MAA 6617 Course Notes Spring 2025

1. Normed vector spaces

In this section F stands for either R or C. Let X be a vector space over F.

1.1. Definitions and preliminary results.

Definition 1.1. A normed vector space X = (X , ∥ · ∥) consists of a vector space X over
F together with a norm ∥ · ∥ : X → [0, ∞) (see definition 5.1 – which does not change
with R replaced by C). We often denote the normed vector space as X , with the norm
∥ · ∥ implicit.

As we noted before, using the properties of a norm, it is straightforward to check

that d : X × X → [0, ∞) defined by
d(x, y) := ∥x − y∥

T
is a metric on X . The resulting topology is the norm topology and it is the default
topology on X .
AF
Definition 1.2. A normed vector space X is a Banach space if it is complete (with its
norm topology). □

Definition 1.3. Two norms ∥ · ∥1 , ∥ · ∥2 on X are equivalent if there exist constants

C, c > 0 such that
c∥x∥1 ≤ ∥x∥2 ≤ C∥x∥1 ,
DR

for all x ∈ X . □
Remark 1.4. Equivalent norms determine the same topology on X and the same
Cauchy sequences (Problem 1.4). In particular, it follows that if X is equipped with
two equivalent norms ∥ · ∥1 , ∥ · ∥2 then it is complete (a Banach space) in one norm if
and only if it is complete in the other.
Equivalence of norms is an equivalence relation on the set of norms on X . □

The next proposition is simple but fundamental; it says that the norm and the
vector space operations are continuous in the norm topology.
Proposition 1.5 (Continuity of vector space operations). Let X be a normed vector
space over F.
a) If (xn ) converges to x in X , then (∥xn ∥) converges to ∥x∥ in R.
b) If (kn ) converges to k in F and (xn ) converges to x in X , then (kn xn ) converges
to kx in X .
c) If (xn ) converges to x and (yn ) converges to y in X , then (xn + yn ) converges to
x + y in X .
1
2

Proof. The proofs follow readily from the properties of the norm, and are left as exercises.
□

The following proposition gives a convenient criterion for a normed vector space to
be complete.

Definition 1.6. Given a sequence (xn ) from a normed vector space X , the expression
P PN
n=1∞ xn denotes the sequence (sN = n=1 xn ), called the sequence of partial sums of
the series. The series converges Pif the sequence of partial sums converges in the norm
topology. In this case we use n=1∞ xn to also denote the limit of this sequence and
call it the sum.
Explicitly, the series ∞
P
n=1 xn converges means there is an x ∈ X such that for each
ϵ > 0 there is an N such that ∥sn − x∥ < ϵ for all n ≥ N.
The series ∞
P P∞
n=1 xn converges absolutely if the series n=1 ∥xn ∥ converges (in the
normed vector space (R, | · |)). □

Proposition 1.7. A normed space (X , ∥ · ∥) is complete if and only if every absolutely

convergent series in X is convergent.

Before proving the Proposition we collect two lemmas. A definition is needed for
the first.

P 1.8. A sequence (yk ) from a normed vector space X is super-cauhcy if the

Definition
series ∞k=1 (yk+1 − yk ) converges absolutely.

Lemma 1.9. If (xn ) is a Cauchy sequence from a normed vector space X , then there is
a subsequence (yk ) of (xn ) that is super-cauchy.

Proof. With ϵ = 21 , there exists an N1 such that ∥xn − xm ∥ < 12 for all m, n ≥ N1 since
(xn ) is Cauchy. Assuming N1 < N2 < · · · < Nk have been chosen so that ∥xn − xm ∥ < 21j
1
for 1 ≤ j ≤ k, there is an Nk+1 < Nk such that ∥xn − xm ∥ < 2k+1 since (xn ) is Cauchy.
Hence by recursion we have constructed a (strictly)increasing sequence of integers Nk
such that ∥xn −xm ∥ < 21k for all m, n ≥ Nk . Set yk = xNk and note that ∥yk+1 −yk ∥ < 21k ,
from which it follows that (yk ) is a super-cauchy subsequence of (xn ). □

The proof will also use the following standard lemma from advanced calculus.

Lemma 1.10. If (xn ) is a Cauchy sequence from a metric space (X, d) and if (xn ) has
a subsequence (yk ) that converges to some x, then (xn ) converges to x.

Proof of Proposition 1.7. First suppose X is complete and ∞

P
PN n=1 xn is absolutely con-
th
vergent. Write sN = n=1 xn for the N partial sum and let ϵ > 0 be given. Since
P∞ P∞
n=1 ∥xn ∥ is convergent, there exists an L such that n=L ∥xn ∥ < ϵ. If N > M ≥ L,
 3

then
N
X N
X
∥sN − sM ∥ = xn ≤ ∥xn ∥ < ϵ.
n=M +1 n=M +1
Thus the sequence (sN ) is Cauchy in X , hence convergent by the completeness hypoth-
esis.
Conversely, suppose every absolutely convergent series in X is convergent and that
(xn ) is given Cauchy sequence from X. By Lemma 1.9 there is a super-cauchy subse-
quence (yk ) of (xn ). Since (yk ) is super-cauchy, the series ∞
P
k=1 (yk+1 − yk ) is absolutely
convergent and hence, by hypothesis, convergent in X . Thus there is an z ∈ X such
that the sequence of partial sums
X n
(yk+1 − yk ) = yn+1 − y1
k=1

converges to z. Rearranging, (xNn+1 = yn+1 ) converges to x = z+y1 . Hence (xn ) is Cauchy

and has a convergent subsequence. Thus (xn ) converges (to x) by Lemma 1.10. □

1.2. Examples.

1.2.1. Euclidean space. Observe that the Euclidean norm on the complex vector space
Cn agrees with the Euclidean norm on the real vector space R2n (via that natural real
linear isomomorphism R2 → C sending (x, y) to x + iy). Thus, Fn with the usual
1/2
Euclidean norm ∥(x1 , . . . xn )∥ = ( nk=1 |xk |2 ) is a Banach space.
P

The vector space Fn can also be equipped with the ℓp -norms

n
!1/p
X
p
∥(x1 , . . . xn )∥p := |xk |
k=1
∞
for 1 ≤ p < ∞, and the ℓ -norm
∥(x1 , . . . xn )∥∞ := max(|x1 |, . . . |xn |).
For 1 ≤ p < ∞ and p ̸= 2, it is not immediately obvious that ∥ · ∥p defines a norm.
We will prove this assertion later. It is not too hard to show that all of the ℓp norms
(1 ≤ p ≤ ∞) are equivalent on Fn (though the constants c, C depend on the dimension
n). For instance, for n ∈ N,
∥x∥∞ ≤ ∥x∥2 ≤ ∥x∥1 ≤ n∥x∥∞ .
The first and third inequalities are evident. For the middle inequality, observe
Xn n
X
2
(∥x∥1 ) = |xj | |xk | ≥ |xj |1 = ∥x∥22 .
j,k=1 j=1

Given a normed vector space X = (X , ∥ · ∥), denote its closed unit ball by
X1 = {x ∈ X : ∥x∥ ≤ 1}.
4

It is instructive to sketch the closed unit ball in R2 with the three norms above.
It turns out that any two norms on a finite-dimensional vector space are equivalent.
As a corollary, every finite-dimensional normed space is a Banach space. See Problem 1.5.
Lemma 1.11. If ∥·∥1 and ∥·∥2 are norms on X and there is a constant C > 0 such that
∥x∥1 ≤ C∥x∥2 for all x ∈ X , then the mapping ι : (X , ∥ · ∥2 ) → (X , ∥ · ∥1 ) is (uniformly)
continuous.

Proof. For x, y ∈ X , we have ∥ι(x) − ι(y)∥1 = ∥ι(x − y)∥1 = ∥x − y∥1 ≤ C∥x − y∥2 . □
Proposition 1.12. If ∥x∥ is a norm on Rn , then ∥x∥ is equivalent to the Euclidean
norm ∥ · ∥2 .

Sketch of proof. Let {e1 , . . . , en } denote the usual basis for Rn . Given x = aj ej ∈ Rn ,
P
X X
∥x∥ ≤ |aj | ∥ej ∥ = |aj |∥ej ∥ ≤ M ∥x∥1 ≤ n M ∥x∥2 ,
where M = max{∥e1 ∥, . . . , ∥en ∥}. It now follows that the map ι : (Rn , ∥ · ∥2 ) → (Rn , ∥ · ∥)
is continuous and therefore so is the map f : (Rn , ∥ · ∥2 ) → [0, ∞) defined by f (x) = ∥x∥.
Since
S n−1 = {x ∈ Rn : ∥x∥2 = 1}
(the unit sphere) is compact in Rn , by the Extreme Value Theorem, f attains its infimum;
that is, there is a point p ∈ S n−1 such that f (p) ≤ f (x) for all x ∈ S n−1 . But f (p) =
∥p∥ > 0 since p ̸= 0. Let c = f (p) = ∥p∥. We conclude that if ∥x∥2 = 1 then ∥x∥ ≥ c∥x∥2 ,
from which it follows by homogeneity that ∥x∥ ≥ c∥x∥2 for all x ∈ Rn . □
Corollary 1.13. All norms on a finite dimensional vector space are equivalent. Further,
if V is a finite dimensional normed vector space, then V1 is compact and V is a Banach
space.

Proof. Suppose V is a normed vector space of dimension n and let {v1 , . . . , vn } denote
a basis for V . The function ∥ · ∥′ : V → [0, ∞) defined by
X X
∥v∥′ = ∥ aj vj ∥′ = |aj |
is easily seen to be a norm.
Now let ∥ · ∥ be a given norm on V. This norm induces a norm ∥ · ∥∗ on Rn given by
X X
∥ aj ej ∥∗ = ∥ aj vj ∥.
Since all norms in Rn are equivalent, the norm ∥ · ∥∗ is equivalent to the norm ∥ · ∥1 .
Hence there exist constants 0 < c < C such that
X X X X X
c∥v∥′ = c |aj | = c∥ aj ej ∥1 ≤ ∥ aj ej ∥∗ ≤ C∥ aj ej ∥1 = C |aj | = C∥v∥′ .
P P
Thus, as ∥ aj ej ∥∗ = ∥ aj vj ∥,
c∥v∥′ ≤ ∥v∥ ≤ C∥v∥′
 5

for all v ∈ V. Thus all norms on V are equivalent.

Further, by definition, f : (V, ∥ · ∥) → (Rn , ∥ · ∥∗ ) is bijective and isometric. Thus,
f −1 is continuous, f −1 (S) where S is the unit ball in (Rn , ∥·∥∗ , is the unit ball in (V, ∥·∥)
and is compact as its the continuous image of a compact set. It is now routine to pass
from compactness of the unit ball in (V, ∥ · ∥) to completeness of (V, ∥ · ∥). □

1.2.2. The Banach space of bounded functions. If V is a vector space over F and ∅ ̸= T
is a set, then F (T, V ), the set of functions f : T → V is a vector space over F under
pointwise operations; e.g., if f, g ∈ F (T, V ) then f + g : T → V, is the function defined
by (f + g)(t) = f (t) + g(t).
Definition 1.14. A subset R of a normed vector space X is bounded if there is a C
such that ∥x∥ ≤ C for all x ∈ R; that is, R ⊆ CX1 .
A function f : T → X is bounded if f (T ) ⊆ X is bounded.
Let Fb (T, X ) denote the vector space (subspace of F (T, X )) of bounded functions
f : T → X.
Remark 1.15. The function ∥ · ∥∞ : Fb (T, X ) → [0, ∞) defined by
∥f ∥∞ = sup{|f (t)| : t ∈ T }
is a norm on Fb (T, X ) as you should verify. Let d∞ denote the resulting metric:
d∞ (f, g) = ∥f − g∥∞ .
Note that convergence of a sequence in the metric space (Fb (T, X ), d∞ ) is uniform
convergence; in particular, a sequence is Cauchy in Fb (T, X ) if and only if it is uniformly
Cauchy. (Exercise.)
Proposition 1.16. If X is a Banach space, then Fb (T, X ) is also Banach space.

Proof. We are to show Fb (T, X ) is complete, assuming X is complete. Accordingly,

suppose (fn ) is a Cauchy sequence from Fb (T, X ) and X is complete. In particular,
given ϵ > 0 there is an N such that d∞ (fn , fm ) = sup{∥fn (t) − fm (t)∥ : t ∈ T } < ϵ. It
follows that, for each s ∈ T, the sequence (fn (s)) is a Cauchy in X and hence converges
to some x ∈ X . Define f : T → X by f (s) = x. It remains to see that f is bounded and
(fn ) converges to f.
Since Cauchy sequences are bounded and (fn ) is Cauchy in the metric space Fb (T, X ),
there is a C such that
sup{∥fn (t)∥ : t ∈ T } = d∞ (fn , 0) ≤ C
for all n. It follows from Proposition 1.5 that (∥fn (t)∥)n converges to |f (t) and hence
∥f (t)∥ ≤ C for all t ∈ T. Thus f is bounded; that is f ∈ Fb (T, X ).
It only remains to show that (fn ) converges to f in Fb (T, X ). To do so let ϵ > 0
be given. There is an N such that if m, n ≥ N, then ∥fn (t) − fm (t)∥ < ϵ for all t ∈ T.
6

Given s ∈ T, there is an M ≥ N such that ∥fm (s) − f (s)∥ < ϵ for all m ≥ N. Since,
(fm (s) − fn (s))m converges (with m) to (f (s) − fn (s)) in X , another application of
Proposition 1.5 gives (∥fm (s) − fn (s)∥)m converges to ∥f (s) − fn (s)∥. Thus
∥f (s) − fn (s)∥ ≤ ϵ,
for all s ∈ T. Hence d∞ (f, fn ) = ∥f − fn ∥ ≤ ϵ and the proof is complete. □

There are important Banach spaces of continuous functions. Before going further,
we remind the reader of the following result from advanced calculus.
Theorem 1.17. Suppose X, Y are metric spaces, (fn ) is a sequence fn : X → Y and
x ∈ X. If each fn is continuous at x and if (fn ) converges uniformly to f , then f is
continuous at x. Hence if each fn is continuous, then so is f.

Proof. Let x and ϵ > 0 be given. Choose N such that if n ≥ N and y ∈ X, then
dY (fn (y), f (y)) < ϵ. Since fN is continuous at x, there is a δ > 0 such that if dX (x, y) < δ,
then dY (fN (x), fN (y)) < ϵ. Thus, if dX (x, y) < δ, then
dY (f (x), f (y)) ≤ dY (f (x), fN (x)) + dY (fN (x), fN (y)) + dY (fN (y), f (y))
< 3ϵ,
proving the theorem. □

Given a normed vector space Y, let Cb (X, Y) denote the subspace of Fb (X, Y) con-
sisting of continuous functions. Since uniform convergence is the same as convergence in
the normed vector space (Fb (X, Y), d∞ ), by Theorem 1.17, Cb (X, Y) is a closed subspace
of Fb (X, Y). In particular, in the case Y is a Banach space, so is Cb (X, Y).
When X be a compact metric space, let C(X) = C(X, F) denote the set of con-
tinuous functions f : X → F. Thus C(X) is a subspace of Fb (X, F) and we endow
C(X) with the norm it inherits from Fb (X, F). Since F is complete, C(X) is a Banach
space. Of course here we could replace F by a Banach space X and obtain the analogous
conclusion for the space C(X, X ).
Now let X be a locally compact metric space. In this case, a function f : X → F
vanishes at infinity if for every ϵ > 0, there exists a compact set K ⊆ X such that
/ |f (x)| < ϵ. Let C0 (X) denote the subspace of Fb (X, F) consisting of continuous
supx∈K
functions f : X → F that vanish at infinity. Then C0 (X) is a normed vector space with
the norm it inherits from C(X) (equivalently Fb (X, F). It is routine to check that C0 (X)
is complete.

1.2.3. L1 spaces over R. Let (X, M , µ) be a measure space and let L1 (µ) denote the
(real) vector space of (real-valued) absolutely integrable functions on X. The function
∥ · ∥1 : L1 (µ) → [0, ∞) defined by
Z
∥f ∥1 := |f | dm
X
 7

for f ∈ L1 (µ) is a norm on L1 (µ), provided we agree to identify f and g when f = g a.e.
(Indeed the chief motivation for making this identification is that it makes ∥ · ∥1 into a
norm.
Proposition 1.18. The real vector space L1 (µ) is a Banach space.

We will construct a complex vector space analog of L1 (µ) a bit later.

Proof. It suffices to verify the hypotheses of Proposition 1.7. Accordingly suppose

P∞ P∞
n=1 fn is absolutely convergent (so that n=1 ∥fn ∥1 < ∞). By Tonelli’s summation
theorem,
Z X∞ ∞ Z
X X∞
|fn | dm = |fn | dm = ∥fn ∥1 < ∞.
n=1 n=1 n=1
P∞ 1
Thus the function g := n=1 |fn | belongs to L and is thus finite m-a.e. In particular
the sequence of partial sums sN = N
P
n=1 fn is a sequence of measurable functions with
|sN | ≤ g that converges pointwise a.e. to a measurable function f . Hence by the DCT
and its corollary, f ∈ L1 and the partial sums (sN )N converge to f in L1 . □

1.2.4. Complex L1 (µ) spaces. In this subsection we describe the extension of L1 (µ) to a
complex vector space of complex valued functions (equivalence classes of functions).
Again we work on a fixed measure space (X, M , µ). As a topological space, C and
R , are the same. A function f : X → C = R2 is measurable if and only if it is M − B2
2

measurable. Measurability of f can also be described in terms of the real and imaginary
parts of f.
Proposition 1.19. Suppose (X, M ) is a measurable space and f : X → C. Writing
f : X → C as f = u + iv, where u, v : X → R, the function f is measurable if and only
if both u and v are.
Moreover, if f is measurable, then so is |f | : X → [0, ∞).

We begin with the following elementary lemma whose proof is left to the reader.
Lemma 1.20. Suppose (X, M ) is a measure space and Y and Z are topological spaces.
If f : X → Y is M − BY measurable and g : Y → Z is BY − BZ measurable, then g ◦ f
is M − BZ measurable. In particulr, the result holds if g is continuous.

Sketch of proof of Proposition 1.19. The Borel σ-algebra B2 is generated by open rect-
angles; that is, a set U ⊆ C is open if and only if it is a countable union of open rectangles
(with rational vertices even). For an open rectagle I = J × K = (a, b) × (c, d) observe
that
f −1 (J) = u−1 (J) ∩ v −1 (K).
Thus, if u and v are measurable, then f −1 (J) ∈ M . Consequently, f is measurable.
Hence if u, v are both measurable, then so is f.
8

Now suppose f is measurable. In this case

M ∋ f −1 ((t, ∞) × R) = {u > t}.
Since the sets {(t, ∞) : t} generate B1 , u is measurable. By symmetry v is measurable.
To prove the second statement, since f is measurable and g : C → [0, ∞) defined
by g(z) = |z| is continuous, the function g ◦ f = |f | is measurable by Lemma 1.20. □
Definition 1.21. A measurable f : X → C is integrable (or absolutely integrable) if |f |
is integrable.
Remark 1.22. From the inequalities
|Ref |, |Imf | ≤ |f | ≤ |Ref | + |Imf |
it follows that f : X → C is (absolutely) integrable if and only if Ref and Imf are.
Definition 1.23. If f is complex-valued and absolutely integrable (that is, f is mea-
surable and |f | is integrable), we define the integal of f by
Z Z Z
f = Ref + i Imf.
R
We also write ∥f ∥1 := X |f | dµ in the complex case. Finally, we write L1 = L1 (µ) to
denote the set of absolutely integrable complex-valued functions on X.

Generally, we leave it to context to indicate if we are considering the real or complex

version of L1 ; but for the following theorem we temporarily adopt the notation L1R and
L1C to distinguish between the real and complex vector space versions of L1 (µ).
Theorem 1.24 (L1 as a C normed vector space). The set L1C of is a vector space over
C (with the usual addition and scalar multiplication of functions). Morever, if f, g ∈ L1C
and c ∈ C, then
R
(a) the
R mapping
R Λ : L1 → C defined by Λ(f ) = f is linear;
(b) f ≤ |f |.
(c) ∥cf ∥1 = |c|∥f ∥1 .
(d) ∥f + g∥1 ≤ ∥f ∥1 + ∥g∥1 .

Sketch of proof. Write f = u + iv and g = x + iy. In particular, u, v, x, y are all L1R .

Given c = a + ib, the functions au, bv, av, bu are all L1R and so are au − bv and av + bu
since L1R is a real vector space. Therefore, cf = (au − bv) + i(av + bu) is in L1C . A similar,
but easier, argument shows f + g is in L1C . Hence L1C is a vector space over C. Moreover,
since the integral is real linear on L1R ,
Λ(cf ) = Λ((au + bv) + i(av + bu)) = Λ((au + bv)) + iΛ((av + bu))
= aΛ(u) + bΛ(v) + i[aΛ(v) + bΛ(u)]
= (a + ib)[Λ(u) + iΛ(v)] = cΛ(f ).
 9

Likewise Λ(f + g) = Λ(f ) + Λ(g). Thus Λ is C-linear on L1C and item (a) is proved.
R
If R f = 0, then f = 0 almost everywhere and the last three items hold. Otherwise,
write f = reit in polar coordinates and observe
Z
−it
e f ∈ R+ .

Thus, from the definition and linearity of the integral

Z Z Z Z
−it −it −it
R+ ∋ e f = e f = real e f + i image e−it f.

image e−it f = 0 and using results for L1R ,

R
Thus
Z Z Z Z Z Z
−it −it −it −it
f = e f = e f = real e f ≤ | real e f | ≤ |f |,

proving item (b).

Next, Z Z Z
|cf | = |c| |f | = |c| |f |.

Hence item (c) holds. Similarly,the triangle inequality, item (d), follows from |f + g| ≤
|f | + |g| (pointwise). □
Remark 1.25. Proposition 1.24 says ∥ · ∥1 is a semi-norm on L1 . As usual, we identify
functions that differ by a null vector; that is, f ∼ g if ∥f − g∥1 = 0; equivalently,
identifying functions that are equal a.e., we obtain a normed complex vector space of L1
functions (which of course are not actually functions).

1.2.5. Sequence spaces. Define

c0 := {f : N → F| lim |f (m)| = 0}
m→∞

ℓ∞ := {f : N → F| sup |f (m)| < ∞}

m∈N
X∞
ℓ1 := {f : N → F| |f (m)| < ∞}.
m=0

Note that ℓ∞ = Fb (N, F) and is a Banach space with the norm

∥f ∥∞ = sup |f (m)|.
m
∞ ∞
Further, c0 ⊆ ℓ is the subspace C0 (N) of ℓ again with the norm ∥ · ∥∞ . In particular,
c0 is a Banach space.
Observe that ℓ1 is the space (N, P (N), c), where c is counting measure on N and
∥ · ∥1 is the corresponding ℓ1 norm. Since only set of measure zero in this measure space
is the emptyset, two functions in ℓ1 = L1 (c) are equivalent if and only if they are equal.
10

Along with these spaces it is also helpful to consider the vector space
c00 := {f : N → F|f (n) = 0 for all but finitely many n}
Notice that c00 is a vector subspace of each of c0 , ℓ1 and ℓ∞ . Thus it can be equipped with
either the ∥ · ∥∞ or ∥ · ∥1 norms. It is not complete in either of these norms, however.
What is true is that c00 is dense in c0 and ℓ1 (but not in ℓ∞ ). (See Problem 1.11).

1.2.6. Lp spaces. Again let (X, M , m) be a measure space. For 1 ≤ p < ∞ let Lp (m)
denote the set of measurable functions f for which
Z 1/p
p
∥f ∥p := |f | dm <∞
X

(again we identify f and g when f = g a.e.). It turns out that this quantity is a norm on
Lp (m), and Lp (m) is complete, though we will not prove this yet (it is not immediately
obvious that the triangle inequality holds when p > 1).
Choosing (X, M , µ) = (N, P (N), c), counting measure on N, obtains the sequence
spaces ℓp ; that is, the F-vector space of functions f : N → F such that
∞
!1/p
X
∥f ∥p := |f (n)|p <∞
n=1
p
and this quantity is a norm making ℓ into a Banach space.
When p = ∞, we define L∞ (µ) to be the set of all functions f : X → F with the
following property: there exists M > 0 such that
(1) |f (x)| ≤ M for µ − a.e. x ∈ X;
as for the other Lp spaces we identify f and g when there are equal a.e. When f ∈ L∞ ,
let ∥f ∥∞ be the smallest M for which (1) holds. Then ∥ · ∥∞ is a norm making L∞ (µ)
into a Banach space.

1.2.7. Subspaces and products. If (X , ∥ · ∥) is a normed vector space and Y ⊆ X is a

vector subspace, then the restriction of ∥ · ∥ to Y is clearly a norm on Y. If X is a
Banach space, then (Y, ∥ · ∥) is a Banach space if and only if Y is closed in the norm
topology of X . (This is just a standard fact about metric spaces—a subspace of a
complete metric space is complete in the restricted metric if and only if it is closed.)
Definition 1.26. A subspace Y of a normed vector space X is a closed vector subspace
of X , denoted Y ≤ X . The terminology linear manifold is used synonymously with
vector subspace.

If X , Y are vector spaces then the algebraic direct sum is the vector space of ordered
pairs
X ⊕ Y := {(x, y) : x ∈ X , y ∈ Y}
 11

with entrywise operations. If X , Y are equipped with norms ∥ · ∥X , ∥ · ∥Y , then each of

the quantities
|∥(x, y)∥∞ := max(∥x∥X , ∥y∥Y ),
(2) ∥(x, y)∥1 := ∥x∥X + ∥y∥Y
q
∥(x, y)∥2 := ∥x∥2X + ∥y∥2Y

is a norm on X ⊕ Y. These three norms are equivalent; indeed it follows from the
definitions that
∥(x, y)∥∞ ≤ ∥(x, y)∥2 ≤ ∥(x, y)∥1 ≤ 2∥(x, y)∥∞ .
If X and Y are both complete, then X ⊕ Y is complete in each of these norms. The
resulting Banach spaces are denoted X ⊕∞ Y, X ⊕1 Y and X ⊕2 Y.
Since the three norms in the previous paragraph are equivalent, the resulting spaces
are indistinguishable topologically. There is a more abstract description of this topology.

Definition 1.27. Given topological spaces (X, τ ) and (Y, σ), the product topology on the
Cartesian product X × Y is the smallest topoology that makes the coordinate projections
πX : X × Y → X and πY : X × Y → Y defined by πX (x, y) = x, πY (x, y) = y continuous.
That is, the topology generated by the sets U × Y and X × V for open sets U ⊆ X and
V ⊆ Y.

Proposition 1.28. Suppose (X, τ ) and (Y, σ) are topological spaces. The collection of
sets
B = {U × V : U ⊆ X, V ⊆ Y are open}
is a base for the product topology.
For normed vector spaces X and Y, the product topology on X × Y is metrizable and
is the norm topology on X × Y with any of the norms of equation (2). Consequentely,
a sequence zn = (xn , yn ) from X × Y converges (in the product topology) if and only if
both (xn ) and (yn ) converge; and zn converges to z = (x, y) if and only if (xn ) converges
to x and (yn ) converges to y. In particular, if X and Y are Banach spaces, then so is
X × Y.

It is evident how to extend the discussion here to finite products. The product topol-
ogy is the default topology on (finite) products of Banach spaces (and more generally
normed vector spaces).

1.2.8. Qoutient spaces. If X is a normed vector space and M is a proper subspace, then
one can form the algebraic quotient X /M, defined as the collection of distinct cosets
{x + M : x ∈ X }. From linear algebra, X /M is a vector space under the standard
operations. Let π : X → X /M denote the quotient map.
12

Proposition 1.29. If M is a closed subspace of a normed vector space X , then the

quantity
∥π(x)∥ = ∥x + M∥ := inf ∥x − y∥
y∈M

is a norm on X /M. Moreover, if X is a Banach space, then so is X /M.

The norm in Proposition 1.29 is called the quotient norm. Geometrically, ∥x + M∥

is the distance in X from x to the closed set M. The assumption that M is closed in
needed to ensure that the quotient norm is indeed a norm. For instance M = C([0, 1])
is dense subspace of L1 ([0, 1]) (with Lebesgue measure) and hence for any

inf ∥f − g∥ = 0
g∈M

for all f ∈ L1 ([0, 1]).

Proof. We will verify a couple of the axioms of a norm for the quotient norm, leaving the
remainder of the proof as an exercise. First suppose x ∈ X and ∥π(x)∥ = 0. It follows
that there is a sequence (mn ) form M such that (∥x − mn ∥) converges to 0; that is, (mn )
converges to x. Since M is closed, x ∈ M and hence π(x) = 0.
Now let x, y ∈ X and ϵ > 0 be given. There exists m, n ∈ M such that

∥x − m∥ ≤ ∥π(x)∥ + ϵ, ∥y − n∥ ≤ ∥π(y)∥ + ϵ.

Hence

∥π(x)+π(y)∥ = ∥π(x+y)∥ ≤ ∥x+y−(m+n)∥ ≤ ∥x−m∥+∥y−n∥ ≤ ∥π(x)∥+∥π(y)|∥+2ϵ,

from which it follows that the triangle inequality holds and we have proved the quotient
norm is indeed a norm.
To prove X /M is complete (with the quotient norm) under the assumption that
P
X is a Banach space (complete), suppose (yn ) is a sequence from X /M and yn is
1
absolutely convergent. For each n there exists xn ∈ X such that ∥xn ∥ ≤ ∥yn ∥ + n2 and
P
π(xn ) = yn . It follows that xn is absolutely convergent. Since X is a Banach space
the sequence of partial sums sN = N
P
n=1 xn converges to some x ∈ X . In partiulcar,

N
X
∥sN − x∥ ≥ ∥π(sN − x)∥ = ∥π(sN ) − π(x)∥ = ∥ yn − π(x)∥.
n=1
P
Since (∥sN − x∥) converges to 0, it follows that yn converges to π(x). Hence X /M is
complete by Proposition 1.7. □

More examples are given in the exercises and further examples will appear after the
development of some theory.
 13

1.3. Linear transformations between normed spaces.

Definition 1.30. Let X , Y be normed vector spaces. A linear transformation T : X → Y
is bounded if there exists a constant C ≥ 0 such that ∥T x∥Y ≤ C∥x∥X for all x ∈ X .

Remark 1.31. Note that in Definition 1.30 it suffices to require that ∥T x∥Y ≤ C∥x∥X
just for all x ̸= 0, or for all x with ∥x∥X = 1 (why?). □

The importance of boundedness and the following simple proposition is hard to

overstate. Recall, a mapping f : X → Y between metric spaces is Lipschitz continuous
if there is a constant C > 0 such that d(f (x), f (y)) ≤ Cd(x, y) for all x, y ∈ X. A simple
exercise shows Lipschitz continuity implies (uniform) continuity.
Proposition 1.32. If T : X → Y is a linear transformation between normed spaces,
then the following are equivalent:
(i) T is bounded.
(ii) T is Lipschitz continuous.
(iii) T is uniformly continuous.
(iv) T is continuous.
(v) T is continuous at 0.
Moreover, in this case,
∥T ∥ := sup{∥T x∥ : ∥x∥ = 1}
∥T x∥
= sup{ : x ̸= 0}
∥x∥
= inf{C : ∥T x∥ ≤ C∥x∥ for all x ∈ X }
and ∥T ∥ is the smallest number (the infimum is attained 1) such that
(3) ∥T x∥ ≤ ∥T ∥ ∥x∥
for all x ∈ X .

Proof. Suppose T is bounded; that is, there exists a C ≥ 0 such that ∥T x∥ ≤ C∥x∥ for
all x ∈ X . Thus, if x, y ∈ X , then, ∥T x − T y∥ = ∥T (x − y)∥ ≤ C∥x − y∥ by linearity of
T . Hence (i) implies (ii). The implications (ii) implies (iii) implies (iv) implies (v) are
evident.
The proof of (v) implies (i) exploits the homogeneity of the norm and the linearity
of T and not nearly the full strength of the continuity assumption. By hypothesis, with
ϵ = 1 there exists δ > 0 such that if ∥x∥ = ∥x − 0∥ < 2δ, then ∥T x∥ = ∥T x − T 0)∥ < 1.
Given a nonzero vector x ∈ X , the vector δx/∥x∥ has norm less than δ, so
 
δx ∥T x∥
1> T =δ .
∥x∥ ∥x∥
1The suprema need not be attained.
14

Rearranging this we find ∥T x∥ ≤ (1/δ)∥x∥ for all x ̸= 0.

Assuming T is bounded, it is immediate that sup{∥T x∥ : ∥x∥ = 1} exists (and is a
real number). From homogeneity of the norm, it is also clear that
∥T x∥
sup{∥T x∥ : ∥x∥ = 1} = sup{ : x ̸= 0}.
∥x∥
Likewise assuming T is bounded the set S = {C : ∥T x∥ ≤ C∥x∥ for all x ∈ X } ⊆ [0, ∞)
is not empty and bounded below (by 0) and hence the infimum exists. From the definition
of ∥T ∥ we see that ∥T ∥ ∈ S. Hence the infimum is at most ∥T ∥. On the other hand, if
C ′ < inf S, then there is an x ∈ X such that ∥T x∥ > C ′ ∥x∥ so that ∥T x∥
∥x∥
> C ′ . Thus
C ′ < ∥T ∥. □

The set of all bounded linear operators from X to Y is denoted B(X , Y). It is a
vector space under the operations of pointwise addition and scalar multiplication. The
quantity ∥T ∥ is easily seen to be a norm. It is called the operator norm of T .
Problem 1.1. Prove the ∥ · ∥1 and ∥ · ∥∞ norms on c00 are not equivalent. Conclude
from your proof that the identity map on c00 is bounded from the ∥ · ∥1 norm to the
∥ · ∥∞ norm, but not the other way around.
Problem 1.2. Consider c0 and c00 equipped with the ∥ · ∥∞ norm. Prove there is no
bounded operator T : c0 → c00 such that T |c00 is the identity map. (Thus the conclusion
of Proposition 1.34 can fail if Y is not complete.)
Proposition 1.33. For normed vector spaces X and Y, the operator norm makes
B(X , Y) into a normed vector space that is complete if Y is complete.

Proof. That B(X , Y) is a normed vector space follows readily from the definitions and
is left as an exercise.
Suppose now Y is complete, and let Tn be a cauchy sequence in B(X , Y). Let
E = X1 denote the closed unit ball in X . For x ∈ E,
(4) ∥Tn x − Tm x∥ = ∥(Tn − Tm )x∥ ≤ ∥Tn − Tm ∥∥x∥ ≤ ∥Tn − Tm ∥.
Hence the sequence (Tn |E ) is a cauchy sequence (so uniformly cauchy) from Cb (E, Y),
the space of bounded continuous functions from B to Y. Since Y is complete, there is
an F ∈ Cb (E, Y) such that (Tn |E ) converges to F in Cb (E, Y) and moreover ∥F (x)∥ ≤
C := sup{∥Tn ∥ : n} < ∞. See Subsection 1.2.2. An exercise shows, given x, y ∈ B and
c ∈ F if x + y ∈ B and c x ∈ B, then F (x + y) = F (x) + F (y) and F (cx) = cF (x).
Hence F extends, by homogeneity, to a linear map T : X → Y such that ∥T ∥ ≤ C and,
by equation (4), (Tn ) converges to T in B(X, Y ). □

If T ∈ B(X , Y) and S ∈ B(Y, Z), then two applications of the inequality (3) give,
for x ∈ X ,
∥ST x∥ ≤ ∥S∥∥T x∥ ≤ ∥S∥∥T ∥∥x∥
 15

and it follows that ST ∈ B(X , Z) and ∥ST ∥ ≤ ∥S∥∥T ∥. In the special case that Y = X
is complete, B(X ) := B(X , X ) is an example of a Banach algebra.
The following proposition is very useful in constructing bounded operators—at least
when the codomain is complete. Namely, it suffices to define the operator (and show
that it is bounded) on a dense subspace.
Proposition 1.34 (Extending bounded operators). Let X , Y be normed vector spaces
with Y complete, and E ⊆ X a dense linear subspace. If T : E → Y is a bounded linear
operator, then there exists a unique bounded linear operator Te : X → Y extending T (so
Te|E = T ). Further ∥Te∥ = ∥T ∥.

