ISSN: 1072-6691. URL: http://ejde.math.txstate.edu or http://ejde.math.unt.edu
OPTIMAL CONTROL FOR SYSTEMS GOVERNED BY PARABOLIC EQUATIONS WITHOUT INITIAL CONDITIONS
WITH CONTROLS IN THE COEFFICIENTS
MYKOLA BOKALO, ANDRII TSEBENKO Communicated by Suzanne Lenhart
Abstract. We consider an optimal control problem for systems described by a Fourier problem for parabolic equations. We prove the existence of solutions, and obtain necessary conditions of the optimal control in the case of final observation when the control functions occur in the coefficients.
1. Introduction
Optimal control of determined systems governed by partial differential equations (PDEs) is currently of much interest. Optimal control problems for PDEs are most completely studied for the case in which the control functions occur either on the right-hand sides of the state equations, or the boundary or initial conditions. So far, problems in which control functions occur in the coefficients of the state equations are less studied. A simple model of such type problem is the following.
Let Ω be a bounded domain in Rn with piecewise smooth boundary Γ, T >0, Q:= Ω×(0, T), Σ := ∂Ω×(0, T). A state of controlled system for given control v ∈ U := L∞(Q) is defined by a weak solution y =y(v) = y(x, t;v), (x, t) ∈ Q, from the spaceL2(0, T;H01(Ω))∩C([0, T];L2(Ω)), of the problem
yt−∆y+vy=f ∈L2(Q), y
Σ= 0, y
t=0=y0∈L2(Ω).
The cost functional is
J(v) :=ky(·, T;v)−z0(·)k2L2(Ω)+µkvk2L∞(Q) ∀v∈U,
whereµ >0,z0∈L2(Ω) are given. An optimal control problem is to find a function u∈U∂ :=
v∈U :v≥0 a. e. onQ such that J(u) = inf
v∈U∂J(v).
This problem is nonlinear, since the dependence between the state and the control is nonlinear.
The direct generalization of this problem is given as only one among many other problems which were considered in monograph [19]. Other various generalizations
2010Mathematics Subject Classification. 35K10, 49J20, 58D25.
Key words and phrases. Optimal control; problems without initial conditions;
evolution equation.
c
2017 Texas State University.
Submitted December 7, 2016. Published March 14, 2017.
1
of this problem were investigated in many papers, including [1, 2, 4, 5, 9, 11, 15, 17, 21, 22, 25, 26] where the state of controlled system is described by the initial- boundary value problems for parabolic equations.
In [1, 22, 25, 26] the state of controlled system is described by linear parabolic equations and systems, while in [1, 22] control functions appears as coefficients at lower derivatives, and in [25, 26] the control functions are coefficients at higher derivatives. In [22] the existence and uniqueness of optimal control in the case of final observation was shown and a necessary optimality condition in the form of the generalized rule of Lagrange multipliers was obtained. In [1] the authors proved the existence of at least one optimal control for system governed by a system of general parabolic equations with degenerate discontinuous parabolicity coefficient.
In papers [25, 26] the authors consider cost function in general form, and as special case it includes different kinds of specific practical optimization problems. The well-posedness of the problem statement is investigated and a necessary optimality condition in the form of the generalized principle of Lagrange multiplies is estab- lished in this papers.
In [2, 9, 11, 15, 17, 21] the authors investigate optimal control of systems gov- erned by nonlinear PDEs. In particular, in [2] the problem of allocating resources to maximize the net benefit in the conservation of a single species is studied. The pop- ulation model is an equation with density dependent growth and spatial-temporal resource control coefficient. The existence of an optimal control and the uniqueness and the characterization of the optimal control are established. Numerical simu- lations illustrate several cases with Dirichlet and Neumann boundary conditions.
In [9] the problem of optimal control of a Kirchhoff plate is considered. A bilinear control is used as a force to make the plate close to a desired profile taking into the account, a quadratic cost of control. The authors prove the existence of an optimal control and characterize it uniquely through the solution of an optimal- ity system. In [12] the optimal control problem is converted to an optimization problem which is solved using a penalty function technique. The existence and uniqueness theorems are investigated. The derivation of formula for the gradient of the modified function is explained by solving the adjoint problem. Paper [17]
presents analytical and numerical solutions of an optimal control problem for quasi- linear parabolic equations. The existence and uniqueness of the solution are shown.
The derivation of formula for the gradient of the modified cost function by solving the conjugated boundary value problem is explained. In [18] the authors consider the optimal control of the degenerate parabolic equation governing a diffusive pop- ulation with logistic growth terms. The optimal control is characterized in terms of the solution of the optimality system, which is the state equation coupled with the adjoint equation. Uniqueness for the solutions of the optimality system is valid for a sufficiently small time interval due to the opposite time orientations of the two equations involved. In [21] optimal control for semilinear parabolic equations without Cesari-type conditions is investigated.
In this article, we study an optimal control problem (see (4.1), (3.2), (3.4), (4.2), (4.3) below) for systems whose states are described by problems without initial conditions or, other words, Fourier problems for parabolic equations. The model example of considered optimal control problem is a problem which differs from the previous one (see beginning of this section) by the following facts: the initial moment is−∞and, correspondingly, the state equation and control functions are
considered in the domainQ= Ω×(−∞, T), a boundary condition is given on the surface Σ =∂Ω×(−∞, T), while the initial condition is replaced by the condition
t→−∞lim ky(·, t)kL2(Ω)= 0.
