M383C Exercises

Kenneth DeMason, June 11, 2020

2 Normed Linear Spaces

1. Suppose that X is a vector space.
a) If A, B ⊂ X are convex, show that A + B and A ∩ B are convex. What about
A ∪ B and A \ B?

b) Show that 2A ⊂ A + A. When is it true that 2A = A + A?

Solution:
a) Let A, B ⊂ X be convex. Let x, y ∈ A + B. Then there exist ax , ay ∈ A and
bx , by ∈ B such that x = ax + bx and y = ay + by . Let t ∈ [0, 1]. Then,

tx + (1 − t)y = t(ax + bx ) + (1 − t)(ay + by ) = [tax + (1 − t)ay ] + [tbx + (1 − t)by ].

Since A, B are convex, [tax + (1 − t)ay ] ∈ A and [tbx + (1 − t)by ] ∈ B. Thus

tx + (1 − t)y can be written as the sum of an element of A and an element of
B. By definition, tx + (1 − t)y ∈ A + B, and hence A + B is convex.

Let x, y ∈ A ∩ B. Then x, y ∈ A. Since A is convex, for any t ∈ [0, 1] we have

tx + (1 − t)y ∈ A. Similarly, x, y ∈ B and B is convex. So tx + (1 − t)y ∈ B.
Thus tx + (1 − t)y ∈ A ∩ B and hence A ∩ B is convex.

A ∪ B need not be convex. For example, let A = [−2, −1] and B = [1, 2].
Then for x = −1, y = 1, t = 1/2 we have
1
tx + (1 − t)y = (1 − 1) = 0 ∈
/ A ∪ B.
2
Similarly, the difference A\B need not be convex. For example, take A = [−1, 1]
and B = {0}. Then A \ B = [−1, 0) ∪ (0, 1], and for x = −1, y = 1, t = 1/2,
1
tx + (1 − t)y = (1 − 1) = 0 ∈
/ A \ B.
2

b) Let x ∈ 2A. Then there exists a y ∈ A such that x = 2y. On the other hand,

x = 2y = y + y,

so that x is written as the sum of two elements of A. Thus x ∈ A + A, and

2A ⊂ A + A.

If A is convex, then 2A = A + A. To see this, let z ∈ A + A. Then z = x + y

where x, y ∈ A. In order for z to be in 2A, it must be that (x + y)/2 ∈ A.
However, if A is convex, then tx + (1 − t)y ∈ A for any x, y ∈ A and t ∈ [0, 1].
 M383C Exercises

Taking t = 1/2 shows that, for the chosen x, y, we have (x+y)/2 ∈ A as desired.

From the above, it is clear that 2A = A + A iff for all x, y ∈ A, (x + y)/2 ∈ A.

This convexity-type property is called midpoint convexity. It is equivalent to
the following: Let D denote the set of dyadic rationals. That is,
(∞ )
X xn
D= n
xn = 0 or 1 .
n=1
2
Say A has the property dyadic convexity if for all x, y ∈ A and d ∈ D,
dx + (1 − d)y ∈ A.
Clearly, 1/2 ∈ D. So, if A is dyadic convex, it is midpoint convex. We now
show that midpoint convex implies dyadic convex. Let x, y ∈ A and d ∈ D.
Then ∞
X xn
d=
n=1
2n
Note that d, 1 − d are complementary dyadic rationals in the following sense. If
xn denotes swapping 0 with 1 (that is, xn = 1 if xn = 0 and xn = 0 if xn = 1),
then ∞ ∞
X xn X xn
d= n
, (1 − d) =
n=1
2 n=1
2n
which follows since 1 = 1/2 + 1/4 + 1/8 + ....
We show that dx + (1 − d)y ∈ A by stages. Fix an N > 1 ∈ N. Define
z = (x + y)/2 ∈ A since A is midpoint convex. Set z0 as
z + xN x + xN y
z0 = ,
2
which is in A by midpoint convexity (either xN = 1 and xN = 0 or vice-verse).
Set z1 as
z0 + xN −1 x + xN −1 y
z1 = ,
2
which is similarly in A. Continuing in this fashion, define zk by
zk−1 + xN −k x + xN −k y
zk = ,
2
until k = N − 1. All the zk , in particular zN −1 , are in A by midpoint convexity.
Expanding zN −1 gives
 
zN −2 x1 xN −k 1 zN −3 + x2 x + x2 y x1 x1
zN −1 = + x+ y= + x+ y
2 2 2 2 2 2 2
zN −3 + x2 x + x2 y x1 x1
= 2
+ x+ y
 2 2 2  
1 zN −4 + x3 x + x3 y hx
1 x2 i x1 x2
= 2 + + 2 x+ + 2 y = ...
2 2 2 2 2 2
"N −1 # "N −1 #
z X xn X xn
= N −1 + n
x + n
y
2 n=1
2 n=1
2
 M383C Exercises

We may repeat this process for any N . Crucially, zN −1 is always in A. So,

taking the limit as N → ∞, we get
"∞ # "∞ #
X xn X xn
z∞ = n
x+ n
y = dx + (1 − d)y ∈ A.
n=1
2 n=1
2

2. Let (X, d) be a metric space.

a) Show that
ρ(x, y) = min(1, d(x, y))
is also a metric.

b) Show that U ⊂ X is open in (X, d) if and only if U is open in (X, ρ).

c) Repeat the above for

d(x, y)
σ(x, y) = .
1 + d(x, y)
Solution:

a) • Since d(x, y) is a metric, d(x, y) ≥ 0 for all x, y. Thus ρ(x, y) ≥ 0 since it

is either 1 or d(x, y).
• Suppose x = y. Then d(x, y) = 0, and since 0 < 1, ρ(x, y) = d(x, y) = 0.
Suppose now that ρ(x, y) = 0. Then it must be that d(x, y) = 0 (since
ρ(x, y) is either 1 or d(x, y)), and hence x = y because d is a metric.
• Let x, y ∈ X. Then

ρ(x, y) = min(1, d(x, y)) = min(1, d(y, x)) = ρ(y, x)

so ρ is symmetric.
• Let x, y ∈ X. Let z ∈ X. By definition of ρ,

ρ(x, y) ≤ 1

If either d(x, z) or d(y, z) exceeds 1, then ρ(x, z) or ρ(y, z) = 1. Thus

ρ(x, z) + ρ(y, z) > 1 ≥ ρ(x, y).

Now assume both d(x, z), d(y, z) ≤ 1. So ρ(x, z) = d(x, z) and ρ(y, z) =
d(y, z). Then by the triangle inequality of d,

ρ(x, y) ≤ d(x, y) ≤ d(x, z) + d(y, z) = ρ(x, z) + ρ(y, z).

Thus ρ is a metric.
 M383C Exercises

b) First observe if R ≤ 1, we have BRd (x) = BRρ (x). Let y ∈ BRd (x). Then,
d(x, y) < R ≤ 1. Thus, ρ(x, y) = d(x, y) < R. Hence, y ∈ BRρ (x) and BRd (x) ⊂
BRρ (x). Now let y ∈ BRρ (x). Then d(x, y) ≤ ρ(x, y) < R, and y ∈ BRd (x). So
BRρ (x) ⊂ BRd (x) and BRρ (x) = BRd (x).
Let U be open in (X, d). Fix an x ∈ U . Then there exists an R > 0 such that
BRd (x) ⊂ U . We may take R sufficiently small, for example R ≤ 1. The above
then shows BRρ (x) = BRd (x) ⊂ U , and U is open in (X, ρ). Essentially the same
proof shows U open in (X, ρ) implies U is open in (X, d).
c) We first show σ is a metric on X.
• Since d(x, y) is a metric, d(x, y) ≥ 0 for all x, y. Thus, 0 ≥ σ(x, y) ≤ 1.
• Suppose x = y. Then d(x, y) = 0 and σ(x, y) = 0/1 = 0. Suppose now
that σ(x, y) = 0. Then
d(x, y)
σ(x, y) = =0 ⇔ d(x, y) = 0(1 + d(x, y)) = 0.
1 + d(x, y)

• Let x, y ∈ X. Then
d(x, y) d(y, x)
σ(x, y) = = = σ(y, x),
1 + d(x, y) 1 + d(y, x)
so σ is symmetric.
• Observe that the function t/(1 + t) is monotone increasing for nonnegative
t. (Write t/(1 + t) as 1 − 1/(1 + t), and note that 1/(1 + t) is monotone
decreasing). Thus, since d(x, y) ≤ d(x, z) + d(z, y),
d(x, y) [d(x, z) + d(z, y)]
σ(x, y) = ≤
1 + d(x, y) 1 + [d(x, z) + d(y, z)]
d(x, z) d(z, y)
= +
1 + d(x, z) + d(y, z) 1 + d(x, z) + d(y, z)
d(x, z) d(z, y)
≤ + = σ(x, z) + σ(z, y)
1 + d(x, z) 1 + d(z, y)

Now we show that σ generates the same topology as d. Observe that, since
1 + d(x, y) ≥ 1, we have
d(x, y)
σ(x, y) = ≤ d(x, y)
1 + d(x, y)
So that for any R > 0, BRd (x) ⊂ BRσ (x). Now let y ∈ BRσ (x). Then,
d(x, y) R
<R ⇔ d(x, y) < R + Rd(x, y) ⇔ d(x, y) <
1 + d(x, y) 1−R
d
so, if R < 1, we have BR/(1−R) (x) ⊂ BRσ (x). Since open balls in (X, σ) are open
in (X, d) and vice verse, σ and d generate the same topology.
 M383C Exercises

3. Let X be a NLS, x0 be a fixed vector in X, and α 6= 0 a fixed scalar. Show that

the mappings x 7→ x + x0 and x 7→ αx are homeomorphisms of X onto itself.

Solution: The topology on X is induced by the metric d(x, y) = kx − yk, where

k · k is the norm on X.
i) T (x) = x + x0 is a homeomorphism. Let x, y ∈ X be such that T (x) = T (y).
Then x + x0 = y + y0 , and x = y. So, T is injective.
Let y ∈ X. Then x := y − x0 ∈ X (since X is a vector space), and T (x) =
x + x0 = y − x0 + x0 = y. So T is surjective.
Let U ⊂ X be open. It is easy to see that T −1 (x) = x − x0 . Let x ∈ T −1 (U ) =
U −x0 . Then x+x0 ∈ U and there exists an r > 0 such that Br (x+x0 ) ⊂ U . We
show Br (x) ⊂ T −1 (U ). Let y ∈ Br (x). Since y ∈ Br (x), we have kx − yk < r.
Then, kx + x0 − (y + x0 )k < r. It follows that y + x0 ∈ Br (x + x0 ) ⊂ U . By
definition of T −1 , we have T −1 (y + x0 ) = y + x0 − x0 = y. So, y ∈ T −1 (U ). A
similar argument (replacing x0 with −x0 ) shows that T −1 is continuous.
ii) S(x) = αx is a homeomorphism. Let x, y ∈ X be such that S(x) = S(y). Then
αx = αy. Since α 6= 0, we may divide by it and conclude x = y. So, S is
injective.
Let y ∈ X. Since α 6= 0 and X is a vector space, x := y/α ∈ X. Thus
S(x) = α(y/α) = y, and S is surjective.
Let U ⊂ X be open. We see that S −1 (x) = x/α. Then, S −1 (U ) = 1/αU . Let
x ∈ S −1 (U ). Then αx ∈ U , and there exists an r > 0 such that Br (αx) ⊂ U .
We show Br/α (x) ⊂ S −1 (U ). Let y ∈ Br/|α| (x). Then,
r
kx − yk < ⇔ |α|kx − yk < r ⇔ kαx − αyk < r
|α|
and αy ∈ Br (αx) ⊂ U . Thus, αy ∈ U . It follows that S −1 (αy) = y ∈ S −1 (U ).
4. Show that if X is a NLS, then X is homeomorphic to Br (0) for fixed r. [Hint:
consider the mapping x →
7 xr/(1 + kxk)].

