Decoupling and Pole Assignment

in Linear Multivariable Systems:

A Geometric Approach
W. M. WONHAM AND A. S. MORSE

IN

the theory of isolated dynamical systems (flows on manifolds, etc.) the notion of an invariantset serves a crucial role for
classifying the nature of a system. A subset of the state space is
said to be invariantif startingin it, the trajectorymust remainin it
for all time. For control systems, this property is not particularly
useful. For example, if a system is controllable, in the sense that
any initial state can be drivento any terminal state, then there are
obviouslyno non-trivial invariantsets in this sense. However, if
one modifies"must remain" into "can remain" in this definition,
then a very useful notion is obtained. This is the key observation
that underlies the notion of invariance introduced in the paper
by Wonhamand Morse (see also [1]).
Wonham and Morse consider finite dimensional linear sys=
tems. The linear subspace V of the state space X of
Ax + Bu is said to be controlledinvariant[or (A, B)- invariant,
as they call it] if for any Xo E X there exists an input u(.) such
that the resulting solution x() with x(O) = Xo satisfies x(t) E V
for all t E IR. It turns out that this condition is equivalent to
(A, B)-invariance (AV S; V + im(B)) and to the possibility of
making the subspace invariantby state feedback (3F such that

1tx

(A

+ BF)V

S; V).

This concept, together with the related notion of controllability subspace and the dual notions of conditional invariance,
opened the road to the elegant "geometric" theory of linear systems and a wide range of applications,especially to problems of
disturbance decoupling,non-interactingcontrol, and regulation.
Indeed, the decoupling problem was the main motivationfor the
developmentof the geometric approach.The early contributions
to the study of this problem are reviewed in Wonham's seminal
book [20, Sec. 9.10]. The problem of disturbance decoupling
with output feedback was completely solved in the linear case
in [16] and, incorporating stability,in [18].
Soon after the appearance of the notion of controlled invariance for finite dimensional linear systems, it was generalized in
many directions, notably to infinitedimensional systems [4], to
"almost" versions of these notions [19] (that made it possible
to treat approximate decoupling problems and high gain feedback), and to nonlinear systems. It also sparked a totally new
approach to discrete event systems; see, e.g., [15].

In the late 70s and in the early 80s the notion of controlled invariance was extended to nonlinear systems, and this extension
opened the way to the systematic development of methods for
the design of nonlinear feedback laws. In the context of nonlinear systems, the notion of invariantsubspace generalizes in two
distinctways:invariantsubmanifoldsandinvariantdistributions.
Givena vector field f, an f -invariantsubmanifoldis a "surface"
with the property that any integral curve of f which intersects
this surfaceis entirelycontainedin it, whilean f -invariantdistribution is, essentially, a partition of the state space into a "family
of surfaces" with the property that the flowof f carries surfaces
into surfaces (if, in particular, one of the surfaces of this family
contains an equilibrium of f, then this surface is an invariant
submanifold,otherwiseit is not). In light of this, the extensionof
the concept of controlled invarianceto nonlinear systems leads
to the notions of controlledinvariantsubmanifolds/distributions
.as objects that can be rendered invariant by means of suitable
feedback laws. Thus, for instance, in a control system
d
dt x

= f(x) + g(x)u

with state x E lRn and u E R, a submanifold/distribution is controlled invariant if there exists a feedback law u = a(x) which
renders it invariantfor the "closed-loop" system
d

dt x

= f(x) + g(x)a(x)

In the nonlinearcase it is the notionof controlledinvariantdistribution (in contrast to invariantlinear manifold, around which
the linear theory is centered)that plays a fundamentalrole in the
analysis and solution of the problems of disturbancedecoupling
and noninteracting control. Decoupling a fixed output from a
fixed disturbance input is achieved by imposing, by feedback,
the existence of an invariant distribution that renders the influence of the disturbance unobservable by that particular output.
This fruitful domain of research was opened by the works [9]
and [6], and then continued by various authors (see, e.g., [13],
[10], [5]). The problem of noninteracting control with stability
required some extra effort in view of the existence of certain

321

obstructions that do not have a counterpart in the case of linear

systems, but eventually was solvedin [8], [17],and [2].
The notion of invariant manifold led to the introduction of
the conceptof zero dynamics, the nonlinearanalogue of the dynamicsassociated withthe numerator of a transferfunction. The
zero dynamics of a nonlinearsystemis a dynamical systemdescribingall (forced) statetrajectories that are consistentwiththe
constraint that the output is identically zero. As in the case of
linear systems, it can be shown (under appropriate "regularity"
hypotheses) that those trajectories are actuallyfree trajectories
of a feedback-modified system,whosestate-space is a submanifold (a controlled invariant submanifold) of the original state
space. If all such trajectories converge as time tends to 00 to an
equilibrium, the system has properties analogous to those of a
linear minimum-phase system. This notion, developed in a few
preliminary conference paperswhoseresultsare summarized in
[3],hada significant impactin theanalysis of theproblemof output trackingand, aboveall, in the systematic designof adaptive
feedback laws for systemsaffected by parameteruncertainties,
as shownin the monographs [12], [11].
REFERENCES

[1] G. BASILE AND G. MARRo, "Controlled and conditioned invariant subspaces in linear system theory," J. Optimization Th. & Appl. 3:306-315,
1969.
[2] S. BA'ITILOTII, "A sufficient conditionfor noninteracting controlwith stability via dynamicstate feedback," IEEETrans. Aut.Contr., AC-36:10331045, 1991.
[3] C.I. BYRNES AND A. ISIDORI, "Asymptotic stabilization of minimum-phase
nonlinearsystems," IEEETrans. Aut. Contr., AC-36:1122-1137, 1991.
[4] R.F. CURTAIN AND H. J. ZWART, An Introduction to Infinite-Dimensional
LinearSystemsTheory, Springer-Verlag (Berlin), 1995.
[5] W.P. DAYAWANSA, D. CHENG, T. J. TARN, AND W.M. BOOTHBY, "Global

([, g)-invariance of nonlinearsystems," SIAM J. Contr. Optimiz.,26:11191132, 1988.

