arXiv:2307.03064v1 [math.

FA] 6 Jul 2023

A journey into Matrix Analysis

 2  2
1 4 5 1  2
2 6 7 = U 2 U ∗ + V 4 5 V ∗ + W 6 7 W ∗
2
 
8 9
3 8 9 3

Jean-Christophe Bourin
2
Outline of the thesis

Matrix Analysis is essential in many areas of mathematics and sciences. This is the
main topic of this thesis which presents a substantial part of my research. There are 9
chapters.

Chapter 1 is an introductory chapter, some results from the period 1999-2010 are given.

Chapters 2–8 are the central part of the thesis. Each chapter presents an article (with a
blue title). This article is complemented with an additional section, Around this article.
We may divide these chapters into three groups.

Chapters 2-4 deal with matrix inequalities, Chapter 2 is concerned with norm
inequalities and logmajorization and Chapters 3-4 with functional calculus and a
unitary orbit technique that I started to develop in 2003.

Chapter 5 is a time-break in infinite dimensional Hilbert space operators

Chapters 6-8 establish several decompositions for partitioned matrices, especially

for positive block matrices. Some norm inequalities involving the numerical range
are derived.

Chapter 9 is for students; a proof of the Spectral Theorem for bounded operators is
derived from the matrix case.

The thesis is divided into two parts of similar size. The first part establishes matrix
inequalities involving symmetric norms, eigenvalues and unitary orbits. These results are
also used in the second part, dealing with operator diagonals of Hilbert space operators,
partitioned matrices, and numerical ranges.

3
4
Contents

I Matrix inequalities, norms and unitary orbits 9

1 Some results as an introduction 13
1.1 Rearrangement inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Subadditivity and superadditivity . . . . . . . . . . . . . . . . . . . . . 15
1.3 Diagonal blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Missing topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 References of Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Majorization and Perspective 19

2.1 Log-majorization and log-convexity . . . . . . . . . . . . . . . . . . . . . 19
2.2 Araki type inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 With positive operators . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 With normal operators . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Hölder type inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Proof of Theorem 2.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Hilbert space operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7 References of Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3 Unitary Orbits and Functions 39

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 A matrix Jensen type inequality . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Jensen type inequalities via unitary orbits . . . . . . . . . . . . . 41
3.2.2 Comments and references . . . . . . . . . . . . . . . . . . . . . . 49
3.3 A matrix subadditivity inequality . . . . . . . . . . . . . . . . . . . . . . 53
3.3.1 Sub/super-additivity inequalities via unitary orbits . . . . . . . . 53
3.3.2 Comments and references . . . . . . . . . . . . . . . . . . . . . . 57
3.4 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 References of Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4 Around Hermite-Hadamard 63
4.1 Elementary scalar inequalities . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 The extremal property for matrices . . . . . . . . . . . . . . . . . . . . . 65
4.3 Majorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 References of chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5
6 CONTENTS

II Operator diagonals and partitionned matrices 73

5 Pinchings and Masas 77
5.1 The pinching theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Sums in a unitary orbit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3 Conditional expectation onto a masa . . . . . . . . . . . . . . . . . . . . 82
5.3.1 Conditional expectation of general operators . . . . . . . . . . . . 82
5.3.2 A reduction lemma . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3.3 Conditional expectation of idempotent operators . . . . . . . . . . 85
5.4 Unital, trace preserving positive linear maps . . . . . . . . . . . . . . . . 86
5.5 Pinchings in factors ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.6 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.6.1 Essential numerical range . . . . . . . . . . . . . . . . . . . . . . 89
5.6.2 Proof of the pinching theorem . . . . . . . . . . . . . . . . . . . . 90
5.6.3 Müller-Tomilov’s theorem . . . . . . . . . . . . . . . . . . . . . . 93
5.7 References of Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

6 Partial trace 95
6.1 Introduction and a key lemma . . . . . . . . . . . . . . . . . . . . . . . . 95
6.2 Direct sum and partial trace . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.2.1 Around Hiroshima’s theorem . . . . . . . . . . . . . . . . . . . . . 97
6.2.2 Around Rotfel’d inequality . . . . . . . . . . . . . . . . . . . . . . 100
6.3 Proof of Theorem 2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.4 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.4.1 Complex matrices and small partitions . . . . . . . . . . . . . . . 103
6.4.2 Separability criterion . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.5 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.6 References of Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

7 Block matrices and numerical range 109

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.2 Eigenvalue inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3 Norm inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.4 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.5 References of Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

8 Block matrices and Pythagoras 119

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.2 A Pythagorean theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.3 Compression onto a hyperplane . . . . . . . . . . . . . . . . . . . . . . . 125
8.4 Four and five blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.5 Concave or convex functions . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.6 Around this article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.7 References of Chapter 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
CONTENTS 7

9 The spectral Theorem 133

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9.2 The matrix case, and two natural definitions for operators . . . . . . . . 133
9.3 Continuous functional calculus . . . . . . . . . . . . . . . . . . . . . . . . 134
9.4 Spectral projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
9.5 The spectral theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8 CONTENTS
 Part I

Matrix inequalities, norms and

unitary orbits

9
 11

f ′′ (t) ≤ 0, f (0) ≥ 0, A ≥ 0, Z ∗ Z ≥ I

=⇒

Tr f (Z ∗AZ) ≤ Tr Z ∗f (A)Z
12
Chapter 1

Some results as an introduction

This chapter gives some theorems that are not discussed in the thesis. We will only
mention their relation with the other chapters. This chapter covers a significant part of
my work in the period 1999-2010.

1.1 Rearrangement inequalities

In an old paper [18] (see [25] for a simplified proof) I obtained the following inequality
for the Hilbert-Schmidt norm k · k2 .
Denote by Mn the space of n-by-n matrices and by M+ n the positive semidefinite
+
cone. A pair (A, B) in Mn is called a monotone pair if A = f (C) and B = g(C) for
two non-decreasing functions f (t), g(t) and some C ∈ M+ n . If f (t) is non-decreasing and
g(t) is non-increasing, then we say that (A, B) is antimonotone.

Theorem 1.1.1. Let Z ∈ Mn be a normal matrix. If (A, B) is a monotone pair in M+

n
then,
kAZBk2 ≤ kZABk2 .
If (A, B) is antimonotone, then the inequality reverses.

The result still holds for Hilbert space operators whenever A is Hilbert-Schmidt.
Letting Z be a rank one projection, we recapture Chebyshev inequality for pairs of
non-decreasing functions on (0, 1),
Z 1 Z 1 Z 1
f (t) dt g(t) dt ≤ f (t)g(t) dt.
0 0 0

Letting Z be a unitary matrix, we recapture von Neumann trace inequality (1937) for
general matrices X, Y ∈ Mn ,
n
X
|Tr XY | ≤ µj (X)µj (Y )
j=1

where µ1 (X) ≥ · · · ≥ µn (X) are the singular values of X.

13
14 CHAPTER 1. SOME RESULTS AS AN INTRODUCTION

Hence studying the product AZB is useful. Chapter 2 will be devoted to the study of
the functional (s, t) 7→ As ZB t for a general matrix Z and any pair of positive matrices
A, B. This will provide a number of new inequalities extending some famous inequalities.
Two consequences of Theorem 1.1.1 for the operator norm k · k∞ are given in [18]. If
(A, B) is a monotone pair in M+ n , then

kAEBk∞ ≤ kEABk∞ (1.1.1)

for all projections E, and √

kASBk∞ ≤ 2kSABk∞
for all semi-unitary matrices S. Here semi-unitary means that S ∗ S = SS ∗ = F for some
projection F . This suggests the following conjecture for any symmetric norm k · k (i.e.,
unitarily invariant norm).
Conjecture A. Let Z ∈ Mn be a normal matrix. If (A, B) is a monotone pair in M+
n,
then, for all symmetric norms,
√
kAZBk ≤ 2kZABk.

Conjecture A is supported by the remarkable result recently proved by Eric Ricard

(private communication) : Conjecture A holds true if we take the constant 8.
The norm inequality (1.1.1) can be considerably extended. In [19] I give the singular
value inequality:

Theorem 1.1.2. Let (A, B) be a monotone pair in M+

n and let E ∈ Mn be a projection.
Then for some unitary V ∈ Mn ,

|AEB| ≤ V |EAB| V ∗

From this result follow several eigenvalue inequalities for compressions onto subspaces,
and more generally for unital, positive linear map Φ. For instance :

Corollary 1.1.3. Let (A, B) be a monotone pair in M+ n and let Φ : Mn → Md be a

positive, unital linear map. Then for some unitary V ∈ Md ,

Φ(ABA) ≤ V Φ(A)Φ(B)Φ(A)V ∗ .

This was recorded in a joint paper with Ricard [48]. Concerning Rearrangement
inequalities, one may state another conjecture [26].
Conjecture B. Let A, B ∈ M+
n and p, q > 0. Then, for all symmetric norms,

kAp B q + B p Aq k ≤ kAp+q + B p+q k.

Several authors, including Audenaert, Bhatia, Kittaneh, proved some very special
cases of the conjecture, but the general case is still open.
1.2. SUBADDITIVITY AND SUPERADDITIVITY 15

1.2 Subadditivity and superadditivity

In 1969, Rotfel’d stated a remarkable trace inequality for all non-negative concave func-
tions f (t) defined on [0, ∞) and any pair A, B ∈ M+n,

Tr f (A + B) ≤ Tr f (A) + Tr f (B).
In a joint paper with Aujla [12] we give the stronger statement,
f (A + B) ≤ Uf (A)U ∗ + V f (B)V ∗ (1.2.1)
for some unitary matrices U, V ∈ Mn . This result and a number of related Jensen’s type
inequalities are discussed at length in Chapter 3.
The Rotfel’d trace inequality can also be extended as a norm inequality. In the paper
written with Uchiyama [49], we obtain the following theorem.
Theorem 1.2.1. If f (t) is a nonnegative concave function on [0, ∞) and A, B ∈ M+
n,
then, for all symmetric norms,
kf (A + B)k ≤ kf (A) + f (B)k.

This result was first proved for operator concave functions by Ando and Zhan (1999).
Our proof (2007) is completely different, and does not use the theory of operator mono-
tone functions. A part of our proof comes from [24] where the following theorem was
obtained.
Theorem 1.2.2. Let f (t) be a nonnegative concave function on [0, ∞), let A ∈ M+
n,
and let Z ∈ Mn be expansive. Then, for all symmetric norms,
kf (Z ∗ AZ)k ≤ kZ ∗ f (A)Zk.

Here Z expansive means that Z ∗ Z ≥ I. For the trace, the function f (t) is not
necessarily increasing, and we get the trace inequality [22] of Page 11. We may gather
Theorems 1.2.1 and 1.2.2 into a single statement [30] :
Theorem 1.2.3. In the space Mn , let {Ai }m m
i=1 be positive and let {Zi }i=1 be expansive.
Let f (t) be a non-negative concave function on [0, ∞). Then, for all symmetric norms,
m
! m
X X
∗
f Zi Ai Zi ≤ Zi∗ f (Ai )Zi .
i=1 i=1

A basic principle for symmetric norms shows that Theorem 1.2.3 entails a reversed
inequality in case of convex functions.
Corollary 1.2.4. In the space Mn , let {Ai }m m
i=1 be positive and let {Zi }i=1 be expansive.
Let g(t) be a non-negative convex function on [0, ∞) with g(0) = 0. Then, for all
symmetric norms, !
Xm Xm
∗ ∗
Zi g(Ai )Zi ≤ g Zi Ai Zi .
i=1 i=1
16 CHAPTER 1. SOME RESULTS AS AN INTRODUCTION

Another extension of Theorem 1.2.1 is given in [27].

Theorem 1.2.5. If f (t) is a nonnegative concave function on [0, ∞) and A, B ∈ Mn
are normal matrices, then, for all symmetric norms,
kf (|A + B|)k ≤ kf (|A|) + f (|B|)k.

A number of corollaries follow.

Corollary 1.2.6. If f (t) is a nonnegative concave function on [0, ∞) and Z ∈ Mn has
the Cartesian decomposition Z = A + iB, then, for all symmetric norms,
kf (|Z|)k ≤ kf (|A|) + f (|B|)k.

Corollary 1.2.7. If f (t) is a nonnegative concave function on [0, ∞) and Z ∈ Mn ,

then, for all symmetric norms,
kf (|Z + Z ∗ |)k ≤ kf (|Z|) + f (|Z ∗ |)k.

1.3 Diagonal blocks

Chapters 6-8 deal with partitioned matrices and their diagonal blocks. Chapter 5 con-
cerns operator diagonals of Hilbert space operators.
My earlier result on operator diagonals of block matrices [20] is :
Theorem 1.3.1. Let A ∈ M2n . Then, there exists some B ∈ Mn such that
 
B ⋆
A≃ .
⋆ B

Here, the stars hold for unspecified entries, and ≃ means unitarily congruence. The
simplest (and well-known) case is for A ∈ M2 , and it is the key for the proof of the
Hausdorff-Toeplitz theorem (1918) ensuring that the numerical range is convex. The
proof of Theorem 1.3.1 cannot cover the case of Hilbert space operators. However we
may propose a conjecture.
Conjecture C. Let A be an operator acting on a infinite dimensional separable Hilbert
space H. Then, for some subspace S ⊂ H,
AS ≃ AS ⊥ .

Conjecture D. There exists A ∈ M6 such that, for any B ∈ M2 , A is not unitarily

equivalent to a matrix of the form
 
B ⋆ ⋆
⋆ B ⋆.
⋆ ⋆ B

Chapters 6-8 consider positive partitioned matrices; the matrix inequalities and the
decomposition of Chapter 3 come into play.
1.4. MISSING TOPICS 17

1.4 Missing topics

To keep a reasonable lenght for the thesis, I do not give any results involving the matrix
geometric mean and I do not develop the topic of symmetric antinorms. In a paper1
devoted to the study of antinorms, a version of (1.2.1) for τ -measurable operators is
given (this is nontrivial when the algebra is not a factor). In a paper2 devoted to the
geodesics associated to the geometric mean, the following exotic Hölder inequality is
obtained,
m
! m
! 1/p m
! 1/q
X X X
sinh Ai Bi ≤ sinh Api sinh Biq ,
i=1 i=1 i=1

for all positives matrices Ai , Bi , such that Ai Bi = Bi Ai , (i = 1, . . . , m), conjugate

exponents p, q > 1 and symmetric norms k · k. A related result states that
m
!
X
(t1 , . . . , tm ) 7→ Tr log Xi∗ Atii Xi
i=1

is jointly convex on Rm , where Ai > 0 and Xi is invertible (i = 1, . . . , m). Hence, the ge-
ometric mean provides results without the geometric mean ! A well-known phenomenon
since the fundamental paper by Ando (1979).

1.5 References of Chapter 1

[18] J.-C. Bourin, Some inequalities for norms on matrices and operators, Linear Algebra
Appl. 292 (1999), no. 1–3, 139–154.

[19] J.-C. Bourin, Singular values of compressions, restrictions and dilations, Linear Algebra
Appl. 360 (2003), 259–272.

[20] J.-C. Bourin, Total dilations, Linear Algebra Appl. 368 (2003), 159–169.

[22] J.-C. Bourin, Convexity or concavity inequalities for Hermitian operators. Math. In-
equal. Appl. 7 (2004), no. 4, 607–620.

[24] J.-C. Bourin, A concavity inequality for symmetric norms, Linear Algebra Appl. 413
(2006), 212-217.

[25] J.-C. Bourin, Matrix versions of some classical inequalities. Linear Algebra Appl. 416
(2006), no. 2–3, 890–907.

[26] J.-C. Bourin, Matrix subadditivity inequalities and block-matrices, Internat. J. Math.
20 (2009), no. 6, 679–691.

[27] J.-C. Bourin, A matrix subadditivity inequality for symmetric norms, Proc. Amer.
Math. Soc. 138 (2010), no. 2, 495–504.
1
Bourin-Hiai 2015
2
Bourin-Shao 2020
18 CHAPTER 1. SOME RESULTS AS AN INTRODUCTION

[30] J.-C. Bourin and E.-Y. Lee, Concave functions of positive operators, sums, and con-
gruences, J. Operator Theory 63 (2010), 151–157.

[48] J.-C. Bourin and E. Ricard, An asymmetric Kadison’s inequality, Linear Algebra Appl.
433 (2010) 499–510.

[49] J.-C. Bourin and M. Uchiyama, A matrix subadditivity inequality for f (A + B) and
f (A) + f (B), Linear Algebra Appl. 423 (2007), 512–518.
Chapter 2

Majorization and Perspective

Matrix Inequalities
from a two variables functional [36]
Abstract. We introduce a two variables norm functional and establish its joint log-convexity.
This entails and improves many remarkable matrix inequalities, most of them related to the
log-majorization theorem of Araki. In particular: if A is a positive semidefinite matrix and
N is a normal matrix, p ≥ 1 and Φ is a sub-unital positive linear map, then |AΦ(N )A|p is
weakly log-majorized by Ap Φ(|N |p )Ap . This far extension of Araki’s theorem (when Φ is the
identity and N is positive) complements some recent results of Hiai and contains several special
interesting cases such as a triangle inequality for normal operators and some extensions of the
Golden-Thompson trace inequality. Some applications to Schur products are also obtained.
Keywords. Matrix inequalities, Majorization, Positive linear maps, Schur products.
2010 mathematics subject classification. 47A30, 15A60.

2.1 Log-majorization and log-convexity

Matrices are regarded as non-commutative extensions of scalars and functions. Since
matrices do not commute in general, most scalars identities cannot be brought to the
matrix setting, however they sometimes have a matrix version, which is not longer an
identity but an inequality. These kind of inequalities are of fundamental importance in
our understanding of the noncommutative world of matrices. A famous, fifty years old
example of such an inequality is the Golden-Thompson trace inequality: for Hermitian
n-by-n matrices S and T ,
Tr eS+T ≤ Tr eS/2 eT eS/2 .
A decade after, Lieb and Thirring [77] obtained a stronger, remarkable trace inequality:
for all positive semidefinite n-by-n matrices A, B ∈ M+
n and all integers p ≥ 1,

Tr (ABA)p ≤ Tr Ap B p Ap . (2.1.1)

This was finally extended some fifteen years later by Araki [7] as a very important
theorem in matrix analysis and its applications. Given X, Y ∈ M+
n , we write X ≺wlog Y

19
20 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

when the series of n inequalities holds,

k
Y k
Y
λj (X) ≤ λj (Y )
j=1 j=1

for k = 1, . . . n, where λj (·) stands for the eigenvalues arranged in decreasing order. If
further equality occurs for k = n, we write X ≺log Y . Araki’s theorem considerably
strenghtens the Lieb-thirring trace inequality as the beautiful log-majorization

(ABA)p ≺log Ap B p Ap (2.1.2)

for all real numbers p ≥ 1. In particular, this ensures (2.1.1) for all p ≥ 1.
Log- and weak log-majorization relations play a fundamental role in matrix analysis,
a basic one for normal operators X, Y ∈ Mn asserts that

|X + Y | ≺wlog |X| + |Y |. (2.1.3)

This useful version of the triangle inequality belongs to the folklore and is a byproduct
of Horn’s inequalities, see the proof of [26, Corollary 1.4].
This article aims to provide new matrix inequalities containing (2.1.2) and (2.1.3).
These inequalities are given in Section 2. The first part dealing with positive operators
is closely related to a recent paper of Hiai [66]. The second part of Section 2 considers
normal operators and contains our main theorem (Theorem 7.3.1), mentioned in the
Abstract.
Our main idea, and technical tool, is Theorem 2.1.2 below. It establishes the log-
convexity of a two variables functional. Fixing one variable in this functional yields a
generalization of (2.1.2) involving a third matrix Z ∈ Mn , of the form

(AZ ∗ BZA)p ≺wlog Ap Z ∗ B p ZAp .

We will also derive the following weak log-majorization which contains both (2.1.2) and
(2.1.3) and thus unifies these two inequalities.

Proposition 2.1.1. Let A ∈ M+

n and let X, Y ∈ Mn be normal. Then, for all p ≥ 1,

|A(X + Y )A|p ≺wlog 2p−1 Ap (|X|p + |Y |p )Ap .

Letting X = Y = B in Proposition 2.1.1 we have (2.1.2), more generally,

|AXA|p ≺log Ap |X|p Ap (2.1.4)

for all A ∈ M+ n and normal matrices X ∈ Mn . When X is Hermitian, this was noted
by Audenaert [8, Proposition 3]. If A is the identity and p = 1, Proposition 2.1.1 gives
(2.1.3). From (2.1.4) follows several nice inequalities for the matrix exponential, due
to Cohen and al. [54], [55], including the Golden-Thompson trace inequality and the
elegant relation
|eZ | ≺log eRe Z (2.1.5)
2.2. ARAKI TYPE INEQUALITIES 21

for all matrices Z ∈ Mn , where Re Z = (Z + Z ∗ )/2, [54, Theorem 2].

Fixing the other variable in Theorem 2.1.2 below entails a Hölder inequality due
to Kosaki. Several matrix versions of an inequality of Littlewood related to Hölder’s
inequality will be also obtained.
The two variables in Theorem 2.1.2 are essential and reflect a construction with
the perspective of a convex function. Recall that a norm on Mn is symmetric whenever
kUAV k = kAk for all A ∈ Mn and all unitary U, V ∈ Mn . For X, Y ∈ M+ n , the condition
X ≺wlog Y implies kXk ≤ kY k for all symmetric norms. We state our log-convexity
theorem.

Theorem 2.1.2. Let A, B ∈ M+n and Z ∈ Mn . Then, for all symmetric norms and
α > 0, the map
αp
(p, t) 7→ At/p ZB t/p
is jointly log-convex on (0, ∞) × (−∞, ∞).

Here, if A ∈ M+
n is not invertible, we naturally define for t ≥ 0, A
−t
:= (A + F )−t E
where F is the projection onto the nullspace of A and E is the range projection of A.
The next two sections present many hidden consequences of Theorem 2.1.2, several
of them extending (2.1.2) and/or (2.1.3), for instance,
p
T + T∗ p |T |
p
+ |T ∗ |p p
A A ≺wlog A A
2 2

for all A ∈ M+n , p ≥ 1, and any T ∈ Mn . The proof of Theorem 2.1.2 is in Section 4.
The last section provides a version of Theorem 2.1.2 for operators acting on an infinite
dimensional Hilbert space.

2.2 Araki type inequalities

2.2.1 With positive operators
To obtain new Araki’s type inequalities, we fix t = 1 in Theorem 2.1.2 and thus use the
following special case.
Corollary 2.2.1. Let A, B ∈ M+
n and Z ∈ Mn . Then, for all symmetric norms and
α > 0, the map
αp
p 7→ A1/p ZB 1/p
is log-convex on (0, ∞).

We may now state a series of corollaries extending Araki’s theorem.

Corollary 2.2.2. Let A, B ∈ M+ n and p ≥ 1. Let Z ∈ Mn be a contraction. Then, for
all symmetric norms and α > 0,

k(AZ ∗ BZA)αp k ≤ k(Ap Z ∗ B p ZAp )α k.

22 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

Let I be the identity of Mn . A matrix Z is contractive, or a contraction, if Z ∗ Z ≤ I,

equivalently if its operator norm satisfies kZk∞ ≤ 1.

Proof. The function f (p) = k|B 1/p ZA1/p |2αp k is log-convex, hence convex on (0, ∞), and
bounded since Z is contractive, 0 ≤ f (p) ≤ kBk2α 2α
∞ kAk∞ kIk. Thus f (p) is nonincreasing,
so f (1) ≥ f (p) for all p ≥ 1. Replacing B by B and A by Ap completes the proof.
p/2

Let k · k{k} , k = 1, . . . , n, denote the normalized Ky Fan k-norms on Mn ,

k
1X
kT k{k} = λj (|T |).
k j=1

Since, for all A ∈ M+

n,
( k )1/k
1/α
Y
lim kAα k{k} = λj (A)
α→0+
j=1

we obtain from Corollary 3.4.2 applied to the normalized Ky Fan k-norms, with α → 0+ ,
a striking weak-log-majorization extending Araki’s theorem.

Corollary 2.2.3. Let A, B ∈ M+

n and p ≥ 1. Then, for all contractions Z ∈ Mn ,

(AZ ∗ BZA)p ≺wlog Ap Z ∗ B p ZAp .

If Z = I, we have the determinant equality and thus Araki’s log-majorization (2.1.2).

Corollary 3.4.2 and 2.2.3 are equivalent. Our proof of these extensions of Araki’s theorem
follows from the two variables technic of Theorem 2.1.2. It’s worth mentioning that
Fumio Hiai also obtained Corollary 2.2.3 in the beautiful note [66]. Hiai’s approach is
based on some subtle estimates for the operator geometric mean.
For X, Y ∈ M+ n , the notation X ≺
wlog
Y indicates that the series of n inequalities
holds,
k
Y Yk
νj (X) ≥ νj (Y )
j=1 j=1

for k = 1, . . . n, where νj (·) stands for the eigenvalues arranged in increasing order. The
following so-called super weak-log-majorization is another extension of Araki’s theorem.
A matrix Z is expansive when Z ∗ Z ≥ I.

Corollary 2.2.4. Let A, B ∈ M+

n and p ≥ 1. Then, for all expansive matrices Z ∈ Mn ,

(AZ ∗ BZA)p ≺wlog Ap Z ∗ B p ZAp .

