Fuzzy Sets and Systems 157 (2006) 1211 – 1228

www.elsevier.com/locate/fss

A new fuzzy Lyapunov function approach for a Takagi–Sugeno

fuzzy control system design
Bong-Jae Rhee, Sangchul Won∗
Department of Electrical and Electronic Engineering, Pohang University of Science and Technology (POSTECH), San 31, Hyojadong,
Namgu, Pohang, Gyungbuk 790-784, Republic of Korea

Received 24 March 2005; received in revised form 31 July 2005; accepted 21 December 2005
Available online 23 January 2006

Abstract
In this paper, a new fuzzy Lyapunov function approach is presented for a class of continuous-time Takagi–Sugeno fuzzy control
system. The proposed fuzzy Lyapunov function is formulated as a line-integral of a fuzzy vector which is a function of the state, and
it can be regarded as the work done from the origin to the current state in the fuzzy vector field. Unlike the approaches using a fuzzy
blending of multiple quadratic Lyapunov functions, the time-derivatives of membership functions do not appear in the proposed
approach. The effectiveness of the proposed approach is shown through numerical examples.
© 2006 Elsevier B.V. All rights reserved.

Keywords: Fuzzy Lyapunov function; Takagi–Sugeno fuzzy control system; Stability analysis; Fuzzy controller design

1. Introduction

Fuzzy control based on Takagi–Sugeno (TS) fuzzy model [14] has been used widely to control nonlinear systems
because TS fuzzy model efficiently represents a nonlinear system with a set of linear subsystems [5,6,8–13,15,16,19].
For the stability analysis of TS fuzzy control systems, many researchers have presented the conventional quadratic
Lyapunov function approaches to find a constant positive definite matrix of a quadratic Lyapunov function satisfying the
stability conditions of all subsystems [5,8,12,13,15–17,19]. Recently, in order to find more relaxed stability conditions,
a fuzzy Lyapunov function approach was introduced which utilized a fuzzy blending of multiple quadratic Lyapunov
functions [6,9–11]. However, in the stability analysis, time-derivatives of membership functions appear, and their upper
bounds have to be known prior to design a fuzzy controller. Since the time-derivatives are proportional to not only
states but also inputs, it is practically difficult to evaluate the upper bounds prior. Furthermore, the upper bounds of the
time-derivatives imply that the states are bounded.
In this paper, a new fuzzy Lyapunov function approach is presented for the stability analysis and the fuzzy controller
synthesis of a class of the continuous-time TS fuzzy control system. The proposed Lyapunov function is formulated
as a line-integral of a fuzzy vector along a path from the origin to the current state. Therefore, the time-derivatives of
membership functions do not appear in the stability analysis of the TS fuzzy control system, and the Lyapunov function
can be regarded as the work done in the fuzzy vector field. Because the vector is a fuzzy blending of the vectors which are

∗ Corresponding author. Tel.: +82 54 279 2221; fax: +82 54 279 8119.
E-mail addresses: [email protected] (B.-J. Rhee), [email protected] (S. Won).

0165-0114/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.fss.2005.12.020
1212 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

functions of the states in the subsystems, the Lyapunov function can be regarded as a fuzzy Lyapunov function. Moreover,
since the conventional quadratic Lyapunov function is a specific case of the proposed fuzzy Lyapunov function, the
proposed stability conditions are more relaxed than those of the conventional Lyapunov function approaches. To show
the effectiveness of the proposed fuzzy Lyapunov function approach, numerical examples are presented.
This paper is organized as follows. The new fuzzy Lyapunov function candidate for a class of the continuous-time
TS fuzzy control system is introduced in Section 2. The stability conditions for the autonomous TS fuzzy system is
presented in Section 3. The fuzzy controller design method is presented in Section 4. Finally, concluding remarks are
offered in Section 5.

2. Fuzzy Lyapunov function candidate

2.1. A Takagi–Sugeno fuzzy control system

Consider a class of the continuous-time TS fuzzy control system which can be described by the following fuzzy
rules:
Ri : If x1 (t) is M̄1i and · · · xj (t) is M̄ji · · · and xn (t) is M̄ni ,
then ẋ(t) = Ai x(t) + Bi u(t), i = 1, . . . , r, (1)
where Ri denotes the ith fuzzy rule, x(t) = [x1 (t), . . . , xn ∈ R is the state vector, u(t) = [u1 (t), . . . , um
(t)]T n
∈
(t)]T
Rm is the input vector, Ai ∈ Rn×n and Bi ∈ Rn×m are system matrices, M̄ji denotes an xj -based fuzzy set used for the
ith fuzzy rule, and r is the total number of fuzzy rules. Here, as premise variables, we use xj s, but not zj s which are
used in general description and are functions of x and u. Note
that xj s are chosen as the premise variables.
Let rj be the number of xj -based fuzzy sets. Then, r = nj=1 rj . However, in the ith fuzzy rule in (1), the fuzzy
set notation does not specify which of xj -based fuzzy sets is used. In order to construct the proposed fuzzy Lyapunov
function, a specific fuzzy set notation is required. Therefore, we rewrite (1) as



