See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/251655495

Optimality conditions for state-constrained PDE control problems with time-

dependent controls

Article in Control and Cybernetics · January 2008

CITATIONS READS

34 78

4 authors:

Juan Carlos De los Reyes Pedro Martín Merino

Centro de Modelización Matemática (MODEMAT) Escuela Politécnica Nacional
74 PUBLICATIONS 1,753 CITATIONS 18 PUBLICATIONS 115 CITATIONS

SEE PROFILE SEE PROFILE

Joachim Rehberg Fredi Tröltzsch

Weierstrass Institute for Applied Analysis and Stochastics Technische Universität Berlin
88 PUBLICATIONS 1,613 CITATIONS 224 PUBLICATIONS 5,592 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Pedro Martín Merino on 11 August 2014.

The user has requested enhancement of the downloaded file.

 OPTIMALITY CONDITIONS FOR STATE-CONSTRAINED
PDE CONTROL PROBLEMS WITH TIME-DEPENDENT
CONTROLS

J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

Abstract. The paper deals with optimal control problems for semilinear ellip-
tic and parabolic PDEs subject to pointwise state constraints. The main issue
is that the controls are taken from a restricted control space. In the parabolic
case, they are Rm -vector-valued functions of the time, while the are vectors of
Rm in elliptic problems. Under natural assumptions, first- and second-order
sufficient optimality conditions are derived. The main result is the extension
of second-order sufficient conditions to semilinear parabolic equations in do-
mains of arbitrary dimension. In the elliptic case, the problems can be handled
by known results of semi-infinite optimization. Here, different examples are
discussed that exhibit different forms of active sets and where second-order
sufficient conditions are satisfied at the optimal solution.

1. Introduction
This paper is a further contribution to second-order optimality conditions in the
optimal control of partial differential equations. A large number of papers has been
devoted to this issue so far and conditions of this type are used as an essential
assumption in publications on numerical methods.
Sufficient conditions were investigated first for problems with control constraints.
Their main focus was on second-order optimality conditions that are close to the
associated necessary ones, for instance, by Bonnans [5], Casas, Unger and Tröltzsch
[12], Goldberg and Tröltzsch [18]; see also the examples for the control of PDEs
in the recent monograph by Bonnans and Shapiro [6]. The situation changes, if
pointwise state constraints are given. Here, the theory is essentially more difficult
as the Lagrange multipliers associated with the state constraints are Borel measures.
Therefore, the associated theory is less complete than that for control constraints.
Although considerable progress has been made in this issue, cf. Casas, Tröltzsch
and Unger [13], Raymond and Tröltzsch [29], Casas and Mateos [10], Casas, De los
Reyes and Tröltzsch [9], there remain open questions, if pointwise state constraints
are formulated in the whole domain: In the elliptic case of boundary or distributed
control with pointwise state-constraints, the conditions are sufficiently general only
for spatial dimension 2 or 3, respectively, [13, 9]. For parabolic problems, only
distributed controls in one-dimensional domains can be handled in full generality,
[29].
The difficulties mentioned above are intrinsic for problems with pointwise state
constraints and it seems that they cannot be entirely avoided. However, a review
on the application of optimal control problems for partial differential equations
shows that the situation is often easier in practice: In many cases, the control
function depends only on finitely many parameters that may also depend on time
in parabolic problems.
For instance, in all applications the group of the fourth author has been engaged
so far, the controls are finite-dimensional in this sense. This finite-dimensionality
seems to be characteristic for real applications of control theory. This concerns
1
2 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

the cooling of steel profiles by water discussed in [16], [31], [32], some problems of
flow control, cf. Henning and King [22], the sublimation process for the production
of SiC single crystals, cf. [26], and local hyperthermia in tumor therapy, [15]. In
all of these problems, the controls are represented by finitely many real values,
namely, the intensities of finitely many spray nozzles in cooling steel, amplitude
and frequency of controlled suction or blowing in flow control, frequency and power
of the electrical current in sublimation crystal growth, and the energy of finitely
many microwave antennas in local hyperthermia.
In some cases, these finitely many real values depend on time. Moreover, in all
the applications mentioned above, pointwise state constraints are very important.
Problems of this type are the main issue of our paper. We address the following
points:
First, we consider semilinear parabolic equations with distributed controls of the
Pk
type f (x, t) = i=1 ei (x)ui (t), where the functions ei are bounded and the control
functions ui are taken from a set of admissible controls. Thanks to the boundedness
of the fixed functions ei , we are able to extend a result of [9] for one-dimensional
parabolic problems to domains Ω of arbitrary dimension. Although our regularity
results can be extended to controls f ∈ L2 (0, T ; L∞ (Ω)), we need the restriction to
the finite-dimensional ansatz above to deal with second-order sufficient optimality
conditions, cf. Remark 2.
Moreover, we consider different problems with finite-dimensional controls u ∈
Rm . If the control is a vector of Rm , pointwise state constraints generate a semi-
infinite optimization problem. The associated constraint set is infinite by its na-
ture. The Lagrange multipliers for the state-constraints remain to be Borel mea-
sures. Therefore, this class of problems is sufficiently interesting for the numerical
analysis. This setting belongs to the class of semi-infinite optimization problems.
Here, we are able to invoke the known theory of first- and second-order optimality
conditions. In the case of partial differential equations, this theory needs special
adaptations that are briefly sketched. Moreover, we present different examples
of state-constrained control problems, where second-order sufficient conditions are
satisfied at the optimal solution. These problems can be used to test numerical
methods and show how diverse the active set may look like.

2. Semilinear parabolic problems

2.1. Problem statement. We consider the following distributed optimal control
problem with time-dependent control functions,
 Z Z

 min J(u) = L(x, t, y(x, t), u(t)) dxdt + `(x, t, y(x, t)) dS(x)dt


 u∈Uad Q Σ

 Z
+ r(x, y(x, T )) dx






 Ω


subject to

(P1) (2.1)
 y + A y(x, t) + d(x, t, y(x, t)) = Pm e (x) u (t) in Q,



 t i=1 i i



 ∂ ν y(x, t) = 0 on Σ,



 y(x, 0) − y 0 (x) = 0 in Ω,



g(x, t, y(x, t)) 6 0 for all (x, t) ∈ K ⊂ Q̄,


where u = (u1 , . . . , um )> and Uad is defined by

Uad = {u ∈ L∞ (0, T ; Rm ) : ua (t) ≤ u(t) ≤ ub (t) a.e. t ∈ [0, T ]}.

 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 3

In this setting, Ω is a subset of Rn (n > 1) with boundary Γ, we have set Q =

Ω × (0, T ), Σ = Γ × (0, T ), A is a second-order elliptic differential operator, K is
a non-empty compact subset of Q̄, and T > 0 is a fixed real number. Moreover,
functions L : Q × R × Rm → R, d : Q × R → R, ` : Σ × R → R, r : Ω × R → R,
g : K × R → R, an initial state y0 ∈ C(Ω̄), fixed functions ei : Ω → R, i = 1, . . . , m,
and fixed bounds ua , ub ∈ L∞ (0, T ; Rm ) are given such that ua (t) ≤ ub (t) holds
a.e. on (0, T ) in the componentwise sense. The symbol ∂ν denotes the derivative
in the direction of the outward unit normal ν at Γ.

2.2. Main assumptions and well-posedness of the state equation. In the

parabolic case, we rely on the following assumptions:

(A.1) The set Ω ⊂ Rn is an open and bounded Lipschitz domain in the sense of
Nečas [28]. The differential operator A is defined by
n
X
Ay(x) = − ∂xj (aij (x)∂xi y(x))
i,j=1

with coefficients aij ∈ L∞ (Ω) satisfying

n
X
λA |ξ|2 ≤ aij (x)ξi ξj ∀ξ ∈ Rn , ∀ x ∈ Ω
i,j=1

for some λA > 0.

(A.2) (Carathèodory type assumption) For each fixed pair (x, t) ∈ Q = Ω × (0, T )
or Σ = Γ×(0, T ), respectively, the functions d = d(x, t, y) and ` = `(x, t, y) are twice
partially differentiable with respect to y. For all fixed y ∈ R, they are Lebesgue
measurable with respect to (x, t) ∈ Q, or x ∈ Σ respectively.
Analogously, for each fixed pair (x, t) ∈ Q, L = L(x, t, y, u) is twice partially
differentiable with respect to (y, u) ∈ Rm+1 . For all fixed (y, u) ∈ Rm+1 , L is
Lebesgue measurable with respect to (x, t) ∈ Q.
The function g = g(x, t, y) is supposed to be twice continuously differentiable
∂g ∂2g
with respect to y on K × R, i.e. g, , and are continuous on K × R.
∂y ∂y 2

(A.3) (Monotonicity) For almost all (x, t) ∈ Q or (x, t) ∈ Σ, respectively, and all
y ∈ R it holds that
∂d
(x, t, y) ≥ 0.
∂y

(A.4) (Boundedness and Lipschitz properties) The functions ei , i = 1, . . . , m, be-

long to L∞ (Ω). There is a constant C0 and, for all M > 0, a constant CL (M ) such
that the estimates
∂d ∂2d
|d(x, t, 0)| + | (x, t, 0)| + | 2 (x, t, 0)| 6 C0
∂y ∂y
∂2d ∂2d
| 2 (x, t, y1 ) − 2 (x, t, y2 )| 6 CL (M ) |y1 − y2 |
∂y ∂y
hold for almost all (x, t) ∈ Q and all |yi | 6 M , i = 1, 2. The functions `, and g
are assumed to satisfy these boundedness and Lipschitz properties on Σ and Q,
4 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

respectively. The function r is assumed to obey these assumptions with x ∈ Ω

substituted for (x, t) ∈ Q. Analogously,
|L(x, t, 0, 0)| + |L0 (x, t, 0, 0)| + |L00 (x, t, 0, 0)| 6 C0
|L00 (x, t, y1 , u1 ) − L00 (x, t, y2 , u2 )| 6 CL (M ) (|y1 − y2 | + |u1 − u2 |)
hold for almost all (x, t) ∈ Q and all |yi | 6 M , |ui | 6 M , i = 1, 2. Here, L0 and L00
denote the gradient and the Hessian matrix of L with respect to (y, u) ∈ Rm+1 .

