$binder 1

C:\work2011\shamov\yulia7\00704983.
pdf
1110 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 8, AUGUST 1998
Hence the feedback interconnection of two stable subsystems with bounding functions '1 and '2 is stable since the internal signals can be bounded for any bounded exogenous input. V. CONCLUSION A boundedness result for nonlinear unity-feedback interconnections with stable subsystems is introduced in Theorem 1. Each subsystem is associated with a nondecreasing bounding function. The result reduces to translating two curves that denote the boundaries of feasible regions and seeking bounded intersections. No further analysis or construction of functions other than the original pair of bounding functions is required. The application of this result has a simple two-dimensional graphical interpretation. REFERENCES
[1] C. A. Desoer and M. Vidyasagar, Feedback Systems: InputOutput Properties. New York: Academic Press, 1975. [2] C. A. Desoer and C. A. Lin, Nonlinear unity feedback systems and parametrization, Analysis and Optimization of Systems, Lecture Notes in Control and Information Sciences, A. Bensoussan and J. L. Lions, Eds. Berlin: Springer Verlag, vol. 62, 1984. [3] D. J. Hill, A generalization of the small-gain theorem for nonlinear feedback systems, Automatica, vol. 27, no. 6, pp. 10431045, 1991. [4] A. R. Teel, On graphs, conic relations, and inputoutput stability of nonlinear feedback systems, IEEE Trans. Automat. Contr., vol. 41, pp. 702709, May 1996. [5] M. Vidyasagar, Nonlinear Systems Analysis, 2nd ed. Englewood Cliffs, NJ: Prentice Hall, 1993.
Consider the system described by the equations
x = Ax + Bu; _ y = Cx (1) where x is the state vector, u is the control vector, y is the measurement vector, and A; B; and C are constant matrices. Let Ei and Ek denote the orthogonal projection matrices on the Im C T and Ker C , respectively, i.e., Ei = C + C; Ek = I 0 Ei (2)
and the Hamiltonian matrix be dened as
01 B T A 0 BR01 H (S ) = Q 0 S T R01S 0(ABRBR01 S )T : S 0
(3)
Theorem 3.1 of the paper is stated as follows. Let Ei and Ek be the projection matrices dened in (2) and H (s) the Hamiltonian matrix in (3). Then, the following statements are equivalent. i) System (1) is stabilizable via static output feedback. ii) There exist matrices Q > 0; R > 0; and L of compatible dimensions such that the following algebraic matrix equation: AT P + P A 0 Ei (PB + LT )R01 (B T P + L)Ei + Q = 0 (4) has a unique solution P > 0 and K = R01 (L + B T P )Ei is stabilizing. iii) The pair (A; B ) is stabilizable (by state feedback) and there exist P > 0; Q > 0; R > 0; and a matrix L in (4) such that the Hamiltonian matrix in (3) has no pure imaginary eigenvalues for S = LEi 0 B T P Ek . First, let us see a counterexample. Consider the system (1) with
A=
A New Necessary and Sufcient Condition for Static Output Feedback Stabilizability and Comments on Stabilization via Static Output Feedback
Yong-Yan Cao, You-Xian Sun, and Wei-Jie Mao
which is not stable since its two eigenvalues are 1 and 01. It is not difcult to nd that this system cannot be stabilized via static output feedback when a > 0, while it can be when a < 0. Let a = 0:5 and
0 1
1 0
B=
1 0
C = [1 0a]
R = 1;
L = [0:2 P
=
1 7]
: ;
5: :88 Q = 0496 0404 : :88 4:
The algebraic matrix equation (4) has a unique solution

Abstract In this paper, a counterexample of the above-mentioned paper1 is reported. It is pointed out that one of the conditions for a linear system to be stabilizable via static output feedback is not correct. A modied necessary and sufcient condition for this problem is also presented. Index TermsStabilization, static output feedback.
and then S = [00:2 0:9]. So the output feedback gain is F = 1:6. But the eigenvalues of the closed-loop system A 0 BF C are 02.362 and 0.762. Obviously, it is not stable. So the statements i) and iii) in Theorem 3.1 of the paper are not equivalent. II. A CORRECTED STABILIZABILITY CONDITION In this section, we give a new necessary and sufcient condition for the existence of a stabilizing static output feedback gain matrix. The following lemma is well known [1], [2]. Lemma 1: Let the linear time-invariant system (1) be given. Then, the following statements are equivalent. 1) System (1) is stabilizable via state feedback. 2) There exist matrices Q > 0 and R > 0 of compatible dimensions such that the following algebraic Riccati equation (ARE): PA + AT P 0 PBR01 BT P + Q = 0 has a unique solution
01:7
18
01:7 > 0 1:8
I. INTRODUCTION Stabilization of linear systems by static output feedback is a problem that is practically important and theoretically appealing. Recently, Trono-Neto and Kucera presented two necessary and sufcient conditions for the existence of a stabilizing static output feedback gain matrix,1 but one of them is incorrect.
Manuscript received April 26, 1996; revised September 17, 1996. This work was supported by the National Natural Science Foundation of China under Grant 69604007. The authors are with the Institute of Industrial Process Control and National Laboratory of Industrial Control Technology of Zhejiang University, Hangzhou, 310027, China (e-mail: yycao@iipc.zju.edu.cn). Publisher Item Identier S 0018-9286(98)05800-0. 1 A. Trono-Neto and V. Kucera, IEEE Trans. Automat. Contr., vol. 38, pp. 764765, 1993.
P > 0.
00189286/98$10.00 1998 IEEE
C:\work2011\shamov\yulia7\00704983.pdf
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 8, AUGUST 1998 1111
3) There exist matrices P > 0 and R > 0 of compatible dimensions satisfying the following algebraic Riccati inequality (ARI):
P A + AT P
0 P BR01 BT P < 0:
Theorem 1: Let Ei and Ek be the projection matrices dened in (2). Then, the following statements are equivalent. 1) System (1) is stabilizable via static output feedback. 2) There exist matrices Q > 0; R > 0; and L of compatible dimensions such that the constrained algebraic matrix equation
Now we consider the LQ control problem with the cross term, which we take as the basis for our development. Namely, we consider the cost functions
AT P + P A 0 Ei (P B + LT )R01 (B T P + L)Ei + Q = 0 S = LEi 0 B T P Ek Q 0 S T R01 S > 0;

has a unique solution P > 0. 3) There exist matrices P > 0; R > 0 and L of compatible dimensions satisfying the following modied ARI with suitable cross term
v(x0 ) =
(x
T Qx + 2uT Sx + uT Ru) dt
(5)
where Q > 0; R > 0; and S are constant weighting matrices of compatible dimensions satisfying
Q 0 S T R01 S > 0:
(6)
AT P
T 01 +S R S
+ PA
0 Ei (P B + LT )R01 (BT P + L)Ei

< 0:
(10)
The solution for the LQ control problem associated with (1), (5), and (6) is [1]
P A + AT P
u = 0Kx; 0 (BT P
K = R01 (B T P + S ); T 01 T + S ) R (B P + S ) + Q = 0:
(7)
Then F = R01 (L + B T P )C + is a stabilizing output feedback gain matrix. Proof: From the former part of the proof1 and Lemma 2, the necessity is obvious and there is only a need to prove the sufciency. It is not difcult to nd that the ARE (4) can be written as
In the paper, the authors ignored the inequality constraint (6). Lemma 2: Let (1) be given. Then, the following statements are equivalent. 1) System (1) is stabilizable via state feedback. 2) There exist matrices Q > 0; R > 0; and S of compatible dimensions such that the ARE (7) with the cross term satisfying inequality (6) has a solution P > 0. 3) There exist matrices P > 0; R > 0; and S of compatible dimensions satisfying the following modied ARI with the cross term:
P (A 0 BR01 S ) + (A 0 BR01 S )T P 0 P BR01 BT P + Q 0 S T R01 S = 0:

From [1], the state feedback K = R01 (S + B T P ) = R01 (L + B T P )C + C = R01 (L + B T P )Ei = F C stabilizes (1). In fact, the second condition of statement ii) of Theorem 3.1 in the paper has been modied to inequality constraint (6). Let us continue the above example with a = 01. It is not difcult to nd that we can nd (4) and (6) will hold if
P A + AT P
0 (BT P + S )T R01 (BT P + S ) + S T R01 S < 0:

(8)
L = [0 P
1:5];
Proof: From Lemma 1, we know that if (A; B ) is stabilizable, then there exist matrices R > 0 and Q1 > 0 such that ARE
2 = 0:5
0:5 1
R = 1; > 0; Q
S = [0
3 = 1 1 3
1:5]
> 0:
P A + AT P
0 P BR01 BT P + Q1 = 0
So F = 2 is a stabilizing output feedback gain. In fact, both eigenvalues of the closed-loop system A 0 BF C are 01 and 01. III. CONCLUDING REMARKS The main point of this paper is to correct the result of the above-mentioned paper. A necessary and sufcient condition for stabilizability of a linear time-invariant plants via static output feedback is given using the forms of ARE and ARI. We notice that the ARI (10) is a quadratic matrix inequality. Because of the negative sign in the 0Ei M T R01 MEi ; (M = B T P +L) term, (10) cannot be simplied to a linear matrix inequality (LMI). But we may use an iterative LMI algorithm to solve it since this ARI will be reduced to an LMI problem, which may be solved numerically very efciently [2], if M and R are previously xed. This is also one of our research interests.
has a unique solution P > 0. Because Q1 > 0, we can always nd some S such that
Q1 + P BR01 S + S T R01 B T P > 0:

For example, let S = S0 ; S0 = [Im 0n0m ], where
(9)
jj <
Dene
min (Q1 ) : T max P BR01 S0 + S0 R01 B T P

+S
Q = Q1 + P BR01 S + S T R01 B T P
T R01 S
then Q > 0; Q 0 S T R01 S > 0; and ARE (4) hold. Therefore, 1) implies 2). Note that the ARE (4) can be rewritten as
REFERENCES
[1] B. D. O. Anderson and J. B. Moore, Linear Optimal Control. Englewood Cliffs, NJ: Prentice-Hall, 1971. [2] B. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory. Philadelphia, PA: SIAM, 1994. [3] G. Gu, On the existence of linear optimal control with output feedback, SIAM J. Contr. Optimiz., pp. 711719, 1990.
P (A 0 BR01 S ) + (A 0 BR01 S )T P 0 P BR01 BT P + Q 0 S T R01 S = 0:

From Lemma 1, we know 2) implies that 1) holds. The rest of the proof is omitted because the equivalence of 2) and 3) is obvious.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 6, JUNE 2004 995
Optimization of Static Output Feedback Using Substitutive LMI Formulation

Atsushi Fujimori
AbstractThis note proposes a new design tool for optimizing static output feedback using a linear matrix inequality (LMI) formula called substitutive LMI. A matrix inequality derived from static output feedback is not usually linear. Adding a positive definite term including auxiliary variables, the matrix inequality is transformed into an LMI with respect to the positive definite matrix and the static output feedback gain. An iterative calculation algorithm is given to solve the substitutive LMI. In this note, designs of the static output feedback gain are shown in the frame of and syntheses. A numerical example is shown to demonstrate the effectiveness of the proposed technique.
where x 2 Rn is the state, w 2 Rr is the external signal, u 2 Rm is the input, z 2 Rq is the controlled output and y 2 Rp is the measured output. The following assumption is holds on the generalized plant (1): A1) (A; B2 ) is stabilizable and (C2 ; A) is detectable. The objective in this note is to design a static output feedback law given by
u t
( ) = 0F y(t)
(2)
which stabilizes the closed-loop system and satisfies specified control performances such as H1 , H2 norm constraints, and so on. A. Substitutive LMI Based on Quadratic Stability In this section, static output feedback design using the substitutive LMI is explained in the frame of the quadratic stability, where D21 = 0. Substituting (2) into (1), the closed-loop system is stable only if there 1 exists a Lyapunov function V (x) = xT P x, P 2 Rn2n > 0 such that
V <
and syntheses, iterative calculation, linear matrix Index Terms inequality (LMI), static output feedback.
I. INTRODUCTION Static output feedback design is one of open problems in control community. A number of papers treating static output feedback design have been published. Unlike state feedback design, a static output feedback gain which stabilizes the closed-loop system is not always found even if the plant is controllable and/or observable [1][4]. Recently, design methods of static output feedback using linear matrix inequality (LMI) have been proposed by many researchers [5][10]. Crusius and Trofino [5] presented LMI conditions for static output feedback and extended them to the cases of discrete-time systems, decentralized output feedback, H1 control, and robust control systems with polytopic uncertainties. Benton and Smith [7] presented another LMI formula, in which any iterative calculation is not required. Cao et al. [8] discussed a simultaneous stabilization by static output feedback. Ghaoui et al. [9] and Iwasaki [10] developed iterative algorithms for designing static output feedback gains and fixed-order controllers. This note proposes a new design tool for optimizing static output feedback using an LMI formula called substitutive LMI. A matrix inequality derived from static output feedback is not usually linear. Adding a positivedefinite term including auxiliary variables, the matrix inequality is transformed into an LMI with respect to the positive definite matrix and the feedback gain. An iterative calculation algorithm is given to solve the substitutive LMI. The iterative calculation presented in this note is related to LQ control problems by Cao et al. [8] and Shimomura and Fujii [11]. This note refines the LMIs derived in [8] and [11] and extends them to the designs of static output feedback in the frame of H1 and H2 syntheses. A numerical example is shown to demonstrate the effectiveness of the proposed technique. In this note, the n 2 n unit matrix is written as In , and an n 2 n positive definite matrix is denoted as P 2 Rn2n > 0. II. STATIC OUTPUT FEEDBACK BY SUBSTITUTIVE LMI This section proposes the substitutive LMI for static output feedback. Consider a generalized plant (1)
(3)
holds. Expressing (3) in terms of a matrix inequality, we have
1 8(P; F C2 ) = P (A 0 B2 F C2 ) + (A 0 B2 F C2 )T P
Instead of (4), consider a sufficient matrix inequality
<
0:
(4)
8(P; F C2 ) + 0T 50 < 0 T 1 R 0 1 L 0 R01 B2 P 0= 5= 0 R M 0 F C2
(5)
where L 2 Rm2n and M 2 Rm2n are introduced to solve (5) as an LMI and are called substitutive variables in this note. R 2 Rm2m > 0 is a positivedefinite weighting matrix and is given according to each control specification. Expanding (5), we have
T T T 8(P; L)+3+ F C2 0 R01 B2 P R F C2 0 R01 B2 P 1 8(P; L) = P (A 0 B2 L)+(A 0 B2 L)T P < 0 1 3 = LT RL + M T RM 0 M T RF C2 0 (F C2 )T RM:
<
0
(6)
Using the Schur complement [12], (6) is transformed into
8(P; L) + 3 01 B2 P T F C2 0 R
FC
T 2 0 R01 B2 P 01 0R
T
<
0:
(7)
_ ( ) = Ax(t) + B1 w(t) + B2 u(t) ( ) = C1 x(t) + D11 w(t) + D12 u(t) y (t) = C2 x(t) + D21 w (t)
x t z t
If L and M are fixed, (7) is an LMI with respect to P and F and can be solved by an LMI solver. In this note, (7) is referred as substitutive LMI. Similar techniques were used in [8] and [11] which were based on LQ control problems. The substitutive LMI proposed in this note is an extended formula to a general control problem. If the following relations hold:
L
(1)
Manuscript received June 17, 2003; revised February 21, 2003 and October 21, 2003. Recommended by Associate Editor A. Bemporad. The author is with the Department of Mechanical Engineering, Shizuoka University, Hamamatsu 432-8561, Japan (e-mail: tmafuji@ipc.shizuoka.ac.jp). Digital Object Identifier 10.1109/TAC.2004.829633
T 0 01 2 = 0 0 2 =0
R B P M FC
(8) (9)
then (5) coincides with (4). Taking into consideration this in the calculation algorithm shown in the next section, L and M will be given so as to satisfy (8) and (9) as much as possible.
0018-9286/04$20.00 2004 IEEE
996 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 6, JUNE 2004
The dual form of (7) can be also derived. Defining Q = P 01 , (4) is equivalent to the following matrix inequality:
At Step 4), the following norms of the difference between the (k)th and (k 0 1)th with respect to P (k) (Q(k) ) and F (k) :
1 9(Q; B2 F ) = (A 0 B2 F C2 )Q + Q(A 0 B2 F C2 )T
9(Q; B2 F ) + 262T < 0 1 T 2 = K 0 QC2 W 01 N 0 B2 F
< 0:
(10)
Similar to (5), a sufficient matrix inequality for (10) is given by
1 = 1 (k) = ( Pk)
F
P (k) 0 P (k01) F (k) 0 F (k01)
(k Q )
1 =
Q(k) 0 Q(k01)
(16)
1 6=
(11)
where K 2 Rn2p and N 2 Rn2p are the substitutive variables and are introduced to solve (11) as an LMI. W 2 Rp2p > 0 is a positivedefinite weighting matrix. Expanding (11) and using the Schur complement, (11) is transformed into
may be indexes for the judging the convergence of the iterative calculation. In H1 and H2 syntheses shown in the following section, H1 or H2 norm of the closed-loop system can be used for this. Strictly speaking, the convergence of the proposed calculation is not guaranteed. However, once a feasible solution is found, P (k) , F (k) , L(k) , and M (k) are adjusted to reduce the sufficiency of the substitutive LMI (7); that is, to satisfy (8) and (9). This implicitly indicates the convergence of the calculation. III. SUBSTITUTIVE LMI FORMULAS FOR H1
AND
T 9(Q; K )+1 B2 F 0 QC2 W 01 <0 T W 01 T 01 B2 F 0 QC2 0W 1 9(Q; K ) =(A 0 KC2 )Q + Q(A 0 KC2 )T < 0 1 1 = KW K T + N W N T 0 N W (B2 F )T 0 B2 F W N T :
H2 SYNTHESES
This section presents the substitutive LMIs for designing static output feedback gains in H1 and H2 control syntheses. (12) The generalized plant for H1 synthesis is given by (1). Using Doyles notation [14], the transfer function of the closed-loop system Tzw (s) is written as A. H1 Synthesis
In this note, (7) is called standard form, while (12) is called dual form. B. Iterative Calculation Algorithm of Substitutive LMI This section shows an iterative calculation algorithm for solving the substitutive LMIs (7) and (12). A basic idea of the algorithm is as follows. If the substitutive variables L, M , K and N are fixed, (7) is then an LMI with respect to P and F , while (12) is an LMI with respect to Q and F . Therefore, after L, M , K and N are given appropriately, (7) or (12) can be solved by a conventional LMI solver [13]. These manipulations are iterated until the solutions are converged. The iterative calculation algorithm is given as follows, where the superscript k means the number of iteration and the equations in the round brackets indicate the calculation of the dual form. Step 1) Give a positivedefinite weighting matrix R > 0(W > 0). Set k = 1. Select an initial value L(1) (K (1) ) such that A 0 B2 L(1) (A 0 K (1) C2 ) is stable. M (1) (N (1) ) is given as
Tzw (s) =
1 =
A 0 B2 F C 2 C1 0 D12 F C2 Acl Bcl : Ccl Dcl
B1 0 B2 F D21 D11 0 D12 F D21

(17)
The objective of H1 synthesis is to stabilize the closed-loop system and satisfy H1 norm constraint
kTzw (s)k1 < for > 0
(18)
where is an upper bound of H1 norm and is given in advance. 1) Standard Form: (18) can be represented as a matrix inequality [14] T P Acl + AT P P Bcl Ccl cl 1 T T 81 = (19) Bcl P 0 Ir Dcl < 0:
M (1)
Step 2) Step 3)
= L(1)
N (1)
= K (1)
(13)
Ccl
Dcl
0 Iq
Solve LMI (7) [LMI (12)] with respect to P (k) > 0(Q(k) > 0) and F (k) . L(k+1) and M (k+1) (K (k+1) and N (k+1) ) are given by
Consider the following matrix inequality which is given by adding a positivesemidefinite term to (19):
M (k+1) =F (k) C2
T L(k+1) =R01 B2 P (k)
N (k+1) = B2 F (k) :
T K (k+1) = Q(k) C2 W 01
(14) (15)
T 81 +01 501 < 0 T 1 L 0 R01 B2 P U 0 F D21 0 5 = R 0 1 (20) 01 = M 0 F C2 0 R 0 0 L 2 Rm2n , M 2 Rm2n , and U 2 Rm2r are the substitutive variables. R 2 Rm2m > 0 is a positivedefinite weighting matrix. Ex-
Step 4) If it is judged that the solution is converged, stop the calculation. Then, the static output feedback gain is obtained as F (k) . Otherwise set k ! k + 1 and go to Step 2). Some comments should be given on the aforementioned algorithm. From assumption A1), an initial value L(1) (K (1) ) which stabilizes A 0 B2 L(1) (A 0 K (1) C2 ) is always found because L(1) (K (1) ) corresponds to the state feedback (full-order observer) gain. However, it is not guaranteed whether P (k) > 0 (Q(k) > 0) and F (k) satisfying LMI (7) [LMI (12)] are always found at Step 2) or not. It depends on L(1) (K (1) ) and R(W ). If LMI (7) [LMI (12)] could not be solved during some iteration, it is needed to change L(1) (K (1) ) and/or R(W ).
panding (20) and using the Schur complement, (20) is transformed into (21)
L U0F D21 1 31 = M T RM 0 M T RF C2 0 (F C2 )T RM
8(P; L)+31 (B 1 0 B 2 U )T P C1 0 D12 F C2 T F C2 0 R01 B2 P
3 3 0 Ir 3 D11 0 D12 F D21 0 Iq

0 0 0
0R01
0
3 3 3
0R01
(21)
3 3 3 3
<0
where 3 indicates the elements that are readily inferred by symmetry.
2) Dual Form: (18) can be represented as another matrix inequality [14]
Equation (26) can be then represented as the following matrix inequalities [14]
91
1 =
T Acl Q QAcl Ccl Q T Bcl
T QCcl 0 Iq T Dcl
Bcl Dcl 0 Ir
< :
(22)
Consider the following matrix inequality which is given by adding a positivesemidefinite term to (22)
1 T 82 = P Acl + AT P + Ccl Ccl < 0 cl P P B1 >0 T B1 P X trX < 2 :
(28a) (28b) (28c)
91 +2162
1 21 =
K0 2 V 0 D12 F
T 1< QC T W 01
N 0 B2 F
0 0
0 1 6= W W 0
Equations (28b) and (28c) are LMIs with respect to P and X 2 r2r > , while (28a) is a BMI with respect to P and F . Then, the substitutive LMI is applied to only (28a) as follows:
(23)
82 +
where
T L 0 R01 B2 P
T R L 0 R01 B2 P <
(29)
K 2 n2p , N 2 n2p , and V 2 q2p are the substitutive variables. W 2 p2p > is a positivedefinite weighting matrix. Expanding (23) and using the Schur complement, (23) is transformed into (24), as shown at the bottom of the page. In 1 synthesis, another substitutive variable U or V is introduced. At Step 3) of the iterative calculation, U (k+1) V (k+1) is given by
R R
0 )
1 T = D12 D12 > 0:
(30)
Expanding (29) and using the Schur complement, (29) is transformed into
T F C2 0 R01 B2 P
U (k+1)
= F (k) D21 H
V (k+1)
= D12 F (k)
8(P; L)+32
(25)
1 T 32 = C1T C1 + LT RL 0 C1T D12F C2 0 (F C2 )T D12 C1 :
T F C2 0 R01 B2 P 01 0R
<
0
(31)
The generalized plant for 2 synthesis is given by (1), where the . Thereclosed-loop system Tzw s must be strictly proper Dcl fore, the following assumptions have to hold. A2) D11 . A3) Either of the following is imposed on D12 or D21 . D12 m. a) D21 and and D21 p. b) D12 The objective of 2 synthesis is to stabilize the closed-loop system and satisfy 2 norm constraint:
B. H2 Synthesis
()
= 0)
=0 H
LMIs (28b), (28c), and (31) are simultaneously solved by the iterative calculation. 2) Dual Form: Suppose that assumptions A1), A2), and A3b) are held. The closed-loop system Tzw s is given by
()
=0 =0
rank rank
= =
Tzw s
( )=
Acl C1
Bcl
(32)
Equation (26) can be then represented as the following matrix inequalities [14]:
kTzw (s)k2 < ;
for > 0
(26)
assumption A3). 1) Standard Form: Suppose that assumptions A1), A2), and A3a) are held. The closed-loop system Tzw s is given by
where is an upper bound of H2 norm and is given in advance. In H2 synthesis, the standard or the dual form is derived according to
1 T T 92 = Acl Q + QAcl + Bcl Bcl < 0 T Q QC1 >0 C1 Q Y trY < 2 :
(33a) (33b) (33c)
()
Equations (33b) and (33c) are LMIs with respect to Q and Y 2 q2q > , while (33a) is a BMI with respect to Q and F . Then, The substitutive LMI is applied to only (33a) as follows.
Tzw s
( )=
Acl Ccl
B1
(27)
92 +
T T K 0 QC2 W 01 W K 0 QC2 W 01
<
(34)
9(Q; K ) + 11 3 3 3 3 0 Iq 3 3 3 (C 1 0 V C 2 )Q T T 0 I 3 3 (B1 0 B2 F D21 ) (D11 0 D12 F D21 ) r 0 0 0W 01 3 (B2 F 0 QC2 W 01 )T (V 0 D12 F )T 0 0 0W 01 KT 1 N W N T 0 N W (B F ) T 0 B F W N T : 11 = 2 2
<
0
(24)
998 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 6, JUNE 2004
where
W
1 T = D21 D21 > 0:
(35)
Expanding (34) and using the Schur complement, (34) is transformed into
T 0 QC2 W 01 < 0 T T B2 K 0 QC2 W 01 0W 01 1 B B T + KW K T 0 B (B F D )T 0 B F D B T : 12 = 1 1 1 2 21 2 21 1
9(Q; K )+12
B2 F
(36)
That is, (33b), (33c), and (36) are simultaneously solved by the iterative calculation. R and W given by (30) and (35) are positive definite matrices to save the substitutive variables. Even if R and W were not given by (30) and (35), the substitutive LMIs can be derived with another substitutive variables M and N such as the quadratic stability and 1 synthesis.
IV. NUMERICAL EXAMPLE To demonstrate the proposed design method of static output feedback by the substitutive LMI, a numerical example is shown in this section. The state-space data of (1) were given by the longitudinal motion of a helicopter [10], [15]
Fig. 1. Convergence evaluation by norms of difference synthesis, standard form.
and
00:0366 0:0271 0:0188 00:4555 0:0482 01:01 0:0024 04:0208 A= 0:1002 0:3681 00:707 1:42
B1
= = = =
C1 D11 D21
0 0 1 0 1 0 0 0:4422 0:1761 0 0 0 B2 = 3:5446 07:5922 0 0 0 05:52 4:49 0 0 0 0 0 0 1 0 0 C2 = 1 0 0 0 0 0 0 1 0 1 0 0 0:5 0 0 D12 = 1 0 0 1 0 0 1 0 0 :1 0 : 0 0 0 :1 =4 =3 H
(37)
The size of the variable vectors in (1) was that n ,r ,m , and p . The eigenvalues of A was 00.2325, 02.0727, and : 6 j : ; that is, the controlled plant was unstable. Using this data, the static output feedback gain F was designed by 1 and 2 syntheses.
=2 0 27579 H
q
=2 2 5758
=2
Fig. 2.
norm of closed-loop system.
synthesis, standard form.
system kTzw k1 with respect to the iteration. kTzw k1 finally converged to 1.183. The static output feedback gain was obtained as
F
A. H1 Synthesis
The static output feedback gain F with 1 norm constraint was designed by the standard and the dual forms. R in (21) was given by I2 . The initial value L(1) was obtained by an optimal regulator whose weights for the state and the input were, respectively, given by I4 and
I2
0 02970 4: = 00::3892 05004 :244
(1) =
0:9266 00:01474 00:9622 01:387 00:02248 00:8448 0:1886 0:7135
The eigenvalues of the closed-loop system was 055.03, 00.3138 and 0 : 6j : . Fig. 3 shows a plot of kTzw k1 , where the dual form of 1 synthesis was used. W in (24) was given by I2 . Similar to Fig. 2, the iterative calculation successfully converged and kTzw k1 at the twentieth was 1.207. The static output feedback gain was obtained as
0 2345
1 340
(k) (k) Fig. 1 shows plots of the norms of the difference kP k and kF k with respect to the number of iteration, where the maximum singular value was used as the norm. The both norms almost converged until the tenth iteration. Fig. 2 shows a plot of 1 norm of the closed-loop
0 05341 1: = 0::01364 01159 :551
The eigenvalues of the closed-loop system was 017.10, 00.1307 and 0 : 6j : .
0 2191
1 322
was given to solve the substitutive LMI. For the sake of the page limitation, this note presented formulas of H and H2 syntheses by the substitutive LMI. The substitutive LMIs for the pole placement constraints and the impulse response can be also derived. In addition to the numerical example shown in this note, I applied the substitutive LMI to several numerical examples in [5], [7], [8], and [16]. The static output feedback gains for the fourth-order plant [5] and the third-order plants [7] and [8] were obtained by the substitutive LMI, but the gain for the eighth-order plant [16] could not be found. It was depended on the initial and updated values of L and K in the iterative calculation. One of future subjects to research is to improve the solvability of the substitutive LMI with respect to the initial values.
REFERENCES
[1] A. Trofino and V. Kucera, Stabilization via static output feedback, IEEE Trans. Automat. Contr., vol. 38, pp. 764765, May 1993. [2] V. Kucera and C. E. Souza, A necessary and sufficient condition for output feedback stabilizability, Automatica, vol. 31, pp. 13571359, Sept. 1995. [3] V. L. Syrmos, C. T. Abdallah, P. Dorato, and K. Grigoriadis, Static output feedbackA survey, Automatica, vol. 33, pp. 125137, Feb. 1997. [4] Y.-Y. Cao, Y.-X. Sun, and W.-J. Mao, A new necessary and sufficient condition for output feedback stabilizability and comments on stabilization via static output feedback, IEEE Trans. Automat. Contr., vol. 43, pp. 11101111, Aug. 1998. [5] C. A. R. Crusius and A. Trofino, Sufficient LMI conditions for output feedback control problems, IEEE Trans. Automat. Contr., vol. 44, pp. 10531057, May 1999. [6] J. C. Geromel, C. C. de Souza, and R. E. Skelton, LMI numerical solution for output feedback stabilization, in Proc. Amer. Control Conf., 1994, pp. 4044. [7] R. E. Benton and D. Smith, Static output feedback stabilization with prescribed degree of stability, IEEE Trans. Automat. Contr., vol. 43, pp. 14931496, Oct. 1998. [8] Y.-Y. Cao, Y.-X. Sun, and J. Lam, Simultaneous stabilization via static output feedback and state feedback, IEEE Trans. Automat. Contr., vol. 44, pp. 12771282, June 1999. [9] L. E. Ghaoui, F. Oustry, and M. AitRami, A cone complementarity linearization algorithm for static output-feedback and related problems, IEEE Trans. Automat. Contr., vol. 42, pp. 11711176, Aug. 1997. [10] T. Iwasaki, The dual iteration for fixed-order control, IEEE Trans. Automat. Contr., vol. 44, pp. 783788, Apr. 1999. [11] T. Shimomura and T. Fujii, An iteration method for mixed and control design with uncommon LMI solutions, in Proc. Amer. Control Conf., 1999, pp. 32923296. [12] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory. Philadelphia, PA: SIAM, 1994, vol. 15, SIAM Studies in Applied Mathematics. [13] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control Toolbox. Natick, MA: The Math Works Inc., 1995. [14] C. Scherer, P. Gahinet, and M. Chilali, Multiobjective output-feedback control via LMI optimization, IEEE Trans. Automat. Contr., vol. 42, pp. 896911, July 1997. [15] W. E. Schmitendorf, A design methodology for robust stabilizing controllers, AIAA J. Guid., Control, Dyna., vol. 10, no. 2, 1987. [16] X. A. Wang, Grassmannian central projection and output feedback pole assignment of linear systems, IEEE Trans. Automat. Contr., vol. 41, pp. 786794, June 1996.
Fig. 3.
H
1
norm of closed-loop system.
synthesis, dual form.
The H norm of the closed-loop system achieved by the standard form was almost the same as that by the dual form, but the static output feedback gains were different in this numerical example. should be given as a larger value than that of kTzw k finally achieved. Otherwise, P (k) > 0 and F (k) which satisfied (21) were not found by an LMI solver [13]. In this numerical example, in (21) and (24) was given by 2 and 1.5, respectively.
B. H2 Synthesis The static output feedback gain F with H2 norm constraint was next designed for the same helicopter model (37), where we reset D11 = 0 and D12 = 0. Then, assumptions A2) and A3a) held. Applying the standard form of H2 synthesis, H2 norm of the closed-loop system kTzw k2 converged to 0.5747. The feedback gain was obtained as
F =
2:7599 04:4460
11:846 019:370 :
The eigenvalues of the closed-loop system was 0190.49, 00.1404, and 00:3029 6 j1:322. Moreover, the dual form of H2 synthesis was examined, where we reset D11 = 0 and D21 = 0 in (37). kTzw k2 converged to 2.879. The feedback gain was obtained as
F =
68:800 99:475 27:864 19:141
The eigenvalues of the closed-loop system was 0216.77, 026.92, and 00:3366 6 j2:104. V. CONCLUDING REMARKS This note has proposed a new design tool for optimizing static output feedback using an LMI formula called substitutive LMI. A matrix inequality for static output feedback is not usually linear. Adding a positive definite term including auxiliary variables, the matrix inequality was transformed into an LMI with respect to the positive definite matrix and the static output feedback gain. An iterative calculation algorithm
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 12, DECEMBER 2004 2271
[17] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequality in System and Control Theory. Philadelphia, PA: SIAM, 1994. [18] J. K. Hale and S. M. Verduyn Lunel, Introduction to Functional Differential Equations (Applied Mathematical Sciences, Vol. 99). New York: Springer-Verlag, 1993. [19] L. E. Elsgolts and S. B. Norkin, Introduction to the Theory and Application of Differential Equations With Deviating Arguments (Mathematics in Science and Engineering: Vol. 105). New York: Academic, 1973. control of systems with parameter uncer[20] L. Xie, Output feedback tainty, Int. J. Control, vol. 63, no. 4, pp. 741750, 1996.
is guaranteed to converge. We show in Section IV that, like the existing approach, the new approach is also a dual optimal output feedback LQR theory. Numerical examples are given in Section V to illustrate the effectiveness of the algorithm and the new method. We conclude the note in Section VI. II. FIXED-STRUCTURE CONTROL LAW A. Review of Existing Results Equations (4)(7) are the design equations from [2]; see also [3] and [4]. Note that K = F C .
min
0
C
(x
T Qx + 2uT N x + uT Ru) dt
A Convergent Algorithm for Computing Stabilizing Static Output Feedback Gains

Jen-te Yu
AbstractWe revisit the approach by Cao et al. that uses a fixed-structure control law to find stabilizing static output feedback gains for linear time-invariant systems. By performing singular value decomposition on the output matrix, together with similarity transformations, we present a new stabilization method. Unlike their results that involve a difficult modified Riccati equation whose solution is coupled with other two intermediate matrices that are difficult to find, we obtain Lyapunov equations. We present a convergent algorithm to solve the new design equations for the gains. We will show that our new approach, like theirs, is a dual optimal output feedback linear quadratic regulator theory. Numerical examples are given to illustrate the effectiveness of the algorithm and validate the new method. Index TermsDual linear quadratic regulator (LQR), Lyapunov equation, modified Riccati equation, static output feedback.
T A P + PA
N
i=
=
01 (B T P + L)C + F = R 01 (B T P + N ): K = R
LE
i0
(3) i T P + L)T R01 (B T P + L)E + Q = 0 (4) Ei (B i

k=
I E
T R01 N
0
R >
(2)
BP E
(5) (6) (7)
It is not clear how L and N can be generated by a systematic method, as they are all coupled with P the solution of the modified Riccati equation that is difficult to solve. In addition, N has to satisfy a matrix inequality constraint that involves the other two weighting matrices Q and R, as given in (2). B. New Perspective on the Same Control Law Consider again F = K C + , where K is stabilizing, and perform singular value decomposition on the output matrix C . In the following, U and V are unitary matrices, and S contains singular values of C . We partition V into two parts.
I. INTRODUCTION Consider the following problem. Given a continuous linear time-invariant system
x _ x
C BF C
= = =
U SV A V
TU = I
C V
BK C
2 <n
Ax ;
+ Bu
u
2 <m
Cx y
2 <r
E [x(0)x(0) ] =
X >
(V
T AV
T BK V S + S ) V T :
TV =
BK V S
T U SV T
(8) (9)
(1)
where x(t) is the state, u(t) is the input, y (t) is the output, and E[ 1 ] is the expected value. Find a static output feedback gain F such that u = 0F y stabilizes the given system; see [1], and [5][11]. In this note, we revisit the approach by Cao et al. [2] that uses a fixed-structure control law to find the stabilizing output feedback gains; see also [3] and [4]. The main drawbacks of the aforementioned approach are that its solution involves a difficult modified Riccati equation (or inequality [2]), together with two other matrices (coupled with the solution of the modified Riccati equation) that are difficult to find. To overcome these difficulties we develop a new stabilization method. In this note K denotes the full state feedback gain, and F stands for the output feedback gain. The superscript + denotes the MoorePenrose inverse. A norm with subscript Fro denotes the Frobenius norm. The note is outlined as follows. In Section II, we briefly review the results of [2][4], and present a new stabilization method. In Section III, we present a simple iterative algorithm to solve the new design equations for the gains that
Manuscript received December 16, 2003; revised March 14, 2004. Recommended by Associate Editor Hong Wang. The author is at 895 St. Charles Dr. #5, Thousand Oaks, CA 91360 USA (e-mail: jyuinfinity@hotmail.com). Digital Object Identifier 10.1109/TAC.2004.838482
Let us define the terms in the transformed domain and investigate the closed-loop matrix A 0 BF C
^ A =
V
T AV
^ A
^ B
TB
^ K2 ]
^ K
V1
^ K1 =
A
K V1 V
^ K2 =
K V2
BF C
0 ^[ ^
2<
r
n2r
KV
^ = [ K1
V2
B K1
T: V
2<
^ K2 ] (10) n2(n0r)
(11)
^ One shall see that K2 would be lost. To avoid this information loss, ^ we may impose a constraint: K2 = 0, which is K V2 = 0. As V is T + V V T = I and the following: unitary, we have V1 V1 2 2
KC
K V1 V 1
T =K
V2 V 2
K V2 V 2 :
(12)
C. New Set of Design Equations Theorem 1: Suppose the given system (1) is static output feedback stabilizable and detectable. There exists a static stabilizing output feedback gain F = K C + , if there exists a static stabilizing state feedback gain K that satisfies
(A (A
0 0
BK )Y BK )
TP + P
+ Y (A
0 ( 0
A
BK )
T +X =0 Y =YT >0 (13) T RK + Q = 0; P = P T > 0 BK ) + K
(14)
0018-9286/04$20.00 2004 IEEE
2272 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 12, DECEMBER 2004
10
K 0 R01 B T P I 0 V2 V2T Y 01 V2
01
V2T Y 01
= 0:
(15)
III. A CONVERGENT ALGORITHM We now present an iterative algorithm to solve the design equations for the gain. In the sequel, subscript i denotes the index of iteration, and " stands for the preset tolerance. 1) Initialization: Solve a standard Riccati equation and obtain the Kalman gain as the initial gain. 2) ith iteration: Solve the Lyapunov equations for Yi and Pi , respectively, in
Proof: We may formulate the static output feedback stabilization problem as follows:
(xT Qx + uT Ru) dt subject to KV2 = 0:

0
J
x = Ax + Bu y = Cx u = 0Kx _
minE
E[x(0)x(0)T ] = X Q0 R>0
Define J as the cost to be minimized, and we get
=E
Suppose there exists a matrix P
= P T > 0 so that d(xT P x)=dt = 0xT (K T RK + Q)x

1 =E 0[d(xT P x)=dt] dt 0 = E[xT (0)P x(0) 0 xT (1)P x(1)]:
xT (K T RK + Q)x dt :
(A 0 BKi )Yi + Yi (A 0 BKi )T + X = 0 (A 0 BKi )T Pi + Pi (A 0 BKi ) + KiT RKi + Q = 0:

Evaluate the gain increment
1Ki = R01 BT Pi I 0 V2
(16)
V2T Yi01 V2
01
V2T Yi01
0 Ki :
then
Update the gain by Ki+1 Ki i Ki where < i < , and i is chosen so that A 0 BKi+1 is asymptotically stable. . i Set i Ki C + . Other3) If kKi V2 k < ", stop the procedure, and let F wise, go to 2). Since Ki+1 Ki i Ki , we get the following:
+ 1
= +1 =
+ 1 Ki+1 V2 = Ki V2 + i 1Ki V2 = (1 0 i )Ki V2 :

kKi+1 V2 k = j1 0 i j kKi V2 k 0 < i < 2 ) ilim kKi V2 k = 0: !1
For an asymptotically stable system
J = tracefE[xT (0)P x(0)]g = trace(P X ) xT (A 0 BK )T P x + xT P (A 0 BK )x + xT (Q + K T RK )x = 0:

(17) The previous expression must hold for any x. As a result
As a result (24) (25)
(A 0 BK )T P + P (A 0 BK ) + K T RK + Q = 0:
Define the Hamiltonian H as
(18)
In order to get a successful implementation, it is important that we maintain the closed-loop asymptotic stability for all steps during the solution procedure. We can achieve this by the following theorem. Theorem 2: The following conditions guarantee the asymptotic stability of A 0 BKi+1
= trace fP X + [(A 0 BK )T P + P (A 0 BK ) +K T RK + Q]Y g + 2trace V2T K T M

@H @P
: (19)
Matrices Y and M are Lagrange multipliers. The stationary conditions are as follows:
Proof: Suppose M1 is asymptotically stable and satisfies the following Lyapunov equation:
0 < i < 2kX 01=2 B11 i Yi X 01=2 k2 K 1 0 < i < 2kX 01=2B1Ki Yi X 01=2 kFro + W M1T + Z = 0
p p
Z
(26) (27)
=0
@H @Y
=0
@H @K
=0
@H @M
= 0:
(20)
M1 W
= Z T > 0:
T M2 :
(28)
From which, we get
(A 0 BK )Y + Y (A 0 BK )T + X = 0 (A 0 BK )T P + P (A 0 BK ) + K T RK + Q = 0 RKY 0 B T P Y + MV2T = 0 KV2 = 0:

K
0. Take
(21) (22) We get
For a vector v that is compatible to a matrix G, we have v T GT Gv
G = (1=
2)Z 1=2 0 2Z 01=2 W

T M2 v
(29)
(1=2)vT Zv + 2vT (M2)W Z 01 W

vT
Suppose
The first two stationary equations are exactly (13) and (14). From (21), we obtain
T W M2
+ (M2 )W
v: (30)
= R01
BT P
0 MV2T Y 01
(23)
T Z > 4(M2 )W Z 01 W M2 :
(31)
Combining (22) and (23), we obtain
R01 B T P
From which, we get
0 MV2T Y 01
T
V2 = 0 :
We then get
(1=2)vT Zv + (1=2)vT Zv > vT

T T T
01 M = B P V2 V2 Y 01 V2
T
Substitution of M into (23) yields (15).
) v Zv > v W M2 + (M2 )W v ) 0vT M1 W + W M1T v > vT W M2T + (M2 )W ) 0 > vT [W (M1 + M2 )T + (M1 + M2 )W ]v:
T W M2
+ (M2)W
v v
11
This implies that M1 M2 is asymptotically stable. We used the sufficient condition
Matrix S consists of two parts: S and nonsingular. As a result
= [S1 0], where S1 is diagonal

0 C + = V1 S 1 1 U T : 01
4 2
Equivalently
T W M2
T Z 01 W M2 < Z:
= U S1 V1T
CT
= V1 S1 U T
The right-hand side of (35) now becomes

T T Z 01=2 W M2 Z 01=2 < I
4 2 4 2
<
2 T Z 01=2 W M2 Z 01=2 < 1 2
T Z 01=2 W M2 Z 01=2
Y C T (CY C T )01 = Y V1 S1 U T U S1 V1T Y V1 S1 U T
= Y V1
(32)
V1T Y V1
01 01 T S1 U :
follows directly from (32). As a result of k 1 k2 k 1 kFro , we obtain (27). From (24). we can see that faster convergence can be achieved, if we choose i close to 1, while taking into account the asymptotic stability condition (26) or (27). Typically, it is assumed that X I, which greatly simplifies (26) and (27).
2kZ 01=2M2 W Z 01=2 k2 Recall A 0 BKi+1 = A 0 BKi 0 i B 1Ki . Compare (13) with (28), and let M1 = A 0 BKi ; W = Yi ; Z = X; M2 = 0B 1Ki , then (26) =
The left-hand side of (35) becomes
I 0 V2 V2T Y 01 V2
01
V2T Y 01 C + V2T Y 01 V2
= I 0 V2
Now, we need to show
01
0 V2T Y 01 V1 S1 1 U T : 01
V2T Y 01 V1 :
Y V1 V1T Y V1
01
= I 0 V2
V2T Y 01 V2
The conclusion follows directly from (34). V. ILLUSTRATIVE NUMERICAL EXAMPLES We give three examples (see Figs. 16) to illustrate the effectiveness of the new method. Example 2, taken from [7], is not static output feedback stabilizable. Example 3 is not static output feedback stabi05 , except for lizable for any h 2 < (defined below). We set " 04 . The gain K0 stands for the initial the third example, where " gain (the Kalman gain), K is the measure of error from the optimal gain. Eigenvalues are denoted by s. It is interesting to note that the closed-loop eigenvalues of Examples 2 and 3 fall on the imaginary axis. The best the algorithm can do for these two examples is not allowing the closed-loop eigenvalues to go to the right complex plane. Example 1:
IV. DUAL OPTIMAL OUTPUT FEEDBACK LQR THEORY We now show that our approach, like the previous one, is a dual optimal output feedback LQR theory. The optimal output feedback LQR gain F can be obtained by solving three coupled matrix equations
identical. We will be using the following equality:
(A 0 BF C )Y + Y (A 0 BF C )T + X = 0 (A 0 BF C )T P + P (A 0 BF C ) + (F C )T R(F C ) + Q = 0 F = R01 B T P Y C T (CY C T )01 where Y = Y T > 0; P = P T > 0. As K = KC + C = F C , we only need to show that the gains are
I 0 V2 V2T Y 01 V2
Equivalently
= 10
= 10
01
V2T Y 01 V1 V1T Y V1 = Y V1 :
(33)
I 0 V2 V2T Y 01 V2
01
V2T Y 01 V1
= Y V1
V1T Y V1
01
04 09 07 A = 06
: (34)
Before we proceed, let us derive the aforementioned equality. Consider the left-hand side of (33)
I 0 V2 V2T Y 01 V2
01
V2T Y 01
V1 V1T Y V1 I 0 V2 V2T Y V1
= I 0 V2 = I 0 V2 + V2 = I 0 V2 = Y V1 :
V2T Y 01 V2 V2T Y 01 V2
01 01
V2T Y 01
V2T Y 01 0 V2 V2T
V2T Y 01 V2
01
V2T Y 01 V2 V2T Y V1
V2T Y 01 V2
01
V2T Y 01 Y V1
We used V2T V1 in the previous derivation, which results from I the fact that V is unitary, and the off-diagonal elements of V T V are zero, i.e., V2T V1 . By comparison, it is obvious that we need to show that the following equality holds:
=0 =0
I 0 V2 V2T Y 01 V2
01
V2T Y 01 C + = Y C T (CY C T )01 : (35)
2 5 0 3 :9 0 :5 2 :0 0 :5 0 :1 0 1 :0 B = 0 2 :5 0 0 :5 0 1 :0 1 :0 2 :5 2 :0 01:0 00:05 3 6 0 5 02 01 0 7 5 C= 01 04 0 7 01 06 0 5 03 Q = I 7 = X R = I2 (A) = f018:4797; 11:0100 6 j 5:4496; 04:6867 6 j 9:7361 4:8710; 05:0381g 06:5953 2:1799 18:0050 7:7135 K0 = 01:7234 5:0666 12:3486 0:6713 2:7930 14:7524 0:2250 10:2665 6:0697 13:1018 (A 0 BK0 ) = f018:4033; 011:6814 6 j 6:2995 04:7938 6 j 9:5050; 06:5629; 05:2828g 4:4999 18:5063 33:6949 4:9221 K= 1:9301 5:9048 5:6425 0:4170
02 07 03 03 0 05
08 5 01 08 4 06 03 02 2 6 7 5 2 10 01 8 1 2 3 07 6 03 0 4 6 1 8 04 6 0 9 0 2 04 8 3 1 9 06 3
2274 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 12, DECEMBER 2004
12
Fig. 1. Trajectory of Frobenius norm of KV Example 1.
Fig. 2.
Trajectory of scaling factor Example 1.
28:6039 24:3573 13:8375 5:8308 2:9928 4:5799 0: 0: :051 :015 1K = 1001 2 00014 00020 00116 00067 :129 :280 0: 0: 00:025 00:056 0:018 00:029 0:217 00:230 00:0844 04:7532 F = KC + = 0:3026 01:0222 (A 0 BF C ) = f016:8986 6 j 33:7676; 01:7898 6 j 12:1583 01:9086 6 j 2:9663; 03:4386g:
Example 2:
0 0 1 0 0 0 0 0 1 B= 0 A= 01 1 0 0 1 1 01 0 0 0 0 1 CT = 0 0 (A) = f0; 0; 0 6 j 1:4142g Q = I4 = X R = 1 K0 = [1:7212 00:3070 2:1077 1:1365] (A 0 BK0 ) = f00:6801 6 j 0:4961; 00:3738 6 j 1:3624g K = [0 0:2651 0 0] F = KC + = 0:2651 (A 0 BF C ) = f0 6 j 0:3778; 0 6 j 1:3628g: 0 1 B = 1 C T = 1=h 1 0 0 0h (A) = f01:0; 1:0g Q = I2 = X R = 1 K0 = [2:4142 2:4142] (A 0 BK0 ) = f01:4142; 01:0g h = 500 K = [0:0001 1:5992] F = KC + = 00:0032 (A 0 BF C ) = f0 6 j 0:7741g:
A=
VI. CONCLUSION
Fig. 3.
Trajectory of gain K Example 1.
Example 3:
Fig. 4.
Trajectory of
1K Example 1.
We gave a new perspective on an existing fixed-structure control law, and proposed a new method for computing stabilizing static output feedback gains. The new method only involves Lyapunov equations,
and does not have the technical difficulties residing in the previous approach. We presented an iterative algorithm to solve the new design
13
Fig. 5. Trajectory of Frobenius norm of
KV Example 2.
[4] A. Pimpalkhare and B. Bandyopadhyay, Comments on Stabilization via static output feedback, IEEE Trans. Automat. Contr., vol. Sept., pp. 11481148, 1994. [5] J. Geromel, C. de Souza, and R. Skelton, Static output feedback controllers: Stability and convexity, IEEE Trans. Automat. Contr., vol. 43, pp. 120125, Jan. 1998. , LMI numerical solution for output feedback stabilizations, in [6] Proc. Amer. Control Conf., vol. 1, June 1994, pp. 4044. [7] M. Mesbahi, A semi-definite programming solution of the least order dynamic output feedback synthesis problem, in Proc. Conf. Decision and Control, Dec. 1999, pp. 18511856. [8] F. Leibfritz. Computational design of stabilizing static output feedback controllers. [Online]. Available: http://www.mathematik.uni-trier.de/ ~leibfritz/ [9] M. de Oliveira and J. Geromel, Numerical comparison of output feedback design methods, in Proc. Amer. Control Conf., vol. 1, June 1997, pp. 7276. [10] F. Leibfritz. Trust region methods for solving the optimal output feedback design problem. [Online]. Available: http://www.mathematik.unitrier.de/~leibfritz/ [11] L. Ghaoui, F. Oustry, and M. AitRami, A cone complementarity linearization algorithm for static output-feedback and related problems, IEEE Trans. Automat. Contr., vol. 42, pp. 11711176, Aug. 1997.
Stability of a Riccati Equation Arising in Recursive Parameter Estimation Under Lack of Excitation
Alexander Medvedev
AbstractStability properties of the Riccati equation in a recently suggested antiwindup algorithm for recursive parameter estimation are analyzed. Convergence of the resulting dynamic system is implied by that of a linear time-varying difference matrix equation. By means of converging matrix products theory, the linear mapping associated with the system is shown to be a paracontraction with respect to a certain norm. Therefore, measured in that norm, the solution to the matrix equation will not diverge notwithstanding excitation properties of the data. Thus the purpose of anti-windup is achieved. Index TermsLyapunov methods, Riccati equations, recursive estimation, stability.
Fig. 6. Trajectory of Frobenius norm of
KV Example 3.
equations for the output feedback gains, which is guaranteed to converge. We showed that the new approach, like the previous one, is a dual optimal output feedback LQR theory. Numerical examples clearly illustrated the effectiveness of the algorithm, and validated the new method. Based on the presented framework, our future work will be to come up with a convergent algorithm that generates optimal output feedback gains in the sense of LQR. ACKNOWLEDGMENT The author would like to thank the reviewers for their valuable comments and suggestions. REFERENCES
[1] V. Syrmos, C. Abdallah, and P. Dorato, Static output feedback: A survey, in Proc. 33rd IEEE Conf. Decision and Control, 1994, pp. 837842. [2] Y. Cao, Y. Sun, and W. Mao, A new necessary and sufficient condition for static output feedback stabilizability and comments on stabilization via static output feedback, IEEE Trans. Automat. Contr., vol. 43, pp. 11101111, Sept. 1998. [3] A. Trofino-Neto and V. Kucera, Stabilization via static output feedback, IEEE Trans. Automat. Contr., vol. 38, pp. 764765, June 1993.
I. PRELIMINARIES Consider the following difference Riccati equation typical to parameter estimation of regression models: P (t 0 1)'(t)'T (t)P (t 0 1) P (t) = P (t 0 1) 0 (1) + Q (t ) r(t) + 'T (t)P (t 0 1)'(t)
where P ( 1 ) 2 Rn2n ; Q( 1 ) 2 Rn2n ; Q( 1 ) = QT ( 1 ); Q( 1 ) T n 0; P (0) = P (0); P (0) 0; '( 1 ) 2 R is a regressor vector, r (t) is a positive scalar, and t 2 f1; 2; . . . ; 1g. This equation is typically associated with parameter estimation in the regressor model
y (t) = 'T (t) + e(t)
(2)
Manuscript received September 18, 2003; revised March 31, 2004 and July 30, 2004. Recommended by Associate Editor A. Garulli. This work was supported in part by the Royal Swedish Academy of Sciences and Jernkontoretthe Swedish Steel Producers Association. The author is with the Department of Information Technology, Uppsala University, SE-751 05 Uppsala, Sweden (e-mail: Alexander.Medvedev@it.uu.se). Digital Object Identifier 10.1109/TAC.2004.838481
0018-9286/04$20.00 2004 IEEE
14
1678 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 10, OCTOBER 2006
15
the value of K . For the Van der Pol system, the optimization problem . In this case, (8) remains feasible even if K ! 1, i.e., K 01 the solution of problem (8) provides a slightly poor estimate of the : . However, the stability region is the whole attracting set, R2 , so that the limit cycle is globally stable.
=0
An Improved ILMI Method for Static Output Feedback Control With Application to Multivariable PID Control
Yong He and Qing-Guo Wang
= 6 4912
V. CONCLUSION This note presents a systematic procedure for estimating the attracting set of a class of nonlinear systems called the extended Lurie system. The extended Lurie system consists of systems that can be written like the Lurie problem, but with just one nonlinearity that violates the sector condition. The procedure was applied to two classical nonlinear systems: the forced Dufng equation and the Van der Pol system. The numerical results show that the proposed procedure provides good estimates of the attracting sets. Moreover, for the Van der Pol system the attracting set was proved to be the whole R2 . The calculus of matrix P , using the LMI problems in (7) and (8), tends to improve the stability region estimate since the negative term to be minimized is on the denominator of . However, that improvement is bounded since the norm of matrix P is on the numerator of . APPENDIX
AbstractAn improved iterative linear matrix inequality (ILMI) algorithm for static output feedback (SOF) stabilization problem without introducing any additional variables is proposed in this note. The proposed controller design ILMI algorithm is also extended to solve the SOF problem. They are applied to the multivariable PID controllers. Numerical examples show that the proposed algorithms yield better results and faster convergence than the existing ones. control, linear matrix inequality (LMI), multivariIndex Terms able PID control, stabilization, static output feedback (SOF).
I. INTRODUCTION The static output feedback (SOF) plays a very important role in control theory and applications. Recently, it has attracted considerable attention (see, e.g., [1][7] and the references therein). Yet, it is still left with some open problems. Unlike the state feedback case, a SOF gain which stabilizes the system is not easy to nd. Linear matrix inequality (LMI) [8] is one of the most effective and efcient tools in controller design and a great deal of LMI-based design methods of SOF design have been proposed over the last decade [9][19]. Among these methods, an iterative linear matrix inequality (ILMI) method was proposed by Cao et al. [13] and later employed to solve some multivariable PID controller design problems [20], [21]. In this context, a new additional variable was introduced such that the stability condition becomes a sufcient one. The iterative algorithm in [13] tried to nd a sequence of the additional variables such that the sufcient condition is close to the necessary and sufcient one. The similar idea is used in the so-called substitutive LMI method by [19]. Both of them set the additional matrix variables at the current step with the matrices derived in the preceding step. With additional variables, the dimensions of the LMIs become higher. We observe that it is possible that the matrices derived in the preceding step can be used in the next one directly without introducing the additional variables and the dimensions of the LMIs need not be increased. In addition, we can develop some efcient ways to get suitable initial values in the iterative procedure, which the existing approaches have not been thought of. In this note, a new ILMI algorithm is proposed for SOF stabilization problem without introducing any additional variables, and assisted with a separate ILMI algorithm to nd good initial variables. The algorithms for SOF stabilization are also extended to solve the SOF H1 control problem. They are applied to multivariable PID control. Numerical examples show the effectiveness and an improvement of the algorithms over the existing methods.
: [a; b] 2 D R 2 R ! R be a C function 1 8 1 8 , where and W (t; x) = [x h(t; x) ] 8 6 80 6 is symmetric negative denite. Then, W (t; x) x [1 0 86 8 ]x 8(t; x) 2 [a; b] 2 D, where [1 0 860 8 ] is symmetric negative
Theorem 3: Let h
T n p
h(t;x)
denite.
REFERENCES
[1] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory. Philadelphia, PA: SIAM, 1994. [2] M. Dellnitz and A. Hohmann, A subdivision algorithm for the computation of unstable manifolds and global attractors, Numerische Mathematik, vol. 75, pp. 293317, 1997. [3] J.-Y. Dieudelot and P. Borne, An estimation of the stability domain of a fuzzy controled pendulum using overlapping attractors and vector norms, in Proc. IEEE Int. Conf. Systems, Man, and Cybernetics, 2001, vol. 4, pp. 22452249. [4] M. F. Gameiro and H. M. Rodrigues, Applications of robust synchronization to communication systems, Appl. Anal., vol. 79, pp. 2145, 2001. [5] W. M. Haddad and V. Kapila, Absolute stability criteria for multiple slope-restricted monotonic nonlinearities, IEEE Trans. Autom. Control, vol. 40, no. 2, pp. 361365, Feb. 1995. [6] S. Heidari and C. L. Nikias, Characterizing chaotic attractors using fourth-order off-diagonal cumulant slices, in Proc. 27th Asilomar Conf. Signal, Systems, and Computers, 1993, vol. 1, pp. 466470. [7] R. A. Horn and C. R. Johnson, Matrix Analysis. New York: Cambridge Univ. Press, 1996. [8] H. K. Khalil, Nonlinear Systems. Upper Saddle River, NJ: PrenticeHall, 1992. [9] J. LaSalle and S. Lefschetz, Stability by Liapunovs Direct Method With Applications. New York: Academic, 1991. [10] R. K. Miller, A. N. Michel, and G. S. Krens, Stability analysis of limit cycles in nonlinear feedback systems using describing functions: improved results, IEEE Trans. Circuit Syst., vol. CAS-31, no. 6, pp. 561567, Jun. 1984. [11] H. M. Rodrigues, L. F. C. Alberto, and N. G. Bretas, On the invariance principle: generalizations and applications to synchronization, IEEE Trans. Circuits Syst. I, vol. 47, no. 5, pp. 730739, May 2000. [12] J. Argyris, G. Faust, and M. Haase, An Exploration of Chaos. Amsterdan, The Netherlands: North-Holland, 1994.
Manuscript received April 3, 2005; revised January 12, 2006 and April 23, 2006. Recommended by Associate Editor M. Kothare. The work with Y. He was supported in part by the National Science Foundation of China under Grant 60574014 and in part by the Doctor Subject Foundation of China under Grant 20050533015. Y. He is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore 119260, Singapore. He is also with the School of Information Science and Engineering, Central South University, Changsha 410083, China. Q.-G. Wang is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore 119260, Singapore (e-mail: elewqg@nus.edu.sg). Digital Object Identier 10.1109/TAC.2006.883029
0018-9286/$20.00 2006 IEEE
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 10, OCTOBER 2006 1679
16
II. SOF STABILIZATION Consider the following system:
The key point in our algorithm is to nd an initial P . The P which satises (4) cannot be derived using LMI due to unknown F . By setting V1 = P B2 F , (4) becomes
T P A + AT P + V1 C2 + C2 V1T < 0:
x = Ax(t) + B1 !(t) + B2 u(t) _ z (t) = C1 x(t) + D11 !(t) + D12 u(t) y(t) = C2 x(t) + D21 !(t)
(6)
(1)
where x(t) 2 Rn is the state vector, ! (t) 2 Rr is the external signal, u(t) 2 Rm is the controlled input, z 2 Rq is the controlled output, and y(t) 2 Rp is the measured output. A, B1 , B2 , C1 , C2 , D11 , D12 , and D21 are constant matrices with appropriate dimensions. The following assumption is made on (1). Assumption 1: (A; B2 ) is stabilizable and (C2 ; A) is detectable. The SOF stabilization problem is to nd a SOF controller
However, (6) ignores B2 so that this P does not take into account all the information. On the other hand, (4) is transformed to the following inequality by pro- and postmultiply L = P 01 :
(A + B 2 F C 2 )L + L (A + B 2 F C 2 )
< 0:
(7)
By setting V2 =
F C2 L, (7) becomes AL + LAT

