Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

www.elsevier.com/locate/cam

A steepest descent method for vector optimization

L.M. Graña Drummonda,1 , B.F. Svaiterb,∗,2
a FACC–UFRJ, Av. Pasteur 250, CEP 22290–240, Rio de Janeiro, RJ, Brazil
b Instituto de Matemática Pura e Aplicada (IMPA), Estrada Dona Castorina 110, CEP 22460-320, Rio de Janeiro, RJ, Brazil

Received 24 March 2003; received in revised form 3 June 2004

Abstract
In this work we propose a Cauchy-like method for solving smooth unconstrained vector optimization problems.
When the partial order under consideration is the one induced by the nonnegative orthant, we regain the steepest
descent method for multicriteria optimization recently proposed by Fliege and Svaiter. We prove that every accu-
mulation point of the generated sequence satisfies a certain first-order necessary condition for optimality, which
extends to the vector case the well known “gradient equal zero” condition for real-valued minimization. Finally,
under some reasonable additional hypotheses, we prove (global) convergence to a weak unconstrained minimizer.
As a by-product, we show that the problem of finding a weak constrained minimizer can be viewed as a particular
case of the so-called Abstract Equilibrium problem.
© 2004 Elsevier B.V. All rights reserved.
Keywords: Pareto optimality; Vector optimization; Steepest descent; K-convexity; Quasi-Féjer convergence

1. Introduction

In multicriteria optimization, several objective functions have to be minimized simultaneously. Usually,

no single point will minimize all given objective functions at once (i.e., there does not exist an ideal
minimizer), and so the concept of optimality has to be replaced by the concept of Pareto-optimality
or efficiency. A point is called Pareto-optimal or efficient, if there does not exist a different point with

∗ Corresponding author. Tel.: +55-21-2529-5112; fax: +55-21-2529-5129.

E-mail addresses: [email protected] (L.M. Graña Drummond), [email protected] (B.F. Svaiter).
1 Partially supported by FAPERJ under Grant E-26/152.107/1999-Bolsa.
2 Partially supported by CNPq Grant 302748/2002-4, PRONEX-Optimization and FAPERJ.

0377-0427/$ - see front matter © 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2004.06.018
396 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

smaller than or equal objective function values, such that there is a decrease in at least one objective
function value. Applications for this type of problem can be found in engineering design [12] (mainly
truss optimization [8]), location science [5], statistics [6], management science [13] (specially portfolio
analysis [24]), etc.
Among the main solution strategies for multicriteria optimization problems, we mention the scalar-
ization approaches [15,16,19,21,23]. Here, one or several parameterized single-objective (i.e., classical)
optimization problems are solved. Frequently, some parameters have to be specified in advance, leav-
ing the modeler and the decision-maker with the burden of choosing them. Moreover, in the weighting
method, for example, bad choices of these parameters can lead to unbounded scalar problems. Other
scalarization techniques are parameter-free [7,8,21] but try to compute a discrete approximation to the
whole set of Pareto-optimal points.
Parameter-free multicriteria optimization techniques use in general an ordering of the different criteria,
i.e., an ordering of importance of the components of the objective function vector. In this case, the ordering
has to be specified. Moreover, the optimization process is usually augmented by an interactive procedure
[20], adding an additional burden to the task of the decision-maker.
In a recent paper, Fliege and Svaiter [14] proposed a Pareto descent method for multiobjective optimiza-
tion. This procedure is parameter-free and relies upon a suitable extension for vector-valued functions
of the classical steepest descent direction. Neither ordering information nor weighting factors for the
different objective functions is assumed to be known in this new method, which may be interpreted as a
“Cauchy’s method” for multicriteria optimization.
We recall that the steepest descent method (also known as gradient or Cauchy’s method) is one of
the oldest and more basic minimization schemes for scalar unconstrained optimization. Despite its com-
putational shortcomings, like, for instance, “hemstitching” phenomena, the Cauchy’s method can be
considered among the most important procedures for minimization of real-valued functions defined on
Rn , since it is the departure point for many other more sophisticated and efficient algorithms. For instance,
it is partially used in some “globally convergent” modifications of Newton’s method for unconstrained
optimization. Here, “globally convergent” means that all sequences produced by the method have de-
creasing objective function values, and that all accumulation points of these sequences are critical points.
We refer the reader to [9], where the “double dog-leg” method is discussed. The simple idea of decreasing
the value of the objective function is also used in many other modifications of Newton’s method. We
refer the reader again to [9] for a very clear exposition. It remains an open question how to extend more
efficient procedures, as Newton’s method, to vector optimization.
The purpose of this paper is to take a step further on the direction of Fliege and Svaiter’s [14] work.
Based on their ideas, we present a Cauchy-like method for smooth vector optimization. In this setting,
the partial order is induced by a general closed convex pointed cone K, with nonempty interior in a
finite-dimensional space. Our procedure depends on the choice of an arbitrary initial point, as well as on
a certain compact set which characterizes the positive polar cone (see Section 3). On the final remarks
we make more comments on this issue. When the cone is the nonnegative orthant, our procedure turns
out to be the very same proposed by them for the unconstrained case. Our convergence results extend
theirs and, furthermore, under some additional (and quite reasonable) hypotheses, we also show that all
sequences produced by the method converge to a weakly efficient point, no matter how poor is the initial
guess. We point out that we are not attempting to find the set of all efficient or weak efficient optima.
Regarding the importance of vector optimization, we point out that even though the vast majority of
real life problems formulated as vector-valued problems deals with the component-wise partial order, i.e.,
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 397

