660 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-12, NO.

6, DECEMBER 1967

On Pole Assignment inMulti-InputControllable

Linear Systems

A&stracf-It is shown that controllability of an open-loop system THEOREM

is equivalent to the possibility of assigning an arbitrary set of poles
to the transfermatrix of the closed-loop system, formed by means of The pair ( A , B ) is controllable i f an.d only <f,f o r every
suitable linear feedback of the state. As an application of this result, choice of the set d , there is a matrix C m c h that A+BC
it isshown that an open-loop system can be stabilized by linear feed- has A f o r its set of eigenvalues.
back if and only if the unstable modes of its system matrix are con-
trollable. A dual of this criterion is shown to be equivalent to the
In other words, controllability is equivalent to the
existence of an observer of Luenberger's type for asymptotic state property that the closed-loop transfer matrix
identification.
INTRODUCTIOX [SI - (A +X)]-'B
Consider the system
can be assigned an arbitrary set of poles b>- a suitable
dx(t)/dt = A r ( t ) + Bu(t). (1)
choice of the feedback "gain" matrix C.
Here and in the following, all vectors and matrices have If B is an n vector (m= l), the result stated is well
real-valued elements, and all matrices are constants. In known and is obviousafter a change of basislvhich
(l),A and B are matrices of dimension n X n and n X m , transforms d tocompanion(rationalcanonical)form
respectively; x is an n vector denoting the state and u and B to the form
is an m vector which,as usual, denotesan external input. B = (0, 0, . . . , 0, 1)'
Consider the closed-loop system defined by
u=cI+P wheretheprimedenotestranspose.Suchachoice of
basis is always possible if ( A , B ) is controllable; for the
for some m X n matrix C and a new external input 21. details see, for instance, Lefschetz,['l p. 113.
Then (1) becomes To prove the result in the general case, the (pair A ,B )
dx(t)/dt = ( A +
BC)x(l) Be(t).+ (2) is first transformed to a canonical form in which
multi-input system is exhibited as a triangular array,
the
In applications it is often desirable to choose C so t h a t of nhich the diagonal blocks are systems of the type
the matrix A +BC has special properties, for example, just described. This transformation was used by Lan-
stability. Intuitively i t is clear that the possibility of genhopr21and laterin a dual contextby Bass andGural61
such a choice depends on the controllability, in an ap- and Luenberger.1'1 Langenhop proved a theorem similar
propriate sense, of the state x with respect to '11. This to that stated above,in the case where ti = Fx +Gzc and
paper is concerned with the property of "poleassign- where the elements of the parameter matrices may be
ability," w-hich is shown to be equivalent to controllabil- arbitrary complex numbers. The present restriction to
ity of (1) in the usual sense. matrices with real-valued elements apparently prevents
T o be precise, consider the follou-ing. Let the immediate application of Langenhop's result; but
A = { X l J . . . ,x,) this difficulty is overcome by use of preliminary results.
be an arbitrary set of n complex numbers X i , such that
any X i with I m X i # O appears in A in a conjugate pair. PROOF
OF THEOREM
Alsorecall that the pair ( A , B ) is(completely)con- A . Sufi.ciency
trollable if and only if the n X m n matrix Theargument of Langenhopt?] applies
without
K = [B, A B , . . , dn-'B] change. In fact, letXI, . . , X, be any distinctreal num-
e

berssuchthatdet(A - A i l ) #O, i = 1, , a. By a ~ -
is of full rank n. Then the main result to be proved is sumption there are n vectors . r i # O , i = l , * * * , n, and
the follon-ing. an m X n matrix C such that
Manuscript received February 17, 1967; revised August 10, 1967.
This research was supportedin part by the Office of Aerospace
Research, US Air Force, under AFOSR Grant AF-.lFOSR-693-66,
(-4 + BC - XJ)Zi = 0, i = 1, ... J ''
in part by NSF under Grant GK-967, and in part by NAS.4 under or
Grant NGR-40-002-015.
The author is with the Div. of Applied Mathematics, Center for
Dynamical Systems, Brown University, Providence, R. I., presently - ( A - Xil)-1BCxi = .2i, i = 1, . . . ,n . (3)
on leave with NASA Electronics Research Center, Cambridge, hlass.
WONHAM:
ASSIGNMENT
POLE IN CONTROLLABLE
SYSTEMS 66 1

