StableAdaptiveController

Design, Part 11:

Proof of Stability
KUMPATI S. NARENDRA, FELLOW, IEEE, YUAN-HA0 LIN, STUDENT
AND LENA S. VALAVANI

I.

INTRODUCTION

PAPER we present a proof of the stability in

Iof NthethatTHIS
large of an adaptive system which
is a modification
originallysuggestedbyMonopoli
[l] and later
refined by Narendra and Valavani [2]. The proof of stabil-

ity applies to both continuous and discrete systems.

In [2] the stability problem was clearly defined and a
conjecture was made that the planttogetherwiththe
controllerwouldbestable
if theaugmentederroris
bounded. Recently, while applying the same approach to
discrete systems, Lin and Narendra [3] suggested a new
error model based on the third prototype discussed in [2]
of Kudva and
(Lemma 1) and on the earlierwork
Narendra [4]. This model contains an additional feedback
term similar to that found in much of the. adaptive observer literature [5], [6] and is particularly useful in proving the stability in the large of discrete adaptive systems
[q.The successful resolution of the latter problem providesthemotivationfor
using asimilarmodel
in the
continuous case as well.
Ever since the publication of [l] there has been a great
deal of interest in the problem of designing stable adaptive controllers. In [ t i ] Feuer and Morse presented such a
controller but its extreme complexity precludes its use in
practical applications. Since then, several authors [9]-[13]
have also attempted this problem for both discrete and
continuous systems. In [9] the stability in the large of a
Manuscript received April 27, 1979; revised December 10, 1979. Paper
recommended by J. L. Speyer, Past Chairman of the Stochastic Control
Committee. This work was supported by the Office of Naval Research
under Contract N00014-76-C-0017.
K S. Narendra and Y.-H. Lin are with the Departmentof Engineering
and Applied S
c
e
in
c
e
,Yale University, New Haven, CT 06520.
L. S. Valavani was with the Department of Engineering and Applied
ff 06520. She is now with the
Science, Yale University, New Haven,
Department of Electrical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139.

MEMBER, IEEE,

simple discrete adaptive controller is consideredand both

discrete and continuous controllers are discussed in [lo].
In [ 111 and [12]the continuousadaptiveproblem
is
treated when the input to the system is "sufficiently rich."
A proof of stability for the continuouscase is proposed in
[ 131 for the adaptive control problem considered[I]
inand
[2]andalsoinvolvestheadditionalfeedbacksignaldescribed earlier.
The problem discussed in this paper is essentially the
same as that in [2]. The philosophy as well as the specific
for theadditional
structure of thecontroller(except
feedback term) and the adaptive laws are the same as in
[2]. Hence, these aspects of the problem are discussed in
Section I1 only in a condensed form and the reader is
referred to [2] for all details. The remaining sections are
devoted to the proof of stability with which this paper is
primarily concerned. The simple proof givenin Section V
can be used, as mentioned earlier, for both continuous as
well as discrete systems with only minor modifications
but
only the continuous case is discussed here. The reader is
referred to a companion paper [Afor a description of the
stability problem in the discrete case.
The augmented input to the reference model was first
introduced in [l] in ordertoavoiddifferentiation.The
concept of positive realness which is central to the development of stable adaptive laws led to the introduction of
the operator' PL($)=L(s)$(t)L-'(s)in 121. Much of the
difficulty encounteredin proving the stability of the adaptive system can be traced to the characteristics of P,(B)
and therelatedoperator
FL(8)=L-'(s)B(r)L(s). Since
PL(8)= pL(fJ)=8, when 8 is a constant, itwas felt that the
adaptivecontrollerwould
be stable if O(t) changes
"slowly" with time. This intuitive assumption can now be
justified by the stability analysis presented tin
his paper.
In Section I1 the structure of the adaptive controller is
discussedbriefly and theprincipalstabilityquestion
is
restated in Section 111. While state-variable equations are
used wherever appropriate, the emphasis throughout the
paper is on input-output descriptions. SectionIV is essentiallyself-contained and all the mathematicalconcepts
needed are developed here. Using these results the proof
of the adaptive system is shown to follow directly in
Section V.
'Throughout the paper both differential equations and transfer functions are used in the arguments and, depending on the context, "s" is
used as a differential operator or the Laplace W o r m variable.

