Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

php

Chapter 5

Parametric Lyapunov Functions for Uncertain Systems: The Multiplier Approach

Minyue Fu and Soura Dasgupta
5.1 Introduction
This chapter proposes a parametric multiplier approach to deriving parametric Lyapunov functions for robust stability analysis of linear time-invariant (LTI) systems involving uncertain parameters. This new approach generalizes the traditional multiplier approach used in the absolute stability literature, where the multiplier is independent of the uncertain parameters. The main result provides a general framework for studying multiaffine Lyapunov functions. It is shown that these Lyapunov functions can be found using LMI techniques. Several known results on parametric Lyapunov functions are shown to be special cases. Our focus is on the existence of parametric Lyapunov functions. Our motivation for considering parametric Lyapunov functions stems from two facts. First, they can offer less conservative robust stability conditions than parameter-independent Lyapunov functions that are used in the quadratic stability theory. Second, they can be applied to stability analysis of systems with time-varying parameters. In the latter case, one possible approach is to find a parametric Lyapunov function that ensures that the "frozen" uncertain system (i.e., the uncertain system with "frozen" parameters) is robustly stable with certain stability margin. This margin can then be used to determine the "average" time variation for the parameters that could be withstood without losing stability. The use of parametric Lyapunov functions can be traced back to the work of Parks [325], who proves the RouthHurwitz stability condition for polynomials directly using a Lyapunov matrix called Hermite matrix. The unique feature of the Hermite matrix is that it is bilinear in the coefficients of the polynomial. The well-known Popov criterion [337], when specialized to dealing with a single constant uncertain parameter rather than sector-bounded nonlinearity, yields a Lyapunov function that depends on the un95

96

Fu and Dasgupta

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

certain parameter affinely. In fact, most absolute stability criteria given in the literature give, explicitly or implicitly, a parametric Lyapunov function when the uncertainty or nonlinearity they consider reduces to constant uncertain parameters. Recently, Dasgupta et al. [92] studied the existence of parametric Lyapunov functions for the following family of systems:
(5.1)

i(t) = A(q)x(t) = (Ao + bc T (q))x(t), q E Q,

where b, c(q) E

(5.2)

Q = {q = [q1, ... , qm] T : qti < q j < qi }

and h(q) is affine in the elements of q. This type of uncertainty is called rank-one uncertainty. It is shown in [92] that the uncertain system above is robustly stable, i.e., stable V q E Q, if and only if there exists some constant stable matrix F and a compatibly

dimensioned vector w such that the augmented uncertain system

(5.3) x(t) = H(q)x(t) _ [ 0
c"()

] (t), q E Q,

admits a parametric Lyapunov function P(q) that depends on q multiaffinely. A procedure for constructing F, w, and P(q) is given in [92]. It is also shown how this result is used in robust stability analysis for linear systems with time-varying parameters. In Feron, Apkarian, and Gahinet [135] and Haddad and Bernstein [183], a more general class of uncertain matrices
P

(5.4)

A(q) = Ao +
i =1

giA2

are considered along with the so-called generalized Popov criterion to generate parametric Lyapunov functions. This kind of uncertain system is also studied by Gahinet et al. [149] using an affine Lyapunov matrix
P

(5.5)

P(q) = Po +
i =1

giPi ,

where Pi, i = 0, 1, ... , p, are symmetric. In contrast to [92], the robust stability tests given by the generalized Popov criterion and an affine Lyapunov matrix above are conservative

in general.
In this chapter, we study a more general family of uncertain systems described by (5.6)

(t) = A(q)x(t) = (Ao - BD -1 (q)C(q))x(t), q E Q,

where the nominal matrix Ao E R" x " is assumed to be stable, B E IV x m is a constant full-rank matrix, in < n, C(q) E Rmx" and D(q) E 7Zm x m are affine in q as follows:
P P

(5.7)

C(q) =

E qZG'^, D(q) = Do + s=1 =1

and D(q) is invertible V q E Q. Also, we seek a multiaffine Lyapunov matrix P(q) _ PT (q) > 0 of the form
P

(5.8)

P(q) = Po +
i=1

+ E gsqjP^j + .. .
ij

5.2. Multipliers and robust stability such that Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php (5.9)

97

(Ao BD -1 (q)C(q)) T P(q) + P(q)(Ao BD -1 (q)C(q)) <0 Vq E Q.

