Você está na página 1de 7

596

IEEE TRANSACTIONS ON MJTOMATIC

CONTROL, VOL.

ac-16,

NO.

6, DECEDER 1971

An Introduction to Observers
DAVID G . LUENBERGER,

SENOR MEMBER,

IEEE

Abstract-Observerswhichapproximately
reconstruct missing
state-variable information necessary for control are presented in an
introductory manner. The special topics of the identity observer, a
reduced-order observer, linear functional observers, stability properties, and dual observers are
discussed.

ha.s as its inputs the inputs and available outputs of the


system whose state is to be approxima.ted and has a state
vector t,hat is linearly related to the desired approximation. The observer is a dynamic syst,em whose characteristics are somewhat free to be determined by the designer, and it is through its introduction that dyna.mics
I. INTRODUCTION
ent.er the overall two-phase design procedure when the
T IS OFTEN convenient when designing feedback entire state is not available.
controI systems to assumeinitially that t.he entire
The observer wa.s first proposed and developed in [ l ]
state vector of the system to be controlled isavailable
andfurther developed in [2]. Since theseearlypapers,
through measurement. Thus for the linear time-invariant svhich concentrat,ed on observers for purely deterministic
system governed by
continuous-time lineartime-invariant system, observer
i ( t ) = Ax(t)
Bu(t)
(1.1) theory has been ext,ended by several researchers to include
time-va.rying systems,discretesystems,
andstochastic
where X is an n X 1 state vector, u is an I X 1 input systems [3]-[18]. The effect of an observeronsystem
vector, A is an n X n system matrix, and B is an n X T performance (with respect to a quadratic cost functional)
distribution matrix, one might, design a feedback law of has been examined [ 5 ] , [19]-[22]. New insight,s have been
the form u(t) = t ( x ( t ) , t ) which could be implemented if obtained, and t.he theory has been substantially strea.mx ( t ) mere availa.ble. This is, for example, precisely the lined [23]-[25]. At. t.he same t.ime, since 1964, observers
form of control lam that resu1t.sfrom solution of a quadratic have formed an integral part of numerous control system
loss optimization problem posed for the syst,em (l), designs of which a small percentage have been explicitly
from design techniques that place poles at. prespecified report,ed [26]-[31]. The simplicity of its design and its
point,s, and from numerous other techniques t,hat, insure
resolution of the difficulty imposed by missing measurestabilit,y and insome sense improve system performance.
mentsmake the observer an attract.ivegeneral design
If the entirestate vectorcannot be measured, as is component [ F A ] , [32], [33].
typical in most complex systems, the control law deduced
In addition to their practical utilit,y, observers offer a
in t,he form u(t) = dr(x(t), t) cannot be implemented. uniquetheoreticalfascination.
The associated theoryis
Thus either
a
nen- approachthat,
direct.12: accounts intimatelyrelated
t.0 the fundamentallinearsystem
for the nonavailabi1it.y of the entire state vector must be concepts of cont,rollability, observability,dynamic
redevised, or a suitable approximation to the state vector sponse, andst.ability, and provides a simple set.ting in
must be determined t,hat can be substitut,ed into t>hecon- which all of these concepts interact. This t,heoretical
trol law. I n almost every situation the latt.er
approach, richness has made t.he observer an at,tractive area of rethat of developing and using an approximate state vector, search.
is vastly simpler than a new direct att,ack on the design
This paper discusses the basic elements of observer
problem.
design from an elementaryvienToint.
For simplicity
Adopting this point of view, t,hat an approximate state attention isrestricted, as in the earlypapers, to detervector will be substitutedforthe
unavailable state, ministic continuous-time linear time-invariant systems.
result,s in the decomposit,ion of a control design problem The approa.ch t.aken in this paper, hon-ever, is influenced
into t x o phases. The first phase is design of the control substa,ntially by the simplification and insights derived
law assuming that the statevector is available. This may from the work of several other a.ut.hors during the past
be based on opt,imization or other design techniques and seven years. In order t.o highlight t,he new t.echniques and
typically resu1t.s in a control law without dynamics. The to provide the opport.unit,y for comparison with the
second phase is t.he design of a syst.em that produces an old, many of the example syst,ems presented in t,his paper
approximation to the state vector. This syst.em, n-hich are the same as in theearlier papers.
in a deterministic setting is called an observer, or Luen11. BASIC
THEORS
berger observer to distinguish it from t,he Iialman filber,

Manuscript received July 21, 1971. Paper reconunended by


R. W. Brockett,, Associate Guest Editor. Thisresearch was supported
in part by theNational
Science Foundation under GrantGK
16125.
Theauthor is withthe Depart,ment. of Engineering-Economic
Systems, School of Engineering, Stanford University, Stanford,
Calif. 9430.5, current.ly on leaveat Office of Science and Technology,
Execut,ive Office of the President,, Washington, D.C.

