CHAPTER 3

Operators on a Hilbert Space

This chapter is a review of basic results concerning operators on a Hilbert space.

The main reference for this chapter is the book of Reed-Simon [RS].
Throughout this chapter we let H be a separable Hilbert space and we denote
by L(H) its C ∗ -algebra of continuous endomorphisms.

3.1. Polar Decomposition

Definition 3.1.1. An operator T ∈ L(H) is said to be positive if can be written
in the form T = S ∗ S with S ∈ L(H).
We denote by L(H)+ the set of positive elements of L(H).
Proposition 3.1.2. For T ∈ L(H) the following are equivalent:
(i) T is positive.
(ii) There exists S ∈ L(H) selfadjoint such that T = S 2 .
(iii) T is selfadjoint and Sp T ⊂ [0, ∞).
(iv) hT ξ, ξi ≥ 0 for all ξ ∈ H.
Proof. We shall prove the implications (i) ⇒ (iv), (ii) ⇒ (i), (iii) ⇒ (ii) and
(iv) ⇒ (iii), which will give the proposition.
• (i) ⇒ (iv): If T = S ∗ S for some S ∈ L(H) then, for all ξ ∈ H,

hT ξ, ξi = hS ∗ Sξ, ξi = hSξ, Sξi = kSξk2 ≥ 0,

proving that (i) ⇒ (iv).
• (ii) ⇒ (i): If T = S 2 with S ∈ L(H) selfadjoint, then T = S ∗ S and hence T is
positive. Thus (ii) ⇒ (i).
• (iii) ⇒ (ii): Assume that T is selfadjoint and Sp T ⊂ [0, ∞). Then the function
√ √
t is defined and continuous on Sp T ,√and so we can define S := T by functional
continuous calculus. Notice that, as t takes real values on Sp T , it follows from
Remark 1.5.4 that S is selfdajoint. Recall that the continuous functional √ 2calculus
f → f (T ) is a homomorphism √ of algebrs from C(Sp T ) to L(H). Since ( t) = t on
Sp T , it follows that S 2 = ( T )2 = T , and hence T satisfies (ii). Thus (iii) ⇒ (ii).
• (iv) ⇒ (iii): Assume that hT ξ, ξi ≥ 0 for all ξ ∈ H. Then T is selfadjoint. Indeed,
for all ξ, η in H, we have
4hT ξ, ηi = hT (ξ +η), ξ +ηi−hT (ξ −η), ξ −ηi−ihT (ξ +iη), ξ +iηi+ihT (ξ −iη), ξ −iηi,
from which we see that hT ξ, ηi = hT η, ξi, and hence hT ξ, ηi = hξ, T ηi.
Let us now show that Sp T ⊂ [0, ∞). Since T is selfadjoint, and hence Sp T ⊂ R
by Proposition 1.3.4, we only need to show that T − λ is invertible for all λ < 0.
1
Thus let λ ∈ (−∞, 0) and let ξ ∈ H. As hT ξ, ξi ≥ 0 we have
|λ|kξk2 = −λhξ, ξi ≤ h(T − λ)ξ, ξi ≤ k(T − λ)ξkkξk.
Thus,
(3.1) k(T − λ)ξk ≤ |λ|kξk ∀ξ ∈ H.
This implies that ker(T − λ) = {0}. As T − λ is selfadjoint we then see that
im(T − λ) = (ker(T − λ))⊥ = H.
In fact, the inequality (3.1) also implies that im(T − λ) is closed. Indeed, if
η = lim(T − λ)ξn , then (3.1) implies that the sequence (ξn )n≥0 is Cauchy in H, and
hence ξn converges to some ξ in H. Then η = lim(T − λ)ξn = T ξ, showing that η
is contained in im(T − λ). Thus ker(T − λ) = 0 and im(T − λ) = H, i.e., T − λ is
bijective. Recall that by the open mapping theorem any bijective continuous linear
map of H onto H has a continuous inverse (see [Fo, p. 162]), so T −λ is an invertible
element of L(H). Thus Sp T ⊂ [0, ∞), and hence T satisfies (iii), showing that (iv)
implies (iii). The proof is complete. 
Corollary 3.1.3. The set of positive operators L(H)+ is a positive cone of
L(H), i.e.,
λ1 T1 + λ2 T2 ∈ L(H)+ ∀Tj ∈ L(H)+ ∀λj ≥ 0.
Proof. For j = 1, 2 let Tj ∈ L(H)+ and let λj ∈ [0, ∞). Using the character-
ization (iv) of Proposition 3.1.2, we see that, for all ξ ∈ H,
h(λ1 T1 + λ2 T2 )ξ, ξi = λ1 hT1 ξ, ξi + λ2 hT2 ξ, ξi ≥ 0,
proving that λ1 T1 + λ2 T2 is a positive operator. 
Corollary 3.1.4. Let T ∈ L(H) be normal and let f ∈ C(Sp T ) be non-
negative. Then the operator f (T ) is positive.
Proof. Since f is real-valued it follows from Remark 1.5.4 that f (T ) is selfad-
joint. Moreover, by (1.13) we have Sp f (T ) = f (Sp T ) ⊂ [0, ∞), so it follows from
Proposition 3.1.2 (iii) that f (T ) is positive. 
Let T ∈ L(H). Then T ∗ T is a positive operator, so by the previous proposition
∗
T T is selfadjoint and its spectrum is contained in√[0, ∞). Therefore, by continuous
functional calculus we can define its square root T ∗ T as an element of L(H). It
follows from Corollary 3.1.4 that |T | is a positive operator.
√
Definition 3.1.5. For all T ∈ L(H) the operator T ∗ T is denoted |T | and is
called the modulus of T .
Lemma 3.1.6. Let T ∈ L(H). Then
(i) |T | is the unique element of L(H)+ whose square is T ∗ T .
(ii) We have
(3.2) k|T |ξk = kT ξk ∀ξ ∈ H,
and hence ker |T | = ker T .
Proof. The continuous functional
√ 2 calculus f → f (T ∗ T ) is a ∗-homomorphism
√
∗
from C(Sp T T ) to L(H). As ( t) = t on [0, ∞) it follows that |T |2 = ( T ∗ T )2 =
T ∗T .
2
 Let S ∈ L(H)+ be such that S 2 = T . Since S√is positive, by Proposition 3.1.2
