Journal of Applied Mathematics & Bioinformatics, vol.3, no.2, 2013, 45-64 ISSN: 1792-6602 (print), 1792-6939 (online) Scienpress Ltd, 2013 On Nondifferentiable Nonlinear Programming I. Husain 1 and Santosh K. Srivastava 2 Abstract In this paper, we obtain optimality conditions for a class of nondifferentiable nonlinear programming problems with equality and inequality constraints in which the objective contains the square root of a positive semidefinite quadratic function and is ,therefore, not differentiable. Using Karush-Kuhn-Tucker optimality, Mond-Weir dual to this problem is constructed and various duality results are validated under suitable generalized invexity hypotheses. A mixed type dual to the problem is also formulated and duality results are similarly derived. Mathematics Subject Classification: 90C30, 90C11, 90C20, 90C26 Keywords: Nondifferentiable Nonlinear Programming, Generalized Invexity, Mixed type duality, Square root function 1 2 Department of Mathematics, Jaypee University of Engineering and Technology, Guna (M.P), India, e-mail:ihusain11@yahoo.com Department of Mathematics, Jaypee University of Engineering and Technology, Guna (M.P), India Article Info: Received : December 14, 2012. Revised : February 12, 2013 Published online : June 30, 2013 46 On Nondifferentiable Nonlinear Programming 1 Introduction In mathematical programming, there are a large number of researchers who have discussed duality for a problem involving the square root of a positive semidefinite. Mond studied Wolfe type duality while Chandra et. al. [1] investigated Mond-Weir type duality for a class of nondifferentiable mathematical programming problem containing square root term. Subsequently several research papers have appeared in the literature, namely Chandra and Husain [2] and references exist there. The popularity of this kind of problem seems to originate from the fact that, even though the objective functions and or constraint functions are non-smooth, a simple representation for the dual problem may be found. The nonsmooth mathematical programming deals with much more general kind of functions by means of generalized subdifferential, or quasidifferentials. However, the square root of a positive semidefinite quadratic form is one of the four cases of a nondifferentiable function for which subdifferential can explicitly be written. In this paper, we obtain optimality conditions for a class of nondifferentiable nonlinear programming problem having square root term in the objective function with equality and inequality constraints. It is to be remarked here, a constrained optimization problem with equality and inequality constraints represents a majority of real life problems, and hence it is important. For this class of problem, Mond-Weir duality is investigated using generalized invexity assumptions. A mixed type dual problem to this problem is also constructed to obtain various duality results. 