the one which arises from the Paretian cone, there are many others that require preference orders induced
by closed convex cones other than the nonnegative orthant. Such cones have been recently analyzed, for
example in [1] (based on the theoretical results of [2]), where problems of portfolio selection in security
markets require finding weak Pareto minimal points with respect to feasible portfolio cones, which are
nonlattice, that is to say cones with more extreme rays than the ambience space dimension.
The outline of this work is as follows. In Section 2, we introduce the unconstrained K-minimization
problem and discuss a necessary (but in general nonsufficient) first-order optimality condition, called
K-criticality. This condition extends to vector optimization the classical “gradient equal zero” condition
for scalar minimization. The notion of K-criticality allows us to derive a generic K-descent scheme. We
discuss an Armijo-like strategy for choosing the stepsizes. In Section 3, we characterize the negative cone
and its interior in terms of a convex function. This function is used to define the steepest descent direction,
as well as its approximations. We present the complete steepest descent method for vector optimization,
with approximations of the steepest descent direction, and show its well definedness. In Section 4, we
discuss the convergence of the method. We prove that all the cluster points of any sequence generated
by the algorithm are K-critical. In Section 5, we discuss the relationship between vector steepest descent
direction and the scalarization approach. In Section 6, assuming K-convexity of the objective function
and a certain very reasonable condition on the objective, we prove, based upon the notion of quasi-
Fejér convergence, that, with the pure steepest descent direction choice or even with other more general
directions, called s-compatibles, the method is globally convergent to a weak unconstrained K-minimizer.
In Section 7, using the scalar function defined in Section 3, we show that the problem of finding a weak
constrained K-minimizer of a vector-valued function can be viewed as a particular case of the well known
abstract equilibrium problem. Finally, in Section 8 we make some final remarks about our work.

2. Basic definitions

Let K be a closed pointed convex cone of Rm , with nonempty interior. The partial order in Rm induced
by K,
K , is defined by:

u
K v if v − u ∈ K.

Consider also the following relation induced by int(K) in Rm , ≺K :

u≺K v if v − u ∈ int(K).

Given a continuously differentiable function F : Rn → Rm , we consider the problem of finding an

unconstrained K-minimizer (or K-optimum) of F, i.e., a point x ∗ ∈ Rn such that there exists no other
x ∈ Rn with F (x)
K F (x ∗ ) and F (x) = F (x ∗ ). In other words, we are seeking unconstrained minimizers
for F in the partial order induced by the cone K. We denote this problem as

min K F (x). (1)

A necessary, but, in general, nonsufficient, condition for K-optimality of a point x ∈ Rn is

−int(K) ∩ Image(J F (x)) = ∅, (2)

398 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

where J F (x) stands for the Jacobian of F at x and Image(J F (x)) stands for the image of Rn by the linear
operator J F (x) (see, for instance, [19]). Observe that (2) generalizes to vector optimization the classical
condition “gradient equal zero” for the real-valued case.
A point x is K-critical if it satisfies (2). Therefore, if a point x is not K-critical, there exists a direction
v ∈ Rn satisfying

J F (x)v ∈ −int(K) , (3)

that is to say J F (x)v≺K 0. It is a well known fact that such v is a K-descent direction for the objective
F. Actually, it holds that (see [19]), if v satisfies (3), there exists t¯ > 0 such that,

F (x + tv)≺K F (x) for all t ∈ (0, t¯). (4)

If v is a K-descent direction at x, we say that t > 0 satisfies the “Armijo-like” rule for
∈ (0, 1) if

F (x + tv)
K F (x) +
tJ F (x)v.

As in the scalar case, given a “descent” direction, the Armijo rule is satisfied for some t’s.

Proposition 2.1. Let
∈ (0, 1). If J F (x)v≺K 0, then there exists t¯ > 0 such that,

F (x + tv)≺K F (x) +
tJ F (x)v

for all t ∈ (0, t¯).

Proof. Since F is differentiable, we have

F (x + tv) = F (x) + t (J F (x)v + R(t)), (5)

with limt→0 R(t) = 0. We are assuming that J F (x)v ∈ −int(K). Since
∈ (0, 1), (1 −
)J F (x)v ∈
−int(K). Hence, there exists t¯ > 0 such that, for all t ∈ (0, t¯], R(t) is small enough, so that

R(t) + (1 −
)J F (x)v ∈ −int(K).

Equivalently,

R(t)≺K − (1 −
)J F (x)v for all t ∈ (0, t¯].

Combining this K-inequality with (5) the conclusion follows. 

Proposition 2.1 opens the way for defining a generic K-descent method using the Armijo-like rule.

Algorithm 1 (Generic K-descent method).

1. Take
∈ (0, 1), x 0 ∈ Rn . Set k := 0.

2. If x k is K-critical STOP. Otherwise,
3. Find v k a K-descent direction at x k .
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 399

4. Find tk > 0 such that

F (x k + tk v k )
K F (x k ) +
tk J F (x k )v k .
5. Set x k+1 := x k + tk v k , k := k + 1 and GOTO 2.

Only very general properties can be proved for this algorithm.

Proposition 2.2. Let {x k } be an infinite sequence generated by Algorithm 1. If x̄ is an accumulation point

of {x k } then
F (x̄)
K F (x k )
for all k and limk→∞ F (x k ) = F (x̄). In particular, F is constant in the set of accumulation points of {x k }.

Proof. We are supposing that an infinite sequence {x k } was generated by Algorithm 1. Therefore, all x k
are non K-critical and {F (x k )} is K-decreasing. By assumption, there is a subsequence {x kj } converging
to x̄. Take any k ∈ N. For j large enough kj > k and
F (x kj )
K F (x k ).
Taking the limit for j → ∞ we get F (x̄)
K F (x k ). Let x̂ be another accumulation point of {x k }. Then
there exists a subsequence {x kp } converging to x̂. Since
F (x̄)
K F (x kp ),
letting p → ∞ we get F (x̄)
K F (x̂). By the same reasoning, F (x̂)
K F (x̄). Since K is pointed, these
two K-inequalities imply that F (x̂) = F (x̄). 

Regarding the choice of the stepsizes tk , if they are taken too small, the generated sequence {x k } may
converge to a non K-critical point. In order to choose a suitable steplength at iteration k, we prescribe the
usual backtracking procedure:

3a. Set t := 1
3b. If F (x k + tv k )
K F (x k ) +
tJ F (x k )v k then
tk = t, END (of backtracking)
else
3c. Set t := t/2, GOTO 3b.

