2172 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO.

9, SEPTEMBER 2011

where using 5= (In 0 Dcy D22 )01 [15] I. Masubuchi, A. Ohara, and N. Suda, “LMI -based output feedback
controller design using a convex parametrization of full-order con-
Ap + B2 5Dcy C2 B2 5Cc trollers,” in Proc. Amer. Control Conf., 1995, pp. 3473–3477.
A
Bcy (C2 + D22 5Dcy C2 ) Ac + Bcy D22 5Cc [16] S. Sajjadi-Kia and F. Jabbari, “Modified anti-windup compensators for
Cz =
stable linear systems,” IEEE Trans. Autom. Control, vol. 54, no. 8, pp.
C1 + D12 5Dcy C2 D12 5Cc 1934–1939, Aug. 2009.
Cu
5Dcy C2 5Cc
B1 + B2 Duw
Bw
Bcw + Bcy (D22 Duw + D21 )
Dzw =
D11 + D12 Duw
A New Model Transformation of Discrete-Time
Duw
Systems With Time-Varying Delay and
5Dcy D21 +5Dcw Its Application to Stability Analysis
Bq B
Xianwei Li and Huijun Gao, Senior Member, IEEE
Dzq Dz =
Duq Du Abstract—This technical note focuses on analyzing a new model transfor-
0B2 + B2 Duq B2 Du mation of uncertain linear discrete-time systems with time-varying delay
Bcy D22 Duq Bcy D22 Du In +[ 0n 2n ] and applying it to robust stability analysis. The uncertainty is assumed
to be norm-bounded and the delay intervally time-varying. A new com-
D12 Duq 0 D12 D12 Du
: parison model is proposed by employing a new approximation for delayed
state, and then lifting method and simple Lyapunov–Krasovskii functional
05Dcy D22 5[0n 2n In ] method are used to analyze the scaled small gain of this comparison model.
This new approximation results in a much smaller error than the existing
ones. Based on the scaled small gain theorem, new stability criteria are pro-
posed in terms of linear matrix inequalities. Moreover, it is shown that the
obtained conditions can be established through direct Lyapunov method.
REFERENCES Two numerical examples are presented to illustrate the effectiveness and
superiority of our results over the existing ones.
[1] G. Grimm, A. R. Teel, and L. Zaccarian, “Results on linear LMI -based
external anti-windup design,” in Proc. IEEE Conf. Decision Control, Index Terms—Asymptotic stability, discrete-time system, scaled small
2002, pp. 299–304. gain theorem, time-delay system.
[2] N. Kappor, A. R. Teel, and P. Daoutidis, “An anti-windup design for
linear systems with input saturation,” Automatica, vol. 34, no. 5, pp.
559–574, 1998. I. INTRODUCTION
[3] E. Mulder, M. V. Kothare, and M. Morari, “Multivariable anti-windup
controller synthesis using linear matrix inequalities,” Automatica, vol. Since practical systems, such as biology, mechanics and viscoelas-
37, no. 9, pp. 1407–1416, 2001. ticity, often encounter time-delay phenomena, the problems of time-
[4] G. Grimm, I. Hatfield, A. R. Teel, M. C. Turner, and L. Zaccarian, “An- delay systems have drawn considerable attention over the past decades,
tiwindup for stable linear systems with input saturation: An LMI -based such as controller design [1]–[3], state estimation [4]–[7], and espe-
synthesis,” IEEE Trans. Autom. Control, vol. 48, no. 9, pp. 1509–1525,
Sep. 2003.
cially the fundamental stability analysis [8]–[10]. Discrete-time sys-
[5] G. Herrmann, M. C. Turner, and I. Postlethwaite, “Some new results on tems have been fairly concerned due to their strong engineering back-
anti-windup-conditioning using the weston-postlethwaite approach,” ground [11]. As for the stability of discrete-time systems with a con-
in Proc. 43rd Conf. Decision Control, 2004, pp. 5047–5052. stant delay, a necessary and sufficient condition has been obtained [12]
[6] T. Hu, A. R. Teel, and L. Zaccarian, “Regional anti-windup compen-
by using the lifting method. For the time-varying delay case, a great
sation for linear systems with input saturation,” in Proc. Amer. Control
Conf., Portland, OR, 2005, pp. 3397–3402. number of stability criteria have been developed and the main concern
[7] L. Zaccarian and A. R. Teel, “Nonlinear scheduled anti-windup design is to reduce the conservatism of these conditions, especially, by the di-
for linear systems,” IEEE Trans. Autom. Control, vol. 49, no. 11, pp. rect Lyapunov method [13]–[18].
2055–2061, Nov. 2004. Another powerful but indirect research framework, borrowed from
[8] I. E. Kose and F. Jabbari, “Scheduled controllers for linear systems with
bounded actuators,” Automatica, vol. 39, no. 8, pp. 1377–1387, 2003.
the robust control theory [19], [20], is the input-output (IO) approach,
[9] S. Galeani, S. Onori, and L. Zaccarian, “Nonlinear scheduled control with which the stability of the original time-delay system is studied
for linear systems subject to saturation with application to anti-windup by way of that of the interconnection of two subsystems. This ap-
control,” in Proc. IEEE Conf. Decision Control, New Orleans, LA, proach was firstly introduced for constant delay [21], [22], and then
2007, pp. 1168–1173.
[10] F. Wu, K. M. Grigoriadis, and A. Packard, “Anti-windup controller
design using linear parameter-varying control methods,” Int. J. Control, Manuscript received September 21, 2010; revised January 09, 2011 and April
vol. 73, no. 12, pp. 1104–1114, 2000. 04, 2011; accepted April 15, 2011. Date of publication April 25, 2011; date of
[11] F. Wu and M. Soto, “Extended anti-windup control schemes for LTI current version September 08, 2011. This work was supported in part by Na-
and LPV systems with actuator saturations,” Int. J. Robust Nonlin. Con- tional Natural Science Foundation of China under Grants 60825303, 60834003,
trol, vol. 14, no. 15, pp. 1255–1281, 2004. and 61028008, in part by the 973 Project under Grant 2009CB320600 and in
[12] B. Lu, F. Wu, and S. W. Kim, “Linear parameter-varying antiwindup part by the Key Laboratory of Integrated Automation for the Process Industry,
compensation for enhanced flight control performance,” J. Guid., Con- Northeastern University, Ministry of Education, China. Recommended by As-
trol, Dyn., vol. 28, no. 3, pp. 494–505, 2005. sociate Editor F. Wu.
[13] S. Sajjadi-Kia and F. Jabbari, “Scheduled static anti-windup augmen- The authors are with the Research Institute of Intelligent Control and
tation synthesis for open-loop stable plants,” in Proc. Amer. Control Systems, Harbin Institute of Technology, Harbin 150001, China (e-mail:
Conf., Baltimore, MD, 2010, pp. 6751–6756. [email protected]; [email protected]).
[14] S. Sajjadi-Kia and F. Jabbari, “Multiple stage anti-windup agumenta- Color versions of one or more of the figures in this technical note are available
tion synthesis for open-loop stable plants,” in Proc. IEEE Conf. Desi- online at http://ieeexplore.ieee.org.
cion Control, Atlanta, GA, 2010, pp. 1281–1286. Digital Object Identifier 10.1109/TAC.2011.2146850

