128 3 Dual Spaces, Transposes and Adjoints

Next, let K := C. Treating X as a linear space over R, we may find, as

above, a real-linear continuous functional u from X to R such that u(x) > 0
for all x ∈ E. Define f (x) := u(x) − iu(ix) for x ∈ X. As we have seen in
Lemma 3.3, f is a continuous complex-linear functional from X to C such
that Re f (x) = u(x) for all x ∈ X, and so Re f (x) > 0 for all x ∈ E. #
$
If, in addition to the hypotheses of the above proposition, there is a sub-
space Y of X such that Y ∩ E = ∅, then there is a continuous linear functional
f on X such that f (y) = 0 for all y ∈ Y and Re f (x) > 0 for all x ∈ E. See
Exercise 11 of this chapter.
The following result is often considered as a companion result of the Hahn-
Banach extension theorem
Theorem 3.11. (Hahn-Banach separation theorem) Let X be a normed
space over K. Let E1 and E2 be disjoint convex subsets of X with E1 open.
Then there are a continuous linear functional f on X and t ∈ R such that

Re f (x1 ) < t ≤ Re f (x2 ) for all x1 ∈ E1 and x2 ∈ E2 .

Proof. Let E := E2 − E1 = {x2 − x1 : x1 ∈ E1 and x2 ∈ E2 }. Since E1 ∩ E2 =

∅, we see that 0 (∈ E. Also,! it is easy to see that E is a convex subset of
X. Further, since E = x2 ∈E2 (x2 − E1 ), and the set x2 − E1 is open for
every x2 ∈ E2 , it follows that E is an open subset of X. By Proposition 3.10,
there is a continuous linear functional f on X such that Re f (x) > 0 for all
x ∈ E, that is, Re f (x1 ) < Re f (x2 ) for all x1 ∈ E1 and x2 ∈ E2 . Since E1
and E2 are convex subsets of X, it follows that Re f (E1 ) and Re f (E2 ) are
convex subsets of R, that is, they are intervals in R. Also, since E1 is an open
subset of X, Re f (E1 ) is an open interval in R. Hence there is t ∈ R such that
Re f (x1 ) < t ≤ Re f (x2 ) for all x1 ∈ E1 and x2 ∈ E2 . #
$
Let f be a linear functional on a linear space X, and let t ∈ R. Then the
subset {x ∈ X : Re f (x) = t} of X is called a real hyperplane in X. If X is
a normed space and f is continuous, then this real hyperplane is closed in X.
The Hahn-Banach separation theorem says that

E1 ⊂ {x ∈ X : Re f (x) < t} and E2 ⊂ {x ∈ X : Re f (x) ≥ t}.

In this sense, the closed real hyperplane ‘separates’ the disjoint convex subsets
E1 and E2 of X, of which E1 is open.

3.2 Dual Spaces and Riesz Representation

Let X be a normed space over K. The space BL(X, K) of all bounded linear
functionals on X is called the dual space of X, and it is denoted by X " .
Let X be finite dimensional. Then all linear maps defined on X are con-
tinuous. Let x1 , . . . , xn constitute a basis for X. Then the linear functionals
3.2 Dual Spaces and Riesz Representation 129

f1 , . . . , fn on X, introduced in Section 0.2, which satisfy fj (xi ) = δi,j , i, j =

1, . . . , n are in X " . Since the subset {f1 , . . . , fn } is linearly independent, and
since f = f (x1 )f1 + · · · + f (xn )fn for every linear functional f on X, we
see that f1 , . . . , fn constitute a basis for X " ; it is called the dual basis of
X " relative to x1 , . . . , xn . In particular, the dimension of X " is equal to the
dimension of X.
Let x"1 , x"2 ∈ X " . By the very definition of a function, x"1 = x"2 if and only if
x1 (x) = x"2 (x) for all x ∈ X. Next, let x1 , x2 ∈ X. If x1 (= x2 , that is, x1 −x2 (=
"

0, then by Proposition 3.7 (i), there is x" ∈ X " such that x" (x1 − x2 ) (= 0, that
is, x" (x1 ) (= x" (x2 ). Thus x1 = x2 if and only if x" (x1 ) = x" (x2 ) for all x" ∈ X " .
This interchangeability between x ∈ X and x" ∈ X " explains the nomenclature
‘dual space’ of X " .
The operator norm on X " is given by

+x" + := sup{|x" (x)| : x ∈ X and +x+ ≤ 1} for x" ∈ X " .

