Rivista di matematica per le scienze economiche e sociali - Anno 15 ~ Fascicolo 1"

D I N I D E R I V A T I V E S IN O P T I M I Z A T I O N - P A R T I (*)

G. G I O R G I
Dip. di Ricerche Aziendali
Univ. of Pavia (Italy)

S. K O M I . ~ S I
Fac. of Economics Jauns Pannonius
University Pdcs (Hungary)

Versione definitiva pervenuta il 18/02/1992

This paper, published in two parts, is mainly concerned with general properties of Dini deriva-
fives of functions of one and several variables and with some applications of this topic to the study of
generalized convexity and generalized optimality conditions for mathematical programming problems.
In part I the basic definitions and properties are given, with reference both to functions of one
real variable and to functions of several real variables. In this part special attention is given to the
restatement of the basic theorems of the classical analysis to nondifferentiable functions, in terms of
Dini derivatives.
In part I1 we use these derivatives in order to define some classes of nondifferentiable generalized
convex functions and the class of generalized upper quasidifferentiable functions. This part concludes
with the development of optimality conditions for a nonsmooth programming problem, expressed in
terms of the tools prevously introduced.

Introduction

T h e classical concept o f derivative (Fr6chet-, Gfiteaux- and directional) has been

proved one o f the most useful tools in mathematics, as a large variety o f problems has
been described and solved by means o f this topic. In solving classical optimization prob-
lems the use o f derivatives is inevitable. ~Classical,, means in this context that the func-
tions i n v o l v e d in the problem are differentiable or continuously differentiable or higher

(*) This research was supported by the Italian Ministry of University and Scientific Research and
by the National Science Foundation of the Hungarian Academy of Sciences and Arts (grant OTKA
354/86). The authors have shared their work as follows: Chapters 1 and 3 are attributed to the first
author, whereas Chapters 2 and 4 are attributed to the second author. The Introduction and Chapter
5 is common.
order differentiable. Besides treating the classical smooth problems, the mathematicians
got many impulses in the last decades from other sciences (mainly from economics, en-
geneering, etc.) in order to treat nonsmooth, nondifferentiable optimization problems.
The analysis of such problems has definitely required some generalizations of the
derivative. Since the early 1960's much effort has gone into the development of a gener-
alized kind of differentiation that can be useful in the analysis of optimization problems.
The subject has grown very rapidly since then.
Our aim in this paper is to collect and to state the most relevant (old and new) results
concerning Dini detivative~ the roots of whose study stretch back to the end o f the last
century [16, 39]. Besides gathering together the relevant results concerning this subject
we provide some new results as well and reformulate some known results in a more
complete or more general form.
Chapter 1 is concerned with the Dini derivatives of a function of one variable. After
summing up some basic properties and calculus rules, some extensions and general-
izations of the classical Rolle and Lagrange Mean Value Theorems are discussed. The
common feature of these kinds of results lays in establishing the relationships between
the increment of a function on a closed interval and the values of its Dini derivatives.
These results allow to characterize monotonicity and continuity in terms of the Dini
derivatives. At the end of this chapter the classical second order optimality conditions
are extedend for twice Dini differentiable functions.
Chapter 2 is concerned with the study of different Dini derivative concepts for func-
tions of several variables. The basic properties of the directional Dini derivatives and the
directional Dini-Hadamard derivatives (the contingent derivatives) are reviewed. At the
end of this chapter the calculus rules for max-functions and an implicit function theorem
for nondifferentiable functions are presented.
In Chapter 3 quasiconvexity and pseudoconvexity are characterized in terms of di-
rectional Dini derivatives.
Generalized convexity, as its name points out, is an extension of the convexity con-
cept. Quasidifferentiability provides another path in extending the convexity concept
(and the differentiability concept, as well). Chapter 4 is devoted to the study of general-
ized upper quasidifferentiability, a recent extension of the quasidifferentiability notion.
In Chapter 5 optimality conditions are derived, for both unconstrained and
constrained mathematical programming problems described by generalized upper qua-
sidifferentiable functions.
Owing to the relevance of the subject with respect to the present line of researches
in optimization theory and for the reader's convenience, we have preferred to give the
proofs of several results, which has compelled us to divide the paper in two parts: part
I, composed by the Introduction and Chapters 1 and 2; part II, composed by the same
Introduction and the remaining Chapters 3, 4 and 5.
1. Dini derivatives of function of single variable
Let f ( t ) be a real valued function defined over the open interval ( a, b). Let t o E
E (a, b). We define the four Dini derivatives of f ( t ) at t o as follows:

D§ f (to) = lim sup f(t) - f(t~ , D § (to) = lim inf f(t) - f (to)
t-.t o. t -- t o t-.to, t -- t o

D - f (to) = l i m s u p f ( t) - f (to) , D f (to) = lim inf f ( t) - f (to)

t-.to- t -- to t~to- t -- t o

D + f ( t o ) is the upper right Dini derivative, D+ f ( t o ) is the lower right Dini derivative,
D - f ( t o) is the upper left Dini derivative, D_ f ( t o ) is the lower left Dini derivative
of f ( t ) evaluated at t o .
It is obvious that the following inequalities always hold:

D§ <_D*: (to) and D_: (to)<_D-/(to).

We allow infinite limits in the above definitions, so the Dini derivatives may attain
the infinite values +o,o, - o o , as well. The typical feature of Dini derivatives is that they
always exist and admit very useful calculus rules, as well. The following calculus rules
come from the properties of the ,limsup>~ and <,liminf~ operations and from the definition
of the Dini derivatives. For a larger collection of such rules we refer the reader to the
book of Mc Shane [41 ].
Let f l ( t ) and f z ( t ) be arbitrary functions defined on a neighbourhood of t o E R .
Then
D+ [fl + f2] (to) -( D+fl (to) + D+f2 (to) (la)

and
D+ [fl+ f2] (to)-> D+f, (to)+ D+iz (to); (Ib)
in particular if f~(to) exists then

D" [:~ + A] (to)= D':, (to)+ :; (to). (lc)

Let f ( t ) be continuous and g(t) be continuously differentiable at t o . If g(t0) _> 0

then
D + [ f 0 ] (to) = f (to) g' (to) + g (to) D+ f (to), (2a)

and if g(to) <_ 0 then

z~§ :ol (to)= : (to)o'(to)+ o (to)D.: (to). (2b)

For arbitrary f ( t ) we have the following rules:

D+[-f](t) = -D+f(t), D + [ - f ] ( t ) = -D+ f ( t ) , (3a)

D-t-f](t) = -D_f(t), D_[-f](t) = -D-f(t). (3b)

The following statement provides the link beetwen the Dini derivatives and the deriva-
tive of a function of one variable: f ( t ) is differentiable at t o if and only if the four
Dini derivatives at t o are equal and finite. Then

:' (to) -- ,9+f (to) = ,9§ (to) = ,D- f (to) = ,9_f (to).

Nondifferentiable functions having finiteDini derivatives are closely related to the

differentiable functions: they are continuous and almost everywhere differentiable.

THEOREM 1.1. Let f ( t) be delfned on a neighborhood of t o . If each of the four Dini

derivatives is l~nite at to, then the difference quotient

/ ( t ) - / (t o)
t - to

is bounded around to, consequently f ( t) is continuous at t o .

P r o o f If the function f ( t ) is not continuous at to, then there exist e > 0 and a
sequence t , 4 t o such that I f ( t , ) - f ( t 0 ) l > c. It follows that

I f(t')-'f(t~ 4+oo

for t , 4.4 to and this implies that at least one of the four Dini derivatives is not finite.
Q.E.D.