0018-9286/$26.00 © 2011 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011 2173

extended to the time-varying case [23]–[26]. A key to tackling time- formulation, and further through an IO approach, to seek new delay-
varying delay with this approach is to find a proper approximation dependent sufficient conditions for the robustly asymptotic stability of
for x(k 0 dk ) (for discrete-time systems) such that the approxima- system S for all dk 2 fh1 ; h1 + 1; 1 1 1 ; h2 g and 6(k) 2 6 .
tion error is as small as possible, where dk is the time-varying delay. The adopted IO approach actually is the application of the SSG the-
To mention a few, [23] adopted x(k 0 ((h1 + h2 )=2)) as the approxi- orem; for more details, we refer readers to [24, Ch.8]. Here, we just
mation of x(k 0 dk ) while [25] employed x(k 0 h1 ) with h1 and h2 recall the most elemental notion on the SSG theorem. Consider an in-
being the upper and lower bounds of delay, respectively. Very recently, terconnected system consisting of two subsystems
a two-term approximation idea, (x(t 0 h1 ) + x(t 0 h2 ))=2, was briefly
used for the continuous-time delayed systems in [27] and the resulting (S1 ) : y(t) = Gw(t); (S2 ) : (t) = 1 y (t) (3)
error was smaller than the one-term ones above. However, no result re-
where the forward S1 is a known linear time-invariant system (LTI)
garding the application of this approximation method in discrete-time
delayed systems has been reported, and moreover, [27] only considered
G
with operator mapping w to y , the feedback S2 is an unknown linear
the nominal systems, which motivate the research of this brief. y w
1
time-varying one with operator 1 2 D = f1 : k1k  1g and 1
In this technical note, a new model transformation will be analyzed y(t) 2 , w(t) 2 . As a direct result of the small gain theorem
and applied for the stability analysis of uncertain discrete-time systems [20], [28], [29], a sufficient condition for the robustly asymptotic sta-
with a time-varying delay in the state. One of the main technical con- bility of the interconnection in (3) is given as follows.
tributions is that a new comparison model is introduced by performing Lemma 1 (SSG Theorem): Consider (3) and assume S1 is internally
the approximation for x(k 0 dk ) as (x(k 0 h1 ) + x(k 0 h2 ))=2. Such
G 0 1
stable. The closed-loop system is robustly asymptotically stable for all
an approximation gives a much smaller upper bound of the l2 -induced 1 2 D if kTy   Tw 1 k < 1 holds for some matrices fTw ; Ty g 2
T with
fTw ; Ty g 2 w2w 2 y2y : Tw ; Ty
norm of the uncertainty resulted from the approximation error. Then,
the lifting method [12] and the Lyapunov–Krasovskii functional (LKF) T =1 nonsingular;
approach are employed to derive sufficient conditions for the scaled
small gain (SSG) of a subsystem so that the robustly asymptotic sta- Tw  1  Ty01 11 : (4)
bility of the original time-delay system is guaranteed. Moreover, we
show that the obtained criteria based on the IO approach can also be
established by applying the direct Lyapunov method to the original III. MAIN RESULTS
time-delay systems. Two examples will be given to illustrate the ad-
vantage of our results over the existing ones. In this section, we focus on transforming system S in (1) into the
G G
Notation: 1  2 represents the series connection of mapping 1 G interconnection of two subsystems as in (3) and then analyzing the SSG
G
and 2 . The superscript “T ” stands for matrix transpose. The notation of S1 . Consequently, stability criteria for S can be established by virtue
I
P > 0( 0) means that matrix P is positive (semi)definite, denotes an of the SSG theorem.
identity matrixwith appropriatedimension.Inblockmatrices, weuse“3”
A. A New Model Transformation
todenotethetermsthatcanbededucedbysymmetryand diagf1 1 1g stands
for a block-diagonal matrix. kxkl =
1 1 T
k=0 x (k)x(k) denotes the
Note that dk , 1A(k) and 1Ad (k) are all uncertain. The objective
1
l2 norm of series x(k) and k 1 k represents the l2 -induced norm of a of model transformation is to pull these uncertainties out of the original
time-delay system S in (1) so that S1 is an LTI while 1 includes all
transfer function matrix or a general operator.
the uncertainties.
To pull out the uncertainty in dk , we express x(k 0 dk ) as
II. PROBLEM FORMULATION AND PRELIMINARIES
1 h12
As widely investigated in [14], [16] and [12] et al., we consider the x(k 0 dk ) = [x(k 0 h1 ) + x(k 0 h2 )] + wd (k) (5)
2 2
following uncertain linear discrete-time system S with an interval time
delay in the state: where (1=2)[x(k 0 h1 ) + x(k 0 h2 )] is regarded as the approximation
1
of x(k 0 dk ) and (h12 =2)wd (k) is the approximation error with h12 =
(S ) : x(k + 1) = (A + 1A(k)) x(k) h2 0 h1 . By (5), system S can be written as
+ (Ad + 1Ad (k )) x(k 0 dk ); 1
x(k) = '(k); k = 0h2 ; 0h2 + 1; 1 1 1 ; 0: (1)
x(k + 1) = (A +1A(k)) x(k) + (Ad +1Ad (k ))
2
h12
2
Here, x(k) 2 n is the state vector; A; Ad 2 n n are constant
2[x(k 0 h1 )+ x(k 0 h2 )] +
2
(Ad +1Ad (k )) wd (k ): (6)
nominal system matrices and 1A(k) and 1Ad (k) are time-varying 1
real matrix functions; dk is the time delay, time-varying and satisfying By calculation and defining yd (k) = x(k + 1) 0 x(k), it is easily
1  h1  dk  h2 , where h1 and h2 are known constants; f(k ); k = found that
0h2 ; 0h2 + 1; . . . ; 0g is the given initial condition sequence. 2 1
1A(k ) and 1Ad (k ) are assumed to be of the usual norm-bounded wd (k) = x(k 0 dk ) 0 [x(k 0 h1 ) + x(k 0 h2 )]
h12 2
type as [1A(k) 1Ad (k)] = E 6(k)[F Fd ], where 6(k) is a real un-
certain matrix function satisfying 6(k) 2 6 with 1
0 01
k d 0 01
k h
= yd (i) 0 yd (i)
h12 0 0
1
6=
T
I
6(k ) : 6 (k )6(k )  m ; the elements of 6(k ) are
i=k h
0 0
k h 1
i=k d

