Joseph Muscat 2015 1

Topological Vector Spaces and

Algebras
[Link]@[Link]
1 June 2016

1 Topological Vector Spaces over R or C

Recall that a topological vector space is a vector space with a T0 topology such
that addition and the field action are continuous. When the field is F := R or
C, the field action is called scalar multiplication.
Examples:
• RA , such as sequences RN , with pointwise convergence.
• Sequence spaces ℓp (real or complex) with topology generated by Br =
P p
p p
{ (an ) : n |an | < r }, where p > 0.
R p
• Lebesgue spaces Lp (A) with Br = { f : A → F, measurable, p |f |p < r }
(p > 0).
• Products and quotients by closed subspaces are again topological vector
spaces.
If πi : Y → Xi are linear maps, then the vector space Y with the ini-
tial topology is a topological vector space, which is T0 when the πi are
collectively 1-1.
The set of (continuous linear) morphisms is denoted by B(X, Y ). The mor-
phisms B(X, F) are called ‘functionals’.

+, ∗, → Finitely- Locally Bounded First

Generated Separable countable
Top. Vec. Spaces ///// Lp 0 < p < 1 ℓp [0, 1] (ℓp )N (ℓp )R
Locally Convex ///// Lp p > 1 L∞ RN , C(Rn ) RR pointwise, ℓ2weak
2
Inner Product ///// L ℓ2 [0, 1] ///// /////
Locally Compact Rn ///// ///// ///// /////

1. A set is balanced when |λ| 6 1 ⇒ λA ⊆ A.

(a) The image and pre-image of balanced sets are balanced.
(b) The closure and interior are again balanced (if A ∈ T0 ; since λA◦ =
(λA)◦ ⊆ A◦ ); as are the union, intersection, sum, scaling, and prod-
uct A × B of balanced sets.
 Joseph Muscat 2015 2

(c) Hence every set generates largest and smallest balanced sets,
[ \
V =: bal(A) ⊆ A ⊆ Bal(A) := V = { λa : |λ| 6 1, a ∈ A }
V ⊆A V ⊇A
V bal. V bal.
S
Bal(A) is open if A is, and V V ◦ is balanced open in A; hence X
has a topological base of balanced open sets.
(d) Balanced sets are star-shaped hence path-connected.
2. X is path-connected and locally connected. There are no open subspaces
(clopen) except for X.
3. Connected open sets are path-connected (since a boundary point of a
path-connected component would be surrounded by a balanced open set).
4. A convex set is one which contains every line segment joining any two of
its points,

0 6 t 6 1 ⇒ (1 − t)C + tC = C
⇔ 0 6 s, t ⇒ sC + tC = (s + t)C

For example, subspaces.

(a) Convexity is preserved by linear images and pre-images.
(b) Convex sets are connected.
(c) Convex(A + λB) = Convex(A) + λConvex(B)
(d) The closure, interior, sum, scaling, and product are convex (e.g. tC ◦ +
(1 − t)C ◦ is open in C).
(e) The intersection of convex sets is again convex; hence every set gen-
erates its convex hull, the smallest convex set containing it,
X
Convex(A) = { t1 a1 + · · · + tn an : ti = 1, ti > 0, ai ∈ A }.
i

(f) If A open or balanced, then so is Convex(A) (but Convex(A)◦ 6=

Convex(A◦ )).
If A is convex, then so is bal(A) (not Bal(A)). S
If K1 , K2 are compact convex, then so is Convex(K1 ∪K2 ) = t∈[0,1] (1−
t)K1 + tK2 (as the continuous image of [0, 1] × K1 × K2 ).
A polyhedron is Convex(F ) of a finite set F ; a simplex is when F is
independent. A cone is Convex(A ∪ { x }) where A is of dimension n − 1.
But the convex hull, even of a compact set, need not be closed (e.g. the
PN
compact set of sequences xn := (1, . . . , n1 , 0, . . .) and 0; then n=1 xn /N →
( n1 )); convex sets (6= X) may be dense in X (e.g. c00 in ℓ1 , { f ∈ C[0, 1] :
f (0) = 1 } in L1 [0, 1]).
 Joseph Muscat 2015 3

5. An extreme subset A of a convex set C satisfies Convex(C rA) ⊆ C rA;

the intersection of extreme sets is extreme. In particular extreme points
do not
P lie on proper line segments in C, e ∈ / Convex(C re), equivalently,
e = i ti ai ⇒ ∃i, e = ai ( ⇒ ∀i, e = ai ).
If C = Convex(E), then E is minimal ⇔ E is the set of extreme points.
There need not exist any extreme points, and the set of extreme points
need not be closed.
6. Recall that a set is bounded when λA → 0 as λ → 0, i.e.,

∀U ∈ T0 , ∃r > 0, Br A ⊆ U.

For a balanced set this is equivalent to ∃λ, A ⊆ λU . The only bounded

subspace is 0.
Given a fixed open set U ∈ T0 , the extent of a bounded set can be gauged
by
NU (A) := inf{ r > 0 : A ⊆ rU }
(a) NU (λA) = |λ| NU (A) when U is balanced
(b) NU (A + B) 6 NU (A) + NU (B) when U is convex
(c) NU (A) = 0 ⇔ A = { 0 } when U is bounded
7. Any balanced convex open neighborhood of 0 generates a semi-norm NC (x)
and conversely, C = { x ∈ X : NC (x) < 1 }. (But there need not exist any
non-trivial ones.)
S
8. If Ti : X → Y are morphisms such that iSTi x is bounded for all x ∈ K a
non-meagre bounded convex subset, then i Ti K is bounded.
S
Proof: Let Ac := { x ∈ K : ∀i, Ti x ∈ cW̄ } closed; then K = c Ac , so
some Ac contains an interior point x0 + V . But K ⊆ x0 + V /t for some
t < 1, so xt := tx + (1 − t)x0 ∈ K ∩ (x0 + V ), then tK ⊆ Ac + (1 − t)Ac ,
so tTi K ⊆ cW̄ + cW̄ ⊆ cU .

Dual Space
1. The dual space is X ∗ := B(X, F). A linear map φ : X → F is continuous
iff ∃V ∈ T0 , |φV | 6 1.
For any balanced convex C ∈ T0 , let NC (φ) := inf{ r > 0 : |φC| 6 r } =
supNC (x)<1 |φ(x)|; then |φ(x)| 6 NC (φ)NC (x).
(X × Y )∗ ∼
= X ∗ × Y ∗ via (φ, ψ)(x, y) := φx + ψy.
Note: When F = C, the real and imaginary parts of a functional are not
independent: Im φ(x) = − Re φ(ix).
 Joseph Muscat 2015 4

2. Every linear map Y → F, which is bounded with respect to some semi-

norm, NC (φ) < ∞, can be extended to all of X with NC (φ̃) = NC (φ).
Proof: φ can be extended from Y to Y + [[v]] by φ(y + λv) = φ(y) +
λc for some c ∈ F. Given |φ(y)| 6 N (φ)N (y); require a c such that
|φ(y) + c| 6 N (φ)N (y + v), which is possible when φ is real-valued since
φ(y1 ) − φ(y2 ) 6 N (φ)(N (y1 + v) + N (y2 + v)). For complex φ = φ1 + iφ2 ,
then φ2 (y) = −φ1 (iy), so both can be extended. Let φ̃ be a maximal
extension of φ (exists by Hausdorff’s maximality); its domain is X else
can extend further by the above.
3. Weak convergence: Every pair (x, φ) ∈ X × Y ∗ gives a functional on
operators: (x, φ) 7→ φT x. Hence they induce a ‘weak’ convergence
Ti ⇀ T ⇔ ∀x ∈ X, ∀φ ∈ Y ∗ , φTi x → φT x,
In particular,
xi ⇀ x ⇔ ∀φ ∈ X ∗ , φ(xi ) → φ(x),
φi ⇀ φ ⇔ ∀x ∈ X, φi (x) → φ(x) (weak-*).

The topology induced by this convergence is generated from the sub-basic

balanced convex open subsets Ur,x,φ := { T : |φT x| < r }, hence is locally
convex but not necessarily T0 , nor locally bounded (Ur,x,φ ⊇ ker φ) except
when finite dimensional. However, X ∗ is a T0 topological vector space
since X separates points of X ∗ .
Morphisms preserve weak convergence, xi ⇀ x ⇒ T xi ⇀ T x.
Note that if Ti → T in Y X , pointwise, i.e., ∀x, Ti x → T x, then Ti ⇀ T .
w
xn → x ⇔ xn ⇀ x and { xn : n ∈ N } is totally bounded; Ā ⊆ A .
Many properties of subsets have weak analogues e.g. weakly bounded when
∀φ ∈ X ∗ , φA is bounded in F (A bounded ⇒ A weakly bounded).
4. If Ti ⇀ T and Si x → Sx, ∀x then Ti S ⇀ T S; if φSi → φS then Si Ti ⇀
ST .
5. There are links between a space and its dual, via the adjointly related
polar of a subset in X and the pre-polar of a subset in X ∗ ,
⊥
A ⊥ := { φ ∈ X ∗ : NA (φ) = sup |φA| 6 1 } = ConvexBal(A)
⊥
Φ := { x ∈ X : sup |Φx| 6 1 }
Φ ⊆ A ⊥ ⇔ |ΦA| 6 1 ⇔ A ⊆ ⊥ Φ
A ⊥ is balanced, convex, and weak-closed in X ∗ (and ⊥ Φ in X).
When U ∈ T0 (X), U ⊥ is weak*-compact.
X
Proof: J : U ⊥ → BF (compact), φ 7→ (φx)x∈X is clearly an embedding.
J(U ⊥ ) is closed: J(φi ) → f ⇔ ∀x, φi x → f (x), hence f is linear with
∀x ∈ U, |f (x)| 6 1, so f ∈ U ⊥ . Thus U ⊥ is compact in X ∗ .
 Joseph Muscat 2015 5

6. Similarly, annihilator and pre-annihilator

⊥
A⊥ := { φ ∈ X ∗ : φA = 0 } = [[A]] ,
⊥
Φ := { x ∈ X : Φx = 0 } = ⊥ [[Φ]], Φ ⊆ A⊥ ⇔ ΦA = 0 ⇔ A ⊆ ⊥ Φ
(A ∪ B)⊥ = A⊥ ∩ B ⊥ , A⊥ + B ⊥ ⊆ (A ∩ B)⊥ .

They are weak-closed subspaces of X ∗ and X respectively. For A un-

bounded, A ⊥ = A⊥ .
7. Every morphism T : X → Y has an adjoint morphism T ∗ : Y ∗ → X ∗
defined by T ∗ φ := φ ◦ T .
Then

T A ⊆ B ⇒ T ∗ B ⊥ ⊆ A⊥ ,
ker T ∗ = (im T )⊥ , (S + λT )∗ = S ∗ + λT ∗ ,
im T ∗ ⊆ (ker T )⊥ , (ST )∗ = T ∗ S ∗ .

T 7→ T ∗ is not weakly continuous but Ti∗ ⇀ T ∗ ⇒ Ti ⇀ T .

8. A continuous projection (idempotent) on a complete space decomposes
it into the product of closed subspaces X ∼
= M × N (M = ker P , N =
im P = ker(1 − P )).
9. If M is a closed subspace of finite codimension, then X ∼
= M × N (using
representatives πn xn = en ).

