MTL 411: Functional Analysis

Lecture B: Hahn-Banach theorem and its consequences

In this lecture, we discuss the Hahn-Banach theorem which is one of the major theorems
in the functional analysis. The theorem guarantees that every continuous linear functional
on a subspace can be extended to the whole space with norm conservation.

1 Hahn-Banach theorems
Theorem 1.1. Let M be a proper subspace of a real normed linear space X and
f : M → R be a continuous linear functional. Then there exists a continuous linear functional
fe : X → R such that

(i) the restricted functional fe|M = f , i.e., fe(x) = f (x), for all x ∈ M ;

(ii) ||fe|| = ||f ||.

Proof. The idea of the proof is to expand the space M step-by-step, and correspondingly
construct a continuous linear extension, then use the Zorn’s Lemma to complete the argument.
Without loss of generality, we assume that |f (x)| ≤ ||x|| for all x ∈ M and ||f || = 1 (why?).
Let z ∈ X \ M. For x, y ∈ M , consider
f (x − y) ≤ ||x − y|| ≤ ||x + z|| + || − y − z||
−||y + z|| − f (y) ≤ ||x + z|| − f (x).
Notice that in the above inequality, LHS is indepedent of x and RHS is indepedent of y.
Therefore, we get finite numbers s and t such that
s = sup (−||y + z|| − f (y)) ≤ ||x + z|| − f (x), for each x ∈ M
y∈M
s ≤ inf (||x + z|| − f (x)) = t. (1.1)
x∈M

Choose any number r satisfying s ≤ r ≤ t. Now consider the space

Mz = {x + αz : x ∈ M, α ∈ R}.
By the choice of z, every element in Mz has a unique representation of the form x + αz.
Define
h(x + αz) = f (x) + αr for x + αz ∈ Mz .
It is easy to see that h is linear and h|M = f. Further, it is to verify that
|h(w)| ≤ ||w||, ∀w ∈ Mz .
Indeed, for α > 0, we have from (1.1)
x x
r ≤ + z − f ( ) (∀x ∈ M )

α α
=⇒ f (x) + αr ≤ ||x + αz||
=⇒ h(x + αz) ≤ ||x + αz|| (∀x ∈ M ).

1
For other values of α, use the lower inequality of (1.1) and verify that

|h(x + αz)| ≤ ||x + αz||, ∀x + αz ∈ Mz .

Therefore, h is continuous linear on Mz . Since ||h|| ≤ 1 and h|M = f, we get ||h|| = ||f || = 1.
If Mz = X, then we are done. If not we repeat the above procedure with the subspace
Mz , then we get a subspace Mz0 and a continuous linear functional h0 such that Mz ⊂ Mz0 ,
h0 |Mz = h and ||h0 || = ||h||. In this manner, we can repeat the procedure but there is no
clue when/how to stop. So we apply the Zorn’s lemma carefully to guarantee a existence of
desired continuous linear functional on X.
Define

P = {(M̃ , h̃) : M ⊂ M̃ , h̃ is a continuous linear extension of f on M̃ , ||h̃|| = ||f ||}.

Observe that P is non-empty. Define a relation 6 on P:

we say (M 0 , h0 ) 6 (M 00 , h00 ) if M 0 ⊂ M 00 and h00 |M 0 = h0 .
It is easy to show that (P, 6) is a partial ordered set (POSET), i.e., the relation 6 is reflexive,
antisymmetric, and transitive.
Let {(Mλ , hλ ) : λ ∈ Λ} be a totally ordered set, where Λ is a indexing set. Then the pair
(M , b
c h) defined by
c = ∪λ∈Λ Mλ and b
M h(x) = hλ (x), x ∈ Mλ , λ ∈ Λ
is an upper bound for {(Mλ , hλ ) : λ ∈ Λ}. (why? Is (M h) in P? ).
c, b
Since every totally ordered set has an upper bound in P, the POSET (P, 6) has a maximal
f, fe) in P by the Zorn’s lemma.
element (M
Claim. Mf = X.
Suppose M f 6= X, then M f is a proper subspace of X. Then we can repeat the above
procedure and construct a pair (M f0 , fe0 ) ∈ P such that (M f0 , fe0 ). This contradicts
f, fe)) 6 (M
to the fact that (M
f, fe) is a maximal element.

The theorem is also true for complex normed linear spaces. Suppose X is a complex
normed linear space and f : X → C be (complex) continuous linear functional on X. Then

f (x) = u(x) + iv(x)

where u is real part of f and v is imaginary part of f . It is easy to show that

f (x) = u(x) − iu(ix)

and u : X → R is (real) continuous linear functional. (Recall, if X is vector space over C,

then X is also vector space over R). Moreover, we can state a general lemma.