Lebesgue measurableg 1 1
(2) = (i)yd (i) (7)
h12 0
2 2
and E 2 n l , F; Fd 2 m n are known matrices representing the
i=k h

structure of uncertainty. where

The objective of this technical note is to introduce and analyze a new 1 1; when i  k 0 dk 0 1
model transformation which converts system S into the interconnection
(i) =
01; when i > k 0 dk 0 1.
2174 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011

1
For brevity, operator d is used to denote the mapping d yd 1 : ! wd B. SSG Analysis and Robust Stability Conditions
in (7). Now, (6) has only two known constant delays and the uncertainty
1
in dk has been moved into d . Furthermore, the following result shows
1 in (9), we specify the matrix set T in (4) as
As for
1
that the l2 -induced norm of d is bounded by one.
T=
1
ffTw ; Ty g : Tw = diagfT; Ilg; Ty = diagfT; Img;
1 : ! k1 k1  1 T 2 n2n nonsingular; scalar  6= 0
Lemma 2: Operator d yd wd satisfies d .
Proof: By using Jensen inequality [30] and exchanging the sum (10)
order, we have, under the zero initial condition
S 0 k1
k  
for simplification. Note that 1 in (9) still has two state delays,
1 i0h 01 i0h 01 1
kwdkl2 = h1122
thus testing the satisfaction of Ty G Tw 1 < cannot be
(j )ydT (j ) (j )yd (j ) solved directly by the bounded real lemma in [31]. However, since
i=0 j =i0h j =i0h
S
delays are both constant, the lifting method [12] can be employed to
1 i0h 01
 h1122 (h2 0 h1) 2 (j )ydT (j )yd (j )
further convert 1 to be delay-free; moreover, one can choose some

k   0 k1
simple but efficient LKFs to derive sufficient LMI conditions ensuring
i=0 j =i0h Ty G Tw 1 0
< for a given real > . Accordingly, we have

= h112
0h 01 1 T S
yd (i + j )yd (i + j ) S
two LMI methods testing the SSG of 1 .

j =0h i=0 k  
stable and satisfies Ty G Tw 1 00k1
Theorem 1: Given a scalar > , system 1 in (9) is asymptotically
< for some Tw ; Ty T f g2
0h 01 1 T
 h112 yd (i)yd (i) = kyd kl2
in (10) if either of the following two conditions holds:
(i) There exist symmetric matrices < P 0 2
(h +1)n (h +1)n 2
j =0h i=0 2
2
,
0 <S n n 0
and a scalar " > such that
which implies k1d k1  1. This completes the proof. 2 0 22 < 0
GT 1 G (11)
Further by introducing
where 21 = diagfP; S; "Im g, 22 = diagfP; g
( ) = F x(k) + 12 Fd [x(k 0 h1 ) + x(k 0 h2 )]
2
S; " 2 Il and
y6 k

