JOURNAL OF OPTIMIZATION THEORY AND APPLICATIONS: Vol. 131, No. 3, pp.

365–382, December 2006 (

C 2006)

DOI: 10.1007/s10957-006-9150-4

Semivectorial Bilevel Optimization Problem:

Penalty Approach
H. BONNEL1 AND J. MORGAN2

Communicated by H. P. Benson
Published Online: 5 December 2006

Abstract. We consider a bilevel optimization problem where the upper level

is a scalar optimization problem and the lower level is a vector optimization
problem. For the lower level, we deal with weakly efficient solutions. We
approach our problem using a suitable penalty function which vanishes over
the weakly efficient solutions of the lower-level vector optimization problem
and which is nonnegative over its feasible set. Then, we use an exterior penalty
method.

Key Words. Bilevel optimization, multiobjective optimization, optimiza-

tion over the efficient set, penalty methods.

1. Introduction

In this paper, we consider a bilevel problem, which we call semivectorial

bilevel optimization problem, where the upper level is a scalar optimization prob-
lem and the lower level is a vector optimization problem. Therefore, we deal with
the following problem:
(BLW ) min f (x, y),
x,y
s.t. (x, y) ∈ U ⊆ X × Y, (upper-level constraints)
 
y ∈ w-ARG MIN

C {F (x, y ) | y ∈ S(x)},
y
(lower-level w-efficiency)
where X and Y are Hausdorff topological spaces, Z is a real Banach space, C ⊆ Z is
a pointed convex cone [i.e. R+ C + C ⊆ C and C ∩ (−C) = {0}] with topological
interior int C = ∅, f (the upper-level objective) is a real-valued map on X ×
1 (Formerly Serban Bolintinéanu) Professor, University of New Caledonia, ERIM, Nouméa, New
Caledonia. This author thanks the University of Naples Federico II for its support and the Department
of Mathematics and Statistics for its hospitality.
2 Professor, University of Naples Federico II, Department of Mathematics and Statistics, Napoli, Italy.

365
0022-3239/06/1200-0365/0

C 2006 Springer Science+Business Media, Inc.
366 JOTA: VOL. 131, NO. 3, DECEMBER 2006

Y, U (the upper-level feasible set) is a subset of X × Y, F (the lower-level vector

objective) is a map from X × Y to Z, and S : X → → Y is a set-valued map. For each
x ∈ X, S(x) stands for the lower-level feasible set and w-ARGMINy  C {F (x, y  ) |
y  ∈ S(x)} is the set of the weakly-efficient solutions (see Section 2 for details)
associated to the lower-level vector optimization problem:

(LL)(x) MINIMIZE

C F (x, y ), s.t. y  ∈ S(x).
y
We assume that there exists (x̌, y̌) ∈ U such that
y̌ ∈ w-ARGMINy  C {F (x̌, y  ) | y  ∈ S(x̌)}.
Problem (BLw ) covers in particular the following important problems.
(A) Optimization over a weakly-efficient set,
(OES) min f1 (y), s.t. y ∈
1 ,
y

where
1 is the set of weakly-efficient solutions associated to the vector

optimization problem
MINIMIZE C F1 (y), s.t. y ∈ S1 ,
y

with f1 : Y → R, F1 : Y → Z, S1 ⊆ Y .
This problem is an important tool in the decision-making theory. The main
difficulty to solve Problem (OES) is due to its nonconvexity. Even when all the
functions are linear and S1 is a polyhedron, Problem (OES) is not convex. Another
difficulty is given by the fact that
1 , the feasible set of Problem (OES), is not
known explicitly. For these reasons, Problem (OES) was studied intensively during
the last two decades by many authors. Thus, the all linear case (i.e., when
1 is
a linear functional and the vector optimization problem has linear objectives and
linear constraints) has been studied theoretically and some algorithms have been
proposed in Refs. 1–5. Some nonlinear cases dealing also with theoretical aspects
and algorithms can be found in Refs. 6–13. For a survey paper and an extensive
bibliography, see Ref. 14.
Problem (OES) is obtained as a particular case of Problem (BLw ) considering
X = {0}, U = X × Y , and for every y ∈ Y ,
f (0, y) = f1 (y), F (0, y) = F1 (y), S(0) = S1 .

(B) The usual scalar bilevel programming problem,

(SBP) min f (x, y), s.t. G(x, y) ≤ 0, y ∈
(x),
x,y

where
(x) is the solution set of the lower-level scalar minimization

problem
min

F (x, y  ), s.t. H (x, y  ) ≤ 0,
y
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 367

with
x ∈ Rnx , y ∈ Rny , f, F : Rnx +ny → R, G : Rnx +ny → Rnu ,
H : Rnx +ny → Rnl .
We obtain Problem (SBP) from Problem (BLw ) considering X = Rnx , Y =
R , Z = R, C = R+ ,
ny

U = {(x, y) ∈ X × Y | G(x, y) ≤ 0},

and for each x ∈ X,
S(x) = {y ∈ Y | H (x, y) ≤ 0}.
This problem has been investigated by one of the authors (in particular, see
Refs. 15–18) and by many authors (see for example Ref. 19 and Ref. 20 for an
extensive and recent bibliography with more than 400 references).
In this work, we approach Problem (BLw ) using a suitable penalty function
h : X × Y → R ∪ {+∞}. This function allows us to describe, for each x ∈ X,
every weakly efficient solution y of Problem (LL(x)) by the equality h(x, y) =
0. Moreover, we have h(x, y) ≥ 0 for every feasible solution y of (LL(x)). In
Theorems 3.1 and 3.2, we obtain, under different assumptions, the continuity of h
over the feasible set of Problem (BLw ); then, we use an exterior penalty method
in Theorems 3.3 and 3.4.