The tool we use for establishing parametric Lyapunov functions is the multiplier approach, which is popularly used in the absolute stability literature. The basic idea of this approach is to find a transfer matrix of certain structure, called multiplier, such that the cascade of the multiplier and some transfer matrix related to the uncertain system is strictly positive real. What makes our approach unique is the use of parametric multipliers, whereas in the literature only constant multipliers are used. Our main result gives a sufficient condition for the existence of a multiaffine Lyapunov function in terms of the existence of an affine multiplier with a special structure. It turns out that this multiplier can be found by solving a set of linear matrix inequalities (LMIs). The solution to the LMIs automatically provides a multiaffine Lyapunov function for the given uncertain system. We then focus on the analysis of the conservatism of the method we propose. Two schemes have been analyzed in detail, namely, the so-called generalized Popov criterion for the multiparameter case, which has been studied by many authors, and the so-called affine quadratic stability (AQS) test for the single parameter case. It is found that the former renders a constant multiplier, and the latter results in a constant multiplier in the single parameter case and an affine multiplier in general. Subsequently, a frequency domain interpretation of these two schemes is given.

5.2 Multipliers and robust stability

To motivate the multiplier approach, we consider the following uncertain transfer matrix associated with (5.6): (5.10)

G(s, q) = C(q)(sI - A o )

1B

+ D(q).
-

Lemma 5.1. The system (5.6) is stable if and only if G ' (s, q) is stable. Proof 5.1. The characteristic polynomial of the system (5.6) is given by det(sI - A o + BD -1 (q)C(q)) = det(sI - Ao) det(I + (sI - Ao)
-

1 BD 1 (q)C(q))
-

= det(sI - Ao) det(I + D -l (q)C(q)(sI - A o )

1 B)

= det(sI - Ao) det(D -1 (q)) det(D(q) + C(q)(sI - Ao) -1 B). Hence, the eigenvalues of the system (5.6) coincide with the poles of G' 1 (s,q) modulo the stable eigenvalues of A 0
. K

In view of the result above, we are interested in finding an affine transfer matrix, called affine multiplier, of the form (5.11)

K(s,q) = Ck(q)(sI - A o ) 1 B + Dk(q) = (Cko + E,P 1 giCki)(sI Ao) -1 B + (Dko +

-

giDki)

such that the transfer matrix below (5.12)

H(s, q) = K(s, q)G -1 (s, q)

is strictly positive real (SPR) V q E Q. A few remarks are in order.

98

Fu and Dasgupta 1. A direct consequence of the existence of such a multiplier is that G -1 (s, q) and hence the system (5.6) are robustly stable. This follows from the trivial fact that an SPR function is stable. If B and C(q) are rank one, which is slightly more general than the uncertain system (5.1) where D(q) = 1, the corresponding G(s, q) is a single-input singleoutput transfer function depending on q affinely. That is, the numerator coefficients of G(s, q) are affine in q and the denominator of G(s, q) is parameter independent. In this case, it is known that G -1 (s, q) is robustly stable if and only if a parameter-independent multiplier K(s) (possibly with a high order) exists to render K(s)G -1 (s,q) robustly SPR; see Anderson et al. [12]. The parametric Lyapunov functions given in [92] are actually based on the existence of such a multiplier. The multipliers we allow in this chapter are more general than [92] in the sense that they are paremeter dependent but more restrictive in the sense that they share the same Ao as G(s, q). As we will see later, this restriction is used to assure that the multiplier will yield a multiaffine Lyapunov matrix (5.8). 3. Although the multiplier approach is a frequency domain method, it has a natural state-space domain interpretation. This interpretation is achieved using a version of the well-known Kalman-Yakubovich-Popov (KYP) lemma called the parametric KYP lemma that is to be introduced in the next section. This result is the key device for deriving a multiaffine Lyapunov function from an affine multiplier. It is this lemma that motivates the use of parametric multipliers.

