SIAM J. OPTIM.

c 2013 Society for Industrial and Applied Mathematics


Vol. 23, No. 1, pp. 153–169

STRONG DUALITY IN CONE CONSTRAINED NONCONVEX

Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

OPTIMIZATION∗
FABIÁN FLORES-BAZÁN† AND GIANDOMENICO MASTROENI‡

Abstract. In this paper we deepen the analysis of the conditions that ensure strong duality
for a cone constrained nonconvex optimization problem. We first establish a necessary and sufficient
condition for the validity of strong duality without convexity assumptions with a possibly empty
solution set of the original problem, and second, via Slater-type conditions involving quasi interior or
quasirelative interior notions, various results about strong duality are also obtained. Our conditions
can be used where no previous result is applicable, even in a finite dimensional or convex setting.

Key words. quasi-relative interior, strong duality, nonconvex optimization

AMS subject classifications. 90C26, 90C30, 49N15, 90C46

DOI. 10.1137/120861400

1. Introduction. The purpose of this paper is to develop the analysis of con-

ditions ensuring the existence of strong duality for a cone constrained nonconvex
optimization problem. Such conditions are based on the notion of quasi-relative inte-
rior [3] that recently has received great attention in the literature. Our results unify
and extend to the nonconvex case some analogous theorems that have been obtained
under suitable convexity assumptions on the functions involved [6, 7, 15].
Let X be a real locally convex topological vector space, Y be a Hausdorff locally
convex topological vector space (t.v.s.), P ⊆ Y be a nonempty closed convex cone
with possibly empty topological interior, and C be a nonempty subset of X. Given
f : C → R and g : C → Y , let us consider the cone constrained minimization problem
.
(1.1) μ= inf f (x).
g(x)∈−P
x∈C

The (Lagrangian) dual problem associated to (P ) is

.
(1.2) ν = sup inf [f (x) + λ∗ , g(x)],
λ∗ ∈P ∗ x∈C

where P ∗ is the positive polar cone of P . We say that problem (1.1) has a (La-
grangian) zero duality gap if the optimal values of (1.1) and (1.2) coincide, that is,
μ = ν. Problem (1.1) is said to have strong duality if it has a zero duality gap and
problem (1.2) admits a solution. Our task is to characterize this property without
convexity assumptions and under minimal hypotheses. It is known that some con-
straints qualification (CQ) is needed, which may be of Slater-type, or interior-point
condition, to get strong duality. In some other situation, the validity of strong duality
requires a so called closed-cone CQ. Such CQ often restricts some applications.
∗ Received by the editors January 5, 2012; accepted for publication (in revised form) November

13, 2012; published electronically January 29, 2013.

http://www.siam.org/journals/siopt/23-1/86140.html
† Departamento de Ingenierı́a Matemática and CI2 MA, Facultad de Ciencias Fı́sicas y

Matemáticas, Universidad de Concepción, Casilla 160-C, Concepción, Chile (fflores@ing-mat.udec.

cl). This author’s research was supported in part by CONICYT-Chile through FONDECYT 112-0980
and BASAL projects, CMM, Universidad de Chile.
‡ Department of Informatics, University of Pisa, Largo B. Pontecorvo, 3, 56127 Pisa, Italy

(gmastroeni@di.unipi.it).
153

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 154 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

More precisely, when X = C = Rn and P = [0, +∞[ with g being a nonidentically

zero quadratic function, the authors in [17] prove that (1.1) has strong duality for each
quadratic function f if and only if there exists x̄ ∈ Rn such that g(x̄) < 0, which is
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

the classical Slater condition. A similar result is proved in [21, Theorem 3.3] when
X = Rn , C = {x ∈ Rn : Hx = d}.
Similarly, when g is P -convex (g(tx1 + (1 − t)x2 ) ∈ tg(x1 ) + (1 − t)g(x2 ) − P for
every t ∈ ]0, 1[ and all x1 , x2 ∈ C, provided C is convex) and continuous, it is proved
in [4] that (1.1) has strong duality for each f ∈ X ∗ if and only if a certain CQ holds.
This CQ involves the epigraph of the support function of C and the epigraph of the
conjugate of the function x → λ∗ , g(x). This CQ is also equivalent to the fact that
(1.1) has strong duality for each continuous and convex function f [16].
Stable zero duality gaps in convex programming (g is continuous and P -convex,
and f is a lower semicontinuous proper convex function), which means that strong
duality holds for each linear perturbation of f , were characterized in terms of a similar
CQ as above; see [18, 20] for details.
Several sufficient conditions of the zero duality gap have been also established in
the literature; see [12, 1, 2, 31, 4, 6, 7].
Unlike some of the above results (those including closed-cone CQ [4, 16]), which
involve conditions on g and C that guarantee (1.1) has strong duality for every f in
a certain class of functions, our approach allows us to derive conditions on f , g, and
C, jointly, that ensure (1.1) has strong duality under no convexity assumption. Thus,
we provide results where none of those in [12, 4, 18, 16, 5, 6, 7, 19] is applicable.
At the same time, because of many applications, our purpose is also to consider
convex cones P possibly with empty topological interior. This happens for instance
if (1 < p < +∞ and Ω being an open bounded set in Rn )
.
P = Lp+ = {u ∈ Lp (Ω) : u ≥ 0 for a.e. x ∈ Ω},
or if P is of the form P = Q × {0} with int Q = ∅. The former case appears when
dealing with constrained best interporlation problems; see the nice work by Qi [27]
(see also [24]).
A good substitute for the topological interior is the quasi interior and even the
quasi-relative interior. Borwein and Lewis in [3] introduced the quasi-relative interior
of a convex set A ⊆ Y , although the concept of quasi interior was introduced earlier.
We use both notions, and since the sets considered are not necessarily convex, the
convex hull arises naturally.
The paper is structured as follows. Section 2 provides the basic definitions, nota-
tion, and preliminaries on quasi (relative) interior of convex sets. In section 3 we first
establish a characterization of strong duality without any additional assumption; then
we present two main theorems on the validity of strong duality under no convexity as-
sumptions, which extend and unify existing results in the literature. Furthermore, we
show instances where no previous result is applicable. Consequences and comparison
with other existing results are discussed in section 4.
2. Basic notation and preliminaries. Throughout the paper, Y is a real
Hausdorff locally convex topological vector space, its topological dual space is Y ∗ ,
and ·, · denotes the duality pairing between Y and Y ∗ .
A set P ⊆ Y is said to be a cone if tP ⊆ P ∀ t ≥ 0; given A ⊆ Y , cone(A) stands
for the smallest cone containing A, that is,

cone(A) = tA,
t≥0

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 155

whereas cone(A) denotes the smallest closed cone containing A: obviously cone(A) =
cone(A), where A denotes the closure of A. Additionally, we set
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