The problem without initial conditions for evolution equations describes pro- cesses that started a long time ago and initial conditions do not affect on them in the actual time moment. Such problem were investigated in the works of many mathematicians (see [3, 7, 24] and bibliography there).
As we know among numerous works devoted to the optimal control problems for PDEs, only in [4, 5] the state of controlled system is described by the solution of Fourier problem for parabolic equations. In the current paper, unlike the above two, we consider optimal control problem in case when the control functions occur in the coefficients of the state equation. The main result of this paper is existence of the solution of this problem.
The outline of this article is as follows. In Section 2, we give notation, definitions of function spaces and auxiliary results. In Section 3, we prove existence and uniqueness of the solutions for the state equations. Furthermore, we construct a priori estimates for the weak solutions of the state equations. In Section 4, we formulate the optimal control problem. Finally, the existence and necessary conditions of the optimal control are presented in Section 5.
2. Preliminaries
Let n be a natural number, Rn be the linear space of ordered collections x= (x1, . . . , xn) of real numbers with the norm|x|:= (|x1|2+. . .+|xn|2)1/2. Suppose that Ω is a bounded domain in Rn with piecewise smooth boundary Γ. SetS :=
(−∞,0],Q:= Ω×S, Σ := Γ×S.
Denote byL∞loc(Q) the linear space of measurable functions onQsuch that their restrictions to any bounded measurable setQ0 ⊂Qbelong to the spaceL∞(Q0).
LetX be an arbitrary Hilbert space with the scalar product (·,·)X and the norm k · kX. Denote byL2loc(S;X) the linear space of measurable functions defined onS with values inX, whose restrictions to any segment [a, b]⊂S belong to the space L2(a, b;X).
Letω∈R,α∈C(S) be such thatα(t)>0 for allt∈S,γ=αorγ= 1/α, and letX be as above. Put by definition
L2ω,γ(S;X) :=n
f ∈L2loc(S;X) : Z
S
γ(t)e2ωR0tα(s)dskf(t)k2Xdt <∞o . This space is a Hilbert space with respect to the scalar product
(f, g)L2
ω,γ(S;X)= Z
S
γ(t)e2ωR0tα(s)ds(f(t), g(t))Xdt and the norm
kfkL2
ω,γ(S;X):=Z
S
γ(t)e2ωR0tα(s)dskf(t)k2Xdt1/2
.
For an interval I, we denote by Cc1(I) the linear space of continuously differ- entiable functions on I with compact supports (if I = (t1, t2), then we will write Cc1(t1, t2) instead ofCc1((t1, t2))).
LetH1(Ω) :={v∈L2(Ω) :vxi∈L2(Ω) (i= 1, n)}be a Sobolev space, which is a Hilbert space with respect to the scalar product (v, w)H1(Ω):=R
Ω
∇v∇w+vw dx
and the corresponding norm kvkH1(Ω) := R
Ω
|∇v|2+|v|2 dx1/2
, where∇v = (vx1, . . . , vxn), |∇v|2=Pn
i=1|vxi|2. UnderH01(Ω) we mean the closure inH1(Ω) of the spaceCc∞(Ω) consisting of infinitely differentiable functions on Ω with compact supports. Denote by H−1(Ω) the dual space of H01(Ω), that is, the space of all continuous linear functionals onH01(Ω).
We suppose (after appropriate identification of functionals), that the spaceL2(Ω) is a subspace of H−1(Ω). Identifying spacesL2(Ω) and L2(Ω)0
, we obtain con- tinuous and dense embeddings
H01(Ω)⊂L2(Ω)⊂H−1(Ω). (2.1) Note, that in this casehg, viH1
0(Ω)= (g, v) for everyv ∈H01(Ω), g∈L2(Ω), where (·,·) is the scalar product on L2(Ω) and h·,·iH1
0(Ω) is the scalar product for the duality H−1(Ω), H01(Ω). Therefore, further we use the notation (·,·) instead of h·,·iH1
0(Ω). We define
K:= inf
v∈H01(Ω), v6=0
R
Ω|∇v|2dx R
Ω|v|2dx . (2.2)
It is well known that the constantKis finite and coincides with the first eigenvalue of the eigenvalue problem
−∆v=λv, v|∂Ω= 0. (2.3)
From (2.2) it clearly follows the Friedrichs inequality Z
Ω
|∇v|2dx≥K Z
Ω
|v|2dx ∀v∈H01(Ω). (2.4) Further, an important role will be played by the following statement, which is a well-known result (see, e.g. [10, Theorem 3, p. 287]), but we reformulate it according to our needs.