[6] R.M. HIRSCHORN, "(A, B)-invariant distributions and disturbancedecoupiing of nonlinearsystems," SIAM J. Contr. Optimiz., 19:1-19. 1981.
[7] A. ISIDORI, Nonlinear Control Systems: An Introduction, 3rd ed., SpringerVerlag (Berlin), 1995.
[8] A. ISIDORI AND J. W.GRIZZLE, "Fixed modesand nonlinearnoninteracting
control with stability," IEEETrans. Aut. Contr., AC-33:907-914, 1988.
[9] A. ISIDORI, A. J. KRENER, C. GORI GIORGI, AND S. MONACO, "Nonlinear
decouplingvia feedback: A differential geometricapproach," IEEETrans.
Aut. Contr., AC-26:331-345, 1981.
[10] AJ. KRENER, "(Adf,g), (adf,g) and locally (adf,g)-invariant and controllabilitydistributions," SIAM J. Contr. Optimiz., 23:523-549, 1985.
[11] M.KRSTIC, I. KANELLAKOPOULOS AND P.KOKOTOvIc,NonlinearAdaptive
Control Design, Wiley(NewYork), 1995.
[12] R.MARINO AND P.TOMEI, Nonlinear Control Design:Geometric, Adaptive,
Robust,PrenticeHall (New York), 1995.
[13] H. NUMEUER AND AJ. VAN DER SCHAFf, "Controlledinvariance for nonlinear systems," IEEETrans Aut. Contr., AC-27:904-914, 1982.
[14] H. NUMEIJER AND AJ. VAN DER SCHAFf, Nonlinear Dynamical Control
Systems, Springer-Verlag (New York), 1990.
[15] PJ. RAMADGE AND W.M. WONHAM, "On the supremal controllable sublanguageofa givenlanguage," SIAMJ. Contr. Optimiz., 25:637--659,1987.
[16] J'M, SCHUMACHER, "Compensator synthesis using (C, A, B)-pairs,"
IEEE Trans. Aut. Contr., AC-25:1133-1138, 1980.
[17] K.G. WAGNER, "Nonlinearnoninteraction with stabilityby dynamicstate
feedback," SIAMJ. Contr. Optimiz., 29:609--622,1991.
[18] lC. WILLEMS AND C. COMMAULT, "Disturbancedecouplingby measurement feedbackwith stabilityor pole placement," SIAMJ. Contr. Optimiz.,
19:490-504, 1981.
[19] J .C. WILLEMS, "Almost invariant subspaces: An approach to high gain
feedback design-Part I: Almost controlled invariant subspaces, Part IT:
Almost conditionally invariant subspaces," IEEE Trans. Aut. Contr.,
AC-26:235-252, 1981,and 27:1071-1085, 1982.
[20] W.M. WONHAM, Linear Multivariable Control: A Geometric Approach,
1st edition, Lecture Notes in Economics and Mathematical Systems,
Vol. 101, Springer-Verlag (New York), 1974; 2nd edition, Applications
in Mathematics series, Volume 10, Springer-Verlag (New York), 1979.

A.I. &J.C.W

322

DECOUPLING AND POLE ASSIGNMENT IN LINEAR

MULTIVARIABLE SYSTEMS: A GEOMETRIC APPROACH
W. M. WONHAMt AND A. S. MORSE:

1. Introduction. The current interest in linear multivariable control has

led to several algebraic results with important applications to system synthesis.
In particular, the problem of decoupling of individual system outputs by means
of state variable feedback was studied by Rekasius [1], Falb and Wolovich [2]
and Gilbert [3]; the problem of realizing arbitrary pole locations in the closed
loop system transfer matrix was investigated by Wonham [4] and Heymann [5].
In the present article, new results are obtained along these lines. In 3, the problem
of neutralizing the effect of disturbances with respect to a specified group of
output variables is solved. In 4, the concept of a controllability subspace is
introduced and its relation to pole assignability is investigated. This material
is preliminary to the formulation of a general problem of output decoupling
in 5. In 6 and 7, necessary and sufficient conditions for decoupling are obtained in two specialcases; the results of 7 complement and extend those obtained
previously in [1], [2J and [3]. In each case, the problem of pole assignment is
solved completely.
Our viewpoint is that such problems are usefully treated in a geometric
framework in which both definitions and results become intuitively transparent.
In this way, entanglement at the outset in a thicket of algebraic calculations is
avoided. Of course, for applications, it is necessary to translate the geometric
criteria into matrix operations suitable for computation. This matter will be
considered in a future article.
2. Notation. The control system of interest is specified by the differential
equation
(2.1)

X(I)

= Ax(.t) + Bu(t'

with x an n-vector, u an m-vector and A, B constant matrices of dimension, respectively, n x nand n x m. Here and below, all vectors and matrices have realvalued elements. Script letters denote linear subspaces: 8" is real n-space; 'I '1 is
the orthogonal complement of the subspace .'1"'; 0 denotes both the vector zero
and the zero subspace.
If K is a matrix, {K} or .K is the range of K, and. i' '(K) is the null space of
K. If K is of dimension J.l x v and 'I.' ' c till, we write K- 1 1 . for the subspace

{z:zet8''',KzE 1"} C s:
The controllable subspace of the pair (A, B), written {Alat}, is defined as
{AI~}

= &4 + A~ + ... +

An-lYl.

Received by the editors February 3~ 1969, and in revised form June 4. 1969.
t Office of Control Theory and Application, NASA Electronics Research Center, Cambridge.
Massachusetts 02139. The work of this author was supported by the National Aeronautics and Space
Administration while he held an NRC postdoctoral resident research associateship.
: Office of Control Theory and Application, NASA Electronics Research Center. Cambridge..
Massachusetts 02139.
Reprinted with permission from SIAM Journal on Control, W. M. Wonham and A. S. Morse,
"Decoupling and Pole Assignment in Linear Multivariable Systems: A Geometric
Approach" Vol. 8, 1970, pp.1-18.

