Nonlinear Analysis 39 (2000) 955 – 976
www.elsevier.nl/locate/na
Some results on integration of subdi erentials
Zili Wu, Jane J. Ye ∗
Department of Mathematics and Statistics, University of Victoria, Victoria, B.C., Canada V8W 3P4
Received 13 January 1998; accepted 28 June 1998
Keywords: Nonsmooth analysis; Integration of subdi erentials; Clarke subdi erential; Michel–
Penot subdi erential; Lower Dini subdi erential; Frechet subdi erential; m-subdi erential and
@∗ -subdi erential
1. Introduction
Integration of subdi erentials is a fundamental problem in nonsmooth analysis. The
problem is whether or not the condition that the subdi erential of f contains the
subdi erential of g implies that f and g di er only by a constant. For this problem,
perhaps the earliest result is Theorem 24.9 in Rockafellar [15], which asserts that if
f and g are two closed proper convex functions such that @c g(x) ⊆ @c f(x) ∀x ∈ R n ;
where @c is the subdi erential in the sense of convex analysis, then f and g di er by
a constant.
Some progress on the study of integration of subdi erentials for nonconvex functions has been made over the last decade or so. Rockafellar [16] proved that if
f; g : R n → R are locally Lipschitz and f is Clarke regular, and if the Clarke subdi erential @g(x) ⊆ @f(x) ∀x ∈ R n , then g(x) = f(x) + C for some constant C. This result was generalized to upper–upper and upper–lower regular functions in any Banach
space by Correa and Thibault [10]. Some classes of nonconvex functions which can
be determined by their Clarke subdi erentials (i.e., any two functions from the same
class di er only by a constant if their Clarke subdi erentials coincide) have also been
found. Qi [14] showed that a primal function whose domain is connected is determined
by its Clarke subdi erential up to an additive constant. Poliquin [13] found a class of
functions called primal lower nice functions which can be determined by their Clarke
subdi erentials (hence their proximal subdi erentials since they coincide for this class
Corresponding author. Tel.: +1-250 -721-7437; fax: +1-250 -721-8962.
E-mail address: janeye@uvic.ca (J.J. Ye)
∗
0362-546X/00/$ - see front matter ? 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 3 6 2 - 5 4 6 X ( 9 8 ) 0 0 2 6 3 - 6
956
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
of functions) up to a constant. Other results on integration of subdi erentials include
those of Borwein [2], Correa and Jofre [9], Thibault and Zagrodny [17].
Although most of these integration results concern the Clarke subdi erential, integrating subdi erentials other than that of Clarke is also important. Clarke and Redhe er [6]
proved that if g : R n → (−∞; ∞] is a lower semicontinuous function, f is a continuously twice di erentiable function on R n and @ g(x) ⊆ @ f(x), where @ denotes the
proximal subdi erential, then f(x) = g(x)+C for some constant C. They also indicated
that if f is di erentiable and @F g(x) ⊆ @F f(x), where @F is the Frechet subdi erential,
then f and g di er only by a constant.
In this paper, we study primarily the integration of @∗ -subdi erentials, a class of subdi erentials which includes the lower Dini subdi erential, the Frechet subdi erential
and the m-subdi erential. We prove that if lower semicontinuous functions f : X → R,
g : X → (−∞; ∞] are such that g − f is lower semicontinuous on a nonempty open
convex subset U of X , and for each x in U , @∗ f(x) ∪ @∗ (−f)(x) is nonempty and
bounded, then g di ers from f by a constant on U if and only if for some subdi erential @0 and any x in U; @∗ g(x) ⊆ @∗ f(x) for x with @∗ f(x) 6= ∅; @∗ (−f)(x) ⊆ @∗ (−g)(x)
for x with @∗ f(x) = ∅ and @0 (g − f)(x) ⊆ @∗ (g − f)(x), where @0 h is a subdi erential
having the property that @0 h(x) ⊆ {0} on U implies that h is constant on U . This result
serves to unify and extend several results in the literature. In particular, a consequence
of the above result is a generalization of Rockafellar’s result on integrability of regular
functions, i.e., if f is a locally Lipschitz function and g is an extended-valued lower
semicontinuous function on an Asplund space or a separable Banach space and if f is
Michel–Penot regular (weaker than Clarke regular) and @− g(x) ⊆ @− f(x), where @−
denotes the lower Dini subdi erential (weaker than @g(x) ⊆ @f(x)) for all x, then f
and g di er by a constant. We also obtain similar results for a separately regular function, i.e., a bivariate function on a product of two Asplund spaces (or two separable
Banach spaces) X × Y which is Clarke regular as a function of x and y separately. Our
main result also extends Clarke and Redhe er’s result by allowing f to be any local
Lipschitz function whose proximal subdi erential or Frechet subdi erential is nonempty
everywhere.
We organize the paper as follows. In Section 2, we give conditions under which
some @∗ -subdi erentials are nonempty and some results on the calculus of these subdi erentials. In Section 3, we discuss our integration results in detail.
Throughout this paper, X is a real Banach space whose open unit ball and dual
space are denoted by B and X ∗ , respectively. H denotes a real Hilbert space.
2. Calculus of subdi erentials
We brie y review some well-known notions of subdi erentials. Let f : X → R be
Lipschitz of rank L near x ∈ X .
• The Clarke derivative of f at x in the direction v is de ned by
f◦ (x; v) := lim sup
y→x
t→0+
f(y + tv) − f(y)
:
t
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
957
• The Michel–Penot derivative of f at x in the direction v is de ned by
f⋄ (x; v) := sup lim sup
y
t→0+
f(x + ty + tv) − f(x + ty)
:
t
• The Clarke subdi erential of f at x is the set
@f(x) := { ∈ X ∗ : h; vi ≤ f◦ (x; v) ∀v ∈ X }:
• The Michel–Penot subdi erential of f at x is the set
@⋄ f(x): = { ∈ X ∗ : h; vi ≤ f⋄ (x; v) ∀v ∈ X }:
Since the Clarke and Michel–Penot derivatives are both sublinear functions of v and
their values f◦ (x; v) and f⋄ (x; v) are bounded by Lkvk, their corresponding subdi erentials are nonempty.
Let f : X → R ∪ {∞} be lower semicontinuous (l.s.c.) at x ∈ dom f := {x ∈ X : f(x)¡
∞} and v be in X .
• The lower Dini derivative of f at x in the direction v is de ned by
f− (x; v) := lim inf
+
t→0
f(x + tv) − f(x)
:
t
• The lower Dini subdi erential of f at x is the set
@− f(x) := { ∈ X ∗ : h; vi ≤ f− (x; v) ∀v ∈ X }:
• The Frechet subdi erential of f at x is the set
(
@F f(x) :=
∈ X ∗ : f(y) − f(x) + (ky − xk) ≥ h; y − xi
for some (t) with lim+
t→0
(t)
= 0; ¿0 and any y in x + B :
t
To unify certain notions of subdi erentials, we de ne the m-subdi erential as in
Clarke et al. [7].
