A brief introduction to trace class operators

Christa Hawthorne
December 2015

Contents
1 Introduction 1

2 Preliminaries 1

3 Trace class operators 2

4 Duals 8

1 Introduction
The trace is a useful number we can attach to an operator on a finite-dimensional space; a natural question
when studying operators on an arbitrary Hilbert space is whether we can formulate a meaningful generalization
of the trace. The trace class operators provide an appropriate setting for generalizing the finite-dimensional
trace. They are of further interest as the dual of the compact operators and a predual of the bounded
operators.
In this paper, I introduce the trace class operators and verify the above claim about their dual and
predual. This paper is largely based on the exposition in [2]; I have, however, avoided defining and using
Hilbert-Schmidt operators, and thus deviated somewhat from my primary source. Where possible, I adapted
proofs involving Hilbert-Schmidt operators to work solely with the trace class operators; elsewhere, I found
unrelated proofs.
Section 2 covers some prerequisite definitions and results. Section 3 defines the trace class operators,
covers their elementary properties, and relates them back to the original goal of finding an appropriate
generalization of the finite-dimensional trace. Section 4 proves that the trace class operators are the dual of
the compact operators and a predual of the bounded operators.

2 Preliminaries
In this section I present some standard definitions and results that we will make use of.
Unless otherwise stated, H refers to an arbitrary Hilbert space over F ∈ { C, R }.
The following theorem is usually proven using the functional calculus or spectral theorem, as in [1]. A
proof by approximations can be found in [3, Section 104].
Theorem 2.1. Suppose A ∈ B(H) with A ≥ 0. Then there is a unique B ∈ B(H) with B ≥ 0 such that
B 2 = A.
Definition 2.2. Suppose A ∈ B(H). We define |A| to be the unique B ∈ B(H) with B ≥ 0 such that
B 2 = A∗ A.

Remark 2.3. |A| is the unique positive B ∈ B(H) such that |A|g = ∥Ag∥ for all g ∈ H.

1
 We would like an analogue of the polar decomposition of a matrix. Unfortunately, outside of the finite-
dimensional context, there does not always exist a true isometry U such that A = U |A|. (Consider, for
example, the backwards shift A ∈ B(ℓ2 ) given by A(x1 , x2 , x3 , . . . ) = (x2 , x3 , . . . ).) All is not lost, however;
we simply need to relax our notion of isometry to the following:
Definition 2.4. We say W ∈ B(H) is a partial isometry if W ↾ ker(W )⊥ is an isometry.
Fact 2.5. If W ∈ B(H) is a partial isometry, then so is W ∗ .
We then get the following generalization of the polar decomposition of a matrix:
Theorem 2.6 (Polar decomposition). Suppose A ∈ B(H). Then there is a partial isometry W ∈ B(H) with
ker(W ) = ker(|A|) and A = W |A|.
The proof is similar to the finite-dimensional case.
Remark 2.7. We then have that ker(W )⊥ = ker(|A|)⊥ = Ran(|A|∗ ) = Ran(|A|) (since A ≥ 0), and W ↾
Ran(A) is an isometry. In particular, we get W ∗ A = W ∗ W |A| = |A|.
TODO 1. Missing |A| in Ran
Finally, a notational convention:
Definition 2.8. For g, h ∈ H, we define g ⊗ h∗ ∈ B00 (H) be (g ⊗ h∗ )v = ⟨v, h⟩g.
Remark 2.9. ∥g ⊗ h∗ ∥ = ∥g∥∥h∥.
It is a straightforward exercise in linear algebra to show that B00 (H) is spanned by operators of the form
g ⊗ h∗ .

3 Trace class operators

In this section, I introduce the trace class operators and their properties. [2, Section 18] introduces the
Hilbert-Schmidt operators early in the section, so the order and content of the proofs for the most part differs
pretty strongly from that of [2].
Recall that in finite dimensions, we can recover the trace of a matrix A as
n
X
⟨Aei , ei ⟩
i=1

where { e1 , . . . , en } is an orthonormal basis. We might hope to extend this to an arbitrary Hilbert space H
by setting X
tr(A) = ⟨Ae, e⟩
e∈E

for A ∈ B(H) (where E is some orthonormal basis for H). We run into two problems: the sum may diverge,
or it may differ depending on the choice of E.
Example 3.1. If H is infinite-dimensional and I ∈ B(H) is the identity map, then for any orthonormal basis E
of H we have X X
⟨Ie, e⟩ = 1=∞
e∈E e∈E