Observe that the above backtracking procedure has always finite termination, thanks to Proposition 2.1.
Moreover
tk = max{2−j | j ∈ N, F (x k + 2−j v k )
K F (x k ) +
2−j J F (x k )v k }.
The main problem now is the choice of v k . For classical optimization, we have m = 1, K = R+ . In
this case, the natural choice is the steepest descent direction v k = −∇F (x k ), which happens to be the
solution of
min v, ∇F (x k ) + (1/2)v2 , v ∈ Rn . (6)
400 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

For multiobjective optimization (m  1), where K is the positive orthant Rm

+ , Fliege and Svaiter [14]
proposed to take v k as the solution of

min max v, ∇Fi (x k ) + (1/2)v2 , v ∈ Rn , (7)

i=1,...,m

where F (x) = (F1 (x), . . . , Fm (x)). Observe that (6) is a particular case of (7) when m = 1.
In the following section, we will extend the notion of steepest descent direction for the partial order

K and propose the K-steepest descent method.

3. K-steepest descent method

The positive polar cone of A ⊆ Rm is the set

A∗ = {w ∈ Rm | y, w  0 ∀y ∈ A}.

Since K is a closed convex cone, K = K ∗∗ (see [22, Theorem 14.1]) and

−K = {y ∈ Rm | y, w  0 ∀w ∈ K ∗ }

and

−int(K) = {y ∈ Rm | y, w < 0 ∀w ∈ K ∗ \{0}}.

The convex hull of A ⊆ Rm will be denoted by conv(A), and the cone generated by A will be denoted by
cone(A).
From now on, we assume that we have a compact set C ⊆ Rm such that:
0∈/ C, (8)
cone(convC) = K ∗ . (9)

As int(K) = ∅ and C ⊆ K ∗ \{0}, it follows that 0 ∈

/ conv(C). Therefore
−K={u ∈ Rm | u, w  0 ∀w ∈ C}, (10)
−int(K)={u ∈ Rm | u, w < 0 ∀w ∈ C}. (11)

In classical optimization, K = R+ and we may take C = {1}. For multiobjective optimization, K and
K ∗ are the positive orthant of Rm and we may take C as the canonical basis of Rm . If K is a polyhedral
cone, C may be taken as a finite set of extremal rays of K ∗ . For a generic K (closed pointed convex cone
with nonempty interior), the set

C = {w ∈ K ∗ |w1 = 1}

(where w1 = |w1 | + · · · + |wm |) will satisfy conditions (8), (9).

Define now  : Rm → R,

(y) := sup y, w. (12)

w∈C
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 401

In view of (10)–(11) and the compactness of C, the function  gives a “scalar” characterization of −K
and −int(K) :
−K = {y ∈ Rm | (y)  0} (13)
and
−int(K) = {y ∈ Rm | (y) < 0}. (14)
In the following lemma we establish some elementary properties of the function  which will be used in
the sequel.

Lemma 3.1. (i) Let y, y  ∈ Rm , then (y + y  ) (y) + (y  ) and (y) − (y  ) (y − y  ). (ii) Let
y, y  ∈ Rm , if y≺K y  (y
K y  ), then (y) < (y  ) ((y) (y  )). (iii) The function  : Rn → R is
Lipschitz continuous.

Proof. Item (i). The first inequality holds trivially and the second follows from the first. Item (ii). The
fact that y − y  ∈ −int(K) (y − y  ∈ −K) is equivalent to (y − y  ) < 0 ((y − y  )  0), according to
(14) and (13). Hence, the result follows from (i).
Item (iii). By virtue of what was established in (i), (y)− (y  ) (y −y  ) and (y  )− (y) (y  −y).
Hence
|(y) − (y  )|  sup{(y − y  ), (y  − y)}.
Therefore, from (12) and Cauchy–Schwartz inequality,
|(y) − (y  )|  Ly − y  , (15)
where, L := sup{w | w ∈ C}, and the result follows. 

Define now for x ∈ Rn , fx : Rn → R as

fx (v) :=(J F (x)v)
= sup w, J F (x)v. (16)
w∈C

From (14) it follows that v is a K-descent direction at x, if and only if, fx (v) < 0. Therefore, x is K critical
if and only if fx (v)  0 for all v ∈ Rn .
We can now extend the notion of steepest descent direction to the vector case (with K an arbitrary cone
satisfying the conditions stated at the beginning of Section 2).

Definition 3.2. Given x ∈ Rn , the K-steepest descent direction (for F) at x, denoted by vx is the solution
of
min fx (v) + (1/2)v2 , v ∈ Rn . (17)
The optimal value of this problem will be denoted by x .

Remark 1. In the scalar minimization case, where F : Rn → R and K = R+ , taking C = {1}, the
K-steepest descent direction is exactly the classical steepest descent direction, i.e., vx = −∇F (x).
402 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

Remark 2. For multicriteria optimization, where K = Rm m

+ , with C given by the canonical basis of R ,
we retrieve the steepest descent direction proposed in [14].

Since v  → fx (v) is a real-valued closed convex function, vx and x are well defined. Furthermore,
as F is continuously differentiable and  is Lipschitz continuous, the mapping (x, v)  → fx (v) is also
continuous.

Lemma 3.3.

1. If x is K-critical then vx = 0, x = 0.
2. If x is not K-critical then vx = 0, x < 0,
fx (vx ) < − (1/2)vx 2 < 0
and vx is a K-descent direction.
3. The mappings x → vx , x → x are continuous.

Proof. Item 1. If x is K-critical, then fx (v)  0 ∀v ∈ Rn . Since fx (0) = 0, the conclusion follows.
Item 2. If x is not K-critical, then, for some v ∈ Rn , fx (v) < 0. Observe that fx (·) is positive homoge-
neous of degree 1. Taking
t˜ = −fx (v)/v2 , ṽ = t˜v,
we get
fx (ṽ) + (1/2)ṽ2 = t˜fx (v) + (1/2)t˜2 v2
= − (1/2)fx (v)2 /v2 < 0.
Hence x < 0. The other statements of item 2 now follow trivially.
Item 3. Take x 0 ∈ Rn and  > 0. Define
S := {v ∈ Rn | vx 0 − v = }.
Note that vx 0 is optimal for (17) with x = x 0 . From (16) it follows that fx (·) is convex, so the objective
function on the minimization problem (17) is strongly convex with modulus 1/2. It follows that,
fx 0 (v) + (1/2)v2  fx 0 (vx 0 ) + (1/2)vx 0 2 + (1/2)2 ∀v ∈ S.
Since the mapping (x, v) → fx (v) is continuous, and S is compact, using this equation we conclude that
there exists  > 0 such that, if x − x 0  , then
fx (v) + (1/2)v2 > fx (vx 0 ) + (1/2)vx 0 2 ∀v ∈ S.
Take now x ∈ Rn , x − x 0  . As v → fx (v) + (1/2)v2 is convex, we conclude from the above
inequality that vx , the minimizer of fx (·)+(1/2)·2 is not in the region v−vx 0  , hence vx −vx 0  .
Continuity of x follows now trivially. 