De nition 2.1. Let f : X → R ∪ {∞} be l.s.c. at x ∈ dom f. Let m : [0; ∞) → [0; ∞) be a
modulus function (m-function), that is, m is nondecreasing and limt→0+ m(t) = m(0) = 0.
The m-subdi erential of f at x is the set
@m f(x) := { ∈ X ∗ : f(y) − f(x) + Mm(ky − xk)ky − xk ≥ h; y − xi
for some M ¿0; ¿0 and any y in x + B}:
Common m-subdi erentials include the Fenchel subdi erential of f at x:
@c f(x) := { ∈ X ∗ : f(y) − f(x) ≥ h; y − xi ∀y ∈ x + B};
958
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
the proximal subgradient of f at x:
@ f(x) := { ∈ X ∗ : f(y) − f(x) + M ky − xk 2 ≥ h; y − xi
for some M ¿0; ¿0 and any y in x + B}
and the s-Holder subdi erential (s¿0) of f at x:
@HS f(x) := { ∈ X ∗ : f(y) − f(x) + M ky − xk1+s ≥ h; y − xi
for some M ¿0; ¿0 and any y in x + B}
whose special case (where s = 1) is Lipschitz smooth @LS f(x) which in Hilbert space
coincides with Rockafellar’s proximal subgradient.
Remark 2.2. From the de nition, it is easy to see that the following relations hold:
(i) Let f : X → R ∪ {∞} be l.s.c. at x ∈ dom f. Then for any m-function m(t),
@c f(x) ⊆ @m f(x) ⊆ @F f(x) ⊆ @− f(x):
Moreover, if f : X → R ∪ {∞} is convex, then
@c f(x) = @m f(x) = @F f(x) = @− f(x):
(ii) Let U be a nonempty open convex subset of X and f : U → R bounded above
on a neighborhood of some point of U . Then f is convex if and only if for
each x ∈ U , f is Lipschitz near x and for any m-function m(t),
@c f(x) = @m f(x) = @F f(x) = @− f(x) = @⋄ f(x) = @f(x):
Generally, if f is Lipschitz near x, then
@c f(x) ⊆ @m f(x) ⊆ @F f(x) ⊆ @− f(x) ⊆ @⋄ f(x) ⊆ @f(x):
(iii) A function f : X → R is Frechet di erentiable at x ∈ X if and only if for some
m-function m(t), both @m f(x) and @m (−f)(x) are nonempty, if and only if both
@F f(x) and @F (−f)(x) are nonempty. In this case,
@m f(x) = @F f(x) = {Df(x)};
where Df(x) is the Frechet derivative of f at x.
By the de nition of subdi erentials, it is easy to prove the following scalar multiplication rule and sum rule.
Proposition 2.3. Let f; g : X → R be l.s.c. at x ∈ dom f ∩ dom g: Then
(i) for any ¿0, @∗ (f)(x) = @∗ f(x):
(ii) @∗ f(x) + @∗ g(x) ⊆ @∗ (f + g)(x);
where @∗ denotes @− , @m or @F and by convention A + B = ∅ if either A or B is empty.
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
959
Analogous to Gâteaux and Frechet di erentiability, we de ne m-di erentiability as
follows:
De nition 2.4. For some m-function m(t), a function f : X → R is m-di erentiable at
x ∈ X if there exist in X ∗ , M ¿0 and ¿0 such that
|f(y) − f(x) − h; y − xi | ≤ Mm(ky − xk)ky − xk
∀y ∈ x + B:
Clearly, f is m-di erentiable at x ∈ X if and only if both @m f(x) and @m (−f)(x)
are nonempty.
It turns out that when one of two functions in the sum rule is “di erentiable”, the
inclusion in the sum rule becomes equality. We omit the proof which is similar to that
of [7, Lemma 2.2].
Proposition 2.5. Let f; g : X → R ∪ {∞} be l:s:c: at x ∈ dom f ∩ dom g:
(i) If g is m-di erentiable at x, then
@m (f ± g)(x) = @m f(x) ± @m g(x):
(ii) If g is Gâteaux di erentiable at x, then
@− (f ± g)(x) = @− f(x) ± @− g(x):
(iii) If g is Frechet di erentiable at x, then
@F (f ± g)(x) = @F f(x) ± @F g(x):
Remark 2.6. From Proposition 2:5, we conclude that adding a “di erentiable” function does not change nonemptiness of the subdi erentials, that is, if g : X → R ∪ {∞}
is m-di erentiable (Gateaux
di erentiable or Frechet
di erentiable) at x, then @m f(x)
ˆ
−
m
(@ f(x) or @F f(x)) is nonempty if and only if @ (f + g)(x) (@− (f + g)(x) or
@F (f + g)(x)) is nonempty.
Unlike the Clarke subdi erential of a function which is nonempty when the function
is locally Lipschitz, the subdi erentials considered above may be empty. We now investigate some sucient conditions for nonemptiness of subdi erentials. The following
is one for the nonemptiness of lower Dini subdi erential.
Proposition 2.7. Let f : X → R be Lipschitz near x. Suppose that there exists in
X ∗ satisfying
lim inf {h − ; vi: ∈ @f(y) and y →v x} ≥ 0
∀v ∈ X;
where y →v x denotes that y goes to x in the direction v. Then must be in @− f(x):
Proof. Let ¿0 be small enough such that f is locally Lipschitz on x + B: For
any nonzero vector v, let 0¡t¡=kvk: Then y := x + tv is in x + B: By Lebourg
960
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
Mean-Value Theorem [5, Theorem 2.3.7], there exist in (0; 1) and in @f(x + tv)
such that
f(y) − f(x) = h; y − xi:
Thus
f(x + tv) − f(x)
− h; vi = h − ; vi:
t
Denote z := x + tv. Then is in @f(z) and by assumption,
lim inf {h − ; vi: ∈ @f(z) and z →v x} ≥ 0
∀v ∈ X:
Hence,
lim inf
+
t→0
f(x + tv) − f(x)
− h; vi ≥ 0
t
∀v ∈ X;
which implies that is in @− f(x):
Recall that f : X → R is said to be Clarke regular at x provided that for all v ∈ X , the
usual one-sided derivative f′ (x; v) exists and is equal to the Clarke derivative f◦ (x; v)
(see e.g. [5]). Clearly, if f is Lipschitz near x, then f is Clarke regular at x if and
only if @− f(x) = @f(x): In this case, since the Clarke subdi erential is nonempty, so
is the lower Dini subdi erential. Similarly, we de ne Michel–Penot (m-C or m-MP)
regular function as below.
De nition 2.8. Let f : X → R be Lipschitz near x.
• f is said to be Michel–Penot regular at x if @− f(x) = @⋄ f(x).
• f is said to be m-C regular at x if @m f(x) = @f(x) for some m-function m(t).
