TOPICS FROM ANALYSIS - PART 1/3

T.K.SUBRAHMONIAN MOOTHATHU

Contents

1. Michael’s selection theorem 1

2. Schauder bases in Banach spaces 5
3. Unconditional bases and orthogonal systems in Banach spaces 10
4. Convexity, smoothness, and equivalent norms 13
5. Krein-Milman theorem 19

A few advanced topics from Analysis will be briefly touched upon in this course. In this first
part, we will discuss certain topics from Functional Analysis. In the second and third parts of the
note, we will concentrate mainly on certain topics related to Measure Theory. Therefore,

Pre-requisites: Functional Analysis and Measure Theory.

For this first part of the note, some basic references are the following books: (i) Benyamini
and Lindenstrauss, Geometric Functional Analysis - I, (ii) Fabian et al, Functional Analysis and
Infinite Dimensional Geometry, (iii) Lindenstrauss and Tzafriri, Classical Banach Spaces - I, (iv)
Rudin, Functional Analysis.

Topological spaces considered in this note are assumed to be Hausdorff, unless specified otherwise.

1. Michael’s selection theorem

Question: Let P(Y ) denote the power set, i.e., set of all subsets, of a given set Y . Let X, Y be
topological spaces, and F : X → P(Y ) \ {∅} be a function (often F is called a multifunction from
X to Y ). When can we get a continuous map f : X → Y such that f (x) ∈ F (x) for every x ∈ X?

Example: Let X be the closed unit disc in R2 , Y be the unit circle in R2 , and F : X → P(Y ) be
F (x) = {x} if x ∈ Y and F (x) = Y if x ∈ X \ Y . Then F does not have a continuous selection,
i.e., there does not exist continuous f : X → Y with f (x) ∈ F (x) for every x ∈ X, since the unit
circle is not a retract of the closed unit disc (a subset Y of a topological space X is called a retract
of X if there is a continuous map f : X → Y with f |Y = IY ).
1
2 T.K.SUBRAHMONIAN MOOTHATHU

Example: Let X1 ⊂ X2 and Y be topological spaces, and g : X1 → Y be a continuous map that

does not have a continuous extension ge : X2 → Y (for instance, g : (0, 1) ⊂ [0, 1] → R given by
g(x) = 1/x does not have a continuous extension to [0, 1]). Let F : X1 → P(Y ) be F (x) = {g(x)}
if x ∈ X1 and F (x) = Y if x ∈ X2 \ X1 . Then F does not have a continuous selection.

Michael’s theorem will give a sufficient condition for the existence of a continuous selection. The
proof will use the notion of a partition of unity on a paracompact space.

Definition: Let X be a topological space. A collection {Aj : j ∈ J} of subsets of X is said to be

locally finite if every x ∈ X has a neighborhood that intersects only finitely many Aj ’s. We say X
is paracompact if every open cover of X has a locally finite open refinement covering X. Clearly,
compact (Hausdorff) spaces are paracompact.

Exercise-1: [Stone’s theorem] Every metric space X is paracompact. [Hint: First suppose X is
separable, and let U be an open cover for X. Each point in X has a neighborhood whose closure
is contained in some U ∈ U. Hence we can find a countable open cover {Vn : n ∈ N} of X such
∪n−1
that for each n ∈ N there is Uσ(n) ∈ U with Vn ⊂ Uσ(n) . Let W1 = Uσ(1) and Wn = Uσ(n) \ j=1 Vj
for n > 1. If x ∈ X, and nx ∈ N is the smallest with x ∈ Vnx ⊂ Uσ(nx ) , then x ∈ Wnx and hence
W := {Wn : n ∈ N} is an open cover for X, refining U. If x ∈ X, and mx ∈ N is the smallest with
x ∈ Vmx , then the neighborhood Vmx of x does not intersect Wn for n > mx , and thus W is locally
finite. For a proof for the general case, see the articles M.E.Rudin, A new proof that metric spaces
are paracompact, and D.Ornstien, A new proof of the paracompactness of metric spaces.]

Definition: Let X be a topological space and gj : X → [0, 1] be continuous for j ∈ J. We say

{gj : j ∈ J} is a partition of unity on X if the collection {supp(gj ) : j ∈ J} is locally finite, and
∑
if j∈J gj (p) = 1 for every p ∈ M , where supp(gj ) = {x ∈ X : gj (x) ̸= 0}. A partition of unity
{gj : j ∈ J} on X is said to be subordinate to an open cover U of X if for each j ∈ J, there is
U ∈ U with supp(gj ) ⊂ U .

Exercise-2: [Existence of partition of unity] Let X be a paracompact space and U be an open cover
of X. Then there is a partition of unity on X subordinate to U. [Hint: See my notes Introduction
to Manifolds.]

Let X, Y be topological spaces. If g : X → Y is a discontinuous function and if F : X → P(Y ) is

defined as F (x) = {g(x)}, then evidently F has no continuous selection. Therefore, when we look
for a continuous selection for a multifunction F , we should impose some sort of continuity on F .

Definition: Let X, Y be topological spaces and F : X → P(Y ) be a function. We say F is inner semi-
continuous (or upper semi-continuous) if for every open set V ⊂ Y , the set {x ∈ X : F (x) ⊂ V }
 TOPICS FROM ANALYSIS - PART 1/3 3

is open in X. We say F is outer semi-continuous (or lower semi-continuous) if for every open set
V ⊂ Y , the set {x ∈ X : F (x) ∩ V ̸= ∅} is open in X.

Example: Let X be a topological space and Y ⊂ X. Define F : X → P(Y ) as F (y) = {y} for y ∈ Y
and F (x) = Y for x ∈ X \ Y . If Y is open in X, then F is upper semi-continuous. If Y is closed
in X, then F is lower semi-continuous.

[101] [Michael’s selection theorem] Let X be a paracompact space, Y be a Banach space, and
F : X → P(Y ) be a lower semi-continuous function such that F (x) is a nonempty closed convex
subset of Y ∀ x ∈ X. Then ∃ a continuous map f : X → Y with f (x) ∈ F (x) for every x ∈ X.

Proof. Step-1 : [approximate selection] We claim that given ε > 0 there is a continuous map
h : X → Y with dist(h(x), F (x)) < ε for every x ∈ X. Note that U := {Uy : y ∈ Y } is an
open cover for X if we we put Uy = {x ∈ X : F (x) ∩ B(y, ε) ̸= ∅}. Let {gj : j ∈ J} be a partition of
unity on X subordinate to U, and assume supp(gj ) ⊂ Uyj . Define the continuous map h : X → Y
∑
as h(x) = j∈J gj (x)yj . For x ∈ X, let Jx = {j ∈ J : gj (x) ̸= 0}, which is a finite set. We have
dist(yj , F (x)) < ε for j ∈ Jx . Since F (x) is convex, the ε-neighborhood B(F (x), ε) of the set F (x)
is also convex and hence h(x) ∈ B(F (x), ε) since h(x) is a convex combination of yj ’s with j ∈ Jx .

Step-2 : Let F0 = F and h0 be the approximate selection for F0 given by Step-1 for ε = 1. Check
that F1 : X → P(Y ) defined as F1 (x) = F0 (x) ∩ B(h0 (x), 1) satisfies the hypothesis of F0 . Then by
Step-1 applied with ε = 2−1 , we get a continuous map h1 : X → Y with dist(h1 (x), F1 (x)) < 2−1
for every x ∈ X. Next define F2 : X → P(Y ) as F2 (x) = F0 (x) ∩ B(h1 (x), 2−1 ), and obtain by
Step-1 a continuous map h2 : X → Y with dist(h2 (x), F2 (x)) < 2−2 for every x ∈ X, and so on.
The sequence (hn ) of continuous maps from X to Y satisfies ∥hn (x) − hn−1 (x)∥ < 2−n+1 + 2−n and
dist(hn (x), F (x)) < 2−n for every x ∈ X. The first property implies that (hn (x)) is Cauchy in the
Banach space Y so that the limit f (x) := limn→∞ hn (x) exists and that f : X → Y defined in this
manner is continuous (verify). The second property implies f (x) ∈ F (x) since F (x) is closed. 

Definition: Let X be a topological space and f : X → R be a map. We say f : X → R is upper

semi-continuous (respectively, lower semi-continuous) if for every c ∈ R, the set {x ∈ X : f (x) < c}
(respectively, {x ∈ X : f (x) > c}) is open in X. Note that f : X → R is lower semi-continuous iff
the multifunction x 7→ (−∞, f (x)] is lower semi-continuous as per our earlier definition.

Example: (i) f : R → R given by f (x) = ⌊x⌋ (integer part of x) is upper semi-continuous. (ii) Let
X be a topological space, U ⊂ X be open, and A ⊂ X be closed. Then the indicator function 1U
is lower semi-continuous and 1A is upper semi-continuous.
4 T.K.SUBRAHMONIAN MOOTHATHU

[102] [Some direct applications of Michael’s selection theorem] (i) Let X be a paracompact space,
and f1 , f2 : X → R be functions such that f1 is upper semi-continuous, f2 is lower semi-continuous,
and f1 ≤ f2 . Then there is a continuous function f : X → R with f1 ≤ f ≤ f2 .
(ii) Let X be a Banach space and K ⊂ X be a nonempty closed convex subset. Then K is a retract
of X, i.e., there is a continuous map f : X → K such that f |K = IK .
(iii) [A sort of generalization of Tietze extension theorem] Let A be a closed subset of a paracompact
space X, and f : A → Y be a continuous map into a Banach space Y . If K ⊂ Y is the closed
convex hull of f (A), then f has a continuous extension fe : X → K.
(iv) [Existence of continuous right inverse] Let T : Y → X be a surjective bounded linear operator
of Banach spaces. Then there is a continuous map f : X → Y such that T ◦ f = IX .

