, | |
) is its topo-
logical dual space endowed with the weak
A
|x x
|
the distance of x from A and
(x, A) =
0 if x A
+ if x , A
the indicator function of A. We denote the topological interior, the
closure and the convex hull of A by int A, cl A and conv A respectively.
The set K X is a cone if tK K for each t > 0; the recession cone of
A is the set
0
+
A = v X : A + tv A, t > 0
where 0
+
= . The extendedvalued function f : X IR + is
said to be proper if domf ,= , where
domf = x X : f(x) < +
is the (eective) domain of f. We denote by c(X) the class of the
proper extendedvalued functions dened on X, by T(X) c(X) the
subclass of the lower semicontinuous functions and by L(X) T(X) the
subclass of the locally Lipschitz functions. We recall that each convex
and continuous function f : X IR is directionally dierentiable, i.e.
for each point x X and for each direction v X the limit
f
(x, v) = lim
t0
f(x + tv) f(x)
t
exists and it is nite; moreover the subdierential of f at x is
f(x) = x
: f(x
) f(x) +x
, x
x), x
X.
A mean value theorem 3
2. Mean value theorems
In the last years many nonsmooth generalizations of the classical mean
value theorem for a dierentiable function were stated via dierent di-
rectional derivatives. In this section we collect the main results. The
rst one is based on the concept of Dini derivative. Given f c(X),
the lower Dini derivative of f at x domf in the direction v X is
D
f(x, v) = liminf
t0
f(x + tv) f(x)
t
.
Theorem 2.1 (Diewert (1981)) Given a function f T(X), for each
a, b domf with a ,= b there exists x
0
[a, b) such that
D
f(x
0
, b a) f(b) f(a).
Such a result was presented in a weaker form also by Borwein and
Strojwas (1989) and by Penot (1988). An important application of
Diewerts mean value theorem is due to Komlosi (1995) that shows
how quasiconvexity, pseudoconvexity and strict pseudoconvexity of lower
semicontinuous functions can be characterized via the quasimonotonic-
ity, pseudomonotonicity and strict pseudomonotonicity of the lower Dini
derivative.
A dierent mean value theorem has been stated by Lebourg (1975) for
locally lipschitzian functions. It is well-known that for each f L(X)
the directional derivative of Clarke (1975) of f at x domf in the
direction v X
f
(x, v) = limsup
x
x,t0
f(x
+ tv) f(x
)
t
is continuous and sublinear and therefore the set
f(x) = x
: f
(x, v) x
, v), v X
is a nonempty convex compact set called Clarke subdierential of f
at x. The following mean value theorem plays a fundamental role in
nonsmooth analysis theory.
Theorem 2.2 (Lebourg (1979)) Given a function f L(X), for each
a, b X with a ,= b there exist x
0
(a, b) and x
f(x
0
) such that
x
0
, b a) = f(b) f(a).
Borwein et al (1987) proved that, under the same assumptions, it is
possible to substitute the Clarke subdierential
(x, v) = sup
uX
limsup
f(x+tu)f(x), t0
f(x + tu + tv) f(x + tu)
t
.
Both of the theorems that we have just described are in exact form:
the next one, instead, will be an approximate mean value theorem
and it is based on the directional derivative introduced by Rockafellar
(1980). Given f T(X) the ClarkeRockafellar directional derivative
of f at x domf in the direction v X is
f
(x, v) = sup
r>0
limsup
x
f
x, t0
inf
v
v+rB
f(x
+ tv
) f(x
)
t
where x
f
x means x
x and f(x
f(x) = x
: f
(x, v) x
, v), v X.
Theorem 2.3 (Zagrodny (1988)) Given a function f T(X), for
each a, b domf with a ,= b there exist x
0
[a, b) and two sequences
x
k
X with x
k
x
0
,
x
k
X
with x
f(x
k
)
such that
liminf
k+
x
k
, b a) f(b) f(a),
liminf
k+
x
k
, b x
k
)
|b x
0
|
|b a|
(f(b) f(a)).
Analogously to Diewerts mean value theorem, Zagrodnys one is crucial
in order to study the convexity of a lower semicontinuous function (see
for instance the paper of Correa et al (1992)). We end this section
describing a particular mean value theorem due to Clarke and Ledyaev
(1994) that give an estimate of the rate of growth of a function f in
multiple directions simultaneously.