Ri : If x1 (t) is M1 i1 and · · · xj (t) is Mj ij · · · and xn (t) is Mn
in ,
then ẋ(t) = Ai x(t) + Bi u(t), i = 1, . . . , r (2)
where
ij specifies which xj -based fuzzy set is used in the ith fuzzy rule. For example,
ij = k means that the kth


fuzzy set among xj -based fuzzy sets is used in the ith fuzzy rule, therefore 1

ij
rj for any i. Let j ij (xj (t)) be the


membership function of Mj ij . The normalized membership function is



j ij (xj (t))
j ij (xj (t)) = rj
ij (3)

ij =1 j (xj (t))

and


0
j ij (xj (t))
1, (4)
rj


j ij (xj (t)) = 1. (5)

ij =1

Then, the membership function hi (x(t)) of the ith fuzzy rule becomes

n


hi (x(t)) = j ij (xj (t)), (6)
j =1

and

0
hi (x(t))
1, (7)

 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1213


r
hi (x(t)) = 1. (8)
i=1

For simplicity, we omit t in x(t) and hi (x(t)), then (2) can be rewritten as
ẋ = A(x)x + B(x)u, (9)
where

r
A(x) = hi (x)Ai , (10)
i=1


r
B(x) = hi (x)Bi . (11)
i=1

2.2. Fuzzy Lyapunov function candidate

To find the fuzzy Lyapunov function which does not associate with the time-derivatives of membership functions in
the stability analysis, consider a line-integral function such that

V (x) = 2 f () · d, (12)
(0,x)

where (0, x) is a path from the origin 0 to the current state x,  ∈ Rn is a dummy vector for the integral, f (x) ∈ Rn
is a vector which is a function of the state x, (·) denotes an inner product, and d ∈ Rn is an infinitesimal displacement
vector. In the literature [7], Lyapunov functions are physically regarded as energy-like functions in general. However,
if f (x) is regarded as a force vector at a state x, V (x) in (12) can be regarded as the work done from the origin to the
current state x in the force vector field. Since the work can be represented by the potential energy difference between the
origin 0 and the state x, V (x) can be regarded as an energy-like function. To be a Lyapunov function candidate, V (x)
has to satisfy the following conditions: (1) V (x) is a continuously differentiable function, (2) V (x) is positive-definite,
and (3) V (x) is radially unbounded [7]. However, if V (x) is dependent on the path (0, x), the condition (2) and (3)
would not be satisfied. Therefore, we need following lemma for the path-independent V (x).

Lemma 1. Let f (x) = [f1 (x), . . . , fn (x)]T . A necessary and sufficient condition for V (x) to be a path-independent
function is
jfi (x) jfj (x)
= (13)
jxj jxi
for i, j = 1, . . . , n.

Proof. It is the condition for a line-integral to be path-independent. See [20]. 

Consider the fuzzy vector f (x) with the following fuzzy rules.



Ri : If x1 is M1 i1 and · · · xj is Mj ij · · · and xn is Mn
in ,
then f (x) = Pi x, i = 1, . . . , r, (14)
where Pi ∈ R n×n
is a positive definite symmetric matrix such as
Pi = P̄ + Di > 0, (15)
⎡ ⎤
0 p12 · · · p1n
⎢ p12 0 · · · p2n ⎥
⎢ ⎥
P̄ = ⎢ . .. . . .. ⎥ , (16)
⎣ .. . . . ⎦
p1n p2n · · · 0
1214 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228
⎡
i1 ⎤
d11 0 ··· 0
⎢ 0 d
i2 ··· 0 ⎥
⎢ 22 ⎥
Di = ⎢ . .. .. . ⎥. (17)
⎣ .. . . .. ⎦

in
0 0 · · · dnn
Therefore, the off-diagonal elements of Pi ’s are same, but the diagonal elements are different according to the fuzzy
sets in the premise parts of the fuzzy rules. For the fuzzy rules with the different xj -based fuzzy sets, the j th diagonal
elements of Pi ’s are independent each other, but for the fuzzy rules with the same xj -based fuzzy sets, the j th diagonal
elements of Pi ’s are same. For any different two fuzzy rules, there is at least one different diagonal elements in Pi ’s.
The above fuzzy vector can be rewritten simply as
f (x) = P(x)x, (18)
where

r
P(x) = hi (x)Pi > 0. (19)
i=1

Theorem 1. The fuzzy vector f (x) in (18) satisfies the condition (13) in Lemma 1.

Proof. See Appendix A. 

Theorem 2. V (x) in (12) with the fuzzy vector f (x) in (18) is a Lyapunov function candidate.

Proof. See Appendix B. 

Remark 1. Since f (x) is a fuzzy vector, V (x) is regarded as a fuzzy Lyapunov function candidate. f (x), satisfying
the condition (13), is called as the conservative force in physics [18].

Remark 2. If all Di s are chosen as a common diagonal matrix D̄ such that P̄ + D̄ > 0, V (x) becomes a conventional
quadratic Lyapunov function which is V (x) = x T (P̄ + D̄)x. It can be easily shown from the proof of Theorem 2, and
it implies that the conventional quadratic Lyapunov function is a specific case of the proposed Lyapunov function.