A . Almost any System is an Observer


Initially, consider t,he problem of observing afree
system SI,i.e., a syst,em with zero input. If t,he available
output,s of this system are used as inputs to drive anot,her
system Sz, t.he second system will almostalwaysserve
as an observer of the first system In that. its state will

LUENBERGER: AN INTRODUCTION TO

OBSERSERS

597

and Sz,the observer, is governed by

i(t)
Fig. 1. A simple observer.

tend to t.ra.ck a linear tra.nsformat.ion of the state of t,he


first, system (see Fig. I ) . This result forms t,he basis of
observer theory and explains Ivhy there is a great. deal of
freedom
the in
design of an observer.

z(t)

Tx(t)

+ eFf[z(O) - T x ( O ) ](.2 . 1 )

Proof: We may write immediately

Substituting TA - FT

i(t) - Ti(t)

H t,his becomes
=

F[z(t) - T x ( t ) ]

which has (2.1) as a solution.

It should be notjed t,hat t.he two systems SI and Sp


need not, have the mnle dimension. Also, it can be shown
[ l ] that there is aunique solution T t,o theequation
TA - FT = H if A and F have no common eigenvalues.
Thus a.ny syst>emSzhaving different, eigenvalues from A
is an observer for SIin thesense of Theorem 1 .
Next, we n0t.e that the result of Theorem 1 for free
syst.ems can be easily extended to forced systems by
including the input. in t,he observer as well as the original
system. Thus if SI is governed by

Fz(t)

+ Gg(t)

(2.5)

then H = GC. In designing the observer the m X n


matrix C is fixed and t.he n X m matrix G is arbitrary.
Thus an ident,itv observer is determineduniquelyby
selection of G a,nd takes t,he form

i(t)

( A - GC)z(t)

+ Gg(t).
(2.6)

havearbit.rarydynamics
if the original systemis completely observable. First, recall that a syst.em (2.4)
. , is
completely observable if the matrix

has rank n. Generally, if an n X n mat.rix A andan m X n


matrix C satisfyt,his condition we shall say (C,A ) is
completely observable.

Lenanza 1: Corresponding to t,he real matrices C and A ,


then the set of eigenvalues of A - GC can be made to
correspond to the set of eigenvalues of any n X n real
matrixbysuitable
choice of the real mat,rix C if and
only if ( C , A )is completely observable.

Thislemma, which is now a cornerstone of linear


syst,em theory, was developed in severa.1 st,epsovera
period of nearlya decade. For the case 112 = 1 , corresponding to single out,put systems, early statements can
be found in Kalman [34] and Luenberger [ l ] , [35]. The
general resultis implicit.ly contained in Luenberger [ 2 ] ,
X(t) = Ax(t) h ( t )
(2.2) [ 3 6 ] , and the problem is treabed definitively in Wonham
[ 3 7 ] . A nice proof is given byGopinath [%I. (It was
a system Sp governed by
recent,ly pointed out to me t,hat, Popov [38] published a
proof of a result of this type in 1964.) Calcu1at)ionof the
i ( t ) = Fz(t) H x ( ~ ) TBu(t)
(2.3)
appr0priat.e G ma.t,rix t.0 achieve given eigenvalue place\\-ill satisfy (2.1).Therefore, an observer for a. system can ment for a high-dimensional multivariablesystemcan,
be designed by first assuming the system is free and then however, be a difficult cornputrational chore.
incorporating the inputs asin (2.3).
The result. of this basic lemma translat,es directly into a
result on observers.
B . Identity Observer
Th.eorem 2: An identity observer having arbitrary
An obviously convenient observer would be one in
dynamics can be designed for alinear
t.ime-invariant
which the tra.nsformation T relating t.he st.ate of the
system if and only if the syet,emis completely observable.
observer to the st,ateof the original system is t,he identity
t.ransformat.ion. This requires t.hat t.he observer S, be of
I n practice, the eigenvalues of tlhe observer a.re selected
the same dynamic order asthe original syst.em S1 a,nd t.hat to be negative, so t.hat t.he state of the observer will
(with T = I ) F = A - H . Specification of such an ob- converge to t>hestate of the observed system, and t.hey are
server rests therefore on specificat,ionof t,he matrix H.
chosen t,o be somewhat, more negative t.han the eigenThe matrix H is determined part.ly by the fixed out,put values of the observed system so t.hat convergence is
structure of t,he original system and partly by the input
faster than other system effects. Theoretically, the eigenstructure of the observer. If SI, w3h an m-dimensiona,l valuescan be moved arbitrarily t,oward minus infinity,
output, vectory: is governed by
yielding ext.remely rapid convergence. Thistends, however, t,o make the observer act like a differentiator and
i ( t(2.4a)
) = Ax(t)
t,hereby become highly sensitive to noise, and to introduce
= C4t)
(2.4b) other difficukies. The general problem of select,ing good

