Professional Documents
Culture Documents
1. Lyusternik-Graves theorem
The concept of metric regularity and its sibling’s openness with linear rate and
Aubin continuity have their roots in the Banach open mapping principle which,
according to the original treatise of Banach [3], p. 150, says that the surjectivity
of a linear and bounded mapping A, acting from a Banach space (X, · ) to a
Banach space (Y, · ), is equivalent to the existence of a positive constant κ such
that for any y ∈ Y there exists x ∈ X with Ax = y and also x ≤ κy. In terms
of the distance d(x, C) from a point x to a set C the latter condition is equivalent
to d(x, A−1 (y)) ≤ κy − Ax for all (x, y) ∈ X × Y, a property of the mapping A
which is nowadays known as metric regularity.
The Banach open mapping theorem has been extended to nonlinear mappings
in the work of Graves [13] who obtained the following result: let f : X → Y be a
function which is continuous in a neighborhood of x̄, let A : X → Y be a surjective
linear and bounded mapping, and let κ be the constant in the Banach principle for
the mapping A. Let f − A be Lipschitz continuous in a ball IB ε (x̄) centered at x̄
with radius ε > 0 with Lipschitz constant μ such that κμ < 1. Then, in terms of
ȳ = f (x̄), if y satisfies y − ȳ ≤ (κ−1 − μ)ε, then the equation f (x) = y has a
solution x in IB ε (x̄). A slight modification in the original proof of Graves (see [9,
p. 277]) gives us the condition
κ
(1) d(x, f −1 (y)) ≤ y − f (x) for any (x, y) near (x̄, ȳ),
1 − κμ
which is the definition of local metric regularity of a function f at x̄ for ȳ.
Received by the editors December 24, 2009 and, in revised form, March 08, 2010.
2010 Mathematics Subject Classification. Primary 49J53; Secondary 47J22, 49J40, 49K40,
90C31.
Key words and phrases. Metric regularity, openness, Lyusternik-Graves theorem, fixed point,
Ekeland principle.
The first author was supported by National Science Foundation grant DMS 1008341.
2010
c American Mathematical Society
Reverts to public domain 28 years from publication
521
522 ASEN L. DONTCHEV AND HÉLÈNE FRANKOWSKA
The next important step in generalizing the Banach principle is due to Milyutin
(see [5]), who observed that the mapping A in Graves’ theorem can be replaced
by any function h which is open with linear rate, a property that turned out to be
equivalent to metric regularity. In terms of metric regularity, Milyutin’s theorem
says that if a function h is metrically regular at x̄ for ȳ = h(x̄) with constant κ and,
for a function f with f (x̄) = h(x̄) = ȳ, the difference f − h is Lipschitz continuous
with constant μ such that κμ < 1, then f is metrically regular at x̄ for ȳ with
constant κ/(1 − κμ), as in (1). Milyutin and his coauthors [5] linked this result
with a theorem by Lyusternik [16], which characterizes the tangent manifold to the
kernel of a function at a given point, a result weaker than Graves’ theorem. One
should note that both Lyusternik and Graves, as well as Milyutin, used in their
proofs iterative schemes that resemble the Picard iteration or, even more directly,
the Newton method.
In the last several decades, metric regularity has been recognized as a basic
property in the general area of optimization, which serves as a major constraint
qualification condition1 in deriving optimality conditions, and even more impor-
tantly, is very instrumental in obtaining error bounds for perturbed minima and
proving convergence of algorithms for solving optimization problems and beyond.
For a comprehensive treatment of these developments together with historical re-
marks, see the recent book [9], as well as the earlier survey [14].
To proceed, let us first fix the notation. Throughout f : X → Y means that f is
a function (a single-valued mapping) while F : X → → Y denotes a general mapping
which may be set-valued. The graph of F is the set gph F = (x, y) ∈ X ×Y y ∈
F (x) , and the inverse of F is the mapping y → F −1 (y) = x | y ∈ F (x) . The
o
closed ball centered at x with radius r is IB r (x) while the open one is IB r (x); when
X is a linear metric space, the closed unit ball centered at zero is IB and the open
o
one is IB. By convention, for x ∈ X, IB ∞ (x) = X. A set C is said to be locally
closed (resp., complete) at x̄ ∈ C when there exists r > 0 such that C ∩ IB r (x̄)
is closed (resp., complete). The metric in a metric space X is denoted by dX .