Proof. By a limit argument, we may assume invertibility of A and B. Taking inverses,

and using that Z −1 is contractive, Corollary 2.2.4 is then equivalent to Corollary 2.2.3.
2.2. ARAKI TYPE INEQUALITIES 23

Corollaries 2.2.3-2.2.4 imply a host of trace inequalities. We say that a continuous

function h : [0, ∞) → (−∞, ∞) is e-convex, (resp. e-concave), if h(et ) is convex, (resp.
concave) on (−∞, ∞). For instance, for all α > 0, t 7→ log(1 + tα ) is e-convex, while
t 7→ log(tα /(t + 1)) is e-concave. The equivalence between Corollary 2.2.3 and Corollary
2.2.5 below is a basic property of majorization discussed in any monograph on this topic
such as [13] and [67].

Corollary 2.2.5. Let A, B ∈ M+

n , Z ∈ Mn , and p ≥ 1.

(a) If Z is contractive and f (t) is e-convex and nondecreasing, then

Tr f ((AZ ∗ BZA)p ) ≤ Tr f (Ap Z ∗ B p ZAp ).

(b) If Z is expansive and g(t) is e-concave and nondecreasing, then

Tr g((AZ ∗ BZA)p ) ≥ Tr g(Ap Z ∗ B p ZAp ).

We will propose in Section 4 a proof of Theorem 2.1.2 making use of antisymmetric

tensor powers, likewise in the proof of Araki’s log-majorization. We will also indicate
another, more elementary way, without antisymmetric tensors. The antisymmetric ten-
sor technic goes back to Hermann Weyl, cf. [13], [67]. We use it to derive our next
corollary.
Corollary 2.2.6. Let A, B ∈ M+
n and Z ∈ Mn . For each j = 1, . . . , n, the function
defined on (0, ∞)
1/p
p 7→ λj (Ap Z ∗ B p ZAp )
converges as p → ∞.

Proof. We may assume that Z is contractive. As in the proof of Corollary 3.4.2 we then
see that the function g(p) = λp1 (A1/p Z ∗ B 1/p ZA1/p ) is log-convex and bounded, hence
nonincreasing on (0, ∞). Therefore g(p) converges as p → 0 and so g(1/p) converges
1/p
as p → ∞. Thus p 7→ λ1 (Ap Z ∗ B p ZAp ) converges as p → ∞. Considering k-th
antisymmetric tensor products, k = 1, . . . , n, we infer the convergence of
k
1/p 1/p
Y
λj (Ap Z ∗ B p ZAp ) = λ1 (∧k A)p ∧k Z ∗ (∧k B)p ∧k Z(∧k A)p


p 7→
j=1

1/p
and so, the convergence of p 7→ λj (Ap Z ∗ B p ZAp ) as p → ∞, for each j = 1, 2, . . . .

When Z = I, Audenaert and Hiai [9] recently gave a remarkable improvement of

Corollary 2.2.6 by showing that p 7→ (Ap B p Ap )1/p converges in Mn as p → ∞. We do
not know whether such a reciprocal Lie-Trotter limit still holds with a third matrix Z
as in Corollary 2.2.6.
It is possible to state Corollary 2.2.3 in a stronger form involving a positive linear
map Φ. Such a map is called sub-unital when Φ(I) ≤ I.
24 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

Corollary 2.2.7. Let A, B ∈ M+

n and p ≥ 1. Then, for all positive linear, sub-unital
map Φ : Mn → Mn ,
(AΦ(B)A)p ≺wlog Ap Φ(B p )Ap .

Proof. We may assume (the details are given, for a more general class of maps, in the
proof of Corollary 2.3.7) that
Xm
Φ(X) = Zi∗ XZi
i=1

where m = n2 and Zi ∈ Mn , i = 1, . . . , m, satisfy m ∗

P
i=1 Zi Zi ≤ I. Corollary 2.2.7 then
follows from Corollary 2.2.3 applied to the operators Ã, B̃, Z̃ ∈ Mmn ,
     
A 0n · · · 0n B 0n · · · 0n Z1 0n · · · 0n
0n 0n · · · 0n   0n B · · · 0n   Z2 0n · · · 0n 
à =  .. .. . . ..  , B̃ =  .. .. . . ..  , Z̃ =  .. .. . . .. 
     
. . . . . . . .  . . . .
0n 0n · · · 0n 0n 0n · · · B Zm 0n · · · 0n

where 0n stands for the zero matrix in Mn .

Corollary 2.2.7 can be applied for the Schur product ◦ (i.e., entrywise product) in
Mn .

Corollary 2.2.8. Let A, B, C ∈ M+

n and p ≥ 1. If C has all its diagonal entries less
than or equal to one, then

(A(C ◦ B)A)p ≺wlog Ap (C ◦ B p )Ap .

Proof. The map X 7→ C ◦ X is a positive linear, sub-unital map on Mn .

Corollary 2.2.8 with the matrix C whose entries are all equal to one is Araki’s log-
majorization. With C = I, Corollary 2.2.8 is already an interesting extension of Araki’s
theorem as we may assume that B is diagonal in (2.1.2). We warn the reader that the
super weak-log-majorization, for A, B ∈ M+ p
n and p ≥ 1, (A(I ◦B)A) ≺
wlog
Ap (I ◦B p )Ap
does not hold, in fact, in general, det2 I ◦ B < det I ◦ B 2 .

2.2.2 With normal operators

To obtain Proposition 2.1.1 we need the following generalization of Corollary 2.2.7.

Theorem 2.2.9. Let A ∈ M+ n and let N ∈ Mm be normal. Then, for all positive linear,
sub-unital maps Φ : Mm → Mn , and p ≥ 1,

|AΦ(N)A|p ≺wlog Ap Φ(|N|p )Ap .

2.2. ARAKI TYPE INEQUALITIES 25

Proof. By completing, if necessary, our matrices A and N with some 0-entries, we may
assume that m = n and then, as in the proof of Corollary 2.2.7, that Φ is a congruence
map with a contraction Z̃, Φ(X) = Z̃X Z̃ ∗ . Now, we have with the polar decomposition
N = U|N|,

|AZ̃N Z̃ ∗ A| = AZ̃|N|1/2 U|N|1/2 Z̃ ∗ A

≺log AZ̃|N|Z̃ ∗ A

by using Horn’s log-majorization |XKX ∗ | ≺wlog XX ∗ for all X ∈ Mn and all contrac-
tions K ∈ Mn . Hence, from Corollary 2.2.3, for all p ≥ 1,
p
|AZ̃N Z̃ ∗ A|p ≺log AZ̃|N|Z̃ ∗ A ≺wlog Ap Z̃|N|p Z̃ ∗ Ap

which completes the proof.

We are in a position to prove Proposition 2.1.1 whose m-variables version is given

here.

Corollary 2.2.10. Let A ∈ M+n and let X1 , · · · , Xm ∈ Mn be normal. Then, for all
p ≥ 1, ! p !
m
X Xm
A Xk A ≺wlog mp−1 Ap |Xk |p Ap .
k=1 k=1

Proof. Applying Theorem 7.3.1 to N = X1 ⊕ · · · ⊕ Xm and to the unital, positive linear

map Φ : Mmn → Mn ,  
S1,1 · · · S1,m m
 .. . . ..  7→ 1 X S
 . . .  k,k
m k=1
Sm,1 · · · Sm,m
yields Pm p Pm p
Xk p k=1 |Xk |
A k=1
A ≺wlog A Ap
m m
which is equivalent to the desired inequality.

A special case of Corollary 2.2.10 deals with the Cartesian decomposition of an arbi-
trary matrix.

Corollary 2.2.11. Let X, Y ∈ Mn be Hermitian. Then, for all p ≥ 1,

|A(X + iY )A|p ≺wlog 2p−1 Ap (|X|p + |Y |p )Ap

where the constant 2p−1 is the best possible.

26 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

To check that 2p−1 is optimal, take A = I ∈ M2n and pick any two-nilpotent matrix,
 
0 T
X + iY = .
0 0
For a single normal operator, Corollary 2.2.10 gives (2.1.4) as we have equality for
the determinant. This entails the following remarkable log-majorization for the matrix
exponential.
Corollary 2.2.12. Let A, B ∈ Mn . Then,
eA+B ≺log eRe A/2 eRe B eRe A/2 .

Corollary 2.2.12 contains (2.1.5) and shows that when A and B are Hermitian we
have the famous Thompson log-majorization, [91, Lemma 6],
eA+B ≺log eA/2 eB eA/2 (2.2.1)
which entails
keA+B k ≤ keA/2 eB eA/2 k
for all symmetric norms. For the operator norm this is Segal’s inequality while for the
trace norm this is the Golden-Thompson inequality. Taking the logarithms in (2.2.1),
we have a classical majorization between A + B and log eA/2 eB eA/2 . Since t 7→ |t| is
convex, we infer, replacing B by −B that
kA − Bk ≤ k log(eA/2 e−B eA/2 )k
for all symmetric norms. For the Hilbert-Schmidt norm, this is the Exponential Metric
Increasing inequality, reflecting the nonpositive curvature of the positive definite cone
with its Riemannian structure ([15, Chapter 6]).
Corollary 2.2.12 follows from (2.1.4) combined with the Lie Product Formula [13, p.
254] as shown in the next proof. Note that Corollary 2.2.12 also follows from Cohen’s
log-majorization (2.1.5) combined with Thompson’s log-majorization (2.2.1), thus we
do not pretend to originality.
Proof. We have a Hermitian matrix C such that, using the Lie Product Formula,

n
eA+B = eRe A+Re B+iC = lim e(Re A+Re B)/2n eiC/n e(Re A+Re B)/2n .
n→+∞

On the other hand, by (2.1.4), for all n ≥ 1,

n
e(Re A+Re B)/2n eiC/n e(Re A+Re B)/2n ≺log eRe A+Re B

so that
eA+B ≺log eRe A+Re B .
Using again the Lie Product Formula,

n
eRe A+Re B = lim eRe A/2n eRe B/n eRe A/2n ,
n→+∞

combined with (2.1.4) (or (2.1.2)) completes the proof.

2.2. ARAKI TYPE INEQUALITIES 27

Theorem 2.2.9 is the main result of Section 2 as all the other results in this section
are special cases. One more elegant extension of Araki’s inequality follows, involving an
arbitrary matrix.
Corollary 2.2.13. Let A ∈ M+
n and p ≥ 1. Then, for any T ∈ Mn ,
p
T + T∗ |T |p + |T ∗ |p p
A A ≺wlog Ap A.
2 2

Proof. It suffices to apply Theorem 7.3.1 to

 
0 T
N=
T∗ 0
and to the unital, positive linear map Φ : M2n → Mn ,
 
B C B+C +D+E
7 → .
D E 2

We apply Theorem 7.3.1 to Schur products in the next two corollaries.

Corollary 2.2.14. Let A ∈ M+
n and let X, Y ∈ Mn be normal. Then, for all p ≥ 1,

|A(X ◦ Y )A|p ≺wlog Ap (|X|p ◦ |Y |p )Ap .

Proof. We need to see the Schur product as a positive linear map,

X ◦ Y = Φ(X ⊗ Y )
where Φ : Mn ⊗ Mn → Mn merely consists in extracting a principal submatrix. Setting
N = X ⊗ Y in Theorem 7.3.1 completes the proof.

We note that Corollary 2.2.14 extends (2.1.4) (with X in diagonal form and Y = I)
and contains the classical log-majorization for normal operators,
|X ◦ Y | ≺wlog |X| ◦ |Y |.
As a last illustration of the scope of Theorem 7.3.1 we have the following result.
Corollary 2.2.15. Let A ∈ M+
n and p ≥ 1. Then, for any T ∈ Mn ,

|A(T ◦ T ∗ )A|p ≺wlog Ap (|T |p ◦ |T ∗ |p )Ap .

Proof. We apply Corollary 2.2.14 to the pair of Hermitian operators in M2n ,

0 T∗
   
0 T
X= , Y =
T 0 T∗ 0
with A ⊕ A in M+2n . We then obtain
p
A(T ◦ T ∗ )A A(|T |p ◦ |T ∗ |p )A
  
0 0
≺wlog
A(T ◦ T ∗ )A 0 0 A(|T |p ◦ |T ∗ |p )A
which is equivalent to the statement of our corollary.
28 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

2.3 Hölder type inequalities

Now we turn to Hölder’s type inequalities. Fixing p = 1 in Theorem 2.1.2, we have the
following special case.
Corollary 2.3.1. Let A, B ∈ M+
n and Z ∈ Mn . Then, for all symmetric norms and
α > 0, the map
α
t 7→ At ZB t
is log-convex on (−∞, ∞).

This implies a fundamental fact, the Löwner-Heinz inequality stating the operator
monotonicity of tp , p ∈ (0, 1).
Corollary 2.3.2. Let A, B ∈ M+ p p
n . If A ≥ B, then A ≥ B for all p ∈ (0, 1).

Proof. Corollary 2.3.1 for the operator norm, with Z = I, α = 2, and the pair A−1/2 , B 1/2
in place of the pair A, B shows that f (t) = kA−t/2 B t A−t/2 k∞ is log-convex. Hence for
p ∈ (0, 1), we have f (p) ≤ f (1)p f (0)1−p . Since f (0) = 1 and by assumption f (1) ≤ 1,
we obtain f (p) ≤ 1 and so Ap ≥ B p .
Corollary 2.3.1 entails a Hölder inequality with a parameter. This inequality was first
proved by Kosaki [74, Theorem 3]. Here, we state it without the weight Z.
Corollary 2.3.3. Let X, Y ∈ Mn and p, q ≥ 1 such that p−1 + q −1 = 1. Then, for all
symmetric norms and α > 0,

k|XY |α k ≤ k|X|αp k1/p k|Y |qα k1/q .

Proof. Let A, B ∈ M+ n with B invertible. By replacing B with B

−1
and letting Z = B
t 1−t α
in Corollary 2.3.1 show that t 7→ k|A B | k is log-convex on (−∞, ∞). Thus, for t ∈
(0, 1), k|At B 1−t |α k ≤ kAα kt kB α k1−t . Then, choose A = |X|p, B = |Y ∗ |q , t = 1/p.

More original Hölder’s type inequalities are given in the next series of corollaries.
Corollary 2.3.4. Let A ∈ M+
n and Z ∈ Mn,m . Then, for all symmetric norms and
α > 0, the map

αp
(p, t) 7→ Z ∗ At/p Z
is jointly log-convex on (0, ∞) × (−∞, ∞).

Proof. By completing, if necessary, our matrices with some 0-entries, we may suppose
m = n and then apply Theorem 2.1.2 with B = I.

, am ) and w = (w1 , · · · , wm ) be two m-tuples in R+

Corollary 2.3.5. Let a = (a1 , · · · P
and define, for all p > 0, kakp := ( m p 1/p
i=1 wi ai ) . Then, for all p, q > 0 and θ ∈ (0, 1),

kak 1 ≤ kakθ1 kak1−θ

1 .
θp+(1−θ)q p q
2.3. HÖLDER TYPE INEQUALITIES 29

1/2 1/2
Proof. Fix t = 1 and pick A = diag(a1 , . . . am ) and Z ∗ = (w1 , . . . , wm ) in the previous
corollary.

Corollary 2.3.5 is the classical log-convexity of p → k · k1/p , or Littlewood’s version

of Hölder’s inequality [59, Theorem 5.5.1]. The next two corollaries, seemingly stronger
but actually equivalent to Corollary 2.3.4, are also generalizations of this inequality.
Corollary 2.3.6. Let Ai ∈ M+n and Zi ∈ Mn,m , i = 1, . . . , k. Then, for all symmetric
norms and α > 0, the map
( k )αp
t/p
X
(p, t) 7→ Zi∗ Ai Zi
i=1

is jointly log-convex on (0, ∞) × (−∞, ∞).

The unweighted case, Zi = I for all i = 1, . . . , k, is especially interesting. With

t = α = 1, it is a matrix version of the unweighted Littlewood inequality.
Proof. Apply Corollary 2.3.4 with A = A1 ⊕ · · · ⊕ Ak and Z ∗ = (Z1∗ , . . . , Zk∗ ).

Corollary 2.3.7. Let A ∈ M+ + +

m and let Φ : Mm → Mn be a positive linear map. Then,
for all symmetric norms and α > 0, the map
αp
(p, t) 7→ Φ(At/p )


is jointly log-convex on (0, ∞) × (−∞, ∞).

Proof. When restricted to the ∗-commutative subalgebra spanned by A, the map Φ has
the form m X n
X
∗
Φ(X) = Zi,j XZi,j (2.3.1)
i=1 j=1

for some rank 1 or 0 matrices Zi,j ∈ Mm,n , i = 1, . . . , m, j = 1, . . . , n. So we are in the

range of the previous
Pcorollary. To check the decomposition (2.3.1), ∗write the spectral
decomposition A = m i=1 λi (A)Ei with rank one projections Ei = xi xi for some column
vectors xi ∈ Mm,1 and set Zi,j = xi Ri,j where Ri,j ∈ M1,n is the j-th row of Φ(Ei )1/2 .

The above proof shows a classical fact, a positive linear map on a commutative domain
is completely positive. Our proof seems shorter than the ones in the literature. We close
this section with an application to Schur products.
Corollary 2.3.8. Let A, B ∈ M+
n . If p ≥ r ≥ s ≥ q and p + q = r + s, then, for all
symmetric norms and α > 0,

k{Ar ◦ B s }α kk{As ◦ B r }α k ≤ k{Ap ◦ B q }α kk{Aq ◦ B p }α k

and
k{Ar ◦ B s }α k + k{As ◦ B r }α k ≤ k{Ap ◦ B q }α k + k{Aq ◦ B p }α k.
30 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

Proof. By a limit argument we may assume invertibility of A and B. Let w := (p + q)/2.

We will show that the maps

t 7→ k{Aw+t ◦ B w−t }α k, t 7→ k{Aw−t ◦ B w+t }α k (2.3.2)

are log-convex on (−∞, ∞). This implies that the functions

f (t) = k{Aw+t ◦ B w−t }α kk{Aw−t ◦ B w+t }α k

and
g(t) = k{Aw+t ◦ B w−t }α k + k{Aw−t ◦ B w+t }α k
are convex and even, hence nondecreasing on [0, ∞). So we have

f ((r − s)/2) ≤ f ((p − q)/2) and g((r − s)/2) ≤ g((p − q)/2)

which prove the corollary.

To check the log-convexity of the maps (2.3.2), we see the Schur product as a positive
linear map acting on a tensor product,A ◦ B = Ψ(A ⊗ B). By Corollary 2.3.7, the map

t 7→ k{Φ(Z t )}α k

is log-convex on (−∞, ∞) for any positive matrix Z ∈ Mn ⊗ Mn and any positive linear
map Φ : Mn ⊗ Mn → Mn . Taking Z = A ⊗ B −1 and

Φ(X) = Ψ(Aw/2 ⊗ B w/2 · X · Aw/2 ⊗ B w/2 )

we obtain the log-convexity of the first map t 7→ k{Aw+t ◦ B w−t }α k in (2.3.2). The
log-convexity of the second one is similar.

2.4 Proof of Theorem 2.1.2

In the proof of the theorem, we will denote the k-th antisymmetric power ∧k T of an
operator T simply as Tk . The symbol k · k∞ stands for the usual operator norm while
ρ(·) denotes the spectral radius. Given A ∈ M+ ↓
n we denote by A the diagonal matrix
with the eigenvalues of A in decreasing order down to the diagonal, A↓ = diag(λj (A)).

Proof. Recall that if A ∈ M+ n is not invertible and t ≥ 0, we define A

−t
as the generalized
t −t −t
inverse of A , i.e., A := (A + F ) E where F is the projection onto the nullspace of A
and E is the range projection of A. With this convention, replacing if necessary Z by
EZE ′ where E is the range projection of A and E ′ that of B, we may and do assume
that A and B are invertible.
Let
Yk
gk (t) := λj (|At ZB t |) = kAtk Zk Bkt k∞
j=1
2.4. PROOF OF THEOREM 2.1.2 31

Then
(t+s)/2 (t+s)/2 1/2
gk ((t + s)/2) = kAk Zk Bkt+s Zk∗ Ak k∞
= ρ1/2 (Atk Zk Bkt+s Zk∗ Ask )
≤ kAtk Zk Bkt+s Zk∗ Ask k1/2
∞

≤ kAtk Zk Bkt k1/2 s ∗ s 1/2

∞ kBk Zk Ak k∞

= {gk (t)gk (s)}1/2 .

Thus t 7→ gk (t) is log-convex on (−∞, ∞) and so (p, t) 7→ gkp (t/p) is jointly log-convex
on (0, ∞)×(−∞, ∞). Indeed, its logarithm p log gk (t/p) is the perspective of the convex
function log gk (t), and hence is jointly convex. Therefore
 
(p+q)/2 (t + s)/2
gk ≤ {gkp (t/p)gkq (s/q)}1/2 (2.4.1)
(p + q)/2
for k = 1, 2, . . . , n, with equality for k = n as it then involves the determinant. This is
equivalent to the log-majorization
p+q q
t+s t+s 2 p ↓
A p+q ZB p+q ≺log |At/p ZB t/p | 2 ↓ As/q ZB s/q 2

which is equivalent, for any α > 0, to the log-majorization

t+s t+s α p+q
2
αp
↓ αq
↓
A p+q ZB p+q ≺log At/p ZB t/p 2
As/q ZB s/q 2

ensuring that
t+s t+s α p+q
2
αp
↓ αq
↓
A p+q ZB p+q ≤ At/p ZB t/p 2
As/q ZB s/q 2
(2.4.2)

for all symmetric norms. Thanks to the Cauchy-Schwarz inequality for symmetric norms,
we then have
t+s t+s α p+q
2 αp 1/2 αq 1/2
A p+q ZB p+q ≤ At/p ZB t/p At/q ZB t/q (2.4.3)

which means that

αp
(p, t) 7→ At/p ZB t/p
is jointly log-convex on (0, ∞) × (−∞, ∞).

Denote by Ik the identity of Mk and by detk the determinant on Mk . If n ≥ k, Θ(k, n)

stands for the set of n × k isometry matrices T , i.e, T ∗ T = Ik . One easily checks the
variational formula, for A ∈ Mn and k = 1, . . . , n,
k
Y
λj (|A|) = max |detk V ∗ AW | .
V,W ∈Θ(k,n)
j=1
32 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

From this formula follow two facts, Horn’s inequality,

k
Y k
Y
λj (|AB|) ≤ λj (|A|)λj (|B|)
j=1 j=1

for all A, B ∈ Mn and k = 1, . . . , n, and, making use of Schur’s triangularization, the

inequality
Y k Yk
λj (|AB|) ≤ λj (|BA|)
j=1 j=1

whenever AB is normal (indeed, by Schur’s theorem we may assume that BA is upper

triangular with the eigenvalues of BA, hence of AB, down to the diagonal and our
variational formula then gives the above log-majorization). This shows that the proof
of Theorem 2.1.2 can be written without the machinery of antisymmetric tensors.
The novelty of this proof consists in using the perspective of a one variable convex
function. One more perspective yields the following variation of Corollary 2.2.1.

Corollary 2.4.1. Let A, B ∈ M+

m , let Z ∈ Mm . Then, for all symmetric norms and
α > 0, the map
α p
p 7→ A1/p ZB 1/p

is log-convex on (0, ∞).

Proof. By Theorem 2.1.2 with fixed p = 1, the map t 7→ log (k|At ZB t |α k) is convex on
(0, ∞), thus its perpective
 p

|At/p ZB t/p |α |At/p ZB t/p |α


(p, t) 7→ p log = log

is jointly log-convex on (0, ∞) × (0, ∞). Now fixing t = 1 completes the proof.

From this corollary we may derive the next one exactly as Corollary 2.3.7 follows
from Theorem 2.1.2. This result is another noncommutative version of Littlewood’s
inequality ([59, Theorem 5.5.1]).

Corollary 2.4.2. Let Φ : Mm → Mn be a positive linear map and let A ∈ M+

m . Then,
for all symmetric norms and α > 0, the map
p
p 7→ Φα (A1/p )

is log-convex on (0, ∞).

2.5. HILBERT SPACE OPERATORS 33

2.5 Hilbert space operators

In this section we give a version of Theorem 2.1.2 for the algebra B of bounded linear
operators on a separable, infinite dimensional Hilbert space H. We first include a brief
treatment of symmetric norms for operators in B. Our approach does not require to
discuss any underlying ideal, we refer the reader to [89, Chapter 2] for a much more
complete discussion.
We may define symmetric norms on B in a closely related way to the finite dimensional
case as follows. Let F be the set of finite rank operators and F+ its positive part.

Definition 2.5.1. A symmetric norm k · k on B is a functional taking value in [0, ∞]

such that:

(1) k · k induces a norm on F.

(2) If {Xn } is a sequence in F+ strongly increasing to X, then kXk = limn kXn k.

(3) kKZLk ≤ kZk for all Z ∈ B and all contractions K, L ∈ B.