2. Preliminaries

Let · denote the norm of Z and let ·, · denote the duality scalar product
between Z ∗ (the topological dual of Z) and Z. The extended space Z̄ = Z ∪
{−∞C , +∞C } has been introduced in Ref. 21. Recall that a neighborhood of +∞C
is a set N ⊆ Z̄ containing a + C ∪ {+∞C }, for some a ∈ Z, and its opposite −N
is a neighborhood of −∞C . The partial-order relation, compatible with the linear
structure of Z, given by
∀z, z ∈ Z, z ≤C z ⇐⇒ z − z ∈ C,
and the transitive relation
∀z, z ∈ Z, z <C z ⇐⇒ z − z ∈ int C,
are extended respectively to Z̄ by
∀z ∈ Z, −∞C <C z <C +∞C , −∞C ≤C z ≤C +∞C .
Note that the embedding Z ⊆ Z̄ is continuous and dense.
Let G be an extended-valued map from Y to Z ∪ {+∞C }, which is proper
(i.e., not identically equal to +∞C ) and let S1 ⊆ Y be a nonempty set.
The effective domain of G is
dom (G) = {y ∈ Y | G(y) = +∞C },
368 JOTA: VOL. 131, NO. 3, DECEMBER 2006

the positive polar cone of C is

C + = {λ ∈ Z ∗ | λ, z ≥ 0, ∀z ∈ C},
and we denote
C0+ = {λ ∈ Z ∗ | λ, z > 0, ∀z ∈ C\{0}}.
We extend by continuity every λ ∈ C + \{0} to Z̄, setting (see Ref. 21, 22 for more
details)
λ, ±∞C  = ±∞.
We denote by T̄ the topological closure in the topological space Z̄ of the set
T ⊆ Z̄. The weakly-infimal set of a subset T ⊆ Z̄ is the set
w-INFC (T ) = {z ∈ T̄ |  ∃v ∈ T , v <C z}.
We recall the following result (see Ref. 21).

Proposition 2.1. Let T be a nonempty subset of Z̄. The following state-

ments are equivalent:
(i) −∞C ∈ w-INFC (T ).
(ii) w-INFC (T ) = {−∞C }.
(iii) There exists a sequence (zn )n∈N in T such that zn → −∞C (for the
topology of Z̄).
(iv) −∞C ∈ T̄ .

For the vector optimization problem of the form

MINIMIZE C G(y), s.t. y ∈ S1 ,
y

a point a ∈ Y is called:
(i) weakly-efficient (or weakly-Pareto) solution if a ∈ S1 and there is no
y ∈ S1 verifying
G(y) <C G(a) [in other words, G(a) ∈ w-INFC (G(S1 ))];
(ii) efficient (or Pareto) solution if a ∈ S1 and there is no y ∈ S1 verifying
G(y) ≤C G(a) and G(y) = G(a).
The set of weakly-efficient solutions is denoted by w-ARGMINC {G(y) | y ∈
S1 }. To solve the vector optimization problem
(VOP) MINIMIZE C G(y), s.t. y ∈ Y,
y

means to find the set of the weakly-efficient solutions.

 JOTA: VOL. 131, NO. 3, DECEMBER 2006 369

Notice that a particular constrained vector optimization problem,

(CVOP) MINIMIZE C G0 (y), s.t. y ∈ S1 ,
y

where S1 ⊆ Y, S1 = Y is the feasible set and G0 is a map from S1 to Z, is equivalent

to the unconstrained extended-valued vector optimization problem (VOP) with


G0 (y), if y ∈ S1 ,
G(y) =
+∞C , if y ∈ Y \S1 .
Problems (CVOP) and (VOP) are equivalent in the sense that they have the
same weakly-efficient solutions and the same weakly-infimal set.

3. Penalty Approach for Problem (BLW )

From now on, we suppose that the ordering cone C is also closed. For any
→ B, where A and B are some sets, we denote by
set-valued map  : A →
Gr() := {(a, b) ∈ A × B | b ∈ (a)}
the graph of  and by
dom() := {a ∈ A | (a) = ∅}
the domain of .
Note first that Problem (BLW ) is equivalent to the problem
(GMW ) min f (x, y), s.t. (x, y) ∈ Gr (
w ) ∩ U,
x,y

→ Y is defined, for each x ∈ X, by

where the set-valued map
w : X →


w (x) = w-ARGMIN

C F (x, y ).
y ∈S(x)

In this section, we use a penalty function (inspired from Ref. 9) which allows
us to describe the graph of the set-valued map x →
w (x) as an equality constraint.
Note that Reference 9 deals with the problem of optimizing a scalar function over
the weakly-efficient set [Problem OES, see Introduction] in finite-dimensional
spaces, while the present paper deals with the much more general problem of
semivectorial bilevel optimization, where the function F depends also on the
variable x of the upper level. Moreover, in this paper, we work in non-necessarily
finite-dimensional spaces.
In the sequel, e is a fixed vector belonging to int C. We consider the function
h : X × Y → R ∪ {+∞} given, for each (x, y) ∈ X × Y , by
h(x, y) = sup{t ∈ R | ∃y  ∈ S(x) : F (x, y  ) ≤C F (x, y) − te}. (1)
370 JOTA: VOL. 131, NO. 3, DECEMBER 2006

Thus, h(x, y) represents the optimal value of the maximization problem

(MPx,y ) max t, s.t. y  ∈ S(x), F (x, y  ) ≤C F (x, y) − te.
(t,y  )∈R×Y

Lemma 3.1. We have the following statements:

(i) h(Gr(S)) ⊆ R+ ∪ {+∞}.
(ii) Gr(
W ) = h−1 ({0}) ∩ Gr(S).
In other words, for every (x, y) ∈ Gr(S), we have:
(i ) h(x, y) ≥ 0,
(ii ) h(x, y) = 0 ⇐⇒ (x, y) ∈ Gr(
W ).

