A construction of a maximal monotone extension of a monotone map

Jean-Pierre CROUZEIX, Eladio Oca na Anaya Wilfredo Sosa Sandoval January 18, 2006

Abstract A proof based on the axiom of choice shows that any monotone map has maximal monotone extensions but this proof is not constructive. In this paper, we give a construction of such an extension. The process is based on some generic properties of (maximal) monotone maps given before.

Key words: Multivalued maps, Monotonicity, Maximal Monotonocity. AMS classication: 47H05, 47H04

Introduction

The history of (maximal) monotone multi-valued maps started in the sixties with the pioneering works of Minty [9, 10, 11]. A classical survey on the monotonicity of maps dened on Hilbert spaces is the book by Brezis [3]. Very recent surveys are due to Auslender and Teboulle [1] and Rockafellar
LIMOS, Universit e Blaise Pascal, France. Instituto de Matematica y Ciencias Anes, Lima, Per u, the research of this author has been done when preparing his Ph.D. in Universit e Blaise Pascal, France. Instituto de Matematica y Ciencias Anes, Lima, Per u.

and Wets [17]. In these books, the readers will nd the most important results and a comprehensive bibliography. In the present paper we consider multi-valued maps dened on nite dimensional spaces. In contrast with the classical literature, we tackle the monotonicity via the graphs of the maps instead of the maps themselves. We think that, in the same manner that the real essence of the convexity of a function is more present in the convexity of its epigraph than in the analytic denition, the true nature of monotonicity is more in the graphs than in the maps. Any monotone map has a maximal monotone extension. The classical proof uses the axiom of choice (or the Zorns lemma) and thereby is not constructive. In the last section, we show how to build a maximal monotone extension of a monotone map), and thereby we obtain a constructive proof of the existence. In the previous sections, after a short background on (maximal) monotonicity, we establish some basic properties on maximal monotone extensions used in the construction. In the paper X is a nite dimensional space and X is its dual space. Of course X can be identied to X , nevertheless we prefer to distinguish X and X in order to set in evidence their two dierent roles. Given a closed convex set C of X and x C , we denote respectively by TC (x) and NC (x) the tangent and the normal cone at x to C .

Denitions and basic results

A set F X X is said to be monotone if

and it is said maximal monotone if for any monotone subset G of X X containing F we have F = G. Let us associate with F the set F X X dened by F := { (x, x ) X X : x y , x y 0 for all (y, y ) F }. 2

By construction, F is closed since it is an intersection of closed subsets. The following properties are direct consequences of the denitions. Proposition 2.1 Assume that F and G are two subsets of X X . a) If F G then G F . b) F is monotone if and only if F F . c) If F G and G is monotone then F is monotone and G F . d) F is maximal monotone if and only if F = F . e) F is monotone if and only if cl (F) is monotone. f ) If F is maximal monotone then it is closed. In particular, b) and c) say that if F is monotone then F contains all monotone extensions of F and in particular all the maximal monotone extensions. In association with G X X , we introduce the multivalued maps and dened respectively on X and X by (x) := {x X : (x, x ) F } and (x ) := {x X : (x, x ) F }. Thus G can be considered as the graph of both and . If G is (maximal) monotone, then and are said (maximal) monotone. In connection with the construction of G from G, we associate with and the maps (x) := {x X : (x, x ) G} and (x ) := {x X : (x, x ) G}. The domains of and are the projections of G on X and X , dom () := {x : (x) = } = proj X (G), dom ( ) := {x : (x ) = } = proj X (G). Clearly, = and, thanks to Proposition 2.1, the equality = (equivalently = ) holds if and only if the multivalued map is maximal monotone. 3

Proposition 2.2 Let G X X . Then, for any (x, x ) X X , the sets (x) and ( )(x ) are closed and convex. In particular, (x) and (x ) are closed and convex when G is maximal monotone. Proof. By denition, (x) =
(y,y )G

{ x : x y , x y 0 }

and ( )(x ) =
(y,y )G

{ x : x y , x y 0 } .

These sets are closed and convex as intersections of half spaces. Given A X , we denote by co (A) and co (A) the convex hull and the closed convex hull of A. Next, given a multivalued map : X X , we dene the multivalued maps co and co by (co )(x) = co ((x)) and (co )(x) = co ((x)) for all x X . The following result is a direct consequence of Proposition 2.1. Proposition 2.3 If one of the maps , co and co is monotone, so are the other ones. Also Proposition 2.3 implies that if G is maximal monotone, then (x) and (x ) are closed and convex for all x X, x X .