Lemma 1.2. Let X be a complex normed linear space. Then f is a (complex) continuous
linear functional on X if and only if there exists a unique real continuous linear functional u
on X such that f (x) = u(x) − iu(ix) and ||f || = ||u||.

Proof. Observe that for x ∈ X, choose α ∈ C with |α| = 1 such that

|f (x)| = αf (x) = f (αx) = u(αx).

The remaining details are exercise.

From Theorem 1.1 and Lemma 1.2, we can conclude the following theorem.

2
Theorem 1.3. Let M be a proper subspace of a complex normed linear space X and f :
M → C be a (complex) continuous linear functional. Then there exists a continuous linear
functional fe : X → C such that

(i) the restricted functional fe|M = f , i.e., fe(x) = f (x), for all x ∈ M ;

(ii) ||fe|| = ||f ||.

Proof. By Lemma 1.2, choose the real part of f which is real continuous linear functional u
on M and ||f || = ||u||. Now apply Theorem 1.1 with the real continuous linear functional u
on M , then there exist a ue ∈ X 0 such that u
e|M = u and ||eu|| = ||u||. Again, using Lemma
e 0
1.2, we get f ∈ X such that the real part of f is u
e e and ||fe|| = ||e
u|| = ||u|| = ||f ||. Also
e(x) − ie
f (x) = u
e u(ix) = u(x) − iu(ix) = f (x), for x ∈ M.

2 Consequence of Hahn-Banach thereom

First application, we see that the dual space X 0 6= {0} if X 6= {0}.

Corollary 2.1. Let X be a normed linear space and x0 ∈ X a non-zero vector. Then there
are non-trivial continuous linear functionals on X. In particular, there exists a f ∈ X 0 such
that ||f || = 1 and f (x0 ) = ||x0 ||.

Proof. Define M = {αx0 : α ∈ K}. Define f : M → K by f (αx0 ) = α||x0 ||. It is clear that
f is continuous linear functional on M . Then by Hahn-Banach theorem, there exists fe ∈ X 0
such that fe|M = f on M and ||fe|| = ||f || = 1.

From the above corollary, we get the following results.

Corollary 2.2. Let X be a normed linear space. Then

|f (x)|
||x|| = sup .
06=f ∈X 0 ||f ||

Proof. For 0 6= f ∈ X 0 , we get

|f (x)|
|f (x)| ≤ ||f || ||x|| =⇒ ||x|| ≥
||f ||
|f (x)|
=⇒ ||x|| ≥ sup .
06=f ∈X 0 ||f ||

If x = 0, then the result is trivial. If x 6= 0, then the result follows by corollary 2.1.

Corollary 2.3. Let X be a normed linear space. If f (x) = 0 for all f ∈ X 0 , then x = 0.

Corollary 2.4. Let Y be a proper closed subspace of a normed space X. Let x0 ∈ X \ Y be

arbitrary and
δ = inf ||x0 − y|| = dist(x0 , Y ).
y∈Y

Then there exists a continuous linear functional fe on X 0 such that

fe|Y = 0, fe(x0 ) = δ, and ||fe|| = 1.

3
Proof. The proof is similar to corollary 2.1. Consider the subspace M = {y +αx0 : y ∈ Y, α ∈
K}. Define f : M → K by
f (y + αx0 ) = αδ.
Notice that the number δ > 0 (why?), hence f 6= 0. Now consider

|f (y + αx0 )| ≤ |α|δ = |α| inf ||x0 − y||, ∀α ∈ K

y∈Y
≤ |α|||x0 − y||, ∀y ∈ Y
≤ ||αx0 − αy||, ∀y ∈ Y
≤ ||αx0 + y||, ∀y ∈ Y , α ∈ K. (Why?)

Using the Hahn-Banach theorem, the result follows (verify).

Corollary 2.5. Let T : H → H be a bounded linear operator on a Hilbert space H. If

T = T ∗ , then ||T || = sup{| hT x, xi | : ||x|| = 1}.

