OPTIMAL PROBLEM OF REGULAR COST FUNCTION FOR RETARDED SYSTEM
JONG-YEOUL PARK, JIN-MuN JEONG AND YOUNG-CHEL KWUN
ABSTRACT. We study the optimal control problem of system gov- erned by retarded functional differential
, 1°
x (t) = Aox(t)+A1x(t - h)+ -ha(s)A2x(t+s)ds+Bou(t)
in Hilbert spaceH. After the fundamental facts of retarded system and the description of condition so called a weak backward unique- ness property are established, the technically important maximal principle and the bang-bang principle are given. its corresponding linear system.
1. Introduction
In this paper we deal with the control problem for retarded func- tional differential equation:
(1.1)
~x(t)=Aox(t)+A1X(t - h)+
I:
a(s)A2x(t+s)ds+Bou(t),
(1.2) x(O) =gO, x(s) =g1(s), sE [-h,0)
in Hilbert space H. We investigate the optimization of control func- tions appearing as the cost function with particular objective.
Received March 18, 1997.
1991 Mathematics Subject Classification: Primary 49B20; Secondary 93C20.
Key words and phrases: optimal control problem, retarded functional differen- tial equation, maximal principle, bang-bang principle.
This work was supported by the Basic Science Research Institute Program, Ministry of Education, 1996, Project No. BSRI-96-1410.
(2.1)
We solve the optimization problem by introducing the structural operatorF and the transposed dual system in the sense of S. Nakagiri
[7].
In section 2, we consider some the regurality and a formular rep- resentation for functional differential equations in Hilbert spaces. We establish a form of the mild solution which is described by the integral equation in terms of fundamental solution using structural operator.
In section 3, we shall give a cost function, which is called the feedback control law for regulator problem and consider results on the existence and uniqueness of optimal control on some admissible set. After consid- ering the relation between the operatorAl and the structural operator F, we will give the condition so called a weak backward uniqueness property. Maximal principle and bang-bang principle for the given technologically important cost function are also derived.
2. Functional differential equation with time delay
Let H be a Hilbert spaces and V be a dense subspace in H. The norm on V(resp. H) will be denoted by 11 . 11 (resp. I· I) and the corresponding scalar products will be denoted by ((', ·))(resp. (', .)).
Assume that the injection of V into H is continuous. The antidual ofV is denoted by V*, and the norm ofV* by 11 . 11*. Identifying H with its antidual we may consider that H is embedded in V*. Hence we have V c H c V* densely and continuously. Ifthe operator Ao is a bounded linear operator from V to V* and generates an analytic semigroup, then it is easily seen that
H = {x EV* :
I
T IIAoetAoxlI;dt<oo},for the timeT >0 where 11 .11* is the norm of the element ofV* . .The realization ofAoinH which is the restriction ofAoto
D(Ao)= {u EV : Aou E H}
is also denoted by Ao. Therefore, in terms of the interpolation theory we can see that
(2.2) (V,V*)J.2 ,2=H
and hence we can also replace the interpolation space F in the paper G. Blasio, K. Kunisch and A. Sinestrari [2] with the space H. Hence, from now on we derive the same results of [2]. Leta(u, v) be a bounded sesquilinear form defined inV x V satisfying Garding's inequality
Let Aobe the operator associated with a sesquilinear form (Aou, v) = -a(u, v), u, v E V
Then Aogenerates an analytic semigroup in both H and V* and so the equation (1.1) and (2.2) may be considered as an equation in both Hand V*.
Let the operatorsA1andA2be a bounded linear operators fromV to V*. The functiona(·) is assume to be a real valued Holder continuous in [-h,0] and the controller operator Bo is a bounded linear operator from some Hilbert space U to H. Under these conditions, from (2.2) and Theorem 3.3 of [2] we can obtain following result.
Let Z denote the product reflexive space H x L2(-h,0;V) with the norm
The adjoint space Z* of Z is identified with the product space H x L2(0, T;V*) via duality pairing
where (.,.) denotes the duality pairing.
PROPOSITION 2.1. Let 9= (9°,91) E Z andu E L2(0,T; U). Then for each T >0, a solution x of the equation (1.1) and (1.2) belongs to
L2(0,T;V) nW1,2(0,T;V*) c C([O,T]; H).
where Wl,2(0,T; V*) denotes the SobolevspaceofV*-valued measur- able functions on (0,T) such that itselfandits distributional derivative belongto£2(0,T; V*).
