1.

Banach spaces
Definition 1.1. A B ∗ -space or a normed space is a vector space X (over R or C) equipped
with a norm ∥·∥. A B-space or a Banach space is a complete B ∗ -space.
Exercise 1.4.7. Let X be a B ∗ space. Then, X is a Banach Pspace if and only if for every
sequence {xn }n∈N ⊂ X satisfying ∞ ∞
P
n=1 ∥xn ∥ < ∞ the sum n=1 x n is convergent.
Pn
P∞ that X is a Banach space. Denote by sn the finite
Proof. Step 1. Assume P sum k=1 xk .
For any ϵ > 0, as n=1 ∥xn ∥ < ∞, there is an N (ϵ) ∈ N such that n≥N (ϵ) ∥xn ∥ < ϵ.
Then, for any n, m ≥ N (ϵ), we have
X
∥sn − sm ∥ ≤ ∥xn ∥ < ϵ,
n≥N (ϵ)

that is, {sn }n∈N is a Cauchy sequence

P∞ in X . Since X is complete, {sn }n∈N converges to
some s ∈ X , which is the desired n=1 xn .

Step 2. Let {xn }n∈N be an arbitrary Cauchy sequence in X . Then, for every k ∈ N, there
is an N (k) ∈ N such that
1
∥xm − xn ∥ < k for any m, n ≥ N (k).
2
Notice that
xN (k) = (xN (k) − xN (k−1) ) + (xN (k−1) − xN (k−2) ) + · · · + (xN (2) − xN (1) ) + xN (1) ,
and that the subsequence {xN (k) }k∈N satisfies that
∞ ∞
X X 1
xN (k+1) − xN (k) < < ∞.
2k
k=1 k=1
By assumption, the sequence 
xN (k) − xN (1) k∈N
is convergent, hence the subsequence {xN (k) }k∈N is convergent. As {xn }n∈N is a Cauchy
sequence, it converges to the same limit of the subsequence {xN (k) }k∈N . □
Exercise 1.4.11. Let X be a normed space. A function φ : X → R is called convex if
φ(λx + (1 − λ)y) ≤ λφ(x) + (1 − λ)φ(y) for ∀ x, y ∈ R and ∀ 0 ≤ λ ≤ 1.
Show that a local minimum of φ must be a global minimum of φ.
Proof. Let x0 be a local minimum of φ. Then, there exists an ϵ > 0 such that
φ(x0 ) = min φ(x)
x∈B(x0 ,ϵ)

where B(x0 , ϵ) is the open ball centred at x0 with radius ϵ. Let y0 be a global minimum of
φ, that is,
φ(y0 ) = min φ(x).
x∈X
For every 0 ≤ λ ≤ 1, as φ is convex, we have
φ(y0 ) ≤ φ(λx0 + (1 − λ)y0 ) ≤ λφ(x0 ) + (1 − λ)φ(y0 ).
Pick λ sufficiently close to 1, then λx0 + (1 − λ)y0 belongs to the open ball B(x0 , ϵ). Since
φ(x0 ) is a local minimum, we get
φ(x0 ) ≤ φ(λx0 + (1 − λ)y0 ) ≤ λφ(x0 ) + (1 − λ)φ(y0 ),
1
2

which implies that

φ(x0 ) ≤ φ(y0 ),
hence φ has a global minimum φ(x0 ) at x0 . □
Exercise 1.4.14. Denote by C0 the set of all sequences of real numbers whose limits are
0, i.e.,
C0 := {x = {ξn }n∈N ∈ R∞ | lim ξn = 0}.
n→∞
Let ∥·∥ be the norm on C0 defined as follows:
∥x∥ = max |ξn | for every x = {ξn }n∈N ∈ C0 .
n≥1