its spectrum is contained in [0, ∞),√and hence √ t2 = t on Sp S. Therefore, by
2 ∗
continuous functional calculus S = S = T T = |T |. Thus |T | is the unique
element of L(H)+ whose square is T ∗ T .
As |T | is selfadjoint |T |∗ |T | = |T |2 = T ∗ T . Therefore, for all ξ ∈ H,
k|T |ξk2 = h|T |ξ, |T ||ξ|i = h|T |∗ |T |ξ, ξi = hT ∗ T ξ, ξi = hT ξ, T ξi = kT ξk2 ,
proving (3.2). This immediately implies that |T | and T have same kernel. 
Proposition 3.1.7 (Polar Decomposition). Let T ∈ L(H). Then there exists
a unique U ∈ L(H), called the phase of T , such that
(i) T = U |T |;
(ii) ker U = ker |T |.
Proof. Since |T is selfadjoint (ker |T |)⊥ = im |T |, and hence |T | is a bijection
from im |T | onto itself. Let |T |−1 : im |T | → im |T | be its inverse and denote by U
the linear map T |T |−1 : im |T | → H. Then, for any ξ ∈ im |T |, we have
kU ξk2 = hT |T |−1 ξ, T |T |−1 ξi = hT ∗ T |T |−1 ξ, |T |−1 ξi = h|T |2 |T |−1 ξ, |T |−1 ξi
= h|T |ξ, |T |−1 ξi = hξ, |T ||T |−1 ξi = kξk2 .
Thus U uniquely extends to an isometric linear map U : im |T | → H. Extending U
to be 0 on ker |T | = (im |T |)⊥ we then get a continuous endomorphism U : H → H
whose null space is ker |T |.
If ξ ∈ im |T |, then U |T |ξ = T |T |−1 |T |ξ = ξ, and hence T = U |T | on im |T | by
continuity. Moreover, as by Lemma 3.1.6 ker |T | = ker T , we have U |T |ξ = 0 = T ξ
for all ξ in ker |T | = (im |T |)⊥ . Therefore, we see that T = U |T | on H, showing
that U satisfies the conditions (i) and (ii) of the proposition.
Let V ∈ L(H) be such that T = V |T | and ker V = ker |T |. If ξ ∈ ker |T |, then
obviously V ξ = 0 = U ξ. If ξ ∈ im |T |, then V ξ = V |T ||T |−1 ξ = T |T |−1 ξ = U ξ, so
by continuity V = U on im |T |. It follows from this that V = U on H, and hence
U is the unique element of L(H) such that T = U |T | and ker U = ker |T |, giving
the proposition. 
Proposition 3.1.8. Let T ∈ L(H) have polar decomposition T = U |T | and
denote by Π0 (T ) (resp. Π0 (T ∗ )) the orthogonal projection onto ker T (resp. ker T ∗ ).
(i) The range of U is im T .
(ii) We have
U ∗ U = 1 − Π0 (T ) and U U ∗ = 1 − Π0 (T ∗ ),
so that U is a partial isometry and has norm 1 unless T = 0.
(iii) If T is injective and has dense range, then U is unitary.
(iv) We have
|T | = U ∗ T, T ∗ = U T U ∗, |T ∗ | = T U ∗ = U |T |U ∗ .
(v) The phase of T ∗ is U ∗ .
Proof. As U vanishes on ker |T | we see that im U = U (ker |T |). Notice that
by its construction in the proof of Proposition 3.1.7 the operator U is isometric
on (ker |T |)⊥ and agrees with T |T |−1 on im |T |. In particular, it follows from
Lemma 1.1.8 that U ((ker |T |)⊥ ) is closed and U induces a unitary operator from
3
(ker |T |)⊥ onto U ((ker |T |)⊥ ) = im U . As (ker |T |)⊥ = im T we then see that
im U = U (im |T |). Since U = T |T |−1 on im |T | we have U (im |T |) ⊂ im T , and
hence im U ⊂ im T . However, as T = U |T | we also have im T ⊂ im U , and as im U
is closed we see that im T is contained is im U , and so the range of U is im T .
As abovementioned U induces a unitary operator from (ker |T |)⊥ onto im U =
im T . Since by Lemma 3.1.6 ker |T | = ker T this shows that U is a unitary endo-
morphism of H when T is injective and has dense range. In any case, as we have
im T = (ker T ∗ )⊥ we see that U induces a unitary operator from (ker T )⊥ onto
(ker T ∗ )⊥ . Since ker U = ker |T | = ker T we then deduce that U ∗ U is the orthogo-
nal projection onto (ker T )⊥ and U U ∗ is the orthogonal projection onto (ker T )⊥ ,
that is, U ∗ U = 1 − Π0 (T ) and U U ∗ = 1 − Π0 (T ∗ ). In particular, if T 6= 0 then
kU k2 = kU ∗ U k = k1 − Π0 (T )k = 1, i.e., kU k = 1.
Notice that, as ker |T | = ker T , we have |T |(1 − Π0 (T )) = (1 − Π0 (T ))|T | = |T |,
and hence U ∗ T = U ∗ U |T | = (1 − Π0 (T ))|T | = |T |. Moreover,
(U |T |U ∗ )2 = U |T |U ∗ U |T |U ∗ = U |T |(1 − Π0 (T ))|T |U ∗ = (U |T |)(U |T |)∗ = T T ∗ .
As |T | is positive, for any ξ ∈ H, we have hU |T |U ∗ ξ, ξi = h|T |U ∗ ξ, U ∗ ξi ≥ 0, and
hence U |T |U ∗ is positive. Thus U |T |U ∗ is a positive operator whose square is equal
to T T ∗ , so using Lemma 3.1.6 we see that U |T |U ∗ = |T ∗ |. Since T = U |T | this
also shows that |T ∗ | = T U ∗ .
Notice that ker U ∗ = (im U )⊥ = (im T )⊥ = ker T ∗ = ker |T ∗ |. Moreover,
U ∗ |T ∗ | = U ∗ U |T |U ∗ = (1 − Π0 (T ))|T |U ∗ = |T |U ∗ = (U |T |)∗ = T ∗ ,
so by Proposition 3.1.7 the phase of T ∗ is U ∗ . Notice that the above equalities in-
clude the equality T ∗ = |T |U ∗ . As |T | = U ∗ T we see that T ∗ = U ∗ T U ∗ , completing
the proof. 