Sketch of proof. Recall, if X, Y are metric spaces, Y is complete, D ⊆ X is dense and

f : D → Y is uniformly continuous, then f has a unique continuous extension f˜ :
X → Y . Moreover, this extension can be defined as follows. Given x ∈ X, choose a
sequence (xn ) from D converging to x and let f˜(x) = lim f (xn ) (that the sequence f (xn )
is Cauchy follows from uniform continuity; that it converges from the assumption that
Y is complete and finally it is an exercise to show f˜(x) is well defined independent of
the choice of (xn )). Thus, it only remains to verify that the extension Te of T is in fact
linear and ∥T ∥ = ∥Te∥. Both are routine exercises. □
Example 1.35. Equip c0 and c00 with the sup norm, ∥ · ∥∞ and consider the identity
map ι : c00 → c00 . If T is an extension of ι to the completion c0 of c00 (in the sup norm),
then, letting sn ∈ c00 denote the sequence sn (m) = m1 for m ≤ n and sn (m) = 0 for
m > n, the sequence (sn ) is converge in c0 to the sequence s with s(m) = m1 for all m.
Hence (T (sn ) = sn ) converges to some t ∈ c00 . But now there is a K such that t(k) = 0
for all k ≥ K so that ∥sn − t∥ ≥ K1 for all n ≥ K, a contradiction. This example shows
completeness of Y is essential in Proposition 1.34. □

Definition 1.36. A bounded linear transformation T ∈ B(X , Y) is said to be invertible

if it is bijective (being bijective, automatically T −1 exists and is a linear transformation)
and T −1 is bounded from Y to X . Two normed spaces X , Y are said to be (boundedly)
isomorphic if there exists an invertible linear transformation T : X → Y.
Example 1.37. As an example, given equivalent norms ∥ · ∥1 and ∥ · ∥2 on a vector
space X , the identity mapping ι : (X , ∥ · ∥1 ) → (X , ∥ · ∥2 ) is (boundedly) invertible and
witnesses the fact that these two normed vector spaces are boundedly isomorphic.
Definition 1.38. An operator T : X → Y such that ∥T x∥ = ∥x∥ for all x ∈ X is an
isometry. Note that an isometry is automatically injective and if it is also surjective then
it is automatically invertible and T −1 is also an isometry. The normed vector spaces are
isometrically isomorphic if there is an invertible isometry T : X → Y.
Example 1.39. If X is a finite dimensional vector space and T : X → X is an isometry,
then T is onto. However, when X is not finite dimensional, an isometry need not be
16

surjective. As examples, let ℓp = ℓp (N) denote the sequence spaces from Subsection 1.2.5.
The linear map S : ℓp (N) → ℓp (N) defined by Sf (n) = 0 if n = 0 and f (n − 1) if n > 0
(for f = (f (n))n ∈ ℓp ) is the shift operator . It is straightforward to verify that S is an
isometry but not onto.
Example 1.40. Following up on the previous example, a linear map T : X → X can
be one-one and have dense range without being (boundedly) invertible. Let en ∈ ℓ2 (N)
denote the function en (m) = 1 if n = m and 0 otherwise for non-negative integers
0 ≤ m, n. The set of c00 = { N 2
P
n=0 an en : N ∈ N, cn ∈ F} is dense in ℓ (N) and the
mapping D : c00 → ℓ2 (N) defined by
N N
X X an
D( an e n ) = en
n=0 n=0
n+1
is easily seen to be bounded with ∥D∥ = 1. It is also injective. Hence D extends to an
injective bounded operator, still denoted D, from ℓ2 → ℓ2 , with ∥D∥ = 1. The range of
D contains {en : n ∈ N} and is thus dense in ℓ2 (N).
Since ∞
P 1 2
P∞ 1 2
n=0 | n+1 | < ∞, the vector f = n=1 n+1 en is in ℓ (N). On the other hand,
if g ∈ ℓ2 (N) and Dg = f, then
1 1
g(n) = (Dg)(n) = f (n) =
n+1 n+1
2
P
and thus g(n) = 1 for all n; however, since |g(n)| = ∞, we obtain a contradiction.
Hence f is not in the range of D.

1.4. Examples.
(a) If X is a finite-dimensional normed space and Y is any normed space, then every
linear transformation T : X → Y is bounded. See Problem 1.16.

(b) Let X denote c00 equipped with the ∥ · ∥1 norm, and Y denote c00 equipped with the
∥ · ∥∞ norm. Then the identity map idX ,Y : X → Y is bounded (in fact its norm is
equal to 1), but its inverse, the identity map ιY,X : Y → X , is unbounded. To verify
this claim, For positive integers n, let fn denote the element of c00 defined by
(
1 if m ≤ n
fn (m) =
0 if m > n.
Now ∥ιY,X (fn )∥1 = n, but ∥fn ∥∞ = 1.

(c) Consider c00 with the ∥ · ∥∞ norm. Let a : N → F be any function and define a linear
transformation Ta : c00 → c00 by
(5) Ta f (n) = a(n)f (n).
The mapping Ta is bounded if and only if M = supn∈N |a(n)| < ∞, in which case
∥Ta ∥ = M . In this case, Ta extends uniquely to a bounded operator from c0 to c0
 17

by Proposition 1.34, and one may check that the formula (5) defines the extension.
All of these claims remain true if we use the ∥ · ∥1 norm instead of the ∥ · ∥∞ norm.
In this case, we get a bounded operator from ℓ1 to itself.

(d) Define S : ℓ1 → ℓ1 as follows: given the sequence (f (n))n from ℓ1 let Sf (1) = 0 and
Sf (n) = f (n − 1) for n > 1. (Viewing f as a sequence, S shifts the sequence one
place to the right and fills in a 0 in the first position). This S is an isometry, but is
not surjective. In contrast, if X is finite-dimensional, then the rank-nullity theorem
from linear algebra guarantees that every injective linear map T : X → X is also
surjective.

(e) Let C ∞ ([0, 1]) denote the vector space of functions on [0, 1] with continuous deriva-
tives of all orders. The differentiation map D : C ∞ ([0, 1]) → C ∞ ([0, 1]) defined by
df
Df = dx is a linear transformation. Since, for t ∈ R, we have Detx = tetx , it follows
that there is no norm on C ∞ ([0, 1]) such that dxd
is bounded.

1.5. Problems.

Problem 1.3. Prove Proposition 1.5.

Problem 1.4. Prove equivalent norms define the same topology and the same Cauchy
sequences.

Problem 1.5. (a) Prove all norms on a finite dimensional vector space X are equivalent.
P P
Suggestion: Fix a basis e1 , . . . en for X and define ∥ ak ek ∥1 := |ak |. It is routine
to check that ∥ · ∥1 is a norm on X . Now complete the following outline.
(i) Let ∥ · ∥ be the given norm on X . Show there is an M such that ∥x∥ ≤ M ∥x∥1 .
Conclude that the mapping ι : (X , ∥ · ∥1 ) → (X , ∥ · ∥) defined by ι(x) = x is
continuous;
(ii) Show that the unit sphere S = {x ∈ X : ∥x∥1 = 1} in (X , ∥ · ∥1 ) is compact in
the ∥ · ∥1 topology;
(iii) Show that the mapping f : S → (X , ∥ · ∥) given by f (x) = ∥x∥ is continuous
and hence attains its infimum. Show this infimum is not 0 and finish the proof.
(b) Combine the result of part (a) with the result of Problem 1.4 to conclude that every
finite-dimensional normed vector space is complete.
(c) Let X be a normed vector space and M ⊆ X a finite-dimensional subspace. Prove
M is closed in X .

Problem 1.6. Finish the proofs from the examples subsections.

Problem 1.7. A function f : [0, 1] → F is called Lipschitz continuous if there exists a

constant C such that
|f (x) − f (y)| ≤ C|x − y|
18

for all x, y ∈ [0, 1]. Define ∥f ∥Lip to be the best possible constant in this inequality.
That is,
|f (x) − f (y)|
∥f ∥Lip := sup
x̸=y |x − y|
Let Lip[0, 1] denote the set of all Lipschitz continuous functions on [0, 1]. Prove ∥f ∥ :=
|f (0)| + ∥f ∥Lip is a norm on Lip[0, 1], and that Lip[0, 1] is complete in this norm.
Problem 1.8. Let C 1 ([0, 1]) denote the space of all functions f : [0, 1] → R such that
f is differentiable in (0, 1) and f ′ extends continuously to [0, 1]. Prove
∥f ∥ := ∥f ∥∞ + ∥f ′ ∥∞
is a norm on C 1 ([0, 1]) and that C 1 is complete in this norm. Do the same for the norm
∥f ∥ := |f (0)| + ∥f ′ ∥∞ . (Is ∥f ′ ∥∞ a norm on C 1 ?)
Problem 1.9. Let (X, M ) be a measurable space. Let M (X) denote the (real) vector
space of all signed measures on (X, M ). Prove the total variation norm ∥µ∥ := |µ|(X)
is a norm on M (X), and M (X) is complete in this norm.
Problem 1.10. Prove, if X , Y are normed spaces, then the operator norm is a norm on
B(X , Y).
Problem 1.11. Prove c00 is dense in c0 and ℓ1 . (That is, given f ∈ c0 there is a sequence
fn in c00 such that ∥fn − f ∥∞ → 0, and the analogous statement for ℓ1 .) Using these
facts, or otherwise, prove that c00 is not dense in ℓ∞ . (In fact there exists f ∈ ℓ∞ with
∥f ∥∞ = 1 such that ∥f − g∥∞ ≥ 1 for all g ∈ c00 .)
Problem 1.12. Prove c00 is not complete in the ∥ · ∥1 or ∥ · ∥∞ norms. (After we have
studied the Baire Category theorem, you will be asked to prove that there is no norm
on c00 making it complete.)
Problem 1.13. Consider c0 and c00 equipped with the ∥ · ∥∞ norm. Prove there is no
bounded operator T : c0 → c00 such that T |c00 is the identity map. (Thus the conclusion
of Proposition 1.34 can fail if Y is not complete.)
Problem 1.14. Prove the ∥ · ∥1 and ∥ · ∥∞ norms on c00 are not equivalent. Conclude
from your proof that the identity map on c00 is bounded from the ∥ · ∥1 norm to the
∥ · ∥∞ norm, but not the other way around.
Problem 1.15. a) Prove f ∈ C0 (Rn ) if and only if f is continuous and lim|x|→∞ |f (x)| =
0. b) Let Cc (Rn ) denote the set of continuous, compactly supported functions on Rn .
Prove Cc (Rn ) is dense in C0 (Rn ) (where C0 (Rn ) is equipped with sup norm).
Problem 1.16. Prove, if X , Y are normed spaces and X is finite dimensional, then
every linear transformation T : X → Y is bounded. Suggestion: Let d denote the
dimension of X and let {e1 , . . . , ed } denote a basis. The function ∥ · ∥1 on X defined by
P P
∥ xj ej ∥1 = |xj | is a norm. Apply Problem 1.5.
 19

Problem 1.17. Prove the claims in Example 1.4(c).

Problem 1.18. Let g : R → F be a (Lebesgue) measurable function. The map M g :

f → gf is a linear transformation on the space of measurable functions. Prove, if
/ L∞ (R), then there is an f ∈ L1 (R) such that gf ∈
g ∈ / L1 (R). Conversely, show if
g ∈ L∞ (R), then Mg is bounded from L1 (R) to itself and ∥Mg ∥ = ∥g∥∞ .

Problem 1.19. Prove the claims about direct sums.

Problem 1.20. Let X be a normed vector space and M a proper closed subspace.
Prove for every ϵ > 0, there exists x ∈ X such that ∥x∥ = 1 and inf y∈M ∥x − y∥ > 1 − ϵ.
(Hint: take any u ∈ X \ M and let a = inf y∈M ∥u − y∥. Choose δ > 0 small enough so
a u−v
that a+δ > 1 − ϵ, and then choose v ∈ M so that ∥u − v∥ < a + δ. Finally let x = ∥u−v∥ .)
Note that the distance to a (closed) subspace need not be attained. Here is an
example. Consider the Banach space C([0, 1]) (with the sup norm of course and either
real or complex valued functions) and the closed subspace
Z 1
T = {f ∈ C([0, 1]) : f (0) = 0 = f dt}.
0

Using machinery in the next section it will be evident that T is a closed subspace of
C([0, 1]). For now, it can be easily verified directly. Let g denote the function g(t) = t.
Verify that, for f ∈ T , that
Z Z
1
= g dt = (g − f ) dt ≤ ∥g − f ∥∞ .
2
In particular, the distance from g to T is at least 21 .
Note that the function h = x − 12 , while not in T , satisfies ∥g − h∥∞ = 12 .
On the other hand, for any ϵ > 0 there is an f ∈ T so that ∥g − f ∥∞ ≤ 12 + ϵ
(simply modify h appropriately). Thus, the distance from g to T is 12 . Now
R verify, using
the inequality above, that h is the only element of C([0, 1]) such that h dt = 0 and
∥g − h∥∞ = 12 .

Problem 1.21. Prove, if X is an infinite-dimensional normed space, then the unit ball
ball(X ) := {x ∈ X : ∥x∥ ≤ 1} is not compact in the norm topology. (Hint: use the
result of Problem 1.20 to construct inductively a sequence of vectors xn ∈ X such that
∥xn ∥ = 1 for all n and ∥xn − xm ∥ ≥ 12 for all m < n.)

Problem 1.22. (The quotient norm) Let X be a normed space and M a proper closed
subspace.
a) Prove the quotient norm is a norm.
b) Show that the quotient map x → x + M has norm 1. (Use Problem 1.20.)
c) Prove, if X is complete, so is X /M.
20

Problem 1.23. A normed vector space X is called separable if it is separable as a metric

space (that is, there is a countable subset of X which is dense in the norm topology).
Prove c0 and ℓ1 are separable, but ℓ∞ is not. (Hint: for ℓ∞ , show that there is an
uncountable collection of elements {fα } such that ∥fα − fβ ∥ = 1 for α ̸= β.)

END FALL TERM

2. Linear functionals and the Hahn-Banach theorem

If there is a fundamental theorem of functional analysis, it is the Hahn-Banach

theorem. The theorem is somewhat abstract-looking at first, but its importance will be
clear after studying some of its corollaries.
Definition 2.1. Let X be a normed vector space over the field F. A linear functional
on X is a linear map L : X → F. The dual space of X , denoted X ∗ is the space B(X , F)
of bounded linear functionals on X .
Remark 2.2. Since F = R or C is complete, the vector space of bounded linear func-
tionals is itself a Banach space (complete normed vector space) and is known as the .
It is not yet obvious that X ∗ need be non-trivial (that is, that there are any bounded
linear functionals on X besides 0). One corollary of the Hahn-Banach theorem is there
exist enough bounded linear functionals on X to separate points.

2.1. Examples. This subsection contains some examples of bounded linear functionals
and dual spaces.
Example 2.3. For each of the sequence spaces c0 , ℓ1 , ℓ∞ , for each n the map f → f (n) is
a bounded linear functional. That is, λn : X → F defined by λn (f ) = f (n) for f : N → F
in X , where X is any one of c0 , ℓ1 , ℓ∞ , is continuous since in each case it is immediate
that
|λn (f )| = |f (n)| ≤ ∥f ∥X .
Example 2.4. Given g ∈ ℓ1 , if f ∈ c0 , then
X∞ X∞
(6) |f (n)g(n)| ≤ ∥f ∥∞ |g(n)| = ∥g∥1 ∥f ∥∞ .
n=0 n=0
P∞
Thus n=0 f (n)g(n) converges and we obtain a functional Lg : c0 → F defined by
∞
X
(7) Lg (f ) := f (n)g(n).
n=0

The inequality of equation (6) says Lg is bounded (continuous) and ∥Lg ∥ ≤ ∥g∥1 . More-
over, it is immediate that Φ : ℓ1 → c∗0 defined by Φ(g) = Lg is bounded and linear and
∥Φ∥ ≤ 1. In fact, Φ is onto so that every bounded linear functional on c0 is of the form
Lg for some g ∈ ℓ1 .
 21

Proposition 2.5. The map Φ : ℓ1 → c∗0 defined by Φ(g) = Lg is an isometric isomor-

phism from ℓ1 onto the dual space c∗0 .

Proof. We have already seen that each g ∈ ℓ1 gives rise to a bounded linear functional
Lg ∈ c∗0 via
∞
X
Lg (f ) := g(n)f (n),
n=0

that ∥Lg ∥ ≤ ∥g∥1 and the the mapping Φ is bounded and linear. We will prove simul-
taneously that this map is onto and that ∥Lg ∥ ≥ ∥g∥1 .
Let L ∈ c∗0 . We will first show that there is unique g ∈ ℓ1 so that L = Lg . Let
en ∈ c0 be the indicator function of n, that is
en (m) = δnm .
Define a function g : N → F by
g(n) = L(en ).
We claim that g ∈ ℓ1 and L = Lg . To see this, fix an integer N and define h = hN :
N → F by
(
g(n)/|g(n)| if n ≤ N and g(n) ̸= 0
h(n) =
0 otherwise.
Thus h = N
P
n=0 h(n)en . Further, by h ∈ c00 ⊆ c0 and ∥h∥∞ ≤ 1. Now
N
X N
X
|g(n)| = h(n)g(n) = L(h) = |L(h)| ≤ ∥L∥∥h∥ ≤ ∥L∥.
n=0 n=0

It follows that g ∈ ℓ1 and ∥g∥1 ≤ ∥L∥. By construction L = Lg when restricted to c00 , so

by the uniqueness of extensions of bounded operators, Proposition 1.34, L = Lg . Thus
the map g → Lg is onto and
∥g∥1 ≤ ∥L∥ = ∥Lg ∥ ≤ ∥g∥1 . □

Example 2.6. Given g ∈ ℓ∞ , if f ∈ ℓ1 , then equation (6) shows |Lg (f )| ≤ ∥g∥∞ ∥f ∥1 ,

where Lg is defined as in equation (7). Thus ∥Lg ∥ ≤ ∥g∥∞ and we obtain a bounded
linear map Ψ : ℓ∞ → (ℓ1 )∗

Proposition 2.7. The map Ψ is an isometric isomorphism from (ℓ1 )∗ onto ℓ∞ .

Proof. The proof follows the same lines as the proof of the previous proposition; the
details are left as an exercise. □

Remark 2.8. The same mapping g → Lg also shows that every g ∈ ℓ1 gives a bounded
linear functional on ℓ∞ , but it turns out these do not exhaust (ℓ∞ )∗ (see Problem 2.12).
22

Regarding ℓ1 and ℓ∞ as L1 and L∞ for counting measure on N, it is not surprising

that, given a measure space (X, M , µ), a function g ∈ L∞ (µ) (see Subsection 1.2.6 for
the definition of L∞ (µ)) defines a linear functional Lg : L1 (µ) → F by
Z
Lg (f ) := f g dm
X
1
for f ∈ L (µ) is a bounded linear functional of norm at most M . We will prove in
Section 4 that the norm of Lg is in fact ∥g∥∞ , and every bounded linear functional on
L1 (m) is of this type (at least when m is σ-finite). □
Example 2.9. A regular Borel measure µ on a locally compact set X such that µ(K) <
∞ for compact subsets of X determines a linear functional λ : Cc (X) → F by
Z
λ(f ) = λµ (f ) = f dµ.
X

An f ∈ Cc (X) is a positive function (really non-negative), written f ≥ 0, if f (x) ≥ 0

for all x ∈ X. The linear functional λµ is a positive linear functional in the sense that if
f ∈ Cc (X) is positive, then λµ (f ) ≥ 0.
As a second example, let X = [0, 1] and note that the mapping I : C([0, 1]) → C
defined by
Z 1
I(f ) = f dx,
0
where the integral is in the Riemann sense, is a positive linear functional on C([0, 1]).

END Monday 2025-01-13

Theorem 2.10 (Riesz-Markov Representation Theorem: positive version). Let X =
(X, τ ) be a locally compact Hausdorff space. If λ : Cc (X) → C is a positive linear
functional, then there exists a unique Borel measure µ on the Borel σ-algebra BX , such
that Z
λ(f ) = f dµ

for f ∈ Cc (X). Moreover, µ is regular in the sense that

(i) if K ⊆ X is compact, then µ(K) < ∞;
(ii) if E ∈ BX , then µ(E) = inf{µ(U ) : E ⊆ U, U open}; and
(iii) if E ∈ BX and µ(E) < ∞, then µ(E) = sup{µ(K) : K ⊆ E, K compact}.
Remark 2.11. In general elements of Cc (X)∗ correspond to signed measures that will
appear later in these notes. □

2.2. Continuous linear functionals. For linear functionals, we can add to the list of
equivalent conditions of Proposition 1.32. In particular, the proof of the equivalence of
items (a) and (c) in Proposition 2.12 requires a map into the scalar field.
 23

Proposition 2.12. If X is a normed vector space and if λ : X → F is a non-zero (not

identically zero) linear functional , then the following are equivalent.
(a) λ is continuous;
(b) λ is bounded;
(c) ker λ is closed;
(d) ker λ ̸= X .

Proof. Items (a) and (b) are equivalent by Proposition 1.32 and it is evident that item (a)
implies item (c).
Suppose item (b) does not hold. Thus, there exists a sequence (fn ) from X such
that ∥fn ∥ ≤ 1, but |λ(fn )| ≥ n. Choose e ∈ X with λ(e) = 1 and let
fn
hk = e −
λ(fn )
and note that (hk ) converges to e but λ(hk ) = 0 for all k. Thus (hk ) is a sequence from
ker λ that converges to a point not in ker λ. Thus item (c) implies item (b).
Since ker λ ̸= V (since λ is not the zero map), item (c) implies item (d). Now
suppose item (c) does not hold. Thus there exists an f ̸∈ ker λ and a sequence (fn )
from ker λ that converges to f. Without loss of generality, λ(f ) = 1. Given g ∈ X , the
sequence
gn = (g − λ(g)f ) + λ(g) fn
converges to g and λ(gn ) = 0. Thus g ∈ ker λ and we conclude X = ker λ. □
Remark 2.13. Note that Proposition 2.12 remains true with λ−1 ({a}) in place of ker λ,
for any choice of a ∈ F. For instance, in the proof that item (c) implies item (b), simply
require λ(e) = a + 1 instead of λ(e) = a/ In the proof that item (d) implies item (c),
suppose the sequence (fn ) converges to f and λ(fn ) = a, but λ(fn ) = b ̸= a. In this
case, given g ∈ X , let gn = (g − cf ) + cfn < where c = a−λ(g)
a−b
. The details are left as an
exercise.
As a corollary, Proposition 2.12 extends to linear maps from a normed space X into
a finite dimensional normed space as an easy argument shows.
If X is infinite dimensional, the result is false. Just choose a basis B for X and
let B0 = {b1 , b2 , . . . } denote a countable subset of B. Define L : X → X by declaring
L(bn ) = n bn and L(b) = b for b ∈ B \ B0 and extend by linearity. Thus L is one-one so
that ker L = {0} is closed, but L is not bounded (and so not continuous). □

We close this subsection with the following result that should be compared with
item (a) from Subsection 1.4 followed by an example.
Proposition 2.14. If V is an infinite dimensional normed vector space, then there exists
a linear map f : V → F that is not continuous.
24

For a Banach space X , there are notions of a basis that reference the norm. For
instance, a Schauder basis for X is a sequence (en )∞ n=1 such that
P∞for each x ∈ X there
exists a unique choice of scalars xn ∈ F such that the series n=1 xn en converges to
x. Forgetting the norm structure, a Hamel basis B ⊆ X for X is a basis in the sense
of linear algebra. Explicitly, letting F00 (B) denote the functions a : B → F such that
ab = a(b) is zero for all but finitely many b ∈ B, the set B is a Hamel basis for X if for
each v ∈ X there exist is a unique function a ∈ F00 (B) such that
X finite
X
v= ab b = ab b.
b∈B b∈B
In this case any choice of c : B → F determines uniquely a linear functional λ : X → F
via the rule X
λ(v) = c b ab ,
b∈B
where cb = c(b). Often this process is described informally as: let λ(b) = c(b) and extend
by linearity. Finally, an argument using Zorn’s Lemma, which we will soon encounter
in the proof of the Hahn-Banach Theorem, shows that every vector space has a basis.
While it is true that every basis for a vector space V has the same cardinality, all that
we need to make sense of the statement V is an infinite dimensional vector space is the
fact that V has a basis that is infinite, then all bases for V are infinite, which is an
immediate consequence of the fact that all bases for a finite dimensional vector space
have the same cardinality. Thus, we can take the statement X is infinite dimensional
to mean that X has a Hamel basis B that contains a countable set B0 .

Proof of Proposition 2.14. Let B denote a Hamel basis for V. By assumption, B has a
countable subset B0 . Write B0 = {b1 , b2 , . . . } (so choose a bijection ψ : N → B0 ) and
assume, without loss of generality that ∥bj ∥ = 1. Let λ : V → F denote the linear
functional determined by λ(bj ) = j for bj ∈ B0 and λ(b) = 0 for b ∈ B \ B0 and observe
that λ is not bounded. □
Example 2.15. Let X denote an infinite dimensional normed vector space and suppose
f : X → F denote a discontinuous linear functional. An exercise shows that f −1 ({1}) =
X in addition to ker f = X (see Proposition 2.12). Let U = {f ≤ 0} ⊆ ker λ and note
V = X \ U = {f > 0} ⊇ f −1 ({1}). Now U and V are disjoint convex sets such that
X = U ∪ V and U = X = V .

2.3. The Hahn-Banach Extension Theorem. To state and prove the Hahn-Banach
Extension Theorem, we first work in the setting F = R, then extend the results to the
complex case.
Definition 2.16. Let X be a real vector space. A Minkowski functional is a function
p : X → R such that p(x + y) ≤ p(x) + p(y) and p(λx) = λp(x) for all x, y ∈ X and
nonnegative λ ∈ R.
 25

For examples, if L : X → R is any linear functional, then the function p : X → R

defined by p(x) := |L(x)| is a Minkowski functional; and if ∥ · ∥ is a seminorm on X ,
then p : X → R defined by p(x) = ∥x∥ is a Minkowski functional.
Theorem 2.17 (The Hahn-Banach Extension2 Theorem, real version). Let X denote
a real vector space, p a Minkowski functional on X , and M a subspace of X . If L is a
linear functional on M such that L(x) ≤ p(x) for all x ∈ M, then there exists a linear
functional L′ on X such that
(i) L′ |M = L (L′ extends L)
(ii) L′ (x) ≤ p(x) for all x ∈ X (L′ is dominated by p).
Remark 2.18. In the statement of Theorem 2.17, X is a vector space, not a normed vec-
tor space and correspondingly M is a subspace in the sense of linear algebra (sometimes
referred to as a linear manifold). □

The proof will invoke Zorn’s Lemma, a result that is equivalent to the axiom of
choice (as well as the well-ordering principle and the Hausdorff maximality principle).
A partial order ⪯ on a set S is a relation that is reflexive, symmetric and transitive;
that is, for all x, y, z ∈ S
(i) x ⪯ x,
(ii) if x ⪯ y and y ⪯ x, then x = y, and
(iii) if x ⪯ y and y ⪯ z, then x ⪯ z.
We call S, or more precisely (S, ⪯), a partially ordered set or poset. A subset T of
S is totally ordered , if for each x, y ∈ T either x ⪯ y or y ⪯ x. A totally ordered subset
T is often called a chain. An upper bound z for a chain T is an element z ∈ S such that
t ⪯ z for all t ∈ T . A maximal element for S is a w ∈ S that has no successor; that is
there does not exist an s ∈ S such that s ̸= w and w ⪯ s. An upper bound for a subset
A of S is an element s ∈ S such that a ⪯ s for all a ∈ A.
Theorem 2.19 (Zorn’s Lemma). Suppose S is a partially ordered set. If every chain in
S has an upper bound, then S has a maximal element.

END Wednesday 2025-01-15

The following Lemma is at the heart of the proof of Theorem 2.17.
Lemma 2.20. With the hypotheses of Theorem 2.17, if x ∈ X \ M, then the conclusion
of Theorem 2.17 holds with the subspace M + Rx in place of X .

Proof. For any m1 , m2 ∈ M, by hypothesis,

L(m1 ) + L(m2 ) = L(m1 + m2 ) ≤ p(m1 + m2 ) ≤ p(m1 − x) + p(m2 + x).
2Thereis also the Hahn-Banach Separation Theorem. Both theorems are often simple called the (sic)
Hahn-Banach Theorem.
26

Rearranging gives, for m1 , m2 ∈ M,

L(m1 ) − p(m1 − x) ≤ p(m2 + x) − L(m2 )
and thus
sup {L(m) − p(m − x)} ≤ inf {p(m + x) − L(m)}.
m∈M m∈M
Now choose any real number λ satisfying
sup {L(m) − p(m − x)} ≤ λ ≤ inf {p(m + x) − L(m)}.
m∈M m∈M

In particular, for m ∈ M,
L(m)−λ ≤ p(m − x)
(8)
L(m)+λ ≤ p(m + x).
Let N = M + Rx and define L′ : N → R by L′ (m + tx) = L(m) + tλ for m ∈ M
and t ∈ R. Thus L′ is linear and agrees with L on M by definition. Moreover, by
construction and equation 8,
L(m − x) ≤ p(m − x)
(9)
L(m + x) ≤ p(m + x).
We now check that L′ (y) ≤ p(y) for all y ∈ M + Rx. Accordingly, suppose y ∈ N
so that there exists m ∈ M and t ∈ R such that y = m + tx. If t = 0 there is nothing
to prove. If t > 0, then, in view of the second inequality of equation (9),
 m  m
′ ′
L (y) = L (m + tx) = t L( ) + λ ≤ t p( + x) = p(m + tx) = p(y)
t t
and a similar estimate, using the first inequality of equation (9), shows that
L′ (m + tx) ≤ p(m + tx)
for t < 0. We have thus successfully extended L to a linear map L′ : N → R satisfying
L′ (n) ≤ p(n) for all n ∈ N and the proof is complete. □

We make one further observation before turning to the proof of the Hahn-Banach
Theorem. If T is a totally ordered set and (Nα )α∈T are subspaces of a vector space X
that are nested increasing in the sense that Nα ⊆ Nβ for α ⪯ β, then N = ∪α∈T Nα is
again a subspace of X . By contrast, if X is a normed vector space and Nα are (closed)
subspaces of X , then N will not necessarily be a (closed) subspace of X .

Proof of Theorem 2.17. Let L denote the set of pairs (L′ , N ) where N is a subspace of
X containing M, and L′ is an extension of L to N obeying L′ (y) ≤ p(y) on N . Declare
(L′1 , N1 ) ⪯ (L′2 , N2 ) if N1 ⊆ N2 and L′2 |N1 = L′1 . This relation ⪯ is a partial order on L;
that is (L, ⪯) is a partially ordered set. Further, Lemma 2.20 says if (L′ , N ) is maximal
element, then N = X .
An exercise shows, given any increasing chain (L′α , Nα ) in L has as an upper bound
(L′ , N ) in L, where N := α Nα and L′ (nα ) := L′α (nα ) for nα ∈ Nα . By Zorn’s Lemma
S
 27

the collection L has a maximal element (L′ , N ) with respect to the order ⪯ and the
proof is complete. □

The use of Zorn’s Lemma in the proof of Theorem 2.17 is a typical - one knows how
to carry out a construction one step a time, but there is no clear way to do it all at once.
As an exercise, use Zorn’s Lemma to prove that if V Is a vectors space and S ⊆ V is a
linearly independent set, then there is a basis B for V such that B ⊇ S.
In the special case that p is a seminorm, since L(−x) = −L(x) and p(−x) = p(x)
the inequality L ≤ p is equivalent to |L| ≤ p.
Corollary 2.21. Suppose X is a real normed vector space, M is a subspace of X , and
L is a bounded linear functional on M. If C ≥ 0 and |L(x)| ≤ C∥x∥ for all x ∈ M,
then there exists a bounded linear functional L′ on X extending L such that ∥L′ ∥ ≤ C.

Proof. Apply the Hahn-Banach theorem with the Minkowski functional p(x) = C∥x∥.
□

Before obtaining further corollaries, we extend Thoerem 2.17 to complex normed

spaces. First, if X is a vector space over C, then trivially it is also a vector space over
R, and there is a simple relationship between the R- and C-linear functionals.
Lemma 2.22. Let X be a vector space over C.
(a) If L : X → C is a C-linear functional, then u(x) = real L(x) defines an R-linear
functional on X and L(x) = u(x) − iu(ix).
(b) Conversely, if u : X → R is R-linear then L(x) := u(x) − iu(ix) is C-linear.
(c) If L : X → C is a C-linear functional, p : X → R is a seminorm, and u = real L,
then |u(x)| ≤ p(x) for all x ∈ X if and only if |L(x)| ≤ p(x) for all x ∈ X .

Proof. Problem 2.5.

To prove the last statement, it is immediate that |u(x)| ≤ |L(x)| for all x ∈ X .
Conversely, given x there is a unimodular α such that αL(x) = |L(x)|. Hence,
|L(x)| = L(αx) = |u(αx)| ≤ p(αx) = |α| p(x) = p(x). □
Remark 2.23. Note that in passing from the real to the complex case, we must give
up the generality of a Minkowski functional and instead content ourselves with the
seminorm p.

END Friday 2025-01-17

Theorem 2.24 (The Hahn-Banach Theorem, complex version). Let X denote a complex
vector space, p a seminorm on X , and M a subspace of X . If L : M → C is a C-linear
functional satisfying |L(x)| ≤ p(x) for all x ∈ M, then there exists a C-linear functional
L′ : X → C such that
28

(i) L′ |M = L and
(ii) |L′ (x)| ≤ p(x) for all x ∈ X .

Sketch of proof. Using the Lemma 2.22 and its notation: The proof consists of applying
the real Hahn-Banach theorem (Corollary 2.21) to the R-linear functional u = ReL to
obtain a real linear functional u′ : X → R extending u and satisfying u′ (x) ≤ p(x) for all
x ∈ X . The resulting complex functional L′ associated to u′ is then a desired extension
of L. The details are left as an exercise. □

The following corollaries are quite important, and when the Hahn-Banach theorem
is applied it is usually in one of the following forms:
Corollary 2.25. Let X be a normed vector space over F (either R or C).
(i) If M ⊆ X is a subspace and L : M → F is a bounded linear functional, then there
exists a bounded linear functional L′ : X → F such that L′ |M = L and ∥L′ ∥ = ∥L∥.
(ii) (Linear functionals detect norms) If x ∈ X is nonzero, there exists L ∈ X ∗ with
∥L∥ = 1 such that L(x) = ∥x∥.
(iii) (Linear functionals separate points) If x ̸= y in X , there exists L ∈ X ∗ such that
L(x) ̸= L(y).
(iv) (Linear functionals detect distance to subspaces) If M ⊆ X is a closed subspace
and x ∈ X \ M, there exists L ∈ X ∗ such that
(a) L|M = 0;
(b) ∥L∥ = 1; and
(c) L(x) = dist(x, M) = inf y∈M ∥x − y∥ > 0.
(v) if L is a linear submanifold of X and x ∈ X , then x ∈ L if and only if λ(x) = 0
for every λ ∈ X ∗ for which L ⊆ ker λ.

Proof. To prove item (i) consider the (semi)norm p(x) = ∥L∥ ∥x∥. By construction,
|L(x)| ≤ p(x) for x ∈ M. Hence, there is a linear functional L′ on X such that
L′ |M = L and |L′ (x)| ≤ p(x) for all x ∈ X . In particular, ∥L′ ∥ ≤ ∥L∥. On the other
hand, ∥L′ ∥ ≥ ∥L∥ since L′ agrees with L on M.
For item (ii), let M be the one-dimensional subspace of X spanned by x. Define a
x
functional L : M → F by L(t ∥x∥ ) = t. In particular, |L(y)| = ∥y∥ for y ∈ M and thus
∥L∥ = 1. By (i), the functional L extends to a functional (still denoted L) on X such
that ∥L∥ = 1.
An application of item (ii) to the vector x − y proves item (iii).
To prove item (iv), let δ = dist(x, M). Since M is closed, δ > 0. Define a functional
L : M + Fx → F by L(y + tx) = tδ for y ∈ M and t ∈ F. Since for t ̸= 0 and y ∈ M,
∥y + tx∥ = |t|∥t−1 y + x∥ ≥ |t|δ = |L(y + tx)|,
by Hahn-Banach we can extend L to a functional L ∈ X ∗ with ∥L∥ ≤ 1.
 29

Let M = L. Thus M is a closed subspace of X . If λ ∈ X ∗ and L ⊆ ker λ, then, by

continuity, M ⊆ ker λ proving one direction of item (v). For the remaining direction,
apply item (iv) to x ∈ X \M to obtain a λ ∈ X ∗ such that M ⊆ ker λ, but λ(x) ̸= 0. □

2.4. The bidual and reflexive spaces. Note that since X ∗ is a normed space, we
can form its dual, denoted X ∗∗ , and called the bidual or double dual of X . There is
a canonical relationship between X and X ∗∗ . Each fixed x ∈ X gives rise to a linear
functional x̂ : X ∗ → F via evaluation,
x̂(L) := L(x).
Since |x̂(L)| = |L(x)| ≤ ∥L∥ ∥x∥, the linear functional x̂ is in X ∗∗ and ∥x̂∥ ≤ ∥x∥.
Corollary 2.26. (Embedding in the bidual) The map x → x̂ is an isometric linear map
from X into X ∗∗ .