Since then P(A)x=O, and so a10 (a divides p). Since a is the

n mp of E , 01a. Hence P=a, and so b generates E. T h e
- ( A - XI)-l = Pj(X)Ai-' proof is complete.
j=l Let B be an n X m matrix as before, and let { B ]
denote the subspace of E spanned by the column vec-
for suitable rational functions pi, ( 3 ) implies
tors of B. Obviously if { B ] contains a generator of E ,
n then ( A , B ) is controllable. Since the mp of E may be
pj(Xi)dj-'BCai = Xi, i = 1, * * . ,n. (4) of degreelessthan n , theconversestatementisin
j= 1
general false. However, for a cyclic space the following
Sincethe X i aredistinct,theeigenvectors xi of is true.
A +BC are linearly independent. Thus (4)states that LEMMA 3 : Let ( A , B ) be controllable and let E be
the range of K is the whole space, that is, K has rank n. cyclic. There exists an n vector bE { B ] such that ( A , b )
i s controllable.
B. Necessity Proof: For i = 1, . -
, m let P i be the mp of b;, and
T h e proof of necessity leans on the theory of cyclic E i the cyclicsubspacegeneratedby bi. Thus E = El
subspaces(Gantmacher,r31ch. 7). Forcompleteness, + - +E, and, if a is the mp of E ,
preliminary results are collected below. a = LCM(&, * * - ,Dm).
1) PreliminaryResults: Denotethecoordinate n
spaceby E andletthe n X n matrix A be fixed. If Let c be a generator of E. Therearepolynomials
b E E and b#O, let v(b) be thegreatestintegersuch 71, . . . , y m such that bi=y,(A)c, i = l , . . . , m. T h e
that the vectors existence of realnumbers rl, . , rm will beproved,
such that rlbl+ * * * +rmbm generates E . By Lemma 2,
6, A6, . . . , A'(')-'b i t is enough to choose the ri so t h a t
are linearly independent, and let Eb denote their span.
Ea is the cyclicsubspace generated byb. E is cyclic if
Y(X) = rlyl(X) * * rmrm(X) + +
there exists b E E such that Eb = E , that is, such that is coprime with CY.
( A , b ) is controllable. T h e minimal polynomial (mp) of For this, note first that
b is the (unique) monic polynomialP@) of lowest degree
such that P ( A )b = 0. Thus b generates E if and only if GCD f 4 4 , r1(Q,* ' ,rm<A>) = 1. ( 3
the mp of b is of degree n. T h e m p of E is, as usual, the In fact, if K is the GCD on the left, then
(unique)monicpolynomial a @ ) of leastdegreesuch
t h a t a ( A )= 0 ; a @ )is the least common multiple (LCM) ff = K4