The following statement is a special case of the Denjoy-Young-Saks theorem [ 12, 37,
44].

THEOREM 1.2. Let f ( t) be detlned on the open interval I and the Dini derivaO'ves
D + f ( t ) , D + f ( t ) , D - f ( t ) , D _ f ( t ) betiniteateachpointofI. Then f ( t ) isdiffer-
entiable on I almost everywhere.

The following example shows that if all of the four Dini derivatives are not finite,
continuity is in general not ensured.

6
EXAMPLE 1. Let us consider the following function:

f ( t) = f o if t>0 and t is irrattbnal

if t < O or t is rational.

f ( t ) fails to be continuous at t o = 0 ; an easy calculation shows that D - f ( O ) ,

D _ f ( O ) = O, D+ f(O) = O, D+f(O) = - o o .

Generalized mean value theorems

We have several extensions of the main theorems of the analysis to nondifferentiable
functions that have finite Dini derivatives [5, 15, 22, 27, 32, 33, 40, 41,43]. Mean value
theorems play a central role in the classical theory of the Differential Calculus. Now we
provide some extensions of the classical Rolle and Lagrange theorems for nondifferen-
tiable functions.

Rolle-type theorems. The common feature of these kinds of theorems lays in es-
tablishing special properties of the derivative (or generalized derivatives) of a function
defined on a closed interval and assuming the same values at the endpoints of that given
interval.

The folliwing lemma plays a crucial role in extending several statements of the clas-
sical analysis.

LEMMA 1.3. [5, Lemma 1] Let f ( t ) be defined on the open interval I . I f D+ f ( t o ) >
> 0 and D _ f ( t o ) > O, where t o E I, then f ( t ) islocallysttictlyincreasingat t o .
I f D+ f ( to) < 0 and D - f ( t o) < 0 then f ( t) is locally strictly decreasing at to 9

Proof. Assume that D + f ( t o) > 0 and D r ( t o ) > 0 . Then there exist t',t" E I,
t' < t o < t" such that
f ( O - f (to)
>0
t - to

for every t E ( t', t") , t :t to 9 It follows that

f(t) < f(to) whenever t E (t',to)

f ( t ) > f (to) whenever t E (to,t").

The second part of the theorem can be prove,de by the same reasoning. Q.E.D.

Now we show that if a function is strictly monotone at each point of an interval,

then it is strictly monotone over the whole given interval (this is a result of the classical
analysis, although specified in few text-books).
LEMMA 1.4. Let f ( t) be locally stn'ctly increasing (decreasing) at every point os the
dosed interval [ a, b]. Then f ( t) is stn'ctly increasing (decreasing) on [ a, b] .

Proof. Let f ( t ) be locally strictly increasing at each point of [ a, b]. Assume for
contradiction that there exist t l , t 2 E [a,b], suchthat t I < t 2 and f ( t l ) > f ( t 2 ) .
Set
T= {tit1 <t<t2 and f(tl) <f(t)}
and
to = sup 7'.
Since f ( t ) is locally strictly increasing at t I and t2, therefore T -~ O and t I <
< to < t2 9
Since f ( t ) is locally strictly increasing at to, therefore there exists 6 < O, such
that
to - 6 < t < to implies f ( t ) < f (to) (4a)
and
to < t < to + 6 implies f (to) < : ( t ) . (4b)
Bythedefinitionof to,t o < t impliesthat t E T, consequently f ( t) < f ( t 1) . Taking
into account (4b) it follows that forany t o < t < to+6 onehas f ( t o ) < f ( t ) < f ( t l)
and thus
: (to)< : (t,).
Ontheotherhand thereexists t' E (to -6, to) such that t' E T .Taking imo account
(4a) and the definitionof 7", one has f(tl) <_ f(t') < f(to), which contradicts the
above inequality f( ti) > f(to) 9 This contradictionproves the statement of the above
theorem.
The statement on strictly decreasing functions is proved in a similar manner. Q.E.D.

The following Rolle-type theorem, due to E. Castagnoli, does not impose any extra
requirements on the function.

THEOREM 1.5. [5] Let f ( t ) be detined on the closed interval [ a, b] and let f ( a ) =
= f ( b ) . Then there exist t 1, t2 E [ a, b] such that

either D+f (tl) < 0 or D _ : (tl) <_ 0

and
either D + f (t2) > 0 or D - : (t2) > 0.

Proof. Assume for contradiction that there does not exist t~ with the required prop-
erty. Then D+f(t) > 0 and D _ f ( t ) > 0 for all t E [a,b]. It follows by L e m m a
1.3 that f ( t ) is locally strictly increasing at each point of [ a, b] and by Lemma 1.4
that it is strictly increasing on [ a, b], which contradicts the assumption f ( a ) = f ( b ) .
The second part of the theorem can be proved by the same reasoning. Q.E.D.

If we impose some special properties on the function we can obtain more information
about the behaviour of the dini derivatives.

THEOREM 1.6. (Generalized Rolle theorem) Let f ( t) be defined on [ a, b] and let

f ( a ) = f ( b).
i) L e t f ( t ) assume a minimal value m on [ a, b] .
I f m < f ( a ) = f ( b ) thenthereexists t I E ( a , b ) suchthat

D- y (tl)< 0 < D+f (tl), (5a)

whereas i f m = f ( a ) = f ( b ) then one has

D-/(b) < 0 < D+y(a). (5b)

ii) L e t f ( t ) assume a maximal value M on [a, b] . I f M > f ( a ) = f ( b ) then

there exists t2 E (a, b) such that

D+f(t2) < 0 < D_f(t2),

whereas i f M = f ( a ) = f ( b ) then one has

D§ <0 <D f(b).

Proof. i) By assumption there exists t I E [a,b] suchthat f ( t l) = m . If m < f ( a )

then t 1 E ( a , b ) . Since in this case wehave f ( t ) >__f ( t 1) forall t E [a,b], therefore
(5a) holds. In the other case, when m = f ( a ) , one has f ( t ) > f ( a ) = f ( b ) for all
= t E [ a, b], from which (5b) follows immediately.
The proof of statement ii) is carried on by similar arguments. Q.E.D.

REMARK. If f ( t ) is continuous on [a, b] and differentiable on (a, b) then the above

lemma coincides with the classical Rolle theorem, whereas if f ( t ) is continuous only,
then the statement of the above lemma is due to D.B. Goodner [22, Theorem 1]; for
semicontinuous function it gives a result of W.E. Diewert [15].

One can also find Rolle-type statements involving Dini derivatives and continuous
functions in some other papers (see [39, 43]). Now we present a more complete version
of the said results.
THEOREM 1.7. Let the function f ( t ) be defined and continuous on the closed interval
[ a, b] . A s s u m e that f ( a ) = f ( b ) . Then there exist intennediate values t l, t2, t3, t4 E
E (a, b) such that

D+f(tl)>O, D+f(t2)<O, D_f(t3)>_O, D-f(t,)<O. (6)

Proof. According to a well known theorem, the set of the function values is a closed
interval:
{f(t) :t E [a,b]} = [m,M].

Case i): m = M .
In this case f ( t ) is constant over [ a, b] so all of the Dini derivatives are equal to
0 for every t E (a, b).
Case ii): M > f ( a ) .
In this case there exists d E (a,b) suchthat f ( d ) = M . Let f ( a ) < y < M .
Since f ( t ) is continuous there is a last point d I in [a,d] and a first point d 2 in [d,b]
where the function f ( t) assumes the value y ( f ( d 1) = f ( d 2) = ~1) . Put t I = dl,
t 2 -- d, t 3 = d, t 4 = d,2 . It is not difficult to show that the inequalities (6) are fulfilled
with this choice.
Case iii): m < f ( a ) .
The proof goes by the same arguments of ii). Q.E.D.