3.2. Spectral Theorem and Borel Functional Calculus

Let T ∈ L(H) be a normal operator (i.e., T ∗ T = T T ∗ ) and set S = Sp T .
Theorem 3.2.1 (Spectral Theorem; see [RS, Thm. VII.3]). There exist a fi-
nite measure space (X, µ), a unitary operator U : H → L2µ (X), and a bounded
measurable function f on X, in such way that
(3.3) U T U ∗ξ = f ξ ∀ξ ∈ L2µ (X).
For F ∈ L∞ (X) denote by TF the multiplication operator by F , i.e., the
operator TF ∈ L(L2µ (X)) defined by
TF ξ = F ξ ξ ∈ L2µ (X).
For instance, Eq. (3.3) says that U T U ∗ = Tf .
The essential range of F consists of all λ ∈ C such that
µ(λ −  < F < λ + ) > 0 ∀ > 0.
It can be shown that Sp TF agrees with the essential range of F . In particular, we
see that the essential range of Tf is S. Thus, without any loss of generality, we may
assume that the range of f is S.
We endow L∞ (X) with its usual norm, i.e.,
kF kL∞ (X) = {λ; λ in the essential range of |F |} ∀F ∈ L∞ (R).
4
We also endow L∞ (X) with the involution F → F given by complex conjugation.
This turns L∞ (X) into a commutative unital C ∗ -algebra. Then it is not difficult
to check that the map F → TF is a ∗-homomorphism from L∞ (X) to L(L2µ (X)).
Moreover, as TF is a normal operator, we have
kTF k = sup |λ| = kF kL∞ (S) .
λ∈Sp TF

