Chapter 3

Characterization of best
approximations

In this chapter we study properties which characterite solutions of the approximation

problem. There is a big difference in the treatment of this question in Hilbert and Banach
spaces. A harmonization is possible by considering semi-inner products.
The material presented is standard and can be found in textbooks on approximation
theory and convex analysis; see for instance Cheney [6], Deutsch [8], Holmes [11] and
Singer [17]. For the material concerning the semi-inner products we refer to Dragomir [9],
Giles [10], Penot and Ratsimahalo [15] and Rosca [16].

3.1 Characterization of best approximations in Hilbert

spaces
Theorem 3.1 (Kolmogorov’s criterion). Let H be a Hilbert space and let C be a nonempty
closed convex subset of H . Then for x ∈ H and w ∈ C the following conditions are
equivalent:

(a) w = PC (x) .
(b) �x − w|y − w� ≤ 0 for all y ∈ C .
(c) �x − w�2 + �y − w�2 ≤ �x − y�2 for all y ∈ C .

Proof:
(a) =⇒ (b) Let y ∈ C and t ∈ (0, 1] . Then ty + (1 − t)w ∈ C and we have

0 ≤ �x − (ty + (1 − t)w)�2 − �x − w�2 = −2t�x − w|y − w� + t2 �y − w�2 .

Dividing by t and using t ↓ 0 we obtain (b) .

(b) =⇒ (c) Using (b) we obtain

�x−y�2 = �x−w+w−y�2 = �x−w�2 +�w−y�2 +2�x−w|w−y� ≥ �x−w�2 +�y−w�2 .

(c) =⇒ (a) Obviously. �

47
 Kolmogorov’s criterion in Hilbert spaces1 , i.e. the equivalence (a) ⇐⇒ (b) is saying
that the best approximation w to x is such that x − w forms an obtuse angle with all
direction y − w from w into the set C . Another way of looking at this criterion is to
define, for x ∈ H\C the hyperplane

Hx := {y ∈ H : �x − PC (x)|y� = 0} .

One then notes that the translate Hx +PC (x) supports the convex set C at PC (x) . Lateron,
we will present Kolmogorov’s criterion also for Banach spaces. Here the formulation and
the proof is much more delicate.
The equivalence (a) ⇐⇒ (c) leads to the so called strong uniqueness of PC (x):

�x − PC (x)�2 ≤ �x − y�2 − �y − PC (x)�2 for all y ∈ C .

Theorem 3.2. Let H be a Hilbert space and let C be a nonempty closed convex subset of
H . Then the metric projection PC has the following properties:

(1) PC (x) is a singleton for each x ∈ H .

(2) �PC (x) − PC (x � )�2 + �(I − PC )(x) − (I − PC )(x � )�2 ≤ �x − x � �2 for all x, x � ∈ X .

(3) PC : X −→ C is nonexpansive.

(4) PC : X −→ C is uniformly continuous.

(5) PC is idempotent, i.e. PC ◦ PC = PC .

(6) PC is positively homogeneous, i.e. PC (tx) = tPC (x) for all t > 0, x ∈ H, if C is a
cone.

Proof:
Ad (1) C is a Chebyshev set; see Theorem 2.17.
Ad (2) Let w = PC (x), w � = PC (x � ) . Then

�x − x � �2 = �PC (x) − PC (x � ) + (I − PC )(x) − (I − PC )(x � )�2

= �PC (x) − PC (x � )�2 + �(I − PC )(x) − (I − PC )(x � )�2
+2�PC (x) − PC (x � )|(I − PC )(x) − (I − PC )(x � )�
≥ �PC (x) − PC (x � )�2 + �(I − PC )(x) − (I − PC )(x � )�2

where we have used (b) in Theorem 3.1.

Ad (3) Follows from (2) .
Ad (4) Follows from the nonexpansivity; see (3) .
Ad (5) Follows from the fact that PC is the identity on C .
Ad (6) Let w = PC (x) . Choose t > 0 and y ∈ C . Then tw, t−1 y ∈ C and we obtain

�tx − tw� = t�x − w� ≤ t�x − t−1 y� = �tx − y� .

This shows tPC (x) = tw = PC (tx) . �

1
Historically, it was developed mainly for the best approximation in the space of continuous functions
in an interval. The underlying Banach space C[a, b] is not reflexive and not strictly convex.

48
 Theorem 3.3. Let H be a Hilbert space and let U be a closed subspace of H . Then the
metric projection PU das the following properties:

(1) PU (x) is a singleton for each x ∈ H .

(2) PU (x) = x for each x ∈ U .

(3) PU is idempotent, i.e. PU = PU ◦ PU .

(4) For each x ∈ H we have: w = PU (x) if and only if �x − w|u� = 0 for all u ∈ U .

(5) x − PU (x) ∈ U⊥ for each x ∈ X .

(6) PU is selfadjoint, i.e. �PU (x)|y� = �x|PU (y)� for all x, y ∈ H .

(7) �PU (x)� ≤ �x� for each x ∈ H .

(8) PU is a linear mapping.

(9) PU is a linear nonexpansive.

(10) �PU (x)|x� ≥ 0 for every x ∈ H .

(11) If C is a closed affine subset of H then PC is affine and weakly continuous.

Proof:
Ad (1), (2), (3) are clear because U is Chebyshev set; see Theorem 2.17.
Ad (4) From (b) in Theorem 3.1 we conclude �x − PU (x), u� for all u ∈ U .
Ad (5) Follows from (4)
Ad (6) Using (5) we obtain

�PU (x)|y� = �PU (x)|y − PU (y) + PU (y)� = �PU (x)|PU (y)� .

By symmetry reasons we obtain �x|PU (y)� = �PU (x)|PU (y) .

Ad (7) We have for x ∈ H due to (5)

�x�2 = �PU (x) + (x − PU (x))�2 = �PU (x)�2 + �x − PU (x)�2 ≥ �PU (x)�2 .

Hence �x� ≥ �PU (x)� .

Ad (8) Let x, y ∈ H, a, b ∈ R . Then by (5) x − PU (x), y − PU (y) ∈ U⊥ . Hence

ax + by − (aPU (x) + bPU (y)) = a(x − PU (x)) + b(y − PU (y)) ∈ U⊥ .

Therefore w := aPU (x) + bPU (y) = PU (ax + by) ∈ U by (4) .

Ad (9) Follows from (7) and (8) .
Ad (10) Follows with (6) as follows:

�PU (x)|x� = �PU (PU (x))|x� = �PU (x)|PU (x)� = �PU (x)� ≥ 0 .

Ad (11) Let C = x0 + U with a closed linear subspace U . Let x ∈ H . Then

�(x − x0 ) − u� = �x − (x0 + u)� for all u ∈ U .

49
This implies x0 + PU (x − x0 ) = Px0 +U (x) . Let

L : H � x �−→ Px0 +U (x) − Px0 +U (θ) = x0 + PU (x − x0 ) − Px0 +U (θ) ∈ H .

Since PU is a linear continous mapping (see (8),(9)), L is linear and continuous too.
Let x ∈ H and let x = w − limn xn . Then

lim�xn |L∗ y� = �x|L∗ y� , lim�Lxn |y� = �Lx|y� for all y ∈ H .

n n

This shows w − limn Lxn = Lx and therefore w − limn Px0 +U (xn ) = Px0 +U (x) . �

Theorem 3.4 (Reduction principle). Let H be Hilbert space, let C be a convex subset and
let U ⊂ H be a linear subspace which is a Chebyshev set. Suppose that C ⊂ U . Then we
have:

(a) PC ◦ PU = PC = PU ◦ PC .
(b) dist(x, C)2 = dist(x, U)2 + dist(PU (x), C)2 for all x ∈ H .

Proof:
Notice that PC (x) is empty or a singleton due to the fact that C is a convex subset of a
Hilbert space. Moreover PU (x) is a singleton for each x ∈ H since U is a Chebyshev set.
Ad (a) PU ◦ PC = PC since C ⊂ U . Let x ∈ H, y ∈ C . Then y ∈ U and x − PU (x) ∈ U⊥
due to (5) in Theorem 3.3. Therefore

�x − y�2 = �x − PU (x)�2 + �PU (x) − y�2 . (3.1)

We conclude that y minimizes u �−→ �x − u�2 iff y minimizes v �−→ �PU (x) − v�2 .
This shows that PC (x) exists iff PC (PU (x)) exists and PC (x) = PC (PU (x)) holds true. We
obtain that PC (x) �= ∅ iff PC (PU (x)) �= ∅ . This implies (a) .
Ad (b) Follows from (3.1). �

The interpretation of the reduction principle is the following: To find the best approx-
imation to x from C we first project x onto U and then project PU (x) onto C ; the result
is PC (x) .

Corollary 3.5. Let H be a Hilbert space, let C ⊂ X be a nonempty closed convex subset
of H . Then PC : X −→ C satifies

�PC (x) − PC (y)�2 ≤ �PC (x) − PC (y)|x − y� , x, y ∈ H . (3.2)

Proof:
Let x, y ∈ H . From Kolmogorov’s criterion we have:

�x − PC (x)|PC (y) − PC (x)� ≤ 0 , �y − PC (y)|PC (x) − PC (y)� ≤ 0 .

Adding these two inequalites implies (3.2). �

The property (3.2) of metric projections in Hilbert spaces is called firmly nonexpan-
sivity. We shall come back to this property in a general context in Chapter 7. Notice
that we obtain from (3.2) the nonexpansivity of PC .

50
3.2 Duality mapping
Let X be a Banach space. We set:

J(x) := {λ ∈ X ∗ : �λ, x� = �λ��x�, �λ� = �x�} , x ∈ X .