2.3. Well-posedness of the state equation.

2.3.1. State equation and existence of optimal controls. In this subsection, we show
that the control-to-state mapping G : u 7→ y is well defined for all admissible
controls u. Moreover, we show certain continuity properties of G. Later, we will
also discuss the differentiability of G. In all what follows, we denote the state
function y associated with u by yu , i.e. yu = G(u).
Theorem 1. Suppose that the assumptions (A.1) – (A.4) are satisfied. Then,
for every u ∈ Lq (0, T ; Rm ) with q > n/2 + 1, the state equation (2.1) has a unique
solution yu ∈ C(Q̄)∩W (0, T ). If uk * u weakly in Lq̃ (0, T ; Rm ) with q̃ = max{q, 2},
then yuk → yu strongly in C(Q̄).
Proof. The functions ei are bounded, hence the right-hand side
m
X
v(x, t) = ei (x) ui (t)
i=1

of equation (2.1) belongs to Lq (Q) with q > n2 + 1. Therefore, existence, uniqueness

and continuity of the solution y ∈ W (0, T ) of (2.1) follow from Casas [8]. It remains
to show that uk * u weakly in Lq̃ (0, T, Rm ) implies yuk → yu strongly in C(Q̄).
To this aim, we first re-write the parabolic equation as
dyk Pm
+ A yk = vk := i=1 ei (x) uk,i (t) − d(x, t, yk ) in Q,
dt (2.2)
∂ν yk = 0 on Σ,
yk (x, 0) = y0 (x) in Ω,
where we introduced yk := yuk . The sequence of right-hand sides (vk ) is bounded
in Lq̃ (Q). Therefore, the boundedness of the sequence yk in C(Q̄) is a standard
conclusion, [8], so that (d(x, t, yk )) is bounded in L∞ (Q). In view of this, the
sequence (vk ) is bounded in L2 (0, T ; Lp (Ω)), where p is taken sufficiently large to
meet the assumptions of Theorem 3. Therefore, (vk ) contains a sub-sequence (vkl )
with vkl * v weakly in L2 (0, T ; Lp (Ω)), l → ∞,
W.l.o.g we can assume y0 = 0, since the solution ŷ of ŷt + Aŷ = 0, ∂ν ŷ = 0 and
ŷ(x, 0) = y0 (x) can be subtracted from the sequence yk . This fixed function does
not influence the convergence properties.
Thanks to Theorem 3, the mapping vk 7→ yk defined by (2.2) is linear and
continuous from L2 (0, T ; Lp (Ω)) to the Hölder space C α (Q) for some α > 0. By
the compactness of the injection C α (Q) ,→ C(Q̄), (ykl ) converges uniformly to some
in L∞ (Q) ,→ L2 (0, T ; Lp (Ω)) so
y ∈ C(Q̄). This, implies (d(x, t, ykl )) → (d(x, t, y))P
m
that the right-hand side converges weakly to v = i=1 ei ui − d(·, y). Therefore, y
is associated with this function v and solves the semilinear equation with control
u. By uniqueness, it must hold y = yu , hence the same result is obtained for any
subsequence of (vk ), and a standard argument yields yk → y in C(Q̄) as k → ∞. 
It is obvious that only those parts of the assumption (A.4) are needed in the
theorem that are related to d and dy .
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 5

Theorem 2. Assume that the assumptions (A.1) – (A.4) are fulfilled, the function
L = L(x, t, y, u) is convex with respect to u ∈ Rm and the set of feasible controls is
nonempty. Then the control problem (P1) has at least one solution.
This theorem is a standard consequence of Theorem 1 and the lower semiconti-
nuity of J that needs the convexity of L with respect to u.

2.3.2. Hölder regularity for linear parabolic equations. The results of this subsection
are needed for the extension of second-order sufficient conditions to the case of n-
dimensional domains, if the controls depend only on time. To derive second-order
sufficient conditions, linearized equations are considered for L2 -controls, and the
regularity of associated states is derived in Theorem 3, which has already been used
in the proof of Theorem 1. The theorem is proved by recent results on maximal
parabolic regularity. Throughout this section, Ω is allowed to be a Lipschitz domain
in the sense of Grisvard [21]. This is more general than domains with Lipschitz
boundary in the sense of Nečas [28].
We consider the linear parabolic problem
dz
+ A z + c0 z = v in Q,
dt (2.3)
∂ν z = 0 in Σ,
z(x, 0) = 0 in Ω,
where c0 ∈ L∞ (Q) is given fixed. Later, c0 stands for the partial derivative ∂d/∂y
taken at the optimal state. To deal with equation (2.3), we provide some basic facts
on maximal parabolic regularity.
We denote by A ∈ L∞ (Ω; Rn×n ) the matrix function A(x) = (aij (x)), with aij ,
i, j ∈ {1, . . . , n} defined in (A.1). Associated with our differential operator A, the
linear and continuous operator −∇ · A∇ : H 1 (Ω) → H 1 (Ω)0 is defined by
Z
< −∇ · A∇v, w >:= A ∇v · ∇w dx ∀ v, w ∈ H 1 (Ω). (2.4)
Ω

The restriction of this operator to the spaces Lp (Ω) with p ≥ 2 is denoted by Ap .

Its domain is given by
D(Ap ) = {y ∈ H 1 (Ω) : A y ∈ Lp (Ω)}
equipped with the associated graph norm. It is known that Ap incorporates a
(homogeneous) Neumann boundary condition, see [17] Ch. II.2 or [14] Ch. 1.2.
Let us next recall the concept of maximal regularity: Let X be a Banach space
and A denote a closed operator with dense domain D ⊂ X equipped with the
graph norm. Moreover, let S = (T0 , T ) ⊂ R be a bounded interval. Suppose
r ∈ (1, ∞). Then A is said to satisfy maximal parabolic Lr (S; X)-regularity iff
for any f ∈ Lr (S; X) there is a unique function w ∈ W 1,r (S; X) ∩ Lr (S; D) that
satisfies
∂w
+ Aw = f, w(T0 ) = 0. (2.5)
∂t
By W 1,r (S; X), we denote the set of those elements from Lr (S; X) whose distribu-
tional derivative also belongs to Lr (S; X). If X, Y are Banach spaces which form
an interpolation couple, then we denote by [X, Y ]θ the corresponding complex in-
terpolation space and by (X, Y )θ,r the real interpolation space, see [30]. Below we
well employ the following fact: There is a continuous injection
E : W 1,r (S; X) ∩ Lr (S; D) ,→ C(S̄; (X, D)1− r1 ,r ), (2.6)
see [4] Ch. III Thm. 4.10.2, cf. also [30] Ch. 1.8. Moreover, for any η ∈ (0, 1 −
1
r ) there is a continuous embedding (X, D)1− r ,r ,→ [X, D]η and, consequently, a
1
6 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

continuous embedding
C(S̄; (X, D)1− r1 ,r ) ,→ C(S̄; [X, D]η ). (2.7)
Next, we need the following
Lemma 1. Assume τ ∈ (0, 1 − 1r ). Then there is an index κ > 0 such that
W 1,r (S; X) ∩ Lr (S; D) continuously embeds into C κ (S; [X, D]τ ).
Proof. First of all, the estimate
Z t Z t  r1 Z t  r−1
r
kw(t) − w(t0 )kX = k w0 (s) dskX ≤ kw0 (s)krX ds ds ≤
t0 t0 t0
r−1
≤ kwkW 1,r (S;X) |t − t0 | r

gives us a (continuous) embedding from W 1,r (S; X) into C δ (S; X), where δ = r−1r .
Let η be a number from (τ, 1− 1r ). Then, putting λ = τη we obtain by the reiteration
theorem for complex interpolation, see [30] Ch. 1.9.3,
kw(t) − w(s)k[X,D]τ kw(t) − w(s)k[X,[X,D]η ]λ
δ(1−λ)
≤c ≤
|t − s| |t − s|δ(1−λ)
kw(t) − w(s)k1−λ
≤c X
kw(t) − w(s)kλ[X,D]η ≤
|t − s|δ(1−λ)
 kw(t) − w(s)k 1−λ  λ
X
≤ c1 δ
2kwkC(S̄;[X,D]η ) ≤ c2 kwkW 1,r (S;X)∩Lr (S;D) .
|t − s|

Corollary 1. Assume that A satisfies maximal parabolic Lr (S; X)-regularity and
let τ ∈ (0, 1 − 1r ) be given. Then the mapping that assigns to every f ∈ Lr (S; X)
the solution of (2.5) is continuous from Lr (S; X) into C κ (S; [X, D]τ ) for a certain
κ > 0.
Proof. The mapping
∂
+ A : W 1,r (S; X) ∩ Lr (S; D) ∩ {w | w(T0 ) = 0} → Lr (S; X)
∂t
is continuous and bijective. Hence, the inverse is also continuous by the open
mapping theorem. Combining this with the preceding lemma gives the assertion.

Now we are able to show our main result on parabolic regularity: Hölder regu-
larity can be obtained for functions which exhibit only to L2 -regularity in time, if
their spatial regularity is Lp and p is sufficiently large.
Theorem 3. If f belongs to Lr (S; Lp (Ω)) with r > 1 and sufficiently large p, then
the solution w of
∂w
+ Ap w = f, w(0) = 0 (2.8)
∂t
is from a space C κ (S; C β (Ω)). Moreover, the mapping f 7→ w is continuous from
Lr (S; Lp (Ω)) to C κ (S; C β (Ω)).
Proof. We apply the general results on maximal parabolic regularity to our operator
Ap . It is known that Ap enjoys maximal parabolic Lr (S; Lp (Ω))-regularity for
every p ∈ (1, ∞) and r > 1. We refer to [19], Thm. 7.4. Moreover, the following
interpolation result is known: If θ ∈ (0, 1) and
n
β := θα − (1 − θ) > 0, (2.9)
p
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 7

then
[Lp (Ω), C α (Ω)]θ ,→ C β (Ω). (2.10)
The result is shown in [19], see the proof of Thm. 7.1. In general, the domain of Ap
is difficult to determine. However, the following embedding result is helpful: For
p > n2 , there exists α > 0 such that the continuous embedding

dom(Ap ) ,→ C α (Ω) (2.11)

holds true. This result is proved in [20]. Keeping α > 0 fixed we can increase p
so that also (2.9) is satisfied. Clearly, (2.10) and (2.11) remain true. Taking now
into account Corollary 1 for X = Lp (Ω), then the assertion follows from (2.10) and
(2.11). 