Proof. First, from the definition we see that

|x̂(L)| = |L(x)| ≤ ∥L∥∥x∥
so x̂ ∈ X ∗∗ and ∥x̂∥ ≤ ∥x∥. It is straightforward to check (recalling that the L’s are
linear) that the map x → x̂ is linear. Finally, to show that ∥x̂∥ = ∥x∥, fix a nonzero
x ∈ X . From Corollary 2.25(i) there exists L ∈ X ∗ with ∥L∥ = 1 and L(x) = ∥x∥. But
then for this x and L, we have |x̂(L)| = |L(x)| = ∥x∥ so ∥x̂∥ ≥ ∥x∥, and the proof is
complete. □
Definition 2.27. A Banach space X is called reflexive if the map ˆ : X → X ∗∗ is
surjective.

In other words, X is reflexive if the mapˆ is an (isometric) isomorphism of X with

∗∗
X . For example, every finite dimensional Banach space is reflexive (Problem 2.6).
Reflexive spaces often have nice properties. For instance, the distance from a point to a
(closed) subspace is attained.
Needless to say, the proof of the Hahn-Banach theorem is thoroughly non-constructive,
and in general it is an important (and often difficult) problem, given a normed space X ,
to find some concrete description of the dual space X ∗ . Usually doing so means finding
a Banach space Y and a bounded (or, better, isometric) isomorphism T : Y → X ∗ .
Example 2.28. Recall ℓ1 = c∗0 isometrically and ℓ∞ = (ℓ1 )∗ isometrically by Proposi-
tions 2.5 and 2.7 Moreover it is straightfowrard to show that under the identification of
Corollary 2.26, the canonical map c0 → c∗∗ 0 corresponds to the natural inclusion of c0
∞ ∞
into ℓ . Since c0 is separable, but ℓ is not, c0 is not reflexive.
Example 2.29 (Banach Limits). The set
 
c = f : N → F lim f (n) exists ,
n→∞
30

is a subspace of ℓ∞ . The function L : c → F defined by

L(f ) = lim f (n)
n

is a linear functional and it satisfies |L(f )| ≤ ∥f ∥∞ . Hence L is continuous and ∥L∥ ≤ 1.

On the other hand, letting o : N → F denote the function that is constantly equal to
1, we see 1 = |L(0)| = ∥o∥. Hence ∥L∥ = 1. Thus, by the Hahn-Banach Extension
Theorem, L extends to a bounded linear functional on all of ℓ∞ of norm 1. Any such
extension of L is a Banach limit.
By example 2.6, elements λ ∈ (ℓ1 )∗ are precisely of the form Lf for some f ∈ ℓ∞ ,
f (n)g(n) for g ∈ ℓ1 . Thus g ∈ ℓ1 thus determines an element gb in
P
where Lf (g) =
(ℓ1 )∗∗ by
gb(f ) = Lf (g).
Let on : N → F denote the function on (j) = 0 for j ≤ n and on (j) = 1 for j > n and
observe,
X∞
gb(on ) = g(n)
j=n+1
and L(on ) = 1 for all n (where L is a Banach limit). It follows that L ̸= gb and therefore
the natural embedding of ℓ1 into (ℓ1 )∗∗ = (ℓ∞ )∗ is not onto.
In fact more is true. Namely, there is no isometric isomorphism between ℓ1 and
(ℓ∞ )∗ . As an outline of a proof, show, if X is a normed vector space and X ∗ is separable,
then X is also separable. This fact, applied to ℓ∞ , shows (ℓ∞ )∗ is not separable. Since
ℓ1 is separable, the result follows. □

END Monday 2025-01-24

Remark 2.30. After we have studied the Lp and ℓp spaces in more detail, we will see
that Lp is reflexive for 1 < p < ∞.
We note in passing that if X is reflexive, then its dual X ∗ has a unique predual:
that is, if Y is another Banach space and Y ∗ is isometrically isomorphic to X ∗ , then
in fact Y is isometrically isomorphic to X . However this conclusion can fail when X
is not reflexive; for example it turns out that ℓ1 does not have a unique predual. See
Problems 2.10 and 2.15. □

The embedding into the bidual has many applications; one of the most basic is the
following.
Proposition 2.31 (Completion of normed spaces). If X is a normed vector space, then
there is a Banach space X and an isometric linear map ι : X → X such that the image
ι(X ) is dense in X .

Proof. Embed X into X ∗∗ via the map x → x̂ and let X be the closure of the image of
X in X ∗∗ . Since X is a closed subspace of a complete space, it is complete. □
 31

The space X is called the completion of X . It is unique in the sense that if Y is

another Banach space and j : X → Y embeds X isometrically as a dense subspace of Y,
then Y is isometrically isomorphic to X . The proof of this fact is left as an exercise.

2.5. Dual spaces and adjoint operators. [Optional] Let X , Y be normed spaces
with duals X ∗ , Y ∗ . If T : X → Y is a linear transformation and f : Y → F is a linear
functional, then T ∗ f : X → F defined by
(10) (T ∗ f )(x) = f (T x)
is a linear functional on X . If T and f are both continuous (that is, bounded) then the
composition T ∗ f is bounded, and more is true:

Theorem 2.32. Let T : X → Y be a bounded linear transformation. The function

T ∗ : Y ∗ → X ∗ defined, for f ∈ Y ∗ , . Then:
(i) For f ∈ Y ∗ the function T ∗ f defined by the formula (10) is in X ∗ .
(ii) The mapping T ∗ : Y ∗ → X ∗ is a bounded linear map with ∥T ∗ ∥ = ∥T ∥.

Proof. Since T is assumed bounded, for a fixed f ∈ Y ∗ and all x ∈ X

|T ∗ f (x)| = |f (T x)| ≤ ∥f ∥∥T x∥ ≤ ∥f ∥∥T ∥∥x∥.
It follows that T ∗ f is bounded on X (thus, belongs to X ∗ ) and
(11) ∥T ∗ f ∥ ≤ ∥f ∥∥T ∥.
Thus T ∗ maps Y ∗ into X ∗ and it is straightforward to verify that T ∗ is linear. Moreover,
the inequality of equation (11) also shows that T ∗ is bounded and ∥T ∗ ∥ ≤ ∥T ∥.
It remains to show ∥T ∗ ∥ ≥ ∥T ∥. Toward this end, let 0 < ϵ < 1 be given and choose
x ∈ X with ∥x∥ = 1 and ∥T x∥ > (1 − ϵ)∥T ∥. Now consider T x. By the Hahn-Banach
Theorem (Corollary 2.25(i)), there exists f ∈ Y ∗ such that ∥f ∥ = 1 and f (T x) = ∥T x∥.
For this f ,
∥T ∗ ∥ ≥ ∥T ∗ f ∥ ≥ |T ∗ f (x)| = |f (T x)| = ∥T x∥ > (1 − ϵ)∥T ∥.
Hence, ∥T ∗ ∥ ≥ (1 − ϵ)∥T ∥. Since ϵ was arbitrary, ∥T ∗ ∥ ≥ ∥T ∥. □

END Monday 2025-01-27 also covered Proposition 2.12

2.6. Duality for Sub and Quotient Spaces. [Optional. Not covered Spring 2025]
The Hahn-Banach Theorem allows for the identification of the duals of subspaces and
quotients of Banach spaces. Informally, the dual of a subspace is a quotient and the dual
of a quotient is a subspace. The precise results are stated below for complex scalars, but
they hold also for real scalars.
32

Given a (closed) subspace M of the Banach space X , let π denote the map from X
to the quotient X /M. Recall (see Problem 1.22), the quotient is a Banach space with
the norm,
∥z∥ = inf{∥y∥ : π(y) = z}.
In particular, if x ∈ X , then
∥π(x)∥ = inf{∥x − m∥ : m ∈ M}.
It is evident from the construction that π is continuous and ∥π∥ ≤ 1. Further, by
Problem 1.20 (or see Proposition 2.33 below) if M is a proper (closed) subspace, then
∥π∥ = 1. In particular, π ∗ : (X /M)∗ → X ∗ (defined by π ∗ λ = λ ◦ π) is also continuous.
Moreover, if x ∈ M, then
π ∗ λ(x) = λ(π(x)) = 0.
Let
M⊥ = {f ∈ X ∗ : f (x) = 0 for all x ∈ M}.
(M⊥ is called the annihilator of M in X ∗ .) Recall, given x ∈ X , the element x̂ ∈ X ∗∗
is defined by x̂(τ ) = τ (x), for τ ∈ X ∗ . In particular,
M⊥ = ∩x∈M ker(x̂)
and thus M⊥ is a closed subspace of X ∗ . Further, if λ ∈ (X /M)∗ , then π ∗ λ ∈ M⊥ .
Proposition 2.33 (The dual of a quotient). The mapping ψ : (X /M)∗ → M⊥ defined
by
ψ(λ) = π ∗ λ
is an isometric isomorphism; i.e., the mapping π ∗ : (X /M)∗ → X ∗ is an isometric
isomorphism onto M⊥ .

Informally, the proposition is expressed as (X /M)∗ = M⊥ .

Proof. The linearity of ψ follows from Theorem 2.32 as does ∥ψ∥ = ∥π∥ ≤ 1. To prove
that ψ is isometric, let λ ∈ (X /M)∗ be given. Automatically, ∥ψ(λ)∥ ≤ ∥λ∥. To prove
the reverse inequality, fix r > 1. Let q ∈ X /M with ∥q∥ = 1 be given. There exists an
x ∈ X such that ∥x∥ < r and π(x) = q. Hence,
|λ(q)| = |λ(π(x))∥ = ∥ψ(λ)(x)∥ ≤ ∥ψ(λ)∥ ∥x∥ < r∥ψ(λ)∥.
Taking the supremum over such q shows ∥λ∥ ≤ r∥ψ(λ)∥. Finally, since 1 < r is arbitrary,
∥λ∥ ≤ ∥ψ(λ)∥.
To prove that ψ is onto, and complete the proof, let τ ∈ M⊥ be given. Fix q ∈ X /M.
If x, y ∈ X and π(x) = q = π(y), then τ (x) = τ (y). Hence, the mapping λ : X /M → C
defined by λ(q) = τ (x) is well defined. That λ is linear is left as an exercise. To see that
λ is continuous, observe that
|λ(q)| = |τ (x)| ≤ ∥τ ∥ ∥x∥,
 33

for each x ∈ X such that π(x) = q. Taking the infimum over such x gives shows
|λ(q)| ≤ ∥τ ∥ ∥q∥.
Finally, by construction ψ(λ) = τ. □

Since M⊥ is closed in X ∗ , the quotient space X ∗ /M⊥ is a Banach space. Let

ρ : X ∗ → X ∗ /M⊥ denote the quotient mapping. Suppose λ ∈ M∗ . By Corollary 2.25,
there is an f ∈ X ∗ such that f |M = λ; that is f is a bounded extension of λ (and
indeed f can be chosen such that ∥f ∥ = ∥λ∥). If f and g are two extensions of λ to
bounded linear functionals on X ∗ , then f (x) − g(x) = 0 for x ∈ M. Hence f − g ∈ M⊥
or equivalently, ρ(f ) = ρ(g). Consequently, the mapping φ : M∗ → X ∗ /M⊥ defined
by φ(λ) = ρ(f ) (where f is any bounded extension of λ to X ) is well defined. It is
easily verified that φ is linear. Further, given q ∈ X ∗ /M⊥ , there is an f ∈ X ∗ such that
ρ(f ) = q. In particular, with λ = f |M we have φ(λ) = ρ(f ). Therefore φ is onto.

Proposition 2.34 (The dual of a subspace). The mapping φ : M∗ → X ∗ /M⊥ is an

isometric isomorphism.

Proof. It remains to show that φ is an isometry, a fact that is an easy consequence of the
Hahn-Banach Theorem. Fix λ ∈ M∗ and let q = φ(λ). If f is any bounded extension
of λ to X ∗ , then ∥f ∥ ≥ ∥λ∥. Hence,
∥φ(λ)∥ =∥q∥
= inf{∥f ∥ : f ∈ X ∗ , ρ(f ) = q}
= inf{∥f ∥ : f ∈ X ∗ , f |M = λ}
≥∥λ∥.
On the other hand, by the Hahn-Banach Theorem there is a bounded extension g of λ
with ∥g∥ = ∥λ∥. Thus ∥λ∥ ≤ ∥q∥. □

A special case of the following useful fact was used in the proofs above. If X , Y are
vector spaces and T : X → Y is linear and M is a subspace of the kernel of T , then T
induces a linear map Te : X /M → Y. A canonical choice is M = ker(T ) in which case
Te is one-one. If X is a Banach space, Y is a normed vector space and M is closed, then
X /M is a Banach space.

Lemma 2.35. If X is a Banach space, M is a (closed) subspace of ker(T ), Y is a

normed vector space and T : X → Y is continuous, then the mapping Te is bounded and
∥Te∥ = ∥T ∥.

Proof. Let π : X → X /M denote the quotient map and observe that Teπ = T . Since the
quotient map π has norm 1 (see Problem 1.22), we see that ∥Te∥ ≤ ∥T ∥. For the opposite
34

inequality, let 0 < ϵ < 1 and choose x ∈ X such that ∥x∥ = 1 and ∥T x∥ > (1 − ϵ)∥T ∥.
Then ∥π(x)∥ ≤ 1 and
∥T̃ ∥ ≥ ∥T̃ π(x)∥ = ∥T x∥ > (1 − ϵ)∥T ∥.
Letting ϵ go to zero finishes the proof. □

2.7. Hahn-Banach separation theorems. [Optional. Not covered Spring 2025]

Besides the extension theorem and its corollaries, the other important applications
of the Hahn-Banach theorem consist of various separation theorems. We begin with a
few definitions.
Definition 2.36. Let X be a vector space. A hyperplane in X is a subspace M of
codimension 1. An affine hyperplane is a set of the form x + M ⊆ X, for some fixed
x ∈ X and hyperplane M.

If L : X → F is a nonzero linear functional (bounded or not), the space M = ker L

is a hyperplane, and if we fix any scalar t ∈ F then the set {x ∈ X : L(x) = t} is an affine
hyperplane. Conversely, any hyperplane is the kernel of a nonzero linear functional. (To
see this, observe that if M is a hyperplane in X , then, since it has codimension 1, for
any fixed choice of a vector y ∈ X \ M we can write every x ∈ X uniquely as x = m + ty
with m ∈ M and t ∈ F. Then define L(x) = t.) Consequently, every affine hyperplane
has the form H = {x ∈ X : L(x) = t} for some nonzero linear functional L and some
scalar t.
Lemma 2.37. If M is a hyperplane in a normed vector space X , then M is either
closed, or dense in X .

Proof. It is easy to check that the closure of subspace of X is again a subspace. It follows
that M is a subspace with M ⊆ M ⊆ X . Since M has codimension 1, we must have
either M = M or M = X . □
Proposition 2.38. Let X be a normed vector space and L : X → F a linear functional.
Then L is continuous (that is, bounded) if and only if ker L is closed. Consequenlty, L
is continuous if and only if there exists a nonempty open set U such that U ∩ ker L = ∅.

Proof. Trivially, if L is continuous then ker L is closed. Conversely, suppose M = ker L

is closed. We can then form the quotient space X /M, and since M is a hyperplane
this space is one-dimensional. If we let π denote the quotient map and define L̃ :
X /M → F by L̃π = L, then the linear functional L̃ is continuous (since its domain is
finite-dimensional), and since the quotient map is also continuous, we conclude that L
is continuous.
The second statement follows by combining the first statement with Lemma 2.37.
□
 35

Recall that a set K in a real vector space X is called convex if for every x, y ∈ K and
every 0 ≤ t ≤ 1, we have tx + (1 − t)y ∈ K. Let X be a normed vector space over R and
let U ⊆ X be a convex, open set containing 0. We define a function p : X → [0, +∞) by
1
(12) p(x) = inf{r > 0 : x ∈ U }.
r
(To see that the definition makes sense, observe that since U is open and 0 ∈ U , there
exists δ > 0 so that x ∈ U whenever ∥x∥ < δ. It follows that for every x ∈ X , we
have 1r x ∈ U for all r > ∥x∥
δ
; thus the set appearing in the definition is nonempty.) The
functional p is sometimes called a gauge for the set U , it is important because of the
following lemma.
Lemma 2.39. Let X be a normed vector space over R and let U ⊆ X be a convex open
set containing 0. Then the function p defined by (12) is a Minkowski functional, and
U = {x ∈ X : p(x) < 1}.

Proof. If r, s > 0 then trivially rs x ∈ U if and only if r/s1

x ∈ U , and it follows that
p(sx) = sp(x) for all s > 0. Likewise it is immediate from the definition of p that
p(0) = 0, so that p(sx) = sp(x) for all s ≥ 0. Next we show that p(x) < 1 if and only if
x ∈ U : indeed, if x is in U then since U is open, there is a δ > 0 such that (1 + δ)x ∈ U ,
thus p(x) ≤ (1 + δ)−1 < 1. On the other hand if p(x) < 1 then 1r x ∈ U for some
0 < r < 1, but then since U is convex and 0 ∈ U , we can write x = r( xr ) + (1 − r) · 0 ∈ U .
x y
Finally, let us show that p(x + y) ≤ p(x) + p(y). Fix any r, s > 0 such that r
and s
belong to U . Since U is convex, the convex combination
   
r x s y x+y
+ =
r+s r r+s s r+s
belongs to U , so by what was just proved we have p( x+y
r+s
) < 1. By homogeneity we
conclude that p(x + y) < r + s, and finally by taking the infimum over r and s we get
p(x + y) ≤ p(x) + p(y).
□
Theorem 2.40 (Separation). Let X be a normed vector space over R. If U ⊆ X is a
nonempty, open, convex set, and x ∈ X \ U , then there exists a bounded linear functional
L ∈ X ∗ and a real number a such that L(y) < a = L(x) for all y ∈ U .

Proof. We first assume that 0 ∈ U , the general case will follow by translation. Let
N be the one-dimensional subspace Rx. Define L on N by putting L(x) = 1 and
extending linearly. Let p be the gauge functional for U . Since x ∈
/ U , we have p(x) ≥ 1,
so 1 = L(x) ≤ p(x). Since both L and p are positive homogeneous, we also have
L(tx) ≤ p(tx) for all t ≥ 0. For t < 0, we have L(tx) < 0 ≤ p(tx) (since p ≥ 0 by
definition). Thus, we have L(y) ≤ p(y) for all y in the subspace N . It follows from the
Hahn-Banach theorem that there exists an extension of L to all of X (still denoted L)
36

such that L(y) ≤ p(y) for all y ∈ X . It follows that L(x) = 1 and L(y) ≤ p(y) < 1 for
all y ∈ U . To see that this extension L is bounded, let V = U − x; then V is an open
set in X and L(y) < 0 for all y ∈ V , that is, V ∩ ker L = ∅, so by the second part of
Proposition 2.38 it follows that L is bounded.
Finally, in the case that U does not contain 0, we choose a point x0 ∈ U and apply
the theorem to U ′ := U − x0 and x′ = x − x0 , to obtain a bounded functional such that
L(x) = 1 + L(x0 ) and L(y) < 1 + L(x0 ) for all y ∈ U ; the details are left to the reader.
□
Lemma 2.41. i) If X is a normed vector space over R, K ⊆ X is a convex set,
and x0 is an interior point of K, then for every x ∈ K and every 0 ≤ t < 1, the
point x0 + t(x − x0 ) is an interior point of K.
ii) If K is convex then int(K) is convex.
iii) If K is a closed, convex subset of X and K has nonempty interior, then K is
equal to the closure of its interior.
iv) Let K be a closed, convex subset of X and suppose 0 is an interior point of K. If
p is the gauge functional for the convex set U = int(K), then K = {x ∈ X |p(x) ≤
1}.

Proof. For (i), by translation, there is no loss of generality in supposing that x0 = 0.

Fix x ∈ K, and fix δ such that y ∈ K for all ∥y∥ < δ. For 0 ≤ t < 1, put ϵ = (1 − t)δ.
If ∥z − tx∥ < ϵ, then we can write z = tx + y with ∥y∥ < (1 − t)δ. It follows that
∥(1−t)−1 y∥ < δ, so y ′ = (1−t)−1 y belongs to K. We have thus written z = tx+(1−t)y ′
with x, y ′ ∈ K, so z ∈ K. That is, the open ball B(tx, ϵ) is contained in K.
(ii) follows immediately from (i).
For (iii), since x = lim(x0 + tn (x − x0 )) for any sequence tn increasing to 1, we see
that every x ∈ K is a limit of interior points.
For (iv), let p(x) ≤ 1. If p(x) < 1 then x ∈ U and thus x ∈ K. If p(x) = 1, then by
the definiton of p we have tx ∈ U for every 0 ≤ t < 1, so taking a sequence of scalars
tn increasing to 1, we get that x belongs to the closure of U so x ∈ K. Conversely, if
x ∈ K, then by part (i) tx ∈ U for every 0 ≤ t < 1, so p(x) ≤ 1. □
Corollary 2.42 (Strict separation). Let X be a normed vector space over R. If K ⊆ X
is a closed, convex set with nonempty interior, and x ∈ X \K, then there exists a bounded
linear functional L ∈ X ∗ , and real numbers a < b such that L(y) ≤ a < b = L(x) for all
y ∈ K.

Proof. Again we assume that 0 is an interior point of K, and leave the general case to the
reader. Let U = int(K) and let p be the gauge functional for U ; by the lemma we have
K = {x ∈ X |p(x) ≤ 1}. Thus, if x ∈/ K, then p(x) > 1. We may then choose a number
0 < t < 1 so that tx ∈
/ K. Applying the previous separation theorem to U and tx, we
obtain a bounded linear functional L and a real number a such that L(y) < a = L(tx)
 37

for all y ∈ U , we put b := at = L(x). Since L(y) < a on U , and L is continuous, and
K is the closure of U (by item (iii) of the Lemma), we conclude that L(y) ≤ a for all
a ∈ K, which completes the proof. □

2.8. Problems.
Problem 2.1. Prove, if X is any normed vector space, {x1 , . . . xn } is a linearly indepen-
dent set in X , and α1 , . . . αn are scalars, then there exists a bounded linear functional f
on X such that f (xj ) = αj for j = 1, . . . n. (Recall linear maps from a finite dimensional
normed vector space to a normed vector space are bounded.)
Problem 2.2. Let X , Y be normed spaces and T : X → Y a linear transformation.
Prove T is bounded if and only if there exists a constant C such that for all x ∈ X and
f ∈ Y ∗,
(13) |f (T x)| ≤ C∥f ∥∥x∥;
in which case ∥T ∥ is equal to the best possible C in (13).
Problem 2.3. Let X be a normed vector space. Show that if M is a closed subspace of
X and x ∈ / M, then M + Fx is closed. Use this result to give another proof that every
finite-dimensional subspace of X is closed.
Problem 2.4. Prove, if M is a finite-dimensional subspace of a Banach space X , then
there exists a closed subspace N ⊆ X such that M ∩ N = {0} and M + N = X . (In
other words, every x ∈ X can be written uniquely as x = y + z with y ∈ M, z ∈ N .)
Hint: Choose a basis x1 , . . . xn for M and construct, using Problem 2.1 and the Hahn-
Banach Theorem, bounded linear functionals f1 , . . . fn on X such that fi (xj ) = δij . Now
let N = ∩ni=1 ker fi . (Warning: this conclusion can fail badly if M is not assumed finite
dimensional, even if M is still assumed closed. Perhaps the first known example is that
c0 is not complemented in ℓ∞ , though it is nontrivial to prove.)
Problem 2.5. Prove Proposition 2.22.
Problem 2.6. Prove every finite-dimensional Banach space is reflexive.
Problem 2.7. Let B denote the subset of ℓ∞ consisting of sequences which take values
in {−1, 1}. Show that any two (distinct) points of B are a distance 2 apart. Show, if C
is a countable subset of ℓ∞ , then there exists a b ∈ B such that ∥b − c∥ ≥ 1 for all c ∈ C.
Conclude ℓ∞ is not separable. Prove there is no isometric isomorphism Λ : c0 → ℓ∞ .
Problem 2.8. [This problem belongs in the section with signed measures] Prove, if µ
is a finite regular (signed) Borel measure on a compact Hausdorff space, then the linear
function Lµ : C(X) → R defined by
Z
Lµ (f ) = f dµ
X
38

is bounded (continuous) and ∥Lµ ∥ = ∥µ∥ := |µ|(X). (See the Riesz-Markov Theorem
for positive linear functionals.)

Problem 2.9. Let X and Y be normed vector spaces and T ∈ L(X , Y).
a) Consider T ∗∗ : X ∗∗ → Y ∗∗ . Identifying X , Y with their images in X ∗∗ and Y ∗∗ ,
show that T ∗∗ |X = T .
b) Prove T ∗ is injective if and only if the range of T is dense in Y.
c) Prove that if the range of T ∗ is dense in X ∗ , then T is injective; if X is reflexive
then the converse is true.
d) Assuming now that X and Y are Banach spaces, prove that T : X → Y is
invertible if and only if T ∗ is invertible, in which case (T ∗ )−1 = (T −1 )∗ .

Problem 2.10. a) Prove that if X is reflexive, then X ∗ is reflexive. (Hint: let

ι : X → X ∗∗ be the canonical inclusion; by assumption ι is invertible. Compute
(ι−1 )∗ .)
b) Prove that if X is reflexive and M ⊆ X is a closed subspace, then M is reflexive.
c) Prove that a Banach space X is reflexive if and only if X ∗ is reflexive.
d) Prove that if X is reflexive and Y is another Banach space with Y ∗ isometrically
isomorphic to X ∗ , then Y is isometrically isomorphic to X . (This conclusion can
fail if X is not reflexive; see Problem 2.15.)

Problem 2.11. Prove, if X is a Banach space and X ∗ is separable, then X is separable.

[Hint: let {fn } be a countable dense subset of X ∗ . For each n choose xn such that
∥xn ∥ = 1 and |fn (xn )| ≥ 21 ∥fn ∥. Show that the set of Q-linear combinations of {xn } is
dense in X .]

Problem 2.12. a) Prove there exists a bounded linear functional L ∈ (ℓ∞ )∗ with
the following property: whenever f ∈ ℓ∞ and limn→∞ f (n) exists, then L(f ) is
equal to this limit. (Hint: first show that the set of such f forms a subspace
M ⊆ ℓ∞ ). Such an L is a Banach limit.
b) Show that such a functional L is not equal to Lg for any g ∈ ℓ1 ; thus the map
T : ℓ1 → (ℓ∞ )∗ given by T (g) = Lg is not surjective.
c) Give another proof that T is not surjective, using Problem 2.11.

Problem 2.13. Let X be a normed space and let K ⊆ X be a convex set. (Recall,
this means that whenever x, y ∈ K, then 12 (x + y) ∈ K; equivalently, tx + (1 − t)y ∈ K
for all 0 ≤ t ≤ 1.) A point x ∈ K is called an extreme point of K whenever y, z ∈ K,
0 < t < 1, and x = ty + (1 − t)z, then y = z = x. (That is, the only way to write x as a
convex combination of elements of K is the trivial way.)
a) Let X be a normed space and let B = ball(X ) denote the (closed) unit ball of
X . Prove that x ∈ B is not an extreme point of B if and only if there exists a
nonzero y ∈ B such that ∥x ± y∥ ≤ 1.
 39

b) Prove that if X and Y are normed spaces, and T : X → Y is a surjective

linear isometry, (so that X and Y are isometrically isomorphic) then T induces
a bijection between the extreme points of ball(X ) and ball(Y).
c) Let ℓpn denote the (real) Banach space Rn equipped with the ℓp norm, 1 ≤ p ≤ ∞.
Prove that ℓ12 and ℓ∞2 are isometrically isomorphic, but that there is no isometry
1 ∞
between ℓ3 and ℓ3 .
Problem 2.14. a) Show that the extreme points of the unit ball of ℓ1 are precisely
the points of the form λen where |λ| = 1 and en is the sequence which is 1 in the
nth entry and 0 elsewhere. (See Problem 2.13).
b) Determine the extreme points of the unit ball of ℓ∞ .
c) Show that the unit ball of c0 has no extreme points.
Problem 2.15. Let  
c= f : N → F lim f (n) exists ,
n→∞

and equip c with the supremum norm ∥f ∥∞ := sup |f (n)|.

a) Show that c∗ ∼
= ℓ1 isometrically.
b) Prove that c is boundedly isomorphic to c0 .
c) Prove that c is not isometrically isomorphic to c0 . (Hint: examine the extreme
points of the unit balls of c and c0 ; see Problems 2.13 and 2.14 .)
(This problem provides an example of Banach spaces X and Y such that X and Y are
not isometrically isomorphic, but X ∗ and Y ∗ are. So in general we cannot recover X
(isometrically) from X ∗ . In fact the situation is worse, ℓ1 has isometric preduals which
are not even boundedly isomorphic to c0 , but the construction is more involved and
outside the scope of these notes.)

3. The Baire Category Theorem and applications

This section contains three important applications of the Baire category theorem in
functional analysis. These are the Principle of Uniform boundedness (also known as the
Banach-Steinhaus theorem), the Open Mapping Theorem, and the Closed Graph The-
orem. (In learning these theorems, keep careful track of what completeness hypotheses
are needed.)

3.1. Baire’s Theorem. Recall, a set D in a metric space X is dense (in X) if D = X.

Lemma 3.1 below should be familiar. We will use the notation B(x, r), for the open ball
of radius r > 0 center to the point x in a metric space X = (X, d)
B(x, r) = {y ∈ X : d(x, y) < r};
and F ◦ for the interior of a subset F of a metric space X.
Lemma 3.1. Suppose X is a metric space.
40

(a) For a subset D ⊆ X of X, the following are equivlalent:

(i) D is dense in X;
(ii) Dc does not contain a nonempty open set ((Dc )◦ = ∅;
(iii) if ∅ ̸= U is open, then D ∩ U ̸= ∅.
(b) If U ⊆ X is open and x ∈ U, then there is an r > 0 such that B(x, r) ⊆ U.
(c) A subset F of X is closed with empty interior if and only if F c is open and dense.

Theorem 3.2 (The Baire Category Theorem). Suppose X is a complete metric space.
(a) If (Un )∞ ∞
n=1 is a sequence of open dense subsets of X, then ∩n=1 Un ̸= ∅.

(b) If (Fn )n is a sequence of closed sets with empty interior, then ∪Fn ̸= X.
Remark 3.3. We will actually prove that ∩Un is dense in X. This conclusion is in fact
equivalent to the conclusion that ∩Un ̸= ∅.
Theorem 3.2 is true if X is a locally compact Hausdorff space and there are connec-
tions between the Baire Category Theorem and the axiom of choice. □

The following lemma should be familiar from advanced calculus. It will be used in
the proof of Theorem 3.2.
Lemma 3.4. Let X be a complete metric space and suppose (Cn ) is a sequence of subsets
of X. If
(i) each Cn is nonempty;
(ii) (Cn ) is nested decreasing;
(iii) each Cn is closed; and
(iv) (diam(Cn )) converges to 0,
then there is an x ∈ X such that
{x} = ∩Cn .
Moreover, if (xn ) is a sequence from X and xn ∈ Cn for each n, then (xn ) converges to
some x.

Proof of Theorem 3.2. The two items are equivalent, but for our purposes it is enough to
show item (a) implies item (b). To this end, suppose item (a) holds, (Fn ) is a sequence
of closed sets such that Fn◦ = ∅ for all n and X = ∪Fn . Taking complements, ∅ = ∩Fnc .
By Lemma 3.1, the sets Fnc are open and dense in X by Lemma 3.1. Hence X is not
complete and therefore item (a) implies item (b). Thus it suffices to prove item (a).
To prove item (a), let (Un )∞
n=1 be a sequence of open dense sets in X and let I = ∩Un .
To prove I is dense, it suffices to show that I has nontrivial intersection with every
nonempty open set W by Lemma 3.1. Fix such a W . Since, by Lemma 3.1, U1 is dense,
 41

there is a point x1 ∈ W ∩ U1 . Since U1 and W are open, there is a radius 0 < r1 < 1 such
that the B(x1 , r1 ) is contained in W ∩ U1 by Lemma 3.1. Similarly, since U2 is dense
and open there is a point x2 ∈ B(x1 , r1 ) ∩ U2 and a radius 0 < r2 < 12 such that

B(x2 , r2 ) ⊆ B(x1 , r1 ) ∩ U2 ⊆ W ∩ U1 ∩ U2 .

Continuing inductively, since each Un is dense and open there is a sequence of points
(xn )∞ 1
n=1 and radii 0 < rn < n such that

B(xn , rn ) ⊆ B(xn−1 , rn−1 ) ∩ Un ⊆ W ∩ (∩nj=1 Un ).

The sequence of sets (B(xn , rn )) satisfies the hypothesis of Lemma 3.4 and X is complete.
Hence there is an x ∈ X such that

x ∈ ∩n B(xn , rn ) ⊆ W ∩ I. □

3.2. Category. Baire’s theorem is used as a kind of pigeonhole principle: the “thick”
complete metric space X cannot be expressed as a countable union of “thin” closed sets
without interior.

Definition 3.5. A subset E of a metric space X is nowhere dense (in X) if its closure
has empty interior; that is (E)◦ = ∅.
A set F in a metric space X is first category (or meager ) if it can be expressed as a
countable union of nowhere dense sets. In particular, a countable union of first category
sets is first category.
A set G is second category if it is not first category.

Corollary 3.6 (The Baire Category Theorem restated). If X is a complete metric space,
then X is not a countable union of nowhere dense sets; that is, X is of second category
in itself.

Proof. Suppose X = ∪∞ n=1 En where each En is nowhere dense. It follows that X =

∞
∪n=1 Fn , where each Fn = En is closed and with empty interior. Hence, by Theorem 3.2
item (b), X is not complete. □

Corollary 3.7. An infinite dimensional Banach space can not have a countable basis.
In particular, there is no norm on c00 that makes it a Banach space; ditto for the vector
space of polynomials.

Proof. The proof is left as an exercise. As a starting point, show, if M is finite dimen-
sional subspace of an infinite dimensional Banach space X , then M is nowhere dense in
X. □
42

3.3. The Principle of Uniform Boundedness.

Theorem 3.8 (The Principle of Uniform Boundedness (PUB)). Suppose X , Y are

normed spaces and {Tα : α ∈ A} ⊆ B(X , Y) is a collection of bounded linear trans-
formations from X to Y. Let B denote the set
(14) B := {x ∈ X : sup ∥Tα x∥ < ∞}.
α

If B is of the second category (thus not a countable union of nowhere dense sets) in X,
then
sup ∥Tα ∥ < ∞.
α

In particular, if X is complete and if the collection {Tα : α ∈ A} is pointwise bounded,

then it is uniformly bounded.

Proof. For notational convenience, set M (x) = supα ∥Tα x∥ < ∞.

For each integer n ≥ 1 consider the set
Vn := {x ∈ X : M (x) > n}.
Since each Tα is continuous (bounded), the sets Vn are open. (Indeed, for each α the map
x → ∥Tα x∥ is continuous from X to R, so if ∥Tα x∥ > n for some α then also ∥Tα y∥ > n
for all y sufficiently close to x.) Let En denote the complement of Vn and observe that
B = ∪∞ n=1 En . Since B is assumed to be of the second category, there is an N such that
◦
(EN ) is not empty. Since EN is closed, it follows that EN has nonempty interior; i.e.,
there is an x0 ∈ EN and r > 0 so that B(x0 , r) ⊆ EN . α and every ∥x∥ < r, expressing
x = (x − x0 ) + x0 as the sum of two elements of B(x0 , r) gives
∥Tα x∥ ≤ ∥Tα (x − x0 )∥ + ∥Tα x0 ∥ ≤ N + N.
That is, if ∥x∥ < r, then M (x) ≤ 2N . By rescaling we conclude that if ∥x∥ < 1, then
∥Tα x∥ ≤ 2N/r for all α and thus supα ∥Tα ∥ ≤ 2N/r < ∞. □

The following result is one of the many corollaries to the PUB.

