Nonlinear Analysis, Theory, Methods & Applications, Vol. 30, No. 3, pp.

1487-1496, 1997
Proc. 2nd World Congress of Nonlinear Analysts
Pergamon © 1997Elsevier ScienceLtd
Printed in Great Britain. All rights reserved
0362-546X/97 $17.00 + 0.00
PII: S0362-546X(97)00213-7

ON CONE CONVEXITY OF SET-VALUED MAPS 1

DAISHI KUROIWA t, TAMAKI TANAKAt,2 and T R U O N G XUAN DUC HA§

? Department of M a t h e m a t i c s and Computer Science, Interdisplinary Faculty of Science and Engineering,
Shimane University, Matsue 690, Japan;
t Department of Information Science, Faculty of Science, Hirosaki University, Hirosaki 036, Japan; and
§ Hanoi Institute of Mathematics, P.O.Box 631, Boho, Hanoi, Vietnam

K e y words and phrases: Set-valued analysis, convexity of set-valued maps, cone-convexity, cone-
convexlikeness, quasi cone-convexity, properly quasi cone-convexity, and naturally quasi cone-
convexity.

1. I N T R O D U C T I O N

How is the concept of convexity of set-valued map defined? This paper is concerned with
such convexity, and some kinds of relationship between two sets with respect to a convex
cone. If y is a vector-valued function from a vector space into another (ordered) vector space,
we consider several notions based on vector-ordering for two vectors. It is natural and quite
popular ill vector optimization theory as well as applied mathematics; e.g., [9, 12, 13, 14].
On the other hand, the case of set-valued map is not so simple, because we should consider
relationship between two image sets. Up to now, some generalizations for convexity of vector-
valued function into set-valued version are proposed to extend optimal conditions in the area
of optimization theory; [2, 3, 4, 5, 8, 10, 11, 15, 16]. Such generalizations are natural and useful
for optimization problems, but there is no detail report about unified theory for convexity of
set-valued map except few fiterature; [6, 7]. Therefore, the aim of this paper is to give a unified
report on such convexity, that is, we define five kinds of cone convexity for set-valued maps
as generalizations of some convexities for vector-valued maps, and we investigate relationship
among such cone convexities.
The organization of the paper is as follows. In Section 2, we consider eight kinds of re-
lationships (in fact, six different relationships) between two sets in an ordered vector space
with respect to a convex cone, which are regarded as modifications of the order relation on
an ordered space. For the convenience of notation, we propose useful symbols to define such
set-relations. Based on each of the six relationships, we define, in Section 3, five categories
of cone convexity for set-valued maps as generalizations of some convexities for vector-valued
maps; convexity, convexlikeness, quaziconvexity, properly quasiconvexity and naturally quasi-
convexity for set-valued maps. We discuss them ill the corresponding subsections. It is simple
to define convexities, convexlikenesses and quasiconvexities of set-valued maps, however, the
1This paper is based on the first author's doctor thesis.
2This work was based on research 08740129 supported by Grant-in-Aid for Scientific Research from the
Ministry of Education, Science and Culture of Japan. This research has been partially supported by T h e
Telecommunications Advancement Foundation.
1487
1488 Second World Congress of Nonlinear Analysts

concepts of the others, t h a t is, properly quasiconvexities and naturally quasiconvexities for
set-valued m a p s are more complicated. Because convexity, convexlikeness, and quasiconvexity
for vector-valued maps are represented by conditions between two vectors, however, prop-
erly quasiconvexity and naturally quasiconvexity are defined by conditions between a vector
and a subset. Moreover, we investigate some relations among those kinds of cone convexity
for set-valued maps. Especially, we show that, among some kinds of cone convexity for set-
valued maps, there are similar relations to those among the corresponding cone convexities
for vector-valued maps.

