Sei sulla pagina 1di 4

GENERALIZED RICCATI EQUATION AND SPECTRAL FACTORIZATION

FOR DISCRETE-TIME DESCRIPTOR SYSTEM


Tohru Katayama and Yoshikazu Onuki
Department of Applied Mathematics and Physics
Faculty of Engineering, Kyoto University
Kyoto 606-01, Japan

Abstract. This paper considers a spectral factorization for discrete-time descriptor systems by using a gen-
eralized Riccati equation (GARE) for an LQ regulator problem. A novelty of the present factorization approach
is that it can be applied to the spectral factorization associated with transfer matrices no longer proper. For
the state-space models, this algorithm can be applied to strictly proper matrices with the A-matrix singular.
We also present a numerical method for computing the optimal gain and the solution of GARE based on the
generalized eigenvalue problems.
Keywords. Descriptor system; Generalized Riccati equation; LQ regulator; Spectral factorization; General-
ized eigenvalue problems.

INTRODUCTION solution for given inputs {Ut} and . suitable bound-


ary values Xo and XN. If there are no anticausal
For more than a decade, considerable interest has modes, the descriptor system is called causal. More-
been directed to the analysis, control, and estimation over,we assume that (E, A, B) is controllable, namely
of descriptor systems. Lewis (1986) has presented a his- rank[A - >'E Bl = n, '1;/>' E GU {oo}, where G is a set
torical review of linear singular systems, together with of complex numbers.
a survey of basic results on singular systems. Bender Now if we apply a linear feedback control Ut = K Xt,
and Laub (1987a,b) have treated the linear-quadratic then the closed-loop system is given by
optimal regulator .of descriptor systems for both the
continuous-time case and the discrete-time case based
(2)
on the singular value decompos ition (SVD) coordinate
system and the Hamiltonian minimi zation approach .
In this paper, we consider a spectral factoriza- If a(z) := det(zE - A - Bl<) :f: 0, then the feedback
tion problem for a discrete-time linear descriptor sys- system is regular. The closed-loop descriptor system
tem based on a generalized algebraic Riccati equation is stable, if the descriptor vector Xt tends to zero as t
(GARE) for the infinite-horizon LQ regulator problem. goes to infinity. In this case, we write (E, AK) is stable,
Using the solution of GARE, we develop a spectral fac- where AK := A + B K. A necessary and sufficient
torization algorithm, valid even if the original trans- condition for the stability of the closed-loop system is
fer matrix involved is not proper. The present algo- that all the roots of a(z) = 0 lie inside the unit disk .
rithm is applicable to the case where A-matrix of the It is well known (Lewis 1986) that if the descriptor
original transfer matrix involved is singular, so that system is controllable, there exists a feedback control
the limitation of the results of Gu et al. (1989) is that attains a desired pole assignment, removing the
relaxed. We also present a non-recursive numerical anticausal modes. Thus we say that a feedback control
method for GARE based on two generalized eigen- is admissible if the closed-loop system is (a) regular,
value problems (GEPs) associated with the Hamilto- (b) causal, and (c) stable. In this . case, we also say
nian equation (Bender and Laub 1987a, Van Dooren that (E, AK) and l< are admissible.
1981). The LQ regulator problem is to find a feedback gain
J( that minimizes the cost function
DESCRIPTOR SYSTEM
\Ve consider a linear descriptor system (3)
EXt+! = ' AXt + BUt, t =0, 1" " (1)
where Xt is the n x 1 descriptor vector, Ut is the m x 1 where it is assumed that the weighting matrix
control vector, and E E R nxn , A E Rnxn, B E Rnxm.
Wo Q
=
[ S R
ST]
is non-negative definite .
=
We assume that E is singular with rankE r ::; n, and
that the system is regular, i.e. det(zE - A) :f: O. In Suppose that
general, the descriptor system has exponential modes,

~]=n+m
nondynamic modes and anticausal modes. But the reg- A - >'E
rank [ C (4)
ularity condition assures that the uniqueness of the

2nd IFAC \VO.-)<6h 0 1> on


SYSTEM S TR U CTURE ANI) CONTROL
56 3 - :§ Scpt c rnh cl" l Q02 , PIlAGUE. C ZECHOSLOVAKIA
where). = ei9 (0 ~ 8 ~ 211'). Then the GARE associ- where Q = erC, S = DTC, R = DT D. Taking the
ated with the LQ regulator problem is given by · inverse of (9) yields

x(BTXA+S)+Q (5)
where R+BT X B > o. It can be shown (Katayama and
Onuki 1992) that the optimal feedback gain is given by

(6) Suppose that GARE of (5) has an admissible solution


X, and define Tl and T2 as
and that (E, AK) is admissible. It should be noted
that the solution X to (5) does not always exist, but
ET X E is unique and non negative definite. It will be
shown in Theorem 2 that the gain minimizing (3) is
not unique. Equation (5) has appeared in the LQG
context (Arnold and Laub 1984) and more recently in Then, using T 1 , T 2 , we get
the estimation context (Nikoukhah et al. 1989).

SPECTRAL FACTORIZATION
A 0 B
The spectral factorization algorithm derived by Gu
et al. (1989) is based on the assumption that for the _ [E 0T 0 1 -I(TVI( ET I(TV
~ I~ ] - { 0 A 0 , -VI( zBT V
transfer matrix G(z) = [ involved, A is non- o0 0 [ 0 0 Im
singular and D is of full column rank. In this section,
by using the descriptor system representation and the where V = BT X B + R.
GARE of (5), we extend the spectral factorization re- Thus the inverse of (10) gives another realization of
sult to the case where G(z) is not proper. ~(z), namely,
We consider the descriptor system (1) together with
an output equation VI = CXI + DUI, where VI is a p X 1
K~ti
A 0
E 0 T - T
output vector with p ;:: rn, and E, A, B, C, Dare 'It(z)={[OA T ], -I< VI<E
[
]} (11)
constant matrices of appropriate dimensions. Then the -VI< zB
transfer matrix is given by G(z) = D+ C(zE - A)-l B.
In the sequel, according to (Hara and Sugie 1991), the It is easy to see that (11) is factorized as
transfer matrix of a descriptor system is expressed as

{E, [ I ]}
~ ~ := D + C(zE - A)-l B (7)
Theorem 1: Suppose that (E, A, B) is controllable
Now 've' consider the spectral factorization of and ~(z) > 0 for Izl =
1. Also, suppose that GARE
of (5) has an admissible solution. Then the spectral
G- (z )G( z) associated with the transfer matrix G( z) =
{E, [ ~ I:~ ]}, where (E,A) is stable. We assume where
factorization of ~(z) is given by ~(z) = G;(z)Go(z),

that (4) holds. This implies that G(z) has no zeros on Go(z) = {E, V [ ~( I ~ ]} (12)
the unit circle (Izl = 1). Define ~(z) := G-(z)G(z).
Then, it follows that and I( is determined by (6) and V = Vl/2. 0
It should be noted that the descriptor representation
(8) of transfer matrix of (7) has conveniently been em-
ployed to handle the case where A and/or R = DT D
The spectral factorization problem is to find an outer may be singular. Hence, for the state-space model
function Go(z) such that ~(z) = G;(z)Go(z), where (E = In), Theorem 1 is an extension of the earlier
Go(z) is square and stable together with its inverse. result (Gu et al. 1989) to the case where A is singular
=
Since G-(z) cT(z-l), it follows that and/or D does not have full column rank.
For (E, A) is not stable, we can factorize the de-

~(z) ={[ E 0] [A
0 AT , -Q ET
0 nominator of (11) by applying the similar technique in
(9) order to complete the spectral factorization. This will
S zB be presented elsewhere.

57
\Ve see under the assumption of Theorem 1 that the Thus, from (14), we get
inner-outer factorization of G( z) is given by G( z) =
Gi(Z)Go(z), where EU 1 W = AU1 + BU3
=
AT U2 W -QU1 + E T U2 STU3 (15)
I
-

A + BK
, z ) -- {E , [ C+DK
G·( B ] V- 1 }
D (13) =
-BT U2 W SU1 + RU3

where W is the r x r Jordan form corresponding to the


It may be noted that the above factorization re- r stable eigenvalues.
sult is applicable to the factorization of llI(z) = /2 1- We first develop a method for computing an ad-
G-(z)G(z) with the obvious modification of Q, R, S. missible gain K and the optimal cost Jmin. We
present a lemma showing that EU1 has full column
SOLUTION OF GENERALIZED ALGE-
rank, namely, rankEUl = r based on an SVD coor-
BRAIC RICCATI EQUATION
dinate system. We assume that the descriptor sys-
In this section, in order to solve the regulator prob- tem is given by an SVD coordinate system such that
lem and spectral factorization, we consider the GEP
associated with the two-point boundary value problem
M EN = [~ ~], M AN = [~~~ ~~~], M B ~
derived from the Hamiltonian formulation of the regu-
lator problem (Bender and Laub 1987a). [ Bl]
B2 '
NTQN - [Qll Qrl] SN - [S S]
- Q21 Q22' - 1 2,

The two-point boundary value problem arising from where M, N are nonsingular.
the optimal regulator problem of minimizing (3) sub- Lemma 3: Suppose that the following rank condi-
ject to (1) is given by tions hold:
All ] = r
rank [ Qll (16)

[~ oo ] [XI+l]
o UI+1
/1+1 = [A
-Q
S
0
ET
0
rank
[
An
Qn SrB2] = n - r + m (17)
where /1 is the n x 1 codescriptor vector. 'Ve consider S2 R
the GEP
Then, EU 1 and hence U1 has full column rank.
Mx = )"Lx (14)
Proof: See Katayama and Onuki (1992). 0
where The condition~ of (16) and (17) imply that P(z) =
[ A ~'E ~] hM full column conk a. , = 0, '0
L=[! that the descriptor system has no zeros at z = o.
We observe that there are many sufficient conditions
Lemma 1: Suppose that (E, A, B) is controllable,
ensuring the condition (17) . An obvious one is that
and (4) holds . Then no eigenvalues of GEP of (14) lie
on the unit circle. 0
Lemma 2: Consider the GEP of (14). If)" is an
[~:2 SI] is positive definite . Another one is to
assume that R > 0 and An is nonsingular. There
eigenvalue, then 1/).. is also an eigenvalue. 0
We see that L has m + (n - r) zero eigenvalues, since
is also a possibility thatB2has full column rank and
Qn > 0, so that the rank condition (17) is satisfied
the last m columns of L are zeros , and rankE = r.
even if R == O.
Hence, by Lemmas 1 and 2, the eigenvalues of GEP
Since, under the assumption of Lemma 3, EU1 has
may be arranged as
full column rank, we can choose an n x (n - r) matrix
0,···,0, )..1,"', )..r, 1/)..1,"" 1/)..r,oo, .. ·, 00, 00, ··· ,00 E2 such that E := [EU 1 E 2] is nonsingular. Then we
~ ~ '----..-----" '----.--' ' - . - - " define an n X (n - r) matrix V2 and an m x (n - r)
n-r r r n-r m
matrix V3 by
with 1)..;1 < 1, i = 1,···,r.
For GEP of(14), let U be the (2n + m) x r matrix of
the generalized eigenvectors and the generalized prin-
cipal vectors corresponding to the r stable eigenvalues
It should be noted that the solutions V2, V3 may not
(Van Dooren 1981, Pappas et al. 1980). Then U, a
exist for arbitrarily given E 2 . But since (E, A, B) is
basis for the stable eigenspace, can be partitioned into
controllable including z = 00, it is possible to choose
three submatrices
E2 such that (18) has a solution.
Theorem 2: Define IT = [U1 - V2 ], and the gain
matrix K by
KIT = [U3 V3] (19)