Proposition 3.12. Let X be a normed space. Then

+x+ = sup{|x" (x)| : x" ∈ X " and +x" + ≤ 1} for all x ∈ X.

Further, if Y is a normed space, and F ∈ BL(X, Y ), then

+F + = sup{|y " (F (x))| : x ∈ X, +x+ ≤ 1 and y " ∈ Y " , +y " + ≤ 1}.

Proof. Let x ∈ X. By the basic inequality (Theorem 2.10 (i)), |x" (x)| ≤
+x" + +x+ for all x" ∈ X " . Also, by a consequence of the Hahn-Banach ex-
tension theorem (Proposition 3.7 (i)), there is x" ∈ X " such that x" (x) = +x+
and +x" + ≤ 1. Hence desired the expression for +x+ follows.
Next, let F ∈ BL(X, Y ). Then for x ∈ X, it follows that +F (x)+ =
sup{|y " (F (x))| : y " ∈ Y " and +y " + ≤ 1}. Hence

+F + = sup{+F (x)+ : x ∈ X and +x+ ≤ 1}

= sup{|y " (F (x))| : x ∈ X, +x+ ≤ 1 and y " ∈ Y " , +y " + ≤ 1},

as desired. #
$

We shall find the dual spaces of some important normed spaces, and also indi-
cate how the dual spaces of some others can be found. This task is important
because if we know the dual space of a normed space, we can rephrase ques-
tions about the given space in a way that can throw light from a different
angle, and possibly lead us to answers. For example, in Remark 3.8 (i), we
have seen that a criterion for the approximation of an element of a normed
space by elements belonging to a fixed subspace of that normed space can be
given in terms of the elements of the dual space that vanish on that subspace.
The ‘duality’ between X and X " does not mean that results about X and
X are always symmetric in nature. For instance, since K is a (one dimensional)
"
130 3 Dual Spaces, Transposes and Adjoints

Banach space, Theorem 2.10 (iii) shows that X " = BL(X, K) is a Banach
space, even if X is not a Banach space. Actually, the consideration of the dual
space of a dual space gives us a neat way of constructing a completion of a
normed space, and of an inner product space, as we now show.
Let X be a normed space, and let x ∈ X. Define jx : X " → K by
jx (x" ) := x" (x) for x" ∈ X " .
Clearly, jx is linear, and +jx + := sup{|jx (x" )| : x" ∈ X " and +x" + ≤ 1} = +x+
by Proposition 3.7 (i). Hence jx belongs to the dual space (X " )" of X " . We shall
denote the second dual space (X " )" of the normed space X by X "" . Define
J : X → X "" by J(x) := jx for x ∈ X. Then J is linear and +J(x)+ = +x+
for all x ∈ X, that is, J is a linear isometry from X into X "" . It is called the
canonical embedding of X into X "" . If J(X) = X "" , then we say that the
normed space X is reflexive. We shall show that a Hilbert space is reflexive.
Let Xc denote the closure of J(X) in X "" . (If X is a Banach space, then
J(X) is a closed subspace of X "" , and so Xc = J(X).) Since X "" := BL(X " , K)
is a Banach space, and since Xc is closed in X "" , we see that Xc is a Banach
space, and J : X → Xc is a linear isometry such that J(X) is dense in Xc .
The Banach space Xc is called the completion of the normed space X. It is
unique in the following sense. If X1 is a Banach space, and J1 : X → X1 is a
linear isometry such that J1 (X) is dense in X1 , then there is a linear isometry
Φ from Xc onto X1 . This can be seen as follows.