Lemma 2.1. Suppose that a functionz∈L2(t1, t2;H01(Ω)), witht1< t2, satisfies Z t2
t1
Z
Ω
−zψϕ0+ (g0ψ+
n
X
i=1
giψxi)ϕ dx dt= 0, (2.5) forψ∈H01(Ω),ϕ∈Cc1(t1, t2), wheregi∈L2(Ω×(t1, t2)) (i= 0, n). Then
(1) the derivative zt of the function z in the sense D0(t1, t2;H−1(Ω)) (the dis- tributions space) belongs toL2(t1, t2;H−1(Ω)), furthermore for a.e. t∈(t1, t2),
zt(·, t) =−g0(·, t) +
n
X
i=1
gi(·, t)
xi inH−1(Ω), (2.6) 1
2 d
dtkz(·, t)k2L2(Ω)= (zt(·, t), z(·, t)), (2.7) Z t2
t1
kzt(·, t)k2H−1(Ω)dt≤
n
X
i=0
kgik2L2(Ω×(t1,t2)); (2.8)
(2) the function z belongs to the space C([t1, t2];L2(Ω)) and for all τ1, τ2 ∈ [t1, t2] (τ1< τ2)and for everyθ∈C1([t1, t2]),q∈L2(t1, t2;H01(Ω)) we have
1 2θ(t)
Z
Ω
|z(x, t)|2dx
t=τ2 t=τ1−1
2 Z τ2
τ1
Z
Ω
|z|2θ0dx dt +
Z τ2
τ1
Z
Ω
g0z+
n
X
i=1
gizxi θ dx dt= 0,
(2.9)
Z τ2
τ1
zt(·, t), q(·, t) dt+
Z τ2
τ1
Z
Ω
g0q dx dt+
n
X
i=1
Z τ2
τ1
Z
Ω
giqxidx dt= 0. (2.10) Proof. As it has already been mentioned, this lemma follows directly from the well- known result. But for clarity we re-present schematically some points of the proof.
The first statement is: Since the spaces L2(t1, t2;H01(Ω)), L2(t1, t2;H−1(Ω)) can be identified with subspaces of the space of distributions D0(t1, t2;H−1(Ω)), then it allows us to speak about derivatives of functions from L2(t1, t2;H01(Ω)) in the senseD0(t1, t2;H−1(Ω)) and their belonging to the spaceL2(t1, t2;H−1(Ω)).
Let us rewrite equality (2.5) in the form
− Z t2
t1
Z
Ω
zψϕ0dx dt=− Z t2
t1
Z
Ω
(g0ψ+
n
X
i=1
giψxi)ϕ dx dt, (2.11) forψ∈H01(Ω),ϕ∈Cc1(t1, t2). According to the definition of the derivative of dis- tributions fromD0(t1, t2;H−1(Ω)), (2.11) implies existence ofztand its belonging to the spaceL2(t1, t2;H−1(Ω)), then according to [10, Theorem 3, p. 287] identity (2.7) holds. From (2.11) for almost allt∈(t1, t2) we have
zt(·, t), ψ(·)
=− Z
Ω
g0(x, t)ψ(x) +
n
X
i=1
gi(x, t)ψxi(x)
dx, (2.12) that is, (2.6) holds.
From (2.12), using the Cauchy-Schwarz inequality, for almost allt∈(t1, t2) we obtain
zt(·, t), ψ(·)
≤ kg0(·, t)kL2(Ω)kψ(·)kL2(Ω)+
n
X
i=1
kgi(·, t)kL2(Ω)kψxi(·)kL2(Ω)
≤Xn
i=0
kgi(·, t)k2L2(Ω)
1/2
kψ(·)kH1(Ω).
(2.13)
From (2.13) it follows that for almost allt∈(t1, t2) the following estimate is valid kzt(·, t)k2H−1(Ω)≤
n
X
i=0
kgi(·, t)k2L2(Ω), which easily implies (2.8).
Let us prove the second statement of Lemma 2.1. The fact that the functionz belongs to the spaceC([t1, t2];L2(Ω)) follows directly from [10, Theorem 3, p. 287]
according to the first statement.
Since for a.e. t ∈ S the function q(·, t) ∈H01(Ω), we can take ψ(·) = q(·, t) in (2.12) and obtain
zt(·, t), q(·, t)
=− Z
Ω
g0(x, t)q(x, t) +
n
X
i=1
gi(x, t)qxi(x, t)
dx, t∈S. (2.14) Integrating this inequality by t over (τ1, τ2) for arbitrary τ1, τ2 ∈ S, we obtain (2.10).
Takingq(·, t) =θ(t)z(·, t),t∈S, in (2.10) and integrating over (τ1, τ2), we obtain Z τ2
τ1
θ(t) zt(·, t), z(·, t) dt+
Z τ2
τ1
Z
Ω
g0z+
n
X
i=1
gizxi θ dx dt= 0. (2.15) Using (2.7) and integration by parts, we have
Z τ2
τ1
θ(t) zt(·, t), z(·, t) dt= 1
2 Z τ2
τ1
θ(t)d
dtkz(·, t), z(·, tk2L2(Ω)dt
= 1
2θ(t)kz(·, t)k2L2(Ω)
t=τ2
t=τ1−1 2
Z τ2
τ1
θ0kz(·, t)kL2(Ω)dt,
which, together with (2.15), gives (2.9).