If g ∈ C(S) then g ◦ f is again a bounded measurable function on X. In fact,

the continuous functional calculus for f is just g → g ◦ f . Since the maps F → TF
(resp., F → U ∗ TF U ) is an isometric ∗-homomorphisms from L∞ (X) to L(L2µ (X))
(resp., L(H)), it follows that, for all g ∈ C(S),
(3.4) g(T ) = g(U ∗ Tf U ) = U ∗ g(TF )U = U ∗ Tg◦f U.
We then can extend the definition of g(T ) for any any bounded Borel function g
on Sp T by letting
g(T ) := U ∗ Tg◦f U.
This defines a bounded operator on H.
Theorem 3.2.2 (Borel Functional Calculus; see [RS, Thm. VII.2]). The fol-
lowing holds.
(1) The map g → g(T ) is a ∗-homomorphism from L∞ (S) to L(H) such that
(3.5) kg(T )k ≤ kgkL∞ (S) ∀g ∈ L∞ (S).
(2) If (gn )n≥0 is a bounded sequence in L∞ (S) such that gn → g a.e., then
gn (T ) → g(T ) strongly (i.e., gn (T )ξ → g(T )ξ for all ξ ∈ H).
(3) If g ∈ L∞ (S) is real-valued (resp., non-negative), then g(T ) is selfadjoint
(resp., positive).

3.3. Unbounded Operators

The spectral theorem and the Borel functional calculus can be extended to
unbounded operators as follows.
Definition 3.3.1. An (unbounded) operator on H is a linear operator T :
D(T ) → H, where the domain D(T ) is a subspace of H. T
An operator T on H is said to be densily defined when its domain D(T ) is a
dense subspace of H. An operator S is said to be an extension of T , and we write
T ⊂ S, if D(T ) ⊂ D(S) and S agrees with T on D(T ).
The graph of an operator T is defined to be
G(T ) = {(ξ, η) ∈ H ⊕ H; η = T ξ}.
The graph of T is a subspace of H ⊕ H. We say that T is closed when G(T ) is a
closed subspace of H ⊕ H. We say that T ∗ is closable when G(T ) is the graph of
an operator T . In this case we call T the closure of T . This is the smallest closed
extension of T .
If T is densily defined, then its adjoint is the operator T ∗ with graph
G(T ∗ ) = {(ξ, η) ∈ H ⊕ H; hT ζ, ξi = hζ, ηi ∀ζ ∈ D(T )}.
Since G(T ∗ ) is a closed subspace, we see that T ∗ is always a closed operator.
The resolvent set of a closed operator T consists of all λ ∈ C such that
(i) T − λ : D(T ) → H is a bijection.
5
 (ii) The inverse (T − λ)−1 : H → D(T ) is bounded.
The spectrum of T , denoted Sp T , is the complement of the resolvent set.
It can be shown that Sp T is a closed subset of C (which may be empty) and
the resolvent λ → (T − λ)−1 is analytic from C \ Sp T to L(H) (provided we regard
the inverses (T − λ)−1 as elements of L(H)).
A densily defined operator T is said to be symmetric is T ⊂ T ∗ . It is said to
be selfadjoint if T = T ∗ . Note that any symmetric operator is closable. We also
say that T is essentially selfadjoint when T is symmetric and closable.
Proposition 3.3.2 ([RS, Thm. VIII.3]). Let T be a symmetric operator on H.
Then the following are equivalent:
(i) T is selfadjoint.
(ii) T is closed and ker(T ± i) = {0}.
(iii) im(T ± i) = H.
Corollary 3.3.3 (see [RS]). Let T be a symmetric operator on H. Then the
following are equivalent:
(i) T is essentially selfadjoint.
(ii) ker(T ± i) = {0}.
(iii) im(T ± i) = H.
Let T be a selfadjoint (unbounded) operator on H.
Theorem 3.3.4 (Spectral Theorem; see [RS, Thm. VIII.4]). There exist a
measured space (X, µ) with µ(X) < ∞, a unitary operator U : H → L2µ (X), and a
measurable real-valued function f on X such that
U (D(T )) = {ξ ∈ L2µ (X); f ξ ∈ L2µ (X)},
U T U ∗ξ = f ξ ∀ξ ∈ U (D(T )) .
If g is a bounded Borel function on R, then we define g(T ) as the bounded
operator on L(H) given by
g(T ) := U ∗ Tg◦f U.
Theorem 3.3.5 (Borel Functional Calculus; [RS, Thm. VIII.5]). The following
holds.
(1) The map g → g(T ) is a ∗-homomorphism from L∞ (R) to L(H) such that
(3.6) kg(T )k ≤ kgkL∞ (R) ∀g ∈ L∞ (R).
(2) If (gn )n≥0 is a bounded sequence in L∞ (R) such that gn → g a.e., then
gn (T ) → g(T ) strongly.
(3) If g ∈ L∞ (R) is real-valued (resp., non-negative), then g(T ) is selfadjoint
(resp., positive).
More generally, if g is a possibly unbounded Borel function on R, then g(T )
makes sense as an unbounded operator as follows. The domain of g(T ) is
 