The reader familiar to the theory of symmetrically normed ideals may note that our
definition of a symmetric norm is equivalent to the usual one. More precisely, restricting
k · k to the set where it takes finite values, Definition 2.5.1 yields the classical notion of
a symmetric norm defined on its maximal ideal.
Definition 2.5.1 shows that a symmetric norm on B induces a symmetric norm on
Mn for each n, say k · kMn . In fact k · k can be regarded as a limit of the norms k · kMn ,
see Lemma 2.5.6 for a precise statement, so that basic properties of symmetric norms
on Mn can be extended to symmetric norms on B. For instance the Cauchy-Schwarz
inequality also holds for symmetric norms on B, (with possibly the ∞ value) as well as
the Ky Fan principle for A, B ∈ B+ : If A ≺w B, then kAk ≤ kBk for all symmetric
norms. In fact, even for a noncompact operator A ∈ B+ , the sequence {λj (A)}∞ j=1 and
↓
the corresponding diagonal operator A = diag(λj (A)) are well defined, via the minmax
formulae (see [65, Proposition 1.4])

λj (A) = inf {kEAEk∞ : E projection with rank(I − E) = j − 1}

E

The Ky Fan principle then still holds for A, B ∈ B+ by Lemma 2.5.6 and the obvious
property
λj (A) = lim λj (En AEn )
n→∞

for all sequences of finite rank projections {En }∞

n=1 strongly converging to the identity.
k
Q k
Note also that we still have k ∧ Ak∞ = j=1 λj (A).
Thus we have the same tools as in the matrix case and we will be able to adapt the
proof of Theorem 2.1.2 for B. The infinite dimensional version of Theorem 2.1.2 is the
following statement.
34 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

Theorem 2.5.2. Let A, B ∈ B+ , let Z ∈ B. Then, for all symmetric norms and α > 0,
the map
αp
(p, t) 7→ At/p ZB t/p
is jointly log-convex on (0, ∞) × (0, ∞). This map takes its finite values in the open
quarter-plan
Ω(p0 , t0 ) = {(p, t) | p > p0 , t > t0 }
for some p0 , t0 ∈ [0, ∞], or on its closure Ω(p0 , t0 ).

Note that, contrarily to Theorem 2.1.2, we confine the variable t to the positive half-
line. Indeed, when dealing with a symmetric norm, the operators A and B are often
compact, so that, for domain reasons, we cannot consider two unbounded operators such
as A−1 and B −1 .
Proof. Note that AZB = 0 if and only if Aq ZB q = 0, for any q > 0. In this case,
our map is the 0-map, and its logarithm with constant value −∞ can be regarded as
convex. Excluding this trivial case, our map takes values in (0, ∞] and it makes sense to
consider the log-convexity property. We may reproduce the proof of Theorem 2.1.2 and
obtain (2.4.1) for all k = 1, 2, · · · . This leads to weak-logmajorizations and so to a weak
majorization equivalent (Ky Fan’s principle in B) to (2.4.2), with possibly the ∞ value
on the right side or both sides. The Cauchy-Schwarz inequality for symmetric norms in
B yields (2.4.3) (possibly with the ∞ value). Therefore our map is jointly log-convex.
To show that the domain where it takes finite values is Ω(p0 , t0 ) or Ω(p0 , t0 ), it suffices
to show the following two implications:
αp
Let 0 < t < s and 0 < p < q. If At/p ZB t/p < ∞, then
αp
(i) As/p ZB s/p < ∞, and
αq
(ii) At/q ZB t/q < ∞.

Since 0 < t < s ensures that, for some constant c = c(s, t) > 0,

λj (|At/p ZB t/p |) ≥ cλj (|As/p ZB s/p |)

for all j = 1, 2, . . ., we obtain (i). To obtain (ii) we may assume that Z is a contraction.
Then arguing as in the proof of Corollary 3.4.2 we see that the finite value map
αp
p 7→ At/p ZB t/p

is nonincreasing for all Ky-Fan norms. Thus this map is also nonincreasing for all
symmetric norms. This gives (ii).

Exactly as in the matrix case, we can derive the following two corollaries.
Corollary 2.5.3. Let A, B ∈ B+ and p ≥ 1. Then, for all contractions Z ∈ B,

(AZ ∗ BZA)p ≺wlog Ap Z ∗ B p ZAp .

2.5. HILBERT SPACE OPERATORS 35

Corollary 2.5.4. Let A, B ∈ B+ and let Z ∈ B be a contraction. Assume that at

least one of these three operators is compact. Then, if p ≥ 1 and f (t) is e-convex and
nondecreasing,
Tr f ((AZ ∗ BZA)p ) ≤ Tr f (Ap Z ∗ B p ZAp ).

Here, we use the fact that for X ∈ K+ and a nondecreasing continuous function
f : [0, ∞) → (−∞, ∞), we can define Tr f (X) as an element in [−∞, ∞] by

k
X
Tr f (X) = lim f (λj (X)).
k→∞
j=1

Given a symmetric norm k · k on B, the set where k · k takes a finite value is an ideal.
We call it the maximal ideal of k · k or the domain of k · k. From Theorem 3.4.5 we
immediately infer our last corollary.

Corollary 2.5.5. Let A, B ∈ B+ and Z ∈ B. Suppose that AZB ∈ J, the domain of a

symmetric norm. Then, for all q ∈ (0, 1), we also have |Aq ZB q |1/q ∈ J.

Following [89, Chapter 2], we denote by J(0) the k · k-closure of the finite rank opera-
tors. In most cases J = J(0) , however the strict inclusion J(0) ⊂6= J may happen. We do
not know whether we can replace in the last corollary J by J(0) .
We close our article with two simple lemmas and show how the Cauchy-Schwarz
inequality for the infinite dimensional case follows from the matrix case.

Lemma 2.5.6. Let k · k be a symmetric norm on B and let {En }∞ n=1 be an increasing
sequence of finite rank projections in B, strongly converging to I. Then, for all X ∈ B,
kXk = limn kEn XEn k.

Proof. We first show that kEn Xk → kXk as n → ∞. Since kEn Xk = k(X ∗ En X)1/2 k
and (X ∗ En X)1/2 ր |X| by operator monotonicity of t1/2 , we obtain limn kEn Xk = kXk
by Definition 2.5.1(2). Similarly, limk kEn XEk k = kEn Xk, and so limn kEn XEk(n) k =
kXk, and thus, by Definition 2.5.1(3), limp kEp XEp k = kXk.

Lemma 2.5.7. Let k · k be a symmetric norm on B and let {En }∞ ∞

n=1 and {Fn }n=1 be
two increasing sequences of finite rank projections in B, strongly converging to I. Then,
for all X ∈ B, kX ∗ Xk = limn kEn X ∗ Fn XEn k.

Proof. By Definition 2.5.1(2)-(3), the map n → 7 kEn X ∗ Fn XEn k is nondecreasing. By

Definition 2.5.1(2), for any integer p, its limit is greater than or equal kEp X ∗ XEp k. By
Lemma 2.5.6, the limit is precisely kX ∗ Xk.
36 CHAPTER 2. MAJORIZATION AND PERSPECTIVE

Let X, Y ∈ B, let k · k be a symmetric norm on B, and let {En }∞

n=1 be as in the above
lemma. Let Fn be the range projection of Y En . We have by Lemma 2.5.6
kX ∗ Y k = lim kEn X ∗ Y En k = lim kEn X ∗ Fn Y En k.
n n

Let Hn be the sum of the ranges of En and Fn . This is a finite dimensional subspace,
say dim Hn = d(n). Applying the Cauchy-Schwarz inequality for a symmetric norm on
Md(n) , we obtain, thanks to Lemma 2.5.7,
kX ∗ Y k = lim kEn X ∗ Fn Y En kMd(n)
n
1/2 1/2
≤ lim kEn X ∗ Fn XEn kMd(n) kEn Y ∗ Fn Y En kMd(n)
n

= kX ∗ Xk1/2 kY ∗ Y k1/2 .
Thus the Cauchy-Schwarz inequality for a symmetric norm on B follows from the
Cauchy-Schwarz inequality for symmetric norms on Mn . Of course, the two previ-
ous lemmas and this discussion are rather trivial, but we wanted to stress on the fact
that Theorem 3.4.5 is essentially of finite dimensional nature. However, it would be also
desirable to extend these results in the setting of a semifinite von Neumann algebra.

2.6 Around this article

Let Φ : Mn → Mm be a positive linear map and let N ∈ Mn be normal. Then there
exists a unitary V ∈ Mm such that
Φ(|N|) + V Φ(|N|)V ∗
|Φ(N)| ≤ (2.6.1)
2
and
1
|Φ(N)| ≤ Φ(|N|) + V Φ(|N|)V ∗ .
4
These two inequalities and several consequences are proved in [38], [39]. As an applica-
tion for the Schur product of two normal matrices A, B ∈ Mn , one may infer that
1
|A ◦ B| ≤ |A| ◦ |B| + V (|A| ◦ |B|)V ∗
4
for some unitary V ∈ Mn , where the constant 1/4 is optimal. Another interesting
consequence of (2.6.1) is the following improvement of the Russo-Dye theorem stating
that every positive linear map attains its norm at the identity: if Z ∈ Mn is a contraction,
then
Φ(I) + V Φ(I)V ∗
|Φ(Z)| ≤
2
for some unitary V ∈ Mn . Applying this to the Schur product with S ∈ M+ n yields some
exotic eigenvalue inequalities such as
λ3 (|S ◦ Z|) ≤ δ2 (S)
where λ3 (·) stands for the third largest eigenvalue, and δ2 (·) for the second largest
diagonal entry.
2.7. REFERENCES OF CHAPTER 2 37

2.7 References of Chapter 2

[7] H. Araki, On an inequality of Lieb and Thirring, Let. Math. Phys. 19 (1990 )167-170.

[8] K. Audenaert, On the Araki-Lieb-Thirring Inequality, Int. J. Inf. Syst. Sci. 4 (2008),
78-83.

[9] K. Audenaert and F. Hiai, Reciprocal Lie-Trotter formula, Linear Mult. Algebra, in
press

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York,

1996.

[15] R. Bhatia, Positive Definite Matrices, Princeton University press, Princeton 2007.

[26] J.-C. Bourin, Matrix subadditivity inequalities and block-matrices, Internat. J. Math.
20 (2009), no. 6, 679–691.

[36] J.-C. Bourin and E.-Y. Lee, Matrix inequalities from a two variables functional, Inter-
nat. J. Math. 27 (2016), no. 9, 1650071, 19 pp.

[38] J.-C. Bourin and E.-Y. Lee, Positive linear maps on normal matrices, Internat. J.
Math. 29 (2018), no. 12, 1850088, 11 pp.

[39] J.-C. Bourin and E.-Y. Lee, On the Russo-Dye theorem for positive linear maps. Linear
Algebra Appl. 571 (2019), 92–102.

[54] J.E. Cohen, Spectral inequalities for matrix exponentials, Linear Algebra Appl. 111
(1988) 25-28.

[55] J.E. Cohen, S. Friedland, T. Kate, and F. Kelly, Eigenvalue inequalities for products
of matrix exponentials, Linear Algebra Appl. 45 (1982) 55-95.

[59] D.J.H. Garling, Inequalities - A journey into linear analysis. Cambridge University
Press, Cambridge, 2007.

[66] F. Hiai, A generalization of Araki’s log-majorization. Linear Algebra Appl. 501 (2016),
1–16.

[67] F. Hiai, D. Petz, Introduction to Matrix Analysis and applications. Universitext,

Springer, New Delhi, 2014.

[74] H. Kosaki, Arithmetic-geometric mean and related inequalities for operators. J. Funct.
Anal. 156 (1998), no. 2, 429-451.

[89] B. Simon, Trace ideal and their applications, Cambridge University Press, Cambridge,
1979.
38 CHAPTER 2. MAJORIZATION AND PERSPECTIVE
Chapter 3

Unitary Orbits and Functions

Unitary orbits of Hermitian operators with convex or

concave functions [31]
Abstract. This short but self-contained survey presents a number of elegant ma-
trix/operator inequalities for general convex or concave functions, obtained with a uni-
tary orbit technique. Jensen, sub or super-additivity type inequalities are considered.
Some of them are substitutes to classical inequalities (Choi, Davis, Hansen-Pedersen) for
operator convex or concave functions. Various trace, norm and determinantal inequali-
ties are derived. Combined with an interesting decomposition for positive semi-definite
matrices, several results for partitioned matrices are also obtained.
Keywords: Operator inequalities, positive linear map, trace, unitary orbit, convex function,
symmetric norm, anti-norm.
AMS subjects classification 2010: Primary 15A60, 47A30, 47A60

3.1 Introduction
The functional analytic aspect of Matrix Analysis is evident when matrices or operators
are considered as non-commutative numbers, sequences or functions. In particular, a
significant part of this theory consists in establishing theorems for Hermitian matrices
regarded as generalized real numbers or functions. Two classical trace inequalities may
illustrate quite well this assertion. Given two Hermitian matrices A, B and a concave
function f (t) defined on the real line,
 
A+B f (A) + f (B)
Tr f ≥ Tr (3.1.1)
2 2

and, if further f (0) ≥ 0 and both A and B are positive semi-definite,

Tr f (A + B) ≤ Tr f (A) + Tr f (B). (3.1.2)

The first inequality goes back to von-Neumann in the 1920’s, the second is more subtle
and has been proved only in 1969 by Rotfel’d [88]. These trace inequalities are matrix
versions of obvious scalar inequalities.

39
40 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

The aim of this short survey is to present in a unified and self-contained way two
recent significant improvement of the trace inequalities (3.1.1)-(3.1.2) and some of their
consequences. Our unitary orbit method is also used to prove some basic facts such as
the triangle inequality for Schatten p-norms or Minkowski’s determinantal inequality.
By operator, we mean a linear operator on a finite dimensional Hilbert space. We use
interchangeably the terms operator and matrix. Especially, a positive operator means a
positive (semi-definite) matrix. Consistently Mn denotes the set of operators on a space
of dimension n and M+ n stands for the positive part. As many operator inequalities, our
results lie in the scope of matrix techniques. Of course, there are versions for operators
acting on infinite dimensional, separable Hilbert spaces (and operator algebras); we will
indicate the slight modifications which might then be necessary.
The rest of this introduction explains why inequalities with unitary orbits are rel-
evant for inequalities involving functional calculus of operators such as the concavity-
subadditivity statements (3.1.1) and (3.1.2).
That inequalities with unitary orbits naturally occur can be seen from the following
two elementary facts. Firstly, If A, B ∈ M+ n are such that A ≥ B (that is A − B is
positive semi-definite) then, whenever p > 1, it does not follow in general that Ap ≥ B p .
However, for any non-decreasing function f (t), the eigenvalues (arranged in decreas-
ing order and counted with their multiplicities) of f (A) are greater or equal to the
corresponding ones of f (B). By the min-max characterization of eigenvalues, this is
equivalent to
f (A) ≥ Uf (B)U ∗ (3.1.3)
for some unitary U ∈ Mn . Secondly, if A ∈ M+ n and C ∈ Mn is a contraction, then we
have C AC ≤ UAU for some unitary U ∈ Mn , i.e., the eigenvalues of C ∗ AC are smaller
∗ ∗

or equal to those of A. Note also that C ∗ AC = V A1/2 CC ∗ A1/2 V ∗ ≤ A for some unitary
V , since T T ∗ and T ∗ T are unitarily congruent for any operator T . The reading of this
paper does not require more knowledge about matrices, see [13] for a good background.
The most well-known matrix inequality involving unitary orbits is undoubtedly the
triangle inequality due to Thompson [92]: If X and Y are two operators in Mn , then
|X + Y | ≤ U|X|U ∗ + V |Y |V ∗ (3.1.4)
for some unitary U, V ∈ Mn . Here |X| := (X ∗ X)1/2 is the positive part of X occurring
in the polar decomposition X = V |X| for some unitary V . By letting
 1/2   
A 0 0 0
X= , Y =
0 0 B 1/2 0
√ √ √
where A, B ∈ M+ n , the triangle inequality (3.1.4) yields A + B ≤ K AK ∗ + L BL∗
for some contractions K, L ∈ Mn . Thus, for some unitaries U, V ∈ Mn
√ √ √
A + B ≤ U AU ∗ + V BV ∗ . (3.1.5)
√
This inequality for the function t is a special case of the main theorem of Section 3.
If f (t) is convex on [0, ∞), then (3.1.2) is obviously reversed. In case of f (t) = tp
with exponents p ∈ [1, 2] a much stronger inequality holds,
p
Ap + B p

A+B
≤ , (3.1.6)
2 2
3.2. A MATRIX JENSEN TYPE INEQUALITY 41

this says that tp is operator convex for p ∈ [1, 2], and this is no longer true if p > 2.
However by making use of (3.1.3) and (3.1.6) we get, for any p > 1,
p
Ap + B p ∗

A+B
≤U U , (3.1.7)
2 2

for some unitary U ∈ Mn . In fact, if we assume that (3.1.7) holds for p ∈ [2n , 2n+1 ], n a
positive integer, then it also holds for 2p ∈ [2n+1 , 2n+2 ] since
2p  2 p
A + B2 A2p + B 2p ∗ ∗

A+B
≤ U0 U0∗ ≤ U0 U1 U1 U0
2 2 2
for some unitary U0 , U1 . Inequality (3.1.7) may serve as a motivation for Section 2. It
is worthwhile to notice that, in contrast with the theory of operator convex functions,
our methods are rather elementary.

3.2 A matrix Jensen type inequality

3.2.1 Jensen type inequalities via unitary orbits
In this section we present some extension of (3.1.1). The most general one involves a
unital positive linear map. A linear map Φ : Mn → Md is unital if Φ(I) = I where I
stands for the identity of any order, and Φ is positive if Φ(A) ∈ M+ +
d for all A ∈ Mn .
The simplest case is given when d = 1 by the map

A 7→ hh, Ahi (3.2.1)

for some unit vector h (our inner product is linear in the second variable). Restricting
this map to the diagonal part (more generally, to any commutative ∗-subalgebra) of Mn ,
we have n
X
A 7→ hh, Ahi = wi λi (A) (3.2.2)
i=1

where the λi (A)’s are the eigenvalues of the normal operator A and the wi ’s form a
probability weight. For this reason, unital positive linear maps are regarded as non-
commutative versions of expectations. If A is Hermitian, and f (t) is a convex function
defined on the real line, the Jensen’s inequality may be written in term of the map Φ in
(3.2.1)-(3.2.2) as
f (hh, Ahi) ≤ hh, f (A)hi. (3.2.3)
The map (3.2.1) is a special case of a compression. Given an n-dimensional Hilbert
space H and a d-dimensional subspace S ⊂ H, we have a natural map from the algebra
L(H) of operators on H onto the algebra L(S), the compression map onto S,

A 7→ AS := EA|S , A ∈ L(H),

where E denotes the ortho-projection onto S. Identifying L(H) with Mn by picking an

orthonormal basis of H and L(S) with Md via an orthonormal basis S, we may consider
42 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

compressions as unital positive linear maps acting from Mn onto Md , and they are then
represented as
A 7→ J ∗ AJ, A ∈ Mn ,
where J is any n-by-d matrix such that J ∗ J = I, the identity of order d.
In view of (3.2.3) it is quite natural to compare for a convex function f (AS ) and
f (A)S when A is a Hermitian on H, i.e, a Hermitian in Mn . In this setting, the Jensen
inequality (3.2.3) is adapted by using unitary orbits on S. This is actually true for any
unital positive linear maps, as stated in Theorem 2.1 below. This is the main result of
this section. The notation Mn {Ω} stands for the Hermitian part of Mn with spectra in
an interval Ω of the real line.

Theorem 3.2.1. Let Φ : Mn → Md be a unital positive linear map, let f (t) be a convex
function on an interval Ω, and let A, B ∈ Mn {Ω}. Then, for some unitary U, V ∈ Md ,
UΦ(f (A))U ∗ + V Φ(f (A))V ∗
f (Φ(A)) ≤ .
2
If furthermore f (t) is monotone, then we can take U = V . The inequality reverses for
concave functions.
The next corollaries list some consequences of the theorem. This statement for posi-
tive linear maps contains several Jensen type inequalities. The simplest one is obtained
by taking Φ : M2n → Mn ,  
A X A+B
Φ := .
Y B 2
With X = Y = 0, Theorem 3.2.1 then says:
Corollary 3.2.2. If A, B ∈ Mn {Ω} and f (t) is a convex function on an interval Ω,
then, for some unitaries U, V ∈ Mn ,
   
A+B 1 f (A) + f (B) ∗ f (A) + f (B) ∗
f ≤ U U +V V .
2 2 2 2
If furthermore f (t) is monotone, then we can take U = V .

From this corollary we can get a generalization of the famous Minkowski inequality,

det1/n (A + B) ≥ det1/n A + det1/n B, A, B ∈ M+

n. (3.2.4)

A proof is given after the proof of Corollary 3.2.10 below. Equivalently, (3.2.4) says that
the Minkowski functional X 7→ det1/n X is concave on the positive cone M+ n . Combined
with the concave version of Corollary 3.2.2, this concavity aspect of (3.2.4) is improved
as:
Corollary 3.2.3. If f (t) is a non-negative concave function on an interval Ω and if
A, B ∈ Mn {Ω}, then,

det1/n f (A) + det1/n f (B)

 
1/n A+B
det f ≥ .
2 2
3.2. A MATRIX JENSEN TYPE INEQUALITY 43

Corollary 3.2.2 deals with the simplest convex combination, the arithmetic mean of
two operators. Similar statements holds for weighted means of several operators. In
fact these means may even have operator weightsP(called C ∗ -convex combinations). An
m-tuple {Zi }mi=1 in Mn is an isometric column if m ∗
i=1 Zi Zi = I. We may then perform
∗
P m ∗
the C -convex combination i=1 Zi Ai Z i . If all the Ai ’s are Hermitian operators in
Mn {Ω} for some interval Ω, then so is m ∗
P
i=1 Zi Ai Zi . Hence, Corollary 3.2.2 is a very
special case of the next one.

Corollary 3.2.4. Let {Zi }m m

i=1 be an isometric column in Mn , let {Ai }i=1 be in Mn {Ω}
and let f (t) be a convex function on Ω. Then, for some unitary U, V ∈ Mn ,
m
! ( m
! m
! )
X 1 X X
f Zi∗ Ai Zi ≤ U Zi∗ f (Ai )Zi U ∗ + V Zi∗ f (Ai )Zi V ∗ .
i=1
2 i=1 i=1

If furthermore f (t) is monotone, then we can take U = V . The inequality reverses for
concave functions.

If all the Ai ’s are zero except the first one, we obtain an inequality involving a
congruence Z1∗ A1 Z1 with a contraction Z1 (that is Z1∗ Z1 ≤ I). We state the concave
version in the next corollary. It is a matrix version of the basic inequality f (za) ≥ zf (a)
for a concave function with f (0) ≥ 0 and real numbers z, a with z ∈ [0, 1].

Corollary 3.2.5. Let f (t) be a concave function on an interval Ω with 0 ∈ Ω and

f (0) ≥ 0, let A ∈ Mn {Ω} and let Z be a contraction in Mn . Then, for some unitaries
U, V ∈ Mn ,
U (Z ∗ f (A)Z) U ∗ + V (Z ∗ f (A)Z) V ∗
f (Z ∗ AZ) ≥ .
2
If furthermore f (t) is monotone, then we can take U = V .

For a sub-unital positive linear map Φ, i.e., Φ(I) ≤ I, it is easy to see that Theorem
3.2.1 can be extended in the convex case when f (0) ≤ 0, and in the concave case, when
f (0) ≥ 0 (this sub-unital version is proved in the proof of Corollary 3.2.7 below). This
also contains Corollary 3.2.5. The above results contains some inequalities for various
norms and functionals, as noted in some of the corollaries and remarks below. For
instance we have the following Jensen trace inequalities.

Corollary 3.2.6. Let f (t) be a convex function defined on an interval Ω, let {Ai }m
i=1 be
m
in Mn {Ω}, and let {Zi }i=1 be an isometric column in Mn . Then,
m
! m
X X
Tr f Zi∗ Ai Zi ≤ Tr Zi∗ f (A)i Zi . (3.2.5)
i=1 i=1

If further 0 ∈ Ω and f (0) ≤ 0, we also have

Tr f (Z1∗ A1 Z1 ) ≤ Tr Z1∗ f (A1 )Z1 . (3.2.6)

44 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

A typical example of positive linear map on Mn is the Schur multiplication map

A 7→ Z ◦ A with an operator Z ∈ M+ n . Here Z ◦ A is the entrywise product of A and
Z. The fact that the Schur multiplication with Z ∈ M+ n is a positive linear map can be
easily checked by restricting the Schur product to positive rank ones operators. Hence,
Theorem 3.2.1 contains results for the Schur product. In particular, the sub-unital
version yields:

Corollary 3.2.7. Let f (t) be a concave function on an interval Ω with 0 ∈ Ω and

f (0) ≥ 0, and let A ∈ Mn {Ω}. If Z ∈ M+n has diagonal entries all less than or equal to
1, then, for some unitaries U, V ∈ Mn ,

U (Z ◦ f (A)) U ∗ + V (Z ◦ f (A)) V ∗
f (Z ◦ A) ≥ .
2
If furthermore f (t) is monotone, then we can take U = V .

Proof. Let Ψ : Mn → Md be a positive linear map and suppose that Ψ is sub-unital,

i.e., Ψ(I) = C for some contraction C ∈ M+
d . Then the map Φ : Mn+1 → Md ,

..
" #
A . 7→ Ψ(A) + b(I − C)
... b

is unital. Thus, by Theorem 3.2.1, If A ∈ Mn {Ω} where Ω contains 0 and if f (t) is

concave on Ω,

UΦ(f (A ⊕ 0))U ∗ + V Φ(f (A ⊕ 0))V ∗

f (Φ(A ⊕ 0)) ≥
2
for some unitary U, V ∈ Md , equivalently,

U{Ψ(A) + f (0)(I − C)}U ∗ + V {Ψ(A) + f (0)(I − C)}V ∗

f (Ψ(A)) ≥
2
hence, if further f (0) ≥ 0, the sub-unital form of Theorem 3.2.1 :

UΨ(f (A))U ∗ + V Ψ(f (A))V ∗

f (Ψ(A)) ≥ .
2
Applying this to the sub-unital map Ψ : A 7→ Z ◦ A yields the corollary.

Corollary 3.2.7 obviously contains a trace inequality companion to (3.2.6). By making

use of (3.2.4) we also have the next determinantal inequality.