REMARK. Note that Theorem 1.7 is not a consequence of Theorem 1.6. The arguments
used in the proof of Theorem 1.6 are not sufficient to obtain Theorem 1.7.

Mean value-type theorems. These kinds of results establish relationships between

the increment of a function on a closed interval and the Dini derivatives. The following
theorem has been proved by W.E. Dierwert [15, Theorem 1 and Corollary 1].

THEOREM 1.8. (Generalized Mean Value Theorem) Let the function f ( t) be defined
on the closed interval [ a, b] . Put

f(b) - f(a)
a = and h(t) = f ( t ) - f ( a ) - a ( t - a). (7)
b-a

/) //" f ( t ) is lower semicontinuous on [ a, b] and

m = i n f { h ( t ) : a < t < b} < O,

then there, exists t I E ( a, b) such that

D-/(t~) < a< D j ( t l ) ,

10
whereas i f m = O, then one has

D - f ( b ) < ~ < D+ f ( a ) .

ii) / f f ( t ) is upper semicontinuous on [ a, b] and

M=sup(h(t) :a<_t<b)>0,

then there exists t 2 E ( a, b) such that

D§ < < D_S .

whereas i f M = O, then one has

D § f(a) < ~ < D_ S(b).

Proof. By the generalized Weierstrass theorem a lower (upper) semicontinuous func-

tion assumes a minimum (maximum) on a closed interval, hence the function h(t) sat-
isfies the conditions of Theorem 1.6. Taking into account that

D + h ( t ) = D§ f ( t ) - o,, D-h(t) = D-f(t) - c~,

D + h ( t ) = D + f ( t ) - or, D _ h ( t ) = D f ( t ) - c~

the statement of the above theorem is immediately obtained by means of Theorem 1.6.
Q.E.D.

REMARK. If f ( t ) is differentiable then the above theorem gives the Lagrange theorem,
and coincides with Goodner's result in the case of continuous function with finite Dini
derivatives [22, Theorem 2].

Owing to further developments we reformulate the statement of the above theorem

for lower semicontinuous functions as follows.

COROLLARY 1.9. L e t f ( t) be l o w e r semicontinuous on the closed interval [ a, b].

Then there exist t I E [ a, b) an t 2 E (a,b) such that

.j >_ and D-y 0.) -

It should be mentioned that in contrast with the classical mean value theorem (for dif-
ferentiable functions) in case of semicontinuous functions it is no longer possible to think
of t I and t 2 as intermediate values of [ a, b]. The following example demonstrates
this fact.

11
EXAMPLE 2. Let us consider the following lower semicontinuous function:

f ( t) =
t-r~ if
if
o<,<1,
t = O,
t=l.

I t i s e a s y t o c h e c k t h a t 4 = 1, D + f ( t ) = 0 forM1 t E ( 0 , 1 ) and D+f(O) = +oo.

So the ordypoint satisfying D + f ( t) >_ 4 is t = 0 .

B.M. Glover has made an attempt so that t I is an intermediate value, by' impos-
ing an additional property on f ( t ) . He has stated [21, Remark 3.2] that if t h e lower
semicontinuous function f ( t ) assumes a maximum over [ a, b] then there exists an
intermediate value t I E (a,b) such that D + f ( t l ) >_ a . Example 2 is a counterex-
ample for Glover's statement. The next theorem provides a possible way to transform
the Glover's assertion into a true statement. For this it is enough to impose the G l o v e r ' s
condition on the auxiliary function h ( t ) , instead of f ( t ) , where h(t) is the function
defined in (7).

THEOREM 1.10. Let the function f ( t) be defined on the closed interval [ a, b] .

i) I f f ( t) is lower semicontinuous and h( t) assumes a maximum o v e r [ a, b]
then there exist interme&'ate values tl , t 2 E ( a, b) such that

D . f (t,) > 4 and m-/(t1)<_4.

ii) I f f ( t) is upper semicontinuous and h( t) assumes a minimum o v e r [ a, b]

then there exist intermediate values t~ , t 2 E ( a, b) such that

D+f(t,)<_4 and D_f(t2) > 4.

Proof.. i) First we prove that h(t) can neither be strictly increasing nor strictly de-
creasing on (a, b).
Suppose for contradiction that h(t) is strictly decreasing on (a, b). Furthermore
h(t) is also lowersemicontinuous on [ a,b] and h(t) > h( a) = h(b) for all t E
E ( a , b ) . From these facts it follows that h(t) does not assume a maximal value on
[a, b]. This contradiction proves that h(t) cannot be strictly decreasing on (a, b).
One can prove by the same arguments that h(t) can neither be strictly increasing on
(a, b).
It follows that if h(t) is either nonincreasing or nondecreasing on (a, b), t h e n there
exists a point t o E (a, b) and a neighborhood of it where h(t) is constant. In this
case D+ h ( t 0) = D - h ( t 0) = 0 and thus the statement of the above theorem holds for
t ! = t o and t 2 = t o .

12
 At last assume now that h(t) fails to be monotone on ( a , b ) . Then there exist
s l , s 2 E ( a , b ) , a I < s 2 and Px,P2 E ( a , b ) , p t < P2 such that

h (sl) < h (s2) and h (Pl) > h (P2).

By Corollary 1.9 there exist t I E [ s I , s 2) and t 2 E (Pl ,P2] such that

D+h(tl)= D+f(tl)-ot> - h(s2)-h(sl) >0 and

8 2 - - 81

h (P2) - h (Pl)
D - h (t2) = D - f (t2) - a < < 0,
P2 - Pl

so the statement of the theorem is satisfied with t I and t 2 . The proofofii) is analogous
to the proof of i). Q.E.D.

The following theorem is an immediate consequence of the above theorem, however

it can also be deduced from Theorem 1.7, applied to the function h ( t ) .

THEOREM 1.11. Let the function f ( t ) be delincd and continuous on the closed interval
[ a, b] . Then there exist tl , t2 , t3 , t 4 E ( a, b) such that

D+f(tl)>a , D+f (t2) < a, D_f(t3)>c~ , D - f (t4) < c~ (8)

where the quantity ct is defined in (7).

It is worth mentioning that an extension of the Cauchy mean value theorem and the
Taylor-formula, involving Dini derivatives, has been elaborated by E. Castagnoli in [5].

Monotonicity and Dini derivatives

The differential calculus provides a very efficient tool to describe the increasing or
decreasing behaviour of differentiable functions. For nondifferemiable functions similar
characterizations are available using the Dini derivatives.
It is quite obvious that if f ( t ) is nonincreasing on (a, b) then for all t E (a, b)

D+ f ( O < 0 and D-f(t) < 0

and if f ( t ) is nondecreasing on ( a, b) then for all t E ( a, b)

D + f ( t ) >>0 and D f(t) >0.

Now we prove sufficient conditions of (strict) monotonicity for special classes of

functions. The following theorem is an immediate consequence of Lemma 1.3 and 1.4.

13
THEOREM 1.12. [5, Lemma 1] Let f ( t) be dcfincd on ( a, b) .
i) l f D+f(t) > 0 and D _ f ( t ) > 0 forMl t E (a,b) then f ( t ) isstrictly
increasing on ( a, b).
it') If D+ f ( t ) < 0 and D - f ( t ) < 0 forall t E (a,b) then f ( t ) isstn'ctly
decreasing on ( a, b).

