1.4.

THE HAHN BANACH THEOREM 17

1.4 The Hahn Banach Theorem

Definition 1.4.1. Suppose that V is a vector space over K. A real-valued
function p on V , satisfying

• p(0) = 0,

• p(x + y) ≤ p(x) + p(y), and

• p(λx) = λp(x) for λ > 0,

is called a sublinear functional on V .

Note that 0 = p(0) ≤ p(x) + p(−x), so that p(−x) ≥ −p(x).

Theorem 1.4.2. [The analytic Hahn-Banach Theorem, real version]

Suppose that p is a sublinear functional on a real vector space V , that W
is a linear subspace of V and that f is a linear functional on W satisfying
f (y) ≤ p(y) for all y ∈ W . Then there exists a linear functional g on V such
that g(x) = f (x) for all x ∈ W (g extends f ) and such that g(y) ≤ p(y) for
all y ∈ V (control is maintained).

Theorem 1.4.3. [The analytic Hahn-Banach Theorem, general version]

Suppose that p is a seminorm on a real or complex vector space V , that W
is a linear subspace of V and that f is a linear functional on W satisfying
|f (x)| ≤ p(x) for all x ∈ W . Then there exists a linear functional g on V
such that g(x) = f (x) for all x ∈ W (g extends f ) and such that |g(y)| ≤ p(y)
for all y ∈ V (control is maintained).

Corollary 1.4.4. Let X be a normed space Y a subspace and y ∗ ∈ Y ∗ . Then

there exists an extension x∗ of y ∗ to an element in X ∗ with %x∗ % = %y ∗ %.

Proof. Put p(x) = %y ∗ %%x%.

Corollary 1.4.5. Let X be a normed space Y a subspace and x ∈ X with

h = dist(x, Y ) > 0. Then there exists an x∗ ∈ X ∗ , with x∗ |Y ≡ 0 and
x∗ (x) = 1.

Proof. Consider Z = {y + ax : y ∈ Y and a ∈ K}. Note that every z ∈ Z

has a unique representation z = y + ax, with y ∈ Y and a ∈ K. Indeed, if
y1 + a1 x = y2 + a2 x, with y1 , y2 ∈ Y and a1 , a2 ∈ K, then we observe that
a1 = a2 , because otherwise x = (y1 − y2 )/(a1 − a2 ) ∈ Y , and thus, y1 = y2 .
18 CHAPTER 1. SOME BASIC BACKGROUND

We define f : Z → K, y + ax (→ a. The unique representation of each

z ∈ Z implies that f is linear. The functional f is also continuous. Indeed,
assume zn = yn + an x → 0, if n → ∞, but inf n∈N |an | ≥ ε for some ε > 0,
then
!z yn " !z "
n n
dist(x, Y ) = dist − , Y ≤ dist , Y → 0 if n → ∞,
an an an

which contradicts our assumption.

We can therefore apply the Hahn-Banach Theorem 1.4.2 to the linear
functional f on Z and the norm p(x) = %f %Z ∗ %x%.

Corollary 1.4.6. Let X be a normed space and x ∈ X. Then there is an

x∗ ∈ X ∗ , %x∗ % = 1, so that *x∗ , x+ = %x%.

Proof. Let p(x) = %x% and f (αx) = α%x%, for αx ∈ span(x) = {ax : a ∈
K}.

Definition 1.4.7. [The Canonical Embedding, Reflexive spaces]

For a Banach space we put X ∗∗ = (X ∗ )∗ (the dual space of the dual space of
X).
Consider the map

χ : X → X ∗∗ , with χ(x) : X ∗ → K, *χ(x), x∗ + = *x∗ , x+, for x ∈ X.

The map χ is well defined (i.e. χ(x) ∈ X ∗∗ for x ∈ X), and since for x ∈ X

%χ(x)%X ∗∗ = sup |*x∗ , x+| ≤ %x%,

x∗∈BX ∗

it follows that %χ%L(X,X ∗∗ ) ≤ 1 . By Corollary 1.4.6 we can find for each x ∈ X

an element x∗ ∈ BX ∗ with *x∗ , x+ = %x%, and thus %χ(x)%X ∗∗ | = %x%X .
It follows therefore that χ is an isometric embedding of X into X ∗∗ . We
call χ the canonical embedding of X into X ∗∗ .
We say that X is reflexive if χ is onto.