A possible choice for v k in Algorithm 1 is vx k , i.e., the K-steepest descent direction at x k . Since
the computation of vx requires the solution of (17), it would be interesting to work with approximated
solutions of this problem.
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 403

Definition 3.4. Let  ∈ [0, 1). We say that v is a -approximate K-steepest descent direction at x ∈ Rn if
fx (v) + (1/2)v2  (1 − )x ,
or equivalently


fx (v) + (1/2)v2 − fx (vx ) + (1/2)vx 2 |x |.

Observe that the (exact) K-steepest descent direction at x is always a -approximate K-steepest descent
direction, because we assume  ∈ [0, 1). The exact K-steepest descent direction at x is the unique  = 0-
approximate K-steepest descent direction.

Lemma 3.5. Let  ∈ [0, 1). If v is a -approximate K-steepest descent direction at x, then
vx − v2  2|x |.

Proof. The function

v → fx (v) + (1/2)v2
is strongly convex with modulus 1/2. Since vx is the minimizer of this function,


fx (v) + (1/2)v2 − fx (vx ) + (1/2)vx 2  (1/2)vx − v2 .
Using Definition 3.4, the conclusion follows. 

Let  ∈ [0, 1) be a prespecified tolerance. From Lemmas 3.3 and 3.5, it follows that v = 0 is a -
approximate K-steepest descent direction at x if, and only if, x is K-critical. Note also that, if x is not
K-critical and v is a -approximate K-steepest descent direction at x, then v is a K-descent direction, and
in particular v = 0.
Now we formally state the K-steepest descent method (with K-Armijo rule, implemented with back-
tracking). This algorithm is a particular case of Algorithm 1.

Algorithm 2 (K-steepest descent method).

1. Choose
∈ (0, 1),  ∈ [0, 1), x 0 ∈ Rn . Set k := 0.

2. If x k is K-critical (i.e., if fx k (v)  0 for all v ∈ Rn ) STOP. Otherwise,
3. Compute v k , a -approximate K-steepest descent direction at x k .
4. Compute the steplength tk ∈ (0, 1] in the following way:
tk := max{2−j | j ∈ N, (F (x k + 2−j v k ) − F (x k ) −
2−j J F (x k )v k )  0}.
5. Set x k+1 := x k + tk v k , k := k + 1 and GOTO 2.

Observe that if x k is not K-critical, then v k obtained in step 3 is a K-descent direction and tk in step 4 is
well defined. Moreover, such tk may be obtained by a backtracking procedure, as discussed previously.
Note that, by virtue of (13), in step 4 we have,
tk := max{2−j | j ∈ N, F (x k + 2−j v k )
K F (x k ) +
2−j J F (x k )v k }. (18)
404 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

Furthermore, x k+1 will satisfy F (x k+1 )
K F (x k ). So, the objective values sequence {F (x k )} is K-
nonincreasing.
We finish this section with a generalization of Proposition 2.1, which is a simple consequence of the
fact that F is continuously differentiable.

Proposition 3.6. Let
∈ (0, 1), x and v such that J F (x)v≺K 0. Then there exist tˆ, ,  > 0 such that,
v  is a K-descent direction at x  ,
F (x  + tv  )≺K F (x  ) +
tJ F (x  )v 

for any t ∈ (0, tˆ), x  ∈ B(x; ), v  ∈ B(v;  ).

Proof. By assumption, J F (x)v ∈ −int(K). So, there exists  > 0 such that

J F (x)v + y ∈ −int(K) ∀y ∈ Rm , y . (19)

Since JF is continuous, there exist 1 , 2 > 0 such that if x  − x 1 , v  − v 2 , then

J F (x  )v  − J F (x)v /2 (20)

and so

J F (x  )v  ≺K 0.

Continuity of JF also implies that

F (z + tu) = F (z) + tJ F (z)u + tR(z, tu),

with limt→0 R(z, tu) = 0 uniformly for z and u in compact sets. Therefore, there exists tˆ > 0 such that,
for t ∈ (0, tˆ), x  − x 1 , v  − v 2 ,

R(x  , tv  )  (/2)(1 −
). (21)

Now, assume that t ∈ (0, tˆ), x  − x 1 , v  − v 2 . Then

F (x  + tv  )=F (x  ) + tJ F (x  )v  + tR(x  , tv  )
= F (x  ) + t
J F (x  )v  + t[(1 −
)J F (x  )v  + R(x  , tv  )].
Defining

u := J F (x  )v  + (1 −
)−1 R(x  , tv  ), (22)

we have

F (x  + tv  ) = F (x  ) + t
J F (x  )v  + t (1 −
)u. (23)

It suffices to see that u≺K 0. Observe that

u = J F (x)v + ỹ, (24)

 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 405

where ỹ := J F (x  )v  − J F (x)v + (1 −
)−1 R(x  , tv  ). Using (20) and (21),

ỹ  J F (x  )v  − J F (x)v + (1 −
)−1 R(x  , tv  )
.
Therefore, using (19) and (24) we conclude that u≺k 0 and so, the result follows from (23). 

4. Convergence analysis: the general case

From now on, {x k }, {v k }, {tk } are sequences generated by Algorithm 2. If the algorithm terminates after
a finite number of iterations, it terminates at a K-critical point. In this section, we suppose that an infinite
sequence {x k } is generated. So, in view of Lemma 3.3, Definition 3.4 and Proposition 3.6, for all k,
x k < 0,
fx k (v k ) + (1/2)v k 2  (1 − )x k < 0,
F (x k+1 )
K F (x k ) +
tk J F (x k )v k
K F (x k ).

In particular the sequence {F (x k )} is K-decreasing.

