J. Math. Anal. Appl.

320 (2006) 787–794

www.elsevier.com/locate/jmaa

The strong Ekeland variational principle

Tomonari Suzuki 1
Department of Mathematics, Kyushu Institute of Technology, Sensuicho, Tobata, Kitakyushu 804-8550, Japan
Received 5 January 2005
Available online 31 August 2005
Submitted by C.E. Chidume

Abstract
In this paper, we consider the strong Ekeland variational principle due to Georgiev [P.G. Georgiev,
The strong Ekeland variational principle, the strong drop theorem and applications, J. Math. Anal.
Appl. 131 (1988) 1–21]. We discuss it for functions defined on Banach spaces and on compact metric
spaces. We also prove the τ -distance version of it.
© 2005 Elsevier Inc. All rights reserved.

Keywords: The Ekeland variational principle; Quasiconvex function; τ -Distance

1. Introduction

Throughout this paper we denote by N the set of all positive integers and by R the set
of all real numbers.
In [4,5], Ekeland proved the following, which is called the Ekeland variational princi-
ple.

Theorem 1. (Ekeland [4,5]) Let X be a complete metric space with metric d. Let f be a
function from X into (−∞, +∞] which is proper lower semicontinuous and bounded from

E-mail address: suzuki-t@mns.kyutech.ac.jp.

1 The author is supported in part by Grants-in-Aid for Scientific Research from the Japanese Ministry of
Education, Culture, Sports, Science and Technology.

0022-247X/$ – see front matter © 2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2005.08.004
788 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794

below. Then for u ∈ X and λ > 0, there exists v ∈ X such that f (v)  f (u) − λ d(u, v)
and f (w) > f (v) − λ d(v, w) for every w ∈ X \ {v}.

This principle is equivalent to Caristi’s fixed point theorem [2,3] and nonconvex mini-
mization theorem according to Takahashi [18]. So this principle is one of generalizations
of the Banach contraction principle [1].
In 1988, Georgiev proved the following interesting generalization of Theorem 1, which
is called the strong Ekeland variational principle. See also the proof in [20].

Theorem 2. (Georgiev [6]) Let X be a complete metric space with metric d. Let f be a
function from X into (−∞, +∞] which is proper lower semicontinuous and bounded from
below. Then for u ∈ X, λ > 0 and δ > 0, there exists v ∈ X satisfying the following:

(i) f (v) < f (u) − λ d(u, v) + δ;

(ii) f (w) > f (v) − λ d(v, w) for all w ∈ X \ {v};
(iii) if a sequence {xn } in X satisfies limn (f (xn ) + λ d(v, xn )) = f (v), then {xn } converges
to v.

In 2001, the author introduced the notion of τ -distance to generalize Theorem 1 and to
improve the results in Tataru [19], Zhong [21,22] and others.
Let X be a metric space with metric d. Then a function p from X × X into [0, ∞) is
called a τ -distance on X [10] if there exists a function η from X × [0, ∞) into [0, ∞) and
the following are satisfied:

(τ 1) p(x, z)  p(x, y) + p(y, z) for all x, y, z ∈ X;

(τ 2) η(x, 0) = 0 and η(x, t)  t for all x ∈ X and t ∈ [0, ∞), and η is concave and con-
tinuous in its second variable;
(τ 3) limn xn = x and limn sup{η(zn , p(zn , xm )): m  n} = 0 imply p(w, x) 
lim infn p(w, xn ) for all w ∈ X;
(τ 4) limn sup{p(xn , ym ): m  n} = 0 and limn η(xn , tn ) = 0 imply limn η(yn , tn ) = 0;
(τ 5) limn η(zn , p(zn , xn )) = 0 and limn η(zn , p(zn , yn )) = 0 imply limn d(xn , yn ) = 0.

We note that the metric d is a τ -distance on X. Many useful examples and properties
are stated in [7–17]. The following is the τ -distance version of Theorem 1.

Theorem 3. [10] Let X be a complete metric space and let p be a τ -distance on X. Let f
be a function from X into (−∞, +∞] which is proper lower semicontinuous and bounded
from below. Then the following hold:

(A) For each u ∈ X, there exists v ∈ X such that f (v)  f (u) and f (w) > f (v)−p(v, w)
for all w ∈ X \ {v};
(B) for each λ > 0 and u ∈ X with p(u, u) = 0, there exists v ∈ X such that f (v) 
f (u) − λp(u, v) and f (w) > f (v) − λp(v, w) for all w ∈ X \ {v}.
 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794 789

In this paper, we consider the strong Ekeland variational principle (Theorem 2). We
discuss it for functions defined on Banach spaces and on compact metric spaces. Also,
using the notion of τ -distance, we generalize Theorem 2.