The following example shows that e.g. condition D+ f ( t ) > 0 alone does not ensure
the strict monotonicity of the related function.

EXAMPLE 3. Let us consider the following function:

t if t E [0,1)
f(t)= t-1 if tel1,21.

It is easy to check that D+f(t) > 0 on ( 0 , 2 ) , however, f ( t ) fails to be strictly

increasing on ( O, 2).

The following theorem shows that for semicontinuous functions the nonnegativity
(nonpositivity) of certain Dini derivatives ensures the monotonocity.

THEOREM 1.13. [ 15, Corollary 4 and 5] Let f ( t ) be det~ned on [ a, b].

i) Let f ( t) be lower semicontinuous on [ a, b] . f f D - f ( t) >_ 0 ( D - f ( t) > O)
for ail t E ( a, b] then f ( t) is nondecreasing (stn'ctly increasing) on ( a, b] .
f f D+f(t) <_ 0 ( D + f ( t ) < O) forall t E [a,b) then f ( t ) isnonincreasing
(stn'ctly decreasing) on [ a, b).
ii) Let f ( t) be upperscmicontinuous on [ a, b] . f f D+ f ( t) >_0 ( D+ f ( t) > 0)
for all t E [ a, b) then f ( t) is nondecreasing (stn'ctly increasing) on [ a, b).
It' D _ f ( t ) < 0 ( D _ f ( t ) < O) forall t E (a,b] then f ( t ) isnonincreasing
(stn'ctly decreasing) on ( a, b].

ProOf. The proof makes uses of the Generalized Mean Value Theorem (Theorem
1.8). We provide a proof of the first statement in i). The other statements can be proved
similarly.

Assume that f ( t ) is lower semicontinuous and condition D - f ( t ) >__0 ( D - f ( t ) >

> 0) holds for all t E (a, b]. Assume for contradiction that f ( t ) is not nonincreasing
(strictly decreasing). Then there exist s l , s 2 E (a,b] suchthat a I < s 2 and f ( s l ) >
~> f ( s 2 ) ( f ( a l ) > f ( 8 2 ) ) . By Corollary 1.9 there exists t I E (sl, 82 ] SUCh that

n-/(t,) < / I < o o).

8 2 ~ 81

14
 This contradiction proves the first statement of i). Q.E.D.

The following theorem shows that for continuous functions the nonnegativeness
(nonpositiveness) of any of the Dini derivatives ensures monotonicity.

THEOREM 1.14. let f ( t) be defined and continuous on the open intervM ( a, b). If for
every t E ( a, b) any of the following conditions holds, then f ( t) is a nondecreasing
(strictly increasing) function on ( a, b) :

D+f(t) > 0 (D+f(t) > 0 ) , (9a)

D+f(t) >_0 (D+f(t) > 0 ) , (9b)

D-f(t) > 0 (D-f(t) > 0), (9c)

D_f(t) > 0 (D f ( t ) > 0 ) , (9d)

Proof. The statement of the above theoremis an immediate consequence of Theorem

1.13; it can also be deduced from Theorem 1. I 1 by means of indirect arguments.Q.E.D.

Applying the above theorem to the function s(t) = - f ( t ) and taking into account
the identities (3), one can characterize nonincreasing and strictly decreasing functions
via the Dini derivatives.
Similar results can be found in [33, 36, 39, 40, 41].

Lipsclfftz-contl"nuity
It is obvious that Lipschitz continuous functions have finite and bounded Dini deriva-
tives. Now we prove the converse of this statement.

LEMMA 1.15. let f ( t) be defined and continuous on ( a , b ) .

i) I f any of the Dini derivatives D +f ( t) , D+ f ( t ) , D - f ( t ) , D _ f ( t) is bounded
from above with upper bound K, then for every tl , t 2 E ( a, b) it is

: (t~) - : (t,) <_ K (t~ - t,).

ii) I f any of the Dini derivatives is bounded from below with lower bound k, then
forevery tl,t2 E ( a,b) it is

/ (t~) - / (t,) > k (t~ - t , ) .

15
 Proof. Assume that D +f ( t ) < K for all t E ( a, b). Consider the auxiliary function
g(t) = f ( t ) - K t . Since D+g(t) = D + f ( t ) - K < 0 for all t E ( a , b ) , therefore
by Theorem 1.14 g(t) is nonincreasing, that is a < t I < t 2 < b implies

: (t,) - Kt, = 9 (t,) _ 9 (t2) = : (t~) - Kt~,

which is just the thesis of i); ii) is proved in a similar way. Q.E.D.

As a corollary we can formulate the following statement.

THEOREM 1.16. [36] Let f ( t ) be detined and continuous on ( a, b). I f some o f its Dini
derivatives is bounded on ( a, b) then f ( t) is Lipsclu'tz continuous on ( a, b) .

Proos Assume, e.g., that

ID+f(t) l <_ K for every t E (a, b). (10)

By the previous results, (10) implies that

- K (t2 - t , ) _ f(t~) - f(t,) _< K (t~ - q )

for every t l, t 2 E (a, b), which means that f ( t ) is Lipschitz continuous on (a, b)
with Lipschitz constant K . Q.E.D.

First and second order optimality conditions

The following first order necessary optimality conditions in terms of Dini derivatives
are immediate consequences of the definitions.

THEOREM 1.17. lct f ( t ) bedcfinedon (a,b) andlet t o E (a,b) .

i) I f f ( t) assumes a local minimum at to then

D.f(to) >0 and D-f (to) <_0.

ii) I f f ( t) assumes a local maximum at t o then

D_f (to) _> 0 and o§ _< 0

Now we present a generalization of the classical second order sufficient optimality

conditions. This result is a modified version of the one elaboreted by E. Castagnoli and
based on the concept of second order Dini derivatives [5, Teorema 8].

16
 Second order Dini derivatives are defined in a very natural and usual way: they are
the Dini derivatives of the Dini derivative functions.
The symbol D§ D§ f ( t 0), for instance, denotes the lower right Dini derivative of the
function D § at t o . For the existence (which is not equivalent to the finiteness!)
of D§ o) the only requirement is the finiteness of D § is some fight hand
side neighborhood [to,t" ) of t o . It follows that the existence of D§247 does
not imply the continuity of f ( t ) at t o . Moreover obviously the function f ( t ) may
have at most 16 second order Dini derivatives at a given point.

THEOREM 1.18. Let f ( t ) bedefinedon (a,b) andlet t o E (a,b) .

a) Let f ( t) be upper semicontinuous on ( a, b). Then f ( t) attains a stn'ct local
minimum at t o if tbe following conditions hold:
(rai) D.f(to) > 0 and D § 2 4 7 (to) > O,
(mii) D - f (to) < 0 and D D - f (to) > 0 .
b) Let f ( t) be lower semicontinuous on ( a, b) . Then f ( t ) attains a stn'ct local
maximum at t o if the following conditions hold:
(Mi) D_f(to) >0 and D+D_f(to) <0,
(Mii) D§ <0 and D-D§

Proof. a): Condition D+D+f(t o) > 0 ensures the existence of t" E (to, b) such
that for each t o < t < t" one has

0 <_ D+f(to) < D § < D§

Since f ( t ) is upper semicontinuous therefore by Theorem 1.13 it follows that f ( t )
is strictly increasing on (t0, t " ) .
Similarly condition D _ D - f ( t o) > 0 ensures the existence of t' E ( a , t 0 ) such
that for each t' < t < t o one has

D _ f ( t ) <_ D - f ( t ) < D - f (to) <_ O.