The distance from a point x to a nonempty set C in a metric space (X, dX ) is
denoted by d(x, C); d(x, C) = inf v∈C dX (x, v). We set d(x, C) = ∞ iff C = ∅.
The excess from a set C to a set D is e(C, D) = supx∈C d(x, D) and the Pompeiu-
Hausdorff distance between the sets C and D is h(C, D) = max{e(C, D), e(D, C)}.
The infimum distance between points of two sets C and D in (X, dX ) is denoted by
d(C, D); d(C, D) = inf x∈C,v∈D dX (y, v). Given a mapping Φ : X → → X, we denote
Fix(Φ) = {x ∈ X | x ∈ Φ(x)}.
The definition of metric regularity of a general set-valued mapping is as follows:
Definition 1 (metric regularity). Given two metric spaces (X, dX ) and (Y, dY ), a
mapping F : X → → Y is said to be metrically regular at x̄ for ȳ when ȳ ∈ F (x̄) and
there is a constant κ > 0 together with neighborhoods U of x̄ and V of ȳ such that
(2) d(x, F −1 (y)) ≤ κd(y, F (x)) for all (x, y) ∈ U × V.
The infimum of κ over all such combinations of κ, U and V is called the regularity
modulus for F at x̄ for ȳ and is denoted by reg(F ; x̄| ȳ).
1 The main aim of Lyusternik in [16] was deriving a Lagrange multiplier rule for an abstract
It is now well known that the metric regularity is equivalent to two other prop-
erties. The first one is the openness with linear rate defined as follows: A mapping
F : X → → Y is said to be open with linear rate or linearly open at x̄ for ȳ when
ȳ ∈ F (x̄) and there are constants ρ > 0 and h0 > 0 together with neighborhoods
U of x̄ and V of ȳ such that
o o
(3) IB hρ (y) ⊂ F (IB h (x)) for all (x, y) ∈ gph F ∩ (U × V ), h ∈ (0, h0 ).
It is important to note that while the neighborhoods U and V in (2) and (3)
may be different, the constant ρ in (3) could be any positive number smaller than
1/ reg(F ; x̄| ȳ). The standard openness, that is, the property that F (U ) is open when
U is open, is implied by (3), but the converse is not true. Linear openness postulates
openness around the reference point with balls having proportional radii. In the
literature the openness with linear rate is sometimes called the covering property;
e.g., a mapping F is open with rate ρ if it is covering with constant ρ.
The second equivalent property is the Aubin continuity.2 A mapping Ψ : X → →Y
is said to be Aubin continuous at x̄ for ȳ when there are a constant κ and neigh-
borhoods U of x̄ and V of ȳ such that
(4) e(Ψ(x) ∩ V, Ψ(x )) ≤ κdX (x, x ) for all x, x ∈ U.
The infimum of κ over all κ, U and V for which (4) holds is denoted Lip(Ψ; x̄| ȳ).
It turns out that a mapping F is metrically regular at x̄ for ȳ with constant κ
(or, equivalently, linearly open with rate 1/κ) if and only if its inverse F −1 is
Aubin continuous at ȳ for x̄; moreover, Lip(F −1 ; ȳ | x̄) = reg(F ; x̄| ȳ). If Ψ in (4)
is single-valued, say, a function s with s(x̄) = ȳ, then (4) simply means that s is
Lipschitz continuous near x̄, and then Lip(s; x̄|s(x̄)) equals the Lipschitz modulus
of s, denoted lip(s; ȳ), which is the lower limit of the Lipschitz constants of s taken
over all neighborhoods of x̄.
If V = Y in (4), then the mapping Ψ becomes Lipschitz continuous in U with
respect to the Pompeiu-Hausdorff distance; that is,
h(Ψ(x), Ψ(x )) ≤ κdX (x, x ) for all x, x ∈ U.
If in addition U = X, we say that the mapping Ψ is globally Lipschitz continuous.
We will also need a partial version of the Aubin continuity. A mapping Ψ :
P ×X → → Y is said to be partially Aubin continuous with respect to x uniformly in
p at (p̄, x̄) for ȳ if ȳ ∈ Ψ(p̄, x̄) and there is a nonnegative constant κ together with
neighborhoods Q for p̄, U of x̄ and V of ȳ such that
e(Ψ(p, x) ∩ V, Ψ(p, x )) ≤ κdX (x, x ) for all x, x ∈ U and p ∈ Q.
The infimum of κ over all such combinations of κ, Q, U and V is called the par-
tial Lipschitz modulus of Ψ with respect to x uniformly in p and is denoted by
x (Ψ; (p̄, x̄)| ȳ).