Remark. There Banach spaces X for which X and X ∗∗ are isometrically iso-
morphic, but not via the canonical embedding. An Example by R. C. James
will be covered in Chapter 3.
1.4. THE HAHN BANACH THEOREM 19

Definition 1.4.8. [The adjoint of an operator]

Assume that X and Y are Banach spaces and T : X → Y a linear and
bounded operator. Then the operator
T ∗ : Y ∗ (→ X ∗ , y ∗ (→ y ∗ ◦ T,
(i.e. *T ∗ (y ∗ ), x+ = *y ∗ ◦ T, x+ = *y ∗ , T (x)+ for y ∗ ∈ Y ∗ and x ∈ X)
Proposition 1.4.9. Assume X and Y are Banach spaces and T : X → Y a
linear and bounded operator. Then T ∗ is a bounded linear operator from Y ∗
to X ∗ , and %T ∗ % = %T %.
Moreover if T is surjective T ∗ is an isomorphic embedding, and if T is
an isomorphic embedding T ∗ is surjective.
We want to formulate a geometric version of the Hahn- Banach Theorem.
Definition 1.4.10. A subset A of a vector space V is called convex if for
all a, b ∈ A and all λ ∈ [0, 1] also λa + (1 − λ)b ∈ A.
If A ⊂ V we define the convex hull of A by
#$ %
conv(A) = C : A ⊂ C ⊂ V, C convex
&' n n
' (
= λj aj : n ∈ N, λj ∈ [0, 1], ai ∈ A, for i = 1, . . . n, and λj = 1 .
j=1 j=1

A subset A ⊂ V is called absorbing if for all x ∈ V there is an 0 < r < ∞

so that x/r ∈ A. For an absorbing set A we define the Minkowski functional
by
µA : V → [0, ∞), x (→ inf{λ > 0 : x/λ ∈ A}.
Lemma 1.4.11. Assume C is a convex and absorbing subset of a vector
space V . Then µC is a sublinear functional on V , and
(1.1) {v ∈ V : µC (v) < 1} ⊂ C ⊂ {v ∈ V : µC (v) < 1}.
If V is a normed linear space and if there is an ε > 0 so that εBV ⊂ C,
then µA is uniformly continuous (and thus bounded on BV ).
Proof. Since C is absorbing 0 ∈ C and µC (0) = 0. If u, v ∈ V and ε > 0
is arbitrary, we find 0 < λu < µC (u) + ε and 0 < λv < µC (v) + ε, so that
u/λu ∈ C and v/λv ∈ C and thus
u+v λu u λv v
= + ∈ C,
λ u + λv λ u + λv λu λ u + λv λv
20 CHAPTER 1. SOME BASIC BACKGROUND

which implies that µC (u + v) ≤ λu + λv ≤ µC (u) + µC (v) + 2ε, and, since,

ε > 0 is arbitrary, µC (u + v) ≤ µC (u) + µC (v).
Finally for λ > 0 and u ∈ V
&r λu (
µC (λv) = inf{r > 0 : λv/r ∈ C} = λ inf : ∈ C = λµC (u).
λ r
To show the first inclusion in (1.1) assume v ∈ V with µC (v) < 1, there
is a 0 < λ< 1 so that v/λ ∈ C, and, thus,
v
v = λ + (1 − λ)0 ∈ C.
λ
The second inclusion is clear since for v ∈ C it follows that v = λv ∈ C.
If V is a normed linear space and εBV ⊂ C, it follows from the sublinearity
of µC , for u, v ∈ V , that
%u − v%
µC (u) − µC (v) ≤ µC (u − v) ≤ ,
ε
#u−v#
and similarly µC (v) − µC (u) ≤ ε
.