• f is said to be m-MP regular at x if @m f(x) = @⋄ f(x) for some m-function m(t).
Obviously, if a function is Clarke regular, then it is Michel–Penot regular. But the
reverse is not true. For example, the function f is de ned by
( 2
x · sin 1x if x 6= 0;
f(x) =
0
if x = 0
is Michel–Penot regular at 0 but not Clarke regular at this point since @− f(0) = @⋄ f(0)
= {0} =
6 [−1; 1] = @f(0):
The following proposition gives a sucient condition for m-subdi erential to be
nonempty.
Proposition 2.9. Let f : X → R be Lipschitz near x. If for some m-function m(t),
in X ∗ satis es
lim
inf
y→x
∈@⋄ f(y)
h − ; y − xi
¿−∞;
m(ky − xk)ky − xk
(1)
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
961
then is in @m f(x): And hence if f is m-MP hypomonotone at x, i.e., inequality (1)
holds for all in @⋄ f(x), then f is m-MP regular at x.
Proof. Let ¿0 be small enough such that f is locally Lipschitz on x + B: Then for
any y in x + B; by the Mean-Value Theorem [3, Theorem 19], there exist in (0; 1)
and in @⋄ f(x + (y − x)) such that
f(y) − f(x) = h; y − xi:
Denote z = x + (y − x). Then
h − ; y − xi
f(y) − f(x) − h; y − xi
=
m(ky − xk)ky − xk
m(ky − xk)ky − xk
=
h − ; (z − x)=i
m(ky − xk)(kz − xk=)
=
m(kz − xk)
h − ; z − xi
·
:
m(ky − xk) m(kz − xk)kz − xk
By assumption,
lim
inf
z→x
∈@⋄ f(z)
h − ; z − xi
¿−∞:
m(kz − xk)kz − xk
Thus
lim inf
y→x
f(y) − f(x) − h; y − xi
¿−∞;
m(ky − xk)ky − xk
which implies that is in @m f(x):
Furthermore, if inequality (1) holds for each in @⋄ f(x), then by Remark 2.2(ii),
we have
@m f(x) = @F f(x) = @⋄ f(x):
Therefore, f is m-MP regular at x.
By Proposition 2.9 and the fact that @⋄ f(x) ⊆ @f(x) (see Remark 2.2(ii)), we
can easily obtain the following result which implies that for an m-C hypomonotone
function, which is introduced as a generalization of hypomonotone function de ned in
Rockafellar [16], its m-subdi erential is always nonempty.
Corollary 2.10. Let f : X → R be Lipschitz near x. If for some m-function m(t), in
X ∗ satis es
lim
inf
y→;x
∈@f(y)
h − ; y − xi
¿ − ∞;
m(ky − xk)ky − xk
(2)
962
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
then is in @m f(x): And hence if f is m-C hypomonotone at x, i.e., inequality (2)
holds for all in @f(x), then f is m-C regular at x.
Note that a hypomonotone function is an m-C hypomonotone function when m(t) = t:
Based on Corollary 2.10, for hypomonotone function f, we have @ f(x) = @f(x),
which was proved by Poliquin [13] when X = R n .
In nonsmooth analysis, we often deal with the pointwise maximum function of several functions in the form of f(x) = max{f1 (x); : : : ; fl (x)}. For such a function, we
present two cases in which the subdi erential of f can be expressed by those of fi ’s
and hence its subdi erential is nonempty.
Proposition 2.11. Let X be a real re exive Banach space. Suppose that fi : X → R is
continuous for i = 1; : : : ; l and f(x) := max{f1 (x); : : : ; fl (x)}. Denote
I (x) := {1 ≤ i ≤ l: fi (x) = f(x)}
∀x ∈ X:
If for each i in I (x), fi is Frechet di erentiable at x ∈ X , then for some m-function
m(t),
@m f(x) = @F f(x) = co{Dfi (x): i ∈ I (x)};
where co denotes the convex hull, Dfi (x) is the Frechet derivative of fi at x for each
i in I (x):
Proof. According to Remark 2.2(i), @m f(x) ⊆ @F f(x). Hence it suces to prove that
for some m-function m(t),
@F f(x) ⊆ co{Dfi (x): i ∈ I (x)} ⊆ @m f(x):
By Remark 2.2(iii), it is straightforward to show that for some m-function m(t),
co{Dfi (x): i ∈ I (x)} ⊆ @m f(x):
Hence we only need to prove @F f(x) ⊆ co{Dfi (x): i ∈ I (x)}. By contradiction suppose
that is in @F f(x) but not in co{Dfi (x): i ∈ I (x)}: Then there exist M ¿0, 1 ¿0 and
(t) with limt→0+ (t)=t = 0 such that
f(y) − f(x) + (ky − xk) ≥ h; y − xi
∀y ∈ x + 1 B:
Denote S1 = {} and
X
X
i Dfi (x): i ≥ 0;
i = 1 :
S2 =
i∈I (x)
(3)
i∈I (x)
Obviously, S1 and S2 are two disjoint closed convex subsets of X ∗ . Also S1 is compact.
By the separation theorem [8, Theorem 3.9, Ch. IV], S1 and S2 can be strictly separated,
that is, there exist a nonzero vector p∗∗ in X ∗∗ and ¿0 such that kp∗∗ k = 1 and
h p∗∗ ; i ≥ + sup{h p∗∗ ; i: ∈ S2 }:
963
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
Since X is re exive, there is p in X satisfying kpk = kp∗∗ k = 1 and
h p∗∗ ; x∗ i = hx∗ ; pi
∀x∗ ∈ X ∗ :
Thus,
h; pi ≥ + sup{h; pi: ∈ S2 }:
(4)
On the other hand, by the continuity of fi for i = 1; : : : ; l, there exists 2 ¿0 such
that
f(y) = max{fi (y): i ∈ I (x)}
∀y ∈ x + 2 B:
For each i in I (x), the Frechet di erentiability of fi at x implies
fi (y) = fi (x) + hDfi (x); y − xi + ky − xk i (y − x)
∀y ∈ X;
where limy→x i (y − x) = 0 for each i in I (x). Therefore for any y in x + 2 B,
f(y) = max{fi (y): i ∈ I (x)}
= max{fi (x) + hDfi (x); y − xi + ky − xk i (y − x): i ∈ I (x)}
= max{f(x) + hDfi (x); y − xi + ky − xk i (y − x): i ∈ I (x)}
= f(x) + max{hDfi (x); y − xi + ky − xk i (y − x): i ∈ I (x)}:
Taking = min{1 ; 2 } and y = x + (1=n)p, then by inequality (3), we have
1
1
1
1
+ h; pi for n¿ :
f x + p ≥ f(x) −
n
n
n
And, hence,
max hDfi (x); pi +
i
1
1
p : i ∈ I (x) ≥ h; pi − n
n
n
1
for n¿ :
Letting n → ∞ yields
max{hDfi (x); pi: i ∈ I (x)} ≥ h; pi:
Thus,
sup {h; pi: ∈ S2 } = sup
X
i hDfi (x); pi: i ≥ 0;
i∈I (x)
X
i∈I (x)
≥ max{hDfi (x); pi: i ∈ I (x)} ≥ h; pi;
which contradicts inequality (4). Therefore,
@F f(x) ⊆ co{Dfi (x): i ∈ I (x)}
and the proof of the proposition is complete.
i = 1
964
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
The following proposition shows that the class of m-C regular functions is closed
under ( nite) pointwise maximization.