Proof. Verify that each multifunction F given below is lower semi-continuous with F (x) closed and
convex for each x. Then apply [101]:
(i) Let F : X → P(R) be F (x) = [f1 (x), f2 (x)].
(ii) Let F : X → P(X) be F (x) = {x} for x ∈ K, and F (x) = K for x ∈ X \ K.
(iii) Let F : X → P(Y ) be F (x) = {f (x)} for x ∈ A, and F (x) = K for x ∈ X \ A.
(iv) Let F : X → P(Y ) be F (x) = T −1 (x). 

Remark: Let X be paracompact, and A ⊂ U ⊂ X, where A is closed and U is open in X. Applying

[102](i) with f1 = 1A and f2 = 1U , we get a continuous map f : X → [0, 1] such that f (A) = {1}
and f (X \ U ) = {0}.

Next we present a slightly non-trivial application about continuous extensions - this is also a
generalization of Tietze extension theorem, in the special case of compact metric spaces. Let X
be a compact metric space, and C(X, R) denote the Banach space of all continuous real-valued
functions on X, where the norm is the supremum norm.

[103] Let ∅ ̸= A ⊂ X be closed in a compact metric space X. Then there is a linear isometry
T : C(A, R) → C(X, R) such that T 1 = 1 and (T g)|A = g for every g ∈ C(A, R).

Proof. Using Riesz representation theorem, the set Γ of all Borel probability measures on A can
be thought of as a subspace of the unit ball of the dual space C(A, R)∗ . Then Γ is compact and
convex w.r.to the weak* topology (the smallest topology such that the evaluation maps ϕ 7→ ϕ(a)
on C(A, R)∗ are continuous for each a ∈ A). Moreover, if {fn : n ∈ N} is a countable dense subset
of the unit ball of the separable space C(A, R), the affine homeomorphism µ 7→ (2−n ⟨µ, fn ⟩) =
∫
(2−n A fn dµ) identifies (Γ, weak*) with a compact convex subset of the closed unit ball Bl2 of l2 .
Through this identification, we assume Γ ⊂ Bl2 ⊂ l2 . For a ∈ X, let µa ∈ Γ denote the pointed
measure on A at a. We have µa (B) = 1 if a ∈ B and µa (B) = 0 if a ∈ A \ B for any Borel
 TOPICS FROM ANALYSIS - PART 1/3 5
∫
set B ⊂ A. Therefore, µa (g) = A g(x)dµa (x) = g(a) for g ∈ C(A, R). We claim that there is a
continuous map f : X → (Γ, weak*) ⊂ Bl2 ⊂ l2 such that f (a) = µa for every a ∈ A.

Assume that the claim is true. Since f (x) ∈ C(A, R)∗ , we can perform the evaluation ⟨f (x), g⟩
for g ∈ C(A, R). Define T : C(A, R) → C(X, R) as T g(x) = ⟨f (x), g⟩. If (xn ) → x, then
f (xn ) → f (x) in weak* topology, and in particular if we consider the evaluation at g, we get
(⟨f (xn ), g⟩) → ⟨f (x), g⟩. Therefore T g ∈ C(X, R) indeed. It is clear that T is linear. Since
∥f (x)∥ ≤ 1 for every x ∈ X, we have |(T g)(x)| ≤ ∥f (x)∥∥g∥ ≤ ∥g∥ and hence ∥T g∥ ≤ ∥g∥, and
thus ∥T ∥ ≤ 1. Since T 1 = ⟨f (·), 1⟩ = 1, we conclude ∥T ∥ = 1. Finally, for a ∈ A and g ∈ C(A, R),
we have (T g)(a) = ⟨f (a), g⟩ = ⟨µa , g⟩ = g(a).

It remains to prove the claim, and for this we will use [101]. Let F : X → P(l2 ) be F (a) = {µa }
for a ∈ A and F (x) = Γ for x ∈ X \ A. Then F (x) is compact and convex for every x ∈ X. Let us
verify that F is lower semi-continuous. Let V ⊂ l2 be open, and x ∈ X be such that F (x) ∩ V ̸= ∅.
If x ∈ X \ A, then F (y) = Γ = F (x) for every every y ∈ X \ A and hence F (y) ∩ V ̸= ∅ for every
y ∈ X \ A. If x ∈ A, and if fn ’s are the maps used earlier for our affine homeomorphism, then
(fn (x)) ∈ V , and it is clear that (fn (y)) ∈ V for y sufficiently close to x since ∥fn ∥ = 1 for every
n ∈ N. These observations imply that {x ∈ X : F (x) ∩ V ̸= ∅} is open in X, proving that F is
lower semi-continuous. By [101], there is a continuous map f : X → (Γ, weak*) ⊂ l2 such that
f (x) ∈ F (x) for every x ∈ X. In particular, f (a) = µa for every a ∈ A. 

Remark: The above proof works also for the case where X is a compact Hausdorff space and
∅ ̸= A ⊂ X is a metrizable closed subset since we did not use the metrizability of X anywhere.

Exercise-3: If K is a compact metric space, let C(K) = C(K, R) or C(K, C), which is a Banach
space w.r.to the supremum norm. Let X be any separable Banach space.
(i) There is an isometric (linear) embedding of X in C(A), where A is the Cantor space.
(ii) [Banach-Mazur theorem] There is an isometric (linear) embedding of X in C[0, 1].
[Hint: (i) By Alouglu’s theorem etc., the closed unit ball Γ of X ∗ is a compact metrizable space
w.r.to weak* topology. Sending x ∈ X to the evaluation map at x gives an isometric embedding of
X in C(Γ). As Γ is a compact metric space, there exists a continuous surjection h : A → Γ. Then
f 7→ f ◦ h is an isometric embedding of C(Γ) in C(A). (ii) By [103], C(A) embeds isometrically in
C[0, 1]; or show directly that sending f ∈ C(A) to its extension fe ∈ C[0, 1] defined by the condition
that the graph of fe is linear on each open interval of [0, 1] \ A, is an isometric embedding.]

2. Schauder bases in Banach spaces

An application of Baire category theorem says that any Hamel basis of an infinite dimensional
Banach space X must be uncountable. If we require only that every element of X has a uniquely
6 T.K.SUBRAHMONIAN MOOTHATHU

expression as a countably infinite linear combination (rather than a finite linear combination) of
basis elements, then we can get a countable ‘basis’ for X in many cases.

Definition: A sequence (en )∞ n=1 in an infinite dimensional Banach space X is a Schauder basis if
∑∞
for every x ∈ X, there is a unique sequence (αn )∞
n=1 of scalars with x = n=1 αn en , where the last
∑k
equality means limk→∞ ∥x − n=1 αn en ∥ = 0 (similar definition for a finite sequence in a finite
dimensional Banach space). The uniqueness part implies that en ̸= 0, and hence (en /∥en ∥)∞
n=1 is a
normalized Schauder basis whenever (en )∞
n=1 is a Schauder basis for X. Note that a Schauder basis
is not just a set: the order of elements is important.
∑∞
[104] Let X be a Banach space with a Schauder Basis {en : n ∈ N}. For x = n=1 αn en ∈ X,
∑
define Pk x = kn=1 αn en and ∥x∥0 = supk∈N ∥Pk x∥. Then,
(i) ∥ · ∥0 is a norm on X and ∥x∥ ≤ ∥x∥0 for every x ∈ X.
(ii) (X, ∥ · ∥0 ) is a Banach space, and the two norms ∥ · ∥, ∥ · ∥0 are equivalent.
(iii) For each k ∈ N, Pk : (X, ∥ · ∥) → (X, ∥ · ∥) (called a natural projection) is a bounded linear
operator, and C := supk∈N ∥Pk ∥ < ∞.
(iv) Let ϕk be the kth coordinate functional on X associated to the Schauder basis (en ), given by
∑
ϕk ( ∞ ∗
n=1 αn en ) = αk . Then ϕk ∈ X and 1 ≤ ∥ϕk ∥∥ek ∥ ≤ 2C.
(v) Let Yk = span{en : n ∈ N \ {k}}. Then ek ∈
/ Yk for k ∈ N.

Proof. (i) For a fixed x ∈ X, the sequence (Pk x) is bounded since (Pk x) → x, and thus ∥x∥0 < ∞.
Easy to check that ∥ · ∥0 is a norm, and ∥x∥ = limk→∞ ∥Pk x∥ ≤ supk∈N ∥Pk x∥ = ∥x∥0 .
∑∞
(ii) Let (x(j)) be a Cauchy sequence in (X, ∥ · ∥0 ), where x(j) = n=1 αn (j)en . For any y =
∑∞
n=1 βn en , we have |βn |∥en ∥ = ∥Pn (y) − Pn−1 (y)∥ ≤ ∥Pn (y)∥ + ∥Pn−1 (y)∥ ≤ 2∥y∥0 . Therefore,
|αn (j) − αn (i)|∥en ∥ ≤ 2∥x(j) − x(i)∥0 , which shows that (αn (j))∞
j=1 is Cauchy for each n ∈ N. Let
∑∞
αn = limj→∞ αn (j). We will show x := n=1 αn en exists and (x(j)) → x in (X, ∥x · ∥0 ). Given
∑
ε > 0, ∃ j0 ∈ N with ε > ∥x(m) − x(j)∥0 = supk∈N ∥ kn=1 (αn (m) − αn (j))en ∥ ∀ m, j ≥ j0 .
∑
Letting m → ∞, we obtain supk∈N ∥ kn=1 (αn − αn (j))en ∥ ≤ ε for every j ≥ j0 . (*)
∑k
Since x(j0 ) ∈ X, there is j1 ≥ j0 such that ∥ n=j+1 αn (j0 )en ∥ ≤ ε for every k > j ≥ j1 . Hence
∑
for k > j ≥ j1 , by (*) and triangle inequality we have ∥ kn=j+1 αn en ∥
∑ ∑ ∑
≤ ∥ kn=1 (αn − αn (j0 ))en ∥ + ∥ jn=1 (αn − αn (j0 ))en ∥ + ∥ kn=j+1 αn (j0 )en ∥ ≤ ε + ε + ε = 3ε.
∑
This shows x := ∞ n=1 αn en exists in (X, ∥ · ∥) and (*) shows that (x(j)) → x in (X, ∥ · ∥0 ).
Thus (X, ∥ · ∥0 ) is a Banach space. Now I : (X, ∥ · ∥0 ) → (X, ∥ · ∥) is a bijection that is continuous
since ∥x∥ ≤ ∥x∥0 . By Inverse mapping theorem (a corollary of Open mapping theorem), I is an
isomorphism. If we let C = ∥I −1 ∥, we get ∥x∥ ≤ ∥x∥0 ≤ C∥x∥ for every x ∈ X, and this gives the
equivalence of the two norms.

