Fixed Point Theory and Applications

This Provisional PDF corresponds to the article as it appeared upon acceptance. Fully formatted PDF and full text (HTML) versions will be made available soon.

A New Extragradient Method for Generalized Variational Inequality in Euclidean Space

Fixed Point Theory and Applications 2013, 2013:139 doi:10.1186/1687-1812-2013-139

Haibin Chen (Chenhaibin508@163.com)

ISSN Article type Submission date Acceptance date Publication date Article URL

1687-1812 Research 2 March 2013 13 May 2013 29 May 2013

http://www.fixedpointtheoryandapplications.com/content/2013/1/139

This peer-reviewed article can be downloaded, printed and distributed freely for any purposes (see copyright notice below). For information about publishing your research in Fixed Point Theory and Applications go to http://www.fixedpointtheoryandapplications.com/authors/instructions/ For information about other SpringerOpen publications go to http://www.springeropen.com

2013 Chen This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

FPTA_269_edited

[05/28 11:51]

1/12

A new extragradient method for generalized variational inequality in Euclidean space

Haibin Chen

Abstract In this paper, we extend the extragradient projection method proposed in [JOTA, 119, 167-183, 2003] for the classical variational inequalities to the generalized variational inequalities. For this algorithm, we rst prove its expansion property of the generated sequence with respect to the starting point and then show that the existence of the solution to the problem can be veried through the behavior of the generated sequence. The global convergence of the method is also established under mild conditions. Keywords: generalized variational inequalities; extragradient method; solution existence; multivalued mapping AMS Subject Classication(2000): 90C30, 15A06.

Introduction
n

Let F (x) be a multi-valued mapping from Rn into 2R with nonempty values, and let X be a nonempty closed convex set in Rn . The generalized variational inequalities [1, 2], abbreviated as GVI, are to nd a vector x X such that there exists F (x ) satisfying , x x 0, x X, (1.1)

where , stands for the Euclidean inner product of vectors in Rn . The solution set of problem (1.1) is denoted by X . Certainly, the GVI reduces to the classical variational inequalities (VI) when F is a single-valued mapping, which has been well studied in the past decades [3, 4].
This Corresponding

work was supported by the Natural Science Foundation of China (Grant No. 11171180,11101303). author. School of Management Science, Qufu Normal University, Rizhao Shandong, China, 276800.

E-mail: chenhaibin508@163.com.

FPTA_269_edited

[05/28 11:51]

2/12

For the GVI, theories and solution methods have been extensively considered in the literature [2, 5, 6, 7, 8, 9, 10, 11, 12], and it is well known that the existence of solutions is an important topic for the GVI [1]. Generally, there are mainly two approaches to attack the solution existence problem of the GVI. One is an analytic approach which rst reformulates the GVI as a well-studied mathematical problem and then invokes an existence theorem for the latter problem [13]. The second is a constructive approach in which the existence can be veried by the behavior of the proposed method. The algorithm that is considered in this paper belongs to the second approach. First, we give a short summary to the constructive approach on the existence theory for the VI. For this approach, the equivalence between the existence of solutions to the VI problem and the boundedness of the sequence generated by some modied extragradient methods was rst established by Sun in [14]. Later, Wang et al. [4] established the same theory by a new type of extragradient-type method. Furthermore, the generated sequence possesses an expansion property with respect to the starting point and converges to a solution point if the solution set of the VI is nonempty. Now a question is posed naturally: as the GVI problem is an extension of the VI, can this theory be extended to the GVI? This constitutes the main motivation of the paper. In this paper, inspired by the work in [4], we propose a new type of extragradient projection method for the problem GVI. We rst establish the existence results for the GVI under pseudomonotonicity and continuity assumption of the underlying mapping F , and then show the global convergence of the proposed method. The rest of this paper is organized as follows. In Section 2, we give some related concepts and conclusions. In Section 3, we present the description of the algorithm and establish some properties of the algorithm. The global convergence of the sequence is also established.

Preliminaries

For a nonempty closed convex set K Rn and a vector x Rn , the orthogonal projection of x onto K , i.e., arg min{ y x | y K }, is denoted by PK (x). In what follows, we state some well-known properties of the projection operator which will be used in the sequel. Lemma 2.1 [15] Let K be a nonempty, closed and convex subset in Rn . Then, for any x, y Rn and z K , the following statements hold: (i) PK (x) x, z PK (x) 0; 2

FPTA_269_edited

[05/28 11:51]
2

3/12

(ii) PK (x) PK (y )

xy

PK (x) x + y PK (y ) 2 .