2 Statement of the Problem and Related Pre-requisites We consider the following nondifferentiable nonlinear programming problems: I. Husain and Santosh K. Srivastava (EP): 47 Minimize f ( x ) + ( xT B x ) 1/ 2 subject to g ( x ) ≤ 0, (1) h ( x ) = 0, (2) Where (i) f : R n → R, g : R n → R m and h : R n → R p are continuously differentiable functions. (ii) B is an n × n symmetric positive semi-definite matrix. We recall the following definitions of generalized invexity which will be used to derive various duality results. Definitions 2.1: (i) A function φ : R n → R is said to be quasi-invex with respect to a vector function η = η ( x,u ) , if φ ( x ) ≤ φ ( u ) ⇒ η T ( x,u ) ∇ φ ( u ) ≤ 0 . (ii) A function φ is said to be pseudo-invex with respect to a vector function η = η ( x,u ) , if η T ( x,u ) ∇ φ ( u ) ≥ 0 ⇒ φ ( x ) ≥ φ ( u ) . (iii) φ is said to be the strictly pseudoinvex with respect to η if x ≠ u, η T ( x,u ) ∇ φ ( u ) ≥ 0 ⇒ φ ( x ) > φ ( u ) Equivalently, if φ ( x ) ≤ φ ( u ) ⇒ η T ( x,u ) ∇ φ ( u ) < 0. We shall make use of the generalized Schwartz inequality [3] (x T B ω ) ≤ ( xT B x ) 1 / 2 ( ω T B ω ) 1 / 2 The equality holds if, for λ ≥ 0 , B x = λ B ω . The function φ ( x ) = ( xT B x ) , 1/2 48 On Nondifferentiable Nonlinear Programming being convex and everywhere finite, has a subdifferential in the sense of convex analysis. The subdifferential of ∂ ( xT B x ) 1/2 = (x T B x) 1/2 is given by {Bw x Bw = ( x B x ) T T 1/2 } , where w ∈ R n , and wT B w ≤ 1 . We also require the Mangasarian-Fromovitz constraint qualification which is described as the following: Let x ∈ Ω be the set of feasible solution of the problem (EP), that is, Ω= {x ∈ R n } g ( x ) ≤ 0, h ( x )= 0 , and by A ( x ) , the set of inequality active = A( x ) constraint indices, that is, j , g j ( x ) 0} , {= where x ∈ Ω . We say the Mangasarian-Fromovitz constraint qualification holds at x ∈ Ω when the equality constraint gradients ∇h1 ( x ) , ∇h2 ( x ) ,..., ∇hp ( x ) are linearly 0 and independent and there exist a vector d ∈ R n such that ∇h ( x ) d = ∇g j ( x ) d < 0, for all j ∈ A ( x ) . 3 Optimality Conditions In this section, optimality conditions for the problem (EP) are obtained. Theorem 3.1 (Fritz John Necessary Optimality Condition) If x is an optimal solution of (EP), then there exist r ∈ R, y ∈ R m , z ∈ R p and w ∈ R n such that I. Husain and Santosh K. Srivastava 49 r∇f ( x ) + B ω + ∇yT g ( x ) + ∇z T h ( x ) = 0, yT g ( x ) = 0, xT B w=( xT B x ) , 1/2 ω T B ω ≤ 1, ( r, y ) ≥ 0 ( r , y, z ) ≠ 0. Proof: The problem (EP) may be written as Minimize φ= ( x ) f ( x ) +ψ ( x ) , Subject to − g ( x ) ∈ R+m , h ( x ) = 0. Where R+m is the non-negative orthant of R m and the nondifferentiable convex function ψ = ( xT B x ) . 