In general, J is a set-valued mapping; we write therefore J : X ⇒ X ∗ . This mapping

is called the (normalized) duality mapping2 . The duality map plays a crucial role
in solving (nonlinear) problems in the setting of Banach spaces, due to the lack of an
inner product in a Banach space. We will see that this map is also helpful in describing
properties of the metric projections and the geometry of Banach spaces. Let us collect
some facts concerning the duality map.
Lemma 3.6. Let X be a Banach space. Then we have:
(1) J(x) is nonempty for each x ∈ X and J(θ) = {θ} .

(2) J is a homogenous mapping, i.e. J(ax) = aJ(x) for all x ∈ X , a ≥ 0 .

(3) J(x) is a convex set for all x ∈ X .

(4) J(x) is a weak∗ compact set.

(5) The duality map is monotone in the following sense:

�λ − µ, x − y� ≥ 0 for x, y ∈ X , λ ∈ J(x), µ ∈ J(y) . (3.3)

(6) The duality map is surjective if and only if every functional λ ∈ X ∗ attains its maxi-
mum on S1 .

(7) The range of J is dense in X ∗ .

(8) J is surjective if and only if X is reflexive.

(9) If X is a Hilbert space then the duality map is the identity.

Proof:
Ad (1) Evidentely, J(θ) = {θ} . Let x ∈ S1 . Then there exists λ ∈ X ∗ with �λ� = 1 such
hat �λ, x� = �x� = �x�|λ� (continuation theorem of Hahn-Banach/see the preliminaries
in the preface). This shows that λ ∈ J(x) . For an arbitrary x ∈ X \{θ} apply the reasoning
above to x�x�−1 .
Ad (2) This is easily verified by distinguishing the case a = 0, a �= 0 . . Notice a∅ =
∅, a ∈ R .
Ad (3) For x = θ nothing has to be proved. Let x �= θ, λ, µ ∈ J(x) and σ := aλ + (1 − a)µ
with a ∈ [0, 1] . Then �x�−1 �σ� ≤ 1 and therefore

�x� ≥ ��x�−1 σ, x� = a�x� + (1 − a)�x� = �x� .

This shows �x�2 = �σ, x� and �σ� = �x� .

Ad (4) J(x) is a closed subset of the weak∗ compact of the ball with radius �x� and
2
The concept of a duality mapping in a Banach space X was introduced by Beurling and Livingston
[3].

51
therefore weak∗ compact (Theorem of Banach-Alaoglu).
Ad (5) Let x, y ∈ X and λ ∈ J(x), µ ∈ J(y) . Then
�λ − µ, x − y� ≥ = �λ, x� − �λ, y� − �µ, x� + �µ, y�
≥ �x�2 + �y�2 − (�λ��y� − �µ��x�) = (�x� − �y�)2 (3.4)
Now, we read off the monotonicity.
Ad (6) Let λ ∈ X ∗ . Then there exists x ∈ X with �λ, x� = �λ� : see the preliminaries in
the preface. Let y := �λ�x . Then λ ∈ J(y) since �λ� = �y� and
�λ, y� = �λ, �λ�x� = �λ��λ, x� = �λ�2 = �y�2 .
Let λ ∈ X ∗ and let y ∈ X ∗ with J(y) = λ . If y = θ then λ = θ and nothing has to be
proved. Let y �= θ . Then
�λ, y�y�−1 � = �y�−1 �λ, y� = �y� = �λ� ,
which implies that λ attains its norm in �y�−1 y ∈ S1 .
Ad (7) This is the theorem of Bishop-Phelps; see [7].
Ad (8) This follows from (5) by using the Theorem of James which says that a Banach
space is reflexive iff every functional λ ∈ X ∗ attains its maximum on S1 .
Ad (9) Let λ ∈ J(x) . By the Riesz-representation theorem we have y ∈ X with �y|x� =
�λ, x� = �x�2 , �λ� = �y� = �x� . By the parallelogram identity it is easily seen that
y = x. �
Example 3.7. Let for 1 < p < ∞ and let X := Lp (Ω) be the Banach space of p-integrable
functions on the measurable set Ω ⊂ Rn . In this case, Lp (Ω)∗ = Lq (Ω) with 1/p+1/q = 1
and one has
J(x) = {�x�p−2 |x|p−1 sign(x)} , x �= θ .
Later on, we find a way for computing the result above. �
Lemma 3.8. Let X be a Banach spoace with duality mapping JX . Let x ∈ X and (xn )n∈N
such that
lim�JX (xn ) − JX (x), xn − x� = 0 .
n
n
Then x = limn x .
Proof:
See [5]. �
Lemma 3.9. Let X be a Banach spoace with duality mapping JX . Let (sn )n∈N be a non-
increasing sequence of positive numbers with lim sn = ∞ and let (xn )n∈N . Assume
�sn JX (xn ) − sm JX (x), xn − xm � =≤ 0 , m, n ∈ N .
Then (�xn �)n∈N is a nondecreasing sequence. If it is bounded then (xn )n∈N is convergent.
Proof:
See [4]. �

As we know, the duality mapping J on a Banach space is set-valued in general. This

is more or less a consequence of the continuation theorem of Hahn-Banach. To illustrate
this consider the following example.

52
Example 3.10. Consider X := R2 endowed with the l1 -norm and let U be the one-
dimensional subspace R1 × {θ} . Then µ : U � (x, y) �−→ x ∈ R is a linear continuous
functional on U with �µ� = 1 = �µ, (1, 0)�. Each continuation of µ to a continuous
functional λ on R2 has the representation λ : R2 � (x, y) �−→ αx + βy ∈ R . It is easy
to verify that such a functional λ has the property λ|U = µ and �λ� = 1 if |β| ≤ 1 ; notice
that the dual norm on R2 is the l∞ -norm. Thus, for each such λ we have

λ ∈ X∗ , �λ, (1, 0)� = 1 = �(1, 0)�2 , �λ� = 1 = �(1, 0)� ,

so that λ ∈ J((1, 0)) . �

Example 3.11. Let K be a compact topological space and let C(K) be the space of con-
tinuous functions f from K into R . C(K) is a Banach space under the supremum norm
� · �∞ : �f�∞ := supξ∈K |f(ξ)| . Its dual space C(K)∗ can be represented in different ways.
In the case K = [0, 1] the representation as the space BV[0, 1] of functions with bounded
variations is approriate. Each function g of bounded variation leads to a Stieltjes-integral
on C(K) the elements of C(K)∗ have the following presentation:
�
∗
∀ λ ∈ C(K) ∃ g ∈ BV[0, 1] (�λ, f� = fdg) .
K

In the general case, we have to consider the set of signed Borel measure µ on K .
Each such measure µ has a decomposition µ = µ+ − µ− where µ+ , µ− are nonnegative
measures; |µ| := µ+ + µ− is called the total variation of µ . Now, the dual space C(K)∗
can be described as follows: For each λ ∈ C(K)∗ there exists a uniquely determined finite
and regular signed Borel measure with
�
�λ, f� = fdµ , f ∈ C(K) .
K

Finite“ means that the range of µ is contained in R and regular“ means that the measure
” ”
µ(A) of each Borel subset A of K can be approximated from interior by the measure of
open subsets. The norm in C(K)∗ under this representation is given by the total variation.
Examples of functionals in C(K)∗ are Dirac’s point measures δξ , ξ ∈ K : �δξ , f� = f(ξ) .
Now the duality map J on C(K) is given as follows:
�
J(f) = {µ : µ finite, regular, signed measure on K, fdµ = |µ|�f�∞ , |µ| = �f�∞ } .
K

If we choose the function f ≡ 1 then J(f) is the set of probability measures on the Borel
sigma algebra of K . �

The property (4) in Lemma 3.6 has a nice consequence concerning the continuity of J
on the unit sphere.

Lemma 3.12. Let X be a Banach space with the duality mapping J and let x ∈ S1 with
#J(x) = 1 . Suppose we have sequences (xn )n∈N , (λn )n∈N with �xn � = 1, n ∈ N, lim xn =
x, λn ∈ J(xn ), n ∈ N . Then w∗ − limn λn = J(x) .

53
Proof:
We have �λn � = �xn � = �λn , xn � = 1, n ∈ N . Since the unit ball in X ∗ is σ(X∗ , X) com-
pact (Theorem of Alaoglu-Banach) the sequence (λn )n∈N posesses σ(X ∗ , X ) convergent
subsequences. Let (λnk )k∈N be such a subsequence of (λn )n∈N ; λ := w∗ − limk λnl . Then
�λ� ≤ 1 and

|�λ, x� − 1| = |�λ, x� − �λnk , xnk �|

≤ |�λ, x� − �λnk , x�| + |�λnk , x� − �λnk , xnk �|
≤ |�λ, x� − �λnk , x�| + �xnk − x�

By taking k → ∞ this implies �λ, x� = 1, �λ� = 1 and therefore λ ∈ J(x) . Since #J(x) = 1
we conclude that λ = w∗ − limn λn = J(x) .
�

In several cases, when one works in specific Banach spaces like lp , Lp (Ω), W 1,p (Ω), 1 <
p < ∞, a generalization of the normalized duality map is appropriate. In a more general
consideration, one defines a duality mapping as

Jh : X � x �−→ {λ ∈ X∗ : �λ, x� = �x��h(x)�, �λ� = �h(x)�} ⊂ X∗ .

Here, h is called a gauge-function. It is a non-decreasing map h : [0, ∞) −→ [0, ∞)

and has the properties h(0) = 0, h(s) > 0 for s > 0 and lims→∞ h(s) = ∞ . We set

G := {h : [0, ∞) −→ [0, ∞) : h a gauge-function} .