Remark 2.1 In fact, (i) in Lemma 2.1 provides also a sucient and necessary condition for a vector u K to be the projection of the vector x; i.e., u = PK (x) if and only if u x, z u 0, z K. Denition 2.1 Let K be a nonempty subset of Rn . A multi-valued mapping F : K 2R is said to be (i) monotone if and only if u v, x y 0, x, y K, u F (x), v F (y );
n

(ii) pseudomonotone if and only if, for any x, y K, u F (x), v F (y ), u, y x 0 = v, y x 0.

Now let us recall the denition of a continuous multi-valued mapping F . Denition 2.2 Assume that F : X 2R is a multi-valued mapping, then (i) F is said to be upper semicontinuous at x X if for every open set V containing F (x), there is an open set U containing x such that F (y ) V for all y X U ; (ii) F is said to be lower semicontinuous at x X if given any sequence {xk } converging to x and any y F (x), there exists a sequence y k F (xk ) that converges to y . F is said to be continuous at x X if it is both upper semicontinuous and lower semicontinuous at x. For the simplicity of our description, we list the assumptions needed in the sequel. Assumption 2.1 Suppose that X is a nonempty closed convex set in Rn . The multi-valued mapping F : X 2R is pseudomonotone and continuous on X with nonempty compact convex values.
n n

Main results

For x Rn , F (x), we rst dene the projection residue r(x, ) = x PX (x ). It is well known that the projection residue is related intimately to the solution set X . Proposition 3.1 For x X and F (x), they solve problem (1.1) if and only if x = P X (x ). 3

FPTA_269_edited

[05/28 11:51]

4/12

The basic idea of the designed algorithm is as follows. At each step of the algorithm, compute the projection residue r(xk , k ) at iterate xk . If it is a zero vector, then stop with xk being a solution of the GVI; otherwise, nd a trial point y k by a back-tracking search at xk along the residue r(xk , k ), and the new iterate is obtained by projecting x0 onto the intersection of X with two halfspaces respectively associated with y k and xk . Repeat this process until the projection residue is a zero vector. Algorithm 3.1 Choose , (0, 1), x0 X , k = 0. Step1: Given the current iterate xk X , if r(xk , ) = 0 for some F (xk ), stop; else take any k F (xk ) and compute z k = PX ( xk k ) . Then, let y k = (1 k )xk + k z k , where k = mk , with mk being the smallest nonnegative integer m satisfying: k F (xk m r(xk , k )) such that k , r ( xk , k ) r ( xk , k ) 2 .
0 1 H 2 X (x ), where Step2: Let xk+1 = PHk k 1 = {x Rn | x y k , k 0, k F (y k )}, Hk 2 = {x Rn | x xk , x0 xk 0}. Hk

(3.1)

Select k = k + 1 and go to step1. Now, we rst discuss the feasibility of the stepsize rule (3.1). Lemma 3.1 If xk is not a solution of problem (1.1), then there exists the smallest nonnegative integer m satisfying (3.1). Proof. By the denition of r(xk , k ) and Lemma 2.1, we know that PX (xk k ) (xk k ), xk PX (xk k ) 0, which implies k , r ( x k , k ) r ( xk , k ) Since (0, 1), we get
m 2

> 0.

(3.2)

lim (xk m r(xk , k )) = xk .

FPTA_269_edited

[05/28 11:51]

5/12

By this and the fact that F is lower semicontinuous, there exists m F (xk m r(xk , k )) such that
m

lim m = k .

So,
m

lim m , r(xk , k ) = k , r(xk , k ) r(xk , k )

> 0,

which implies the conclusion.

1 The following lemma shows that the halfspace Hk in Algorithm 3.1 strictly separates xk and the

solution set X if X is nonempty.

