[1] J. M. BORWEIN, ‘Continuity and differentiability properties of convex operators’, Proc. London Math. Soc. (3) 44 (1982) 420-444.
[2] J. M. BORWEIN, ‘Subgradients of convex operators’, Math. Operationsforsch. Statist. Ser. Optim. 15 (1984) 179-191.
[3] W. W. BRECKNER and G. ORBAN, ‘On the continuity of convex mappings, Rev. Anal. Numer. Theor. Approx. 6 (1977) 117-123.
[4] M. M. FEL’DMAN, ‘About the sufficient conditions of the existence of supporting operators to sublinear mappings’, Sibirsk. Mat. t . 16 (1975) 132-138 (Russian).
[5] V. L. LEVIN, ‘The subdifferentials of convex mappings and composed functional’, Sibirsk. Mat. It. 13 (1972) 1295-1303 (Russian).
[6] C. W. MCARTHUR, ‘ In what spaces is every closed normal cone regular?’, Proc. Edinburgh Math. Soc. (2) 17 (1970) 121-125.
[7] C. W. MCARTHUR, ‘Convergence of monotone nets in ordered topological vector spaces’, Studia Math. 34(1970) 1-16.
[8] A. B. NGMETH, ‘Some differential properties of convex mappings’, Mathematica (Cluj) 22 (45) (1980) 107-114.
[9] A. B. NEMETH, ‘Sequential regularity and directional differentiability of convex operators are equivalent’, preprint, Babes-Bolyai University 1984.
[10] A. B. NEMETH, ‘On some universal subdifferentiability properties of ordered vector spaces’, preprint, Babes-Bolyai University 1985.
[11] N. S. PAPAGFORGIU, ‘Nonsmooth analysis on partially ordered vector spaces: part 1—convex case’ Pacific J. Math 107 (1983) 403-485.
[12] A. L. PERESSINI, Ordered topological vector spaces (Harper & Row, New York 1967).
[13] H. SCHAFFER, Banach lattices and positive operators (Springer, Berlin 1974).
[14] M. VALADIER, ‘Sous-differentiabilite de fonctions convexes a valuers dans un espace vectoriel ordonne’ Math. Scand. 30 (1972) 65-74.
[15] J. ZOWE, ‘Subdifferentiability of convex functions with values in ordered topological vector spaces’ Math. Scand. 34 (1974) 69-83.
Paper (preprint) in HTML form
1987-Nemeth-, On the subdifferentiability of convex operators
ON THE SUBDIFFERENTIABILITY OF CONVEX OPERATORS
A. B. NÉMETH
Introduction
Since its inception in the papers of Valadier [14] and Levin [5], the theory of vector-valued convex analysis has been concerned with operators having values in order-complete vector lattices. The special interest of these spaces is motivated by the nice subdifferentiability properties of convex operators acting in such spaces. Besides considerable progress in this direction (see [11] and the references therein), there are results concerning more general ordered vector spaces. Zowe [15], and recently Borwein [1], have obtained results when conditions are imposed on the domain space of operators. Another approach is to build up the subgradients using directional minorants (or directional derivatives, when a topology is given). This method, initiated by Fel'dman [4] (see also [8,2]), has the advantage that conditions are imposed only on the range space. This approach suggests the natural question of how to characterize the ordered (topological) vector spaces in which every convex operator has directional minorants (directional derivatives) at each point and in every direction. Such a characterization can be done in terms of classical ordered (topological) vector spaces, and this is the principal result of the present note. It furnishes, using Fel'dman's theorem, a necessary and sufficient condition for every convex operator with values in an ordered vector space admitting isotone real functionals to have nice subdifferentiability properties.
The author expresses his gratitude to the referee and to J. M. Borwein for many valuable suggestions which improved the original version. Their remark, that the directional minorability of convex operators implies the Archimedian property of the space led to the present improved version of Theorem 1.