Corollary 3.2.8. Let f (t) be a non-negative concave function on an interval Ω, 0 ∈ Ω,

and let A ∈ Mn {Ω}. If Z ∈ M+ n has diagonal entries all less than or equal to 1, then,

det f (Z ◦ A) ≥ det Z ◦ f (A). (3.2.7)

3.2. A MATRIX JENSEN TYPE INEQUALITY 45

Some other consequences of Theorem 3.2.1 are given below in Corollary 3.2.10 and
in Subsection 2.2, as well as references and related results.
We turn to the proof of Theorem 3.2.1. Thanks to the next lemma, we will see that it
is enough to prove Theorem 3.2.1 for compressions. By an abelian ∗-subalgebra A of Mm
we mean a subalgebra containing the identity of Mm and closed under the involution
A 7→ A∗ . Any abelian ∗-subalgebra A of Mm is spanned by a total family of ortho-
projections, i.e., a family of mutually orthogonal projections adding up to the identity.
A representation π : A → Mn is a unital linear map such π(A∗ B) = π ∗ (A)π(B).

Lemma 3.2.9. Let Φ be a unital positive map from an abelian ∗-subalgebra A of Mn

to the algebra Mm identified as L(S). Then, there exists a space H ⊃ S, dim H ≤ nm,
and a representation π from A to L(H) such that

Φ(X) = (π(X))S .

Proof. A is generated by a total family of k projections Ei ,Pi = 1, . . . , k (say Ei are rank

one, that is k = n). Let Ai = Φ(Ei ), i = 1, . . . , n. Since ni=1 Ai is the identity on S,
we can find operators Xi,j such that
 1/2 1/2

A1 ... An
 X1,1 ... Xn,1

V =
 
.. .. ..
 . .  .
X1,n−1 . . . Xn,n−1

is a unitary operator on F = ⊕n S. Let Ri be the block matrix with the same i-th
column than V and with all other entries 0. Then, setting Pi = Ri Ri∗ , we obtain a total
family of projections on F satifying Ai = (Pi )S . We define π by π(Ei ) = Pi .

In the following proof of Theorem 3.2.1, and in the rest of the paper, the eigenvalues
of a Hermitian X on an n-dimensional space are denoted in non-increasing order as
λ1 (X) ≥ · · · ≥ λn (X).

Proof. We consider the convex case. We first deal with a compression map. Hence Mn is
identified with L(H) and Φ(A) = AS where S is a subspace of H. We may find spectral
subspaces S ′ and S ′′ for AS and a real r such that

(a) S = S ′ ⊕ S ′′ ,

(b) the spectrum of AS ′ lies on (−∞, r] and the spectrum of AS ′′ lies on [r, ∞),

(c) f is monotone both on (−∞, r] ∩ Ω and [r, ∞) ∩ Ω.

46 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

Let k be an integer, 1 ≤ k ≤ dim S ′ . There exists a spectral subspace F ⊂ S ′ for AS ′

(hence for f (AS ′ )), dim F = k, such that

λk [f (AS ′ )] = min hh, f (AF )hi

h∈F ; khk=1

= min{f (λ1 (AF )) ; f (λk (AF ))}

= min f (hh, AF hi)
h∈F ; khk=1

= min f (hh, Ahi)

h∈F ; khk=1

where at the second and third steps we use the monotony of f on (−∞, r] and the fact
that AF ’s spectrum lies on (−∞, r]. The convexity of f implies

f (hh, Ahi) ≤ hh, f (A)hi

for all normalized vectors h. Therefore, by the minmax principle,

λk [f (AS ′ )] ≤ min hh, f (A)hi

h∈F ; khk=1

≤ λk [f (A)S ′ ].

This statement is equivalent (by unitary congruence to diagonal matrices) to the exis-
tence of a unitary operator U0 on S ′ such that

f (AS ′ ) ≤ U0 f (A)S ′ U0∗ .

(Note that the monotone case is established.) Similarly we get a unitary V0 on S ′′ such
that
f (AS ′′ ) ≤ V0 f (A)S ′′ V0∗ .
Thus we have
   ∗ 
U0 0 f (A)S ′ 0 U0 0
f (AS ) ≤ .
0 V0 0 f (A)S ′′ 0 V0∗

Besides we note that, still in respect with the decomposition S = S ′ ⊕ S ′′ ,

         
f (A)S ′ 0 1 I 0 I 0 I 0 I 0
= f (A)S + f (A)S .
0 f (A)S ′′ 2 0 I 0 I 0 −I 0 −I

So, letting    
U0 0 U0 0
U= and V =
0 V0 0 −V0
we get
Uf (A)S U ∗ + V f (A)S V ∗
f (AS ) ≤ (3.2.8)
2
for some unitary U, V ∈ L(S), with U = V if f (t) is convex and monotone. This proves
the case of compression maps.
3.2. A MATRIX JENSEN TYPE INEQUALITY 47

Next we turn to the case of a general unital linear map Φ : Mn → Mm . Let A be

the abelian ∗-subalgebra of Mn spanned by A. By restricting Φ to A and by identifying
Mm with L(S), Lemma 3.2.9 shows that Φ(X) = (π(X))S for all X ∈ A. Since f
and π commutes, f (π(A)) = π(f (A)), we have from the compression case some unitary
U, V ∈ L(S) = Mm such that,

f (Φ(A)) = f ((π(A))S )
U(f (π(A))S U ∗ + V (f (π(A))S V ∗
≤
2
U(π(f (A))S U ∗ + V (π(f (A))S V ∗
=
2
UΦ(f (A))U ∗ + V Φ(f (A))V ∗
= ,
2
where we can take U = V if the function is convex and monotone.

The following is an application of Theorem 3.2.1 to norm inequalities. A norm k · k

on Mn is a symmetric norm if kAk = kUAV k for all A ∈ Mn and all unitary U, V ∈ Mn .
These norms are also called unitarily invariant norms. They contain the Schatten p-
norms k · kp , 1 ≤ p < ∞, defined as kAkp = {Tr |A|p }1/p . The polar decomposition
shows that a symmetric norm k · k is well defined by its value on the positive cone M+
n.
+
The map on Mn , A 7→ kAk is invariant under unitary congruence and is subadditive.
There are also some interesting, related superadditive functionals. Fix p < 0. The map
X 7→ kX p k1/p is continuous on the invertible part of M+ +
n . If X ∈ Mn is not invertible,
p 1/p +
setting kX k := 0, we obtain a continuous map on Mn .

Corollary 3.2.10. Let A, B ∈ M+

n and let p < 0. Then, for all symmetric norms,

k (A + B)p k1/p ≥ kAp k1/p + kB p k1/p . (3.2.9)

Proof. We will apply Theorem 3.2.1 to the monotone convex function on (0, ∞), t 7→ tp .
First, assume that A, B ∈ M+ p p
n are such that kA k = kB k = 1 and let s ∈ [0, 1]. Then,
thanks to Theorem 3.2.1 (or Corollary 3.2.4),

k(sA + (1 − s)B)p k ≤ ksAp + (1 − s)B p k ≤ skAp k + (1 − s)kB p k = 1,

hence
k(sA + (1 − s)B)p k1/p ≥ 1. (3.2.10)
Now, for general invertible A, B ∈ M+n , insert A/kA k
p 1/p
and B/kB p k1/p in place of
A, B in (3.2.10) and take
kAp k1/p
s= .
kAp k1/p + kB p k1/p
This yields (3.2.9).
48 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

Corollary 3.2.10 implies Minkowski’s determinantal inequality (3.2.4). Indeed, in

1 1/n
(3.2.9) take the norm on M+ n defined by kAk := n Tr A, and note that det A =
limpր0 kAp k1/p . Hence, the superadditivity of A 7→ kAp k1/p for p < 0 entails the super-
additivity of A 7→ det1/n A.
If we apply Theorem 3.2.1 (or Corollary 3.2.2) to the convex function on the real line
t 7→ |t| we obtain: If A, B ∈ Mn are Hermitian, then

U(|A| + |B|)U ∗ + V (|A| + |B|)V ∗

|A + B| ≤
2
for some unitaries U, V ∈ Mn . In fact, we can take U = I and this remains true for
normal operators A, B. This is shown in the proof of the following proposition.

Proposition 3.2.11. If f (t) is a nondecreasing convex function on [0, ∞) and if Z ∈ Mn

has a Cartesian decomposition Z = A + iB, then, for some unitaries U, V ∈ Mn ,

Uf (|A| + |B|)U ∗ + V f (|A| + |B|)V ∗

f (|Z|) ≤ .
2

Proof. let X, Y be two normal operators in Mn . Then, the following operators in M2n
are positive semi-definite,

|X| X ∗ |Y | Y ∗
   
≥ 0, ≥ 0,
X |X| Y |Y |

and consequently
|X| + |Y | X ∗ + Y ∗
 
≥ 0.
X +Y |X| + |Y |
Next, let W be the unitary part in the polar decomposition X + Y = W |X + Y |. Then

|X| + |Y | X ∗ + Y ∗
  
∗ I
I −W ≥ 0,
X +Y |X| + |Y | −W

that is
|X| + |Y | + W ∗ (|X| + |Y |)W − 2|X + Y | ≥ 0.
Equivalently,
|X| + |Y | + W ∗ (|X| + |Y |)W
|X + Y | ≤ . (3.2.11)
2
Letting X = A and Y = iB, and applying f (t) to both sides of (3.2.11), Corollary 2.2
completes the proof since f (t) is nondecreasing and convex.

Proposition 3.2.12. If f (t) is a nondecreasing convex function on [0, ∞) and if A, B ∈

Mn are Hermitian, then, for some unitaries U, V ∈ Mn ,

Uf (A+ + B+ )U ∗ + V f (A+ + B+ )V ∗
f ((A + B)+ ) ≤ .
2
3.2. A MATRIX JENSEN TYPE INEQUALITY 49

Proof. Here A+ := (A + |A|)/2. Note that A + B ≤ A+ + B+ . Let E be the projection

onto ran (A + B)+ and let F be the projection onto ker(A + B)+ Since (A + B)+ =
E(A + B)E, we have

(A + B)+ ≤ E(A+ + B+ )E + F (A+ + B+ )F,

equivalently
(A+ + B+ ) + W (A+ + B+ )W ∗
(A + B)+ ≤ (3.2.12)
2
where W = E − F is a unitary. Applying Corollary 2.2 completes the proof.

3.2.2 Comments and references

In this second part of Section 2, we collect few remarks which complete Theorem 3.2.1
and the above corollaries. Good references for positive maps and operator convex func-
tions are the nice survey and book [64] and [13].

Remark 3.2.13. Theorem 3.2.1 appears in [23]. It is stated therein for compressions
maps and for the case of or ∗-convex combinations given in Corollaries 3.2.4 and 3.2.5
(the monotone case was earlier obtained in [22]). That the compression case immedi-
ately entails the general case of an arbitrary unital positive map is mentioned in some
subsequent papers, for instance in [12] where some inequalities for Schur products are
pointed out. From the Choi-Kraus representation of completely positive linear maps,
readers with a background on positive maps may also notice that Corollary 3.2.5 and
Theorem 3.2.1 are equivalent. For scalar convex combinations and with the assumption
that f (t) is non-decreasing, Theorem 3.2.1 is first noted in Brown-Kosaki’s paper [50];
with these assumptions, it is also obtained in Aujla-Silva’s paper [11].

Remark 3.2.14. Let g(t) denote either the convex function t 7→ |t| or t 7→ t+ . Let
A, B ∈ Mn be Hermitian. Then (2.11) and (2.12) show that
 
A+B g(A) + g(B) g(A) + g(B) ∗
g − ≤V V
2 4 4
for some unitary V ∈ Mn . It would be interesting to characterize convex functions for
which such a relation holds.

Remark 3.2.15. Theorem 3.2.1 holds for operators acting on infinite dimensional spaces,
with an additional rI term. We state here the monotone version. H and S are two
separable Hilbert spaces and r > 0 is fixed. Let Φ : L(H) → L(H) be a unital positive
linear map, let f (t) be a monotone convex function on (−∞, ∞) and let A, B ∈ L(H)
be Hermitian. Then, for some unitary U ∈ L(S),

f (Φ(A)) ≤ UΦ(f (A))U ∗ + rI. (3.2.13)

The proof is given in the first author’s thesis when Φ is a compression map, this entail
the general case. For convenience, the proof is given at the end of this section.
50 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

Remark 3.2.16. The trace inequality (3.2.5) is due to Hansen-Pedersen [63], and the spe-
cial case (3.2.6) is due to Brown-Kosaki. Corollary 3.2.4 considerably improves (3.2.5):
In case of a monotony assumption on the convex function f (t), we have eigenvalue
inequalities; and, in the general case we may still infer the majorization relation
" m
!# " m #
X X
σk f Zi∗ Ai Zi ≤ σk Zi∗ f (A)i Zi , k = 1, . . . , n, (3.2.14)
i=1 i=1

where σk [X] := kj=1 λj [X] is the sum of the k largest eigenvalues of a Hermitian X.
P
In fact, the basic relation σk [X] = max Tr XE, where the maximum runs over all rank
k projections E, shows that σk [·] is convex, increasing on the Hermitian part of Mn so
that (3.2.14) is an immediate consequence of Theorem 3.2.1. The theorem also entails
(see [23] for details) a rather unexpected eigenvalue inequality:
" m
!# m
#
X X
∗ ∗
λ2k−1 f Zi Ai Zi ≤ λk Zi f (A)i Zi , 1 ≤ k ≤ (n + 1)/2.
i=1 i=1

Remark 3.2.17. Choi’s inequality [52] claims: for an operator convex function f (t) on
Ω,
f (Φ(A)) ≤ Φ(f (A)) (3.2.15)
for all A ∈ Mn {Ω} and all unital positive linear map. Thus Theorem 3.2.1 is a substitute
of Choi’s inequality for a general convex function. In the special case of a compression
map, then (3.2.15) is Davis’ inequality [56], a famous characterization of operator con-
vexity. The most well-known case of Davis’ inequality is for the inverse map on positive
definite matrices, it is then an old classical fact of Linear Algebra. Exactly as Theorem
3.2.1 entails Corollary 3.2.4, Choi’s inequality contains Hansen-Pedersen’s inequality
[62], [63]: If f (t) is operator convex on Ω, then
m
! m
X X
∗
f Zi Ai Zi ≤ Zi∗ f (A)i Zi
i=1 i=1

for all unitary columns {Zi }mi=1 in Mn and Ai ∈ Mn {Ω}, i = 1, . . . , m. For operator
concave functions, the inequality reverses. A special case is Hansen’s inequality [61]: if
f (t) is operator concave on Ω, 0 ∈ Ω and f (0) ≥ 0, then

f (Z ∗ AZ) ≥ Z ∗ f (A)Z (3.2.16)

for all A ∈ Mn {Ω} and all contractions Z ∈ Mn .

Remark 3.2.18. Hansen’s inequality (3.2.16) may be formulated with an expansive op-
erator Z ∈ Mn , i.e., Z ∗ Z ≥ I; then (3.2.16) obviously reverses. We might expect that
in a similar way, Corollary 3.2.5 or the Brown-Kosaki trace inequality reverses. But
this does not hold. Corollary 3.2.5 can not reverse when Z is expansive, even under the
monotony assumption on f (t). An unexpected positivity assumption is necessary, and
3.2. A MATRIX JENSEN TYPE INEQUALITY 51

we must confine to weaker inequalities, such as trace inequalities: if f (t) is a concave

function on the positive half-line with f (0) ≥ 0, then,

Tr f (Z ∗ AZ) ≤ Tr Z ∗ f (A)Z

for all A ∈ M+
n and all expansive Z ∈ Mn . For a proof, see [22] and also [24], [30] where
remarkable extensions to norm inequalities are given.

Remark 3.2.19. Lemma 3.2.9 is a part of Stinespring’ s theory of positive and completely
positive linear maps in the influential 1955 paper [90]. The proof given here is somewhat
original and is taken from [12]. Note that in the course of the proof, we prove
P Naimark’s
dilation theorem: If {Ai }ni=1 are positive operators on a space S such that ni=1 Ai ≤ I,
then there exist some mutually orthogonal projections {Pi }ni=1 on a larger space H ⊃ S
such that (Pi )S = Ai , (1 ≤ i ≤ n).

Remark 3.2.20. Given a symmetric norm k · k and p < 0, the functionals defined on M+ n,
A 7→ kAp k1/p , are introduced in [29] and called derived anti-norms. Corollary 3.2.10 is
given therein, [29, Proposition 4.6]. The above proof is much simpler than the original
one. For more details and many results on anti-norms and derived anti-norms, often
in connection with Theorem 3.2.1, see [28] and [29]. Several results in these papers
are generalizations of Corollary 3.2.4. By using (3.1.7) and arguing as in the proof of
Corollary 3.2.10, we may derive the triangle inequality for the Schatten p-norms on M+n,
i.e.,

{Tr (A + B)p }1/p ≤ {Tr Ap }1/p + {Tr B p }1/p , A, B ∈ M+

n , p > 1. (3.2.17)

Remark 3.2.21. The inequality (3.2.11) for normal operators can be extended to general
A, B ∈ Mn , with a similar proof, as
|A| + |B| + V (|A∗ | + |B ∗ |)V ∗
|A + B| ≤ (3.2.18)
2
for some unitary V ∈ Mn . This is pointed out in [48]. This is still true for operators
A, B in a von Neumann algebra M with V a partial isometry in M. If M is endowed
with a regular trace, this gives a short, simple proof of the triangle inequality for the
trace norm on M. Inequality (3.2.11) raises the question of comparison |A + B| and
|A| + |B|. The following result is given in [76]. Let A1 , · · · , Am be invertible operators
with condition numbers dominated by ω > 0. Then
ω+1
|A1 + · · · + Am | ≤ √ (|A1 | + · · · + |Am |).
2 ω
Here the condition number of an invertible operator A on a Hilbert space is kAkkA−1 k−1 .
Note that the bound is independent of the number of operators. Though it is a rather
low bound, it is not known whether it is sharp. Combining (3.2.18) and (3.2.17) we get
the triangle inequality for the Schatten p-norms on the whole space Mn ,

{Tr |A + B|p }1/p ≤ {Tr |A|p }1/p + {Tr |B|p }1/p , A, B ∈ Mn , p > 1.
52 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

It remains to give a proof of the infinite dimensional version (3.2.13) of the monotone
case of Theorem 2.1, the non-monotone case following in a similar way to the finite
dimensional version. As for finite dimensional spaces, we may assume that Φ is a
compression map, thus we consider a subspace S ⊂ H and the map A 7→ AS . By
replacing f (t) by f (−t) and A by −A, we may also assume that f (t) is nondecreasing.
If X is a Hermitian on H, we define a sequence of numbers {λk (X)}∞ k=1 ,

λk (X) = sup inf hh, Xhi

{F : dim F =k} {h∈F : khk=1}

where the supremum runs over k-dimensional subspaces. Note that {λk (X)}∞ k=1 is a
non-increasing sequence whose limit is the upper bound of the essential spectrum of X.
We also define {λ−k (X)}∞
k=1 ,

λ−k (X) = sup inf hh, Xhi.

{F : codim F =k−1} {h∈F : khk=1}

Then, {λ−k (X)}∞ k=1 is a nondecreasing sequence whose limit is the lower bound of the
essential spectrum of X. The following fact (a) is obvious and fact (b) is easily checked.
(a) If X ≤ Y , then λk (X) ≤ λk (Y ) and λ−k (X) ≤ λ−k (Y ) for all k = 1, · · · .
(b) if r > 0 and X, Y are Hermitian, λk (X) ≤ λk (Y ) and λ−k (X) ≤ λ−k (Y ),for all
k = 1, . . . , then X ≤ UY U ∗ + rI for some unitary U.
These facts show that, given r > 0, two Hermitians X, Y with X ≤ Y , and a continuous
nondecreasing function φ, there exists a unitary U such that φ(X) ≤ Uφ(Y )U ∗ + rI.
By fact (b) it suffices to show that
λk (f (AS )) ≤ λk (f (A)S ) (3.2.19)
and
λ−k (f (AS )) ≤ λ−k (f (A)S ) (3.2.20)
for all k = 1, · · · . Now, we prove (3.2.20) and distinguish two cases:
1. λ−k (AS ) is an eigenvalue of AS . Then, for 1 ≤ j ≤ k, λ−j (f (AS )) are eigenvalues for
f (AS ). Consequently, there exists a subspace F ⊂ S, codimS F = k − 1, such that
λ−k (f (AS )) = min hh, f (AS )hi
{h∈F : khk=1}

= min f (hh, AS hi)

{h∈F : khk=1}

≤ inf hh, f (A)hi ≤ λ−k (f (A)S )

{h∈F : khk=1}

where we have used that f is non-decreasing and convex.

2. λ−k (AS ) is not an eigenvalue of AS (so, λ−k (AS ) is the lower bound of the essential
spectrum of AS ). Fix ε > 0 and choose δ > 0 such that |f (x) − f (y)| ≤ ε for all x, y
are in the convex hull of the spectrum of A with |x − y| ≤ δ. There exists a subspace
F ⊂ S, codimS F = k − 1, such that
λ−k (AS ) ≤ inf hh, AS hi + δ.
{h∈F : khk=1}
3.3. A MATRIX SUBADDITIVITY INEQUALITY 53

Since f is continuous nondecreasing we have f (λ−k (AS )) = λ−k (f (AS )) so that, as f is

nondecreasing,  
λ−k (f (AS )) ≤ f inf hh, AS hi + δ .
{h∈F : khk=1}

Consequently,
λ−k (f (AS )) ≤ inf f (hh, AS hi) + ε,
{h∈F : khk=1}

so, using the convexity of f and the definition of λ−k (·), we get

λ−k (f (AS )) ≤ λ−k (f (A)S ) + ε.

By letting ε −→ 0, the proof of (3.2.20) is complete. The proof of (3.2.19) is similar.

Thus (3.2.13) is established.

3.3 A matrix subadditivity inequality

3.3.1 Sub/super-additivity inequalities via unitary orbits
This section deals with some recent subadditive properties for concave functions, and
similarly superadditive properties of convex functions. The main result is:

Theorem 3.3.1. Let f (t) be a monotone concave function on [0, ∞) with f (0) ≥ 0 and
let A, B ∈ M+
n . Then, for some unitaries U, V ∈ Mn ,

f (A + B) ≤ Uf (A)U ∗ + V f (B)V ∗ .

Thus, the obvious scalar inequality f (a + b) ≤ f (a) + f (b) can be extended to positive
matrices A and B by considering element in the unitary orbits of f (A) and f (B). This
inequality via unitary orbits considerably improves the famous Rotfel’d trace inequality
(3.1.2) for a non-negative concave function on the positive half-line, and its symmetric
norm version
kf (A + B)k ≤ kf (A)k + kf (B)k (3.3.1)
for all A, B ∈ M+
n and all symmetric norms k · k on Mn .
Of course Theorem 3.3.1 is equivalent to the next statement for convex functions:

Corollary 3.3.2. Let g(t) be a monotone convex function on [0, ∞) with g(0) ≤ 0 and
let A, B ∈ M+
n . Then, for some unitaries U, V ∈ Mn ,

g(A + B) ≥ Ug(A)U ∗ + V g(B)V ∗ . (3.3.2)

Proof. It suffices to prove the convex version, Corollary 3.3.2. We may confine the proof
to the case g(0) = 0 as if (3.3.2) holds for a function g(t) then it also holds for g(t) − α
for any α > 0. This assumption combined with the monotony of g(t) entails that g(t)
has a constant sign ε ∈ {−1, 1}, hence g(t) = ε|g|(t).
We may also assume that A + B is invertible. Then

A = X(A + B)X ∗ and B = Y (A + B)Y ∗

54 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

where X = A1/2 (A+B)−1/2 and Y = B 1/2 (A+B)−1/2 are contractions. For any T ∈ Mn ,
T ∗ T and T T ∗ are unitarily congruent. Hence, using Corollary 3.2.5 we have two unitary
operators U0 and U such that

g(A) = g(X(A + B)X ∗ )

≤ U0 Xg(A + B)X ∗ U0∗
= εU ∗ (|g|(A + B))1/2 X ∗ X(|g|(A + B))1/2 U,

so,
Ug(A)U ∗ ≤ ε(|g|(A + B))1/2 X ∗ X(|g|(A + B))1/2 . (3.3.3)
Similarly there exists a unitary operator V such that

V g(B)V ∗ ≤ ε(|g|(A + B))1/2 Y ∗ Y (|g|(A + B))1/2 . (3.3.4)

Adding (3.3) and (3.4) we get

Ug(A)U ∗ + V g(B)V ∗ ≤ g(A + B)

since X ∗ X + Y ∗ Y = I.

The following corollary is matrix version of another obvious scalar inequality.

Corollary 3.3.3. Let f : [0, ∞) → [0, ∞) be concave and let A, B ∈ Mn be Hermitian.

Then, for some unitaries U, V ∈ Mn ,

Uf (A)U ∗ − V f (B)V ∗ ≤ f (|A − B|).

Proof. Note that

A ≤ |A − B| + B.
Since f (t) is non-decreasing and concave there exists unitaries W, S, T such that

W f (A)W ∗ ≤ f (|A − B| + B) ≤ Sf (|A − B|)S ∗ + T f (B)T ∗.

Hence, we have
Uf (A)U ∗ − V f (B)V ∗ ≤ f (|A − B|)
for some unitaries U, V.