Theorem 1.4.12. [The Geometric Hahn-BanachTheorem, general version]

Let C be a non empty, closed convex subset of a Banach space X and let
x0 ∈ X \ C.
Then there is an x∗ ∈ X ∗ so that

sup .(*x∗ , x+) < .(*x∗ , x0 +).

x∈C

Proof. We first assume that K = R and we also assume w.l.o.g. that 0 ∈ C

(otherwise pass to C −x and x0 −x for some x ∈ C). Put ε := dist(x0 , C) > 0
and put D = {x ∈ X : dist(x, C) ≤ ε/2}. From Lemma 1.4.11 it follows that
µD is a bounded sublinear functional on X, and dist(x0 , D) ≥ ε/2.
On the one dimensional space Y = span(x0 ) define

f : Y → R, αx0 (→ αµC (x0 ).

Then f (y) ≤ µC (y) for all y ∈ Y (if y = αx0 , with α > 0 this follows from the
positive homogeneity of µC , and if α < 0 this is clear). By Theorem 1.4.2 we
can extend f to a linear function x∗ , defined on all of X, with x∗ (x) ≤ µC (x)
for all x ∈ X. Since µC is bounded on BX it follows that x∗ ∈ X ∗ .
1.4. THE HAHN BANACH THEOREM 21

Moreover, since x0 [1 − 4#xε 0 # ] /∈ D (otherwise dist(x0 , D) ≤ ε/4 and thus

dist(x0 , C) ≤ 2ε + dist(x0 , D) < ε), it follows that f (x0 ) = µD (x0 ) > 1. This
proves our claim in the case that K = R.
If K = C we first choose g, by considering X to be real Banach space, and
then put f (x) = g(x) − ig(ix). It is then easily checked that f is a complex
linear bounded functional on X.

Exercises

1. Prove Proposition 1.4.9 .

2. Let X be a Banach space with norm % ·% . Show that µBX = % ·% .

3.∗ Show that there is an x∗ ∈ %∗∞ so that

a) %x∗ % = 1,
b) *x∗ , x+ = limi→∞ xi , for x = (xi ) ∈ c = {(ξi ) : limi→∞ ξi exists}
c) If x = (ξi ) ∈ %∞ , and ξi ≥ 0, for i ∈ N, then *x∗ , x+ ≥ 0, and
d) If x = (ξi ) ∈ %∞ and x' = (ξ2 , ξ3 , . . .) then *x∗ , x' + = *x∗ , x+

4. Show that %1 is not isomorphic to a subspace of c0 .

22 CHAPTER 1. SOME BASIC BACKGROUND

1.5 Finite dimensional Banach spaces

Theorem 1.5.1. [Auerbach bases]
If X = (Kn , % · %) is an n-dimensional Banach space, then X has a basis
x1 , x2 , . . . xn for which there are functionals x∗1 , . . . x∗n ∈ X ∗ , so that

a) %xj % = %x∗j % = 1 for all j = 1, 2 . . . n,

b) for all i, j = 1, 2 . . . n
)
if i = j,
*x∗i , xj + = δ(i,j) = .
/ j.
if i =

We call in this case (xj , x∗j ) an Auerbach basis of X.

Proof. We consider the function

Det :X n = X X × X- → K,
* × +,
n times
(u1 , u2 , . . . un ) (→ det(u1 , u2 , . . . un ).

Thus, we consider ui ∈ Kn , to be column vectors and take for u1 , u2 , . . . un ∈

Kn the determinant of the matrix which is formed by vectors ui , for i =
1, 2, . . . n. Since (BX )n is a compact subset of X n with respect to the prod-
uct topology, and since Det is a continuous function on X n we can choose
x1 , x2 , . . . xn in BX so that
. . . .
.Det(x1 , x2 , . . . xn ). = max .Det(u1 , u2 , . . . un )..
u1 ,u2 ,...un ∈BX

By multiplying x1 by the the appropriate number α ∈ K, with |α| = 1, we

can assume that

Det(x1 , x2 , . . . xn ) ∈ R and Det(x1 , x2 , . . . xn ) > 0.