. 
cone+ (A) = tA.
t>0

Evidently, cone(A) = cone+ (A) ∪ {0} and therefore cone(A) = cone+ (A).
Some elementary properties of cones are collected in the next lemma, where co(A),
int A stand for the convex hull of A which is the smallest convex set containing A
.
and topological interior of A, respectively. We denote R+ = [0, +∞[.
Given a convex set A ⊆ Y and x ∈ A, NA (x) stands for the normal cone to A at
x, defined by NA (x) = {ξ ∈ Y ∗ : ξ, a − x ≤ 0 ∀ a ∈ A}. We say that x ∈ A is a (see,
for instance, [7])
(a) quasi-interior point of A, denoted by x ∈ qi A, if cone(A − x) = Y , or
equivalently, NA (x) = {0};
(b) quasi-relative interior of A, denoted by x ∈ qri A, if cone(A − x) is a linear
subspace of Y , or equivalently, NA (x) is a linear subspace of Y ∗ .
For any convex set A, we have that [24, 7] qi A ⊆ qri A and int A = ∅ implies
int A = qi A. Similarly, if qi A = ∅, then qi A = qri A. Moreover [3], if Y is a finite
dimensional space, then qi A = int A and qri A = ri A, where ri A means the relative
interior of A, which is the interior with respect to the affine hull of A, denoted by
aff A.
We recall the definition of pointedness for a cone that is not necessarily convex
(see, for instance, [29]).
Definition 2.1. A cone P ⊆ Y is called “pointed” if x1 + · · · + xk = 0 is
impossible for x1 , x2 , . . . , xk in P unless x1 = x2 = · · · = xk = 0.
It is easy to see that a cone P is pointed if and only if co(P ) ∩ (−co(P )) = {0} if
and only if 0 is a extremal point of co(P ).
The positive polar of the convex cone P ⊆ Y is defined by
.
P ∗ = {y ∗ ∈ Y ∗ : y ∗ , x ≥ 0 ∀x ∈ P }.

Lemma 2.2. Let ∅ = M ⊆ Y . The following relations hold:

(a) co(M ∪ {0}) ⊆ cone(co M ).
(b) cone(co(M ∪ {0})) = cone(co M ).
(c) If ∅ = N ⊆ Y is a convex set, then co(M + N ) = co(M ) + N.
(d) 0 ∈ qri[co(M ∪ {0})] ⇐⇒ 0 ∈ qri[cone(co M )].
(e) 0 ∈ qi[co(M ∪ {0})] ⇐⇒ 0 ∈ qi[cone(co M )].
Proof. (a) Let ȳ ∈ co(M ∪ {0}). Then
p

ȳ = αi mi for some αi ≥ 0, mi ∈ M, i = 1, . . . , p.
i=1
p
If αpi = 0, then ȳ =
i=1 0 ∈ cone M ⊆ cone(co M ).
If i=1 αi > 0, then
 p  p
  α
ȳ = αi p i mi ∈ cone(co M ).
i=1 i=1 i=1 αi

(b) The inclusion (⊇) is obvious; the other comes from (a).

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 156 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

q
(c) (⊆) Let mi ∈ M, pi ∈ N, i = 1, . . . , q, αi ≥ 0, i=1 αi = 1. Then
q
 q
 q

Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

αi (mi + pi ) = αi mi + αi pi ∈ co(M ) + N.
i=1 i=1 i=1
q
(⊇) Let mi ∈ M, i = 1, . . . , q, αi ≥ 0, i=1 αi = 1, p ∈ N .
Then
q
 q

αi mi + p = αi (mi + p) ∈ co(M + N ).
i=1 i=1

(d) It is a consequence of the following equalities obtained from (b):

cone[co(M ∪ {0})] = cone(co M ) = cone[cone(co M )].
Therefore, cone[co(M ∪ {0})] is a linear subspace of Y if and only if cone[cone(co M )]
is a linear subspace of Y , or, equivalently, 0 ∈ qri[cone(co M )].
(e) It is analogous to the proof of part (d).
The following separation theorem is a direct consequence of the equivalent char-
acterization of the quasi-relative interior and of Lemma 2.2. As mentioned in [3], the
quasi-relative interior of a set M consists of the points of x ∈ M for which it is not
possible to find a supporting hyperplane to M at x.
Theorem 2.3. Let ∅ = M ⊆ Y . Then, 0 ∈ qri[cone(co M )] (or, equivalently,
0 ∈ qri[co(M ∪ {0})]) if and only if there exists x∗ ∈ Y ∗ \ {0} such that x∗ , x ≤ 0
∀ x ∈ M with strict inequality for some x̄ ∈ M .
Proof. The necessity part is as follows. Since 0 ∈ co(M ∪ {0})), 0 ∈ qri[co(M ∪
{0})] if and only if Nco(M∪{0}) (0) is not a linear subspace of Y ∗ , i.e., there exists
x∗ ∈ Y ∗ \ {0} such that
x∗ , x ≤ 0 ∀x ∈ co(M ∪ {0}),
and, furthermore, there is x̂ ∈ co(M ∪ {0}) such that
(2.1) x∗ , x̂ < 0.
Since x̂ ∈ co(M ∪ {0}), for some integer p ≥ 1,
p

x̂ = αi xi , αi ≥ 0, xi ∈ M, i = 1, . . . , p.
i=1

It follows that there exists at least one i ∈ 1, . . . , p, such that x∗ , xi  < 0; otherwise
(2.1) would be contradicted, which proves the necessity part.
The sufficiency is straightforward.
The next result [3] is a useful characterization of the quasi-relative interior.
Theorem 2.4 (see [3, Theorem 3.10]). Let Y be as above partially ordered by a
convex cone P with P − P = Y . Then
y ∈ qri P ⇐⇒ y ∈ P and y ∗ , y > 0, ∀y ∗ ∈ P ∗ \ {0}.
Note that in case P is a closed convex cone, the condition y ∈ P can be omitted.
Proposition 2.5. Let A ⊆ Y be a convex set. Then,
(a) cone(A − A) = cone A − cone A provided 0 ∈ A;
(b) [0 ∈ qi A] ⇐⇒ [0 ∈ qi(A − A) and 0 ∈ qri A].

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 157

Proof. (a) It is straightforward.

(b) From (a) it follows that
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(2.2) cone(A − A) = cone(A) − cone(A) = cone(A) − cone(A).

This along with the equivalence 0 ∈ qi A ⇐⇒ 0 ∈ A and cone(A) = Y allow us to get

cone(A − A) = Y , and therefore

[0 ∈ qi A] =⇒ [0 ∈ qi(A − A) and 0 ∈ qri A].

The converse implication follows from (2.2) as well. Indeed, if 0 ∈ qri A, then cone(A)
is a linear subspace of Y , and therefore from (2.2) it follows that cone(A − A) =
cone(A). If additionally 0 ∈ qi(A − A), then Y = cone(A − A) = cone(A). Hence
0 ∈ qi A, and the proof is completed.
Notice that (b) can also be found in [15].
Proposition 2.6. Let P ⊆ Y be a convex cone such that P − P = Y . Then

qri P = qi P.

Proof. We only need to prove that qri P ⊆ qi P . Since y ∈ qri P if and

only if 0 ∈ qri(P − y) and y ∈ P , by virtue of the previous proposition, we need
to check that 0 ∈ qi(P − y − (P − y)) = qi(P − P ), which is true by assumption:
cone(P − P ) = P − P = Y .
3. Lagrangian strong duality: Main results and regularity conditions.
Let Y be as in the preceding section and X be a Hausdorff topological vector space,
f : X → R, g : X → Y , C ⊆ X, and P be a closed and convex cone in Y . We consider
the problem
.
(3.1) μ = inf f (x),
x∈K
.
where K = {x ∈ C : g(x) ∈ −P }. We assume that μ is finite and that the feasible
region K of (3.1) is nonempty.
Notice that the requirement of f taking finite values is not restrictive since no
additional structure like convexity or closedness on C is imposed. Thus for functions
.
f : X → R ∪ {+∞} the set C may be the effective domain dom f = {x ∈ X : f (x) <
+∞}.
.
Let F = (f, g) and consider the set
.
Eμ = F (C) − μ(1, 0) + (R+ × P ).

