Functional Analysis and Applications

Hahn Banach Theorem

K. Sreenadh
Indian Institute of Technology Delhi

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 1 / 22

Dual spaces
Definition
Let X be a normed linear space. A bounded linear map from X → R is
called functional. The set of all functionals is called the dual of X ,
denoted by X ∗ OR X ′ .

Example
Pn
X = Rn . We may define the functional f on X as , for x = i=1 ci ei

n
X
f (x) = ci f (ei )
i=1

n
X X 1 X 1
Indeed, |f (x)| = | ci f (ei )| ≤ ( ci2 ) 2 ( f (ei )) 2
i=1
X 1
≤ ∥x∥A, where A = ( f (ei )2 ) 2
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 2 / 22
Dual spaces
Theorem
The dual of Rn is isomorphic to Rn

Proof: Let M bePthe dual of Rn . From Linear algebra, we see that if

f ∈ M, and x = ni=1 xi ei implies
n
X
f (x) = xi f (ei ),
i=1
By Cauchy-Scwartz inequality,
X 1
=⇒ |f (x)| ≤ ( f (ei )2 ) 2 ∥x∥
X
=⇒ ∥f ∥ ≤ ( f (ei )2 )1/2
The equality in Cauchy-Scwartz inequality is achieved if
x = (f (e1 ), f (e2 ), ...f (en )), in this case,
X X X
|f (x)| = f (ei )2 = ( f (ei )2 )1/2 ( f (ei )2 )1/2
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 3 / 22
 X
∴ ∥f ∥ = ( f (ei )2 )1/2 (1)

That is for every f ∈ (Rn )′ we get γ ∈ Rn such that

∥f ∥ = ∥γ∥

Now we define a map ϕ : M → Rn as

ϕ(f ) = (f (e1 ), f (e2 ), ...f (en ))

Then ϕ is linear transformation. Easy to see that it is 1 − 1.

ONTO: For every ξ ∈ Rn , we can get f ∈ M as
X
f (x) = ξk xk .

Isometry: follows from (1). Therefore

(Rn )′ = Rn

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 4 / 22

Theorem
The dual of ℓ1 is ℓ∞ .
Proof: Let f ∈ (ℓ1 )′ . Then for x = {xn } and the Schauder basis {en }, we
can write
X∞
x= xk ek
k=1

|f (ek )| ≤ ∥f ∥∥ek ∥ = ∥f ∥ (2)

n
X Xn
continuity of f =⇒ f (x) = f ( lim xk ek ) = lim f ( xk ek )
n→∞ n→∞
k=1 k=1
n
X ∞
X
= lim xk f (ek ) = xk f (ek )
n→∞
k=1 k=1
X
∴ |f (x)| ≤ (sup |f (ek )|) |xk | = ∥x∥1 sup |f (ek )|
=⇒ ∥f ∥ ≤ sup |f (ek )| (3)
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 5 / 22
From (2) and (3) we get

∥f ∥ = sup |f (ek )|

Hence the map f 7→ {f (ek )} is an isometry. Also given a sequence

{βk } ∈ ℓ∞ , we define g ∈ (ℓ1 )′ as
X
g (x) = xk βk

Then X
|g (x)| ≤ (sup |βk |) |xk | = ∥x∥1 sup(|βk ).

=⇒ g ∈ (ℓ1 )′
Hence f 7→ {f (ek )} is an onto isomometry.
1 1
Exercise: For 1 < p < ∞, the dual of ℓp is ℓq where p + q = 1.

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 6 / 22

The Second Dual
The second dual X ′′ , the dual of X ′ is also a Banach space. Moreover,
for each x ∈ X , we can define x̂ ∈ X ′′ as
x̂(f ) = f (x) ∀f ∈ X ′
Then
|x̂(f )| = |f (x)| ≤ ∥f ∥∥x∥
=⇒ ∥x̂∥ ≤ ∥x∥
that is the embedding (identity map) from X into X ′′ : x 7→ x̂ is
continuous. This is called natural embedding
Later in this lecture we will show show that
∥x̂∥ = ∥x∥

Definition
If X = X ′′ , that is the natural embedding is ONTO, we say X is Reflexive
space.
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 7 / 22
Extensions
We are interested in the following
1 Is X ′ non-empty?
2 Is X ′ rich, in the sense that if x ̸= y , does there exists f ∈ X ′ such
that f (x) ̸= f (y ).
For example, if we take x0 ∈ X and M = span(x0 ). Then we may define
f ∈ M ′ as
f (x) = α, x = αx0
Our questions can be answered if we can extend this f to the whole of X .
The answer is affirmative and is the Hahn-Banach theorem. For, we need
Definition
A function p : X → R is called semi-linear functional if

p(x + y ) ≤ p(x) + p(y ) ∀x, y ∈ X

p(αx) = αp(x), ∀α ≥ 0, x ∈ X .
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 8 / 22
Lemma
Let p be a semi-liner functional on X . Let Y be a subspace of X and
f ∈ Y ′ such that
f (x) ≤ p(x)
Suppose x0 ∈ X \Y , then f can be extended as g ∈ Z ′ = [span(Y , x0 )]′
such that
g (x) = f (x), ∀x ∈ Y
g (z) ≤ p(z), ∀z ∈ Z .