323

Thus, {AltI} is the largest subspace of 8" which the control u() in (2.1) can
influence. Observe that {Alai} is an A-invariant subspace of In.
With (2.1), we consider the auxiliary equation

(2.2)

y(t) = Hx(t),

where H is a constant q x n matrix. The vector y is the output.

Equations (2.1) and (2.2) play no essentialrole but serve to guide the investigation.
3. Localization of disturbances. In place of (2.1), consider the perturbed
system

= Ax(t) + Bu(t) + De(t),

d matrix and c;( ) is a disturbance input. If u(t) =

(3.1)

x(t)

where D is a constant n x
Cx(t)
+ v(t) (where v( ) is an external control input), then the output y( ) will be unaffected by all possible c;( ) if and only if {A + BCIP}} c ~K(H). This suggests
the problem: given A, B, !J) c tI", .IV c tfn, under what conditions does there
exist an m x n matrix C such that {A + BCI9}} c ..AI? If C exists, the effect of
disturbances is, in an algebraic sense, localized to ./V:
THEoREM 3.1. There exists C such that {A + BCI~} c ,iV if and only if
~ c 1/;where l ' is the maximalsubspace such that

(3.2)
Furthermore
(3.3)

j/"

is given by "I/' =

1,'(0)

.,N'~

j"{lI),

j"(i)

where

rv: 1) n A - l(aJ +,,"(i-1}),

i= 1,2,,v,
and v = dim .A":
Here and below, "maximal" ('''minimal'') mean l.u.b. (g.l.b.) with respect to
the usual partial ordering of subspaces by inclusion.
To prove the theorem we need two auxiliary facts.
LEMMA 3.1. Let XiEIPI, UiEsm, i = 1, "', N, and write X = (Xl"'" XN),
U = (U1'' UN)' Thereexistsanm x n matrix Csuch that Cx, = Uhi = 1",, N,
if and only if %(X) c ..(U). C always exists if the Xi are linearly independent.
The simple proof is omitted.
LEMMA 3.2. Let "f'"' c tin. Thereexists an m x n matrix C such that (A + Be) 1, .
c 1''"' if and only if Ar" c 91 + 'I':
Proof. Necessity is clear. For sufficiency, let VI' , VII be a basis of 1<
Then AVi == BUi + Wi for some Ui E 8 m and Wi E r; Choose C, by Lemma 3.1,
such that CVi = -Uh i = 1, ... , u; then (A + BC)v; = Wi'
Proof of Theorem 3.1. For sufficiency, (3.2) implies l' C .. f' and A l' . c 91
+ "f': By Lemma 3.2, there exists C such that (A + BC)"f' c l' ~ Then

{A

+ BCI,q)}

{A

+ Belt '}

= 1" c . ,~<

The maximal property of 'j' ' was not required.

For necessity write {A + BC'~} = If: Then
(3.4)

lI'~

c .'V:

A }f" c
324

d4

1/':

If '~is the class of all jf' c tin which satisfy (3.4), then clearly 0 E 1Y and 11' is
closed under addition. Hence, 1f'''contains a (unique) maximal member 'II: Then
~ c 'II'" c 'I" and r: satisfies (3.2).
To prove the second statement of the theorem, observe that 1"'(0):::> r;
and if ; '(i- t) ~ 1~ then ,,'(0 ::> 1/ A - 1(~ + 1"') = '1/~ Thus, t"(i) ::':) '/.''' for all
i; and since ;,'(i) c t (i-1). there is a least integer j such that ~.(;) = r '(j) if i ~ j.
Since 'f '(j) :::> ;'- and 1,/"0) satisfies (3.4), tJ 'Cj) = 1- ~ Clearly, 0 ~ j ~ v: and if
f} c 1-' we even have 0 ~ j ~ v - dim ~.
Remark 1. Theorem 3.1 depends essentially on the fact that the class 11'
determined by (3.4), or equivalently

n A - 1(~ + 11' ")} ~

'1Y = {lI": 1r c . J'

has a maximal element 1': Furthermore, 1" is defined constructively by means of

(3.3). This fact will be used without special comment in the following sections.
4. CoatroUabiUty subspaces. In regard to the system (2.1), suppose that a
subspace 91 c I" is selected and that it is desired to modify the system in such a
way that ~, but no larger subspace, is completely controllable. This aim is to be
realizedby feedback of state variablesand by formingsuitable linear combinations
of control variables: that is, by setting u = ex + Kv, where K is an m x Ill'
matrix for some m' ~ m. Then (2.1) becomes

x=

(A

{A

+ BCI{BK}}

+ BC)x + BKt'

and we require
(4.1)

= .-Jt,

Condition (4.1) can be expressed more neatly by noting that {BK: c tI and the
following.
LEMMA 4.1. If rJ c !M and {AI~} = 91, then {AI~ n JI} = ~jt. Conoersely, {!'
{AliM n ~} = fJI, thereexists a matrix K such that {AI{ BK }} = iJt.
Proof. {AI!f} = fJI implies rI c ~, so !f c 91 n ~, and thus ..Jt = l AI~:
c {AlflI n ~}. Also, A9t c ~ implies A(aJ n at) c .~; by induction Ai(:jJ n .11)
c ~,j = 1,2, ,and so {AI~
at} c YI.
For the converse, let bi , i = 1, ... , m, be the ith column of B and let :"i~
j = I, . , m'} be a basis of 81 n 91. Then

rj

L k;jb;,

= I ~ , .... m',

i= 1

for suitable ki j , and we set K = [ki j ] . This completes the proof of the lemma.
By Lemma 4.1, we can pose the synthesis problem as follows:
Given A, Band rJI, find conditions for the existence of C such that
(4.2)

{A

+ BCI~

n 9t} = JI.

