MINIMAX PROGRAMMING
UNDER NEW INVEXITY ASSUMPTIONS
MARIA VIORICA S¸TEF ˘ANESCU and ANTON S¸TEF ˘ANESCU
We consider the minimax programming problem and focus on optimality condi- tions and duality properties. Some of previously known theoretical results are now restated in terms of (Φ, ρ)-invexity and related properties.
AMS 2000 Subject Classification: 90C26, 90C46.
Key words: minimax programming, invexity, duality.
1. INTRODUCTION
The theory of mathematical programming has grown remarkably after generalized convexity has been used in the settings of optimality conditions and duality theory. In 1981, Hanson [4] showed that both weak duality and Kuhn- Tucker sufficiency for optimum hold when convexity is replaced by a weaker condition. This condition, called invexity by Craven [2], was further studied for more general problems and was a source of a vast literature. After the works of Hanson and Craven, other types of differentiable functions have been introduced with the intent of generalizing invex functions from different points of view. Hanson and Mond [5] introduced the concept of F-convexity and Jeyakumar [3] generalized Vial’s ρ-convexity ([7]) by introducing the concept ofρ-invexity. The concept of generalized (F, ρ)-convexity, introduced by Preda [6], is in turn an extension of the above properties and was used by several authors to obtain relevant results.
(F, ρ)-convexity was recently generalized to (Φ, ρ)-invexity by Caristi et al. [1] to extend the main theoretical results of mathematical programming.
Now we consider the minimax programming problem involving (Φ, ρ)- invex functions, and obtain optimality conditions and duality results.
2. (Φ, ρ)-INVEXITY AND RELATED EXTENSIONS Letϕ:X0 →Rbe a differentiable function (X0 ⊆Rn), anda∈X0. In the definitions below, an element of the (n+ 1)-dimensional Euclidean space Rn+1 is represented as an ordered pair (y, r), with y ∈ Rn and r ∈ R,
REV. ROUMAINE MATH. PURES APPL.,52(2007),3, 367–376
and Φ is a real-valued function defined onX0×X0×Rn+1such that Φ(x, a, .) is convex onRn+1 and Φ(x, a,(0, r))≥0 for every (x, a)∈X0×X0 andr ∈R+. Now, fix a nonvoid subsetX of X0,a point a∈X0, and a real number ρ.Following [1] we recall some basic notions and define some new extensions of them.
Definition 1. ϕis said to be (Φ, ρ)-invexat awith respect to X if (1) ϕ(x)−ϕ(a)≥Φ(x, a,(∇ϕ(a), ρ)), ∀x∈X.
Remark. For Φ(x, a,(y, ρ)) =F(x, a, y) +ρd2(x, a),withF sub-linear in the third argument, (Φ, ρ)-invexity reduces to Preda’s (F, ρ)-convexity. More- over, an example in [1] shows that (Φ, ρ)-invexity may hold when other known invexity-type conditions are not satisfied.
Definition 2. ϕis said to bepseudo (Φ, ρ)-invex at awith respect to X ifϕ(x)−ϕ(a)≥0 whenever Φ(x, a,(∇ϕ(a), ρ))≥0 for somex∈X.
Definition 3. ϕis said to bequasi(Φ, ρ)-invexat awith respect toX if Φ(x, a,(∇ϕ(a), ρ))≤0,whenever ϕ(x)−ϕ(a)≤0 for some x∈X.
Definition 4. ϕ is said to be semistrict quasi (Φ, ρ)-invex at a with respect to X if Φ(x, a,(∇ϕ(a), ρ)) < 0, whenever ϕ(x)−ϕ(a) < 0 for some x∈X.
ϕ is said to be (Φ, ρ)-invex (pseudo (Φ, ρ)-invex, quasi (Φ, ρ)-invex, semistrict quasi (Φ, ρ)-invex) on X0 if it is (Φ, ρ)-invex (pseudo (Φ, ρ)-invex, strictly pseudo (Φ, ρ)-invex, quasi (Φ, ρ)-invex, semistrict quasi (Φ, ρ)-invex) at afor everya∈X0.
Obviously, (Φ, ρ)-invexity implies pseudo,quasi and semistrict quasi (Φ, ρ)- invexity.