2. R E L A T I O N S H I P BETWEEN TWO SETS WITH RESPECT TO CONES

Throughout the paper, let Z be an ordered topological vector space (ordered t.v.s., for short)
with the vector ordering _~c induced by a convex cone C: for x, y E Z,

x -<c Y if y - x E C. (2.1)

The convex cone C is assumed not to be pointed but to be solid, that is, its topological interior
i n t C is nonempty; hence, C O := (int C) U {0} is a pointed convex cone and induces another
antisymmetric vector ordering -<¢0 weaker than -<c in Z. Also, F is said to be a set-valued
m a p from X into Z if F is a m a p from X into 2 z, which is the power set of Z, and also we
write F : X ~ Z. Moreover, for a set-valued m a p F : X ~ Z we use the following symbols:

Graph(F):={(x,y) [ x E X , y E F ( x ) } ; DomF:={xEX [F(x)~}. (2.2)

In this paper, we consider several generalizations of convexity of vector-valued function

into t h a t of set-valued map. With respect to convexity of function there are two ways of gener-
alization. One is a generalization based on values of set-valued m a p F, that is, a prescription
of relationship between two sets f (Axl + (1 - A)x2) and AF (xl) + (1 - A)F (x2); the other is
a generalization based on equivalent characteristic sets of set-valued m a p F, that is, prescrip-
tions by epigraph of F, image set of F, and lower level set of F. This paper's approach is the
former, because the latter generalization is included in the former as mentioned in Section 3.
Now, we start with discussion on set-relationship, that is, we introduce eight kinds of
relationships between two sets in an ordered vector space with respect to a convex cone. This
classification is based on two ideas for set-relation.
First, with respect to relationship between two vectors a, b E Z, one of the followings
holds:
(i) a E b + C (equivalently b E a - C); (iii) b E a + C (equivalently a E b - C);

(ii) a • b q- C (equivalently b ¢ a - C); (iv) b ¢ a -b C (equivalently a ¢ b - C).

These relationships are summarized as b~ca, b ~c a or a~cb, a ~c b, t h a t is, one vector

is dominated by the other vector or otherwise. In the case of relationship between a n o n e m p t y
set A C Z and a vector b E Z, a different situation is observed; we have two domination
structure

(i) for all a E A, a<_cb;

(ii) there exists a E A such t h a t a<_cb.
 Second World Congress of Nonlinear Analysts 1489

The first relation means the vector b dominates the whole set A from above with respect to
the vector ordering <c. The second relation means the vector b is dominated from below by
an element of the set A. If the set A is singleton, they are coincident with each other. These
relationships are denoted by b C A ~ C and b E AYC, respectively, where
Aff~C:= A a + C and A~C:= U a+C. (2.3)
aEA aEA

Analogously, we use the following notations for a nonempty set B C Z:

BAC:= A b-C=B~(-C) and BUC:= U b-C=By(-C). (2.4)
bEB bEB

It is easy to see that A ~ C C At~C and BLTC C BUC, and also that A ~ B = A + B and
A Y B = A - B.
Secondly, we consider the relationship between two nonempty sets in Z, which is strongly
concerned with intersection and inclusion in set theory. Given nonempty sets A , B C Z,
exactly one of following conditions holds: (i) A A B = 0; (ii) A M B # O. The latter case
includes its special cases A C B and A D B.
By using above two ideas, we classify the relationship between two nonempty sets A, B E Z
in the sense that A is (partially) dominated from above by B or A (partially) dominates B
from below:
(i) A C BAC; (v) Ag~C D B;

(ii) A M (BAG) # O; (vi) (A~C) M B # 0;

(iii) A y e D B; (vii) A C BUC;

(iv) (At~C) M B # 0; (viii) A M ( B y e ) # 0.

Since conditions (i) and (v) coincide and conditions (iv) and (viii) coincide, we define six
kinds of classification for set-relationship; see Figure 1.
DEFINITION 2.1. For nonempty sets A , B E Z and a convex cone C in Z, we denote

• AACDBbyA<g) B; • (A~C) MB#ObyA_<gV)B;

• AM(BAC)#ObyA<(~) B; • AcBUCbyA_<~)B;

• A Y C D B by A _<(Cfii)B; • (AyC) M B # O b y A _ < ~ i ) B .