0018-9286/80/0600-00$00.75

01980 IEEE

NARENDRA et

441

d.:STABLE ADAlTIW! CONTROLLER DESIGN

Structure of the Controller (n* <2)

The basic structure of the adaptive controller used in

the following sections is shown in Fig. 1. The controller
consists of a gainco and two auxiliary signal generators Fl
and F,. F, contains ( n - 1) parameters ci ( i = 1,2,-. ,
n - 1) and F2 contains n parameters 4 (j
=0,1,2,. ' ,
n - 1). The 2n adjustableparameters are denotedby
the elements of a parameter vector

1-

[ Co(t),e*(t)]

= [ CO(t),Cl(t),.
A

,c,-,(t),$(t),d,(r),. * ,d,-,(t)].
(5)

F, and Fz are described by the vector differential quatiOnS

A plant P to be controlled is completely represented by
the input-output pair { u(t),yP(t)}and can be modeled by
a linear time-invariant system

6(l)=Au(l) + b,u
w(l) = c Tu(l)

ip=ApXp(t)+bpU(t)
Yp= h,TX,(t)

(1)

where 4 is an (n X n) matrix and % and bp are n-vectors.

The transfer functionof the plant is Wp(s)where

where A is an (n - 1) X (n - 1) stablematrix, c *=
[c1,cz,-",cn-1], anddT=[d~,dzy...,d,-l].
Defining the vectorof transfer functions V(s) as

V(S) (d--h)-'b,
with Wp(s)strictly proper, Zp(s)is a monic polynomial of
degree m( Q n - l), %(s) is a monic polynomial of degree
n, and J$ is aconstant gain parameter. We further
assume:
1) the sign of E$,
2) an upperbound on the order n of the plant,
3) the relativedegree n*( = n- m) of the plant are
known, and that
4) the monk polynomial Z,(s) is Hurwitz.
A model M represents the behavior desired from the
The
plant when itis augmented with a suitable controller.
modelhasareference
input r(t) whichisuniformly
bounded and an outputy,(t). The transfer functionof the
model, denoted by W,(s) may be represented as

V;.(s),the ith component of V(s), denotesthetransfer

function fromu(t) to ujl)(r) ory,(t) to u,!')(t). For constant
values of the control parameters the transfer functionsof
F, and F2 are W&) and Wz(s),where

and N(s) contains Z,(s) as a factor. If (h,b,) is in

companion form,ci and 4 ( i = 1,2, ,n - 1) represent the
coefficients of s"-+-l of thenumeratorpolynomials of
W,(s)and W2(s),respectively.
Denoting the principal signals of interest r(t), dl)(?),
yp(t),and o@( t) as the elements of a vector W ( t )

---

zT(t)
where ZM(s)is amonic Hurwitz polynomial of degree

m,Q m ,R,(s) is a monic Hurwitz polynomial of degree n

and K, is a constant. W,(s) is completely specified and
the aim of thedesign is togeneratesuitablebounded
inputs u(t) to the plant so that the deviation of the plant
from the desired behavior as measured by an error signal
cl(t) where

I
tends to zero at t+co.

tYp(t)--YM(t)l

(4)

(7)

[r(t),o*(t)]

A [ r(t)yu(1)'(t),yp(t),u(2)*(t)]

(9)2

the inputto the plant,as shown in Fig. 1, is

~ ( t ) e'(t)is(t)-ae1(t)~;(t)r7j(t)

r=rT>o, a>o.
(10)

% (5) and (9) the Zn-dimensionalvectms.&t) and qi)are defmed. In

the main discussions in the paper, as e x p h e d later, only the (2n - 1)-

dimensional vectors O(t) andw(r) are ust;d.

442

IBBE TR4NSACIIONS ON AUTOhfATIC CONTROL, VOL. AC-25, NO. 3, JUNE 1980

I
Fig. 2. The enur model.

The only difference between the controller structure described in (5)-(10) and that given in [2] is the existence of
the second term in the right-hand sideof (10).
Following theresults of [2] it can beshown that a
constant control parameter vector #* exists such that if
&c)~8*the transfer function of the plant together with
the controller matches that of the model exactly. If e(t)
represents the state error between model and plant, the
error differential equations are

- aEl(t)o
~(t)rqt)]

i ( t ) =ACe(t) b C [ G T ( t ) q t )

el(f)=h:(t)

1)

(1

where W&) h:(zI -A,)- 'b, =$/ KM WM(s)is strictly

positive real, and +((t)=f?(t)- is the parameter error
vector ([2, equations (18), (19)] and Fig. 2). The adaptive
laws

e*

&I)=
-rcl(t)w(t)

assure the boundedness of

(12)

a?)and ~ ( tand
) that eI(f)+O

as t+oo (see Section 111 for proof)? Since the statesof the
model are bounded, this also assures the boundedness of

laws. In such a case, d ( t ) = -I'e,(t){(t), where L-'(s)o(t)