Theorem 2.4 (Clarke and Ledyaev (1994)) Let C
1
, C
2
X be two
nonempty closed convex bounded sets with at least one compact and f
L(X). For each > 0 there exist x
0
conv (C
1
C
2
) and x
f(x
0
)
such that
inf
C
1
f sup
C
2
f < x
0
, x
1
x
2
) + , x
1
C
1
, x
2
C
2
.
A mean value theorem 5
If in addition both C
1
and C
2
are both compact sets, then
min
C
1
f max
C
2
f x
0
, x
1
x
2
), x
1
C
1
, x
2
C
2
.
There are some interesting applications to calculus, ow invariance, and
generalized solutions to partial dierential equations.
3. Axiomatic derivatives and subdierentials
The concept of the directional derivative and the subdierential of a
convex function were used with advantage for treating convex optimiza-
tion problems. Since more than thirty years much eort was made to
establish similar concept in the nonconvex nonsmooth case with the in-
troduction of modications of the directional derivative. In accordance
with such investigations, Elster and Thierfelder (1988) proposed an ax-
iomatic appoach for constructing generalized directional derivatives of
arbitrary functions: the basic idea is the fact that the epigraphs of the
dierent directional derivatives of a function f c(X) can be consid-
ered as a cone approximations of the epigraph of f. The rst step was
to nd an axiomatic denition of abstract local cone approximations.
Denition 3.1 A setvaled map
K : 2
X
X ;X
is said local cone approximation if to each set A X and each point
x X a cone K(A, x) is associated such that the following properties
hold:
(1) K(A, x) = K(Ax, 0),
(2) K(A (x + rB), x) = K(A, x) for each r > 0,
(3) K(A, x) = for each x , cl A,
(4) K(A, x) = X for each x int A,
(5) K((A), (x)) = (K(A, x)) for each linear homeomorphism :
X X,
(6) 0
+
A 0
+
K(A, x) for each x cl A.
The local cone approximation K is said to be isotone if for each A
1
A
2
X and for each x X we have K(A
1
, x) K(A
2
, x).
The second step was to use the concept of local cone approximation in
order to describe generalized directional derivatives.
6
Denition 3.2 Let K be a local cone approximation, f c(X) and
x domf; the positively homogeneous function
f
K
(x, ) : X IR
dened
f
K
(x, v) = inf y IR : (v, y) K(epi f, (x, f(x)))
is said Kdirectional epiderivative of f at x. We assume inf = +.
From the local property (2) of K we deduce that for each pair f
1
, f
2
c(X) of functions that coincide in a suitable neighbourhood of a point
x X then
f
K
1
(x, v) = f
K
2
(x, v), v X.
Using these two notions, general optimality conditions and duality re-
sults with respect to nonsmooth optimization problems could be de-
rived (see for instance Elster and Thierfelder (1988), Castellani and
Pappalardo (1995)). A dierent axiomatic appoach in nonsmooth opti-
mization lies in the identication of the minimal main properties of the
subdierentials. Abstract classes of subdierentials were considered by
Ioe (1984), Correa et al (1994), Thibault and Zagrodny (1995). We x
our attention on the paper of Aussel et al (1995) that introduced a class
of generalized subdierential with much less restrictive properties.
Denition 3.3 A setvalued map
: c(X) X ;X
2
(x) =
k
k
|x x
k
|
2
where
k
k
= 1,
k
0 and x
k
X
is a convergent sequence
are dierentiable. Given f T(X), for each a, b X with a domf
and a ,= b, and for each r f(b) there exist x
0
[a, b) and two sequences
x
k
X with x
k
f
x
0
,
x
k
X
with x
k
f(x
k
)
such that
liminf
k+
x
k
, b a) r f(a),
liminf
k+
x
k
, x x
k
)
|x x
0
|
|b a|
(r f(a))
for each x = x
0
+ t(b a) with t > 0.
Since for any norm the functions d
2
[a,b]
,
2
L(X), then any norm is
f(x) + (x, ) + t
k
d
[a,b]
+
1
k
2
= inf
f + t
k
d
[a,b]
+
1
k
2
f(x
0
);
in fact,
if x
k
, we have
f(x) + t
k
d
[a,b]
(x) +
1
k
2
f(x
0
)
1
k
2
+
1
k
2
= f(x
0
),
if x
k
, we have
f(x) + t
k
d
[a,b]
(x) +
1
k
2
+ t
k
r
k
+
1
k
2
f(x
0
).