3. Stability conditions for an autonomous TS fuzzy system

Consider an autonomous TS fuzzy system described by the following fuzzy rules:

Ri : If x1 is M1 i1 and · · · xj is Mj ij · · · and xn is Mn
in ,
then ẋ = Ai x, i = 1, . . . , r. (20)
The fuzzy system can be rewritten as
ẋ = A(x)x, (21)
where A(x) is defined in (10). To derive the stability conditions for the above fuzzy system, the following lemma is
necessary.

Lemma 2 (Tanaka et al. [13]). Let the maximum number of the fuzzy rules, which are activated simultaneously at any
instant, be s. Then, 1 < s
r. The membership functions of the fuzzy rules satisfy the following inequality:
r 
r 
(s − 1) h2i (x) − 2 hi (x)hj (x)0. (22)
i=1 i=1 j >i

for any x.
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1215

Now, the stability conditions are presented by the following theorem.

Theorem 3. The TS fuzzy system (20) is asymptotically stable if there exist P̄ , Di and X 0 satisfying

Pi = P̄ + Di > 0, (23)

Pi Ai + ATi Pi + (s − 1)X < 0, (24)

Pi Aj + ATj Pi + Pj Ai + ATi Pj − 2X
0, i<j (25)

for i, j = 1, . . . , r. Here, P̄ and Di ’s are defined in (16) and (17) respectively.

Proof. The time-derivative of V (x) is

V̇ (x) = x T {P(x)A(x) + AT (x)P(x)}x. (26)

Define M(x) = P(x)A(x) + AT (x)P(x). Then,


r 
r
M(x) = hi (x)hj (x)(Pi Aj + ATj Pi )
i=1 j =1


r 
r 
= h2i (x)(Pi Ai + ATi Pi ) + hi (x)hj (x)(Pi Aj + ATj Pi + Pj Ai + ATi Pj ). (27)
i=1 i=1 j >i

Using (24), (25) and Lemma 2,


r 
r 
M(x)
h2i (x)(Pi Ai + ATi Pi ) + 2 hi (x)hj (x)X
i=1 i=1 j >i
r

h2i (x)(Pi Ai + ATi Pi + (s − 1)X)
i=1
< 0. (28)

Therefore,

V̇ (x) = x T M(x)x < 0 (29)

for all x  = 0. 

Remark 3. Unlike the fuzzy Lyapunov function approach in [6,9–11], the upper bound of the time-derivatives of
membership functions is not required in the proposed stability conditions. If the constraint that D1 = D2 = · · · = Dr
is added, Theorem 3 is equivalent with the stability conditions of the conventional Lyapunov function approaches
[8,13,16,17,19]. Therefore, the proposed stability conditions are more relaxed than those of the conventional Lyapunov
function approaches.

Remark 4. For the simple derivations of the stability conditions, the relaxation method presented in [13] is applied.
However, for the better relaxation, the other methods in [8,17,19] can be applied, and the corresponding stability
conditions can be derived in similar ways.

Remark 5. Since the stability conditions are expressed in terms of LMIs, Pi ’s can be found numerically using convex
optimization techniques [1].

To show the effectiveness of the proposed stability conditions, an example is presented.

1216 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

Fig. 1. The fuzzy vector f (x) for Example 1.

200

160

120
b

80

40

0
-10 -8 -6 -4 -2 0
a

Fig. 2. The stability-guaranteed region of the proposed fuzzy Lyapunov function approach.

Example 1. Consider the following TS fuzzy system:

R1 : If x1 is M11 and x2 is M21 then ẋ = A1 x,

R2 : If x1 is M11 and x2 is M22 then ẋ = A2 x,
R3 : If x1 is M12 and x2 is M21 then ẋ = A3 x,
R4 : If x1 is M12 and x2 is M22 then ẋ = A4 x, (30)
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1217

200

160

120

b
80

40

0
-10 -8 -6 -4 -2 0
(a) a

200

160

120
b

80

40

0
-10 -8 -6 -4 -2 0
(b) a

Fig. 3. The stability-guaranteed region of the fuzzy Lyapunov function approach in [11]: (a) |dhi /dt|
0.85 for i = 1, . . . , 4; (b) |dhi /dt|
2.57
for i = 1, . . . , 4.

where
−5 −4 −4 −4
A1 = , A2 = ,
−1 a 1
5 (3b− 2) 1
− 4)
5 (3a
−3 −4 −2 −4
A3 = , A4 = , (31)
1
5 (2b− 3) 1
5 (2a − 6) b −2
for the given constants a and b. Then,

11 = 1,
12 = 1,
21 = 1,
22 = 2,
31 = 2,
32 = 1,
41 = 2,

42 = 2. (32)
For i = 1, 2, the normalized membership functions of Mi1 and Mi2 are given as
⎧
⎨ (1 − sin(xi ))/2 for |xi |
/2,
1i (xi ) = 0 for xi > /2, (33)
⎩
1 for xi < −/2,
1218 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

200

160

120

b
80

40

0
-10 -8 -6 -4 -2 0
a

Fig. 4. The stability-guaranteed region of the conventional Lyapunov function approaches in [8,13,16,17,19].

6
16

16
16

3
4
4

1
x2

0
16

4 1
16
4

-3 16
16

-6
-2 -1 0 1 2
x1

Fig. 5. The contour plot of V (x) and a state-trajectory for a = −3 and b = 50.