1/2 A Fritz John theorem [4] shown that the necessary condition for a minimum of (EP) at x are the existence of Lagrange multipliers r ∈ R, y ∈ R m , z ∈ R p and ω∈ R n such that 0 ∈ r ∂ φ ( x ) + r ∂ψ ( x ) + ∂ yT g ( x ) + ∂z T h ( x ) This implies 0 ∈ r {∇f ( x )} + r{B= ω ω ∈ Rn , x T B w (x T B x ) , ω T B ω ≤ 1} + {∇yT g ( x ) + ∇z T h ( x )} 1/2 which implies r ( ∇f ( x ) + B ω ) + ∇ yT g ( x ) + ∇ z T h ( x ) = 0, yT g ( x ) = 0, xT B w=( xT B x ) , 1/2 ω T B ω ≤ 1, ( r , y ) ≥ 0, ( r , y, z ) ≠ 0. 50 On Nondifferentiable Nonlinear Programming Thus the theorem follows. Karush-Kuhn-Tucker type optimality conditions can be deduced from the above Fritz John optimality condition if Mangasarian-Fromovitz Constraint Qualification holds at x ∈ Ω. The following theorem gives the Karush-Kuhn-Tucker types optimality conditions. Theorem 3.2 (K-K-T optimality condition) If x is an optimal solution of (EP) and satisfies Mangasarian-Fromovitz constraint qualification, then there exist r ∈ R, y ∈ R m , z ∈ R p and ω ∈ R n such that ∇f ( x ) + B ω + ∇yT g ( x ) + ∇z T h ( x ) = 0, yT g ( x ) = 0, xT B w=( xT B x ) , 1/2 ω T B ω ≤ 1, y ≥ 0. 4 Mond-Weir type Duality In this section, we present the Mond-Weir type dual to (EP). Using Karush-Kuhn-Tucker necessary optimality conditions, Wolfe type dual to the following problem was formulated in [5]. Problem: (P) Minimize f ( x ) + ( xT B x ) 1/ 2 Subject to g ( x) ≤ 0 . Dual (WD): Maximize f ( u ) + yT g ( u ) − u T ( ∇yT g ( u ) + ∇f ( u ) ) I. Husain and Santosh K. Srivastava 51 subject to ∇f ( u ) + B ω + ∇y T g ( u ) = 0 , ω T B ω ≤ 1, y ≥ 0. The problem (WD) is a dual to (P) assuming that f and g are convex: Further, Chandra et. al. [1] in order to weaken the convexity requirements in [5] for duality to hold, formulated the following Mond-Weir type dual to the problem (P). (M-WD): Maximize f ( u ) + u T B ω subject to ∇f ( u ) + B ω + ∇y T g ( u ) = 0 , yT g ( u ) ≥ 0, ω T B ω ≤ 1, y ≥ 0. and established duality results assuming that f (.) + (.) B ω is pseudoconvex for T all ω ∈ R n and yT g (.) is quasiconvex. Later Mond and Smart [6] generalized the results by Zhang and Mond [7] and Chandra et. al. [1] to invexity conditions. Here, we propose the following Mond-Weir type dual to the problem (EP) to study duality: (M-WED): Maximize f ( u ) + u T B ω Subject to ∇f ( u ) + B ω + ∇yT g ( u ) + ∇z T h ( u ) = 0, (3) ω T B ω ≤ 1, (4) yT g ( u ) ≥ 0, (5) z T h ( u ) ≥ 0, (6) y ≥ 0. (7) 52 On Nondifferentiable Nonlinear Programming ( u, y, ω ) Theorem 4.1 (Weak Duality) Let x be feasible for (EP) and feasible for (M-WED). If for all feasible ( x, u , y, ω ) , f (.) + (.) B ω T is pseudoinvex, yT g ( .) and zT h ( .) are quasi-invex with respect to the same η , then infimum (EP) ≥ supremum (M-WED) Proof: Since x is feasible for (EP) and ( u , y, ω ) is feasible for (M-WED), we have yT g ( x ) ≤ yT g ( u ) (8) zT h ( x ) ≤ zT h ( u ) (9) By quasi-invexity of yT g (.) and zT h ( .) with respect to the same η , (8) and (9) respectively yield η T ( x , u ) ∇y T g ( u ) ≤ 0 (10) η T ( x , u ) ∇z T h ( u ) ≤ 0 (11) Combining (10) and (11), we have η T ( x,u ) ( ∇yT g ( u ) + ∇z T h ( u ) ) ≤ 0 which because of (3) gives η T ( x , u ) ( ∇f ( x ) + B ω ) ≥ 0 By pseudoinvexity of f ( .) + ( .) B ω , this implies T f ( x ) + xT B ω ≥ f ( u ) + u T B ω Since ω T B ω ≤ 1, by the generalized Schwartz inequality, it follows that f ( x ) + ( xT B x ) 1/2 ≥ f ( u ) + uT B ω giving infimum (EP) ≥ supremum (M-WED). I. Husain and Santosh K. Srivastava 53 Theorem 4.2 (Strong Duality) If x is an optimal solution of (EP) and MFCQ holds at x , then there exist y ∈ R m , z ∈ R p and ω ∈ R n such that ( x , y, ω , z ) is feasible for (M-WED) and the corresponding values of (EP) and (M-WED) are equal. If, also f (.) + (.) B ω is pseudoinvex for all ω ∈ R n , yT g ( .) and zT h ( .) T are quasi-invex with respect to the same η . Then ( x , y, ω , z ) is an optimal solution of (M-WED). Proof: Since x is an optimal solution of (EP) and MFCQ holds at x , then from Theorem 3.2, then there exist y ∈ R m , z ∈ R p and ω ∈ R n such that ∇f ( x ) + B ω + ∇yT g ( x ) + ∇z T h ( x ) = 0, yT g ( x ) = 0, xT Bω =( xT B x ) , 1/2 ω T B ω ≤ 1, y ≥ 0. Since x is feasible for (EP) and (M-WED), h ( x ) = 0, which implies z T h ( x ) = 0, ∇f ( x ) + B ω + ∇yT g ( x ) + ∇z T h ( x ) = 0, yT g ( x ) = 0, ω T B ω ≤ 1, y ≥ 0. and If f ( .) + ( .) B ω T is pseudoinvex for all ω ∈ R n , yT g (.) and z T h (.) are quasi-invex with respect to the same η . Then from Theorem 4.1, ( x , y, ω , z ) is an optimal solution for (M-WED). The following is another dual to the following (EP): (M-WED1): Maximize subject to f ( u ) + uT B ω 54 On Nondifferentiable Nonlinear Programming ∇f ( u ) + B ω + ∇yT g ( u ) + ∇z T h ( u ) = 0, yi gi ( u ) ≥ 0, i = 1, 2,..., m z T h ( u ) ≥ 0, ω T B ω ≤ 1, y ≥ 0. This is a dual of (EP) if f ( .) + ( .) B ω is pseudo-invex and each yi gi , T i = 1, 2,..., m is quasi-invex and zT h ( u ) is quasi-invex with respect to the same η . It is remarked here that if gi ( .) is quasi-invex with respect to same η , yi ≥ 0, then yi gi is quasi-invex with respect to η . The problem (M-WED1) is a dual to (EP) if f (.) + (.) B ω is pseudoinvex and each gi ,i = 1, 2 ,..m is quasi-invex and T zT h ( .) is quasi-invex with respect to the same η . Theorem 4.3 (Strict Converse duality) Assume that f (.) + (.) B ω for all T ω ∈ R n is strictly pseudoinvex, yT g (.) is quasi-invex and z T h (.) is quasi-invex with respect to the same η . Assume also that (EP) has an optimal solution x which satisfies Mangasarian-Fromovitz constraint qualification. If ( u , y, z , ω ) is an optimal solution. Then u is an optimal solution of (EP) with u = x . Proof: We assume that u ≠ x and exhibit a contradiction. Since x is an optimal solution of (EP), it implies from Theorem 4.2 that there exists y ∈ R m , z ∈ R p and ω ∈ R n such that ( u , y, z , ω ) ( x , y, z , ω ) is an optimal solution of (M-WED1). Since is also an optimal solution of (M-WED1), it follows that f ( x ) + xT Bω = f (u ) + u B ω This, by strict pseudoinvexity of f (.) + (.) B ω yields T I. Husain and Santosh K. Srivastava 55 η T ( x , u ) ( ∇f ( u ) + B ω ) < 0 (12) From the constraints of (EP) and (M-WED), we have yT g ( x ) ≤ yT g ( u ) , zT h ( x ) ≤ zT h ( u ) , which by quasi-invexity of yT g (.) and z T h (.) with respect to the same η , yield η T ( x , u ) ∇y T g ( u ) ≤ 0 (13) η T ( x , u ) ∇z T h ( u ) ≤ 0 (14) From (12), (13) and (14), we have η T ( ∇f ( x ) + B ω ) + ∇ yT g ( u ) + ∇ z T h ( u ) < 0, contradicting the feasibility of ( u , y, z , ω ) for (M-WED). Hence u = x. Theorem 4.4 (Converse duality) Let ( x , y , z , ω ) be optimal to (M-WED) at which ( A1 ) : the matrix ∇ 2 ( f ( x ) + y T g ( x ) + z T h ( x ) ) , is positive or negative definite and ( A 2 ) : the vectors ∇ y T g ( x ) and ∇ z T h ( x ) are linearly independent. If, for all feasible ( x, u , y , z , ω ) , f (.) + (.) B ω is pseudoinvex, yT g ( .) and T z T h (.) are quasi-invex with respect to the same η , then x is an optimal solution of (EP). Proof: By Theorem 3.1, there exist τ ∈ R, θ ∈ R n , α ∈ R, β ∈ R, η ∈ R m and γ ∈ R such that 56 On Nondifferentiable Nonlinear Programming τ ( ∇f + B ω ) + θ T ( ∇ 2 f ( u ) + ∇ 2 y T g ( u ) + ∇ 2 z T h ( u ) ) + α ∇y T g + β ∇z T h ( u ) = 0, θ ∇g ( u ) + α g + η = 0, (15) (16) Implies θ T ∇y T g ( u ) = 0, (17) θ T ∇z T h ( u ) = 0, (18) 0, τ ( xT B ) + θ B − 2 γ B ω = (19) α yT g = 0, (20) β zT g = 0, (21) γ (1 − ω B ω ) = 0, T (22) η T y = 0, (23) (τ , α , β , γ ,η ) ≥ 0, (24) (τ , α , β , r ,η ,θ ) ≠ 0. (25) Using (3) in (15), we have τ ( ∇y T g ( u ) + ∇z T h ( u ) ) + θ ∇ 2 ( f ( u ) + y T g ( u ) + z T h ( u ) ) 0, + α ∇y T g ( u ) + β ∇ 2 ( z T h ( u ) ) = 0, ( α − τ ) ∇y T g ( u ) + ( β − τ ) ∇ 2 z h ( u ) + θ T ∇ 2 ( f ( u ) + y T g ( u ) + z T h ( u ) ) = ⇒ (α − τ ) θ T ∇y T g ( u ) + ( β − τ ) θ T ∇ 2 z T h ( u ) + θ T ∇ 2 ( f ( u ) + y T g ( u ) + z T h ( u ) ) = 0, Which because of (17) and (18), yields θ T ∇ 2 ( f ( u ) + yT g ( u ) + zT h ( u ) ) = 0, By ( A1 ) , it follows that θ = 0. Then (15) implies 0. (α − τ ) ∇ y T g ( u ) + ( β − τ ) ∇ z T h ( u ) = I. Husain and Santosh K. Srivastava 57 0, By the linear independent of ∇ y T g ( x ) and ∇ z T h ( x ) , this gives (α − τ ) = 0. and ( β − τ ) = Let= θ 0, τ > 0 . Then = α 0,= β 0. Consequently (16) implies η = 0. The relation (19) and (22) together imply that γ = 0. This leading to a contradiction to (25). Hence τ > 0, α > 0 and β > 0. Also= θ 0, τ > 0 and (19) give xT B = 2γ τ B ω. Hence x T B ω = ( x T B x ) 1/2 (ω T Bω) . 1/2 If γ > 0, then (22) give ω T B ω = 1 and so x T B ω = ( x T B x ) . 1/2 If γ = 0, (19) gives B x = 0. so we still get (x T Bω) =( xT B x ) . 1/2 (26) Thus, in earlier case, we obtain (x T Bω) =( xT B x ) . 