Proposition 2.12. Let fi : X → R be Lipschitz near x for i = 1; : : : ; l and f(x) =
max{f1 (x); : : : ; fl (x)}: Denote
I (x) := {1 ≤ i ≤ l: fi (x) = f(x)}
∀x ∈ X:
If for some m-function mi , fi is mi -C regular for each i in I (x); then f is m-C regular
and
@m f(x) = co{@m fi (x): i ∈ I (x)};
where m(t) = max{mi (t): i ∈ I (x)}:
Proof. Since @mi fi (x) ⊆ @m fi (x) ⊆ @fi (x) and @mi fi (x) = @fi (x) for each i in I (x),
@m fi (x) = @fi (x): By Remark 2.2(ii) and [5, Proposition 2.3.12], we have
@m f(x) ⊆ @f(x) ⊆ co{@fi (x): i ∈ I (x)} = co{@m fi (x): i ∈ I (x)}:
So it suces to show that
co{@m fi (x): i ∈ I (x)} ⊆ @m f(x):
We suppose that i is in @m fi (x) for i in I (x): Then there exist M ¿0 and ¿0
such that
fi (y) − fi (x) + Mm(ky − xk)ky − xk ≥ hi ; y − xi
∀y ∈ x + B:
Thus,
f(y) − f(x) + Mm(ky − xk)ky − xk ≥ hi ; y − xi
∀y ∈ x + B;
m
m
which implies
P that i is in @ f(x) for any i in I (x): Since @ f(x) is convex, for any
i ≥ 0 with i∈I (x) i = 1;
X
i i ∈ @m f(x):
i∈I (x)
This is what we need to prove.
The next proposition indicates that the sum of two m-C regular functions is still an
m-C regular function and the m-subdi erential of sum function is equal to the sum of
m-subdi erentials of these two functions.
Proposition 2.13. If f; g : X → R are m-C regular at x, then f +g is also m-C regular
at x and
@m (f + g)(x) = @m f(x) + @m g(x):
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
965
Proof. Since @m f(x) = @f(x) 6= ∅; @m g(x) = @g(x) 6= ∅; by Proposition 2.3, Remark
2.2(ii) and Proposition 2.3.3 in [5], we have
@m f(x) + @m g(x) ⊆ @m (f + g)(x) ⊆ @(f + g)(x)
⊆ @f(x) + @g(x) = @m f(x) + @m g(x):
It follows that f + g is also m-C regular and the equality stated holds.
The following is a chain rule for m-subdi erentials.
Proposition 2.14. Let gi : X → R be Lipschitz of rank L near x0 in X for i = 1; : : : ; l
and f : Rl → R ∪ {∞} be l.s.c. Suppose that g(x0 ) := (g1 (x0 ); : : : ; gl (x0 ))t is in dom f
and i ≥ 0 (i = 1; : : : ; l) whenever := (1 ; : : : ; l )t is in @m f(g(x0 )) for some m-function
m(t). Then
l
X
i @m gi (x0 ) ⊆ @m1 (f ◦ g)(x0 );
i=1
where m1 (t) = max{m(t); m(lLt)}:
Proof. Let be in @m f(g(x0 )). Then for some m-function m(t), there exist M1 ¿0 and
1 ¿0 such that
f(y) − f(g(x0 )) + M1 m(ky − g(x0 )k)ky − g(x0 )k
≥
l
X
i (yi − gi (x0 ))
∀y ∈ g(x0 ) + 1 Bl ;
i=1
where y = (y1 ; : : : ; yl )t and Bl is the open unit ball in Rl .
Let g be Lipschitz of rank L near x0 . Then for 1 , there is 2 ¿0 such that
|gi (x) − gi (x0 )| ≤ Lkx − x0 k¡1 =l ∀x ∈ x0 + 2 B:
If i is in @m gi (x0 ) for i = 1; : : : ; l; then there exist M2 ¿0 and (2 ¿)¿0 such that
gi (x) − gi (x0 ) ≥ hi ; x − x0 i −M2 m(kx − x0 k)kx − x0 k
∀x ∈ x0 + B:
Since for any x in x0 + B;
lLkx − x0 k ≥
l
X
|gi (x) − gi (x0 )| ≥ kg(x) − g(x0 )k;
i=1
we have
f(g(x)) − f(g(x0 )) + M1 lLm(lLkx − x0 k)kx − x0 k
≥ f(g(x)) − f(g(x0 )) + M1 m(kg(x) − g(x0 )k)kg(x) − g(x0 )k
966
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
≥
l
X
i (gi (x) − gi (x0 ))
i=1
≥
* l
X
i i ; x − x0
i=1
+
−
l
X
i M2 m(kx − x0 k)kx − x0 k;
i=1
which implies that
l
X
i i ∈ @m1 (f ◦ g)(x0 ):
i=1
Hence,
l
X
i @m gi (x0 ) ⊆ @m1 (f ◦ g)(x0 ):
i=1
3. Integration of subdi erentials
In this section, we prove our main results on integration of subdi erentials and show
that our results unify and extend many results in the literature.
First consider the following example from Poliquin [13]. The lower semicontinuous
functions de ned by
f(x) =
(
0 if x ≤ 0;
1 if x¿0
and
g(x) =
(
0 if x ≤ 0;
2 if x¿0
do not di er by the same constant on the intervals (−∞; 0] and (0; ∞) even though
they have the same proximal subgradient everywhere. Observe that g − f is l.s.c. and
@ f(x) is nonempty for all x but @ f(0) = [0; ∞) is not bounded. However, if g − f is
l.s.c., @ f(x) is nonempty and bounded, and @ g(x) ⊆ @ f(x), then f and g must di er
by only a constant and the result is actually true for the class of @∗ -subdi erentials
de ned as follows.
De nition 3.1. Let f; g : X → R ∪ {∞} be lower semicontinuous at x ∈ dom f.
(1) The @∗ -subdi erential of f at x, denoted by @∗ f(x), is a subset of X ∗ satisfying
the following properties:
(p1 ) @∗ f(x) = @c f(x) whenever f is convex.
(p2 ) @∗ f(x) ⊆ @f(x) when f is Lipschitz near x.
(p3 ) 0 ∈ @∗ f(x) when x is a local minimum of f.
(p4 ) @∗ f(x) + @∗ g(x) ⊆ @∗ (f + g)(x).
