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

net/publication/220258767

Optimal Control Problems with Mixed Control-State Constraints

Article in SIAM Journal on Control and Optimization · January 2000

DOI: 10.1137/S0363012999357926 · Source: DBLP

CITATIONS READS
57 408

2 authors:

Nadir Arada Jean-Pierre Raymond

43 PUBLICATIONS 757 CITATIONS

Université Toulouse III - Paul Sabatier
134 PUBLICATIONS 3,329 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Nadir Arada on 27 May 2014.

The user has requested enhancement of the downloaded file.

SIAM J. CONTROL OPTIM. c 2000 Society for Industrial and Applied Mathematics

Vol. 39, No. 5, pp. 1391–1407

OPTIMAL CONTROL PROBLEMS WITH MIXED CONTROL-STATE

CONSTRAINTS∗
N. ARADA† AND J.-P. RAYMOND†
Abstract. We consider control problems governed by semilinear parabolic equations in the
presence of pointwise mixed control-state constraints. We obtain optimality conditions with finitely
additive measures as multipliers associated to the mixed constraints. We study the regularity of the
multipliers for different problems. In particular, when a monotonicity or a separation condition is
satisfied, we prove that multipliers are bounded measurable functions.

AMS subject classifications. 49K20, 93C20

Key words. optimal control, mixed control-state constraints, pointwise constraints, regular
multipliers

PII. S0363012999357926

1. Introduction. This paper deals with control problems governed by parabolic

equations in the presence of pointwise constraints on the control and the state vari-
ables. We are interested in optimality conditions for these problems, and in regularity
properties of Lagrange multipliers associated with mixed control-state constraints.
For optimization problems in Banach spaces the regularity of the multipliers is, in
general, deduced from the following heuristic statement.
(HS) The Lagrange multiplier associated with a constraint belongs to
the dual space of the space in which the constraint set has a nonempty
interior (or a finite codimension).
For bounded controls, pointwise mixed control-state constraints are well posed in
an L∞ -space. Thus, according to (HS), the corresponding Lagrange multiplier must
belong to a dual space of the form (L∞ ) .
In some situations this heuristic statement does not give the best regularity result
for the Lagrange multipliers. Let us give an example. The multiplier corresponding
to pointwise state constraints is, in general, a bounded Radon measure [7], [10], [15],
[12]. Indeed, pointwise state constraints are well posed in the space of continuous
functions. For problems considered in [7], [15], the multiplier associated with a state
constraint of the form y ≥ 0 is a Radon measure. If the previous state constraint is
slightly perturbed and is replaced by the control-state constraint y + εv ≥ 0 (with
ε > 0 and v ≥ 0), then the multiplier may be a bounded measurable function (see
Corollary 5.6). This is clearly a better regularity result than the one derived from
(HS).
This kind of result is known since 1962 for the control of ordinary differential
equations, and is stated in [14]. The case of ordinary differential equations is very
specific since it is often assumed that optimal controls are piecewise continuous [14],
[11]. Even under the assumption of piecewise continuous controls, the “informal”
Theorem 4.1 in [11] has not, to the authors’ knowledge, been proved fully in the
literature (see [11, p. 185]).

∗ Received by the editors June 18, 1999; accepted for publication (in revised from) June 17, 2000;

published electronically December 28, 2000.

http://www.siam.org/journals/sicon/39-5/35792.html
† UMR-CNRS 5640, UFR MIG, Université Paul Sabatier, 31062 Toulouse Cedex 4, France

([email protected], [email protected]).
1391
1392 N. ARADA AND J.-P. RAYMOND

The situation is more complicated for partial differential equations. Indeed, in

this case, it is not realistic to assume that optimal controls are piecewise continuous.
Very recently, using a duality method, Bergounioux and Tröltzsch [5] have proved the
existence of regular Lagrange multipliers for a linear control problem with constraints
of bottleneck type. An extension to the case of semilinear equations is obtained in
[6]. The case of integral control-state constraints is studied in [8].
To our knowledge, there are no general optimality conditions for control prob-
lems governed by partial differential equations in the presence of pointwise mixed
control-state constraints. In this paper we first prove optimality conditions for con-
trol problems with constraints of the form

g(y, v) ∈ C ⊂ (L∞ )
(mixed control-state constraints)

(in this setting g = (g1 , . . . , g
) is a vector valued function). In section 3, we do not

make any particular assumption on mixed state-control constraints, and the corre-
sponding multiplier belongs to a space of the type ((L∞ )
) (Theorem 3.1). In sections
4 and 5, we prove the existence of regular Lagrange multipliers in the following cases:
• = 1 and (gv (y, v))−1 belongs to L∞ (Theorem 4.1),
• mixed control-state constraints of the form

gi (y, v) ≤ 0 for i = 1, . . . , (Theorems 5.2, 5.5, 5.7).

Let us denote by ζ̄i the multiplier of the constraint gi (y, v) ≤ 0 corresponding to a so-
lution (ȳ, v̄) of problem (P). (The control problem (P) is defined in section 1.1.) With
optimality conditions and the Hahn–Banach extension theorem, we first establish that
the sum Σ
i=1 giv 
(ȳ, v̄)ζ̄i has the same regularity as the adjoint state.
In section 5.1 we study the case when = 2 and when (g1 , g2 ) satisfies a separation
condition of the form g1 (ȳ, v̄) + g2 (ȳ, v̄) ≤ −ε < 0. We are able to prove that the
supports of ζ̄1 and ζ̄2 are disjoints. We deduce that each multiplier has the same
regularity as the adjoint state, and we use a bootstrap argument to obtain the best
regularity result (Theorem 5.2).
In section 5.2 we suppose that g satisfies a monotonicity condition of the form

giv (y, v) ≤ 0 for i = 1, . . . , . We prove that, for i = 1, . . . , , the additive measures

giv (ȳ, v̄)ζ̄i are nonnegative. With a decomposition theorem for nonnegative additive

measures, we prove that each term giv (ȳ, v̄)ζ̄i has the same regularity as the adjoint
state, and we can conclude with a bootstrap argument (Theorem 5.5).
In section 5.3 we study a problem in which the separation and monotonicity
conditions are coupled (Theorem 5.7). These results are applied to examples for which
the regularity of multipliers cannot be deduced from results known in the literature.
The separation condition seems to be new (see Remark 5.1). The monotonicity
condition is similar to regularity conditions stated for the control of ordinary differ-
ential equations (condition (b) in [4]).
For clarity, we consider only the case of a boundary control, but our method can
be extended to problems with distributed and boundary controls. Our method is
general and may be extended to other problems.
1.1. Setting of the control problem. Consider the semilinear parabolic equa-
tion
∂y ∂y
(1.1) + Ay + Φ(·, y) = 0 in Q, + Ψ(·, y, v) = 0 on Σ, y(0) = yo in Ω,
∂t ∂nA
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1393

where Ω is a bounded domain in RN , Q = Ω×]0, T [, Σ = Γ×]0, T [, Γ is the boundary

of Ω, T > 0, v ∈ L∞ (Σ) is a boundary control, A is a second order elliptic operator,
Φ and Ψ are Carathéodory functions (assumptions are specified in section 1.2), and
yo is a fixed function in C(Ω). Constraints of the form
(1.2) g(·, y(·), v(·)) ∈ C
are imposed on the pair (y, v), where g is a continuous mapping from Σ × R2 into R
,
and C is a closed convex subset of (L∞ (Σ))
with a nonempty interior in (L∞ (Σ))
.
The paper is concerned with the control problem
(P) inf{J(y, v) | (y, v) ∈ C(Q) × L∞ (Σ), (y, v) satisfies (1.1) and (1.2)},
where the cost functional is given by