(iii) Pk is linear, and ∥Pk x∥ ≤ ∥x∥0 ≤ C∥x∥ for every x ∈ X and k ∈ N, if C is as above.
 TOPICS FROM ANALYSIS - PART 1/3 7
∑∞
(iv) Clearly ϕk is linear. For x = n=1 αn en ∈ X, we have |ϕk (x)|∥ek ∥ = ∥αk ek ∥ = ∥Pk x −
Pk−1 x∥ ≤ ∥Pk x∥ + ∥Pk−1 x∥ ≤ 2C∥x∥. Also 1 = ϕk (ek ) ≤ ∥ϕk ∥∥ek ∥.

(v) Yk ⊂ ker(ϕk ) and ϕk (ek ) = 1 ̸= 0. 

Remark: If (en ) is a Schauder basis of a Banach space X, then C := supk∈N ∥Pk ∥ is called the basis
constant of (en ). We have C ≥ 1 and ∥x∥ ≤ ∥x∥0 ≤ C∥x∥ for every x ∈ X. The basis constant
depends on the norm, and it can be shown that C = 1 w.r.to the norm ∥ · ∥0 .

Remark: Let X be a Banach space having a Schauder basis (en ). Then any compact linear operator
T : X → X is the limit (in the operator norm) of a sequence of finite rank operators Tk : X → X
(k ∈ N). In fact, if Pk is the natural projection associated to (en ), we may take Tk = Pk T .

[105] [Converse of [104]] Let X be a Banach space and (en ) be a sequence in X \ {0} with Y :=
∑ ∑
span{en : n ∈ N} dense in X. Suppose there is C ≥ 1 such that ∥ kn=1 αn en ∥ ≤ C∥ m n=1 αn en ∥
whenever k ≤ m for scalars α1 , . . . , αm . Then (en ) is a Schauder basis for X.
∑ ∑k
Proof. Let Pk : Y → Y be Pk ( mn=1 αn en ) = n=1 αn en for k ≤ m. Then Pk is a bounded linear
projection on Y with ∥Pk ∥ ≤ C, and hence (by uniform continuity) Pk extends to a bounded
linear projection Pk on X with ∥Pk ∥ ≤ C. Linear functional ϕk on Y defined by the condition
that ϕk (x)ek = Pk x − Pk−1 x satisfies ∥ϕk ∥ ≤ 2C/∥ek ∥, and hence ϕk also extends to a bounded
linear functional ϕk on X. Note that ϕk (en ) = 1 if k = n and ϕk (en ) = 0 if k ̸= n. Given
x ∈ X and ε > 0, choose y ∈ Y with ∥x − y∥ < 2C ε
, and then choose k ∈ N large enough with
∑k
Pk y = y. Now ∥x − n=1 ϕn (x)en ∥ = ∥x − Pk x∥ ≤ ∥x − y∥ + ∥Pk (y − x)∥ ≤ 2C
ε Cε
+ 2C ≤ ε. This
∑∞ ∑∞ ∑∞
shows x = n=1 ϕn (x)en . Uniqueness of the coefficients: if x = n=1 βn en = n=1 ϕn (x)en , then
applying ϕk we get βk = ϕk (x). 

[106] Let (en ) be a Schauder basis in a reflexive Banach space X, and (ϕn ), (Pk ) be coordinate
functionals and natural projections respectively. Let Pk∗ : X ∗ → X ∗ be Pk∗ ψ = ψ ◦ Pk . Then,
∑
(i) Pk∗ ψ = kn=1 ψ(en )ϕn ∈ span{ϕ1 , . . . , ϕk } for every ψ ∈ X ∗ (true for any Banach space).
(ii) (Pk∗ ψ(x)) → ψ(x) for every ψ ∈ X ∗ and x ∈ X (true for any Banach space).
(iii) (ϕn ) is a Schauder basis for X ∗ with (Pk∗ ) as the sequence of natural projections. In particular,
(Pk∗ ψ) → ψ in X ∗ for every ψ ∈ X ∗ .
∑ ∑k
(iv) ∞ n=1 αn en ∈ X whenever sup{∥ n=1 αn en ∥ : k ∈ N} < ∞.

∑∞ ∑ ∑
Proof. (i) For x = n=1 ϕn (x)en , Pk∗ ψ(x) = ψ(Pk x) = ψ( kn=1 ϕn (x)en ) = kn=1 ψ(en )ϕn (x).

(ii) (Pk∗ ψ(x)) = (ψ(Pk x)) → ψ(x) since (Pk (x)) → x and ψ is continuous.

(iii) Part (ii) together with reflexivity says that (Pk ψ) → ψ weakly for every ψ ∈ X ∗ . Hence X ∗ is
the weak closure of Γ := span{ϕn : n ∈ N} by (i). Since the norm closure and weak closure coincide
8 T.K.SUBRAHMONIAN MOOTHATHU

for a vector subspace (more generally for a convex subset) of a Banach space, X ∗ = Γ. Since
∥Pk∗ ∥ ≤ ∥Pk ∥ (in fact, ∥Pk∗ ∥ = ∥Pk ∥ by reflexivity), we have C := sup{∥Pk∗ ∥ : k ∈ N} < ∞. Hence
∑ ∑ ∑m
by (i), ∥ kn=1 αn ϕn ∥ = ∥Pk∗ ( m n=1 αn ϕn )∥ ≤ C∥ n=1 αn en ∥ for k ≤ m. So (ϕn ) is a Schauder
basis for X ∗ by [105].
∑k
(iv) Let (αn ) be a fixed sequence of scalars with sup{∥xk ∥ : k ∈ N} < ∞, where xk = n=1 αn en .
Since X is reflexive and separable (by the presence of a Schauder basis), the closed unit ball of X -
and hence any closed ball of X - is compact and metrizable in the weak topology. Hence (xk ) has
a weakly convergent subsequence, say (xkj ) → x weakly in X. In particular, for n ≤ kj we have
∑
ϕn (x) = limj→∞ ϕn (xkj ) = limj→∞ αn = αn . In other words, x = ∞
n=1 αn en . 

Remark: There is also a converse with (iii) and (iv) implying reflexivity. The equivalence ‘reflexive
⇔ (iii)+(iv)’ is known as James’ theorem. If a Banach space X is not reflexive, then in the place
of [106](iii), we can say that (ϕn ) is a Schauder basis for the closure of its span.

Exercise-4: (i) Let X, Y be Banach spaces, (en ) be a Schauder basis of X, and (e′n ) be a Schauder
Basis of Y . Then there is an isomorphism T : X → Y with T en = e′n for every n ∈ N (if this
∑
happens, we say (en ) and (e′n ) are equivalent) iff the following property holds: ∞n=1 αn en converges
∑∞
in X iff n=1 αn e′n converges in Y . In particular, any two orthonormal bases of a separable Hilbert
∑ ∑∞
space H are equivalent since ∞ n=1 αn en converges in H iff n=1 |αn | < ∞.
2

(ii) Let (en ) be a Schauder basis in a Banach space X, and (ϕn ) be the sequence of coordinate
∑
functionals. If (zn ) is a sequence in X such that ∞
n=1 ∥ϕn ∥∥en − zn ∥ < 1, then (zn ) is a Schauder
basis for X equivalent to (en ).
(iii) Let X be a Banach space with Schauder basis (en ), and {xk : k ∈ N} be any countable dense
subset of X. Then some subsequence (xkn ) of (xk ) is a Schauder basis for X equivalent to (en ).
(iv) [This is needed for [107] below] Let Y be a finite dimensional vector subspace of an infinite
dimensional Banach space X, and ε > 0. Then there is x in the unit sphere SX of X such that
∥y∥ ≤ (1 + ε)∥y + αx∥ for every y ∈ Y and every scalar α.
∑ ∑∞
[Hint: (i) For ⇐, define T ( ∞ n=1 αn en ) =
′
n=1 αn en and use closed graph theorem to show T
∑∞
and similarly T −1 are bounded. (ii) Let T : X → X be T x = n=1 ϕn (x)(en − zn ). Then T
is bounded linear with ∥T ∥ < 1 so that I − T is invertible. And (I − T )ek = zk . (iii) Choose
xkn sufficiently close to en so that part (ii) becomes applicable. (iv) Enough to show there is
x ∈ SX such that ∥y + βx∥ ≥ 1/(1 + ε) for every y ∈ SY and scalar β. Since SY is compact,
∪
there are y1 , . . . , yk ∈ SY such that SY ⊂ kj=1 B(yj , ε/2). Choose ϕj ∈ SX ∗ with ϕj (yj ) = 1
∩
and pick a unit vector x ∈ kj=1 ker(ϕj ). Given y ∈ Sy , let yj be with ∥y − yj ∥ < ε/2. Then
∥y + βx∥ ≥ ∥yj + βx∥ − ε/2 ≥ |ϕj (yj + βx)| − ε/2 = 1 − ε/2 ≥ 1/(1 + ε) for ε ∈ (0, 1).]
 TOPICS FROM ANALYSIS - PART 1/3 9

Definition: Let X be a Banach space. A sequence (en ) in X is called a basis sequence if (en ) is
a Schauder basis for the closure of its span. Let (en ) be a basis sequence in X. A sequence (yk )
in X is a block basis sequence of (en ) if there exist an increasing sequence 0 = t0 < t1 < t2 < · · ·
of integers and a sequence (αn ) of scalars such that if we let Ak = {n : tk−1 < n ≤ tk }, then
∑
yk = n∈Ak αn xn for each k ∈ N. In this case, it can be seen that (yk ) is also a basis sequence
(with a possibly smaller span) in X and the basis constant of (yk ) is ≤ the basis constant of (en ).

