Journal of Mathematical Analysis and Applications 240, 92 104 1999. Article ID jmaa.1999.6593, available online at http:r rwww.idealibrary.

com on

P-functions, Quasi-convex Functions, and Hadamard-type Inequalities 1

C. E. M. Pearce
Department of Applied Mathematics, The Uni ersity of Adelaide, Adelaide, South Australia, 5005, Australia

and A. M. Rubinov
School of Information Technology and Mathematical Sciences, Uni ersity of Ballarat, Gear A enue, Mount Helen, P.O. Box 663, Ballarat, Victoria 3353, Australia Submitted by Themistocles M. Rassias Received February 5, 1999

We establish some results concerning P-functions from the standpoint of abstract convexity. In particular, we show that the set of all P-functions on a segment is the least set closed under pointwise sum, supremum, and convergence and containing the class of all nonnegative quasi-convex functions on that segment. Further, generalizations are derived of a recent inequality of Hadamard type for 1999 Academic Press P-functions.

1. INTRODUCTION A nonnegative function p defined on the segment S is said to be a function of P type w1x or simply, a P-function. if p x q 1 y . y . F p x . q p y . ; x , y g S , 0 F F 1.

Let S s w a, b x and PS be the class of P-functions defined on S. It has been proved in w1x that for an integrable function f g PS , we have the
1

This research has been supported by Australian Research Council Grants. 92

0022-247Xr99 $30.00
Copyright 1999 by Academic Press All rights of reproduction in any form reserved.

HADAMARD-TYPE INEQUALITIES

93

Hadamard-type inequality f

aqb 2

H bya a

f x . dx F 2 f a . q f b . . .

In this paper we prove the following generalization of the left side of this inequality. Assume for the sake of simplicity that w a, b x s w0, 1x. If f is an integrable P-function and u g 0, 1., then f u. F

H min u , 1 y u . 0

f x . dx .

In fact we present a version of this inequality for an integral with respect to an atomless probability measure defined on the Borel -algebra of subsets of the segment w0, 1x. For nonnegative quasi-convex functions we show that this also holds for an arbitrary not necessarily atomless. probability measure. These results are the subject of Section 4. More generally, we study links between P-functions and nonnegative quasi-convex functions, which form an important class of generalized convex functions see, for example, w5x.. It is well known that the sum of quasi-convex functions is not necessarily quasi-convex. The cone hull of the set of all quasi-convex functions defined on a segment S is a very large set, containing, for example, all functions of bounded variation. The cone hull of the set Qq of all nonnegative quasi-convex functions is also very broad, but one can find nonnegative functions of bounded variation which do not belong to this set. The pointwise supremum of a family of elements of Qq is again an element of Qq. One of the important problems of the theory of quasi-convex functions is to describe the least cone containing Qq which is closed in the topology of pointwise convergence and contains pointwise suprema of all families of its elements. In Section 3 we show that this cone coincides with the set PS of all P-functions defined on S. We use methods of abstract convexity see, for example, w3, 4, 7x.. Our approach is based on the description of small supremal generators of the sets under consideration.

2. PRELIMINARIES First we recall some definitions from abstract convexity. Let s y, q. be a real line and q s j q4 . Consider a set X and a set H of functions h: X defined on X . A function f : X q is called abstract con ex with respect to H or H-con ex . if there exists a set

94

PEARCE AND RUBINOV

U ; H such that f x . s sup h x . : h g U 4 . The set s f , H . s h g H : h x . F f x . for all x g X 4 is called the support set of a function f with respect to H. Clearly f is H-convex if and only if f x . s sup h x . : h g s f , H .4 for all x g X . Let Y be a set of functions f : X q . A set H ; Y is called a supremal generator of the set Y if each function f g Y is abstract convex with respect to H. We consider only nonnegative functions defined on the real line and mapping into w0, qx. Recall that a function f defined on is called quasi-con ex if f x q 1 y . y . F max f x ., f y .. for all x , y g R and g 0, 1.. A function f is quasi-convex if and only if its lower level sets x : f x . F c4 are segments for all c g . Let us give some examples. EXAMPLE 1 See for Instance w6x.. Let H1 be a set of two-step functions h of the form h x . s