If such a C exists, we call Yl a controllability subspace of the pair (A, B). Observe
that at = 0 and fJl = {AI~} are controllability subspaces.
Controllability subspaces can be characterized as follows.
325

THEOREM

4.1. Let A, B, .~ c I" be fixed. iJI. is a controllability subspace

(A, B) if and only if

(4.3)

;~

A.JI c

~r

+ .Jt

and
.~

(4.4)

= if,

where ~ is the minimal subspace such that

~ = .'if

(4.5)
Furthermore, :i =

.\jI(P), where

p = dim :JI and

.~(O)

(4.6)

n (A~ + :M).

= 0,
;=1 . 2.,1l.

WriteC for the classof matrices C such that (A +

theorem we need two preliminary results.
LEMMA 4.2. Let ~ c .rJl. For all C E C,

BC~ c

.'iI. To prove the

+ (A + BC).j = .:JI n (Abf + :JI).

Proof. Let C E C. Then (A + BC).j c .rN and A.J + .M = (A + BC~.J + .~.
Jl

By the modular distributive rule for subspaces,

.:Jt

LEMMA

n (A~ + ;M) = .if n [(A + Be).} + .A]

= (A + BC~~ + .~ n ;JI,

4.3. I)' C E C then

i

L (A + BCP-l(~ n :it) =

(4.7)

;.jf(i).

j= 1

where the sequence [JI(l) is defined by (4.6).

Proof. Equation (4.7) is true for i = I, If it is true for i = k - I.. then by
Lemma 4.2,
Il

L (A

+ BCY- l(aI n ..JI) = ~ n .JI + (A + BC)af(Il-

I)

j= 1

= .JI n (A9r(k-l) + .rM)

= .;J1(11.,.
Proof of Theorem 4.1. By Lemma 3.2, C is nonempty if and only if (4.3) is
true. Let
.lJI = {A

(4.8)

+ BCltf nat},

Then C e C. By Lemma 4.3,

.. jI =

L" (A + BCY-

(~

j= 1

326

n .:Jr.) = .Jttll) =

;:JI(P.

Conversely, if .'1t = ~(n). then (4.8) is true for every C E C. It remains to show that
(4.5) has the minimal solution g,(p). By induction on i in (4.6), it is seen that
Jt(i) c ~. i = 1,2. for every solution .j of (4.5), and that the sequence .jf(it is
monotone nondecreasing. Hence, there is Ji ~ P such that .ljI(l) = ..Jt(Il) for ; ~ JL:

in particular, .tjf(p) c if and ;jf(II) satisfies (4.5),

Remark 2. If:~ is a controllability subspace, then it was proved incidentally
that
~~

= {A + BCI.qd n ~jf~

for every C such that (A + BC).~ c ~jf. This fact will be used later without special
mention.
Consider now the problem of assigning the eigenvalues of the restriction of
A + Be to fJt. It will be shown that there is complete freedom of assignment and
that simultaneously the control v introduced earlier can be made a scalar: i.e., in
(4.1) K can be made an m-vector (111' = 1). For this, recall [4] that a subspace .1' is
A-cyclic if there exists x E fl' such that {AI {x l} = PI'; that is. if .Uj' contains a
generator x. Thus we can take m' = I if and only if.~ can be made (A + BC)-cyclic
and ~ n PA contains a generator.
THEOREM 4.2. Let (4.3) and (4,4) hold, and let ~1' , ~/' be arbitrary real
numbers (p = dim 91). Then C canbe chosensuchthat (4,2);s trueand.'1I is (A + Be)cyclic with characteristic polynomial
I'

(4.9)

L ~iAi-l,

; ..p -

i=1

lf 0 b e fJI

fJI is arbitrary, C call be chosen so that, ill addition.. b generales .11.

Proof', By Lemma 4.3 and Theorem 4.1, C is nonempty and

{A + BClbit

(4.10)

n .jf} = .jf

for every C E C, Choose C 1 E C arbitrarily and write A + Be 1 = A I ' Let

b 1 = b E!JI ~ and let PI be the largest integer such that the vectors

bt,Atb l ,
are independent. Put r 1 = b, and rj

'I

A/ll-lb l

= A1rj - 1 + b1,j = 2,, Pl' Then riE.1f

and the ri are independent. If PI < p, choose b 2 E~:Jt

n .A such that r 1" , '."."

b2 are independent; such a b2 exists by (4.7). Let P2 be the greatest integer such
that

are independent, and define

r p t + i = A1' PI + i -

+ h2 "

I ....... P2'

Then r. , , .. , 'Pl are independent and in~. Continuing thus, we obtain eventually
r l ' , rP independent and in .~, with the property
ri+l

where hi E Jt

= Atri + bi ,

i = I....... p - I.

n 94. Now let C2 be chosen such that

Be 2';'= hi'
327

; = l , ... " p,

where bp E;;t n JI is arbitrary. Since bi = BUi for suitable Ui' and the r, are
independent, Lemma 3.1 guarantees that C2 exists. The situation now is that
i=I~.p-l

and

By independence of the r i'

+ Be 2H r 1 }-;. = .11;
A + B((~l + e 2 ) with generator

.: A 1

E.*

that is, YI is cyclic relative to

r, = b,
n ;14.
It is wellknown [4] that now an n-vector c can be found such that A + B(C1 + ('2)
+b1c' (restricted to (1) has the characteristic polynomial (4.9). Setting b, = Bg
for suitable gEl"', it follows that the matrix
C =

(~I

+ C 2 + gc'

has all the required properties.

Remark 3. The result that any nonzero vector in dI ;~ can serveas generator
is an extension of the useful lemma in [5].
Remark 4. If 9t = 8", (4.3) holds automatically and (4.4) amounts to
{AI&4I} = tin, i.e., complete controllability of (A, B). Then Theorem 4.2 yields the
known result [4] that controllability implies pole assignability. The construction
just used furnishes a simpler proof of this fact than that in [4J.
It will be necessary later to compute the maximal controllability subspace
contained in a given subspace .C/. For this, let '1~ be the maximal subspace of .(/
which is (A + BC)-invariant for some C (recall Remark 1 following Theorem 3.1 ):
and let C("Y)be the class of C for which (A + BC)"f~' c -1':'
THEOREM 4.3. If C E C("Y), the subspace

(4.11)

jj

= {A + BCI31 n

17'}

is the maximal controllability subspace in .V}.

