SIAMJ.CONTROLANDOPTIMIZATION Vol. 31, No. 5,pp. 1167-1182, September1993
1993Society for Industrial and Applied Mathematics
006
OPTIMAL CONTROL FOR INTEGRODIFFERENTIAL
EQUATIONS
OF PARABOLIC TYPE*
GIUSEPPE DA PRATO AND AKIRA ICHIKAWA$
Abstract. Quadratic control problems for integrodifferential equationsof parabolic typeare considered. Astate-spacerepresentationofthesystemisobtainedby choosinganappropriate prod-uct space. Byusingthe standardmethod basedonRiccati equation,aunique optimalcontrolover afinite horizonandunderastabilizabilitycondition is obtainedandthe quadraticproblemover an infinite horizon is solved. It is shown that the approachis also valid for someintegrodifferential equationsofdifferenttypes. Twoexamplescoveredbythemodelaregiven.
Keywords, optimalcontrol,stabilizability, integrodifferentiMequations
AMS subject classifications. 93D15, 93C22,93C25
1. Introduction.
Let
H
andU
beHilbert spaces. Consider the controlsystem(1)
y’(t)
Ay(t)
+
f
K(t
r)y(r)dr
+
Bu(t),
where
A
istheinfinitesimalgeneratorofananalytic semigroupetAinH.We
denote byD(A)
thedomain ofA
andbyID(A)
the
graphnormofA. g(.)
is anL(D(A);
H)-valued operator, and B E
L(U;
H).
Under suitable conditions(see
Hypothesis 1below)
there exists aresolvent operator(see
[3], [13],
[16])
associatedwith(1)
and a unique classical solution to(1).
For
each ue L
2(0,
T;
U)
wecandefinea mild solution to(1)
inC([0, T];
H).
We
thenwish to minimizethe functionalT
a(u)
{IMy(t)l
2+
lu(t)l
2}
dt+
(Gy(T),y(T)}
over all u
L2(0,T;
U).
HereM
L(H;
Ho),
Ho
is a Hilbert space, andG
L+
(H)
isthespaceof selfadjoint nonnegativeoperatorsonH.
Underastabilizabilitycondition
(see
Hypothesis 4below)
we alsowishto minimizethefunctional(3)
J(u)
{IMy(t)l
2+
lu(t)l
2}
dtover all u
L2(0,
cx:);U).
To
our knowledge thereis no direct method to solve theseproblems. Inthispaperwegiveastate-spacerepresentationof
(1)
similar tothosein[15]
and[19].
As
in[9]-[11]
wethen reduceourproblemsto linearquadratic problems of standard type[1].
We
recallsome fundamental results concerningthe resolvent operator associated with(1).
It is convenientto introduce equations(4)
y’(t)
Ay(t)
+
f
K(t
r)y(r)
dr,Receivedby the editorsOctober 22, 1990; acceptedfor publication (inrevised form) January
17, 1992.
ScuolaNormMeSuperioredi Pisa,Piazza deiCavMieri, 7,56100, Pisa, Italy.
Departmentof Electrical Engineering, ShizuokaUniversity,Hamamatsu432,Japan.
1167
(5)
y’(t)
Ay(t)
/f
K(t
r)y(r)dr
+
f(t),
(0)
o,In
[6]
and[16]
the existenceofaresolventoperator for(4)
isshown underthefollowingconditions.
Hypothesis 1.
(i)
g(.) e
LI(0,
c;L(D(A); H)).
(ii)
For all hED(A),
theLaplace transformK(.)h
canbeextendedto asectorS--
{
EC"
w,arg(-w)l <
},
where wR, ]/2,[.
(iii)
There existfl
el0,
1]
and c>
0 such thatIAz(.)h]
cIhlD(A),
A
e S,
h
e
D(A).
The followingresult isproved in
[3]
and[16].
THEOREM 1.1. There exists an analytic resolvent operator
R(t)
e L(H;
D(A)),
t
O,
such that(i)
R(t)yo
is continuousfor
anyYoe H
andR(O)
I.(ii)
For
each Yoe
D(A)
andT>
O,
R(t)yo e
C([0, T];
D(A))
AC1([0,
T]; H)
andit
satisfies (4).
(iii)
For
each Yoe
D(A)
andf
e
C([O,T];H)
(a-Hhlder
continuous), y(t),
given byy(t)
R(t)yo
+
R(t
r)
f
(r)
dr,is a unique classical solution
(see
[3]),
inC([0, T];
H)
AC(]0, T]; D(A))
C
(]0,
T];
H).
(iv)
There existro
>
0 ando
e
],
]
such thatfor
anye
S withI1
ro,
I
’1
<
o, hn
opror
-
A- K()
D(A)
-
H
nvnbe nd(
A
K()))
-1e
L(H;
D(A))
coincides with the Laplacetransform of
R(t).