Solution: Fix an r > 0. Define T (x) = xr/(1 + kxk). Let x, y ∈ X be such that
T (x) = T (y). Then,
 
xr yr 1 + kxk
= ⇔ x= y.
1 + kxk 1 + kyk 1 + kyk
Geometrically, this means that x, y are parallel (x = λy for some λ ∈ F). Taking
norms of both sides shows that
 
1 + kxk
kxk = (kyk) ⇔ kxk + kxkkyk = kyk + kxkkyk ⇔ kxk = kyk.
1 + kyk
So, x, y are parallel and have the same norm. This implies that x = λy where |λ| = 1.
Applying T shows that
yr λ(yr)
= T (y) = T (x) = T (λy) = ,
1 + kyk 1 + |λ|kyk
 M383C Exercises

and so,
y(1 + |λ|kyk) = y(λ + λkyk).
Since 1 + |λ|kyk is real and positive, it follows that λ is real and positive. Hence
λ = 1. Thus x = y and T is injective.
Let y ∈ Br (0). Then kyk < r so that r − kyk > 0. Now, x := αy ∈ X where
α = 1/(r − kyk) > 0. It follows that

α(yr) α(yr) yr yr
T (x) = T (αy) = = = = = y,
1 + |α|kyk 1 + αkyk 1/α + kyk r − kyk + kyk
and T is surjective.
Since d is a metric, d(x, 0) = kxk is continuous. Thus T (x) = ridX (x)/(1 + kxk) is
continuous, as the quotient of two continuous functions. Note that 1 + kxk 6= 0, so
that T (x) is continuous everywhere. Now, T −1 (x) = idX (x)/(r − kxk), which is the
quotient of two continuous functions and hence continuous. Since x ∈ Br (0) in this
case, kxk > r and r − kxk = 6 0. So T −1 is continuous on all of Br (0).

5. In Rd show that any two norms are equivalent. Hint: Consider the unit sphere,
which is compact.

Solution: We need only check that kxk is equivalent to

d
X
kxk1 := |xn |,
n=1

i.e. the l1 norm on Rd . To see this, note that norm equivalence is transitive. That
is, let k · ka and k · kb be equivalent to k · k1 . Then there exist ca , Ca and cb , Cb (all
positive) such that
ca k · k 1 ≤ k · k a ≤ C a k · k 1 ,
cb k · k 1 ≤ k · k b ≤ C b k · k 1 .
Since k · kb ≤ Cb k · k1 ,
   
k · kb ca
k · k a ≥ ca k · k 1 ≥ ca = k · kb .
Cb Cb

Similarly, since cb k · k1 ≤ k · kb ,
   
k · kb Ca
k · k a ≤ Ca k · k 1 ≤ Ca = k · kb ,
cb cb

and thus k · ka is equivalent to k · kb (with c = ca /Cb > 0 and C = Ca /cb > 0). Let
{en }dn=1 be a basis for Rd . Then by the triangle inequality,
d
X
kxk = kx1 e1 +...+xd ed k ≤ kx1 e1 k+...+kxd ed k = |x1 |ke1 k+...+|xd ked k = |xn |ken k.
n=1
 M383C Exercises

Let M = max ken k. Then,

d
X d
X
kxk ≤ |xn |ken k ≤ |xn |M = M kxk1 .
n=1 n=1

This inequality shows that balls in (X, k · k) are open in (X, k · k1 ). Then if U is
open in R, k · k−1 (U ) is open in (X, k · k) (by continuity of the norm in its generated
topology), and by the above, k · k−1 (U ) is open in (X, k · k1 ). So k · k is continuous on
(X, k·k1 ). Let m = inf x∈S11 kxk. Since k·k is a continuous function on the compact set
S11 (because it’s continuous in (X, k·k1 )), it attains its minimum. Thus there exists an
x0 ∈ S11 such that kx0 k = m. This shows that m > 0 (since 0 ∈ / S11 ). Now let x ∈ X
be nonzero. Then x/kxk1 ∈ S11 and by definition of the infimum, k(x/kxk1 )k ≥ m.
Hence
k(x/kxk1 )k ≥ m ⇔ kxk ≥ mkxk1
If x = 0, we trivially have mkxk1 ≤ kxk. Together, for all x ∈ X,

mkxk1 ≤ kxk ≤ M kxk1

so that k · k and k · k1 are equivalent.

6. Let X and Y be NLS over the same field, both having the same finite dimen-
sion n. Then prove that X and Y are topologically isomorphic, where a topological
isomorphism is defined to be a mapping that is simultaneously an isomorphism and
a homeomorphism.

Solution: It suffices to show that X, Y ' Fn , where ' denotes topologically iso-
morphic. This follows since ' is an equivalence relation.
Let {ei }ni=1 be a basis for X and define T : X → Fn by
n
X
T (x) = (x1 , ..., xn ) where x = xi ei
i=1

is the unique representation of x in the basis {ei }. Now define k · k1 := kT (·)kl1 . Then
k · k1 is a norm on X. Since X is finite dimensional, all norms on it are equivalent
(Prop 2.10 in notes). Then there exist c, C > 0 such that

ck · k1 ≤ k · k ≤ Ck · k1 .

We first show that T is an isomorphism of vector spaces (that is, T is an invertible

linear map). Let x, y ∈ X and α ∈ F. Then
n
X n
X
x= xi e i , y= yi ei .
i=1 i=1
 M383C Exercises

Now we have
n n
! n
!
X X X
T (x + αy) = T xi ei + α yi ei =T (xi + αyi )ei
i=1 i=1 i=1
= (x1 + αy1 , ..., xn + αyn ) = (x1 , ..., xn ) + (αy1 , ..., αyn )
= (x1 , ..., xn ) + α(y1 , ..., yn ) = T (x) + αT (y)

where the last few equalities follow from the vector space structure on Fn . So, T
is linear.PTo show T is invertible, note that for x = (x1 , ..., xn ) ∈ Fn , the map
S(x) = ni=1 xi ei is such that ST = idX and T S = idF n . So, S = T −1 and T is
invertible. Otherwise, we show that T is bijective as follows:
Pn that x ∈ X is such that T (x) = 0. Then xi = 0 for all 1 ≤ i ≤ n. So,
Suppose
x = i=1 xi ei = 0. Thus ker(T ) = {0} and T is injective. By the rank-nulllity
theorem, it follows that range(T ) has dimension n, and hence must coincide with Fn .
Then, T is bijective.
So, T is an invertible linear map, and is hence an isomorphism. We must show now
that T is a homeomorphism. It remains to prove that T and T −1 are continuous. But
for x ∈ X,
1
kT (x)kl1 = kxk1 ≤ kxk
c
by definition of k · k1 and the equivalence of norms. Similarly for x ∈ Fn ,

kT −1 (x)k ≤ CkT −1 (x)k1 = CkT T −1 (x)kl1 = Ckxkl1

so T and T −1 are bounded linear maps, and hence are continuous.

7. Show that (C[a, b], k · k∞ ), the set of real-valued continuous functions in the
interval [a, b] with the sup-norm (L∞ -norm), is a Banach space.

Solution: First recall that the sup-norm is given by

kf k∞ = sup |f (x)| < ∞

x∈[a,b]

where finiteness comes from the fact that |f | is a continuous function on a compact
set. Let {fn }∞
n=1 ⊂ C[a, b] be a Cauchy sequence. Let x ∈ [a, b]. Then,

|fn (x) − fm (x)| ≤ sup |fn − fm | = kfn − fm k∞

x∈[a,b]

and it follows that {fn (x)}∞

n=1 is a Cauchy sequence in F. Since F is complete, the
sequence converges to some point – define f by

f (x) = lim fn (x).

n→∞

Let  > 0 and N such that for m, n ≥ N ,

kfn − fm k∞ < .
 M383C Exercises

It follows that for all x ∈ [a, b],

|fn (x) − fm (x)| < .

Taking the limit as m → ∞ in the above shows that

|fn (x) − f (x)| ≤ ,

and therefore
kfn − f k∞ = sup |fn (x) − f (x)| ≤ sup  = 
x∈[a,b] x∈[a,b]

So fn → f in norm. This means that fn → f uniformly. Since fn are continuous,

and the uniform limit of continuous functions is continuous, f is continuous. Thus
f ∈ C[a, b], and C[a, b] is complete.

8. If f ∈ Lp (Ω) show that

Z Z
kf kp = sup f g dx = sup |f g| dx
Ω Ω

where the supremum is taken over all g ∈ Lq (Ω) such that kgkq ≤ 1 and 1/p+1/q = 1,
where 1 ≤ p, q ≤ ∞.

Solution: First note that the above equalities are trivially true if f = 0 a.e. So,
throughout, we will assume that f 6= 0. By Hölder’s inequality,

kf gk1 ≤ kf kp kgkq

where f ∈ Lp (Ω) and g ∈ Lq (Ω) with 1/p + 1/q = 1 and 1 ≤ p, q ≤ ∞. Assuming

that kgkq ≤ 1,
Z Z
|f g| dx ≤ kf kp ⇒ sup |f g| dx ≤ kf kp .
Ω Ω

By the trivial inequality Z Z

f dx ≤ |f | dx

it is easily seen that

Z Z
sup f g dx ≤ sup |f g| dx ≤ kf kp .
Ω Ω

To prove equality, we must appeal to several cases.

• [p = 1, q = ∞]. Let g(x) = sgn f (x), where sgn(x) is defined as

1
 x>0
sgn(x) = 0 x=0

−1 x<0

 M383C Exercises

It is clear that kgk∞ = 1 so that g ∈ Lq (Ω) = L∞ (Ω), and satisfies the condition
kgkq ≤ 1. Moreover, Z Z
f g dx = |f | dx = kf k1
Ω Ω
so that the sup is attained.

• [1 < p, q < ∞]. Define g(x) by

|f (x)|p−1 sgn f (x)

g(x) =
kf kp−1
p

Observe that q(p − 1) = pq(1 − 1/p) = pq(1/q) = 1. Then,

kf kpp
Z Z
q 1 p q 1 p
kgkq = |f | | sgn(f )| dx = |f | dx = =1
kf kpp Ω kf kpp Ω kf kpp

by noting trivial properties of the sgn function. Then g ∈ Lq (Ω) and kgkq ≤ 1.
Finally,

|f |p−1 f sgn f kf kpp

Z Z Z
1 p
f g dx = dx = |f | = = kf kp ,
Ω Ω kf kp−1
p kf kp−1
p Ω kf kp−1
p

and the sup is attained.