Proof. Let (x, y) ∈ Gr(S). It follows that (0, y) is feasible for the maxi-
mization problem defining the function h; hence, (i) holds.
Let (x, y) ∈ Gr(
W ). We have that y ∈
W (x); i.e., y is a weakly-efficient
solution to Problem (LL)(x). It follows that y ∈ S(x); i.e., (x, y) ∈ Gr(S). From
(i), we have that h(x, y) ≥ 0. If h(x, y) = 0, then we can find y  ∈ S(x) and a real
number t > 0 such that F (x, y  ) ≤C F (x, y) − te, hence F (x, y  ) <C F (x, y).
This is in contradiction to the fact that y is a weakly-efficient solution for Problem
(LL)(x); therefore, h(x, y) = 0.
Conversely, let (x, y) ∈ Gr(S) such that h(x, y) = 0. If (x, y) ∈/ Gr(
W ), then
y is not a weakly-efficient solution for Problem (LL)(x); hence, there exists some
y  ∈ S(x) such that F (x, y  ) <C F (x, y). Thus, using the continuity of the map
t → t · e, there exists some t > 0 such that F (x, y  ) ≤C F (x, y) − te. Therefore,
h(x, y) > 0 and we get a contradiction.

Using Proposition 2.1 and the definition of h, we obtain easily the following
remark.

Remark 3.1. The function h has only finite values ⇐⇒ −∞C ∈ /

w-INFC (F (Gr(S)). This happens for instance when w-INFC (F (Gr(S)) ∩ Z = ∅.

Theorem 3.1. Let Z be endowed with the norm (strong) topology. Let X
and Y be Hausdorff topological spaces. Assume the following:

(A1) For any x ∈ X, there exist a neighborhood Nx of x, a neighborhood

NYx of the set ∪x  ∈Nx S(x  ), and two compact sets Kx ⊆ X and KYx ⊆ Y
such that
 
Nx × NYx ∩ Gr(S) ⊆ Kx × KYx .
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 371

(A2) The restriction of the map F to Gr(S) is continuous.

(A3) The set-valued map S : X → → Y is continuous3 and has closed val-
ues; i.e., for any x ∈ X, S(x) is a closed subset of Y 4 . Then, the re-
striction of the function h to Gr(S) is finite and continuous at every
point.

Proof. We will use the following classical result (see Ref. 23, Proposition
3.2; see also Ref. 24).

Proposition 3.1. Let X̃ and V be Hausdorff topological spaces, let  :

V→→ X̃ be a closed set-valued map (i.e., the set Gr() is a closed subset of
V × X̃), and let ϕ : X̃ × V → R be a continuous function. Consider the optimal
value function v : V → R̄ and the maximizer set-valued map M : V →→ X̃, given
for any u ∈ V by
v(u) = sup ϕ(x̃, u),
x̃∈(u)
M(u) = argmax ϕ(x̃, u).
x̃∈(u)

Let u0 ∈ V be given. Assume that:

(i) There exists α ∈ R and a compact set K ⊆ X̃5 such that, for every u in
a neighborhood of u0 , the level set
levα ϕ(·, u) = {x̃ ∈ (u) | ϕ(x̃, u) ≥ α}
is nonempty and contained in K.
(ii) For any neighborhood NX̃ of the set M(u0 ), there exists a neighborhood
NV of u0 such that Nx̄ ∩ (u) = ∅ for all u ∈ NV .
Then, the following consequences arise:
(a) The optimal value function v is continuous at u0 .

3 This
means that:
(a) S is upper semicontinuous at each x0 ∈ X; i.e., for any neighborhood N of S(x0 ), there exists a
neighborhood N  of x0 such that

x ∈ N  =⇒ S(x) ⊆ N.

(b) S is lower semicontinuons at each x0 ∈ X; i.e., for any y0 ∈ S(x0 ) and any neighborhood N of
y0 , there exists a neighborhood N  of x0 such that

x ∈ N  =⇒ S(x) ∩ N = ∅.
4 According to (A1), we have in fact that S(x) is compact.
5 Of course, α and K may depend on u0 .
372 JOTA: VOL. 131, NO. 3, DECEMBER 2006

(b) M is upper semicontinuous at u0 .

To prove the theorem, consider in Proposition 3.1 X̃ = R × Y, V = Gr(S).