The relation between the normalized duality map and the generalized duality map Jh is
easily seen to be
h(�x�)
Jh (x) = J(x) , x ∈ X .
�x�
Duality mappings will be used as a main tool for studying the properties of the metric
projection. We restrict ourselves to a consideration using a normalized duality mapping
only.
Lemma 3.13. Let X be a Banach space such that the dual space X ∗ is a strictly convex
Banach space. Then the duality map J is a single-valued mapping.
Proof:
Let x ∈ X . If x = θ then J(x) = {θ} . Suppose that x �= θ . Let λ, µ ∈ J(x) . Then
�λ� = �µ� = �x� and

2�λ��x� = 2�x�2 = �λ + µ, x� ≤ �λ + µ��x� .

From this follows 2�λ� ≤ �λ + µ� and therefore by symmetry �λ� + �µ� ≤ �λ + µ� . This
implies �λ� + �µ� = �λ + µ� . By the strict convexity of X ∗ we get λ = µ ; see Lemma
2.14. �
Lemma 3.14. Let X be a Banach space with duality mapping J . Then the following
conditions are equivalent:
(a) X is strictly convex.

54
(b) For all λ ∈ X ∗ \{θ} there exists at most one x ∈ S1 with �λ� = �λ, x� .
(c) J is strictly monotone, i.e.

�λ − µ, x − y� > 0 for all x, y ∈ X , x �= y, λ ∈ J(x), µ ∈ J(y) .

Proof:
(a) =⇒ (b) Let λ ∈ X ∗ \{θ} . We may assume �λ� = 1 . Let x, y ∈ X with

1 = �x� = �y�, x �= y, �λ, x� = �λ, y� = 1 .

Then �x + y� ≥ �λ, x + y� = 2 and hence X cannot be strictly convex.

(b) =⇒ (a) Let x, y ∈ S1 , x �= y, �x + y� = 2 . Then there exists λ ∈ X ∗ with

�λ� = 1 , 2 = �x + y� = �λ, x + y� = �λ, x� + �λ, y� .

This implies �λ, x� = �λ, y� = 1 and condition (b) is violated.

(a) =⇒ (c) Assume that J is not monotone. Then there exist x, y ∈ X , x �= y,
λ ∈ J(x), µ ∈ J(y) with �λ − µ, x − y� = 0 . Then from (3.4) �y� = �x� = �λ� = �µ� .
Since x �= y this implies �x� = �y� �= 0 . From (3.4) (first line) we obtain

l := �x�2 − �x��λ, y�y�−1 � = �µ, x�x�−1 � − �y�2 =: r .

Clearly, l ≥ 0, r ≤ 0 . If l �= 0 then we see that r > 0, a case which cannot occur. If l = 0

then λ attains its norm on two different points, namely x, y, and hence X is not strictly
convex (see Lemma (3.14)).
(c) =⇒ (a) Assume that X is not strictly convex. Then there exist λ ∈ X ∗ \{θ} and
x, y ∈ X , x �= y, with

�λ, x� = �λ� = �λ, y� , �x� = �y� = 1 .

We may assume �λ� = 1 . Then λ ∈ J(x), λ ∈ J(y) and �λ − λ, x − y� = 0 . Therefore J is

not strictly monotone. �
Lemma 3.15. Let X be reflexive and suppose that the spaces X and X ∗ are strictly
convex. Let J be the duality map of X and let J∗ be the duality map of X ∗ . Then

J : S1,X −→ S1,X ∗ is bijective with J∗ ◦ J = I, J ◦ J∗ = I .

Proof:
Let λ ∈ X ∗ . If λ = θ then λ is in the range of J since J(θ) = θ . Assume now λ �= θ and
consider µ := λ�λ�−1 . Since X is reflexive there exists z ∈ X with �z� = 1, �λ, z� = 1 .
Then �µ� = �z�, �µ, z� = �µ��z� and we conclude that µ ∈ J(z) . Since X ∗ is strictly
convex J is single-valued and we have J(z) = µ . With x := �λ�z we have J(x) = J(�λ�z) =
�λ�J(z) = �λ�µ = λ . Now, the proof of surjectivity is complete.
Assume: J is not injective. Then there exist u, v ∈ X , u �= v, with λ := J(u) = J(v) . We
have �λ� = �u�, �λ, u� = �λ��u�, �λ� = �v�, �λ, v� = �λ��v� . We set with t ∈ (0, 1)
x := tu + (1 − t)v . Then �λ, x� = �λ�(t�u� + (1 − t)�v�) and therefore �λ�(t�u� + (1 −
t)�v�) = |�λ, x�| ≤ �λ��x�, t�u� + (1 − t)�v� ≤ �x� . Obviously, �x� ≤ t�u� + (1 − t)�v� .
Thus, we have
�tu + (1 − t)v� = t�u� + (1 − t)�v� .

55
We set w := u�u�−1 , z : v�v�−1 . Since �u� = �v� = λ� we have

�w� = �z� = 1, �tw + (1 − t)z� = t�w� + (1 − t)�z�, w �= z .

This shows that the segment [w, z] is contained in S1,X ∗ . This is contradicting the as-
sumption that X ∗ is strictly convex.
Now we have that J is bijective. From the definition of J and J∗ we conclude J∗ ◦ J =
I, J ◦ J∗ = I . �

3.3 Smoothness
Let us begin with a helpful result concerning convex real functions.

Lemma 3.16. Let X be a vector space and let g : X −→ R be a convex function which
satisfies g(θ) = 0 . Let y ∈ X . Then the mapping (0, ∞) � t �−→ g(ty)/t ∈ R is
monotone non-decreasing.

Proof:
Let 0 < s ≤ t and set a := s/t ∈ [0, 1] . Then

g(sy)/s = g(aty + (1 − a)θ)/s ≤ (ag(ty) + (1 − a)g(θ))/s = g(ty)/t .

Consider in a normed space X the norm-mapping

ν : X � x �−→ �x� ∈ R , x ∈ X ,

and the associated right-derivative and left-derivative of ν (under the assumption

that they exist):

�x + ty� − �x� �x + ty� − �x�

ν+� (x, y) := lim , ν−� (x, y) := lim , x, y ∈ X .
t↓0 t t↑0 t

Lemma 3.17. Let X be a Banach space with norm-mapping ν . Let x ∈ X \{θ} . Then we
have:

(1) ν+� (x, y), ν−� (x, y) exist for all y ∈ X .

�x + ty� − �x�
(2) ν−� (x, y) = lim t = −ν+� (x, −y) for all y ∈ X .
t↑0

(3) ν+� (x, ay) = aν+� (x, y), ν−� (x, ay) = aν−� (x, y) for all y ∈ X , a ≥ 0 .

(4) ν+� (x, u + w) ≤ ν+� (x, u) + ν+� (x, w), ν−� (x, u + w) ≥ ν−� (x, u) + ν−� (x, w) , u, w ∈ X .

(5) |ν+� (x, y)| ≤ �y�, |ν−� (x, y)| ≤ �y� for all y ∈ X .

(6) ν+� (x, x) = ν−� (x, x) = 1 .

56
Proof:
Ad (1) Let x, y ∈ X . We apply Lemma 3.16 to g(y) := �x + y� − �x� and obtain that
�x + ty� − �x�
dx,y : R\{0} � t �−→ t is monotone non-decreasing in (0, ∞) . In the same
way, dx,−y is monotone non-decreasing in (0, ∞) . This implies by a simple argumentation
that dx,y is monotone non-decreasing in R\{0} . We conclude that
�x + ty� − �x�
ν+� (x, y) = lim
t↓0 t
exists. A similar argumentation shows that ν−� (x, y) exists.
Ad (2) See the proof of (1) .
Ad (3), (4) This is easy to show.
Ad (5) Follows from
−�y� ≤ ν−� (x, y) ≤ ν+� (x, y) ≤ �y�, y ∈ X .
Ad (6) Obviously. �
Definition 3.18. Let X be a Banach space with norm mapping ν . Then the norm in X
is called Gateaux differentiable in x if ν+� (x, y) = ν−� (x, y) for all y ∈ X . If the norm
is Gateaux differentiable in x we set ν � (x, y) := ν+� (x, y), y ∈ X , and we call ν � (x, y) the
Gateaux derivative of the norm in x in direction y . �
Clearly, a norm cannot be Gateaux differentiable in θ .
Lemma 3.19. Let X be a Banach space with norm-mapping ν and let the norm be
Gateaux differentiable in x ∈ X \{θ} . Then we have:
(1) ν � (x, ·) is linear.
(2) �ν � (x, y)� ≤ �y� for all y ∈ X .
(3) ν � (x, ·) ∈ X∗ , ν � (x, x) = �x� and �ν � (x, ·)� = 1 .
Proof:
Ad (1) Follows from (2), (3), (4) in Lemma 3.17.
Ad (2) See (5) in Lemma 3.17
Ad (3) We know from (1), (2) that ν � (x, ·) ∈ X ∗ and �ν � (x, ·)� ≤ 1 . Since ν � (x, x) = �x�
we have �ν � (x, ·)� = 1 . �
Lemma 3.20. Let X be a Banach space with norm-mapping ν and let the norm be
Gateaux differentiable in x ∈ X \{θ} . Then we have �x�ν � (x, ·) ∈ J(x) .
Proof:
Due to Lemma 3.19 we have for x ∈ X \{θ}
ν � (x, ·) ∈ {λ ∈ X ∗ : �λ� = 1, �λ, x� = �x�} .
�

We set
Σ(x) := {λ ∈ X ∗ : �λ, x� = �x�, �λ� = 1} , x ∈ X \{θ} .
Since for x �= θ �x�−1 J(x) = Σ(x) we know that Σ(x) is nonempty for all x ∈ X \{θ} .
Above we have seen ν � (x, ·) ∈ Σ(x) if ν is Gateaux differentiable in x .