Proof. By (4.2) and Lemma 4.1, ~ is a controllability subspace. Furthermore,
by Lemma 4.3 with C(1,,4) in place of C. ~ is independent of C E C(T) and so is
uniquely defined. Now suppose
;j = {A

+ BCI&it n di},

.lI c

.Y

fA is (A + Be)-invariant and 17" is maximal, there follows .iI c '1~. Let

f" = ;j E9 f~ By the construction used in proving Lemma 3.2. a matrix C exists

Since

such that

cs c.
s

xE.i~

Then C E C(Jil, and

~ = {A
c {A

+ BCI~ n ~}
+ BCI8I n 1~'}

= ~i;

that is.. ~ is maximal.

328

s. DecoupUng of output variables: Problem statement. Consider the output

equation (2.2), with
H=

(S.l)

where Hi is of dimension qi x n.. i = I, ... k, k

(2.2) can be written
'l

2" til + .. , + 'It = 'I. Then

i = 1, ... k .

(5.2)

'I

where Yl is a q,-vector. The vectors Yi may be regarded as physically significant

groups of scalar output variables. It may therefore be desirable to control completelyeach of the output vectors Yi individually, without affecting the behavior
of the remaining Yj, j =F i. This end is to be achieved by linear state-variable feedback togetherwith the assignment of a suitablegroup of control inputs to each Yi'
That is, in (2.1) we set
k

(5.3)

= ex + L

i= 1

For Vi to control Yi completely, we must have

(5.4)

where Pltj is the range of Hi' Since the ith control r, is to leave the outputs Yj'
j :#: i, unaffected, we require also
(5.5)

Recalling the equivalence of (4.1) and (4.2), we can express conditions (5.4)
and (5.S) more neatly as follows. Write tI" = I and
(5.6)

Then our problem is: Given A, Band . ~ i ,

bility subspaces ~ I '

'l

L k .
'I

ti, find a matrix C and controlla-

.Jt" with the properties:

, .

(5.7)

~~i = {A

(5.8)

.Jti

(5.9)

:Jt i c:

+ BC'~

+ . t; = 8,

n . Jj~

n .Jtd- .

; = L . k .
i = 1..
k.
; = 1. ., k .

i*i

Here (5.8) and (5.9) are equivalent, respectively, to (5.4) and (5.5).
The relations (5.7}-(5.9) provide a geometric formulation of the problem of
simultaneous decoupling and complete control of the output vectors ..vi" . . " J'k'
Thus stated, the problem definition is both natural and intuitively transparent.
We observe that the output matrices Hi play no role beyond specification of
thesubspaces. f'jOl Since the Hi need have no special structure, the. "; are similarly
329

unrestricted. Nevertheless, we shall rule out trivialities by tacitly assuming:

(i) . ii #: I, i = 1, , k,
(ii] The subspaces ~ t are mutually independent. 1 In particular, the . i are
distinct and
0

(5.10)

i :F 0,

; = 1~ . , k.

(iii) The pair (A, B) is completely controllable, i.e., {AI~~} = if.

For if (i) fails, then for some i, . t i = I'; that is, Hi = 0 and Yi == O. If (ii] fails,
then for some i,
, t 't

n L .r OJ =F 0
j~i

or, by taking orthogonal complements,

.'i+

n.i'j~8
j~i

and (5.8) must fail. For (iii), if {AI~} = 8 1 #: 8 we can write I = 1'1

8 2 and

(2.1) as

Xl = A1x1

+ A 3 X 2 + Blu,

X2 = A 2 X 2 ,

"1'

where X;E"i, i = 1,2, and {Allbf!} =

The problem is unrealistic unless A 2
isstable(i.e., the pair (A, B)is stabilizable [4]). Hence, we may assume X2(t) == 0 and
take as starting point

The problem can then be reformulated with 8 1 in placeof I.

We tum now to the determination of necessary and sufficient conditions for
the existence of a solution to (5.7)--(5.9) in two special, but interesting, cases.
In the following sections, ~i denotes the maximal controllability subspace
such that
(5.11 )

:li en

..Jj~

= 1,, ~ k.

j*i

The ~i are constructed according to Theorem 4.3.

6. Deeoupllng when rank (H)
(6.1)

= n.

Our assumption is equivalent to

n" .i: = o.
i= I

That is, there is a one-to-one mapping of state variables into output variables.
THEOREM 6.1. If (6.1) holds, then the problem (5.7)-(5.9) has a solution {r and
only if

; = 1, ..... k .

(6.2)
1

Equivalently.. the row spaces or the Hi are mutually independent.

330

Proof. If the problem has a solution :JI" i = 1, .. , k, then by maximality

of the ~i' i = 1, ... , k, there follows:A i c .ii' and (6.2) follows from (5.8).
Conversely, suppose (6.2) holds. The jii are mutually independent; for.. by
(5.11) and (6.1),

s, n L~Il [n .'j] n [L n "f\:] en. "j n " i =

O.

Il~i

j~i

Il:l=i

V~1l

j~i

Let C, be chosen such that

~j;i

= {A + BCi/:A n ~d"

L,,. k .

Since the ~jii are independent there exists" by Lemma 3.1" a matrix C such that
= C;r (reti i , i = 1. ..... k).. i.e,..

Cr

(A

+ BC)r =

+ Bei)r .

(A

re.ji"

i = I . . k .

Then

J; = -fA +

BCI~

n JJ i } ..

i = L " .. k :

and C. together with the ~,' satisfy (5.7)-(5.9).