• [p = ∞, q = 1]. Let  > 0 and let E be a set of positive measure such that
|f (x)| > kf k∞ −  > 0. Such a set exists, since kf k∞ is the smallest number
such that
µ{|f | > kf k∞ } = 0.
That E can be chosen to have finite measure follows from the σ-finiteness of µ.
Now set g = χE sgn f /µ(E). Then,
Z Z
1 1 µ(E)
kgk1 = χE | sgn f | dx = χE dx = =1
µ(E) Ω µ(E) Ω µ(E)

where the second to last equality follows since, if x is such that f (x) = 0, then
/ E. Thus g ∈ L1 (Ω) = Lq (Ω) and kgkq ≤ 1 as required. Then,
x∈
Z Z Z
1 1
f g dx = χE |f | dx = χE |f | dx
Ω µ(E) Ω µ(E) Ω
Z Z
1
= |f | dx ≥ (kf k∞ − ) dx = kf k∞ − 
µ(E) E E

Taking  → 0 shows that Z

kf k∞ ≤ f g dx
Ω
 M383C Exercises

So, in all the above cases, there exists a g ∈ Lq (Ω) with kgkq ≤ 1 such that
Z
kf kp ≤ f g dx .
Ω
d
9. Suppose that Ω ⊂ R has finite measure and 1 ≤ p ≤ q ≤ ∞.
a) Prove that if f ∈ Lq (Ω), then f ∈ Lp (Ω) and
kf kp ≤ (µ(Ω))1/p−1/q kf kq .

b) Prove that if f ∈ L∞ (Ω), then

lim kf kp = kf k∞ .
p→∞

c) Prove that if f ∈ Lp (Ω) for all 1 ≤ p < ∞, and there is K > 0 such that
kf kp ≤ K, then f ∈ L∞ (Ω) and kf k∞ ≤ K.
Solution: Throughout, we assume µ(Ω) 6= 0. Otherwise, all equalities are trivial since
every function f is a.e. equal to 0 on Ω.
a) First, we may assume that p < q, otherwise (in the case p = q), the conclusion
and inequality are trivially true. Now, observe that for any r ∈ [1, ∞], χΩ ∈
Lr (Ω). This is easy to see since, for r finite,
Z Z
r r
kχΩ kr = |χΩ | dx = χΩ dx = µ(Ω) < ∞
Ω Ω

For r = ∞, it is easy to see that kχΩ k∞ = 1. Now,

Z Z
p p
kf kp = |f | dx = |f |p |χΩ | dx.
Ω Ω

Observe that q/p and 1/(1 − p/q) are conjugate exponents since
1 1 p p
+ = +1− =1
q/p 1/(1 − p/q) q q
Furthermore, since p < q, p/q < 1 and 1 > 1 − p/q > 0. It follows that
1/(1 − p/q) ∈ (1, ∞). Next, f p ∈ Lq/p (Ω), since
Z p/q Z p/q
p q/p q
kf kq/p = |f | dx = |f | dx = kf kpq < ∞
Ω Ω

Thus, f p ∈ Lq/p (Ω) and χΩ ∈ L1/(1−p/q) (Ω). By Hölder’s inequality,

Z
|f |p χΩ dx = kf p χΩ k1 ≤ kf kq/p kχΩ k1/(1−p/q) = µ(Ω)1−p/q kf kpq .
Ω

In total,
kf kpp ≤ µ(Ω)1−p/q kf kpq
from the conclusion follows, since x 7→ x1/p is monotone increasing and hence
preserves inequalities.
 M383C Exercises

b) By a), since f ∈ L∞ (Ω), f ∈ Lp (Ω) for all 1 ≤ p ≤ Ω, and

kf kp ≤ µ(Ω)1/p−1/∞ kf k∞ = µ(Ω)1/p kf k∞ .
Since µ(Ω) 6= 0, taking p → ∞ gives
lim sup kf kp = kf k∞ .
p→∞

Let Et = {x ∈ Ω | |f (x)| > t}. Then,

|f |p kf kpp
Z Z Z
1 p
µ(Et ) = 1 dx ≤ p
dx ≤ |f | dx =
Et Et t tp X tp
since |f (x)|/t > 1 on Et . This is known as Chebyshev’s inequality. Now let
t ∈ (0, kf k∞ ). Then µ(Et ) > 0 (otherwise, t < kf k∞ would be such that |f | ≤ t
a.e., a contradiction of the definition of kf k∞ ). Then by Chebyshev’s inequality,
kf kp ≥ tµ(Et )1/p .
Since Ω has finite measure, all subsets do. It follows that µ(Et )1/p → 0 as
p → ∞. Hence,
lim inf kf kp ≥ t,
p→∞

and taking t → kf k∞ gives

lim inf kf kp ≥ kf k∞ .
p→∞

Thus,
kf k∞ ≤ lim inf kf kp ≤ lim kf kp ≤ lim sup kf kp ≤ kf k∞
p→∞ p→∞ p→∞

and limp→∞ kf kp = kf k∞ .
c) The above proof showed that
lim inf kf kp ≥ kf k∞
p→∞

regardless whether f ∈ L∞ (Ω) or not. But, kf kp ≤ K, so that

kf k∞ ≤ lim inf kf kp ≤ lim inf K = K.
p→∞ p→∞

So, f ∈ L∞ (Ω) and kf k∞ ≤ K.

10. Finite dimensional matrices.
a) Let M n×m be the set of matrices with real valued coefficients aij , for 1 ≤ i ≤ n
and 1 ≤ j ≤ m. For every A ∈ M n×m , define
|Ax|Rn
kAk = maxm .
x6=0∈R |x|Rm
Show that (M n×m , k · k) is a NLS.
 M383C Exercises

b) Each A ∈ M n×n defines a linear map of Rn into itself. Show that

kAk = max y T Ax,

|x|=|y|=1

where y T is the transpose of y.

c) Show that each A ∈ M n×n is continuous.

d) Prove that the convex hull is convex, and that it is the intersection of all convex
subsets of X containing A.

e) If X is a normed linear space, prove that the convex hull of an open set is open.

f) If X is a normed linear space, is the convex hull of a closed set always closed?

g) Prove that if X is a normed linear space, then the convex hull of a bounded set
is bounded.

Solution:

a) • Since each of | · |Rn , | · |Rm are nonnegative, we have that the quotient is
always nonnegative. Therefore kAk ≥ 0.
• Suppose A = 0, the zero matrix. Then by simple matrix multiplication,
it is seen for any x ∈ Rm that Ax = 0. Thus |Ax|Rn = 0 for all x, and
kAk = 0. Suppose now that kAk = 0. Then for all x 6= 0 ∈ Rm ,

|Ax|Rn
≤0 ⇔ |Ax|Rn ≤ 0|x|Rm = 0
|x|Rm

On the other hand, as just discussed, |Ax|Rn is always nonnegative. Thus,

Ax = 0 for all x ∈ Rm , which by definition means that A = 0.
• Let α ∈ R. Then
|αAx|Rn |Ax|Rn |Ax|Rn
kαAk = maxm = maxm |α| = |α| maxm = |α|kAk.
x6=0∈R |x|Rm x6=0∈R |x|Rm x6=0∈R |x|Rm

• Appealing to the triangle inequality of | · |Rn ,

|Ax + Bx|Rn |Ax|Rn |Bx|Rn

kA + Bk = maxm ≤ maxm +
x6=0∈R |x|Rm x6=0∈R |x|Rm |x|Rm
|Ax|Rn |Bx|Rn
≤ maxm + maxm = kAk + kBk
x6=0∈R |x|Rm x6=0∈R |x|Rm

where the second inequality comes from the fact that, for any U, V , max(U +
V ) ≤ max(U ) + max(V ).
 M383C Exercises

b) As a remark, recall that hy, xi = y T In x, where h·, ·i is the standard inner product
on Rn and In is the n × n identity matrix. Therefore, y T Ax = y T In AX =
hy, Axi. Now, by Cauchy-Schwarz, and assuming x, y have unit norm,

hy, Axi ≤ |y||Ax| = |Ax|

Taking the maximum of both sides shows that

max y T Ax ≤ max |Ax| = kAk.

|x|=|y|=1 |x|=1

Now for the reverse inequality,

|Ax|2
 
T Ax
max y Ax ≥ max , Ax = max = max |Ax|
|x|=|y|=1 |x|=1,Ax6=0 |Ax| |x|=1,Ax6=0 |Ax| |x|=1,Ax6=0

where the inequality comes from taking the maximum over smaller sets. Now,
if Ax = 0 then |Ax| = 0, and hence does not contribute to the max (unless
A = 0, in which case there is nothing to prove). Thus,

max y T Ax ≥ max |Ax| = max |Ax| = kAk.

|x|=|y|=1 |x|=1,Ax6=0 |x|=1

c) Note that M n×n is a finite dimensional vector space, so all norms on it are
equivalent. Define
Xn
kAk1 = max |aij |.
1≤j≤n
i=1

It can easily be checked that k · k1 is a norm on M n×n . Now, clearly, kAk1 < ∞
for all matrices A. So A is bounded, and hence continuous. Since all norms
generate the same topology on a finite dimensional vector space, we see that A
is continuous in k · k.
2
d) Note that M n×n ' Rn , so continuing forward we work with subsets of Rd .
Define the convex hull of a set Ω ⊂ Rd by
( n n
)
X X
co(Ω) = λi xi | n ∈ N, xi ∈ Ω, λi ≥ 0, λi = 1
i=1 i=1

This is clearly a generalization of the condition tx + (1 − t)y ∈ Ω whenever

t ∈ [0, 1] and x, y ∈ Ω (take n = 2, x1 = x, x2 = y, λ1 = t, λ2 = 1 − t). We can
then prove the following proposition:
Pn
Proposition: A set C ⊂ Rd P is convex iff i=1 λi xi ∈ C for all n ∈ N, x1 , ..., xn ∈
C, and λ1 , ..., λn ≥ 0 with ni=1 λi = 1.

Proof: (⇐) Take the case n = 2, then the discussion above shows that C is
convex.
 M383C Exercises

P C be convex. Let {xi }, {λi } satisfy the required properties. We show

(⇒) Let
that ni=1 λi xi ∈ C. We prove this inductively. Suppose it holds for n − 1; we
show it holds for n. Note that we may take n ≥ 2, since the case n = 1 is trivial.
Further, we may take λi > 0 (since, otherwise, wePmay omit the point xi from
n−1
the sum and decrease n). Define λ = 1 − λn = i=1 λi . By the assumption
λi > 0, we see that λ 6= 0. Define µi = λi /λ for 1 ≤ i ≤ n − 1. Then µi ≥ 0 and
n−1 n−1
X 1X λ
µi = λi = = 1
i=1
λ i=1 λ
Pn−1
By the inductive hypothesis, y := i=1 µi xi ∈ C. By convexity of C,

λy + (1 − λ)xn ∈ C

But,
n−1 n
λX X
λy + (1 − λ)xn = λi xi + λn xn = λ i xi .
λ i=1 i=1

This completes the proof of the proposition. Note that, rephrased, this says
that C is convex iff co(C) = C.
Pn Pn
Now, we show that co(Ω) is convex. Let p = i=1 λi pi and q = i=1 µi pi
be in co(Ω). Note that we can take pi in both by adding terms with coefficients
of zero to each sum. Now, for λ ∈ [0, 1], we have
n
X
λp + (1 − λ)q = (λλi + (1 − λ)µi )pi
i=1

Observe that λλi + (1 − λ)µi ≥ 0 for all 1 ≤ i ≤ n since λ, λi , µi ≥ 0. Next,

n
X n
X n
X
(λλi + (1 − λ)µi ) = λ λi + (1 − λ) µi = λ + (1 − λ) = 1
i=1 i=1 i=1

since p, q ∈ co(Ω). It follows that λp + (1 − λ)q ∈ co(Ω), and co(Ω) is convex.

We now show that co(Ω) = ∩α∈A Cα where {Cα }α∈A is the set of convex sets Cα
containing Ω. Clearly, Ω ⊂ co(Ω) (by taking n = 1 and λ = 1), and co(Ω) is
convex. Hence, co(Ω) ∈ {Cα }, and ∩α∈A Cα ⊂ co(Ω). Let C ∈ {Cα }. Then C is
convex, and by the proposition C = co(C). But, by definition of C, Ω ⊂ C, and
hence co(Ω) ⊂ co(C) = C. Since C was arbitrary, co(Ω) ⊂ C for all C ∈ {Cα },
and hence co(Ω) ⊂ ∩α∈A Cα .

