• No results found

An Algorithm Based on Resolvant Operators for Solving Positively Semidefinite Variational Inequalities

N/A
N/A
Protected

Academic year: 2020

Share "An Algorithm Based on Resolvant Operators for Solving Positively Semidefinite Variational Inequalities"

Copied!
15
0
0

Loading.... (view fulltext now)

Full text

(1)

Volume 2007, Article ID 76040,15pages doi:10.1155/2007/76040

Research Article

An Algorithm Based on Resolvant Operators for Solving

Positively Semidefinite Variational Inequalities

Juhe Sun, Shaowu Zhang, and Liwei Zhang

Received 16 June 2007; Accepted 19 September 2007 Recommended by Nan-Jing Huang

A new monotonicity,M-monotonicity, is introduced, and the resolvant operator of an M-monotone operator is proved to be single-valued and Lipschitz continuous. With the help of the resolvant operator, the positively semidefinite general variational inequality (VI) problem VI (Sn

+,F+G) is transformed into a fixed point problem of a nonexpan-sive mapping. And a proximal point algorithm is constructed to solve the fixed point problem, which is proved to have a global convergence under the condition thatFin the VI problem is strongly monotone and Lipschitz continuous. Furthermore, a convergent path Newton method is given for calculating-solutions to the sequence of fixed point problems, enabling the proximal point algorithm to be implementable.

Copyright © 2007 Juhe Sun et al. This is an open access article distributed under the Cre-ative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. Introduction

In recent years, the variational inequality has been addressed in a large variety of prob-lems arising in elasticity, structural analysis, economics, transportation equilibrium, op-timization, oceanography, and engineering sciences [1,2]. Inspired by its wide applica-tions, many researchers have studied the classical variational inequality and generalized it in various directions. Also, many computational methods for solving variational inequal-ities have been proposed (see [3–8] and the references therein). Among these methods, resolvant operator technique is an important one, which was studied in the 1990s by many researchers (such as [4,6,9]), and further studies developed recently [3,10,11].

(2)

proved to be Lipschitz-continuous. Applying the resolvant operator technique, we trans-form the positively semidefinite variational inequality (VI) problemVI(Sn+,F+G) into a fixed point problem of a nonexpansive mapping and suggest a proximal point algo-rithm to solve the fixed point problem. Under the condition thatFin theVIproblem is strongly monotone and Lipschitz-continuous, we prove that the algorithm has a global convergence. To ensure the proposed proximal point algorithm is implementable, we in-troduce a path Newton algorithm whose step size is calculated by Armijo rule.

In the next section, we recall some results and concepts that will be used in this paper. InSection 3, we introduce the definition of anM-monotone operator, and discuss prop-erties of this kind of operators, especially the Lipschitz continuity of the resolvant opera-tor of anM-monotone operator. InSection 4, we construct a proximal point algorithm, based on the results inSection 3, forVI(Sn+,F+G), and prove its global convergence. To ensure that the proposed proximal point algorithm inSection 4is implementable, we in-troduce a path Newton algorithm, inSection 5, in which the step size is calculated by Armijo rule.

2. Preliminaries

Throughout this paper, we assume thatSndenotes the space ofn×nsymmetric matrices andSn+denote the cone ofn×nsymmetric positive semidefinite matrices. ForA,B∈Sn, we define an inner productA,B =tr(AB) which induces the normA =A,A. Let 2Sn denote the family of all the nonempty subsets ofSn. We recall the following concepts, which will be used in the sequel.

Definition 2.1. LetA,B,C:Sn→Snbe single-valued operators and letM:Sn×Sn→Snbe mapping.

(i)M(A,·) is said to beα-strongly monotone with respect toAif there exists a con-stantα >0 satisfying

M(Ax,u)−M(Ay,u),x−y≥αx−y2, x,y,uSn; (2.1)

(ii)M(·,B) is said to beβ-relaxed monotone with respect toBif there exists a constant β >0 satisfying

M(u,Bx)−M(u,By),x−y≥ −βx−y2, x,y,uSn; (2.2)

(iii)M(·,·) is said to beαβ-symmetric monotone with respect toAandBifM(A,·) is α-strongly monotone with respect toA; andM(·,B) isβ-relaxed monotone with respect toBwithα≥βandα=βif and only ifx=y, for allx,y,u∈Sn;

(iv)M(·,·) is said to beξ-Lipschitz-continuous with respect to the first argument if there exists a constantξ >0 satisfying

(3)

(iv)Ais said to bet-Lipschitz-continuous if there exists a constantt >0 satisfying

Ax−Ay ≤tx−y, ∀x,y∈Sn; (2.4)

(vi)Bis said to bel-cocoercive if there exists a constantl >0 satisfying

Bx−By,x−y ≥lBx−By2, x,ySn; (2.5)

(vii)Cis said to ber-strongly monotone with respect toM(A,B) if there exists a con-stantr >0 satisfying

Cx−Cy,M(Ax,Bx)−M(Ay,By)≥rx−y2, x,ySn. (2.6)

In a similar way to (v), we can define the Lipschitz continuity of the mappingM with respect to the second argument.

