)
Lecture Notes for Spring ‘08
Jeﬀ Schenker
Michigan State University
Email address: jeffrey@math.msu.edu
Contents
Comments and course information
Part 1. HahnBanach Theorem and Applications
Lecture 1. Linear spaces and the Hahn Banach Theorem
Lecture 2. Geometric HahnBanach Theorems
Lecture 3. Applications of HahnBanach
Part 2. Banach Spaces
Lecture 4. Normed and Banach Spaces
Lecture 5. Noncompactness of the Ball and Uniform Convexity
Lecture 6. Linear Functionals on a Banach Space
Lecture 7. Isometries of a Banach Space
Homework I
Part 3. Hilbert Spaces and Applications
Lecture 8. Scalar Products and Hilbert Spaces
Lecture 9. RieszFrechet and LaxMilgram Theorems
Lecture 10. Geometry of a Hilbert space and GramSchmidt process
Part 4. Locally Convex Spaces
Lecture 11. Locally Convex Spaces and Spaces of Test Functions
Lecture 12. Generation of a LCS by seminorms and Fr´echet Spaces
Lecture 13. The dual of an LCS
Lecture 14. Spaces of distributions
Lecture 15. Applications: solving some PDE’s
Lecture 16. The Dirichlet problem
Part 5. Weak Convergence and Weak Topology
iii
iv CONTENTS
Lecture 17. Dual of a Banach space
Lecture 18. RieszKakutani theorem
Lecture 19. Weak convergence
Lecture 20. Weak sequential compactness, weak
∗
convergence and the weak
topology
Lecture 21. An application: positive harmonic functions
Presentation topics
Homework II
Part 6. Convexity
Lecture 22. Convex sets in a Banach space
Lecture 23. Convex sets in a Banach space (II)
Lecture 24. KreinMilman and StoneWeierstrass
Lecture 25. Choquet type theorems
Part 7. Bounded Linear Maps
Lecture 26. Bounded Linear Maps
Lecture 27. Principle of Uniform Boundedness and Open Mapping Theorem
Lecture 28. The Spectrum of a Linear Map
Lecture 29. Some examples
Part 8. Compact Linear Maps
Lecture 30. Compact Maps
Lecture 31. Fredholm alternative
Lecture 32. Spectral Theory of Compact Maps
Homework III
Part 9. Compact Linear Maps in Hilbert Space
Lecture 33. Compact Symmetric Operators
Lecture 34. MinMax
Lecture 35. Functional calculus and polar decomposition
Comments and course information
These are lecture notes for Functional Analysis (Math 920), Spring 2008. The text for
this course is Functional Analysis by Peter D. Lax, John Wiley & Sons (2002), referred to
as “Lax” below. In some places I follow the book closely in others additional material and
alternative proofs are given.
Other excellent texts include
• M. Reed and B. Simon, Methods of Modern Mathematical Physics Vol. I: Functional
Analysis, Academic Press (1980).
• W. Rudin, Functional Analysis, McGrawHill, 2nd ed. (1991).
(As needed, these will be referred to below as “Reed and Simon” and “Rudin” respectively.)
v
Part 1
HahnBanach Theorem and Applications
LECTURE 1
Linear spaces and the Hahn Banach Theorem
Reading: Chapter 1 and ¸3.1 of Lax
Many objects in mathematics — particularly in analysis — are, or may be described in
terms of, linear spaces (also called vector spaces). For example:
(1) C(`) = space of continuous functions (R or C valued) on a manifold `.
(2) ¹(l) = space of analytic functions in a domain l ⊂ C.
(3) 1
p
(j) = ¦j integrable functions on a measure space /. j¦.
The key features here are the axioms of linear algebra,
Definition 1.1. A linear space A over a ﬁeld 1 (in this course 1 = R or C) is a set on
which we have deﬁned
(1) addition: r. µ ∈ A → r +µ ∈ A
and
(2) scalar multiplication: / ∈ 1. r ∈ A → /r ∈ A
with the following properties
(1) (A. +) is an abelian group (+ is commutative and associative and ∃ identity and
inverses.)
• identity is called 0 (“zero”)
• inverse of r is denoted −r
(2) scalar multiplication is
• associative: c(/r) = (c/)r,
• distributive: c(r +µ) = cr +/µ and (c +/)r = cr +/r,
and satisﬁes 1r = r.
Remark. It follows from the axioms that 0r = 0 and −r = (−1)r.
Recall from linear algebra that a set of vectors o ⊂ A is linearly independent if
n
j=1
c
j
r
j
= 0 with r
1
. . . . . r
n
∈ o =⇒ c
1
= = c
n
= 0
and that the dimension of A is the cardinality of a maximal linearly independent set in A.
The dimension is also the cardinality of a minimal spanning set, where the span of a set o
is the set
span o =
_
n
j=1
c
j
r
j
: c
1
. . . . . c
n
∈ R and r
1
. . . . . r
n
∈ o
_
.
and o is spanning, or spans A, if span o = A.
More or less, functional analysis is linear algebra done on spaces with inﬁnite dimension.
Stated this way it may seem odd that functional analysis is part of analysis. For ﬁnite
dimensional spaces the axioms of linear algebra are very rigid: there is essentially only
11
12 1. LINEAR SPACES AND THE HAHN BANACH THEOREM
one interesting topology on a ﬁnite dimensional space and up to isomorphism there is only
one linear space of each ﬁnite dimension. In inﬁnite dimensions we shall see that topology
matters a great deal, and the topologies of interest are related to the sort of analysis that
one is trying to do.
That explains the second word in the name ”functional analysis.” Regarding “functional,”
this is an archaic term for a function deﬁned on a domain of functions. Since most of
the spaces we study are function spaces, like C(`), the functions deﬁned on them are
“functionals.” Thus “functional analysis.” In particular, we deﬁne a linear functional to be
a linear map / : A → 1, which means
/(r +µ) = /(r) +/(µ) and /(cr) = c/(r) for all r. µ ∈ A and c ∈ 1.
Often, one is able to deﬁne a linear functional at ﬁrst only for a limited set of vectors
1 ⊂ A. For example, one may deﬁne the Riemann integral on 1 = C[0. 1], say, which
is a subset of the space 1[0. 1] of all bounded functions on [0. 1]. In most cases, as in the
example, the set 1 is a subspace:
Definition 1.2. A subset 1 ⊂ A of a linear space is a linear subspace if it is closed
under addition and scalar multiplication: µ
1
. µ
2
∈ 1 and c ∈ 1 =⇒ µ
1
+cµ
2
∈ 1 .
For functionals deﬁned, at ﬁrst, on a subspace of a linear space of R we have
Theorem 1.1 (Hahn (1927), Banach (1929)). Let A be a linear space over R and j a
real valued function on A with the properties
(1) j(cr) = cj(r) for all r ∈ A and c 0 (Positive homogeneity)
(2) j(r +µ) ≤ j(r) +j(µ) for all r. µ ∈ A (subadditivity).
If / is a linear functional deﬁned on a linear subspace of 1 and dominated by j, that is
/(µ) ≤ j(µ) for all µ ∈ 1 , then / can be extended to all of A as a linear functional dominated
by j, so /(r) ≤ j(r) for all r ∈ A.
Example. Let A = 1[0. 1] and 1 = C[0. 1]. On 1 , let /()) =
_
1
0
)(t)dt (Riemann
integral). Let j : 1 → R be j()) = sup¦[)(r)[ : r ∈ [0. 1]¦. Then j satisﬁes (1) and (2)
and /()) ≤ j()). Thus we can extend / to all of 1[0. 1]. We will return to this example and
see that we can extend / so that /()) ≥ 0 whenever ) ≥ 0. This deﬁnes a ﬁnitely additive
set function on all(!) subset of [0. 1] via j(o) = /(χ
S
). For Borel measurable sets it turns
out the result is Lebesgue measure. That does not follow from HahnBanach however.
The proof of HahnBanach is not constructive, but relies on the following result equivalent
to the axiom of choice:
Theorem 1.2 (Zorn’s Lemma). Let o be a partially ordered set such that every totally
ordered subset has an upper bound. Then o has a maximal element.
To understand the statement, we need
Definition 1.3. A partially ordered set o is a set on which an order relation c ≤ / is
deﬁned for some (but not necessarily all) pairs c. / ∈ o with the following properties
(1) transitivity: if c ≤ / and / ≤ c then c ≤ c
(2) reﬂexivity: if c ≤ c for all c ∈ o.
(Note that (1) asserts two things: that c and c are comparable and that c ≤ c.) A subset 1
of o is totally ordered if r. µ ∈ 1 =⇒ r ≤ µ or µ ≤ r. An element n ∈ o is an upper bound
for 1 ⊂ o if r ∈ 1 =⇒ r ≤ n. A maximal element : ∈ o satisﬁes : ≤ / =⇒ : = /.
1. LINEAR SPACES AND THE HAHN BANACH THEOREM 13
Proof of HahnBanach. To apply Zorn’s Lemma, we need a poset,
o = ¦extensions of / dominated by j¦ .
That is o consists of pairs (/
t
. 1
t
) with /
t
a linear functional deﬁned on a subspace 1
t
⊃ 1
so that
/
t
(µ) = /(µ). µ ∈ 1 and /
t
(µ) ≤ j(µ). µ ∈ 1
t
.
Order o as follows
(/
1
. 1
1
) ≤ (/
2
. 1
2
) ⇐⇒ 1
1
⊂ 1
2
and /
2
[
Y
1
= /
1
.
If 1 is a totally ordered subset of o, let (/. 1 ) be
1 =
_
¦1
t
: (/
t
. 1
t
) ∈ 1¦
and
/(µ) = /
t
(µ) for µ ∈ 1
t
.
Since 1 is totally ordered the deﬁnition of / is unambiguous. Clearly (/. 1 ) is an upper
bound for 1. Thus by Zorn’s Lemma there exists a maximal element (/
+
. 1
+
).
To ﬁnish, we need to see that 1
+
= A. It suﬃces to show that (/
t
. 1
t
) has an extension
whenever 1
t
,= A. Indeed, let r
0
∈ A. We want /
tt
on 1
tt
= ¦cr
0
+µ : µ ∈ 1. c ∈ R¦. By
linearity we need only deﬁne /
tt
(r
0
). The constraint is that we need
c/
tt
(r
0
) +/
t
(µ) ≤ j(cr
0
+µ)
for all c. µ. Dividing through by [c[, since 1
t
is a subspace, we need only show
±/
tt
(r
0
) ≤ j(µ ±r
0
) −/
t
(µ) (1.1)
for all µ ∈ 1
t
. We can ﬁnd /
tt
(r
0
) as long as
/
t
(µ
t
) −j(µ
t
−r
0
) ≤ j(r
0
+µ) −/
t
(µ) for all µ. µ
t
∈ 1
t
. (1.2)
or equivalently
/
t
(µ
t
+µ) ≤ j(r
0
+µ) +j(µ
t
−r
0
) for all µ. µ
t
∈ 1
t
. (1.3)
Since
/
t
(µ
t
+µ) ≤ j(µ
t
+µ) = j(µ
t
−r
0
+µ +r
0
) ≤ j(r
0
+µ) +j(µ
t
−r
0
).
(1.3), and thus (1.2), holds. So we can satisfy (1.1).
In ﬁnite dimensions, one can give a constructive proof involving only ﬁnitely many choices.
In inﬁnite dimensions the situation is a quite a bit diﬀerent, and the approach via Zorn’s
lemma typically involves uncountably many “choices.”
LECTURE 2
Geometric HahnBanach Theorems
Reading: ¸3.2, ¸3.3 of Lax.
We may use HahnBanach to understand something of the geometry of linear spaces. We
want to understand if the following picture holds in inﬁnite dimension:
Figure 2.1. Separating a point from a convex set by a line hyperplane
Definition 2.1. A set o ⊂ A is convex if for all r. µ ∈ o and t ∈ [0. 1] we have
tr + (1 −t)µ ∈ o.
Definition 2.2. A point r ∈ o ⊂ A is an interior point of o if for all µ ∈ A ∃ε 0 s.t.
[t[ < ε =⇒ r +tµ ∈ o.
Remark. We can a deﬁne a topology using this notion, letting l ⊂ A be open ⇐⇒
all r ∈ l are interior. From the standpoint of abstract linear algebra this seems to be
a “natural” topology on A. In practice, however, it has way too many open sets and we
work with weaker topologies that are relevant to the analysis under considerations. Much of
functional analysis centers around the interplay of diﬀerent topologies.
We are aiming at the following
21
22 2. GEOMETRIC HAHNBANACH THEOREMS
Theorem 2.1. Let 1 be a nonempty convex subset of A, a linear space over R, and
suppose 1 has at least one interior point. If µ ,∈ 1 then ∃ a linear functional / : A →R s.t.
/(r) ≤ /(µ) for all r ∈ 1. (2.1)
with strict inequality for all interior points r of 1.
This is the “hyperplane separation theorem,” essentially validates the picture drawn
above. A set of the form ¦/(r) = c¦ with / a linear functional is a “hyperplane” and the sets
¦/(r) < c¦ are “half spaces.”
To accomplish the proof we will use HahnBanach. We need a dominating function j.
Definition 2.3. Let 1 ⊂ A be convex and suppose 0 is an interior point. The gauge
of 1 (with respect to the origin) is the function j
K
: A →R deﬁned as
j
K
(r) = inf¦c : c 0 and
r
c
∈ 1¦.
(Note that j
K
(r) < ∞ for all r since 0 is interior.)
Lemma 2.2. j
K
is positive homogeneous and subadditive.
Proof. Positive homogeneity is clear (even if 1 is not convex). To prove subadditivity
we use convexity. Consider j
K
(r +µ). Let c. / be such that r,c. µ,/ ∈ 1. Then
t
r
c
+ (1 −t)
µ
/
∈ 1 ∀t ∈ [0. 1].
so
r +µ
c +/
=
c
c +/
r
c
+
/
c +/
µ
/
∈ 1.
Thus j
K
(r +µ) ≤ c +/, and optimizing over c. / we obtain the result.
Proof of hyperplane separation thm. Suﬃces to assume 0 ∈ 1 is interior and
c = 1. Let j
K
be the gauge of 1. Note that j
K
(r) ≤ 1 for r ∈ 1 and that j
K
(r) < 1 if r is
interior, as then (1 + t)r ∈ 1 for small t 0. Conversely if j
K
(r) < 1 then r is an interior
point (why?), so
j
K
(r) < 1 ⇐⇒ r ∈ 1.
Now deﬁne /(µ) = 1, so /(cµ) = c for c ∈ R. Since µ ,∈ 1 it is not an interior point and
so j
K
(µ) ≥ 1. Thus j
K
(cµ) ≥ c for c ≥ 0 and also, trivially, for c < 0 (since j
K
≥ 0). Thus
j
K
(cµ) ≥ /(cµ)
for all c ∈ R. By HahnBanach, with 1 the one dimensional space ¦cµ¦, / may be extended
to all of r so that j
K
(r) ≥ /(r) which implies (2.1).
An extension of this is the following
Theorem 2.3. Let H. ` be disjoint convex subsets of A, at least one of which has an
interior point. Then H and ` can be separate by a hyperplane /(r) = c: there is a linear
functional / and c ∈ R such that
/(n) ≤ c ≤ /(·)∀n ∈ H. · ∈ `.
2. GEOMETRIC HAHNBANACH THEOREMS 23
Proof. The proof rests on a trick of applying the hyperplane separation theorem with
the set
1 = H −` = ¦n −· : n ∈ H and · ∈ `¦
and the point µ = 0. Note that 0 ,∈ 1 since H ∪ ` = ∅. Since 1 has an interior point
(why?), we see that there is a linear functional such that /(r) ≤ 0 for all r ∈ 1. But then
/(n) ≤ /(·) for all n ∈ H, · ∈ `.
In many applications, one wants to consider a vector space A over C. Of course, then A
is also a vector space over R so the real HahnBanach theorem applies. Using this one can
show the following
Theorem 2.4 (Complex HahnBanach, Bohenblust and Sobczyk (1938) and Soukhom
linoﬀ (1938)). Let A be a linear space over C and j : A → [0. ∞) such that
(1) j(cr) = [c[j(r)∀c ∈ C. r ∈ A.
(2) j(r +µ) ≤ j(r) +j(µ) (subadditivity).
Let 1 be a C linear subspace of A and / : 1 →C a linear functional such that
[/(µ)[ ≤ j(µ) (2.2)
for all µ ∈ 1 . Then / can be extended to all of A so that (2.2) holds for all µ ∈ A.
Remark. A function j that satisﬁes (1) and (2) is called a seminorm. It is a norm if
j(r) = 0 =⇒ r = 0.
Proof. Let /
1
(µ) = Re /(µ), the real part of /. Then /
1
is a real linear functional and
−/
1
(iµ) = −Re i/(µ) = Im/(µ), the imaginary part of /. Thus
/(µ) = /
1
(µ) −i/
1
(iµ). (2.3)
Clearly [/
1
(µ)[ ≤ j(µ) so by the real HahnBanach theorem we can extend /
1
to all of
A so that /
1
(µ) ≤ j(µ) for all µ ∈ A. Since −/
1
(µ) = /
1
(−µ) ≤ j(−µ) = j(µ), we have
[/
1
(µ)[ ≤ j(µ) for all µ ∈ A. Now deﬁne the extension of / via (2.3). Given µ ∈ A let
θ = arg ln /(µ). Thus /(µ) = e
iθ
/
1
(e
−iθ
µ) (why?). So,
[/(µ)[ = [/
1
(e
−iθ
µ)[ ≤ j(µ).
Lax gives another beautiful extension of HahnBanach, due to Agnew and Morse, which
involves a family of commuting linear maps. We will cover a simpliﬁed version of this next
time.
LECTURE 3
Applications of HahnBanach
Reading: ¸4.1, ¸4.2 of Lax.
To get an idea what one can do with the HahnBanach theorem let’s consider a concrete
application on the linear space A = 1(o) of all real valued bounded functions on a set o.
1(o) has a natural partial order, namely r ≤ µ if r(:) ≤ µ(:) for all : ∈ o. If 0 ≤ r then r
is nonnegative. On 1(o) a positive linear functional / satisﬁes /(µ) ≥ 0 for all µ ≥ 0.
Theorem 3.1. Let 1 be a linear subspace of 1(o) that contains µ
0
≥ 1, so µ
0
(:) ≥ 1
for all : ∈ o. If / is a positive linear functional on 1 then / can be extended to all of 1 as
a positive linear functional.
This theorem can be formulated in an abstract context as follows. The nonnegative
functions form a cone, where
Definition 3.1. A subset 1 ⊂ A of a linear space over R is a cone if tr + :µ ∈ 1
whenever r. µ ∈ 1 and t. : ≥ 0. A linear functional on A is 1nonnegative if /(r) ≥ 0 for
all r ∈ 1.
Theorem 3.2. Let 1 ⊂ A be a cone with an interior point r
0
. If 1 is a subspace
containing r
0
on which is deﬁned a 1 ∩1 positive linear functional /, then / has an extension
to A which is 1positive.
Proof. Deﬁne a dominating function j as follows
j(r) = inf¦/(µ) : µ −r ∈ 1. µ ∈ 1 ¦.
Note that µ
0
− tr ∈ 1 for some t 0 (since µ
0
is interior to 1), so
1
t
µ
0
− r ∈ 1. This
shows that j(r) is well deﬁned. It is clear that j is positive homegeneous. To see that it is
subadditive, let r
1
. r
2
∈ A and let µ
1
. µ
2
be so that µ
j
−r
j
∈ 1. Then µ
1
+µ
2
−(r
1
+r
2
) ∈ 1,
so
j(r
1
+r
2
) ≤ /(µ
1
) +/(µ
2
).
Minimizing over µ
1,2
gives subadditivity.
Since /(r) = /(r −µ) +/(µ) ≤ /(µ) if r ∈ 1 and µ −r ∈ 1 we conclude that /(r) ≤ j(r)
for all r ∈ 1 . By HahnBanach we may extend / to all of A so that /(r) ≤ j(r) for all r.
Now let r ∈ 1. Then j(−r) ≤ 0 (why?), so −/(r) = /(−r) ≤ 0 which shows that / is
1positive.
The theorem on 1(o) follows from the Theorem 3.2 once we observe that the condition
µ
0
≥ 1 implies that µ
0
is an interior point of the cone of positive functions. The linear
functional that one constructs in this way is monotone:
r ≤ µ =⇒ /(r) ≤ /(µ).
which is in fact equivalent to positivity. This can allow us to do a little bit of analysis, even
though we haven’t introduced notions of topology or convergence.
31
32 3. APPLICATIONS OF HAHNBANACH
To see how this could work, let’s apply the result to the Riemann integral on C[0. 1].
We conclude the existence of a positive linear functional / : 1[0. 1] → R that gives /()) =
_
1
0
)(r)dr. This gives an “integral” of arbitrary bounded functions, without a measurability
condition. Since the integral is a linear functional, it is ﬁnitely additive. Of course it is
not countably additive since we know from real analysis that such a thing doesn’t exist.
Furthermore, the integral is not uniquely deﬁned.
Nonetheless, the extension is also not arbitrary, and the constraint of positivity actually
pins down /()) for many functions. For example, consider χ
(a,b)
(the characteristic function
of an open interval). By positivity we know that
) ≤ χ
(a,b)
≤ p. ). p ∈ C[0. 1] =⇒
_
1
0
)(r)dr ≤ /(χ
(a,b)
) ≤
_
1
0
p(r)dr.
Taking the sup over ) and inf over p, using properties only of the Riemann integral, we see
that /(χ
(a,b)
) = / −c (hardly a surprising result). Likewise, we can see that /(χ
U
) = [l[ for
any open set, and by ﬁnite additivity /(χ
F
) = 1 −/(1
c
) = 1 −[1
c
[ = [1[ for any closed set
([ [ is Lebesgue measure). Finally if o ⊂ [0. 1] and
sup ¦/(1) : 1 closed and 1 ⊂ o¦ = inf ¦/(l) : l open and l ⊃ o¦
then /(χ
S
) must be equal to these two numbers. We see that /(χ
S
) = [o[ for any Lebesgue
measurable set. We have just painlessly constructed Lebesgue measure from the Riemann
integral without using any measure theory!
The rigidity of the extension is a bit surprising if we compare with what happens in
ﬁnite dimensions. For instance, consider the linear functional /(r. 0) = r deﬁned on 1 =
¦(r. 0)¦ ⊂ R
2
. Let 1 be the cone ¦(r. µ) : [µ[ ≤ αr¦. So / is 1 ∩ 1 positive. To extend /
to all of R
2
we need to deﬁne /(0. 1). To keep the extension positive we need only require
µ/(0. 1) + 1 ≥ 0
if [µ[ ≤ α. Thus we must have [/(0. 1)[ ≤ 1,α, and any choice in this interval will work.
Only when α = ∞ and the cone degenerates to a half space does the condition pin /(0. 1)
down. So some interesting things happen in ∞ dimensions.
A second example application assigning a limiting value to sequences
a = (c
1
. c
2
. . . .).
Let 1 denote the space of all bounded Rvalued sequences and let 1 denote the subspace of
sequences with a limit. We quickly conclude that there is an positive linear extension of the
positive linear functional lim to all of 1. We would like to conclude a little more, however.
After all, for convergent sequences,
limc
n+k
= limc
n
for any /.
To formalize this property we deﬁne a linear map on 1 by
1(c
1
. c
2
. . . .) = (c
2
. c
3
. . . .).
(1 is the “backwards shift” operator or “left translation.”) Here,
Definition 3.2. Let A
1
. A
2
be linear spaces over a ﬁeld 1. A linear map 1 : A
1
→ A
2
is a function such that
1(r +cµ) = 1(r) +c1(µ) ∀r. µ ∈ A
1
and c ∈ 1.
3. APPLICATIONS OF HAHNBANACH 33
Thus lim1a = lima for a ∈ 1. We would like the extension to have this property. Can
this be done?
The answer is given by the following a simpliﬁed version of the Theorem of Agnew and
Morse given in Chapter 3 of Lax:
Theorem 3.3. Let A be a linear space over R and let 1 : A → A be a linear map. Let
j : A →R be a positive homogeneous, subadditive function invariant under 1:
j(1r) = j(r) ∀r ∈ A.
Let / be a linear functional deﬁned on a subspace 1 ⊂ A invariant under 1,
1µ ∈ 1 ∀µ ∈ 1.
If / is invariant under 1 — /(1µ) = /(µ) — and dominated by j then / may be extended to
all of A so that the extension is also invariant under 1 and dominated by j.
Remark. Applying this to the space 1 of bounded sequences and letting
j(a) = limsup a.
we ﬁnd that there is an extension of the linear functional lim from 1 to all of 1 such that
for all a = (c
1
. c
2
. . . .)
(1) lim
n→∞
c
n+k
= lim
n→∞
c
n
for all /.
(2) liminf a ≤ lima ≤ limsup a . (The ﬁrst inequality follows from the second applied
to −a.)
Proof. Let us deﬁne
¡(r) = inf j(¹r).
where the inﬁmum is over all convex combinations of powers of 1,
¹ =
n
j=0
c
j
1
j
(3.1)
where 1
0
= 1, the identity operator on A, 1r = r, and c
j
is any ﬁnite, nonnegative sequence
with
j
c
j
= 1. Clearly ¡(r) is positive homogeneous and ¡(r) ≤ j(r). Also, since / and 1
are invariant under any ¹ of the form (3.1) we have
/(µ) ≤ ¡(µ) ∀µ ∈ 1.
To apply HahnBanach with ¡ as the dominating function we need to show it is sub
additivie. To this end, let r. µ ∈ A and let ¹. 1 be of the form (3.1) so that
j(¹r) ≤ ¡(r) +c and j(1µ) ≤ ¡(µ) +c.
Then ¹ ◦ 1 = 1 ◦ ¹ is of the form (3.1) and
¡(r +µ) ≤ j(¹1(r +µ)) ≤ j(1¹r) +j(¹1µ) ≤ j(¹r) +j(1µ) ≤ ¡(r) +¡(µ) + 2c.
Thus ¡ is subadditive.
So now we know that an extension exists to all of A with /(r) ≤ ¡(r) for all r. However,
¡(r −1r) ≤ j
_
1
:
n
j=0
1
j
(1 −1)r
_
=
1
:
j(r −1
n+1
r) ≤
1
:
(j(r) +j(−r)) → 0.
So ¡(r −1r) = 0 for all r and thus /(r) −/(1r) = 0 for all r.
Part 2
Banach Spaces
LECTURE 4
Normed and Banach Spaces
Reading: ¸5.1 of Lax.
The HahnBanach theorem made use of a dominating function j(r). When this function
is nonnegative, it can be understood roughly as a kind of “distance” from a point r to the
origin. For that to work, we should have j(r) 0 whenever r ,= 0. Such a function is called
a norm:
Definition 4.1. Let A be a linear space over 1 = R or C. A norm on A is a function
 : A → [0. ∞) such that
(1) r = 0 ⇐⇒ r = 0.
(2) r +µ ≤ r +µ (subadditivity)
(3) cr = [c[ r for all c ∈ 1 and r ∈ A (homogeneity).
A normed space is a linear space A with a norm .
The norm on a normed space gives a metric topology if we deﬁne the distance between
two points via
d(r. µ) = r −µ .
Condition 1 guarantees that two distinct points are a ﬁnite distance apart. Subadditivity
gives the triangle inequality. The metric is
(1) translation invariant: d(r +.. µ +.) = d(r. µ)
and
(2) homogeneous d(cr. cµ) = [c[d(r. µ).
Thus any normed space is a metric space and we have the following notions:
(1) a sequence r
n
converges to r, r
n
→ r, if d(r
n
. r) = r
n
−r → 0.
(2) a set l ⊂ A is open if for every r ∈ l there is a ball ¦µ : µ −r < c¦ ⊂ l.
(3) a set 1 ⊂ A is closed if A ¸ 1 is open.
(4) a set 1 ⊂ A is compact if every open cover of 1 has a ﬁnite subcover.
The norm deﬁnes the topology but not the other way around. Indeed two norms 
1
and 
2
on A are equivalent if there is c 0 such that
c r
1
≤ r
2
≤ c
−1
r
2
∀r ∈ A.
Equivalent norms deﬁne the same topology. (Why?)
Recall from real analysis that a metric space A is complete if every Cauchy sequence r
n
converges in A. In a normed space, a Cauchy sequence r
n
is one such that
∀c 0∃` ∈ N such that :. : ` =⇒ r
n
−r
m
 < c.
A complete normed space is called a Banach space.
41
42 4. NORMED AND BANACH SPACES
Not every normed space is complete. For example C[0. 1] with the norm
)
1
=
_
1
0
[)(r)[dr
fails to be complete. (It is, however, a complete space in the uniform norm, )
u
=
sup
x∈[0,1]
[)(r)[.) However, every normed space A has a completion, deﬁned abstractly as
a set of equivalence classes of Cauchy sequences in A. This space, denoted A is a Banach
space.
Examples of Normed and Banach spaces
(1) For each j ∈ [1. ∞) let
/
p
= ¦j summable sequences¦ = ¦(c
1
. c
2
. . . .) [
∞
j=1
[c
j
[
p
< ∞¦.
Deﬁne a norm on /
p
via
a
p
=
_
∞
j=1
[c
j
[
p
_1
p
.
Then /
p
is a Banach space.
(2) Let
/
∞
= ¦bounded sequences¦ = 1(N).
with norm
a
∞
= sup
j
[c
j
[. ()
Then /
∞
is a Banach space.
(3) Let
c
0
= ¦sequences converging to 0¦ = ¦(c
1
. c
2
. . . .) [ lim
j→∞
c
j
= 0¦.
with norm (). Then c
0
is a Banach space.
(4) Let
T = ¦sequences with ﬁnitely many nonzero terms¦
= ¦(c
1
. c
2
. . . .) [ ∃` ∈ `such that : ≥ ` =⇒ c
n
= 0¦.
Then for any j ≥ 1, T
p
= (T. 
p
) is a normed space which is not complete. The
completion of T
p
is isomorphic to /
p
.
(5) Let 1 ⊂ R
d
be a domain and let j ∈ [1. ∞).
(a) Let A = C
c
(1) be the space of continuous functions with compact support in
1, with the norm
)
p
=
__
D
[)(r)[
p
dr
_1
p
.
EXAMPLES OF NORMED AND BANACH SPACES 43
Then A is a normed space, which is not complete. Its completion is denoted
1
p
(1) and may be identiﬁed with the set of equivalence classes of measurable
functions ) : 1 →C such that
_
D
[)(r)[
p
dr < ∞ (Lebesgue measure).
with two functions ). p called equivalent if )(r) = p(r) for almost every r.
(b) Let A denote the set of C
1
functions on 1 such that
_
D
[)(r)[
p
dr < ∞ and
_
D
[∂
j
)(r)[
p
dr < ∞. , = 1. . . . . :.
Put the following norm on A,
)
1,p
=
_
_
D
[)(r)[
p
dr +
n
j=1
_
D
[∂
j
)(r)[
p
dr.
_1
p
.
Then A is a normed space which is not complete. Its completion is denoted
\
1,p
(1) and is called a Sobolev space and may be identiﬁed with the sub
space of 1
p
(1) consisting of (equivalence classes) of functions all of whose ﬁrst
derivatives are in 1
p
(1) in the sense of distributions. (We’ll come back to this.)
Note that a ∈ /
1
is summable,
j
c
j
≤ a
1
.
Theorem 4.1 (H¨ older’s Inqequality). Let 1 < j < ∞ and let ¡ be such that
1
j
+
1
¡
= 1.
If a ∈ /
p
and b ∈ /
q
then
ab = (c
1
/
1
. c
2
/
2
. . . .) ∈ /
1
and
¸
¸
¸
¸
¸
∞
j=1
c
j
/
j
¸
¸
¸
¸
¸
≤ a
p
b
q
.
Proof. First note that for two nonnegative numbers c. / it holds that
c/ ≤
1
j
c
p
+
1
¡
/
q
.
For j = ¡ = 2 this is the familiar “arithmeticgeometric mean” inequality which follows since
(c−/)
2
≥ 0. For general c. / this may be seen as follows. The function r → exp(r) is convex:
exp(tr + (1 −t)µ) ≤ t exp(r) + (1 −t) exp(µ). (Recall from calculus that a C
2
function ) is
convex if )
tt
≥ 0.) Thus,
c/ = exp(
1
j
ln c
p
+
1
¡
ln /
q
) ≤
1
j
exp(ln c
p
) +
1
¡
exp(ln /
q
) =
1
j
c
p
+
1
¡
/
q
.
Applying this bound coordinate wise and summing up we ﬁnd that
ab
1
≤
1
j
a
p
p
+
1
¡
b
q
q
.
44 4. NORMED AND BANACH SPACES
The result follows from this bound by “homogenization:” we have
a
p
. b
q
=⇒ ab ≤ 1.
so _
_
_
_
_
a
a
p
b
b
q
_
_
_
_
_
1
≤ 1
from which the desired estimate follows by homogeneity of the norm.
Similarly, we have
Theorem 4.2 (H¨older’s Inequality). Let 1 < j < ∞ and let ¡ be the conjugate exponent.
If ) ∈ 1
p
(1) and p ∈ 1
q
(1) then )p ∈ 1
1
(1) and
_
D
[)(r)p(r)[dr ≤ )
p
p
q
.
LECTURE 5
Noncompactness of the Ball and Uniform Convexity
Reading: ¸5.1 and ¸5.2 of Lax.
First a few more deﬁnitions:
Definition 5.1. A normed space A over 1 = R or C is called separable if it has a
countable, dense subset.
Most spaces we consider are separable, with a few notable exceptions.
(1) The space ` of all signed (or complex) measures j on [0. 1], say, with norm
j =
_
1
0
[j[(dr).
Here [j[ denotes the total variation of j,
[j[(¹) = sup
Partitions A
1
, . . . , An of A
n
j=1
[j(¹
j
)[.
This space is a Banach space. Since the point mass
δ
x
(¹) =
_
1 r ∈ ¹
0 r ,∈ ¹
is an element of ¹ and δ
x
−δ
y
 = 2 if r ,= µ, we have an uncountable family of
elements of ` all at a ﬁxed distance of one another. Thus there can be no countable
dense subset. (Why?)
(2) /
∞
is also not separable. To see this, note that to each subset of ¹ ⊂ N we may
associate the sequence χ
A
, and
χ
A
−χ
B