+ h212 Fd wd (k); A (0)1 1A

2 d (0)2 1A
2 d
h
2 Ad E
Ih 0 0 0
w6 (k) = 6(k)y6 (k)
n
(8) G = 0
A I (0)1 (0)2
1A 1A ; h
2 d 2 d 2 Ad E
the uncertainty in 1A(k) and 1Ad (k) can be pulled out and (6) can F (0)3 1(0)4 1 h
2 Fd 0
2 Fd 2 Fd
be simply rewritten as (0)1 = 0n2(h 01)n; (0)2 = 0n2(h 01)n;
( + 1) = Ax(k) + 21 Ad [x(k 0 h1 ) + x(k 0 h2 )]
x k
(0)3 = 0m2(h 01)n ; (0)4 = 0m2(h 01)n :
(ii) There exist symmetric matrices 0 < P 2 n2n , 0 < Qi 2
+ h212 Ad wd (k) + Ew6 (k): n2n
, 0 < Ri 2 n2n , i = 1, 2, 0 < S 2 n2n and a scalar
" > 0 such that

S
This together with (7) and (8) ultimately composes the interconnection

S
formulation of system in (1).
41 [ 42T P h1 43T R1 h2 43T R2 43T S "44T ] 0
2f + 1 111 g 6( ) 2 6
Proposition 1: System in (1) is robustly asymptotically stable for
all dk h1 ; h1 ; ; h2 and k if the interconnection of 3 diagf0P; 0R1 ; 0R2 ; 0S; 0"Ig < (12)
the following two subsystems is robustly asymptotically stable

81
where
( + 1)
x k
h
2 Ad E ()
 k
(S1 ) : yd (k) = 82 h
2 Ad E wd (k) ; 41 = 4311 diagf4[ R121; 413R;20 02 S;00] " 2 Ig ;
y6 (k) 83 w6 (k)
411 = 0 P + Q1 + Q2 0 R1 0 R2 ;
h
2 Fd 0
G 412 = 0 Q1 0 R1 ; 413 = 0Q2 0 R2 ;
(S2 ) : ( ) = 1d [1] 0
wd k yd k()
()
w6 k 0 6(k) y6 k() (9) 42 = [ A 12 Ad 12 Ad h2 Ad E ] ;
1 43 = [ A 0 I 12 Ad 12 Ad Ad E ] ;
where
44 = [ F 12 Fd 12 Fd h2 Fd 0 ] :

81 = A
1 Ad 1 Ad ; 82 = A 0 I 1 Ad 1 Ad ; Proof: (Proof of (i)) Denoting
2 2 2 2 ( ) = col fx(k); x(k 0 1); 111 ; x(k 0 h2 )g
x k
1
83 = F 2 Fd 2 Fd ; 1
and using the lifting method in [12], it is easily verified that 1 in (9) S
 (k) = col fx(k); x(k 0 h1 ); x(k 0 h2 )g : is equivalent to the following augmented state-space model:

In light of the assumption of 6(k) in (2) and Lemma 2, it follows

S
that the l2 -induced norm of 2 in (9) from input to output is bounded ( + 1)
x k (k)
x
yd (k) = G wd (k) : (13)
S
by one. Thus based on Lemma 1, in the following, we naturally focus
on studying the internal stability and SSG of system 1 in (9). y6 (k) w6 (k)
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011 2175

Similarly, denote GT as the augmented state-space parameter of the Now, by combining Theorem 1 with Lemmas 1, 2 and Proposition
G
series Ty   Tw01 , and it is obvious that 1, it is in a position to present two sufficient conditions ensuring the
robustly asymptotic stability of system S in (1) as follows.
GT = diag fIh n ; Ty g Gdiag Ih n ; Tw01 : Theorem 2: System S in (1) is robustly asymptotically stable for all
dk 2 fh1 ; h1 + 1; 1 1 1 ; h2 g and 6(k) 2 6 in (2) if
From Lemma 5.1 in [32], S1 is stable and kTy  G  Tw01k1 < if (i) Theorem 1-(i) holds for  1;
and only if (ii) or Theorem 1-(ii) holds for  1.
For system S without uncertainty in the system matrices, i.e.
GTT diagfP; In+m gGT 0 diagfP; 2 In+l g < 0
(S~) : x(k + 1) = Ax(k) + Ad x(k 0 dk ) (17)
holds for some P > 0, which is equivalent to (11) by pre- and post-
I I
multiplying it by diagf h n ; TwT g and diagf h n ; Tw g, respectively,
and letting S = T T T and " =  2 .
the interconnection in (9) is reduced as follows:

(Proof of (ii)) Instead of the state lifting method, consider a simple x(k + 1) 81 h Ad (k)
LKF candidate for S1 in (9) as follows: (S~1 ) : = 2
;
yd (k) 82 h
2 Ad wd (k)
V (k) = V1 (k) + V2 (k) + V3 (k) (14)
G~
where (S~2 ) : wd (k) = 1d [yd (k)] :: (18)
2 k01
V1 (k) = xT (k)Px(k); V2 (k) = xT (i)Ql x(i); Accordingly, the following results regarding the internal stability and
l=1 i=k0h the SSG of S~1 and the asymptotic stability of S~ can be deduced from
2 01 k01
V3 (k) = hl ydT (j)Rl yd (j) Theorem 1 and Theorem 2, respectively.
Corollary 1: Given a scalar > 0, system S~1 in (18) is asymptoti-
cally stable and satisfies kT  ~  T 01 k1 < for some nonsingular
l=1 i=0h j =k+i
G
with positive matrices P , Qi and Ri , i = 1, 2, to be determined. Taking T 2 n2n if either of the following two conditions (holds:
the forward difference of V (k) and using Jensen inequality [30], it can (i) There exist symmetric matrices 0 < P 2 h +1)n2(h +1)n ,
be obtained that under the zero input 0 < S 2 n2n such that
1V (k)  T (k) 4^ 1 +81T P81 +82T h12 R1 +h22R2 82 (k) (15) G~ T 2~ 1 G~ 0 2~ 2 < 0
where 2 ~ 1 = diagfP; S g, 2~ 2 = diagfP; 2 S g and
^1
where 4 = 4311 diag[Rf41 12R;4
2]
13 g
and 81 ,
82 and (k) are de- A (0)1 12 Ad (0)2 12 Ad h A
d
fined in (9). By the Schur complement, (12) implies that 1V (k) < 0,
2

i.e., system S1 is asymptotically stable.