2i (xi ) = 1 − 1i (xi ). (34)

Then, the membership functions of the fuzzy rules are

h1 (x) = 11 (x1 )12 (x2 ), h2 (x) = 11 (x1 )22 (x2 ), h3 (x) = 21 (x1 )12 (x2 ), h4 (x) = 21 (x1 )22 (x2 ). (35)
The fuzzy vector f (x) for this example is shown in Fig. 1, where the relationship between
ij , membership functions,
and the diagonal elements of Pi are illustrated well.
Using the LMI control toolbox of MatLab [4], the stability was investigated for various values of a and b.
Fig. 2 shows the stability-guaranteed region of the proposed approach. Fig. 3 shows the regions of the fuzzy Lya-
punov function approach in [11] with two different bounds of dhi /dt’s. It shows that the region is expanded as the
bound is reduced, but the stability-guaranteed region in the state space is shrunken because dhi /dt’s are proportional
to the states. Fig. 4 shows the region of the conventional Lyapunov function approaches in [8,13,16,17,19]. As we can
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1219

see, Theorem 3 provides relaxed stability conditions without the upper bound of dhi /dt’s. For a = −3 and b = 50,
Pi ’s are found as
3.2525 0.2290 3.2525 0.2290 7.1747 0.2290
P1 = , P2 = , P3 = ,
0.2290 1.2305 0.2290 0.5767 0.2290 1.2305
7.1747 0.2290
P4 = . (36)
0.2290 0.5767

Fig. 5 shows the contour plot of V (x) and a state-trajectory from an initial state toward the origin.

4. Fuzzy controller design

The fuzzy controller is designed using the concept of parallel distributed compensation (PDC) which shares the
fuzzy rules with the TS fuzzy control system as follows [16]:



Ri : If x1 is M1 i1 and · · · xj is Mj ij · · · and xn is Mn
in ,
then u = Fi x, i = 1, . . . , r (37)
where Fi ∈ Rm×n is a local feedback gain matrix. Let


r
F(x) = hi (x)Fi . (38)
i=1

Then, the control input can be rewritten simply as

u = F(x)x, (39)

and the closed-loop system becomes

ẋ = {A(x) + B(x)F(x)}x. (40)

Define

Giii = Pi (Ai + Bi Fi ), (41)

Giij = 13 {Pi (Ai + Aj + Bi Fj + Bj Fi ) + Pj (Ai + Bi Fi )}, (42)

Gij k = 16 {Pi (Aj + Ak + Bj Fk + Bk Fj ) + Pj (Ai + Ak + Bi Fk + Bk Fi )

+Pk (Ai + Aj + Bi Fj + Bj Fi )}. (43)
Then, the stability conditions are derived by the following theorem.

Theorem 4. The TS fuzzy control system (2) with the fuzzy controller (37) is asymptotically stable if there exist P̄ , Di ,
Fi , and X 0 satisfying

Pi = P̄ + Di > 0, (44)

Giii + GTiii + (s − 1)X < 0, (45)

Giij + GTiij + 13 (s − 3)X
0, i  = j, (46)

Gij k + GTij k − X
0, i<j <k (47)

for i, j , k = 1, . . . , r. Here, P̄ and Di ’s are defined in (16) and (17), respectively.

1220 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228


Proof.. Because rk=1 hk (x) = 1, (22) in Lemma 2 can be rewritten as
⎛ ⎞ 

r r  
r
⎝(s − 1) h2i (x) − 2 hi (x)hj (x)⎠ hk (x)
i=1 i=1 j >i k=1


r 
r  
r 
r
= (s − 1) h3i (x) + (s − 1) h2i (x)hk (x) − hi (x)hj (x)hk (x)
i=1 i=1 k=i i=1 j =i k=1


r 
r  
r 
= (s − 1) h3i (x) + (s − 3) h2i (x)hj (x) − 6 hi (x)hj (x)hk (x)
i=1 i=1 j =i i=1 j >i k>j

0. (48)
The time-derivative of V (x) is
V̇ (x) = x T {P(x)A(x) + AT (x)P(x) + P(x)B(x)F(x) + F T (x)B T (x)P(x)}x. (49)
Define M(x) = P(x)A(x) + AT (x)P(x) + P(x)B(x)F(x) + F T (x)B T (x)P(x) and Ḡij k = Pi (Aj + Bj Fk ). Then,
Gij k = Gikj = Gj ik = Gj ki = Gkij = Gkj i
= 16 (Ḡij k + Ḡikj + Ḡj ik + Ḡj ki + Ḡkij + Ḡkj i ), (50)
and M(x) can be rewritten as
r  r  r
M(x) = hi (x)hj (x)hk (x)(Ḡij k + ḠTij k )
i=1 j =1 k=1


r 
r 
r
= hi (x)hj (x)hk (x)(Gij k + GTij k )
i=1 j =1 k=1

r 
r 
r  
= h3i (x)(Giii + GTiii ) + 3 h2i (x)hj (x) Giij + GTiij
i=1 i=1 j =1
j  =i

r 
+6 hi (x)hj (x)hk (x)(Gij k + GTij k ). (51)
i=1 j >i k>j

Using (45)–(47) and (48),

r 
r  
r 
M(x)
h3i (x)(Giii + GTiii ) + 3 h2i (x)hj (x)(Giij + GTiij ) + 6 hi (x)hj (x)hk (x)X
i=1 i=1 j =i i=1 j >i k>j
r 
r 

h3i (x)(Giii + GTiii + (s − 1)X) + 3 h2i (x)hj (x){(Giij + GTiij ) + 13 (s − 3)X}
i=1 i=1 j =i
< 0. (52)
Therefore,
V̇ (x) = x T M(x)x < 0 (53)
for all x  = 0. 