Example 3.2. Let H = ℓ2R . Consider A ∈ B(H) given by

(
e2n+1 i = 2n
Aei =
e2n i = 2n + 1

Then ⟨Aen , en ⟩ = 0 for all n ∈ N; so

∞
X
⟨Ae, e⟩ = 0
n=0

2
 1
converges absolutely. But F0 = { 2− 2 (e2n + e2n+1 ) : n ∈ N } is an orthonormal set, and Af = f for all
f ∈ F0 ; so
∞
1 1
X X
⟨Af, f ⟩ = ⟨A(2− 2 (e2n + e2n+1 )), 2− 2 (e2n + e2n+1 )⟩
f ∈F0 n=0
∞
1
X
= 2− 2 ∥e2n + e2n+1 ∥2
n=0
X∞
= 1
n=0
=∞

So X
⟨Af, f ⟩
f ∈F

does not converge absolutely for any orthonormal basis F extending F0 . The Riemann rearrangement theorem
then yields that, by reordering the basis, we can make the sum converge to any real number (or diverge).
If we want our notion of trace to make sense, we need to restrict our attention to a class of operators for
which we will be able to prove the trace is finite and independent of the chosen orthonormal basis.
Definition 3.3. T ∈ B(H) is trace class if
X
⟨|T |e, e⟩ < ∞
e∈E

for some orthonormal basis E of H. We write B1 (H) for the set of trace class operators on H.
The following theorem follows quickly from Parseval’s identity; see [2, Corollary 18.2].
Theorem 3.4. Suppose T ∈ B1 . Then
X X
⟨|T |e, e⟩ = ⟨|T |f, f ⟩ < ∞
e∈E f ∈F

for all orthonormal bases E and F of H.

This justifies the following definition:
Definition 3.5. Suppose T ∈ B1 (H). We define
X
∥T ∥1 = ⟨|T |e, e⟩
e∈E

where E is some orthonormal basis for H.

For now, this is just a notational convenience; however, we will see in Theorem 3.15 that (B1 (H), ∥·∥1 ) is
a normed linear space.
The following theorem provides some justification for my implicit assertion that B1 (H) is the class of
operators we wish to work in if we want a sensible notion of trace.
Theorem 3.6. Suppose T ∈ B1 (H). Then X
⟨T e, e⟩
e∈E

is finite and independent of the choice of orthonormal basis E.

We need two more results before we can prove this.
I am indebted to [4, Lemma 6.3.1] for an idea in proving the following lemma and corollary.

3
 1
Lemma 3.7. Suppose T ∈ B1 (H). Then |T | 2 ∈ K(H).
1
Proof. Suppose T ∈ B1 (H). We show that |T | 2 can be approximated in ∥·∥ by finite rank operators.
Suppose ε > 0.

1
Claim 3.8. There do not exist infinitely many linearly independent g ∈ H such that |T | 2 g ≥ ε.

1
Proof. Suppose otherwise; then we may let E0 be an infinite orthonormal set such that |T | 2 e ≥ ε for all
e ∈ E0 . Let E be an orthonormal basis extending E0 . Then
X
∥T ∥1 = ⟨|T |e, e⟩
e∈E
X
≥ ⟨|T |e, e⟩
e∈E0
12 2
X
= |T | e
e∈E0
X
≥ ε2
e∈E0
=∞

contradicting our assumption that T ∈ B1 (H). Claim 3.8

1
We may then choose a finite orthonormal E0 such that if |T | 2 g ≥ ε, then g ∈ span E0 . Extend E0 to an
orthonormal basis E of H. Define F ∈ B00 (H) by
( 1
|T | 2 e e ∈ E0
Fe =
0 e ∈ E \ E0