Corollary 2. If r > 1 and p is sufficiently large, then for all v ∈ Lr (S; Lp (Ω)), the
weak solution of (2.3) belongs to C α (Q̄) with some α > 0. The mapping v 7→ z is
continuous from Lr (0, T ; Lp (Ω)) to C α (Q̄).
∂
Proof. The operator ∂t + Ap is a topological isomorphism between Lr (S; D(Ap )) ∩
W (S; L (Ω)) ∩ {z : z(T0 ) = 0} and Lr (S; Lp (Ω)), and is therefore a Fredholm
1,r p

operator of index zero. Obviously, the multiplication operator induced by c0 is

∂
bounded on L2 (S; Lp (Ω)); hence, the domain of ∂t + Ap + c0 equals the domain of
∂
∂t +A p – which, due to Theorem 3, compactly embeds into Lr (S; Lp (Ω)). Thus, the
∂
multiplication operator induced by c0 is relatively compact with respect to ∂t + Ap .
∂
By a well known perturbation theorem, ∂t + Ap + c0 then also must be a Fredholm
operator and is also of index zero (see Kato [23], Ch. IV.5.3). Let us show that it
is injective: Let z be a solution of ∂z ∂z
∂t + Ap z + c0 z = 0 or, equivalently, ∂t + Ap z =
−c0 z. As a solution of this parabolicR t equation with right hand side −c0 z, z has
then the representation z(t) = − 0 e(s−t)Ap c0 (s, ·)z(s) ds. But the semigroup
e−tAp is contractive on Lp (Ω) (see [19] Thm. 4.11) and c0 is essentially bounded;
∂
thus Gronwall’s lemma yields z ≡ 0. Therefore, ∂t + Ap + c0 is injective. Because
it is Fredholm and, additionally, of index zero, it is also surjective. Consequently,
∂
the inverse operator maps Lr (S; Lp (Ω)) continuously into the domain of ∂t + Ap ,
and this is continuously embedded into a space C α (Q̄). 

Notice that in particular L2 (0, T ; L∞ (Ω)) is mapped continuously into a space

α
C (Q̄). This is what we need for the discussion of second-order sufficient conditions.
The theorem refers to the PDE (2.3) with homogeneous initial condition. It is
obvious that the result extends to inhomogeneous Hölder continuous initial data.

Remark 1. The Hölder regularity might also be deduced from Ladyzhenskaya et

al. [25] under the assumption that Γ is a Lipschitz boundary in the sense of Nečas
[28], cf. also [8]. Under this assumption, the continuity of the state function for our
restricted class of controls was also shown in [29]. The use of maximal parabolic
regularity permits to extend these results to Lipschitz domains in the sense of Gris-
vard [21]. Moreover, this approach is not restricted to Neumann conditions. It also
allows for mixed boundary conditions and, of course, for Dirichlet conditions. Last,
but not least, our approach essentially shortens and unifies the associated proofs.

2.4. Necessary optimality conditions. The control-to-state mapping G(u) =

yu , G : L∞ (0, T ; Rm ) → C(Q̄) ∩ W (0, T ), and the reduced objective functional
J are of class C 2 from L∞ (0, T ; Rm ) to their image spaces, provided that the
assumptions (A.1)-(A.4) are satisfied. This follows by the arguments of [8].
8 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

We define, for v ∈ L∞ (0, T ; Rm ), the function zv as the unique solution to

dzv ∂d
 Pm
 dt + Azv + ∂y (x, t, yu )zv = i=1 ei (x)vi (t) in Q,


∂ν zv = 0 in Σ, (2.12)



y(x, 0) = 0 in Ω.
Then G (u), G : L (0, T ; R ) → C(Q̄) ∩ W (0, T ) is given by G0 (u)v = zv . More-
0 ∞ m

over, for v1 , v2 ∈ L∞ (0, T ; Rm ), we introduce zvi = G0 (u)vi , i = 1, 2, and obtain

G00 (u)v1 v2 = zv1 v2 , where zv1 v2 is the solution to
dzv1 v2 ∂d ∂2d


 + Azv1 v2 + (x, t, yu )zv1 v2 + 2 (x, t, yu )zv1 zv2 = 0 in Q,
 dt ∂y ∂y
 ∂ν zv1 v2 = 0 in Σ,


zv1 v2 (x, 0) = 0 in Ω.
(2.13)
The adjoint state ϕ0u ∈ W (0, T ) associated with u and J is introduced as the
unique solution to
dϕ ∂d ∂L


 − + A∗ ϕ + (x, t, yu )ϕ = (x, t, yu , u) in Q,



 dt ∂y ∂y
 ∂`
∂ν ϕ = (x, t, yu ) in Σ, (2.14)

 ∂y


 ∂r

 ϕ(x, T ) = (x, yu (x, T )) in Ω,
∂y
where A∗ is the formally adjoint operator to A. Standard computations show
Z TX m Z   
∂L
J 0 (u)v = (x, t, yu , u) + ϕ0u ei (x) dx vi (t) dt, (2.15)
0 i=1 Ω ∂ui
Z  2
00 ∂ L ∂2L
J (u)v1 v2 = 2
(x, t, y u , u)z v 1 z v 2 + (x, t, yu , u) · (zv1 v2 + zv2 v1 )
Q ∂y ∂y∂u
∂2L ∂2d

+v1> 2 (x, t, yu , u)v2 − ϕ0u 2 (x, t, yu )zv1 zv2 dxdt,
∂u ∂y
Z 2
∂ `
+ 2
(x, t, yu )zv1 zv2 dS dt
Σ ∂y
∂2r
Z
+ 2
(x, yu (x, T ))zv1 (x, T )zv2 (x, T ) dx.
Ω ∂y
(2.16)
As in (2.16), we will use a dot to denote the inner product of Rm . Notice that the
zvi and ϕ0u depend on (x, t), while the vi depend on t only.
We require the following linearized Slater condition: There exists a function
u0 ∈ L∞ (0, T ; Rm ) with ua (t) ≤ u0 (t) ≤ ub (t) for a.e. t ∈ (0, T ) such that
∂g
g(x, t, ȳ(x, t)) + (x, t, ȳ(x, t))zu0 −ū (x, t) < 0 ∀(x, t) ∈ K. (2.17)
∂y
Inserting t = 0, this imposes a condition on the initial state y0 . Therefore, to satisfy
(2.17), we have to assume that
g(x, 0, y0 (x)) < 0 ∀x ∈ Ω̄ with (x, 0) ∈ K. (2.18)
Defining the Hamiltonian H by
m
hX i
H(x, t, y, u, ϕ) = L(x, t, y, u) + ϕ ei (x) ui − d(x, t, y) ,
i=1
the first-order necessary conditions admit the following form:
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 9

Theorem 4. Let ū be a local solution of (P1). Suppose that the assumptions (A.1)
– (A.4) hold and assume the Slater condition (2.17) with some u0 ∈ L∞ (0, T ; Rm ),
ua (t) ≤ u0 (t) ≤ ub (t) for a.e. t ∈ (0, T ). Then there exists a measure µ̄ ∈ M (K)
and a function ϕ̄ ∈ Ls (0, T ; W 1,σ (Ω)) for all s, σ ∈ [1, 2) with (2/s) + (n/σ) > n + 1
such that
 − dϕ̄ + A∗ ϕ̄ + ∂d (x, t, ȳ) ϕ̄ = ∂L (x, t, ȳ, ū) + ∂g (x, t, ȳ)µ̄ ,

 |Q



 dt ∂y ∂y ∂y
 ∂` ∂g
∂ν ϕ̄(x, t) = (x, t, yu (x, t)) + (x, t, ȳ(x, t))µ̄|Σ ,

 ∂y ∂y


 ∂r ∂g

 ϕ̄(x, T ) = (x, ȳ(x, T )) + (x, T, ȳ(x, T ))µ̄|Ω×{T }
∂y ∂y
(2.19)
for a.a. x ∈ Ω, t ∈ (0, T ), where µ̄|Q , µ̄|Σ , and µ̄|Ω×{T } denote the restrictions of µ̄
to Q, Σ, and Ω × {T }, respectively,
Z
(z(x, t) − g(x, t, ȳ(x, t))dµ̄(x, t) ≤ 0 ∀z ∈ C(K) with z(x, t) ≤ 0 ∀(x, t) ∈ K,
K
(2.20)
and, for almost all t ∈ (0, T ),
Z
∂H
(x, t, ȳ(x, t), ū(t), ϕ̄(x, t)) dx · (u − ū(t)) ≥ 0 ∀ u ∈ [ua (t), ub (t)]. (2.21)
Ω ∂u

This theorem follows from Casas [8] using the admissible set
m
X
Ũad = {v | v = ei (·)ui (·), ua,i (t) ≤ ui (t) ≤ ub,i (t) a.e. on (0, T )}
i=1

and writing the associated variational inequality in terms of u. The inequality

(2.20) implies the well-known complementary slackness condition
Z
g(x, t, ȳ(x, t))dµ̄(x, t) = 0. (2.22)
K

Notice that ∂H/∂u is a m−vector function. Since we need the integrated form of
H and its derivatives at the optimal point, we introduce the vector function
Z
∂H
Hu (t) = (x, t, ȳ(x, t), ū(t), ϕ̄(x, t)) dx (2.23)
Ω ∂u

and the (m, m)-matrix valued function

Z 2
∂ H
Huu (t) = 2
(x, t, ȳ(x, t), ū(t), ϕ̄(x, t)) dx (2.24)
Ω ∂u

with entries
Hui uj (t) := (H̄uu )ij (t), i, j ∈ {1, . . . , m}.
The (reduced) Lagrange function is defined in a standard way by
Z Z
L(u, µ) = L(x, t, yu (x, t), u(t)) dxdt + `(x, t, yu (x, t)) dSdt
QZ Z Σ

+ r(x, yu (x, T )) dx + g(x, t, yu (x, t)) dµ(x, t).

Ω K

For later use, we establish the second-order derivative of L with increments vi ∈

L∞ (0, T ; Rm ), i = 1, 2. The expressions below contain functions zvi (x, t) and vi (t),
10 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

but we suppress their arguments for short.

∂2L
Z  2
∂ L
2
(u, µ)v1 v2 = 2
(x, t, yu , u)zv1 zv2
∂u Q ∂y

∂2L ∂2L
+ (x, t, yu , u) · (zv1 v2 + zv2 v1 ) + v1> (x, t, yu , u)v2
∂y∂u ∂u2
∂2d ∂2r
 Z
−ϕu 2 (x, t, yu )zv1 zv2 dxdt + 2
(x, yu (x, T ))zv1 (x, T )zv2 (x, T ) dx
∂y Ω ∂y
Z T 2
∂2g
Z
∂ `
+ 2
(x, t, y u )z v 1 z v 2 dt + 2
(x, t, yu )zv1 zv2 dµ̄(x, t),
0 ∂y K ∂y
(2.25)
where ϕu is the solution of (2.19) with u taken for ū, yu for ȳ, and µ for µ̄,
respectively.

