4 Linear operators and linear functionals

The next section is devoted to studying linear operators between normed

spaces.

Definition 4.1. Let V and W be normed spaces over a field F. We say that
T : V → W is a linear operator if T is linear (that is, T (x+y) = T (x)+T (y)
for all x, y ∈ V and T (λx) = λT (x) for all x ∈ V and λ ∈ F).

Definition 4.2. Let V and W be normed spaces. We say that a linear

operator T : V → W is bounded if there exists M > 0 such that kxk ≤ 1
implies kT (x)k ≤ M . In such cases we define the operator norm by

kT k = sup{kT (x)k : x ∈ V and kxk ≤ 1}.

Example. 1. Not all operators are bounded. Let V = C([0, 1]) with
R 1/2
1 2 dx
respect to the norm kf k = 0 |f (x)| . Consider the linear
operator T : V → C given by T (f ) = f (0). We can see that this
operator
R 1 is unbounded by defining functions where fn where fn (0) = n
but 0 |fn |2 dx = 1.

2. On the otherhand if we define T : `∞ → `∞ by T ({xn }n∈N ) = T ({yn })

where yn = xn+1 we can see that we have a bounded operator. Since
if kxk∞ ≤ 1 we have that kT (x)k ≤ 1. Moreover if we take x =
(0, 1, 0, 0, . . .) we can see that kxk∞ = 1 and kT (x)k∞ = 1 and so
kT k = 1.

Our first key result related bounded operators to continuous operators.

Theorem 4.3. Let V and W be normed spaces and T : V → W a linear

operator. Then the following statements are equivalent

1. T is bounded

2. T is continuous

3. T is continuous at 0.

Proof. We first show 1 implies 2. Suppose that T is bounded and let x ∈ V .

Let  > 0 and fix δ = /kT k. For y such that kx − yk ≤ δ we have that

kT (x − y)k ≤ kT kkx − yk ≤ .

The fact that 2 implies 3 is obvious. Finally suppose that T is continuous

at 0 and choose δ such that if y ∈ V and kyk ≤ δ, then kT (x)k ≤ 1. Now
if kxk ≤ 1, then kδxk ≤ δ and so kT (δx)k ≤ 1. Thus kT (x)k ≤ δ −1 . So we
can conclude that T is bounded.

1
Definition 4.4. Let V and W be normed spaces. We let B(V, W ) be the
space of Bounded linear operators T : V → W .

Theorem 4.5. Let V be a normed space and W a Banach space. The space
B(V, W ) is a Banach space with respect to the operator norm.

Proof. It’s a routine exercise to show that B(V, W ) is a normed space. Now
let {Tn }n∈N be a Cauchy sequence in B(V, W ). Let x ∈ V and consider the
sequence Tn (x). Let  > 0 and choose N such that for all n, m ≥ N we have
that kTn − Tm k ≤ . We then have that

kTn (x) − Tm (x)k = k(Tn − Tm )(x)k ≤ kxkkTn − Tm k ≤ kxk.

So Tn (x) is a Cauchy sequence in W , and since W is Banach, we may define

T (x) = limn→∞ Tn (x).
We now need to show that T ∈ B(V, W ). Well T is linear since each
Tn is linear. So let  > 0 and choose N ∈ N such that for n, m ≥ N we
have that kTn − Tm k ≤ . Take x ∈ V such that kxk ≤ 1 and note that for
n, m ≥ N , we have

kT (x) − Tn (x)k ≤ kT (x) − Tm (x)k + kTm (x) − Tn (x)k ≤ kT (x) − Tm (x)k + .

Since this holds for all m ≥ N and limm→∞ kT (x) − Tm (x)k = 0 we have
that kTn (x) − T (x)k ≤  for all n ≥ N . Thus kT (x)k ≤ kTN (x)k +  and so
T is bounded. We can also conclude that for n ≥ N , kTn − T k ≤  and so
Tn converge to T in B(V, W ). Hence, B(V, W ) is a Banach space.

Unless it is otherwise specified when we talk about B(V, W ) we mean

the normed space where the norm is the operator norm.

Linear functionals and Dual spaces

We now look at a special class of linear operators whose range is the field F.

Definition 4.6. If V is a normed space over F and T : V → F is a linear

operator, then we call T a linear functional on V .

Definition 4.7. Let V be a normed space over F. We denote B(V, F) = V ∗

(recall that B(V, F) is the the space of bounded operators for V to F). We
can V ∗ the dual space of V .

Corollary 4.8 (Corollary to Theorem 4.5). For any normed space V we

have that V ∗ is a Banach space.

In the case of `p space where p 6= ∞ we can now determine exactly what

the dual space is.