Then if g + h ∈ (span E0 ) + (span E0 )⊥ = H, then

 1   1   1  1
|T | 2 − F (g + h) = |T | 2 − F g + |T | 2 − F h = |T | 2 h < ε

1 1
So |T | 2 can be approximated in ∥·∥ by finite rank operators. So |T | 2 ∈ K(H). Lemma 3.7
Corollary 3.9. B1 (H) ⊆ K(H).
1 1
Proof. Suppose T ∈ B1 (H). Let T = W |T | be the polar decomposition. Then T = W |T | 2 |T | 2 . But
1
|T | 2 ∈ K(H), and K(H) is a two-sided ideal. So T ∈ K(H). Corollary 3.9
We are now ready to prove Theorem 3.6:
Proof of Theorem 3.6. Suppose T ∈ B1 (H). Then by Corollary 3.9 we have that |T | is compact. So there is
an orthonormal basis E diagonalizing |T |; i.e. there are se ≥ 0 such that |T |e = se e for all e ∈ E. Suppose F
is any orthonormal basis. Let T = W |T | be the polar decomposition of T . Then
X X X
|se ⟨W e, e⟩| ≤ |se |∥W e∥∥e∥ ≤ |se | = ∥T ∥1 < ∞
e∈E e∈E e∈E

so X
se ⟨W e, e⟩
e∈E

4
converges absolutely, and we need not worry about rearranging terms. Then
* +
X X X
se ⟨W e, e⟩ = se W e, ⟨e, f ⟩f
e∈E e∈E f ∈F
XX
= ⟨f, e⟩⟨W |T |e, f ⟩
e∈E f ∈F
XX
= ⟨W |T |⟨f, e⟩e, f ⟩
f ∈F e∈E
* +
X X
= W |T | ⟨f, e⟩e, f
f ∈F e∈E
X
= ⟨W |T |f, f ⟩
f ∈F
X
= ⟨T f, f ⟩
f ∈F

Theorem 3.6
This allows us to make the following definition:
Definition 3.10. Suppose T ∈ B1 (H). We define the trace of T to be
X
tr(T ) = ⟨T e, e⟩
e∈E

for some orthonormal basis E.

I devote the remainder of this section to demonstrating properties of B1 (H) that we will use in Section 4.
Remark 3.11. B00 (H) ⊆ B1 (H).
Theorem 3.12. B00 (H) is dense in B1 (H) with respect to ∥·∥1 .
Proof. Suppose T ∈ B1 (H). Then by Corollary 3.9 we have that |T | ∈ K(H). But |T | ≥ 0. So, by the spectral
theorem for compact normal operators, we have some orthonormal basis E of H that diagonalizes |T |; i.e.
there exist se ≥ 0 such that for all e ∈ E we have |T |e = se e.
Suppose ε > 0. Then since T ∈ B1 (H), there is some finite E0 ⊆ E such that
X X
se = ⟨|T |e, e⟩ > ∥T ∥ − ε
e∈E0 e∈E0

and in particular X
se < ε
e∈E\E0

Define F ∈ B00 (H) by (

|T |e e ∈ E0
Fe =
0 e ∈ E \ E0
Then X
|T | − F = se e ⊗ e∗ ≥ 0
e∈E\E0

So  
|T | − F  = |T | − F
 

and XD  E X X X
|T | − F = |T | − F e, e = ⟨(|T | − F )e, e⟩ = ⟨|T |e, e⟩ = se < ε
 
e∈E e∈E e∈E\E0 e∈E\E0

Theorem 3.12

5
Proposition 3.13. Suppose T ∈ B1 (H). Then ∥T ∗ ∥1 = ∥T ∥1 ; in particular, T ∗ ∈ B1 (H).
Proof. Let T = W |T | be the polar decomposition of T . Then
(T ∗ )∗ T ∗ = T T ∗ = W |T ||T |W ∗ = W |T |W ∗ W |T |W ∗ = (W |T |W ∗ )2
 1
∗  1 
Also W |T |W ∗ = |T | 2 W ∗ |T | 2 W ∗ ≥ 0. So, by uniqueness of square roots, we have that |T ∗ | = W |T |W ∗ .
Now, let E0 be an orthonormal basis for ker(W ∗ ); let E be an orthonormal basis for H extending E0 . Let
F0 = { W ∗ e : e ∈ E \ E0 }. Then Fact 2.5 yields that W ∗ is a partial isometry, and thus F0 is orthonormal;
let F be an orthonormal basis for H extending F0 . Then
X
∥T ∗ ∥1 = ⟨|T ∗ |e, e⟩
e∈E
X
= ⟨W |T |W ∗ e, e⟩
e∈E
X
= ⟨|T |W ∗ e, W ∗ e⟩
e∈E
X
= ⟨|T |f, f ⟩
f ∈F0
X
= ⟨|T |f, f ⟩
f ∈F
= ∥T ∥1
∗ ⊥
(since f ∈ F \ F0 satisfies f ∈ Ran(W ) = ker(W ) = ker(|T |)). Proposition 3.13
Lemma 3.14. Suppose A ∈ B(H) and T ∈ B1 (H). Suppose E is an orthonormal basis for H diagonalizing
|T |; that is, suppose there is se ≥ 0 such that |T |e = se e for all e ∈ E. Then
X
|⟨AT e, e⟩| ≤ ∥A∥∥T ∥1
e∈E