Definition 2.2. LetA,B:Sn→Sn,M:Sn×Sn→Snbe mappings.Mis said to be coercive with respect toAandBif

lim

x→∞

M(Ax,Bx),x

x =+∞. (2.7)

Definition 2.3. LetA,B:Sn→Sn,M:Sn×SnSnbe mappings.Mis said to be bounded

with respect toAandBifM(A(P),B(P)) is bounded for every bounded subsetPofSn. M is said to be semicontinuous with respect toAandBif for any fixedx,y,z∈Sn, the

functiont→ M(A(x+ty),B(x+ty)),zis continuous at 0+. Definition 2.4. T:Sn→2Snis said to be monotone if

x−y,u−v ≥0, ∀u,v∈Sn,x∈Tu, y∈Tv; (2.8)

and it is said to be maximal monotone ifT is monotone and (I+cT)(Sn)=Sn for all

c >0, whereIdenotes the identity mapping onSn.

3.M-Monotone operators

In this section, we introduceM-monotonicity of operators and discuss its properties. Definition 3.1. LetA,B:Sn→Snbe single-valued operators,M:Sn×SnSna mapping,

andT:Sn→2Sna multivalue operator.Tis said to beM-monotone with respect toAand BifTis monotone and (M(A,B) +cT)(Sn)=Snholds for everyc >0.

Remark 3.2. IfM(A,B)=H, then the above definition reduces toH-monotonicity, which was studied in [5]. IfM(A,B)=I, then the definition ofI-monotonicity is just the maxi-mal monotonicity.

Remark 3.3. LetTbe a monotone operator and letcbe a positive constant. IfT:Sn→2Sn is anM-monotone operator with respect toAandB, every matrixz∈Sncan be written

(4)

Proposition3.4. LetM beαβ-symmetric monotone with respect toAandBand letT: Sn2Sn

be anM-monotone operator with respect toAandB, thenTis maximal monotone.

Proof. SinceT is monotone, it is sufficient to prove the following property; inequality

x−y,u−v ≥0 for (v,y)Graph(T) implies that

x∈Tu. (3.1)

Suppose, by contradiction, that there exists some (u0,x0)Graph(T) such that

x0−y,u0−v

0, (v,y)∈Graph(T). (3.2)

SinceTisM-monotone with respect toAandB, (M(A,B) +cT)(Sn)=Snholds for every

c >0, there exists (u1,x1)Graph(T) such that

MAu1,Bu1

+cx1=M

Au0,Bu0

+cx0∈Sn. (3.3)

It follows form (3.2) and (3.3) that 0≤cx0−x1,u0−u1

= −MAu0,Bu0

−MAu1,Bu1

,u0−u1

= −MAu0,Bu0

−MAu1,Bu0

,u0−u1

−MAu1,Bu0

−MAu1,Bu1

,u0−u1

≤ −(α−β)u0−u1

0,

(3.4)

which yieldsu1=u0. By (3.3), we have thatx1=x0. Hence (u0,x0)Graph(T), which is a contradiction. Therefore (3.1) holds andT is maximal monotone. This completes the

proof.

The following example shows that a maximal monotone operator may not beM -monotone for someAandB.

Example 3.5. LetSn=S2,T=I, andM(Ax,Bx)=x2+ 2Exfor allxS2, whereEis an identity matrix. Then it is easy to see thatIis maximal monotone. For allx∈S2, we have that

M(A,B) +I

(x)2=x2+ 2Ex+x2=x2+ 2E2=trx2+ 2E2 8, (3.5)

which means that 0 ¯(M(A,B) +I)(S2) andIis notM-monotone with respect toAand B.

Proposition3.6. LetT:Sn→2Snbe a maximal monotone operator and letM:Sn×Sn

Snbe a bounded, coercive, semicontinuous, andαβ-symmetric monotone operator with

re-spect toAandB. ThenTisM-monotone with respect toAandB.