Let x(t; g,u) be the solution of (1.1) and (1.2) with initial value 9= (gO,gl) EZ and controlu E £2(0,T; U). According to S. Nakagiri [7], we define the fundamental solution W(t) for (1.1) and (1.2) by
W(t) 0 = {X(t;(gO,O),O), t~O
g O t < 0
for gO EH. Here, we note that
x(t; (gO,0), 0) = G(t)+
Lt
G(t - s)(Al W(s - h)+[Oh a(-T)A2W(r+s)d-T)ds, t ~ 0
where G(t) is an analytic semigroup generated by Ao. Since we as- . sume that a(·) is Holder continuous, the fundamental solution ex- ists as seen in [11]. It is also known that W(t) is strongly contin- uous and AoW(t) and dW(t)/dt are strongly continuous except at t = nh, n = 0, 1, 2, ....
For each t > 0, we introduce the structural operator F(·) from H x £2(0,T; V) to H X £2(0,T; V*) defined by
Fg = (gO,Flgl),
Flgl = Algl (-h - s)+
l
s a(r)A2gl(r - s)dr-h
for 9= (90,f/) E Z. The solution x(t) = x(t;g,u) of (1.1) and (1.2) is represented by
.
x(t) = W(t)gO +1°
Ut(s)gl(s)ds+it
W(t - s)Bou(s)ds-h
°
where
Ut(s) = W(t - s - h)Al +iSh W(t - s+r)a(r)A2dr for t ~O.
PROPOSITION 2.2. IfAl : V --+ V* isan isomorphism, then F : Z --+ Z* isan isomorphism.
Proof For f E Z* the element 9E Z satisfyinggO = fO and gl( -h - s) +ISh a(T)Al1A2l(T - S)dT= Allfl(s)
is the unique solution of Fg = f. The integral equation mentioned above is of Volterra type, and so it can be solved by successive approx-
imation method. 0
The following result isobtained from Lemma 5.1 in [8].
LEMMA 2.1. Letf E V(O,T;H), 1:5p:5 00. If
it
W(t - s)f(s)ds= 0, 0:5t :5T,then f(t) = 0 a.e. 0 :5t :5T.
THEOREM2.1. LetAl be anisomorphism. Then the solutionx(t;g,0) is identically zero on a positivemeasure containingzeroin [-h,T] for T :2:h ifand onlyifgO = 0 andgl== O.
Proof With the change of variable and Fubini's theorem we obtain [Ohllt (S)gl(S)dS
= [OhW(t - s - h)A1gl(S)ds
+ [ :([ShW(t - s+T)a(T)A2dT)l(s)ds
= [OhW(t+s){AlX(_h,O)(s)gl(-h - s)
+J:h a(T)A2(T)gl(T - S)dT}ds
= [OhW(t+s)[Flg1](s)ds.
Thus the mild solutionx(t;g,0) is represented by
Thus, we have that x(O) = W(O)gO = gO = 0 in H. Because that Al is an isomorphism and, we obtain that PI is isomorphism from Proposition2.2. Therefore from Lemma 2.1 x(t;g,0) = 0 if and only if
gO =0 andgl ==O. 0
Let 1= [0,T], T> 0be a finite interval. We introduce the trans- posed system which is exactly same asin S. Nakagiri[8]. Let qo E X*,
qi E £1(1;H). The retarded transposed system in H is defined by·
(2.3)
dyd(t)t +Aoy(t) +Aiy(t+h)+
1°
-ha(s)A2y(t - s)ds+qi(t) =0 a.e. tEl, (2.4)
y(T) =qo, y(s) =0 a.e. sE (T, T +h].
Let W*(t) denote the adjoint of W(t). Then as proved inS. Nakagiri [8], the mild solution of(2.3) and (2.4) is defined as follows:
T
y(t) =W*(T - t)(qo) +
1
W*(~- t)qi(~)d€,for t E 1 in the weak sense. The transposed system is used to present a concrete form of the optimality conditions for control optimization problems.
COROLLARY 2.1. The solution y(t) is identicallyzero on a positive measure containingT in [T, T +h] ifand onlyifqo = 0 andqi = o.
Ifthe equation (2,3) and (2.4) satisfies the result in Corollary 2.1, the equation(2.3) and (2.4) issaid to have a weak backward uniqueness property.