Lip
For U = X and V = Y , Definition 1 gives us global metric regularity, in which
the reference point (x̄, ȳ) can be dropped. Also, the property in (3) becomes global
linear openness or a global covering, while, as already mentioned, Aubin continuity
(4) reduces to global Lipschitz continuity. As in the local case, these three global
properties are equivalent to each other (for the Lipschitz continuity of the inverse
mapping). The global properties are quite different from their local counterparts,
and we will see the difference more explicitly in what follows.
2 Originally introduced in [2] under the name pseudo-Lipschitz continuity.
524 ASEN L. DONTCHEV AND HÉLÈNE FRANKOWSKA
Proof. The assumptions for the mappings F and Ψ yield the existence of a positive
constant α such that one of the following conditions holds: (i) gph F ∩ (IB α (x̄) ×
IB α (ȳ)) is complete and gph Ψ ∩ (IB α (x̄) × IB α (ȳ)) is closed; (ii) gph Ψ ∩ (IB α (x̄) ×
IB α (ȳ)) is complete and gph F ∩ (IB α (x̄) × IB α (ȳ)) is closed; (iii) gph(F −1 Ψ) ∩
(IB α (x̄) × IB α (x̄)) is complete. Also,
(9) d(x, F −1 (y)) ≤ κd(y, F (x)) for all (x, y) ∈ IB α (x̄) × IB α (ȳ)
and
(10) e(Ψ(x) ∩ IB α (ȳ), Ψ(x )) ≤ μdX (x, x ) for all (x, x ) ∈ IB α (x̄).
Pick any positive reals a, b and ε so that the following system of inequalities holds:
⎧
⎪
⎪
κ(μ + ε) + ε < 1,
⎪
⎪
⎪
⎪
⎪
⎨ 1
[a + κb + ε] + a ≤ α,
(11) 1 − (κ(μ + ε) + ε)
⎪
⎪
⎪
⎪
⎪
⎪ 1
⎪
⎩ [(μ + ε)(a + κb + ε)] + b ≤ α.
1 − (κ(μ + ε) + ε)
We will now prove that (8) holds with U = IB a (x̄) and V = IB b (ȳ). Fix x ∈
IB a (x̄). If Ψ(x) ∩ IB b (ȳ) = ∅, then the right side of (8) (with V = IB b (ȳ)) is +∞
and there is nothing more to prove. Let y ∈ Ψ(x) ∩ IB b (ȳ). From (9), there exists
z 1 ∈ F −1 (y) such that
(12) dX (z 1 , x) ≤ d(x, F −1 (y)) + ε ≤ κd(y, F (x)) + ε.
Furthermore,
dX (z 1 , x) ≤ d(x, F −1 (y)) + ε ≤ dX (x, x̄) + d(x̄, F −1 (y)) + ε
(13)
≤ a + κ d(y, F (x̄)) + ε ≤ a + κdY (y, ȳ) + ε ≤ a + κb + ε.
Then, using (11), we also have
(14) dX (z 1 , x̄) ≤ dX (z 1 , x) + dX (x, x̄) ≤ 2a + κb + ε ≤ α.
If z 1 = x, then x ∈ F −1 (Ψ(x)), which is the same as x ∈ Fix(F −1 Ψ). Then (8)
holds automatically, since its left side is 0. Let z 1 = x. Since dX (z 1 , x) > 0, there
exists y 1 ∈ Ψ(z 1 ) such that
dY (y 1 , y) ≤ d(y, Ψ(z 1 )) + εdX (z 1 , x).
From (10), remembering that y ∈ Ψ(x) ∩ IB b (ȳ), we obtain
(15) dY (y 1 , y) ≤ e(Ψ(x) ∩ IB b (ȳ), Ψ(z 1 )) + εdX (z 1 , x) ≤ (μ + ε)dX (z 1 , x).
Then, using (11) and (13),
(16) dY (y 1 , ȳ) ≤ dY (y, ȳ) + (μ + ε)dX (z 1 , x) ≤ b + (μ + ε)(a + κb + ε) ≤ α.
By induction, we construct sequences of points z k and y k , with z 0 = x and
0
y = y, such that, for k = 0, 1, 2, . . .,
z k+1 ∈ F −1 (y k ) with dX (z k+1 , z k ) ≤ (κ(μ + ε) + ε)k dX (z 1 , x),
(17)
y k+1 ∈ Ψ(z k+1 ) with dY (y k+1 , y k ) ≤ (μ + ε) dX (z k+1 , z k ).