= { ( t ) . However, sincethe controller is to be differentiator
is notpossible.Therefore,the
free,suchaprocedure
following approach described in [2] is used. Corresponding to every feedback signalf?,(t)oi(t)
to the plant, asignal
[ei(t)-P,(Bi(t))]oi(t) isfedback into themodelto p r e
serve the form of the error differential equations (1) and
hence the adaptive laws (12). This results in the structure
of the adaptive controller described later in this section.
The complexity of the controller (
asgiven in [2l)is also
determined by the prior knowledge of the gain l$ of the
plant. We therefore deal with the simpler case first, when
6 is known, and laterconsiderthecasewhen
Kp is
unknown. In Section V, forclarity or presentation,the
proof of stability is also given separately for thetwo cases.
Case i) (5 known): With no loss of generality, we can
assume here thatKM = K,,= 1. This implies that co= 1 and
of
hence only 2n- 1 control parameters [the elements e(t)]
need to be adjusted.
~f
the input signalt i t h e plant is
U( t )=

e =(t)w( t )+ r( t )

(13)

the plant output can be expressed as

r,(t)= ~ M ( S ) [ r ( t ) + r p T ( t ) W ( t ) ] .

(14)

The model output is augmented by the additional output

y,( t ) where

uu(t)= W , ( s ) W [{ r . - ' ( s ) e ( t ) - e ( t ) L - ' ( s ) } * ~ ( t )

the plant states.

+ 4r)S=(t)rS(t)](15)
For a relative degree n* =2, the modification suggested
in [ 11and [2] can again beused in this case also to develop and L-'(syW(r)= S(t) (Fig. 3).
The total output of the model y ( t ) is the s u m of the
adaptive control laws of the form (12) [2, equation (2211.
The principal difficulty, therefore,arises when the relative desired output y M ( t )= W&(s)r(t)and the augmented outdegree n* is greater thanor equal to three. In such a case, put y,(t); the augmented errorel(t)
an auxiliary signal corresponding to each control paramee l w ki Y,(t) - A t ) = 4 t )
(16)
ter has to be used as an input to the modelas described in
[21.
It is worth pointing outthat for the case when n* = 1 or is given by
2, the second term on the right-hand side of(10) is not
el(O = w M ( s ) WcpT(t)S(t)
[
- ael(r)tT(t)rS(t)] (17)
needed to prove global stability of the adaptive loop. It is
as (11)with G(t) and e l ( t )
only when n* > 3 that this term is found to be important which has thesameform
in the proof of stability as shown in the following sections. replaced by {(t) and e,(t), respectively.
The adaptivelaw for updating the control parameters
is
now
Structure of the Controller (n* > 3)

-m

If n* 3, with no loss of generality [21], let L(s) represent a Hurwitz polynomial in ccs"such that W,(s)L(s) is
a strictly positive real transfer function. If every parameter Oi(t) inthecontroller
in Fig. 1 is replacedby an
operator PL(Oi(t))= L(s)Oi(t)L-'(s), the same procedure
as that outlined earlier can be used to derive the adaptive
is the errm between plant and (
u
n
a
u
g
m
e
n
t
e
d
)
model outputs. In
W o n III e&) is the output error when the model output is augmented.
The form of the differential equations (11) and (12), however,
remains
samethe
in the two cases.
3c1(#)

&t)= -rel(t)S(t).

(18)

Case ii) ($ unknown): While the error equations a p

pear to be considerably more involved when
I$ is unknown, they can be readily reduced to the same form as
before by using a change of variables (asshown in [2D.
The input to the plant is now given by
u( t ) =

eyt)Z( t).

(19)

The main difficulty here is that K, is different from KM

and hence an additional paramek #l(t) has to be i
k

NARBNDRAet

ai.: STABLE m

443

m CONTROLLER DESIGN

rlt)

cc
Y

Fig. 3. Adaptive controller for n* > 3 and $ known.

troduced in series with the auxiliary inputto the model to

obtain the error equations in the form given in (17). For
eachsignal Oj(f)wi(f) f e d back into theplant a signal
L(s)+,,(t)[L-'(s)O,(t)-O,(t)L-'(s)]w,(t)has to be f e d back
into the model (or +,(f)[L-'(s)8,(t)-f?,(t)L-'(s)]wj(t)
into
WM(s)L(s)for practical realization). The 2n+ 1 parame,2n) and #,(t) have to be adaptively
ters e,(?) ( i = 1,2;
adjusted so that +,(t)+E$/& and O,(t)+O,* as t+m.
The augmented error e l ( t ) and the adaptivelaws are
now given by [2]

--

el(t)=

% [ %f(s)Us)] ( [ B T ( t ) C ( t )+ w ) m ]
- alel(t)CT(t)rf(~)](20)