Proof. Let T = W |T | be the polar decomposition of T . Then

X X 
|⟨AT e, e⟩| = ⟨AW ∗ |T |e, e⟩

e∈E e∈E
X
= se |⟨AW ∗ e, e⟩|
e∈E
X
≤ se ∥AW ∗ e∥∥e∥
e∈E
X
≤ se ∥A∥
e∈E
X
= ∥A∥ se
e∈E
X
= ∥A∥ ⟨|T |e, e⟩
e∈E
= ∥A∥∥T ∥1
Lemma 3.14
Theorem 3.15. B1 (H) is a subspace of B(H), and ∥·∥1 is a norm on B1 (H).
(In fact, (B1 (H), ∥·∥) is a Banach space; however, we won’t see this until Corollary 4.6.)
A proof of the above theorem can be found in [2, Theorem 18.11]. (Though on the surface Conway
appears to rely on properties of Hilbert-Schmidt operators, the only non-trivial use is to show that trace
class operators are compact, which we verified in Corollary 3.9.) The only part that is not straightforward is
verifying the triangle inequality; this follows from a somewhat tedious manipulation of sums.
The following remark then follows easily:

6
Remark 3.16. tr : B1 (H) → F is linear.
Theorem 3.17. Suppose A ∈ B(H) and T ∈ B1 (H). Then ∥AT ∥1 ≤ ∥A∥∥T ∥1 and ∥T A∥1 ≤ ∥A∥∥T ∥1 ; in
particular, B1 (H) is a two-sided ideal of B(H).
Proof. Suppose A ∈ B(H) and T ∈ B1 (H).
Claim 3.18. ∥AT ∥1 ≤ ∥A∥∥T ∥1 .
Proof. Let T = W |T | be the polar decomposition of T ; let AT = V |T | be the polar decomposition of AT .
Then
TODO 2. really?
Corollary 3.9 yields that T ∈ K(H); thus |T | = W ∗ T ∈ K(H). But |T | ≥ 0; so, by the spectral theorem
for compact operators, we have an orthonormal basis E of H that diagonalizes |T |. But then Lemma 3.14
yields
X
∥AT ∥1 = ⟨|AT |e, e⟩
e∈E
X
= |⟨|AT |e, e⟩|
e∈E
X
= |⟨V ∗ AT e, e⟩|
e∈E
≤ ∥V ∗ A∥∥T ∥1
≤ ∥A∥∥T ∥1
Claim 3.18
Claim 3.19. ∥T A∥1 ≤ ∥A∥∥T ∥1 .
Proof. Proposition 3.13 and Claim 3.18 yield
∥T A∥1 = ∥(A∗ T ∗ )∗ ∥1 = ∥A∗ T ∗ ∥1 ≤ ∥A∗ ∥∥T ∗ ∥1 = ∥A∥∥T ∥1
Claim 3.19
Theorem 3.17
Theorem 3.20. Suppose T ∈ B1 (H) and A ∈ B(H). Then |tr(AT )| ≤ ∥T ∥1 ∥A∥.
Proof. Corollary 3.9 yields that there is an orthonormal basis E for H diagonalizing |T |; Lemma 3.14 then
yields that
 
X 
|tr(AT )| =  ⟨AT e, e⟩
 
 
e∈E
X
≤ |⟨AT e, e⟩|
e∈E
≤ ∥A∥∥T ∥1
Theorem 3.20
Remark 3.21. It is easy to see that if T ∈ B1 (H), then tr(T ) = tr(T ∗ ).
Corollary 3.22. Suppose T ∈ B1 (H) and A ∈ B(H). Then |tr(T A)| ≤ ∥T ∥1 ∥A∥.
Proof. Note that
|tr(T A)| = |tr((A∗ T ∗ )∗ )| = |tr(A∗ T ∗ )| ≤ ∥A∗ ∥∥T ∗ ∥1 = ∥A∥∥T ∥1
(using Theorem 3.20 and Proposition 3.13). Corollary 3.22
I end this section by noting a neat generalization of one of the properties of the trace in finite dimensions
that we will neither prove nor make use of:
Theorem 3.23. Suppose T ∈ B1 (H) and A ∈ B(H). Then tr(AT ) = tr(T A).