Since is a closed set, applying Ekelands variational principle to the
lower semicontinuous function
f(x) + (x, ) + t
k
d
[a,b]
(x)
with =
1
k
2
and =
1
k
, there exists x
k
such that, for each x ,
we have
|x
k
x
0
| <
1
k
, (1.4)
f(x
k
) + t
k
d
[a,b]
(x
k
) f(x
0
), (1.5)
f(x
k
) + t
k
d
[a,b]
(x
k
) f(x) + t
k
d
[a,b]
(x) +
1
k
|x x
k
|. (1.6)
From (1.4) we deduce x
k
x
0
,= b, from the lower semicontinuity of
f and (1.5) we deduce f(x
k
) f(x
0
), and nally from (1.6)we deduce
that the function
k
(x) = f(x) + (x, ) + t
k
d
[a,b]
(x) +
1
k
|x x
k
|
assumes minimum at x
k
. Since x
k
int for k large enough, from
Theorem 4.1, for each v X we have
0
H
k
(x
k
, v)
K
k
(x
k
, v)
(f + (, ))
K
(x
k
, v) + t
k
d
[a,b]
(x
k
, v) +
1
k
|v|. (1.7)
A mean value theorem 11
Moreover f + (, ) and f coincide in a suitable neighbourhood of x
k
and hence (1.7) becomes
f
K
(x
k
, v) + t
k
d
[a,b]
(x
k
, v) +
1
k
|v| 0. (1.8)
Let y
k
[a, b] be such that
d
[a,b]
(x
k
) = |x
k
y
k
|.
For each x
k
d
[a,b]
(x
k
) we have
x
k
, b x
k
) d
[a,b]
(b) d
[a,b]
(x
k
) = d
[a,b]
(x
k
) 0;
moreover, since d
[a,b]
L(X) with constant 1, we have |x
k
|
1 and
then
x
k
, b y
k
) = x
k
, b x
k
) +x
k
, x
k
y
k
)
d
[a,b]
(x
k
) +|x
k
|
|x
k
y
k
|
d
[a,b]
(x
k
) + d
[a,b]
(x
k
) = 0.
Since y
k
[a, b), for k large enough, there exists
k
(0, 1] such that
a =
1
k
(y
k
(1
k
)b) and then
x
k
, b a) =
1
k
x
k
, b y
k
) 0.
Choosing x
k
d
[a,b]
(x
k
) such that
x
k
, b x
k
) = d
[a,b]
(x
k
, b x
k
), k IN,
from (1.8) we deduce
f
K
(x
k
, b x
k
) t
k
d
[a,b]
(x
k
, b x
k
)
1
k
|b x
k
|
= t
k
x
k
, b x
k
)
1
k
|b x
k
|
1
k
|b x
k
|,
and hence
liminf
k+
f
K
(x
k
, b x
k
) liminf
k+
|b x
k
|
k
= 0.
Choosing x
k
d
[a,b]
(x
k
) such that
x
k
, b a) = d
[a,b]
(x
k
, b a), k IN,
12
we deduce
f
K
(x
k
, b a) t
k
d
[a,b]
(x
k
, b a)
1
k
|b a|
= t
k
x
k
, b a)
1
k
|b a|
1
k
|b a|,
and hence
liminf
k+
f
K
(x
k
, b a) liminf
k+
|b a|
k
= 0.
Then the proof is complete.
As immediate consequence we deduce the following result.
Corollary 4.1 Let f T(X) and K be a local cone approximation
satisfying the same assumptions in Theorem 4.3; then, for each a, b X
with a domf, and for each r f(b) there exist x
0
[a, b) and a
sequence x
k
domf with x
k
f
x
0
such that
liminf
k+
f
K
(x
k
, b a) r f(a), (1.9)
liminf
k+
f
K
(x
k
, b x
k
)
|b x
0
|
|b a|
(r f(a)). (1.10)
Proof. Let us dene the lower semicontinuous function
F(x) = f(x) +x
, x)
where x
satises
x
, a b) = r f(a).
Notice that
F(a) = f(a) +x
, a)
= f(a) +x
, b) + r f(a)
x
, b) + f(b) = F(b)
and
f
K
(x, v) F
K
(x, v) x
, v).
Applying Theorem 4.3 to the function F, there exist x
0
[a, b) and a
sequence x
k
domf with x
k
f
x
0
such that, from (1.2),
liminf
k+
f
K
(x
k
, b a) liminf
k+
F
K
(x
k
, b a) x
, b a)
= r f(a) + liminf
k+
F
K
(x
k
, b a)
r f(a).
REFERENCES 13
Moreover, from (1.3), we deduce
liminf
k+
f
K
(x
k
, b x
k
) liminf
k+
F
K
(x
k
, b x
k
) x
, b x
k
)
= x
, x
0
b) + liminf
k+
F
K
(x
k
, b x
k
)
x
, x
0
b).