X̃ is endowed with the product topology and V is considered as a topological
subspace of the product space X × Y . Also, for any x̃ = (t, y  ) ∈ X̃ and u =
(x, y) ∈ V , we put
ϕ(x̃, u) = ϕ(t, y  , x, y) := t,
(u) = (x, y) := {(t, y  ) ∈ R × Y | y  ∈ S(x), F (x, y  ) ≤C F (x, y) − te}.
Note that the optimal value function v from Proposition 3.1 coincides with the
restriction of h to Gr(S). Obviously, ϕ is continuous from X̃ × V to R and, for
any u = (x, y) ∈ V , we have (0, y) ∈ (u), hence dom () = V .
Consider a net ((xi , yi ), (ti , yi ))i ∈ Gr() which converges to an element
((x, y), (t, y  )) ∈ V × X̃. Thus, for each i,
yi ∈ S(xi ), F (xi , yi ) − ti e − F (xi , yi ) ∈ C.
Let VY be a closed neighborhood of the set S(x). Since S is upper semicon-
tinuous at x, there exists a neighborhood Vx of x such that, for any x  ∈ Vx , we
have S(x  ) ⊆ VY . Thus, there exists i0 such that, for all i ≥ i0 , we have xi ∈ Vx ,
hence S(xi ) ⊆ VY . It follows that yi ∈ VY for i ≥ i0 , hence y  ∈ VY (because
VY is closed). On the other hand, any closed set in a Hausdorff topological
space coincides with the intersection of all closed neighborhoods of it. Thus,
y  ∈ S(x). Moreover, the continuity of F and the fact that C is closed imply that
F (x, y) − te − F (x, y  ) ∈ C, hence ((x, y), (t, y  )) ∈ Gr(). This shows that 
is closed.
Let
u0 = (x0 , y0 ) ∈ Gr(S).
To verify (i) of Proposition 3.1, according to (A1), consider the neighborhoods
Nx0 , NYx0 , and the compact set KYx0 . Take α = 0, and the compact subset K =
[0, β] × KYx0 of X̃, where
 
β = 2 max{ F (x, y) : (x, y) ∈ Nx0 × NYx0 ∩ Gr(S)}/d(e, Z\C). (2)
Note that β is well defined, since the numerator exists and is finite accord-
ing to (A1), (A2) by the Weierstrass theorem and that the distance d(e, Z\C)
from e to the set Z\C is strictly positive because e ∈ int C. Let us prove
that the neighborhood Vu0 = (Nx0 × NYx0 ) ∩ V of u0 has the property required
in (i):
∀u = (x, y) ∈ Vu0 , lev0 ϕ(·, u) := {x̃ ∈ (u) | ϕ(x̃, u) ≥ 0} is nonempty and
contained in K. (3)
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 373

Let
u = (x, y) ∈ Vu0 .
Since
lev0 ϕ(·, u) = {(t, y  ) ∈ R × Y | y  ∈ S(x), t ≥ 0, F (x, y  ) ≤C F (x, y) − te},
and since (x, y) ∈ Gr(S), we have that (0, y) ∈ lev0 ϕ(·, u); hence, lev0 ϕ(·, u) is not
empty. Let (t, y  ) ∈ lev0ϕ (·, u). For any λ ∈ C + , λ ∗ = 1( · ∗ denotes the norm
of the dual space Z ∗ of the Banach space Z), we have λ, e ≥ d(e, Z\C) > 0
[see Ref. 21, Lemma 2.2]; hence,
t ≤ λ, F (x, y) − F (x, y  )/λ, e ≤ F (x, y) − F (x, y  ) /λ, e ≤ β. (4)
Since y  ∈ S(x) ⊆ KYx0 , we obtain that (t, y  ) ∈ K; hence, (3) holds.
To verify (ii), remark first that M(u0 ) is not empty because ϕ(·, u0 ) is contin-
uous and lev0 ϕ(·, u0 ) is compact.
Let NX̃ be a neighborhood of M(u0 ) and let (t0 , y0 ) ∈ M(u0 ). There exist
δ > 0 and a neighborhood Vy0 of y0 in Y such that
[t0 − δ, t0 + δ] × Vy0 ⊆ NX̄ . (5)
Since e ∈ int C, there exists r > 0 such that e − (1/r)BZ ⊂ C, where BZ = −BZ
is the open unit ball of Z. We can write BZ ⊂ re − C; i.e.,
∀h0 ∈ Z, h0 < 1 =⇒ h0 ≤C re. (6)
Let ε = δ/2r. The continuity of F [restricted to Gr(S)] at (x0 , y0 ) and (x0 , y0 ),
the
lower semicontinuity of S at x0 , and the fact that y0 ∈ S(x0 ) ∩ Vy0 = ∅ ensure the
existence of a neighborhood Vx0 of x0 and of a neighborhood Vy0 of y0 such that:
(a) For any x ∈ Vx0 , we have S(x) ∩ Vy0 = ∅.
(b) For any (x, y) ∈ (Vx0 × Vy0 ) ∩ Gr(S), we have F (x, y  ) − F (x0 , y0 ) <
ε.
(c) For any (x, y  ) ∈ (Vx0 × Vy0 ) ∩ Gr(S), we have F (x, y  ) − F (x0 , y0 ) <
ε.
Take
NV = (Vx0 × Vy0 ) ∩ Gr(S).
Let (x, y) ∈ NV and consider according to (a) an element ŷ0 ∈ S(x) ∩ Vy0 . From
(c), there exists h1 ∈ BZ such that
F (x, ŷ0 ) = F (x0 , y0 ) + εh1 .
From (b), there exists h2 ∈ BZ such that
F (x, y) = F (x0 , y0 ) + εh2 .
374 JOTA: VOL. 131, NO. 3, DECEMBER 2006

Since (t0 , y0 ) ∈ M(u0 ), we have

F (x0 , y0 ) ≤C F (x0 , y0 ) − t0 e.
It follows that
F (x, ŷ0 ) ≤C F (x, y) − t0 e + ε(h1 − h2 ).
On the other hand, h1 − h2 ∈ 2BZ ; hence, using (6), we obtain
ε(h1 − h2 ) ≤C 2εre = δe.
Thus,
F (x, ŷ0 ) ≤C F (x, y) − (t0 − δ)e.
We obtain that
(t0 − δ, ŷ0 ) ∈ (u) ∩ NX̃ .