J ! ! Xc
X! J(X)
!! ""
!! "
!! Φ0 ""
J1 !!" # $""" Φ
X1
Define Φ0 : J(X) → X1 by
Φ0 (J(x)) := J1 (x) for x ∈ X.
Then Φ0 ∈ BL(J(X), X1 ), where J(X) is dense in Xc , and X1 is a Banach
space. By Theorem 2.10 (iv), there is Φ ∈ BL(Xc , X1 ) such that Φ(J(x)) =
Φ0 (J(x)) for all x ∈ X. Since +Φ0 (J(x))+ = +J1 (x)+ = +x+ = +J(x)+ for all
x ∈ X, it follows that Φ is also a linear isometry. Also, since Φ(Xc ) is both
closed as well as dense in X1 , we obtain Φ(Xc ) = X1 .
In case (X, -·, ·.) is an inner product space, we introduce an inner product
on the completion Xc of X as follows. For x"" and y "" in Xc , let J(xn ) → x""
and J(yn ) → y "" in X "" , and let -x"" , y "" .c := limn→∞ -xn , yn .. It is easy to
see that -· , ·.c an inner product on Xc , and -J(x), J(y).c = -x, y. for all
x, y ∈ X. Thus the completion of an inner product space is a Hilbert space.
For example, let p ∈ {1, 2}, and let X := C([a, b]) with the norm given by
"# b $1/p
p
+x+p := |x(t)| dt for x ∈ X.
a
3.2 Dual Spaces and Riesz Representation 131

Since X is dense in Lp ([a, b]) by Proposition 0.25, and since Lp ([a, b]) is com-
plete by Proposition 1.23 (iv), Lp ([a, b]) can be identified with the completion
of X. Thus if p := 1, the completion of the normed space X is the Banach
space L1 ([a, b]), and if p := 2, the completion of the inner product space X is
the Hilbert space L2 ([a, b]).
By considering the completion of a normed space (or of an inner product
space), we can extend to incomplete normed spaces (or to incomplete inner
product spaces), results that are initially proved only for Banach spaces (or
only for Hilbert spaces). See Exercises 16 and 17 of this chapter.
The canonical embedding J : X → X "" introduced earlier is useful in a
variety of ways. We give an instance below.

Theorem 3.13. (i) (Resonance theorem) Let X be a normed space and

E ⊂ X. Then E is a bounded subset of X if and only if x" (E) is a bounded
subset of K for every x" ∈ X " .
(ii) Let X and Y be normed spaces, and let F : X → Y be a linear map. Then
F is continuous if and only if y " ◦F is continuous for every y " ∈ Y " .

Proof. (i) Suppose E is a bounded subset of X. Then there is α > 0 such that
!
+x+ ≤ α for all x ∈ E. If x" ∈ X , then |x" (x)| ≤ +x" + +x+ ≤ +x" + α for all
x ∈ E, and so x" (E) is a bounded subset of K.
Conversely, suppose x" (E) is a bounded subset of K for every x" in X " . Let
J : X → X "" be the canonical embedding of X into X "" . Then {J(x) : x ∈ E}
is a subset of X "" := BL(X " , K). Now X " is a Banach space, and for each
fixed x" ∈ X " , the set {J(x)(x" ) : x ∈ E} = x" (E) is bounded. By the uniform
boundedness principle (Theorem 2.20), J(E) is a bounded subset of X "" . Since
+J(x)+ = +x+ for each x ∈ X, it follows that E is a bounded subset of X.
(ii) Let the linear map F : X → Y be continuous. Since every y " ∈ Y " is
continuous, we see that y " ◦F is continuous. Conversely, let y " ◦F be continuous
for every y " ∈ Y " . Let E := {F (x) : x ∈ X and +x+ ≤ 1} ⊂ Y . Then y " (E) is
bounded for every y " ∈ Y " by Proposition 2.2. By (i) above, E is a bounded
subset of Y . Thus the linear map F is continuous by Proposition 2.2. #
$

The resonance theorem proved above says that for a subset E of a normed
space X, the boundedness of E in X is the same thing as the boundedness of
x" (E) for every x" ∈ X " . However, for a sequence (xn ) in X, the convergence
of (x" (xn )) for every x" ∈ X " may not imply the convergence of (xn ) in X.
For example, if X := #2 , then we conclude from Example 3.19 (i) that (x" (en ))
" "
√ x ∈ X . But obviously (en ) does not converge in X,
converges to 0 for every
since +en − em +2 = 2 for all n (= m ∈ N. (See Exercise 19 of this chapter.)
Before computing dual spaces of some well-known normed spaces, we prove
two additional facts about X and X " .

Proposition 3.14. Let X be a normed space.

132 3 Dual Spaces, Transposes and Adjoints

(i) Let X0 be a dense subspace of X. For x" ∈ X " , let F (x" ) denote the restric-
tion of x" to X0 . Then the map F is a linear isometry from X " onto X0" .
(ii) If X " is separable, then so is X.

