These criteria are much less conservative and include some existing results
as their special cases. Finally, numerical examples show that significant improvement using
the delay decomposition approach is achieved over some existing method even for coarse delay
decomposition.

1. INTRODUCTION For comparison, we use a numerical example for system

(1) with
Time-delays are frequently encountered in many fields 
−2 0
 
−1 0

of science and engineering, including communication net- A= , Ad = .
0 −0.9 −1 −1
work, manufacturing systems, biology, economy and other
areas (Hale and Verduyn Lunel, 1993; Gu et al., 2003). The analytical delay limit for stability for the numerical
During the last two decades, the problem of stability example is calculated as ranalytical = 6.17258. In the
of retarded and neutral type have been the subject of existing literature, in order to derive a delay-dependent
considerable research efforts. Many significant results have stability criterion, one transforms system (1) into a system
been reported in the literature. For the recent progress, the with a distributed delay, i.e.
 t
reader is referred to Gu et al. (2003) and the references ẋ(t) = (A + Ad )x(t) − Ad [Ax(ξ) + Ad x(ξ − r)]dξ. (3)
therein. t−r
Choose a Lyapunov function
For the stability issue of time-delay systems, in the existing
literature, there are two kinds of Lyapunov-Krasovskii V (t, xt ) = xT (t)P x(t), P = P T > 0, (4)
functionals, i.e. complete Lyapunov-Krasovskii functionals and apply Razumikhin Theorem to obtain rmax = 0.9041
and simple Lyapunov-Krasovskii functionals, for estimat- (Gu et al., 2003). As pointed out by Gu et al. (2003)
ing the maximum time-delay bound the system can toler- (Example 5.3 in Gu et al. (2003)), for this example, the
ate and still retain stability. Using the complete Lyapunov- stability of system (1) is equivalent to that of system (3).
Krasovskii functionals (Gu, 2001; Gu et al., 2003), one The conservatism of the result is due to the application
can obtain the maximum time-delay bound which is very of the Razumikhin Theorem. For some system (1) with
close to the analytical delay limit for stability. Employ- different system’s matrices, the model transformation (3)
ing the simple Lyapunov-Krasovskii functionals usually may induce additional dynamics (Gu et al., 2003). To
yields conservative results. However, the results based on reduce the conservatism, instead of transforming system
the simple Lyapunov-Krasovskii functionals can be easily (1) into (3), one transforms it into  t
applied to controller synthesis and filter design. Hence, it is ẋ(t) = (A + Ad )x(t) − Ad ẋ(ξ)dξ. (5)
still an attractive topic for finding some simple Lyapunov- t−r
Krasovskii functionals, by which one can have less conser- Then choosing a Lyapunov-Krasovskii
 t functional
vative results.
V (t, xt ) = xT (t)P x(t) + xT (ξ)Qx(ξ)dξ
In order to introduce a new approach to stability analysis t−r
0 t
for time-delay systems, we consider the retarded system
 
+ xT (ξ)ATd RAd x(ξ)dξdθ, (6)
ẋ(t) = Ax(t) + Ad x(t − r), (1) −r t+θ
with the continuous vector valued initial condition where P = P T > 0, Q = QT > 0, R = RT > 0, and
x(θ) = φ(θ), ∀θ ∈ [−r, 0]. (2) using the bounding technique for some cross term yield
 This work was supported in part by Central Queensland University rmax = 4.3588 (Park, 1999). This result was also derived
for the Research Advancement Awards Scheme Project “Robust by decomposing delayed term matrix as Ad = Ad1 + Ad2
Fault Detection, Filtering and Control for Uncertain Systems with in Han (2002). In Fridman (2001), the author introduced
Time-Varying Delay” (Jan 2006 - Dec 2008). a descriptor transformation

978-3-902661-00-5/08/$20.00 © 2008 IFAC 2607 10.3182/20080706-5-KR-1001.1356

17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

ẋ(t) = y(t), (7) variables apart from Lyapunov-Krasovskii functional ma-

 t trices variables. One way to solve the problem is to choose
y(t) = (A + Ad )x(t) − Ad y(ξ)dξ. (8) a new Lyapunov-Krasovskii functional. For this purpose,
t−r we propose the following new simple Lyapunov-Krasovskii
A Lyapunov-Krasovskii functional was chosen as functional N  t−(i−1)h
T 
V (t, xt )=xT (t)P x(t)+ xT (ξ)Qi x(ξ)dξ
    
x(t) I 0 P1 0 x(t)
V (t, xt )= t−ih
y(t) 0 0 P2 P3 y(t) i=1
 t N  −(i−1)h  t
xT (ξ)Qx(ξ)dξ ẋT (ξ)(hRi )ẋ(ξ)dξdθ