By Theorem 1.13 it follows that f ( t ) is strictly decreasing on (t t, t o ) . Since 0 <
< D+f(to) < +co and - c o < D - f ( t o) < 0 therefore conditions
f ( t o ) = i n f { f ( t ) : t E (to,t")} and f ( t o ) = i n f { f ( t ) : t (5 ( t ' , t o ) }
should hold, which means that f ( t ) attains a strict local minimum at t o (and that f ( t )
is continuous at t o ).
The proof of part b)can be accomplished by the same resoning. Q.E.D.

The following example shows that if we drop the semicontinuity condition imposed
on f ( t ) in the above theorem, then conditions (mi) and (mii) are no longer sufficient
for strict local optimality.

17
 EXAMPLE 4. Let

it) = ( t - l / n - l/r~
0
- t - 1 / n - llr~
if
ir
if
lln<_t<ll(n-1)
t=0
-ll(n- 1) < t < - I / n ,

for n = 1 , 2 , . . . This function fails to be upper semicontinuous or any neighborhood

o f t o = 0 . An easy calculation shows that

D§ for all rE(0,1)

D§247
D§ =0 S
D-f(t)=l for all tE(0,1)'~
D D-f(O) +oo.
D § f(O) = 0 S
Note that conditions (mi) and (mii) hold for f ( t ) at t o = 0 but this point fails m
be a local minimum point of f ( t ) .
Now we formulate a variant of the previous theorem, the proof of which can be ac-
complished by the same arguments.

THEOREM 1.19. Let f ( t) be detined on (a,b) and let t o E ( a , b ) .

a) Let f ( t) be upper semicontlnuous on ( a, b). Then f ( t) attains a strict local
minimum at t o i f the following conditions hold:

(si) D§ >0 and D§247 < O,

(sii) D_f(to) < 0 and D _ V _ y ( t o ) > O.

b) Let f ( t) be lower semicontinuous on ( a, b). Then f ( t) attains a strict local

maximum at t o i f the following conditions hold:

(si) D-f(to) 20
and D+D-f(to) <0,
(Sii) D+y(to) <_0 and D_D+y(t0) <0.

REMARK. Mixing the above conditions one gets further second order suffcient optimal-
ity criteria. For example for upper semicontinuous functions any of the following pairs

{(mi),m(ii)}, {(si),(raii)}, {(mi),(sii)}, {(si),(sii))

is sufficient for t o to be a strict local minimum point of f ( t ) .

18
2. Dini derivatives of functions of several variables
Let the function f ( x ) be defined on the open set S C R n, where n > 1. Let
a E S . There are several ways to extend the notion of the Dini derivatives for the case
of several variables.

Radial Dini derivatives

A possible extension is to apply the original concept of the Dini derivative radially,
which means that this notion is applied for the single variable functions of the following
type
s( t) = f ( a + td) , t E R; a, d E Rn; a + td E S.
This approach leads to the concept of the radial Dini derivatives as follows:

D+ f ( a; d) = lim sup f ( +td) - f ( a) ,

t..-,O + t

D+f(a; d) = lim inf f ( a + tar) - f ( a )

t~O + t '

f ( a + td) - f ( a )
D - f ( a ; d) = lim sup t J
t~O -

f ( a + td) - f ( a)
D _ f ( a ; d) = lim inf
t--,O - t

The first two derivatives, D + f ( a ; d) and D + f ( a ; d), are usually called as the upper
and lower directional Dini derivatives of f ( x ) at a .
Since for each d E R" one has

D-f(a;d) = -D+f(a;-d) and D_f(a;d) = -D+f(a;-d) (11)

therefore it is quite obvious to deal only with the directional Dini derivatives D + f ( a; d)
and D+ f ( a ; d ) .

The Dini-Hadamard derivatives

Another extension of the Dini derivative concept is the Dini-Hadamard derivative.
We define the upper and lower directional Dini-Hadamard derivatives of f ( x ) at a
along the direction d by

d + f ( a ; d ) = lim sup f ( a + tu) - f ( a ) ,

l..d) § t
u-'-~

d+f(a; d) = lim inf f ( a + tu) - f ( a )

1"40+ t
u-.-.d

The following example shows that the radial Dini derivative and the Dini-Hadamard
derivative can be different.

19
EXAMPLE 5. Let

f (:rl z2 ) -- ( 0 if a; 2 = 0
' 5 1 + Z2 if :2172 r 0.

Let a = ( 0 , 0 ) and d = el 9 An easy calculation shows that

d+.f (a; e,) = 1 and D+f (a; e,) = O.

Elcmentary properties
The following statements are immediate consequences of the definitions.

d + [ - f l ( a ; d) = -d.f(a; d), d + [ - f ] ( a ; d) = -d§ d),

(12)
d+f(a; d) < D+f(a; d) < D+f(a; d) < d+f(a; d).

THEOREM 2.1. d+f( a; d)( d§f( a; d) ) is a lower (upper) semicontinuous function of

the direction d E R " .

Proof. First we show that d+f(a; d) is lower semicontinuous in d, which is equiv-

alent to the following property: for all c E R" the upper level set

M = {die < d.f(a; d)}

is open. By definition
d+f(a; d) = lim sd(c~,r),
r--~ +

where
f ( a + tu) - f ( a)
aa(o~, r) = inf
I[u-*qt<~ ~;
o<l<r

From the nondecreasing behaviour of so( o~, r) it follows that

sd(a,r) < d+f(a; d)

for any positive numbers o~ and r .

Assume that M 51 0 and let d E M be arbitrary. Taking into account the above
properties of the function sd(c~, r) it follows that there exist positive numers o~0,r 0
such that
c < 8d (60, r0).

20
 Let G denote the open ball around d with radius a0 9 Now we show that G C M
proving that M is open set.
Let z E G be arbitrarily chosen and consider a neighbourhood N of z, with radius
/~, contained in G . From the inclusion N C G it follows that

c< s~ (ao,~o) --_s. (~,~o) -< d§ z),

which proves that z E M .
The upper semicontinuity of d+ f ( a ; d) can be proved by the same reasoning.Q.E.D.

It is worth pointing to that condition

D§ d) = D+ f(a; d) (13)
is equivalent to the existence of the limit

f'( a; CO = t--.o.lim f ( a + td)t - f( a) , (14)

which is the well known directional derivative of f ( z ) at a .

Similarly the condition
d+f( a; d) = d* f ( a; d)
is equivalent to the existence of the limit

dr(a; d) = lim
,-.o. f ( a + tu)t - f ( a) ' (15)
u,...d

which is called the uniform dimctionM derivative of f ( x ) at a.

If the (uniform) directional derivative exists for every d E R" then the function
f ( x ) is called (uniformly) directionally differentiable at a .
The function f ( z ) is called (uniformly) G~teaux differendable at a ifthe (uniform)
directional derivative f'(a; d) ( dr( a; cO) exists and is a linear function of the direction
dE R".
The following example shows that the G~teaux differentiability does not imply the
Frdchet ditTerentiability.

EXAMPLE 6. Let us consider the following Peano-type function:

= ~'0, if y=O or lYl>x2

f(x,y)
11 otherwhise.
Let a = ( 0 , 0 ) . It is easy to check that f ' ( a; d) = 0 for all d E R n and thus f ( x, y)
is Gfiteaux differentiable at the origin. But f ( z , y) fails to be continuous at ( 0 , 0 ) ,
therefore it is not Fr6chet differentiable at the origin.
The following theorem shows that uniform G~teaux differentiability is equivalent to
the Frgchet differentiability.