(5)

toAandB, it follows from [9, Corollary 32.26] thatM(A,B) +cT is surjective, that is, (M(A,B) +cT)(Sn)=Snholds for everyc >0. Thus,Tis anM-monotone operator with

respect toAandB. The proof is complete. Theorem3.7. LetMbe anαβ-symmetric monotone with respect toAandBand letT be anM-monotone operator with respect toAandB. Then the operator(M(A,B) +cT)1is single-valued.

Proof. For any givenu∈Sn, letx,y(M(A,B) +cT)1(u). It follows thatM(Ax,Bx) + u∈Txand−M(Ay,By) +u∈T y. The monotonicity ofTandMimplies that

0≤−M(Ax,Bx) +u−−M(Ay,By) +u,x−y

= −M(Ay,By)−M(Ax,Bx),x−y

≤ −−β)u0−u1

0.

(3.6)

From the symmetric monotonicity ofM, we get thatx=y. Thus (M(A,B) +cT)−1 is single-valued. This completes the proof. Definition 3.8. LetMbe anαβ-symmetric monotone with respect toAandBand letTbe anM-monotone operator with respect toAandB. The resolvant operatorJcTM:Sn→Snis

defined by

JM

cT(u)=M(A,B) +cT−1(u), ∀u∈Sn. (3.7)

Theorem3.9. LetM(A,B)beα-strongly monotone with respect toAandβ-relaxed mono-tone with respect toBwithα > β. Suppose thatT:Sn→2Sn is anM-monotone operator. Then the resolvant operatorJcTM :Sn→Snis Lipschitz-continuous with constant1/(α−β), that is,

JM

cT(u)−JcTM(v)α1βu−v, ∀u,v∈Sn. (3.8)

Since the proof ofTheorem 3.9is similar as that of [5, Theorem 2.2], we here omit it.

4. An algorithm for variational inequalities

Let F,G:Sn+→Sn be operators. Consider the general variational inequality problem VI(Sn+,F+G), defined by findingu∈Sn+such that

F(u) +G(u),v−u≥0, ∀v∈Sn+. (4.1)

We can rewrite it as the problem of findingu∈Sn

+such that

0∈G(u) +T(u), (4.2)

whereT≡F+ᏺ(·;Sn

(6)

Proposition4.1. LetF,G:Sn+→Snbe continuous and letM:Sn×Sn→Snbe a bounded, coercive, semicontinuous, andαβ-symmetric monotone operator with respect toA:Sn→Sn

andB:Sn→Sn. Then the following two properties hold for the mapT≡F+ᏺ(·;Sn+): (a)JcTM(M(Ax,Bx)−cG(x))=Sol(Sn+,Fcx), whereFcx(y)=M(Ay,By)−M(Ax,Bx) +

c(F(y) +G(x));

(b)IfFis monotone, thenTisM-monotone with respect toAandB. Proof. We have that the inclusion

y∈JM

cTM(Ax,Bx)−cG(x)=M(A,B) +cT−1M(Ax,Bx)−cG(x) (4.3)

is equivalent to

M(Ax,Bx)∈M(A,B) +cF+c·;Sn+

(y) +cG(x), (4.4)

or in other words,

0∈M(Ay,By)−M(Ax,Bx) +cF(y) +G(x)+ᏺy;Sn+

. (4.5)

This establishes (a).

By [10, Proposition 12.3.6], we can deduce thatT is maximal monotone, it follows fromProposition 3.6, we get thatT isM-monotone with respect toAandB. This

com-pletes the proof.

Lemma4.2. LetMbe anαβ-symmetric monotone with respect toAandBand letTbe an M-monotone operator with respect toAandB. Thenu∈Sn

+is a solution of0∈G(u) +T(u) if and only if

u=JM

cTM(Au,Bu)−cG(u), (4.6)

whereJcTM=(M(A,B) +cT)−1andc >0is a constant.

In order to obtain our results, we need the following assumption.

Assumption4.3. The mappingsF,G,M,A,Bsatisfy the following conditions. (1)FisL-Lipschitz-continuous andm-strongly monotone.

(2)M(A,·)isα-strongly monotone with respect toA; andM(·,B)isβ-relaxed monotone with respect toBwithα > β.

(3)M(·,·)isξ-Lipschitz-continuous with respect to the first argument andζ -Lipschitz-continuous with respect to the second argument.

(4)Aisτ-Lipschitz-continuous andBist-Lipschitz-continuous.