Taking into account (15), we see that z 1 and y 1 satisfy (17) for k = 0. Suppose
that for some n ≥ 1 we have generated z 1 , z 2 , . . . , z n and y 1 , y 2 , . . . , y n satisfying
(17). Note that z 0 ∈ IB α (x̄) and y 0 ∈ IB α (ȳ). We will first show that z i ∈ IB α (x̄)
LYUSTERNIK-GRAVES THEOREM AND FIXED POINTS 527
and y i ∈ IB α (ȳ) for all i = 1, 2, . . . , n. Indeed for i = 1 this follows from (14) and
(16). Utilizing (17), for i ≥ 2 we have
i−1
dX (z i , x) ≤ j=0 dX (z j+1 , z j )
i−1
(18) 1
≤ (κ(μ + ε) + ε)j dX (z 1 , x) ≤ dX (z 1 , x),
j=0
1 − (κ(μ + ε) + ε)
Hence, by (11),
(μ + ε)(a + κb + ε)
dY (y i , ȳ) ≤ + b ≤ α.
1 − (κ(μ + ε) + ε)
If z n = z n−1 , then z n ∈ F −1 (y n−1 ) and y n−1 ∈ Ψ(z n ). Hence z n ∈ F −1 (Ψ(z n )).
Then, through (12) and (18), we get
d(x, Fix(F −1 Ψ)) ≤ dX (x, z n )
1
≤ dX (z 1 , x)
(19) 1 − (κ(μ + ε) + ε)
1
≤ κd(y, F (x)) + ε .
1 − (κ(μ + ε) + ε)
Since the left side of this inequality does not depend on the ε on the right and y
was arbitrarily chosen in Ψ(x) ∩ IB b (ȳ), by letting ε go to 0 we obtain (8) with
U = IB a (x̄) and V = IB b (ȳ). Set z n+1 = z n , y n+1 = y n .
Assume next that z n = z n−1 . Since z n ∈ F −1 (y n−1 ) ∩ IB α (x̄), from (9) there
exists z n+1 ∈ F −1 (y n ) such that
dX (z n+1 , z n ) ≤ d(z n , F −1 (y n )) + εdX (z n , z n−1 )
≤ κd(y n , F (z n )) + εdX (z n , z n−1 ) ≤ κdY (y n , y n−1 ) + εdX (z n , z n−1 ).
Then, by invoking the induction hypothesis (17) for k + 1 = n,
dX (z n+1 , z n ) ≤ κ(μ + ε)dX (z n , z n−1 ) + εdX (z n , z n−1 )
= (κ(μ + ε) + ε)dX (z n , z n−1 ) ≤ (κ(μ + ε) + ε)n dX (z 1 , x).
By repeating the argument in (19), if z n+1 = z n , we get (8). Then set y n+1 = y n .
If z n+1 = z n , since y n ∈ Ψ(z n ) ∩ IB α (ȳ), by (10), there exists y n+1 ∈ Ψ(z n+1 ) such
that
(20) dY (y n+1 , y n ) ≤ (μ + ε) dX (z n+1 , z n ).
This completes the induction step, and hence (17) holds for all k.
528 ASEN L. DONTCHEV AND HÉLÈNE FRANKOWSKA
We have already shown that if z k = z k−1 for some k, then (8) holds true. Suppose
now that z k+1 = z k for all k. By virtue of the inequality for z k+1 in (17), we see
for any natural n and m with m < n that
n−1 n−1
dX (z n , z m ) ≤ dX (z k+1 , z k ) ≤ (κ(μ + ε) + ε)k dX (z 1 , x)
k=m k=m
1
dX (z , x)
≤ (κ(μ + ε) + ε)m .
1 − (κ(μ + ε) + ε)
But then, from (19) and the above inequalities,
n−1 n−1
dY (y n , y m ) ≤ dY (y k+1 , y k ) ≤ (μ + ε) dX (z k+1 , z k )
k=m k=m
1
(μ + ε)dX (z , x)
≤ (κ(μ + ε) + ε)m .