21
THEOREM 2.2. I f them exists g E R" such that the condiSon

d . f ( a; d) = d* f ( a; d) = (g,d) (16)

holds [or every d E R n, then f ( x) is Frdchet di[ferentiable at a . (Here < .,. >
denotes the inner product on R n .)

Proof. The statement follows from the fact that the Frrchet differentiability

lira f ( x ) - f ( a ) - (g, x - a) = 0
=-'* - all

is equivalent to the following conditoin (which is equivalem m (t6)): for all d E R '~

lira f ( a + tu) - f ( a ) = ( g , g ) . Q . E . D .
t "*O § t
i*..*d

It follows that the upper (lower) directional Dini derivative could not coincide with its
Dini-Hadamard counterpart, however the following statement holds tree.

THEOREM 2.3. I f f ( z) is locally Lipschitz around a, then for every d E R"

d+ f ( a ; d) = D* f ( a ; d) and d . f ( a ; d) = D . f ( a ; d).

Proof. Let f ( z ) be Lipschitz continuous on G, a neighbourhood of a, with Lips-

chitz constant L . Let d E R" be arbitrary. Than there exist /5 > 0 and r > 0 such
that foreach u E R * and t E R satisfying conditions I[ u - d I1< 8 and 0 < t < r,
one has a + td, a + tu E G ; consequently

I f ( a + tu) - f ( a + td)l S Lltl. II u - d II,

which gives the following relation:

l i m f ( a + td) - / ( a + tu) = O.
=--.o* t
=-..4

Applying this fact and the properties of the limsup operation we get the desired result:

/ ( a + tu) - / ( a )
d*f(a; d ) = lim sup
i-.0 §
e---d

= lim sup f ( a + td) - f ( a ) + lira sup f ( a + tu) - f ( a + td) = D + f ( a ; cO.

,...o 9 t ,-.o. t
=-.~ u-,4

22
The other formula can be proved in a same way. Q.E.D.

It follows that for locally Lipschitz functions the uniform directional differentiability
coincides with the directional differentiability. Consequently if the locally Lipschitz
function f ( x ) is in addition G~teaux differentiable, then f ( ~ ) is Fr6chet differentiable
as well. It is obvious tha if f ( z ) is Fr6chet differentiable at a, then the Dini derivatives
at a are all finite for every direction d E -R'~ 9 The converse of this statement is not true
but we have the following very interesting result due to S.G. Slobodnik.

THEOREM 2.4. [42] Let M C R" . If the Dini derivatives D§ f ( z; d) , D +f( z; d) are
finite for any x E M, and for any direction d E R", then f ( x ) is Frdchet differentiable
Mmost everywhere on M .

The lack of the gradient in the nondifferentiable case usually leads to the introduction
of some kinds of subdifferentials. Following J.P. Penot [32] let us introduce the following
two sets
P+f(a) = {g e R '~ : (g,d) < d+f(a;d)},
P+ y(a) = {g E R" : (g,d) > d+f ( a ; d ) } ,

which are called the Dini Subdifferential and the Dini superdifferentiM of f ( x ) at a,
respectively. It is not difficult to show that these sets can also be described as follows:

P+f( a) = {g E R" : lim inf f ( z) - f ii( za)_ -al(g'z

[ - >0},

P+ f ( a ) = { 9 E R~ : limsup f ( z ) - f (, ,az)- a , , <0}

It follows that
P§ n P§162 r
iff f ( ~ ) is Frfchet differentiable at a, when

P+ :(a) n P§ = (f'(a)),

where f'(a) denotes the gradient of f ( z ) at a .

Apart from obvious smooth and convex cases, it is rather usual that P § = O at
some point. (The same is of course true for Dini superdifferential.) The following two
propositions describe some elementary but very important properties of Dini subdiffer-
entials. In particular, the first one shows how useful the Dini sub- and superdifferential
is in characterizing local minima and local maxima.

23
THEOREM 2.5. [14, 24, 31]
a) I f f ( x) attains a local minimum at a, then the following conditions hold:

0 E P+f(a), (17)

d+/(a; d) > 0 forall dER". (18)

b) Conversely, frO E P+f ( a) , then for nny r > 0 the function f ( z) + rl lx - all
attains a stn'ct local minimum at a .
c) finny o f the following conditions holds then f ( x ) attains a strict local m inimum
at a :
0 E int P § (19)

d§ >0 forall d E R*, d ~ O . (20)

Proof. Assume that f ( x ) attains a lcoal minimum at a . Then it is obvious that

condition (18) has to be ful filled. Since conditions (17) and (18) are equivalent, therefore
a) has been proved.
To prove b), assume ab absurdo that f ( x ) + rllz - all does not attain a strict local
minimum at a . Then there exists a sequence x,, ---, a, x,, ~ a such that f ( x , ) +
§ -all _< f(~). Set t~ = Ilxn -all, ~ -- t g a ( x ~ - a ) . Then II"~ll-- 1 andwe
can assume that u, converges to some d, IId[I -- 1 . It follows that

d§ a; d) < lim inf t ; 1 ( f (a + t , u , ) - f ( a) ) =

= lim inf t ; l ( f ( x n ) - f ( a ) ) < - r < O,

in contradiction with the assumption that 0 E P§ a).

Statement c) can be proved by adopting the same reasoning applied to prove part b),
taking into account that conditions (19) and (20) are in fact equivalent, due to the lower
semicontinuity of d§ f ( a ; d). Q.E.D.

REMARKS. i) Applying Theorem 2.3 and 2.5 one can deduce the following result of A.
Ben-Tal and J. Zowe [45] from the above theorem: if f ( x ) is Lipschitz continuous in
a neighborhood of a and f ' ( a ; d ) > 0 foreach d E R", d S / 0 , then f ( x ) attains a
strict local minimum at a . In [45] an example is given, where the Lipschitz continuity
cannot be weakende to just continuity, of course if the directional derivative f ' ( a; d),
which is a radial concept, is used.
ii) Obviously, analogous statements can be proved for maximization problems, as
well.

24
THEOREM 2.6. [17, 24] Let f ( z ) be a lower semicontinuous function defined on R" 9
Then P+ f ( z ) ~r D on a dense subset o f dom f .

Proos Let a 9 d o m f , let r > 0 and set g(x) = ( r - IIz - all~) -' 9 If r is
so small that f ( x ) is bounded from below on the neighbourhood of a with radius r,
then f ( x ) + g ( z ) attains its minimal value on this neighbourhood at some point u
(because g(x) ---, oc as [[z - a[[ ---* r ). Since g ( z ) is continuously differentiable on
its domain, it follows that d+f(u; d)+ < g'(u), d > > 0 for all d 9 R", that means
that - g ' ( u) 9 P+f( u) . Q.E.D.

The Dini-Hadamard derivatives are also called contingent derivafives since there is a
well known relationship between the lower Dini-Hadamard derivatives und the so called
contingent cone and between the upper Dini-Hadamard derivatives and the so called
cone of interior displacements. For the details we refer to [3, 18, 24]. Let E C R " and
e E c I E . (~(cl >>is the sign for the closure operation).
The contingent cone K ( E , e) to E at e E cIE is given by

K ( E , e ) = {v e R'~ : 3 (vi) ,-q (ti) such that v i---, v,ti ---* O+e + t,vi 9 E ) .