Foreach Y0 H andu
L2(0,
T;
U)
(6)
y(t)
R(t)yo
+
R(t
r)Bu(r)
driswell definedand is in
C([0, T]; H).
It is amild solutionof(1)
in thesensey(t)
etAyo
+
e(t-r)AK(r
s)y(s)
dsdr+
e(t-r)ABu(r)
dr.Notethat the cost function
(2)
makes sense for the mild solution.For later use we establish additional properties of
R(t)
that are not given in[3].
LetDA(c,2),
ce
]0,
1[
be the real interpolation space betweenD(A)
and H.Consider theproblem
(8)
y’(t)
Ay(t)
+
f(t),
y(O)
yo.OPTIMALCONTROLFOR INTEGRODIFFERENTIALEQUATIONS 1169 THEOREM 1.2.
(i)
Let
Yo EDA(1/2,2),
and letf
L2(0,
T; H).
Then the mild solutionof (8)
lies inL2(O,T;D(A))
NWI’2(O,T;H)
CC([O,T];DA(5,
There exists a unique solution y to
(4)
inL2(0,
T;
D(A))
W1,2(0,
T;
H).
HenceR(t)yo, R
,
f
e
L2(0,
T;
D(A))
W,2(0,
T; H).
2).
Then the mild(ii)
Letyoe
D(A),
f
e
W’2(O,T;H),
andAyo+f(O) e
DA(,
solution
of (8)
lies inWI’2(O,T;D(A))
W2’2(O,T;H)
c
CI([O,
TI;DA(1/2,2)).
Moreover,
there exists aunique solutiony to(4)
inW,2(0,
T;
D(A))
W2,2(0,
T; H).
Hence
R(t)yo,
R,f
e
W1’2(0,
T;
D(A))
W2,2(0,
T;
H).
Proof.
Thefirst assertion in(i)
iswell known[17].
Toshow thesecond assertionof
(i)
we considerthe corresponding integralequation of the type(7).
For asmall Twe apply a contraction-mapping theorem on
L2(O,T;
D(A)).
The general case then follows by splitting the interval into small subintervals. The first assertion in(ii)
isprovedas in
[15].
The second part of(ii)
follows byraisingregularity and consideringa contraction mappingon
WI’2(0,
T;
D(A)).
See[15]
for details. DTo
givea state-spacerepresentation[8],
[9]
of(1)
weconsider(9)
y’(t)
Ay(t)
+
K(t-
r)y(r)
dr,y(O)
y(),
e]
cx),0[,
ye
L2(-,
0;D(A)).
Wenowrewrite this as(10)
y’(t)
Ay(t)
+
K(t-
r)y(r)
dr+
f(t),
where
f(t)
fo
K(t
O)y
()
dOe
L2(0,
;
H).
Hypothesis2.
K(.)
e
L2(0,
cx;L(D(A);
H)).
Ifwe assume Hypothesis 2 istrue, thentheoperator]C definedby
(11)
]CyK(-O)y
(O)dO,
ye L
2(-,
0;D(A))
lies in
L(L2(-cx,
0;D(A)),
H).
(-oc,
0;D(A))
sinceMoreover,
f
W,2(O, T;
H)
for any yW
1,2f’(t)
K(t)y(O)
+
K(t-O)y’l(O)dO
e
L2(O,T;H).
Note alsof(0)
following corollary.
Using these observations and Theorem 1.2, we have the
COROLLARY
1.3.(i)
For each Yo EDA(1/2;2)
andYlL2(-oc,
O;D(A))
thereexists a unique solution to
(10) (and
hence to(9))
in the spaceL2(O,T;D(A))
WI,2(O,T;H).
(ii)
Assume
Hypothesis 2, and let yle
W’2(-oc,
0;D(A)),
y(O)
Yo, andAyo
+
]y
DA(
1/2
2).
Then there exists aunique solution to(9)
inW’2(-oc,
O;D(A))
NW2’2(O,T;H)
CCI([O,T];DA(1/2;2)).
Proof.
Part
(i)
follows directly from Theorem1.2(i).
Under Hypothesis 2f
W,2(0,T;
H)
and the assumptions in(ii)
of Theorem 1.2 are satisfied. Hence(10)
hasaunique solution inW1,2(0,
T;
D(A))
NW2’2(0,
T; U).
Sincey(0)
YoD(A),
thereexists aunique solution to(5)
inW1,2(-c,
T;
D(A)).
DNow
wewrite(9)
in theform(12)
y’(t)
Ay(t)
/K(-O)y(t
/)
dO,u(0)
uo,y(O)
Yl(0),
0e]
--(:X),0[.
Thisis adelayequation with infinitedelay.