As shown in Figure 1, all implications among the set-relations are easily verified.
PROPOSITION 2.1. For nonempty sets A , B E Z and a convex cone C in Z, the following
statements hold:
• A G g ) B implies A <(~) B; • A <_g) B implies A _<(~v) B;

• A _<(~) B implies A <(~fii) B; • A _<(~v) B implies A < ~ ) B;

• A _<(~fii) B implies A _<~i) B; • A < ~ ) B implies A _<~i) B.

1490 SecondWorldCongressof NonlinearAnalysts

3. C A T E G O R I Z E D CONVEXITY FOR SET-VALUED MAPS

In this paper, (cone) convexity of function is generalized in the following two ways: One is
based on prescriptions of relationship between two sets F (Axl + (1 - A)x2) and h E (Xl) + (1 -
A)F (x2); the other is based on prescriptions by epigraph of F, image set of F, and lower level
set of F. Epigraph convexity, Image-set convexity, and lower level-set convexity are concerned
with convexity, convexlikeness, and quasiconvexity of set-valued map, respectively.
Using the six kinds of relationships between two nonempty sets introduced in Section 2,
we consider some different concepts with respect to six different set-relations <(C
a)- (k = i, . . . ,
vi) for each convexity of set-valued map as generalizations of those of vector-valued function.
We categorize such generalized convexities into five class, that is, convexity, convexlikeness,
quasiconvexity, properly quasiconvexity, naturally quasiconvexity; and this section consists of
four subsections related to them.

3.1. CONVEXITY AND CONVEXLIKENESS OF SET-VALUED MAP

A vector-valued function f : X -* Z is said to be C-convex ([13, 14]) if for every xl, x2 E X

and A E (0, 1),
f (AXl + (1 - A)x2) _<c Af(xl) + (1 - A)f(x2), (3.1)
which is equivalent to the following condition:

G r a p h ( f ) + {Ox} × C is a convex set. (3.2)

Whenever Z --- R and C -- R+, C-convexity above is the same as the ordinary convexity of a
real-valued function. Based on the six different set-relations _<(~)--(k = i, . . . , vi), we propose
the following generalization of convexity (3.1) to set-valued map.
DEFINITION 3.2. For each k = i, . . . , vi, a set-valued map F : X -,z Z is said to be t y p e (k)
c o n v e x if for every xl, x2 C D o m F and A E (0, 1),

f (Ax 1 -4- (1 - )t)x2) _<(~) AF(Xl) A- (1 - A)F(x2). (3.3)

By Proposition 2.1, we have some imphcations among convexities above:

PROPOSITION 3.2. For a set-valued map F : X ~.~ Z, the following relationships hold:
type (i) convex ~ type (ii) convex --~ type (iii) convex

type (iv) convex , type (v) convex --~ type (vi) convex

The set G r a p h ( F ) + ({Ox} × C) is said to be the epigraph of set-valued map F, and then we
have the following result on the epigraph convexity.

PROPOSITION 3.3. A set-valued map F : X ~ Z is type (iii) convex if and only if its epigraph
is convex.

REMARK 3.1. In [6], four notions of convexity of set-valued map are defined, which are in-
cluded in Definition def:convex-1.
 Second World Congress of Nonlinear Analysts 1491

Next, we proceed to the convexlikeness of set-valued map. A vector-valued function f :

X ---, Z is said to be C-convexlike ([13, 14]) if for every Xl,X2 E X and ,k E (0, 1), there exists
x E X such t h a t
f ( x ) _<c A f ( x l ) + (1 -- A)f(x2), (3.4)
which is equivalent to the following condition:

f ( X ) + C is a convex set. (3.5)

Based on the six different set-relations <(~) (k = i, . . . , vi), we propose the following gener-
alization of convexlikeness (3.4) to set-valued map.