∞
= 1
if ¹ ,= 1.
(3) /
p
is separable for 1 ≤ j < ∞.
(4) 1
p
(1) is separable for 1 ≤ j < ∞.
(5) 1
∞
(1) is not separable.
Noncompactness of the Unit Ball
Theorem 5.1 (F. Riesz). Let A be a normed linear space. Then the closed unit ball
1
1
(0) = ¦r : A ≤ 1¦ is compact if and only if A is ﬁnite dimensional.
Proof. The fact that the unit ball is compact if A is ﬁnite dimensional is the Heine
Borel Theorem from Real Analysis.
To see the converse, we use the following
51
52 5. NONCOMPACTNESS OF THE BALL AND UNIFORM CONVEXITY
Lemma 5.2. Let 1 be a closed proper subspace of a normed space A. Then there is a
unit vector . ∈ A, . = 1, such that
. −µ
1
2
∀µ ∈ 1.
Proof of Lemma. Since 1 is proper, ∃r ∈ A ¸ 1 . Then
inf
y∈Y
r −µ = d 0.
(This is a property of closed sets in a metric space.) We do not know the existence of a
minimizing µ, but we can certainly ﬁnd µ
0
such that
0 < r −µ
0
 < 2d.
Let . =
x−y
0
x−y
0

. Then
. −µ =
r −µ
0
−r −µ
0
 µ
r −µ
0

≥
d
2d
=
1
2
.
Returning to the proof of the Theorem, we will use the fact that every sequence in a
compact metric space has a convergent subsequence. Thus it suﬃces to show that if A is
inﬁnite dimensional then there is a sequence in 1
1
(0) with no convergent subsequence.
Let µ
1
be any unit vector and construct a sequence of unit vectors, by induction, so that
µ
k
−µ
1
2
∀µ ∈ span¦µ
1
. . . . . µ
k−1
¦.
(Note that span¦µ
1
. . . . . µ
k−1
¦ is ﬁnite dimensional, hence complete, and thus a closed sub
space of A.) Since A is ﬁnite dimensional the process never stops. No subsequence of µ
j
can be Cauchy, much less convergent.
Uniform convexity
The following theorem may be easily shown using compactness:
Theorem 5.3. Let A be a ﬁnite dimensional linear normed space. Let 1 be a closed
convex subset of A and . any point of A. Then there is a unique point of 1 closer to . than
any other point of 1. That is there is a unique solution µ
0
∈ 1 to the minimization problem
µ
0
−. = inf
y∈K
µ −. . ()
Try to prove this theorem. (The existence of a minimizer follows from compactness; the
uniqueness follows from convexity.)
The conclusion of theorem does not hold in a general inﬁnite dimensional space. Nonethe
less there is a property which allows for the conclusion, even though compactness fails!
Definition 5.2. A normed linear space A is uniformly convex if there is a function
c : (0. ∞) → (0. ∞), such that
(1) c is increasing.
(2) lim
r→0
c(:) = 0.
(3)
_
_
1
2
(r +µ)
_
_
≤ 1 −c(r −µ) for all r. µ ∈ 1
1
(0), the unit ball of A.
UNIFORM CONVEXITY 53
Theorem 5.4 (Clarkson 1936). Let A be a uniformly convex Banach space, 1 a closed
convex subset of A, and . any point of A. Then the minimization problem () has a unique
solution µ
0
∈ 1.
Proof. If . ∈ 1 then µ
0
= . is the solution, and is clearly unique.
When . ,∈ 1, we may assume . = 0 (translating . and 1 if necessary). Let
: = inf
y∈K
µ .
So : 0. Now let µ
n
∈ 1 be a minimizing sequence, so
µ
n
 → :.
Now let r
n
= µ
n
, µ
n
, and consider
1
2
(r
n
+r
m
) =
1
2 µ
n

µ
n
+
1
2 µ
m

µ
m
=
_
1
2 µ
n

+
1
2 µ
m

_
(tµ
n
+ (1 −t)µ
m
)
for suitable t. So tµ
n
+ (1 −t)µ
m
∈ 1 and
tµ
n
+ (1 −t)µ
m
 ≥ :.
Thus
1 −c(r
n
−r
m
) ≥
1
2
_
:
µ
n

+
:
µ
m

_
→ 1.
We conclude that r
n
is a Cauchy sequence, from which it follows that µ
n
is Cauchy. The
limit µ
0
∈ lim
n
µ
n
exists in 1 since A is complete and 1 is closed. Clearly µ
0
 = :.
Warning: Not every Banach space is uniformly convex. For example, the space C(1) of
continuous functions on a compact set 1 is not uniformly convex. It may even happen that
_
_
_
_
1
2
() +p)
_
_
_
_
∞
= 1
for unit vectors ) and p. (They need only have disjoint support.) Lax gives an example of
a closed convex set in C[−1. 1] in which the minimization problem () has no solution.
It can also happen that a solution exists but is not unique. For example, in C[−1. 1] let
1 = ¦functions that vanish on [−1. 0]¦. and let ) = 1 on [−1. 1]. Clearly
sup
x
[)(r) −p(r)[ ≥ 1 ∀p ∈ 1.
and the distance 1 is attained for any p ∈ 1 that satisﬁes 0 ≤ p(r) ≤ 1.
LECTURE 6
Linear Functionals on a Banach Space
Reading: ¸8.1 and ¸8.2 of Lax
Definition 6.1. A linear functional / : A → 1 on a normed space A over 1 = R or C
is bounded if there is c < ∞ such that
[/(r)[ ≤ c r ∀r ∈ A.
The inf over all such c is the norm / of /,
/ = sup
x,=0, x∈X
[/(r)[
r
. ()
Theorem 6.1. A linear functional / on a normed space is bounded if and only if it is
continuous.
Proof. It is useful to recall that
Theorem 6.2. Let A. 1 be metric spaces. Then ) : A → 1 is continuous if and only if
)(r
n
) is a convergent sequence in 1 whenever r
n
is convergent in A.
Remark. Continuity =⇒ the sequence condition in any topological space. That the
sequence condition =⇒ continuity follows from the fact that the topology has a countable
basis at a point. (In a metric space, 1
2
−n(r), say.) Speciﬁcally, suppose the function is not
continuous. Then there is an open set l ⊂ 1 such that )
−1
(l) is not open. So there is
r ∈ )
−1
(l) such that for all : 1
2
−n(r) ,⊂ )
−1
(l). Now let r
n
be a sequence such that
(1) r
n
∈ 1
2
−n(r) ¸ )
−1
(l) for : odd.
(2) r
n
= r for : even.
Clearly r
n
→ r. However, )(r
n
) cannot converge since lim
k
)(r
2k
) = )(r) and any limit
point of )(r
2k+1
) lies in the closed set l
c
containing all the points )(r
2k+1
).
In the normed space context, note that
[/(r
n
) −/(r)[ ≤ / r
n
−r .
so boundedness certainly implies continuity.
Conversely, if / is unbounded then we can ﬁnd vectors r
n
so that /(r
n
) ≥ :r
n
. Since
this inequality is homogeneous under scaling, we may suppose that r
n
 = 1,
√
:, say. Thus
r
n
→ 0 and /(r
n
) → ∞, so / is not continuous.
The set A
t
of all bounded linear functionals on A is called the dual of A. It is a linear
space, and in fact a normed space under the norm (). (It is straightforward to show that
() deﬁnes a norm.)
Theorem 6.3. The dual A
t
of a normed space A is a Banach space.
61
62 6. LINEAR FUNCTIONALS ON A BANACH SPACE
Proof. We need to show A
t
is complete. Suppose /
m
is a Cauchy Sequence. Then for
each r ∈ A we have
[/
n
(r) −/
m
(r)[ ≤ /
n
−/
m
 r .
so /
n
(r) is a Cauchy sequence of scalars. Let
/(r) = lim
n→∞
/
n
(r) ∀r ∈ A.
It is easy to see that / is linear. Let us show that it is bounded. Since [ /
n
 −/
m
 [ ≤
/
n
−/
m
 (this follows from subadditivity), we see that the sequence /
n
 is Cauchy, and
thus bounded. So,
[/(r)[ ≤ sup
n
/
n
 r
and / is bounded. Similarly,
[/
n
(r) −/(r)[ ≤ sup
m≥n
/
n
−/
m
 r .
so
/
n
−/ ≤ sup
m≥n
/
n
−/
m

and it follows that /
n
→ /.
Of course, all of this could be vacuous. How do we know that there are any bounded
linear functionals? Here the HahnBanach theorem provides the answer.
Theorem 6.4. Let µ
1
. . . . . µ
N
be ` linearly independent vectors in a normed space A
and α
1
. . . . . α
N
arbitrary scalars. Then there is a bounded linear functional / ∈ A
t
such that
/(µ
j
) = α
j
. , = 1. . . . . `.
Proof. Let 1 = span¦µ
1
. . . . . µ
N
¦ and deﬁne / on 1 by
/(
N
j=1
/
j
) =
N
j=1
/
j
α
j
.
(We use linear independence here to guarantee that / is well deﬁned.) Clearly, / is linear.
Furthermore, since 1 is ﬁnite dimensional / is bounded.
(Explicitly, since any two norms on a ﬁnite dimensional space are equivalent, we can ﬁnd
c such that
j
[/
j
[ µ
j
 ≤ c
_
_
_
_
_
j
/
j
µ
j
_
_
_
_
_
.
Thus, /(µ) ≤ c sup
j
[α
j
[ µ for µ ∈ 1 .)
Thus / is a linear functional on 1 dominated by the norm . By the HahnBanach
theorem, it has an extension to A that is also dominated by , i.e., that is bounded.
A closed subspace 1 of a normed space A is itself a normed space. If A is a Banach
space, so is 1 . A linear functional / ∈ A
t
on A can be restricted to 1 and is still bounded.
That is there is a restriction map 1 : A
t
→ 1
t
such that
1(/)(µ) = /(µ) ∀µ ∈ 1.
It is clear that 1 is a linear map and that
1(/) ≤ / .
6. LINEAR FUNCTIONALS ON A BANACH SPACE 63
The HahnBanach Theorem shows that 1 is surjective. On the other hand, unless 1 = A,
the kernel of 1 is certainly nontrivial. To see this, let r be a vector in A ¸ 1 and deﬁne /
on span¦r¦ ∪ 1 by
/(cr +µ) = c ∀c ∈ 1 and µ ∈ 1.
Then / is bounded on the closed subspace span¦r¦ ∪ 1 and by HahnBanach has a closed
extension. Clearly 1(/) = 0. The kernel of 1 is denoted 1
⊥
, so
1
⊥
= ¦/ ∈ A
t
: /(µ) = 0 ∀µ ∈ 1 ¦.
and is a Banach space.
Now the quotient space A,1 is deﬁned to be the set of “cosets of Y,”
¦1 +r : r ∈ A¦.
The coset 1 +r is denoted [r]. The choice of label r is, of course, not unique as r +µ with
µ ∈ 1 would do just as well. It is a standard fact that
[r
1
] + [r
2
] = [r
1
+r
2
]
gives a well deﬁned addition on 1,A so that it is a linear space.
Lemma 6.5. If 1 is a closed subspace of a normed space then
[r] = inf
y∈Y
r +µ
is a norm on A,1 . If A is a Banach space, so is A,1 .
Proof. Exercise.
A bounded linear functional that vanishes on 1 , that is an element / ∈ 1
⊥
, can be
understood as a linear functional on A,1 since the deﬁnition
/([r]) = /(r)
is unambiguous. Thus we have a map J : 1
⊥
→ (A,1 )
t
deﬁned by J(/)([r]) = /(r).
Conversely, there is a bounded linear map Π : A → A,1 given by Π(r) = [r] and any
linear functional / ∈ (A,1 )
t
pulls back to a bounded linear functional / ◦ Π in 1
⊥
. Clearly
J(/ ◦ Π) = /. Thus we have, loosely, that
(A,1 )
t
= 1
⊥
.
LECTURE 7
Isometries of a Banach Space
Reading: ¸5.3 of Lax.
Definition 7.1. Let A. 1 be normed spaces. A linear map 1 : A → 1 is bounded if
there is c 0 such that
1(r) ≤ c r .
The norm of 1 is the smallest such c, that is
1 = sup
x,=0
1(r)
r
.
Theorem 7.1. A linear map 1 : A → 1 between normed spaces A and 1 is continuous
if and only if it is bounded.
Remark. The proof is a simple extension of the corresponding result for linear function
als.
An isometry of normed spaces A and 1 is a map ` : A → 1 such that
(1) ` is surjective.
(2) `(r) −`(µ) = r −µ.
Clearly translations 1
u
: A → A, 1
u
(r) = r + n are isometries of a normed linear space. A
linear map 1 : A → 1 is an isometry if 1 is surjective and
1(r) = r ∀r ∈ A.
A map ` : A → 1 is aﬃne if `(r) − `(0) is linear. So, ` is aﬃne if it is the
composition of a linear map and a translation.
Theorem 7.2 (Mazur and Ulam 1932). Let A and 1 be normed spaces over R. Any
isometry ` : A → 1 is an aﬃne map.
Remark. The theorem conclusion does not hold for normed spaces over C. In that
context any isometry is a real aﬃne map (`(r) −`(0) is real linear), but not necessarily
a complexaﬃne map. For example on C([0. 1]. C) the map ) → ) (complex conjugation) is
an isometry and is not complex linear.
Proof. It suﬃces to show `(0) = 0 =⇒ ` is linear. To prove linearity it suﬃces to
show
`
_
1
2
(r +µ)
_
=
1
2
(`(r) +`(µ)) ∀r. µ ∈ A.
(Why?)
Let r and µ be points in A and . =
1
2
(r +µ). Note that
r −. = µ −. =
1
2
r −µ .
71
72 7. ISOMETRIES OF A BANACH SPACE
so . is “halfway between r and µ.” Let
r
t
= `(r). µ
t
= `(µ). .
t
= `(.).
We need to show 2.
t
= r
t
+µ
t
. Since ` is an isometry,
r
t
−.
t
 = µ
t
−.
t
 =
1
2
r
t
−µ
t
 .
and all of these are equal to
1
2
r −µ. So .
t
is “halfway between r
t
and µ
t
.” It may happen
that
1
2
(r
t
+µ
t
) is the unique point of 1 with this property (in which case we are done). this
happens, for instance, if the norm in 1 is strictly subadditive, meaning
βr
t
,= αµ
t
=⇒ r
t
+µ
t
 < r
t
 +µ
t
 .
In general, however, there may be a number of points “halfway between r
t
and µ
t
.”
So, let
¹
1
= ¦n ∈ A : r −n = µ −n =
1
2
r −µ¦.
and
¹
t
1
= ¦n
t
∈ 1 : r
t
−n
t
 = µ
t
−n
t
 =
1
2
r
t
−µ
t
¦.
Since ` is an isometry, we have ¹
t
1
= `(¹
1
). Let d
1
denote the diameter of ¹
1
,
d
1
= sup
u,v∈A
1
n −· .
This is also the diameter of ¹
t
1
. Now, let
¹
2
=
_
n ∈ ¹
1
: · ∈ ¹
1
=⇒ n −· ≤
1
2
d
1
_
.
the set of “centers of ¹
1
.” Note that . ∈ ¹
2
since if n ∈ ¹
1
then 2. −n ∈ ¹
1
:
r −(2. −n) = n −µ = r −n = µ −(2. −n) .
Similarly, let
¹
t
2
=
_
n
t
∈ ¹
t
1
: ·
t
∈ ¹
t
1
=⇒ n
t
−·
t
 ≤
1
2
d
1
_
.
Again, since ` is an isometry we have ¹
t
2
= `(¹
2
). In a similar way, deﬁne decreasing
sequences of sets, ¹
j
and ¹
t
j
, inductively by
¹
j
= ¦n ∈ ¹
j−1
: · ∈ ¹
j−1
=⇒ n −· ≤
1
2
diam(¹
j−1
)¦.
and
¹
t
j
= ¦n
t
∈ ¹
t
j−1
: ·
t
∈ ¹
t
j−1
=⇒ n
t
−·
t
 ≤
1
2
diam(¹
t
j−1
)¦.
Again `(¹
j
) = ¹
t
j
and . ∈ ¹
j
since ¹
j−1
is invariant under inversion around .: n ∈
¹
j−1
=⇒ 2. −n ∈ ¹
j−1
. Since diam(¹
j
) ≤ 2
1−j
d
1
we conclude that
∞
j=1
= ¦.¦. and
∞
j=1
¹
t
j
=
_
1
2
(r
t
+µ
t
)
_
.
Since .
t
∈ ¹
t
j
for all ,, () follows.
Homework I
Homework I
Due: February 15, 2008
(1) (Ex. 1, Ch.1) Let A be a linear space. Let ¦o
α
: α ∈ Ω¦ be a collection of
subspaces. (Here Ω is some index set, not necessarily ﬁnite or countable.) Prove
that
(a) The sum
α
o
α
= ¦r
1
+ + r
n
: r
j
∈ o
α
j
. α
j
∈ Ω. , = 1. . . . . :. : ∈ N¦
is a subspace.
(b) The intersection ∩
α∈Ω
o
α
is a subspace.
(c) The union ∪
α
o
α
is a subspace provided Ω is a totally ordered set and α ≤ β
=⇒ o
α
⊂ o
β
.
(2) (Ex. 2, Ch. 4) The Cesaro means of a bounded sequence c
n
are the numbers
`
n
(c) =
1
:
n
j=1
c
j
.
A bounded sequence c is called Cesaro summable if lim
n→∞
`
n
(c) exists.
(a) Show that any convergent sequence c is Cesaro summable and
lim
n→∞
`
n
(c) = lim
n→∞
c
n
.
(b) Give an explicit example of a bounded Cesaro summable sequence that is not
convergent.
(c) Show there is a Banach limit LIM such that for any bounded Cesaro summable
sequence LIM
n→∞
c
n
= lim
n→∞
`
N
(c). (A Banach limit is a linear functional
on the space /
∞
of bounded sequences which agrees with the limit on convergent
sequence, is invariant under shifts, and is bounded above and below by liminf
and limsup.)
(3) (Ex. 2, Ch.5) Let A be a normed linear space and 1 a subspace of A. Prove that
the closure of 1 is a linear subspace of A.
(4) (Ex. 3, Ch.5) Show that if A is a Banach space and 1 is a closed subspace of A,
then the quotient space A,1 is complete.
(5) Let ` : A → 1 be a continuous map between normed spacse A and 1 such that
`(0) = 0 and `
_
1
2
(r +µ)
_
=
1
2
`(r) +
1
2
`(µ). ∀r. µ ∈ A.
Show that ` is linear.
(6) Let / ∈ c
0
be a bounded linear functional on c
0
. Prove that there is a sequence
b ∈ /
1
such that
/(a) =
∞
j=1
/
j
c
j
.
Conclude that c
0
is isometrically isomorphic to /
1
.
(7) Let / ∈ /
1
be a bounded linear functional on /
1
. Prove that there is a sequence
b ∈ /
∞
such that
/(a) =
∞
j=1
/
j
c
j
.
Conclude that /
1
is isometrically isomorphic to /
∞
.
(8) Show that /
p
is isometrically isomorphic to /
q
, 1 < j < ∞ and 1,j + 1,¡ = 1.
(9) Let / = ¦a ∈ /
p
: c
2n
= 0. ∀:¦ ⊂ /
p
. Show that / is a closed subspace and that
/
p
,/ is isometrically isomorphic to /
p
.
(10) Let A
1
. A
2
be Banach spaces, with norms 
1
. 
2
. Deﬁne the direct sum
A
1
⊕ A
2
= ¦(r. µ) : r ∈ A
1
and µ ∈ A
2
¦.
with coordinatewise addition and scalar multiplication. For each j ∈ [1. ∞] deﬁne
the norm
(r. µ)
p
=
_
(r
p
1
+µ
p
2
)
1/p
1 ≤ j < ∞.
max¦r
1
. µ
2
¦ j = ∞.
Show that A
1
⊕A
2
is a Banach space under 
p
and all these norms are equivalent.
(11) Consider A = 1
1
(R) + 1
2
(R) = ¦p + / : p ∈ 1
1
(R) and / ∈ 1
2
(R)¦. Given
j ∈ [1. ∞), deﬁne
)
inf
p
= inf¦(p
p
L
1
+/
p
L
2
)
1/p
: ) = p +/. p ∈ 1
1
and / ∈ 1
2
¦.
Show that )
inf
p
is a norm and that all of these norms are equivalent. (Challenge:
Is A a Banach space?)
(12) Consider 1 = 1
1
(R) ∩1
2
(R), which is a subspace of A from the previous problem.
(a) For each j ∈ [1. ∞), deﬁne
)
+
p
=
1
2
_
)
p
L
1
+)
p
L
2
_
1/p
.
for ) ∈ 1 . Show that 
+
p
is a norm and all of these norms are equivalent.
(b) Show that 1 is a Banach space.
(c) Show that 1 is dense in A under any of the norms 
inf
p
.
(d) Show that )
inf
p
≤ )
+
p
. Is there a constant c such that )
+
p
≤ c )
inf
p
?
Part 3
Hilbert Spaces and Applications
LECTURE 8
Scalar Products and Hilbert Spaces
Reading: ¸6.1 and ¸6.2 of Lax.
Definition 8.1. A scalar product on a linear space A over R is a real valued function
¸. ) : A A →R with the following properties
(1) Bilinearity: r → ¸r. µ) and µ → ¸r. µ) are linear functions.
(2) Symmetry: ¸r. µ) = ¸µ. r).
(3) Positivity: ¸r. r) 0 if r ,= 0. (Note that ¸0. 0) = 0 by bilinearity.
A (complex) scalar product on a linear space A over C is a complex valued function
¸. ) : A A →C with the properties
(1) Sesquilinearity: r → ¸r. µ) is linear and µ → ¸r. µ) is skewlinear,
¸r. µ +cµ
t
) = ¸r. µ) +c
¸r. µ
t
) .
(2) Skew symmetry: ¸r. µ) = ¸µ. r)
.
(3) Positivity: ¸r. r) 0 for r ,= 0.
Given a (real or complex) scalar product, the associated norm is
r =
_
¸r. r).
Remark. A complex linear space is also a real linear space, and associated to any
complex inner product is a real inner product:
(r. µ) = Re ¸r. µ) .
Note that the associated norms are the same, so the metric space structure is the same
whether or not we consider the space as real or complex. Note that
(ir. µ) = (r. iµ). ()
and the real and complex inner products are related by
¸r. µ) = (r. µ) −i(ir. µ) = (r. µ) + i(r. iµ). ()
Conversely, given any real inner product on a complex linear space which satisﬁes (), ()
gives a complex inner product.
We have not shown that the deﬁnition r =
_
¸r. r) actually gives a norm. Homogeneity
and positivity are clear. To verify subadditivity we need the following important Theorem.
Theorem 8.1 (CauchySchwarz). A real or complex scalar product satisﬁes
[¸r. µ)[ ≤ r µ .
with equality only if cr = /µ.
81
82 8. SCALAR PRODUCTS AND HILBERT SPACES
Remark. A corollary is that
r = max
u=1
[ ¸r. n) [.
from which follows subadditivity
r +µ ≤ r +µ .
Proof. It suﬃces to consider the real case, since given r. µ we can always ﬁnd θ so that
¸
e
iθ
r. µ
_
= e
iθ
¸r. µ) is real. Also, we may assume µ ,= 0.
So let ¸. ) be a real inner product and t ∈ R. Then
r +tµ
2
= r
2
+ 2t ¸r. µ) +t
2
µ
2
.
Minimizing the r.h.s. over t we ﬁnd that,
t
min
= −
¸r. µ)
µ
2
.
and
0 ≤ r
2
−
¸r. µ)
2
µ
2
.
The CauchySchwarz inequality follows.
Another important, related result, is the parallelogram identity
r +µ
2
+r −µ
2
= 2 r
2
+ 2 µ
2
.
von Neumann has shown that any norm which satisﬁes the parallelogram law comes from
an inner product.
Definition 8.2. A linear space with a scalar product that is complete in the induced
norm is a Hilbert space.
Any scalar product space can be completed in the norm. It follows from the Schwarz
inequality that the scalar product is cts. in each of its factors and extends uniquely to the
completion, which is thus a Hilbert space.
Examples:
(1) /
2
is a Hilbert space with the inner product
¸c. /) =
j
c
j
/
j
.
which is ﬁnite by the H¨older inequality.
(2) C[0. 1] is an inner product space with respect to the inner product
¸r. µ) =
_
1
0
r(t)µ(t)
dt.
It is not complete. The completion is known as 1
2
[0. 1] and can be associated with
the set of equivalence classes of Lebesgue square integrable functions.
Remark. There is no standard as to which factor of the inner product is skewlinear.
In the physics literature, it is usually the ﬁrst factor; in math it is usually the second.
8. SCALAR PRODUCTS AND HILBERT SPACES 83
Definition 8.3. Two vectors in an inner product space are orthogonal if
¸r. µ) = 0.
The orthogonal complement of a set : is
o
⊥
= ¦· : ¸·. µ) = 0 ∀µ ∈ 1 ¦.
Lemma 8.2. Given any set o in a Hilbert space, o
⊥
is a closed subspace.
Proof. That o
⊥
is a subspace is clear. That it is closed follows, from continuity of the
inner product in each factor, since if ·
n
→ ·, ·
n
∈ o
⊥
, then
¸·. µ) = lim
n
¸·
n
. µ) = 0 for µ ∈ o.
Theorem 8.3. Let H be a Hilbert space, 1 a closed subspace of H and 1
⊥
the orthogonal
complement of 1 . Then
(1) Any vector r ∈ H can be written uniquely as a linear combination
r = µ +·. µ ∈ 1 and · ∈ 1
⊥
.
(2) (1
⊥
)
⊥
= 1.
To prove this theorem, we need
Lemma 8.4. Given a nonempty closed, convex subset 1 of a Hilbert space, and a point
r ∈ H, there is a unique point µ in 1 that is closer to r than any other point of1.
Proof. This follows if we show that H is uniformly convex, by the Theorem of Clarkson
from lecture 5. Let r. µ be unit vectors. It follows from the parallelogram law that
_
_
_
_
1
2
r +µ
_
_
_
_
2
= 1 −
1
4
r −µ
2
.
so
_
_
_
_
1
2
r +µ
_
_
_
_
≤ 1 −
_
1 −
_
1 −
1
4
r −µ
2
_
. ¸¸ .
x−y
.
Proof of Theorem. According to the Lemma there is a unique point µ ∈ 1 closest
to a given point r ∈ H. Let · = r −µ. We claim that ¸·. µ
t
) = 0 for any µ
t
∈ 1 . Indeed, we
must have
·
2
≤ · +tµ
t