Set  ∞ 
X ξn
M := x = {ξn }n∈N ∈ C0 =0 .
2n
n=1
(1) Show that M is a closed linear subspace of C0 .
(2) Let x0 = (2, 0, . . . , 0, . . . ). Show that
inf ∥x0 − x∥ = 1,
x∈M
and that ∥x0 − y∥ > 1 for every y ∈ M .
Proof. Step 1. For (1), the verification that M is a linear subspace of C0 is trivial, so it
suffices to show that M is a closed subset of C0 . Let xn = {ξnk }k∈N be a sequence in M
such that
lim xn = x ∈ C0 .
n→∞
We shall show that x is also contained in M , and so M is closed. Write x = {ξn }n∈N , then
it suffices to show that
∞
X ξn
= 0.
2n
n=1
For an arbitrary ϵ > 0, there exits an N (ϵ) ∈ N such that
∞
X 1 ϵ
n
< .
2 2
n=N (ϵ)+1

We can also assume that ∥xn − x∥ < 1 for every n ≥ N (ϵ). Moreover, there exists an
N (ϵ)′ ≥ N (ϵ) such that ∥xn − x∥ < ϵ/2N (ϵ) for every n ≥ N (ϵ)′ , that is,
ϵ
max |ξnk − ξk | < for every n ≥ N (ϵ)′ .
k∈N 2N (ϵ)
Thus, for every n ≥ N (ϵ)′ , we have
∞ ∞ N (ϵ) ∞
X ξnk X ξk X ξnk − ξk X ξnk − ξk
− = +
2k 2 k 2k 2k
k=1 k=1 k=1 k=N (ϵ)+1
N (ϵ) ∞
X |ξnk − ξk | X 1
≤ +
2k 2k
k=1 k=N (ϵ)+1
N (ϵ)
X ϵ ϵ
< + = ϵ.
2N (ϵ) 2
k=1
 3
P∞ k
However, as xn ∈ M , the sum k=1 ξnk /2 is equal to 0, hence, for every ϵ > 0,
∞
X ξk
< ϵ,
2k
k=1
P∞ k
which is equivalent to saying that k=1 ξk /2 = 0.

Step 2. We prove (2) in this step. Assume that ∥x0 − y∥ < 1 for some y = {ξn }n∈N ∈ M .
Then, |ξn | < 1 for any n ≥ 2 and |ξ1 | > 1, which implies that
∞ ∞
|ξ1 | 1 X ξn X |ξn | 1
(1) > and n
≤ n
< ,
2 2 2 2 2
n=2 n=2

hence it is not possible to make y ∈ M . If ∥x0 − y∥ = 1 for some y = {ξn }n∈N , then |ξ1 | ≥ 1
and |ξn |n∈N ≤ 1 for any n ≥ 2. If |ξ1 | > 1 or |ξn | < 1 for some n ≥ 2, the same reason as
(1) shows that y ̸∈ M , so we can conclude that |ξn | = 1 for any n ∈ N. This contradicts
the hypothesis that limn→∞ ξn = 0, hence y ̸∈ M . Thus, we have ∥x0 − y∥ > 1 for every
y ∈ M . Moreover, set
1
xn = (1 − , −1, . . . , −1, 0, . . . , 0, · · · ) ∈ C0 .
2n | {z }
n copies

It is easy to see that

n+1
1 X 1
(1 − ) − = 0,
2n 2k
k=2
so xn ∈ M for every n ∈ N. Furthermore, since
1
∥x0 − xn ∥ = 1 + → 0 as n → ∞,
2n
we see that inf x∈M ∥x0 − x∥ = 1 as required. □

2. The Minkowski functional (gauge)

Recall that definition of Minkowski functional.
Definition 2.1. Let X be a linear space (over R or C). Let C ⊂ X be a convex subset
such that 0 ∈ C. The Minkowski functional associated to C (or the gauge of C) is defined
as the function valued in the extended real numbers:
 
x
PC : X → R ∪ {+∞}, x 7→ inf λ > 0 ∈C .
λ
Definition 2.2. Let (X , ρ) be a metric space and E ⊂ X a subset. We say that x is an
interior point of E if there exists BX (x, ϵ) for some ϵ > 0 such that BX (x, ϵ) ⊂ E, where
BX (x, ϵ) := {y ∈ X | ρ(x, y) < ϵ}.
Denote by int(E) the set of all the interior points of E.
Lemma 2.3. Let X be a B ∗ -space. Let C be a convex subset of X such that 0 ∈ int(C).
Then, the Minkowski functional PC associated to C is R-valued and uniformly continuous.
4