57
Lemma 3.21. Let X be a Banach space with norm-mapping ν . Let x ∈ S1 and let λ ∈ X ∗ .
Then the following statements are equivalent:

(a) λ ∈ Σ(x) .
(b) ν−� (x, y) ≤ �λ, y� ≤ ν+� (x, y) for all y ∈ X .

Proof:
(a) =⇒ (b) Let y ∈ X , t > 0 . Then

�λ, x + ty� − �λ, x� �λ, x + ty� − �x� �x + ty� − �x�

�λ, y� = = ≤
t t t
and
�x − ty� − �x� �x + t(−y)� − �x�
=− ≤ −�λ, −y� = �λ, y� .
−t t
This implies ν−� (x, y) ≤ �λ, y� ≤ ν+� (x, y) for all y ∈ X .
(b) =⇒ (a) From (5), (6) in Lemma 3.17 we conclude λ ∈ Σ(x) . �

Lemma 3.22. Let X be a Banach space with norm-mapping ν . Let x ∈ X \{θ}, z ∈ X

and s ∈ [ν−� (x, z), ν+� (x, z)] . Then there exists λ ∈ X ∗ with

ν−� (x, y) ≤ �λ, y� ≤ ν+� (x, y) for all y ∈ X , �λ, z� = s , �λ� = 1 , �λ, x� = 1 . (3.5)

Proof:
We shall construct λ by the Hahn-Banach theorem. Set Z := span({z}) and define µ on
Z by
�µ, az� := as , a ∈ R .
Using the properties (2), (3) in Lemma 3.17 it is easy to prove

ν−� (x, u) ≤ �µ, u� ≤ ν+� (x, u) for all u ∈ Z .

We know from (3) in Lemma 3.17 that ν+� (x, ·) is subadditive, i.e. ν+� (x, u + w) ≤
ν+� (x, u) + ν+� (x, w) for all u, w ∈ X . Hence, by the Hahn-Banach Theorem we may
extend µ from Z to the space X and obtain a linear functional λ ∈ X � such that

�λ, u� = �µ, u� for all u ∈ Z and �λ, y� ≤ ν+� (x, y) for all y ∈ X .

Using (2) in Lemma 3.17 we obtain

ν−� (x, y) ≤ �λ, u� ≤ ν+� (x, y) for all y ∈ X

and this implies due to (5) in 3.17

|�λ, y�| ≤ �y� for all y ∈ X .

Therefore λ ∈ X ∗ and �λ� ≤ 1 . Since ν−� (x, x) = ν+� (x, x) = 1 we have �λ� = 1 . �

Definition 3.23. A Banach space X is called Gateaux differentiable if the norm is

Gateaux-differentiable in each x ∈ X \{θ} . �

58
 Clearly, a Hilbert space H is Gateaux differentiable. The Gateaux derivative ν � (x, ·)
of the norm in x ∈ H\{θ} is given as follows:
x
ν � (x, y) = � |y� , y ∈ H .
�x�
Gateaux differentiability of a space is an analytic property of the norm-mapping. Now,
we want to show that this property is related to the observation that in an euclidean space
a boundary point of the unit sphere allows to draw a unique line through this point which
is tangent to the sphere. We know ν � (x, ·) ∈ Σ(x) . Each λ ∈ Σ(x) defines a family of
hyperplanes
Hλ,a := {z ∈ X : �λ, z� = a} , a ∈ R .
The hyperplane Hλ,a for a := �x� touches the ball B�x� in X and we say that it supports
B�x� in x .
Definition 3.24. Let X be a Banach space.
(a) X is smooth at x ∈ S1 (or x ∈ S1 is point of smoothness of X ) if #Σ(x) = 1 .
(b) X is smooth if X is smooth at every x ∈ S1 .
�
Clearly, a Hilbert space H is smooth since we already know that #Σ(x) = 1 for all
x ∈ H\{θ} .
Theorem 3.25. Let X be a Banach space. Then we have:
(a) If X ∗ is strictly convex then X is smooth.
(b) If X ∗ is smooth then X is strictly convex.
Proof:
Ad (a) Assume that X is not smooth. Then there exists x0 ∈ S1 and λ, µ ∈ Σ(x0 ) with
λ �= µ . We obtain

�λ� + �µ� ≥ �λ + µ� ≥ �λ + µ, x0 � = 2 = �λ� + �µ�

which implies that X ∗ is not strictly convex.

Ad (b) Assume that X is not strictly convex. Then there exist x, y ∈ S1 so that � 12 (x +
y)� = 1 . Choose λ ∈ Σ( 12 (x + y)) . Then

1 1 1 1
1 = �λ, (x + y)� = �λ, x� + �λ, y� ≤ (�x� + �y�) = 1 .
2 2 2 2
Therefore we have �λ, x� = �λ, y� = 1 . Consider now x, y as elements in X ∗∗ . Then we
conclude that X ∗ is not smooth at λ . �
Theorem 3.26. Let X be a Banach space. Let x ∈ X \{θ} . Then the following statements
are equivalent:
(a) The duality map is single-valued.
(b) X is smooth at x .

59
(c) The norm is Gateaux differentiable in x .

Proof:
(a) ⇐⇒ (b) This follows from the fact that X is smooth at x ∈ S1 iff #SSigma(x) = 1 .
(b) =⇒ (c) If the norm is not Gateaux differentiable at x then there exists a z ∈ X with
ν−� (x, z) < ν+� (x, z) . Then by Lemma 3.22 there exist at least two different functionals
λ1 , λ2 ∈ Σ(x) . Therefore x is not a point of smoothness.
(c) =⇒ (b) Since ν−� (x, y) = ν+� (x, y) for all y ∈ X the only functional λ ∈ Σ(x) is the
functional λ with �λ, y� = ν � (x, y) = ν+� (x, y) for all y ∈ X . �

Theorem 3.27. Let X be a reflexive Banach space with the duality map J . Then the
following conditions are equivalent:

(a) J is single-valued.
(b) X ∗ is strictly convex.
(c) X is Gateaux differentiable.
(d) X is smooth.

Proof:
(a) =⇒ (b) We know #Σ(x) = 1 for all x ∈ S1 . Let λ, µ ∈ X ∗ with

�λ� = �µ� = 1 and �λ + µ� = 2 .

Since X ∗ is reflexive there exists x0 ∈ X with

1 1
�x0 � = 1 and 1 = � (λ + µ)� = � (λ + µ), x0 � .
2 2
Since |�λ, x0 �| ≤ 1, |�µ, x0 �| ≤ 1, we obtain �λ, x0 � = �µ, x0 � = 1 = �x0 � . This shows
λ, µ ∈ Σ(x0 ) and therefore λ = µ .
(b) =⇒ (c) Let x ∈ S1 . If the norm is not Gateaux differentiable in x then there exists
z ∈ X with ν−� (x, z) < ν+� (x, z) . Choose s1 , s2 with

ν−� (x, z) < s1 < s2 < ν+� (x, z) .

By Lemma 3.22 there exist λ1 , λ2 ∈ X ∗ with �λ1 , z� �= �λ2 , z�, �λ1 � = �λ2 � = 1, and

ν−� (x, y) ≤ �λ1 , y� ≤ ν+� (x, y) , ν−� (x, y) ≤ �λ2 , y� ≤ ν+� (x, y) , for all y ∈ X .

Then
�λ1 + λ2 � ≥ �λ1 + λ2 , x� = �λ1 , x� + �λ2 , x� ≥ ν−� (x, x) + ν−� (x, x) = 2
and we conclude that X ∗ is not strictly convex by Lemma 3.14.
(c) ⇐⇒ (d) See Theorem 3.26.
(d) =⇒ (a) If #J(x) = #Σ(x) ≥ 2 for x ∈ S1 then X cannot be smooth at x .
�

Without reflexivity the implication (d) =⇒ (b) in Theorem 3.27 is not true. One
obtains a counterexample by remorming l1 such that the resulting space lr1 is strictly
convex and l∗1 = l∞ = (lr1 )∗ ; see for instance [14].

60
3.4 Characterization of best approximations in
Banach spaces
Now, we want to derive similar results as in Section 3.1 for the case that the approximation
problem is formulated in a Banach space. First of all, we introduce the tangential and
normal cone of a nonempty closed convex set.

Definition 3.28. Let X be a Banach space, let C be a nonempty closed convex subset of
X and let z ∈ C .

(a) T (z, C) := ∪{s(u − z) : u ∈ C, s ≥ 0} is called the tangential cone of C at x .

(b) If X is a Hilbert space H we set N(z, C) := {y ∈ H : �y|u − z� ≤ 0 for all u ∈ C}
and N(z, C) is called the normal cone of C at z .

Clearly, T (z, C), N(z, C) are nonempty closed convex cones. One can easily see that
T (z, C) is the smallest closed convex cone which contains alle vectors u − z, u ∈ C . The
normal cone is the polar cone of the tangential cone. For interior points z of C both cones
are not very interesting: T (z, C) = H, N(z, C) = {θ} .
As we know from Kolmogorov’s criterion, if x ∈ X and w ∈ C then

w = PC (x) ⇐⇒ x − w ∈ N(w, C) (3.6)

This form of Kolmogorov’s criterion will be now transfered to the Banach space case. The
way we do that is to use optimality conditions for convex programs.