Remark: Let X be a Banach space. A sequence (en ) in X \ {0} is a basis sequence iff there is C ≥ 1
∑ ∑
such that ∥ kn=1 αn en ∥ ≤ C∥ m
n=1 αn en ∥ for every k ≤ m and scalars (αn ) by [105].

[107] (i) Let X be a Banach space (possibly without a Schauder basis) and ε > 0. Then there
is a normalized basis sequence in X with basis constant ≤ 1 + ε. In particular, X has an infinite
dimensional closed vector subspace with a Schauder basis.
(ii) Let X be a Banach space with a Schauder basis (en ). If (xm ) is a sequence in X with inf{∥xm ∥ :
m ∈ N} > 0 and (xm ) → 0 weakly, then some subsequence (xmk ) is a basis sequence equivalent to
a block basis sequence of (en ).

∏∞
Proof. (i) Let εn > 0 be chosen with n=1 (1 + εn ) ≤ 1 + ε. Let e1 ∈ SX (unit sphere of X). Having
chosen e1 , . . . , en ∈ SX , put Yn = span{e1 , . . . , en } and by Exercise-4(iv) choose en+1 ∈ SX such
that ∥y∥ ≤ (1 + εn )∥y + αen+1 ∥ for every y ∈ Yn and scalar α. Then for k < m and scalars (αn )
∑ ∑ ∑m
we have ∥ kn=1 αn en ∥ ≤ (1 + εk ) · · · (1 + εm−1 )∥ m n=1 αn en ∥ ≤ (1 + ε)∥ n=1 αn en ∥.
∑
(ii) Let C ≥ 1 is the basis constant of (en ) and ∞ k=1 δk =: δ < 1 be a convergent series of positive
reals with inf{∥xm ∥ : m ∈ N} > 2δ > 0. Let (Pk ) be the sequence of natural projections associated
to (en ). We claim that there are increasing sequences 1 ≤ m1 < m2 < . . . and 0 = t0 < t1 < t2 < . . .
δδk
of integers such that ∥xmk − (Ptk − Pt(k−1) )xmk ∥ < for every k ∈ N. Assume the claim and let
2C
δδk
yk = (Ptk − Pt(k−1) )xmk ∈ span{en : tk−1 < n ≤ tk }. Since ∥xmk ∥ > 2δ and ∥xmk − yk ∥ < ≤ δ,
2C
we have ∥yk ∥ > δ > 0. Thus (yk ) is a block basis sequence of (en ), and the basis constant of (yk )
is ≤ C. If (ψk ) is the sequence of coordinate functionals for (yk ), then ∥ψk ∥δ ≤ ∥ψk ∥∥yk ∥ ≤ 2C
∑ ∑∞
by [104](iv) and therefore ∥ψk ∥ ≤ 2C/δ. Hence ∞ k=1 ∥ψk ∥∥xmk − yk ∥ ≤ k=1 δk = δ < 1. By
Exercise-4(ii), (xmk ) is a basis sequence equivalent to (yk ).
δδ1
It remains to prove the claim. Let m1 = 1, t0 = 0 and choose t1 ∈ N with ∥xm1 − Pq1 xm1 ∥ <
,
2C
where Pt0 := 0. Now assume we have chosen 1 ≤ m1 < · · · < mk and 0 = t0 < t1 < · · · <
tk with the required property. Since (xm ) → 0 weakly, and since Ptk is a linear combination
of the first tk coordinate functionals, we have limm→∞ Ptk xm = 0. Choose mk+1 > mk with
δδk+1 δδk+1
∥Ptk xm(k+1) ∥ < , and then choose tk+1 > tk with ∥xm(k+1) − Pt(k+1) xm(k+1) ∥ < . Then
4C 4C
δδk+1
∥xm(k+1) − (Pt(k+1) − Ptk )xm(k+1) ∥ ≤ ∥xm(k+1) − Pt(k+1) xm(k+1) ∥ + ∥Ptk xm(k+1) ∥ < . 
2C
10 T.K.SUBRAHMONIAN MOOTHATHU

Remark: (i) There exists a separable Banach space that does not have a Schauder Basis - this is a
non-trivial result of P.Enflo. (ii) A sequence (xm ) as in [107](ii) may not exist in a Banach space;
for instance, any weakly null sequence (xm ) in l1 satisfies limm→∞ ∥xm ∥ = 0.

3. Unconditional bases and orthogonal systems in Banach spaces

∑∞ ∑∞
Definition: A series n=1 xn in a Banach space is said to be unconditionally convergent if n=1 xσ(n)
converges for every permutation σ of N. A sequence (en ) in X is said to be an unconditional basis
∑
of X if every x ∈ X has a unique expression x = ∞n=1 αn en that converges unconditionally. Thus
every unconditional basis is a Schauder basis.

[108] The following are equivalent for a sequence (xn ) in a Banach space X:
∑
(i) ∞ n=1 xσ(n) converges for every permutation σ of N.
∑
(ii) ∀ ε > 0, ∃ n0 ∈ N such that ∥ n∈A xn ∥ < ε for any finite subset A ⊂ {n0 , n0 + 1, n0 + 2, . . .}.
∑
(iii) ∞ k=1 xnk converges for any strictly increasing sequence (nk ) of natural numbers.
∑∞
(iv) n=1 εn xn converges for every choice of signs εn ∈ {−1, 1}.
∑
(v) ∞ n=1 αn xn converges for any bounded sequence (αn ) of scalars.
∑∞
(vi) n=1 |ϕ(xn )| converges uniformly in ϕ for ϕ belonging to the closed unit ball of X ∗ .

Proof. (i) ⇔ (ii): If (ii) holds, then the partial sums of the series in (i) are Cauchy and hence (i)
holds. If (ii) does not hold, there exist ε > 0 and a sequence of finite subsets (Ak ) of N such that
∑
max Ak < min Ak+1 and ∥ n∈Ak xn ∥ ≥ ε for every k ∈ N. If σ is a permutation on N that satisfies
∑
σ −1 (Ak ) is a set of |Ak | consecutive integers for every k ∈ N, then the partial sums of ∞
n=1 xσ(n)
do not form a Cauchy sequence and hence (i) fails.
∪
(iii) ⇒ (ii): If (ii) fails, and if Ak ’s are as above, then writing ∞
k=1 Ak = {m1 < m2 < . . .} we see
∑∞
that the partial sums of the series j=1 xmj do not form a Cauchy sequence and hence (iii) fails.

(iv) ⇒ (iii): Let A = {n1 < n2 < · · · }, and let εn = 1 or −1 according to whether n ∈ A or n ∈
/ A.
∑∞ ∑∞ ∑∞
Then k=1 xn = (1/2)[ n=1 xn + n=1 εn xn ] converges.

(v) ⇒ (iv): Trivial.

(vi) ⇒ (v): Suppose M := sup{|αn | : n ∈ N} < ∞. Given k < m, by Hahn-Banach theorem

∑ ∑m ∑m
there is ϕ ∈ X ∗ with ∥ϕ∥ = 1 such that ϕ( mn=k αn xn ) = ∥ n=k αn xn ∥. Then ∥ n=k αn xn ∥ =
∑m ∑m ∑m
ϕ( n=k αn xn ) ≤ n=k |ϕ(αn xn )| ≤ M n=k |ϕ(xn )|.

(ii) ⇒ (vi): Given ε > 0, choose n0 ∈ N as in (ii), and consider ϕ ∈ X ∗ with ∥ϕ∥ ≤ 1.
∑
Let m > k ≥ n0 and A+ = {k ≤ n ≤ m : Re(ϕ(xn )) ≥ 0}. Then n∈A+ Re(ϕ(xn )) =
∑ ∑
Re(ϕ( n∈A+ xn )) ≤ ∥ n∈A+ xn ∥ < ε since ∥ϕ∥ ≤ 1. Similarly, if A− = {k ≤ n ≤ m :
 TOPICS FROM ANALYSIS - PART 1/3 11
∑ ∑
Re(ϕ(xn )) < 0}, then −( n∈A− Re(ϕ(xn ))) < ε and hence mn=k |Re(ϕ(xn ))| < 2ε. In the same
∑m ∑m ∑m ∑
way, n=k |Im(ϕ(xn ))| < 2ε. So, n=k |ϕ(xn )| ≤ n=k |Re(ϕ(xn ))|+ mn=k |Im(ϕ(xn ))| < 4ε. 

Remarks and Examples: (i) It follows from [108] that (en ) is an unconditional basis for a Banach
space X iff (eσ(n) ) is a Schauder basis for X for every permutation σ of N.
∑ ∑∞
(ii) Let ∞n=1 xn be an unconditionally convergent series in a Banach space X, and let x = n=1 xn .
∑∞
We claim that n=1 xσ(n) = x (the same limit) for every permutation σ of N. Given ε > 0, choose
∑
n0 ∈ N such that ∥ n∈A xn ∥ < ε for every finite subset A ⊂ {n0 + 1, n0 + 2, . . .} by [108] and
∑ ∑
also such that ∥ n>m xn ∥ < ε for every m ≥ n0 by the convergence of ∞ n=1 xn . Choose n1 ≥ n0
with {1, . . . , n0 } ⊂ {σ(1), . . . , σ(n1 )}. Then for any k > n1 , letting m = max{σ(n) : 1 ≤ n ≤
k} ≥ n1 ≥ n0 and A = {1, . . . , m} \ {σ(n) : 1 ≤ n ≤ k} ⊂ {n0 + 1, n0 + 2, . . .}, we have
∑ ∑ ∑
∥x − kn=1 xσ(n) ∥ ≤ ∥ n∈A xn ∥ + ∥ n>m xn ∥ < ε + ε = 2ε, proving the claim.