This set or its conic hull arises in a natural way when dealing with duality results
or in deriving alternative theorems; see [22, 12, 8, 15, 10]. Giannessi [13] used it in
a systematic manner for a constrained extremum problem giving rise to the image
space analysis.
Proposition 3.1. The following assertions hold.
(a) Assume that μ ∈ f (K). Then, μ = inf x∈K f (x) if and only if

(3.2) Eμ ∩ H = ∅,
.
where H = {(u, v) ∈ R × Y : u < 0, v ∈ −P }. Furthermore,

Eμ ∩ H = ∅ ⇐⇒ cone(Eμ ) ∩ H = ∅.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 158 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

(b) inf x∈K f (x) = −∞ if and only if

Eρ ∩ H = ∅ ∀ ρ ∈ R,
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(3.3)

where Eρ is Eμ with μ replaced by ρ.

Proof. Since (b) is obvious, we only prove (a). We preliminarly observe that (3.2)
is equivalent to the fact that the system

(3.4) f (x) − μ + t < 0, g(x) + p ∈ −P, (x, t, p) ∈ C × R+ × P

is impossible.
Let μ = inf x∈K f (x) and assume that (3.2) does not hold. Then, there exists a
solution (x̃, t̃, p̃) ∈ (C × R+ × P ) of system (3.4), i.e.,

f (x̃) < μ − t̃ ≤ μ, g(x̃) ∈ −p̃ − P ∈ −P,

which contradicts the definition of μ. Conversely, assume that (3.4) is impossible;

then, in particular, setting t = 0 and p = 0, we have that

f (x) ≥ μ ∀x ∈ K

and, since μ ∈ f (K), (3.1) holds.

For the equivalence, one implication is obvious, whereas the other follows from
(0, 0) ∈ H.
Strong duality for (3.1) requires the existence of a linear continuous functional
that separates the sets Eμ and H. Actually, we need more than that as the next
theorem asserts. Let us introduce the Lagrangian

L(γ ∗ , λ∗ , x) = γ ∗ f (x) + λ∗ , g(x)

associated with (3.1).

Theorem 3.2. The following assertions are equivalent:
(a) Strong duality holds for (3.1), i.e., there exists λ∗0 ∈ P ∗ such that

(3.5) inf f (x) = inf L(1, λ∗0 , x).

x∈K x∈C

(b) cone(co Eμ ) ∩ (−R++ × {0}) = ∅.

Proof. Assume that strong duality holds; then,

(3.6) f (x) − μ + λ∗0 , g(x) ≥ 0 ∀ x ∈ C,

which implies

f (x) + t − μ + λ∗0 , g(x) + p ≥ 0 ∀ x ∈ C, ∀ t ≥ 0, ∀ p ∈ P,

i.e.,

u + λ∗0 , v ≥ 0 ∀ (u, v) ∈ Eμ .

It follows that

u + λ∗0 , v ≥ 0 ∀ (u, v) ∈ cone(co Eμ ).

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 159

Moreover, observe that

u + λ∗0 , v < 0 ∀ (u, v) ∈ −R++ × {0},

Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

so that (b) follows.

Vice versa, assume that (b) holds. Let A := (−1, 0) + (] − ρ, ρ[×V (0)), where
V (0) is an open convex neighborhood of 0Y and ρ > 0. Then, cone+ (A) is an open
convex set and by (b) it follows that

cone(co Eμ ) ∩ cone+ (A) = ∅

for a suitable choice of ρ > 0 and V (0). By the separation theorem for convex sets in
a t.v.s., there exist (γ0∗ , λ∗0 ) ∈ (R × Y ∗ ), (γ0∗ , λ∗0 ) = (0, 0), such that

(3.7) γ0∗ u + λ∗0 , v ≥ 0 ∀(u, v) ∈ cone Eμ ,

(3.8) γ0∗ u + λ∗0 , v ≤ 0 ∀(u, v) ∈ cone A.

Let us prove that γ0∗ = 0. By contradiction, suppose that γ0∗ = 0; then from (3.8)
it follows that λ∗0 , v ≤ 0 ∀v ∈ V (0), which implies λ∗0 = 0, thus contradicting
(γ0∗ , λ∗0 ) = (0, 0). Therefore, γ0∗ = 0, and with no loss of generality, we can assume
γ0∗ = 1, since by (3.8) (at the point (−1, 0) ∈ A), we have −γ0∗ ≤ 0. Then, (3.7)
implies

(3.9) u + λ∗0 , v ≥ 0 ∀(u, v) ∈ Eμ ,

and, in turn,

f (x) − μ + λ∗0 , g(x) ≥ 0 ∀x ∈ C,

so that

inf L(1, λ∗0 , x) ≥ μ.

x∈C

Let us prove that λ∗0 ∈ P ∗ . For fixed x0 ∈ C, we obtain from (3.7)

γ0∗ (f (x0 ) − μ) + λ∗0 , g(x0 ) + λ∗0 , p ≥ 0 ∀p ∈ P,

which implies λ∗0 ∈ P ∗ . Then

inf L(1, λ∗0 , x) ≤ μ,

x∈C

which completes the proof.

We now establish other theorems of a different nature about strong duality, which
involve generalized Slater conditions and quasi (relative) interior. To that purpose,
we first prove a necessary and sufficient condition for the existence of Fritz John
multipliers, and afterwards, under Slater-type conditions, the desired strong duality
result is established. Example 4.3 below shows the difference between both results.
The next theorem yields the existence of an hyperplane which may separate not
in a proper sense (0, 0) from cone(co Eμ ). Instead, Theorem 3.4 below provides a
hyperplane which separates them properly.
Theorem 3.3. Let us consider problem (3.1) and assume that μ is finite. The
following assertions are equivalent:

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 160 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

(a) There exist (γ0∗ , λ∗0 ) ∈ R+ × P ∗ , (γ0∗ , λ∗0 ) = (0, 0), such that

γ0∗ inf f (x) = inf L(γ0∗ , λ∗0 , x).

Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

x∈K x∈C

(b) (0, 0) ∈ qi[cone(co Eμ )].

(c) (0, 0) ∈ qi[co(Eμ ∪ {(0, 0)})].
In case qi(co Eμ ) = ∅, any of the previous conditions proves the pointedness of
cone[qi(co Eμ )]. Consequently, if int(co Eμ ) = ∅, then (a) is equivalent to the pointed-
ness of cone[int(co Eμ )].
Proof. By Lemma 2.2(e) it follows that (b) and (c) are equivalent. Assume that
(b) is fulfilled. Since (0, 0) ∈ cone(co Eμ ), (b) holds if and only if Ncone(coEμ) (0, 0)
= {(0, 0)}, i.e., there exists (0, 0) = (−γ0∗ , −λ∗0 ) ∈ Ncone(co Eμ )(0, 0) such that

(γ0∗ , λ∗0 ), (u, v) ≥ 0 ∀ (u, v) ∈ cone(co Eμ ),

or, equivalently,

(3.10) (γ0∗ , λ∗0 ), (u, v) ≥ 0 ∀ (u, v) ∈ Eμ .