Let X be a Banach space, let C be a nonempty closed convex subset of X and let
x ∈ X . The approximation problem may be reformulated as follows:

Minimize (f + δC )(u) subject to u ∈ X . (3.7)

Here f : X � y �−→ �x − y� ∈ R and

�
0 if y ∈ C
δC (y) := .
∞ if y ∈
/C

Of course, f : X −→ R is a convex function and δC : X −→ R := R ∪ {∞} is convex

function too, since both functions have an epigraph which is a convex subset on X × R .
Set G := f + δC . G is a convex function too. A supporting hyperplane of G in a point
(y0 , G(y0 )) ∈ epi(G) with G(y0 ) < ∞ is the graph of a mapping

X � y �−→ �λ, y − y0 � + G(y0 ) ∈ R (λ �= θ)

such that
epi(G) ⊂ H+ := {(y, s) : �λ, y − y0 � + G(x0 ) ≤ s} .
This leads us to the subdifferential of a convex function.

61
Definition 3.29. Let X be a Banach space and let F : X −→ R convex. If y0 ∈
dom(F) := {y : F(y) < ∞} then the set

∂F(y0 ) := {λ ∈ X∗ : �λ, y − y0 � + F(y0 ) ≤ F(y) for all y ∈ X (3.8)

is called the subdifferential of F in y0 . �

The subdifferential induces a set-valued mapping

∂F : X ⇒ X ∗ .

If F is continuous in a point y0 ∈ dom(G) then ∂G(y0 ) is nonempty; see Theorem 10.44

in Chapter 10.

Let us come back to the minimization problem (3.7). Suppose that w ∈ C belongs to
dom(g) , g := f + δC . Then we have the obvious observation that

w ∈ PC (x) holds true iff θ ∈ ∂g(w) . (3.9)

Now the application of the subdifferential formula

∂(f + δC )(w) = ∂f(w) + ∂δC (w) (3.10)

is allowed since w ∈ dom(f) ∩ dom(δC ) and f is continuous in w ; see Theorem 10.49 in

Chapter 10. Consequentely, if x ∈ X and w ∈ C then

w ∈ PC (x) holds true iff − ∂f(w) ∩ ∂δC (w) �= ∅ . (3.11)

Next we analyze ∂f(w), ∂δC (w) .

Lemma 3.30. Let X be a Banach space and let C be a nonempty closed convex subset of
X . Let w ∈ C . Then
∂δC (w) = N(w, C) (3.12)
where N(w, C) := {µ ∈ X∗ : �µ, z� ≤ 0 for all z ∈ T (w, C)} is the so called normal cone
of C at w .
Proof:
Suppose λ ∈ ∂δC (w) . Then

�λ, y − w� ≤ δC (y) − δC (w) for all y ∈ X .

This implies �λ, u − w� ≤ 0 for all u ∈ C and hence, λ ∈ N(w, C) . �

Lemma 3.31. Let X be a Banach space, let x ∈ X and let f : X � y �−→ �x − y� ∈ R .
If x − w �= θ then
∂f(w) = Σ(x − w) . (3.13)
Proof:
Clearly, ∂f(w) is nonempty since f is convex and f is continuous in w ; see Theorem 10.47
in Chapter 10. Suppose µ ∈ ∂f(w) . Then

�µ, h� ≤ �x − (w + h)� − �x − w� for all h ∈ X . (3.14)

62
Clearly,
�µ, h� ≤ �x − w� + �h� − �x − w� = �h� for all h ∈ X
and
�µ, x − w� ≤ �x − x� − �x − w�
which implies �µ� = 1 since x − w �= θ . Setting h := w − x in (10.27) we obtain

�µ, w − x� ≤ �x − (2w − x)� − �x − w� = �x − w�

and we obtain for ρ := −µ

ρ ∈ X∗ , �ρ� = 1, �ρ, x − w� = �x − w�, i.e. ρ ∈ Σ(x) . (3.15)

Conversely, if λ ∈ Σ(x − w), then

�x−(w+h)�−�x−w� = �x−(w+h)�−�λ, x−w� ≥ �λ, x−(w+h)�−�λ, x−w� = �λ, h�

for all h ∈ X . This shows λ ∈ ∂f(w) . �

Theorem 3.32. Let X be a Banach space and let C be a nonempty closed convex subset
of X . Let x ∈ X , w ∈ C . Then the following conditions are equivalent:

(a) w ∈ PC (x) .
(b) J(x − w) ∩ N(w, C) �= ∅ .

Proof:
If x − w = θ then x ∈ PC (x) and λ := θ ∈ J(x − w) ∩ N(w, C) . If x − w �= θ then we have
shown above w ∈ PC (x) iff Σ(x−w)∩N(w, C) �= ∅ . But as we know, J(x−w) = {�x−w�µ :
µ ∈ Σ(x − w)} and therefore J(x − w) ∩ N(w, C) �= ∅ iff Σ(x − w) ∩ N(x − w, C) �= ∅ . �

In Lemma 3.20 we have shown that in a Banach space X the Gateaux dervative ν � (x, ·)
at the point x ∈ X \{θ} belongs to Σ(x) (when it exists). Hence, �x�ν � (x, ·) ∈ J(x) .
1
Theorem 3.33 (Asplund, 1966). Let X be a Banach space and let j : X � x �−→ 2
�x�2 ∈
R . Then
∂j(x) = J(x) for all x ∈ X . (3.16)

Proof:
We follow [1]. Notice that ∂j(x) is nonempty for all x ∈ X since j is continuous. Moreover,
for x = θ nothing has to be proved. Let x ∈ X \{θ} .
Let λx ∈ ∂j(x) . Choose any y ∈ S�x� . Then �λx , y − x� ≤ 0, i.e. �λx , y� ≤ �λx , x� . This
implies
�λx ��x� = sup �λx , y� ≤ �λx , x� .
y∈S�x�

Hence, we have shown

�λx ��x� = �λx , x� for all λx ∈ ∂j(x), for all x ∈ X \{θ} . (3.17)

63
Let z ∈ S1 and let t > 0, s > 0 . Let λtz ∈ ∂j(tz) . Then using (3.17) we obtain

1 2 1 2
t − s ≥ �λsz , tz� − �λsz ��sz�
2 2
= �λsz , tz� − �λsz ��sz� + �λsz ��tz� − �λsz ��tz�
= �λsz �(t − s) + (�λsz , tz� − �λsz ��tz�)
= �λsz �(t − s) + t(�λsz , sz� − �λsz ��sz�)/s
≥ �λsz �(t − s)

This shows the property λw ∈ J(w) on the ray from θ through z . From this we conclude
that λx ∈ J(x) for all x ∈ J(x) .
Conversely, suppose that λx ∈ J(x) . Then by the definition of J(x)

1 1
�x�2 + �λx , y − x� ≤ �x�2 + �λx �(�y� − �x�)
2 2
and
1 1 1 1 1 1
�x�2 +�λx �(�y�−�x�) ≤ �x�2 +�λx ��y�−�x�2 ≤ �x�2 + �λx �2 + �y�2 −�x�2 = �y�2 .
2 2 2 2 2 2
This shows λx ∈ ∂j(x) . �

Now, we know the subdifferential of the norm mapping ν : X � x �−→ �x� ∈ R and
j : X � x �−→ 12 �x�2 ∈ R:

∂ν(θ) = B1 (3.18)
∂ν(x) = Σ(x) if x �= θ (3.19)
∂j(x) = J(x) (3.20)

The assertion (3.18) follows from the observation

�λ, y − θ� ≤ ν(y − θ) − ν(θ) = �y�, y ∈ X

for all λ ∈ ∂ν(θ) .

3.5 Semi-inner products and characterization of best

approximations
The characterization results in Hilbert spaces and Banach spaces look very different. In
this section we want to bring together these results by using semi-inner products in Banach
spaces.3 This can be done by introducing the concept of semi-inner products defined by
Lumer [13], Giles [10] and used in approximation problems by Penot [15]. Here we follow
mainly Dragomir [9]; see also Rosca [16].
3
Again, remember that we discuss our results strictly under the completeness assumption. Since we
discuss the question of characterization of best approximations it would be sufficient to consider normed
spaces only.

64
Definition 3.34. Let X be a real vector space4 . A mapping [·|·] : X × X −→ R is called
a semi-inner product if the following properties are satisfied:

(1) [x + y|z] = [x|z] + [y|z] for all x, y, z ∈ X ;

(2) [ax|y] = a[x|y], [x|ay] = a[x|y] for all x, y ∈ X, a ∈ R ;

(3) [x|x] ≥ 0 for all x ∈ X and [x|x] = 0 implies x = θ ;

(4) |[x|y]|2 ≤ [x|x] [y|y] for all x, y ∈ X .

Corollary 3.35. Let X be a real vector space and [·|·] : X × X −→ R be a semi-inner

product on X . Then the mapping � · � : X −→ R defined by
1
�x� := [x|x] 2 , x ∈ X,

is a norm on X .

Proof:
This can easily be shown. For convenience we verify the triangle inequality. Let x, y ∈ X
with �x + y� �= 0 ; the case �x + y� = 0 is trivial. Then

�x + y�2 = [x + y|x + y] = [x|x + y] + [y|x + y] ≤ �x��x + y� + �y��x + y�

and dividing by �x + y� the triangle inequality follows. �

Corollary 3.36. Let X be a real vector space and [·|·] : X × X −→ R be a semi-inner

product on X . Let x ∈ X . Then the mapping λ : X −→ R, defined by

λ(y) := [y|x] ∈ R
1
is a functional in X ∗ with �λ� = �x� when X is endowed with the norm [·|·| 2 (see Corollary
3.35.