Separability
The size of a space can be assessed by the minimum cardinality of a set A
such that X = [[A]].
1. X is separable ⇔ A is countable.
Pn
Proof: For any x + U , let V + · · · + V + W ⊆ U , i=1 λi ai P ∈ x + W;
then
P ∃ǫ i , B ǫ i a i ⊆ V , and ∃q
P i ∈ Q + iQ, qi ∈ λi + B ǫ ; thus i qi ai ∈
i (λi + B ǫ )a i ⊆ x + W + i V ⊆ x + U .
P
2. A topological basis is a list of vectors en such that every x = n αn en for
some unique αn . More strongly, en is a Schauder basis when x 7→ αn (x)
are continuous. Such spaces are essentially P sequence spaces x ↔ (an ). A
functional is then of the form φx = n bn an (where bn = φen ).
3. For a separable vector space, U ⊥ (U ∈ T0 ) is a compact metric space.
P
Proof: If xn are dense in U , then kφkw := n 21n |φxn | is a metric on U ⊥ ,
with φi ⇀ φ ⇔ kφi − φkw → 0.
 Joseph Muscat 2015 6

1.1 Quasi-Normed Spaces

are vector spaces with topology induced by a translation-invariant metric d(x, y) =
|x − y|, equivalently, first countable; axiomatically, this quasi-norm satisfies

|x + y| 6 |x| + |y|, |−x| = |x|, |x| = 0 ⇔ x = 0

λn → λ and xn → x ⇒ |λn xn | → |λx|

This last condition can be achieved if, for example, |λx| 6 |λ||x|. Note that
by starting with a balanced local base, the quasi-norm can be chosen to also
be balanced, i.e., |λ| 6 1 ⇒ |λx| < |x| (see the construction of the norm in
topological groups). As in groups, can be completed. A topological vector space
may have more than one inequivalent quasi-norm.
P |(anm )|1
• RN . More generally, arrays of real numbers such that |(anm )| := n 21n 1+|(a nm )|1
,
P
where |(anm )|1 := m |anm | are finite.
R
• L0 (A) with |f |E := E (|f | ∧ 1), i.e., sub-basic open sets Vǫ,δ := { f : µ{ x :
|f (x)| > δ } < ǫ }.
• If πi : Y → Xi are linear maps to a finite number of quasi-normed spaces
(one ofP the πi is 1-1), then the vector space Y can be given the quasi-norm
|y| := i |πi y|.
• Products have the quasi-norm |(x, y)| = |x|+|y| (among others); for count-
P |x|n
able products can take |x| = n 21n 1+|x| .
n

• Quotients have the quasi-norm |x + M | = inf a∈M |x + a|.

1. As in all normed groups, the quasi-norm is continuous and Br +Bs ⊆ Br+s .

The norm constant of concavity is

|x + y|
c := sup 6 2.
|x| ∨ |y|

(But xi ⇀ x 6⇒ kxi k → kxk.)

2. By continuity of scalar multiplication, ∀r, ∃ǫ, s, t < ǫ ⇒ tBs ⊆ Br .
3. The open mapping theorem of topological groups applies S between com-
plete quasi-normed spaces even if not separable: T X = n nT Br , so T Br
contains some open ball; the remaining part of the proof remains valid.
In particular, a bijective morphism is an isomorphism.
4. Closed Graph Theorem: A linear map is continuous iff its graph is closed
in X × Y , i.e., xn → x and T xn → y ⇒ y = T x.
Proof: The graph is itself complete quasi-normed; the projection πX : G →
−1
X is an isomorphism by the open mapping theorem, and T = πY ◦ πX .
 Joseph Muscat 2015 7

5. Isomorphism Theorems for complete spaces: X/ ker T ∼

= im T if im T is
closed (via the continuous map x + ker T 7→ T x).
Hence X+Y ∼Y = X , X×Y ∼
X∩Y = X,
Y
X/Z ∼ X
= . Y /Z Y

6. The totally bounded sets are the metrically bounded sets that are arbi-
trarily close to finite-dimensional subspaces.
Proof: K ⊆ F + Bǫ ⊆ [[F ]] + Bǫ . Conversely, if K ⊆ Br and K ⊆ Y + Bǫ ,
then K ⊆ Y ∩ Br+ǫ + Bǫ ⊆ F + B2ǫ since in finite dimensions balls are
totally bounded.

1.2 Locally Bounded Spaces

when there is a bounded open set; equivalently, a single (balanced bounded)
set B generates the topology by translations and scalar multiplications, x + λB
(λ 6= 0). Hence is first countable.
Examples:
• ℓp and Lp (A) (p > 0).
Quotients are again locally bounded. An infinite product of topological
vector spaces is not locally bounded.
S
1. X = NB = n nB
2. There is a c > 0 such that B + B ⊆ cB; rB + sB ⊆ c(r ∨ s)B.
Proof: V + V ⊆ B, and rB ⊆ V , so r(B + B) ⊆ B.
3. There is an equivalent quasi-norm satisfying |λx| = |λ|p |x| (0 < p 6 1,
cp = 2).
P P
Proof: Let |x| := inf{ ni=1 ν(xi ) : i xi = x }, ν(x) := NB (x) p
Pn, ν̄(x) :=
r
2 P> ν(x). Note ν(x + y) 6 2(ν(x) ∨ ν(y)). Claim: ν( i=1 xi ) 6
2 i ν̄(xi ), since take ν(xi ) in decreasing order; if ν(xj ) 6 2ν(xj+1 ) then
ν(xj +xj+1 ) 6 2ν(xj ) 6 ν̄(xj )+ ν̄(xj+1 ); if 2ν(xi+1 ) 6 ν(xi ) for all P i, then
2 n
ν(x1 + · · P
· + xn ) 6 2ν(x
P 1 ) ∨ 2 ν(x2 ) ∨ · · · ∨ 2 ν(xn ) = 2ν(x1 ) 6 2 i ν̄(xi ).
Hence ν( i xi ) 6 4 i ν(xi ) and 14 ν(x) 6 |x| 6 ν(x).
4. A subset is bounded iff metrically bounded, i.e., covered by some x + rB.
1/p x
5. Every vector has a magnitude and direction (unit vector): x = |x| |x|1/p
.
P
6. If en are bounded and (an ) ∈ ℓp then n an en converges absolutely.
7. A linear map is continuous iff
(a) ∃c > 0, T BX ⊆ cBY . It can be measured by N (T ) := NBY (T BX )
(b) T maps bounded sets to bounded sets (“bounded map”).
 Joseph Muscat 2015 8

N (0) = 0, N (I) = 1, N (T −1 ) > N (T )−1 .

1
Proof: If xn → 0 then T xn = |xn | p T |xnx|n1/p → 0.

8. For every proper closed subspace Y and 0 6 c < 1, there is a unit x such
that |x + Y | = c. The cosets of Y up to a distance of 1 intersect the unit
sphere.
Proof: Let |y + Y | = c; the image of the map z 7→ |y + z|, Y → R,
contains ]c, ∞[, hence some |y + z| = 1.
9. The boundary of Br is Sr := { x : |x| = r }, so B r = { x : |x| 6 r };
w
moreover S r = B r in infinite dimensions.
Tn
Proof: Any neighborhood T i=1 Vǫi ,φi of x ∈ B contains the infinite di-
mensional subspace Y := i ker φi . So there is a unit y ∈ S such that
y+Y = x+Y.
10. Balls are not totally bounded except in finite dimensions. Infinite dimen-
sional totally bounded sets have no interior.
Proof: If B ⊆ Y + ǫB and Y 6= X then there is x ∈ B, |x + Y | > ǫ.

2 Locally Convex Spaces

when there is a base of convex open sets (can be assumed balanced).
Examples:
• RA with sub-base Vx,n := { f : A → R, |f (x)| < n1 }.
S
• C(Ω) with Ω = n Kn a σ-compact topological space, and with the sub-
1
base Vn,m := { f ∈ C(Ω) : |f Kn | < m }.
1
• C ∞ (Ω), with sub-base Vn,k,m := { f ∈ C ∞ (Ω) : |f (k) Kn | < m }.
• B(X, Y ) for topological vector spaces, with weak topology (and indistin-
guishable morphisms identified). In particular, dual spaces X ∗ .

1. If A is bounded or totally bounded, then so is Convex(A).

Proof: A ⊆ F +V ; Convex(F ) ⊆ F ′ +V as a compact set; so Convex(A) ⊆
F′ + V + V ⊆ F′ + U.
2. Separating hyperplanes: A compact convex set K and a disjoint closed
convex set C can be separated by a real functional, φK < α < φC. In
particular X ∗ separates points from closed subspaces.
Proof: A point x can be separated from an open convex set U ∈ T0 using an
extension of the functional φ(λx) := λ; φ is continuous since |φbal(U )| 6 1.
K and C can be separated by (K +V )∩(C +V ) = ∅, V convex; let x0 ∈ K,
 Joseph Muscat 2015 9

y0 ∈ C; x0 − y0 can be separated from the open convex neighborhood

U := (K −x0 +V )−(C −y0 +V ). Hence φ(K +V )−φ(C +V ) = φU −1 < 0,
so φ(K + V ) < φ(C + V ).
3. A closed convex set is weakly closed (if x ∈
/ C̄ then can find φ that sepa-
rates x from C).
Hence, if xi ⇀ x then ∃yi ∈ Convex(xi ), yi → x.
4. ⊥
(A ⊥ ) = ConBal(A), ⊥ (A⊥ ) = [[A]],
w w w
( ⊥ Φ) ⊥ = ConBal(Φ) , (⊥ Φ)⊥ = [[Φ]] ; hence im T ∗ = (ker T )⊥ .
Proof: If x ∈
/ CB(A) =: F ∋ 0, it can be separated from it by a functional,
φF < α < φx; so ψ := φ/α extended to F, satisfies |ψF | < 1 < |ψx| since
F is balanced; so ψ ∈ A ⊥ and x ∈
/ ⊥ (A ⊥ ).
5. Weakly bounded subsets iff bounded.
Proof: |x∗∗ φ| 6 cφ for each x ∈ A; for φ ∈ V ⊥ compact convex, |V ⊥ x| =
|x∗∗ V ⊥ | 6 c; ∴ 1c A ⊆ ⊥ (V ⊥ ) = V̄ ⊆ U .
6. A functional achieves its largest value on a compact convex subset (as |φ|
or Re φ) at an extreme point.
Proof: If |φ| takes its max value α at b, and x = sa + tb ∈ K then
φ(x) 6 sφ(a) + tα, so φ(a) = a = φ(b).
7. A compact convex set has extreme points and they generate the set:
Convex(E) = K.
Proof: For any extreme set A (starting with K), as long as it has distinct
points, can find φ ∈ X ∗ which separates them. Let φ achieve its maximum
α on the closed set F ; then F is an extremeTsubset. Hence can form a
maximal nested chain of extreme closed sets; i Fi is closed extreme and
minimal, hence contains a single (extreme) point. If x ∈ K rC(E) then
a functional separates them, φ(x) > φC(E), so the max of φ contains an
extreme point not in E.
8. Every finite dimensional subspace M induces a decomposition X ∼
= M ×N
(using the dual functionals δi ).
9. A linear map T : X → Y is continuous when for any open convex D ⊆ Y ,
there is an open convex U ⊆ X, such that NV (T U ) < ∞.
10. X is embedded in X ∗∗ .
Proof: x 7→ x∗∗ is 1-1 since for x 6= 0, let x =
6 U convex, so separate x
from U by a functional φ; x∗∗ (φ) = φ(x) 6= 0, so x∗∗ 6= 0.
P Q P
11. ( i Xi )∗ ∼
= i Xi∗ , via (φi ) 7→ i φi .
12. If there is a countable base of convex balanced sets Cn , then the space is
P NCn (x)
quasi-normed by |x| := n 21n 1+N C (x)
.
n
 Joseph Muscat 2015 10