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

5.3 Parametric KYP lemma

This section presents a version of the well-known KYP lemma, which is instrumental in dealing with systems with uncertain parameters. To prepare for the result, we first introduce a known generalized KYP lemma by Willems [419]. Lemma 5.2 (generalized KYP lemma). Given A E Rnxn B E Rnxk, and symmetric Sl E R(n+k)x ( n+k) , there exists a symmetric matrix P E Rnxn such that
,

(5.13)

r AT PT PA PB L B P 0

+1<O

if and only if there exists some e> 0 such that (5.14)

[BT ((sI - A) -1 )` I]SZ I

(sI

- I)-1B

J < 0 b' Re[s] > e.

Further, if A is Hurwitz stable and the Lupper left n x n block of St is positive semidefinite, then P above, when it exists, is positive definite. Our desired result is given as follows in Lemma 5.3. Lemma 5.3 (parametric KYP lemma). Given matrices A E Rnxn, B E Rnxm, m < n, a hyperrectangular set Q C RP, a parametric matrix 1 (q) E R(n+m) x (n+m)
described by
P

(5.15)

f (q) = clT (q) = SIM(q) +

q; nii^

5.3. Parametric KYP lemma where S m (q) is multiaffine in q and Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php (5.16) Dii>0, i =1,...,p.

Then the following conditions are equivalent:

(i) There exists e > 0 such that (5.17) VqEQ. (ii) There exists a multiaffine matrix (5.18) such that (5.19) VgEQ. [1(q) P(q) = P T (q) E R""", q E Q, [BT ((sI - A) -1) I] 1 (q) [ (sI - A)B <0VRe[s]>-e

= I

[ATP() + P(q)A P(q)B ] + Q(q) <0 BT P(q) 0

(iii) The inequality (5.17) holds at all vertices of Q.

(iv) The inequality (5.19) holds at all vertices of Q. Proof 5.2. The equivalences (i)<=>(iii) and (ii)<=>(iv) are obvious due to the assumption in (5.16). The implication (ii)=>(i) follows directly from Lemma 5.2. More precisely, for each fixed q E Q, there exists (q) > 0 such that (5.17) holds V Re[s] > E(q). Since Q is a compact set, e = nu (q)
qEQ

exists and is positive. Subsequently, (5.17) holds for Re[s] >_ -e. To show (iii)=>(iv), we apply Lemma 5.2 and induction. Without loss of generality, we assume Q = [0, 1]P. Suppose (iii) holds for p = 1; i.e., Q = [0, 1]. Let Po and P1 be any solutions to (5.19) at q = 0 and q = 1, respectively. (The solutions are guaranteed to exist by Lemma 5.2.) Also denote the corresponding II(q) by Ho and 111. Define (5.20) P(q) = (1 - q)Po + qPi,

which is affine in q, symmetric, and positive definite. It is obvious that the corresponding 11(q) is negative definite at q = 0 or 1. That is, (iv) holds for p = 1. Suppose (iii)=>(iv) for p = k; we need to prove it for p = k + 1. Assume (iii) holds for p = k + 1 and write q = ( q k , q1). Then, (5.17) holds V qk E [0, 1] k , qk+l = 0, and qk+l = 1. By the assumption, there exists p0(qk) and P1(gk), both multiaffine, symmetric, and positive definite, such that the corresponding 1I(q), denoted by fl(qk) and Ill (qk), are negative definite V qk E [0, 1] k . Now we apply the same "trick" for p = 1 again. That is, define (5.21)
P(q) = ( 1 - qk+l)p (q k ) + qk +1PI(g k )

Then P(q) is multiaffine, symmetric, and positive definite. Also, it is straightforward to verify (5.19) at qk E [0, 1] k , qk+l = 0, and 1, when the above P(q) is applied. That is, (iv) holds for p = k + 1. 0

100

Fu and Dasgupta

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Remark 5.1. It is important to know that the proof above gives an algorithm for constructing the multiaffine Lyapunov matrix P(q) required for the parametric KYP lemma. Alternatively, P(q) can be searched using SDP algorithms because (5.19) is an LMI at each vertex of Q. The latter is efficient especially when only a few terms of P;, Ptij, ... , are sought.

