OPERATOR THEORY - PART 1/3

T.K.SUBRAHMONIAN MOOTHATHU

Contents

1. What is more general than an eigenvalue? 2

2. Banach algebras and Complex Analysis tools 3
3. Common properties of the spectrum 7
4. Holomorphic functional calculus 10
5. Commutative case: maximal ideals and multiplicative functionals 14
6. Spectrum relative to a closed subalgebra 18

Main pre-requisites: Functional Analysis, Complex Analysis, Measure Theory, Topology.

Two major topics of our study are: (i) spectrum, and (ii) functional calculus. Let us give a quick
introduction. If T : Ck → Ck is a linear map, then the set of eigenvalues of T is a finite subset
of C, and we may understand a lot about T by studying this finite set and related issues. When
we go to infinite dimension, to a bounded linear operator T of a Banach space, we can associate
a compact subset of C that plays the role of the set of eigenvalues. This compact subset is called
P
the spectrum of T . If T is an operator and p(z) = ni=0 ai z i is a complex polynomial, then we may
P
define the operator p[T ] := ni=0 ai T i . Functional calculus is about defining the operator f [T ] for
more general complex-valued functions f and studying the many useful consequences.

To simplify our exposition, the following will be assumed at most of the places, even though this
will exclude some interesting operators from our study:

(i) The underlying space X is a Banach space (sometimes, a Hilbert space).

(ii) The underlying field of this Banach space is C.
(iii) The linear operator T is from X to X (rather than from X to another space Y ).
(iv) T is defined on the whole of X, and T is bounded (equivalently, continuous).

Also, it is convenient to use the language of operator algebras while developing the theory, but
we will not be interested in operator algebras per se.
1
2 T.K.SUBRAHMONIAN MOOTHATHU

1. What is more general than an eigenvalue?

Definition: If X is a (complex) Banach space, let Γ(X) = {T : X → X : T is linear and bounded}.

Here, we use the notation Γ to remind us of T . Note that Γ(X) is a Banach space with respect to
the operator norm kT k := sup{kT xk : x ∈ X, kxk ≤ 1}). Once in a while, we may have to consider
Γ(X, Y ) := {T : X → Y : T is linear and bounded} also, where X, Y are Banach spaces.

Question: What should we study about T ∈ Γ(X)?

In Linear Algebra, we study about eigenvalues. For T ∈ Γ(Ck ), note that a complex number α
is an eigenvalue ⇔ ker(αI − T ) 6= {0} ⇔ αI − T is not invertible in Γ(Ck ). When dim(X) = ∞,
for T ∈ Γ(X), injectivity of αI − T does not imply the invertibility of αI − T in Γ(X).

Definition: Let X be a (complex) Banach space. The spectrum σ(T ) of T ∈ Γ(X) is defined as
σ(T ) = {α ∈ C : αI − T is not invertible in Γ(X)}. Any α ∈ σ(T ) is called a spectral value of T .
The point spectrum σp (T ) of T is defined as σp (T ) = {α ∈ C : α is an eigenvalue of T } = {α ∈ C :
ker(αI − T ) 6= {0}}.

Remark: Any T ∈ Γ(Ck ) has at least one eigenvalue since the characteristic polynomial has at
least one complex root. But the concept of a characteristic polynomial is not available in infinite
dimension. In fact, when dim(X) = ∞, a bounded linear operator may not have any eigenvalue at
all; for example consider the right shift T : l2 → l2 given as (x0 , x1 , x2 , . . .) 7→ (0, x0 , x1 , x2 , . . .).

Remark: Clearly σp (T ) ⊂ σ(T ) always by definition, and equality holds when dim(X) < ∞. Strict
inclusion is possible when dim(X) = ∞; see the example below.

[101] [Diagonal operator] Let X be a Hilbert space with orthonormal basis {en : n ∈ N}, let (βn )
P P∞
be a bounded sequence in C, and let T ∈ Γ(X) be defined as ∞ n=1 αn en 7→ n=1 βn αn en . Then,
(i) σp (T ) = {βn : n ∈ N}.
(ii) σ(T ) = {βn : n ∈ N}.

Proof. (ii) Observe that T is invertible in Γ(X) ⇔ βn 6= 0 for every n ∈ N and (1/βn ) is bounded ⇔
0∈
/ {βn : n ∈ N}. For α ∈ C, αI − T is also a diagonal operator with sequence (α − βn ). So αI − T
is not invertible ⇔ 0 ∈
/ {α − βn : n ∈ N} ⇔ α ∈
/ {βn : n ∈ N}. Hence σ(T ) = {βn : n ∈ N}. ¤

[102] [Corollary] Let X be an infinite dimensional separable Hilbert space and let K ⊂ C be a
nonempty compact set. Then there is T ∈ Γ(X) such that σ(T ) = K.

Proof. Let {βn : n ∈ N} be a countable dense subset of K. Then (βn ) is bounded, and we may
consider the diagonal operator corresponding to the sequence (βn ). ¤

Remark: Later we will see that σ(T ) is always nonempty and compact for any T ∈ Γ(X).
 OPERATOR THEORY - PART 1/3 3

Remark: Let X, Y be Banach spaces and let T1 ∈ Γ(X), T2 ∈ Γ(Y ) be conjugate in the sense that
there is an isomorphism S ∈ Γ(X, Y ) with T2 = ST1 S −1 . Then αIY − T2 = αSS −1 − ST1 S −1 =
S(αIX − T1 )S −1 , and it follows that σ(T1 ) = σ(T2 ) and σp (T1 ) = σp (T2 ). Using [102] (or directly
from [101]), we deduce that there are uncountably many pairwise non-conjugate diagonal operators.

[103] [Multiplication operator] Let (Y, µ) be a σ-finite measure space and fix g ∈ L∞
C (Y, µ). Let
L2 = L2C (Y, µ), and define T ∈ Γ(L2 ) as T (f ) = gf , i.e., T (f )(y) = g(y)f (y). Then,
(i) σp (T ) = {α ∈ C : µ[g −1 (α)] > 0}.
(ii) σ(T ) = {α ∈ C : µ[g −1 (B(α, ²))] > 0 for every ² > 0} (called the essential range of g).

Proof. (i) Let A = g −1 (α) ⊂ Y . If µ[A] = 0 and T f = αf , then f (y) = 0 for every y ∈ Y \ A
and hence f = 0 in L2 . So α ∈
/ σp (T ). If µ[A] > 0, choose B ⊂ A with 0 < µ[B] < ∞ by the
σ-finiteness of the measure. Then χB ∈ L2 \ {0} and T χB = αχB so that α ∈ σp (T ).
(ii) Observe that T is invertible in Γ(L2 ) with inverse f 7→ f /g ⇔ 1/g ∈ L∞
C (Y, µ) ⇔ there is ² > 0
such that µ[g −1 (B(0, ²))] = 0. For α ∈ C, αI − T is multiplication by α − g. Hence αI − T is
invertible in Γ(L2 ) ⇔ there is ² > 0 such that µ[g −1 (B(α, ²))] = 0. ¤

Remark: If Y = N and µ is the counting measure, then L2C (Y, µ) = lC

2 (N), and the multiplication

operator reduces to a diagonal operator. Moreover, any infinite dimensional separable Hilbert space
2 (N). So [101] is just a special case of [103].
is isomorphic to lC

2. Banach algebras and Complex Analysis tools

For any T ∈ Γ(X) (where X is a Banach space), the spectrum σ(T ) satisfies some common
properties. To derive these properties, it is convenient to adopt an abstract approach using the
basic theory of Banach algebras (defined below). We use the notation Γ for a Banach algebra and
s, t for elements of Γ to remind us of Γ(X) and S, T ∈ Γ(X).

Definition: (i) An algebra is a pair (Γ, ·), where Γ is a complex vector space, and ‘·’ is an associative,
bilinear, binary operation (called multiplication or product) on the space Γ. Here, ‘associative’
means (t1 · t2 ) · t3 = t1 · (t2 · t3 ) for every ti ∈ Γ, and ‘bilinear’ means linear in each variable
separately. Often we will just write st for s · t when s, t ∈ Γ.
(ii) An algebra Γ is called a normed algebra if (Γ, k · k) is a normed space and the norm k · k is
submultiplicative in the sense that kstk ≤ kskktk for every s, t ∈ Γ. If in addition, the norm is
complete, then Γ is called a Banach algebra.
(iii) A Banach algebra Γ is said to be unital if it has a multiplicative identity, say e. To avoid some
pathologies, we assume e 6= 0, and in particular this implies Γ 6= {0}. We will always suppose that
kek = 1 (after normalizing the norm).
4 T.K.SUBRAHMONIAN MOOTHATHU

Examples: (i) C with the usual operations and norm is a unital Banach algebra. (ii) If X is a
complex Banach space, then Γ(X) is a unital Banach algebra, where the multiplication is compo-
sition, and I is the multiplicative identity. Since kST k ≤ kSkkT k for T, S ∈ Γ(X), the operator
norm on Γ(X) is indeed submultiplicative. (iii) If K 6= ∅ is a compact Hausdorff space, then
C(K, C) = {f : K → C : f is continuous} is a unital Banach algebra with respect to the pointwise
operations and the sup-norm. The multiplicative identity is the constant function 1.