&)- - r e , ( t ) E ( t )
= -uedtk(4

(21)

where

function that the state error vector e(t) or e(t) as well as

the parameter error vector
Nt) are bounded. Whenn* < 2
(and no auxiliary input to the modelis used) the outputof
the model yM(t)is bounded and hence the boundednessof
the plant outputcan be concluded. This, in turn, results in
a,(t)+O
as t+m. However,when
n* > 3 we cannot
assume that themodel output (y(t)=y,(t)+yM(f))is
bounded and this leads to the principal stability question
of interest in this paper.
If the input to the plant u(t) is given by (13)(I$known)
or (19) (I$ unknown), the error model is defined by (17)
or (20) and the corresponding adaptive laws are specified
by (18) and (21), the problem of stability is to conclude
the boundedntxs of the plant output yJt) and hence all
the relevant signalsin the system. Demonstratingthat this
is indeedthecase is the pMcipal contribution of this
paper. We now proceed to state thestabilityproblem
is known.
analyticallyforthe case when
The Error Model (Iyp Known)

f(t)=L-'(s)w(t);

Thedifferentialequationsdescribingtheaugmented
error e(t) may be expressed as [2]

i(t)=[ L - l ( ~ ) ~ t ) - ~ ( r ) L - ' ( s ) ] T wand

(t)
KM
44t)= 1- -#do.

I$

P=A,e(t)=bv(t)

(22)

The stability problemthat arises is the same whetherKp is

known or unknown and is stated in the next section.
111.

STATEMENT OF THE STABILITY PROBLEM

As shown in this section, when the adaptive laws (12),

(18), or (19) are used (for the cases n* (2, n* > 3 ($

known) and n* > 2 ($ unknown), respectively) it is possible to demonstrate through the existence of a Lyapunov

e,(t)=hTe(t)
4 f ) = + T ( t ) S ( f ) - .ST(WS(t)el(t)

a>O

(23)

where
h T ( s ~ - - ~ , ) - ' bw,(s)L(s)
=
is SPR

(24)

[compare with (17)J. The corresponding adaptive lawsare

io)
-re,(t)S(t).
=

(18)

For the case n* < 2, the equations are the same with y(t)

lEBE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-25,NO. 3, JUNE 1980

444

replaced by w ( t ) replaced by w(t) and the boundedness of

the plant output assures the boundedness of w(t) in the
corresponding error equations. A difficulty arises in this
case since y,(i) and hence {(f) cannot be a s s u m e d to be
bounded.
known
By theKalman-Yacubovichlemma[14]itis
that if condition (24) is satisfied, a real matrixP = P >0,
a vector 4, and an E> 0 exist such that for any matrix
N=N~>O

Fig. 4. Plant feedback loop and the error model when $ is known.

ATP+ PA,= - 4 q T - &

Pb= h

(25)

are simultaneously satisfied. Defining a Lyapunov

function for the set of differential equations (23), (18) as

Fig. 5. Plant feedback loop and the error model when $ is unhown.

V(e,+) A e'(t)pe(t)++'(t)r-I+(t),
L-'(s)w(t) and the plant feedback loop together with the
error model may be representedas shown in the Fig. 4.
The Stability Problem: It is then to show that the plant
P(e,+)= -eT(t)[qqT+~]e(t)
feedback loop with +(f) adjusted according to the adap-2a[ e,(t)l(t)]'I'[e,(t)l(t)] < O
for QI >O. (26) tive law (1 8) and I$ E E2 is stable in thelarge.
When I$ is unknown, the error equations (23) have to
It follows that the error modelis uniformly stableand e(t)
be
replaced by those corresponding to (20) but the plant
[hence, el(t)] and +(t) areuniformlyboundedforany
feedback
loop remains essentially the same (co(t) is also
finite initial conditions.Further, since V(t) is a noninadjusted)
(Fig. 5). Hence, the nature of the stability probcreasing function whichis bounded below, it converges to
lem
is
the
same in both cases.
some limit P. [For n* < 2 these conditions are adequate
to assure that e(t)+O as t+co]. By(26)since V(e,+) is
uniformly bounded we have
Iv. MATHEMATICAL PRELIMINARlES

we obtain using (25)

e(-),e1(.)Y(.)~C2.

(27)

since at)=-rel(t)<(t),(27)implies that

&)E

C2

(28)

the importance of the additional feedback term

ael(t)fT(t)rS(
t ) in the error model isnow apparent. Since
V(e,+) contains the term - e:(t)<T(t)rg(t),it is now possible to conclude that
&-)E C2 which is central to the
proof of stability.
The Plant Feedback Loop

The proof of stability of the adaptivecontrol problem is

discussed in Section V and is based on the six lemmas
given in this section. The definitions and related mathe
matical concepts needed for the proof of the lemmas are
developed here.
Definition I :
denotesthespace
of functionswhich
are bounded for finite time,i.e.,
E~=(f:R+~R/supJf(7)~<~,forall~R+).
t >r

Remark I : Sincetheparametererrorvector
+(t) is
uniforrnly bounded all the signals in this paper can grow

at most exponentially and, therefore, belong to this space.