Define for i = 1, . . . n

Det(x1 , . . . xi−1 , x, xi+1 , . . . , xn )

x∗i : X → K, x (→ ,
Det(x1 , x2 , . . . xn )
1.5. FINITE DIMENSIONAL BANACH SPACES 23

It follows that x∗i is a linear functional on X (taking determinants is linear

in each column), and

*x∗i , xi + = 1,
. Det(x , . . . x , x, x , . . . , x ) .
. 1 i−1 i+1 n .
%x∗i % = sup |*x∗i , x+| = sup . .=1
x∈BX x∈BX Det(x 1 , x 2 , . . . x n )
(by the maximality of Det(x1 , x2 , . . . xn ) on (BX )n ),
Det(x1 , . . . xi−1 , xj , xi+1 , . . . , xn )
*x∗i , xj + = = 0 if i /= j, i, j ∈ {1, 2 . . . n}
Det(x1 , x2 , . . . xn )
(by linear dependence of columns)

which finishes our proof.

Corollary 1.5.2. For any two n-dimensional Banach spaces X and Y it

follows that
dBM (X, Y ) ≤ n2 .

Remark. Corollary 1.5.2 Is not the best result one can get. Indeed from the
following Theorem of John (1948) it is possible to deduce that for any two
n-dimensional Banach spaces X and Y it follows that

dBM (X, Y ) ≤ n.

Theorem 1.5.3. [John’s theorem]

Let X = (Kn , % ·% ) be an n-dimensional Banach space. Then there is an
invertible matrix T so that
√
B#2 ⊂ T (BX ) ⊂ nB#2 .

Theorem 1.5.4. For any Banach space X

X is finite dimensional ⇐⇒ BX is compact.

Proof. The implication “⇒” was already noted in the remark in Section 1.2
the implication “⇐” will follow from the following Proposition.

Proposition 1.5.5. The unit ball of every infinite dimensional Banach space
X contains a 1-seperated sequence.
24 CHAPTER 1. SOME BASIC BACKGROUND

Proof. By induction we choose for each n ∈ N an element xn ∈ Bx , so

that %xj − xn % ≥ 1, for j = 1, 2 . . . n − 1. Choose an arbitrary x1 ∈ SX .
Assuming x1 ,x2 ,. . . xn−1 has been chosen, let F = span(x1 , . . . xn−1 ), (the
space generated by xj , j = 1, 2 . . . , n − 1). X/F is infinite dimensional, thus
there is a z ∈ X so that

1 = %z%X/F = inf %z + y% = inf %z + y% = min %z + y%.

y∈F y∈F,#y#≤1+#z# y∈F,#y#≤1+#z#

We can therefore choose xn = z + y so that y ∈ F and

%z + y% = min %z + ỹ% = 1,
ỹ∈F,#ỹ#≤1+#z#

it follows that

1 = %xn %X/F ≤ %xn − xj % for all j = 1, 2 . . . n − 1.

Remark. With little bit more work (see Exercise 3) one can we find in the
unit ball of each infinite dimensional Banach space X a sequence (xn ) with
%xm − xn % > 1, for all m /= n in N.
A much deeper result by J. Elton and E. Odell (see [EO]) says that for
each Banach space X there is a ε > 0 and a sequence (xn ) ⊂ BX with
%xm − xn % ≥ 1 + ε, for all m /= n in N.
Definition 1.5.6. An operator T : X → Y is called a finite rank operator
if T (X) is finite dimensional. In this case we call dim(T (X)) the rank of T
and denote it by rk(T ).
For y ∈ Y and x∗ ∈ X ∗ we denote the operator

X → Y, x (→ y*x∗ , x+

by y ⊗ x∗ . Clearly, y ⊗ x∗ is of rank one.

Proposition 1.5.7. Assume that X and Y are Banach spaces and that
T : X → Y is a linear bounded operator of finite rank n. Then there are
x∗1 , x∗2 . . . , x∗n ∈ X and y1 , y2 . . . yn in Y so that
n
'
T = yj ⊗ x∗j .
j=1
1.5. FINITE DIMENSIONAL BANACH SPACES 25

Exercises:

1. Prove Corollary 1.5.2 using the existence of Auerbach bases. Prove the
claim in the following remark, by using John’s Theorem.

2.∗ Find two spaces X and Y which are not isometric to each other, but
for which dBM (X, Y ) = 1.

3.∗ Prove that in the unit ball of each infinite dimensional Banach space
X there is a sequence (xn ) with %xm − xn % > 1, for all m /= n in N.

4. Prove Proposition 1.5.7.

√
5.∗ For n ∈ N prove that dBM (%n2 , %n1 ) = dBM (%n2 , %n∞ ) = n.