Using the above K-inequality, Lemma 3.1 and the positive homogeneity of , it follows that for all k,
(F (x k+1 )) (F (x k ) +
tk J F (x k )v k )
(F (x k )) + (
tk J F (x k )v k )
= (F (x k )) +
tk (J F (x k )v k )
= (F (x k )) +
tk fx k (v k )
 
(F (x k )) +
tk (1 − )x k − (1/2)v k 2 . (25)

As a consequence of this scalar inequality we obtain the following lemma.

Lemma 4.1. If {F (x k )} is K-bounded from below, (i.e., if there exists ȳ such that ȳ
K F (x k ) for all k)
then,
 
tk |x k | < ∞, tk v k 2 < ∞.

Proof. Adding inequality (25) from k = 0 to n we get,

n

(F (x n+1 )) (F (x 0 )) +
tk ((1 − )x k − (1/2)v k 2 )
k=0
n
  
= (F (x 0 )) −
tk (1 − )|x k | + (1/2)v k 2 .
k=0

If ȳ
K F (x k ) for all k, then (ȳ) (F (x k )) for all k and the conclusion follows. 

Now we are in conditions of studying the convergence properties of Algorithm 2.

406 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

Theorem 4.2. All accumulation points of {x k } are K-critical.

Proof. Let x̄ be an accumulation point of {x k }. Then there exists a subsequence {x kj } converging to x̄,
lim x kj = x̄.
j →∞

Note that {vx kj } and {x kj } are bounded because they converge to vx̄ and x̄ , respectively. Therefore, using
Lemma 3.5 we conclude that {v kj } is also bounded. So (refining the original sequence if necessary), we
may also assume that {v kj } converges to some v̄,
lim v kj = v̄.
j →∞

For all k, fx k (v k ) + (1/2)v k 2  (1 − )x k . Taking limits along k = kj for j → ∞ we get,

fx̄ (v̄) + (1/2)v̄2  (1 − )x̄ . (26)
Recall that Algorithm 2 is a particular case of Algorithm 1, so by Proposition 2.2, F (x̄) is a K-lower
bound for {F (x k )}. Now we may apply Lemma 4.1 to conclude that,
lim tkj x kj = 0, (27)
j →∞
lim tkj v kj  = 0. (28)
j →∞

We claim that
v̄ = 0. (29)
Suppose, for contradictory purposes, that v̄ = 0. As x̄  0, using (26) we get fx̄ (v̄) < 0. Using Proposition
3.6 we conclude that there exists tˆ > 0 such that, for j large enough (greater than some j0 )
F (x kj + tv kj )
K F (x kj ) +
tJ F (x kj )v kj ∀t ∈ [0, tˆ). (30)
Now we will show that for j larger than such j0 ,
2tkj  min{1, tˆ}. (31)
Indeed, let j > j0 . If tkj = 1, the claim holds. If tkj < 1, then, this stepsize was obtained by a backtracking
procedure, where the “previous” possible stepsize 2tkj does not satisfy the descend condition, or equiva-
lently, 2tkj does not belong to the kj th set in (18). Now using (30) we conclude that 2tkj  tˆ and the claim
holds also in this case.
The assumption v̄  = 0 also implies, by (28), that limj →∞ tkj = 0, which contradicts (31).
To end the proof, use (26) and (29) to obtain x̄  0. Since x  0 for any x, we conclude that x̄ = 0
and x̄ is K-critical. 

5. Scalarization and Algorithm 2

A very useful method for solving problem (1) is the so-called scalarization procedure. The method is
quite elegant; it consists of minimizing a certain scalar function, as explained in the sequel. Take some
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 407

w ∈ int(K ∗ ) and define g : Rn → R,

g(x) = w, F (x). (32)

Then, solutions of

min g(x), x ∈ Rn

are also solutions of (1). So, we only need to minimize a (smooth) scalar function, and for this problem
there are many efficient algorithms. The choice of w ∈ int(K ∗ ) is of capital importance. Indeed, for
very well behaved problems, many choices of w lead to unbounded scalar minimizations problems. For
example, in multiobjective optimization, let F : R → R2 ,
  
F (x) = x, 1 + x 2 .

Note that F is component-wise convex. In this context, K = K ∗ = R2+ , and y
K y  means yi  yi , for
i = 1, 2. Hence, such F is K-convex. If we take w = (w1 , w2 ) > 0, with w1 > w2 , then the scalarized
problem is unbounded. Of course, algorithms for choosing w are quite desirable.
Once we have a very simple example in which the “wrong” choice of w breaks down the method, a very
natural question is how does the K-steepest descent method behave in this example. First of all, observe
that, in the above example, the set of K-critical points is given by the halfline (−∞, 0]; furthermore, it
coincides with the K-optimal set. Let {x k } be the sequence generated by Algorithm 2. If this sequence is
finite, then the last iterate is K-critical, and therefore optimal. Suppose that the algorithm does not stop.
Observe that, in this example, for any x̃ ∈ R2 , the set

{x ∈ R2 | F (x)  F (x̃)}

is bounded.As {F (x k )} is K-nonincreasing, it follows that {x k } is bounded. So, it has accumulations points,

all of which, by Theorem 4.2, are K-critical (hence K-optimal). Furthermore, if x̄, x̂ are accumulation
points, then F (x̄) = F (x̂). This readily implies (in this particular example) x̂ = x̄. So {x k } converges to
a solution.
It would be desirable to combine some ideas presented in the preceding sections with the scalarization
method. Indeed, they are connected. Observe that, for g defined as in (32),

∇g(x) = J F (x)t w.