Proof. (i) Let x" ∈ X " . It is clear that F (x" ) ∈ X0" and +F (x" )+ ≤ +x" +. In
fact, if x ∈ X and xn → x, where xn ∈ X0 for all n ∈ N, then
% %
|x" (x)| = % lim x" (xn )% ≤ +F (x" )+ lim +xn + = +F (x" )+ +x+,
n→∞ n→∞

so that +F (x" )+ = +x" +. Also, the map F : X " → X0" is linear. By Theorem
2.10 (iv), every x"0 ∈ X0" has a unique norm-preserving extension x" to X.
Hence the map F is onto.
(ii) Let X " be separable. Suppose X (= {0}. As we have mentioned in
Section 0.3, a nonempty subset of a separable metric space is separable. Let
then {x"1 , x"2 , . . .} be a countable dense subset of {x" ∈ X " : +x" + = 1}. Since
+x"j + = 1, there is xj ∈ X such that +xj + = 1 and |x"j (xj )| > 1/2 for each
j ∈ N. Let D := {k1 x1 + · · · + kn xn : n ∈ N and Re kj , Im kj ∈ Q for j =
1, . . . , n}. Then D is a countable subset of X, and clearly it is dense in the
subspace Y := span {x1 , x2 , . . .} of X. To show that Y is dense in X, we
consider x" ∈ X " such that x" (y) = 0 for all y ∈ Y . Assume for a moment that
x" (= 0, and let y " := x" /+x" +. Then +y " + = 1. Hence there is j ∈ N such that
+x"j − y " + < 1/2, and so |x"j (xj ) − y " (xj )| < 1/2. Then

& ' 1 1
|y " (xj )| = |x"j (xj )− x"j (xj )−y " (xj ) | ≥ |x"j (xj )|−|(x"j (xj )−y " (xj )| > − = 0.
2 2
Now xj ∈ Y and x" (xj ) = +x" +y " (xj ) (= 0 in contradiction to our assumption
that x" (y) = 0 for all y ∈ Y . This shows that if x" ∈ X " and x" (y) = 0 for all
y ∈ Y , then x" = 0. By Proposition 3.7 (ii), Y = X, that is, Y is dense in X.
Thus D is a countable dense subset of X, and so X is separable. #
$

The converse of part (ii) of the above proposition does not hold, as we shall
see in Example 3.19 (ii).
Now we shall calculate the dual of a Hilbert space by using the projection
theorem (Theorem 1.34).
Theorem 3.15. (Riesz representation theorem, 1907) Let H be a Hilbert
!
space, and let f ∈ H . Then there is a unique yf ∈ H, called the representer
of f , such that f (x) = -x, yf . for all x ∈ H. Further, the map T : H " → H
defined by T (f ) := yf , f ∈ H " , is a conjugate-linear isometry from H " onto H.

Proof. Since Z(f ) is a closed subspace of H, the projection theorem (Theorem

1.34) shows that H = Z(f ) ⊕ Z(f )⊥ . If f = 0, then let yf := 0, so that
f (x) = 0 = -x, 0. for all x ∈ H.
Next, let f (= 0. Then Z(f ) (= H, that is, Z(f )⊥ (= {0}. Let z ∈ Z(f )⊥ be
such that z (= 0 so that f (z) (= 0. For x ∈ H, consider
3.2 Dual Spaces and Riesz Representation 133

f (x)
w := x − z.
f (z)
Clearly, w ∈ Z(f ), and so -w, z. = 0, that is, -x, z. = f (x)-z, z./f (z). Then
f (z)
f (x) = -x, z. = -x, yf . for all x ∈ X,
-z, z.

where yf := f (z)z/-z, z.. To prove the uniqueness of yf ∈ H, let y ∈ H be

such that f (x) = -x, y. for all x ∈ H as well. Then -x, y. = f (x) = -x, yf .
for all x ∈ H. Hence y = yf .
For f and g in H " ,

(f + g)(x) = f (x) + g(x) = -x, yf . + -x, yg . = -x, yf + yg . for all x ∈ H.

Hence yf + yg is the representer of f + g ∈ H " , that is, T (f + g) = T (f ) + T (g).