5.4 Main results on parametric Lyapunov functions

Under a "convexity condition," our main result below provides a necessary and sufficient condition for the existence of a multiplier of the form (5.11). This condition automatically renders a multiaffine Lyapunov matrix. Theorem 5.1. Given the uncertain system in (5.6), suppose there exists an af!'ine multiplier K(s, q) of the form (5.11) such that the transfer matrix H(s, q) in (5.15) is SPR at all vertices of Q. In addition, the convexity condition below is satisfied:

(5.22)

He

[ L DT J [Cki Dki] <_ 0, i = 1, ... p.

t

Then, the following properties hold:

(i) H(s, q) is SPR `d q E Q.

(ii) H(s, q) has the following nth- order realization

('

Ao + BD - 1 (q)C(q)) - ' )5 23 H ( s, q) = (Ck(q) - Dk(q)D '(q)C(q))(sI xBD-1(q) + Dk(q)D -1 (q).

(iii) There exists a multiaffine P(q) = PT (q) of (5.8) to establish the robust SPR property of H(s, q), i.e.,

(5.24)

n(q) _

AT (q)P(q) T+ P(q)A(q) 11 12(q) 1122(q) J n12(q)

1<0

holds V q E Q, where

11 12(q) = P(q)BD -1 (q) - C' (q) + C T (q)D - T (q)Dk (q),

1122(q) = - (Dk(q)D -1 (q) + D -T (q)Dk (q)) (iv) (5.24) holds V q E Q if and only if it holds at all corners of Q.
(v) The same P(q) above is a Lyapunov matrix for establishing the robust stability of

(5.6).
Conversely, if there exists P(q) of the form (5.8) and a multiplier K(s, q) of the form (5.11) such that the convexity condition (5.22) is satisfied and that the LMI (5.24) holds at all vertices of Q, then K(s, q)G -1 (s, q) is SPR `d q E Q. Remark 5.2. The property (iv) shows that the multiafflne Lyapunov matrix P(q) can be searched using LMI techniques. Indeed, the matrix in (5.24) is linear in Po, P;, F11,..., (defined in (5.8)). Subsequently, the existence of such P(q) is equivalent so that the 21' LMIs (5.24), one for each corner of Q, are feasible.

5.4. Main results on parametric Lyapunov functions Now the proof of Theorem 5.1 follows.
Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

101

Proof 5.3. Define

11 (q) _ - [ D q) Dk(q)] - [ D'(q) ] [C(q) D(q)]. ( T q) ] [Ck(

Then, K(s, q)G -1 (s, q) being SPR V q E Q is equivalent to the existence of > 0 such that -2He[G*(s, q)K(s, q)] = [B T ((sI - Ao) -1 ) * I] 1Z(q) I (sI - Ao) -1 B 1 <0

V Re[s] > -e and q E Q. Note that the quadratic terms in 11(q) are given by

= -2He I DT L Z

[Ck^ Dki]] > 0, i = 1,. .. , p.

] ]

Using Lemma 5.3, the SPR condition above is equivalent to the existence of a multiaffine P(q) such that the following holds at all vertices of Q: q) + P(q)Ao ] - 2He I D T (q) ] [C(q) Dk(q)]] < 0. P( 0 )B [ Ao P(BT P(q)

Pre- and postmultiplying both sides by

I 0
- D - '(q)C(q) D -1 (q)

and its transpose, respectively, will yield (5.24).