2. Functions on Banach spaces

We assume that X is the dual space of some Banach space. Let f be a function from X
into (−∞, +∞]. We recall that f is said to be lower semicontinuous in the weak∗ topology
if
 
f w∗ -lim xα  lim inf f (xα )
α α
for every weakly∗ convergent net {xα } in X. f is called quasiconvex if
 
x ∈ X: f (x)  α
is a convex subset of X for every α ∈ R. It is obvious that convexity implies quasiconvexity.
In this section, we shall show that a stronger conclusion of Theorems 1 and 2 holds
when f is a quasiconvex function defined on a reflexive Banach space. We first prove the
following.

Theorem 4. Let f be a function from a Banach space X into (−∞, +∞] which is proper
lower semicontinuous in the weak∗ topology and bounded from below. Then for u ∈ X and
λ > 0, there exists v ∈ X satisfying the following:

(i) f (v)  f (u) − λu − v;

(ii) f (w) > f (v) − λv − w for all w ∈ X \ {v};
(iii) if a net {xα } in X satisfies limα (f (xα ) + λv − xα ) = f (v), then {xα } converges
weakly∗ to v.

Proof. From Theorem 1, there exists v ∈ X satisfying (i) and (ii). We note that f (v) < ∞
from (ii). We shall show that such v also satisfies (iii). Let {xα } be a net in X with
limα (f (xα ) + λv − xα ) = f (v). Then we have
lim sup xα   v + lim sup v − xα 
α α
f (xα ) + λv − xα  1
 v + lim sup − inf f (x)
α λ λ x∈X
1 
= v + f (v) − inf f (x)
λ x∈X
< ∞.
Thus, without loss of generality, we can consider {xα } as bounded. Arguing by contradic-
tion, we assume that {xα } does not converge weakly∗ to v. Then from Alaoglu’s theorem,
there exist a subnet {yβ } of {xα } and y ∈ X such that {yβ } converges weakly∗ to y, and
y = v. Since x → v − x is also lower semicontinuous in the weak∗ topology, we obtain
 
f (y) + λv − y  lim f (yβ ) + λv − yβ  = f (v).
β
This contradicts to (ii). So we obtain the desired result. 2
790 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794

Using Theorem 4, we can prove the following.

Theorem 5. Let f be a quasiconvex function from a reflexive Banach space X into

(−∞, +∞] which is proper lower semicontinuous and bounded from below. Then for
u ∈ X and λ > 0, there exists v ∈ X satisfying (i), (ii) in Theorem 4, and

(iii) if a sequence {xn } in X satisfies limn (f (xn )+λv −xn ) = f (v), then {xn } converges
strongly to v.

Proof. We first note that from the reflexivity of X and quasiconvexity of f , the notions
of lower semicontinuity in the strong, the weak, and the weak∗ topology are equivalent.
Thus, for u ∈ X and λ > 0, there exists v ∈ X with (i)–(iii) in Theorem 4. We note that
f (v) < ∞. Now, let us prove (iii) in this theorem. Let {xn } be a sequence in X satisfying
limn (f (xn ) + λv − xn ) = f (v). From (iii) in Theorem 4, such {xn } converges weakly
to v. Arguing by contradiction, assume that {xn } does not converge strongly to v. Then
there exists a subsequence {yk } of {xn } and ε > 0 such that
yk − v > 2ε and f (yk ) + λv − yk  < f (v) + λε
for all k ∈ N. Since f (yk ) < f (v) − λε for k ∈ N, and v is an element of the closed convex
hull of {yk : k ∈ N}, we obtain f (v)  f (v) − λε. This is a contradiction. This completes
the proof. 2

3. τ -Distance versions

In this section, using the notion of τ -distance, we generalize Theorem 2.

Let X be a metric space with metric d and let p be a τ -distance on X. Then a sequence
{xn } in X is called p-Cauchy if there exists a function η from X × [0, ∞) into [0, ∞) satis-
fying (τ 2)–(τ 5) and a sequence {zn } in X such that limn sup{η(zn , p(zn , xm )): m  n} = 0.
We know the following.