c 0

x G d , x - d

with g 1, y14 , c G 0, d g . Then H1 is a supremal generator of the set Qq of all nonnegative quasi-convex functions. Indeed H1 ; Qq; since the pointwise supremum of a family of quasi-convex functions is again quasi-convex, it follows that each H1-convex function belongs to Qq. Now consider a function q g Qq and the family of level sets Sc s x : q x . F c4 with c G 0. Since q is quasi-convex it follows that S c is a segment for each c G 0. Let x o g . Assume for the sake of definiteness that q x o . ) 0 and let 0 - c - q x o .. Since x o f S c it follows that there exist g y1, 14 such that x o ) x for all x g Sc . Let d s x o . Then c [ inf x G d q x . F q x o .. Since Sc ; x : x - d4 it follows that the inequality x G d implies q x . ) c. Hence c F c. Let h x . s

c 0

x G d . x - d

Clearly h F q and h x o . G c G c. Since c is an arbitrary number such that 0 - c - q x o . it follows that q x . s sup h x . : h g s f , H1 .4 for all x g . EXAMPLE 2 See, For Example w6x.. Let H2 be the set of all two-step functions of the form h x . s

c 0

x ) d , x F d

HADAMARD-TYPE INEQUALITIES

95

with the same , c, and d as in Example 1. Then H2 is a supremal l generator of the Qq of all lower semicontinuous nonnegative quasi-convex functions. EXAMPLE 3. Let H3 be the set of all functions h of the form h x . s

c 0

xsu , x/u

with u g , c G 0. Then H3 is a supremal generator of the set of all nonnegative functions defined on . EXAMPLE 4. Let H4 be the set of all Urysohn peaks on , that is, continuous functions g : q of the form

0 g x . s~c affine affine

< x y u< G xsu , uy-x-u u-x-xq

where u g , c G 0, and ) 0. It is easy to check that H4 is a supremal generator of the set of all functions that are lower semicontinuous on .

3. P-FUNCTIONS A function p: w0, qx is called a function of type P w1x or P-function. if p x q 1 y . y . F p x . q p y . for all g 0, 1 . and x , y g . 1.

Denote by P the set of all P-functions. Let us point out some properties of a function f g P. 1. If i ) 0, i s 1, . . . , m. and m m . is 1 i s 1, then f is 1 i x i F f x i .. This can be proved by induction. 2. The set dom f s x g : f x . - q4 is a segment. Indeed let xys inf dom f , xqs sup dom f . Suppose xy- x - xq. Then there exist points x 1 , x 2 g dom f such that x g x 1 , x 2 .. It follows from the definition of P that f x . - q, that is, x g dom f . 3. If the set x : f x . s 04 is nonempty, then it is clearly a segment. m is 1 Let S ; be a segment and PS the set of all P-functions defined on the segment S and mapping into w0, qx. We have P s PS with S s .

96

PEARCE AND RUBINOV

Let S / . For each function f defined on S, consider its extension fq defined by fq x . s

f x q

xgS . xfS

2.

Clearly f g PS if and only if fq g P. Let S be a segment. It is easy to check that the class PS enjoys the following properties. 1. PS is a cone: if f 1 , f 2 g PS , then f 1 q f 2 g PS ; if ) 0, f g PS , then f g PS ; 2. PS is a complete upper semilattice: if f . g A is a family of functions from PS and f x . s sup g A f x ., then f g PS ; 3. PS is closed under pointwise convergence. The classes PS are extremely broad. We now describe some subclasses of PS . 1. Each quasi-convex nonnegative function defined on S belongs to PS . In particular nonnegative convex, increasing, and decreasing functions defined on S belong to PS . 2. Let h be a bounded function defined on S. Then there exists a number c ) 0 such that the function f x . s h x . q c belongs to PS . Indeed, let c s sup x , y g S , z g w x , y x h z . y h x . y h y .. We have for each x , y g S and g w0, 1x that h x q 1 y . y . F h x . q h y . q c. Let f x . s h x . q c. Then f x q 1 y . y . s h x q 1 y . y . q c F h x . q c . q h y . q c . s f x . q f y . , that is, f g PS . We now describe a small supremal generator of the set P. Applying this generator and the extension defined by 2., we can easily describe a supremal generator of the class PS for a segment S g . Let T be the set of all collections t s u; c1 , c 2 4 with u g and nonnegative c1 , c 2 . For t s u; c1 , c 2 4 g T , consider the function h t de-

HADAMARD-TYPE INEQUALITIES

97

fined on by

h x . s~c
t

c c q c
1 1 2

x-u xsu. x)u

3.