598

IEEE
ON TRANSACTIONS

-w

-4UTOMATIC CONTROL, DECEMBER

1971

s+2

S+l

Fig. 2. A second-order syst.em.

Fig. 3. Struct.ure of reduced-order observer.

eigenvalues is still not completely resolved but thepractice


of placing t,hem SO that the observer is slightly faster than The
reduced-order observer was firstintroduced in
the rest of t,he (closed-loop) sgrstem seems t.0 be a good one. [I]. The simple development present,ed in this section is
Example: Consider t,he system shown in Fig. 2. This due to Gopinath [251.
state-variable
has
representation.
We aga.in consider the system

X(t) = Ax(t)
(2.7a)
(2.7b)

An identity observer isdeterminedby

specifying the

observer input vect,or

The resulting observer system matrix is

y ( t ) = CX(t)

X2

+ (3 +

g1)X

+2 + +
g1

g2

0.

(2.9)

Suppose we decide to make the observer have two eigenvaluesequal to -3. This would give the charact.eristic
equation (X
3)2 = X2
6X
9 = 0. Matching coefficientsfrom (2.9) yields g1 = 3, g2 = 4.The observer is thus
governed by

21

[ --4
5

-1 1 1 2221 1

3 .I!
-1-

(3.2b)

I;[

is nonsingular and using the variables X = Mx.) It is then


convenient to partit.ion t.he state vector as
x

+ +

(3.2a)

and assume without loss of generality that the m outputs


of the system are linearly independent-or equiva.lentlg
that the outputdistribution matrix C has rank-m. I n this
case it can also be assumed, by possibly introducing a
cha.nge of coordina,tes, that. the matrix C takes the form
C = [ZiO],i.e., C is partitioned into a.n m X m identity
matrix and anm X (n - mn) zero matrix. (An appropriate
change of coordinates is obtained by selecting an (n - m)
x n mat.rix D in such a. way that

M =
which has corresponding characteristic equation

+ Bu(t)

y]
W

and accordingly writmethe syst.emin theform

+ A l d t ) + Blu(t)
Aas(t) +
+ B2~(t).

k(t) = Ally(t)
G(t) =

A22~l(t)

(3.3a)
(3.3b)

The idea of the construction is thenasfollow.The


vect.or y(t) is available for measurement, and if n-e difThe ident,ity observer ahhough possessing an ample ferentiate it., so is k(t). Since u(t) is also measureable
measure of simplicity also possesses a certain degree of (3.3a) provides t.he measurement & ~ ( i )for t,he syst.em
redundancy. The redundancy &ems from the fact. that (3.3b) u-hich has state vector w(t) andinput. Aely(t)
while the observer const,ructs an estimat,e of the entire
B2u(t).An ident,ity observer of order n - m is constructed
st.a.te,part of t,he state asgiven by the system outputs are for (3.3b) using this measurement. Later t>henecessity to
already available by direct measurement. This redundancydifferent,iat,ey is circumvent,ed.
can be eliminat.ed and an observer of lower dimension but.
The justifica6ion of t.he const.ruct.ion is based on the
still of arbkrary dynamics can be constructed.
following lemma [%I.
The basic construction of a reduced-order observer is
Lemma. 2: If (C, A ) is completely observable, then so is
shown in Fig. 3. If y ( t ) is of dimension m , an observer of
order n - m is constructed with state z(t) t.hat, approxi- (An, Am).
mates Tx(t) for some m x n mat,rix T , as in Theorem 1.
The validitmyof t.his stat.ement is, in view of the preceding
Then an estimate i ( t ) of x ( t ) can be determined through
discussion, int-uitivelyclear. It ran beeasily proved direct.ly
by applying the definition of comp1et.e observability.
To construct the observer we initially define it in the
form
provided thatthe indicat,ed part.itioned mnt,rix is in- h(t) = (A22 - LAn)zir(t) A P I Y ( ~ ) B 2 ~ ( t )
vertible. Thus the T associat,ed with the observer must
L ( ! m - A d t ) ) - LBlU(t). (3.4)
have n - m rows that are linearly independent of t.he
rows of C.
I n view of Lemmas 1 and 2, L can be selected so that

111. REDUCED
DIMEKSIOK
OBSERVER

599

LUENBERGER: AN INTRODUCTION TO OBSERVERS


Y

82 -LE,

Fig. 6 . Observer for second-order system.