3. MINIMAX PROGRAMMING
The problem to be considered here is the minimax programming problem (M) : min max
1≤i≤p fi(x), g(x)≤0, x∈X0,
where X0 is a nonvoid open subset of Rn, fi, i ∈ {1,2, . . . , p}, and gj, j ∈ {1,2, . . . , m}, are differentiable real-valued functions defined on X0; f and g are the vector functions f = (f1, . . . , fp) : X0 → Rp, g = (g1, g2, . . . , gm) : X0 →Rm,and the symbol “≤” stands both for the usual inequality inRand for the component-wise inequality in a multi-dimensional Euclidean space
LetX be the set of all feasible solutions X ={x∈X0 |g(x)≤0}
of problem (M). The problem (M1) below is “equivalent” to (M):
(M1) : minν, g(x)≤0, f(x)≤νe, x∈X0. Hereeis the p-dimensional vector with all entries 1.
The equivalence means that if x is a feasible (optimal) solution of (M), then (x, ν),whereν = max
1≤i≤p fi(x),is a feasible (optimal) solution of (M1) and, conversely, if (x, ν) is a feasible (optimal) solution of (M1) then xis a feasible (optimal) solution of (M).
Note that (M1) is a smooth problem while (M) is not, and then we will appeal to (M1) as a substitute of (M) when differentiability is a necessary condition.
Another restatement of our minimax programming problem is (M2) : minF(x), x∈X0,
where ∆ is the standard (p −1)-dimensional simplex of Rp, ∆ = {µ = (µ1, . . . , µp) |µi≥0, i= 1, . . . , p,p
i=1µi = 1}, and F(x) = sup
(µ,λ)∈∆×Rm+ L(x, µ, λ), withL(x, µ, λ) =p
i=1µifi(x) +m
j=1λjgj(x).
It is obvious that if X =∅, then (M) and (M2) have the same optimal solutions. In fact, if X = ∅, then O(M) = O(M1) = O(M2). (Here, O(P) is the generic notation for the set of all optimal solutions of problem (P).)
In the next two sections we will show that some fundamental results con- cerning optimality and duality for the minimax programming can be restated in terms of (Φ, ρ)-invexity and related properties.
For the sake of concision we will use the shorthands “(Φ, ρ)-invexity (pseudo-, quasi-, etc.) at a” to mean “(Φ, ρ)-invexity (pseudo-, quasi-, etc.) at a, with respect to X”. When (Φ, ρ)- invexity is defined with respect to another set, this will be specified.
4. OPTIMALITY CONDITIONS
In this section we extends some necessary and sufficient conditions for optimality under various (Φ, ρ)-invexity conditions defined above.
Since (M1) is a smooth scalar optimization problem, we can follow the usual procedure involving the associated Lagrange function in order to obtain the Kuhn-Tucker conditions.
Definition5. ((a, ν), µ, λ)∈X×R×Rp+×Rm+ is said to be aKuhn-Tucker point of(M1) if
(2)
p i=1
µi∇fi(a) + m j=1
λj∇gj(a) = 0,
(3)
m j=1
λjgj(a) = 0,
(4) f(a)≤νe,
(5) µ, f(a)=ν,
(6) e, µ= 1.
Note that ifJ(a) stands for the set of indices of all active restrictions at a∈X, J(a) ={j |gj(a) = 0},then (2) and (3) become, respectively,
(7)
p i=1
µi∇fi(a) +
j∈J(a)
λj∇gj(a) = 0,
(8)
j∈J(a)
λjgj(a) = 0.
In what follows we will use both versions of these conditions.
SetM(a) =
i|fi(a) = max
1≤≤p f(a) .
Definition6. A triple (a, µ, λ) ∈X×Rp+×Rm+ is said to be aKuhn-Tucker point of(M) if it verifies (3) and the conditions
i∈M(a)
µi∇fi(a) +
j∈J(a)
λj∇gj(a) = 0,
(9)
i∈M(a)
µi= 1.
First, we will prove that the Kuhn-Tucker conditions are satisfied by optimal solutions of (M) (or (M1)) if these problems also satisfy some (Φ, ρ)- invexity conditions.
Theorem 1. Let (a, ν) be an optimal solution of (M1). Assume that Slater’s constraint qualification is satisfied for all restrictions indexed inJ(a).
Assume that for eachj∈J(a)the function gj is semistrict quasi (Φ, ρj)-invex ata for some ρj ≥0.Then there exist µ∈Rp+ and λ∈Rm+ such that (2)–(6) are verified by ((a, ν), µ, λ).
Proof. Since the problem (M1) is smooth, the Fritz-John necessary con- ditions claim the existence of non-negative multipliersw∈R, u∈Rp,v∈Rm such that
(10) w− e, u= 0, p
i=1
ui∇fi(a) + m j=1
vj∇gj(a) = 0, f(a)≤νe, u, f(a)−νe= 0, v, g(a)= 0,
(11) w+
p i=1
ui+ m j=1
vj >0.