So, co(Ω) is convex, and is equal to the intersection of all convex sets containing
Ω.
 M383C Exercises

e) Let U ⊂ X be open. Let x ∈ co(U ). Then,

n
X n
X
x= λ i xi , xi ∈ U, λi ≥ 0, λi = 1
i=1 i=1

By the sum condition of {λi }, there exists a λk 6= 0. WLOG, we may choose

this to be λ1 . Now define f : X → X by
n
!
1 X
f (z) = z− λ i xi
λi i=2

It is clear that f (z) is continuous (it’s a polynomial – actually, it is a homothety

−1 −1
Pn (x − λ1 x1 )/(1 − λ1 )). Thus, f (U ) is open. But, fPn(U ) = λ1 U +
fixing
i=2 λi xi . In particular, since x1 ∈ U , we have that λ1 x1 + i=2 λi xi = x ∈
−1 −1
f (U ). Furthermore, it is clear that f (U ) ⊂ co(U ). Thus co(U ) is open.

f) No. For example, take Ω = {(x, y) | y ≥ 1/(1 + x2 )}. Then Ω is closed, but
co(Ω) = R2y>0 , which is open.

g) A set B is bounded if there exists BR (0) containing B. Since X is a normed

linear space, BR (0) is convex. Hence,

co(B) ⊂ co(BR (0)) = BR (0)

and co(B) is bounded.

11. Prove that if X is a normed linear space and B = B1 (0) is the unit ball, then X is
infinite dimensional if and only if B contains an infinite collection of non-overlapping
balls of diameter 1/2.

Solution: We need the following lemma due to Riesz:

Lemma [Riesz’ Lemma]: Let M be a closed, proper subspace of X and 1 > r > 0.
Then there exists an x ∈ X such that kxk = 1 and inf m∈M kx − mk ≥ r.

Proof: Let x1 ∈ / M and set R = inf m∈M kx1 − mk. Then R > 0 (if not, there
exists a sequence mn ∈ M such that kx1 − mn k → 0, which means that mn → x1
in norm. But M is closed so that x1 ∈ M , a contradiction). Let  > 0. Then there
exists an m1 ∈ M such that kx1 − m1 k < R + . Set x = (x1 − m1 )/kx1 − m1 k, which
clearly has unit norm. Then,
m1 x1 1
inf kx−mk = inf m+ − = inf kmkx1 − m1 k + m1 − x1 k
m∈M m∈M kx1 − m1 k kx1 − m1 k kx1 − m1 k m∈M
But M is a proper subspace so that mkx1 − m1 k + m1 ∈ M . Thus,
1 R
inf kx − mk = inf km − x1 k >
m∈M kx1 − m1 k m∈M R+
 M383C Exercises

For a prescribed r < 1, there exists an  > 0 such that R/(R + ) = r, completing the
proof.

Next, observe that if X is infinite dimensional, then for any x1 , ..., xn ∈ X, the
set M = Span(x1 , ..., xn ) is a closed, proper subspace of X. Clearly M is finite di-
mensional, and hence (by Exercise 6) is topologically isomorphic to Fn . But Fn is
closed, and so M is closed too. Further, since M is finite dimensional, it is a proper
subspace of X.

So, let x1 ∈ X have unit norm. Then M1 = Span(x1 ) is a closed, proper sub-
space of X. Hence, for r = 2/3, there exists by Riesz’ Lemma an x2 ∈ / M1 such that
d(x2 , M1 ) ≥ 2/3 and kx2 k = 1. Let M2 = Span(x1 , x2 ). By the same logic, there
exists an x3 ∈ / M2 such that d(x3 , M2 ) ≥ 2/3 and kx3 k = 1. Continuing inductively
to obtain x1 , x2 , ... all with unit norm such that for any m, n, d(xm , xn ) ≥ 1/2. Now
let B i denote the ball centered at 3/4xi with radius 1/4. Then B i ∩ B j are disjoint
for i 6= j, since if x ∈ B i ∩ B j ,
3 3 3 3 3 3 1 1 1
xi − xj ≤ xi − x + x − xj ≤ xi − x + x − xj < + =
4 4 4 4 4 4 4 4 2
But, by construction,
3 3 1
kxi − xj k ≥ 2/3 ⇔ xi − xj ≥
4 4 2

Next, B i ⊂ B for all i. To see this, let x ∈ B i . Then,

3 3 3 3 1 3
kxk ≤ x − xi + xi ≤ x − xi + xi ≤ + = 1
4 4 4 4 4 4
by using the fact that xi was chosen so that kxi k = 1. Hence, there exists an infinite
collection of non-overlapping balls with diameter 1/2 (equivalently, radius 1/4).

For the other direction, suppose X is finite dimensional. Let {B i } be a non-overlapping

collection of balls with diameter 1/2. We show that such a collection cannot exist.
Since X is finite dimensional, the closed unit ball is compact. Consider the sequence
{xi } where each xi is the center of B i . Then there exists a convergent subsequence,
which we will still denote by {xi }. Then for  = 1/4, there exists an N such that
for m, n ≥ N , kxm − xn k < 1/4. It follows that xm ∈ B n , so that xm ∈ B n ∩ B m
(similarly for xn ). Thus the balls are overlapping.

12. Prove that a subset A of a metric space (X, d) is bounded if and only if ev-
ery countable subset of A is bounded.

Solution: Clearly if A is bounded, then every subset of A (countable or not) is

bounded. Hence we need only prove the reverse direction. Suppose that every count-
able subset of A is bounded. To achieve a contradiction, suppose that A itself is not
 M383C Exercises

bounded. Then, for every R > 0, A ( BR (0). In particular, this holds for R = n ∈ N.
Now let xn ∈ A \ Bn (0). Then à = {xn } is a countable subset of A, and hence is
bounded. On the other hand, for any R > 0, there exists an N > R, and thus
xN ∈
/ BR (0) (by definition of xN ). So, Ã is not bounded, a contradiction.

13. Consider (lp , k · kp ).

a) Prove that lp is a Banach space for 1 ≤ p ≤ ∞. Hint: Use that R is complete.
b) Show that k · kp is not a norm for 0 < p < 1. Hint: First show the result on R2 .
Solution:
a) Throughout, we will use the fact that x 7→ xp and x 7→ x1/p are monotone
nondecreasing for all p. Hence, they preserve order.
First suppose that p ∈ [1, ∞). Then k · kp is defined by

∞
! p1
X
kxkp = |xi |p ,
i=1

where x = (x1 , x2 , ...). Let {xk }∞ p

k=1 be a Cauchy sequence in l . Then, for  > 0,
there exists a K > 0 such that for m, n ≥ K,
∞
X
|xm n p m n p p
i − xi | = kx − x kp <  .
i=1

Since all the |xm k

i − xi | are nonnegative, this implies that, for fixed i,

|xm n p p
i − xi | <  ,

and thus, for fixed i, there exists a K > 0 such that for all m, n ≥ K,

|xm n
i − xi | < .

This means that for fixed i, {xki } is Cauchy. But, this is a sequence in F, which
is complete. So, it converges to some x∞ i . Define x
∞
by

x∞ = (x∞ ∞
1 , x2 , ...).

We must show that xk → x∞ in norm, and that x∞ ∈ lp . Since {xk } is Cauchy,

for any  > 0, there exists K such that for m, n ≥ K, and for any I,
I
! p1
X
|xm n p
i − xi | ≤ kxm − xn kp < .
i=1

Letting m → ∞ gives
I
! p1
X
|x∞ n p
i − xi | < .
i=1
 M383C Exercises

Finally, letting I → ∞ gives

kx∞ − xn k < ,

and xk → x in norm.
Again since {xk } is Cauchy, it is bounded (say, by C > 0). Then for any I,

I
! p1
X
|xki |p ≤ kxk kp ≤ C.
i=1

Taking the limit as k → ∞ gives

I
! p1
X
|x∞
i |
p
≤ kxk kp ≤ C,
i=1

and then taking the limit as I → ∞ gives

kx∞ kp ≤ C

so that x∞ ∈ lp .

For p = ∞, let {xk } be Cauchy in l∞ . Let  > 0. Then there exists a K > 0
such that for m, n ≥ K,
sup |xni − xm
i | < 
i

which implies, for any fixed i, that

|xni − xm n m
i | ≤ sup |xi − xi | < 
i

and hence {xki } is Cauchy in F. Since F is complete, this sequence converges,

and define x∞ as above. Since {xk } is Cauchy, it is bounded, and therefore

sup |xki | ≤ C
i

Then for fixed i,

|xki | ≤ C,
and taking the limit as k → ∞ gives

|x∞
i | ≤ C.

Hence, taking the sup over i of both sides,

kx∞ k∞ = sup |x∞

i | ≤ C
i

so that x∞ ∈ l∞ . Now, as before, for m, n ≥ K we have

kxm − xn k∞ = sup |xm n

i − xi | < .
i
 M383C Exercises

Thus, for fixed i,

|xm n
i − xi | < 

Taking the limit as m → ∞ and then suping over i gives

kx∞ − xn k∞ < 

so that xk → x∞ in norm.

b) We show that the triangle inequality fails. Consider (1, 0) and (0, 1), which
both have norm 1 for any p. But,
1 1
k(1, 0) + (0, 1)kp = (1p + 1p ) p = 2 p

Since 0 < p < 1, we have that 1/p > 1, and thus

1
k(1, 0) + (0, 1)kp = 2 p ≥ 2 = 1 + 1 = k(1, 0)kp + k(0, 1)kp

The same result holds in lp for (1, 0, 0, 0...) and (0, 1, 0, 0, ...).

14. If an infinite dimensional vector space X is also a NLS and contains a sequence
{en }∞
n=1 with the property that for every x ∈ X there is a unique sequence of scalars
{αn }∞
n=1 such that

kx − (α1 e1 + ... + αn en )k → 0 as n → ∞,

then {en }∞
n=1 is called a Schauder basis for X, and we have the expansion of x

∞
X
x= αn e n .
n=1

a) Find a Schauder basis for lp , 1 ≤ p < ∞.

b) Show that if a NLS has a Schauder basis, then it is separable. [Remark: The
converse is not true.]

Solution:

a) Define ei by (ei )n = δin . We easily get an expansion for x ∈ X by

∞
X
x = (x1 , x2 , ...) = xn e n .
n=1

Then,
N
X ∞
X
x− xn en = xn e n → 0
n=1 N −1

as N → ∞, since the sum becomes empty.

 M383C Exercises

b) Note that a Schauder basis {en } is countable, so its span is countable. Then, it
suffices to show that Span{en } dense in X. Let x ∈ X. Then by definition of a
Schauder basis, there exists an expansion of x,
∞
X
x= αn en .
n=1

Let xN be defined by
N
X
xN = αn en ,
n=1

which by definition of a Schauder basis satisfies

kx − xN k → 0 as N → ∞.

Therefore, {xN } is a sequence in Span{en } such that xN → x in norm. Hence,

Span{en } is dense in X.
15. Let Y be a subspace of a vector space X. The coset of an element x ∈ X with
respect to Y is denoted by x + Y and is defined to be the set

x + Y = {z ∈ X | z = x + y for some y ∈ Y }.

Show that the distinct cosets form a partition of X. Show that under algebraic
operations defined by

(x1 + Y ) + (x2 + Y ) = (x1 + x2 ) + Y and λ(x + Y ) = λx + Y,

for any x1 , x2 , x ∈ X and λ in the field, these cosets form a vector space. This space
is called the quotient space of X by (or modulo Y , and it is denoted X/Y ).