(5)Gisγ-Lipschitz-continuous ands-strongly monotone with respect toM(A,B). Remark 4.4. LetAssumption 4.3hold and

c−γs2

s2γ2(ξτ+ζt)2(αβ)2

γ2 , s

(7)

We can deduce that

JM

cTM(Ax,Bx)−cG(x)−JcTMM(Ay,By)−cG(y) α1

−βM(Ax,Bx)−M(Ay,By)−c

G(x)−G(y)

(ξτ+ζt)22cs+c2γ2

α−β x−y

≤ x−y,

(4.8)

which implies thatJcTM(M(A,B)−cG) is nonexpansive. Then, it is natural to consider the recursion

xk+1JM

cTMAxk,Bxk−cGxk, (4.9)

which is desired to converge to a zero ofG+T. Actually, this can be proved to be true. However, based onLemma 4.2, we construct the following proximal point algorithm for VI(Sn+,F+G).

Algorithm4.5 Data. x0Sn,c

0>0,ε00, andρ0>0. Step 1. Setk=0.

Step 2. Ifxk∈Sol(Sn+,F+G), stop. Step 3. Findwksuch thatwkJM

ckT(M(Axk,Bxk)−ckG(xk)) ≤εk. Step 4. Setxk+1(1ρ

k)xk+ρkwkand selectck+1,εk+1andρk+1. Setk←k+ 1and go to Step 1.

The following theorem fully describes the convergence ofAlgorithm 4.5for finding a solution toVI(Sn+,F+G).

Theorem 4.6. Suppose thatAlgorithm 4.5 holds. LetM be bounded, coercive, semicon-tinuous, and αβ-symmetric monotone with respect to A and B; and let F be monotone and Lipschitz-continuous. Letx0Snbe given, let{εk} ⊂[0,)satisfyE

k=1εk<∞,

{ck} ⊂(cm,), wherecm>0and

ck−γs2

<

s2γ2(ξτ+ζt)2β)2

γ2 , s

2> γ2(ξτ+ζt)2β)2 , (4.10)

which implies that

L=

α−β−(ξτ+ζt)22c

ks+ck2γ2

α−β+ 3(ξτ+ζt)22c

ks+ck2γ2

2 >0. (4.11)

If{ρk} ⊆[Rm,RM], where0< Rm≤RM≤pL, for all p∈[2, +), then the sequence{xk}

(8)

Proof. We introduce a new map

QkIJM ckT

M(A,B)−ckG. (4.12)

Clearly, any zero ofG+F+ᏺ(·;Sn+), being a fixed point ofJcMkT(M(A,B)−ckG), is also a zero ofQk. Now, let us prove thatQkisL-cocoercive.

Forx,y∈Snwe know that

Qk(x)Qk(y),xy =x−y−JM

ckT

M(Ax,Bx)−ckG(x)−JcMkT

M(Ay,By)−ckG(y),x−y = x−y2JM

ckT

M(Ax,Bx)−JM ckT

M(Ay,By)−ckG(x)−G(y),x−y

≥ x−y2 1

α−βM(Ax,Bx)−M(Ay,By)−ck

G(x)−G(y)x−y

≥ x−y2 1 α−β

(ξτ+ζt)22c

ks+ck2γ2x−y2

=

1

(ξτ+ζt)22cks+c2

2

α−β

x−y2,

(4.13)

Qk(x)Qk(y)2

=x−y−JM ckT

M(Ax,Bx)−ckG(x)−JcMkT

M(Ay,By)−ckG(y)2 = x−y22xy,JM

ckT

M(Ax,Bx)−ckG(x)−JcMkT

M(Ay,By)−ckG(y)

+JcMkTM(Ax,Bx)−ckG(x)−JcMkT

M(Ay,By)−ckG(y)2

≤ x−y2+ 2

(ξτ+ζt)22cks+c2

2

α−β x−y2

+

(ξτ+ζt)22cks+c2

2

α−β x−y2

=

⎛ ⎝1 + 3

(ξτ+ζt)22c

ks+ck2γ2

α−β

xy2.

(4.14)

Inequalities (4.13) and (4.14) imply that

Qk(x)Qk(y),xy

⎡ ⎣1

(ξτ+ζt)22c

ks+c22

α−β

⎤ ⎦ ⎡ ⎣1+3

(ξτ+ζt)22c

ks+c22

α−β

⎤ ⎦

1

Qk(x)Qk(y)2

=LQk(x)Qk(y)2 .

(9)

For allk, we denote byxkthe point computed exactly by the resolvent. That is,

xk+11ρ

kxk+ρkJcMkT

MAxk,Bxkc

kGxk. (4.16)

For every zerox∗ofT, we obtain

xk+1x2

=xkρ

kQkxk−x∗2 =xkx2

2ρkQkxk−Qkx∗,xk−x∗+ρk2Qkxk2 ≤xkx2

kLQkxk2+ρ2kQkxk2 ≤xkx2

−ρk2L−ρkQkxk2 ≤xkx2

−Rm2L−RMQkxk2 ≤xkx2

.

(4.17)

Sincexk−xk ≤ρkεk, we get that

xk+1xxk+1x+xk+1xk+1

≤xkx+ρ

kεk

≤x0x+k

i=0 ρiεi

≤x0x+pLE.

(4.18)

Therefore, the sequence{xk}is bounded. On the other hand, we have that

xk+1x2

=xk+1x+xk+1xk+12

=xk+1x2

+ 2xk+1−x∗,xk+1xk+1+xk+1xk+12

≤xk+1x2

+ 2xk+1−x∗xk+1−xk+1+xk+1−xk+12

≤xkx2

+ 2ρkεkx0−x∗+pLE+ρ22k −Rm2L−RMQkxk2.

(4.19)

LettingE2=

i=0εk2<∞, we have for everyk,

xk+1x2

≤x0x2

+ 2pLEx0x+pLE

+p2L2E2−Rm2L−RM k

i=0

Qkxk2

. (4.20)

Passing to the limitk→ ∞, one has thati=0Qk(xk)2<∞, implying that

lim

k→∞Q

(10)

According to Remark 3.3, for every k, there exists a unique pair (yk,vk) in gphT such thatzk=M(Axk,Bxk)c

kG(xk)=M(Ayk,Byk) +ckvk. ThenJcMkT(M(Axk,Bxk) ckG(xk))=yk. So thatQk(xk)0 implies that (xk−yk)0,vk→0.

Sinceckis bounded away from zero, it follows thatck−1Qk(xk)0. Sincexkis bounded,

it has at least a limit point. Letx∞ be such a limit point and assume that the subse-quence{xki:kik}converges tox∞. It follows that{yki:kik}also converges tox∞. For every (y,v) in gphTby the monotonicity ofT, we have thaty−yk,vvk0.

Let-tingki(∈k)→ ∞, we get thaty−y∞,v−vk ≥0. We see thatTisM-monotone due to Proposition 4.1, this implies that (x,−G(x∞))gphT, that is,G(x)T(x). This

completes the proof.

5. Solving an approximate fixed point toJcMkT

How to calculatewkat Step 3 is the key inAlgorithm 4.5. Ifεk=0, this amounts to the

exact solution ofVI(Sn+,Fk), where

Fk(x)=M(Ax,Bx)−MAxk,Bxk+ckF(x) +G(xk. (5.1)

Now, we consider the case ofεk>0. We can prove thatJcMkT(M(Axk,Bxk)−ckG(xk)) is the unique solution of theVI(Sn+,Fk). Hence,wk is an inexact solution of theVI(Sn+,Fk) satisfying dist(wk, Sol(Sn

+,Fk))≤εk.

Lemma5.1. LetF,G,M,A,Bsatisfy all the conditions ofAssumption 4.3. Then a constant c(k)>0exists such that

distwk, SolSn+,Fk

≤c(k)FknatSn +

wk. (5.2)

Proof. ByAssumption 4.3, we can easily get that Fk isL(k)-Lipschitz-continuous and η(k)-strongly monotone, whereL(k)=ξτ+ζt+ckLandη(k)=α−β+ckm, that is,

Fk(x)Fk(y)L(k)xy,

Fk(x)−Fk(y),x−y≥η(k)x−y2, ∀x,y∈Sn+.

(5.3)

Letr=(Fk)natSn+(wk), where (Fk)natSn+ is the natural map associated with theVI(Sn+,Fk). We have thatwk−r=ΠSn+(wk−Fk(wk)), that is,

y−wk+r,F

kwk−r≥0, ∀y∈Sn+. (5.4) For allx∗∈Sol(Sn

+,Fk) andwk−r∈Sn+, we also have that

wkrx,F

kx∗≥0. (5.5)

From (5.4) and (5.5), we get that

x∗wk,F

kwk−r+r,Fkwk−r

=x∗wk,Fkwkxwk,r+r,Fkwk+r2

0, (5.6)

(11)

Adding (5.6) and (5.7), we deduce that

η(k)wkx2

≤x∗wk,F

kwk−Fkx∗ ≤ −wkx,rr2+r,F

kx∗−Fkwk ≤ rwkx+rL(k)wkx

=1 +L(k)wk−x∗r.