2
= ·
2
+ 2t Re ¸·. µ
t
) +t
2
µ
t

2
for any t. In other words the function
0 ≤ 2t Re ¸·. µ
t
) +t
2
µ
t

2
for all t.
which can occur only if Re ¸·. µ
t
) = 0. Since this holds for all µ
t
∈ 1 we get ¸·. µ
t
) = 0 by
complex linearity.
Thus the decomposition r = µ +· is possible. Is it unique? Suppose r = µ +· = µ
t
+·
t
.
Then µ −µ
t
= · −·
t
∈ 1 ∩ 1
⊥
. But . ∈ 1 ∩ 1
⊥
=⇒ ¸.. .) = 0 so . = 0.
Part (3) is left as a simple exercise.
LECTURE 9
RieszFrechet and LaxMilgram Theorems
Reading: ¸6.3 of Lax.
We have already seen that for ﬁxed µ ∈ H, a Hilbert space, the map /
y
(r) = ¸r. µ) is a
bounded linear functional — boundedness follows from CauchySchwarz. In fancy language
µ → /
y
embeds H into H
, the dual of H. In fact, since
/
y
 = sup
x
[¸r. µ)[
r
= µ .
again by CauchySchwarz, this map is an isometry onto it’s range. In a real Hilbert space,
this is a linear map; in a complex Hilbert space, it is skewlinear:
/
y+αy
= /
y
+α
/
y
.
The question now comes up whether we get every linear functional in H
this way? The
answer turns out to be “yes.”
Theorem 9.1 (RieszFrechet). Let /(r) be a bounded linear functional on a Hilbert space
H. Then there is a unique µ ∈ H such that
/(r) = ¸r. µ) .
Before turning to the proof, let us state several basic facts, whose proof is left as an
exercise:
Lemma 9.2.
(1) Let A be a linear space and / a nonzero linear functional on A. Then the null
space of / is a linear subspace of codimension 1. That is, if1 = ¦µ : /(µ) = 0¦
then there exists r
0
,∈ 1 and any vector r ∈ A may be written uniquely as
r = αr
0
+µ. α ∈ 1 and µ ∈ 1.
(2) If two linear functionals /, : share the same null space, they are constant multiples
of each other: / = c:.
(3) If A is a Banach space and / is bounded, then the nullspace of / is closed.
Proof. If / = 0 then µ = 0 will do, and this is the unique such point µ.
If / ,= 0, then it has a null space 1 , which by the lemma is a closed subspace of co
dimension 1. The orthogonal complement 1
⊥
must be one dimensional. Let ˆ µ be a unit
vector in 1
⊥
— it is unique up to a scalar multiple. Then :(r) = ¸r. ˆ µ) is a linear functional,
with nullspace 1 . Thus / = α: and we may take µ = α
ˆ µ.
To see that µ is unique, note that if ¸r. µ) = ¸r. µ
t
) for all r then µ −µ
t
 = 0, so
µ = µ
t
.
In applications, one is often given not a linear functional, but a quadratic form:
91
92 9. RIESZFRECHET AND LAXMILGRAM THEOREMS
Definition 9.1. Let H be a Hilbert space over R. A bilinear form on H is a function
1 : H H →R such that
r → 1(r. µ) and µ → 1(r. µ)
are linear maps. A skewlinear form on a Hilbert space H over C is a map 1 : H H →C
such that
r → 1(r. µ) is linear, and µ → 1(r. µ) is skewlinear.
A quadratic form refers to a bilinear form or a skewlinear form depending on whether the
ﬁeld of scalars is R or C. A quadratic form 1 on H is bounded if there is a constant c 0
such that
[1(r. µ)[ ≤ c r µ .
and is bounded from below if there is a constant / 0 such that
[1(µ. µ)[ ≥ / µ
2
.
Theorem 9.3 (LaxMilgram). Let H be a Hilbert space, over R or C, and let 1 be a
bounded quadratic form on H that is bounded from below. Then ever bounded linear functional
/ ∈ H
may be written
/(r) = 1(r. µ). for unique µ ∈ H.
Proof. For ﬁxed µ, r → 1(r. µ) is a bounded linear functional. By RieszFrechet there
exists . : H → H such that
1(r. µ) = ¸r. .(µ)) .
It is easy to see that the map µ → .(µ) is linear. Thus the range of .,
ran . = ¦.(µ) : µ ∈ H¦ .
is a linear subspace of H.
Let us prove that ran . is a closed subspace. Here we need the fact that 1 is bounded
from below. Indeed,
1(µ. µ) = ¸µ. .(µ)) .
so
/ µ
2
≤ µ .(µ) =⇒ / µ ≤ .(µ) .
If µ
n
is any sequence then
µ
n
−µ
m
 ≤ /
−1
.(µ
n
) −.(µ
m
) .
Thus .(µ
n
) → .
0
=⇒ µ
n
Cauchy =⇒ µ
n
→ µ
0
, and it is easy to see we must have
.
0
= .(µ
0
). Thus .
0
∈ ran ., so ran . is closed.
Now we show that ran . = H. Since ran . is closed it suﬃces to show ran .
⊥
= ¦0¦. Let
r ⊥ ran .. It follows that
1(r. µ) = ¸r. .(µ)) = 0 ∀µ ∈ H.
Thus 1(r. r) = 0 and so r = 0 since r
2
≤ /
−1
[1(r. r)[ .
Since ran . = H we see by RieszFrechet that any linear functional / may be written
/(r) = ¸r. .(µ)) = 1(r. µ) for some µ. Uniqueness of µ follows as above, since if 1(r. µ) =
1(r. µ
t
) for all r we conclude that µ −µ
t
 = 0 since 1 is bounded from below.
APPLICATION: RADONNIKODYM THEOREM 93
Application: RadonNikodym Theorem
Theorem 9.4 (RadonNikodym). Let `, Σ be a measurable space on which we have
deﬁned two ﬁnite nonnegative measures j and ν. If j(¹) = 0 =⇒ ν(¹) = 0 for all ¹ ∈ Σ
then there exists a Σ−measurable function / : ` → [0. ∞) such that
ν(¹) =
_
A
/dj. ¹ ∈ Σ.
Proof. Consider the measure j +ν and the Hilbert space 1
2
(j +ν),
)
2
=
_
M
)
2
d(j +ν).
Deﬁne a linear functional
/()) =
_
M
)dj.
Since
[/())[
2
≤ j(`)
_
M
)
2
dj ≤ j(`) )
2
.
it follows that / is bounded (on 1
2
(j +ν)). Thus ∃p ∈ 1
2
(j +ν) such that
_
M
)dj =
_
M
)pd(j +ν).
Rewrite this as,
_
M
)(1 −p)dj =
_
M
)dν. ()
Let 1 = ¦r ∈ ` : p(r) ≤ 0¦ and plug ) = χ
F
into (). Then
j(1) ≤
_
F
(1 −p)dj =
_
F
pdν ≤ 0.
Thus j(1) = 0. Likewise let G = ¦r ∈ ` : p(r) 1¦ and plug ) = χ
G
into (). If
j(G) 0 then
0
_
G
(1 −p)dj =
_
G
pdν ν(G).
which is a contradiction. Thus j(G) = 0. Hence, after modifying p on the j null set 1 ∪ G
we may assume that () holds with
0 < p(r) ≤ 1.
But then given 1 ∈ Σ, plugging ) = p
−1
χ
E
into () we get
ν(1) =
_
M
)pdν =
_
M
χ
E
1 −p
p
dj =
_
E
1 −p
p
dj.
Thus take / = (1 −p),p.
LECTURE 10
Geometry of a Hilbert space and GramSchmidt process
Reading: ¸6.4 of Lax.
Recall that the linear span of a set o in a linear space A is the collection of ﬁnite linear
combinations of elements of o:
span o =
_
n
j=1
α
j
r
j
: r
j
∈ o. α
j
∈ 1. , = 1. . . . . :. : ∈ N
_
.
This is also the smallest subspace containing o:
span o = ∩¦1 : 1 ⊂ A is a subspace and o ⊂ 1 ¦.
If A is a Banach space, it is natural to look at the smallest closed subspace containing o:
span o = ∩¦1 : 1 ⊂ A is a closed subspace and o ⊂ 1 ¦.
Proposition 10.1. Let A be a Banach space. Then span o = span o.
The proof is left as an exercise.
In a Hilbert space we have a geometric characterization of span o:
Theorem 10.2. Let o ⊂ H be any subset of a Hilbert space H. Then
span o =
_
o
⊥
_
⊥
.
That is, µ ∈ span o if and only if µ is perpendicular to everything that is perpendicular to o:
¸µ. .) = 0 for all . such that ¸r. .) = 0 for all r ∈ o.
Proof. Recall that a closed subspace 1 satisﬁes (1
⊥
)
⊥
= 1 . Thus it suﬃces to show
(span o)
⊥
= o
⊥
. Since o ⊂ span o we clearly have (span o)
⊥
. On the other hand, if . ∈ o
⊥
.
Thus . is perpendicular to span o and by continuity of the scalar product . ⊥ span o =
span o. Thus o
⊥
⊂ (span o)
⊥
.
Definition 10.1. A collection of vectors o in an inner product space H is called
orthonormal if
¸r. µ) =
_
1 r = µ ∈ o
0 r ,= µ. r. µ ∈ o.
An orthonormal collection o is called an orthonormal basis if span o = H.
Lemma 10.3. Let o be an orthonormal set of vectors in a Hilbert space H. Then the
span o consists of all vectors of the form
r =
∞
j=1
α
j
r
j
. r
j
∈ o. , = 1. . . . . ∞. ()
101
102 10. GEOMETRY OF A HILBERT SPACE AND GRAMSCHMIDT PROCESS
where the α
j
are square summable:
∞
j=1
[α
j
[
2
< ∞.
The sum converges in the Hilbert space:
_
_
_
_
_
r −
n
j=1
α
j
r
j
_
_
_
_
_
→ 0.
and
r
2
=
∞
j=1
[α
j
[
2
.
Furthermore, the sum may be written
r =
y∈S
¸r. µ) µ.
In particular, ¸r. µ) , = 0 for only countably many elements µ ∈ o.
Remark. Most orthonormal sets encountered in practice are countable, so we would
tend to write o = ¦r
1
. . . . . ¦ and
r =
∞
j=1
¸r. r
j
) r
j
.
However, the lemma holds even for uncountable orthonormal sets.
Proof. It is clear that all vectors of the form () are in span o = span o. Furthermore
vectors of this form make up a subspace, which is easily seen to be closed. (Exercise: show
that this subspace is closed. This rests on the fact that a subset of a complete metric space
is closed iﬀ it is sequentially complete.) By deﬁnition span o is contained in this subspace.
Thus the two subspaces are equal.
The remaining formulae are easy consequences of the form ().
Theorem 10.4. Every Hilbert space contains an orthonormal basis.
Proof. We use Zorn’s Lemma. Consider the collection of all orthonormal sets, with
o ≤ 1 iﬀ o ⊂ 1. This collection is nonempty since any unit vector makes up a one element
orthonormal set.
A totally ordered collection has an upper bound — the union of all sets in the collection.
Thus there is a maximal orthonormal set. Call it o
max
.
Suppose span o
max
A. Then, span o
⊥
max
is a nontrivial closed subspace. Let µ ∈
span o
⊥
max
be a unit vector. So o
max
∪ ¦µ¦ is an orthonormal set contradicting the fact that
o
max
is maximal.
Corollary 10.5 (Bessel’s inequality). Let o be any orthonormal set in a Hilbert space
H (not necessarily a basis), then
y∈S
[¸r. µ)[
2
≤ r
2
for all r ∈ H.
Equality holds for every r if and only if o is a basis.
10. GEOMETRY OF A HILBERT SPACE AND GRAMSCHMIDT PROCESS 103
If the Hilbert space H is separable — it contains a countable dense set — then any
orthonormal basis is countable. In this case we can avoid Zorn’s Lemma. The foundation of
this is the GramSchmidt process.
Theorem 10.6 (GramSchmidt Process). Let µ
j
be a sequence of vectors in a Hilbert
space. Then there is an orthonormal sequence r
j
such that
span¦µ
1
. . . . . µ
n
¦ ⊂ span¦r
1
. . . . . r
n
¦.
Proof. The proof is constructive. We may assume, without loss, that ¦µ
1
. . . . . µ
n
¦ is
linearly independent for each :. (Otherwise throw out vectors µ
n
until this is the case.)
Then µ
1
,= 0 so let
r
1
=
µ
1
µ
1