J(y, v) = F (x, t, y) dx dt + G(s, t, y, v) ds dt + L(x, y(T )) dx.
Q Σ Ω

Let us give a simple example for which results of the paper may be applied. Consider
the state equation
∂y ∂y
− ∆y = 0 in Q, + |y|3 y = v on Σ, y(0) = yo in Ω,
∂t ∂n
 2
with4 yo ≥ 0, v ≥ 0, and the functional J(y, v) = Ω a(x)|y(x, T ) − yd (x)| dx +
Σ
y (s, t) dsdt. The nonnegative function  a plays the role of a weight, yd is a desired
profile of temperature, and the term Σ y 4 (s, t) dsdt may be interpreted as a term
penalizing a too-high temperature on the boundary. In the above radiation boundary
4
condition, the control v corresponds to yext , where yext is the exterior temperature. In
4
industrial processes it may be important that the difference yext −y 4 be bounded from
4
above. Thus a constraint of the form 0 ≤ v ≤ y + c on Σ, with c > 0, is meaningful.
Setting g1 (y, v) = v − |y|3 y − c, g2 (y, v) = −v, the above constraint is equivalent to
g(y, v) = (g1 (y, v), g2 (y, v)) ∈ C, where C = {z ∈ (L∞ (Σ))2 | z ≤ 0} (z ≤ 0 must be
understood componentwise). The existence of solutions for the corresponding problem
(P) can be obtained by standard arguments. In Corollary 5.3 we obtain optimality
conditions for a class of problems including this example, with multipliers belonging
to L∞ (Σ). In our knowledge, the existence of bounded measurable multipliers for
this example cannot be deduced from known results in the literature. Examples
of constraints for which we obtain the same results are given in section 5. Other
examples of equations and functionals satisfying the assumptions of the paper are
given in [15].
1.2. Assumptions and notation. We suppose that Ω is of class C 2 (the bound-
ary Γ of Ω is an (N − 1)-dimensional manifold of class C 2 such that Ω lies locally on
one side of Γ). The operator A is of the form
N

Ay(x) = − Di (aij (x)Dj y(x))
i,j=1

(Di denotes the partial derivative with respect to xi ), with coefficient aij belonging
to C 1 (Ω) and satisfying the conditions
N

aij (x) = aji (x) for all i, j ∈ {1, . . . , N }, m0 |ξ|2 ≤ aij (x)ξi ξj
i,j=1
1394 N. ARADA AND J.-P. RAYMOND

for every ξ ∈ RN and every x ∈ Ω, with m0 > 0. The conormal derivate of y with
∂y
respect to A is denoted by ∂n A
, that is

∂y 
(s, t) = aij (s)Dj y(s, t)ni (s),
∂nA i,j

where n = (n1 , . . . , nN ) is the unit normal to Γ outward Ω.

To prove the existence of multipliers in (Lk (Σ))
we need some regularity assump-
tions on the data. They are specified below. Some of them depend on k. Observe
that k = ∞ is allowed (see (A3)–(A5) below). Throughout the text, the exponent k
denotes a given fixed exponent belonging to ]1, ∞].
(A1) Φ is a Carathéodory function from Q×R into R. For almost every (x, t) ∈ Q,
Φ(x, t, ·) is of class C 1 . The following estimates hold:

|Φ(x, t, 0)| ≤ η(|y|), Co ≤ Φy (x, t, y) ≤ η(|y|),

where Co ∈ R and η is a nondecreasing function from R+ into R+ .

(A2) Ψ is a Carathéodory function from Σ×R2 into R. For almost every (s, t) ∈ Σ,
Ψ(s, t, ·) is of class C 1 , and it satisfies

|Ψ(s, t, 0, v)| + |Ψv (s, t, y, v)| ≤ η(|y|) η(|v|),

Co ≤ Ψy (s, t, y, v) ≤ η(|y|) η(|v|).

(A3) F is a Carathéodory function from Q × R into R. For almost all (x, t) ∈ Q,

F (x, t, ·) is of class C 1 . The following estimates hold:

|F (x, t, y)| ≤ F1 (x, t)η(|y|), |Fy (x, t, y)| ≤ F2 (x, t)η(|y|),

where F1 ∈ L1 (Q), F2 ∈ Lk1 (Q), with k1 > N(N+1+2k +2)k

if k < ∞, and k1 > N2 + 1 if
k = ∞.
(A4) G is a Carathéodory function from Σ × R2 into R. For almost all (s, t) ∈ Σ,
G(s, t, ·) is of class C 1 . The following estimates hold:

|G(s, t, y, v)| ≤ G1 (s, t)η(|y|)η(|v|), |Gv (s, t, y, v)| ≤ G2 (s, t)η(|y|)η(|v|),

|Gy (s, t, y, v)| ≤ G3 (s, t)η(|y|)η(|v|),

where G1 ∈ L1 (Σ), G2 ∈ Lk (Σ), with 1 < k ≤ ∞, G3 ∈ Lk2 (Σ) with k2 > (N +1)k
N +1+k if
k < ∞, and k2 > N + 1 if k = ∞.
(A5) L is a Carathéodory function from Ω × R into R. For almost all x ∈ Ω,
L(x, ·) is of class C 1 , and

|L(x, y)| ≤ L1 (x)η(|y|), |Ly (x, y)| ≤ L2 (x)η(|y|),

where L1 ∈ L1 (Ω), L2 ∈ Lk3 (Ω), with k3 > NN+1 k

if k < ∞, and k3 = ∞ if k = ∞.
2

(A6) g is a function from Σ × R into R . For every (s, t) ∈ Σ, g(s, t, ·) is of class
C 1 , and for every y ∈ R, g(·, y) is bounded measurable on Σ. The functions g, gy ,
and gv are bounded on compact subsets of Σ × R2 .
For simplicity, we often write g(y, v) in place of g(s, t, y, v). We set Ω0 = Ω × {0},
ΩT = Ω × {T }, ΓT = Γ × {T }. For every 1 ≤ r ≤ ∞, the usual norms in the spaces
Lr (Ω), Lr (Q), Lr (Σ) will be denoted by || · ||r,Ω , || · ||r,Q , || · ||r,Σ . The Hilbert space
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1395

W (0, T ; H 1 (Ω), (H 1 (Ω)) ) = {y ∈ L2 (0, T ; H 1 (Ω)) | dy 2 1 

dt ∈ L (0, T ; (H (Ω)) )} will be
denoted by W (0, T ). Throughout what follows, we denote by C a generic constant. If
O is a locally compact subset of RN +1 , Cb (O) denotes the space of bounded continuous
functions on O, and C0 (O) denotes the space of continuous functions vanishing at
infinity. The dual space of C0 (O) will be denoted by Mb (O) (the space of bounded
Radon measures  on O). If µ belongs to Mb (O) and y belongs to Cb (O), we set
µ, yb,O = O y dµ. For σ > 1, k ∧ σ stands for min(k, σ). For simplicity, ·, ·∗,Σ
stands for the duality pairing between the spaces ((L∞ (Σ))
) and (L∞ (Σ))
.
2. State and adjoint equations. We begin this section by recalling some ex-
istence, uniqueness, and regularity results concerning the state equation (1.1).
Definition 2.1. A function y ∈ L2 (0, T ; H 1 (Ω)) ∩ C([0, T ]; L2 (Ω)) is a weak
solution of (1.1) if and only if Φ(·, y(·)) ∈ L1 (Q), Ψ(·, y(·), v(·)) ∈ L1 (Σ), and
 

N

−y ∂z
+ aij Di yDj z + Φ(·, y)z  dx dt − yo z(0) dx
Q ∂t i,j=1 Ω


= − Ψ(·, y, v)z ds dt for all z ∈ C 1 (Q) such that z(T ) = 0.
Σ

Theorem 2.2 (see [16, Theorem 3.1]). Let us suppose that (A1)–(A2) are satis-
fied. Equation (1.1) admits a unique weak solution yv ∈ W (0, T ) ∩ C(Q) satisfying

yv C(Q) ≤ C(v∞,Σ + yo C(Ω) + 1).

Now, let us recall some existence, uniqueness, and regularity results for the adjoint
equation. Let (a, b) be in L∞ (Q) × L∞ (Σ), and consider the terminal boundary value
problem

∂p ∂p
(2.1) + Ap + ap = µQ in Q, + bp = µΣ on Σ, p(T ) = µΩT on Ω,
∂t ∂nA

where µ = µQ +µΣ +µΩT is a bounded Radon measure on Q\Ω0 , µQ is the restriction

of µ to Q, µΣ is the restriction of µ to Σ, and µΩT is the restriction of µ to ΩT .
Definition 2.3. A function p ∈ L1 (0, T ; W 1,1 (Ω)) is a weak solution of (2.1) if
and only if
 

N

p ∂z
+ aij Dj pDi z + azp dxdt + bpz dsdt = µ, zb,Q\Ω0
Q ∂t i,j=1 Σ

for all z ∈ C 1 (Q) satisfying z(0) = 0 on Ω.