Proof: Let Z = span(Y , x0 ). If g ∈ Z ′ such that g = f on Y , and g ≤ p

on Z . Theng is of the form
g (z) = g (x + αx0 ) = g (x) + αg (x0 ) = f (x) + cα, c = g (x0 )
g (z) ≤ p(z) =⇒ f (x) + cα ≤ p(x + αx0 ), ∀α ∈ R
That is, we need to find c such that g (z) = f (x) + cα and
f (x) + αc ≤ p(x + αx0 ), ∀α (4)
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 9 / 22
If α > 0, this (4) is same as
1 1 x x
c≤ p(x + αx0 ) − f (x) = p( + x0 ) − f ( ), ∀x ∈ Y
α α α α
That is, (by subspace property)
c ≤ p(x + x0 ) − f (x), ∀x ∈ Y
If α < 0, (4) is same as

1 1
p(x + αx0 ) − f (x)
c≥
α α
x x
= −p(− − x0 ) − f ( )
α α
That is, (by subspace property)
c ≥ −p(−x − x0 ) − f (x), ∀x ∈ Y
From the above two assertions, we need to find c such that
− p(−x − x0 ) − f (x) ≤ c ≤ p(x + x0 ) − f (x) (5)
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 10 / 22
to get such c, let x, y ∈ Y

f (y ) − f (x) = f (y − x) ≤ p(y − x)
= p(y + x0 − x − x0 )
≤ p(y + x0 ) + p(−x − x0 )

That is,
−f (x) − p(−x − x0 ) ≤ p(y + x0 ) − f (y )
LHS is independent of y and RHS is independent of x. So we take
supremum on LHS and infimum on RHS Let

s = sup{−f (x) − p(−x − x0 )} t = inf{p(y + x0 ) − f (y )}

Then s, t ∈ R and s ≤ t.
So we can choose c such that

s≤c ≤t

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 11 / 22

Hahn-Banach Theorem
Theorem
Let p be a semi-liner functional on X and let M be a proper subspave of
X . Suppose f ∈ M ′ such that f (x) ≤ p(x). Then ∃g ∈ X ′ such that

g (x) = f (x) and g (x) ≤ p(x) ∀x ∈ X .

Proof: Let C be the collection of all pairs (g , Z ) where Z is a subspace

containing M and g ∈ Z ′ such that g ≤ p on Z .
By previous lemma, C is non-empty.
Now we define the order relation on C:
(g1 , Z1 ) ≤ (g2 , Z2 ) if Z1 ⊂ Z2 , and g1 = g2 on Z1
Now consider an ordered chain
{(gi , Zi ), i ∈ I }
i.e., for every pair i, j ∈ I ,
Either (gi , Zi ) ≤ (gj , Zj ) OR (gi , Zi ) ≥ (gj , Zj )
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 12 / 22
Let us consider the subspace
[
Z= Zi
i∈I

and define g on Z as
g = gj on Zj
Then g ∈ Z ′ and

g (x) = f (x), x ∈ M and g (x) ≤ p(x), x ∈ Z

That is equivalent to saying

(g , Z ) ∈ C and (g , Z ) ≥ (gi , Zi ), ∀i ∈ I

Therefoe every ordered chain in C has an upper bound. Then we use

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 13 / 22

Theorem
Zorn’s Lemma
If every chain has an upper bound in a partially ordered set, then there
exists a maximal element.
Hence there exists a maximal element (say) (g , Z ) ∈ C.
We claim: Z = X
Suppose not. Then ∃x0 ∈ X and x0 ̸∈ Z . Then by the previous lemma, we
can extend g to Y = span(Z , x0 ) as h. Then

(h, Y ) ∈ C, and (h, Y ) ≥ (g , Z )

This contradicts the maximality of (g , Z ).

Hence Z = X and this proves the theorem.

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 14 / 22

Theorem
Let Y be a subspace of X and let f ∈ Y ′ . Then ∃g ∈ X ′ such that g = f
on Y and ∥f ∥ = ∥g ∥.