of the mp fib) ofthevectors b E E . Evidently E is y i = &, i = 1, . . . 2 fa

cyclic only if the degree of a@) is n. T h e converse may
be stated as follows. for suitable polynomials di and Ti. Then
LEMMA 1 (Gantmacher,[31p . 180, Theorem 2 ) : If the
di(A)bi == B(A)K(A)Ti(A)C
mp a @ ) of E i s of degree n, then E i s c y c l i ~that
, is, there
exists b E E such that the mp of b i s a @ ) . = -ji(A)a(A)c
In particular, the condition of Lemma 1 holds if the =o, i = 1,. . . m.
7
eigenvalues of A are distinct.
Twomonicpolynomials @,$ are coprime if their Thus Pildi, i = l , . . . , m ; hence,bydefinition of the
greatestcommondivisor (GCD) is 1. Then (Birkhoff cy1
LCM, 8,t h a t is, K = 1.
andiLIa~Lane,[~l p. 75, Theorem 12) therearepoly- Finally, observe that CY and y are coprime if and only
nomials p and u such that if r@j) ZO, j = 1, . . . n, where the X j are the (complex)
P ( m 0 ) + .(W(V 1.
zeros of a@).By (5), the quantities yi@,), i = l , -- ,
m, cannot all vanish for any Xi. Therefore, numbers ri
LEMMA 2 (Birkhofand MacLane,141 p . 324, The- with the required property exist.
orem 19, Coroll~ryy): Let E be cyclic, with mp a , and let c The next observation will be useful; the simple proof
be a generator of E . If b = y ( A ) c for some polynonzial y , is omitted.
and i f a and y are coprime, then b is also a generator of E. LEMMA 4: If ( A , B ) i s controllableand C isan
T h e converse is also true, butwill not be needed. m X n matrix, then ( A+BC, B ) i s controllable.
Proof: Let /3 be the mp of b. Since a and y are Also of use is the concept of congruence; for details
coprime, there exist p and u such that 1 =yp+cru. Thus see Gantmacher,r31 ch. 7.3. Let E1 be an invariant sub-
space of E , t h a t is, A x E E 1 whenever x E E 1 .4
, subspace
c = y(A)p(A)c + cu(A)a(A)c = p ( A ) b . El is invariant (mod El) if x€E2 implies A x = y l + y S ,
If x E E is arbitrary, then for some polynomial 0, where yl EEl and y~EEP.Write x =0 (mod El) if x E El.
Vectors X I , -- , xk are linearlyindependent (mod E l )
x = B(A)c = B(A)p(A)b; if the relation rlxl+ - +rkxk=O (mod El) implies
662 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, DECEMBER 1967

r l = . . . = r k =o, for all sets of scalars 11, . , Y k . For -- where the summation does not appear in case k = v i ;
some b e E 1 let v be the greatest integer such that the t h a t is,
vectors eivi = b,, i = 1, - - 9 t. (8b)
b, Ab, * * , Ap'b
I t is clear t h a t for each i the vectors eit, k = 1, . , vi,
are linearly independent (mod E l ) , and let E? be their areindependentlinearcombinations of thevectors
span. Then E2 is cyclic (mod E l ) ; clearly E2 is invariant Aj-lbi, j = 1, . . , vi, and thus the entire set of eil, is a
+

(mod El). T h e relative mp of E2 (mod El) is the (unique) basis for E. Observe t h a t for 2 5 k 4v;
monic polynomial 7 of least degree such t h a t q ( A ) x=0
(mod El) for all xEE2. A e i k = ei,k-l + aaeiri, i = 1, . . ,d . (Sa)

2 ) Proof of Necessity: I t isnowpossible toprove Furthermore, by use of the fact that the relative mp
necessity. Let bi denote the ith column of B : of b , (mod &+ . +&-I)
+ is of degree vi, there follows
B = [bl, . . , b,].
*
(see (7))
i-1
+
Vk
Let E; be the cyclic subspace generated by b;, and put Aeil = w e + yijkekj, i = 1, . - ,t (9b)
ni= dim(&). Since ( A , B ) is controllable, k-1 j=1

E=El+-.-+Em for certain scalars- y i j k , and where the double summation

b u t in general the Ei are not independent. Holvever, E does not appear if i = 1.
can be written as a direct sum Next, using (9), compute the form of A , regarded as
a lineartransformationin E , relative to the basis
E=i?1G?Ez@ - . * @ E t , t 6 ~ (6) ell, - , elv,, e21, . * , eZv2, . . , e,l, . . . , e,,,. T h a t is,
wherethearecertainsubspaces of the Ej. To see let H be the matrix with column vectorseik in the order
this,assume bl#O anddefine 21 =El. If bzeE1 let just written, and let T=H-'. Then T A P 1 = A has the
vg, 1 $vesn2,be the greatest integer such that the vec- block form
tors r A1 A12 .. A l t 1
bl, Abl, . . * , A"'-'bl, ba, Ab2, . . . , A'F'ba
I I
arelinearlyindependent,andlet be thespan of the
-
vectors Aj-lbp, j = 1, . , v2.T h u s & is cyclic (mod El)
and has relative mpof degree u2. Continuing in this nay,
for i>=2 define vi, 1 Pvi Sni, to be the greatest integer In (10) the matrices Ai which occur on the diagonal of
such that the vectors d are v i x v ; companion matrices
bl, . . . , Api-l-lbi-l, bi, Abi, . . A*i-lbi .
0 1 0 0 * e -

arelinearlyindependent,andlet & bethespan of

A j-1bi, j = 1, -- , yi. Then Bi is cyclic (mod .&@ . . .
@ Bipl),with relative mp of degree vi. If at any stage (1 1)
biEkl@ -e,&, i% 2, then b; isskipped.Byre- . . .
orderingthecolumns of B , if necessary,itcanbe ...
-
arranged that for i= 1, . . , t the vectors bi generate
ai1 a;2

independent subspaces &, whereeither t = m or and the matricesi3;j have dimensionv i X v j .