(iii) If (en ) is an orthonormal basis of a Hilbert space X, then (en ) is an unconditional basis. This is
∑ ∑∞
because ∞ n=1 αn en converges in X iff n=1 |αn | < ∞, and because of the fact that absolute con-
2

vergence implies unconditional convergence for a series of real numbers (this is true more generally
for a series in a finite dimensional Banach space). Similarly if en ∈ lp is (0, . . . , 0, 1n , 0, . . .), then (en )
∑
is an unconditional basis for lp for 1 ≤ p < ∞ since the convergence of ∞ p
n=1 αn en in l is equivalent
∑ ∑m ∑m
to the (absolute) convergence of the real series ∞ n=1 |αn | as ∥
p
n=k αn en ∥ =
p
n=k |αn | .
p

∑
(iv) For a series ∞ n=1 xn in a Banach space, absolute convergence implies unconditional conver-
gence. The converse is not true even in a Hilbert space: if (en ) is an orthonormal basis, then
∑ ∞ en ∑ 1
n=1 converges unconditionally (since ∞ n=1 2 < ∞) but not absolutely.
n n
∑∞ ∑
(v) Let c0 = { n=1 αn en ∈ l : lim αn = 0}, which is a closed subspace of l∞ . Let yn = ∞
∞
j=1 ej .
∑∞ ∑∞ ∑m
Given x = n=1 αn en , we note x = n=1 (αn − αn+1 )yn . Since ∥ n=k (αn − αn+1 )yn ∥∞ =
∑k ∑m
∥ n=1 (αk − αm+1 )en + n=k+1 (αn − αm+1 )en ∥∞ = max{|αn − αm+1 | : k ≤ n ≤ m} and since
∑ ∑∞ ∑∞
∥x− m n=1 (αn −αn+1 )yn ∥∞ = ∥αm+1 n=1 en ∥∞ = |αm+1 |, the series n=1 (αn −αn+1 )yn is indeed
∑∞
convergent in c0 with n=1 (αn − αn+1 )yn = x. Consequently (yn ) is a Schauder basis for c0 . But
∑ yn
(yn ) is not an unconditional basis since the series ∞ n=1 does not converge in c0 even though
n
∑∞ (−1)n yn
n=1 is convergent (check).
n
∑
(vi) Let c = { ∞ n=1 αn en ∈ l
∞ : (α ) is convergent}, which is also a closed subspace of l∞ . Let
n
∑∞ ∑
yn = (0, . . . , 0, 1, 1, . . .). Given x = n=1 αn en , we note x = ∞ n=1 (αn − αn−1 )yn (where α0 := 0)
| {z }
n−1
whose validity can be verified as in the c0 case. Consequently (yn ) is a Schauder basis for c (in
particular, c is separable). But (yn ) is not an unconditional basis. Let (αn ) be the sequence
12 T.K.SUBRAHMONIAN MOOTHATHU

∑ ∑ ∑
(1, 1, 1/2, 1/2, 1/3, 1/3, . . .). Then we have ∥ kn=1 αn yn ∥ = kn=1 αn so that ∞ n=1 αn yn does not
∑∞ ∑∞ e2n−1
converge in c. On the other hand, n=1 (−1) n+1 αn yn converges to n=1 ∈ c (check).
n
(vii) A deep result of Gowers and Maurey says that there exists an infinite dimensional Banach
space in which no infinite dimensional closed vector subspace possesses an unconditional basis.

Exercise-5: Let X be a Banach space and (en ) be a sequence in X \ {0} with dense span. Then
the following are equivalent: (i) (en ) is an unconditional basis for X.
∑ ∑
(ii) There is a constant C1 ≥ 1 such that ∥ n∈A αn en ∥ ≤ C1 ∥ m n=1 αn en ∥ for every m ∈ N, every
∅ ̸= A ⊂ {1, . . . , m}, and every sequence (αn ) of scalars.
∑ ∑m
(iii) There is a constant C2 ≥ 1 such that ∥ m n=1 εn αn en ∥ ≤ C2 ∥ n=1 αn en ∥ for every m ∈ N,
every choice of signs εn ∈ {−1, 1}, and every sequence (αn ) of scalars.
∑ ∑
[Hint: (i) ⇒ (ii): For finite A ⊂ N, PA : X → X defined as PA ( ∞ n=1 αn en ) = n∈A αn en is
∑∞
bounded linear by Closed graph theorem. For x = n=1 αn en ∈ X, choosing n0 as in [108](ii) for
∑ 0
ε = 1, we note sup{∥PA (x)∥ : A ⊂ N finite}∥ ≤ ∥ nn=1 αn en ∥ + 1. By the Uniform boundedness
∑m ∑ ∑
principle, supA ∥PA ∥ < ∞. (ii) ⇒ (iii): ∥ n=1 εn αn en ∥ ≤ ∥ εn =1 αn en ∥ + ∥ εn = −1 αn en ∥. (iii)
∑ ∑ ∑m
⇒ (ii): Let εn = ±1 for n ∈ A and n ∈ / A. Then n∈A αn en = (1/2)( m n=1 αn en + n=1 εn αn en ).
∑k ∑k
(ii) & (iii) ⇒ (i): Taking A = {1, . . . , k} for k < m, ∥ n=1 εn αn en ∥ ≤ C2 ∥ n=1 αn en ∥ ≤
∑
C1 C2 ∥ m n=1 αn en ∥. So (en ) is an unconditional basis for X by [106] and [108](iv).]

Seminar topics: [More examples of bases] (i) The Haar system is a Schauder basis for Lp [0, 1] for
1 ≤ p < ∞, and is an unconditional basis for Lp [0, 1] for 1 < p < ∞. The Haar system {fn : n ∈ N}
j−1 j
is defined as follows. Let L(k, j) = ( k , k ] for k ∈ N and 1 ≤ j ≤ 2k . Define f1 ≡ 1 and
2 2
f2k +j = 1L(k,2j−1) − 1L(k,2j) for k ≥ 2 and 1 ≤ j ≤ 2k−1 .
(ii) C[0, 1] has a Schauder basis {gn : n ∈ N} called the Faber-Schauder basis defined as follows:
∫x
g1 ≡ 1 and gn (x) = 0 fn−1 (y)dµ(y) for n > 1, where {fn : n ∈ N} is the Haar system given above.
(iii) The trigonometric system is a Schauder basis for Lp (S1 ) for 1 < p < ∞, and is an unconditional
basis for Lp (S1 ) for p = 2.

Definition: Let X be a Banach space. Sequences (en ) in X and (ϕn ) in X ∗ are said to be a bi-
orthogonal system for X if ϕm (en ) = δmn for every m, n (here (en ) need not be a basis sequence).
The standard example of a bi-orthogonal system is a Schader basis (en ) together with the sequence
(ϕn ) of coordinate functionals.

If X is a finite dimensional vector space, then any basis of X together with the dual basis of X ∗
will satisfy the bi-orthogonality condition. The highlight of Exercise-6 below is that we can also
control the norms of these vectors.
 TOPICS FROM ANALYSIS - PART 1/3 13

Exercise-6: [Auerbach basis] Let X be a k-dimensional Banach space, k < ∞. Then X has
a bi-orthogonal system (en )kn=1 , (ϕn )kn=1 with ∥en ∥ = 1 = ∥ϕn ∥ for 1 ≤ n ≤ k. [Hint: Let
Y = SX := {x ∈ X : ∥x∥ = 1}, which is compact. Fix a basis {z1 , . . . , zk } ⊂ Y k for X. Let
∑k
f : Y k → R be f (y1 , . . . , yk ) = |det([aij ]k×k )|, where aij ’s are such that yj = i=1 aij zi . If f
f (e1 , . . . , em−1 , y, em+1 , . . . , ek )
attains maximum at (e1 , . . . , ek ) ∈ Y k , define ϕm (y) = for y ∈ Y .]
f (e1 , . . . , ek )
The importance of the notion of a bi-orthogonal system is that even when a separable Banach
space X has no Schauder basis, it is possible to obtain a bi-orthogonal system (en ), (ϕn ) for X such
that (en ) (respectively (ϕn )) behaves almost like a Schauder basis for X (respectively X ∗ ). This is
the content of the next result (the case dim(X) < ∞ is covered in Exercise-6).

[109] [Markushevich] Let X be an infinite dimensional separable Banach space. Then there is a
bi-orthogonal system (en ), (ϕn ) on X such that:
(i) span{en : n ∈ N} is dense in X.
(ii) x ∈ X and ϕn (x) = 0 for every n ∈ N ⇒ x = 0; in other words, span{ϕn : n ∈ N} is weak*-dense
in X ∗ . If X ∗ is also separable, we can choose (ϕn ) so that span{ϕn : n ∈ N} is dense in X ∗ .

Proof. Let yn ∈ X \ {0} be with span{yn : n ∈ N} dense in X. Since X is separable Banach, the
closed unit ball Γ of X ∗ is compact and metrizable in the weak* topology by Alaoglu’s theorem.
Let {ψn : n ∈ N} ⊂ Γ \ {0} be a countable weak*-dense subset of Γ (if X ∗ is also separable, then
we may also assume span{ψn : n ∈ N} is dense in X ∗ ). Then the first assertion of (ii) holds for
{ψn : n ∈ N}. We will choose (en ) and (ϕn ) inductively. At the odd steps, we will choose en first
and then ϕn . At the even steps, we will choose ϕn first and then en .

First and second steps: Let k1 = 1 and e1 = yk1 . Let m1 be the smallest with ψm1 (e1 ) ̸= 0, and let
ϕ1 = ψm1 (e1 )−1 ψm1 . Let m2 be the smallest with ψm2 ∈
/ span{ϕ1 }, let ϕ2 = ψm2 − ψm2 (e1 )ϕ1 , let
k2 be such that ϕ2 (yk2 ) ̸= 0, and e2 = ϕ2 (yk2 )−1 [yk2 − ϕ1 (yk2 )e1 ].

General step for n odd : Let kn be the smallest with ykn ∈ / span{e1 , . . . , en−1 }, let en = ykn −
∑n−1 −1
∑n−1
j=1 ϕj (ykn )ej , let mn be such that ψmn (en ) ̸= 0, and ϕn = (ψmn (en )) [ψmn − j=1 ψkn (ej )ϕj ].

