142 A. ADDOU and A.

BENBRIK

where z is a solution of the state equation

 N
 ∂z ∂z

 + ∆2 z = i=1 ui(t) in Q = Ω×]0, T [,
 ∂t ∂xi
(E) ∂z

 z= =0 on Σ = Γ×]0, T [,

 ∂ν
z(x, 0) = z0 (x) in Ω.
We restrict ourselves to the study of the above problem and solve it by
the standard variational method, which can be extended to more general
situations.
After fixing the functional framework, in the first section we prove a
priori estimates for solutions of the equations describing the system. In the
second section, we study the existence and uniqueness of solutions for this
type of equations. The last section is devoted to proving the existence of an
optimal control u and obtaining optimality conditions and the uniqueness
of u.
Functional framework. Denote V = H02 (Ω); let V  = H −2 (Ω) be dual
space and let H = L2 (Ω) be identified with its dual H  . (Ω is a bounded
domain of RN with smooth boundary Γ.)
Then we have the Gelfand triple
V → H → V  ;
all the embeddings are continuous, dense, and compact.
We introduce the space

W = w ∈ L2 (0, T ; V ); ẇ ∈ L2 (0, T ; V  )
endowed with the norm of the graph
wW = wL2(0,T ;V ) + ẇL2 (0,T ;V
) .
It is well known [6] that:
• W is a Banach space.
• Every w ∈ W is continuous a.e. on [0, T ] with values in H.
• The embedding W → L2 (0, T ; H) is compact.
We now consider the equation
ż + Az(t) =u(t)Bz(t) + f(t),
z(0) =z0 ,
where
• A : V → V  , ϕ → ∆2 (ϕ) is a linear continuous operator and ϕ =
a(ϕ), where a(ϕ) = A(ϕ), ϕ V
V is equivalent to the norm defined
in H02 (Ω);
 OPTIMAL CONTROL OF A BILINEAR DISTRIBUTED SYSTEM 143

∂ϕ
• B = (B1 , . . . , BN ), where Bi : V → H, ϕ → , is linear and
∂xi
continuous.
As the control space, we take U = L2 (0, T ; RN ) and set
N
∂z
u(t)Bz(t) = ui(t) (t);
∂xi
i

∂z
ui(t) (t) are then in L2 (0, T ; V  ) and the choice of U is compatible with
∂xi
the equation. The performance index becomes

J(u) = u2U + z2L2(0,T ;H) .

1. A priori estimates
• The triple consisting of equation with the operator Au = A−u(t)B,
f ∈ L2 (0, T ; V  ) and the initial state z0 , is denoted by E(z0 , u, f).
• The spaces L2 (0, T ; V ), L2 (0, T ; V  ), and L∞ (0, T ; H) are denoted
by L2 (V ), L2 (V  ), and L∞ (H), respectively.
We recall that by a solution of E(z0 , u, f), we mean a function z : [0, T ] → V
such that


 ż + Az(t) = u(t)Bz(t) + f(t),
E(z0 , u, f) z(0) = z0 ,

 z ∈ W.

Lemma 1.1. Assume that z0 ∈ H, u ∈ U = L2 (0, T ; RN ), and f ∈

L (V  ). Then if z is a solution of E(z0 , u, f), the following inequalities
2

hold:
1
(E1 ) zL2 (V ) ≤ √ |z0 | + fL2 (V
) ;
2

√
(E2 ) zL∞(H) ≤ |z0 | + 2fL2 (V
) ;

1  √ 
(E3 ) żL2 (V
) ≤ √ |z0 | + fL2 (V
) α1 + 2uU + fL2 (V
)
2

(where α1 is such that AξV
≤ α1 ξV ).