The plant together with the controller may be described
Definition 2: Let x( -),y(-)E
We denote y ( t )=
by the differential equations
o[x(r)]if there exists a continuous function/3(-)such that
P(t)+O as t - m andy(t)=&t)x(t).
i(t)=A,x(t)+bm[+T(t)~(t)+r(f)]
Remark 2:
w( t )= C,x( t )
i) If y(t)= o[y(t)],theny(t)+O as t+m.
ii) If Ir(Ol=o ~ s u ~ ~ > , l ~ theny(t)-+O
(4ll,
as t+m.
or equivalently by a vector transfer function
i
i
i
) If ly(t)l=x(t)+ o[sup,>,ly(~)1]and x ( t ) is uniformly bounded, theny ( t ) is also uniformly bounded.
Definition 3: Let x ( - ) , y ( - )E
.:E If there exists a constant M >O such that ly(t)l <Mlx(t)l, then we denote

e:.

y(t)=O[x(t)].

[where V(s) is defined in (711 in the forward path

gain vector Hi) in thefeedbackpath.Further,

Remark 3: If the input to a linear exponentially stable

system is x( .)E:E and the corresponding outputis y( .),

NARLNDRA et

al.: STABLE AD-

445

CONTROLLER DESIGN

then

lu(t)l=O[S
tU
>r
P Ix(.'I].

(30)

In particular, the input and output of an asymptotically Fig. 6. An asymptotically stablefeedback system with a timevarying
stable linear time-invariant system satisfy (30).
feedback gain /3(t);b(,)E E' or '
E or #I(.)-&.
Definition 4: Let x ( . ) , y ( . ) ~ C : . If y(t)=O[x(t)] and
x(t)=O[y(t)], then we say that x(t) andy(t)are equivalent
Lemma 2: Let x ( . ) E' (or C') and p(.)E.:E
If
and denote this by
x( .)p( is the input to H(s) and y ( .) the corresponding
output, then
x(t)-Y(t).
e)

Definition 5: Let x(.),y(-) E e.:

If sup,>,ly(~)lSUP~>~~X
we( Tsay
) ~ ,that x(t) andy(t) grow at the same
rate.
Remark 4:
i) Two signals which are equivalent growat the same
rate but the converse is not true.
ii) Two signals whichare uniformly bounded growat
the same rate. Hence, this definition is of interest mainly
for unbounded signals.
iii) If x( - ) , y ( - ) are continuous,onlyone
of the
following three conditions can hold:
SUP 1
t >r

-w-

SUP lY(.)lY

.[
t >7

SUP Ir(.)l=

SUP IX(.)I].

t >r

or SUPIx(.)I
t >r

iv)

[ SUPI).(.I

t>7

= o [ SUP lu(7)1].
t >r

= o [ SUP

t >7

1 >r

lr(.)I]
or SUP IX(+

t >r

SUP lU(7)l)
t>r

f >r

t >T

v)Let ann-dimensionalvectorx(-)EE:beunbounded, then there exists at least one component

such that

xJt)

S~PIXi,(~)l-SUPIIx(.)II*
t >7

If p(.) E C' or E' or p(t)+O as r+m, then:

i) y(t)+O as t+m, if x(r)+O as r+m, and
ii) y ( t ) isuniformlybounded,
if x(r) is uniformly
bounded.
Proofi The proof is a direct consequence of Lemma
2, the linearity of H and Remark 2.
Remark 5: If in Lemma 3 1x(t)1 = o [ s ~ p , ~ ~ ( y ( T ) the
~],
same results holdand x(t),y(t)+O.
Remark 6: The signals considered in the proof of stability in Section V are all generated by differential equationswithboundedcoefficients
(also see Remark l). If
x(t) is such a signal and it satisfies the condition

t >r

In the following Lemmas 1-3, x( -),y(-)E :C and H(s)

is the rational transfer function of an asymptotically stable system and is strictly proper. All initial conditions are
assumed to be identically zero!
Lemma I: Let x ( . ) and y ( - ) be the input and output,
respectively, of a system with transfer function
H(s). Then
i) x(.) E S + y ( - ) , j ( . ) E Ep

i
i
)
iii)

x ( . ) E C' or C'+y(t)+O
x(t)+O
as t+m+y(t)+O

1< p < m
as r-+m

4~ the initial conditions are nonzero, the output in n em ma 2 contains

an additional exponentiallydecaying term,

SUP lXt.)l]
t >r

Lemma 4 shows the input-output relationshipof dynami-

c a l systems with transfer functions having left half-plane

(l.h.p.) zeros with x ( t ) as the input.
Lemma 4: Let x ( - ) and y ( - ) be the input and output,
respectively, of a system with a rational transfer function
whose zeros are inthe open left half plane. If x( -)be such
that
i) x(-)EE:

ast+m.