Some properties of

In complement with the last section, we give less immediate results on . The rst of them is well known when = , i.e., when is maximal monotone. Its proof is quite easy. Proposition 3.1 Let : X X be a monotone multivalued map and C = co dom (). Assume that x dom () C. Then, NC ( x) coincides with the recession cone of ( x), i.e., NC ( x) = (( x)) . 4

It follows that (x) = (x) + NC (x) for all x C. Proof. By denition w NC ( x) if and only if x + tw y , x y 0 for all (y, y ) graph (), t 0 and x ( x). Thus, NC ( x) consists of the vectors w such that x + tw ( x) for all x ( x) and t 0. This condition means that NC ( x) coincides with the recession cone of the closed convex set (x). For the forthcoming results, we need the following technical lemma. Lemma 3.1 Let D = co {x0 , x1 , , xn } be a n-dimensional simplex of R n , V be a closed convex subset of int (D). Then for any x V and any c R n the following linear program
n n

inf [
u i=0

ui : ui 0 and
i=0

ui (xi x) = c ]

(lpl)

is feasible and has one unique optimal solution. Denote by u(c, x) this optimal solution, then the function (c, x) u(c, x) is continuous on X V . Proof. Because x V int (D), the n vectors (xi x), i = 1, 2, , n are linearly independent. Hence there exist uniquely dened i (x) > 0, i = 1, 2, , n, such that
n

(x x0 ) =
i=1

i (x)(xi x).

Also, there are uniquely dened i (c, x) R , i = 1, 2, , n such that

c=
i=1

i (c, x)(xi x).

Furthermore the functions i and i are continuous on V and R n V respectively. 5

Thus, u is feasible for (lpl) if and only if u = (u0 , u1 , , un ) 0 and i (c, x) = ui u0 i (x), i = 1, 2, , n or equivalently u0 0 and i (c, x) + u0 i (x) = ui 0, i = 1, 2, , n. Then (lpl) is reduced to the one variable linear program
n

inf [ u0 (1 +
u0 i=1

i (x)) : u0 0, u0

i (c, x) , i = 1, 2, , n ]. i ( x )

(lpl1 )

Since 1 + n i=1 i (x) > 0, (lpl1 ) has one unique optimal solution, this is also the case for (lpl). The optimal solution of (lpl) is u0 (c, x) = max{0, max [
i

ui (c, x) = u0 (c, x)i (x)

i (c, x) : i = 1, 2, , n ]}, i (x) for i = 1, 2, , n.

The continuity of the function (c, x) u(c, x) is a consequence of the continuity of the functions i and i . In the following results we assume that the interior of the convex hull of the domain of is not empty. These results will be generalized in the next section with the relative interior in place of the interior. Theorem 3.1 Let : X X be monotone. Denote by D the interior of the convex hull of the domain of . Then for all x D there exists a compact K X and a neighbourhood V of x such that = (x) K for all x V . It follows that D is contained in the interior of the domain of . Proof. i) Assume that X = R n . Denote by G the graph of . Since x i > 0 and (xi , x belongs to D, there exist t i ) G for i = 0, 1, , n such that
n

x =
i=0

i xi , 1 = t 6

n i=0

i t

and the n vectors (xi x0 ), i = 1, 2, , n, are linearly independent . Let some > 0 be such that
n

i for i = 0, 1, , n. Let V be dened by <t

V = {x =
i=0

t i xi : 1 =
i=0

ti

and

ti for all i}.

Then V is a compact convex neighbourhood of x and is contained in D. n Given c R and x V , let us dene (c, x) = sup [ c, x : x (x) ]
x

with by convention (c, x) = if (x) = . Then, (c, x) (c, x) where (c, x) = sup [ c, x : x , xi x x i , xi x , i = 0, 1, 2, , n ].
x

(Pe )

The dual of the linear program (Pe ) is the problem

(c, x) = inf [ u

ui x i , xi x : ui 0 and

ui (xi x) = c ].
i=0

(De )

i=0

The feasible set of (De ) is the same that the feasible set of the problem (lpl) considered in Lemma 3.1. Hence, (De ) is feasible, there is no duality gap between the two problems problems (Pe ) and (De ), and
n

(c, x) = (c, x) (c, x) =

ui (c, x) x i , xi x

where u(c, x) is the optimal solution of the problem of (lpl). Next, dene M = sup [ (c, x) : x V, c 1 ] and K = { x : x M }.
x,c

Since V is compact and the function u is continuous on R n V , M is nite and K is a compact convex set. 7

Then for all c with c 1 and all x V , one has (c, x) (c, x) (c, x) M. Dene the quantities A(x) = sup[ x : x (x) ],
x x

(1)

B (x) = sup[ x : x x i , xi x 0, Then,

i = 0, 1, , n ].