13. Let K be a compact convex subset of X, and T : K → K is continuous and

affine, then T has a fixed point T x = x (proof: let Tn := (1+. . .+T n−1 )/n,
so Tn K is compact; so ∃x ∈ K, ∀n, x ∈ Tn K ie ∃xn , x = Tn xn ; so x−T x =
(xn − T n xn )/n → 0 since xn − T n xn ∈ K + K is compact). CHECK
14. If K convex compact and f : K → K continuous then f has a fixed point
f (x) = x; (also, amenable locally compact T2 groups acting continuously
on a convex compact set has a fixed point Gx = x)
Proof: K ⊆ F +V ⊆ [[F ]]+V ; let fV := πV ◦f : Convex(F ) → Convex(F ).
Then by Brouwer’s fixed point theorem, fV (xV ) = xV ∈ Convex(F ). For
some subsequence, xn → x∗ , hence

x∗ − f (x∗ ) = x∗ − xn + fVn (xn ) − f (xn ) + f (xn ) − f (x∗ ) ∈ V + V + f V ⊆ U

3 Normed Spaces
have scale-homogeneous norms kλxk = |λ| kxk; equivalently they are the locally
convex locally bounded vector spaces (with norm NB (x)). The unit ball BX
generates the topology via the convex bounded balls Br (x) = x + rBX . As in
quasi-normed spaces, can be completed (called a Banach space).
Examples:
• ℓ∞ , the space of bounded sequences, with k(an )k∞ := supn |an |; its closed
subspace c0 of sequences that converge to 0.
P
• ℓ1 , the space of absolutely summable sequences, with k(an )k1 := n |an |.
• Lp (A), p > 1
• L∞ (A), and its closed subspace of bounded continuous functions Cb (A).
• C(K) with sup norm, K compact T2 . Every Banach space is embedded
in some C(K).

Quotients and finite products are also normed.

1. T : X → Y linear is continuous iff it is Lipschitz, kT xk 6 ckxk.

2. B(X, Y ) is a normed space with kT k = supkxk=1 kT xk,

kT xk 6 kT kkxk

It is complete when Y is. In particular, X ∗ is also a complete normed

space.

kS + T k 6 kSk + kT k, kλT k = |λ|kT k, kIk = 1, kST k 6 kSkkT k

Proof: If Tn is Cauchy, then so are (Tn x).

 Joseph Muscat 2015 11

3. X is isometrically embedded in B(X): fix unit a ∈ X, φ ∈ X ∗ , φa = 1,

let Px := xφ; so x = Px a, T Px = PT x .
4. im T is closed ⇔ im T ∗ is closed, in which case im T ∗ = (ker T )⊥ (weak*-
closed). So T invertible ⇒ T ∗ invertible.
Proof: If φ ∈ (ker T )⊥ , then can define ψ(T x) := φx, extended to all of Y ;
T ∗ ψ = φ. Conversely, let T̃ : X → im T , T̃ x := T x, so T̃ ∗ is 1-1. Separate
C := T̃ BX from any other y by ψ, |ψ T̃ x| 6 r < |ψy| for x ∈ B X ; so
r < kψkkyk 6 1c kT̃ ∗ ψkkyk 6 rc kyk, so kyk > c; hence T̃ BX contains some
open ball, so T̃ is onto, i.e., im T is closed.
5. T is onto ⇔ kT ∗ φk > ckφk,
T is an embedding ⇔ kT xk > ckxk.
Proof: T is onto implies im(T ∗ ) is closed and T ∗ 1-1, hence by the open
mapping theorem, kT ∗φk > ckφk.
6. B X ∗ = BX⊥
, hence weak*-closed bounded subsets of X ∗ are weak*-compact.
∗
(So X is meagre when infinite dimensional.)
Similarly, B X = ⊥ BXP ; its weak topology is metrizable when X ∗ is sepa-
rable (using kxkw := n 21n |φn x|).

7. (Krein) If K is weakly compact, then so is Convex(K).

(Eberlein-Shmulian) Weakly compact iff every sequence has a weakly con-
vergent subsequence.
8. Every Banach space is embedded in C(K) for some compact T2 space
K (take K = B X ∗ ) and hence P embedded in some ℓ∞ (A); and covered
by some ℓ1 (A) (via (ai )i∈A 7→ i ai xi , xi dense in B). For example,
separable Banach spaces are embedded in C(2N ) (Cantor space) and ℓ∞ ,
and covered by ℓ1 .

9. X ∗ is not separable if X isn’t.

Proof: If φn is dense in X ∗ , then |φn xn | > (kφn k − ǫ) for some unit
xn . If M := [[xn ]] 6= X, then ψM = 0 with kψ − φn k < ǫ, so |φn xn | =
|(ψ − φn )xn | 6 ǫ.
10. kx + ker φk = |φx|/kφk (since kφk = supa∈ker φ |λ||φx|/kλx + ak).
kxk = supkφk=1 |φx| = kx∗∗ k, hence X is isometrically embedded in X ∗∗ .
T ∗∗ extends T .

kT k = sup |φT x| = kT ∗k
kφk=1
kxk=1

11. If Y is a closed subspace, then (X/Y )∗ ∼

= Y ⊥ (via φ(x + Y ) := φx) and
∗ ⊥ ∼ ∗
X /Y = Y (via φ 7→ φ|Y ).
12. If Ti satisfy kTi xk 6 cx then Ti are equicontinuous, hence kTi k 6 c.
 Joseph Muscat 2015 12

13. If Ti are weakly bounded, |φTi x| 6 cφ,x , then Ti are bounded, kTi k 6 c.
In particular if Tn ⇀ T then kT k 6 lim inf kTn k.
Proof: kT k = sup |φT x| = sup limn→∞ |φTn x| 6 limn→∞ kTn k.
14. A morphism is called a compact operator when it maps bounded sets to
totally bounded sets; equivalently, if xn is a bounded sequence in X, then
T xn has a Cauchy subsequence; or xn ⇀ x ⇒ T xn → T x.
(a) The space of compact operators forms a closed *-ideal in B(X, Y ).
(b) im T is separable.
S S
Proof: T B ⊆ Tn B+(T −Tn)B ⊆ F +ǫB+ǫB. im T = T n nB = n nT B
separable.
Examples include finite rank operators T : X → FN : they are the only
compact operators with closed range (by open mapping theorem, T B is
open and totally bounded in im T ).
15. A Fredholm operator is a morphism whose kernel is finite dimensional and
image is finite co-dimensional. Its index is

index(T ) := dim ker T − dim(im T )⊥

π R ι
T : X → X/ ker T → im T → Y with R an isomorphism.
The product and adjoint are again Fredholm,

index(ST ) = index(S) + index(T ), index(T ∗ ) = −index(T ).

T is Fredholm ⇔ it is invertible up to compact operators (since T R−1 = I,

R−1 T = I − P ).
If index(T ) = 0 then T is 1-1 ⇔ T is onto.
16. In a space with a Schauder basis, the coefficients depend continuously on
x.
Pn
Proof: Let |||x||| := supn k i=1 αi ei k > kxk, complete; hence I : X|||||| →
Pn Pn−1
Xkk has continuous inverse and |αn (x)| = k αi ei − αi ei k 6
2|||x||| 6 ckxk.

T ⊤ is defined on the space B = { φ : φ ◦ T continuous } ⊆ Y ∗ ; when B is

dense in Y ∗ , then T and T ⊤ are closed, T ⊤⊤ = T ; if T is 1-1 and densely onto,
then T ⊤ is 1-1 and T ⊤ −1 = T −1 ;
⊤

3.1 Reflexive Banach Spaces

are spaces for which x 7→ x∗∗ is an isomorphism X ∗∗ ∼
= X.
qP P
Example: Arrays of numbers with aij = 0 for j > i and k(aij )k := 2
j( i |aij |) <
∞.
 Joseph Muscat 2015 13

Closed qsubspaces, the dual space X ∗ , quotients, countable products with

P 2
k(xn )k := n kxn kXn < ∞ are again reflexive.

1. A⊥ can be identified with ⊥

A; and T ∗∗ with T , since T ∗∗ x∗∗ = (T x)∗∗ .
im T ∗ = (ker T )⊥ .
2. X reflexive iff X ∗ reflexive. (The weak and weak-* topologies of X ∗ coin-
cide.)
Proof: If φ∗∗ ∈ X ⊥ then φx = φ∗∗ (x∗∗ ) = 0, so φ = 0.
3. Weakly closed bounded subsets are weakly compact.
Proof: B X = B X ∗∗ is weak*-compact in X ∗∗ , hence weak compact in X.
w
4. S = B using sequences.
Proof: Let vn ∈ S, kvn − vm k > 12 . Then ∃vn ⇀ v; yn := vn+1 − vn ⇀ 0,
xn := x + kyλnnk yn (λn 6 2) such that kxn k = 1; then xn ⇀ x.

5. Any functional attains its norm somewhere on S.

Proof: Let |φxn | → kφk, xn ∈ B; then for a subsequence, xn ⇀ x, so
φxn → φx and |φx| = kφk; kxk = 1.
6. A weakly closed subset has a closest point to any other point.
Proof: Let kyn − xk → d := inf{ ky − xk : y ∈ F }; yn bounded, so
∃yn ⇀ y; ∴ |φ(y − x)| = limn→∞ |φ(yn − x)| 6 dkφk and ky − xk 6 d.
7. X is weakly complete, i.e., every weakly Cauchy sequences converges
weakly (let Ψ(φ) := limi φxi , so Ψ = x∗∗ ; then xi ⇀ x).
8. Ti ⇀ T ⇒ Ti∗ ⇀ T ∗ .

3.2 Uniformly Convex Banach Spaces

are Banach spaces such that kx + yk/2 → 1 ⇒ kx − yk → 0 uniformly on unit
vectors,
x+y
∀ǫ > 0 ∃δ > 0, ∀x, y ∈ B X , 1 − δ < ⇒ kx − yk < ǫ
2
Example:
• ℓp and Lp (A) 1 < p.

1. The set of extreme points of a closed ball is its sphere.

2. xn → x ⇔ xn ⇀ x and kxn k → kxk.
Proof: yn := xn
kxn k ⇀ x
kxk =: y; let φy = 1 = kφk. Then 1 > |φ( yn +y
2
m
)| →
1, so yn +y
2
m
→ 1, kyn − ym k → 0, and yn → y. Hence xn = kxn kyn →
kxky = x.
 Joseph Muscat 2015 14

3. For any closed convex set, the point closest to x is unique.

Proof: yn ⇀ y, kyk = d; so yn → y. If v is another closest point then
1 6 y+v
2d ← 12 kŷn + v̂n k 6 1; hence kŷn − v̂n k → 0 and y = v.
4. X is reflexive.
Proof: Given unit Ψ ∈ X ∗∗ ; let kφk k = 1, Ψ(φk ) → 1. B X is dense
in B X ∗∗ , so ∃xn , unit, φ(xn ) → Ψ(φ). Then 1 > |φ( xn +x
2
m
)| → 1, so
xn +xm
2 → 1, kxn − x m k → 0, xn → x.