Remark: There are many other interesting examples of algebras and Banach algebras (for instance,
the disc algebra), but we ignore them for the moment.

Definition: Let Γ be a unital Banach algebra. We say t ∈ Γ is invertible in Γ (or simply invertible
when Γ is understood) if there is s ∈ Γ such that ts = st = e. If t is invertible, the inverse is unique
and is denoted as t−1 . We define Inv(Γ) := {t ∈ Γ : t is invertible in Γ}.

[104] [Open group of invertible elements] Let Γ be a unital Banach algebra. Then,
(i) (Inv(Γ), ·) is a group (generally non-abelian).
(ii) If t ∈ Γ and ktk < 1, then e − t ∈ Inv(Γ) and (e − t)−1 = e + t + t2 + t3 + · · · .
(iii) If t ∈ Inv(Γ), s ∈ Γ and kt − sk < 1/kt−1 k, then s ∈ Inv(Γ). Therefore, Inv(Γ) is open in Γ.
(iv) The map t 7→ t−1 from Inv(Γ) to itself is a homeomorphism.
(v) If t1 , . . . , tn ∈ Γ are commuting elements, then t1 · · · tn ∈ Inv(Γ) iff tj ∈ Inv(Γ) for every j.

Proof. (i) is clear.

P
(ii) By submultiplicity, ktn k ≤ ktkn . So ∞ n
n=0 t is absolutely convergent by hypothesis. Hence
P∞ n P∞ n P∞ n
n=0 t ∈ Γ since Γ is Banach. Now it is easy to verify that (e−t)( n=0 t ) = ( n=0 t )(e−t) = e.

(iii) s = t − (t − s) = [e − (t − s)t−1 ]t. Also, k(t − s)t−1 k ≤ kt − skkt−1 k < 1 so that (ii) applies.

(iv) The map is clearly bijective. By symmetry, it now suffices to prove continuity of the map.
Fix t ∈ Inv(Γ) and let s ∈ B(t, 1/2kt−1 k). Then s ∈ Inv(Γ) by part (iii). Moreover, ks−1 k ≤
ks−1 − t−1 k + kt−1 k ≤ ks−1 (t − s)t−1 k + kt−1 k ≤ 21 ks−1 k + kt−1 k implying ks−1 k ≤ 2kt−1 k. Now
ks−1 − t−1 k = ks−1 (t − s)t−1 | ≤ ks−1 kkt − skkt−1 k ≤ 2kt−1 k2 kt − sk, giving continuity at t.

(v) The implication ⇐ is clear. For ⇒, by induction it suffices to consider the case n = 2. So
let s, t ∈ Γ be commuting elements with st ∈ Inv(Γ), and let u = (st)−1 = (ts)−1 . Note that
tu = etu = (uts)tu = ut(stu) = ute = ut. Therefore, from stu = e = uts = tus, we see s−1 = tu.
Similarly, su = us is the inverse of t. ¤

Convention: Let Γ be a unital Banach algebra for the rest of this section. We denote the Banach
space dual of Γ by Γ∗ .

[105] For s, t ∈ Γ, we have s = t iff φ(s) = φ(t) for every φ ∈ Γ∗ (i.e., Γ∗ separates points of Γ).
 OPERATOR THEORY - PART 1/3 5

Proof. This is a consequence of Hahn-Banach theorem. ¤

For Γ-valued maps defined on open subsets of C, there is a parallel theory of Complex Analysis.
We sketch below what we need from this parallel theory.

Definition: Let U ⊂ C be open (and nonempty).

f (z) − f (w)
(i) We say f : U → Γ is holomorphic on U if limz→w exists in Γ for each w ∈ U . Here
z−w
note that (z − w)−1 is a complex scalar and f (z) − f (w) is an element of Γ.
(ii) We say f : U → Γ is weakly holomorphic if the function φ ◦ f : U → C is holomorphic in the
usual sense for every φ ∈ Γ∗ .

[106] Let U ⊂ C be open. Then, f : U → Γ is holomorphic iff f is weakly holomorphic.

Proof. Suppose f is holomorphic, let φ ∈ Γ∗ and w ∈ U . Since φ is linear and continuous,

µ ¶
φ(f (z)) − φ(f (w)) f (z) − f (w)
lim = φ lim = φ(f 0 (w)).
z→w z−w z→w z−w

This shows φ◦f is holomorphic on U . For a proof of the converse part, see p.79 of Rudin, Functional
Analysis. We omit the proof since we will not need this converse. ¤

[107] [Liouville’s theorem] If f : C → Γ is holomorphic and bounded, then f is constant.

Proof. Suppose f is not a constant, and let z, w ∈ C be with f (z) 6= f (w). Since Γ∗ separates points
of Γ, there is φ ∈ Γ∗ such that φ(f (z)) 6= φ(f (w)) and thus φ ◦ f : C → C is not a constant. But
φ ◦ f is holomorphic by [106]. Also φ ◦ f is bounded since f is bounded and kφ(f (z))k ≤ kφkkf (z)k.
This contradicts the classical Liouville’s theorem in Complex Analysis. So f must be constant. ¤

Next we clarify how to integrate certain Γ-valued maps. There are different concepts of integra-
tion for Γ-valued maps. For our purpose, a simple analogue of Riemann integration will suffice.

Definition: Let f : [a, b] → Γ be a map, and consider a partition P = {a = a0 ≤ a1 ≤ · · · ≤ an = b}

P
of [a, b]. Let R(f, P ) = nk=1 (ak − ak−1 )f (ak ) ∈ Γ be the Riemann sum of f with respect to P .
Also let kP k = max{|ak − ak−1 | : 1 ≤ k ≤ n}. We say f is integrable if there is t ∈ Γ such
that limn→∞ R(f, Pn ) = t for any sequence (Pn ) of partitions of [a, b] satisfying kPn k → 0. If this
Rb
happens we write a f (x)dx = t.

Exercise-1: Let f : [a, b] → Γ be continuous and ² > 0. Let λ > 0 be such that |f (x) − f (y)| < ²
whenever |x − y| < λ. If P1 , P2 are partitions of [a, b] such that kPi k < λ for i = 1, 2, then
kR(f, P1 ) − R(f, P2 )k ≤ 2(b − a)². [Hint: Let P3 be a common refinement of P1 and P2 . It suffices
to show kR(f, Pi ) − R(f, P3 )k ≤ (b − a)² for i = 1, 2. Let P1 = {a = a0 ≤ a1 ≤ · · · ≤ an = b} and fix
k ∈ {1, . . . , n}. The restriction of P3 to [ak−1 , ak ] has the form {ak−1 = b0 ≤ b1 ≤ · · · ≤ bm = ak }
6 T.K.SUBRAHMONIAN MOOTHATHU
Pm
for some m depending on k. Then k(ak − ak−1 )f (ak ) − j=0 (bj − bj−1 )f (bj )k ≤ (ak − ak−1 )².
Summing over k, we get kR(f, P1 ) − R(f, P3 )k ≤ (b − a)².]
Rb Rb
Exercise-2: Let f : [a, b] → Γ be continuous. Then, (i) f is integrable. (ii) φ( a f ) = a φ ◦ f for
any bounded linear φ : Γ → Γ1 , where Γ1 is any unital Banach algebra (in particular, Γ1 can be C
Rb Rb Rb Rb
or Γ). (iii) t( a f ) = a (tf ) and ( a f )t = a (f t) for every t ∈ Γ.

[Hint: (i) Let (Pn ) be a sequence of partitions of [a, b] such that kPn k → 0. Then (R(f, Pn ))
Rb
is Cauchy in Γ, and its limit in Γ is independent of (Pn ), all by Exercise-1. (ii) φ( a f ) =
Rb
φ(limn→∞ R(f, Pn )) = limn→∞ φ(R(f, Pn )) = limn→∞ R(φ ◦ f, Pn ) = a φ ◦ f . (iii) s 7→ ts and
s 7→ st are bounded linear maps from Γ to itself so that (ii) applies.]

Definition: Let U ⊂ C be open, f : U → Γ be holomorphic and let δ be a C 1 path in U , i.e.,

δ : [a, b] → U is a continuously differentiable map. Then note that δ 0 (f ◦ δ) : [a, b] → Γ is
continuous, and hence integrable by Exercise-2. Imitating what we do in Complex Analysis, we
R R R Rb
define the path integral δ f as δ f = δ f (z)dz := a δ 0 (x)f (δ(x))dx ∈ Γ (here note that δ 0 (x) ∈ C
R
and f (δ(x)) ∈ Γ). By additivity, we may also define δ f for piecewise C 1 paths δ in U .

Definition: Let U ⊂ C be open and f : U → Γ be holomorphic. In applications we will have to

integrate f over a collection of finitely many piecewise C 1 paths δ1 , . . . , δn in U . Symbolically we
R P R
write ∆ := δ1 + · · · + δn for convenience, and ∆ f := nj=1 δj f . If δ1 , . . . , δn are piecewise C 1
closed paths in U and if ∆ = δ1 + · · · + δn , then we say ∆ is a closed contour in U .