(5.8)

Hence,wkxη(k)1(1 +L(k))r. This implies that

distwk, SolSn+,Fk

≤c(k)FknatSn+wk, (5.9)

wherec(k)=η(k)−1(1 +L(k)). This completes the proof.

Consequently, the computation ofwk can be accomplished by obtaining an inexact solution ofVI(Sn+,Fk) satisfying the residual condition

Fknat

Sn +

wk 1

c(k)εk. (5.10)

We note that the operatorSn

+(·) is directionally differentiable and strongly semismooth everywhere (see, e.g., [12]). IfFk(·) is continuously differentiable, then we get that

FknatSn+(w)=w−

Sn+

w−Fk(w) (5.11)

is directionally differentiable.

In what follows, we present the following path Newton method for solving the equa-tion (Fk)natSn

+(w

k)=0.

Algorithm5.2

Data. w0Sn,γ(0, 1), andρ(0, 1).

Step 1. Setj=0. Step 2. If(Fk)natSn

+(w

j)=0, stop.

Step 3. Select an elementVj∈∂[(Fk)natS+n(wj)]and consider the corresponding pathpj(·)= wj(·)V1

j (Fk)natSn+(wj)with domainIj=[0, ¯τj)for someτ¯j(0, 1]. Find the smallest non-negative integerijsuch that withi=ij,ρiτ¯j∈Ijand

Fknat

Sn +

pjρiτ¯

j≤1−γρiτ¯jFknatSn +

wj. (5.12)

Step 4. Setτj=ρijτ¯j,wj+1=pjj), and j←j+ 1; go to Step 2.

Theorem5.3. Let F,G,M,A, andBsatisfy all the conditions of Assumption 4.3. If for allw∈Sn

+every matrix in∂[(Fk)natS+n(w)]is nonsingular, then the sequence{wj}generated byAlgorithm 5.2has at least one accumulation point and every accumulation point of the sequence{wj}is the zero point of(F

(12)

Proof. The nonnegative sequence{(Fk)natSn+(wj)}is monotonically decreasing; thus it is bounded. It follows fromLemma 5.1that the sequence{wj}is bounded. That implies

that{wj}has at least one accumulation point.

Assume that there exists a subsequence{wjm}of{wj}converging tow∗such that

FknatSn +

w∗=0, (5.13)

that is, there exist positive constantsδandηsatisfying

Fknat

Sn+wjm≥η, ∀wjm∈Bw∗. (5.14) By the strong semismoothness of (Fk)natSn+, we have that

FknatSn+(w+h)−FknatSn+(w)−V(h)=Oh2, ∀w∈Bw∗,δ,h∈Sn+, (5.15)

whereV∈∂(Fk)natSn

+(w) is nonsingular and there is a positive constantcsuch that sup

w∈B(w∗,δ) V∈∂(Fk)nat

Sn+(w)

maxV,V−1≤c. (5.16)

Lettingτ∈(0, ¯τjm),h=pjm(τ)−wjm, andVjm∈∂[(Fk)natSn+(wjm)], we have that Vjm

pjm(τ)wjm=F

knatSn +

pjm(τ)F

knatSn +

wjmOpjm(τ)wjm2. (5.17)

It follows that

Fknat

Sn+pjm(τ)−Vjmpjm(τ)−wjm−FknatSn+wjm

pjm(τ)−wjm =

Opjm(τ)wjm2

pjm(τ)−wjm . (5.18) So we can choose a positivet∗small enough so that

o(t)

t

1−γ

c , ∀t∈

0,t∗ . (5.19)

From the definition ofpjm, we know that there exists a constantτ∗(0, ¯τj

m] small enough so thatpjm(τ)wjmt, for allτ(0,τ], which implies that

opjm(τ)wjm

pjm(τ)−wjm 1−γ

c , ∀τ∈

0,τ∗ . (5.20)

It follows from (5.16), (5.18), (5.19), and (5.20) that

Fknat

Sn +

pjm(τ)Vj m

pjm(τ)wjm+Fknat

Sn +

wjm

+pjm(τ)wjmopjm(τ)−wjm pjm(τ)−wjm

(1−τ)FknatSn +

wjm+τV−1

jmFk

nat

Sn +

wjm1−γ

c

(1−τγ)FknatSn +

wjm, τ(0,τ∗].