+ + (13)
t−r i=1 −ih t+θ
 0 t
where h = r/N and N is the number (a positive integer) of
+ y T (ξ)ATd RAd y(ξ)dξdθ, (9)
−r t+θ
divisions of the interval [−r, 0], and h is the length of each
division; and xt is defined as xt = x(t + θ), ∀θ ∈ [−r, 0]
where P1 = P1T > 0, Q = QT > 0 and R = RT > 0. and P = P T > 0, Qi = QTi > 0, Ri = RiT > 0 (i =
Use this model transformation and bounding technique 1, 2, · · · , N ).
for cross terms (Park, 1999) to obtain rmax = 4.4721
in Fridman and Shaked (2002). In He et al. (2004a,b), It should be pointed out that if N = 1, the Lyapunov-
the authors introduce some slack variables (free-weighting Krasovskii functional (13) reduces to (11) employed in Han
matrices) to derive the same result. (2005a) by setting Q = Q1 and R = R1 . Based on (13),
one can have a delay-dependent stability criterion (see
In Han (2005a), the author avoided using model transfor- Corollary 5 in this paper). Applying this new criterion,
mation on system (1) and chose the following Lyapunov- one obtains the maximum allowed delay bound as rmax =
Krasovskii functional 5.7175 for N = 2; rmax = 5.9678 for N = 3; and rmax =
 t
V (t, xt ) = xT (t)P x(t) + xT (ξ)Qx(ξ)dξ 6.0569 for N = 4 and so on, which significantly improve the
t−r result rmax = 4.4721 in the above mentioned references.
 t One can clearly see that we have made very significant
+ (r − t + ξ)ẋT (ξ)(rR)ẋ(ξ)dξ, (10) steps towards the analytical delay limit for stability of the
t−r system.
which is equivalent to t Since the delay interval [−r, 0] is decomposed into N uni-

V (t, xt )=xT (t)P x(t) + xT (ξ)Qx(ξ)dξ form segments, for the purpose of distinguishing from the
t−r
 0  t
existing methods, in what follows, we refer the approach
+ ẋT (ξ)(rR)ẋ(ξ)dξdθ (11) based on the decomposition idea as a delay decomposition
−r t+θ approach.
where P = P T > 0, Q = Q > 0 and R = RT > 0. Instead
T
In this paper, we will use a delay decomposition approach
of using the bounding technique for some cross term, the proposed above to study the stability of linear neutral
author used the
 following bounding (Han, 2005a) systems. We will derive new stability criteria which include
t
some existing results as their special cases. These new
− ẋT (ξ)(rR)ẋ(ξ)dξ stability criteria will be much less conservative than some
t−r
T  existing results.
−R R
  
x(t) x(t)
≤ (12) Notation: Rn denotes the n-dimensional Euclidean space.
x(t − r) R −R x(t − r)
Rm×n is the set of all m × n real matrices. The notation
to derive the maximum allowed delay bound as rmax = P > 0 (P ≥ 0) means that P is symmetric and positive
4.4721. Compared with the above mentioned results, the definite (positive semi-definite). For symmetric matrices P
most advantage of the result in Han (2005a) is that the and Q, the notation P > Q (P ≥ Q) means that matrix
stability condition which was formulated in an LMI form, P − Q is positive definite (positive semi-definite). I is an
was very simple and easily applied to controller design, identity matrix of appropriate dimensions. λi (A) is the ith
and did only include the Lyapunov-Krasovskii functional eigenvalue of a real matrix A. · stands for the Euclidean
matrices variables P, Q and R, which means that no addi- vector norm or the induced matrix 2-norm as appropriate.
tional matrix variable was involved. From the computation For an arbitrary matrix U and two symmetric matrices
point of view, it is clear to see that testing the result in Han P and Q, the symmetric term in a symmetric matrix is
(2005a) is less time-consuming than some existing results
 