7
4 Duals
Besides being of independent interest, trace class operators arise naturally as the dual of the compact
operators; a corollary of this will be that (B1 (H), ∥·∥1 ) is a Banach space. In this section, I demonstrate this;
I further show that the bounded operators are the dual of the trace class operators.
The proofs of the theorems are almost entirely due to [2, Section 19]; the only modification beyond style
and presentation that I made was to explicitly prove injectivity in Theorem 4.2.
The following easy technical lemma will prove useful.
Lemma 4.1. Suppose A ∈ B(H), g, h ∈ H. Then tr(A(g ⊗ h∗ )) = ⟨Ag, h⟩.

Proof. Recall that g ⊗ h∗ ∈ B00 (H) ⊆ B1 (H), so the trace is well-defined.

Case 1. Suppose h = 0. Then tr(A(g ⊗ h∗ )) = 0 = ⟨Ag, h⟩.
h
Case 2. Suppose h ̸= 0. Let E be an orthonormal basis extending ∥h∥ . Then
X
tr(A(g ⊗ h∗ )) = ⟨A(g ⊗ h∗ )e, e⟩
e∈E
X
= ⟨e, h⟩⟨Ag, e⟩
e∈E
  
h h
= , h Ag,
∥h∥ ∥h∥
 
h
= ∥h∥ Ag,
∥h∥
= ⟨Ag, h⟩

Lemma 4.1

We now proceed to our main theorems.

Theorem 4.2. The dual of K(H) is (B1 (H), ∥·∥1 ).
Proof. Let Ψ : B1 (H) → K(H)∗ be ΨT (K) = tr(T K). I claim that Ψ is an isometric isomorphism.
Claim 4.3. ΨT is linear and Ψ is linear.
Proof. To see that ΨT is linear, note that

ΨT (aK + L) = tr(T (aK + L)) = tr(aT K + T L) = a tr(T K) + tr(T L) = aΨT (K) + ΨT (L)

by Remark 3.16. To see that Ψ is linear, note that

ΨaS+T (K) = tr((aS + T )K) = tr(aSK + T K) = a tr(SK) + tr(T K) = aΨS (K) + ΨT (K)

again by Remark 3.16. Claim 4.3

Claim 4.4. ∥ΨT ∥ ≤ ∥T ∥1 for all T ∈ B1 (H).
Proof. This is just Corollary 3.22. Claim 4.4
So Ψ is indeed a linear map B1 (H) → K(H)∗ .
Claim 4.5. Ψ is bijective and ∥ΨT ∥ ≥ ∥T ∥1 .

8
Proof. Suppose α ∈ K(H)∗ . We define a sesquilinear form on H by [g, h] = α(g ⊗ h∗ ); it is easily seen that
this is sesquilinear and that |[g, h]| ≤ ∥α∥∥g∥∥h∥ for all g, h ∈ H. Thus there is a unique S ∈ H such that
[g, h] = ⟨Sg, h⟩ = tr(S(g ⊗ h∗ )) for all g, h ∈ H (where the last equality follows from Lemma 4.1).
Note that the uniqueness immediately implies that Ψ is injective, since if T1 , T2 ∈ B1 (H) satisfy ΨT1 =
ΨT2 = α, then

tr(T1 (g ⊗ h∗ )) = ΨR (g ⊗ h∗ )
= α(g ⊗ h∗ )
= [g, h]

and likewise we have tr(T2 (g ⊗ h∗ )) = [g, h]; uniqueness then yields that T1 = T2 .
Now, let S = W |S| be the polar decomposition of S. Suppose E is an orthonormal basis for H; suppose
E0 ⊆ E is finite. Then
 