We can employ Theorem 3.3.1 to get an elegant inequality for positive block-matrices,
 
A X
∈ M+ n+m , A ∈ M+ +
n , B ∈ Mm ,
X∗ B

which nicely extend (3.3.1). To this end we need an interesting decomposition lemma
for elements in M+
n+m .
3.3. A MATRIX SUBADDITIVITY INEQUALITY 55

Lemma 3.3.4. For every matrix in M+

n+m written in blocks, we have a decomposition
     
A X A 0 ∗ 0 0
=U U +V V∗ (3.3.5)
X∗ B 0 0 0 B

for some unitaries U, V ∈ Mn+m .

Proof. To obtain this decomposition of the positive semi-definite block matrix, factorize
it as a square of positive matrices,
    
A X C Y C Y
=
X∗ B Y∗ D Y∗ D

and observe that it can be written as

     
C 0 C Y 0 Y 0 0
+ = T ∗ T + S ∗ S.
Y∗ 0 0 0 0 D Y∗ D

Then, use the fact that T ∗ T and S ∗ S are unitarily congruent to

   
∗ A 0 ∗ 0 0
TT = and SS = ,
0 0 0 B

completing the proof of the decomposition.

Combined with Theorem 3.3.1, the lemma yields a norm inequality for block-matrices.
A symmetric norm on Mn+m induces a symmetric norm on Mn , via kAk = kA ⊕ 0k.

Corollary 3.3.5. Let f (t) be a non-negative concave function on [0, ∞). Then, given
an arbitrary partitioned positive semi-definite matrix,
 
A X
f ≤ kf (A)k + kf (B)k
X∗ B

for all symmetric norms.

Proof. From (3.3.5) and Theorem 3.3.1, we have

     
A X f (A) 0 ∗ f (0)I 0
f =U U +V V∗
X∗ B 0 f (0)I 0 f (B)

for some unitaries U, V ∈ Mn+m . The result then follows from the simple fact that
symmetric norms are nondecreasing functions of the singular values.

Applied to X = A1/2 B 1/2 , this result yields the Rotfel’d type inequalities (3.1.2)-
(3.3.1), indeed,   1/2   1/2
B 1/2
 
A X A 0 A
=
X∗ B B 1/2 0 0 0
56 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

is then unitarily equivalent to (A + B) ⊕ 0. In case of the trace norm, the above result
may be restated as a trace inequality without any non-negative assumption: For all
concave functions f (t) on the positive half-line and all positive block-matrices,
 
A X
Tr f ≤ Tr f (A) + Tr f (B).
X∗ B
The case of f (t) = log t then gives Fisher’s inequality,
 
A X
det ≤ det A det B.
X∗ B
Theorem 3.3.1 may be used to extend another classical (superadditive and concavity)
property of the determinant, the Minkowski inequality (3.2.4). We have the following
extension:
Corollary 3.3.6. If g : [0, ∞) → [0, ∞) is a convex function, g(0) = 0, and A, B ∈ M+
n,
then,
det1/n g(A + B) ≥ det1/n g(A) + det1/n g(B).

As another example of combination of Theorem 3.3.1 and (3.3.5), we have:

Corollary 3.3.7. Let f : [0, ∞) → [0, ∞) be concave and let A = (ai,j ) be a positive
semi-definite matrix in Mn . Then, for some rank one ortho-projections {Ei }ni=1 in Mn ,
n
X
f (A) ≤ f (ai,i )Ei .
i=1

Proof. By a limit argument, we may assume that A is invertible, and hence we may also
assume that f (0) = 0, indeed if the spectrum of A lies in an interval [r, s], r > 0, we
may replace f (t) by any concave function on [0, ∞) such that f˜(0) = 0 and f˜(t) = f (t)
for t ∈ [r, s]. By a repetition of (3.3.5) we have
n
X
A= ai,i Fi
i=1

for some rank one ortho-projections {Fi }ni=1 in Mn . An application of Theorem 3.3.1
yields
Xn
f (A) ≤ Ui f (ai,i Fi )Ui∗
i=1

for some unitary operators {Ui }ni=1 . Since f (0) = 0, for each i, Ui f (ai,i Fi )Ui∗ = f (ai,i )Ei
for some rank one projection Ei .

Corollary 3.3.7 refines the standard majorization inequality relating a positive semi-
definite n-by-n matrix and its diagonal part,
d
X
Tr f (A) ≤ f (ai,i ).
i=1
3.3. A MATRIX SUBADDITIVITY INEQUALITY 57

3.3.2 Comments and references

Remark 3.3.8. Theorem 3.3.1, Corollaries 3.3.2 and 3.3.2 are from [12]. In case of
positive operators acting on an infinite dimensional, separable Hilbert space, we have a
version of Theorem 3.3.1 with an additional rI term in the RHS, as in (2.13).

Remark 3.3.9. The decomposition of a positive block-matrix in Lemma 3.3.4 is due to

the authors. It is used in [75] to obtain the norm inequality stated in Corollary 3.3.5.
The next two Corollaries 3.3.6 and 3.3.7 are new, though already announced in [28].

Remark 3.3.10. The concavity requirement on f (t) in Rotfel’d inequality (3.1.2) and
hence in Theorem 3.3.1 cannot be relaxed to a mere superadditivity assumption; indeed
take for s, t > 0,
 √   √ 
1 s st 1 s − st
A= √ , B= √ ,
2 st t 2 − st t

and observe that the trace inequality Tr f (A + B) ≤ Tr f (A) + f (B) combined with
f (0) = 0 means that f (t) is concave.

Remark 3.3.11. There exists a norm version of Rotfel’d inequality which considerably
improves (3.3.1). If f : [0, ∞) → [0, ∞) is concave and A, B ∈ M+
n , then

kf (A + B)k ≤ kf (A) + f (B)k

for all symmetric norms. The case of operator concave functions is given in [6] and the
general case is established in [49], see also [30] for further results. Concerning differences,
the following inequality holds

kf (A) − f (B)k ≤ kf (|A − B|)k

for all symmetric norms, A, B ∈ M+ n , and operator monotone functions f : [0, ∞) →

[0, ∞). This is a famous result of Ando [4]. Here the operator monotonicity assumption
is essential, see [10] for some counterexamples. A very interesting paper by Mathias
[79] gives a direct proof, without using the integral representation of operator monotone
functions.

Remark 3.3.12. There exists alsoPsome subaditivity results involving convex functions
m k
[28]. For instance: Let g(t) = k=0 ak t be a polynomial of degree m with all non-
negative coefficients. Then, for all positive operators A, B and all symmetric norms,

kg(A + B)k1/m ≤ kg(A)k1/m + kg(B)k1/m .

Remark 3.3.13. It is not known wether the monotonicity assumption in Theorem 3.1
can be deleted, i.e., wether the theorem holds for all concave functions f (t) on [0, ∞)
with f (0) ≥ 0.
58 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

3.4 Around this article

We will see several applications of this article in the next chapters. Here we recall
some results from [40] improving the Clarkson-McCarthy inequalities for p ≥ 2. These
inequalities show that the unit ball for the Schatten p-norm is uniformly convex,
p p
A+B A−B kAkpp + kBkpp
+ ≤ , p ≥ 2,
2 p 2 p 2

i.e.,
p p
A+B A−B Tr |A|p + Tr |B|p
Tr + Tr ≤ , p ≥ 2. (3.4.1)
2 2 2
Thus, if kAkp = kBkp = 1 and kA − Bkp = ε,
A+B
≤ (1 − (ε/2)p )1/p , p ≥ 2,
2 p

which estimates the uniform convexity modulus of the Schatten p-classes for p ≥ 2.
Some nice extensions of these inequalities have been given by Bhatia-Holbrook [16]
and Hirzallah-Kittaneh [69]. The next series of corollaries provide several eigenvalue
inequalities completing those of Bhatia-Holbrook and Hirzallah-Kittaneh. First, we
state the following improvement of (3.4.1).
Theorem 3.4.1. Let A, B ∈ Mn and p > 2. Then there exists two unitarie U, V ∈ Mn
such that p p
A+B A−B |A|p + |B|p
U U∗ + V V∗ ≤ .
2 2 2

Proof. Note that

p p/2
|A|2 + |B|2 + A∗ B + B ∗ A

A+B
=
2 4
and p/2
p
|A|2 + |B|2 − (A∗ B + B ∗ A)

A−B
= .
2 4
Now, recall Corollary 3.3.2 : Given two positive matrices X, Y and a monotone convex
function g(t) defined on [0, ∞) such that g(0) ≤ 0, we have

g(X + Y ) ≥ U0 g(X)U0∗ + V0 g(Y )V0∗ (3.4.2)

for some pair of unitary matrices U0 and V0 . Applying this to g(t) = tp/2 ,
|A|2 + |B|2 + A∗ B + B ∗ A
X=
4
and
|A|2 + |B|2 − (A∗ B + B ∗ A)
Y = ,
4
3.4. AROUND THIS ARTICLE 59

we obtain p/2 p p
|A|2 + |B|2

A+B A−B
≥ U0 U0∗ + V0 V0∗ . (3.4.3)
2 2 2
Next, recall Corollary 3.2.2 : Given two positive matrices X, Y and a monotone convex
function g(t) defined on [0, ∞) , we have
 
g(X) + g(Y ) X +Y
≥ Wg W∗ (3.4.4)
2 2

for some unitary matrix W . Applying this to g(t) = tp/2 , X = |A|2 and Y = |B|2 , we
get
p/2
|A|p + |B|p
 2
|A| + |B|2
≥W W ∗. (3.4.5)
2 2
Combining (3.4.3) and (3.4.5) completes the proof with U = W U0 and V = W V0 .

Corollary 3.4.2. Let A, B ∈ Mn and p > 2. Then, for all k = 1, 2, . . . , n,

k k p k p
|A|p + |B|p
   
X X A+B X A−B
λ↑j ≥ λ↑j + λ↑j .
j=1
2 j=1
2 j=1
2

Corollary 3.4.3. Let A, B ∈ Mn and p > 2. Then, for all k = 1, 2, . . . , n,

( k )1/k ( k )1/k ( k )1/k
Y ↑  |A|p + |B|p  Y ↑  A + B p Y ↑  A − B p
λj ≥ λj + λj .
j=1
2 j=1
2 j=1
2

Here λ↑1 (X) ≤ λ↑2 (X) ≤ · · · ≤ λ↑n (X) stand for the eigenvalues of X ∈ M+
n arranged
in the nondecresaing order. These two corollaries follow from Theorem 3.4.1 and the
fact that the functionals on M+ n

k
X
X 7→ λ↑j (X)
j=1

and ( k )1/k
Y
X 7→ λ↑j (X)
j=1

are two basic examples of symmetric anti-norms, see [28], [29].

The next corollary follows from Theorem 3.4.1 combined with a classical inequality
of Weyl for the eigenvalues of the sum of two Hermitian matrices.
Corollary 3.4.4. Let A, B ∈ Mn and p > 2. Then, for all j, k ∈ {0, . . . , n − 1} such
that j + k + 1 ≤ n,
 p p p
|A| + |B|p
  
↓ ↓ A+B ↑ A−B
λj+1 ≥ λj+k+1 + λk+1 .
2 2 2
60 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS

For a monotone concave function g(t) defined on [0, ∞) such that g(0) ≥ 0, the
inequalities (3.4.2) and (3.4.4) are reversed. Applying this to g(t) = tq/2 , 2 > q > 0, the
same proof than that of Theorem 3.4.1 gives the following statement.

Theorem 3.4.5. Let A, B ∈ Mn and 2 > q > 0. Then, for some unitaries U, V ∈ Mn ,
q q
A+B ∗ A−B ∗ |A|q + |B|q
U U +V V ≥ .
2 2 2

By using Weyl’s inequality, this theorem yields an interesting eigenvalue estimate.

Corollary 3.4.6. Let A, B ∈ Mn and 2 > q > 0. Then, for all j, k ∈ {0, . . . , n − 1}
such that j + k + 1 ≤ n,
 q q q
|A| + |B|q
  
↓ ↓ A+B ↓ A−B
λj+k+1 ≤ λj+1 + λk+1 .
2 2 2

3.5 References of Chapter 3

[4] T. Ando, Comparison of norms |||f (A) − f (B)||| and kf (|A − B|)|||. Math. Z. 197
(1988), no. 3, 403-409.

[10] K. Audenaert and J. S. Aujla, On Ando’s inequalities for convex and concave functions,
arXiv:0704.0099v1.

[11] J. S. Aujla and F. C. Silva, Weak majorization inequalities and convex functions, Linear
Algebra Appl. 369 (2003), 217-233.

[12] J.S. Aujla and J.-C. Bourin, Eigenvalue inequalities for convex and log-convex func-
tions, Linear Algebra Appl. 424 (2007), 25–35.

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York,

1996.

[16] R. Bhatia and J. Holbrook, On the Clarkson-McCarthy inequalities, Math. Ann. 281
(1988), no. 1, 7–12.

[22] J.-C. Bourin, Convexity or concavity inequalities for Hermitian operators. Math. In-
equal. Appl. 7 (2004), no. 4, 607–620.

[23] J.-C. Bourin, Hermitian operators and convex functions, J. Inequal. Pure Appl. Math.
6 (2005), Article 139, 6 pp.

[24] J.-C. Bourin, A concavity inequality for symmetric norms, Linear Algebra Appl. 413
(2006), 212-217.

[28] J.-C. Bourin and F. Hiai, Norm and anti-norm inequalities for positive semi-definite
matrices, Internat. J. Math. 63 (2011), 1121-1138.

[29] J.-C. Bourin and F. Hiai, Jensen and Minkowski inequalities for operator means and
anti-norms. Linear Algebra Appl. 456 (2014), 22–53.
3.5. REFERENCES OF CHAPTER 3 61

[30] J.-C. Bourin and E.-Y. Lee, Concave functions of positive operators, sums, and con-
gruences, J. Operator Theory 63 (2010), 151–157.

[31] J.-C. Bourin and E.-Y. Lee, Unitary orbits of Hermitian operators with convex or
concave functions, Bull. Lond. Math. Soc. 44 (2012), no. 6, 1085–1102.

[40] J.-C. Bourin and E.-Y. Lee, Clarkson-McCarthy inequalities with unitary and isometry
orbits, Linear Algebra Appl. 601 (2020), 170–179.

[48] J.-C. Bourin and E. Ricard, An asymmetric Kadison’s inequality, Linear Algebra Appl.
433 (2010) 499–510.

[49] J.-C. Bourin and M. Uchiyama, A matrix subadditivity inequality for f (A + B) and
f (A) + f (B), Linear Algebra Appl. 423 (2007), 512–518.

[50] L. G. Brown and H. Kosaki, Jensen’s inequality in semi-finite von Neuman algebras,
J. Operator theory 23 (1990), 3–19.

[52] M.-D. Choi, A Schwarz inequality for positive linear maps on C ∗ -algebras, Illinois J.
Math. 18 (1974), 565–574.

[56] C. Davis, A Schwarz inequality for convex operator functions, Proc. Amer. Math. Soc.
8 (1957), 42-44.

[61] F. Hansen, An operator inequality, Math. Ann. 246 (1979/80), no. 3, 249–250.

[62] F. Hansen and G. K. Pedersen, Jensen’s inequality for operators and Löwner’s theorem,
Math. Ann. 258 (1982), 229–241.

[63] F. Hansen and G. K. Pedersen, Jensen’s operator inequality, Bull. London Math. Soc.
35 (2003), no. 4, 553–564.

[64] F. Hiai, Matrix Analysis: Matrix Monotone Functions, Matrix Means, and Majoriza-
tion (GSIS selected lectures), Interdisciplinary Information Sciences 16 (2010), 139–248.

[69] O. Hirzallah and F. Kittaneh, Non-commutative Clarkson inequalities for unitarily

invariant norms, Pacific J. Math. 202 (2002), no. 2, 363–369.

[75] E.-Y. Lee, Extension of Rotfel’d Theorem, Linear Algebra Appl. 435 (2010), 735–741.

[76] E.-Y. Lee, How to compare the absolute values of operator sums and the sums of
absolute values ?, to appear in Operator and Matrices.

[79] R. Mathias, Concavity of monotone matrix functions of finite order, Linear and Mul-
tilinear Algebra 27 (1990), no. 2, 129-138.

[88] S. Ju. Rotfel’d, The singular values of a sum of completely continuous operators,
Topics in Mathematical Physics, Consultants Bureau, Vol. 3 (1969) 73-78.

[90] W. F. Stinespring, Positive functions on C ∗ -algebras, Proc. Amer. Math. Soc. 6, (1955).
211-216.

[92] R.-C. Thompson, Convex and concave functions of singular values of matrix sums,
Pacific J. Math. 66 (1976), 285–290.
62 CHAPTER 3. UNITARY ORBITS AND FUNCTIONS
Chapter 4

Around Hermite-Hadamard

Matrix inequalities and majorizations around Hermite-

Hadamard’s inequality [43]
Abstract. We study the classical Hermite-Hadamard inequality in the matrix setting. This
leads to a number of interesting matrix inequalities such as the Schatten p-norm estimates

1/p
kAq kpp + kB q kpp ≤ k(xA + (1 − x)B))q kp + k(1 − x)A + xB)q kp

for all positive (semidefinite) n × n matrices A, B and 0 < q, x < 1. A related decomposition,
with the assumption X ∗ X + Y ∗ Y = XX ∗ + Y Y ∗ = I, is
2n
1 X
(X ∗ AX + Y ∗ BY ) ⊕ (Y ∗ AY + X ∗ BX) = Uk (A ⊕ B)Uk∗
2n
k=1

for some family of 2n × 2n unitary matrices Uk . This is a majorization which is obtained by

using the Hansen-Pedersen trace inequality.
Keywords. Positive definite matrices, block matrices, convex functions, matrix inequalities.
2010 mathematics subject classification. 15A18, 15A60, 47A30.

4.1 Elementary scalar inequalities

Extending basic scalar inequalities, for instance |a + b| ≤ |a| + |b|, to matrices lies at the very
heart of matrix analysis. Here, we are interested in the elementary inequality which supports
the Hermite-Hadamard inequality. This classical theorem can be stated as follows:
Proposition 4.1.1. Let f (t) be a convex function defined on the interval [a, b]. Then,
  Z 1
a+b f (a) + f (b)
f ≤ f ((1 − x)a + xb) dx ≤ .
2 0 2
In spite of its simplicity, the Hermite-Hadamard inequality is a powerful tool for deriving
a number of important inequalities; see the nice paper [83] and references therein.
The first inequality immediately follows from the convexity assumption
 
a+b f ((1 − x)a + xb) + f (xa + (1 − x)b)
f ≤ (4.1.1)
2 2

63
64 CHAPTER 4. AROUND HERMITE-HADAMARD

The second inequality is slightly more subtle; it follows from the extremal property

f ((1 − x)a + xb) + f (xa + (1 − x)b) ≤ f (a) + f (b) (4.1.2)

which requires the convexity assumption twice. This is the key for Proposition 4.1.1 and it
has a clear geometric interpretation; (4.1.2) is equivalent to the increasingness of

ϕ(t) := f (m + t) + f (m − t)

with m = (a + b)/2 and t ∈ [0, b − m]. In fact, if we assume that f (t) is C 2 and observe that
for t ∈ [0, b − m], Z m+t
′ ′ ′
ϕ (t) = f (m + t) − f (m − t) = f ′′ (s) ds,
m−t

we can estimate ϕ′ (t) with f ′′ (s) ≥ 0.

The extremal property (4.1.2) of f (t) says that for four points in [a, b],

p ≤ s ≤ t ≤ q, p + q = s + t ⇒ f (s) + f (t) ≤ f (p) + f (q). (4.1.3)

Let us see now what can be said for matrices. Important matrix versions of (4.1.1) are
well-known. Let Mn denote the space of n × n matrices and Ms.a
n its self-adjoint (Hermitian)
part with the usual order ≤ induced by the positive semidefinite cone M+ n . We recall [31,
Corollary 2.2].

Theorem 4.1.2. Let A, B ∈ Ms.a n with spectra in [a, b] and let f (t) be a convex function on
[a, b]. Then, for some unitaries U, V ∈ Mn ,
   
A+B 1 f (A) + f (B) ∗ f (A) + f (B) ∗
f ≤ U U +V V .
2 2 2 2

If furthermore f (t) is monotone, then we can take U = V .

Theorem 4.1.2 is a major improvement of the classical trace inequality of von Neumann
(around 1920),  
A+B f (A) + f (B)
Tr f ≤ Tr
2 2
which entails the following trivial extension of Proposition 4.1.1.

Proposition 4.1.3. Let f (t) be a convex function defined on the interval [a, b] and let A, B ∈
Ms.a
n with spectra in [a, b]. Then,
  Z 1
A+B f (A) + f (B)
Tr f ≤ Tr f ((1 − x)A + xB) dx ≤ Tr .
2 0 2

What about matrix versions of the equivalent scalar inequalities (4.1.2) and (4.1.3)? There
is no hope for (4.1.3) : in general the trace inequality

Tr f (P ) + Tr f (Q) ≤ Tr f (S) + Tr f (T )

does not hold for all Hermitian matrices P, Q, S, T with spectra in [a, b] and such that

P ≤ S ≤ T ≤ Q, P + Q = S + T.
4.2. THE EXTREMAL PROPERTY FOR MATRICES 65

In the matrix setting (4.1.2) and (4.1.3) are not equivalent. This paper aims to establish
two matrix versions of the extremal inequality (4.1.2). Doing so, we will obtain several new
matrix inequalities.
For an operator convex functions h(t) on [a, b], and A, B ∈ Ms.a
n with spectra in this interval,
the matrix version of (4.1.2) (as well as Proposition 4.1.1) obvioulsy holds,

h((1 − x)A + xB) + h(xA + (1 − x)B) ≤ h(A) + h(B) (4.1.4)

for any 0 < x < 1. Our results hold for much more general convex/concave functions and have
applications to eigenvalue inequalities that cannot be derived from (4.1.4), even in the case of
the simplest operator convex/concave function h(t) = t.
We often use a crucial assumption: our functions are defined on the positive half-line and
we deal with positive semidefinite matrices. In the matrix setting, the interval of definition of
a function may be quite important; for instance the class of operator monotone functions on
the whole real line reduces to affine functions.

4.2 The extremal property for matrices

For concave functions, the inequality (4.1.2) is reversed. Here is the matrix version. An
isometry U ∈ M2n,n means a 2n × n matrix such that U ∗ U = I, the identity of Mn .

Theorem 4.2.1. Let A, B ∈ M+ n , let 0 < x < 1, and let f (t) be a monotone concave function
on [0, ∞) with f (0) ≥ 0. Then, for some isometry matrices U, V ∈ M2n,n ,

f (A) ⊕ f (B) ≤ U f ((1 − x)A + xB)U ∗ + V f (xA + (1 − x)B)V ∗ .

Proof. Let x = sin2 θ, 1 − x = cos2 θ, consider the unitary Hermitian matrix

√ √ 
1 − xI √ xI
R := √
xI − 1 − xI

and note the unitary congruence

   
A 0 (1 − x)A + xB ⋆
R R= (4.2.1)
0 B ⋆ xA + (1 − x)B

where the stars hold for unspecified entries.

 Now, recall the decomposition [31, Lemma 3.4] : Given any positive semidefinite matrix
C X
partitioned in four blocks in Mn , we have
X∗ D
     
C X C 0 ∗ 0 0
= U0 U + V0 V∗
X∗ D 0 0 0 0 D 0

for some unitary matrices U0 , V0 ∈ M2n . Applying this to (4.2.1), we obtain

     
A 0 (1 − x)A + xB 0 ∗ 0 0
= U1 U1 + V1 V∗ (4.2.2)
0 B 0 0 0 xA + (1 − x)B 1

for two unitary matrices U1 , V1 ∈ M2n .

66 CHAPTER 4. AROUND HERMITE-HADAMARD

Next, recall the subadditivity inequality [31, Theorem 3.4] : Given any pair of positive
semidefinite matrices S, T ∈ Md , we have

f (S + T ) ≤ U2 f (S)U2∗ + V2 f (T )V2∗

for two unitary matrices U2 , V2 ∈ Md . Applying this to (4.2.2) yields

     
f (A) 0 f ((1 − x)A + xB) 0 ∗ f (0)I 0
≤U U +V V∗
0 f (B) 0 f (0)I 0 f (xA + (1 − x)B)

for two unitary matrices U, V ∈ M2n . This proves the theorem when f (0) = 0.
To derive the general case, we may assume that f (t) is continuous (a concave function on
[0, ∞) might be discontinuous at 0). Indeed, it suffices to consider the values of f (t) on a finite
set, the union of the spectra of the four matrices A, B, (1−x)A+xB and xA+(1−x)B. Hence
we may replace f (t) by a piecewise affine monotone concave function h(t) with h(0) ≥ 0. Now,
since h(t) is continuous, a limit argument allows us to suppose that A and B are invertible.
Therefore, letting λ↓n (Z) denote the smallest eigenvalue of Z ∈ Ms.a
n ,

r := min{λ↓n (A), λ↓n (B)} > 0.

We may then replace h(t) by hr (t) defined as hr (t) := h(t) for t ≥ r, hr (0) := 0 and hr (s) :=
h(r) rs for 0 ≤ s ≤ r. The function hr (t) is monotone concave on [0, ∞) and vanishes at 0, thus
the case f (0) = 0 entails the general case.

Let λ↓j (Z), j = 1, 2, . . . n, denote the eigenvalues of Z ∈ Ms.a

n arranged in the nonincreasing
order.