Fig. 4.

order system with a. single output so a first-order observer


with an arbitmry eigenvalue can be constructed. The C
matrix already has the required form, C = 1 0. In this

Reduced-order observer using derivative.

case A22 - GAB = -1 - G, which gives t.he eigenvalue


of the observer. Let. us select G = 2 so that the observer
will haveits
eigenvalue equal t.0 -3. The resulting
observer a.ttachedto thesystem is shown in Fig. 6.

IIT.OBSERVING

Fig. 5. R.educed-order observer.

Azz - EAzl has arbitrary eigenvalues. The configuration


of this observer is shown in Fig. 4.
The requireddifferentiation of y can be avoided by
modifying the block diagram of Fig. 4 to that of Fig. 5 ,
which is equivalent at thepoint, ~. This yields the desired
final form of t,he observer, which can be written

i(t)

+ ( A n - LAlz)~%(t)
+ (Azl - LAll)y(t) + (Bz - LBl)u(t)

= (Azz -

LAlz)z(t)

(3.5)

A SINGLE

LINEARFUNCTIONAL

For some applications an estimate of a single 1inea.r


functional E = ax of the st,at,eis a.11 t,hat isrequired.
For example, a linear t,ime-invaria,nt, control law for a
single input, syst?emis by definition determined simply by a
linear functional of the system stat.e. The quest,ion arises
then as 60 m-het.her a less complex observer can be constructed to yield an estimate of a given linear functional
tha.n an observer that estimat,es t.he entire st.ate.Of course,
again, it is desired to have freedom in the selection of the
eigenvalues of the observer.
A major result for bhis problem [ 2 ] is that. any given
linear functiona,l of the state, say, E = QX, can be estimat.ed m&h an observer having v - 1 arbitrary eigenvalues.
Here v is the obsemability index [ 2 ] defined as the least
positive integer for which the matrix

[CIAC;(Ar)2Ci. . \ ( A r ) y - l C r ]

with
z(t) = G ( t ) - 5 / ( t ) .

(3.6)

For this observer T = [-LIZ]. This const,ruction enables


us 60 st,ate thefollowing theorem.
Theorem 3: Corresponding t.0 an nth-order completely
controllable
linear
time-invariant
system
having
m
linearlyindependent outputs astat.e observer of order
n - m ca.n be constructed having arbit.rary eigenvalues.

has rank n. Since for any completely observa.ble system


v - 1 5 n - m and for many systems v - 1 is far less
than n - m, observing a single linear functional of t,he
stat.e may be f a r simpler than observing the entire &ate
vector.
The genera.1 form of the observer is exactly analogous
to a reduced-order observer for the entire state vector.
The estimate of E = ax is defined by

It is important to understand that the explicit. form of


2 ( t ) = by(t)
cz(t)
(4-1)
the reduced-order observer given here, obt,ained by partiOioning the system, is o d y one way to find the observer.
i ( t ) = Fz(t) Hx(t)
TBu(t)
(4.2)
I n a.ny specific instance, ot,her techniques such as transforming to canonical form or simply hypot,hesizing the where F,H , T , B a.re as in Sect,ion 11-A and where b and c
general structure and solving for the unknown parameters are vectors satisfying bC
cT = a.
may be algebraically simpler. Theorem 3 guarantees that,
Again the important, result is that the observer need
suchmethods will always yield an appropriate result,. only haveorder v - 1. The precise design technique is
The preceding method used in the derivation is, of course, dict,ated by considerations of convenience.
often a. convenient one.
We illust.rate the general result 1vit.h a single example.
Example: Consider the syst,em shown in Fig. 2 and The method used in t.his example ca,n, however, be applied
t,reated in the exa.mple of Sect,ion 11. This is a second- to any multivariable system.

+
+

600

IEEE TRANS.4CTIONS ON AUTOXL4TIC COSTROL, DECEMBER

1971

observer it,self and those t,hxt mouldbe obtained if the


control lax- could be directly implemented. Thus an observer does not change the closed-loopeigenvalues of a
design but merely adjoins its own eigenvalues. Similar
results hold for systems with non1inea.r control laws [a].
Suppose we have the system

Fig. 7. A fourth-order system.