So, the steepest descent direction for g at x is −J F (x)t w. Take x̄ ∈ Rn . We claim that for a suitable
w ∈ K ∗ , the steepest descent direction for the scalarized objective function g at x̄ coincides with the
K-steepest descent direction vx̄ . To prove this claim, define
 = conv(C).
C

Obviously, C is a convex compact set and (y) = sup w, y. The K-steepest descent direction at x̄ is
w∈C
the solution of (17) with x = x̄, which may also be written as


minn max w, J F (x̄)v + (1/2)v2 . (33)

v∈R w∈C
408 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

The dual of this problem is

max minn (w, J F (x̄)v + (1/2)v2 ). (34)
 v∈R
w∈C

Trivially,

arg minn w, J F (x̄)v + (1/2)v2 = −J F (x̄)t w,

v∈R
minn w, J F (x̄)v + (1/2)v2 = (−1/2)J F (x̄)t w2 .
v∈R

Hence, problem (34) may be simplified to

max −(1/2)J F (x̄)t w2 (35)

w∈C

or, equivalently,
min (1/2)J F (x̄)t w2 . (36)

w∈C

 is convex and compact, problem (34) has always a solution, say wx̄ (which may not be unique)
Since C

and there is no duality gap. In particular, (vx̄ , wx̄ ) is a saddle point of w, J F (x̄)v+(1/2)v2 in Rn × C:
w, J F (x̄)vx̄  + (1/2)vx̄ 2  wx̄ , J F (x̄)vx̄  + (1/2)vx̄ 2  wx̄ , J F (x̄)v + (1/2)v2
 So,
for all (v, w) ∈ Rn × C.
vx̄ = −J F (x̄)t wx̄ (37)
and wx̄ ∈ C ⊆ K ∗ . Taking w = wx̄ in (32) we get −∇g(x) = −J F (x)t wx̄ , and so −∇g(x̄) = vx̄ , as we
claimed.
Since there is no duality gap in (33)–(35), using (37) we get,
x̄ = −(1/2)vx̄ 2 ,

where x̄ is the optimal value of problem (17) for x = x̄, so

(J F (x̄)vx̄ ) = −vx̄ 2 .

Now we will prove that approximate solutions of the dual problem (34) (or (35), (36)) yield approximate
solutions of (33), i.e., approximate K-steepest descent directions at x.

 and
Proposition 5.1. Let x ∈ Rn be non K-critical and  ∈ (0, 1). There exists  > 0 such that, if w̃ ∈ C
(1/2)J F (x)t w̃2 − (1/2)J F (x)t wx 2 ,
then ṽ = −J F (x)t w̃ is a -approximate K-steepest descent direction at x.

Proof. Define
ε = −x .
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 409

As x is non K-critical, x < 0, and so ε > 0. Since the objective function on (12) and JF are continuous,
there exists  > 0 such that
v − vx   ⇒ (J F (x)v) + (1/2)v2 (J F (x)vx ) + (1/2)vx 2 + ε.
So, v − vx   implies that v is a -approximate K-steepest descent direction at x.
Optimality of wx for (36), convexity of C and the hypothesis w̃ ∈ C
 imply
J F (x)t (w̃ − wx ), J F (x)t wx   0.
Hence, using the equalities ṽ = −J F (x)t w̃, vx = −J F (x)t wx ,
J F (x)t w̃2  J F (x)t wx 2 + J F (x)t (w̃ − wx )2
= J F (x)t wx 2 + ṽ − vx 2 .
Therefore,

ṽ − vx   J F (x)t w̃2 − J F (x)t wx 2 .

So, it is enough to take  = 2 /2 > 0. 

Let w ∈ C  be an approximate solution of (35) (or (36)). A question of practical relevance is whether
v = −J F (x)t w̃ is a -approximate K-steepest descent direction at x. In our next proposition we give a
sufficient condition for -approximation.
 and define v = −J F (x)t w. If
Proposition 5.2. Take  ∈ [0, 1), w ∈ C
(J F (x)v)  − (1 − /2)v2 ,
then v is a -approximate K-steepest descent direction at x.

Proof. We already know that (17) (or (33)) and (35) are a primal-dual pair of problems. Since w is dual
feasible,
−(1/2)J F (x)t w2 x .
Therefore, making the substitution v = −J F (x)t w we get
(1 − )(−1/2)v2  (1 − )x . (38)
Hence, if
(J F (x)v)  − (1 − /2)v2 ,
from (38) it follows that v = −J F (x)t w is a -approximate K-steepest descent direction. 

In the following section, we will show a theoretical advantage of using descent directions as discussed
in Propositions 5.1, 5.2. From now on, we will say that v is scalarization compatible, or s-compatible if
there exist some w ∈ C  such that
v = −J F (x)t w = ∇x w, F (x).
Note that vx , the exact K-steepest descent direction at x, is always s-compatible.
410 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

6. Convergence analysis: the K-convex case

Throughout this section we will assume that F is K-convex, i.e.,

F (x + (1 − )x  )
K F (x) + (1 − )F (x  )

for all x, x  ∈ Rn and all  ∈ [0, 1]. Under this additional assumption we have the following extension of
the classical gradient inequality to the vector case:

F (x) + J F (x)(x  − x)
K F (x  ),

for any x, x  ∈ Rn (see [19, Lemma 5.2]).

As in the general case, here also optimality implies K-criticality. A point x ∗ ∈ Rn is a weak uncon-
strained K-minimizer (or weak K-optimum) of F, or weak Pareto minimal element for F (see [19]), if
there is no x ∈ Rn with F (x)≺K F (x ∗ ). A well known fact is that, under the K-convexity assumption on
the objective function F, K-criticality and weak optimality are equivalent conditions.
If the algorithm has finite termination, the last iterate is K-critical and hence a weak unconstrained
K-minimizer of F. We will study the case in which the algorithm does not have finite termination and
therefore produces infinite sequences {x k }, {v k } and {tk }. Let us now establish the additional assumptions
under which we will prove full convergence of {x k } to a K-critical point or, in view of the above discussion,
to a weak unconstrained K-minimizer of F.

A1. Every K-decreasing sequence in the image of F

{y k } ⊆ {F (x) | x ∈ Rn }

is K-bounded below by a point in the image of F.

 such that,
A2. All v k ’s are scalarization compatible, i.e., there exists a sequence {w k } in C

v k = −J F (x k )t wk , k = 0, 1, . . .

Some comments concerning the generality/restrictiveness of these assumptions are in order. Regarding
Assumption A1, in the case of classical unconstrained (convex) optimization, this condition is equivalent
to existence of solutions of the optimization problem. This assumption, known as K-completeness, is
standard for ensuring existence of efficient points for vector optimization problems (see [19, Section
3]). Assumption A2 deals with the implementation of the algorithm rather than with F. This assumption
holds if each v k is the exact K-steepest descent direction at x k (see (37)). Of course, Assumption A2
applies to the sequence of directions {v k } prescribed by the algorithm, i.e., we are assuming that v k is an
s-compatible -approximate K-steepest descent direction for all k.
We will need the following technical lemma in order to prove that the K-steepest descent method is
convergent.