If f ( x ) is a function defined on R" and E = e p i f (the epigraph of f ( x ) ) , e =

= ( a , f ( a ) ) then
( d , a ) 9 K ( E , e ) r o~ >_ d §
It follows that the epigraph of the Dini-Hadamard lower derivative d§ d) co-
incides with the contingent cone to E = e p i f at e = ( a , f ( a ) ) . Furthermore the
following relation holds:

P+f(a) = (g 9 R" : ( g , - 1 ) 9 K * ( E , e ) } ,

where K ' ( E , e) denotes the polar cone o f K ( E , e ) .

The cone o f interior displacements to E at e 9 is given by

I t f ( E, e) = R " \ If( R "~ \ E, e).

It follows from this definition that v 9 I K ( E, e) iff for any sequences (vi) , ti)
such that v i ---. v and t i ~ 0* one has e + t i v i 9 E for i large enough. One can
prove that if E = epi f and e = ( a, f ( a) ) then

(d,a) 9 r c~ > d + f ( a ; d ) .

It follows that the strict epigraph of the Dini-Hadamard upper derivative d + f ( a ; d)

coincides with the cone of interior displacements to E = e p i f at the point e =
= ( a, f ( a ) ) . Furthermore the following relation holds:

P+f(a) = (g 9 R" : ( - g , 1) 9 I K ' ( E , e ) } .

25
 The contingent cone K ( E , e) allows to extend the optimality conditions (18) and
(20) for a relative local minimum. By adopting the same reasoning applied to prove
Theorem 2.5 the following statements can be proved.

THEOREM 2.7. L e t f ( x) be deigned on a set ~ C R* and let E C t) , E not

necezsadly open.
i) I f a E E is a point o f looM minimum o f f ( x) on E, then

D + f ( a ; d) >_ 0,Vd E K ( E , a ) .

ii) I r a E E and it is

d§ >O, VdE K(E,a), dqO,

then a is a point o f stn'ct local minimum o f f ( z) on E .

The following two results are needed for our further investigations.

Implicit function theorem

Let f ( z ) be defined on the open set S C R ~ 9 We introduce the following notations.
Consider a direction d E R '~, d ~ 0 . Denote by L the n - 1 dimensional subspace
orthogonal to d. Then every x E R ~ has a unique representation as x = u + v d , where
u E h and v E R . We may regard u as an element of R '~-1 and we may identify x
with ( u, v) .

THEOREM 2.8. [27, Theorem 3] Let f ( x) be continuous on the open set S . Let
D +f ( a ; d) < 0 for a certain a = ( Uo, v o) E 8 and let D § f ( z; d) be upper semicon-
tinuous in z at a . Then there exist convex neighbourhoods G and N o f a and u o
respectively and a unique function h,,d( u) delined on N such that:
0 x=(u,v) EG andf(z)=f(a)c~uEN andh,,a(u)=v,
ii) x = ( u, v) E G and f ( x) < f ( a ) v~uEN andh~,~(u) < v ,
iii) z = ( u, v) E G and f ( x) > f ( a) c~ u E N and h~.a( u) > v .

Proof. From the upper semicontinuity of D § d) at a it follows that there exists

anopenball B around a such that D § < 0 foreach z E B .
Choose a positive number ~ such that zl, x 2 E B, where

:-- . - (.o, % - and :-- o § Z d - - ("o, % + a ) .

Define z ( t ) := a + td and define s(t) := f ( x ( t ) ) for t E [0,2/~] . Since

D+a(t) = D + f ( z ( t ) ; d) < 0 therefore by Theorem 1.14 a(t) is strictly decreasing
on ( 0 , 2 # ) . Consequently f ( z I) > f ( a ) > f ( x 2) .

26
 From the continuity of f ( z ) at z I and x 2, it follows that there exists a convex
neighbourhood N of u 0 (an open ball around u 0 ) such that for every u 6 N we
have
f ( ( u, v0 - 8) ) > f ( a ) > f ( ( u, vo + 8))
and
G : N • (~o - Z, ~o + ~) C s.
Let x = ( u , v ) 6 G . Then u 6 N and v 6 ( % - / ~ , v 0 + 8 ) . Define z =
= ( u , v o - f l ) . Repeating our former reasoning, we find that f ( z + td) is a strictly
decreasing function on ( 0 , 2 / ~ ) . Furthermore, since f ( z + tar) is continuous in t, the
equation f ( a ) = f ( z + td) has a unique solution in [ 0 , 2 B], which we denote by
h~ u) 9
Statements i)-ii) are immediate consequences of the strictly decreasing property of
the function f ( z + td) . Q.E.D.

REMARK. The conclusion of the above theorem remains valid if we replace D + f ( x; d)

with D + f ( x ; d) and a similar implicit function theorem can be proved for the case
when D + f ( a ; d) > 0 and D § f ( z ; d) is lower semicontinuous in z at a or when
D + f ( a; d) > 0 and D +f( z; d) is lower semicontinuous in z at a.

Dini derivatives o f max-functions

LEMMA 2.9. Let the functions h i ( z ) , j = 1 , 2 , . . . , k be radially continuous at a 6

6 R " . Let
h(z) = max{hi(z) : ] = 1,...,k}.

Define
I(z) = (]: h(z) = h/z)}.

Then
D§ d) = rnax{D§ d) : ] 6 I ( a ) }

and
D§ = max { D § hi( a; d) : 1" 6 I ( a ) } .

Proof. Let j 6 I ( a ) be arbitrary. Then for small t > 0 one has

h i ( a + td) - h i ( a ) <_ h( a + td) - h( a).

It follows that
(
max l D * h i ( a ; d ) : j 6 I ( a ) ~ _< D § d) (21a)

27
 and
max{D+hi(a;d) : j E I(a)} < D+h(a;d). (21b)

On the other hand for all ] E I ( a ) and i ~ I ( a ) we have the following inequality

hi(a) < h(a) = h i ( a ) .

Since h i and h i ( z ) are radially continuous at a, therefore there exists T > 0

such that for all 0 < t < T and for all ] E I ( a ) and i r I ( a )

h i ( a + td) < h i ( a + td).

It means that for all 0 < t < T we have that I ( a + tar) C /'( a ) . Without loss of
generality we may assume that there exists and index p such that p E I ( a + t d ) for all
0 < t < T . It follows that h(a + td) coincides with hp(a + tar) for all 0 < t < T,
consequently D+ h( a; d) = D§ hp( a, d) and D § h( a; d) = D § hp( a; d) which proves
that the inequalities (21a) and (21b) are always satisfied with the equality sign.

Acknowledgement
The authors thank an anonymous refers for many useful suggestions and improve-
ments. Obviously the usual caveat applies.

REFERENCES

[1] KJ. ARROW,A.C. EENTHOVEN,Quasicuncave programming, Econometrica 29, 1961,779-

800.
[2] M. AVRIEL,W.E. DIEWERT,S. SCHAIBLE,W.T. ZIEMBA,Generalized concavity, Plenum Press,
New York, 1987.
[3] M.S. BAZARAA,J.J. GOODE,M.Z. NAStiED, On the cone o f tangents with applicatio~ to math-
ernaticalprogramming, J.O.T.A. 13, 1974, 389-426.
[4] J. BORDE,J.E CROUZEIX,Continuity properties o f the normal cone to the level sets o f a quasi-
convex function, J.O.T.A. 66, 1990, 415-429.
[5] E. CASTAGNOLI,Estensioni di alcuni classici teoremi dell 'analisi al caso di funzioni non deri-
vabili, Ist. Mat. Fin. Univ. di Parma, serie II, 11, 1983.
[6] M.G. CINQUINI,G. GtORt31,Funzioni convesse generalizzate e loro impiego in prograrnmazione
matematica, Pubblicazioni, Universit~ degli studi di Pavia, Facolt~ di Economia e Commercio,
8, 1988/89.
[7] EH. CLARKE,Optimization and nonsmooth anaysis, John Wiley, New York, 1983.
[8] J.P. CROUZEIX, Some differentiabilityproperties o f quasiconvex functions on R ~ , in Optimiza-
tion and Optimal Control, A. Auslender, W. Oetfli, J. Stoer (eds.). Proceedings of a conference
held at Oberwolfach, March 16/22/1980, Springer Vedag, Berlin, 1981, 9-20.
[9] J.P. CROUZE , Continuity and differentiability properties of quasiconvex functiona on R ~ ,
in Generalized concavity in optimization and economics, S. Schaible, W.T. Ziemba (eds.).
Academic Press, New York, 1981, 109-130.