3.3 Inner Product Spaces

p
have a norm induced by an inner product, kxk = hx, xi, where

hx, y + zi = hx, yi + hx, zi, hy, xi = hx, yi,

hx, λyi = λhx, yi, hx, xi = 0 ⇔ x = 0,
hx, xi > 0.

Equivalently, a normed space that satisfies the parallelogram law

2 2 2 2
kx + yk + kx − yk = 2kxk + 2kyk .

Can be completed by taking h[xn ], [yn ]i := limn→∞ hxn , yn i (called a Hilbert

space).
Isometric morphisms preserve the inner product, hP x, P yi = hx, yi. Unitary
morphisms are the automorphisms, i.e., invertible isometries. Conformal mor-
phisms preserve orthogonality hx, yi = 0 ⇒ hT x, T yi = 0; hence are multiples
of isometries.
Example: ℓ2 and L2 (A).
Subspaces, products have inner products:

h(x1 , y1 ), (x2 , y2 )iX×Y := hx1 , x2 iX + hy1 , y2 iY

For a ‘complexified’ real inner product space, X + iX, hx, yi = g(x, y) + iω(x, y)
with g, ω real bilinear non-degenerate forms on X 2 , but g is symmetric and ω
skew-symmetric.
2 2 2
1. (a) kx + yk = kxk + 2 Re hx, yi + kyk .
2 2 2 2
(b) hx, yi = 14 (ky + xk + iky + ixk − ky − xk − iky − ixk ).
(c) |hx, yi| 6 kxkkyk, so the inner product is continuous (but not neces-
sarily weakly continuous). (Take x = hy,xi
hy,yi y + z with hz, yi = 0.)
x+y 2 x−y 2
(d) Uniformly convex (since for x, y ∈ B, 2 + 2 = 1).
2. X ∗ ∼
= X via x 7→ hx, ·i (onto since φ(x)y − φ(y)x ∈ ker φ = x⊥ ).
Hence A⊥ = { x ∈ X : ha, xi = 0, ∀a ∈ A }; A ∩ A⊥ ⊆ 0.
T ∗ acts on X as hT ∗ x, yi = hx, T yi; (λT )∗ = λ̄T ∗ .
 Joseph Muscat 2015 15

3. There are linear orthogonal projections onto closed subspaces, so closed

subspaces are complemented, X ∼ = Y × Y ⊥.
If M, N are complete orthogonal subspaces, then so is M + N =∼ M × N.
To find the best approximate solution for T x = y in x, solve T ∗ T x = T ∗ y
(since y − T x ∈ (im T )⊥ ).
2
4. T ∗ T has kernel ker T , closed image im T ∗ and norm kT k .
5. A frame is a set of (unit) vectors ei such that the norm khei , xikℓ2 (I) is
equivalent to kxk. Then [[ei ]] = X.
2
The associated Fourier seriesP operator∗ F : X → ℓ (I), x 7→ (hei , xi)i∈I is
∗
1-1; its adjoint is F (ai ) = i ai ei ; F F > c > 0 hence has a continuous
inverse.
Each frame has a dual ‘biorthogonal’ frame ẽi := (F ∗ F )−1 ei , with an
associated Fourier operator F̃ = F (F ∗ F )−1 , and hei , ẽj i = δij
X X
∀x ∈ X, x = hei , xie˜i = hẽi , xiei .
i i
P
F̃ F ∗ is an orthogonal projection onto im F ⊆ ℓ2 , so among all i αi ei = x,
kF̃ xkℓ2 6 k(αi )kℓ2 .
Proof: hẽi , xi = hei , (F ∗ F )−1 xi = F (F ∗ F )−1 x. F = F̃ F ∗ F , so im F =
im F̃ .
A Riesz frame is a linearly independent frame (equivalent to an uncondi-
tional Schauder basis)
6. An orthonormal basis is a maximal set of orthonormal vectors ei , hei , ej i =
δij (exists). Hence [[E]] = X (since E ⊥ = 0).
P 2
P
i ai ei converges ⇔ (ai ) ∈ ℓ ⇔ i ai ei converges weakly; hence ei is
a self-dual frame and F is an isomorphism:
X
x= hei , xiei , hx, yi = hF x, F yiℓ2
i

Hence every Hilbert space is isomorphic to some ℓ2 (I), via x 7→ F x; the

separable Hilbert spaces are ℓ2 and Fn .
U∗ D V
7. Any compact operator is diagonalizable T = V DU ∗ , X 7→ ℓ2 7→ ℓ2 7→
Y ; T un = λn vn , T ∗ vn = λn un . Thus, any compact operator can be
approximated by a matrix.
Proof: T ∗ T and T T ∗ share the same non-zero (positive) eigenvalues λ2n →
0, with orthonormal eigenvectors un ; vn := T un are also orthonormal.

Any solution of T x = y is given by hun , xi = hvn , yi/λn , assuming the

latter coefficients are in ℓ2 .
 Joseph Muscat 2015 16

3.4 Symplectic Spaces

are vector spaces with a symplectic form ω : X 2 → R such that

ω(x, y + z) = ω(x, y) + ω(x, z), ω(y, x) = −ω(x, y),

ω(x, λy) = λω(x, y), ∀y, ω(x, y) = 0 ⇔ x = 0.

The symplectic morphisms preserve this form

ω(T x, T y) = ω(x, y)

1. Every symplectic space is isomorphic to some V ×V ∗ with ω((u, φ), (v, ψ)) :=
ψ(u) − φ(v).
2. A⊥ := { x : ω(a, x) = 0 ∀a ∈ A }. A ⊆ B ⊥ ⇔ B ⊆ A⊥ , so A ⊆ A⊥⊥ .
Y is isotropic when Y ⊆ Y ⊥ ; in this case, Y ⊥ /Y is also symplectic. It
can be extended to a Lagrangian subspace, Y = Y ⊥ .
3. Y is a symplectic subspace of X iff Y ∩ Y ⊥ = 0.

4 Finite Dimensional Spaces, RN

They are the locally compact topological vector spaces; equivalently, a totally
bounded open set exists.
Proof: Let K be a compact (bounded) balanced neighborhood of 0; then
1
K⊆F+ T 2 K for some finite F with MS := [[F ]]; so K ⊆ 12 K + M ⊆ 21r K + M ,
1
so K ⊆ r (M + 2r K) = M and X = r 2r K ⊆ M .
X is isomorphic to Euclidean space FN with the inner product hx, yi =
PN
n=1 ān bn . In particular, all norms
PN
are equivalent and complete.
Proof: T : FN → X, (ak ) 7→ k=1 ak ek is continuous, since (ak ) 7→ ai 7→
ai ei is continuous. Conversely, let f (v) := kT vk continuous; then 0 ∈ / fS
compact, where S is the unit sphere of FN , i.e., [0, c[ ⊆ f S, c 6 kT vk/kvk.

1. Totally bounded ⇔ bounded

Compact ⇔ closed and bounded
xn → x ⇔ xn ⇀ x
T linear are compact and Fredholm.
2. If K is compact then so is Convex(K).
 
P 1 ··· 1
Proof: Let x = i ti vi ; the matrix has a null vector if k >
P P v1 · · · vk P
n + 1, i.e.,P∃αi , i αi = 0, i αi vi = 0; β := min ti /|αi |; then i (ti −
βαi ) = 1, i (ti − βαi )vi = x but has less terms.
3. A∗ = Ā⊤ . Unitary matrices have orthonormal columns.
 Joseph Muscat 2015 17

α
4. The Hausdorff measure satisfies µα (λE) = |λ| µα (E). Also µα+β (E ×
F ) > cα,β µα (E)µβ (F ). Borel sets are µα -measurable; countable sets are
µα -null.
Normalized µn (n ∈ N) are called Lebesgue measures: cardinality, length,
area, volume, etc..
5. The dimension of E is dim(E) := inf{ α : µα (E) = 0 }.
dim(A ∪ B) = max(dim A, dim B),
A ⊆ B ⇒ dim A 6 dim B,
dim(E × F ) 6 dim E + dim F .
 Joseph Muscat 2015 18

5 Topological Algebras over R or C

A topological algebra is a topological ring +, λ, · that contains F in its center.
Thus it is a topological vector space with continuous +, λ, ·.
The morphisms are those maps which preserve +, λ, ·,

φ(x + y) = φ(x) + φ(y), φ(λx) = λφ(x), φ(xy) = φ(x)φ(y)

must be continuous with kφk = 1 (the automorphisms form a closed Lie sub-
group of GL(X) with Lie algebra Der(X)). The morphisms X → C (if there
b
are any) are called characters; they form the set X.
Examples:
• RA with f g(x) := f (x)g(x).
• B(X) for X a topological vector space.
Products are again a topological algebra.

1. 1

6 Normed Algebras
A normed algebra is a topological algebra with a norm such that

kx + yk 6 kxk + kyk, kλxk = |λ|kxk,

kxyk 6 kxkkyk, k1k = 1

Can be completed so that [xn ][yn ] = [xn yn ]; it is then called a Banach
algebra. If kxyk 6 ckxkkyk then there is an equivalent norm with c = 1.
Examples:
1. C(K) with K compact.
2. L1 (G) with convolution; in particular, ℓ1 = L1 (Z).
3. Cn with convolution and 1-norm.
4. B(X) for X a Banach space; contains the closed ideal of compact oper-
ators. Every normed algebra is embedded in some B(X) via a 7→ La ,
La (x) := ax.
5. H quaternions, with absolute value as norm.
Products are again normed algebras (with ∞-norm).
 Joseph Muscat 2015 19

1. The state space is S(X) := { φ ∈ X ∗ : φ1 = 1 = kφk }, a weak*-compact

convex set.

S(x + y) ⊆ Sx + Sy, S(x + λ) = S(x) + λ, S(λx) = λSx, S1 = { 1 }

Proof: S is weak*-closed in the weak*-compact B X ∗

2. The spectrum of an element is σ(x) := { λ ∈ C : x − λ is not invertible }.
It is a non-empty compact subset of C, with largest extent ρ(x) and small-
est extent ρ(x−1 )−1 (or 0). It depends continuously on x:
Proof: σ(x)c = f −1 GL(X) open; if |λ| > ρ(x) then ρ(x/λ) < 1, so x − λ =
−λ(1−x/λ) is invertible. If xn → x, then σ(xn ) is eventually in σ(x)+ǫB.
k(x − λ)−1 k > 1/d(λ, σ(x)). When an algebra is enlarged, the interior of
σ(x) decreases, and its boundary increases; ultimately, the result is the
‘singular spectrum’ of x − λ that are topological divisors of zero.
b is weak*-compact in S,
3. The character set X
b + y) ⊆ Xx
X(x b + Xy,
b b
X(xy) b Xy),
⊆ (Xx)( b b = {1}
X1

b ⊆ σ(x) ⊆ Sx ⊆ kxkB̄
Xx

Proof: X b is weak*-closed. If y := x − λ is not invertible, then 1 ∈/ [[y]], so

b and y is invertible, then
there is a φ ∈ S, φ[[y]] = 0, i.e., φx = λ. If φ ∈ X
φx − λ = φy 6= 0.
4. The extreme points of S are called pure states, SE , and their weak*-closure
W̄ . They generate the state space
w
S = Convex(SE ) , Sx = Convex(SE x)

Thus the largest value of Sx is achieved by a pure state.