Proof. Since the set {| hT x, xi | : ||x|| = 1} is bounded by ||T || (why?), the supremum α =
{| hT x, xi | : ||x|| = 1} exists and α ≤ ||T ||.
Let x, y ∈ H, and ||x|| = ||y|| = 1. Consider

hT (x ± y), x ± y)i = hT x, xi ± hT x, yi ± hT y, xi + hT y, yi
= hT x, xi ± hT x, yi ± hy, T ∗ xi + hT y, yi
= hT x, xi ± 2 Re hT x, yi + hT y, yi . (∵ T = T ∗ )

Thus
hT (x + y), x + y)i − hT (x − y), x − y)i = 4 Re hT x, yi .
From this, we get

4 Re hT x, yi ≤ α(||x + y||2 + ||x − y||2 ) (∵ hT z, zi ≤ α||z||2 )

≤ 2α(||x||2 + ||y||2 )
≤ 4α (∀||x|| = ||y|| = 1). (2.1)

Since | hT x, yi | = eiθ hT x, yi = T x, e−iθ y and ||e−iθ y|| = 1, we get



| hT x, yi | ≤ α, ∀||x|| = ||y|| = 1 [ Using (2.1)]

sup | hT x, yi | ≤ α,
||y||=1
||T x|| ≤ α, ∀||x|| = 1 [ Using corollary 2.2 ]
=⇒ ||T || ≤ α.

3 Bidual space
Let X be a normed linear space and its dual space X 0 . The dual space (X 0 )0 of X 0 is called
the bidual or second dual or double dual space of X. We write X 00 = (X 0 )0 .
Definition 3.1. A normed space X is said to be reflexive if X is isometrically isomorphic
to X 00 .
Remark. Recall that the dual space of a normed space is always complete. Hence if X is
reflexive then X must be complete. In other words, incomplete spaces are not reflexive.
Examples.

4
 1. We know that (`p )0 ' `q , 1 < p, q < ∞, 1
p + 1
q = 1. Hence the sequence space `p ,
1 < p < ∞, is reflexive.

2. The space `1 is not a reflexive space. We know that (`1 )0 ' `∞ . Now the question is
(`∞ )0 ' `1 ? Answer is NO. (USE THEOREM 3.3 BELOW)
Here we show that the natural isometry map is not onto:
First we can observe that for given y = (yk ) ∈ `1 , we can define a continuous linear
functional fy by
∞
X
fy (x) = xk yk , x = (xk ) ∈ `∞ (3.1)
k=1
|fy (x)| ≤ ||x||∞ ||y||1 and ||fy || = ||y||1 .

Therefore, we can define the natural isometry map T from `1 into (`∞ )0 by y 7→ fy . Now
we show that it is not surjective, that is, we construct a continuous linear functional
on `∞ which is not of the form (3.1). We make use of Hahn-Banach theorem in the
construction.
On the subspace c (the set of convergence sequences), we define a functional g by

g(x) = lim xk , x = (xk ) ∈ c

k→∞
|g(x)| ≤ ||x||∞ .

Using Hahn-Banach theorem, we get a continuous linear functional ge on `∞ such that

ge|c = g and ||e
g || = ||g||. Observe that ge(en ) = g(en ) = lim en (k) = 0 for all n ∈ N. If ge
k→∞
is of the form (3.1), then ge must be zero but ge 6= 0 continuous linear functional (why?).
Thus T is not surjective.

Now we show that a normed linear space X is isometrically isomorphic to a subspace of

X 00 .In some sense, X 00 is bigger space and contains X.

Theorem 3.2. Let X be a normed linear space. Then the space X is isometrically isomorphic
to a subspace of the bidual space X 00 .

Proof. For each (fixed) x ∈ X, define a functional Fx on X 0 by

Fx (f ) = f (x), f ∈ X 0 .

Observe that for each f in X 0 , we associate one scalar value f at the point x. It is easy to
verify that Fx is linear on X 0 . Consider

|Fx (f )| = |f (x)| ≤ ||f || ||x||, ∀f ∈ X 0 .

Thus Fx is continuous and ||Fx || ≤ ||x||. By corollary 2.1 to the Hahn-Banach theorem, there
exists a f ∈ X 0 such that f (x) = ||x|| and ||f || = 1, we get ||Fx || = ||x||. Therefore, the map
T : X → X 00 is defined by
T (x) = Fx
is a linear isometry and hence X is isometrically isomorphic to the subspace T (X) = X.
e

e is a subset of X 00 . If X
In general the identified subspace X e = X 00 , then X is reflexive.

Theorem 3.3. (Separability) If the dual space X 0 of a normed space X is separable, then X
itself is separable.

5
Proof. Recall that every subspace of a separable metric space is separable. Since X 0 is
separable, the unit ball U = {f ∈ X 0 : ||f || = 1} is separable subspace. Let {fn } be a
countable dense subset of U . Pick xn ∈ X such that ||xn || = 1 and |fn (xn )| ≥ 21 .
Consider M = Span{xn }. By definition M is separable. We show that M = X.
Suppose not, then M is a proper closed subspace of X. Then by corollary 2.4 to the
Hahn-Banach theorem, ∃ g ∈ X 0 such that ||g|| = 1 and g|M = 0. In particular, we have
g ∈ U and g(xn ) = 0 for all n ∈ N.
Consider
1
≤ |fn (xn )| = |fn (xn ) − g(xn )|
2
≤ ||fn − g|| ||xn || = ||fn − g||.