2.5. Second-order sufficient optimality conditions. The regularity results of

the preceding section at hand, we are able to discuss second-order sufficient condi-
tions for arbitrary dimensions of Ω. To establish them and to show their sufficiency,
we follow the recent paper [9] by Casas et al. . The presentation here is similar to
the one in [9], but there occur some differences due to the appearance of vector-
valued control functions. We do not assume that the reader is familiar with the
results of [9]. Therefore, we sketch the main steps of the analysis. For an easier
comparison with the arguments presented in [9], we adopt also part of the notation
used there.
First, we define the cone of critical directions associated with ū by
Cū = {h ∈ L2 (0, T ; Rm ) : h satisfies (2.26), (2.27) and (2.28) below},

 ≥ 0 if ūi (t) = ua,i (t),
∀i ∈ {1, . . . , m}, hi (t) = ≤ 0 if ūi (t) = ub,i (t), (2.26)
= 0 if Hu,i (t) 6= 0,

∂g
(x, t, ȳ(x, t))zh (x, t) ≤ 0 if g(x, t, ȳ(x, t)) = 0, (2.27)
∂y
Z
∂g
(x, t, ȳ(x, t))zh (x, t) dµ̄(x, t) = 0. (2.28)
K ∂y

The vector function Hu was defined in (2.23).

Moreover, we define, for fixed τ > 0 and all i ∈ {1, . . . , m}, the sets of ”suffi-
ciently active control constraints”
Eiτ = {t ∈ [0, T ] : |Hu,i (t)| ≥ τ }. (2.29)
In the next theorem, we write diag(χEiτ (t)) for the matrix diag(χE1τ (t), . . . , χEm
τ (t)).

The second-order sufficient optimality conditions for ū are stated in the following
result:
Theorem 5. Let ū be a feasible control of problem (P1) that satisfies, together with
the associated state ȳ and (ϕ̄, µ̄) ∈ Ls (0, T ; W 1,σ (Ω)) × M (K) for all s, σ ∈ [1, 2)
with (2/s) + (n/σ) > n + 1, the first-order conditions (2.19)-(2.21). Assume in
addition that there exist constants ω > 0, α0 > 0, and τ > 0 such that for all
α > α0
 
d> Huu (t) + α diag(χEiτ (t)) d ≥ ω|d|2 a.e. t ∈ [0, T ], ∀ d ∈ Rm , (2.30)
∂2L
(ū, µ̄)h2 > 0 ∀h ∈ Cū \ {0}. (2.31)
∂u2
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 11

Then there exist ε > 0 and δ > 0 such that, for every admissible control u of
problem (P1), the following inequality holds
δ
J(ū) + ku − ūk2L2 (0,T ;Rm ) ≤ J(u) if ku − ūkL∞ (0,T ;Rm ) < ε. (2.32)
2
Proof. The proof is by contradiction. It follows the one presented in Casas et al.
[9] for an elliptic control problem. Nevertheless, we sketch the main steps, since
there are some essential changes due to the different nature of our problem.
Suppose that ū does not satisfy the quadratic growth condition (2.32). Then
there exists a sequence {uk }∞ ∞ m
k=1 ⊂ L (0, T ; R ) of feasible controls for (P1) such
∞ m
that uk → ū in L (0, T ; R ) and
1
J(ū) +kuk − ūk2L2 (0,T ;Rm ) > J(uk ) ∀k. (2.33)
k
Define ρk = kuk − ūkL2 (0,T ;Rm ) and
1
hk = kuk − ūkL2 (0,T ;Rm ) .
ρk
Since khk kL2 (0,T ;Rm ) = 1, a weakly converging subsequence can be extracted.
W.l.o.g. we can assume that hk * h weakly in L2 (0, T ; Rm ), k → ∞. Now,
the proof is split into several steps.
Step 1: It is shown that
∂L
(ū, µ̄)h = 0. (2.34)
∂u
The arguments are the same as in [9]. Moreover, they are analogous to the
classical proof for finite-dimensional problems. Therefore, we omit them.
Step 2: h ∈ Cū . We have to confirm (2.26)–(2.28). It is easy to verify that h
satisfies the sign conditions of (2.26). To see that hi (t) vanishes, where Hu,i (t) 6= 0,
we notice that (2.21) implies after a standard discussion that, for all i ∈ {1, . . . , m}
|Hu,i (t)hi (t)| = Hu,i (t)hi (t) ≥ 0 for a.a. t ∈ [0, T ]. (2.35)
Therefore,
Z m
T X Z T
∂L
|Hu,i (t) hi (t)| dt = Hu (t) · h(t) dt = (ū, µ̄)h = 0
0 i=1 0 ∂u
must hold in view of (2.34). This implies hi (t) = 0 if Hu,i (t) 6= 0, hence (2.26) is
verified.
The proof of (2.27) is fairly standard and follows from
yū+ρk hk − ȳ
zh = G0 (ū)h = lim in C(Q̄),
k→∞ ρk
and from g(x, t, yū+ρk hk (x, t)) = g(x, t, yk (x, t)) ≤ 0 for every (x, t) ∈ K, since uk
is feasible. Notice that g(x, t, ȳ(x, t)) = 0 holds for all (x, t) considered in (2.27).
It remains to show (2.28). We get from the complementary slackness condition
(2.22) that
Z Z
∂g 1  
(·, ȳ)zh dµ̄ = lim g(·, yū+ρk hk ) − g(·, ȳ) dµ̄ =
K ∂y k→∞ ρk K
Z
1
= lim g(·, yuk ) dµ̄ ≤ 0, (2.36)
k→∞ ρk K

since µ̄ ≥ 0 and g(x, t, yuk (x, t)) ≤ 0 on K. On the other hand, (2.33) yields
J(ū + ρk hk ) − J(ū) J(uk ) − J(ū) ρk
J 0 (ū)h = lim = lim ≤ lim = 0. (2.37)
k→∞ ρk k→∞ ρk k→∞ k
12 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

Now, (2.34), (2.36), (2.37) and the simple identity

Z
∂L 0 ∂g
(ū, µ̄)h = J (ū)h + (x, t, ȳ(x, t))zh (x, t) dµ̄(x, t)
∂u K ∂y
imply that Z
0 ∂g
J (ū)h = (x, t, ȳ(x, t))zh (x, t) dµ̄(x, t) = 0,
K ∂y
hence (2.28) holds and we have shown h ∈ Cū .
Step 3: h = 0. In view of (2.31), it suffices to show
∂2L
h>
(ū, µ̄)h ≤ 0. (2.38)
∂u2
For this purpose, we perform a second-order Taylor expansion of the Lagrangian,
∂L ρ2 ∂ 2 L
L(uk , µ̄) = L(ū, µ̄) + ρk (ū, µ̄)hk + k (wk , µ̄)h2k , (2.39)
∂u 2 ∂u2
where wk is an intermediate point between ū and uk . Re-writing this equation, we
get
∂L ρ2 ∂ 2 L
ρk (ū, µ̄)hk + k (ū, µ̄)h2k
∂u 2 ∂u2
ρ2k ∂ 2 L ∂2L
 
= L(uk , µ̄) − L(ū, µ̄) + (ū, µ̄) − (wk , µ̄) h2k . (2.40)
2 ∂u2 ∂u2
Moreover, we mention that (2.33) can be written in terms of L as
ρ2k
L(uk , µ̄) − L(ū, µ̄) ≤ . (2.41)
k
Taking into account the expression (2.25) of the second derivative of the Lagrangian,
the assumptions (A.1)-(A.4), Theorem 1, (2.12) and the fact that uk → ū in
L∞ (0, T ; Rm ) and khk kL2 (0,T ;Rm ) = 1, we get that
 2
∂2L ∂2L ∂2L

∂ L 2
(ū, µ̄) − (wk , µ̄) h k ≤ (ū, µ̄) − (wk , µ̄) khk k2L2 (0,T ;Rm )
∂u2 ∂u2 ∂u2 ∂u2 B(L2 (0,T ;Rm ))

∂2L ∂2L
= 2
(ū, µ̄) − (wk , µ̄) →0 when k → ∞, (2.42)
∂u ∂u2 B(L2 (0,T ;Rm ))
where B(L2 (0, T ; Rm )) is the space of quadratic forms in L2 (0, T ; Rm ).
From (2.35) and the definition of hk we know that Hu,i (t) hk,i (t) ≥ 0 for a.a.
t ∈ [0, T ] and all i ∈ {1, . . . , m}, therefore
Z T m Z
∂L X
(ū, µ̄)hk = Hu (t) · hk (t) dt ≥ |Hu,i (t)||hk,i (t)| dt
∂u 0 τ
i=1 Ei
Xm Z
≥ τ |hk,i (t)| dt. (2.43)
i=1 Eiτ

For any ε > 0, there exists kε such that

kρk hk kL∞ (0,T ;Rm ) = kū − uk kL∞ (0,T ;Rm ) < ε ∀k ≥ kε ,
therefore
ρ2k h2k,i (t)
≤ ρk |hk,i (t)| ∀k ≥ kε , a.e. t ∈ [0, T ].
ε
From this inequality and (2.43) it follows that
m Z m Z
∂L X ρ2 τ X
ρk (ū, µ̄)hk ≥ ρk τ |hk,i (t)| dt ≥ k h2k,i (t) dt. (2.44)
∂u i=1 Ei
τ ε i=1 Ei
τ
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 13

Collecting (2.40)-(2.42) and (2.44) and dividing by ρ2k /2 we obtain for any k ≥ kε
m Z
2τ X ∂2L 2 ∂2L ∂2L
h2k,i (t) dt + 2 (ū, µ̄)h2k ≤ + 2
(ū, µ̄) − (wk , µ̄) .
ε i=1 Eiτ ∂u k ∂u ∂u2 B(L2 (0,T ;Rm ))
(2.45)
Let us consider the left-hand side of this inequality. First of all we notice that from
(2.25) and Z 2
∂ L
Huu (t) = 2
(x, t, ȳ(x, t), ū(t)) dx (2.46)
Ω ∂u
it follows that
Z T
∂2L
Z 
2 >
(ū, µ̄)h k = h k (t) H uu (t)h k (t) + 2 L̄ uy (x, t)z hk (x, t) dx · h k (t) dt
∂u2 0 Ω
Z Z
+ L̄yy (x, t)zh2k (x, t) dxdt + `¯yy (x, t)zh2k (x, t) dsdt
Q Σ
Z Z 2
∂ g
+ 2
(x, t, ȳ(x, t))zh2k (x, t) dµ̄(x, t),
r̄yy (x)zh2k (x, T ) dx +
Ω K ∂y
where Huu (t) was defined in (2.24) and L̄uy , L̄yy , `¯yy , and r̄yy are defined by
∂2L ∂2L
L̄uy (x, t) = (x, t, ȳ(x, t), ū(t)), L̄yy (x, t) = (x, t, ȳ(x, t), ū(t)),
∂u∂y ∂y 2
∂2` ∂2r
`¯yy (x, t) = (x, t, ȳ(x, t)), r̄yy (x) = (x, ȳ(x, T )).
∂y 2 ∂y 2
2
The first integral of ∂∂uL2 (ū, µ̄)h2k needs special care. We notice that
m Z Z T
2τ X  2τ 
h2k,i (t) dt = hk (t)> diag(χEiτ (t)) hk (t)dt. (2.47)
ε i=1 Eiτ 0 ε
Thanks to assumption (2.30), it holds that
 2τ 
d> Huu (t) + diag(χEiτ (t)) d ≥ ω |d|2 ∀ 0 < ε < ε0 , (2.48)
ε
if ε0 is taken sufficiently small. Therefore, the matrix function Huu (t)+ 2τ ε diag(χEi (t))
τ

is uniformly positive definite, and hence we infer that

Z T  2τ 
lim inf h>
k (t) Huu (t) + diag(χEiτ )(t)) hk (t) dt
k→∞ 0 ε
Z T  2τ 
≥ h> (t) Huu (t) + diagχEiτ (t)) h(t) dt. (2.49)
0 ε
∂d
Finally, taking into account that, by Corollary 2 with c0 = ∂y (·, ȳ), zhk → zh
strongly in C(Q̄), we deduce from (2.45)-(2.49) and (2.42)
Z T Z
 2τ 
h> (t) Huu (t) + diagχEiτ (t)) h(t) dt + 2 L̄uy (x, t)zh (x, t) · h(t) dxdt
0 Z ε Z Q