It is easy to check that h t g P for all t g T . Let H be the set of all functions of the form h t with t g T . Clearly H is a conic set, that is, if h g H and ) 0, then h g H. The following statement describes a certain extremal property of elements h g H. PROPOSITION 1. Let h g H , h s h t with t s u; c1 , c 2 4 . If f g P, f F h, and f u. s h u., then f s h. Proof. Let f g P, f F h, and f u. s h u.. Take a point y - u and find a point z ) u and a number g 0, 1. such that u s y q 1 y . z. Since f g P it follows that c1 q c 2 s h u . s f u . F f y . q f z . F h y . q h z . s c1 q c 2 . Hence f y . q f z . s c1 q c 2 . Since f y . F c1 and f z . F c 2 it follows that f y . s c1. In the same manner, we can show that f . s c 2 for an arbitrary point ) u. PROPOSITION 2. H is a supremal generator of P. Proof. Let f g P and u g . First assume that u g dom f . Let ) 0 and cX1 s inf x - u f x ., cX2 s inf x ) u f x .. We now check that f u. y 2 F cX1 q cX2 . Let points x 1 - u and x 2 ) u be such that f x 1 . F cX1 q and f x 2 . F cX2 q , respectively. Then f u. y 2 F f x 1 . q f x 2 . y 2 F cX1 q cX2 . Take nonnegative numbers c1 and c 2 such that c1 F cX1 , c 2 F cX2 , and c1 q c 2 s f u. y 2 . Consider the function h t with t s u; c1 , c 2 4 g T . It follows from the definition of t that h t F f and h t u. s f u. y 2 . Thus f u. s sup h u. : h g H , h F f 4 for all u g dom f . Consider now a point u f dom f . Assume for the sake of the definiteness that u F inf dom f . Let t s u; c1 , c 2 4 g T , where c 2 s inf x 2 g f x . and c1 is an arbitrary positive number. Then h t F f . So f u. s q s sup h g H , h F f h u.. Remark 1. This proposition may be compared with the examples of Section 2. Indeed we can consider H as a certain mixture of two-step functions from Example 1 and pointed functions from Example 3. Clearly the function h t is a function of bounded variation for each t g T . Therefore h t can be represented as the sum of increasing and

98

PEARCE AND RUBINOV

decreasing functions. We now show that for each t g T the function h t can be represented as the sum of nonnegative increasing and decreasing functions. PROPOSITION 3. Let t s u; c1 , c 2 4 g T . Then there exist a nonnegati e 2 increasing function h1 t and a nonnegati e decreasing function h t such that 1 2 ht s ht q ht . Proof. Let 0 h1 t x. s c 2 x-u xGu; xFu . x)u

h2 t x. s

c1 0

t 1 2 it is easy to check that h1 q h2 t s h t , h t is an increasing function, and h t is a decreasing function.

Let c Qq . be the cone hull of the set Qq of all nonnegative quasi-convex functions, that is, the set of all functions f of the form f s q1 q q2 where q1 , q2 g Qq. It follows from Proposition 3 that H ; c Qq .. be the upper semilattice generated by H , that is, the set of all Let H functions h of the form

h x . s

is 1, . . . , m

max

hi x . ,

h i g H , i s 1, . . . , m ; m s 1, 2, . . . ,

and H be the upper semilattice generated by c Qq ., that is, the set of all functions h of the form h x . s
is 1, . . . , m

max

q i x .

q i g c Qq . , i s 1, . . . , m ; m s 1, 2, . . . .

s cl H , where cl A is the closure of the set A PROPOSITION 4. P s cl H in the topology of pointwise con ergence. . s : Proof. Let f g P and s f , H hgH h F f 4 be the support set of . We can consider s f , H . as a directed set with respect f with respect to H to the natural order relation: h1 G h 2 if h1 x . G h 2 x . for all x g . Since .4 f x . s sup h x . : h g s f , H . 4 s sup h x . : h g s f , H
x g . ,

.4 is increasing, it follows that and the generalized sequence h: h g s f , H . f x is the pointwise limit of this generalized sequence. Thus . P ; cl H
4.

; H. As P is a cone and an upper Since H ; c Qq . it follows that H semilattice and the set Qq is contained in P it follows that H ; P. Since P is closed in the topology of pointwise convergence, it follows that ; cl H ; P. The desired result follows from this inclusion and 4.. cl H