Remark 5.3 (LMI solution). The inequalities (5.22) and (5.24) represent a finite set of LMIs (p from (5.22) and 2P from (5.24)). These LMIs are jointly convex in Po, Pi , P^^..., Ck i , and Dk i . Thus, efficient LMI algorithms, such as interior-point algorithms, can be applied to compute the affine multiplier and and the associated multiaffine Lyapunov matrix. The following result shows that the form of the affine multiplier can be simplified to have one fewer pair of Ckti, Dki. In particular, for single parameter uncertain systems, the existence of affine multipliers is equivalent to the existence of a constant multiplier, provided that the convexity condition mentioned above is satisfied. Theorem 5.2. Given the uncertain system in (5.6), suppose there exists an affine multiplier K(s, q) of the form (5.11) such that the transfer matrix H(s, q) in (5.15) is SPR at all vertices of Q and the convexity condition in (5.22) is satisfied. In addition, assume that [Cr , DP ] has rank m (full rank). Then there exists EP E Rm x m such that
(5.25)

[Ckp Dk p ] = EP [CP , DP ], EP + ET < 0.

Further, let
(5.26) and
P-1 P-1 giCki (sI - Ao) -1 B + Dko +
i =1 i =1

Cki = Cki - EPC:, Dki = Dki - EPDi,

i = 0,1, .. . , p - 1,

(5.27)

K(s, q) = Oko +

giDki

Then K(s, q)G -1 (s, q) is SPR V q E Q. In particular, ifp = 1 then the resulting multiplier K(s, q) is parameter independent.

102 The proof of the theorem above hinges upon the following lemma. Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php Lemma 5.4. Suppose A, B E Rnx k , k < n with rank(A) = k and
(5.28)

Fu and Dasgupta

ABT + BA T > 0.
Xk

Then there exists some H E W (5.29)

such that

B = AH, rank(H) = rank(A), H + H T > 0.

Proof 5.4. Let T E R" x "` be a nonsingular transformation matrix such that

TA=[

Ik

and denote
B=[ i
]

=TB.

Then (5.28) implies (5.30) [ Ok

]

[Bi B2 ] + B2
[

Ii
[

0] = I

Bl B2B1 02

J > 0.

This results in B2 = 0. Consequently,

TB = I ^ k

B i =TAB 1 .

So, H = Bl and B = AH and rank(B) = rank(H). Finally, H + HT >_ 0 follows from (5.28) and the full-rank property of A. So (5.29) holds. K Now the proof of Theorem 5.2 follows. Proof 5.5. Using Lemma 5.4, the convexity condition (5.22) and the full-rank property of [Cr , Dp ] imply that Ep exists to satisfy (5.25). Since K(s, q)G -1 (s, q) is SPR `d q E Q, there exists e > 0 such that He[G* (s, q)K(s, q)] > 0, Re[s] > -e, q E Q. It follows that
He[G* (s, q)K(s, q)] > He[G* (s, q)K(s, q)] + G* (s, q)He[EP ]G(s, q)

= He[G*(s, q)(K(s, q) + Ep G(s, q))] = He[G* (s, q)K(s, q)] > 0, Re[s] > -e, q E Q.

Remark 5.4. The assumption that [Cr , Dp ] is full rank is always satisfied for the single parameter case. This condition, however, may not be satisfied in general. Whenever this condition holds, the result above shows that the LMIs (5.22) and (5.24) are simplified without any harm by setting [Ckp Dk p ] = 0. It is, however, difficult to simplify further because the resulting [Cki, Dkz] may not meet the convexity condition. Remark 5.5. The result above shows how to simplify the multiplier. But it makes no implication that similar simplification can be made for the Lyapunov matrix. For example, it is not claimed that a constant multiplier renders a constant Lyapunov matrix.

5.5. Comparison with other schemes

103

5.5 Comparison with other schemes

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php Theorem 5.1 gives a sufficient condition for the existence of a multiaffine Lyapunov matrix that is of the same order as the plant. In this section, we show that two other known schemes for obtaining parametric Lyapunov functions are special cases of Theorem 5.1. One of these two schemes is a generalized Popov criterion derived based on the so-called

S-procedure [135, 149, 203, 183, 380], and the other is called the AQS test by Gahinet et al. [149]. We show that the former gives a parameter-independent multiplier, while the latter gives a parameter-independent multiplier in the single parameter case and an affine multiplier in general.