Proof:
The linearity follows from the linearity of [·|·] in the first argument. From |λ(y)| = |[y|x]| ≤
1 1 1
[x|x] 2 [y|y] 2 we conclude �λ� ≤ [x|x] 2 = �x� and hence, λ is continuous. On the other
�
1 1
hand, we have [x|x] = |λ(x)| ≤ �λ�[x|x] 2 , i.e. �λ� ≥ [x|x] 2 = �x� .

Definition 3.37. Let X be a Banach space with norm � · � and endowed with a semi-inner
product [·|·] . Then we say that [·|·] generates the norm if �x� = [x|x] 2 for all x ∈ X . �
1

Definition 3.38. Let X be a Banach space and let J : X ⇒ X∗ be the duality map. Then
a mapping J̃ : X −→ X∗ with J̃(x) ∈ J(x), x ∈ X , is called a section of J . Such a section
J̃ is called a homogeneous section if the mapping x �−→ J̃(x) is homogeneous, i.e.
J̃(ax) = aJ̃(x), x ∈ X , a ∈ R . �
4
Semi-inner products can be defined for vector spaces with complex scalars. The homogenity property
has to be adapted.

65
Lemma 3.39. Let X be a Banach space with duality map J . Then the following statements
hold:

(a) The duality map allows homogeneous sections.

(b) The duality map has a uniquely determined homogeneous section if and only if J is
single-valued.
(c) Every semi-inner product [·|·] on X which generates the norm in X is of the form

[x|y] = �J̃(y), x� for all x, y ∈ X , (3.21)

where J̃ is a homogenous section of the duality map J on X .

Proof:
Ad (a) Let z ∈ S1 . Choose (exactly one) λz ∈ X∗ with �λz � = 1, �λz , z� = 1 . Then
λz ∈ Σ(z) . Now, if x = az, a ∈ R, z ∈ S1 , set λx := aλz .
Let x, y ∈ X , x �= θ, r ∈ R, r �= 0 . Then for r > 0

�rλx , y� = �r�x�λx�x�−1 , y� = ��x�λrx�rx�−1 , y� = �λrx , y�

and for r > 0

�rλx , y� = �(−r)λx , −y� = �λ−rx , −y�

= ��r(−x)�λr(−x)�r(−x)�−1 , −y� = �r�x�λx�x�−1 , y� = �rλx , y� .

Moreover,

�λx , y� = ��x�λx�x�−1 , y� ≤ �x��y� , �λx , x� = ��x�λx�x�−1 , x� = �x�2 .

Therfore, a homogeneous section J̃ of J is defined by setting

J̃(x) = λx , x ∈ X .

Ad (b) From the proof in (a) it is clear that homogeneous sections are not uniquely
determined if #J(x) ≥ 2 for some x ∈ X \{θ} .
Ad (c) Let J̃ be a homogeneous section of J . Define the mapping

[·|·] : X × X −→ R , [x|y] := �J̃(y), x� .

Then | · |·] is linear in the first argument and homogeneous in each argument. We have

[x|x] = �J̃(x), x� = �x�2 = �J̃(x)��x� ≥ 0, x ∈ X ,

and therefore [x, x] = 0 iff x = θ . Since

|[x|y]|2 = |�J̃(y), x�|2 ≤ �J̃(y)�2 �x�2 = �y�2 �x�2 = [x|x]|y|y]

[·|·] is a semi-inner product which generates the norm.

Conversely, let [·|·] be a semi-inner product on X which generates the norm of X . Define
J̃ : X −→ X ∗ by
�J̃(y), x� := [x|y] , x ∈ X .

66
Then
�J̃(x), x� = [x|x] = �x�2 , �J̃(y)� = �x�, x, y ∈ X ,
Consequentely, we have

�J̃(x), x� = �x�2 and �J̃(x)� = �x� .

Moreover y �−→ J̃(y) is homogeneous section. �

Lemma 3.39 says that every Banach space can be endowed with a semi-inner product.

Corollary 3.40. Let X be a smooth Banach space. Then the duality map is single-valued
and the uniquely determined semi-inner product which generates the norm is given by

[x, y] := �J(y), x� , x, y ∈ X . (3.22)

Proof:
We know that in a smooth Banach space the duality map is single-valued; see Theorem
3.27. Therefore by Lemma 3.39 the semi-inner product is uniquely determined. �

Example 3.41. Consider the Banach space X := R3 endowed with the l1 -norm. It is
easy to check that by setting
3
� xk y k
[x|y] := �y�1 , x = (x1 , x2 , x3 ), y = (y1 , y2 , y3 ) .
k=1,yk � 0
=
|yk |

a semi-inner product in X is defined which generates the l1 norm in R3 . �

With the help of a semi-inner product which generates the norm in a Banach space we
can reformulate the property of strictly convexity.

Theorem 3.42. Let X be a Banach space and let [·|·] be a semi-inner product which
generates the norm in X . Then the following statements are equivalent:

(a) X is strictly convex.

(b) If x, y ∈ X \{θ} with [x|y] = �x��y� then there exists a > 0 with x = ay .

Proof:
(a) =⇒ (b) Let J̃ be a homogeneous section of J which generates the semi-inner product.
Let x, y ∈ X \{θ} with [x|y] = �x��y� . Then [x|y] = �J̃(y), x� = �x��y� . This implies
x y
�J̃(y), � = �y� = �J̃(y)� , �J̃(y), � = �y� = �J̃(y)� .
�x� �y�

Since X is strictly convex, every continuous linear functional attains its norm on at most
one point; see Lemma 3.14. This implies
x y
=
�x� �y�

67
and we have x = ay with a = �x��y�−1 .
(b) =⇒ (a) Let x, y ∈ X \{θ} with �x + y� = �x� + �y� . Then

[x|x + y] ≤ �x��x + y� , [y|x + y] ≤ �y��x + y� . (3.23)

Suppose that in (3.23) one of inequalities above is strict. Then, by addition, we obtain

�x + y�2 = [x + y|x + y] = [x|x + y] + [y|x + y] < (�x� + �y])�x + y� = �x + y�2

which is a contradiction. Therefore, [x|x + y] = �x�|�x + y� and by (b) we obtain

x = a(x + y) with a > 0 . Since obviously a �= 0 we have x = sy with s = (1 − a)−1 . �
Lemma 3.43. Let X be a Banach space and let [·|·] be a semi-inner product which gen-
erates the norm in X . Then for all x, y ∈ X , t > 0, x �= θ, x + ty �= θ,
[y|x] �x + ty� − �x� [y|x + ty]
≤ ≤ (3.24)
�x� t �x + ty�
Proof:
Let J̃ be a homogeneous section which generates [·|·]:

[x|y] = �J̃(y), x�, x, y ∈ X .

Let x, y ∈ X , t > 0, x �= θ, x + ty �= θ .

�x + ty� − �x� �x + ty��x� − �x�2 �J̃(x), x + ty� − �x�2

= ≥
t t�x� t�x�
�J̃(x), y� [y|x]
= =
�x� �x�
and
�x + ty� − �x� �x + ty�2 − �x��x + ty�
=
t t�x + ty�
�x + ty�2 − �J̃(x + ty), x�
≤
t�x + ty�
�x + ty�2 − �J̃(x + ty), x + ty� + �J̃(x + ty), ty�
=
t�x + ty�
�J̃(x + ty), y� [y|x + ty]
= =
�x + ty� �x + ty�
�
Theorem 3.44. Let X be a Banach space and let [·|·] be a semi-inner product which
generates the norm in X . Then the following statements are equivalent:
(a) X is smooth.
(b) X is Gateaux-differentiable.
(c) The semi-inner product [·|·] which generates the norm in X is uniquely determined.

68
(d) The duality map of X is single-valued.
(e) limt→0 [y|x + ty] exists for all x ∈ X \{θ}, y ∈ X .
Additionally, if one of the above statements is true, the limit in (e) is given as [y, x] .
Proof:
Let J be the duality map of X and let Let J̃ be a homogeneous section which generates
[·|·] .
(a) ⇐⇒ (b) Theorem 3.26.
(a) =⇒ (c) If X is smooth, #J(x) = 1 for all x ∈ X ; see Theorem 3.26. Therefore there
exists exactly one homogeneous section of J which generates the semi-inner product.
(c) =⇒ (d) Obviously, there exists only one homogeneous section of J and by Lemma
3.39, #J(x) = 1 for all x ∈ X .
(d) =⇒ (e) Let x ∈ X \{θ}, y ∈ X .
Let (tn )n∈N be a sequence in (0, ∞) with limn tn = 0 . Then (J̃((x + tn y)�x + tn y�−1 ))n∈N
is a sequence in the unit ball of X ∗ . By Lemma 3.12 limt↓0 J̃((x + tn y)�x + tn y�−1 )) =
J̃(x�x�−1 ) . This shows

lim J̃((x + ty�x + ty�−1 )) = J̃(x�x�−1 )

t↓0

which implies limt→0 [y|x + ty] = [y|x] .