+ B2 V2 + V 2
T BT
2
< 0:
(8)
u ( t) = F y ( t)
where F given by
(2)
2 Rm2p such that the closed-loop system with !(t) = 0

x = ( A + B 2 F C 2 ) x ( t) _
(3)
According to the idea of the cone complementary linearization method [11], L = P 01 yields P L = I , which is relaxed with the following LMI:
P I I L
0
(9)
is stable. As we all know, the closed-loop system (3) is stable if and only if there exists a P = P T > 0 such that
and the linearized version of trace(P L) is minimized. Then, an iterative algorithm to nd an initial P is stated as follows. Algorithm 1
P (A + B2 F C2 ) + (A + B2 F C2 )T P < 0:
(4)
Condition (4) is a BMI which is not a convex optimal problem. An ILMI method was proposed in [13], where a new variable X was introduced such that the stability condition becomes a sufcient one when X 6= P . The algorithm presented in [13] tried to nd some X close to P by using an iterative method and the iterative procedure carries between P and X . On the other hand, a substitutive LMI formulation was proposed in [19], where some new variables such as L and M were introduced such that the stability condition also becomes sufcient one when L 6= R01 B2 P or M 6= F C2 . The algorithm presented in [19] also tried to nd some L and M close to R01 B2 P and F C2 , respectively, by using iterative method. It is clear that so introduced variables are given with the information on P and F in the last step in the iterative procedure, that is, when deriving P , they employed the P obtained in the last step to express these additional variables. In fact, the information on P can be used directly without introducing additional variables. For example, when we derive a P in this step, we can employ it to derive the F in the next step. In the contrary, the F is also used to derive P , and so on. Therefore, these variables in the iterative procedures in [13] and [19] are unnecessary and the iteration can be carried out between P and F directly. Base on this idea, we propose a new algorithm as follows. As mentioned in [13], if
Step 1) Set i = 1 and P0 = I and L0 = I . Step 2) Derive a Pi and Li by solving the following optimization problem for Pi , Li , V1 and V2 : OP1: Minimize trace(Pi 3 Li01 + Li 3 Pi01 ) subject to the following LMI constraints:
T Pi A + AT Pi + V1 C2 + C2 V1T < 0 (10) T + B V + V T BT < 0 ALi + Li A (11) 2 2 2 2 Pi I 0: (12) I Li If trace(Pi Li ) 0 n < "1 , a prescribed tolerance, an initial P = Pi is found, stop.
Step 3)
Step 4) If the difference of two iterations satises trace(Pi Li ) 0 trace(Pi01 Li01 ) < "2 , a prescribed tolerance, the initial P may not be found, stop. Step 5) Set i = i + 1, goto Step 2). If an initial P can not be found by Algorithm 1, the SOF control problem for system (1) with ! (t) = 0 may not have solutions. On the other hand, after an initial P is found, an ILMI algorithm that stabilizes system (1) with ! (t) = 0 using SOF is stated as follows. Algorithm 2 Step 1) Set i = 1 and P1 = P as obtained from Algorithm 1. Step 2) Solve the following optimization problem for F with given Pi : OP1: Minimize i subject to the following LMI constraint:
P (A + B2 F C2 ) + (A + B2 F C2 )T P 0 P < 0
(5)
holds, the closed-loop system matrix A + B2 F C2 has its eigenvalues in the strict left-hand side of the line =2 in the complex s-plane. If a 0 satisfying (5) can be found, the SOF stabilization problem is solved.
P i (A + B 2 F C 2 ) + ( A + B 2 F C 2 )T P i 0 i P i < 0:
(13)
17
TABLE I SOF RESULTS (EXAMPLE 1)
Step 3) If i 0, F is a stabilizating SOF gain, stop. Step 4) Set i = i + 1. Solve the following optimization problem for Pi with given F : OP2: Minimize i subject to the aforementioned LMI constraint (13). Step 5) If i 0, F is a stabilization SOF gain, stop. Step 6) Solve the following optimization problem for Pi with given F and i : OP3: Minimize trace(Pi ) subject to the previous LMI constraint (13). Step 7) If kPi 0 Pi01 k=kPi k < , a prescribed tolerance, goto Step 8), else set i = i + 1 and Pi = Pi01 , then goto Step 2). Step 8) The system may not be stabilizable via SOF, stop. Remark 1: The discussions on the iterative procedure and convergence of Algorithm 2 follow those in [13]. Remark 2: If C2 = I , the SOF stabilization problem reduces to a state feedback problem. In fact, the initial P derived in Algorithm 1 is also the solution of state feedback stabilization problem since the state feedback gain F can be derived by F = V2 L01 = V2 P . On the other hand, the F can also be obtained directly by OP1 in Algorithm 2 when the previous initial P is given. Thus, the state feedback stabilization gain without conservativeness can also be derived using our Algorithms 1 and 2. Remark 3: The algorithm presented in our note is different from those in [11]. The stopping criterion in [11] is given in terms of "eof which depends on selected and , where and should be sufciently small. However, it is difcult to determine how small the values should be as they depend on a particular situation under consideration. If inappropriate and are selected, the algorithm may not be convergent. In our procedure, Algorithm 1 nds an initial P for a given "1 , where the selection of "1 is not crucial as P obtained is not the nal solution but will be elaborated in Algorithm 2. With the initial P , Algorithm 2 produces a static output feedback gain matrix F which can guarantee the stability of closed-loop system (3). The stopping condition is < 0, where need not be specied a prior. Overall, our procedure is easier to use. Example 1: [13] Consider the SOF stabilization problem of system (1) with ! (t) = 0 and the following parameter matrices:
0 1 1 0 1 0
Example 2: [7] Consider the SOF stabilization problem of system (1) with ! (t) = 0 and the following parameter matrices:
A=
B2 = C2 =
02 08 5 01 08 4 07 06 03 02 2 6 03 7 5 2 10 01 03 8 1 2 3 07 0 05 6 03 04 6 1 2 8 04 6 09 02 04 5 8 3 1 9 06 3 0 3 :9 2 0 :1 0 2 :5 0 1 2 :5 0 1 0 :5 0 :5 0 1 0 0 :5 1 2 00:05 3 6 05 02 01 07 5 : 01 04 07 01 06 05 03
04 09 07 06
(15)
00:8871 4:9310 using Algorithm 2 with two itera00:6576 0:9869 tions. In this case, = 00:7481 and the eigenvalues of the closed-loop system are 07:8846 6 j 36:0334, 00:6354 6 j 12:2411, 00.3742, 04:9779 6 j 6:3825. The convergence speed is faster than that in [7].
F is found as
III.
After more than 250 iterations, the algorithm in [7] converges. Algorithm 1 yields an initial P with only four iterations. Then, a SOF gain
H1 SYNTHESIS
An ILMI algorithm presented in the preceding section is now employed to solve the SOF H1 control problem. The objective of the SOF H1 synthesis is to to nd a SOF controller (2) such that the transfer function of the closed-loop system satises H1 norm constraint
kTz! (s)k1 < ; for > 0:

(16) can be represented as a matrix inequality [13]
(16)
T P Acl + AT P P Bcl Ccl cl TP T Bcl 0 I Dcl < 0 Ccl Dcl 0 I

where
(17)
A=
B2 =
C2 = [1 ]:
(14)
The open-loop system is unstable since the eigenvalues are 1 and 01. The calculation results for SOF stabilization using the method in [13] and our algorithms are listed in Table I. It is noted that the iteration number in form of l + k in the table means that l is the iteration number of Algorithm 1 to nd an initial P and k is the iteration number of Algorithm 2 to nd a SOF gain. It can be seen that our convergence speed is greatly faster than that in [13].
Acl = A + B2 F C2 Bcl = B1 + B2 F D21 Ccl = C1 + D12 F C2 Dcl = D11 + D12 F D21 :

Similarly to Algorithm 2, we propose an algorithm to obtain the solution of matrix inequality (17) for a given > 0. It relies on, like as Algorithm 1, an algorithm for nding an initial P for SOF H1 control problem.
18
TABLE II SOF AND PID-CONTROLLER AND THEIR PERFORMANCES (EXAMPLE 4)
Algorithm 3
Step 1) Set i = 1 and P0 = I and L0 = I . Step 2) Derive a Pi and Li by solving the following optimization problem for Pi , Li , V1 , and V2 : OP1: Minimize trace(Pi 3 Li01 + Li 3 Pi01 ) subject to the LMI constraints (18) and (19), as shown at the bottom of the page, and (20)
Step 3) If trace(Pi Li ) 0 n < "1 , a prescribed tolerance, an initial P = Pi is found, stop. Step 4) If the difference of two iterations satises trace(Pi Li ) 0 trace(Pi01 Li01 ) < "2 , a prescribed tolerance, the initial P may not be found, stop. Step 5) Set i = i + 1, goto Step 2). After an initial P is found, an ILMI algorithm for SOF H1 control problem for system (1) is stated as follows. Algorithm 4
Pi I
I Li
0:
(20)
Step 1) Set i = 1 and P = P1 as obtained from Algorithm 3.
T T T T Pi A + AT Pi + V1 C2 + C2 V1T Pi B1 + V1 D21 C1 + C2 F T D12 T P + DT V T T + DT F T DT B1 i 0 I D11 2 1 21 12 < 0 C1 + D12 F C2 D11 + D12 F D21 0 I T T T ALi + Li AT + B2 V2 + V2T B2 B1 + B2 F D21 C1 Li + V2T D12 T + DT F T DT T + DT F T B T 0 I D11 B1 12 < 0 21 2 21 0 I D11 + D12 F D21 Li C1 + D12 V2
(18)
(19)
19
Step 2) Solve the following optimization problem for F with given Pi : OP1: Minimize i subject to the LMI constraint (21), as shown at the bottom of the page, where
by Algorithm 3 after 24 iterations. Then the SOF H1 norm converges to 1.144 after two iterations using Algorithm 4. The resulting SOF gain is
811 = Pi
A + AT P
+ Pi B 2 F C 2
+ C T F T BT P
2 2
0 Pi :
Step 3) If i 0, F is a stabilization SOF H1 control gain for , stop. Step 4) Set i = i + 1. Solve the following optimization problem for Pi with given F : OP2: Minimize i subject to the above LMI constraint (21). Step 5) If i 0, F is a stabilizating SOF H1 control gain for , stop. Step 6) Solve the following optimization problem for Pi with given F and i : OP3: Minimize trace(Pi ) subject to the above LMI constraint (21). Step 7) If kPi 0 Pi01 k=kPi k < , a prescribed tolerance, goto Step 8), else set i = i + 1 and Pi = Pi01 , then goto Step 2). Step 8) It may not be decided by this algorithm whether SOF H1 control problem is solvable, stop. Example 3: [19] Consider system (1) with the following parameter matrices:
0:0976 03:8054 00:4191 4:6958 : In this case, = 08:2776 2 1007 and the eigenvalues of the closedloop system are 050.1599, 00.2781, 00:2431 6 j 1:3534.
F=
IV. PID CONTROL Motivated by its popularity in industry, let us consider now the following PID controller:
t
0
u(t) = F1 y(t) + F2
_ y()d + F3 y(t)
(22)
instead of SOF controller (2), where F1 ; F2 ; F3 2 Rm2p are gain matrices to be designed. Without lose of generality, D21 is set as zero. It is noted that a new method is proposed in [20] to transform this PID controller design problem to a SOF control problem. In order to using the method proposed in [20], the following assumption is needed. Assumption 2: The matrix I 0 F3 C2 B2 is invertible. Denote x(t) =
x ( t) 0 y ( )d
t
and y (t) =
posite system with (1) and (22) is described by
C 2 x ( t) t y()d C2 Ax(t)
0
. The com-
00:0366 0:0271 0:0188 00:4555 0:0482 01:01 0:0024 04:0208 A= 0:1002 0:3681 00:707 1:42
B1 = C1 = D11 = D21 =
0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 :5 0 0 0 1 0 0 0 :1 0 0 0 0 :1
1 0:4422 3:5446 B2 = 05:52 0 1 0 C2 = 0 1 1 0 D12 = 0 1

:
0 0:1761 07:5922 4:49 0 0 0 0 0
_ x = A x ( t) + B 1 ! ( t) + B 2 u ( t) 1x(t) + D11 !(t) + D12 u(t) z ( t) = C 2x(t) y ( t) = C

where
(23)
T T T T C2 = C21 C22 C23 D11 = D11 D12 = D12 :

The controller reduces to
A 0 B B B1 = 1 B2 = 2 0 0 C 0 C21 = [C2 0] C22 = [0 I ] C23 = [C2 A 0]
A=
C1 = [C1 0]
The SOF H1 norm obtained in [19] is 1.183. For matrix P is obtained as
= 1:144, the initial

where
u ( t) = F y ( t) Fi = (I 0 F3 C2 B2 )01 Fi ; F = [ F 1 F 2 F 3 ]:
i = 1 ; 2; 3
(24)
0:7130 00:4054 P1 = 00:3283 01:1658
00:4054 00:3283 01:1658

2:2769 1:4029 1:2100 1:4029 1:5941 1:0533 1:2100 1:0533 3:2305
BT P
1
C1 + D12 F C2
+ DT
811
21
F T BT P
2
T T T Pi B1 + Pi B2 F D21 C1 + C2 F T D12 T + DT F T DT 0 I D11 21 12 < 0 D11 + D12 F D21 0 I
(21)
20
1:0000 0 0 0 1 0 0 0 C1 = [0 1 0 0 0 0] C2 = 0 0 0 1 0
A=
00:0266 036:6170 018:8970 032:0900 3:2509 00:7626 0:0001 01:8997 0:9831 00:0007 00:1708 00:0050 0:0123 11:7200 02:6316 0:0009 031:6040 22:3960
0 0 0 0 0 0 0 0 0 0 0
0 0 0 B1 = 0 0 0 030:0000 0 30 0 030:0000 0 0 D11 = [0] D12 = [1 1] D21 = : 0
B2 =
0 0 0 0 0 0 0 0 30 0 0 30
(25)
Once the composite matrix F = [F1 F2 F3 ] is found, the original PID controller gain can be recovered from F3 = F3 (I + C2 B2 F3 )01 F2 = (I 0 F3 C2 B2 )F2 F1 = (I 0 F3 C2 B2 )F1 : As stated in [20], the invertibility of matrix I + C2 B2 F3 is guaranteed by the following Lemma. Lemma 1: Matrix I + C2 B2 F3 is invertible if and only if Assump tion 2 holds, where F3 and F3 are related to each other by F3 = (I 0 F3 C2 B2 )01 F3 ; or F3 = F3 (I + C2 B2 F3 )01 :
Under Assumption 2, one easily see that Algorithms 1 and 2 can be employed to derive the stabiliziable PID control gains, F1 , F2 , and F3 . Similarly, Algorithms 3 and 4 can be used to derive the PID H1 controller for a given performance > 0. Example 4: [20] One of the state space realization of the aircraft controller system model is as system (1) with the following parameters in (25), as shown at the top of the page. The calculation results using our algorithms and the method in [13] and [20] for SOF control and PID control are listed in Table II. It can be seen that the convergence speed is faster than that in [13], [20]. On the other hand, Remark 3 is demonstrated in this example. For example, when "1 in Algorithm 3, which is similar to those in [11], is set as 1003 , two iterations obtain an initial P . However, based on this P , F is not a stabilization H1 control gain for given = 1:001. On the contrary, the corresponding F in Table II is derived after 14 iterations by using Algorithm 4. V. CONCLUSION New ILMI algorithms for SOF stabilization and H1 control are proposed in this note, which avoid introduction of the additional variables, leading to lower dimensions of the LMIs than [13] and [19]. In particular, an algorithm to derive an initial value is also given. The algorithms are also applied to design the multivariable PID controller. Numerical examples show that the proposed algorithms produces better result and/or faster convergence than the existing ones.
REFERENCES
[1] D. S. Bernstein, Some open problems in matrix theory arising in linear systems and control, Linear Alg. Appl., vol. 162164, pp. 409432, 1992. [2] A. Trono and V. Kucera, Stabilization via static output feedback, IEEE Trans. Autom. Control, vol. 38, no. 5, pp. 764765, May 1993.
[3] V. Kucera and C. E. Souza, A necessary and sufcient condition for output feedback stabilizability, Automatica, vol. 31, pp. 13571359, Sep. 1995. [4] M. Fu and Z. Q. Luo, Computational complexity of a problem arising in xed order output feedback design, Syst. Control Lett., vol. 30, pp. 209215, 1997. [5] V. L. Syrmos, C. T. Abdallah, P. Dorato, and K. Grigoriadis, Static output feedbackA survey, Automatica, vol. 33, pp. 125137, 1997. [6] H. Gao, J. Lam, C. Wang, and Y. Wang, Delay-dependent output-feedback stabilisation of discrete-time systems with time-varying state delay, Proc. Inst. Elect. Eng. Control Theory Appl., vol. 151, pp. 691698, 2004. [7] J. T. Yu, A convergent algorithm for computing stabilizing static output feedback gains, IEEE Trans. Autom. Control, vol. 49, no. 12, pp. 22712275, Dec. 2004. [8] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory. Philadelphia, PA: SIAM, 1994, vol. 15, SIAM Studies in Applied Mathematics. [9] J. C. Geromel, C. C. de Souza, and R. E. Skelton, LMI numerical solution foroutput feedback stabilization, in Proc. Amer. Control Conf., 1994, pp. 4044. [10] T. Iwasaki and R. E. Skelton, The XY-centring algorithm for the dual LMI problem: a new approach to xed-order control design, Int. J. Control, vol. 62, pp. 12571272, 1995. [11] L. E. Ghaoui, F. Oustry, and M. AitRami, A cone complementarity linearization algorithm for static output-feedback and related problems, IEEE Trans. Autom. Control, vol. 42, no. 8, pp. 11711176, Aug. 1997. [12] R. E. Benton and D. Smith, Static output feedback stabilization with prescribed degree of stability, IEEE Trans. Autom. Control, vol. 43, no. 10, pp. 14931496, Oct. 1998. [13] Y. Y. Cao, J. Lam, and Y. X. Sun, Static output feedback stabilization: An ILMI approach, Automatica, vol. 34, pp. 16411645, 1998. [14] Y. Y. Cao, Y. X. Sun, and W. J. Mao, A new necessary and sufcient condition foroutput feedback stabilizability and comments on stabilization via static output feedback, IEEE Trans. Autom. Control, vol. 43, no. 8, pp. 11101111, Aug. 1998. [15] Y. Y. Cao, Y. X. Sun, and J. Lam, Simultaneous stabilization via static output feedback and state feedback, IEEE Trans. Autom. Control, vol. 44, no. 6, pp. 12771282, Jun. 1999. [16] C. A. R. Crusius and A. Trono, Sufcient LMI conditions for output feedback control problems, IEEE Trans. Autom. Control, vol. 44, no. 5, pp. 10531057, May 1999. [17] F. Leibfritz, An LMI-based algorithm for designing suboptimal static output feedback controllers, SIAM J. Control Optim., vol. 39, pp. 17111735, 2001. [18] M. Rotunno and R. A. de Callafon, A Bundle method for solving the xed order control problem, in Proc. 41st IEEE Conf. Decision and Control, Las Vegas, NV, Dec. 2002, pp. 31563161. [19] A. Fujimori, Optimization of static output feedback using substitutive LMI formulation, IEEE Trans. Autom. Control, vol. 49, no. 6, pp. 995999, Jun. 2004. [20] F. Zheng, Q. G. Wang, and T. H. Lee, On the design of multivariable PID controllers via LMI approach, Automatica, vol. 38, pp. 517526, 2002. [21] C. Lin, Q. G. Wang, and T. H. Lee, An improvement on multivariable PID controller design via iterative LMI approach, Automatica, vol. 40, pp. 519525, 2004.
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 16, NO. 3, JUNE 2008 615
21
Control Synthesis of Singularly Perturbed Fuzzy Systems

Guang-Hong Yang, Senior Member, IEEE, and Jiuxiang Dong
AbstractThis paper considers the problem of designing stabicontrollers for nonlinear singularly perturbed syslizing and tems described by TakagiSugeno fuzzy models with the consideration of the bound of singular perturbation parameter . For the synthesis problem of simultaneously designing the bound of and stabilizing or controllers, linear matrix inequalities (LMI)based methods are presented. For evaluating the upper bound of subject to stability or a prescribed performance bound constraint for the resulting closed-loop system, sufcient conditions are developed, respectively. For the stabilizing and control synthesis without the consideration of improving the bound of , new design methods are also given in terms of solutions to a set of LMIs. Examples are given to illustrate the efciency of the proposed methods. Index Terms performance, linear matrix inequalities (LMIs), nonlinear control systems, singularly perturbed systems, stabilizing control, state feedback control, TakagiSugeno (T-S) fuzzy models.
I. INTRODUCTION
N CONTROL engineering applications, it is well known that the multiple time-scale systems or known as singularly perturbed systems often raise serious numerical problems. For the purpose of avoiding the difculties linked with the stiffness of the equations involved in the design, the singular perturbation design method has been developed [1], where singular perturbation with a small parameter, say , is exploited to determine the degree of separation between slow and fast modes of the system, and the so-called reduction technique is proposed to control of handle these systems. For the stabilization and linear singularly perturbed systems, many important advances have been achieved, see [1][6] and the references therein. In particular, the fundamental results are given in [1], [4], and [5].
Manuscript received April 19, 2006; revised January 11, 2007. This work was supported in part by the Program for New Century Excellent Talents in University (NCET-04-0283), by the Funds for Creative Research Groups of China (No. 60521003), by the Program for Changjiang Scholars and Innovative Research Team in University (No. IRT0421), by the State Key Program of National Natural Science of China under Grant 60534010, by the Funds of National Science of China under Grant 60674021, and by the Funds of the Ph.D. program of MOE, China, under Grant 20060145019. G.-H. Yang is with the College of Information Science and Engineering, Northeastern University, Shenyang 110004, China, and also with the Key Laboratory of Integrated Automation of Process Industry, Northeastern University, Ministry of Education, Shenyang 110004, China (e-mail: yangguanghong@ise.neu.edu.cn; yang_guanghong@163.com). J. Dong is with the College of Information Science and Engineering, Northeastern University, Shenyang 110004, China (e-mail: dongjiuxiang@ise.neu. edu.cn; dong_jiuxiang@163.com). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TFUZZ.2007.905911
In recent years, there have been some attempts to address the control problem for nonlinear singularly perturbed syscontrol for a class of singularly tems. In [7] and [8], the perturbed systems with nonlinearity in the slow variables is excontrol problem for afne amined. A local state feedback nonlinear singularly perturbed systems is studied in [9]. Howcontrol design for general nonlinear singularly ever, the perturbed systems still remains as an open research subject. An important approach to the synthesis problems for nonlinear systems is to model the considered system as Takagi and Sugeno (T-S) fuzzy systems [10], which are locally linear timeinvariant systems connected by IF-THEN rules. In [11] and [12], it has shown that the T-S fuzzy systems can approximate any continuous functions at any preciseness, which shows that the T-S fuzzy models can approximate a wide class of nonlinear systems. As a result, the conventional linear system theory can be applied to analysis and synthesis of the class of nonlinear control systems. In recent years, the T-S fuzzy control systems have been studied extensively, and many signicant advances have been achieved (see [13][15] and the references therein). For nonlinear singularly perturbed systems, some control synthesis problems have been studied [16][19]. In [16] and [18], control of nondesign methods for the stabilization and linear singularly perturbed systems via state feedback are given in terms of solutions of linear matrix inequalities (LMIs) [20], respectively. A robust state feedback control design is presented output feedin [17]. An LMI-based method of designing back controllers for uncertain fuzzy singularly perturbed systems is presented in [19]. For the effective applications of the design methods for singularly perturbed control systems, the accurate knowledge of the stability bound of a singularly perturbed system (i.e., the ) is very important. The characsystem is stable for terization and computation of the stability bound have attracted considerable efforts for the past over two decades [21][28]. In general, there are two classes of methods to characterize and compute the stability bounds, one is based on frequency domain transfer functions and another is based on state space models. Both of the two methods can provide the exact bounds as shown in [22], [25], and [26]. However, the issue of how to improve the bound in controller designs has not been addressed in the literature, which undoubtedly is very important for the applications of singularly perturbed system theory. This paper is concerned with the problem of designing stacontrollers for nonlinear singularly perturbed bilizing and systems described by T-S fuzzy models with the consideration of the bound of singular perturbation parameter . Two LMI-based methods are presented for simultaneously designing
1063-6706/$25.00 2007 IEEE
616 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 16, NO. 3, JUNE 2008
22
the bound of and stabilizing or controllers for a fuzzy singularly perturbed system, respectively. For the issue of computing the bound of singularly perturbed parameter , sufcient conditions are derived for evaluating the upper bound of subject to stability or a prescribed performance bound con, restraint for the resulting closed-loop system for spectively, where the upper bound can be obtained by solving a generalized eigenvalue problem (GEVP) [20]. For the problem controllers without the considof designing stabilizing and eration of improving the bound of , design methods are also given in terms of solutions to LMIs. The paper is organized as follows. In Section II, the system description, the considered problems, and preliminary lemmas are presented. Section III considers the problem of designing stabilizing controllers and the bound of , and the case for controller design is studied in Section IV. Section V gives examples to illustrate the effectiveness of the new proposed methods. Finally, Section VI concludes this paper. Notation: For a matrix , is dened as the largest singular value of . For a square matrix , is dened as
is the grade of membership of it is assumed that
in
, where
Let
then (2) is said to be normalized membership functions. The T-S fuzzy model of (1) is inferred as follows:
The symbol
within a matrix represents the symmetric entries
. . .
. . .
..
. . . (3)
II. SYSTEM DESCRIPTION AND PROBLEM STATEMENT A. System Description The class of nonlinear singularly perturbed systems under consideration is described by the following fuzzy system model:
The system (3) can be rewritten as follows:
(4) where
IF is THEN
and
is
is
(1) where variables, are fuzzy sets, are the premise and are the state vectors, is the control input, is the disturbance, is the controlled output, the matrices , , , , , , , , , , and are of appropriate is a dimensions, is the number of IF-THEN rules, and small constant. Denote In this paper, the concept of parallel distributed compensation (PDC) is used to design fuzzy controller, i.e., the designed fuzzy controller shares the same fuzzy sets with the fuzzy model in
YANG AND DONG: CONTROL SYNTHESIS OF SINGULARLY PERTURBED FUZZY SYSTEMS 617
23
the premise parts (more details can be found in [13]). For the fuzzy model (1), the following state feedback controller [13] is adopted:
C. Preliminaries The following preliminaries will be used in the sequel. be a constant. For the fuzzy control system (7), let If (7) is asymptotically stable, and for any (the space of square integrable functions) and , the following inequality holds:
IF is THEN
and
is
is (5)
Because the controller rules are same as plant rules, we obtain the state feedback controller as follows: (6) Combining (6) and (4), then the resulted closed-loop system is given as follows:
then the system (7) is said to be with an equal to [29], [30]. Denote
-norm less than or
then the closed-loop system (7) can be rewritten as follows: (7) (8) B. Problem Statement In this paper, the following problems will be addressed. Controller Design With Consideration of Bound of : , and an as big as possible (i) Find is such that the closed-loop system (7) with asymptotically stable for any and all satisfying (2). be a given constant. Find (ii) Let and an as big as possible such that the closed-loop system (7) is asymptotically stable and with an -norm and all less than or equal to for any satisfying (2). Evaluation of Bound of : be given. Find an as (iii) Let big as possible such that the closed-loop system (7) with is asymptotically stable for any and all satisfying (2). and be given. Find an (iv) Let as big as possible such that the closed-loop system (7) is -norm less than or asymptotically stable and with an equal to for any and all satisfying (2). Remark 1: Problems (i) and (ii) are concerned with simuland nding the upper taneously designing bound of with guaranteeing the stability and performance of the closed-loop system (7), respectively. Problems (iii) and (iv) are related to the problem of nding the upper bound of subject to that the closed-loop system (7) is asymptotically stable or with the constraint of an performance bound when are given. Moreover, the controller design problems without the consideration of bound of will also be studied in Sections III and IV, respectively. The following lemma from [29] gives a sufcient condition for the system (8) to be with an -norm less than or equal to . Lemma 1: [29] Consider the system (8). If there exists a positive denite matrix such that
holds. Then the system (8) is asymptotically stable and with an -norm less than or equal to . Lemma 2: If there exist symmetric matrices , , , , and matrices , , such that the following LMIs hold: (9) (10) (11) where
then . . . . . . (12)
24
Proof:
renders the singularly perturbed fuzzy system (7) with an norm less than . . . . III. STABILITY BOUND AND STABILIZATION In this section, a method of simultaneously designing the upper bound of and stabilizing controller gains is derived where the upper bound of singularly perturbed parameter can be improved, which provides a solution to Problem (i). For solving Problem (iii), a method of computing the upper bound of singularly perturbed parameter subject to the stability of the closed-loop system is presented. Moreover, a technique for designing stabilizing controllers for singularly perturbed fuzzy systems without consideration of improving the bound of singularly perturbed parameter is also presented in Section III-C. A. Design of Stability Bound of and Stabilizing Controllers The following theorem presents a method of simultaneously designing the upper bound of and stabilizing controller gains. Theorem 1: If there exist matrices , , , , , , , , , and positive scalars , , , , with
. . .
. . .
. . .
..
. . .
From (9) to (11), it follows (12). Lemma 3: [16] If there exist matrices with
and
where and lowing LMIs:
are symmetric matrices, satisfying the fol(13) (14)
where , , , , , matrices, satisfying the following LMIs:
are symmetric (17) (18) (19) (20) (21) (22)
such that, for then there exists a scaler state feedback controller (6) with
, the
renders the singularly perturbed fuzzy system (7) asymptotically stable. , if there exist matrices Lemma 4: [18] For given and , , with
(23) (24) (25) (26)
where and are symmetric matrices, satisfying the following LMIs, shown in (15) and (16) at the bottom of this page, such that, for , the then there exists a scaler state feedback controller (6) with
where
(15)
He
(16)
25
Combining it with , , (28) can be minimized by solving the optimization problem minimize subject to
(17)(26)
(29)
and denote (27) where
are positive weighting constants to where , be chosen. Since the constraints (17)(26) are of LMIs, the optimization problem can be effectively solved via LMI Control Toolbox [31]. Regarding the issue of how to choose the , generally, one can weighting scalars . It choose bigger for rendering smaller should be pointed out that the upper bound obtained by solving the previous optimization problem may be conservative. After obtaining stabilizing controller gains, a less conservative bound of can be obtained by Theorem 2 in Section III-B. B. Computation of Stability Bound of In this subsection, assume that the controller has been designed. The following theorem gives a technique to estimate the upper bound of singularly perturbed parameter subject to the stability of the closed-loop system. , Theorem 2: If there exist matrices , , , , and a constant with
then for
, the state feedback controller (6) with
renders the singularly perturbed fuzzy system (7) asymptotically stable. Proof: See Appendix A. Remark 2: Theorem 1 presents sufcient conditions under of singularly perturbed parameter which an upper bound and stabilizing controller gains can be obtained. From (27) and , , can be minimized by solving the following optimization problem: minimize subject to (17)(26) (28)
where , , , are symmetric matrices, satisfying (9)(11), and the following LMIs: (30) (31)
are positive weighting constants to be where , chosen. However, the optimization problem cannot be solved . Consider directly due to the term
(32)
(33) , , , . where , the singularly perturbed closed-loop Then, for fuzzy system (7) is asymptotically stable, where
(34)
(35) From (25) and (26), we have Proof: From condition (30) and (31), we have (36) then it follows that By , and (36), it follows that for (37)
26
Pre- and post multiplying (37) by
On the other hand, from (32) and (33), we have
and its transpose, then the following inequality holds: for Consider the following Lyapunov function:
for Applying (9)(11), (39) and Lemma 2, it follows . . . . . .
(39)
combined with (38), we have then for Thus, for , the closed-loop singularly perturbed fuzzy system (7) is asymptotically stable. Remark 3: By Theorem 2, an upper bound of can be obtained by solving inequalities (9)(11), (30)(33) for . The optimization problem Minimize subject to (9), (10), (11), (30)(33)
is a generalized eigenvalue problem (GEVP) [20], which can be effectively solved using LMI Control Toolbox [31]. The problem of computing the bound of was considered in [17], so that the system where a method of nding an interval was derived. However, to search for is stable for is related to -dependent computation, which cannot small avoid the difculties linked with the stiffness of the equations involved in the design. C. Stabilizing Controller Design Without Considering Stability Bound In this section, a technique for designing stabilizing controllers for singularly perturbed fuzzy systems without consideration of improving the bound of singularly perturbed parameter , is given as follows. , Theorem 3: If there exist matrices , , , , , with
where , , , satisfying the following LMIs:
are symmetric matrices,
(40) . . . . . . (38) where , are given by (34) and (35). (41) (42) (43) (44)
27
where
(49) (50) From (47), we have that there exist a scalar for such that
then there exists a scaler such that, for state feedback controller (6) with
, the where that
(51) , . By (48)(50), (51) and Lemma 2, it follows
renders the singularly perturbed fuzzy system (7) asymptotically stable. Proof: By i.e., and , , we choose invertible formula, it follows that , then by matrix
Since that
and
, there exists a scaler , such for , which implies combined with (45), then we have
Choose Lyapunov function
for which implies that the closed-loop singularly perturbed fuzzy system (7) is asymptotically stable, and from (46), we obtain
then
(45) On the other hand, substitute (46) for in (44) and pre- and post multiplying (44) by and , then we can obtain (47) where
Thus, the proof is complete. Remark 4: Theorem 3 presents a method for designing state feedback stabilizing controllers for singularly perturbed fuzzy systems. However, the upper bound of singularly perturbed parameter is not addressed in the design. The following theorem shows that the new method given in Theorem 3 is less conservative than that given in Lemma 3 [16]. Theorem 4: If the condition in Lemma 3 holds, then the condition in Theorem 3 holds. Proof: Assume that the condition in Lemma 3 holds, then , choose , , . Then it follows that conditions (40)(43) hold from (13) and (14). Moreover, from , the following inequalities hold: (14) and the chosen
Now, pre- and post multiplying (41)(43) by follows that
and
, then it
. . .
. . .
..
. . .
(48)
28
which implies that (44) holds. Therefore, the condition of Theorem 3 holds. Thus, the proof is complete. IV. BOUND OF AND CONTROL
Denote
In this section, the results in Section III are extended to the case that the closed-loop system is required to be with an performance bound. Solutions to Problem (ii) and (iv) formulated in Section II-B are, respectively, given in Section IV-A and IV-B. Moreover, Section IV-C also presented a new method for controllers without consideration of improving designing the bound of singularly perturbed parameter . A. Design for Bound of and Performance
(62) where (63) then for , the singularly perturbed closed-loop fuzzy system (8) via state feedback controller
conIn this subsection, a method is given for designing trollers with consideration of improving the bound of singularly perturbed parameter . Theorem 5: For given , if there exist matrices , , , , , , , , , and positive scalar variables , , , , , with
where
is asymptotically stable and with an norm less than . Proof: See Appendix A. Remark 5: If the conditions of Theorem 5 hold, then is an upper bound of singularly perturbed parameter , and
where , , , , , matrices, satisfying the following LMIs:
are symmetric
(52) (53) (54) (55) (56) (57) The problem of minimizing can be reduced to solving the following optimization problem with LMI constraints: Minimize subject to (52)(61) where , , be chosen. , , and are positive weighting constants to
B. Computation of Upper Bound of In this subsection, we assume that the state feedback gains are given. Then the following theorem gives a method to estimate the upper bound of singularly perturbed parameter subject to performance bound. the closed-loop system with an , if there exists , , , Theorem 6: For a given , , and a constant with
(58) (59) (60) (61) where , are the same as in Theorem 3, and
where , , , are symmetric matrices satisfying (9)(11) and the following LMIs: (64) (65) (66) (67)
29
where
C.
Controller Design
In this subsection, a new method is given for designing controllers, but without consideration of improving the bound of singularly perturbed parameter . Theorem 7: For given , if there exist matrices , , , , , , with
where
where , , , and trices satisfying the following LMIs:
are symmetric ma-
then, for , the singularly perturbed closed-loop fuzzy system (8) is asymptotically stable and with an norm less than . Proof: From (64) to (67), it follows that, for any
where
(68) and
By (9)(11), and Lemma 2, which further implies that . . . On the other hand . . . then there exists a scaler such that, for , the singularly perturbed fuzzy system (7) via the state feedback controller (6) with
(69)
. . .
. . .
is asymptotically stable and with an norm less than . Proof: It is similar to Theorem 3, and omitted here. The following theorem shows that the new method given in Theorem 7 is less conservative than that given in Lemma 4 [18]. Theorem 8: If the condition of Lemma 4 holds, then the condition of Theorem 7 holds. Proof: It is similar to Theorem 4 and omitted here. V. EXAMPLE We consider an inverted pendulum controlled by a motor via a gear train. It can be described by the following state equations [32]:
Thus, from Lemma 1 and (69), the conclusion follows. Remark 6: When the state feedback gains are given, Theorem 6 gives a method to estimate the upper bound of singularly perturbed parameter subject to that the closed-loop system performance bound. is required to be with a prescribed An upper bound of can be obtained by solving inequalities (9)(11) and (64)(67) for . The optimization problem Minimize subject to (9), (10), (11), (64)(67)
is also a generalized eigenvalue problem (GEVP) [20], which can be effectively solved using LMI Control Toolbox [31].
30
TABLE I UPPER BOUNDS OF VIA THEOREM 6
where , , , is the control input, is the disturbance input, is the motor torque constant, is the back emf constant, and is the gear ratio. The parameters for the plant are given as 9.8 m/s , 1 m, 1 kg, , 0.1 Nm/A, 0.1 Vs/rad, 1 and mH. Note that the inductance represents the small parasitic parameter in the system. Then, we get
Choose the membership functions of the fuzzy sets as
Lemma 4, Theorems 5 and 7 are applicable for designing fuzzy controllers for the system. By using Lemma 4 and Theorem 7, the obtained optimal performance indexes are and , respectively, and the corresponding controller gains are given as follows:
(70) This fuzzy model exactly represents the dynamics of the non. A T-S fuzzy linear mechanical system under model can be obtained as follows Plant Rule 1: is IF THEN and Plant Rule 2: is IF THEN (72) By using Theorem 6, the upper bounds of subject to guaranteeing performance bound can be estimated, and shown in Table I. From Table I, it is easy to see that the new proposed design given by Theorem 5 gives a considerable improvement of upper bounds of . The upper bounds of given by Lemma 4 and Theorem 7 are very small. Assume the initial states are zero, and the disturbance input is shown in Fig. 1. The simulation results of the signal output with the controller gains (70)(72) are given in Figs. 24, respectively. The simulations for the square root of ratio of the regulated output energy to the disturbance input noise energy are depicted in Figs. 57. From the simulation results, it can be seen that the fuzzy controller (72) guarantees good performance of the resulting closed-loop system, while the controllers (70) and (71) give poor system responses. , then the controller gains are given as follows: (71) Now we apply Theorem 5 to design a fuzzy the system. Choose weighting scalars as controller for
where
31
Fig. 1. Disturbance input w (t) used during simulation.
Fig. 4. Trajectory of z (t) via controller (72).
Fig. 5.
(t)z (t)dt=
(t)w (t)dt via controller (70).
Fig. 6.
(t)z (t)dt=
32
which further implies that for Let , then can be expressed as follows: (76)
So (76) is equivalent to (77) Pre- and postmultiplying (77) by

Fig. 7.
(t)z (t)dt=
VI. CONCLUSION In this paper, we have studied the problem of designing stacontrollers for nonlinear singularly perturbed bilizing and systems described by T-S fuzzy models with the consideration of improving the bound of singular perturbation parameter . The main contribution is as follows. For the synthesis problem of simultaneously designing the bound of and stabilizing or controllers, LMI-based methods are presented. For evaluating the upper bound of subject to stability or a prescribed performance bound constraint for the resulting closed-loop system, LMI-based sufcient conditions are developed, respeccontrol synthesis without the tively. For the stabilizing and consideration of improving the bound of , new design methods are also given in terms of solutions to a set of LMIs. The examples have shown the efciency of the proposed methods. APPENDIX A PROOF OF THEOREM 1 Proof: It consists of three parts. In Parts 1 and 2, we prove that (78) and (92) hold, respectively. In Part 3, the proof is completed by using (78) and (92). Part 1: First, from conditions (25) and (26), we have (73) From (27) and (73), it follows (74) then (75) Pre- and postmultiplying (75) by and its transpose and applying the Schur complement, we obtain
and its transpose, then the following inequality holds: (78) Part 2: From (25), we can obtain (79) From (27), we have (80) By (17)(19), (23), and Lemma 2, it follows that
i.e., (81)
Applying the Schur complement to (81), we have
i.e., (82)
33
Notice that
Applying Lemma 2 to the previous inequality, then
which implies that (87) (83) (88) From (85) and (88), we have
where are the same as in (34). Applying (82) to (83), it follows that (84)
(89) By (80) and (84), we have Substituting into (89), then we have
which implies that (85) Pre- and postmultiplying (24) by pose, it follows that and its transwhere are same as in (35). From (87), then
(90) Pre- and postmultiplying (90) by , then we can obtain (91)
. . . Multiplying (86) by follows ,
. . .
(86)
Combining it and (91), that yields
and summing them, it
for . . . Combining it with (79), we can obtain . . . Part 3: Choose Lyapunov function
(92)
Then, by (78) and (92), it follows that . . . . . . . . . for for
and . . .
. Thus, the proof is complete.
34
APPENDIX B PROOF OF THEOREM 5 Proof: From (58), (62), and Part 1 of the proof of Theorem 3, we have for
From (63), it follows that (97) Combining (96), (97), and (93), we have
where . By (62), it follows that (93) On the other hand, from (60), we have where which further implies that for (98)
(94) Substituting , , then (94) becomes (95) where Thus, from (98) and (99), it follows: (100) By (61) and Lemma 2, we can obtain (101) shown at the bottom of the page. Applying the Schur complement to (101), it follows that For brief expression, we also denote On the other hand, from (59), we have (99)
Applying the Schur complement to (95), we obtain where which implies that (96)
(102)
(101)
35
and
By (99), (100), and (102), we have (103) Pre- and postmultiplying (103) by and , then we obtain
i.e.,
Then, by Lemma 1, the proof is completed. REFERENCES

[1] P. V. Kokotovic, J. OReilly, and H. K. Khalil, Singular Perturbation Methods in Control: Analysis and Design. Orlando, FL: Academic Press, 1986. [2] E. Fridman, A descriptor system approach to nonlinear singularly perturbed optimal control problem, Automatica, vol. 37, no. 4, pp. 543549, 2001. [3] E. Fridman, Effects of small delays on stability of singularly perturbed systems, Automatica, vol. 38, no. 5, pp. 897902, 2002. [4] Z. Pan and T. Basar, H -optimal control for singularly perturbed systems. Part I: Perfect state measurements, Automatica, vol. 29, no. 2, pp. 401423, 1993. [5] Z. Pan and T. Basar, H -optimal control for singularly perturbed systems. II. Imperfect state measurements, IEEE Trans. Autom. Control, vol. 39, no. 2, pp. 280299, Feb. 1994. [6] P. Shi and V. Dragan, Asymptotic H control of singularly perturbed systems with parametric uncertainties, IEEE Trans. Autom. Control, vol. 44, no. 9, pp. 17381742, Sep. 1999. [7] Z. Pan and T. Ba, Time-scale separation and robust controller design for uncertain nonlinear singularly perturbed systems under perfect state measurements, Int. J. Robust Nonlinear Control, vol. 6, no. 7, pp. 585608, 1996. [8] H. D. Tuan and S. Hosoe, On linear robust controllers for a class of nonlinear singular perturbed systems, Automatica, vol. 35, no. 4, pp. 735739, 1999. [9] E. Fridman, State-feedback H control of nonlinear singularly perturbed systems, Int. J. Robust Nonlinear Control, vol. 11, no. 12, pp. 11151125, 2001. [10] T. Takagi and M. Sugeno, Fuzzy identication of systems and its applications to modeling and control, IEEE Trans. Syst., Man. Cybern., vol. 15, no. 1, pp. 116132, Jan. 1985. [11] S. G. Cao, N. W. Rees, and G. Feng, Analysis and design for a class of complex control systems part I: Fuzzy modelling and identication, Automatica, vol. 33, no. 6, pp. 10171028, 1997. [12] G. Feng, S. G. Cao, and N. W. Rees, An approach to H control of a class of nonlinear systems, Automatica, vol. 32, no. 10, pp. 14691474, 1996. [13] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality Approach. New York: Wiley, 2001. [14] X.-J. Ma, Z.-Q. Sun, and Y.-Y. He, Analysis and design of fuzzy controller and fuzzy observer, IEEE Trans. Fuzzy Syst., vol. 6, no. 1, pp. 4151, Feb. 1998. [15] K. Tanaka, T. Hori, and H. O. Wang, A multiple Lyapunov function approach to stabilization of fuzzy control systems, IEEE Trans. Fuzzy Syst., vol. 11, no. 4, pp. 582589, Aug. 2003.
[16] H. Liu, F. Sun, and Z. Sun, Stability analysis and synthesis of fuzzy singularly perturbed systems, IEEE Trans. Fuzzy Syst., vol. 13, no. 5, pp. 273284, Oct. 2005. [17] T. H. S. Li and K.-J. Lin, Stabilization of singularly perturbed fuzzy systems, IEEE Trans. Fuzzy Syst., vol. 12, no. 5, pp. 579595, Oct. 2004. [18] W. Assawinchaichote and S. K. Nguang, H fuzzy control design for nonlinear singularly perturbed systems with pole placement constraints: An LMI approach, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 1, pp. 579588, Jan. 2004. [19] P. Shi, H output feedback control design for uncertain fuzzy singularly perturbed systems: An LMI approach, Automatica, vol. 40, no. 12, pp. 21472152, Dec. 2004. [20] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix inequalities in system and control theory. Philadelphia, PA: Society for Industrial and Applied Mathematics (SIAM), 1994. [21] N. R. Sandell, Robust stability of systems with application to singular perturbations, Automatica, vol. 15, no. 4, pp. 467470, 1979. [22] W. Feng, Characterization and computation for the bound in linear time-invariant singularly perturbed systems, Syst. Control Lett., vol. 11, no. 3, pp. 195202, 1988. [23] B.-S. Chen and C.-L. Lin, On the stability bounds of singularly perturbed systems, IEEE Trans. Automat. Control, vol. 35, no. 11, pp. 12651270, Nov. 1990. [24] Z. H. Shao and M. E. Sawan, Robust stability of singularly perturbed systems, Int. J. Control, vol. 58, no. 6, pp. 14691476, 1993. [25] S. Sen and K. B. Datta, Stability bounds of singularity perturbed systems, IEEE Trans. Autom. Control, vol. 38, no. 2, pp. 302304, Feb. 1993. [26] D. Mustafa and T. N. Davidson, Block bialternate sum and associated stability formulae, Automatica, vol. 31, no. 9, pp. 12631274, 1995. [27] L. Cao and H. M. Schwartz, Complementary results on the stability bounds of singularly perturbed systems, IEEE Trans. Autom. Control, vol. 49, no. 11, pp. 20172021, Nov. 2004. [28] Z. H. Shao, Robust stability of two-time-scale systems with nonlinear uncertainties, IEEE Trans. Autom. Control, vol. 49, no. 2, pp. 258261, Feb. 2004. [29] A. Isidori, H control via measurement feedback for afne nonlinear systems, Int. J. Robust Nonlinear Control, vol. 4, pp. 553574, 1994. [30] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control. Upper Saddle River, NJ, USA: Prentice-Hall, 1996. [31] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control Toolbox. Natick, MA: The MathWorks, Inc., 1995. [32] S. H. Z. Maccarley, State-feedback control of non-linear systems, Int. J. Control, vol. 43, pp. 14971514, 1986.
Guang-Hong Yang (SM03) received the B.S. and M.S. degrees in mathematics from Northeast University of Technology, Shenyang, China, in 1983 and 1986, respectively, and the Ph.D. degree in control engineering from Northeastern University, Shenyang, China (formerly, Northeast University of Technology), in 1994. He is currently a Professor with the College of Information Science and Engineering, Northeastern University. From 1986 to 1995, he was a Lecturer/Associate Professor with Northeastern University. In 1996, he was as a Postdoctoral Fellow with the Nanyang Technological University, Singapore. From 2001 to 2005, he was a Research Scientist/Senior Research Scientist with the National University of Singapore, Singapore. His current research interests include fault tolerant control, fault detection and isolation, nonfragile control systems design, and robust control. Dr. Yang is an Associate Editor for the International Journal of Control, Automation, and Systems (IJCAS), and an Associate Editor on the Conference Editorial Board of the IEEE Control Systems Society.
Jiuxiang Dong received the B.S. degree in mathematics and applied mathematics and the M.S. degree in applied mathematics from Liaoning Normal University, Dalian, China, in 2001 and 2004, respectively. He is currently pursing the Ph.D. degree at Northeastern University, Shenyang, China. His research interests include fuzzy control, robust control, and fault-tolerant control.
36
104 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 17, NO. 1, FEBRUARY 2009
37
Multiobjective Control for TS Fuzzy Singularly Perturbed Systems