Corollary 3.9. Suppose X is a Banach space and Y is a normed vector space. If (Tn )
is a sequence of bounded operators Tn : X → Y that converges pointwise to a (necessarily
linear) map T : X → Y, then T is bounded.

Outline of proof. In a metric space, convergent sequences are bounded. Hence (Tn x)n is
bounded in Y for each x ∈ X . Thus the set X = {x ∈ X : sup{∥Tn x∥ : n ∈ N} < ∞}
is of second category in X . Thus C = sup{∥Tn ∥ : n ∈ N} < ∞. Thus the proof reduces
to showing ∥T x∥ ≤ C∥x∥ for all x ∈ X , a task that is left as an exercise for the gentle
reader. □
 43

3.4. Open mapping. Given a subset B of a vector space X and a scalar s ∈ F, let
sB = {sb : b ∈ B}. Similarly, for x ∈ X , let B − x = {b − x : b ∈ B}. Let X , Y be
normed vector spaces and suppose T : X → Y is linear. If B ⊆ X and s ∈ F is nonzero,
then T (sB) = sT (B) and further, an easy argument shows T (sB) = s T (B). It is also
immediate that if B is open, then so is B − x.
Recall that if X, Y are topological spaces, a mapping f : X → Y is called open if
f (U ) is open in Y whenever U is open in X. In particular, if f is a bijection, then f is
open if and only if f −1 is continuous. In the case of normed linear spaces the condition
that a linear map is open is refined by the Open Mapping Theorem, Theorem 3.11 below.
But first a lemma.
Lemma 3.10 (Translation and Dilation lemma). Let X , Y be normed vector spaces, let
B denote the open unit ball of X , and let T : X → Y be a linear map. The following are
equivalent.
(i) The map T is open;
(ii) T (B) contains an open ball centered at 0;
(iii) there is an s > 0 such that T (sB) contains an open ball centered at 0; and
(iv) T (sB) contains an open ball centered at 0 for each s > 0.

In the proof of this lemma and that of Theorem 3.11 to follow, we use B X (x, r) and
B Y (y, s) to denote the open balls centered to x and y with radii r and s in X and Y
respectively when needed to avoid ambiguity.

Proof. This result is more or less immediate from the fact that, for fixed z0 ∈ X and
r ∈ F, the translation map z → z + z0 and the dilation map z → rz are continuous in a
normed vector space.
The implication item (i) implies item (ii) is immediate. The fact that T (sB) =
sT (B) for s > 0 readily shows items (ii), (iii) and (iv) are all equivalent.
To finish the proof it suffice to show item (iv) implies item (i). Accordingly, suppose
item (iv) holds and let U ⊆ X be a given open set. To prove that T (U ) is open, let
y ∈ T (U ) be given. There is an x ∈ U such that T (x) = y. There is an s > 0 such that
the ball B(x, s) lies in U ; that is B(x, s) ⊆ U . The ball B(0, s) = B(x, s) − x is an open
ball centered to 0. By hypothesis there is an r > 0 such that B Y (0, r) ⊆ T (B(0, s)). By
linearity of T ,
B Y (y, r) =B Y (0, r) + y ⊆ T (B(0, s)) + y
=T (B(0, s)) + T (x) = T (B(0, s) + x) = T (B(x, s)) ⊆ T (U ).
Thus T (U ) is open and the proof is complete. □

Theorem 3.11 (Open Mapping). Suppose that X is a Banach space, Y is a normed

vector space and T : X → Y is bounded. If the range of T is of second category, then
44

(i) T (X ) = Y;
(ii) Y is complete (so a Banach space); and
(iii) T is open.
In particular, if X , Y are Banach spaces, and T : X → Y is bounded and onto, then
T is an open map.

Proof. Assuming T is open and letting B denote the open unit ball in X , by Lemma 3.10,
there is an r > 0 such that B Y (0, r) ⊆ T (B). Hence,
Y = ∪∞ Y
n=1 B (0, n r) ⊆ T (nB) ⊆ T (X ),

so that item (i) holds. (Here the superscript Y is used to emphasize this ball is in Y.)
That is, item (iii) implies item (i).
To prove item S∞ r centered at x in X .
S∞ (iii), let B(x, r) denote the open ball of radius
Trivially X = n=1 B(0, n) and thus range T = T (X ) = n=1 T (B(0, n)). Since the
range of T is assumed second category, there is an N such that T (B(0, N )) is second
category and hence somewhere dense. In other words, T (B(0, N )) has nonempty interior.
By scaling (see Lemma (3.10)), T (B(0, 1)) has nonempty interior. Hence, there exists
p ∈ Y and r > 0 such that T (B(0, 1)) contains the open ball B Y (p, r). It follows that
for all ∥y∥ < r,
y = (y + p) − p ∈ T (B(0, 1)) + T (B(0, 1)) ⊆ T (B(0, 2)),
where we have used −T (B(0, 1)) = T (B(0, 1)). In other words,
B Y (0, r) ⊆ T (B(0, 2)).
By scaling, it follows that, for n ∈ N,
r 1
B Y (0, ) ⊆ T (B(0, )).
2n+1 2n
END Friday 2025-01-31
We will use the hypothesis that X is complete to prove B Y (0, 4r ) ⊆ T (B(0, 1)), which,
by Lemma 3.10, implies T is open. Accordingly let y such that ∥y∥ < 4r be given. Since
y is in the closure of T (B(0, 12 )), there is a y1 ∈ T (B(0, 21 )) such that ∥y − y1 ∥ < 8r . Since
y − y1 ∈ B Y (0, 8r ) it is is in the closure of T (B(0, 41 )). Thus there is a y2 ∈ T (B(0, 41 ))
such that ∥(y − y1 ) − y2 ∥ < 16 r
. Continuing in this fashion produces a sequence (yj )∞ j=1
from Y such that,
(a) ∥y − nj=1 yj ∥ ≤ 2n+2 r
P
; and
1
(b) yn ∈ T (B(0, 2n ))
for all n. It follows the sequence (sm = m
P
j=1 yj )m converges to y. P Further, for each
j there is an xj ∈ B(0, 2j ) such that yj = T xj . Thus, setting tm = m
1
j=1 xj , we have
 45

T sm = tm and
∞
X ∞
X
∥xj ∥ < 2−k = 1,
j=1 k=1
P∞
thus the sequence (tm )m converges to some x with ∥x∥ ≤ j=1 ∥xj ∥ < 1, that is,
x ∈ B(0, 1). It follows that y = T x by continuity of T . Consequently y ∈ T (B(0, 1))
and the proof of item (iii) is complete.
To prove item (ii), let M denote the kernel of T and Te the mapping Te : X /M → Y
determined by Teπ = T ; that is Te(π(x)) = T (x) for x ∈ X . By construction Te is one-one
and by Lemma 2.35, it is continuous. Further its range is the same as the range of T ,
namely Y, and is thus second category. Hence, by what has already been proved, Te is
an open map. and consequently Te−1 is continuous. Hence X /M and Y are isomorphic
(though of course not necessarily isometrically isomorphic) as normed vector spaces.
Therefore, since X /M is complete (Proposition 1.29), so is Y. □

Note that the proof of item (ii) in the Open Mapping Theorem shows, in the case
that in the case that T is one-one and its range is of second category, that T is onto and
its inverse is continuous. In particular, if T : X → Y is a continuous bijection and Y is
a Banach space (so the range of T is second category), then T −1 is continuous.

Corollary 3.12 (The Banach Isomorphism Theorem). If X , Y are Banach spaces and
T : X → Y is a bounded bijection, then T −1 is also bounded (hence, T is an isomor-
phism).

Proof sketch. Note that when T is bijection, T is open if and only if T −1 is continuous.
The result thus follows from the Open Mapping Theorem and Proposition 1.32. □

The following examples show that the hypothesis that X and Y are Banach spaces,
and not just normed vector spaces, is needed in Corollary 3.12.

Example 3.13. This example shows that the assumption that the range of T is second
category in Y is necessary in Theorem 3.11.
Let X denote the Banach space ℓ1 and let Y = ℓ1 as a linear manifold in c0 with
the c0 (sup) norm. So Y is a normed space, but not a Banach space. The identity map
ιX toY is a bijection. It is also continuous since the supremum norm of an element of ℓ1
dominates its ℓ1 norm. Let G = B X (0, 1) ⊆ X denote the (open) unit ball in X. Thus
G is open in X . Given r > 0 choose n ∈ N such that n > 2r and let x = 2r nj=1 ej , where
P

ej ∈ ℓ1 is the function ej : N → F defined by

(
1 if m = j
ej (m) =
0 if m ̸= j.
46

Observe that ∥x∥∞ < r, but ∥x∥1 > 1. Hence x ∈ B Y (0, r), but x ∈/ G. Thus B Y (0, r) ̸⊆
G for any choice of r > 0, which means 0 is not an interior point of G (in Y). Hence G
is not open in Y.
The argument above of course shows that the range of ι is not second category Pn in
1
Y. Here is simple direct proof of this fact. First note that the sequence gn = n j=1 ej
converges to 0 in c0 and gn ∈ ℓ1 with ∥gn ∥1 = 1. For N ∈ N, let BN = {f ∈ ℓ1 : ∥f ∥1 ≤
N } ⊆ Y. Verify that each BN is closed in Y. On the other hand, given N, the sequence
fk = f + 3N gk converges to f in c0 and so in Y, but ∥f + 3N gk ∥ ≥ 3N − ∥f ∥ ≥ 2N > N.
Thus f is not in the interior of BN and so BN is nowhere dense and Y = ∪∞ N =1 BN .

Example 3.14. This example shows that the assumption X is a Banach space can not
be relaxed to X is simply a normed vector space in Theorem 3.11.
Let Y be an infinite dimensional Banach space. Let λ be a discontinuous linear
functional, whose existence is the content of Proposition 2.14. As an exercise, show that
the function ∥ · ∥∗ : Y → [0, ∞) given by
∥x∥∗ = ∥x∥Y + |λ(x)|
is a norm on Y. Let X denote the normed space Y with this norm; that is X = (Y, ∥·∥∗ ).
Let T : X → Y denote the identity map (so bijective). Let G denote the unit ball in X .
We claim 0 is not in the interior of G as a subset of Y. Indeed, since λ is not continuous,
there is a sequence (xn ) of unit vectors in Y such that |λ(x)| ≥ n. Consequently, given
r > 0 and choosing n sufficiently large, ∥ 2r xn ∥Y < r, but ∥ 2r xn ∥∗ > 1. Thus B(0, r) ̸⊆ G
proving the claim.
A consequence of the argument is that X is not a Banach space. To verify this fact
directly, let y ∈
/ ker λ be given. By Proposition prop:bdd-lf-iff-cns, there is a sequence
(xn ) from ker λ that converges to x (in Y). Since ∥xn ∥∗ = ∥x∥Y , the sequence (xn ) is
Cauchy in X . However, since ∥xn − y∥∗ = ∥xn − y∥Y − λ(y), the sequence (xn ) does
not converge to y in X . Now suppose (xn ) converged to some z ∈ X . Thus z ∈ Y and
∥xn − z∥∗ = ∥xn − z∥Y + λ(z) converges to 0 from which it follows that z = y and the
proof is complete.
This result depends on the axiom of choice. In this proof, choice is smuggle in
through the appeal to Propostion 2.14, whose proof in turn depended on the existence
of a Hamel basis, which in turn uses Zorn’s Lemma (choice). □
END Monday 2025-02-03 - except had not discussed the Banach Isomorphism The-
orem.

3.5. The Closed Graph Theorem. Recall that the Cartesian product X × Y of Ba-
nach spaces X and Y with its default product topology from Subsection 1.2.7. In
particular, the product topology on X × Y is the coarsest topology that makes both
coordinate projections πX and πY from X × Y to X and Y respectively continuous. This
topology is the same as that determined by the norms in equation (2).
 47

Definition 3.15. The graph of a linear map T : X → Y between normed vector spaces
is the set
G(T ) := {(x, y) ∈ X × Y : y = T x}.
Observe that since T is a linear map, G(T ) is a linear subspace of X × Y. The transfor-
mation T is closed if G(T ) is a closed subset of X × Y. □

It is an easy exercise to show that G(T ) is closed if and only if whenever (xn , T xn )
converges to (x, y), we have y = T x. Problem 3.2 gives an example where G(T ) is closed,
but T is not continuous. On the other hand, the next theorem says that if X , Y are
complete (Banach spaces), then G(T ) is closed if and only if T is continuous.
Theorem 3.16 (The Closed Graph Theorem). If X , Y are Banach spaces and T : X →
Y is closed, then T is bounded.

Proof. We prove T closed implies T is bounded, leaving the easy converse as an exercise.
With the norm ∥(x, y)∥∞ = max{∥x∥, ∥y∥} the vector space X ×Y is a Banach space with
the product topology. The coordinate projections πX , πY are bounded with norm one.
Let π1 , π2 be the coordinate projections πX , πY restricted to G(T ); explicitly π1 (x, T x) =
x and π2 (x, T x) = T x. Note that π1 is a bijection between G(T ) and X and in particular
π1−1 (x) = (x, T x). By hypothesis G(T ) is a closed subset of a Banach space and hence
a Banach space. Thus π1 is a bounded linear bijection between Banach spaces and
therefore, by Corollary 3.12, π1−1 : X → G(T ) is bounded. Since π2 is bounded, π2 ◦π1−1 :
X → Y is continuous. To finish the proof, observe π2 ◦ π1−1 (x) = π2 (x, T x) = T x. □

3.6. Problems.
Problem 3.1. Show that there exists a sequence of open, dense subsets Un ⊆ R such
that m( ∞
T
n=1 Un ) = 0.

Problem 3.2. Consider the linear subspace D ⊆ c0 defined by

D = {f ∈ c0 : lim |nf (n)| = 0}
n→∞

and the linear transformation T : D → c0 defined by (T f )(n) = nf (n).

a) Prove T is closed, but not bounded. b) Prove T is bijective and T −1 : c0 → D is
bounded (and surjective), but not open. c) What can be said of D as a subset of c0 ?
Problem 3.3. Suppose X is a vector space equipped with two norms ∥ · ∥1 , ∥ · ∥2 such
that ∥ · ∥1 ≤ ∥ · ∥2 . Prove that if X is complete in both norms, then the two norms are
equivalent.
Problem 3.4. Let X , Y be Banach spaces. Provisionally, say that a linear transforma-
tion T : X → Y is weakly bounded if f ◦ T ∈ X ∗ whenever f ∈ Y ∗ . Prove, if T is weakly
bounded, then T is bounded.
48

Problem 3.5. Let X , Y be Banach spaces. Suppose (Tn ) is a sequence in B(X , Y) and
limn Tn x exists for every x ∈ X . Prove, if T is defined by T x = limn Tn x, then T is
bounded.

Problem 3.6. Suppose that X is a vector space with a countably infinite basis. (That
is, there is a linearly independent set {xn } ⊆ X such that every vector x ∈ X is
expressed uniquely as a finite linear combination of the xn ’s.) Prove there is no norm
on X under which it is complete. (Hint: consider the finite-dimensional subspaces
Xn := span{x1 , . . . xn }.)

Problem 3.7. The Baire Category Theorem can be used to prove the existence of
(very many!) continuous, nowhere differentiable functions on [0, 1]. To see this, let En
denote the set of all functions f ∈ C[0, 1] for which there exists x0 ∈ [0, 1] (which may
depend on f ) such that |f (x) − f (x0 )| ≤ n|x − x0 | for all x ∈ [0, 1]. Prove the sets En
are nowhere dense in C[0, 1]; the Baire Category Theorem then shows that the set of
nowhere differentiable functions is second category. (To see that En is nowhere dense,
approximate an arbitrary continuous function f uniformly by piecewise linear functions
g, whose pieces have slopes greater than 2n in absolute value. Any function sufficiently
close to such a g will not lie in En .)

Problem 3.8. Let L2 ([0, 1]) denote the Lebesgue measurable functions f : [0, 1] → C
such that |f |2 is in L1 ([0, 1]). It turns out, as we will see later, that L2 ([0, 1]) is a linear
manifold (subspace of the vector space L1 ([0, 1])), though this fact is not needed for this
problem.
Let gn : [0, 1] → R denote the function which takes the value n on [0, n13 ] and 0
elsewhere. Show,
R
(i) if f ∈ L2 ([0, 1]), then limn→∞ gn f dm R= 0;
(ii) Ln : L1 ([0, 1]) → C defined by Ln (f ) = gn f dm is bounded, and ∥Lg ∥ = n;
(iii) conclude L2 ([0, 1]) is of the first category in L1 ([0, 1]).

Problem 3.9. A Banach space of functions on a set X is a vector subspace B of the

space of complex-valued functions on X with a norm ∥ ·∥ making B a Banach space such
that, for each x ∈ X, the mapping Ex : B → C defined by Ex (f ) = f (x) is continuous
(bounded) and if f (x) = 0 for all x ∈ X, then f = 0.
Suppose g : X → C. Show, if gf ∈ B for each f ∈ B, then the linear map
Mg : B → B defined by Mg f = gf is bounded.

Problem 3.10. Suppose X is a Banach space and M and N are closed subspaces.
Show, if for each x ∈ X there exist unique m ∈ M and n ∈ N such that
x = m + n,
then the mapping P : X → M defined by P x = m is bounded.
 49

Problem 3.11. Let X be a Banach space and M ⊆ X a closed subspace. A linear

transformation P : X → M is called a bounded projection if it is bounded and P (m) = m
for all m ∈ M. Prove that if M is a closed subspace and there exists a bounded
projection P : X → M, then there exists a closed subspace N ⊆ X such that M ∩ N =
{0} and X = M + N . Show also that in this case there exists a bounded projection
Q : X → N.

Remark: Given a closed subspace M ⊆ X , we say M is (topologically) comple-

mented if there exists a closed subspace N ⊆ X such that M∩N = {0} and M+N = X .
Taken together, the last two problems show that a closed subspace M ⊆ X is comple-
mented if and only if there is a bounded projection P : X → M. Not every subspace of
a Banach space is necessarily complemented, for example c0 is not complemented in ℓ∞ ,
though this is nontrivial to prove.
Problem 3.12. Here, for definiteness we take the scalar field R.
Let T : ℓ∞ → ℓ∞ denote the backward shift operator defined by T f (n) = f (n + 1)
A bounded linear functional λ : ℓ∞ → R satisfying,
(i) if f ∈ ℓ∞ and (f (n)) converges, then λ(f ) = limn→∞ f (n); and
(ii) λ(T f ) = λ(f )
is a Banach Limit.
Prove
(a) Banach limits exist.
(b) If λ is a Banach limit and f ∈ ℓ∞ , then
lim inf f (n) ≤ λ(f ) ≤ lim sup f (n).
A sequence f ∈ ℓ∞ for which (f (n)) does not converge, but λ(f ) = µ(f ) for all
Banach limits λ and µ is almost convergent. Show that g defined by g(n) = (−1)n is
almost convergent. (Suggestion: given a Banach limit λ, consider λ(g + T g).
Problem 3.13. Prove that Q is not a Gδ set.

4. Lp spaces

Throughout this section, (X, M , µ) is a measure space and X ̸= ∅.

Definition 4.1. For 0 < p < ∞, let Lp (µ) denote the space of measurable functions
f : X → F that satisfy
Z 1/p
p
∥f ∥p := |f | dµ < ∞.
X

Lemma 4.2. Suppose 0 < p < ∞. If f, g ∈ Lp (µ) and c ∈ F, then ∥f + g∥pp ≤

2p ∥f ∥pp + ∥g∥pp and ∥cf ∥p = |c| ∥f ∥p . Hence Lp (µ) is a vector space.
50

Later we will show, for 1 ≤ p ≤ ∞, that ∥ · ∥p is a semi-norm on Lp (µ).

Sketch of proof. The equality ∥cf ∥p = |c| ∥f ∥p is immediate. As a pointwise inequality,

|f + g| ≤ |f | + |g| ≤ 2 max{|f |, |g|}. Hence,

|f + g|p ≤ 2p max{|f |, |g|}p = 2p max{|f |p , |g|p } ≤ 2p (|f |p + |g|p ),

from which the rest of the result follows. □

END Wednesday 2025-02-05

Definition 4.3. A measurable function f : X → F is essentially bounded if there is a

t > 0 such that
µ({|f | > t}) = 0
and let L∞ (µ) denote the set of essentially bounded functions on (X, M , µ).
Define ∥ · ∥∞ : L∞ (µ) → [0, ∞) by

(15) ∥f ∥∞ = inf{t > 0 : µ({|f | > t}) = 0}.

Proposition 4.4. The set L∞ (µ) is a vector space. Further, the infimum in equa-
tion (15) is attained and ∥ · ∥∞ is a semi-norm on L∞ (µ).

It is customary to write Lp instead of Lp (µ) when the µ is understood (or generic).

Proof. It is evident that, if c ∈ F and f ∈ L∞ then cf ∈ L∞ and ∥cf ∥∞ = |c| ∥f ∥∞ .

Now suppose f, g ∈ L∞ . Let s, t > 0 be given such that s > ∥f ∥∞ and t > ∥g∥∞ . By
definition, the (measurable) sets A = {|f | > t} and B = {|g| > s} have measure 0.
Let C = {|f + g| > s + t}. Hence A ∪ B has measure 0 and by the triangle inequality
Ac ∩ B c ⊆ C c . Thus C ⊆ A ∪ B and hence C has measure 0. Thus f + g ∈ L∞ (µ) and
∥f + g∥ ≤ s + t. It now follows that ∥f + g∥ ≤ ∥f ∥∞ + ∥g∥∞ and hence L∞ is a vector
space and ∥ · ∥∞ is a semi-norm on L∞ .
That the infimum is attained in equation (15) is left as an (easy) exercise based
upon the fact that a countable union of sets of measure zero has measure zero. □

We record the following simple observation for later use - often without comment.

Lemma 4.5. If 0 < p ≤ ∞ and f ∈ Lp (µ), then ∥f ∥p = 0 if and only if f = 0 almost

everyewhere.
R
Proof. For 0 < p < ∞, by assumption
R g = |f |p is unsigned g = ∥f ∥pp . Since g = 0
almost everywhere if and only if g = 0, the result follows. □
 51

4.1. Conjugate indices and the inequalities of Young, Holder and Minkowski.
We now restrict our attention to 1 ≤ p ≤ ∞.
Definition 4.6. The conjugate index or dual exponent to 1 < p < ∞ is the unique
1 < q < ∞ satisfying
1 1
+ = 1.
p q
The dual index to p = ∞ is q = 1; and the dual index to p = 1 is q = ∞. □

Note that (p − 1)q = p and likewise (q − 1)p = q. The significance of dual indices is
apparent in the following result.
Lemma 4.7 (Young’s inequality). If a, b are nonnegative numbers and 1 < p, q < ∞
are dual indices, then
ap bq
ab ≤ +
p q
q p
and equality holds if and only if b = a .

Proof. If a or b is 0 there is nothing to prove. So suppose a, b > 0. Define ψ : R → R

by ψ(t) = ap(1−t) bqt . A bit of rearranging gives ψ(t) = ap · exp(ct), where c = log(bq /ap ).
The function ψ is infinitely differentiable and
ψ ′′ (t) = c2 ψ(t) > 0.
Thus ψ is convex. In particular, (using the fact that p1 + 1q = 1)
ap bq
   
1 1 1 1 1
(16) ψ =ψ · 0 + · 1 ≤ ψ(0) + ψ(1) = + .
q p q p q p q
For the case of equality, note that ψ(t) is strictly convex unless c = 0 (ap = bq ), in which
case ψ is constant. □

For an alternate geometric proof of Lemma 4.7, see Problem 2.9.

Theorem 4.8 (Hölder’s inequality). Suppose 1 ≤ p ≤ ∞ and q is the conjugate index
to p. If f ∈ Lp and g ∈ Lq , then f g ∈ L1 , and
(17) ∥f g∥1 ≤ ∥f ∥p ∥g∥q .

Further, assuming 1 ≤ p < ∞ and f ∈ Lp (µ), if ∥f ∥p ̸= 0, then there exists a

g ∈ Lq (µ) such that
(i) ∥g∥q = 1;
(ii) f g ≥ 0; and
(iii)
Z
(18) ∥f g∥1 = f g = ∥f ∥p ∥g∥q = ∥f ∥p .
52

If µ has the property that every set of positive measure contains a set of positive,
but finite, measure and f ∈ L∞ (µ), then

∥f ∥∞ = sup{∥f g∥1 : g ∈ L1 (µ), ∥g∥1 = 1}.

Remark 4.9. For f ∈ L1 (µ) of course equality holds in equation (17) with g = 1.
The assumption that (X, M , µ) has the property that every (measurable) set of
positive measure contains a set of finite measure is needed as the following example
shows. For the measure space ({0}, {∅, {0}}, µ), where µ(∅) = 0 and µ({0}) = ∞, we
have L1 (µ) = {0} and thus ∥f g∥1 = 0 for all f ∈ L∞ (µ) and g ∈ L1 (µ). □

Proof of Theorem 4.8. The proof is easy in the cases p = ∞ or p = 1. Now suppose
1 < p < ∞.
If ∥f ∥p = 0, then f = 0 a.e. by Lemma 4.5. Hence f g = 0 a.e. and, by another
application of Lemma 4.5, ∥f g∥1 = 0. Thus the inequality of equation (17) holds. By
symmetry, the same is true for g. Hence we may assume ∥f ∥p ̸= 0 ̸= ∥g∥q .
By homogeneity we may assume ∥f ∥p = ∥g∥q = 1. We are to show
Z
|f g| dµ ≤ 1.

Applying Lemma 4.7 gives

1 1
(19) |f (x)g(x)| ≤ |f (x)|p + |g(x)|q .
p q
Integrating (19) with respect to µ and applying the normalizations on p, q, f, g gives the
inequality of equation (17). Further, observe, in the case 1 < p, q < ∞, that equality
holds in Hölders in inequality if and only if equality holds a.e. µ in equation (19) if and
only if |f |p = |g|q a.e. µ by Lemma 4.7.
To prove the further portion of the theorem, suppose 1 < p < ∞ and f ∈ Lp (µ)
satisfies ∥f ∥p = 1. Let g = |f |p f −1 (interpreting g as 0 when f is 0). From |g|q =
|f |(p−1)q = |f |p it follows that g ∈ Lq (µ) and ∥g∥q = 1. Further, f g = |f |p so that
∥f g∥1 = 1 = ∥f ∥p and hence equation (18) holds. (Note that a small tweak to this
argument also handles the case p = 1.)
For the last statement, suppose every subset S of X with µ(S) = ∞ contains a set
T for which 0 < µ(T ) < ∞ and let f ∈ L∞ (µ) be given. Without loss of generality,
∥f ∥∞ = C > 0. Given 0 < ρ < C, the set E = {|f | > ρ} has positive measure. Thus
1
there is a set F ⊆ E such that 0 < µ(F ) < ∞. Let g = µ(F χ , where χF is the
) F
1
characteristic function of F. Observe g ∈ L (µ) and ∥g∥1 = 1. Moreover, |f g| ≥ ρg and
hence ∥f g∥1 ≥ ρ∥g∥1 and the result follows. □

END Friday 2025-02-07

 53

Example 4.10. One can get a more intuitive feel for what Hölder’s inequality says by
examining it in the case of step functions. Let E, F be sets of finite, positive measure
and put f = 1E , g = 1F . Then ∥f g∥1 = µ(E ∩ F ) and
∥f ∥p ∥q∥q = µ(E)1/p µ(F )1/q ,
1
so Hölder’s inequality can be proved easily in this case using the relation p
+ 1q = 1 and
the fact that µ(E ∩ F ) ≤ min(µ(E), µ(F )).

Corollary 4.11 (Minkowski’s inequality). Let (X, M , µ) be a measure space and sup-
pose 1 ≤ p ≤ ∞. If f, g ∈ Lp (µ), then
∥f + g∥p ≤ ∥f ∥p + ∥g∥p .
Hence ∥ · ∥p is a semi-norm on Lp (µ).

Proof. The result has already been established for p = 1 and p = ∞ so suppose 1 <
p < ∞ and let q denote the conjugate index to p. By Lemma 4.2, f + g ∈ Lp (µ). The
result is vacuous if f + g = 0 (almost everywhere); equivalenlty, ∥f + g∥p = 0. Now
suppose ∥f + g∥p ̸= 0. By Theorem 4.8, there is an h ∈ Lq (µ) such that ∥h∥q = 1 and
∥(f + g)h∥1 = ∥f + g∥p . On the other hand, two more applications of Theorem 4.8 and
the fact that ∥ · ∥1 is a semi-norm gives,
∥f + g∥p = ∥(f + g)h∥1 ≤ ∥f h∥1 + ∥gh∥1 ≤ ∥f ∥p ∥h∥1 + ∥g∥p ∥h∥1 = ∥f ∥p + ∥g∥p . □

4.2. The Lebesuge spaces Lp (µ). The proof of the following proposition, based on
Lemma 4.5 is left to the gentle reader.

Proposition 4.12. The set N (ν) = {f ∈ Lp (µ) : ∥f ∥p = 0} is a subspace of Lp (µ) and

the function ∥ · ∥p descends to a norm on the quotient space Lp (µ)/N (µ).

Definition 4.13. The normed vector space (Lp (µ)/N (µ), ∥ · ∥p ) (for 1 ≤ p ≤ ∞) is
denoted Lp (µ) and is known as a Lebesgue space.

Suppose 1 ≤ p ≤ ∞ and q is the conjugate index to p. Fix g ∈ Lq (µ). For f ∈ Lp (µ),

Hölder’s inequality (Theorem 4.8) implies gf ∈ L1 (µ) and moroever ∥f g∥1 ≤ ∥f ∥p ∥g∥q .
Thus, we obtain a bounded linear functional Lg : Lp (µ) → F of norm at most ∥g∥q
defined by
Lg (f ) = gf.
Hence we obtain a bounded map (with norm at most one) Φ : Lq (µ) → Lp (µ)∗ .

Proposition 4.14. For 1 < p ≤ ∞, the mapping Φ : Lq (µ) → Lp (µ)∗ defined by

Φ(g) = Lg is isometric.
When p = 1, if µ is σ-finite, then Φ : L∞ (µ) → L1 (µ)∗ is isometric.
54

Remark 4.15. Returning to the example in Remark 4.9 where L1 (µ) = {0}, the map
Φ : L∞ (µ) → L1 (µ)∗ is not one-one. Since in this case L1 (µ)∗ = {0}, but L∞ (µ) = F
isometrically, Φ is the zero map and, in particular, Φ(1) = 0.
On the other hand, in the case p = 1 it suffices to assume that (X, M , µ) has the
property that every set S such that µ(S) = ∞ contains a subset T with 0 < µ(T ) < ∞.
Later we will see that the map Φ in Proposition 4.14 is an isometric isomorphism
for 1 ≤ p < ∞, with the proviso that µ is σ-finite in the case p = 1.
Problems 2.12 and 4.6 says that Φ need not be onto in the case that p = ∞.

Proof of Proposition 4.14. Let g ∈ Lq (µ) be given. If ∥g∥q = 0, then Lg = 0 so that the
results holds, even when p = 1 without conditions on the measure space (X, M , µ).
Now suppose 1 < p ≤ ∞ and ∥g∥q ̸= 0. As already observed, ∥Ψ(g)∥ = ∥Lg ∥ ≤ ∥g∥q ,
for g ∈ Lq (µ). On the other hand, since 1 ≤ q < ∞, for 0 ̸= g ∈ Lq (µ), the moreover
portion of Hölder’s Inequality (Theorem 4.8) gives a function f ∈ Lp (µ) such that
∥f ∥p = 1 and Z
Lg (f ) = f g = ∥g∥q .
Hence ∥Lg ∥ ≥ ∥g∥q and thus ∥Lg ∥ = ∥g∥q .
In the case p = 1 and µ satisfies the hypothesis of the additional hypotheses given,
Hölder’s inequality implies that, for each 0 ≤ ρ < ∥g∥∞ , there is an f ∈ L1 (µ) such
that ∥f ∥1 = 1 and |Lg (f )| = ∥f g∥1 > ρ. Thus ∥Lg ∥ ≥ ∥g∥∞ and consequently ∥Lg ∥ =
∥g∥∞ . □
Proposition 4.16. For 1 ≤ p ≤ ∞, the normed vector space Lp (µ) is a Banach space.

Proposition 4.16 is a near immediate consequence of the following lemma.

Lemma 4.17. Suppose 1 ≤ p ≤ ∞ and (fn )∞ p
n=1 is a sequence from L (µ). If for each
ϵ > 0 there is an N such that if m, n ≥ N, then ∥fn − fm ∥p < ϵ, then there is an
f ∈ Lp (µ) such that
(a) the sequence (∥fn − f ∥p )n converges to 0;
(b) there is subsequence (fnk ) of (fn ) that converges to f pointwise almost everywhere.

END Monday 2025-02-10

Sketch of proof. The proof for the case 1 ≤ p < ∞ is very much like the case p = 1 that
has already been established and is just sketched here.
There is a subsequence (gk )∞
k=1 of (fn ) such that ∥gk+1 −gk ∥ < 2
−k
for k ≥ 1. Setting
P∞
g0 = 0, the series k=0 ∥gk+1 − gk ∥p converges. (The subsequence (gk ) is super-Cauchy.)
Let m
X
hm = |gk+1 − gk |
k=0
 55

and let h denote the pointwise limit (in [0, ∞]) of the non-negative increasing sequence
(hm ). By the Monotonne Convergence Theorem,
Z Z
h = lim hpm .
p

Thus,
(20) ∥h∥p = lim ∥hm ∥p .s
The inequality,
m
" ∞
#p
X X
∥hm ∥p ≤ ∥gk+1 − gk ∥p ≤ ∥gk+1 − gk ∥p <∞
k=0 k=0
p
and equation (20) implies h ∈ L . Thus
∞
X
h= |gk+1 − gk |
k=0

is finite almost everywhere and hence,

m
X
(gk+1 − gk ) = gm+1
k=0

also converges almost everywhere to a measurable function f. That is, the sequence (gk )
converges pointwise to f. Further, since |f | ≤ h and h ∈ Lp , it follows that f ∈ Lp .
By construction, for m fixed, if n > m, then ∥gn − gm ∥p < 21−m and (|gn − gm |p )n
converges pointwise almost everywhere to |f − gm |p . Thus, by Fatou,
Z Z
p p
∥f − gm ∥p = |f − gm | ≤ lim inf |gn − gm |p = lim inf ∥gn − gm ∥pp < 21−m .

Thus the sequence (∥f − gm ∥p )m converges to 0.

A standard fact that, in a metric space X, if (xn ) is Cauchy and if there is an x ∈ X
and a subsequence (xnk ) of (xn ) that converges to x, then (xn ) converges to x. Thus,
from what has been proved, (fn ) converges to f in Lp and the subsequence (gk ) of (fn )
converges to f pointwise, completing the proof for 1 ≤ p < ∞.
The case p = ∞ follows from the fact that, for g ∈ L∞ (µ), the set {|g| > ∥g∥∞ }
has measure 0 (Proposition 4.4) so that it can be assumed that |g| is bounded by ∥g∥∞ .
From here the proof is very much like the proof of completeness of the space Fb (X, F) of
bounded functions on a set X with the supremum norm. (See Proposition 1.16) In par-
ticular, a Cauchy sequence (fn ) converges pointwise almost everywhere (no subsequence
is needed). The details are left to the reader. □
Corollary 4.18. If (fn ) is a sequence from Lp (µ) (1 ≤ p ≤ ∞) that converges to f
in Lp (µ), then there is a subsequence (fnk ) of (fn ) that converges to f pointwise almost
everywhere.
56

Proof. The sequence (fn ) (viewed as representative of their respective equivalence classes)
satisfies the hypotheses of Lemma 4.17. Hence there is a g ∈ Lp (µ) and subsequence (gk )
of (fn ) that converges to g both in Lp (µ) and pointwise almost everywere. By uniqueness
of limits, g = f as elements of Lp (µ); that is, almost everywhere. Thus (gk ) converge to
f pointwise almost everywhere. □

Corollary 4.19. Suppose 1 ≤ p ≤ ∞ and (fn ) is a sequence from Lp (µ). If (fn ) con-
verges to f in Lp (µ) and to g pointwise a.e., then f = g a.e.; that is, the pointwise limit
and Lp (µ) limit are the same (almost everywhere).