5. Except for X = C, there are non-zero topological divisors of zero (else as
σ(x) has non-empty boundary, x = λ ∈ C).
6. a is a quasi-nilpotent (or radical element), i.e., 1 − xa is invertible for all
x, iff ρ(xa) = 0, ∀x. Then σ(x + a) = σ(x).
Proof: y + a = y(1 + y −1 a) is invertible since ρ(y −1 a) = 0, so λ ∈
/ σ(x +
a) ⇔ 0 ∈/ σ(x − λ).
7. If f is analytic on an open set around σ(x), then define
I
1
f (x) := f (z)(z − x)−1 dz
2πi

(a) ax = xb ⇒ f (a)x = xf (b), so f (x−1 ax) = x−1 f (a)x

(b) xy = yx ⇒ f (x)g(y) = g(y)f (x).
 Joseph Muscat 2015 20

(c) The map f 7→ f (x) is a Banach-algebra-morphism C ω (σ(x)) → X.

b ψf (x) = f (ψx).
(d) σ(f (x)) = f (σ(x)); for ψ ∈ X,

Proof: If d(λ, f σ(x)) > 0, then (f (z) − λ)−1 is analytic. If f (x) − f (λ)
has an inverse y, then (x − λ)F (x)y = 1 = yF (x)(x − λ), where F (z) =
(f (z) − f (λ))/(z − λ).
If x satisfies f (x) = 0, then σ(x) ⊆ { λ : f (λ) = 0 }. For example,
idempotents have spectrum { 0, 1 }; nilpotents { 0 }.
8. If f is analytic on an open annulus Rr then it is a Laurent series with
1
H kf k
coefficients an = 2πi f (z)z −1−n dz (so |an | 6 Rn∞ for n ∈ N).
For σ(x) ⊂ Rr,
X∞
f (x) = an xn
n=−∞
P
Proof: (z − x)−1 = n xn /z 1+n .
9. If σ(x) = σ1 ∪ · · · ∪ σn , each enclosed by a simple curve, then there are
idempotents ei := 1σi (x), such that 1 = e1 + · · · + en , σ(xei ) = σi .
10. Exponential function
x2 xn x
ex := 1 + x + + ···+ + · · · = lim (1 + )n
2 n! n→∞ n
d tx
(a) e0 = 1, (ex )−1 = e−x , enx = (ex )n , dt e = etx x.
x y 1
(b) ex+y = limn→∞ ((1 + 2n )(1 + n
2n )) ; e e = ex+y+ 2 [x,y]+...;
x y
x+y
if xy = yx then e = ex ey .
(c) ex = cosh x + sinh x, even/odd parts. tanh x := sinh x(cosh x)−1 .
(d) The exponential function is periodic with purely imaginary period
τ i; π := τ /2. Then
eiπ + 1 = 0
(e) e2πix = cos(2πx) + i sin(2πx), so sin(x + y) = sin x cos y + cos x sin y,
cos(x + y) = cos x cos y − sin x sin y;
11. For any continuous derivative D, etD is an automorphism of X; in partic-
ular etDx y = etx ye−tx .
P 1 n n P 1 n n P 1 m m
Proof: etD (xy) = n n! t (D xy+· · ·+xDn y) = n n! t D x m m! t D y.
x2 x3
12. Logarithm function For ρ(x) < 1, let ln(1 + x) := x − 2 + 3 + ··· +
(−1)n+1 n
n x + ···.
n ln(1+x)
Then e = (1 + x)n , so let (1 + x)p := ep ln(1+x) (p ∈ C), then
 
p p(p − 1) 2 p n
(1 + x) = 1 + px + x + ··· + x + ···
2! n
 Joseph Muscat 2015 21

More generally, given

R z any simple path “branch cut” from 0 to ∞ (typically
−R+ ), let ln z := 1 w1 dw (along a path that does not intersect the branch
cut). Then eln x = x = ln ex , xp := ep ln x
b where F (x) = x̂, x̂(ψ) := ψx ∈ σ(x),
13. Gelfand Transform: F : X → C(X),
is a morphism,

x[
+ y = x̂ + ŷ, c = λx̂,
λx cy = x̂ŷ,
x b
1 = 1, fd b.
(x) = f ◦ x

The kernel of F contains all elements with ρ(x) = 0 and all commutators.

6.1 B(X)
1. An morphism J : B(X) → B(Y ) induces a morphism L : X → Y ; if J is an
isomorphism, then so is L, with J(T ) = LT L−1. Hence all automorphisms
of B(X) are inner; they form the Lie group GL(X).
Proof: X ⊂ ∼ B(X) via x 7→ Px . J(Pa ) = bψ = Pb for some unit b, ψ,
ψb = 1, since they have the same kernel and image. Hence J(Px ) =
J(Px Pa ) = J(Px )Pb = PJ(Px )b ; L(x) := J(Px )b; invertible when J is.
2. The center of B(X) is F.
Proof: T (xφ) = (xφ)T , so T x = λx.
3. There are no proper radical elements: For every T 6= 0 there is S := xφ
such that (1 − ST )x = 0, so 1 ∈ σ(ST ).
4. There are no characters unless X = C.
Proof: Let M be a two-dimensional (complemented) subspace, and Eij
a basis for B(M ). Then Eii Ejj = 0, Eii Eij = Eij , Ejj = Eij Eji , so
ψEij = 0, ∀i, j.
5. The spectrum of T ∈ B(X) splits into the
• eigenvalues when T − λ is not 1-1 (a left divisor of zero);
• the continuous spectrum with T − λ 1-1 and dense (a left topological
divisor of zero);
• the residual spectrum (otherwise; a right divisor of zero).
It includes approximate eigenvalues, i.e., (T − λ)xn → 0 for some unit xn
(i.e., T − λ is a left topological divisor of zero).
6. Distinct eigenvalues have linearly independent eigenspaces.
P Y Y
Proof: If v := n αn en = 0 then 0 = (T − λn )v = αk (λk − λn )ek .
n6=k n6=k

7. σ(T ∗ ) = σ(T ), σr (T ) ⊆ σp (T ∗ ) ⊆ σp (T ) ∪ σr (T ), σc (T ∗ ) ⊆ σc (T ).
When X is reflexive, σr (T ∗ ) ⊆ σp (T ) and σc (T ∗ ) = σc (T ).
 Joseph Muscat 2015 22

8. Recall that if T ∈ B(X) has finite ascent and descent (see Universal Algebras)
then every x ∈ X can be represented uniquely by some T n y, modulo
ker T n , i.e., X = ker T n ⊕ im T n .
9. The compact operators form a closed ideal, so B(X)/K is a Banach alge-
bra; contains the ideal F (X) of finite-rank operators.
10. If K is a compact operator, then 1 + K is Fredholm of finite ascent and
descent, its spectrum is a countable set of eigenvalues whose only possible
limit point is 0, and each non-zero eigenvalue has a finite dimensional
extended eigenspace.
Proof: If 1 + K has infinite ascent/descent, then can choose separated unit
xn ∈ ker(1+K)n or im(1+K)n , so Kxn is not Cauchy. T −λ = λ(1−T /λ).
Similarly, can choose separated unit eigenvectors, so T en = λn en → λen
has no Cauchy subsequence unless λ = 0. (T − λ)n is still Fredholm.
T ∗ has the same non-zero eigenvalues and eigenspace dimensions as T ,
ker(S ∗ ) = im(S)⊥ ∼
= Y / im S ∼
= ker S.

6.2 Commutative Banach algebras

Example: Z(Z(x)) for any x ∈ X.

1. The only simple commutative Banach algebra is C (the closed ideal Xa is

0 or contains 1).
2. The radical consists of elements with zero spectrum, ρ(x) = 0 (since
ρ(xy) 6 ρ(x)ρ(y)).
b 6= ∅.
3. Any maximal ideal is the kernel of some character; so X
Proof: I = ker π for π : X → X/I; if I is maximal, X/I is simple, i.e., C.
4. σ(x + y) ⊆ σ(x) + σ(y), σ(xy) ⊆ σ(x)σ(y) (in Z(Z(x, y))).
b since ker F = J .
5. X/J is embedded in C(X),
b = σ(x),
b = Xx
im x kb
xkC(X) b
b = sup |Xx| = ρ(x), xd b−1 .
−1 = x

Proof: If λ ∈ σ(x) then x − λ ∈ I = ker φ maximal, φx = λ.

6. The Banach algebras that are embedded in some C(K) are those that
2
satisfy kxk 6 ckx2 k for all x. In particular, they are commutative and
have trivial J .
n 2−n
Proof: kxk 6 ckx2 k → cρ(x) = ckb xk, so J = 0; kxyk 6 cρ(yx) 6
ckyxk; let F (z) := e−zx aezx , analytic, then kF (z)k 6 ckak, hence F (z) =
a, i.e., xa = ax.
Those that are isometrically embedded in C(X) b are the commutative semi-
simple Banach algebras, equivalently kx2 k = kxk2 .
 Joseph Muscat 2015 23

7. Dex = ex , D cosh x = sinh x, D sinh x = cosh x, D cos x = − sin x,

D sin x = − cos x.

7 Involution algebras
are the normed algebras with an involution ∗ : X → X,

x∗∗ = x,
(x + y)∗ = x∗ + y ∗ , (xy)∗ = y ∗ x∗ , i∗ = −i,
kx∗ k = kxk

So ∗ is a (continuous) anti-automorphism. A complete involution algebra is

called a C ∗ -algebra. The ∗-morphisms preserve involution φ(x∗ ) = φ(x)∗ .
Example: Cb (R) with f ∗ (t) := f (−t). Products are again involutive with
(x, y)∗ = (x∗ , y ∗ ).
A ∗-sub-algebra/ideal has to be closed under involution.
An element is called normal when x∗ x = xx∗ , i.e., x∗ ∈ Z(x); e.g. x + eiθ x∗ .
It is called self-adjoint when a∗ = a; e.g. x∗ x, x + x∗ , i(x − x∗ ). It is unitary
when u∗ = u−1 ; e.g. x∗ x−1 when x is normal, in particular eia when a is self-
adjoint.

1. 1∗ = 1∗ 1 = (1∗ 1)∗ = 1, so the involution on C is conjugation.

2. (x−1 )∗ = (x∗ )−1 , σ(x∗ ) = σ(x)∗ .
If x is nilpotent, radical, divisor of zero, or topological divisor of zero, then
so is x∗ .
If x∗ x and xx∗ are both invertible then so is x: x−1 = (x∗ x)−1 x∗ =
x∗ (xx∗ )−1 .
3. Any element can be written as a + ib, with a, b self-adjoint, called the real
and imaginary parts; kak, kbk 6 kxk.
x∗ = a − ib, x∗ x = (a2 + b2 ) + i[a, b], xx∗ = (a2 + b2 ) − i[a, b];
x is normal ⇔ ab = ba, unitary ⇔ ab = ba and a2 + b2 = 1.
4. Polarization identity: For ω := e2πi/N ,
N
1 X n
x∗ y = ω (x + ω n y)∗ (x + ω n y)
N n=1

N
1 X
x∗ x + y ∗ y = (x + ω n y)∗ (x + ω n y)
N n=1

5. (a) The closed ∗-sub-algebra generated by x is C[x, x∗ ] (non-commuting

polynomials).
 Joseph Muscat 2015 24

(b) Z(A∗ ) = Z(A)∗ , so Z(A) is a closed ∗-sub-algebra when A∗ = A.