Proposition 3.2. Under the hypotheses of Theorem 3.1, if in addition dom

(S) = X, then we have that dom (
W ) = X.

Proof. Let us consider x ∈ X. The set S(x) is compact, nonempty, and

the function y → F (x, y) is continuous on S(x). Hence, using Theorem 3.3 and
Lemma 3.5, pp. 46-47, from Ref. 25, we obtain that Problem (LL)(x) has an
efficient solution. It is obvious that every efficient solution is also a weakly-
efficient one. Thus, we obtain that
W (x) = ∅.

The following result, which is a consequence of Theorem 3.1, uses more

classical hypotheses and takes into account the upper-level constraints set U ⊆
X × Y.
Let pX : X × Y → X and pY : X × Y → Y denote the canonical projec-
tions defined by pX (x, y) = x, PY (x, y) = y for all (x, y) ∈ X × Y . Consider the
topological subspaces of X and Y respectively, defined by
X̂ = pX (Gr(S) ∩ U ) and Ŷ = pY (Gr(S) ∩ U )
→ Ŷ defined by Ŝ(x) =
respectively. Consider also the set-valued map Ŝ : X̂ →
S(x) ∩ Ŷ for all x ∈ X̂.

Theorem 3.2. Let Z be endowed with the norm (strong) topology. Let X
and Y be Hausdorff topological spaces. Assume the following:
(B1) The set Gr(S) ∩ U is compact.
(B2) The restriction of the map F to Gr(Ŝ) is continuous.
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 375

→ Ŷ is continuous and Gr(Ŝ) is closed.

(B3) The set-valued map Ŝ : X̂ →

Then, the restriction of the function h to Gr(S) ∩ U is finite and continuous

at every point.

Proof. It is obvious from (B1) that the topological spaces X̂ and Ŷ are
compact spaces. Now, we apply Theorem 3.1 replacing the spaces X and Y by X̂
and Ŷ respectively. Consequently, F is restricted to X̂ × Ŷ and S is replaced by
Ŝ : X̂ →→ Ŷ . Note that (A1) of Theorem 3.1 is satisfied taking Nx = Kx = X̂ and
NYx = KYx = Ŷ .
Therefore, the restriction of h to Gr(Ŝ) is continuous. On the other hand,
Gr(Ŝ) ⊇ Gr(Ŝ) ∩ U and Gr(Ŝ) ∩ U = Gr(S) ∩ U according to the definition of
X̂, Ŷ , Ŝ.

Next, for each positive real number r, we consider the following penalized
problem:
(Pr ) min (f (x, y) + rh(x, y)) , s.t. (x, y) ∈ U ∩ Gr(S),
x,y

which is a scalar minimization problem.

Note that a point (x, y) ∈ X × Y verifies the constraints of Problem (Pr ) iff
(a) (x, y) ∈ U ; i.e., (x, y) is feasible for the upper-level constraints;
(b) y ∈ S(x); i.e., y is feasible for the constraints of the lower-level vector
optimization problem associated to a given x ∈ X.
The interesting point is that the penalty term rh(x, y) forces y to be (close to)
a weakly-efficient point for the lower level vector problem associated to a given
x ∈ X.
Now, we can state the main result based on an exterior penalty method.

Theorem 3.3. Let X and Y be two reflexive Banach spaces. Assume that
X ⊆ X and Y ⊆ Y are weakly closed subsets. We consider X (resp. Y ) as a
Hausdorff topological space endowed with the trace of the weak topology of X
(resp. Y). We make all the assumptions of Theorem 3.1. Moreover, we suppose
that:
(A) The set Gr(S) ∩ U is closed in X × Y (i.e., with respect to the product
topology).
(B) The restriction of f to Gr(S) ∩ U is continuous (i.e., it is continuous
on the topological subspace Gr(S) ∩ U of the topological space X × Y
endowed with the product topology).
376 JOTA: VOL. 131, NO. 3, DECEMBER 2006

(CH) f (x, y) → +∞, as (x, y) ∈ Gr(S) ∩ U, x X + y Y → +∞.

Then:
(i) For every r > 0, Problem (Pr ) has at least a solution (xr , yr ).
(ii) The generalized sequence6 (xr , yr )r↑∞ is bounded [in the product
Banach space X × Y; i.e., the real generalized sequence ( xr X +
yr Y )r↑∞ is bounded]; hence, it admits at least a cluster point (i.e.,
a cluster point in X × Y with respect to the weak topology; obviously,
this point belongs to X × Y , and is a cluster point of the sequence
(xr , yr )r↑∞ with respect to the topology of X × Y ).
(iii) Any cluster point of the generalized sequence (xr , yr )r↑∞ solves Prob-
lem (BLw ).

Proof. We adapt the classical proof (see for example Ref. 26, pp. 366–
369) to our special problem in not-necessarily finite-dimensional spaces. Problem
(BLw ) is equivalent to the problem of minimizing f over the set
A := Gr(
W ) ∩ U,
which has been assumed to be nonempty in the Introduction. Let us denote
B = Gr(S) ∩ U.
Obviously, A ⊆ B; hence, B is nonempty. From Theorem 3.1, we have that the
restriction hB of h to B is continous. Moreover, according to Lemma 3.1, the
function h is positive over B. Also, we assumed that the restriction fB of f to B
is continuous. Thus, Problems (Pr ) and (BLW ) can be written equivalently as
min(fB (·, ·) + rhB (·, ·)), over B, (7)
respectively
min fB (·, ·), over A. (8)
Note that A = h−1 B ({0}); hence, A is closed in B; therefore, A is closed in X × Y
because B is closed in X × Y .
Let (x̂, ŷ) ∈ A and consider the sets
 = {(x, y) ∈ A | fB (x, y) ≤ fB (x̂, ŷ)},
B̂ = {(x, y) ∈ B | fB (x, y) ≤ fB (x̂, ŷ)}.