2
Example. Consider the space `1 . We can find a norm preserving linear
bijection (isomorphism for Banach spaces) between theP dual space (`1 )∗ and
`∞ . For x = (xn )n∈N define fx : `1 → F by fx (y) = ∞ n=1 xn yn . We can
see that fx is linear and for y ∈ `1 , |fx (y)| ≤ kxk∞ kyk1 , so fx is a bounded
linear functional. We can define T : `∞ → (`1 )∗ by T (x) = fx and easily see
that it is a injective linear map.
To show that T is surjective let f ∈ (`1 )∗ and consider the elements en ∈
`1 with nth term 1 and all other terms 0. We have that ∞
1
Pck = (f (en ))n≥1 ∈ `
and for any y = (yn )n∈N ∈ ` note that y = limk→∞ n=1 yn en . So by the
continuity of f ,
k
X ∞
X
f (y) = lim f (en )yn = f (en )yn = fc (y).
n→∞
n=1 n=1

Finally we need to show T preserves norms. To do this fix x ∈ `∞ and

note we have already shown that kfx k ≤ kxk∞ . Now let  > 0 and n such
that |xn | ≥ kxk∞ − . We have that

|fx (en )| ≥ kxk∞ − 

and since ken k1 = 1 we can conclude that kfx k = kxk∞ . Thus for all x ∈ `∞ ,
kT (x)k = kxk∞ .

For Hilbert spaces, it turns out that Hilbert spaces can be identified with
their own dual spaces.

Theorem 4.9 (Riesz-Frechet). Let H be a Hilbert space. For any f ∈ H ∗

there exists an unique y ∈ H such that f (x) = hx, yi for all x ∈ H and
kf k = kyk.

Proof. The case f = 0 is obvious. Let’s assume that f 6= 0.

We first address the existence of such a y. Let f ∈ H 0 . Let

A = ker(f ) = {x ∈ H : f (x) = 0}

and note that A is a closed subspace of H and A 6= H. So A⊥ is a closed

subspace of H which does not just contain 0 (since H = A ⊕ A⊥ ). Thus
we can now pick z ∈ A⊥ such that f (z) = 1. For x ∈ H we can write
x = (x − f (x)z) + f (x)z and we’ll have that x − f (x)z ∈ A and f (x)z ∈ A⊥ .
Thus for all x ∈ H we have

hx, zi = hx − f (x), zi + hf (x)z, zi = f (x)kzk2 .

We can now let y = z/kzk2 to get for all x ∈ H f (x) = hx, yi.
To show that such a y is uniquely determined we simply note that if
hx, yi = hx, zi for all x ∈ H, then hx, y − zi = 0 for all x ∈ H and so y = z.

3
Finally we need to show kf k = kyk which is again trivial for y = 0. For
y 6= 0 we have by Cauchy-Schwarz that

|hx, yi| ≤ kxkkyk

and so kf k ≤ kyk. Moreover

|f (y/kyk)| = |hy/kyk, yi| = kyk

and so kf k = y.

For general Banach spaces or normed spaces it is much harder to char-

acterise the dual space (even `∞ is more difficult than these examples).
Later on in the unit you will see the Hahn-Banach theorem which gives a
very useful way of defining linear functionals (for instance it gives a way of
showing that there are nonzero functional on any non-trivial normed space,
surprisingly this is a non-trivial fact).

Inverse operators, adjoints and unitary operators

We now turn to the question of when operators have inverses and then move
on to the notion of adjoint operators and finally the notion of a unitary
operator. In finite dimension vector spaces we can write a linear operator
as a matrix and use the determinant to determine whether it is invertible
or not. In general normed spaces things are quite a bit more complicated.

Definition 4.10. Let V be a normed space. Let Iv ∈ B(V, V ) be the identity

operator on V where
IV (x) = x for all x ∈ V.

Definition 4.11. Let V and W be normed space. We say that T ∈ B(V, W )

is invertible if there exists S ∈ B(W, V ) such that

T S = IW and ST = IV

and we call S = T −1 the inverse of T .

If V is a finite dimensional vector space and T ∈ B(V ) = B(V, V ) then

the following are all equivalent to T being invertible

1. T is injective

2. T is surjective

3. There exists S ∈ B(V ) such that ST = IV .

4. There exists T ∈ B(V ) such that T S = IV .

4
For normed spaces none of these conditions all equivalent to an operator
being invertible.

Example. Let V = `∞ and T : `∞ → `∞ be the shift map defined by

T ((x1 , x2 , . . . , )) = (x2 , x3 , . . .))

and S : `∞ → `∞ be defined by

S((x1 , x2 , . . .)) = (0, x1 , x2 , . . .).

We can see that T is surjective, but not injective and that T S = I but not
ST = I.

However for Banach spaces there is a criteria which can be used to find
examples of invertible operators.

Theorem 4.12. Let V be a Banach space and A ∈ B(V ) with kAk < 1.
We have that I − A is invertible and
∞
X
−1
(I − A) = An .
n=0

(note that we take A0 = I.)