P U P U
denoted by *, i.e. = .
in the literature. However, the result rmax = 4.4721 is ∗ Q UT Q
not close enough to the analytical delay limit for stability
ranalytical = 6.17258 and work needs to be done to arrive 2. STABILITY OF LINEAR NEUTRAL SYSTEMS
at a value much closer to the analytical delay limit for
stability. Therefore, the natural question is: How can one Consider the following linear neutral system
improve the result by using simple Lyapunov-Krasovskii d
functionals? Answer to this question will significantly en- [x(t) − Cx(t − τ )] = Ax(t) + Ad x(t − r), (14)
dt
hance the stability analysis and controller synthesis of with
time-delay systems. It seems that using the existing sim- x(θ) = φ(θ), ∀θ ∈ [− max{r, τ }, 0]. (15)
ple Lyapunov-Krasovskii functionals can not realize the where x(t) ∈ Rn is the state vector of the system; r > 0
outcome even if one introduces more additional matrices is the constant discrete delay and τ > 0 is the constant

2608
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

neutral delay; φ(·) is a continuous vector valued initial hAT R1 hAT R2 · · · hAT RN
⎛ ⎞
function; A ∈ Rn×n , Ad ∈ Rn×n and C ∈ Rn×n are ⎜ 0 0 ··· 0 ⎟
constant matrices. 0 0 ··· 0
⎜ ⎟
(3)
⎜ ⎟
Π =⎜ .. .. .. .. ⎟,
Define xt as xt = x(t + θ), ∀θ ∈ [−max{r, τ }, 0], and the
⎜
. . . .
⎟
operator Dxt = x(t) − Cx(t − τ ). We assume that
⎜ ⎟
⎝ 0 0 ··· 0 ⎠
Assumption 1. |λi (C)| < 1 ( i = 1, 2, · · · , n). T T T
hAd R1 hAd R2 · · · hAd RN
(5)
Π = hC T R1 hC T R2 · · · hC T RN ,

Choose a Lyapunov-Krasovskii functional candidate as
N  t−(i−1)h
T Π(6) = diag ( −R1 −R2 · · · −RN ) .
xT (ξ)Qi x(ξ)dξ

V (t, xt )=(Dxt ) P (Dxt )+
i=1 t−ih
In order to prove Proposition 2, we need the following
N  −(i−1)h  t lemma.
ẋT (ξ)(hRi )ẋ(ξ)dξdθ

+ Lemma 3. (Han, 2005a) For any constant matrix W ∈
i=1 −ih t+θ
Rn×n , W = W T > 0, scalar r > 0, and vector valued
t t
function ẋ : [−r, 0] → Rn such that the following integra-
 
T
+ x (ξ)S1 x(ξ)dξ + ẋT (ξ)S2ẋ(ξ)dξ (16)
t−τ t−τ
tion is well defined, then
 t
where h = r/N and N is the number of divisions of the
− ẋT (ξ)(rW )ẋ(ξ)dξ
interval [−r, 0], and h is the length of each division; and t−r
P = P T > 0, Qi = QTi > 0, Ri = RiT > 0 (i = −W W
  
x(t)
1, 2, · · · , N ), Sj > 0 (j = 1, 2). ≤ xT (t) xT (t − r)

.
W −W x(t − r)
We now state and establish the following result.
Proposition 2. Under Assumption 1, for given scalars Proof of Proposition 2. Taking the derivative of V (t, xt )
r > 0 and τ > 0, the system described by (14) and with respect to t along the trajectory of (14) yields
(15) is asymptotically stable if there exist real n × n V̇ (t, xt ) = xT (t)(AT P + P A + S1 )x(t)
matrices P > 0, Qi > 0, Ri > 0 (i = 1, 2, · · · , N ) and
Sj > 0 (j = 1,⎛2) such that +2xT (t)P Ad x(t − N h) − 2xT (t)AT P Cx(t − τ )
Π(1) Π(2) 0 Π(3) Π(4) −2xT (t − N h)ATd P Cx(t − τ )
⎞
⎜ ∗ −S1 0 0 0 ⎟
⎜ (5) T
⎟ −xT (t − τ )S1 x(t − τ ) − ẋT (t − τ )S2 ẋ(t − τ )
Π=⎜ ∗
⎜ ∗ −S2 Π C S2 ⎟ ⎟ < 0, (17)
N
⎝ ∗ ∗ ∗ Π(6) 0 ⎠
xT (t − (i − 1)h)Qi x(t − (i − 1)h)