6. The kernel of a ∗-morphism and the radical J are closed ∗-ideals.
7. The normal elements form a closed subset containing C: if x is normal, so
are x∗ , αx, x + α, x±n .
Z(x∗ ) = Z(x). If q ∈ Z(x) is a quasi-nilpotent, then x + q is not normal
unless q = 0.
Proof: For y ∈ Z(x∗ ), let αx = a+ib, F (α) := e−αx yeαx = e−a−ib yea+ib =
∗ ∗
e−2ib ye2ib is bounded kF (z)k 6 kyk, so constant; i.e., eᾱx y = yeᾱx .
8. The self-adjoints form a real closed sub-space (Jordan algebra) containing
R: a + b, (ab + ba)/2 (e.g. b ∈ R), a±n , i[a, b], are again self-adjoint.
9. The unitaries form a closed sub-group of the invertible elements G(X)
(closed under ∗ but not a normal sub-group), containing eiR .

8 C ∗-algebras
2
are ∗-algebras such that kx∗ xk = kxk .
p 2
1. For normal elements, kx2 k = kx∗ xx∗ xk = kxk , so ρ(x) = kxk.
Sx = Convex(σ(x)). The only normal quasi-nilpotent is 0.
Proof: If λ ∈/ Convex(σ(x)) then can separate by a ball z + rB. So
|φx − z| = |φ(x − z)| 6 kx − zk < |λ − z| for φ ∈ S.
p
2. kxk = ρ(x∗ x), so the norm is unique. The involution is also unique.
p
3. Semi-simple: There are no radical elements, as kqk = ρ(q ∗ q) = 0.
4. S preserves involution, φ(x∗ ) = φ(x)∗ , kφk 6 1, and separates points.
Sx∗ = (Sx)∗ .
Proof: If a∗q= a and φ(a) = α + iβ, then |β + t| 6 |φ(a + it)| 6 ka + itk =
2 2
ρ(a+it) = kak + t2 , so (2t+β)β 6 kak and β = 0. φ(x∗ ) = φ(a−ib) =
φ(x)∗ . σ(a) ⊆ S(a) = 0 ⇒ a = 0. kφxk2 = ρ(φ(x∗ x)) 6 ρ(x∗ x) = kxk2 .
c∗ = x
5. The Gelfand transform preserves involution: x b∗ .
6. If x is normal, C[x, x∗ ] ≡ C(σ(x)), via F : p(x, x∗ ) 7→ p(b b∗ ).
x, x
In particular, can define f (x) for any f ∈ C(σ(x)) via f (x) := F −1 f F x.
Then f ∗ (x) = f (x)∗ , σ(f (x)) = f (σ(x)), and if xy = yx then f (x)g(y) =
g(y)f (x). For example, |x|.
7. The self-adjoints are the normal elements with Sa ⊆ R (since φ(a∗ − a) =
0).
Let a 6 b when S(b − a) > 0. Then
 Joseph Muscat 2015 25

(a) α 6 a 6 β ⇔ Sa ⊆ [α, β]
(b) a + c 6 b + c; if a, b > 0 commute, then ab > 0.
(c) a = a+ + a− , |a| = a+ − a− , a+ a− = 0, a− 6 a 6 a+ 6 |a| 6 kak.
(d) a ∨ b = a + (b − a)+ , a ∧ b = a − (a − b)+ ; hence a (+, ∨)-group lattice.
(e) a 6 b ⇒ x∗ ax 6 x∗ bx, in particular x∗ x > 0.
(f) For φ ∈ S, φ(x∗ y) is a semi-inner product, φ(x∗ ax) 6 φ(x∗ x)kak and
|φ(x)|2 6 φ(x∗ x) (since a 6 kak).
(g) If φ 6 ψ, φ ∈ S, ψ ∈ X,b then φ = ψ.
b is part of the extreme points of S.
(h) X
Proof: x∗ x = a+ + a− , so (xa− )∗ (xa− ) = a3− 6 0; let xa− = b + ic, then
0 6 2(b2 +c2 ) = (xa− )∗ (xa− )+(xa− )(xa− )∗ 6 0 and xa− = 0; √hence
3
√a− =
∗ ∗ ∗ ∗
(xa− ) (xa− ) = 0, and x x = a+ > 0. a > 0 ⇒ x ax = ( ax) ( ax).
If φ 6 ψ then |φ(x)|2 6 φ(x∗ x) 6 |ψ(x)|2 , so ker ψ ⊆ ker φ and ψ = φ.
b then |φ1 (x)|2 + |φ2 (x)|2 6 φ1 (x∗ x) + φ2 (x∗ x) =
If ψ = 21 (φ1 + φ2 ) ∈ X,
1
2ψ(x x) = 2 |φ1 (x)+φ2 (x)|2 , hence |φ1 (x)−φ2 (x)|2 = 0 and φ1 = φ2 = ψ.
∗

1 1 1 1 1 1 1 1
For example, 0 6 a 6 b ⇒ b− 4 a 2 b− 2 a 2 b− 4 6 1 ⇒ 0 6 b− 4 a 2 b− 4 6
1 1
1 ⇒ 0 6 a 2 6 b 2 . A map which preserves + and ∗ automatically
preserves 6 (since a 6 b ⇔ b − a = x∗ x). A bijective ∗-morphism is an
isomorphism.
8. For unitary u,

(a) kuk = 1, kuxk = kxk = kxuk.

(b) They are the normal elements with σ(u) ⊆ eiR .
(c) The inner automorphism by αu is a ∗-automorphism.

Proof: σ(u−1 ) = σ(u∗ ) = σ(u)∗

9. A normal element is idempotent iff self-adjoint with σ(e) ⊆ { 0, 1 }.
10. Polar decomposition:
√ Every invertible element can be written uniquely as
x = ur, where r = x∗ x > 0, u := xr−1 unitary.
11. Every C ∗ -algebra is embedded in some B(H).
Proof: Map a ∈ X to Ja : (xφ )φ∈S 7→ (axφ )φ∈S , where xφ is a coset
of Mφ := { x : φ(x∗ x) = 0 }. Hence X embeds in B(ℓ2 (X/Mφ )). Note
hxy, zi = hy, x∗ zi.
A state ψ is pure iff for any state φ, 0 6 λφ 6 ψ ⇒ φ = αψ.
Proof. If ψ = tψ1 + (1 − t)ψ2 , then 0 6 tψ1 6 ψ, so tψ1 = λψ so
ψ1 = ψ = ψ2 .
Conversely, if 0 6 φ 6 ψ then 0 6 φ1 6 1; if φ1 = 0 then |φT | 6 φkT k = 0
so φ = 0; if φ1 = 1 then (ψ − φ)1 = 0 so ψ − φ = 0; if 0 < φ1 < 1 then
ψ−φ φ
ψ = (1 − φ1) 1−φ1 + φ1 φ1 , so φ/φ1 = ψ.
 Joseph Muscat 2015 26

12. A tensor algebra is the free (unital) algebra generated by a vector space
V , so that any morphism from V extends to tensors on it.
(a) Every element decomposes into sub-components of different grades
x = α + v + v2 + · · · + with α ∈ F, v ∈ V , v2 ∈ V ⊗ V , etc. The
grade-0 part is called its real part: Re(x) := α; Re(xy) = Re(yx).
1 P
(b) Exterior product: v1 ∧ · · · ∧ vn := n! σ sgn(σ)vσ(1) · · · vσ(n) )

wv − vw
w∧v = = −v ∧ w, v ∧ v = 0
2
T (v1 ∧· · ·∧vn ) := T v1 ∧· · ·∧T vn (in finite dimensions T ω = det(T )ω).
(c) Inversion (an involution) (vr∗ = (−1)r(r−1)/2 vr )

(α + v + v2 + · · · )∗ := α + v − v2 − v3 + · · ·

(d) The algebra splits in two parts X + ⊕ X − , i.e., the even and odd
grades: x = x+n(x)
2 + x−n(x)
2 , where n : v 7→ −v. A product of r
vectors gives an element in X ± depending on whether r is even/odd,
so X + is a sub-algebra.
(e) The symmetric algebra is the commutative algebra of the quotient of
tensors by the ideal generated by the commutators; it is isomorphic
to F[V ].
13. Conjecture: The only closed ∗-sub-algebra that separates extreme points
of S is X

8.1 B(H)
1. A ∗-automorphism is of type T 7→ LT L−1 where L is non-zero multiple of a
Hilbert space isomorphism. The isometric ones are the unitary operators.
2. Distinct eigenvalues in σ(T ) and σ(T ∗ )∗ have orthogonal eigenspaces.
Proof: (λ − µ)hx, yi = hx, T yi − hT ∗ x, yi = 0.
3. The mean value of T in the direction x is hx, T xi (it minimizes kT x − λxk;
a functional on T ). The numerical range W (T ) is the set of mean values of
T . W (I) = { 1 }, W (λT + z) = λW (T )+ z, W (T ∗ ) = W (T )∗ , W (S + T ) ⊆
W (S) + W (T ).
W (T ) is a convex subset of C satisfying

σ(T ) ⊆ W (T ) ⊆ S(T ).

Proof: Let 0 < α := d(λ, W (T )) 6 k(T − λ)xk, so T − λ is 1-1 with closed

image; as is T ∗ − λ∗ ; so T − λ is invertible.
 Joseph Muscat 2015 27

4. Uncertainty principle: For a fixed unit x, there is a semi-inner-product,

Cov(S, T ) := hSx, T xi − hSx, xihx, T xi

p
and semi-norm σT := Cov(T, T ), then

| Cov(S, T )| 6 σS σT

σT 6 21 diam(σ(T )), σT = 0 ⇔ x is an eigenvector of T .

5. Normal operators:
(a) kT ∗ xk = kT xk
(b) ker T ∗ = ker T = ker T 2 are T and T ∗ invariant.
(c) im T is dense ⇔ T is 1-1
(d) T is an embedding ⇔ invertible
(e) S(T ) = W (T ) = Convex(σ(T ))
(f) σ(T ) has no residual spectrum, and isolated points are eigenvalues.
(g) Eigenvalues of T and T ∗ are conjugate; no extended eigenvectors.
6. Self-adjoint : S 6 T ⇔ hx, Sxi 6 hx, T xi, ∀x.
√
7. Polar decomposition: Every T = U R, where R = T ∗ T and U (Rx) := T x
is an isometry on im T . Then T ∗ = RU ∗ = U ∗ T U ∗ , kRk = kT k. T is
normal ⇔ R = T U ∗ , unitary ⇔ T = U invertible.

Hence ideals are automatically ∗-ideals since T ∗ = U ∗ T U ∗ .

8. Unitaries: Every unitary is of the type eiA with A self-adjoint.
(U = B + iC, C = V |C|, A := V arccos(B))
2 2 2
Un ⇀ U ⇔ Un x → U x (since kUn x − U xk = kUn xk + kU xk −
2 Re hU x, Un xi → 2kxk2 − 2 Re kU xk2 = 0).
(Stone): any one-parameter group of normal operators which is weakly
continuous in t must be of the type etT with T normal and Re(σ(T ))
bounded above; for unitary operators, eitA ; more generally any unitary
representation of a locally compact
R T2 abelian group which is weakly con-
tinuous in t is of the form Ux = χ(x)dEχ ).
9. Ergodic theorem: If T normal, kT k = 1, then T n x → y (Cesaro) such
that T y = y.