Similarly, for f ∈ H " and k ∈ K, (k f )(x) = k f (x) = k-x, yf . = -x, k yf . for
all x ∈ H. Hence k yf is the representer of kf ∈ H " , that is, T (k f ) = k T (f ).
Also, for f ∈ H " ,

+f + = sup{|-x, yf .| : x ∈ H and +x+ ≤ 1} = +yf +

since |-x, yf .| ≤ +x+ +yf + for all x ∈ X by the Schwarz inequality (Proposition
1.12), and -yf , yf . = +yf +2 . Thus +T (f ) = +f +. Finally, given y ∈ H, define
fy (x) := -x, y. for x ∈ X. Then fy ∈ H " , and y is the representer of fy , that
is, T (fy ) = y. Thus T is a conjugate-linear isometry from H " onto H. #
$
Remark 3.16. The Riesz representation theorem does not hold for an in-
complete inner product space. For example, let X := c00 with the usual inner
product. Define f : X → K by
∞
( x(j)
f (x) := , x ∈ X.
j=1
j

Then f is linear. By letting n → ∞ in the Schwarz inequality for numbers

(Lemma 0.4),
"( ∞ $" (∞ $
1 π2
|f (x)|2 ≤ 2
|x(j)| 2
= +x+2 .
j=1
j j=1
6
√
Thus f ∈ (c00 )" and +f + ≤ π/ 6. But f has no representer in c00 , because if
y ∈ c00 , then f (ej ) = 1/j, while -ej , y. = y(j) = 0 for all large j ∈ N. !
Corollary 3.17. Let H be a Hilbert space.
(i) Let yf denote the representer of f ∈ H " . Define -f, g." = -yg , yf . for
f, g ∈ H " . Then -· , ·." is an inner product on H " , -f, f ." = +f +2 for all
f ∈ H " , and H " is a Hilbert space.
134 3 Dual Spaces, Transposes and Adjoints

(ii) H is a reflexive normed space.

Proof. (i) Let f, g, h in H " , and let k ∈ K. Then

-f + g, h." = -yh , y(f +g) . = -yh , yf + yg . = -f, h." + -f, g." ,

and
-kf , g." = -yg , ykf . = -yg , kyf . = k-f, g." .
Also, -f, g." = -yg , yf . = -yf , yg . = -g, f ." . Further, -f, f ." = -yf , yf . ≥ 0,
and -f, f ." = -yf , yf . = 0 if and only if yf = 0, that is, f = 0. Thus -· , ·."
is an inner product on H " . Since H is complete, and the map f 3−→ yf is
an isometry from H " onto H, we see that H " is also complete. Thus H " is a
Hilbert space.
(ii) Consider the canonical embedding J : H → H "" . Let φ ∈ H "" = (H " )" .
By (i) above, H " is a Hilbert space, and by the Riesz representation theorem
(Theorem 3.15), there is a representer g ∈ H " of φ, that is,

φ(f ) = -f, g." = -yg , yf . = f (yg ) = J(yg )(f ) for all f ∈ H " .

Thus φ = J(yg ). This shows that the map J is onto, that is, H is a reflexive
normed space. #
$

As a consequence of the Riesz representation theorem, we give a simple

proof of the Hahn-Banach extension theorem for a Hilbert space. Moreover,
we show that a Hahn-Banach extension is unique in this case.

Theorem 3.18. (Unique Hahn-Banach extension) Let H be a Hilbert

space, G be a subspace of H, and let g ∈ G" . Then there is a unique f ∈ H "
such that f (y) = g(y) for all y ∈ G, and +f + = +g+.

Proof. By Theorem 2.10 (iv), we can assume, without loss of generality, that G
is a closed subspace of H. Then G is a Hilbert space, and there is a representer
yg ∈ G of g ∈ G" by the Riesz representation theorem (Theorem 3.15). Define
f : H → K by f (x) := -x, yg . for x ∈ H. Then f is linear, f (y) = -y, yg . =
g(y) for all y ∈ G, and +f + = +yg + = +g+. Hence f is an Hahn-Banach
extension of g.
To prove the uniqueness of the Hahn-Banach extension of g, let f˜ ∈ H "
be such that f˜(y) = g(y) for all y ∈ G, and +f˜+ = +g+. Let ỹ ∈ H denote
the representer of f˜. Then +ỹ+ = +f˜+ = +g+ = +yg + and -yg , ỹ. = f˜(yg ) =
g(yg ) = -yg , yg .. Now

+yg − ỹ+2 = +yg +2 − 2Re -yg , ỹ. + +ỹ+2 = 2+yg +2 − 2Re -yg , yg . = 0,

and so ỹ = yg . (Compare Exercise 11 of Chapter 1.) Thus for all x ∈ H,

f˜(x) = -x, ỹ. = -x, yg . = f (x), that is, f˜ = f. #
$
3.2 Dual Spaces and Riesz Representation 135

The uniqueness of the Hahn-Banach extension of a continuous linear func-

tional on a subspace of a Hilbert space H also follows from Proposition 3.6,
since the dual H " of H is an inner product space by Corollary 3.17 (i), and
inner product spaces are strictly convex. (See Exercise 12 of Chapter 1.)