HADAMARD-TYPE INEQUALITIES

99

Remark 2. We have proved that a function f belongs to P if and only if this function can be represented as the pointwise limit of a generalized sequence f . where each f is a finite maximum of the functions represented as the sum of two nonnegative quasi-convex functions. It also follows from Propositions 2 and 3 that each f g P can be represented as the supremum of a family of functions belonging to c Qq .. Since P is a cone and a complete upper semilattice and Qq; P, it follows that P coincides with the set of all functions which can be represented in such a form. Let Pl be the set of all l.s.c. functions belonging to P. We now describe a supremal generator of Pl consisting of continuous functions. Consider the set S of all collections s s u; c1 , c 2 ; 4 , where u; c1 , c 2 4 g T and ) 0. For s g S define the continuous function l s by l s x . s c1 x F u y . , l s u . s c1 q c 2 , l s x . s c2 x G u q . , l s is affine on segments w u y , u x and w u , u q x . Denote the set of all functions l s with s g S by L. PROPOSITION 5. L is a supremal generator of Pl . Proof. Let f g Pl and u g dom f . Since f is l.s.c., it follows that for each ) 0 there exists ) 0 such that f x . ) f u. y if < x y u < - . There exist numbers c1 F inf x - uy f x . and c 2 F inf x ) uq f x . such that f u. y s c1 q c2 , for the same reasons as those given in the proof of Proposition 2. Let s s u; c1 , c 2 ; 4 . It follows from the definition of the numbers c1 , c 2 , and that l s F f . Since l s u. s f u. y it follows that f u. s sup l u. : l g L, l F f 4 . It is easy to check that this equality also holds for points u f dom f . PROPOSITION 6. Each function l g L can be represented as the sum of increasing and decreasing continuous functions. Proof. Let l s l s with s s u; c1 , c 2 ; x. A simple calculation shows that l s l 1 q l 2 where l1 s

~ ~

0c

xFuy
2

c c
2 1

xyuq.

x g u y , u. , x)u xFu

l2 s

c1

u q y x.

x g u, u q . x)uq

100

PEARCE AND RUBINOV

It is easy to check that l 1 , l 2 are continuous, l 1 is increasing, and l 2 is decreasing. For the same reasons as those in the proof of Proposition 4, it follows s cl L and cl Pl s cl L s cl L where L is the set of all that Pl ; cl L functions that can be presented as the maximum of a finite family of elements of L and L is the set of all functions that can be presented as the maximum of the finite family of the sum of two continuous quasi-convex functions.

4. INEQUALITIES OF HADAMARD TYPE We begin with the following principle of preser ation of inequalities w3x. PROPOSITION 7. Let Y be a set of functions defined on a set X and equipped with the natural order relation. Let H be a supremal generator of Y . Further , let a be an increasing functional defined on Y and u g X . Then h u . F a h . for all h g H if and only if f u . F a f . for all f g Y . Proof. We have f u . s sup h u . : h g s f , H . 4 F sup a h . : h g s f , H . 4 F a sup h : h g s f , H . 4 . s a f . .

We now establish some inequalities of Hadamard type for P-functions by applying the principle of preservation of inequalities. Consider the Borel -algebra of subsets of the segment w0, 1x and a measure , that is, a nonnegative -additive function defined on . Assume that w0, 1x s 1. Let Po be the set of all measurable with respect to . functions f g P such that dom f s w0, 1x. Since H consists of Borel-measurable functions on and H is a supremal generator of P, it follows that H is a supremal generator of Po as well. Let I : Po w0, qx be the functional defined by I f . s For y g w0, 1. consider the functions e1 y x. s

H0

f d .

1 0

xFy ; x)y

e2 y x. s

0 1

xFy . x)y

HADAMARD-TYPE INEQUALITIES

101

Let g 1 y . s

H0

1 1

e y d s w 0, y x . , g 2 y . s

H0

1 2

e y d s y , 1x . , y g 0, 1 . .

By definition, set g 11. s 1, g 2 1. s 0. Clearly g 1 is increasing, g 2 is decreasing, and g 1 y . q g 2 y . s w0, 1x. s 1 for all y g w0, 1x. Let us calculate I h t . for h t g H. LEMMA 1. Let be an atomless measure, that is, x 4. s 0 for each x g w0, 1x. Let h t g H be the function corresponding to a collection t s y ; c1 , c 2 4 . Then I h t . s c1 g 1 y . q c 2 g 2 y .. Proof. Consider the function h t given by c ht x . s 1

c2

xFy s x)y

c1 e 1 y x. c2 e 2 y

xFy x)y

x.

Since is atomless, it follows that I h t . s I h t .. We have I ht . s

H0

ht d s

H0

1 2 h t d s H c1 e 1 y d q H c2 e y d 0 y

s c1 g 1 y . q c 2 g 2 y . . For u g 0, 1., consider the number

u s
It easy to check that

c1G 0, c 2G 0

min

c1 g 1 u . q c 2 g 2 u . c1 q c 2

5.

u s min g 1 u . , g 2 u . . .

6.