~G = Ih n 0 0
Let S > 0, " > 0 and A 0 I (0)1 12 Ad (0)2 21 Ad h A
2 d
1
with (0)1 and (0)2 in (12);
J =1 ydT (k)Syd (k) 0 2 wdT (k)Swd (k) (ii) There exist symmetric matrices 0 < P 2 n2n , 0 < Qi 2
k=0 n2n , 0 < R 2 n2n , i = 1, 2, and 0 < S 2 n2n such that
i
+"y6T (k)y6 (k) 0 " 2 w6T (k)w6 (k) :
4~ 1 [ 4~ 2 P h14~ 3T R1 h2 4~ 3T R2 4~ 3T S ] < 0
T
Again, taking the forward difference of V (k) and using Jensen in- 3 diagf0P; 0R1 ; 0R2 ; 0S g
equality [30], we have, under zero initial condition where

4~ 1 = 4311 diag[fR4112 ; 4R132 ; 00 ] 2 S g ;

1
J  J + V (1) 0 V (0) = [J(k) + 1V1 + 1V2 + 1V3]
1
k=0
4~ 2 = [ A 12 Ad 12 Ad h2 Ad ] ;
  T (k) 41 + 42T P42 + 43T h12 R1 + h22 R2 43 4~ 3 = [ A 0 I 12 Ad 12 Ad h2 Ad ]
k=0
+43T S43 + "44T 44 (k) (16) and 411 , 412 and 413 are defined in (12).
Corollary 2: System S~ in (17) is asymptotically stable for all dk 2
where fh1 ; h1 + 1; 1 1 1 ; h2 g if
(k) = col fx(k); x(k 0 d1 ); x(k 0 d2 ); wd (k);w6 (k)g : (i) Corollary 1-(i) holds for  1;
(ii) or Corollary 1-(ii) holds for  1.
From the definition of J , if letting S = T T T and " =  2 with 1 > 0, Remark 2: With the idea of “pulling out uncertainties,” it is easy
G
J < 0 means that kTy   Tw01k1 < . By the Schur comple- to extend our results to multiple time-varying delay case, for which,
ment, it follows from (16) that (12) is a sufficient condition ensuring the uncertainty of each delayed state will be pulled out and forms a
G
kTy   Tw01 k1 < . The proof is completed. feedback channel similar to the diagram in Fig. 1. in [21]. Besides the
Remark 1: The condition (ii) is conservative for testing the internal uncertainties included in delay and system matrices, the input-output
stability and the given SSG level of S1 . Thus results deduced from (ii) performance of the original system can also be considered in such a
will be more conservative than that from (i). However, the number of formulation, such as the standard l2 -induced norm gain [23]. This fa-
variable (NoV) of (i) depends on h2 and generally is larger than that of cilitates to extend our results to many other analysis and synthesis prob-
(ii); hence, (ii) is expected to have lighter computational burden. lems of delayed systems.
2176 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011

time-domain sufficient condition guaranteeing the SSG condition kT 

G~  T 01k1 < 1 for S~1 .
Remark 3: A special case is to choose V (k) as in (14), then
Theorem 3 shows that Corollary 2-(ii) can be obtained by the di-
rect Lyapunov method. Note that by this method, free-weighting
matrices are often introduced to help estimate an upper bound of
0(1=h12 ) i0=h00 1 T
h yd (k + i)Syd (k + i) [14]–[18], while we do
not need any such weighting matrix in (20). Hence, the NoV of
Corollary 2-(ii) has been remarkably reduced to 3n2 + 3n, which is
much smaller than, respectively, 11:5n2 + 2:5n in [14], 9n2 + 3n in
[18], 13n2 + 5n in [15], [16] and 31n2 + 7n in [17] (m = 1). The
computational advantage is quite obvious.
2) Comparison of Different Approximation Methods: At first,
we compare the approximation errors of some other approximation
methods with that of ours in detail. In [25] where h1 = 0 was assumed,
the adopted approximation was x(k 0 dk ) ' x(k) = x(k 0 h1 ) and
then an upper bound of the l2 -induced norm of the operator 1d was
exposed as k1d k1  h2 = h12 in Lemma 2 in [25]. A similar treat-
Fig. 1. versus when = 16. ment in the continuous-time case can also be found in [25] and [22]. In
[23], another better approximation x(k 0 dk ) ' x(k 0 (h1 + h2 )=2)
(assuming (h1 + h2 )=2 is an integer) was introduced and the given p
C. Some Further Comments upper bound of the l2 -induced norm of 1d was k1d k1  (h12 = 2)
1) Direct Lyapunov Method: Although the basic research tools at- which could be deduced from Theorem 2.3 in [23]. However, if no
tacking the stability of system S in (1) are model transformation and normalization is performed to the error wd (k) in (6), i.e., the coeffi-
the SSG theorem, we could show, similar to [27], that the stability cri- cient h12 =2 before it was moved into it, the upper bound in Lemma 2
teria presented can also be obtained by applying the direct Lyapunov would be restated as k1d k1  (h12 =2). Therefore, the “size” of the
method. For brevity, only consider the nominal system S~ in (17). error for the approximation in (5) is quantitatively smaller than those
Theorem 3: For the subsystem S~1 in (18), if there exists an LKF in some existing literature.
V (k) and S > 0 satisfying As the fundamental purpose, to apply Theorem 2 or Corollary 2 to
test the stability of system S is to check whether the conditions in these
1V (k) + ydT (k)Syd (k) 0 wdT (k)Swd (k) < 0 (19) results are satisfied for a given  1 or not. Hence, the manner of mea-