5.5.1 Generalized Popov criterion

Consider the following system: (5.31) Or, equivalently,
p

x = ( Ao + E giAti x.

i = Aox+^Aiw i ,
i =1

yi = x,
qjyj, i =

(5.32)

1 ,...,p,

where q = (ql, ... , qp ) E Q = [-1, 1]P (without loss of generality), Ao, Az E R", i = 1, ... , p, and A0 is asymptotically stable. The generalized Popov criterion seeks a Lyapunov function of the following form [149]:
p (5.33)
i'

V (x, t) = 2x T Po + P gi,Pi x + ( 1 q 2 ) I t yi' Siyidr,

where Si = Sr >_ 0 and (Po + > i giPP) = (Po + > s giPj) T > 0 for all qj E [-1, 1]. Note that this function appears to be more general than an affine quadratic Lyapunov function due to the integral terms. The following sufficient condition is established [135, 149], which is generalized from
[183, 203, 380].

Lemma 5.5. The system (5.31) is robustly stable with the Lyapunov function (5.33) if there exists symmetric matrices Po, Pd = diag{P2}, S = diag{Si} >_ 0 and skew symmetric matrix T = diag{Ti} such that the following LMI is satisfied:
( 5.34) I A0Po+PoAo+C T SC PoB+AoC T Pd+CTT L BT Po+PdCAoTC BT CT Pd+PdCBS ] <0'

where B = [Ai,A 2 ,...,Ap ] and C T = [I,I,...,I]. What we intend to show here is that the condition in Lemma 5.5, if satisfied, will lead to a parameter-independent multiplier K(s), which in turn leads to a multiaffine Lyapunov function without an integral term (cf. (5.33)). Theorem 5.3 is the result.

104

Fu and Dasgupta

Theorem 5.3. Consider the uncertain system (5.32) and suppose that the condition in Lemma 5.5 is satisfied. Define A = diag {q;I}, C(q) = AC,
Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(5.35) (5.36)

G(s, q) = I - C(q)(sI - Ao) -1 B, K(s) = -(TC + PdCAO)(sI - Ao) -1 B + (S - PdCB).

Then K(s) is invertible and K(s)G -1 (s, q) is SPR `d q E Q. Proof 5.6. Suppose the condition in Lemma 5.5 holds. A straightforward application of Lemma 5.3 yields G*(s)SG(s) - S + (sPd + T)G(s) + G*(s)(sPd + T)* <0 `d Re[s] > - E for some e > 0, where G(s) = C(sI - Ao) - 'B. Define X(s) = S +G*(s)(sPd + T), Y(s) = G*(s)SG(s) - S + (sPd + T)G(s) + G* (s) (sPd +T)*. It is straightforward to verify that X(s)G(s, q) + G*(s, q)X*(s) _ -Y(s) + G*(s)S 1 ' 2 (I - A*A)S 1 / 2 G(s) +(5 112 AS lf 2 G(s)) * (S" 2 AS1/2G(s)) `d Re[s] > - E, A. > -Y(s) >0 That is, (5.37)

He[(S + G*(s)(sPd + T))G(s, q)] > 0 V Re[s] > - E

holds `d A. The idea above is in fact borrowed from [277]. It is straightforward to verify that an alternative representation of K(s) in (5.36) is given by K(s) = S (sPa + T)C(sI A o )- 1 B. Noting that S and Pd are symmetric and T is skew symmetric, we have that (5.37) is equivalent to He[G*(s, q)K(s)] = He[K *(s)G(s, q)] > 0 `d Re[s] > e. That is, K(s)G -1 (s, q) is SPR d q E Q. O

Remark 5.6. Note that the Lyapunov function in (5.33) involves integral terms. However, the existence of a multiplier K(s) in (5.36) implies that an alternative Lyapunov function can be found that does not involve integral terms. This claim follows from property (iv) in Theorem 5.1. The trade-off is that the new Lyapunov function may involve multiaffine terms. But in the special case where p = 1, we can conclude that the Lyapunov function obtained using Theorems 5.3 and 5.1 is affine.