Example 4.20. [The typewriter sequence] Define fn : [0, 1] → R by

fn (x) = χ( n−2k , n+1−2k ] , for 2k ≤ n < 2k+1 ,
2k 2k

(here χ is the indicator function) viewed as functions in L1 (m) for Lebesgue measure m
on [0, 1]. The sequence (fn ) converges to 0 in L1 (m), but does not converge pointwise
anywhere. On the other hand, the subsequence (gk = f2k ) converges to 0 pointwise
(everywhere).

Example 4.21. Let (X, M , µ) denote a measure space and suppose h : X → F is

a measurable function. Given 1 ≤ p, r ≤ ∞, if hf ∈ Lr for each f ∈ Lp , then the
linear mapping Mh : Lp → Lr is bounded. As an example of what more can be said, if
µ(X) < ∞ and p = 2 = r, then h ∈ L∞ .
Use the Closed Graph Theorem as follows. Suppose (fn , Mh fn ) is a sequence that
converges to (f, g) in Lp ×Lr . Apply Corollary 4.18 to (fn ) and f to obtain a subsequence
(gk ) of (fn ) converging to f pointwise a.e. and of course in Lp . Apply Corollary 4.18 to
(hgk ) and g to deduce g = hf. Now use Closed Graph.
For the bit about L2 , make an argument like the one at the end of the proof of
Hölder’s inequality.

4.3. Problems.

Problem 4.1. Suppose f : [0, A] → [0, ∞) is differentiable, strictly increasing and

f (0) = 0. Prove, for each 0 < a ≤ A, that
Z x Z f (x)
f + f −1 = xf (x).
0 0
Rx R f (x)
[Suggestion: Differentiate g(x) = 0
f + 0
f −1 − xf (x).] Deduce Young’s inequality.

Problem 4.2. [Truncation of Lp functions] Suppose f is an unsigned function in Lp (µ),

1 < p < ∞. For t > 0 let
Et = {x : f (x) > t}.
Show:
 57

(a) For each real number t > 0, the horizontal truncation 1Et f belongs to Lq for all
1 ≤ q ≤ p.
(b) For each real number t > 0, the vertical truncation ft := min(f, t) belongs to Lq for
all p ≤ q ≤ ∞.
(c) Every f ∈ Lp , 1 < p < ∞, can be decomposed as f = g + h where g ∈ L1 and
h ∈ L∞ .
Problem 4.3. Suppose f ∈ Lp0 ∩L∞ for some p0 < ∞. Prove f ∈ Lp for all p0 ≤ p ≤ ∞,
and limp→∞ ∥f ∥p = ∥f ∥∞ .
Problem 4.4. Prove fn → f in the L∞ norm if and only if fn → f essentially uniformly,
and that L∞ is complete.
Problem 4.5. Show that Lp (R) ̸⊆ Lq (R) for any pair p, q.
Problem 4.6. Consider L∞ (R).
a) Show that M := C0 (R) is a closed subspace of L∞ (R) (more precisely, that the
set of L∞ functions that are a.e. equal to a C0 function is closed in L∞ ). Prove
there is a bounded linear functional λ : L(R)∞ → F such that λ|M = 0 and
λ(1R ) = 1.
b) Prove there is no function g ∈ L1 (R) such that λ(f ) = R f g dm for all f ∈ L∞ .
R

(Hint: look at the restriction of λ to C0 (R).)

5. Hilbert Space

5.1. Inner product spaces.

Definition 5.1. Let V denote a vector space over C. A function ⟨·, ·⟩ : V × V → C is a
inner product if, for all f, g, h ∈ V and c ∈ C,
(a) ⟨f, f ⟩ ≥ 0;
(b) ⟨f, f ⟩ = 0 if and only if f = 0;
(c) ⟨f + cg, h⟩ = ⟨f, h⟩ + c ⟨g, h⟩;
(d) ⟨g, f ⟩ = ⟨f, g⟩.

END Wednesday 2025-02-12

Proposition 5.2. An inner product on a vector space V satisfies the the Cauchy–
Schwarz inequality,
|⟨f, g⟩|2 ≤ ⟨f, f ⟩ ⟨g, g⟩.
Equality holds if and only if f and g are linearly dependent.
p
The function ∥ · ∥ : V → C defined by ∥f ∥ = ⟨f, f ⟩ is a norm on V and, with this
notation, the CS inequality becomes
|⟨f, g⟩| ≤ ∥f ∥ ∥g∥.
58

Further, ∥f + g∥ = ∥f ∥ + ∥g∥ if and only if either f = 0 or there is a t ≥ 0 such

that g = tf.
Remark 5.3. Given an inner product space V = (V, ⟨·, ·⟩), we endow it with the norm
- and hence metric - arising from the inner product.
Lemma 5.4 (Joint continuity of the inner product). Let H be an inner product space
equipped with its norm topology. If (xn ) converges to x and (yn ) converges to y in H,
then (⟨xn , yn ⟩) converges to ⟨x, y⟩.

Proof. By Cauchy-Schwarz,
|⟨xn , yn ⟩ − ⟨x, y⟩| ≤ |⟨xn , yn − y⟩| + |⟨xn − x, y⟩| ≤ ∥xn ∥∥yn − y∥ + ∥xn − x∥∥y∥ → 0,
since ∥xn − x∥, ∥yn − y∥ → 0 and the sequence ∥xn ∥ is bounded. □
Definition 5.5. A Hilbert space H over Fpis an inner product space over F that is
complete in the metric d(x, y) = ∥x − y∥ = ⟨x − y, x − y⟩. (Here, as usual, F is either
C or R.)
We continue to use the notation M ≤ H to indicate M is a (closed) subspace of H
from Definition 1.26.
Example 5.6 (Fn ). It is easy to check that the standard scalar product on Rn is an
inner product; it is defined as usual by
n
X
(21) ⟨x, y⟩ = xj y j
j=1

where we have written x = (x1 , . . . xn ); y = (y1 , . . . yn ). Similarly, the standard inner

product of vectors z = (z1 , . . . zn ), w = (w1 , . . . wn ) in Cn is given by
n
X
(22) ⟨z, w⟩ = zj wj .
j=1

(Note that it is necessary to take complex conjugates of the w’s to obtain positive
definiteness.)
It is straightforward to check that equations (21) and (22) define inner products on
R and Cn respectively that induce the Euclidean norm. Since these Euclidean spaces
n

are complete, they are Hilbert spaces.

Example 5.7 (L2 (µ)). Let (X, M , µ) be a measure space. Given f, g ∈ L2 (µ), by
Hölder’s inequality (Theorem 4.8), the function f g ∈ L1 (µ) and ∥f g∥ ≤ ∥f ∥2 ∥g∥2 .
From here it is a simple exercise to verify that the Banach space L2 (µ) is the inner
product space with the inner product,
Z
(23) ⟨f, g⟩ = f g dµ.
X
 59

That is, equation (23) is an inner product and the norm on L2 (µ) is the norm derived
from this inner product.

Example 5.8 (ℓ2 (N)). Let

∞
X
2
ℓ (N) = {(a1 , a2 , . . . an , . . . ) | an ∈ F, |an |2 < ∞}.
j=1

This space is L2 (c) for the measure space (N, P (N), c), where c is counting measure. In
particular, ℓ2 (N) is a Hilbert space with the inner product,
∞
X
(24) an b n
n=1

for sequences a = (a1 , a2 , . . . ) and b = (b1 , b2 , . . . ) in ℓ2 .

Note too, that example 5.6, is the special case of L2 (ν) for ν equal to counting
measure on P ({1, 2, . . . , n}).

5.2. Orthogonality. In this section we show that many of the basic features of the
Euclidean geometry of Fn extend naturally to the setting of an inner product space.

Definition 5.9. Let H be an inner product space.

(i) Two vectors x, y ∈ H are orthogonal if ⟨x, y⟩ = 0, written x ⊥ y.
(ii) Two subsets A, B of H are orthogonal if x ⊥ y for all x ∈ A and y ∈ B, written
A ⊥ B.
(iii) A subset A of H is orthogonal if x ⊥ y for each x, y ∈ A with x ̸= y and is
orthonormal if also ⟨x, x⟩ = 1 for all x ∈ A.
(iv) The orthogonal complement of a subset E of H is

E ⊥ = {x ∈ H : ⟨x, e⟩ = 0 for all e ∈ E}.

The proof of the following lemma is an easy exercise. Indeed, the first item follows
immediately from Lemma 5.4 and the second from the positive definiteness of a norm.

Lemma 5.10. If E is a subset of an inner product space H, then

(i) E ⊥ is a closed subspace of H;
(ii) E ∩ E ⊥ ⊆ {0}; and
(iii) E ⊆ (E ⊥ )⊥ = E ⊥⊥ .

Theorem 5.11 (The Pythagorean Theorem). If H is an inner product space and

f1 , . . . fn are mutually orthogonal vectors in H, then

∥f1 + · · · + fn ∥2 = ∥f1 ∥2 + · · · + ∥fn ∥2 .

60

Proof. When n = 2, we have

∥f1 + f2 ∥2 = ∥f1 ∥2 + ⟨f1 , f2 ⟩ + ⟨f2 , f1 ⟩ + ∥f2 ∥2
= ∥f1 ∥2 + ∥f2 ∥2 .
The general case follows by induction. □

Suppose V is a vector space over F. A function [·, ·] : V × V → F satisfying the

axioms of items c and d is bilinear form in the case F = R and a sesquilinear form when
F = C. If it also satisfies item a, then it is positive semi-defininite.
Theorem 5.12 (The Parallelogram Law). If [·, ·] is a bilinear (resp. sesquilinear) form
on a vector space over R (resp. C) and f, h ∈ V, then
(25) [f + g, f + g] + [f − g, f − g] = 2 ([f, f ] + [g, g]) .
In particular, if H is an inner product space, then
(26) ∥f + g∥2 + ∥f − g∥2 = 2(∥f ∥2 + ∥g∥2 ).

Proof. By linearity (resp. sesquilinearity),

(27) [f ± g, f ± g] = [f, f ] ± [f, g] ± [g, f ] ± [g, g]
Adding these two equations together gives the identity of equation (25).
In the case of Hilbert space, equation (26) follows from equation (25) by the defini-
tion of the norm coming from the inner product. □

Subtracting, instead of adding, in the proof of the Parallelogram Law gives the
polarization identity
2 ([f, g] + [g, f ]) = [f + g, f + g] − [f − g, f − g].
Theorem 5.13 (The Polarization identity). If [·, ·] is a bilinear form on a vector space
over Rand f, h ∈ V, then
(28) 4[f, g] = [f + g, f + g] − [f − g, f − g].
In particular, if H is an inner product space over R, then
1
∥f + g∥2 − ∥f − g∥2 .


(29) ⟨f, g⟩ =
4
If [·, ·] is a sesquilinear form on a vector space over Cand f, h ∈ V, then
3
X
(30) 4[f, g] = ik [f + ik g, f + ik g].
k=0
If H is a complex Hilbert space, then
3
X
(31) 4⟨f, g⟩ = ik ⟨f + ik g, f + ik g⟩.
k=0
 61

Remark: Note that, in a Hilbert space, the polarization identity says that the inner
product is determined by the norm.
An elementary (but tricky) theorem of von Neumann says, in the real case, that if
H is any vector space equipped with a norm ∥ · ∥ such that the parallelogram law (26)
holds for all f, g ∈ H, then H is an inner product space with inner product given by
formula (29) in the case of real scalars and formula (31) in the case of complex scalars.
(The proof is simply to define the inner product by equation (29) or (31), and check
that it is indeed an inner product.)

5.3. Best approximation in Hilbert space.

Definition 5.14. A subset K of a vector space V is convex if whenever a, b ∈ K and
0 ≤ s, t sum to 1, it follows that sa + tb ∈ K as well. (Geometrically, this means that
when a, b lie in K, so does the line segment joining them.)
A normed vector space X is strictly convex if x, y ∈ X and ∥x + y∥ = ∥x∥ + ∥y∥,
then either x = 0 or there is a t ≥ 0 such that y = tx.
Example 5.15. Subspaces and balls (B(x, r)) in a normed vector spaces are convex.
The closure and interior of a convex set are convex.
Remark 5.16. Hilbert spaces are strictly convex. The Lebesgue spaces Lp are convex
for 1 < p < ∞, but not for p = 1, ∞.
That a normed vector space is strictly convex if and only if x ̸= y and ∥x∥ = ∥y∥ = 1,
then ∥ 12 (x + y)∥ < 1 offers an explanation for the terminology.

END Friday 2025-02-14 (though we had not finished with the remark immediately
above)
Proposition 5.17. Suppose X is a strictly convex normed vector space. If K ⊆ X is
convex, h ∈ X and there exists a y, z ∈ K such that
∥h − y∥ = dist(h, K) = inf{∥h − k∥ : k ∈ K} = ∥h − z∥,
then z = y.

Proof. Let d = dist(h, K). By convexity, k = y+z 2

∈ K and by the triangle inequality,
1 1 1
d ≤ ∥h − k∥ = ∥ (h − y) + (h − z)∥ ≤ [∥h − y∥ + ∥h − z∥] = d.
2 2 2
Hence equality holds in the triangle inequality. Without loss of generality, h−y ̸= 0 and,
by strict convexity, there is a t ≥ 0 such that h−y = t(h−z). Since ∥h−y∥ = d = ∥h−z∥,
it follows that t = 1 and therefore y = z. □
Theorem 5.18. Suppose H is a Hilbert space. If ∅ ̸= K ⊆ H is a closed, convex,
nonempty set, and h ∈ H, then there exists a unique vector k0 ∈ K such that
∥h − k0 ∥ = dist(h, K) := inf{∥h − k∥ : k ∈ K}.
62

Proof. Uniqueness follows from Propositions 5.2 and 5.17.

Let d = dist(h, K) = inf k∈K ∥h − k∥. First observe, if x, y ∈ K, then, by convexity,
so is v = x+y
2
and in particular, ∥h − v∥2 ≥ d2 . Hence, by the parallelogram law, applied
to f = x−h
2
and g = y−h
2
,
2 2
x−y 1 x+y
= ∥x − h∥2 + ∥y − h∥2 −


−h
(32) 2 2 2
1
≤ ∥x − h∥2 + ∥y − h∥2 − d2 .


2

There exists a sequence (kn ) in K so that (∥kn − h∥) converges to d. Given ϵ > 0
choose N such that for all n ≥ N , ∥kn − h∥2 < d2 + ϵ2 . By (32), if m, n ≥ N then
2
km − kn 1 1
< (2d2 + ϵ2 ) − d2 = ϵ2 .
2 2 2

Consequently ∥km − kn ∥ < ϵ for m, n ≥ N and (kn ) is a Cauchy sequence. Since H is

complete, (kn ) converges to a limit k ∈ H, and since K is closed, k ∈ K. Since (kn − h)
converges to (k − h) and ∥kn − h∥ converges to d it follows, by continuity of the norm,
that ∥k − h∥ = d. □

The most important application of the preceding approximation theorem is in the

case when K = M is a (closed) subspace of the Hilbert space H. What is significant
is that in the case of a subspace, the minimizer k has an elegant geometric description,
namely, it is obtained by “dropping a perpendicular” from h to M . This geometric
interpretation is the content of the next theorem, whose statement uses Theorem 5.18.
Recall M ≤ H to means that M is a (closed) subspace of H.

Theorem 5.19. Suppose H is a Hilbert space, M ≤ H, and h ∈ H. If f0 is the unique

element of M such that ∥h−f0 ∥ = dist(h, M ), then (h−f0 ) ⊥ M . Conversely, if f0 ∈ M
and (h − f0 ) ⊥ M , then ∥h − f0 ∥ = dist(h, M ).

Proof. Let f0 ∈ M with ∥h − f0 ∥ = dist(h, M ) be given. Given f ∈ M , for t ∈ R, let

λ = t⟨h − f0 , f ⟩. Since f0 + λf ∈ M ,

0 ≤ ∥h − (f0 + λf )∥2 − ∥h − f0 ∥2
=∥(h − f0 ) + λf ∥2 − ∥h − f0 ∥2
= − 2 real λ ⟨h − f0 , f ⟩ + |λ|2 ∥f ∥2
2
= −2t + t2 ∥f ∥2 ⟨h − f0 , f ⟩
 

for all t. Thus |⟨h − f0 , f ⟩| = 0.

 63

Conversely, suppose f0 ∈ M and (h − f0 ) ⊥ M . In particular, we have (h − f0 ) ⊥

(f0 − f ) for all f ∈ M . Therefore, by Theorem 5.11, for all f ∈ M
∥h − f ∥2 = ∥(h − f0 ) + (f0 − f )∥2
= ∥h − f0 ∥2 + ∥f0 − f ∥2 ≥ ∥h − f0 ∥2 .
Thus ∥h − f0 ∥ = dist(h, M ). □
Corollary 5.20. If H is a Hilbert space and M ≤ H, then (M ⊥ )⊥ = M .

Proof. By Lemma 5.10, M ⊆ (M ⊥ )⊥ . Now suppose that x ∈ (M ⊥ )⊥ . By Theorem 5.19

applied to x and M , there exists m ∈ M such that x − m ∈ M ⊥ . On the other hand,
both x and m are in (M ⊥ )⊥ and thus by Lemma 5.10, x − m ∈ (M ⊥ )⊥ . Hence x − m = 0
by Lemma 5.10, and x ∈ M . □

If E is a subset of the Banach space X, and E is the collection of all closed subspaces
N of X such that E ⊆ N , then
M = ∩N ∈E N
is the smallest closed subspace containing E.
Corollary 5.21. If E is a subset of H, then (E ⊥ )⊥ is equal to the smallest closed
subspace of H containing E. In particular, if E is a linear manifold (vector subspace)
in H, then E = (E ⊥ )⊥ .

Proof. The proof uses Lemma 5.10 freely. In particular, E ⊆ (E ⊥ )⊥ and (E ⊥ )⊥ is a

closed subspace. If M is a closed subspace containing E, then E ⊥ ⊇ M ⊥ and hence
(E ⊥ )⊥ ⊆ (M ⊥ )⊥ = M by Corollary 5.20.
For the last statement, from the fact that E and E ⊥⊥ are both the smallest closed
subspace containing the linear maniforld E. See Corollary 5.20. □

END Monday 2025-02-17

Corollary 5.22. A vector subspace E of a Hilbert space H is dense in H if and only if
E ⊥ = {0}.
Proposition 5.23. Suppose M, N ≤ H. If M and N are orthogonal, then M + N is
closed. In particular, M + N is again a subspace of H.

Proof. It suffices to prove that M + N is complete. Accordingly suppose (mk + nk ) is a

Cauchy sequence from M + N . From orthogonality, for k, ℓ ∈ N,
∥mk − mℓ ∥2 + ∥nk − nℓ ∥2 = ∥(mk + nk ) − (mℓ + nℓ )∥2
and hence (mk ) and (nk ) are both Cauchy. Since H is complete and M, N are closed,
M and N are each complete. Thus (mk ) converges to some m ∈ M and (nk ) converges
to some n ∈ N and thus (mk + nk ) converges to m + n ∈ M + N . □
64

Definition 5.24. Given subspaces M, N ≤ H of a Hilbert space H, the notation M ⊕N

is used for M + N in the case M and N are closed subspaces and M ⊥ N and is called
the orthogonal direct sum. Hence, M ⊕ N indicates that M, N are orthogonal closed
subspaces of H.

The following corollary should be compared with Problem 3.10.

Corollary 5.25. If M ≤ H, then H = M ⊕ M ⊥ .

Proof. Given x ∈ H, there exists m ∈ M such that x − m ∈ M ⊥ by Theorem 5.19.

Hence x = m + (x − m) ∈ M ⊕ M ⊥ . □
Example 5.26. In a Banach space, a best approximation to a subspace need not exist as
the following example illustrates. Consider the real Banach space C([0, 1]), the subspace
R
U = ker λ1 ∩ker λ2 where λj are the linear functionals on C([0, 1]) defined by λ1 (f ) = f
and λ2 (f ) = f (1). Since these linear functionals are bounded with norm 1, the linear
manifold U is closedR (so a subspace). Let f = 1 − x and observe, for g ∈ U, that
(f − g)(1) = 0 and (f − g) = 12 . Thus the average of f − g is 12 but (f − g)(1) < 21 .
Consequently, there is a point p ∈ [0, 1] such that (f − g)(p) > 12 and we conclude that
there does not exists a g ∈ U such that ∥g − f ∥ = 12 .
1
−ϵ+2ϵ2
Given 0 < ϵ < 21 , choose 0 < δ = 2 1+2ϵ < 1 Let γ = 12 − ϵ − δ and let g = gϵ denote
the piecewise linear function that takes values 12 − ϵ − x for 0 ≤ x ≤ δ, and then connects
the points (δ, γ) to (1, 0) (draw the picture). By construction, g ∈ C([0, 1]) and g(1) = 0.
Further, δ was chosen to insure that g = 0. Thus g ∈ U and ∥f − g∥∞ = 21 + ϵ. Hence
R

dist(f, U ) = 12 but there does not exist a g ∈ U where this distance is achieved.
Example 5.27. This example show that in a Banach space, there can be more than
one closest point from a point to a subsapce.
Consider the real Banach space (R2 , ∥ · ∥∞ ) (thus ∥(x1 , x2 )∥∞ = max{|x1 |, |x2 |}. Let
M = {(x1 , 0) : x1 ∈ R} ⊆ R2 and note M is a subspace of R2 . Let y = (0, 1) and observe
dist(y, M ) = 1 and this distance is attained for each (x, 0) ∈ M with |x| ≤ 1.

END Wednesday 2025-02-19 - we also discussed Corollary 4.19 and example 4.20.

5.4. The Riesz Representation Theorem and Hilbert space adjoint operators.
In this section we investigate the dual H ∗ of a Hilbert space H. One way to construct
bounded linear functionals on Hilbert space is as follows. Given a vector g ∈ H define,
Lg (h) = ⟨h, g⟩.
Indeed, linearity of L is just the linearity of the inner product in the first entry, and the
boundedness of L follows from the Cauchy-Schwarz inequality,
|Lg (h)| = |⟨h, g⟩| ≤ ∥g∥∥h∥.
 65

So ∥Lg ∥ ≤ ∥g∥, but in fact it is easy to see that ∥Lg ∥ = ∥g∥; just apply Lg to the unit
vector g/∥g∥ (assuming g ̸= 0). Hence, L : H → H ∗ defined by g 7→ Lg is a conjugate
linear isometry (thus linear in the case of real scalars).
In fact, it is clear from linear algebra that every linear functional on Fn takes the
form Lg . More generally, every bounded linear functional on a Hilbert space has the form
just described.
Theorem 5.28 (The Riesz RepresentationTheorem). If H is a Hilbert space and λ :
H → F is a bounded linear functional, then there exists a unique vector g ∈ H such that
λ = Lg . Thus the conjugate linear mapping L is isometric and onto.

Proof. It has already been established that L is isometric and in particular one-one.
Thus it only remains to show L is onto. Accordingly, let λ ∈ H ∗ be given. If λ = 0, then
λ = L0 . So, assume λ ̸= 0. Since λ is continuous, by Proposition 1.32 ker λ = λ−1 ({0})
is a proper, closed subspace of H. Thus, by Theorem 5.19 (or Corollary 5.25) there exists
a nonzero vector f ∈ (ker λ)⊥ and by rescaling we may assume λ(f ) = 1.
Given h ∈ H, observe
λ(h − λ(h)f ) = λ(h) − λ(h)λ(f ) = 0.
Thus h − λ(h)f ∈ ker λ and consequently,
0 = ⟨h − λ(h)f, f ⟩
= ⟨h, f ⟩ − λ(h)⟨f, f ⟩.
f
Thus λ = Lg , where g = ∥f ∥2
and the proof is complete. □

5.4.1. Duality for Hilbert space. In the case F = R the Riesz representation theorem
identifies H ∗ with H. In the case F = C, the mapping sending λ ∈ H ∗ to the vector
h0 is conjugate linear and thus H ∗ is not exactly H (under this map). However, it is
customary when working in complex Hilbert space not to make this distinction. This
convention creates some conflicts that must kept in mind. For instance, given Banach
spaces X and Y and a bounded linear map T : X → Y, the adjoint of T, denoted T ∗
is the uniquely determined (by Hahn-Banach) linear map T ∗ : Y ∗ → X ∗ defined by
T f = f ◦ T so that T f (x) = f (T (x)) for x ∈ X . See Theorem 2.32 Because of our
conjugate linear identification of H with H ∗ , the notion of the adjoint of a operator
in the context of Hilbert space differs from the of operators between Banach spaces as
described in the following proposition.
Proposition 5.29. If H, K are Hilbert spaces and T : H → K is a bounded operator,
then there is a unique bounded operator S : K → H satifying,
⟨T h, k⟩ = ⟨h, Sk⟩.
Moreover, ∥S∥ = ∥T ∥.
66

Definition 5.30. The operator S associated to T in Proposition 5.29 is the (Hilbert

space) adjoint of T, denoted T ∗ (sic). Note that Proposition 5.29 says ∥T ∗ ∥ = ∥T ∥. It
is easy to verify that (T ∗ )∗ = T and if c ∈ F, then (cT )∗ = cT ∗ .

The following elementary lemma will be used in the proof of Proposition 5.29 and
elsewhere without comment.
Lemma 5.31. Suppose H, K are Hilby spaces and Y : H → K. If ⟨T h, k⟩ = 0 for all
h ∈ H and k ∈ K, then Y = 0

Proof. Given h choose k = T h and use positive definitiness of the inner product. □
Lemma 5.32. Suppose H and K are Hilbert spaces. A linear mapping T : H → K is
bounded if and only if there is a C such that |⟨T h, k⟩| ≤ C∥h∥ ∥k∥ for all h ∈ H and
k ∈ K. Moreover, in this case ∥T ∥ is the smallest such C.

Proof sketch. Suppose C exists. Given h ∈ H and choosing k = T h gives,

∥T h∥2 = |⟨T h, T h⟩| ≤ C∥h∥ ∥T h∥.
If T h ̸= 0, it follows that ∥T h∥ ≤ C∥h∥. Hence T is bounded and ∥T ∥ ≤ C.
Assuming T is bounded, it is immediate from the Cauchy-Schwarz inequality that
|⟨T h, k⟩| ≤ ∥T ∥ ∥h∥ ∥k∥.
Assuming T is bounded, the argument in the first paragraph shows ∥T ∥ is the
smallest possible C. □

Proof of Proposition 5.29. Define S : K → H as follows. Given k ∈ K, observe that the

mapping λ : H → C defined by λ(f ) = ⟨T f, k⟩ is (linear and) continuous. Hence, by the
Reisz Representation Theorem (for Hilbert space), 5.28, there is a vector Sk such that
⟨T f, h⟩ = λ(f ) = ⟨f, Sk⟩.
It is an exercise to verify that S is linear.
Conversely, if S ′ : K → H is linear and
⟨T f, k⟩ = ⟨f, S ′ k⟩
for all f, ∈ H and k ∈ K, then ⟨(S − S ′ )k, f ⟩ = 0 for all f ∈ H and k ∈ K and hence
S ′ = S.
Finally to prove that S is bounded and ∥S∥ = ∥T ∥, observe, given h ∈ H and k ∈ K
that
|⟨h, Sk⟩| = |⟨T h, k⟩| ≤ ∥T ∥∥h∥ ∥k∥.
Hence, by Lemma 5.32, S is bounded and ∥S∥ ≤ ∥T ∥. By symmetry, it follows that
∥T ∥ ≤ ∥S∥. Hence equality holds. □

Further properties of the adjoints on Hilbert space appear in Problem 5.2.

A bounded operator T on a Hilbert space H is self-adjoint or hermitian if T ∗ = T .
 67

Proposition 5.33. If T is a bounded self-adjoint operator on a Hilbert space H, then

[·, ·] : H × H → F defined by [f, g] = ⟨T f, g⟩ is a bilinear/sesquilinear form on H. If, in
addition, ⟨T h, h⟩ = 0 for all h ∈ H, then T = 0.

Proof. Define [·, ·] : H × H → F by

[f, g] = ⟨T f, g⟩.
Since T is self-adjoint, [g, f ] = ⟨T g, f ⟩ = ⟨g, T f ⟩ = ⟨T f, g⟩ = [f, g], from which it
follows that [·, ·] is a bilinear/sesquilinear form on H. Hence, by the polarization identity
(Theorem 5.13),
X3
4⟨T f, g⟩ = ik ⟨T (f + ik g), f + ik g⟩ = 0,
k=0
for all f, g ∈ H. By Lemma 5.31, T = 0. □

5.4.2. Projections. Returning to Theorem 5.19, if M ≤ H and h ∈ H, there exists

a unique f0 ∈ M such that (h − f0 ) ⊥ M . We thus obtain a well-defined function
P : H → H (or, we could write P : H → M ) defined by
(33) P h = f0 .
That is, P h is characterized by P h ∈ M and (h − P h) ∈ M ⊥. If the space M needs to
be emphasized we will write PM for P .
Definition 5.34. A bounded operator Q on a Hilbert space H (meaning Q : H → H is
linear and bounded) is a projection if Q∗ = Q and Q2 = Q. □

The following Theorem says if Q is a projection, then Q = PN , where N is the

range of Q; that is, Q is uniquely determined by its range, justifying the use of the in
Definition 5.34; and conversely, if M ≤ H, then PM is a projection (onto M ).
Theorem 5.35. Suppose M ≤ H. The mapping P = PM is a projection with range M .
Moreover, if Q is a projection with range N , then
(i) if h ∈ N , then Qh = h;
(ii) ∥Qh∥ ≤ ∥h∥ for all h ∈ H;
(iii) N ≤ H;
(iv) N ⊥ is the kernel of Q;
(v) I − Q is a projection with range N ⊥ ; and
(vi) Q = PN .
Definition 5.36. For M ≤ H and Q the operator PM is called the orthogonal projection
of H on M and, for h ∈ H, the vector PM h is the orthogonal projection of h onto M .

Proof. In view of Corollary 5.25, M ⊕ M ⊥ = H, from which it follows readily that P is

a linear map.
68

Evidently P maps into M and if f ∈ M, then P f = f and hence P maps onto M

and P P f = P f (and so P 2 = P ).
If h ∈ H, then h = P h + (h − P h). But (h − P h) ∈ M ⊥ and P h ∈ M , and thus, by
the Pythagorean Theorem

∥h∥2 = ∥h − P h∥2 + ∥P h∥2 .

Hence ∥P h∥ ≤ ∥h∥. In particular, P is a bounded operator on H. (See also Problem

3.10.)
Given g, f ∈ H, since g − P g is orthogonal to M and P f is in M ,
⟨P f, P g⟩ =⟨P f, P g⟨+⟨P f, (g − P g)⟩
=⟨P f, g⟩ = ⟨f, P ∗ g⟩.

On the other hand, by the same reasoning

⟨P f, P g⟩ =⟨P f + (I − P )f, P g⟩
=⟨f, P g⟩.

Hence P ∗ = P and all the claims about P have now been proved.
Turning to Q, suppose Q is a projection and let N denote the range of Q. Since
Q2 = Q it follows that Qh = h for h ∈ N (the range of Q). Also from Q2 = Q we have
Q(I − Q) = 0. Thus if h, f ∈ H, then

⟨Qh, (I − Q)f ⟩ = ⟨h, Q(I − Q)f ⟩ = 0.

Choosing f = h, it follows that h = Qh + (I − Q)h is an orthogonal decomposition and

hence ∥Qh∥ ≤ ∥h∥ and so Q is continuous.
If (hn ) is a sequence from the range of Q that converges to h ∈ H, then, by continuity
of Q, the sequence (hn = Qhn ) converges to Qh and thus h = Qh so that the range of
Q is closed.
Next, f ∈ N ⊥ if and only if

0 = ⟨Qh, f ⟩ = ⟨h, Qf ⟩

for every h ∈ H; if and only if Qf = 0. Thus N ⊥ = ker(Q).

An easy argument shows I − Q is a projection too. In particular, f is in the range
of I − Q if and only if (I − Q)f = f . On the other hand (I − Q)f = f if and only if
Qf = 0. Thus the range of I − Q is the kernel of Q. Finally, given h ∈ H, we have
Qh ∈ N and h − Qh = (I − Q)h ∈ N ⊥ . Thus Q = PN . □

END Friday 2025-02-21 Though we had not proved the assertions about Q in Propo-
sition 5.29 nor Proposition 5.33.
 69

5.5. Orthonormal Sets and Bases. Recall, a subset E of a Hilbert space H is or-
thonormal if ∥e∥ = 1 for all e ∈ E, and if e, f ∈ E and e ̸= f , then e ⊥ f .
Definition 5.37. An orthonormal set is maximal if it is not contained in any larger
orthonormal set. A maximal orthonormal set is called an (orthonormal) basis or a Hilbert
space basis for H.
Proposition 5.38. An orthonormal set E is maximal if and only if E ⊥ = {0} if and
only if the span E = H.

An subset E of a Hilbert space H is a complete orthonormal set if E is orthonormal

and E ⊥ = {0}. Thus, E is a Hilbert space basis if and only if E is a complete orthonormal
set.

Proof. Suppose E is not maximal. Hence there is an orthonormal set F ⊇ E and a vector
f ∈ F \ E. In particular, 0 ̸= f ∈ E ⊥ . Conversely, if 0 ̸= f ∈ E ⊥ , then F = E ∪ { ∥ff ∥ }
is an orthormal set that properly contains E and hence E is not maximal.
For the second part, from what has been proved, E is maximal if and only if E ⊥ =
{0} if and only if E ⊥⊥ = H. On the other hand span E = E ⊥⊥ by Corollary 5.25 □
Remark 5.39. It must be stressed that a basis in the above sense need not be a basis
in the sense of linear algebra; that is, a basis for H as a vector space. In particular,
it is always true that an orthonormal set is linearly independent (Exercise: prove this
statement), but in general an orthonormal basis need not span H. In fact, if E is an
infinite orthonormal subset of H, then E does not span H. See Problem 3.6.
If E is an orthonormal set in a Hilbert space H, then E is a basis for the Hilbert
space span E.
Example 5.40. Here are some common examples of orthonormal bases.
(a) Of course the standard basis {e1 , . . . , en } is an orthonormal basis of Fn .

(b) In much the same way we get a orthonormal basis of ℓ2 (N); for each n define
(
1 if k = n
en (k) =
0 if k ̸= n
It is straightforward to check that the set E = {en }∞
n=1 is orthonormal. In fact, it is
a basis. To see this, notice that if h : N → F belongs to ℓ2 (N), then ⟨h, en ⟩ = h(n),
and hence if h ⊥ E, we have h(n) = 0 for all n, so h = 0.

(c) Let H = L2 [0, 1]. Consider for n ∈ Z the set of functions E = {en (x) = e2iπnx : n ∈
Z}. An easy exercise shows this set is orthonormal. Though not obvious, it is in
fact a basis. (See Problem 5.6.) Here is an outline of a proof. Given a (Lebesgue)
measurable set E ⊆ [0, 1], by regularity there exists an open set U and a closed set
70

F such that F ⊆ E ⊆ U and m(U \ F ) < ϵ, where m is Lebesgue measure. Let

K = [0, 1] \ U and define f : [0, 1] → R by
d(t, K)
f (t) = ,
d(t, F ) + d(t, K)
where d(t, S) = inf{|t − s| : s ∈ S} is the distance from a point t to the set S ⊆ [0, 1].
Because F, K are compact, the infima in these distance are attained. In particular,
f (t) = 1 for t ∈ F, and f (t) = 0 for t ∈ K, while otherwise 0 ≤ f (t) ≤ 1. It follows
that
Z 1
|f − χE |2 dm ≤ µ(U \ F ) < ϵ,
0

Since simple functions are dense in L2 ([0, 1]) (an exercise), it follows that con-
tinuous functions are too. Stone Weierstrass implies that the span of E (the set of
trigonmetric polynomials) is uniformly dense in C([0, 1]).

END Monday 2025-02-24

Theorem 5.41. Every Hilbert space H ̸= {0} has an orthonormal basis.

Proof. The proof is essentially the same as the Zorn’s lemma proof that every (non-
trivial) vector space has a basis. Let H be a Hilbert space and E the collection of
orthonormal subsets of H, partially ordered by inclusion. Since H ̸= (0), the collection
E is not empty. If (Eα ) is an ascending chain in E, then it is straightforward to verify
that ∪α Eα is an orthonormal set, and is an upper bound for (Eα ). Thus by Zorn’s
lemma, E has a maximal element, say E. □

Proposition 5.42 (Bessel’s Inequality). If E is an orthornormal set in a Hilbert space

H, then, for each h ∈ H,
X
|⟨h, e⟩|2 ≤ ∥h∥2 .
e∈E

In particular, Eh = {e ∈ E : ⟨h, e⟩ =
̸ 0} is at most countable.