10. Compact operators

(a) B(H) contains the closed subalgebra C ⊕ K.
(b) Every ideal contains the simple ideal KF of finite-rank operators.
 Joseph Muscat 2015 28

(c) The compact operators form the closed ideal K = KF ; so B(X)/K

is simple (its invertible elements are the Fredholm operators). It is
maximal when X ∼ = ℓ2 .
(d) T has a matrix consisting of blocks of type
 
λ
 .. 
1 . 
 
 . . 
 . .
. . 
1 λ
Compact normal operators are diagonalizable.
∗ ⊥ 2
(e) T x =
Py, 1if y ∈ (ker T ) and heσ , yi/σ ∈ ℓ , then the solutions are
x = σ σ heσ , yieσ + ker T , else no solutions.
Proof: Given T ∈ I and T a = b unit; let Exy := xy ∗ for any unit y. Then
Exy = Exb T Eay ∈ I. As a compact operator, on each finite dimensional
eigenspace, T = λ + (T − λ). As kernel basis for the nilpotent A := T − λ
pick u, Au, . . . , An−1 u, etc.
11. There are various closedP
ideals contained in K: Let the trace of an operator
be defined by tr(T ) := i hei , T ei i; it is well-defined independently of ei
when tr(|T |) < ∞.
(a) tr(S + T ) = tr(S) + tr(T ), tr(λT ) = λ tr(T ), tr(T ∗ ) = tr(T )∗ .
(b) Trace class operators: kT k1 := tr |T | < ∞, kT k1 = k(σn )kℓ1 .
2
(c) Hilbert-Schmidt operators: kT k2 :=qtr(T ∗ T ) < ∞; complete inner-
P
product hS, T i := tr(S ∗ T ); kT k2 = 2
ij |hej , T ei i| = k(σn )kℓ2 .
1
(d) Schatten operators: kT kp := (tr |T |p ) p = k(σn )kℓp < ∞.
1 1
(e) Hölder’s inequality: kST kr 6 kSkp kT kq where p + q = 1r .

12. Spectral Theorem: For T normal and f ∈ L∞ (σ(T )),

Z
f (T ) := f (λ)dPλ ∈ B(H)
σ(T )
R
meaning hx, f (T )yi = σT f dhx, P (E)yi, where P (E) is an orthogonal
projection measure, i.e., for any measurable subsets of σT , P (E ∩ F ) =
P (E)P (F ), P (E ∪ F ) = P (E) + P (F ) for E, F disjoint, P (En ) ⇀ P (E)
for En → E, P (σ(T )) = I. f (T ) = U −1 f (λ)U where U : H → H is the
unitary operator x 7→ Pλ x; then
(f + g)(T ) = f (T ) + g(T ), (λf )(T ) = λf (T ), (f g)(T ) = f (T )g(T ),

f¯(T ) = f (T )∗ , f ◦g(T ) = f (g(T )), f[

(T ) = f ◦T̂, kf (T )k 6 kf kL∞ (σ(T ))
Finite Dimensions: Square Matrices
 Joseph Muscat 2015 29

13. The nearest number to a matrix (in the 2-norm) is tr(T )/n.
14. The quasi-nilpotents (radical) are the nilpotents.
15. The matrices with distinct eigenvalues are dense and open in Mn (C) (since
T = D + N is close to D′ + N where D′ has distinct eigenvalues).
16. If p(x) = det(TQ− x), then Q
p(T ) = 0
(since p(T ) = i pi (Ti ) = i Ani i = 0, pi (x) = (x − λ)n ).
17. Self adjoint matrices: If T , with eigenvalues λi , is restricted to P T P where
P is a projection to a sub-space M of one dimension less than M (for
example, by removing the kth row and column), then the new eigenvalues
are interlaced
λ1 6 µ1 6 λ2 6 µ2 6 λ3 6 · · · 6 λn

18. Positive matrices, amn > 0. W (T ) has its largest extent for a positive real
x.
p √
19. n |det T | 6 n maxi,j |Tij |; the maximum is achieved
 by the Hadamard
∗ Hk Hk
matrices: HH = nI, H0 = [1], Hk+1 = )
Hk −Hk

8.2 Commutative C ∗ -algebras

Equivalently, every element is normal.
Examples:
• L∞ (A) of bounded measurable functions, with usual product and f ∗ (a) =
f (a).
• Cb (X), bounded continuous functions, when X is a locally compact T2
space; contains the closed ideal C0 (X). For example, C(K) for K compact;
e.g. C(S), ℓ∞ = Cb (N), Cn = C(n).
• The generated subalgebra Z(A ∪ A∗ ); Z(x) for a normal element.

1. X ≡ C(K) via the Gelfand map. The state space consists of the positive
Radon measures. The characters are the Dirac functionals δx (f ) = f (x).
2. The self-adjoints form a real Banach lattice algebra. They correspond to
the real-valued functions.
3. The unitaries correspond to unit-valued functions.
4. Stone-Weierstraß: Any ∗-subalgebra that separates points is dense in X.
 Joseph Muscat 2015 30

8.3 Finite Dimensional Algebras

Equivalently a regular Banach algebra (i.e., every element is regular ∀a, ∃x, axa =
a).
It can be given the non-degenerate bilinear form hx, yi := tr(x∗ y) where the
elements are considered as matrices.
They are the reflexive C ∗ -algebras. Proof: If X is infinite dimensional then
there an x ∈ X with K := σ(x) ⊇ A countably infinite; so X ⊇ C ∗ (x) ∼ =
C(K) ⊇ C(A) ∼ = c, which is not reflexive.
The ∗-simple finite-dimensional C ∗ -algebras are Mn (C) and Mn (C)2 (with
(x, y)∗ = (y ∗ , x∗ ).) Of these the only commutative ones are n = 1, i.e., C and
C2 .

8.3.1 Frobenius Algebras

are finite-dimensional algebras with a non-degenerate bilinear form such that
hxy, zi = hx, yzi.
Examples: Mn (F) with hx, yi := tr(xy).

8.3.2 Geometric Algebras

A geometric algebra is the algebra generated by a real/complex finite-dimensional
vector space V such that v 2 ∈ R for v ∈ V . Note that q(v) := v 2 is thus a
quadratic form.
Let g := [hai , aj i] = RDR∗ , with D consisting of p 1s, q −1s and r 0s; the
orthogonal columns (in Euclidean sense) of R form an orthogonal basis ei (wrt
the bilinear form); so ej ei = ±ei ej or 0.
The algebra has dimension 2dim V , generated by the orthogonal basis ei · · · ej
(1 6 i < · · · < j 6 n, adding 1 separately).
As tensors, the elements are
graded. The elements of grade r give an nr -dimensional subspace. The highest
grade subspace is one-dimensional, called the pseudo-scalars, generated by ω =
e1 · · · en .
vw + wv
hx, yi := Re(x∗ y) = αβ + + ···
2
Note vw + wv = (v + w)2 − v 2 − w2 ∈ R.

vw = hv, wi + v ∧ w, hα + v, α + vi = α2 + v 2
h1, vi = 0, hv, wi = 0 ⇔ vw = −wv
hx, yzi = hy ∗ x, zi = hxz ∗ , yi
vvr = v · vr + v ∧ vr
vvr −(−1)r vr v vvr +(−1)r vr v
where v·vr := 2 , v∧vr = 2 (by induction); more generally

vr vs = vr · vs + · · · + vr ∧ vs

where vr · vs has grade |r − s|, up by two grades, to the highest grade r + s.

 Joseph Muscat 2015 31

1. X + is a geometric sub-algebra.
1
2. 2 (uvw + wvu) = hv, wiu − hw, uiv + hu, viw
3. u · (v ∧ w) = hu, wiv − hu, viw,
u · (v1 ∧ v2 ∧ v3 ) = hu, v1 iv2 ∧ v3 − hu, v2 iv1 ∧ v3 + hu, v3 iv1 ∧ v2 , etc.
4. Hodge duality: ∗x := −ωx.
∗vr = vn−r = −ωvr = −(−1)r(n−1) vr ω, so there is a correspondence
between r-vectors and (n − r)-vectors.
∗(xy) = ∗(x)y; e.g. vr ×ws := ∗(vr ∧ws ) = ∗vr ·ws , u×(v×w) = −u·(v∧w),
∗(vr · ws ) = ∗(vr ) ∧ ws .
5. For any morphism T , y ∗ T (x) = T ∗ (y) ∗ x. Eigenvectors can be extended
to T vr = λvr .
6. Rotation by θ in e1 , e2 plane: x 7→ rxr∗ , where r = ±ee2 e1 θ/2 (called a
‘rotor’).
Reflection along direction e is v 7→ (eve)∗ = −eve.
Inversion is v 7→ v −1 = v/v 2 .

Exterior algebra: v 2 = 0 for all v ∈ V . For all u, v, hu, vi = 0, so

uv = u ∧ v.
Non-degenerate geometric algebras: v 2 = 0 ⇒ v = 0. Hence the
Clifford algebra is Cℓp,q (R) or Cℓn (C).
There is a conjugation x 7→ axa∗ −1 .

X = Cℓp,q (R)
p p+1 p+2
Y = Cℓq,p (R)
q X +
Xp+1,q ∼
= Y Cℓ2,0 ⊗ Y
q+1 +
Xp,q+1 ∼
= X Cℓ1,1 ⊗ X
q+2 Cℓ0,2 ⊗ Y
(
e′i ⊗ e′′1 ⊗ e′′2 i6p
Proof: Use the maps J : ei 7→ for a basis e′i of
1 ⊗ e′′i−q i>q
(
e′i ⊗ e′′1 ⊗ e′′2 i6p
Cℓp,q (R) and e′′i of Cℓ2,0 (R) = M2 (R); or J : ei →
7 ; or
1 ⊗ e′′i−p i>p

 ′ ′′ ′′
e i ⊗ e 1 e 2 i 6 p or p + 1 < i 6 p + q + 1
J : ei 7→ 1 ⊗ e1 ′′ .
i=p+1


1 ⊗ e′′2 i=p+q+2

It follows that Cℓp+1,q ∼

= Cℓq+1,p , Cℓp,q+4 ∼
= Cℓp+4,q , Cℓp+8,q ∼
= M16 (Cℓp,q ); if
∼
p − q = 1 (mod 4) then Cℓp+i,q = Cℓp,q+i .
Hence the first few geometric algebras over R are (note that Mn (R) ⊗ F ∼ =
Mn (F), C ⊗ H ∼= M2 (C), H ⊗ H ∼ = M4 (R))
 Joseph Muscat 2015 32

p − q − 1 (mod 8) 0 ±1 ±2 ±3 4
Cℓp,q (R) R(m)2 R(m) C(m) H(m) H(m)2
where F(n) := M2n (F).
Similarly, Cℓn (C) ∼
= C(n) or C(n)2 , Cℓn+2 ∼
= M2 (Cℓn ).
Proposition 1

The finite-dimensional real division algebras are R, C, and H.

The only complex finite dimensional division algebra is C.