X X 
⟨|S|e, e⟩ =  ⟨W ∗ Se, e⟩
 
 
e∈E0 e∈E
 0 
X 
= ⟨Se, W e⟩
 
 
e∈E0 
X 
= [e, W e]
 
 
e∈E
 0 
X 
∗ 
= α(e ⊗ (W e) )

 
e∈E0 !
 X 
∗ 
= α e ⊗ (W e) 

 
e∈E
0
X
≤ ∥α∥ e ⊗ e∗ ∥W ∥


e∈E0
≤ ∥α∥

Thus X
∥S∥1 = ⟨|S|e, e⟩ ≤ ∥α∥ < ∞
e∈E

and S ∈ B1 (H). Furthermore, as remarked above, we have ΨS (g ⊗ h∗ ) = [g, h] = α(g ⊗ h∗ ) for all g, h ∈ H.
But B00 (H) is spanned by elements of the form g ⊗ h∗ , and B00 (H) is ∥·∥-dense in K(H); furthermore, ΨS
and α are linear functionals that are ∥·∥-continuous and agree on elements of the form g ⊗ h∗ . So ΨS and α
agree on all of K(H), and ΨS = α.
So Ψ is surjective. Furthermore, by injectivity, we have S is the unique preimage of α; this, combined
with our earlier note that ∥S∥1 ≤ ∥α∥, shows that ∥ΨT ∥ ≥ ∥T ∥1 for all T ∈ B1 (H). (Note that we need the
uniqueness of the preimage to conclude this for all T ∈ B1 (H); otherwise, all we have is that ΦS has some
preimage R with ∥ΨR ∥ ≥ ∥R∥.) Claim 4.5
Putting the claims together, we see that Ψ is a bijective linear isometry. So Ψ is an isometric isomorphism.
Theorem 4.2
Corollary 4.6. (B1 , ∥·∥1 ) is a Banach space.

Theorem 4.7. The dual of (B1 (H), ∥·∥1 ) is B(H).

Proof. Let Φ : B(H) → B1 (H)∗ be ΦA (T ) = tr(AT ). I claim that Φ is an isometric isomorphism.
That Φ is a linear map B(H) → B1 (H)∗ follows exactly as in Theorem 4.2.
Claim 4.8. ∥ΦA ∥ ≤ ∥A∥.

9
Proof. This is just Theorem 3.20. Claim 4.8
Claim 4.9. ∥ΦA ∥ ≥ ∥A∥.
Proof. Suppose ε > 0; we will show that ∥ΦA ∥ > ∥A∥−ε. Pick non-zero g ∈ H such that ∥Ag∥ > (∥A∥−ε)∥g∥.
Then Lemma 4.1 yields

ΦA (g ⊗ (Ag)∗ ) = tr(A(g ⊗ (Ag)∗ ))

= ∥Ag∥2
> (∥A∥ − ε)∥g∥∥Ag∥
= (∥A∥ − ε)∥g ⊗ (Ag)∗ ∥

So ∥ΦA ∥ > ∥A∥ − ε for all ε > 0. So ∥ΦA ∥ ≥ ∥A∥. Claim 4.9
Claim 4.10. Φ is surjective.
Proof. This follows similarly to the proof of Theorem 4.2, albeit more simply. Suppose α ∈ B1 (H)∗ . Consider
the sesquilinear form on H given by [g, h] = α(g ⊗ h∗ ). We again have that this is a bounded, sesquilinear
form; we again get A ∈ B(H) such that [g, h] = ⟨Ag, h⟩ = tr(A(g ⊗ h∗ )) = ΦA (g ⊗ h∗ ). We again have that
B00 (H) is spanned by elements of the form g ⊗ h∗ ; Theorem 3.12 gives us that B00 (H) is ∥·∥1 -dense in B1 (H).
Furthermore, we have that ΦA and α are linear and ∥·∥1 -continuous, and they agree on elements of the form
g ⊗ h∗ ; so ΦA and α agree on all of B1 (H), and ΦA = α. So Φ is surjective. Claim 4.10

So Φ is a bijective linear isometry. So Φ is an isometric isomorphism. Theorem 4.7

References
[1] John B. Conway. A Course in Functional Analysis. Springer-Verlag, 1990 (cit. on p. 1).
[2] John B. Conway. A Course in Operator Theory. American Mathematical Society, 2000 (cit. on pp. 1–3,
6, 8).
[3] Frigyes Riesz and Béla Szőkefalvi-Nagy. Functional Analysis. Courier Corporation, 1990 (cit. on p. 1).
[4] Nik Weaver. Mathematical Quantization. CRC Press, 2001 (cit. on p. 3).

10