X(t)

+ Bu(t)

Ax(t)

(5.la)

-2

XI

and the control law


x3

If it were possible t,o realize this control law by use of

Fig.8. Funct,ional observer.

available
measurements
(which
would
be
possible if
= RC for some R), then the closed-loop system would
be governed by

Example: Consider t,he fourt,h-order systemshownin


Fig. 7. This system nith available measurements x1 a.nd
X(t) = ( A B K ) x ( t )
(5.3)
x3 has observability index 2. Thus any linear functional
can be observed T1-it.h a first-order observer. Let us decide and hence its eigenvalueswould be the eigenvalues of
to construct an observer with a single eigenvalue equal to A
BK.
-3 to observe the functional xz x.%.
Now if the control cannot be realized directJy, an obInitially neglecting theinput
'LC we hypothesize an
server of the form
observer of the form
i ( t ) = Fz(t) Gp(t) TBu(t)
(5.4a)
i = -32
91x1
93x3.
~ ( t=
) Ki(t) = Ez(t)
Dg(t)
(5.4b)
According to Theorem 1this has an associated T satisfying
where
r-2
1
o
TA - FT = GC
(5.5a)
T I
3T =
0 93
(4.3)
K = ET DC
(5.5b)
-1
0
0 0
canbeconstructed.
From ourprevious theory ( C ,A )
If T = tl tz t3 t4, we would like tz = 1, t4 = 1. Sub- completely observableis sufficient. for t.here to be G ,
E, D! F , T satisfying (5.5) with F having arbit,rary eigenst,itut.ingthese valuesin (4.3) we obtain the equation
values. Sett.ing u(t) = K f ( t )leads to the composit,esystem

+
+

01

; -;-; :] +

9'

;] [GCA ++ BDC
TBDC
=

that ca.n be solved for the fourunknowns tl, g1? t3, g3.
This resultsin tl = - 1, t3 = - 3, g1 = -2, g3 = - 5.
From this t,he final observer shonn in Fig. 8 is deduced by
inspection.