+
∗ ∗ ∗ ∗ −S2
i=1
where ⎛ (1) N
Π11 R1 0 · · · 0
⎞
xT (t − ih)Qi x(t − ih)

P Ad −
⎜ ∗ Π(1) R · · · 0 0
2 i=1
⎟
⎜ 22 ⎟
(1)
∗ ∗ Π · · · 0 0 N
⎜ ⎟
(1)
⎜ 33 ⎟
ẋT (t)(h2 Ri )ẋ(t) + ẋT (t)S2 ẋ(t)

Π =⎜ . ⎜ . . . . . ⎟ +
⎜ .. .. .. .. .. .. ⎟
i=1
⎟
⎝ ∗ (1)
∗ ∗ · · · ΠN N
⎜ ⎟
RN ⎠ N  t−(i−1)h
(1)
ẋT (ξ)(hRi )ẋ(ξ)dξ.

∗ ∗ ∗ ··· ∗ ΠN +1 N +1 − (18)
i=1 t−ih
with
(1) Δ
Π11 = AT P + P A + Q1 − R1 + S1 , Rewrite system (14) as
(1) Δ
Π22 = −Q1 − R1 + Q2 − R2 , ẋ(t) = Ax(t) + Bx(t − N h) + C ẋ(t − τ ). (19)
(1) Δ Then we have
Π33 = −Q2 − R2 + Q3 − R3 ,
ẋT (t)(h2 Ri )ẋ(t) = η T (t)T (h2 Ri )η(t), (20)
..
. and
(1) Δ ẋT (t)S2 ẋ(t) = η T (t)T S2 η(t), (21)
ΠN N = −QN −1 − RN −1 + QN − RN ,
where
(1) Δ
ΠN +1 N +1 = −QN − RN , η T (t) = η1T (t) · · · η2T (t) η3T (t) ,


with
and
η1T (t) = xT (t) xT (t − h) xT (t − 2h) ,

T T
−A P C −A S2
⎛ ⎞ ⎛ ⎞
0 0 η2T (t) = xT (t − (N − 1)h) xT (t − N h) ,
⎜ ⎟ ⎜ ⎟ 
0 0
⎜ ⎟ ⎜ ⎟
Π(2) = ⎜ (4) η3T (t) = xT (t − τ ) ẋT (t − τ ) ,
⎜ ⎟ ⎜ ⎟ 
.. ⎟ , Π = ⎜ .. ⎟,
. .
⎜ ⎟ ⎜ ⎟
and
⎜ ⎟ ⎜ ⎟
⎝ 0 ⎠ ⎝ 0 ⎠
T T  = ( A 0 0 · · · 0 Ad 0 C ) .
−Ad P C −Ad S2
Use Lemma 3 to obtain

2609
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

t−(i−1)h (1)
···
 ⎛ ⎞
T Ξ11 R1 0 0 P Ad
− ẋ (ξ)(hRi )ẋ(ξ)dξ ⎜ ∗ Ξ(1) R
t−ih ⎜ 22 2 ··· 0 0 ⎟
⎟
(1)
−Ri Ri ⎜ ∗ ∗ Ξ33 ···
 
0 0
⎜ ⎟
T
≤ ζ (t) Ξ(1)
⎟
ζ(t), (22) =⎜
Ri −Ri ⎜ .. .. .. .. .. .. ⎟
.
⎟
⎜ . . . . . ⎟
x(t − (i − 1)h)
 
⎝ ∗ (1)
∗ ∗ · · · ΞN N
⎜ ⎟
where ζ(t) = . Then from (18), (20), RN ⎠
x(t − ih) (1)
∗ ∗ ∗ · · · ∗ ΞN +1 N +1
(21) and (22), we have
(1) (1) (1) (1)
Π(1) Π(2) 0 with Ξ11 = Π11 − S1 , Ξii = Πii (i = 2, · · · , N + 1), and
⎛ ⎞