∗ 2 2
D(g(T )) = U {ξ ∈ Lµ (X); (g ◦ f )ξ ∈ Lµ (X)} ,

and we define g(T ) by the formula

g(T )ξ := U ∗ ((g ◦ f )U ξ) ∀ξ ∈ D(g(T )).
6
For instance, if g(t) = t, then g(T ) = T . In addition, it is not hard to check that,
if (gn ) be a sequence of Borel functions on R such that gn → g and |gn | ≤ |g|, then
as n → ∞ we have
gn (T )ξ −→ g(T )ξ ∀ξ ∈ D(g(T )).
Example 3.3.6. Let ∆ = −(∂x21 + . . . + ∂x2n ) be the (positive) Laplacian on Rn .
We shall regard ∆ as an operator on the space S 0 (Rn ) of tempered distributions on
Rn . Denoting by u → û the Fourier transform on S 0 (Rn ), we have
(3.7) (∆u)∧ = |ξ|2 û ∀u ∈ S 0 (Rn ).
Thus, we can regard ∆ as an unbounded operator on the Hilbert space H = L2 (Rn )
with domain,
D(∆) = {u ∈ L2 (Rn ); |ξ|2 û ∈ L2 (Rn )}.
Notice that D(∆) agrees with the Sobolev space W 2,2 (Rn ). As such ∆ is a selfad-
joint operator with spectrum [0, ∞).
Denote by U the unitary operator on L2 (Rn ) defined by
n
U v = (2π)− 2 v̂ ∀v ∈ L2 (Rn ).
Then it follows from (3.7) that
∆v = U ∗ T|ξ|2 U v ∀v ∈ W 2,2 (Rn ).
Therefore, we see that in this example the spectral theorem follows from elementary
Fourier-analytic considerations.
If g is a bounded Borel function on R, then g(∆) is given by
Z
g(∆)v(x) = (U ∗ Tg(|ξ|2 ) U v)(x) = (2π)−n eix.ξ g(|ξ|2 )û(ξ)dξ ∀v ∈ L2 (Rn ).

The following operators are of special interest:

- The heat semigroup e−t∆ , t ≥ 0.
- The wave group eit∆ , t ∈ R.
- The complex powers ∆z , <z ≥ 0.
These operators can be similarly defined by Borel functional calculus for any self-
adjoint unbounded operators with nonnegative spectrum.

3.4. Compact Operators

In the sequel for any r > 0 we denote by B(0, r) the (open) unit ball of H of
radius r about the origin.
Definition 3.4.1. An operator T ∈ L(H) is said to be compact when T (B(0, 1)
is compact in H.
We denote by K the space of compact operators.
The following lemma will be useful to study compact operators.
Proposition 3.4.2. Let T ∈ L(H). Then the following are equivalent:
(i) T is a compact operator.
(ii) For any bounded sequence (ξn )n≥0 ⊂ H there is a subsequence (ξnk )k≥0
such that the sequence (T ξnk )k≥0 converges in norm.
(iii) For any sequence (ξn )n≥0 ⊂ H converging weakly to 0 the sequence (T ξn )n≥0
converges to 0 in norm.
7
 (iv) There is an orthonormal basis (ξn )n≥0 of H such that
lim kT|EN⊥ k = 0,
N →∞

where we have denoted by EN the span of ξ0 , . . . , ξN −1 .

(v) For any orthonormal basis (ξn )n≥0 of H,
lim kT|EN⊥ k = 0,
N →∞

where we have denoted by EN the span of ξ0 , . . . , ξN −1 .