V. CLOSED-LOOP
PROPERTIES
Once an observerha.sbeen
construct.ed for a 1inea.r
system which produces an estimate of t,he state vector or
of a linear transformat.ion of the stat.e vector it is important, to consider the effect induced by using this estimate
in place of t.he true value called for by a cont,rol lau-. Of
paramount importance in this respect is the effect, of an
observeron
the closed-loop stability properbies of t.he
system. It would be undesirable, for example, if an otherwise stable cont.ro1designbeca.me unsta.ble when it was
realized by
introduction
of an observer. Observers,
fortunat.ely, do not dist.urb stabilit,y properties when they
are introduced. I n this section we show that if a linear
time-invariantcontrollaw is realized u-ith an observer,
the resulting- eigenvalues
of the srst,em are t.hnse
he
~~~-- - of
_ - t---~

"1.

+BETBE] z

(5.6)

This whole structurecan be simplified byintroducing


= z - Tx and using x and E as coordinates. Then (5.6)
becomes, using (5.5)

(5.7)
Thus the eigenvalues of the composite system arc those of

BK and of F.
m e note that in view of Lemma. 1 (applied in its dual
form) if the syst,em (5.1) is completely controllable it is
possible t,o select. K t.0 place the closed-loopeigenvalues
arbit,rarilg. If this control Ian- is not. realizablebut, the
systemiscomplet.dyobservable,
an observer (of some
order no great,er than n - m ) can be const.ruct,edso that
t,hecontrol law can be estimated. Sincc t,heeigenvalues
of the observer are also arbitrary the eigenvalues of the
completecomposite system may beseltct,edarbitrarily.
We thereforest.ate
t.he following important result of
linear syst,em theory [I], [2].
A

Theorem 4: Correspondingnth
an
to

order completely

a.nd hence t,he plant follows the free system. This tracking
property can be used t.o definea closed-loop system for the
plant.
Rather than fix a.t,t,entionon t,he fact that only cert,ain
outputs of the
plant
are
available, we concentratme
on the
fact. that only certain inputs, as defined by B, are availa.ble.
If we ha.d comp1et.e freedom as to where input.s could be
supplied, t.he output, limitation would not much matter.
Indeed, if t,he out,put y(t) = Cx(t) could be fed t.0 t.he
system in theform

Fig. 9. A general servodesign.

Fig. 10. Compensator for example.

i(t) =

Ax@)

+ Ly(t)

03-31

t.hen t,he eigenvalues of the system would be the eigenLC. By Lemma 1, if the s:stem is observIralues of A
able L can be selected to place the eigenvalues arbitrarily.
Thedual observer can be thought of assupplying an
approximation to the desired inputs.
To achieve t,he desired result we construct, thedual
Alt.hough this eigenvalue result for linear t.ime-invariant, observer in the form
systems is of great theoretical interest, it. should be kept
z ( t ) = Fz(t) M w ( t )
(6.4a.)
in mind t,hat.t,he more general key result is that. stabilitg
is not afect,edby a (stable) observer. Thus even for nonu(t) = y ( t )
CPz(t)
(6.4b)
linear or t,ime-varying cont,rol laws an observer can supply
u(t) = Jz(t)
Nw(t)
(6.4~)
a suit,able estimate.
Example: Suppose a feedback cont.rol spst,em is t.0 be where
designed for the syst.enl shown in Fig. 2 so t.hat its output
AP - PF = BJ
(6.h)
closely tracks a dist,urbance input d. The general form of
design is shown in Fig. 9.
L = P M + BN.
(6.5b)
For the particular syst,em shon-n in Fig. 2 let us decide
to design a control law that places the eigenvalues at. Equations (6.5) are dual to (5.5) and will have solution
- 1 f i. It is easily found t,hat u = - 2x1
.x2 will acJ, MI N , F wit.h F having arbit,rary eigenvalues if (6.1)
complish this. If this ~ R Wis implemented with t,he first.- is conlpletely controllable.
order observer construct,ed earlier, we obtain the overall
The composite system is
system shown in Fig. 10, which can be verified t,o have
A
BNC BJ
BNCP x .
(6.6)
eigenvalues - 3, - 1 i, - 1 -i.
=
MC
F
MCP
z]
VI. DUALOBSERVERS
Introducing n = x
P z and using z and n for coordinates
The funda.menta1 propert): of one syst.em observing
yields the comp0sit.e syst,em
another can be applied in a reverse direction to obtain a
special kind of controller. Such a c~nt~roller,
called a dual
i]
=
LC
observer, wa.s introduced by Brasch [33].
F n]
z
2
MC
Suppose in Fig. 1 the systenl S2is the given system and
SI is a syst,em tha.t we construct t,o control S2. MTehave which is the dual of (5.7). The eigenvalues of the comshown that the system SZ t.ends to follow SI and hence posite system a,re thus seen to be the eigenvalues of A
LC and the eigenvalues of F. We may therefore st,at.ethe
SI ca,n be considered as governing the behavior of S,.
dual of Theorem 4.