A
more generaldelayequation, butwith finitedelay, was considered in[15],
anda semigroupwasconstructedontheproductspace
DA(1/2;
2)
n2(-r,
0;D(A)).
ToobtainasimilarresultweassumeHypothesis 2andrewrite
(12)
asy’(t)
Ay(t)
+
t:yt,
(13)
y(0)
Yo,Y(O)
Yl(O),
e]
--CX:),0[.
where
yt(’)
y(t
+
.).
The following result is a modification of[15,
Thms. 4.1 and4.2]
forthe casewithinfinitedelay.THEOREM 1.4.
Assume
Hypotheses 1 and 2 aretrue, and let y be theuniquesolu-tion
of(12)
foryoe
DA(-12;
2)
andyle
L2(-oc,
O;D(A)),
which lies inL2(O,T;
D(A))
NW1,2(0,
T;
H) for
anyT>
O. Then the map(14)
S(t).
(
yo)
(y(t)
Yl
Yt(’)
)
on
Z
DA(
1/2
2)
xL2(-oc,
0;D(A))
is astronglycontinuoussemigroup.Its
infinites-imalgeneratoris given by
(5)
y
a_
dO(16)
D(.4)
y(O)
{(y0,y) e
Z:yle
W1,2(-oc,
O;D(A)),
yo,
Ayo
+
t:y e
DA(1/2;
2)}
OPTIMAL CONTROL FOR INTEGRODIFFERENTIALEQUATIONS 1171
Proof.
The only difference between(13)
and the equation in[15]
isthelength ofthememory involved.
Hence
one couldrepeat the proofs in[15].
However,
we shallgive a different proof for the characterization ofthe generator.
Note
first that thestrong continuity and the semigroup property of
S(t)
follow from Corollary1.3(i).
We
now show that4,
given by(15)
and(16)
is the infinitesimal generator of thesemigroup
_S(t).
Choose[Y0,
yl]’
e
D(,4);
then_S(t)[y0, yl]’= [y(t), yt(’)],
wherey(t)
isthesolutionof(13)
and henceof(9).
Then byCorollary 1.3(ii)
wehavelimY(t)-Y
(
)
t-o t
y’(O)
Ayo
+
Yl
inDA
;2limYt
(’)
Y(’)
dy__1
t--,o t dO in L
2
(-oc,
0;D(A)).
This implies that theinfinitesimal generator ofthesemigroup
_S_S(t)
coincides with J[ onD(J()
andisan extensionof,4.To
seethat4
isin factthe generatorweneed toshowonly that the resolvent set of
4
isnonempty. NowchooseA
>
0and considerThis isequivalent to
(/
4)[y0,
Yl]’
[Z0, Zl]
e
Z.AYo
Ahy
zo
EDA
(1/2;
2),
dylAy
dOz
5 2(-oc,
0;D(A)).
The second equationyields
/o
Yl
(0)
eAOyl (0)
A- eA(O-)z(r/)
dr/.
Setting
yl(O)
Yo and substitutingyl intothe first equation, weobtain/o
(,x
A-
Ce’)o
zo
+
e(-’z(v)
dv
=:o.
Noting
that/Ce’o
’(A)o,
wehave(-
4- g(a))o
o.
By
virtue of Theorem1.1(iv)
this is solvable for anyA
> ro
ando
(-
A-K(A))-Io
D(A).
Then/o
Yl
(0)
eA0(A-
.4-/(/))--1
0-[-e)(0--v/)Z
l(r/)
dr/
liesinW,2(-c,
0;D(A)).
We
also haveAyo
+
1Cy Ay0zo
DA(
1/2
2).
Hence A
p(J[)
and ,4isthe infinitesimalgenerator of the semigroup_S(t).
[:]Remark 1.5.
Let
A
A0
/A1,
whereA0
is selfadjoint and negative. IfA1
EL(D(-A)/2;H),
then we can replaceDA(1/2;2)
byD(-A)
/2 andDA(--1/2;2)
byD(-A*)
1/2(see 2).
In
[14]
aspecialcaseofthedelayequation in[15]
wasconsidered and aquadraticcontrolproblem on
DA(1/2;
2)
xL2(-r,
0;D(A))
wassolved.Ifwe take
B
e
L(U;DA(1/2;2)),
M e
L(DA(
-2;
2),H0),
andG
e
L+(DA(1/2;2)),
thenbyusing the semigroup
__S(t)
inTheorem 1.4 wecansolveourcontrolproblemas in[14];
however, thestate spaceZ
isnot convenient in applications, andwe wishtotake theinitialvalueY0 in
H
rather than inDA
(1/2;2).
Moreover,
ourcost functionals(2)
or(3)
aremore natural,aswecan seefrom examples(see
Example5.1).