3. Well-posedness of the problem without initial conditions for linear parabolic equations
Consider the equation yt−
n
X
i,j=1
aij(x, t)yxi
xj +a0(x, t)y=f(x, t), (x, t)∈Q, (3.1) wherey:Q→Ris an unknown function and data-in satisfies conditions:
(A1) a0, aij ∈L∞loc(Q), aij =aji (i, j = 1, n), a0(x, t)≥0 for a. e. (x, t) ∈Q, there exists a function α ∈ C(S) such that α(t) > 0 for all t ∈ S and Pn
i,j=1aij(x, t)ξiξj ≥α(t)|ξ|2for every ξ∈Rn and for a. e. (x, t)∈Q;
(A2) f ∈L2loc(S;L2(Ω)).
Additionally, we impose the boundary condition y
Σ= 0 (3.2)
on a solution of equation (3.1).
Definition 3.1. A weak solution of problem (3.1), (3.2) is a function y which belongs toL2loc(S;H01(Ω))∩C(S;L2(Ω)) and satisfies
Z Z
Q
n−yψϕ0+
n
X
i,j=1
aijyxiψxjϕ+a0yψϕo dx dt
= Z Z
Q
f ψϕ dx dt, ψ∈H01(Ω), ϕ∈Cc1(−∞,0).
(3.3)
In other words: a weak solution of problem (3.1), (3.2) is the function y which belongs toL2loc(S;H01(Ω))∩C(S;L2(Ω)) withyt∈L2loc(S;H−1(Ω)), and satisfies
yt−
n
X
i,j=1
(aijyxi)xj+a0y=f in L2loc(S;H−1(Ω)).
Remark 3.2. There may exist many weak solutions of problem (3.1), (3.2). E.g., the functionsyc(x, t) =cv(x)e−Kt, (x, t)∈Q(c∈R), wherevis an eigenfunction of problem (2.3) corresponding to the first eigenvalue, are weak solutions of problem (3.1), (3.2) when aij = δij, a0 = 0 and f = 0, where δij is Kronecker’s delta (i, j= 1, n). Therefore, to ensure uniqueness of the weak solution of (3.1) satisfying condition (3.2), we have to impose some additional conditions on solutions, for instance, some restrictions on their behavior ast→ −∞.
We will consider the problem of finding the weak solution of (3.1), (3.2) satisfying the analogue of the initial condition
t→−∞lim eωR0tα(s)dsky(·, t)kL2(Ω)= 0, (3.4) whereω∈Ris given.
We will briefly call this problem by problem (3.1), (3.2), (3.4), and the function y is called the solution of problem (3.1), (3.2), (3.4).
Theorem 3.3. Suppose that condition(A1)holds,Kis a constant defined by (2.2).
The following two statements hold:
(1) Ifω≤K then (3.1),(3.2),(3.4)has at most one weak solution.
(2) Ifω < K and
f ∈L2ω,1/α(S;L2(Ω)), (3.5)
then there exists a unique weak solution of (3.1),(3.2),(3.4), it belongs to the space L2ω,α(S;H01(Ω)) and the following estimates are satisifed
eωR0τα(s)dsky(·, τ)kL2(Ω)≤C1kfkL2
ω,1/α(Sτ;L2(Ω)), τ ∈S, (3.6) kykL2
ω,α(Sτ;H01(Ω))≤C2kfkL2
ω,1/α(Sτ;L2(Ω)), τ∈S, (3.7) whereSτ:= (−∞, τ] (τ ∈(−∞,0],S0=S),C1, C2 are positive constants depend- ing only onK andω.
Remark 3.4. In the particular case of equation (3.1), which was considered in Remark 3.2, we haveα(t) = 1, therefore condition (3.4) takes on the form:
eωtky(·, t)kL2(Ω)→0 ast→ −∞.
Obviously in this case for the nonzero solutions of (3.1), (3.2), (3.4), indicated in Remark 3.2, we have limt→−∞eKtkyc(·, t)kL2(Ω) =C, where C is a nonzero con- stant; limt→−∞eωtkyc(·, t)kL2(Ω)= +∞, ifω < K; limt→−∞eωtkyc(·, t)kL2(Ω)= 0, if ω > K. This means that the condition ω ≤ K is essential for ensuring the uniqueness of the weak solution of (3.1), (3.2), (3.4), i.e., it cannot be simplified.
Proof of Theorem 3.3. In the proof we use the same technique as in the proofs of corresponding results in [4,5]. Nevertheless, we present the proof, because it is important for us to obtain more precise estimates of the solution of (3.1), (3.2), (3.4) and to track how this solution depends on the coefficient (which serves as a control in the following sections).
Let us prove the first statement of Theorem 3.3. Assume the opposite. Let y1, y2 be two weak solutions of (3.1), (3.2), (3.4). Substituting them one by one into integral identity (3.3) and subtracting the obtained equalities, for the difference z:=y1−y2 we obtain
− Z Z
Q
zψϕ0dx dt+ Z Z
Q
Xn
i,j=1
aijzxiψxj+a0zψ
ϕ dx dt= 0, (3.8)
for allψ∈H01(Ω),ϕ∈Cc1(−∞,0).