Examples 3.19. (i) Let ) H := #2 . Consider y := (y(1), y(2), . . .) ∈ #2 , and

∞
define fy (x) := -x, y. = j=1 x(j)y(j) for x := (x(1), x(2), . . .) ∈ #2 . As we
have seen in the proof of Theorem 3.15, fy ∈ H " and +fy + = +y+2 = +y+2 .
The map Φ(y) := fy , y ∈ #2 , gives a linear isometry from #2 to (#2 )" . Now let
f ∈ (#2 )" . By Theorem 3.15, f = fy , where y := yf . Therefore Φ is a linear
isometry from #2 onto (#2 )" .
Next, let H := L2 , where L2 denotes L2 ([a, b]). Consider y ∈ L2 , and define
*b
fy (x) := -x, y. = a x(t)y(t)dm(t) for x ∈ L2 . As we have seen in the proof of
Theorem 3.15, fy ∈ H " and +fy + = +y+2 = +y+2 . The map Φ(y) := fy , y ∈ L2 ,
gives a linear isometry from L2 to (L2 )" . Now let f ∈ (L2 )" . By Theorem 3.15,
f = fy , where y := yf . Therefore Φ is a linear isometry from L2 onto (L2 )" .
(ii) Let us find (#1 )" . Taking a cue from the description of (#2 )" given in
∞
)∞above, let us consider y := (y(1), y(2),1. . .) ∈ # , and define
(i) )∞ fy (x) :=
j=1 x(j)y(j) for x := (x(1), x(2), . . .) ∈ # . Then the series j=1 x(j)y(j)
is summable in K, and |fy (x)| ≤ +x+1 +y+∞ for all x ∈ # . Also, fy is clearly
1

a linear functional on #1 . Hence fy ∈ (#1 )" and +fy + ≤ +y+∞ . In fact, since
+fy + ≥ |fy (ej )| = |y(j)| for all j ∈ N, we obtain +fy + = +y+∞ . The map
Φ(y) := fy , y ∈ #∞ , gives a linear isometry from #∞ to (#1 )" . Now let f ∈ (#1 )" ,
and define y := (f (e1 ), f (e2 ), . . .). Since |f (ej )| ≤ +f + for all j ∈ N, we see
that y ∈ #∞ , and f = fy . Therefore Φ is a linear isometry from #∞ onto (#1 )" .
Next, let L1 denote L1 ([a, b]), and L∞ denote L∞ ([a, b]). To find (L1 )" , let
*b
us consider y ∈ L∞ , and define fy (x) := a x(t)y(t)dm(t) for x ∈ L1 . Then
*b
the Lebesgue integral a x(t)y(t)dm(t) exists, and |fy (x)| ≤ +x+1 +y+∞ for all
x ∈ L1 . Also, fy is clearly a linear functional on L1 . Hence fy ∈ (L1 )" and
+fy + ≤ +y+∞ . To show +y+∞ ≤ +fy +, that is, |y(t)| ≤ +fy + for almost all
t ∈ [a, b], we let En := {t ∈ [a, b] : |y(t)| > +f !y∞+ + (1/n)} for n ∈ N, and
E
)∞ := {t ∈ [a, b] : |y(t)| > +f y +}. Since E = n=1 En , we obtain m(E) ≤
n=1 m(E n ). It is enough to prove that m(E n = 0 for each n ∈ N. For
)
n ∈ N, let cEn denote the characteristic function of En , and xn := (sgn y)cEn .
Then xn y = |y|cEn , |xn | = cEn and +xn +1 = m(En ). Hence
" $ # # b
1
+fy + + m(En ) ≤ |y|dm = xn y dm = fy (xn ) ≤ +fy + m(En ),
n En a