Lemma 6.1. Suppose that F is K-convex and that v k is scalarization compatible (at x k ). If F (x̂)
K F (x k )
then

x̂ − x k+1 2  x̂ − x k 2 + x k+1 − x k 2 .

 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 411

 such that
Proof. By assumption, there exists some wk ∈ C
v k = −J F (x k )t wk .
Using the K-convexity of F we have F (x k ) + J F (x k )(x̂ − x k )
K F (x̂). Since F (x̂)
K F (x k ), we get
J F (x k )(x̂ − x k )
K 0.
Taking into account that wk ∈ K ∗ and using the above results we get
−(v k )t (x̂ − x k ) = (w k )t J F (x k )(x̂ − x k )  0.
Recall that x k+1 = x k + tk v k , with tk > 0. Therefore
(x k − x k+1 )t (x̂ − x k )  0,
which implies the desired inequality, because
x̂ − x k+1 2 = x̂ − x k 2 + x k − x k+1 2 + 2(x k − x k+1 )t (x̂ − x k ). 

Before stating our convergence result, we recall that a sequence {y k } ⊂ Rm is quasi-Fejér convergent
[10,11] to a set U ⊂ Rm if for every u ∈ U there exists a sequence {k } ⊂ R, k  0 such that
y k+1 − u2  y k − u2 + k for all k = 1, 2, . . . ,
with
∞

k < ∞.
k=1

We will also need the following result concerning quasi-Fejér convergent sequences, whose proof can
be found in [4,18].

Theorem 6.2. If the sequence {y k } is quasi-Fejér convergent to a nonempty set U ⊂ Rm , then {y k } is

bounded. If furthermore a cluster point y of {y k } belongs to U, then limk→∞ y k = y.

In [4], it is proved that the steepest descent method for smooth (scalar) convex minimization, with
stepsize obtained using backtracking and Armijo rule, is globally convergent to a solution (under the sole
assumption of existence of optima). We will extend those results to the K-steepest descent method, using
the same techniques as in [4].

Theorem 6.3. Suppose that F is K-convex and that Assumptions A1, A2 hold. Then {x k } converges to a
K-critical point x ∗ .

Proof. First of all, observe that all results of Section 4 are still valid under the additional assumptions of
this theorem. In particular, {F (x k )} is a K-decreasing sequence, so, using Assumptions A1, there exists
an x̄ ∈ Rn such that,
F (x̄)
K F (x k ) ∀k ∈ N. (39)
412 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

Now observe that 0 < tk  1 for all k. Hence

1 k+1
tk v(x k )2 = x − x k 2  x k+1 − x k 2 . (40)
tk
Therefore, from (39), (40) and Lemma 4.1 it follows that
∞

x k − x k+1 2 < ∞.
k=1

Define

L := {x ∈ Rn | F (x)
K F (x k ) ∀k ∈ N}.

Note that x̄ ∈ L, so L is nonempty. Using Assumption A2 and Lemma 6.1, we conclude that for any
x ∈ L (and k ∈ N),

x − x k+1 2  x − x k 2 + x k − x k+1 2 .

Since ∞ k
k=1 x − x
k+1 2 < ∞, we conclude that the sequence {x k } is quasi-Fejér convergent to the set

L. As L is nonempty, from Theorem 6.2 it follows that {x k } has accumulation points. Let x ∗ be one of
them. By Proposition 2.2, x ∗ ∈ L. Then, once again by virtue of Theorem 6.2, it follows that the whole
sequence {x k } converges to x ∗ . We finish the proof by observing that Theorem 4.2 guarantees that x ∗ is
K-critical. 

7. Vector optimization and the abstract equilibrium problem

In this section, we discuss the connection between the problem of seeking a weak constrained K-
minimizer and the Abstract Equilibrium problem. We will see that by means of the function , defined
in (12), the first problem can be viewed as a particular case of the second one.
The Weak Constrained K-Minimization problem [19] is defined in the following way:
Given closed convex pointed cones with nonempty interior K ⊂ Rm , Ki ⊂ Rmi , i = 1, 2, . . . , r, the
corresponding induced orders: x≺K y if y − x ∈ int(K), u
Ki v if v − u ∈ Ki , and
F : Rn → Rm , K-convex and continuously differentiable,
Gi : Rn → Rmi , Ki -convex, for all i = 1, 2, . . . , r,
find x ∗ ∈ M = {x ∈ Rn | Gi (x)
Ki 0, for 1  i  r} such that there does not exist any other x ∈ M with,
F (x)≺K F (x ∗ ).
The Abstract Equilibrium problem [3] can be stated in the following way: Given M, a nonempty closed
convex set in a Hausdorff topological vector space and

f :M ×M →R

such that
f (·, y) is upper semicontinuous for all y ∈ M,
f (x, ·) is convex and lower semicontinuous for all x ∈ M,
f (x, x) = 0 for all x ∈ M,
 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414 413

find x̄ ∈ M with the property that

f (x̄, y)  0 for all y ∈ M.
Taking
M = {x ∈ Rn | Gi (x)
Ki 0 for 1  i  r}
as in the weak constrained K-minimization problem, and
f (x, y) := (F (y) − F (x)),
where  was defined in (12), we have that the solution set of this equilibrium problem is exactly the same
as the solution set of the weak constrained K-minimization problem. Observe that the weak unconstrained
K-minimization problem is a particular case of the constrained one and, therefore, can also be solved by
means of methods for solving the equilibrium problem (and conversely). So we can find K-critical points
for problem (1) via the equilibrium problem formulation.