3. Optimality for the regular cost function
In this section, the optimal control problem is to find a control u which minimizes the cost function
T
J(u) = (Gx(T) , X(T))H +
1
«D(t)x(t),X(t))H+(Q(t)u(t) ,U(t))u)dt wherex(·) isa solution of (1.1) and (1.2), G E B(H) isself adjoint and nonnegative, and D E B=(O,T;H, H) which is a set of all essentially bounded operators on (0,T) and Q E B=(O,T;U, U) are self adjoint and nonnegative, with Q(t) ~ m for some m > 0, for almost all t.Let us assume that there exists no admissible control which satisfies Gx(T;g,u) # 0.
THEOREM 3.1. LetUad be closed convex in L2(0, T;U). Then there exists a unique element u EUad such that
(3.1) J(u) = inf J(V).
VEUa.d
Moreover, it is holds the following inequality:
IT
(BoY(s)+Q(S)U(S),V(S) - u(s))ds ~ 0where yet) is a solution of(2.3) and (2.4) for initial condition that yeT) = Gxu(T) and yes) =
°
for s E (T, T +h] substituting qi(t) by D(t)xu(t). That is, yet) satisfies the following transposed system:(3.2)
dyd(t)t +A(jy(t)+Aiy(t+h)+
1°
-ha(s)A2y(t - s)ds+D(t)xu(t) =
°
a,e. t E I,(3.3)
yeT) = Gxu(T), yes) =
°
a.e. sE (T, T +h]inthe weak sense.
Proof. Let x(t) = x(t;9, 0). Then it holds that J(v) = 1r(v, v)
where
1r(u,v) =(Gxu(T),xv(T))H
+
I
T((D(t)xu(t),xv(t))H+ (Q(t)u(t),v(t))u)dt The form 1r(u,v) is a continuous bilinear form inL2(0,T;U) and from assumption of the positive definite of the operatorQ we have1r(v,v) ~ cllvl12 V E L2(0,T;U).
Therefore in virtue of Theorem 1.1 of Chapter 1 in [6] there exists a unique uE L2(0,T;U) such that (3.1). Ifu is an optimal control (cf.
Theorem 1.3. Chapter 1 in [6]), then
(3.4) l (u)(v - u) ~ 0 u E Uad,
where .I(u)v means the Frechet derivative of J at u, applied to v. It is easily seen that
x~(t)(v- u)= (v - u,x~(t))
= xv(t) - xu(t).
Since
l (u)(v - u)= 2(Gxu(T),xv(T) - xu(T)) + 21
T
(D(t)xu(t), xv(t) - xu(t)) + 2(Q(t)u(t), v(t) - u(t))dt, (3.4)-is equivalent to that
I
T(BoW*(T - s)(Gxu(T),v(s) - v(s))ds+T . T
1
(Bo 1
W*(t - s)D(t)xu(t)dt+Qu(s), v(s) - u(s))ds~ o.
Hence
yes) =W*(T - s)Gxu(T)+
iT
W*(t - s)D(t)xu(t)dt is solves (3.2) and (3.3).Let the admissible set Uad be
Uad = {uE L2(0,T;U): u(t) E W}
o
where W is a bounded subset in U.
COROLLARY 3.1 (MAXIMAL PRINCIPLE). Let W be bounded and Q= o. Ifu be an optimal solution for J then
max(v,AuIBoq(t)) = (u(t),AUIBoq(t))
vEW
almost everywhere in 0 :::; t leT where q(t) = -yet) and yet) is given byin Theorem 3.1.
Proof. We note that if Uad is bounded then the set of elements u E Uad such that (3.1) is a nonempty, closed and convex set in Uad.
Let t be a Lebesgue point ofu, v E Uad and t < t +€ < T. Further, put
( ) {
V, if t < s< t +€
v€ S =
u(s), otherwise.
Then Substitutingv€ forv in (3.4) and dividing the resulting inequality by €, we obtain
11
t+€- (AUIBoq(s),u(s) - v(s))ds ~O.
€ t
Thus by letting€--7 0,the proof is complete. o
Thus from Theorem 3.1 the result is obtained.
THEOREM 3.2 (BANG-BANG PRINCIPLE). Let Bobe one to one
mapping. Then the optimal control u(t)isa bang-bangcontrol, Le, u(t) satisfies u(t) E 8W for almost allt where 8W denotes the boundary
ofUad.