Proof: Define semilinear functional p : X → R as

p(x) = ∥f ∥∥x∥

Then f (x) ≤ p(x). By Hahn-Banach theorem, ∃g ∈ X ′ such that

g (x) ≤ p(x) = ∥f ∥∥x∥

Also,

g (−x) ≤ p(−x) =⇒ −g (x) ≤ p(−x)

=⇒ g (x) ≥ −p(−x)
=⇒ g (x) ≥ −∥f ∥∥x∥

Therefore
|g (x)| ≤ ∥f ∥∥x∥ =⇒ ∥g ∥ ≤ ∥f ∥ (6)
K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 15 / 22
On the other side we also have

∥f ∥ = sup{|f (y )| : ∥y ∥ ≤ 1}
≤ sup{|g (x)| : ∥x∥ ≤ 1}
= ∥g ∥ (7)

From (6) and (7), we get ∥f ∥ = ∥g ∥

Theorem
Let X be a normed space. If x ∈ X , then ∃ f ∈ X ′ such that ∥f ∥ = 1 and
f (x) = ∥x∥. In particular, we say X ′ separates points of X .

Proof: If x = 0 then any f ∈ X ′ will do.

If x ̸= 0 then by Hahn-Banach theorem, X ′ is non-empty.
Let
Y = {αx : α ∈ R} and g (αx) = α∥x∥
Then g ∈ Y ′ and ∥g ∥ = 1.

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 16 / 22

By Hahn Banach theorem, there exists extension f of g such that

∥f ∥ = ∥g ∥ = 1 and f (x) = g (x) on Y .

To see that X ∗ separates X ,

Let x, y ∈ X such that x ̸= y . By above arguments, there exists f ∈ X ′
with
f (x) − f (y ) = f (x − y ) = ∥x − y ∥ =
̸ 0.
Hence
f (x) ̸= f (y ).

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 17 / 22

Theorem
X ′ separates a proper subspace of X and a point outside this subspace.

Proof: Let Y be a proper subspace of X and let x0 ̸∈ Y .

Then ∃r > 0 such that

∥x − x0 ∥ > r , ∀x ∈ Y .

Consider the subspace Z = span(Y , x0 ). Define f : Z → R by

f (x + αx0 ) = α.

Then f is linear and

f (x) = 0 ∀x ∈ Y and f (x0 ) = 1. (8)

We claim that f is also bounded. Then by Hahn Banach theorem this f

can be extended to X and (8) holds for the extension.

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 18 / 22

Claim: f ∈ Z ′
We note that
x x
∥x − x0 ∥ > r , ∀x ∈ Y =⇒ ∥ − x0 ∥ > r (as ∈ Y)
−α −α
x
=⇒ ∥ + x0 ∥ > r
α
Therefore,
x
r |f (x + αx0 )| = r |α| ≤ |α|∥ + x0 ∥
α
= ∥x + αx0 ∥

1
∴ |f (x + αx0 )| ≤ ∥x + αx0 ∥
r
1
=⇒ ∥f ∥ ≤ =⇒ f ∈ Z ′
r

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 19 / 22

Natural embedding
Let X ′′ be the dual of X ′ . Then for each x ∈ X , we define x̂ ∈ X ′′ as

x̂(f ) = f (x)

Then
∥x̂∥ ≤ ∥x∥
By Hahn-Banach theorem, ∃f ∈ X ′ such that

f (x) = ∥x∥, ∥f ∥ = 1.

therefore for such f

∥x∥ = f (x) = |x̂(f )| ≤ ∥∥x̂∥∥f ∥ = ∥x̂∥

=⇒ ∥x∥ ≤ ∥x̂∥
∴ x 7→ x̂ : X → X ′′ is an isometry

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 20 / 22

Note that X ′′ is always Banach space for any Normed space X . Hence the
embedding need not be Onto.
Definition
If X is a Banach space such that the embedding of X into X ′′ is onto.
Then we say X is Reflexive Banach space.

Definition
X ′′ is called completion of X

Then we have the following ”dual of UBP”

Theorem
Let A ⊂ X such that {f (x) : x ∈ A} is bounded subset of R for each
f ∈ X ′ . Then A is bounded.

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 21 / 22

Proof: Recall UBP:

{Tx, T ∈ A ⊂ X ′ } isbounded =⇒ A is bounded in X ′

By the natural embedding A may be considered as subset of X ′′ as
 ⊂ X ′′ .

{f (x) : x ∈ A} is bounded =⇒ {x̂(f ), x̂ ∈ Â} is bounded

Hence by UBP, Â is bounded in X ′′ .
By isometry x 7→ x̂ we get

A is bounded in X

K. Sreenadh (IIT Delhi ) Functional Analysis and Applications 22 / 22