The proof of Lemma 1 can be found in any good texton

linear systems [ 151.

W =o[

(31)
t >r

then

446

IEEE TXWSACTIONS

'

P r u u j Without
loss
of generality
let
~ ( s=
)
H1(s)H2(s)where H l ( s ) has the same numberof poles and
zeros and Hz(s) has only poles. Let H2(s)x(t)= XI(?) and
H,(s)x,(t)=y(t). Since all the zeros of H,(s) are in the
open 1.h.p. H;'(s) is stable and hence

S
lt >=r O ~
t >~
7 P I X r ~ ~ ) supIr(.r)l].

(33)

If x(?) grows without bound there exists a sequence

such that
Ix(t,)l=
suplx(7)1

i=0,1,2,.

{ti}

ON AUTOMATIC CONTROL,

AC-25,NO. 3, JUNE 1980

VOL.

Consider the feedback loop TI defined by

d(')=Au(')+ b u

(6)

u(t)=OT(t)u(t)+r(t).

t 13)

and
Since r ( t ) isuniformlyboundedand
y,(t) and d 2 ) ( t )
satisfy (36) and (37), if u(t) is unbounded, dl)(?)
is also
unbounded and by Remark 4 there is at least one variable
u,")(t), 1 <io Q n - 1 which grows at the same rate as the
input u ( t ) or

/U(T)~.
SU~U$;)(T)~-SUP
I]U(~)(T)~~-SUP

ti >r

t $7

and

t >r

Since TI is avectordifferential
coefficients

x( ti)>0.
In view of relation ii) in (31) it follows that given any ti
there exists an interval A > 0 such that x ( t i + A ) >clx(ti)
for some constant c1>O. Since

(38)

t >r

Id$;)(?)/ =o[ sup llo'1'(7)11]

equation withbounded

=o[ sup l U p ( T ) / .
t >r

t>7

(39)

Since W,(s)is a transfer function with zeros in the open

left half plane and
(40)

up(?) = W,(s)up(t)
it follows from Lemma 4 and (39) that
where

sup Iu$;'(.)l=O[ sup I W l , ' ) ( T ) l ] .

(41)

t >r

t>7

From (38) and (41)

If it is assumed that xl(ti)=o[sup,>Tx(.r)]wehave a
contradiction by (34) since xl(ti +A) >clczx(ti) for all i=
0,1,2,. * . Hence,

t>
suplx(t)l=O[ st >r
~ P l x l ~ ~ ~ lSUPlY(?.)l]
] 7= o [

(35)

by (33) which proves Lemma 4.

While various generalizations of thefollowinglemma
can bestated,Lemma
5 is adequate fortheproof
of
stability in SectionV.
Lemma 5: In the plant feedback loop (Fig. 3) defined
by (1) and (Q, the input to the plant and the states of Fl
and F2 cannot grow faster than the output of the plant,
1.e.,
u(t),u!"(t),u!2'(t)=O[ SUPIYP(T)l]

i = 1,2,. * * ,n- 1.

t >r

sup lu(T)l=o[ SUPlUf)(T)l]

I >r

which is acontradictionto
marks 4iii) and 4iv)

t>T

(37). Hence,followingRe-

ut?) = q SUP IY,(4lj

t >r

which proves the lemma.

Remark 7: Lemma 5 isneededsincewe
cannot directlyconclude that u(t)=O[~up~>~ly,(~)1].
In theerror
model,although e,(t) is uniformly bounded, we cannot
concludethat v ( t ) isuniformlyboundedforthesame
reason.
Lemma 6 is central to the proof ofstability inSection V
and relatesthe operator FL(+)=L-'(s)+(t)L(s) to the
pure gain +(t)[2, equation (lo)].
Let L(s) be a Hurwitz polynomial in "s" of degree r.
With no loss of generalityS assume

Pro03 Since ufl)(t) and uf2)(f) are thestates of the

asymptotically stable systems Fl and Fz with inputs u(t)
r
and up(?),respectively,wehave
U~~)(~)=O[SU~~>,~Y,(T)~].
L(s)=
(s+a,)
i= 1
Hence, it suffices to show that
where a, is real. Defining A&) as
u(t)=O[ SUPIYPtT)l].
t >7
i
A,(s) L(s)/
(s + ai),
j=1
We show in what follows that

II

(42)

II

y , ( t ) = o [ suPlu(T)l]
t>7

i=1,2;.-,r-l;

(36)

A,(s)=L(s)

us)

5~f
contains complex conjugate zeros, the form of the expansion
of the operator L-'@L remains thesame but thecoefficients of the
polynomials have to be modified. A,(s) is, however,stable for all i=

leads to a contradiction. By (36) we have

(37)

0,1;.., r-l,e.g.,ifL(s).=s2+2as+bandhascomplexzerosL-'cPL=

cP-[(S+LI)/SZ+2as+b)EP-[l/(s2+2as+b)Ep(s+a).