˜
˜
(2) The @-subdi
erential of f at x, denoted by @f(x),
is a subset of X ∗ satisfying
the following properties [1]:
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
967
˜
@f(x)
= @c f(x) whenever f is convex.
˜
0 ∈ @f(x) when x is a local minimum of f.
˜ + g)(x) ⊆ @f(x)
˜
˜
@(f
+ @g(x)
whenever g is a real-value convex continuous
˜
˜
˜
function which is @-di erentiable at x, i.e., both @g(x)
and @(−g)(x)
are
nonempty.
(3) The @∗ -subdi erential of f at x, denoted by @∗ f(x), is a subset of X ∗ satisfying
the following properties [17]:
(P1 ) @∗ f(x) = @c f(x) whenever f is convex.
(P2 ) @∗ f(x) = @∗ g(x) whenever f = g on a neighbourhood of x.
(P3 ) 0 ∈ @∗ f(x) when x is a local minimum of f.
(P4 ) @∗ (f + g)(x) ⊆ @∗ f(x) + @∗ g(x) for any continuous convex function g.
(i)
(ii)
(iii)
By Remark 2.2, Propositions 2.3 and 2.5, the lower Dini subdi erential, the Frechet
˜
subdi erential and the m-subdi erential are both @∗ -subdi erentials and @-subdi erentials while the Clarke subdi erential and the Michel–Penot subdi erential are
˜
both @-subdi
erentials and @∗ -subdi erentials.
˜
De nition 3.2 (Aussel et al. [1]). A norm k · k on X is said to be @-smooth
if the
˜
following functions are @-di erentiable:
2
min{kx − ck 2 : c = a + (1 − )b; ∈ [0; 1]} where a; b ∈ X .
(i) d[a;
b] (x) :=
P
P
(ii) △2 (x) := n n kx − n k 2 , where n n = 1; n ≥ 0, and (n ) converges in X .
˜
˜
A Banach space X admits a @-smooth
renorm if it has an equivalent @-smooth
norm.
−
It was pointed out in [1] that a norm is @F -smooth (resp. @ -smooth, @LS -smooth) if
and only if it is Frechet di erentiable (resp. Gâteaux di erentiable, Lipschitz smooth)
o the origin and that re exive Banach spaces (resp. separable Banach spaces, Hilbert
spaces and Lp (2 ≤p¡∞) spaces) admit a @F -smooth renorm (resp. @− -smooth renorm,
@LS -smooth renorm).
We use @0 h to denote any subdi erential of h for which the inclusion @0 h(x) ⊆ {0}
on a nonempty open convex subset U of X implies that h is a constant on U . For
˜
example, @∗ -subdi erentials in a Banach space, @-subdi
erentials in a Banach space
˜
with a @-smooth renorm and the Frechet subdi erential in an Asplund space (that is,
the Banach space on which every continuous convex function is generically Frechet
di erentiable) are all @0 -subdi erentials according to the following lemma which is
very useful to our result.
Lemma 3.3. Let U be a nonempty open convex subset of X and h : U → (−∞; ∞]
be l:s:c: Then h is constant if and only if
(i) @∗ h(x) ⊆ {0} ∀x ∈ U when X is a Banach space [17, Theorem 1.2].
˜
˜
(ii) @h(x)
⊆ {0} ∀x ∈ U when X is a Banach space with a @-smooth
renorm [1, Theo−
rem 5.2]. (In particular, h is constant if and only if @ h(x) ⊆ {0} ∀x ∈ U when X
is a separable Banach space and h is constant if and only if @ h(x) ⊆ {0} ∀x ∈ U
when X is a Hilbert space [7, Corollary 3.9] or a Lp (2 ≤ p¡∞) space.)
(iii) @F h(x) ⊆ {0} ∀x ∈ U when X is an Asplund space [12, Corollary 8.9].
To prove Theorem 3.5 we need the following cancellation rule.
968
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
Lemma 3.4. Let A and B be nonempty subsets of X ∗ , and let B be bounded. Suppose
A + B ⊆ B. Then
A = {0}:
Proof. Suppose that there exists a nonzero element a in A. Since a + b is in B for any
b in B, na + b is in B for any n in N (natural number set) and any b in B: By the
boundness of B, there is a positive M such that kbk∗ ≤ M for all b in B: In particular,
kna + bk∗ ≤ M
∀n ∈ N and b ∈ B:
On the other hand, since kak∗ 6= 0, for any n in N with n¿2M=kak∗ , we have
kna + bk∗ ≥ nkak∗ − kbk∗ ≥ nkak∗ − M ¿M;
which contradicts the boundness of B:
Now, we state and prove our main results in this section as below. In the following
context we assume that (−f) is l.s.c. at x whenever @∗ (−f)(x) is used.
Theorem 3.5. Let U be a nonempty open convex subset of X; f : U → R and g : U →
(−∞; ∞] be such that g − f is lower semicontinuous. Suppose that for each x in U ,
@∗ f(x) ∪ @∗ (−f)(x) is nonempty and bounded. Then for some constant C and any x
in U , f(x) = g(x) + C if and only if for some @0 -subdi erential,
@∗ g(x) ⊆ @∗ f(x) ∀x ∈ U s:t: @∗ f(x) 6= ∅;
(5)
@∗ (−f)(x) ⊆ @∗ (−g)(x)
(6)
∀x ∈ U s:t: @∗ f(x) = ∅
and
@0 (g − f)(x) ⊆ @∗ (g − f)(x) ∀x ∈ U:
(7)
Proof. The “only if ” part is straight from the de nition of @∗ -subdi erential and
Lemma 3.3. We prove the “if ” part of this result as follows.
Suppose that Eqs. (5) –(7) hold. Consider the function h : U → (−∞; ∞] de ned by
h(x) = g(x) − f(x) ∀x ∈ U:
For any x in U with @∗ f(x) 6= ∅,
@0 h(x) + @∗ f(x) ⊆ @∗ h(x) + @∗ f(x) (by Eq: (7))
⊆ @∗ g(x) (by (p4 ) of De nition 3:1)
⊆ @∗ f(x) (by Eq: (5))
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
969
and for any x in U with @∗ f(x) = ∅, since @∗ (−f)(x) 6= ∅,
@0 h(x) + @∗ (−g)(x) ⊆ @∗ h(x) + @∗ (−g)(x) (by Eq: (7))
⊆ @∗ (−f)(x) (by (p4 ) of De nition 3:1)
⊆ @∗ (−g)(x) (by Eq: (6));
which implies by Lemma 3.4 that @0 h(x) ⊆ {0} ∀x ∈ U . Thus by the de nition of
@0 -subdi erential, for some constant C and any x in U , h(x) = C, that is, f(x) =
g(x) + C:
Remark 3.6. It is natural to ask whether the condition that @∗ f(x) ∪ @∗ (−f)(x) is
nonempty and bounded everywhere in Theorem 3.5 can be weaken to the condition
that it is satis ed on a set that is big in the sense of either measure or category. This
question is answered partially in the case @∗ = @HS by Borwein and Wang [4], who
gave an example where f is Lipschitz continuous on R and @HS f(x) ∪ @HS (−f)(x) is
densely nonempty but there exist in nitely many Lipschitz functions g on R di ering
by more than a constant, such that @HS g(x) ⊆ @HS f(x) for all x ∈ R and @HS (−g)(x) ⊆
@HS (−f)(x) for all x ∈ R.