1. Prerequisites
An ordered vector space is, by definition, a pair ( Y,KY, K ), where YY is a real vector space and KK is a cone in it, that is, a subset having the properties (i) K+K sub KK+K \subset K, (ii) tK sub Kt K \subset K for each non-negative real number tt and (iii) K nn(-K)={0}K \cap(-K)=\{0\}. The order relation induced by KK in YY is defined by putting u <= vu \leqslant v whenever v-u in Kv-u \in K. Then <=\leqslant is a reflexive, transitive and antisymmetric relation which is invariant with respect to translations and to multiplication by non-negative real numbers. The cone KK is called the positive cone of the ordered vector space.
The ordered vector space ( Y,KY, K ) is Archimedian if y <= 0y \leqslant 0 whenever y <= tzy \leqslant t z for some z in Kz \in K and all t > 0t>0.
The space ( Y,KY, K ) is said to admit an isotone functional if there exists f:Y rarrRf: Y \rightarrow \mathbb{R} having the property that for any uu and vv in YY, the relations u <= v,u!=vu \leqslant v, u \neq v imply that f(u) < f(v)f(u)<f(v).
The space ( Y,KY, K ) is said to have the monotone sequence (monotone net) property if every decreasing sequence (net) in YY with a lower bound has an infimum.
If we suppose that YY is a topological vector space, then ( Y,KY, K ) will be called an ordered topological vector space. All the topological vector spaces we shall deal with are supposed Hausdorff.
The ordered topological vector space ( Y,KY, K ) is called normal if there is a base of neighbourhoods VV of 0 with
V=(V-K)nn(K-V).V=(V-K) \cap(K-V) .
The ordered topological vector space ( Y,KY, K ) is called regular (sequentially regular) if every decreasing net (sequence) having a lower bound has a limit in YY. If ( Y,KY, K ) is regular (sequentially regular) and if KK is closed, then ( Y,KY, K ) has the monotone net (sequence) property by [12, Corollary II.3.2]. Observe that in this case ( Y,KY, K ) is Daniell space. An ordered topological vector space is called (countably) Daniell if it has the monotone net (sequence) property and every decreasing net (sequence) with a lower bound converges to its infimum.
Let FF be an operator from a real vector space XX to (Y,K)(Y, K). Then FF is called convex if
F(tx_(1)+(1-t)x_(2)) <= tF(x_(1))+(1-t)F(x_(2))F\left(t x_{1}+(1-t) x_{2}\right) \leqslant t F\left(x_{1}\right)+(1-t) F\left(x_{2}\right)
for all x_(1),x_(2)x_{1}, x_{2} in xx and all tt in [0,1][0,1].
The directional minorant of FF at x_(0)x_{0} in the direction hh is defined by
grad F(x_(0);h)=i n f_(t > 0)t^(-1)(F(x_(0)+th)-F(x_(0)))\nabla F\left(x_{0} ; h\right)=\inf _{t>0} t^{-1}\left(F\left(x_{0}+t h\right)-F\left(x_{0}\right)\right)
when this infimum exists.
Suppose that ( Y,KY, K ) is an ordered topological vector space. Then the directional derivative of FF at x_(0)x_{0} in the direction hh is the limit
is increasing on R\\{0}\mathbb{R} \backslash\{0\} (see for example [5] or [14]). Hence, when KK is closed and F(x_(0);h)F\left(x_{0} ; h\right) exists, grad F(x_(0);h)\nabla F\left(x_{0} ; h\right) also exists (this follows from [12, Corollary II.3.2]).
Let L(X,Y)L(X, Y) be the space of linear operators from XX to YY. The set
del F(x_(0))={A in L(X,Y):Ax <= F(x_(0)+x)-F(x_(0))" for all "x in X}\partial F\left(x_{0}\right)=\left\{A \in L(X, Y): A x \leqslant F\left(x_{0}+x\right)-F\left(x_{0}\right) \text { for all } x \in X\right\}
is called the subdifferential of FF at x_(0)x_{0}. The elements of del F(x_(0))\partial F\left(x_{0}\right) are called subgradients of FF at x_(0)x_{0}.