1 in Algorithm 3.1 separates the point xk from the set X . Lemma 3.2 If X = , the halfspace Hk

Moreover,
1 X, k 0. X Hk

Proof. By the denition of r(xk , k ) and Algorithm 3.1, we know y k = (1 k )xk + k z k = xk k r(xk , k ), which can be written k r(xk , k ) = xk y k . Then, by this and (3.1), we get k , x k y k > 0,
1 1 and (3.3), we get xk / Hk . where k is a vector in F (y k ). So, by the denition of Hk

(3.3)

On the other hand, for any x X and x X , we have , x x 0, F ( x ) . Since F is pseudomonotone on X , we get , x x 0, F (x). Let x = y k in (3.4), for any k F (y k ), we have k , y k x 0,
1 1 . Moreover, it is easy to see that X Hk X, k 0. which implies x Hk 1 2 Hk X and thus The following lemma says that if the solution set is nonempty, then X Hk 1 2 Hk X is a nonempty set. Hk

(3.4)

FPTA_269_edited

[05/28 11:51]

6/12

1 2 Lemma 3.3 If the solution set X = , then X Hk Hk X for all k 0 under Assumption 2.1. 2 for all k 0. The proof will be Proof. From the analysis above, it is sucient to prove that X Hk

given by induction. Obviously, if k = 0,

2 = Rn . X H0

Now, suppose that

2 X Hk

holds for k = l 0. Then X Hl1 Hl2 X. For any x X , by Lemma 2.1 and the fact that xl+1 = PHl1 Hl2 X (x0 ), it holds that x xl+1 , x0 xl+1 0.
2 Thus X Hl2 +1 . This shows that X Hk for all k 0, and the desired result follows. 1 2 Hk X is also nonempty from the For the case that the solution set is empty, we have that Hk

following lemma, which implies the feasibility of Algorithm 3.1.

1 2 Hk X = for all k 0 under Assumption 2.1. Lemma 3.4 Suppose that X = , then Hk

Before proving Lemma 3.4, we present a fundamental existence result for the GVI problem dened over a compact convex region [16]. For the sake of completeness, we give the proof process. Lemma 3.5 Let X Rn be a nonempty bounded closed convex set and let the multi-valued mapping F : X 2R be lower semicontinuous with nonempty closed convex values. Then the solution set X is
n

nonempty. Proof. Since the multi-valued mapping F is lower semicontinuous and has nonempty closed convex values, by Michaels selection theorem (see, for instance, Theorem 24.1 in [17]), it admits a continuous selection; that is, there exists a continuous mapping G : X Rn such that G(x) F (x) for every x X . Since X is a nonempty bounded closed convex set, the VI(X, G), which consists of nding an x X such that G(x), y x 0, y X, has a solution (see Lemma 3.1 in [18]), i.e., the solution set X of the problem VI(X, G) is nonempty. It follows from X X that X is nonempty. 6

FPTA_269_edited

[05/28 11:51]

7/12

1 2 Proof of Lemma 3.4. On the contrary, suppose that there exists k0 1 such that Hk Hk X = . 0 0

Then there exists a positive number M such that {xk | 0 k k0 } B (x0 , M ) and {xk k | 0 k k0 , k F (xk )} B (x0 , M ), where B (x0 , M ) = {x Rn | x x0 M }. Let Y = X B (x0 , 2M ) and consider the GVI(F, Y ). From Lemma 3.5, we know that the solution set
1 2 }, {Hk }, and {xk }, we Y of GVI(F, Y ) is nonempty. In order to avoid confusion with the sequence {Hk

2 } and {x 1 }, {H k }, respectively, when Algorithm 3.1 is denote the three corresponding sequences by {H k k applied to GVI(F, Y ) with a starting point x0 . We claim that 0 , x 1 , , x k0 ; (i) the set has at least k0 + 1 elements: x 1 = H 1, H 2 = H 2 for 0 k k0 ; (ii) x k = xk , H k k k k (iii) xk0 is not a solution of GVI(F, Y ). 1 H 2 X = , so H 1 H 2 X = , which Since Y = , using Lemma 3.3, we know that H k0 k0 k0 k0
1 2 Hk X = . contradicts the supposition that Hk 0 0

From Lemma 3.4, if the solution set of problem (1.1) is empty, then Algorithm 3.1 generates an innite sequence. More generally, we have the following conclusion. Theorem 3.1 Suppose that Assumption 2.1 holds. Assume that Algorithm 3.1 generates an innite sequence {xk }. If the solution set X is nonempty, then the sequence {xk } is bounded and all its cluster points belong to the solution set. Otherwise,
k+

lim

x k x0 = +

if the solution set X is empty. Proof. First, we suppose that the solution set is nonempty. Since
0 1 H 2 X ( x ) , xk+1 = PHk k

by Lemma 3.3 and the denition of the projection, it holds that xk+1 x0 x x0 for any x X . So, {xk } is a bounded sequence. 7