We have to prove thatw+p
i=1ui >0.
Otherwise, if w= 0 and u= 0,it follows from (11) thatv0=m
j=1vj >
0.Settingvj =vj/v0, j = 1,2, . . . , m,equation (7) becomes
j∈J(a)vj∇gj(a) = 0.It then follows from the properties of Φ that
0≤Φ
x, a,
j∈J(a)
vj∇gj(a),
j∈J(a)
vjρj
≤
j∈J(a)
vjΦ(x, a,(∇gj(a), ρj)) for everyx∈X.
Let x∗ ∈ X satisfy Slater’s conditions gj(x∗) < 0, ∀j ∈ J(a). Then, because eachgj is semistrict quasi (Φ, ρj)-invex, we have
j∈J(a)
vjΦ(x∗, a,(∇gj(a), ρj))<0, so that we have reached a contradiction.
Now, observe that the above inequality together (10) say thate, µ= 1, where µi = ui/w. Finally, set λj = w1vj, j = 1, . . . , m, and the proof is complete.
Theorem2. Let a be an optimal solution of (M). Assume that Slater’s constraint qualification is satisfied for all restrictions indexed inJ(a).Assume that for eachj∈J(a) the functiongj is semistrict quasi(Φ, ρj)-invex atafor someρj ≥0.Then there exist µ∈Rp+ and λ∈Rm+ such that (2), (3)and (9) are verified by (a, µ, λ).
Proof. Since a∈O(M) we have (a, ν) ∈O(M1),where ν = max
1≤i≤p fi(a).
Then, by Theorem 1, there exist µ∈ Rp+ and λ∈ Rm+ such that ((a, ν), µ, λ) is a Kuhn-Tucker point of (M1). The definition of ν makes (4) trivial, and it follows from (5) thatµi= 0 if i /∈M(a).Hence (6) reduces to (9).
The next two results concern the sufficiency of Kuhn-Tucker conditions when (Φ, ρ)-invexity is assumed
Theorem3. Let ((a, ν), µ, λ) be a Kuhn-Tucker point of (M1).Assume for each j ∈ J(a), that p
i=1µifi is pseudo (Φ, ρ0)-invex at a, gj is quasi (Φ, ρj)-invex at a, and ρ0+
j∈J(a)λjρj ≥ 0. Then a ∈ O(M) and (a, ν) ∈ O(M1).
Proof. By (4), (a, ν) is a feasible solution of (M1). Let (x, ν) be an arbitrary feasible solution of (M1). Two situations should be considered:
(i)λ= 0; (ii) λ= 0.
In case (i), we have
(12) Φ
x, a,
p
i=1
µi∇fi(a), ρ0
≥0.
If case (ii) holds, letw= 1 +
j∈J(a)λj,so that we have 1/wΦ
x, a,
p
i=1
µi∇fi(a), ρ0
+
j∈J(a)
λj/wΦ(x, a,(∇gj(a), ρj))≥
≥Φ
x, a,
1/w p i=1
µi∇fi(a) +
j∈J(a)
λj/w∇gj(a), ρ0/w+
j∈J(a)
λjρj/w
≥0.
Since the gj are quasi (Φ, ρj)-invex at a and gj(x)−gj(a) =gj(x) ≤ 0 for eachj ∈ J(a),we have
j∈J(a)λj/wΦ(x, a,(∇gj(a), ρj))≤0, so that we arrive again to (12).
Now, by (12), (Φ, ρ0)-invexity implies the inequality p
i=1
µifi(x)− p i=1
µifi(a)≥0.
But fi(x) ≤ν, p
i=1µifi(a) = ν and p
i=1µi = 1. Therefore, the inequality above implies the inequalityν−ν≥0.
Theorem4. Let (a, µ, λ) be a Kuhn-Tucker point of (M). Assume that fi is semistrict quasi (Φ, ρ0i)-invex at afor each i∈M(a) gj is quasi (Φ, ρj)- invex at a for each j ∈ J(a), and
i∈M(a)µiρ0i+
j∈J(a)λjρj ≥ 0. Then, a∈ O(M).
Proof. By Definition 5,ais a feasible solution of (M). Putν = max
1≤i≤p fi(a).