Corollary 4.2.2. Let A, B ∈ M+ n , let 0 < x < 1, and let f (t) be a nonnegative concave
function on [0, ∞). Then, for j = 0, 1, . . . , n − 1,

λ↓1+2j (f (A ⊕ B)) ≤ λ↓1+j (f (xA + (1 − x)B)) + λ↓1+j (f ((1 − x)A + xB))

and   
A+B
λ↓1+j {(f (xA + (1 − x)B)) + (f ((1 − x)A + xB))} ≤ 2λ↓1+j f .
2

Proof. The first inequality is a straighforward consequence of Theorem 4.2.1 combined with
the inequalities of Weyl [13, p. 62] : For all S, T ∈ Ms.ad and j, k ∈ {0, . . . , d − 1} such that
j + k + 1 ≤ d,
λ↓1+j+k (S + T ) ≤ λ↓1+j (S) + λ↓1+k (T ).
The second inequality is not new; it follows from Theorem 8.5.1.

Corollary 4.2.3. Let A, B ∈ M+ n , let 0 < x < 1, and let f (t) be nonnegative concave function
on [0, ∞). Then, for all p ≥ 1,

1/p
kf (A)kpp + kf (B)kpp ≤ kf (xA + (1 − x)B))kp + kf ((1 − x)A + xB)kp .
4.2. THE EXTREMAL PROPERTY FOR MATRICES 67

Proof. From Theorem 4.2.1 we have

kf (A) ⊕ f (B)kp ≤ kU f ((1 − x)A + xB)U ∗ + V f (xA + (1 − x)B)V ∗ kp .

The triangle inequality for k · kp completes the proof.

Corollary 4.2.3 with f (t) = tq reads as the following trace inequality.

Corollary 4.2.4. Let A, B ∈ M+

n and 0 < x < 1. Then, for all p ≥ 1 ≥ q ≥ 0,

{Tr Apq + Tr B pq }1/p ≤ {Tr (xA + (1 − x)B)pq }1/p + {Tr ((1 − x)A + xB))pq }1/p .

Choosing in Corollary 4.2.4 q = 1 and x = 1/2 yields McCarthy’s inequality,

Tr Ap + Tr B p ≤ Tr (A + B)p .

This shows that Theorem 4.2.1 is already significant with f (t) = t. Our next corollary, for
convex functions, is equivalent to Theorem 4.2.1.

Corollary 4.2.5. Let A, B ∈ M+ n , let 0 < x < 1, and let g(t) be a monotone convex function
on [0, ∞) with g(0) ≤ 0. Then, for some isometry matrices U, V ∈ M2n,n ,

g(A) ⊕ g(B) ≥ U g((1 − x)A + xB)U ∗ + V g(xA + (1 − x)B)V ∗ .

Since the Schatten q-quasinorms k · kq , 0 < q < 1, are superadditive functionals on M+

n,
Corollary 4.2.5 yields the next one.

Corollary 4.2.6. Let A, B ∈ M+ n , let 0 < x < 1, and let g(t) be a nonnegative convex function
on [0, ∞) with g(0) ≤ 0. Then, for all 0 < q < 1,

1/q
kg(A)kqq + kg(B)kqq ≥ kg(xA + (1 − x)B))kq + kg((1 − x)A + xB)kq .

From the first inequality of Corollary 4.2.2 we also get the following statement.

Corollary 4.2.7. Let A, B ∈ M+ n and let f (t) be a nonnegative concave function on [0, ∞).
Then, for j = 0, 1, . . . , n − 1,
Z 1
λ↓1+2j (f (A ⊕ B)) ≤ 2 λ↓1+j (f (xA + (1 − x)B)) dx.
0

Up to now we have dealt with convex combinations (1 − x)A + xB with scalar weights. It
is natural to search for extensions with matricial weights (C ∗ -convex combinations). We may
generalize Theorem 4.2.1 with commuting normal weights.

Theorem 4.2.8. Let A, B ∈ M+ n and let f (t) be a monotone concave function on [0, ∞) with
f (0) ≥ 0. If X, Y ∈ Mn are normal and satisfy XY = Y X and X ∗ X + Y ∗ Y = I, then, for
some isometry matrices U, V ∈ M2n,n ,

f (A) ⊕ f (B) ≤ U f (X ∗ AX + Y ∗ BY )U ∗ + V f (Y ∗ AY + X ∗ BX)V ∗ .

68 CHAPTER 4. AROUND HERMITE-HADAMARD

Proof. The proof is quite similar to that of Theorem 4.2.1 except that we first observe that
the 2n × 2n matrix  
X Y
H :=
Y −X
is unitary. Indeed for two normal operators, XY = Y X ensures X ∗ Y = Y X ∗ and a direct
computation shows that H ∗ H is the identity in M2n . We then use the unitary congruence
 ∗
X AX + Y ∗ BY
  
∗ A 0 ⋆
H H= (4.2.3)
0 B ⋆ Y ∗ AY + X ∗ BX
where the stars hold for unspecified entries.

Hence, in the first inequality of Corolloray 4.2.2 and in the series of Corollaries 4.2.3-4.2.6,
we can replace the scalar convex combinations (1 − x)A + B and xA + (1 − x)B by C ∗ -convex
combinations X ∗ AX + Y ∗ BY and Y ∗ AY + X ∗ BX with commuting normal weights. Here we
explicitly state the generalization of Corollary 4.2.5.
Corollary 4.2.9. Let A, B ∈ M+n and let g(t) be a monotone convex function on [0, ∞) with
g(0) ≤ 0. If X, Y ∈ Mn are normal and satisfy XY = Y X and X ∗ X + Y ∗ Y = I, then, for
some isometry matrices U, V ∈ M2n,n ,

g(A) ⊕ g(B) ≥ U g(X ∗ AX + Y ∗ BY )U ∗ + V g(Y ∗ AY + X ∗ BX)V ∗ .

4.3 Majorization
The results of Section 2 require two essential assumptions : to deal with positive matrices and
with subadditive (concave) or superadditive (convex) functions. Thanks to these assumptions,
we have obtained operator inequalities for the usual order in the positive cone.
The results of this section will consider Hermitian matrices and general convex or concave
functions. We will obtain majorization relations. We also consider C ∗ -convex combinations
more general than those with commuting normal weights.
We recall the notion of majorization. Let A, B ∈ Ms.a n . We say that A is weakly majorized
by B and we write A ≺w B, if
k k
λ↓j (A) ≤ λ↓j (B)
X X

j=1 j=1

for all k = 1, 2, . . . n. If furthemore the equality holds for k = n, that is A and B have the
same trace, then we say that A is majorized by B, written A ≺ B. See [13, Chapter 2] and [64]
for a background on majorization. One easily checks that A ≺w B is equivalent to A + C ≺ B
for some C ∈ M+ n . We need two fundamental principles:

(1) A ≺ B ⇒ g(A) ≺w g(B) for all convex functions g(t).

(2) A ≺w B ⇐⇒ Tr f (A) ≤ Tr f (B) for all nondecreasing convex functions f (t). Equiva-
lently A ≺ B ⇐⇒ Tr g(A) ≤ Tr g(B) for all convex functions g(t).

Lemma 4.3.1. Let A, B ∈ Ms.a n and let g(t) be a convex function defined on an interval
containing the spectra of A and B. If X, Y ∈ Mn satisfy X ∗ X + Y ∗ Y = XX ∗ + Y Y ∗ = I,
then,
Tr {g(X ∗ AX + Y ∗ BY ) + g(Y ∗ AY + X ∗ BX)} ≤ Tr {g(A) + g(B)} .
4.3. MAJORIZATION 69

Proof. By the famous Hansen-Pedersen trace inequality [67], see also [31, Corollary 2.4] for a
generalization,

Tr {g(X ∗ AX + Y ∗ BY ) + g(Y ∗ AY + X ∗ BX)}

≤ Tr {X ∗ g(A)X + Y ∗ g(B)Y ) + Y ∗ g(A)Y + X ∗ g(B)X}
= Tr {(g(A) + g(B))(XX ∗ + Y Y ∗ )} = Tr {g(A) + g(B)}

where the first equality follows from the cyclicity of the trace.

Theorem 4.3.2. Let A, B ∈ Ms.a ∗ ∗ ∗ ∗

n and X, Y ∈ Mn . If X X + Y Y = XX + Y Y = I, then,
2n
for some unitary matrices {Uk }k=1 in M2n ,

2n
1 X
(X ∗ AX + Y ∗ BY ) ⊕ (Y ∗ AY + X ∗ BX) = Uk (A ⊕ B)Uk∗ .
2n
k=1

Proof. From Lemma 4.3.1, we have the trace inequality

Tr g ((X ∗ AX + Y ∗ BY ) ⊕ (Y ∗ AY + X ∗ BX)) ≤ Tr g(A ⊕ B)

for all convex functions defined on (−∞, ∞). By a basic principle of majorization, this is
equivalent to
(X ∗ AX + Y ∗ BY ) ⊕ (Y ∗ AY + X ∗ BX) ≺ A ⊕ B. (4.3.1)
By [42, Proposition 2.6], the majorization in Ms.a
d , S ≺ T , ensures that (and thus is equivalent
to)
d
1X
S≤ Vj T Vj∗
d
j=1

for d unitary matrices Vj ∈ Md . Applying this to (4.3.1) completes the proof.

Remark 4.3.3. We can prove the majorization (4.3.1) in a different way by oberving that our
assumption on X and Y ensures that the map Φ, defined on Md , (here d = 2n),

X ∗ AX + Y ∗ BY
   
A C 0
Φ := ,
D B 0 Y ∗ AY + X ∗ BX)

is a positive linear map unital and trace preserving. Such maps are also called doubly stochas-
tic. It is a classical result (see Ando’s survey [5, Section 7] and references therein) that we
have
Φ(Z) ≺ Z (4.3.2)
for every doubly stochastic map on Md and Hermitian Z. Here, our map is even completely
positive (it is a so called quantum channel). It is well known in the litterature ([5, Theorem
7.1]) that we have then
Xm
Φ(Z) = tj Uj∗ ZUj
j=1
70 CHAPTER 4. AROUND HERMITE-HADAMARD

for some convex combination 0 < tj ≤ 1, m

P
j=1 tj = 1, and some unitary matrices Uj (these
scalars ti and matrices Ui depend on Z). In 2003, Zhan [94] noted that we can take m = d.
That we can actually take an average,
d
1X ∗
Φ(Z) = Uj ZUj ,
d
j=1

follows from the quite recent observation [42, Proposition 2.6] mentioned in the proof of The-
orem 4.3.2. For quantum channels, one has the Choi-Kraus decomposition (see [15, Chapter
3])
Xd2
Φ(A) = Ki AKi∗
i=1
Pd2
∗
Pd2 ∗
for some weights Ki ∈ Md such that i=1 Ki Ki = i=1 Ki Ki = I. So, the proof of
Lemma 3.1 shows that, for quantum channels, one may derive the fundamental majoriza-
tion (4.3.2) from two results for convex functions: the basic principle of majorisation and
Hansen-Pedersen’s trace inequality.

Theorem 4.3.2 combined with Theorem 4.1.2 provide a number of interesting operator
inequalities. The next corollary can be regarded as another matrix version of the scalar
inequality (4.1.2). We will give a proof independent of Theorem 4.1.2.

Corollary 4.3.4. Let A, B ∈ Ms.a n and let f (t) be a convex function defined on an interval
containing the spectra of A and B. If X, Y ∈ Mn satisfy X ∗ X + Y ∗ Y = XX ∗ + Y Y ∗ = I,
then, for some unitary matrices {Uk }2n
k=1 in M2n ,

2n
∗ ∗ ∗ 1 X ∗
f (X AX + Y BY ) ⊕ f (Y AY + X BX) ≤ Uk f (A ⊕ B) Uk∗ .
2n
k=1

In particular, for the absolute value, we note that

4n
∗ ∗ ∗ 1 X ∗
|X AX + Y BY | ⊕ |Y AY + X BX| ≤ Uk |A ⊕ B| Uk∗ .
2n
k=1

Proof. The majorization (4.3.1) and the basic principle of majorizations show that

f (X ∗ AX + Y ∗ BY ) ⊕ f (Y ∗ AY + X ∗ BX) ≺w f (A ⊕ B)

for all convex functions defined on an interval containing the spectra of A and B. Thus

f (X ∗ AX + Y ∗ BY ) ⊕ f (Y ∗ AY + X ∗ BX) + C ≺ f (A ⊕ B)

for some positive semidefinte matrice C ∈ M+

2n . Hence, as in the previous proof,

2n
1 X
f (X ∗ AX + Y ∗ BY ) ⊕ f (Y ∗ AY + X ∗ BX) + C ≤ Uk f (A ⊕ B) Uk∗
2n
k=1

for some family of unitary matrices Uk ∈ M2n .

4.3. MAJORIZATION 71

Corollary 4.3.5. Let A, B ∈ M+ ∗ ∗ ∗ ∗

n . If X, Y ∈ Mn satisfy X X + Y Y = XX + Y Y = I,
then,
det(X ∗ AX + Y ∗ BY ) det(Y ∗ AY + X ∗ BX) ≥ det A det B.

Proof. Since the classical Minkowski functional Z 7→ det1/2n Z is concave on M+

2n , the result
is an immediate consequence of Theorem 4.3.2.

The case A = B and X, Y are two orthogonal projections reads as the classical Fisher’s
inequality.
Corollary 4.3.4 yields inequalities for symmetric norms and antinorms on M+ 2n . Symmetric
norms, k · k, also called unitarily invariant norms, are classical objects in matrix analysis. We
refer to [13], [64], [67, Chapter 6] and, in the setting of compact operators, [89]. The most
famous examples are the Schatten p-norms, 1 ≤ p ≤ ∞, and the Ky Fan k-norms.
Symmetric anti-norms k · k! are the concave counterpart of symmetric norms. Famous
examples are the Schatten q-quasi norms, 0 < q < 1 and the Minkowski functional considered
in the proof of Corollary 4.3.5. We refer to [28] and [29, Section 4] for much more examples.

Corollary 4.3.6. Let f (t) and g(t) be two nonnegative functions defined on [a, b] and let
A, B ∈ Mn be Hermitian with spectra in [a, b]. If X, Y ∈ Mn satisfy X ∗ X + Y ∗ Y = XX ∗ +
Y Y ∗ = I, then :

(i) If f (t) is concave, then, for all symmetric antinorms,

kf (X ∗ AX + Y ∗ BY ) ⊕ f (Y ∗ AY + X ∗ BX)k! ≥ kf (A ⊕ B)k! .

(ii) If g(t) is convex, then, for all symmetric norms,

kg (X ∗ AX + Y ∗ BY ) ⊕ g (Y ∗ AY + X ∗ BX)k ≤ kg (A ⊕ B)k .

Proof. Since symmetric antinorms are unitarily invariant and superadditive, the first assertion
follows from the version of Corollary 4.3.4 for concave version. The second assertion is an
immediate consequence of Corollary 4.3.4.

We close this section by mentioning Moslehian’s weak majorization which provides a matrix
version of the first inequality of Proposition 4.1.1. We may restate [81, Corollary 3.4] as
inequalities for symmetric and antisymmetric norms.

Proposition 4.3.7. Let f (t) and g(t) be two nonnegative functions defined on [a, b] and let
A, B ∈ Mn be Hermitian with spectra in [a, b].

(i) If f (t) is concave, then, for all symmetric antinorms,

  1
A+B
Z
f ≥ f ((1 − x)A + xB) dx .
2 ! 0 !

(ii) If g(t) is convex, then, for all symmetric norms,

  1
A+B
Z
g ≤ g((1 − x)A + xB) dx .
2 0
72 CHAPTER 4. AROUND HERMITE-HADAMARD

4.4 References of chapter 4

[5] T. Ando, Majorization, doubly stochastic matrices, and comparison of eigenvalues,
Linear Algebra Appl. 118 (1989), 163-248.

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York,

1996.

[15] R. Bhatia, Positive Definite Matrices, Princeton University press, Princeton 2007.

[28] J.-C. Bourin and F. Hiai, Norm and anti-norm inequalities for positive semi-definite
matrices, Internat. J. Math. 63 (2011), 1121-1138.

[29] J.-C. Bourin and F. Hiai, Jensen and Minkowski inequalities for operator means and
anti-norms. Linear Algebra Appl. 456 (2014), 22–53.

[31] J.-C. Bourin and E.-Y. Lee, Unitary orbits of Hermitian operators with convex or
concave functions, Bull. Lond. Math. Soc. 44 (2012), no. 6, 1085–1102.

[42] J.-C. Bourin and E.-Y. Lee, A Pythagorean Theorem for partitioned matrices, Proc.
Amer. Math. Soc., in press.

[43] J.-C. Bourin and E.-Y. Lee, Matrix inequalities and majorizations around Hermite-
Hadamard’s inequality, preprint

[64] F. Hiai, Matrix Analysis: Matrix Monotone Functions, Matrix Means, and Majoriza-
tion (GSIS selected lectures), Interdisciplinary Information Sciences 16 (2010), 139–248.

[67] F. Hiai, D. Petz, Introduction to Matrix Analysis and applications. Universitext,

Springer, New Delhi, 2014.

[81] M. S. Moslehian, Matrix Hermite-Hadamard type inequalities, Houston J. Math. 39

(2013), no. 1, 177–189.

[83] C. P. Niculescu; L.-E. Persson, Old and new on the Hermite-Hadamard inequality, Real
Anal. Exchange 29 (2003/04), no. 2, 663–685.

[89] B. Simon, Trace ideal and their applications, Cambridge University Press, Cambridge,
1979.

[94] X. Zhan, The sharp Rado theorem for majorizations, Amer. Math. Monthly 110 (2003)
152–153.
 Part II

Operator diagonals and

partitionned matrices

73
 75

U (A + B)U ∗ + V (A + B)V ∗
 
A X
=
X B 2
76
Chapter 5

Pinchings and Masas

For a background on the essential numerical and the proof of the pinching theorem used in
this article, see Section 5.6 “Around this article”.

Pinchings and positive linear maps [37]

Abstract. We employ the pinching theorem, ensuring that some operators A admit any
sequence of contractions as an operator diagonal of A, to deduce/improve two recent theorems
of Kennedy-Skoufranis and Loreaux-Weiss for conditional expectations onto a masa in the
algebra of operators on a Hilbert space. Similarly, we obtain a proof of a theorem of Akeman
and Anderson showing that positive contractions in a continuous masa can be lifted to a
projection. We also discuss a few corollaries for sums of two operators in the same unitary
orbit.
Keywords: Pinching, essential numerical range, positive linear maps, conditional expectation
onto a masa, unitary orbit.
2010 Mathematics Subject Classification. 46L10, 47A20, 47A12.

5.1 The pinching theorem

We recall two theorems which are fundamental in the next sections to obtain several results
about positive linear maps, in particular conditional expectations, and unitary orbits. These
theorems were established in [21], we also refer to this article for various definitions and
properties of the essential numerical range We (A) of an operator A in the algebra L(H) of
all (bounded linear) operators on an infinite dimensional, separable (real or complex) Hilbert
space H.
We denote by D the unit disc of C. We write A ≃ B to mean that the operators A and
B are unitarily equivalent. This relation is extended to operators possibly acting on different
Hilbert spaces, typically, A acts on H and B acts on an infinite dimensional subspace S of H,
or on the spaces H ⊕ H or ⊕∞ H.

Theorem 5.1.1. Let A ∈ L(H) with We (A)L⊃ D and {Xi }∞ i=1 a sequence in L(H) such that
supi kXi k < 1. Then, a decomposition H = ∞ H
i=1 i holds with AHi ≃ Xi for all i.

Of course, the direct sum refers to an orthogonal decomposition, and AHi stands for the
compression of A onto the subspace Hi .

77
78 CHAPTER 5. PINCHINGS AND MASAS

Theorem 5.1.1 tells us that we have a unitary congruence between an operator in L(⊕∞ H)
and a “pinching” of A,
∞
M X∞
Xi ≃ Ei AEi
i=1 i=1

for some sequence of mutually orthogonal infinite dimensional projections {Ei }∞i=1 in L(H)
summing up to the identity I. Thus {Xi }∞i=1 can be regarded as an operator diagonal of A.
In particular, if X is an operator on H with kXk < 1, then, A is unitarily congruent to an
operator on H ⊕ H of the form,  
X ∗
A≃ . (5.1.1)
∗ ∗
For a sequence of normal operators, Theorem 5.1.1 admits a variation. Given A, B ⊂ C,
the notation A ⊂st B means that A + rD ⊂ B for some r > 0.

Theorem 5.1.2. Let A ∈ L(H) and let {Xi }∞ i=1 be a sequenceLof normal operators in L(H)
∞
such that ∪∞
i=1 W (Xi ) ⊂ st W e (A). Then, a decomposition H = i=1 Hi holds with AHi ≃ Xi
for all i.

If all the operators are self-adjoint, this is true for the strict inclusion in R (if A, B ⊂ R,
the notation A ⊂st B then means that A + rI ⊂ B for some r > 0, where I = [−1, 1]). This
is actually an easy consequence of Theorem 5.1.1 or Theorem 5.1.2.

Corollary 5.1.3. Let A ∈ L(H) be self-adjoint and let {Xi }∞ i=1 be a sequence of self-adjoint
L∞
operators in L(H) such that ∪∞ i=1 W (Xi ) ⊂ st W e (A). Then, a decomposition H = i=1 Hi
holds with AHi ≃ Xi for all i.

To get Corollary 5.1.3 from Theorem 5.1.2, let We (A) = [a, b] and write A = A1 + A2 +
A3 + A4 where Ai Aj = 0 if i 6= j, with a ∈ σe (A1 ) ∩ σe (A2 ) and b ∈ σe (A3 ) ∩ σe (A4 ). Apply
Theorem 5.1.2 to the normal operator à = (1 + i)A1 + (1 − i)A2 + (1 + i)A3 +L(1 − i)A4 as
∞
∪∞i=1 W (Xi ) ⊂st conv{(1 ± i)a; (1 ± i)b} ⊂ We (Ã). We get a decomposition H = i=1 Hi with
ÃHi ≃ Xi for all i. Therefore, taking real parts we also have AHi ≃ Xi .
In Section 3, our concern is the study of generalized diagonals, i.e., conditional expectations
onto a masa in L(H), of the unitary orbit of an operator. The pinching theorems are the good
tools for this study; we easily obtain and considerably improve two recent theorems, of Kennedy
and Skoufranis for normal operators, and Loreaux and Weiss for idempotent operators. For
self-adjoint idempotents, i.e., projections, and continuous masas, we obtain a theorem due to
Akemann and Anderson. Section 4 deals with an application to the class of unital, positive
linear maps which are trace preserving. Section 5 collects a few questions on possible extension
of Theorems 5.1.1 and 5.1.2 in the setting of von Neumann algebras.
The next section gives applications which only require (5.1.1). These results mainly focus
on sums of two operators in a unitary orbit.

5.2 Sums in a unitary orbit

We recall a straightforward consequence of (5.1.1) for the weak convergence, [21, Corollary
2.4].
5.2. SUMS IN A UNITARY ORBIT 79

Corollary 5.2.1. Let A, X ∈ L(H) with We (A) ⊃ D and kXk ≤ 1. Then there exists a
sequence of unitaries {Un }∞
n=1 in L(H) such that

wot lim Un AUn∗ = X.

n→+∞

Of course, we cannot replace the weak convergence by the strong convergence; for instance
if A is invertible and kXhk < kA−1 k−1 for some unit vector h, then X cannot be a strong
limit from the unitary orbit of A. However, the next best thing does happen. Moreover, this
is even true for the ∗-strong operator topology.

Corollary 5.2.2. Let A, X ∈ L(H) with We (A) ⊃ D and kXk ≤ 1. Then there exist two
sequences of unitaries {Un }∞ ∞
n=1 and {Vn }n=1 in L(H) such that

Un AUn∗ + Vn AVn∗
∗ sot lim = X.
n→+∞ 2

Proof. From (5.1.1) we have  

X −R
A≃ .
−S T
Hence there exist two unitaries U, V : H → H ⊕ H such that

U AU ∗ + V AV ∗
 
X 0
= . (5.2.1)
2 0 T

Now let {en }∞n=1 be a basis of H and choose any unitary Wn : H⊕H → H such that Wn (ej ⊕0) =
ej for all j ≤ n. Then  
X 0
Xn := Wn Wn∗
0 T
strongly converges to X. Indeed, {Xn } is bounded in norm and, for all j, Xn ej → Xej . Taking
adjoints,  ∗ 
X 0
Xn∗ = Wn Wn∗ ,
0 T∗
we also have Xn∗ → X strongly. Setting Un = Wn U and Vn = Wn V and using (5.2.1) completes
the proof.

Remark 5.2.3. Corollary 5.2.2 does not hold for the convergence in norm. We give an example.
Consider the permutation matrix  
0 0 1
T = 1 0 0
0 1 0
and set A = 2(⊕∞ T ) regarded as an operator in L(H). Then We (A) ⊃ D, however 0 cannot be
a norm limit of means of two operators in the unitary orbit of A. Indeed 0 cannot be a norm
limit of means of two operators in the unitary orbit of (A + A∗ )/2 as (A + A∗ )/2 = 2I − 3P
for some projection P .

Remark 5.2.4. The converse of Corollary 5.2.2 holds: if A ∈ L(H) has the property that any
contraction is a strong limit of a mean of two operators in its unitary orbit, then necessarily
We (A) ⊃ D. This is checked by arguing as in the proof of Corollary 5.4.1.
80 CHAPTER 5. PINCHINGS AND MASAS

We reserve the word “projection” for self-adjoint idempotent. A strong limit of uniformly
bounded idempotent operators is still an idempotent; thus, the next corollary is rather sur-
prising.