.
Clearly span¦r
1
¦ = span¦µ
1
¦.
Now, suppose we are given r
1
. . . . . r
n−1
such that
span¦r
1
. . . . . r
n−1
¦ = span¦µ
1
. . . . . µ
n−1
¦.
Let
r
n
=
µ
n
−
n−1
j=1
¸µ
n
. r
n−1
) r
n−1
_
_
_µ
n
−
n−1
j=1
¸µ
n
. r
n−1
) r
n−1
_
_
_
.
This is OK since µ
n
,=
n
j=1
¸µ
n
. r
n−1
) ∈ span¦µ
1
. . . . . µ
n−1
¦. Clearly
span¦r
1
. . . . . r
n
¦ = span¦µ
1
. . . . . µ
n
¦.
By induction, the result follows.
Corollary 10.7. Let H be a separable Hilbert space. Then H has a countable orthonor
mal basis.
Finally, let us discuss the isometries of Hilbert spaces.
Theorem 10.8. Let H and H
t
be Hilbert spaces. Given an orthonormal basis o for H,
an orthonormal set o
t
⊂ H
t
and a onetoone onto map ) : o → o
t
, deﬁne a linear map
H → H
t
via
y∈S
α
y
µ
T
f
−−−→
y∈S
α
y
)(µ).
Then 1 is a linear isometry onto span o
t
⊂ H
t
. Furthermore, any isometry of H with a
subspace of H
t
is of this form.
Corollary 10.9. Two Hilbert spaces are isomorphic iﬀ their orthonormal bases have
equal cardinality. In particular, every Hilbert space is isomorphic with /
2
(o) for some set o.
Any separable, inﬁnite dimensional Hilbert space is isomorphic to /
2
.
Remark. For an arbitrary set o, /
2
(o) is deﬁned to be the set of functions ) : o → R
or C such that
y∈S
[)(µ)[
2
< ∞.
Note that ) ∈ /
2
(o) =⇒ ¦µ : )(µ) ,= 0¦ is countable.
The proof of these results is left as an exercise.
Part 4
Locally Convex Spaces
LECTURE 11
Locally Convex Spaces and Spaces of Test Functions
Reading: ¸13.13.2 and ¸B.1 of Lax.
A Banach space is one example of a topological vector space (TVS), which is a linear
space A together with a topology on A such that the basic operations of addition and scalar
multiplication are continuous functions.
Definition 11.1. A topological vector space is a linear space A with a Hausdorﬀ topol
ogy such that
(1) (r. µ) → r + µ is a continuous map from A A (with the product topology) into
A.
(2) (/. r) → /r is a continuous map from 1 A (with the product topolgoy, 1 = R or
C) into A.
Remark. Recall that a Hausdorﬀ space is one in which points may be separated by open
sets: given r. µ ∈ A, r ,= µ there are disjoint open sets l. \ , l ∩ \ = ∅ such that r ∈ l
and µ ∈ \ .
Theorem 11.1. Let A be a TVS and let l ⊂ A be open. Then
(1) For any r ∈ A, l −r = ¦µ : µ +r ∈ l¦ is open.
(2) For any scalar / ,= 0, /l = ¦µ : /
−1
µ ∈ l¦ is open
(3) Every point of l is interior: given r ∈ l and µ ∈ A there is c 0 such that for
any scalar t with [t[ < c we have r +tµ ∈ l.
Proof. The set l − r is the inverse image of l under the map µ → µ + r. Thus
(1) follows from continuity of the map µ → µ + r which follows from joint continuity of
(µ. r) → µ +r. (Why?)
Likewise (2) follows from continuity of µ → /
−1
µ.
Since l − r is open it suﬃces to suppose r = 0 ∈ l. For ﬁxed µ ∈ A the map t → tµ
is continuous. (Why?) Thus ¦t : tµ ∈ l¦ is open. Since this set contains t = 0 it must
contain an interval (−c. c) (or an open ball at the origin if the ﬁeld of scalars is C).
The class of TVSs is rather large. However, almost all of the TVSs important to analysis
have the following property:
Definition 11.2. A locally convex space (LCS) is a TVS A such that every open set
containing the origin contains an open convex set containing the origin. That is, there is a
basis at the origin consisting of open convex sets.
Given a LCS, A, we deﬁne the dual A
t
to be the set of all continuous linear functional
on A. A LCS space shares the property of separation of points by linear functionals:
Theorem 11.2. Let A be a LCS and let µ ,= µ
t
be points of A. There is a linear
functional / ∈ A
t
such that
/(µ) ,= /(µ
t
).
111
112 11. LOCALLY CONVEX SPACES AND SPACES OF TEST FUNCTIONS
Proof. Of course, we use the HahnBanach theorem. Speciﬁcally the hyperplane sepa
ration Theorem 2.1.
First, it suﬃces to suppose the ﬁeld of scalars is R, for if we construct a suitable real
linear functional /
r
on a complex LCS we can complexify it
/(r) = /
r
(r) −i/
r
(ir).
Now, without loss we suppose that µ
t
= 0. Since the topology on A is Hausdorﬀ, there is
an open set l ÷ µ
t
with µ ,∈ l. Since A is locally convex, we may suppose l to be convex.
Replacing l with l ∩(−l) we may assume l is symmetric about 0, so r ∈ l =⇒ −r ∈ l.
Since all points of l are interior, Theorem 2.1 asserts the existence of a linear functional /
with 1 = /(µ) and /(r) < 1 for r ∈ l. In fact, the proof shows that
/(r) ≤ j
U
(r) ∀r ∈ A.
where j
U
is the gauge function of l,
j
U
(r) = inf¦t 0 : t
−1
r ∈ l¦.
We need to show that / is continuous. It suﬃces to show /
−1
(c. /) is open for any
c < / ∈ R. Let t ∈ (c. /). Let r
0
be any point with /(r
0
) = t. Then, /
−1
(c − t. / − t) =
/
−1
(c. /) − r
0
÷ 0. (Why?) Thus it suﬃces to suppose c < 0 < / and show that /
−1
(c. /)
contains an open neighborhood at 0. Let t = min¦−c. /¦. The given r ∈ l,
/(tr) ≤ j
U
(tr) = tj
U
(r) < t and /(−tr) ≤ j
U
(−tr) = tj
U
(−r) < t.
Thus tl ⊂ /
−1
(−t. t) ⊂ (c. /).
Converse to this construction is the following idea
Theorem 11.3. Let A be linear space and let 1 be any collection of linear functionals
on A that separates points: for any µ. µ
t
∈ A there is / ∈ 1 such that /(r) ,= /(r
t
). Endow
A with the weakest topology such that all elements of 1 are continuous. Then A is a LCS,
and the dual of A is
A
t
= span 1 = ¦ﬁnite linear combinations of elements of 1¦.
Remark. Recall that a topology on A is a collection  of subsets of A that includes
A and ∅, and is closed under unions and ﬁnite intersections. The intersection of a family of
topologies is also a topology. Thus the weakest topology with property ¹ is the intersection
of all topologies with property ¹.
Proof. Exercise.
What are some examples of LCSs? First oﬀ, any Banach space is locally convex, since
the open balls at the origin are a basis of convex sets. But not every LCS has a norm which
is compatible with the topology. By far the most important examples, though are socalled
spaces of text functions and their duals, the socalled spaces of distributions.
LECTURE 12
Generation of a LCS by seminorms and Fr´echet Spaces
Reading: ¸5.2 of Reed and Simon (or just these notes)
Test functions
The theory of distributions — due to Laurent Schwarz — is based on introducing a
LCS of “test functions” A and it’s dual A
t
, a space of “distributions.” The test functions
are “nice:” we can operate on them arbitrarily with all the various operators of analysis –
diﬀerentiation, integration etc. Using integration, we embed A → A
t
via a map φ → /
φ
:
/
φ
(ψ) =
_
R
d
φ(r)ψ(r)dr.
Thus we think of a (test) function both as a map and as an “averaging” procedure. A key
identity is integration by parts
/
φ
(∂
i
ψ) = −/
∂
i
ψ
(φ).
This suggests that we deﬁne
∂
i
/(ψ) = −/(ψ)
for any distribution / ∈ A
t
. One typically uses a function notation for a distribution, writing
/(ψ) =
_
1(r)ψ(r).
even if the “function” 1 doesn’t exist.
A common, and useful, space of test functions is
C
∞
c
(R
d
) = ¦C
∞
functions on R
d
with compact support.¦.
We wish to topologize this set so that a sequence n
k
converges n if the supports of all n
k
are contained in some ﬁxed compact 1 ⊂ R
d
and if for any choice of multindex α =
(α
1
. . . . . α
d
) ∈ N
d
we have
1
α
n
k
= ∂
α
1
1
∂
α
d
d
n
k
→ 1
α
n uniformly in 1.
What is really going here is this. Deﬁne for each : ≥ 0 and each multiindex α a
seminorm of C
∞
c
(R
d
),
j
n,α
(n) = sup
[x[≤2
n
[1
α
n(r)[ .
(Recall that a seminorm on a linear space A is a map j : A → [0. ∞) which is positive
homogeneous (j(cr) = [c[j(r)) and subadditive (j(r + µ) ≤ j(r) + j(µ)). It is allowed
that j(r) = 0 for r ,= 0. ) It may happen that n
n,α
vanishes even if n ,= 0, however the
collection separates points (Why?):
Definition 12.1. A collection of seminorms o separates points if
j(n) = 0 ∀j ∈ ( =⇒ n = 0.
121
122 12. GENERATION OF A LCS BY SEMINORMS AND FR
´
ECHET SPACES
Now endow A = C
∞
0
(R
d
) with the smallest topology such that A is a TVS and each of
the seminorms j
n,α
is continuous.
Claim. C
∞
c
(R
d
) with this topology is a LCS
Warning: this is not the standard topology on C
∞
c
(R
d
). See below.
Generation of an LCS by seminorms
The claim follows from the following general result.
Definition 12.2. Let A be a linear space and o a family of functions ) : A → `, ` a
topological space (usually ` = R or C). The TVS topology generated by o is the weakest
topology on A such that A is a TVS and all the functions in o are continuous.
Theorem 12.1. Given a linear space A and a collection of seminorms o that separates
points, the TVS topology generated by o makes A is locally convex.
Conversely, given a LCS A and ( a neighborhood base at the origin consisting of convex,
symmetric sets, the LCS topology on A is the TVS topology generated by o = ¦j
U
: l ∈ (¦
are continuous.
Proof. First given an LCS space A let us show that the gauge function j
U
of a convex,
symmetric neighborhood of the origin l is continuous. To begin, note that
j
−1
U
[0. /) = ¦r ∈ A : r ∈ /l¦ = /l
is open for each /. Next consider the sets j
−1
U
(/. ∞). Let r be in this set and let α = j
U
(r).
Consider the open neighborhood \ = r + (α −/)l then for µ = r + (α −/)µ
t
∈ \ we have
j
U
(µ) ≥ j
U
(r) −(α −/)j
U
(µ
t
) /.
So \ ⊂ j
−1
U
(/. ∞) and thus the set is open. Continuity of j
U
follows since the sets [0. /),
(/. ∞) as / ranges over (0. ∞) generate the topology on [0. ∞).
Now, any topology under which every j ∈ o is continuous certainly contains the collection
( = ¦j
−1
[0. /) : / ∈ (0. ∞)¦.
consisting of convex, symmetric sets with the origin as an interior point. Consider the
smallest TVS topology containing this collection. It is easily seen to be locally convex. (We
need the fact that o separates points to get the Hausdorﬀ property.) Since j = j
U
for
l = j
−1
[0. 1) we see from the above argument that all j ∈ o are continuous in this topology.
Thus this is the TVS topology generated by o.
Conversely, let T denote a given LCS topology on A. Thus T is certainly a topology un
der which A is a TVS and all elements of o = ¦j
U
: l a convex, symmetric neighborhood of 0¦
are continuous. To prove it is the weakest such, we must show that any such topology con
tains T . Any l ∈ ( may be written as l = j
−1
U
([0. 1)). Thus any topology under which
A is a TVS and all j
U
are continuous certainly contains ( and all its translates, and thus
T .
METRIZABLE LCSS 123
Metrizable LCSs
Above we claimed that the topology on C
∞
0
could be given in terms of uniform conver
gence of sequences of functions. However, in a general LCS sequential convergence may not
specify the topology — a set may fail to be closed even it contains the limits of all conver
gent sequences of its elements — because there may not be a countable neighborhood base
at the origin. (Don’t worry too much about this.) However, if the origin has a countable
neighborhood base then it turns out that the LCS is actually metrizable, so in particular
sequential convergence speciﬁes the topology.
Theorem 12.2. Let A be an LCS. The following are equivalent
(1) A is metrizable
(2) A has a countable neighborhood basis at the origin ( consisting of convex, symmetric
sets
(3) the topology on A is generated by a countable family of seminorms.
Proof. The equivalence of (2) and (3) is established by associating to convex, symmetric
neighborhoods of the origin the corresponding gauge function and vice versa. The details
are left as an exercise.
To show (1) =⇒ (2), suppose A is metrizable. Then A has a countable neighborhood
basis (this is a property of metric spaces), and in particular a countable neighborhood basis
at the origin. Since A is a LCS we may ﬁnd a convex, symmetric open set contained in each
of the basis sets, thus obtaining a convex, symmetric, countable neighborhood basis at the
origin.
To show (2) =⇒ (1), suppose A has a countable neighborhood basis as indicated, and
let T denote it’s topology. Since ( is countable we may assume, without loss, that it is
a decreasing sequence ( = ¦l
1
⊃ l
2
⊃ ¦. (Order the elements of ( and take ﬁnite
intersections l
k
→ l
1
∩ ∩ l
k
.) Let j
j
(r) = j
U
j
(r), so j
k
(r) ≥ j
j
(r) if / ≥ ,. Deﬁne a
metric
d(r. µ) =
∞
j=1
2
−j
j
j
(r −µ)
1 +j
j
(r −µ)
.
and the metric topology T
d
. (Recognize this? Why is this a metric? Note that the collection
¦j
U
: l ∈ (¦ separates points since ( is a basis.)
Clearly,
¦r : d(r. 0) < 2
−j−1
¦ ⊂ l
j
.
Thus T ⊂ T
d
(since any T open set contains a translate of some dball centered at each of
its points). On the other hand, if r ∈ tl
k
d(r. 0) <
k
j=1
2
−j
t
1 +t
+
∞
j=k+1
2
−j
≤
1
2
t
1 +t
+ 2
−k−1
.
Thus
2
−k
l
k
⊂ ¦r : d(r. 0) < 2
−k
¦.
which shows that T
d
⊂ T . Thus T = T
d
and A is metrizable.
Definition 12.3. A Fr´echet space is a complete, metrizable, locally convex linear space.
Remark. Recall that a metric space is complete if every Cauchy sequence converges.
LECTURE 13
The dual of an LCS
It turns out that C
∞
c
(R
d
), although a metric space with the metric suggested above,
is not complete. The completion is the space of C
∞
0
functions that together with all their
derivatives vanish at ∞, which could be topologized with the seminorms
j
α
()) = sup
x
[1
α
)(r)[ .
The correct way to think of C
∞
c
(R
d
) is as an “inductive limit,” which is as the union
∪
n
C
∞
([−2
n
. 2
n
]). There is an inductive limit topology on this space, making it a com
plete, but nonmetrizable, LCS. Each of the spaces C
∞
([−2
n
. 2
n
]) is a Fr´echet space, but
clearly the inductive limit topology is diﬀerent that the (metrizable) LCS topology generated
above. It is still true that a sequence converges according to the criteria given above, but
this doesn’t give the whole picture as the space isn’t separable! (See Reed and Simon chapter
V.)
A remedy for this, that is often suﬃcient, is to work with the following somewhat larger
space
Definition 13.1. The Schwarz space o(R
d
) consists of every C
∞
function ) on R
d
such
that
j
α,β
()) = sup
x∈R
d
[r
α
1
β
)(r)[ < ∞.
for every pair of multiindices α. β ∈ N
d
.
Remark. • r
α
= r
α
1
1
r
α
d
d
.
• Note that o(R
d
) ⊂ 1
p
(R
d
) for every j.
• Likewise ) ∈ o(R
d
) =⇒ 1
α
) ∈ 1
p
(R
d
) for every j.
Corollary 13.1. The space o(R
d
) with the topology generated by the seminorms j
α,β
is a Fr´echet space.
The dual space of a LCS
It is useful to use the inner product notation to denote the pairing between elements of
A and linear functionals
/(r) = ¸r. /) .
(Note that this inner product is linear in both factors even if we are dealing with complex
spaces.)
Given a linear space A and a linear space of linear functionals 1 on A that separates
points we have seen that there is a LCS topology on A such that 1 is the dual of A. This
topology is called the 1weak topology on A and is denoted σ(A. 1).
131
132 13. THE DUAL OF AN LCS
On the other hand, given an LCS, we can think of A as a collection of linear functionals
on A
, associating to r ∈ A the map
/ → ¸r. /) .
The Aweak topology on A
, σ(A
. A), is also called the weak
toplogy. It is generated by
the family of seminorms
j
x
(/) = [¸r. /)[ .
Theorem 13.2. If A is an LCS then (A
. σ(A
. A))
= A.
Remark. Recall that σ(A. A
) is the given LCS topology on A so we also have (Aσ(A. A
))
=
A
. Thus for any LCS (A
)
= A, provided we topologize A
with the weak
toplogy. If A
is a Banach space we also have a norm topology on A
, which is substantially stronger than
the weak
topology and with respect to which this identity may not hold. For instance,
(1) As Banach spaces c
0
= /
1
and /
1
= /
∞
and /
∞
, which includes Banach limits, is
strictly larger than /
1
.
(2) As LCS spaces c
0
= /
1
and (/
1
. σ(/
1
. c
0
)
= c
0
, etc.
The moral of the story is topology matters.
The following theorem is useful for determining if a linear functional is continuous.
Theorem 13.3. Let A be a LCS generated by a family of seminorms o. Then a linear
functional / ∈ A
t
if and only if there is a constant C 0 and a ﬁnite collection j
1
. . . . . j
n
∈ o
such that
[/(r)[ ≤ C
n
j=1
j
j
(r) ∀r ∈ A.
Proof. (⇒) If / is continuous then l = /
−1
(−1. 1) an open, convex, symmetric neigh
borhood of the origin in A. By virtue of the fact that o generates the topology on A, since
l is open we have
n
j=1
¦r : j
j
(r) < ε¦ ⊂ l
for some ﬁnite collection j
1
. . . . . j
n
. Thus,
\ = ¦r :
n
j=1
j
j
(r) < ε¦ ⊂ l.
Now, given r let t =
2
ε
n
j=1
j
j
(r). Then
n
j=1
j
j
(t
−1
r) =
ε
2
.
so t
−1
r ∈ \ ⊂ l. Thus
¸
¸
/(t
−1
r)
¸
¸
< 1 =
n
j=1
j
j
(r)
n
j=1
j
j
(r)
=
2
ε
n
j=1
j
j
(t
−1
r).
Multiplying through by t we get the desired bound with C = 2,ε.
THE DUAL SPACE OF A LCS 133
(⇐) Since / is linear, it suﬃces to show that / is continuous at 0. That is we must show
that /
−1
(−ε. ε) contains an open set containing the origin for each ε 0. But clearly
¦r : C
n
j=1
j
j
(r) < ε¦ ⊂ /
−1
(−ε. ε).
LECTURE 14
Spaces of distributions
Tempered Distributions
The dual of o(R
d
) denoted o
(R
d
) is the space of tempered distributions. Here are some
examples:
(1) o(R
d
) ⊂ o
(R
d
) where we associate to a function φ ∈ o(R
d
) the distribution
ψ → ¸ψ. φ) =
_
R
d
ψ(r)φ(r)dr.
(2) More generally, a function 1 ∈ 1
1
loc
(R
d
) that is polynomially bounded in the sense
that j(r)
−1
1(r) ∈ 1
1
(R
d
) for some positive polynomial j 0 may be considered as
a tempered distribution
ψ → ¸j:i. 1) =
_
R
d
ψ(r)1(r)dr.
(3) Similarly, any polynomially bounded Borel measure j, with
_
j(r)
−1
d[j[(r) < ∞.
is a tempered distribution:
¸ψ. j) =
_
R
d
ψ(r)dj(r).
To go further we need the following generalization of Theorem 13.3 from the last lecture:
Theorem 14.1. Let A. 1 be a LCSs generated by a families of seminorms o. T respec
tively. Then a linear map 1 : A → 1 is continuous if and only if for any seminorm ¡ ∈ o
there is a constant C 0 and a ﬁnite collection j
1
. . . . . j
n
∈ o such that
¡(1r) ≤ C
n
j=1
j
j
(r) ∀r ∈ A.
Corollary 14.2. For each , = 1. . . . . d, diﬀerentiation ∂
j
is a continuous map from
o → o.
Now we deﬁne ∂
j
: o
→ o
. Note that
¸∂
j
ψ. 1) = −¸ψ. ∂
j
1) .
whenever 1 is a C
1
function of polynomial growth. Thus deﬁne for arbitrary / ∈ o
:
¸ψ. ∂
j
/) = −¸∂
j
. /) .
Proposition 14.3. So deﬁned, ∂
j
: o
→ o
is a continuous map.
Proof. Exercise
141
142 14. SPACES OF DISTRIBUTIONS
Thus we have the following generalization of the above examples:
• Let α ∈ ^
d
be a multiindex and let 1 be a polynomial 1
1
bounded function. Then
1
α
1 is a tempered distribution:
¸ψ. 1
α
1) = (−1)
α
¸1
α
ψ. 1) = = (−1)
α
_
R
d
1
α
ψ(r)1(r)dr.
Theorem 14.4 (Structure Theorem for Tempered Distributions). Let / ∈ o(R
d
) be a
tempered distribution. Then there is a polynomially bounded continuous function p and a
multiindex α ∈ N
d
such that / = 1
α
p.
For the proof see Reed and Simon, Ch. V.
For example, we now understand in a precise sense the identity
δ(r) =
d
2
dr
2
[r[.
More generally, in d = 2, we have the identity
δ(r) =
1
2π
∆ln [r[.
This can be veriﬁed as follows. Note that ∆ln [r[ = 0 if r ,= 0. Thus
¸φ. ∆ln [r[) = 0 if φ(r) = 0 for [r[ < c.
Let / be a compactly supported function that is 1 in the neighborhood of the origin, then
¸φ −/φ. ∆ln [r[) = 0. Thus it suﬃces to suppose φ ∈ C
∞
c
(R
d
). For such φ
¸φ. ∆ln [r[) =
_
[x[<M
∆φ(r) ln [r[dr = lim
ε↓0
_
ε<[x[<M
∆φ(r) ln [r[dr.
since ln [r[ is locally integrable. Now integrate by parts:
_
ε<[x[<M
∆φ(r) ln [r[dr = −
_
ε<[x[<M
r
[r[
2
∇φdr −ln ε
_
[x[=ε
r
[r[
∇φ(r)dσ(r).
where in the ﬁrst integral
x
[x[
2
is ∇ln [r[ and in the second integral −
x
[x[
is an outward facing
normal on ¦[r[ = ε¦ and dσ(r) is the length measure on the circle. Continuing, we ﬁnd that
_
ε<[x[<M
∆φ(r) ln [r[dr = =
1
ε
_
[x[=ε
φ(r)dσ(r) −C(ε ln ε) → πφ(0).
Remark. In a similar fashion, in dimension d ≥ 3,
δ(r) = −∆c
d
[r[
2−d
.
with c
−1
d
= (d −2) area of the unit sphere ¦[r[ = 1 : r ∈ R
d
¦.
Other spaces of distributions
Consider the scale of spaces
C
∞
c
(R
d
) ⊂ o(R
d
) ⊂ C
∞
0
(R
d
) ⊂ C
∞
(R
d
).
Each has a natural LCS topology such that it is a complete space. The middle two are
Fr´echet spaces and we have discussed their topologies already.
OTHER SPACES OF DISTRIBUTIONS 143
On C
∞
c
(R
d
) we put the inductive limit topology as follows. For each :, let Ω
n
=
(−2
n
. 2
n
)
d
. The space A
n
= C
∞
0
(Ω
n
) is a Fr´echet space generated by the seminorms
j
n,α
(n) = sup
x∈ωn
[1
α
n(r)[ . ()
Note that A
n
⊂ A
n+1
, the embedding is continuous, and A
n
is a closed subspace of A
n+1
.
Put of C
∞
c
(R
d
) = A = ∪
n
A
n
the LCS topology which has the following convex neighborhood
base at the origin
( = ¦l ⊂ A : l is convex, symmetric, all points of l are interior, and l∩A
n
is open for each :¦.
Theorem 14.5. With this topology C
∞
c
(R
d
) is a LCS and is complete. The topology does
not depend on the choice of sets Ω
n
and a sequence n
n
∈ C
∞
c
(R
d
) converges if and only if
there is ` such that n
n
∈ A
N
for all : and n
n
converges in A
N
.
For the proof see Reed and Simon.
A distribution is an element of T
(R
d
) = (C
∞
c
(R
d
))
. Since any tempered distribution 1
acts on C
∞
c
(R
d
) and furthermore if ¸φ. 1) = 0 for all φ in C
∞
c
then 1 = 0 (note that C
∞
c
(R
d
)
is dense in o), we have the embedding o
⊂ T. Distributions in T need not be bounded at
∞ – any locally integrable function, like e
e
e
x
2
is a distribution. Also, the structure theorem
doesn’t hold in this context, however we have
Theorem 14.6. Let 1 be a distribution in T(R
d
) then there is a sequence )
α
of continuous
functions, indexed by multiindices, such that
1 =
α
1
α
)
α
.
where for any n ∈ C
∞
c
(R
d
) only ﬁnitely many terms contribute to the sum
¸n. 1) =
α
¸n. 1
α
)
α
) .
Roughly speaking, the order of the distribution can become unbounded at ∞.
Likewise we may put a topology on C
∞
(R
d
) given by uniform convergence on compact
subsets. That is generated by the family of seminorms (), where now n need not have
compact support. This makes c(R
d
) = C
∞
(R
d
) into a Fr´echet space. Let us denote it’s dual
by c
. We have the inclusions
c
⊂ T
0
⊂ o
⊂ T
.
where T
0
is the dual of C
∞
0
(R
d
). Distributions in c
have compact support, where
Definition 14.1. A distribution vanishes on an open set l ⊂ R
d
if ¸φ. 1) = 0 whenever
φ has compact support in l. The support of a distribution is the smallest closed set 1 such
that 1 vanishes on R
d
¸ 1.
Distributions in T
0
are “bounded,” in the sense that they may be written as ﬁnite
sums of derivatives of ﬁnite measures. Distributions in o are “unbounded at ∞,” but only
polynomially so.
LECTURE 15
Applications: solving some PDE’s
The Poisson equation
The electric potential φ produced by a charge distribution ρ is known to satisfy the
Poisson equation:
−∆φ(r) = ρ(r). r ∈ R
3
with the permitivity of space set = 1. Let us use the fact that
−∆
1
4π
[r[
−1
= δ(r)
to solve this equation. We need the following notion.
Definition 15.1. Given two functions φ. ψ the convolution of φ and ψ is
φ ∗ ψ(r) =
_
R
d
φ(r −µ)ψ(µ)dµ.
whenever the integral is deﬁned.
Proposition 15.1. The convolution product is abelian, φ∗ψ = ψ∗φ and is a continuous
map of 1
1
1
1
→ 1
1
, o o → o. and T T → T.
Proof. Exercise.
Theorem 15.2 (Young’s Inequality). Convolution is a continuous map from 1
p
1
q
→
1
r
with 1,j + 1,¡ = 1 + 1,:, 1 ≤ j. ¡. : ≤ ∞.
Proof. Let ) ∈ 1
r
, 1,:
t
+ 1,: = 1. Then
[
_
)(r)φ ∗ ψ(r)dr[ ≤
_ _
[)(r)φ(r −µ)ψ(µ)[drdµ.
Write the integrand as α(r. µ)β(r. µ)γ(r. µ) with
α(r. µ) = [)(r)[
r
p
[φ(r −µ)[
q
p
β(r. µ) = [φ(r −µ)[
q
r
[ψ(µ)[
p
r
γ(r. µ) = [)(r)[
r
q
[ψ(µ)[
p
q
and 1,j + 1,j
t
= 1, 1,¡ + 1,¡
t
= 1. Note that 1,j
t
+ 1,¡
t
= 1,:
t
= 1 −1,:. So, by Holder,
_ _
[)(r)φ(r −µ)ψ(µ)[drdµ ≤
__ _
[)(r)[
r
[φ(r −µ)[
q
drdµ
_ 1
p
__ _
[φ(r −µ)[
q
[ψ(µ)[
p
drdµ
_1
r
__ _
[)(r)[
r
[ψ(µ)[
p
drdµ
_ 1
q
= )
r
p
+
r
q
r
φ
q
p
+
q
r
q
ψ
p
r
+
p
q
p
.
151
152 15. APPLICATIONS: SOLVING SOME PDE’S
Now note that if φ ∈ C
∞
and ) is integrable and compactly supported, say, then φ ∗ )
is C
∞
. Indeed
∂
j
(φ ∗ )) = (∂
j
φ) ∗ ).
Likewise, if φ ∈ o and ) is integrable with compact support — or more generally polynomial
decay at ∞, so j(r)) ∈ 1
1
for any polynomial j — then φ ∗ ) ∈ o. If also ψ ∈ o then
¸ψ. φ ∗ )) =
_
ψ ∗
¯
). φ
_
.
where
¯
)(r) = )(−r).
Thus we deﬁne the convolution 1 ∗ ) of a ), a compactly supported integrable function, and
a distribution 1 to be the distribution
¸ψ. 1 ∗ )) =
_
ψ ∗
¯
). 1
_
.
Theorem 15.3. If ) ∈ 1
1
has compact support and 1 is a tempered distribution, then
1 ∗ ) ∈ o
is a tempered distribution, and if 1 ∈ 1
1
loc
∩ o
∗
then 1 ∗ ) ∈ 1
1
loc
as well and
¸ψ. 1 ∗ )) =
_
R
d
R
d
ψ(r)1(r −µ))(µ)drdµ.
The map φ. 1 → 1 ∗ φ is a continuous map of o o
→ C
∞
∩ o
.
Proof. The ﬁrst two statements are clear. To show that 1 ∗φ is C
∞
if φ ∈ o, note that
this follows if 1 is a function since
∂
j
_
φ(r −µ)1(µ)dµ =
_
∂
j
φ(r −µ)1(µ).
More generally, we may write a general distribution as 1
α
p with p continuous, and integrate
by parts to see that
1 ∗ φ(r) =
_
1
α
φ(r −µ)p(µ)dµ.
Continuity of the map is easy.
Now, let 1(r) =
1
4π
[r[
−1
. This is a locally integrable function, so if ρ ∈ 1
1
loc
we have
φ(r) =
1
4π
_
1
[r −µ[
ρ(µ)dµ
as a (weak) solution of the Poisson equation. It is not the unique solution since we can add
to it any harmonic function l which satisﬁes ∆l = 0. For example, l(r) = r
1
is harmonic.
So is l(r) = r
1
r
2
. In fact, one can show that any Harmonic tempered distribution is a
polynomial. (The easy way to do this is to use the Fourier Transform.)
More generally we have the following theorem due to Weyl
Theorem 15.4 (Weyl). Let 1 be a distribution that satisﬁes ∆1 = 0 on an open set
1 ∈ R
n
, so ¸∆φ. 1) = 0 for φ ∈ C
∞
c
(1). Then 1 is a C
∞
function in 1: there is
p ∈ C
∞
(1) such that ¸φ. 1) = ¸φ. p) for φ ∈ C
∞
c
(1).
For the proof, see appendix B.
LECTURE 16
The Dirichlet problem
Reading: ¸7.2 of Lax
Suppose we want to solve Poisson’s equation for φ that vanishes outside an open set
1 ⊂ R
d
containing the support of ρ. That is
_
∆φ(r) = ρ(r) r ∈ 1
φ(r) = 0 r ,∈ 1.
It is useful to use Hilbert space methods. Let us set up the inner products
¸). p)
0
=
_
D
)(r)p(r)dr and ¸). p)
1
=
_
D
d
j=1
∂
j
)(r)∂
j
p(r)dr.
Take these over C
∞
c
(1), which is incomplete but completes to 1
2
(1) with respect to the
ﬁrst and to a Sobolev space in the second.
Lemma 16.1. For any ) ∈ C
∞
c
(1)
)
0
≤
diam(1)
√
d
)
1
where diam(1) = sup
x,y∈D
[r −µ[ is the width of 1.
Proof. This is a calculus exercise. Extend ) to be identically 0 outside 1. Note that
)(r) =
_
x
1
x
(0)
1
∂
1
)(µ. r
2
. . . . . r
d
)dµ.
where r
(0)
= (r
(0)
1
. r
2
. . . . . r
d
) ∈ ∂1. Thus by CauchySchwarz,
)(r)
2
≤ diam(1)
_
R
∂
1
)(µ. r
2
. . . . . r
d
)dµ.
Integrating over r ∈ 1 now gives
)
2
0
≤ diam(1)
2
∂
1
)
2
0
.
Similarly,
)
2
0
≤ diam(1)
2
∂
j
)
2
0
.
Averaging these results gives the Lemma.
Let H
(0)
1
denote the completion of C
∞
0
(1) in the norm 
1
. This is a Hilbert space.
Lemma 16.2. Every element ) ∈ H
(0)
1
may be identiﬁed with a locally integrable function
such that
(1) )(r) = 0 if r ,∈ 1
(2) ) ∈ 1
2
(1)
161
162 16. THE DIRICHLET PROBLEM
(3) For , = 1. . . . . d, ∂
j
) (in the sense of distributions) is a locally integrable function
in 1
2
(1).
Furthermore we have
¸). p)
1
=
D
j=1
¸∂
j
). ∂
j
p)
0
and
¸). ∂
j
p)
0
= −¸∂
j
). p)
whenever ). p ∈ H
(0)
1
.
Proof. The inequality )
0
≤ const.. )
1
shows that any Cauchy sequence of function
in the 
1
norm is Cauchy in 1
2
. So any sequence )
n
∈ C
∞
c
(1) which converges in the H
(0)
1
also converges in 1
2
. Identify the limit ) with the corresponding element of 1
2
(1), extended
to be zero outside 1. (1) and (2) follow.
To derive (3), note that for p ∈ C
∞
c
¸p. ∂
j
)
n
) = −¸∂
j
p. )
n
)
by integration by parts. The r.h.s. converges to ¸∂
j
p. )) using CauchySchwarz. On the other
hand, ∂
j
)
n
converges in the 1
2
norm to a function /
j
. We conclude that /
j
= ∂
j
) in the
sense of distributions, so ∂
j
) ∈ 1
2
(1).
To show the formulas, note that they hold for elements of C
∞
c
(1) and follow for the
). p ∈ H
(0)
1
by taking limits.
Now, ﬁx ρ ∈ 1
2
(1). This function gives rise to a linear functional / on H
(0)
1
⊂ 1
2
(1) by
/(n) = ¸n. ))
0
.
Since
[/(n)[ ≤ n
0
)
0
≤
diam(1)
√
d
)
0
n
1
.
this is a bounded linear functional. By RieszFrechet there is an element φ ∈ H
(0)
1
such that
¸n. ρ)
0
= ¸n. φ)
1
.
for all n ∈ H
(0)
1
. Specializing to n ∈ C
∞
c
(1), we ﬁnd that
) = −∆φ
in the sense of distributions.
In a similar way, we may write down a bilinear form
1(n. ·) =
_
D
_
i,j
¹
i,j
(r)∂
i
n(r)∂
j
·(r) +n(r)
i
1
i
(r)∂
i
·(r) +:(r)n(r)·(r)
_
dr.
Let us check the hypotheses of LaxMilgram. Clearly 1 is bilinear and bounded. To estimate
1(n. n) from below, suppose that ¹
i,j
(r) is pointwise positive deﬁnite, that is
i,j
¹
i,j
(r)λ
i
λ
j
≥ σ(r)
i
λ
2
i
.
Then
1(n. n) ≥
_
D
_
σ(r) [∇n(r)[
2
−[n(r)[ [F(r)[ [∇n(r)[ +:(r)n(r)
2
_
dr.
16. THE DIRICHLET PROBLEM 163
Now suppose the quadratic function
σ(r)t
2
−[F(r)[ t: +:(r):
2
≥ δ(t
2
+:
2
).
uniformly in r, for some δ 0. This amounts to the requirement that σ(r) 0, :(r) 0
and the discriminant
sup
x
[F(r)[
2
−4σ(r):(r) < 0.
Then
1(n. n) ≥ δ(n
2
1
+n
2
0
) ≥ δ n
2
1
.
Thus, by LaxMilgram given ρ ∈ 1
2
(1) we may ﬁnd φ ∈ H
(0)
1
such that
¸·. ρ)
0
= 1(·. φ) ∀· ∈ H
(0)
1
.
Such φ is a distributional solution to the equation
−
i,j
∂
i
¹
i,j
(r)∂
j
φ(r) +
i
1
i
(r)∂
i
φ(r) +:(r)φ(r) = ρ(r). r ∈ 1.
One might ask, in what sense do the elements of H
(0)
1
vanish outside 1? The complete
answer to this question involves the theory of Sobolev spaces. An elementary answer is the
following:
Lemma 16.3. Let 1 be a hypercube [−1. 1]
d
and let ) ∈ H
(0)
1
. Then
lim
δ→0
1
[¦dist(r. ∂1) < δ¦[
_
dist(x,∂D)<δ
[)(r)[ = 0.
Remark. [ [ denotes Lebesgue measure, so [¦dist(r. ∂1) < δ¦[ = C(δ),
Proof. The set of points within distance δ of ∂1 is a union of slabs such as o = ¦r ∈
1 : −1 < r
1
< −1+δ¦. Let ) ∈ C
∞
c
(1) and consider the integral of [)(r)[ over this slab.
Integrating by parts, we have
_
S
[)(r)[dr =
_
S
∂
1
[)(r)[(−1 +δ −r
1
)dr.
since ) vanishes on the boundary. Now ∂
1
[)(r)[ ≤ [∂
1
)(r)[. Thus,
_
S
[)(r)[dr ≤
_
S
[∂
1
)(r)[(−1 +δ −r
1
)dr ≤ δ[o[
1
2
)
1
≤ C )
1
δ
3
2
.
since [o[ = C(δ). Adding up the contributions from all slabs we get
_
¡dist(x,∂D)<δ¦
[)(r)[dr ≤ C )
1
δ
3
2
.
Taking limits, this estimate follows, with constant C independent of ), for ) ∈ H
(0)
1
. Since
[dist(r. ∂1) < δ[ ≥ cδ the result follows.
Remark. We were a bit wasteful. We could have averaged over just Σ
δ
= ¦dist(r. Σ) <
δ¦ with Σ a piece of the boundary with codimension 2. In that case [Σ
δ
[ · δ
2
and
_
Σ
δ
[)(r)[dr ≤ Cδ
2
__
Σ
δ
[∇)(r)[
2
dr
_1
2
.
The integral on the r.h.s. goes to zero, so the average of [)[ vanishes also on codimension 2
sets.
Part 5
Weak Convergence and Weak Topology
LECTURE 17
Dual of a Banach space
Reading: ¸8.2 8.3 of Lax.
Recall that the dual of a Banach space A is the space A
t
of bounded linear functionals
on A. Recall that we put a norm topology on A
t
by deﬁning
/ = sup r ,= 0
[/(r)[
r
.
Likewise, we have the following dual characterization of the norm on A:
Theorem 17.1. For every r ∈ A we have
r = max / ,= 0. / ∈ A
t
[/(r)[
/
.
Proof. Since / µ ≥ [/(µ)[ the l.h.s. is no smaller than the r.h.s. Thus we need only
produce an / such that [/(r)[ = r /. Deﬁne / ﬁrst on the one dimensional subspace
span¦r¦ by /(tr) = t r. Since this functional is norm bounded by 1 on this subspace it
has an extension (by HahnBanach) to the whole space with this property.
We also have the weak
∗
topology on A
t
, the weakest LCS topology such that all elements
of A are continuous functionals on A
t
. Recall that (A
t
. wk
∗
)
∗
= A. One might wonder if
(A
t
)
t
= A as Banach spaces. In fact this does not hold in general.
Definition 17.1. A Banach space is called reﬂexive if (A
t
)
t
= A. That is if every
bounded linear functional on A
t
is of the form / → /(r) for some r ∈ A.
Many important spaces are reﬂexive, but not all. For instance:
Theorem 17.2. c
t
0
= /
1
, /
t
1
= /
∞
, and /
t
∞
/
1
.
Proof. Let / be a linear functional on c
0
. Evaluating / on the sequence e
k
(:) = 1 if
/ = : and 0 otherwise produces a sequence
b(:) = /(e
n
).
If a ∈ c
0
then
a = lim
n→∞
n
j=1
c(,)e
j
.
so by linearity and continuity
/(a) = lim
n→∞
n
j=1
c(,)b(,).
171
172 17. DUAL OF A BANACH SPACE
That is for any a ∈ C
0
the sequence ab is summable and
/(a) =
∞
j=1
a(,)b(,).
To see that b ∈ /
1
, take a
n
(,) = e
−i arg b(j)
if , ≤ : and 0 otherwise. Thus
n
j=1
[b(,)[ = /(a
n
) ≤ / .
It follows that b ∈ /
1
. (Conversely, it is clear that any b ∈ /
1
gives a linear functional on
c
0
.)
The same idea works to prove /
t
1
= /
∞
. Finally, it is clear that /
1
⊂ /
t
∞
, however no
Banach limit such as deﬁned in the third lecture can be written as a scalar product with
something in /
1
.
However,
Theorem 17.3. Every Hilbert space is reﬂexive.
Proof. In this case A
t
= A by RieszFr´echet.
For 1
p
spaces we have.
Theorem 17.4. Let A. j be a ﬁnite measure space. For 1 ≤ j < ∞ the dual of 1
p
(A)
is 1
q
(A) where
1
p
+
1
q
= 1.
Remark. The result, but not the proof, extends to arbitrary measure spaces for 1 <
j < ∞. If j = 1 (so ¡ = ∞) the result holds in σﬁnite measure spaces.
Proof. The H¨older inequality shows that 1
q
(A) → 1
p
(A)
t
, via the pairing
¸). p) =
_
X
)(r)p(r)dj(r). ) ∈ 1
p
and p ∈: 1
q
.
Suppose now that / is a linear functional on 1
p
(A). For each measurable set ¹ we have
χ
A
∈ 1
p
(A). So
ν(¹) = /(χ
A
)
deﬁnes a ﬁnitely additive set function. In fact, it is countably additive since / is continuous,
and if ¹
1
. ¹
2
. . . . are pairwise disjoint then
n
j=1
χ
A
j
→ χ
∪A
j
in 1
p
as one may readily show. Furthermore ν(¹) = 0 if j(¹) = 0 since then χ
A
= 0 in 1
p
. Thus
ν << j and by RadonNikodym there is function p ∈ 1
1
such that
/(χ
A
) = ν(¹) =
_
A
p(r)dj(r) =
_
X
χ
A
(r)p(r)dj(r).
By taking limits of simple functions we have
/()) =
_
A
)(r)p(r)dj(r) ()
whenever ) ∈ 1
∞
, using dominated convergence.
17. DUAL OF A BANACH SPACE 173
It remains to show that p ∈ 1
q
for then () extends to all of 1
p
by density of 1
∞
. To
this end, ﬁx t 0 and let
)
t
(r) =
_
[g(x)[
q
g(x)
0 < [p(r)[ < t
0 [p(r)[ = 0 or [p(r)[ ≥ t.
Clearly [)
t
(r)[ ≤ t
q−1
so )
t
∈ 1
∞
⊂ 1
p
. Thus
_
¡[g(x)[<t¦
[p(r)[
q
dj(r) =
_
)
t
(r)p(r)dj(r) ≤ /
__
[)
t
(r)[
p
dj(r)
_1
p
.
But [)
t
(r)[
p
= [p(r)[
pq−p
= [p(r)[
q
on ¦[p(r)[ < t¦. Thus
__
¡[g(x)[<t¦
[p(r)[
q
dj(r)
_
1−
1
p
≤ /
and the result follows.
Corollary 17.5. 1
p
(A) is reﬂexive for 1 < j < ∞.
This result also follows from
Theorem 17.6 (Milman 1938). Any uniformly convex Banach space is reﬂexive
In general 1
1
is not reﬂexive: (1
1
)
t
= 1
∞
but 1
∞
contains linear functionals that are
not in 1
1
. Notice where the proof breaks down. Given a linear functional / on 1
∞
we can
deﬁne a set function
ν(¹) = /(χ
A
)
as above. It is certainly additive, and absolutely continuous since if j(¹) = 0 then χ
A
= 0
in 1
∞
. It is not, however, countably additive since
n
j=1
χ
A
j
,→ χ
∪A
j
in 1
∞
.
The inequality (1
∞
) ,= 1
1
also follows from:
Theorem 17.7. Let A be a Banach space over C. If A
t
is separable so is A.
Proof. Let ¦/
n
¦ be a countable dense subset of A
t
. For each : there is r
n
∈ A such
that
r
n
 = 1 and /
n
(r
n
) ≥
1
2
/
n
 .
It suﬃces to show span¦r
n
¦ is dense in A.
Suppose contrarily that span¦r
n
¦ , = A. Then there is a nonzero linear functional / ∈ A
t
such that /(r
n
) = 0 for all :. We may assume that / = 1. However, we can ﬁnd : such
that / −/
n
 ≤
1
4
, say. Thus /
n
 ≥
3
4
and
0 = /(r
n
) = /(r
n
) −/
n
(r
n
) +/
n
(r
n
) ≥
1
2
/
n
 −/ −/
n
 ≥
1
8
.
Thus no such / exists and we must have span¦r
n
¦ = A.
LECTURE 18
RieszKakutani theorem
Theorem 18.1 (RieszKakutani). Let Q be a compact Hausdorﬀ space, C(Q) the space
of continuous real valued functions on Q with the max norm. Then C(Q)
t
= `(Q) = set of
signed Borel measures of ﬁnite total variation.
That is to every bounded linear functional / ∈ C(Q)
t
is associated a unique Borel measure
: such that
/()) =
_
Q
)d:.
Furthermore the norm of / is the total variation / =
_
Q
[d:[.
Remark. i) A Borel measure is a measure on sets in the Borel σalgebra which is the
smallest σalgebra containing the open sets on Q. ii) The dual of C(Q; C) is the set of
complex Borel measures, : = :
r
+ i:
i
with :
r
. :
i
∈ `(Q). iii) Compactness is crucial
here. (More on this later.)
Before proving this theorem, let us prove
Theorem 18.2. Given / ∈ C(Q) there is a unique deomposition / = /
+
− /
−
with /
±
positive linear functionals and / = /
+
(1) +/
−
(1).
Proof. Let C(Q)
+
= set of nonnegative functions in C(Q). For ) ∈ C(Q)
+
deﬁne
/
+
()) = sup¦/(/) : 0 ≤ / ≤ )¦.
It is clear that /
+
(t)) = t/
+
()) for t ≥ 0 and that /
+
()) ≥ /(0) = 0 for ) ∈ C(Q)
+
. Given
), p in C(Q)
+
their sum is also in C(Q)
+
. Clearly
/
+
() +p) ≥ sup¦/(/
1
) +/(/
2
) : 0 ≤ /
1
≤ ) and 0 ≤ /
2
≤ p¦ = /
+
()) +/
+
(p).
The opposite inequality clearly follows from the following result
Claim. Given ). p ∈ C(Q) and 0 ≤ / ≤ ) +p we can write / = /
1
+/
2
with /
1,2
∈ C(Q),
0 ≤ /
1
≤ ) and 0 ≤ /
2
≤ p.
Proof of Claim. Let /
1
= min¦). /¦. So /
1
is continuous and 0 ≤ /
1
≤ ). Let
/
2
= / −/
1
. Since /
1
≤ /, /
2
≥ 0. When /
1
= ) we have /
2
≤ ) +p −) = p. On the other
hand if /
1
= / we have /
2
= 0 so /
2
≤ p.
Thus /
+
(t) + :p) = t/
+
()) + :/
+
(p) whenever t. : ≥ 0 and ). p ∈ C(Q)
+
. Deﬁne /
+
on
all of C(Q) by
/
+
()) = /
+
()
+
) −/
+
()
−
). with )
+
= max¦). 0¦ and )
−
= min¦). 0¦.
It is not hard to see that /
+
is linear. Now set /
−
= /
+
−/ and note that
/
−
()) = sup¦/(/) −/()) : 0 ≤ / ≤ )¦ = sup¦/(/) : −) ≤ / ≤ 0¦ ≥ 0.
for ) ∈ C(Q)
+
.
181
182 18. RIESZKAKUTANI THEOREM
To prove / = /
+
(1) +/
−
(1) note ﬁrst that /
±
 = /
±
(1) since these are positive linear
functionals. Thus, / ≤ /
+
 +/
−
 = /
+
(1) +/
−
(1). On the other hand by the deﬁnition
of /
+
and the corresponding inequality for /
−
we have
/
+
(1) +/
−
(1) = sup¦/(/) : 0 ≤ / ≤ 1¦ + sup¦/(/) : −1 ≤ / ≤ 0¦
= sup¦/(p) : −1 ≤ p ≤ 1¦ ≤ / .
We are now ready to prove the RieszKakutani Theorem. By the splitting of a linear
functional into positive and negative parts, it suﬃces to show
Theorem 18.3 (RieszKakutani). Let / be a positive linear functional on C(Q) with Q
a compact Hausdorﬀ space. Then there is a unique positive Borel measure : on Q such that
/()) =
_
Q
)(r)d:(r) and / = :(Q). Conversely, any positive Borel measure : gives a
positive linear functional on C(Q) via ¸). :) =
_
Q
)(r)d:(r).
Proof. First, it is clear that any positive Borel measure : gives rise to a positive linear
functional by /()) =
_
Q
)(r)d:(r). (Continuous functions are Borel measurable.) Since
the functional is positive / = /(1) = :(Q).
It remains to show that every positive linear functional is of this form. So let / ∈ C(Q)
t
be
given. We would like to deﬁne :(o) = /(χ
S
). for o a measurable set. However, χ
S
,∈ C(Q)
(unless o is both open and closed, so a union of connected components). Thus we do the
next best thing: given an open set l we take
:(l) = sup¦/()) : ) ≺ l¦.
where, ) ≺ l indicates that ) ∈ C(Q), supp ) ⊂ l and 0 ≤ ) ≤ 1.
Claim. Let l
1
. l
2
⊂ Q be a disjoint pair of open sets. Then :(l
1
∪l
2
) = :(l
1
)+:(l
2
).
Proof of claim. Note that given )
j
≺ l
j
, we have )
1
+ )
2
≺ l
1
∪ l
2
and conversely
given ) ≺ l
1
∪ l
2
, we have )χ
U
j
≺ l
j
. The identity follows.
Thus : is ﬁnitely additive on open sets. So far the construction works on an arbitrary
compact space, but the open sets do not form a σalgebra. A ﬁrst step is to deﬁne : on
closed sets 1 by
:(1) = :(Q) −:(1) = inf¦/()) : 1 ≺ )¦.
where 1 ≺ ) if 0 ≤ ) ≤ 1 and )(r) = 1 on 1.
To proceed we need the following result which uses the assumption that Q is a compact
Hausdorﬀ space:
Claim. If 1 ⊂ l, 1 closed and l open, then
:(1) = inf¦/()) : 1 ≺ ) ≺ l¦ ≤ sup¦/()) : 1 ≺ ) ≺ l¦ = :(l).
Proof of claim. It is clear that
:(1) ≤ inf¦/()) : 1 ≺ ) ≺ l¦ ≤ sup¦/()) : 1 ≺ ) ≺ l¦ ≤ :(l).
To obtain equality on the two ends, we use Urysohn’s lemma, valid here since a compact
Hausdorﬀ space is normal (disjoint closed sets may be separated by open sets):
Theorem 18.4 (Uryshohn’s Lemma). Given two disjoint closed sets 1
0
. 1
1
⊂ Q there is
a continuous function ) : Q → [0. 1] such that ) ≡ 0 on 1
0
and ) ≡ 1 on 1
1
.
18. RIESZKAKUTANI THEOREM 183
Given )
1
≺ l, let 1
1
= 1 ∪ supp )
1
, and let \ be open with 1
1
⊂ \ ⊂ 1 ⊂ l — such
\ exists because Q is normal. Now let 1
0
= \
c
. Then ) ≡ , on 1
j
, , = 0. 1 clearly satisﬁes
)
1
≤ ) and 1 ≺ ) ≺ l. We conclude that
sup¦/()) : 1 ≺ ) ≺ l¦ = sup¦/()) : ) ≺ l¦ = :(l).
Similarly given )
1
~ 1, let 1
1
= 1 and 1
0
= l
c
. Pick )
2
∈ C(Q) with 1 ≺ )
2
≺ l and
set ) = )
1
)
2
. Clearly 1 ≺ ) ≺ l and ) ≤ )
1
. Thus /()) ≤ /()
1
) so
inf¦/()) : 1 ≺ ) ≺ l¦ = inf¦/()) : 1 ≺ )¦.
It now follows that
Claim. If 1 is closed then :(1) = inf¦:(l) : 1 ⊂ l and l open¦.
If l is open then :(l) = sup¦:(1) : 1 ⊂ l and 1 closed¦.
Proof of claim. Clearly :(1) ≤ inf¦:(l) : 1 ⊂ l¦. To prove the converse, note
that given c 0 there is ) ~ 1 such that /()) ≤ :(1) + c. Let l = ¦r : )(r)
1 − c¦. Note that p ≺ l =⇒ p ≤
1
1−
) so /(p) ≤
1
1−
/()). Thus :(l) ≤
1
1−
/()). Thus
(1 −c):(l) ≤ :(1) +c. It follows that inf¦:(l) : 1 ⊂ l¦ ≤ :(1).
The opposite identity, :(l) = sup¦:(1)¦, follows by taking complements.
Now given arbitrary o ⊂ Q deﬁne
:
+
(o) = inf¦:(l) : o ⊂ l and l open¦.
:
−
(o) = sup¦:(1) : o ⊃ 1 and 1 closed¦.
Clearly :
−
(o) ≤ :
+
(o). If these two numbers are equal deﬁne
:(o) = :
±
(o) if :
+
(o) = :
−
(o).
Claim. The collection Σ = ¦o ⊂ Q : :
+
(o) = :
−
(:)¦ is a σalgebra containing all
Borel sets and : deﬁnes a countably additive measure on this σalgebra.
Proof of Claim. Clearly Σ contains all closed and all open sets. Thus once we show
it is a σalgebra it is immediate that it contains all Borell sets. Clearly o ∈ Σ =⇒ o
c
∈ Σ.
Thus we need only show that Σ is closed under countable unions. This is left as an exercise
as is countable additivity of :.
To complete the proof, we must show that
_
)d: = /()). It suﬃces to prove this for )
with 0 ≤ ) ≤ 1. Note that
_
Q
)d: =
_
1
0
:¦r : )(r) ≥ t¦dt =
_
1
0
:¦r : )(r) t¦dt.
Let us show that
_
)d: ≤ /()). Note that
_
Q
)d: ≤
n
j=1
1
:
:
_
r : )(r) ≥
, −1
:
_
.
Let p
j;n
be functions with
_
)(r) ≥
j−1
n
_
≺ p
j;n
≺
_
)(r)
j−2
n
_
. Such functions exist by
Urysohn’s Lemma. So
_
Q
)d: ≤
n
j=1
1
:
/(p
j;n
) = /(
n
j=1
1
:
p
j;n
).
184 18. RIESZKAKUTANI THEOREM
Now for every r ∈ Q
n
j=1
1
:
p
j;n
(r) =
]nf(x)+1
j=1
1
:
+C(
1
:
) =
1
:
¸:)(r) +C(
1
:
).
We conclude that ¸
¸
¸
¸
¸
n
j=1
1
:
p
j;n
(r) −)(r)
¸
¸
¸
¸
¸
= C(
1
:
).
so
/(
n
j=1
1
:
p
j;n
) −→ /()).
and
_
)d: ≤ /()).
To show the reverse inequality, note that
_
Q
)d: ≥
n
j=1
:
_
)(r)
,
:
_
.
Thus,
_
Q
)d: ≥ /(
n
j=1
1
:
/
j;n
).
with ¦)(r) ≥
j+1
n
¦ ≺ /
j;n
≺ ¦)(r)
j
n
¦. Then
n
j=1
1
:
/
j;n
(r) =
]nf(x)−1
j=1
1
:
+C(
1
:
) =
1
:
¸:)(r) +C(
1
:
).
Again
j
/
j;n
→ ), so
_
)d: ≥ /()). completing the proof.
A word about the noncompact case. Suppose A is a locally compact Hausdorﬀ space,
so A is a Hausdorﬀ space such that every point is contained in an open set with compact
closure. Then we can consider several spaces C
c
(A) ⊂ C
0
(A) ⊂ C
b
(A) ⊂ C(A) where
C
b
(A) = ¦bounded continuous functions¦. The middle two are Banach spaces in the sup
norm. The ﬁrst and last are LCS spaces: C
c
(A) has an inductive limit topology obtained
from writing it as C
c
(A) = ∪
U
C
0
(l) where the union is over open sets l with compact
closure, C(A) has a topology generated by the seminorms j
K
()) = sup
x∈K
[)(r)[ for compact
1. Regarding the duals of these spaces, we have
C
c
(A)
t
= `(A) = ¦Borel measures : such that [:[(1) < ∞ for any compact 1¦.
C
0
(A)
t
= `
0
(A) = ¦ﬁnite Borel measures on A¦.
C
b
(A)
t
= ¦ﬁnite Borel measures on the Stone
˘
Cech compactiﬁcation of A¦.
C(A)
t
= ¦compactly supported Borel measures¦.
LECTURE 19
Weak convergence
Reading: ¸10.1 and 10.2 of Lax.
We have already deﬁned weak topologies in the general context of LCS spaces. Let us
now look at them in the special case of Banach spaces. Consider a Banach space A and
it’s Banach space dual A
t
. So A
t
is a collection of linear functionals on A. The weakest
topology on A so that every element of A
t
is continuous is called the weak topology on A,
denoted σ(A. A
t
).
A sequence ¦r
n
¦ in A converges weakly to r — converges in the weak topology — if
/(r
n
) → /(r) for every / ∈ A
t
.
Sometimes this is denoted
r
n
÷ r.
or
wk − lim
n→∞
r
n
= r.
This notion is weaker than strong convergence, which is in norm:
r
n
−r → 0. or r
n
→ r.
That is more sequences converge weakly than converge strongly. For instance
χ
[n,n+1]
÷ 0 in 1
p
(R). 1 < j < ∞.
but
_
_
χ
[n,n+1]
_
_
L
p
= 1.
(Note that χ
[n,n+1]
does not converge weakly in 1
1
.)
Proposition 19.1. Let ¦r
n
¦ be an orthonormal sequence in a Hilbert space H. Then
r
n
÷ 0.
Remark. Since r
n
 = 1, r
n
does not converge strongly to 0.
Proof. Fix µ ∈ H. By Bessel’s inequality
n
[ ¸µ. r
n
) [
2
≤ µ
2
.
we see that ¸µ. r
n
) → 0. By the Riesz theorem on linear functionals on a Hilbet space,
r
n
÷ 0.
Theorem 19.2. Suppose ¦r
n
¦ ∈ A, a Banach space, satisﬁes
(1) r
n
are uniformly bounded: sup
n
r
n
 < ∞.
(2) lim/(r
n
) = /(r) for / ∈ 1
t
with 1
t
dense in A
t
.
Then r
n
÷ r.
Proof. This is an easy approximation argument and is left as an exercise.
191
192 19. WEAK CONVERGENCE
The interesting thing is that the converse is true: weakly convergent sequences are
uniformly bounded. To prove this we will use
Theorem 19.3 (Principle of Uniform Boundedness for a complete metric space). Let A
be a complete metric space and T a collection of real valued continuous functions on A. If
T is bounded at each point r ∈ A,
[)(r)[ ≤ `(r) < ∞ for all ) ∈ T.
then there is an open set l ⊂ A and a constant ` < ∞ such that
[)(r)[ ≤ ` for all r ∈ l and ) ∈ T.
Proof. This result follows from the Baire Category Theorem of topology:
Theorem 19.4 (Baire Category Theorem). A complete metric space is not the union of
a countable number of nowhere dense sets.
Remark. Recall that the interior of a set o is the largest open set contained in o, that
is
int o = o
o
=
_
¦l ⊂ o : l open¦.
and that a set o is nowhere dense if it’s closure o has empty interior. Thus a closed set is
nowhere dense if o ⊂ o
c
. For the proof see Reed and Simon Chapter III or any book on
point set topology.
To prove the PUB, note that by assumption
A =
_
n
¦r : [)(r)[ ≤ : for all ) ∈ T¦.
Thus, at least one of the (closed) sets ¦r : [)(r)[ ≤ : ∀) ∈ T¦ has nonempty interior,
which is to say it contains an open set l. This is the open set claimed in the theorem.
Suppose A is a Banach space and each function ) ∈ T is subadditive ()(r + µ) ≤
)(r) + )(µ)) and absolutely homogeneous ()(cr) = [c[)(r)). For instance each ) could be
of the form )(r) = [/(r)[ for some linear functional. Then
Theorem 19.5 (Principle of Uniform Boundedness for subadditive functionals). Let A
be a Banach space and let T be a collection of realvalued continuous, subadditive, absolutely
homogeneous functions on A. Suppose for each r ∈ A, [)(r)[ ≤ `(r) < ∞ for all ) ∈ T.
Then the function ) ∈ T are uniformly bounded in the sense that there is c < ∞ such that
[)(r)[[ ≤ c r for all r ∈ A and ) ∈ T.
Proof. Clearly the hypotheses of the PUB for metric spaces applies. Let l be the open
set claimed and let r
0
∈ l. Since l is open there is c 0 such that µ < c =⇒ r
0
+µ ∈ l.
Now consider µ with µ < c. We have, for ) ∈ T,
)(µ) = )(µ +r
0
−r
0
) ≤ )(µ +r
0
) +)(r
0
) ≤ 2`.
Thus for arbitrary r ∈ A and ) ∈ T,
)(r) =
2 r
c
)(
c
2 r
r) ≤
4`
c
r .
19. WEAK CONVERGENCE 193
Corollary 19.6. Let A be a Banach space and let L be a collection of bounded linear
functionals that is pointwise bounded, so /(r) ≤ `(r) for all / ∈ L, then there is a constant
c < ∞ such that
/ ≤ c for all / ∈ L.
Corollary 19.7. Let A be a Banach space and let o ⊂ A be a weakly precompact
subset of A. Then there is a constant c < ∞ such that
r ≤ c for all r ∈ o.
In particular, any weakly convergent sequence is norm bounded.
Remark. Recall that a set o is precompact if o is compact. Thus any sequence in a
precompact set has a convergent subsequence, but the limit may lie outside o.
Proof. Think of points of A as functions on A
t
. Since o is weakly compact /(r) must be
bounded for each / as r ranges over o (otherwise we could ﬁnd a weakly divergent sequence).
By the PUB there is a constant c such that r ≤ c for all r ∈ o.
A function ) : A → R, A a topological space, is called lower semicontinuous if ¦r :
)(r) t¦ is open for each t ∈ R. Such a function satisﬁes )(r) ≤ liminf
n
)(r
n
) for any
convergent sequence r
n
→ r. (Note that for each c 0 the set ¦µ : )(µ) )(r) − c¦ is
open and thus eventually contains r
n
so liminf r
n
≥ )(r) −c.)
Theorem 19.8 (Weak lower semicontinuity of the norm). Let A be a Banach space. The
norm  is weakly lower semicontinuous. In particular, if r
n
÷ r in A then
r ≤ liminf r
n
 .
Remark. 1) This should remind you of Fatou’s lemma from measure theory. 2) We have
already seen that the norm is not continuous, since it may “jump down” in a limit.
Proof. Fix t ≥ 0. Let A
t
1
denote the unit ball ¦/ : / ≤ 1¦ in A
t
. Note that r t
if and only if there is a linear functional / ∈ A
t
1
with [/(r)[ t . Thus
¦r t¦ = ∪
∈X
1
¦r : /(r) t¦
is weakly open.
LECTURE 20
Weak sequential compactness, weak
∗
convergence and the weak
topology
Definition 20.1. A subset C of a Banach space A is called weakly sequentially compact
if any sequence of pints in C has a weakly convergent subsequence, whose weak limit is in
C.
Recall that sequential compactness is, in general, a strictly weaker notion than compact
ness. They are equivalent, however, in metric spaces. In the present context, A is metrizable
in the σ(A. A
t
) topology if and only if A
t
is separable.
Proposition 20.1. A weakly sequentially compact set is bounded
Proof. Use the PUB. Details left as an exercise.
Theorem 20.2. In a reﬂexive Banach space A the closed unit ball is weakly sequentially
compact.
Remark. We will see that the unit ball is, in fact, weakly compact in a reﬂexive Banach
space, and more generally in the dual of a Banach space. However, the proof of that result
is far less constructive than the following.
Proof. Let ¦µ
n
¦ be a sequence of points in the unit ball. Let 1 be the closed linear
span span¦µ
n
¦. Since A is reﬂexive, it follows that 1 is reﬂexive, since
Theorem 20.3 (Thm. 15 of Chapter 8 in Lax). Any closed subspace of a reﬂexive space
is reﬂexive.
Proof. See Lax.
Since 1 = 1
tt
is separable, it follows that 1
t
is separable. So 1
t
contains a dense
countable subset ¦:
j
¦. Consider the array of scalars
c
i,j
= :
i
(µ
j
).
The i
th
row is bounded by :
i
. Starting with the ﬁrst row we pick a subsequence µ
j
(1)
k
so
that :
1
(µ
j
(1)
k
) converges. Reﬁne this subsequence again and again to produce subsequences
µ
j
(n)
k
for each : so that :
1
(µ
j
(n)
k
). . . . . :
n
(µ
j
(n)
k
) all converge. Now let
.
n
= µ
j
(n)
n
.
Clearly :
j
(.
n
) converges as : → ∞ for each ,. By density of ¦:
j
¦ in 1
t
it follows that
lim
n→∞
:(.
n
) = µ(:)
exists for each : ∈ 1
t
. This limit is clearly a linear functional of : and since
:(.
n
) ≤ : .
n
 ≤ : .
201
202 20. WEAK SEQUENTIAL COMPACTNESS, WEAK
∗
CONVERGENCE AND THE WEAK
TOPOLOGY
the linear functional is bounded. Since 1 is reﬂexive, we see that there is µ ∈ 1 such that
µ(:) = :(µ). Since the restriction of / ∈ A
t
to 1 gives an element of 1
t
, we have .
n
÷ µ
in A.
We may also consider a weak topology on the dual A
t
of a Banach space. That is the
weak ∗ topology σ(A
t
. A). A sequence n
n
of linear functionals is said to be weak ∗ convergent
to n if
limn
n
(r) = n(r) for all r ∈ A.
also denoted
wk
∗
− lim
n→∞
n
n
= n.
Weak
∗
convergence of measures is also known as vague convergence. If A is reﬂexive then
weak
∗
convergence is the same as weak convergence, but in general the weak
∗
topology is
strictly weaker than the weak topology since the latter makes all linear functionals in A
tt
continuous.
Theorem 20.4. A weak
∗
convergent sequence n
n
is uniformly bounded and
n ≤ liminf n
n
 .
if n = wk
∗
−limn
n
.
Proof. Exercise.
Definition 20.2. A subset C of a dual Banach space A
t
is weak
∗
sequentially compact
if every sequence of points in C has a weak
∗
convergent subsequence with weak
∗
limit in C.
Theorem 20.5 (Helly 1912). Let A be a separable Banach space. Then the closed unit
ball in A
t
is weak
∗
sequentially compact.
Proof. Given n
n
∈ A
t
with n
n
 ≤ 1 and a countable dense subset ¦r
n
¦ of A, we can
use the diagonal process to select a subsequence ·
n
of n
n
so that
lim
n→∞
·
n
(r
k
)
exists for every r
k
. By density of ¦r
k
¦ this extends to all of A:
lim
n→∞
·
n
(r) = ·(r)
for all r ∈ A. One readily veriﬁes that · is linear and bounded, so it is the desired limit.
In fact, more is true. The unit ball in A
t
is weak
∗
compact, even if A is no separable:
Theorem 20.6 (Alaoglu). Let A
t
be the dual of a Banach space A. The unit ball of A
t
is wk
∗
compact.
Proof. Let 1 be the unit ball in A
t
. Let 1 be the (uncountable) product space:
1 =
x∈X
1
x
. 1
x
= [−r . r].
By the Tychonov theorem 1 is compact in the product topology. To complete the proof, we
embed 1 as a closed subset of 1.
The inﬁnite product space 1 is the collection of all functions 1 : A → R such that
1(r) ∈ 1
x
for all r. Given / ∈ 1, [/(r)[ ≤ / r ≤ r so /(r) ∈ 1
x
for every r. Thus
1 ⊂ 1.
20. WEAK SEQUENTIAL COMPACTNESS, WEAK
∗
CONVERGENCE AND THE WEAK
TOPOLOGY 203
Now the product topology on 1 is just the weakest topology such that coordinate eval
uation 1 → 1(r) is continuous for every r. The restriction of this topology to 1 is just the
wk
∗
topology on 1.
Thus we have embedded 1 as a subset of the compact space 1. It suﬃces to show that
1 is closed. For each r. µ ∈ A and t ∈ R, let
Φ
x,y;t
(1) = 1(r +tµ) −1(r) −1(µ).
a continuous map of 1 into the ﬁeld of scalars. Clearly 1 ⊂ Φ
−1
x,y;t
(¦0¦) and Φ
−1
x,y;
(¦0¦) is a
closed set. Thus
1 ⊂
x,y,t
Φ
−1
x,y;
(¦0¦).
so every element of 1 is linear. Since any 1 ∈ 1 is also bounded by r, [1(r)[ ≤ r, we
conclude that 1 = 1.
Clearly Alaoglu’s theorem implies Helly’s theorem. However, the proof of Helly’s theorem
is much more useful. Often times what one really wants is to ﬁnd a convergent sequence.
The proof Helly’s theorem gives you an idea how to construct it; Alaoglu’s theorem just tells
you it is there.
Corollary 20.7. The unit ball in a reﬂexive space is weakly compact.
Remark. In fact weak compactness of the unit ball is equivalent to reﬂexivity, a result
due to Eberlein (1947) and Smulyan (1940).
In particular, this result applies to any Hilbert space and to 1
p
, 1 < j < ∞. The unit
ball in 1
∞
is weak
∗
compact since 1
∞
= (1
1
)
t
. The unit ball in 1
1
is not weakly compact.
Here is what happens in 1
1
. Consider, for example, 1
1
([0. 1]), and let
)
n
(r) = :χ
[0,
1
n
]
(r).
So )
n

L
1
= 1 and for any continuous function p ∈ C([0. 1]) T
_
1
0
p(r))
n
(r)dr → p(0).T Thus
wk
∗
lim)
n
dr = δ(r)dr in `([0. 1]) but )
n
has no weak limit in 1
1
. Of course, the sequence
)
n
has a weak
∗
convergent subsequence in 1
∞
([0. 1])
t
, which shows the existence of a linear
functional on 1
∞
that restricts to p → p(0) for continuous functions p. (We could have used
the Hahn Banach theorem to get this.)
Here is another example. On 1
1
([0. ∞)) let
)
n
(r) =
1
:
χ
[0,n]
(r).
Again )
n

L
1
= 1. As measures )
n
dr ÷ 0 in `
0
([0. ∞)), that is
_
∞
0
)
n
(r)p(r)dr −→ 0 p ∈ C
0
([0. ∞)).
however, )
n
does not converge weakly to zero in 1
∞
. Indeed for the constant function p ≡ 1,
_
∞
0
)
n
(r)p(r) = 1.
LECTURE 21
An application: positive harmonic functions
Reading: ¸11.6 in Lax
Positive harmonic functions
We may apply weak
∗
compactness to prove the following:
Theorem 21.1 (Herglotz). Let n be a function on the open unit disk 1 = ¦[.[ < 1¦ such
that
(1) n(.) ≥ 0 for all . ∈ 1
(2) n is harmonic in 1, that is it satisﬁes the mean value property
n(.) =
1
πε
2
_
Dε(z)
n(u)d:(u).
for all . ∈ 1 and ε < 1 − [.[. Here : is Lebesgue measure on the disk and 1
ε
(.)
is the disk of radius ε centered at ..
Then there is a unique ﬁnite, nonnegative, Borel measure j on ∂1 = ¦[.[ = 1¦ such that
n(.) =
_
∂D
1 −[.[
2
[. −u[
2
dj(u). ()
Conversely, any such function is a nonnegative Harmonic function on the disk.
Remark. The theorem implies [n(.)[ ≤ const.,(1 − [.[), so nonnegative harmonic
functions cannot blow up arbitrarily at ∂1. One might wonder if a similar theorem holds
for, say, real valued Harmonic functions. That is, given n Harmonic and real valued does
there exist a signed measure j such that n is the Poisson integral of j? A moment’s thought
shows that the answer is “No!”, for it is easy to construct a real valued harmonic function
which violates the estimate [n(.)[ ≤ const.,(1 −[.[). For example, n(.) = Re 1,(1 −.)
2
.
Proof. Note that n is continuous. To see this, ﬁrst observe that the mean value property
implies that n is locally integrable. Next, observe that
n(. +/) −n(.) =
1
πε
2
__
Dε(z+h)
n(u)d:(u) −
_
Dε(z)
n(u)d:(u)
_
.
By dominated convergence the integral on the r.h.s. converges to zero as / converges to zero.
Since n is continuous, we may diﬀerentiate
π:
2
n(0) =
_
Dr(0)
n(.)d:(.) =
_
r
0
_
2π
0
n(:e
iθ
)dθ:d:
with respect : and conclude that for every :
n(0) =
1
2π
_
2π
0
n(:e
iθ
)dθ.
211
212 21. AN APPLICATION: POSITIVE HARMONIC FUNCTIONS
Thus for each : ∈ (0. 1) the measure dj
r
(θ) = (2π)
−1
n(:e
iθ
)dθ on the circle ∂1 has mass
n(0). Think of these measures as elements of the dual to C(∂1). By Helly’s theorem, we
may ﬁnd a weak
∗
convergent subsequence j
rn
. That is, there is a Borel measure j ∈ C(∂1)
t
such that
_
∂D
)(θ)dj(θ) = lim
n→∞
1
2π
_
∂D
)(θ)n(:e
iθ
)dθ
for every ) ∈ C(∂1).
To complete the proof, we will use the identity
n(.) =
1
2π
_
2π
0
:
2
−[.[
2
[:e
iθ
−.[
2
n(:e
iθ
)dθ. ()
valid for . ∈ 1
r
(0). Let us defer the proof for the moment and show how () implies the
representation (). Fix . and let
)
r,z
(e
iθ
) =
:
2
−[.[
2
[:e
iθ
−.[
2
.
It is easy to see that )
r,z
→ )
z
uniformly as : → 1, where
)
z
(e
iθ
) =
1 −[.[
2
[e
iθ
−.[
2
.
Thus the weak
∗
convergence j
rn
→
_
j then implies
n(.) = lim
n
_
∂D
)
rn;z
(e
iθ
)dj
rn
(θ) =
_
2π
0
)
z
(e
iθ
)dj(θ).
The identity () is a classical formula, which may be veriﬁed in a number of ways. One
of these is as follows. Let ·(.) denote the integral on the r.h.s. It is easy to show that · is
harmonic in 1
r
(0) — for this it suﬃces to show that ([u[
2
− [.[
2
),[u − .[
2
is harmonic in
1
[w[
(0) for ﬁxed u. Furthermore, it is not too hard to show that
lim
s↑r
·(:e
iθ
) = n(:e
iθ
).
since for any continuous function ) on the circle
1
2π
lim
s↑r
_
2π
0
:
2
−:
2
[:e
iθ
−:e
iφ
[
2
)(e
iφ
)dφ = )(e
iθ
).
(Exercise: verify this formula.) Thus n(.)−·(.) is a harmonic function on 1
rn
(0), continuous
up to the boundary and identically equal to zero there. It follows from the maximum
principle, applied to n −· and · −n, that n −· = 0 throughout. (The maximum principle
is a straightforward consequence of the mean value property and continuity.)
HerglotzRiesz Theorem
An important application of the above is the following:
Theorem 21.2 (HerglotzRiesz). Let 1 be an analytic function in the unit disk 1 such
that Re 1 ≥ 0 in 1. Then there is a unique nonnegative, ﬁnite, Borel measure j on ∂1
such that
1(.) =
_
2π
0
e
iθ
+.
e
iθ
−.
j(dθ) + i Im1(0).
HERGLOTZRIESZ THEOREM 213
Conversely every analytic function in the disk with positive real part can be written in this
form.
Proof. First apply the Herglotz theorem to Re 1. Let
G(.) =
_
2π
0
e
iθ
+.
e
iθ
−.
j(dθ).
So G and 1 are analytic functions on the disk whose real parts agree. It follows that 1 −G
is constant and imaginary. However G(0) = Re 1(0) so 1(.) −G(.) = i Im1(0).
The theorem is often used in the following form
Theorem 21.3. Let 1 be an analytic map from the upper half plane ¦. : Im. 0¦
into itself. Then there is a unique nonnegative Borel measure j on R and a nonnegative
number ¹ ≥ 0 such that
_
R
1
1 +r
2
dj(r) < ∞
and
1(.) = ¹. + Re 1(i) +
_
R
1 +r.
r −.
1
1 +r
2
dj(r). (  )
Furthermore
¹ = lim
z→∞
1(.)
.
.
and
dj(r) = wk
∗
lim
y↓0
1
π
Im1(r + iµ)dr.
If lim
z→∞
(1(.) −¹.) = 1 exists and is real, and if lim
z→∞
.(1(.) −¹. −1) exists then j
is a ﬁnite measure and
1(.) = ¹. +1 +
_
R
1
r −.
dj(r).
Remark. Note that
1 +r.
r −.
1
1 +r
2
=
1
r −.
−Re
1
r −i
.
Proof. Consider the function
G(ζ) = −i1
_
i
1 −ζ
ζ + 1
_
.
This is an analytic map from the disk into the right half plane. By the HerglotzRiesz
theorem
1
_
i
1 −ζ
ζ + 1
_
= i
_
2π
0
e
iθ
+ζ
e
iθ
−ζ
dν(θ) + Re 1(i)
Now let . = i(1 −ζ),(ζ + 1), so ζ = (1 + i.),(1 −i.) and
1(.) = i
_
2π
0
e
iθ
(1 −i.) + 1 + i.
e
iθ
(1 −i.) −1 −i.
dν(θ) + Re 1(i).
Now we deﬁne a map φ : ∂1 ¸ ¦−1¦ →R via
φ(e
iθ
) = i
1 −e
iθ
1 + e
iθ
.
214 21. AN APPLICATION: POSITIVE HARMONIC FUNCTIONS
and let ¯ j = φ;ν, that is
_
)d¯ j =
_
) ◦ φdν
for functions ) ∈ C
0
(R). Now given p ∈ C(∂1), p − p(−1)1 vanishes at −1 and may be
written as
p −p(−1)1 = ) ◦ φ
with ) = (p −p(−1)1) ◦ φ
−1
. Thus
_
∂D
pdν =
_
(p −p(−1)1) ◦ φ
−1
d¯ j +p(−1)ν(∂1).
Since ν(∂1) = Im1(i), we conclude that
1(.) = Re 1(i) +. Im1(i) +
_
R
_
i
(1 + ir)(1 −i.) + (1 −ir)(1 + i.)
(1 + ir)(1 −i.) −(1 −ir)(1 + i.)
−.
_
d¯ j(r).
since φ
−1
(r) = (1 + ir),(1 −ir). After simplifying, this gives
1(.) = ¹. + Re 1(i) +
_
R
1 +r.
r −.
d¯ j(r).
with ¹ = Im1(i) − ¯ j(R). The representation (  ) follows with dj(r) = (1 +r
2
)d¯ j(r).
The identity
¹ = lim
z→∞
1(.)
.
holds since
1
.
_
R
1 +r.
r −.
d¯ j(r) −→ 0.
Furthermore, if lim
z→∞
.(1(.) −¹. −1) exists for some real number 1 then in particular
lim
t→∞
t(Im1(it) −i¹t) = lim
t→∞
t
_
R
Im
1 + itr
r −it
d¯ j(r)
exists and is ﬁnite. The integrand on the r.h.s. is
t
2
r
2
+t
2
+t
2
r
2
t
2
+r
2
=
t
2
r
2
+t
2
(1 +r
2
)
converges pointwise, monotonically to (1 +r
2
). Thus j is a ﬁnite measure, and
1(.) = ¹. + Re 1(i) +
_
R
1
r −.
dj(r) −Re
_
R
1
r −i
dj(r).
One checks now that
lim
z→∞
(1(.) −¹.) = Re 1(i) −Re
_
R
1
r −i
dj(r).
Presentation topics
Possible presentation topics
Please choose a presentation topic and discuss it with me by March 31. Presentations
will be 20 minutes in length and given in class April 16, 18, 21, 23 and 25.
(1) ¸9.1 Completeness of weighted powers in C
0
(R).
(2) ¸9.2 M¨ untz Approximation Theorem.
(3) ¸11.2 Divergence of Fourier Series.
(4) ¸11.3 Approximate quadrature.
(5) ¸11.5 Existence of solutions to P.D.E.’s
(6) ¸14.3 Completely monotone functions
(7) ¸14.7 Theorems of Carath´eodory and Bochner
(8) ¸16.3.2 Hilbert Transform
(9) ¸16.3.3 Laplace Transform
(10) ¸16.4 Solution operators for hyperbolic equations
(11) ¸16.5 Solution operator for the heat equation
(12) ¸22.4 Operators deﬁned by parabolic equations
Homework II
Homework II
Due: March 31, 2008
(1) (Ex. 1, Ch.6) Show that a norm that satisﬁes the parallelogram law comes from an
inner product.
(2) (Ex. 3, Ch. 6) Show that /
2
is complete.
(3) Let A be a reﬂexive Banach space and 1 ⊂ A a closed subspace of A. Show that
(1
⊥
)
⊥
= 1 .
(4) (Ex. 5, Ch. 8) Let A. j be a measure space with j a positive measure. Show that
if j(A) = 1 then )
p
is an increasing function of j
(5) Prove Theorem 11.3 of these notes.
(6) Prove that −∆[r[
2−d
= (d −2)[o
d−1
[δ(r) in the sense of distributions on R
d
, d ≥ 3,
where ∆ = Laplacian, [o
d−1
[ is the area of the unit sphere ¦[r[ = 1 : r ∈ R
d
¦ and
δ(r) is the Dirac delta “function.”
(7) Consider the map 1 deﬁned on Schwarz functions on the real line by the “principle
value integral:”
1(ψ) = P.V.
_
R
1
r
ψ(r)dr = lim
ε↓0
__
∞
ε
1
r
ψ(r)dr +
_
−ε
−∞
1
r
ψ(r)dr.
_
Show that 1 is a tempered distribution. Convolution with 1 is the “Hilbert trans
form.”
(8) (Thm. 20.4 of the notes) Let A
t
be the dual of a Banach space. Show that / → /
is weak
∗
lower semicontinuous and that
/ ≤ liminf
n→∞
/
n
 .
(9) Prove Prop. 20.1 of these notes.
(10) Show that the unit ball in an inﬁnite dimensional Hilbert space contains inﬁnitely
many disjoint translates of a ball of radius
√
2,4. Conclude that there is no non
trivial translation invariant measure on an inﬁnite dimensional Hilbert space.
(11) A subset o of a Banach space is weakly bounded if for all / ∈ A
t
, sup
x∈S
[/(r)[ <
∞. Prove that o is weakly bounded if and only if it is strongly bounded, i.e.,
sup
x∈S
r < ∞.
212
(12) Let A be a locally convex space.
(a) Let l be a compact neighborhood of 0. Show that one can ﬁnd r
1
. . . . . r
n
so
that l ⊂ ∪
n
i=1
(r
i
+
1
2
l), and thus a ﬁnite dimensional linear space 1 ⊂ A with
l ⊂ 1 +
1
2
l.
(b) Prove that l ⊂ 1 + (
1
2
)
m
l for any :.
(c) Prove that l ⊂ 1 = 1 .
(d) Conclude that A = 1 = 1 is ﬁnite dimensional.
Thus, any locally compact, locally convex, space is ﬁnite dimensional.
(13) Let A be a reﬂexive Banach space. Is the open unit ball in A open in the weak
topology?
Part 6
Convexity
LECTURE 22
Convex sets in a Banach space
Reading: ¸8.4 and Ch. 12 of Lax
Definition 22.1. The support function o
M
: A
t
→R of a subset ` of a Banach space
A over R is the function
o
M
(/) = sup
y∈M
/(µ).
The support function o
M
of a set ` has the following properties:
(1) Subadditivity, o
M
(/ +:) ≤ o
M
(/) +o
M
(:).
(2) o
M
(0) = 0.
(3) Positive homogeneity, o
M
(c/) = co
M
(/) for c 0.
(4) Monotonicity: for ` ⊂ `, o
M
(/) ≤ o
N
(/).
(5) Additivity: o
M+N
= o
M
+o
N
. (Recall that ` +` = ¦r +µ[r ∈ ` and µ ∈ `¦.)
(6) o
−M
(/) = o
M
(−/)
(7) o
M
= o
M
.
The proof of these is left as an exercise.
In addition, we have
(9) o
˘
M
= o
M
,
where
Definition 22.2. The closed convex hull of a subset ` of a Banach space, denoted
˘
`
is the smallest closed convex set containing `.
Remark.
˘
` is also the closure of the convex hull
ˆ
`, where the convex hull is the
smallest convex set containing `. (Exercise)
Let us prove property (9), assuming the other properties. First by (8) it suﬃces to show
o
ˆ
M
= o
M
. Since ` ⊂
ˆ
`, by (5) we have o
M
≤ o
ˆ
M
. However,
ˆ
` = ¦
n
j=1
c
j
r
j
: r
j
∈ `. c
j
0. and
j
c
j
= 1¦.
(Exercise: prove that this is in fact the smallest convex set containing `.) So for any point
in
ˆ
` we have
/(
n
j=1
c
j
r
j
) =
n
j=1
c
j
/(r
j
) ≤ o
M
(/).
Thus o
ˆ
M
≤ o
M
.
Here are some examples:
i. If ` = ¦r
0
¦, o
M
is just evaluation at r
0
.
ii. If ` = 1
R
(0) then o
M
(/) = 1/ .
221
222 22. CONVEX SETS IN A BANACH SPACE
iii. If ` = 1
R
(r
0
) then ` = ¦r
0
¦ +1
R
(0) so o
M
(/) = 1/ +/(r
0
).
iv. If ` is a closed subspace then o
M
(/) = 0 for / ∈ `
⊥
and ∞ for all other /.
Note that in the last example the set ` is unbounded. For bounded sets, o
M
: A
t
→R,
however in general we deﬁne o
M
as a map from A
t
→R∪¦∞¦. We extend the order relation
and arithmetic to R ∪ ¦∞¦ by r ≤ ∞ and r + ∞ = ∞ for all r and c∞ = ∞ for c 0.
This extended function satisﬁes all of the above properties.
If ` is bounded, then o
M
(/) ≤ const. / and is therefore continuous in the norm
topology, since by subadditivity
[o
M
(/) −o
M
(/
t
)[ ≤ max¦o
M
(/ −/
t
). o
M
(/
t
−/)¦ ≤ const. / −/
t
 .
This fails if ` is unbounded, and also in the weak
∗
topology. Nonetheless, we have the
following additional property
(10) o
M
is weak
∗
lower semicontinuous
Indeed, since it is a sup of weak
∗
continuous functions, we have
¦/ : o
K
(/) t¦ =
_
z∈K
¦/ : /(.) t¦ .
(Weak
∗
continuity of a R∪¦∞¦ valued function is deﬁned in the same way as for a R valued
function.)
Theorem 22.1. The closed convex hull
˘
` of a subset ` of a Banach space A over R
is equal to
˘
` = ¦. : /(.) ≤ o
M
(/) for all /¦.
Proof. Since o
M
= o
˘
M
it follows that /(.) ≤ o
M
(/) for all . ∈
˘
`.
Now, suppose . ,∈
˘
`. Since
˘
` is closed there is an open ball 1
R
(.) with 1
R
(.)∩
˘
` = ∅.
By the geometric HahnBanach theorem we can ﬁnd a linear functional /
0
and c ∈ R such
that
/
0
(n) ≤ c < /
0
(·) for all n ∈
˘
` and · ∈ 1
R
(.).
In particular, if r ≤ 1 then
/
0
(−r) = −/
0
(r +.) +/
0
(.) ≤ /(.) −c .
so /
0
 ≤ 1
−1
(/
0
(.) − c). Thus /
0
is bounded. From the deﬁnition of the norm of a linear
functional, we have
inf
x<R
/
0
(r) = −1/
0
 .
Since for . +r ∈ 1
R
(.) we have c ≤ /
0
(. +r), we ﬁnd that
/
0
(.) ≥ c +1/
0
 .
It follows that
/
0
(.) ≥ o
M
(/
0
) +1/
0
 .
Thus /
0
is a linear functional such that /
0
(r) o
M
(/
0
).
The theorem shows that a closed, convex set 1 is speciﬁed as the set
1 = ¦. : Φ
K
(.) ≤ 0¦
where
Φ
K
(.) = sup
:≤1
[/(.) −o
K
(/)].
22. CONVEX SETS IN A BANACH SPACE 223
Since o
K
: A
t
→ R ∪ ¦∞¦, the function Φ
K
is initially deﬁned as a map A → R ∪ ¦−∞¦.
However, note that Φ(.) = −∞ for some . if and only if o
K
(/) = ∞ for all /, in which case
Φ(.) = −∞ for all . and 1 = A. For any proper closed, convex set 1 there is some / such
that o
K
(/) < ∞ and Φ
K
: A →R.
Since o
K
is weak
∗
lower semicontinuous, it follows that, for ﬁxed ., /(.)−o
K
(/) is weak
∗
upper semicontinuous, that is for each t
¦/ : /(.) −o
K
(/) < t¦
is weak
∗
open. This observation is relevant, since ¦/ ≤ 1¦ is compact, and
Proposition 22.2. Let 1 be a compact topological space and let 1 : 1 →R∪¦−∞¦ be
upper semicontinuous. Then 1 is bounded from above and attains it’s maximum.
Remark. We did this for lower semicontinuous functions, but it didn’t make it into the
notes, so let’s prove it again.
Proof. The sets ¦1(r) < t¦ are increasing, open, and cover 1. By compactness 1 ⊂
¦1(r) < t¦ for some t. So 1 is bounded from above. Now let t
m
= sup
x∈K
1(r). Suppose
1(r) < t
m
for all r. Then the sets ¦1(r) < t¦ for t < t
m
cover A. By compactness there is
then some t < t
m
such that 1 ⊂ ¦1(r) < t¦, contradicting t
m
= sup
x∈K
1(r). So there is a
point r
m
such that t
m
= 1(r
m
).
It follows that,
Φ
K
(.) = max
≤1
[/(.) −o
K
(/)].
The function Φ
K
(.), being a sup of weakly continuous functions, is in turn weakly lower
semicontinuous. In particular,
1 = ¦. : Φ
K
(.) ≤ 0¦
is weakly closed! Thus we have obtained the following theorem:
Theorem 22.3 (Theorem 2, ¸12 of Lax). A convex set 1 of a Banach space is closed in
the norm topology if and only if it is closed in the weak topology.
This theorem is astounding, since there are certainly strongly closed sets that are not
weakly closed. (Weakly closed =⇒ strongly closed for any set.) For instance the complement
of an open ball ¦r : r ≥ 1¦ is norm closed but not weakly closed. (Exercise: prove this.)
LECTURE 23
Convex sets in a Banach space (II)
Reading: ¸8.4
Recall that we showed last time that a closed convex set 1 in a Banach space is equal to
1 = ¦r : Φ
K
(r) ≤ 0¦ .
where
Φ
K
(r) = max
≤1
[/(r) −o
K
(/)]. o
K
(/) = sup
z∈K
/(.).
The weak lower semicontinuity of Φ
K
showed that 1 is weakly closed. As a corollary we
have
Corollary 23.1. If A is reﬂexive, then a bounded, norm closed, convex subset 1 is
weakly compact.
Remark. This may fail if A is not reﬂexive. For instance in 1
1
(R) the set 1 of non
negative functions ρ ≥ 0 with integral
_
ρ = 1 is convex, norm closed and bounded. However,
it is not weakly compact.
Similarly, this may fail if A is a dual and we take the weak
∗
topology on A. For instance,
inside `
0
(R) – the space of ﬁnite measures on R – the space of probability measures is norm
closed, bounded, and convex but is not weak
∗
closed.
All of the suggests that Φ
K
(r) might be a decent measure of how far a point r is from
1. In fact, it is precisely the distance of r to 1!
Theorem 23.2. Let 1 be a closed, convex subset of a Banach space A. Then
Φ
K
(r) = inf
u∈K
r −n .
Proof. Suppose r ∈ 1. Then /(r) −o
K
(/) ≤ 0 for all / so the maximum is attained at
/ = 0 and Φ
K
(r) = 0.
If r ,∈ 1 and n ∈ 1 and / ≤ 1, then
o
K
(/) ≥ /(n) = /(n −r) +/(r) ≥ /(r) −n −r .
Thus
n −r ≥ sup
≤1
[/(r) −o
K
(/)].
On the other hand, if 1 < inf
u∈K
n −r, then the convex set 1+1
R
(0) still has positive
distance from r. Thus by the Theorem of last lecture there is /
0
∈ A
t
such that
o
K
(/
0
) +1/
0
 = o
K+B
R
(0)
(/
0
) < /
0
(r).
This inequality is homogeneous under positive scaling, so we may take /
0
 = 1 to conclude
1 < /
0
(r) −o
K
(/
0
) ≤ sup
≤1
[/(r) −o
K
(/)].
As 1 was any number less than inf
u∈K
n −r the reverse inequality follows.
231
232 23. CONVEX SETS IN A BANACH SPACE (II)
Let us turn all of this around. Suppose we are given a function o : A
t
→ R ∪ ¦∞¦.
Consider the set
1 = ¦. ∈ A : /(.) ≤ o(/) for all / ∈ A
t
¦.
Then 1 is clearly convex and weakly closed, and it’s support function
o
K
(/) = sup
x∈K
/(.) ≤ o(/).
Can it happen that the inequality is strict? Of course it can, as the function o is arbitrary.
However, if we assume that o is, like o
K
, positive homogeneous, subadditive, maps 0 to 0
and is weak
∗
lower semicontinuous then the answer is “no,” at least if A is reﬂexive.
Theorem 23.3. Let A be a reﬂexive Banach space and let o : A
t
→R∪¦∞¦ be a weak
∗
lower semicontinuous function which is positive homogeneous, subadditive, and maps 0 to
0. Then
o(/) = sup
x∈K
/(r).
where
1 = ¦r : /(r) ≤ o(/) for all /¦
.
Proof. To begin, let us prove the theorem under the additional restriction that o is
bounded: [o(/)[ ≤ β /. Then 1 is clearly bounded, since r ∈ 1 =⇒ r ≤ β. Since
o(0) = 0 the identity holds for / = 0. So ﬁx a nonzero linear functional /
0
. Let us deﬁne a
linear functional 1 ∈ A
tt
, the double dual, via HahnBanach. Begin on the onedimensional
subspace span¦/
0
¦ and let
1(t/
0
) = to(/
0
).
By positive homogeneity and subadditivity 1(t/
0
) ≤ o(t/
0
) for all t ∈ R. (Note that
o(−/
0
) ≥ −o(/
0
).) Thus by HahnBanach there is a linear functional 1 on A
t
which satisﬁes
1(/) ≤ o(/)
for every / ∈ A
t
. Since o(/) ≤ β / this functional is bounded. As A is reﬂexive there is
. ∈ A such that 1(/) = /(.). Since /(.) ≤ o(/) for all / this point . ∈ 1, and since
o(/
0
) = 1(/
0
) = /
0
(.).
we have
o(/
0
) = max
x∈K
/
0
(r).
As /
0
was arbitrary, this completes the proof for o bounded.
To extend this to unbounded o, note that
−∞ < inf
≤1
o(/) ≤ 0.
by wk
∗
lower semicontinuity. Let −β = inf
≤1
o(/). Then by positive homogeneity,
o(/) ≥ −β / .
that is o is bounded from below.
Now, for each c deﬁne
1
= ¦r : /(r) ≤ o
(/) for all /¦ .
23. CONVEX SETS IN A BANACH SPACE (II) 233
where
o
(/) = inf
1
,
2
∈X
:
1
+
2
=
_
o(/
1
) +
1
c
/
2