Proof. Since 0 ∈ int(C), there exists an r > 0 such that BX (0, r) ⊂ C. Then, for every
0 ̸= x ∈ X , we have
rx
∈ C.
2 ∥x∥
Hence, C is an absorbing subset of X , which implies that PC is an R-valued function, and
2 ∥x∥
PC (x) ≤ for every x ∈ X \ {0}.
r
For an arbitrary pair of points x, y ∈ X , since
PC (x) ≤ PC (x − y) + PC (y), and
PC (y) ≤ PC (y − x) + PC (x),
we can conclude that
2
|PC (x) − PC (y)| ≤ max{PC (x − y), PC (y − x)} ≤ · ∥x − y∥ .
r
Thus, PC : X → R is uniformly continuous. □
Exercise 1.5.1. Let X be a normed space, and let E be a proper, convex subset of X
containing 0 ∈ X as an interior point, i.e., 0 ∈ int(E). Denote by PE the Minkowski
functional associated to E. Show that
(1) x ∈ int(E) ⇐⇒ PE (x) < 1, and that
(2) int(E) = E.
Proof. Step 1. First, assume that x ∈ int(E). Then, there exists an ϵ > 0 such that
BX (x, ϵ) ⊂ E, which implies that
x
∈E
1−δ
for some 0 < δ ≪ 1, hence PE (x) < 1. Conversely, assume that PE (x) < 1. By Lemma 2.3,
PE is continuous, so there exists an r > 0 such that
BX (x, r) ⊂ (PE )−1 (−∞, 1) ,

that is, for every y ∈ BX (x, r), we have PE (y) < 1. Set py := 1/PE (y), then py > 1, and
there exists δ > 0 such that
py − δ > 1 and that (py − δ)y ∈ E.
As C is convex and 0 ∈ C, we see that y ∈ E, i.e., the whole BX (x, r) is contained in E.
Thus, x ∈ int(E). This shows (1).

Step 2. Pick x ∈ int(E), then there exists a sequence of points xn ∈ int(E), n ∈ N, such
that limn→∞ xn = x. Since PE is continuous by Lemma 2.3, PE (x) ≤ 1 as PE (xn ) < 1 for
every xn by Step 1. On the other hand, assume that PE (x) = 1. If x ∈ E, then there is a
sequence of points yn on the line segment L0,x connecting 0 and x such that
x
lim yn = x and yn := for some αn > 1.
n→∞ αn
Then, we have that
1 1
PE (yn ) = · PE (x) = < 1.
αn αn
By (1), every yn belongs to int(E), and so x ∈ int(E) by the choice of {yn }n∈N . If x ̸∈ E,
there exists a sequence of real numbers λn > 1 such that
x
lim λn = 1 and xn := ∈ E for every n ∈ N.
n→∞ λn
 5

Again, we see that

 
x 1 1
PE (xn ) = PE = · PE (x) = < 1,
λn λn λn
which shows that xn ∈ int(E) for every n ∈ N. We can also conclude that x ∈ int(E).
Therefore, int(E) is equal to the set C := {x ∈ X | PE (x) ≤ 1}.

Step 3. As int(E) ⊂ E, we have int(E) ⊆ E. Conversely, pick an arbitrary point x ∈ E,

then there exists a sequence of points xn ∈ E such that limn→∞ xn = x. By Lemma 2.3,
we have
PE (x) = lim PE (xn ) ≤ 1
n→∞
since PE (xn ) ≤ 1 for every n ∈ N. Thus, x ∈ C, and so E ⊆ C. By Step 2, this is equivalent
to saying that E ⊆ int(E), hence int(E) = E. □

Santai Qu
Institute of Geometry and Physics, University of Science and Technology of China,
Hefei, Anhui Province, China
Email: santaiqu@ustc.edu.cn