Chunyu Yang and Qingling Zhang
AbstractThis paper investigates the problem of multiobjective control for a class of TakagiSugeno (TS) fuzzy singularly perturbed systems. Based on a linear matrix inequality (LMI) approach, a state feedback controller that depends on the singular perturbation parameter is developed such that: 1) the H performance of the resulting closed-loop system is less than or equal to some prescribed value; 2) the closed-loop poles of each local system are within a prespecied LMI stability region; and 3) for a given upper bound for the singular perturbation parameter , both 1) and 2) are guaranteed for all (0, ]. It is shown that the dependent controller is well dened for any (0, ], and can be reduced to an -independent one if is sufciently small. Finally, a practical example is given to show the feasibility and effectiveness of the obtained method. Index TermsH performance, linear matrix inequality (LMI), pole placement, singular perturbation bound, Takagi Sugeno (TS) fuzzy singularly perturbed systems.
I. INTRODUCTION N MODERN control systems, such as economic models, motor control systems, convectiondiffusion systems, power systems, and magnetic-ball suspension systems, small parameters are often involved, which can lead to high dimensionality and ill-conditioned numerical issues in the system analysis and controller design. To deal with these problems, the theory of singular perturbation has been developed in the past four decades (see [1][8] and the references therein). Stability bound problem in a singularly perturbed system (SPS) has attracted much attention (see, e.g., [9][14]). This problem is usually described as characterizing and computing an upper bound 0 for the singular perturbation parameter , such that the stability of the SPS is ensured for all (0, 0 ) or (0, 0 ]. The stability bound problem of linear SPSs has been solved perfectly, and some frequency- and time-domain methods have been derived to provide the largest upper bound for (see, e.g., [4], [5], and [13]). For nonlinear SPSs, the stability bound problem is a complex and challenging topic. Some methods to estimate the upper bound for have been obtained in [8], [10][12], and [14]. Stabilization bound problem has also been considered (see [15][18]). The aim is to design a controller to enlarge the stability bound. The authors in [15][17] have
Manuscript received December 19, 2007; revised May 30, 2008; First published September 9, 2008; current version published February 4, 2009. This work was supported by the Natural Science Foundation of China under Grant 60574011. The authors are with the Institute of Systems Science, Northeastern University, Shenyang 110004, China, and also with the Key Laboratory of Integrated Automation of Process Industry, Northeastern University, Ministry of Education, Shenyang 110004, China (e-mail: chunyuyang@yahoo.cn; qlzhang@mail.neu.edu.cn). Digital Object Identier 10.1109/TFUZZ.2008.2005404
discussed the stabilization bound problem for linear SPSs and presented some valuable results. Recently, the robust stabilization problem of a class of nonlinear SPSs has been considered in [18], where a state feedback controller is designed by solving two independent Lyapunov equations, and the stability bound of the closed-loop system is then computed by using the existing methods given in [14]. In this two-step design procedure, stability bound is not regarded as one of design objects. To the best of our knowledge, stability bound synthesis of nonlinear SPSs remains as an open area. Further, it is also a signicant problem to provide a bound for , such that the stability and other performances of the SPS are guaranteed. In the past several decades, nonlinear control approaches based on the TakagiSugeno (TS) model have been extensively studied and successfully applied [19][21]. The main reasons are as follows: 1) TS model has been shown to be a universal approximator for a wide class of nonlinear systems [22], [23] and 2) the design is usually formulated in the linear matrix inequality (LMI) framework that can be efciently solved by the existing tools [24]. Recently, many researchers have focused on the analysis and design of TS fuzzy SPSs. Stability analysis and stabilization problems for both continuous- and discrete-time TS fuzzy SPSs were investigated in [25], and some LMI-based approaches were derived. H control for TS fuzzy SPSs was considered in [26] and [27], and LMI-based design methods were obtained. To get a satisfactory transient behavior, H control for TS fuzzy SPSs with pole placement constraints was considered in [28] and [29]. The obtained design methods were expressed in terms of -independent LMIs. Using the results in [25][29], the stability and H performance of the resulting closed-loop systems are only ensured for sufciently small singular perturbation parameter . Since no reliable upper bound for can be determined, the effectiveness of the obtained controllers has to be validated by trial and error. Such a problem has been recognized by several researchers. Based on the separation of states into slow and fast ones, a composite fuzzy H controller was constructed in [30] and [31], and the upper bound for the singular perturbation parameter is then determined. In these results, the controller design procedures are not concerned with the bound of the singular perturbation parameter . Moreover, since these approaches involve the separation of states into slow and fast ones, they cannot be applied to nonstandard SPSs. H control for TS fuzzy SPSs with the consideration of enlarging the bound of was rst studied in [32], where stabilizing and H controllers with the consideration of maximizing the bound of were designed. However, in the design procedures, there are ve weighting constants required to be chosen subjectively, which will result in conservatism and inconvenience in
1063-6706/$25.00 2009 IEEE
YANG AND ZHANG: MULTIOBJECTIVE CONTROL FOR TS FUZZY SINGULARLY PERTURBED SYSTEMS 105
38
applications. In practice, the singular perturbation parameter is usually very small, and a rough bound for the parameter is often known. Thus, it is signicant to regard a predened bound for as one of design objects. This paper considers the problem of multiobjective control for TS fuzzy SPSs. The problem consists of H control, pole placement, and singular perturbation bound design. Specically, given an H performance bound > 0, an LMI stability region D, and an upper bound for the singular perturbation parameter , this paper will construct an -dependent state feedback controller, such that (0, ], the L2 -gain of the mapping from the exogenous input noise to the regulated output is less than or equal to , and the poles of each subsystem are all within the LMI stability region D. Two subproblems of the multiobjective control are discussed, and the main problem is then solved. An -dependent state feedback controller is designed by solving a set of -independent LMIs. It is shown that the controller is well dened (0, ]. If is sufciently small, the controller can be reduced to an -independent one. At last, an inverted pendulum controlled by a dc motor via a gear train is used to illustrate the obtained approach. The rest of this paper is organized as follows. In Section II, the problems under consideration are dened. The main results are given in Section III. The multiobjective control problem is reduced to the feasibility of a set of -independent LMIs. In Section IV, an illustrative example is given to show the effectiveness and advantage of the obtained method. Section V concludes this paper and proposes a new topic for our future work. II. PROBLEM FORMULATION Consider a TS fuzzy SPS, in which the ith rule is formulated as follows: Plant rule i : IF v1 (t) is Mi1 , v2 (t) is Mi2 , . . . , v (t) is Mi THEN E()x(t) = Ai x(t) + B1i w(t) + B2i u(t) z(t) = Ci x(t) + Di u(t) for i = 1, 2, . . . , r (1)
It is assumed in this paper that

r
wi (v(t)) 0,
i=1
wi (v(t)) > 0,
i = 1, 2, . . . , r
t 0.
Let i (v(t)) = Then

r
wi (v(t)) , r i=1 wi (v(t))
i = 1, 2, . . . , r.
i (v(t)) 0,
i=1
i (v(t)) = 1,
i = 1, 2, . . . , r
t 0.
(2) For the convenience of notations, we denote i = i (v(t)), i = 1, 2, . . . , r. Then, the TS fuzzy model (1) is inferred as follows:
r
E()x(t) =
i=1 r
i [Ai x(t) + B1i w(t) + B2i u(t)]
z(t) =
i=1
i [Ci x(t) + Di u(t)].
(3)
Throughout the paper, it is assumed that the singular perturbation parameter is available for feedback. Based on the concept of parallel distributed compensation (PDC), the state feedback fuzzy controller is described by Controller rule i : IF v1 (t) is Mi1 , v2 (t) is Mi2 , . . . , v (t) is Mi THEN u(t) = Ki ()x(t) for i = 1, 2, . . . , r. (4)
Because the controller rules are same as the plant rules, the state feedback controller is given as follows:
r
u(t) =
i=1
i Ki ()x(t).
(5)
where E() = diag{I, I}, > 0 is the singular perturbation parameter, Mij (i = 1, 2, . . . , r, j = 1, 2, . . . , ) are fuzzy sets, r is the number of fuzzy rules, v(t) = [ v1 (t) v1 (t) v (t) ]T is the premise vector that may depend on states in many cases, is the number of premise variables, x(t) n is the state, u(t) s is the input, w(t) p is the disturbance that belongs to L2 [0, ), z(t) s is the controlled output, and Ai , B1i , B2i , Ci , Di are constant matrices with appropriate dimensions. Denote
Remark 2.1: In many SPSs, the singular perturbation parameter can be measured. In these cases, is available for feedback, which has attracted much attention. For example, -dependent controllers were designed for TS fuzzy SPSs in [27] and [29]. Since is usually very small, an -dependent controller may be ill-conditioning as tends to zero. Thus, it is a key task to ensure the obtained controller to be well dened. This problem will be discussed later. Substituting (5) into (3) yields the closed-loop system
r r
E()x(t) =
i=1 j =1 r r
i j [(Ai + B2i Kj ())x(t) + B1i w(t)]
z(t) = Mik (vk (t)), i = 1, 2, . . . , r

i=1 j =1
i j [(Ci + Di Kj ())x(t)].
(6)
wi (v(t)) =
k =1
where Mik (vk (t) is the grade of membership of vk (t) in Mik .
The following denitions that can be found in [28] will be used in this paper.
39
Denition 2.1: Given > 0, a system of the form (3) is said to be with an H -norm less than or equal to if
Tf 0
satisfying the following LMIs Z1 > 0 Z1 + Z3 Z5 Z1 + Z3 Z5

T Z5 T Z5
(10) (11) (12) i = 1, 2, . . . , r i = 1, 2, . . . , r 1i<jr 1i<jr k = 1, 2, . . . , r k = 1, 2, . . . , r

T Z5 Z4
z T (t)z(t) dt 2
0
Tf
wT (t)w(t) dt
(7)
Z2
>0 >0
holds for x(0) = 0, where Tf is the terminal time of control and x(0) denotes the initial condition of system (3). Denition 2.2: A subset D of the complex plane C is called an LMI region if there exist a symmetric matrix L Rdd and a matrix M Rdd such that D = {z = x + jy C : fD (z) < 0} where the characteristic function fD (z) is given as follows: fD (z) = L + M z + M T z. (9) (8)
Z2 + 2 Z4 ii 1 ij 1 i 2 j i 1 j ) 2 k 4
ii < 0, 1 + + + < 0, < 0, < 0,
(13) (14) (15) (16) (17) (18)
ij 1
j i 1
(i 2 k 3
k < 0, 3 + < 0,
In this paper, multiobjective control for TS fuzzy SPSs is formulated as follows: Problem 1: Given an H performance bound , an LMI stability region D, and an upper bound for the singular perturbation parameter , design a state feedback controller of the form (5), such that for all (0, ], the closed-loop system (6) is with an H -norm less than or equal to , and the closed-loop poles of each local system are all within the LMI stability region D. Remark 2.2: Synthesis problems for TS fuzzy SPSs have been investigated by many researchers (see [26][32]). Among these papers, only H control was addressed in [26], [27], [30], and [31]. H control with pole placement constraints was investigated in [28] and [29]. H control and singular perturbation bound design were simultaneously considered in [32]. This paper will consider a more general case. Problem 1 consists of H control, pole placement, and singular perturbation bound design. To solve Problem 1, the following subproblems will be studied. Problem 2: Given an H performance bound and an upper bound for the singular perturbation parameter , design a state feedback controller of the form (5), such that for all (0, ], the closed-loop system (6) is asymptotically stable and with an H -norm less than or equal to . Problem 3: Given an LMI stability region D and an upper bound for the singular perturbation parameter , design a state feedback controller of the form (5), such that for all (0, ], the closed-loop poles of each local system are all within the LMI stability region D. III. MAIN RESULTS In this section, the main results will be derived by using the basic lemmas given in the Appendix. For solving Problem 1, Problems 2 and 3 are considered. The following theorem provides a solution to Problem 2. Theorem 3.1: Given an H performance bound and an upper bound for the singular perturbation parameter , if there exT ist matrices Zk (k = 1, 2, . . . , 5) with Zk = Zk (k = 1, 2, 3, 4), T Fi (i = 1, 2, . . . , r) and Yij with Yij = Yj i (i, j = 1, 2, . . . , r)
where U1 = Z1 Z5 0 Z2 U2 = Z3 0
T T ij = U1 AT + FjT B2i + Ai U1 + B2i Fj i 1
1 T B1i B1j Yij , 2 .. .
i, j = 1, 2, . . . , r i = 1, 2, . . . , r Y1r . . . Yr r Cr U1 + Dr Fk
T i = U2 AT + Ai U2 , 2 i Y11 . . . k = 3 Yr 1 C1 U1 + D1 Fk
T T T T U1 C1 + Fk D1 . . . , T T T T U1 Cr + Fk Dr I
k = 1, 2, . . . , r
k 4 =
0 . . . 0 C1 U2
.. .
0 . . . 0 Cr U2
T T U2 C1 . . . , T T U2 Cr 0
k = 1, 2, . . . , r
then, for any (0, ], the closed-loop system (6) with Ki () = Fi (U1 + U2 )1 (i = 1, 2, . . . , r) is asymptotically stable and with an H -norm less than or equal to . Proof: Suppose that LMIs (10)(18) are feasible. From LMIs (13)(18), it follows that ii + i < 0 (0, ], 1 2 ij 1 j i 1 (i 2 j ) 2 < 0 (0, ], i = 1, 2, . . . , r (19) + + + 1i<jr (20)
40
and k + k < 0 3 4 (0, ], k = 1, 2, . . . , r. (21)
Substituting Z() = U1 + U2 into inequalities (19)(21) shows that ii () a ij () a and Y11 . . . Yr 1 C1 Z() + D1 Fk .. . Y1r . . . Yr r Cr Z() + Dr Fk Z
T
From (2) and (27), it follows that X11 () . r . . k 0> Xr 1 () k =1 C1 + D1 Kk ()
.. .
X1r () . . . Xr r () Cr + Dr Kk ()
<0 <0
(0, ], (0, ],
i = 1, 2, . . . , r 1i<jr
(22) (23)
j i () a
T T T C1 + Kk ()D1 . . . T T T Cr + Kk ()Dr I
X11 () . . . Xr 1 ()
r k =1
.. .
. . . <0 T T T T Z ()Cr + Fk Dr I (0, ], k = 1, 2, . . . , r, (24)
T ()C1
T Fk
T D1
k (C1 + D1 Kk ()) X1r () . . . Xr r ()
r k =1
k (Cr + Dr Kk ())
T T T k (C1 + Kk ()D1 ) . . . . r T T T k =1 k (Cr + Kk ()Dr ) r k =1
T where ij () = Z T ()AT + FjT B2i + Ai Z() + B2i Fj + a i T i, j = 1, 2, . . . , r. 1/ 2 (B1i B1j Yij ), Using Lemma 1.3, LMIs (10)(12) imply
E()Z() = Z T ()E() > 0
(0, ]
I Applying the Shur complement to the previous inequality, one gets r (C T + K T ()DT )
k =1 k 1 k 1
which indicates that Z() is nonsingular for any (0, ]. Pre- and postmultiplying (22) and (23) by Z T () and its transpose, respectively, yield ii () < 0 b and 1i<jr (26) where ij = (Ai + B2i Kj ())T P () + P T ()(Ai + B2i Kj b T ()) + 1/ 2 (P T ()B1i B1j P () Xij ()), Kj () = Fj Z 1 1 (), P () = Z (), and Xij () = Z T ()Yij Z 1 (). Pre- and postmultiplying (24) by diag{Z T (), . . . , T Z (), I} and its transpose, respectively, yield + <0 X11 () . . . Xr 1 () C1 + D1 Kk () .. . X1r () . . . Xr r () Cr + Dr Kk () <0 T T T Cr + Kk ()Dr I . . . (27)
T C1 T + Kk T ()D1
(0, ],
i = 1, 2, . . . , r
(25)
. . .
r k =1 T T T k (Cr + Kk ()Dr )
ij () b
j i () b
(0, ],
r k =1
T T T k (C1 + Kk ()D1 ) T . . .
T T T k (Cr + Kk ()Dr ) X11 () X1r () . . < 0. .. . . + . . .
r k =1
(28)
Xr 1 () Then, for all x 1 x T . 0 . . r x

n
Xr r ()
, it holds that .. . X1r () 1 x . . . . . . r x Xr r ()
(0, ],
k = 1, 2, . . . , r.
X () 11 . . . Xr 1 ()
41
1 x T . + . . r x
r r k =1
r k =1
T T T k (C1 + Kk ()D1 ) . . . T T T k (Cr + Kk ()Dr )
=
i=1
2 xT (t)[(Ai + B2i Ki )T P () i + P T ()(Ai + B2i Ki ) + +

i=1 i< j
r k =1
T T T k (C1 + Kk ()D1 ) T . . . T k (Cr
r k =1 r
T Kk
T ()Dr
1 x . . . r x (29)
1 T T P ()B1i B1i P ()]x(t) 2

r r
i j xT (t)
=
i=1 j =1
i j xT Xij ()x + z T z
[(Ai + B2i Kj + Aj + B2j Ki )T P () + P T ()(Ai + B2i Kj + Aj + B2j Ki ) + 1 T T P ()B1i B1j P () 2

r r
which indicates
r r
i j xT Xij ()x z T z.
i=1 j =1
(30) + +
i=1
Furthermore, it can be seen that

r r
1 T T P ()B1j B1i P ()]x(t) 2 i=1
i j
j =1
i j xT Xij ()x < 0

i=1 j =1
x = 0.
(31)
1 T T x (t)P T ()B1i B1j P ()x(t) 2

r T i [wT (t)B1i P ()x(t)
Using Lemma 1.3, LMIs (10)(12) imply E()Z() = Z T ()E() > 0 which shows that E()P () = P T ()E() > 0 Choose a Lyapunov function as V (x(t)) = xT (t)E()P ()x(t). Then V (x(t)) = xT (t)E()P ()x(t) + xT (t)P T ()E()x(t)
r r
(0, ] (0, ].
+ xT (t)P T ()B1i w(t)].
(33)
From inequalities (25), (26), (30), and (33), it follows that

r
V (x(t)) (32)
i=1
2 xT (t)Xii ()x(t) i
r r
+
i=1 i< j r r
i j xT (t)(Xij () + Xj i ())x(t) 1 T T P ()B1i B1j 2
=
i=1 j =1
i j [(Ai + B2i Kj )x(t) + B1i w(t)]

r r
i=1 j =1
i j xT (t)
r
P ()x(t) + xT (t)P T ()
i=1 j =1
i j
P ()x(t) +
i=1
T i [wT (t)B1i P ()x(t)
[(Ai + B2i Kj )x(t) + B1i w(t)]

r r r r
+ xT (t)P T ()B1i w(t)] =

i=1 j =1
=
i=1 j =1
i j x (t) [(Ai + B2i Kj )T P () + P T ()(Ai + B2i Kj ) +

r T B1j P ()]x(t) i=1 j =1 T B1i B1j P ()x(t) r r
i j xT (t)Xij ()x(t) + 2 wT (t)w(t) 1 T i B1i P ()x(t) i=1 1 T i B1i P ()x(t) i=1

r r T
w 1 T P ()B1i 2 i j xT (t) 1 T P () 2
i=1 j =1
w
r r
i j xT (t)Xij ()x(t) + 2 wT (t)w(t) (34)
+
i=1 T
T i [wT (t)B1i P ()x(t) T
z T (t)z(t) + 2 wT (t)w(t) which shows V (x(t)) z T (t)z(t) + 2 wT (t)w(t).
+ x (t)P ()B1i w(t)]
(35)
42
Integrating both sides of (35) from 0 to Tf , with x(0) = 0, yields

Tf 0
Q11 . . . Qr 1 where V1 = U1 = Z1 0 Z1 Z5 0 0 0 Z2
z (t)z(t) dt
T
2 0
Tf
.. .
Q1r . . <0 . Qr r Z3 Z5 Z3 0
T Z5 Z2 T Z5 Z4
(42)
w (t)w(t) dt.
Inequalities (31) and (34) indicate that V (x(t)) < 0 x = 0 when w(t) 0. This implies that the closed-loop system with w(t) 0 is asymptotically stable. This completes the proof. Remark 3.1: LMIs (10) and (11) indicate that Z1 > 0 and Z2 > 0. As a result, the matrix U1 = [ Z 1 Z0 ] is nonZ5 2 singular. In addition, the proof of Theorem 3.1 has shown that Z() = U1 + U2 is nonsingular for all (0, ]. Then, Ki () = Fi (U1 + U2 )1 is always well dened for all 1 (0, ] and lim0 + Ki () = Fi U1 . Thus, if is sufciently small, controller (5) can be reduced to an -independent one. This remark is also suitable for Theorems 3.2 and 3.3. Remark 3.2: H control for TS fuzzy SPSs with the consideration of singular perturbation bound has been studied in [30][32]. The design approaches given by [30] and [31] involve the separation of states into slow and fast ones, which limits the applications of these methods to standard SPSs. An H controller with the consideration of maximizing the bound for the singular perturbation parameter was designed in [32] by an LMI-based algorithm. To perform this algorithm, one has ve weighting constants to choose subjectively, which may results in conservatism and inconvenience in practice. Though this paper has not proved that the design method described by Theorem 3.1 is less conservative than that given by [32], the example in Section IV will show the advantage of Theorem 3.1. This advantage may result from: 1) the information on the singular perturbation parameter used in the fuzzy controller (5) and 2) Lyapunov function (32) reduced to the one used by [32] if we set Z3 = 0 and Z4 = 0. Thus, the Lyapunov function, based on which Theorem 3.1 is derived, is more general than the one used by [32]. Remark 3.3: The proof of Theorem 3.1 is partially motivated by that of [21, Th. 2]. Now, a controller of the form of (5) will be constructed such that, for any admissible singular perturbation parameter , the resulting closed-loop system (6) with w(t) 0 is D-stable. Theorem 3.2: Given an LMI stability region D and an upper bound for the singular perturbation parameter , if there exT ist matrices Zk (k = 1, 2, . . . , 5) with Zk = Zk (k = 1, 2, 3, 4), T Qij with Qij = Qj i (i, j = 1, 2, . . . , r), and Fi (i = 1, 2, . . . , r) satisfying LMIs (10)(12) and
ii S1 ii S1
V2 = U2 =
V3 =
0 0
0 Z4
ij S1 = L V1 + M (Ai U1 ) + M T (Ai U1 )T
+ M (B2i Fj ) + M T (B2i Fj )T Qij

i S2 = L V2 + M (Ai U2 ) + M T (Ai U2 )T
S3 = L V 3 and denotes the Kronecker product of the matrices; then for any (0, ], the poles of each subsystem of system (6) with Ki () = Fi (U1 + U2 )1 (i = 1, 2, . . . , r) are all within the given LMI region D. Proof: Suppose LMIs (10)(12) and (36)(42) are feasible. ij i Let Z() = U1 + U2 and ij = S1 + S2 + 2 S3 , i, j = 1, 2, . . . , r; then E()Z() = V1 + V2 + 2 V3 and ij = L V1 + M (Ai U1 ) + M T (Ai U1 )T + M (B2i Fj ) + M T (B2i Fj )T + (L V2 + M (Ai U2 ) + M T (Ai U2 )T ) + 2 L V3 Qij = L (V1 + V2 + 2 V3 ) + M [Ai (U1 + U2 ) + B2i Fj ] + M T [Ai (U1 + U2 ) + B2i Fj ]T Qij = L [E()Z()] + M [(Ai + B2i Kj ())Z()] + M T [(Ai + B2i Kj ())Z()]T Qij . Using Lemma 1.2, LMIs (36)(38) imply that ii < 0 (0, ], (0, ], i = 1, 2, , r. 1 i < j r. (44) (43)
Similarly, using Lemma 1.2, LMIs (39)(41) show that ij + j i < 0 Denote FD = L X() + M
r r
(45)
< 0, < 0,
i = 1, 2, . . . , r i = 1, 2, . . . , r i = 1, 2, . . . , r 1i<jr 1i<jr 1i<jr
(36) (37) (38) (39) (40) (41)
i S2
ii i S1 + S2 + 2 S3 < 0, ij j S1 + S1 i < 0, ij j j i S1 + S1 i + (S2 + S2 ) < 0, ij j j i S1 +S1 i +(S2 +S2 )+22 S3 < 0,
i=1 j =1 r r
i j E 1 ()(Ai + B2i Kj ())X() + M T i j E 1 () (Ai + B2i Kj ())X()

i=1 j =1 T
(46)
where X() is a matrix with appropriate dimensions.
43
By computation, it holds that [I E()] FD [I E()]

r r
=
i=1 j =1
i j {L [E()X()E()] + M [(Ai + B2i Kj ())X()E()] + M T [E()X()(Ai + B2i Kj ())T ]}. (47)
Let X() = Z()E 1 (). Using Lemma 1.3, LMIs (10)(12) imply that E()Z() = Z T ()E() > 0 which shows X() = X T () > 0 Substituting X() = Z()E
r r 1
(0, ] (48)
(0, ]. () into (47) gives
[I E()] FD [I E()] =
i=1 j =1
i j {L [E()Z()] + M [(Ai + B2i Kj ())Z()] + M T [Z T ()(Ai + B2i Kj ())T ]}

r
=
i=1
2 {L [E()Z()] i + M [(Ai + B2i Ki ())Z()] + M T [Z T ()(Ai + B2i Ki ())T ]}

r r
+
i=1 i< j
i j {2L [E()Z()] + M [(Ai + B2i Kj ())Z()] + M T [Z T ()(Ai + B2i Kj ())T ] + M [(Aj + B2j Ki ())Z()] + M T [Z T ()(Aj + B2j Ki ())T }. (49)
Inequalities (44), (45), and (49) imply that [I E()] FD [I E()]

r r
From (48), (51), and Lemma 1.1, it follows that the closedloop system (6) with w(t) 0 is D-stable, i.e., the poles of each subsystem of system (6) with Ki () = Fi (U1 + U2 )1 (i = 1, 2, . . . , r) are all within the given LMI region D. This completes the proof. Theorems 3.1 and 3.2 provide solutions to Problems 2 and 3, respectively. By Theorems 3.1 and 3.2, Problem 1 can be solved by the following theorem. Theorem 3.3: Given an H performance bound , an LMI stability region D, and an upper bound for the singular perturbation parameter , if there exist matrices Zk (k = T 1, 2, . . . , 5) with Zk = Zk (k = 1, 2, 3, 4), Yij with Yij = T Yj i (i, j = 1, 2, . . . , r), Qij with Qij = QTi (i, j = 1, 2, . . . , r), j and Fi (i = 1, 2, . . . , r) satisfying LMIs (10)(18) and (36) (42), then for any (0, ], the closed-loop system (6) with Ki = Fi (U1 + U2 )1 (i = 1, 2, . . . , r) is with an H -norm less than or equal to , and the poles of each subsystem of system (6) are within the given LMI region D. Remark 3.4: H control for TS fuzzy SPSs with pole placement constraints via state and output feedback controllers has been studied in [28] and [29], respectively. Using these methods, the obtained controllers are only valid for the case that is sufciently small and the upper bounds for have to be determined by trial and error. While in Theorem 3.3, the prescribed upper bound is one of design objects, which is very signicant in practice. Remark 3.5: Besides the approaches based on the TS fuzzy model [25][32], there have been various design methods for nonlinear SPSs in particular forms. Stabilization controller design methods were proposed in [33][36] by using sliding mode control, feedback linearization, integral control, and gain scheduling, respectively. However, these methods can only be applied to standard nonlinear SPSs since they are based on decomposing the original system into two reduced-order subsystems in different time scales. H control for nonlinear SPSs was investigated in [37][39]. In particular, a class of SPSs being nonlinear only on the slow variables was examined in [37] and [38]. A local state feedback H control problem for an afne nonlinear SPS was addressed in [39]. Each of the aforementioned designs presents a formula to compute the bound for the perturbation parameter or shows the existence of the upper bound. In this paper, a given upper bound for is one of the design objects, which represents an advantage of this paper over literature [33][39]. Furthermore, the obtained method can be applied to a wide class of nonlinear SPSs including nonstandard ones, and the LMI-based design procedure is easy to perform. IV. EXAMPLE
<
i=1 j =1
i j Qij T .. . 1 Q1r . . . . . . . Qr r r This section considers an inverted pendulum controlled by a dc motor via a gear train whose physical model can be found in [28] and [40]. The system can be described by the following state equations [28], [40]: x1 (t) = x2 (t) + 0.1w(t) x2 (t) = g N Km sin x1 (t) + x3 (t) l ml2
Q11 1 . . = . . . . r Qr 1
(50)
Inequalities (42) and (50) indicate that [I E()] FD [I E()] < 0 (0, ], which shows that FD < 0 (0, ]. (51)
44
TABLE I UPPER BOUNDS FOR
La x3 (t) = Kb N x2 (t) Ra x3 (t) + u(t) + w(t) z(t) = 0.1x1 (t) + 0.1u(t) (52) where x1 (t) = p (t), x2 (t) = p (t), x3 (t) = Ia (t), u(t) is the control input, w(t) is the disturbance input, km is the motor torque constant, Kb is the back emf constant, and N is the gear ratio. The parameters for the plant are given as g = 9.8 m/s2 , l = 1 m, m = 1 kg, N = 10, Km = 0.1 N m/A, Kb = 0.1 V s/rad, Ra = 1 , and La = mH. Note that the inductance La represents the small parameter in the system. Substituting the parameters into (52) gives x1 (t) = x2 (t) + 0.1w(t) x2 (t) = 9.8 sin x1 (t) + x3 (t) x3 (t) = x2 (t) x3 (t) + u(t) + w(t) z(t) = 0.1x1 (t) + 0.1u(t). (53)
Fig. 1.
LMI conic sector region of pole location.
As in [32], the membership functions of the fuzzy sets are chosen as follows: |x1 (t)| |x1 (t)| M1 (x1 (t)) = 1 M2 (x1 (t)) = . Then, the following TS fuzzy model can exactly represent the dynamics of nonlinear SPS (53) under x1 (t) Plant rule 1 : IF x1 (t) is M1 (x1 (t)), THEN E()x(t) = A1 x(t) + B11 w(t) + B21 u(t) z(t) = C1 x(t) + D1 u(t) Plant rule 2 : IF x1 (t) is M2 (x1 (t)), THEN E()x(t) = A2 x(t) + B12 w(t) + B22 u(t) z(t) = C2 x(t) + D2 u(t) where 1 0 0 0 1 0 E() = 0 1 0 A1 = 9.8 0 1 0 0 0 1 1 0 1 0 0.1 1 B11 = 0 A2 = 0 0 0 1 1 1 0.1 0 0 B21 = 0 B22 = 0 B12 = 0 1 1 1 C1 = [ 0.1 D1 = 0.1 0 0] C2 = [ 0.1 0 0]
Fig. 3. Simulation for
t 0
Fig. 2.
Simulation for
t 0
z T (s)z(s) ds/
t 0
w T (s)w(s) ds with =
0.02, x(0) = 0, and w(t) = sin(10t).
(54)
z T (s)z(s) ds/
t 0
w T (s)w(s) ds with =
D2 = 0.1.
0.2, x(0) = 0, and w(t) = sin(10t).
45
Fig. 4. Pole map of each local system with = 0.02. (a) i = 1 and j = 1: poles at 2.4789, 64.7421 + 10.4377 i, and 64.7421 10.4377 i. (b) i = 2 and j = 2: poles at 100.5430, 2.7396, and 43.3492. (c) i = 1 and j = 2: poles at 100.4619, 2.3880, and 43.7819. (d) i = 2 and j = 1: poles at 2.7997, 64.5817 + 9.9423 i, and 64.5817 9.9423 i.
The fuzzy controller is described as follows: Plant rule 1 : IF x1 (t) is M1 (x1 (t)), THEN u(t) = K1 ()x(t) Plant rule 2 : IF x1 (t) is M2 (x1 (t)), THEN u(t) = K2 ()x(t) (55)
that the upper bounds obtained by Theorem 3.1 are bigger than those given by [32]. B. Case II: D Is a Conic Sector Region In this section, the closed-loop poles of each local system will be placed within an LMI conic sector region with = 45 (see Fig. 1). Note that the LMI conic sector region can be described by (8) with L=0 M= sin() cos() . cos() sin()
where K1 () and K2 () are controller gains to be determined. The following two cases are considered to illustrate the design method developed in this paper. Case I is used to show the advantage of the obtained method over the one given by [32]. Case II is utilized to show the applicability of the proposed controller design method. A. Case I: D is the Open Left-Half Plane In this case, Problem 1 is equivalent to Problem 2. By Theorem 3.1, the upper bounds for subject to guaranteeing various H performance bounds are shown in Table I. The problem of H control for system (54) with the consideration of singular perturbation bound design was studied in [32] and the upper bounds for were given (see Table I). It can be seen
Solving the LMIs in Theorem 3.3 with = 0.3 and = 0.5 gives 1.8335 4.7635 0 0 U1 = 4.7635 12.8429 0.1042 36.7990 65.7377 U2 = 3.4179 0 0.9586 3.4179 8.2240 0 9.0002 31.0419 0.1042 2.6154 40.9112 58.4247 . 36.7990
F1 = 3.7414 F2 = 8.1129
46
Fig. 5. Pole map of each local system with = 0.2. (a) i = 1 and j = 1: poles at 69.7388, 1.6183, and 7.3209. (b) i = 2 and j = 2: poles at 75.2666, 4.3815 + 1.6563 i, and 4.3815 1.6563 i. (c) i = 1 and j = 2: poles at 75.2495, 1.5142, and 7.2658. (d) i = 2 and j = 1: poles at 69.7593, 4.4594 + 1.7353 i, and 4.4594 1.7353 i.
Then, the fuzzy controller is given by

2
u(t) =
i=1
i (t)Ki ()x(t)
(56)
emerges when the singular perturbation parameter is subject to uncertainties. We will consider this problem in our future work. In addition, Lemmas 1.2 and 1.3 are important in the derivation of our main results. We expect that they are also useful in other research areas of singularly perturbed systems. APPENDIX PRELIMINARY LEMMAS Lemma 1.1: Given a dynamic system x(t) = Ax(t) and an LMI region D, the system is D-stable, i.e., (A) D if there exists a matrix X Rn n with X = X T > 0 such that L X + M (AX) + M T (AX)T < 0 where (A) is the set of eigenvalues of A and denotes the Kronecker product of the matrices. Lemma 1.2: For a positive scalar and symmetric matrices S1 , S2 , and S3 with appropriate dimensions, inequality S1 + S2 + 2 S3 > 0 holds for all (0, ], if S1 0 S1 + S2 > 0 S1 + S2 + 2 S3 > 0. (58) (59) (60) (57)
where 1 (t) = M1 (x1 (t)), 2 (t) = M2 (x1 (t)), K1 () = F1 (U1 + U2 )1 , and K2 () = F2 (U1 + U2 )1 . Apply controller (56) to system (54) and assume x(0) = 0 and w(t) = sin(10t). The simulation results corresponding to = 0.02 and = 0.2 are shown in Figs. 25. It is easy to see that, in either case, the H -norm is less than the prescribed value 0.5, and the closed-loop poles of each local system are within the LMI conic sector region with = 45 . V. CONCLUSION AND REMARKS This paper has investigated multiobjective control for a class of TS fuzzy singularly perturbed systems. An LMI-based approach to design an -dependent state feedback controller has been proposed. By virtue of this method, the obtained controller can ensure that, for any singular perturbation parameter in the given bound, the prescribed H performance bound is satised, and the closed-loop poles of each subsystem are all within the predened LMI stability region. It should be noted that the approach presented in this paper requires that the perturbation parameter is available for feedback. Consequently, a signicant and challenging problem
47
Proof: Given 1 [0, 1), multiplying inequalities (59) and (60) by 1 and 1 1 , respectively, yield 1 S1 + 1 S2 0 and (1 1 )S1 + (1 1 )S2 + (1 1 )2 S3 > 0. Adding (61) to (62) shows that S1 + S2 + (1 1 )2 S3 > 0 holds for any 1 [0, 1). Similarly, inequalities (58) and (63) imply that S1 + (1 2 )S2 + (1 2 )(1 1 )2 S3 > 0 (64) (63) (62) (61)
REFERENCES
[1] E. H. Abed, A new parameter estimate in singular perturbations, Syst. Control Lett., vol. 6, no. 3, pp. 193198, 1985. [2] H. K. Khalil, Stability analysis of nonlinear multiparameter singularly perturbed systems, IEEE Trans. Autom. Control, vol. 32, no. 3, pp. 260 263, Mar. 1987. [3] B. S. Chen and C. L. Lin, On the stability bounds of singularly perturbed systems, IEEE Trans. Autom. Control, vol. 35, no. 11, pp. 12651270, Nov. 1990. [4] W. Feng, Characterization and computation for the bound in linear time-invariant singularly perturbed systems, Syst. Control Lett., vol. 11, no. 3, pp. 195202, 1988. [5] S. Sen and K. B. Datta, Stability bounds of singularly perturbed systems, IEEE Trans. Autom. Control, vol. 38, no. 2, pp. 302304, Feb. 1993. [6] L. Saydy, New stability/performance results for singularly perturbed systems, Automatica, vol. 32, no. 6, pp. 807818, 1996. [7] S. J. Chen and J. L. Lin, Maximal stability bounds of singularly perturbed systems, J. Franklin Inst., vol. 336, no. 8, pp. 12091218, 1999. [8] H. D. Tuan and S. Hosoe, Multivariable circle criteria for multiparameter singularly perturbed systems, IEEE Trans. Autom. Control, vol. 45, no. 4, pp. 720725, Apr. 2000. [9] N. R. Sandell Jr., Robust stability of systems with application to singular perturbations, Automatica, vol. 15, no. 4, pp. 467470, 1979. [10] H. K. Khalil, Asymptotic stability of nonlinear multiparameter singularly perturbed systems, Automatica, vol. 17, no. 6, pp. 797804, 1981. [11] A. Saberi and H. Khalil, Quadratic-type lyapunov functions for singularly perturbed systems, IEEE Trans. Autom. Control, vol. 29, no. 6, pp. 542 550, Jun. 1984. [12] K. Khorasani and M. A. Pai, Asymptotic stability of nonlinear singularly perturbed systems using higher order corrections, Automatica, vol. 21, no. 6, pp. 717727, 1985. [13] L. Cao and H. M. Schwartz, Complementary results on the stability bounds of singularly perturbed systems, IEEE Trans. Autom. Control, vol. 49, no. 11, pp. 20172021, Nov. 2004. [14] Z. H. Shao, Robust stability of two-time-scale systems with nonlinear uncertainties, IEEE Trans. Autom. Control, vol. 49, no. 2, pp. 258261, Feb. 2004. [15] T. H. S. Li and J. H. Li, Stabilization bound of discrete two-time-scale systems, Syst. Control Lett., vol. 18, no. 6, pp. 479489, 1992. [16] W. Q. Liu, M. Paskota, and V. Sreeram, K. L. Teo, Improvement on stability bounds for singularly perturbed systems via state feedback, Int. J. Syst. Sci., vol. 28, no. 6, pp. 571578, 1997. [17] J. S. Chiou, F. C. Kung, and T. H. S. Li, An innite -bound stabilization design for a class of singularly perturbed systems, IEEE Trans. Circuits Syst. I, Fundam Theory Appl., vol. 46, no. 12, pp. 15071510, Dec. 1999. [18] Z. H. Shao and M. E. Sawan, Stabilisation of uncertain singularly perturbed systems, Proc. Inst. Electr. Eng. Control Theory Appl., vol. 153, no. 1, pp. 99103, 2006. [19] T. Takagi and M. Sugeno, Fuzzy identication of systems and its applications to modeling and control, IEEE Trans. Syst., Man, Cybern., vol. 15, no. 1, pp. 116132, Jan./Feb. 1985. [20] B. S. Chen, C. S. Tseng, and H. J. Uang, Mixed H 2 /H fuzzy output feedback control design for nonlinear dynamic systems: An LMI approach, IEEE Trans. Fuzzy Syst., vol. 8, no. 3, pp. 249265, Jun. 2000. [21] X. Liu and Q. Zhang, Approaches to quadratic stability conditions and H control designs for TS fuzzy systems, IEEE Trans. Fuzzy Syst., vol. 11, no. 6, pp. 830839, Dec. 2003. [22] S. G. Cao, N. W. Rees, and G. Feng, Analysis and design for a class of complex control systems part I: Fuzzy modelling and identication, Automatica, vol. 33, no. 6, pp. 10171028, 1997. [23] S. G. Cao, N. W. Rees, and G. Feng, Analysis and design for a class of complex control systems part II: Fuzzy controller design, Automatica, vol. 33, no. 6, pp. 10291039, 1997. [24] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory. Philadelphia, PA: SIAM, 1994. [25] H. Liu, F. Sun, and Z. Sun, Stability analysis and synthesis of fuzzy singularly perturbed systems, IEEE Trans. Fuzzy Syst., vol. 13, no. 2, pp. 273284, Apr. 2005. [26] H. Liu, F. Sun, and Y. Hu, H control for fuzzy singularly perturbed systems, Fuzzy Sets Syst., vol. 155, no. 2, pp. 272291, 2005. [27] W. Assawinchaichote, S. K. Nguang, and P. Shi, H output feedback control design for uncertain fuzzy singularly perturbed systems: An LMI approach, Automatica, vol. 40, no. 12, pp. 21472152, 2004.
holds for any 2 [0, 1). For any , (0, ], there exist 1 , 2 [0, 1), such that = (1 1 ), = (1 2 ). Then, it follows from (64) that S1 + S2 + S3 > 0 (65) holds for any , (0, ]. Thus, inequality (57) holds for any (0, ]. Lemma 1.3: If there exist matrices Zi (i = 1, 2, . . . , 5) with Zi = ZiT (i = 1, 2, 3, 4) satisfying the LMIs (10)(12), then E()Z() = Z T ()E() > 0 where Z() = Proof: Denote S1 = then E()Z() = Z1 + Z3 Z5
T Z5
(0, ]
T Z5
(66)
Z1 + Z3 Z5 Z3 Z5
Z2 + Z4
T Z5
(67)
Z1 0
0 0 S2 =
Z2
S3 =
0 0
0 Z4
Z2 + 2 Z4
= S1 + S2 + 2 S3 = Z T ()E(). From inequalities (10)(12), it follows that S1 0 S1 + S2 > 0 S1 + S2 + 2 S3 > 0. Then, using Lemma 1.2, one gets S1 + S2 + 2 S3 > 0, (0, ]. (72) (69) (70) (71) (68)
As a result, (68) and (72) implies (66). This completes the proof. ACKNOWLEDGMENT The authors would like to thank the Editor and the anonymous reviewers for their valuable comments.
48
[28] W. Assawinchaichote and S. K. Nguang, H fuzzy control design for nonlinear singularly perturbed systems with pole placement constraints: An LMI approach, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 1, pp. 579588, Feb. 2004. [29] W. Assawinchaichote and S. K. Nguang, Fuzzy H output feedback control design for singularly perturbed systems with pole placement constraints: An LMI approach, IEEE Trans. Fuzzy Syst., vol. 14, no. 3, pp. 361371, Jun. 2006. [30] T. H. S. Li and K. J. Lin, Stabilization of singularly perturbed fuzzy systems, IEEE Trans. Fuzzy Syst., vol. 12, no. 5, pp. 579595, Oct. 2004. [31] T. H. S. Li and K. J. Lin, Composite fuzzy control of nonlinear singularly perturbed systems, IEEE Trans. Fuzzy Syst., vol. 15, no. 2, pp. 176187, Apr. 2007. [32] G. H. Yang and J. X. Dong, Control synthesis of singularly perturbed fuzzy systems, IEEE Trans. Fuzzy Syst., vol. 16, no. 3, pp. 615629, Jun. 2008. [33] J. Alvarez-Gallegos and G. Silva-Navarro, Two-time scale sliding-mode control for a class of nonlinear systems, Int. J. Robust Nonlinear Control, vol. 7, no. 9, pp. 865879, 1997. [34] H. L. Choi, Y. S. Shin, and J. T. Lim, Control of nonlinear singularly perturbed systems using feedback linearisation, Proc. Inst. Electr. Eng. Control Theory Appl., vol. 152, no. 1, pp. 9194, 2005. [35] J. Wang, J. Wang, and H. Li, Nonlinear PI control of a class of nonlinear singularly perturbed systems, Proc. Inst. Electr. Eng. Control Theory Appl., vol. 152, no. 5, pp. 560566, 2005. [36] J. W. Son and J. T. Lim, Robust stability of nonlinear singularly perturbed system with uncertainties, Proc. Inst. Electr. Eng. Control Theory Appl., vol. 153, no. 1, pp. 104110, 2006. [37] Z. Pan and T. Basar, Time-scale separation and robust controller design for uncertain nonlinear singularly perturbed systems under perfect state measurements, Int. J. Robust Nonlinear Control, vol. 6, no. 7, pp. 585 608, 1996. [38] H. D. Tuan and S. Hosoe, On linear robust H controllers for a class of nonlinear singular perturbed systems, Automatica, vol. 35, no. 4, pp. 735739, 1999. [39] E. Fridman, State-feedback H control of nonlinear singularly perturbed systems, Int. J. Robust Nonlinear Control, vol. 11, no. 12, pp. 11151125, 2001. [40] S. H. Zak and C. A. Maccarley, State-feedback control of non-linear systems, Int. J. Control, vol. 43, no. 5, pp. 14971514, 1986.
Chunyu Yang received the B.Sc. degree in 2002 from the Department of Mathematics, Northeastern University, Shenyang, China, where he is currently working toward the Ph.D. degree. His current research interests include descriptor systems, robust control, and fuzzy control.
Qingling Zhang received the B.Sc. and M. Sc. degrees from the Department of Mathematics, Northeastern University, Shenyang, China, in 1982 and 1986, respectively, and the Ph.D. degree from the Department of Automatic Control, Northeastern University, in 1995. During 1997, he was a Postdoctoral Fellow in the Department of Automatic Control, Northwestern Polytechnical University, Xian, China. Since 1997, he has been a Professor at Northeastern University. He is also a member of the University Teaching Advisory Committee of the National Ministry of Education. He visited Hong Kong University, Sydney University, Western Australia University, Niigata University, Pohan University of Science and Technology, Seoul University, Alberta University, Lakehead University, and Wisor University as a Research Associate, Research Fellow, Senior Research Fellow, and Visiting Professor, respectively. He has authored or coauthored six books and more than 230 papers about control theory and applications. Prof. Zhang received 14 prizes from central and local governments for his research. He has also received the Golden Scholarship from Australia in 2000.
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008 1371
49
H Filtering for Fuzzy Singularly Perturbed Systems

Guang-Hong Yang, Senior Member, IEEE, and Jiuxiang Dong
AbstractThis paper considers the problem of designing H lters for fuzzy singularly perturbed systems with the consideration of improving the bound of singular-perturbation parameter . First, a linear-matrix-inequality (LMI)-based approach is presented for simultaneously designing the bound of the singularly perturbed parameter , and H lters for a fuzzy singularly perturbed system. When the bound of singularly perturbed parameter is not under consideration, the result reduces to an LMI-based design method for H ltering of fuzzy singularly perturbed systems. Furthermore, a method is given for evaluating the upper bound of singularly perturbed parameter subject to the constraint that the considered system is to be with a prescribed H performance bound, and the upper bound can be obtained by solving a generalized eigenvalue problem. Finally, numerical examples are given to illustrate the effectiveness of the proposed methods. Index TermsFuzzy ltering, H ltering, linear matrix inequalities (LMIs), singularly perturbed systems, TakagiSugeno (T-S) fuzzy models.
I. I NTRODUCTION N THE PAST 30 years, singularly perturbed systems have been intensively studied, where singular perturbation with a small parameter, for example , is exploited to determine the degree of separation between slow and fast modes of the system. For the purpose of avoiding the difculties linked with the stiffness of the equations involved in the design, the singularperturbation design techniques have been developed [1]. For the problem of lter design for linear singularly perturbed systems, many signicant advances have been achieved (see [2][4] and the reference therein). In particular, the decomposition solution of H lter gain for linear singularly perturbed systems is given in [3], and a method of designing reducedorder H optimal lters for systems with slow and fast modes is proposed in [4]. However, the problem of designing H lters for nonlinear singularly perturbed systems still remains as open research subject.
Manuscript received June 8, 2007; revised January 31, 2008 and April 26, 2008. This work was supported in part by the Program for New Century Excellent Talents in University under Grant NCET-04-0283, by the Funds for Creative Research Groups of China under Grant 60521003, by the Program for Changjiang Scholars and Innovative Research Team in University under Grant IRT0421, by the State Key Program of National Natural Science of China under Grant 60534010, by the Funds of National Science of China under Grant 60674021, by the 111 Project under Grant B08015, and by the Funds of Ph.D. program of the Ministry of Education, China under Grant 20060145019. This paper was recommended by Associate Editor W. J. Wang. The authors are with the Key Laboratory of Integrated Automation of Process Industry (Ministry of Education), and College of Information Science and Engineering, Northeastern University, Shenyang 110004, China (e-mail: yangguanghong@ise.neu.edu.cn; yang_guanghong@163.com; dongjiuxiang@ ise.neu.edu.cn; dong_jiuxiang@163.com). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TSMCB.2008.927277
An important approach to address the synthesis problems for nonlinear systems is to model the considered system as Takagi and Sugeno (T-S) fuzzy systems [5], which are locally linear time-invariant systems connected by IFTHEN rules. As a result, the conventional linear-system theory can be applied to analysis and synthesis of the class of nonlinear control systems. In recent years, the T-S fuzzy control systems have been studied extensively [6][10]. For the nonlinear-ltering problem via T-S fuzzy models, some numerically effective design methods have been obtained [11][14]. In [11], an H fuzzy-ltering design for nonlinear discrete-time systems is proposed, and Chen et al. [12] present a method of designing mixed H2 /H lters for equalization/detection of nonlinear communication systems using fuzzy interpolation and linear-matrix-inequality (LMI) techniques [15]. The H fuzzy-ltering problem with D stability constraints is studied in [13], and Feng [14] gives design methods for robust H fuzzy ltering of nonlinear systems. For nonlinear singularly perturbed systems, the earlier methods will lead to the ill-conditioning resulting from the interaction of slow and fast dynamic modes. For the fuzzy control and ltering problems of nonlinear singularly perturbed systems, some progresses have been achieved [16][18], [20], [21]. In [16] and [17], design methods for the stabilization and H control of nonlinear singularly perturbed systems via state feedback are given in terms of solutions of LMIs, respectively. LMI-based methods for designing robust controllers and determining allowable perturbation bounds are given in [18]. In [19], a method for determining allowable perturbation bounds is presented via some algebra inequalities. A solution to the nonlinear H fuzzy-ltering problem with pole-placement constraints is presented in [20]; a method based on Riccatti equation for multiparameter singularly perturbed systems is presented in [21]. The aforementioned results are based on LMIs or Riccatti equations; the advantage of the existing methods is that some performance indexes can be optimized. However, the upper bound of the singular-perturbation parameter is not addressed at the stage of designing lters or controllers. In this paper, a lter will be designed with the consideration to improve the upper bound of singular-perturbation parameter . It is well known that the accurate knowledge of the stability bound of a singularly perturbed system (i.e., the system is stable for [0, ]) is very important for applications. The characterization and computation of the stability bound have attracted considerable efforts for over two decades in the past ([22][24] and the reference therein). In general, there are two classes of methods to characterize and compute the stability bounds: One is based on frequency-domain transfer functions, and another is based on state-space models. Both methods can provide the exact bounds as shown in [22][24]. However,
1083-4419/$25.00 2008 IEEE
1372 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008
50
the issue of how to improve the bound by controller or lter design has not been addressed in the literature, which undoubtedly is very important for the applications of singularly perturbed system theory. Furthermore, for nonlinear systems, no result has been presented on how to evaluate the upper bound of singularly perturbed parameter subject to the constraint that the considered system is to be with a prescribed performance bound. This paper is concerned with the problem of designing H lters for fuzzy singularly perturbed systems with the consideration of improving the bound of singular-perturbation parameter . First, an LMI-based approach is presented for simultaneously designing the bound of and H lter for a fuzzy singularly perturbed system. When the bound of singularly perturbed parameter is not under consideration, the result reduces to a new LMI-based design method for H ltering of fuzzy singularly perturbed systems, which can provide less or at least the same conservative results than existing method [20]. Furthermore, a method is given for evaluating the upper bound of singularly perturbed parameter subject to the considered system with a designed lter to be with a prescribed performance bound for (0, ], where the upper bound can be obtained by solving a generalized eigenvalue problem (GEVP) [15]. Although a method of nding an interval (0 ] so that the system is stable and satises an H performance index for (0 ] is given in [19], the H performance index in [19] is dependent on Lyapunov matrices and different from the general denition for H performance index. In this paper, a general H performance index (independent of Lyapunov matrices) is considered. Moreover, a convex method of nding an interval [ , ] is proposed in [18]; however, the search for small is related to -dependent computation, which cannot avoid the difculties linked with the stiffness of the equations. Therefore, a convex method of nding an interval (0, ] in this paper is more effective. This paper is organized as follows. In Section II, the system description and some preliminaries are presented. Section III considers the H lter-design problem with the consideration of improving the bound of singular-perturbation parameter , and the estimation of the upper bound of singular-perturbation parameter with satisfying H performance requirement. Section IV gives examples to illustrate the effectiveness of the new proposed methods. Finally, Section V concludes this paper. II. P ROBLEM S TATEMENT , N OTATION , AND P RELIMINARIES In this section, we rst give the notations used in this paper.
and E T denotes (E 1 )T if E is nonsingular. The symbol within a matrix represents the symmetric entries H
11 T H12
[Hij ]rr =: . . . T H1r B. System Description
H12 H22 . . .
T H2r
.. .
H1r H2r . . . . Hrr
The nonlinear singularly perturbed system under consideration is described by the following fuzzy-system model: Plant Rule i : IF v1 (t) is Mi1 and v2 (t) is Mi2 , . . . , vp (t) is Mip THEN E x(t) = Ai x(t) + B1i w(t) z(t) = C1i x(t) y(t) = C2i x(t) + D21i w(t) where E = In1 n1 0 0 In2 n2 (1)
x(t) Rn , n = n1 + n2 is the state vector, z(t) Rnz is the controlled output vector, y(t) Rny is the measurable output vector, and w(t) Rnw is the exterior disturbance vector. r is the number of IFTHEN rules, v(t) Rp1 are the premise variables, > 0 is the singular-perturbation parameter, and Mij are the fuzzy sets. By using the fuzzy-inference method with a singleton fuzzier, product inference, and center average defuzziers, the nal T-S fuzzy model is obtained as
r
wi (v(t)) (Ai x(t) + B1i w(t)) E x(t) = i=1

r r
wi (v(t))
i=1
wi (v(t)) (C1i x(t)) z(t) = i=1

r r
wi (v(t))
i=1
wi (v(t)) (C2i x(t) + D21i w(t)) y(t) =

i=1 r
wi (v(t))
i=1
where
p
wi (v(t)) = A. Notation We use the following standard notations. For a matrix M , M is dened as the largest singular value of M . For a square matrix E, He(E) is dened as He(E) = E + E T
j=1
Mij (vj (t)) .
Mij (vj (t)) is the grade of membership of vj (t) in Mij , where it is assumed that
r
wi (v(t)) > 0,
i=1
wi (v(t)) 0; i = 1, 2, . . . , r.
YANG AND DONG: H FILTERING FOR FUZZY SINGULARLY PERTURBED SYSTEMS 1373
51
Denote i (t) = wi (v(t))

r
which can be rewritten as the follows: E (t) = AF ()(t) + BF ()y(t) zF (t) = CF ()(t) where
r r
wi (v(t))
i=1
(5)
then 0 i (v(t)) 1
r
AF () =
i=1 j=1 r
i (t)j (t)AF ij
i (v(t)) = 1.
i=1
BF () =
i=1 r
i (t)BF i
i (v(t)) (which is also denoted as i ), 1 i r are said to be normalized membership functions. Therefore, the singularly perturbed fuzzy system (1) can be rewritten as follows:
r r
CF () =
i=1
i (t)CF i .
Let i B1i w(t) ze (t) = z(t) zF (t) xe (t) = x(t) . (t)
E x(t) =
i=1 r
i Ai x(t) +
i=1
z(t) =
i=1 r
i C1i x(t)
r
Augmenting the system (2) with the states of the lter (5), then the following ltering-error systems can be obtained: i D21i w(t)
i=1
y(t) =
i=1
i C2i x(t) +
xe (t) = Ae ()xe (t) + Be ()w(t) ze (t) = Ce ()xe (t) where (6)
i.e., E x(t) = A()x(t) + B1 ()w(t) z(t) = C1 ()x(t) y(t) = C2 ()x(t) + D21 ()w(t) where
r
Ae () = (2) Be () =
E 1 A() E 1 BF ()C2 () E 1 B1 () E BF ()D21 ()

1
0 E 1 AF ()
Ce () = [ C1 () A() =
i=1 r
CF () ] .
i Ai
B1 () =
i=1 r
i B1i
For the lter-error system (6), let > 0 be a constant. If (6) is asymptotically stable, and for any w(t) L2 [0, ) (the space of square-integrable functions) and x(0) = 0, the following inequality holds:
T ze (t)ze (t)dt 0
C1 () =
i=1 r
i C1i
2 0
wT (t)w(t)dt
C2 () =
i=1 r
i C2i
D21 () =
i=1
i D21i .
(3)
In this paper, we use the following lter, which is from [20], for the H ltering for fuzzy singularly perturbed systems:
r r r
E (t) =
i=1 j=1 r
i j AF ij (t) +
i=1
i BF i y(t)
zF (t) =
i=1
i CF i (t)
(4)
then the system (6) is said to be with an H -norm or L2 -gain less than or equal to . In this paper, the upper bound of the singularly perturbed parameter is the main topic to be addressed. A method will be given for designing a lter with simultaneously improving the upper bound of the singularly perturbed parameter and satisfying H performance requirement. Furthermore, an LMI-based method is proposed to estimate the upper bound of singular-perturbation parameter of the singularly perturbed fuzzy system (6) satisfying H performance-bound requirement. Remark 1: It should be pointed out that the existing lterdesign method [20] is only considering the H performance requirement, which might result in a very small stability bound of so that the designed lter may not be applicable in practice.
52
C. Preliminary Lemmas The following lemma, essentially from [26], gives a sufcient condition for the system (6) to be with an H -norm less than or equal to . Lemma 1: For a given scalar > 0, if there exists matrices P = P T , satisfying the following matrix inequality: Me1 () = P Ae () + AT ()P e
T + P Be ()Be ()P +
then 1 I T . . . r I where 1 T C ()Ce () < 0 2 e (7)

l J l = Jij i Jii = 0,
1 I . l J l . 0 . l=1 r I
r
(14)
rr
for 1 l r.
then the H -norm of the system (6) from exogenous input to the lter error is less than or equal to . Lemma 2: For a given scalar > 0, if there exist matrices S = S T and N = N T , AF (), BF (), and CF (), satisfying the following matrix inequalities: N <0 S > 0
T Me3 () = Me3a ()+Me3a () 1
Proof: 1 I T . . . 1 I . l J l . . l=1 r I r I Jl 1 I T 11 r . . . = l . . . l T l=1 r I J1r

r r r 2 i i i j Jij + Jij i=1 j=i+1 r r j j 2 i i j Jjj + Jij + Jij i=1 j=i+1 r r r j l i i j l He Jij + Jil + Jjl . i=1 j=i+1 l=j+1 T T j + Jii
(8)
.. .
l J1r
l Jrr
1 I . . . r I
S E B1 () + (S N )E 1 B1 ()+N E 1 BF ()D21 () S E 1 B1 () 1 (S N )E B1 ()+N E 1 BF ()D21 () + 1 [ C1 ()CF () 2 C1 () ]T (9)