144 A. ADDOU and A. BENBRIK

Proof.
1. We multiply equation E(z0 , u, f) by z(t) and integrate the result over
[0, T ]:
T T T
ż(t), z(t) dt + Az(t), z(t) dt = u(t)Bz(t), z(t) dt +
0 0 0
T
+ f(t), z(t) dt.
0

T
Since (u(t)Bz(t), z(t))V,V
= 0 and z2L2 (V ) = 0
Az(t), z(t) dt, we have
1
z2L2 (V ) ≤ |z0 |2 + zL2(V ) fL2 (V
) .
2
Thus,
 1 2 1 1
zL2 (V ) − fL2 (V
) ≤ |z0 |2 + f2L2 (V
) .
2 2 4
This implies (E1 ).
2. Multiplying E(z0 , u, f) by z and integrating over [0, t], we obtain
1 1
|z(t)|2 ≤ |z0 |2 + zL2 (V ) fL2 (V
) ;
2 2
now (E1 ) implies (E2 ).
3. Using the duality L2 (V ), L2 (V  ) defined by
T
Ψ, ϕ = Ψ(t), ϕ(t) V
,V dt (Ψ ∈ L2 (V  ) and ϕ ∈ L2 (V )),
0

for ξ ∈ L (V ), we have
2

T  T  T 

     
 ż(t), ξ(t) dt ≤  Az(t), ξ(t) dt +  u(t)Bz(t), ξ(t) +
     
0 0 0
T 
 
+  f(t), ξ(t) ;

0

hence
  
ż, ξ  ≤ α1 zL2 (V ) .ξL2(V ) + zL∞ (H) uL2 (0,T ;RN ) ξL2 (V ) +

+ fL2 (V
) ξL2(V ) .
The definition of żL2 (V
) and inequality (E2 ) yield (E3 ).
 OPTIMAL CONTROL OF A BILINEAR DISTRIBUTED SYSTEM 145

2. Evolution equation
Proposition 2.1. Assume that u ∈ U and z0 ∈ H. The equation
E(z0 , u, 0) admits a unique solution z in L2 (V ) ∩ L∞ (H) with ż ∈ L2 (V  ).
Proof.
1. If zi (i = 1, 2) is a solution of E(z0 , u, f), then z = z1 − z2 satisfies
ż + Az = uBz,
z(0) = 0.
For t ∈ [0, T ], we have
t
1
|z(t)|2 ≤ |u(τ )||z(τ )|2 dτ.
2
0

Using the Gronwall lemma, we obtain

z1 = z2 .
2. The existence follows from the standard application of the Galerkin
method (see [6] and [8]) and the a priori estimates given by Lemma 1.1. We
note that z ∈ C([0, T ]; H).
Proposition 2.2. The mapping Θ : U → C([0, T ]; H), u → z (a solution
of E(z0 , u, 0)) is differentiable in the Fréchet sense, and Θu .h satisfies the
equation
ẏ(t) + Ay(t) = u(t)By(t) + h(t)Bz(t),
y(0) = 0,
where z = Θ(u).
Proof.
(i) First, we note that the mapping L : h → y is linear. Since y is a
solution of equation E(0, u, hBz), the a priori estimates (E1 ), (E2 ), and
(E3 ) yield
yL2 (V ) ≤ hBzL2 (V
) ≤ K1 hU zL∞ (H)
and
√
ẏL2 (V
) ≤ hBzL2 (V
) (1 + α1 + 2uU ) ≤ K2 hU zL∞ (H) .
Thus,
yC([0,T ];H) ≤ K3 hU .
It follows that L : h ∈ U → y ∈ C([0, T ]; H) is continuous.
(ii) Set
zh = Θ(u + h),
z = zh − z.
146 A. ADDOU and A. BENBRIK

Then z satisfies
ż(t) + Az(t) = u(t)Bz(t) + h(t)Bzh (t),
E(0, u, hBzh )
z(0) = 0.