+ L̄yy (x, t)zh2 (x, t) dxdt + `¯yy (x, t)zh2 (x, t) dsdt
Q Σ
∂2g
Z Z
2
+ r̄yy (x)zh (x, T ) dx + 2
(x, t, ȳ(x, t))zh2 (x, t) dµ̄(x, t) ≤ 0.
Ω K ∂y
This expression can be written as
m Z
2τ X ∂2L
h2i (t) dt + (ū, µ̄)h2 ≤ 0,
ε i=1 Eiτ ∂u2
which along with (2.31) and the fact that h ∈ Cū implies that h = 0.
14 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

Step 4: hk → 0 strongly in L2 (0, T ; Rm ). We have already proved that hk * 0

weakly in L2 (0, T ; Rm ), therefore zhk → 0 strongly in C(Q̄) by Corollary 2. In view
of (2.45) and khk kL2 (0,T ;Rm ) = 1 we get
Z T
0 < ω = ω lim sup |hk (t)|2 dt
k→∞ 0
Z T  2τ 
≤ lim sup h>
k (t) Huu (t) + diag(χEiτ (t)) hk (t) dt
k→∞ 0 ε
(
2 ∂2L ∂2L
≤ lim sup + 2
(ū, µ̄) − (wk , µ̄)
k→∞ k ∂u ∂u2 B(L2 (0,T ;Rm ))
Z h i
− 2 L̄uy (x, t)zh (x, t) · h(t) − L̄yy (x, t)zh2 (x, t) dxdt
Q
Z Z
− `¯yy (x, t)zh2 (x, t) dsdt − r̄yy (x)zh2 (x, T ) dx
Σ Ω
∂2g
Z
− 2
(x, t, ȳ(x, t))zh2 (x, t) dµ̄(x, t) = 0.
K ∂y
Thus we have got a contradiction so that our hypotheses (2.33) cannot be true. 
Remark 2. A study of the proof reveals, why the special form of time-dependent
controls is important: The sequence (hk ) converges in L2 (0, T ; Rm ). This space
is associated with the domain (0, T ), where the controls can vary (notice that the
functions ei ∈ L∞ (Ω) are fixed). Controls of L2 (0, T ; Rm ) are mapped continuously
to continuous state functions. This fact is essential for the theory to work. For more
general control functions f ∈ L2 (0, T ; L∞ (Ω)), the associated domain of variability
is Ω×(0, T ). Here, the space L2 (Ω×(0, T )) would underly the proof. However, state
functions associated with controls of L2 (Ω × (0, T )) are not in general continuous
so that the proof does not go through.
Let us briefly comment on the consequences of the assumption (2.30) for the
matrix function Huu . To this aim, we consider a certain index set I ⊂ {1, . . . , m}
and assume that \
E= Eiτ (2.50)
i∈I
is non-empty. To avoid a re-numbering, let us assume that I = {1, . . . , k}. We
write Huu + α diag(χE (t)) as a block matrix in the form
 
A(t) + α Id B(t)
Huu (t) + α diag(χE (t)) = .
B > (t) C(t)
Then (2.30) is satisfied on the set E, if C(t) is uniformly positive definite there:
There must exist β > 0 such that
ξ > C(t)ξ ≥ β |ξ|2 ∀ξ ∈ Rm−k ,
but no condition on A(t) is needed. This is seen as follows: We split d ∈ Rm in
d1 ∈ Rk , containing the first k components and d2 ∈ Rm−k with the remaining
m − k components. Then
 
d> Huu (t) + α diag(χE (t)) d = d> 2 > >
1 A(t)d1 + α |d1 | + 2d1 B(t)d2 + d2 C(t)d2

≥ α |d1 |2 − cA |d1 |2 − cB |d1 |2 − ε|d2 |2 + β|d2 |2

α β
≥ |d1 |2 + |d2 |2 ≥ ω|d|2
2 2
by the Young inequality, if α is sufficiently large, ε > 0 is taken sufficiently small,
and ω = min{α, β}/2.
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 15

Tm
In the particular case k = m, i.e. on E = i=1 Eiτ , we do not need Smany condition
on Huu , while Huu must be uniformly positive definite on [0, T ] \ i=1 Eiτ , where
no strong activity helps.
Certainly, there are many possible index sets I, and the discussion of all different
cases is mainly a matter of combinatorics. We do not dwell upon this point. Instead,
let us discuss the case m = 2, where Huu + α diag(χE1τ , χE2τ ) is given as follows:
– In E1τ ∩ E2τ , we have
 
H11 (t) + α H12 (t)
Huu + α diag(χE1 , χE2 ) =
τ τ .
H12 (t) H22 (t) + α
This matrix is uniformly positive definite on E1τ ∩ E2τ for all sufficiently large α.
Therefore, the matrix Huu satisfies (2.30) on this subset without any further as-
sumption on positive definiteness of Huu .

– In E1τ \ (E1τ ∩ E2τ ), it holds

 
H11 (t) + α H12 (t)
Huu + α diag(χE1τ , χE2τ ) =
H12 (t) H22 (t)
so that we need H22 (t) ≥ β > 0 to make the matrix positive definite on E1τ \ (E1τ ∩
E2τ ) for sufficiently large α. An analogous condition must be imposed on H11 (t) on
E2τ \ (E1τ ∩ E2τ ).

– In [0, T ] \ (E1τ ∪ E2τ ), the matrix Huu (t) must be uniformly positive definite to
satisfy (2.30), since diag(χE1τ , χE2τ ) vanishes here.

Finally, we mention the case of a diagonal matrix Huu (t) = diag(Hui ui (t)). Here,
(2.30) is satisfied, if and only if
∀i ∈ {1, . . . , m} : Hui ui (t) ≥ ω ∀t ∈ [0, T ] \ Eiτ .
This happens in the standard setting where u appears only in a Tikhonov regular-
ization term, i.e.
L = L(x, t, y) + λ|u|2 .

3. Semilinear elliptic case

3.1. Problem statement. In this section, the following elliptic problem with
pointwise state constraints is considered:
 Z Z

 min J(u) = L(x, y(x), u) dx + `(x, y(x)) dS(x)


 u∈Uad Ω Γ



 subject to


(P2) (3.1)


 −∆y(x) + y(x) + d(x, y(x), u) = 0 in Ω



 ∂ν y(x) + b(x, y(x)) = 0 on Γ,



g(x, y(x)) 6 0 for all x ∈ Ω̄,


where Uad := {u ∈ Rm : ua 6 u 6 ub }. In this setting, Ω ⊂ Rn (n > 2)

is a bounded domain with Lipschitz boundary Γ. Moreover, sufficiently smooth
functions L, d : Ω × R × Rm → R, g : Ω̄ × R → R, and b, ` : Γ × R → R are given; ua
and ub with ua 6 ub are vectors of Rm . We adopt the notation from the preceding
sections.
Similar to [10], [12], the following assumptions are imposed on L, `, d, b:
16 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

(A.5) (Carathèodory type assumption) For each fixed x ∈ Ω or Γ respectively,

the functions L = L(x, y, u), ` = `(x, y), d = d(x, y, u), b = b(x, y), are of class C 2
with respect to (y, u). For all fixed (y, u) or fixed y, respectively, they are Lebesgue
measurable with respect to the variable x ∈ Ω, or x ∈ Γ respectively. The function
g = g(x, y) is supposed to be twice continuously differentiable with respect to y on
Ω̄ × R.
(A.6) (Monotonicity) For almost all x ∈ Ω, or x ∈ Γ, respectively, and any fixed

u ∈ Uad when applies, it holds

∂d ∂b
(x, y, u) ≥ 0, (x, y) ≥ 0.
∂y ∂y

(A.7) (Boundedness and Lipschitz properties) There is a constant C0 and, for all

M > 0, an CL (M ) such that the estimates

|d(x, 0, 0)| + |d0 (x, 0, 0)| + |d00 (x, 0, 0)| 6 C0
|d00 (x, y1 , u1 ) − d00 (x, y2 , u2 )| 6 CL (M )(|y1 − y2 | + |u1 − u2 |)
hold for almost all x ∈ Ω, all u, ui ∈ Uad , and all |yi | 6 M , i = 1, 2.
The functions L, ϕ, b, and g are assumed to satisfy (A.7) as well.
3.2. First-order necessary conditions. In this section, we consider optimality
conditions for problem (P2). Since the controls are in Rm and infinitely many
pointwise state constraints are given, this problem belongs to the class of semi-
infinite mathematical programming problems.
Therefore, the first- and second-order optimality conditions might be deduced
from the theory of semi-infinite programming. Nevertheless, the transfer of these
results to the control of PDEs needs the handling of the associated partial dif-
ferential equations and to discuss the differentiability properties of the underlying
control-to-state mappings so that these facts are worth mentioning.
First, we derive some basic results for the control-to-state mapping that are
standard for problems with control functions appearing linearly on the right hand
side of the PDE. Since our setting includes controls appearing nonlinearly, we have
to slightly update the arguments, and we state them for convenience of the reader.
Theorem 6. Let Ω ⊂ Rn be a bounded and Lipschitz domain. Suppose that the
conditions (A.5) and (A.6) are satisfied. Then, for all u ∈ Uad , the partial differ-
ential equation
−∆y(x) + y(x) + d(x, y(x), u) = 0 in Ω
(3.2)
∂ν y(x) + b(x, y(x)) = 0 on Γ,

has a unique solution yu ∈ H 1 (Ω) ∩ C(Ω̄) and the estimate

kyu kH 1 (Ω) + kyu kC(Ω̄) 6 cM (3.3)
holds with a constant cM that does not depend on u, if u belongs to Uad .
Proof. The result follows by setting
˜ y) := d(x, y, u),
d(x,
and applying regularity results for semilinear elliptic equations obtained in [3] or
[7]. Notice that d˜ is a monotone function with respect to y. 
Remark 3. Due to (3.3), the control-to-state mapping G : Rm → H 1 (Ω) ∩ C(Ω̄)
that assigns to each u ∈ Uad the solution yu of equation (3.2), satisfies the bound-
edness condition kG(u)kH 1 (Ω)∩C(Ω̄) 6 cM for all u ∈ Uad .
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 17