We recall an existence theorem for parabolic equations with measures as data.
Theorem 2.4 (see [15, Theorem 4.1]). Let (a, b) be in L∞ (Q) × L∞ (Σ), and let
µ be in Mb (Q \ Ω0 ). Equation (2.1) admits a unique solution p in L1 (0, T ; W 1,1 (Ω)).
N
For every (δ, d, β) satisfying δ ≥ 1, d ≥ 1, β ≥ 0, and 2d + 1δ > N2 + β2 , p belongs to
δ β,d
L (0, T ; W (Ω)) and

||p||Lδ (0,T ;W β,d (Ω)) ≤ C||µ||Mb (Q\Ω0 ) ,

1396 N. ARADA AND J.-P. RAYMOND

where C ≡ C(Ω, T, δ, d, M ) is a positive constant independent of a and b, and M is

an upper bound for ||a||∞,Q + ||b||∞,Σ . In particular, the trace of p on Σ satisfies
N +1
||p|Σ ||Lσ (Σ) ≤ C||µ||Mb (Q\Ω0 ) for every 1 ≤ σ < .
N
Moreover, there exists a function in L1 (Ω), denoted by p(0), such that




∂z ∂z
+ Az + az p dx dt + + bz p ds dt + z(0)p(0) dx
Q ∂t Σ ∂nA Ω

= µ, zb,Q\Ω0 for all z ∈ Y,

where Y = {y ∈ W (0, T ) | ∂y q ∂y ∞
∂t + Ay ∈ L (Q), ∂nA ∈ L (Σ), y(0) ∈ C(Ω)}.
Proposition 2.5. Set Λ(µQ , µΣ , µΩT ) = p|Σ , where p is the solution of (2.1)
corresponding to (µQ , µΣ , µΩT ). The mapping µQ → Λ(µQ , 0, 0) is continuous from
(N +2)β N
Lβ1 (Q) into Lβ (Σ), with β1 > N +1+2β if 1 ≤ β < ∞, and β1 > 2 + 1 if β = ∞. The
mapping µΣ → Λ(0, µΣ , 0) is continuous from L (Σ) into L (Σ), with β2 > (N
β2 β +1)β
N +1+β
if 1 ≤ β < ∞, and β2 > N + 1 if β = ∞.
The mapping µΩT → Λ(0, 0, µΩT ) is continuous from Lβ3 (Ω) into Lβ (Σ), with
β3 > NN+1β
if 1 ≤ β < ∞, and β3 = ∞ if β = ∞.
In particular, p|Σ belongs to Lk (Σ) if (µQ , µΣ , µΩT ) belongs to Lk1 (Q) × Lk2 (Σ) ×
Lk3 (Ω), where k, k1 , k2 , and k3 are the exponents in assumptions (A3)–(A5).
Proof. The proof may be performed by using estimates on the analytic semigroup
as in [17], Propositions 3.1 and 3.2.
3. Main results. Define the Hamiltonian function H : Σ × R4 −→ R by

H(s, t, y, v, p, α) = α G(s, t, y, v) + p Ψ(s, t, y, v).

We shall say that (ȳ, v̄) is regular if there exists v ∈ L∞ (Σ) such that

(3.1) g(ȳ, v̄) + gy (ȳ, v̄)(zv̂ − zv̄ ) + gv (ȳ, v̄)(v − v̄) ∈ int C,

where zv (with v = v̂ or v = v̄) is the solution of


 ∂z

 + Az + Φy (·, ȳ)z = 0 in Q,
∂t
(3.2)

 ∂z
 + Ψy (·, ȳ, v̄)z = −Ψv (·, ȳ, v̄) v on Σ, z(0) = 0 in Ω.
∂nA
In (3.1) “int” denotes the interior for the topology of (L∞ (Σ))
. The qualification
condition (3.1) is of Mangasarian–Fromowitz type. It appears in [19, (3.6), p. 20].
Theorem 3.1. Suppose that (A1)–(A6) are fulfilled. If (ȳ, v̄) is a solution of (P),
then there exist ᾱ ∈ [0, 1], ζ̄ ∈ ((L∞ (Σ))
) , and p̄ ∈ L1 (0, T ; W 1,1 (Ω)), such that the
following conditions hold.
• Nontriviality condition :

(3.3) (ᾱ, ζ̄) = 0.

• Complementarity condition:

(3.4) ζ̄, z − g(ȳ, v̄)∗,Σ ≤ 0 for all z ∈ C.

 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1397

• Adjoint equation:

 ∂ p̄

 − + Ap̄ + Φy (·, ȳ)p̄ = −ᾱFy (·, ȳ) in Q,

 ∂t

(3.5) ∂ p̄
 + Ψy (·, ȳ, v̄)p̄ = −ᾱ Gy (·, ȳ, v̄) − [gy (ȳ, v̄)∗ ζ̄]|Σ on Σ,

 ∂n A



p̄(T ) = −ᾱLy (·, ȳ(T )) − [gy (ȳ, v̄)∗ ζ̄]|Γ×{T } on Ω,

where [gy (ȳ, v̄)∗ ζ̄]|Σ is the restriction of gy (ȳ, v̄)∗ ζ̄ to Σ, [gy (ȳ, v̄)∗ ζ̄]|Γ×{T } is the re-
striction of gy (ȳ, v̄)∗ ζ̄ to Γ × {T }, and gy (ȳ, v̄)∗ ζ̄ is the bounded Radon measure on
Σ ∪ (Γ × {T }) defined by

gy (ȳ, v̄)∗ ζ̄, zb,Σ∪Γ×{T } = ζ̄, gy (ȳ, v̄)z∗,Σ for all z ∈ C0 (Σ ∪ (Γ × {T })).

• Optimality condition for v̄:

(3.6) Hv (s, t, ȳ, v̄, p̄, ᾱ)χ ds dt + ζ̄, gv (ȳ, v̄) χ∗,Σ = 0 for all χ ∈ L∞ (Σ).
Σ

Moreover, if there exists v satisfying (3.1), then we can take ᾱ = 1.

Remark 3.2. Here gy (ȳ, v̄) is a linear continuous operator from C0 (Σ ∪ (Γ × {T }))
into (L∞ (Σ))
, and gy (ȳ, v̄)∗ denotes the corresponding adjoint operator.
Proof. Let us set

A = {(z, λ) ∈ (L∞ (Σ))
× R | z = g(ȳ, v̄) + gy (ȳ, v̄)(zv − zv̄ ) + gv (ȳ, v̄)(v − v̄),
λ = Jy (ȳ, v̄) (zv − zv̄ ) + Jv (ȳ, v̄)(v − v̄) for some v ∈ L∞ (Σ)},
B = int C×] − ∞, 0[,

where zv̄ (respectively, zv ) is the solution of (3.2) corresponding to v̄ (respectively, v).

The sets A and B are convex, and B is open. Let us prove that A ∩ B = ∅. Argue by
contradiction, and suppose that there exists vo ∈ L∞ (Σ) such that

(3.7) g(ȳ, v̄) + gy (ȳ, v̄)(zvo − zv̄ ) + gv (ȳ, v̄)(vo − v̄) ∈ int C,

(3.8) λo = Jy (ȳ, v̄) (zvo − zv̄ ) + Jv (ȳ, v̄)(vo − v̄) < 0.

Set vρ = v̄ + ρ(vo − v̄). Let yρ be the solution of (1.1) corresponding to vρ , and set
gρ = g(ȳ, v̄) + ρ1 (g(yρ , vρ ) − g(ȳ, v̄)). Due to (3.7) and (3.8), there exists 0 < ρo < 1
such that
J(yρ , vρ ) − J(ȳ, v̄)
gρ ∈ int C and <0 for all 0 < ρ ≤ ρo < 1.
ρ
Therefore, for every 0 < ρ ≤ ρo < 1, we have

g(yρ , vρ ) = ρ gρ + (1 − ρ) g(ȳ, v̄) ∈ int C and J(yρ , vρ ) < J(ȳ, v̄) = inf(P).