Thusweneeda representation ofoursystem
(1)
in alarger space.2. The semigroup model. Let
DA(--c,2),
c]0,1[,
be the extrapolationspace of
A
(see
[2]).
To take Y0 inH
rather than inDA(1/2;
2)
we replace H (re-spectively,D(A))
byDA(--1/2;2)
(respectively,DA(1/2;2))
and assume, in addition toHypothesis 1, the following hypothesis.
Hypothesis 3.
K(.)
L2(O,
oc);
L(DA(-1/2;
2);
DA(1/2;
2)).
Then theoperator
K
in(11)
belongstoL(L2(-oc,
0;DA(-;
2));
DA(---;
2)).
By
translationweobtain all resultssimilarto thosein1.
In
particular, we state results correspondingto Corollary 1.3 and Theorem 1.4, respectively.THEOREM 2.1.
(i)
For each Yo H and y L2(-oo,
0;On
(-12;
2))
there exists a unique solution to(10)
in1,2
2))
CC([0 T]
H).
L2(O,T;DA(;2))
CW
(O,T;DA(
1/2;
(ii)
Assume
Hypothesis 3, and let ye
WI’2(--oo,
O;DA(1/2;2)),
y(O)
Yo, andAyo
+
yl H. Then there exists a unique solution to(9)
inW’2(-oo,
T;DA(1/2;2))
3W2’2(O,T;DA(-1/2;2))
CCI([O,T];H).
THEOREM 2.2.
Assume
Hypothesis 1 and3 withH (respectively
D(A))
replaced2)).
Lete
H,
lete
L2(-c,
0DA(5;
2)),
and byDA(--1/2;2)
(respectively
DA(;
Yo Yllet
y(t)
be thesolutionof
(9)
(and
henceof
(13))
given in Theorem2.1(i). Define
the map on Z UxL2(-c,
0;DA(1/2;2))
(17)
S(t).
(yo)__,
(y(t)
)"
Then
S(t)
is astronglycontinuoussemigroup onZ,
anditsinfinitesimal
generatoris given by(18)
4(
yO(
AyO--Kyl)
dO(19)
D(A)
{
(yo,
ye
Z yle W
’2(-oc,
0;DA
(1/2;
2)),
y(0)
Y0,Ayo
+
Ky
H).
1173
OPTIMALCONTROLFORINTEGRODIFFERENTIAL EQUATIONS
Next
we expressS(t)
by usingthe resolvent operator.We
write(9)
as(20)
y’(t)
Ay(t)
+
K(t-
r)y(r)
dr+
K1
(t)yl,
where
2
(
(1))(
gl(t)yl
g(t-O)yl(O)dO e
L20,c;DA
--;2
NC
[O,T];DA
-;2
Thesolutionof(20)
can bewrittenas(t)
R(t)o
+
R(t
r)K1
(r)yl
dr Set(t)"-(11(t)
SI
(t)
S(t)
theny(t)
Sll(t)yo
-
Sl2(t)yl.
Thuswe haveSl
(t)o
R(t)o,
Similarly,we have
’12(t)yl
R(t
r)K1
(r)yl
dr.(s(t)0)(.)
R(t
+
.)o,
(S22(t)yl)(’)
R(t-
r)K(r)yl
dr.Hypotheses 1and 3 comefrom physical examples suchasExample5.1. Ifweassume,
insteadof Hypothesis 3, the following hypothesis, wehave, infact, Corollary 2.3.
2))).
Hypothesis 3’.g(.) e
L2(O,
oc;L(DA(1/2;2);H)))
NL2(O, cx;L(H;DA(-;
then wecan findasemigroup onH xL2(-cx),
0;H).
COROLLARY 2.3.
Assume
Hypothesis3’.
Then.for
each Yo EH
and yln2(-c,
0;H)
there exists a unique solutiony(t)
to(13)
in.for
anyT>
O. The map(21)
2))
C([0,
T] H)
L2(O,T;DA(;
1
(.1
)
is a strongly continuous semigroup on Z H x
L2(-oc,
0;H).
generatoris given by
(22)
Yl
a_
dOIts
infinitesimal
(23)
D(A)
((Y0,
Yl)
e
Z:Yl (wl’2(-oo,
O;H),yi(O)
yo,Ayo
/Kyle
H}.
If
K(.)
EL2(0,
c:H),
thenA
neednot be analytic. Hypothesis3".K(.)
e
L2(0,
oc;L(H)).
COROLLARY 2.4. Let
A
be anyinfinitesimal
generatorof
a strongly continu-ous semi-group on H.Assume
Hypothesis 3.
Thenfor
each yo H and ylL2(-cx),
0;H)
there existsaunique solutiony(t)
to(13)
inC([0, T];
H)
for
anyT>
O.Define
the mapS(t)
as in(21).