Remark 5. By Theorem 4.2, the C, can be chosen so that A + BC;, restricted
to ~i' has any desired spectrum. Hence, the same is true for A + BC. Furthermore.
there exists b, E &f ~i such that

~i

= :A + BCI{b'::'"

7. DecOlipUng when rank (8)

= k,

i =

L ., .. k .

Our assumption is equivalent to

dim:M = k,

(7.1)

Here the situation has been simplified by narrowing the choice of generating
subspaces fJI s; The same assumption was made in [1], [2] and [3]" with the
additional restriction that the outputs Yi be scalars.
THEOREM 7.1. //(7.1) holds, then the problem (5.7>-(5.9) has a solution ~r and
only ~r

(7.2)

ji

+".i = 8,

; = L ., ... k .

and
k

(7.3)

;;f

= L

n s;

i= 1

Furthermore, if C, Jt l '

. ,

:Jtk is any solution" then

(7.4)

Proof", Part 1. Suppose C,

= L . k.

~ l ' ... " .il k

is a solution. The necessity of (7,2)

follows, as in the proof of Theorem 6.1. To verify (7.3), write

~ n .II; = :iI; E9 [ !11 n .</1; n j~j .lIi ]

331

L. k .

The ~i are mutually independent; in fact,

HI,

L !/I

i_'

j C

:Mi

L ;M n ~j c

n (~ n .it;) n

.oA;

i-'

L~.i =

o.

jei

Recall that the:A i are (A + Be)-invariant. Then

= ~A + B('l:jfi~' + ji~

.11;

where

.Ai C{A +

Bcl L ilti}C L .Jt L n .1"~ c. t;.

!j

j C

*' i

j i Il

j ;

*j

Therefore. by (5.8),
{A

BCI~':'

+ . t ; = A,

and since . Ji #: 4 there follows

~i

"# 0, i = 1....... k. Therefore

L ~i = L dim ~i =
Ie

"

i=1

i= 1

dim

k:

so

(7.5)

.~

= ~A I ~

...

Ea :M"

"and
dim..M;

= 1..

l .... . k .

Since ~, c: fM n :iii C ~ n Jilt it follows that (7.3) is true.

Proof. Part 2. To verify (7.4),it is enough to show that the subspaces ~M
are independent. For then..
dim (fjI

n Ji

i)

n .iI;

; = 1....... k .

= 1..

and so

(7.6)

l.. ..... k .

Assuming (7.6) is true, let.ji = ~i G) ~i and choose Ci by Lemma 3. t . such that

(A

Then C, E C(aI,)

+ BCi).,ii C

'~i"

eir

= C,.,

re .if;,

n C(~,), so that
;it i

= {A + BCil~ n A'd"
= fA + BCil;~ n ~d'
=

.A;

which proves (7.4).

We proceed to show that the.-A

n ~ii are independent. Write

J;r = L s;
j;

332

1.. ... " k .

It is even true that

:An.lin~jr=O,

(7.7)

i=L.k.

On the contrary, suppose (7.7) fails for, say, ; = 1. If dim ( j

:J4n~l

(7.8)

If dim (at

n ~j 1) = L then

c.iT.

n ~1) ~ 2, and

(7.9)

;A

n ji . ~r,

i = J,

0"

k,

then

Jl = 1, . , k - 1, that is,

dim

[.f,= ~ n ';;i] ~
1

3;

and by induction
dim

[.r ~ n !.Ii] ~
,=1

I,

a contradiction. Thus (7.9) is false; combining this result with (7.8) there follows
(7.10)

&f

n d(l. c fl,:

for some cxe(l,, k). It will be shown below that there exists C2 such that
(7.11)

(A

+ BCac)9l(l

~2;

(A

+ BC(l~:

c ;~:.

Assuming (7.11) is true, we have

~(I = {A

+ BC,JBI

n :i(l.}

{A + BC21~:} c ~: c .. i';.

and therefore (7.2) fails for i = ~. With this contradiction, (7.7) is established.
It remains to verifythe existence of C2 For this we need the following result.
LEMMA 7.1. Let 'IJ"~ 11" be arbitrary. There exists C such that
0

(A

+ Be)l-"

j ~

(A

Be) 'II' c

'JI'

if and only if'

A1,'c9l+

1~

A11'cal+

'II~

A(1'

n ~"') c

:11

1"

n ~:

Proof. Necessity is obvious. For sufficiency, write

l'

'H"

= tAO ~ (1 . n ';f") ffi

333

1i~

where 1-' c 1 ~ ,*~' c 11 ~ By the construction of Lemma 3.2, C can be chosen such
that

n 'II ') c

(A

(A

+ BC){' c:

(A

BC)( t .

t'

1~

eci c

'N~

This completes the proof of the lemma.

~:. Clearly A;j2 c J8 +
Consider now
fA = ~ n :12 + aJ n ~:, and so

s;

A(:i:l

n ;1:) c

n 'H ~

s; A-l: c

~11

+ .j:.

By (7.3)

.12 ) n (~ + ,~:)
= 1~ + (;M + ~ 2) n .j:
= 111 + (~ n ~j: + j2) n ~:
= .eJI + ~2 n Ji:.
(:M +

By applying Lemma 7.1, the existence of Cf/. is finally established.

Proof. Part 3. We now prove that (7.2) and (7,3) are sufficient conditions for
existence of a solution. Let fi be the maximal subspace such that
1~'

(7,12)

en. tj,

i = 1.. ..... k.

j:Fi

It is enough to check that the f; are compatible, in the sense that there exists C
such that
(A

+ Be')1;' c

t~.

i=

L,~k.

;=

1.. .. _k .

We show first that the subspaces

L 1~'

1~'" =

j'i

are compatible. From (7.12) there follows

A

where ~i = 1M
that

n c: at + '-:1
= :JI

n ~. +

= ~i

'f~1

(by (7.3))

t~,

n t~. By Lemma 3.2.. there exist B, with {Bd' = Jli't and

('i"

such

i = L . k .