T
[ C1 ()CF ()
C1 () ] < 0
where Me3a () is given in (10), shown at the bottom of the page. Then, the system (6) is asymptotically stable, and the H -norm of the system (6) from exogenous input to the lter error is less than or equal to . Proof: See Appendix. Remark 2: Lemma 2 presents a sufcient condition for the system (6) to be with an H performance bound, and the conditions can be converted into LMIs by using appropriate variable transformations. However, the resulting design conditions are -dependent and may become ill-conditioned when is sufciently small, which always is the case for singularly perturbed systems. j Lemma 3: If there exist symmetric matrices Jii , 1 i = l j r and matrices Jij , 1 i < j r, 1 l r, such that the following LMIs hold:
j i i Jii + Jij + Jij j j i Jjj + Jij + Jij T T
From (11)(13), it follows (14). The following result from [20] gives a sufcient condition for H ltering of fuzzy singularly perturbed systems via the lter (4). Lemma 4: For a given > 0 [20], if there exist matrices X, Y , AF 0ij , BF 0i , CF 0i , 1 i, and j r, with X= X11 X21 0 X22 Y = Y11 Y21 0 Y22
where X11 , X22 , Y11 , and Y22 are symmetric matrices, satisfying the following LMIs: X11 0 I 0 ii < 0, 0 X22 0 I I 0 Y11 0 0 I >0 0 Y22 1ir 1i<jr
> 0, > 0, > 0,
1i<jr 1i<jr 1i<j<lr
(11) (12) (13)
(15)
(16a) (16b)
He
l Jij
j Jil
i Jjl
ij + ji < 0,
Me3a () =
S E 1 A() (S N )E 1 A()+N E 1 AF ()+N E 1 BF ()C2 ()
S E 1 A() (S N )E 1 A()+N E 1 BF ()C2 ()
(10)
53
where ij is solved as shown at the bottom of the page. Then, there exists an > 0 such that the H -norm of the system (6) with AF ij = N T AF 0ij Y T Ai X BF 0j C2i X (M T )1 BF i = N T BF 0i CF i = CF 0i (M )
T 1
3 I S21 4 I I 6 I N21
T S21 I
<0 <0 <0 <0 5 I I I N22 <0
(21)
I N11
T N21 I
(22) (23)
(17)
is less than or equal to for (0, ], where M = N11 0 M11 M12 and N = satisfy N M T = I Y X. 0 M22 N21 N22 III. M AIN R ESULTS A. Filter Design With Improving the Upper Bound of The following theorem is concerned with the problem of designing a lter with the consideration of improving the upper bound of singular-perturbation parameter and satisfying H performance requirement. Theorem 1: If there exist matrices S, Sa , N , Na , , AF ij , F i , CF i , 1 i, j r, J l , 1 l r, 1 i < j r, J l , and B ij ii 1 i = l r and scalar variable i , i , and 1 i 6, with S= Sa = S11 S21 0 0 0 S22
T S21 0
0 S T B1i 2 6 I (S T N T )B1i + BF j D21i < 0 I 2 F j D21i )T (B 2 I <0 BF j D21i I 2 T 3 I 0 Sa B1i 2 T T 0 3 I Sa Na B1i < 0 T T B1i Sa B1i (Sa Na ) I 2 T T T 5 I 0 AF ij + C2i BF j 2 T T <0 0 5 I C2i BF j AF ij + BF j C2i BF j C2i I
2 6 I l [ij ]rr + Jij
I I 4 I
(24)
(25)
(26)
(27) (28) (29)
N=
N11 N21
0 N22
Me5aii + where
rr T Me5aii < 1 I
< 0,
l = 1, . . . , r
T 0 N21 Na = 0 0
l where S11 , S22 , N11 , N22 , , Jii , and 1 i = l r are symmetric matrices, satisfying the following LMIs: j i i Jii + Jij + Jij i Jjj T
ij = Hij + 0 0 Me5aii =
0 0 0 0 0 0
T Sa Ai (Sa Na )T Ai
> 0, > 0,
1i<jr 1i<jr 1i<j<lr 2 I I I S22 <0
(18) (19) (20)
T Sa Ai (Sa Na )T Ai
j Jij
j Jij
S T Ai S T Ai Me4aij = F j C2i +(S N)TAi (S N)TAi + BF j C2i AF ij + B where Hij is given in (30), shown at the bottom of the page, and denote = max{1 2 3 , 4 5 6 , 4 7 } (31)
j l i He Jij + Jil + Jjl > 0,
1 I I
I S11
<0
Ai X + X T Ai AF 0ij + AT i ij = T B1i C1i X CF 0j
T T T T Ai Y + Y Ai + BF 0j C2i + C2i BF 0j
T T T B1i Y + D21i BF 0j
I 0
2 I
, 1ijr
C1i
Hij =
He(Me4aij +Me4aji ) 2
S T (B1i +B1j ) 2 (S T N T )(B1i +B1j )+BF j D21i +BF i D21j 2
T T T T C1i CF i +C1j CF j 2 T T C1i +C1j 2
(30)
0 2 I
54
where
T T 7 = 1 +2 Na (N 1 )T 5 +26 3 + Na (N 1 )T 2 T + 3 + Na (N 1 )T 2 2
bound 1/ obtained by solving the earlier optimization problem may be conservative. After obtaining the lter gains, a lessconservative bound of can be obtained by Theorem 2 in the next section. B. Estimation of the Upper Bound of In this section, assume that the lter has been designed. The following theorem gives a technique to estimate the upper bound of singularly perturbed parameter subject to the ltering-error system (6) with an H performance bound. Theorem 2: Consider the system (6), and let > 0 be a given l constant. If there exist a scalar and matrices S, Sa , N , Na , Jii , l 1 i = l r, Jij , 1 i < j r, and 1 l r with S= Sa = S11 S21 0 0 0 S22
T S21 0
(32)
Then, for (0, (1/ )], the H -norm of the system (6) with AF ij = N T AF ij BF i = N T BF i , 1ijr (33)
is less than or equal to . Proof: See Appendix. Remark 3: Theorem 1 presents sufcient conditions under which an upper bound (1/) of singularly perturbed parameter and lter gains can be obtained. From (31) and i > 0, i > 0, and 1 i 6, we have found that can be minimized by solving the following optimization problem:
6 6
N= Na =
N11 N21
0 N22
minimize
i=1
ai i +
i=1,i=4
i i + 4 N T N T b b a (34)
T 0 N21 0 0
subject to (18)(29)
l where S11 , S22 , N11 , N22 , Jii , and 1 i = l r are symmetric matrices, satisfying the following LMIs:
where ai , i , and 1 i 6 are positive weighting constants to b be chosen. However, the optimization problem cannot directly T be solved due to the term 4 Na N T . Consider b
T Na N T =
Y1 < S11 Y1 S21

T S21 S22
(35) >0 (36) (37) >0

T
0 N21
0 0
1 N11 0
1 T 1 N11 N21 N22 1 N22
Y2 < N11 Y2 N21

T N21 N22
= =
0 1 N21 N11
1 T 1 N21 N11 N21 N22
(38) 1i<jr 1i<jr 1i<j<lr 1lr

rr
0 0 0 0 + 1 1 T 1 N21 N11 0 0 N21 N11 N21 N22

1 N21 N11
j i i Jii + Jij + Jij j j i Jjj + Jij + Jij
> 0, > 0,
(39) (40) (41) (42) 1lr (43)
1 T 1 N21 N11 N21 N22 2 1 N11 1 N22 .
N21 From (22), we have

1 N11 4
1 N11 + N21
j l i He Jij + Jil + Jjl > 0, l [Eij ]rr + Jij
rr
< 0,
1 N22 5
N21
6 .
l [Eaij ]rr < [Eij ]rr + Jij
Then, it follows that

T Na N T
6 4 + 4 5 6 .
where
11 Eij Eij =
Combining it with i > 0, i > 0, and 1 i 6, (34) can be indirectly minimized by solving the optimization problem
6
minimize
i=1
(ai i + bi i )
subject to (18)(29)
12 Eij 22 Eij
13 Eij 23 Eij 33 Eij
ij 44 Eij
14 Eij 24 Eij E 34
(44)
where ai , bi , and 1 i 6 are positive weighting constants to be chosen. Since the constraints (18)(29) are of LMIs, the optimization problem can be effectively solved via LMI Control Toolbox [25]. It should be pointed out that the upper
1 11 Eij = He(S T Ai +S T Aj ) 2 1 12 T T Eij = He S T Ai +AT ij N +C2i BF j N +AT (S N ) F i 2

T T + S T Aj +AT ji N +C2j BF i N +AT (S N ) F j
55
1 13 Eij = S T (B1i +B1j ) 2

14 Eij =
For (0, 1/], from (46), it follows that

1
1 T T T C T CF i +C1j CF j 2 1i
S11 S21
T S21 S22
> 0,
for
0,
1 .
1 22 Eij = He (S N )T Ai + N T BF j C2i + (S N )T Aj 2 + N BF i C2j

23 Eij = T
Multiplying the earlier inequality by , then, yields S = S11 S21

T S21 S22
> 0,
for
0,
1 .
(47)
1 (S T N T )(B1i +B1j ) 2 + N T BF j D21i +N T BF i D21j
Similarly, from condition (37) and (38), we can obtain N = N11 N21
T N21 N22
< 0,
for
0,
1 .
(48)
24 Eij =
1 T C T +C1j 2 1i
33 Eij = I
On the other hand, by (39)(43) and applying Lemma 3, it follows that:

r r
34 44 Eij = 0 Eij = 2 I 11 12 13 Eaij Eaij Eaij 22 23 Eaij Eaij Eaij = 33 Eaij 11 Eaij = 12 Eaij =
14 Eaij 24 Eaij 34 Eaij 44 Eaij
i=1 j=1
i j Eij < 0 (45)

r r r r
i j Eaij <
i=1 j=1 i=1 j=1
i j Eij
1 T T S Ai +Sa Aj 2 a 1 T T T S Ai +AT ij Na +C2i BF j Na +AT (Sa Na ) F i 2 a

T T T Sa Aj +AT ji Na +C2j BF i Na +AT (Sa Na ) F j
which implies that

r r
i j ( Eaij + Eij ) < 0,

i=1 j=1
0,
1 .
(49)
1 T 13 Eaij = Sa (B1i +B1j ) 2

14 Eaij = 0 22 Eaij =
From (44) and (45), then (49) can be rewritten as (50), shown at the bottom of the page, where Me3a () is same as in (10). Applying Schur complement to (50), then it follows that: Me3 () < 0 (51)
1 T He (Sa Na )T Ai +Na BF j C2i +(Sa Na )T Aj 2

T + Na BF i C2j
23 Eaij =
1 2
T T T Sa Na (B1i +B1j )+Na BF j D21i T + Na BF i D21j
where Me3 () is same as in (9). By (47), (48), and (51) and applying Lemma 2, the conclusion follows. Thus, the proof is complete. Remark 4: By Theorem 2, an upper bound of can be obtained by solving the following optimization problem: Minimize subject to (35)(43). It is a GEVP [15], which can be effectively solved using LMI Control Toolbox [25]. The problem of computing the bound of was considered in [18], where a method of nding an interval [ , ], > 0 so that the system is stable for [ , ] was derived. However, the search for small is related to -dependent computation, which cannot avoid the difculties linked with the stiffness of the equations. Moreover, a method of nding an interval (0 ] so that the system is stable and satises an H performance index
24 33 Eaij = 0 Eaij = 0
34 44 Eaij = 0 Eaij = 0.
Then the H -norm of the system (6) is less than or equal to for (0, ] with = 1/. Proof: From condition (35) and (36), we have S11 S21
T S21 S22
> 0.
(46)
T Me3a ()+Me3a ()
S T N T
S T E 1 B1 () 1 E B1 ()+N E 1 BF ()D21 () I
T T C1 ()CF () T C1 () < 0, 0 2 I
for
0,
(50)
56
for (0 ] is given in [19]. However, the H performance index in [19] is dependent on Lyapunov matrices and is different from the general denition for H performance index, which is independent of Lyapunov matrices and is exploited in Theorem 2. Some comparisons among the earlier methods are shown in Example 1. C. Comparison With the Existing Results In this section, rst, based on Theorem 1, a lter-design condition without the consideration of improving the upper bound of singular-perturbation parameter will be given as a corollary of Theorem 1. Then, a comparison between the corollary and the existing result in [20] will be presented in Theorem 3. Corollary 1: 1) For a given scalar > 0, if there exist matrices S, N , l l AF ij , BF i , CF i , 1 i, j r, Jii , 1 i = l r, Jij , 1 i < j r, and 1 l r with S= S11 S21 0 S22 N= N11 N21 0 N22
Proof: 1) Assume that the conditions of 1) in Corollary 1 hold, then from (52), we have that there exist i , 1 i 6 such that 1 I < S11 1 1 I < N11 4 1 I < S22 2
T S21 S21 < 3 I T N21 N21 < 6 I.
1 I < N22 5
Applying Schur complement to the earlier inequality yields (21) and (22). Moreover, it is easily obtained that there exist i , 1 i = 4 6 such that (24)(27) and (29) hold. On the other hand, if (54) holds, then there exists a very small scalar > 0 such that
l [Hij ]rr + Jij rr
+ rdiag[Ia , Ia , . . . , Ia ]rr < 0, l = 1, . . . , r (58)
where
I2n2n Ia = 0 0
0 Inw nw 0
0 0 Inz nz
l where S11 , S22 , N11 , N22 , Jii , and 1 i = l r are symmetric matrices, satisfying the following LMIs:
and we can obtain the following inequality: rdiag[I0 , I0 , . . . , I0 ]rr < rdiag[Ia , Ia , . . . , Ia ]rr where (59)
S11 > 0 S22 > 0,

j i i Jii + Jij + Jij j j i Jjj + Jij + Jij T
N11 < 0 > 0, > 0,
N22 < 0
(52) (53)
1i<jr 1i<jr
j l i He Jij + Jil + Jjl > 0, l [Hij ]rr + Jij rr
1i<j<lr 1lr (54)
I2n2n I0 = 0 0
0 0nw nw 0
0 0 0nz nz
< 0,
Moreover, we can also obtain the following inequality: [Rbij ]rr < rdiag[I0 , I0 , . . . , I0 ]rr (60)
where Hij is the same as in (30); then, there exists an > 0 such that the H -norm of the system (6) with (33) is less than or equal to for (0, ]. 2) For a given scalar > 0, if there exist matrices S, N , AF ij , BF i , CF i , 1 i, and j r with S= S11 S21 0 S22 N= N11 N21 0 N22
where Rbij = I0 . From (58)(60), it follows that

l [Hij ]rr + Jij rr
+ [Rbij ]rr < 0,
l = 1, . . . , r
i.e.,
l [Hij + Rbij ]rr + Jij rr
< 0,
l = 1, . . . , r.
(61)
where S11 , S22 , N11 , and N22 are symmetric matrices, satisfying the following LMIs: S11 > 0, Hii < 0, S22 > 0, N11 < 0, 1i<jr N22 < 0 (55) (56) (57)
1ir
Hij + Hji < 0,
where Hij is the same as in (30); then, there exists an > 0 such that the H -norm of the system (6) with (33) is less than or equal to for (0, ]. Furthermore, if the conditions of 2) hold, then the ones of 1) hold.
0 0 Let = I2n2n . Then Rbij = 0 0 0 . 0 0 0 Let ij = Hij + Rbij , then (61) implies that (28) holds. Because = I2n2n , then there exists 4 > 0 such that (23) holds. Therefore, if the conditions of 1) in Corollary 1, then the conditions of Theorem 1 hold. Furthermore, from Theorem 1, it follows that there exists an > 0 such that the H -norm of the system (6) with (33) is less than or equal to for (0, ].
57
2) Assume that the conditions of 2) in Corollary 1 hold, then l l there exist Jii = 0, 1 i = l r and Jij = Hij 1 i < j r, 1 l r such that
j Jii i i + Jij + Jij T T
Applying the Schur complement to (15) yields Y11 0

1 X11 0 Y22 0
0 1 X22
> 0.
> 0, > 0,
1i<jr 1i<jr 1i<j<lr l = 1, . . . , r
Combining it and (64), then it follows that

1 N11 = X11 Y11 < 0 1 N22 = X22 Y22 < 0.
j j i Jjj + Jij + Jij
j l i He Jij + Jil + Jjl > 0, l [Hij ]rr + [Jij ]rr < 0,
Pre- and postmultiplying (16) by diag[S T I I I] = diag[X T I I I] and its transpose, then we can obtain ii < 0, ij + ji < 0, 1ir 1i<jr
which implies that (53) and (54) hold. Then, the conditions of 1) in Corollary 1 hold, then the conclusion follows. Corollary 1 provides two lter-design methods without the consideration of improving the upper bound of . By the proof of Corollary 1, it can be seen that 1) in Corollary 1 is less conservative than 2). Moreover, the equivalence between the existing result given in [20] (i.e., Lemma 4) and 2) in Corollary 1 will be given in the following theorem, which shows that the new lter-design method given by 1) in Corollary 1 is less conservative than the existing one [20]. Theorem 3: The conditions of Lemma 4 hold if, and only if, the conditions of 2) in Corollary 1 hold. Proof: (=): Assume that the conditions of Lemma 4 hold. Choose S = X 1 N = X 1 Y CF i = CF 0i X 1 . (62)
where ij is solved as shown at the bottom of the page. Combining it and (62) and (63), then we have Hii < 0, Hij + Hji < 0, 1ir 1i<jr
i.e., (56) and (57) hold. Thus, the conditions of 2) in Corollary 1 hold. (=): Assume that the conditions of 2) in Corollary 1 hold. Then choose X = S 1 Y =SN CF 0i = CF i S 1 . (65)
AF 0ij = AF ij S 1 + BF j C2i S 1 + (S N )T Ai S 1 BF 0i = BF i (66)
AF ij = AF 0ij X 1 BF 0j C2i Y T Ai BF i = BF 0i (63)
From (55) and (65) and considering the lower triangular structure of S and N , we have
1 X11 0
Because X and Y have the lower triangular structures, then

1 S11 = X11 1 S22 = X22 1 N11 = X11 Y11 1 N22 = X22 Y22 .
0 1 X22
>0 0 1 X22 Y22 0 N22 < 0. = S11 Y11 0 0 S22 Y22
1 X11 Y11 0
(64)
N11 0
From (15), we have Y11 0 0 Y22 >0 X11 0 0 X22 > 0.
Applying Schur complement to the earlier inequality, then it follows that (15) holds. Moreover, pre- and postmultiplying (56) and (57) by diag[X T I I I] and its transpose, then we can obtain Hii < 0, 1ir 1i<jr
Combining it and (64), then, yields S11 =

1 X11
> 0 S22 =
1 X22
> 0.
Hij + Hji < 0,
S T Ai + AT S i AF 0ij S + AT S i ij = T B1i S C1i CF 0j S
T T T T Ai Y + Y Ai + BF 0j C2i + C2i BF 0j
T T T B1i Y + D21i BF 0j
I 0
2 I
C1i
58
these parameters, the dynamic of the tunnel diode circuit can be rewritten as x1 (t) = 0.1x1 (t) 0.5x3 (t) + 10x2 (t) + w1 (t) 1 x2 (t) = x1 (t) 0.5x2 (t) y(t) = Sx(t) + w2 (t) z(t) = x1 (t).
Fig. 1. Tunnel-diode-circuit diagram.
(67)
where Hij is solved as shown at the bottom of the page. Combining it and (65) and (66), then we have ii < 0, ij + ji < 0, 1ir 1i<jr
For the sake of simplicity, we will use as few rules as possible. Assuming that |x1 (t)| 3, the nonlinear network system (67) can be approximated by the following T-S fuzzy model:
2
E x(t) =
i=1 2
i (t)Ai x(t) + B1 w(t)
i.e., (16) holds. Thus, the conditions of Lemma 4 hold. IV. E XAMPLES In this section, two numerical examples are given. The rst one shows the effectiveness of the new design method with improving the upper bound of singular-perturbation parameter . The second one further shows the new given design method [1) in Corollary 1] without considering that the bound of singularperturbation parameter can give less-conservative results than the existing method in [20] (Lemma 4). Example 1: Consider a tunnel diode circuit shown in Fig. 1, where the tunnel diode is characterized by
3 iD (t) = 0.01vD (t) + 0.05vD (t).
y(t) =
i=1
i (t)C2i x(t) + D21 w(t)
z(t) = x1 (t) where i (t) is the normalized time-varying fuzzy weighting function for each rule i = 1, 2, x(t) = [xT (t), xT (t)]T 1 2 A1 = A2 = B1 = 0.1 1 4.6 1 1 0 0 0 0] 0] 10 0.5 10 0.5
C1 = [ 1
Assuming that the inductance L is the parasitic parameter and letting x1 (t) = vC (t) and x2 (t) = iL (t) be the state variables, we have C x1 (t) = 0.01x1 (t) 0.05x3 (t) + x2 (t) + 0.1w1 (t) 1 L x2 (t) = x1 (t) Rx2 (t)
C21 = C22 = [ 1 D21 = [ 0 E = 1]
1 0 . 0
y(t) = Sx(t) + w2 (t) z(t) = x1 (t) where w(t) = [w1 (t) w2 (t)]T is the disturbance noise input, y(t) is the measurement output, z(t) is the controlled output, and S = [1 0] is the sensor matrix. The parameters of the circuit are C = 100 mF, R = 0.5 , and L = H = 0.01 H. With
Fig. 2 shows the plot of the membership functions for rules 1 and 2. First, we apply Theorem 1 to design lter with weighting scalars a1 = 0.001 a5 = 10 000 b3 = 0.8 a2 = 0.0011 a6 = 0.01 a3 = 0.001 b1 = 0.1 a4 = 10 000
b2 = 0.1
b4 = 1 b5 = 0.0005
b6 = 0.07
Ai X + X T AT i AF ij X + BF j C2i X + (S N )T Ai X + AT i Hij = T B1i C1i X CF j X
T T T T Ai Y + Y Ai + BF j C2i + C2i BF j
T T T B1i Y + D21i BF j
I 0
2 I
C1i
59
ESTIMATED BOUNDS
TABLE I WITH (0
] BY
THEOREM 2
Fig. 2.
Membership functions (x1 (t)) and (x1 (t)).
and = 0.05. The corresponding gain matrices are given as follows: AF 11 = AF 12 = AF 21 = AF 22 = BF 1 = BF 2 = 3.7234 0.7649 4.1219 0.7892 7.6107 0.6166 5.2345 0.7637 3.1647 0.1557 0.3256 0.0169 0.0001 ] 0.0001 ] . (68) 10.0593 0.5029 10.0578 0.5028 10.0578 0.5028 10.0590 0.5028
CF 1 = [ 0.9372 CF 2 = [ 0.8578
Next, applying Lemma 4, i.e., the approach in [20] [which is equivalent to 2) in Corollary 1] for designing a lter, the lter gain matrices are given as follows: AF 11 = AF 12 = AF 21 = AF 22 = BF 1 = BF 2 = 60.0861 1.1746 6.0267 0.7924 5.9058 0.7924 39.9183 1.1834 0.1803 0.0000 0.1803 0.0000 0.2871 ] 0.2058 ] (69) 66.7726 0.5873 40.9452 0.3962 40.9459 0.3962 70.7685 0.5917
Since Example 1 is the simplest fuzzy model (i.e., fuzzy model l with two fuzzy rules), the introduced variables Jij , 1 i, j, and l r of 1) in Corollary 1 do not play a role for reducing conservatism, and the computational results show that 1) in Corollary 1 and Lemma 4 [or 2) in Corollary 1] give the same H performance index. In addition to the lter-design technique in Theorem 1, a new method for estimating the upper bound of singularperturbation parameter is also given in Theorem 2. In order to illustrate the effectiveness of both the new lter-design technique (Theorem 1) and the method for estimating the upper bound of (Theorem 2), the bounds of of the singularly perturbed fuzzy system with the lter gains (68) and (69) will be, respectively, estimated by Theorem 2 and the methods in [18] and [19]. Note that the type of bounds [ ] with > 0 is considered in [18], where a weighting matrix Q is involved. The T T C1 ()C1 () C1 ()CF () weighting matrix is Q = T T CF ()C1 () CF ()CF () for this example. Unfortunately, the condition of [18] is infeasible for all or (0, +); the reason is that, in the AT P + P Aij + Q P ij condition, block matrices P 1I 2 T P + P Aij + Q P Aij and [18, Eq. (53)] have to P 1I 2 be less than zero, where = diag[I I I I], = I I]. diag[I I However, if a smaller Q is exploited, for example, 0.12 Q is considered, then the condition in [18] is feasible, and the corresponding results are given in Table II. Moreover, the obtained upper bounds by Theorem 2 and the methods in [19] are shown in Tables I and III, respectively. For comparisons with the existing methods, we have the following observations. 1) From Tables IIII, it can be seen that Theorem 1 is effective to obtain larger upper bounds of singular-perturbation parameter than the existing lter-design method (Lemma 4). 2) For the methods for estimating the upper bounds of singular-perturbation parameter (Theorem 2 and the ones in [18] and [19]), Tables IIII show that Theorem 2 can give larger estimated bounds of singular-perturbation parameter for this example than the existing methods in [18] and [19]. 3) Note that the singular-perturbation parameter = 0.01 in the example, for 1 (t) = 1 and 2 (t) = 0 Ae () = E 1 A1 E 1 BF 1 C21 0 . E 1 AF 11
CF 1 = [ 0.4318 CF 2 = [ 0.5857 = 0.0497.
The eigenvalues of Ae () with the designed lter gain (69) are 25.0500 + 19.4293i, 25.0500 19.4293i,
60
TABLE II ESTIMATED BOUNDS AND WITH [ ] BY THE METHOD OF [18]
ESTIMATED BOUNDS
TABLE III WITH (0 ] BY THE METHOD OF [19]
Fig. 3. Trajectory of 1 (t) with the gain (69).
Fig. 4.
Trajectory of ze = z (t) z(t) with gain (68).
0.6771 + 65.6805i, and 0.6771 65.6805i. Therefore, the designed lter by Lemma 4 (the existing method) cannot be applied. The fact is also illustrated by the trajectory of the state 1 (t) of the lter in Fig. 3. Therefore, the new proposed method with improving the upper bounds of is necessary. Following these comparisons, some simulations will be given to illustrate the effectiveness of the new proposed method. Assume that the initial state x(0) = [0 0]T and the disturbance w(t) = sin(t)[1/(t + 1) 1/(t2 + 1)], the state 1 (t) of the lter with the gain (69) is given in Fig. 3. It can be seen that 1 (t) is not convergent, which shows that the existing method [2) in Corollary 1, i.e., Lemma 4] does not work. The reason is that the existing method [2) in Corollary 1] does not address the issue of improving the bound of singularperturbation parameter when the lter is designed, and the bound of singular-perturbation parameter of the singularly perturbed system with the designed lter gain (69) is less than the practical singularly perturbed parameter = 0.01 for this example. Therefore, the resulting system is unstable. Using the gain (68) given by Theorem 1 (the new proposed method), the trajectory of ze = z (t) z(t), i.e., the lter error, is shown in Fig. 4; it is shown that the lter error ze = z (t) z(t) tends to zero, which means that the lter estimation z ensures that fast-tracking characteristic. For some instant T , the ratio of
T T T 0 ze (t)ze (t)dt/ 0
Fig. 5.
T 0
T ze (t)ze (t)dt/
T 0
wT (t)w(t)dt with gain (68).
wT (t)w(t)dt can show the in-
uence of disturbance w(t) on the lter error ze = z (t) z(t), and the plot of the ratio is shown in Fig. 5. It can be seen that the ratio tends to a constant value 0.016, which is less than the prescribed value = 0.05 (an obtained H performance-index requirement). These simulations show that the new proposed method is effective.
61
Example 2: Consider the singularly perturbed fuzzy system (3) with r = 2 A1 = B11 C11 C21 0 10 2 8 0.6 0 = 1 0 = [1 0] = [0 3] 1] 2 6.1 3 9 1 0 B12 = 2 0 C12 = [ 1 1 ] C22 = [0 3] A2 = D212 = [0 1] E = 1 0 0 .
Choose S = X 1 and M = X . Then from (72), it follows that N = X 1 Y . Since 1 is invertible, then from (73), it follows that P > 0 is equivalent to X I I Y > 0. (74)
D211 = [2
Applying 1) and 2) in Corollary 1 to the example, then the obtained optimal H performance bounds are = 2.0760 l and = 2.2881, which implies that the slack variables Jij of 1) in Corollary 1 are helpful for less-conservative results. Since Lemma 4 (the method in [20]) is equivalent to 2) in Corollary 1 (see Theorem 3), the example also shows that the new technique [1) in Corollary 1] can give less-conservative results than the existing method [20]. V. C ONCLUSION In this paper, we have investigated the problem of designing H lters for fuzzy singularly perturbed systems with the consideration of improving the bound of singular-perturbation parameter. An LMI-based approach is presented for simultaneously designing the bound of the singularly perturbed parameter and H lters for a fuzzy singularly perturbed system. When the bound of singularly perturbed parameter is not under consideration, the result reduces to an LMI-based design method for H ltering of fuzzy singularly perturbed systems, which is less conservative than the existing one. Furthermore, a method is given for evaluating the upper bound of singularly perturbed parameter subject to the constraint that the considered system is to be with a prescribed H performance bound, and the upper bound can be obtained by solving a GEVP. Numerical examples are given to illustrate the effectiveness of the proposed methods. A PPENDIX Proof of Lemma 2: Proof: Let the symmetric matrix P in Lemma 1 have the following decomposition: P = Y NT N P 1 = X MT M (70)
Applying Schur complement to (74), then we have proof that (8) holds. Therefore, (74) holds when (8) holds. Denote Me1 () = P Ae () + AT ()P e
T + P Be ()Be ()P +
1 T C ()Ce (). 2 e
Pre- and postmultiplying the earlier equality by T and 1 , 1 it follows that Me2 () = T Me1 ()1 1
T = Me2a () + Me2a ()
+ +
YE
E 1 B1 () B1 () + N E 1 BF ()D21 ()
T
E 1 B1 () 1 Y E B1 () + N E 1 BF ()D21 () 1 [ C1 ()X CF ()M T 2 C1 () ]T
[ C1 ()X CF ()M T where, from (8), it follows that S 0
C1 () ]
(75)
0 is invertible. I S 0 Pre- and postmultiplying (75) by and its transpose, 0 I then yields Me2a (), shown at the bottom of the next page, and Me3 () = S 0 S 0 Me2 () I 0 0 I
T =Me3a () + Me3a ()
where M and N are invertible. Denote X 1 = MT I 0 I 2 = 0 Y . NT (71)
+ +
(S N )E
S E 1 B1 B1 ()+N E 1 BF ()D21 ()
T
S E 1B1 () 1 (S N )E B1 ()+N E 1BF ()D21 () 1 [ C1 ()CF () 2 C1 () ]T
Then 1 and 2 are invertible, and the following equalities hold: P 1 = 2 M NT = I X Y X I T P 1 = 1 I Y (72) . (73)
[ C1 ()CF ()
C1 () ]
where Me3a () is given by (10). From (9) and the earlier equality, it follows that Me3 () < 0. Since M = X is invertible, 1 is invertible.
62
From Me1 () = T 1 S 0 0 I
1
Multiplying the earlier inequality by , then it follows that Me3 () S 0 0 I

1
1 < 0 1
S =
S11 S21
T S21 S22
> 0,
for
0,
1 .
(76)
then we have that Me1 () < 0. Applying Lemma 1, the conclusion follows. Thus, the proof is complete. Proof of Theorem 1: Proof: In order to give a brief expression, we use the following notations:
r r r
Similarly, from condition (22), we have N = N11 N21

T N21 N22
< 0,
for
0,
1 .
(77)
Part 2: Denote Me5ab = 0 [ N T AF + N T BF C2 T Na (N 1 )T N T BF C2 ] . (78) Then

T Me5ab Na (N 1 )T
A=
i=1 r
i Ai
B1 =
i=1
i B1i
r
C1 =
i=1
i C1i
C2 =
i=1 r
i C2i
r
D21 =
i=1
i D21i
r
AF =
i=1 j=1 r
i j AF ij
BF =
i=1 r r
i BF i
N T AF + N T BF C2
N T BF C2
I.
(79)
CF =
i=1 r
i CF i
AF =
i=1 j=1 r
i j AF ij
BF =
i=1
i BF i
CF =
i=1
i CF i .
Multiplying (27) by i j and summing them, we have 2 T 5 I 0 AT + C2 BF F 2 T T < 0. 0 5 I C2 BF AF + BF C2 BF C2 I Applying the Schur complement to the earlier inequality, then it follows that
T T AT + C2 BF F [ AF + BF C2 T T C2 BF 2 BF C2 ] 5 I
Let AF ij = N T AF ij BF i = N T BF i .
The proof consists of four parts. In part 1, we prove that (76) and (77) hold. In parts 2 and 3, we prove that (91) and (93) holds, respectively. Finally, the proof is completed in part 4 by using (76), (77), (91), and (93). Part 1: First, from condition (21), we have
1 S11 < 1 1 S22 < 2 T S21 S21 < 3 .
which implies AT + BF C2 F i.e., N T AT + N T BF C2 F Combining it and (79), we have

T Me5ab Na (N 1 )T 5 I.
BF C2
< 5
N T BF C2
< 5 .
Combining it and (31), then, yields

T 1 S112 S21 S22 S21 S112 1 1
1 2 3
(80)
which implies
T 1 S112 S21 S22 S21 S112 1 2 3 I I. 1 1
Pre- and postmultiplying the earlier inequality by S11 and its transpose, and applying the Schur complement, then we can obtain
1
1/2
Multiplying (24) by i j and summing them, we obtain 2 6 I 0 S T B1 2 6 I (S T N T )B1 + BF D21 < 0. I Applying Schur complement to the earlier inequality, then, yields S T B1 (S N )B1 + BF D21
T T
S11 T S21
T S21 S22
> 0,
for
0,
1 .
S T B1 (S N )B1 + BF D21
T T
T 2 6 I
Me2a () =
E 1 A()X Y E 1 A()X + N E 1 AF ()M T + N E 1 BF ()C2 ()X YE

1
E 1 A() A() + N E 1 BF ()C2 ()
63
which implies that S T B1 (S N )B1 + BF D21

T T
From (82)(84), it follows that < 6 . (81)

T Me5b 6 3 + Na (N 1 )T 2 I T Sa B1 T B1 + Na BF D21 T + Na (N 1 )T
(85)
Denote Me5b = S T B1 (S N )B1 + N T BF D21

T T T Sa
T Na
Then Me5b
T T T Sa Na B1 + Na BF D21
T Sa B1
Sa B1 T T Sa Na B1
BF D21 (86)
T = 3 + Na (N 1 )T 2 .
S T B1 (S N )B1 + N T BF D21
T T T T S a Na T Sa B1 T B1 + Na BF D21 T
Multiplying (29) by i , 1 i r and summing them, then we have I

T Me5aa + Me5aa = i=1 r T i Me5aai + Me5aai
S T B1 (S N )B1 + N T BF D21
T T T Sa B1 T T Sa Na B1 T
= 0 T Na BF D21
T
T Sa A() (Sa Na )T A()
T Sa A() (Sa Na )T A() T
I +
S T B1 (S N )B1 + N T BF D21
T T T Sa B1 T T Sa Na B1 T T + Na (N 1 )T
T T Sa A() Sa A() (Sa Na )T A() (Sa Na )T A()
< 1 I. BFD21 I.
(87)
Now, consider Me5 as follows:

T T Me5 = Me5a + Me5a + Me5b + Me5b
Combining it and (81), then, yields Me5b 6

T Sa B1 T T Sa Na T T + Na (N 1 )T
B1
BF D21
I. (82)
T Sa
T Na
T Sa B1 T B1 + Na BF D21 T
Multiplying (25) by i j and summing them, then it follows that

2 2 I F D21 B
T T S a Na
T Sa B1 T B1 + Na BF D21
<0
T T T = Me5aa + Me5aa + Me5ab + Me5ab + Me5b + Me5b
which implies that BF D21 2 . Similarly, from (26), we can obtain

T Sa T Sa B1 T Na B1
+ (83)
T Sa
T Na
T Sa B1 T B1 + Na BF D21 T
T T S a Na
(88)
3 .
(84)
and Me5a and Me5b , shown at the bottom of the page, where Me5aa and Me5ab are, respectively, the same as in (78) and (87),
Me5a = Me5b =
T Na AF
T Sa A T Na BF C2 +
T Sa A T (Sa Na ) A (Sa Na ) A + Na BF C2 T T T Sa B1 T B1 + Na BF D21 T
S T B1 T T (S N )B1 + N T BF D21
T T S a Na
64
and from (80) and (85)(87), we have

T T T Me5 Me5aa + Me5aa + Me5ab + Me5ab + Me5b + Me5b
Applying the Schur complement to (92), then, yields Me4 + < 0 where
T Me4 = Me4a + Me4a
(93)
T Sa
T Na
T Sa B1 T B1 Na BF D21
T T = 1 +2 Na (N 1 )T 5 +26 3 + Na (N 1 )T 2 T + 3 + Na (N 1 )T 2 2
S T B1 (S N )B1 N T BF D21
T T T
= 7 I where 7 is the same as in (32). By the denition of in (31), it follows that 4 7 . On the other hand, from (23), we can obtain
1 4 I <
S T B1 T T (S N )B1 N T BF D21 1 + 2 [ C1 CF C1 ]T [ C1 CF Part 4: From (91) and (93), it follows: (89) Me4 + Me5 < 0, 0,
C1 ] < 0.
4 > 0
which implies that 1 < 4 . Then, from (89) and (90), it yields 2 Me5 2
1 1
(90)
where the denitions of Me5 , Me4a , and Me4 are given in (88), (92), and (93), respectively. Consider the solution for Me4 + Me5 , shown on the next page. Combining it and (33), then it follows that (94), shown on the page following the next page, where S and N are the same as in (47) and (48). Pre- and postmultiplying the earlier inequality by T 1 S 0 0 I
T
7 Me5 I I 4 7 I.
Combining it and (89), then we have 2 Me5 2 I

1 1
which is equivalent to Me5 which implies that Me5 , for 0, 1 (91)
and its transpose, then we have (95), shown on the page following the next one. Consider P is same as in (70) and choose X = S 1 , M = X = S 1 , then from (72), we have N = S Y and (96), shown on the page following the next one. Therefore T 1 S 0 0 I
T
(Me4 + Me5 )
S 0
0 I
1 1 0, 1
= T Me2 ()1 < 0, 1 1
where Me5 is the same as in (88). Part 3: From condition (28) and Lemma 3, we obtain
r r
where Me2 () is same as in (75). Moreover, it follows that T Me2 ()1 = Me1 () from (75). Then 1 1 Me1 < 0, 0, 1 . (97)
i j ij < 0
i=1 j=1
i.e., S T B1 T M +Me4a + T T e4a (S N )B1 + BF D21 I where Me4a
T T C1 CF T C1 < 0 0 2 I (92)
Finally, from (76) and (77), we have S > 0 and N < 0, respectively. Note that X = S 1 , N = S Y , then X > 0, X 1 Y < 0, applying Schur complement, it follows that X I > 0. Combining it and (73), then we can obtain I Y P > 0. Then, by P > 0, (97), and Lemma 1, it follows that the H norm of the system (6) with AF i = N T AF i BF i = N T BF i
STA STA = . T AF + BF C2 + (S N ) A (S N )T A+ BF C2
is less than or equal to for (0, (1/ )].
65
T Me4 + Me5 = Me4a + Me4a +
T T
T T
+ + = + + + +
1 [C1 CF C1 ]T [C1 CF C1 ] + 2
2 T T S a Na T Sa B1 T B1 + Na BF D21
T T Me5a + Me5a + Me5b + Me5b T Sa B1 T B1 + Na BF D21 T
T T S a Na
STA AF + BF C2 + (S N )T A
STA STA STA + T T AF + BF C2 + (S N ) A (S N )T A+ BF C2 (S N ) A+ BF C2 S T B1 T T (S N )B1 N T BF D21

T T T
S T B1 T T (S N )B1 N T BF D21
T Na AF
1 [C1 CF C1 ]T [C1 CF C1 ] 2
+ +
T Sa A T Na BF C2 + T Sa A T Na BF C2 +
T Sa A T (Sa Na ) A (Sa Na ) A + Na BF C2 T Sa A T (Sa Na ) A (Sa Na ) A + Na BF C2 T T T Sa B1 T B1 + Na BF D21 T T
T Na AF
S T B1 T T (S N )B1 + N T BF D21 S T B1 T T (S N )B1 + N T BF D21

2 T Sa
T T S a Na
T T S a Na
T Sa B1 T B1 + Na BF D21 T Sa B1 T B1 + Na BF D21
T Na
T Sa
T Na
STA STA F + BF C2 + (S N )T A (S N )T A + BF C2 A +
T Sa A T Na BF C2 + T Sa A T (Sa Na ) A (Sa Na ) A + Na BF C2 T T T
T Na AF
STA STA F + BF C2 + (S N )T A (S N )T A + BF C2 A +
T Sa A T Na BF C2 +
T Na AF
T Sa A T (Sa Na ) A (Sa Na ) A + Na BF C2 T T
1 [C1 CF C1 ]T [C1 CF C1 ] 2 S T B1 T T (S N )B1 N T BF D21 +

T T
S T B1 T T (S N )B1 N T BF D21
T Sa
S T B1 (S N )B1 + N T BF D21 S T B1 T T (S N )B1 + N T BF D21

2 T T S a Na T Sa B1 T B1 + Na BF D21
T Na
T T S a Na
T Sa B1 T B1 + Na BF D21 T Sa B1 T B1 + Na BF D21
T T S a Na
< 0,
0,
66
Me4 + Me5 = +
N E
S T E 1 A AF + N E BF C2 + (S N )T E 1 A
T 1
S T E 1 A (S N ) E 1 A + N T E 1 BF C2
T
N E 1 + 2 [C1 CF C1 ]T [C1 CF C1 ] + = +
T T T
T 1
S T E 1 A (S N ) E 1 A + N T E 1 BF C2
T
S T B1 + (S N )B1 N T BF D21 S T B1 + T T (S N )B1 N T BF D21 N E

1 T 1
T Sa
T Na
T Sa B1 T B1 + Na BF D21 T Sa B1 T B1 + Na BF D21 T
T T S a Na
S T E 1 A (S N ) E 1 A + N T E 1 BF C2
T
S T E 1 A N T E 1 AF + N T E 1 BF C2 + (S N )T E 1 A 1 + 2 [C1 CF C1 ]T [C1 CF C1 ] + < 0, S T E 1 B1 T T S N E 1 B1 + N T E 1 BF D21 1 0,
S T E 1 A (S N )T E 1 A + N T E 1 BF C2
ST N T
S T E 1 B1 E 1 B1 + N T E 1 BF D21
(94)
T 1
S 0
0 I
(Me4 + Me5 )
T 1 1
S 0
T
0 I
1
1 1 E 1 A (S N ) E A + N T E 1 BF C2
T 1
= T 1
N E +
AF S
1
E 1 AS 1 + N E BF C2 S 1 + (S N )T E 1 AS 1
1 T 1
N E AF S 1 + 2 [C1 S 1 CF S 1 C1 ]T [C1 S 1 CF S 1 C1 ] + < 0, S N

T T
E 1 AS 1 + N E BF C2 S 1 + (S N )T E 1 AS 1
E 1 A (S N ) E A + N T E 1 BF C2
T 1
E 1 B1 E B1 + N T E 1 BF D21
1
S N
T
E 1 B1 E B1 + N T E 1 BF D21
1
1 1 (95)
0,
T 1
S 0
0 I
(Me4 + Me5 )
T 1 T
S 0
0 I
1 1 YE
1 1
= T 1 +
N E
T 1
E 1 AX AF M + N T E 1 BF C2 X + Y E 1 AX
T
E 1 A A + N T E 1 BF C2
T
N E 1 + 2 [C1 X CF M C1 ]T [C1 X CF M C1 ] + < 0, E 1 B1 1 Y E B1 + N T E 1 BF D21 0, 1
E 1 AX AF M + N T E 1 BF C2 X + Y E 1 AX
YE
E 1 A A + N T E 1 BF C2
E 1 B1 1 Y E B1 + N T E 1 BF D21
1 1 (96)
67
R EFERENCES
[1] P. Kokotovic, H. K. Khalil, and J. OReilly, Singular Perturbation Methods in Control: Analysis and Design. New York: Academic, 1986. [2] Z. Gajic and M. T. Lim, A new ltering method for linear singularly perturbed systems, IEEE Trans. Autom. Control, vol. 39, no. 9, pp. 1952 1955, Sep. 1994. [3] X. Shen and L. Deng, Decomposition solution of H lter gain in singularly perturbed systems, Signal Process., vol. 55, no. 3, pp. 313 320, Dec. 1996. [4] M. T. Lim and Z. Gajic, Reduced-order H optimal ltering for systems with slow and fast modes, IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 47, no. 2, pp. 250254, Feb. 2000. [5] T. Takagi and M. Sugeno, Fuzzy identication of systems and its applications to modeling and control, IEEE Trans. Syst., Man, Cybern., vol. SMC-15, no. 1, pp. 116132, Jan. 1985. [6] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analysis. Hoboken, NJ: Wiley, 2000. [7] X. J. Ma, Z. Q. Sun, and Y. Y. He, Analysis and design of fuzzy controller and fuzzy observer, IEEE Trans. Fuzzy Syst., vol. 6, no. 1, pp. 4151, Feb. 1998. [8] K. Tanaka, T. Hori, and H. O. Wang, A multiple Lyapunov function approach to stabilization of fuzzy control systems, IEEE Trans. Fuzzy Syst., vol. 11, no. 4, pp. 582589, Aug. 2003. [9] H.-N. Wu and K.-Y. Cai, Robust fuzzy control for uncertain discrete-time nonlinear Markovian jump systems without mode observations, Inf. Sci., vol. 177, no. 6, pp. 15091522, Mar. 2007. [10] W.-J. Wang, Y.-J. Chen, and C.-H. Sun, Relaxed stabilization criteria for discrete-time T-S fuzzy control systems based on a switching fuzzy model and piecewise Lyapunov function, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 37, no. 3, pp. 551559, Jun. 2007. [11] C. S. Tseng and B. S. Chen, H fuzzy estimation for a class of nonlinear discrete-time dynamic systems, IEEE Trans. Signal Process., vol. 49, no. 11, pp. 26052619, Nov. 2001. [12] B. S. Chen, C. L. Tsai, and D. S. Chen, Robust H and mixed H2 lters for equalization designs of nonlinear communication systems: Fuzzy interpolation approach, IEEE Trans. Fuzzy. Syst., vol. 11, no. 3, pp. 384398, Jun. 2003. [13] S. K. Nguang and W. Assawinchaichote, H ltering for fuzzy dynamical systems with D stability constraints, IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 50, no. 11, pp. 15031508, Nov. 2003. [14] G. Feng, Robust H ltering of fuzzy dynamic systems, IEEE Trans. Aerosp. Electron. Syst., vol. 41, no. 2, pp. 658670, Apr. 2005. [15] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix inequalities in systems and control theory, in SIAM Studies in Applied Mathematics. Philadelphia, PA: SIAM, 1994. [16] H. Liu, F. Sun, and Z. Sun, Stability analysis and synthesis of fuzzy singularly perturbed systems, IEEE Trans. Fuzzy. Syst., vol. 13, no. 2, pp. 273284, Apr. 2005. [17] W. Assawinchaichote and S. K. Nguang, H fuzzy control design for nonlinear singularly perturbed systems with pole placement constraints: An LMI approach, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 1, pp. 579588, Feb. 2004. [18] T. H. S. Li and K. J. Lin, Stabilization of singularly perturbed fuzzy systems, IEEE Trans. Fuzzy. Syst., vol. 12, no. 5, pp. 579595, Oct. 2004. [19] T.-H. S. Li and K.-J. Lin, Composite fuzzy control of nonlinear singularly perturbed systems, IEEE Trans. Fuzzy. Syst., vol. 15, no. 2, pp. 176187, Apr. 2007.
[20] W. Assawinchaichote and S. K. Nguang, H ltering for fuzzy singularly perturbed systems with pole placement constraints: An LMI approach, IEEE Trans. Signal Process., vol. 52, no. 6, pp. 16591667, Jun. 2004. [21] H. Mukaidani, A numerical algorithm for nding solution of sign-indenite algebraic Riccati equations for general multiparameter singularly perturbed systems, Appl. Math. Comput., vol. 189, no. 1, pp. 255270, Jun. 2007. [22] N. R. Sandell, Jr., Robust stability of systems with application to singular perturbations, Automatica, vol. 15, no. 4, pp. 467470, Jul. 1979. [23] L. Cao and H. M. Schwartz, Complementary results on the stability bounds of singularly perturbed systems, IEEE Trans. Autom. Control, vol. 49, no. 11, pp. 20172021, Nov. 2004. [24] Z. H. Shao, Robust stability of two-time-scale systems with nonlinear uncertainties, IEEE Trans. Autom. Control, vol. 49, no. 2, pp. 258261, Feb. 2004. [25] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control Toolbox. Natick, MA: MathWorks, Inc., 1995. [26] A. Isidori, H control via measurement feedback for afne nonlinear systems, Int. J. Robust Nonlinear Control, vol. 4, no. 4, pp. 553574, Apr. 1994.
Guang-Hong Yang (SM04) received the B.S. and M.S. degrees in mathematics and the Ph.D. degree in control engineering from Northeastern University, Shenyang, China, in 1983, 1986, and 1994, respectively. From 1986 to 1995, he was a Lecturer/Associate Professor with Northeastern University. In 1996, he was a Postdoctoral Fellow with the Nanyang Technological University, Singapore, Singapore. From 2001 to 2005, he was a Research Scientist/Senior Research Scientist with the National University of Singapore, Singapore. He is currently a Professor with the College of Information Science and Engineering, and Key Laboratory of Integrated Automation of Process Industry (Ministry of Education), Northeastern University. His current research interests include fault-tolerant control, fault detection and isolation, nonfragile control systems design, and robust control. Dr. Yang is an Associate Editor for the International Journal of Systems Science and the International Journal of Control, Automation, and Systems. He is also an Associate Editor of the Conference Editorial Board of the IEEE Control Systems Society.
Jiuxiang Dong was born on January 2, 1978, in Shenyang, China. He received the B.S. degree in mathematics and applied mathematics and the M.S. degree in applied mathematics from Liaoning Normal University, Dalian, China, in 2001 and 2004, respectively. He is currently working toward the Ph.D. degree in the College of Information Science and Engineering and Key Laboratory of Integrated Automation of Process Industry (Ministry of Education), Northeastern University, Shenyang. His research interests include fuzzy, robust, and fault-tolerant controls.
68
69
Automatica 43 (2007) 1597 1604 www.elsevier.com/locate/automatica
Brief paper
Parameterization of all stabilizing H static state-feedback gains: Application to output-feedback design

J. Gadewadikar a, , Frank L. Lewis a , L. Xie b , V. Kucera c , M. Abu-Khalaf a
a Automation and Robotics Research Institute, University of Texas at Arlington, USA b School of Electrical and Electrical Engineering, Nanyang Technological University, Singapore c Center for Applied Cybernetics, and Control Engineering at Czech Technical University, Prague, Czech Republic
Received 8 December 2005; received in revised form 10 October 2006; accepted 5 February 2007 Available online 6 July 2007
Abstract This paper presents a simplied parameterization of all H static state-feedback controllers in terms of a single algebraic Riccati equation and a free parameter matrix. As a special case, necessary and sufcient conditions for the existence of an static output-feedback gain are given. An efcient computational algorithm is given and its correctness proven. No initial stabilizing output-feedback gain is needed. The technique is used to design an H lateraldirectional command augmentation system for the F-16 aircraft. 2007 Elsevier Ltd. All rights reserved.
Keywords: Controller parameterization; Static output feedback; H control
1. Introduction The parameterization of all stabilizing controllers and the static output-feedback control problems are two of the most researched and written about issues in modern control. Static output feedback is indeed a special case of the former problem where the controller gain is restricted to lie in some subspace. The computation of optimal H controllers of prescribed order and of static output-feedback controllers are non-convex problems, and so are difcult to confront. Main objective of this work is to present a simpler parameterization to achieve disturbance attenuation with a static state-feedback compensator; the approach is bringing H control, static compensation, and parameterization together. First, an introduction of the relevant work is given below.
1.1. Parameterization of H dynamic compensators The parameterization of all stabilizing polynomial controllers was rst given by Youla, Jabr, and Bongiorno (1976) and Kucera (1975, 1979) in terms of solutions to a Diophantine equation. The H approach (Doyle, Glover, Khargonekar, & Francis, 1989) has shown its effectiveness in the design of controllers for modern systems. A parameterization of H dynamic compensators for the state-feedback and full-information cases was given by Zhou (1992) in terms of the Hamiltonian matrix and the YoulaKucera parameterization. Parameterization of H dynamic compensators was given by Iwasaki and Skelton (1995). Henrion, Kucera, and Molina-Cristobal (2005) use the YoulaKucera parameterization to provide a method for designing low order compensators by simultaneous optimization over the numerator and denominator of the compensator. Application is made to H control. 1.2. Static output feedback A survey of static output-feedback control is given by Syrmos, Abdallah, Dorato, and Grigoriadis (1997). Conditions often involve two Riccati equations coupled by a spectral
This paper was not presented at any IAFC meeting. This paper was recommended for publication in revised form by Associate Editor Sam Ge under the direction of Editor Miroslav Krstic. Corresponding author. Tel.: +1 817 272 5967; fax: +1 817 272 5989. E-mail addresses: jyotirmay@gmail.com (J. Gadewadikar), lewis@uta.edu (F.L. Lewis), ELHXIE@ntu.edu.sg (L. Xie), kucera@fel.cvut.cz (V. Kucera), abukhalaf@arri.uta.edu (M. Abu-Khalaf). 0005-1098/$ - see front matter 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.automatica.2007.02.005
70
1598
J. Gadewadikar et al. / Automatica 43 (2007) 1597 1604
radius condition (Lewis & Syrmos, 1995). Moerder and Calise (1985) presented an algorithm for computing the optimal H2 static output-feedback gain that is in standard use, along with a convergence proof. This algorithm is extended and used for aircraft control design in (Stevens & Lewis, 2003). An initial stabilizing output-feedback gain is needed, which is difcult to determine for practical systems. Recent results address these issues. Yu (2004) provides a solution for the optimal static output-feedback problem in terms of two AREs plus a projective coupling condition, but the AREs are standard ones involving state feedback, not output feedback. An algorithm is presented that requires an initial stabilizing state-feedback gain, which is easy to nd using standard LQR methods. A parameterization of all static output-feedback gains is given by Yasuda, Skelton, and Grigoriadis (1993) in terms of such equations. LMI approaches have been presented by Iwasaki and Skelton (1995), where a parameterization was also presented. In Colaneri, Geromel, and Locatelli (1997) and Geromel, de Souza, and Skelton (1998), a parameterization of static output-feedback gains is given in terms of the blocks of a symmetric matrix derived from coupled matrix inequalities. Also provided is a condition in terms of a matrix and its inverse. A numerical procedure is given to determine these quantities. LMI conditions for optimal and H static output feedback are given by Shaked (2003). El Ghaoui and Balakrishnan (1994) propose an iterative algorithm for designing LTI controllers with prespecied structure based on LMIs. Sufcient LMI conditions for OPFB control problem are presented in (Crusius and Trono, 1999). Work by Scherer (1995) was one of the rst that discussed the H problem with mixed H2 /H objectives. Trono-Neto and Kucera (1993) solve the static outputfeedback problem using inverse optimal control in terms of a single ARE projected onto nullspace perpendicular of C and a free parameter matrix.
to solve the H static output-feedback problem for discrete systems. To summarize; a lot of work has been pursued, however several problems are still open, most of the approaches though theoretically appealing, are hard to implement, are difcult to solve for higher-order systems, and can impose numerical problems. In this paper we present a parameterization of all stabilizing H controllers for continuous-time systems in terms of a single ARE and a free parameter matrix. The form of the result is simple and can be used to select the parameter matrix to satisfy additional design criteria. The result is used to develop necessary and sufcient conditions for existence of a solution to the H static output feedback problem for continuous-time systems. An algorithm is provided that does not require an initial stabilizing output-feedback gain, but only an initial stabilizing state-feedback gain, which is easy to nd. The correctness of the algorithm is proven, namely that if it converges, it yields the correct solution. The results are used to simply design an H lateraldirectional command augmentation system (CAS) for the F-16 aircraft given in Stevens and Lewis (2003). 2. Parameterization of all stabilizing static state-feedback gains 2.1. System description and denitions Consider the linear time-invariant (LTI) system in Fig. 1 x = Ax + Bu + Dd, x Rn, u Rm, y = Cx, y Rp (1)
and a performance output z(t) that satises z(t)