Proof.
P For a finite subset F of E, observe that h is the sum of the orthogonal vectors
f = e∈F ⟨h, e⟩e and h − f. Hence,
X
∥h∥2 = ∥f ∥2 + ∥h − f ∥2 ≥ ∥f ∥2 = |⟨h, e⟩|2 .
e∈F

Thus,
X X
∥h∥2 ≥ sup{ |⟨h, e⟩|2 : F ⊆ E, |F | < ∞} = |⟨h, e⟩|2 . □
e∈F e∈E
 71

5.6. Convergent series in Hilbert space and basis expansions. This section be-
gins with a discussion of convergence of infinite series in Hilbert space before turning to
basis expansions and Parseval’s equality.
We have already encountered ordinary convergence and absolute convergence in our
discussion of completeness: recall that the series ∞
P PN
n=1 hn converges if limN →∞ n=1 hn
exists;
P∞ its limit h is called the sum of the series. The series converges absolutely if
n=1 ∥hn ∥ < ∞ and absolute convergence implies convergence.

∞
P∞ 5.43. Suppose H is a Hilbert space and (hn )n=1 is a sequence from H. The
Definition
series n=1 hn is unconditionally convergent
P∞ if there exists an h ∈ H such that for
each bijection φ : N → N the series n=1 hφ(n) converges to h. (In other words, every
reordering of the series converges, and to the same sum.)

Remark 5.44. Of course absolute convergence implies unconditional convergence. For

ordinary scalar series, or in a finite dimensional Hilbert space such as Fn , unconditional
convergence implies absolute convergence; however in infinite dimensional Hilbert space
unconditional convergence need not imply absolute convergence as example 5.46 follow-
ing the proof of Theorem 5.45 shows.

Theorem 5.45. Suppose E = {e1 , e2 , . . . } ⊆ H is a countable orthonormal set and (an )

is a sequence of complex numbers. The following are equivalent.

(i) the series ∞

P
P∞ j=1 aj ej converges;
2
(ii) j=1 |aj | converges; and
(iii) the series ∞
P
j=1 aj ej converges unconditionally.

If ∞
P
j=1 aj ej converges to g, then ⟨g, ej ⟩ = aj for all j.
Further, if h ∈ H, then the series
∞
X
(34) ⟨h, ej ⟩ej
j=1

is unconditionally convergent and, letting g denote the (unconditional) sum,

⟨g, ej ⟩ = ⟨h, ej ⟩

for all j.

P∞
Proof. Let sn denote the partial sums of the series j=1 aj ej ,
n
X
sn = aj ej .
j=1
72

Since H is complete, the series ∞

P
j=1 aj ej converges (meaning (sn ) converges) if and only
if for each ϵ > 0 there is an N so that for all m ≥ n ≥ N ,
Xm
2
(35) ∥sm − sn ∥ = |aj |2 < ϵ
j=n+1

(meaning (sn ) is Cauchy) if and only if the series m 2

P
j=1 |aj | converges. Hence items (i)
and (ii) are equivalent.
Now suppose sn = nj=1 aj ej converges to, say, g and φ : N → N is a permutation
P
Pn
(bijection). Let s′n = j=1 aφ(j) eφ(j) . Given ϵ > 0, choose N so that (35) holds. In
particular,
X ∞
|aj |2 ≤ ϵ.
j=N +1
3
Now choose M ≥ N so that
JN = {1, 2, . . . N } ⊆ {φ(1), φ(2), . . . φ(M )}.
For n ≥ M let Jn = {1, . . . n} and Jn′ = {φ(1), . . . φ(n)} and let Gn denote their
symmetric difference; that is Gn = (An \ Bn ) ∪ (Bn \ An ). From
X X X X
sn − s′n = aj ej = aj e j − aj e j ,
j∈Jn j∈Jn′ j∈Jn \Jn′ j∈Jn′ \Jn

it follows that X
∥sn − s′n ∥2 = ∥ak ∥2 .
k∈Gn
On the other hand Gn ⊆ since JN ⊆ Jn , Jn′ . Therefore,
JNc ,
X X∞
′ 2 2
∥sn − sn ∥ = |ak | ≤ |ak |2 ≤ ϵ.
k∈Gn N +1

Hence (s′n ) converges to g too. Hence item (ii) implies item (iii) and the proof of the
first part of the theorem is complete, since evidently item (iii) implies item (i).
Now suppose ∞
P Pn
j=1 aj ej converges to g and set sn = j=1 aj ej . Using Lemma 5.4,
since (sn ) converges to g, for each m, the sequence (⟨sn , em ⟩)n converges to ⟨g, em ⟩. On
the other hand, ⟨sn , em ⟩ = am for n ≥ m. Hence ⟨g, em ⟩ = am .
|⟨h, ej ⟩|2
P
For h ∈ H Bessel’s inequality, Theorem 5.42, implies the convergence of
P
and thus, by what has already been proved, the series ⟨h, ej ⟩ ej converges (uncondi-
tionally) to some g ∈ H and ⟨g, em ⟩ = ⟨h, em ⟩ for all m. □
Example 5.46. Suppose {e1 , e2 , . . . } is a countable orthonormal set in a Hilbert space
H. The series ∞
X 1
ej
j=1
j
3For instance M = max φ−1 (JN ).
 73

is Cauchy (verify this as an exercise) and hence converges to some h ∈ H. From

Theorem 5.45 it follows that ⟨h, ej ⟩ = 1j and the series above converges unconditionally
to h. On the other hand, this series does not converge absolutely and hence unconditional
convergence does not imply absolute convergence.

END Wednesday 2025-02-26

There is another notion of convergence in Hilbert space.
Definition 5.47. Suppose ∅ ̸= S ⊆ H and let F denote the collection of finite subsets
of S. The series X
s
s∈S
converges as a net if there exists h ∈ H such that for every ϵ > 0 there exists an F ∈ F
such that for every F ⊆ G ∈ F,
X
∥ s − h∥ < ϵ.
s∈G

Often S is presented as an indexed set, so that S = {hi : i ∈ I} for some set I, in which
P
case the series is written as i∈I hi .
Proposition 5.48. If E is an orthonormal subset of a Hilbert space H and h ∈ H, then
the series X
⟨h, e⟩e
e∈E
converges (as a net). Moreover, if g is the limit (as a net) of this series, then, for each
e ∈ E,
⟨g, e⟩ = ⟨h, e⟩.

Proof. Let Eh = {e ∈ E : ⟨h, e⟩ =

̸ 0}. From Bessel’s inequality, Proposition 5.42, Eh is at
most countable. Suppose Eh is countable and choose an enumeration, Eh = {e1 , e2 , . . . }.
By Theorem 5.45, the series
X∞
⟨h, ej ⟩ej
j=1

converges unconditionally to some g ∈ H and Pnmoreover ⟨g, ej ⟩ = ⟨h, ej ⟩ for all j. On

the other hand, since the partial sums sn = j=1 ⟨h, ej ⟩ej converge to h (in norm), for
each e ∈ E \ Eh , the sequence (0 = ⟨sn , e⟩)n converges to ⟨g, e⟩ and so ⟨g, e⟩ = 0. Hence
⟨g, e⟩ = ⟨h, e⟩ for all e ∈ E. In particular, Eg = Eh .
P
To prove the series e∈E ⟨h, e⟩e converges to g as a net, let ϵ > 0 be given. There
is an N so that
XN
∥g − ⟨h, ej ⟩ej ∥ < ϵ
j=1
74

and hence
∞
X
|⟨h, ej ⟩|2 < ϵ2 .
j=N +1

Let F = {e1 , . . . , eN }. If G ⊆ E is finite and F ⊆ G, then, letting T = G \ F,

X X X
∥ ⟨h, e⟩e∥2 = ∥ ⟨h, e⟩e + ⟨h, e⟩e∥2
e∈T e∈Eh ∩T e∈(E\Eh )∩T
X ∞
X
2
=∥ ⟨h, e⟩e∥ ≤ |⟨h, ej ⟩|2 < ϵ2 .
e∈Eh \F j=N +1

Hence
X X X
∥g − ⟨h, e⟩ e∥ ≤ ∥g − ⟨h, e⟩∥ + ∥ ⟨h, e⟩∥
e∈G e∈F e∈T
N
X X
= ∥g − ⟨g, ej ⟩ ej ∥ + ∥ ⟨g, e⟩e∥ < 2ϵ.
j=1 e∈T
P
Hence e∈E ⟨h, e⟩ converges as a net to g. □

Item (d) in the following theorem is known as Parseval’s equality. For each h ∈ H,
the series in item (b) converges (as a net) to some g ∈ H by Proposition 5.48.

Theorem 5.49. If E ⊆ H is an orthonormal set, then the following are equivalent:

(a) E is a (orthonormal) basis for H;
P
(b) h = e∈E ⟨h, e⟩e for each h ∈ H;
P
(c) ⟨g, h⟩ = e∈E ⟨g, e⟩⟨e, h⟩ for each g, h ∈ H; and
(d) ∥h∥2 = e∈E |⟨h, e⟩|2 for each h ∈ H.
P

Proof. Suppose E is an orthonormal set in H, but item (b) does not hold. Thus there
P
is an h ∈ H such that h ̸= e∈E ⟨h, e⟩e. By Proposition 5.48,
X
⟨h, e⟩e
e∈E

converges (as a net) to some g ∈ H and moreover ⟨g, e⟩ = ⟨h, e⟩ for all e ∈ E. By
assumption, f = g − h ̸= 0. On the other hand,
⟨f, e⟩ = ⟨g − h, e⟩ = 0,
and thus E ⊥ ̸= {0} so that, by Proposition 5.38, E is not maximal. Hence item (a)
implies item (b).
 75

Now suppose item (b) holds and let h, g ∈ H be given. Given ϵ, choose a finite
subset F of E such that if F ⊆ G ⊆ E, then
X X √
∥h − ⟨h, e⟩e∥, ∥g − ⟨g, e⟩e∥ < ϵ
e∈G e∈G
and observe, using the Cauchy-Schwarz inequality,
X X
ϵ > ⟨h − ⟨h, e⟩e, g − ⟨g, f ⟩f
e∈G f ∈G
X
= ⟨h, g⟩ − ⟨h, e⟩ ⟨e, g⟩ .
e∈G

Hence item (b) implies item (c).

Item (d) follows from item (c) by choosing g = h. Finally, suppose that item (a)
does not hold. In that case there exists a unit vector h ∈ H such that h is orthogonal
to E. Thus ⟨h, e⟩ = 0 for all e ∈ E so that
X
|⟨h, e⟩|2 = 0 ̸= 1 = ∥h∥2
e∈E

and item (d) does not hold. □

Given a set E, let V denote the vector space of finite linear combinations of elements
of E and define an inner product on V by declaring ⟨e, f ⟩ = 0 if e, f ∈ E and e ̸= f and
⟨e, e⟩ = 1 for e ∈ E. The completion H of V (see Proposition 2.31) is a Hilbert space,
denoted ℓ2 (E).
Corollary 5.50. If E is a basis for a Hilbert space H, then H is isomorphic, as a Hilbert
space, to ℓ2 (E).

5.7. Gram-Schmidt and Hilbert space dimension.

Theorem 5.51. Let {e1 , . . . en } be an orthonormal set in H, and let M = span{e1 , . . . en }.
The orthogonal projection P = PM onto M is given by, for h ∈ H,
Xn
(36) Ph = ⟨h, ej ⟩ej .
j=1
Pn
Proof. Given h ∈ H, let g = j=1 ⟨h, ej ⟩ej . Since g ∈ M, it suffices to show (h−g) ⊥ M .
For 1 ≤ m ≤ n,
* n +
X
⟨h − g, em ⟩ =⟨h, em ⟩ − ⟨h, ej ⟩ej , em
j=1
Xn
= ⟨h, em ⟩ − ⟨h, ej ⟩⟨ej , em ⟩
j=1

= ⟨h, em ⟩ − ⟨h, em ⟩ = 0.
76

It follows that h − g is orthogonal to {e1 , . . . , en } and hence to M . □

END Friday 2025-02-28

Theorem 5.52 (Gram-Schmidt process). If (fn )∞ n=1 is a linearly independent sequence

in H, then there exists an orthonormal sequence (en )∞
n=1 such that span{f1 , . . . fn } =
span{e1 , . . . en } for each n.

Proof. The proof proceeds by induction. Put e1 = f1 /∥f1 ∥ and note {e1 } is an orthormal
set and span{e1 } = span{f1 }. Assuming e1 , . . . en have been constructed satisfying the
P of the theorem, let P = PMn where Mn = span{e1 , . . . , en } and let gn+1 =
conditions
fn+1 − nj=1 ⟨fn+1 , ej ⟩ej = fn+1 − P f. By Theorem 5.51 gn+1 is orthogonal to Mn . It is
also not 0 by the independence assumption on the fj . Let en+1 = ∥ggn+1
n+1
∥
. □

Corollary 5.53. Suppose H is a Hilbert space. If H has finite dimension n ≥ 1 as

a vector space, then there exists an orthonormal set {e1 , . . . , en } in H that spans H.
Conversely, if there is a positive integer n and an orthonormal set {e1 , . . . , en } that
spans H, then H has finite dimension n as a vector space.
In particular, if H contains a finite maximal orthonormal set, then every maximal
orthornormal set in E has the same cardinality and moreover this cardinality is the
dimension of H as a vector space.

Remark 5.54. If H has a finite orthonormal basis E = {e1 , . . . , en }, then by Theorem

5.49(b), E spans (in the sense of linear algebra) and is therefore a vector space (Hamel)
basis for H. Hence H has dimension n as a vector space and further every orthonormal
basis of H has exactly n elements.
On the other hand, if H has an infinite orthonormal basis E, then it contains an
infinite linearly independent set (the basis E) and so has infinite dimension as a vector
space. □

Theorem 5.55. Any two bases of a Hilbert space H have the same cardinality.

The proof uses some basic facts about cardinality. Two sets A and B have the same
cardinality, written |A| = |B| if there is a bijection f : A → B. If there is a one-one
map f : A → B we write |A| ≤ |B|. By the Cantor–Schröder–Bernstein theorem, if
|A| ≤ |B| and |B| ≤ |A|, then |A| = |B|. If A is an infinite set, then |A × N| = |A|4
and if A is an infinite set and Ba is an at most countable set for each a ∈ A, then
| ∪a∈A Ba | ≤ |A × B| = |A|. The theorem says if E, F are orthonormal bases for H, then
|E| = |F |.

4That|S × S| = |S| for an infinite set S in full generality requires the axiom of choice. On the other
hand, since |N| = |N × N the proof of Theorem 5.55 given below shows if H has a countable orthonormal
basis, then all orthonormal basis of H are countable.
 77

Proof. Suppose E, F are orthonormal bases for H. If E is finite, then E is a basis in

the vector space sense and thus H is finite dimensional as a vector space. Since F is
orthonormal, it is linearly independent and hence |F | ≤ |E|. Thus F is also a vector
space basis for H and so |F | = |E|. By symmetry, either both E and F are finite and
have the same cardinality or both are infinite. Accordingly suppose both are infinite.
Fix e ∈ E and consider the set
Fe = {f ∈ F | ⟨f, e⟩ =
̸ 0}.
Since F is orthonormal, each Fe is at most countable by Proposition 5.42 and since E
S
is a basis, each f ∈ F belongs to at least one Fe . Thus e∈E Fe = F , and
[
|F | = Fe ≤ |E × N| = |E|,
e∈E

where the last equality holds since E is infinite.

By symmetry, |E| ≤ |F | and the proof is complete. □

In light of this theorem, we make the following definition.

Definition 5.56. The (orthogonal) dimension of a Hilbert space H is the cardinality
of any orthonormal basis, and is denoted dim H. If dim H is finite or countable, H is
separable and in this case the terminology H is a separable Hilbert space is commonly
used.
Corollary 5.57. Suppose H is a Hilbert space. If H is finite dimensional as a vector
space, then H is separable as a metric space.
If H is not finite dimensional as a vector space, then H is separable as a metric
space if and only if there is a a countable orthonormal set E = {e1 , e2 , . . . } such that
span E = H.

Proof. We consider the complex case of F = C, the real case being similar. If H has
a countable orthonormal basis E = {e1 , e2 , . . . }, then the set D = { nj=1 aj ej : aj ∈
P

Q + iQ, n ∈ N} is dense in H since E ⊆ D so that H = span E ⊆ span D ⊆ H.

The proof that a finite dimensional Hilbert space is separable as a metric space is
similar to the proof above.
Now suppose a basis of E contains uncountably many√elements (and thus all bases
of E are uncountable by Theorem 5.55). Since ∥e − f ∥ = 2 for all e, f ∈ H such that
e ̸= f, if C is a countable subset of H, then E ̸⊆ ∪c∈C B(c, 1) and hence C is not dense.
Thus H is not separable as a metric space. □

5.8. Weak convergence. [Optional] In addition to the norm topology, Hilbert spaces
carry another topology called the weak topology. In these notes we will stick to the
seperable case and just study weakly convergent sequences.
78

Definition 5.58. Let H be a seperable Hilbert space. A sequence (hn ) in H converges

weakly to h ∈ H if for all g ∈ H,
⟨hn , g⟩ → ⟨h, g⟩.

The Cauchy-Schwarz inequality implies if (hn ) converges to h in norm, then (hn )

converges weakly to h. However, when H is infinite-dimensional, the converse can fail.
For instance, let {en }∞
n=1 be an orthonormal basis for H. Then (en ) converges to 0
weakly. (The proof is an exercise, see Problem 5.9). On the other hand, the sequence
(en ) is not norm convergent, since it is not Cauchy. In this section weak convergence is
characterized as “bounded coordinate-wise convergence” and it is shown that the unit
ball of a separable Hilbert space is weakly sequentially compact.

Proposition 5.59. Let H be a Hilbert space with orthonormal basis {ej }∞

j=1 . A sequence
(hn ) in H is weakly convergent if and only if
i) supn ∥hn ∥ < ∞, and
ii) limn ⟨hn , ej ⟩ exists for each j.

Proof. Suppose (hn ) converges to h weakly. For each n

Ln (g) = ⟨g, hn ⟩
is a bounded linear functional on H. Since, for fixed g, the sequence |Ln (g)| converges, it
is bounded. Thus, the family of linear functionals (Ln ) is pointwise bounded and hence,
by the Principle of Uniform boundedness, sup ∥hn ∥ = sup ∥Ln ∥ < ∞, showing (i) holds.
Item (ii) is immediate from the definition of weak convergence.
Conversely, suppose (i) and (ii) hold, let M = sup ∥hn ∥. Define
ĥj = lim⟨hn , ej ⟩.
We will show that j |ĥj |2 ≤ M (so that the series
P P
ĥj ej is norm convergent in H);
we then define h to be the sum of this series and show that hn → h weakly.
For positive integers J and all n,
J
X
|⟨hn , ej ⟩|2 ≤ ∥hn ∥2 ≤ M 2
j=1

by Bessel’s inequality. Thus,

J
X J
X J
X
|ĥj |2 = lim |⟨hn , ej ⟩|2 = lim |⟨hn , ej ⟩|2 ≤ M 2 .
n n
j=1 j=1 j=1

Thus j |ĥj |2 ≤ M 2 and therefore the series j ĥj ej is norm convergent to some h ∈ H
P P

such that ⟨h, ej ⟩ = ĥj by Theorem 5.45. By Theorem 5.49, ∥h∥ ≤ M .

 79

Now wePprove that (hn ) converges to h weakly. Fix g ∈ H and let ϵ > 0 be given.
Since g = j ⟨g, ej ⟩ej (where the series is norm convergent) there exists an positive
integer J large enough so that
J
X ∞
X
g− ⟨g, ej ⟩ej = ⟨g, ej ⟩ej < ϵ.
j=1 j=J+1
PJ
Let g0 = j=1 ⟨g, ej ⟩ej , write g = g0 + g1 , observe ∥g1 ∥ < ϵ and estimate,
|⟨hn − h, g⟩| ≤ |⟨hn − h, g0 ⟩| + |⟨hn − h, g1 ⟩|.
By (ii), the first term on the right hand side goes to 0 with n, since g0 is a finite sum of
ej ’s. By Cauchy-Schwarz, the second term is bounded by 2M ϵ. As ϵ was arbitrary, we
see that the left-hand side goes to 0 with n. □

It turns out, if (hn ) converges to h weakly, then ∥h∥ ≤ lim inf ∥hn ∥ and further, still
assuming (hn ) converges weakly to h, ∥h∥ = lim ∥hn ∥ if and only if (hn ) converges to h
in norm. See Problem 5.9.
Theorem 5.60 (Weak compactness of the unit ball in Hilbert space). If (hn ) is a
bounded sequence in a separable Hilbert space H, then (hn ) has a weakly convergent
subsequence.

Proof. Using the previous proposition, it suffices to fix an orthonromal basis (ej ) and
produce a subsequence (hnk )k such that ⟨hnk , ej ⟩ converges for each j. This is a standard
“diagonalization” argument, and the details are left as an exercise (Problem 5.11) □

5.9. Problems.
Problem 5.1. Prove the complex form of the polarization identity: if H is a Hilbert
space over C, then for all g, h ∈ H
1
∥g + h∥2 − ∥g − h∥2 + i∥g + ih∥2 − i∥g − ih∥2


⟨g, h⟩ =
4
Problem 5.2. (Adjoint operators) Let H be a Hilbert space and T : H → H a bounded
linear operator.
a) Prove there is a unique bounded operator T ∗ : H → H satisfying ⟨T g, h⟩ =
⟨g, T ∗ h⟩ for all g, h ∈ H, and ∥T ∗ ∥ = ∥T ∥.
b) Prove, if S, T ∈ B(H), then (aS +T )∗ = aS ∗ +T ∗ for all a ∈ F, and that T ∗∗ = T .
c) Prove ∥T ∗ T ∥ = ∥T ∥2 .
d) Prove kerT is a closed subspace of H, (ranT ) = (kerT ∗ )⊥ and kerT ∗ = (ranT )⊥ .
Problem 5.3. Let H, K be Hilbert spaces. A linear transformation T : H → K is
called isometric if ∥T h∥ = ∥h∥ for all h ∈ H, and unitary if it is a surjective isometry.
Prove the following:
80

a) T is an isometry if and only if ⟨T g, T h⟩ = ⟨g, h⟩ for all g, h ∈ H, if and only if

T ∗ T = I (here I denotes the identity operator on H).
b) T is unitary if and only if T is invertible and T −1 = T ∗ , if and only if T ∗ T =
T T ∗ = I.
c) Prove, if E ⊆ H is an orthonormal set and T is an isometry, then T (E) is an
orthonormal set in K.
d) Prove, if H is finite-dimensional, then every isometry T : H → H is unitary.
e) Consider the shift operator S ∈ B(ℓ2 (N)) defined by
(37) S(a0 , a1 , a2 , . . . ) = (0, a0 , a1 , . . . )
Prove S is an isometry, but not unitary. Compute S ∗ and SS ∗ .

Problem 5.4. For any set J, let ℓ2 (J) denote the set of all functions f : J → F such
that j∈J |f (j)|2 < ∞. Then ℓ2 (J) is a Hilbert space.
P

a) Prove ℓ2 (I) is isometrically isomorphic to ℓ2 (J) if and only if I and J have the
same cardinality. (Hint: use Problem 5.3(c).)
b) Prove, if H is any Hilbert space, then H is isometrically isomorphic to ℓ2 (J) for
some set J.

Problem 5.5. Let (X, M , µ) be a σ-finite measure space. Prove the simple functions
that belong to L2 (µ) are dense in L2 (µ).

Problem 5.6. (The Fourier basis) Prove the set E = {en (t) := e2πint |n ∈ Z} is an
orthonormal basis for L2 [0, 1]. (Hint: use the Stone-Weierstrass theorem to prove that
the set of trigonometric polynomials P = { N 2πint
P
n=−M cn e } is uniformly dense in the
space of continuous functions f on [0, 1] that satisfy f (0) = f (1). Then show that this
space of continuous functions is dense in L2 [0, 1]. Finally show that if fn is a sequence
in L2 [0, 1] and fn → f uniformly, then also fn → f in the L2 norm.)

Problem 5.7. Let (gn )n∈N be an orthonormal basis for L2 [0, 1], and extend each function
to R by declaring it to be 0 off of [0, 1]. Prove the functions hmn (x) := 1[m,m+1] (x)gn (x −
m), n ∈ N, m ∈ Z form an orthonormal basis for L2 (R). (Thus L2 (R) is separable.)

Problem 5.8. Let (X, M , µ), (Y, N , ν) are σ-finite measure spaces, and let µ×ν denote
the product measure. Prove, if (fm ) and (gn ) are orthonormal bases for L2 (µ), L2 (ν)
respectively, then the collection of functions {hmn (x, y) = fm (x)gn (y)} is an orthonromal
basis for L2 (µ × ν). Use this result to construct an orthonormal basis for L2 (Rn ), and
conclude that L2 (Rn ) is separable.

Problem 5.9. (Weak Convergence)

a) Prove, if (hn ) converges to h in norm, then also (hn ) converges to h weakly.
(Hint: Cauchy-Schwarz.)
 81

b) Prove, if H is infinite-dimensional, and (en ) is an orthonormal sequence in H,

then en → 0 weakly, but en ̸→ 0 in norm. (Thus weak convergence does not
imply norm convergence.)
c) Prove (hn ) converges to h in norm if and only if (hn ) converges to h weakly and
∥hn ∥ → ∥h∥.
d) Prove if (hn ) converges to h weakly, then ∥h∥ ≤ lim inf ∥hn ∥.
Problem 5.10. Suppose H is countably infinite-dimensional (separable Hilbert space).
Prove, if h ∈ H and ∥h∥ < 1, then there is a sequence hn in H with ∥hn ∥ = 1 for all n,
and (hn ) converges to h weakly, but hn does not converge to h strongly.
Problem 5.11. Prove Theorem 5.60.
Problem 5.12. Prove, if (an ) is a sequence of complex numbers, then the following are
equivalent.
P
(1) Pn∈N an converges as a net;
(2) P∞ n=1 an converges unconditionally;
(3) ∞ n=1 an converges absolutely.
P∞
Problem 5.13. Suppose (h n ) is a sequence from a Hilbert space H. Show, if n=1 hn
converges absolutely, then ∞
P
n=1 nh converges unconditionally and as a net.
Problem
P∞ 5.14. Suppose H is a Hilbert space and P (hj ) is a sequence from H. Show,
h
j=1 j converges unconditionally if and only if j∈N hj converges as a net. (Warning:
showing unconditional convergence implies convergence as a net is challenging.)

6. Signed measures

In this section we consider measures with codomain F (either R of C) instead of

[0, ∞].

6.1. Definitions, examples and elementary properties.

Definition 6.1. Let (X, M ) be a measurable space. A signed measure or an F-measure
is a countably additive function ρ : M → F; that is, if (En )∞
n=1 is a disjoint sequence of
measurable sets, then
X∞ ∞
[
(38) ρ(En ) = ρ( En ).
n=1 n=1

Sometimes the terminology positive .measure is used instead of simply measure to

indicate ρ takes values in [0, ∞] and then finite positive measure indicates ρ takes values
in [0, ∞).
Remark 6.2. Several remarks are in order before proceeding.
(a) Choosing En = ∅ for all n obtains ρ(∅) = ∞
P
n=1 ρ(∅). Hence ρ(∅) = 0.
82

(b) Since ρ(∅) = 0, it follows that the countable additivity condition also includes finite
additivity by choosing En = ∅ as needed to pass from a finite set of sets to a
countable one.
(c) Given a permutation π : N → N, since ρ( ∞
S S∞
n=1 En ) = ρ( n=1 Eπ(n) ), the series in
equation (38) converges unconditionally and hence, by Riemann’s rearrangement
theorem, absolutely.
Alternately, one can see directly that the series in equation (38) must converges
absolutely as follows. First take F = R and let Let K+ = {k : ρ(Ek ) ≥ 0} and let
K− = {k : ρ(Ek ) < 0}. The collections A± = {Ek : k ∈ K± } are at most countable
and their elements are pairwisde disjoint. Hence,
X X
±ρ(∪k∈K± Ek ) = ± ρ(Ek ) = |ρ(Ek )|,
k∈K± k∈K±

where the fact that the order of summation is immaterial for series with non-negative
terms has been used. The complex case follows from the real case, since real ρ and
image ρ are both real-measures and |ρ(E)| ≤ | real ρ(E)| + | image ρ(E)|.
(d) The argument in item (c) also proves the following in the case F = R. Given disjoint
sets E1 , . . . , En ∈ M , let A = ∪{Ek : ρ(Ek ) ≥ 0} and B = ∪{Ek : ρ(Ek ) < 0} and
observe
(i) ρ(A), −ρ(B) ≥ 0;
(ii) E = ∪Ej = A ∪ B;
(iii) ρ(E) = ρ(A) − ρ(B); and
P
(iv) |ρ(Ej )| = |ρ(A)| + |ρ(B)| = ρ(A) − ρ(B).
(e) In the case F = R the theory can be developed allowing ρ to take values in either
(−∞, ∞] or [−∞, ∞) (so as to avoid ∞ − ∞). We will eschew this extra generality.
(f) If ρ1 , . . . , ρn are finite positive measures on a measure space (X, M ) and a1 , . . . , an ∈
j=1 aj ρj is an F-measure on M . In this way MF (M ), the set of
Pn
F, then ρ =
measures on (X, M ) becomes a vector space.
(g) If µ is a finite positive measure on a measure space (X, M ) and F ∈ M , then

ρ(E) = µ(E ∩ F ) − µ(E ∩ F c )

defines an R-measure on M .

END Monday 2025-03-03, though we had not yet discussed items (d), (e) and (f) in
the remark above
 83

Proposition 6.3. If (X, M , µ) is a measure space and f ∈ L1 (µ), then the function
µf : M → F defined by
Z
µf (E) = f dµ
E
is a signed measure.
In particular, if F = R and f = f + − f − is the decomposition of f into is positive
and negative parts, then
µf = µf + − µf − .

Proof. First consider the case F = R. If f is unsigned, then we have already seen µf
is a finite positive measure. Dropping the assumption that f is unsigned, consider the
decomposition of f into is positive and negative parts, f = f + − f − . Each of µf ± is a
finite positive measure and hence so is µf = µf + − µf − .
For the complex case, write f = g + ih, for real valued functions h, g ∈ L1 (µ) and
apply the already proven real case of to each of h and g. □

Remark 6.4. The measure µf is often denoted f dµ.

Note that in the real case the real case, by choosing E+ = {f ≥ 0} and E− =
{f < 0}, the proof actually shows there exists finite positive measures µ± such that
µf = µ+ − µ− and also such that µ± (E) = µf (E ∩ E± ) and of course E+ ∩ E− = ∅. In
particular, if F ⊆ E± , the µ± (F ) ≥ 0. □

6.2. Total variation.

Definition 6.5. The total variation, |ρ|, of an F-measure ρ on a measure space (X, M )
is the function |ρ| : M → [0, ∞] defined for E ∈ M by
Xn
|ρ|(E) = sup{ |ρ(Ej )| : n ∈ N, {E1 , . . . , En } is a measurable partition of E}.
j=1

Remark 6.6. In the notation of Definition 6.5, for E ∈ M ,

(i) |ρ(E)| ≤ |ρ|(E);
(ii) ρ(E) = |ρ(E)| if ρ is a positive measure;
(iii) |ρ(E)| = 0 if and only if ρ(A ∩ E) = 0 for all A ∈ M ;
(iv) If F ∈ M and F ⊆ E, then |ρ|(F ) ≤ |ρ|(E);
(v) |ρ|(E) = sup{ nj=1 |ρ(Ej )| : E1 , . . . , En ∈ M are disjoint and ∪nj=1 Ej ⊆ E}.
P

(vi) If µ is a (positive) measure on (X, M ) and |ρ(f )| ≤ µ(F ) for all F ∈ M , then
|ρ| ≤ µ in the sense that |ρ|(E) ≤ µ(E) for all E ∈ M .
(vii) If F = R, then |ρ(E)| = sup{|ρ(A)| + |ρ(B)| : A, B ∈ M , A ∩ B = ∅, A ∪ B ⊆ E}.
84

Proposition 6.7. SupposeR (X, M , µ) is a measure space. If h ∈ L1 (µ), then |h dµ| =

|h| dµ; that is |h dµ|(E) = E |h| dµ for all E ∈ M .

Before proving the theorem, we establish a couple of lemmas.

Lemma 6.8. In the context of Proposition 6.7, |h dµ| ≤ |h| dµ.

Proof. For notational ease, let µh = h dµ. Given E ∈ M and a measurable partition
{E1 , . . . , En } of E,
Xn n
X Z n Z
X n
X
|µh (Ej )| = h dµ ≤ |h| dµ = µ|h| (Ej ).
j=1 j=1 Ej j=1 Ej j=1

Thus |h dµ|(E) ≤ |h| dµ(E) as desired. □

END Wednesday 2025-03-05

Lemma 6.9. Proposition 6.7 holds for measurable simple functions h.

Proof. Since h is simple, there exists a measuarable partition {F1 , . . . , Fm } of X and

scalars c1 , . . . , cm ∈ F such that h = m 1
P
k=1 ck χFk . (Since h ∈ L (µ), for each k either
µ(Fk ) < ∞ or ck = 0.) Let E ∈ M and a measurable partition {E1 , . . . , En } of E be
given. Thus,
m,n
X
hχE = ck χEj ∩Fk
j,k=1

and {Ej ∩ Fk : 1 ≤ j ≤ n, 1 ≤ k ≤ m} is a measurable partition of E. Further,

m,n m,n Z m,n
X X X
|µh (Ej ∩ Fk )| = | h dµ| = |ck |µ(Ej ∩ Fk )
j,k=1 j,k=1 Ej ∩Fk j,k=1
n
XXZ m X n Z
= |h| dµ = |h| dµ
j=1 k=1 Ej ∩Fk j=1 Ej
Z
= |h| dµ = µ|h| (E).
E

Thus |h dµ|(E) ≥ |h| dµ(E). An application of Lemma 6.8 completes the proof. □

Proof of Proposition 6.7. Let ϵ > 0 be given. Since simple functions are dense in L1 (µ),
there is a simple function g ∈ L1 (µ) such that ∥h − g∥1 < ϵ. For a measurable set E,
Z Z Z
µ|g| (E) = |g| dµ ≥ |h| dµ − |h − g| dµ ≥ µ|h| (E) − ∥h − g∥1 > µ|h| (E) − ϵ.
E E E

Thus |g| dµ(E) > |h| dµ(E) − ϵ.

 85

Similarly, for a measurable partition {E1 , . . . , En } of E,

Xn n
X Z Xn Z Xn Z
|µh (Ej )| = | h dµ| ≥ | g dµ| − | (g − h) dµ|
j=1 j=1 Ej j=1 Ej j=1 Ej

Xn n Z
X
≥ |µg (Ej )| − |g − h| dµ
j=1 j=1 Ej

Xn n
X
≥ |µg (Ej )| − ∥g − h∥1 ≥ |µg (Ej )| − ϵ.
j=1 j=1

Thus, |h dµ|(E) ≥ |g dµ|(E) = ϵ and using Lemma 6.9, |h dµ|(E) ≥ |g dµ|(E) − ϵ =

|g| dµ(E) − ϵ > |h| dµ(E) − 2ϵ. It follows that |h dµ|(E) ≥ |h| dµ(E). □
Proposition 6.10. If ρ is an F-measure on a measurable space (X, M ), then |ρ| is a
(positive) measure on M .

Later we will see that |ρ| is a finite measure.

Proof. Since ρ(∅) = 0 it follows that |ρ|(∅) = 0.