_
.
It is left as an exercise to show, for each c 0, that o
is positive homogeneous, subadditive
and maps 0 to 0. Note also that
−β / ≤ o
(/) ≤
1
c
/ . and o
(/) ≤ o(/).
so o
is bounded and smaller than o.
Now, in fact,
1
= 1 ∩ 11
(0).
Indeed, given r ∈ 1 ∩ 1
1/
(0) we have
/(r) = /
1
(r) +/
2
(r) ≤ o(/
1
) +
1
c
/
2

if / = /
1
+ /
2
, so /(r) ≤ o
(/) and r ∈ 1
. On the other hand if r ∈ 1
then /(r) ≤ o(/)
and
1
/ for all / so r ∈ 1 ∩ 1
1/
(0).
It follows that
1 = ∪
1
..
so
sup
x∈K
/(r) = sup
sup
x∈K
/(r) = sup
o
(/).
Thus, it suﬃces to show, for ﬁxed /, that
o(/) = sup
o
(/).
To show this, note that o
increases as c decreases, so
o
0
(/) := lim
→0
o
(/) = sup
>0
o
(/)
exists and (since o
≤ o) satisﬁes
0 ≤ o
0
(/) ≤ o(/).
Furthermore, for each c we can ﬁnd /
such that
o
(/) ≤ o(/ −/
) +
1
c
/
 ≤ o
(/) +c.
Since o(/ −/
) ≥ −β / −β /
 we see that
−β / +
_
1
c
−β
_
/
 ≤ o
(/) +c.
Consider the following cases: (1) /
 ,c is bounded as c → 0 or (2) /
 ,c is unbounded as
c → 0 . In case (2) lim
→0
o
(/) = o
0
(/) = o(/) = ∞. On the other hand, in case (1) /
→ 0
so by weak
∗
lower semicontinuity of o we ﬁnd that
o(/) ≤ liminf
→0
o(/ −/
) ≤ o
0
(/) −liminf
→0
1
c
/
 ≤ o
0
(/).
which completes the proof.
LECTURE 24
KreinMilman and StoneWeierstrass
Reading: ¸13.3
Definition 24.1. An extreme subset o of a convex set 1 is a subset o ⊂ 1 such that
(1) o is nonempty and convex
(2) If r ∈ o and r = tµ + (1 −t). with µ. . ∈ 1 then µ. . ∈ o.
An extreme point is a point r ∈ 1 such that ¦r¦ is an extreme subset.
The following is a classical result due to Carath´eodory:
Theorem 24.1. Every compact convex subset 1 of R
N
has extreme points, and every
point of 1 can be written as a convex combination of (at most) ` + 1 extreme points.
Remark. The proof is left as an exercise. Use induction on `. The case ` = 1 is easy!
Theorem 24.2 (Krein and Milman). Let A be a locally convex space, 1 a nonempty,
compact, convex subset of A. Then
(1) 1 has at least one extreme point
(2) 1 is the closure of the convex hull of its extreme points.
Proof. Consider the collection c of all nonempty closed extreme subsets of 1. Since
1 ∈ c this collection is nonempty. Partially order c by inclusion. We wish to apply Zorn’s
lemma to see that c has a “maximal” element, i.e., a set that is minimal with respect to
conclusion.
Let T ⊂ c be a totally ordered subcollection. That is T = ¦1
ω
: ω ∈ Ω¦ with Ω some
totally ordered index set and 1
α
⊂ 1
β
if α ≥ β. Clearly ∩T is a candidate for an “upper
bound.” To see this we must show that ∩T is nonempty, closed, and extreme.
Clearly ∩T is closed. Furthermore,one easily shows that the intersection of an arbitrary
family of extreme sets is extreme provided it is nonempty. Thus we need only show T is
nonempty. Here we use compactness of 1 in a crucial way.
Recall that a family T of closed sets is said to have the ﬁnite intersection property (FIP)
if any ﬁnite collection 1
1
. . . . 1
n
∈ T has nonempty intersection: 1
1
∩ ∩ 1
n
,= ∅, and
there is the following result connecting the FIP and compactness:
Theorem 24.3. A topological space ` is compact if and only if every collection T of
closed sets with the FIP satisﬁes
∩T ,= ∅.
Proof. Suppose ` is compact and ∩T = ∅. Then ∪ = ` with / = ¦1
c
: 1 ∈ T¦.
Thus ` = ∪
n
j=1
1
c
j
for some ﬁnite collection. Thus ∩
n
j=1
1
j
= ∅ and T does not have the FIP.
Conversely, if T is a collection of closed sets with the FIP and nonetheless ∩T = ∅, then
 = ¦1
c
: 1 ∈ ¦¦ is an open cover of ` with no ﬁnite subcover so ` is not compact.
241
242 24. KREINMILMAN AND STONEWEIERSTRASS
Regarding our collection T , it clearly has the FIP since it is totally ordered, so any ﬁnite
collection 1
1
. . . . 1
n
∈ T has a minimal element. Thus ∩T , = ∅.
Sp c has minimal elements. We claim that any such minimal element is a one point set.
Indeed, suppose 1 ⊂ c contains two distinct points (it must also contain the line segment
joining them). Then there is a continuous linear functional / on A that separates these
points. Let ` ⊂ 1 be the set
` = ¦r ∈ 1 : /(r) = max
z∈E
/(r)¦.
Then ` is a nonempty, proper, closed subset of 1. It is clearly convex and one easily shows
it is an extreme subset of 1. It follows that ` is an extreme subset of 1 (why?) and ` 1
so 1 is not minimal.
This proves (1): 1 has at least one extreme point. (It might have only one: 1 could be
the set ¦x¦.) Since any closed extreme subset 1 of 1 is itself a closed convex set we ﬁnd
that every extreme subset of 1 has an extreme point r. It is easy to see that an extreme
point of 1 is also an extreme point of 1 (since 1 is an extreme subset). Thus
Every closed, extreme subset of 1 contains an extreme point of 1.
Let 1
e
denote the set of extreme points,
´
1
e
its convex hull, and
˘
1
e
its closed convex
hull which is the closure of
´
1
e
. One easily shows that
˘
1
e
is convex. (Exercise)
Clearly
´
1
e
⊂ 1, so since 1 is closed
˘
1
e
⊂ 1. On the other hand if . ,∈
˘
1
e
then
there is an open set l with . ∈ l ⊂ 1
e
. We may take l to be convex. By the geometric
HahnBanach there is a linear functional / and c ∈ R such that
/(r) < c ≤ /(µ) for all r ∈ l and µ ∈
˘
1
e
.
(As l is open, all points of l are interior, so the ﬁrst inequality is strict.) The gauge function
of l −.,
j
U−z
(r) = inf¦t : r,t ∈ l −.¦.
is a continuous seminorm on A (why?). If r,t ∈ l −. we have
1
t
/(r) = /(
r
t
+.) −/(.) < c −/(.).
Thus
/(r) ≤ (c −/(.))j
U−z
(r).
and / is a continuous linear functional because
Lemma 24.4. A linear functional on locally convex space is continuous if and only if it
is bounded with respect to some continuous seminorm.
Proof. Exercise.
Since 1 is compact / achieves its minimum on 1. Let 1 be the set of minimizers. Then
1 is closed, convex and extreme. (Why?) By the above derived result 1 contains an extreme
point. Thus
min
x∈K
/(r) = min
x∈Ke
/(r) /(.).
Thus . ,∈ 1.
An interesting application of this theorem is:
24. KREINMILMAN AND STONEWEIERSTRASS 243
Theorem 24.5 (StoneWierstrass). Let o be a compact Hausdorﬀ space and C(o) the
set of real valued continuous functions on o. Let 1 ⊂ C(o) be an subalgebra, that is 1 is
linear subspace and ). p ∈ 1 =⇒ )p ∈ 1. If the constant function 1 ∈ 1 and if for any
pair of points j. ¡ ∈ o there is a function ) ∈ 1 with )(j) ,= )(¡), then 1 is dense in C(o)
in the max norm.
This theorem is due to Stone and generalizes the classical result of Weierstrass on ap
proximation of continuous functions on an interval with polynomials: o = [0. 1] and 1 =
polynomials.
Sketch of proof. Consider the collection ^ of all ﬁnite measures j on o such that
_
)dj = 0 for all ) ∈ 1. Then o is dense if and only if / = ¦0¦. (Why?)
By construction ^ is weak
∗
closed. So 1 = ^ ∩ 1
1
(0) is weak
∗
compact. Furthermore
it is convex. Suppose 1 contains a nonzero measure j. Then 1 must contain a nonzero
extreme point j. Since j is extreme we must have j = 1 (otherwise we could write j as
a linear combination of 0 and some multiple of j).
Suppose such a j exists. Since 1 is an algebra
_
)pdj = 0 for all ). p ∈ 1. Thus
pdj ∈ ^ for all p ∈ 1. Now let p be a continuous function on o with 0 < p(j) < 1 for all
j ∈ o. Let
c =
_
pd [j[ . / =
_
(1 −p)d [j[ .
So c. / 0 and c +/ = 1 and pdj,c, (1 −p)dj,/ ∈ 1. Since
dj = c
p
c
dj +/
(1 −p)
/
dj
we must have pdj,c = dj (recall that j is an extreme point).
Consider the support of j:
supp j = ¦j : [j[(l) 0 for any open neighborhood of j¦.
Since dj = pdj,c for any p ∈ 1 that satisﬁes 0 < p(j) < 1 we must have p ≡ c on supp j.
(Why?) Suppose j and ¡ are distinct points in o. Then there is a function p ∈ 1 such that
0 < p < 1 and p(j) ,= p(¡). (Just add a large constant to a function in 1 that separates j
and ¡). Thus at most one of the points j. ¡ lies in the support of j. That is the support of
j is a single point supp j¦j
0
¦! Since [j[(1) = j = 1 we have
_
)(j)dj(j) = )(j
0
) or
_
)(j)dj(j) = −)(j
0
).
However, we have
_
1dj = 0 which is a contradiction, so ^ = ¦0¦ and 1 is dense.
Following the above proof, we ﬁnd
Theorem 24.6. The extreme points of the unit ball ¦j :
_
d[j[ ≤ 1¦ ⊂ /(o) are the
point masses ±δ(j −j
0
).
Theorem 24.7. If ¹ ⊂ C(o) is a proper closed subalgebra that separates points of o
then ¹ = ¦) : )(j
0
) = 0¦ for some j
0
∈ o.
The algebras ¹
p
= ¦) : )(j) = 0¦ are exactly the maximal ideals of C(o). It is no
accident that the set of maximal ideals is the set of proper subalgebras that separate points
and is in one to one correspondence with extreme points of the unit ball in C(o)
t
which is
in one to one correspondence with o.
LECTURE 25
Choquet type theorems
Reading: ¸13.4 and 14.10
In ﬁnite dimensions, Caratheodory’s Theorem shows that points of a compact set may be
expressed as a convex combination of extreme points: no more than ` +1 points are needed
in dimension `. (A simplex shows that this number is optimal.) The following generalizes
this idea to LCS’s:
Theorem 25.1. Let A be an LCS and 1 a nonempty compact, convex subset of A. For
any n ∈ 1 there is a Borel probability measure j
u
on 1
e
the closure of the set of extreme
points such that
/(n) =
_
Ke
/(r)dj
u
(r) ()
for all / ∈ A
t
.
Remark. The integral in the theorem is understood as the identity
n =
_
Ke
rdj
u
(r).
“in the weak sense.” This expresses n as a generalized convex combination of points of 1
e
.
Lax presents without proof a sharper theorem , due to Choquet, in which the measure j
u
is deﬁned on 1
e
provided 1 is metrizable. Any representation of type () is called a “Choquet
decomposition.” Such representations need not be unique, even in ﬁnite dimensions, for
example if 1 is a disk in R
2
.
Proof. The general idea of the proof is to show that given a function ) on 1
e
we can
“evaluate it” at a pont n ∈ 1 and produce a bounded linear functional: ) → )(n). We then
represent this linear functional as a measure on 1
e
via RieszKakutani.
There are two complications: (1) not every function ) on 1
e
can be extended in a
reasonable, or unique way to n ∈ 1 and (2) RieszKakutani does not apply since 1
e
may
not be compact. The second complication is easily dealt with by replacing 1
e
with it’s
closure 1
e
which, being a closed subset of a compact space (1) is compact.
As for the ﬁrst: any linear functional / ∈ A
t
can be evaluated at n. Furthermore, being
continuous and linear, / achieves its max and min over 1 on the set of extreme points 1
e
.
(We saw this in the proof of KreinMillman.) That is
min
x∈Ke
/(r) ≤ /(n) ≤ max
x∈Ke
/(r) for all n ∈ 1.
It follows that if /
1
(r) = /
2
(r) for r ∈ 1
e
then /
1
= /
2
on 1, so any linear functional
is determined on 1 by it’s restriction to 1
e
. Let 1 ⊂ C(1
e
) be the subspace of maps
) : 1
e
→R with
)(r) = /(r)
251
252 25. CHOQUET TYPE THEOREMS
for some / ∈ A
t
, or )(r) = c for some c ∈ R. Deﬁne a linear functional 1 on 1 by
1()) =
_
/(n) )(r) = /(r)
c )(r) = c.
Then 1 is a positive linear functional on 1 and 1 contains the constant functions. By
Theorem 3.1 of these notes, we may extend 1 as a positive linear functional on the space of
all functions. Restricting this linear functional to C(1
e
), we obtain by RieszKakutani that
1()) =
_
Ke
)(r)dj
u
(r) for all ) ∈ C(1
e
)
for some j
u
. In particular () holds. Since j
u
(1) = 1(1) = 1 j
u
is a probability measure.
The RieszKakutani theorem, which was used in the above proof, is itself an example of
this theorem. Indeed, let 1 ⊂ C(o)
t
be the closed unit ball of the dual of C(o) with o a
compact Hausdorﬀ space. We saw last time that the extreme points 1
e
are point evaluations:
/
p,±1
()) = ±1 )(j)
for some j ∈ o. Thus 1
e
is in one to one correspondence with o ¦−1. +1¦. Exercise:
show that 1
e
= o ¦−1. +1¦ as topological spaces, where 1
e
has the weak
∗
topology and
o ¦−1. +1¦ has the product topology using the discrete topology on ¦−1. +1¦ (all sets –
all 4 of them – are open). The above theorem shows that any linear functional / ∈ 1, with
norm no larger than 1, can be written as a combination
/()) =
_
S¡±1¦
/
p,σ
())d:(j. σ) =
_
S
)(j)d:(j. +) −)(j)d:(j. −1) =
_
S
)(j)dj(j)
with dj(j) = d:(j. +) −d:(j. −1).
Measure preserving maps
Let Ω be a compact metric space and let 1 : Ω → Ω be a homeomorphism. Consider the
set 1 of all probability measures on Ω that are invariant with respect to 1:
1 = ¦j ∈ /(Ω) : j ≥ 0. j(Ω) = 1. and j(o) = j(1(o)) for all measurable o¦ .
We say that j is ergodic under 1 provided 1(o) ⊂ o =⇒ j(o) = 0 or 1.
The following is one of the most important applications of Choquet type theorems.
Theorem 25.2. The set 1 is nonempty, convex and compact in the weak
∗
topology.
The extreme points of 1 (which exist by Krein Millman) are the ergodic measures. Every
invariant measure can be represented uniquely as an average of ergodic measures.
Sketch of proof. Convexity of 1 is easy. Since 1 is clearly a subset of the unit ball
in /(Ω) it is compact once it is closed. Weak
∗
closure follows since
j ∈ 1 ⇐⇒
_
Ω
)(1
−1
(ω))dj(ω) =
_
Ω
)(ω)dj(ω)
for all ) ∈ C(Ω). To see that 1 is nonempty, pick a point ω
0
∈ Ω and consider the sequence
of measures ν
n
given by
_
)(ω)dν
n
(ω) =
1
:
n
j=0
)(1
−j
(ω
0
)).
MEASURE PRESERVING MAPS 253
This is a sequence of probability measures on Ω. Since the unit ball is weak
∗
compact there
is a weak
∗
convergent subsequence. The limit ν of this subsequence satisﬁes ν ≥ 0, ν(Ω) = 1,
and
_
)(1
−1
(ω))dν(ω) = lim
k→∞
1
:
k
n
k
j=0
)(1
−j−1
(ω
0
))
= lim
k→∞
1
:
k
n
k
j=0
)(1
−j
(ω
0
) −
1
:
k
)(ω
0
) +
1
:
k
)(1
−n
k
−1
(ω
0
)) =
_
)(ω)dν(ω).
Suppose j ∈ 1 is not ergodic. Then there must be a disjoint union Ω
1
∪ Ω
2
= Ω with
1(Ω
j
) ⊂ Ω
j
and j(Ω
j
) 0. Let j
j
be the restriction of j to Ω
j
renormalized to be a
probability measure:
j
j
(o) =
j(o ∩ Ω
j
)
j(Ω
j
)
.
Then each j
j
is invariant under 1 and
j = j(Ω
1
)j
1
+j(Ω
2
)j
2
so j is not extreme.
Conversely, suppose j is not extreme, so
j = c:
1
+ (1 −c):
2
with :
1
,= :
2
∈ 1 and 0 < c < 1. Suppose :
2
<< :
1
. Then by RadonNikodym there is
) ∈ 1
1
(:
1
), ) ,≡ 1, such that
dj = (c + (1 −c)))d:
1
.
Since j and :
1
are invariant under 1 we must have )(1(ω)) = )(ω) for :
1
almost every ω.
Since ) ,≡ 1 the sets Ω
1
= ¦ω : )(ω) ≤ 1¦ and Ω
2
= ¦ω : )(ω) 1¦ are separately invariant
under 1 and both have positive j measure. So j is not ergodic.
If :
2
,<< :
1
then the situation is even happier. For then there is some set o such that
:
1
(o) = 0 but :
2
(o) 0. Let Ω
1
= ∪
∞
j=0
1
j
(o). Then :
1
(Ω
1
) = 0, :
2
(Ω
1
) ≥ :
2
(o) 0
and Ω
1
is invariant under 1. Since :
1
(Ω
1
) = 0 we must have j(Ω
1
) ≤ 1 − c so j(Ω
2
) 0
with Ω
2
= Ω ¸ Ω
1
. Thus j is not ergodic.
The representation of j as a unique integral over ergodic measures is clearly a Choquet
type representation – we need not take the closure of extreme points since the space of
measures is metrizable in the weak
∗
topology as it is the dual of a separable Banach space
(this was “Choquet’s theorem” which we did not prove). The uniqueness requires a separate
argument, for which Lax refers a paper of Oxtoby in Bull. AMS 58 (1952).
Part 7
Bounded Linear Maps
LECTURE 26
Bounded Linear Maps
Definition 26.1. Let A and 1 be Banach spaces. A linear map ` : A → 1 is bounded
if there is some 0 ≤ c < ∞ such that
`r
Y
≤ c r
X
.
The smallest such c is called the norm, or operator norm, of `, denoted `.
Theorem 26.1. A linear map ` : A → 1 is continuous if and only if it is bounded.
Proof. Bounded implies Lipschitz continuous since
dist(`r. `µ) = `r −`µ ≤ ` r −µ = ` dist(r. µ).
On the other hand if ` is not bounded then there is a sequence r
n
such that
`r
n
 :r
n
 .
As this inequality is invariant under scaling we may take r
n
 = 1,
√
: so r
n
→ 0 but
`r
n
 → ∞.
Theorem 26.2. The operator norm has the following properties
• c` = [c[ ` for all c ∈ 1 (homogeneity)
• ` ≥ 0 and ` = 0 if and only ` ≡ 0 (positivity)
• ` +1 ≤ ` +1 (subadditivity)
That is the operator norm is a norm.
Note that the norm is deﬁned to be
` = sup
x,=0
`r
r
.
By homogeneity this is the same as
` = sup
x=1
`r .
If ` is bounded on a normed space which is incomplete then ` has an extension to the
completion (as does any continuous function) which is also bounded, with the same norm.
Definition 26.2. Let L(A. 1 ) denote the set of all bounded linear maps from A to 1 .
Theorem 26.3. Let A be a normed space and let 1 be a Banach space. Then L(A. 1 )
is a Banach space under the operator norm.
Proof. Since  is a norm, we need to show completeness. Suppose `
n
is a Cauchy se
quence. For each r ∈ A it follows that `
n
r is a Cauchy sequence in 1 , since `
n
r −`
m
r ≤
261
262 26. BOUNDED LINEAR MAPS
`
n
−`
m
 r. By completeness of 1 there is a limit. Call this limit `r. It is easy to see
that ` is linear. Moreover, if r = 1,
`r = lim
n
`
n
r ≤ limsup
n→∞
`
n
 < ∞.
so ` is ﬁnite and ` is bounded.
The operator norm topology on L(A. 1 ) is called the uniform topology. There are two
other natural topologies on this space
Definition 26.3. The strong topology on L(A. 1 ) is the weakest TVS topology such
that all functions ` → `r are continuous from L(A. 1 ) → 1 . The weak topology (or
weak operator topology) on L(A. 1 ) is the weakest TVS topology such that all maps ` →
/(`(r)), for / ∈ 1
t
and r ∈ A, are continuous.
Weak and strong sequential convergence are deﬁned similarly. A sequence `
n
converges
strongly if
: −lim
n
`
n
r exists in 1 for every r.
and converges weakly if
wk −lim
n
`
n
r exists in 1 for every r.
Note that the weak operator topology on L(A. 1 ) is potentially much weaker than the weak
topology on L(A. 1 ) as a Banach space: the linear functionals ` → /(`(r)) are just one
kind of linear functional on L(A. 1 ).
Definition 26.4. The transpose of a linear operator ` ∈ L(A. 1 ) is the map `
t
∈
¸
t
. A
t
deﬁned by
¸r. `
t
/) = ¸`r. /) .
where we use the dual pairing notation ¸µ. /) = /(µ).
Definition 26.5. The null space of a linear operator ` is the set
`
M
= ¦r ∈ A : `r = 0¦.
The range of ` is the set
1
M
= ¦`r : r ∈ A¦.
Theorem 26.4. Let ` ∈ L(A. 1 ) with A and 1 normed spaces. Then
(1) `
t
is bounded and `
t
 = `
(2) The nullspace of `
t
is the annihilator of the range of `:
`
M
= 1
⊥
M
.
(3) The nullspace of ` is the annihilator of the range of `
t
:
`
M
= 1
⊥
M
= ¦r : ¸r. /) = 0 if / ∈ 1
M
¦.
(4) (c` +/`)
t
= c`
t
+/`
t
Proof. Exercise, or see Lax.
As a corollary we see that `
M
and `
M
are weak and weak
∗
closed respectively. Since
these sets are subspaces this is the same as being closed. This is the ﬁrst part of
Theorem 26.5. Let ` ∈ L(A. 1 ) with A and 1 normed spaces. Then
26. BOUNDED LINEAR MAPS 263
(1) `
M
is a closed linear subspace of A.
(2) `, regarded as a map
`
0
:
A
`
M
→ 1
is onetoone, bounded, with `
0
 = ` and 1
M
0
= 1
M
.
Proof. The facts that `
0
is onetoone and 1
M
0
= 1
M
are general facts about linear
maps. To see that `
0
is bounded, recall that A,`
M
is the set of equivalence classes [r] with
r ∼ µ if r −µ ∈ `
M
and we put a norm on this space by
[r] = inf
y∈N
M
r +µ .
Since `
0
[r] = `r = `(r +µ) for any µ ∈ `
M
we have
` = sup
x,=0
`r
r
= sup
x,=0
sup
y∈N
M
`(r +µ)
r +µ
= sup
[x],=0
`
0
[r]
[r]
= `
0
 .
Regarding convergence of adjoints we have the following
Proposition 26.6. If wk −lim`
n
= ` then wk −lim`
t
n
= `
t
.
Proof. Exercise.
However, this does not hold for strong limits or uniform limits. For example, on any /
p
space, 1 < j < ∞, let o
j
be the ,
th
forward shift, the map o
j
∈ L(/
p
. /
p
) given by
o(c
0
. c
1
. . . .) = (0. . . . . 0
. ¸¸ .
j zeroes
. c
0
. c
1
. . . .).
Then o
j
converges to zero weakly, but
o
j
a
p
= a
p
.
so o
j
does not converge to zero strongly. On the other hand the adjoint o
t
j
is the backwards
, shift,
o
t
j
(c
0
. c
1
. . . .) = (c
j
. c
j+1
. . . .).
This map converges to zero strongly since
_
_
o
t
j
a
_
_
p
→ 0
for 1 < j < ∞. (It does not converge to zero strongly for j = 1.)
LECTURE 27
Principle of Uniform Boundedness and Open Mapping Theorem
Reading: ¸15.3–15.5
As for linear functionals, we have the following useful criteria for strong/weak conver
gence:
Proposition 27.1. Let `
n
∈ L(A. 1 ) be a sequence of bounded maps between Banach
spaces A and 1 . Suppose that `
n
are uniformly bounded:
sup
n
`
n
 < ∞.
(1) If `
n
r converges in norm for all r in a dense subset of A then `
n
converges in
the strong operator topology.
(2) If `
n
r converges weakly for all in a dense subset of A then `
n
converges in the
weak operator topology.
Proof. Exercise.
As in the case of linear functionals, boundedness turns out to be necessary for convergence
as well:
Theorem 27.2 (Principle of Uniform Boundedness). Let A and 1 be Banach spaces and
let / ⊂ L(A. 1 ) be a collection of bounded linear maps. If for each r ∈ A and / ∈ 1
t
there
is a constant c(r. /) such that
[¸`r. /)[ ≤ c(r. /) for all ` ∈ /.
then / is uniformly bounded, i.e., there is c < ∞ such that
` ≤ c for all ` ∈ /.
Sketch of proof. This follows very closely the proof of the PUB for linear functionals,
Corollary 19.6 of these notes. First apply that result to conclude that for each r ∈ A there
is c(r) with `r ≤ c(r) for all ` ∈ /. Then consider the collection of real values,
continuous, positive homogeneous functions )(r) = `r on the Banach space A. Apply
Theorem 19.5 to conclude that `r ≤ c r for all ` ∈ /.
The PUB follows from the Baire category theorem. This theorem also implies several
results which show that bounded linear maps have a number of useful properties in addition
to continuity.
Theorem 27.3. Let A and 1 be Banach spaces and ` : A → 1 a bounded linear map
of A onto 1 . Then the image of the unit ball `1
1
(0) contains an open ball around the
origin in 1 .
Corollary 27.4 (Open mapping theorem). A bounded linear map ` from A onto 1 ,
with A and 1 Banach spaces, is open that is `(l) is open in 1 for any open subset l ⊂ A.
271
272 27. PRINCIPLE OF UNIFORM BOUNDEDNESS AND OPEN MAPPING THEOREM
Proof. Let l be open in A. Then given µ ∈ `(l) we have µ = `r for some r ∈ l.
Since l is open r+c1
1
(0) ⊂ l for some c 0. Thus `r+c`1
1
(0) ⊂ `l so `l contains
an open ball centered at µ = `r.
Proof of Theorem. Since ` maps A onto 1 we have
1 = ∪
∞
n=1
`1
n
(0).
By the Baire category theorem at least one of the sets `1
n
(0) is dense in some open set.
That is there are δ 0 and µ ∈ 1 such that `1
n
(0) ∩ (µ +1
δ
(0)) is dense in 1
δ
(0). Since
` is onto, there is r ∈ A such that `r = µ. So `(1
n
(0) − r) is dense in 1
δ
(0). By the
triangle inequality,
1
n
(0) −r ⊂ 1
n+x
(0).
By homogeneity we conclude that for every : 0, `1
r
(0) is dense in 1
αr
(0) with
α = δ,(: +r).
Now let µ ∈ 1
α
(0) ⊂ 1 . Then there is r
1
∈ 1
1
(0) such that
µ −`r
1
 ≤
1
2
α
Let µ
1
= µ −`r
1
. So µ
1
∈ 11
2
α
(0). Thus there is r
2
∈ 11
2
(0) such that µ
1
−`r
2
 ≤
1
4
α.
That is
µ −`(r
1
+r
2
) ≤
1
4
α.
Following this procedure construct, by induction, a sequence r
1
. r
2
. . . . with
(1) r
j
∈ 1 1
2
j−1
(0) ⊂ A
(2)
_
_
_µ −`
n
j=1
r
j
_
_
_ ≤
1
2
j
α.
Now the partial sums
n
j=1
r
j
converge as : → ∞ to some point r ∈ A with
r ≤
∞
j=1
2
1−j
= 2.
Clearly µ = `r. Thus 1
α
(0) ⊂ `1
2
(0) and the result follows by homogeneity.
Theorem 27.5. Let A and 1 be Banach spaces and ` : A → 1 a bounded, onetoone
map of A onto 1 . Then the inverse map `
−1
is bounded from 1 → A.
Note in particular that
`r ≥
1
`
−1