Choosing a basis {VI' ... , vll }- for

1~'

+ .,. +

t~',

we define C such that

BCv,. =

L ec;
;= I

334

\' = I . , .

}l.

Then

(A + BC)"f7 = (A + ec, + Ji BiCi) ",-,1'

+ BiC i )"f71 +

c (A

L B4j
i*i

(7.13)
c

'f-r

L 1j'
ji

This proves compatibility of the

1li. Now define

1'; = n 'f"j,

= 1..','. k .

= 1.. ..

't

k.

Clearly, 'I i :::> f;', i = I, ~ k. By (7.13),

(7.14)

(A

+ Be)'! i c i'i.

i = 1,, . k..

and, furthermore by the second condition of (7.12),

r,c

(7.15)

n L n
j~i

,t~

= n . I j.l

2:1:j m'2

j*i

By (7.14) and (7.15), the j'j satisfy the conditions imposed on the '1~' in (7.12).
Since the "fi are maximal, there results 'I i c j~, and, therefore, 1; = '1~' .
i = 1, "., k.
Remark 6. If the conditions of Theorem 7.1 are satisfied, then

(7.16)

.t

,-I

iii =

. {A + BC/f n

,= 1

= {A

;ji}

= {A + BCj. f n iii}
,=1

+ BClaI} = {Alai} = 8.

We turn now to the problem of pole assignment. In contrast to the situation

of 6, it is no longer possible, in general, to vary the spectrum of A + Be on each

:Ai independently. The following example shows that certain eigenvalues of

A + Be may even be fixed for all admissible C.
Let tf = 8 3 k = 2 and

A =

I 0 0]I
[001
1 I

= [:

~l

0]
Ii = [~ .
l.

2 This identity and its dual,

rule for subspaces.

L n L = L, are readilyestablished by using the (modular)distributive

335

It is easily checked that (S.7H5.9) have the (unique) solution

and that C must have the form

C

with arbitrary c 1 '

C2'

det (A

[('1 0 OJ

Then

+ Be -

A./) = (1

"2

A..HI - A.)( 1 +

('1 -

('2 -

A).

Observe that the eigenvalue A. = I, belonging to the eigenvector (0'1 1'10)' of A

BC",

is fixed,
To discuss the present case in general, we introduce a suitable decomposition
of 8. Assume that the problem of (5.7H5.9) has a solution C, .j l ' ... , .11k .. and
let C denote the class of matrices C for which (A + BC~~i C ~j;, ; = L .... k.
We know that the spaces :l; are the unique solutions: for simplicity of notation"
write Jt; for JI/. Define
k

(7.17)

n *r,

tfo =

i= 1

and let tl i be any subspace such that

.lJti=8i~(.'JtintfO)'

(7.18)

;= I.. ,k.

In the following, J denotes the set of indices (1, ... , k), J o the set (0, 1, ... , k).
In intersections and summations involving 91's, the index ranges over J: in those
involving 1'5, the index ranges over J 0
LEMMA 7.2. The subspaces tl i have the properties
(7.19)

4 o E9 8 1 ~

(7.20)

(A

~8k

=I .

+ BC)tIi c li+ 4 o ,

ieJ o .. CECa

Proof. Assertion (7.20) is obvious by the fact that the :iii are (A
invariant. For (7.19), observe first that
9l i

n tI0 = !JIi n n .~1 n fJtr = :Jli n ;Jt;

j:;';

and so, if i e J,

8i

n (8

L tS

j)

c:

s, n (8 0 + .<Jtj)

j*O
j:Fi

n ~r
= G n :Jt n ,CJtr

= A'i

(7.21)

= o.
Now for arbitrary subspaces ,Yi, i = 1, 2, 3, if
,y!

n ('~2 + ,Y3) =

.V1 n '</2
336

+ ,Y~ n .</3,

+ Be)-

then
Applying this fact and using (7.21) we have
81

Ie

n ( 8 0 + i~2 8) = 0 = 8 1 n A'o + 8 1 n j~2 8'j.

and therefore
")
Ion ( t!.+j~2'i
=/o n / . + / o n JIi. 28j
Ie

n L lSj.

= 10

j=2

Repetition of this argument yields, after k - 2 steps"

10 n (11 + jt2 Ii) = 1 0 n It = O.

(7.22)

Equations (7.21)" (7.22) state that the 8 i l E J 0' are independent. Finally" by (7.16) .
'l

Ie"

L Ii = L (8
i=O

+ 80)

i=1

L ~.;ti = I.
i=1

Remark 7. If the 91; are independent, then 8 0 = 0 and G; = Jt;9; E J.

For; E J o let Pi be the projection on 8; along
8 j and now let C E C be
fixed.
LEMMA 7.3. Let BI n '~i = {hi}' ieJ. Then

k;

tI; =

't

+ BC~{Pibd'}'
i e J,
Proof By (7.18) and (7.19), ti = Pial By (7.18) and (7.20),

(7.23)

{~(A

P/Jti

= Pi L (A + BCy-l{b i }
j= 1

= L [Pi(A + Bc)]j-l{Pibi:'
j= I

= {~(A + BC~{~bd'}'
LEMMA 7.4.

(7.24)

.4

n8

= O.

Proof. By (7.3) and (7.7),

~ n 40 = ( i

i= 1

Y4

n ::11;) n

= (iM n 91 1 +
= (

n.:Jt1

j= 1

i ~ n rJl

i)

;=2

i ~ n .11;) n niN!

i=2

j=2

337

n .':1It n

n,1
lt

j=2

This completes the proof of Lemma 7.4.

Next let C = C 1 be a fixed member ofC, let C2 E C, and write D = C 2 - C' l :
thus A + Be 2 = A + BC I + BD. Now bie;~i c ~';i + ~IO (jEJ): and (7.20)
yields BDtl i c Ii + 1 0 (ieJ o): therefore
(7.25)

i.j e J;

r ).