Using the continuity of functions, we can extend Theorem 3.5 to a connected subset
of a Banach space.
Corollary 3.7. Let Ui be a nonempty open convex subset of X for i = 1; : : : ; l and
Sl
V := i=1 U i be connected, where U i is the closure of Ui . Suppose that f; g : V → R
are continuous and that for any x in Ui (i = 1; : : : ; l); @∗ f(x) ∪ @∗ (−f)(x) is nonempty
and bounded. Then for some constant C and any x in V , f(x) = g(x) + C if and only
if for some @0 -subdi erential and any x ∈ Ui (i = 1; : : : ; l),
@∗ g(x) ⊆ @∗ f(x) for x with @∗ f(x) 6= ∅;
@∗ (−f)(x) ⊆ @∗ (−g)(x)
for x with @∗ f(x) = ∅
and
@0 (g − f)(x) ⊆ @∗ (g − f)(x):
Example 3.8. Let U1 = {(x; y): x¿0; y¿0}, U2 = {(x; y): x¡0; y¿0}, U3 = {(x; y):
x¡0; y¡0} and U4 = {(x; y): x¿0; y¡0}. By applying Corollary 3.7 to Ui
(i = 1; 2; 3; 4) and @0 = @∗ = @ , we see that the following locally Lipschitz function
f : R 2 → R de ned by
f(x; y) = |x| − |y|
970
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
is the uniquely locally Lipschitz
{(1; −1)}
{(−1; −1)}
@ f(x; y) =
{(−1; 1)}
{(1; 1)}
function satisfying f(0; 0) = 0 and
if (x; y) ∈ U1 ;
if (x; y) ∈ U2 ;
if (x; y) ∈ U3 ;
if (x; y) ∈ U4 :
Note that since @ f(0; 0) = ∅ and @ (−f)(0; 0) = ∅, we can not directly apply
Theorem 3.5 to U = R 2 to get the above conclusion.
From the proof of Theorem 3.5, if @∗ f(x) is nonempty and bounded on U , then we
will not need the information of @∗ (−f).
Theorem 3.9. Let U be a nonempty open convex subset of X , f : U → R and g : U →
(−∞; ∞] be such that g−f is lower semicontinuous. Suppose that @∗ f(x) is nonempty
and bounded for each x in U . Then for some constant C and any x in U , f(x) = g(x)+
C if and only if for some @0 -subdi erential and any x in U ,
@∗ g(x) ⊆ @∗ f(x) and
@0 (g − f)(x) ⊆ @∗ (g − f)(x):
Remark 3.10. (i) For a locally Lipschitz function f, due to the fact that @f(x) is
nonempty and bounded and contains @∗ f(x), @∗ f(x) is also bounded. So the condition
that @∗ f(x) is bounded in Theorem 3.9 can be omitted if f is locally Lipschitz.
(ii) According to Lemma 3.3, we can take @ (resp. @− , @F ) as @0 in Theorems 3.5
and 3.9 when X is a Hilbert space (resp. a separable space, an Asplund space). In
Hilbert space, the inclusion @ (g − f)(x) ⊆ @∗ (g − f)(x) is satis ed automatically by
many @∗ -subdi erentials including Frechet subdi erentials, lower Dini subdi erentials
and m-subdi erentials for all modulus functions m which satisfy
lim inf
+
t→0
m(t)
¿0:
t
Next, we apply Theorems 3.5 and 3.9 to obtain some results with particular @∗ subdi erentials and particular spaces and explore the relationship between our results
and the existing results.
The rst consequence of Theorem 3.9 is an integration result on pointwise maximum
function of nite number of functions.
Corollary 3.11. Let U be a nonempty open convex subset of a real re exive Banach
space X . Let g : U → (−∞; ∞] be lower semicontinuous, fi : U → R be continuous
for i = 1; : : : ; l and f(x) := max{f1 (x); : : : ; fl (x)}. If for each i in I (x), fi is Frechet
di erentiable at x ∈ U . Then for some constant C and any x in U , f(x) = g(x) + C
if and only if for any x in U ,
@F g(x) ⊆ @F f(x):
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
971
Proof. By Proposition 2.11, for all x ∈ U , @F f(x) = co{Dfi (x): i ∈ I (x)}, which means
that @F f(x) is nonempty and bounded at each x ∈ U . Recall that a re exive
Banach space is an Asplund space. By Lemma 3.3(iii), one can take @0 = @F . So taking
@∗ = @0 = @F in Theorem 3.9 completes the proof.
The following question was asked on [11, p. 104]: Let f; g : R n → R be locally
Lipschitz with f(0) = 0 = g(0) and @ g(x) ⊆ @ f(x) for all x. Does it follow that
f and g are identical? The negative answer was given by Borwein and Wang [4]
who found in nitely many Lipschitz functions g and a Lipschitz function f satisfying
@ g(x) ⊆ @ f(x) for all x ∈ R but f and g di er by more than a constant. In their example, @ f(x) is not always nonempty. However, if @ f(x) is nonempty for every x,
then we can apply Theorem 3.9 to the proximal subgradient and obtain the following
corollary which not only gives a positive answer under the condition that @ f(x) is
nonempty but also extends a result of Clarke and Redhe er [6] in which H = R n and
f ∈ C 2 (U ).
Corollary 3.12. Let U be a nonempty open convex subset of a Hilbert space H
or a Lp (2 ≤ p¡∞) space, f : U → R locally Lipschitz and g : U → (−∞; ∞] lower
semicontinuous. Suppose that for any x in U , @ f(x) is nonempty. Then for some
constant C and any x in U , f(x) = g(x) + C if and only if @ g(x) ⊆ @ f(x).
Proof. By Lemma 3.3(ii), @ can be taken as @0 . Hence taking @∗ = @0 = @ in
Theorem 3.9 completes the proof.
Similarly, by Lemma 3.3(iii) and taking @∗ = @0 = @F in Theorem 3.9, we can obtain
the following corollary with a weaker condition than that of Clarke and Redhe er [6]
in which X = R n and f is di erentiable.
Corollary 3.13. Let U be a nonempty open convex subset of an Asplund space X ,
f : U → R locally Lipschitz and g : U → (−∞; ∞] lower semicontinuous. Suppose that
for any x in U , @F f(x) is nonempty. Then for some constant C and any x in U ,
f(x) = g(x) + C if and only if @F g(x) ⊆ @F f(x).
Recall that @F f(x) ⊆ @− f(x) and that f is Michel–Penot regular at x means that
@ f(x) is nonempty and bounded. By Lemma 3.3(ii), one can take @0 = @F when the
space is an Aspund space and @0 = @− when the space is a separable Banach space.