The pair (yρ , vρ ) is admissible for (P) and we have a contradiction. Thus,
A ∩ B = ∅. From a geometric version of the Hahn–Banach theorem (the Eidelheit
theorem [18]), there exists (ᾱ, ζ̄) ∈ R × ((L∞ (Σ))
) , such that

(3.9) ᾱ λ1 + ζ̄, z1 ∗,Σ > ᾱ λ2 + ζ̄, z2 ∗,Σ for all (z1 , λ1 , z2 , λ2 ) ∈ A × B,

1398 N. ARADA AND J.-P. RAYMOND

(3.10) ᾱ λ1 + ζ̄, z1 ∗,Σ ≥ ᾱ λ2 + ζ̄, z2 ∗,Σ for all (z1 , λ1 , z2 , λ2 ) ∈ A × B.

• Due to (3.9), (ᾱ, ζ̄) = 0. We easily see that ᾱ is nonnegative. Indeed if ᾱ < 0,
setting z1 = g(ȳ, v̄), λ1 = 0, and letting λ2 tend to −∞, we obtain a contradiction.
Thus ᾱ is nonnegative. For z fixed in C, by setting z1 = g(ȳ, v̄), z2 = z, λ1 = λ2 = 0
in (3.10), we establish (3.4).
• Let v ∈ L∞ (Σ). By setting z1 = g(ȳ, v̄) + gy (ȳ, v̄)(zv − zv̄ ) + gv (ȳ, v̄)(v − v̄),
λ1 = Jy (ȳ, v̄) (zv − zv̄ ) + Jv (ȳ, v̄)(v − v̄), z2 = g(ȳ, v̄), and λ2 = 0 in (3.10), we obtain

ᾱJy (ȳ, v̄) (zv − zv̄ ) + gy (ȳ, v̄)∗ ζ̄, zv − zv̄ b,Σ∪ΓT
(3.11)
+ ᾱJv (ȳ, v̄)(v − v̄) + ζ̄, gv (ȳ, v̄)(v − v̄)∗,Σ ≥ 0 for all v ∈ L∞ (Σ).

Since the above inequality is satisfied for all v ∈ L∞ (Σ), the inequality can be
replaced by an equality. Let p̄ be the weak solution of (3.5). With the Green formula
of Theorem 2.4, we have



ᾱ Fy (x, t, ȳ) (zv − zv̄ ) dx dt + ᾱ Ly (x, ȳ(T )) (zv − zv̄ )(T ) dx
Q Ω


+ ᾱ Gy (s, t, ȳ, v̄) (zv − zv̄ ) ds dt + gy (ȳ, v̄)∗ ζ̄, zv − zv̄ b,Σ∪ΓT
Σ


= p̄ Ψv (s, t, ȳ, v̄) (v − v̄) ds dt for all v ∈ L∞ (Σ).
Σ

This equality together with (3.11) gives (3.6).

• Finally, if there exists v ∈ L∞ (Σ) such that (3.1) is satisfied, then by set-
ting z1 = z2 = g(ȳ, v̄) + gy (ȳ, v̄)(zv̂ − zv̄ ) + gv (ȳ, v̄)(v − v̄) in (3.9), we prove that
ᾱ = 0.
4. Regularity of multipliers for purely mixed constraints. Throughout
what follows, (ȳ, ū) stands for an optimal solution to (P). In the following two sections,
we want to prove that the multiplier ζ̄ may be identified with η̄ dsdt, where η̄ belongs
to (Lk (Σ))
. In this case, (3.3)–(3.6) are rewritten in the form


(4.1) (ᾱ, η̄) = 0, η̄ (z − g(ȳ, v̄)) ds dt ≤ 0 for all z ∈ C,
Σ


 ∂ p̄

 − + Ap̄ + Φy (·, ȳ)p̄ = −ᾱFy (·, ȳ) in Q,

 ∂t

(4.2) ∂ p̄
 + Ψy (·, ȳ, v̄)p̄ = −ᾱGy (·, ȳ, v̄) − gy (ȳ, v̄)∗ η̄ on Σ,

 ∂n


A

p̄(T ) = −ᾱLy (·, ȳ(T )) on Ω,

(4.3) Hv (s, t, ȳ(s, t), v̄(s, t), p̄(s, t), ᾱ) = −gv (ȳ, v̄)∗ η̄(s, t) for almost all (s, t) ∈ Σ.

In this section, we consider the control problem (P) when = 1 and when the following
regularity condition is satisfied.
(A7) = 1 and the function (gv (ȳ(·), v̄(·)))−1 belongs to L∞ (Σ).
Theorem 4.1. Let (ȳ, v̄) be a solution to (P). Suppose that (A1)–(A7) are
fulfilled. There exist ᾱ ∈ [0, 1], η̄ ∈ Lk (Σ) (k is the exponent in assumption (A3)),
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1399

and p̄ ∈ L1 (0, T ; W 1,1 (Ω)), such that (4.1)–(4.3) hold. Moreover, if (3.1) is satisfied,
then we can take ᾱ = 1.
Proof. Step 1. Due to Theorem 3.1, there exist ᾱ ∈ [0, 1], ζ̄ ∈ (L∞ (Σ)) , and
p̄ ∈ L1 (0, T ; W 1,1 (Ω)), such that (3.3)–(3.6) hold. From Proposition 2.5, we know
that
N +1
(4.4) p̄|Σ ∈ Lσ (Σ) for all σ < .
N
Let σ be such that σ < NN+1 . From (4.4), with assumptions (A2) and (A4), we can
easily see that Hv (·, ȳ, v̄, p̄, ᾱ) belongs to Lk∧σ (Σ). Consider the continuous linear


operators S : L(k∧σ) (Σ) → R ((k ∧ σ) is the exponent conjugate to k ∧ σ) and
K : L∞ (Σ) → R, defined by




S(ϕ) = − Hv (s, t, ȳ, v̄, p̄, ᾱ) ϕ ds dt for all ϕ ∈ L(k∧σ) (Σ),
Σ

K(χ) = ζ̄, gv (ȳ, v̄) χ∗,Σ for all χ ∈ L∞ (Σ).

The optimality condition (3.6) can be rewritten as S(χ) = K(χ) for all χ ∈ L∞ (Σ).
Due to the Hahn–Banach extension theorem, there exists µ̄ ∈ Lk∧σ (Σ) such that


(4.5) µ̄ χ ds dt = ζ̄, gv (ȳ, v̄) χ∗,Σ for all χ ∈ L∞ (Σ).
Σ

Moreover, we have




µ̄ ϕ ds dt = S(ϕ) for all ϕ ∈ L(k∧σ) (Σ).
Σ

If we set η̄ = gv (ȳ, v̄)−1 µ̄, due to assumption (A7), η̄ belongs to Lk∧σ (Σ), and (4.5)
is equivalent to the following equation:


(4.6) η̄χ ds dt = ζ̄, χ∗,Σ for all χ ∈ L∞ (Σ).
Σ

The complementarity condition, the adjoint equation, and the optimality condition
for v̄ follow from (3.4), (3.5), (3.6), and (4.6). Let us prove the nontriviality condition.
If ᾱ = 0, the proof is complete. If ᾱ = 0, then due to (3.3) we have ζ̄ = 0, and from
(4.6) it follows that η̄ = 0.
Step 2. Let us prove that η̄ belongs to Lk (Σ). Let σ1 be such that 1 < σ1 < NN+1 .
Due to step 1 and assumption (A6) the function gy (ȳ, v̄)η̄ belongs to Lk∧σ1 (Σ).
From Assumptions (A3)–(A5) and from Proposition 2.5, it follows that p̄|Σ belongs
to Lσ2 (Σ) for all σ2 satisfying