1/2 Therefore, from (26), we have f ( x ) + ( xT B x ) 1/2 = f ( x ) + ( x T B ω ). The equality of objective values follows. If, for all feasible ( x , u, y, z , ω ) , f (.) + (.) B ω is pseudoinvex, yT g (.) and T z T h (.) is quasi-invex with respect to the same η , then by Theorem 4.1 , it implies x is an optimal solution of (EP). 5 Mixed type duality = Let M 2,..., m} , L {1, 2,...l} , Iα {1,= ⊆ M, α = 0,1, 2,..., r 58 On Nondifferentiable Nonlinear Programming r with Iα = Iβ φ, α ≠ β Jα ∩ J β= φ , (α ≠ β ) and and Iα = M and Jα ⊆ L, α = 0,1, 2,3..., r with α =0 r Jα = L . α =0 In relation to (EP), consider the problem. Mix (ED): Maximize f ( u ) + u T B ω + ∑ y g (u ) + ∑ z h (u ) i i∈I 0 i j j∈J 0 j subject to ∇f ( u ) + B ω + ∇ y T g ( u ) + ∇z T h ( u ) = 0 (27) ∑ y g ( u ) ≥ 0, α= 1, 2,..., r. (28) ∑ z h ( u ) ≥ 0, α= 1, 2,..., r. (29) i∈Iα j∈Jα i i j j ωT B ω ≤ 1 (30) (31) y≥ 0 Theorem 5.1 (Weak Duality) Let x be feasible and (ED). If for all feasible ( u, y, z, ω ) feasible ( x, u, y, z, ω ) , f (.) + (.)T B ω + ∑ yi gi (.) + ∑ z j h j (.) i∈I 0 pseudoinvex, ∑ y g (.), α = 1, 2,..., r i∈Iα i i is quasi-invex for is j∈J 0 ∑ z h (.), and j∈Jα j j α = 1, 2,..., r is quasi-invex with respect to the same η , then f ( x ) + ( xT B x ) 1/2 ≥ f (u ) + u T B ω + ∑ y g (u ) + ∑ z h (u ) i∈I 0 i i j∈J 0 j j That is, infimum (EP) ≥ supremum Mix (ED). Proof: Since x is a feasible for (ED) and we have ( u, y, z , ω ) is feasible for Mix (ED), I. Husain and Santosh K. Srivastava ∑ y g ( x ) ≤ ∑ y g (u ) i i∈Iα i i∈Iα By quasi-invexity of i 59 , α= 1, 2...r i ∑ y g (.) , α = 1, 2,..., r , i∈Iα i i this implies η T ( x, u ) ∇ ∑ yi gi ( u ) ≤ 0, α = 1, 2,..., r ∈ i I α η T ( x, u ) ∇ ∑ yi gi ( u ) ≤ 0, α = 1, 2,..., r ∈ − i M I T f (.) + (.) B ω + ∑ yi (.) gi (.) + ∑ z j (.) h j (.) is pseudo-invex. (32) 0 i∈I 0 Also j∈J 0 1, 2,...r . ∑ z h ( x ) ≤ ∑ z h (u ) , α = j j∈Jα j j∈Jα j j By quasi-invexity of ∑ z j h j (.) , α = 1, 2,...r j∈Jα This gives η T ( x, u ) ∇ η T ( x, u ) ∇ ( ∑ j∈L − J 0 ∑ z h (u ) j∈Iα j j 1, 2,..., r ≤ 0, α = 1, 2,..., r z j h j ( u )) ≤ 0, α = (33) From the (27), it follows that η T ( x, u ) [∇ f ( u ) + B ω + ∇∑ yi gi ( u ) + ∇ ∑ z j h j ( u )] ≥ 0 i∈I 0 The pseudoinvexity of f (.) + (.) B ω + j∈J 0 ∑ y g (u ) + ∑ z h (u ) , i∈I 0 i i f ( x ) + ( xT B ω ) ≥ f ( u ) + u T B ω + j∈J 0 j this yield j ∑ y g (u ) + ∑ z h (u ) i∈I 0 i i j∈J 0 j j Since ω T B ω ≤ 1, by the generalized Schwartz inequality, this implies f ( x ) + ( xT B x ) 1/2 ≥ f ( u ) + u T B ω + ∑ y g (u ) + ∑ z h (u ) i∈I 0 i i implying, infimum (EP) ≥ supremum Mix (ED). j∈J 0 j j 60 On Nondifferentiable Nonlinear Programming Theorem 5.2 (Strong Duality) If x is an optimal solution of (EP) and Mangasarian-Fromovitz Constraint Qualification is satisfied at x , then there exist y ∈ R m , z ∈ R l and ω ∈ R n such that ( x, u , z , ω ) is feasible for Mix (ED) and the corresponding value of (EP) and Mix(ED) are