Then it is a strongly continuous semigroup onZ
H
L2(-oc,
0;H)
withgenerator(22),
(23).
See
[12]
for more general cases of integrodifferential operators whereA
is notanalytic.
3. Quadratic control on finite horizon.
Now
weconsider(13)
with controly’(t)
Ay(t)
+
Kyt
+
Bu(t)
(ea)
(0)
o
()
(),
e]
-,
0],
where Yl
e
L2(-oc,
O;DA(1/2,2)).
Then bysettingz(t)=
[y(t),yt(’)]’
weobtain(25)
z(0)
[0,],
where
Foreach u
e
L2(0,
T;
U)
wedefinethe mild solutionof(25)
by(26)
z(t)
S(t)[yo,
y]’
-t-S(t
r)Bu(r)
dr.The mild solution
(6)
of(1)
correspondsto thefirst component ofz(t)
ofthe specialcase Yl 0, i.e.,
(27)
z(t)
S(t)[yo,
0]
+
S(t
r)Bu(r)
dr. The cost functional(2)
can be rewritten(28)
TJ(u)
[]/z(t)l
2+
lu(t)l
2]
dt+
((z(T),
z(T)),
where M-0e
L(Z;
Ho)
and 0 0e
(z).
OPTIMAL CONTROL FOR INTEGRODIFFERENTIALEQUATIONS 1175
The controlproblem
(26),
(28)
is astandard quadraticproblem[1]
inthe state-space form[8],
[9].
As
iswellknown, the optimalcontrol is given by thefeedback law(29)
_u-B*Q(t)z(t),
where
Q
istheuniqueselfadjoint nonnegative solutionoftheRiccati equation(3o)
Q’+A*Q+QA+*-Q*Q=O,
Q(T)
G.
Setting(t)
Q(t)
Q21(t)
we canwrite
(29)
inthe formQ(t)
)
Q22(t)
(31)
u_(t)
-B*[Qll
(t)y(t)
+
Q2(t)yt].
The minimal cost correspondingto uis
(32)
j(_u)
(Q(O)(
Y0)yl
(
Y0)).yl
Hence theminimalcost for the problem
(1),
(2)
is givenby(33)
J(u_)
(Q(O)yo, yo).
Summingup, wehave the followingtheorem.
THEOREM 3.1.
Assume
Hypothesis 1 and3 are true. Thenthere exists a uniqueoptimal control
.for
the problem(1),
(2).
It isgiven by thefeedback
law(31),
and theminimal cost is given by
(33).
For the control problem
(1),
(2),
whereK(.)
satisfies either Hypothesis 3’ orHypothesis
3",
thefeedbacklaw(31)
isstilloptimalandthe optimal cost isgivenby(33).
4.
uadratic
controloninfinite horizon. Hereweconsiderthe controlprob-lem
(1),
(3).
Toavoid the trivial casewemake the followingassumption for(27).
Hypothesis4. ForeachY0 EHand y EL2(-(x),
0;DA(1/2,2)
there existsacontrolu
L2(0,
cx);U)
suchthat(34)
J(u)
Irz(t)l
2+
lu(t)l
2 dt<
cx.Laterwegive sufficient conditionsfor Hypothesis 4. Let
QT(t)
be the solution ofthe Riccati equation
(30)
withQT(T)
0. Thenthe followingisknown.PROPOSITION 4.1.
Assume
Hypotheses 3 and 4 are true. Then there exists astrong limit
Qo
of
QT. Q
is the minimal nonnegative solutionof
the algebraicRiccati equation
(35)
A*Q
+
QA
+
M*M
QBB*Q
O.If
Jr,
M is detectable, thenQ
is theunique nonnegative solutionof
(35). Moreover,
A-
BB*Q generates an exponentially stable semigroup onZ.THEOREM 4.2.
Assume
Hypotheses3 and4 aretrue. Then there exists a uniqueoptimal control
.for (26),
(34).
It is given by thefeedback
law(36)
__u-B*Qz(t),
and the minimal cost is
(37)
J(-)-"
IQcx)(
ylY
)
(yOyl)1"
In
particular, the optimal controlfor
theproblem(1),
(2)
is given by(38)
u_(t)
-B*[Q11y(t)
+
Q12yt]
and
(39)
J(_u)
(Qllyo, yo),
where
Qll Q12
)
Q=
Q:
Q.:
If
Q12(y)
fo
Q12(-0)y(0)dO
forsomeQ2e
L2(0,
oc;L(DA(1/2,2);
H)),
thenthe optimalclosed-loop system correspondingto
(38)
is(40)
y’(t)
(A
BB*Q)y(t)
+
[K(t
r)
BB*Q12(t
r)]y(r)
drandofthe same form as
(1).
If(jr,M)
isdetectable, thenthe resolvent operator for(40)
isexponentially stable.A
sufficient condition for Hypothesis 4 can be found in[7].