1 − (κ(μ + ε) + ε)
This gives us that both sequences {z k } and {y k } satisfy the Cauchy condition and
z k ∈ IB α (x̄), y k ∈ IB α (ȳ) for all k ≥ 1. Suppose that gph F ∩ (IB α (x̄) × IB α (ȳ)) is
complete. Since (z k , y k−1 ) ∈ gph F ∩ (IB a (x̄) × IB α (ȳ)) the sequence is convergent
to, say, (z, u) ∈ gph F . Note that (z k , y k ) ∈ gph Ψ × (IB α (x̄) × IB α (ȳ)) which
by assumption is closed and then (z, u) ∈ gph Ψ. But then z ∈ Fix(F −1 Ψ). Now
let gph Ψ ∩ (IB α (x̄) × IB α (ȳ)) be complete. Since (z k , y k ) ∈ gph Ψ ∩ (IB α (x̄) ×
IB α (ȳ)) this sequence converges, and then, as in the preceding case, we conclude
that z ∈ Fix(F −1 Ψ). Finally, if gph F −1 Ψ ∩ (IB α (x̄) × IB α (x̄)) is complete, noting
that (z k , z k+1 ) ∈ gph F −1 Ψ ∩ (IB α (x̄) × IB α (x̄)), we obtain convergence of {z k } to
a point z ∈ Fix(F −1 Ψ).
Utilizing (12) and (18), we finally obtain
d(x, Fix(F −1 Ψ)) ≤ dX (z, x) = lim dX (z k , x)
k→∞
1
≤ dX (z 1 , x)
1 − (κ(μ + ε) + ε)
1
≤ κd(y, F (x)) + ε .
1 − (κ(μ + ε) + ε)
Since y was arbitrarily chosen in Ψ(x) ∩ IB b (ȳ), taking the limit as ε → 0 gives us
(8) with U = IB a (x̄) and V = IB b (ȳ).
If the assumptions for F and Ψ hold globally, then we can take any x̄ ∈ X and
ȳ ∈ Y , and choosing α = +∞ repeat the proof with a = b = +∞.
In the following theorem we present a slight extension of Theorem 5 in which the
reference points for the mappings F and Ψ are different and which can be proved
in a way parallel to the proof of Theorem 5 with only minor adjustments.
Theorem 6. Let (X, dX ), (Y, dY ) be metric spaces and α, κ and μ be positive
constants such that κμ < 1. Consider any two mappings F : X → → Y and Ψ :
→
X → Y and points (x̄, ȳ) ∈ gph F and (x̄, ȳ ) ∈ gph Ψ such that
¯ ¯
dY (ȳ, ȳ¯) ≤ α/2 and U := IB α (x̄) ∩ IB α (x̄
¯) = ∅.
Assume that either one of the sets gph F ∩ (U × IB α (ȳ)) and gph Ψ ∩ (U × IB α (ȳ¯)) is
complete while the other is closed or the set gph(F −1 Ψ) ∩ (U × U ) is complete. Also
assume that F is metrically regular at x̄ for ȳ with constant κ and neighborhoods
LYUSTERNIK-GRAVES THEOREM AND FIXED POINTS 529
¯
IB α (x̄) and IB α (ȳ), that is, (9) holds, and assume that Ψ is Aubin continuous at x̄
for ȳ¯ with constant μ and neighborhoods IB α (x̄ ¯) and IB α (ȳ¯); that is,
e(Ψ(x) ∩ IB α (ȳ¯), Ψ(x )) ≤ μdX (x, x ) for all (x, x ) ∈ IB α (x̄
¯).
Let a and b be any positive reals that satisfy
a + κb μa + μκb
(21) max + a, 2 +b < α.
1 − κμ 1 − κμ
Then for any x ∈ IB a (x̄) ∩ IB a (x̄
¯),
κ
(22) d(x, Fix(F −1 Ψ)) ≤ d(Ψ(x) ∩ IB b (ȳ) ∩ IB b (ȳ¯), F (x)).
1 − κμ
Next comes a version of Theorem 5 which covers the case when Ψ depends on a
parameter, which can again be proved by slightly modifying the proof of Theorem 5
in order to take into account the dependence of the parameter.
Theorem 7. Let (X, dX ), (Y, dY ) be metric spaces, let P be a metric space, and
let κ and μ be positive constants such that κμ < 1. Consider any two mappings F :
X→ → Y and Ψ : P × X → → Y , a point p̄ ∈ P and a point (x̄, ȳ) ∈ gph F ∩ gph Ψ(p̄, ·)
such that the following conditions hold:
(i) for any p near p̄, either one of the sets gph F and gph Ψ(p, ·) is locally com-
plete while the other is locally closed at (x̄, ȳ), or the set gph F −1 (Ψ(p, ·)) is locally
complete at (x̄, x̄);
(ii) F is metrically regular at x̄ for ȳ with reg(F ; x̄| ȳ) < κ;
(iii) Ψ is partially Aubin continuous with respect to x uniformly in p at (p̄, x̄)
x (Ψ; (p̄, x̄)| ȳ) < μ < 1/κ.
for ȳ with Lip
Then there exist neighborhoods Q of p̄, U of x̄ and V of ȳ such that for any p ∈ Q
and x ∈ U ,
κ
(23) d(x, Fix(F −1 (Ψ(p, ·)))) ≤ d(Ψ(p, x) ∩ V, F (x)).