Remark 4. Due to our assumptions, the weak solution of (3.2) lies in the space
Yq,p = {y ∈ H 1 (Ω) : −∆y + y ∈ Lq (Ω), ∂ν y ∈ Lp (Γ)}
for all p, q ≥ 1. Yq,p is known to be continuously embedded in H 1 (Ω) ∩ C(Ω̄) for
each q > n/2 and each p > n − 1, see [3] or [13].
Remark 5. The operator G : Uad → H 1 (Ω) ∩ C(Ω̄) obeys the Lipschitz property
ky1 − y2 kH 1 (Ω) + ky1 − y2 kC(Ω̄) 6 CL (M )|u1 − u2 | (3.4)
m
for all yi such that yi = G(ui ), with ui ∈ R and |ui | ≤ M for i = 1, 2.
Remark 6. Under Assumption (A.7), the control-to-state mapping G : Rm →
H 1 (Ω) ∩ C(Ω̄) is twice continuously Fréchet-differentiable. For arbitrary elements
ū and u of Uad , and for ȳ = G(ū), the function y = G0 (ū)v is the unique solution
of the problem
∂d ∂d

 −∆zv + zv +
 (x, ȳ, ū)zv = − (x, ȳ, ū)v in Ω
∂y ∂u (3.5)
∂b

 ∂ν zv + (x, ȳ)zv = 0 on Γ.
∂y
Moreover, y satisfies the inequality
kzv kH 1 (Ω) + kzv kC(Ω̄) 6 c∞ |v| (3.6)
for some constant c∞ independent of u.
The function zv1 v2 , defined by zv1 v2 = G00 (u)[v1 , v2 ], is the unique solution of

∂d
 −∆zv1 v2 + zv1 v2 + (x, yu , u)zv1 v2 = −(yu1 , u> 00 > >
1 ) d (x, yu , u) (yu2 , u2 ) in Ω


∂y
∂b ∂2b (3.7)

 ∂ν zv1 v2 + (x, yu )zv1 v2 = − 2 (x, yu )yu1 yu2 on Γ.
∂y

∂y
The proofs of these results are fairly standard. For control functions appearing
linearly in the right hand side of (3.2), it is given in [11]. It can also be found in
[33]. The adaptation to the vector case needed here is more or less straightforward.
We omit the associated arguments.
We first note that the set feasible set of this problem is closed and bounded in
Rm due to the continuity of g. Therefore, the reduced functional f defined above is
continuous and compactness guarantees existence of an optimal control ū provided
that the feasible set is non-empty.
Notice that this existence result is not in general true for control functions ap-
pearing nonlinearly. In the sequel, ū stands for an optimal control with state
ȳ = G(ū). Later, in the context of second-order conditions, it is again a candidate
for local optimality. Henceforth we assume the following linearized Slater condition:

(A.8) (Regularity Condition) There exists u0 ∈ Uad such that

g(x, ȳ(x)) + gy (x, ȳ(x))zu0 −ū (x) < 0 for all x ∈ Ω̄, (3.8)
where ȳ satisfies (3.2).
It is well known, see [24], that this condition guarantees the existence of a nonneg-
ative Lagrange multiplier µ in the space M (Ω̄) = C(Ω̄)∗ of regular Borel measures
on Ω̄, and an associated adjoint state ϕ̄ ∈ W 1,σ (Ω) for all 1 ≤ σ < n/(n − 1),
defined by the adjoint equation

∂d ∂L ∂g
 −∆ϕ̄ + ϕ̄ + (·, ȳ, ū)ϕ̄ = (·, ȳ, ū) + µ̄Ω (·, ȳ) in Ω


∂y ∂y ∂y
(3.9)
∂b ∂` ∂g
∂ν ϕ̄ + (·, ȳ)ϕ̄ = (·, ȳ) + µ̄Γ (·, ȳ) on Γ,



∂y ∂y ∂y
18 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

where µ̄Ω and µ̄Γ denote the restrictions of µ̄ ∈ M (Ω̄) to Ω and Γ, respectively.
Following [7] or [13], we know that equation (3.9) admits a unique solution ϕ̄ ∈
W 1,σ (Ω), for all σ < n/(n − 1). Moreover the variational inequality
HT
u (u − ū) > 0, ∀u ∈ Uad , (3.10)
m
holds, where the vector Hu ∈ R is defined by Hu = (Hu,i )i=1,...,m , with
Z
∂L ∂d
Hu,i := (x, ȳ(x), ū) − ϕ̄(x) (x, ȳ(x), ū) dx,
Ω ∂ui ∂ui
and the complementarity condition
Z
g(x, ȳ(x))dµ̄(x) = 0 (3.11)
Ω̄
is satisfied.
For the ease of later computations, we introduce a different form of the Lagrange
function L : Yq,p × Rm × W 1,σ (Ω) × M (Ω̄) → R,
Z
L(y, u, ϕ, µ) =J(y, u) − (−∆y + y + d(x, y, u))ϕ dx
Ω
Z
− (∂ν y + b(x, y))ϕ dS + hµ, g(·, y)iΩ , (3.12)
Γ
which accounts also for the state-equation. Clearly, it holds that L = L, if y satisfies
the state equation. Moreover, it is known that it holds for the partial derivatives
of L with respect to y and u
Ly (ȳ, ū, ϕ, µ)y = 0 ∀y ∈ H 1 (Ω) (3.13)
Lu (ȳ, ū, ϕ, µ)(u − ū) = Hu (u − ū) > 0
T
∀u ∈ Uad . (3.14)
Therefore, the Lagrange function is an appropriate tool to express optimality con-
ditions in a convenient way, in particular it is useful to verify second-order sufficient
conditions by checking L00 (ȳ, ū, ϕ̄, µ̄)[(zh , h)]2 > 0 for all zh satisfying the linearized
equation (3.5). This second derivative of L with respect to (y, u) is expressed by
L00 (ȳ, ū,ϕ, µ)[(y1 , u1 ), (y2 , u2 )] =
Z 2
∂ L ∂2L ∂2L T∂ L
2
2 y1 y2 + ∂u∂y y2 · u1 + ∂y∂u · u2 y1 + u1 u2 dx
Ω ∂y ∂u2
Z 2
∂2d ∂2d
Z
∂ `
+ 2 y 1 y 2 dS − 2 y 1 y 2 + y2 · u 1
Γ ∂y Ω ∂y ∂u∂y
∂2d ∂2d
∂2b
Z
+ · u2 y1 + u1 T 2 u2 ϕ dx − 2 y1 y2 ϕ dS
∂y∂u ∂u Γ ∂y
∂2g
 
+ µ, 2 (·, ȳ)y1 y2 , (3.15)
∂y Ω
where the derivatives of L and d are taken at x, ȳ, and ū, respectively.
3.3. Second-Order Sufficient Optimality Conditions. In this section, we dis-
cuss second-order sufficient conditions for problem (P2). Since our controls belong
to Rm , the low regularity of the adjoint state does not cause troubles in the esti-
mations of L00 . Therefore, second-order sufficient conditions can be obtained for
arbitrary dimension of the domain.
The proof of these conditions can either be performed analogous to Section 2.4.
or derived by transferring the known second-order conditions from the theory of
semi-infinite programming problems to our control problem.
Therefore, we state the second-order sufficient optimality conditions without
proof. Let d = (Hu ) denote the first-order derivative of the Lagrangian function
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 19

with respect to u introduced in (3.10). For convenience, we introduce the following

sets: A+ := {i : di > 0}, A− := {i : di < 0}, and A := A+ ∪ A− . Let τ := min{|di | :
i ∈ A} > 0. We also define the critical cone associated with ū
Cū = {h ∈ Rm : hi = 0 ∀i ∈ A, and satisfies (3.16)-(3.18)}

> 0 if ūi = ua,i
hi = (3.16)
6 0 if ūi = ub,i
∂g
(x, ȳ(x))zh (x) 6 0, if g(x, ȳ(x)) = 0 (3.17)
∂y
Z
∂g
(x, ȳ(x))zh (x)dµ = 0, (3.18)
Ω̄ ∂y
where zh stands for G0 (ū)h.
Remark 7. It is not difficult to see that this definition of the critical cone Cū is
equivalent with the one used for semi-infinite programing in Bonnans and Shapiro
[6]. We only include the restrictions on the control in the restriction function of
the semi-infinite problem.
Theorem 7. Let ū be a feasible control for problem (P 2) with associated state ȳ
satisfying the first-order necessary conditions formulated in Section 3.2. Assume
that
∂2L
h> 2 (ū, µ)h > 0, ∀h ∈ Cū \{0}. (3.19)
∂u
Then there exist ε > 0 and δ > 0 such that for every feasible control of (P 2)
δ 2
J(ū) + |u − ū| 6 J(u)
2
holds for all feasible controls with |u − ū| < ε.
Remark 8. As mentioned above, this theorem might be proven as for our parabolic
problem. The proof is even simpler, since our controls belong to the space Rm , where
the unit sphere is compact. Moreover, it follows from regularity results Casas [7]
or Alibert and Raymond [3] that the control-to-state mapping u 7→ y is of class
C 2 from Rm to H 1 (Ω) ∩ C(Ω̄). However, the second-order conditions follow also
from the theory of semi-infinite programming, cf. Bonnans and Shapiro [6], if the
associated results are re-written in terms of partial differential equations.
In the same way, parabolic problems for the type
 Z Z

 min J(y, u) = L(x, t, y(x, t)) dxdt + `(x, t, y(x, t), u) dS(x)dt


 u∈Uad Q Σ



 subject to




yt − ∆y(x, t) + d(x, t, y(x, t)) = 0 in Q, (3.20)

∂ y(x, t) + b(x, t, y(x, t), u) = 0 on Σ,

ν



y(x, 0) − y0 (x) = 0 on Ω,








g(x, t, y(x, t)) 6 0 for all (x, t) ∈ Q̄
can be dealt with, where Uad is defined as before. Notice that here the control
appears in the boundary condition, where second-order sufficient conditions have
not yet been proven for control functions under the presence of pointwise state
constraints. Here, the control-to-state mapping is from Rm to C(Q̄), cf. Casas
[8]. Again, the second-order conditions can also be derived from the conditions for
semi-infinite programming in [6].
20 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