Corollary 5.2.5. Fix α > 0. There exists an idempotent Q ∈ L(H) such that for every
X ∈ L(H) with kXk ≤ α we have two sequences of unitaries {Un }∞ ∞
n=1 and {Vn }n=1 in L(H)
for which
∗ sot lim Un QUn∗ + Vn QVn∗ = X.
n→+∞

Proof. Let a > 0, define a two-by-two idempotent matrix

 
1 0
Ma = (5.2.2)
a 0

and set Q = ⊕∞ Ma regarded as an operator in L(H). Since the numerical range W (·) of
 
0 0
2 0

is D, we infer that W (2α−1 Ma ) = We (2α−1 Q) ⊃ D for a large enough a. The result then
follows from Corollary 5.2.2 with A = 2α−1 Q and the contraction α−1 X.

Corollary 5.2.5 does not hold for the convergence in norm.

Proposition 5.2.6. Let X ∈ L(H) be of the form λI + K for a compact operator K and
/ {0, 1, 2}. Then X is not norm limit of Un QUn∗ + Vn QVn∗ for any sequences of
a scalar λ ∈
unitaries {Un }∞ ∞
n=1 and {Vn }n=1 and any idempotent Q in L(H).

Proof. First observe that if {An }∞ ∞

n=1 and {Bn }n=1 are two bounded sequences in L(H) such
that An − Bn → 0 in norm, then we also have A2n − Bn2 → 0 in norm; indeed

A2n − Bn2 = An (An − Bn ) + (An − Bn )Bn .

Now, suppose that λ 6= 1 and that we have the (norm) convergence,

Un QUn∗ + Vn QVn∗ → λI + K.

Then we also have

Wn QWn∗ − (−Q + λI + Un∗ KUn ) → 0 (5.2.3)
where Wn := Un∗ Vn . Hence, by the previous observation,

(Wn QWn∗ )2 − (−Q + λI + Un∗ KUn )2 → 0,

that is
Wn QWn∗ − (−Q + λI + Un∗ KUn )2 → 0. (5.2.4)
Combining (5.2.3) and (5.2.4) we get

(−Q + λI + Un∗ KUn ) − (−Q + λI + Un∗ KUn )2 → 0

5.2. SUMS IN A UNITARY ORBIT 81

hence
(−2 + 2λ)Q + (λ − λ2 )I + Kn → 0
for some bounded sequence of compact operators Kn . Since λ 6= 1, we have
λ
Q= I +L
2
for some compact operator L. Since Q is idempotent, either λ = 2 or λ = 0.

The operator X in Proposition 5.2.6 has the special property that We (X) is reduced to a
single point. However Proposition 5.2.6 may also hold when We (X) has positive measure.

Corollary 5.2.7. Let Q be an idempotent in L(H) and z ∈ C \ {0, 1, 2}. Then, there exists
α > 0 such that the following property holds:
If X ∈ L(H) satisfies kX − zIk ≤ α, then X is not norm limit of Un QUn∗ + Vn QVn∗ for
any sequences of unitaries {Un }∞ ∞
n=1 and {Vn }n=1 in L(H).

Proof. By the contrary, zI would be a norm limit of Un QUn∗ + Vn QVn∗ for some unitaries
Un , Vn , contradicting Proposition 5.2.6.

More operators with large numerical and essential numerical ranges are given in the next
proposition. An operator X is stable when its real part (X + X ∗ )/2 is negative definite
(invertible).

Proposition 5.2.8. If X ∈ L(H) is stable, then X is not norm limit of Un QUn∗ + Vn QVn∗ for
any sequences of unitaries {Un }∞ ∞
n=1 and {Vn }n=1 and any idempotent Q in L(H).

Proof. We have a decomposition H = Hs ⊕ Hns in two invariant subspaces of Q such that

Q acts on Hs as a selfadjoint projection P , and Q acts on Hns as a purely nonselfadjoint
idempotent, that is AHns is unitarily equivalent to an operator on F ⊕ F of the form
 
I 0
QHns ≃ (5.2.5)
R 0

where R is a nonsingular (i.e., a zero kernel) positive operator on a Hilbert space F, so

 
I 0
Q≃P⊕ . (5.2.6)
R 0

Let Y be a norm limit of the sum of two sequences in the unitary orbit of Q. If the purely
non-selfadjoint part Hns is vacuous, then Y is positive, hence Y 6= X. If Hns is not vacuous,
(5.2.6) shows that
 
2I R
Q + Q∗ ≃ 2P ⊕
R 0
   
I I R 0
≃ 2P ⊕ + .
I I 0 −R

This implies that k(Q + Q∗ )+ k ≥ k(Q + Q∗ )− k, therefore Y + Y ∗ cannot be negative definite,

hence X 6= Y .
82 CHAPTER 5. PINCHINGS AND MASAS

It is known [85] that any operator is the sum of five idempotents. We close this section by
asking whether Corollorary 5.2.5 admits a substitute for Banach space operators.

Question 5.2.9. Let X be a separable Banach space and T ∈ L(X ), the linear operators on
X . Do there exist two sequences {Pn }∞ ∞
n=1 and {Qn }n=1 of idempotents in L(X ) such that
T = sot limn→+∞ (Pn + Qn ) ?

5.3 Conditional expectation onto a masa

5.3.1 Conditional expectation of general operators
Kennedy and Skoufranis have studied the following problem: Let X be a maximal abelian ∗-
subalgebra (masa) of a von Neumann algebra M, with corresponding expectation EX : M → X
(i.e., a unital positive linear map such that EX (XM ) = XEX (M ) for all X ∈ X and M ∈ M).
Given a normal operator A ∈ M, determine the image by EX of the unitary orbit of A,

∆X (A) = { EX (U AU ∗ ) : U a unitary in M }.

In several cases, they determined the norm closure of ∆X (A). In particular, [73, Theorem 1.2]
can be stated in the following two propositions.

Proposition 5.3.1. Let X be a masa in L(H), X ∈ X, and A a normal operator in L(H). If

σ(X) ⊂ convσe (A), then X lies in the norm closure of ∆X (A).

Proposition 5.3.2. Let X be a continuous masa in L(H), X ∈ X, and A a normal operator

in L(H). If X lies in the norm closure of ∆X (A), then σ(X) ⊂ convσe (A).

Since we deal with normal operators, σ(X) ⊂ convσe (A) means W (X) ⊂ We (A). Proposi-
tion 5.3.2 needs the continuous assumption. It is a rather simple fact; we generalize it in Lemma
5.3.5: Conditional expectations reduce essential numerical ranges, We (EX (T )) ⊂ We (T ) for all
T ∈ L(H). Thus, the main point of [73, Theorem 1.2] is Proposition 5.3.1 which says that
if W (X) ⊂ We (A) then X can be approximated by operators of the form EX (U AU ∗ ) with
unitaries U . With the slightly stronger assumption W (X) ⊂st We (A), Theorem 5.1.2 guar-
antees, via the following corollary, that X is exactly of this form. Furthermore the normality
assumption on A is not necessary.

Corollary 5.3.3. Let X be a masa in L(H), X ∈ X and A ∈ L(H). If W (X) ⊂st We (A), then
X = EX (U AU ∗ ) for some unitary operator U ∈ L(H).

we note a simple fact: Let {Pi }∞

Proof. First,P i=1 be a sequence of orthogonal projections in
X such that ∞ i=1 Pi = I, and let Z ∈ L(H) such that Pi ZPi ∈ X for all i. Then, we have a
strong sum
X∞
EX (Z) = Pi ZPi .
i=1

Now, denote by Hi the range of Pi and assume dim Hi = ∞ for all i. We have W (XHi ) ⊂
W (X), hence
∪∞i=1 W (XHi ) ⊂st We (A).
5.3. CONDITIONAL EXPECTATION ONTO A MASA 83

We may then apply Theorem 5.1.2 and get a unitary U on H = ∞

L
i=1 Hi such that
 
XH1 ∗ ··· ···
 ∗ XH2 ∗ · · ·
∗
 
A ≃ U AU =   ... ∗
..
.
.. 
.
.
.. ..
 
.. .
. . . ..

Since 0 ⊕ · · · ⊕ XHi ⊕ 0 · · · ∈ X for all i, the previous simple fact shows that
∞
M
EX (U AU ∗ ) = XHi = X.
i=1

We remark that Corollary 5.3.3 also covers the assumption W (X) ⊂ We (A) of Proposition
5.3.1. Indeed, We (A) ⊂st We (A + D) for some normal operator D with arbitrarily small norm,
and we may apply Corollary 5.3.3 to X and A + D.
Kadison’s article [71] completely describes the diagonals of a projection, thus, providing
in his terminology a carpenter theorem for discrete masas L(H). Our method can be used
to obtain a similar statement for continuous masas in L(H) given in the next corollary. This
result is due to Akemann and Anderson, see [1, Corollary 6.19]. The case of a masa in a
type-II1 factor is solved in Ravichandran’s paper [87].

Corollary 5.3.4. Let X be a continuous masa in L(H) and let X ∈ X be a positive contraction.
Then there exists a projection P in L(H) such that X = EX (P ).

Proof. We may suppose that 0 < kXhk < 1 for all unit vectors h ∈ H; otherwise decompose
X as Q + X0 for some L projection Q ∈ X and consider X0 in place of X. Then we have a
decomposition H = ∞ i=1 Hi , whereLeach summand Hi is the range of some projection in X,
∞
such that the decomposition X = i=1 XHi satisfies
L W (XHi ) ⊂st [0, 1] for all i. We may
further decompose each Hilbert space Hi as Hi = ∞ j=1 Hi,j where Hi,j is the 
range of some
L∞ L∞ 
nonzero projection in X. We have a corresponding decomposition X = i=1 j=1 XHi,j
Since ∪∞j=1 W (XHi,j ) ⊂ W (XHi ) ⊂st [0, 1], Corollary 5.1.3 yields some projection Pi ∈ L(Hi ),
We (P ) = [0, 1], such that  
XHi,1 ∗ ··· ···
 ∗ XHi,2 ∗ · · ·
 
Pi =  ..
 .. .. .
 . ∗ . .
.. ..

.. .
. . . ..
Letting P = ∞
L
i=1 Pi , we have X = EX (P ).

5.3.2 A reduction lemma

The following result extends Proposition 5.3.2, the “easy” part of Kennedy-Skoufranis’ theorem
[73, Theorem 1.2].

Lemma 5.3.5. If X is a masa in L(H) and Z ∈ L(H), then W (EX (Z)) ⊂ W (Z) and
We (EX (Z)) ⊂ We (Z).
84 CHAPTER 5. PINCHINGS AND MASAS

Proof. (1) Assume Z is normal. We may identify the unital C ∗ -algebra A spanned by Z with
C 0 (σ(Z)) via a ∗-isomorphism ϕ : C 0 (σ(Z)) → A with ϕ(z 7→ z) = Z. Let h ∈ H be a unit
vector. For f ∈ C 0 (σ(Z)), set
ψ(f ) = hh, EX (ϕ(f ))hi.
Then ψ is a positive linear functional on C 0 (σ(Z)) and ψ(1) = 1. Thus ψ is a Radon measure
induced by a probabilty measure µ,
Z
ψ(f ) = f (z) dµ(z).
σ(Z)

We then have hh, EX (Z)hi = ψ(z) ∈ conv(σ(Z)). Since conv(σ(Z)) = W (Z), we obtain
W (EX (Z)) ⊂ W (Z).
(2) Let Z be a general operator in L(H) and define a conditional expectation
E2 : L(H ⊕ H) → X ⊕ X
by    
A C EX (A) 0
E2 = .
D B 0 EX (B)
From the first part of the proof, we infer
   
EX (Z) 0 Z C
W (EX (Z)) ⊂ W ⊂W
0 EX (B) D B
 
Z C
whenever is normal. Since we have, by a simple classical fact [60],
D B
\  Z C 
W (Z) = W
D B
 
Z C
where the intersection runs over all B, C, D such that is normal, we obtain W (EX (Z)) ⊂
D B
W (Z).
(3) We deal with the essential numerical range inclusion. We can split X into its discrete
part D and continuous part C with the corresponding decomposition of the Hilbert space,
X = D ⊕ C, H = Hd ⊕ Hc .
We then have
We (EX (Z)) = conv {We (ED (ZHd )); We (EC (ZHc ))} . (5.3.1)
We have an obvious inclusion
We (ED (ZHd )) ⊂ We (ZHd ). (5.3.2)
On the other hand, for all compact operators K ∈ L(H),
We (EC (ZHc )) = We (EC (ZHc ) + KHc ) = We (EC (ZHc + KHc )) ⊂ W (ZHc + KHc )
by the simple folklore fact that a conditional expectation onto a continous masa vanishes on
compact operators and part (2) of the proof. Thus, when K runs over all compact operators,
we obtain
We (ED (ZHc )) ⊂ We (ZHc ). (5.3.3)
Combining (5.3.1), (5.3.2) and (5.3.3) completes the proof.
5.3. CONDITIONAL EXPECTATION ONTO A MASA 85

5.3.3 Conditional expectation of idempotent operators

For discrete masas, unlike continuous masas [72], there is a unique conditional expectation,
which merely consists in extracting the diagonal with respect to an orthonormal basis. In a
recent article, Loreaux and Weiss give a detailed study of diagonals of idempotents in L(H).
They established that a nonzero idempotent Q has a zero diagonal with respect to some
orthonormal basis if and only if Q is not a Hilbert-Schmidt perturbation of a projection (i.e.,
a self-adjoint idempotent). They also showed that any sequence {an } ∈ l∞ such that |an | ≤ α
for all n and, for some an0 , ak = an0 for infinitely many k, one has an idempotent Q such
that kQk ≤ 18α + 4 and Q admits {an } as a diagonal with respect to some orthonormal basis
[78, Proposition 3.4]. Using this, they proved that any sequence in l∞ is the diagonal of some
idempotent operator [78, Theorem 3.6], answering a question of Jasper. This statement is in
the range of Theorem 5.1.1. Further, it is not necessary to confine to diagonals, i.e., discrete
masas, and the constant 18α + 4 can be improved; in the next corollary we have 3 as the best
constant when α = 1.

Corollary 5.3.6. Let X be a masa in L(H) and α > 0. There exists an idempotent Q ∈ L(H),
such that for all X ∈ X with kXk < α, we have X = EX (U QU ∗ ) for some unitary operator
U ∈ L(H). If α = 1, kQk = 3 is the smallest possible norm.

Proof. As in the proof of Corollary 5.2.5 we have an idempotent Q such that We (Q) ⊃ αD,
hence the first and main part of Corollary 5.3.6 follows from Corollary 5.3.3. The remaining
parts require a few computations.
To obtain the bound 3 when α = 1 we get a closer look at ⊕∞ Ma with Ma given by (5.2.2)
where a is a positive scalar. We have

W (Ma ) = hh, Ma hi : h ∈ C2 , khk = 1



= |h1 |2 + ah2 h1 : |h1 |2 + |h2 |2 = 1 ,



√
hence, with h1 = reiθ , h2 = 1 − r 2 eiα ,
[ n p o
W (Ma ) = r 2 + ar 1 − r 2 ei(θ−α) : θ, α ∈ [0, 2π] .
0≤r≤1

√
Therefore W (Ma ) is a union of circles Γr with centers r 2 and radii ar 1 − r 2 . To have
D ⊂ W (Ma ) it is necessary and sufficient that −1 ∈ Γr for some r ∈ [0, 1], hence

1 + r2
a= √ . (5.3.4)
r 1 − r2

Now we minimize a = a(r) given by (5.3.4) when r ∈ (0, 1) and thus obtain the matrix Ma∗
with smallest norm such that W (Ma∗ ) ⊃ D. Observe that a(r) → +∞ as r → 0 and as r → 1,
and
r 2 (1 − r 2 )3/2 a′ (r) = 3r 2 − 1.
√ √
Thus a(r) takes its minimal value a∗ when r = 1/ 3. We have a∗ = 2 2, hence

kMa∗ k = 3.
86 CHAPTER 5. PINCHINGS AND MASAS

Now, letting Q = ⊕∞ Ma∗ , we have We (Q) = W (Ma∗ ), so that Q is an idempotent in L(H)

such that We (Q) ⊃ D, and thus by Corollary 5.3.3 any operator X such that kXk < 1 satifies
EX (U QU ∗ ) = X for some unitary U .
It remains to check that if Q is an idempotent such that Corollary 5.3.6 holds for any oper-
ator X such that kXk < 1, then kQk ≥ 3. To this end, we consider the purely nonselfadjoint
part QHns of Q in (5.2.5),  
I 0
QHns ≃ .
R 0
We have We (Q) ⊃ D if and only if We (QHns ) ⊃ D. By Lemma 5.3.5 this is necessary. We
may approximate We (QHns ) with sligthly larger essential numerical ranges, by using a positive
diagonalizable operator Rε such that Rε ≥ R ≥ Rε − εI, for which
  ∞  !
I 0 M 1 0
We = We
Rε 0 an 0
n=1

where {an }∞ n=1 is a sequence of positive scalars, the eigenvalues of Rε . By the previous step
of the proof, this essential numerical range contains D if and only if lim an ≥ a∗ . If this holds
for all ε > 0, then kQk ≥ 3.

5.4 Unital, trace preserving positive linear maps

Unital positive linear maps Φ : Mn → Mn , the matrix algebra, which preserve the trace play
an important role in matrix analysis and its applications. These maps are sometimes called
doubly stochastic [5].
We say that Φ : L(H) 7→ L(H) is trace preserving if it preserves the trace ideal T and
Tr Φ(Z) = Tr Z for all Z ∈ T .
Corollary 5.4.1. Let A ∈ L(H). The following two conditions are equivalent:
(i) We (A) ⊃ D.
(ii) For all X ∈ L(H) with kXk < 1, there exists a unital, trace preserving, positive linear
map Φ : L(H) → L(H) such that Φ(A) = X.
We may further require in (ii) that Φ is completely positive and sot- and wot-sequentially
continuous.

Proof. Assume (i). By Theorem 5.1.1 we have a unitary U : H → ⊕∞ H such that

 
X ∗ ··· ···
∗ X ∗ · · ·
A ≃ U AU ∗ = 
 
 ... ∗ X
.. 
. .
.. .. . .
 
.
. . . ..

Now consider the map Ψ : L(⊕∞ H) → L(H),

 
Z1,1 Z1,2 · · · ∞
Z2,1 Z2,2 · · · X
→
7 2−i Zi,i
.. ..
 
.. i=1
. . .
5.4. UNITAL, TRACE PRESERVING POSITIVE LINEAR MAPS 87

and define Φ : L(H) → L(H) as Φ(T ) = Ψ(U T U ∗ ). Since both Ψ and the unitary congruence
with U are sot- and wot-sequentially continuous, and trace preseverving, completely positive
and unital, so is Φ. Further Φ(A) = X.
Assume (ii) and suppose that z ∈/ We (A) and |z| < 1 in order to reach a contradiction. If
iθ −iθ
z = |z|e , replacing A by e A, we may assume 1 > z ≥ 0. Hence,

We ((A + A∗ )/2) ⊂ (−∞, z]

and there exists a selfadjoint compact operator L such that

A + A∗
≤ zI + L.
2

This implies that X := 1+z 2 I cannot be in the range of Φ for any unital, trace preserving
positive linear map. Indeed, we would have

X + X∗ A + A∗
 
1+z
I= =Φ ≤ zI + Φ(L)
2 2 2

which is not possible as Φ(L) is compact.

In the finite dimensional setting, two Hermitian matrices A and X satisfy the relation
X = Φ(A) for some positive, unital, trace preserving linear map if and only if X is in the
convex hull of the unitary orbit of A. In the infinite dimensional setting, if two Hermitian
A, X ∈ L(H) satisfy We (A) ⊃ [−1, 1] and kXk ≤ 1, then X is in the norm closure of the
unitary orbit of A. This is easily checked by approximating the operators with diagonal
operators. Such an equivalence might not be brought out to the setting of Corollary 5.4.1.

Question 5.4.2. Do there exist A, X ∈ L(H) such that We (A) ⊃ D, kXk < 1, and X does
not belong to the norm closure of the convex hull of the unitary orbit of A ?

Here we mention a result of Wu [93, Theorem 6.11]: If A ∈ L(H) is not of the form scalar
plus compact, then every X ∈ L(H) is a linear combination of operators in the unitary orbit
of A.
If one deletes the positivity assumption, the most regular class of linear maps on L(H)
might be given in the following definition.

Definition 5.4.3. A linear map Ψ : L(H) → L(H) is said ultra-regular if it fulfills two
conditions:

(u1) Ψ(I) = I and Ψ is trace preserving.

(u2) Whenever a sequence An → A for either the norm-, strong-, or weak-topology, then we
also have Ψ(An ) → Ψ(A) for the same type of convergence.

Any ultra-regular linear map preserves the set of essentially scalar operators (of the form
λI +K with λ ∈ C and a compact operator K). For its complement, we state our last corollary.

Corollary 5.4.4. Let A ∈ L(H) be essentially nonscalar. Then, for all X ∈ L(H) there exists
a ultra-regular linear map Ψ : L(H) → L(H) such that Ψ(A) = X.
88 CHAPTER 5. PINCHINGS AND MASAS

Proof. An operator is essentially nonscalar precisely when its essential numerical range is not
reduced to a single point. So, let a, b ∈ We (A), a 6= b. By a lemma of Anderson and Stampfli
[3], A is unitarily equivalent to an operator on H ⊕ H of the form
 
D ∗
B=
∗ ∗

where D = ⊕∞
n=1 Dn , with two by two matrices Dn ,
 
an 0
Dn =
0 bn

such that an → a and bn → b as n → ∞. We may assume that, for some α, β > 0, we have
α > |an |+|bn | and |an −bn | > β. Hence there exist γ > 0 and two by two intertible matrices Tn
such that, for all n, W (Tn Dn Tn−1 ) ⊃ D and kTn k + kTn−1 k ≤ γ. So, letting T = (⊕∞
n=1 Tn ) ⊕ I,
we obtain an invertible operator T on H ⊕ H such that We (T BT −1 ) ⊃ D.
Hence we have an invertible operator S on H such that We (SAS −1 ) ⊃ D. Therefore we
may apply Corollary 5.4.1 and obtain a wot- and sot-sequentially continuous, unital, trace
preserving map Φ such that Φ(SAS −1 ) = X. Letting Ψ(·) = Φ(S · S −1 ) completes the
proof.

We cannot find an alternative proof, not based on the pinching theorem, for Corollaries
5.4.1 and 5.4.4.
If we trust in Zorn, there exists a linear map Ψ : L(H) → L(H) which satifies the condition
(u1) but not the condition (u2). Indeed, let {ap }p∈Ω be a basis in the Calkin algebra C =
L(H)/K(H), indexed on an ordered set Ω, whose first element ap0 is the image of I by the
canonicalPprojection π : L(H) → C. Thus, for each operator X, we have a unique decomposition
π(X) = p∈Ω (π(X))p ap with only finitely many nonzero terms. Further (π(X))p0 = 0 if X is
compact, and (π(I))p0 = 1. We then define a map ψ : L(H) → L(H ⊕ H) by
 
X 0
ψ(X) = .
0 (π(X))p0 I

Letting Ψ(X) = V ψ(X)V ∗ where V : H ⊕ H → H is unitary, we obtain a linear map Ψ :

L(H) → L(H) which satifies (u1) but not (u2): it is not norm continuous.
Let ω be a Banach limit on l∞ and define a map φ : l∞ → l∞ , {an } 7→ {bn }, where
b1 = ω({an }) and bn = an−1 , n ≥ 2. Letting Ψ(X) = φ(diag(X)), where diag(X) is the
diagonal of X ∈ H in an orthonormal basis, we obtain a linear map Ψ which is norm continuous,
satisfies (u1) but not (u2): it is not strongly sequentially continuous.
However, it seems not possible to define explicitly a linear map Ψ : L(H) → L(H) satisfying
(u1) but not (u2).

5.5 Pinchings in factors ?

We discuss possible extensions to our results to a von Neumann algebra R acting on a separable
Hilbert space H. First, we need to define an essential numerical range WeR for R. Let A ∈ R.
If R is type-III, then WeR (A) := We (A). If R is type-II∞ , then
\
WeR (A) := W (A + K)
K∈T
5.6. AROUND THIS ARTICLE 89

where T is the trace ideal in R (we may also use its norm closure K, the “compact” operators
in R, or any dense sequence in K)

Question 5.5.1. In Corollaries 5.2.1, 5.2.2 and 5.2.5, can we replace L(H) by a type-II∞ or
-III factor R with WeR ?

Question 5.5.2. In Corollaries 5.3.3 and 5.3.6, can we replace L(H) by a type-II∞ or -III
factor R with WeR ?

Question 5.5.3. In Corollaries 5.4.1 and 5.4.4, can we replace L(H) by a type-II∞ factor R
with WeR ?

Recently, Dragan and Kaftal [58] obtained some decompositions for positive operators in
von Neumann factors, which, in the case of L(H) were first investigated in [33]-[34] by using
Theorem 5.1.1. This suggests that our questions dealing with a possible extension to type-II∞
and -III factors also have an affirmative answer. In fact, it seems pausible that Theorem 5.1.1
and Theorem 5.1.2 admit a version for such factors and this would affirmatively answer these
questions.
Let R be a type-II∞ or -III factor.
P∞
Definition 5.5.4. A sequence {Vi }∞
i=1 of isometries in R such that i=1 Vi Vi
∗ = I is called
an isometric decomposition of R.

Conjecture 5.5.5. Let A ∈ R with WeR (A) ⊃ D and {Xi }∞ i=1 a sequence in R such that
supi kXi k < 1. Then, there exists an isometric decomposition {Vi }∞ ∗
i=1 of R such that Vi AVi =
Xi for all i.

5.6 Around this article

5.6.1 Essential numerical range
Let us give three equivalent definitions of the essential numerical range We (A) of an operator
A acting on the Hilbert space H.