(e) =⇒ (b) We conclude from Lemma 3.43 that ν−� (x, y) = ν+� (x, y) which shows that
ν is Gateaux differentiable. �
Definition 3.45. Let X be a Banach space and let [·|·] : X × X −→ R be a semi-inner
product on X . The semi-inner product is called continuous iff

lim[x|y + tx] = [x|y] for all x, y ∈ S1 .

t→0

�
Theorem 3.46. Let X be a Banach space and let [·|·] : X × X −→ R be a semi-inner
product on X which generates the norm. Let C be a nonempty closed convex set and
x ∈ X , w ∈ C . Then the following statements are equivalent:
(a) w ∈ PC (x) .
(b) [(u − w|x − w − t(u − w)] ≤ 0 for all u ∈ C and t ∈ [0, 1] .
(c) [u − w|x − w] ≤ 0 for all u ∈ C .
Proof:
(a) =⇒ (b) Assume by contradiction, [(u − w|x − w − t(u − w)] > 0 for some u ∈ C
and t ∈ [0, 1] . Clearly, u �= w . Set z := tu + (1 − t)w . Then z ∈ C and we have

�x − z� = �x − w − t(u − w)�−1 [x − w − t(u − w)|x − w − t(u − w)]

= �x − w − t(u − w)�−1 [x − w − t(u − w)|x − w − t(u − w)]
−t[u − w|x − w − t(u − w)]
< �x − w − t(u − w)�−1 [x − w|x − w − t(u − w)]
≤ �x − w�

69
This is a contradiction to the best approximation property of w .
(b) =⇒ (c) Take t := 0 in (b) .
(c) =⇒ (a) We have for all u ∈ C:

�x − w�2 = [x − w|x − w]
= [x − u|x − w] + [u − w|x − w]
≤ [x − u|x − w]
≤ �x − u��x − w�

We conclude �x − w� ≤ �x − u� for all u ∈ C . Hence, w ∈ PC (x) and since C is a

Chebyshev set w = PC (x) .
�
Definition 3.47. Let X be a Banach space and let [·|·] : X × X −→ R be a semi-inner
product on X . Let K be a nonempty subset of X . Then we set

K◦ := {x ∈ X : [y|x] ≤ 0 for all y ∈ K}

and we call K◦ the dual cone (relative to [·|·]) of K . �

Theorem 3.48. Let X be a Banach space and let [·|·] : X × X −→ R be a semi-inner
product on X which generates the norm. Let C be a nonempty closed convex set and
x ∈ X , w ∈ C . Then the following statements are equivalent:
(a) w ∈ PC (x) .
(b) x − w − t(y − w) ∈ (C − w)◦ for all y ∈ C and t ∈ [0, 1] .
(c) x − w ∈ (Cw )◦ .
Proof:
Obviously with Theorem 3.46. �

Let X be a Banach space with norm � · � . Since j : X � x �−→ 1

2
�x� ∈ R is a convex
function the following limits exist:

�y + tx�2 − �y�2 �y + tx�2 − �y�2

[x, y]+ := lim , [x, y]− := lim ; (3.25)
t↓0 2t t↑0 2t
apply Lemma 3.16. The binary functions [·|·]+ , [·|·]− are called the superior semi-inner
product and the inferior semi-inner product associated to the norm in X respectively.
It is not difficult to see that for all x, y ∈ X , t > 0, s < 0 we have

�y + tx�2 − �y�2 �y + sx�2 − �y�2

lim ≥ [x, y]+ ≥ [x, y]− ≥ lim . (3.26)
t↓0 2t s↑0 2s
For details see [9].
Theorem 3.49. Let X be a Banach space with norm �·� . Then we have for all x, y ∈ X :
(a) [x, y]+ = sup{�λ, y� : λ ∈ J(y)} .
(b) [x, y]− = inf{�λ, y� : λ ∈ J(y)} .

70
Proof:
See Dragomir [9]. �

[·, ·]+ , [·, ·]− may be use to develop Kolmogorov-like results for best approximations in
Banach spaces. This is done in [15].

A semi-inner product [·|·] in a real vector space X max be used to introduce an orthog-
onality relation via x orthogonal to y“ if and only if [y|x] = 0 . This is not a symmetric
”
relation and transitivity of the orthogonality of vectors is not guaranteed in general.

Example 3.50. Consider the space R3 endowed with the l1 -norm; see Example 3.41.
Then the semi-inner product
3
�
[x|y] := �y� |yi |−1 xi yi
i=1,yi �=0

generates the norm. Consider x := (−2, 1, 0), y := (1, 1, 0) . Then we have [y|x] = 0,
[x|y] = −2 and we see that x is orthogonal to y but y is not orthogonal to x . �

There are other various concepts of orthogonality in normed spaces. In general, they
are modeled along a property which holds in inner product spaces. Let (X, �·|·�) be an
inner product space. Here are three properties in an inner product space which may serve
as a model for orthogonality in normed spaces:

(1) x⊥y ⇐⇒ �x|y� = 0 .

(2) x⊥y ⇐⇒ �x + y� = �x − y�

(3) x⊥y ⇐⇒ �x� ≤ �x + ty� for all t ∈ R .

The orthogonality via (1) can be generalized to Banach spaces by saying

x is orthogonal to y iff �J(y), x� = 0 .

where we suppose that J is single-valued. This is the concept which corresponds to semi-
inner products; see above. We denote this orthogonality as J-orthogonality.
Orthogonality in the sense of Birkhoff in normed spaces is derived from (3):

x is orthogonal to y iff �x� ≤ �x + ty� for all t ∈ R .

In general, Birkhoff-orthogonality is not a symmetric property too, i.e. x⊥y does not
imply y⊥x . Nevertheless, Birkhoff-orthogonality may be used to generalize orthogonal
projections from Hilbert spaces to general Banach spaces; see for instance [9, 12].

71
3.6 Appendix: Convexity II
We want to consider the following optimization problem:

Minimize (f + g)(x) , x ∈ X . (3.27)

Here X is a Banach space and f, g : X −→ R ^ are proper closed convex functions. In
order to associate to this problem a dual problem we consider the process of conjugation
for convex functions.
Definition 3.51. Let X be a Banach space and let f : X −→ R ^ be a convex function.
Then we set
f∗ (λ) := sup(�λ, x� − f(x)) , λ ∈ X∗
x∈X
∗
and call f the conjugate of f . �
Conjugation defines a convex function f∗ : X∗ −→ R̄ . It is obvious that f∗ : X −→
^ if f is proper. An immediate consequence of the definition is the Fenchel-Young
R
inequality
f(x) + f∗ (λ) ≥ �λ, x� , x ∈ X, λ ∈ X∗ . (3.28)
This is an extension to non-quadratic convex functions f of the inequality
1 1
�x�2 + �y�2 ≥ �x|y�
2 2
in the euclidean space Rn .
We also note that f∗ is convex and weak∗ -lower semicontinuous because it is the supremum
of the family of affine continuous functions (�·, x� − f(x))x∈X . When X is a Hilbert space
then the conjugate may be considered as a function on the Hilbert space itself and the
only function f which is selfdual (f = f∗ ) is the function x �−→ 12 �x�2 .

We set
f∗∗ (x) := sup (�λ, x� − f∗ (λ)) , x ∈ X,
λ∈X∗

and call f∗∗ the double conjugate of f . One has the remarkable duality (see [2])

f = f∗∗ ⇐⇒ f convex and lower semicontinuous

The most important theorem concerns the Fenchel-Rockafellar duality which is the
content of the following theorem. To formulate this duality we associate to the primal
problem (10.33) a dual problem using the conjugates of f and g:

Maximimize (−f∗ (−λ) − g∗ (−λ)) , λ ∈ X∗ (3.29)

and we set
p := inf (f(x) + g(x)) , d := sup (−(f∗ + g∗ )(−λ)) ,
x∈X λ∈X∗

p ∈ R̄, d ∈ R̄ are the values of the primal and dual problem, respectively.
Corollary 3.52 (Weak duality). Let X be a Banach space and let f, g : X −→ (−∞, ∞]
be convex. Then p ≥ d .

72
Remark 3.53. In case of d < p we say that there exists a duality gap. Without further
assumptions we cannot guarantee that there is no duality gap. If there is no duality gap
the identity p = d can be used to find a lower bound for p since we have

−(f∗ + g∗ )(−λ) ≤ d = p for all λ ∈ X∗ .

An upper bound for p is easily found:

p ≤ f(x) + g(x) for all x ∈ X .

Theorem 3.54 (Fenchel, 1949, Rockafellar, 1966). Let X be a Banach space and let
f, g : X −→ R ^ be convex. We assume that one of the functions f, g is continuous in
0
some point x ∈ dom(f) ∩ dom(g) . Then

p = inf (f(x) − g(x)) = max∗ (−(f∗ + g∗ )(−λ)) = d . (3.30)

x∈X λ∈X

Remark 3.55. The duality asserts that in the presence of a so called primal constraint
qualification such as f (or g) is continuous at a point x0 ∈ dom(f) ∩ dom(g)“ one has
”
p := inf (f(x) + g(x)) = d := sup (−f∗ + g∗ )(−λ))
x∈X λ(∈X∗

and the supremum for d is a maximum. If we add a bf dual constraint qualification such
as f∗ (or g∗ ) is continuous at a point λ0 ∈ dom(f∗ ) ∩ dom(g∗ )“ one has that the infimum
”
for p is a minimum and we have that the primal and the dual problem are solvable and
that there exists no duality gap. �

Let X be a Banach space. In the theory of monotone operators T : X ⇒ X∗ a very

important role is played by the so called duality mapping. It essentially does in Banach
space the job done by the identity in Hilbert spaces.

Definition 3.56. Let X be a Banach space. The duality mapping J = JX : X ⇒ X∗ is

defined as follows:

J(θ) = θ , J(x) = {µ ∈ X∗ |�µ, x� = �x�2 , �µ� = �x�} for x �= θ . (3.31)

Expression (10.38) is equivalent to

J(x) = {µ ∈ X∗ |�µ, x� ≥ �x�2 , �µ� ≤ �x�} for x ∈ X . (3.32)

One can show directly from the definition, using the Hahn-Banach, that J(x) is a non-
empty, weak∗ -closed convex subset of X for each x ∈ X.