6 Recallthat a generalized sequence (zr )r↑∞ in a topological Hausdorff space Z is a function r → zr

from R+ to Z. The generalized sequence (zr )r↑∞ is convergent to z ∈ Z if, for every neighborhood
N of z, there exists r0 > 0 such that, for all r > r0 , we have zr ∈ N . A cluster point of this generalized
sequence is an element a ∈ Z such that, for any neighborhood N of a and for any r0 > 0, the set
[r0 , +∞[∩{r ∈ R+ | zr ∈ N } is not empty.
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 377

Obviously  and B̂ are nonempty (because (x̂, ŷ) ∈  ⊆ B̂) weakly closed sets
in X × Y (because they are closed in B) and bounded in X × Y according to
(CH). Thus, Â and B̂ are weakly compact sets in X × Y, hence compact sets in
the topological space X × Y . Let us denote
Lr = {(x, y) ∈ B | fB (x, y) + rhB (x, y) ≤ fB (x̂, ŷ)}.
Since hB ≥ 0, it follows that Lr ⊆ B̂. Hence, Lr is compact in X × Y and
nonempty [it contains (x̂, ŷ) because hB (x̂, ŷ) = 0]. Thus, for each r > 0, the
problem
min(fB (·, ·) + rhB (·, ·)), over Lr ,
admits an optimal solution (xr , yr ) which is obviously an optimal solution for
Problem (Pr ).
Also, there exists minimizers (x ∗ , y ∗ ) ∈ Â and (x̄, ȳ) ∈ B̂ of fB over the
compact sets  and B̂ respectively. It is obvious that (x ∗ , y ∗ ) solves Problem
(BLw ) and that
f∗ := f (x̄, ȳ) = min f (x, y).
(x,y)∈B

Since, for each r > 0, (xr , yr ) ∈ B̂, we obtain that the generalized sequence
(xr , yr )r↑∞ is bounded in X × Y.
For each r > 0, we have
f∗ ≤ f (xr , yr ) ≤ f (xr , yr ) + rh(xr , yr ) ≤ f (x ∗ , y ∗ );
hence,
0 ≤ h(xr , yr ) ≤ [f (x ∗ , y ∗ ) − f∗ ]/ r.
Thus, we obtain that h(xr , yr ) tends to 0 as r → +∞. Let (x̃, ỹ) be a weak cluster
point of the generalized sequence (xr , yr )r↑∞ in X × Y, hence (x̃, ỹ) ∈ X × Y .
Let (xrn , yrn )n∈N , with rn → +∞, be a sequence convergent to (x̃, ỹ) in X × Y
(of course, there exists such a sequence). By the continuity of hB , we obtain that
hB (x̃, ỹ) = 0, hence (x̃, ỹ) ∈ A. Also, for each (x, y) ∈ A, n ∈ N, we have
f (xrn , yrn ) ≤ f (xrn , yrn ) + rn h(xrn , yrn ) ≤ f (x, y);
hence, letting n → ∞, we obtain f (x̃, ỹ) ≤ f (x, y).

Next, under the more classical assumptions of Theorem 3.2, we obtain a

similar result but in topological spaces.

Theorem 3.4. Assume that the hypotheses of Theorem 3.2 hold and that
the restriction of f to Gr(S) ∩ U is continuous. Then:
(i) For every r > 0, Problem (Pr ) has at least a solution (xr , yr ).
378 JOTA: VOL. 131, NO. 3, DECEMBER 2006

(ii) Any cluster point (which exists) of the generalized sequence (xr , yr )r↑∞
solves Problem (BLw ).

Proof. Since now we assume that Gr(S) ∩ U is compact, one can adapt
easily and simplify the proof of Theorem 3.3. For the sake of clarity, we give the
proof.
Problem (BLW ) is equivalent to the problem of minimizing f over the set
A := Gr(
W ) ∩ U , which has been assumed to be nonempty in the Introduction.
Let us denote
B = Gr(S) ∩ U.
Obviously A ⊆ B, hence B is nonempty. From Theorem 3.2, we have that the
restriction hB of h to B is continuous. Moreover, according to Lemma 3.1, the
function h is positive over B. Also, we assumed that the restriction fB of f to B
is continuous. Thus, Problems (Pr ) and (BLW ) can be written equivalently as
min(fB (·, ·) + rhB (·, ·)), over B, (9)
respectively
min fB (·, ·), over A. (10)
Note that A = h−1
B ({0}),
hence A is closed in B, therefore A is compact.
For each r > 0, Problem (Pr ) admits an optimal solution (xr , yr ) ∈ B. Also,
there exists minimizers (x ∗ , y ∗ ) ∈ A and (x̄, ȳ) ∈ B of fB over the compact sets
A and B respectively. It is obvious that (x ∗ , y ∗ ) solves Problem (BLW ) and we
denote f∗ := f (x̄, ȳ).
For each r > 0, we have
f∗ ≤ f (xr , yr ) ≤ f (xr , yr ) + rh(xr , yr ) ≤ f (x ∗ , y ∗ );
hence,
0 ≤ h(xr , yr ) ≤ [f (x ∗ , y ∗ ) − f∗ ]/r.
Thus, we obtain that h(xr , yr ) tends to 0 as r → +∞. Let (x̃, ỹ) be a cluster point
of the generalized sequence (xr , yr )r↑∞ ; let (xrn , yrn )n∈N , with rn → +∞, be a
sequence convergent to (x̃, ỹ) (of course there exists such a sequence). By the
continuity of hB , we obtain that hB (x̃, ỹ) = 0, hence (x̃, ỹ) ∈ A. Also, for each
(x, y) ∈ A, n ∈ N, we have that
f (xrn , yrn ) ≤ f (xrn , yrn ) + rn h(xrn , yrn ) ≤ f (x, y);
hence, letting n → ∞, we obtain
f (x̃, ỹ) ≤ f (x, y).

