1220 IEBE TRANSACTIONS ON AUTOMATIC CONTROL., VOL AC-25, NO.

6, DECEMBER 1980

is analytic in C+ ; note, however, that it has a pole at every zero of x(-). A Two-Stage Design of Linear Feedback Controls
Furthermore, N,,is analytic in C, because it is a polynomialin sihence,
it is an entire [Link], by (9), for all z E Z [ P ] with R. G. PHILLIPS
ZEC,,

rank[ f i y , ( z ) ] < n o (12)

rank[N',(z)]<no;
because
thus, Z[i]cZ[fiyu]. 0 IM-RODUCIION
In this paper we decompose the feedback design of the n-dimensional
CONCLUDING REMARKS system
1) F2r exponentially stable closed-loop systems, everyC +-zero of $e
plant P isAneceSsAarily a zero of the closed-loop transfer function Hyy
whatever C and FAmaybe; if any C + -zero of the plant is not a zero of a
specified stable H,,, then theclosed-loopsystem is unstable. Conse-
into two reduced4rder subsystems designs, In (1) the control u is an
quently, y y sound design method must impose this limitation on the
m-vector and Aii are ( n i X n i ) matrices; i=1,2; n , + n 2 = n . In most
required Hyu(see,e.g., [lo] and [l2D.
applications adesiredeffect of feedback is to move the eigenvalues
2) In fact, from (?) and (IO), it follows that if z € Z [ i ] an$ x ( t ) # O ,
farther left in the complex plane, which also increases their magnitudes.
then z is a zero of H,,.-Consequently, if z E C - is a F r o of-P, the only
In the proposed two-stage design procedure the n I eigenvalues, which
way to remove it fromHa, (by appropriate choice of C and F)is to have
werelargerin the open-loopsystem (l), also become the n1 larger
it cancelled out by a C --zero of the closed-loop characteristic poly-
eigenvalues of thedesignedclosed-loop system. In other words, the
nomial x , i.e., by a pole of the function (11).
requirement
3) Some a d d i t i d information is available from the proof. By (9), if
Z E Z [ P ] and x(z)#O, then ~ A , ~ > ~ A i ~ , i = ,nl;
l,~-- j=nl+l,..-,n (2)

R[ & < 4 ] c R [ N p r ( ~ ) l , is satisfied at each design stage. Similar two-stage designs can be based
on any block-diagonalizingtransformation. The main tool of the design
proposed hereis an explicitly invertible transformation. Its properties are
thus emphasizing that zeros of transmissionshave"position" (inthe reviewed [Link] the design procedure is presented. Finally, we show
transfer function matrix) as well as "value" (in the complex plane) [I 11.
that earlier proposed decompositionssuch as [I]-[3], applicable to singu-
4) Since the proof required onlythat transfer functions be factorizable
larly perturbed systems, are special cases of this procedure when eigen-
and that the notion of characteristic polynomial (function) be defined in value separation (2) is sufficiently large.
terms of these factors, it is obvious that this theorem applies to the
dbrrihred case (see, e.g., [I31 and [I41and use the methods of [3D.
5) A reviewer pointed out that the theorem above is a consequence of AN EXPLICITLY [Link]
the generalsynthesis conditions [15, Theorem 6.21 of Pernebo. Our It can easily be verified that the transformation
derivation is more direct and easily extended to the distributed case.

REFERENCES
T= [ L -"I
I2

C. A. Dsm and I. D. Schulman, "Zeros and poles of matrix pansfer fundions where Ii is the ni Xn, identity, i = 1,2; Mis n 1 X n 2 ; and L i s n2 X n l has
and their dynamical interpretation," IEEE Tmnr. Circuits %st., vol. CAS-21, pp. the following properties.
3-8. Jan. 1974.
C A Desoer and Y. T. Wan&"Linear timeinvariant robustservomechaaism 1) For any L and M the inverse of T is
problem: A seli-containedexposition," in Adumced Control and &mtSyrtemr,
vol. 16, C.T. h n d c s , Ed. New Academic, to be published-
F. [Link], V. H. L. Cheng, and C k Desoer, "Dynamic interpretation of poles
and hansmission zeros for distributed multivariable systemr5" submittedto the 1980
European Confemce on Circuit Theory and Design, Warsaw, Poland.
H. H. Roxnbrock, State-Space and Mdtieruirrble 7 k o r y . New York: Wiley, 1970.
T-'=
[ fl, "I
I,-LM '

A. G . J. MacFarlane and N. Karcantas . "Poles and zeros of linear multivariable

systems: A survey of the algebraic, geometric and complex-variable theory," Inr. 3.
Contr., vol. 24, pp. 33-74, July 1976.
E.J. Davison and S. [Link]%"Roperlies and calculation of transmission mos of
linear multivariable systems," Auromarica, vol. 10, pp. 643-658, Dec. 1974.
A. C. Pugh, Trammission and system zeros," Inr. 1. Contr. vol. 26, pp. 315-324,
Aug. 19n.
[Link], Linear %stems. Englewood Cliffs, NJ: PRnti-Hall, 1980.
P. G. Ferreira and S.P. Bhamcharyya,. ' nblocking zeros." IEEE [Link].
O
Conv., VOL AGZZ,pp. 258-259, Apr. 1977.
C. A. Desm and M. [Link], "Desii of multivariablefeedback systems with
stable plant," Univ. California, Berkeley, ERL Memo M80/13. 1980.
E. [Link], "A new prooss interaction concept: Pinned zeros." prcprinL
R A. Poluiktov. "Constrainsdue to the plant in synthesis problemsof multidimen-
sional closed-loop systems," Automat. Remote Conrr. (Iransl. from A t d i
Telemekhm*ka. voL 27, pp. 26-33, Mar. 1966). where
F. M. callier and C A. Desoer, "An algebra of W er functions for distniuted
Linear time-invariant systems," IEEE [Link] % st. pp.
, CAS-25, 651-661,
voL
SepI. 1978; corrections in vol. CAS-26, p. 360, May 1979. ManuscriptreceivedMay 14, 1979. ?his work was supported in pan by the U S .
- ,"Matrix fractionrepresentation theory for convolution systems." presented at Department of Energy, Electric Energy Sysiems Division, under Contract EX-76Z-01-
the 4th Conf. Math Theory of Networks and Syst, Delft, The Nethexlands, July 2088, and in part by the Joint Services Electronics Program (US. Army, U.S. Navy, and
1979. [Link] Force) under Contract DUG-29-78-c-0016.
L. Pernebo,"Algebraic control theory for Linear multivariable syskms," Ph.D. The author is with the W o n and Control Laboratory. Coordinated Science Labora-
dissertation, Dep. Automat. Contr, Lund InsL Technd., Luna Sweden, 1978. tory. University of Illinois Urbana, IL 61801.
IEEE TRANSACTIONS ON AUTOMAT~CCONTROL, VOL. AC-25, NO. 6,DECEMBER 1980 1221

of this approach depends on the computational effort needed to obtain

these transformations and their inverses. The advantages of (3) are, firsf
its explicit inverse (4)and, second, the recursive determination of L and
M [[Link]
Note that lA(Fl)l> JA(Fz)Ias required by (2)assures that M is uniquely
determined by (6). It is also useful to note that L=O when A,, =O and
M=O when A,, =O. Rapidly convergent iterative methods are available Lk+l=(A22Lk+LkA1ZLk-A21)A~1 (16)
[4]for computing L and [Link] also that, in general, a permutation of Lo = -A,,A,
states in (I) is necessary to isolate fast and slow states. An algorithm to
accomplish this is given in (41. and then M from

THETWO-STAGE
DESIGN M~+I=A~~M~(A~~+LAIz)+A~~AIzLM~
For clarity we assume that completecontrollability of (l), which M,=-A,~A,,.
implies the same for the subsystems in 0, that is, for the pain ( F , , G I )
and (Fz,G2), and consider the pole placement problem for (I) using a A similar recursion is used for 2. convergence rate of these iterations
state feedback control u = k k is known [4]to be of order ck where e is the "separation ratio"
In the first stage we design an m X n , feedback matrix HI to place the
eigenvalues of F, + G , H , at the desired n , locations. The substitution of
u=u, +uz = [ H , O ] y + u z (10)

into (7)yields When this ratio is small, as in singularly perturbed systems, then a few
iterations are typically sufficient.
This property is illustrated by a power system control problem [6]
whose system matrices are

-5 0 0 0 4.75 0 0 0
0 -2 0 0 0 -2 0 0
-0.08 -0.11 -3.99 -0.93 0 -0.07 IO -9.1
0 0 132 -1.39 0 0 0 -0.28
A=
0 0 0 0 -0.2 0 0 0
0.17 0 0 0 0 -0.17 0 0
0.2 0 0 0 -0.5 0
10.09 0.01
0 -0.06 0.12 0 0.01 0 -0.11

Now we apply to (1 1) the transfynation (3) and denote it by ?. In (1 1)

the block "1,2" is zero; hence, M=O and only
&[O
0 0
0 IO 0 0 0 0 0
0 0 4 0 0 0 1
-~(F,+G,H,)-F,~+G,H,=o with open-loop eigenvalues of

needs to be solved for [Link] transformation z = Fy yields A, = -4.35 A5.6 = -0.13k0.21j

A,= -5.00 A, = -0.17

A, = -2.00 A, = -0.2
A, = - 1.39.
A& we note that the pair (5, G, +LIG,) is completely controllable
The minimum open-loop separation ratio is about 1:5 for lA51:IA41, that
and we proceed to the secondstage in which we design an m X n ,
is, with n, =n2 =4. After two iterations L is approximated by
feedback matrix Hz to place the eigenvalues of Fz +(Gz +LG,)H, at the
desired nz locations. Substituting the feedback control 0 0 0
u,=[O H2]z (13) L 1 = [-0.001
0.034 -0.001
0 0
0.048 -0.001
0.002 0.007
0.014 0.092
into (10) results in

u = ( [ H , O]T+[O H , ] f = T ) x . ( 14) and M by

Expressing T and 2 according to (4) we obtain the final form of the 0.989 0 0
feedback matrix K i n u=Kx: -0.137 1.091 0
-0.006 -0.0122.359 -1.966
HI-H2~)(Z-ML):-HlM+Hz(Z+h4)].
(15) 0.006 0.032
2.48 - 1.625
resulting in
Although this two-stagedesign is presented in terms of a N1 state
feedback, it also can be used in output feedback design. 0 0
0
APPROXIMATE DESIGN
0.W 1.323 - 1.363
The form (14)of the final result is valid for any transformations Tand
9 block diagonalizing (1) and (1 l), respectively. However, the usefulness 0
1222 TRANSAcIlONS ON AUTOMATIC CONTROL, VOL. AC-25,NO. 6, DECEMBW 1980

For desired eigenvalues is q c ) close to the block diagonal form (7)where

A,,,(des)= - 8 + 2 j A, =A, -A2,A,'AI2 Bo =B2 -A2,A,'B1.
h,(des)= -6 HI can now be designed to place the eigenvaluesof A , , +B,Hl at X,
h,(des) = -4. locations. The resultingclosed-loopeigenvalues will be of theform
,,,X( +O(c))/c. The partially closed-loop system of (33)
is now givenas
A feedback gain H I is obtained as
=[
-20.865-0:615
Hl
0 009 -0.777
-0.016
0

To design a feedback H2 for the pair ( 4 ,G2+iG,),we first find i,

-
All + B l H , +€El,
EBOH,+c2A0A2,A,'
EA,
-'A12A2,A,2A,2
-C~A,A~,A,~A~~ ly
-0.020 6.178 0
0.007 -0.033 0 0 where
(26)
-0.001
-0.002
0.057
0.132
0 0.035 -[Link] 0.015 O I Ell = A 1 2 A 2 , A , 1 + A ~ 1 A , 2 B , H , .
resulting in Next, approximatingI? in by
- 0.2 0 0 ~ = ~ B ~ H ~+ (B A~ H
, ,, ) - ' + O ( ~ ~ ) (35)
-0.167
0.164 -0.006
F 2 = [ -0.003 0-0.439
-0.015
and using y = fz, (34)becomes
0.00s 0.023 0.137 -0.262

[ G~-LC,] *= [ 0.7 0.045

0
-0.001
0.031
0.088
0.002
1 '
(28)
.
cz= [ A , , +BiH1 + € E , , + w e z )
o(r2) cAO+O(c2)
For desired eigenvalues of
h,,,(des)= - 2 + l j
X7(des)= - 1
where
&(des)= -05
E12 = -Ai2A2,AIi2-4,2.
a feedback gain H2 is obtain&
The feedback H2 is designed forthepair ( A o , B o - i ( B l / c ) ) . The
-2.615 -49.15 0 resulting closed-loop eigenvaluescorresponding to this subsystem will be
H2=[
-0.234 -1.142 . -2.266
-39.916 of the form X,,,+O(r). Substitution of(31),(32), and (35) into (15)
gives the compoate feedback matrix
The actualclosed-loop eigenvalues using the composite feedback (14) are
h i , 2 ( C L ) = -7.99+ 1.99j h5,6(CL)= - 2 2 lj
h,(CL) = - 6.09 X 7(CL) = - 1.03 This parallels the results of [l] where HI and H2 were the results of
two reduced-order linear optimal regulator problems.
h,(CL)= -3.93 Xs(CL)= -0.5 In [2] and q c ) independence of the two stages was accomplished
using a particular form of the control u2 in the design of H2, namely,
and are close to their desired locations, the worsterror beiig
u2=(Z+HIAfiiBl)a (38)

resulting in the control matrix of (36):

The approach
APPLICATTONSTO

presented
here
offers
simpler
a
SINGULAR PFXTURBATIONS
way to
known results for singularly perturbed systems[ 1L [2].
Consider the system
obtain some
(4+ ' A , ' A 1 2 ~ 0 ) ( ' + ~ i A I 1 1 ~ 1 )
~Bo(Z-H,(A,,+B,H,)-1B,)(Z+H,A~'Bl)+~~2)

Usinga kuown matrixidentity, it can be shown thatthe matrix

1
0. (39)

€&[ 4 2 1 € 4 2

If in this case L and M are approximated by

[+.I a ] .
multiplying €Bo in (39) is Z , and hence the subsystem pair for the second
(30) stage is ( A , + w e ) , Bo +we)). Since it is independent of H I up to q c )
terms, the need for taking HI into account in the second stage has been
eliminated. The composite feedback matrix of (15) now becomes

M = -Ai1Al2 (32)whichparallelsthe results of[2] and [3].

then the transformed system
CONCLUSIONS
.
V=[
+€A12A2,Afi1
A,, -€A,2A2,Afi2AU
c2AoA2,A,' €A,-, - C ~ A , A ~ ~ A ~ ~ A , ~

[
Iy An explicitly invertible transformation enables a reduced-order eigen-
value
placement
problem to be solved
in two stages. The
only require-
ment on the open-loop system is that the spectrum be disjoint Earlier
+ B , +c4i%2B0
1 (33)decompositionsapplicable to singularlyperturbed systems appear as
special cases of this two-stage procedure when the eigenvalueseparation
IEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL AC-25,NO. 6, DECEMBER 1980 1223

wassufficientlylarge. A similar designprocedure for discrete-time where Im and ker are the image and the kernel of linear maps. Let vf be
systems is developed in [;1. Finally, the decomposition is applicable to the supremal output-nulling invariant subspace. The supremal output-
designcriteria other than eigenvaluelocations and output feedback nulling controllabilitysubspace R* is defined by [2]’
problems.
R*=(A+BPl@n V * ) (2)

REFERENCES where V* and E* satisfy (1) and 4 is the range of the map B restricted to
kerE, i.e., B is spanned by the columns of BG, where the c o l u m n s of GI
R A. Yackel and P. V. Kokotovic, “A boundary layermethd for the matrix Riccati span ker E. If E=O, V* and R* are the supremal (A, B)-invariant and
equation,” IEEE Tram. A w o m . Conrr., voL AG18, pp. 17-23, Feb. 1973.
J. H. Chow and P. V. Kokotovic, “Eigenvalue placement in two-time-scale system5” controllability subspaces in ker C. In the following, V* and R* w li be
in Pror. IFAC @mp. Lurge Scale %st., Udine, Italy, 1976, pp. 321-326. referred to as the q r e m a l (A, B)-imariant and conirollabili@subqmces,
-, “A decomposition of near-optimum regulars for systems with slow and fast
modes,”IEEE Tram. A u f o m . Conlr, vol. AG21, pp. 701-705, Oct 1976. respectively.
[Link], “Iterative algorithms for the lime scale separation of Linear dynamical If the given system S is completely controllable, then the dimensions
systems,” submitted for publication.
P. V. Kokotovic, “A Riccati equation for block-diagonalization of ill-conditioned of V* and R* can be explicitly determined [l] as functions of n, the
system” IEEE Tram. A w o m . Confr, vol. AC-20, pp. 812-814, Dec. 1975. order of the system; q, the number of zeros; and d,, the degree of the
[Link]. ‘*DyynimIstate space models of electric power systems,” Dep. EL- determinant of a polynomial matrix X ( s ) ; these quantities are defined
hg., Univ. Illinols Urbana, 1971.
R [Link]$ “Twc-lime-scale discnte systems,” MS. thesis, Coordinated sei Lab, below for completeness.
Univ. Illinois, Urbana, Rep. R-839, Feb. 1979. The q zeros of S (also called invariant zeros) are those zi (multiplicity
included) for which