Note that (3.10) is equivalent to

γ0∗ (f (x) + t − μ) + λ∗0 , g(x) + p ≥ 0 ∀ t ∈ R+ , ∀ x ∈ C, ∀ p ∈ P.

Since (γ0∗ , λ∗0 ) ∈ (R+ × P )∗ = R+ × P ∗ , the previous inequality yields

γ0∗ f (x) + λ∗0 , g(x) + p ≥ γ0∗ μ ∀ x ∈ C, ∀ p ∈ P.

Hence

γ0∗ inf f (x) ≤ inf L(γ0∗ , λ∗0 , x).

x∈K x∈C

The reverse inequality is obvious, so that we obtain (a).

Vice versa, if (a) holds, then the previous relations show that (3.10) is fulfilled
for a suitable (0, 0) = (γ0∗ , λ∗0 ) ∈ R+ × P ∗ , so that we obtain (b).
For the last part we proceed as follows. Let x, − x ∈ cone[qi(co Eμ )], x = 0.
Thus, x, − x ∈ cone+ [qi(co Eμ )]. Then 0 = x + (−x) ∈ cone+ [qi(co Eμ )]. Therefore
(0, 0) ∈ qi(co Eμ ), and so Y = cone(co Eμ ) = cone[cone(co Eμ )]. This contradicts
(b), proving the desired implication. For the reverse implication, simply notice that
pointedness of cone[int(co Eμ )] implies that (0, 0) ∈ int(co Eμ ). Then, use a standard
separation theorem to derive (a).
A similar theorem using quasi-relative interior is obtained next.
Theorem 3.4. Let us consider problem (3.1) and assume that μ is finite. The
following assertions are equivalent:
(a) There exist (γ0∗ , λ∗0 ) ∈ R+ × P ∗ , (γ0∗ , λ∗0 ) = (0, 0), x̃ ∈ C, t̃ ≥ 0, and p̃ ∈ P
such that

γ0∗ inf f (x) = inf L(γ0∗ , λ∗0 , x) and γ0∗ (f (x̃) + t̃) + λ∗0 , g(x̃) + p̃ > μγ0∗ .
x∈K x∈C

(b) (0, 0) ∈ qri[cone(co Eμ )].

(c) (0, 0) ∈ qri[co(Eμ ∪ {(0, 0)})].
In case qri(co Eμ ) = ∅, any of the previous conditions implies the pointedness of
cone[qri(co Eμ )].

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 161

Proof. The equivalences are consequences of Theorem 2.3. The remaining part
follows a similar reasoning as in the preceding theorem.
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Looking at Theorems 3.3 and 3.4, we realize that strong duality is obtained under
the nonverticality of the linear functional (γ0∗ , λ∗0 ), that is, we need γ0∗ > 0. It holds
whenever a Slater-type condition is imposed as the following two theorems show.
Theorem 3.5. Assume that μ is finite and cone(co(g(C)) + P ) = Y , i.e., 0 ∈
qi(co(g(C)+P )). Then, any of the assumptions (b) or (c) of Theorem 3.3 is equivalent
to (3.5) for some λ∗0 ∈ P ∗ . In such a situation,

(3.11) inf f (x) = inf f (x).

x∈K λ∗ ,g(x)≤0
0
x∈C

Hence, if x̄ is a solution to (3.1), then λ∗0 , g(x̄) = 0.

Proof. As regards (3.5), from Theorem 3.3, we have only to prove that γ0∗ > 0.
If, on the contrary, γ0∗ = 0, then

0 = inf L(0, λ∗0 , x) ≤ λ∗0 , g(x) ∀ x ∈ C.

x∈C

It implies that λ∗0 , v ≥ 0 ∀ v ∈ g(C) + P , which yields

λ∗0 , v ≥ 0 ∀ v ∈ cone(co(g(C)) + P ).

Therefore, by assumption, we obtain λ∗0 = 0, which cannot happen as (γ0∗ , λ∗0 ) = (0, 0).
Hence γ0∗ > 0, and the conclusion follows.
For the equality in (3.11), we observe that the inequality ≥ is obvious. The reverse
inequality is a consequence of (3.5):

inf f (x) ≥ inf L(1, λ∗0 , x) ≥ inf L(1, λ∗0 , x) = inf f (x).
λ∗
0 ,g(x)≤0 λ∗
0 ,g(x)≤0 x∈C x∈K
x∈C x∈C

Now, we consider a Slater-type condition involving the quasi-relative interior of

the set co(g(C) + P ).
Theorem 3.6. Assume that μ is finite and 0 ∈ qri(co(g(C) + P )). Then, any of
the assumptions (b) or (c) of Theorem 3.4 is equivalent to the following:

(3.12) inf f (x) = inf L(1, λ∗0 , x)

x∈K x∈C

for some λ∗0 ∈ P ∗ . In such a case, if x̄ is a solution to (3.1), then λ∗0 , g(x̄) = 0.
Proof. By Theorem 3.4, we have only to prove that γ0∗ > 0, taking into account
that, in such a case, the second assertion in Theorem 3.4(a)

γ0∗ (f (x̃) + t̃) + λ∗0 , g(x̃) + p̃ > γ0∗ μ

for some (γ0∗ , λ∗0 ) ∈ R+ × P ∗ , x̃ ∈ C, t̃ > 0, p̃ ∈ P , is automatically satisfied if the

feasible set K is nonempty.
Lemma 2.2(d) proves that (b) and (c) of Theorem 3.4 are equivalent. By The-
orem 2.3, where we have set M = Eμ , (b) or (c) holds if and only if there exists
(0, 0) = (γ0∗ , λ∗0 ) ∈ R × Y ∗ such that

(γ0∗ , λ∗0 ), (u, v) ≤ 0 ∀ (u, v) ∈ Eμ

with strict inequality for some (ū, v̄) ∈ Eμ .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 162 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

Suppose on the contrary that γ0∗ = 0; then λ∗0 = 0 and

λ∗0 , v ≤ 0 ∀ v ∈ co(g(C) + P ),
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(3.13)

i.e., λ∗0 ∈ Nco(g(C)+P ) (0), recalling that 0 ∈ co(g(C) + P ) because the feasible set is
nonempty. Since 0 ∈ qri(co(g(C) + P )) is equivalent to saying that Nco(g(C)+P ) (0) is
a linear subspace, −λ∗0 ∈ Nco(g(C)+P ) (0), and it follows that

λ∗0 , v = 0 ∀ v ∈ co(g(C) + P ),

which contradicts that strict inequality holds in (3.13) for v = v̄.