Remark: Note that the image [∆] of a closed contour ∆ in U need not be connected. For example,
if δj parametrizes {z ∈ C : |z| = j} for j = 1, . . . , n (n ≥ 2) and ∆ = δ1 + · · · + δn , then as per our
definition ∆ is a closed contour in C, and the image of ∆ is a union of n disjoint circles.

[108] Let U ⊂ C be open, f : U → Γ be holomorphic, and let ∆ be a closed contour in U . Then,

R R
(i) φ( ∆ f ) = ∆ φ ◦ f for every φ ∈ Γ∗ .
R R R R
(ii) t( ∆ f ) = ∆ (tf ) and ( ∆ f )t = ∆ (f t) for every t ∈ Γ.

Proof. Use Exercise-2 and the definition of path integral. ¤

Exercise-3: Let ∅ =
6 K ⊂ U ⊂ C, where K is compact and U is open. Then there is a closed
contour ∆ in U \ K such that the winding number n(∆, w) = 1 for every w ∈ K and n(∆, w) = 0
1 R 1
for every w ∈ C \ U , where n(∆, w) = ∆ dz. [Hint: Consider a grid of squares (of equal
2πi z−w
size) covering U . Since K is compact, there is r > 0 such that B(w, r) ⊂ U for every w ∈ K. So we
may assume any square in the grid intersecting K is contained in U . Consider all squares from the
grid intersecting K, parametrize their edges in the anticlockwise direction, and cancel out opposite
edges. What remains will be a collection of finitely many piecewise C 1 closed paths δ1 , . . . , δn in
U \ K. Take ∆ = δ1 + · · · + δn . For the details, see p.269 of Rudin, Real and Complex Analysis.]
 OPERATOR THEORY - PART 1/3 7

[109] [Cauchy’s integral theorem and formula] Let U ⊂ C be open and nonempty, f : U → Γ be
holomorphic, and let ∆ be a closed contour in U such that n(∆, w) = 0 for every w ∈ C \ U . Then,
R
(i) ∆ f = 0 ∈ Γ.
1 R f (z)
(ii) n(∆, w)f (w) = dz for every w ∈ U \ [∆].
2πi ∆ z − w

Proof. Let φ ∈ Γ∗ be arbitrary. Use the linearity of φ, the interchangeability property from [108],
R
and the classical Cauchy’sµtheorem and formula
¶ (applied to φ ◦ f ) to conclude that φ( ∆ f ) = φ(0)
1 R f (z)
and φ (n(∆, w)f (w)) = φ dz . Now note that Γ∗ separates points of Γ. ¤
2πi ∆ z − w

3. Common properties of the spectrum

Recall that if X is a Banach space, the Γ(X) is a unital Banach algebra. For T ∈ Γ(X), the
spectrum σ(T ) satisfies some common properties. We derive them below using the language of
Banach algebras.

Definition: Let Γ be a unital Banach algebra. For t ∈ Γ, the spectrum σ(t) of t is σ(t) = {α ∈
C : αe − t ∈
/ Inv(Γ)}. Its complement ρ(t) := C \ σ(t) is the resolvent set of t. The function
R : ρ(t) → Γ given by R(α) = (αe − t)−1 is the resolvent function of t. If it is necessary to specify
t, we will write Rt (α) for R(α).

Exercise-4: Let Γ be a unital Banach algebra and t ∈ Γ. Then (i) ρ(t) is open in C, (ii) |α| > ktk
implies α ∈ ρ(t) for every α ∈ C. [Hint: Use [104]. Note that B(αe − t, 1/kR(α)k) ⊂ Inv(Γ) for
every α ∈ ρ(t), and eα − t = α−1 (e − t/α) ∈ Inv(Γ) when |α| > ktk.]

[110] [Properties of the resolvent function] Let Γ be a unital Banach algebra and t ∈ Γ. Consider
the resolvent function R : ρ(t) → Γ of t given by R(α) = (αe − t)−1 . Then,
(i) [Resolvent equation] R(α) − R(β) = −(α − β)R(α)R(β) for every α, β ∈ ρ(t).
(ii) R is holomorphic.
(iii) dist(α, σ(t)) ≥ 1/kR(α)k for every α ∈ ρ(t).
(iv) kR(α)k → 0 as |α| → ∞.

Proof. (i) R(α) = R(α)(βe − t)R(β) = R(α)[αe − t + (β − α)e]R(β) = [e + (β − α)R(α)]R(β) =

R(β) + (β − α)R(α)R(β).
R(α) − R(β)
(ii) By [104](iv), R is continuous. Now, limα→β = −R(β)2 by the resolvent equation.
α−β
(iii) By [104](iii), we have B(αe − t, 1/kR(α)k) ⊂ Inv(Γ) for every α ∈ ρ(t). So for any β ∈ σ(t),
we get |β − α| = k(βe − t) − (αe − t)k ≥ 1/kR(α)k.

(iv) Note that kR(α)k = |α−1 |k(e − t/α)−1 k. As |α| → ∞, we have t/α → 0 and therefore
k(e − t/α)−1 k → kek = 1 since taking the inverse is a continuous map by [104](iv). ¤
8 T.K.SUBRAHMONIAN MOOTHATHU

Remark: From [110](iii) we get kR(α)k → ∞ as dist(α, σ(t)) → 0. Hence ρ(t) is the maximal
region in which the resolvent R of t is holomorphic.

Remark: Recall that to show T ∈ Γ(Ck ) has at least one eigenvalue, we use the fundamental
theorem of algebra, which in turn is usually proved using Liouville’s theorem. Below we use the
Banach algebra version of Liouville’s theorem to show every t ∈ Γ has at least one spectral value.

[111] [Fundamental theorem of spectrum] Let Γ (6= {0}) be a unital Banach algebra and t ∈ Γ.
Then, σ(t) is a nonempty compact subset of C contained in B(0, ktk).

Proof. Exercise-4 tells us that σ(t) is a closed subset of C contained in B(0, ktk), and hence σ(t) is
compact also. It remains to show σ(t) 6= ∅.
Let if possible σ(t) = ∅. Then ρ(t) = C and the resolvent R : C → Γ, R(α) = (αe − t)−1 , is
holomorphic. By [110](iv), there is M > 0 such that kR(α)k ≤ 1 for every α ∈ C with |α| > M .
Moreover, R is bounded on the compact set B(0, M ). Thus R is bounded on C and hence is
constant by Liouville’s theorem, [107]. On the other hand, if α 6= β, then αe − t 6= βe − t, and
therefore R(α) 6= R(β) so that R cannot be a constant function. ¤

Exercise-5: [Gelfand-Mazur theorem] Let Γ be a unital Banach algebra. If Inv(Γ) = Γ \ {0}, then
Γ is isometrically isomorphic to C as Banach algebras. [Hint: Define h : C → Γ as h(α) = αe. If
t ∈ Γ \ h(C), then αe − t ∈ Γ \ {0} = Inv(Γ) for every α ∈ C, implying σ(t) = ∅, a contradiction.
Hence h(C) = Γ. Next, verify that h is a linear isometry, and h(αβ) = h(α)h(β).]

[112] [Small perturbations do not enlarge the spectrum much] Let Γ be a unital Banach algebra,
t ∈ Γ and let U ⊂ C be an open neighborhood of σ(t). Then ∃ ² > 0 such that σ(s) ⊂ U ∀ s ∈ B(t, ²).

Proof. Otherwise there is a sequence (tn ) in Γ converging to t such that there is αn ∈ σ(tn ) \ U for
every n ∈ N. Since |αn | ≤ ktn k and (tn ) is convergent, we get that (αn ) is a bounded sequence.
By passing onto a subsequence we may assume (αn ) converges to some α ∈ C, and necessarily
α ∈ C \ U ⊂ ρ(t) since U is open. Thus αe − t ∈ Inv(Γ). Now (αn e − tn ) → αe − t and Inv(Γ) is
open, and so we have αn e − tn ∈ Inv(Γ) for all large n ∈ N , a contradiction since αn ∈ σ(tn ). ¤

Remark: However, the spectrum can shrink considerably under small perturbations. The following
is possible: there is t ∈ Γ with σ(t) = B(0, 1) such that for every ² > 0, there is s ∈ B(t, ²) with
σ(s) = S1 (see Problem 102 in Halmos, A Hilbert Space Problem Book).
Pn j
Exercise-6: Let Γ be a unital Banach algebra and t ∈ Γ. If p(z) = j=0 aj z is a complex
P
polynomial, define p[t] = nj=0 aj tj ∈ Γ. Then,
(i) (αp + βq)[t] = αp[t] + βq[t], and (pq)[t] = p[t]q[t] = q[t]p[t].
(ii) σ(p[t]) = p(σ(t)).
(iii) p[t] ∈ Inv(Γ) ⇔ p has no zeroes in σ(t).
 OPERATOR THEORY - PART 1/3 9

(iv) If t is an idempotent (i.e., t2 = t), then σ(t) ⊂ {0, 1}.