(1) We (A) = ∩W (A + K), the intersection running over the compact operators K

(2) Let {En } be any sequence of finite rank projections converging strongly to the identity
and denote by Bn the compression of A to the subspace En⊥ . Then We (A) = ∩n≥1 W (Bn )

(3) We (A) = {λ | there is an orthonormal system {en }∞

n=1 with limhen , Aen i = λ}.

It follows that We (A) is a compact convex set containing the essential spectrum of A,
Spe (A). The equivalence between these definitions has been known since the early seventies
if not soone. The very first definition of We (A) = is (1); however (3) is also a natural notion
and easily entails convexity and compactness of the essential numerical range.
90 CHAPTER 5. PINCHINGS AND MASAS

5.6.2 Proof of the pinching theorem

Recall that an operator mean an element of the algebra L(H) of all bounded linear operators
acting on the usual (i.e. complex, separable, infinite dimensional) Hilbert space H. We will
denote by the same letter a projection and the corresponding subspace. Thus, if F is a
projection and A is an operator, we denote by AF the compression of A by F , that is the
restriction of F AF to the subspace F . Given a total sequence of nonzero mutually orthogonal
projections {En }, we consider the pinching
∞
X ∞
M
P(A) = En AEn = AEn .
n=1 n=1

If {An } is a sequence of operators acting on separable Hilbert

L∞ spaces with An unitarily equiv-
alent to AEn for all n, we also naturally write P(A) ≃ n=1 An . Our main result [21] for
operator diagonals can then be stated as:

Theorem 5.6.1. Let A be an operator with We (A) ⊃ D and let {An }∞

n=1 be a sequence of
operators such that supn kAn k∞ < 1. Then, we have a pinching
∞
M
P(A) ≃ An .
n=1

We need two lemmas. The first one is Theorem 5.6.1 for a single strict contraction:

Lemma 5.6.2. Let A be an operator with We (A) ⊃ D and let X be a strict contraction. Then
there exists a projection E such that AE = X.

The second Lemma is a refined version of the first one:

Lemma 5.6.3. Let a ≥ 1 and 1 > ρ > 0 be two constants. Let h be a norm one vector,
let X be a strict contraction with kXk∞ < ρ and let B be an operator with kBk∞ ≤ a and
We (B) ⊃ D. Then, there exist a number ε > 0, only depending on ρ and a, and a projection
E such that:
(i) dim E = ∞ and BE = X,
(ii) dim E ⊥ = ∞, We (BE ⊥ ) ⊃ D and kEhk ≥ ε.

Proof of Theorem 5.6.1. The proof is organized in five steps:

Step 1. Some preliminaries are given.
Step 2. Proof of Lemma 5.6.2 in the special case when X is normal, diagonalizable.
Step 3. Proof of Lemma 5.6.2 in the general case.
Step 4. Proof of Lemma 5.6.3.
Step 5. Conclusion.

1. Preliminaries
We shall use a sequence {Vk }k≥1 of orthogonal matrices acting on spaces of dimensions 2k .
This sequence is built up by induction:
   
1 1 1 1 Vk−1 Vk−1
V1 = √ then Vk = √ for k ≥ 2.
2 −1 1 2 −Vk−1 Vk−1
5.6. AROUND THIS ARTICLE 91

Given a Hilbert space G and a decomposition

2 k
M
G= Hj with H1 = · · · = H2k = H,
j=1
N
we may consider the unitary (orthogonal) operator on G : Wk = Vk I, where I denotes the
identity on H,
Now, let B : G → G be an operator which, with respect to the above decomposition of G,
has a block diagonal matrix  
B1
B=
 .. .

.
B2k
We observe that the block matrix representation of Wk BWk∗ has its diagonal entries all equal
to
1
(B1 + . . . B2k ) .
2k
So, the orthogonal operators Wk allow us to pass from a block diagonal matrix representation
to a block matrix representation in which the diagonal entries are all equal.

2. Proof of Lemma 5.6.2 when X is normal, diagonalizable.

Let {λn (X)}n≥1 be the eigenvalues of X repeated according to their multiplicities. Since
|λn (X)| < 1 for all n and We (A) ⊃ D, we may find a norm one vector e1 such that he1 , Ae1 i =
λ1 (T ). Let F1 = [span{e1 , Ae1 , A∗ e1 }]⊥ . As F1 is of finite codimension, We (AF1 ) ⊃ D. So,
there exists a norm one vector e2 ∈ F1 such that he2 , Ae2 i = λ2 (T ). Next, we set F2 =
[span{e1 , Ae1 , A∗ e1 , e2 , Ae2 , A∗ e2 }]⊥ , . . . . If we go on like this, we exhibit an orthonormal
system {en }n≥1 such that, setting E = span{en }n≥1 , we have AE = X.

3. Proof of Lemma 5.6.2 in the general case.

The contraction Y = (1/kXk∞ )X can be dilated in a unitary

−(I − Y Y ∗ )1/2
 
Y
U=
(I − Y ∗ Y )1/2 Y∗
thus X can be dilated in a normal operator N = kXk∞ U with kN k∞ < ρ. This permits to
restrict to the case when X is a normal strict contraction. So, let X be a normal operator
with kXk∞ < ρ < 1. We remark with the Berg-Weyl-von Neumann theorem, that X can be
written as
X =D+K (1)
where D is normal diagonalizable, kDk∞ = kXk∞ < ρ, and K is compact with an arbitrarily
small norm. Let K = ReK + iImK be the Cartesian decomposition of K. We can find an
integer l, a real α and a real β such that decomposition (1) satisfies:
a) the operators αD, βReK, βImK are dominated in norm by ρ,
b) there are positive integers m, n with 2l = m + 2n and
1
X= (mαD + nβReK + nβiImK). (2)
2l
More precisely we can take any l such that [2l /(2l − 2)].kXk∞ < ρ. Next, assuming kKk∞ <
ρ/2l , we can take m = 2l − 2, n = 1, α = 2l /(2l − 2) and β = 2l .
92 CHAPTER 5. PINCHINGS AND MASAS

Let then T be the diagonal normal operator acting on the space

2 l
M
G= Hj with H1 = · · · = H2l = H,
j=1

and defined by      
m m+n l
2
M M M M M
T = Dj   Rj   Sj 
j=1 j=m+1 j=m+n+1

where Dj = αD, Sj = βReK and Sj = βiImK.

We note that kT k∞ < ρ < 1 and that the operator Wl T Wl∗ , represented in the preceding
decomposition of G, has its diagonal entries all equal to X by (2). Hence, applying the
preceding step to T yields Lemma 5.6.2.

4. Proof of Lemma 5.6.3.

Let a ≥ 1 and let 1 > ρ > 0 be two constants. We take an arbitrary norm one vector h and
any operator B satisfying to the assumptions of Lemma 5.6.3. We can show, using the same
reasoning as that applied in the above Step 2, that we have an orthonormal system {fn }n≥0 ,
with f0 = h, such that:
a) hf2j , Bf2j i = 0 for all j ≥ 1.
b) {hf2j+1 , Bf2j+1 i}j≥0 is a dense sequence in D.
c) If F = span{fj }j≥0 , then BF is the normal operator
X
hfj , Bfj ifj ⊗ fj .
j≥0

Setting F0 = span{f2j }j≥0 and F0′

= span{f2j+1 }j≥0 , we then have:
a) With respect to the decomposition F = F0 F0′ , BF can be written
L
 
BF 0 0
BF = .
0 BF0′
b) We (BF0′ ) ⊃ D and h ∈ F0 .
We can then write a decomposition of F0′ , F0′ = ∞
L
j=1 Fj where
Lfor each index j, Fj commutes
with BF and We (BFj ) ⊃ D; so that the decomposition F = ∞ j=0 Fj yields a representation
of BF as a block diagonal matrix,
M∞
BF = BF j .
j=0
Since We (BFj ) ⊃ D when j ≥ 1, the same reasoning as in Step 3 entails
L ′that for any sequence
{Xj }j≥0 of strict contractions we have decompositions (†) Fj = Gj Gj allowing us to write,
for j ≥ 1,  
Xj ∗
BF j = .
∗ ∗
Since kXk∞ < ρ < 1 and kBk∞ ≤ a, we can find an integer l only depending on ρ and a, as
well as strict contractions X1 , . . . , X2l , such that
 
l −1
2X
1 
X = l BF 0 + Xj  . (3)
2
j=1
5.6. AROUND THIS ARTICLE 93

Considering decompositions (†) adapted to these Xj , we set

 
l −1
M 2M
G = F0  Gj  .
j=1

With respect to this decomposition,

 
BF 0
 X1 
BG =  .
 
..
 . 
X2l −1

Then we deduce from (3) that the block matrix Wl BG Wl∗ has its diagonal entries all equal to
X.
Ll
Summary: h ∈ G and there exists a decomposition G = 2j=1 Ej , in which l depends only
on ρ and a, such that BEj = X for each j. Thus we have an integer j0 such that, setting
Ej0 = E, we have
1
BE = X and kEhk ≥ √ .
2l
√
Taking ε = 1/ 2l ends the proof of Lemma 5.6.3.

4. Conclusion.
Fix a dense sequence {hn } in the unit sphere of H and set a = kAk∞ . We claim that the
statement (i) and (ii) of Lemma 5.6.3 ensure
Pthat there exists a sequence of mutually orthogonal
projections {Ej } such that, setting Fn = j≤n Ej , we have for all integers n:
(∗) An = AEn and We (AFn⊥ ) ⊃ D (so dim Fn⊥ = ∞),
(∗∗) kFn hn k ≥ ε.
In Lemma 5.6.3, set a = kAk∞ . Replacing B by A, Lemma 5.6.3 proves (∗) and (∗∗) for
n = 1. Suppose this holds for an N ≥ 1. Let ν(N ) ≥ N + 1 be the first integer for which
FN hν(N ) 6= 0. Note that kAF ⊥ k∞ ≤ kAk∞ . We apply Lemma 5.6.3 to B = AF ⊥ , X = AN +1
N N
and h = FN hν(N ) /kFN hν(N ) k. We then deduce that (∗) and (∗∗) are still valid for N + 1.
Therefore (∗) and (∗∗) hold for allP∞n. Denseness of {hn } and (∗∗) show that Fn strongly
increases to the identity I so that j=1 Ej = I as required. 

5.6.3 Müller-Tomilov’s theorem

Define the diagonal set ∆(A) of an operator A as the scalars λ ∈ C such that

λ = hen , Aen i, n = 1, 2, . . .

for some orthonormal basis {en }∞

n=1 . One has

int We (A) ⊂ ∆(A) ⊂ We (A)

The set ∆(A) is an analytic set, I do not know examples where ∆(A) is not Borel. The
boundary of ∆(A) might be quite complicated. In 2003 [21], I asked wether ∆(A) is always
convex. Müller and Tomilov give a positive answer in the recent paper [82].
94 CHAPTER 5. PINCHINGS AND MASAS

5.7 References of Chapter 5

[1] C.A. Akemann and J. Anderson. Lyapunov theorems for operator algebras. Mem.
Amer. Math. Soc., 94 no 458, 1991.

[3] J.H. Anderson and J.G. Stampfli, Commutators and compressions, Israel J. Math. 10
(1971), 433–441.

[5] T. Ando, Majorization, doubly stochastic matrices, and comparison of eigenvalues,

Linear Algebra Appl. 118 (1989), 163-248.

[21] J.-C. Bourin, Compressions and pinchings, J. Operator Theory 50 (2003), no. 2, 211-
220.

[33] J.-C. Bourin and E.-Y. Lee, Sums of Murray-von Neumann equivalent positive opera-
tors, C. R. Math. Acad. Sci. Paris 351 (2013), no. 19-20, 761-764.

[34] J.-C. Bourin and E.-Y. Lee, Sums of unitarily equivalent positive operators, C. R.
Math. Acad. Sci. Paris 352 (2014), no. 5, 435–439.

[37] J.-C. Bourin and E.-Y. Lee, Pinchings and positive linear maps, J. Funct. Anal. 270
(2016), no. 1, 359–374.

[58] C. Dragan and V. Kaftal, Sums of equivalent sequences of positive operators in von
Neumann factors, preprint, arXiv:1504.03193.

[60] P.R. Halmos, Numerical ranges and normal dilations, Acta Sci. Math. (Szeged) 25
(1964) 1–5.

[71] R. Kadison, The Pythagorean theorem. II. The infinite discrete case, Proc. Natl. Acad.
Sci. USA 99 (2002), no. 8, 5217-5222.

[72] R. Kadison and I. Singer, Extensions of pure states, Amer. J. Math. (1959), 383-400.

[73] M. Kennedy and P. Skoufranis, The Schur-Horn Problem for Normal Operators, Proc.
London Math. Soc., in press, arXiv:1501.06457.

[78] J. Loreaux and G. Weiss, Diagonality and idempotents with applications to problems
in operator theory and frame theory, J. Operator Theory, in press, arXiv:1410.7441.

[82] V. Müller, Y. Tomilov, In search of convexity: diagonals and numerical ranges, Bull.
London Math soc., 53 (2021), no. 4, 1016–1029.

[85] C. Pearcy and D. Topping, Sums of small numbers of idempotents, Michigan J. Math.
14 (1967) 453–465.

[87] M. Ravichandran, The Schur-Horn theorem in von Neumann algebras, preprint, arXiv:1209.0909.

[93] P.Y. Wu, Additive combination of special operators, Banach Center Publ. 30 (1994),
337-361.
Chapter 6

Partial trace

Decomposition and partial trace of positive matrices

with Hermitian blocks [32]
Abstract. Let H = [As,t ] be a positive definite matrix written in β × β Hermitian blocks and
let ∆ = A1,1 + · · · + Aβ,β be its partial trace. Assume that β = 2p for some p ∈ N. Then, up
to a direct sum operation, H is the average of β matrices isometrically congruent to ∆. A few
corollaries are given, related to important inequalities in quantum information theory such as
the Nielsen-Kempe separability criterion.
Keywords: Positive definite matrices, norm inequalities, partial trace, separable state.
AMS subjects classification 2010: 15A60, 47A30, 15A42.

6.1 Introduction and a key lemma

Positive semi-definite matrices partitioned in two by two blocks occur as an efficient tool in
matrix analysis, sometimes a magic tool ! − according to Bhatia’s famous book [13]. These
partitions allow to derive a lot of important inequalities and those with Hermitian blocks shed
much light on the geometric and harmonic matrix means. Partitions into a larger number of
blocks are naturally involved with tensor products, in the theory of positive linear maps and
in their application in quantum physics.
This article deals with positive matrices partitioned in Hermitian blocks. By using unitary
or isometry congruences, we will improve some nice majorisations, or norm estimates, first
obtained in the field of quantum information theory.
For partitioned positive matrices, the diagonal blocks play a quite special role. This is
apparent in a rather striking decomposition due to the authors [31].

Lemma 6.1.1. For every matrix in M+

n+m written in blocks, we have a decomposition
     
A X A 0 ∗ 0 0
=U U +V V∗
X∗ B 0 0 0 B

for some unitaries U, V ∈ Mn+m .

This lemma leads to study partitions via unitary congruences. It is the key of the subsequent
results. A proof and several consequences can be found in [31] and [28]. Of course, Mn is the
algebra of n × n matrices with real or complex entries, and M+ n is the positive part. That

95
96 CHAPTER 6. PARTIAL TRACE

is, Mn may stand either for Mn (R), the matrices with real entries, or for Mn (C), those with
complex entries. The situation is different in the next statement, where complex entries seem
unavoidable.

Theorem 6.1.2. Given any matrix in M+ 2n (C) written in blocks in Mn (C) with Hermitian
off-diagonal blocks, we have
 
A X 1
= {U (A + B)U ∗ + V (A + B)V ∗ }
X B 2

for some isometries U, V ∈ M2n,n (C).

Here Mp,q (C) denote the space of p rows and q columns matrices with complex entries, and
V ∈ Mp,q (C) is an isometry if p ≥ q and V ∗ V = Iq . Even for a matrix in M+2n (R), it seems
essential to use isometries with complex entries ! The result, due to Lin and the authors, is
based on Lemma 6.1.1, a proof is in [46] and implicitly in [45].
There is no evidence whether a positive block-matrix H in M+ 3n ,
 
A X Y
H = X B Z 
Y Z C

with Hermitian off-diagonal blocks X, Y, Z, could be decomposed as

1
H= {U ∆U ∗ + V ∆V ∗ + W ∆W ∗ }
3
where ∆ = A+B +C and U, V, W are isometries. In fact, this would be surprising. However, a
quite nice decomposition is possible by considering direct sum copies: this provides a substitute
to Theorem 6.1.2 for partitions into an arbitrary number of blocks; it is the main result of this
article.
These decompositions entail some nice inequalities. Lemma 6.1.1 yields a simple estimate
for all symmetric (or unitarily invariant) norms,
 
A X
≤ kAk + kBk. (6.1.1)
X∗ B

Recall that a symmetric norm on Mm satisfies kAk = kU Ak = kAU k for all A ∈ Mm and all
unitaries U ∈ Mm . This obviously induces a symmetric norm on Mn , 1 ≤ n ≤ m. The most
familiar symmetric norms are the Schatten p-norms, 1 ≤ p < ∞,

kAkp = {Tr (A∗ A)p/2 }1/p , (6.1.2)

and, with p → ∞, the operator norm. In general, the sum of the norms kAk + kBk can not
be replaced in (6.1.1) by the norm of the sum kA + Bk. However, Theorem 6.1.2 implies the
following remarkable corollary.

Corollary 6.1.3. Given any matrix in M+

2n written in blocks in Mn with Hermitian off-
diagonal blocks, we have  
A X
≤ kA + Bk
X B
for all symmetric norms.
6.2. DIRECT SUM AND PARTIAL TRACE 97

This is the simplest case of Hiroshima’s theorem, discussed in the next section. There are
some positive matrices in M6 partitioned in blocks in M3 , with normal off-diagonal blocks X,
X ∗ , such that  
A X
> kA + Bk∞ .
X∗ B ∞
Hence the assumptions are rather optimal.
In Section 2, we state our decomposition and derive several inequalities, most of them
related to Hiroshima’s theorem. The proof of the decomposition is given in Section 3. A
discussion of previous results for small partitions and some remarks related to quantum infor-
mation are given in the last section.

6.2 Direct sum and partial trace

A typical example of positive matrices written in blocks are formed by tensor products. Indeed,
the tensor product A ⊗ B of A ∈ Mβ with B ∈ Mn can be identified with an element of
Mβ (Mn ) = Mβn . Starting with positive matrices in M+ +
β and Mn we then get a matrix in
M+βn partitioned in blocks in Mn . In quantum physics, sums of tensor products of positive
semi-definite (with trace one) occur as so-called separable states. In this setting of tensor
products, the sum of the diagonal block is called the partial trace (with respect to Mβ ). We
will use this terminology.

Theorem 6.2.1. Let H = [As,t ] ∈ M+ βn be written in β × β Hermitian blocks in Mn and let

Pβ
∆ = s=1 As,s be its partial trace. If β is dyadic, then, with m = 2β , we have
β
m 1X
⊕ H= Vk (⊕m ∆) Vk∗
β
k=1

where {Vk }βk=1 is a family of isometries in Mmβn,mn .

Here the spaces Mn and Mp,q denote either the real or complex spaces of matrices. By a
dyadic number β, we mean β = 2p for some p ∈ N.
This theorem has strong links with quantum information theory (QIT) as detailed in Section
4. Researchers in QIT may like to restate the theorem by replacing direct sums with tensors
products, ⊕m H → Im ⊗ H and ⊕m ∆ → Im ⊗ ∆. Tensoring in identity means that an operator
on a Hilbert space H is lifted to an operator acting on F ⊗ H where F is an auxiliary Hilbert
space, an ancilla space in the QIT terminology.

6.2.1 Around Hiroshima’s theorem

A straightforward application of Theorem 6.2.1 is the following beautiful result first proved by
Hiroshima in 2003 (see Section 4 for the complete form of Hiroshima’s theorem and is relevance
in quantum physics).
+
Pα 6.2.2. Let H = [As,t ] ∈ Mαn be written in α × α Hermitian blocks in Mn and let
Corollary
∆ = s=1 As,s be its partial trace. Then, we have

kHk ≤ k∆k
98 CHAPTER 6. PARTIAL TRACE

for all symmetric norms.

Proof. By completing H with some zero rows and columns, we may assume that α = β is
dyadic. Theorem 6.2.1 then implies, with m = 2β ,

k ⊕m Hk ≤ k ⊕m ∆k

for all symmetric norms, which is equivalent to the claim of the corollary.

By the Ky Fan principle, Corollary 6.2.2 is equivalent to the majorisation relation

j
X j
X
λi (H) ≤ λi (∆)
i=1 i=1

for all j = 1, . . . , αn. (we set λj (A) = 0 when A ∈ M+

d and j > d).
Theorem 6.2.1 says much more than this majorisation. For instance, we may completes
these eigenvalue relations with the following ones.
+
Pα 6.2.3. Let H = [As,t ] ∈ Mαn be written in α × α Hermitian blocks in Mn and let
Corollary
∆ = s=1 As,s be its partial trace. Then, we have

λ1+βk (H) ≤ λ1+k (∆)

for all k = 0, . . . , n − 1, where β is the smallest dyadic number such that α ≤ β.

+
Proof. By completing H ∈ M+ αn with some zero blocks, we may assume that H ∈ Mβn .
Theorem 6.2.1 then yields the decomposition
β
m 1X
⊕ H= Vk (⊕m ∆) Vk∗
β
k=1

where {Vk }βk=1 is a family of isometries in Mmβn,mn and m = 2β . We recall a simple fact,
Weyl’s theorem: if Y, Z ∈ Md are Hermitian, then

λr+s+1 (Y + Z) ≤ λr+1 (Y ) + λs+1 (Z)

for all nonnegative integers r, s such that r + s ≤ d − 1. When Y, Z are positive, this still holds
for all nonnegative integers r, s with our convention (λj (A) = 0 when A ∈ M+ d and j > d).
From the previous decomposition we thus infer

λ1+βk (⊕m H) ≤ λ1+k (⊕m ∆) (6.2.1)

for all k = 0, 1, . . .. Then, observe that for all A ∈ M+

d and all j = 0, 1, . . .,

λ1+j (⊕m A) = λh(1+j)/mi (A) (6.2.2)

where hui stands for the smallest integer greater than or equal to u. Combining (6.2.1) and
(6.2.2) we get
λh(1+βj)/mi (H) ≤ λh(1+j)/mi (∆)
for all j = 0, 1, . . .. Taking j = km, k = 0, 1 . . . completes the proof.
6.2. DIRECT SUM AND PARTIAL TRACE 99

The above proof actually shows more eigenvalue inequalities.

+
Pα 6.2.4. Let H = [As,t ] ∈ Mαn be written in α × α Hermitian blocks in Mn and let
Corollary
∆ = s=1 As,s be its partial trace. Then, we have

1
λ1+βk (S) ≤ λ1+k1 (∆) + · · · + λ1+kβ (∆)
β

where k1 + · · · + kβ = βk and β is the smallest dyadic number such that α ≤ β.

Corollary 6.2.2 implies the following rearrangement inequality.

Corollary 6.2.5. Let {Si }αi=1 be a commuting family of Hermitian operators in Mn and let
T ∈ M+
n . Then,
Xα Xα
2
Si T Si ≤ T Si2 T
i=1 i=1

for all symmetric norms.

Proof. Define a matrix Z ∈ Mαn by

 
T S1
Z = XX ∗ =  ...  S1 T
 
··· Sα T .


T Sα

Hence Z = [T Si Sj T ] is positive and partitioned in Hermitian blocks in Mn , with diagonal

blocks T Si2 T , 1 ≤ i ≤ α. Thus, for all symmetric norms,

α
X
kZk ≤ T Si2 T
i=1

Since XX ∗ and X ∗ X have same symmetric norms for any rectangular matrix X, we infer
α
X α
X
Si T 2 Si ≤ T Si2 T
i=1 i=1

as claimed.

From Corollary 6.2.3 we similarly get the next one.

Corollary 6.2.6. Let {Si }αi=1 be a commuting family of Hermitian operators in Mn and let
T ∈ M+
n . Then,
α α
! !
X X
λ1+βk Si T 2 Si ≤ λ1+k T Si2 T
i=1 i=1

for all k = 0, . . . , n − 1, where β is the smallest dyadic number such that α ≤ β.

100 CHAPTER 6. PARTIAL TRACE

6.2.2 Around Rotfel’d inequality

Given two Hermitian matrices A, B in Mn and a concave function f (t) defined on the real
line,  
A+B f (A) + f (B)
Tr f ≥ Tr (6.2.3)
2 2
and, if further f (0) ≥ 0 and both A and B are positive semi-definite,

Tr f (A + B) ≤ Tr f (A) + Tr f (B). (6.2.4)

The first inequality goes back to von-Neumann in the 1920’s, the second is more subtle and
has been proved only in 1969 by Rotfel’d [88]. These trace inequalities are matrix versions of
obvious scalar inequalities. Theorem 6.2.1 yields a refinement of the Rotfel’d inequality for
families of positive operators {Ai }αi=1 by considering these operators as the diagonal blocks of
a partitioned matrix H as follows.

Corollary 6.2.7. Let H = [As,t ] ∈ M+αn be written in α × α Hermitian blocks in Mn . Then,

we have
α α
!
X X
Tr f As,s ≤ Tr f (H) ≤ Tr f (As,s )
s=1 s=1

for all concave functions f (t) on R+ such that f (0) ≥ 0.

Proof. Note that if these inequalities hold for a non-negative concave function f (t) with f (0) =
0, then they also hold for the function f (t) + c for any constant c > 0. Therefore it suffice to
consider concave functions vanishing at the origine. This assumption entails that

f (V AV ