W81
EXACT MODEL MATCHING BY STATE FEEDBACK AND DYNAMIC COMPENSATION
B. C. Mooreand L. M. Si l ver mans
Department of Electrical Engineering 
Systems
University of Southern California
Los Angeles, California 90007
114
Definition 2. 1: Let 5, be the c l a s s of a l l r e a l m a  B??BG= CB CBG. By expanding the lefthand side
t r i x p a i r s ( F , G). Then (A, B, C, D), i s s a i d to be of (2. 5) for a general k and substituting @aJBfor
feedback equivalent (3,e uivalent) to (a, R, @, B ) (C+DF)(AtBF)JBG forall j < k, one can see that
i f and only if there e x i s t s k i , such that DG= B and
where F = GIF.
..
for k =<,1 , . , y ( y defined as in Lemma 2. 1) 3. Summ ar y and Pr oper t i es
of the Structure Algorithm
Proof: By Lemma 2. 1, (A, B, C, D), is Tzequiv The structure algorithm sequentially yields
alent to (Q,R,@, R ) , if and only i f there exists a set of quantities which are invariant under
(F,G)E 5, such that DG = B, and (F,G) transformations (G nonsingular). Unlike
the invariants utilized in [ l ] and [2], operations
(C+DF)(AtBF)kBC= W B .
k = 0, 1 , . , y. (2. 5 ) on the controllability matrix (an s X s r matrix)
are not required. Rather, operations on ma
For k = 0, this implies that DFBG=@n CBG, or t r i c e s of decr eas i ng s i ze, s t ar t i ng with D (an
mX r m at r i x) , ar e per f or med. A s um mar y of
115
this algorithm together with several of its prop p a c e of 3,. with this choice for F,G note that
e r t i e s is given below. C1 % 0 and Dl 4[Iq, 01 f o r all $ Let A:,
Let qo = rank D and let E, b e the submatrix 
A BDf C, then it follows that R, C,Ao= [O CL,]!
formed from the first qo independent rows ofD. Hence
Then So nonsingular exists such that i f
y o ( t ) = S,,y(t) = C, x ( t ) t Dou(t)
then D, = So D = [ 5
conformably with De a s C, = [
'.
: 0 ] Co = S %is partitioned
Ci ] z!
Continu . (3.1)
ing recursively, suppose after kt1 steps we have
It is c l e a r the? that the nonzero rows of C",Ai
y,(t) = C,x(t) t Dku(t) a r e given by C, since R is a pe_rmutation matrix
with C, = [z; e:]' =[E: E, 1
for each i. Let L1+, = C i * C,' ]'and let B be
and D,
and BLhave 9, rows aqd rank D = q,
[ C: ( C,A)'] , Jwl= [ 4' (C,B)'!',
Ol'where
Let E, + =
and let q,+,=
. , be the first integer such that rank LB= rank L
Then L can be viewed as the observability m f
1.
t r i x of &e pair (4, C,) with observability index
rank Jwl. Let S&, be theunique permutation
matrix such that the first qnl rows of SZ+,J,, B.
Let y (t) be partitioned conformably with D,
contains the first ab,, independent rows of J,,, a s y,(t) = t y k ( t ) f,(t)]'. An important structural
with the relative oyder maintained and its list' Lnvairant under 5,feedback i s the dependence of
mq,, rows contain the remaining rows of J
yi (t) on 7,(t) for all i. This is made evident by
also with the relative order maintained. S&,,Xks the following theorem:
the form
Theorem 3.1: Let Be= BB: and let A, be defined
above. Then, for x(t,) = x, in (2. 1) and for all u
then
116
4. Application of the Structure Algorithm
@ O Q t B o for 0 Sk S y  1 and the lemma follows.
Let S, and N, be the quantities defined by
the structure algorithm applied toJA, B, C, D), The main result follows directly by Lemmas
and define@,= [@kc:]'= S & , [ T l B i ] ' = S O B 4.1 and 4.2. Let CY=a1 .if rank q a = m((A, B, C, D),
where the "barred" quantities are the top qa rows right invertible), and let o= y a s defined m Lem
of SC, a_nd ,S,B. Now define recursively, for kg 0 ma 4. 2 i f q, < m.
8 L j @" : (@kg)'], = [b; B)'_l, 
[@k+l: @r+l I 1 =  5 k + 1 8 k + 1 L 4 4  1 =[TLl : B:+l 15'='  Theorem 4. 1: Let (F, G)E 5, then (A, B, C, D),
S,+,$, where 4+1
and b a r e the f i r s t q k + l PA+ (ff,B,@,S ) Uif and only if
rows of S,+18,+1 and S,+; jkcl respectively.
It is clear that If (A, B, C, D), i s y1 ecpiv
d e n t to (Q,B,C, B ) , then it must be true that& = 0
f or all is 0 since the output equation f, (t)=,C,x(t)
isinvariantunderfeejback.Moreover, if =0
for 0 S i S a and rank B, = q, then
Theorem 4.1 provides a considerably more
N,w(t) = [w:(t) ~ G ( t ) ] ' = C , z ( t ) t B , v ( t ) efficient test for aaequivalence than that given
in $ 2 (and consequently those given in [ l ] and [ 2 ] ) .
Furth_ermo_re, for any nonsingular matrix G such A set of readily computed necessary conditigns
that D, G= B,, we canlas in Theorem 3.1 azd a r e provided by (4.1) and the fact that rank Ba
Equatjon ( 3 . 2 ) define B z = GI D$,Q0=a~b$@, , must be qa by (4. 2). If these conditions are sat
Bf+$ and write isfied, we then must determine i f a pair (F, G) E
8, satisfying (4.2) exists. However, this is in
general a simpler problem than that originally
posed sinceadditional constraints on G have been
specified (D,G=Ba, rather than DG= b). In par
ticular, i f q a = r (in which case (A, B, C,D), i s
Q, (8) isthe transfer function of the system
(ffo,Bo,@a,O)* 
l ef t j nver t i bl e [ 4] ) G iuniquely
s specified a s G =
Drl& leaving an unconstrained set of l i near equ a
tions in F. This important case is detailed below,
Lemma 4.1: Let (F,G) Es,. T h p H(A, B, C, D)
With G = E i l & a n d F = G'IF, it follows f r om
A
' H(G B,@, b ) i f and onlyif,B, = 0 for OSiLa, Theorem (2. 11) that (4. 2) holds i f and only if for
H(A,B, C,, D, 1 Le, H(Q,B,C, ,Sa ) and Q, ( a ) =
Pa ( 8 ) .  .
k = 0, 1, , , , y Jy def i l ed l or the r educeds ys t ems)
FAkBG=B'l[@,dCBC,ABGEJ,,C$+JBFAJlBG]
Let 6 = BG then this can be r ewr i t t en as
Proof: Sufficiency i s c l e a r fro m t h e above dis
cussion. The fact that the first and second con i?[$I A$': *!
0 .
AY$] = [ Q o : Q, : * * *
ditions a r e n e c e s s a r y follows fro m the (F, G) in
variants discussed in $ 3 . The third condition is where
necessarysince G nonsingular implies that Qk

= b~'(&a'BZaAkB*ZFlflJm~,1)
k
H(O,B,@,,Sa) mu8 t haKe iEdependent rows and j
thus that Pa ( 8 ) H(Q,B,C, ,Sa = Qa (s)H(Q,B,& ,>, 1 Notice that the Q, ' s a r e computed sequentially s o
i f and only i f Pa (4= Q, ( a ) . that nonexistence of F can be checked after each
Let 6 and 8 be the respective controllabil computation,
itiee indices of (Ao,Bp) and (Qo,B,). Let p be a s
\"'")E]
computed and expanded a s i n (4. 3).