Since hBzh is in L2 (V  ), the a priori estimate (E2 ) gives

√
zL∞ (H) ≤ 2hBzh L2 (V
) ≤ KhU zh L∞ (H) ;
consequently,
zL∞ (H) ≤ K  hU .
(iii) Now, we set
w = z − y.
Then w satisfies
ẇ(t) + Aw(t) = u(t)Bw(t) + h(t)Bz(t),
E(0, u, hBz)
w(0) = 0.
Since w ∈ C([0, T ]; H), we deduce that
wC([0,T ];H) ≤ Kh2L2(0,T ;RN ) ,
i.e.,
Θ(u + h) − Θ(u) − Θu .hC([0,T ];H) ≤ Kh2L2 (0,T ;RN ) .

3. Optimal control
3.1. Existence.

Proposition 3.1. Assume that z0 ∈ H. Let J the function be defined as

J(u) = z2L2 (H) + u2U ,
where z(t) is the unique solution of the equation
ż(t) + Az(t) = u(t)Bz(t),
z(0) = z0 .
Then there exists a pair (z, u) ∈ C([0, T ]; H) × U such that z is a solution
of E(z0 , u, 0) and u minimizes J in U.

Proof. Let
I = inf J(u),
and let (un )n be a minimizing sequence. Then J(un ) is bounded; it follows
that
un L2 (0,T ;RN ) ≤ const.
 OPTIMAL CONTROL OF A BILINEAR DISTRIBUTED SYSTEM 147

We deduce from the a priori estimates that

zn L2 (V ) ≤ M1 , zn L∞ (H) ≤ M2 ,
A(zn )L2(V
) ≤ M3 , un B(zn )L2 (V
) ≤ M4 ,
żn Lp
(V
) ≤ M5 ,
where Mi are positive constants.
By passing to subsequences if necessary, we assume that:
un " u weakly in L2 (0, T ; RN );
zn " z weakly∗ in L∞ (H);
zn " z weakly in L2 (V );
A(zn ) " χ weakly in L2 V  );
un B(zn ) " Λ weakly in L2 (V  );
żn " Ψ weakly in L2 (V  ).
We now pass to the limit in the state equation
z˙n + Azn (t) = un (t)Bzn (t).
1. Using the convergence σ(D(0, T ; V ); D (0, T ; V  )) (see [7]), we obtain
Ψ = ż.
2. The operator A : L2 (V ) → L2 (V  ), y → Ay(t), is a linear strongly
continuous and, therefore, weakly continuous. It follows that
A(z) = χ.
3. The embedding W → L2 (H) is compact, therefore, the sequence
(zn ) admits a subsequence (again denoted by zn ) which converges strongly
in L2 (H).
∂y
Furthermore, the operator L2 (H) → L2 (V  ) : y → , 1 ≤ i ≤ N , is
∂xi
linear and continuous; thus
∂zn ∂z
→ strongly in L2 (V  ).
∂xi ∂xi
Since un " u weakly in L2 (0, T ; RN ) and un is independent of x, it follows
that
uBz = Λ.
4. By the classical argument (see [5]) we verify that z(0) = z(0). In
conclusion, 
ż + Az(t) = u(t)Bz(t),
z(0) = z0 .
148 A. ADDOU and A. BENBRIK

Now, using the lower semicontinuity of the norms, we deduce that

J(u) ≤ lim inf J(un ) = I;
n

then (z, u) is an optimal pair.

3.2. Adjoint state and optimality conditions.

Proposition 3.2. The function J(u) is differentiable and we have

N    
∂z
∀h ∈ U dJ(u).h = 2 hi , ui + p, ,
i=1
∂xi V
,V L2 (0,T )

where p is a solution of the adjoint equation


 − dp + Ap(t) = −u(t)Bp(t) + z(t),
dt

p(T ) = 0, p ∈ L2 (V ), and ṗ ∈ L2 (V  ).
Remarks.
 ∂z   
∂z
(1) p, denotes the function: [0, T ] → R, t → p(t), (t)
∂xi ∂xi V
,V
in L2 (0, T ), because
∂z
p ∈ L2 (V ) and ∈ L∞ (V  ) (z ∈ L∞ (H)).
∂xi
(2) If we set