From (3.4) it follows that e2ωR0tα(s)ds
Z
Ω
|z(x, t)|2dx→0 ast→ −∞. (3.9) According to Lemma 2.1 withθ(t) = 2e2ωR0tα(s)ds,t∈R, (3.8) implies that
e2ωR0τ2α(s)ds Z
Ω
|z(x, τ2)|2dx−e2ωR0τ1α(s)ds Z
Ω
|z(x, τ1)|2dx
−2ω Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|z|2dx dt + 2
Z τ2
τ1
Z
Ω
e2ωR0tα(s)dsh Xn
i,j=1
aijzxizxj+a0|z|2i
dx dt= 0, whereτ1, τ2∈S (τ1< τ2) are arbitrary numbers.
Taking into account condition (A1) and inequality (2.4), we obtain e2ωR0τ2α(s)ds
Z
Ω
|z(x, τ2)|2dx−e2ωR0τ1α(s)ds Z
Ω
|z(x, τ1)|2dx + 2(K−ω)
Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|z|2dx dt≤0.
(3.10)
Sinceω≤K, from (3.10) we obtain e2ωR0τ2α(s)ds
Z
Ω
|z(x, τ2)|2dx≤e2ωR0τ1α(s)ds Z
Ω
|z(x, τ1)|2dx. (3.11) In (3.11) fix τ2 and let τ1 to −∞. According to condition (3.9) we obtain the equality
e2ωR0τ2α(s)ds Z
Ω
|z(x, τ2)|2dx= 0.
Sinceτ2∈S is an arbitrary number, we have z(x, t) = 0 for a. e. (x, t)∈Q, that is, y1(x, t) = y2(x, t) = 0 for a. e. (x, t)∈ Q. The resulting contradiction proves the first statement.
Let us prove the second statement. First we determine a priori estimates of a weak solution of (3.1), (3.2), (3.4). According to Lemma 2.1, condition (3.3) implies
1 2θ(τ2)
Z
Ω
|y(x, τ2)|2dx−1 2θ(τ1)
Z
Ω
|y(x, τ1)|2dx
−1 2
Z τ2
τ1
Z
Ω
|y|2θ0dx dt+ Z τ2
τ1
Z
Ω
h Xn
i,j=1
aijyxiyxj +a0|y|2i θ dx dt
= Z τ2
τ1
Z
Ω
f yθ dx dt,
(3.12)
whereθ∈C1(S) is an arbitrary function,τ1, τ2∈S(τ1< τ2) are arbitrary numbers.
Further assume thatθ(t)≥0 for allt∈S.
Using the Cauchy inequality with “ε“:
ab≤ ε 2a2+ 1
2εb2, a, b∈R, ε >0,
we estimate the right side of (3.12) as follows:
Z τ2
τ1
Z
Ω
f yθ dx dt ≤ ε
2 Z τ2
τ1
Z
Ω
α|y|2θ dx dt+ 1 2ε
Z τ2
τ1
Z
Ω
[α]−1|f|2θ dx dt, (3.13) whereε >0 is arbitrary.
From condition (A1) we obtain Z τ2
τ1
Z
Ω
n
X
i,j=1
aijyxiyxj +a0|y|2
θ dx dt≥ Z τ2
τ1
Z
Ω
α|∇y|2θ dx dt, (3.14) where∇y:= (yx1, . . . , yxn) is the gradient of y.
According to (3.13) and (3.14), equality (3.12) implies 1
2θ(τ2) Z
Ω
|y(x, τ2)|2dx−1 2θ(τ1)
Z
Ω
|y(x, τ1)|2dx
−1 2
Z τ2
τ1
Z
Ω
|y|2θ0dx dt+ Z τ2
τ1
Z
Ω
α|∇y|2θ dx dt
≤ε 2
Z τ2
τ1
Z
Ω
α|y|2θ dx dt+ 1 2ε
Z τ2
τ1
Z
Ω
[α]−1|f|2θ dx dt, whereε >0 is arbitrary.
Takingθ(t) = 2e2ωR0tα(s)ds witht∈S, we obtain e2ωR0τ2α(s)ds
Z
Ω
|y(x, τ2)|2dx−e2ωR0τ1α(s)ds Z
Ω
|y(x, τ1)|2dx
−2ω Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|y|2dx dt+ 2 Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|∇y|2dx dt
≤ε Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|y|2dx dt+1 ε
Z τ2
τ1
Z
Ω
[α(t)]−1e2ωR0tα(s)ds|f|2dx dt.
By the above inequality and using (2.4), we obtain e2ωR0τ2α(s)ds
Z
Ω
|y(x, τ2)|2dx−e2ωR0τ1α(s)ds Z
Ω
|y(x, τ1)|2dx +χ(K, ω, ε)
Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|∇y|2dx dt
≤1 ε
Z τ2
τ1
Z
Ω
[α(t)]−1e2ωR0tα(s)ds|f|2dx dt,
(3.15)
where
χ(K, ω, ε) :=
(2(K−ω)−ε
K if 0< ω < K,
2K−ε
K ifω≤0.
Takingε=Kifω≤0, andε=K−ωif 0< ω < K in (3.15), we obtain e2ωR0τ2α(s)ds
Z
Ω
|y(x, τ2)|2dx−e2ωR0τ1α(s)ds Z
Ω
|y(x, τ1)|2dx +C3
Z τ2
τ1
Z
Ω
α(t)e2ωR0tα(s)ds|∇y|2dx dt
≤C4
Z τ2
τ1
Z
Ω
[α(t)]−1e2ωR0tα(s)ds|f|2dx dt,
(3.16)
whereC3>0,C4>0 are constants depending only onK andω.