'k+l ["']G=
ekB L 0 z Sk+l['ZrB ]
t
. 
since F,=G'a;' , a_nd the s o n z e r o LOWS of
Co& a n d &,e
a r e Ck and , then G & B o =
117
the dynamic case. It will now be shown that a
finite subset of these constraints is also sufficient
f o r the existence of a dynamic compensator. We
whlc exists
118
Theorem 5.3: (A, B, C , D ), i s dy33amically equiv form. ) If this i s the case, F is chosen so that n
alent to (a,B,@, B ) , i f and only i f Bk = 0 for k = of the poles of (6. 1 ) a r e moved to equal the poles
0,1,. *.
,n s s . of (6.2), and the remaining (sn) ar_e used to can
cel excess numerator zeros. That B k = p P  l B =0
Proof: Necessity follows directly from Lemma for 0 S k S a 1 i s a sufficient condition for dynamic
6.1. equivalence is clear since this implies s  a l s;
The sufficiency proof is as follows, First n4, o r s t t S 5 n t i , , which means that
i t is shownthat there exists a r o e r r a t i o n a l = H'(A, B, C, D)H(c7,9,@,B ) does not have more
matrix &(s) such that H(A, B, C*s)H(O,B,@,B),,, zeros than poles.
Let the structure algorithm be ap lied to Now let us consider two examples, the first
(A*, B*, C*, D*),+n . Then, since jk=
0 for OSkP of which was considered by Wolovich [ l ] .
*s, the f i r s t nts steps are exactly the same as
those of the algorithm applied to (A, B, C, D),
Hence q 6 = 9". Letting Hl(s) = H(A, B, C, D ) m , a n d
. Example 1: Consider the system representation
(A, B, C, Dl3 where
H,(s) = H(c~,%,C,B),, we see from Theorem 6.1
and Corollary 6.1 that H, (.s)Hf(s)H,(.s) = H,(s).
H ~ ( s ) H , ( s )i s a proEer rational matrix since it
can be written P ( s ) N & ( s ) w h e r e P ( s ) j s % p r o p e r
rational matrix, and N,H,(s) = H(Q,B,C,, &),,
Let (Al, B,, C,, D l ) & be a realization for
H$(s)H,(s). One can easily verify that z=k and
yields ~ ~ F , ~ G ) = H , ( s )t H , ( s ) H , ( s ) =
H(LtB*F,gG,
H(C7, B, @, B )
6.InterpretationsandExamples
To answer the question, we apply the structure
F o r the case in which ( z . 1 ) and ( 2 . 2 ) a r e s i n  algorithm to (A, B, C, D), and H(s).
gle inputsingle output systems, the necessary
and sufficient conditions of $4 and $ 5 have clear Step 1: Apply the f i r s t s t e p of the structure algor
interpretations. Let ithm:
b o s a l t b l s a 1 1 t * * * t b
B, H(A, C, Dl, = (6. 1)
s~+a,s~lt***ta,
~os%pls~llt..*t
H(a, 6,C, B ) , =
pa a (6.2)
snt cylsnlt *tan
where we assume, for the sake of illustration, The first necessary condition, Beb = 0, is satisfied
that s > 4 and n > A,. Applyingthe s t ru c t u re a l 
gorithm gives No= l , N,= d k / d t r , a n d a = s  a , is Step 2 : Second step of the Structure Algorithm:

the first integer i for which CAI 1B #_O.Also,_
D,= CAa1B,B6=W61B,andCr= CAk,& = C?@,,B,=
@1B for 1 S k e a . Since 3, is a s c a l a r , the_
necessary conditiols for Z,equivalence that a,= 0
f o r 0 d k I a 1 and B6 # 0 simply imply that s  a , ,
the relative order [8] of H(A, B, C, D), , equals
n  a,, the r e l a t i v e o rd e r of H(Q,R,C, B ) , . This is
an understandable condition since relative order
is invariant under feedback.