Remark 6. From (49), it can be said that {A(x), B(x)} has to be stabilizable for any x = 0 for the existence of the
feedback gains. It implies that if {A(x), B(x)} is uncontrollable for an x, the uncontrollable states have to be stable.

From the above theorem, it can be said that Remark 3 is also effective in the stabilization problem. However, because
the stability conditions of Theorem 4 are unfortunately BMIs (bilinear matrix inequalities), it is difficult to find Pi ’s
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1221

Im(s)

θ
−β 0 Re(s)

Fig. 6. LMI region for pole placement in s-domain [2,3].

25

20

15
b

10

0
-20 -15 -10 -5 0 5 10
a

Fig. 7. The stabilizable region of the proposed fuzzy Lyapunov function approach.

25

20

15
b

10

0
-20 -15 -10 -5 0 5 10
a

Fig. 8. The stabilizable region of the conventional Lyapunov function approaches [8,13,16,17, 19].
1222 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

6
1 0.1

6
0.04

0.1
1
x2
0 0.1
6
0.0
4
0.0
0.01

0.04 6
0.1
-1

0.16

-2
-2 -1 0 1 2
x1

Fig. 9. The contour plot of V (x) and the state-trajectories toward the origin for a = 2 and b = 1.

and Fi ’s in general. Therefore, we present an algorithm to find Pi and Fi ’s using convex optimization techniques.
The key idea of the algorithm is simple. Firstly, using a locally-available conventional quadratic Lyapunov function
approach, we find the feedback gains which guarantee the local stability of the fuzzy subsystems, not the global stability.
Then, we select them as the candidates of Fi ’s. Secondary, we replace Fi ’s in (45)–(47) with the candidates. Then,
the inequality (45)–(47) become LMIs, and we check whether there exist feasible Pi ’s using convex optimization
techniques. If there exist, the candidates are selected as the feedback gain Fi ’s. Otherwise, new candidates are found
which are more robust, then because the local stability-guaranteed region by the candidates is expanded in the state-
space, the possibility of the candidates to satisfy the stabilization conditions of Theorem 4 increases. The algorithm
runs iteratively until the feedback gains Fi ’s are found, or it stops when the candidates satisfying the given robustness
constraint are not found because the candidates satisfying more robustness constraint cannot exist. In general, even
though there can exist multiple solutions satisfying the local stability conditions, only one solution can be found using
LMI toolboxes. Therefore, using the additional constraint on the upper bound of the difference between positive definite
matrices of the locally-available Lyapunov functions, more candidates can be found. The upper bound is reduced as the
algorithm runs.
The detailed algorithm is introduced in Appendix C. To show the effectiveness of the proposed fuzzy controller
design method, an example is presented.

Example 2. For a TS fuzzy control system, let us design the fuzzy controller to place all poles of the subsystems in
the shaded region of Fig. 6 in s-domain. The region is described as the intersection of a disk centered at the origin
with radius r̄ = 10, a sector form with an angle  = /4, and a half-plane Re(s) < − = −5. We modify the
searching algorithm for the pole-placement constraint, and the algorithm is presented in Appendix D. The TS fuzzy
control system has the following fuzzy rules:
R1 : If x1 is M11 then ẋ = A1 x + B1 u,
R2 : If x1 is M12 then ẋ = A2 x + B2 u, (54)
where
2 −10 a −5 1 b
A1 = , A2 = , B1 = , B2 = (55)
2 0 1 2 1 2

for the given constants a and b. The normalized membership functions of M11 and M12 are given in (33) and (34). Then,

11 = 1,
21 = 2,
12 =
22 = 1, (56)

 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1223

and P1 and P2 are

1 p
d11 2 p
d11
12 12
P1 = 1 , P2 = 1 . (57)
p12 d22 p12 d22
The stabilizable regions for the various values of a and b are found numerically using the LMI control toolbox of
MatLab [4]. Fig. 7 shows the stabilizable region of the proposed fuzzy Lyapunov function approach. Fig. 8 shows the
stabilizable region of the conventional Lyapunov function approaches [8,13,16,17,19]. The fuzzy Lyapunov function
approach in [11] cannot stabilize the system for any a, b and upper bound of the time-derivatives of membership
functions, even without the pole-placement constraint. As we can see, the proposed fuzzy controller design method is
more feasible. For a = 2 and b = 1, Fi ’s are found as

F1 = [6.5717 − 22.3122], F2 = [3.5460 − 9.0271], (58)

and Pi ’s as
0.0988 −0.1216 0.1439 −0.1216
P1 = , P2 = . (59)
−0.1216 0.1918 −0.1216 0.1918
Fig. 9 shows the contour plot of V (x) and the state-trajectories from two initial states toward the origin.