DEFINITION 3.3. For each k = i, . . . , vi, a set-valued map F : X -~ Z is said to be t y p e (k)

c o n v e x l i k e if for every Xl,X2 E D o m F and A E (0, 1), there exists x C D o m F such t h a t

F ( x ) <(~) ~ F ( x l ) + (1 - ~)F(x~). (3.6)

By Proposition 2.1, we have some implications among convexlikeness above:

PROPOSITION 3.4. For a set-valued map F : X -~ Z, the following relationships hold:

type (i) convexlike ~ type (ii) convexlike * type (iii) convexlike

type (iv) conve Uke type (v) ¢onve ike , type (vi) conve ike

T h e set F ( D o m F ) + C is said to be the image set of set-valued map F , and then we have the
following result on the image-set convexity.

PROPOSITION 3.5. A set-valued map F : X ---+ Z is type (iii) convexlike if and only if its
image set is convex.

PROPOSITION 3.6. For a set-valued map F : X -,~ Z and each k = i, . . . , vi, type (k)
convexity implies type (k) convex_likeness.

3.2. QUASI CONVEXITY OF SET-VALUED MAP

A vector-valued function f : X --~ Z is said to be quasi C-convex ([13, 14]) if it satisfies one
of the following two equivalent conditions:
• ( L u c ' s q u a s i C - c o n v e x i t y ) for every Xl,X2 • X and A • (0, 1),

/ (Axl + (1 - A)x2) <c z, for all z • C ( f ( x l ) , f(x2)), (3.7)

where C ( f ( x l ) , f(x2)) -- {z • Z [ f(Xx) <c z, f(x2) <c z };

• ( F e r r o ' s q u a s i C - c o n v e x i t y ) for each z • Z, the set

/ - 1 = {x • x I • z - c} (3.8)

is convex or empty.
1492 SecondWorldCongressof NonlinearAnalysts

Based on the six different set-relations _<(~) (k = i, . . . , vi), we propose two ways of general-
ization of quasi C-convexities (3.7) and (3.8) to set-valued map.
First, to define Luc's type quasiconvexity of set-valued map we introduce the following
sets. For a set-valued map F : X --+ Z and xl,x2 E D o m F , we denote, respectively, the
dominated set from below by sets F(xl) and F(x2) and the set of points dominating sets
F(xl) and F(x2) simultaneously from above by
Cr( F( xl), F(x2) ) = ( F( xl)~C) Q ( F(x2)t~C) , (3.9)

and
Cu(F(xl), f(xz) ) = (F(xl)AC) n ( f(x2)~C) . (3.10)
When F is a single-valued map, we can verify that

CL(f(xl), F(x~)) = C~(F(x~), F(x2)) = C(F(xl), F(x~)). (3.11)

By using two sets and the six different set-relations <(~) (k = i, . . . , vi), we consider gener-
alization of quasi C-convexity (3.7), but types (iv)-(vi) generalizations are meaningless since
the following conditions (3.12) and (3.13) are trivial in the cases.

DEFINITION 3.4. For each k = i, ii, iii, a set-valued map F : X -,z Z is said to be

• type (k)-lower q u a s i c o n v e x if for every xl, x2 E D o m e and A E (0, 1),

F (Ax~ + (1 - ;~)x2) <_(~) Cz(F(xl), F(x2)); (3.12)

• t y p e ( k ) - u p p e r q u a s i e o n v e x if for every xl, x2 • D o m e and ~ • (0, 1),

F ($xl + (1 - A)x2) _<(~) Ctr(f(xl),f(z2)). (3.13)

Second, we define Ferro's type quasiconvexity of set-valued map.

DEFINITION 3.5. A set-valued map F : X -.z Z is said to be

• F e r r o t y p e ( - 1 ) - q u a s i e o n v e x if for every z • Z,

F- (z - c ) := • x I f ( x ) n (z - C) # 0} (3.14)
is convex or empty;

• F e r r o t y p e ( + l ) - q u a s i c o n v e x if for every z • Z,

f + X ( z - C):= {z • X I f ( x ) c ( z - C)} (3.15)

is convex or empty.