3.4. Examples. Let us finally illustrate the situation by some examples, which
exhibit different types of active sets and some kinds of nonlinearities.
Example 1. Here we study the optimal control problem
 1 2 1 2
 min4 J(y, u) = ky − yd kL2 (Ω) + |u − ud |



 u∈R 2 2



 subject to


(E1) P4


 −∆y(x) + y(x) + 71 y(x)3 = i=1 ui ei (x) in Ω



 ∂ν y(x) = 0 on Γ,



y(x) > b(x) for all x ∈ Ω̄ = [0, 1] × [0, 1],


where
( (
2x1 + 1, x1 < 12 , 1, x ∈ Ωi ,
b(x) = and ei = i = 1, 2, 3, 4,
2 x1 > 12 , 0 otherwise,

and Ω1 = (0, 12 ] × (0, 21 ], Ω2 = (0, 12 ] × ( 12 , 1), Ω3 = ( 12 , 1) × [ 12 , 1), and Ω4 =

( 21 , 1) × (0, 12 ). We shall see that the optimal state is ȳ = 2, hence the active set for
this problem is Ω3 ∪ Ω4 .
The first-order conditions for this problem are as follows: For the optimal solu-
tion (ȳ, ū) in H 2 (Ω)×R4 of the optimal control problem (E1), there exist a Lagrange
multiplier µ̄ ∈ M (Ω̄) and an adjoint state ϕ̄ ∈ W 1,σ (Ω) such that
P4
−∆ȳ + ȳ + 17 ȳ 3 = i=1 ūi ei in Ω






 ∂ν ȳ = 0 on Γ,
 12



 −∆ϕ̄ + ϕ̄ + 7 ϕ̄ = ȳ − yd − µ̄ in Ω



 ∂ν ϕ̄ = 0 on Γ,
 Z Z >
(3.21)

 (ū − u d ) + ϕ̄e1 dx, . . . , ϕ̄e4 dx = 0,
Ω Ω



 Z


(ȳ − b)dµ = 0, ȳ(x) > b(x) ∀x ∈ Ω̄




 Ω̄



and µ̄ > 0.
The following quantities satisfy the optimality system:
1

22  −x21 + , x1 < 12 ,

ȳ(x) = 2, ū =
>
[1, 1, 1, 1] , ϕ̄(x1 , x2 ) = 2
7  1,

x1 > 12 ,
4
together with the Lagrange multiplier µ̄ = δx1 ( 12 ), where δxi (z) denotes the Dirac
measure with respect to the variable xi , concentrated at xi = z.
In fact, it is easy to see that ȳ and ū fulfill the state equation in (3.21). Since ϕ̄
does not depend on x2 , we find that ∆ϕ̄ = ∂x21 ϕ̄ = δx1 ( 21 ) − ψ(x1 ), where
(
2, x1 < 12 ,
ψ(x1 ) =
0, x1 > 12 .
Therefore,
(
12 1 2 + (1 + 12 1 2 1
7 )( 2 − x1 ), x1 < 2 ,
−∆ϕ̄ + ϕ̄ + ϕ̄ = −δx1 ( ) + 1 12
7 2 4 (1 + 7 ), x1 > 12 ,
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 21

and a simple computation shows that this is equal to ȳ − yd − µ̄, with

( >
(x21 − 12 )(1 + 12 1 
7 ) x1 < 2 , 5 5 1 1
yd = u d = ū + , , , .
2 − 14 (1 + 127 ) x1 > 12 , 48 48 16 16
To check the second order conditions notice that ϕ̄ 6 12 , hence we have
Z
2 2 12 2
L (ȳ, ū, ϕ̄, µ̄)[(zh , h)] = kzh kL2 (Ω) + |h| −
00 2
ϕ̄zh dx
Ω 7
Z
2 1 2
≥ |h| + z 2 dx > |h|
7 Ω h
for all h ∈ R4 .
Example 2. This example includes a semilinear elliptic equation with controls ap-
pearing nonlinearly. The active set has measure zero. The problem is the following

1 1

2 2

 min J(y, u) = ky − yd kL2 (Ω) + |u − ud |



 u∈R2 2 2


 subject to



(E2) −∆y(x) + y(x) + 51 y(x)3 = 15 (u1 e1 (x) + u2 e2 (x))2 in Ω (3.22)






 ∂ν y(x) =0 on Γ,




y(x) > b(x) for all x ∈ Ω̄ = [0, 1] × [0, 1],


where
(
1 1
2 − 4 (x1 + x2 ), x1 + x2 > 1,
b(x) = 1
4 (x1 + x2 ), x1 + x2 < 1,
( (
1 1
8, x1 + x2 > 1, 4, x1 + x2 < 1,
e1 = e2 =
0 otherwise, 0 otherwise.
The optimality system is as in the previous example, with the gradient equation
" R #
2 (ū 1 e1 (x) + ū 2 e2 (x))ϕ̄e1 dx
(ū − ud ) + RΩ
5 (ū e (x) + ū2 e2 (x))ϕ̄e2 dx
Ω 1 1
" R #
2 Ω
ū1 e1 (x)2 ϕ̄ dx
= (ū − ud ) + = 0. (3.23)
ū e (x)2 ϕ̄ dx
R
5 Ω 2 2

To satisfy the optimality system, we define the quantities

> (
−(x1 2 − 2x1 + 2), x1 + x2 > 1,

1 9
ȳ(x) = , ū = 9, , ϕ̄(x) =
4 2 −(x2 2 + 1), x1 + x2 < 1,
note that ϕ̄ satisfies ∂ν ϕ̄ = 0. Moreover, we define the Lagrange multiplier µ̄ by the
positive measure
µ̄ = 2(1 − x1 )δx2 (1 − x1 ) + 2x2 δx1 (1 − x2 ).
where δx2 (1 − x1 ) and δx1 (1 − x2 ) are the Dirac measures with respect to x2 and x1
concentrated at x2 = 1 − x1 and x1 = 1 − x2 , respectively. These quantities satisfy
the state equation in (3.22) and the identity (3.23). The active set is the diagonal
of the unit square: x2 = 1 − x1 . Let us discuss the adjoint equation. It holds
(
∂2ϕ̄ −2, x1 + x2 > 1,
(x1 , x2 ) = 2x2 δx1 (1 − x2 ) + ,
∂x1 2 0, x1 + x2 < 1,
22 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

and
(
∂2ϕ̄ 0, x1 + x2 > 1,
(x1 , x2 ) = 2(1 − x1 )δx2 (1 − x1 ) +
∂x2 2 −2, x1 + x2 < 1.

Inserting the defined functions in the adjoint equation, we get

3
−∆ϕ̄ + ϕ̄ + ȳ2ϕ̄ =
5 (
83
2− 80 (x1 2 − 2x1 + 2), x1 + x2 > 1,
2(1 − x1 )δx2 (1 − x1 ) + 2x2 δx1 (1 − x2 ) + 83
2− 80 (x2 2 + 1), x1 + x2 < 1.

The right-hand side of the last equation must be equal to ȳ − yd − µ̄, hence we define

83 7
(
80 (x1 2 − 2x1 + 2) − 4 x1 + x2 > 1,
yd = 83
80 (x2 2 + 1) − 74 x1 + x2 < 1.

1 21 27 >
 
From (3.23), we find ud = ū − 10 64 , 32 . Finally, we confirm the second-order
sufficient conditions. Since 0 > ϕ̄ > −2, we have for all h ∈ R2 \{0} that
Z Z
2 6 2
L (ȳ, ū, ϕ̄, µ̄)[(zh , h)] = |h| +
00 2
zh2 (1
− ȳ ϕ̄) dx + (h1 e1 (x) + h2 e2 (x))2 ϕ̄ dx
Ω 5 5 Ω
Z Z
2 4
> |h| + zh2 dx − ((h1 e1 (x))2 + (h2 e2 (x))2 ) dx
Ω 5 Ω
2 4 1 2 2 19 2
> |h| − ( ) |Ω||h| > |h| .
5 4 20

In the last estimate, we have used ei (x) ≤ 1/4 for i = 1, 2.

Example 3. This is a slight modification of an example in [27], where the optimal

state is active in one single point.

 1 2 1 2
 min3 J(y, u) = ky − yd kL2 (Ω) + |u − ud |
2 2



 u∈R


 subject to



(E3) P3 (3.24)


 −∆y(x) + y(x) + y(x)3 = i=1 ui ei (x) in Ω



 ∂ν y(x) = 0 on Γ,



y(x) > 2 − |x|2 for all x ∈ Ω̄ = B1 (0),


In this case, we define Ω1 , Ω2 , Ω3 as the subsets of Ω̄ determined by (r, θ) ∈

[0, 1] × [0, π/2], (r, θ) ∈ (0, 1] × (π/2, 3π/2), (r, θ) ∈ (0, 1] × [3π/2, 2π), respectively.
Furthermore we set
  
1 in Ω1 , 2 in Ω2 , 10 in Ω3 ,
e1 (x) = e (x) = e (x) =
0 elsewhere, 2 0 elsewhere, 3 0 elsewhere.
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 23

The associated optimality system is


3
P3
 −∆ȳ + ȳ + ȳ = i=1 ui ei (x) in Ω

∂ν y(x) = 0 on Γ,







 2
 −∆ϕ̄ + ϕ̄ + 3ȳ ϕ̄ = ȳ − yd − µ̄ in Ω



∂ν ϕ̄(x) = 0 on Γ,




 Z Z Z >




 (ū − ud ) + ϕ̄e1 (x) dx, ϕ̄e2 (x) dx, ϕ̄e3 (x) dx = 0

 Z Ω Ω Ω

(ȳ − 2 + |x|2 )dµ̄ = 0, ȳ(x) > 2 − |x|2 ∀x ∈ Ω̄




Ω̄




µ̄ > 0.


It is not difficult to check that ȳ ≡ 2 and the optimal control ū = [10 5 1]> satisfy
the state equation. The active set consists of the single point x = (0, 0). The
Lagrange multiplier µ = δ0 satisfies the complementarity condition, where δ0 is
the Dirac measure concentrated at the origin. Let us define as adjoint state ϕ̄ =
1 1
log |x| − |x|2 , then we have that ∂ν ϕ̄ = 0 at the boundary,
2π 4π
 
1 1 1
−∆ϕ̄ + ϕ̄ + 3ȳ 2 ϕ̄ = − δ0 + 13 log |x| − |x|2 ,
π 2π 4π
and it is easy to confirm that the right-hand side of the last identity is equal to
ȳ − yd − µ̄, if we define
   >
1 1 1 317 37 1
yd = 2 − − 13 |x|2 − log |x| , ud = , , .
π 4π 2π 32 8 16
Therefore, the adjoint equation and the optimality system are satisfied. The second
order sufficient conditions are fulfilled, too. We have
Z
2 2
L (ȳ, ū, ϕ̄, µ̄)[(zh , h)] = |h| +
00 2
zh2 (1 − 12ϕ̄) dx> |h|
Ω
2
for all h ∈ R , since ϕ̄ 6 0.
Example 4. In this example we consider a problem with some coefficients of the
equation as controls.
1 1

2 2

 min2 J(y, u) = ky − yd kL2 (Ω) + |u − ud |

 u∈R

 2 2


subject to






 −∆y(x) + (u + π 2 )y(x) + u y(x)3 = f (x) in Ω

 1 2
(E4) ∂ ν y(x) = 0 on Γ, (3.25)




 (
(2 − 3x1 )2 − 41 x1 6 12 ,




 y(x) 6 b(x) := for all x ∈ Ω̄ = [0, 1] × [0, 1],
0 x1 > 12 ,





>

0 6 u = [u1 , u2 ] .


