JJ II

(1)

volume 4, issue 2, article 28, 2003.

Received 18 November, 2002;

accepted 4 February, 2003.

Communicated by:A.M. Rubinov

Abstract Contents

JJ II

J I

Home Page Go Back

Close Quit

Journal of Inequalities in Pure and Applied Mathematics

GENERALIZED QUASI-VARIATIONAL INEQUALITIES AND DUALITY

JACQUELINE MORGAN AND MARIA ROMANIELLO

Dipartimento di Matematica e Statistica Università di Napoli Federico II Via Cinthia, 80126 Napoli, Italy.

EMail:morgan@unina.it

Dipartimento di Organizzazione Aziendale e Amministrazione Pubblica Università della Calabria

Via Pietro Bucci, 87036 Cosenza, Italy.

EMail:mromanie@unina.it

c

2000Victoria University ISSN (electronic): 1443-5756 126-02

(2)

Generalized Quasi-Variational Inequalities and Duality

Jacqueline Morgan and Maria Romaniello

Title Page Contents

JJ II

J I

Go Back Close

Quit Page2of18

J. Ineq. Pure and Appl. Math. 4(2) Art. 28, 2003

http://jipam.vu.edu.au

Abstract

We present a scheme which associates to a generalized quasi-variational inequality a dual problem and generalized Kuhn-Tucker conditions. This scheme allows to solve the primal and the dual problems in the spirit of the classical Lagrangian duality for constrained optimization problems and extends, in non necessarily finite dimentional spaces, the duality approach obtained by A.

Auslender for generalized variational inequalities. An application to social Nash equilibria is presented together with some illustrative examples.

2000 Mathematics Subject Classification:65K10, 49N15, 91A10.

Key words: Generalized quasi-variational inequality, Primal and dual problems, Gen- eralized Kuhn-Tucker conditions, Banach space, Social Nash equilibrium, Subdifferential.

1. Introduction

Let X be a real Banach space with dualX^∗ or, more generally, let X andX^∗ be two real locally convex topological vector spaces, duals with respect to a product of dualityh·,·i(see [14, p. 336]).

IfA and K are two set-valued operators from X toX^∗ and from X to X, respectively, we are interested to the following variational problem (in short (V P)):

findx^∗ ∈X such that x^∗ ∈K(x^∗) and there exists (V P)

z^∗ ∈A(x^∗)satisfying hz^∗, x−x^∗i ≥0, for allx∈K(x^∗).

This problem, called Generalized Quasi-Variational Inequality ([16], [8], [12], ...), generalizes the following problems:

– variational inequalities as introduced by G. Stampacchia [17] (see also [2], [6], [11], ...)

– generalized variational inequalities ([2], [5], [11], ...) – quasi-variational inequalities ([6], [12], ...)

and describes various economic and engineering problems (see Section 3and, for example, [1], [7], [10]).

Existence results for solutions of such a problem have been given in [8] and [16], while stability of the following problem (equivalent to (V P) under suitable assumptions):

(4)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page4of18

(V P)⁰ : findx^∗ ∈X such thatx^∗ ∈K(x^∗) and inf

z^∗∈A(x^∗)

hz^∗, x−x^∗i ≥0, for allx∈K(x^∗) has been investigated in [12].

Differently, to our knowledge there exists no results concerning a duality scheme or a numerical method which solves a generalized quasi-variational inequality. Nevertheless, in the case of generalized variational inequalities, for constraints defined by a finite number of inequalities and in finite dimensional spaces, A. Auslender introduced in [2] a duality scheme which associates to the Primal Problem another generalized variational inequality (with only constraints of positivity) for which an algorithm has been developed (see [3]).

In this paper, we extend to generalized quasi-variational inequalities in non necessarily finite dimensional spaces the duality approach obtained by Auslen- der for generalized variational inequalities. More precisely we present a scheme which associates to the variational problem (V P):

– a dual problem, called (DV P) – Generalized Kuhn-Tucker Conditions

which allows us to solve (V P) and (DV P) in the spirit of the classical La- grangian duality for constrained optimization problems. From a numerical point of view, we point out that the dual problem (DV P) has a special structure which allows to apply the algorithm introduced in [3] for generalized variational inequalities.