and so m(En ) = 0 for each n ∈ N, as desired. This shows that +fy + = +y+∞ .
The map Φ(y) := fy , y ∈ L∞ , gives a linear isometry from L∞ to (L1 )" . Now
let f ∈ (L1 )" . For t ∈ (a, b], let ct denote the characteristic function of the
subinterval (a, t] of [a, b], so that ct ∈ L1 . Define z : [a, b] → K by z(a) := 0,
and z(t) := f (ct ) for t ∈ (a, b]. It can be shown that the function z is absolutely
136 3 Dual Spaces, Transposes and Adjoints

continuous on [a, b]. By the fundamental theorem of calculus for Lebesgue

integration (Theorem 0.23), z " (t) exists for almost all t ∈ [a, b], z " ∈ L1 , and
*t
z(t) = a z " dm for all t ∈ [a, b]. Define y := z " . It can be seen that y ∈ L∞
*b
and f (x) = a x(t)y(t)dm(t) for all x ∈ L1 , that is, f = fy . Therefore Φ is a
linear isometry from L∞ onto (L1 )" .
We observe that #1 and L1 are separable, but their dual spaces are not.
∞ "
)∞to find ∈ (# ) . Consider y := (y(1), y(2),
(iii) Let us attempt . . .) ∈ #1 ,
∞
and define fy (x) := j=1 x(j)y(j) for x := (x(1), x(2), . . .) ∈ # . Then the
)∞
series j=1 x(j)y(j) is summable in K, and |fy (x)| ≤ +x+∞ +y+1 for all x ∈ #∞ .
Also, fy is clearly a linear functional on #∞ . Hence fy ∈ (#∞ )" and +fy + ≤
+y+1 . In fact, if we let ) x := (sgn y(1), sgn y(2), . . .), then +x+∞ ≤ 1, and
∞
since +fy + ≥ |fy (x)| = j=1 |y(j)| = +y+1 , we obtain +fy + = +y+∞ . The
map Φ(y) := fy , y ∈ #1 , gives a linear isometry from #1 to (#∞ )" . However,
this isometry is not onto. To see this, consider the closed subspace c0 of #∞ ,
and let a := (1, 1, . . .) (∈ c0 . By Proposition 3.7 (ii), there is f ∈ (#∞ )" such
that f (x) = 0 for all x ∈ c0 and f (a) = 1. If y ∈ #1 and f = fy , then
y(j) = fy (ej ) = f (ej ) = 0 for all j ∈ N, that is, y = 0, and so fy = 0.
But f (= 0. Thus f (= fy for any y ∈ #1 . In fact, (#∞ )" cannot linearly
homeomorphic to #1 . This follows by noting that #1 is separable, but #∞ is
not (Example 0.7 (ii))), and if the dual space of a normed space X is separable,
then X itself must be separable (Proposition 3.14 (ii)).
In a similar manner, we can see that if y ∈ L1 , and we let fy (x) :=
*b
a
x(t)y(t)dm(t) for x ∈ L∞ , then fy ∈ (L∞ )" and +fy + ≤ +y+1 . In fact, if we
let x(t) := sgn y(t) for t ∈ [a, b], then +x+∞ ≤ 1, and since +fy + ≥ |fy (x)| =
*b
a
|y(t)| = +y+1 , we obtain +fy + = +y+∞ . The map Φ(y) := fy , y ∈ L1 , gives
a linear isometry from L1 to (L∞ )" . However, this isometry is not onto. To see
this, define g : C([a, b]) → K by g(x) = x(b) for x ∈ C([a, b]), and let f be a
Hahn-Banach extension of g to L∞ . Let (bn ) be a sequence in [a, b) such that
bn → b. For n ∈ N, let xn (t) := 0 if t ∈ [a, bn ] and xn (t) := (t − bn )/(b − bn ) if
t ∈ (bn , b]. Then xn ∈ C([a, b]) and +xn +∞ ≤ 1 for all n ∈ N. Also, xn (t) → 0
for each t ∈ [a, b) and xn (b) = 1. If y ∈ L1 and f = fy , then fy (xn ) = f (xn ) =
*b
g(xn ) = xn (b) = 1 for all n ∈ N, whereas fy (xn ) = a xn (t)y(t)dm(t) → 0 by
the dominated convergence theorem (Theorem 0.18 (ii)). Thus f (= fy for any
y ∈ L1 . In fact, (L∞ )" cannot linearly homeomorphic to L1 .
(iv) Let X := C([a, b]) with the sup norm + · +∞ . To find X " , consider
Y := BV ([a, b]), the linear space of all functions of bounded variation, along
with the norm +y+BV := |y(a)| + V (y), where V (y) denotes the total vari-
ation of y ∈ BV ([a, b]). (See Exercise 26 of Chapter 1.) Fix y ∈ Y , and
*b
define fy (x) := a x(t)dy(t) for x ∈ X, where dy denotes the Riemann-Stiltjes
integration with respect to y. Then fy is a linear functional on X. Also,
|fy (x)| ≤ +x+∞ +y+BV for all x ∈ X. Hence f ∈ X " and +fy + ≤ +y+BV , but an
equality may not hold here. We, therefore, replace BV ([a, b]) by its subspace
N BV ([a, b]) consisting of all normalized functions of bounded varia-
3.3 Transposes and Adjoints 137