r .
Thus bounded, onetoone and onto =⇒ ` is bounded from below! This is somehow
amazing because we were not given any quantitative information on the inverse just it’s
existence.
Proof. Since `1
1
(0) ⊃ 1
d
(0) for some d 0 we have
`
−1
µ ∈ 1
1
(0)
for all µ ∈ 1
d
(0). By homogeneity `
−1
 ≤ 1,d.
Definition 27.1. A map ` : A → 1 is closed if whenever r
n
→ r in A and `r
n
→ µ
in 1 then `r = µ.
27. PRINCIPLE OF UNIFORM BOUNDEDNESS AND OPEN MAPPING THEOREM 273
Theorem 27.6 (Closed Graph Theorem). Let A and 1 be Banach spaces and ` : A →
1 a closed linear map. Then ` is bounded.
Proof. Let the graph G of ` by the set of all pairs (r. `r) with r ∈ A. By linearity
of `, G is a linear space under coordinatewise addition and scalar multiplication. Give G
a norm by deﬁning
(r. `r) = r +`r .
Since ` is closed, the space G is complete in this norm. Now let 1 : G → A be the
map 1(r. `r) = r. So 1 is bounded (1 ≤ 1), onetoone and onto. It follows that
1
−1
r = (r. `r) is bounded. Thus there is a constant c < ∞ such that
(r. `r) ≤ c r .
So
`r ≤ (c −1) r .
Definition 27.2. Let A be a linear space and let 
1
and 
2
be two norms on A. The
norms 
j
. , = 1. 2 are called compatible if whenever a sequence converges in both norms
the limits are equal.
For example, on 1
1
∩ 1
2
the 1
1
and 1
2
norms are compatible.
Corollary 27.7. Let A be a linear space and let 
j
,, = 1. 2 be compatible norms. If
A is complete in both norms then the two norms are equivalent: there is c ∈ (0. ∞) such
that c
−1
r
1
≤ r
2
≤ c r
1
.
Proof. Let A
j
, , = 1. 2, be the Banach space which is A under norm 
j
, , = 1. 2.
The identity map 1 : A
1
→ A
2
is closed by compatibility of the norms. Thus 1 is bounded.
Likewise 1 : A
2
→ A
1
is bounded.
LECTURE 28
The Spectrum of a Linear Map
Reading: ¸15.5 and 20.1
First some observations about composition of linear maps. If ` : A → 1 and ` : 1 → 2
are linear maps their composition (``)r = `(`r) is a map `` : A → 2.
Theorem 28.1. If ` and ` are bounded then so is ``. Furthermore
(1) `` ≤ ` `
(2) (``)
t
= `
t
`
t
.
Proof. Exercise, or see Lax.
We now consider some general properties of linear maps from a Banach space A into
itself. Let L(A) = L(A. A). A map ` ∈ L(A) is invertible if ` maps A onto A and is
1−1. We saw last time that it follows from the open mapping theorem that `
−1
is bounded
as well. Let (L(A) denote the set of all invertible maps in L(A).
Proposition 28.2. If 1. 1 ∈ (L(A) then so is 11 and (11)
−1
= 1
−1
1
−1
.
Theorem 28.3. (L(A) is an open set in the uniform topology.
Proof. We must show that if 1 is invertible then so is every map of the form 1 + ¹
with ¹ < c for some c 0. Since 1 + ¹ = 1(1 + 1
−1
¹) and 1
−1
¹ ≤ 1
−1
 ¹ it
suﬃces to prove this for 1 = 1 the identity map.
The basic idea is to make use of the geometric series to write (1 +¹)
−1
as
∞
n=0
(−1)
n
¹
n
.
Since
¹
n
 ≤ ¹
n
this series converges once ¹ < 1. But then
(1 +¹)
N
n=0
(−1)
n
¹
n
= 1 + (−1)
N
¹
N+1
→ 1
as ` → ∞ so (1 +¹) is invertible.
Definition 28.1. The resolvent set ρ(`) of a linear operator ` is the set of λ ∈ C such
that (λ1 −`) ∈ (L(A). The spectrum σ(`) of ` is the complement of ρ(`), that is the
set of λ such that λ1 −` is not invertible.
Theorem 28.4. The resolvent set ρ(`) is open and contains the set ¦λ : [λ[ `¦.
Furthermore, if [λ[ ` then
_
_
(λ1 −`)
−1
_
_
≤
1
[λ[ −`
. ()
281
282 28. THE SPECTRUM OF A LINEAR MAP
Proof. It follows from the previous theorem that the resolvent set ρ(`) is open. Also
following the proof of the previous theorem we see that if λ ` then
(λ1 −`)
−1
=
∞
n=0
λ
−(n+1)
`
n
.
so λ ∈ ρ(`). The estimate () follows by summing the r.h.s. of
_
_
(λ1 −`)
−1
_
_
≤
∞
n=0
[λ[
−(n+1)
`
n
.
Theorem 28.5. For any λ ∈ ρ(`) we have
_
_
(λ1 −`)
−1
_
_
≥
1
dist(λ. σ(`))
.
Proof. For [.[ < 1, (λ1 −`)
−1
 the geometric series
((λ +.)1 −`)
−1
=
∞
n=0
(−.)
n
(λ1 −`)
−1−n
is absolutely convergent, so λ +. ∈ ρ(`).
It follows that for any r ∈ A and / ∈ A
t
the function
)
x,
(λ) = ¸(λ1 −`)
−1
r. /).
has a convergent power series expansion at each point of ρ(`). Thus )
x,
is analytic in ρ(`).
From () we have
[)
x,
(λ)[ ≤ r /
1
[λ[ −`
for large [λ[. Thus )
x,
vanishes at inﬁnity. By Liouville’s theorem a bounded entire function
is constant. Thus, we have either )
x,
≡ 0 or ρ(`) ,= C.
Theorem 28.6. σ(`) is a nonempty closed subset of the disk ¦λ : [λ[ ≤ `¦.
Proof. It follows from the previous proof that σ(`) is closed and contained in ¦[λ[ ≤
`¦. Suppose σ(`) is empty, so ρ(`) = C. It follows that )
x,
≡ 0 for all r ∈ A and
/ ∈ A
t
. This however is a contradiction as then (λ1 − `)
−1
= 0, so 0 is an invertible
operator.
Theorem 28.7. σ(`) = σ(`
t
).
Since (λ1 −`)
t
= λ1 −`
t
, this follows easily from
Lemma 28.8. 1 ∈ (L(A) if and only if 1
t
∈ (L(A
t
).
Proof. If 1 is invertible then 11 = 11 = 1
X
with 1 = 1
−1
. By taking adjoints
1
t
1
t
= 1
t
1
t
= 1
X
.
so 1
t
−1
= [1
−1
]
t
.
Supposes now that 1
t
is invertible. Then ran 1
t
= A
t
, so the null space of 1 (which
is the annihlator of ran 1
t
) is ¦0¦. Thus 1 is onetoone. To see that 1 is onto, note that
since `
K
= ¦0¦ it follows that ran 1 is dense. Thus it suﬃces to show ran 1 is closed. To
28. THE SPECTRUM OF A LINEAR MAP 283
prove this, note that 1
tt
∈ (L(A
tt
) by the ﬁrst part of the theorem and that 1
tt
r = 1r
for r ∈ A. However, A is a norm closed subspace of its double dual, since the norm on
the double dual is the same as the norm on A (by the dual characterization of the norm
r = sup
[¸r. /)[). Applying the following lemma with 1 = A
tt
and 1 = [1
tt
]
−1
it follows
that 1 has closed range.
Lemma 28.9. Let 1 ∈ L(A. 1 ) and suppose there is 1 ∈ L(1. A) such that 11 = 1
X
.
Then `
X
= ¦0¦ and ran 1 is closed.
Proof. Clearly `
X
= ¦0¦. On the other hand if 1r
n
→ µ is a convergent sequence in
ran 1, then r
n
→ 1µ so 1r
n
→ 11µ = µ and µ ∈ ran 1.
In ﬁnite dimensions the spectrum of a linear operator is the same as the set of eigenvalues.
This follows since if A is ﬁnite dimensional then 1 ∈ L(A) is invertible if and only if it is
onetoone `
X
= ¦0¦. In inﬁnite dimension, an operator 1 may fail to be invertible for a
number of reasons:
(1) `
K
may be nontrivial.
(2) ran
K
is contained in a proper closed subspace of A, in which case `
K
is nontrivial.
(3) ran
K
is a proper dense subspace of A
It follows that λ ∈ σ(`) must satisfy (at least) one of the following:
(1) λ is an eigenvalue. That is there is nonzero r such that `r = λr.
(2) The range of λ1 −` is contained in a proper closed subspace of A, in which case
λ is an eigenvalue of `
t
.
(3) The range of λ1 −` may be a proper dense subspace of A.
All three possibilities occur, so spectrum — and spectral theory — in inﬁnite dimensions
is a good deal more complicated than in ﬁnite dimensions.
LECTURE 29
Some examples
Reading: ¸20.2¸20.3
Shifts
Let 1 and 1 denote the right and left shifts on c
0
, sequences that vanish at ∞,
1(c
1
. c
2
. . . .) = (0. c
1
. c
2
. . . .)
1(c
1
. c
2
. . . .) = (c
2
. c
3
. . . .).
Let 1
p
, 1
p
denote the restriction of these operators to /
p
, 1 ≤ j < ∞ and let 1
∞
, 1
∞
denote
the natural extension of these to /
∞
. Note that we have
1
t
= 1
1
and 1
t
= 1
1
.
1
t
p
= 1
p
and 1
t
p
= 1
p
.
1,j + 1,j
t
= 1, 1 ≤ j < ∞.
Proposition 29.1. σ(1) = σ(1) = σ(1
p
) = σ(1
p
) = ¦[λ[ ≤ 1¦, for each 1 ≤ j ≤ ∞.
Proof. First note that 1 = 1 = 1
p
 = 1
p
 = 1, so all of the various spectra are
contained in the unit disk. On the other hand for each [λ[ < 1 the sequence
(λ. λ
2
. λ
3
. . . .)
is an eigenvector of the left shift with eigenvalue λ. Since this vector lies in ∩
p
/
p
= /
1
and
the spectrum is closed we have σ(1) = σ(1
p
) = ¦[λ[ ≤ 1¦. Since 1
1
= 1
t
and 1
p
= 1
t
p
,
1 < j ≤ ∞, it follows that σ(1
p
) = ¦[λ[ ≤ 1¦ for 1 ≤ j ≤ ∞. Likewise, since 1
1
= 1
t
it
follows that σ(1) = σ(1
1
) = ¦[λ[ ≤ 1¦.
Now consider the space c
o
(Z) of two sided sequences. We may deﬁne right and left shifts
here as well
¯
1(. . . . c
−1
. c
0
. c
1
. . . .) = (. . . . c
−2
. c
−1
. c
0
. . . .)
¯
1(. . . . c
−1
. c
0
. c
1
. . . .) = (. . . . c
0
. c
1
. c
2
. . . .).
Note that these operators are isometries and inverses of each other
¯
1
¯
1 =
¯
1
¯
1 = 1.
Proposition 29.2. σ(
¯
1) = σ(
¯
1) = ¦[λ[ = 1¦.
We use
Lemma 29.3. Let ` ∈ L(A) be an isometry onto A. Then σ(`) ⊂ ¦[λ[ = 1¦.
291
292 29. SOME EXAMPLES
Proof. Since `r = r for all r, ` is 1 − 1. As ` is also onto, 0 ∈ ρ(`). As
` = 1, it follows that σ(`) ⊂ ¦[λ[ ≤ 1¦ and that ¦[λ[ < 1¦ ⊂ ρ(`), as the geometric
series
(λ −`)
−1
= −
∞
n=0
λ
n
`
−n−1
converges there.
Proof of Proposition. It follows from the lemma that the two spectra are subsets
of the circle. We show that ran(λ −
¯
1) ,= c
0
(Z) for each λ of modulus one. The proof for
¯
1
is similar.
Suppose (λ1 −
¯
1)a = b. Then the coeﬃcients satisfy
λc
j
−c
j−1
= /
j
.
Since c
j−1
−λ
−1
c
j−2
= λ
−1
/
j−1
it follows that
λc
j
−λ
−1
c
j−2
= (/
j
+λ
−1
/
j−1
).
Continuing we conclude that
λc
j
−λ
−n
c
j−n−1
=
n
m=0
λ
−m
/
j−m
.
Since c
j−n−1
→ 0 as : → ∞ we ﬁnd that
c
j
= lim
n→∞
n
m=0
λ
−m−1
/
j−m
.
Thus, for instance, the sequence /
j
= λ
−j 1
[j[+1
cannot be in ran(λ1 −
¯
1) since
n
m=0
λ
−m−1
/
j−m
= λ
−j−1
n
m=0
1
[, −:[ + 1
diverges as : → ∞.
Note an amusing point of the proof: we actually derived a formula for the inverse of
(λ1 −
¯
1). The point is that this inverse is deﬁned only on a dense domain in A.
Similarly, we have
Proposition 29.4. Let
¯
1
p
,
¯
1
p
be the right and left shifts on /
p
(Z) for 1 ≤ j ≤ ∞. Then
σ(
¯
1
p
) = σ(
¯
1
p
) = ¦[λ[ = 1¦.
The case j = ∞ is easy since then every point of the circle is an eigenvalue of each shift.
The remaining cases are left as an exercise.
Volterra Integral Operators
Consider the operator of integration
\ )(r) =
_
x
0
)(µ)dµ
on the Banach space A = C[0. 1].
VOLTERRA INTEGRAL OPERATORS 293
Theorem 29.5. \ is a bounded operator, \  ≤ 1, and σ(\ ) = ¦0¦. 0 is not an
eigenvalue of \ .
Proof. It is easy to verify that \  ≤ 1. Clearly \ ) = 0 if and only if ) = 0 so 0 is not
an eigenvalue of \ . Lax’s proof that σ(\ ) = ¦0¦ relies on results from the theory of Banach
algebras which we have not proved yet, so let us proceed directly. We will give a formula for
the inverse of λ1 −\ . Note that
\
n
)(r) =
_
x
0
_
y
1
0
. . .
_
y
n−1
0
)(µ
n
)dµ
n
dµ
n−1
dµ
1
.
It follows that
\
n
) ≤ )
_
1
0
_
y
1
0
. . .
_
y
n−1
0
)(µ
n
)dµ
n
dµ
n−1
dµ
n
=
1
:!
) .
Thus the geometric series
∞
n=0
1
λ
n+1
\
n
is norm convergent with norm bounded by [λ[
−1
e
[λ[
−1
. Clearly
(λ1 −\ )
∞
n=0
1
λ
n+1
\
n
= 1.
Thus λ1−\ is boundedly invertible for all λ ,= 0. Since the spectrum is nonempty we must
have σ(\ ) = ¦0¦.
One can also verify directly that \ is not invertible. After all \ )(0) = 0 so ran \ is
contained in the space of functions that vanish at 0. Note that \
t
is the map on /[0. 1]
which maps a measure dj to the measure j(r. 1]dr. That is \
t
(j)(o) =
_
S
j((r. 1])dr for
any measurable set. Thus 0 is an eigenvalue of \
t
with eigenvector δ
0
dr.
Part 8
Compact Linear Maps
LECTURE 30
Compact Maps
Motivation: integral operators
Let 1 be a compact Hausdorﬀ space and j a Borel measure on 1. Suppose 1(:. t) is a
continuous function on 1 1 and deﬁne a map K ∈ L(C(1)) by
K)(:) =
_
T
1(:. t))(t)dj(t).
Since 1 is continuous on a compact space, hence uniformly continuous and bounded, the
image K) is continuous and satisﬁes
max
s
[K)(:)[ ≤ const. max
t
[)(t)[.
Thus K is indeed a bounded map from C(1) into itself.
Such maps are certainly on of the most important examples of bounded linear maps.
(There are 1
1
, 1
2
, etc. analogues.) The map is not typical, however, because it has one
additional important property. Suppose [)(t)[ ≤ 1 for all t. Then
[K)(:) −K)(:)[ ≤ [j[(1) sup
t
[1(:. t) −1(:. t)[.
Since the right hand side is ﬁnite and converges to zero as : → : we ﬁnd that
Lemma 30.1. The collection ¦K) : )
C(T)
≤ 1¦ is equicontinuous. That is, given
: ∈ 1 and c 0 there is a neighborhood l of : such that [K)(:) − K)(:)[ ≤ c whenever
: ∈ l and )
C(T)
≤ 1.
Since the collection is also uniformly bounded, by the ArzelaAscoli theorem
Theorem 30.2. The operator K maps the unit ball in C(1) to a precompact set.
Thus integral operators as above have the rather special property of mapping the unit
ball of the Banach space onto a “small subset” of the Banach space. This has many useful
and important consequences.
Compact operators
Definition 30.1. A map C ∈ L(A. 1 ) is compact if the image C1
1
(0) of the unit ball
in A is precompact in 1 (in the norm topology).
Remark. Recall that a subset o of a complete metric space is precompact if its closure
is compact. Equivalently, every sequence in o has a Cauchy subsequence.
Proposition 30.3. If C
1
and C
2
are precompact subsets of a Banach space then
(1) C
1
+C
2
is precompact.
(2) `C
1
is precompact for any ` ∈ L(A. 1 ) with 1 a Banach space.
(3) the convex hull of C
1
is precompact.
301
302 30. COMPACT MAPS
Let ((A. 1 ) = ¦ compact maps from A → 1 ¦.
Theorem 30.4. ((A. 1 ) is closed subspace of L(A. 1 ). Furthermore if M ∈ L(1. 2)
and N ∈ L(\. A) then MCN ∈ ((\. 2) for any C ∈ ((A. 1 ).
Proof. The proof that ((A. 1 ) is a subspace is left as an exercise, based on the propo
sition above, as is the proof that MC is compact if C is. To see that CN is compact,
note that CN1
1
⊂ C1
N
, so
1
N
CN1
1
is precompact and hence so is CN (by (2) of the
proposition).
Finally, suppose C
n
→ C in L(A. 1 ). Let r
j
be a sequence in the unit ball of A. Let
r
j;1
be a subsequence such that C
1
r
j;1
converges as , → ∞. By induction construct r
j;n
for
each : such that
(1) r
j;n
is a subsequence of r
j;n−1
(2) lim
j
C
n
r
j;n
= µ
n
.
It follows that
µ
m
= lim
j
C
m
r
j;n
for any : ≤ :. Note that µ
m
−µ
n
= lim
j
(C
m
−C
n
)r
j;n
, so
µ
m
−µ
n
 ≤ C
m
−C
n
 .
Thus µ
n
are Cauchy and have a limit µ. For each :, let ,
n
be such that
µ
n
−C
n
r
jn;n
 ≤ C
n
−C .
Consider the diagonal sequence r
jn;n
. Then
Cr
jn;n
−µ ≤ C−C
n
 +C
n
r
jn;n
−µ
n
 +µ
n
−µ ≤ 3 C
n
−C → 0.
Thus Cr
jn;n
is a convergent subsequence of Cr
j
and so C1
1
(0) is compact.
Let ((A) = ((A. A).
Theorem 30.5. Let C ∈ ((A), let I be the identity map on A, and let T = I −C. Then
(1) `
T
is ﬁnite dimensional.
(2) There is an integer i such that
`
T
k = `
T
i for / ≥ i.
(3) ran T is closed.
Proof. Note that r ∈ `
T
iﬀ r = Tr. Thus the unit ball of `
T
is contained in T1
1
(0).
Thus the closed unit ball of `
T
is compact. So `
T
is ﬁnite dimensional. (See Lecture 5.)
Note that `
T
k+1 ⊃ `
T
k, and that each subspace is ﬁnite dimensional (since T
k
is compact
for any /). Suppose this sequence never stabilizes. By Lemma 5.2 there is then a vector r
k
for each / such that
r
k
∈ `
T
k . r
k
 = 1 and r
k
−r
1
2
for all r ∈ `
T
k−1.
Then, if : < :,
C(r
n
−r
m
) = r
n
−Tr
n
−r
m
+Tr
m
.
Now Tr
n
. r
m
. Tr
m
∈ `
T
n−1 (since : < :), so
Cr
n
−Cr
m
 ≥
1
2
.
COMPACT OPERATORS 303
So Cr
n
has no Cauchy subsequence, which is a contradiction.
To prove ran T is closed, let µ
k
= Tr
k
and suppose µ
k
→ µ. We may shift r
k
as we like
by elements of the `
T
. So r
k
may get very large. However, if we consider the distance
d
k
= inf
u∈N
T
r
k
−n .
from r
k
to the `
T
, then I claim we have
sup
k
d
k
< ∞.
Indeed, suppose not. Then, passing to a subsequence so d
k
→ ∞ and choosing n
k
so
r
k
−n
k
 ≤ 2d
k
we have
0 = lim
k
µ
k
d
k
= lim
k
T
r
k
−n
k
d
k
.
since µ
k
are bounded. As r
k
−n
k
 ,d
k
≤ 2 we may pass again to a subsequence to conclude
that
C
r
k
−n
k
d
k
converges. Thus
lim
k
r
k
−n
k
d
k
= . = lim
k
C
r
k
−n
k
d
k
= C..
so T. = 0, that is . ∈ `
T
. But then r
k
−n
k
−d
k
. ≥ d
k
so
_
_
_
x
k
−u
k
d
k
−.
_
_
_ ≥ 1 contradicting
the convergence derived above. Thus sup
k
d
k
< ∞.
Now since d
k
are bounded, we may choose n
k
∈ `
T
so that r
k
−n
k
 ≤ 2 sup
k
d
k
. Now
pass to a subsequence so that
C(r
k
−n
k
)
converges. But then
lim
k
(r
k
−n
k
) = µ + lim
k
C(r
k
−n
k
).
Thus lim
k
(r
k
−n
k
) = r exists and µ = Tr ∈ ran T.
LECTURE 31
Fredholm alternative
Reading: ¸21.1
Theorem 31.1 (Fredholm Alternative). Let C ∈ ((A). Then T = I −C satisﬁes
dim`
T
= dim(A, ran T).
Remark. The result is known as the “Fredholm Alternative” because Fredholm proved
this result in the context of integral operators. The “alternative” is that either one is an
eigenvalue of C, or the equation
µ = r −Cr
is uniquely solvable for r for any µ.
Corollary 31.2. Let C ∈ ((A). Then a nonzero point λ ∈ C is in the spectrum of C
if and only if λ is an eigenvalue of ﬁnite algebraic and geometric multiplicity, where
geometric multiplicity = dim`
λI−C
and
algebraic multiplicity = sup
j
dim`
(λI−C)
j .
Proof. Apply the previous two theorems to the compact map λ
−1
C.
Caution: a compact map need not have eigenvalues. The Volterra operator of the last
lecture is an example.
Recall that dimA,1 is the codimension of 1 , denoted The identity
dim`
M
= codimran `
is valid for any operator ` ∈ L(A) if A is ﬁnite dimensional. In an inﬁnite dimensional
setting, this identity need not hold even if both sides are ﬁnite. If both sides are ﬁnite, the
index of ` is the diﬀerence
ind T = dim`
M
−codimran `.
Thus the alternative says that the index of a I −C is zero for C compact.
To prove the theorem we will use the following
Lemma 31.3. Let C ∈ ((A) and suppose 1 ⊂ A is an invariant subspace for C, so
C1 ⊂ 1 . Then
¯
C[r] = [cr]
is a compact map from A,1 → A,1 .
Proof. Exercise.
311
312 31. FREDHOLM ALTERNATIVE
Proof of Theorem. First suppose dim`
T
= 0. We need to show that ran T = A.
If, on the contrary, A
1
= ran T is a proper subspace of A. Since T is onetoone, A
2
= TA
1
is a proper subspace of A
1
. By induction, with A
k
= T
k
A, we ﬁnd that A
1
⊃ A
2
⊃ A
3
with proper inclusion at every step. Now A
1
is closed by Thm 30.5 from last time. Likewise
A
k
= T
k
A and T
k
= I −
k
j=1
_
k
j
_
(−1)
j−1
C
j
= I −compact, so A
k
is closed. Thus, we may
ﬁnd a sequence of vectors r
j
such that
r
j
∈ A
j
. r = 1 . and dist(r
j
. A
j+1
)
1
2
.
It follows, if : < :, that
Cr
m
−Cr
n
 = x
m
−Tr
m
−r
n
+Tr
n