b , E & @ . . . @E,, s = t + l , . . . , m. Since Aj-Ib;E& Finally, T B = 3, where
8 . e&, if j 2 v i + l , 1Sisnz, i t follows t h a t (6) is
true and thus V I + . . . + v t =n.
Next A is transformed toa convenient standard form. Here
Let
vi
pi - aij~+l, i = 1, .'97 ('I)
j= 1

be the relative mp of bi (mod &+ . . . +&-I) or the

absolute mp of bl in case i = 1. Following LangenhopL21 vi)' = n - (VI + * + Vi),

introduce vectors
and Bo= T[bt+l, * , b,] is an nX (vz - t ) matrix, which
P i does not appear if t = m.
eik = AYsLx-b, - aijA*S1bj, K = 1, 2 , - , vi To verify (lo), (ll),and (12) in detail, observe that
j&+1
i = I, 2 , - . - ,1 @a) Tx is the column of components of x in the basis { e i k ] .
BLE
IN ASSIGNMENT
WONHAM: POLE SYSTEMS 663

Thus the 9th column of Ax

is a list of the components, matrices Pi@) are of fixed dimension. If, for example,
in this basis, of d e i p , where eik is the pth basis vector. t = 2 and P1, P2 are each of dimension 1 X 1, i t is impos-
This listcanberead off from (9). Similarly 6; is the sible to assign an arbitrary complex pair of eigenvalues
column of components of bi, so t h a t (12) follows a t once by independent ad~ustmentof P I , p2. This is the added
from (8b). difficulty arising fromthe restriction toreal parameters.
137th A , B in canonical form, i t will be shown finally T o proceed, let C= Cl+C2, where C1 is of the form (13)
that the eigenvalues of A + B C can be prescribed arbi- and is subjectonlytotherequirementthat all the
trarily.Put c= CT-l andobservethatthematrix eigenvalues of A+BC1bedistinct. Ithasjust been
J + B C is similar under T to -4 + B C . From now on, the seen thatsuch a choice of C1 ispossible. Then,by
tilde on 6 , 3, c will be dropped. Next, take C in the Lemma 1, the space E is cyclic relative to the matrix
form A +BCl. By Lemma 4, (A+BCl, B ) is controllable.
Applying Lemma 3, one can find an 72 vector bE { B 1
such that ( A+BC1, b ) is controllable. S o w b = Bg for a
suitable m vector g. Thus if u is taken in the form
u = ClX + gz + w
L --%---J \x-ith z a scalar and w an r~ vector, then the system (1)
where becomes

Finally, because the theorem is true in the single-input

In C the upper block is of dimension t X n and the lower case, can be chosen in the form
block of dimension (m-t) Xn. From (10) through (13) i t
is seen t h a t z = cfx

1 such that the matrix A +BCl+bc’ has the desired set

of eigenvalues A . T h a t is, the “gain” matrix
c = c1 + gcf
l o -P O ) 1 has
the
required
property.
Thetheorem is proved.

1
where, for i = 1, .
Pi@) = A i 4- Q i G i f - XI 1) The construction in the proof of necessity can be
--x 1 0 . . . . . . . . . * o summarized in terms of a block diagram (Fig. 1). After
o . . . . . . . choosing a suitable set of state variables (the canonical
1
. . . . . . . . . . . . . . :o form (8)-(12)) the designer constructs an inner feedback
loop (through the Cl matrix) which renders the system
cyclic, t h a t is, controllable by a single input. T h e de-
()....... 0 --x 1 signer then picks a suitable input in the form gz (Bg=b
is a generator of E ) and completes the outer loop bv
ail + ~ i l , + ci?, . . , aivi
* + -X
~ i v , setting z = G ’ X toa~hievetherequireddisposition of
poles.
Let $,(A) be the determinant of Pi@),t h a t is,