then Vcl (k) = V (k) + (1=h12 ) i0=h0h01 k01 y T (j )Sy (j ) is an

suring the improvement of a approximation method could be, for given
j =k+i d d delay bounds, to compare the minimum allowable with that by other
LKF candidate to directly demonstrate the asymptotic stability of the methods, or for a given , to compare the allowable delay bounds with
closed-loop system formed by S~1 and S~2 , or that of S~. that by other methods (of course, before carrying out the comparison,
Proof: By virtue of Jensen inequality, taking the forward differ- the l2 -induced norm of 1d should be normalized). Fortunately, by the
ence of Vcl (k) along the trajectory of closed-loop system consisting of lifting method, it is convenient to numerically calculate the allowable
S1 and S2 , we have based on, for example, Corollary 1-(i). For facilitating comparison,
1
0h 01 let S~h and S~h denote the counterparts of S~1 in (18) but the approx-
1Vcl (k) = 1V (k) + yT (k)Syd (k) imation methods are, respectively, x(k 0 h1 ) in [25] and x(k 0 hm )
h12 i=0h d in [23], where hm = [(h1 + h2 )=2] denotes the maximum integer less
0ydT (k + i)Syd (k + i) than or equal to (h1 + h2 )=2. Similar to Corollary 1-(i), we have the
following results.
= 1V (k) + ydT (k)Syd (k) Lemma 3: Given a scalar > 0, S~h (S~h ) is asymptotically stable
1
0h 01 G
and satisfies kT  ~  T 01 k1 < for some nonsingular T 2 n2n if
0 yT (k + i)Syd (k + i) there exist symmetric matrices 0 < P 2 (h +1)n2(h +1)n (0 < P 2
h12 i=0h d (h +1)n2(h +1)n
), 0 < S 2 n2n such that G ~ T diagfP; S gG~0
= 1V (k) + ydT (k)Syd (k)
2
diagfP; S g < 0 is satisfied, where
1
k0h 01 A 0 n2(h 01)n Ad h12 Ad
0
h12 i=k0h
(i)ydT (i)S(i)yd (i) G~ = I h n 0 h n2n 0 h n2n
A 0 I 0n2(h 01)n Ad h12 Ad
 1V (k) + ydT (k)Syd (k) ~=
(G
0 2
1
k0h 01
(i)yd (i) S
T
A 0n2(h 01)n Ad h12 (h2 0 hm )Ad
h12 i=k0h Ih n 0h n2n 0h n2n ):
k0h 01 A 0 I 0n2(h 01)n Ad h12 (h2 0 hm )Ad
2 (i)yd (i)
i=k0h In the later section, we will present some numerical comparisons,
= 1V (k) + ydT (k)Syd (k) 0 wdT (k)Swd (k): (20) which clearly show the great improvement of the approximation
method and the deduced stability criteria in the technical note.
From (19) and (20), we have 1Vcl (k) < 0, which means that the However, theoretically thoroughly comparing the SSGs of S1 or S~1
closed-loop system is asymptotically stable based on the Lyapunov resulting from different approximation methods is fairly challenging.
stability theory. The proof can be completed by noting that (19) is a In addition, due to h1  hm  h2 , an obvious fact is that the number
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011 2177

of variables of Lemma 3 for S~h , Lemma 3 for S~h and Corollary

1-(i) for S~1 are ranked in the increasing order. Hence, the results
deduced from the two-term approximation method generally have
heavier computational burden than that from the existing one-term ap-
proximation ones. However, this case is not always true for Corollary
1-(ii) since its NoV is independent of h1 and h2 . These comments will
be clearly illustrated by the numerical results, too.
Remark 4: h1 = 0 is assumed for Theorem 2 in [25], which seems to
be contradictive from h1 > 0 for our approximation method. However,
note that the key technical step in [25] is to approximate x(k 0 dk ) by
x(k), and due to that x(k) = x(k 0 h1 ) when h1 = 0, the underlying
idea in [25] indeed is to approximate x(k 0 dk ) by x(k 0 h1 ). This
point reveals that the approximation method in [25] can be immediately
extended to the case of h1 > 0 (this case has already been briefly
discussed for single-input-single-output systems in [25, Sec.4.2]), and
the first condition in Lemma 3 is developed right for this case.