5.5. Comparison with other schemes

105

5.5.2 The AQS test

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Consider the uncertain system (5.31) and the affine Lyapunov matrix (5.38)
P

P(q) = Po +
i =1

giPi

The AQS test proposed in Gahinet et al. [149] amounts to finding such P(q) that the following two conditions are satisfied:
P T

(5.39) (5.40)

Ao +
i =1

giAi

P(q) + P(q) Ao +
i =1

giAi

< 0 b' qi = 1,

ATPi +PPA. > 0 `di=1,...,p.

The use of the constraint (5.40), which adds to the conservatism of the method, is to assure that (5.39) holds `d q E Q. The AQS test requires us to solve the set of LMIs above. The result below gives an interpretation of the AQS in terms of the multiplier approach. Theorem 5.4. Given the uncertain system (5.31), suppose there exists an affine Lyapunov matrix of the form (5.38) such that (5.39) holds. Define
P

(5.41) and (5.42)

G(s,q) = I

giAi (sI Ao) i

K(s, q) =

o+

E gjPi (sI Ao)

-1 .

Then, K(s, q)G -1 (s, q) is SPR V q E Q. The convexity condition for K(s, q) as stated in Theorem 5.1 holds when (5.40) is satisfied. Further, for the single parameter case, let A l be decomposed into (5.43)
where B E Rnxm C H E Rk x c such that
,

Al = BC,
E

Rmxn are full-rank

matrices and m < n. Then there exists some

> 0,

(5.44)

BTP1 = HC, rank(H) = rank(B T Pl ), H + H T

[- 1,

and K(s)G(s, q) -1 is SPR V q E

1], where
-

(5.45) and (5.46)

G(s, q) = I qC(sI Ao )

1B

K(s) = B T Po(sI Ao) -1 B + H.

Proof 5.7. The properties about K(s, q)G -1 (s, q) are trivial. The existence of H follows from Lemma 5.4. To show that K(s)G -1 (s, q) is SPR V q E [- 1, 1], we take p = 1 and note that (5.39) implies the existence of r > 0 such that
He[(sI Ao qBC)*(Po + qPi )] > 0 V Re[s] > E, q E [-1, 1].

106

Fu and Dasgupta

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Pre- and postmultiplying the above by the full-rank matrix (sI - A0 )B and its Hermitian gives He[G*(s, q)(B T Po(s1- Ao) -1 B + qB T Pl(sI - Ao) -1 B)] > 0. Using (5.44), the above becomes 0 < He[G *(s, q)(B T Po(sI - A o ) - 1 B + gHC(s1- A o ) - 1 B)] = He[G *(s, q)(B T Po (sI - A o ) - 1 B + H - HG(s, q))] < He[G *(s, q)(B T Po(sI - A o ) - 1 B + H)] = He[G *(s, q)K(s)] d Re[s] > - e, q E [-1,1]. So K(s)G -1 (s) is SPR `d q E [-1, 1]. K

5.6 Generalization to discrete-time systems

Consider the uncertain discrete-time system (5.47) x(k + 1) = A(q)x(k) = (Ao + BD - 1 (q)C(q))x(k),

where all the entries are the same as in the continuous-time case. Searching for parametric Lyapunov functions in the discrete-time case seems more involved than in the continuoustime case. The reason is that the stability requirement in the discrete-time case becomes finding P(q) = PT (q) > 0 such that (5.48) AT (q)P(q)A(q) - P(q) <0 V q E Q.

Even when A(q) and P(q) are both affine, the inequality above involves cubic terms. Alternatively, the following equivalent condition to (5.48) can be analyzed: (5.49} -P(q) AT (q) A(q) -P-1(q) < 0.