We impose nonnegativity of the components of u to have a monotone operator in

the state equation. The function f is given by
2
f (x) = 3π 2 cos(πx1 ) + cos(πx1 )3 .
3
24 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

The optimality system is



 −∆ȳ + (ū1 + π 2 )ȳ + ū2 ȳ 3 = f (x) in Ω
∂ν y(x) = 0 on Γ,








−∆ϕ̄ + (ū1 + π 2 )ϕ̄ + 3ū2 ϕ̄ȳ 2 = ȳ − yd + µ̄ in Ω




∂ν ϕ̄(x) = 0 on Γ






" R #!>
Ω
ϕ̄ȳ dx
 (ū − ud ) − R (u − ū) > 0 ∀0 6 u ∈ R2 ,


3
ϕ̄ȳ dx




 Ω

 Z

(ȳ(x) − b(x))dµ̄ = 0, ȳ(x) 6 b(x) ∀x ∈ Ω̄






 Ω̄

with µ̄ > 0.


We define
 2
 x1 − 1 , x 6 1 ,
" #
2
π 1 2
ȳ(x) = cos(πx1 ), ū = 2 , ϕ̄(x) = 2 8
 0, 1
3 x1 > 2 ,
and the Lagrange multiplier
1
µ̄ =
δx (1/2).
2 1
Since ȳ(x) = b(x) at x1 = 1/2, the state is active in the set: {( 21 , x2 ) : 0 6 x2 6 1}.
Inserting ϕ̄ in the adjoint equation, we obtain
ȳ − yd + µ̄ = −∆ϕ̄ + ϕ̄(ū1 + π 2 + 3ū2 ȳ 2 )
  2 
2 2 x1 1
−1 + 2(π + ȳ ) − x1 6 12 ,

1 
2 8
= δx1 (1/2) +
2  0 x1 > 12 .


Therefore, adjoint equation is satisfied with the choice

  
 1 − x2 − 1 π 2 + cos(πx1 )2 , x1 6 1 ,


1 2
yd = cos(πx1 ) + 4
 0 1
x1 > 2 .


The variational inequality holds true with

 >
1 2 20 1
ud = π 2 + 3 , + .
π 3 27 π 3
To verify the second-order sufficient conditions, we note that − 18 < ϕ̄ 6 0, and
compute L00 . This leads to the expression
 >  
Z zh zh
L00 (ȳ, ū, ϕ̄, µ̄)[(zh , h)]2 =  h1  H  h1  dx > 0
Ω h2 h2
for all h = [h1 , h2 ]> ∈ R2 \{0}, with
−3ȳ 2 ϕ̄
 
1 − 4ȳ ϕ̄ −ϕ̄
H= −ϕ̄ 1 0 
−3ȳ 2 ϕ̄ 0 1
Positive definiteness of the matrix H holds for all x ∈ [0, 1], because it is a sym-
metric diagonal dominant matrix with positive diagonal.
 CONTROL WITH FINITE-DIMENSIONAL CONTROL SPACE 25

References
[1] S. Agmon. On the eigenfunctions and on the eigenvalues of general elliptic boundary value
problems. Comm. Pure Appl. Math., 15:119–147, 1962.
[2] S. Agmon, A. Douglis, L. Nirenberg. Estimates near the boundary for solutions of elliptic
partial differential equations satisfying general boundary conditions. I. Comm. Pure Appl.
Math., 12:623–727, 1959. II. Comm. Pure Appl. Math., 17:35–92, 1964.
[3] J.-J. Alibert and J.-P. Raymond. Boundary control of semilinear elliptic equations with dis-
continuous leading coefficients and unbounded controls. Numer. Funct. Anal. and Optimiza-
tion, 3&4:235–250, 1997.
[4] Amann, H. Linear and quasilinear parabolic problems. Basel-Boston-Berlin, Birkhäuser, 1995.
[5] F. Bonnans. Second-order analysis for control constrained optimal control problems of semi-
linear elliptic systems. Appl. Math. and Optimization, 38:303–325, 1998.
[6] F. Bonnans and A. Shapiro. Perturbation Analysis of Optimization Problems. Springer-
Verlag, New York, 2000.
[7] E. Casas. Control of an elliptic problem with pointwise state constraints. SIAM J. Control
and Optimization, 4:1309–1322, 1986.
[8] E. Casas. Pontryagin’s principle for state-constrained boundary control problems of semilinear
parabolic equations. SIAM J. Control and Optimization, 35:1297–1327, 1997.
[9] E. Casas, J.C. de los Reyes, and F. Tröltzsch. Sufficient second-order optimality conditions
for semilinear control problems with pointwise state constraints. submitted, 2007.
[10] E. Casas and M. Mateos. Second order sufficient optimality conditions for semilinear elliptic
control problems with finitely many state constraints. SIAM J. Control and Optimization,
40:1431–1454, 2002.
[11] E. Casas and F. Tröltzsch. Second order necessary and sufficient optimality conditions for
optimization problems and applications to control theory. SIAM J. Optimization, 13:406–431,
2002.
[12] E. Casas, F. Tröltzsch, and A. Unger. Second order sufficient optimality conditions for a
nonlinear elliptic control problem. Z. für Analysis und ihre Anwendungen (ZAA), 15:687–
707, 1996.
[13] E. Casas, F. Tröltzsch, and A. Unger. Second order sufficient optimality conditions for some
state-constrained control problems of semilinear elliptic equations. SIAM J. Control and
Optimization, 38(5):1369–1391, 2000.
[14] Ciarlet, P.G. The finite element method for elliptic problems, Studies in Mathematics and its
Applications, North Holland, Amsterdam-New York-Oxford, 1979.
[15] P. Deuflhard, M. Seebass, D. Stalling, R. Beck, and H.-C. Hege. Hyperthermia treatment
planning in clinical cancer therapy: Modelling, simulation, and visualization. In A. Sydow, ed-
itor, Computational Physics, Chemistry and Biology, pages 9–17. Wissenschaft und Technik-
Verlag, 1997.
[16] K. Eppler and F. Tröltzsch. Fast optimization methods in the selective cooling of steel. In
M. Grötschel, S. O. Krumke, and J. Rambau, editors, Online optimization of large scale
systems, pages 185–204. Springer-Verlag, 2001.
[17] Gajewski, H., Gröger, K., Zacharias, K. Nichtlineare Operatorgleichungen und Operatordif-
ferentialgleichungen. Akademie-Verlag, Berlin, 1974.
[18] H. Goldberg and F. Tröltzsch. Second order optimality conditions for a class of control prob-
lems governed by nonlinear integral equations with application to parabolic boundary control.
Optimization, 20:687–698, 1989.
[19] J. A. Griepentrog, H. C. Kaiser, J. Rehberg. Heat kernel and resolvent properties for second
order elliptic differential operators with general boundary conditions on Lp . Adv. Math. Sci.
Appl., 11:87–112, 2001.
[20] J. A.Griepentrog. Linear elliptic boundary value problems with non-smooth data: Cam-
panato spaces of functionals. Math. Nachr., 243: 19–42, 2002.
[21] Grisvard, P. Elliptic Problems in Nonsmooth Domains. Pitman, Boston 1985.
[22] L. Henning and R. King. Drag reduction by closed-loop control of a separated flow over a
bluff body with a blunt trailing edge. 44th IEEE Conference on Decision and Control and
European Control Conference ECC 2005, Seville, Spain. accepted.
[23] Kato, T. Perturbation theory for linear operators. Corr. printing of the 2nd ed., Grundlehren
der mathematischen Wissenschaften. Springer, Berlin-Heidelberg-New York.
[24] S. Kurcyusz J. Zowe. Regularity and stability for the mathematical programming problem in
Banach spaces. Applied Mathematics and Optimization, (5):49–62, 1979.
[25] Ladyzhenskaya, O. A., Solonnikov, V. A, and Ural’ceva, N. N. Linear and Quasilinear Equa-
tions of Parabolic Type. American Math. Society, Providence, R.I. 1968.
26 J.C. DE LOS REYES‡ P. MERINO‡ J. REHBERG? F. TRÖLTZSCH†

[26] C. Meyer and P. Philip. Optimizing the temperature profile during sublimation growth of
sic single crystals: Control of heating power, frequency and coil position. Crystal Growth &
Design, 5:1145–1156, 2005.
[27] C. Meyer, U. Prüfert, and F. Tröltzsch. On two numerical methods for state-constrained
elliptic control problems. Technical report, Institut für Mathematik, Technische Universität
Berlin, 2005. Report 5-2005, submitted.
[28] Nečas, J. Les méthodes directes en théorie des equations elliptiques. Academia, Prague 1967.
[29] J.-P. Raymond and F. Tröltzsch. Second order sufficient optimality conditions for nonlin-
ear parabolic control problems with state constraints. Discrete and Continuous Dynamical
Systems, 6:431–450, 2000.
[30] Triebel, H. Interpolation theory, Function Spaces, Differential Operators. North Holland,
Amsterdam-New York-Oxford, 1978.
[31] F. Tröltzsch, R. Lezius, R. Krengel, and H. Wehage. Mathematische Behandlung der opti-
malen Steuerung von Abkühlungsprozessen bei Profilstählen. In K.H. Hoffmann, W. Jäger,
T. Lohmann, and H. Schunck, editors, Mathematik– Schlüsseltechnologie für die Zukunft,
Verbundprojekte zwischen Universität und Industrie, pages 513–524. Springer-Verlag, 1997.
[32] F. Tröltzsch and A. Unger. Fast solution of optimal control problems in the selective cooling
of steel. ZAMM, 81:447–456, 2001.
[33] Tröltzsch, F. Optimale Steuerung partieller Differentialgleichungen – Theorie, Verfahren und
Anwendungen. Vieweg, 2005.
‡ Department of Mathematics, EPN Quito, Ecuador
? Weierstrass-Institut für Angewandte Analysis und Stochastik (WIAS), Berlin, Ger-

many
† Institut für Mathematik, TU Berlin, Germany

View publication stats