A(x) = sup [ c, x : x (x), c 1 ] = sup[ (c, x) : c 1]

and B (x) = sup[ (c, x) : c 1].

The inequalities A(x) B (x) M for all x V deduced from (1) imply (x) {x : x , xi x x i , xi x , i = 0, 1, , n} K y ) = . Then, ii) Assume for contradiction that y V and ( y) = = (
(x,x )F

(2)

{y R n : y x , y x 0} K.

Since K is compact, there exist a positive integer q and points (xj , x j) F for j = n + 1, n + 2, , n + q such that

j =n+q

{ y : y x j , y xj 0} K.

= Next, in view of (2),

j =n+1

j =n+q

=
j =0

{y : y x j , y xj 0}.

(3)

Consider the n (n + q +1) matrix Q with columns (y xi ), i = 0, 1, n + q , and the vector a R n+q+1 with components aj = x j , y xj , then (3) is equivalent to the following assertion There is no y R n such that Qt y a. 8

In view of a theorem on alternatives, this condition is equivalent to u R n+q+1 such that u 0, Qu = 0 and a, u > 0. Without loss of generality, we assume that y = ui xi and
n+q n+q

ui = 1. Then Qu = 0 implies

a, u a, u

=
j =0

uj x j,
i=0 n+q

ui x i x j ,

=
i,j =0

u i uj x i xj , x i xj .

Because the (xi , x i ) belong to the graph of which is monotone, one has xi xj , xi xj 0 for all i, j in contradiction with u 0 and a, u > 0. Let us apply this theorem to the particular case where is maximal monotone. Then = and we recover the fundamental result which says that (x) is locally bounded and not empty on the interior of the domain of . The next theorem also generalizes a result known for maximal monotone maps to monotone maps. Theorem 3.2 Let : X X be monotone. Assume that the interior of the convex hull of its domain is not empty. Denote by D the domain of . Then the sets int (D) and cl (D) are convex. Moreover cl (D) = cl (int (D)) and int (D) = int (cl (D)). Proof. Fix some a in the interior of the convex hull. i) Let us prove that x cl (int (D)) for any x cl (D). Consider some . For each k , consider some x sequence {xk }k D converging to x k (xk ). Next consider the map k dened from its graph as follows graph (k ) = graph () {(xk , x k )}. Then k is monotone, dom (k ) = dom () {xk } and k . Let yk = 1 xk + k (a xk ), then by construction, xk dom (k ), the sequence {yk }k 9

converges to x and is contained in the interior of the convex hull of domain of k . It follows from Theorem 3.1 that yk belongs to the interior of the domain of k and therefore to the interior of the domain of . ii) Let us prove that int (D) is convex. Let x , y int (D) and (0, 1), we must prove that z = x + (1 ) y int (D). There exists > 1 such that x = a + ( x a) and y = a + ( y a) belong to int (D). Then z = a + ( z a) = x + (1 )y . Let x (x) and y (y ). We distinguish two cases. ii-a) First case: x y , x y 0. Take : X X such that graph () = graph () {(x, x )} {(y, y )}. By construction, is monotone and contains and applying Theorem 3.1 we obtain the state. ii-b) Second case: x y , x y < 0. We shall prove that z = x + (1 )y (z ) that is to say that A = z w , z w 0 for any (w, w ) graph (). Let us compute A. A = (x y ) + (y w ), (1 )(y x) + (x w) ; + (1 ) y w , y w + w x + y w , x w ; A = (1 ) x y , y x + x w , x w + + (1 ) y w , y w . The assumptions imply A 0, hence (z, z ) graph (). Next, dene : X X by graph () = graph () {(z, z )}. Then, is monotone, contains and z belongs to the interior of the convex hull of its domain. Proceeds as above. iii) The other claims of the theorem are direct consequences of i) and ii). Now, we turn our interest in some generic properties of monotone maps. For that, we introduce the following denition. 10