These sets are said to be the lower level sets of set-valued map F, and Ferro type (-1)-
quasiconvexity and Ferro type (+l)-quasiconvexity are provided by convexity of their sets,
respectively. By Proposition 2.1. and simple demonstration, we have the following interesting
implications among quasiconvexities above, including the level-set convexity.
 Second World Congressof Nonlinear Analysts 1493

PROPOSITION 3.7. For a set-valued map F : X ~-+ Z, the following relationships hold:
type (i)-lower type (i)-upper Ferro type
quasiconvex quasiconvex (+ 1)-quasiconvex

type (ii)-lower type (ii)-upper

quasiconvex quasiconvex

Ferro type type (iii)-lower type (iii)-upper

( - 1)-quasiconvex quasiconvex quasiconvex

3.3, PROPERLY QUASI CONVEXITY OF SET-VALUED MAP

A vector-valued function f : X --~ Z is said to be properly quasi C-convex ([13, 14]) if for
every xl, x2 E X and A e (0, 1),

.:(AXl + (i - A)x2) <-c f(xl) or f (Axl + (i - A)x2) ~c/(x2). (3.16)

This condition can be described in another way, / (Axl + (1 - A)x2) E ( { f ( x l ) , / ( x 2 ) } - C),

and hence various types of generalization of the properly quasiconvexity can be considered,
but we concentrate upon a generalization of properly quasi C-convexity (3.16) to set-valued
map.
DEFINITION 3.6. For each k = i, . . . , vi, a set-valued map F : X ~.z Z is said to be t y p e (k)
p r o p e r l y q u a s i c o n v e x if for every Xl, x2 E D o m F and A E (0, 1),

k) F(xl) or F (Axl + (1 - A)x2) _<(~) F(x:).

F (Axl + (1 - A)x2) _<(C (3.17)

By Proposition 2.1, we have some implications among properly quasiconvexities above:

PROPOSITION 3.8. For a set-valued map F : X --~ Z, the following relationships hold:
type (i) properly quasiconvex type (iv) properly quasiconvex

type (ii) properly quasiconvex type (v) properly quasiconvex

type (iii) properly quasiconvex type (vi) properly quasiconvex

3.4. NATURALLY QUASI CONVEXITY OF SET-VALUED MAP

A vector-valued function f : X --~ Z is said to be naturally quasi C-convex ([13, 14]) if for
every xl, x2 e X and A E (0, 1), there exists # E [0, 1] such that

f (Axl + (1 - A)x2) <-c #f(xl) + (i - p)f(x2). (3.18)

This condition can be described in another way, f (Axl + (1 - A)x2) E (co {f(xl), f(x2)} - C),
and hence various types of generalization of the naturally quasiconvexity can be considered,
but we concentrate upon a generalization of naturally quasi C-convexity (3.18) to set-valued
map.
1494 Second WorldCongressof Nonlinear Analysts

DEFINITION 3.7. For each k = i, . . . , vi, a set-valued map F : X -.~ Z is said to be t y p e (k)
n a t u r a l l y q u a s i e o n v e x if for every xl,x2 E D o m F and )~ E (0, 1), there exists # E [0, 1]
such that
V (Axl + (1 - A)x2) <(~) # F ( x l ) + (1 - #)F(x2). (3.19)

By Proposition 2.1, we have some implications among naturally quasiconvexities above:

PROPOSITION 3.9. For a set-vaiued map F : X -,~ Z, the following relationships hold:

type (i) naturally quasiconvex type (iv) naturally quasiconvex

l
type (ii) naturally quasiconvex type (v) naturally quasiconvex
l
type (iii) naturally quasiconvex type (vi) naturally quasiconvex

Finally, we have the following results on the relationships among the generalized convexi-
ties of set-valued map introduced in the paper.

THEOREM 3.1. For a set-valued map F : X -,z Z, the following statements hold:

(i) For each k = i, . . . , vi, type (k) convexity implies type (k) convexlikeness;
(ii) For each k = i, . . . , vi, type (k) convexity implies type (k) naturally quasiconvexity;
(iii) For each k = i, . . . , vi, type (k) properly quasiconvexity implies type (k) naturally
quasiconvexity;
(iv) Type (iii) naturally quasiconvexity implies type (iii)-lower quasiconvexity;
(v) Type (vi) naturally quasiconvexity implies type (ii)-upper quasiconvexity;
(vi) Assume that C is a closed convex cone and that F is an upper semicontinuous and
convex-valued set-valued map. If F is type (iii) naturally quasiconvex then it is
also type (iii) convexlike.