(-1)y.(\) t / = hl’i - (cyij f 6jj)Aj-l. (19

j=1

orm
angular
the From of A +BC-XI i t follows t h a t w D X

+ BC - XI) = p l ( x ) . . . p t ( x ) .
det (-4
Also, i t is clear from (14) that the zeros of p i @ ) can be
I
assignedarbitrarily(subject to conjugacy of complex
zeros) by proper choice of the real coefficients cij. This
shows that, in particular, the system (1) can always be
stabilized by appropriate choice of C.
T o show that an arbitrary set of eigenvalues A can
Fig. 1. Construction in proof after choice of
be assigned, account must be taken of the fact that the canonical state variables.
664 AUTOMATIC
TRANSACTIONS
ON IEEE CONTROL, DECEMBER 1967

2) In practice there may be many ways of choosing where A+= A P + is the restriction of A to E+. Similarly,
state variables and the quantities CI,g, c to achieve a
given assignment of poles. T o exploit this freedom via i = A-z +PBu
suitable criteria of design is an interesting problem of and here A- is stable. Now,1 if { K )>E+ then
current research (see, e.g., Kalmanr5J).
3) It is worth noting that more than one [‘canonical” E+ = P{Rf
form (10)-(12) may exist for a given ( A , B ) pair. T h a t
is, the matrix A of (10) may not reveal the intrinsic = { [ P B ,P A B , . . , P i l * l B ] )
structure of A , as do the usual rational canonical de- = { [ P B , A + P B , . . . , (A+)*lP+B]j.
compositions into block-diagonal forms (Gantmacher,r31
ch. 7). The drawbackof such block-diagonal representa- Thusthepair(A+, P B ) is controllable in E+, and
tions here is that, in general, t h e cyclic subspaces cor- there exists a matrix C+ suchthat A++P+BC+ is
responding to individual blocks need not have genera- stable. Let C= C+P+ and put u = Cx. Then
tors in the subspace { B ]. If they do, the corresponding
controls are completely noninteracting.
+ P+BC+)y
jt = ( A +
i = P B C + y + -4-2. (17)
TO STABILIZATION
APPLICATIOH Equation (17) exhibits A +BC in triangular form with
Referring to (2), i t is natural to say that the pair both diagonal blocks stable, that is, A+BC is stable.
( A , 3) is stabilizable if there exists an m X n matrix C ~ o n v e r s e ~suppose
y 1
{ K DE+. Since { K ] is in-
such that the matrix A +BC is stable, that is, all its variant and a+,a- are coprime,
eigenvalues havenegative real parts.Obviously, if
( A , B ) is controllable then ( A , B ) is stabilizable. I t mill
be shown that stabilizability is equivalent to the weaker Thus the subspaceP+ { K ) is properly contained in E+;
property that at least the “unstablemodes” of A be con- in other words, the pair (A+, P+B) is not controllable.
trollable. In E+ let Q denotetheorthogonalprojection on the
T o be precise, let the mp ab)of E , with respect to A , A+-invariant subspace P+ { K 1.
Then QA+ = A+Q,
be factored as I - Q# 0, and ( I - Q)P+B= 0. Writing
&(X) = (Y+(X)a-(X) YI = QYl YZ= (I - Q)Y
where all zeros of a+@) lie in the closed right half com- it follows from (16) t h a t
plex plane and allzerosof a-@) lie in the openleft
half-plane. Define the subspace of unstable modes 32 = (I - Q)Y
E+ = { x : R: E E , &+(A)R:= oj = (1- Q>A+Orl +YJ
and let K denote the controllability matrix (see Intro-
= (I - Q) A+y2.
duction). Since the eigenvalues of ( I - Q ) A + are included among
those of A+, i t is clear that the subsystem with state
PROPOSITIOK : .The pair (-4, 3 ) i s stabilizable if and vector yz cannot be stabilized by any choice of control
only if the range of K contains E+. u. T h e proof is complete.
Proof: Define
0 = ( x : x E E, a-(A)R: = 01.
APPLICATIOK
TO STATE IDENTIFICATION
Since a+,CY- are coprime it follows (Gantmacher,L31 p. Consider (l), defined for t 2 0 , togetherwiththe
179, Theorem 1) t h a t E = E+$E-. Let P+(P-) denote auxiliary equation
theprojectionon E+(E-) along E-(E+), andwrite
y = P+x, z = P-x. From (1) r(t) = Rx(t), t 20 (18)