General step for n even: Let mn is the smallest with ψmn ∈ / span{ϕ1 , . . . , ϕn−1 }, let ϕn = ψmn −
∑n−1 −1
∑n−1
j=1 ψmn (ej )ϕj , let kn be such that ϕn (ykn ) ̸= 0, and en = ϕn (ykn ) [ykn − j=1 ϕj (ykn )ϕj ].

By construction, span{y1 , . . . , yn } ⊂ span{e1 , . . . , e2n } and span{ψ1 , . . . , ψn } ⊂ span{ϕ1 , . . . , ϕ2n }

for every n ∈ N, and from this we may deduce properties (i) and (ii). Verify that ϕm (en ) = δmn . 

In the above construction we do not have control over the norms of en and ϕn . See p.44 of
Lindenstrauss and Tzafriri, Classical Banach Spaces - I, for a proof of the following:
14 T.K.SUBRAHMONIAN MOOTHATHU

[109′ ] [Ovespian and Pelczynski] In [109], we can choose (en ) and (ϕn ) in such a manner that the
additional property sup{∥en ∥∥ϕn ∥ : n ∈ N} < ∞ is also satisfied.

4. Convexity, smoothness, and equivalent norms

Convention and aim: Let X be a real Banach space throughout this section. Let SX = {x ∈ X :
∥x∥ = 1} and BX = {x ∈ X : ∥x∥ ≤ 1}. Our aim here is to examine a little bit the following: how
different is the geometry of a Banach space from that of a Hilbert space/Euclidean space?

Definition: We say X is strictly convex if ∥x + y∥ < 2 for any two distinct x, y ∈ SX . We say X is
smooth if for every x ∈ X \ {0}, there exists a unique ϕ ∈ SX ∗ with ϕ(x) = ∥x∥ (here ϕ is called a
supporting functional of x). Equivalently, X is smooth if for every x ∈ SX , there exists a unique
ϕ ∈ SX ∗ with ϕ(x) = 1.

Examples: (i) A Hilbert space is both smooth and strictly convex. Hint: If x, y ∈ SX are distinct,
then y = αx + z, where α ∈ (−1, 1) and ⟨x, z⟩ = 0. Then |⟨x, y⟩| = |α| < 1. So ∥x + y∥2 < 4.
(ii) C([0, 1], R) is neither strictly convex nor smooth. Hint: ∥1∥ = ∥t∥ = ∥(1 + t)/2∥ = 1. If ϕc is
the evaluation functional f 7→ f (c), then ∥ϕc ∥ = 1 and ϕc (1) = 1 for every c ∈ [0, 1].
(iii) l1 is neither strictly convex nor smooth. Hint: ∥e1 + e2 ∥ = 2; ∥ϕ1 ∥ = 1 = ∥ϕ1 + ϕ2 ∥ and
ϕ1 (e1 ) = 1 = (ϕ1 + ϕ2 )(e1 ).
(iv) l∞ is neither strictly convex nor smooth. Hint: ∥e1 + e2 ∥ = 1 = ∥e1 ∥ and ∥(e1 + e2 ) + e1 ∥ = 2;
ϕj (e1 + e2 ) = 1 for j = 1, 2.
(v) Using the norm ∥(x, y)∥ = |x| + |y| on R2 , we see that a finite dimensional Banach space need
not be strictly convex or smooth.
(vi) For 1 < p < ∞, it can be shown that lp is strictly convex (one should examine the equality
case of Minkowski inequality); hence lp is also smooth by duality and Exercise-9 below.

Definition: Let A ⊂ X be a convex set. We say z ∈ A is an extreme point of A if A \ {z} is convex,

i.e., if z is not a proper convex combination of two distinct points in A.
Exercise-7: The following are equivalent:
(i) X is strictly convex.
(ii) ∥x + y∥ = ∥x∥ + ∥y∥ ⇒ {x, y} is linearly dependent for x, y ∈ X.
(iii) Every z ∈ SX is an extreme point of BX .
[Hint: (i) ⇒ (ii): Suppose ∥x + y∥ = ∥x∥ + ∥y∥. Let α = ∥x∥, β = ∥y∥. and assume 0 < α ≤ β.
Then 2 ≥ ∥α−1 x+β −1 y∥ ≥ ∥α−1 x+α−1 y∥−∥α−1 y−β −1 y∥ = α−1 (α+β)−β(α−1 −β −1 ) = 2 so that
α−1 x = β −1 y, or βx − αy = 0. (ii) ⇒ (iii): Let x ̸= y be in SX , α ∈ (0, 1) and z = αx + (1 − α)y. If
z ∈ SX , then 1 = ∥αx + (1 − α)y∥ ≤ α∥x∥ + (1 − α)∥y∥ = 1, which implies {αx, (1 − α)y} is linearly
dependent. Then {x, y} is linearly dependent, forcing y = −x. Hence ∥z∥ < 1, a contradiction.]
 TOPICS FROM ANALYSIS - PART 1/3 15

Definition: A function f : X → R is said to be Gateaux differentiable at x ∈ X if there is ϕ ∈ X ∗

f (x + ty) − f (x)
such that limt→0 = ϕ(y) for every y ∈ X (understand this as the existence of
t
directional derivative in all directions). If the above limit exist uniformly for y ∈ SX , then f is
said to be Frechet differentiable at x. We will not discuss these concepts in detail, but Exercise-8
below says that smoothness of (X, ∥ · ∥) is equivalent to the Gateaux differentiability of the norm
function ∥ · ∥ : X → R at every x ∈ X \ {0}.

Exercise-8: [Reason for the name smooth] Let x ∈ SX and ϕ ∈ SX ∗ be such that ϕ(x) = 1. Let
∥x+ty∥−∥x∥
y ∈ X and f (t) = t for t ∈ R \ {0}. Then,
(i) f (−t) ≤ ϕ(y) ≤ f (t) ≤ ∥y∥ for t ∈ (0, ∞).
(ii) f is increasing on (0, ∞) and decreasing on (−∞, 0). Consequently, the limits exist in the
following inequality: limt→0− f (t) ≤ ϕ(y) ≤ limt→0+ f (t) ≤ ∥y∥.
∥x+ty∥−∥x∥
(iii) X is smooth at x ⇔ the limit limt→0 t exists (and is equal to ϕ(y)) for every y ∈ X.
[Hint: (i) tϕ(y) = 1+tϕ(y)−1 = ϕ(x+ty)−∥x∥ ≤ ∥x+ty∥−∥x∥. (ii) . For 0 < s < t < ∞, we have
t∥x+sy∥ ≤ s∥x+ty∥+(t−s)∥x∥, which implies f (s) ≤ f (t). (iii) If the limit does not exist for some
y ∈ X, then y ∈ X \ span{x}, and there is b ̸= ϕ(y) with limt→0− f (t) < b < limt→0+ f (t) ≤ ∥y∥.
By Hahn-Banach there is ψ ∈ SX ∗ with ψ(x) = 1 and ψ(y) = b ̸= ϕ(y) so that X cannot be smooth
at x. If the limit exists for every y ∈ X, then ϕ(y) has to be this limit, which specifies ϕ uniquely.]

Exercise-9: (i) If X ∗ is strictly convex, then X is smooth.

(ii) If X ∗ is smooth, then X is strictly convex.
(iii) For a reflexive Banach space, strict convexity and smoothness are dual properties.
[Hint: (i) If smoothness fails at x ∈ SX , and if ϕ ̸= ψ are unit functional with ϕ(x) = 1 = ψ(x),
then 2 = (ϕ + ψ)(x) ≤ ∥ϕ + ψ∥ so that X ∗ is not strictly convex. (ii) If X is not strictly convex
and if ∥x + y∥ = 2 for unit vector x ̸= y, choose ϕ ∈ SX ∗ with ϕ( x+y
2 ) = 1 by Hahn-Banach. Then
ϕ(x) = 1 = ϕ(y) so that by considering x, y as members of SX ∗∗ we see X ∗ fails to be smooth at ϕ.]

Definition: Two norms ∥ · |, ∥ · ∥′ on X are equivalent if there are 0 < a ≤ b such that a∥x∥′ ≤
∥x∥ ≤ b∥x∥′ for every x ∈ X. Note that one inequality is sufficient: consider I : X → X, where the
domain and range are equipped with different norms, and use open mapping theorem.

Exercise-10: Let T : X → Z be an injective bounded linear operator into a strictly convex Banach
space Z. Then the following norms on X are strictly convex and equivalent to the norm of X:
√
(i) µ(x) := ∥x∥ + ∥T x∥. (ii) ρ(x) := ∥x∥2 + ∥T x∥2 .
[Hint: (i) Clearly ∥x∥ ≤ µ(x) ≤ (1+∥T ∥)∥x∥. If µ(x+y) = µ(x)+µ(y), then ∥x+y∥+∥T x+T y∥ =
∥x∥ + ∥T x∥ + ∥y∥ + ∥T y∥, which forces ∥x + y∥ = ∥x∥ + ∥y∥ and ∥T x + T y∥ = ∥T x∥ + ∥T y∥ by
triangle inequality. From the latter and Exercise-7, we deduce {T x, T y} is linearly dependent.
√
Then {x, y} is linearly dependent since T is injective. (ii) Clearly ∥x∥ ≤ ρ(x) ≤ 1 + ∥T ∥2 ∥x∥. If
16 T.K.SUBRAHMONIAN MOOTHATHU

ρ(x + y)2 = (ρ(x) + ρ(y))2 , we may deduce ∥x + y∥2 + ∥T x + T y∥2 = (∥x∥ + ∥y∥)2 + (∥T x∥ + ∥T y∥)2
since 2ρ(x)ρ(y) ≥ 2∥x∥∥y∥ + 2∥T x∥∥T y∥. Now argue as in (i).]