Proof: Any x ∈ X satisfies a polynomial 0 = (x − α) · · · (x2 − 2βx + γ);
hence x ∈ R or it satisfies x2 − 2βx + γ = 0. For x ∈
/ R, x has only two complex
eigenvalues λ, λ̄, so x2 ∈ R ⇔ λ + λ̄ = 2β = 0 ⇔ tr(x) = 0. Hence X is a
geometric algebra.
For a geometric division algebra, e2 = 0 ⇒ e = 0, e2 = 1 ⇒ (e+1)(e−1) =
0 ⇒ e ∈ R; if e2i = −1, then (1 − e1 e2 e3 )(1 + e1 e2 e3 ) = 0. So the only
possibilities are Cℓ0 = R, Cℓ0,1 = C, Cℓ0,2 = H.


(There is also the octonion algebra O which is weakly associative, x2 y =

x(xy), yx2 = (yx)x).

8.3.3 Finite-dimensional Complex Lie algebras

Example: The skew-adjoint matrices u(n), satisfying A∗ Q = −QA, where
Q(x, y) is linear in y and anti-linear in x.
Solvable Lie algebras are embedded in the upper-triangular matrices b(n).
Semi-simple Lie algebras are products of simple Lie algebras. These are

Simple Lie algebra sl(n) so(2n + 1) so(2n) sp(2n) g2 f4 e6 e7 e8

Corresp. Weyl group An−1 Bn Dn Cn G2 F4 E6 E7 E8

(They are classified because the Weyl group of reflections along the root
vectors form certain Coxeter groups). so(3) ∼
= R3 (with cross-product).

8.3.4 Finite-dimensional Jordan algebras

P
The formally real Jordan algebras (i.e., i x2i = 0 ⇒ xi = 0) are classified -
they are the product of the simple ones, i.e.,
1. “Real”, the self-adjoint operators on RN ;
2. “Complex”, the self-adjoint operators on CN ;
3. “Quaternionic”, the self-adjoint operators on HN ;
 Joseph Muscat 2015 33

4. “Octonion”, the self-adjoint operators on O3 (exceptional case);

5. “Spin factor”, R × RN with (s, x) ∗ (t, y) = (st + x · y, sy + tx).
The first 4 examples all have x ∗ y = (xy + yx)/2. Their projections are RP N −1 ,
CP N −1 , HP N −1 , OP 2 .

9 Examples
Finite Dimensional Spaces
Pn
1. Euclidean space with inner product hx, yi := i=1 āi bi . Euclidean theo-
rems apply.
2. Taxicab metric k(a, b)k := |a| + |b|. Although its topological properties
are the same as the Euclidean case, its metric properties are different.
There are many shortest paths between two points; the angle between
two unit vectors can be taken to be the length of arc on the unit circle;
equilateral triangles need not be equiangular, SAS triangles need not be
congruent; ‘conics’ as d(x, a) = ed(x, b), as sum/difference of distances
from two points being constant, or as distance from line d(x, L) = ed(x, a);
circles may touch at a whole line.
3. Dual numbers: the exterior
 algebra
 on R: a + bǫ with ǫ2 = 0. (a + bǫ)∗ =
a b
a − bǫ. Isomorphic to . It is a local ring. For any differentiable
0 a
function, f (a + bǫ) = f (a) + f ′ (a)bǫ.
   
0 1 0 −i
4. Cℓ3 (R) = M2 (C), can be represented by the Pauli matrices 1 0 , i 0 ,
 
1 0
0 −1
(they generate sl(2)). Contains the quaternions (as σi /i).
   
σ 0 0 −1
5. H = Cℓ0,2 (R), can be represented by i = 0i σ where σi = 1 0
, and
      i
0 −σ 1 0 0 1
j, k = σ 0 where σj = 0 −1 , σk = 1 0 .
Sequence Spaces
P 1 |an |
6. RN with pointwise convergence. Has quasi-norm n 2n 1+|an | . Locally
convex, but not locally bounded.
7. ℓ∞ of bounded sequences with norm supn |an |, and involution (an )∗ :=
(a∗n ), hence a C ∗ -algebra. Its dual is ba, so not reflexive; not separable. It
is injective, i.e., it is complemented in any larger Banach space (via pro-
jection x 7→ (πi x) where πi are extensions of the coordinate projections).
Weak convergence implies pointwise iff weak* convergence.
c is the closed subspace of convergent sequences (not complemented in
ℓ∞ ); isomorphic to c0 , the subspace of sequences that converge to 0, a Ba-
nach algebra; isomorphic to cs, the space of convergent series with norm
 Joseph Muscat 2015 34

P ∼ bv). k(an ) + c0 k = lim sup |an |. Its

k(an )kcs := supn | i>n ai | (cs∗ = n
1
dual is ℓ , so not reflexive; Schauder basis en , so separable. Not weak com-
plete, e.g. (1, . . . , 1, 0, . . .) is weak Cauchy but does not converge weakly.
en ⇀ 0. It is the only separable injective Banach space. The closed unit
ball of c0 is not weak compact and has no extreme points; the closed unit
ball of c has extreme points ±1. The character space consists of δi .
P
8. ℓ1 , the space of absolutely summable series with norm k(an )k := n |an |,
a Banach algebra. Dual space is ℓ∞ , so not reflexive; Schauder basis en ,
so separable. Weak*-convergence iff pointwise convergence and bounded.
Weak convergence of sequences iff norm convergence, implies pointwise
iθ
convergence. The closed unit P∞ball hasn extreme points e en . The charac-
ters are B C , with ψ(an ) = n=0 an z ‘generating function’.
P
ℓ1 (Z) has characters S 1 and ψ(θ) = n∈Z an z n ; σ(an ) = im b(an ); (an )
P
has a ∗-inverse iff n an einθ 6= 0 for all θ. Can be made into a C ∗ -algebra
with (an ) = (ān ) and norm kxk = kLx k, embedded in B(ℓ2 ).
∗

pP
9. ℓp , p > 1, with norm k(an )k := p n |an |p . I : ℓp → ℓq is continuous
for q 6 p; (Pitt) Every operator ℓp → ℓq is compact when q < p; hence
∗
ℓp 6∼= ℓq . Dual space is ℓp where 1p + p1∗ = 1, so reflexive; uniformly
convex; Schauder basis en , so separable. Weak convergence iff pointwise
convergence and bounded. The set { en : n ∈ N } is closed (discrete) but
en ⇀ 0; { en } ∪ { 0 } is weakly compact. n1/p en 6⇀ 0 (since unbounded)
but 0 is a weak limit point of the sequence (∀N, ∃n > N, n1/p en ∈ Vx,ǫ ).
The compact operators form the only closed ideal (p > 1).

P
ℓ2 has inner product h(an ), (bn )i := ān bn .
n
P
10. ℓp , 0 < p < 1, with quasi-norm k(an )k := n |an |p . Locally bounded,
separable (via en ), not locally convex. Dual space is isometric to ℓ∞ via
1
usual x 7→ x∗ . The set n1−p en is totally bounded but its convex hull is
PN 1
unbounded (e.g. n=1 n1−p en /N ).
11. James’ space: subspace of c0 with norm

sup k(an2 − an1 , · · · , ank − ank−1 , ank+1 , 0, . . .)kℓ2 ,

(ni )∈O

where O is any odd sequence of (increasing) integers. Complete, separable

with en as a conditional Schauder basis. Not reflexive even though X ∼ =
X ∗∗ .
12. ba, the space of finitely additive signed measures on N, with norm kµk :=
supE⊆N µ(E) − inf E⊆N µ(E). Not separable. Although the unit ball is
weak*-compact it is not sequentially compact, e.g. e∗n acting on ℓ∞ has no
weak*-convergent subsequence.
 Joseph Muscat 2015 35

Contains the closed subspace P bv, of sequences of bounded variation with

norm k(an )kbv := |a1 | + n |an+1 − an |; isomorphic to ℓ1 via (an ) 7→
(a1 , . . . , an+1 − an , . . .). en 6⇀ 0.
Function Spaces
R1
13. L1 [0, 1], space of functions with norm kf k1 := 0 |f |. Dual space is
L∞ [0, 1], so not reflexive; separable by polynomials. Weakly sequentially
complete: every weakly Cauchy sequence converges weakly. The closed
unit ball has no extreme points.
R 2π
L1 (S 1 ) has character space Z, ψn (an ) = 0 einθ f (θ) dθ; the Gelfand map
are the Fourier coefficients.
R
L1 (R) has character space R, ψξ (f ) = eixξ f (x) dx; the Gelfand map is
the Fourier transform.
R∞
L1 (R+ ) has character space R+ ×iR, ψz (f ) = 0 e−zx f (x) dx; the Gelfand
map is the Laplace transform.
qR
1 ∗
14. Lp [0, 1], 1 < p, with norm kf kp := p 0 |f |p . Dual space is Lp where
1 1
p + p∗ = 1, so reflexive; uniformly convex since

∗ ∗
2(kf kp + kgkp )p−1 6 kf + gkp + kf − gkp 6 2(kf kp + kgkp ), (p 6 2)

(reversed inequalities for p > 2); separable. I : Lp [0, 1] → Lq [0, 1] is con-

tinuous for q 6 p, with meagre image (unit ball has no interior in Lq ).
The closed unit ball has its boundary as extreme points.
R1
L2 [0, 1] has inner product hf, gi := 0 f¯g; isomorphic to ℓ2 . The Hilbert-
Schmidt operators are the integral operators with kernel in L2 [0, 1]2 .
15. Lp [0, 1], 0 < p < 1. Locally bounded, but there are no non-trivial open
convex subsets; hence trivial dual space (no morphisms into a locally con-
vex space); the only weakly closed subspaces are 0 and X. No Schauder
basis.
16. L∞ [0, 1], space of bounded (ae) functions with norm kf k∞ := supx a.e. |f (x)|.
Isomorphic to ℓ∞ ; not separable. The closed unit ball has extreme points
|f | = 1 a.e..

17. L0 [0, 1], the space of measurable functions with fn → 0 when ∀ǫ >
0, µ{ x : |fn (x)| > ǫ } → 0 as n → ∞.
18. C(Ω), the space of continuous functions with complete quasi-norm: if (fn )
is Cauchy, then (fn ) is Cauchy in each C(Ki ), so fn → f in Ki ; take f as
P |fn −f |i
patch of all these f ’s; then |fn − f | = i 21i 1+|f n −f |i
1
< m , i.e., fn → f
in C(Ω).
 Joseph Muscat 2015 36

C(K) is separable iff K is metrizable (similarly C0 (X)). Dual space con-

sists of regular Borel measures of bounded variation (not separable: un-
countable δt ). Weak-convergence iff pointwise and bounded. The closed
unit ball has extreme points δx , x ∈ K.
C[0, 1] with involution f ∗ (t) = f (t), a C ∗ -algebra; has character space
[0, 1], δt ; its Gelfand map is the identity, σ(f ) = im f . The closed ideals
correspond to closed subsets of [0, 1] as IA = { f : f A = 0 }. σ(f ) = im(f ).
C(RN ). Locally convex but not locally bounded; not separable (contains
ℓ∞ ). The closed unit ball has extreme points ±1 (or |f | = 1 if over C).
Matrix Algebras
19. B(ℓ2 ), not separable (contains ℓ∞ ).

20. B(c
P 0 ). Each eigenvalue belongs to a closed disk about Tii of radius
j6=i |Tji |.