For the sake of completenesswequotesomeresultsfrom[7].
LetK(.)
beamaximalanalyticextensionof the Laplace transform of
K,
and let gt0 be its domainofdefinition.We
setp0
{A
e
gt" 3(A-
A-
(,))-1},
F(A)
(A-
A-
())-1
forA e
P0,and wedenote by pl theset ofall isolated removable singularity of
F(.).
Moreover,
weset
p--poUpi,
F(A)
limz_,F(z),
e p\po.
Definethe generalized spectrum a
C\p.
If0
is apole ofF(.)
oforderm0, weset,for/k sufficientlyclose to
A0,
F(A)
Sn(A-
A0)
n+
Q(A-
A0)
--,
n=0 n=0
OPTIMALCONTROL FORINTEGRODIFFERENTIAL EQUATIONS 1177 where
and
C(A0,
e)
isthecirclewith center ,0 having sufficiently smallradius>
0. Letw>
0be such that a;3{,
E C:aezk-w}
q},
and leta+(w)
a3{
C
Re,>
-w},
a_(w)
aN{,
C
ReA
<
-w}.
We can make the followingassumption.
Hypothesis 5.
(i)
a+(w)
{,1,...
,,g},
where foreachj 1,...,g,
Aj is apole ofF(.)
of order mj<
oo.(ii)
TheresiduesRi,k,
kO,
1,...,mj,ofF(-)
atA
,
arefinite-rankoperators.The conditionbelow iscalled a Hautuscondition.
Hypothesis 6.
Range
Q,k
3 Ker B*{0}
for all j 1,2,...,N and k 0,1,...,m.
Let
X
beaBanach space, and letC,([0, oo[; X)
bethe space of bounded contin-uous functionsx(t)
in Xwith propertysupt>0IIx(t)etllx
<
+oo.
Under Hypothesis5it isshown
[7]
that Hypothesis6holdsifand onlyif thefollowing istrue:For
eachY0e
Hthereexistsacontrolue
C([0, oo[; U)
(in
fact,ue
C([0, oo[; U))
such that ye
C,([0, oo[;
H),
where yisthesolutionof(1).
Hence
ifHypotheses 5 and6hold,then thecontrolproblem
(1),
(3)
iswell defined. Modifying slightlytheproof of[7,
Thm.2.3],
we can show that under Hypothesis 6 system(13)
is stabilizable in the above sense.We
have, in fact, the following result, the proofof which wassuggestedto us by
A.
Lunardi.THEOREM 4.3.
If
Hypothesis 6 holds, then the system(13)
is stabilizable, i.e.,for
eachYoe
H there exists a controlte
C([0, oo[;
U)
such thatye
C([0, oo[;
H)
for
somew]0, w0[.
Proof.
DefineR()
/
(,) mj--1tt
ketF(d)
eJ
k---.
Qj,k
k--O and NRT0
j--1Then as in
[7,
Prop.
1.1]
we can show that problem(13)
isstabilizable in the abovesense if and only if for each Y0 E
H
and ylL2(-,
0,DA(1/2,2))
there exists uC([0,
c[; U)
suchthat(41)R
(t)yo
+
R
’+
(t
s)K
(s)y
dsn+"
(t
s)Bu(s)
ds, t>_O, whereK(.)
is as given in2.
First, weassumeRe
Aj>
0, j 1,2,...,N,
andshow(41)
withw w0. If thereexists,kj withRe
,
0,thenwecansetv(t)
cRy(t)
for sufficientlysmall>
0 and reduce theproblemto thecase for which w w0.
As
in theproof ofTheorem2.3 in[7],
wecan show that(41)
isequivalent toNote
that thesecond term is well defined sinceKI(.)yl
is bounded and ReAy
>
0.Since the functions t
eJtt
"
are linearly independent,(42)
is equivalent toFu
Q[yo,
gl(.)y],
whereN F"
C([0,
+[;
U)
H
K,
KE
my,
j--1Fu
e-
8(--s)k-’Qy,k
Bu(s)ds
k--n andQ"
H--
HK,
Q(yo, K(.))y)
Qi,y
+
,=
e(--s)k-nQy,kKl(S)ylds
j-1 N;n---O m.-ISincethe range ofFand
Q
arefinite-dimensional,(42)
holds ifandonlyif(43)
KerQ*
Ker F*.Foreach
(hjn)
(hjn)y=l
N;n=O m-Ie
HK wehave Nmj-lm-l-kfOoT(X
(_8)
kF*(hn)U
-k!e-
<
u(s)’B*Q;’+hh’+
>
ds j=l k=O h=O--
Q*(hjn)(yo,
gl(’))y)
N mj-1(Yo,
O;hhh}
j--1 h----O N m.-1m.-l-k/EE
E
j--1 k--0 h--0 +(K1
(8)y
Q,k+nhy,k+n)ds
(--)
k! so thatKerF*
(hyn)
e
HK" B*Qhhyn
j 1,...,N,
kO,
1,...,my
1\ h--k
{(hy,)
e
HK B*Qyhhy,,j
* 1,...,N,k0,1,...,my
1},
OPTIMAL CONTROL FOR INTEGRODIFFERENTIAL EQUATIONS 1179
N mi-1
ger
Q*
D{(hj)
e
HE"
E E
Q
* h0}.
j-1 h-0
Now
wecan showthat(13)
isstabilizable andis equivalentto(43).
It
is easyto seethat if
(13)
is stabilizable, then(43)
holds. Conversely, assume that(23)
holds and * * for some jo,ho.
Set hjn
5j,joSn,noh;
thenB*Qhjn
0 for let h EKer
B
Qjoho
eachj 1,...,N,h
0,...,m
1.By
(43)
wehave N mj-1O;o o
Q;o o
o.
j=l h--O
Therefore, for eachj0,
h0
wehaveKerQjono
BQjono,
and(13)
isstabilizable. F1Nowconsider the detectability of
,4,
M. It isusefulto considerthe following:’
(t)
-A*(t)
ft
TK*(r
t)(r)
dr,(44)
v_(T)
Suppose
wehaveclassical solutionsfor(4)
and(44).
Then by differentiating(y(t),
_(t)},
integrating from t 0 to
T.,
and usingthe Fubinitheorem weobtain(y(T),
r]l(yo,
q_(O) ).
Hence
(44)
isthe adjoint system of(4).
Equations(44)
canbe also written’
(t)
A*r(t)
+
f
K*(t
r)y(r)
dr,Thusthe detectability of
(jr,
M)
istranslatedintothe stabilizability of the following system:(46)
’(t)
A*(t)
+
K*(t-
r)(r)
dr+
My(t).
If A* and K* have properties similar to those of
A
andK
we can obtain sufficient conditions for detectability. Finally, wenote that we canalso solve control problems for inhomogeneous systems asin[4],
and[5].
Example 4.4.
Let
be a bounded open domain inR
withC
boundary Consider the heat equation in materials of the fading-memory type introduced byNunziato
[20]:
(47)
O
yt x
+
cO
oo
bo-
-
(t
r)y(r,
x)
drcoAy(t,
x)
7(t
r)Ay(r,
x)
dr+
Bou(t, x),
t>
O,
xe
it,e
Fy(t,x)=O,
t>O,
xcOgt,
where
Fy
yorFy
Oy/On. y(t,
x)
representsthetemperature at x E 12 at timet, bo, andco
are positive constants, and u is theheat supply. and"
are completelymonotonekernelswith
Z(t)
-,(d),
(t)
-(d),
where #and uare positive Borel measures with compact supportsupp#andsupp u contained in
]a, oc[,
a>
0. Then the heat equation(47)
can be written as(1)
inH
L2(12)
with 1 Ah7-(coAh -/3(0)h),
o0D(A)
{h
e
H" Ahe
H,
Fh0},
1K(t)h
7-(-’(t)h
-(t)Ah),
oo
1B
0B0.
Let
(.)
and(.)
be analytic extensions of the Laplace transforms of and-
toC\supp
#andC\supp
,
respectively. Thenpo
e
C.co
-() #
0,x(0
+
(x))
0-()
#-F()
bo
(o+2())
co-()R(
(b
+/())
A)
0()
where
{-An }
isthe decreasingsequenceof the eigenvalues ofA.
Hence{"
ReA>0}={
0
ifFy=y
on Oft,{0}
ifFy=
O0n
on 0a.If
Fy
y, then(47)
with u 0 is stable. Hence our control problems(1),
(2)
and
(1),
(3)
are well defined. IfFy
n’
then(47)
with u 0 is not stable, butsatisfies Hypothesis 5. This implies that if
h(x)
ho (constant
different from0)
Ker B*,
then(47)
is stabilizable. IfBub(x)u(t),
b(.)
H,
and u is a scalar,then
fa
b(x)dx #
0implies stabilizability. Hencethe quadraticproblems(1),
(3)
and, of course,(1),
(2)
are well defined. See[6]
for more discussionsonthis example andother examples of system
(1).
See
also[19]
for heat equations withmemory.Thefollowing examplewasintroduced tousby
J.
Zabczykand is covered byourmodel.