2
= x T Qx + uT Ru,
(2)
1.3. Motivation Solutions to the parameterization of all stabilizing controllers and the output-feedback problems are generally complex, depending for instance on coupled Riccati equations, LMIs, or on a matrix and its inverse belonging to some sets. Geromel and Peres (1985) give a sufcient condition for a stabilizing static output feedback in terms of a single ARE with a free parameter matrix, plus a condition on the form of the gain matrix. A computational algorithm is proposed. Kucera and de Souza (1995) provide necessary and sufcient conditions for stabilizing static output feedback control using these constructions. Gadewadikar, Lewis, and Abu-Khalaf (2006) extend these conditions to obtain necessary and sufcient conditions for the existence of an H static output feedback control. Geromel, Yamakami, and Armentano (1989) provide similar results for the discrete-time case for decentralized feedback and output feedback. de Souza and Xie (1992) provide for the discrete-time case a parameterization of all stabilizing H controllers in terms of similar constructions, and use that result
with QT = Q 0, R T = R > 0. It is assumed that C has full row rank, a standard assumption to avoid redundant measurements. The system L2 gain is said to be bounded or attenuated by if
0 0
z(t) d(t)
2 dt 2 dt
T T 0 (x Qx + u Ru) dt T 0 (d d) dt
(3)
for any non-zero energy-bounded disturbance input d. Call the minimum gain for which this occurs. Doyle et al. (1989) and Scherer (1992) show procedure to calculate using a quadratically convergent algorithm. For linear systems, there
z y x x = Ax + Bu + Dd y = Cx z
2
= xT Qx + uT Ru
u = Kx
Fig. 1. System description.
71
1599
are explicit formulae to compute , see, e.g., Chen (2000). Throughout this paper we shall assume that is xed and > . 2.2. Bounded L2 gain design problem Dening a constant state-feedback control as u = Kx, (4)
for (Q1/2 )T [(Q1/2 )T (R 1/2 K)T (1/ )P D] and L [L BR 1/2 0]. It follows that if pair (A, Q1/2 ) is detectable then pair (Ac , Q1/2 ) is detectable as well. Moreover Q 0, and P 0. Therefore, using the Lyapunov stability criteria (Wonham, 1985, Lemma 12.2), Eq. (9) implies closed-loop stability. Finally for z = [Q1/2 x R 1/2 u]T one has zT z = x T Qx, and P Ac + AT P + c 1
2
for the system described in Section 2.1, it is desired to nd the state-variable feedback (SVFB) gain K such that the closedloop system Ac A BK is asymptotically stable and the L2 gain is bounded by a prescribed value > . The following main theorem shows necessary and sufcient conditions for the parameterization of all stabilizing H static state-feedback gains. To the best of our knowledge, it provides a simpler parameterization than most other results in the literature. Theorem 1 (Parameterization of all H SVFB gains). Assume that (A, Q1/2 ) is detectable and (A, B) is stabilizable. Then K is a state feedback that stabilizes system (1) and guarantees L2 gain bounded by > , if and only if there exists a parameter matrix L such that K = R 1 (B T P + L), where P = PT 0, is a solution to 1
2
PDDT P + Q = 0,
(11)
and system L2 gain boundedness then follows from Van der Schaft (1992, Theorem 2). A special case of this result shows when there exists a static output-feedback (OPFB) gain F that stabilizes the system (1) with bounded L2 gain. OPFB requires a restricted form of the gain matrix K that has C as a right divisor. Dene the control input for OPFB as u = F y = F Cx = Kx which yields the closed-loop system A0 (A BF C). Equations for OPFB design are generally complicated, consisting of three coupled matrix equations, as given in Lewis and Syrmos (1995), Moerder and Calise (1985), and Yu (2004). The next result provides simplied equations for OPFB design. The corollary holds for a prescribed matrix Q and follows directly from Theorem 1. Corollary 1 (Existence of H OPFB for a given Q). Consider a specied Q 0 such that (A, Q1/2 ) is detectable, (A, B) is stabilizable, and a specied value of > . Then there exists an OPFB gain F such that A0 (A BF C) is asymptotically stable with bounded L2 gain if and only if there exists a parameter matrix L such that F C = R 1 (B T P + L), where P = P T 0, is a solution to 1
2
(5)
PA + AT P + Q + + LT R 1 L = 0.
PDDT P PBR1 B T P (6)
Proof. Necessity: Suppose that there exists a state-feedback gain K that stabilizes the closed loop system Ac A BK and satises L2 gain bounded by . Consider the equation P Ac + AT P + c 1
2
PDDT P + Q + K T RK = 0.
(7)
Considering the denition (2), from Knobloch, Isidori, and Flockerzi (1993, Theorem 2.3.1), closed-loop stability and L2 gain boundedness implies that Eq. (7) has a unique symmetric solution such that P 0. Rearranging Eq. (7) and completing the square will yield PA + A P + Q +
T
(12)
P A + AT P + Q +
PDDT P (13)
PDD P PBR 2
T
B P (8)
PBR1 B T P + LT R 1 L = 0.
+ (K R 1 B T P )T R(K R 1 B T P ) = 0.
Substituting the gain dened by (5) in (8) yields (6). Sufciency: Dene Q Q + K T RK + (1/ 2 )PDDT P and Q Q + K T RK. Suppose that (5) and (6) hold, then (7) follows, so that P Ac + AT P + Q = 0. c (9)
The next result gives necessary and sufcient conditions for the existence of a stabilizing OPFB with bounded L2 gain. It follows from Corollary 1 by setting Q=C T C. It is a renement of the result in Kucera and de Souza (1995), and extends that result to solve the H OPFB problem. Corollary 2 (Existence of H OPFB). Assume that (A, B) is stabilizable. Then, for a given > , there exists an OPFB gain such that A0 (A BF C) is asymptotically stable with L2 gain bounded by if and only if: (i) (A, C) is detectable and there exist matrices L and P = P T 0 such that
We claim that detectability of (A, Q1/2 ) implies detectability of (Ac , Q1/2 ). Note that detectability of (A, Q1/2 ) implies that 1/2 is stable for some L. One can write A LQ A LQ1/2 = Ac LQ1/2 (10)
72
1600
(ii) F C = R 1 (B T P + L). (iii) P A + AT P + C T C + 1

2
PDDT P (14)
The following algorithm is based on Yu (2004), where, however, three coupled equations must be solved. It solves for a state feedback at each iteration, and then projects that SVFB onto nullspace perpendicular of C. OPFB design algorithm.
. Set n = 0, L = 0. Solve a standard 1. Initialize: Fix 0 (e.g. LQR) Riccati equation for given Q and R and obtain a stabilizing SVFB gain as initial gain K0 . Dene closed-loop matrix A0 = A BK 0 . 2. nth iteration: Solve ARE for P T Pn (An ) + (An )T Pn + Q + Kn RK n
PBR1 B T P + LT R 1 L = 0.
Proof. Necessity: To prove necessity, set Q = C T C, K = F C and proceed with the necessity proof of Theorem 1. Note further that stability of (A BF C) implies detectability of (A, C). Sufciency: Suppose that (i)(iii) hold, then the sufciency proof of Theorem 1 holds with Q = C T C 0. Remarks. Note that under the special form Q = C T C, the performance output is given by z
2
1
2
Pn DD T Pn = 0,
(16)
= y
+ u 2.
update K, project onto nullspace perpendicular of C

T Kn+1 = R 1 (B T Pn + Ln )(I V2 V2 ),
Note that a stabilizability condition on (A, B) is hidden in the assumption of existence of a solution P = P T 0 to (14). 3. Solution algorithm for H output feedback Most existing iterative algorithms for static OPFB design require the determination of an initial stabilizing OPFB gain, which can be very difcult for practical systems, and the solution of three coupled design equationstwo Riccati equations and a spectral radius condition. Here is presented an algorithm to solve the two coupled design equations in Corollaries 1 and 2. The algorithm starts with a stabilizing SVFB gain, which can easily be found using standard computational tools (e.g. LQR). It is shown that the algorithm, if it converges, solves the outputfeedback problem. Since C has full row rank, the right inverse is dened as C + C T (CC T )1 , which is best computed using the SVD C = USVT = U [S0 0]
T V1 T V2
(17)
update L Ln+1 = RK n B T Pn , update closed-loop system matrix An+1 = A BK n+1 . (19) (18)
3. Check convergence. If converged, go to step 4 otherwise set n = n + 1 go to step 2. 4. End. Set K = Kn+1 and compute OPFB gain F = KV 1 (S0 )1 U T . The convergence can be checked using the norm of Pn+1 Pn e.g. Pn+1 Pn < , here is a small number and operator denotes the matrix norm. The next result shows the correctness of the algorithm, namely, that if it converges, it provides the H OPFB gain. Lemma 1. If this algorithm converges, it provides the H OPFB gain. Proof. Clearly, at convergence (7) holds for Pn+1 = Pn P . Substitution of Eq. (17) into Eq. (18) yields
T Ln+1 = RR 1 (B T Pn + Ln )(I V2 v2 ) B T Pn .
(15)
Then CV =C[V1 V2 ]=[C1 0], so that CV 1 =C1 , and CV 2 =0. T T Note that I = V1 V1 + V2 V2 . Therefore, (i) C + = V S + U T = V1 (S0 )1 U T . T T (ii) V1 V1 = I V2 V2 = C + C is the projection onto nullspace perpendicular of C.
T T (iii) V2 V2 = I V1 V1 = I C + C is the projection onto the nullspace of C.
The equation F C = R 1 (B T P + L) has an exact solution F if and only if

T 0 = (B T P + L)(I C + C) = (B T P + L)V2 V2 .
At convergence Ln+1 = Ln L, so that L = (B T P + L)(I T V2 V2 ) B T P , therefore B T P + L = (B T P + L)C + C. (20)
This guarantees that there exists a solution F to (12) given by F = KC + = R 1 (B T P + L)C + = KV 1 (S0 )1 U T . Remark. Note that the proof guarantees that, if the algorithm converges, then there exists a solution F to (12).
Then the solution is given by F = R 1 (B T P + L)C + = R 1 (B T P + L)V1 (S0 )1 U T .
73
1601
4. H output-feedback design for F-16 regulator In this example we demonstrate the effectiveness of the proposed static H OPFB design technique on a realistic design example. It is desired to design a lateraldirectional (e.g., roll damper/yaw damper) CAS for the F-16 dynamics from Stevens and Lewis (2003) linearized at the nominal ight condition in Table 3.6-3 (Stevens & Lewis, 2003) (VT = 502 ft/s, 300 psf dynamic pressure, cg at 0.35c). This CAS shown in Fig. 2 has two input channels, and includes actuators and a washout lter. GW is the washout circuit, the transfer function Ga represents an equivalent transfer function for differential actuation of the left and right ailerons, and Gr is the rudder actuator. The design requires the selection of eight control gains. The structured nature of the CAS means that OPFB, not SVFB, must be used. The lateral states are sideslip , bank angle
, roll p, and yaw rate r. Additional states a and r are introduced by the aileron and rudder actuators a = (20.2/(s + 20.2))ua , r = (20.2/(s + 20.2))ur . A washout lter rw = (s/(s + 1))r is used, with r the yaw rate and rw the washed out yaw rate. The washout lter state is denoted xw . The entire state vector is x=[
p r
a r
given by Eq. (1), with

0.3220 0.0640 0.0364 0.9917 0.003 0.0008 0 0 0 1 0.0037 0 0 0 0 3.6784 0.6646 0.7333 0.1315 0 30.6492 8.5396 A= 0 0.0254 0.4764 0.0319 0.0620 0 , 0 0 0 0 20.2 0 0 0 0 0 0 0 20.2 0 0 0 0 57.2958 0 0 1
0 0 0 B = 0 20.2 0 0 The output is
0 0 0 0 , 0 20.2 0
0 D = 0 0
0 0 0
0 0 0
0 0 0
1 0 0
0 1 0
0 T 0 . 1
x w ]T .
0 rw p 0 y= = 57.2958
0 0 0 57.2958
0 57.2958 0 0
57.2958 0 0 0
0 0 0 0
0 0 0 0
1 0 x. 0 0
The control inputs are the rudder and aileron servo inputs so that u = [ua ur ]T . The disturbance is given by d(t) = [da (t) dr (t) dn (t)]T , where da (t) affects the aileron actuation, dr (t) the rudder actuation, and dn (t) the washout lter state. Disturbances are applied to the control inputs, e.g. wind gusts, and to the measurements, i.e. measurement noise. The full-state variable model of the aircraft plus actuators, washout lter, disturbance, and control dynamics is of the form
ua + 2x4 OPFB r1 Gain r 2 matrix u + r
Ga
a r Aircraft p
Gr
The factor of 57.2958 is added to convert angles from radians to degrees. The feedback control is output feedback of the form u = F y, so that the F is a 2 4 matrix. That is, eight feedback gains must be selected. In this example (A, C) are observable. For the computation of the H output-feedback gain F it is necessary to select Q, R, and . The proposed algorithm makes it very easy and fast to perform the design for different values of Q, R, . We selected Q = C T C and R = I . If the resulting gain F is not suitable in terms of time responses and closed-loop poles, the elements of Q and R can be changed and the design repeated. These values were found suitable in this example. The gain parameter denes the desired L2 bound. For the initial design, a fairly large is selected. If the algorithm converges, the parameter may be reduced. If is taken too small the algorithm will not converge since the ARE has no positive semidenite solution. After some design repetitions, which were performed very quickly using the algorithm; we found the smallest value of the gain to be =1.499. To compare this with popular design methods, Chen (2000), smallest value of gain both with State-feedback and Dynamic OPFB compensator is = 0.996. H static output-feedback methods have bigger s than state-feedback methods because the static output feedback is a subset of state feedback. The results for the gain are
K= 95.1780 19.7080 56.0021 10.1009 47.1518 9.3495 17.6280 50.5608 0 0 0 0 0.3077 , 0.8825
rw
GW
F=
Fig. 2. Lateral/directional augmentation system.
0.3077 0.8825
0.8230 0.1632
1.6612 0.3440
0.9774 . 0.1763
74
1602

104 102
Fro Norm P
100 10-2 10-4 10-6
Fig. 3. Trajectory of the gain K.
10
15
20
25
iteration
Fig. 6. Trajectory of Frobenius norm Pn+1 Pn log scale.
1.5 1 0.5 Time response L 0 -0.5 -1 -1.5 -2 5 10 15 iteration

Fig. 4. Trajectory of the parameter matrix L.
Dutch-roll states and r 2 1.5 1 0.5 0 -0.5 0 1 2 3 Time Roll mode states and p 1 Time response 4 5 6 7
20
25
14000 12000 10000 8000 6000 4000 2000 0 -2000 5 10 15 iteration 20 25
0.5 0 -0.5 0 1 2 3 Time

Fig. 7. Closed loop response.
Fig. 5. Trajectory of algebraic Riccati equation solution P.
outperforms the H2 optimal OPFB controller when disturbances are introduced. The resulting closed-loop poles are at s = 11.3762 25.4311i, 15.1711, 4.1338, 1.3423 1.0468i, 1.1350. Convergence properties can be observed in Figs. 36. Note that the norm in Fig. 6 was used in the stopping criterion for the algorithm The resulting gains are applied to the system, and step disturbances d(t) are introduced in simulations to verify disturbance rejection properties of the design. The resulting time responses shown in Figs. 7 are very good. The design procedure based on solving two coupled equations is signicantly easier than methods based on solving three coupled equations, e.g. as described in Lewis and Syrmos (1995) and Moerder and Calise (1985). Moreover, the H static OPFB controller generally Acknowledgments The authors would like to acknowledge Carlos de Souza, for his enthusiasm and ongoing insight into the problems of output feedback and controller parameterization. This research was supported by NSF grant ECS-0501451 and ARO grant W91NF-05-1-0314, and the Ministry of Education of the Czech Republic project 1M0567. References
Chen, B. M. (2000). Robust and H Control. Berlin: Springer. Colaneri, P., Geromel, J. C., & Locatelli, A. (1997). Control Theory and design, an RH2 and RH viewpoint. San Diego: Academic Press Interscience.
75
J. Gadewadikar et al. / Automatica 43 (2007) 1597 1604 Crusius, C. A. R., & Trono, A. (1999). Sufcient LMI conditions for output feedback control problems. IEEE Transactions on Automatic Control, 44(5), 10531057. Doyle, J. H., Glover, K., Khargonekar, P., & Francis, B. (1989). State solutions to standard H2 and H control problems. IEEE Transactions on Automatic Control, 34(8), 831847. de Souza, C. E., & Xie, L. (1992). On the discrete-time bounded real lemma with application in the characterization of static state feedback H controllers. Systems & Control Letters, 18(1), 6171. El Ghaoui, L., & Balakrishnan, V. (1994). Synthesis of xed-structure controllers via numerical optimization. In Proceedings of IEEE Conference on Decision and Control, FL, pp. 26782683. Gadewadikar, J., Lewis, F. L., & Abu-Khalaf, M. (2006). Necessary and sufcient conditions H static output-feedback control. Journal of Guidance Dynamics and Control, 29(4), 915920. Geromel, J. C., & Peres, P. L. D. (1985). Decentralised load frequency control. IEE Proceedings, 132(5), 225230. Geromel, J. C., de Souza, C. C., & Skelton, R. E. (1998). Static output feedback controllers: Stability and convexity. IEEE Transactions on Automatic Control, 43(1), 120125. Geromel, J. C., Yamakami, A., & Armentano, V. A. (1989). Structural constrained controllers for discrete-time linear systems. Journal of Optimization Theory and Applications, 61(1), 7394. Henrion, D., Kucera, V., & Molina-Cristobal, A. (2005). Optimizing simultaneously over the numerator and denominator polynomials in the YoulaKucera Parametrization. IEEE Transactions on Automatic Control, 50(9), 13691374. Iwasaki, T., & Skelton, R. E. (1995). All xed order H controllers: observer based structure and covariance bounds. IEEE Transactions on Automatic Control, 40(3), 512516. Knobloch, H. W., Isidori, A., & Flockerzi, D. (1993). Topics in control theory. Berlin, Germany: Birkhauser. Kucera, V. (1975). Stability of discrete linear feedback systems. Preprints 6th IFAC Congress, Vol. 1, paper 44.1, Boston. Kucera, V. (1979). Discrete linear control: The polynomial equation approach. Chichester: Wiley. Kucera, V., & de Souza, C. E. (1995). A necessary and sufcient condition for output feedback stabilizability. Automatica, 31(9), 13571359. Lewis, F. L., & Syrmos, V. L. (1995). Optimal control. 2nd ed., New York: Wiley-Interscience. Moerder, D. D., & Calise, A. J. (1985). Convergence of a Numerical Algorithm for calculating optimal output feedback gains. IEEE Transactions on Automatic Control, 30(9), 900903. Scherer, C. W. (1992). H -optimization without assumptions on nite or innite zeros. SIAM Journal on Control Optimization, 30, 143166. Scherer, C. W. (1995). Multiobjective H2 /H Control. IEEE Transactions on Automatic Control, 40(6), 10541062. Shaked, U. (2003). An LPD approach to robust H2 and H static output-feedback design. IEEE Transactions on Automatic Control, 48(5), 866872. Stevens, B. L., & Lewis, F. L. (2003). Aircraft control and simulation. 2nd ed., New York: Wiley Interscience. Syrmos, V. L., Abdallah, C. T., Dorato, P., & Grigoriadis, K. (1997). Static output feedbacka survey. Automatica, 33(2), 125137. Trono-Neto, A., & Kucera, V. (1993). Stabilization via static output feedback. IEEE Transactions on Automatic Control, 38(5), 764765. Van der schaft, A. J. (1992). L2 gain analysis of nonlinear systems and nonlinear state feedback H control. IEEE Transactions on Automatic Control, 37(6), 770784. Wonham, W. M. (1985). Linear multivariable control. Berlin, Germany: Springer. Yasuda, K., Skelton, R., & Grigoriadis, K. (1993). Covariance controllers: a new parameterization of the class of all stabilizing controllers. Automatica, 29, 785788.
1603
Youla, D. C., Jabr, H. A., & Bongiorno, J. J. (1976). Modern WienerHopf design of optimal controllers, Part II: The multivariable case. IEEE Transactions on Automatic Control, AC-21, 319338. Yu, J.-T. (2004). A convergent algorithm for computing stabilizing static output-feedback gains. IEEE Transactions on Automatic Control, 49(12), 22712275. Zhou, K. (1992). On the parameterization of H controllers. IEEE Transactions on Automatic Control, 37(9), 14421445. Jyotirmay Gadewadikar received his Bachelors degree in Electronics and Instrumentation Engineering from SGS Institute of Technology and Science, Indore, India in 1997. He then worked for Tata Steel as a Senior Ofcer and was involved with installation and commissioning in Cold Rolling Mill Project. He received the Masters of Science in Electrical Engineering in 2003. He then joined Automation and Robotics Research Institute (ARRI) for Systems & Controls Research, Prototyping & Implementation of Control Algorithms on Electromechanical Systems, taught Control System Design Capstone course, and was involved in the set up and development of undergraduate controls lab, he was also at Qualcomm San Diego as an Intern. His research work includes H control, output feedback control, helicopter UAV control, electromechanical systems, and Automated Testing and Measurement tools for optimization of Wireless Telecommunication Networks. F.L. Lewis, Fellow IEEE, Fellow UK Institute of Measurement & Control, PE Texas, UK Chartered Engineer, is a Distinguished Scholar Professor and Moncrief-ODonnell Chair at University of Texas at Arlingtons Automation & Robotics Research Institute. He obtained the Bachelors Degree in Physics/EE and the MSEE at Rice University, the M.S. in Aeronautical Engineering from Univ. W. Florida, and the Ph.D. at Ga. Tech. He works in feedback control, intelligent systems, and sensor networks. He is author of ve US patents, 174 journal papers, 286 conference papers, and 11 books. He received the Fulbright Research Award, NSF Research Initiation Grant, and ASEE Terman Award. Received Outstanding Service Award from Dallas IEEE section, selected as Engineer of the year by Ft. Worth IEEE Section. Listed in Ft. Worth Business Press Top 200 Leaders in Manufacturing. Lihua Xie received the B.E. and M.E. degrees in electrical engineering from Nanjing University of Science and Technology in 1983 and 1986, respectively, and the Ph.D. degree in electrical engineering from the University of Newcastle, Australia, in 1992. Dr. Xie is currently a Professor with the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore. He held teaching appointments in the Department of Automatic Control, Nanjing University of Science and Technology from 1986 to 1989. His current research interests include robust control and estimation, networked control systems, time delay systems, control of disk drive systems and sensor networks. In these areas, he has published several papers and co-authored (with C. Du) the monograph H-innity Control and Filtering of Two-dimensional Systems (Springer, 2002). Dr. Xie is currently an associate editor of IEEE Transactions on Automatic Control, Automatica, IEEE Transactions on Circuits and SystemsII, and Journal of Control Theory and Applications. He is also a member of the Editorial Board of IEE Proceedings on Control Theory and Applications. He served as an associate editor of International Journal of Control, Automation and Systems from 2004 to 2006 and the Conference Editorial Board, IEEE Control Systems Society from 2000 to 2005. He is a Fellow of IEEE and a Fellow of Institution of Engineers, Singapore.
76
1604
J. Gadewadikar et al. / Automatica 43 (2007) 1597 1604 of the Czech Technical University in 2006, and French Government Medal Chevalier dans lordre des Palmes Acadmiques in 2006. He is an Honorary Professor at the Northeastern University, Shenyang, China (1996) and received Doctor honoris causa degrees from Universit Paul Sabatier, Toulouse (2003) and Universit Henri Poincar, Nancy (2005).
Vladimr Ku era was born in Prague, Czech c Republic, in 1943. He graduated in Electrical Engineering from the Czech Technical University in Prague in 1966, and received the CSc. and DrSc. degrees in Control Engineering from the Czechoslovak Academy of Sciences in 1970 and 1979, respectively. The research interests of V. Ku era include linc ear systems, optimal and robust control. He has contributed to the theory of the Riccati equation, pioneered the use of polynomial equations in control system design, and paved the way to the parametrization of all stabilizing controllers. From 1970 to 1990 he was a Research Scientist, and from 1990 to 1998 the Director of the Institute of Information Theory and Automation, Academy of Sciences of the Czech Republic. Since 1995 he has been Professor of Control Engineering, and from 2000 to 2006 the Dean of the Faculty of Electrical Engineering, Czech Technical University in Prague. V. Ku era is the author of four books, including Discrete Linear Control: c The Polynomial Equation Approach (Wiley, 1979) and Analysis and Design of Discrete Linear Control Systems (Prentice-Hall, 1991), and published 300 research papers. He is Past President of IFAC, Fellow of IEEE, and Fellow of the Czech Academy of Engineering. He was awarded the National Prize of the Czech Republic in 1989, Automatica Best Paper Award in 1990, Medal of the Ministry of Education of the Czech Republic in 2000, Felber Gold Medal
Murad Abu-Khalaf was born in Jerusalem, Palestine in 1977. He obtained his B.S. in Electronics and Electrical Engineering from Bo azii University in Istanbul, Turkey in g 1998, and the M.S. and Ph.D. in Electrical Engineering from The University of Texas at Arlington in 2000 and 2005, respectively. His research interest is in the areas of nonlinear control, optimal control, neural network control, and adaptive intelligent systems. He is the author/co-author of one book, two book chapters, eight journals papers and 15 refereed conference proceedings. He is a member of IEEE member, and a member of Etta Kappa Nu honor society, and is listed in Whos Who in America. His research interests are in the areas of nonlinear control, optimal control, neural network control, adaptive intelligent systems.
77
Information Sciences 177 (2007) 30053015 www.elsevier.com/locate/ins
Static output feedback controller design for fuzzy systems: An ILMI approach
Dan Huang, Sing Kiong Nguang
*
The Department of Electrical and Computer Engineering, The University of Auckland, Private Bag 92019 Auckland, New Zealand Received 15 July 2004; received in revised form 4 February 2007; accepted 11 February 2007
Abstract This paper examines the problem of static output feedback control of a TakagiSugeno (TS) fuzzy system. The existence of a static output feedback control law is given in terms of the solvability of bilinear matrix inequalities. An iterative algorithm based on the linear matrix inequality is proposed to compute the static output feedback gain. To reduce the conservatism of the design, the structural information of the membership function of the fuzzy rules is incorporated. Numerical examples are used to illustrate the validity of the methods. 2007 Elsevier Inc. All rights reserved.
Keywords: Static output feedback control; TakagiSugeno fuzzy model; Linear matrix inequality
1. Introduction Over the past two decades, static output feedback problem has attracted considerable attentions of many researchers [24,11,12]. The problem can be stated as follows: given a system, nd a static output feedback so that the closed-loop system is stable. Normally, the existence of a full order output feedback control law is given in terms of the solvability of two convex problems. However, the synthesis of a static output feedback gain or a xed order controller is much more dicult. The main rationale is that the separation principle does not hold in such cases. A comprehensive survey on static output feedback can be found in [12]. The static output feedback problem is important in its own right, because static controllers are less expensive to be implemented and more reliable in practice. In [12] the authors show that any dynamic output feedback problem can be transformed into a static output feedback problem. Hence, the static output feedback formulation is more general than the full order dynamic output feedback formulation, that is, the static output formulation can be used to design a full order dynamic controller, but the converse is not true.
Corresponding author. E-mail address: sk.nguang@auckland.ac.nz (S.K. Nguang).
0020-0255/$ - see front matter 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.ins.2007.02.014
78
3006
D. Huang, S.K. Nguang / Information Sciences 177 (2007) 30053015
Recently, a great amount of eort has been devoted to describing a nonlinear system using a Takagi Sugeno fuzzy model; see [1,510,1317,20]. In this fuzzy model, local dynamics in dierent state space regions are represented by local linear systems. The overall model of the nonlinear system is obtained by blending of these linear models through nonlinear fuzzy membership functions. Unlike conventional modelling techniques which use a single model to describe the global behavior of a nonlinear system, fuzzy modelling is essentially a multi-model approach in which simple sub-models (typically linear models) are fuzzily combined to described the global behavior of a nonlinear system. The TS fuzzy model has been proved to be a very good representation for a certain class of nonlinear dynamic systems. Based on this fuzzy model, a number of systematic model-based fuzzy control design methodologies have been developed. The major drawback of the above mentioned papers [1,510,1317,20] is that their design methodologies do not incorporate membership function characteristics, which may lead to conservative design methodologies. Motivated by this drawback and the simplicity of static output feedback controller, in this paper, input membership function characteristics are incorporated into the static output feedback control design. To the best of our knowledge, the problem of static output feedback for TS fuzzy models with membership functions characteristics has not been studied in the literature. We show that the existence of a static output feedback control law can be expressed in terms of the solvability of bilinear matrix inequalities (BMIs). In order to compute a solution to these BMIs, an iterative algorithm based on the linear matrix inequality has been developed. The rest of this paper is organized as follows. System description and notations are given in Section 2. Main results are presented in Section 3. The validity of our approach is demonstrated by examples in Section 4. Finally, conclusions are given in Section 5. 2. System description and notations A fuzzy dynamic model has been proposed by Takagi and Sugeno [13] to represent local linear input/output relations of nonlinear systems. This fuzzy linear model is described by IFTHEN rules and has been shown to be able to approximate a large class of nonlinear systems. As in [14,15,18], we examine a TS fuzzy model, in which the ith rule is formulated as follows: Plant Rule i: IF z1 t is Mi1 and and zp t is Mip, THEN _ xt Ai xt Bi ut; yt C i xt; 2:1
where i 1; . . . ; r, r is the number of fuzzy rules; zk t are premise variables, Mik are fuzzy sets, k 1; . . . ; p; p is the number of premise variables; xt 2 Rn ; ut 2 Rm ; yt 2 Rl denote state, control input, and output, respectively. Matrices Ai 2 Rnn ; Bi 2 Rnm and C i 2 Rln are of appropriate dimensions. Remark 2.1. There are two major ways in TS fuzzy modelling. One is the TS fuzzy model identication [13,16,17] using inputoutput data, and the other is the TS fuzzy model construction, by the idea of sector nonlinearity [14,15,18]. By using center-average defuzzifer, product inference and singleton fuzzifer, the local models can be integrated into a global nonlinear model: _ xt Ahxt Bhut; yt Chxt; where Ah
r X i1
2:2
hi ztAi ;
Bh
r X i1
hi ztBi ;
Ch
r X i1
hi ztC i
79
3007
and zt z1 t; z2 t; . . . ; zp tT ; p Y xi zt M ik zk t; xi zt P 0;
k1 r X i1 r X i1
xi zt > 0;
xi zt hi zt Pr ; i1 xi zt
hi zt P 0;
hi zt 1:
Here, M ik zk t denote the grade of membership of zk t in Mik. For the nonlinear plant represented by (2.2), the fuzzy output feedback controller is inferred as ut Khyt; 2:3 Pr where Kh i1 hi ztK i . Ki in each plant rule is a local controller gain to be designed. Applying the fuzzy controller (2.3) to the global fuzzy plant (2.2), we have the following closed-loop system: _ xt Ah BhKhChxt: 2:4
For the sake of notational convenience, xt, ut, and hi zt are denoted as x, u, and hi, respectively. Note that in the symmetric block matrices, we use (*) as an ellipsis for terms that are induced by symmetry. Before ending this section, we recall the S-procedure which will be used in the proof. Denote the set G fgg and let Fg; Y1 g; . . . ; Ym g be some functional or functions. Further dene domain H: H fg 2 G : Y1 g 6 0; . . . ; Ym g 6 0g: The S-procedure [19] states that Fg < 0 8g 2 H, if exist k1 P 0; . . . ; km P 0 such that m X kk Yk g < 0: Fg
k1
3. Main results In this section, we shall present our procedure for designing a static output feedback control gain for the system (2.4). A static output feedback design methodology without input membership function characteristics is presented in Section 3.1. Section 3.2 provides a static output feedback design methodology with input membership function characteristics. 3.1. Controller design without membership function characteristics Before presenting the rst result, the following Lemma is derived. Lemma 3.1. The system (2.4) is asymptotically stable by means of a static output feedback if there exist symmetric matrix P > 0 and matrices Kis satisfy the following conditions: " # T AT P PAi PBi BT P Y ii i i < 0 i 1; . . . ; r; 3:1 BT P K i C i I i 1 0 T 3 2 Ai P PAi C B 7 6 T C B 7 6 T7 6 1 B PBi Bj P C Y 7 C 6 2B ij T B A P PAj C 7 < 0 i < j; 6 3:2 j A @ 7 6 7 6 T 7 6 PBj Bi P 5 4 T 1 Bi P K i C j BT P K j C i I j 2
80
3008
Y 11
Y 12 Y 22
.. .
Y 1r
1 3:3
BY B 12 B . B . @ . Y 1r
Y 2r C C . C > 0: . C . A Y rr
Y 2r
Proof. Let us choose Lyapunov function candidate as V x xT Px. The time derivative of V x along the system (2.4) is _ _ _ V x xT Px xT P x xT fAh BhKhChgT Px xT P fAh BhKhChgx 6 xT fAh BhKhCh P P Ah BhKhCh Ch Kh KhChgx xT fAh P P Ah P BhBh P Bh P KhCh Bh P KhChgx: From here, we can see that as long as W Ah P P Ah P BhBh P Bh P KhCh Bh P KhCh < 0; the fuzzy system (2.4) is stable. Using the Schur complement, W is equivalent to the following quadratic matrix equality: " # T T T Ah P P Ah P BhBh P W BhT P KhCh I " # T r r X X AT P PAi PBi BT P i j hi hj BT P K i C j I i1 i1 i " # T r T T X A P PAi PBi Bi P h2 i i BT P K i C i I i1 i " # T 1 r X AT P PAi PBi BT P AT P PAj PBj BT P j i i j 2 hi hj 2 1 BT P K i C j BT P K j C i I i<j i j 2 r r X X 6 h2 Y ii 2 hi hj Y ij i
i1 i<j T T T T T T T T T T T T T
3:4
3:5
3:6
h1
1T 0
Y 11
Y 12 Y 22
Y 1r
10
h1
1 3:7
Bh C B 2C B . C B . C @ . A hr
BY B 12 B . B . @ . Y 1r
Y 2r CB h2 C CB C . CB . C: .. . CB . C . A@ . A . Y rr hr h
Y 2r
Thus, if (3.3) holds, (3.7) is less than zero, i.e., W < 0, and the fuzzy system (2.4) is stable.
Note that the matrix inequalities (3.1) and (3.2) are bilinear matrix inequalities (BMI) and cannot be solved with ease by a convex optimization algorithm. An iterative algorithm (ILMI) based on the linear matrix inequality (LMI) has been employed to solve this BMI problem in [2,5]. To apply an ILMI method, we need to accommodate the negative quadratic term PBi BT P by introducing i T T an additional design matrix variable X. Using the fact X P BhBh X P P 0 for any X and P of the same dimension, we obtain X T BhBh P P BhBh X X T BhBh X 6 P BhBh P :
T T T T
3:8
Plugging (3.8) into (3.6) and following the same procedure of the proof afterwards, we obtain the following theorem:
81
3009
Theorem 3.1. The fuzzy system (2.4) is asymptotically stable by means of a static output feedback if there exist symmetric matrix P > 0 and matrices Kis and a matrix X satisfy the following conditions: " # Gii T < 0; i 1; . . . ; r; 3:9 T ii I " # T Gij < 0; i < j; 3:10 T ij I 1 0 Y 11 Y 12 Y 1r C B B Y 12 Y 22 Y 2r C C B C B 3:11 B . . C > 0; .. . C B . . A . . @ Y 1r Y 2r Y rr where Gii AT P PAi X T Bi BT P PBi BT X X T Bi BT X Y ii ; i i i i 1 Gij AT P PAi X T Bi BT P PBi BT X X T Bi BT X j j j 2 i AT P PAj X T Bj BT P PBj BT X X T Bj BT X Y ij ; j i i i T ii BT P K i C i ; i 1 T ij BT P K i C j BT P K j C i : j 2 i Remark 3.1. If the auxiliary variable X is xed, the BMI (3.9) and (3.10) reduce to LMIs in P and Kis and can be solved eciently. However, in general, xing the auxiliary variable X yields no solution to the LMIs. Thus, to relax those LMIs and make them feasible, a new term aP is introduced in (3.9) and (3.10) as follows: " " Gii aP T ii Gij aP T ij
T
# <0 # <0 i < j: 3:13 i 1; . . . ; r; 3:12
I
T
The ILMI algorithm [2], as will be shown later, repeats the search for P and Kis and updates the auxiliary variable X by decreasing a until a becomes negative. If a < 0 is reached, then the BMIs (3.9) and (3.10) have a solution. Iterative linear matrix inequality (ILMI) algorithm Step 1: Solve the following LMIs that is jointly convex in Q; F : Q QT > 0; " QAT Ai Q F T Bi BT Bi BT F i i i BT F i
T
# <0 i 1; . . . ; r:
3:14 3:15
Set t 1 and X t FQ1 and choose a0 to be a big enough number. Step 2: Solve the following optimization problem in P and Kis using the auxiliary variable Xt determined in the previous step to obtain at:
82
3010
Minimize Subject to
at P > 0; " Gii at P T ii " Gij at P T ij e Y > 0:
# <0 # <0 i < j; i 1; . . . ; r;
T I
Step 3: If at < 0, P and Kis obtained in Step 2 are a feasible solution to the BMIs and stop. Step 4: Solve the following optimization problem in P and Kis using at and Xt determined in Step 2: Minimize Subject to traceP P > 0; " # T Gii at P < 0 i 1; . . . ; r; T ii I " # Gij at P T < 0 i < j; T ij I e Y > 0: Step 5: If kX t P k < d, go to Step 6, where d is a predetermined small value. Else set t t 1 and X t P , then go to Step 2. Step 6: The system may not have a feasible solution and stop. Remark 3.2 1. Note that (3.9) and (3.11) imply H ii AT P PAi X T Bi BT P PBi BT X X T Bi BT X < 0; i i i i
1
3:16
which provides an initial guess for X. By premultiplying and postmultiplying (3.16) with Q P and dening F XQ, we can see that H ii < 0 is equivalent to LMI (3.15) by the Schur complement. What Step 1 does is to seek a X corresponds to the necessary condition of (3.9) and uses it as an initial guess for the iterative algorithm. 2. aP is introduced in (3.9) and (3.10) to relax the LMIs. (3.12) and (3.13) correspond to the following Lyapunov inequality: V x xT Px; _ V x 6 aV x:
If a is negative, the BMI (3.10) is feasible and the fuzzy system (2.4) is stable. 3. The optimization problem in Step 2 is a generalized eigenvalue minimization problem. This step guarantees the progressive reduction of at. Step 4 is to guarantee the convergence of the algorithm. 4. Sometimes, at+1 can be greater than at in Step 2 which may be due to the implementation problem of the LMI program. In this situation, we must set at1 at . Owing to the eect of some numerical errors in Step 4, the optimization problem in Step 4 may be infeasible. In such case, let at at Dat , where D is a small positive real number, and go to Step 4 to solve the optimization problem again. 3.2. Controller design with membership function characteristics In the previous result, the membership function characteristics have been neglected. This information is crucial in many cases and may render less conservative stability conditions. Based on the so-called (outer)
83
3011
ellipsoidal approximation algorithm, a new result is derived in this subsection by incorporating the membership function characteristics. Before proceeding with the development of the controller design technique, the following denition is needed. Denition 3.1. Dene Rij as the region where the fuzzy rule i and fuzzy rule j are activated: Rij fxjhi xhj x > 0g: This paper assumes that each region Rij i; j 62 I n ) can be outer approximated by a union of ellipsoids Eijk for k 1; . . . ; m; where m is the number of ellipsoids. This assumption implies that there exist matrices Eijk ; f ijk such that Rij
m [ k1
Eijk ;
where Eijk fxjkEijk x fijk k 6 1g:
3:17
Notice that each Eijk can also be represented as LMI forms: # T " T Eijk Eijk T x x Eijk 6 0: T T 1 fijk Eijk 1 fijk fijk 1
3:18
Remark 3.3. Suppose that Rij fxjd 1 6 cT x 6 d 2 g, then it is easy to see that we can take Eij1 2cT =d 2 d 1 and fij1 d 2 d 1 =d 2 d 1 . Using this denition and Theorem 3.1, we have the following theorem. Theorem 3.2. The fuzzy system (2.4) is asymptotically stable by means of a static output feedback if for i 1; . . . ; r, i < j 6 r, and k 1; . . . ; m, there exist symmetric matrix P > 0 and matrices Kis and scalars kijk P 0, and matrix X satisfy the following conditions: 9 " # T > Gii > > < 0; > > = T ii I " # when 0 2 Eijk ; 3:19 > > Gij T > > <0 > ; T ij I 9 2 3 T T Gii kiik S iik > > > 6 > T 7 4 > T ii I 5 < 0 > > > > = kiik Y iik 0 kiik Z iik when 0 62 Eijk ; 3:20 2 3 T T > Gij kijk S ijk > > > 6 > T 7 4 > T ij I 5 < 0 > > > ; kijk Y ijk 0 kijk Z ijk 1 0 Y 11 Y 12 Y 1r C BY B 12 Y 22 Y 2r C B . 3:21 . C > 0; .. B . . C @ . . A . Y 1r where S ijk ET Eijk ; ijk
T Y ijk fijk Eijk ; T Z ijk 1 fijk fijk :
Y 2r
Y rr
84
3012
Proof. Following from the proof of Lemma 3.1, it can be shown that r r X X _ V xt 6 h2 xT tGii T T T ii Y ii xt 2 hi hj xT tGij T T T ij Y ij xt: i ii ij
i1 i<j
Notice that in Theorem 3.1, conditions (3.9) and (3.10), respectively, imply that xT tGii T T T ii xt < 0; ii and xT tGij T T T ij xt < 0; ij xT tGii T T T ii xt < 0; ii and xT tGij T T T ij xt < 0; ij xt 6 0; xt 2 Rij : xt 6 0; xt 2 Rn : xt 6 0; xt 2 Rn
These requirements may lead to conservatism. In actually fact, we only need xt 6 0; xt 2 Rii
One way to introduce this condition into the stability conditions is through the use of S-procedure. When T 0 2 Eijk , that is, Eijk contains the origin, the term 1 fijk fiik becomes positive which implies that for the LMIs to hold, kijk must be zero; see (3.20). However, the LMIs are no longer strictly feasible. Hence, to avoid nonstrictly feasible, when 0 2 Eijk , the membership function characteristics are not incorporated. Therefore, we obtained (3.19) which is the same as in Theorem 3.1. When 0 62 Eijk , that is, Eijk does not contain the origin, by the use of the S-procedure, we have xT tGii T T T ii xt kiik Eiik < 0; ii and xT tGij T T T ij xt kijk Eijk < 0; ij k 1; . . . ; m: 3:23 k 1; . . . ; m 3:22
By Schur complements, we have conditions given in (3.20). Notice that conditions given in (3.19) imply that the origin is globally stable. h Remark 3.4. Clearly from (3.22) and (3.23), when kijk 7!0, the conditions given in Theorem 3.2 become the same as in Theorem 3.1. Therefore, if there exists a solution to Theorem 3.1, there always exists a solution to Theorem 3.2, but the conserve is not true. Hence, Theorem 3.2 is less conservative than Theorem 3.1. The ILMI algorithm can be applied to Theorem 3.2. 4. Numerical examples In this section, two design examples and their computer simulation are provided to illustrate the validity of the designs obtained in the previous section. Example 1 (Lorenz chaotic system). To design a fuzzy static output feedback controller, the Lorenz chaotic system needs to be represented by a TS fuzzy model. An exact TS fuzzy modelling [15] is employed to construct a TS fuzzy model for the Lorenz chaotic system. The method utilises the concept of sector nonlinearity. For more details, see [18,14]. The following Lorenz chaotic system with the input term will be consider in the sequel: _ x1 t ax1 t ax2 t ut; _ x2 t cx1 t x2 t x1 tx3 t; _ x3 t x1 tx2 t bx3 t; 4:1
where a 10; b 8=3; c 28, x1 t; x2 t and x3 are the state variables, and ut is the control input. Assume that x1 t 2 NN , the nonlinear terms x1 tx3 t and x1 tx2 t can be expressed as
85
3013
x1 x3 t h1 x1 tNx3 t h2 x1 tNx3 t and x1 x2 t h1 x1 tNx2 t h2 x1 tNx2 t; where h1 x1 t x1 t N 2N and h2 x1 t x1 t N : 2N
Then, using the above membership functions, we can have the following TS fuzzy model which exactly represents (4.1) under the assumption on bounds of the state variable x1 t 2 NN . _ xt yt
2 X i1 2 X i1
hi ztAi xt hi ztC i xt;

T
2 X i1
hi ztBi ut; 4:2
where xt x1 tx2 tx3 t , 2 3 a a 0 6 7 A1 4 c 1 N 5; 0 N b
6 A2 4 c 0
a 1 N
7 N 5; 8=3
2 3 1 6 7 B1 B2 4 0 5 ; 0
C1 C2 1
0 :
The TS fuzzy model (4.2) exactly represents (4.1) under the assumption on bounds of the state variable x1 t 2 NN where N > 0. However, this assumption is not strict because of two reasons. It is well know that state variables of the chaotic system are bounded. In addition, N can be set to any value. Even if the nonlinear equations of the Lorenz chaotic system are unknown, recently developed fuzzy modelling techniques [13,16,17] using observed data can be utilised to obtain fuzzy models. Through the simulations of (4.1), we learn that x1 t 2 30; 30, hence N is set be 30. Using Theorem 3.1, we obtain the static output feedback gains as K 1 1220, K 2 1278:3, and K 3 1751:7. Simulation result is shown in Fig. 1 for the initial condition x1 ; x2 ; x3 10; 10; 10. Example 2 (Nonlinear mass-spring-damper system). Consider a nonlinear mass-spring-damper mechanical system with a nonlinear spring:
10 x (t)
1
x (t)
2
x3(t)
States
5 10 15 0
10
Time (sec)
Fig. 1. Lorenzs chaotic system.
86
3014
a
Degree of membership
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 0.5 0 0.5 1 1.5
0.4
x (t)
1
0.3 0.2
x (t)
2
The states
h1(x2(t))
0.1 0 0.1 0.2 0.3 0.4 0.5 0 10 20 30 40 50 60 70 80 90 100
h (x (t))
2 2
X2(t)
Time (sec)
Membership function
Simulation result
Fig. 2. Nonlinear mass-spring-damper system: (a) membership function; (b) simulation result.
x_1 t 0:1125x1 t 0:02x2 t 0:67x3 t ut; 2 x_2 t x1 t; yt x2 t;
4:3
_ where x2 t is the springs displacement and x1 t x2 t. The term 0:67x3 is due to the nonlinearity of the 2 spring. The spring is attached to a wall, therefore the springs displacement x2 t is physically constrained by the length of the spring and the wall. The length of the spring could be any value, in this paper, we assume x2 t 2 1; 1:5. The lower limit is due to the minimum length that the spring can be compressed. Same as Example 1, the concept of sector nonlinearity [15] is employed to construct an exact TS fuzzy model for the mass-spring-damper system. Using the fact that x2 t 2 1; 1:5, this nonlinear term can be expressed as 0:67x3 t h1 x2 t0x2 t h2 x2 t1:5075x2 t; 2 where h1 x2 t 1 x t and h2 x2 t x t; see Fig. 2a. 2:25 2:25 Using h1 x2 t and h2 x2 t, we obtain the following TS fuzzy model which exactly represents (4.3) under the assumption on bounds of the state variable x2 t 2 11:5: _ xt yt
2 X i1 2 X i1
2 2
hi ztAi xt hi ztC i xt;
2 X i1
hi ztBi ut; 4:4
where xt x1 tx2 tT , 0:1125 0:02 A1 ; 1 0
A2
0:1125 1:5275 1 0
; B1 B2
1 0
C1 C2 0
1 :
S S From the membership functions given in Fig. 2a, we have R11 1 E11k , R12 1 E13k , and R22 k1 k1 S1 k1 E33k , where E111 E121 E221 x2 t 1:5x2 t 1 6 0. In the form given in (3.18), we obtain T 111 T 121 T 221 00:8; f 111 f121 f221 0:2: Applying Theorem 3.2, a solution to the BMIs given in Theorem 3.2 is found with a 0:0015 to be 3:3633 0:36961 P 0:36961 2:6421 and the static output feedback gains K 1 0:75351; K 2 0:75348:
87
3015
Fig. 2b depicts the simulation result with the initial condition x0 0:1; 0:5. One can see that the closed system is asymptotically stable. 5. Conclusion Static output feedback control designs for a TS fuzzy system have been provided in this paper. The existence of a static output feedback control law has been expressed in terms of the solvability of bilinear matrix inequalities. To compute a solution to the BMIs, an iterative algorithm based on the linear matrix inequality has been proposed. To reduce the conservatism of the design, the structural information of the rule base has been incorporated. Numerical examples have been provided to illustrate the validity of our design. References
[1] W. Assawinchaichote, S.K. Nguang, P. Shi, Robust H1 fuzzy lter design for uncertain nonlinear singularly perturbed systems with Markovian jumps: an LMI approach, Information Sciences 177 (7) (2007) 16991714. [2] Y.Y. Cao, J. Lam, Y.X. Sun, Static output feedback stabiliztion: an ILMI approach, Automatica 34 (12) (1998) 16411645. [3] W. Chang, J.B. Park, Y.H. Joo, G.R. Chen, Static output-feedback fuzzy controller for Chens chaotic system with uncertainties, Information Sciences 151 (2003) 227244. [4] I.N. Kar, Design of static output feedback controller for uncertain systems, Automatica 35 (1) (1999) 169175. [5] E. Kim, H. Lee, New approaches to relaxed quadratic stability condition of fuzzy control systems, IEEE Transactions on Fuzzy Systems 8 (5) (2000) 523534. [6] J. Li, H.O. Wang, D. Niemann, K. Tanaka, Dynamic parallel distributed compensation for TakagiSugeno fuzzy systems: An LMI approach, Information Sciences 123 (34) (2000) 201221. [7] S.K. Nguang, W. Assawinchaichote, H1 ltering for fuzzy dynamical systems with pole placement constraints, IEEE Transactions on Circuits and Systems I 50 (11) (2003) 15031508. [8] S.K. Nguang, P. Shi, Robust H1 output feedback control design for fuzzy dynamic systems with quadratic D stability constraints: an LMI approach, Information Sciences 176 (15) (2006) 21612191. [9] S.K. Nguang, P. Shi, H1 fuzzy output feedback control design for nonlinear systems: an LMI approach, in: Proc. 40th IEEE Conf. on Decision and Control, Orlando, USA, 2001, pp. 25012506. [10] S.K. Nguang, P. Shi, Stabilisation of a class of nonlinear time-delay systems using fuzzy models, in: Proc. 39th IEEE Conf. on Decision and Control, Sydney, Australia, 2000, pp. 44154419. [11] E. Prempain, I. Postlethwaite, Static output feedback stabilisation with H1 performance for a class of plants, Systems and Control Letter 43 (6) (2001) 159166. [12] V.L. Syrmos, C.T. Abdallah, P. Dorato, K. Grigoriadis, Static output feedback: a survey, Automatica 33 (2) (1997) 125137. [13] T. Takagi, M. Sugeno, Fuzzy identication of systems and its applications to medeling and control, IEEE Transactions of System, Man, and Cybernetics 15 (1) (1985) 116132. [14] K. Tanaka, T. Ikeda, H.O. Wang, Fuzzy regulators and fuzzy observers: Relaxed stability conditions and LMI-based design, IEEE Transactions on Fuzzy Systems 6 (2) (1998) 250265. [15] T. Taniguchi, K. Tanaka, H. Ohtake, H.O. Wang, Model construction, rule reduction, and robust compensation for generalized form of TakagiSugeno fuzzy ystems, IEEE Transactions on Fuzzy Systems 9 (4) (2001) 525538. [16] L.X. Wang, A Course in Fuzzy Systems and Control, Prentice-Hall, Englewood Clis, NJ, 1997. [17] L. Wang, R. Langari, Building sugeno-type models using fuzzy discretization and orthogonal parameter estimation techniques, IEEE Transactions on Fuzzy Systems 3 (4) (1995) 454458. [18] H.O. Wang, K. Tanaka, M.F. Grin, An approach to fuzzy control of nonlinear systems: stability and design issues, IEEE Transactions on Fuzzy Systems 4 (1) (1996) 1423. [19] V.A. Yakubovich, S-procedure in nonlinear control theory, Vestnik Leningradskogo Universiteta, Ser. Matematika (1971) 6277. [20] S.S. Zhou, G. Feng, J. Lam, S.Y. Xu, Robust H1 control for discrete-time fuzzy systems via basis-dependent Lyapunov functions, Information Sciences 174 (34) (2005) 197217.
88
89
SIAM J. CONTROL OPTIM. Vol. 39, No. 6, pp. 17111735
c 2001 Society for Industrial and Applied Mathematics
AN LMI-BASED ALGORITHM FOR DESIGNING SUBOPTIMAL STATIC H2 /H OUTPUT FEEDBACK CONTROLLERS