Now suppose E1 , E2 , . . . is a disjoint sequence from M and let E = ∪∞ k=1 Ej . Given
measurable partitions {Ek,1 , . . . , Ek,nk } of Ek , for each N ∈ N the collection of sets
{Ek,j : 1 ≤ k ≤ N, 1 ≤ j ≤ nk } is a finite disjoint collection of measurable sets such
nk
that ∪Nk=1 ∪j=1 Ek,j ⊆ E. Thus, using Remark 6.6 item v,
nk
N X
X
|ρ(Ek,j )| ≤ |ρ|(E).
k=1 j=1
Pnk
Since j=1 |ρ(Ek,j )| ≤ |ρ|(Ek ) for each k, it follows that
N
X
|ρ|(Ek ) ≤ |ρ|(E)
k=1
and therefore ∞
X
|ρ|(Ek ) ≤ |ρ|(E).
k=1

To prove the reverse inequality, let {F1 , . . . , Fn } be a given measurable partition of

Pj,kn = Fj ∩Ek for 1 ≤ j ≤ n are disjoint measurable subsets Ek . By Remark 6.6
E. Thus, E
item (v), j=1 |ρ(Fj ∩ Ek )| ≤ |ρ|(Ek ) for each k. Therefore, since ρ is an F-measure,
n
X n
X ∞
X ∞
n X
X
|ρ(Fj )| = | ρ(Fj ∩ Ek )| ≤ |ρ(Fj ∩ Ek )|
j=1 j=1 k=1 j=1 k=1
X∞ Xn ∞
X
= |ρ(Fj ∩ Ek )| ≤ |ρ|(Ek ).
k=1 j=1 k=1
86
P∞
Thus |ρ|(E) ≤ k=1 |ρ|(Ek ) and the proof is complete. □
Proposition 6.11. If ρ is an F measure on the measurable space (X, M ), then |ρ| is a
finite measure. Equivalently, |ρ|(X) < ∞.
The proof of Proposition 6.11 uses the following lemma whose proof, based upon
disjointification, is the same as for the case of a finite positive measure.
Lemma 6.12. If ρ is an F-measure on the measurable space (X, M ) and E0 ⊇ E1 ⊇
E2 . . . is a decreasing sequence of measurable sets, then
ρ(∩∞
n=1 En ) = lim ρ(En ).
n→∞
Similarly, if E0 ⊆ E1 ⊆ E2 . . . is an increasing sequence from M , then
ρ(∪∞
n=1 En ) = lim ρ(En ).

END Friday 2025-03-07

Proof of Proposition 6.11. As a first step in proving the result in the case that ρ is a
real measure, suppose E ∈ M and |ρ|(E) = ∞ and let C > 0 be given. By item (vii)
of Remark 6.6, there exists disjoint sets A, B ∈ M such that A, B ⊆ E and |ρ(A)| +
|ρ(B)| ≥ 2(C + |ρ(E)|). Thus, without loss of generality,
|ρ(A)| ≥ C + |ρ(E)|.
Since |ρ| is a measure (Theorem 6.11),
|ρ|(A) + |ρ|(E \ A)| = |ρ|(E) = ∞,
so that either |ρ|(A) = ∞ or |ρ|(E \ A)| = ∞. We will show that |ρ(A)|, |ρ(E \ A)| ≥ C,
which shows that E contains a subset F ∈ M such that |ρ|(F ) = ∞ and |ρ(F )| ≥ C.
By construction, |ρ(A)| ≥ C. From ρ(E \ A) + ρ(A) = ρ(E), it follows that
|ρ(E \ A)| = ρ(E) − ρ(A) = ρ(A) − ρ(E) ≥ |ρ(A)| − |ρ(E)| ≥ C.
To prove the proposition still assuming ρ is a real measure, it suffices to show
|ρ|(X) < ∞. Arguing by contradiction, suppose |ρ|(X) = ∞. Choosing A = X and
C = 1, there is a measurable set E1 ⊆ X such that |ρ(E1 )| ≥ 1 and |ρ|(E1 ) = ∞.
Suppose now measurable sets E1 ⊇ E2 ⊆ E3 · · · ⊇ Em have been constructed such
that |ρ(En )| ≥ n and |ρ|(En ) = ∞ for 1 ≤ n ≤ m. It follows that, with A = Em and
C = m + 1, there is a measurable set Em+1 ⊆ Em such that |ρ(Em+1 )| ≥ m + 1 and
|ρ|(Em+1 ) = ∞. Thus recursion produces a nested decreasing sequence of measurable sets
(En )∞
n=1 such that |µ(En )| ≥ n. An application of Lemma 6.12 produces the contradiction
that |ρ(∩En )| = ∞.
To complete the proof, suppose now ρ is a F-measure. From what is already proved,
| real ρ|(X), | image ρ|(X) < ∞. On the other hand, |ρ(F )| ≤ | real ρ(F )| + | image ρ(F )|.
Hence |ρ|(E) ≤ | real ρ|(E) + | image ρ|(E) and thus |ρ|(X)| real ρ|(X) + | image ρ|(X) <
∞. □
 87

6.3. Banach spaces of measures.

Proposition 6.13. Suppose (X, M ) is a measurable space. The mapping ∥·∥ : MF (M ) →
[0, ∞) defined by ∥ρ∥ = |ρ|(X) is a norm on the space of measures.

Proof. Suppose ρ, τ ∈ MF (M ) and c ∈ F. From Remark 6.6 item (vii), ∥ρ∥ = 0 if and
only if ρ = 0. It is straightforward to verify that ∥cρ∥ = |c| ∥ρ∥.
Finally, to prove the triangle inequality, simply note that
∥ρ + τ ∥ = |(ρ + τ )(X)| = |ρ(X) + τ (X)| ≤ |ρ(X)| + |τ (X)| = ∥ρ∥ + ∥τ ∥. □
Proposition 6.14. Suppose (X, M ) is a measurable space. The normed space MF (M )
is a Banach space.

It is straightforward to show that if (ρn )n is a Cauchy sequence from MF (M ),

then, for each E ∈ M the sequence (ρn (E))n converges. That ρ : M → F given by
ρ(E) = limn ρn (E) that ρ ∈ MF (M ) and (ρn ) converges to ρ (in the normed vector space
MF (M )) is left to the gentle reader.

6.4. The Hahn decomposition.

Definition 6.15. Suppose ρ is an R-measure on a measurable space (X, M ). A set E ∈
M is totally positive (resp. totally negative) for ρ if ρ(F ∩ E) ≥ 0 (resp. ρ(F ∩ E) ≤ 0)
for all F ∈ M ; the set E totally null if ρ(F ∩ E) = 0 for all F ∈ M .
Remark 6.16. A set E is totally null for ρ if and only if it is both totally positive and
totally negative for ρ.
A set E ∈ M is totally positive for ρ if and only if ρ(F ) ≤ ρ(E) for all F ⊆ E.
(Consider E \ F ).
If E is totally positive for ρ, then ρe : M → F defined by ρe(F ) = ρ(E ∩ F ) is a finite
positive measure.
If (En )n is a sequence of totally positive sets, then ∪∞
n=1 En is also totally positive;
that is X+ is totally positive for ρf and X− is totally negative for ρf . □
Example 6.17. In the context of Proposition 6.7, decompose a real-valued function f ∈
L1 (µ) into its positive and negative parts f = f + − f − , the sets X+ := {x : f + (x) > 0}
and X− := {x : f − (x) > 0} are disjoint and totally positive for µf .

Theorem 6.18 (Hahn Decomposition Theorem). If ρ is an R-measure on the measurable

space (X, M ), then there exists a partition of X into disjoint measurable totally positive
sets X = X+ ∪ X− .
The decomposition is unique in the sense that if X+′ , X−′ is another such pair, then
X+ ∆X+′ and X− ∆X−′ are totally null for ρ.
88

END Monday 2025-03-10

The following lemma will be used in the proof of Theorem 6.18
Lemma 6.19. Suppose ρ is an R-measure on a measurable space (X, M ). If ρ(G) > 0,
then there exists a subset E ⊆ G such that E is totally positive and ρ(E) ≥ ρ(G).

The proof uses the greedy algorithm.

Proof. For notational convenience, let E1 = G.

If E1 is totally positive, then there is nothing to prove. Otherwise, there is a
measurable set H ⊆ E1 such that ρ(H) < 0 and thus ρ(E1 \ H) > ρ(E1 ); that is,
there is a measurable set F ⊆ E1 such that ρ(F ) > ρ(E1 ). Thus the set
1
J1 = {n ∈ N+ : there is an F ⊆ E1 such that ρ(F ) ≥ ρ(E1 ) + }
n
is nonempty and therefore has a smallest element n1 . Choose E2 such that ρ(E2 ) ≥
ρ(E1 ) + 1/n1 . If E2 is totally positive, then the proof is complete. Otherwise, let n2
denote the smallest element of
1
J2 = {n ∈ N+ : there is an F ⊆ E2 such that ρ(F ) ≥ ρ(E2 ) + }
n
and choose E3 ⊆ E2 such that ρ(E3 ) ≥ ρ(E2 ) + /n2 . (Note that n2 ≥ n1 and it could
1

be the case that n2 = n1 .) Either this recursion terminates after finitely many steps
producing a totally positive subset E of E1 with ρ(E) ≥ ρ(E1 ); or it generates a nested
decreasing sequence of measurable sets (Ej ) and a sequence of positive integers (nj ) such
that
1
ρ(Ej+1 ) ≥ ρ(Ej ) + ,
nj
where
1
(39) nj = min{n ∈ N+ : there is an F ⊆ Ej such that ρ(F ) ≥ ρ(Ej ) + }.
n
In particular,
j
X 1
(40) ρ(Ej+1 ) ≥ + ρ(E1 ).
k=1
nk
Assuming this latter case, let X+ = ∞
T
j=1 Ej . We will show that ρ(X+ ) > ρ(E1 ) and X+
is totally positive.
By Lemma 6.12, ρ(Ej ) increases to ρ(E) > ρ(G). Thus, by equation (40) and
the assumption that ρ is a finite measure, the sequence (nj )j converges to infinity (as
otherwise ρ(E) = ∞). To show that E must be totally positive, suppose, by way of
contradiction, there exists a measurable F ⊆ E such that ρ(F ) > ρ(E). There is an
m ∈ N+ such that ρ(F ) > ρ(E) + 1/m. There is a j such that nj > m. Now F ⊆ Ej and
1
ρ(F ) > ρ(E) + 1/m > ρ(Ej ) + > ρ(Ej ) + 1/m,
m
 89

contradicting the choice of nj in equation (39) and completing the proof. □

Proof of Theorem 6.18. If ρ(G) ≤ 0 for all G ∈ M , then ρ is totally negative and the
choices X+ = ∅ and X− = X satisfies the conclusion of the theorem.
Otherwise, by Proposition 6.11,
∞ > a = sup{ρ(G) : G ∈ M } > 0.
For each n ∈ N such that a > 1/n, there exists a set Gn such that ρ(Gn ) > a − 1/n > 0.
By Lemma 6.19, there exists a totally positive set En such that En ⊆ Gn and ρ(En ) ≥
ρ(Gn ) > 0. Let Fm = ∪m
n=1 En and note that Fm is totally positive by Remark 6.16. Thus
ρ(Fm \ Em+1 ) ≥ 0 and therefore
ρ(Fm+1 ) = ρ(Em+1 ) + ρ(Fm \ Em+1 ) ≥ ρ(Em+1 ).
By Lemma 6.12, (ρ(Fm ))m converges to ρ(E) where E = ∪∞ Fm = ∪∞ En . On the other
hand, (ρ(Fm )) converges to a. Hence ρ(E) = a. By Remark 6.16, E is totally positive.
Further, if F ⊆ E c , then a ≥ ρ(E ∪F ) = ρ(E)+ρ(F ) = a+ρ(F ) and therefore ρ(F ) ≤ 0.
Hence E c is totally negative and {E, E c } partition X.
For the final statement, observe if F ⊆ X+ \ X+′ = X+ ∩ X−′ , then ρ(F ) ≥ 0 since F
is a subset of the totally positive set X+ ; also ρ(F ) ≤ 0 since F is a subset of the totally
negative set X−′ . Hence ρ(F ) = 0 and thus X+ \X+′ is totally null. The remaining details
are left to the gentle reader. □

6.5. The Jordan decomposition.

Definition 6.20. Suppose ρ is an F-measure on a measurable space (X, M ). A set E
is a support set for ρ if E c is totally null for ρ. Two signed measures ρ, σ are mutually
singular , denoted ρ⊥σ, if they have disjoint support sets; that is, there exists disjoint
measurable sets E and F such that E c is totally null for ρ and F c is totally null for σ.
Remark 6.21. Two positive measures ρ and σ on the same measurable space (X, M )
are mutually singular if and only if there exists disjoint (measurable) sets E and F such
that ρ(E c ) = 0 = σ(F c ) (in which case it can be assumed that F = E c if desired).
Example 6.22. Let m denote Lebesgue measure on (R, L) (where L is the sigma-algebra
of Lebesgue measurable subsets of R) and let δ : L → R denote point mass at 0; that is
(
1 if 0 ∈ E
δ(E) =
0 if 0 ∈ / E.
It is immediate that m⊥δ.

Theorem 6.23 (Jordan Decomposition). If ρ is an R-measure on (X, M ), then there

exist unique positive measures ρ+ , ρ− such that ρ+ ⊥ρ− and ρ = ρ+ − ρ− . Moreover,
|ρ| = ρ+ + ρ− .
90

Proof. Let X = X+ ∪ X− be a Hahn decomposition for ρ (Theorem 6.18) and define

ρ± : M → R by ρ± (E) = ±ρ(E ∩ X± ). It is immediate from the properties of the Hahn
decomposition that ρ+ , ρ− have the desired properties; uniqueness is left as an exercise.
To prove the last statement let τ = ρ+ + ρ1 . For E ∈ M ,
|ρ|(E) ≥ |ρ(E ∩ X+ )| + |ρ(E ∩ X− )| = ρ+ (E) + ρ− (E) = τ (E).
On the other hand,
τ (F ) = ρ+ (F ) + ρ− (F ) = |ρ(F ∩ X+ )| + |ρ(F ∩ X− )|
≥ |ρ(F ∩ X+ ) + ρ(F ∩ X− )| = |ρ(F )|
for F ∈ M and hence τ ≥ |ρ| by Remark 6.6 item (vi). □

END Wednesday 2025-03-12

Example 6.24. Referring to Proposition 6.3 and example 6.17, it is now immediate that
the decomposition µf = µf + − µf − is the Jordan decomposition of µf . Thus the Jordan
decomposition theorem is analogous to the decomposition of a real-valued function into
its positive and negative parts.

6.6. The Radon-Nikodym derivative.

Definition 6.25. Suppose (X, M , µ) is a measure space. An F-measure ρ : M → F
is absolutely continuous with respect to µ, written ρ ≪ µ provided ρ(E) = 0 whenever
E ∈ M and µ(E) = 0.
Remark 6.26. Given a measure space (X, M , µ) and an f ∈ L1 (µ), the measure µf :
M → F defined by Z
µf (E) = f dµ
E
(see Proposition 6.3) is absolutely continuous with respect to µ. That is µf = f dµ ≪ µ.
If ρ is an F-measure on (X, M ), then ρ ≪ |ρ|. □
Theorem 6.27 (Radon-Nikodym). Suppose µ and ν are σ-finite positive measures on a
measurable space (X, M ). If ν ≪ µ, then there exists (an essentially unique) measurable
function h : X → [0, ∞) such that ν = µf ; that is
Z
ν(E) = h dµ
E
for all E ∈ M .
In the case that ν is finite, h ∈ L1 (µ).
dν
The function h is the Radon-Nikodym derivative of ν with respect to µ, denoted dµ
.
Corollary 6.28. Suppose (X, M , µ) is a σ-finite measure space and ρ : M → F is an
F-measure. If ρ ≪ µ, then there exists an h ∈ L1 (µ) such that ρ = h dµ.
 91

Proof. Suppose ρ is an R-measure and let ρ = ρ+ − ρ− denote its Jordan decomposi-

tion. It is routine to check if µ(E) = 0, then both ρ± (E) = 0. Two applications of
Theorem 6.27 produces unsigned functions h± ∈ L1 (µ) such that ρ± = h± dµ. Hence the
function h = h+ − h− is in L1 (µ) and ρ = h dµ.
Finally, if ρ is a C-measure, then real ρ and image ρ are R-measures. Thus there
exists h, g ∈ L1 (µ) such that real ρ = h dµ and image ρ = g dµ. Hence (h + ig) ∈ L1 (µ)
and ρ = (h + ig) dµ. □

Corollary 6.29. If ρ is an F-measure on a measurable space (X, M ), then, there exists

an h ∈ L1 (µ) such that |h| = 1 a.e. |ρ| and ρ = h dµ; that is
Z
ρ(E) = h d|ρ|
E

for all E ∈ M .

Proof. Note that ρ is absolutely continuous with respect to µ = |ρ| and hence the Radon-
Nikodym Theorem, Theorem 6.27, produces an h ∈ L1 (µ) such that ρ = h dµ. To see
that |h| = 1 almost everywhere µ, note that Proposition 6.7 implies µ = |ρ| = |h| dµ. In
particular, Z
(1 − |h|) dµ = 0
E
and the result follows by choosing E = {|h| =
̸ 1}. □

A consequence of the Lebesgue-Radon-Nikodym theorem is the existence of condi-

tional expectations.

Corollary 6.30. Suppose (X, M , µ) be a σ-finite measure space (µ a positive measure),

N a sub-σ-algebra of M , and ν = µ|N is σ-finite. If f ∈ L1 (µ) then there exists
g ∈ L1 (ν) (unique modulo ν-null sets) such that
Z Z
f dµ = g dν
E E

for all E ∈ N . (The function g is called the conditional expectation of f on N .)

Sketch of proof. Since f is M -measurable, it is also N measurable and moreover f ∈

L1 (ν) Thus, we may define ρ : N → F by
Z Z
(41) ρ(E) = f dν = f dµ.
E E

It is immediate that ρ is absolutely continuous with respect to ν. Thus, by Corollary 6.28,

there is an essentially unique g ∈ L1 (ν) such that equation 41 holds. □

END Friday 2025-03-14

92

Proof of Theorem 6.27. As a first step, assume µ and ν are finite positive measures.
This step is key and the proof given here is due to von Neumann.
Let m = ν + µ. Thus m is a finite positive measure and in particular 1 ∈ L2 (m),
where L2 (m) denotes the real Hilbert space of real valued square integrable (with respect
to m) functions. Given f ∈ L2 (m), observe that the Cauchy-Schwarz inequality gives
Z Z
(42) |f | dν ≤ |f | 1 dm = ⟨|f |, 1⟩L2 (m) ≤ ∥f ∥2 ∥1∥2 .

Hence it is sensible to define φ : L2 (m) → R by

Z
φ(f ) = f dν

and moreover, the estimate of equation (42) gives |φ(f )| ≤ ∥1∥2 ∥f ∥2 so that φ is a
bounded linear functional (with norm at most ∥1∥2 ).
By the Reisz representation Theorem (for Hilbert space), Theorem 5.28, there exists
a g ∈ L2 (m) (real-valued) such that
Z Z Z Z
f dν = φ(f ) = f g dm = f g dµ + f g dν.

and therefore,
Z Z
(43) f (1 − g) dν = f g dµ

for all f ∈ L2 (µ). With G1 = {g ≥ 1}, equation (43) we have

Z Z Z
0 ≤ µ(G1 ) = dµ = χG1 g dµ = χG1 (1 − g) dν ≤ 0.
G1

Hence µ(G1 ) = 0. Since ν ≪ µ, it also is the case the ν(G1 ) = 0 = m(G1 ).5 Now let
Gn = {g < − n1 } and observe, again using equation (43) that
Z Z
1
− µ(Gn ) ≥ χGn g dµ = χGn (1 − g) dν ≥ 0.
n
Thus µ(Gn ) = 0 and hence µ(G0 ) = 0, where G0 = ∪Gn = {g < 0}. As before it follows
that ν(G0 ) = 0 = m(G0 ).
Since m(G0 ) = 0 = m(G1 ), it is harmless to assume, as we now do, that 0 ≤ g < 1
1
pointwise. Let ψ = 1−g , set h = gψ and note both ψ and h are unsigned. The sequence
(ψn ) defined by
ψn = ψχ{ψ≤n}
is a pointwise increasing sequence of bounded non-negative functions that converges
pointwise to ψ. Thus each ψn ∈ L2 (m) and by the monotone convergence theorem
5Compare with the proof of Theorem 6.31 in Subsection 6.7, where absolute continuity is not assumed.
 93

(twice) and equation (43) the sequence

Z Z
ψn (1 − g)χE dν = ψn g χE dµ
R R
converges to both χE dν and hχE dµ; that is
Z
ν(E) = h dµ = µh (E)
E

for all E ∈ M . Choosing E = X shows h ∈ L1 (µ) completing the proof in this special
case that both ν and µ are finite.
We now sketch a proof of the case that both ν and µ are σ-finite (positive) measures.
Since ν and µ are σ-finite, there exists a sequence X1 ⊆ X2 ⊆ . . . of measurable
sets of finite measure such that X = ∪Xn . For n ∈ N define µn : M → [0, ∞) by
µn (E) = µ(E ∩Xn ) and define νn similarly. The pair (νn , µn ) are finite positive measures
and νn ≪ µn . Hence, by what has already been proved and with n = 1, there exists
h1 ∈ L1 (µ1 ) such that dν1 = h1 dµ1 . Without loss of generality, we assume h1 = 0 on
Xnc . With n = 2, there is an h2 ∈ L2 (µ) such that dν2 = h2 dµ2 and h2 = 0 on X2c .
Moreover, since h1 and h2 agree µ a.e. on X1 , we also assume, without loss of generality,
that h1 = h2 on X1 . Continuing in this fashion constructs an increasing sequence (hn ) of
unsigned functions that converges pointwise to some h and satisfies νn = hn dµn for each
n. Finally, given E ∈ M , let En = E ∩ En and apply the monotone convergence theorem
to the sequences χEn and hχEn and the measures ν and µ respectively to conclude that
the sequence
Z Z
χEn dν = νn (E) = hn dµn (E) = hχEn dµ
R
converges to both ν(E) and to E h dµ. Hence ν = h dµ. In the case that ν is finite,
choosing E = X and using ν(X) < ∞ gives h ∈ L1 (µ). □

6.7. The Lebesgue decomposition.

Theorem 6.31 (Lebesgue Decomposition - positive measure version). Suppose (X, M , µ)

is a σ-finite measure space. If ν is a finite positive measure on (X, M ), then there exist
unique positive measures νa and νs such that
(i) νa ≪ µ;
(ii) νs ⊥µ; and
(iii) ν = νa + νs .
Moreover, there exists a measurable set F such that
(i) νa (E) = ν(E ∩ F ) = 0;
(ii) νs (E) = ν(E ∩ F c ) = 0; and
(iii) µ(F c ) = 0.
94

Remark 6.32. The result holds if ν is assumed σ-finite, a result that follows easily from
the case of ν finite. The details are left to the interested reader. □

The uniqueness asserted in Theorem 6.31 is a consequence of the following lemma.

Lemma 6.33. Suppose (X, M , µ) is a measure space and ρ : M → R is a measure. If
ρ ≪ µ and ρ⊥µ, then ρ = 0.

Proof. Since ρ⊥µ, there exist a set F ∈ M such that ρ(E) = ρ(E ∩ F c ) and µ(E) =
µ(E ∩ F ) for all E ∈ M . Hence, for E ∈ M ,
µ(E ∩ F c ) = µ(E ∩ F c ∩ F ) = 0.
Since ρ ≪ µ, it follows that ρ(E ∩ F c ) = 0. Thus both F and F c are totally null for ρ.
Hence ρ = 0. □

Proof of Theorem 6.31. Let m = ν + µ. In particular, m is σ-finite and ν ≪ m. Hence,

by the Radon Nikodym Theorem, Theorem 6.27. there is a uniquely (a.e. m) determined
unsigned function g : X → [0, ∞) such that ν = mg . Thus, for all measurable E,
Z Z
ν(E) = g d(µ + ν) = g dm
E E
An easy argument shows g ≤ 1 a.e. m.
Let F = {g < 1} and note F c = {g = 1}.6 Define νa (E) = ν(E ∩ F ) and νs (E) =
ν(E ∩ F c ) for E ∈ M . Both are positive measures, νs (F ) = 0 and ν = νa + νs . Next,
Z
c
ν(F ) = g d(µ + ν) = µ(F c ) + ν(F c ).
Fc
Since ν(F ) ∈ [0, ∞) it follows that, µ(F c ) = 0 and thus νs ⊥µ.
c

To prove νa is absolutely continuous with respect to µ, suppose E is measurable and

µ(E) = 0. Letting Fn = {g < 1 − n1 } ⊆ F (for positive integers n),
1 1
ν(E ∩ Fn ) ≤ (1 −
)[µ + ν](E ∩ Fn ) = (1 − )ν(E ∩ Fn ).
n n
Hence ν(E ∩ Fn ) = 0. Since E ∩ F = ∪(E ∩ Fn ) it follows that νa (E) = ν(E ∩ F ) = 0
and therefore νa ≪ µ.
To prove uniqueness, suppose ν = ρa + ρs . Since these are finite measures, ρa − νa =
νs − ρs . Now the R-measure on the right hand side is singular with respect to µ while
the R-meaure on the left hand side is absolutely continuous with respect to µ. Hence,
by Lemma 6.33, both are 0. □

END Monday 2025-03-24

The Lebesgue decomposition easily extends to the case of F-measures
6In the case that ν ≪ µ ≪ m the set F c is m-null. Compare with the proof of Theorem 6.27.
 95

Theorem 6.34 (Lebesgue Decomposition - F-measure version). Suppose (X, M , µ) is

a σ-finite measure space. If ρ : M → F is an F-measure, then then there exist unique
measures ρa and ρs such that ρa ≪ µ and ρs ⊥µ and ρ = ρa + ρs .

Proof. Let ν = |ρ|. By Theorem 6.31 there exists measures νs and νa such that νa ⊥µ
and νs ≪ µ. Moreover, there exists a measurable set F such that νa (E) = ν(E ∩ F )
and νs (E) = ν(E ∩ F c ) for E ∈ M and µ(F c ) = 0. In particular, νs (F ) = 0. Define
ρa , ρs : M → F by ρa (E) = ρ(E ∩ F ) and ρs (E) = ρ(E ∩ F c ). If E ⊆ F is measurable,
then ρs (E) = ρ(E ∩ F c ) = ρ(∅) = 0 and hence F is totally null for ρs . Thus ρs ⊥µ, since
also µ(F c ) = 0. On the other hand, if µ(E) = 0, then, writing E = (E ∩ F ) ∪ (E ∩ F c )

|ρa (E)| = |ρ(E ∩ F )| ≤ |ρ|(E ∩ F ) = νa (E) = 0,

since νa ≪ µ. Hence ρa (E) = 0 and we conclude that ρa ≪ µ. By construction ρ = ρa +ρs .

Once again, uniqueness follows from Lemma 6.33. □

6.8. Duality for Lebesgue spaces - conclusion. This subsection contains a sketch
of a proof, based on the Radon-Nikodym Theorem (Theorem 6.27), that the isometric
map of Proposition 4.14 is in fact onto (unitary). Recall, given 1 ≤ p ≤ ∞ and a
g ∈ Lq (µ), where q is the conjugate index to p, that for f ∈ Lp (µ), Hölder’s inequality
(Theorem 4.8) implies gf ∈ L1 (µ) and moroever ∥f g∥1 ≤ ∥f ∥p ∥g∥q . Thus, we obtain a
bounded linear functional Lg : Lp (µ) → F of norm at most ∥g∥q defined by

Lg (f ) = gf.

Let Φ : Lq (µ) → Lp (µ)∗ denote the bounded map (with norm at most one) given by
Φ(g) = Lg .

Theorem 6.35. If (X, M , µ) is a σ-finite measure space and 1 ≤ p < ∞, then the
mapping Φ : Lq (µ) → Lp (µ)∗ defined by Φ(g) = Lg is an isometric isomorphism.

Recall, Problems 2.12 and 4.6 says that the result fails in the case of p = ∞. Likewise
the result fails for p = 1 without the σ-finite hypothesis. See Remark 4.15.

Proof of Theorem 6.35 in the case of a finite measure. Proposition 4.14 says Φ is iso-
metric. Thus it remains to show that Φ is onto under the assumption that µ is a
finite (positive) measure. Let φ ∈ Lp (µ)∗ be given. Define ν : M → F as follows.
Given E ∈ M the function χE ∈ L1 (µ) since µ is finite. Set ν(E) = φ(χE ). In par-
ticular, ν(∅) = 0. To prove that ν is countably additive and hence an F-measure,
∞ ∞
supposePn(En )n=1 is a sequence of disjoint measurable sets and let E = ∪j=1 Ej . Let
sn = j=1 1Ej . In particular, (sn ) increases pointwise with limit s = 1E . Further,
0 ≤ (s − sn )p ≤ 1 ∈ Lp (µ) and thus, by dominated convergence, (sn ) converges to s in
96

Lp (µ).7 Using continuity and linearity of φ,

n
X ∞
X
ν(E) = φ(s) = lim φ(sn ) = lim ν(Ej ) = ν(Ej ).
n→∞ n→∞
j=1 j=1

If µ(E) = 0, then χE = 0 in Lp and therefore ν(E) = 0. Thus, the measure

ν is absolutely continuous with respect to µ. Consequently, by the Radon-Nikodym
Theorem, there exists an h ∈ L1 (µ) such that
Z
(44) φ(χE ) = ν(E) = h dµ.
E
Temporarily, view Lh as defined (and continuous) on L∞ (µ). (See Proposition 4.14.)
If s is a measurable simple function, then, by equation (44),
Z
φ(s) = s h dµ.
X

Now suppose f is a bounded unsigned measurable function. Since µ is a finite

measure, f is in Lp (µ) as well as L∞ (µ). Hence, both λh (f ) and φ(f ) are defined.
There exists a sequence (sn ) of measurable simple functions 0 ≤ sn ≤ f such that (sn )
converges to f uniformly and therefore in both Lp (µ) and L∞ (µ) (again because µ is
finite). It follows
φ(f ) = lim φ(sn ) = Lh (sn ) = Lh (f ).
It now follows that if f is bounded and measurable, then φ(f ) = Lh (f ).
To prove h ∈ Lq , first assume p > 1. For positive integers N , let EN = {|h| ≤ N }
and let hN = hχEn . Thus hN is bounded and so is fN = hN |hN |q−2 (where we set
/ EN ). Thus all are in each Lr (µ) since µ is finite. By Lp (µ) continuity
fN (x) = 0 if x ∈
of φ,
Z Z
(45) ∥fN ∥p ∥φ∥ ≥ |φ(fN )| = |Lh (fN )| = fN h dµ = |hN |q dµ = ∥hN ∥qq .
X X

On the other hand, ∥fN ∥p = ∥hN ∥q−1

q ,
and combining this equality with (45) we see that
∥hN ∥q ≤ ∥φ∥. By monotone congvergence, h ∈ Lq (µ) and moreover ∥h∥q ≤ ∥φ∥.
h
In the p = 1 case, put Et = {|h| > t} and let ft = |h| χEt . Thus ∥ft ∥1 = µ(Et ) for
∞
all t, and, since f ∈ L (µ),
Z Z
(46) µ(Et ) ∥φ∥ = ∥φ∥ ∥ft ∥1 ≥ |φ(ft )| = |Lh (ft )| = ft h dµ = |h| dµ ≥ t µ(Et ).
X Et
∞
Hence µ(Et ) = 0 for t > ∥φ∥ and thus h ∈ L (µ) and in fact ∥φ∥ ≥ ∥h∥∞ .
Now that we know h ∈ Lq , it follows that Lh is continuous. It also agrees with φ
on simple functions. Since simple functions are dense in Lq (µ), the conclusion φ = Lh
follows. □
7On the other hand, (sn ) does not necessarily converge to s in L∞ (µ).
 97

END Wednesday 2025-03-26

The following lemmas will be used to prove Theorem 6.35 in the case µ is σ-finite.

Lemma 6.36. If µ (X, M , µ) is a σ-finite measure space, then there exists a measurable
function w ∈ L1 (µ) such that 0 < w(x) < 1 for all x.

Proof. Write X = ∪∞ n=1 Xn , a countable P

union of disjoint measurable sets of finite mea-
sure. Let wn = 2n (µ(Xn )+1) 1Xn and w = ∞
1
n=1 wn . □

Lemma 6.37. Suppose µ is a σ-finite measure, w ∈ L1 (µ) and 0 < w(x) < 1 for all x.
Let τ denote the measure w dµ.
1
For 1 < p < ∞, a measurable function f is in Lp (µ) if and only if g = w− p f ∈ Lp (τ )
and in this case ∥f ∥p = ∥g∥p ; that is, the mapping Ψp : Lp (τ ) → Lp (µ) defined by
1
Ψp (f ) = w p f
is a (linear) isometric isomorphism.
1
Proof. It is easy to check that Ψp is isometric with inverse given by f 7→ w− p f . □

Proof of Theorem 6.35. For the σ-finite case with p > 1, let w be as in Lemma 6.36.
Likewise, let τ = w dµ. By Lemma 6.37 the mappings Ψr : Lp (τ ) → Lp (µ) defined by
1
Ψr h = w r h are linear isometric isomorphisms. Thus, ψ = φ ◦ Φp is a bounded linear
functional on Lp (τ ). Since τ is a finite measure, by what is already proved, there is a
1
g ∈ Lq (τ ) such that ψ = Lg . Let h = Ψq g = w q h. Thus h ∈ Lq (µ) and ∥h∥q = ∥g∥q .
− p1
Moreover, if f ∈ Lp (µ), then F := Ψ−1 p f = w f ∈ Lp (τ ) and
φ(f ) =ψ(F ) = Lg (F )
Z Z
= F g dτ = F g w dµ
Z Z
1 1
= w p F (w q g) dµ = f h dµ = Lh (f ).

In the case p = 1, write X = ∪∞

n=1 Xn , where Xn are measurable sets of finite measure
and apply what has already been proven in the to the measure space (Xn , Mn , µn ), where
Mn = {E ∈ M : E ⊆ Xn } and µn = µ|Mn . The details are left to the gentle reader. □

6.9. Problems.

Problem 6.1. a) Prove Proposition 6.30. b) In the case µ = Lebesgue measure on

[0, 1), fix a positive integer k and let N be the sub-σ-algebra generated by the intervals
[ kj , j+1
k
) for j = 0, . . . k − 1. Give an explicit formula for the conditional expectation g
in terms of f . c) Show that the σ-finite hypothesis on ν is needed.
98

7. The Fourier transform

We assume all functions are complex-valued unless stated otherwise.

Definition 7.1. [The Fourier transform] Let f ∈ L1 (R). The Fourier transform of f is
the function F(f ) = fb : R → C defined at each t ∈ R by
Z ∞
(47) f (t) :=
b f (x)e−2πitx dx.
−∞

The terminology Fourier transform is often used for the mapping that sends f to fb. □

Note that fb makes sense, since f ∈ L1 (R) and, for each t, the function exp(−2πitx) ∈
L∞ (R). In fact, |fb(t)| ≤ ∥f ∥1 so that fb ∈ Fb (R), the Banach space of bounded functions
on R with the supremum norm, and ∥fˆ∥∞ ≤ ∥f ∥1 .
The basic properties of the Fourier transform listed in the following proposition stem
from two basic facts: (1) Lebesgue measure is translation invariant; and (2) that, for
each t ∈ R, the function
χt : x → exp(2πitx)
is a character of the additive group (R, +);8 that is, χt is a homomorphism from R into
the mulitplicative group of unimodular complex numbers. Explicitly for all x, y, t ∈ R
χt (x + y) = χt (x)χt (y).
Example 7.2. Given real numbers a < b, let f = χ[a,b] (that the interval is closed, open,
or neither is not important here) and verify
( −2πibt −2πiat
−e
ie 2πt
t ̸= 0
fb(t) =
b−a t = 0.
Note that the derivative of exp(−2πiαt) at t = 0 is −2πiα. In particular, for b > 0 and
1
f = 2b χ[−b,b] ,
fb(t) = sinc(2πbt).
Since for k ∈ N,
Z (k+1)π
sin(x) 2
≥ ,
kπ (k + 1) k+1
the sinc function is not in L (R); that is, in general f ∈ L1 (R), does not imply fb ∈
1

L1 (R). □

The following example will be used later when the Poisson kernel for the upper half
plane is introduced.
8The
characters of the multiplicative group T (the unit circle in the complex plane) are parameterized
by Z with n ∈ Z corresponding to the character χn (γ) = γ n ; that is χn (eit ) = eint (for t ∈ R).
Proceeding in this way lead to the theory of Fourier Series.
 99

Example 7.3. For a > 0, let

(48) Qa (t) = e−2aπ|t|
(the extra factor of 2π turns out to be a convenient normalization).
Z
1 a
Q
ca (−t) = Qa (t)e2πitx dt = .
R π a2 + x 2
Note, this function Q is, in a sense, a smoother version of the indicator function from
example 7.2. It has the virtue that is Fourier transform is in L1 (R). □

7.1. Basic Properties. Before going further we introduce some notation: for fixed
y ∈ R and a function f : R → C, define fy (x) := f (x − y).
Proposition 7.4 (Basic properties of the Fourier transform). Let f, g ∈ L1 (R) and let
α ∈ R.
\
(a) (Linearity) cf + g = cfb + gb
(b) (Translation) fby (t) = e−2πity fb(t)
(c) (Modulation) If g(x) = e2πiαx f (x), then gb(t) = fb(t − α)
(d) (Reflection) If g(x) = f (−x), then gb(t) = fb(t).
(e) (Scaling) If λ > 0 and g(x) = f (x/λ) then gb(t) = λfb(λt).