The next two theorems provide certain regularity conditions on cone(co Eμ )
under the Slater-type assumptions. It is really an important and interesting fact
since, for instance, in finite dimension, such a Slater assumption guarantees that
int[cone(co Eμ )] = ∅ under the hypoyheses of Theorem 3.5.
Proposition 3.7.
(a) If 0 ∈ qi[co(g(C)+P )], then qi[co(Eμ ∪{(0, 0)})] = ∅ and so qi[cone(co Eμ )] = ∅.
(b) If 0 ∈ qri[co(g(C)+P )], then qri[co(Eμ ∪{(0, 0)})] = ∅ and qri[cone(co Eμ )] = ∅.
Proof. (a) Since we assume that the feasible region of (3.1) is nonempty, there
exists x̃ ∈ C such that 0 ∈ g(x̃) + P . (Therefore, our assumption is equivalent to
saying that 0 ∈ qi[co(g(C) + P )].) Let g(x̃) = −p̃, where p̃ ∈ P . We will prove that

(3.14) (f (x̃) + t − μ, g(x̃) + p̃) ∈ qi(co(Eμ ∪ {(0, 0)})) ∀ t > 0.

Let t > 0 be fixed and consider

(u∗ , v ∗ ) ∈ Nco(Eμ ∪{(0,0)}) ((f (x̃) + t − μ, g(x̃) + p̃)).

Therefore u∗ ∈ R, v ∗ ∈ Y ∗ , and

(3.15) u∗ (u − ((f (x̃) + t − μ)) + v ∗ , v − (g(x̃) + p̃) ≤ 0 ∀(u, v) ∈ co(Eμ ∪ {(0, 0)}).
. .
Setting u = f (x̃) + 2t − μ, v = g(x̃) + p̃, we obtain −u∗ 2t ≤ 0, and setting u =
3 ∗t ∗
f (x̃) + 2 t − μ, v = g(x̃) + p̃, we obtain u 2 ≤ 0. Since t > 0, it must be u = 0.
Therefore, (3.15) becomes

(3.16) v ∗ , v − (g(x̃) + p̃) ≤ 0 ∀v ∈ co(g(C) + P ).

Since g(x̃) + p̃ = 0, (3.16) implies that

v ∗ , v − (g(x̃) + p̃) ≤ 0 ∀v ∈ cone[co(g(C) + P )] = Y,

so that it must be v ∗ = 0.
Hence,

(3.17) Nco(Eμ ∪{(0,0)}) ((f (x̃) + t − μ, g(x̃) + p̃)) = {(0, 0)},

which proves (3.14).

In order to complete the proof, we simply observe that co(Eμ ∪ {(0, 0)}) ⊆
cone(co Eμ ) implies that qi[co(Eμ ∪ {(0, 0)})] ⊆ qi[cone(co Eμ )] = ∅; see Lemma 2.2(a).
(b) With the same x̃ ∈ C and p̃ ∈ P as in (a), we will prove that

(3.18) (f (x̃) + t − μ, g(x̃) + p̃) ∈ qri(co(Eμ ∪ {(0, 0)})) ∀ t > 0.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 163

Indeed, similar to (a), we can check that

(3.19) v ∗ , v − (g(x̃) + p̃) ≤ 0 ∀ v ∈ co(g(C) + P ),
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

which says v ∗ ∈ Nco(g(C))+P (g(x̃) + p̃). By assumption, 0 = g(x̃) + p̃ ∈ qri(co(g(C)) +

P ), so that by the equivalent characterization of the quasi-relative interior, Nco(g(C)+P )
(g(x̃) + p̃) is a linear subspace of Y ∗ . Then we have also −v ∗ ∈ Nco(g(C)+P ) (g(x̃) + p̃),
i.e.,
(3.20) −v ∗ , v − (g(x̃) + p̃) ≤ 0 ∀v ∈ co(g(C)) + P.
From (3.19) and (3.20), it follows that
(3.21) (0, ±v ∗ ) ∈ Nco(Eμ ∪{(0,0)}) ((f (x̃) + t − μ, g(x̃) + p̃)).
We have already proved that if (u∗ , v ∗ ) ∈ Nco(Eμ ) ((f (x̃)+t−μ, g(x̃)+ p̃)), then u∗ = 0.
Thus, (3.21) implies that Nco(Eμ ∪{(0,0)}) ((f (x̃) + t − μ, g(x̃) + p̃)) is a linear subspace
and (3.18) holds.
The second part of the proof follows a similar reasoning as above. More precisely,
we observe that
(u∗ , v ∗ ) ∈ Ncone(co Eμ ) ((f (x̃) + t − μ, g(x̃) + p̃)) for some t > 0
implies that u∗ = 0, so that v ∗ ∈ Ncone(co(g(C)+P )) (g(x̃) + p̃). The assumption 0 ∈
qri(co(g(C) + P )) is equivalent to the fact that Ncone(co(g(C)+P )) (g(x̃) + p̃) is a linear
subspace (see Remark 3.9) which implies that Ncone(co Eμ ) ((f (x̃) + t − μ, g(x̃) + p̃)) is
a linear subspace, i.e., (f (x̃) + t − μ, g(x̃) + p̃) ∈ qri[cone(co Eμ )], which completes the
proof.
Remark 3.8. Denoted by aff(A) the affine hull of A, one of the referees proposes
the following alternate proof of the second part of (b). It follows observing that if
A ⊆ B and aff(A) = aff(B), then one has qri(A) ⊆ qri(B). We apply it to the sets
A = co(Eμ ∪ {(0, 0)}), B = cone(co Eμ ). However, our proof is self-contained.
Remark 3.9. By Lemma 2.2(d) and since K = ∅, the hypothesis 0 ∈ qri(co(g(C)+
P )) is equivalent to 0 ∈ qri(cone(co(g(C) + P ))); likewise, by Lemma 2.2(e), the
hypothesis 0 ∈ qi(co(g(C) + P )) is equivalent to 0 ∈ qi(cone(co(g(C) + P ))).
We end this section by noting that our results are closely related with saddle point
conditions for the Lagrangian L(1, λ∗ , x) in case the infimum of (3.1) is attained. The
saddle point characterization of strong duality clarifies the importance of such a prop-
erty for nonconvex optimization problems both from the theoretical and algorithmic
point of view.
Theorem 3.10. Let Y be locally convex and assume that 0 ∈ qi(co(g(C) + P )).
Then, μ is attained at x̄ ∈ K and any of the assumptions (b) or (c) of Theorem 3.3
holds if and only if there exists λ∗0 such that (λ∗0 , x̄) ∈ P ∗ × C is a saddle point for
L(1, λ∗ , x) on P ∗ × C, i.e.,
(3.22) L(1, λ∗ , x̄) ≤ L(1, λ∗0 , x̄) ≤ L(1, λ∗0 , x) ∀(λ∗ , x) ∈ P ∗ × C.
Proof. It is enough to recall that (λ∗0 , x̄) is a saddle point for L(1, λ∗ , x) on (P ∗ ×C)
if and only if x̄ ∈ K, f (x̄) = inf x∈C L(1, λ∗0 , x) and λ∗0 , g(x̄) = 0, so that the thesis
follows from Theorem 3.5.
Similarly, from Theorem 3.6 we obtain the following result.
Theorem 3.11. Let Y be locally convex and assume that 0 ∈ qri(co(g(C) + P )).
Then, μ is attained at x̄ ∈ K and any of the assumptions (b) or (c) of Theorem 3.4
holds if and only if there exists λ∗0 such that (λ∗0 , x̄) ∈ P ∗ × C is a saddle point for
L(1, λ∗ , x) on P ∗ × C.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 164 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

4. Some consequences and comparison with other existing results. We

observe that the convex hull appearing in the results of the previous section can
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

be deleted everywhere simply by requiring the convexity of the sets cone Eμ and
cone(g(C) + P ), since in this situation,

cone Eμ = cone(co Eμ ), cone(co(g(C)) + P ) = cone(g(C) + P ).