N +1 N +1 1
(4.7) σ 1 < σ2 and < + .
2σ1 2σ2 2

Let σ2 satisfy (4.7). From the regularity of p|Σ , we deduce that Hv (·, ȳ, v̄, p̄, ᾱ) belongs
to Lk∧σ2 (Σ). From (4.3), it follows that gv (ȳ, v̄)η̄ belongs to Lk∧σ2 (Σ). With (A7), we
deduce that η̄ belongs to Lk∧σ2 (Σ), and with (A6) that gy (ȳ, v̄)η̄ belongs to Lk∧σ2 (Σ)
for all σ2 satisfying (4.7). After a finite number of iterations, we can prove that η̄
belongs to Lk (Σ). The proof is complete.
1400 N. ARADA AND J.-P. RAYMOND

Remark 4.2. Let us observe that (3.1) can be easily verified when C = {z ∈
L∞ (Σ) | z ≤ 0}, and Ψ is of the form Ψ(·, y, v) = ψ(·, y) − v. For ε > 0, let us set
wε = v̄ − εgv (ȳ, v̄)−1 + gv (ȳ, v̄)−1 gy (ȳ, v̄)zv̄ . Let ξε be the solution to the equation

∂ξ
+ Aξ + Φy (·, ȳ)ξ = 0 in Q,
∂t
∂ξ
+ ψy (·, ȳ)ξ + gv (ȳ, v̄)−1 gy (ȳ, v̄)ξ = wε on Σ, ξ(0) = 0 in Ω.
∂nA

It is clear that ξε is the solution of (3.2) corresponding to v = v̄ − εgv (ȳ, v̄)−1 −

gv (ȳ, v̄)−1 gy (ȳ, v̄)(ξε − zv̄ ), which yields to

g(ȳ, v̄) + gy (ȳ, v̄)(zv̂ − zv̄ ) + gv (ȳ, v̄)−1 (v − v̄) = −ε + g(ȳ, v̄) ≤ −ε.

Therefore, the pair (v, zv̂ ) satisfies the condition (3.1). In this case, we can set
ᾱ = 1 in the statement of Theorem 4.1.
5. Other regularity results. In this section, we are concerned with problem
(P) when C = ({z ∈ L∞ (Σ) | z ≤ 0})
. In this case, (4.1) is equivalent to


(5.1) (ᾱ, η̄) = 0, η̄ ≥ 0, η̄ g(ȳ, v̄) ds dt = 0.
Σ

In section 5.1, we suppose that = 2 and g satisfies the following separation condition.
(A8) There exists ε > 0 such that g1 (ȳ, v̄) + ε ≤ −g2 (ȳ, v̄) almost everywhere

(a.e.) on Σ. Moreover, we suppose that, for i = 1, 2, (giv (ȳ(·), v̄(·)))−1 belongs to
∞
L (Σ).
Remark 5.1. In examples studied in section 5 (Corollary 5.3 and examples in
section 5.3), we are able to verify the qualification condition (3.1) by using both the
separation condition and the properties of the state equation. But the separation
condition is of a different nature since it gives the regularity of the multipliers (which
is not the case of condition (3.1)), and in general it does not give optimality conditions
in qualified form (which is the case of condition (3.1)).
With (A8) we are able to prove that the supports of the multipliers associated
with the two constraints are disjoint. To prove such a result, we use the isomorphism
between (L∞ (Σ)) and the space of bounded Radon measures on Σ# , where Σ# is a
compact Hausdorff space [9].
In section 5.2, we suppose that g satisfies a monotonicity condition of the following
form.
 
(A9) For i = 1, . . . , , giv (ȳ, v̄) ≥ 0 a.e. on Σ, and (giv (ȳ(·), v̄(·)))−1 belongs to
∞
L (Σ).
In this case, the regularity of multipliers follows from properties of nonnegative
additive measures, and from the Radon–Nikodym theorem.
In section 5.3, we study a problem where the above separation and monotonicity
conditions are coupled.
5.1. Regularity of multipliers with a separation assumption.
Theorem 5.2. Let (ȳ, v̄) be a solution to (P). Suppose that (A1)–(A6) and (A8)
are fulfilled. Then, there exist ᾱ ∈ [0, 1], η̄ ∈ (Lk (Σ))2 , and p̄ ∈ L1 (0, T ; W 1,1 (Ω))
satisfying (4.2), (4.3), and (5.1). If, in addition, v satisfies (3.1), then we can take
ᾱ = 1.
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1401

Recall that k is the exponent introduced in the introduction. It is known that

(L∞ (Σ)) can be identified with the space of bounded finitely additive measures van-
ishing on zero-measure sets [20, Theorem 2.3]. Below we identify (L∞ (Σ)) with
M(Σ# ), the space of Radon measures on Σ# , where Σ# is a compact Hausdorff space.
This identification is useful to characterize the supports of elements of (L∞ (Σ)) .
Proof. Step 1. Due to Theorem 3.1, there exist ᾱ ∈ [0, 1], ζ̄ = (ζ̄1 , ζ̄2 ) ∈
((L∞ (Σ)) )2 , and p̄ ∈ L1 (0, T ; W 1,1 (Ω)) such that (3.3)–(3.6) hold. Condition (3.4)
may be rewritten as

(5.2) ζ̄i ≥ 0 and ζ̄i , gi (ȳ, v̄)∗,Σ = max {ζ̄i , z∗,Σ | z ∈ C} = 0 for i = 1, 2.

Step 2. Let us denote by So the closed unit sphere of (L∞ (Σ)) (for the weak-star
topology), and let us set

Σ# = {q ∈ So | q = 0, q, h∗,Σ = q, f ∗,Σ q, g∗,Σ if h = g f a.e. on Σ}.

It is well known [9, Theorem 11, p. 445] that Σ# is a compact Hausdorff space.
Moreover, there exists an isometric homomorphism τ from L∞ (Σ) onto C(Σ# ). The
isomorphism τ maps nonnegative functions into nonnegative functions and is an alge-
braic isomorphism in the sense that if χ = χ1 χ2 a.e. on Σ, then τ (χ) = τ (χ1 )τ (χ2 ). If
f is an arbitrary real continuous function, and χ is in L∞ (Σ), then τ (f (χ)) = f (τ (χ)).
Hence, for i = 1, 2, the measure ζ̄i ∈ (L∞ (Σ)) can be identified with ζ̄ i ∈ M(Σ# )
(the space of Radon measures on Σ# ), via the formula

ζ̄i , ψΣ# = ζ̄i , τ −1 (ψ)∗,Σ for all ψ ∈ C(Σ# ),

where ·, ·Σ# denotes the duality pairing between M(Σ# ) and C(Σ# ), and where
τ −1 is the inverse mapping of τ . (For more details see [2].)
Let us prove that under the separation condition (A8), the supports of ζ̄1 and ζ̄2
(denoted by supp ζ̄ i ) are disjoint. The condition (5.2) is rewritten as

ζ̄ i ≥ 0 and ζ̄ i , τ (gi (ȳ, v̄))Σ# = 0 for i = 1, 2.

Let us set Σi = {q ∈ Σ# | τ (gi (ȳ, v̄))(q) =  q, gi (ȳ, v̄) ∗,Σ = 0}. Since the mapping
q → τ (gi (ȳ, v̄))(q) is continuous for the weak-star topology of (L∞ (Σ)) , Σi is closed
for this topology. Therefore we have

(5.3) suppζ̄ i ⊂ Σi .

On the other hand, due to the positivity property of τ and due to (A8), we have

(5.4) τ (g1 (ȳ, v̄)) (q) − ε = τ (g1 (ȳ, v̄) − ε) (q) ≤ τ (g2 (ȳ, v̄)) (q) for all q ∈ Σ# .

From (5.3) and (5.4), we deduce that supp ζ̄1 ∩ supp ζ̄2 = ∅.
Step 3. Now, let us establish the regularity of gv (ȳ, v̄)∗ ζ̄. First, notice that (3.6)
can be stated in the form


Hv (·, ȳ, v̄, p̄, ᾱ)χ ds dt + ζ̄1 , τ (g1v
 
(ȳ, v̄)) τ (χ)Σ# + ζ̄2 , τ (g2v (ȳ, v̄)) τ (χ)Σ# = 0
Σ
1402 N. ARADA AND J.-P. RAYMOND

for all χ ∈ L∞ (Σ). Consider the linear operators S : L(k∧σ) (Σ) → R and K :
L∞ (Σ) → R defined by




S(ϕ) = − Hv (·, ȳ, v̄, p̄, ᾱ) ϕ ds dt for all ϕ ∈ L(k∧σ) (Σ),
Σ

 
K(χ) = ζ̄1 , τ (g1v (ȳ, v̄)) τ (χ)Σ# + ζ̄2 , τ (g2v (ȳ, v̄)) τ (χ)Σ# for all χ ∈ L∞ (Σ).