Lemma 3.57. Let X be a Banach space and let j : X −→ R be the function defined by
x �−→ 12 �x�2 . Then
∂j(x) = J(x) , x ∈ X . (3.33)

73
Proof:
Let µ ∈ ∂j(x). It suffices to consider the case x �= θ since ∂j(θ) = {θ} and J(θ) = {θ} .
Then
1 1
�y�2 ≥ �x�2 + �µ, y − x� , x ∈ X . (3.34)
2 2
Let t > 0 and u ∈ X be arbitrary, and replace in (10.40) y by x + tu:
1 1 1
t�µ, u� ≤ �x + tu�2 − �x�2 ≤ t�x��u� + t2 �u�2 .
2 2 2
Dividing through by t and passing to the limit t ↓ 0 we obtain �µ, u� ≤ �x��u� . Since u
was arbitrary this implies �µ� ≤ �x� . On the other hand, let y = tx in (10.40):
1 2
(t − 1)�x�2 ≥ (t − 1)�µ, x� . (3.35)
2
For 0 < t < 1 we obtain from (10.41) 12 (t + 1)�x�2 ≤ �µ, x�. Letting t → 1 we have
�µ, x� ≥ �x�2 , completing the proof that µ ∈ J(x) .
For other way around, assume now that µ ∈ J(x) . Let us estimate the righthand side in
(10.40) using the properties of µ :
1 1 1
�µ, y� − �µ, x� + �x�2 ≤ �x��y� − �x�2 + �x�2 ≤ �y�2 .
2 2 2
�
Definition 3.58. Let X be a Banach space and let C be asubset of X . Then the set

core(C) := {x ∈ C : ∪r>0 (r(C − x)) = span(C − x)

is called the strong relative interior of C . �

Clearly, int(C) ⊂ core(C) .
Definition 3.59. Let X be a Banach space and let f, g ∈ Γ0 (X ) . Then the inf-convolution
f � g is defined as follows:

(f � g)(x) := inf (f(u) + g(x − u)) , x ∈ X .

u∈X

�
Lemma 3.60. Let X be a Banach space and let f, g ∈ Γ0 (X ), h := f � g . Then

epi(h) = epi(f) + epi(g)

if the infimum in f � g is attained for all x ∈ X .

Proof:
�
Theorem 3.61. Let X be a Banach space and let f, g ∈ Γ0 (X ) . Suppose that the constraint
qualification
θ ∈ core(dom(f) − dom(g))
is satisfied. Then

74
(a) (f + g)∗ = f∗ � g∗ .
(b) For all λ ∈ X ∗ exists µ ∈ X ∗ with (f + g)∗ (λ) = f∗ (µ) + g∗ (λ − µ) .
(c) ∂(f + g) = ∂f + ∂g .

Proof:

3.7 Conclusions and comments

3.8 Exercises
1.) Let X be a Banach space and let C be a closed convex cone. Then for w ∈ X the
following conditions are equivalent:
(a) w ∈ PC (x) .
(b) There exists λ ∈ J(x − w) ∩ C◦ such that �l, w� = 0 . Here C◦ := �µ ∈ X∗ :
�µ, u� ≤ 0 for all u ∈ C} is the polar cone of C .
2.) Let H be a Hilbert space, let A, B be a closed subsets of H and let x ∈ H . Show:
If PB (x) ∈ A, then PA (x) = PB (x) .
3.) Let X be a Banach space and let H := Hλ,α := {x ∈ X : �λ, x� = α} be a hyperplane
(λ ∈ X ∗ , λ �= θ, α ∈ R). Show dist(x, H) = |�λ, x� − α|�λ�−1 for all x ∈ X .
4.) Let X be a Banach space and let H := Hλ,α := {x ∈ X : �λ, x� = α} be a hyperplane
(λ ∈ X ∗ , λ �= θ, α ∈ R) which is a Chebyshev set. Show that PH is linear.
5.) Definiere f : R � r �−→ 12 dist(r, 2Z) ∈ R . Set C := graph(f) and consider R2
endowed with the l1 -norm. Show that C is a nonconvex Chebyshev set.
6.) Let R2 endowed with the maximum norm. Show that the orthogonality relation of
Birkhoff is not symmetric relation.
7.) Show that in an inner product space (X, �·|·�) for x, y ∈ X the following conditions
are equivalent:
(a) �x|y� = 0
(b) �x� ≤ �x + ty� for all t ∈ R .
8.) Let X be a normed space and let x, y ∈ X . Set Ly := span({y}) . Show that the
following conditions are equivalent:
(a) x⊥y in the sense of Birkhoff.
(b) �x� = dist(x, Ly ) .
9.) Let X ∗ be a strictly convex Banach space with duality map J . Show that x, y ∈ X
are orthogonal in the sense of Birkhoff iff that x, y are J-orthogonal.
10.) Let X ∗ be a strictly convex Banach space with duality map J . Show: If x1 , . . . , xm ∈
X are linearly independent then x1 , . . . , xm are pairwise J-orthogonal.

75
11.) Let X be a Banach space with a single-valued duality mappings J : X ⇒ X ∗ , J∗ :
X ∗ ⇒ X . If J is strongly monotone, i.e.

�J(x) − J(y), x − y� ≥ b�x − y�2 for all x, y ∈ X

with a positive constant b, then J is Lipschitz-continuous iff J∗ is strongly monotone.

12.) Let X be a smooth Banach space. Consider

1 �x + ty� − �x� �x + ty� − �x�

g : X × X −→ R , g(x, y) := �x�(lim + lim ).
2 t↓0 t t↑0 t
Then:
(a) g(x, ·) ∈ J(x) for all x �= θ .
(b) An angle φ ∈ [0, π] between x, y is defined by

g(x, y) + g(y, x)
cos(φ) =
2�x��y�

(c) If X is a Hilbert space H then the angle above is the usual in inner product
spaces.
13.) Let H be a Hilbert space and let C be a nonempty closed convex subset of H .
Suppose that U : H −→ H is a surjective linear isometry. Then U(C) is a
nonempty closed convex subset of H and PU(C) = U ◦ PC ◦ U∗ .
14.) Let H be a Hilbert space and let C be a nonempty closed convex cone of H . The
set C◦ := {x ∈ H : �x|y� ≤ 0 for all y ∈ C} is called the polar cone of C . Verify
P C◦ = I − P C .
15.) Show x = PC (x + z) if z ∈ N(x, C) .
16.) Let H be a Hilbert space and let S ⊂ H be a nonempty set. S is called a sun if
for all x ∈ H the following assertion hold:

PS (x) = PS (u) for all u ∈ {PS (x) + t(x − PS (x)) : t ≥ 0} .

Let C ⊂ H be a no(f + g)∗ 0f∗ � g∗ .nempty Chebyshev set. Prove the equivalence
of the following conditions:
(a) C is convex.
(b) C is a sun.
(c) PC is a nonexpansive mapping.
17.) Let X be a Banach space and let [·|·] be a semi-inner product which generates the
norm in X . Then the following statements are equivalent:
(a) X is smooth.
[x|x + ty] − �x�2
(b) limt→0 t exists for all x ∈ X \{θ}, y ∈ X .
18.) Give an elementary proof of the fact that each convex function f : R −→ R is
continuous.

76
Bibliography

[1] E. Asplund. Positivity of duality mappings. Amer. Math. Soc., 73:200–203, 1966.
[2] J. Baumeister. Funktionalanalysis, 2013. Skriptum einer Vorlesung, Goethe–Universität
Frankfurt/Main.

[3] A. Beuerling and A.E. Livingston. A theorem on duality mappings in Banach spaces. Ark.
Mat., 4:405–411, 1962.

[4] H. Brézis, M.G. Crandall, and A. Pazy. Perturbations of nonlinear maximal monotone sets
in Banach spaces. Comm. Pure Appl. Math., 23:123–144, 1970.

[5] F. Browder. Nonlinear operators and nonlinear equations of evolution in Banach spaces.
Proc. Symp. Pure and Appl. Math. 18. Providence, 1976.

[6] E.W. Cheney. Introduction to approximation theory, 2nd ed. AMS publsihing, Providence,
1982.

[7] I. Cioranescu. Geometry of Banach Spaces, Duality Mappings and Nonlinear Problems.
Kluwer, Dordrecht, 1990.

[8] F. Deutsch. Best approximation in inner product spaces. Springer, New York, 2001.

[9] S.S. Dragomir. Semi-inner products and applications. Nova Sciene Hauppage, 2004.

[10] J.R. Giles. Classes of semi-inner-product spaces. Trans. of Amer. Math. Soc., 129:436–446,
1967.

[11] R. Holmes. A Course on Optimization and Best Approximation. Springer, 1971.

[12] S. Kinnunen. On projections and birkhoff-james orthogonality in Banach spaces. Nieuw.

Arch. Wisk., (IV), 2:235–255, 1984.

[13] G. Lumer. Semi-inner product spaces. Trans. Amer. Math. Soc., 100:29–43, 1961.

[14] R.E. Megginson. An introduction to Banach space theory. Springer, New York, 1998.

[15] J.-P. Penot and R. Ratsimahalo. Characteritations of metric projections in Banach spaces
and applications. 1998. Preprint.

[16] I. Rosca. Semi produits scalaires et réprésentaion du type riesz pour les functionélles
linéaires te bornées sur les espace normés. C. R. Acad. Sci. Paris, 283 (19), 1976.

[17] I. Singer. The theory of best approximation and functional analysis, volume 13 of Series in
applied mathematics. SIAM, Philadelphia, 1974.

77