(v) If t is nilpotent (i.e., tk = 0 for some k ∈ N), then σ(t) = {0}.
[Hint: (ii) Easy for constant polynomials. So assume n = deg(p) ≥ 1. Let α ∈ C, and q(z) =
α − p(z) = c(α1 − z) · · · (αn − z), where c ∈ C \ {0} and αj ∈ C are the solutions to p(z) = α.
Correspondingly, q[t] = αe − p[t] = c(α1 e − t) · · · (αn e − t). By [104](v), we have αe − p[t] ∈ Inv(Γ)
⇔ αj e − t ∈ Inv(Γ) for every j ⇔ αj ∈
/ σ(t) for every j ⇔ α ∈
/ p(σ(t)).
(iii) If p(z) = b(β1 − z) · · · (βn − z), then p[t] ∈ Inv(Γ) iff βj e − t ∈ Inv(Γ) for every j.
(iv) and (v): Let p(z) = z 2 − z, q(z) = z k . Then p(σ(t)) = q(σ(t)) = σ(0) = {0} by (ii).]

Remark: We know σ(t) ⊂ B(0, ktk). Is ktk the smallest radius with this property? Definitely
not; for instance, if t 6= 0 is nilpotent, then ktk > 0, but σ(t) = {0} by Exercise-6 (an example
is T ∈ Γ(Ck ) given by T (ej ) = ej+1 for 1 ≤ j < k and T (ek ) = 0). It is natural to look for the
smallest radius of a closed ball centered at 0 enclosing the spectrum. This number gives a crude
measure of how big the spectrum is, and this number will be related to the sequence (ktn k1/n ).

Definition: Let Γ be a unital Banach algebra. The spectral radius of an element t ∈ Γ is defined as
r(t) := inf{r > 0 : σ(t) ⊂ B(0, r)} = sup{|α| : α ∈ σ(t)} = max{|α| : α ∈ σ(t)}.

Remark: (i) r(t) ≤ ktk by [111], and r(t) < ktk is possible as noted above. (ii) r : Γ → R may not
be continuous. It is possible to have t ∈ Γ with r(t) > 0 such that for every ² > 0 there is s ∈ B(t, ²)
with r(s) = 0 (see Problem 104 in Halmos, A Hilbert Space Problem Book). On the otherhand, by
[112] we have semi-continuity in the sense that if (tn ) → t, then r(t) ≥ lim supn→∞ r(tn ).

[113] [Spectral radius formula] Let Γ be a unital Banach algebra and t ∈ Γ. Then we have
r(t) = limn→∞ ktn k1/n = inf n∈N ktn k1/n .

Proof. If α ∈ σ(t), then αn ∈ σ(tn ) by Exercise-6(ii). Hence |αn | ≤ ktn k, or |α| ≤ ktn k1/n . This
shows r(t) ≤ inf n∈N ktn k1/n ≤ lim inf n→∞ ktn k1/n . Now consider a real number a > r(t), and we
will show lim supn→∞ ktn k1/n ≤ a to complete the proof. Note that it suffices to show (tn /an ) is
bounded, for if ktn k ≤ M an with M > 0, then lim supn→∞ ktn k1/n ≤ lim supn→∞ M 1/n a = a.
Let b > max{a, ktk}, A = {z ∈ C : r(t) < |z| < b}, and A1 = {z ∈ C : ktk < |z| < b}. Since
A ⊂ ρ(t), the resolvent function R(z) = (ze − t)−1 is holomorphic in A by [110]. In the smaller
P
annulus A1 , we have R(z) = (1/z)(e − t/z)−1 = (1/z) ∞ n
n=0 (t/z) as |z| > ktk. Hence, for φ ∈ Γ
∗
P∞
and z ∈ A1 , (φ ◦ R)(z) = (1/z) n=0 φ(tn )z −n . This Laurent series expansion should hold in the
bigger annulus A also, since φ ◦ R : A → C is holomorphic. As a ∈ A, in particular (φ ◦ R)(a) =
P P∞
(1/a) ∞ n −n = (1/a)
n=0 φ(t )a
n n n n
n=0 φ(t /a ). The convergence of the series implies (φ(t /a )) is
bounded. As φ ∈ Γ∗ is arbitrary, we deduce that (tn /an ) is bounded in Γ by the Principle of
uniform boundedness (applied to {Sn : n ∈ N}, where Sn : Γ∗ → C is Sn φ = φ(tn /an )). ¤
10 T.K.SUBRAHMONIAN MOOTHATHU
P∞
Remark: For t ∈ Γ, we know R(α) = (αe − t)−1 = (1/α) n=0 (t/α)
n for every α ∈ C with
|α| > ktk. Since Γ∗
separates points of Γ, the above proof gives the following improvement:
P∞
R(α) = (1/α) n=0 (t/α)n for every α ∈ C with |α| > r(t).

Exercise-7: Let (wn ) be a bounded sequence in C and let T : l2 → l2 be the weighted shift given by
(x0 , x1 , x2 , . . .) 7→ (w1 x1 , w2 x2 , . . .). Then,
(i) kT k = sup{|wn | : n ∈ N} and kT n k = sup{|wk wk+1 · · · wn+k−1 | : k ∈ N}.
(ii) T is a compact operator ⇔ (wn ) → 0.
(iii) If (wn ) → 0, then r(T ) = 0 so that σ(T ) = {0}.
(iv) If wn 6= 0 ∀ n ∈ N, then T is not nilpotent and σp (T ) = ∅.
[Hint: (i) Check by induction T n x = (w1 · · · wn xn , w2 · · · wn+1 xn+1 , w3 · · · wn+2 xn+2 , · · · ). (ii) If
|wnk | ≥ λ > 0 for every k ∈ N, then (T enk ) has no convergent subsequence. If (wn ) → 0, then there
are finite rank operators Tn with (Tn ) → T . (iii) Show limn→∞ kT n k1/n = 0 by the help of (i).]

4. Holomorphic functional calculus

Let Γ be a unital Banach algebra and t ∈ Γ. In Exercise-6, we defined p[t] ∈ Γ for a complex
polynomial p and showed that σ(p[t]) = p(σ(t)). Can this result be extended to more functions?
There are two immediate generalizations to polynomials: (i) power series, (ii) rational functions.
For f in any of these classes, we may try to imitate what we did for polynomials to define f [t].
But there is another approach via the Cauchy integral that will define f [t] for a bigger class of
holomorphic functions. First let us make note of a necessary condition.

Remark: If f (z) = 1/z, then f is holomorphic on C\{0}. The natural definition for f [t] is f [t] = t−1 .
1
This is valid if t ∈ Inv(Γ), or equivalently if 0 ∈
/ σ(t). More generally, if f (z) = , then f is
z−α
−1
holomorphic on C \ {α} and we may define f [t] ∈ Γ as f [t] = (t − αe) provided α ∈ / σ(t). In view
of this observation, we ask a restricted question below, and it will be answered in the affirmative.

Question: Let Γ be a unital Banach algebra and t ∈ Γ. Let U ⊂ C be an open neighborhood of

σ(t) and f : U → C be holomorphic. Can we define f [t] ∈ Γ in some natural manner?

Definitions: (i) Let K ⊂ U ⊂ C, where K is a nonempty compact set and U is open. We say ∆ is a
closed contour in U encircling K if ∆ is a closed contour in U \ K such that n(∆, w) = 1 for every
w ∈ K and n(∆, w) = 0 for every w ∈ C \ U . Recall from Exercise-3 that such a closed contour
always exists. (ii) If U ⊂ C is open, let H(U ) = {f : U → C : f is holomorphic}.

Exercise-8: Let U ⊂ C be open, f ∈ H(U ), and let a ∈ U be with f (a) = 0. Then there is g ∈ H(U )
such that f (z) = (z − a)g(z) for every z ∈ U . [Hint: Use power series to find the required g ∈ H(B)
for some ball B = B(a, ²) ⊂ U , and extend g to U by setting g(z) = f (z)/(z − a) for z ∈ U \ B.]

In [114] and [115] below, it is convenient to use the notation z instead of α for an element of C.
 OPERATOR THEORY - PART 1/3 11

[114] Let Γ be a unital Banach algebra, t ∈ Γ. Let ∆ be a closed contour in C encircling σ(t), and
1 R
let R be the resolvent function of t given by R(z) = (ze − t)−1 . Then, p(z)R(z)dz = p[t] for
2πi ∆
any complex polynomial p.

Proof. By linearity, it suffices to consider p(z) = z k . Let g : ρ(t) → Γ be the holomorphic map
g(z) = z k R(z). If ∆1 , ∆2 are two closed contours in C encircling σ(t), then n(∆1 − ∆2 , w) = 0
R R R
for every w ∈ σ(t), and hence ∆ g = 0, or ∆1 g = ∆2 g by Cauchy’s integral theorem [109]. So
we may assume ∆ is the usual parametrization for the circle {z ∈ C : |z| = b}, where b > ktk.
P P∞ k−1−n n
Now for z ∈ [∆], we have g(z) = z k−1 (e − t/z)−1 = z k−1 ∞ n=0 z
−n tn =
n=0 z t . And
R P∞ R k−1−n n
∆ g(z)dz = n=0 ∆ z t dz since uniform convergence on the compact set [∆] allows to
interchange integration and summation. If n 6= k, then z k−n tn /(k − n) is a primitive of z k−1−n tn ,
R R R
and therefore ∆ z k−1−n tn dz = 0. Hence ∆ g(z)dz = ∆ z −1 tk dz = 2πitk . ¤

Remark: The result above guarantees that the definition of f [t] below agrees with the earlier
definition when f is a polynomial.