With arguments similar to those of Theorem 4.1, we can prove the existence of a
function µ ∈ Lk∧σ (Σ) satisfying


 
(5.5) ζ̄1 , τ (g1v (ȳ, v̄)) τ (χ)Σ# + ζ̄2 , τ (g2v (ȳ, v̄)) τ (χ)Σ# = µ χ ds dt
Σ

for all χ ∈ L∞ (Σ). Since supp ζ̄1 and supp ζ̄2 are two disjoint compact subsets of
Σ# , there exists ψo ∈ C(Σ# ), such that

0 ≤ ψo ≤ 1, ψo ≡ 1 on supp ζ̄1 , and ψo ≡ 0 on supp ζ̄2 .

Letting χ̃ be in L∞ (Σ) and setting χ1 = τ −1 (ψo ) χ̃ and χ2 = (1 − ψo ) χ̃ in (5.5), we

obtain


 
(5.6) ζ̄i , giv (ȳ, v̄)χ̃∗,Σ = ζ̄i , τ (giv (ȳ, v̄)) τ (χ̃)Σ# = µ̄i χ̃ ds dt
Σ

for i = 1, 2, with µ̄1 = µ τ −1 (ψo ) and µ̄2 = µ τ −1 (1 − ψo ). It is clear that

µ̄1 and µ̄2 belong to Lk∧σ (Σ) for all σ < NN+1 . Let us set η̄i = (giv 
(ȳ, v̄))−1 µ̄i
for i = 1, 2. We have proved that (4.2), (4.3), and (5.1) are satisfied with η̄ =
(η̄1 , η̄2 ) ∈ Lk∧σ (Σ). We conclude with a bootstrap process as in step 2 of the proof of
Theorem 4.1.
Consider the following example:

(P1 ) inf {J(y, v) | v ∈ L∞ (Σ), (y, v) satisfies (1.1) and 0 ≤ v ≤ γ(y) + c},

where Ψ is of the form Ψ(·, y, v) = ψ(·, y)−v. It is a particular case of (P) correspond-
ing to g1 (s, t, y, v) = v−γ(s, t, y)−c(s, t) and g2 (y, v) = −v. We suppose that c belongs
to L∞ (Σ), and γ is defined either by γ(s, t, y) = b(s, t)y, or by γ(s, t, y) = φ(y), where
b belongs to L∞ (Σ), b ≥ 0, and φ is a nondecreasing function of class C 1 .
Corollary 5.3. Let (ȳ, v̄) be a solution of (P1 ), and suppose that (A1)–(A5)
are fulfilled. Suppose in addition that there exists ε > 0 such that

(5.7) γ(ȳ) + c ≥ ε a.e. on Σ.

Then, there exist η̄ ∈ (Lk (Σ))2 and p̄ ∈ L1 (0, T ; W 1,1 (Ω)), such that (4.2), (4.3), and
(5.1) are satisfied with ᾱ = 1.
Remark 5.4. Observe that if yo ≥ 0, then ȳ ≥ 0. If φ is nonnegative on R+ and
if c ≥ ε, then condition (5.7) is satisfied. The case when γ(s, t, y) = b(s, t)y is studied
in [6]. The proof in [6] is based on duality techniques. Here it is a direct consequence
of Theorem 5.2.
Proof. The separation condition (A8) is nothing else than (5.7). Therefore, due
to Theorem 5.2, we can state optimality conditions in nonqualified form. To prove
the corollary, we have only to check that there exists v̂ ∈ L∞ (Σ) such that (3.1) is
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1403

satisfied. If γ(s, t, y) = b(s, t)y, we set β = b, and if γ(s, t, y) = φ(y), we set β = φ (ȳ).
Therefore β ≥ 0. For λ > 0, we set wλ = v̄ − λε − βzv̄ . Let ξλ be the solution to

∂ξ ∂ξ
+ Aξ + Φ̄y ξ = 0 in Q, + ψ̄y ξ − βξ = wλ on Σ, ξ(0) = 0 in Ω,
∂t ∂nA

where Φ̄y = Φy (·, ȳ) and ψ̄y = ψy (·, ȳ). Observe that ξλ = zṽλ is the solution of (3.2)
for vλ = v̄ − λε + β(ξλ − zv̄ ), and

(5.8) −β(zṽλ − zv̄ ) + ṽλ − v̄ = −λε.

By making the difference between the equation satisfied by zv̄ and the equation sat-
isfied by ξλ , we can prove that

(5.9) ξλ = zṽλ ≤ zv̄ and ||zṽλ − zv̄ ||C(Q) ≤ Cλε.

By combining (5.9) and (5.8), we prove that there exists a constant C̄ > 0 such that

(5.10) ||
vλ − v̄||∞,Σ ≤ C̄ λε.

Set Σ0 = {(s, t) ∈ Σ | v̄ = 0}, Σλ = {(s, t) ∈ Σ | 0 < v̄ ≤ 2C̄λε}, and let χλ

(respectively, χ0 ) be the characteristic function of Σλ (respectively, Σ0 ). Set vλ =
λεχ0 + (v̄ + λε)χλ + vλ (1 − χ0 − χλ ). We claim that the pair (zv̂λ , vλ ) satisfies (3.1)
for λ small enough.
• On Σ0 ∪ Σλ , we have vλ ≥ λε > 0. Due to (5.10), on Σ \ (Σ0 ∪ Σλ ) we have
vλ = vλ ≥ v̄ − |
vλ − v̄| ≥ C̄λε. Thus we have vλ ≥ min(λε, C̄λε).
• With (5.10) we obtain

||zv̂λ − zṽλ ||C(Q) ≤ C||(v̄ − vλ + λε)(χλ + χ0 )||∞,Σ ≤ Cλε.

Now, from (5.9) we deduce ||zv̂λ −zv̄ ||C(Q) ≤ ||zv̂λ −zṽλ ||C(Q) +||zṽλ −zv̄ ||C(Q) ≤ Cλε.
Due to (5.7), we have

v̄ − γ(ȳ) − c − β(zv̂λ − zv̄ ) + (vλ − v̄) ≤ 2C̄λε − ε − β(zv̂λ − zv̄ ) + λε on Σ0 ∪ Σλ .

Therefore, we can choose λ > 0 small enough to have

ε
v̄ − γ(ȳ) − c − β(zv̂λ − zv̄ ) + (vλ − v̄) ≤ − on Σ0 ∪ Σλ .
2
Since ṽλ = −λε + β(zṽλ − zv̄ ) < 0 on Σ0 (because of (5.9)), we have vλ − ṽλ ≥ 0 on
Σ \ Σλ . Hence if ζλ is the solution to (3.2) for v = (vλ − ṽλ )χλ = (v̄ + λε − ṽλ )χλ , we
1
have zv̂λ − zṽλ ≥ ζλ , and ||ζλ ||C(Q) ≤ C||(v̄ + λε − ṽλ )χλ ||r,Σ ≤ Cλε LN (Σλ ) r , with
r > N + 1, and LN (Σλ ) is the Lebesgue measure of Σλ . With (5.8) and the bound
for ζλ , we have the following estimate on Σ \ (Σ0 ∪ Σλ ):

v̄ − γ(ȳ) − c − β(zv̂λ − zv̄ ) + (vλ − v̄) ≤ −β(zṽλ − zv̄ ) + (

vλ − v̄) − β(zv̂λ − zṽλ )
1
≤ −λε + ||β||∞,Σ ||ζλ ||∞,Σ ≤ −λε + Cλε LN (Σλ ) r .
1
Since −λε + Cλε LN (Σλ ) r < − 12 λε for λ > 0 small enough, the proof is
complete.
1404 N. ARADA AND J.-P. RAYMOND

5.2. Regularity of multipliers with a monotonicity assumption.

Theorem 5.5. Let (ȳ, v̄) be a solution of (P). Suppose that (A1)–(A6) and (A9)
are fulfilled. Then, there exist ᾱ ∈ [0, 1], η̄ ∈ (Lk (Σ))
, and p̄ ∈ L1 (0, T ; W 1,1 (Ω)),
such that conditions (4.2), (4.3), and (5.1) are satisfied.
Proof. Due to Theorem 3.1, there exist ᾱ ∈ [0, 1], ζ̄ = (ζ̄1 , . . . , ζ̄
) ∈ ((L∞ (Σ))
) ,
and p̄ ∈ L1 (0, T ; W 1,1 (Ω)) such that (3.3)–(3.6) hold. With arguments similar to
those of the proof of Theorem 5.2, we show that

(5.11) ζ̄i ≥ 0, ζ̄i , gi (ȳ, v̄)∗,Σ = 0 for all i = 1, . . . , .