Taking into account (3.4) and (3.5), we let τ1 → −∞ in (3.16). As a result, adoptingτ2=τ ∈S, we obtain
e2ωR0τα(s)ds Z
Ω
|y(x, τ)|2dx+C3 Z τ
−∞
Z
Ω
α(t)e2ωR0tα(s)ds|∇y|2dx dt
≤C4 Z τ
−∞
Z
Ω
[α(t)]−1e2ωR0tα(s)ds|f|2dx dt.
(3.17)
Hence, using inequality (2.4), we easily obtain estimates (3.6) and (3.7).
Now let us prove the existence of a weak solution of problem (3.1), (3.2), (3.4).
First, for eachm∈Nwe defineQm:= Ω×(−m,0],fm(·, t) :=f(·, t), if−m < t≤0, and fm(·, t) := 0, if t ≤ −m, and consider the problem of finding a function ym∈L2(−m,0;H01(Ω))∩C([−m,0];L2(Ω)) satisfying the initial condition
ym(x,−m) = 0, x∈Ω, (3.18) (as an element of spaceC([−m,0];L2(Ω))) and equation (3.1) inQm in the sense of integral identity; that is,
Z Z
Qm
n−ymψϕ0+
n
X
i,j=1
aijym,xiψxjϕ+a0ymψϕo dx dt=
Z Z
Qm
fmψϕ dx dt, forψ∈H01(Ω),ϕ∈Cc1(−m,0).
The existence and uniqueness of the solution of this problem easily follows from the known results (see, for example, [16]). For everym∈Nwe extendymby zero for the entire setQand keep the same notationymfor this extension. Note that for eachm∈N, the functionymbelongs toL2(S;H01(Ω))∩C(S;L2(Ω)) and satisfies integral identity (3.3) withfmsubstituted forf, i.e.,
Z Z
Q
n−ymψϕ0+
n
X
i,j=1
aijym,xiψxjϕ+a0ymψϕo dx dt=
Z Z
Q
fmψϕ dx dt, (3.19) forψ ∈H01(Ω), ϕ∈ Cc1(−∞,0). Consequently, we have shown that ym is a weak solution of problem (3.1), (3.2), (3.4) withfmsubstituted forf. Therefore, forym
we obtain estimates similar to (3.6), (3.7), in particular, forτ∈S, e2ωR0τα(s)dskym(·, τ)k2L2(Ω)≤C1
Z τ
−∞
[α(t)]−1e2ωR0tα(s)dskf(·, t)k2L2(Ω)dt, (3.20) Let us take identity (3.19) with alternating m = k and m = l, where k, l are arbitrary positive integers, l > k, and then subtract the obtained identities. As a result, we obtain the same identity as (3.19) with zk,l := yk−yl, fk,l :=fk −fl instead of ym and fm, respectively. Finally taking into account that the function zk,lsatisfies conditions (3.2) and (3.4), replacingywithzk,l, we see that the function zk,lis a weak solution of the problem, which differs from problem (3.1), (3.2), (3.4) only in that instead ofy and f, there arezk,land fk,l, respectively. Thus, forzk,l we have estimates similar to (3.6), (3.7), i.e.
e2ωR0τα(s)dskyk(·, τ)−yl(·, τ)k2L2(Ω)
≤C1
Z −k
−l
[α(t)]−1e2ωR0tα(s)dskf(·, t)k2L2(Ω)dt, τ∈S,
(3.21)
kyk−ylkL2
ω,α(S;H01(Ω))≤C2
Z −k
−l
[α(t)]−1e2ωR0tα(s)dskf(·, t)k2L2(Ω)dt. (3.22)
Condition (3.5) implies that the right-hand sides of inequalities (3.21) and (3.22) tend to zero when k and l tend to +∞. This means that the sequence {ym}∞m=1 is a Cauchy sequence in the spaceL2ω,α(S;H01(Ω)) andC(S;L2(Ω)). Consequently, we obtain the existence of the functiony ∈L2ω,α(S;H01(Ω))∩C(S;L2(Ω)) such that
ym −→
m→∞y strongly inL2ω,α(S;H01(Ω)) and C(S;L2(Ω)). (3.23) Note that (3.23) implies
ym −→
m→∞y, ym,xi −→
m→∞yxi (i= 1, n) strongly inL2loc(S;L2(Ω)). (3.24) Let us show that the function y is a weak solution of (3.1), (3.2), (3.4). To do this, first we letm→ ∞in identity (3.19), taking into account (3.24) and the definition of the functionfm. Consequently, we obtain identity (3.3). Now, taking into account (3.23), we let m→+∞in (3.20). From the resulting inequality and condition (3.5), we obtain condition (3.4). Hence, we have proven thatyis a weak
solution of problem (3.1), (3.2), (3.4).
4. Formulation of the optimal control problem and main result LetU :=L∞(Q) be a space of controls andU∂ be a convex and closed subset of {v∈U :v≥0 a. e. inQ}. We suppose that U∂ is the set of admissible controls.
We assume that the state of the investigated evolutionary system for a given controlv∈U∂ is described by a weak solution of (3.1), (3.2), (3.4) whena0=ea0+v, whereea0∈L∞loc(Q) is a given function such thatea0≥0 a. e. inQ. Then, equation (3.1) has the form
yt−
n
X
i,j=1
aij(x, t)yxi
xj+ (ea0(x, t) +v(x, t))y=f(x, t), (x, t)∈Q. (4.1) The specified problem will be called problem (4.1), (3.2), (3.4). The weak solution y of (4.1), (3.2), (3.4) for a given control v, denoted by y, or y(v), or y(x, t), (x, t)∈Q, ory(x, t;v), (x, t)∈Q. Further, we assume that conditions (A1), (3.5) and the inequalityω < K hold. From the previous section (see Theorem 3.3), we immediately obtain the existence and uniqueness of the weak solution of problem (4.1), (3.2), (3.4) and its estimates (3.6), (3.7).
We assume that the cost functional has the form
J(v) =ky(·,0;v)−z0(·)k2L2(Ω)+µkvkL∞(Q), v∈U∂, (4.2) wherez0∈L2(Ω),µ≥0 if U∂ is bounded, andµ >0 otherwise.
We consider the following optimal control problem: find a controlu∈U∂ such that
J(u) = inf
v∈U∂J(v). (4.3)
We call this problem (4.3), and its solutions will be calledoptimal controls.
The main results of this paper are the following.
Theorem 4.1 (Existence of an optimal control). With the above assumptions in this section, a set of optimal controls of problem (4.3) is nonempty and ∗-weakly closed inL∞(Q).
Theorem 4.2 (Necessary conditions of an optimal control). Let U∂ be bounded, µ= 0, and
α(t)≥α0= const. >0 for a.e. t∈S. (4.4) Then an optimal control of problem (4.3)satisfies the relations
y∈L2ω,α(S;H01(Ω)), yt∈L2loc(S;H−1(Ω)), yt−
n
X
i,j=1
(aijyxi)xj + (ea0+u)y=f inL2loc(S;H−1(Ω)), y
Σ= 0, lim
t→−∞eωR0tα(s)dsky(·, t)kL2(Ω)= 0,
(4.5)
p∈L2−ω,1/α(S;H01(Ω)), pt∈L2loc(S;H−1(Ω)),
−pt−
n
X
i,j=1
(aijpxi)xj + (ea0+u)p= 0 in L2loc(S;H−1(Ω)), p
Σ= 0, p(·,0) =y(·,0)−z0(·),
(4.6)
Z Z
Q
yp(v−u)dx dt≤0 ∀v∈U∂. (4.7) Since y belongs to L2ω,α(S;H01(Ω)), and p belongs to L2−ω,1/α(S;H01(Ω)), the product py belongs to L1(Q), and thus the left-hand side of inequality (4.7) is well-defined.
Problem (4.6) is called an adjoint problem, its solution is called an adjoint state and is introduced in order to characterize an optimal control.
5. Proof of main results
Proof of Theorem 4.1. Since the cost functionalJ is bounded below, there exists a minimizing sequence {vk} in U∂: limk→∞J(vk) = infv∈U∂J(v). This and (4.2) imply that the sequence{vk} is bounded in the spaceL∞(Q), that is
ess sup(x,t)∈Q|vk(x, t)| ≤C5, (5.1) whereC5 is a constant, which does not depend onk.
Since for eachk∈Nthe functionyk:=y(vk) (k∈N) is a weak solution of (4.1), (3.2), (3.4) forv=vk, the following identity holds:
Z Z
Q
n−ykψϕ0+
n
X
i,j=1
aijyk,xiψxjϕ+ (ea0+vk)ykψϕo dx dt
= Z Z
Q
f ψϕ dx dt, ψ∈H01(Ω), ϕ∈Cc1(−∞,0).
(5.2)
According to Theorem 3.3 we have the estimates e2ωR0τα(s)dskyk(·, τ)k2L2(Ω)≤C1kfkL2
ω,1/α(Sτ;L2(Ω)), τ∈S, (5.3) kykkL2
ω,α(Sτ;H10(Ω))≤C2kfkL2
ω,1/α(Sτ;L2(Ω)). (5.4)
Taking into account the first statement of Lemma 2.1, from (5.2) for arbitrary τ1, τ2∈S (τ1< τ2) we obtain
Z τ2
τ1
kyk,tk2H−1(Ω)dt≤ Z τ2
τ1
Z
Ω
Xn
j=1
n
X
i=1
aijyk,xi
2+|(ea0+vk)yk−f|2
dx dt. (5.5)
By condition (A1), (3.5), (5.1), and (5.4), estimate (5.5) implies Z τ2
τ1
kyk,t(·, t)k2H−1(Ω)dt≤C6, (5.6) whereτ1, τ2∈S (τ1< τ2) are arbitrary,C6>0 is a constant which depends onτ1 andτ2, but does not depend on k.