[115] [Holomorphic functional calculus] Let Γ be a unital Banach algebra, t ∈ Γ, let U ⊂ C be an

open neighborhood of σ(t), and let ∆ be a closed contour in U encircling σ(t). Let R(z) = (ze−t)−1
1 R
for z ∈ ρ(t). For f ∈ H(U ), define f [t] ∈ Γ as f [t] = f (z)R(z)dz. Then,
2πi ∆
(i) f [t] is well-defined (i.e., the integral actually exists in Γ).
(ii) The value f [t] is independent of the particular choice of ∆.
(iii) [Homomorphism property] For f, g ∈ H(U ) and a, b ∈ C, we have (af + bg)[t] = af [t] + bg[t]
and (f g)[t] = f [t]g[t] = g[t]f [t].
(iv) [Continuity] If f, fn ∈ H(U ) and (fn ) → f uniformly on all compact K ⊂ U , then (fn [t]) → f [t].
P∞ n
(v) If f ∈ H(U ) has a power series expansion f (z) = n=0 an (z − a) for every z ∈ U , then
P P∞
f [t] = ∞ n
n=0 an (t − ae) (in particular, if a = 0, then f [t] =
n
n=0 an t ).

(vi) If f = p/q ∈ H(U ), where p, q are complex polynomials, then f [t] = p[t]q[t]−1 = q[t]−1 p[t].
(vii) [Spectral mapping theorem] σ(f [t]) = f (σ(t)) for every f ∈ H(U ).
(viii) [Composition property] Let f ∈ H(U ), let V ⊂ C be an open neighborhood of f (U ) and let
g ∈ H(V ). Then (g ◦ f )[t] = g[f [t]].

Proof. (i) V := U \ σ(t) is an open neighborhood of [∆], and the map z 7→ f (z)R(z) is holomorphic
on V (see [110]). Hence the integral exists and so f [t] is well-defined.

(ii) For i = 1, 2, let ∆i be a closed contour in U encircling σ(t). Put ∆ = ∆1 − ∆2 , let V = U \ σ(t),
and let F : V → Γ be the holomorphic map F (z) = f (z)R(z). For i = 1, 2, we have n(∆i , w) = 0
12 T.K.SUBRAHMONIAN MOOTHATHU

for w ∈ C \ U and n(∆i , w) = 1 for w ∈ σ(t) so that n(∆, w) = n(∆1 , w) − n(∆2 , w) = 0 for every
R R R
w ∈ C \ V . Therefore, ∆1 F − ∆2 F = ∆ F = 0 by Cauchy’s integral theorem, [109].

(iii) Linearity is clear. To prove the product rule, the trick is to choose one contour in the ‘interior’
of another. Let V = {w ∈ U \ [∆] : n(∆, w) = 1} (the ‘interior’ of [∆]). Then V is open and
σ(t) ⊂ V . Let Ω be a closed contour in V encircling σ(t). By the resolvent identity [110](i), we get
R R R R R(w) − R(z)
(2πi)2 f [t]g[t] = ∆ Ω f (z)g(w)R(z)R(w)dwdz = ∆ Ω f (z)g(w) dwdz
z−w
· ¸ · ¸
R R R(w) R R R(z)
= ∆ Ω f (z)g(w) dwdz + ∆ Ω f (z)g(w) dwdz = I1 + I2 , say.
z−w w−z
By our choice of V and Ω, n(∆, w) = 1 for every w ∈ [Ω] and n(Ω, z) = 0 for every z ∈ C \ V (in
particular for every z ∈ [∆]). Also, Fubini’s theorem allows the interchange of the integrals since
the integrand is continuous. Therefore, by Cauchy’s integral theorem and formula [109],
· ¸
R R f (z) R
I1 = Ω ∆ dz g(w)R(w)dw = Ω [2πif (w)]g(w)R(w)dw = (2πi)2 (f g)[t], and
z − w· ¸
R R g(w) R
I2 = ∆ f (z)R(z) Ω dw dz = ∆ f (z)R(z) × 0 dz = 0.
w−z
This establishes f [t]g[t] = (f g)[t]. Since f g = gf , we also get (f g)[t] = g[t]f [t].

(iv) By hypothesis, (fn R) → f R uniformly on [∆]. This allows to interchange limit and integration.
P
(v) If pk is the polynomial pk = kn=0 an (z − a)n , then (pk ) → f uniformly on compact subsets of
U . So we may apply (iv) and [114].

(vi) p[t] = (f q)[t] = f [t]q[t] by part (iii), and hence f [t] = p[t]q[t]−1 . Similarly, f [t] = q[t]−1 p[t].

(vi) Let α ∈ f (σ(t)), and let β ∈ σ(t) be such that f (β) = α. By Exercise-8, there is g ∈ H(U ) such
that α − f (z) = (β − z)g(z) = g(z)(β − z), and hence obtain αe − f [t] = (βe − t)g[t] = g[t](βe − t).
Since β ∈ σ(t), we have βe − t ∈ / Inv(Γ) and therefore αe − f [t] ∈ / Inv(Γ) by [104](v). Thus α ∈
1
σ(f [t]). Conversely suppose α ∈/ f (σ(t)). Then h(z) := is holomorphic in a neighborhood
α − f (z)
of σ(t). We see that h[t] is the inverse of αe − f [t], implying α ∈
/ σ(f [t]).

(viii) Let Ω be a closed contour in V encircling σ(f [t]) = f (σ(t)). Then we have 2πig(f [t]) =
R −1
Ω g(w)(we − f [t]) dw. Let U1 ⊂ U be an open neighborhood of σ(t) such that f (U1 ) is contained
in the ‘interior’ of Ω, and let ∆ be a closed contour in U1 encircling σ(t). For w ∈ [Ω], we have
R R(z)
(we − f [t])−1 = (2πi)−1 ∆ dz since z 7→ 1/(w − f (z)) is holomorphic in U1 . Substituting
w − f (z)
this in the first integral, and using Cauchy’s integral formula,
· ¸ · ¸
R R −1 R(z) R −1
R g(w)
2πig(f [t]) = Ω ∆ g(w) (2πi) dz dw = ∆ (2πi) Ω w − f (z) dw R(z)dz
w − f (z)
R
= ∆ g(f (z))R(z)dz = 2πi(g ◦ f )[t]. ¤

Seminar topic: Runge’s theorem from Complex Analysis.

 OPERATOR THEORY - PART 1/3 13

Remark: If U ⊂ C is open, then {f ∈ H(U ) : f is a rational function} is dense in H(U ) with

respect to uniform convergence on compact sets by a corollary of Runge’s theorem. Consequently,
f [t] is uniquely determined for f ∈ H(U ) once we define p[t] for polynomials p (see properties (iv)
and (vi) above). Thus the holomorphic functional calculus is unique.

Remark: Let Γ be a unital Banach algebra, t ∈ Γ, and U ⊂ C be an open neighborhood of σ(t). If

s ∈ Γ commutes with t, then s commutes with p[t] for every polynomial p. Using the approximation
by rational functions, we conclude sf [t] = f [t]s for every f ∈ H(U ).

Remark: Let Γ be a unital Banach algebra and t ∈ Inv(Γ). Then 0 ∈

/ σ(t) so that the function 1/z
is holomorphic in a neighborhood of σ(t). By spectral mapping theorem, σ(t−1 ) = {1/α : α ∈ σ(t)}.
Hence 1/r(t−1 ) = inf{|α| : α ∈ σ(t)} and therefore σ(t) ⊂ {α ∈ C : 1/r(t−1 ) ≤ |α| ≤ r(t)}.

Remark: By [115], f [t] is defined for all entire functions f . In particular, exp[t] is defined and
P
we also have exp[t] = ∞ n
n=0 t /n!. We may note: (i) exp[0] = e, (ii) exp[s + t] = exp[s]exp[t] for
commuting elements s, t ∈ Γ (to see this, consider the RHS as a product of two power series, group
together terms of the same degree, and observe that sk tn−k /(k!(n − k)!) = nC
ks
k tn−k /n!), (iii)
exp[t] ∈ Inv(Γ) for every t ∈ Γ and exp[t]−1 = exp[−t].

[116] Let Γ be a unital Banach algebra, t ∈ Γ, let U ⊂ C be an open neighborhood of σ(t) and let
f ∈ H(U ). Then,
(i) If f [t] = 0 ∈ Γ, then f |σ(t) ≡ 0.
(ii) f [t] ∈ Inv(Γ) ⇔ f has no zeroes in σ(t).

Proof. (i) f (σ(t)) = σ(f [t]) = σ(0) = {0}.

(ii) Suppose α ∈ σ(t) and f (α) = 0. Then there is g ∈ H(U ) such that f (z) = (α − z)g(z) =
g(z)(α−z) by Exercise-8. Hence f [t] = (αe−t)g[t] = g[t](αe−t). Since αe−t ∈
/ Inv(Γ), we conclude
f [t] ∈
/ Inv(Γ) by [104](v). Conversely assume f has no zeroes in σ(t). Then h := 1/f is holomorphic
in a neighborhood of σ(t). And e = (f h)[t] = f [t]h[t] = h[t]f [t] so that f [t]−1 = h[t]. ¤

Observe that if K ⊂ C is compact, then exactly one connected component of C\K is unbounded.