A = (1 ) x y , y x + x w + w y , x w +

Denition 3.1 Given : X X and S dom (), we dene the map S : X X from its graph graph (S ) = cl [graph () (S X )]. It is clear that if is monotone, so are S and co (S ). In the following results we denote by C the closed convex hull of dom () and we assume that the interior of C is not empty. The rst result considers points in the interior of C and the next two results points in the boundary. Theorem 3.3 Let : X X be monotone and S dom (). Let V C be open, convex and such that cl (V S) = cl (V). Then, and co S coincide on V . It follows rst that is monotone on V , next that any maximal monotone map containing coincides with on V . Proof. i) Let x V . Since ( x) co S ( x) and co S is monotone, one has co (S ( x)) ( x). We prove the reverse inclusion by contradiction. As x)). In view of separation sume that a ( x) exist such that a / co (S ( theorems, there exists a vector d, d = 1, such that sup[ d, a : co (S ( x)) ] < 0. (4)

Theorem 3.1 implies that a compact set K X and an open neighborhood W V of x exist such that (x) K for all x W . Since cl (V S) = cl (V), there exist a sequence of vectors {dk } X converging to d and a sequence of positive real numbers {tk } converging to 0 such that xk = x + tk dk S W for all k . Since = (xk ) K for all k and K is compact, we can nd x K and a subsequence {(xkl , x x, x ). Without loss of kl )} converging to ( generality, we assume that the whole sequence converges. Then x S ( x). Next, since a ( x) and the sequence {(xk , xk )} is contained in graph (), one has 1 0. x x a , xk x k a , dk = tk k Passing to the limit when k goes to +, we obtain x a , d 0, in contradiction with (4). 11

ii) The monotonicity of implies that the graph of co (S ) is monotone, hence i) implies that is monotone on V . Proposition 3.2 Let : X X be monotone. Assume that x bd (C) and there exist S dom () and an open neighborhood V of x such that cl (V S) = cl (V C). Assume in addition that there exist d int (TC ( x)) and a sequence {(xk , xk )}k (S X ) graph () such that xk = x + tk dk for all k , tk 0+ , dk d and xk + when k +. Then, )}k graph () such S ( x) = ( x) = . Moreover, for any sequence {(yk , yk that the sequence {yk }k converges to x , yk goes to + when k goes to +. Proof. i) Since S ( x) ( x), it is enough to show that ( x) = . Assume, for contradiction, that there is some x ( x). Without loss of generality, x k } converges to some w . It results we assume that the whole sequence { x k from the assumptions that, for all k , x x 1 k , dk = xk xk t k x k Passing to the limit, one obtains 0. w , d (5) x , xk x 0. k x

On the other hand, for all (z, z ) graph () with z int (C) and for all k , one has x k z , xk z 0. Hence, passing to the limit, one obtains < 0 because w , x z 0. One deduces that w NC ( x). Then w , d int (TC ( w = 1 and d x)), in contradiction with (5). ii) If yk does not go to + when k goes to +, then there are y and a subsequence of {(yk , yk )}k converging to ( x, y ). Then ( x, y ) graph (), in contradiction with the rst part of the theorem. Theorem 3.4 Let : X X be monotone. Assume that x bd (C) and that there exist a subset S dom () and an open neighborhood V of x satisfying cl (V S) = cl (V C). Then, ( x) = co (S ( x))) + NC ( x). 12

Proof. The inclusion co (S ( x)) + NC ( x) ( x) is a consequence of Propositions 2.2 and 3.1. Let us prove the reverse inclusion. Firstly, we prove that co (S ( x)) + NC ( x) is closed. Since we are faced with the sum of two closed convex sets it is sucient to prove that (see [13]) NC ( x) [co (S ( x))] = 0. Note that the inclusion ( x) co (S ( x))) and Proposition 3.1 imply [co (S ( x))] (( x)) = NC ( x). On the other hand NC ( x) NC ( x) = {0} because int (C) = . Therefore NC ( x) [co (S ( x))] NC ( x) NC ( x) = {0}. Next, assume, for contradiction, that there exists x ( x) such that x / co (S ( x)) + NC ( x). Since co (S ( x)) + NC ( x) is closed and convex, in view of separation theorems, a vector d, d = 1, exists such that x + x : x S ( x sup[ d, x) , x x) ] < d, . 1 2 1 2 NC ( Since NC ( x) is a cone, x : x NC ( 0 = sup [ d, x) ] , 2 2 TC ( which says that d x). Next, (6) and (7) imply x x sup [ d, : x x) ] < 0. 1 1 S ( Since int (C) is not empty, there exists d int (TC ( x)) such that sup [ d, x : x x) ] < 0. 1x 1 S (
Since d int (TC ( x)), there is a sequence {(xk , x k )} (S X ) graph () such that tk 0+ , dk d and xk = x + tk dk int (C) for all k . Since ( x) is not empty, Proposition 3.2 implies that the sequence {xk } is bounded. Without loss of generality it can be assumed that the whole sequence {x k} converges to some x . Then x S ( x) ( x).