Proof. Each of these properties is verified by elementary transformations of the integral

defining fb; the details are left as an exercise. □
Proposition 7.5. If f ∈ L1 (R), then fb is continuous and bounded (fb ∈ Cb (R)) and
∥fb∥∞ ≤ ∥f ∥1 . In particular, the mapping L1 (R) ∋ f 7→ fb ∈ Cb (R) is a bounded linear
map of norm at most 1.

Proof. Fix t ∈ R and a sequence tn → t. The sequence f (x)e−2πitn x converges to

f (x)e−2πitx pointwise on R, and since trivially |f (x)e−2πitn x | ≤ |f (x)| for all n, we have
by dominated convergence

Z ∞
fb(t) = f (x)e−2πitx dx
Z−∞
∞
= lim [f (x)e−2πitn x ] dx
−∞ n→∞
Z ∞
= lim f (x)e−2πitn x dx
n→∞ −∞

= lim fb(tn ).
n→∞

The second statement of the theorem follows immediately from the estimate supt∈R |fb(t)| ≤
∥f ∥1 . □
100

In fact, fb always belongs to C0 (R), a result that is known as the Riemann-Lebesgue

Lemma. To prove it we first need the following result, which we will apply often (recall
the notation fy (x) := f (x − y)):
Lemma 7.6 (Translation is continuous on Lp ). If 1 ≤ p < ∞ and f ∈ Lp (R), then
limy→0 ∥fy − f ∥p = 0. In particular, if (yn ) converges to y, then (fyn )n converges to fy
in Lp (R).

Sketch. We sketch the proof of an approximation argument, leaving the details as an

exercise. Let X ⊆ Lp denote the set of all f for which the conclusion of the theorem is
true.
Verify X is a vector space and contains χI for all finite intervals I. By Littlewood’s
first principle, a measurable set of finite measure is nearly a finite union of intervals. Thus
X contains the indicator functions of all sets of finite measure and therefore all simple Lp
functions. Since simple Lp functions are dense in Lp , it suffices to show that X is closed.
Toward this end, note if f, g ∈ Lp and ∥f − g∥p < ϵ, then ∥f − g∥p = ∥fy − gy ∥p < ϵ for
all y ∈ R by the translation invariance of Lebesgue measure. Now suppose that g is in
the closure of X and let ϵ > 0 be given. Choose f ∈ X with ∥f − g∥p < ϵ, and choose
δ > 0 so that ∥fy − f ∥p < ϵ for all |y| < δ. Then for all |y| < δ,
∥gy − g∥p < ∥gy − fy ∥p + ∥fy − f ∥p + ∥f − g∥p < 3ϵ.
Thus g ∈ X as well and hence X is closed. The proof is finished. □

END Friday 2025-03-28

Lemma 7.7 (The Riemann-Lebesgue Lemma). If f ∈ L1 (R), then fb ∈ C0 (R).

Proof. From Proposition 7.5, fb is continuous.

The proof here that f vanishes at infinity appeals to the continuity of translation
in L1 (Lemma 7.6), and a simple trick. First, since e−πi = −1,
Z Z  
−2πit(x+(1/2t)) 1
(49) f (t) = − f (x)e
b dx = − f x − e−2πixt dx.
R R 2t
Combining equation (49) with the usual definition of fb, we have
Z   
1 1
fb(t) = f (x) − f x − e−2πixt dx.
2 R 2t
Thus
1
|fb(t)| ≤ ∥f − f 1 ∥1 .
2 2t

By Lemma 7.6 ∥f − f 1 ∥1 → 0 as t → ±∞. □

2t

Continuing our catalog of basic properties, we see that the Fourier transform also
interacts nicely with differentiation.
 101

Proposition 7.8 (Multiplication becomes differentiation). Suppose f ∈ L1 (R). If

g(x) := xf (x) belongs to L1 (R), then fb is differentiable and
−1 d b
gb(t) = f (t),
2πi dt
for all t ∈ R.

The proof uses the standard estimate,

|1 − eit | ≤ |t|
for t real and dominated convergence.

Proof. For real numbers s ̸= t

Z ∞ −2πisx
fb(s) − fb(t) e − e−2πitx
(50) = f (x) dx.
s−t −∞ s−t
The estimate
e−2πisx − e−2πitx
≤ 2π|x|
s−t
holds for all s ̸= t. Thus, by the assumption xf (x) ∈ L1 , a dominated convergence
argument in (50) shows that the limit as s → t exists and moreover
Z ∞ −2πisx
fb(s) − fb(t) e − e−2πitx
lim = lim f (x) dx
s→t s−t s→t −∞ s−t
Z ∞
= (−2πi)e−2πitx xf (x) dx
∞
= −2πib
g (t).
Thus fb is differentiable and the claimed formula holds. □

Note that if f ∈ L1 and also g(x) := xn f (x) ∈ L1 for some integer n ≥ 1, then
x f (x) belongs to L1 for all 0 ≤ k ≤ n. The previous proposition can then be applied
k

inductively to conclude:
Corollary 7.9. If f ∈ L1 and g := xn f ∈ L1 , then fb is n times differentiable, and
 k
−1
k
x f=
d fb(k) for each 0 ≤ k ≤ n.
2πi

One also expects a theorem in the opposite direction: the Fourier transform should
convert differentiation to multiplication by the independent variable. Under reasonable
hypotheses, this is the case.
Proposition 7.10. If f ∈ C0 (R) and f ′ is continuous and in L1 , then
F(f ′ )(t) = fb′ (t) = 2πitfˆ(t).
102

Proof. Compute
Z ∞
fb′ (t) = f ′ (x)e−2πitx dx
−∞
Z b
= lim f ′ (x)e−2πitx dx
b→∞ −b
 Z b 
−2πibt 2πibt −2πixt
= lim [f (b)e − f (−b)e ] + 2πit f (x)e dx
b→∞ −b

= 2πitfb(t),
where the second equality follows from the Dominated Convergence Theorem, the third
using integration by parts, and the fourth from the C0 (R) assumption on f and another
application of Dominated Convergence. □

Example 7.11. For a > 0 let g = ga denote the Gaussian,

2
g(x) := e−πax .
(The factor of π will be convenient given our choice of normalization in the definition of
the Fourier transform.)
Rather than computing the transform of ga directly, we exploit Propositions 7.8 and
7.10. We may also assume a = 1 since the general case follows from this by scaling
(Proposition 7.4(e)). Note that h ∈ L1 (R) and g ′ = −2πh. Thus,
g )′ (t) = −2πib
(b h(t)
1 ′
= −2πiF(− g)
(51) 2π
= i2πitĝ(t)
= −2πtĝ(t),
where the first equality follows from Proposition 7.8, the second from g ′ = −2πh and
the third from Proposition 7.10. It follows from equation (51) and the product rule that
d πt2
(e gb(t)) = 0.
dt
2
Hence the function eπt gb(t) is constant. To evaluate the constant, we set t = 0 and use
the well-known Gaussian integral
Z ∞
2
gb(0) = e−πx dx = 1.
−∞

We note in passing that F(h1 ) = −ih1 too.

As a final remark, the F(H4n g1 ) = H4n g1 , where Hn are (appropriately normalized)
hermite polynomials. □
 103

7.2. Convolution and the Fourier transform. The last set of basic properties of the
Fourier transform concern its interaction with convolution, which we now introduce. If
f, g are measurable functions on R, the convolution of f and g is the function
Z
(52) (f ∗ g)(x) := f (x − y)g(y) dy
R

defined at each x for which the integral makes sense. In particular, if f ∈ L∞ and
g ∈ L1 , then f ∗ g is defined on all of R. Observe, using the invariance of Lebesgue
measure with respect to x → −x and a simple change of variable,
Z
(53) f ∗ g(x) = g(x − y)f (y) dy = g ∗ f (x).
R

The next most basic facts about convolution are the following.
Proposition 7.12. If f, g ∈ L1 (R) and g ∈ L1 (R), then
(a) The function H : R2 → C defined by H(x, y) = f (x − y)g(y) is (jointly) measurable
and in L1 (R2 ) and ∥H∥1 = ∥f ∥1 ∥g∥p ;
(b) f ∗ g is defined for almost every x ∈ R;
(c) f ∗ g is measurable;
(d) f ∗ g ∈ L1 (R); and
(e) ∥f ∗ g∥1 ≤ ∥f ∥1 ∥g∥.

Proof. That H is jointly measurable as a function of x and y is left as an easy exercise.

By Tonelli
ZZ Z Z 
(54) |H(x, y)| dxdy = |g(y)| |f (x − y)| dx dy = ∥f ∥1 ∥g∥1 ,
R R

where we have used the translation invariance R of Lebesgue measureR in the second equal-
1 2
ity. Hence H is in L (R ). Thus by Fubini, RR|f (x − y)g(y)| dy = R |H(x, y)| dy is finite
for almost every x ∈ R and the function x 7→ R f (x − y)g(y) dy is measurable (and real)
for almost every x; that is f ∗ g is defined almost everywhere, in L1 and
Z
|f ∗ g(x)| ≤ |H(x, y)| dy.
R

Hence, by equation (54),

Z Z ZZ
∥f ∗ g∥1 = f (x − y)g(y) dy dx ≤ |H(x, y)| dydx = ∥f ∥1 ∥g∥1 . □
R R

We will also have use of the following result when studying the L2 theory of the
Fourier transform on R. It holds more generally with 2 replaced by 1 ≤ p ≤ ∞.
Proposition 7.13. If g ∈ L1 (R) and f ∈ L2 (R), then f ∗ g is defined almost everywhere
and in L2 (R) and ∥f ∗ g∥2 ≤ ∥f ∥1 ∥g∥2 .
104

Sketch of proof. First suppose f, g : X → [0, ∞). Thus the function f ∗ g : R → [0, ∞]
is defined everywhere. If also h ∈ L2 , then hf−t ∈ L1 for each t and thus, using Tonelli
and Cauchy-Schwartz,
Z Z Z Z Z Z
|f ∗ g(x)h(x)| dx = |h(x)| f (x − t)g(t)dt dx = g(t) |h(x)|f (x − t)dx dt
R R R R R
≤ ∥g∥1 ∥f ∥2 ∥h∥2 .

It follows that the function (f ∗ g)h is in L1 (R) for each h ∈ L2 . In particular, f ∗ g is

finite almost everywhere. Further, the mapping λ : L2 (R) → C defined by
Z
λ(h) = (f ∗ g)h
R

is continuous (with norm at most ∥g∥1 ∥f ∥2 ). Hence, by the Riesz Representation The-
orem (Theorem 5.28), there is an L2 (R) function ψ such that
Z Z
λ(h) = hψ = h (f ∗ g).
R R

Thus f ∗ g = ψ almost everywhere and so f ∗ g is in L2 and ∥f ∗ g∥2 = ∥λ∥ ≤ ∥g∥1 ∥f ∥2 .

Finally, dropping the assumption that f, g map into [0, ∞), from what is already
proved, |f | ∗ |g| ∈ L2 (R). Thus, since |f ∗ g| ≤ |f | ∗ |g| pointwise, it follows that f ∗ g ∈
L2 (R) and ∥f ∗ g∥2 ≤ ∥|f | ∗ |g|∥2 = ∥g∥1 ∥f ∥2 . □

END Monday 2025-03-31

Proposition 7.14. Let f, g, h ∈ L1 (R).

a) (Commutativity) f ∗ g = g ∗ f .
b) (Associativity) (f ∗ g) ∗ h = f ∗ (g ∗ h).
c) (Distributivity) (f + g) ∗ h = f ∗ g + f ∗ h.
d) (Scalar multiplication) If c ∈ C, then (cf ) ∗ g = c(f ∗ g).

Remark 7.15. The properties listed in Proposition 7.14, taken together, say L1 (R) with
the usual addition of functions and convolution as multiplication is a commutative ring.
(In fact it has even more structure, that of a Banach algebra, but we will not pursue
this direction in this course). □

We can now describe how convolution behaves under the Fourier transform.

Proposition 7.16 (Convolution becomes multiplication). Let f, g ∈ L1 (R). Then

∗ g(t) = fb(t)b
f[ g (t).
 105

Proof. By virtue of Proposition 7.12 the function G(x, y) = f (x − y)g(y)e−2πixt is in

L1 (R2 ) and thus we can use Fubini to compute f[ ∗ g(t):
Z Z 
f ∗ g(t) =
[ f (x − y)g(y) dy e−2πixt dx
R R
Z Z 
−2πixt
= g(y) f (x − y)e dx dy
R R
Z
= fb(t)e−2πiyt g(y) dy
R

= fb(t)b
g (t),
where we have used Proposition 7.4(b) to obtain the third equality. □

Given what we have proved so far, it follows that the Fourier transform is a ring
homomorphism from L1 (R) (with addition and convolution) to C0 (R) (with pointwise
addition and multiplication). We will see later that the Fourier transform is injective.
It turns out that it is not surjective, however. (See Problem 7.7.)
The following basic properties of convolution are immediate. Their proofs are left
as an exercise.

7.3. The Poisson kernel for the upper half plane. Let us fix the notation
1 a
(55) Pa (x) := .
π a + x2
2

Notice that P1 (x) is nonnegative and R P1 (x) dx = 1. Moreover, Pa (x) = a1 P1 ( xa ).

R

Remark 7.17. The function Pa (x) (viewed as a function of the two arguments (a and
x) is known as the Poisson kernel .
Viewing (x, a) ∈ R2 = C, the set UHP = {(x, a) : x ∈ R, a > 0} is the upper half
plane. Thus Pa (x) = P (x, a) determines a function P : UHP → R. It is not hard to
2 2
verify that P (x, a); that is ∂∂ 2Pa + ∂∂ 2Px = 0. □
Definition 7.18. An L1 approximate unit is a collection of functions ϕλ ∈ L1 (R) indexed
by λ > 0 such that:
(a) Rϕλ (t) ≥ 0 almost everywhere, for each λ,
(b) R ϕλ (t) dt = 1 for all λ, and
(c) For each fixed δ > 0, we have ∥1|t|>δ ϕλ ∥1 → 0 as λ → 0.
Proposition 7.19. The Poisson kernel {Pa }a>0 is an L1 approximate unit.
Theorem 7.20. If 1 ≤ p < ∞ and f ∈ Lp (R), then Pa ∗ f converges to f in Lp (R).

Only the cases p = 1, 2 of Theorem 7.20 are needed for the purposes here and that
is what is proved below. The proof uses the following lemma.
106

Lemma 7.21. Suppose ϕλ is an L1 (R) approximate unit and g : R → C is a bounded

measurable function. If g is continuous at a point x ∈ R, then

lim (g ∗ ϕλ )(x) = g(x).

λ→0

R
Proof. Using the trick g(x) = ϕλ (y)g(x) dy (item (a)),
R
Z
(g ∗ ϕλ )(x) − g(x) = (g(x − y) − g(x)) ϕλ (y) dy.
R

By positivity of ϕλ (item (b)),

Z
(56) |(g ∗ ϕλ )(x) − g(x)| ≤ |g(x − y) − g(x)|ϕλ (y) dy
R

To estimate the right-hand side, let ϵ > 0 be given. By the continuity of g at x, choose
δ > 0 so that |g(x − y) − g(x)| < ϵ when |y| < δ. We then split the integral in (56) into
two integrals, over the regions |y| ≤ δ and |y| > δ:
Z
|g(x − y) − g(x)|ϕλ (y) dy
R
Z Z
= |g(x − y) − g(x)|ϕλ (y) dy + |g(x − y) − g(x)|ϕλ (y) dy.
{|y|≤δ} {|y|>δ}

The first integrand is bounded by ϵϕλ , so

Z Z
|g(x − y) − g(x)|ϕλ (y) dy ≤ ϵ ϕλ (y) dy ≤ ϵ
{|y|≤δ} {|y|≤δ}
R
since R ϕλ (y) dy = 1. The second integrand is bounded by 2∥g∥∞ χ{|y|>δ} ϕλ (y), so goes
to 0 as λ → 0 by item (c) in the definition of approximate unit. □

END Wednesday 2025-04-02

Proof of Theorem 7.20. Let p = 1 or p = 2. By continuity of translation in Lp (R)

(Lemma 7.6), the function
Z
p
h(t) = ∥f − ft ∥p = |f (x) − f (x − t)|p dx
R

is continuous and in particular it is continuous at 0. It is also in L∞ (R) since h(t) =

∥f − ft ∥p ≤ 2∥f ∥p for all t. Consequently, as a function of a > 0, the function Pa ∗ h(0)
has limit h(0) at 0 by Lemma 7.21.
For a > 0 the function Pa is in L1 (R) and f is Lp (R) and thus, by Proposition 7.12 in
the L1 (R) case and Proposition 7.13 in the L2 (R) case, Pa ∗ f is measurable and defined
 107

almost everywere and in Lp (R). Using items (a) and (b) of the approximate identity
property of the Poisson kernel,
Z
|f (x) − Pa ∗ f (x)| = |f (x) − f (x − t)Pa (t) dt|
R
(57) Z
= | (f (x) − f (x − t)) Pa (t) dt|.
R
Let µ denote the measure Pa (t)dt. In particular µ(R) = 1 and therefore applying the
Cauchy-Schwarz inequality to (f (x) − f (x − t) and the function 1 in the Hilbert space
L2 (µ) when p = 2 and the obvious inequality when p = 1, equation (57) gives
Z  p1
p
(58) |f (x) − Pa ∗ f (x)| ≤ |f (x) − f (x − t)| Pa (t)dt
R

Hence by Tonelli and equation (58),

Z Z Z
p
|f (x) − Pa ∗ f (x)| dx ≤ |f (x) − f (x − t)|p Pa (t)dt dx
R R R
Z Z 
p
= |f (x) − f (x − t)| dx Pa (t) dt
ZR R

= ∥f (x) − f (x − t)∥pp Pa (t) dt

ZR
= h(t)Pa (t) dt
R
Z
= h(t)Pa (0 − t) dt = Pa ∗ h(0).
R
Therefore, ∥f − Pa ∗ f ∥p tends to 0 as a tends to 0+ and the proof is complete. □

7.4. Inversion and uniqueness. In this section we study the problem of recovering
f from fb. Loosely, the Fourier transform can be thought of as a resolution of f as a
superposition of sinusoidal functions e2πitx ; the value of fb(t) measures the “amplitude”
of f in the “frequency” t. This suggests that a formula like
Z
(59) f (x) = fb(t)e2πitx dt
R
1
ought to hold, at least if fb ∈ L . If we formally substitute the definition of fb and switch
the order of integration, we are confronted with
Z Z 
2πi(x−u)t
f (u) e dt du
R R

and the inner integral is not convergent, regardless of any assumption on fb. In fact (59)
does hold when fb ∈ L1 , but a more delicate argument is necessary. So, the goal of this
section will be to prove:
108

Theorem 7.22 (Fourier inversion, L1 case). If f and fb belong to L1 , then

Z
(60) f (x) = fb(t)e2πixt dt
R

for almost every x ∈ R.

Remark 7.23. If both f and fb are in L1 , then, by Theorem 7.22,

Z
f (x) = fb(−t)e−2πixt dt.
R

Thus f is the Fourier transform of the L1 function fb(−x) and therefore f ∈ C0 (R). By
symmetry fb ∈ C0 (R) too.
From f (x) = F fb(−x) it follows that F 4 = I. □

The inversion formula implies that L1 functions are determined by their Fourier
transforms.

Corollary 7.24. Suppose f, g ∈ L1 . If fb = gb, then f = g a.e.

Proof. From the inversion theorem, if f ∈ L1 and fb = 0, then f = 0. By the linearity of

− g = fb − gb, and the corollary follows.
the Fourier transform, f[ □

/ L1 ; and this is often the case;

So, in principle, f is fully determined by fb, even if fb ∈
e.g., for b > 0 the the function f = χ[−a,a] from example 7.2. To recover f from fb in
these cases, we turn to summability methods; in fact summability methods will already
be of use in proving the inversion theorem. The idea is this: suppose we have a divergent
integral
Z
h(t) dt
R

where the function h is, say, locally L , but not L1 . We might try to make sense of the
1

integral as
Z a
lim h(t) dt,
a→+∞ −a

effectively we have introduced the cutoff function ψa (t) := 1[−a,a] , which is positive,
integrable, and increases to 1 pointwise as a → ∞. Given any family of functions ψa
with these three propertes, we can consider the integrals
Z
h(t)ψa (t) dt.
R

The square cutoff χ[−a,a] has some undesirable properties; e.g., its Fourier transform is
not L1 (and not of constant sign).
 109

We will work first the smoother cutoff functions Qa (t) from example 7.3. Note that
example also computes the inverse Fourier transform of Qa for a > 0 as
Z
1 a
(61) Qa (t)e2πitx dt = = Pa (x),
R π a + x2
2

the Poisson kernel. Further, Qa (t) increases pointwise to the constant function 1 as a
tends to 0.
We are now able to compute the integral (59) modified by the cutoff function Qa (t):
Proposition 7.25. If f ∈ L1 , then for all a > 0 and all x ∈ R
Z
(f ∗ Pa )(x) = Qa (t)fb(t)e2πitx dt.
R
R
In particular, ga R
Qa (t)fb(t)e2πitx dt converges to f in L1 .

Remark 7.26. The last statement of Proposition 7.25 recovers f from fb, but only
in the L1 norm. The proposition says nothing about the pointwise covergence of the
reguralized integrals. In fact, it is true that the integrals (??) converge to f a.e., but
this requires a more delicate argument. In the proof below, it is shown that there is a
sequence an such that gan converges to f pointwise (almost everywhere.

Proof. For a > 0 and x fixed, let G(t, y)[a, x] = Qa (t)f (y)e−2πi(x−y)t . Observe that
∥f ∥1
Z
|G(t, y)[a, x]|dydt = ∥f ∥1 ∥Qa ∥1 = .
πa
Thus G[a, x] is L1 (R2 ) and hence we can apply Fubini (explaining the role of the cut-off
function Qa )
Z Z Z
2πix
Qa (t)f (t)e
b dt = Qa (t) f (y)e2πi(x−y)t dy dt
R
ZR ZR
= Qa (t) f (x − y)e2πiyt dy dt
ZR R
Z
= f (x − y) Qa (t)e2πiyt dt dy
R R

= (f ∗ Pa )(x),
where the second equality comes from a change of variable; the third from Fubini, and
the last from equation (61). □

Proof of Theorem 7.22. Assume f, fb ∈ L1 (R). Define

Z
g(x) = fb(t)e2πitx dt.
R

We are to show g = f a.e.

110

For a > 0 and x ∈ R, let

ha,x (t) = Qa (t)fb(t)e2πitx .
and for a > 0, let Z Z
ga (x) = Qa (t)fb(t)e2πitx dt = ha,x dt.
R R
From Proposition 7.25 ga (x) = f ∗ Pa (x). Fix a sequence an → 0. Since Qan increases
pointwise to the constant function 1, for fixed x, the sequence ha,x converges pointwise
to hx = fbte2πitx and at the same time |ha,x (t)| ≤ ∥hx (t)∥ = |fb(t)∥ for all t; that is, ha,x
is dominated by the L1 (R) function fb. Thus, by the dominated convergence theorem,
Z Z
lim gan (x) = lim hn,x dt = hx dt = g(x).
n n R R

On the other hand, gn = f ∗ Pan converges to f in L1 (R) by Theorem 7.20. By

Lemma 4.17, there is a subsequence (gnk )k of (gn ) that converges to f pointwise al-
most everywhere. Thus f = g almost everywhere as claimed. □
Remark 7.27. Observe that the above proof did not really use the explicit form of Pa ;
rather the point was that Qa (t) = {e−2aπ|t| }a>0 was a cutoff function (uniformly bounded
and converging pointwise to the constant function 1) whose Fourier transform {Pa } was
an L1 approximate unit. Any other cutoff function with this property could have been
used. □

END 2025-04-04

7.5. The L2 theory. In this section we study the Fourier transform on L2 . There is an
immediate problem, of course, since by Problem 4.5 L2 ̸⊆ L1 , so the integral (47) need
not be defined. However, we can observe that L1 ∩ L2 is dense in L2 (why?), and start
there.
Lemma 7.28. If f ∈ L1 ∩ L2 , then fb belongs to L2 and ∥fb∥2 = ∥f ∥2 .

Proof. Let fe(x) := f (−x). Since f, fe ∈ L1 , the convolution g = f ∗ fe is defined a.e. and
g ∈ L1 by Proposition 7.12. Now
Z Z
(62) g(x) = f (x − y)f (−y) dy = f (x + y)f (y) dy.
R R
2
Since f and f−x are both in L (R) equation (62) can be interpreted as g(x) = ⟨f−x , f ⟩L2
almost everywhere. By Lemma 7.6, the map x → f−x is continuous from R into L2 and
of course the vector f determines a continuous linear functional. Thus g is a continuous
function of x, and g(0) = ∥f ∥22 . By Cauchy-Schwarz again,
|g(x)| ≤ ∥f−x ∥2 ∥f ∥2 = ∥f ∥22 ,
so g is bounded.
 111

Let, as before Qa (t) = exp(−2aπ|t|). Since g ∈ L1 (R) we can apply Proposition 7.25
to compute Z
(g ∗ Pa )(0) = Qa (t)b
g (t) dt.
R
As g is continuous, by Lemma 7.21
Z
(63) ∥f ∥22 = g(0) = lim (g ∗ Pa )(0) = lim Qa (t)b
g (t) dt.
a→0 a→0 R

Let us compute the limit of this last integral in a different way. Recall that by
definition g = f ∗ fe, so by Propositions 7.16 and 7.4(d),
gb(t) = |fb(t)|2 .
Making this substitution in the integral in (63) and applying the monotone convergence
theorem (recall 0 < Qa (t) ≤ 1 converges pointwise increasing to 1 as a → 0+),
Z Z Z
2 2
∥f ∥2 = lim Qa (t)b g (t) dt = lim lim Qa (t)|f (t)| dt =
b |fb|2 .
a→0 R a→0 a→0 R R

Consequently fb ∈ L2 (R) and ∥fb∥2 = ∥f ∥2 . □

Theorem 7.29 (The Fourier transform on L2 ). There is a unique bounded linear trans-
formation F : L2 → L2 satisfying the following conditions:
(a) For all f ∈ L1 ∩ L2 , Ff = fb.
(b) (The Plancherel theorem) ∥Ff ∥2 = ∥f ∥2 for all f ∈ L2 .
(c) The mapping f → Ff is an Hilbert space isomorphism of L2 onto L2 .
(d) (The Parseval identity) ⟨f, g⟩ = ⟨Ff, Fg⟩ for all f, g ∈ L2 .

Remark: Note that when f, g ∈ L1 ∩ L2 , the Parseval identity reads

Z Z
f (x)g(x) dx = fb(t)b
g (t) dt.
R R

Proof of Theorem 7.29. By Lemma 7.28, the map f → fb is bounded linear transforma-
tion from a dense subspace of L2 into L2 . Thus, since the codomain L2 is complete,
by Proposition 1.34 the map f → fb has a unique bounded linear extension to a map
F : L2 → L2 . Hence item (a) holds and (b) follows since ∥f ∥2 = ∥Ff ∥2 on a dense set
(namely L1 ∩ L2 ). item (d) follows from item (b) by Polarization (Theorem 5.13). See
Problem 5.3(a). It remains to prove item (c); what we must show is that F is onto.
We show that F has dense range; combined with the fact that F is an isometry, it
follows that F is in fact onto. (The proof of this last assertion is left as an exercise). Let
M denote the set of all functions g ∈ L2 such that g = fb for some f ∈ L1 ∩ L2 . Clearly
the range of F contains M , so it will suffice to prove that M is dense, or equivalently,
that M ⊥ = {0}.
112

Recall the cutoff functions Qa (x) = e−2aπ|x| , a > 0 introduced in equation (48).
The functions e2πibx e−2aπ|x| belong to L1 ∩ L2 for all a > 0 and b ∈ R, so their Fourier
transforms Z
Pa (t − b) = e2πibx Qa (x) e−2πitx dx
R
belong to M . So, let h ∈ M ⊥ be given and let H(x) = h(−x). Thus,
Z Z
(Pa ∗ H)(−b) = Pa (−b − t)h(−t) dt = Pa (t − b)h(t) dt = 0
R R

for all b. Theorem 7.20 implies Pa ∗ H converges to H in L2 . Hence h = 0 and conse-

quently M is dense in L2 and the proof is finished. □
Theorem 7.30 (L2 inversion). Let f ∈ L2 . Define
Z N Z N
−2πixt
ϕN (t) = f (x)e dx, ψN (t) = (Ff )(t)e2πixt dt.
−N −N
Then ∥ϕN − Ff ∥2 → 0 and ∥ψN (t) − f ∥2 → 0 as N → ∞.

Proof. Let fN := 1[−N,N ] f . Then fN ∈ L1 ∩ L2 , and ϕN = fc N . An application of

dominated convergence shows that (fN )N converges to f in the L2 norm. Hence, (ϕN =
2
N = FfN )N converges in L to Ff by Theorem 7.29.
fc
The statement for ψN is proved by similar methods and is left as an exercise (Prob-
lem 7.9). □
Remark 7.31. It is important to note that, for a general function f ∈ L2 , its Fourier
transform is defined only as an element of L2 . In particular it is defined only a.e., and
cannot be evaluated at points. It is customary to write fb for Ff when f ∈ L2 , with the
understanding that the integral definition is only valid when f ∈ L1 ∩ L2 .

END Monday 2025-04-07

7.6. Problems.
Problem 7.1. Prove Proposition 7.4
Problem 7.2. Complete the proof of Lemma 7.6.
Problem 7.3. Prove, if E ⊆ [0, 1] has positive Lebesgue measure, then the set
E − E = {x − y : x, y ∈ E}
contains an interval centered at the origin. (Hint: let −E = {−x : x ∈ E} consider the
function h(x) = 1−E ∗ 1E .)
R
Problem 7.4. Suppose ϕ is an unsigned L1 function with ϕ = 1, and let ϕλ (x) =
1
ϕ λx .


λ

a) Prove {ϕλ }λ>0 is an L1 approximate unit.

 113

b) Give a simpler proof of Lemma 7.21 by making a change of variables in equa-

tion (56).
Problem 7.5. a) Prove, if f ∈ Cc1 (R) and g is a compactly supported L1 function,
then f ∗ g is C 1 with compact support. (Hint: justify differentiation under the
integral sign.)
b) By induction, conclude that if f ∈ Cc∞ (R) and g ∈ L1 is compactly supported,
then f ∗ g ∈ Cc∞ (R).
c) Conclude that Cc∞ (R) is dense in Lp for all 1 ≤ p < ∞. (Suggestion: Construct
a C ∞ approximate identity with compact support.)
Problem 7.6. Compute the integral in Lemma 7.3.
Problem 7.7. This problem gives a proof that the Fourier transformb: L1 → C0 (R) is
not surjective.
a) Draw a picture of hn := 1[−n,n] ∗ 1[−1,1] and determine its C0 (R) norm.
b) Show that hn is, up to a multiplicative constant independent of n, the Fourier
transform of the L1 function
sin 2πx sin 2πnx
fn := .
x2
(Hint: you can compute integrals, or use the L1 inversion theorem.)
c) Show that ∥fn ∥1 → ∞ as n → ∞. Conclude that the Fourier transform is not
surjective. (Hint: if it were surjective... .)
Problem 7.8. Suppose that f ∈ L1 , f is differentiable a.e., f ′ ∈ L1 , and f (x) =
Rx ′
−∞
f (y) dy for a.e. x ∈ R. Prove fb′ = 2πitfb(t).
Problem 7.9. Complete the proof of Theorem 7.30.
Problem 7.10. Let φλ be an L1 (T) approximate unit. Prove, if f ∈ C(T), then
f ∗ φλ → f uniformly as λ → 0.
Problem 7.11. State and prove an analog of Proposition 7.4 for Fourier series.
Problem 7.12. Show if f : R → R is twice continuously differentiable and has compact
support, then fb ∈ L1 (R). Now show the Fourier transform F : L1 (R) → C0 (R) has
dense range. (It is not onto by Problem 7.7.)
 Index
B(X , Y), 14 bidual, 29
B(x, r), 39 bilinear form, 60
B X (x, r), 43 bounded, 13
C(X), 6 bounded linear transformation, 13
C0 (X), 6
Cb (X, Y), 6 category, 41
F (T, V ), 5 Cauchy–Schwarz inequality, 57
F ◦ , 39 chain, 25
Fb (T, X ), 5 character, 98
L1 approximate unit, 105 closed transformation, 47
Lp (µ), 53 complete orthonormal set, 69
M ≤ H, 58 completion, 31
M ⊕ N , 64 complex-valued measurable function, 7
PM , 67 conditional expectations, 91
M(M ), 82 conjugate index, 51
ℓ∞ -norm, 3 converges absolutely, 2, 71
ℓp -norms, 3 converges as a net, 73
dν convolution, 103
dµ , 90
F-measure, 81 cutoff function, 108
Lp (µ), 49
µf , 83 dense, 10, 39
⊥, 59 double dual, 29
ρ ≪ µ, 90 dual exponent, 51
|ρ|, 83 dual space, 20
f dµ, 83
fy , 99 equivalent norms, 1
(Hilbert space) adjoint, 66 essentially bounded, 50
(orthogonal) dimension, 77
finite positive measure, 81
(orthonormal) basis, 69
first category, 41
converges weakly to h ∈ H, 78 Fourier transform, 98
convex, 61
isometric, 79 Gaussian, 102
maximal, 69 graph, 47
shift operator, 80 greedy algorithm, 88
unitary, 79
Hahn Decomposition Theorem, 87
absolutely, 90 Hamel basis, 24
absolutely integrable, 8 hermitian, 66
adjoint, 65 Hilbert space, 58
algebraic direct sum, 10 Hilbert space basis, 69
algebraic quotient, 11 horizontal truncation, 57
almost convergent, 49
inner product, 57
backward shift, 49 integrable, 8
Banach algebra, 15, 104 interior, 39
Banach Limit, 49 invertible linear transformation, 15
Banach limit, 30, 38 isometrically isomorphic normed spaces, 15
Banach space, 1 isometry, 15
Banach space of functions, 48 isomorphic normed spaces, (boundedly), 15
114
 115

Jordan Decomposition Theorem, 89 strictly convex, 61

subspace, 10
Lebesgue space, 53 summability methods, 108
linear functional, 20 super-cauhcy, 2
linear manifold, 10 support set, 89
linear transformation, bounded, 13
Lipschitz continuous, 13, 17 total variation, 83
total variation norm, 18
meager, 41 totally null, 87
Minkowski functional, 24 totally ordered, 25
mutually singular, 89 totally positive, 87
norm topology, 1 unconditionally convergent, 71
nowhere dense, 41 unit ball, 3
unit sphere, 17
open mapping, 43 upper bound, 25
open mapping theorem, 43 upper half plane, 105
operator norm, 14
orthogonal, 59 vanishes at infinity, 6
orthogonal complement, 59 vertical truncation, 57
orthogonal direct sum, 64
orthogonal projection of H on M , 67 weak topology, 77
orthonormal, 59, 69 weakly bounded, 47

Zorn’s Lemma, 25
Parseval’s equality, 74
partial order, 25
partial sums, 2
partially ordered set, 25
Poisson kernel, 105
poset, 25
positive function, 22
positive linear functional, 22
positive semi-defininite, 60
projection, 67

quotient norm, 12

Radon-Nikodym derivative, 90
reflexive, 29
regular, 22
Riemann-Lebesgue Lemma, 100

Schauder basis, 24
second category, 41
self-adjoint, 66
separable, 20
separable Hilbert space, 77
series converges, 2
sesquilinear form, 60
shift operator, 16
signed measure, 81
signed measures, 22
smallest closed subspace containing, 63