Solution: The cosets form a partition since ' defined by x+Y ' z +Y iff x−z ∈ Y is
an equivalence relation. Alternatively, since Y is a subspace, 0 ∈ Y . Thus, x ∈ x + Y ,
and each coset is nonempty.
The above also easily implies that
[
x + Y = X,
x∈X

since each coset is a subset of X, and each point x ∈ X is in the coset x + Y .

Suppose now that x1 + Y and x2 + Y are distinct cosets. We show that they are
disjoint. Suppose there exists an x ∈ (x1 + Y ) ∩ (x2 + Y ). Then x = x1 + y1 and
x = x2 + y2 for some y1 , y2 ∈ Y . Now let z ∈ x1 + Y . Then z = x1 + y for some
y ∈ Y . By the above, x1 = x2 + y2 − y1 so that

z = x1 + y = (x2 + y2 − y1 ) + y = x2 + (y2 − y1 + y).

Since Y is a subspace, y2 − y1 + y ∈ Y , and z ∈ x2 + Y . The reverse inclusion is

similar, and hence x1 + Y = x2 + Y , a contradiction.
 M383C Exercises

Thus, the cosets form a partition.

We now show that the set of cosets forms a vector space. To this end, we must
show eight properties.

i) (Associativity of addition). Let x1 , x2 , x3 ∈ X. Then,

[(x1 + Y ) + (x2 + Y )] + (x3 + Y ) = [(x1 + x2 ) + Y ] + (x3 + Y ) = ((x1 + x2 ) + x3 ) + Y

= (x1 + (x2 + x3 )) + Y = (x1 + Y ) + [(x2 + x3 ) + Y ]
= (x1 + Y ) + [(x2 + Y ) + (x3 + Y )].

ii) (Commutativity of addition). Let x1 , x2 ∈ X. Then,

(x1 + Y ) + (x2 + Y ) = (x1 + x2 ) + Y = (x2 + x1 ) + Y = (x2 + Y ) + (x1 + Y ).

iii) (Identity under addition) Let 0 denote the 0 element of X. Then for any x ∈ X,

(x + Y ) + (0 + Y ) = (x + 0) + Y = x + Y.

iv) (Inverse under addition). Let x ∈ X. Since X is a vector space, −x ∈ X. Then,

(x + Y ) + (−x + Y ) = (x − x) + Y = 0 + Y.

v) (Compatibility of scalar multiplication with field multiplication). Let a, b ∈ F.

Then, for any x ∈ X,

a(b(x + Y )) = a(bx + Y ) = abx + Y = (ab)(x + Y ).

vi) (Identity under multiplication). Let 1 be the multiplicative identity of F. Then

for any x ∈ X,
1(x + Y ) = (1x) + Y = x + Y

vii) (First distributivity law). Let a ∈ F and x1 , x2 ∈ X. Then,

a((x1 + Y ) + (x2 + Y )) = a((x1 + x2 ) + Y ) = (a(x1 + x2 ) + Y )

= (ax1 + ax2 ) + Y = (ax1 + Y ) + (ax2 + Y )
= a(x1 + Y ) + a(x2 + Y ).

viii) (Second distributivity law). Let a, b ∈ F and x ∈ X. Then,

(a + b)(x + Y ) = (a + b)x + Y = (ax + bx) + Y = (ax + Y ) + (bx + Y )

= a(x + Y ) + b(x + Y ).

Thus, the set of cosets of X (under Y ) is a vector space.

 M383C Exercises

16. Let Y be a closed subspace of a NLS X. Show that a norm on the quotient space
X/Y is given for x̂ ∈ X/Y by

kx̂kX/Y = inf kxkX .

x∈x̂

Solution:

• Evidently for any x ∈ x̂, since k · kX is a norm, inf x∈x̂ ≥ 0.

• Suppose x̂ is the zero coset, x̂ = 0 + Y . Since Y is a subspace, 0 ∈ Y , and

therefore 0 ∈ x̂. But k0kX = 0, so that the inf is zero. Thus, kx̂kX/Y = 0. Now
suppose that kx̂kX/Y = 0. Then there exists a sequence {xn } ⊂ x̂ such that
kxn kX → 0. Since xn ∈ x̂ for all n, we may write them as xn = x + yn (where
x̂ = x + Y ) for some yn ∈ Y . Since Y is closed,

xn − x = yn → y ∈ Y.

Now, the xn converge in X to 0 so that, on the other hand,

yn = xn − x → 0 − x = −x.

Since limits are unique, −x ∈ Y . It follows that x ∈ Y , and so x + Y = 0 + Y .

• Observe that the definition given is equivalent to the following. Let x̂ = x + Y .

Then,
kx̂kX/Y = inf kx − ykX = d(x, Y ).
y∈Y

This follows since every z ∈ x̂ is of the form x − y for some y ∈ Y , so

inf kzkX = inf kx − ykX .

z∈x̂ y∈Y

Now let α 6= 0 ∈ F. Then,

kαx̂kX/Y = inf kαx − ykX = inf kαx − αykX = |α| inf kx − ykX = |α|kx̂kX/Y ,
y∈Y y∈Y y∈Y

since Y is a subspace. If α = 0, then αx̂ = (αx) + Y = 0 + Y is the zero coset,

and
kαx̂kX/Y = 0 = |α|kx̂kX/Y .

• The triangle inequality follows from the triangle inequality of k · kX and of inf.
Let xˆ1 = x1 + Y and xˆ2 = x2 + Y . Then,

kxˆ1 + xˆ2 kX/Y = inf kx1 + x2 − ykX = inf kx1 + x2 − y1 − y2 kX

y∈Y y1 ,y2 ∈Y
≤ inf kx1 − y1 kX + kx2 − y2 kX
y1 ,y2 ∈Y
≤ inf kx1 − y1 kX + inf kx2 − y2 kX = kxˆ1 kX/Y + kxˆ2 kX/Y .
y1 ∈Y y2 ∈Y
 M383C Exercises

17. If X and Y are NLS, then the product space X × Y is also a NLS with any of
the norms
k(x, y)kX×Y = max(kxkX , kykY )
or, for any 1 ≤ p < ∞,
1
k(x, y)kX×Y = (kxkpX + kykpY ) p .

Why are these norms equivalent?

Solution: Denote the max norm by k · k and the p-norms by k · kp . We will show three
inequalities. Fix a p. We have that, for any (x, y) ∈ X × Y ,

(kxkX + kykY )p ≤ (2 max{kxkX , kykY )p ≤ 2p (k(x, y)kp ).

Next, k(x, y)kp is one of kxkpX or kykpY . Thus, we may add the other to it. Since both
are nonnegative, this either increases or does not change the value. Hence using the
previous inequality,

k(x, y)kp1 ≤ 2p (k(x, y)kp ) ≤ 2p (kxkpX + kykpY ) = 2p k(x, y)kpp .

It follows that
k(x, y)k1 ≤ 2k(x, y)kp .
By employing the same trick,

k(x, y)k ≤ kxkX + kykY = k(x, y)k1 .

Finally,
1 1 1
k(x, y)kp = (kxkpX + kykpY ) p ≤ (2k(x, y)kp ) p = 2 p k(x, y)k.
Together, we conclude that
1 1
k(x, y)k 1 ≤ k(x, y)kp ≤ k(x, y)k ≤ k(x, y)k1
21+1/p 21/p
This shows that k · k1 is equivalent to k · kp for any p and k · k. Thus they are all
equivalent (by transitivity).

Another way to see this is that they all define norms on R2 , which is finite dimen-
sional, and so all norms are equivalent.

18. If X and Y are Banach spaces, prove that X × Y is a Banach space.

Solution: Let (X, k · kX ) and (Y, k · kY ) be Banach spaces. Equip X × Y with the
norm k(x, y)k1 = kxkX + kykY . Let {(xn , yn )} be Cauchy in (X × Y, k · k1 ). Then for
any  > 0 there exists N such that for m, n ≥ N ,

kxn − xm kX + kyn − ym kY = k(xn − xm , yn − ym )k1 = k(xn , yn ) − (xm , ym )k1 < .

 M383C Exercises

Then, the sequences {xn } and {yn } are Cauchy in (X, k · kX ) and (Y, k · kY ), re-
spectively. Thus, they converge, say to x, y. In the above inequality, take m → ∞.
Then,
kxn − xkX + kyn − ykY = k(xn , yn ) − (x, y)k1 < 
so that (xn , yn ) → (x, y) in norm. Thus, (X × Y, k · k1 ) is complete.

19. Let T : C[0, 1] → C[0, 1] be defined by

Z t
(T x)(t) = x(τ ) dτ.
0

Find the range R(T ) of T , and show that T is invertible on its range, T −1 : R(T ) →
C[0, 1]. Is T −1 linear and bounded?

Solution: The range of T is the set of all functions f on [0, 1] such that f (0) = 0, f 0
exists, and f 0 is continuous. That is, R(T ) = C 1 [0, 1]. To see this, note that (T x)(t)
gives a C 1 function since
   
d d d
(T x)(t) = t x(t) − 0 x(0) = x(t)
dt dt dt

and by definition, x(t) is continuous on [0, 1]. Moreover,

Z 0
(T x)(0) = x(τ ) dτ = 0.
0

So that T x is a C 1 function on [0, 1] such that (T x)(0) = 0. Now, if f ∈ C 1 [0, 1] with

f (0) = 0, then f 0 is continuous on [0, 1] and
Z t
0
(T f )(t) = f 0 (τ ) dτ = f (t) − f (0) = f (t).
0

So, T surjects onto the set of C 1 [0, 1] functions which fix 0. The inverse has already
been found – it is (T −1 x)(t) = d/dtx(t). Since the differential operator is linear, T −1
is linear. However, T −1 is not bounded. For example, define fn by

fn (x) = sin(nx)

Then,
fn0 (x) = n cos(nx).
It follows that
kfn kC[0,1] = 1, kT −1 fn kC[0,1] = kfn0 kC[0,1] = n.
Therefore, there cannot be a C > 0 such that

kT −1 f kC[0,1] ≤ Ckf kC[0,1]

 M383C Exercises

for all f ∈ C 1 [0, 1] with f (0) = 0.

20. Show that on C[a, b], for any y ∈ C[a, b] and scalars α and β, the function-
als Z b
f1 (x) = x(t)y(t) dt and f2 (x) = αx(a) + βx(b)
a
are linear and bounded.