1 − κμ
Now we will derive Theorems 1-4 stated in Section 1.
Proof I of Theorem 1. For a function g as in the statement, the local Lipschitz
continuity of g at x̄ implies that, for some r > 0, the graph of the restriction g|IB r (x̄)
is closed.
Pick α > 0, κ̃ > κ, μ̃ > μ such that κ̃μ̃ < 1,
d(x, F −1 (y)) ≤ κ̃d(y, F (x)) for all (x, y) ∈ IB α (x̄) × IB α (ȳ)
and
dY (g(x), g(x )) ≤ μ̃dX (x, x ) for all x, x ∈ IB α (x̄).
Without loss of generality, let g(x̄) = 0. Choose positive a and b such that (21)
holds with κ, μ replaced by κ̃ and μ̃ respectively and in addition aμ̃ < b. Let γ > 0
satisfy aμ̃ + γ ≤ b and γ ≤ α/2. Pick p ∈ IB γ (ȳ) and apply Theorem 6 for F and
Ψ(x) = −g(x) + p, with the so-chosen α, a and b, and with ȳ¯ = p and x̄ ¯ = x̄. Then
x ∈ Fix(F Ψ) iff x ∈ F (−g(x) + p), which is the same as x ∈ (F + g)−1 (p).
−1 −1
and
dY (−g(x) + p, p) ≤ dY (−g(x), −g(x̄)) ≤ μ̃a ≤ b,
the intersection (−g(x) + p) ∩ IB b (ȳ) ∩ IB b (p) is nonempty and hence consists of just
the point −g(x) + p. But then the right side of (24) equals d(p, (g + F )(x)). Since
κ̃ and μ̃ can be arbitrarily close to κ and μ, respectively, the proof is complete.
Proof II of Theorem 1. We will now derive Theorem 1 from the parametric The-
orem 7. As in the preceding proof, let g(x̄) = 0. We apply Theorem 7 with
Ψ(p, x) = −g(x) + p and p̄ = ȳ. Then Lip x (Ψ; (p̄, x̄)| ȳ) = lip(g; x̄). Choosing κ̃
and μ̃ as in Proof I, we obtain from Theorem 7 the existence of neighborhoods Q
of p̄, U of x̄ and V of ȳ such that (23) holds for any p ∈ Q and x ∈ U . Noting that
Fix(F −1 (−g(·) + p)) = (F + g)−1 (p) and, adjusting V if necessary so that for any
x ∈ U , p ∈ Q the intersection (−g(x) + p) ∩ V is nonempty and hence consists of
the single point −g(x) + p, we obtain from (23) that for any x ∈ U , p ∈ Q,
κ̃
d(x, (F + g)−1 (p)) ≤ d(p, (g + F )(x)).
1 − κ̃μ̃
Since κ̃ and μ̃ can be arbitrarily close to κ and μ, the proof follows.
Proof of Theorem 2. Apply Theorem 6 with Y = X, Ψ = Φ, α = 2c, μ = λ,
¯ and F the identity. Fix ε > 0 such that d(x̄, Φ(x̄)) + ε < α2 (1 − μ)
ȳ = ȳ¯ = x̄ = x̄
and define b = d(x̄, Φ(x̄)) + ε. Then Φ(x̄) ∩ IB b (x̄) = ∅. Pick any a > 0 satisfying
α
a + b < (1 − μ).
2
Then condition (21) holds and we can apply Theorem 6, obtaining for x = x̄ that
1
d(x̄, Fix(Φ)) ≤ inf dX (z, x̄).
1 − μ z∈Φ(x̄)∩IB b (x̄)
By the choice of b, the set Φ(x̄) ∩ IB b (x̄), over which the infimum is taken on the
right, is nonempty; hence
b α
d(x̄, Fix(Φ)) ≤ < = c.
1−μ 2
Therefore Φ has a fixed point in IB c (x̄).