PROOF. The assertions (i)-(iii) are obvious. The assertions (iv) and (v) can be easily verified
based on the following facts: given sets A and B, we have

CL(A, B) C , A + (1 - , ) B + C (3.20)

for all # E [0, 1], and

A u B c z - C (3.21)

for all z E Car(A, B). The proof of the assertion (vi) is contained in [6].
These results are similar to those of vector-valued versions; see [13, 14].
 Second World Congress of Nonlinear Analysts 1495

REFERENCES

1. AUBIN J.-P. & F R A N K O W S K A H., Set-Valued Analysis, Birkh{iuser, Boston (1990).

2. BORWEIN J., Multivalued Convexity and Optimization: A Unified Approach to Inequality and Equality
Constraints, Math. Programming, 13, 183-199 (1977).
3. C O R L E Y H.W., Existence and Lagrangian Duality for Maximizations of Set-Valued Functions, J. Op-
tim. Theory. Appl., 54, 489-501 (1987).
4. CORLEY H.W., Optimality Conditins for Maximizations of Set-Valued Functions, J. Optim. The-
ory. Appl., 58, 1-10 (1988).
5. HA T.X.D., Demicontinuity, Cone Convexity and Saddle Point Theorem for Set-Valued Maps, Preprint
of Hanoi Institute of Mathematics, 15 (1995).
6. K U R O I W A D., Convexity for Set-Valued Maps, Appl. Math. Lett., 9, 97-101, 1996.
7. KUROIWA D., Convexity for Set-Valued Maps and Optimization, Doctoral thesis, Niigata University,
Niigata (1996).
8. LUC D.T., An Existence Theorem in Vector Optimization, Math. Oper. Res., 14, 693-699 (1989).
9. LUC D.T., Theory of Vector Optimization, Lecture Notes in Economics and Mathematical Systems, 319,
Springer-Verlag, Berlin (1989).
10. LUC D.T. & VARGAS C., A Saddlepoint Theorem for Set-Valued Maps, Nonlinear Analysis: Theory,
Methods ~ Appl., 18, 1-7 (1992).
11. TAN K.K., YU J., & YUAN X.Z., Existence Theorems for Saddle Points of Vector-Valued Maps,
J. Optim. Theory. Appl., 89, 731-747 (1996).
12. T A N A K A T., Cone-Convexity of Vector-Valued Functions, Science Reports of Hirosaki University, 37,
170-177 (1900).
13. T A N A K A T., Generalized Quasiconvexities, Cone Saddle Points, and Minimax Theorem for Vector-
Valued Functions, J. Optim. Theory. Appl., 81,355-377 (1994).
14. T A N A K A T., Cone-Quasieonvexity of Vector-Valued Functions, Science Reports of Hirosaki University,
42, 157-163 (1995).
15. TANINO T. & SAWARAGI Y., Duality Theory in Multiobjective Programming, J. Optim. The-
ory. Appl., 27, 509-529 (1979).
16. TANINO T. & SAWARAGI Y., Conjugate Maps and Duality in Multiobjective Optimization, J. Op-
tim. Theory. Appl., 31, 473-499 (1980).
17. YU P.L., Cone Convexity, Cone Extreme Points, and Nondominated Solutions in Decision Problems with
Multiobjectives, J. Optim. Theory. Appl., 14, 319-377 (1974).
1496 Second World Congress of Nonlinear Analysts

A (BI~C
<l ) A Ir~C ) B
I
%
(i)
A< C B

AI~ (B[~ C ) ~
A <(ii)n
--C v
A <(iv)R
--C ""

A~.]C) B A~ B~]C
• " " <(v) R

-]

Figure h Six kinds of classification for set-relationship