447

if { ( t ) is a function which is r-times differentiable, then

[ L-'(s)+(t)L(s)]S(t)
=

$40-

5'

A 3 s ) & ( t ) A i +I($)

i=O

{ ( t ) + 4 1 ) (43)

where ~ ( t is) an exponentially decaying term due to the

initial conditions. Fig.7 shows thetwo equivalent formsin
(43)and the corresponding signals.
Lemma 6: If L(s) is defined as in (42) and L(s)S(t)=
w(t) with w(-)~f!:
and ;P(.)EE2nE;, then

:[-

y,(t)=

Proof: The proof follows directly by usingthe expansion (43),Lemma 2, and Remark 3 since

I[%
1
+ [t
I

W,(s)

= wM(s)r(t)

Ai(s)S(t)=Ai(s)L-I(s)w(t)=O[suplo(.r)l].

%(t)+p(t)qt)

-W,(s)L(s)

t>r

v.

PROOF OF

By the same argumentsas case i)

STABILl'lY

The proof of stability of the plant feedback loop follows

directly from Lemmas 5 and 6 in Section IV.
The motivation for the proof may be described
qualitatively as follows: assuming that o(t) and hence { ( t ) are
unbounded, Lemma 5 assures us that they cannot grow
faster than the plant output y,(t). Using Lemma 6, y,(t) From Fig. 5
can be shown to be the sum of three signals, W,(s)r((t),
S=(t)f(t)=v(t)-a;=(t)f(t)-$(t)z(t).
el(t), and y,(t) where the first two are bounded and the
third term byLemmas 2, 3, and 5 is ~[sup,,~!y,(.r))]. But
Hence, we have a contradiction and y,(t) is umformly
$(t)=[ L - ' ( s ) p ( t ) - p ( t ) L - ' ( s ) ] o ( t )
bounded. This, in turn, impliesthat w(t), [ ( t ) and x,(t) are
uniformly bounded.
=L-'(s)G=(t)L(s)f(t)-p(t)f(f)
Proof:
Casei)n*>3andIr$=KM=l(Fig.4): Byl..emma5,
= o[ SUP IYJ.)l].
if we can show that y,(t) is uniformly bounded, then all
t>r
signals in the system are uniformly bounded. From (14)
Since $(t) is uniformly bounded, again we conclude that

u,(t)

=W

M ( M f )+ W M ( S ) + = ( M t )

where W,(s)r(t) is a uniformly bounded signal.

(44)

y,(t) is uniformly bounded.

We have established that w(t) is uniformly bounded,
hence the input to the error model y(t) is also uniformly

448

IEEE TRANSACTIONS ON AUTOMATIC

bounded. It follows from (23) and (26) and the definition

of { ( t ) that e(t), k( t), {(t), and f ( t ) are alluniformly
boundedandhencethesecondderivative
of theLyapunovfunction f ( t ) isalsouniformlybounded.
This
assures that f ( t ) is uniformly continuous and hence

lim

r+eo

f(<t)=
lim e($)=
t+oo

lim &(Z)=O.

r+m

We alsonotice that theaugmented input signaltothe

model [ L - l ( s ) + T ( t )- T ( t ) L - l ( s ) ] ~ ( ta) n d
ae,(t){ T(t)I{(t) alsotend to zero as t+w. Hence,the
error
e,(t)=y,(t)-y,(t)--+O

as t+oo.

VI. CONCLUSION
The paper presents a complete proof of the stability in
the large of continuous adaptive control systems. Except
for an additional term, the controller structure is identical
to that givenin [2]. Whenthespecifiedmodeltransfer
function W,(s) is not positive real, an operator PL(0)=
L(s)O(t)L-(s)was used in [2]to generate a stable adaptivelawforaparameter
O(t). Almost allthestability
questions that arise in the adaptive control problem can
be traced to this
operator as well as a related operator
P,(O)=L-(s)O(t)L(s)and the extent to which these approximate a pure gain O(t). Intuition suggests that when
O ( t ) is constrained to varyslowlyinsomesense(e.g.,
d( .)+O or d( -)E the operator FL(fl)would approximate O(t). The proof given in Section V may be considered to be a mathematical justification of this statement.
In thissense, it isbasedon our intuition regarding the
behavior of the adaptive loop and the augmented inputs
[l] and
which originally led to the schemes suggested in
[21*

e)