For future reference we shall use Greek letters for alternatives in the following known result concerning the existence of directional minorants and directional derivatives.
Proposition 1. ( alpha\alpha ) If ( Y,KY, K ) is an ordered vector space with the monotone sequence property, then each convex operator with values in ( Y,KY, K ) has a directional minorant in every direction at every point of its domain.
( beta\beta ) If ( Y,KY, K ) is a sequentially regular ordered topological vector space, then each convex operator with values in ( Y,KY, K ) has a directional derivative in every direction at every point of its domain.
The proof is straightforward. The alternative ( alpha\alpha ) appears, for instance, in [14] and in [1, Proposition 3.7(a)].
Since (beta)(\beta) was not explicitly stated in this form (usually the normality of KK is also required-see for example [14, Theorem 5.1] or [1, Proposition 3.7(c)]) we give its proof.
Let FF be a convex operator from the vector space XX to (Y,K)(Y, K). Let x_(0)x_{0} and hh be fixed in XX. The operator phi\phi defined by (*) is increasing on R\\{0}\mathbb{R} \backslash\{0\}. If we assume that lim_(t↘0)phi(t)\lim _{t \searrow 0} \phi(t) does not exist, then we get a neighbourhood UU of 0 in YY for which we can construct a decreasing sequence (t_(n))\left(t_{n}\right) of real numbers converging to 0 such that
phi(t_(2k-1))-phi(t_(2k))!in U quad" for each "k.\phi\left(t_{2 k-1}\right)-\phi\left(t_{2 k}\right) \notin U \quad \text { for each } k .
Now, since the sequence (phi(t_(n)))\left(\phi\left(t_{n}\right)\right) is decreasing and F(x_(0))-F(x_(0)-h)F\left(x_{0}\right)-F\left(x_{0}-h\right) is a lower bound for it, we get a contradiction with the hypothesis.
Remark. For some ordered topological vector spaces, regularity implies normality. McArthur [7] has shown that every closed regular cone in a Fréchet space is normal. In the case of locally convex spaces he gave in [6] conditions in order that every closed normal cone be regular. There exist ordered normed spaces which are regular but lack normality [9].
2. Main results
We state first our principal results in the form of two theorems.
Theorem 1. Let (Y,K)(Y, K) be an ordered vector space and let XX be a vector space of dimension greater than or equal to 1 . The following assertions are equivalent.
(i) (Y,K)(Y, K) has the monotone sequence property.
(ii) Each convex operator F:X rarr(y,K)F: X \rightarrow(y, K) possesses a directional derivative in every direction at every point.
(iii) Each convex operator F:X rarr(Y,K)F: X \rightarrow(Y, K) possesses a directional minorant at 0 in some non-zero direction.
THEOREM 2. Let (Y,K)(Y, K) be an ordered topological vector space and let XX be a vector space of dimension at least 1 . Then the following assertions are equivalent.
(i) (Y,K)(Y, K) is sequentially regular.
(ii) Each convex operator F:X rarr(Y,K)F: X \rightarrow(Y, K) possesses a directional derivative in every direction at every point.
(iii) Each convex operator F:X rarr(Y,K)F: X \rightarrow(Y, K) possesses a directional derivative at 0 in some non-zero direction.
We shall follow in the proofs a schema proposed by J. M. Borwein (which simplifies essentially our original version in [9]). Consider first some auxiliary results.
Proposition 2. Let ( a_(n)a_{n} ) and ( y_(n)y_{n} ) be sequences in R\mathbb{R} and YY, respectively with a_(n) > a_(n+1)a_{n}>a_{n+1} for each n inNn \in \mathbb{N}. In order that a convex operator f:Rrarr Yf: \mathbb{R} \rightarrow Y such that f(a_(n))=y_(n)f\left(a_{n}\right)=y_{n} exists, it is necessary and sufficient that
f(t)=max_(k inN)f_(k)(t)=f_(n)(t)quad" if "t in[a_(n+1),a_(n)]f(t)=\max _{k \in \mathbb{N}} f_{k}(t)=f_{n}(t) \quad \text { if } t \in\left[a_{n+1}, a_{n}\right]
defines a convex operator with f(a_(n))=y_(n)f\left(a_{n}\right)=y_{n}.
Proposition 3. Let (u_(n))\left(u_{n}\right) be a sequence in KK with u_(n+1) <= t_(n)u_(n)u_{n+1} \leqslant t_{n} u_{n} for some t_(n)t_{n} in ( 0,1 ). Then one can select a sequence ( a_(n)a_{n} ) in R\mathbb{R} decreasing to zero such that the operator ff defined on {a_(n)}\left\{a_{n}\right\} by f(a_(n))=a_(n)u_(n)f\left(a_{n}\right)=a_{n} u_{n} has a convex extension to R\mathbb{R}.
Proof. Inductively, suppose that a_(1),dots,a_(n)a_{1}, \ldots, a_{n} have been selected. By Proposition 2 it is sufficient to find rr small enough so that
Take rr sufficiently small in order to have r//(a_(n)-r) <= delta_(n)r /\left(a_{n}-r\right) \leqslant \delta_{n} as well; then Delta_(n-1) <= Delta_(n)(r)\Delta_{n-1} \leqslant \Delta_{n}(r) and 0 < a_(n+1) < a_(n)0<a_{n+1}<a_{n}, as desired.
Lemma. If each convex operator f:X rarr(Y,K)f: X \rightarrow(Y, K) possesses a directional minorant at 0 in some non-zero direction, then ( Y,KY, K ) is Archimedian.
Proof. We shall show that if (Y,K)(Y, K) is not Archimedian, then there exists a convex operator from R\mathbb{R} to ( Y,K\mathrm{Y}, \mathrm{K} ) without directional minorant at 0inR0 \in \mathbb{R} in the direction 1inR1 \in \mathbb{R}. Observe first that ( Y,KY, K ) is Archimedian if and only if each set of the
form {tx:t > 0}\{t x: t>0\} with x in Kx \in K has an infimum. Indeed, suppose that vv is an infimum of the set {tx:t > 0}\{t x: t>0\}, where x in Kx \in K. Then v >= 0v \geqslant 0 and sx >= rvs x \geqslant r v for arbitrary s > 0s>0 and r > 0r>0. If we fix r=2r=2, then it follows that 2v2 v is a lower bound for the set {tx:t > 0}\{t x: t>0\}, thus v >= 2vv \geqslant 2 v, that is, v <= 0v \leqslant 0. Hence if every set of this kind has an infimum, it must be 0 . But then, if tx >= yt x \geqslant y for some x in Kx \in K and every t > 0t>0, it follows that y <= 0y \leqslant 0, that is, (Y,K)(Y, K) is Archimedian. The converse is immediate. Now assuming that ( Y,KY, K ) is not Archimedian and considering a set {tx:t > 0}\{t x: t>0\}, where x in Kx \in K, without infimum, we define f:Rrarr(Y,K)f: \mathbb{R} \rightarrow(Y, K) by putting f(t)=0f(t)=0 for t <= 0t \leqslant 0 and f(t)=t^(2)xf(t)=t^{2} x for t > 0t>0. Then ff is obviously convex and the set {t^(-1)(f(t)-f(0))=tx:t > 0}\left\{t^{-1}(f(t)-f(0))=t x: t>0\right\} has no infimum, that is, grad f(0;1)\nabla f(0 ; 1) does not exist.
Proof of Theorem 1. Clearly (ii) implies (iii), while (i) implies (ii) by Proposition 1(alpha)1(\alpha). To show that (iii) implies (i) we argue as follows. Let (v_(n))\left(v_{n}\right) be a decreasing sequence in KK. Let u_(n)=v_(n)(n+1)//nu_{n}=v_{n}(n+1) / n. Then 0 <= u_(n) <= (1-1//n^(2))u_(n-1)0 \leqslant u_{n} \leqslant\left(1-1 / n^{2}\right) u_{n-1} for each nn, and Proposition 3 applies. Let F(t)=f(|t|)F(t)=f(|t|) with ff constructed as in Proposition 3. Then d=grad F(0;1)d=\nabla F(0 ; 1) exists by hypothesis and
d=i n f_(n inN)a_(n)^(-1)(F(a_(n))-F(0))=i n f_(n inN)u_(n).d=\inf _{n \in \mathbb{N}} a_{n}^{-1}\left(F\left(a_{n}\right)-F(0)\right)=\inf _{n \in \mathbb{N}} u_{n} .
Since ( Y,KY, K ) must be Archimedian by the preceding lemma, dd also will be an infimum of (v_(n))\left(v_{n}\right). Indeed, we have d <= (1+1//n)v_(m)d \leqslant(1+1 / n) v_{m} for all nn and mm. Fix mm for the moment. Then d-v_(m) <= v_(m)//nd-v_{m} \leqslant v_{m} / n for all nn and hence d <= v_(m)d \leqslant v_{m}, that is, dd is a lower bound for (v_(m))\left(v_{m}\right). Every other lower bound vv of (v_(n))\left(v_{n}\right) will be a lower bound for (u_(n))\left(u_{n}\right) too, and hence v <= dv \leqslant d. This completes the proof.
Proof of Theorem 2. Clearly (ii) implies (iii), while (i) implies (ii) by Proposition 1(beta)1(\beta). For the proof that (iii) implies (i) we consider a decreasing sequence (v_(n))\left(v_{n}\right) in KK. Then, since u_(n)=v_(n)(n+1)//nu_{n}=v_{n}(n+1) / n, we can construct by Proposition 3 a convex function F:R rarr YF: R \rightarrow Y such that F(0)=0F(0)=0 and F(a_(n))=a_(n)u_(n)F\left(a_{n}\right)=a_{n} u_{n} with some sequence (a_(n))\left(a_{n}\right) decreasing to zero in R\mathbb{R}. If we consider now
which certainly exists, we deduce that lim_(n rarr oo)u_(n)\lim _{n \rightarrow \infty} u_{n} exists. But v_(n)=(n//(n+1))u_(n)v_{n}=(n /(n+1)) u_{n} has obviously the same limit, and so Theorem 2 is proved.
3. Application to subdifferentiability
Let XX be a vector space and let ( Y,KY, K ) be an ordered (topological) vector space. Suppose that F:X rarr(Y,K)F: X \rightarrow(Y, K) is a convex operator. We shall say that FF is fully subdifferentiable at x_(0)in Xx_{0} \in X if (i) grad F(x_(0);*)\nabla F\left(x_{0} ; \cdot\right) (or F(x_(0);*)F\left(x_{0} ; \cdot\right) ) is defined on XX, (ii) del F(x_(0))\partial F\left(x_{0}\right) is non-empty, and (iii) for every hh in XX one has the relation grad F(x_(0);h)=max{Ah:A in del F(x_(0))}(:}\nabla F\left(x_{0} ; h\right)=\max \left\{A h: A \in \partial F\left(x_{0}\right)\right\}\left(\right. or {:F(x_(0);h)=max{Ah:A in del F(x_(0))})\left.F\left(x_{0} ; h\right)=\max \left\{A h: A \in \partial F\left(x_{0}\right)\right\}\right).
As a consequence of a result of Fel'dman [4] (see also [8,2][8,2] ) combined with some continuity properties of convex operators in [3,1][\mathbf{3}, \mathbf{1}] we have the following.
Proposition 4. Let FF be a convex operator from the vector space XX to the ordered vector space ( Y,KY, K ) with the monotone net property or to the ordered topological vector space ( Y,KY, K ) which is Daniell. Then FF is fully subdifferentiable at each point of XX.
If XX is a topological vector space, ( Y,KY, K ) is an ordered normal topological vector space which is Daniell, and if FF is continuous at x_(0)x_{0}, then it is fully subdifferentiable at x_(0)x_{0} and all the maps in the definition of the full subdifferentiability are continuous.