(6)

(7)

(8)

13

Then the denition of implies x x k , dk = 1 x x 0 for all k. k , xk x tk

Passing to the limit, one obtains x x , d 0 in contradiction with (8). The combination of theorems 3.3 and 3.4 give birth to the following result. Theorem 3.5 Let : X X be monotone. Assume that the interior of C = co (dom ()) is not empty and there exists S dom () such that cl (S) = C. Then, the multivalued map : X X dened by
co ( (x)) + N (x) S C (x) =

if x C, if x / C,

is the unique maximal monotone map containing with domain contained in C . Proof. Let x C . Then theorems 3.3 and 3.4 imply (x) = (x). Hence (x) = (x). It remains to prove that (x) = for all x / C . Assume, for contradic tion, that there is (a, a ) graph () with a / C . Dene : X X by graph () = graph () {(a, a )}. Then is monotone. (0, 1) be such that c = b+t (ab) bd (C). Denote Let b int (C) and t by D the convex hull of C {a}. Then c int (D). Therefore is bounded in a neighbourhood of c. If (c) is not empty, then (c) = (c) + NC (c) which is unbounded. If (c) is empty, then (b + t(a b)) = (b + t(a b)) becomes with t < t . In both cases we have a contradiction. unbounded when t t Thus coincides with and therefore is maximal monotone. It coincides with on C . We know that contains any monotone map containing . We deduce that is unique maximal monotone map containing with domain contained in C . Remark. This theorem says that if a monotone map is known on a dense set, it has an unique maximal monotone extension on the closure of the 14

set. As a corollary, we recover the fact that the subdierential of a convex function can be recovered from the gradients of the function at points where it is dierentiable.

Dealing with relative interiors

This short section extends the results of section 3 to the case where the interior of the convex hull of the domain of : X X is empty. Through the section we assume that the ane hull of dom () is of the form a (dom ()) = a + X1 where X1 is a linear subspace of X and a dom (). We denote by X2 the orthogonal space to X1 , X2 = {x X : x , x = 0 for all x X1 }. Here again, we distinguish the dierent roles played by the primal and dual variables in denoting by X1 and X2 the dual spaces of X1 and X2 . We associate with the map : X1 X1 dened by
(x1 ) = {x 1 X1 : x2 X2 with x1 + x2 (a + x1 )}.

The proof of the following result is easy, it is left to the reader. Proposition 4.1 is monotone if and only if is monotone. is maximal monotone if and only if is maximal monotone and (a + x1 ) = (x1 ) + X2 for all x1 X1 . By the way of this proposition the results of section 3 are easily extended to (maximal) monotone maps . In particular we straightforwardly recover the following results. Proposition 4.2 Let : X X be maximal monotone. Then the relative interior and the closure of its domain are convex. Moreover the closure of the domain coincides with the closure of its relative interior. 15

Proposition 4.3 Let : X X be monotone. Denote by C the closed convex hull of its domain. Then, for all x ri (C), there is a neighbourhood V of x and a compact set K X1 such that for all x V C it holds
= (x) = (x) + X2 K + X2 .

A construction of a maximal monotone extension

In this section, we show how to construct a maximal monotone extension of a monotone map : X X (or equivalently a maximal extension of its graph G). Unlike the traditional proof on the existence of such an extension which is based on the axiom of choice, we give a constructive proof based on the generic properties seen in Section 3. Of course if a subset G of X X is not monotone, it has no maximal monotone extension. In the other case, the extension is not unique as seen from the following example. Example 5.1 The set G = {(0, 0), (0, 1)} R 2 . The sets G1 = {0} R and G2 = (, 0] {0} {0} [0, 1] [0, ) {1} are two maximal monotone extensions of G. In our construction we have chosen to work on , but the same construction can be done in working on . One get dierent maximal monotone extensions. Let : X X be monotone, let C be the closure of the convex hull of its domain. Assume in a rst time that the interior of C is not empty. Consider some countable set S = {x0 , x1 , , xk , } int (C) such that cl (S) = C, such sets S exist. The rst part of the construction is described in the following algorithm. Algorithm:

16

 Step 0: Dene 0 : X X as
(x) + N (x) C 0 (x) =

if x C, if not .