F. LEIBFRITZ Abstract. We consider the problem of designing a suboptimal H2 /H feedback control law for a linear time-invariant control system when a complete set of state variables is not available. This problem can be necessarily restated as a nonconvex optimization problem with a bilinear, multiobjective functional under suitably chosen linear matrix inequality (LMI) constraints. To solve such a problem, we propose an LMI-based procedure which is a sequential linearization programming approach. The properties and the convergence of the algorithm are discussed in detail. Finally, several numerical examples for static H2 /H output feedback problems demonstrate the applicability of the considered algorithm and also verify the theoretical results numerically. Key words. static controllers, output feedback, suboptimal control, robust control, linear systems, linear matrix inequalities, nonconvex programming AMS subject classications. 90C22, 90C26, 93A99, 93B36, 93B51, 93B52, 93C05, 93D09, 49N99, 65K05 PII. S0363012999349553
1. Introduction. The static or reduced xed order dynamic output feedback control problem that meets desired performance and/or robustness specications is an active research area of the control community. In this paper we consider the computational design of static H2 /H output feedback controllers. This is an important example of a nonconvex control problem. It consists of determining a static output feedback gain which achieves a certain nominal (suboptimal) performance measure subject to a robustness constraint. The static output feedback problems are important, since it is not always possible to have full access to the state vector and a controller must be used which is based only on the available observations. Moreover, they are important because other problems are reducible to some variations of the static output feedback problem and relevant when a simple controller must be used due to cost and reliability. During the past decade, control problems with combined H2 and H design criteria have gained a great deal of attention. Concerning continuous-time systems, [7] provides the solution of standard H2 and H control problems in terms of algebraic Riccati equations, where both state feedback and full order compensator-based output feedback are considered. The design of feedback controllers that satisfy both H and H2 specications is interesting because it oers robust stability and nominal performance. In 1989, Bernstein and Haddad [2] introduced a mixed H2 /H problem. Their approach is to minimize an auxiliary cost subject to an H norm constraint, and this cost yields an upper bound on the H2 norm. The work of [2] is extended in [46] and [6], where another mixed H2 /H problem is addressed. The system considered therein is dual to the BernsteinHaddad setup (see [45]). Other related works on the design of H2 /H controllers by state or full order output feedback can be found, for example, in [23], [33], [35], [38], and [40]. Only [44] considers a mixed H2 /H problem for the static output feedback case. The solvability conditions and
by the editors January 7, 1999; accepted for publication (in revised form) September 5, 2000; published electronically February 28, 2001. http://www.siam.org/journals/sicon/39-6/34955.html FB-IV Mathematics, University of Trier, D-54286 Trier, Germany (leibfr@uni-trier.de). 1711
Received
90
1712
F. LEIBFRITZ
the algorithms discussed in the above literature are based on coupled Riccati and/or Lyapunov equations. Recently, linear matrix inequalities (LMIs) have attained much attention in control engineering [3], [5], [36], since many control problems can be formulated in terms of LMIs and thus solved via convex programming approaches. For example, this includes H [4], [11], [12], [25], H2 [18], [37], and mixed H2 /H [19], [20], [27], [29], [39]. However, the resulting controllers are state feedback or of order nx equal to the plant. The diculties arise if we want to design a static (or reduced xed order) output feedback controller. Then the problem of determining a static output feedback controller including H2 and H can be restated as a linear algebra problem, which involves two coupled LMIs. In this case, the solution of one should be the inverse of the other. The problem is then no longer convex [12], [25], [31], [30], [42], [43], and nding a solution numerically to these nonconvex problems is a dicult task. In this paper we will develop an LMI-based computational procedure for solving a mixed H2 /H problem by a static output feedback controller, which is an extension of the algorithm proposed by Leibfritz [31] for the design of stabilizing static H2 and H output feedback gains. The suboptimal static H2 /H output feedback problem considered in this paper can be necessarily rewritten to a nonconvex, multiobjective programming problem. In particular, this problem consists of minimizing a (nonconvex) functional of the form J (P, Q, Y ) = Tr(P Q) + Tr(Y ) subject to suitably chosen LMI constraints. Then, using the solution of this problem (if any exists), the existence of a static H2 /H output feedback gain F can be tested by solving a suitably chosen LMI feasibility problem in F . Therefore, the problem of nding a suboptimal static H2 /H output feedback gain reduces to an optimization problem with a nonconvex objective over a convex set and an LMI feasibility problem. Similar to Leibfritz [31], we will derive the so-called sequential linear programming matrix method (SLPMM) for solving the resulting nonconvex programming problem. This approach is motivated by successive minimization of a linearization of J (P, Q, Y ) subject to LMI constraints as proposed by [1] for general nonconvex bilinear programming problems. Note that the SLPMM algorithm is closely related to the cone complementarity linearization method developed by [9] for solving the static output feedback stabilization problem. As shown in [31], the theoretical advantages of the SLPMM algorithm over the cone complementarity algorithm are the following. First, the SLPMM algorithm always generates a strictly decreasing sequence of the objective function values which is bounded below by an integer nx , and thus it is convergent. Second, the sequence of iterates generated by the SLPMM is contained in a compact level set, and therefore it is always bounded. Finally, if a corresponding bilinear matrix problem is nonempty, then every accumulation point of the generated sequence solves this bilinear matrix problem. In this case, it is always possible to reconstruct a static output feedback gain from this solution. On the other hand, if there exists no matrix pair satisfying the bilinear matrix problem, then the SLPMM algorithm always terminates with an objective function value of a bilinear matrix inequality minimization problem which is greater than nx . This indicates for the considered plant that there exists no static output feedback controller. But in this case, one can construct a reduced xed order dynamic controller from the computed solution of the bilinear matrix inequality minimization problem. For more details, we refer to Leibfritz [31]. In contrast to these strong theoretical convergence results, the authors in [9] can guarantee only that the sequence of a linear approximation of the objective function values is a monotonically (nonstrict) decreasing sequence, which
91
SUBOPTIMAL STATIC H2 /H OUTPUT FEEDBACK DESIGN
1713
is bounded below by 2nx . Moreover, in the last few years a number of numerical procedures have been proposed for solving static H2 output feedback problems. For example, the LMI-based methods also include the Min-Max algorithm of Geromel, de Souza, and Skelton [16], [17], the XY-centering algorithm of Iwasaki and Skelton [26], and the alternating projection method of Grigoriadis and Skelton [21], but the convergence of these algorithms is not always guaranteed. Particularly, the Min-Max algorithm guarantees only sequences of upper and lower bounds to the maximal and minimal eigenvalues of P Q, which are strictly decreasing, and increasing sequences under strong technical assumptions. In addition, for ensuring the global convergence of the Min-Max algorithm, they must assume that the generated sequences are contained in a compact set. On the other hand, it may occur that the Min-Max algorithm generates an unbounded sequence even if the bilinear matrix problem is nonempty. In this case the Min-Max method breaks down [9], [17]. Since the XY-centering algorithm is closely related to the Min-Max procedure, the global convergence of this approach can be shown only under similar technical assumptions that are as strong as the ones imposed for the Min-Max algorithm [26, Theorem 2]. Finally, the alternating projection method is guaranteed to converge only locally. These observations motivate us to derive the SLPMM algorithm for more complicated problems such as the static H2 /H problem. For this problem class, we will show that the SLPMM procedure behaves theoretically as well as numerically in a similar way to that described above. The paper is organized as follows. Section 2 denes the considered system realization and describes the static H2 /H output feedback problem considered in this paper. Section 3 contains the necessary and sucient conditions for the existence of stabilizing static H output feedback controllers. Moreover, the formulation of the LMI-based nonconvex optimization problem, which must be necessarily solvable if the static H2 /H output feedback problem has a solution, can be found therein. Section 3.1 presents the main part of this paper. Therein we motivate the nonconvex multiobjective programming problem, and, similarly as in [31], we derive the SLPMM algorithm for nding a numerical solution of this problem class. Thereafter, we discuss the properties and global convergence of this procedure. Finally, in section 4, we present several examples for the design of suboptimal static H2 /H output feedback control laws, which will demonstrate the applicability of the SLPMM algorithm applied to this problem class. We also verify numerically the theoretical results and demonstrate the design of reduced xed order dynamic controllers if the algorithm terminates with an optimal value greater than nx + Tr(Y ). We will use the following notation. Ir denotes the (r r) identity matrix. The + set of real symmetric (n n) matrices is denoted by Sn , and Sn describes the cone of symmetric positive denite (n n) matrices. For A Sn , A 0 (A 0) means that A is positive denite (semidenite). Similarly, A 0 (A 0) denotes that A is negative denite (semidenite). For A, B Sn , A B (A B) denotes the n usual Loewner ordering [24]. The symbol Tr(A) = i=1 aii is the trace operator of a matrix A Rnn . ||A||F is the Frobenius norm of a matrix. Finally, ||Tzw || denotes the H norm of a proper and real rational stable transfer matrix, i.e., Tzw RH [10], and ||Tzw ||H2 is the usual H2 norm of a strictly proper and real rational stable transfer matrix, i.e., Tzw RH2 [10]. 2. The static H2 /H output feedback problem. In this section, we focus on the static H2 /H output feedback problem as formulated by Doyle et al. [6] for the full order dynamic output feedback case. However, we take a suboptimal approach for designing a static output feedback controller which is similar to [29]. The problem
92
1714
F. LEIBFRITZ
solved by Doyle et al. has been shown to be a dual problem of Bernstein and Haddad [2] in some sense; see, for example, Yeh, Banda, and Chang [45]. Khargonekar and Rotea [29] have obtained a nice solution to the dual problem for a class of suboptimal full order output feedback compensators. For example, a convex optimization approach is proposed to solve the full order output feedback dual problem. However, the static output feedback case considered in this paper is much more dicult, since the problem is then no longer convex. Consider a nite dimensional linear timeinvariant plant P with the state space realization x(t) = Ax(t) + B0 w0 (t) + B1 w1 (t) + B2 u(t), x(0) = 0, z(t) = C1 x(t) + D10 w0 (t) + D11 w1 (t) + D12 u(t), P y(t) = C2 x(t) + D20 w0 (t) + D21 w1 (t), where x(t) Rnx is the state, wi (t) Rnw , i = 0, 1, are the disturbance inputs, u(t) Rnu is the control input, z(t) Rnz is the regulated output, and y(t) Rny is the measured output of the control system. The static output feedback controller C with the state space realization C u(t) = F y(t)
is to be designed, where F Rnu ny denotes the unknown static output feedback gain. Substituting C into the plant, P yields the corresponding closed loop system given by cl x(t) = AF x(t) + B0F w0 (t) + BF w1 (t), z(t) = CF x(t) + D0F w0 (t) + DF w1 (t), x(0) = 0,
where the closed loop matrices AF , BF , CF , DF , B0F , and D0F are dened as follows: AF = A + B2 F C2 , BF = B1 + B2 F D21 , CF = C1 + D12 F C2 , DF = D11 + D12 F D21 , B0F = B0 + B2 F D20 , D0F = D10 + D12 F D20 . Throughout the whole paper, the following assumptions are imposed on the system. Assumption 2.1. (1) The pair (A, B2 ) is stabilizable, and the pair (A, C2 ) is detectable. (2) The data matrices in P , especially the following, are real constant matrices: A Rnx nx , B1 Rnx nw , B2 Rnx nu , C1 Rnz nx , C2 Rny nx , D11 Rnz nw , D12 Rnz nu , D21 Rny nw , B0 Rnx nw , D10 Rnz nw , D20 Rny nw . (3) F Rnu ny , nu < nx , ny < nx , and rank(B2 ) = nu , rank(C2 ) = ny . We start by considering the analysis problem. Assume the static output feedback law C is xed such that the closed loop system cl is internally stable. For example, there exists a static output feedback gain in the set (2.1) Fs = {F Rnu ny | AF is Hurwitz},
the so-called stability set. Let Tzw1 (s) = CF (sI AF )1 BF +DF (Tzw0 (s) = CF (sI AF )1 B0F + D0F ), s C, denote the closed loop transfer matrix from w1 to z (from w0 to z). The concepts of H norm and H2 norm/cost are well known (cf. [7]). Therefore, we will omit detailed discussion and content ourselves with starting the
93
1715
following denitions for reference. Since all coecient matrices are assumed to be real and F Fs is xed, the transfer matrix Tzw1 RH and the H norm of Tzw1 is dened by (2.2) ||Tzw1 || = sup max (Tzw1 (i)),
R
where max (Tzw1 ()) denotes the largest singular value of Tzw1 and i denotes the imaginary unit. Recall that Tzw0 RH2 and the H2 norm of Tzw0 is nite, for example, ||Tzw0 ||H2 < , if and only if D0F 0. In this case, if Lo denotes the observability Gramian of the pair (AF , CF ), the H2 norm of Tzw0 can be computed by (2.3)
T ||Tzw0 ||2 2 = Tr(B0F Lo B0F ); H
T for example, Lo satises Lyap(Lo , F ) = AT Lo + Lo AF + CF CF = 0. F Now let the scalar > 0 be given and assume that ||Tzw1 || < . Dene RF = T Inw 2 DF DF , and then it is a standard fact (see, for example, [30], [47]), that there exist a unique real symmetric matrix X and a gain F Rnu ny such that RF 0, T (2.4) Ric(X, F ) = AT X + XAF + 1 CF CF F 1 T T T + 1 (XBF + 1 CF DF )RF (BF X + 1 DF CF ) = 0, 1 T T and AF + 1 BF RF (BF X + 1 DF CF ) is Hurwitz. Moreover, X satises (cf. [2])
0 where P fullls Ric(P, F ) (2.5)
Lo
P,
0. Thus, if the H2 norm of Tzw0 is nite, then we have
T T T ||Tzw0 ||2 2 = Tr(B0F Lo B0F ) Tr(B0F XB0F ) Tr(B0F P B0F ). H
Similarly as in [2], [46], [29], or [35], these inequalities motivate us to dene the following auxiliary H2 /H cost function for the linear timeinvariant closed loop system cl : (2.6)
T C (P, F ) = Tr(B0F P B0F ),
which is an upper bound on ||Tzw0 ||2 2 if and only if D0F 0. Moreover, C (P, F ) H Tr(Y ) whenever the symmetric matrices P 0 and Y 0, Y Rnw nw , satisfy Ric(P, F ) 0 and (2.7) (F, P, Y ) := Y P B0F
T B0F P P
0.
Note that (2.7) is equivalent to Y (2.8) Y P B0

T B0 P P
T B0F P B0F
0 and can be stated also as follows: 0 +

T D20 0
0 P B2
D20
FT
T B2 P
0.
+ Thus, for given P Snx , the matrix inequality (2.8) is an LMI in Y and F . By this discussion, it is immediate that ||Tzw1 || < and ||Tzw0 ||2 2 < C (P, F ) if H and only if D0F 0. Hence, the Riccati inequality Ric(P, F ) 0 leads to an H norm bound and an H2 cost upper bound C (P, F ). If D0F = 0, then ||Tzw0 ||2 2 = . In H
94
1716
F. LEIBFRITZ
this case, we do not interpret the auxiliary cost function C (P, F ) as an upper bound on the H2 cost, but we can interpret it as a robust performance measure similar to the results of [46, Theorem 4] and [6, Theorem 1]. The results concerning the optimization of the auxiliary static H2 /H performance measure over the set of all stabilizing static controller gains F satisfying ||Tzw1 || < can be obtained from the following characterization. The proof of this result is very similar to [29, Lemma 2.1] and is thus omitted (see also [6, Theorems 1, 2]). Lemma 2.2. Consider the stable closed loop system cl and let Tzw1 (Tzw0 ) denote the closed loop transfer matrix from w1 to z (from w0 to z). Let > 0 be given and suppose that Tzw0 is strictly proper, i.e., D0F 0. Let Ric() be dened by (2.4). Then there exists a gain F Rnu ny satisfying ||Tzw1 || < if and only if there + exists a pair (P, F ), F Rnu ny , P Snx , such that RF 0 and Ric(P, F ) 0. In this case,
T (2.9) C (P, F ) = inf{Tr(B0F P B0F ) | (P, F ) satisfy RF
0, Ric(P, F ) 0, P
0}.
With the result of Lemma 2.2, our goal is to minimize the auxiliary static H2 /H performance measure over all stabilizing static output feedback gains F that enforce the H constraint. From the previous discussion, this is equivalent to minimizing Tr(Y ) over all matrices F , P , and Y satisfying Ric(P, F ) 0, RF 0, P 0, Y 0, and (F, P, Y ) 0. Thus, we state the suboptimal static H2 /H output feedback problem considered in this paper as the following nonconvex optimization problem: (2.10) min Tr(Y ), subject to (s.t.) P 0, RF 0, Ric(P, F ) 0, Y (F, P, Y ) 0. 0,
In the following sections, we will discuss a procedure for nding solutions to this problem. Note, a solution (P , F , Y ) of (2.10), if any exists, is suboptimal in the sense that
opt T C (X, F ) = min{Tr(B0F XB0F ) | X
0, RF
0, Ric(X, F ) = 0}
(2.11)
T Tr(B0F P B0F )
Tr(Y ),
opt where C (X, F ) denotes the minimal value of the optimal static H2 /H output feedback problem. Finally, note that the results in the following sections can be easily extended to the following static H2 /H output feedback problem formulation:
(2.12)
min Tr(Y ), s.t. P
0, RF 0, Ric(P, F ) 0, Y 0, (F, P, Y ) Lyap(P, F ) 0, D0F = 0.
0,
Here, the goal is to minimize an upper estimate of the optimal static H2 performance subject to the H constraint. 3. Suboptimal static H2 /H output feedback design; LMI approach. This paragraph is devoted to the computational design of a suboptimal static H2 /H output feedback controller. For deriving the LMI-based formulation of the static H2 /H output feedback problem, we need the following result which can be found, for example, in [25]. Lemma 3.1. Let B Rnm , rank(B) = m < n, C Rrn , rank(C) = r < n, and Rnn be given. Then there exists F Rmr satisfying BF C + (BF C)T + 0 if and only if N (B T )T N (B T ) 0 and N (C)T N (C) 0 hold, where N (B T )
95
1717
Rn(nm) , N (C) Rn(nr) , denoting any matrices whose columns form orthonormal bases of the null spaces of B T , C, respectively. By involving Lemma 3.1 and the strict bounded real lemma [47], the suboptimal static H output feedback problem, i.e., nd a static output feedback gain F , if any exists, such that AF is Hurwitz and ||Tzw1 || < , can be transformed to a problem of solving two LMIs coupled through a bilinear matrix equation. Similar results can be found, for example, in [12], [25], and [30]. Theorem 3.2 (existence of static H controllers). Let Fs = , > 0 be given and consider the closed loop system cl with w0 = 0. Then the following are equivalent. (i) There exists a static output feedback gain F Rnu ny such that AF is a Hurwitz matrix and ||Tzw1 || < . + (ii) There exists a pair (F, P ), F Rnu ny , P Snx , satisfying (3.1) BF C + (BF C)T + 0, AT P + P A P B1 T B1 P Inw = C1 D11 T P B2 C1 T D11 , B = 0 , C = [C2 D21 0] . D12 Inz
+ + (iii) There exist matrices P Snx and Q Snx satisfying P Q = I and
(3.2) (3.3)
T NQ
T AQ + QAT + 1 B1 B1 1 T (C1 Q + D11 B1 ) T AT P + P A + 1 C1 C1 1 T (P B1 + C1 D11 )T
T (C1 Q + 1 D11 B1 )T 1 T D11 D11 Inz T P B1 + 1 C1 D11 1 T D11 D11 Inw
NQ 0, NP 0,
T NP
T T where NQ := N ([B2 D12 ]) and NP := N ([C2 D21 ]), denoting any matrices whose T T columns form orthonormal bases of the null spaces of [B2 D12 ] and [C2 D21 ], respectively. Proof. Combining the strict bounded real lemma [47], Theorem 3.1, and using a Schur complement argument yields the desired result with Q = P 1 . The inequalities (3.2) and (3.3) are LMIs in Q and P , respectively, and are therefore convex. But nding P 0 and Q 0 satisfying P Q = I, (3.2), and (3.3) together is a dicult task since
(3.4)
+ H (P, Q, ) := {(P, Q) Snx | (P, Q) satisfying P Q = I, (3.2), (3.3)}
is a nonconvex set, i.e., elements in H (P, Q, ) must be inverse to each other. A numerical algorithm for determining a pair (P, Q) H (P, Q, ) can be found, for example, in Leibfritz [31]. Similarly as in Theorem 3.2, we can eliminate the matrix variable F from (2.8) by using [25, Theorem 1], which is a generalization of Lemma 3.1. In particular, we have the following lemma. + Lemma 3.3. Let Y Rnw nw , Y 0, P Snx , and F Rnu ny . Moreover, nx nw nx nu ny nw let B0 R , B2 R , and D20 R be given. Suppose [0 P B2 ]T (nw +nx )nu T R , rank([0 P B2 ] ) nu , and [D20 0] Rny (nw +nx ) , rank([D20 0]) ny . Then the following statements are equivalent. (i) There exists a triple (F, P, Y ) satisfying (2.8).
96
1718
F. LEIBFRITZ
+ + 0, P Snx , and Q Snx satisfying T B0 Q T B0 P P
(ii) There exist matrices Y Rnw nw , Y P Q = I and (3.5) (3.6)

T Q(Q, Y ) = N ([0 B2 ])T
Y B0 Y P B0
T N ([0 B2 ])
0, 0,
P(P, Y ) = N ([D20 0])T
N ([D20 0])
T where N ([0 B2 ]) and N ([D20 0]) denote any matrices whose columns form orthonorT mal bases of the null spaces of [0 B2 ] and [D20 0], respectively. Proof. Observing that
0 P B2
Inw 0
0 B2
and using [25, Theorem 1], we obtain the desired result with Q = P 1 . The set + (P, Q, Y ) := {(P, Q) Snx , Y Snw | (P, Q, Y ) (3.7) satisfying Y 0, P Q = I, (3.5), (3.6)} is not convex due to the coupling condition P Q = I. Obviously, a suboptimal static H2 /H controller in the sense of the previous section exists if and only if condition (ii) of Theorem 3.2 and (2.8) hold for the same + gain matrix F. Assuming that there exist matrices F, P Snx , Y Snw satisfying + (3.1) and (2.8), there exist matrices (P, Q) Snx and Y Snw satisfying condition (iii) of Theorem 3.2 and condition (ii) of Lemma 3.3. Thus, the static H2 /H output feedback problem can be necessarily transformed to the following bilinear optimization problem: min Tr(Y ), s.t. (P, Q, Y ) H (P, Q, ) (P, Q, Y ). (3.8)
Now suppose that there exists a solution triple (P , Q , Y ) of (3.8) (not necessarily unique); then there exists a suboptimal static H2 /H output feedback controller if and only if there exists an F satisfying (3.1) and (2.8) for P = P and Y = Y . Note, for given (P , Q , Y ), this is an LMI feasibility problem in F . Therefore, the suboptimal static H2 /H problem is solvable if and only if the bilinear optimization problem (3.8) has a solution and the corresponding LMI feasibility problem in F is nonempty. Hence, these observations lead to the following necessary and sucient conditions for the existence of static H2 /H output feedback controllers. Theorem 3.4. Let Fs = , > 0 be given, and consider the closed loop system cl . Then the following statements are equivalent. + (i) There exists a triple (F, P, Y ), F Rnu ny , P Snx , Y Snw , and Y 0 satisfying (3.1) and (2.8). + + (ii) There exist matrices P Snx , Q Snx , Y Snw , and Y 0 satisfying P Q = I, (3.2), (3.3), (3.5), (3.6), and, for all such xed (P, Y ), there is a matrix F Rnu ny satisfying (3.1) and (2.8). Due to this result, the suboptimal static H2 /H problem can be solved by rst nding a solution of (3.8) which is also a necessary condition for the solvability of (2.10). Secondly, if (3.8) has a solution, the static H2 /H output feedback gain, if any exists, can be obtained by solving the corresponding LMI feasibility problem in F . Finally, all gains F which can be reconstructed in this way will be called suboptimal opt due to the relation Tr(Y ) C (P , F ) C (X, F ).
97
1719
3.1. Bilinear LMI-based algorithm. In this subsection we describe a bilinear, multiobjective LMI-based algorithm for nding a triple (P, Q, Y ) satisfying approxi matively (3.8). As dened by (3.4) and (3.7), the sets H (P, Q, ) and (P, Q, Y ) are not convex and not closed. To make these sets convex, the coupling constraint P Q = I can be weakened to the following well-known semidenite programming (SDP)-relaxation: (3.9) P Q1 0 M (P, Q) := P I I Q 0.
Replacing P Q = I in the nonconvex sets H (P, Q, ) and (P, Q, Y ) by M (P, Q) 0 yields a convex approximation of these sets. Moreover, for computational purposes, we prefer to have closed sets. Introducing a positive scalar > 0 and replacing the closed loop matrix AF by A + I + B2 F C2 in Theorem 3.2, we can rewrite the existence conditions of Theorem 3.2 to the following bilinear matrix feasibility problem (cf. Leibfritz [31]). (3.10) Find (P, Q) 0, such that P Q = I, Q (Q, ) 0, P (P, ) 0,
where, for given > 0 and > 0, we dene

T Q (Q, ) = NQ T AQ + QAT + 2Q + 1 B1 B1 1 T C1 Q + D11 B1 T (C1 Q + 1 D11 B1 )T 1 T D11 D11 Inz
NQ ,
(3.11)
T (3.12) P (P, ) = NP
T AT P + P A + 2P + 1 C1 C1 1 T T (P B1 + C1 D11 )
T P B1 + 1 C1 D11 1 T D11 D11 Inw
NP ,
and NQ , NP are dened as in Theorem 3.2. With these denitions, we replace the nonconvex set H (P, Q, ) by the bilinear approximation
+ H (P, Q, , ) = {(P, Q) Snx | (P, Q) satisfying P Q = I, Q (Q, ) (3.13) P (P, ) 0}
0,
and redene the bilinear optimization problem (3.8) as (3.14) min Tr(Y ), s.t. (P, Q, Y ) H (P, Q, , ) (P, Q, Y ).
Then, replacing P Q = I by M (P, Q) 0 in (3.7) and (3.13) yields the following closed and linear approximations to the nonconvex and open sets H (P, Q, ) and (P, Q, Y ), respectively: (3.15) X (P, Q, ) = {(P, Q) Snx | M (P, Q) and X (P, Q, Y ) = {(P, Q) Snx , Y Snw | M (P, Q) (3.16) P(P, Y ) 0}. 0, Y 0, Q(Q, Y ) 0, 0, Q (Q, ) 0, P (P, ) 0}
Observe that the condition P Q = I, which is equivalent to Tr(P Q) = nx , is satised if and only if rank(M (P, Q)) nx , i.e., the nx smallest eigenvalues of the positive
98
1720
F. LEIBFRITZ
semidenite (2nx 2nx ) matrix M (P, Q) are equal to zero. In fact, the constraint P Q = I characterizes the boundary of the set {(P, Q) Snx | P Q1 0} = {(P, Q) Snx | M (P, Q) 0}.
Thus, a feasible matrix triple to the following nonconvex bilinear matrix feasibility problem (3.17) Find (P, Q) P Q = I, Q (Q, ) 0, Y 0, P (P, ) 0 such that 0, Q(Q, Y ) 0, P(P, Y ) 0
can be obtained by searching for boundary points of the linear and closed set X (P, Q, ) X (P, Q, Y ), dened by (3.15) and (3.16), respectively. This suggests the following nonconvex bilinear matrix inequality minimization problem: (3.18) min Tr(P Q), s.t. (P, Q, Y ) X (P, Q, ) X (P, Q, Y ).
Note that there exists a feasible triple (P, Q, Y ) satisfying (3.17) if and only if the optimal value of (3.18) is equal to nx . Since we are interested in a solution triple (P, Q, Y ) of the nonconvex problem (3.14), this observation motivates us to dene the following bilinear, multiobjective programming problem: (3.19) min Tr(P Q) + Tr(Y ), s.t. (P, Q, Y ) X (P, Q, ) X (P, Q, Y ).
This problem combines the objective functionals of (3.14) and (3.18). Obviously, minimizing Tr(P Q) enforces a solution of (3.19) to be close to or on the boundary of the feasible set of (3.19), while minimizing Tr(Y ) drives a solution of (3.19) to be suboptimal for the corresponding static H2 /H output feedback problem. If the triple (P , Q , Y ) is a boundary solution of (3.19) satisfying P Q = I, then we know that (P , Q , Y ) is contained in the feasible set of (3.14) and also satises nx + Tr(Y ) = min{Tr(P Q) + Tr(Y ) | (P, Q, Y ) H (P, Q, , ) (P, Q, Y )}. Thus, the optimal value of (3.14) fullls min{Tr(Y ) | (P, Q, Y ) H (P, Q, , ) (P, Q, Y )} nx Tr(Y ) nx = min{Tr(P Q) + Tr(Y ) | (P, Q, Y ) X (P, Q, ) X (P, Q, Y )} nx . Hence, an optimal solution of (3.19) yields an upper bound to the optimal value of (3.14) and at least a suboptimal solution of (3.14) if and only if the minimal triple (P , Q , Y ) of (3.19) satises P Q = I. Finally, if and only if P Q = I, then the corresponding suboptimal static H2 /H output feedback gain can be reconstructed from (P , Q , Y ) if and only if the following LMI feasibility problem in F is nonempty. Find F Rnu ny , such that (3.20) Y P B0 +
T B0 P P T D20 0
BF C + (BF C)T + F D20 0, 0
and
0 P B2
T 0 B2 P
FT
99
1721
where
AT P + P A + 2P T B1 P = C1
P B1 Inw D11
T P B2 C1 T D11 , B = 0 , C = [C2 D21 0] . D12 Inz
In this case, the discussion in section 3 implies the existence of a suboptimal static H2 /H controller. It also guarantees that AF is Hurwitz, ||Tzw1 || < , and T opt Tr(Y ) Tr(B0F P B0F ) C (X, F ). Thus, such an F is indeed a suboptimal static H2 /H output feedback gain. In what follows we explain a numerical procedure for determining an optimal solution of the bilinear, multiobjective programming problem (3.19). This problem is not convex since the functional Tr(P Q) is, in general, not convex, but it is bilinear. Therefore, in problem (3.19) we minimize a combination of a bilinear and linear matrix functional over a closed convex set. To solve such a problem, a sequential linearization programming approach as proposed by [1] for general nonconvex bilinear programming problems can be used. In particular, the idea of this approach is very simple. Instead of solving the nonconvex problem (3.19) directly, we linearize the bilinear part of the objective functional. Then we minimize successively the resulting (linearized) LMI constrained semidenite programming problems. Algorithm 1 (SLPMM). For given > 0, let X (P, Q, ) X (P, Q, Y ) = . 0 0 0 (0) Determine (P , Q , Y ) X (P, Q, ) X (P, Q, Y ). For k = 0, 1, 2, . . . do (1) Determine (U k , V k , Z k ) as the unique solution of (3.21) min Tr(P Qk + P k Q) + Tr(Y ), s.t. (P, Q, Y ) X (P, Q, ) X (P, Q, Y ).
(2) If Tr(U k Qk + P k V k ) + Tr(Z k ) = 2Tr(P k Qk ) + Tr(Y k ), Stop. (3) Compute [0, 1] by solving (3.22) min[0,1] Tr((P k + (U k P k ))(Qk + (V k Qk ))) + Tr(Y k + (Z k Y k )).
(4) Set P k+1 = (1 )P k + U k , Qk+1 = (1 )Qk + V k , and Y k+1 = (1 )Y k + Z k . This algorithm is similar to the SLPMM procedure proposed by Leibfritz [31] for the design of stabilizing static H2 and suboptimal static H output feedback controllers. The initialization step of Algorithm 1 is an LMI feasibility problem, and (3.21) is an SDP problem with a linear objective functional under LMI constraints. There are many algorithms available for solving such kinds of problems. For example, interior point methods developed for SDPs can be used (cf. [13]). Algorithm 1 terminates if the rst order necessary minimum principle is satised at a (local) minimum of (3.19) [34, section 6.1]. For example, dene (3.23) J (P, Q, Y ) := Tr(P Q) + Tr(Y ). Note that J (P, Q, Y ) is continuous and dierentiable on the cone of symmetric matrices. Hence, J (P, Q, Y )(U, V, Z) = Tr(U Q + P V ) + Tr(Z), J (P, Q, Y )(U, V, Z)(H, G, W ) = Tr(HV + U G)
100
1722
F. LEIBFRITZ
for any U, V, H, G Snx and Z, W Snw . If (P , Q , Y ) denotes a local minimum of J over the closed convex set X (P, Q, ) X (P, Q, Y ), then for any feasible direction (P P , Q Q , Y Y ) = (P, Q, Y ) at (P , Q , Y ) we obtain (3.24) J (P , Q , Y )(P, Q, Y ) = Tr(P Q + P Q) + Tr(Y ) 2Tr(P Q ) Tr(Y ) 0
for all (P, Q, Y ) X (P, Q, ) X (P, Q, Y ) [34, section 6.1, Proposition 1]. Numerically, for a suciently small scalar > 0, we terminate the algorithm if (3.25) k := Tr(U k Qk + P k V k ) + Tr(Z k ) 2Tr(P k Qk ) Tr(Y k ) , k 0.
Moreover, if (P k , Qk , Y k ) X (P, Q, )X (P, Q, Y ) and (U k , V k , Z k ) X (P, Q, ) X (P, Q, Y ) do not satisfy (3.24), we determine a step size parameter [0, 1] by minimizing (3.22) with respect to . Finally, the new iterates (P k+1 , Qk+1 , Y k+1 ) are convex combinations of (P k , Qk , Y k ) and (U k , V k , Z k ), respectively, and therefore, they are also contained in X (P, Q, ) X (P, Q, Y ). Note that the points (U k P k , V k Qk , Z k Y k ), k 0, are descent directions for J at (P k , Qk , Y k ) unless k 0 for some k 0. For proong the convergence of Algorithm 1, we need the following result. Sup pose for xed > 0 that there exist matrices (P 0 , Q0 , Y 0 ) X (P, Q, ) X (P, Q, Y ); then we can dene the following level set: (3.26) (P 0 , Q0 , Y 0 ) := {(P, Q, Y ) X (P, Q, ) X (P, Q, Y ) | J (P, Q, Y ) 0 0 0 J (P , Q , Y )}.
For this level set, we conclude the following lemma. Lemma 3.5. Let > 0 be given, X (P, Q, ) X (P, Q, Y ) be nonempty, and let 0 0 0 (P, Q, Y ) be given. Then the level set (P 0 , Q0 , Y 0 ) is (P , Q , Y ) X (P, Q, ) X compact. Proof. Using [24, Theorem 7.4.10], the closeness of X (P, Q, ), X (P, Q, Y ), and 0 0 0 the denition of (P , Q , Y ) shows the desired result. With the compactness of the level set (P 0 , Q0 , Y 0 ), it is straightforward to show the existence of an optimal solution of (3.19) in this level set. In particular, we have the following lemma. Lemma 3.6. Let X (P, Q, ) X (P, Q, Y ) = and (P 0 , Q0 , Y 0 ) X (P, Q, ) (P, Q, Y ) be given. Then there exists an optimal solution (P , Q , Y ) of (3.19) in X the level set (P 0 , Q0 , Y 0 ). Proof. By the compactness of (P 0 , Q0 , Y 0 ), the continuity of J , and the theorem of BolzanoWeierstrass, the result follows immediately. Thus, Lemma 3.6 ensures the existence of a solution of the multiobjective programming problem (3.19) at least in the compact level set (P 0 , Q0 , Y 0 ). The following lemma is needed in the proof of the Theorem 3.8 given below. Lemma 3.7. Let a > 0, b < 0, and c 1 be given. Then there exists 0 such that the optimal solution of min[0,1] (c + b + (2 /2) a) satises 0 < = b + 1. a
Moreover, = 0 if and only if a b. Proof. The result can be proven by using the convexity of the objective function and the KarushKuhnTucker theorem.
101
1723
The next result states the basic convergence properties of Algorithm 1. It also justies the fact that we need only the existence of a solution of (3.19) in (P 0 , Q0 , Y 0 ). Before stating these properties, we dene the boundary of the set X (P, Q, ) X (P, Q, Y ), given by (3.27) X (P, Q, Y ) := {(P, Q, Y ) X (P, Q, ) X (P, Q, Y ) | P Q = I}. X (P, Q, Y ) = H (P, Q, , ) (P, Q, Y ) X (P, Q, ) X (P, Q, Y ). Hence, the level set (P 0 , Q0 , Y 0 ) also contains all (P, Q, Y ) on the boundary of X (P, Q, ) X (P, Q, Y ). Therefore, similarly as in Lemma 3.6, we can conclude that there also exists an optimal solution of the nonconvex optimization problem (3.14) in the level set (P 0 , Q0 , Y 0 ). Theorem 3.8. Let > 0 and (P 0 , Q0 , Y 0 ) X (P, Q, ) X (P, Q, Y ) be given. Furthermore, assume that {(P k , Qk , Y k )} X (P, Q, ) X (P, Q, Y ) is generated by Algorithm 1. Then the sequence {(P k , Qk , Y k )} is well dened and for all k 0, we have (3.28) nx + Tr(Y ) nx + Tr(Y k+1 ) J (P k+1 , Qk+1 , Y k+1 ) < J (P k , Qk , Y k ) unless J ()() = 0, where (P , Q , Y ) solves (3.19). Thus, {J (P k , Qk , Y k )} con verges to J nx + Tr(Y ) and for all k 0, (P k , Qk , Y k ) (P 0 , Q0 , Y 0 ), i.e., {(P k , Qk , Y k )} is bounded. Finally, J (P k , Qk , Y k ) = nx + Tr(Y k ) if and only if (P k , Qk , Y k ) X (P, Q, Y ) and J ()() = 0 for some k. Proof. For all k 0, we dene the following abbreviations: S k = (P k , Qk , Y k ), S k = (U k P k , V k Qk , Z k Y k ), min and T k = (U k , V k , Z k ). J (S k )(S) If Algorithm 1 terminates at S k X (P, Q, ) X (P, Q, Y ), k 0, then J (S k )(S k ) = J (S k )(T k ) =
SX (P,Q,)X (P,Q,Y )
Obviously,
and J (P k , Qk , Y k )(P P k , Q Qk , Y Y k ) 0 for all (P, Q, Y ) X (P, Q, ) X (P, Q, Y ). Thus, if the algorithm does not terminate in step (2), then (3.29) J (S k )(S k ) = Tr(U k Qk + P k V k ) + Tr(Z k ) 2Tr(P k Qk ) Tr(Y k ) < 0, i.e., S k is a descent direction. Using S k X (P, Q, )X (P, Q, Y ), T k X (P, Q, ) X (P, Q, Y ), and the convexity of X (P, Q, ) X (P, Q, Y ), steps (3) and (4) of Algorithm 1 imply that (P k+1 , Qk+1 , Y k+1 ) X (P, Q, ) X (P, Q, Y ) for all [0, 1]. Hence, the sequence {S k } = {(P k , Qk , Y k )} is well dened by Algorithm 1. To show the strictly decreasing property of {J (P k , Qk , Y k )}, note that for [0, 1], the construction of the algorithm implies
2 J (S k )(S k )(S k ) 2 2 J (S k )(S k )(S k ) J (S k ) + J (S k )(S k ) + 2 1 J (S k ) + J (S k )(S k ) + J (S k )(S k )(S k ), (3.30) 2 where [0, 1] denotes the optimal value of (3.22).
J (S k+1 ) = J (S k ) + J (S k )(S k ) +
102
1724
F. LEIBFRITZ
Assuming J (S k )(S k )(S k ) = 0 and using (3.29), from (3.30) we obtain J (S k+1 ) < J (S k ) for all k 0 unless J (S k )(S k ) 0. Moreover, if J (S k )(S k )(S k ) < 0, then we can conclude that J (S k+1 ) < J (S k ) for all k 0. Finally, if J (S k )(S k )(S k ) > 0, then the problem is convex. Dening a = J ()()(), b = J ()(), and c = J (), Lemma 3.7 implies the existence of 0 such that the solution of (3.22) satises 0 < = But this implies (3.31) 1 2 1 1 J (S k )(S k )(S k ) = J (S k )(S k ) . 2 2 2 J (S k )(S k ) + 1. J (S k )(S k )(S k )
Using (0, 1], 0, (3.29), (3.30), and (3.31), we obtain J (S k+1 ) < J (S k ) unless J (S k )(S k ) 0 for all k 0. Hence, {J (P k , Qk , Y k )} is a strictly decreasing sequence unless J (S k )(S k ) 0. Since Tr(Y k ) 0 and Tr(P k Qk ) nx for all k 0, the sequence {J (P k , Qk , Y k )} is also bounded below by nx + Tr(Y ). Therefore, it converges to a limit J nx +Tr(Y ). Moreover, by the denition of the compact level set (P 0 , Q0 , Y 0 ), we know that {(P k , Qk , Y k )} (P 0 , Q0 , Y 0 ), and thus that the generated sequence {(P k , Qk , Y k )} is bounded. Finally, assume that (P k , Qk , Y k ) X (P, Q, Y ) and J (S k )(S k ) = 0 for some k. Note that this is fullled if and only if P k Qk = I and the minimal point T k of (3.21) satisfy U k = P k , V k = Qk , and Z k = Y k for some k. Thus, J (P k , Qk , Y k ) = nx + Tr(Y k ) nx + Tr(Y ) if and only if (P k , Qk , Y k ) X (P, Q, Y ) and J (S k )(S k ) = 0 for some k. Theorem 3.8 ensures that {J (P k , Qk , Y k )} is a strictly decreasing sequence which is bounded below by nx + Tr(Y ) if and only if (P k , Qk , Y k ) X (P, Q, Y ) and J (S k )(S k ) < 0. On the other hand, J (P k , Qk , Y k ) = nx + Tr(Y k ) nx + Tr(Y ) if and only if (P k , Qk , Y k ) X (P, Q, Y ) and J (S k )(S k ) = 0 for some k 0. Then we know that P k Qk = I and U k = P k , V k = Qk , Z k = Y k for some k. Hence, in this case, the SLPMM algorithm terminates in step (2) at a point satisfying the coupling condition P Q = I. The SLPMM algorithm may be interpreted as a modied version of the cone complementarity algorithm, expect that in each iteration we must also compute a step size parameter. But this further computational work is essential for the strictly decreasing property of the SLPMM. Indeed, the novel aspect of the SLPMM algorithm is that it always generates a strictly decreasing sequence {J (P k , Qk , Y k )}. Moreover, this approach guarantees the boundedness of the iterates (P k , Qk , Y k ) for all k 0. In contrast to this, the cone complementarity algorithm, the XY-centering algorithm, and other related computational methods in the literature do not share these properties. The strictness in the inequality (3.28) and the boundedness of {(P k , Qk , Y k )} is essential to prove the global convergence of the SLPMM algorithm. The following theorem states the global convergence of our method under rather weak assumptions compared to the existing algorithms in the literature. Theorem 3.9. Let > 0 and (P 0 , Q0 , Y 0 ) X (P, Q, ) X (P, Q, Y ) be given. Furthermore, assume that {(P k , Qk , Y k )} X (P, Q, ) X (P, Q, Y ) is generated by Algorithm 1. Then the following hold. (i) The algorithm terminates at some (P k , Qk , Y k ) satisfying (3.24), or every accumulation point (P , Q, Y ) of {(P k , Qk , Y k )} is stationary, i.e., (P , Q, Y ) satisfy (3.24).
103
1725
(ii) Every accumulation point (P , Q, Y ) of {(P k , Qk , Y k )} satises the nonconvex bilinear matrix feasibility problem (3.17) and solves the nonconvex optimization problem (3.14) if and only if the boundary set X (P, Q, Y ) dened in (3.27) is nonempty. Proof. (i) This result follows immediately by making straightforward modications to the proof of Leibfritz [31, Theorem 3.7]. (ii) Note that X (P, Q, Y ) = H (P, Q, , ) (P, Q, Y ) (P 0 , Q0 , Y 0 ) and that (P 0 , Q0 , Y 0 ) is compact. Theorem 3.8 implies limk J (P k , Qk , Y k ) = J nx + Tr(Y ) =: J . Since (P k , Qk , Y k ) (P 0 , Q0 , Y 0 ), we also know that the functions J (P, Q, Y ) = Tr(P Q) + Tr(Y ) and J (P, Q, Y ) = Tr(Y ) attain the (global) minimum values on the compact level set (P 0 , Q0 , Y 0 ); i.e., J = J (P , Q , Y ) nx + Tr(Y ) and J = J (P , Q , Y ) Tr(Y ), respectively. Suppose J > J = nx + Tr(Y ). From the compactness of (P 0 , Q0 , Y 0 ) = and the continuity of J , it follows that for every > 0, there exist a > 0 and a vicinity U (P , Q , Y ) such that U (P , Q , Y ) (P 0 , Q0 , Y 0 ) = and 0 J (P, Q, Y ) J < for all (P, Q, Y ) U (P , Q , Y )(P 0 , Q0 , Y 0 ). Choosing := J nx Tr(Y ) > 0 implies J (P, Q, Y ) < + nx + Tr(Y ) = J . But this implies the existence of an integer k such that for all k k and (P k , Qk , Y k ) 0 0 0 k k k , which contradicts that U (P , Q , Y ) (P , Q , Y ) we have J (P , Q , Y ) < J the sequence {J (P k , Qk , Y k )} converges monotonically to J . Hence J = J = nx + Tr(Y ); i.e., (3.32)
k
lim J (P k , Qk , Y k ) = J lim (J (P k , Qk , Y k ) nx ) = J .
k
Since {(P k , Qk , Y k )} (P 0 , Q0 , Y 0 ) and (P 0 , Q0 , Y 0 ) is compact, we conclude the existence of a convergent subsequence of {(P k , Qk , Y k )}; i.e.,
j
lim (P kj , Qkj , Y kj ) = (P , Q, Y ) (P 0 , Q0 , Y 0 ).
Suppose that this accumulation point is not globally optimal. Then J (P , Q, Y ) > and the sequence {J (P k , Qk , Y k ) nx Tr(Y )} do not J = nx + Tr(Y ) > J tend to zero. This contradicts (3.32). Hence, X (P, Q, Y ) = if and only if every accumulation point satises (3.17) and solves min{Tr(P Q) + Tr(Y ) | (P, Q, Y ) H (P, Q, , ) (P, Q, Y )} = nx + Tr(Y ) H (P, Q, , ) (P, Q, Y )}, = nx + min{Tr(Y ) | (P, Q, Y )
which in turn is equivalent to the solvability of (3.14). From Theorem 3.8, we know that the sequence {J (P k , Qk , Y k )} is strictly decreasing unless J ()() = 0 and is bounded below by nx + Tr(Y ), and hence converges. If in the limit Tr(P Q) = nx , then X (P, Q, Y ) = . In this case, Theorem 3.9 guarantees the existence of an accumulation point which satises the nonconvex bilinear matrix feasibility problem (3.17). Moreover, this point is also a solution of the nonconvex optimization problem (3.14), and therefore, in the limit the necessary condition for the existence of a static H2 /H output feedback gain is fullled.
104
1726
F. LEIBFRITZ
Note that for reconstructing such a gain, it is not necessary that the computed solution of the multiobjective problem (3.19) is the global solution of this problem. It suces that this solution is contained in the boundary set X (P, Q, Y ) dened by (3.27), i.e., it satises P Q = I. Then, this solution triple is at least an upper bound to the global optimal value of (3.14), and the corresponding gain, if any exists, is at least suboptimal. Thus, if Algorithm 1 terminates with a boundary solution of (3.19), i.e., J (P k , Qk , Y k ) = nx + Tr(Y k ) for some k 0, then Theorem 3.8 implies (P k , Qk , Y k ) X (P, Q, Y ) for some k. Moreover, (P k , Qk , Y k ) X (P, Q, Y ) satises min{Tr(Y ) | (P, Q, Y ) H (P, Q, , ) (P, Q, Y )} Tr(Y k ). Then, using the results of the previous section, a corresponding suboptimal static H2 /H output feedback controller exists if and only if the LMI feasibility problem (3.20) in F is nonempty. On the other hand, it may occur that Algorithm 1 terminates at a triple (P k , Qk , Y k ) satisfying the rst order necessary minimum principle, but (P k , Qk , Y k ) X (P, Q, Y ). In this case, we know that Tr(P k Qk ) > nx and a static gain F can not be reconstructed from this triple. But we can reconstruct a reduced dynamic output feedback control law of order nc nx max{nu , ny } from this triple by using a well-known system augmentation technique. For more details, we refer the reader, for example, to Leibfritz [31]. 4. Numerical examples. In this section, several examples are given for test purposes in order to test the SLPMM approach. We present examples for the design of suboptimal static H2 /H output feedback controllers. The SLPMM algorithm as well as the reconstruction of the controller gain F from the corresponding LMIs have been implemented making use of MATLAB 5.0 facilities. Particularly, for determining a feasible solution (P 0 , Q0 , Y 0 ) X (P, Q, ) X (P, Q, Y ) in Algorithm 1, step (0), we have used the LMI control toolbox [14] routine FEASP, which nds a solution to a given system of LMIs, if any exists. Moreover, for solving the linearized minimization problem (3.21) of Algorithm 1, step (1), which is a semidenite programming problem, we have taken the LMI control toolbox procedure MINCX. This solver is an implementation of Nesterov and Nemirovskis projective method for minimizing a linear objective function under LMI constraints as described in [13]. For MINCX we have adjusted the desired relative accuracy on the optimal value to 1012 . We terminated Algorithm 1 if, for a suciently small scalar > 0, the condition (3.25) was fullled. Moreover for the computation of a step size [0, 1] we have used the MATLAB function FMIN. Finally, we have taken the LMI control toolbox function FEASP applied to the LMI feasibility problem (3.20) for reconstructing the static output feedback controller gain F . The following data are given in the tables: the iteration counter k of the SLPMM algorithm; if k = 0, j denotes the total number of iterations for nding a feasible initial point (P 0 , Q0 , Y 0 ) by FEASP; else j is the total number of iterations for solving the semidenite programming problem (3.21) by MINCX with relative accuracy of 1012 ; the objective function value J (P k , Qk , Y k ) = Tr(P k Qk ) + Tr(Y k ) of the corresponding multiobjective optimization problem (3.19); the function values Tr(P k Qk ) and Tr(Y k ); and k , which indicates if the rst order necessary minimum principle is approximatively satised.
105
1727
Table 4.1 Convergence behavior of the SLPMM algorithm for the VTOL helicopter model: static H2 /H case. k 0 1 2 3 j 6 53 45 18 J (P k , Qk , Y k ) 1209.96168 4.21991536 4.21973458 4.21973458 Tr(P k Qk ) 1156.3660285 4.0001534809 4.0000000000 4.0000000000 Tr(Y k ) 53.5956538 0.21976188 0.21973458 0.21973458 k 2.058e + 03 1.808e 04 0.000e + 00
For all test examples we have chosen the data matrices B0 , D10 , D11 , and D20 as follows: B0 = B1 , D10 = D11 = 0, D20 = 0.
This choice guarantees D0F = 0. Therefore, if a static H2 /H controller gain F exists for the corresponding test problem, it is an upper estimate of ||Tzw0 ||2 2 that H enforces ||Tzw1 || < . Example 1. A state space model of the longitudinal motion of a VTOL helicopter is considered as in [28]. The dynamic equations of the helicopter model are linearized around a nominal solution where the given dynamic equation is computed for typical loading and ight conditions of the VTOL helicopter at a certain airspeed [41]. The linearization results in a fourth order linear time-invariant state equation with two control and two unknown input components. The data matrices of the linearized model are given by
0.0366 0.0271 0.0188 0.0482 1.0100 0.0024 A= 0.1002 0.3681 0.7070 0 0 1.0000 0.4422 0.1761 3.5446 7.5922 , B2 = 5.5200 4.4900 0 0 C1 =
1 2
0.4555 0.0468 4.0208 , B1 = 0.0457 0.0437 1.4200 0 0.0218
0 0.0099 , 0.0011 0
2 0
0 1
0 0
0 , C2 = 0 0 D21 = 0.00039
0 , D12 =
1 2
1 0
0 , 1
0.00174 .
The goal is to design a suboptimal static H2 /H output feedback controller C by the SLPMM Algorithm 1 according to the discussion of the previous sections. In the run of Algorithm 1 we have chosen = 0.01, = 0.423722, and = 1010 . Table 4.1 demonstrates the convergence behavior of the SLPMM. It illustrates numerically that {J (P k , Qk , Y k )} is a strictly decreasing sequence if (P k , Qk , Y k ) X (P, Q, Y ) and k < 0, which converges to nx + Tr(Y ) = 4 + 0.21973458. Thus, we have X (P, Q, Y ) = , and equality holds if and only if (P k , Qk , Y k ) X (P, Q, Y ) and k = 0 for some k according to Theorem 3.8. Moreover, Theorem 3.9 guarantees the existence of an accumulation point of {P k , Qk , Y k )} which also solves the nonconvex optimization problem (3.14). Therefore, a suboptimal static H2 /H output feedback controller exists if and only if the LMI feasibility problem (3.20) in F is nonempty. The resulting static gain matrix which satises (3.20) and the closed loop
106
1728 eigenvalues are F = 0.978987 19.43483

T
F. LEIBFRITZ
, (AF ) = 152.04
0.0989
0.2990 0.949i .
T Moreover, ||Tzw1 || = 0.2937866 < and Tr(B0F P B0F ) = 0.02869950 Tr(Y ) = 0.21973458, which shows that F is a suboptimal static H2 /H output feedback gain. opt To compare this solution with the optimal one, we have calculated C (X, F ) = 0.0231044 with an algorithm proposed by Leibfritz [32], [30]. This shows, that (3.19) provides a good suboptimal solution to the problem under consideration. As it can be also veried, the sequence {(P k , Qk , Y k ) (P 0 , Q0 , Y 0 ) and is thus bounded. For example, we have
{||P k ||} = {82.26, 49.82, 49.82, 49.82}, {||Qk ||} = {41.16, 1.159, 1.159, 1.159}, {||Y k ||} = {26.86, 0.1099, 0.1099, 0.1099}, which demonstrates numerically the boundedness of the generated sequence according to Theorem 3.8. The whole computation needs 2.47 CPU seconds on a SUN Ultra 60. Finally, for this example, we demonstrate numerically that the extended static H2 /H output feedback problem (2.12) can be also solved by the LMI approach as discussed in section 3. The SLPMM terminates after four outer iterations with an approximate solution (P, Q, Y ) of the corresponding bilinear programming problem, i.e., Tr(P Q) = 4.0000001. Then, using this solution, the corresponding LMI feasibility problem in F was found to be nonempty. Thus, a static H2 /H output feedback gain F satisfying (2.12) exists, and the resulting gain matrix is given by F = [0.457876 17.38107]T . Example 2 (transport airplane). This example studies the longitudinal motion of a modern transport airplane under VMIN ight conditions [15]. The linearized state space model yields the following data matrices:
0.06254 0.01089 0.07743 0 0 A = 0 0 0 0 0.01888 0.99280 1.67540 0 0 0 0 0 0 0 0.99795 1.31111 1 0 0 0 0 0 0.56141 0.00097 0.00030 0 0 0 0 0 0 0.02751 0.07057 4.25030 0 20.0000 0 0 0 0 0 0 0 0 20 30 0 0 0 0.06254 0.01089 0.07743 0 0 0 0.88206 0 0 0.00123 0.06449 0.10883 0 0 0 0 0.88206 0.00882 0 0 0 0 0 0 0 0.00882 0.88206 ,
0 0 0 0 0 B1 = 0 1.3282 0 0
0 0 0 0 0 0 0 1.62671 68.75283 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0
0 0.00519 0.47604 0 0.00098 0 0 0.00031 1 0.03378 , B2 = 0 , C1 = 2 0 30 0.00519 0 0 0.03086 0 0
0.00519 0.47604 0.00098 0.00031 T 0.03378 C2 = 0 0.00519 0.03086 0
0 0 T , D21 = 0 0
0 0 , 0 1
and D12 = [1/ 2]. The goal is the design of a robust static control law. Dening = 102 , = 0.152032 and = 1012 , Table 4.2 illustrates the convergence behavior of the SLPMM for this example. According to Theorem 3.8, it shows numerically
107
1729
Table 4.2 Convergence behavior of the SLPMM algorithm for the transport airplane: static H2 /H case. k 0 1 2 3 4 5 j 18 58 49 56 86 24 J (P k , Qk , Y k ) 1.766e + 14 131.706754 9.80419099 9.39445809 9.34395876 9.34395876 Tr(P k Qk ) 1.766e + 14 83.966174369 9.4602075306 9.0504993630 9.0000000338 9.0000000338 Tr(Y k ) 7.622e + 07 47.74057934352 0.343983455585 0.343958730681 0.343958722207 0.343958722207 k 3.533e + 14 1.689e + 02 2.843e 01 4.318e 02 0.000e + 00
the strictly decreasing property of the objective function values of (3.19) and the nonemptiness of the boundary of the set X (P, Q, ) X (P, Q, Y ), i.e., in the limit we achieve Tr(P Q) = nx . Thus, X (P, Q, Y ) = and Theorem 3.9 ensures the existence of an accumulation point of {P k , Qk , Y k )} which also solves the nonconvex optimization problem (3.14). Therefore, a suboptimal static H2 /H output feedback controller exists if and only if the LMI feasibility problem (3.20) in F is nonempty. Solving the LMI feasibility problem (3.20) results in the feasible gain F = 2.286293 0.001023 .
The real parts of the closed loop poles range between 0.02686 and 33.4213. MoreT over, ||Tzw1 || = 0.08088935 and Tr(B0F P B0F ) = 0.05135721 Tr(Y ) = 0.34395872. Once again, comparing this solution with the optimal one, we have calculated opt C (X, F ) = 0.0502692 with the algorithm of [32] for solving the optimal static H2 /H output feedback problem (2.11). Finally, the whole computation needs 53.53 CPU seconds. Again, this example demonstrates numerically the theoretical properties of our algorithm and shows that this approach can be used for the design of a suboptimal static H2 /H output feedback gain. Example 3 (EulerBernoulli beam). This example consists of a simple supported EulerBernoulli beam as discussed in [22] (see also [18]). Following [22], the state space model is given by P , with matrices as follows:
A = diag
T B2 = C2 =
0 r 4 0
1 0.02 r 2 0.891 0 0 0
, r = 1, . . . , 5 , 0.5878 0 0 0 0 0.7071 0 0 1 0 , B1 = 0 0 , B2 0101 ,
0.9877
0.309 0 0
C1 =
0.809 0
0.9511 0 0 0.5
0.309 0 , D21 =
0.5878 0 1.9 .
D12 =
Choosing = 0.01, = 3.59251, and = 1010 , Algorithm 1 terminates after 28.59 CPU seconds, and (3.20) provides the static H2 /H output feedback controller gain T F = 0.6221049 with Tr(B0F P B0F ) = 0.5742053 Tr(Y ) = 4.4202 and ||Tzw1 || = 2.179593. Moreover, the closed loop poles are (AF ) = 0.395 24.9i 0.257 15.9i 0.326 8.98i 0.305 0.95 0.059 3.99i .
opt Using the algorithm proposed by [30], [32] yields the optimal cost value C (X, F ) = 0.185282 of (2.11) with Fopt = 1.036985 and ||Tzw1 || = 2.14176.
108
1730
10
1
F. LEIBFRITZ
10
10 Singular Values (db)
10
10
10
10
10
10
10 Frequency (rad/sec)
10
Fig. 4.1. Singular value plot.