Taking @∗ = @− in Theorem 3.9 yields the following result.
−
Corollary 3.14. Let U be a nonempty open convex subset of an Asplund space (or
a separable Banach space), f : U → R locally Lipschitz and g : U → (−∞; ∞] lower
semicontinuous. Suppose that f is Michel–Penot regular at any x in U . Then for
some constant C and any x in U , f(x) = g(x) + C if and only if one of the following
inclusions holds:
(i) @− g(x) ⊆ @− f(x) for all x in U .
(ii) g is locally Lipschitz on U and @⋄ g(x) ⊆ @⋄ f(x) for all x in U .
972
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
Proof. (i) It is a special case of Theorem 3.9 where @0 = @F and @∗ = @− .
(ii) The necessity follows directly from the de nition of Michel–Penot subdi erential.
Suciency is from (i) and the following inclusions
@− g(x) ⊆ @⋄ g(x) ⊆ @⋄ f(x) = @− f(x) ∀x ∈ U:
Note that the class of m − MP regular functions and m − C regular functions are
Michel–Penot regular. According to Proposition 2.9 and Corollary 3.14, the m − MP
hypomonotone functions can be determined by their @− -subdi erentials up to a constant.
By Lemma 3.3(ii), one can take @0 = @ when the space is a Hilbert space. Taking
@∗ = @ in Theorem 3.9 yields the followng result which slightly improves Poliquin’s
result on integrability of hypomonotone functions on R n [13] by not requiring the function g to be a hypomonotone function and with weaker assumption @ g(x) ⊆ @ f(x).
Corollary 3.15. Let U be a nonempty open convex subset of a Hilbert space,
f : U → R locally Lipschitz and g : U → (−∞; ∞] lower semicontinuous. Suppose that
f is hypomonotone at any x in U . Then for some constant C and any x in U ,
f(x) = g(x) + C if and only if one of the following inclusions holds:
(i) @ g(x) ⊆ @ f(x) for all x in U .
(ii) g is locally Lipschitz on U and @g(x) ⊆ @f(x) for all x in U .
In the following result, we will apply Corollary 3.14 to the function f which is
de ned on a nonempty open convex set of X × Y and is regular separately in x and y
or f is regular in x and −f is regular in y. These functions generalize the convex–
convex functions and the convex–concave functions. Denote
f1⋄ ((x; y); u) := sup lim sup
x′
t→0+
f(x + tx′ + tu; y) − f(x + tx′ ; y)
;
t
@⋄1 f(x; y) := { ∈ X ∗ : h; ui ≤ f1⋄ ((x; y); u) ∀u ∈ X }
the partial Michel–Penot derivative of f with respective to x at (x; y) in the direction u
and the partial Michel–Penot subdi erential of f with respective to x at (x; y). Similarly
we use f2⋄ ((x; y); v) (f1− ((x; y); u) and f2− ((x; y); v)) and @⋄2 f(x; y) (@−
1 f(x; y) and
f(x;
y))
to
denote
the
corresponding
limit
and
subdi
erential.
@−
2
Corollary 3.16. Let X and Y be Asplund spaces (or separable Banach spaces) and
U be a nonempty open convex subset of X × Y . Suppose that f : U → R is a locally
Lipschitz function and g : U → (−∞; ∞] is a lower semicontinuous function such that
−
−
⋄
⋄
⋄
f satis es @−
1 f(x; y) = @1 f(x; y) and @2 f(x; y) = @2 f(x; y) (@2 (−f)(x; y) = @2 (−f)
(x; y)) at any (x; y) in U . Then for some constant C and any (x; y) in U , f(x; y) =
g(x; y) + C if and only if for any (x; y) in U , one of the following conditions holds:
−
−
−
−
(i) @−
1 g(x; y) ⊆ @1 f(x; y) and @2 g(x; y) ⊆ @2 f(x; y) (g is continuous and @2 (−g)
−
(x; y)⊆ @2 (−f)(x; y)).
(ii) @⋄1 g(x; y) ⊆ @⋄1 f(x; y) and @⋄2 g(x; y) ⊆ @⋄2 f(x; y) (g is locally Lipschitz and @⋄2 (−g)
(x; y) ⊆ @⋄2 (−f)(x; y)).
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
973
⋄
Proof. “only if ” part is directly from the de nition of @−
i and @i (i = 1; 2). We only
prove “if ” part. Since
−
⋄
⋄
@−
1 g(x; y) ⊆ @1 g(x; y) ⊆ @1 f(x; y) = @1 f(x; y)
and
−
⋄
⋄
@−
2 g(x; y) ⊆ @2 g(x; y) ⊆ @2 f(x; y) = @2 f(x; y)
−
⋄
⋄
(@−
2 (−g)(x; y) ⊆ @2 (−g)(x; y) ⊆ @2 (−f)(x; y) = @2 (−f)(x; y));
condition (ii) implies condition (i). Therefore, we only need to show that condition (i)
is sucient.
−
−
−
Let for any (x; y) in U , @−
1 g(x; y) ⊆ @1 f(x; y) and @2 g(x; y) ⊆ @2 f(x; y) (resp.
−
−
@2 (−g)(x; y) ⊆ @2 (−f)(x; y)) and denote for any x ∈ X and y ∈ Y ,
U (x) := {y ∈ Y : (x; y) ∈ U }
and
U (y) := {x ∈ X : (x; y) ∈ U }:
Then U (x) and U (y) are open convex subsets in X and Y , respectively, and (x; y) ∈ U
if and only if both U (x) and U (y) are nonempty. By Corollary 3.14,
f(x; y) = g(x; y) + c(y)
∀x ∈ U (y):
Note that c(y) is continuous on U (x) for any x ∈ X such that U (x) is nonempty and
that
−
−
−
@−
2 f(x; y) + @ (−c)(y) ⊆ @2 g(x; y) ⊆ @2 f(x; y)
−
−
−
(@−
2 (−f)(x; y) + @ c(y) ⊆ @2 (−g)(x; y) ⊆ @2 (−f)(x; y)):
By Lemma 3.4, @− (−c)(y) ⊆ {0} (@− c(y) ⊆ {0}). Thus c(y) = C for some constant C.
Therefore,
f(x; y) = g(x; y) + C
∀(x; y) ∈ U:
Remark 3.17. (i) Note that if f is Clarke regular and @g(x) ⊆@f(x) at any x ∈ U ⊆ X ,
then f is Michel–Penot regular and @− g(x) ⊆ @− f(x). By Corollary 3.16, f = g + C
on U for some constant C. Therefore, Corollary 3.16 extends Rockafellar’s result on
integrability of regular functions from R n to any Asplund space with weaker condition
@− g(x) ⊆ @− f(x).