Lemma 1. [10] Let X be a metric space and let p be a τ -distance on X. If {xn } is a

p-Cauchy sequence, then {xn } is a Cauchy sequence in the usual sense.

Lemma 2. [10] Let X be a metric space and let p be a τ -distance on X. If a sequence {xn }
in X satisfies limn p(z, xn ) = 0 for some z ∈ X, then {xn } is a p-Cauchy sequence.

Lemma 3. [10] Let p be a τ -distance on a metric space X and let c be a positive real
number. Then a function q from X × X into [0, ∞) defined by q(x, y) = cp(x, y) for all
x, y ∈ X is also a τ -distance on X.

Using the above lemmas and Theorem 3(B), we obtain the following.

Theorem 6. Let X be a complete metric space and let p be a τ -distance on X. Let f be a

function from X into (−∞, +∞] which is proper lower semicontinuous and bounded from
 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794 791

below. Let u ∈ X with p(u, u) = 0. Then for λ > 0 and δ > 0, there exists v ∈ X satisfying
the following:

(i) f (v)  f (u);

(ii) f (v) < f (u) − λp(u, v) + δ;
(iii) f (w) > f (v) − λp(v, w) for all w ∈ X \ {v};
(iv) if a sequence {xn } in X satisfies limn (f (xn ) + λp(v, xn )) = f (v), then {xn } is
p-Cauchy, limn xn = v and p(v, v) = limn p(v, xn ) = 0.

Remark. p(v, v) = 0 does not necessarily hold. If there exists a sequence {xn } satisfying
the assumption of (iv), then p(v, v) = 0 holds.

Proof. In the case of f (u) = ∞, (i) and (ii) hold for all v ∈ X. We also note that (iii)
and (iv) do not depend on u. Thus, it is sufficient to show the conclusion in the case of
f (u) < ∞. We assume f (u) < ∞. We take λ ∈ (0, λ) satisfying
λ − λ  
f (u) − inf f (x) < δ.
λ x∈X

By Theorem 3(B), there exists v ∈ X such that f (v)  f (u) − λ p(u, v) and f (w) >
f (v) − λ p(v, w) for all w ∈ X \ {v}. It is obvious that v satisfies (i). We also have

λ − λ λ − λ
f (v) = 1 + f (v) − f (v)
λ λ

λ − λ   λ − λ
 1+
f (u) − λ p(u, v) − f (v)
λ λ
λ − λ  
= f (u) − λp(u, v) +
f (u) − f (v)
λ
λ − λ  
 f (u) − λp(u, v) + f (u) − inf f (x)
λ x∈X
< f (u) − λp(u, v) + δ.
This is (ii). For w ∈ X \ {v}, we have
f (w) > f (v) − λ p(v, w)  f (v) − λp(v, w).
This is (iii). We finally show (iv). We assume that a sequence {xn } in X satisfies
limn (f (xn ) + λp(v, xn )) = f (v). We note that
f (w)  f (v) − λ p(v, w)
for all w ∈ X. So we have
λp(v, xn ) − λ p(v, xn )
lim sup p(v, xn ) = lim sup
n→∞ n→∞ λ − λ
λp(v, xn ) + f (xn ) − f (v)
 lim
n→∞ λ − λ
= 0.
792 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794

By Lemma 2, {xn } is a p-Cauchy sequence and hence is a Cauchy sequence. Since X is

complete, {xn } converges to some point x ∈ X. From (τ 3), we have p(v, x) = 0. We also
have
 
f (x)  lim inf f (xn )  lim f (xn ) + λp(v, xn ) = f (v).
n→∞ n→∞

Thus, if v = x, then we have

f (x) > f (v) − λ p(v, x) = f (v)  f (x).
This is a contradiction. Hence we obtain v = x. This completes the proof. 2

Proving Theorem 6, we use Theorem 3(B). Conversely, can we prove Theorem 3(B) by
using Theorem 6? In the case that p is the metric d, we can do it because
 
Y = x ∈ X: f (x) + λd(u, x)  f (u)
is a nonempty complete metric space. More generally, we can do it in the case that p is
lower semicontinuous in its second variable. However, the author thinks that it is difficult to
prove Theorem 3(B) from Theorem 6 directly. Thus, we do not call Theorem 6 the ‘strong’
version of Theorem 3(B). On the other hand, we can prove the following, which is stronger
than Theorem 3(A).