Proof On account of Corollary 3.1 it is enough to show that Boq(t)
=1= 0 for almost all t. If Boq(t) = 0 on a set e of positive measure containingT, then q(t) =0 for each tEe. By Corollary 2.1, we have
Gxu(T) = 0, which is a contraction. 0
From now on, we consider the case where Uad = L2(0, T;U). Let xu(t) =x(t; g, 0)+J~W(t - s)Bou(s)ds be solution of (1.1) and (1.2).
DefineT E B(H,L2(0,T;H)) and TT E B(L2(0,T;H),H) by
(T4»(t) = I t W(t - s)4>(s)ds,
TT4>= IT W(T - s)4>(s)ds.
Then we can write the cost function as (3.5)
J(u) =(G(x(T; g,0) +TTBou) , (x(T; g,0)+TTBou))H
+(D(x(·;g,O) +TBou),x(·;g,O) +TBou)p(O,T;H)
+(Qu,U)P(O,T;U)"
The adjoint operators T* and TT are given by
(T*4»(t) = iT W*(s - t)4>(s)ds, (T;4»(t) = W*(T - t)4>.
THEOREM 3.3. Let Uad = L2(0, T;U). Then there exists a unique control u such that (4.1) and
u(t) = _A-1Boy(t)
foralmostallt, where A = Q+BoT* DTB o+BoT;GT;Bo and where y( t) is a solution of (2.3) and (2.4) for initial condition that y(T) = Gx(T) and y(s) = 0 for sE (T, T +h] substituting qi(t) by Dx(t).
Proof The optimal control for J is unique solution of (3.6)
From (3.5) we have
i (u)v =o.
i (u)v =2(G(x(T; g, 0) +TTBOU) , TTBov»
+2(D(x(·; g,0) +TBou) , TBov)
+2(Qu, v)
=2«Q+BoT* DTBo+BT;GTTBo)u,v)
+2(BoT* Dx(·;g,0)+BoT;Gx(T;g,O),v).
Hence (3.6) is equivalent to that
(Au+BoT* Dx(t; g, 0)+BoT;Gx(T; g, 0), v) = 0
since A-I E B=(O,T;H,U) (see Appendix of [3]). Hence from The definitions ofT and TT it follows that
yet) = T* Dx(t; g,0)+T;Gx(T; g, 0)
= W*(T - t)Gx(T) +iT W*(s - t)Dx(t)ds.
Therefore, the proof is complete.
References
o
[1] G. Da Prato andL.Lunardi, Stabilizability of intgrodifferential parabolic equa- tions, J. Integral Eq. and Appl. 2 (1990), 281-304.
[2] G. Di Blasio, K. Kunisch and E. Sinestrari, £2-regularity for parabolic par- tial integrodifferential equations with delay in the highest-order derivative, J.
Math. Anal. Appl. 102 (1984),38-57.
[3] J. S. Gibson, The riccati integral equations for optimal control problems on Hilbert spaces, SIAM J. Control Optim. 17 (1979),537-565.
[4] J. M. Jeong, Stabilizability of retarded functional differential equation in Hilbert space, Osaka J. Math. 28 (1991), 347-365.
[5] J. M. Jeong, retarded functional differential equations with £1_ valued con- troller, Funkcial. Ekvac 36 (1993), 71-93.
[6] J. L. Lions, Optimal Control of Systems Governed by Partial Differential Equations, Springer-Verlag Berlin new-York, 1971.
[7] S. Nakagiri,Structural properties of functional differential equations in Banach spaces, Osaka J. Math. 25 (1988), 353-398.
[8] S. Nakagiri, Optimal control of linear retarded systems in Banach space, J.
Math. Anal. Appl. 120 (1986), 169-210.
[9] T. Suzuki and M. Yamamoto, Observsbility, controllability, and feedback sta- bilizability for evolution equations I, Japan J. Appl. Math. 2 (1985),211-228.
[10] H. Tanabe, Equations of Evolution, Pitman-London, 1979.
[11] H. Tanabe, Fundamental solution of differential equation with time delay in Banach space, Funkcial. Ekvac. 35 (1992), 149-177.
Jong-Yeoul Park
Department of Mathematics Pusan National University Pusan 609-739, Korea Jin-Mun Jeong
Division of Mathematical Sciences Pukyong National University Pusan 608-737, Korea
Young-Chel Kwun
Department of Mathematics Dong-A University
Pusan 604-714, Korea