V̇(t, xt )≤ η T (t)⎝ ∗ −S1 0 ⎠η(t) ⎛

hAT R1 hAT R2 · · · hAT RN −1 hAT RN
⎞
∗ ∗ −S2 ⎜ 0 0 ··· 0 0 ⎟
···
N
0 0 0 0
 ⎜ ⎟
Ξ(2) = ⎜
⎜ ⎟
T T 2 T
 ⎟,
+η (t)  (h Ri )+ S2  η(t). (23) .. .. .. .. ..
⎜
. . . . .
⎟
i=1
⎜ ⎟
⎝ 0 0 ··· 0 0 ⎠
According to Schur
⎞ complement, (17) is equivalent to hATd R1 hATd R2 · · · hATd RN −1 hATd RN
Π(1) Π(2) 0
⎛
⎝ ∗ −S1 0 ⎠ + N T (h2 Ri ) + T S2  < 0. Ξ(3) = diag ( −R1 −R2 · · · −RN −1 −RN ) .

i=1
∗ ∗ −S2
When τ = r, system (14) becomes
Therefore, there exists a scalar λ1 > 0 such that V̇ (t, xt ) ≤ d
−λ1 xT (t)x(t) < 0 for x(t) = 0. Notice that Assumption 1 [x(t) − Cx(t − r)] = Ax(t) + Bx(t − r), (26)
dt
guarantees the operator Dxt is stable. Hence, by Theorem with x(θ) = φ(θ), ∀θ ∈ [−r, 0]. (27)
8.1 (pp. 292-293, Hale and Verduyn Lunel (1993)), the The corresponding Lyapunov-Krasovskii functional candi-
system described by (14) and (15) is asymptotically stable. date is N  t−(i−1)h
This completes the proof. 
Ṽ (t, xt )=(D̃xt )T P (D̃xt )+ xT (ξ)Qi x(ξ)dξ

If N = 1, then (16) becomes the Lyapunov-Krasovskii i=1 t−ih
functional employed in Han (2005b) N  −(i−1)h  t
 t
ẋT (ξ)(hRi )ẋ(ξ)dξdθ

+
V (1) (t, xt )=(Dxt )T P (Dxt )+ xT (ξ)Q1 x(ξ)dξ
t−r i=1 −ih t+θ
 0  t
 t
+ ẋT (ξ)(rR1 )ẋ(ξ)dξdθ + ẋT (ξ)S3 ẋ(ξ)dξ (28)
−r t+θ t−N h
 t  t where D̃xt = x(t)−Cx(t−N h). Then we have the following
T
+ x (ξ)S1 x(ξ)dξ + ẋT (ξ)S2 ẋ(ξ)dξ result.
t−τ t−τ
Proposition 6. Under Assumption 1, for a given scalar r >
By Proposition 2 we have the following corollary.
0, the system described by (26) and (27) is asymptotically
Corollary 4. (Han, 2005b) Under Assumption 1, for given stable if there exist real n × n matrices P > 0, Qi > 0,
scalars r > 0 and τ > 0, the system described by (14) Ri > 0 (i = 1, 2, · · · , N
and (15) is asymptotically stable if there exist real n × n  ) and S3 >0 such that
Ω(1) Ω(2)
matrices P > 0, Q1 > 0, R1 > 0 and Sj > 0 (j = 1, 2) Ω= < 0, (29)
Ω(1)T Ω(3)
such
⎛ that
(1, 1) (1, 2) −AT P C 0 rAT R1 AT S2 where ⎛ (1)
⎞
Ω11 R1 0 · · · 0 P Ad − AT P C 0
⎞
⎜ ∗ (2, 2) −ATd P C 0 rATd R1 ATd S2 ⎟ ⎜ ∗ Ω(1) R · · · 0
⎜ ∗
⎜
∗ −S1 0 0 0 ⎟
⎟ ⎜ 22 2 0 0 ⎟⎟
⎟ < 0, (24) (1)
T T ⎜ ∗ ∗ Ω33 · · · 0
⎜ ⎟
⎜ ∗
⎜
∗ ∗ −S2 rC R1 C S2 ⎟ 0 0 ⎟
∗ ∗ ∗ ∗ −R1 Ω(1) = ⎜ . .. .. . . . .. .. ⎟ ,
⎜ ⎟
0 ⎜ .. . ..
⎝ ⎠
. . . .
∗ ∗ ∗ ∗ ∗ −S2
⎟
⎜ ∗ (1)
∗ ∗ · · · ΩN N
⎜ ⎟
RN 0 ⎟
where ⎜
⎝ ∗ (1)
⎟
Δ ∗ ∗ ··· ∗ ΩN +1 N +1 0 ⎠
(1, 1) = AT P + P A + S1 + Q1 − R1 , ∗ ∗ ∗ ··· ∗ ∗ −S3
Δ Δ
(1, 2) = P Ad + R1 , (2, 2) = −Q1 − R1 . (1) (1) (1)
with Ω11 = Π11 − S1 , Ωii = Πii (i = 2, · · · , N ) and
(1)

(1) (1)
If τ = 0, then system (14) becomes system (1). The follow- ΩN +1 N +1 = ΠN +1 N +1 − ATd P C − C T P Ad , and
ing corollary is immediately implied by using Proposition hAT R1 hAT R2 · · · hAT RN AT S3
⎛ ⎞
2. ⎜ 0 0 ··· 0 0 ⎟
Corollary 5. For a given scalar r > 0, the system described
⎜
⎜ 0 0 ··· 0 0 ⎟⎟
by (1) and (2) is asymptotically stable if there exist real (2)
⎜ .. .. . .. .
.. . ⎟
.. ⎟
Ω =⎜ ⎜
. . ⎟,
n × n matrices P > 0, i > 0 (i = 1, 2, · · · , N )
 Qi(1)> 0,(2)R
such that
⎜ 0 0 · · · 0 0 ⎟
Ξ Ξ ⎜
⎝ hAT R hAT R · · · hAT R AT S ⎠
⎟
Ξ= < 0, (25) d 1 d 2 d N d 3
Ξ(2)T Ξ(3)
hC T R1 hC T R2 · · · hC T RN C T S3
where

2610
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Ω(3) = diag ( −R1 −R2 · · · −RN −S3 ) .

Notice that system (14) can be rewritten as

ẋ(t) = Ax(t) + Bx(t − r) + C ẋ(t − τ ). (30)
For this form, choosing a Lyapunov-Krasovskii functional
candidate as N  t−(i−1)h
V̂ (t, xt )=xT (t)P x(t)+ xT (ξ)Qi x(ξ)dξ


i=1 t−ih
N  −(i−1)h  t
ẋT (ξ)(hRi )ẋ(ξ)dξdθ

+
i=1 −ih t+θ
Fig. 1. Small PEEC model for metal strip
 t
T elements, that is, the partial inductances of the form
+ ẋ (ξ)S4 ẋ(ξ)dξ (31) d
t−τ Lpij dt (ij (t − r)) are coupled by the retarded dependent
we can conclude that current sources of the form pij /pii icj (t − r). This model
usually results in a neutral delay differential equation. The
Proposition 7. For given scalars r > 0 and τ > 0, the
mathematical model for the PEEC shown in Figure 1 can
system (30) is asymptotically stable if there exist real n×n
be given as
matrices P > 0, Qi > 0, Ri > 0 (i = 1, 2, · · · , N ) and
C0 ẏ(t) + G0 y(t) + C1 ẏ(t−r) + G1 y(t − r)

S4 > 0 such that  (1) (2) 
Φ Φ = Bu(t, t − r), t ≥ t0 (33)
Φ= < 0, (32) y(t) = g(t), t ≤ t0
Φ(2)T Φ(3)
where C0 is a diagonal matrix. The associated delay
where
⎛ (1) differential equation of neutral type is
Φ11 R1 0 · · · 0
⎞
P Ad PC 
d
⎜ ∗ Φ(1) R · · · 0
2 0 0 ⎟ [y(t)−Cy(t − r)] = Ay(t)+Ad y(t − r), t ≥ t0 (34)
⎜ 22 ⎟ dt
⎜ ∗
⎜ (1)
∗ Φ33 · · · 0 0 0 ⎟
⎟ y(t) = g(t), t ≤ t0
Φ(1) = ⎜ . .. .. . . . .. .. ⎟
⎜ ⎟
⎜ .. . . . .. . . ⎟ Consider the system (34) with
(1)
⎜ ∗ ∗ ∗ · · · 0 −3
⎜ ⎟
Φ R 0 β 1 2 1
   
NN N
⎟
(1) A = 100× 3 −9 0 , Ad = 100 × −0.5 −0.5 −1 ,
⎜ ⎟
⎝ ∗ ∗ ∗ · · · ∗ ΦN +1 N +1 0 ⎠
∗ ∗ ∗ ··· ∗ ∗ −S4 1 2 −6 −0.5 −1.5 0
−1 5 2
 
(1) (1) (1) (1)
with Φ11 = Π11 − S1 , Φii = Πii (i = 2, · · · , N + 1), and 1
C= × 4 03 .
72 −2 4 1
hAT R1 hAT R2 · · · hAT RN AT S4
⎛ ⎞
⎜ 0 0 ··· 0 0 ⎟ For β = −7, Bellen et al. (1999) studied the stability
⎜
⎜ 0 0 ··· 0 0 ⎟⎟ problem of the system and concluded that the system is
(2)
⎜ .
. .
. . . .
. .. ⎟ asymptotically stable. However, the result is independent
Φ =⎜ ⎜
. . . . . ⎟
⎟
⎜ 0 0 ··· 0 0 ⎟ of the delay. In Han (2005c), the author claimed that the
⎜
⎝ hAT R hAT R · · · hAT R AT S ⎠
⎟ system is asymptotically stable independent of the delay
d 1 d 2 d N d 4 for β ≤ −2.106.
hC T R1 hC T R2 · · · hC T RN C T S4
Applying the Proposition 6 in this paper and the criteria
Φ(3) = diag ( −R1 −R2 · · · −RN −S4 ) . in Han (2005c) and Yue and Han (2004), Table 1 lists the
Remark 8. Notice that for N = 1, the Lyapunov- maximum allowed time-delay rmax for asymptotic stability
Krasovskii functionals (28) and (31) for systems (26) and for different β. From this table, one can clearly see that
(30) reduce to the ones used in Han (2005b). Similar to for N = 1, the results in Han (2005c) are recovered; for
Corollary 4, we can have the corresponding results which N ≥ 2, the results using Proposition 6 can provide much
recover the results in Han (2005b). less conservative results than the criteria in Han (2005c);
Remark 9. Similar to Han (2005b), one can address robust Yue and Han (2004).
stability issue for neutral systems subject to parameter Table 1. Comparison of rmax using different
uncertainty. Due to page limit, it is omitted. methods
3. EXAMPLES Method β = −2.105 β = −2.103 β = −2.1
Han (2005c) 1.0874 0.3709 0.2433
Yue and Han (2004) 1.1413 0.3892 0.2553
In order to show significant improvements over some exist- Proposition 6 (N = 1) 1.0874 0.3709 0.2433
ing results in the literature, we employ a Partial Element Proposition 6 (N = 2) 1.5318 0.5185 0.3381
Equivalent Circuit (PEEC) model, which is recurring an Proposition 6 (N = 3) 1.6238 0.5490 0.3577
increasingly important role in many practical applications, Proposition 6 (N = 4) 1.6567 0.5599 0.3647
especially in combined electromagnetic and circuit analy-
sis. Figure 1 shows a small PEEC model for metal strip We now consider another example to show significant
(Bellen et al., 1999), in which there are delayed circuit improvements over some existing results in the literature.

2611
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Example 10. Consider the system (26) with neutral type. IEEE Trans. Circ. Syst.-I: Fundamental

−0.9 0.2
 
−1.1 −0.2
 Theory App., 46 (1): 212-216, 1999.
A= , Ad = , J.-D. Chen, C.-H. Lien, K.-K. Fan and J.-H. Chou. Criteria
0.1 −0.9 −0.1 −1.1
for asymptotic stability of a lass of neutral systems via a
−0.2 0
 
C= . LMI approach. IEE Proc. - Control Theory Appl., 148:
0.2 −0.1 442-227, 2001.
The analytical delay limit for stability can be calculated E. Fridman. New Lyapunov-Krasovskii functionals for
as ranalytical = 2.2255. Using some existing criteria and stability of linear retarded and neutral type systems.
Proposition 6 in this paper the maximum allowed time- Syst. Control Lett., 43 (4): 309-319, 2001.
delay rmax is compared in Table 2 . It is clear to see that E. Fridman and U. Shaked. A descriptor system approach
for this example the criterion in this paper can provide to H∞ control of linear time-delay systems. IEEE Trans.
much less conservative results than some existing ones Autom. Control, 47 (2): 253-270, 2002.
mentioned in Table 2. From this table, one can also see K. Gu. A further refinement of discretized Lyapunov func-
that as N increases, the results converge to the analytical tional method for the stability of time-delay systems.
delay limit for stability. Int. J. Control, 74 (10): 967-976, 2001.
K. Gu, V.L. Kharitonov, and J. Chen. Stability of time-
Table 2. Comparison of rmax using different delay systems. Birkhäuser, Boston, 2003.
methods J.K. Hale and S.M. Verduyn Lunel. Introduction to
Methods rmax
Functional Differential Equations. Springer-Verlag, New
Lien et al. (2000) 0.3 York, 1993.
Chen et al. (2001) 0.5658 Q.-L. Han. Robust stability of uncertain delay-differential
Fridman (2001) 0.74 systems of neutral type. Automatica, 38 (4): 719-723,
Lien and Chen (2003) 0.8844 2002.
Han (2004) 1.61 Q.-L. Han. A descriptor system approach to robust sta-
He et al. (2004a) 1.6527 bility of uncertain neutral systems with discrete and
Fridman and Shaked (2002) 1.7191 distributed delays. Automatica, 40 (8): 1791-1796, 2004.
Han (2005b) 1.7858 Q.-L. Han. Absolute stability of time-delay systems with
Proposition 6 (N = 1) 1.7858 sector-bounded nonlinearity. Automatica, 41 (12): 2171-
Proposition 6 (N = 2) 2.1077 2176, 2005.
Proposition 6 (N = 3) 2.1708 Q.-L. Han. A new delay-dependent stability criterion for
Proposition 6 (N = 4) 2.1932 linear neutral systems with norm-bounded uncertainties
Proposition 6 (N = 5) 2.2036
in all system matrices. Int. J. Systems Sci., 36 (8): 469-
Proposition 6 (N = 6) 2.2093
Proposition 6 (N = 7) 2.2121
475, 2005.
Proposition 6 (N = 8) 2.2149
Q.-L. Han. Stability analysis for a partial element equiva-
Proposition 6 (N = 9) 2.2164 lent circuit (PEEC) model of neutral type. Int. J. Circ.
Proposition 6 (N = 10) 2.2175 Theory App., 33: 321-332, 2005.
Proposition 6 (N = 15) 2.2201 Y. He, M. Wu, J.H. She, and G.P. Liu. Parameter-
Proposition 6 (N = 20) 2.2210 dependent Lyapunov functional for stability of time-
delay systems with polytopic-type uncertainties. IEEE
Trans. Autom. Control, 49 (5): 828-832, 2004.
4. CONCLUSION
Y. He, M. Wu, J.H. She and G.P. Liu. Delay-dependent
robust stability criteria for uncertain neutral systems
We have considered the problem of stability of linear with mixed delays. Syst. Control Lett., 51: 57-65, 2004.
neutral systems. The main contribution of this paper is C.-H. Lien and J.-D. Chen. Discrete-delay-independent
that we have proposed a delay decomposition approach to and discrete-delay-dependent criteria for a class of neu-
address the problem. The delay decomposition approach tral systems. ASME J. Dyn. Syst. Measu. Control, 125:
has opened a door in the area of time-delay systems how 33-41, 2003.
to use simple Lyapunov-Krasovskii functionals to derive C.-H. Lien, K. W. Yu, and J.G. Hsieh. Stability conditions
much less conservative results for stability analysis and for a class of neutral systems with multiple time delays.
controller design. For a stability issue, employing the J. Math. Anal. Appl., 245: 20-27, 2000.
delay decomposition approach, we have obtained some P.G. Park. A delay-dependent stability criterion for sys-
new stability criteria which are much less conservative tems with uncertain time-invariant delays. IEEE Trans.
than some existing ones using simple Lyapunov-Krasovskii Autom. Control, 44 (4): 876-877, 1999.
functionals. The stability limit can be approached with fine D. Yue and Q.-L. Han. A delay-dependent stability crite-
delay decomposition. rion of neutral systems and its application to a partial
It should be pointed out that one can use the delay element equivalent circuit model. IEEE Trans. Circ.
decomposition approach to address controller design and Syst.-II: Express Briefs, 51 (12): 685-689, 2004.
filtering for time-delay systems. Due to page limit, we have
left these issues in other papers.

REFERENCES
A. Bellen, N. Guglielmi, and A.E. Ruehli. Method for
linear systems of circuit delay differential equations of

2612