Solution:
• We show that f1 is linear and bounded. Let x1 , x2 ∈ C[a, b] and λ ∈ F. Then,
Z b Z b
f1 (x1 + λx2 ) = (x1 (t) + λx2 (t))y(t) dt = x1 (t)y(t) + λx2 (t)y(t) dt
a a
Z b Z b Z b Z b
= x1 (t)y(t) dt + λx2 (t)y(t) dt = x1 (t)y(t) dt + λ x2 (t)y(t) dt
a a a a
= f1 (x1 ) + λf1 (x2 )
following linearity of the integral. To show that f1 is bounded,
Z b Z b Z b
|f1 (x)| ≤ x(t)y(t) dt ≤ |x(t)||y(t)| dt ≤ kxkkyk dt = [(b−a)kyk]kxk
a a a

where k · k = k · kC[a,b] is the sup norm. It follows that kf1 kC[a,b]∗ ≤ (b − a)kyk,
and f1 is bounded.
• We show that f2 is linear and bounded. Let x1 , x2 ∈ C[a, b] and λ ∈ F. Then,
f2 (x1 + λx2 ) = α(x1 (a) + λx2 (a)) + β(x1 (b) + λx2 (b))
= αx1 (a) + λαx2 (a) + βx1 (b) + λβx2 (b)
= [αx1 (a) + βx1 (b)] + λ[αx2 (a) + βx2 (b)]
= f2 (x1 ) + λf2 (x2 ),
and f2 is linear. Now to show that f2 is bounded,
|f2 (x)| ≤ |αx(a) + βx(b)| ≤ |α||x(a)| + |β||x(b)| ≤ (|α| + |β|)kxk.
So, kf2 kC[a,b]∗ ≤ |α| + |β|, and f2 is bounded.
21. Find the norm of the linear functional f defined on C[−1, 1] by
Z 0 Z 1
f (x) = x(t) dt − x(t) dt.
−1 0

Solution: By an easy inequality,

Z 0 Z 1 Z 0 Z 1
|f (x)| ≤ x(t) dt − x(t) dt ≤ x(t) dt + x(t) dt
−1 0 −1 0
Z 0 Z 1 Z 1
≤ |x(t)| dt + |x(t)| dt = |x(t)| dt
−1 0 −1
≤ (1 − (−1))kxk = 2kxkC[−1,1]
 M383C Exercises

Thus, kf kC[−1,1]∗ ≤ 2. To show kf kC[−1,1]∗ = 2, we approximate by functions. Note

that the simple function x(t) = χ[−1,0] − χ[0,1] is such that
Z 0 Z 1
x(t) dt − x(t) dt = 2 = 2kxkC[−1,1] .
−1 0

However, x(t) is not continuous. Instead, define xn (t) on [−1, 1] by


1
 −1 ≤ x ≤ −1/n
xn (t) = −nx |x| ≤ 1/n

−1 1/n ≤ x ≤ 1


Then,    
1 1 1
f (xn ) = 2 1 − +2 =2−
n 2n n
(geometrically, we compute the area of two triangles with the same base length 1/n
and height 1, and the area of two rectangles with the same base length (1 − 1/n) and
height 1). It follows that
 
1
|f (xn )| = 2 − kfn kC[−1,1]
n

Thus, kf kC[−1,1]∗ ≥ 2 − 1/n for all n, and it follows that kf kC[−1,1]∗ = 2. (If it were
less than 2, we could find an N such that 2 − 1/N exceeds the norm, and apply the
above to xN ).

23. The space C 1 [a, b] is the NLS of all continuously differentiable functions defined
on [a, b] with the norm

kxk = sup |x(t)| + sup |x0 (t)|.

t∈[a,b] t∈[a,b]

a) Show that k · k is indeed a norm.

b) Show that f (x) = x0 ((a+b)/2) defines a continuous linear functional on C 1 [a, b].

c) Show that f defined above is not bounded on the subspace of C[a, b] consisting
of all continuously differentiable functions with the norm inherited from C[a, b].

Solution: Note that differentiation is a linear mapping. Most properties follow from
this fact.

a) We show that k · k is a norm.

• First, since |x(t)|, |x0 (t)| ≥ 0 for all t ∈ [a, b], we have that the supremae are
nonnegative and hence k · k ≥ 0. Furthermore, since x, x0 are continuous
on a compact set, they are bounded, and k · k < ∞.
 M383C Exercises

• Suppose that x(t) = 0, the constant zero function on [a, b]. Then x0 (t) = 0,
and
kxk = sup 0 + sup 0 = 0.
t∈[a,b] t∈[a,b]

Suppose now that kxk = 0. Since both supremae are nonnegative, for their
sum to be zero, both must be zero. In particular, for any t ∈ [a, b],
0 ≥ |x(t)| ≤ sup |x(t)| = 0,
t∈[a,b]

and thus x(t) = 0 on [a, b]. So, x is the constant zero function. Note that
we could have not reached the same conclusion if kxk was defined to be
supt∈[a,b] |x0 (t)|, since all constant functions have a derivative of 0.
• Let α ∈ R. Then for any x ∈ C 1 [a, b],
kαxk = sup |(αx)(t)| + sup |(αx)0 (t)| = sup |αx(t)| + sup |αx0 (t)|
t∈[a,b] t∈[a,b] t∈[a,b] t∈[a,b]
0
= |α| sup |x(t)| + |α| sup |x (t)| = |α|kxk.
t∈[a,b] t∈[a,b]

• Let x, y ∈ C 1 [a, b]. Then,

kx + yk = sup |(x + y)(t)| + sup |(x + y)0 (t)| = sup |x(t) + y(t)| + sup |x(t) + y 0 (t)|
t∈[a,b] t∈[a,b] t∈[a,b] t∈[a,b]

≤ sup [|x(t)| + |y(t)|] + sup [|x(t)| + |y(t)|]

t∈[a,b] t∈[a,b]

≤ sup |x(t)| + sup |y(t)| + sup |x0 (t)| + sup |y 0 (t)| = kxk + kyk
t∈[a,b] t∈[a,b] t∈[a,b] t∈[a,b]

which follows from the triangle inequality of | · | and of sets, sup(A + B) ≤

sup(A) + sup(B).
Thus, k · k is a norm.
b) We first show f is a linear functional. Clearly f : C 1 [a, b] → R so that it is a
functional. To show it is linear, let x, y ∈ C 1 [a, b] and α ∈ R. Then,
       
0 a+b 0 0 a+b 0 a+b 0 a+b
f (x + αy) = (x + αy) = (x + αy ) =x + (αy )
2 2 2 2
   
a+b a+b
= x0 + αy 0 = f (x) + αf (y).
2 2
So, f is linear. To show that f is continuous, we show it is bounded. Let
x ∈ C 1 [a, b]. Then,
 
0 a+b
|f (x)| = x ≤ sup |x0 (t)| ≤ sup |x(t)| + sup |x0 (t)| = kxk
2 t∈[a,b] t∈[a,b] t∈[a,b]

where the first inequality follows by definition of the supremum, and the sec-
ond inequality follows since we simply add something nonnegative. Thus,
kf kC 1 [a,b]∗ ≤ 1, and f is bounded, hence continuous.
 M383C Exercises

c) Consider xn (t) = sin(n(t − (a + b)/2)). Then, x0n (t) = n cos(n(t − (a + b)/2))

and   
0 a+b a+b
xn ((a + b)/2) = n cos n − = n cos(0) = n
2 2
On the other hand, kxn kC[a,b] = 1 for all n. So, f cannot possibly be bounded.

24. Suppose X is a vector space. The algebraic dual of X is the set of all linear func-
tionals on X, and is also a vector space. Suppose also that X is a NLS. Show that
X has finite dimension if and only if the algebraic dual and the dual space X ∗ coincide.

Solution: Suppose that X has finite dimension. We need to show that the alge-
braic dual and X ∗ coincide. Clearly, X ∗ (the set of all bounded linear functionals) is
a subset of the algebraic dual. So, we need to show that every linear functional on X
is bounded. To this end, let T be a linear functional and define a norm on X by

kxkT := |T (x)|.

This is easily seen to be a norm by noting that T is linear. Since all norms are
equivalent on a finite dimensional space, there exists a C > 0 such that kxkT ≤ Ckxk.
Hence,
|T (x)| = kxkT ≤ Ckxk,
and T is bounded. Now suppose X is infinite dimensional. We construct a discontin-
uous (not bounded) linear functional. First note that every vector space has a Hamel
basis. This is defined to be a set B such that B is linearly independent, and every
vector x ∈ X can be written as a finite linear combination of vectors in B. By Zorn’s
lemma, every linearly independent set E can be extended to a Hamel basis. So, take
{en }∞
n=1 a linearly independent set (such a set can be chosen inductively). We may
further take ken k = 1. Then extend this to a Hamel basis B. Define T : X → R by

T (en ) = nken k = n,

and zero for all other vectors in B. Since any vector x can be represented as a linear
combination of vectors, we can extend T to a linear functional on X by
N
! N
X X
T̃ (x) = T̃ x n bn = xn T (bn )
n=1 n=1

where bn ∈ B. Some care must be taken to show this is linear (in particular, we must
write x, y as sums over the same basis vectors, adding terms of zero when necessary).
It is clear that
T̃ (en ) = T (en ) = n
so that T̃ is unbounded. So, the algebraic dual and X ∗ do not coincide when X is
infinite dimensional.

25. Let X be a NLS and M a nonempty subset. The annihilator M a of M is

 M383C Exercises

defined to be the set of all bounded linear functionals f ∈ X ∗ such that f restricted
to M is zero. Show that M a is a closed subspace of X ∗ . What are X a and {0}a ?

Solution: Let {fn }∞ a ∗

n=1 be a sequence in M such that fn → f ∈ X . We show
that f ∈ M a . For any fixed x ∈ M , we have

f (x) = lim fn (x) = lim 0 = 0,

n→∞ n→∞

and therefore f vanishes on M . Another way to see M a is closed is as follows. Define

for x ∈ X the element of x̂ of X ∗∗ by

x̂(τ ) = τ (x),

where τ ∈ X ∗ . Then τ vanishes on M if τ ∈ ker x̂ (where x ∈ M ), and M a ⊂ ker x̂.

Thus, \
Ma = ker x̂
x∈M

Each kernel is closed since ker(x̂) = x̂−1 (0), and each x̂ is continuous (they are all in
the double dual, hence bounded). Thus M a is closed.

The annihilator of X is the set of bounded linear functionals which are zero on
X. Thus, X a = {0}, where 0 is the zero function. On the other hand, {0}a is the
set of functionals which are zero at zero. Hence, {0}a is the set of bounded linear
functionals such that f (0) = 0. But, all linear functionals fix zero (by linearity), and
so {0}a = X ∗ .

26. Define the operator T by the formula

Z b
(T f )(x) = K(x, y)f (y) dy.
a

Suppose that K ∈ Lq ([a, b] × [a, b]), where q lies in the range 1 ≤ q ≤ ∞. Determine
the values of p for which T is necessarily a bounded linear operator from Lp (a, b) to
Lq (a, b). In particular, if a and b are both finite, show that K ∈ L∞ ([a, b] × [a, b])
implies T to be bounded on all the Lp -spaces.

Solution: An important result, proved previously, is if f ∈ Lp , then f ∈ Lr for

all 1 ≤ r ≤ p. Consider the case T : Lq (a, b) → Lq (a, b).
If q = ∞, then
Z b Z b
k(T f )(x)kq = sup K(x, y)f (y) dy ≤ sup |K(x, y)||f (y)| dy
x∈[a,b] a x∈[a,b] a

NEED TO FINISH.

27. Let U = Br (0) = {x ∈ X | kxk < r} be an open ball about 0 in a real normed
 M383C Exercises

linear space, and let y ∈

/ cl(U ). Show that there is a bounded linear functional f that
separates U from y. (That is, U and y lie in opposite half spaces determined by f ,
which is to say there is an α such that U lies in {x | f (x) < α} and f (y) > α.)

Solution: Since y ∈/ cl(U ), y ∈ cl(U )c . Hence, y lies in some open set, and there exists
an R > 0 such that BR (y) ⊂ cl(U )c . This exactly means that cl(U ) ∩ BR (y) = ∅.
Hence, U ∩BR (y) = ∅. By the separating hyperplane theorem, there exists an f ∈ X ∗
and γ ∈ R such that f (x) < γ for all x ∈ U , and f (x) > γ for all x ∈ BR (y). In
particular, this implies that U ⊂ {f (x) < γ} and f (y) > γ.

28. Prove that L2 [0, 1] is of the first category in L1 [0, 1]. (Recall that a set is of first
category if it is a countable union of nowhere dense sets, and that a set is nowhere
dense if its closure has an empty interior.) Hint: Show that Ak = {f | kf k2 ≤ k} is
closed in L1 but has empty interior.