By the Compactness Lemma (see [20, Proposition 4.2]), and the compactness of the embeddingH01(Ω)⊂L2(Ω) (see [19, p. 245]), estimates (5.1), (5.4), (5.6) yield that there exist a subsequence of the sequence {vk, yk} (which is also denoted by {vk, yk}) and functionsu∈U∂, andy∈L2ω,α(S;H01(Ω)) such that
vk −→
k→∞u ∗-weakly in L∞(Q), (5.7)
yk −→
k→∞y weakly in L2ω,α(S;H01(Ω)), (5.8) yk −→
k→∞y strongly in L2loc(S;L2(Ω)). (5.9) Note that (5.8) implies
yk −→
k→∞y, yk,xi −→
k→∞yxi (i= 1, n) weakly in L2loc(S;L2(Ω)). (5.10) Let us show that (5.7) and (5.9) yield
Z Z
Q
ykvkψϕ dx dt −→
k→∞
Z Z
Q
yuψϕ dx dt ∀ψ∈H01(Ω),∀ ϕ∈Cc1(−∞,0). (5.11) Indeed, letg:=ψϕ, andt1, t2∈S be such that suppϕ⊂[t1, t2]. Then we have
Z Z
Q
ykvkg dx dt= Z t2
t1
Z
Ω
(ykvk−yvk+yvk)g dx dt
= Z t2
t1
Z
Ω
yvkg dx dt+ Z t2
t1
Z
Ω
(yk−y)vkg dx dt.
(5.12)
From (5.1) and (5.9) it follows that
Z t2
t1
Z
Ω
(yk−y)vkg dx dt
≤Z t2 t1
Z
Ω
|vkg|2dx dt1/2Z t2 t1
Z
Ω
|yk−y|2dx dt1/2
→0 as k→ ∞.
(5.13)
Thus, using (5.7) and (5.13), (5.12) implies (5.11).
Using (5.10) and (5.11), and lettingk→ ∞ in (5.2), we obtain Z Z
Q
n−yψϕ0+
n
X
i,j=1
aijyxiψxjϕ+ (ea0+u)yψϕo dx dt
= Z Z
Q
f ψϕ dx dt, ψ∈H01(Ω), ϕ∈Cc1(−∞,0).
(5.14)
According to Lemma 2.1, identity (5.14) implies that y ∈ C(S;L2(Ω)) and yt ∈ L2loc(S;H−1(Ω)). Hence, the functiony=y(u) is a weak solution of problem (4.1), (3.2). Let us show thaty satisfies condition (3.4). First, we prove the convergence
∀τ∈S: yk(·, τ) −→
k→∞y(·, τ) strongly inL2(Ω). (5.15)
For this purpose, we subtract (5.2) from (5.14). To the resulting identity, we apply Lemma 2.1 with z =y−yk, g0 = (ea0+u)y−(ea0+vk)yk, gi =Pn
j=1aij(yxj − yk,xj) (i= 1, n),θ(t) = 2(t−τ+ 1),τ1 =τ−1,τ2 =τ, whereτ ∈S is arbitrary.
Consequently, Z
Ω
|y(x, τ)−yk(x, τ)|2dx− Z τ
τ−1
Z
Ω
|y−yk|2dx dt +
Z τ
τ−1
Z
Ω
h Xn
i,j=1
aij(yxi−yk,xi)(yxj −yk,xj)i θ dx dt
+ Z τ
τ−1
Z
Ω
(ea0+u)y−(ea0+vk)yk
y−yk
θ dx dt= 0.
(5.16)
Let us transform the last term on the left side of (5.16) as follows:
Z τ
τ−1
Z
Ω
(ea0+u)y−(ea0+vk)yk
y−yk θ dx dt
= Z τ
τ−1
Z
Ω
(ea0+u)y−(ea0+vk)(yk−y+y)
y−yk θ dx dt
= Z τ2
τ1
Z
Ω
(ea0+vk)|y−yk|2+ (u−vk)y(y−yk) θ dx dt.
(5.17)
From (5.16), taking into account (A1) and (5.17), we obtain Z
Ω
|y(x, τ)−yk(x, τ)|2dx+ 2 Z τ
τ−1
Z
Ω
(ea0+vk)|y−yk|2dx dt
≤ Z τ
τ−1
Z
Ω
|y(y−yk)||u−vk|dx dt+ Z τ
τ−1
Z
Ω
|y−yk|2dx dt.
(5.18)
Using (5.1) and Cauchy-Schwarz inequality, (5.18) yields Z
Ω
|y(x, τ)−yk(x, τ)|2dx
≤C7Z τ τ−1
Z
Ω
|y−yk|2dx dt1/2 +
Z τ
τ−1
Z
Ω
|y−yk|2dx dt,
(5.19)
whereC7>0 is a constant which does not depend onk.
From (5.9), according to (5.19), we obtain (5.15). Taking into account (5.15), lettingk→ ∞in (5.3), the resulting inequality, according to condition (3.5), implies
τ→−∞lim e2ωR0τα(s)ds Z
Ω
|y(x, τ)|2dx= 0. (5.20) Hence, we have shown that y = y(u) = y(x, t;u), (x, t) ∈ Q, is the state of the controlled system for the controlu.
It remains to prove thatuis a minimizing element of the functionalJ. Indeed, (5.15) implies
kyk(·,0)−z0(·)k2L2(Ω) −→
k→∞ky(·,0)−z0(·)k2L2(Ω). (5.21) Also, (5.7) and properties of∗-weakly convergent sequences yield
lim inf
k→∞ kvkkL∞(Q)≥ kukL∞(Q). (5.22)