Proof of Theorem 3. Fix z ∈ Y . We apply Theorem 5 to Ψ(·) = −G(·) + z. If
x ∈ X is such that F (x) + G(x) = ∅, then d(z, (F + G)(x)) = ∞ and therefore
d(x, Fix(F −1 Ψ)) ≤ 1−κμ
κ
d(z, G(x) + F (x)). If both F (x) and G(x) are nonempty,
applying Theorem 5 in the global case with Ψ(x) = −G(x) + z, we get
κ
d(x, Fix(F −1 Ψ)) ≤ inf d(z − v, F (x))
1 − κμ v∈G(x)
(25)
κ
= d(z, G(x) + F (x)).
1 − κμ
Note that x ∈ F −1 (−G(x) + z) is equivalent to the existence of y ∈ F (x) and
v ∈ −G(x) such that y = v + z, which, in turn, is equivalent to z = y − v ∈
F (x) − v ∈ F (x) + G(x). Hence Fix(F −1 (−G(·) − z)) = (F + G)−1 (z) and then,
by (25),
κ
d(x, (F + G)−1 (z)) ≤ d(z, (F + G)(x)),
1 − κμ
which completes the proof.
LYUSTERNIK-GRAVES THEOREM AND FIXED POINTS 531
Proof of Theorem 4. From the global version of Theorem 5 with κ = γ −1 , for any
x ∈ X we have
1
d(x, Fix(F −1 (Ψ))) ≤ d(Ψ(x), F (x)).
γ−μ
It remains to note that for any ε > 0 there is a point ξ ∈ Fix(F −1 Ψ) with dX (ξ, x) ≤
d(x, Fix(F −1 (Ψ))) + ε, that is F (ξ) ∩ Ψ(ξ) = ∅, and hence ξ satisfies (7).
One may argue that it is possible to obtain the local results displayed above from
the corresponding global ones by using a truncation of the mapping in question.
Specifically, for an Aubin continuous mapping, the question is whether one can find
a submapping of it which would be Lipschitz continuous. First, observe that if a
mapping F is metrically regular, its restriction
F (x) if x ∈ IB α (x̄),
F|IB α (x̄) (x) =
∅ otherwise
may be not metrically regular on IB α (x̄). Then, e.g., the global linear openness
assumption of Theorem 4 may be lost by taking a restriction.
In [4] it was shown that when a mapping Ψ acting in finite-dimensional spaces
is convex-valued and is Aubin continuous at x̄ for ȳ with constant κ, then, for
some positive α and β, the truncated mapping IB α (x̄) x → Ψ(x) ∩ IB β (ȳ) is
Lipschitz continuous in IB α (x̄) with Lipschitz constant larger than κ. In particular,
taking truncation may increase the Lipschitz constant and lead to a violation of
the inequality κμ < 1.
The following example shows that in general Aubin continuous mappings may
not be Lipschitz continuous after truncation. This example illustrates the essential
difference between the global and the local cases.
and let
∞
K = {(0, 0)} ∪ Ki .
i=1
→ IR2 defined by
Then K is closed. Consider Ψ : IR →
1
x, x ∪ (x(0, 1) + K) for x > 0,
Ψ(x) =
K otherwise.
and to
(τ − γ)θ
ū = (x, y) and δ= .
1+λ
Let uδ = (t, z) be as in the Ekeland principle. Then from (26)
(τ − γ)θ (τ − γ)hθ 2
z + g(t) − u + d((t, z), (x, y)) ≤ y + g(x) − u = .
1+λ 1+λ
In particular, dX (t, x) ≤ hθ. Hence, by (29),
(31) dX (t, x̄) ≤ α + hθ < a
and also
z − ȳ ≤ z − y + β ≤ z + g(t) − u + u − g(t) − y + β
τ −γ
≤ 2 h + γh + β < (2τ − γ)h + β < 2τ h + β < b.
1+λ
Moreover, from (27), for any (x , y ) ∈ K,
(τ − γ)θ
(33) z + g(t) − u ≤ y + g(x ) − u + d((t, z), (x , y )).
1+λ
We will now show that u = z + g(t). Indeed, assume z + g(t) − u > 0, choose
ε ∈ (0, τ ) such that
ε
(34) 1− >θ
τ −γ
and denote
τ −ε
w=− (z + g(t) − u).
z + g(t) − u
Then, from (28), taking into account (31) and (32),
z + hw ∈ F (Bh (t)).
This implies that
z + hw ∈ F (x̃) for some x̃ ∈ IB h (t).
Note that, from (29) and (31),
dX (x̃, x̄) ≤ dX (t, x̄) + h < a.