Moreover, as in the proof of Theorem 4.1, with the Hahn–Banach extension theorem,
we can establish the existence of a multiplier µ̄ belonging to Lk∧σ (Σ) for all σ < NN+1 ,
such that


gv (ȳ, v̄)∗ ζ̄, χ∗,Σ = µ̄ χ ds dt for all χ ∈ L∞ (Σ).
Σ


Due to (A9), for i = 1, . . . , , the additive measures giv (ȳ, v̄)∗ ζ̄i are nonnegative. From
a decomposition theorem for nonnegative finitely additive measures [9, Theorem 8, p.

163], we deduce that giv (ȳ, v̄)∗ ζ̄i admits a unique decomposition

giv (ȳ, v̄)∗ ζ̄i = µ̄i + ν̄i ,

where µ̄i are nonnegative countably additive measures, and ν̄i are nonnegative purely
finitely additive set functions in (L∞ (Σ)) . It follows that





gv (ȳ, v̄)∗ ζ̄ = µ̄i + ν̄i = µ̄ ds dt.
i=1 i=1

Thus, the pair (Σ
i=1 µ̄i , Σ
i=1 ν̄i ) is a decomposition of µ̄ dsdt in a nonnegative count-
ably additive measure and a nonnegative purely finitely additive set function. From
the uniqueness of this decomposition, we deduce that ν̄i = 0. Since µ̄i ≥ 0 for
i = 1, . . . , , Σ
i=1 µ̄i = µ̄ dsdt, and µ̄ belongs to Lk∧σ (Σ) for all σ < NN+1 , there exist
nonnegative functions ξi ∈ Lk∧σ (Σ) such that µ̄i = ξi ds dt (it is a consequence of the

Radon–Nikodym theorem). We set η̄i = (giv (ȳ, ū))−1 ξi . We have




giv (ȳ, v̄)∗ ζ̄i , χ∗,Σ = ξi χ ds dt = 
giv (ȳ, v̄)η̄i χ ds dt
Σ Σ
∞
for all χ ∈ L (Σ). Thus, (4.2), (4.3), and (5.1) are satisfied with η̄ = (η̄1 , . . . , η̄
) ∈
(Lk∧σ (Σ))
. We finish with a bootstrap argument as in the proof of Theorem
4.1.
Consider the following example.

(P2 ) inf {J(y, v) | v ∈ L∞ (Σ), (y, v) satisfies (1.1), v ≤ γi (y) + ci for i = 1, . . . , },

where Ψ is of the form Ψ(·, y, v) = ψ(·, y) − v. It is a particular case of (P) corre-

sponding to gi (y, v) = v − γi (y) − ci for i = 1, . . . , . We suppose that ci belongs
to L∞ (Σ), and γi is defined either by γi (s, t, y) = bi (s, t)y, or by γi (s, t, y) = φi (y),
where bi belongs to L∞ (Σ), and φi is of class C 1 .
Corollary 5.6. Let (ȳ, v̄) be a solution of (P2 ). Suppose that (A1)–(A5) are
fulfilled. Then there exist η̄ ∈ (Lk (Σ))
and p̄ ∈ L1 (0, T ; W 1,1 (Ω)), such that (4.2),
(4.3), and (5.1) are satisfied with ᾱ = 1.
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1405

Proof. The assumptions of Theorem 5.5 are clearly satisfied. We have only to
check that we can take ᾱ = 1. If γi (s, t, y) = bi (s, t)y, we set βi = bi , and if γi (s, t, y) =
φi (y), we set βi = φi (ȳ). For λ > 0 and i = 1, . . . , , we set wiλ = v̄ − λε − βi zv̄ . Let
ξλ be the solution to

∂ξ
+ Aξ + Φy (·, ȳ)ξ = 0 in Q,
∂t
(5.12)
∂ξ
+ ψy (·, ȳ)ξ = min (wiλ + βi ξ) on Σ, ξ(0) = 0 in Ω.
∂nA i=1,...,

Since the function “min” is Lipschitz with respect to its arguments, the existence of a
solution to (5.12) may be proved as in [16] (see also [6] where the same trick is used).
Observe that ξλ is the solution of (3.2) corresponding to vλ = mini=1,...,
(wiλ + βi ξ),
that is, ξλ = zṽλ . Moreover,

v̄ − γi (ȳ) − ci + (
vλ − v̄) − βi (zṽλ − zv̄ ) ≤ (
vλ − v̄) − βi (zṽλ − zv̄ ) ≤ −λε

vλ , zṽλ ).
for i = 1, . . . , . Thus (3.1) is satisfied by (
5.3. Coupling separation and monotonicity conditions. In this section we
study a problem for which the monotonicity and separation conditions are coupled.
We suppose that = 3 and that g = (g1 , g2 , g3 ) satisfies the assumption below.
(A10) There exists ε > 0 such that g1 (ȳ, v̄) + ε ≤ −g2 (ȳ, v̄) and g3 (ȳ, v̄) + ε ≤
−g2 (ȳ, v̄) a.e. on Σ. The pair (g1 (ȳ, v̄), g3 (ȳ, v̄)) satisfies the monotonicity condition

stated in (A9). The function (g2v (ȳ(·), v̄(·)))−1 belongs to L∞ (Σ).
The case of bottleneck type constraints studied in [5], [6] falls into this setting (see
Remark 5.8). We prove the existence of regular multipliers when (A10) is fulfilled.
At the end of the section we give two examples for which optimality conditions are
obtained in qualified form (that is, ᾱ = 1). The existence of regular multipliers for
these examples cannot be deduced from [6].
Theorem 5.7. Let (ȳ, ū) be a solution to (P ). Suppose that (A1)–(A6) and (A10)
are fulfilled. Then, there exist ᾱ ∈ [0, 1], η̄ ∈ (Lk (Σ))3 , and p̄ ∈ L1 (0, T ; W 1,1 (Ω)),
such that (4.2), (4.3), and (5.1) are satisfied. If, in addition, v satisfies (3.1), then we
can take ᾱ = 1.
Proof. Due to Theorem 3.1, there exist ᾱ ∈ [0, 1], ζ̄ ∈ ((L∞ (Σ))3 ) , and p̄ ∈
L1 (0, T ; W 1,1 (Ω)) such that (3.3)–(3.6) are satisfied. In particular, the condition
(3.4) is equivalent to

(5.13) ζ̄i ≥ 0 and ζ̄i , gi (ȳ, v̄)∗,Σ = 0

for i = 1, 2, 3. As in the proof of Theorem 5.2, we identify the additive measures ζ̄i
with measures ζ̄i belonging to M(Σ# ) (the space of Radon measures on Σ# ), and the
optimality condition (3.6) is

3

Hv (·, ȳ, v̄, p̄, ᾱ)χ ds dt + 
ζ̄i , τ (giv (ȳ, v̄)) τ (χ)Σ# = 0
Σ i=1

for all χ ∈ L∞ (Σ). (τ is the isometric homomorphism used in the proof of Theorem
5.2, and ·, ·Σ# denotes the duality pairing between M(Σ# ) and C(Σ# ).) With
the Hahn–Banach theorem, there exists a function µ belonging to Lk∧σ (Σ) for all
1406 N. ARADA AND J.-P. RAYMOND

N +1
1<σ< N , such that

3



ζ̄i , τ (giv (ȳ, v̄)) τ (χ)Σ# = µ χ ds dt for all χ ∈ L∞ (Σ).
i=1 Σ

By using (5.13), and since (g1 (ȳ, v̄), g2 (ȳ, v̄)) and (g2 (ȳ, v̄), g3 (ȳ, v̄)) satisfy the
separation condition stated in (A10), we can prove that

(5.14) supp ζ̄1 ∩ supp ζ̄2 = supp ζ̄2 ∩ supp ζ̄3 = ∅.