(ii) The result of Corollary 3.16 is di erent from that of [10, Proposition 3.7] which
states that if f is upper–upper (resp. upper–lower) regular at any (x; y) ∈ U , then
@g(x; y) ⊆ @f(x; y) ∀x ∈ U if and only if f = g + C on U . By [10, Proposition 1.4],
f is upper–upper (resp. upper–lower) regular at (x; y) if and only if
f◦ ((x; y); (u; 0)) = f− ((x; y); (u; 0))
and
f◦ ((x; y); (0; v)) = f− ((x; y); (0; v));
(−f)◦ ((x; y); (0; v)) = (−f)− ((x; y); (0; v));
974
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
for any u ∈ X and v ∈ Y , from which we have
f1◦ ((x; y); u) = f1− ((x; y); u)
and
f2◦ ((x; y); v) = f2− ((x; y); v)
((−f)◦2 ((x; y); v) = (−f)−
2 ((x; y); v)):
Thus for any (x; y) ∈ U;
−
−
⋄
⋄
⋄
@−
1 f(x; y)= @1 f(x; y) and @2 f(x; y)= @2 f(x; y)(@2 (−f)(x; y)= @2 (−f)(x; y)):
Hence that f is upper–upper regular at (x; y) implies that f is Clarke regular in x
and y, respectively, at this point. So [10, Proposition 3.7] has stronger assumptions on
the function f. However, the conclusions of [10, Proposition 3.7] and Corollary 3.16
are di erent. Since there is no exact relationship between the partial subdi erential
and the Clarke subdi erential, Corollary 3.16 cannot be compared with [10, Proposition 3.7].
Since the composite function of a strictly di erentiable map and a Clarke regular
map is Clarke regular, Corollary 3.16 has the following consequence about integrability
of composite functions.
Corollary 3.18. Let U be a nonempty open convex subset of an Asplund space (or a
separable Banach space) X . Suppose that F is a strictly di erentiable map from U to
Banach space Y and g is a real-valued function de ned on Y and Clarke regular at
F(x) for each x ∈ U . Then for any l.s.c. function f : U → (−∞; ∞], f(x) = g ◦ F(x) +
C ∀x ∈ U if and only if
@− f(x) ⊆ @g− (F(x)) ◦ Ds F(x) ∀x ∈ U;
where Ds F(x) is the strict derivative of F at x.
Proof. By [5, Theorem 2.3.10], f1 := g ◦ F is Clarke regular and
@− f1 (x) = @f1 (x) = @g(F(x)) ◦ Ds F(x) = @− g(F(x)) ◦ Ds F(x)
at each x in U . The result follows from Corollary 3.16.
Using Corollary 3.16, we can recover the result of Rockafellar which says that a
closed proper convex function is integrable in the case where int(dom f) is nonempty.
Corollary 3.19. Let f; g : R n → R ∪ {+∞} be l.s.c. convex functions such that
int(dom f)(the interior of dom f) is nonempty. Suppose that @c g(x) ⊆ @c f(x) for
each x ∈ R n . Then f = g + C on R n for some constant C.
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
975
Proof. Denote U := int(dom f). Then U is nonempty, open and convex and f is convex and locally Lipschitz on U . And hence f is Clarke regular on U: By Corollary 3.16,
f = g + C on U for some constant C which also means that
int(dom f) ⊆ int(dom g):
Now for any y ∈ cl(dom f); let x ∈ int(dom f): Then (1 − )x + y ∈ int(dom f). By
[15, Corollary 7.5.1],
f(y) = lim f((1 − )x + y)
→1−
= lim g((1 − )x + y) + C
−
→1
= g(y) + C:
Since g is l.s.c. convex and @c g(x) ⊆ @c f(x) ∀x ∈ R n , and for any x ∈ int(dom g),
@ g(x) is nonempty,
c
int(dom g) ⊆ int(dom @c g) ⊆ int(dom @c f) ⊆ int(dom f):
Thus,
int(dom f) = int(dom g):
And, hence,
cl(dom f) = cl(dom g):
Therefore for any y not in cl(dom f),
f(y) = ∞ = g(y):
The proof is completed.
References
[1] D. Aussel, J. Corvellec, M. Lassonde, Mean value property and subdi erential criteria for lower
semicontinuous functions, Trans. Amer. Math. Soc. 347 (1995) 4147– 4161.
[2] J. Borwein, Minimal cuscos and subgradients of Lipschtiz functions.
[3] J.M. Borwein, S.P. Fitzpatrick, J.R. Giles, The di erentiability of real functions on normed linear space
using generalized subgradients, J. Math. Anal. 128 (1987) 512–534.
[4] J.M. Borwein, X. Wang, Lipschitz functions on the line with prescribed Holder subdi erentials, preprint.
[5] F.H. Clarke, Optimization and Nonsmooth Analysis, Wiley-Interscience, New York, 1983; reprint,
Classics in Applied Mathematics, vol. 5, SIAM, Philadelphia, PA, 1990.
[6] F.H. Clarke, R.M. Redhe er, The proximal subgradient and constancy, Canad. Math. Bull. 36 (1993)
30–32.
[7] F.H. Clarke, R.J. Stern, P.R. Wolenski, Subgradient criteria for monotonicity, the Lipschitz condition,
and convexity, Canad. J. Math. 45 (1993) 1167–1183.
[8] J.B. Conway, A Course in Functional Analysis, Springer, New York, 1990.
[9] R. Correa, A. Jofre, Tangentially continuous directional derivatives in nonsmooth analysis, J. Optim.
Theory Appl. 61 (1989) 1–20.
976
Z. Wu, J.J. Ye / Nonlinear Analysis 39 (2000) 955 – 976
[10] R. Correa, L. Thibault, Subdi erential analysis of bivariate separately regular functions, J. Math. Anal.
Appl. 148 (1990) 157–174.
[11] P.D. Loewen, Optimal Control Via Nonsmooth Analysis, CRM Proc. & Lecture Notes, American
Mathematical Society, Providence, RI, 1993.
[12] B.S. Mordukhovich, Y. Shao, Nonsmooth sequential analysis in Asplund spaces, J. Math. Anal. Appl.
148 (1990) 157–174.
[13] R.A. Poliquin, Integration of subdi erentials of nonconvex functions, Nonlinear Anal. Theory, Methods
Appl. 17 (1991) 385–398.
[14] Liqun Qi, The maximal normal operator space and integration of subdi erentials of nonconvex functions,
Nonlinear Anal. Theory, Methods Appl. 13 (1989) 1003–1011.
[15] R.T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ, 1970.
[16] R.T. Rockafellar, Favorable classes of Lipschitz-continuous functions in subgradient optimization, in:
Progress in Nondi erentiable Optimization, E. Nurminski (Ed.), IIASA Collaborative Proceedings Series,
International Institute of Applied Systems Analysis, Laxenburg, Austria, 1982, pp. 125–144.
[17] L. Thibault, D. Zagrodny, Integration of subdi erentials of lower semicontinuous functions on Banach
spaces, J. Math. Anal. Appl. 189 (1995) 33–58.