Theorem 7. Let X be a complete metric space and let p be a τ -distance on X. Let f be a

function from X into (−∞, +∞] which is proper lower semicontinuous and bounded from
below. Then for u ∈ X, there exists v ∈ X satisfying the following:

(i) f (v)  f (u);

(ii) f (w) > f (v) − p(v, w) for all w ∈ X \ {v};
(iii) if a sequence {xn } in X satisfies limn (f (xn ) + p(v, xn )) = f (v), then {xn } is
p-Cauchy, limn xn = v and p(v, v) = limn p(v, xn ) = 0.

Proof. We note that (x, y) → (1/2)p(x, y) is also a τ -distance on X. By Theorem 3(A),

there exists v ∈ X such that f (v)  f (u) and f (w) > f (v) − (1/2)p(v, w) for all w ∈
X \ {v}. Hence v satisfies (i). It is not difficult to check that v also satisfies (ii) and (iii).
This completes the proof. 2

Using Theorem 7, we prove the following strong version of Caristi’s fixed point theo-
rem.

Theorem 8. Let X be a complete metric space and let p be a τ -distance on X. Let T be

a mapping on X and let f be a function from X into (−∞, +∞] which is proper lower
semicontinuous and bounded from below. Assume that f (T x) + p(x, T x)  f (x) for every
x ∈ X. Then there exists a fixed point v ∈ X of T satisfying the following:

(i) p(v, v) = 0;
(ii) f (w) > f (v) − p(v, w) for all w ∈ X \ {v};
 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794 793

(iii) if a sequence {xn } in X satisfies limn (f (xn ) + p(v, xn )) = f (v), then limn xn = v and
limn p(v, xn ) = 0.

Proof. By Theorem 7, there exists v ∈ X satisfying (ii) and (iii). We note f (v) < ∞.
From (ii) and the assumption of f (T v) + p(v, T v)  f (v), we obtain v is a fixed point
of T . Since f (v) + p(v, v)  f (v) and f (v) < ∞, we have p(v, v) = 0. This completes
the proof. 2

As a direct consequence of Theorem 8, we obtain the following.

Corollary 1. Let X be a complete metric space and let T be a mapping on X. Let f be a

function from X into (−∞, +∞] which is proper lower semicontinuous and bounded from
below. Assume that f (T x) + d(x, T x)  f (x) for every x ∈ X. Then there exists a fixed
point v ∈ X of T satisfying the following:

(i) f (w) > f (v) − d(v, w) for all w ∈ X \ {v};

(ii) if a sequence {xn } in X satisfies limn (f (xn ) + d(v, xn )) = f (v), then limn xn = v.

4. Functions on compact metric spaces

When X is compact, we prove a stronger result of Theorems 1 and 2.

Theorem 9. Let X be a compact metric space and let p be a τ -distance on X. Assume

that p is lower semicontinuous in its second variable. Let f be a function from X into
(−∞, +∞] which is proper lower semicontinuous and bounded from below. Let u ∈ X
with p(u, u) = 0. Then for λ > 0, there exists v ∈ X satisfying the following:

(i) f (v)  f (u) − λp(u, v);

(ii) f (w) > f (v) − λp(v, w) for all w ∈ X \ {v};
(iii) if a sequence {xn } in X satisfies limn (f (xn ) + λp(v, xn )) = f (v), then {xn } is
p-Cauchy, limn xn = v and p(v, v) = limn p(v, xn ) = 0.

Proof. From Theorem 3(B), there exists v ∈ X satisfying (i) and (ii). We shall show such
v also satisfies (iii). Let {xn } be a sequence in X with limn (f (xn ) + λp(v, xn )) = f (v).
Let {yk } be an arbitrary convergent subsequence of {xn } and put y = limk yk . Then since
x → f (x) + λp(v, x) is lower semicontinuous, we have
 
f (y) + λp(v, y)  lim f (yk ) + λp(v, yk ) = f (v).
k→∞
From this and (ii), v = y holds. Since {yk } is arbitrary, we obtain {xn } converges to v. We
also have
 
f (v)  lim inf f (xn )  lim sup f (xn )  lim f (xn ) + λp(v, xn ) = f (v)
n→∞ n→∞ n→∞

and hence limn f (xn ) = f (v). Therefore limn p(v, xn ) = 0. Thus, we obtain {xn } is
p-Cauchy, and p(v, v) = 0. This completes the proof. 2
794 T. Suzuki / J. Math. Anal. Appl. 320 (2006) 787–794

As a direct consequence of Theorem 9, we obtain the following.