CONTROL,
VOL. AC-25,NO. 3, JUNE 1980

A. Feuer and A. S. Morse,Adaptive control of single-input

single-outputlinearsystem, IEEE Trans. Automat. Contr., vol.
AG23, pp. 557-570, Aug. 1978.
[91 G. C. Goodwin, P. J. Ramadge, and P. E. Caines, Discrete time
multivariable adaptive control, IEEE Trans. Automat. Contr., this
issue, pp. 449-456.
B. Egardt, Stability of model reference adaptive and self-tuning
regulators,Dept.Automat.Contr.,
Lund Inst.Technol.,Tech.
Rep., Dec. 1978
H. Elliott and W. A. Wolovich. Parameter adaptive identification
and control, IEEE Trans. Automat. Contr.,- vol.AC-24,pp.
592-599, Aug. 1979.
G. Kreisselmeier, Algebraic separation in realizing a linear state
feedback control law by means of an adaptive observer, IEEE
Trans. Automat. Contr., vol. AC-25, pp. 238-243, Apr. 1980.
A. S. Morse, Global stability of parameter-adaptive control systems, IEEE Trans. Automat. Contr., this issue, pp. 433-439.
K. S. Narendra and J. H. Taylor, Frequency Domain Criteria for
Absolute Stability. NewYork:Academic, 1973.
C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output
Prqerties. NewYork:Academic, 1975.
K ~ ~ t ~ pS.
a t NPrendrP
i
(S55-M60-SM63-F79)
receivedthePh.D.degreeinappliedphysics
from Harvard University,Cambridge, M A in
January 1959.
At present, he is Professorof Engineering and
Applied Science, Yale University, New Haven,
(3. He has consulted for numerous industries
during the past twenty years including Honeywell, S p r y Rand Research Center, Bell Aero
Systems,Schlumberger, Sikorsky Aircraft, and
Raytheon and is now a consultant to AT & T
Long Lines Division on telephone traffic routing using learning theory.
His present interestsare in the areas of stability theory,adaptive control,
and learning automata. He is theco-author (with J. H. Taylor) of a book
entitled Frequency Domain Criteria for Absolute Stability (Academic,
1973) and is currently working on a book entitled Learning Automata
with M. A. L. Thathachar.
Dr. Narendra is Chairman of the Adaptation, Learning and Pattern
Recognition Committee of the Control Systems Society, the Editor of
Journal of Cybernetics and Information Sciences, and a member of the
AmericanAssociation for theAdvancement of Science. He wasthe
recipient of the 1972 Franklin V. Taylor award of the IEEE Systems,
Man, and Cybernetics Society.

ACKNOWLEDGMENT
The authors would like to thank Profs. Thathachar and
Monopoli for numerous discussions.
REFERENCES

[l] R. V. Monopoli, Model reference adaptive control with an augmentederrorsignal, IEEE Trans. Automat. Contr., vol.AC-19,
pp. 474-484, Oct. 1974.
[2]K. S. Narendra and L. S. Valavani,Stable adaptive controller
design-Dire.ct control, IEEE Trans. Automat. Contr., vol. AC-23,
pp. 570-583, Aug. 1978.
[3]Y.-H.Lin and K. S. Narendra, Anew error model for adaptive
systems, IEEE Trans. Automat. Contr., to be published.
[4]P. Kudva and K. S. Narendra, An identificationprocedurefor
discrete multivariable systems, IEEE Trans. Automat. Contr., vol.
AC-19, pp. 549-552, Oct. 1974.
[5] K. S. Narendra and P. Kudva,Stableadaptiveschemesfor
systems identification and control-Part I, 11, IEEE Trans. @st.,
Cybem, V O ~ .SMC-4, pp. 542-560, NOV. 1974.
[6] R. L. Carroll and D.P.
Lindorff, An adaptiveobserverfor
single-input single-output linear systems, IEEE Trans. Automat.
Contr., vol. AC-18, pp. 428-435, Oct. 1973.
[7]K. S. Narendra and Y.-H. Lin, Stable discrete adaptive control,
IEEE Trans. Automat. Contr., this issue, pp. 456-461.

Lena VaLvani was born in Thessaloniki,Greece,

on February 23,1952.Shereceivedthe
B.S.
degree in physics (cum laude) from Barnard
College, Columbia University, New York, N Y ,
in 1972, and the MS., M.Phil., and W.D. d e
grees in engineering and applied science from
Yale University, New Haven,CT, in 1974,1976,
and 1978, respectively.
She was a postdoctoral fellow in the Department of Engineering and AppliedScience at
Yale in 1978-1979 and is currently with the
Laboratory for Information and Decision Systems, Department of E k e
trical Engineering, Massachusetts Institute of Technology. Her research
interests are inthe areas of adaptive control,stability of dynamical
systems, large-scale systems, and applications of control theory to economic systems.
Dr. Valavani is a member of Sigma Xi.