This condition is often used for quadratic stability analysis when P(q) is restricted to be constant. When parametric Lyapunov matrices are considered, condition (5.49) again seems to be powerless because P -1 (q) is nonlinear in q. In this section, we show that the multiplier idea studied in previous sections can be generalized to discrete-time systems with ease. We first introduce a counterpart of the parametric KYP lemma for discrete time. Lemma 5.6 (discrete-time parametric KYP lemma). Given matrices A, B, SZ(q), Q as in Lemma 5.2, the following two conditions are equivalent: (i) There exists 0 < e < 1 such that (5.50) [BT ((zI - A) -i)* I] 0 (q) I (zl

- A)

-1 B

l < 0

V zl 1 -e, qEQ. (ii) There exists a multiaffine matrix (5.51) such that (5.52) V q E Q. II(q) _ [ P(q) = PT (q) E Rn
x
a,

qEQ

ATP( q )A - P(q) ATP( q )B BTP(q)B ] + 11 (q) <0 BTP( q )A

5.7. Conclusions (iii) The inequality (5.50) holds at all vertices of Q. Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php (iv) The inequality (5.52) holds at all vertices of Q.

107

Proof 5.8. The proof is essentially the same and is based on a discrete-time version of Lemma 5.2. So the details are omitted. K Applying the lemma above, the counterpart of Theorem 5.1 is easily obtained. Theorem 5.5. Given the uncertain system in (5.47), suppose there exists an affine multiplier K(z, q) of the form (5.11) such that the transfer matrix H(z, q) in (5.12) is SPR at all vertices of Q. In addition, the convexity condition in (5.22) is satisfied. Then the following properties hold:

(i) H(z, q) is SPR V q E Q. (ii) H(z, q) has the nth-order realization

(5.53) H(z, q) = (Ck(q) - Dk(q)D
1 (q)C(q))(zI - A o + BD xBD -1 (q) + Dk(q)D -1 (q)
-

1 (q)C(q)) -1

(iii) There exists a multiaffine P(q) = P T (q) to establish the robust SPR property of H(z, q); i.e., (5.54) n(q)

=L

AT (q)P(q)A(q) - P(q) II12(q), I < 0 fl(q) 1122(q).

holds V q E Q, where

11 12(q) = AT (q)P(q)BD -1 (q) - Ck (q) + C T (q)(D T ) -1 (q)Dk ( q),

11 22(q) = (D T ) -1 B T p(q)BD -1 (q) - 2He(Dk(q)D-1(q)].

(iv) The same P(q) above is a Lyapunov matrix for establishing the robust stability of (5.47). Conversely, suppose there exists P(q) of the form (5.8) and a multiplier K(z, q) of the form (5.11) such that the convexity condition (5.22) is satisfied and that the LMI (5.54) holds at all vertices of Q. Then K(z, q)G -1 (z, q) is SPR V q E Q. Proof 5.9. The proof is virtually identical to the continuous-time case. The details are thus omitted. K Remark 5.7. Note that, as in the continuous-time case, (5.54) is affine in Po, Pi, Pu ,. Cki, Dkt. Hence, finding a multiaffine Lyapunov matrix P(q) amounts to solving a finite number of LMIs (p for (5.22) and 2p for (5.54) at the vertices of Q).

5.7 Conclusions
In this chapter, we have studied the use of the multiplier approach to generate parametric Lyapunov functions for linear systems with parameter uncertainty. In the process of doing so, we have derived an extended version of the KYP lemma, the parametric KYP lemma, as a general tool to study the robust stability with parameter uncertainty. Using this lemma, we have provided conditions under which an affinely parameterized multiplier

108

Fu

and

Dasgupta

Downloaded 11/11/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

exists to establish the robust stability of the uncertain system. This type of parametric multiplier then naturally leads to a multiaffine Lyapunov function that can be used to establish robust stability in the state-space domain. Although not studied in this chapter, we point out that parametric Lyapunov functions can also be used in dealing with timevarying parameters; see [92]. Also shown in this chapter is that some previous results in the literature on parametric Lyapunov functions lead to special multipliers. This analysis confirms the generality of our approach. We have also demonstrated the ease of adapting our approach to discrete-time systems despite the observation that uncertain parameters in the discrete-time case appear to be harder to deal with.