28
[10] J.P. CROUZEIX,About differendability o f order one o f quasiconvex functions on R " , J.O.T.A.
36, 1982, 367-385.
[ 11] J.P. CROUZEIX,J.A. FERALND,Criteria for quasiconvexity andpseudoconvexity: relationships
and comparisons, Mathematical Programming 23, 1982, 193-205.
[12] A. DEmOY, M6moire sur les hombres ddrivds des foncdons continues, Journal de Math. 1,
1915, 105-240.
[ 13] V.E DEMYANOV,L.V. VASILtEV,Nondifferentiable Optimization, Optimization Software, Inc.,
New York, 1985.
[14] V.F. DEMYANOV,Nonsmooth analysis and directional derivatives, Working Paper, Universit~
di Pisa, Dipartimento di Matemafica, 163, 1989.
[15] W.E. DmVCEaT,Alternative characterizations o f six kind o f quasiconcavity in the nondifferen-
tiable case with applications to nonsmooth programming, in Generalized Concavity in Opti-
mization and Economics, S. Schaible, W.T. Ziemba (eds.). Academic Press, New York, 1981.
[16] U. DINI, Grundlagen Fureine Theorie der Funktionen einer veranderlichen reellen Gr6sse, B.E.
Teubner, 1892. New Italian edition: Fondamend per la teorica delle funzioni di variabili reali,
Pitagora, 1990.
[17] I. EYd~LAND,G. LEBOURG,Generic Fr6chet differentiability and perturbed optimization prob-
lems, Trans. Amer. Math. Soc. 224, 1976, 193-215.
[18] K.H. ELSTER,J. THIER~LDER,Abstract cone approximation and generalized differendability
in nonsmooth optimization, Optimization 19, 1988, 315-341.
[19] G. GIORGI,Quasiconvex programming revisited, Calcolo 21, 1984, 307-316.
[20] G. GIORGI,A note on quasiconvex functions that are pseudeconvex, Trabajos de Investigacion
Oper. 2, 1987, 80-83.
[21 ] B.M. GLOVER,Generalized convexity in nondifferentiableprogramming, Bull. Austral Math.
Soc. 30, 1984, 193-218.
[22] D.B. GOOD,R, Mean value theorems for functions with finite derivatives, Amer. Math. Mon-
thly 67, 1960, 852-855.
[23] L.M. GRAVES,The theory of functions of real variables, McGraw-Hill Co., New York, 1956.
[24] A.D. IOFFE, Calculus of Dini subdifferentials o f functions and contingent coderivatives o f set
valued maps, Nonlinear Analysis, Theory, Meth., Applic. 8, 1984, 517-539.
[25] S. KARAMARDIAN,Strictly quasiconvex (concave) functions and duality in mathematical pro-
gramming, J. Math. Analysis and Applications 20, 1967, 344-358.
[26] S. KOMLOSI,Second order conditions o f generalized convexity and local optimality in nonlin-
ear programming: the quasi-Hessian approach, Studia Oec. P6cs, 1985.
[27] S. KOMLOSl,Some properties of nondifferentiable psuedoconvex functions, Math. Program-
ming 26, 1983, 232-237.
[28] S. KOMLOSI,On a possible generalization o f Pshenichnyi's quasidifferentiability, Optimization
21, 1990, 3-11.
[29] S. KOMLOSl,Quasiconvex first order approximations and Kuhn- Tucker type optimality condi-
tions, E.g.O.R. 65; 1993, 327-335.
[30] O.L. MANGASARIAN,Pseudoconvex functions, SIAM Journal on Control 3, 1965, 281-290.
[31] B.S. MORDUKHOVICrt,Methods o f approximations in problems of optimization and optimal
control, Nauka, Moscow, 1988, (Russian).
[32] J.E PENOT, Calcul sous-diffdrentiel et optimisation, J. Funct. Anaysis 27, 1978, 248-276.
[33] J.E PENOT, On the mean value theorems, Otimization 19, 1988, 147-156.
[34] B.N. PSHENICHNYI,Necessary conditions for extremal problems, Marcel Dekker, New York,
1971.
[35] R.T. ROCKAFELLAR,Convex analysis, Princeton University Press, Pricenton, 1970.
[36] N. ROUCHE,E HABETS,M. LALOY,Stability theory by Liapunov's direct method, Springer
Verlag, New York Inc., 1977.
[37] S. SAKS, Sur les hombres ddrivds des fonctions, Fundamenta Math. 5, 1924, 98-104.

29
[38] S. SHAIBLE,W.T. ZIE BA (r Generalized concavity in optimization and econotrJics, Aca-
demic Press, New York, 1981.
[39] L. S c ~ , Zur Theotic der stetigen Funktionen einer reellen Ver~derlichen, Acta Mathe-
matica 5, 1884, 183-194.
[40] H.P. SHEFFLER,Mean value pmpem'es of nondifferentiable functions and their application in
nonsmooth analysis, Optimization 20, 1989, 743-759.
[41] E.J. McSHANE,Intcgration, Princeton University Press, Princeton, New Jersey, 1944.
[42] S.G. SLOBODNIK,Thc continuity and differentialpmperties of functions, Dis. D. Phil. Moscow,
1966.
[43] W.H. YOUNG,G.C. YOUNG, On dcrivates and the theorem ofthc mean, Quarterly Journal of
Pure and Appl. Math. 40, 1909, 1-26.
[44] G.C. YOUNG, On infinite derivates, Quarterly Journal of Pure and Appl. Math. 47, 1916,
148-153.
[45] BEN-TAL,J. ZOWE,Directionad derivatives in nonsmooth optimization J.O.T.A. 47, 1 985, 483-
490.
[46] A. CAMBINI,E. CASTAGNOLI,L. MARTEIN,P. MAZZOLENI,S. SrtAIBLE(eds.), Generalized Con-
vexity and Fractional Programming with Economic Applications, Lecture Notes in Economics
and Mathematical Systems 345, 1990. Springer Verlag, Berlin.

RIASSUNTO

I1 presente lavoro, pubblicato in due parti, riguarda le principali propriet~ dei numeri
derivati di Dini (o derivate direzioni di Dini), sia di funzioni di una variabile c h e di pitt
variabili, nonch6 alcune loro applicazioni aUo studio della convessit~t generalizzata ed a
problemi di ottimizzazione vincolata.
Nella prima parte del lavoro si fomiscono le definizioni e le proprieth fondamentali
dei numeri derivati di Dini e vengono riformulati alcuni classici teoremi dell'analisi, con
riferimento a funzioni non differenziabili.
Nella seconda parte tali derivate direzionali vengono applicate nello studio di alcune
classi di funzioni convesse generalizzate non differenziabili e nell'ottenimento di con-
dizioni di ottimalit~ per problemi (non differenziabili) di programmazione matematica.

30