Example: Recall that l1 ⊂ l2 . Let T : l1 → l2 be the inclusion map, which is linear and injective.
If x ∈ l1 is with ∥x∥1 ≤ 1, then ∥x∥22 ≤ ∥x1 ∥1 ≤ 1 since |xn |2 ≤ |xn | when |xn | ≤ 1, and hence
∥T ∥ ≤ 1. In fact, ∥T ∥ = 1 since ∥T e1 ∥2 = 1. Moreover, l2 is strictly convex, being a Hilbert space.
By Exercise-10, ∥x∥ := ∥x∥1 + ∥x∥2 is a strictly convex equivalent norm on l1 .

Question: Given a real Banach space, can we find an equivalent norm that is strictly convex/smooth?
To answer this in [111] a little later, we need some preparation.

[110] [Geometric Hahn-Banach theorem] (i) Let A ⊂ X be closed and convex, and z ∈ X \ A. Then
there is ϕ ∈ X ∗ such that sup{ϕ(x) : x ∈ A} < ϕ(z).
(ii) Let U ⊂ X be convex and open, and z ∈ X \ U . Then there is ϕ ∈ X ∗ such that ϕ(x) < ϕ(z)
for every x ∈ U (but sup{ϕ(x) : x ∈ U } can be equal to ϕ(z)).
(iii) Let C, D ⊂ X be disjoint convex sets with C open. Then there exist ϕ ∈ X ∗ and α ∈ R such
that ϕ(x) < α ≤ ϕ(y) for every (x, y) ∈ C × D. If D is also open, then we have ϕ(x) < α < ϕ(y)
for every (x, y) ∈ C × D.
(iv) Let C, D ⊂ X be convex, int[C] ̸= ∅, and int[C] ∩ D = ∅. Then there exist ϕ ∈ X ∗ \ {0} and
α ∈ R such that ϕ(x) ≤ α ≤ ϕ(y) for every (x, y) ∈ C × D.
(v) Let C, D ⊂ X be disjoint convex sets. If C is compact and D is closed, then there exist ϕ ∈ X ∗
and α < β in R such that ϕ(x) < α < β < ϕ(y) for every (x, y) ∈ C × D.

Proof. Replacing A by A−x and z by z−x for some x ∈ A, we may assume 0 ∈ A. Let λ = dist(z, A)
and K = {x ∈ X : dist(x, A) ≤ λ/2}. Then K is closed, convex, 0 ∈ B(0, λ/2) ⊂ int[K], and
/ K. Let p : X → [0, ∞) be p(x) = inf{α > 0 : α−1 x ∈ K}. We claim the following:
z∈
(a) p(x + y) ≤ p(x) + p(y) for every x, y ∈ X.
(b) p(tx) = tp(x) for every x ∈ X and t > 0.
(c) p(x) ≤ 1 for every x ∈ K.
(d) p(x) ≤ 3∥x∥/λ for every x ∈ X.
(e) p(z) ∈ (1, ∞).
To prove (a), consider x, y ∈ X and ϵ > 0. Enough to show p(x + y) ≤ p(x) + p(y) + 2ϵ. Choose
α, β such that α−1 x, β −1 y ∈ K, 0 < α < p(x) + ϵ, and 0 < β < p(y) + ϵ. Now by convexity,
x+y α x β y
=( ) +( ) ∈ K and hence p(x + y) ≤ α + β ≤ p(x) + p(y) + 2ϵ.
α+β α+β α α+β β
λx
Statements (b) and (c) are easy. We get (d) since p(0) = 0 and ∈ B(0, λ/2) ⊂ K for
3∥x∥
x ∈ X \ {0}. Finally (e) is proved as follows. If α := p(z) ≤ 1, then first we see z/α ∈ K since K
 TOPICS FROM ANALYSIS - PART 1/3 17

is closed, and then we see z ∈ K since K is convex and z is a convex combination of 0, z/α ∈ K.
This is a contradiction. This completes the proof of the properties of p.
Define a linear functional ϕ : span{z} → R as ϕ(αz) = αp(z), note ϕ ≤ p, and use Hahn-Banach
theorem for real vector spaces to get a linear extension ϕ : X → R with ϕ(x) ≤ p(x) for every
x ∈ X. We have ∥ϕ(x)∥ ≤ 3∥x∥/λ by (d) so that ϕ ∈ X ∗ . Moreover, we have ϕ(x) ≤ p(x) ≤ 1 for
every x ∈ A ⊂ K, and ϕ(z) = p(z) ∈ (1, ∞) by (c) and (e).

(ii) After a translation assume 0 ∈ U . Let p : X → R be p(x) = inf{α > 0 : α−1 x ∈ U }, and
proceed as in part (i). Note that p(x) < 1 for x ∈ U since U is open.

(iii) Let U = C − D, z = 0 and apply part (ii) to get ϕ ∈ X ∗ with ϕ(x − y) < ϕ(0) = 0 for every
(x, y) ∈ C × D. Then ϕ(x) < ϕ(y) for every x ∈ C and y ∈ D. Take α = inf{ϕ(y) : y ∈ D}. Since
ϕ ∈ X ∗ \ {0}, it is surjective and hence an open mapping. Therefore ϕ(C) is open and convex in
R. Thus ϕ(C) is an open interval (possibly unbounded). Hence ϕ(x) < α for every x ∈ C. If D is
also open, then similarly note that ϕ(D) is an open interval to conclude α < ϕ(y) for every y ∈ D.

(iv) Since int[C] ̸= ∅ is convex by Exercise-11(i) below, there are ϕ ∈ X ∗ \ {0} and α ∈ R with
ϕ(x) < α ≤ ϕ(y) for every (x, y) ∈ int[C] × D by (iii). Also, C ⊂ int[C] by Exercise-11(ii) below.

(iv) Since the compact set C is contained in the open set X \ D, there is ε > 0 such that the
ε-neighborhood of C is contained in X \ D. Let C1 = {x ∈ X : dist(x, C) < ε/2} and D1 = {x ∈
X : dist(x, D) < ε/2}. Then C1 , D1 are disjoint open convex sets. Apply (iii) to get ϕ ∈ X ∗ and
β ∈ R such that ϕ(x) < β < ϕ(y) for every (x, y) ∈ C1 × D1 . But ϕ(C) must be a compact interval,
being compact and convex. So there is α < β with ϕ(x) < α for every x ∈ C. 

Remark: For complex Banach spaces, an analogue of [110] holds, where one should use Re(ϕ(·))
instead of ϕ(·) in various assertions. Moreover, a similar proof works for locally convex topological
vector spaces, and we note one assertion below for later use:

[110′ ] Let Y be a locally convex real topological vector space, A ⊂ Y be closed and convex, and
z ∈ Y \ A. Then there is a continuous linear functional ϕ : Y → R with sup{ϕ(y) : y ∈ A} < ϕ(z).

Exercise-11: (i) If A ⊂ X is convex, then int[A] is convex.

(ii) If A ⊂ X is convex and int[A] ̸= ∅, then A ⊂ int[A].
(iii) The norm and weak closures coincide for any convex subset A ⊂ X.
(iv) If (xn ) → x weakly in X, there is a sequence (yk ) in conv{xn : n ∈ N} with ∥x − yk ∥ → 0.
[Hint: (i) Let x, y ∈ int[A]. Choose δ > 0 with x+z, y+z ∈ A when ∥z∥ < δ. Let xα = αx+(1−α)y
for α ∈ [0, 1], and note xα + z = α(x + z) + (1 − α)(y + z) to see B(xα , δ) ⊂ A. (ii) Let x ∈ int[A],
y ∈ A, and xα = αx + (1 − α)y for α ∈ (0, 1). Choose δ > 0 with x + z ∈ A for ∥z∥ < δ and show
B(xα , αδ) ⊂ A by noting xα +z = α(x+z)+(1−α)y. Thus y ∈ int[A] since (xα ) → y as α → 0. (iii)
18 T.K.SUBRAHMONIAN MOOTHATHU

w
A ⊂ A always. If z ∈ X \ A, there is ϕ ∈ X ∗ with ϕ(z) > sup{ϕ(x) : x ∈ A} = sup{ϕ(x) : x ∈ A},
w
and consequently z ∈
/ A . (iv) Take A = conv{xn : n ∈ N} (convex hull) and use (iii).]

Remark: Even for a vector subspace Γ ⊂ X ∗ , the norm closure and weak* closure need not be the
same. Start with a non-reflexive Banach space Y , let X = Y ∗ and Γ ⊂ X ∗ = Y ∗∗ be the natural
isometric embedding of Y in Y ∗∗ . Being the isometric image of a Banach space, Γ is norm-closed in
Y ∗∗ ; and Γ ̸= Y ∗∗ since Y is not reflexive. On the other hand, Γ is weak* dense in Y ∗∗ (a property
of the embedding of Y in Y ∗∗ ) and therefore the weak* closure of Γ is Y ∗∗ .

Exercise-12: [Bipolar theorem] For A ⊂ X, let A0 = {ϕ ∈ X ∗ : ϕ(x) ≤ 1 for every x ∈ A} (called

the polar of A). For Γ ⊂ X ∗ , let Γ0 = {x ∈ X : ϕ(x) ≤ 1 for every ϕ ∈ Γ} (called the polar of Γ).
(i) A0 is convex, weak* closed in X ∗ , and 0 ∈ A0 for A ⊂ X.
(ii) Γ0 is convex, weakly closed in X, and 0 ∈ Γ0 for Γ ⊂ X ∗ .
(iii) A ⊂ (A0 )0 and Γ ⊂ (Γ0 )0 always for A ⊂ X and Γ ⊂ X ∗ .
(iv) If A ⊂ X is convex, weakly closed, and 0 ∈ A, then (A0 )0 = A.
(v) If Γ ⊂ X ∗ is convex, weak* closed, and 0 ∈ Γ, then (Γ0 )0 = Γ.
[Hint: (iv) If z ∈ X \ A, by [110] and by the fact that 0 ∈ A, there exist ϕ ∈ X ∗ and α > 0 with
0 ≤ sup{ϕ(x) : x ∈ A} < α < ϕ(z). Replacing ϕ by ϕ/α we get sup{ϕ(x) : x ∈ A} < 1 < ϕ(z).
/ (A0 )0 . (v) If ψ ∈ X ∗ \Γ, then by [110′ ] applied to (X ∗ , weak∗ ), after a positive
Then ϕ ∈ A0 and z ∈
scaling, we get x ∈ X such that sup{ϕ(x) : ϕ ∈ Γ} < 1 < ψ(x). Then x ∈ Γ0 and ψ ∈
/ (Γ0 )0 .]