Hence, applying (33) with x = x̃ and y = z + hw,
(τ − γ)θ λ
z + g(t) − u ≤ z + hw + g(x̃) − u + (h + h w)
1+λ τ
≤ z + hw + g(t) − u + γh + hθ(τ − γ)
(τ − ε)h
= z + g(t) − u 1 − + γh + hθ(τ − γ).
z + g(t) − u
This gives us
(τ − ε)h < γh + θ(τ − γ)h;
534 ASEN L. DONTCHEV AND HÉLÈNE FRANKOWSKA
hence τ − γ − ε < θ(τ − γ), which contradicts the choice of ε in (34). Thus we have
u = z + g(t). Therefore F + g is metrically regular at x̄ for ȳ + g(x̄) with constant
(τ − γ)−1 . Since 1/τ can be arbitrarily close to κ and γ can be arbitrarily close to
μ, we obtain the desired result.
References
[1] A. V. Arutyunov, Covering mappings in metric spaces, and fixed points, Dokl. Akad. Nauk
416 (2007) 151–155 (Russian). MR2450913
[2] J-P. Aubin, Lipschitz behavior of solutions to convex minimization problems, Mathematics
of Oper. Res. 9 (1984) 87–111. MR736641 (86f:90119)
[3] S. Banach, Théorie des Opérations Linéaires, Monografie Matematyczne, Warszawa, 1932.
English translation by North Holland, Amsterdam, 1987. MR0880204 (88a:01065)
[4] D. N. Bessis, Yu. S. Ledyaev and R. B. Vinter, Dualization of the Euler and Hamiltonian
inclusions, Nonlinear Analysis 43 (2001) 861–882. MR1813512 (2002c:49036)
[5] A. V. Dmitruk, A. A. Milyutin and N. P. Osmolovskiı̆, Ljusternik’s theorem and the theory
of the extremum, Uspekhi Mat. Nauk 35, no. 6 (216) (1980) 11–46 (Russian). MR601755
(82c:58010)
[6] A. V. Dmitruk, On a nonlocal metric regularity of nonlinear operators, Control Cybernet.
34 (2005) 723–746. MR2208969 (2006k:49046)
[7] A. V. Dmitruk and A. Y. Kruger, Metric regularity and systems of generalized equations,
J. Math. Anal. Appl. 342 (2008) 864–873. MR2445245 (2009f:49017)
[8] A. L. Dontchev and W. W. Hager, An inverse mapping theorem for set-valued maps, Pro-
ceedings of Amer. Math. Soc. 121 (1994) 481–489. MR1215027 (94h:58020)
[9] A. L. Dontchev and R. T. Rockafellar, Implicit Functions and Solution Mappings, Springer
Mathematics Monographs, Springer, Dordrecht, 2009. MR2515104
[10] A. L. Dontchev, A. S. Lewis and R. T. Rockafellar, The radius of metric regularity, Trans.
Amer. Math. Soc. 355 (2003) 493–517. MR1932710 (2003i:49026)
[11] H. Frankowska, Some inverse mapping theorems, Annales de Inst. H. Poincaré Anal. Non
Linéaire, Section C, 7 (1990) 183–234. MR1065873 (91j:49020)
[12] H. Frankowska, Conical inverse mapping theorems, Bull. Austral. Math. Soc. 45 (1992) 53–
60. MR1147244 (92k:49028)
[13] L. M. Graves, Some mapping theorems, Duke Math. Journal 17 (1950) 111–114. MR0035398
(11:729e)
[14] A. D. Ioffe, Metric regularity and subdifferential calculus, Uspekhi Mat. Nauk 55 (2000),
no. 3 (333), 103–162; English translation in Math. Surveys 55 (2000), 501–558. MR1777352
(2001j:90002)
[15] A. D. Ioffe, Towards variational analysis in metric spaces: metric regularity and fixed points,
Math. Program., Ser. B 123 (2010) 241–252. MR2577330
[16] L. A. Lyusternik, On the conditional extrema of functionals, Mat. Sbornik 41 (1934) 390–
401.
[17] S. B. Nadler, Jr., Multi-valued contraction mapping, Pacific J. Math. 30 (1969) 475–488.
MR0254828 (40:8035)
Mathematical Reviews and the University of Michigan, Ann Arbor, Michigan 48109.
On leave from the Institute of Mathematics, Bulgarian Academy of Sciences, Sofia,
Bulgaria
E-mail address: ald@ams.org