5. Conclusion

In this paper, a new fuzzy Lyapunov function approach is presented for a class of continuous-time TS fuzzy control
system. We propose a new fuzzy Lyapunov function which can be regarded as the work done from the origin to
the current state in a fuzzy force vector field. The stability conditions are derived without the upper bound of the
time-derivatives of membership functions. The effectiveness of the proposed approach is shown through numerical
examples.
As further studies, the research applying the proposed approach to the general TS fuzzy control system is being
investigated.

Appendix A. Proof of Theorem 1

P(x) in (19) can be rewritten as

⎡ ⎤
d11 (x) p12 · · · p1n
⎢ p21 d22 (x) · · · p2n ⎥
⎢ ⎥
P(x) = ⎢ . .. .. .. ⎥ (60)
⎣ .. . . . ⎦
pn1 pn2 · · · dnn (x)
where pkl = plk for k  = l and

r

ik
dkk (x) = hi (x)dkk . (61)
i=1

Because of (5) and (6),

⎡⎛ ⎞ ⎤
 r 
n
dkk (x) = ⎣⎝ 
ij (xj )⎠ d
ik ⎦
j kk
i=1 j =1
⎡⎛ ⎞ ⎤

r1 
rk 
rn 
n
= ··· ··· ⎣⎝ j j (xj )⎠ dkkk ⎦
1 =1 k =1 n =1 j =1
1224 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228
⎛ ⎞ ⎛ ⎞
rj

rk 
n 
=⎝ k k (xk )dkkk ⎠ ⎝ j j (xj )⎠
k =1 j  =k j =1


rk
= k k (xk )dkkk . (62)
k =1

Since dkk (x) is a function of only xk , dkk (x) is rewritten as dkk (xk ). Then, the ith element of f (x) is

n
fi (x) = dii (xi )xi + pik xk . (63)
k=i

For j  = i, taking a partial derivative with respect to xj , it becomes

jfi (x) jdii (xi )
= xi + pij = pij . (64)
jxj jxj
Similarly,
jfj (x)
= pj i . (65)
jxi
Since pij = pj i , f (x) satisfies the condition in (13). 

Appendix B. Proof of Theorem 2

To be a Lyapunov function candidate, V (x) in (12) satisfies the conditions as follows:

Condition (1): Let g(x) = A(x)x + B(x)u, and Lg V (x) denotes the Lie derivative of the function V (x) with respect
to the vector field g(x). Then, the time-derivative of V (x) becomes
V̇ (x) = Lg V (x)
= 2f T (x)g(x)
= f T (x){A(x)x + B(x)u} + {A(x)x + B(x)u}T f (x). (66)
It shows that V (x) is a continuously differentiable function.
Conditions (2) and (3): Because V (x) is path-independent, by the substitution of  = x where ∈ R, V (x) is
rewritten as
 1  1
V (x) = 2 f ( x)x d = 2
T
x T P( x)x d . (67)
0 0

Here, the substitution implies that the line-integral of (12) is done along the straight line from the origin to the state x.
Since Pi is a positive definite matrix, there exist P and P such that
P Pi P > 0 (68)
for i = 1, . . . , r. Then,
P P(x)P > 0 (69)
for any x, and
 1    
V (x)
2 x Px d = 2
T
d x T P x = x T P x. (70)
0 0

Similarly,
V (x)x T P x. (71)
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1225

Therefore,
x T P x V (x)x T P x. (72)
It shows that V (0) = 0, V (x) > 0 for x  = 0, and V (x) → ∞ as |x| → ∞. 

Appendix C. The algorithm searching the feedback gain matrices

The algorithm is composed of two phase. In the first phase, we find the candidates of Fi ’s using a locally-available
conventional quadratic Lyapunov function approach. Note that the approach is not available globally. In the second
phase, by replacing the candidates with Fi ’s in Theorem 4, the stabilization conditions become LMIs, and we check
the availability of the candidates.
Let F̂i be the candidate of Fi . To find F̂i , consider the dynamic equation (9) of the TS fuzzy control system. In
vicinity of any x0 ∈ Rn , the equation can be rewritten as
ẋ = {A(x0 ) + A (x)}x + {B(x0 ) + B (x)}u, (73)
where A (x) = A(x) − A(x0 ) and B (x) = B(x) − B(x0 ), and they can be regarded as the parameter variations.
Then, it is easy to show that there exist A and B such that  A (x)2
A and  B (x)2
B . Define

r
Q(x0 ) = hi (x0 )Qi , (74)
i=1


r
Y(x0 ) = hi (x0 )Yi , (75)
i=1

where Yi ∈ Rm×n , and Qi ∈ Rn×n is a positive definite matrix. Choose a local feedback gain matrix as
F̄(x0 ) = Y(x0 )Q−1 (x0 ). (76)
Then, the closed-loop system becomes
ẋ = {A(x0 ) + A (x) + (B(x0 ) + B (x))F̄(x0 )}x. (77)
Consider a locally-available Lyapunov function candidate
V (x) = x T Q−1 (x0 )x. (78)
Define