IV. ILLUSTRATIVE EXAMPLES

Example 1: Consider the nominal system in (17) as investigated in Fig. 2. versus when = 7.
[14], [16] with
0:80 0:00 00:10 0:00
A= ; Ad = : (21) TABLE I
0:05 0:90 00:20 00:10 COMPARISON OF COMPUTATIONAL COMPLEXITY WITH 8 14
We use this example to illustrate the improvement of our approximation
method and its deduced stability criteria.
Firstly, under the same delay setting, the SSG level of S~1 , S~h
and S~h are calculated by the corresponding Corollary 1 or Lemma 3.
The relationships of versus h1 when h2 = 16 and versus h2 when
h2 = 7 are depicted in Figs. 1 and 2, respectively. From these figures,
one can attain the following observations:
1) S~1 based on our approximation method has much smaller SSG TABLE II
level . For example, when h1 = 10 and h2 = 16 in Fig. 1, CALCULATED UPPER DELAY BOUND FOR DIFFERENT
of S~1 by Corollary 1-(i) is 0.58 while that of S~h and S~h by
Lemma 3 are 1.07 and 0.84, respectively.
2) Moreover, the labeled points show that the stability criterion based
on our approximation method can guarantee larger delay ranges
and thus is less conservative. For example, Fig. 2 shows that for
given h1 = 7, the biggest allowable h1 when (21) is stable is 21
by our method, much larger than 13 and 14 by Lemma 3 for S~h
and S~h , respectively.
3) In addition, numerical results by Corollary 1-(ii) are also drawn in
the figures. Although it is only sufficient for ensuring , it is shown
that this condition is also less conservative than Lemma 3 for S~h
and S~h , which further demonstrates the great improvement of
demonstrates the lower conservatism of this condition. Moreover, the
NoV of Corollary 2-(ii), emphasized by bold number, is only 18, which
our approximation method.
is much fewer than that of some up-to-date methods in [15]–[17].
To compare the computational complexity, Table I lists the NoV and
Example 2: We consider the uncertain time-delay system in (1) bor-
the cost CPU time of different methods for 8  dk  14. It is readily
rowed from [14] with
found that Corollary 1-(i) needs more variables and time to compute
. Generally speaking, the improvement of our approximation method 0:80 0:00 00:10 0:00 
A= ; Ad = ;E = ;
is at the cost of the growth of NoV and computational time. However, 0:00 0:90 00:10 00:10 0
noting the last row of Table I, one knows that through combining our F = [ 1:00 0:00 ]; Fd = [ 0:00 0:00 ]; 1(k) = (k)=  (22)
approximation method with a simple LKF, Corollary 1-(ii) also can
achieve smaller while the needed variables are much fewer than those where j (k)j   and  is given. Obviously, 1T (k)1(k)  1.
in the literature. We desire to find the maximum allowable upper bound (MAUB)
Here, we further compare our stability criteria in Corollary 2 with of  such that system (22) is robustly asymptotically stable for all
some latest ones that is derived by direct Lyapunov method [13]–[18]. j (k)j   with given delay interval d1  dk  d2 . Let 2  dk  7,
For a given d1 , we calculate the maximum allowable upper bound d2 by the MAUBs are  = 0:0830 and  = 0:1901 obtained by [13] and
these methods in the literature and in the technical note. When d1 = 8, [14], respectively. However, using Theorem 2-(i) and (ii) in this tech-
one obtains d2 = 10 by [13], d2 = 14 by [14] and d2 = 15 by [18], nical note yields, respectively,  = 0:1938 and  = 0:2046. More re-
respectively. While using Corollary 2, we obtain d2 = 23 for (i) and 19 sults for different delay intervals are included in Table III, which again
for (ii), respectively. For comparison, more detailed calculated results clearly shows the lower conservatism of our methods. Especially, when
for different d1 are listed in Table II, where italic numbers, d2 by Corol- 8  dk  15, Theorem 2 still can find some allowable  while the ex-
lary 2-(i), are much larger than the ones by the existing methods, which isting methods fail.
2178 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 9, SEPTEMBER 2011