Proof. Taking into account that the implication (v) ⇒ (iv) is immediate, to
prove the proposition we only need to prove the implications (ii) ⇒ (i), (iii) ⇒ (ii),
(iv) ⇒ (iii) and (i) ⇒ (v).
• (ii) ⇒ (i): Assume that T sastifies the condition (ii). Then for any sequence
(ξn )n≥0 ⊂ B(0, 1) the sequence (T ξn )n≥0 admits a convergent subsequence. By
virtue of the Bolzano-Weierstrass criterion this proves that T (B(0, 1)) is compact,
i.e., T is a compact operator. Thus (ii) implies (i).
• (iii) ⇒ (ii): Suppose that T satisfies the condition (iii). Let (ξn )n≥0 ⊂ H be a
bounded sequence, i.e., there exists r > 0 such that ξn ∈ B(0, r) for all n. By
Alaoglu theorem (see [Fo, pp. 169-170]) the ball B(0, r) is precompact with respect
to the weak topology and, as H is separable, the weak topology is metrizable
(see [Fo]), so by the Bolzano-Weierstrass criterion there is a subsequence (ξnk )k≥0
converging weakly to some ξ in H. Then ξnk − ξ converges weakly to 0, so by (iii)
the sequence T (ξnk − ξ) converges to 0 in norm, i.e., T ξnk → T ξ in norm. This
shows that (iii) implies (ii).
• (iv) ⇒ (iii): Suppose that T satisfies (iv). Thus there exists an orthonormal basis
(ξn )n≥0 of H such that if, for any N ∈ N, we denote by EN the span of ξ0 , . . . , ξN −1
then kT|EN⊥ → 0 as N → ∞. In addition, we denote by ΠN the orthogonal projection
onto EN .
Let (ηk )k≥0 ⊂ H be a sequence converging weakly to 0. In particular (ηk )k≥0
is weakly bounded, and hence is bounded in norm by the uniform boundedness
principle and the fact H is isometrically isomorphic to its dual (see [Fo, pp. 163,
174–175 ]). Thus there exists P C > 0 such that kηk k ≤ C for all k ∈ N0 .
Let  > 0. Since ΠN ηk = n<N hξn , ηk iξn we have

(3.8)
X
kT ηk k ≤ kT ΠN ηk k + kT (1 − ΠN )ηk k ≤ |hξn , ηk i|kT ξn k + kT (1 − ΠN )kkηk k
n<N
X
≤ kT k |hξn , ηk i| + CkT|EN⊥ k.
n<N

Since kT|EN⊥ k → 0 as N → ∞ by choosing N large enough we have

(3.9) kT (1 − ΠN )k < .
P
As ηk converges weakly to 0 we see that n<N |hξn , ηk i| goes to 0 as k → ∞, and
hence there exists k0 ∈ N0 such that, for any k ≥ k0 ,
X
(3.10) |hξn , ηk i| < .
n<N

8
Combining (3.8), (3.9) and (3.10) we see that, for all k ≥ k0 , we have
kT ηk k ≤ (ν0 + C).
This shows that T ηk → 0 in norm as k → ∞. Thus (iv) implies (iii).
• (i) ⇒ (v): Suppose that T is a compact operator. Let (ξn )n≥0 be an orthonormal
basis of H. For any N ∈ N we denote by EN the span of ξ0 , . . . , ξN −1 . Assume that
the sequence (kT|EN⊥ k)N ≥1 does not converge to 0 as N → 0. Since this is a non-
increasing sequence of non-negative numbers there is c > 0 such that kT|EN⊥ k > c
⊥
for all N ∈ N. Therefore, for every N ∈ N, there is a unit vector ηN ∈ EN such
that kT ηN k > c.
⊥
P
Let ξ ∈ H. As ηN is contained in EN , and hence ηN = n≥N hξn , ηN iξn , we
have
X X 2  X 2
2 2
|hηn , ξi| ≤ |hηN , ξn ihξ, ξn i| ≤ |hηN , ξn i| |hξ, ξn i| .
n≥N n≥N n≥N

2 2 2
P P
Since n≥N |hηN , ξn i| = kηN k = 1 and n≥N |hξ, ξn i| → 0 as N → ∞, it
follows that hηn , ξi → 0 as N → ∞. Thus ηN converges weakly to 0 as N → ∞. As
T is continuous with respect to the weak topology it follows that T ηN converges
weakly to 0 as N → ∞.
On the other hand, the sequence (T ηN )n≥1 is contained in the image by T of the
unit sphere, which is precompact since T is compact. Therefore, by the Bolzano-
Weierstrass criterion there is a subsequence (ηNk )k≥0 such that T ηNk converges in
norm to some ζ as k → ∞. As T ηNk converges weakly to 0 we must have ζ = 0,
i.e., T ηNk converges to 0 in norm. This contradicts the fact that kT ηN k > c for all
N ∈ N. Therefore, it is not possible for the sequence (kT|EN⊥ k)N ≥1 to not converge
to 0. This proves that kT|EN⊥ k → 0 as N → ∞. Thus (i) implies (v). The proof is
complete. 

Proposition 3.4.3. K is a closed two-sided ideal of L(H).