FPTA_269_edited
2 Since xk+1 Hk , it is obvious that

[05/28 11:51]

8/12

k+1 2 (x ) = xk+1 PH k

from the denition of the projection operator. For xk , since

2 , z xk , x0 xk 0, z Hk 0 2 (x ) from Remark 2.1. Thus, using Lemma 2.1, one has it holds that xk = PHk k+1 0 2 (x 2 (x ) ) PH k PHk 2

xk+1 x0

k+1 0 2 2 (x 2 (x ) PHk ) xk+1 + x0 PHk ;

i.e., xk+1 xk which can be written as xk+1 xk

2 2

xk+1 x0

x k x0 2 ,

+ x k x0

xk+1 x0 2 .

Thus, the sequence { xk x0 } is nondecreasing and bounded, and hence convergent, which implies that
k 1 , we get On the other hand, by xk+1 Hk

lim xk+1 xk

= 0.

(3.5)

xk+1 y k , k 0. Since y k = (1 k )xk + k z k = xk k r(xk , k ), and by (3.6) we have xk+1 y k , k = xk+1 xk + k r(xk , k ), k 0, which implies k r(xk , k ), k xk xk+1 , k 0. Using the Cauchy-Schwarz inequality and (3.1), we obtain k r(xk , k )
2

(3.6)

k r(xk , k ), k xk+1 xk

k .

(3.7)

Since F is continuous with compact values, Proposition 3.11 in [19] implies that {F (y k ) : k N } is a bounded set, and so the sequence { k : k F (y k )} is bounded. By (3.5) and (3.7), it follows that
k

lim k r(xk , k )

= 0.

FPTA_269_edited

[05/28 11:51]

9/12

For any convergent sequence {xkj } of {xk }, denotes its limit by x , i.e.,
j

. lim xkj = x

Without loss of generality, suppose that {kj } has a limit. Then

j

lim kj = 0

or
j

lim r(xkj , kj )

= 0.

For the rst case, by the choice of kj in Algorithm 3.1, we know that , r(xkj , kj ) < r(xkj , kj ) for all F (xkj Since
j k j 2

(3.8)

r(xkj , kj )). lim (xkj kj r(xkj , kj )) = x ,

k j

and F is lower semicontinuous on X , kj F (xkj

j

r(xkj , kj )) such that

lim kj = ,

where is a vector in F ( x). So, by (3.8) we obtain

j

x, ) lim r(xkj , kj ) lim kj , r(xkj , kj ) = , r(

j

= r( x, ) 2 .

(3.9)

Using the similar arguments of (3.2), we have , r( x, ) r( x, ) 2 . Combing this and (3.9), we know that r( x, ) = 0 and thus x is a solution of problem (1.1). For the second case
j

lim r(xkj , kj )

= 0,

it is easy to see that the limit point x of xkj is a solution of problem (1.1). Now, we consider the case that the solution set is empty. Since the inequality xk+1 xk
2

xk+1 x0

xk x0

also holds in this case, the sequence { xk x0 } is still nondecreasing. Next, we claim that
k+

lim

xk x0 = +. 9

FPTA_269_edited

[05/28 11:51]

10/12

Otherwise, { xk x0 } is a bounded sequence. A similar discussion as above would lead to the conclusion that every cluster point of {xk } is a solution of problem (1.1), which contradicts the emptiness of the solution set. Theorem 3.2 Under the assumption of Theorem 3.1, if the solution set X is nonempty, then the sequence {xk } globally converges to a solution x such that x = PX (x0 ); otherwise, limk+ xk x0 = +. That is, the solution set of problem (1.1) is empty if and only if the generated sequence diverges to innity. Proof. First, we suppose that the solution set is nonempty. From Theorem 3.1, we know that the sequence {xk } is bounded and that every cluster point x of {xk } is a solution of problem (1.1). Suppose that the subsequence {xkj } converges to x , i.e.,
j

lim xkj = x .