Assume that the necessary conditions hold. The structure algorithm terminates at this step
Then G, a s c a l a r , is given by5;,s6 = since (A, B, C, D)3 is clearly invertable. 7, has
full row rank and
ca lB 
 Moreover, sincestate
feedback
CA61B bo '
does not change the numerator of (6.1) and can
be used to arbitrarily place the poles, there ex
G=6i1Bl = [: :]
i s t s a solution (F, G) if and only i f the zeros of
( 6 . 2 ) a r e contained in the zeros of ( 6 . 1). (Equa
tion (4.3)will have a solution i f and only i f this is
t r u e , a s can be seen when (A, B) is in companion
119
details of this generalization are presently being
considered.
8. References
[ l ] W. A. Wolovich,"TheApplication of State
Feedback Invariants to Exact Model Match
ing, '' Fifth Annual Princeton Conference on
Information Science and Systems, Princeton,
New J er s ey, M ar ch 1971.
A solution clearly exists.
[Z] S. H. Wang and C. A. Desoer,"TheExact
Step 4: Find AEj,B,,QB, CIA?, Q, and check for a Model Matching of Linear Multivariable
solution to 3~3:AS ] = [Q, : Q, 1. If a solution Systems, ' I submitted for publication.
exists,compute F.
[3] L. M. Silverman,"Inversion of Multivari
able Linear Systems, ' I IEEE Trans. Auto
maticControl, vol. AC14, pp. 270276,
1969.
[4] L. M. Silvermanand R ,J. Payne,"Input
Output Structure of Linear Svstems with
Application to t h e Decoupling Problem,
At this point a unique solution exists: and FA' =Qi SIAM J. onControl, vol. 9, pp. 199233,
..
for i = O , 1,. , y. 1971.
7. Conclusion 51 R. T. Curran and G. F. Franklin, "An
Exact Solution to the Linear Multivariable
An algorithm has been given which solves the ModelFollowingProblem,submittedfor
problem of exact model matching by state feed publication.
back for a class of linear systems, and which
completely solves the problem of model match 61 B. L. Ho and R. E. Kalman,"Effective Con
ing by dynamic feedback. This algorithm is seen struction of Linear StateVariable Models
to be considerably more efficient and complete fromInput/OutputFunctions,Proc.3rd
than the solutions offered in [ l ] and [2] for sev Allerton C o n f . , pp. 449459.
eral reasons. First, no initial coordinate trans
formation is required. Secondly, each step of [7] L. M. Silverman,"Realization of Linear
the algorithm tests a s e t of conditions which a r e Dynamical Systems, 'I to appear IEEE Trans.
n e c e s s a r y f o r the existance of a solution (dynam Automatic Control, 197 1,
ic or state feedback) so that nonexistance may be
detected at any step. If a state feedback solution [8] R. W. Brockett,"Poles,ZerosandFeed
_ .
does not exist, the algorithm can be continued to
t e s t f o r a dynamic solution. The algorithm canbe back: State Space Interpretation, (I IEEE
Trans.AutomaticControl, vol. AC10,
applied directly to (T, B,C, &., o r H(0, n,C, P),,
whichever is available. pp. 129135, 1965.
There are several questions which are still
[9] A. S. Morseand W. M. Wonham,"Decoup
open for consideration. A s mentioned earlier,
ling and Pole Assignment bv Dvnamic Com
our solution gives complete necessary conditions
only for the existence of a solution (F, G) with G
pensation, SIAM J'.
Contrdl, 401. 8, pp.
317337,August 1970.
nonsingular. The general problem of model
matching by state feedback remains unsolved,
F u r t h e r m o re no solution has been given to the
problem of obtaining a dynamic compensator of
minimal order. Another question which has not
yet been resolved is that of designing a stable
compensator. For the case in which H(s) =
H(A, B, C, D) is square and nonsingular it is
c l e a r f r o m the f o r m of the precompensator design
that a sufficient condition for stability is that all
of the zeros of the numerator of detH(s) which
a r e not also poles (of the same degree) of H,(s)
H ( d , B , C , b), lie strictly in the left half plane, In
general, however H t ( s ) i s not unique and the
problem is nontrivial. Some preliminary re
sults have been obtained by the authors and will
be reported elsewhere.
The approach taken here can be generalized to
a broad class of timevariable systems (the class
of "regular systems" as defined in [3]). The
120