Thus,fi(a) =νfor everyi∈M(a).Suppose thatais not optimal. Then there existsx ∈X such that ν = max
1≤i≤p fi(x) < ν. Since (a, µ, λ) is a Kuhn-Tucker point, we have
i∈M(a)
ui∇fi(a) +
j∈J(a)
vj∇gj(a) = 0,
whereui =µi/µ0, i∈M(a),vj =λj/µ0,j∈J(a), µ0 = 1 +
j∈J(a)λj.Also,
i∈M(a)uiρ0i+
j∈J(a)vjρj ≥0, and then Φ
x, a,
i∈M(a)
ui∇fi(a) +
j∈J(a)
vj∇gj(a), p i=1
uiρ0i+
j∈J(a)
vjρj
≥0.
Moreover,
i∈M(a)
uiΦ(x, a,(∇fi(a), ρ0i)) +
j∈J(a)
vjΦ(x, a,(∇gj(a), ρj))≥0 hence
i∈M(a)
uiΦ(x, a,(∇fi(a), ρ0i))≥0.
Fori∈M(a) we havefi(x)−fi(a)≤ν−ν <0,so that Φ(x, a,(∇fi(a), ρ0i))<
0.Sinceu∈Rp+andp
i=1ui >0,we have
i∈M(a)uiΦ(x, a,(∇fi(a), ρ0i))<0, contradicting the previous inequality.
5. DUALITY
In strict formal terms, we will consider here two couples of dual problems defining the Wolfe dual of (M1) and the generalized dual of (M2). Due to the equivalences pointed out in Section 3, both problems are duals of (M).
As we have already observed, problem (M1) is smooth, so that it admits a Wolfe dual
(WM) :
max p
i=1
µifi(y) + m j=1
λjgj(y)
, p
i=1
µi∇fi(y) + m j=1
λj∇gj(y) = 0, y∈X0, µ∈∆, λ∈Rm+. The equivalence of (M) and (M1) allow us to refer to (WM) as to the dual of (M).In fact, the properties listed bellow actually relate by duality the two problems.
Theorem5. Letxbe a feasible solution of(M)and(y, µ, λ)a feasible so- lution of(WM).Assume thatfi is(Φ, ρ0i)-invex at yfor eachi∈ {1,2, . . . , p}, gj is(Φ, ρj)-invex atyfor eachj∈ {1,2, . . . , m},andp
i=1µiρ0i+m
j=1λjρj ≥ 0.Then p
i=1µifi(y) +m
j=1λjgj(y)≤ max
1≤i≤p fi(x).
Proof. Settingw= 1 +m
j=1λj, ui =µi/w, and uj =λj/w, we have Φ
x, y,
p
i=1
ui∇fi(y) + m j=1
vj∇gj(y), p i=1
uiρ0i+ m j=1
vjρj
≥0,
hence p i=1
uiΦ(x, y,(∇fi(y), ρ0i)) + m j=1
vjΦ(x, y,(∇gj(y), ρj))≥0.
Furthermore, the invexity offi and gj implies p
i=1
ui(fi(x)−fi(y)) + m j=1
vj(gj(x)−gj(y))≥0.
Multiplying by w,the last inequality yields p
i=1
µifi(y) + m j=1
λjgj(y)≤ p i=1
µifi(x) + m j=1
λjgj(x).
Since x∈X, it follows that p
i=1
µifi(x) + m j=1
λjgj(x)≤ p i=1
µifi(x)≤ max
1≤i≤pfi(x).
Remark. Under invexity assumptions, if (x, ν) is a feasible solution of (M1), then p
i=1µifi(y) +m
j=1λjgj(y)≤ν.
Theorem6. Let a be an optimal solution of (M). Assume that Slater’s constraints qualification is satisfied for all restrictions indexed in J(a). If, fi is (Φ, ρ0i)-invex on X0, for each for each i ∈ {1,2, . . . , p}, gj is (Φ, ρj)- invex on X0, for each j ∈ {1,2, . . . , m}, where ρ0i, ρj ≥ 0, then there exist µ and λsuch that (a, µ, λ) is an optimal solution of (WM) and p
i=1µifi(a) + m
j=1λjgj(a) = max
1≤i≤p fi(a)
Proof. By Theorem 2 there existµ∈Rp+,andλ∈Rm+ such that (a, µ, λ) is a Kuhn-Tucker point of (M). Then, (a, µ, λ) is a feasible solution of (WM) and sincem
j=1λjgj(a) = 0,we have p
i=1
µifi(a) + m j=1
λjgj(a) =
i∈M(a)
µifi(a) = max
1≤i≤p fi(a) Then by the optimality of (a, µ, λ) follows by Theorem 5.
Let us now generalize duality with respect to (M2).