Exercise-13 below gives a necessary and sufficient condition for a new norm µ∗ on X ∗ to be a
dual norm to some norm µ on X, i.e., to have µ∗ (ϕ) = sup{ϕ(x) : µ(x) ≤ 1} for every ϕ ∈ X ∗ .

Exercise-13: An equivalent norm µ∗ on X ∗ is a dual norm to an equivalent norm on X ⇔ the

closed unit ball Γ := {ϕ ∈ X ∗ : µ∗ (ϕ) ≤ 1} of µ∗ is weak* closed in X ∗ . [Hint: ⇐: Note that
Γ0 := {x ∈ X : |ϕ(x)| ≤ 1 for every ϕ ∈ Γ} since ϕ ∈ Γ iff −ϕ ∈ Γ; and (Γ0 )0 = Γ by Exercise-12.
Define µ(y) = inf{α > 0 : α−1 y ∈ Γ0 }.]

[111] Let X be a separable real Banach space. Then,

(i) X admits an equivalent norm that is strictly convex.
(ii) X admits an equivalent norm that is smooth.
(iii) X admits an equivalent norm that is both strictly convex and smooth.

Proof. (i) Let {xn : n ∈ N} be a dense subset of SX . Let ϕn ∈ SX ∗ be a supporting functional of

xn , i.e., ϕn (xn ) = ∥xn ∥ = 1. Then T : X → l2 defined as T x = (2−n ϕn (x)) is an injective bounded
linear operator (for injectivity show that T x ̸= 0 for every x ∈ SX ). Also l2 is strictly convex.
Hence by Exercise-10, µ on X defined as µ(x) := ∥x∥ + ∥T x∥ is an equivalent strictly convex norm.
 TOPICS FROM ANALYSIS - PART 1/3 19

(ii) Let {xn : n ∈ N} be a dense subset of SX . Then T : X ∗ → l2 defined as T ϕ = (2−n ϕ(xn )) is an

injective bounded linear operator. Hence µ∗ on X ∗ defined as µ∗ (ϕ) := ∥ϕ∥ + ∥T ϕ∥ is an equivalent
strictly convex norm on X ∗ by Exercise-10. We claim that Γ := {ϕ ∈ X ∗ : µ∗ (ϕ) ≤ 1} is weak*
closed in X ∗ . Let ψ be a weak* limit point of Γ. To show ψ ∈ Γ, it suffices to show µ∗ (ψ) ≤ 1 + 4ε
for any ε > 0. Given ε > 0, choose x, y ∈ SX with ∥ψ∥ ≤ |ψ(x)| + ε and ∥T ψ∥ ≤ |T ψ(y)| + ε.
Then choose ϕ ∈ Γ with |ϕ(x) − ψ(x)| ≤ ε and |ϕ(y) − ψ(y)| ≤ ε/∥T ∥. Then ∥ψ∥ ≤ ∥ϕ∥ + 2ε and
∥T ψ∥ ≤ ∥T ϕ∥ + 2ε. Hence µ∗ (ψ) ≤ µ∗ (ϕ) + 4ε ≤ 1 + 4ε. Thus Γ is weak* closed. By Exercise-13,
there is an equivalent norm µ on X such that µ∗ is dual to µ. By Exercise-9, (X, µ) is smooth.

(iii) (Sketch) We may assume X is smooth by part (ii). Let T : X → l2 be an injective bounded
√
linear operator, and define ρ(x) = ∥x∥2 + ∥T x∥2 , which is an equivalent strictly convex norm on
X by Exercise-10. It can also be shown that ρ is smooth since (X, ∥ · ∥), l2 are smooth. 

For further reading: Gateaux and Frechet differentiability, uniformly convex/smooth norm.

5. Krein-Milman theorem

Krien-Milman theorem is true on locally convex topological vector spaces, but we will state it in
the setting of Banach spaces. Let X be a real Banach space throughout this section.

Definition and examples: If C ⊂ X is convex, let ext(C) = {x ∈ C : C \ {x} is convex}, the

collection of extreme points of C.
(i) If {0} ̸= Y ⊂ X is a vector subspace, then ext(Y ) = ∅.
(ii) ext(BX ) ⊂ ext(SX ) always, and equality holds when X is strictly convex.
(iii) For X = l1 , we have ext(BX ) = {±en : n ∈ N}. Hint: If x ∈ SX , xm ̸= 0 ̸= xn for some m ̸= n,
then assume xm > 0 by replacing x with −x if necessary, and note that x = xm y + (1 − xm )z is a
proper convex combination if we put y = em and z = (x − xm em )/(1 − xm ).
∫1
(iv) Consider L1 [0, 1] and let B = {f : ∥f ∥1 := 0 |f | ≤ 1}. If f ∈ B, let y ∈ [0, 1] be with
∫y
0 |f | = 1/2. Then f = (2f · 1[0,y] + 2f · 1[y,1] )/2 so that f ∈
/ ext(B). Thus ext(B) = ∅.
(v) Consider C([0, 1], R) with supremum norm and let B = {f : ∥f ∥ ≤ 1}. Then ext(B) = {1, −1}.
Hint: Let ∥f ∥ = 1 and f ̸= ±1. Fix c ∈ [0, 1] with f (c) ∈ (−1, 0) ∪ (0, 1). Choose g ∈ C([0, 1], R)
with the following properties: f ≤ g ≤ min{1, 2f } when f (x) ≥ 0, max{−1, 2f } ≤ g ≤ f when
f (x) ≤ 0 (in particular g = f when f (x) = 0), and g(c) ̸= f (c). Then putting h = 2f − g, we see
f = (g + h)/2 is a proper convex combination with ∥g∥ = ∥h∥ = 1, and hence f ∈
/ ext(B).

Definition: Let C ⊂ X be convex. A nonempty convex subset A ⊂ C is called an extreme subset

of C if whenever x ∈ A is a convex combination of y, z ∈ C, then y, z ∈ A. Note that C itself and
any singleton extreme point of C are extreme subsets of C.

Exercise-14: (i) If C ⊂ Rn is a nonempty compact convex set, then ext(C) ̸= ∅.

20 T.K.SUBRAHMONIAN MOOTHATHU

(ii) [Finite dimensional Krein-Milman] C = conv(ext(C)), i.e., C is the closed convex hull of its
extreme points, for any nonempty compact convex subset C ⊂ Rn .
(iii) Let A ⊂ B ⊂ C ⊂ X. If A is an extreme subset of B, and B is an extreme subset of C, then
A is an extreme subset of C.
(iv) Let C ⊂ X be convex. If Aj ⊂ C is an extreme subset of C for each j ∈ J and if A :=
∩
j∈J Aj ̸= ∅, then A is an extreme subset of C.
(v) Let C ⊂ X be convex. If A ⊂ C, then ext(A) need not be a subset of ext(C) (consider one ball
inside another). But if A is an extreme subset of C, then ext(A) ⊂ ext(C).
(vi) Let C ⊂ X be nonempty compact and convex. Let ϕ ∈ X ∗ , let [a, b] = ϕ(C) ⊂ R. Then
{x ∈ C : ϕ(x) = a} and {x ∈ C : ϕ(x) = b} are extreme subsets of C.
[Hint: (i) Take x ∈ C with ∥x∥ maximum. (ii) See p.52 of A.Barvinok, A Course in Convexity.]

[112] [Krein-Milman theorem] (i) Let C ⊂ X be nonempty, weakly compact, and convex. Then,
ext(C) ̸= ∅, and C = conv(ext(C)) (closed convex hull of ext(C)).
(ii) Let Γ ⊂ X ∗ be nonempty, weak* compact, and convex. Then, ext(Γ) ̸= ∅, and Γ is the weak*
closed convex hull of ext(Γ).

Proof. We prove only (i), and the proof of (ii) is similar. Let F be the collection of all weakly
compact extreme subsets of C, partially ordered by inclusion, and note C ∈ F. If (Aj )j∈J is a
∩
chain in F, then A′ := j∈J Aj is nonempty, weakly compact, and A′ is an extreme subset of C
by Exercise-14(iv). Hence A′ ∈ F is a lower bound for the chain (Aj ). By Zorn’s lemma, F has a
minimal element, say A. If |A| ≥ 2, then (since X ∗ separates points of X) there is ϕ ∈ X ∗ such
that ϕ(A) = [a, b] ⊂ R and a ̸= b. Then {x ∈ A : ϕ(x) = a}, {x ∈ A : ϕ(x) = b} ∈ F by parts
(iii) and (vi) of Exercise-14, and this contradicts the minimality of A. Thus A is a singleton, say
A = {x}. Then x ∈ ext(C), showing ext(C) ̸= ∅.

Let K = conv(ext(C)). Clearly K ⊂ C, and K is compact and convex. If there is z ∈ C \ K,

then by [110](i), there is ϕ ∈ X ∗ with sup{ϕ(x) : x ∈ K} < ϕ(z). Suppose ϕ(C) = [a, b], and let
E = {x ∈ C : ϕ(x) = b}. Then E is an extreme subset of C by Exercise-14(vi). Moreover E is
weakly compact, being closed in C. So there is y ∈ ext(E) ⊂ ext(C) ⊂ K by the first part of the
proof applied to E and by Exercise-14(v). Since y ∈ K, we have ϕ(y) < ϕ(z). On the other hand,
since z ∈ C and y ∈ ϕ−1 (b), we have ϕ(z) ≤ b = ϕ(y), a contradiction. Hence K = C. 

Seminar topics: Stone-Weierstrass theorem and Choquet’s theorem using Krein-Milman theorem,
see section 13.3 and 13.4 of P.Lax, Functional Analysis.

*****