1
2
.
since Tr
m
+ r
n
− Tr
n
∈ A
m+1
. Thus Cr
n
has no Cauchy subsequence, which is a contra
diction.
If dim`
T
0, then by Thm 30.5 from last time, we have `
T
i+1 = `
T
i for i large enough.
Let ` = `
T
i . So ` is an invariant subspace for T, and since C = I − T we see that `
is invariant for C as well. Thus
¯
C as deﬁned above is a compact map of A,` → A,`.
Consider
¯
T =
¯
I −
¯
C. That, is
¯
T[r] = [1r].
which is welldeﬁned since ` is invariant under 1. Note that `
e
T
= ¦0¦, since if [r] ∈ `
e
T
then 1r ∈ ` =⇒ r ∈ `
T
i+1 = `
T
i = ` so [r] = 0. Thus by ran
¯
T = A,`. Thus for
every µ ∈ A there is r ∈ A and . ∈ ` such that
Tr = µ +..
Thus A = ran T+`. Thus
dim(A, ran T) = dim(`,(` ∩ ran T)).
By ﬁnite dimensional linear algebra
dim(`,(` ∩ ran T)) = dim`
T
.
Theorem 31.4 (Schauder). C ∈ L(A) is compact if and only if C
t
is compact.
Proof. We show C compact =⇒ C
t
is compact. The reverse implication follows by
noting that C is the restriction of C
tt
to the closed subspace A of A
tt
and applying the
following
Lemma 31.5. Let C be compact on A and let 1 ⊂ A be a closed subspace. Then the
restriction of C to 1 is a compact map.
Proof. Exerise.
We must show, if /
n
∈ 1
1
(0) ⊂ A
t
, that C
t
/
n
has a convergent subsequence. Consider
the closure of 1 = C1
1
(0) ⊂ A. This is a compact set and on this set
[/
n
(r) −/
n
(µ)[ ≤ r −µ
since /
n
 ≤ 1. Thus /
n
[
K
are equicontinuous, and bounded, so have a convergent subse
quence by ArzelaAscoli. That is, passing to a subsequence, we have
lim
n→∞
sup
y∈K
[¸µ. /
n
−)(µ)[ = 0
31. FREDHOLM ALTERNATIVE 313
for some cts function ) on 1. Thus
sup
x∈B
1
(0)
[¸r. C
t
/
n
−)(Cr)[ −→ 0.
It follows that /(r) = )(Cr) is a linear functional and the norm limit of C
t
/
n
. So C
t
is
compact.
Theorem 31.6 (Fredholm Alternative). Let C ∈ ((A), T = I −C. Then
(1) r ∈ ran T if and only if /(r) = 0 for all / ∈ `
T
.
(2) dim`
T
= dim`
T
.
Proof. (1) For a general operator T we have
ran T = ¦r : /(r) = 0 fo all / ∈ `
T
¦ .
Since ran T is closed (1) follows.
(2) The null space `
T
is isomorphic to the dual of A, ran T via the pairing
¸[r]. /) = ¸r. /)
which is well deﬁned since / annihlates the ran T. Thus
dim`
T
= codimran T = dim`
T
.
LECTURE 32
Spectral Theory of Compact Maps
Theorem 32.1 (F. Riesz). Let A be a complex Banach space and C ∈ ((A). The
spectrum of C is a denumerable or ﬁnite set of points whose only accumulation point, if any,
is 0. If dimA = ∞ then 0 ∈ σ(C):
σ(C) = ¦λ
j
¦ ∪ ¦0¦.
Furthermore,
(1) Each nonzero λ
j
∈ σ(C) is an eigenvalue of ﬁnite algebraic and geometric multi
plicity.
(2) The resolvent (. −C)
−1
has a pole at each nonzero λ
j
: that is, there is :
j
≥ 1 such
that
(. −λ
j
)
n
j
¸(. −C)
−1
r. /)
is analytic in a neighborhood of λ
j
for every / ∈ A
t
and r ∈ A.
Proof. We have already seen that any nonzero point of the spectrum is an eigenvalue
of ﬁnite multiplicity.
Suppose we have a sequence λ
n
of eigenvalues and eigenvectors
Cr
n
= λ
n
r
n
.
with λ
n
,= λ
m
. Let 1
n
be the linear space of r
1
. . . . . r
n
. Suppose
n
j=1
c
j
r
j
= 0.
Then
n
j=1
c
j
λ
k
j
r
j
= 0
for all /, so
n
j=1
c
j
j(λ
j
)r
j
= 0
for all polynomials j. Since the λ
j
are distinct, we may pick a polynomial that vanishes for
all λ
j
except λ
j
0
for example j(λ) =
j,=j
0
(λ −λ
j
). Thus c
j
= 0 for all ,, so r
j
are linearly
independent. It follows that 1
n−1
is a proper subspace of 1
n
.
The implication of this is the same as above: we ﬁnd a sequence of vectors µ
n
∈ 1
n
such
that
µ
n
 = 1 and µ
n
−µ
1
2
for all µ ∈ 1
n−1
.
Since
µ
n
=
n
j=1
c
(n)
j
r
j
321
322 32. SPECTRAL THEORY OF COMPACT MAPS
we have
Cµ
n
=
n
j=1
c
(n)
j
λ
j
r
j
so
Cµ
n
−λ
n
r
n
=
n−1
j=1
c
(n)
j
(λ
j
−λ
n
)r
j
∈ 1
n−1
.
Thus, for : :,
Cµ
n
−Cµ
m
= λ
n
µ
n
−µ for some µ ∈ 1
n−1
.
Thus
Cµ
n
−Cµ
m
 ≥
[λ
n
[
2
.
Since any subsequence of Cµ
n
has a Cauchy subsequence, we must have λ
n
→ 0. Since λ
n
was an arbitrary sequence of eigenvalues, it follows that 0 is the only (possible) accumulation
point of eigenvalues. Thus there are only ﬁnitely many eigenvalues outside any disk around
the origin. Hence there at most countably many eigenvalues.
It remains to show that the resolvent has poles at λ
j
. Let . be a complex number, then
ﬁnding the resolvent n = (. −C)
−1
r amounts to solving
r = .n −Cn
for r. Suppose . is close to λ
j
. Let ` = `
(λ−C)i
= `
(λ−C)
i+1 with i suﬃciently large. Let
¯
C
be the quotient map on A,`. Since λ
j
is not an eigenvalue of
¯
C (why?), we conclude that
λ
j
−
¯
C is invertible. It follows that . −
¯
C is invertible for . −λ
j
< c for some c 0 (recall
that the set of invertible maps is open) and
_
_
(. −C)
−1
_
_
≤ const. for [. −λ
j
[ < c.
Thus
.[·] −
¯
C[·] = [r]
has a unique solution for an equivalence class [·] = · +` with
[·] ≤ const. [r] .
Thus given r we may ﬁnd : ∈ ` and ·(.) ∈ A such that
(. −C)·(.) = r −:(.).
We may choose ·(.) so that ·(.) is bounded for [. − λ
j
[ < c. It follows that :(.) is
bounded as well:
:(.) ≤ ([.[ +C) ·(.) +r .
Let us now solve
(. −C)µ(.) = :(.)
for µ(.) ∈ ` to obtain n(.) = ·(.) + µ(.). This is a linear algebra problem. We know that
` is an invariant subspace for C and that λ
j
is in the spectrum σ(C [
N
). In fact, λ
j
is the
unique point of the spectrum of C [
N
. Indeed, (λ
j
−C)
i
= 0 on `. It follows that
_
_
(. −C [
N
)
−1
_
_
≤ const.
1
[λ −.[
i
.
(It doesn’t really matter for the proof, but yes it is the same i in both spots.) Hence,
n(.) ≤ ·(.) +µ(.) ≤ const.(1 +[. −λ
j
[
−i
)
32. SPECTRAL THEORY OF COMPACT MAPS 323
for . close to λ
j
.
Thus (.−λ
j
)
i
¸(.−C)
−1
r. /) is analytic in ¦0 < [.−λ
j
[ < c¦ with a removable singularity
at . = λ
j
. That is ¸(. −C)
−1
r. /) has a pole at λ
j
Note that the resolvent may not have a pole at 0. For instance, the Volterra integral
operator has resolvent
(λ −\ )
−1
=
∞
n=0
1
λ
n+1
\
n
for λ ,= 0. This analytic operator valued function has an essential singularity at 0. If one
computes
(λ −\ )
−1
r
n
=
∞
m=0
1
λ
m+1
\
m
r
n
=
∞
m=0
1
λ
m+1
:!
(: +:)!
r
n+m
= :!λ
n−1
∞
m=n
1
:!
_
r
λ
_
m
= :!λ
n−1
e
x/λ
−
n−1
m=0
:!
:!
r
m
λ
n−1−m
.
then one can see the essential singularity very clearly in the ﬁrst term. The second term is
regular and has limit
:r
n−1
as λ −→ 0.
Interestingly, this is −\
−1
r
n
where \
−1
is the densely deﬁned left inverse for \ , namely
diﬀerentiation:
∂
x
\ )(r) = ∂
x
_
x
0
)(µ)dµ = )(r).
Homework III
Homework III
Due: April 30, 2008
This homework assignment deals with elliptic PDE’s. Let Ω ⊂ R
d
be an open set with
compact closure. Suppose we are given a measurable function σ(r) on Ω which is bounded
above and below:
0 < c ≤ σ(r) ≤ c
−1
< ∞ for all r ∈ Ω.
Consider the PDE
∇ σ(r)∇n(r) = )(r) r ∈ Ω. n = 0 on ∂Ω. ()
Our goal is to show that this equation has a unique solution n = o) with o a compact
symmetric operator on 1
2
and to consider the spectral theory of o.
(1) Show for any ) ∈ 1
2
(Ω) that () has a unique solution (in the sense of distributions) n
in 1
2
(Ω). (Hint: use Lax Milgram. The solution is outlined in Lecture 16.)
(2) Let Φ(r) be a smooth, nonnegative, compactly supported, function on R
d
with
_
R
d
Φ(r)dr =
1. Let 1
t
map 1
2
(Ω) into itself by
1
t
)(r) = χ
Ω
(r)
_
R
d
t
−d
Φ(
r −µ
t
))(µ)dµ = χ
Ω
(r)
_
R
d
t
−d
Φ(
µ
t
))(r −µ)dµ.
where we take ) ≡ 0 outside Ω.
(a) Use Minkowski’s inequality:
_
_
dj(r)
¸
¸
¸
¸
_
)(r. µ)dν(µ)
¸
¸
¸
¸
2
_1
2
≤
_
dν(µ)
__
dj(r) [)(r. µ)[
2
_1
2
to show that 1
t
∈ L(1
2
(Ω) with 1
t

L
2
(Ω)→L
2
(Ω)
≤ 1.
(b) Show also that
sup
x
[1
t
)(r)[ ≤ Ct
−d
)
L
2
(Ω)
.
(Hint: the constant will be proportional to Vol(Ω).)
(c) Show that for ) ∈ 1
2
(Ω) and r ∈ Ω
)(r) −1
t
)(r) =
_
R
d
t
−d
Φ(
µ
t
)()(r) −)(r −µ))dµ.
(d) Apply Minkowski’s inequality to conclude, if ) ∈ C
2
c
(Ω) then
) −1
t
)
L
2
(Ω)
≤ Ct )
H
1
0
(Ω)
. ()
where
)
2
H
1
0
(Ω)
=
d
j=1
∂
j
)
2
L
2
(Ω)
.
(Hint: write )(r) − )(r − µ) =
_
1
0
µ ∇)(r − :µ))d:. Make sure you get the factor
of t on the r.h.s!)
322
(e) Conclude that () holds for ) ∈ H
1
0
(Ω) with
H
1
0
(Ω) = ¦completion of C
∞
c
(Ω) in the norm )
2
H
1
=
d
j=1
∂
j
)
2
L
2
¦.
(3) Let )
n
∈ H
1
0
(Ω) be a sequence so that ∂
j
)
n
÷ ·
j
weakly in 1
2
(Ω) for each ,.
(a) Use Lemma 16.1 of Lecture 16 (Lemma 2 of Ch. 7 in Lax) to conclude that )
n
÷ )
weakly in 1
2
(Ω) for some ) ∈ H
1
0
(Ω) and that ∂
j
) = ·
j
. (Hint: ﬁrst use Alaogulu’s
theorem to conclude that )
n
has a weakly convergent subsequence with a limit )
that satisﬁes ∂
j
) = ·
j
for each ,. Next show that any other weakly convergent
subsequence must also converge to ). Conclude that )
n
÷ ).)
(b) Show that
1
t
)
n
(r) −→ 1
t
)(r) as : → ∞.
for every r ∈ Ω. Since [1
t
)
n
(r)[ ≤ Ct
−d
)
n

L
2
(Ω)
by (2b) use dominated convergence
and the principle of uniform boundedness to conclude that
1
t
)
n
−1
t
)
L
2
−→ 0 as : → ∞.
(c) Use this result and () to conclude that
)
n
−)
L
2
−→ 0 as : → ∞.
(d) Combine these to prove
Theorem 32.2 (Rellich). If ` ⊂ H
1
0
(Ω) with uniformly bounded H
1
0
(Ω) norm then
` is precompact in 1
2
(Ω).
(4) Returning to the PDE (), let o : 1
2
(Ω) → 1
2
(Ω) be the map o) = n where n solves
(). Show that o is compact. (Hint: see ¸22.3 where this is done for σ ≡ 1. However, we
have not assumed that Ω has a smooth boundary. But you have just proved the results
needed.)
(5) Show that o is symmetric on 1
2
(Ω). Conclude that there is an orthonormal basis ¦φ
j
¦
for 1
2
(Ω) consisting of eigenfunctions to the linear operator ∇ σ(r)∇. That is, there is
a set ¦φ
j
¦ such that
span¦φ
j
¦ = 1
2
(Ω) . ¸φ
j
. φ
k
)
L
2
(Ω)
= δ
i,j
.
and
∇ σ(r)∇φ
j
(r) = λ
j
φ
j
(r).
Show that the eigenvalues λ
j
are real and diverge to ∞. How are they related to the
eigenvalues of o?
Part 9
Compact Linear Maps in Hilbert Space
LECTURE 33
Compact Symmetric Operators
We now specialize to the case in which the Banach space A is a Hilbert space. For
spectral analysis it is useful to consider a complex Hilbert space H, in which case in place
of the transpose it is natural to consider
Definition 33.1. The adjoint of a bounded linear map 1 ∈ L(H. 1) with H and 1
Hilbert spaces is the linear map 1
†
∈ L(1. H) given by
¸1
†
n. ·)
H
= ¸n. 1·)
K
for all n ∈ 1 and · ∈ H.
Recall that, by the RieszFrechet theorem the dual of a Hilbert space H is isomorphic to
the Hilbert space itself via the conjugate linear map
n → ¸/
u
() = . n).
Thus we have
1
t
/
u
= /
T
†
u
.
Since the correspondence n → /
u
is conjugate linear, it follows that
(1 +co)
†
= 1 +c
o
†
.
We also have
(o1)
†
= 1
†
o
†
and
σ(1
†
) = σ(1)
= ¦λ
: λ ∈ σ(1)¦.
Definition 33.2. A bounded operator 1 ∈ L(H) is called Hermitian (or symmetric, or
selfadjoint) if 1 = 1
†
.
Proposition 33.1. If 1 is Hermitian then
(1) ¸1n. n) is real for all n ∈ H.
(2) If ¸1n. n) = 0 for all n ∈ H then 1 = 0.
(3) σ(1) ⊂ R.
Proof. (1) Note that ¸1n. n) = ¸n. 1n) = ¸1n. n)
. For (2), note that if ¸1n. n) = 0 for
all n then
0 = ¸1(n ±i·). n ±i·) = ±i¸1·. n) ∓i¸1n. ·).
Combining the results we get
0 = ¸1n. ·)
for all n. · ∈ H. Thus 1 ≡ 0.
For (3), we need to show `
zu−T
= ¦0¦ and ran(. −1) = H, for . ∈ C ¸ R. Suppose
.n −1n = 0.
331
332 33. COMPACT SYMMETRIC OPERATORS
It follows that
. n
2
−¸1n. n) = 0.
Since the second term is real we ﬁnd Im. n
2
= 0. Thus n = 0 and n = 0. That is,
`
zu−T
= ¦0¦.
Similarly, suppose n is perpendicular to ran(. −1). Then
0 = ¸(. −1)n. n).
so n = 0. It follows that ran(. −1) is dense. To see that ran(. −1) is closed, note
(. −1)n
2
= ¸(. −1)n. (. −1)n)
= [.[
2
n
2
+1n
2
−.¸n. 1n) −.
¸1n. n)
= [.[
2
n
2
+1n
2
−2 Re .¸1n. n)
= (Im.)
2
n
2
+(Re . −1)n
2
≥ (Im.)
2
n
2
.
Thus, if ran(. − 1)n
n
→ r we ﬁnd that n
n
is Cauchy. since (. − 1) is bounded we must
have (. −1)n = r for n = lim
n
n
n
.
Theorem 33.2. Let H be a inﬁnite dimensional separable Hilbert space. If 1 ∈ L(H) is
Hermitian and compact then there is an orthonormal basis for H consisting of eigenvectors
of 1. That is, there is an orthonormal basis ¦φ
n
¦
∞
n=1
and a sequence of real numbers λ
n
→ 0
such that
1φ
n
= λ
n
φ
n
.
The spectrum σ(1) = ¦λ
n
¦
∞
n=1
.
Remark. If H is notseparable and 1 is compact Hermitian then there is a closed
separable subspace 1 ⊂ H such that 1 : 1 → 1 and
1[
Y
⊥ = 0.
Note that we already have σ(H) = ¦λ
n
¦ ⊂ R. To proceed let us show ﬁrst
Lemma 33.3. Let 1 ∈ ((H). Let λ
+
= max σ(1) and λ
−
= inf σ(1). Then
λ
+
= sup
u=1
¸1n. n) = max
u=1
¸1n. n)
and
λ
−
= inf
u=1
¸1n. n) = min
u=1
¸1n. n).
In particular, if σ(1) = ¦0¦ then 1 ≡ 0.
Proof. Let us prove the identity for λ
+
. The identity for λ
−
follows from the ﬁrst
applied to −1. Let 1 = sup
u=1
¸1n. n). Note that
¸1n. n) ≤ 1n
2
.
It follows that 1 ≥ 0 as otherwise [¸1n. n)[ ≥ [1[ n
2
. so [1[ n ≤ 1n . which contradicts
compactness. (Look at 1n
n
with n
n
an orthonormal basis.)
Let n
n
be a sequence of vectors with n
n
 ≤ 1 so that ¸1n
n
. n
n
) → 1. Passing to a
subsequence, we suppose that n
n
÷ n
+
and 1n
n
→ 1n
+
(strongly). It follows that
¸1n
n
. n
n
) → ¸1n
+
. n
+
).
33. COMPACT SYMMETRIC OPERATORS 333
so ¸1n
+
. n
+
) = 1. Thus the sup is actually a max.
If λ 1 then, for n ≤ 1,
(λ −1)n ≥ λ −¸1n. n)λ
2
≥ λ −1.
It follows that
(λ −1)n ≥ (λ −1) n
2
for all n. Thus λ 1 ≥ 0 is not an eigenvalue of 1. It follows that λ
+
≤ 1.
If 1 = 0, we are done as λ
+
≥ 0. On the other hand, if 1 0 then n
+
,= 0, since
1 = ¸1n
+
. n
+
. Clearly n
+
is a unit vector. Let u ∈ H and consider the function
1
w
(t) =
¸1(n
+
+tu). n
+
+tu)
n
+
+tu
2
.
So 1
w
is smooth and takes its maximum at 0. Diﬀerentiating with respect to t we get
0 = 1
t
w
(0) = ¸1u. n
+
) +¸1n
+
. u) −¸1n
+
. n
+
) [¸u. n
+
) +¸n
+
. u)] .
Thus
0 = 1
t
w
(0) + i1
t
iw
(0) = ¸1n
+
. u) −1¸n
+
. u).
Since this holds for any u, we conclude that 1n
+
= 1n
+
, so 1 ∈ σ(1).
Finally, note if σ(1) = ¦0¦ then ¸1n. n) = 0 for all n so 1 ≡ 0 by the proposition proved
above.
Proof of Theorem. If 1 ≡ 0 there is nothing to show as any orthonormal basis will
do. If 1 is not zero, then it has a nonzero eigenvalue λ
1
by the lemma. Let the corresponding
eigenvector be n
1
. If n ⊥ n
1
then
¸1n. n
1
) = λ
1
¸n. n
1
) = 0.
so 1n ⊥ n
1
. Thus ¦n
1
¦
⊥
is an invariant subspace for 1. Either 1 is identically zero on ¦n
1
¦
⊥
or it has an eigenvector there. Proceed by induction.
More formally, consider the collection o of closed subspaces 1 of H invariant under 1
and such that 1 restricted to 1 has an orthonormal basis of eigenvectors. As above 1
⊥
is
invariant under 1 so either 1 ≡ 0 on 1
⊥
or 1 has a nonzero eigenvector there. Partially
order o by inclusion. It follows that any maximal element 1 of o has 1
⊥
= ¦0¦.
LECTURE 34
MinMax
Let us look at the variational characterization of eigenvalues again. We have
λ
+
= max σ(1) = max
u≤1
¸1n. n).
and a similar identity at the bottom of the spectrum. As it turns out, one can construct
in this way all of the eigenvalues. Indeed, suppose we are given the largest ` − 1 positive
eigenvalues λ
1
≥ λ
2
≥ ≥ λ
N−1
0 of 1 (listed according to multiplicity) together with
the corresponding eigenvectors r
1
. . r
N−1
. Let
λ
N
= max
u⊥x
1
,...x
N−1
, u≤1
¸1n. n). ()
If λ
N
0 then λ
N
is an eigenvalue of 1, λ
N
≤ λ
N−1
, and the corresponding maximizer n
N
is an eigenvector. If λ
N
= 0 then there are no more positive eigenvalues. (Zero may or may
not be an eigenvalue.) Similarly we can ﬁnd the negative eigenvalues and corresponding
eigenvectors by minimizing ¸1n. n).
Theorem 34.1 (MinMax Principle). Let 1 be a compact Hermitian operator on an
inﬁnite dimensional Hilbert space.
(1) Fischer’s principle: If
α
N
= max
S
N
min
u∈S
N
, u=1
¸1n. n).
where the max is taken over all ` dimensional subspaces o
N
, then α
1
≥ α
2
≥ 0
and the nonzero entries of this sequence are the positive eigenvalues of 1 listed
according to multiplicity.
(2) Courant’s principle: The above sequence is also given by
α
N
= min
S
N−1
max
u⊥S
n−1
. u≤1
¸1n. n).
where the max is taken over all ` −1 dimensional subspaces o
N−1
.
Remark. (1) In ﬁnite dimensions, the sequence is no longer nonpositive and gives all
the eigenvalues. In inﬁnite dimensions, the sequence α
j
, being nonpositive and decreasing,
is either positive for all , or eventually 0. The zero entries may not represent eigenvalues.
(2) The negative eigenvalues j
1
≤ j
2
≤ < 0 can of course be found as
j
N
= min
S
N
max
u∈S
N
, u=1
¸1n. n) = max
S
N−1
min
u⊥S
N−1
, u≤1
¸1n. n).
Proof. First note that α
N
≥ 0. Indeed, for any δ 0 there are only ﬁnitely many
eigenvalues λ < −δ (counted according to multiplicity). Thus the orthogonal complement of
the corresponding eigenvectors is inﬁnite dimensional. On this subspace we have ¸1n. n) ≥
−δ n
2
, so α
N
≥ −δ.
341
342 34. MINMAX
Let ¹r
n
= λ
n
r
n
be the positive eigenvalues and eigenvectors listed according to mul
tiplicity with λ
1
≥ λ
2
≥ 0. (The sequence may terminate.) If there are ` positive
eigenvalues and ` `, let o
N
be an `dimensional subspace. By ﬁnite dimensional linear
algebra, we can ﬁnd a unit vector n ∈ o
N
perpendicular to r
1
. . . . . r
M
:
¸n. r
n
) = 0. : = 1. . . . . `.
By () we must have ¸1n. n) ≤ 0 so α
N
≤ 0. On the other hand α
N
≥ 0 so α
N
= 0.
If there are at least ` positive eigenvalues, then with o
N
= span(r
1
. . . . r
N
) we have
min
u∈S
N
,u=1
¸1n. n) = λ
N
.
so α
N
≥ λ
N
. On the other hand if o
N
is any ` dimensional subspace then as above we may
ﬁnd a unit vector n ∈ o
N
which is perpendicular to the ﬁrst ` − 1 eigenvectors. By (),
¸1n. n) ≤ λ
N
. Thus α
N
≤ λ
N
.
Turning now to Courant’s principle, let
β
N
= min
S
N−1
max
u⊥S
N−1
, u≤1
¸1n. n).
Since o
⊥
N−1
is inﬁnite dimensional and 1 is compact it follows that
max
u⊥S
N−1
, u≤1
¸1n. n) ≥ 0.
Thus β
N
≥ 0. Also, note that
β
N
≤ min
S
N−1
min
S
N−2
⊂S
N−1
max
u⊥S
N−2
, u≤1
¸1n. n) = β
N−1
.
First suppose there are at least ` − 1 positive eigenvalues. It follows from () that
β
N
≤ λ
N
= α
N
. since r
1
. . . . . r
N−1
span an ` −1 dimensional subspace.
If λ
N
= 0, it follows that β
M
= 0 = α
M
for ` ≥ `.
On the other hand, suppose there are at least ` positive eigenvalues and let o
N−1
be an
arbitrary ` −1 dimensional subspace. I claim we may ﬁnd a vector in the subspace spanned
by r
1
. . . . r
N
perpendicular to o
N−1
. Indeed, if µ
j
, , = 1. . . . . ` −1, is an orthonormal basis
for o
N−1
then we must solve
N
k=1
¸r
k
. µ
j
)c
k
= 0.
for all ,. Since ` −1 < ` the matrix ¸r
k
. µ
j
) has a nontrivial null space, and the resulting
solution n =
k
c
k
r
k
is perpendicular to o
N−1
. But then
¸1n. n) =
N
k=1
[c
k
[
2
λ
k
≥ λ
N
n
2
.
Thus β
N
≥ λ
N
= α
N
.
The minmax principle is incredibly powerful. It is certainly one of the most important
results in applications of functional analysis. Here is an example of what we can do with it:
Definition 34.1. Let ¹. 1L(H) be Hermitian. We say that ¹ ≥ 1 if
¸¹n. n) ≥ ¸1n. n) ∀n ∈ H.
Remark. This is a partial order on the set of Hermitian operators.
34. MINMAX 343
Theorem 34.2. Let 1. o ∈ ((H) be Hermitian compact operators with 1 ≥ o. If o has
at least ` positive eigenvalues λ
+
1
≥ ≥ λ
+
N
0 then 1 has at least ` positive eigenvalues
j
+
1
≥ ≥ j
+
N
0 and
j
+
N
≥ λ
+
N
.
Similarly, if 1 has at least ` negative eigenvalues j
−
1
≤ ≤ j
−
N
< 0 then o has at least `
negative eigenvalues λ
−
1
≤ ≤ λ
−
N
< 0 and
λ
−
N
≤ j
−
N
.
This theorem is an easy consequence of minmax. It is immediate if o and 1 have the
same eigenvectors, but that is not at all necessary for the relation o ≤ 1. For instance
_
1 0
0 0
_
≤
_
2 1
1 1
_
.
Similarly we have
Theorem 34.3. Let 1 ∈ ((H) be compact Hermitian. If 1 ≥ 0 then all eigenvalues of
1 are ≥ 0.
In fact, we have
Theorem 34.4. Let 1 ∈ L(H) be Hermitian. If 1 ≥ 0 then σ(1) ⊂ [0. ∞).
Proof. Let λ 0. Then
(1 +λ)n
2
= 1n
2
+ 2λ¸n. 1n) +λ
2
n
2
≥ λ
2
n
2
.
It follows that `
T+λ
= ¦0¦ and that ran(1 +λ) is closed. Since ran(1 +λ)
⊥
= `
T+λ
= ¦0¦
it follows that ran(1 +λ) = H. Thus 1 +λ is onetoone, onto and bounded. By the inverse
mapping theorem (27.5 in these notes) (1 +λ)
−1
is bounded. Thus −λ ,∈ σ(1).
LECTURE 35
Functional calculus and polar decomposition
The spectral theorem states that a compact Hermitian operator 1 is of the form:
1 =
∞
n=1
λ
n
¸. φ
n
)φ
n
. ()
Here λ
n
are the eigenvalues and φ
n
is an orthonormal basis. The notation indicates that
1φ =
∞
n=1
λ
n
¸φ. φ
n
)φ
n
.
In fact, since λ
n
→ 0 the sum in () is absolutely convergent:
_
_
_
_
_
1 −
N
n=1
λ
n
¸. φ
n
)
_
_
_
_
_
= sup
n>N
[λ
n
[ −→ 0.
If ) : σ(1) →C is a bounded function we deﬁne
)(1) =
∞
n=1
)(λ
n
)¸. φ
n
)φ
n
.
so
)(1)φ =
∞
n=1
)(λ
n
)¸φ. φ
n
)φ
n
.
It is an easy calculation to see that if )(r) = j(r) is a polynomial in r then )(1) deﬁned
this way agrees with plugging 1 into the polynomial. For instance,
∞
n=1
λ
2
n
¸φ. φ
n
)φ
n
=
∞
n=1
λ
n
¸
∞
m=1
λ
m
¸φ. φ
m
)φ
m
. φ
n
) = 1
2
φ.
by the pairwise orthogonality of φ
m
. The map ) → )(1) is called the functional calculus for
1 and has the following properties:
Theorem 35.1. Let 1 be a compact Hermitian operator. To every bounded function
) : σ(1) →C we assign a unique operator, )(1), such that
(1) )(σ) ≡ 1 =⇒ )(1) = I.
(2) )(σ) = σ =⇒ )(1) = 1.
(3) The map ) → )(1) is an injective homomorphism of the ring of bounded functions
on σ(1) into the L(H):
() +p)(1) = )(t) +p(1). ()p)(1) = )(1)p(1). and )(1) ≡ 0 iﬀ ) ≡ 0 on σ(1).
(4) )(1)
†
= )
(1).
351
352 35. FUNCTIONAL CALCULUS AND POLAR DECOMPOSITION
(5) The map is an isometry:
)
∞
= sup
λ∈σ(T)
[)(λ)[ = )(1) .
This map has the properties
(6) If ) : σ(1) →R then )(1) is Hermitian.
(7) If ) : σ(1) → [0. ∞) then )(1) ≥ 0.
(8) If ) : σ(1) → ¦[.[ = 1¦ then )(1) is a unitary map, that is an isometry of H onto
H.
(9) If lim
λ→0
)(λ) = )(0) = 0 then )(1) is compact.
Remark. Properties (35) show that the map ) → )(1) is an isomorphism of C
algebras, something we haven’t deﬁned yet but that we will see in the fall.
Proof. We already deﬁned the map we will show it is unique in a moment. It is clear
that (1) and (2) hold. It is easy to see that () + p)(1) = )(1) + p(1) and the argument
given for 1
2
above extends to a product )p. Thus (3) holds. To see that (4) holds note that
¸)(1)
†
n. ·) = ¸n. )(1)·) =
_
n.
n
)(λ
n
)
¸·. φ
n
)φ
n
_
=
n
)(λ
n
)
¸n. φ
n
)¸φ
n
. ·) =
_
n
)(λ
n
)
¸n. φ
n
)φ
n
. ·)
_
= ¸)
(1)n. ·).
Since
)(1)n
2
=
n
[)(λ
n
)[
2
[¸n. φ
n
)[
2
≤ )
∞
n
2
and
)(1)φ
n
= )(λ
n
)φ
n
.
(5) follows. (6) and (7) are easy calculations. To see (8) note that if [)(r)[
2
= 1 then
)(1)n
2
=
n
[¸n. φ
n
)[
2
= n
2
.
(9) is an easy exercise.
Finally to see that the map is unique, note that (1), (2), and (3) specify the map for
polynomials j(r). Since 1 is compact, σ(1) is discrete away from 0. Thus a bounded map
on σ(1) is continuous if and only if it is continuous at 0. Thus (5) and StoneWeierstrass
imply that the map for polynomials extends uniquely to bounded functions continuous at 0.
In fact the map is uniquely deﬁned on all bounded functions, but we will defer the proof of
this to next term.
Corollary 35.2. If 1 ≥ 0 and is compact Hermitian then there is a unique positive
square root of 1:
√
1 ≥ 0 and
√
1
2
= 1.
The square root is very useful as it allows us to deﬁne
Definition 35.1. Let 1 be a compact operator on a Hilbert space H. The absolute
value of 1, denoted [1[, is the operator
√
1
†
1.
35. FUNCTIONAL CALCULUS AND POLAR DECOMPOSITION 353
The absolute value [1[ is compact (why?) so it has eigenvalues, which are all nonnegative
since [1[ ≥ 0. The eigenvalues of [1[ are called the singular values of 1. Note that
1φ
2
= ¸φ. 1
†
1φ) = ¸φ. [1[
2
φ) = [1[φ
2
.
It follows that the map [1[φ → 1φ deﬁned on the range of [1[ is a linear isometry. This
extends to the closure of ran [1[. Let this map be denoted l and deﬁne l to be zero on
ran [1[
⊥
= `
[T[
= `
T
. Note that
1 = l[1[.
Theorem 35.3 (Polar decomposition). Every compact operator 1 on a Hilbert space H
may be factored, as
1 = l¹
with ¹ ≥ 0 and l
†
l [
ran A
= I. The map ¹ = [1[ and l is uniquely determined if we specify
l ≡ 0 on ran ¹
⊥
.
Corollary 35.4. Any compact operator 1 on a Hilbert space has a singular value
decomposition
1 =
∞
n=1
j
n
¸. φ
n
)ψ
n
.
with j
n
≥ 0, j
n
→ 0 and φ
n
, ψ
n
(possibly distinct) orthonormal sequences.
Proof. Let j
n
, φ
n
be the eigenvalues/vectors of [1[. Let ψ
n
= lφ
n
with 1 = l[1[.
In fact, the functional calculus, square root and polar decomposition extend to non
compact operators. These are some theorems we will prove next term:
Theorem 35.5. Every positive operator ¹ ≥ 0 on a Hilbert space has a unique positive
square root.
Theorem 35.6 (Polar Decomposition). Every operator 1 ∈ L(H) may be factored as
1 = l¹
with ¹ ≥ 0 and l
†
l [
ran A
= I. The map ¹ = [1[ =
√
1
†
1 and l is uniquely determined if
we specify l ≡ 0 on ran ¹
⊥
= `
T
.
Theorem 35.7 (Spectral Theorem: functional calculus version). To every self adjoint
operator 1 ∈ L(H) there is associated a unique injective C
homomorphism from C(σ(1)) →
L(H) with the properties (1)(8) above. Conversely, given any compact set σ ⊂ R and an
injective C
homomorphism with these properties there is a self adjoint operator 1 so that
the homomorphism is realized as ) → )(1).