q(t) = p(T − t),

v(t) = −u(T − t),
g(t) = z(T − t),
then the adjoint equation becomes
dq
(t) + Aq(t) = v(t)Bq(t) + g(t),
dt
q(0) = 0, q ∈ L2 (V ), q̇ ∈ L2 (V  ).
Therefore, it admits a unique solution.
(3) The differentiability of Θ and the norm ensures the differentiability
of J, and the expression for the derivative is
1 
J (u).h = Θ (u).h, z L2(H) + h, u U.
2
The introduction of the adjoint equation allows us to clarify this derivative
according to h.
 OPTIMAL CONTROL OF A BILINEAR DISTRIBUTED SYSTEM 149

Let p denote the adjoint state; we have

T

Θ (u).h, z L2(H) = y(t), z(t) dt =
0
T
= y(t), −ṗ(t) + Ap(t) + u(t)Bp(t) dt =
0
= −y(T ), p(T ) + y(0), p(0) +
T
+ ẏ(t) + A∗ y(t) + u(t)B ∗ y(t), p(t) dt =
0
T
= h(t)Bz(t), p(t) dt,
0

where y = Θ (u).h, because A∗ = A, B ∗ = −B, y(0) = 0, and p(T ) = 0.

Finally,
1 
J (u).h = hBz, p U + h, u U =
2
N   
∂z
= hi, ui + p, .
i=1
∂xi L2 (0,T )

Corollary 3.1. An optimal control u, the state z, and the corresponding

adjoint state p necessarily satisfy the following relations for 1 ≤ i ≤ N and
t ∈ [0, T ]:
  
 ∂z(t)

 u i (t) = − p(t), ,

 ∂xi V
,V



 N  

 ∂z ∂ ∂

 2
−

 + ∆ z = p(t), z(t) z in Q,

 ∂t
i=1
∂x i ∂x i

N  

 ∂p ∂ ∂

 − + ∆ 2
p = p(t), z(t) p + z(t) in Q,

 ∂t ∂xi ∂xi

 i=1



 z(0) = z0 and p(T ) = 0 in Ω,




 z = ∂z = 0 and p = ∂p = 0 on Σ.
∂ν ∂ν
Proof. This is implied by the result of the preceding proposition and the
fact that the optimal control u minimizing J in U satisfies
∀h ∈ U, dJu .h = 0.
150 A. ADDOU and A. BENBRIK

3.3. Uniqueness of optimal control.

Lemma 3.1. If u1 and u2 are two optimal controls, then

u1 − u2 U ≤ 3|z0 |2 u1 − u2 U .

Proof.
(i) Let z 1 and z 2 be the states corresponding to the optimal controls u1
and u2 , i.e., z j satisfy the equation E(z0 , uj , 0) (j = 1, 2).
The a priori estimates (E1 ) and (E2 ) applied to z j (j = 1, 2) yield
1
z j L2(V ) ≤ √ |z0 |,
2
zj L∞ (H) ≤ |z0 |.

(ii) Let now p1 and p2 be the corresponding adjoint states to u1 and u2 ,

and let

q j (t) = pj (T − t),
vj (t) = −uj (T − t),
gj (t) = z j (T − t).

By using item (2) of Remarks on page 148, we see that q j (j = 1, 2)

satisfy the equation E(0, vj , gj ). Then we can apply the estimates Ei to it;
we have for j = 1, 2:

pj L2(V ) ≤ gj L2(V
) ,

√
pj L∞ (H) ≤ 2gj L2 (V
) .

Thus,
1
pj L2(V ) ≤ √ |z0 |,
2
pj L∞ (H) ≤ |z0 |.

(iii ) It is easy to see that z = z 1 − z 2 satisfies the equation

E(0, u1 , (u1 − u2 )Bz 2 ).