Example 4.5. Considerthe delayequation
(48)
x" (t)
-k(O)x(t)
fk’
(-r)x(t
+
r)dr
+
u(t),
x(O)
xo,’(o)
OPTIMAL CONTROL FOR INTEGRODIFFERENTIALEQUATIONS 1181
which can beregardedasamodel ofthe oscillationofaparticlesuspendedbylight
lin-earlyviscoelastic string,where therelaxationmodulus k"
[0, [--,
R
isdifferentiableand suchthat
k(t)>O,
k’(t)<0,
k"(t)>0,
limk’(t)=0.
Setting y x wecanwrite(48)
as 0-k(0)
0k(-r)
01
0)
0x(t
+
r)dr
+
(0)
1(xo)
y(O)
Yo xlIf k E
L2(0,
cx),
thenthisis aspecialcasecoveredbyTheorem 4.2. Ifk(t)
e-at,
a>
0, the spectrum a isgivenby
a
{"
3
+
aA2
+
k(O)A
+
ak(O)-
10}.
Set(
01)
A=
-k(0)
0 1M=[1,0];
then
(A,
B)
iscontrollableand(M,
A)
isobservable. Hencethe minimizationproblemJ(u)
[Ix(t)l
2+
lu(t)l
2]
dt iswell defined.REFERENCES
[1] R. F. CURTAIN AND A. J. PRITCHARD (1978), InfiniteDimensional LinearSystems Theory,
LectureNotesin Control andInformationSciences, Springer-Verlag,Berlin.
[2] G. DAPRATOAND P. GRISVARD (1984),Maximalregularityforevolution equationsby inter-polation and extrainter-polation, J.Funct.Anal.,58,pp. 107-124.
[3] G. DA PRATO AND M. IANNELL (1985), Existence andregularityfora classof integrodiffer-ential equations ofparabolictype, J.Math. Anal.Appl., 112,pp. 35-55.
[4] G. DA PRATO AND A. ICHIKAWA (1988), Optimal controlfor linear periodic systems, Appl. Math.Optim., 18,pp. 39-66.
[5] (1990), Quadraticcontrolforlineartimevaryingsystems,SIAM J.ControlOptim.,28, pp. 359-381.
[6] G. DAPRATOANDA.LUNARDI (1988),Solvabilityonthereal lineofaclassoflinearVolterra integrodifferential equationsofparabolictype, Ann. Mat.Pura Appl. (4), 150, pp. 67-118. [7] (1990), Stabilizability ofintegrodifferential parabolic equations, J. Integral Equations
Appl., 2,pp. 281-304.
[8] M. C.DELFOUR(1986), Thelinear-quadratic optimalcontrolproblemwithdelaysinstate and control variables: astate space approach,SIAMJ. ControlOptim., 24,pp. 835-883.
[9] M. C. DELFOUR AND J. KARRAKCHOU (1987), State space theory oflinear time invariant systemswithdelaysinstate,control and observationvariables, J. Math. Anal. Appl.,125, pp. 361-450.
[10] M. C. DELFOUR, C. MCCALLA, AND S. K. MITTER (1975), Stability and infinite-time
quadratic costproblemforlinearhereditarydifferentialsystems,SIAM J.Control Optim., 13,pp. 48-88.
[11] M.C. DELFOURAND S.g. MITTER(1974),Controllability, observabilityand optimalfeedback controlofhereditary dierentialsystems,SIAM J. Control Optim., 10,pp. 298-328. [12] W. DESCH, R. (. GRIMMER, AND W. SCHAPPACHER (1986), Wellposedness andwave
prop-agationfora class ofintegrodierentialequations, Report 82, InstitutefiirMathematik, UniversittGraz, Graz, Austria.
[13] W. DESCH AND W. SCHAPPACHER (1985), A semigroup approachto integrodifferential equa-tionsin Banach spaces,J.Integral Equations, 10,pp. 99-110.
[14] G. DI BLASIO(1981), The linear quadratic optimalcontrolproblemfordelaydierential equa-tions,Rend.Acc. Naz.Lincei, 71,pp.156-161.
[15] G. DIBLASIO, KUNISCH, AND E. SINESTRARI (1984), L2-regularityforparabolic partial inte-grodiferential equations withdelayinthe highest order derivative, J.Math. Anal.Appl., 102, pp. 38-57.
[16] A. LUNARDI (1985), Laplace transform methods in integrodiferential equations, J. Integral Equations,10,pp. 185-211.
[17] J. L. LIONSAND E. MAGENES (1968), Problemes auxlimitesnonhomogenes etapplications,
Dunod,Paris.
[18] R. K. MILLER (1974), Linear Volterra integrodierential equations as semigroups, Funkcial.
Ekvac., 17,pp.39-55.
[19] (1979),Anintegrodifferentialequationforrigid heatconductors with memory,J.Math. Anal. Appl., 66,pp. 313-332.
[20] J. W. NUNZIATO (1971), Onheatconductioninmaterials with memory,Quart. Appl.Math.,
29,pp. 187-304.