By construction 0 is monotone and dom (0 ) = dom (). Step k: In the previous step a monotone map k has been obtained with graph (k ) graph (). If dom (k ) S, the multivalued map : x k (x) + NC (x) is maximal monotone (Theorem 3.5). STOP. Otherwise, take p(k ) = min[ p IN : xp / dom (k ) ], and dene a new map k+1 as follows k ( x) k+1 (x) = k (xp(k) ) By construction, k+1 is monotone, dom (k+1 ) = dom (k ) {xp(k) } C, graph () graph (k ) graph (k+1 ), and p(k + 1) p(k ) + 1. Do k = k + 1 and go back to step k . End of algorithm. In the second part of the construction, we consider D = k dom (k ). Since S D C one has cl (D) = C. Next, dene : X X by
(x) k (x) =

if x dom (k ), if x = xp(k) , otherwise .

if x dom (k ), if x / D. 17

By construction is monotone and C = cl (dom ()). Next, construct : X X by (x) = (x) + NC (x) for all x. It folllows from Theorem 3.5 that is maximal monotone, contains and C is the closure of its domain. Other maximal monotone extensions are obtained by considering another S , or in working with instead of , they do not coincide in general. In the case where the interior of C is empty, a maximal monotone extension of is obtained in constructing a maximal monotone extension of the map : X1 X1 dened in section 4, next in taking such that (a + x1 ) = (x1 ) + X2 for all x1 X1 . Here again, is maximal monotone, contains and C is the closure of its domain.

References
[1] A. Auslender and M. Teboulle, Asymptotic cones and functions in optimization and variational inequalities, Springer-Verlag, New York Inc., 2003. [2] G. Birkho, Lattice Theory, Corrected reprint of the 1967 third edition, Amer. Math. Soc. Colloq. Publ. 25, Amer. Math. Soc., Providence, RI, 1979. erateurs Maximaux Monotones et Semigroupes de Con[3] H. Brezis, Op tractions dans les Espaces de Hilbert, North Holland, Amsterdam, 1973. [4] F. Facchinei and J.S. Pang, Finite dimensional variational inequalities and complementarity problems, Vol.1 and Vol.2, Springer-Verlag, New York Inc., 2003. [5] R.I. Kachurovskii, On monotone operators and convex functionals, Uspekhi Matematicheskikh Nauk 15, 1960, 213215. [6] R.I. Kachurovskii, Nonlinear monotone operators in Banach spaces, Uspekhi Matematicheskikh Nauk 23, 1968, 121168.

18

[7] C. Lescarret, Cas daddition des applications maximales dans un espace de Hilbert, C. R. Acad. Sci. Paris, 261, 1965, 11601163. [8] J.-E. Martinez-Legaz and F. Svaiter, Monotone Operators Representable by l.s.c. Convex Functions, Set-Valued Analysis, 13, 2005, 2146. [9] G.J. Minty, Monotone Networks, Proceeding of the Royal Society of London (Serie A) 257, 1960, 194212. [10] G.J. Minty, On the maximal domain of a monotone function, Michigan Mathematical Journal, 8, 1961, 135137. [11] G.J. Minty, Monotone (nonlinear) operators in Hilbert Space, Duke Mathematics Journal 29, 1962, 341346. [12] R. R. Phelps, Convex functions, Monotone Operators, and Dierentiability, Lecture notes in mathematics, 1364, Springer, New York, 1989. [13] R.T. Rockafellar Convex Analysis, Princeton University Press, Princeton, New Jersey, 1970. [14] R.T. Rockafellar On the maximality of sums of nonlinear monotone operators, Trans. Amer. Math. Soc., 149, 1970, 7588. [15] R.T. Rockafellar On the virtual convexity of the domain and range of a nonlinear maximal monotone operator, Mathematische Annalen, 185, 1970, 8190. [16] R.T. Rockafellar Monotone operators and the proximal point algorithm, SIAM J. Control Optim. 14, 1976, 877898. [17] R.T. Rockafellar and R.J.B Wets. Variational Analysis, Springer-Verlag, New York, 1998.

19