Since
x
0
=
|b x
0
|
|b a|
a +
1
|b x
0
|
|b a|
b
we have
liminf
k+
f
K
(x
k
, b x
k
)
|b x
0
|
|b a|
x
, a b) =
|b x
0
|
|b a|
(r f(a)).
This concludes the proof.
5. Conclusions
Mean value theorems for nonsmooth functions have been studied by
means of dierent axiomatic constructions of classes of abstract sub-
dierentials. On the contrary, our analysis is based on the axiomatic
concept of local cone approximation and Kepiderivative introduced by
Elster and Thierfelder (1988) for studying nonsmooth optimization prob-
lems. In this setting we derive an abstract approximate mean value the-
orem that collapses to the wellknown Zagrodnys one choosing K as
the Clarke tangent cone.
References
Aussel D., Corvellec J.N. and Lassonde M. (1995), Mean value property
and subdierential criteria for lower semicontinuous functions, Trans.
Amer. Math. Soc., Vol. 347, pp. 4144-4161.
Borwein J.M., Fitzpatrick S.P. and Giles J.R. (1987), The dierentiabil-
ity of real functions on normed linear space using generalized subgra-
dients, J. Math. Anal. Appl., Vol. 128, pp. 512-534.
Borwein J.M. and Strojwas H.M. (1989), The hypertangent cone, Non-
linear Anal., Vol. 13, pp. 125-144.
Castellani M. and Pappalardo M. (1995), First order cone approxima-
tions and necessary optimality conditions, Optimization, Vol. 35, pp.
113-126.
Clarke F.H. (1975), Generalized gradients and applications, Trans. Amer.
Math. Soc., Vol. 205, pp. 247-262.
14
Clarke F.H. and Ledyaev Yu. S. (1994), Mean value inequalities, Proc.
Amer. Math. Soc., Vol. 122, pp. 1075-1083.
Correa R., Jore A. and Thibault L. (1992), Characterization of lower
semicontinuous convex functions, Proc. Amer. Math. Soc., Vol. 116,
pp. 67-72.
Correa R., Jore A. and Thibault L. (1994), Subdierential monotonicity
as characterization of convex functions, Numer. Funct. Anal. Optim.,
Vol. 15, pp. 531-535.
Diewert W.E. (1981), Alternative characterizations of six kinds of qua-
siconcavity in the nondierentiable case with applications to nons-
mooth programming, in Generalized concavity in optimization and
economics, Schaible S. and Ziemba W.T. (ed.s), Academic Press, pp.
51-93.
Ekeland I. (1974), On the variational principle, J. Math. Anal. Appl.,
Vol. 47, pp. 324-353.
Elster K.H. and Thierfelder J. (1988), Abstract cone approximations and
generalized dierentiability in nonsmooth optimization, Optimization,
Vol. 19, pp. 315-341.
Ioe A.D. (1984), Approximate subdierentials and applications. I: the
nite dimensional theory, Trans. Amer. Math. Soc., Vol. 281, pp. 389-
416.
Komlosi S.(1995), Generalized monotonicity and generalized convexity,
J. Optim. Theory Appl., Vol. 84, pp. 361-376.
Lebourg G. (1975), Valeur moyenne pour gradient generalise, C.R. Acad.
Sci. Paris Ser. A, Vol. 281, pp. 795-797.
Lebourg G. (1979), Generic dierentiability of lipschitzian functions,
Trans. Amer. Math. Soc., Vol. 256, pp. 125-144.
Michel P. and Penot J.P. (1984), Subdierential calculus for lipschitzian
and nonlipschitzian functions, C.R. Acad. Sci. Paris Ser. I Math.,
Vol. 298, pp. 269-272.
Penot J.P. (1988), On the mean value theorem, Optimization, Vol. 19,
pp. 147-156.
Rockafellar R.T. (1980), Generalized directional derivatives and subgra-
dients of nonconvex functions, Canad. J. Math., Vol. 32, pp. 257-280.
Thibault L. (1995), A note on the Zagrodny mean value theorem, Opti-
mization, Vol. 35, pp. 127-130.
Thibault L. and Zagrodny D. (1995), Integration of subdierentials of
lower semicontinuous functions on Banach spaces, J. Math. Anal.
Appl., Vol. 189, pp. 33-58.
Zagrodny D. (1988), Approximate mean value theorem for upper sub-
derivatives, Nonlinear Anal., Vol. 12, pp. 1413-1428.