15
10 Penalty output z(t) / Observed output y(t) 5
10
20
30
40
50 Time t
60
70
80
90
100
Fig. 4.2. Phase portraits of the regulated output z (dashed line) and the observed output y (solid line) for the EulerBernoulli beam under worst case disturbances: Controlled case.
The SLPMM algorithm has generated the sequences {J (P k , Qk , Y k )}4 = {229.5443, 14.4202522, 14.4202000, 14.4202000, 14.4202000}, k=0 {Tr(P k Qk )}4 = {180.32, 10.000029, 10.0000000001, 10.0000000000, 10.0000000000}, k=0 {k }4 = {3.281 102 , 5.224 105 , 1.278 1010 , 0.000}, k=1
which underline the theoretical results stated in Theorems 3.8 and 3.9. Figure 4.1 shows the singular value response of the corresponding closed loop system for the problem under consideration. In Figure 4.2, the phase portraits for the regulated output variable z (dashed line) and the observed output variable y
109
1731
6 5 4 3 Worst case input w(t) 1 2 3 4 0 10 20 30 40 50 Time t 60 70 80 90 100 2 1 0
Fig. 4.3. Worst case disturbances.
250 200
Uncontrolled Penalty output z(t) / Observed output y(t)
150 100 50 0 50 100 150 200 250 0 10 20 30 40 50 Time t 60 70 80 90 100
Fig. 4.4. Phase portraits of the regulated output z (dashed line) and the observed output y (solid line) for the EulerBernoulli beam under worst case disturbances: Uncontrolled case.
(solid line) are displayed for the computed suboptimal static H2 /H output feedback controller under the corresponding worst case inputs w, as illustrated in Figure 4.3. These plots demonstrate that the computed controller gain yields an asymptotically stable closed loop system and satises the robustness constraint, even if the worstcase input aects the system. In contrast to this, the uncontrolled system outputs, as shown in Figure 4.4, oscillate away from the stable equilibrium of the system. Especially, the worst case inputs support this behavior and drive the system to be unstable. Example 4 (reduced order control). In this example we consider a threemassspring system and illustrate our approach for the design of a reduced order compen-
110
1732
F. LEIBFRITZ
sator if the SLPMM algorithm terminates at a point (P, Q, Y ) satisfying Tr(P Q) > nx . The system under consideration consists of three unit masses connected by two linear springs with stiness constant . There we have assumed that only the position of the third mass is measured and a control force acts on the rst mass. The linear state space realization of this system is given by P with the following data matrices:
A= 0 0 0 0 1 0 0 0 0 0 0 0 2 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 , B1 = 0 0 0 0 0 1 , B2 = 0 1 0 0 0 0 , C2 = 0 0 0 0 1 0 T , D12 = 0 0 0 0 0 0 1 ,
C1 = [I6 0]T , D21 = 0, and the state components xi , xi+1 , i = 1, 2, 3, denote the position and the velocity of the mass i, respectively. It is well known that this linear system is not stabilizable by a static output feedback gain, i.e., Fs = , but it is stabilizable by a reduced order controller of order nc = 3, where CD { xc (t) = Ac xc (t) + Bc y(t), u(t) = Cc xc (t) + Dc y(t) denotes the compensator of order nc nx with Ac Rnc nc , Bc Rnc ny , Cc Rnu nc , Dc Rnu ny , and xc (t) Rnc . Using Algorithm 1 with = 0.1, = 50.0869, and = 108 , we can compute a reduced order compensator of the form CD for this example as follows. Since Fs = , the SLPMM algorithm provides a solution triple (P, Q, Y ) of (3.19) satisfying Tr(P Q) > nx , i.e., (P, Q, Y ) X (P, Q, Y ). In particular, the algorithm has generated the following sequences: {J (P k , Qk , Y k )}5 = {10192.8935, 48.3013702, 38.2048006, 37.5019180, 36.7597188, k=0 36.7596674}, {Tr(P k Qk )}5 = {9786.7919, 39.2714948, 30.9417654, 30.8152693, 30.7274670, k=0 30.7274665}, {Tr(Y k )}5 = {406.1016, 9.0298754, 7.2630352, 6.6866487, 6.0322518, k=0 6.0322009}, {k }5 = {1.84 104 , 9.50, 1.24 102 , 9.12 103 , 3.19 109 }. k=1 Obviously, {J (P k , Qk , Y k )} is a strictly decreasing sequence which tends to Tr(P 5 Q5 ) + Tr(Y 5 ) = 30.7274665 + 6.0322009 > nx + Tr(Y ), and Algorithm 1 terminates after six iterations satisfying approximatively the rst order necessary condition. Since X (P, Q, Y ) = , we know that there exists no static gain for the problem under consideration. But it is possible to reconstruct a reduced order compensator from the computed solution triple (P, Q, Y ) of (3.19). For example, rst we compute the eigenvalues of P Q1 0. If there are m eigenvalues of P Q1 which are less than or equal to > 0, 1, then there exists an nc th order dynamic output feedback control law of the form CD such that the eigenvalues of the augmented closed loop system matrix AF = A + B2 F C2 have negative real parts [8, Theorem 3.1], where nc = nx m and the augmented system matrices are dened by [12]: A= (4.1) C2 = A 0 0 C2 0 0nc Inc 0 , B1 = B1 0 , B2 = 0 Inc B2 0 0 D21 , C1 = [C1 0], , F = Ac Cc Bc Dc ,
, D12 = [0 D12 ], D21 =
111
1733
B0 = B1 , D20 = [0 0ny nw ]T , and D10 = D11 = D11 . Second, we decompose the positive semidenite matrix P Q1 as U U T + E by a singular value decomposition, where U Rnx nc , E Rnx nx , ||E|| , and we dene (4.2) P = P UT U Inc , Y = Y.
By [36, Lemma 7.5] we have P 0. Then, replacing in (3.20) the system matrices by their augmented counterparts, the corresponding LMI feasibility problem in F has a solution. Choosing = 1010 yields m = 3, nc = 3, and the resulting third order compensator gain 0.23619 0.59416 1.85859 2.31612 0.52718 0.14806 1.38309 1.58075 Ac Bc . F = = 0.51506 0.91840 0.64521 2.59614 Cc Dc 0.05706 0.01872 0.55757 1.17127 The real parts of closed loop eigenvalues of AF ranges between 0.00963 and 0.02874. T Moreover, Tr(B0F P B0F ) = 6.032201 and ||Tzw1 || = 46.46845 < , which shows nu merically that F is a suboptimal third order H2 /H output feedback compensator gain. Summing up, this example demonstrates that we always can construct from the computed solution of Algorithm 1 at least a reduced order compensator gain which satises the design criteria. Finally, the whole computation time needs 9.62 CPU seconds on a SUN Ultra 60.
REFERENCES [1] K. P. Bennett and O. L. Mangasarian, Bilinear separation of two sets in nspace, Comput. Optim. Appl., 2 (1993), pp. 207227. [2] D. S. Bernstein and W. M. Haddad, LQG control with an H performance bound: A Riccati equation approach, IEEE Trans. Automat. Control, 34 (1989), pp. 293305. [3] S. P. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory, SIAM Stud. Appl. Math. 15, SIAM, Philadelphia, 1994. [4] M. Chilali and P. Gahinet, H design with pole placement constraints: An LMI approach, IEEE Trans. Automat. Control, 41 (1996), pp. 358367. [5] J. Doyle, A. Packard, and K. Zhou, Review of LFTs, LMIs and , in Proceedings of the 30th Conference on Decision and Control, Brighton, England, 1991, pp. 12271232. [6] J. Doyle, K. Zhou, K. Glover, and B. Bodenheimer, Mixed H2 and H performance objectives II: Optimal control, IEEE Trans. Automat. Control, 39 (1994), pp. 15751587. [7] J. C. Doyle, K. Glover, P. P. Khargonekar, and B. A. Francis, State-space solutions to standard H2 and H control problems, IEEE Trans. Automat. Control, 34 (1989), pp. 831847. [8] L. El Ghaoui and P. Gahinet, Rank minimization under LMI constraints: A framework for output feedback problems, in Proceedings of the European Control Conference, Groningen, The Netherlands, 1993, pp. 11761179. [9] L. El Ghaoui, F. Oustry, and M. AitRami, A cone complementarity linearization algorithm for static output feedback and related problems, IEEE Trans. Automat. Control, 42 (1997), pp. 11711176. [10] B. Francis, A Course in H Control Theory, Lecture Notes in Control and Inform. Sci. 88, Springer-Verlag, New York, London, Paris, 1987. [11] P. Gahinet, Explicit controller formulas for LMI-based H synthesis, Automatica J. IFAC, 32 (1996), pp. 10071014. [12] P. Gahinet and P. Apkarian, A linear matix inequality approach to H control, Internat. J. Robust Nonlinear Control, 4 (1994), pp. 421448. [13] P. Gahinet and A. Nemirovski, The projective method for solving linear matrix inequalities, Math. Programming, 77 (1997), pp. 163190.
112
1734
F. LEIBFRITZ
[14] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control Toolbox; For Use with MATLAB, The Math Works Inc., Natick, MA, 1995. [15] D. Gangsaas, K. R. Bruce, J. D. Blight, and U.-L. Ly, Application of modern synthesis to aircraft control: Three case studies, IEEE Trans. Automat. Control, 31 (1986), pp. 9951014. [16] J. C. Geromel, C. C. de Souza, and R. E. Skelton, LMI numerical solution for output feedback stabilization, in Proceedings of the American Control Conference, Baltimore, MD, 1994, pp. 4044. [17] J. C. Geromel, C. C. de Souza, and R. E. Skelton, Static output feedback controllers: Stability and convexity, IEEE Trans. Automat. Control, 43 (1998), pp. 120125. [18] J. C. Geromel and P. B. Gapski, Synthesis of positive real H2 controllers, IEEE Trans. Automat. Control, 42 (1997), pp. 988992. [19] J. C. Geromel, P. L. D. Peres, and S. R. Souza, A convex approach to the mixed H2 /H control problem for discrete-time uncertain systems, SIAM J. Control Optim., 33 (1995), pp. 18161833. [20] A. Giusto, A. Trofino-Neto, and E. B. Castelan, H and H2 design techniques for a class of prelters, IEEE Trans. Automat. Control, 41 (1996), pp. 864871. [21] K. M. Grigoriadis and R. E. Skelton, Low order control design for LMI problems using alternating projection methods, Automatica J. IFAC, 32 (1996), pp. 11171125. [22] W. M. Haddad, D. S. Bernstein, and Y. W. Wang, Dissipative H2 /H controller synthesis, IEEE Trans. Automat. Control, 49 (1994), pp. 827831. [23] W. M. Haddad, V. Kapila, and D. S. Bernstein, Robust H stabilization via parameterized lyapunov bounds, IEEE Trans. Automat. Control, 42 (1997), pp. 241248. [24] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, UK, 1985. [25] T. Iwasaki and R. E. Skelton, All controllers for the general H control problem: LMI existence conditions and state space formulas, Automatica J. IFAC, 30 (1994), pp. 1307 1317. [26] T. Iwasaki and R. E. Skelton, The XY-centering algorithm for the dual LMI problem: A new approach to xed-order control design, Internat. J. Control, 62 (1995), pp. 12571272. [27] I. Kaminer, P. P. Khargonekar, and M. A. Rotea, Mixed H2 /H control for discrete time systems via convex optimization, Automatica J. IFAC, 29 (1993), pp. 5770. [28] L. H. Keel, S. P. Bhattacharyya, and J. W. Howze, Robust control with structured perturbations, IEEE Trans. Automat. Control, 33 (1988), pp. 6877. [29] P. P. Khargonekar and M. A. Rotea, Mixed H2 /H control: A convex optimization approach, IEEE Trans. Automat. Control, 36 (1991), pp. 824837. [30] F. Leibfritz, Static Output Feedback Design Problems, Shaker Verlag, Aachen, Germany, 1998. [31] F. Leibfritz, Computational Design of Stabilizing Static Output Feedback Controllers, Technical Report Trierer Forschungsberichte Mathematik/Informatik 9901, Universitt Trier, a Trier, Germany, 1999. [32] F. Leibfritz, Static Output Feedback Design by Using a Newton-SQP Interior Point Method, Technical Report Trierer Forschungsberichte Mathematik/Informatik 9903, Universitt a Trier, Trier, Germany, 1999. [33] D. J. N. Limebeer, B. D. O. Anderson, and B. Hendel, A Nash game approach to mixed H2 /H control, IEEE Trans. Automat. Control, 39 (1994), pp. 6982. [34] D. G. Luenberger, Introduction to Linear and Nonlinear Programming, Addison-Wesley, Menlo Park, London, Sydney, 1973. [35] D. Mustafa, Combined H /LQG control via the optimal projection equations: On minimizing the LQG cost bound, Internat. J. Robust Nonlinear Control, 1 (1991), pp. 99109. [36] A. Packard, K. Zhou, P. Pandey, and G. Becker, A collection of robust control problems leading to LMIs, in Proceedings of the 30th Conference on Decision and Control, Brighton, England, 1991, pp. 12451250. [37] M. A. Rotea, The generalized H2 control problem, Automatica J. IFAC, 29 (1993), pp. 373 385. [38] M. A. Rotea and P. P. Khargonekar, H2 -optimal control with an H -constraint: The state feedback case, Automatica J. IFAC, 27 (1991), pp. 307316. [39] M. A. Rotea and P. P. Khargonekar, Generalized H2 /H control, in Robust Control Theory, IMA Vol. Math. Appl. 66, B. A. Francis and P. P. Khargonekar, eds., SpringerVerlag, New York, 1995, pp. 81103. [40] A. Saberi, B. M. Chen, S. Peddapullaiah, and U.-L. Ly, Simultaneous H2 /H optimal control: The state feedback case, Automatica J. IFAC, 29 (1993), pp. 16111614.
113
1735
[41] S. N. Singh and A. R. Coelho, Nonlinear control of mismatched uncertain linear systems and application to control of aircraft, Journal of Dynamic Systems, Measurement and Control, 106 (1984), pp. 203210. [42] R. E. Skelton, J. Stoustrup, and T. Iwasaki, The H control problem using static output feedback, Internat. J. Robust Nonlinear Control, 4 (1994), pp. 449455. [43] V. L. Syrmos, C. T. Abdallah, P. Dorato, and K. Grigoriadis, Static output feedbacka survey, Automatica J. IFAC, 33 (1997), pp. 125137. [44] I. Yaesh and U. Shaked, Minimum entropy static output feedback control with an H norm performance bound, IEEE Trans. Automat. Control, 42 (1997), pp. 853858. [45] H.-H. Yeh, S. S. Banda, and B.-C. Chang, Necessary and sucient conditions for mixed H2 and H optimal control, IEEE Trans. Automat. Control, 37 (1992), pp. 355358. [46] K. Zhou, K. Glover, B. Bodenheimer, and J. Doyle, Mixed H2 and H performance objectives I: Robust performance analysis, IEEE Trans. Automat. Control, 39 (1994), pp. 15641574. [47] K. Zhou and P. P. Khargonekar, An algebraic Riccati equation approach to H optimization, Systems Control Lett., 11 (1988), pp. 8591.
114
Information Sciences 179 (2009) 30413058
115
Contents lists available at ScienceDirect
Information Sciences
journal homepage: www.elsevier.com/locate/ins
H1 control design for fuzzy discrete-time singularly perturbed systems via slow state variables feedback: An LMI-based approach
Jiuxiang Dong, Guang-Hong Yang *
College of Information Science and Engineering, Northeastern University, Shenyang 110004, PR China Key Laboratory of Integrated Automation of Process Industry (Ministry of Education), Northeastern University, Shenyang 110004, PR China
a r t i c l e
i n f o
a b s t r a c t
This paper addresses the H1 control problem via slow state variables feedback for discretetime fuzzy singularly perturbed systems. At rst, a method of evaluating the upper bound of singular perturbation parameter with meeting a prescribed H1 performance bound requirement is given. Subsequently, two methods for designing H1 controllers via slow state variables feedback are presented in terms of solutions to a set of linear matrix inequalities (LMIs). In particular, one of them can be used to improve the upper bound of the singular perturbation parameter . Finally, two numerical examples are given to illustrate the effectiveness of the proposed methods. 2009 Elsevier Inc. All rights reserved.
Article history: Received 15 May 2008 Received in revised form 6 February 2009 Accepted 16 March 2009
Keywords: TakagiSugeno (TS) fuzzy systems Discrete-time systems Nonlinear singularly perturbed systems H1 control Linear matrix inequalities Reduced-order control
1. Introduction Slow and fast dynamic phenomena in control systems often occur due to the presence of small parasitic parameters, such as motor control systems, electronic circuits, magnetic-ball suspension systems, and so on. In a state space framework, such systems are commonly modeled by a mathematical description of singular perturbations, where a small parameter is exploited to determine the degree of separation between slow and fast parts of the dynamical system. Due to the very small singular perturbation parameter , the analysis and synthesis approaches for normal systems often lead to ill-conditioned results. Therefore, a so-called reduction technique with a two-step design methodology [17] is widely adopted for overcoming the difculty. Firstly, through the separate stabilization of two lower dimensional subsystems in two different time scales, a composite stabilizing controller is synthesized from separate stabilizing controllers of the two subsystems, where the controller could be determined without the knowledge of the small singular perturbation parameter. In the past several decades, many control problems of singularly perturbed systems have attracted considerable attentions, see the survey paper [23] and the references therein. In control theory, a well-known convex optimization technique, i.e., linear matrix inequality (LMI) technique, has been extensively exploited to solve control problems [3]. In contrast to Riccati approaches, linear matrix inequalities (LMIs) can be formulated as convex optimization problems that are amenable to compute solution and can be solved effectively [3]. Another good feature of LMIs is their ability of adding constraints to the parametrical optimization problem provided
* Corresponding author. Address: College of Information Science and Engineering, Northeastern University, Shenyang 110004, PR China. Tel.: +86 13504182968; fax: +86 24 83681939. E-mail addresses: dongjiuxiang@ise.neu.edu.cn, dong_jiuxiang@163.com (J. Dong), yangguanghong@ise.neu.edu.cn, yang_guanghong@163.com (G.-H. Yang). 0020-0255/$ - see front matter 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.ins.2009.03.012
3042 J. Dong, G.-H. Yang / Information Sciences 179 (2009) 30413058
116
they are themselves linear with respect to unknowns [12]. In particular, for H1 synthesis, it has the merit of eliminating the regularity restrictions attached to the Riccati-based solutions [11]. Motivated by the merits of the LMI formulations, some LMI-based controller design approaches for singularly perturbed systems have been developed in [710,28] recently. On the other hand, there has been a great deal of interest in using TakagiSugeno (TS) fuzzy models to approximate nonlinear systems, and many control problems of nonlinear systems have been widely studied based on TS fuzzy systems, see [14,15,26,29,30] and the references therein. In particular, the controller design conditions for the state feedback [19], static output feedback [21], and dynamic output feedback [2] cases are exploited in terms of solutions of LMIs for fuzzy singularly perturbed systems. Moreover, the methods of designing H1 state feedback controllers with pole placement constraints are given in [1]. In most cases, the fast dynamics of singularly perturbed systems are not adequately modeled and are, therefore, neglected in order to simplify the design [24,6]. In practice the fast variables are not directly measurable sometimes, for example, the exible variables (modeled as fast variables) of the exible link manipulators [22]. Therefore, the study on the problem of designing state feedback controllers by only using slow state variables of singularly perturbed systems is necessary. Because most of the existing synthesis techniques for singularly perturbed systems are independent of for avoiding to obtain ill-conditioned results, it is of great importance to nd the bound of for ensuring the stability of the closed-loop systems. As a result, it has attracted increasing interest in the past several decades. In [5], an approach to characterize and compute the stability bound is presented for continuous-time singularly perturbed systems. By considering critical stability criteria with a bialternate product, systematic approaches to determine the exact stability bound of discrete-time singularly perturbed systems are given in [13,18]. Moreover, an algorithm for nding the upper bound of the singular perturbation parameter for D-stability is presented in [16]. However, by the authors knowledge, the topic of evaluating the upper bounds of singular perturbation parameters for nonlinear discrete-time singularly perturbed systems with meeting H1 performance requirements has not been studied. In this paper, the topic will be addressed by using the two lemmas that are proposed for linear singularly perturbed systems in [8], which are given in Appendices A and B. In this paper, a method of evaluating the upper bound of the singular perturbation parameter for discrete-time fuzzy singularly perturbed systems with meeting a prescribed H1 performance bound requirement is given. Furthermore, two H1 controller design methods via slow state variables feedback are presented in terms of solutions to a set of LMIs. In contrast to the conventional design methods [23], the new design methods are with twofold advantages. One is that the two design methods are based on LMIs, which can eliminate the regularity restrictions attached to the Riccati-based solution. The other is that one of the two methods can be used to improve the upper bound of singularly perturbed parameter at the stage of control design, which implies that the tradeoff between the H1 performance index and the upper bound of the singular perturbation parameter is considered in the design. Thus, the new controller design method can overcome the disadvantage that the allowable upper bound of the singular perturbation parameter of the closed-loop system with the controllers designed by the existing ones is too small to be used. This paper is organized as follows. Section 2 presents system description and some preliminaries. In Section 3, a sufcient condition is derived for evaluating the upper bound of subject to a prescribed H1 performance constraint. Moreover, two new LMI-based H1 controller design methods are presented. In particular, one of them can improve the upper bound of the singular perturbation parameter by designing controllers. The validity of these approaches is illustrated by two numerical examples in Section 4. Finally, Section 5 concludes the paper. Notation: Rn denotes the set which consists of real n-vectors (n 1 matrices). For a symmetric block matrix, (*) is used for the blocks induced by symmetry, for example,
6 4 M 21 M 31
M 11
M 22 M 32
7 6 5 4 M 21 M 33 M 31
M 11
MT 21 M 22 M 32
MT 31 M 33
7 MT 5 32
The superscript T stands for matrix transposition and the notation MT denotes the transpose of the inverse matrix of M. 2. System description and some preliminaries A class of nonlinear singularly perturbed systems under consideration are described by the following fuzzy system model:
Plant Rule i : IF v 1 k is M i1 and
v 2 k is Mi2 ; . . . ; v p k is Mip ; THEN Ai12 x2 k Biw1 wk Biu1 uk x1 k 1 i x2 k 1 A21 x1 k Ai22 x2 k Biw2 wk Biu2 uk zk C iz1 x1 k C iz2 x2 k Dizw wk Dizu uk
Ai11 x1 k
where i 1; 2; . . . ; r, r is the number of IFTHEN rules. M il , 1 6 i 6 r, 1 6 l 6 p are fuzzy sets. v i k are the premise variables, x1 k 2 Rn1 and x2 k 2 Rn2 are respectively the slow and fast state vectors, uk 2 Rnu is the control input, wk 2 Rnw is the disturbance, zk 2 Rnz is the controlled output, the matrices Ai11 , Ai12 , Ai21 , Ai22 , Biw1 , Biw2 , Biu1 , Biu2 , C iz1 , C iz2 , Dizw and Dizu are of appropriate dimensions. > 0 is a singular perturbation parameter, which determines the degree of separation between the slow and fast modes of the system [1].
J. Dong, G.-H. Yang / Information Sciences 179 (2009) 30413058 3043
117
Denote
-i v k
p Y j1
lij v j k
lij v j k is the grade of membership of v j k in Mij , where it is assumed that

r X i1
-i v k > 0; -i v k P 0 i 1; 2; . . . ; r
i1
Let ai v k Pr
-i v k , then -i v k
and
0 6 ai v k 6 1;
r X i1
ai v k 1
ai v k, i 1; . . . ; r are said to be normalized membership functions. Then, the TS fuzzy model of (1) is inferred as follows:
x1 k 1 x2 k 1 zk
r X i1 r X i1 r X i1
ai v k Ai11 x1 k Ai12 x2 k Biw1 wk Biu1 uk ai v k Ai21 x1 k Ai22 x2 k Biw2 wk Biu2 uk

ai v k C iz1 x1 k C iz2 x2 k Dizw wk Dizu uk
which can be rewritten as follows:
xk 1 AakE xk Bw akwk Bu akuk zk C z akE xk Dzw akwk Dzu akuk

where
Aak
r X
ai v kAi ; Bw ak ai v
kBiu ; kDizw ;
r X
ai v kBiw ; ai v kC iz ;
r X i1
Bu ak
i1 r X
C z ak
i1 r X i1
Dzw ak
i1 r X i1
ai v
Dzu ak
ai v kDizu
5
! ! In1 n1 0 x1 k E xk In n 0 x2 k " i # " # 2 2 i i A11 A12 Bw1 Biw Ai Biw2 Ai21 Ai22 " # Biu1 Biu C iz C iz1 C iz2 Biu2
In this paper, the concept of parallel distributed compensation (PDC) is used to design fuzzy controllers, i.e., the designed fuzzy controller shares the same fuzzy sets with the fuzzy model in the premise parts. The more details can be found in [25]. For the fuzzy model (1), the following slow state feedback controller is adopted, where K ia , 1 6 i 6 r are the parameters to be designed.
Control Rule i : IF
v 1 k is Mi1
and
v 2 k is Mi2 ; . . . ; v p k is Mip
6 ! 7
THEN uk
K ia x1 k
r X i1
Because the control rules are the same as the plant rules, the fuzzy controller can be obtained as follows:
uk
where
r X i1
ai v kK ia x1 k
ai v kK i
x1 k x2 k
K i K ia
118
Combining (7) and (4), then the resulting closed-loop system is given as follows:
xk 1 AakE Bu akKakxk Bw akwk zk C z akE Dzu akKakxk Dzw akwk

where
r X i1
Kak
ai v k K ia 0
10
The replacement of x2 by n2 in system (9) will result in the equivalent system
nk 1 E Aak Bu akKaknk E Bw akwk zk C z ak Dzu akKaknk Dzw akwk

where
11
nk
x1 k n2 k
The H1 norm in [20] for nonlinear discrete-time systems is applicable for nonlinear discrete-time singularly perturbed systems. The denition is given as follows: Denition 1 [20]. Given a real number c > 0, it is said that the H1 norm of the closed-loop system (11) is less than or equal to c (i.e., the exogenous signals are locally attenuated by c) if there exists a neighborhood U of x 0 such that for every positive integer N and for every w 2 l2 0; N; Rnw for which the state trajectory of the closed-loop system (11) starting x0 0 remains in U for all k 2 0; N, the response z 2 l2 0; N; Rnz of (11) satises
N X i0
kzk k2 6 c2
N X i0
kwk k2 ; for all N
In this paper, the following problems will be addressed. 2.1. Evaluation of the upper bound of
with meeting stability and H1
performance bound requirement
Let c > 0 be a given constant and the gains K i be given. Find an > 0 as big as possible such that the system (1) with (7) is asymptotically stable and its H1 -norm is less than or equal to c for any 2 0; . 2.2. H1 controller designs without the consideration of improving the upper bound of
Let c > 0 be a given constant. Find gains K i i 1; . . . ; r, and there exists a positive scalar , such that the system (1) with (7) is asymptotically stable and its H1 -norm is less than or equal to c for any 2 0; . 2.3. H1 controller design with the consideration of improving the upper bound of
Let c > 0 be a given constant and > 0 be a prescribed upper bound of the singular perturbation parameter . Find gains K ia i 1; . . . ; r and an > 0 such that the system (1) with (7) is asymptotically stable and its H1 -norm is less than or equal to c for any 2 0; . The following lemmas will be used in this sequel. Lemma 2. If there exists a symmetric positive-denite matrix Pak such that the following LMIs hold,
U11 k; k 1 < 0; U21 k; k 1 U22 k; k 1
f or
2 0;
12
where Aak, Bw ak, Bu ak, C z ak, Dzw ak and Dzu ak are the same as in (5), Kak is the same as in (10), and
U11 k; k 1 Pak Aak Bu akKakT

E Pak 1E Aak Bu akKak 1 T C z ak Dzu akKak
C z ak Dzu akKak
U21 k; k 1 BT akE Pak 1E Aak Bu akKak w

1 DT akC z ak Dzu akKak zw
13
U22 k; k 1 BT akE Pak 1E Bw ak DT akDzw ak cI w zw
119
Then for each singular perturbation parameter 2 0; , the closed-loop system (11) is asymptotically stable and its H1 norm is less than or equal to c. Proof. See Appendix A. h Lemma 3 [8]. For a given positive scalar
, if the following conditions are satised,

14 15 16
aP0 a2 b c < 0 c<0

where a, b and c are constants, then
a2 b c < 0;
f or
2 0;
17
Proof. See Appendix B. h Lemma 4 [8]. For a given positive scalar
and matrices T1 , T2 , T3 , if following conditions are satised,
T1 P 0
18

then
T1 T2 T3 < 0
19 20 f or
T3 < 0
2 T1 T2 T3 < 0;
Proof. See Appendix C. h
2 0;
21
3. Main results In this section, a method of evaluating the upper bound of singularly perturbed parameter subject to the stability of the closed-loop system with meeting H1 performance bound requirements is presented. Moreover, two sufcient conditions for designing H1 controllers are given. In particular, one of them can improve the upper bound of the singular perturbation parameter by designing controllers. 3.1. Computation of stability bound of
subject to an H1
performance bound constraint
Firstly, the following preliminary lemma is needed. Lemma 5. If there exists a symmetric matrix
"
Q ak
Q 11 ak Q T ak 21 Q 21 ak Q 22 ak
such that the following inequalities hold,
# 3 2" ! 3 0 0 0 Q T ak 21 7 7 6 6 0 Q 22 ak 7 6 Q 21 ak 0 7 6 7 6 7 26 0 0 7 6 6 0 0 7 7 6 7 6 7 6 4 0 0 0 5 4 0 0 0 5 0 0 0 0 0 0 0 0 2 3 W11 k; k 1 6 7 0 cI 6 7 6 7 < 0; f or 2 0; 4 W31 k; k 1 Bw ak Q ak 1 5 2
22
W41 k; k 1 Dzw ak
where
cI
W11 k
Q 11 ak 0 0 0
Sak; ak 1 ST ak; ak 1 23
W31 k Aak Bu akKakSak; ak 1 W41 k C z ak Dzu akKakSak; ak 1

then for each singular perturbation parameter equal to c.
2 0; , the system (9) is asymptotically stable and its H1
norm is less than or
120
Proof. See Appendix D. h Based on Lemma 5, a method of evaluating the upper bound of the singular perturbation parameter H1 performance bound constraint is given in the following theorem. Theorem 6. For a given positive scalar
with a prescribed
, if there exist matrices Q i Q i T , Sij , 1 6 i; j 6 r

0 Sij 22 !
" Qi
Q i11 Q i21
Q i21 Q i22
# T ; Sij
S11 Sij 21
satisfying the following LMIs
Miil < 0; 1 6 i; l 6 r 1 1 Miil M ijl M jil < 0; r1 2
24a 1 6 i j 6 r; 1 6 l 6 r 24b 25a 1 6 i j 6 r; 1 6 l 6 r 25b
Kiil < 0; 1 6 i; l 6 r 1 1 Kiil Kijl Kjil < 0; r1 2

where
2" 6 6 6 6 6 M ijl 6 6 6 6 6 4
Q i11 0
0 0
# Sil Sil T
3 7 7 7 7 7 7 7 7 7 7 5 cI
0 cI i jl i j i A Bu K S Bw Q l C iz Dizu K j Sjl Dizw 0 3 2" # " # 3 2 0 0 0 Q i21 T 6 0 0 07 0 0 07 6 7 6 i 7 6 0 Q i22 0 7 6 Q 21 7 6 7 6 7 2 6 6 Kijl 6 0 0 0 07 6 0 0 0 07 7 7 6 7 6 7 6 6 0 0 0 05 4 0 0 0 07 5 4 0 0 0 0 0 0 0 0 3 2" i # Q 11 0 Sil Sil T 7 6 7 6 0 0 7 6 7 6 6 0 cI 7 7 6 7 6 6 Ai Biu K j Sjl Biw Q l 7 7 6 7 6 5 4 Dizw 0 cI C iz Dizu K j Sjl
then for each singular perturbation parameter equal to c. Proof. See Appendix E. h
2 0; , the system (9) is asymptotically stable and its H1
norm is less than or
Remark 7. Theorem 6 presents a method of estimating the upper bound of singularly perturbed parameter subject to the stability of the closed-loop system (9) while satisfying an H1 performance bound requirement. An upper bound of can be obtained by solving the following optimization problem:
Minimize
subject to 24 and 25
which can be effectively solved by using the LMI Control Toolbox [11]. 3.2. H1 controller design In this subsection, two LMI-based methods of designing H1 controllers are given. The two methods are with the merit of eliminating the regularity restrictions attached to the Riccati-based solutions. In particular, one of them can improve the upper bound of by designing controllers.
121
3.2.1. Design without the consideration of improving the upper bound of In the following, an LMI-based design method without considering the improvement of the upper bound of the singular perturbation parameter is given. Theorem 8. If there exist matrices Q i Q i T , Sil , Li , 1 6 i; l 6 r, with
" Q
i
Q i11 Q i21
Q i21 T Q i22
# ; Sil
S11 Sil 21
0 Sil 22
! ;
Li Lia
26
satisfying the following LMIs,
Caiil < 0;
1 6 i; l 6 r 1 1 Caiil Caijl Cajil < 0; r1 2 2" 6 6 6 6 6 6 6 4 # S S 0 Ai Sjl Biu Lj C iz Sjl Dizu Lj

il il T
27a 1 6 i j 6 r; 1 6 l 6 r 3 cI Biw Dizw Q l 0 7 7 7 7 7 < 0; 7 7 5 cI 27b
where
Q i11 0
0 0
Caijl
f or 1 6 i; j; l 6 r
then there exists a sufcient small
> 0 such that for
2 0; , the closed-loop system (9) with

28
K ia Lia S1 ; 11
16i6r
is asymptotically stable and its H1 norm is less than or equal to c. Proof. The proof is easily obtained from Theorem 6 and omitted. h Remark 9. Theorem 8 presents a sufcient condition for designing H1 controllers for discrete-time fuzzy singularly perturbed systems. The method is based on LMIs, and with the merit of eliminating the regularity restrictions attached to the Riccati-based solutions. Example 12 in Section 4 will illustrate the effectiveness of the method. However, in the method, the issue of improving the upper bound of the singular perturbation parameter is not addressed. As a result, the obtained controller might give a very small stability bound so that the resulting closed-loop system is unstable for a practical singular perturbation parameter , see Example 14 in Section 4. In order to overcome the difculty, another new method with the consideration of improving the upper bound of the singular perturbation parameter while satisfying H1 performance constraints will be proposed in the next subsection. 3.2.2. Design with the consideration of improving the upper bound of In this subsection, a new LMI-based H1 controller design method with the consideration of improving the upper bound of the singular perturbation parameter is given as follows: Theorem 10. For a given positive scalar
, if there exist matrices Q i Q i T , Sil , Li , 1 6 i; l 6 r, with

0 Sil 22 ! ; Li Lia 0
" Qi
Q i11 Q i21
Q i21 Q i22
# T ; Sil
S11 Sil 21
satisfying (27b) and the following LMIs,
Cbiil < 0;
1 6 i; l 6 r 1 1 Cbiil Cbijl Cbjil < 0; r1 2 !
29a 1 6 i j 6 r; 1 6 l 6 r 29b
where
Cbijl
6 0 Qi 22 6 6 0 2 6 6 4 0 0
0 0 07 6 6 Q i21 0 7 6 7 0 0 0 7 6 0 6 7 6 0 0 05 4 0 0 0 0 0
2"
Q i21 T
2"
0 0 07 6 7 6 7 6 6 0 0 07 6 7 6 7 6 0 0 05 4 0 0 0
Q i11 0
0 0
# Sil Sil T 0 cI Biw Dizw Q l 0
3 7 7 7 7 7 7 7 5 cI
A S Biu Lj C iz Sjl Dizu Lj
i jl
Then for
2 0; , the closed-loop system (9) with (28) is asymptotically stable and its H1
norm is less than or equal to c.
122
Proof. The proof is easily obtained from Theorem 6 and omitted. h Remark 11. In this section, Theorem 6 provides an LMI-based condition for estimating the upper bound of singularly perturbed parameter of discrete-time singularly perturbed systems subject to H1 performance bound constraints. In contrast to the existing techniques for estimating stability bounds, the new method can be used to estimate the bound of singularly perturbed parameter subject to H1 performance bound constraints. Moreover, two controller design methods are respectively given by Theorems 8 and 10. The two methods can eliminate the regularity restrictions attached the existing Riccatibased solution [23], the fact is shown by Example 12. In particular, the method given by Theorem 10 can be used to improve the upper bound of singularly perturbed parameter by designing controllers, which is illustrated by Example 14. 4. Example In Example 12, a discrete-time fuzzy singularly perturbed system is obtained by discretizing a tunnel diode circuit, which is borrowed from [2]. Because the example does not satisfy the regularity property, the conventional Riccati-based method is not applicable. The new LMI-based method given by Theorem 8 will be applied to the example for illustrating its effectiveness. Moreover, Example 14 is given for better illustrating the effectiveness of the method given by Theorem 10 (i.e., the H1 controller design method with the consideration of improving the upper bound of singularly perturbed parameter ). Example 12. Consider a tunnel diode circuit (Fig. 1), where the diode current is iD t and the diode voltage satisfy that
v D t, and they
iD t 0:2v D t 0:05v 3 t D
Moreover, C, R and L denote the capacitor, the resistance and the inductance, respectively. v C t, v R t and v L t are the capacitor voltage, the resistance voltage and the inductor voltage, respectively. iC t, iR t and iL t are the capacitor current, the resistance current and the inductor current, respectively. ut is the input voltage, wt is the disturbance. Considering Fig. 1 and applying the Kirchoff voltage and current law, we have that
dv C t ic t; RiL t v R t dt diL t L v L t; iC t iL t iD t dt v L t ut v R t v C t wt C
Let x1 t v C t, x2 t iL t. Combining them and (30), then it follows that
30
_ C x1 t 0:2x1 t 0:05x3 t x2 t 1 _ Lx2 t x1 t Rx2 t ut wt zt x1 t 0:1wt
31
where zt is the controlled output. Assume that the parameters of the circuit are C 100 mF, L 1 mH and R = 20 X. With these parameters, (31) can be rewritten as follows:
Fig. 1. The tunnel diode circuit in Example 12.
123
_ x1 t 2x1 t 0:5x3 t 10x2 t 1 _ x2 t 0:1x1 t 2x2 t 0:1ut 0:1wt zt x1 t 0:1wt
32
where xt xT t; xT tT and 104 . Assume that jxtj 6 3 and model the nonlinear system (32) by the sector nonlinear1 2 ity approach in [14], then the following TS fuzzy model can be obtained:
_ E xt zt
where
2 X i1
ai t Aic xt Bicw wt Bicu ut

2 X i1
ai t C icz xt Diczw wt
! ;
A1 c
10
0:1 2 ! 0 B1 B2 ; cu cu 0:1
A2 c
6:9
10
0:1 2
! ;
B1 B2 cw cw
0 0:1
! 1 0 0 !
x2 t 1 , 9
C 1 C 2 1 0; cz cz
D1 D2 0:1; E czw czw
and ai t is the normalized time-varying fuzzy weighting function for each rule i 1; 2, satises a1 t 1 a2 t 1 a1 t.
From (32), it can be seen that x2 t, i.e., the inductor current iL t, changes very fast and exhibits a large range of variation. Therefore, it is very difcult to measure x2 t. On the other hand, the slow variable x1 t changes slowly and exhibits a small range of variation, it can be easily measured. Therefore, the slow state feedback controller is needed for this example. Inhere, we discretize the model with a sampling period T 0:3 s and a zero-order holder, then the following discretetime singularly perturbed model is obtained:
xk 1 zk
where
2 X i1
2 X i1
ai k Aid E xk Bidw wk Bidu uk

ai k C idz xk Didzw wk
!
A1 d B2 dw
! ! 6:8210 34:1039 0:1894 A2 ; B1 d dw 0:0784 0:3921 0:3410 1:7051 0:0405 ! ! ! 0:4547 0:1894 0:4547 ; B1 ; B2 ; C 1 C 2 1 0; dz dz du du 0:0273 0:0405 0:0273 1:5683 7:8415 ;
D1 D2 0:1 dzw dzw

Due to Ddzw ak 0, the regularity property is not satised, the conventional Riccati-based methods [23] are not applicable. Applying Theorem 8 (the H1 controller design method without the consideration of improving the upper bound of the singular perturbation parameter ), then the following results are obtained
Q1
0:4220
0:0308
0:0308 995:8465
! ;
Q2
0:4220
0:0293
0:0293 994:0820
L1 4:4462; a K1 a 10:5348;
L2 6:5378; a K 2 15:4908; copt 0:51639 a
With the obtained controller gains, the upper bound of the singular perturbation parameter of the tunnel diode circuit system are estimated as 0:1772 by using Theorem 6 under the H1 performance constraint c 0:5614. It is bigger than the practical parameter 104 . Therefore, the designed controller can be used and some simulation results are shown in Figs. & 1; 5 6 k 6 10 24 with an initial state x0 0:51T , and wk 0; others: From Figs. 2, and 3, it can be seen that the resulting closed-loop system is asymptotically stable and with a good H1 performance, which further shows the effectiveness of the proposed condition in Theorem 8. Remark 13. Note that Theorem 8 is applied for designing an H1 controller in Example 12 and the upper bound of singularly perturbed parameter for the resulting closed-loop system is more than the practical , hence the controller is applicable. Since the controller design condition in Theorem 8 is without considering the tradeoff of the H1 performance bound c and the stability bound , the obtained controller by Theorem 8 might be not applicable for some singularly perturbed systems. For solving the problem, Theorem 10 can be applied as an alternative, which is shown in Example 14.
124
1 0.8 0.6 0.4
x x
1 2
x(k)
0.2 0 0.2 0.4 0
10
15
20
25
30
Time (sec)
Fig. 2. Trajectories of xk in Example 12.
3.5 3 2.5 2
z w
z(k), w(t)
1.5 1 0.5 0 0.5
10
15
20
25
30
Time (sec)
Fig. 3. Trajectories of zk and wk in Example 12.
Fig. 4. Membership functions ai x1 k for Example 14.
125
Example 14. Consider a numerical example, which is described by (4) with
0:6
1 2 0
2 0 0:3
0:7
1 1 0
3 0 0:3
0:8
6 0 6 A1 6 4 0:1
3 7 7 7; 1 5
6 0 6 A2 6 4 0:1
2 7 7 7 1 5
2:5 0:1 0:2 0:4 2 3 2 3 0 0 6 0:2 7 6 0:3 7 6 7 6 7 2 B1 6 7; Bw 6 7; w 4 0 5 4 0 5 C1 z 0:1 ! 1 0 0 0 ; 0 1 0 0 2 3 1 0 0 0 60 1 0 07 6 7 E 6 7 40 0 05 C2 z 0

where
8 0:5 0:1 0:6 2 3 2 3 0:5 0:3 6 1 7 6 1 7 6 7 6 7 2 B1 6 7; Bu 6 7 u 4 0 5 4 0 5 D1 zw 0 1 0:8 ! ; D2 zw 0:4 ! ! 0 ; D1 D2 zu zu 1:2 0:1 0
0:2
0 0
0:05 and the membership functions ai x1 k, i 1; 2 are given in Fig. 4.
Applying Theorem 8 to the example (the H1 controller design method without the consideration of improving the upper bound of the singular perturbation parameter ), we can obtain the following controller gains and the optimal H1 performance bound.
K 1 0:1544 a
2:1685;
K 2 0:0764 a
1:9145
copt 1:3334
Table 1 Upper bounds of . Theorem 8
33
Theorem 10 0.0527 0.0556 0.0576
c 1:5 c2 c 2:5
0.0014 0.0049 0.0073
x 10
15
x x
1 2
x(k)
10
15
20
25
30
Time (sec)
Fig. 5. Trajectories of xk via the gain (33) in Example 14.
126
Moreover, by using Theorem 10 with 0:05 (the H1 controller design method with the consideration of improving the upper bound of the singular perturbation parameter ), we can obtain
K 1 0:2233 a
2:1834;
K 2 0:7358 a
1:8368
copt 1:4284
34
Then the allowable upper bound of the singular perturbation parameter of the closed-loop system with the controller gain (33) or (34) can be estimated by using Theorem 6. The obtained results are shown in Table 1. From Table 1, it can be seen that,for the larger H1 performance indices,the larger upper bounds of the singular perturbation parameter are achieved by using Theorem 10, which shows that the method of Theorem 10 is effective for improving the upper bound of the singular perturbation parameter .
8 6 4 2
x1 x
2
x(k)
0 2 4 6 0
10
15
20
25
30
Time (sec)
Fig. 6. Trajectories of xk via the gain (34) in Example 14.
30
25
20
The ratio
15
10
1.3572
10
15
20
25
30
Time (sec)
Fig. 7. q Pk T Pk T i0 z izi= i0 w iwi with the gain (34) in Example 14.
127
Note that
0:05, then for the case of a1 k 0, a2 k 1, the eigenvalues of
AakE Bu akKak A2 E B2 K 2 A2 E B2 K 2 0 u u a
with the controller gain (33) designed by using Theorem 8 are 1.1190, 0.6411, 0.4496, 0.1641, which implies that the resulting closed-loop system is unstable. Therefore,the controller cannot be applied. On the other hand, Table 1 shows the achieved upper bounds of of the closed-loop system with the controller gains (34) under different H1 performance requirements. The upper bounds are more than 0:05. The fact shows that Theorem 10 can be used to effectively improve the upper bound of the singular perturbation parameter at the stage of controller design. What it follows, some simulation results will be given in order to further validate the effectiveness of Theorem 10. Assume that the initial state x0 2 8 0 0T and the disturbance
wk
&
2; 3 6 k 6 10 0; others
Figs. 5 and 6 show the trajectories q of the state xk of the closed-loop system with the controller gains (33) and (34), respecPk Pk T T tively. Fig. 7 shows the ratio of i0 z izi= i0 w iwi in the simulation. From Fig. 5, it can be seen that the closed-loop system with the controller gain (33), which is obtained by Theorem 8, is unstable. It can be seen that from Figs. 6 and 7 that the controller gain (34), which is obtained by Theorem 10, can guarantee the stability and meet the H1 performance requirement. These simulation results further show the advantage of the method given by Theorem 10, i.e., Theorem 10 can be used to effectively improve the upper bound of the singular perturbation parameter at the stage of controller design. 5. Conclusion In this paper, the H1 control problem via slow state variables feedback for discrete-time fuzzy singularly perturbed systems has been investigated. Two LMI-based methods for designing H1 controllers via slow state variables feedback are presented, and one of them can be used to improve the upper bound of the singular perturbation parameter , which can overcome the disadvantage in the conventional design methods where the designed controller might not be used because the resulting allowable upper bound of the singular perturbation parameter is too small. Moreover, a method of evaluating the upper bound of a singular perturbation parameter with meeting a prescribed H1 performance bound requirement is given in terms of solutions to a set of LMIs. The effectiveness of the proposed methods has been illustrated by the numerical examples. Acknowledgements This study was supported in part by the Funds for Creative Research Groups of China (No. 60821063), the State Key Program of National Natural Science of China (Grant No. 60534010), National 973 Program of China (Grant No. 2009CB320604), the Funds of National Science of China (Grant No. 60674021), the 111 Project (B08015) and the Funds of PhD program of MOE, China (Grant No. 20060145019). Appendix A. Proof of Lemma 2 Proof. Consider the system (11) (which is equivalent to the system (9)), and let Pak > 0. Choose Lyapunov function
Vk cnT kPaknk
then
Vk 1 Vk zT kzk c2 wT kwk cAak Bu akKaknk Bw akwk E Pak 1E Aak Bu akKaknk Bw akwk cnT kPaknk C z ak Dzu akKaknk Dzw akwk C z ak Dzu akKaknk Dzw akwk c2 wT kwk # ! !T " nk U11 k; k 1 UT k; k 1 nk 21 wk
T T
U21 k; k 1 U22 k; k 1
wk
where /11 k; k 1, /21 k; k 1, /22 k; k 1 are the same as in (13). From (12) and the above equality, then it follows that
Vk 1 Vk zT kzk c2 wT kwk 6 0;
for
2 0;
128
From the above inequality, we have that the system (11) with 2 0; is asymptotically stable in the disturbance-free case. For the initial condition x0 0 and every positive integer N, sum the above inequality from k 0 to N, then we can obtain
VN V0
which implies that
N X k0
1 X k0
zT kzk c2
1 X k0
wT kwk VN
1 X k0
zT kzk c2
1 X k0
wT kwk 6 0;
for
2 0;
zT kzk 6 c2
N X k0
wT kwk;
for
2 0;
Combining it and Denition 1, it follows that the H1 norm of the closed-loop system (11) is less than or equal to c. h Appendix B. Proof of Lemma 3 Proof. Consider the following two cases: (i): If a 0, then from (15) and (16), (17) obviously holds. (ii): If a > 0, we consider the following quadratic function of ,
y a2 b c
35
Since a > 0, y is convex function of [4]. From (15) and (16), it follows that y < 0 and y0 < 0, which further implies that y < 0 for 2 0; , i.e., when a > 0, (17) holds. Thus, the proof is complete. h
Appendix C. Proof of Lemma 4 Proof. For all nonzero vector xk, pre- and post-multiplying (18)(20) by xT k and its transpose, then we have
xT kT1 xk P 0
36
T T
2 T
x kT1 xk x kT2 xk x kT3 xk < 0
37 38
x kT3 xk < 0
Denote axk xT kT1 xk, bxk xT kT2 xk, cxk xT kT3 xk. Substituting axk , bxk and cxk into (36)(38), then it follows that
axk P 0 axk 2 bxk cxk < 0 c xk < 0

From (39)(41) and applying Lemma 3, we can obtain axk 2 bxk cxk < 0 for
39 40 41
2 0; , i.e., for all nonzero vector xk,
2 xT kT1 xk xT kT2 xk xT kT3 xk < 0;
for
2 0;
which implies that matrix inequality (21) holds. Thus, the proof is complete. h
Appendix D. Proof of Lemma 5 Proof. (22) can be rewritten as follows,
W11 k 6 0 6 6 4 W31 k
cI
3 for
Bw ak Q ak 1 W41 k Dzw ak 0 cI
7 7 7 < 0; 5
2 0; :
42
where W31 k, W41 k are the same as in (23) and
W11 k E QakE Sak; ak 1 ST ak; ak 1

where E is same as in (5).
129
From (42), we can obtain that W11 k < 0. Considering the block (3,3) of (42), then yields Q ak > 0. Combining and W11 k < 0, then we have Sak; ak 1 ST ak; ak 1 > 0. Pre- and post-multiply (42) by 3 3 2 1 ST ak; ak 1 0 0 0 S ak; ak 1 0 0 0 7 6 6 0 I 0 0 7 and 6 0 I 0 0 7, respectively. Then it follows that 7 6 4 4 0 0 I 05 0 0 I 05 it 2 0 0 0 I 0 0 0 I
7 6 6 7 0 cI 7 < 0; 6 7 6 Bw ak Q ak 1 5 4 Aak Bu akKak 0 cI C z ak Dzu akKak Dzw ak

where
e W 11 k
3 for
2 0;
43
e W 11 k ST ak; ak 1E Q akE S1 ak; ak 1 S1 ak; ak 1 ST ak; ak 1

Let P1 ak E Q akE , therefore,
Pak > 0
then
44
ST ak; ak 1 Pak P1 ak S1 ak; ak 1 Pak P 0

which implies that
ST ak; ak 1P1 akS1 ak; ak 1 S1 ak; ak 1 ST ak; ak 1 P Pak

Combining it with (43), we can obtain
Pak
cI Bw ak
1 1
E P ak 1E 0 3
1 1
3 for
6 0 6 6 4 Aak Bu akKak
7 7 7 < 0; 5 cI
2 0;
45
C z ak Dzu akKak Dzw ak

Applying Schur complement lemma to (45), then we have
2 6 4
Pak 0
cI
1 1
Aak Bu akKak Bw ak E P ak 1E 2 T 3 C z ak K T akDT ak zu 16 7 4 5 C z ak Dzu akKak Dzw ak 0 DT ak zw
7 5
2 6 4
where
0 11 k 21 k 22 k
1 1
3 7 50; for
2 0;
Aak Bu akKak Bw ak E P ak 1E 1 T 11 k Pak C z ak Dzu akKak
C z ak Dzu akKak 1 T 21 k Dzw akC z ak Dzu akKak
1 22 k cI DT akDzw ak zw
Applying Schur complement lemma to the above inequality, again, then it follows that
! ! AT ak K T akBT ak 11 k T k u 21 E Pak 1E Aak Bu akKak Bw ak 21 k 22 k BT ak w ! U11 k; k 1 UT k; k 1 < 0; for 2 0; 21 U21 k; k 1 U22 k; k 1
where U11 k; k 1, U21 k; k 1 and U22 k; k 1 are the same as in (13).
46
130
Then applying Lemma 2 to (46), we have that, for each singular perturbation parameter asymptotically stable and its H1 norm is less than c. Then the conclusion follows. h Appendix E. Proof of Theorem 6 Proof. Considering the block (3,3) of (24a) and from (24a), it follows that
2 0; ,
the system (11) is
Qi > 0 1 6 i 6 r
which also implies that
Q i22 > 0 1 6 i 6 r
Let
47 3 0 0 07 7 0 0 07 7 7 0 0 05 0 0 0 # 3
0 Q i22
6 0 6 6 0 T 1i 6 6 4 0 2" 0 0
6 6 Q i21 0 6 T 2i 6 0 6 6 4 0 0 T 3il Miil
Q i21 T
0 0 07 7 7 0 0 07 7 7 0 0 05 0 0 0
From (47), it follows that
T 1i P 0
From (25a), we have
48a 48b 48c for
2 T 1i T 2i T 3il < 0
From (24a), we have
T 3il < 0
Applying Lemma 4 to (48a), then yields
2 T 1i T 2i T 3il < 0;
i.e.,
2 0; ;
1 6 i; l 6 r
Kiil < 0;
where
for
2 0; ;
!
1 6 i; l 6 r 2" # 2" # S S
il il T
49a 3 3 cI Biw Dizw Q l 0 7 7 7 7 7 7 7 7 7 5 cI
0 0 6 0 Qi 22 6 6 0 Kiil 2 6 6 4 0 0
0 0 0 07 6 6 Q i21 0 7 6 0 0 0 7 6 7 0 6 7 6 0 0 05 4 0 0 0 0 0
Q i21 T
6 0 0 07 6 7 6 7 6 6 0 0 07 6 7 6 7 6 0 0 05 6 4 0 0 0
Q i11 0
0 0
0 i A Biu K i Sil C iz Dizu K i Sil
Similarly, from (47), (24b) and (25b), we can also obtain
1 1 Kiil Kijl Kjil < 0; r1 2

where
for
2 0; ;
2" 0
1 6 i j 6 r; 1 6 l 6 r # 0 0 0 0 0 0 0 0 3 2" # Sil Sil T 0 A S Biu K j Sjl C iz Sjl Dizu K j Sjl

i jl
49b 3 cI Biw Dizw Q l 0 7 7 7 7 7 7 7 5 cI
Kijl
26
6 0 Qi 22 6 0 6 6 4 0 0
0 0 07 6 6 Q i21 0 7 6 7 0 0 0 7 6 0 6 7 6 0 0 05 4 0 0 0 0 0
Q i21 T
07 6 7 6 7 6 6 07 6 7 6 7 6 05 4 0
Q i11 0
0 0
131
Applying the parameterized linear matrix inequality (PLMI) technique in [27] to (49a), then yields
r r XX i1 j1
ai kaj kKijl < 0; 1 6 l 6 r
Multiplying the above inequality by al k 1 and summing them from l 1 to r, then we can obtain
r r r XXX i1 j1 l1
ai kaj kal k 1Kijl < 0