The estimate (E1 ) implies

zL2 (V ) ≤ (u1 − u2 )Bz 2 L2 (V
) ≤ u1 − u2 U z 2 L∞(H) .

Using (i), we have:

zL2 (V ) ≤ u1 − u2 U |z0 |.

 OPTIMAL CONTROL OF A BILINEAR DISTRIBUTED SYSTEM 151

(iv) Now we set

p = p1 − p2 ,
q = q1 − q2 .

Then q satisfies the equation E(0, v1 , (v1 − v2 )Bq 2 + g1 − g2 ). The estimate

(E1 ) applied to q yields

qL2 (V ) ≤ (v1 − v2 )Bq 2 + g1 − g2 L2(V
) ≤

≤ v1 − v2 U q2 L∞ (H) + g1 − g2 L2 (V
) .

Consequently,

pL2 (V ) ≤ u1 − u2 U p2 L∞ (H) + zL2 (V ) ≤

≤ 2|z0 |u1 − u2 U .

In addition, if u1 and u1 are two optimal controls, we have

ui1 (t) − ui2 (t) = −p1 (t), Bi z 1 (t) + p2 (t), Bi z 2 (t) =
= −p1 (t), Bi (z 1 (t) − z 2 (t)) − (p1 (t) − p2 (t)), Bi z 2 (t)

for i = 1, 2, . . . , N . Taking the norm in U, we obtain

 
u1 − u2 U ≤ zL2 (V ) p1 L∞(H) + pL2 (V ) z 2 L∞ (H) .

Using (i), (ii), (iii), and (iv), we deduce

u1 − u2 U ≤ 3|z0 |2 u1 − u2 U .

1
Corollary 3.2. Under the condition |z0 | < √ , the optimal control is
3
unique.

Proof. In this case, we have

u1 − u2 U < u1 − u2 U .

Thus,
u1 = u2 .
Consequently,
z1 = z 2.
152 A. ADDOU and A. BENBRIK

References

1. V. Barbu and Th. Precupanu, Convexity and optimization in Banach

spaces. In: Mathematics and its applications, Reidel Publishing Com-
pany, 1986.
2. A. Benbrik, Contrôle optimal de systèmes à paramètres répartis
bilinéaires. Thèse de doctorat présentée à l’université de Perpignan,
1987.
3. C. Bruni, G. Dipillo, and G. Koch, Bilinear systems. An application
class of “nearly linear” systems in theory and applications. IEEE, Trans.
Automat. Control AC 19 (1974), No. 4.
4. D. Dan Tiba, Optimal control of nonsmooth distributed parameter sys-
tems. Lecture Notes in Math. 1459, Springer-Verlag.
5. A. V. Fursikov, Optimal control of distributed systems. Theory and ap-
plications. Translations of Math. Monogr. 187, 2000.
6. J. L. Lions, Quelques méthodes de résolution des problèmes aux limites
non linéaires. Dunod, 1969.
7. J. L. Lions and E. Magénes, Problèmes aux limites non homgènes et
applications. Vols. 1–3, Dunod, 1968.
8. J. L. Lions, Contrôle optimal des systèmes gouvernés par des équations
aux dérivées partielles. Dunod, 1968.
9. , Contrôle des systèmes distribuès singuliers. Gauthier-Villars.
10. R. R. Mohler, Bilinear control processes. Academic Press, 1975.
11. A discussion on the atteignability and optimal control for class of bi-
linear system. In: Fourth IFAC Symposium on control of distributed
parametersystems, June 30–July 3 (1986) University of California, Los
Angeles, USA.

(Received 25.06.2001, received in revised form 29.12.2001)

Authors’ address:
Université Mohamed I, Faculté des Sciences,
Département de mathématiques et Informatique,
Oujda, Maroc.
E-mails: {addou,benbrik}@sciences.univ-oujda.ac.ma