58
If II is nonsingular, then f{ is optimal, and the optimal CONCLUSIONS
=
cost Jmin tET X E is independent of the choice of E2
and is obtained by solving ET X Ell [ETU2 0] . = In this paper, based on the GARE for the infinite·
horizon regulator problem for a discrete-time descrip·
Proof: See Katayama and Onuki (1992). 0
We see from Theorem 2 that we can obtain an op- tor system, we have developed a new algorithm for the
timal feedback gain and the optimal cost by using the spectral factorization associated with a discrete-time
generalized eigenvectors and generalized principal vec- descriptor system. Using an SVD coordinate systerr.
tors corresponding to the r stable generalized eigenval- and the GEP, we have also derived a method for com-
ues. But the gain f{ is not unique and, moreover, we puting an optimal feedback gain and a desired spectral
can not construct the solution X of the GARE. factorization.
The spectral factorization problem treated in previ-
REFERENCES
ous section can, however, be solved by using the gain
f{ above. In fact, define 1) Arnold, W. F. and A. J . Laub (1984) . Generalized
eigenproblem algorithms and software for algebraic
(20) Riccati equation. Proc. IEEE, 77,1746-1754.
2) Bender, D. J. and A. J. Laub (1987a) . The linear·
Since G.(z) = H(z)V-l from (13), we see that H(z) quadratic optimal regulator for descriptor systems:
is an un normalized inner matrix with H-(z)H(z) = Discrete-time case. Automatica, 23, 71-85.
V. Evaluating H-(z)H(z) at z =
I, we get V =
H-(1)H(I) Then Go(z) is determined by (12) with 3) Bender, D. J . and A. J . Laub (1987b) . The linear·
quadratic regulator problem for descriptor systems.
V = V 1/ 2 :'
IEEE Trnns. Automat. Contr., AC-32, 672-688.
We now develop a method for constructing an ad-
missible sOlution of the GARE under the assumption 4) Van Dooren, P. (1981) . A generalized eigenvalu€
that for the eigenvalue problem (EP) MT z = .xz, there approach for solving Riccati equations. SIAM J.
exist (n - r) eigenvectors for (n - r) zero eigenvalues Sci. Stat. Comput., 2, 121-135.
due to the' singurality of E .
For EP, let V be the (2n + m) x (n - r) matrix 5) Gu , D. W., M. C. Tsai , S. D. O 'Young and 1.
of the eigenvectors corresponding to the (n - r) zero Postlethwaite (1989). State-space formulae for
eigenvalues. Then V can be partitioned into discrete-time Ho:> optimization . Int. J. Control,
49,1683-1723 .

V -- [ ~l 1 V;
2 , I, v:2 E cn x (n-r) , v:3 E cmx(n-r)
6) Hara, S. and T . Sugie (1991). Inner-outer factoriza-
tion for strictly proper functions with jw-axis zeros.
V3 System fj Control Lett., 16, 179-185.
Thus, for n - r zero eigenvalues , we get from EP 7) Katayama, T . and Y. Onuki (1992) . LQ regulator
and spectral factorization for discrete-time descrip-
ATVI - QV2 + STV3 = 0 tor system. Tech. Report ~92001, Dept. Appl.
EV2 =0 (21) Math . & Physics, Fac. of Eng., Kyoto Univ.
BTVl - SV2 + RV3 = 0
8) Lewis, F. L. (1986). A survey of linear singular
Then we obtain the following theorem that gives a com- systems. Circuits Syst. Signal Process., 5, 3-36.
putation algorithm for solving GARE of (5).
9) Nikoukhah, R., M . B. Adams, A. S. Willsky and
Theorem 3: Suppose that (E, A, B) is controllable,
B. C. Levy (1989). Estimation for boundary-value
and (4) holds . Define
descriptor systems. Circuits Syst. Signal Process.,
(22) 8,25-48.
10) Pappas, T ., A. J . Laub and N. R. SandeIl (1980).
and
On the numerical solution of the discrete-time al-
=
E [EU1 - AV2 + BV3 ] (23)
gebraic Riccati equation. IEEE 7rans. Automat.
Then there exists an admissible solution X of GARE Contr., AC-25, 631--{)41.
of (5) with BT X B + R > 0 if and only if II and E are
nonsingular. Furthermore, an optimal feedback gain
f{ and the solution of GARE are determined by

Proof: See Katayama and Onuki (1992). 0


It should be noted that X of (24) is not necessarily
a unique solution to GARE (5).

59

Potrebbero piacerti anche