tion, that is, functions of bounded variation which vanish at a, and which
are right continuous on (a, b). Then +fy + = +y+BV for y ∈ N BV ([a, b]). The
map Φ(y) := fy , y ∈ N BV ([a, b]), gives a linear isometry from N BV ([a, b]) to
(C([a, b]))" . Further, if f ∈ (C([a, b]))" , then there is a unique y ∈ N BV ([a, b])
such that f = fy . Therefore Φ is a linear isometry from N BV ([a, b]) onto
(C([a, b]))" . We refer the reader to [14, Theorem 6.4.4] for details. !

The results in Examples 3.19 (ii) and (iv) are also known as Riesz rep-
resentation theorems for #1 , L1 ([a, b]) and C([a, b]).

3.3 Transposes and Adjoints

Let X and Y be linear spaces, and let F ∈ BL(X, Y ). Then y " ◦F ∈ X " for
every y " ∈ Y " . The map F " : Y " → X " defined by F " (y " ) := y " ◦F, y " ∈ Y " , is
called the transpose of F . We shall first explain this nomenclature.
Let X be a normed space of dimension n, and let x1 , . . . , xn constitute a
basis for X. Also, let x"1 , . . . , x"n constitute the corresponding dual basis for
X " . Let Y be a normed space of dimension m, and let y1 , . . . , ym constitute
a basis for Y . Also, let y1" , . . . , ym
"
constitute the corresponding dual basis of
"
X . Since

F (xj ) = y1" (F (xj ))y1 + · · · + ym

"
(F (xj ))ym for each j = 1, . . . , n,

the m×n matrix M whose entry in the jth column and the ith row is yi" (F (xj ))
defines the linear map F : X → Y with respect to the basis x1 , . . . , xn for X
and the basis y1 , . . . , ym for Y , as we have seen in Section 0.2. Now since

F " (yi" ) = F " (yi" )(x1 )x"1 + · · · + F " (yi" )(xn )x"n for each i = 1, . . . , m,

we see that the n×m matrix M " whose entry in the ith column and the jth
row is F " (yi" )(xj ) defines the linear map F " : Y " → X " with respect to the basis
y1" , . . . , ym
"
for Y " and the basis x"1 , . . . , x"n for X " . Since F " (yi" )(xj ) = yi" (F (xj ))
for all i = 1, . . . , m, j = 1, . . . , n, we see that M " = M t , the transpose of the
matrix M . Thus if a matrix M defines a linear map F from X to Y , then the
transpose M t of M defines the linear map F " from Y " to X " . Hence the linear
map F " is called the transpose of the linear map F .

Examples 3.20. (i) Let p, r ∈ {1, 2}, and let F ) ∈ BL(#p , #r ). Define ki,j :=
∞
F (ej )(i) for i, j ∈ N. Consider x ∈ # . Since x = j=1 x(j)ej , the continuity
p
)∞
and the linearity of F shows that F (x) = j=1 x(j)F (ej ). Hence the series
)∞
j=1 ki,j x(j) converges in K to y(i) := F (x)(i) for every i ∈ N. Also, y :=
(y(1), y(2), . . .) = F (x) ∈ #r . Thus the infinite matrix M := [ki,j ] defines F .
(This is not the case if p = ∞ and r ∈ {1, 2, ∞}. See Exercise 21 (i) of this
chapter. On the other hand, if p = 1 and r ∈ {1, 2}, then we need not assume
that F is continuous. See Proposition 2.23 and Exercise 29 (iii) of Chapter 2.)