(13)

Consequently,

Fknat

Sn +

pjm(τ)(1τγ)Fknat

Sn +

wjm, τ(0,τ∗]. (5.22)

By the definition of the step-sizeτjm, it follows that there existsξ∈(0,τ∗) such thatτjm≥ ξfor alljm. Indeed, if no suchξexists, then{τjm}converges to zero. This implies that the sequence of integers{ijm}is unbounded. Consequently, by the definition ofijm, we have, for alljmsufficiently large,

Fknat

Sn+pkρijm−1τ¯k>1−γρijm−1τ¯kFknatSn+wk; (5.23) but this contradicts (5.22) withτ≡ρijm−1τ¯

jm. Consequently, the desiredξexists. The in-equality (5.22) implies that

Fknat

Sn +

wjm+11τj mγFk

nat

Sn +

wjm. (5.24)

Passing to the limitm→ ∞, we deduce a contradiction because limm→∞(Fk)natSn+(wjm) η >0 and the sequence{τjm}is bounded away from zero. This yields that (Fk)natSn+(w∗)=0.

This completes the proof.

Remark 5.4. As stated above,Algorithm 5.2generates a sequence converging to the zero point of (Fk)natSn

+, Step 3 inAlgorithm 4.5is implementable. Obviously,Algorithm 5.2stops within a finite number of iterations at awksuch that (5.10) holds.

Example 5.5. Assume that there exists a positive constant ¯csuch that

sup

w∈Sn +

sup

V∈∂Sn+(xk−ck(F(w)+G(xk)))

VJF(w)c.¯ (5.25)

Letck∈(0, 1/c¯). Suppose thatM(Aw,Bw)=w, for allw∈Sn+andFis Lipschitz-continu-ous and strongly monotone. We have (Fk)natSn

+(w)=w−

Sn

+(xk−ck(F(w) +G(xk))), for all w∈Sn

+. Then ∂[(Fk)natSn

+(w)]⊂ {I−ckVJF(w)|V∈∂

Sn +(x

kck(F(w) +G(xk)))},

for all w∈Sn

+. We easily get that every matrix in ∂[(Fk)natSn

+(w)] is nonsingular for all w∈Sn

+. It follows fromTheorem 5.3that every accumulation point of{wk}generated byAlgorithm 5.2is the zero point of (Fk)natSn+.

At first sight, the M-monotonicity of T=F+ᏺ(·,Sn

+) seems having little use be-cause the algorithm based on maximal monotonicity can also solve the VI(Sn+,F+G) directly. However, we will see that in some practical cases the variational inequality us-ingAlgorithm 4.5, which is based onM-monotone operator, is actually much simpler to solve and easier to analyze than using algorithm based on maximal monotone map. We illustrate this by the following example.

Example 5.6. LetF:Sn+→Snbe defined by

F(x)=S(x) + 1

16x, G(x)= 1

8x ∀x∈S

n

(14)

We haveF is (s+ (1/16))-Lipschitz-continuous, (1/16)-strongly monotone, andG is (1/8)-Lipschitz-continuous.

Now, we takeM(Ax,Bx)=Ax+Bx, whereAx=(1 + (ck/16))x andBx= −ckS(x)−

(ck/8)xfor allx∈Snand 0< ck<16. Then, we can easily prove thatM(·,·) is

Lipschitz-continuous with first and second arguments,M(A,B) is bounded and semicontinuous; andAandBare both Lipschitz-continuous. It is also easy to see that

lim

x→∞

M(Ax,Bx),x

x =xlim→∞

−ckS(x) +1−ck/16x,x

x =+, (5.27)

which implies thatM(A,B) is coercive. Also, we can deduce thatM(A,B) is (1 + (ck

/16))-strongly monotone with respect toAandck(s+ (1/8))-relaxed monotone with respect to

Band (1 + (ck/16))> ck(s+ (1/8)), if we lets <(1/ck)(1/16). Also, we can prove thatG

is strongly monotone with respect toM(A,B). We choosex0Sn

+,{εk},{ck}, and{ρk}satisfyingTheorem 4.6and compute{wk}by the residual rule

Fknat

S+

wk=wk Sn

+

wkMAwk,Bwk+MAx

k,Bxk−ckFwk+Gxk

=wk Sn+ −ckS

xk+ 1316ck

!

xk!