M (x, x0 ) = A (x)Q(x0 ) + Q(x0 ) A (x) +

T
B (x)Y(x0 ) + Y
T
(x0 ) T
B (x) (79)
and
M(x0 ) = A(x0 )Q(x0 ) + Q(x0 )AT (x0 ) + B(x0 )Y(x0 ) + Y T (x0 )B T (x0 ) + M (x, x0 ). (80)
Then, the time-derivative of V (x) is
V̇ (x) = x T Q−1 (x0 )M(x0 )Q−1 (x0 )x. (81)
Because C T D + D T C
C T C + −1 D T D for any C, D ∈ Rm×n and > 0, M (x, x0 ) becomes
 
2 2 −1
M (x, x0 )
A + B I + (Q(x0 )Q(x0 ) + Y T (x0 )Y(x0 )) (82)

For a given 0, it is easy to find the upper bounds of A and B satisfying
2 2 −1
( A + B )I + (Q(x0 )Q(x0 ) + Y T (x0 )Y(x0 ))
Q(x0 ). (83)
1226 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

Then, the more positive is given, the more positive upper bounds of A and B are available. For A and B less than
the upper bounds, M(x0 ) becomes

M(x0 )
A(x0 )Q(x0 ) + Q(x0 )AT (x0 ) + B(x0 )Y(x0 ) + Y T (x0 )B T (x0 ) + Q(x0 ). (84)

Then, it can be rewritten as


r
M(x0 )
h2i (x0 )(Ai Qi + Qi ATi + Bi Yi + YiT BiT + Qi )
i=1
r 
r
+ hi (x0 )hj (x0 )(Ai Qj + Aj Qi + Qj ATi + Qi ATj
i=1 j >i
+Bi Yj + Bj Yi + YjT BiT + YiT BjT + (Qi + Qj )). (85)

If there exist Qi > 0, Yi , and X̄ 0 satisfying

Ai Qi + Qi ATi + Bi Yi + YiT BiT + (s − 1)X̄ + Qi < 0 (86)

 
Ai Qj + Aj Qi + Qj ATi + Qi ATj

2X̄, j >i (87)
+Bi Yj + Bj Yi + YjT BiT + YiT BjT + (Qi + Qj )

for i, j = 1, . . . , r, using Lemma 2, M(x0 ) becomes

 r 
r 
r
M(x0 )
h2i (x0 )(Ai Qi + Qi ATi + Bi Yi + YiT BiT + Qi ) + 2 hi (x0 )hj (x0 )X̄
i=1 i=1 j >i
r

h2i (x0 )(Ai Qi + Qi ATi + Bi Yi + YiT BiT + (s − 1)X̄ + Qi )
i=1
< 0. (88)
It shows that V̇ (x) < 0 for x  = 0 in vicinity of x0 if there exist Qi ’s and Yi ’s satisfying (86) and (87). Moreover, it
implies that, when more positive is given, the system (73) becomes more robust against the parameter variations in
vicinity of x0 , but Qi ’s and Yi ’s become less feasible. If Qi ’s and Yi ’s are infeasible, they are infeasible for any more
positive . On the other hand, when Qi ’s and Yi ’s are feasible, there can exist multiple solutions of Qi ’s and Yi ’s, but
only one solution can be generally found using LMI toolboxes. Therefore, using the constraints on Qi ’s as

|Qi − Qj |
(Qi + Qj ), (89)

where  0, we search the multiple solutions. Here, (89) can be regarded as the constraint on the difference between
Q−1 (x0 )s of the locally-available Lyapunov functions in (78) with different x0 s. Because |Qi − Qj |
Qi + Qj for
any positive definite matrices Qi and Qj , we select  = 1 at first, and decrease until  = 0. The candidate F̂i of the
feedback gain matrix for ith subsystem is selected by substituting hi (x0 ) = 1 into (76). Therefore,

F̂i = F̄(x0 )|hi (x0 )=1 = Yi Q−1

i . (90)

The algorithm is summarized as follows:

Algorithm.

Step 1: Set  to one.

Step 2: Set to zero.
Step 3: (Phase 1)
Find Qi > 0, Yi and X̄ satisfying (86), (87), and (89) for i = 1, . . . , r. If Qi and Yi are found, set F̂i = Yi Q−1
i .
Otherwise, go to Step 5.
Step 4: (Phase 2)
 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228 1227

Check whether F̂i ’s are the feedback gain matrices satisfying the stability conditions of Theorem 4. If they satisfy
the conditions, stop the algorithm, and let Fi = F̂i for i = 1, . . . , r. Otherwise, increase and go to Step 3.
Step 5: Decrease . If  < 0, stop the algorithm, and there is no feasible solution. Otherwise, go to Step 2.

In the algorithm, when  = 0 and = 0 are chosen, the local stability conditions in (86) and (87) become same as the
stability conditions of a conventional Lyapunov function approach [13]. Because the stability conditions of Theorem 4
are equivalent with those of the conventional Lyapunov function approach when P1 = · · · = Pr , the algorithm results in
the same solution as the conventional Lyapunov function approach in worst case. It implies that the algorithm provides
wider range of solution space than that of the conventional Lyapunov function approach. Therefore, we propose the
algorithm as a tool to solve the BMI problem of Theorem 4. Here, (86) and (87) are derived using the relaxation method
presented in [13]. For the better relaxation, the other method in [8,17,19] can be applied as mentioned in Remark 4.