To make tshisdiscussion specific suppose t,he plant
Th.eorm 5: Corresponding to an nth-order completely
X(t) = Ax(t)
Bu(t)
(6.la)
controllable and complehely observable
system
(6.1)
feedy ( t ) = Cx(t)
(6.lb) having T linearlyindependent.inputs,adynamic
back system of order n - r can be constructed such t,hat
is driven by the free syst,em
the 2n - r eigenvalues of t.he composite system t,alce any
i ( t ) = Fz(t)
(6.2a.) preassigned values.
controllable and completely observable system (5.1)
having m linearly independent out.puts, a dynamic feedback syst.em of order n - m ca,n be constructed such that,
the 2n - 7% eigenvalues of the composite syst,em take ang
preassigned values.

+
+
+

41

+
+

[" +

01

~ ( t=
) Jz(t)

(6.2b)

1711. CONCLUSIOKS

It has been shown t.hat mising &ate-variable inforwhere AP - PF = BJ for some P . Then from Theorem 1
mation,
not available for measurement,, ca.n besuit,ably
we see that in thiscase the vector n = x
Pz is governed
approximated
by an observer. Generally, as more output,
by the equa.tion
variables are made available] t.he requiredorder of the
n ( t ) = An(t)
observer is decreased.

602

IEEE TRANSACTIONS O N AUTOMATIC CONTROL, DECEMBER

Although the introductorytreat,ment given inthis


pa.per is restrictedtotime-invariantdeterministic
continuous-t,ime 1bea.r syst,ems, much of the t,heory can be
ext,ended t.0 more general sit.uations. The references cited
for this paper should be consulted for these extensions.

1971

[281 P. V. Nadezhdin,The optimal control law in problanswith


arbitrary initial conditions, Eng. Cybern. (USSR), pp. 170174. 1968.
[291 E. E.
APPlicat,ion of observers a n do p t h u m filters to
inertial systems, presented at. the IFAC Symp. Multivariable
Cont,rol SFtems, Dusseldorf, Germany, 1968.
[30] D. Q: Xayne and,:.
Murdock,Modalcontrol
of linear time
mvarlant systems, Int. J . Contr., vol. 11, no. 2, pp. 223-227,
1970.
[31] A.
Foster and P. A. Orner, A design procedure for a class of
distributedparameter
control syst.ems, ASMEPaper
70WA/Aut-6.
[32] D. &I.Wiberg, Schau.ms Outline S e ~ e s State
:
Space and Linear
Systems. New York: McGraw-Hill, 1971, ch. 8.
[33] F. M. Brasch, Jr., Compensators, observers and controllers,
to be published.
[34] R. E. Kalman, Mat.hematica1 descript.ion of linear dynamical
systems, SIAX(SOC.Ind.Appl.Nath.)
J. A p p l . Math.,
ser. A, vol. 1, pp. 15%192, 1963.
[35] D. G. Luenberger, Invertible solutions to the operator equation T A - BT = C, in Proc. Am. Math. Soc., vol. 16, no. 6,
pp. 1226 - 1229, 1965.
[36] --,Canonical
forms for linear
multivariable
system,
IEEE Trans. Automat. Contr. (Short Papers), vol. AC-12,
pp. 290-293, June 196i.
[37] R . X. Wonham, On pole assignment in mult.i-input controllable linearsystem, IEEE Tra.ns.Automat. Contr., vol. AC-12,
pp. 660-665, Dec. 1967.
of automatic
[38] 5. hI. Popov, Hyperst.abilityandoptimality
systems wit,h several control functions, Rev. Roum. Sci. Tech.,
Ser. Electrotechn. Energ., vol. 9, pp. 629-690, 1964.

M.

REFERENCES
111 D. G. Luenberger, Observing the st.ate of alinear system,
IEEE Trans. X i l , Electron., vol. MIL-8, pp. ?&-SO, Apr. 1964.
[2] ---,Observers
for
multivariable
systems,
IEEE Trans.
dubmat. Contr., vol. AC-11, pp. 190-197, Apr. 1966.
[3] M. Aoki and J. R. Huddle, Est,imation of state vector of a
linear stochastic syst.em wit.h a constrained estimator, IEEE
Trans. Automat. Contr. (Short. Papers), vol. AC-12, pp. 432433. A---I I P . 1967.1 W. A. Wolovich, On state estimation of observable systems,
in 1968 Joint Automatic Control Conf., Preprints, pp. 210-220.
J. J. Bongiorno and D. C. Youla, On observers in multivariablecontrolsystems,
Int. J. Contr., vol. 8, no. 3, pp.
221-243, 1968.
H. F. Williams: A solution of the multivariable observer for
lineartimevarying
discrete syst.ems, R e . 2nd Asilomar
Conf.Circuits and Systems, pp. 124-129, 1968.
F. Dellon and P. E. Sarachik,Optimal
control of unstable
linear plants wit.h inaccwible state,, IEEE Tran.s. Automat.
Contr., vol. AC-13, pp. 491-495, Oct. 1968.
R. H. Ash and I. Lee, State est.imation in l i n y r systemsa unified theory of minimum order observers, 3rd Hawaii
Sys. Conf., Jan. 1970.
G. W. Johnson, A deterministictheory of estimation and
control, in 1969 JointAutonatic Control Conf., Preprints,
pp. 155-160; also in IEEE Trans. Automat. Contr. (Short
Papers), vol. AC-14, pp. 380-384, Aug. 1969.
E. Tseand 33. Athans,Optimal
minimal-order observerestimatorsfordiscretelinear
the-varying systems, IEEE
Trans. Automat. Contr., vol. AC-1.5, pp. 416426, -4ug. 1970.
K. G. Brammer, Lower order opt,imal linear filtering of nonstatlonary ra.ndom sequences, I E E E Trans. Automat. Contr.