Corollary 2. Let X be a compact metric space with metric d. Let f be a function from X
into (−∞, +∞] which is proper lower semicontinuous and bounded from below. Then for
u ∈ X and λ > 0, there exists v ∈ X satisfying the following:

(i) f (v)  f (u) − λd(u, v);

(ii) f (w) > f (v) − λd(v, w) for all w ∈ X \ {v};
(iii) if a sequence {xn } in X satisfies limn (f (xn ) + λd(v, xn )) = f (v), then {xn } converges
to v.

References
[1] S. Banach, Sur les opérations dans les ensembles abstraits et leur application aux équations intégrales, Fund.
Math. 3 (1922) 133–181.
[2] J. Caristi, Fixed point theorems for mappings satisfying inwardness conditions, Trans. Amer. Math. Soc. 215
(1976) 241–251.
[3] J. Caristi, W.A. Kirk, Geometric Fixed Point Theory and Inwardness Conditions, Lecture Notes in Math.,
vol. 490, Springer, Berlin, 1975, pp. 74–83.
[4] I. Ekeland, On the variational principle, J. Math. Anal. Appl. 47 (1974) 324–353.
[5] I. Ekeland, Nonconvex minimization problems, Bull. Amer. Math. Soc. 1 (1979) 443–474.
[6] P.G. Georgiev, The strong Ekeland variational principle, the strong drop theorem and applications, J. Math.
Anal. Appl. 131 (1988) 1–21.
[7] O. Kada, T. Suzuki, W. Takahashi, Nonconvex minimization theorems and fixed point theorems in complete
metric spaces, Math. Japan. 44 (1996) 381–391.
[8] N. Shioji, T. Suzuki, W. Takahashi, Contractive mappings, Kannan mappings and metric completeness, Proc.
Amer. Math. Soc. 126 (1998) 3117–3124.
[9] T. Suzuki, Several fixed point theorems in complete metric spaces, Yokohama Math. J. 44 (1997) 61–72.
[10] T. Suzuki, Generalized distance and existence theorems in complete metric spaces, J. Math. Anal. Appl. 253
(2001) 440–458.
[11] T. Suzuki, On Downing–Kirk’s theorem, J. Math. Anal. Appl. 286 (2003) 453–458.
[12] T. Suzuki, Several fixed point theorems concerning τ -distance, Fixed Point Theory Appl. 2004 (2004) 195–
209.
[13] T. Suzuki, Generalized Caristi’s fixed point theorems by Bae and others, J. Math. Anal. Appl. 302 (2005)
502–508.
[14] T. Suzuki, Contractive mappings are Kannan mappings, and Kannan mappings are contractive mappings in
some sense, Comment. Math. Prace Mat., in press.
[15] T. Suzuki, On the relation between the weak Palais–Smale condition and coercivity by Zhong, submitted for
publication.
[16] T. Suzuki, w-Distances and τ -distances, submitted for publication.
[17] T. Suzuki, W. Takahashi, Fixed point theorems and characterizations of metric completeness, Topol. Methods
Nonlinear Anal. 8 (1996) 371–382.
[18] W. Takahashi, Existence theorems generalizing fixed point theorems for multivalued mappings, in: M.A.
Théra, J.B. Baillon (Eds.), Fixed Point Theory and Applications, in: Pitman Res. Notes in Math. Ser.,
vol. 252, Longman, Harlow, 1991, pp. 397–406.
[19] D. Tataru, Viscosity solutions of Hamilton–Jacobi equations with unbounded nonlinear terms, J. Math. Anal.
Appl. 163 (1992) 345–392.
[20] M. Turinici, Strong variational principles and generic well-posedness, Demonstratio Math., in press.
[21] C.-K. Zhong, On Ekeland’s variational principle and a minimax theorem, J. Math. Anal. Appl. 205 (1997)
239–250.
[22] C.-K. Zhong, A generalization of Ekeland’s variational principle and application to the study of the relation
between the weak P.S. condition and coercivity, Nonlinear Anal. 29 (1997) 1421–1431.