Since (supp ζ̄1 ∪ supp ζ̄3 ) and supp ζ̄2 are two disjoint compact subsets of Σ# , by
using the same method as in the proof of Theorem 5.2 we deduce that there exist µ̄1
and µ̄2 belonging to Lk∧σ (Σ) for all 1 < σ < NN+1 , such that


 
(5.15) ζ̄1 , g1v (ȳ, v̄)χ∗,Σ + ζ̄3 , g3v (ȳ, v̄)χ∗,Σ = µ̄1 χ ds dt for all χ ∈ L∞ (Σ),
Σ


(5.16) ζ̄2 , g2v (ȳ, v̄)χ∗,Σ = µ̄2 χ ds dt for all χ ∈ L∞ (Σ).
Σ


Let us set η̄2 = (g2v (ȳ, v̄))−1 µ̄2 . The function η̄2 belongs to Lk∧σ (Σ) for all 1 <
N +1  
σ < N . Since g1v (ȳ, v̄) and g3v (ȳ, v̄) satisfy (A9) with (5.13) and (5.15), as in the
proof of Theorem 5.5, we can establish the existence of two nonnegative functions
ξ¯1 ∈ Lk∧σ (Σ) and ξ¯3 ∈ Lk∧σ (Σ) such that



(5.17) 
g1v (ȳ, v̄)ζ̄, χ∗,Σ = ξ¯1 χ ds dt, 
g3v (ȳ, v̄)ζ̄3 , χ∗,Σ = ξ¯3 χ ds dt
Σ Σ

for all χ ∈ L∞ (Σ). We set η̄1 = (g1v 

(ȳ, v̄))−1 ξ¯1 and η̄3 = (g3v 
(ȳ, v̄))−1 ξ¯3 . The
k∧σ 3
conditions (4.2), (4.3), and (5.1) are satisfied with η̄ ∈ (L (Σ)) . We can conclude
with a bootstrap argument as in the proof of Theorem 4.1. It is clear that if, in
addition, v satisfies (3.1), then we can take ᾱ = 1. The proof is complete.
Finally, consider two examples for which the assumptions of Theorem 5.7 are
satisfied.
Example 1. Set g1 (y, v) = v − β1 y − c1 , g3 (y, v) = v − β3 y − c3 , g2 (y, v) =
−v, where β1 , β3 , c1 and c3 belong to L∞ (Σ), β1 ≥ 0, β3 ≥ 0. Suppose that
Ψ is of the form Ψ(·, y, v) = ψ(·, y) − v. Suppose that there exists ε > 0 such
that min(β1 ȳ + c1 , β3 ȳ + c3 ) ≥ ε. The pairs (g1 (ȳ, v̄), g2 (ȳ, v̄)) and (g2 (ȳ, v̄), g3 (ȳ, v̄))
satisfy the separation condition stated in (A10). The pair (g1 (ȳ, v̄), g3 (ȳ, v̄)) satisfies
the monotonicity condition stated in (A9). Thus (A10) is satisfied and Theorem 5.7
can be applied. Moreover, by combining the arguments of the proofs of Corollaries
5.3 and 5.6, we can prove the existence of v obeying (3.1). Thus we can take ᾱ = 1.
Remark 5.8. If we set β3 = 0 in the above example, we recover the case of
constraints considered in [6].
Example 2. We replace g1 and g3 in the above example by g1 (y, v) = v−φ1 (y)−c1 ,
g3 (y, v) = v − φ3 (y) − c3 , where φ1 and φ3 are nondecreasing functions of class C 1 .
Suppose that there exists ε > 0 such that min(φ1 (ȳ) + c1 , φ3 (ȳ) + c3 ) ≥ ε. Thus (A10)
is satisfied and Theorem 5.7 can be applied. As above, we can take ᾱ = 1.
 CONTROL PROBLEMS WITH MIXED CONSTRAINTS 1407

Acknowledgments. The authors wish to thank the anonymous referees for help-
ful remarks and suggestions that improved the presentation of the paper.

REFERENCES

[1] N. Arada and J.-P. Raymond, Necessary optimality conditions for control problems and the
Stone–C̆ech compactification, SIAM J. Control Optim., 37 (1999), pp. 1011–1032.
[2] N. Arada and J.-P. Raymond, Minimax control of parabolic systems with state constraints,
SIAM J. Control Optim., 38 (1999), pp. 254–271.
[3] A. V. Arutynov, Perturbations of extremal problems with constraints and necessary optimality
conditions, J. Sov. Math., 54 (1991), pp. 1342–1400.
[4] A. V. Arutynov and A. I. Okoulevitch, Necessary optimality conditions for optimal control
problems with intermediate constraints, J. Dynam. Control Systems, 4 (1998), pp. 49–58.
[5] M. Bergounioux and F. Tröltzsch, Optimal control of linear bottleneck problems, ESAIM
Control Optim. Calc. Var., 3 (1998), pp. 235–250.
[6] M. Bergounioux and F. Tröltzsch, Optimal control of semilinear parabolic equations with
state constraints of bottleneck type, ESAIM Control Optim. Calc. Var., 4 (1999), pp. 595–
608.
[7] E. Casas, Boundary control of semilinear elliptic equations with pointwise state constraints,
SIAM J. Control Optim., 31 (1993), pp. 993–1006.
[8] E. Casas, J.-P. Raymond, and H. Zidani, Pontryagin’s principle for local solutions of control
problems with mixed control-state constraints, SIAM J. Control Optim., 39 (2000), pp.
1182–1203.
[9] N. Dunford and J. T. Schwartz, Linear Operators, Part 1, Interscience Publishers, New
York, London, 1958.
[10] H. O. Fattorini, Infinite Dimensional Optimization and Control Theory, Encyclopedia of
Mathematics and its Applications, Cambridge University Press, Cambridge, UK, 1999.
[11] R. F. Hartl, S. P. Sethi, and R. G. Vickson, A survey of the maximum principles for
optimal control problems with state constraints, SIAM Rev., 37 (1995), pp. 181–218.
[12] X. J. Li and J. Yong, Optimal Control Theory for Infinite Dimensional Systems, Birkhäuser,
Boston, Basel, Berlin, 1995.
[13] K. Makowski and L. W. Neustadt, Optimal control problems with mixed control-phase vari-
able equality and inequality constraints, SIAM J. Control, 12 (1974), pp. 184–228.
[14] L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze, and E. F. Mischenko, The
Mathematical Theory of Optimal Processes, Wiley-Interscience, New York, 1962.
[15] J.-P. Raymond, Nonlinear boundary control of semilinear parabolic equations with pointwise
state constraints, Discrete Contin. Dynam. Systems, 3 (1997), pp. 341–370.
[16] J.-P. Raymond and H. Zidani, Hamiltonian Pontryagin’s principles for control problems gov-
erned by semilinear parabolic equations, Appl. Math. Optim., 39 (1999), pp. 143–177.
[17] J.-P. Raymond and H. Zidani, Time optimal problems with boundary controls, Differential
Integral Equations, 13 (2000), pp. 1039–1072.
[18] T. Roubic̆ek, Relaxation in Optimization Theory and Variational Calculus, de Gruyter Ser.
Nonlinear Anal. Appl. 4, Walter de Gruyter, Berlin, 1997.
[19] F. Tröltzsch, Optimality Conditions for Parabolic Control Problems and Applications,
Teubner-Texte Math. 62, B.G. Teubner Verlagsgesellschaft, Leipzig, 1984.
[20] K. Yosida and E. Hewitt, Finitely additive measures, Trans. Amer. Math. Soc., 72 (1952),
pp. 46–66.

View publication stats