Also, using (7.24)

(7.26)

Write
k

(7.27)

L h.jc/j~

BD =

j= 1

where, as before, .~ hj } =
(7.28)

~!d

n ..;tj' Then

PiBDlj=Pibi,litrf'j=O,

ie i,

i#j,

We can now compute the spectrum 1\ of A +

Ai = ~(A

(7.29)

8('2'

B(' d,

i E J 0,

+ Be 2)(1

- Po)

jeJ o .

Define

By (7.26) and (7.29),

(7.30)

Po(A

+ Be 2) =

Po(A

AoPo,

J.

and by (7.28) and (7.29),

k

PitA

+ Be 2 ) =

Ai

+ P;BD

(7.31)

LP

j=O

= Ai

+ Pib;d;P

j,

Suppose A. E A.. with corresponding (complex) eigenvector ~. A brief calculation

from (7.30), (7.31) shows that either (i) for some i E J, Pi~ #- 0 and (Ai + Pibid~)Pi~
= APi~' or (ii) ~ = poe and Aoe = leo Conversely, if Ao~ = Ae for 0 -# ~ E &o - or
(Ai + p;bida~ = A.~ for 0 '# ~ E IJj and some i E J, then AE 1\. Therefore
k

1\ = U Ai'
i=O

where Ai" ; E J o .. is the spectrum of the restriction of Pj(A + Be 2) to & i- By (7.30)"

Ao is independent of the choice of C 2' i.e., is fixed uniquely by the requirement
C E C. On the other hand, for; E J Lemma 7.3 states that 8:j is the controllability
space of the pair (Ai' ~bi). Hence.. any choice of Ai can be realized by appropriate
choice of d, : indeed, for any W E 8 there exists d, such that
I

dx =
I

{W'X
0

for xeA'i'l
for x E L 8'j.
j*i

These results are summarized in the following theorem.

THEOREM 7.2. Let the conditions oj' Theorem 7.1 be satisfied. (r C E C., the
eigenvalues of A + Be can be partitioned into k + 1 disjoint sets

338

where

no =. dim

(.n ]11)'
J= 1

n, = dim (.ji) - dim (~i

n J#r)~

iE J.

The set A o and the integers ni (i E J 0) are fixed for all C E C. The sets Ai (i E J) can
be assigned freely (by suitable choice of C E C) subject only to the requirement
that any Aii with 1m Aij =1= 0 occur in Ai in a conjugate pair.
Remark 8. If basis vectors are chosen in the tS i , then the system differential
equation can be put in a simple "normal" form. Let

and
~x ~

(A

+ BC 2 )x + Br .

Multiplying through by 1'; and using (7.30), (7.31), we obtain

= (Ai + PibidaZi + PiBv.. i E J ,

Zo = Po(A + Be 2)(Z 1 + ... + z,,) + Aozo + PoBt:.
Let K be an m x m (= k x k) matrix such that BK = [b i .. bk ] and put r =
Zi

(7.32~

W == (WI'' WIe)'.

K \V~

Since hiES; Et> tRIo, we have

hi = Pb, + PObi ==

bi + b;o

Adopting nrdimensional representations of the z., etc... we see that (7.32) can be
written as
i, =

(Ai

(7.33)

+ '~)Zi + hiWi~

iE J ~

L AOj : j + Ao=o + Bolt\'.

to =

i'* 1

Equation (7.33) exhibits the system (2.1) as an array of k decoupled subsystems,

each completely controllable by an independent scalar input "';'1 plus one additional subsystem which is driven by the others and by w, Finally, since ;li n &'0
= .'Jti n
c .f', it follows by (5.8) and (7.18) that s, +.~; = ef~ that is, H;&;

= .If;.

atr

Remark 9. The decoupled system is acceptable in practice only if the eigenvalues in the fixed set Ao are all stable. It is possible to check for stability of Ao
as follows. Recall that Jt i C tl i + tf 0 (i E J) and note from (7.20) that A(&;i + a0)
c Ii + tf0 + YI (i E J). Furthermore,
~

+ 80

en. 1j,

i E J.

j=/;i

It follows by Theorem 4.3 and the maximality of the .Jt; (='~i) that
:Jt; = {A

+ BClat n (I'i + 1 0 ) :,

for any C with the property (7.20). That is, (7.20) is both necessary and sufficient
that C e C. Thus, to compute Ao it is necessary only to compute the spectrum of
A + Be0 (restricted to tf 0) where Co is any matrix such that (A + Be0 )($1 0 C ~ ()
339

Concluding remark. This article represents a preliminary investigation of

the generaldecoupling problem formulated in 5.The results for the specialcases
of 6 and 7 suggest the possibility of a complete and detailed geometric theory
of linear multivariable control, in which the concept of controllability subspace
would playa central role. Specific problemsfor future study includenot only that
of 5 but also the problem of decoupling by adjunction of suitable dynamics
(augmentation of the state space), and the problem of sensitivity. As formulated,
decoupling represents a "hard" constraint, an all-or-nothing algebraic property.
Of course, for applications a quantitative approach via "soft" constraints might
also prove rewarding.
It is clear that an adequate qualitative theory of large linear multivariable
systems is currently lacking;and equallyclear that, with computers,such a theory
would find wide application.
REFERENCES
[1] Z. V. REKASIUS. Decouplinx of multivariabte systems by means of state tariable feedback. Proc.
Third AllertonConference on Circuitand SystemTheory, Urbana, Illinois,1965, pp. 439-447.
[2] P. L. FALS AND W. A. WOLOVICH, Decoupling in the design and synthesis of multivariable control
systems, IEEE Trans. Automatic Control, AC-12(1967), pp. 651-659.
[3] E. G. GILBERT, The decoupling of multivariable systemsby statefeedback, this Journal. 7 (1969),
pp.50-63.
[4] W. M. WONHAM, On poleassignment in multi-input controllable linear systems, IEEE Trans. Automatic Control, AC-12 (1967), pp. 660-665.
[5] M. HEYMANN, Pole assignment in multi-input linear systems, lbid., AC-13 (1968). pp. 748-749.

340