η(k)

1 +L(k)εk,

(5.28)

that is,wkcan be computed as follows:

wk

Sn+ −ckS(xk

+ 13ck 16

!

xk!1 +η(k)L(k)εk. (5.29)

It follows fromTheorem 4.6that the sequence{xk}generated byAlgorithm 4.5converges to a solution of

"

S(x) + 1 16x+

1 8x,y−x

#

0, ∀y∈Sn

+. (5.30)

Note that the core of proximal point algorithm is the calculation ofwk. As we have seen, if we use [10, Algorithm 12.3.8], which is based on the maximal monotonicity ofT=

F+ᏺ(·,Sn+),wkwill be computed as

wk

Sn +

xk−ckSwk−c8kxk!1 +η(Lk)(k)εk, (5.31)

(15)

Acknowledgments

This work was supported by the National Natural Science Foundation of China under project Grant no. 10771026 and by the Scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry.

References

[1] W. B. Liu and J. E. Rubio, “Optimal shape design for systems governed by variational inequalities—I: existence theory for the elliptic case,”Journal of Optimization Theory and Ap-plications, vol. 69, no. 2, pp. 351–371, 1991.

[2] W. B. Liu and J. E. Rubio, “Optimal shape design for systems governed by variational inequalities—II: existence theory for the evolution case,”Journal of Optimization Theory and Applications, vol. 69, no. 2, pp. 373–396, 1991.

[3] R. Ahmad, Q. H. Ansari, and S. S. Irfan, “Generalized variational inclusions and generalized re-solvent equations in Banach spaces,”Computers & Mathematics with Applications, vol. 49, no. 11-12, pp. 1825–1835, 2005.

[4] X. P. Ding, “Perturbed proximal point algorithms for generalized quasivariational inclusions,” Journal of Mathematical Analysis and Applications, vol. 210, no. 1, pp. 88–101, 1997.

[5] Y.-P. Fang and N.-J. Huang, “H-monotone operator and resolvent operator technique for varia-tional inclusions,”Applied Mathematics and Computation, vol. 145, no. 2-3, pp. 795–803, 2003. [6] A. Hassouni and A. Moudafi, “A perturbed algorithm for variational inclusions,” Journal of

Mathematical Analysis and Applications, vol. 185, no. 3, pp. 706–712, 1994.

[7] C.-H. Lee, Q. H. Ansari, and J.-C. Yao, “A perturbed algorithm for strongly nonlinear variational-like inclusions,”Bulletin of the Australian Mathematical Society, vol. 62, no. 3, pp. 417–426, 2000.

[8] M. A. Noor, “Implicit resolvent dynamical systems for quasi variational inclusions,”Journal of Mathematical Analysis and Applications, vol. 269, no. 1, pp. 216–226, 2002.

[9] G. X.-Z. Yuan,KKM Theory and Applications in Nonlinear Analysis, vol. 218 ofMonographs and Textbooks in Pure and Applied Mathematics, Marcel Dekker, New York, NY, USA, 1999. [10] F. Facchinei and J.-S. Pang,Finite-Dimensional Variational Inequalities and Complementarity

Problems, vol. 2, Springer, New York, NY, USA, 2003.

[11] Z. Liu, J. S. Ume, and S. M. Kang, “Resolvent equations technique for general variational inclu-sions,”Proceedings of the Japan Academy. Series A, vol. 78, no. 10, pp. 188–193, 2002.

[12] D. Sun and J. Sun, “Semismooth matrix-valued functions,”Mathematics of Operations Research, vol. 27, no. 1, pp. 150–169, 2002.

Juhe Sun: Department of Applied Mathematics, Dalian University of Technology, Dalian, Liaoning 116024, China

Email address:[email protected]

Shaowu Zhang: Department of Applied Mathematics, Dalian University of Technology, Dalian, Liaoning 116024, China

Email address:[email protected]

Liwei Zhang: Department of Applied Mathematics, Dalian University of Technology, Dalian, Liaoning 116024, China

References

Related documents

Some authors indicated that the concentration of leptin in seminal plasma of men with normal semen sample is significantly lower than pathological groups (4)

Table A-5: Mortality hazard ratios (ages 18–64) for first- and second-generation immigrant subgroups by region of origin, France, 1999–2010; native- born individuals with ‘one

odern societies stand in our times in the face of two parallel processes: on one hand, the decline in the Gemeinschaft's geographical significance; on the other

Collectively, studies published thus far demonstrate that alterations in transcript abundance can be detected on a genome-wide scale in the blood of patients with a wide range

There is a growing body of evidence suggesting that barriers to functional recovery are associated with a host of neurocogni- tive impairments in both the early and later course