Indeed if g 1 u. G g 2 u., then u s g 2 u. and if g 1 u. F g 2 u., then u s g 1 u.. Thus 6. holds. THEOREM 1. Let be an atomless measure and u g 0, 1.. Then f u. F for all f g Po .

H min g u . , g u . . 0
1 2

f d

7.

102

PEARCE AND RUBINOV

Proof. Clearly I is an increasing functional defined on the set Po . First we check that 7. holds for all h g H. We consider separately functions h t which are defined by collections t s u; c1 , c 2 . and by collections t s y ; c1 , c 2 . with y / u. Let t s u; c1 , c 2 4 . It follows directly from 5., 6., Lemma 1, and the equality h t u. s c1 q c 2 that h t u . s c1 q c 2 F s 1 min g 1 u . , g 2 u . . 1

c1 g 1 u . q c 2 g 2 u . . s

H0

ht d

H0

ht d .

8.

Assume now that t s y ; c1 , c 2 4 with y / u. It follows from Lemma 1 that I ht . s

H0

h t x . dx s c1 g 1 y . q c 2 g 2 y . .

Let us calculate h t u.. Since y / u it follows that either h t u. s min c1 , c 2 . or h t u. s max c1 , c 2 .. In the first case we have, taking into account that g 1 y . G 0, g 2 y . G 0, and g 1 y . q g 2 y . s 1, that h t u . s min c1 , c 2 . F c1 g 1 y . q c 2 g 2 y . s I h t . . Since g 1 u. F 1, g 2 u. F 1, it follows that ht u. F I ht . F 1 min g 1 u . , g 2 u . . I ht . .

9.

Assume now that h t u. s max c1 , c 2 .. If c1 G c 2 then h t u. s c1 and y ) u, the latter following directly from the definition of the function h t . Since g 1 is an increasing function, we have for y ) u that g 1 u . h t u . F g 1 y . c1 F g 1 y . c1 q g 2 y . c 2 s I h t . . Thus ht u. F 1 g 1 u . I ht . F 1 min g 1 u . , g 2 u . . I ht . .

10 .

If c 2 G c1 , then h t u. s c 2 and y - u. In the same manner we have ht u. F 1 g2 u. I ht . F 1 min g 1 u . , g 2 u . . I ht . .

11 .

HADAMARD-TYPE INEQUALITIES

103

Thus we have verified that the desired inequality 7. holds for all h g H. Since H is a supremal generator of Po and I is an increasing functional, we can conclude, by applying the principle of preservation of inequalities, that 7. hold for all f g Po . Remark 3. Let t s u; c1 , c 2 4 , where u c1 q c 2 . s c1 g 1 u. q c2 g 2 u.. It follows from 8. that the equality ht u. s 1 min g 1 u . , g 2 u . .

holds. Thus the inequality 7. cannot be improved for all P-functions. Let us give an example. EXAMPLE 5. Let be the Lebesgue measure, that is, I f . s H01 f x . dx. Then g 1 y . s y, g 2 y . s 1 y y. It follows from Theorem 1 that f u. F 1 min u , 1 y u .

H0

f x . dx

for all f g Po . In particular we have f

/
2

F2

H0

f x . dx

f g Po . .

12 .

This result was established in w1x. Remark 4. In a similar manner we can prove that 7. holds for each nonnegative quasi-convex function and for each not necessarily atomless. nonnegative measure such that w0, 1x. s 1. We can use for this purpose the supremal generator described in Example 1. Indeed the atomlessness of the measure has been used only for eliminating the special value of the function h t with t s u; c1 , c 2 4 at the point u. We do not need to eliminate this value in the case under consideration.

REFERENCES
1. S. S. Dragomir, J. E. Pecaric, and L. E. Persson, Some inequalities of Hadamard type, Soochow J . Math. 21 1995., 335 341. 2. S. S. Dragomir and C. E. M. Pearce, Quasi-convex functions and Hadamards inequality, Bull. Austral. Math. Soc. 57 1998., 377 385. 3. S. S. Kutateladze and A. M. Rubinov, Minkowski duality and its application, Russian Math. Sur eys 27 1972., 137 192.

104

PEARCE AND RUBINOV

4. D. Pallaschke and S. Rolewicz, Foundations of Mathematical Optimization Convex Analysis without Linearity ., Kluwer Academic, Dordrecht, 1997. 5. J. Ponstein, Seven types of convexity, SIAM Re . 9 1967., 115 119. 6. A. M. Rubinov and B. M. Glover, On generalized quasiconvex conjugation, Contemp. Math. 204, 199 216. 7. I. Singer, Abstract Convex Analysis, Wiley-Interscience, New York, 1997.