Proof. Let (ξn )n≥0 be any sequence converging weakly to 0.
For j = 1, 2 let Tj ∈ K and λj ∈ C. By Proposition 3.4.2 (iii) the sequences
(T1 ξn )n≥0 and (T2 ξn )N converge to 0 in norm, and so (λ1 T1 + λ2 T2 )ξn )n≥0 too
converges to 0 in norm. It then follows from Proposition 3.4.2 (iii) that λ1 T1 + λ2 T2
is a compact operator. Thus K is a subspace of L(H).
Let T ∈ K and let A, B ∈ L(H). Since B is continuous with respect to the
weak topology the sequence (Bξn )n≥0 converges weakly to 0. Since T is compact
Proposition 3.4.2 (iii) insures us that (T Bξn )n≥0 converges to 0 in norm. Then
(AT Bξn )n≥0 converges to 0 in norm too. Thanks to Proposition 3.4.2 (iii), this
shows that AT B is a compact operator. Therefore, we see that K is a two-sided
ideal of L(H).
It remains to show that K is closed. Thus, let (Tk )k≥0 ⊂ K be a sequence such
that Tk → T in L(H) and let us show that T is compact. Let  > 0. Then for
k large enough kT − Tk k < . Since Tk is compact by Proposition 3.4.2 (iii) the
sequence (Tk ξn )n≥0 converges to 0 in norm, and hence there exists N ∈ N, such
that kTk ξn k <  for all n ≥ N . Then, for any n ≥ N , we have
kT ξn k ≤ kTk ξn k + k(T − Tk )ξn k ≤ kTk ξn k + kT − Tk kkξn k < 2.
9
Thus (T ξn )n≥0 converges to 0 in norm, which by Proposition 3.4.2 (iii) shows that
T is compact. This proves that K is closed, completing the proof. 
Proposition 3.4.4. For any T ∈ L(H),
(3.11) T ∈ K ⇐⇒ T ∗ ∈ K ⇐⇒ |T | ∈ K.
Proof. Let T ∈ L(H) and let T = U |T | be its polar decomposition. The
fact that K is a two-sided ideal then implies that if |T | is compact, then so is T .
Likewise, as by Proposition 3.1.8 we have |T | = U ∗ T , if T is compact, then so is
|T |.
Proposition 3.1.8 also tells us that T ∗ = U T U ∗ . Therefore, we also see that
if T is compact then so is T ∗ . Upon substituting T ∗ for T , we see that if T ∗ is
compact, then T is compact too. The proof is complete. 
Proposition 3.4.5. K is a sub-C ∗ -algebra of L(H), and hence is a C ∗ -algebra.
Proof. Since Proposition 3.4.3 tells us that K is a closed two-sided ideal, we
see that K is a closed subalgebra of L(H). As by Proposition 3.4.4 K is closed
under the involution of L(H) it follows that K is a sub-C ∗ -algebra of L(H). 
Theorem 3.4.6 (Riesz-Schauder; see [RS, Thm. VI.15]). Let T ∈ K.
(1) T always contains 0 in its spectrum.
(2) If λ ∈ Sp T \ 0, then λ is an eigenvalue with finite multiplicity.
(3) Sp T is either finite or consists of a sequence of complex numbers converg-
ing to 0.
In the sequel, for any vectors ξ and η in H, we denote by ξ ⊗ η ∗ the element of
L(H) defined by
(3.12) (ξ ⊗ η ∗ )ζ := hη, ξ 0 iξ ∀ζ ∈ H.
∗
Thus in ketbra notation ξ ⊗ η is just the operator |ξihη|. This is an operator of
rank 1. If ξ is a unit vector then ξ ⊗ ξ ∗ is the orthogonal projection onto Cξ.
Theorem 3.4.7 (Hilbert-Schmidt; see [RS, Thm. VI.16]). Let T ∈ K be nor-
mal. Then T diagonalizes in an orthonormal basis, i.e., there exists an orthonormal
basis (ξn )n≥0 of H and a sequence (λn )n≥0 ⊂ C such that
(3.13) T λn = λn ξn ∀n ∈ N0 .
This result allows us to reinterpret the Borel functional calculus for normal
compact operators as follows.
Proposition 3.4.8. Let T ∈ K be normal and let (ξn )n≥0 be an orthonormal
basis of H respect to which T is diagonal, i.e., T λn = λn ξn for all n ∈ N0 . In
addition, let f be bounded function on Sp T = {λn ; n ∈ N0 }.
(1) We have
X
(3.14) f (T ) = f (λn )(ξn ⊗ ξn∗ ),
n≥0

where the series converges strongly.