 JOTA: VOL. 131, NO. 3, DECEMBER 2006 379

In the next remark, we show that, assuming X and Y are Euclidean spaces,
some hypothesis may be simplified.

Remark 3.2. In Theorem 3.3, let us assume that X and Y are Euclidean
spaces. In this case, the weak topology coincides with the strong one; thus, we have
to consider X and Y as closed subsets of X and Y. Also in this case, Assumption
(A1) of Theorem 3.1 may be simplified replacing it with the following:
(A1 ) for any 
x ∈ X, there exists a bounded neighborhood Nx of x such that
the set x  ∈Nx S(x  ) is bounded.

To conclude, for the sake of completeness, we present two examples.

Example 3.1. We give an example of a simple problem which satisfies the

assumptions of Theorem 3.1 but does not satisfy the assumptions of Theorem
3.2. So, let X = Rn and let Y be a compact Hausdorff topological space. We take
U = X × Y and let S : X → → Y be any continuous set-valued map with closed
and nonempty values. Assumption (A1) of Theorem 3.1 is satisfied for Nx =
Kx = the unit closed ball of center x and radius 1 and NYx = KYx = Y . On the
other hand, the set Gr(S) ∩ U is not compact, because its projection on X is equal
to X, which is not compact.

Example 3.2. Since Assumption (A2) of Theorem 3.1 is natural, as well as

all the other assumptions of Theorem 3.3, we give an example of a (polyhedral) set-
valued map S satisfying Assumptions (A1) and (A3) of Theorem 3.1, which seem
more unusual. Thus, in Theorem 3.3, consider X = Rn , Y = Rp , X = Rn+ , Y =
p
R+ and, for every x ∈ X,
S(x) = {y ∈ Y | Ay ≤ x},
where A is a linear map from Rp to Rn (identified with an n × p real matrix (aij )
and of course x and y are written as column vectors). The relation Ay ≤ x means
componentwise inequalities, i.e.,

p
aij yj ≤ xi , for all 1 ≤ i ≤ n,
j =1

in other words,
x − Ay ∈ X.
The only assumption made on the matrix A is that aij > 0 for all i, j .

Let us show that, for all x ∈ X, S(x) is a bounded, nonempty set. Denote
380 JOTA: VOL. 131, NO. 3, DECEMBER 2006

p
αi = min aij , 1 ≤ i ≤ n.
j =1

We have αi > 0 for all i. It is obvious that 0 ∈ S(x) and, for all y ∈ S(x), 1 ≤ i ≤
n, we have

p

p
αi y = αi yj ≤ aij yj ≤ xi ≤ x ; (11)
j =1 j =1

here, · stands for the l 1 norm in Rn or Rp ; i.e., x = ni=1 |xi |. Thus, S(x) ⊆
[0, x1 /α1 ]p ; hence, S(x) is bounded.
Now, we show that S verifies Assumptions (A1) and (A3) of Theorem 3.1.
It is obvious that S(x) is closed for all x ∈ X. Let us show that S is Lipschitz;
i.e., there exists a constant γ > 0 such that, for any x, x  ∈ X and for any y ∈ S(x),
there exists y  ∈ S(x  ) such that y − y  ≤ γ x − x  . Indeed, consider
 
γ = max α1−1 , . . . , αn−1
and let x, x  ∈ X and y ∈ S(x). From (11), we obtain that y ≤ γ x . Thus, if
x = 0, then y = 0 [i.e., S(0) = {0}] and, for any y  ∈ S(x  ), we have
y  − y = y  ≤ γ x  = γ x  − x .
So, we will assume further that x = 0.
If there exists i0 such that xi0 = 0, take y  = 0. We have
y − y  = y ≤ xi0 /αi0 = (xi0 − xi0 )/αi0
≤ (1/αi0 ) x − x  ≤ γ x − x  .
If, for all i, xi > 0, then take y  = ty, where t is defined by t = min{xi /xi |xi >
0} = xi0 /xi0 . We have y  ∈ Y, Ay  ≤ x  , hence y  ∈ S(x  ) and
y − y  = |1 − t| · y
= |xi0 − xi0 | y /xi0
≤ x − x  y /xi0
≤ x − x  αi−1
0

≤ γ x − x  .
The fact that S is Lipschitz implies the continuity of the set-valued map S :
X→→ Y , hence Assumption (A3) is fulfilled.
To prove (A1’) stated in Remark 3.2 [which in our example is equivalent to
(A1)], note that 0 ≤ x ≤ x  ⇒ S(x) ⊆ S(x  ). Denote 1n the column vector in Rn
having all entries equal to 1. Given x ∈ X, we can take
Nx ={x  ∈ X | x  ≤ x + 1n }
 JOTA: VOL. 131, NO. 3, DECEMBER 2006 381

={x  ∈ X | ∀i : xi ≤ xi + 1}
n
= [0, xi + 1].
i=1

Obviously, Nx is a bounded neighborhood of x and

S(x  ) = S(x + 1n ),
x  ∈Nx

which is bounded.