TABLE III [13] H. Gao, J. Lam, C. Wang, and Y. Wang, “Delay-dependent output-
CALCULATED MAXIMUM ALLOWABLE  FOR DIFFERENT DELAY CASES feedback stabilisation of discrete-time systems with time-varying state
delay,” Proc. Inst. Elect. Eng., vol. 151, pp. 691–698, 2004.
[14] H. Gao and T. Chen, “New results on stability of discrete-time systems
with time-varying state delay,” IEEE Trans. Autom. Control, vol. 52,
no. 2, pp. 328–334, Feb. 2007.
[15] Y. He, G. Liu, D. Rees, and M. Wu, “ filtering for discrete-time
systems with time-varying delay,” Signal Process., vol. 89, no. 3, pp.
275–282, Mar. 2009.
[16] Y. He, M. Wu, G.-P. Liu, and J.-H. She, “Output feedback stabilization
for a discrete-time systems with a time-varying delay,” IEEE Trans.
Autom. Control, vol. 53, no. 10, pp. 2372–2377, Nov. 2008.
[17] X. Meng, J. Lam, B. Du, and H. Gao, “A delay-partitioning approach
V. CONCLUSION to the stability analysis of discrete-time systems,” Automatica, vol. 46,
no. 3, pp. 610–614, Mar. 2010.
In this technical note, model transformation and robust stability anal- [18] B. Zhang, S. Xu, and Y. Zou, “Improved stability criterion and its ap-
ysis for discrete-time systems with a time-varying delay in the state plication in delayed controller design for discrete-time systems,” Auto-
have been investigated. A new comparison model has been presented matica, vol. 44, no. 11, pp. 2963–2967, 2008.
[19] J. Doyle, “Analysis of feedback systems with structured uncertainties,”
by employing a novel approximation for delayed state with smaller ap-
Proc. Inst. Elect. Eng. D, vol. 129, no. 6, pp. 242–250, 1982.
proximation error than the existing ones. By applying an input-output [20] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control.
approach based on scaled small gain theorem, new delay-range-depen- Englewood Cliffs, NJ.: Prentice-Hall, 1996.
dent sufficient conditions have been proposed in term of a set of LMIs. [21] Y.-P. Huang and K. Zhou, “Robust stability of uncertain time-delay
In addition, it is shown that the obtained results can be derived by the systems,” IEEE Trans. Autom. Control, vol. 45, no. 11, pp. 2169–2173,
Nov. 2000.
direct Lyapunov method. The given numerical examples have demon- [22] J. Zhang, C. R. Knopse, and P. Tsiotras, “Stability of time-delay sys-
strated the advantage in conservatism and computational complexity of tems: Equivalence between Lyapunov and scaled small-gain condi-
our methods. tions,” IEEE Trans. Autom. Control, vol. 46, no. 3, pp. 482–486, Mar.
2001.
[23] E. Fridman and U. Shaked, “Input-output approach to stability and
-gain analysis of systems with time-varying delays,” Syst. Control
ACKNOWLEDGMENT Lett., vol. 55, pp. 1041–1053, 2007.
[24] K. Gu, V. L. Kharitonov, and J. Chen, Stability of Time-Delay Sys-
The authors are very indebted to the Associate Editor and the anony- tems. Berlin, Germany: Springer-Verlag, 2003.
mous reviewers for their insightful comments and valuable suggestions [25] C.-Y. Kao and B. Lincoln, “Simple stability criteria for systems with
that have helped improve the technical note. time-varying delays,” Automatica, vol. 40, no. 8, pp. 1429–1434, 2004.
[26] C.-Y. Kao and A. Rantzer, “Stability analysis of systems with uncertain
time-varying delays,” Automatica, vol. 43, no. 6, pp. 959–970, Jun.
2007.
REFERENCES [27] K. Gu, Y. Zhang, and S. Xu, “Small gain problem in coupled differ-
ential-difference equations, time-varying delays, and direct Lyapunov
[1] P. Shi, E. K. Boukas, and R. K. Agarwal, “Control of Markovian jump method,” Int. J. Robust Nonlin. Control, vol. 21, no. 4, pp. 429–451,
discrete-time systems with norm bounded uncertainty and unknown 2011.
delay,” IEEE Trans. Autom. Control, vol. 44, no. 11, pp. 2139–2144, [28] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output
Nov. 1999. Properties. New York: Academic, 1975.
[2] F. Wu, “Robust quadratic performance for time-delayed uncertain [29] G. Zames, “On the input-output stability of time-varying nonlinear
linear systems,” Int. J. Robust Nonlin. Control, vol. 13, no. 2, pp. feedback systems-part I: Conditions derived using concepts of loop
153–172, Feb. 2003. gain, conicity, and positivity,” IEEE Trans. Autom. Control, vol. AC-2,
[3] F. Wu and K. M. Grigoriadis, “LPV systems with parameter-varying no. 2, pp. 228–238, Apr. 1966.
time delays: Analysis and control,” Automatica, vol. 37, no. 2, pp. [30] K. Gu, “An integral inequality in the stability problem of time-delay
221–229, Feb. 2001. systems,” in Proc. 39th IEEE Conf. Dec. Control, Sydney, Australia,
[4] P. Shi, “Robust filter design for sampled-data systems with intercon- 2000, pp. 2805–2810.
nections,” Signal Process., vol. 58, no. 2, pp. 131–151, Apr. 1997. [31] P. Apkarian and P. Gahinet, “A convex characterization of gain-sched-
[5] P. Shi, “Filtering on sampled-data systems with parametric uncer- uled controllers,” IEEE Trans. Autom. Control, vol. 40, no. 5, pp.
tainty,” IEEE Trans. Autom. Control, vol. 43, no. 7, pp. 1022–1027, 853–864, May 1995.
Jul. 1998. [32] P. Gahinet and P. Apkarian, “A linear matrix inequality approach to
[6] Z. Wang, B. Huang, and H. Unbehauen, “Robust observer design control,” Int. J. Robust Nonlin. Control, vol. 4, no. 4, pp. 421–448,
of linear state delayed systems with parametric uncertainty: The dis- Jul./Aug. 1994.
crete-time case,” Automatica, vol. 35, no. 6, pp. 1161–1167, 1999.
[7] X. Li, Z. Li, and H. Gao, “Further results on filtering for discrete-
time systems with state delay,” Int. J. Robust Nonlin. Control, vol. 21,
no. 3, pp. 248–270, Feb. 2011.
[8] K. Gu, “A further refinement of discretized lyapunov functional
method for the stability of time-delay systems,” Int. J. Control, vol.
74, no. 10, pp. 967–976, 2001.
[9] Y. S. Moon, P. Park, W. H. Kwon, and Y. S. Lee, “Delay-dependent
robust stabilization of uncertain state-delayed systems,” Int. J. Control,
vol. 74, no. 14, pp. 1447–1455, Sep. 2001.
[10] R. Wang and J. Zhao, “Exponential stability analysis for discrete-time
switched linear systems with time-delay,” Int. J. Innov. Comp. Inf. Con-
trol, vol. 3, no. 6B, pp. 1557–1564, Dec. 2007.
[11] A. Benzaouia and F. Tadeo, “Stabilization of positive switching linear
discrete-time systems,” Int. J. Innov. Comp. Inf. Control, vol. 6, no. 6,
pp. 2427–2438, Jun. 2010.
[12] Y. Xia, G. P. Liu, P. Shi, D. Rees, and E. J. C. Thomas, “New stability
and stabilzation conditions for systems with time-delay,” Int. J. Syst.
Sci., vol. 38, no. 1, pp. 17–24, 2007.