An important class of vector functions implying the convexity of the sets Eμ and
g(C) + P which satisfy more verifiable conditions is that introduced in [22]: given a
convex set C ⊆ X with X as above, a real locally convex topological vector space Z
along with a convex cone Q ⊆ Z, a mapping G : C → Z is called ∗-quasi-convex if
q ∗ , G(·) is quasi-convex ∀ q ∗ ∈ Q∗ . Independently, the author in [30] says that G
is naturally Q-quasi-convex if ∀ x, y ∈ C, G([x, y]) ⊆ [G(x), G(y)] − Q. Both classes
coincide as shown in [8, Proposition 3.9] when int Q = ∅ and [11, Theorem 2.3] for
general Q. See also [23].
It is known from Corollary 3.11 in [8] that every ∗-quasi-convex function G : C →
Z satisfying

∀q ∗ ∈ Q∗ , the restriction of q ∗ , G(·) on any line segment of C

(4.1)
is lower semicontinuous,

is such that G(C) + P is convex, so that G(C  ) + P is also convex for every convex
set C  ⊆ C.
Therefore, by setting F = (f, g) and assuming the convexity of C, the lower
semicontinuity on any line segment of C of q ∗ , F (·) ∀ q ∗ ∈ R+ × P ∗ and the ∗-
quasi-convexity of F : C → R × Y , we get the convexity of F (C) + (R+ × P ) (and
so Eμ is convex as well) and the quasiconvexity of the functions f and p∗ , g(·) on C
∀ p∗ ∈ P ∗ . Hence, f (C) + R+ and g(C) + P are convex sets as well.
Obviously there are vector functions F such that F (C) + (R+ × P ) is convex
without being ∗-quasi-convex. The convexity of F (C) + (R+ × P ) was imposed in
[7, 15]. Hence, our Theorem 3.6 is more general, even in the convex case, than
Theorem 4.4 in [7] and Theorem 10 in [15], since the last two theorems require the
stronger condition 0 ∈ qi(g(C) + P ). This is shown by Example 4.3 below. To be
more precise, Theorem 4.4 in [7] reads as follows.
Theorem 4.1 (see [7, Theorem 4.4]). Suppose that F (C) + (R+ × P ) is convex,
0 ∈ qi(g(C) + P ), and (0, 0) ∈ qri[co(Eμ ∪ {(0, 0)})]. Then, there exists λ∗0 ∈ P ∗ such
that (3.12) holds.
In order to prove the previous theorem, the authors show first that “Fenchel
and Lagrange duality” (so, some convexity assumptions are imposed) are equivalent,
generalizing an earlier result due to Magnanti [25]. Then, from such an equivalence
the strong duality is obtained.
On the other hand, from Proposition 2.5(b), it follows that
(4.2)
0 ∈ qi(co(g(C))+P ) ⇐⇒ 0 ∈ qi[co(g(C))+P −(co(g(C))+P )] and 0 ∈ qri(co(g(C))+P ).

This implies that Theorems 4.2 and 4.4 in [7] are identical provided g(C)+P is convex.
Furthermore, we point out that Theorems 3.5 and 3.6 apply to more general
situations, even to non-quasi-convex functions with equality and inequality constraints
and possibly where argminK f is empty, as the next example shows. Example 4.3
below will show that even in the convex case, our Theorem 3.6 is applicable but
Theorem 4.4 in [7] or Theorem 10 in [15] are not.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 165

Example 4.2. Notice this example shows our approach applies even if int P = ∅.
2
Take C = R2 , P = {0} × R+ , x = (x1 , x2 ), f (x) = x21 + 2e−x2 , and
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

2
g1 (x) = x41 − e−x2 , g2 (x) = x21 − x22 ,

and consider the problem

.
μ = inf{f (x) : g1 (x) = 0, g2 (x) ≤ 0, x ∈ C}.

Thus, P ∗ = R × R+ and μ = 0, although the set of minimizers is empty. Setting

2 2
F (x) = (f (x), g1 (x), g2 (x)), x ∈ C, we obtain F (x) = (x21 + 2e−x2 , x41 − e−x2 , x21 − x22 ).
It follows that
2 2
E0 = {(u, v1 , v2 ) ∈ R3 : u ≥ x21 + 2e−x2 , v1 = x41 − e−x2 , v2 ≥ x21 − x22 , (x1 , x2 ) ∈ R2 },
2
which is nonconvex (see below). Then, because of the condition u ≥ x21 + 2e−x2
∀(x1 , x2 ) ∈ R2 , we have co E0 ⊆ {(u, v1 , v2 ) ∈ R3 : u > 0} and therefore

(0, 0, 0) ∈ int(cone (co(F (C)) − μ(1, 0, 0) + R+ × P )) .

Observe that E0 ∩ H = ∅, where H = {(u, v1 , v2 ) ∈ R3 : u < 0, v1 = 0, v2 ≤ 0}.

Moreover,
2
(g1 , g2 )(C) + P = {(v1 , v2 ) ∈ R2 : v1 = x41 − e−x2 , v2 ≥ x21 − x22 , (x1 , x2 ) ∈ R2 }
√
is nonconvex. In fact, taking (x1 , x2 ) = (0, 0), (0, 2 2), we have that (v1 , v2 ) =
−8
(−1, 0), (−e−8 , −8) belongs to (g1 , g2 )(C)+P but the convex combination ( −1−e 2 , −4)
∈ (g1 , g2 )(C) + P . To prove this, observe that the system

−1−e−8 2
2 = x41 − e−x2 , (x1 , x2 ) ∈ R2 ,
−4 ≥ x21 − x22

is not possible. Otherwise, for a suitable (x1 , x2 ) ∈ R2 it should be

⎧
⎨e−x22 = x4 + 1+e−8 ≥ 1 ,
1 2 2
⎩x2 ≥ x2 + 4 ≥ 4,
2 1

which is clearly impossible. This also proves that E0 is nonconvex.

It is easy to see that

(0, 0) ∈ int(co((g1 , g2 )(C) + P )),

that is, the generalized

√ Slater condition is satisfied. (Actually, taking (x1 , x2 ) =
(0, 0), (1, 0), (1, 2), we have that (v1 , v2 ) = (−1, 0), (0, 1), (1 − e−2 , −1) belongs to
(g1 , g2 )(C) + P , so that the previous relation follows.) On the other hand, given
λ = (λ1 , λ2 ) ∈ R × R+ , we obtain
2 2
L(1, λ, x) = x21 + 2e−x2 + λ1 (x41 − e−x2 ) + λ2 (x21 − x22 )
2
= λ1 x41 + (1 + λ2 )x21 + (2 − λ1 )e−x2 − λ2 x22 .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 166 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

Hence, for λ1 ∈ R and λ2 ≥ 0, we get

⎧
⎨ 0 λ2 = 0, 0 ≤ λ1 ≤ 2,
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

if
inf 2 L(1, λ, x) = 2 − λ1 if λ2 = 0, λ1 > 2,
x∈R ⎩
−∞ if λ2 = 0, λ1 < 0 or λ2 > 0,

and therefore,

max inf L(1, λ, x) = 0 = μ,

(λ1 ,λ2 )∈P ∗ x∈R2

inf L(1, λ∗ , x) = 0, λ∗ = (λ∗1 , 0), 0 ≤ λ∗1 ≤ 2.

x∈R2

The following example shows that even in the convex case, our Theorem 3.6 is
applicable, but Theorem 4.4 in [7] or Theorem 10 in [15] are not.
Example 4.3. This example shows an application to a convex problem with
int P = ∅.
Take C = {(x1 , x2 , x3 ) ∈ R3 : x3 = 0}, P = {0} × {0} × R+ , x = (x1 , x2 , x3 ),
f (x) = x21 ,

g1 (x) = x1 + x2 + x3 , g2 (x) = x1 + x2 − x3 , g3 (x) = x21 + x22 − 1,

and consider the problem

.
μ = inf{f (x) : g1 (x) = 0, g2 (x) = 0, g3 (x) ≤ 0, x ∈ C}.