X ( s ) is a p x p polynomial matrix such that X(s)T(s)has a proper

On the Dimensions of the Supremal ( A ,B)-Invariant right inverse( p < m). i.e.,
and Controllability Subspaces
lim X ( s ) T ( s ) = K , : rank K,=p. (4)
P.J. ANTSAKLIS AND T. W. C . WILLIAMS s+m

Then d, 2 degree (det X(s)). Note that in [I] X ( s ) was taken to be

Ahtmct-lle dimensions of tie s up& output-nulling (A,B)- the “interactor,” amatrixwith special structure satisfying (4) and
m v h t and COntroIlability subspaces of a system are explicitly de-
uniquely defined via a known algorithm. It should be noticed that if S
termined The role of the input and output decoupling zeros h also
tlisasd. can be decoupled via linear state feedback, then d, is the sum of the
“decoupling indices” of S [I]. The main theorem of this paper can now
be stated.
INTRODUCTION Theorem I : dim P =q+dim R* where dim R* is: 1) zero whenp > m,
Among the key concepts of thegeometricapproach [3] are the and2)n-q-dTwhenp<m.*
concepts of the supremal output-nulling ( A , B)-invariant and controlla- Proof: If the system S is completely controllable, then the theorem
bility subspaces [2]P and R*, respectively. Their dimensions havebeen is true as has been shown in [I] (seeTheorem 11 and remarks preceding
explicitly determinedin [I] to be functions of the order n of the system, the example). Before the general case can be shown to be true, several
the number of zeros q and d,, and the degree of the determinant of a properties must be proved.
polynomial matrix derived from the transfer matrix T(s).The assump Assume that thesystem S is not completely controllable. Then its
tions made in [I] of the system beiig completely controllable and T(s) state-space representation can be reduced to
having full rank are relaxed here and it is shown that the same formulas
are valid in the general case (Theorem1 and Lemma 4). This is done by
establishing the relation between R* and the supremal controllability
subspace of: 1) a controllable subsystem (Lemma 2), and 2) a subsystem
withatransfermatrix of full rank (Lemma 4). In addition, the role
where the system S,: {Ac,B,, C,, E } is completely controllable.
played by the input and output decoupling zeros is discussed (Corollary Let V,* and R: denote the supremal (A, Bkinvariant and controllabil-
3) and their importance to the output stabilization problem is indicated. ity subspaces of S,.
Lemma 2: dimR*=dimR:.
MAIN RESULTS Proof: Equation (2) can bewritten as R* =(A +BPIanV*n
Let S denote the system i=Ax+Bu; y = C x + E u where AERnXn, (AIQ)).Notethat
B€RnXm, CERPXn, and EERPXm. Let T(s)=C(sI-A)-’B+E be its
V*n(AIQ)=span withspantg’=~ascan
(5)
be easily seen by restricting the system to the controllable subspace and
transfer matrix whichis assumed to be of full rank, i.e., rank T(s)=
W P ?m)-
A subspace V of the state space is an output-nulling invariant sub-
space [2] if and only if
(z)
u s i n g ( 5 ) . N o w R ~ = ( A c + B , ~ ~ ~ , n whichimpliesthat%*=span
with span r,‘ = R:; therefore,
V,*)
dim R* =dim R:.
The system S, is completely controllable; therefore, dimRr(=dim R*1
Q.E.D.

can be expressed [I] as a function of the order n, of &(A, €GjLnCx”c), q,,

the number of invariant zeros of S,, and dT which is derived from the
transfer matrix T(s) ( S and S, have the same transfer matrix). In order
or, equivalently, if and only if for some F to complete the proofof the theorem, the relations between n , q and
nc, qc must be established. The order and the zeros of S and S, are
(A+BF)VcVcker(C+EF) (1)
related via the input decoupling zeros definedbelow.

Manuscript received February 14, 1980; revised May 26, 1980.

The authors are with theDepartment of Computing and Conlrol, Imperial Cdlege, ‘<AI%) %+A%+ ... + A ” - ’ % .
London. EnglancL ZNote that n-q-d,-Owhenp=m.

0018-9286/80/12W1223$00.75 0 1980 IEEE