The generalized dual of (M) is the scalar optimization problem (GM) : maxG(µ, λ), (µ, λ)∈∆×Rm+,
whereG(µ, λ) = inf
x∈X0
L(x, µ, λ).
The weak duality property follows, with no restriction, from the universal minimax inequality
sup
(µ,λ)∈∆×Rm+ inf
x∈X0
L(x, µ, λ)≤ inf
x∈X0
sup
(µ,λ)∈∆×Rm+ L(x, µ, λ).
Thus,
(13) sup
(µ,λ)∈∆×Rm+ G(µ, λ)≤ inf
x∈X0 F(x).
For a direct duality theorem, (Φ, ρ)-invexity is sufficient.
Theorem 7. Let a ∈ X be an optimum solution of (M2). Assume that Slater’s constraints qualification is satisfied for all restrictions indexed inJ(a).
If fi is (Φ, ρ0i)-invex at a with respect to X0 for each i ∈ M(a) and gj is (Φ, ρj)-invex at a with respect to X0, for each j ∈ J(a) where ρ0i, ρj ≥ 0, then there exist µ and λ such that (µ, λ) is an optimal solution of (GM) and G(µ, λ) =F(a).
Proof. SinceO(M) =O(M2), ais an optimal solution of (M) andF(a) =
1≤i≤pmax fi(a).Then, by Theorem 2, these exist µi ∈Rp+ and λ∈Rm+ such that
i∈M(a)
µi∇fi(a) +
j∈J(a)
λj∇gj(a) = 0,
j∈J(a)
λjgj(a) = 0, and
i∈M(a)µi= 1.
Forx∈X0 we have
i∈M(a)
uiΦ(x, a,(∇fi(a), ρ0i)) +
j∈J(a)
vjΦ(x, a,(∇gj(a), ρj))≥0, where,ui=µi/w, vj =λj/w, w= 1 +m
j=1λj. Now, use the invexity assumptions to obtain
i∈M(a)
ui(fi(x)−fi(a)) +
j∈J(a)
vj(gj(x)−gj(a))≥0.
Let µ∈Rp+ withµi = µi ifi∈M(a) andµi = 0 otherwise. Obviously, µ∈∆ and the last inequality is equivalent to
p i=1
µifi(x) + m j=1
λjgj(x)≥ p
i=1
µifi(a).
Since p
i=1µifi(a) =
i∈M(a)µifi(a) = max
1≤i≤p fi(a) = F(a), and the above inequality holds for every x∈X0,we have
x∈Xinf0
p i=1
µifi(x) + m j=1
λjgj(x)
≥F(a),
i.e. G(µ, λ)≥F(a).Then (13) implies the optimality of (µ, λ) and the equation G(µ, λ) =F(a).
Remark. The last equation is just the minimax equation
x∈Xmin0 sup
(µ,λ)∈∆×Rm+ L(x, µ, λ) = max
(µ,λ)∈∆×Rm+ inf
x∈X0 L(x, µ, λ), so that the above theorem is a minimax theorem forL.
Acknowledgements. This research was partly supported by CNCSIS Grant 27694/2005.
REFERENCES
[1] G. Caristi, M. Ferrara and A. Stefanescu,Mathematical programming with(Φ, ρ)-invexity.
In: Proc. 8th Conf. on Generalized Convexity and Monotonicity, Varese, June 2005.
[2] B.D. Craven,Invex functions and constrained local minima.Bull. Austral. Math. Soc.24 (1981), 357–366.
[3] Y. Jeyakumar,Strong and weak invexity in mathematical programming.Methods Oper.
Res.55 (1985), 109–125.
[4] M.A. Hanson,On sufficiency of Kuhn Tucker conditions.J. Math. Anal. Appl.30(1981), 545–550.
[5] M.A. Hanson and B. Mond,Further generalization of convexity in mathematical program- ming.J. Inform. Optim. Sci. 3(1982), 22–35.
[6] V. Preda,On efficiency and duality for multiobjective programs. J. Math. Anal. Appl.
166(1992), 365–377.
[7] J.P. Vial,Strong and weak convexity of sets and functions. Math. Oper. Res. 8(1983), 231–259.
[8] T. Weir and B. Mond,Generalized convexity and duality in multiple objective program- ming.Bull. Austral. Math. Soc. 39(1989), 287–299.
Received May 2, 2006 Academy of Economic Studies Department of Mathematics
Calea Dorobantilor 15–17, Bucharest, Romania and
University of Bucharest
Faculty of Mathematics and Computer Science Str. Academiei 14
010014 Bucharest, Romania