Thus, μ = 0 and x̄ = (0, 0, 0) is the optimal solution. Setting F (x) = (x21 , x1 + x2 +

x3 , x1 + x2 − x3 , x21 + x22 − 1), it follows that

E0 = {(u, v1 , v2 , v3 ) ∈ R4 : u ≥ x21 , v1 = x1 + x2 + x3 ,
v2 = x1 + x2 − x3 , v3 ≥ x21 + x22 − 1, x ∈ R3 }

is convex, and it is not difficult to check that cone(E0 ) ∩ H = ∅, which implies that
cone(E0 ) ∩ (−R++ × {0}) = ∅, where

H = {(u, v1 , v2 , v3 ) ∈ R3 : u < 0, v1 = 0, v2 = 0, v3 ≤ 0}.

The former equality allows us to apply Theorem 3.2. It is easy to see that

(g1 , g2 , g3 )(C) + P = {(v1 , v2 , v3 ) ∈ R3 : v1 = v2 = x1 + x2 ,

v3 ≥ x21 + x22 − 1, (x1 , x2 ) ∈ R2 }

is a convex set with empty interior so that qi((g1 , g2 , g3 )(C) + P ) = ∅. However, since

(0, 0, 0) ∈ qri((g1 , g2 , g3 )(C) + P ),

Theorem 3.6 can be applied.

The next example shows a problem where the Slater condition does not hold while
(b) of Theorem 3.2 is fulfilled, and therefore strong duality holds.
Example 4.4. Take f (x) = x2 , C = R, g(x) = x2 + x4 , P = R+ and consider the
problem
.
μ = min f (x).
g(x)≤0

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 167

It is easy to see that μ = 0 and x̄ = 0 is the optimal solution. On the other hand,
F (C) = {(x2 , x2 + x4 ) : x ∈ R} = {(u, v) ∈ R × R : v = u + u2 , u ∈ R+ }. Then
F (C) + (R+ × P ) = R2+ is a closed convex cone and (b) of Theorem 3.2 is fulfilled.
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

However, g(C) + P = R+ , and therefore 0 ∈ qi(g(C) + P ) = qri(g(C) + P ).

A simple consequence of Theorem 3.5 is the following.
Corollary 4.5. Assume that μ is finite and 0 ∈ co(g(C)) + qi P . Then, any of
the assumptions (b) or (c) of Theorem 3.3 is equivalent to the existence of λ∗0 ∈ P ∗
such that (3.5) is fulfilled.
Proof. The result follows from Theorem 3.5 once we observe that co(g(C))+qi P ⊆
qi(co(g(C))+P ). Such an inclusion easily follows since any x = q+p with q ∈ co(g(C))
and p ∈ qi P satisfies cone(P − (x − q)) = Y , which yields
Y = cone(P + q − x) ⊆ cone(P + co(g(C)) − x) ⊆ Y,
proving the desired result.
The next corollary is a generalization to the nonconvex case of Theorem 4.1 of
[6].
Corollary 4.6. Let P be a convex cone in Y such that P − P = Y . Assume
that μ is finite and 0 ∈ co(g(C)) + qri P . Then, any of the assumptions (b) or (c) of
Theorem 3.3 is equivalent to the existence of λ∗0 ∈ P ∗ such that (3.5) is fulfilled.
Proof. Since P − P = Y , by Proposition 2.6 we obtain that qri P = qi P so that
our assumptions imply that 0 ∈ co(g(C))+ qi P and from Corollary 4.5 the conclusion
follows.
When the topological interior is employed instead of the quasi-relative interior,
we obtain the following theorem, which is a consequence of Theorems 3.3 and 3.5,
already appeared in [9], and applies to situations when int P may be empty. In what
follows, R++ = ]0, +∞[.
Theorem 4.7. Let us consider problem (3.1) and assume that μ is finite,
int[co(F (C)) + (R+ × P )] = ∅, (or int[co Eμ ] = ∅),
and cone(co(g(C) + P )) = Y . The following assertions are equivalent:
(a) there exists λ∗0 ∈ P ∗ such that (3.12) holds.
(b) cone[int[co(F (C)) − μ(1, 0) + (R+ × P )]] is pointed.
(b ) cone[co(F (C)) − μ(1, 0) + (R++ × int P )] is pointed, provided int P = ∅.
Remark 4.8 (The case of finite dimensional spaces). Theorem 3.4, when special-
ized to finite dimension, reduces to Theorem 3.2 in [14], whereas the finite dimensional
version of Theorem 3.6 strengthens Theorem 3.6 in [14].
Remark 4.9 (Connection with the S-lemma). We now provide a connection with
the well-known S-lemma. This expresses the following: given f, g : Rn → R and
C ⊆ Rn , the question is under which conditions the implication
g(x) ≤ 0, x ∈ C =⇒ f (x) ≥ 0
is satisfied, or, equivalently, when the system g(x) ≤ 0, x ∈ C, f (x) < 0 has no
solution.
The important case when f and g are quadratic with C = Rn was studied by
Yakubovich; see the survey by Pólik and Terlaky in [26]. Its proof uses the Dines
theorem which asserts the convexity of the set {(f (x), g(x)) ∈ R2 : x ∈ Rn } when f
and g are homogeneous quadratic functions. The S-lemma due to Yakubovich says
the following: assume f and g as above and that there is x̄ ∈ Rn such that g(x̄) < 0.
Then, (a) and (b) are equivalent:

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 168 FABIÁN FLORES-BAZÁN AND GIANDOMENICO MASTROENI

(a) There is no x ∈ Rn such that

f (x) < 0, g(x) ≤ 0.

Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(b) There is λ ≥ 0 such that

f (x) + λg(x) ≥ 0 ∀ x ∈ Rn .

Let us sketch a proof. Obviously (b) =⇒ (a) always holds. Assume therefore that
.
(a) holds. This means that g(x) ≤ 0 implies f (x) ≥ 0, that is, 0 ≤ μ = inf g(x)≤0 f (x).
By Proposition 3.1 we have that cone(Eμ ) ∩ H = ∅, where (set F = (f, g))
. .
Eμ = F (Rn ) − μ(1, 0) + R2+ and H = {(u, v) ∈ R2 : u < 0, v ≤ 0}.