8. Final remarks

In this work we proposed for vector unconstrained minimization an extension of the standard steepest
descent method. We showed that all cluster points of the sequences produced by the method satisfy a
certain first-order condition for K-optimality, known as K-criticality. Under K-convexity of the objective
function and assuming a very reasonable condition on the objective function, we proved that we have full
convergence of the method, when performed, for example, with the exact K-steepest descent direction at
each iteration. In this situation, no matter how bad is our initializing point, the method will converge to a
K-critical point, or, in other words, to a weak unconstrained K-minimizer. As in the one-dimensional case,
using the notion of quasi-Fejér convergence, we could prove convergence of the method without assuming
a Lipschitz condition on the objective’s Jacobian and without bounded level sets requirement. Incidentally,
we showed that every weak constrained K-minimization problem can be viewed as a particular equilibrium
problem.
It is worth mentioning that the compact set C ⊂ K ∗ used throughout the whole work is essential for
defining the K-steepest descent direction (or approximations of it). So, somehow, C plays the role of a
sort of “gauge”. Perhaps, it would have been better to define the “gauged” K-steepest descent direction,
or the K-steepest descent direction modulus C, instead of simply the K-steepest descent direction. This
would make the notation more clear but certainly heavier.
Regarding the implementation of the method, we mention that when K (therefore K ∗ ) is finitely
generated, the K-steepest descent direction can be easily computed. In the general case, approximate
K-steepest descent direction can be computed solving approximately (36) and using Proposition 5.1 or
Proposition 5.2. Other possibility is to solve (17) using a bundle method technique [17].
We expect that, in spite of the possible drawbacks of the method, as in the single-valued case, it
will furnish a prototype for more sophisticated and efficient algorithms for solving vector optimization
problems. So we think that a full understanding of the structure and convergence behavior of the K-steepest
descent method is indeed relevant.
The extension to vector optimization of more efficient algorithms, as, for instance, Newton’s and
quasi-Newton methods, and its convergence analysis are left as open problems for future research.
414 L.M. Graña Drummond, B.F. Svaiter / Journal of Computational and Applied Mathematics 175 (2005) 395 – 414

References

[1] C.D. Aliprantis, M. Florenzano, V.F. Martins da Rocha, R. Tourky, Equilibrium analysis in financial markets with countably
many securities, J. Math. Econom. 40 (2004) 683–699.
[2] C.D. Aliprantis, M. Florenzano, R. Tourky, General equilibrium analysis in ordered topological vector spaces, J. Math.
Econom. 40 (3–4) (2004) 247–269.
[3] E. Blum, W. Oettli, From optimization and variational inequalities to equilibrium problems, Math. Student 63 (1–4) (1994)
123–145.
[4] R. Burachik, L.M. Graña Drummond, A.N. Iusem, B.F. Svaiter, Full convergence of the steepest descent method with
inexact line searches, Optimization 32 (1995) 137–146.
[5] E. Carrizosa, E. Conde, M. Muñoz Márquez, J. Puerto, Planar point-objective location problems with nonconvex constraints:
a geometrical construction, J. Global Optim. 6 (1) (1995) 77–86.
[6] E. Carrizosa, J.B.G. Frenk, Dominating sets for convex functions with some applications, J. Optim. Theory Appl. 96 (2)
(1998) 281–295.
[7] I. Das, J.E. Dennis, A closer look at drawbacks of minimizing weighted sums of objectives for Pareto set generation in
multi-criteria optimization problems, Structural Optim. 14 (1) (1997) 63–69.
[8] I. Das, J.E. Dennis, Normal-boundary intersection: a new method for generating the Pareto surface in nonlinear multicriteria
optimization problems, SIAM J. Optim. 8 (3) (1998) 631–657 (electronic).
[9] J.E. Dennis Jr., R.B. Schnabel, Numerical Methods for Unconstrained Optimization and Nonlinear Equations, Society for
Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1996 (Corrected reprint of the 1983 original).
[10] Ju.M. Ermol’ev, On the method of generalized stochastic gradients and stochastic quasi-Féjer sequences, Kibernetika
(Kiev) 1969 (2) (1969) 73–83.
[11] Ju. M. Ermol’ev, A.D. Tuniev, Random Féjer and quasi-Féjer sequences, in: Theory of Optimal Solutions (Proc. Sem.,
Kiev, 1968), No. 2 (Russian), Akad. Nauk Ukrain. SSR, Kiev, 1968, pp. 76–83.
[12] H. Eschenauer, J. Koski, A. Osyczka, Multicriteria Design Optimization, Springer, Berlin, 1990.
[13] G.W. Evans, An overview of techniques for solving multiobjective mathematical programs, Management Sci. 30 (11)
(1984) 1268–1282.
[14] J. Fliege, B.F. Svaiter, Steepest descent methods for multicriteria optimization, Math. Methods Oper. Res. 51 (3) (2000)
479–494.
[15] A.M. Geoffrion, Proper efficiency and the theory of vector maximization, J. Math. Anal. Appl. 22 (1968) 618–630.
[16] Y.Y. Haimes, L.S. Lasdon, D.A. Wismer, On a bicriterion formulation of the problems of integrated system identification
and system optimization, IEEE Trans. Systems Man Cybernet. SMC-1 (1971) 296–297.
[17] J.B. Hiriart-Urruty, C. Lemaréchal, Convex Analysis and Minimization Algorithms, II, Fundamental Principles of
Mathematical Science, Springer, Berlin, 1993.
[18] A.N. Iusem, B.F. Svaiter, M. Teboulle, Entropy-like proximal methods in convex programming, Math. Oper. Res. 19 (4)
(1994) 790–814.
[19] T.D. Luc, Theory of Vector Optimization, Lecture Notes in Economics and Mathematical Systems, vol. 319, Springer,
Berlin, 1989.
[20] K. Miettinen, M.M. Mäkelä, Interactive bundle-based method for nondifferentiable multiobjective optimization: Nimbus,
Optimization 34 (1995) 231–246.
[21] K.M. Miettinen, Nonlinear Multiobjective Optimization, Kluwer Academic Publishers, Norwell, 1999.
[22] R.T. Rockafellar, Convex Analysis, Princeton University Press, New Jersey, 1970.
[23] C. Tammer, K. Winkler, A new scalarization approach and applications in multicriteria, J. Nonlinear Convex Anal. 4 (3)
(2003) 365–380.
[24] D.J. White, Epsilon-dominating solutions in mean-variance portfolio analysis, European J. Oper. Res. 105 (1998)
457–466.