4. Concluding Remarks

In this paper, we have shown in a very general setting that the semivectorial
bilevel optimization problem (BLw ) can be approached using a penalty method.
Of course, it is interesting and useful to see how this method applies in particular
cases, for example when all the functions are linear and we have polyhedral
constraints. Another important case is when all the functions and constraints are
convex. However, the study of all these cases require special developments. This
will be the object of future works.

References

1. BENSON, H. P., Optimization over the Efficient Set, Journal of Mathematical Analysis
and Applications, Vol. 98, pp. 562–580, 1984.
2. BENSON, H. P., A Finite, Nonadjacent Extreme Point Search Algorithm for Optimization
over the Efficient Set, Journal of Optimization Theory and Applications, Vol. 73, pp. 47–
64, 1992.
3. CRAVEN, B. D., Aspects of Multicriteria Optimization, Recent Prospects in Mathemat-
ical Programming, Gordon and Breach, Philadelphia, Pennsylvania, 1991.
4. FÜLÖP, J., A Cutting Plane Algorithm for Linear Optimization over the Efficient
Set, Generalized Convexity, Lecture Notes in Economics and Mathematical Systems,
Springer Verlag, Berlin, Germany, Vol. 405, 1994.
5. PHILIP, J., Algorithms for the Vector Maximization Problem, Mathematical Program-
ming, Vol. 2, pp. 207–229, 1972.
6. AN, L. T. H., TAO, P. D., and MUU, L. D., Numerical Solution for Optimization over the
Efficient Set by D,C, Optimization Algorithms, Operations Research Letters, Vol. 19,
pp. 117–128, 1996.
7. BOLINTINÉANU, S., Optimality Conditions for Minimization over the (Weakly or Prop-
erly) Efficient Set, Journal of Mathematical Analysis and Applications, Vol. 173,
pp. 523–541, 1993.
8. BOLINTINÉANU, S., Minimization of a Quasiconcave Function over an Efficient Set,
Mathematical Programming, Vol. 61, pp. 89–110, 1993.
382 JOTA: VOL. 131, NO. 3, DECEMBER 2006

9. BOLINTINÉANU, S., Necessary Conditions for Nonlinear Suboptimization over the

Weakly-Efficient Set, Journal of Optimization Theory and Applications, Vol. 78, pp.
579–598, 1993.
10. BOLINTINÉANU, S., and EL MAGHRI, M., Pénalisation dans l’Optimisation sur
l’Ensemble Faiblement Efficient, RAIRO-Recherche Opérationnelle/Operations Re-
search, Vol. 31, pp. 295–310, 1997.
11. DAUER, J. P., Optimization over the Efficient Set Using an Active Constraint Approach,
Zeitschrift für Operations Research, Vol. 35, pp. 185–195, 1991.
12. DAUER, J. P., and FOSNAUGH, T. A., Optimization over the Efficient Set, Journal of
Global Optimization, Vol. 7, pp. 261–277, 1995.
13. HORST, R., and THOAI, N. V., Maximizing a Concave Function over the Efficient or
Weakly-Efficient Set, European Journal of Operations Research, Vol. 117, pp. 239–252,
1999.
14. YAMAMOTO, Y., Optimization over the Efficient Set: Overview, Journal of Global
Optimization, Vol. 22, pp. 285–317, 2002.
15. MORGAN, J., Constrained Well-Posed Two-Level Optimization Problems, Nonsmooth
Optimization and Related Topics, Edited by F. Clarke, V. Demyanov, and F. Giannessi,
Plenum Press, New York, NY, pp. 307–326, 1989.
16. LORIDAN, P., and MORGAN, J., A Theoretical Approximation Scheme for Stackelberg
Problems, Journal of Optimization Theory and Applications, Vol. 61, pp. 95–110, 1989.
17. LORIDAN, P., and MORGAN, J., Weak via Strong Stackelberg Problem: New Results,
Journal of Global Optimization, Vol. 8, pp. 263–287, 1996.
18. LIGNOLA, M. B., and MORGAN, J., Stability of Regularized Bilevel Programming Prob-
lems, Journal of Optimization Theory and Applications, Vol. 93, pp. 575–596, 1997.
19. DEMPE, S., Foundations of Bilevel Programming, Kluwer Academic Publishers, Dor-
drecht, Holland, 2002.
20. DEMPE, S., Annotated Bibliography on Bilevel Programming and Mathematical Pro-
grams with Equilibrium Constraints, Optimization, Vol. 52, pp. 333–359, 2003.
21. BOLINTINÉANU, S., Approximate Efficiency and Scalar Stationarity in Unbounded Non-
smooth Convex Vector Optimization Problems, Journal of Optimization Theory and
Applications, Vol. 106, pp. 265–296, 2000.
22. BOLINTINÉANU, S., Vector Variational Principles, ε–Efficiency, and Scalar Stationarity,
Journal of Convex Analysis, Vol. 8, pp. 71–85, 2001.
23. BONNANS, F., and SHAPIRO, A., Perturbation Analysis of Optimization Problems,
Springer, New York, NY, 2000.
24. BANK, B., GUDDAT, J., KLATTE, D., KUMMER, B., and TAMMER, K., Nonlinear Para-
metric Optimization, Academie Verlag, Berlin, Germany, 1982.
25. LUC, D. T., Theory of Vector Optimization, Lecture Notes in Economics and Mathe-
matical Systems, Springer Verlag, Berlin, Germany, 1989.
26. LUENBERGER, D. G., Linear and Nonlinear Programming, 2nd Edition, Addison-
Wesley Publishing Company, Reading, Massachusetts, 1984.