By the Dines theorem [26, Proposition 2.3], F (Rn ) is convex, and therefore Eμ is
convex. It follows that

ri(cone Eμ ) ∩ ri H = ∅ ⇐⇒ ri(cone Eμ ) ∩ ri(H) = ∅,

or, equivalently (recalling that for any nonempty convex sets C1 , C2 ⊆ Rn , ri(C1 +
C2 ) = riC1 + riC2 ; see [28, Corollary 6.6.2]),

(0, 0) ∈ ri[cone Eμ − H] = ri[cone(Eμ − H)] = ri(cone Eμ ).

Therefore, (0, 0) ∈ ri(cone Eμ ). Moreover, we observe that the set g(Rn ) + R+ is

convex since g takes values in R. We can apply Theorem 3.6 to obtain the existence
of λ ≥ 0 such that f (x) + λg(x) ≥ μ ≥ 0 ∀x ∈ Rn .
We end this remark by pointing out that the Dines theorem was extended to the
case when Rn is substituted by a cone K such that K ∪ (−K) is a subspace of Rn in
[19, Theorem 3.2].
Acknowledgments. The authors want to express their gratitude to three ref-
erees for their constructive criticisms and helpful remarks, which have improved the
paper. Part of this research was carried out during a stay of the first author in the
Department of Mathematics, University of Pisa, and a stay of the second author in
the Department of Mathematical Engineering, University of Concepcion during April
and May 2012, supported in part by FONDECYT 112-0980, Chile. Both authors
want to express their gratitude for the kind hospitality of both institutions.

REFERENCES

[1] A. Auslender, Existence of optimal solutions and duality results under weak conditions, Math.
Program. A, 88 (2000), pp. 45–59.
[2] A. Auslender and M. Teboulle, Asymptotic Cones and Functions in Optimization and
Variational Inequalities, Springer Monogr. Math., Springer-Verlag, New York, 2003.
[3] J. M. Borwein and A. S. Lewis, Partially finite convex programming, part I: Quasi relative
interiors and duality theory, Math. Program. A, 57 (1992), pp. 15–48.
[4] R. I. Bot and G. Wanka, An alternative formulation for a new closed cone constraint quali-
fication, Nonlinear Anal., 64 (2006), pp. 1367–1381.
[5] R. I. Bot, S.-M. Grad, and G. Wanka, New regularity conditions for strong and total Fenchel-
Lagrange duality in infinite dimensional spaces, Nonlinear Anal., 69 (2008), pp. 323–336.
[6] R. I. Bot, E. R. Csetnek, and A. Moldovan, Revisiting some duality theorems via the
quasirelative interior in convex optimization, J. Optim. Theory Appl., 139 (2008), pp. 67–
84.
[7] R. I. Bot, E. R. Csetnek, and G. Wanka, Regularity conditions via quasi-relative interior
in convex programming, SIAM J. Optim., 19 (2008), pp. 217–233.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 STRONG DUALITY IN CONE CONSTRAINED OPTIMIZATION 169

[8] F. Flores-Bazán, N. Hadjisavvas, and C. Vera, An optimal alternative theorem and appli-
cations to mathematical programming, J. Global Optim., 37 (2007), pp. 229–243.
[9] F. Flores-Bazán, F. Flores-Bazán, and C. Vera, A complete characterization of strong
Downloaded 11/11/19 to 142.157.252.78. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

duality in nonconvex optimization with a single constraint, J. Global Optim., 53 (2012),

pp. 185–201.
[10] F. Flores-Bazán, F. Flores-Bazán, and C. Vera, Gordan-type alternative theorems and
vector optimization revisited, in Recent Developments in Vector Optimization, Q. H. Ansari
and J. C. Yao, eds., Springer, Berlin, 2012, pp. 29–59.
[11] F. Flores-Bazán and C. Vera, Unifying efficiency and weak efficiency in generalized qua-
siconvex vector minimization on the real-line, Int. J. Optim. Theory Methods Appl., 1
(2009), pp. 247–265.
[12] J. B. G. Frenk and G. Kassay, On classes of generalized convex functions. Gordan-Farkas
type theorems, and Lagrangian duality, J. Optim. Theory Appl., 102 (1999), pp. 315–343.
[13] F. Giannessi, Constrained Optimization and Image Space Analysis, Springer, Berlin, 2005.
[14] F. Giannessi and G. Mastroeni, Separation of sets and Wolfe duality, J. Global Optim., 42
(2008), pp. 401–412.
[15] A. Grad, Quasi-relative interior-type constraint qualifications ensuring strong Lagrange dual-
ity for optimization problems with cone and affine constraints, J. Math. Anal. Appl., 361
(2010), pp. 86–95.
[16] V. Jeyakumar, Constraint qualifications characterizing Lagrangian duality in convex opti-
mization, J. Optim. Theory Appl., 136 (2008), pp. 31–41.
[17] V. Jeyakumar, N. Q. Huy, and G. Y. Li, Necessary and sufficient conditions for S-lemma
and nonconvex quadratic optimization, Optim. Eng., 10 (2009), pp. 491–503.
[18] V. Jeyakumar and G. M. Lee, Complete characterizations of stable Farkas lemma and cone-
convex programming duality, Math. Program A, 114 (2008), pp. 335–347.
[19] V. Jeyakumar, G. M. Lee, and G. Y. Li, Alternative theorems for quadratic inequality systems
and global quadratic optimization, SIAM J. Optim., 20 (2009), pp. 983–1001.
[20] V. Jeyakumar, and G. Y. Li, Stable zero duality gaps in convex programming: Complete dual
characterizations with applications to semidefinite programs, J. Math. Anal. Appl., 360
(2009), pp. 156–167.
[21] V. Jeyakumar and G. Y. Li, Regularized duality for linearly constrained quadratic optimiza-
tion and trust-region problems, J. Global Optim., 49 (2011), pp. 1–14.
[22] V. Jeyakumar, W. Oettli, and M. Natividad, A solvability theorem for a class of qua-
siconvex mappings with applications to optimization, J. Math. Anal. Appl., 179 (1993),
pp. 537–546.
[23] D. Kuroiwa, Convexity for set-valued maps, Appl. Math. Lett., 9 (1996), pp. 97–101.
[24] M. A. Limber and R. K. Goodrich, Quasi interiors, Lagrange multipliers, and Lp spectral
estimation with lattice bounds, J. Optim. Theory Appl., 78 (1993), pp. 143–161.
[25] T. L. Magnanti, Fenchel and Lagrange duality are equivalent, Math. Program A, 7 (1974),
pp. 253–258.
[26] I. Pólik and T. Terlaky, A survey of the S-lemma, SIAM Rev., 49 (2007), pp. 371–418.
[27] H.-D. Qi, Some theoretical aspects on Newton’s method for constrained best interpolation, in
Continuous Optimization: Current Trends and Modern Applications, A. Rubinov and V.
Jeyakumar, eds., Springer, Berlin, 2005, pp. 23–49.
[28] R. T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ, 1970.
[29] R. T. Rockafellar and R. J.-B. Wets, Variational Analysis, Springer-Verlag, Berlin, 1998.
[30] T. Tanaka, General quasiconvexities, cones saddle points and minimax theorem for vector-
valued functions, J. Optim. Theory Appl., 81 (1994), pp. 355–377.
[31] C. Zǎlinescu, Convex Analysis in General Vector Spaces, World Scientific, Singapore, 2002.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.