X , by Lemma 3.3 we have Let x = PX (x0 ). Since x

1 2 Hk X x Hk j 1 j 1

for all j . So, by the iterative sequence of Algorithm 3.1, we have x0 . x k j x0 x Thus, xk j x
2

= =

x k j x 0 + x0 x x kj x 0 x x0
2 2

2 2

+ x0 x
2

+ 2 x k j x0 , x 0 x

+ x0 x

+ 2 x kj x 0 , x0 x .

Letting j , we have x x
2

2 x x0

+ 2 x x 0 , x0 x

, x0 x 0, 2 x x

where the last inequality is due to Lemma 2.1 and the fact that x = PX (x0 ) and x X . So, = P X ( x0 ) . x = x Thus the sequence {xk } has a unique cluster point PX (x0 ), which shows the global convergence of {xk }. For the case that the solution set is empty, the conclusion can be obtained directly from Theorem 3.1.

10

FPTA_269_edited

[05/28 11:51]

11/12

Discussion

Certainly, the proposed extragradient method for the GVI in this paper has a good theoretical property in theory, as the generated sequence not only has an expansion property w.r.t. the starting point, but also the existence of the solution to the problem can be veried through the behavior of the generated sequence. However, the proposed algorithm is not easy to be realized in practice as the termination criterion is not easy to execute. This is an interesting topic for further research. Acknowledgment. The author would like to thank two anonymous reviewers for their insightful comments and constructive suggestions, and Professor Wang Yiju for his careful reading, which helped improve the presentation of the paper.

References
[1] F. Facchinei, J. S. Pang: Finite-Dimensional Variational Inequalities and Complementarity Problems. Springer-Verlag, New York, (2003) [2] F. Q. Xia, N. J. Huang: A projection-proximal point algorithm for solving generalized variational inequalities. J. Optim. Theory Appl. 150, 98-117(2011) [3] Y. Wang, N. Xiu and C. Wang: Unied Framework of Extragradient-type Methods for Pseudomonotone Variational Inequalities. J. Optim. Theory. Appl. 111, 641-656(2001) [4] Y. Wang, N. Xiu and J. Zhang: Modied Extragradient Method for Variational Inequalities and Verication of Solution Existence. J. Optim. Theory. Appl. 119, 167-183(2003) [5] A. Auslender, M. Teboulle: Lagrangian duality and related multiplier methods for variational inequality problems. SIAM J. Optim. 10, 1097-1115(2000) [6] Y. Censor, A. Gibali, S. Reich: The Subgradient Extragradient Method for Solving Variational Inequalities in Hilbert Space. J. Optim. Theory. Appl. 148, 318-335(2011) [7] S. C. Fang, E. L. Peterson: Generalized variational inequalities. J. Optim. Theory Appl. 38, 363383(1982) [8] C. Fand, Y. He: A double projection algorithm for multi-valued variational inequalities and a unied framework of the method. Appl. Math. Comput. 217, 9543-9511(2001) [9] Y. He: Stable pseudomonotone variational inequality in reexive Banach spaces. J. Math. Anal. Appl. 330, 352-363(2007) 11

FPTA_269_edited

[05/28 11:51]

12/12

[10] F. Li, Y. He: An algorithm for generalized variational inequality with pseudomonotone mapping. J. Comput. Appl. Math. 228, 212-218(2009) [11] R. Saigal: Extension of the generalized complementarity problem. Math. Oper. Res. 1, 260-266(1976) [12] G. Salmon, J. J. Strodiot and V. H. Nguyen: A bundle method for solving variational inequalities. SIAM J. Optim. 14, 869-893(2003) [13] Y. He: The Tikhonov Regularization Method for Set-Valued Variational Inequalities. Abst. Appl. Anal. 2012 [14] D. Sun: A class of iterative methods for solving nonlinear projection equations. J. Optim. Theory. Appl. 91, 123140(1996) [15] B. T. Polyak: Introduction to Optimization. Optimization Software Inc, Publications Division, New York (1987) [16] C. Fang, Y. He: An extragradient method for generalized variational inequality. Manuscript submitted (2012) [17] K. Deimling: Nonlinear Functional Analysis. Spriinger-Verlag, Berlin (1985) [18] P. Hartman, G. Stampacchia: On some nonlinear elliptic dierential functional equations. Acta Math. 115, 271-310(1966) [19] J. P. Aubin, I. Ekeland: Applied Nonlinear Analysis. John Wiley Sons Inc., New York (1984)

12