[117] Let Γ be a unital Banach algebra, t ∈ Γ, and suppose 0 belongs to the unbounded component
of C \ σ(t). Then,
(i) [Existence of logarithm] There is s ∈ Γ such that exp[s] = t.
(ii) [Existence of nth root] For every n ∈ N, there is u ∈ Γ such that un = t.

Proof. We claim there is a simply connected open neighborhood U of σ(t) such that 0 ∈
/ U (here,
‘simply connected’ means ‘without holes’). Let V be the unbounded component of C \ σ(t). Let
w ∈ C be with |w| > ktk. Then w ∈ V . Since V is an open connected subset of C, V is polygonally
connected. So there is a polygonal path (i.e., a piecewise linear path) ∆ from 0 to w in V . We
14 T.K.SUBRAHMONIAN MOOTHATHU

may assume ∆ has no self-intersections. Setting U = B(0, |w|) \ [∆], verify the initial claim. Now
we know from Complex Analysis that a branch of logarithm can be defined on such an open set
U . That is, there is f ∈ H(U ) with exp(f (z)) = z for every z ∈ U . Then exp[f [t]] = t by the
composition rule, [115](viii). Take s = f [t] for (i) and u = exp[s/n] for (ii). ¤

Remark: Let Γ be a unital Banach algebra, t ∈ Γ, and suppose ke − tk < 1. Then 0 belongs to
the unbounded component of σ(t) and the conclusions of [117] hold for t. To see this, note that
σ(t − e) ⊂ B(0, 1) and hence σ(t) ⊂ B(1, 1) by applying Exercise-6(ii) with p(z) = z − 1.

Exercise-9: [Application to Linear Algebra] Let T be an invertible k × k complex matrix. Then,

P∞ n
(i) there is a k × k complex matrix S such that T = exp[S] := n=0 S /n!, and (ii) for each
n ∈ N, there is a k × k complex matrix U such that T = U n . [Hint: Since T ∈ Γ(Ck ) is invertible,
0∈
/ σp (T ) = σ(T ). Since σ(T ) is finite, C \ σ(T ) is unbounded and connected. So [117] applies.]

Exercise-10: Let X be a Banach space, T ∈ Γ(X), let U ⊂ C be an open neighborhood of σ(T ) and
let f ∈ H(U ). Then, (i) f (σp (T )) ⊂ σp (f [T ]), and (ii) the reverse inclusion (and hence equality)
holds if f is not a constant on any connected component of U .
[Hint: (i) Let α ∈ σp (T ) ⊂ U . By Exercise-8, there is g ∈ H(U ) such that f (z)−f (α) = g(z)(z −α)
and hence f [T ] − αI = g[T ](T − αI). (ii) Let α ∈ σp (f [T ]) ⊂ σ(f [T ]) = f (σ(T )). Only finitely
many connected components of U can intersect σ(T ) since σ(T ) is compact. By hypothesis and a
property of the set of zeroes of a holomorphic function in a connected region, the set σ(T ) ∩ f −1 (α)
must be finite, say {β1 , . . . , βn }. So α−f (z) = h(z)(β1 −z) · · · (βn −z) for some h ∈ H(U ) having no
zeroes in σ(T ). Then h[T ] ∈ Inv(Γ(X)) by [116](ii) and h[T ]−1 (αI −f [T ]) = (β1 I −T ) · · · (βn I −T )
so that βj I − T is not injective for some j.]

Remark: Another important application of holomorphic functional calculus is Riesz decomposition

theorem. This will be discussed later in connection with the invariant subspace problem.

5. Commutative case: maximal ideals and multiplicative functionals

A Banach algebra Γ is commutative if st = ts for every s, t ∈ Γ. When Γ is a commutative

unital Banach algebra, we will show there exist a nonempty compact Hausdorff space K and a
natural map from Γ to C(K, C). Certain properties of C(K, C) may be transferred to Γ via this
map, thereby providing a better picture of Γ. This has applications even in the non-commutative
setting since a commutative Banach algebra may sit inside a non-commutative one.

Definition (purely algebraic): Let Γ be a commutative, unital Banach algebra. We say Λ ⊂ Γ is

an ideal of Γ if (i) Λ is a vector subspace of Γ, and (ii) tΛ ⊂ Λ for every t ∈ Γ. An ideal Λ 6= Γ
is said to be a proper ideal. A proper ideal Λ of Γ is maximal if there are no proper ideals of Γ
 OPERATOR THEORY - PART 1/3 15

strictly containing Λ. In a general (i.e., not necessarily commutative) unital Banach algebra, we
may define left/right/two-sided ideals by appropriately modifying condition (ii) above.

Remark: For a better understanding, compare with the theory of ideals in commutative rings.

[118] Let Γ be a commutative, unital Banach algebra. Then,

(i) Every proper ideal of Γ is contained in a maximal ideal.
(ii) An ideal Λ ⊂ Γ is proper iff Λ ∩ Inv(Γ) = ∅.
(iii) If Λ ⊂ Γ is a proper ideal, then so is Λ. Consequently, every maximal ideal of Γ is closed in Γ.
(iv) If Λ is a maximal ideal of Γ, then the quotient space Γ/Λ becomes a commutative, unital Banach
algebra with respect to the natural operations and the quotient norm kt+Λk := inf{ks+tk : s ∈ Λ}.
(v) Let t ∈ Γ. Then tΓ is an ideal of Γ. Moreover, t ∈ Inv(Γ) iff tΓ = Γ.
S
(vi) Let MI (Γ) = {all maximal ideals of Γ}. Then MI (Γ) 6= ∅ and Λ∈MI (Γ) Λ = Γ \ Inv(Γ).

Proof. For (i), apply Zorn’s lemma to the collection of all proper ideals containing the given one,
partially ordered by inclusion. Statement (ii) is easy and (iii) essentially follows from (ii) since
Inv(Γ) is open. Statements (iv) and (v) are routine verifications. Now we prove (vi). There is Λ ∈
S
MI (Γ) containing the proper ideal {0} by (i), and so MI (Γ) 6= ∅. We have Λ∈MI (Γ) Λ ⊂ Γ\Inv(Γ)
by (ii). For the other inclusion, consider t ∈ Γ \ Inv(Γ) and get a maximal ideal containing tΓ. ¤

The example below throws some light on the theory that is going to follow.

[119] Let K be a nonempty compact Hausdorff space, and consider the commutative unital Banach
algebra C(K, C) = {f : K → C : f is continuous}. Then,
(i) If a, b ∈ K are distinct, then there is f ∈ C(K, C) such that f (a) = 0 and f (b) = 1.
(ii) Λ ⊂ C(K, C) is a maximal ideal iff there is a ∈ K such that Λ = {f ∈ C(K, C) : f (a) = 0}.
(iii) Λ ⊂ C(K, C) is a closed, proper ideal iff there is a nonempty closed subset A ⊂ K such that
Λ = {f ∈ C(K, C) : f (A) = {0}}.

Proof. (i) Since a compact Hausdorff space is normal, Urysohn’s lemma from Topology gives the
required f . If K is metrizable, directly define f (x) = d(a, x)/[d(a, x) + d(b, x)].

(ii) For a ∈ K, let Λa = {f ∈ C(K, C) : f (a) = 0} which is clearly a proper ideal. Moreover, Λa
is not contained in Λb whenever a 6= b by (i). Since any proper ideal is contained in a maximal
ideal, to prove (ii) now it suffices to show any maximal ideal Λ of C(K, C) is contained in Λa for
some a ∈ K. Consider a maximal ideal Λ. If possible, let there exist fa ∈ Λ with fa (a) 6= 0 for
each a ∈ K. Choose an open neighborhood Ua of a such that fa (x) = 6 0 for every x ∈ Ua . By
Pn P
compactness, obtain a finite cover {Ua1 , . . . , Uan } for K and let f = j=1 |faj |2 = nj=1 faj faj .
Then f ∈ Λ since Λ is an ideal, and f (x) 6= 0 for every x ∈ K so that f ∈ Inv(C(K, C)). This
contradicts the fact that Λ 6= C(K, C) by [118](vi).
16 T.K.SUBRAHMONIAN MOOTHATHU

(iii) ‘⇒’: Let A = {a ∈ K : f (a) = 0 ∀ f ∈ Λ}. Obviously A is closed, and A 6= ∅ since Λ ⊂ Λa for
some a ∈ K by (ii). Putting ΛA = {f ∈ C(K, C) : f (A) = {0}}, we have Λ ⊂ ΛA by definition.
Now consider h ∈ ΛA and ² > 0. If we find g ∈ Λ with kh − gk ≤ 2², then h ∈ Λ by the closedness of
Λ as ² > 0 is arbitrary, completing the proof. For δ > 0, let Uδ = {a ∈ K : |h(a)| < δ} which is an
open neighborhood of A. Assume h 6= 0, and then by assuming ² > 0 to be small enough, suppose
K \ U2² 6= ∅. By Urysohn’s lemma, there is g1 ∈ C(K, C) such that g1 (K) ⊂ [0, 1], g1 (U² ) = {0}
and g1 (K \ U2² ) = {1}. Also, for each a ∈ K \ U² , there is fa ∈ Λ with fa (a) 6= 0 so that as in part
(ii), we can obtain g2 ∈ Λ satisfying g2 (K \ U² ) ⊂ (0, ∞). By Tietze extension theorem, there is
g3 ∈ C(K, C) such that g3 = 1/g2 on K \ U² . Take g = g1 g2 g3 h which belongs to Λ since g2 ∈ Λ.
We get kh − gk ≤ 2² since (g1 g2 g3 )(U2² ) ⊂ [0, 1], and g1 g2 g3 ≡ 1 on K \ U2² . ¤

Remark: (i) Thus the collection of maximal ideals of C(K, C) may be identified with the set K.
(ii) Fixing a ∈ K, note that the map φ : C(K, C) → C given by φ(f ) = f (a) is a bounded linear
functional on C(K, C) satisfying the multiplicative property φ(f g) = φ(f )φ(g).