Solution: Defining Ak as in the hint, let {fn }∞ n=1 be a sequence in Ak such that
fn → f ∈ L1 . By Fatou’s Lemma:
Z 1 Z 1 Z 1
2 2
|f | dx = lim inf |fn | dx ≤ lim inf |fn |2 dx ≤ k.
0 0 n→∞ n→∞ 0

Hence, f ∈ Ak . where fn → f is a subsequence that converges a.e. to f . Let f ∈ Ak

and g ∈ L1 [0, 1] \ L2 [0, 1]. Then f + 1/ng → f in the L1 norm, since

1 1
f+ g−f = kgk1 → 0
n 1 n

where kgk1 is some fixed, finite number. Note, however, that f + 1/ng ∈
/ Ak for any
n. To see this, observe that

2 1
Z
1
f + g = kf k2 + kgk2 + |f g| dx
n 2 n 0

but kgk2 = ∞ since g ∈ / L2 [0, 1], and every other term is nonnegative. Hence,
f + 1/ng ∈ / Ak , and since every (L1 ) neighborhood of f will contain f + 1/ng, it
must be that the neighborhood is not entirely contained in Ak . Thus Ak has empty
interior. Clearly, L2 [0, 1] is the union of all Ak .

Alternatively, one may note that L2 [0, 1] ⊂ L1 [0, 1]. Then, if any Ak has an interior
point f in L1 , we may (via Exercises 3 and 4) find a homeomorphism to Br (f ) ⊂ Ak .
Then, by taking the inverse of this homeomorphism, one may show that the inclusion
operator j : L2 [0, 1] → L1 [0, 1] is surjective. However, this is a contradiction.

29. If a Banach space X is reflexive, show that X ∗ is also reflexive. (∗) Is the
converse true? Give a proof or a counterexample.
 M383C Exercises

Solution: We prove the following lemma first.

Lemma: If Y ⊂ X is closed, and X is reflexive, than Y is reflexive.

Proof: Let ϕ ∈ Y ∗∗ and define ϕ̃ ∈ X ∗∗ by ϕ̃(f ) = ϕ(f |Y ). Since X is reflexive,

ϕ̃ = Ex for some x ∈ X. We show that x ∈ Y , and that ϕ is an evaluation map at x.
Suppose not; then by the Mazur separation lemma, there exists an f : Y → F such
that f (x) 6= 0 and f |Y = 0. But,

f (x) = ϕ̃(f ) = ϕ(f |Y ) = ϕ(0) = 0

a contradiction. So x ∈ Y . By Hahn-Banach, every bounded linear functional f :

Y → F can be written as f˜ : X → F, where f˜|Y = f . Then,

ϕ(f ) = ϕ̃(f˜) = f˜(x) = f (x)

since x ∈ Y . Thus, ϕ is an evaluation map

We now prove the following

Prop: X is reflexive iff X ∗ is reflexive.

Solution: Suppose first that X ∗ is reflexive. Let F : X → X ∗∗ denote the canonical

inclusion. Consider X ∗∗∗ = (X ∗∗ )∗ , and let F ∗ : X ∗ → X ∗∗∗ be the embedding of X ∗
into its double dual. Let η ∈ X ∗∗∗ , and define fη ∈ X ∗ by

fη (x) = η(F (x)).

Since X is reflexive, for each ϕ ∈ X ∗∗ , there exists an x ∈ X such that ϕ = F (x).

Then,
(F ∗ (fη ))(ϕ) = ϕ(fη ) = fη (x).
By definition of fη ,
fη (x) = η(F (x)) = η(φ)
so that F ∗ (fη ) = η. It follows that F ∗ is surjective.
Suppose now that X ∗ is reflexive. Then X ∗∗∗ is reflexive. By the above, X ∗∗ is re-
flexive. But, X ' F (X) ⊂cl X ∗∗ . Applying the lemma, we see that X is reflexive.

30. Let y = (y1 , y2 , y3 , ...) ∈ C∞ be a vector of complex numbers such that ∞

P
i=1 yi xi
∞
converges for every x = (x1 , x2 , x3 , ...) ∈ C0 , where C0 = {xC | xi → 0 as i → ∞}.
Prove that ∞
X
|yi | < ∞.
i=1

Solution: NEED TO FINISH

31. Let X and Y be normed linear spaces, T ∈ B(X, Y ), and {xn }∞

n=1 . If xn * x,
 M383C Exercises

w∗
prove that T xn −→ T x in Y . Thus a bounded linear operator is weakly sequentially
continuous. Is a weakly sequentially continuous linear operator necessarily bounded?
w
Solution: That xn − → x means, for any f ∈ X ∗ , f (xn ) → f (x). So, let g ∈ Y ∗ .
Consider the operator Tg := g ◦ T : X → F. Then, Tg ∈ X ∗ since it is the com-
position of bounded linear maps. Thus, applying weak convergence to Tg , we see
that
g(T xn ) = g ◦ T (xn ) = Tg (xn ) → Tg (x) = g ◦ T (x) = g(T x).
w
In total, for any g ∈ Y ∗ , we have that g(T xn ) → g(T x). Hence, T xn −
→ T x in Y .

Suppose, for any weakly convergent sequence xn , we have that T xn is weakly con-
vergent. Then, by Proposition 2.91 (as a consequence of PUB), it follows that
{kT xn k}∞
n=1 is bounded. Suppose that T is unbounded. Then there is a sequence
xn → 0 in norm (hence, weakly to 0, since strong implies weak convergence) such
that kT xn k → ∞. Then, T xn does not converge weakly (otherwise, it would be
bounded), a contradiction.

32. Suppose that X is a Banach space, M and N are linear subspaces, and that
X = M ⊕ N , which means that

X = M + N = {m + n | m ∈ M, n ∈ N }

and M ∩ N is the trivial linear subspace consisting only of the zero element. Let P
denote the projection of X onto M . That is, if x = m + n, then

P (x) = m.

Show that P is well defined and linear. Prove that P is bounded if and only if both
M and N are closed.

Solution: Let m1 + n1 and m2 + n2 are representations of x. We show that the

representation is actually unique. We have that m1 + n1 = m2 + n2 . After rearrang-
ing, we get m1 − m2 = n2 − n1 . Since M, N are linear subspaces, the LHS is an
element of M , while the RHS is an element of N . The only element shared in both
is the zero element, hence m1 − m2 = 0 = n2 − n1 , and m1 = m2 , n1 = n2 . Hence the
representation is unique, and P is well defined. Let x = mx + nx and y = my + ny be
two elements of X and λ ∈ F. Then,

P (x + λy) = P (mx + λmy + nx + λny ) = mx + λmy = P (x) + λP (y).

Suppose P is bounded (hence continuous). There are several ways to show that
both M and N are closed. First, observe that M = range(P ) and N = ker(P ).
By definition, range(T ) ⊂ M ; on the other hand, every m ∈ M can be written as
m + 0 ∈ X, so that P (m) = m. Every n ∈ N can be written as 0 + n ∈ X, so that
P (n) = P (0 + n) = 0. Hence N ⊂ ker(P ). On the other hand, if x ∈ ker(P ),
then P (x) = 0. This exactly means that if x = m + n, then m = 0. Hence
 M383C Exercises

x = 0 + n = n ∈ X.
So, if P is continuous, the kernel is clearly closed, for ker(P ) = P −1 {0}. To show
that M is closed, note that Q = IX − P is a projection onto N . Then ker(Q) = M =
range(P ) and range(Q) = N = ker(P ). Since P is continuous, Q is continuous, and
therefore its kernel M is closed.

Suppose now that both M and N are closed. Since X is a Banach space, and M
is closed in X, M is a Banach space. By the Closed Graph Theorem, it suffices to
prove that T is closed. So, let xk → x and P xk → y. We must show that y = P x.
Note that xk − P xk ∈ N for all k (since P xk strips off the “M ” part of xk ). Since
N is closed, xk − P xk → x − y ∈ N . Hence, P (x − y) = 0. On the other hand,
P (x − y) = P x − P y so that P x = P y. Now, M is closed so that P xk → y ∈ M , and
therefore P y = y. It follows that y = P x.

33. Let X be a Banach space, Y a NLS, and Tn ∈ B(X, Y ) such that {Tn x}∞ n=1
is a Cauchy sequence in Y . Show that {kTn k}∞n=1 is bounded. If, in addition, Y is a
Banach space, show that if we define T by Tn x → T x, then T ∈ B(X, Y ).

Solution: Since each sequence {Tn x} is Cauchy in Y , they are bounded. Hence,
none of them are unbounded. By PUB, it follows that {kTn k} is uniformly bounded;
say, by M . Then,

kT xk = lim kTn xk ≤ lim kTn kkxk ≤ lim M kxk = M kxk

n→∞ n→∞ n→∞

So, T is bounded.

34. Let X be the normed space of sequences of complex numbers x = {xi }∞ i=1
with only finitely many nonzero terms and norm defined by kxk = supi |xi |. Let
T : X → X be defined by
 
1 1
y = T x = x1 , x2 , x3 , ... .
2 3

Show that T is a bounded linear map, but that T −1 is unbounded. Why does this
not contradict the Open Mapping Theorem?

Solution: If x = 0, then clearly kT xk ≤ M kxk for any M . Suppose x 6= 0. Since

only finitely many terms are nonzero, it follows that kxk = maxi |xi |, and is actu-
ally attained by one of the xi ; say, xn . Then, |xi | ≤ |xn | for all i. It follows that
1/i|xi | ≤ |xn |, so that kT xk ≤ kxk. Hence T has norm at most one. This is actually
the best we can do, for example by taking the sequence x = (1, 0, 0...), for which
T x = x.

The inverse map T −1 is defined by

T −1 x = {x1 , 2x2 , 3x3 , ...} .

 M383C Exercises

Consider the sequence (xk )i = δik . Then, kT −1 xk k = k whereas kxk k = 1 for all k.

This does not contradict the Open Mapping Theorem since X is not a Banach space.
E.g., take xn = (1, 1/2, 1/3, ..., 1/n, 0, 0, ...). Then {xn } is a Cauchy sequence in X,
but does not converge in X; the limit is x = (1, 1/2, 1/3, ...), all of whose entries are
nonzero.

35. Give an example of a function that is closed but not continuous.

Solution: Consider f (x) = 1/(x2 + 1). Then f (R) = (0, 1], which is neither open
nor closed.

36. For each α ∈ R, let Eα be the set of all continuous functions f on [−1, 1] such
that f (0) = α. Show that the Eα are convex, and that each is dense in L2 ([−1, 1]).

Solution: Let f, g ∈ Eα . Let t ∈ [0, 1]. Then,

(tf + (1 − t)g)(0) = tf (0) + (1 − t)g(0) = tα + (1 − t)α = α,

so that tf + (1 − t)g ∈ Eα . Thus, the Eα are convex.

To prove density, we use the fact that a subset of a Hilbert space is dense iff its
orthogonal complement is {0}. The inner product on L2 is precisely
Z 1
hf, gi = f g dx.
−1

The orthogonal complement of Eα is defined by

Eα⊥ = f ∈ L2 [−1, 1] | hf, gi = 0 for all g ∈ Eα .



Note that the polynomials gn = xn + α are all in Eα . Let f ∈ Eα⊥ . Then,

Z 1 Z 1
n
hf, gn i = x f dx + α f dx = 0.
−1 −1

Note that, in particular hf, g0 i = 0 so that

Z 1
xn f dx = 0
−1

for all n. By Stone-Weierstrass, this implies that f = 0.

37. Suppose that X, Y, and Z are Banach spaces and that T : X × Y → Z is

bilinear and continuous. Prove that there is a constant M < ∞ such that

kT (x, y)k ≤ M kxkkyk for all x ∈ X, y ∈ Y.

 M383C Exercises

Is completeness needed here?