equal. f (.) + (.) B ω + ∑ yi gi (.) + ∑ z j h j (.) is pseudoinvex for all ω ∈ R n , i∈I 0 and j∈J 0 If, also, ∑ y g ( .) i∈Iα i i ∑ z h (.) are quasi-invex with respect to the same η , then ( x, y, z, ω ) j∈Jα j j is optimal for Mix (ED). Proof: Since x is optimal solution of (EP) and MFCQ is satisfied at x , then from Theorem 3.2, there exist y ∈ R m , z ∈ R l and ω ∈ R n such that f ( x ) + B ω + ∇ yT g ( x ) + ∇z T h ( x ) = 0, yT g ( x ) ≥ 0, ω T B ω ≤ 1, x Bω = ( x T B x ) , 1/2 y ≥ 0. The relation yT g ( x ) ≥ 0, and z T h ( x ) ≥ 0, are obvious. From the above it implies that ( x , y, z , ω ) is feasible and the corresponding value of (EP) and (Mix ED) are equal. If f ( .) + ( .) Bω + ∑ yi gi ( .) + T i∈I 0 ∑ z h ( .) j∈J 0 j j is quasi-invex, yT g (.) and z T h (.) are quasi-invex with respect to the same η , then from Theorem 5.1, ( x, y , z , ω ) must be an optimal solution of (Mix ED). We now consider some special cases of (Mix ED). If= I 0 φ= , Iα M for some α ∈{1, 2,..., r} and = Jα φ= , α 0,1, 2,..., r. Then (Mix ED) be (WD) and (WD) is dual to (EP) if f (.) + (.) B ω + yT g (.) is T pseudoinvex with respect to the same η . I. Husain and Santosh K. Srivastava 61 If = Jα φ= , α 0,1, 2,..., r. Then (Mix ED) become the Mixed dual to the problem (P) considered by Zhang and Mond [7] and generalized dual to (P) if f ( .) + ( .) B ω + y T g ( .) T is pseudoinvex ∑ y g (.) , α = 1, 2,..., r and i∈Iα i is i quasi-invex with respect to the same η . Theorem5.3 (Converse duality) Let ( x , y,z,ω ) be an optimal solution of Mix (ED) at which ( A1 ) : the matrix ∇ 2 ( f ( x ) + yT g ( x ) + z T h ( x ) ) is positive or negative definite and ( A2 ): the vectors ∑ ∇yi gi ( x ) , i∈Iα 1, 2,..., r are ∑ ∇ z h ( x ), α = j∈Jα j j independent. If f (.) + (.) B ω is pseudoinvex, T linearly ∑ y g ( .) and ∑ z h (.) i∈Iα i i j∈Jα j j are quasi-invex with respect to the same η . Then x is an optimal solution of (EP). Proof: ( τ (∇f ( x ) + B ω + ∑ ∇ yiT gi ( x ) + ∑ ∇z j h j ( x )) + θ T ∇ 2 ( f ( x ) + yT g ( x ) + z T h ( x ) ) i∈I 0 j∈J 0 r r + ∑ pα ∑ ∇yi gi ( x ) + ∑ qα ∑ ∇z j h j ( x ) = 0 ∈ = ∈ α= α i I j J 1 1 α α T τ B x +θ B − 2γ B ω = 0 ) (34) (35) τ gi ( x ) + θ ∇g i ( x ) +η =0, i ∈ Iα , α = 1, 2,..., r (36) τ h j ( x ) + θ T ∇h j ( x )= 0, j ∈ Jα , α= 1, 2,..., r (37) pα ( ∑ yi gi (= x )) 0,= α 1, 2,..., r (38) α 1, 2,..., r qα ( ∑ z j h j (= x )) 0,= (39) η T y= 0 (40) γ (1 − ω T B ω ) = 0 (41) T i∈Iα j∈Jα 62 On Nondifferentiable Nonlinear Programming (τ , p1 , p2 ,..., pr , γ ,η , q1 , q2 ,..., qr ) ≥ 0 (τ , p1 , p2 ,..., pr , γ ,θ ,η , q1 , q2 ,..., qr ) ≠ 0 (42) (43) From (36), it follows that θT ∑ ∇y g ( x ) =0, i i∈Iα α =1, 2,..., r i (44) From (37), it follows that θT 0, α = 1, 2,..., r ∑ ∇Z h ( x ) = J j∈ jα (45) j Using equality constraint of the problem Mix(ED) in (34), we have r r ( pα − τ ) ( ∑ ∇y g ( x )) + ∑ ( qα − τ )( ∑ ∇z h ( x )) ∑ α α = 1 