Definition: Let Γ be a commutative, unital Banach algebra. We say φ ∈ Γ∗ \ {0} is a multiplicative

functional on Γ if φ(st) = φ(s)φ(t) ∀ s, t ∈ Γ. Let MF (Γ) = {all multiplicative functionals on Γ}.
For φ ∈ MF (Γ), note that φ(e) = φ(ee) = φ(e)φ(e) and so φ(e) = 1 (∵ φ(e) 6= 0 since φ 6= 0).

[120] Let Γ be a commutative, unital Banach algebra. Then the map φ 7→ ker(φ) is a bijection
from MF (Γ) to MI (Γ).

Proof. If φ ∈ MF (Γ), then clearly ker(φ) is an ideal of Γ. And ker(φ) 6= Γ since φ 6= 0. If Λ is

an ideal containing ker(φ) and if t ∈ Λ \ ker(φ), then φ(t)e ∈ Inv(Γ) and φ(t)e = [φ(t)e − t] + t ∈
ker(φ) + Λ ⊂ Λ. So Λ = Γ by [118](ii) and hence ker(φ) is maximal.
If φ, ψ ∈ MF (Γ) and if t ∈ Γ is with φ(t) 6= ψ(t), then t − φ(t)e ∈ ker(φ) \ ker(ψ) since
φ(e) = 1 = ψ(e), and this shows φ 7→ ker(φ) is injective.
Next we show surjectivity. Let Λ ∈ MI (Γ), let Γ e = Γ/Λ and let π : Γ → Γ e be the quotient
map which is bounded linear and multiplicative in the sense π(st) = π(s)π(t). If e
t∈Γe \ {0}, then
π −1 (e e is an ideal in Γ properly containing Λ, and hence is equal to Γ. Therefore e
t Γ) tΓe = Γ.
e By
e =Γ
[118](v), Inv(Γ) e∼
e \ {0}, and then Γ = C by Gelfand-Mazur theorem (Exercise-5). Thus we may
think of π as an element of MF (Γ), and clearly ker(π) = Λ. ¤

Remark: Let K be a nonempty compact Hausdorff space and Γ = C(K, C). Combining [119] and
[120] we see that MF (Γ) = {φa : a ∈ K}, where φa : Γ → C is φa (f ) = f (a).

[121] [Corollary] Let Γ be a commutative, unital Banach algebra and t ∈ Γ. Then,

(i) t ∈ Inv(Γ) iff φ(t) 6= 0 for every φ ∈ MF (Γ).
(ii) σ(t) = {φ(t) : φ ∈ MF (Γ)}.
 OPERATOR THEORY - PART 1/3 17

(iii) r(t) = sup{|φ(t)| : φ ∈ MF (Γ)}.

Proof. For (i), use [118](vi) and [120]. By (i), αe − t ∈

/ Inv(Γ) iff 0 = φ(αe − t) = α − φ(t) for some
φ ∈ MF (Γ), proving (ii). And (iii) follows from (ii). ¤

Remark: We have included continuity as one of the requirements in our definition of a multiplicative
functional, but this is really not necessary as indicated below.

Exercise-11: Let Γ be a commutative, unital Banach algebra. If φ : Γ → C is a non-zero linear

map with φ(st) = φ(s)φ(t) for every s, t ∈ Γ, then |φ(t)| ≤ ktk for every t ∈ Γ. Consequently, φ
is continuous, φ ∈ MF (Γ), and kφk = 1. [Hint: As in [120], verify that ker(φ) ∈ MI (Γ). Since
e − t/φ(t) ∈ ker(φ), we have e − t/φ(t) ∈
/ Inv(Γ) and hence kt/φ(t)k ≥ 1. From kφk ≤ 1 and
φ(e) = 1, we get kφk = 1.] [Remark: The conclusions apply to every φ ∈ MF (Γ).]

Definition: Let Γ be a commutative, unital Banach algebra. Because of [120], we identify MI (Γ)
and MF (Γ), and just write M. We call M the Gelfand space of Γ. Members of M are thought of
either as maximal ideals or as multiplicative functionals as is conveneint. For t ∈ Γ, let Et : Γ∗ → C
be the evaluation map Et (φ) = φ(t). The smallest topology on Γ∗ such that Et is continuous for
every t ∈ Γ is known as the weak* topology on Γ∗ . Since M = MF (Γ) ⊂ Γ∗ , the weak* topology
on Γ∗ induces a subspace topology on M. This is called the Gelfand topology on M.

[122] Let Γ be a commutative, unital Banach algebra. Then the Gelfand space M of Γ is a
nonempty compact Hausdorff space with respect to the Gelfand topology. Moreover, M is metriz-
able iff Γ is separable.

6 ∅ by [118](vi). Since the evaluation maps (trivially) separate points of Γ∗ , the weak*
Proof. M =
topology on Γ∗ is Hausdorff, and hence M is Hausdorff. By Exercise-11, M is contained in the
closed unit ball of Γ∗ , and this closed ball is weak* compact by Alaoglu’s theorem from Functional
Analysis. So to prove M is compact, it remains to show M is weak* closed. For this, write M as an
T ³T ´
intersection of closed sets, M = {φ ∈ Γ∗ : Ee (φ) = 1} s,t∈Γ {φ ∈ Γ ∗ : (E − E E )(φ) = 0} .
st s t

The last assertion follows since the weak* topology on the closed unit ball of Γ∗ is metrizable iff Γ
is separable; see some textbook on Functional Analysis. ¤

Definition: Let Γ be a commutative, unital Banach algebra, M be the Gelfand space of Γ equipped
with the Gelfand topology, and consider C(M, C) with the sup-norm. The map E : Γ → C(M, C)
which sends t ∈ Γ to the evaluation map Et ∈ C(M, C), Et (φ) = φ(t), is called the Gelfand mapping
of Γ (so here we think of M as the collection of multiplicative functionals on Γ).

What is stated below is more or less a summary of some of the results above. By a subalgebra
of a Banach algebra we mean a vector subspace that is closed under multiplication - note that by
our definition, a subalgebra may not be topologically closed and may not contain e.
18 T.K.SUBRAHMONIAN MOOTHATHU

[123] [Gelfand mapping theorem] Let Γ be a commutative, unital Banach algebra. Then the Gelfand
space M of Γ is a nonempty compact Hausdorff space with respect to the Gelfand topology, and
the Gelfand mapping E : Γ → C(M, C) satisfies the following properties:
(i) E is a homomorphism of unital Banach algebras (i.e, E is linear, Est = Es Et , and Ee = 1M ).
(ii) kEt k ≤ ktk for every t ∈ Γ and hence E is continuous.
(iii) E(Γ) is a (possibly non-closed) subalgebra of C(M, C) that separates points of M (this means
the trivial fact that if φ, ψ ∈ M are distinct, then there is t ∈ Γ with Et (φ) 6= Et (ψ)).
T T
(iv) ker(E) = φ∈MF (Γ) ker(φ) = Λ∈MI (Γ) Λ (called the radical of Γ).
(v) σ(t) = Et (M) and r(t) = kEt k for every t ∈ Γ.

Remark: Let (K, T ) be a nonempty compact Hausdorff space and let Γ = C(K, C). Then the
Gelfand space M of Γ is bijective with K by [119], so identify M = K as sets. If TG is the Gelfand
topology on M = K, then TG ⊂ T by definition so that the identity map I : (K, T ) → (K, TG ) is
a continuous bijection. Since the domain is compact Hausdorff and the range is Hausdorff, I is a
homeomorphism and hence TG = T . In other words, M = K as topological spaces. Moreover, it
may be seen that the Gelfand mapping E : Γ → C(K, C) is just the identity map.

Exercise-12: [Banach-Stone theorem] Let K1 , K2 be nonempty compact Hausdorff spaces. Then

K1 is homeomorphic to K2 ⇔ C(K1 , C) ∼ = C(K2 , C) as Banach algebras. [Hint: ⇐: M1 is
homeomorphic to M2 (Gelfand spaces), and Mi is homeomorphic to Ki by the above Remark.]

6. Spectrum relative to a closed subalgebra

Let Γ be a unital Banach algebra (not necessarily commutative) and let Π be a closed subalgebra
of Γ containing e. It can happen that an element of Π is invertible in Γ but not in Π, so the spectrum
of an element depends on the underlying Banach algebra. We investigate what happens to the
spectrum when the background is restricted to a closed subalgebra (this investigation is relevant
since we may be able to choose a commutative closed subalgebra, and then Gelfand theory from
the previous section becomes applicable). We will see that the spectral radius does not change,
and the spectrum becomes possibly bigger, but not much, when considered with respect to Π.