(2) f (T ) is a compact operator if and only if
lim f (λn ) = 0.
n→∞
Furthermore, in this case the series (3.14) converges in norm.
10
 Proof. For any n ∈ N0 we have T ξn = λn ξn . Since (ξn )n≥0 is an orthonormal
basis wePdeduce that T ∗ ξn = λn ξn for all n ∈ N0 . Let n ∈ N0 . For any polynomial
f (z) = ajk z j z k , we have
X X
f (T ) = ajk T j (T ∗ )k ξn = ajk λjn (λn )k ξn = f (λn )ξn .
It then follows from (1.14) that, for any f ∈ C(Sp T ), we have
(3.15) f (T )ξn = λn ξn ∀n ∈ N0 .
Since Sp T = {λn ; n ∈ N0 } any function on Sp T is a Borel function. Moreover,
by Theorem 3.4.6 0 is always in Sp T and the non-zero eigenvalues of T have finite
multiplicities, i.e., each non-zero eigenvalue λ appears at most finitely many times
in the sequence (λn )n≥0 . Thus λn → 0 as n → ∞, and hence a function f (λ) on
Sp T is continuous iff limn→∞ f (λn ) = f (0). Let f be a bounded function on Sp T
and, for N ∈ N, let fN be the function on Sp T defined by

f (λn ) if λ = λn with n < N ,
fN (λ) =
f (0) if λ = 0 or λ 6= λn for all n < N .
As limn→∞ fN (λn ) = f (0) = fN (0), we see that fN is a continuous function on
Sp T , and hence it satisfies (3.15). Moreover, for all n ∈ N0 ,
|fN (λn )| ≤ kf k∞ and lim fN (λn ) = f (λn ).
N →∞

Thus (fN )N ≥1 is a bounded sequence in L∞ (Sp T ) which converges pointwise to f .

It then follows from Theorem 3.2.2 and (3.15) that, for all n ∈ N0 ,
(3.16) f (T )ξn = lim fN (T )ξn = lim fN (λn )ξn = f (λn )ξn .
N →∞ N →∞
P
Let ξ ∈ H. Since ξ = n≥0 hξn , ξiξn and f (T ) is continuous, using (3.16) we
get
X X X
f (T )ξ = hξn , ξif (ξn ) = hξn , ξif (λn )ξn = f (λn )(ξn ⊗ ξn∗ )ξ,
n≥0 n≥0 n≥0

which proves (3.14).

Next, for N ∈ N let EN the span of ξ0 , . . . , ξN1 and denote by ΠN be the
orthogonal projection onto EN . Then by Proposition 3.4.2 (iv) the operator f (T )
is compact if and only if
(3.17) kf (T )|EN⊥ k = kf (T )(1 − ΠN )k −→ 0 as N −→ 0.
Set νN = supn≥N |f (λn )|. Since f (T )ξn = f (λn )ξn for all n ∈ N0 , we see that
νN ≤ kf (T )|EN⊥ k = kf (T )(1 − ΠN )k. Conversely, let ξ ∈ H. Since (ξn )n≥0 is an
orthonormal basis, from (3.14) we get
X X
kf (T )(1 − ΠN )ξk2 = |hξn , ξi|2 |f (λn )|2 ≤ νN
2
|hξn , ξi|2 = νN
2
kξk2 .
n≥N n≥0

Thus kf (T )(1 − ΠN )k ≤ νN . Thefore, we have

(3.18) kf (T )|EN⊥ k = kf (T )(1 − ΠN )k = sup |f (λn )|.
n≥N

Combining this with the condition (3.17) then shows that f (T ) is compact if and
only if limn→∞ f (λn ) = 0.
11
 Finally, using (3.14) we get
X
f (T )(1 − ΠN ) = f (λn )(ξn ⊗ ξn∗ ),
n≥N

where the series converges strongly. Thus,

X
f (λn )(ξn ⊗ ξn∗ ) = kf (T )(1 − ΠN )k = sup |f (λn )|.
n≥N
n≥N

Therefore, if limn→∞ f (λn ) = 0, then the series (3.14) converges to f (T ) in norm.

The proof is complete. 
Example 3.4.9. Let T ∈ K be normal. Then, with the notation of Proposi-
tion 3.4.8, we have
X
(3.19) T = λn (ξn ⊗ ξn∗ ),
λn 6=0

where the series converges in norm. Moreover, if we let T = U |T | be the polar

decomposition, then
X X
U= |λn |−1 λn (ξn ⊗ ξn∗ ) and |T | = |λn |(ξn ⊗ ξn∗ ),
λn 6=0 n≥0

where the first series converges strongly (unless T has finite rank, in which case it
is a finite sum), and the second series converges in norm.

Bibliography
[Fo] Folland, G.: Real analysis. 2nd edition. John Wiley & Sons, Inc., New York,
1999.
[RS] Reed, M.; Simon, B.: Methods of modern mathematical physics. I. Functional
analysis. Second edition. Academic Press, Inc., New York, 1980.

12