j = +
=P ( A x B24) where R is a p X n matrix and r represents an observed
vector. Define the n-vector functions(t) by an equation
=P A ( y + +PBu 2) (15) of the form
-
where = d/dt. s(1) = &(I) + Tr(t) + Bu(t), I 2 0
Since E* are A-invariant subspaces,
s(0) = 0
P A = AI’+, P A P = O .
where S and T will be determined below.
T h u s (15) implies
j = A+y +PBu, (16) 1 The notation [ . 1 denotes range of a matrix.
WONHAM: POLE ASSIGNMENT IN CONTROLLABLE SYSTEMS 665

T h e function s ( t ) can serve as an asymptotic estimate REFERENCES

of x ( t ) in the followingsense. If thepair (A’, R’)is [’I S. Lefschetz, Stability of Nonlinear Control Systems. New
stabilizable, then the matrices S and T can be chosen York: Academic Press, 1965.
P I C. E. Langenhop, “On the stabilization of linear systems,”
so t h a t Proc. Amer. Math. Sac., vol. 15, no. 5, pp. 735-742,1964.
L31 F. R.Gantmacher, The Theory of illetrices, vol. 1. New York:
s ( t ) - z(t) -+ 0, t -+ co Chelsea, 1959.
L41 G. BukhoE and S. MacLane, A Suwey of Nodern Algebra.
forallinitialvalues x ( 0 ) . Indeed,choose T so t h a t New York: Macmillan, 1953.
L5I R. E. Kalman, “LVhenis a linear controlsystem optimal?,”
A‘ -R‘T’ is stable and set S = A - TR. Then Trans. A S N E , J. Basic Engvg., ser. D, vol. 86, pp. 51-66, 1964.
[SI R. 1%’. Bass and I. Gura, “High order system design via state-
spaceconsiderations,”Preprints, Joint Automatic Control Conf.
(d/dt)(s - x ) = S(s - x) (Troy, N. Y., June 1965), pp. 311-318.
Lil D. G. Luenberger, “Observers for multivariable systems,”
andtheassertion follows. Thismeansthat if only IEEE Trans. Automatic Control, vol. +C-?l, pp. 190-197, -3pril 1966.
asymptotic identification is required,
then
stabiliz- J . D. Simon, private communlcatlon.
abilitycanreplacethestrongercondition of observ-
ability. Estimators of the type of s(t) havebeendis-
cussed by Luenberger,[’] and this application is due to W. M. Wonham (R.1’61) was born
Simon. Is] in Alontreal,Quebec,Canada,
on November 1, 1934. He re-
COIKLVSIOKS ceived theB.E.degree in engi-
T h e main result of this paper is a proof of the equiva- neering
physics
from
McGill
lence of controllability and pole assignability in finite University,Montreal,in1956,
dimensional dynamic systems with real-valued param- and the Ph.D. degree in control
eters. The two properties mentioned are thus exhibited engineering from the University
as dualconceptsintheirrespectivedomains of time of Cambridge, England, in 1961.
andfrequency.The proof givenislongwinded,but He
has
done
researchand
includes an interesting auxiliary result (Lemma 3) and teachingincontrol atPurdueUniversity,Lafayette,
yields information on canonical structure. The weaker Ind., the Research Institute for Advanced Studies, and
property of stabilizability is characterized as controlla- Brown University, Providence, R. I., where he is cur-
bility of unstablemodes. I t wouldbeinterestingto rently Associate Professor of Xpplied Mathematics and
algorithmizethe pole shiftingconstruction,possibly Engineering. I n addition, he is a Consultant to XASA
along the linesof Bass and Gura,r61 and to find a neater, andan AssociateEditor of the S I A M J o u r m l on
abstract proof of the main theorem. Control.