Appendix D. The algorithm with the pole-placement constraint

Using the concept of D-stability presented in [2] and [3], the constraint to place the poles of the ith subsystem in the
shaded region of Fig. 6 can be expressed in terms of LMIs as
 
−r̄Qi Ḡii
< 0, (91)
ḠTii −r̄Qi
 
sin (Ḡii + ḠTii ) cos (Ḡii − ḠTii )
< 0, (92)
cos (Ḡii − ḠTii ) sin (Ḡii + ḠTii )

Ḡii + ḠTii + Qi < 0, (93)

where

Ḡii = Ai Qi + Bi Yi . (94)

The derivation of (91)–(93) is omitted because it is similar to the result of [5] if the quadratic Lyapunov function of [5]
is replaced with the Lyapunov function Vi (x) = x T Q−1 i x for each ith subsystem. Here, the D-stability constraints are
applied to the local subsystems, not to the global system, therefore, locally-available conventional Lyapunov functions
are used which are independent of the proposed fuzzy Lyapunov function because there can exist many Lyapunov
function showing the stability of the subsystems. Fortunately, the Qi ’s and Yi ’s are compatible with those of Appendix
C. Therefore, inserting the above LMIs into the conditions to find Qi ’s and Yi ’s at Step 3 of the algorithm in Appendix
C, the feedback gain matrices satisfying the pole-placement constraint can be found.

References

[1] S. Boyd, L.E. Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory, SIAM, Philadelphia, PA, 1994.
[2] M. Chilali, P. Gahinet, H∞ design with pole placement constraints: an LMI approach, IEEE Trans. Automat. Control 41 (3) (1996) 358–367.
[3] M. Chilali, P. Gahinet, P. Apkarian, Robust pole placement in LMI regions, IEEE Trans. Automat. Control 44 (12) (1999) 2257–2270.
[4] P. Gahinet, A. Nemirovski, A.J. Laub, M. Chilali, LMI Control Toolbox, The MathWorks Inc, Natick, MA, 1995.
[5] S.K. Hong, R. Langari, An LMI-based H∞ fuzzy control system design with TS framework, Inform. Sci. 123 (1–2) (2000) 163–179.
[6] A. Jadbabaie, A reducntion in conservatism in stability and L2 gain analysis of Takagi-Sugeno fuzzy systems via linear matrix inequalities, in:
Proc. 14th IFAC Triennial World Congress, Beijing, China, 1999, pp. 285–289.
[7] H.K. Khalil, Nonlinear Systems, third ed., Prentice-Hall, Upper Saddle River, NJ, 2002.
[8] E. Kim, H. Lee, New approaches to relaxed quadratic stability condition of fuzzy control systems, IEEE Trans. Fuzzy Systems 8 (5) (2000)
523–534.
[9] K. Tanaka, T. Hori, H.O. Wang, A fuzzy Lyapunov approach to fuzzy control system design, in: Proc. American Control Conf., Arlington, VA,
2001, pp. 4790–4795.
[10] K. Tanaka, T. Hori, H.O. Wang, New parallel distributed compensation using time derivative of membership functions: a fuzzy lyapunov
approach, in: Proc. 40th IEEE Conf. on Decision Control, Orland, FL, 2001, pp. 3942–3947.
1228 B.-J. Rhee, S. Won / Fuzzy Sets and Systems 157 (2006) 1211 – 1228

[11] K. Tanaka, T. Hori, H.O. Wang, A multiple lyapunov function approach to stabilization of fuzzy control systems, IEEE Trans. Fuzzy Systems
11 (4) (2003) 582–589.
[12] K. Tanaka, T. Ikeda, H.O. Wang, Robust stabilization of a class of uncertain nonlinear systems via fuzzy control: quadratic stabilizability,
H∞ control theory and linear matrix inequalities, IEEE Trans. Fuzzy Systems 4 (1) (1996) 1–13.
[13] K. Tanaka, T. Ikeda, H.O. Wang, Fuzzy regulators and fuzzy observers: relaxed stability conditions and LMI-based designs, IEEE Trans. Fuzzy
Systems 6 (2) (1998) 250–265.
[14] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Trans. Systems Man Cybernet. 15
(1985) 116–132.
[15] K. Tanaka, M. Sugeno, Stability analysis and design of fuzzy control systems, Fuzzy Sets and Systems 45 (2) (1992) 135–156.
[16] K. Tanaka, H.O. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality Approach, Wiley, New York, 2001.
[17] M.C.M. Teixeira, E. Assunção, R.G. Avellar, On relaxed LMI-based designs for fuzzy regulators and fuzzy observers, IEEE Trans. Fuzzy
Systems 11 (5) (2003) 613–623.
[18] P.A. Tipler, Physics, Worth, Michigan, 1976.
[19] H.D. Tuan, P. Apkarian, T. Narikiyo, Y. Yamamoto, Parameterized linear matrix inequality techniques in fuzzy control system design, IEEE
Trans. Fuzzy Systems 9 (2) (2001) 324–332.
[20] D.G. Zill, M.R. Cullen, Advanced Engineering Mathematics, second ed., Jones and Bartlett, London, UK, 2000.