This is contradiction to the fact that {fn } is dense in U .

Using the above result, we can easily conclude that `1 is not reflexive (how?).
Problems.

1. Let X = C([a, b]) with the sup norm, and Y be the subspace of X consisting of all
constant functions. Let g ∈ Y 0 defined by g(y) = y(a), ∀y ∈ Y . Define
Z b
1
f (x) = x(t) dt, x ∈ X,
b−a a

then show that f is a Hahn-Banach extension of g.

2. Riesz Lemma: Let X be a norm linear space, and Y be a proper closed subspace of
X. Let r be a real number such that 0 < r < 1. Then there exist some xr ∈ X such
that
kxr k = 1, and r < dist(xr , Y ) ≤ 1.
Show that Riesz Lemma with r = 1 holds in X if and only if every f ∈ X 0 attains its
norm on the unit sphere of X.

3. Let X be a norm linear space and {xn } be a sequence in X. A sequence {xn } is said
to be weakly convergent if ∃ x ∈ X such that x0 (xn ) → x0 (x) for every x0 ∈ X 0 , and
ω
→ x. Let {en } be a standard basis for `p , where 1 ≤ p ≤ ∞. Show
denote it by xn −
ω
→ 0 in `p , for 1 < p ≤ ∞. Further, show that en 9 0 weakly in `1 .
that en −

4. Let c be a set of all convergent sequence, and y := (y1 , y2 , , · · · ) ∈ `1 . Define fy , gy by

∞
X
fy (x) := x n yn , x∈c
n=1

and
∞
X
gy (x) := y1 lim xn + xn yn+1 , x ∈ c.
n→∞
n=1

Then show that fy , gy ∈ c0 and kfy k = kyk1 = kgy k. If F, G : `1 → c0 defined by

F (y) = fy and G(y) = gy , then F and G are isometrices from `1 to c0 . Further, show
that the map G is surjective, but F is not.

5. Let X be a Hilbert space and for f ∈ X 0 , let uf ∈ X be the unique element obtained
0
as in Riesz representation theorem. For f, g ∈ X 0 , let hf, gi = hug , uf i. Prove the
following.

6
 0
(a) h·, ·i is an inner product on X 0 ,
0
(b) X 0 is a Hilbert space with respect to the inner product h·, ·i .

6. Let X and Y be a norm linear space, and F be bounded linear operator from X to Y .
Transpose of F is denoted by F 0 and F 0 : Y 0 → X 0 is defined by F 0 (y 0 ) = y 0 ◦ F .

(a) Show that kF 0 (y 0 )k ≤ kF kky 0 k, ∀ y0 ∈ Y 0.

(b) If there is some γ > 0 such that for every y ∈ Y , there is a sequence {xn } in X
k·kY
with F (xn ) −−−→ y, and kxn kX ≤ γkykY , then show that ky 0 k ≤ γkF 0 (y 0 )k.

7. Let H1 and H2 be Hilbert space and T : H1 → H2 a bounded linear operator. If

M1 ⊂ H1 , and M2 ⊂ H2 , then T (M1 ) ⊂ M2 if and only if T ? (M2⊥ ) ⊂ M1⊥ .

8. Let H be a Hilbert space, and A be a self-adjoint bounded linear operator on H. Show

that A2 ≥ 0 and A ≤ kAkI. Also, if A2 ≤ A, then prove 0 ≤ A ≤ I.
(Note: An operator P on Hilbert space H is said to be positve if hP x, xi ≥ 0, ∀x ∈ H)

9. Suppose C([−1, 1]) is the vector space of countinuous real-valued functions on the in-
terval [−1, 1] with inner product given by
Z 1
hf, gi = f (x)g(x)dx
−1

for f, g ∈ C([−1, 1]). Let φ be the linear functional on C([−1, 1]) defined by φ(f ) =
f (0). Show that there does not exist g ∈ C([−1, 1]) such that φ(f ) = hf, gi for every
f ∈ C([−1, 1]).

10. Consider a space H 1 ([0, 1]) = {f : f is absolutely continuous with f (0) = 0, f 0 ∈

L2 ([0, 1])} and it is a Hilbert space with respect to the inner product
Z 1
hf, gi = f 0 (x)g 0 (x)dx
0

for every f, g ∈ H 1 ([0, 1]). Show that the point evalutation functional Tx : H 1 ([0, 1]) →
C, defined by Tx (f ) = f (x) is continuous. Find the corresponding Riesz element, i.e.
Kx ∈ H 1 ([0, 1]) such that Tx (f ) = hf, Kx i.