Solution: Fix a y ∈ Y and define Ty : X → Z by Ty (x) = T (x, y). Then, as a

restriction of a continuous map, Ty is continuous. It follows that

kT (x, y)k = kTy (x)k ≤ kTy kkxk

Next, consider the set {Ty | y ∈ Y }. Then, this set is pointwise bounded by continuity
of T in the first variable. That is, defining Tx (y) := T (x, y), we have

sup kTy (x)k = sup kT (x, y)k = sup kTx (y)k ≤ kTx k.
y∈B1Y (0) y∈B1Y (0) y∈B1Y (0)

So, by PUB, the Ty are uniformly bounded; say, by M . Since the above was for y
with norm 1, it follows that kTy k ≤ M kyk. Hence,

kT (x, y)k ≤ kTy kkxk ≤ M kxkkyk.

Completeness is not necessary here. It is only necessary if T is instead continuous in

each variable separately (in the product topology, this implies that T is continuous
on the product space, but the topology on a normed product space need not be the
product topology).

38. Prove that a bilinear map is continuous if it is continuous at the origin (0, 0).

Solution: We show that continuity at the origin implies boundedness, and that bound-
edness implies continuity everywhere.
Suppose T is a bilinear map continuous at the origin. Then there exist δ1 , δ2 > 0 such
that if kxk < δ1 and kyk < δ2 ,

kT (x, y)k = kT (x, y) − T (0, 0)k < 1.

Now let x ∈ X and y ∈ Y . Set x̃ = x/kxkδ1 and ỹ = y/kykδ2 . Then,

1
kT (x̃, ỹk ≤ 1 ⇒ kT (x, y)k ≤ kxkkyk.
δ1 δ2
Then, T is bounded.

Suppose now that T is bounded by C > 0. Let  > 0 and fix (x0 , y0 ) ∈ X × Y .
Set δ = min(1, /(C(1 + kx0 k + ky0 k))). Then, if kx − x0 k, ky − y0 k < δ, we have

kT (x, y) − T (x0 , y0 )k = kT (x − x0 , y0 ) + T (x − x0 , y − y0 ) + T (x0 , y − y0 )k

≤ Ckx − x0 kky0 k + Ckx − x0 kky − y0 k + Ckx0 kky − y0 k
< Cδ(δ + kx0 k + ky0 k) ≤ Cδ(1 + kx0 k + ky0 k) ≤ 

So, T is continuous at (x0 , y0 ).

 M383C Exercises

39. Consider X = C[a, b], the continuous functions defined on [a, b] with the maxi-
w
mum norm. Let {fn }∞ n=1 be a sequence in X and suppose that fn −
→ f . Prove that
{fn }∞
n=1 is pointwise convergent. That is,

fn (x) → f (x) for all x ∈ [a, b].

Prove that a weakly convergent sequence in C 1 [a, b] is convergent in C[a, b]. (∗) Is
this still true when [a, b] is replaced by R?
w
→ f weakly if, for all Λ ∈ C[a, b]∗ we have
Solution: By definition, fn −

Λ(fn ) → Λ(f ).

By Proposition 2.91, supn kfn k < ∞. For x ∈ [a, b], let δx be the Dirac measure.
Observe that δx ∈ C[a, b]∗ via
Z b
δx (f ) = f δx = f (x)
a

Hence, δx (fn ) → δx (f ), which precisely means fn (x) → f (x).

Let fn be a weakly convergent sequence in C 1 [a, b]. By Banach Saks, we may upgrade
this to a strongly convergent sequence.
NEED TO FINISH THIS

40. Let X be a normed linear space and Y a closed subspace. Show that Y is
weakly sequentially closed.

Solution: Since Y is convex (it’s a subspace), Corollary 2.104 applies and we see
that the weak and strong closures of Y are equal. But, Y is closed, so the weak
w
closure of Y is Y . Hence Y is weakly closed. Now let yn − → y. Let U be an open
neighborhood of y. Then some tail lies in U , and hence U ∩ Y \ {y} = 6 ∅. Thus
y ∈ cl(Y )w . (It is actually true for any topological space that closed implies sequen-
tially closed).

41. Let X be a normed linear space. We say that a sequence {xn }∞ n=1 ⊂ X is
∞ ∗
weakly Cauchy if {T xn }n=1 is Cauchy for all T ∈ X , and we say that X is weakly
complete if each weak Cauchy sequence converges weakly. If X is reflexive, prove that
X is weakly complete.

Solution: Consider the sequence {Exn }. Since {T xn } is Cauchy in F, it is bounded.

However, T xn = Exn (T ), so that {Exn T } is bounded over xn . It follows by PUB
that {Exn } are uniformly bounded. But, kExn k = kxn k, so that the xn are uniformly
bounded, say by M < ∞. Since each sequence {T xn } is Cauchy in a complete space,
it converges. Define ϕ : X ∗ → F by

ϕ(T ) = lim T (xn ) = lim Exn (T ).

n→∞ n→∞
 M383C Exercises

The map ϕ is linear, and bounded since

kϕ(T )k ≤ lim kxn kkT k = lim sup kxn kkT k ≤ lim M kkT k = M kT k
n→∞ n→∞ n n→∞

So, ϕ ∈ X ∗∗ . Since X is reflexive, there exists an x ∈ X such that Ex = ϕ. But, then

T (x) = ϕ(T ) = lim T (xn )

n→∞

w
so that xn −
→ x.

42. Show that every finite dimensional vector space is reflexive.

Solution: Let X be a d dimensional normed space with d < ∞. Given a basis

{ei }di=1 , we may define a basis on X ∗ by

ei (ej ) = δji

and extending ei linearly to all of X. Now, we can repeat the same thing for X ∗ and
deduce that X ∗∗ has dimension d. Thus, F is a linear, injective map between two d
dimensional spaces. It follows that F is surjective, hence X is reflexive.

43. Show that C[0, 1] is not reflexive.

Solution: Suppose C[0, 1] is reflexive. By Banach-Alaoglu, we see that the unit ball
B1∗∗ is weak-∗ (hence weakly) compact. Since C[0, 1] is reflexive, we see that B1 is
also weakly compact. Hence, for any ϕ ∈ C[0, 1]∗ , there exists an f ∈ C[0, 1] such
that ϕ attains its norm at f (that is, ϕ(f ) = kϕk). Now consider the functional
Z 1/2 Z 1
ϕ(f ) = f dx − f dx
0 1/2

Let us limit our attention to functions with kf k = 1; that is, f is essentially bounded
between −1 and 1. The way to optimize ϕ is to choose the function f = χ[0,1/2] −
χ[1/2,1] , for which ϕ(f ) = 1 = kϕk. Any other function decreases ϕ in one way or
another (cancellation of area, smaller function, etc.) But, the optimal f is not in
C[0, 1]. See Exercise 21 for further discussion.

44. If X and Y are Banach spaces, show that E ⊂ B(X, Y ) is equicontinuous if,
and only if, there is an M < ∞ such that kT k ≤ M for all T ∈ E.

Solution: Suppose first E is equicontinuous. Then, in particular, it is equicontinuous

at 0. Let  = 1. Then there exists a δ > 0 such that if kxk < δ then kT xk <  = 1
for all T ∈ E. Now let x ∈ X. Then,
 
kxk x kxk
kT xk = T δ ≤
δ kxk δ
 M383C Exercises

Then, kT k ≤ 1/δ.
Now suppose the T are uniformly bounded by M . Let  > 0. Set δ = /M . Fix an
x0 ∈ X. Then, if x ∈ X is such that kx − x0 k < δ,

kT x − T x0 k = kT (x − x0 )k ≤ M kx − x0 k < M δ < 

and hence E is equicontinuous at x0 .

45. Let X be a Banach space and T ∈ X ∗ = B(X, F). Identify the range of
T ∗ ∈ B(F, X ∗ ).

Solution: The adjoint is defined by

T ∗ α(x) = α(T x) = (αT )(x)

for α ∈ F. So, the action of T ∗ is precisely to multiply T by some scalar. Thus,

range(T ) = FT ⊂ X ∗ . It should be noted that the dual of F is the set of continuous
functions from F → F. So, really T ∗ should multiply T by such a function. In which
case, the range is actually C[F]T .

46. Let X be a Banach space, S, T ∈ B(X, X), and I be the identity map.
a) Show by example that ST = I does not imply T S = I.
b) If T is compact, show that S(I − T ) = I if, and only if, (I − T )S = I.
c) If S = (I − T )−1 exists for some T compact, show that I − S is compact.
Solution:
a) Let X = l1 . Consider the linear operator T defined by

T (x) = T (x1 , x2 , x3 , ...) = (0, x1 , x2 , x3 , ...)

and the linear operator S defined by

S(x) = S(x1 , x2 , x3 , ...) = (x2 , x3 , ...).

Then, ST (x) = x, but T S(x) = (0, x2 , x3 , ...) = x − x1 e1 . Note that this occurs
since T is injective but not surjective, while S is surjective but not injective.
Such an event can only happen if X is infinite dimensional.
b) An operator T is compact if, for every bounded subset B, the image T (B) has
compact closure. Alternatively, T is compact iff for every sequence {xn }∞ n=1 ⊂
X, {T xn } has a convergent subsequence.
It takes a little work to set up, but there exists a theorem called the Fredholm
alternative. It is stated here:

Proposition: Let T be a compact operator on a Banach space X. Then ex-

actly one of the following holds
 M383C Exercises

i) I − T is invertible with bounded inverse

ii) There exists a nonzero x ∈ X such that T x = x (i.e., 1 is an eigenvalue)
We show that I − T is invertible. Suppose that S(I − T ) = I. Suppose by
way of contradiction that there exists a nonzero x such that T x = x. Then,
S(I − T )x = S(x − x) = S(0) = 0. On the other hand, S(I − T )x = Ix = x,
which implies x = 0. This is a contradiction. The converse is similar (may need
to use adjoints?)

So, in either case I − T is invertible with bounded inverse (by applying the
open mapping principle). It follows in either case that S is the inverse of I − T ,
and hence the other equality holds.
c) Let A = I − S = I − (I − T )−1 . Then, A(I − T ) = (I − T ) − I = −T . Hence,
A = −(I − T )−1 T , which is the product of a compact and bounded operator
(hence compact). To see this, suppose B is the closed unit ball. Let T compact
and S bounded; we show ST is compact. Since S is bounded, it is continuous.
Since T is compact, cl(T B) is compact. Hence, the image under S is compact,
and ST is a compact operator.

For more facts on compact operators see: Appendix

47. Let 1 ≤ p < ∞ and define, for each r ∈ Rd , Tr : Lp (Rd ) → Lp (Rd ) by
Tr (f )(x) = f (x + r).
a) Verify that Tr (f ) ∈ Lp (Rd ) and that Tr is bounded and linear. What is the
norm of Tr ?
b) Show that as r → s, kTr f − Ts f kp → 0. Hint: Use that the set of continuous
functions with compact support are dense in Lp (Rd ) for p < ∞.
Solution:
a) Linearity follows easily by the vector space structure of Lp (Rd ). Let f ∈ Lp (Rd ).
Then, by the translation invariance of Lebesgue measure,
Z 1/p Z 1/p
p p
kTr (f )kp = |f (r + x)| dx = |f (x)| dx = kf kp .
Rd Rd

It follows from this that Tr is bounded with norm 1.

b) Show that as r → s, kTr f − Ts f kp → 0. Hint: Use that the set of continuous
functions with compact support are dense in Lp (Rd ) for p < ∞.
Let f ∈ Lp (Rd ). Then there exists a sequence of continuous functions with
compact support fn such that fn → f . Then,
lim Tr f = lim lim Tr fn = lim lim Tr fn = lim Ts fn = Ts f,
r→s r→s n→∞ n→∞ r→s n→∞

which follows by continuity of fn and the fact that fn → f . By continuity of

the norm, the conclusion follows.