By using Theorems 1 and 2 we can give the following characterization of the ordered vector spaces in which all the convex operators are fully subdifferentiable.
^(^(')){ }^{'} Theorem 3. Let ( Y,KY, K ) be an ordered vector space (respectively an ordered topological vector space with closed positive cone) which admits an isotone functional. Then the following assertions are equivalent.
(i) ( Y,KY, K ) has the monotone sequence property (respectively is countable Daniell).
(ii) Every convex operator FF from a vector space XX to (Y,K)(Y, K) has directional minorants (respectively directional derivatives) in every direction at every point in XX.
(iii) Every convex operator FF from a vector space XX to (Y,K)(Y, K) is fully subdifferentiable at each point of XX.
Proof. Condition (i) implies condition (ii) by Proposition 1. To verify that (ii) implies (iii) we observe that in the presence of an isotone functional the monotone sequence property (respectively the countable Daniell property) implies the monotone net property (respectively the Daniell property), the reasoning being similar to that in [13, Proposition II.4.9]. Hence Proposition 4 can be used. Finally (iii) implies (i) according Theorem 1 (respectively Theorem 2).
Remark. In [4,2][4,2] it was shown that the condition del F(x)!=O/\partial F(x) \neq \varnothing for each convex operator FF with values in ( Y,KY, K ), called the subgradient property of ( Y,KY, K ), does not imply that the space has the monotone sequence property. In the given counterexamples ( Y,KY, K ) is not Archimedian. Our results say nothing about the relationship between the monotone sequence property and either the subgradient or the Archimedian properties. (For other problems concerning the subgradient property see [10].)
References
J. M. Borwein, 'Continuity and differentiability properties of convex operators', Proc. London Math. Soc. (3) 44 (1982) 420-444.
J. M. Borwein, 'Subgradients of convex operators', Math. Operationsforsch. Statist. Ser. Optim. 15 (1984) 179-191.
W. W. Breckner and G. Orbán, 'On the continuity of convex mappings, Rev. Anal. Numér. Théor. Approx. 6 (1977) 117-123.
M. M. Fel'dman, 'About the sufficient conditions of the existence of supporting operators to sublinear mappings', Sibirsk. Mat. Ž. 16 (1975) 132-138 (Russian).
V. L. LEVIN, 'The subdifferentials of convex mappings and composed functionals', Sibirsk. Mat. Ž. 13 (1972) 1295-1303 (Russian).
C. W. McArthur, 'In what spaces is every closed normal cone regular?', Proc. Edinburgh Math. Soc. (2) 17 (1970) 121-125121-125.
C. W. McArthur, 'Convergence of monotone nets in ordered topological vector spaces', Studia Math. 34 (1970) 1-16.
A. B. Németh, 'Some differential properties of convex mappings', Mathematica (Cluj) 22 (45) (1980) 107-114.
A. B. Németh, 'Sequential regularity and directional differentiability of convex operators are equivalent', preprint, Babes-Bolyai University 1984.
A. B. Németh, 'On some universal subdifferentiability properties of ordered vector spaces', preprint, Babes-Bolyai University 1985.
N. S. Papagforgiu, 'Nonsmooth analysis on partially ordered vector spaces: part 1-convex case', Pacific J. Math 107 (1983) 403-485.
A. L. Peressina, Ordered topological vector spaces (Harper & Row, New York 1967).
H. Schaffer, Banach lattices and positive operators (Springer, Berlin 1974).
M. Valadier, 'Sous-differentiabilité de fonctions convexes à valuers dans un espace vectoriel ordonné', Math. Scand. 30 (1972) 65-74.
J. ZowE, 'Subdifferentiability of convex functions with values in ordered topological vector spaces', Math. Scand. 34 (1974) 69-83.
Institutul de Matematica CP 68
3400 Cluj-Napoca
Romania
Received 21 October 1985.
1980 Mathematics Subject Classification 47B55.