2
50
Note that
r r XX j1 l1 r r XX j1 l1 j S aj kal k 1 K ajl 11
"
aj kal k 1K j Sjl
S21
0 Sjl 22
6 j1 6 6P P 4 r r
j1 l1
r P
aj kK ja S11
r r PP j1 l1
3 0
aj kal k 1Sjl 21
r r PP j1 l1
aj kal k 1Sjl 22
7 7 7 5
r P j1
j j kK a
S11 r r 0 4 P P a ka k 1Sjl j l 21
j1 l1
3 5
aj kal k 1Sjl 22
! S11 0 KakSak; ak 1 Kak S21 ak; ak 1 S22 k; k 1

where Kak is the same as in (10) and
2 Sak; ak 1
Then
r r r XXX i1 j1 l1 r 4P r P j1 l1
S11
0
r P r P j1 l1
3 5
aj kal k 1Sjl 21
aj kal k 1Sjl 22
ai kaj kal k 1Biu K j Sjl Bu akKakSak; ak 1 ai kaj kal k 1Dizu K j Sjl Dzu akKakSak; ak 1
r r r XXX i1 j1 l1
Therefore, (50) implies that (22) holds. Combining it and Lemma 5, it follows that for the singular perturbation parameter 2 0; , the system (9) is asymptotically stable and its H1 norm is less than c. Then the conclusion follows. h References
[1] W. Assawinchaichote, S.K. Nguang, H1 fuzzy control design for nonlinear singularly perturbed systems with pole placement constraints: an LMI approach, IEEE Transactions on Systems, Man and Cybernetics, Part B 34 (1) (2004) 579588. [2] W. Assawinchaichote, S.K. Nguang, P. Shi, H1 output feedback control design for uncertain fuzzy singularly perturbed systems: an LMI approach, Automatica 40 (12) (2004) 21472152. [3] S. Boyd, L.E. Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory, Society for Industrial and Applied Mathematics (SIAM), Philadelphia (PA), 1994. [4] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [5] L. Cao, H.M. Schwartz, Complementary results on the stability bounds of singularly perturbed systems, IEEE Transactions on Automatic Control 49 (11) (2004) 20172021. [6] Y. Chen, Y. Liu, Summary of singular perturbation modeling of multi-time scale power systems, in: Transmission and Distribution Conference and Exhibition, IEEE/PES, Asia and Pacic, 2005. [7] H.L. Choi, J.W. Son, J.T. Lim, Stability analysis and control of non-standard nonlinear singularly perturbed system, IEE Proceedings-Control Theory and Applications 153 (6) (2006) 703708. [8] J. Dong, G.H. Yang, Robust H1 control for standard discrete-time singularly perturbed systems, IET Control Theory and Applications 1 (4) (2007) 1141 1148. [9] J. Dong, G.H. Yang, H1 control for nonstandard discrete-time singularly perturbed systems, Automatica 44 (2008) 13851393. [10] E. Fridman, A descriptor system approach to nonlinear singularly perturbed optimal control problem, Automatica 37 (4) (2001) 543549. [11] P. Gahinet, A. Nemirovski, A.J. Laub, M. Chilali, LMI Control Toolbox, The MathWorks, Inc., Natick, MA, 1995. [12] G. Garcia, J. Daafouz, J. Bernussou, The innite time near optimal decentralized regulator problem for singularly perturbed systems: a convex optimization approach, Automatica 38 (8) (2002) 13971406. [13] R. Ghosh, S. Sen, K.B. Datta, Method for evaluating stability bounds for discrete-time singularly perturbed systems, IEE Proceedings-Control Theory and Applications 146 (2) (1999) 227233. [14] K. Tanaka, H.O. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality Approach, John Wiley & Sons Inc., New York, NY, USA, 2001. [15] C. Gong, B. Su, Robust L2 L1 ltering of convex polyhedral uncertain time-delay fuzzy systems, International Journal of Innovative Computing Information and Control 4 (4) (2008) 793802. [16] F.-H. Hsiao, J.-D. Hwang, S.-T. Pan, D-stability problem of discrete singularly perturbed systems, International Journal of Systems Science 34 (3) (2003) 227236.
132
[17] P.V. Kokotovic, J. OReilly, H.K. Khalil, Singular Perturbation Methods in Control: Analysis and Design, Academic Press Inc., Orlando, FL, USA, 1986. [18] T.-H.S. Li, J.-S. Chiou, F.-C. Kung, Stability bounds of singularly perturbed discrete systems, IEEE Transactions on Automatic Control 44 (10) (1999) 19341938. [19] T.H.S. Li, K.-J. Lin, Stabilization of singularly perturbed fuzzy systems, IEEE Transactions on Fuzzy Systems 12 (5) (2004) 579595. [20] W. Lin, C.I. Byrnes, H1 -control of discrete-time nonlinear systems, IEEE Transactions on Automatic Control 41 (4) (1996) 494510. [21] H. Liu, F. Sun, Y. Hu, H1 control for fuzzy singularly perturbed systems, Fuzzy Sets and Systems 155 (2) (2005) 272291. [22] I. Lizarraga, V. Etxebarria, Combined PD-H1 approach to control of exible link manipulators using only directly measurable variables, Cybernetics and Systems 34 (1) (2003) 1931. [23] D.S. Naidu, Singular perturbations and time scales in control theory and applications: an overview, Dynamics of Continuous Discrete and Impulsive Systems-Series B-Applications and Algorithms 9 (2) (2002) 233278. [24] V.R. Saksena, J. Cruz, Stabilization of singularly perturbed linear time-invariant systems using low-order observers, IEEE Transactions on Automatic Control 26 (2) (1981) 510513. [25] K. Tanaka, T. Ikeda, H.O. Wang, Fuzzy regulators and fuzzy observers: relaxed stability conditions and LMI-based designs, IEEE Transactions on Fuzzy Systems 6 (2) (1998) 250265. [26] C.-S. Ting, An observer-based approach to controlling time-delay chaotic systems via TakagiSugeno fuzzy model, Information Sciences 177 (20) (2007) 43144328. [27] H.D. Tuan, P. Apkarian, T. Narikiyo, Y. Yamamoto, Parameterized linear matrix inequality techniques in fuzzy control system design, IEEE Transactions on Fuzzy Systems 9 (2) (2001) 324332. [28] J. Wang, H. Li, Nonlinear PI control of a class of nonlinear singularly perturbed systems, IEE Proceedings-Control Theory and Applications 152 (5) (2005) 560566. [29] Y. Wang, Z. Sun, H1 control of networked control system via LMI approach, International Journal of Innovative Computing Information and Control 3 (2) (2007) 343352. [30] H.-N. Wu, K.-Y. Cai, Robust fuzzy control for uncertain discrete-time nonlinear Markovian jump systems without mode observations, Information Sciences 177 (6) (2007) 15091522.
C:\work2011\shamov\yulia7\c1fd00071c.pdf
View Online
133
Faraday Discussions
Accepted Manuscript
Downloaded by Washington University in St. Louis on 05 October 2011 Published on 14 July 2011 on http://pubs.rsc.org | doi:10.1039/C1FD00071C
This is an Accepted Manuscript, which has been through the RSC Publishing peer review process and has been accepted for publication. Accepted Manuscripts are published online shortly after acceptance, which is prior to technical editing, formatting and proof reading. This new free service from RSC Publishing allows authors to make their results available to the community, in citable form, before publication of the edited article. This Accepted Manuscript will be replaced by the edited and formatted Advance Article as soon as this is available.
More information about Accepted Manuscripts can be found on the Faraday Discussions website at http://blogs.rsc.org/fd/2011/04/05/am/ To cite this manuscript please use its permanent Digital Object Identifier (DOI), which is identical for all formats of publication. For more information about Faraday Discussions, visit the website at www.rsc.org/faraday or contact the Editor, Philip Earis, at faraday-rsc@rsc.org Please note that technical editing may introduce minor changes to the text and/or graphics contained in the manuscript submitted by the author(s) which may alter content, and that the standard Terms & Conditions and the ethical guidelines that apply to the journal are still applicable. In no event shall the RSC be held responsible for any errors or omissions in these Accepted Manuscripts or any consequences rising from the use of any information contained in them.
This manuscript will be presented at the forthcoming Faraday Discussion meeting
Register now to attend this exciting meeting!

www.rsc.org/fd154
Poster abstract deadline 17 June 2011 Early bird registration 17 June 2011 Registration deadline 15 July 2011
www.rsc.org/faraday
Registered Charity Number 207890
134
Protic pharmaceutical ionic liquids and solids: Aspects of Protonics

Jelena Stoimenovski,a Pamela M. Dean,a Ekaterina I. Izgorodina,a Douglas R. MacFarlane* a
a
View Online
School of Chemistry, Monash University, Clayton, Victoria 3800, Australia. *E-mail: Douglas.MacFarlane@monash.edu; Tel: +61 3 9905 4540
A series of new protic compounds based on active pharmaceutical ingredients have been synthesised and characterised. Some of the salts synthesised produced ionic liquids, while others that were associated with rigid molecular structures tended to produce high melting points. The protonic behaviour of these compounds was found to be a major determinant of their properties. Indicator studies, FTIR-ATR and transport properties (Walden plot) were used to probe the extent of proton transfer and ion association in these ionic liquids. While proton transfer was shown to have taken place in all cases, the Walden plot indicated strong ion association in the primary amine based examples due to hydrogen bonding. This was further explored via crystal structures of related compounds, which showed that extended hydrogen bonded clusters tend to form in these salts. These clusters may dictate membrane transport properties of these compounds in vivo.
1. Introduction
Many drug substances on the market are organic acids or bases that have an ability to form salts.1 Converting a drug from its acid or base form to a salt form can often allow a useful manipulation or optimisation of the drugs properties, such as its absorption, pharmacodynamics or pharmacokinetics.2, 3 In addition, by keeping an active ion constant and changing the counterion of the drug salt, either by combining with another active, or other biocompatible ion, one can further alter its chemical and biological properties without changing the structure of the active ingredient. Different salt forms of an active are now accepted to be different drugs, with individual chemical and biological profiles, resulting in varying clinical efficacies and safety.1 Turning the pharmaceutical salt into an ionic liquid (IL) is a further conceptual step that has been discussed recently in this regard.4 ILs are salts with melting points below 100 oC and are typically comprised of cations and anions that have either relatively low symmetry or delocalised charge; these structural properties tend to provide a low, or sometimes nonexistent melting point.5, 6 A wide range of different physical and chemical properties can be designed into the liquid by simply varying the cations and anions of which it is comprised.6 From solvents for chemical reactions,7 to electrolytes for electrochemical cells,8 ILs are now of intense interest in a range of applications in different areas. Recently, recognising that more than half of pharmaceuticals currently on the market are organic salts, ILs have also been proposed as a novel approach to drug design and optimisation in the pharmaceutical world.4, 9-11 In 2007, Rogers and co-workers, in conjunction with our group, filed a
Faraday Discussions Accepted Manuscript
135
series of patents in which the preparation of ILs from known active pharmaceutical ingredients (APIs) was described.12 The concept involved the replacement of what was usually a ballast counterion, for example sodium or bromide, with an alternative counterion that lowered the melting point of the salt such that it was liquid at ambient temperature, thus offering a range of different delivery modalities. If the counterion itself has active pharmaceutical properties then the compound can be considered a dual active, potentially presenting a synergistic combination of effects in a single compound. Other groups have been investigating ILs for related applications including anti-microbial13 and anti-bacterial activity.14-16 However, unlike the broader ILs field where the design or choice of cation and anion is, to an extent, flexible such that the desired low melting point can be achieved, this is not a case with pharmaceutical ILs. In this field the identity of at least one of the ions is specified, hence lowering the melting point of its salts becomes a challenging goal. This challenge has been explored over the last few years using an anti-crystal engineering approach,17 aided by a greater understanding of the lattice energy of crystalline organic salts as revealed by the Madelung constant.18 It was found, not surprisingly, that the extent of specific intermolecular interactions, such as hydrogen bonding, directly influences the phase behaviour of the product. Hence the advantageous combinations of active ion and counterion that can produce the low melting point required to form an IL will typically be those that avoid or disrupt all of these identifiable interactions. Protic Ionic Liquids the proton transfer conundrum Protic Ionic Liquids, PILs, are a subclass of ILs formed by reacting a Brnsted acid with a Brnsted base. Unlike aprotic ILs, PILs exhibit Brnsted acidity due to the exchangeable proton, and can hence be used as solvents for acid catalysed reactions;19-21 for the same reason, they are also found to be suitable proton conducting electrolytes for fuel cells.22 Given their ease of preparation, being simply the product of an acid (HA) - base (B) neutralisation reaction (Scheme 1), HA + B BH+ + A-
View Online
Scheme 1. Neutralisation reaction between an acid and a base this IL family is of significant potential in the broad, multifaceted pharmaceutical application. Many pharmaceuticals are in fact delivered as simple salts of organic bases, or metal ion salts of organic acids. However, the properties manifested by PILs depend predominantly on the degree of proton transfer from the acid to the base as well as the hydrogen bonding in which the proton is involved; the possible outcomes are summarised in Figure 1 and all of these outcomes have been observed in recent examples. For the PIL to be considered a pure salt and not a complex mixture of acid, base and salt, the extent of proton transfer should be at least 99 %.23, 24 Despite the fact that in aqueous solution the proton transfer between the acids and bases usually involved would be strongly complete, the situation is more complex in the neat salt for complex reasons. Yoshizawa et. al.25 originally revealed the problem by correlating observed boiling points with differences in pKaaq values (pKaaq) between the acid and the amine base. On the basis of a comparison of Walden
136
products (explained further below) it appeared that pKaaq 10 was needed for the full proton transfer to be observed (i.e. to achieve the situation in Figure 1c rather than the acid-base mixture of Figure 1a).25
View Online
Figure 1. The four possible outcomes in PILs (a) mostly un-ionised acid and base (b) hydrogen bonded acid-base pairs (no proton transfer) (c) mostly ionised PIL (proton transfer occurred, independent ions), (d) associated PIL (proton transfer occurred but ions associated through hydrogen bonding). To further investigate this phenomenon we explored a series of carefully chosen primary and tertiary amines of very similar pKaaq values.26 We found that restricted proton transfer was clearly exhibited by the tertiary amine base systems, and pKaaq 10 was indeed required to produce substantial proton transfer. On the other hand, the primary amine systems tended to be fully ionised even at pKaaq > 5. In other words, the primary amine based ILs tend to behave much more like aqueous systems. The strong difference in behaviour appears to be related to the hydrogen bond donor ability of the protonated base cation. In the tertiary amine case, only the single, transferred proton is available to hydrogen bond with the anion, whereas the primary base cation can offer multiple hydrogen bond sites. Free energy of solvation calculations from COSMO-RS confirmed this difference in the solvation environment in the two cases.26 However, clearly, additional intermolecular interactions stemming from long alkyl chains, --- interactions and other hydrogen bonding sites in either ion, for example OH functionality, can easily disturb these trends. Hence, introduction of such characteristics into tertiary amine based ILs could induce a stronger degree of proton transfer. These phenomena are crucial to attempts to form PILs based on APIs, since the lack of complete proton transfer may render the formulation nothing more than an acid/base/salt mixture. On the other hand, as we will show, the H-bonding introduced, involving both the transferred proton and also the residual conjugate base anion, can also produce strong ion-association in the PIL, potentially producing a further dimension of property variation in the product. We attempt to describe this complex set of interconnected phenomena surrounding the labile proton, its energetics and dynamics, in terms of a single unified concept protonics. Exploring these phenomena in the context of APIs is the main theme of this paper. Pharmaceutically Active PROTIC Ionic Liquids A number of groups including us have investigated the use of PILs of various compositions as an approach to pharmaceutically active ILs.27, 28 For example, Bica et. al.27 described a series of salts of aspirin, e.g. lidocainium acetylsalicylate, Tg=14C, where lidocaine is a topical anaesthetic and aspirin is a broad analgesic. In the
137
The three pharmaceutical bases were chosen on the basis of their pKaaq values and their structure (Figure 2). Tuaminoheptane, NTH2, a primary amine base with a pKaaq value of 10.50, is used as a nasal decongestant. Amantadine, NAH2, also a primary amine base with a pKaaq value of 10.10, is an antiviral/anti-Parkinsonian drug. 2Pyrrolidinoethanol, (EtOH)pyr, a generally recognised as safe compound, is a tertiary amine with a pKaaq value of 9.44.29
Benzoic acid (BzH) (1)
Salicylic acid (SalH) (2)
Gentisic acid (GenH) (3)
Tuaminoheptane (NTH2) (4)
Amantadine (NAH2) (5)
2-Pyrrolidinoethanol (EtOH)pyr (6)
Figure 2. Acids and bases used in this study
2. Experimental
2.1 General Tuaminoheptane (NTH2), amantadine (NAH2), 2-pyrrolidinoethanol ((EtOH)pyr) and gentisic acid (GenH) were purchased from Aldrich Chemical Company. Benzoic acid
present work we have focussed on salt formation and the issue of proton transfer in a number of families of pharmaceutically active acids and bases. For the purposes of a systematic study, a series of acids and bases (Figure 2) were selected with slight structural variations and physico-chemical properties, with the objective of understanding the factors controlling melting point, H-bonding and proton transfer. A library of nine compounds and four oligomeric PILs was synthesised and characterised. Benzoic, salicylic and gentisic acids were chosen for this study, as these are frequently encountered in pharmaceutical compounds/formulations. Benzoic acid, BzH, a common preservative and pharmaceutical aid (antifungal), is used in a range of products on the market including ointments, mouthwashes and cosmetics. A derivative of benzoic acid, salicylic acid (SalH), is a widely recognized and used keratolytic, also possessing anti-inflammatory, analgesic and antipyretic properties as a sodium salt. Gentisic acid, GenH, another benzoic acid derivative, is used in the pharmaceutical industry as an analgesic and anti-inflammatory.29
View Online
138
(BzH) was purchased from Sigma-Aldrich, while salicylic acid (SalH) was supplied by Chem Supply. The synthesis procedure and the characterisation of the resulting compounds can be found in the Electronic Supplementary Information. 2.2 Instrumental Nuclear Magnetic Resonance (NMR) data were recorded on either a Bruker Avance 400 (9.4 Tesla magnet) with a 5mm broadband autotunable probe with Z-gradients and BACS 60 tube autosampler or on a Bruker DPX-200 spectrometer. All samples were measured as solutions in deuterated dimethylsulfoxide supplied by Cambridge Isotope Laboratories. Multiplicities are reported according to their relative splittings, as a singlet (s), doublet (d), doublet of doublets (d(d)), triplet (t), pentet (p), sextet, multiplet (m) or a broad signal (b).
View Online
Electrospray Mass Spectrometry (ESI) was carried out on the Micromass Platform II API QMS Electrospray Mass Spectrometer with a cone voltage of 25 V or 35 V, using methanol as the mobile phase. Analyses were conducted in both positive (ESI+) and negative (ESI-) modes. Density data was collected using an Anton Paar DMA 5000 density meter. The density of the compounds was determined using the oscillating U-tube principle. The viscosity measurements were performed using a falling ball technique on an Anton Paar AMVn viscosity meter. Conductivity measurements were carried out on a locally designed dip cell probe containing two platinum wires covered in glass. 0.01M KCl solution at 25 oC was used to determine the cell constant. The conductivities were obtained by measuring the complex impedance spectra between 1 MHz and 0.01 Hz on a Solatron SI 1296 dielectric interface. Where crystals were obtained, the reflection intensity data were measured on a Bruker X8 APEX KAPPA CCD single crystal X-ray diffractometer ((EtOH)pyr+ Gen-; NAH3+ Gen-; NAH3+ Sal-; NTH3+ Gen-; (EtOH)pyr+ Bz-; (EtOH)pyr+ Sal-) and a Nonius KAPPA CCD single crystal X-ray diffractometer (NAH3+ Bz-), using graphitemonochromated Mo KR radiation () 0.71073 . Crystals were coated with Paratone N oil (Exxon Chemical Co.,TX, U.S.A.) immediately after isolation and cooled in a stream of nitrogen vapor on the diffractometer. Structures were solved by direct methods using the program SHELXS-97 and refined by full matrix least-squares refinement on F2 using SHELXL-97.25 All non-hydrogen atoms were revealed in the E-map and subsequent difference electron density maps and thus placed and refined anisotropically. All hydrogen atoms were observed in difference syntheses and were either refined isotropically or placed in geometrically idealized positions and constrained to ride on their parent atoms with C-H distances in the range 0.95-1.00 and Uiso(H)) xUeq(C), where x) 1.5 for methyl and 1.2 for all other C atoms.30 Specific details of refinement are contained in the relevant CIF files. The crystal structures of most of the synthesised salts are described here for the first time (CCDC Refcodes: 802629 for NTH3+ Gen-, 802627 for NAH3+ Gen-, 802625 for NAH3+ Sal-, 802628 for (EtOH)pyrH+ Bz-, 802630 for (EtOH)pyrH+ Sal- and 802624 for (EtOH)pyrH+ Gen-. The crystal structures for NAH3+ Bz- has been reported before and can be found in reference 31. However, we have solved the structure of this compound and submitted the report as part of the supplementary information.
139
View Online
Differential scanning calorimetry (DSC) scans were carried out at a heating/cooling rate of 10 oC/min in the range of -150 oC to 200 oC using a T.A Instruments Q100 differential scanning calorimeter. Thermal scans below room temperature were calibrated with the cyclohexane solid-solid transition and melting point at -87.0 C and 6.5 C respectively. Thermal scans above room temperature were calibrated with the indium, tin, lead and zinc. Transition temperatures are reported using the peak maximum of the thermal transition. The FTIR-ATR spectra for all the samples were obtained using Bruker IFS Equinox FTIR system, fitted with a Golden GateTM single bounce diamond. The FTIR spectrometer was connected to a computer equipped with OPUS 6.0 software, which was used for the analysis of spectra.
140
3. Results and Discussion

3.1 Protic Pharmaceutical Salts Table 1 summarises the protic salts made in this work. Of the thirteen compounds synthesised, nine satisfy the definition of an IL by having melting points below 100 o C. Five of these are room temperature ILs. Salts containing the NAH2 base all possess high melting points. ILs containing the (EtOH)pyrH+ ion all have melting points below 100 oC, while those based on the NTH3+ cation have no obvious trend amongst their melting points.
View Online
Gf =0 = Hf - TmSf
=> Tm = Hf /Sf
Eqn 1
Being a long-chain molecule, the NTH3+ cation has more degrees of rotational freedom that can become active on melting and hence exhibits a larger Sf,; this in turn leads to compounds with a lower melting point. On the other hand, NAH3+, a rigid multi-ringed structure, has fewer rotational degrees of freedom that can become active on melting, resulting in a smaller Sf and hence higher melting point of the resulting compounds. Thus, in order to synthesise low melting PILs, the more flexible anion/cation structures are desirable.
When reacted with the same acid, the two primary amines, NTH2 and NAH2, give rise to compounds with significantly different melting points. The compounds based on the NTH3+ cation, a seven carbon molecule, have melting points of 1212oC and below, while those based on the NAH3+ cation, a nine carbon molecule, possess melting points of 2162oC and above. These differences can be understood from a thermodynamic point of view. Given that the change in Gibbs free energy of fusion (Gf) at the melting point is zero by definition, the enthalpy (Hf) and the entropy (Sf) of fusion of the compound determine Tm (equation 1).
141
* Represents compounds with two moles of acid, in some cases two different acids, to one mole of base
Table 1. The protic pharmaceutical compounds prepared in this study including a number that are ILs
142
3.2 Oligomeric Compositions Another domain of PILs that has been explored recently are the protic oligomers, whereby several equivalents of acid were reacted with one equivalent of base to form H-bonded acid-oligomer based ILs (Scheme 2).32 For example, in the case of 2HA + B, the anionic species present appears to be [A-H-A]-. Proton transfer at this stoichiometry can be stronger than at the 1:1 stoichiometry. (A-H)x + B Hx-1(A-)x + B+H
View Online
Scheme 2. Anion oligomer based PIL formation32 Bica et. al. have reported pharmaceutical ILs based on such oligomeric ions; these compositions allow one to alter the dose of one active ingredient over the other active ingredient in the case of dual protic pharmaceutical ILs.33 Recognising that such 2:1 mixtures may also represent buffer type properties,34 they may also have a useful property in controlling pH. A similar approach was investigated in Table 1 in which a second mole of acid, in some cases a different acid, was incorporated into the formulation, in principle forming the dimer acid/anion species. In the case of the formulations involving two different acids HA1 and HA2, there is no intrinsic reason why the dimeric species should be uniquely A1-H-A2; the real situation will be a dynamic equilibrium between all of the possible species, with this overall composition. 3.3 Physical Properties, Proton Transfer and Ion Association As discussed above, certain important physicochemical properties of pharmaceutical acids/bases can be altered by converting them into salts of an appropriate counter ion. For example, in this study the salts of the originally water insoluble/slightly soluble acids and bases have been rendered water soluble in most cases. In the case of the dimer-acid based ILs, an additional acid molecule decreases the water solubility of the compounds. This in turn demonstrates the ease of water solubility manipulation, an important factor in drug delivery. The effects of proton transfer and ion association on physical properties are illustrated by the two salicylate compounds, one based on a primary amine, NTH3+ Sal-, and the other on a tertiary amine, (EtOH)pyrH+ Sal-. Both compounds have low melting points. Indicator studies, the basis of which was described in detail in our previous paper,26 were used to show that both the primary amine, NTH2, as well as the tertiary amine, (EtOH)pyr, have the capability to completely deprotonate the polyprotic indicator-acid, phenol red. This is shown in Figure 3 where the amines have both turned a fuchsia colour in the presence of the indicator acid, which originally exists as an orange coloured solution. One would observe a yellow colour if only a partial deprotonation occurred. The proton transfer behaviour of the (EtOH)pyr, a tertiary amine, is therefore different to the behaviour of the simple tertiary amine cases studied previously,26 presumably as a result of the presence of the hydroxyl group.
143
View Online
Figure 3. The primary amine, NTH2, and tertiary amine, (EtOH)pyr, with indicator phenol red Further evidence of proton transfer in this case is seen in the FTIR-ATR spectra of (EtOH)pyrH+ Sal-, NTH3+ Sal- and the starting materials (EtOH)pyr, NTH2 and SalH displayed in Figure 4. The spectra of the PILs (shown in red) are notably different to those of the starting materials. A peak at 1655 cm-1 in the SalH spectrum (shown in blue), corresponding to the C=O stretch of the acid, is absent in both PIL spectra, which is indicative of complete proton transfer in the ILs. Also, the disappearance of, or shift in, other peaks when compared to the original starting materials are all consistent with strong proton transfer in these PILs.
144
View Online
Figure 4. FTIR-ATR spectra of (a) (EtOH)pyrH+ Sal- (red), (EtOH)pyr (green) and SalH (blue) and (b) NTH3+ Sal- (red), NTH2 (green) and SalH (blue) The effect of ion association is subtly revealed by the bulk properties in these systems. The density data (Figure 5) reveals two distinctly different groups of liquids, where the tertiary amine based PILs (including the 2:1 mixtures) appear to be denser than the primary amine based examples. This in turn implies that ions in the tertiary amine based PILs pack more tightly than in the primary amine based PILs. As will be shown below, the latter have a tendency to form open hydrogen bonded structures that produce lower densities.
145
View Online
Figure 5. Density data of the synthesised PILs The ionic conductivity (Figure 6) also shows two distinctly different groups of compounds; the tertiary amine based PILs have a conductivity approximately one order of magnitude greater than the primary amine based ILs, despite the fact that both families appear to be substantially proton transferred. On the other hand, their viscosities (Figure 6) show no distinct differences. Of all the compounds analysed in this study, the primary amine based PIL, NTH3+ Sal-, displayed the highest viscosity followed by the tertiary amine based PIL, (EtOH)pyrH+ Sal-.
Figure 6. Conductivity (full symbols) and viscosity data (hollow symbols) of the synthesised PILs
146
The transport property and density data can be combined into the Walden plot, to provide a revealing perspective on proton transfer and ion association. The Walden plot25, 35, 36 is a method of comparing transport property data using the well-known Walden rule that states: = k Eqn 2
View Online
Figure 7. Walden plot of transport property data for the synthesised PILs. Each data set represents a temperature series for the compound above its melting point. Also included are some data points for a number of simple inorganic salts at temperatures just above their respective melting points.37
where is the molar conductivity and is the viscosity. An estimate of the constant (k) can be obtained from data for a 0.1M aqueous solution of KCl, which is thought to be a good example of a fully dissociated salt. The Walden rule then appears as a straight line on a plot of log(molar conductivity) versus log(1/viscosity). Previous discussion of this in the literature35 has shown that many ILs fall quite close to the Walden rule line, as do many molten salts. ILs that fall well below the line are thought of as exhibiting signs of ion association in a way that inhibits conduction, but not viscosity. Additionally, in the protic systems incomplete proton transfer can also produce transport properties that fall well below the line. As a reminder of the impact of the logarithmic scale of Figure 7, the dashed line indicates the situation where the molar conductivity is only 10% of that expected from the Walden rule. Such systems are clearly showing a low degree of proton transfer and/or strong signs of ionassociation.
147
3.4 Membrane Transport Properties One of the benefits of synthesising IL forms of known actives is the potential to modulate the membrane permeation properties of an active. Tuning and controlling membrane transport is the key to a variety of novel delivery methods including slow, continuous-release transdermal approaches. Recent work by us38 and others39 provides some insight into the membrane transport properties of these pharmaceutically active IL compounds. It is a well established fact that salts do not readily cross the lipid bilayer unless there are active transport mechanisms available. Hence, it is not immediately obvious that IL forms of an active will be readily membrane permeable. On the other hand, if the IL has a propensity to form neutral species, ion-pairs or clusters then membrane transport may be enhanced as compared to an inorganic salt of the active. To further investigate this hypothesis in respect of these protic examples, we have carried out model membrane transport studies following the method of Tantishaiyakul et. al..39 A full publication describing these results will be presented elsewhere38; we present here an overview of the trends observed. Typical data, for example for NTH3+ Sal- and its parent acid and base, as well as its simple salt forms that are used commercially, show that while the component acid is rapidly transported through the membrane, the sodium salt is not, nor is the sulfate salt (the form normally used pharmaceutically). On the other hand, the neat IL form, despite being a salt, is found to rapidly transport through the membrane over a period of one or two hours. Therefore the IL form is clearly more membrane permeable than the active ingredients that were the starting materials. Interestingly, the IL is not readily transported when dissolved in propylene glycol; this indicates that the situation is not simply one of providing an alternate, but independent, counter-ion. Indeed, Tantishaiyakul et. al. have speculated that ionpairing can play a significant role in the transport of these types of compounds.39 It is interesting to note that NTH3+ Sal- shows signs of strong ion pairing/aggregation by sitting low in the Walden plot and hence this aspect of the behaviour of these PILs may be of importance in their membrane transport properties. The nature of this ion pairing/aggregation is further explored in the next sections. It is also possible that reverse proton transfer to form the neutral acid and base at the point of dissolution into the membrane is the basis of transport partially the result of a relatively smaller pKa gap than the inorganic salts. At this stage it is not possible to discriminate between these hypotheses, however doing so will allow us to understand how
In the Walden plot in Figure 7 the PILs based on the NTH3+ cation display a significant deviation from the ideal line, whereas those containing the (EtOH)pyrH+ cation show better concordance with the Walden rule. The latter is consistent with the infra-red evidence, which indicates strong proton transfer and might be considered a normal example of a PIL. However, the NTH3+ salts appear surprisingly low on the Walden plot, given that the infra-red spectra indicated strong proton transfer. This perhaps indicates that these salts, although proton transferred, are highly associated in the liquid state. Ion pairs and larger clusters that have zero net charge are unable to conduct and therefore cannot contribute to the molar conductivity of the material. The NTH3 Sal case in Figure 7 is exhibiting only 1% of the conductivity that would be expected on the basis of the Walden rule; this indicates that the overwhelming majority of the ions in this case are tied up in ion pairs and clusters. We will return to the nature of association in these salts later, however, we will now present some data showing how association is of some importance in the membrane transport properties of these pharmaceutically active PILs.
View Online
148
modulating the counter acid/base in these systems can be used to control their interaction with membranes. 3.5 Ion association - what can we learn from crystal structures? Crystal structures can also provide some insight into the existence of proton transfer and ion association, and thereby serve as a guide to the existence of these phenomena in related liquid phases. Crystal structures for several of the compounds synthesised here have been solved and can be found in the supplementary information. The two ion-pairs shown in Figure 8 are the asymmetric units of the crystal structures of the tertiary amine based PIL, (EtOH)pyrH+ Sal-, and of the primary amine based PIL, NTH3+ Gen-. As can be seen from the displayed structures, the proton transfer has occurred in both cases. However, we should note that proton transfer in the solid state is likely to be stronger since it incurs a smaller entropic penalty than in the liquid and therefore produces a larger free energy change. This occurs because proton transfer in the neat liquid case causes a sharp loss of the configurational degrees of freedom that are characteristic of the liquid state and have a substantially negative S. Figure 8a also shows that the hydroxyl group of the (EtOH)pyrH+ cation stabilises a resonating carbonyl group in the Sal- anion, through hydrogen bonding, thus supporting the proton transfer. The O. . .H distances observed in this pair are 1.805 and 1.937 , comparable to the hydrogen bond distances in water.
View Online
(a)
(b)
Figure 8. Crystal structures (a) of the tertiary amine based IL, (EtOH)pyrH+ Sal(m.p. 492 oC); N-H. . .O and O-H. . .O distances are shown in red; (b) of the primary amine based salt, NTH3+ Gen- (m.p. 1212 oC) (N-H. . .O distance is shown in red) The asymmetric unit of NTH3+ Gen-, shown in Figure 8b, displays relatively short interionic N-H. . .O contacts of 1.811 , indicating strong hydrogen bonding between the ions. Additional O-H. . .O and C-H. . .O hydrogen bonding is also observed between the cations. Although the proton transfer is clearly complete in both cases, in accord with the FTIR-ATR data, a large difference between their positions in the Walden plot (> 1 orders of magnitude) demands explanation - an explanation that cannot be based on lack of proton transfer. Figure 9 provides a possible answer; NTH3+ Gen- organises itself in the crystal in a cyclic cluster-like complex consisting of two cations and two anions.
149
View Online
Figure 9. Cluster-like complex conformation of the primary amine based IL, NTH3+ GenIf a cyclic arrangement such as this, with its overall zero charge, is reasonably long lived in the liquid state then the molar conductivity will be lower than expected, despite the strong degree of proton transfer. Part of the origin of this cluster-like complex may lie in the inter-ring interactions as seen in Figure S2. . . . stacking between phenyl rings in the anions imposes a certain arrangement of ions in the crystal structure that results in a strong hydrogen-bonded lattice.
Figure 10. Cluster-like complex confirmation of the primary amine based IL, NAH3+ GenAnother example of this phenomenon was found in the analysis of the NAH3+ Gencrystal structure (m.p. = 2502 oC) containing the NAH3+ ion (Figure 10). In this case, a cyclic cluster-like complex is formed with both oxygens of the carboxylic group partaking in hydrogen bonding. Due to spatial constraints, one of the H. . .O bonds is longer than the other by 0.173 . A closer inspection of the packing diagram reveals that the cyclic arrangement extends beyond the cluster-like complex, with oxygens of the COO- group and the hydrogens of the NH3+ group forming distinct channels of hydrogen-bonded interactions (Figure S3). Extended associated structures would also have the effect of placing these compounds low on the Walden plot,
150
despite there being a high degree of proton transfer. Associations such as these would appear to also be crucial to the rapid membrane transport of these salts. Recently we have developed a generalised approach for the calculation of Madelung constants based on crystal structures.18 The Madelung constant, M, represents a measure of the net electrostatic energy in ionic solids and thus determines the stability of the ionic lattice with respect to an isolated ion pair. For stable lattices M is expected to be more than 1, with larger values indicating greater stability of the lattice. The Madelung constant can also be viewed as a geometric parameter; it provides us with a static perspective on cation-anion arrangements that might persist to an extent in the liquid state. Values of the Madelung constant closer to 1 represent ionic solids dominated by ion-pair interactions, whereas larger values indicate an extended lattice of inter-ionic interactions. In this study Madelung constants and electrostatic lattice energies (Table 2) were calculated for the PILs whose crystal structures were resolved in this work. Inspection of Table 2 reveals values that are relatively low; by comparison, a simple symmetrical organic salt such as [NMe4][BF4] has a Madelung constant close to that of NaCl (1.75). Ion association in the solid state is therefore lowering M considerably in all cases. Values for the (EtOH)pyrH+ based salts are closer to 1, indicating that the ion pairs are somewhat more isolated in this case. The electrostatic lattice energies also seem to generally follow the trend in measured melting points, with NAH3+ based salts having the largest lattice energies and highest melting points. As discussed elsewhere,40 this indicates that the electrostatic part of the lattice energy plays a dominant role in the overall energetics of melting of these compounds, despite the obvious influence of H-bonding and ion association in other properties. Table 2. Madelung constants, electrostatic lattice energies and melting points of protic pharmaceutical salts prepared in this work. M NTH3+ GenNAH3+ BzNAH3+ SalNAH3+ Gen(EtOH)pyrH Bz(EtOH)pyrH Sal(EtOH)pyrH Gen1.2307 1.2320 1.2566 1.2358 1.1731 1.0748 1.0831 Electrostatic Lattice energy, kJ mol-1 -551 -624 -625 -616 -546 -493 -487 Tm, C 121 256 216 250 28 49 97
View Online
3.6 Perspectives on Protonics In a number of fields of application, including ionic liquids and solids for fuel cell electrolytes,41 protein stabilising ionic liquids34 and water electrolysis42 the focus is on understanding the energetics and dynamics of protons and how they impact on the properties of the substance. Just as an understanding of the electronic structure of a molecule is vital to understanding its bonding and reactivity, we have found it useful
151
The distinction between the H-bonded acid-base complex and the H-bonded, protontransferred ion-pair becomes very slight in a PIL. In fact, it amounts to simply the matter of where the proton sits on a potential energy surface between the two heavy atoms involved (usually N and O). This is illustrated in Figure 11 for the case of acetic acid and methylpyrrolidine in the gas phase, where the proton transfer amounts to only a ~0.4A shift in an otherwise H-bonded complex. Charge is usually a significant distinction between these two states in a solvent based system, but becomes less so in the PIL, because H-bonding in the neutral complex involves some generation of a partial charge on each atom, while, on the other hand, H-bonding in the ion-pair acts to diminish the unit charge that notionally exists on the heavy atoms. In the case of acetic acid and methylpyrrolidine these minor changes in position and charge do not produce a sufficiently significant free energy change to create a substantial population of the proton transferred species, in either the gas phase or the ionic liquid.26 As we will discuss in more detail below, further solvation interactions are required to push this proton transfer along.
Figure 11. Ab initio calculations for acetic acid and methylpyrrolidine pKaaq = 5. The potential energy surface was generated as a relaxed scan of the O-H bond stretch at the MP2/6-31G+(d,p) level of theory. The optimised points on the PES were then
to adopt a holistic view of protonic behaviour in these contexts in order to fully understand its subtleties. Much of the discussion above has concerned the location of the proton in these protic pharmaceutical systems, the consequences of its transfer from acid to base and the impact of its H-bonding, or lack of, in determining properties. The neat IL provides a rather unique environment in which to observe proton transfer, since there is an almost complete lack of a dielectric solvent to shield the charges from one another and the liquid is so concentrated that the charges are never more than one ion-pair diameter apart. Unlike water and other solvents, in which the protonic properties are heavily dominated by the properties of the solvent, in PILs the IL species themselves are the only acid and base present. The proton energy levels involved are therefore more directly revealed and manipulated, but they are also strongly impacted by the molecular structure of the species involved.
View Online
152
improved at the MP2/aug-cc-pVTZ level of theory. All calculations were performed using GAUSSIAN 0343 Angell and coworkers44, 45 have discussed the use of a proton free-energy scale, first introduced by Gurney46, to assist in understanding the energetics in these materials. In this perspective, the proton site in an acid moiety is considered as one of high proton chemical potential, and the available site on the base moiety as being one of low proton chemical potential. The energy scale involved bears many features in common with electrochemical potential scales; in particular the chemical potential of the proton is the difference in free energies between that of the acid HA and its conjugate base A-. The highest occupied proton level (HOPL) in the system thus corresponds to the species having the highest free energy change available from a proton donation process HA A- + H+. The lowest unoccupied proton level (LUPL) similarly corresponds to the site offering the largest free energy change from a proton addition process. A number of scenarios are illustrated in Figure 12. Note that the LUPL of the base, B, is also the HOPL of the corresponding protonated species BH+.
View Online
Figure 12. Proton Energy levels in (a) a simple, solvent-free acid-base mixture (b) a more complex case in which either the acid or base is multi-protic and (c) a buffer case where the HOPL is ~50% occupied and hence can accept or donate protons to an added strong acid or base to create a buffer action.34 The impact of hydrogen bonding and other solvation interactions on these proton energy levels is the key aspect we focus upon here, being dominated by the absence of, or limited presence of, H-bond donating species in the acid or base. Recognizing that the proton energy levels in Figure 12 actually represent a free energy change in a process that involves both the protonated and the unprotonated species, it is necessary
153
to further deconstruct this diagram to explicitly identify the free energy of each species and then separately consider the effect of solvation on each. Figure 13 shows such an energy level diagram. The diagram begins by considering the isolated gas phase species and, as discussed above, in some cases there may be no proton transfer at all, as illustrated. Then, as shown in Figure 13, providing a source of solvation or additional hydrogen bonding, for example by a primary ammonium cation, impacts on the energy levels and hence ultimately on the proton transfer.26 In the case of the protic systems described here, other H-bonding sites in the molecules have additional solvating effects and hence further alter the respective positions of the proton energy levels available.
View Online
Figure 13. Free energy diagram showing the component molecular free energies that make up the proton energy levels in the gas phase and the effect of solvation on these. Where proton transfer does not take place in the liquid state, despite the expectation from aqueous pKa data that it should, another phenomenon related to H-bonding can occur: un-ionised acid-base complex formation, as shown in Figure 1b. A number of examples from the lidocaine:fatty acid family of liquids have been described recently in the context of pharmaceutically active compounds where the pKaaq values were ~3 and it was shown that proton transfer did not take place.47 In this situation the lack of H-bonding or solvation interactions strongly alters the positions of the individual ionised species in Figure 13 and thereby the HOPL and LUPL levels, to the point that proton transfer is not energetically favoured. However, a complete suppression of the melting point indicated formation, instead, of a hydrogen bonded acid-base complex. Such un-ionised complexes can be thought of as a liquid analogue of the co-crystals that are much studied in the pharmaceutical field. In the liquid co-crystal case the association of the two components produces a stabilising interaction in the liquid state that results in a sharp lowering of the melting point.
154
5. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. P. Heinrich Stahl, C. G. Wermuth and Editors, Handbook of Pharmaceutical Salts; Properties, Selection, and Use, VHCA and Wiley-VCH, 2008. S. M. Berge, L. D. Bighley and D. C. Monkhouse, J. Pharm. Sci., 1977, 66, 119. G. Davies, The Pharmaceutical Journal, 2001, 266, 322-323. J. Stoimenovski, D. R. MacFarlane, K. Bica and R. D. Rogers, Pharm. Res., 2010, 27, 521-526. P. M. Dean, J. M. Pringle and D. R. MacFarlane, Phys. Chem. Chem. Phys., 2010, 12, 9144-9153. R. D. Rogers and K. R. Seddon, Ionic Liquids: Industrial Applications to Green Chemistry, American Chemical Society, 2003. T. Welton, Chem. Rev., 1999, 99, 2071-2083. F. Endres, D. MacFarlane, A. Abbott and Editors, Electrodeposition from Ionic Liquids, Wiley-VCH, 2008. W. L. Hough, M. Smiglak, H. Rodriguez, R. P. Swatloski, S. K. Spear, D. T. Daly, J. Pernak, J. E. Grisel, R. D. Carliss, M. D. Soutullo, J. J. H. Davis and R. D. Rogers, New J. Chem., 2007, 31, 1429-1436. W. L. Hough and R. D. Rogers, Bull. Chem. Soc. Jpn., 2007, 80, 2262-2269. W. L. Hough-Troutman, M. Smiglak, S. Griffin, W. M. Reichert, I. Mirska, J. Jodynis-Liebert, T. Adamska, J. Nawrot, M. Stasiewicz, R. D. Rogers and J. Pernak, New J. Chem., 2009, 33, 26-33. Application: WO Pat., 2006-US39454 2007044693, 2007. A. Cieniecka-Roslonkiewicz, J. Pernak, J. Kubis-Feder, A. Ramani, A. J. Robertson and K. R. Seddon, Green Chem., 2005, 7, 855-862. J. Pernak and J. Feder-Kubis, Chem.-Eur. J., 2005, 11, 4441-4449. J. Pernak, K. Sobaszkiewicz and J. Foksowicz-Flaczyk, Chem.-Eur. J., 2004, 10, 3479-3485. J. Pernak, I. Goc and I. Mirska, Green Chem., 2004, 6, 323-329. P. M. Dean, J. Turanjanin, M. Yoshizawa-Fujita, D. R. MacFarlane and J. L. Scott, Cryst. Growth Des., 2009, 9, 1137-1145. E. I. Izgorodina, U. L. Bernard, P. M. Dean, J. M. Pringle and D. R. MacFarlane, Cryst. Growth Des., 2009, 9, 4834-4839. E. Janus, I. Goc-Maciejewska, M. Lozynski and J. Pernak, Tetrahedron Lett., 2006, 47, 4079-4083.
4. Conclusions A number of protic ionic liquids and salts based on pharmaceutical acids and bases have been synthesised and characterised. FTIR-ATR and Walden plot studies were used to assess the proton transfer and determine the degree of ionicity in the synthesised compounds. It was found that additional hydrogen bonding functional groups in the tertiary amine case can promote proton transfer. In the case of primary amines, formation of cluster-like species containing protonated base/deprotonated acid pairs can result in salts possessing higher melting points and decreased ionicity. These findings are supported by crystal structures and Madelung constant calculations. Nonetheless, for application purposes, PILs with lower ionicity are more desirable for drug delivery as they may cross the membrane barrier more easily than the more ionised species. Membrane studies are currently being carried out in order to further investigate this phenomenon and are showing promising results.
View Online
155
20. 21. 22. 23. 24. 25. 26. 27.

28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43.
Y. Du and F. Tian, J. Chem. Res., 2006, 486-489. Z. Duan, Y. Gu, J. Zhang, L. Zhu and Y. Deng, J. Mol. Catal. A Chem., 2006, 250, 163-168. M. A. B. H. Susan, A. Noda, S. Mitsushima and M. Watanabe, Chem. Commun. (Cambridge, U. K.), 2003, 938-939. B. Nuthakki, T. L. Greaves, I. Krodkiewska, A. Weerawardena, M. I. Burgar, R. J. Mulder and C. J. Drummond, Aust. J. Chem., 2007, 60, 21-28. D. R. MacFarlane and K. R. Seddon, Aust. J. Chem., 2007, 60, 3-5. M. Yoshizawa, W. Xu and C. A. Angell, J. Am. Chem. Soc., 2003, 125, 15411-15419. J. Stoimenovski, E. I. Izgorodina and D. R. MacFarlane, Phys. Chem. Chem. Phys., 2010, 12, 10341-10347. K. Bica, C. Rijksen, M. Nieuwenhuyzen and R. D. Rogers, Phys. Chem. Chem. Phys., 2010, 12, 2011-2017. K. Bica and R. D. Rogers, Chem. Commun. (Cambridge, U. K.), 2010, 46, 1215-1217. R. E. Buntrock, The Merck Index: An Encyclopedia of Chemicals, Drugs, and Biologicals, Fourteenth Edition, edited by Maryadele J. O'Neil, Patricia E. Heckelman, Cherie B. Koch, and Kristin J. Roman, 2007. G. M. Sheldrick, Acta Crystallogr., Sect. A Found. Crystallogr., 2008, A64, 112-122. W. N. Zheng and B. Wang, Acta Crystallogr., Sect. E Struct. Rep. Online, 2009, E65, o2769. K. M. Johansson, E. I. Izgorodina, M. Forsyth, D. R. MacFarlane and K. R. Seddon, PCCP, 2008, 10, 2972-2978. K. Bica and R. D. Rogers, Chem. Commun. (Cambridge, U. K.), 46, 12151217. D. R. MacFarlane, R. Vijayaraghavan, H. N. Ha, A. Izgorodin, K. D. Weaver and G. D. Elliott, Chem. Commun. (Cambridge, U. K.), 2010, 46, 7703-7705. D. R. MacFarlane, M. Forsyth, E. I. Izgorodina, A. P. Abbott, G. Annat and K. Fraser, PCCP, 2009, 11, 4962-4967. K. J. Fraser, E. I. Izgorodina, M. Forsyth, J. L. Scott and D. R. MacFarlane, Chem. Commun., 2007, 3817-3819. S. I. Smedley, The Interpretation of Ionic Conductivity in Liquids, Plenum Press, New York, 1980. J. Stoimenovski and D. R. MacFarlane, In preparation, 2011. V. Tantishaiyakul, N. Phadoongsombut, W. Wongpuwarak, J. Thungtiwachgul, D. Faroongsarng, K. Wiwattanawongsa and Y. Rojanasakul, Int. J. Pharm., 2004, 283, 111-116. U. L. Bernard, E. I. Izgorodina and D. R. MacFarlane, J. Phys. Chem. C, 2010, 114, 20472-20478. U. Rana and et.al., Chem. Commun., 2011, In press. B. Winther-Jensen, K. Fraser, C. Ong, M. Forsyth and D. R. MacFarlane, Adv. Mater. (Weinheim, Ger.), 2010, 22, 1727-1730. M. J. T. Frisch, G. W.; Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J. R.; Montgomery, J. A.; Jr., T. V.; Kudin, K. N.; Burant, J. C.; Millam, J. M.; Iyengar, S. S.; Tomasi, J.; Barone, V.; Mennucci, B.; Cossi, M.; Scalmani, G.; Rega, N.; Petersson, G. A.; Nakatsuji, H.; Hada, M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.; Klene, M.; Li, X.; Knox, J. E.; Hratchian, H. P.;
View Online
156
44. 45. 46. 47.
Cross, J. B.; Bakken, V.; Adamo, C.; Jaramillo, J.; Gomperts, R.; Stratmann, R. E.; Yazyev, O.; Austin, A. J.; Cammi, R.; Pomelli, C.; Ochterski, J. W.; Ayala, P. Y.; Morokuma, K.; Voth, G. A.; Salvador, P.; Dannenberg, J. J.; Zakrzewski, V. G.; Dapprich, S.; Daniels, A. D.; Strain, M. C.; Farkas, O.; Malick, D. K.; Rabuck, A. D.; Raghavachari, K.; Foresman, J. B.; Ortiz, J. V.; Cui, Q.; Baboul, A. G.; Clifford, S.; Cioslowski, J.; Stefanov, B. B.; G. Liu, A. L.; Piskorz, P.; Komaromi, I.; Martin, R. L.; Fox, D. J.; Keith, T.; Al-Laham, M. A.; Peng, C. Y.; Nanayakkara, A.; Challacombe, M.; Gill, P. M. W.; Johnson, B.; Chen, W.; Wong, M. W.; Gonzalez, C.; Pople, J. A., GAUSSIAN 03, Revision C.02;, 2004, Gaussian, Inc., Wallingford CT. J. A. Bautista-Martinez, L. Tang, J. P. Belieres, R. Zeller, C. A. Angell and C. Friesen, J. Phys. Chem. C, 2009, 113, 12586-12593. J.-P. Belieres and C. A. Angell, J. Phys. Chem. B, 2007, 111, 4926-4937. R. W. Gurney, Ionic Processes in Solution, 1953. K. Bica, J. Shamshina, W. L. Hough, D. R. MacFarlane and R. D. Rogers, Chem. Commun. (Cambridge, U. K.), 2011, 47, 2267-2269.
View Online

$binder 1

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

$binder 1

Uploaded by

Copyright:

Available Formats

C:\work2011\shamov\yulia7\00704983.

Consider the system described by the equations

01 B T A 0 BR01 H (S ) = Q 0 S T R01S 0(ABRBR01 S )T : S 0

5: :88 Q = 0496 0404 : :88 4:

The algebraic matrix equation (4) has a unique solution

01:7 > 0 1:8

00189286/98$10.00 1998 IEEE

AT P + P A 0 Ei (P B + LT )R01 (B T P + L)Ei + Q = 0 S = LEi 0 B T P Ek Q 0 S T R01 S > 0;

0 Ei (P B + LT )R01 (BT P + L)Ei

P (A 0 BR01 S ) + (A 0 BR01 S )T P 0 P BR01 BT P + Q 0 S T R01 S = 0:

0 (BT P + S )T R01 (BT P + S ) + S T R01 S < 0:

Q1 + P BR01 S + S T R01 B T P > 0:

min (Q1 ) : T max P BR01 S0 + S0 R01 B T P

P (A 0 BR01 S ) + (A 0 BR01 S )T P 0 P BR01 BT P + Q 0 S T R01 S = 0:

Optimization of Static Output Feedback Using Substitutive LMI Formulation

holds. Expressing (3) in terms of a matrix inequality, we have

8(P; F C2 ) + 0T 50 < 0 T 1 R 0 1 L 0 R01 B2 P 0= 5= 0 R M 0 F C2

Using the Schur complement [12], (6) is transformed into

0018-9286/04$20.00 2004 IEEE

Similar to (5), a sufficient matrix inequality for (10) is given by

P (k) 0 P (k01) F (k) 0 F (k01)

A 0 B2 F C 2 C1 0 D12 F C2 Acl Bcl : Ccl Dcl

B1 0 B2 F D21 D11 0 D12 F D21

kTzw (s)k1 < for > 0

T L(k+1) =R01 B2 P (k)

8(P; L)+31 (B 1 0 B 2 U )T P C1 0 D12 F C2 T F C2 0 R01 B2 P

3 3 0 Ir 3 D11 0 D12 F D21 0 Iq

where 3 indicates the elements that are readily inferred by symmetry.

2) Dual Form: (18) can be represented as another matrix inequality [14]

T Acl Q QAcl Ccl Q T Bcl

1 T 82 = P Acl + AT P + Ccl Ccl < 0 cl P P B1 >0 T B1 P X trX < 2 :

(28a) (28b) (28c)

1 T = D12 D12 > 0:

1 T 32 = C1T C1 + LT RL 0 C1T D12F C2 0 (F C2 )T D12 C1 :

kTzw (s)k2 < ;

1 T T 92 = Acl Q + QAcl + Bcl Bcl < 0 T Q QC1 >0 C1 Q Y trY < 2 :

(33a) (33b) (33c)

9(Q; K ) + 11 3 3 3 3 0 Iq 3 3 3 (C 1 0 V C 2 )Q T T 0 I 3 3 (B1 0 B2 F D21 ) (D11 0 D12 F D21 ) r 0 0 0W 01 3 (B2 F 0 QC2 W 01 )T (V 0 D12 F )T 0 0 0W 01 KT 1 N W N T 0 N W (B F ) T 0 B F W N T : 11 = 2 2

1 T = D21 D21 > 0:

T 0 QC2 W 01 < 0 T T B2 K 0 QC2 W 01 0W 01 1 B B T + KW K T 0 B (B F D )T 0 B F D B T : 12 = 1 1 1 2 21 2 21 1

Fig. 1. Convergence evaluation by norms of difference synthesis, standard form.

0 0 1 0 1 0 0 0:4422 0:1761 0 0 0 B2 = 3:5446 07:5922 0 0 0 05:52 4:49 0 0 0 0 0 0 1 0 0 C2 = 1 0 0 0 0 0 0 1 0 1 0 0 0:5 0 0 D12 = 1 0 0 1 0 0 1 0 0 :1 0 : 0 0 0 :1 =4 =3 H

norm of closed-loop system.

synthesis, standard form.

0 02970 4: = 00::3892 05004 :244

0:9266 00:01474 00:9622 01:387 00:02248 00:8448 0:1886 0:7135

0 05341 1: = 0::01364 01159 :551

The eigenvalues of the closed-loop system was 017.10, 00.1307 and 0 : 6j : .

norm of closed-loop system.

synthesis, dual form.

68:800 99:475 27:864 19:141

A Convergent Algorithm for Computing Stabilizing Static Output Feedback Gains

(3) i T P + L)T R01 (B T P + L)E + Q = 0 (4) Ei (B i

(5) (6) (7)

T +X =0 Y =YT >0 (13) T RK + Q = 0; P = P T > 0 BK ) + K

0018-9286/04$20.00 2004 IEEE

(xT Qx + uT Ru) dt subject to KV2 = 0:

E[x(0)x(0)T ] = X Q0 R>0

Define J as the cost to be minimized, and we get

Suppose there exists a matrix P

= P T > 0 so that d(xT P x)=dt = 0xT (K T RK + Q)x

(A 0 BKi )Yi + Yi (A 0 BKi )T + X = 0 (A 0 BKi )T Pi + Pi (A 0 BKi ) + KiT RKi + Q = 0:

min (Q1 ) : T max P BR01 S0 + S0 R01 B T P

E[x(0)x(0)T ] = X Q0 R>0

(1=2)vT Zv + 2vT (M2)W Z 01 W

) v Zv > v W M2 + (M2 )W v ) 0vT M1 W + W M1T v > vT W M2T + (M2 )W ) 0 > vT [W (M1 + M2 )T + (M1 + M2 )W ]v:

_ x = A x ( t) + B 1 ! ( t) + B 2 u ( t) 1x(t) + D11 !(t) + D12 u(t) z ( t) = C 2x(t) y ( t) = C

T T T T C2 = C21 C22 C23 D11 = D11 D12 = D12 :

A 0 B B B1 = 1 B2 = 2 0 0 C 0 C21 = [C2 0] C22 = [0 I ] C23 = [C2 A 0]