i i∈Iα + θ T ∇2 i = 1 j j∈Jα j 0 ( f ( x ) + y g ( x ) + z h ( x )) = T T (46) Multiplying (46) by θ , and then using (44) and (45), we have θ T ∇ 2 ( f ( x ) + yT g ( x ) + z T h ( x ) )θ = 0. By ( A1 ) , from this we obtain θ = 0. (47) Then (46) implies r ∑ ( pα − τ ) ( ∑ ∇yi gi ( x )) + 1 α= Since the i∈Iα vector ( qα − τ ) ( ∑ ∇z h ( x )) ∑ α {∑ ∇yi gi ( x ), i∈Iα r = 1 j∈Jα j 1, 2,..., r} are ∑ ∇z h ( x ) , α = j∈Jα j j (48) j linearly independent, (48) yields pα − τ= 0, qα − τ= 0, α= 1, 2,..., r If τ = 0, (49) implies pα= 0= qα , α= 1, 2,..., r . (49) From (36) implies η = 0 and from (35) together with (41), we have γ = 0. Hence (τ , θ , p1 , p2 ,..., pr , γ ,η , q1 , q2 ,..., qr ) = 0 contradicting to (43), hence τ > 0. Consequently pα > 0 and qα > 0, α = 1, 2,..., r. Multiplying (36) by yi , i ∈ I 0 and using (40), we have I. Husain and Santosh K. Srivastava 63 τ yi gi (= x ) 0, i ∈ I 0 (50) Multiplying (37) by z j , j ∈ I 0 , we have τ z j h j (= x ) 0, j ∈ J 0 Then from τ > 0, (47) and (48) implies that (51) yi gi ( x= ) 0, i ∈ I 0 (52) z j h j (= x ) 0, j ∈ J 0 (53) Also = θ 0,τ > 0 and (35) implies Bx= 2r τ Bω (54) Hence xT Bω = ( xT B x) 1/2 (ω T Bω) 1/2 (55) If γ > 0, then (41) implies ω T B ω = 1. xT Bω =( xT B x ) 1/2 consequently, (55) yields If γ = 0, then (35), yields B x = 0. So we obtain (x T B x ) 1/2 = x T B ω . Thus in either case, we obtain x T B ω = ( x T B x ) . 1/2 Therefore from (49) and (50), with x T B ω = ( x T B x ) , we have 1/2 f ( x ) + ( xT B x ) 1/2 =f ( x ) + x T B ω + ∑ y g (x) + ∑ z h (x) . i∈ I 0 i i j∈ J 0 j j If the hypotheses of Theorem 5.1 hold then x is an optimal solution of (EP). References [1] S. Chandra, B.D. Craven and B. Mond, Generalized concavity and duality with a square root term, Optimization, 16, (1985), 654-662. [2] S. Chandra and I. Husain, Symmetric dual nondifferentiable programs, 64 On Nondifferentiable Nonlinear Programming Bulletin of the Australian Mathematical Society, 24, (1981), 295-307. [3] F. Riesz and B. Sz.-Nagy, Functional analysis, (translated from the 2nd French edition by Leo F. Boron), Fredrick Ungar Publishing Co., New York, 1955. [4] B.D. Craven and B. Mond, Lagrangean conditions for quasidifferentiable optimization, University of Melbourne, School of Mathematical Sciences, Research Report No. 7, (1976), Survey of Mathematical Programming, North-Holland, Vol. I, pp.177-192, 1979. and Proceedings of IX International Symposium on Mathematical Programming, Budapest, (August, 1976). [5] B. Mond, A class of nondifferentiable mathematical programming problems, J. Math. Anal. Appl., 46, (1974), 169-174. [6] B. Mond and I. Smart, Duality with invexity for a class of nondifferentiable static and continuous programming problems, J. Math. Anal. Appl., 141, (1989), 373-388. [7] J. Zhang and B. Mond, Duality for a nondifferentiable programming problem, Bull. Austral. Math. Soc., 55, (1997), 29-44.