Proof. To start we show that ∞

P n
Pk n
n=0 A ∈ B(V ). Let Ak = n=0 A . We
will show that Ak is a Cauchy sequence in B(V ). Let n, m ≥ N and let
x ∈ V such that kxk ≤ 1. We have that
m m ∞ ∞
X X X X kAkn+1
kAn (x)−Am (x)k = k Ak (x)k ≤ kAk k ≤ kAk k ≤ kAkk .
1 − kAk
k=n+1 k=n+1 k=n+1 k=n+1

So we can see that Ak is a Cauchy sequence and so we can let T = limk→∞ Ak

and note that for all x ∈ V T (x) = limk→∞ Ak (x). We now need to show
T (I − A) = (I − A)T = I. Let x ∈ V we have that

T (I −A)(x) = T (x)−T A(x) = lim Ak (x)−Ak A(x) = lim x−Ak+1 (x) = x

k→∞ k→∞

and we can also show (I − A)T (x) = x.

Corollary 4.13. Let V be a Banach space. The set of invertible operators

is open in B(V ).

Proof. Note that if A, B ∈ B(V ) are invertible then AB is invertible and

that I is invertible so there is always at least one invertible element in B(V ).
We need to show that for any invertible operator A there exists δ > 0 such

5
that if S ∈ B(V ) and kA − Sk < δ then S is invertible. Let A ∈ B(V ) be
invertible with inverse A−1 . Consider

X = {T ∈ B(V ) : kA − T k < (kA−1 k)−1 }.

For S ∈ X we have that

k(S − A)A−1 k ≤ kS − AkkA−1 k < 1.

Thus by Theorem 4.12 we know that I + (S − A)A−1 = SA−1 is invertible.

Thus S = SA−1 A is invertible.

Adjoint, self-adjoint and unitary operators

Definition 4.14. Let V1 and V2 be normed spaces. For T ∈ B(V1 , V2 ), we
define the map T ∗ : V2∗ → V1∗ by f 7→ f ◦ T , f ∈ V2∗ .

Theorem 4.15. The map T ∗ ∈ B(V2∗ , V1∗ ) and kT ∗ k ≤ kT k.

Proof. We first show that T ∗ is linear. For f, g ∈ V2∗ and µ, ν ∈ F, the

functional T ∗ (µf + νg) is defined by x 7→ (µf + νg)(T x) = µf (T x) +
νg(T x) = (µT ∗ (f ) + νT ∗ (g))(x). Hence, T ∗ is linear. Moreover, for f ∈ V2∗
and x ∈ V1 with kxk ≤ 1, |T ∗ (f )(x)| = |f (T x)| ≤ kf kkT xk ≤ kf kkT k. This
shows that kT ∗ k ≤ kT k.

In the context of Hilbert spaces, the adjoint operator can be written

more explicitly using the Riesz-Frechet theorem. Let H1 and H2 be Hilbert
spaces and T ∈ B(H1 , H2 ). For each y ∈ H2 we can define a linear functional
f : H1 → F by fy (x) = hT (x), yi where the inner product is taken in H2 . By
Cauchy-Schwartz it follows that fy ∈ H1∗ and by the Riesz-Frechet Theorem
there exists a unique z ∈ H1 such that hT (x), yi = fy (x) = hx, zi. This
defines the adjoint operator T ∗ : H2 → H1 .

Example. Let T : `2 → `2 be the left shift map T ((x1 , x2 , x3 , . . .)) =

(x2 , x3 , . . .). For y ∈ `2 we have that
∞
X ∞
X
hT (x), yi = xn+1 yn = xn zn
n=1 n=1

where zn = (0, y1 , y2 , . . .). So T ∗ : `2 → `2 is the right shift where

T ∗ (x1 , x2 , . . .) = (x2 , x3 , . . .).

Theorem 4.16. Let H1 and H2 be Hilbert spaces and T ∈ B(H1 , H2 ). We

have that

1. (T ∗ )∗ = T

6
 2. kT ∗ k = kT k

Proof. Let x ∈ H2 and y ∈ H1 . We have that

hx, T (y)i = hT (y), xi = hy, T ∗ (x)i = hT ∗ (x), yi = hx, (T ∗ )∗ (y)i.

This holds for all x ∈ H2 and so T = (T ∗ )∗ . Since kT ∗ k ≤ kT k, we can use

this to deduce that kT k = kT ∗ k.

See Theorem 3.9-4 in Introductory Functional Analysis with Applica-

tions by Kreyszig for more properties of the adjoint.

Definition 4.17. Let H be a Hilbert space and let T ∈ B(H). We call

T self-adjoint (often called Hermitian) if and only if T = T ∗ . We call T
unitary if T ∗ = T −1 .

Example. Take H = `2 , a = (an )n≥1 ∈ `∞ and T : `2 → `2 be given by

T (x) = (an x)n≥1 . For x, y ∈ `2 we have that
∞
X ∞
X
hT (x), yi = (an xn )yn = xn an yn .
n=1 n=1

So we can see that T is self-adjoint if and only if all an ∈ R, and T is unitary

if and only if |an | = 1 for all n.