Definition: Let ∅ 6= K ⊂ C be compact. The polynomial hull of K is [K]P = {z ∈ C : |p(z)| ≤

kpkK for every polynomial p} where kpkK is the sup-norm of p on K. The bounded hull of K is
[K]B = C \ U∞ , where U∞ is the unbounded connected component of C \ K (i.e., [K]B is the union
of K and all bounded connected components of C \ K). Eventually we will show [K]B = [K]P .

Exercise-13: Let ∅ 6= K ⊂ C be compact. Then [K]B , [K]P are compact and K ⊂ [K]B ⊂ [K]P .
[Hint: Consider p(z) = z to see [K]P is bounded. Since ∂[K]B ⊂ ∂K, we have kpk[K]B ≤ kpkK for
any polynomial p by the maximum modulus principle, and therefore [K]B ⊂ [K]P .]
 OPERATOR THEORY - PART 1/3 19

We will use the notations σ(t, Π), ρ(t, Π), and r(t, Π) when the spectrum, resolvent, and spectral
radius are considered with respect to a closed subalgebra Π of Γ. The essence of the result below
is that σ(t, Π) is equal to the union of σ(t, Γ) and some bounded connected components of ρ(t, Γ).

[124] Let Γ be a unital Banach algebra, t ∈ Γ, and let Π be a closed subalgebra of Γ containing
S
{e, t}. Write ρ(t, Γ) = U∞ ∪ ( i∈I Ui ), where U∞ is the unique unbounded connected component
of ρ(t, Γ) and Ui ’s are the bounded connected components of ρ(t, Γ). Then,
(i) r(t, Π) = r(t, Γ).
(ii) ρ(t, Π) is a clopen subset of ρ(t, Γ), and U∞ ⊂ ρ(t, Π).
S
(iii) There is J ⊂ I such that σ(t, Π) = σ(t, Γ) ∪ ( i∈J Ui ) and thus σ(t, Γ) ⊂ σ(t, Π) ⊂ [σ(t, Γ)]B .
(iv) ∂σ(t, Π) ⊂ ∂σ(t, Γ).
(v) If ρ(t, Γ) is connected (in particular, if σ(t, Γ) ⊂ R), then σ(t, Π) = σ(t, Γ).

Proof. (i) r(t) = limn→∞ ktn k1/n which is independent of the underlying closed subalgebra.
(ii) We have ρ(t, Π) ⊂ ρ(t, Γ) since Inv(Π) ⊂ Inv(Γ). We know ρ(t, Π) is open in C, and hence
open in ρ(t, Γ). Now, ρ(t, Π) is closed in ρ(t, Γ) since Π is closed and ρ(t, Π) = R−1 (Π), where
R : ρ(t, Γ) → Γ is R(α) = (αe − t)−1 . By the clopenness, ρ(t, Π) is a union of some of the connected
components of ρ(t, Γ), and U∞ ⊂ ρ(t, Π) since σ(t, Π) is compact.
(iii) This follows from (ii).
S S
(iv) We have ∂σ(t, Π) = ∂ρ(t, Π) = ∂U∞ ∪ ( i∈I\J ∂Ui ) ⊂ ∂U∞ ∪ ( i∈I ∂Ui ) = ∂ρ(t, Γ) = ∂σ(t, Γ).
(v) This follows from (ii). ¤

Remark: As Π becomes smaller and smaller, more and more bounded connected components of
ρ(t, Γ) get included in σ(t, Π). Below we see that when Π becomes the smallest, then all bounded
connected components of ρ(t, Γ) are included in σ(t, Π).

Exercise-14: Let Γ be a unital Banach algebra, t ∈ Γ, and let K = σ(t, Γ). If Π is the closed
subalgebra generated by {e, t}, then σ(t, Π) = [K]B = [K]P . [Hint: By [124] and Exercise-13, it
remains to show [K]P ⊂ KΠ := σ(t, Π). If α ∈
/ KΠ , then (αe−t)−1 ∈ Π = {p[t] : p is a polynomial}.
So there is polynomial p with kp[t] − (αe − t)−1 k < 1/kαe − tk. If q(z) = p(z)(α − z) − 1, then
/ [K]P .]
|q(α)| = 1. But kqkK ≤ kq[t]k = kp[t](αe − t) − ek < 1 since q(K) = σ(q[t], Γ). Thus α ∈

Remark: If ∅ 6= K ⊂ C is compact, by [102] and Exercise-14 we deduce that (i) [K]B = [K]P
always, and (ii) K is polynomially convex (i.e., [K]P = K) iff C \ K is connected.

The next result helps to apply the theory of commutative algebras to the non-commutative ones.

Definition: Let Γ be a unital Banach algebra. A subset Λ ⊂ Γ is said to be commutative if st = ts

for every s, t ∈ Λ. If Λ ⊂ Γ is an arbitrary subset (not necessarily commutative), we define the
commutant Λc := {t ∈ Γ : st = ts for every s ∈ Λ}. The bi-commutant of Λ is Λcc = (Λc )c .
20 T.K.SUBRAHMONIAN MOOTHATHU

[125] Let Γ be a unital Banach algebra and let Λ ⊂ Γ be a commutative set. Then,
(i) Λcc is a commutative, closed subalgebra of Γ containing {e} ∪ Λ.
(ii) σ(t, Λcc ) = σ(t, Γ) for every t ∈ Λcc (in particular, for every t ∈ Λ).

Proof. (i) It is easy to verify Λc is a closed subalgebra of Γ with e ∈ Λc for any set Λ. So the same
argument applied to Λc gives that Λcc is a closed subalgebra of Γ with e ∈ Λcc . Since st = ts for
every s ∈ Λ and t ∈ Λc , we have Λ ⊂ Λcc . Since Λ is commutative, we have Λ ⊂ Λc and therefore
Λc ⊃ Λcc . From this we see that if s, t ∈ Λcc , then s ∈ Λcc and t ∈ Λc so that st = ts.
(ii) Let α ∈ ρ(t, Γ) and let s ∈ Λc . Since t, αe ∈ Λcc , we have αe − t ∈ Λcc and hence (αe − t)s =
s(αe − t). This implies s(αe − t)−1 = (αe − t)−1 s and so (αe − t)−1 ∈ Λcc . Thus α ∈ ρ(t, Λcc ). ¤

Exercise-15: Let Γ be a unital Banach algebra, and let s, t ∈ Γ be such that st = ts. Then,
(i) σ(s + t) ⊂ σ(s) + σ(t) and hence r(s + t) ≤ r(s) + r(t).
(ii) σ(st) ⊂ σ(s)σ(t) and hence r(st) ≤ r(s)r(t).
[Hint: In view of [125], replacing Γ with {s, t}cc we may assume Γ is commutative. For φ ∈ M
(the Gelfand space), note that φ(s + t) = φ(s) + φ(t) and φ(st) = φ(s)φ(t). Now apply [121](ii).]

Remark: Let Γ be a unital Banach algebra. We have mentioned earlier (see the Remark just
before [113]) that the spectral radius function r : Γ → R is not continuous in general. Now
suppose Γ is commutative. Then by Exercise-15 we have r(s) = r(s − t + t) ≤ r(s − t) + r(t), or
r(s)−r(t) ≤ r(s−t) ≤ ks−tk. Interchanging the roles of s and t, finally we have |r(t)−r(s)| ≤ kt−sk,
and thus r is Lipschitz continuous when Γ is commutative.

Remark: The assumption st = ts is necessary in Exercise-15. If we consider the nilpotent elements

S, T ∈ Γ(C2 ) given by S(e1 ) = 0 = T (e2 ), S(e2 ) = e1 and T (e1 ) = e2 , then σ(S) = {0} = σ(T ) and
hence σ(S) + σ(T ) = {0} = σ(S)σ(T ). On the other hand, 1 ∈ σ(S + T ) ∩ σ(ST ) (verify).

Remark: If T is the right shift (x0 , x1 , . . .) 7→ (0, x0 , x1 , . . .) and S is the left shift (x0 , x1 , . . .) 7→
(x1 , x2 , . . .) on l2 , then ST = I is invertible, but T S is not invertible (since it is not injective). In
other words, 0 ∈ σ(T S) \ σ(ST ). However, we can say something positive:

Exercise-16: Let Γ be a unital Banach algebra and let s, t ∈ Γ. Then,

(i) e − st ∈ Inv(Γ) iff e − ts ∈ Inv(Γ).
(ii) σ(st) \ {0} = σ(ts) \ {0}.
[Hint: (i) Informally observe that e + ts + (ts)2 + (ts)3 + · · · = e + t[e + st + (st)2 + · · · ]s. Now for
the formal argument, assume e − st ∈ Inv(Γ) and put u = (e − st)−1 . Then (e − ts)(e + tus) =
e − ts + t[(e − st)u]s = e − ts + ts = e. Similarly, (e + tus)(e − ts) = e. (ii) If α 6= 0, then note that
αe − st = α(e − α−1 st) and apply part (i) to the elements α−1 s and t.]

*****