(Corresp.), vol. AC-13, pp. 198-199, Apr. 1968.
S.D. G. Cumming, Design of observers of reduced dynamics,
Electron. Left., vol. 5, pp. 213-214, 1969.
R. T. K. Chen, On the construction of state observersin
multivariable cont.rolsystems, presented at the Nat.Electron.
Conf., Dec. 8-10, 1969.
L. Novak,The
design of anoptimal observerfor
linear
discrete-t,ime dynamical systems, in Rec.4th dsilomar Conf.
Circuits and Systems, 1970.
Y. 0. Yiiksel and J. J.Bongiorno, Observers for linear multivariable systems Kith applications, t,hk issue, pp. 603-613.
B. E. Bona, Designing observers for time-varying st.ate
systems,in Rec.4thAsi1oma.r
Conf. Circuits a.nd Systems,
1970.
A. K. Newman, Observing nonlinear h e - v a r y i n g systems,
IEEE Trans. Automat. Contr., to be published.
M . 31.Kewmann, A continuous-t.ime reduced-order filter for
estimatingthestate
vector of alinear stochast.ic system,
Int. J . Contr., vol. 11, no. 2, pp. 229-239, 1970.
V. V. S. Sarmaand B. L. Deekshatulu,Optimal
cont.rol
when some of the state variables are not measurable, Int. J .
Contr., vol. 7, no. 3, pp. 251-256, 1968.
B. Porterand >,I. A. Woodhead, Performance of optimal
control systems when some of thestate variables are not
measurable, Int. J . Contr., vol. 8, no. 2, pp. 191-195, 1968.
1211 AI. 11.Newman, Optimal and sub-optimal control using an
observer when some of the statevariables are not measurable,
h t . J. CWtr., VOl. 9, pp. 281-290, 1969.
I. G. Sarnla and C. Jayaraj, On the use of observers in finitetime optimal regulator problems, Int. J . Contr., vol. 11, no. 3,
DD. 489-497. 1970.
[ 231 W . &I. Wonham, Dynamic observers-geometric theory,
IEEE Trans. Autontat. Conk. (Corresp.), vol. 4C-15, pp.
238-259. Am. 1970.
E41 -4.E. Brygon,-Jr., and D. G. Luenberger, The synt.hesis of
regulator logic using state-variable concepts, Proc. IEEE,
vol. 58, pp. 1SO3-1811, Nov. 1970.
B. Gopinat.h, On the control of linear multiple inpuboutput.
systems, Bell Syst. Tech. J., Mar. 1971.
1). K. Frederick, and G. F. Franklin, Design of piecewiselinear swit.ching functionsforrelay
control systems, IEEE
Trans. Automat. Contr., vol. AC-12, pp. 380-387, Bug. 1967.
J. D. Simon and S.K. AIitter, A theory of modal control,
Inform. Contr., vol. 13, pp. 316-353, 1968.
~I

__

David G. Luenberger (S57-kI@-S-MJ7l) was


born in Los Angeles, Cali., on September 16,
1937. H e received the B.S. degree from the
California Inst,it.ute of Technology, Pasadena, in 1959 andthe
M.S. and P b D .
degrees from Stanford
University,
Palo
Alto, Calif., in 1961 and 1963, respectively,
all in electricalengineering.
Since 1963 he has been on t.he faculty of
Stanford Universit,y, where presently he is a
Professor of Engineering-Economic Systems
and of Electrical Engineering. He is also d i a t e d with the Department of Operations Research. His activit.ies have been centered primarily in t,he graduate program, where he has t,aughtcourses in optimization, control, mathematical programming, and information
theory. His research areas have included observability of linear systems, the applicat.ion of functional analysis t.o engineering problems,
optimal cont.ro1, mathematical programming, and optimal planning.
His experience includes summer employment a t Hughes Aircraft
Company and Westinghouse ResearchLaboratories from 1959 t o
1963, and service as a consultant to West.inghouse, Stanford Research Institute, Wolf Management Semices, and Intasa, Inc. This
experience has included work on the control of a large power generating plant, the numerical solution of partid differential equations,
optimization of trajectories,optimal planning problems, and numerous additionalproblems of optimization and control. H e has
helped formdate and solve problems of control, optimization, and
general analysis relating t.0 electric power, defense, water resource
management., telecommunications, air traffic control, education, and
economic planning. He is the author of Optimization by Vector Space
Xethods (Wiley, 1969) and has authored or co-authored over 35
technical papers. Currentlyhe is onleave from Stanford at the
office of Science and Technology, Executive Office of the President,
Washingt.on, D.C.He assist,s the Science Adviser -with program
planning, review and evaluation, and with formulation of science
policy in areas of civilian technology.
Dr. Luenbergeris a member of Sigma Xi, TauBetaPi,the
Society for Industrialand Applied Mathematics,the Operations
Research Society of America, andtheManagement
Science Institute. He served as Chairman (Associate Editor) of the Linear
Systems Comnlit.tee of the IEEE Group on Automatic Cont,rol
from 1969 to 1971.