In Section2, we present the duality scheme and the connections between the primal and the dual problems through the Generalized Kuhn-Tucker Conditions.

(5)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page5of18

In Section3, we apply this method to find Social Nash Equilibria for nonzero- sum games with coupled constraints defined by a finite number of inequalities and we give some illustrative examples.

(6)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page6of18

2. Duality Scheme for (V P )

The scheme presented in this section takes advantage of the particular structure of the set-valued operator K defined by a finite number of inequalities. More precisely, we assume that for allx∈X:

K(x) ={z ∈X/f_j(x, z)≤0, for allj = 1,2, . . . , m}

where:

f_j(x,·) :X →R∪ {+∞} is a proper, closed and (H1)

convex function ([18]) for allj = 1, . . . , m.

Now, for allu∈R^m+, let

(2.1) F (x, y) = (f₁(x, y), . . . , f_m(x, y)) and

(2.2) G(u) = (

−F(x, x) /0∈A(x) +

m

X

j=1

u_j∂₂f_j(x, x) )

where∂₂f_j(x, t)is the subdifferential of the functionf_j(x,·)at the pointt, that is:

∂₂f_j(x, t) ={z ∈X^∗/f_j(x, y)≥f_j(x, t) +hz, y−ti ∀y∈X}

(7)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page7of18

Definition 2.1. The Dual Problem of the problem (V P) (in short (DV P)), is the following generalized variational inequality:

to findu^∗ ∈R^m+ such that there existsd^∗ ∈G(u^∗) (DV P)

satisfying hd^∗, u−u^∗i ≥0, for allu∈R^m+. The problem (DV P) is termed a Dual Problem because we have:

Theorem 2.1. Assume that (H1) is satisfied and that x^∗ is a point of X such that E(x^∗) = ∩^m_j=1dom(f_j(x^∗,·)) is an open subset of X. If (x^∗, u^∗), with u^∗ ∈ R^m+, satisfies the following conditions, called ”Generalized Kuhn-Tucker Conditions”:

(KT)₁ :x^∗ ∈K(x^∗);

(KT)₂ :0∈A(x^∗) +Pm

j=1u^∗_j∂₂f_j(x^∗, x^∗);

(KT)₃ :F (x^∗, x^∗)∈N_R^m

+ (u^∗);

then

(i) x^∗is a solution to (V P) (ii) u^∗is a solution to (DV P).

Proof. First, to prove (i) we observe that:

“(x^∗, z^∗), withz^∗ ∈A(x^∗), solves (V P)”

(8)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page8of18

is equivalent to

“x^∗is a solution to the optimization problem (OP)”

where (OP) is:

(OP) min

x∈K(x^∗)hz^∗, x−x^∗i.

The problem (OP) admits as classical Lagrangian the functionL, fromE(x^∗)×

R^m toR, defined by:

L(x, u) =











hz^∗, x−x^∗i+

m

P

j=1

u_jf_j(x^∗, x) ifx∈E(x^∗) andu∈R^m+

−∞ u /∈R^m+

+∞ otherwise.

So to prove (i), it is sufficient to apply the Theorem 7.5.1 in ([14]) to the problem (OP), taking into account that N_E(x^∗₎(x^∗) = {0} (since E(x^∗) is open) and

∂(hz^∗, x−x^∗i) = z^∗+NE(x^∗)(x^∗).

Now we prove (ii). In light of the assumption (KT)₂, it follows that

−F (x^∗, x^∗) ∈ G(u^∗), where F and G are defined, respectively, by (2.1) and (2.2). So, sinceF (x^∗, x^∗)∈N_R^m₊ (u^∗)by assumption(KT)3, and

N_R^m₊ (u^∗) =

( v ∈R^m+/hv, u−u^∗i ≤0 ∀u∈R^m+ ifu^∗ ∈R^m+

∅ otherwise,

thenu^∗solves the problem (DV P) defined in Definition2.1.

(9)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page9of18

Theorem 2.2. Assume that (H1) is satisfied. Ifx^∗ is a solution to (V P) and if:

(i) E(x^∗) =∩^m_j=1dom(f_j(x^∗,·))is an open subset ofX (ii) ∃y∈Xsuch thatfj(x^∗, y)<0for allj = 1, . . . , m

then, there exists a point u^∗ ∈ R^m+ such that(x^∗, u^∗)satisfies the Generalized Kuhn-Tucker Conditions(KT)₁ to(KT)₃(and thereforeu^∗ solves (DV P) fol- lowing Theorem2.1).

Proof. Letx^∗ be a solution to (V P) andz^∗ ∈A(x^∗)such thathz^∗, x−x^∗i ≥0 for allx∈K(x^∗). By Theorem 7.5.2 in [14], there exists a pointu^∗ ∈R^m+ such that(x^∗, u^∗)is a saddle point for the LagrangianLabove defined. So, it results that:

0∈∂_xL(x^∗, u^∗) =z^∗+

m

X

j=1

u^∗_j∂₂f_j(x^∗, x^∗) which implies that0∈A(x^∗)+Pm

j=1u^∗_j∂₂f_j(x^∗, x^∗). Moreover, sinceL(x^∗, u)≤ L(x^∗, u^∗)for allu∈R^m+:

m

X

j=1

u_j −u^∗_j

f_j(x^∗, x^∗) = hF (x^∗, x^∗), u−u^∗i ≤0 ∀u∈R^m+

that isF(x^∗, x^∗)∈N_R^m

+ (u^∗). Therefore(x^∗, u^∗)satisfies(KT)₁to(KT)₃ and u^∗ solves (DV P).

In light of Theorems 2.1 and 2.2, the variational problem (DV P) can be considered as a dual problem associated to (V P).

(10)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page10of18

Remark 2.1. IfX =Rⁿ, for allx∈X:

K(x) =C ={z ∈X/f_j(z)≤0, for allj = 1,2, . . . , m}

and the generalized quasi-variational inequality comes from an optimization problem defined by a convex and differentiable function, then the previous the- orems reduce to the classical theorems of Convex Mathematical Programming (Theorems 3.2 and 3.3 in [2]).

Remark 2.2. Let us observe that the condition

E(x^∗) =∩^m_j=1dom(f_j(x^∗,·))is an open set of X

has been needed to properly handle convex programs within the formalism of extended valued functions ([14]).

By the previous theorems it follows that, to solve (V P), one can solve the dual problem (DV P) and then, using the generalized Kuhn-Tucker condition (KT)₂, one can find the solutions of problem (V P) proceeding as in the following example.

Example 2.1. If

K(x) = {y∈R/y−2x≤0andx−y≤0}

and

A(x) =











x− ¹₃,0

if0< x < ¹₃ [x,1] if ¹₃ ≤x≤1

∅ otherwise

(11)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page11of18

then the dual problem (DV P) associated to the primal problem (V P) is the easier generalized variational inequality:

to findu^∗ ∈R²+such that there existsd^∗ ∈G(u^∗) satisfying hd^∗, u−u^∗i ≥0, for all u∈R²+

where

G(u₁, u₂) =











0, u₂−u₁+ ¹₃

× {0} if − ¹₃ < u₂ −u₁ <0 ₁

3, u₂ −u₁

× {0} if ¹₃ ≤u₂−u₁ ≤1

∅ otherwise.

The solutions to the problem (DV P) are all the points(0, u₂)such that 1/3≤ u₂ ≤1, so, using the Generalized Kuhn-Tucker Condition(KT)₂, we find that all the pointsx^∗such that 1/3≤x^∗ ≤1are solutions to (V P).

(12)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page12of18

3. Application to Social Nash Equilibria

Let us consider a n-person noncooperative game with coupled constraints, as considered by G. Debreu in [7]. LetYi be a Banach space (or, more generally, a real locally convex topological vector space) and, for the playeri, letX_i ⊆Y_i be the strategy set,J_ifromX =X₁× · · · ×X_ntoRbe the payoff function, and

Ki(x−i) =

yi ∈Xi/f_jⁱ(yi, x−i)≤0,for allj = 1,2, . . . , mi

be the constraints depending on the strategies of the other players, where x−i

is a shorthand for (x_j)_j∈N\{i}. We assume that the players want to minimize their payoff function and play a Social Nash Equilibrium [7] (also called Gen- eralized Nash Equilibrium [10], which is a generalization of the concept of Nash Equilibria [15]). We recall that a Social Nash Equilibrium of the game Γ = {Xi, Ji, Ki}is a pointx^∗ ∈X such that no player can uniterally decrease his payoff given the constraints imposed on him by the other players; that is a point such that:

(SN E) J_i(x^∗)≤J_i x_i, x^∗_−i

for allx_i ∈K_i x^∗_−i

and for alli= 1, . . . n.

It is well known that, under suitable assumptions, the Social Nash Equilibrium problem can be put into the form of a generalized quasi-variational inequality (see for example [6, 4, 11]). More precisely, if we assume that the following condition is satisfied:

(H2) for every x−i ∈ X−i the functionJ_i(·, x−i)is convex and bounded from below onX_i, for alli= 1, . . . , n

(13)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page13of18

then, a pointx^∗is a solution to the problem (SN E) if and only ifx^∗ solves the following system of generalized quasi-variational inequalities:

(SN E)











findx^∗ ∈Xsuch thatx^∗ ∈K₁ x^∗₋₁

× · · · ×K_n x^∗_−n and there existz^∗₁ ∈∂_x₁J₁(x^∗),· · · , z_n^∗ ∈∂_x_nJ_n(x^∗) satisfying

hz₁^∗, x1−x^∗₁i ≥0, for all x1 ∈K1 x^∗₋₁ ...

hz_n^∗, xn−x^∗_ni ≥0, for all xn∈Kn x^∗_−n where∂_x_iJ_i is the subdifferential ofJ_i(·, x−i)for alli= 1, . . . , n.

Now, if we considered the set-valued operator defined onX by:

A(x) =∂_x₁J₁(x)× · · · ×∂_x_nJ_n(x) and

K(x) = {y∈X / y_i ∈K_i(x_−i)∀i= 1, . . . , n}

= {y∈X / fj(x, y)≤0 j = 1, . . . , m}

wherem=m₁+· · ·+m_nand

f_j(x, y) =











f_j¹(y₁, x−1) if j = 1, . . . , m₁ ...

f_jⁱ(y_i, x−i) ifj =Pi−1

r=1m_r+ 1, . . . ,Pi−1

r=1m_r+m_i ...

f_jⁿ(y_n, x−n) ifj =Pn−1

r=1 m_r+ 1, . . . , m

(14)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page14of18

thenx^∗ is a Social Nash Equilibrium forΓif and only if it solves the following generalized quasi-variational inequality:

findx^∗ ∈Xsuch thatx^∗ ∈K(x^∗) and there exists (SN E)

z^∗ ∈A(x^∗)satisfying hz^∗, x−x^∗i ≥0, for allx∈K(x^∗).

If the problem (SN E) satisfies the assumptions (H1) and (H2), we can define the dual problem:

findu^∗ ∈R^m+ such that there existsd^∗ ∈G(u^∗) (DSN E)

satisfying hd^∗, u−u^∗i ≥0, for allu∈R^m+, whereGis the set-valued operator defined by:

G(u) = (

−F (x, x)/0∈∂_x_h J_h(x) +

m

X

j=1

u_j∂_x_hf_j(x, x), for allh= 1, . . . , n )

.

Therefore, we can find the Social Nash equilibria of Γusing the method introduced in section 2, as one can see in the following example:

Example 3.1. Let us consider a two-player gameΓwith J₁(x, y) =x²+ 2x−y²

J₂(x, y) =y²+ 2xy and

K1(y) ={x∈R/ x−y ≤0}

(15)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page15of18

K₂(x) ={y ∈R/2x−y≤0} .

The Social Nash Equilibrium problem associated to this game is equivalent to the following generalized quasi-variational inequality:

find (x^∗, y^∗)∈K(x^∗, y^∗) (SN E)

such that (2x^∗+ 2) (x−x^∗) + (2y^∗+ 2x^∗) (y−y^∗)≥0 for all (x, y)∈K(x^∗, y^∗).

Since

G(u1, u2) = {(2u1+u2+ 4/2,3u1+u2 + 6/2)},

the dual of (SN E) is the easier problem:

findu^∗ ∈R²+such that (DSN E)

(2u^∗₁+u^∗₂+ 4/2) (u₁−u^∗₁) + (3u^∗₁+u^∗₂+ 6/2) (u₂−u^∗₂)≥0 for allu∈R²+.

The unique solution of (DSN E) is(u^∗₁, u^∗₂) = (0,0)and so, by the General- ized Kuhn-Tucker Condition(KT)2, we have that the point (x^∗, y^∗) = (−1,1) is a Social Nash Equlibrium for the gameΓ.

(16)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page16of18

References

[1] C.D. ALIPRANTIS, D. BROWNANDO. BURKINSHAW, Existence and Optimality of Competitive Equilibria, Springer Verlag, Berlin-Heidelbrg- New York-London-Paris-Tokyo, (1988).

[2] A. AUSLENDER, Optimisation. Méthodes Numériques, Masson, Paris- New York-Barcelona, (1976).

[3] A. AUSLENDER AND M. TEBOULLE, Lagrangian duality and related multiplier methods for variational inequality problems, SIAM J. Optim., 10(4) (2000), 1097–1116.

[4] A. BENSOUSSAN, Points de Nash dans le cas de fonctionnelles quadra- tiques et jeux differentials linéaires aN personnes, SIAM J. on Control, 12 (1974), 460-499.

[5] J.P. CROUZEIX, Pseudomonotone variational inequality problems: Exis- tence of solutions, Math. Program., 78 (1997), 305–314.

[6] C. BAIOCCHI AND A. CAPELO, Disequazioni Variazionali e Quasi- variazionali. Applicazioni a Problemi di Frontiera Libera, Quadernidell’

U.M.I., Pitagora Editrice, Bologna, (1978).

[7] G. DEBREU, A social equilibrium existence theorem, Proc. Nat. Acad.

Sci., USA, 38 (1952), 886–893.

[8] X.P. DING ANDK.K. TAN, Generalized variational inequalities and gen- eralized quasi-variational inequalities, J. Math. Anal. Appl., 148 (1990), 497–508.

(17)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page17of18

[9] P.T. HARKERAND J.S. PANG, Finite-dimensional variational inequality and non linear complementarity problems: a survey of theory, algorithms and applications, Math. Program., 48 (1990), 171–220.

[10] T. ICHIISHI, Game Theory for Economic Analysis, Academic Press, New York, (1983).

[11] M.B. LIGNOLA ANDJ. MORGAN, Generalized variational inequalities with pseudomonotone operators under perturbations, J. Optim. Theory Appl., 101 (1999), 213–220.

[12] M.B. LIGNOLA AND J. MORGAN, Convergence of solutions of quasi- variational inequalities and applications, Topol. Methods Nonlinear Anal., 10 (1997), 375–385.

[13] M.B. LIGNOLA AND J. MORGAN, Approximate solutions and alpha- well-posedness for variational inequalities and Nash equilibria, in De- cision and Control in Management Science, Kluwer Academic Science, (2002), 367–378.

[14] P.J. LAURENT, Approximation and Optimisation, Hermann, Paris- London,(1972).

[15] J.F. NASH Jr., Equilibrium points in n-person games, Proc. Nat. Acad.

Sci., USA, 36 (1950), 48–49.

[16] M.H. SHIH ANDK.K. TAN, Generalized quasi-variational inequalities in locally convex vector spaces, J. Math. Anal. Appl., 108 (1985), 333–343.

(18)

Title Page Contents

JJ II

J I

Go Back Close

Quit Page18of18

[17] G. STAMPACCHIA, Variational inequalities, in theory and applications of monotone pperators, Proc. NATO Advanced Study Inst., Edizioni Oderisi, (1968), 101–192.

[18] J.V. TIEL, Convex Analysis. An Introduction Text, John Wiley and Sons, (1984).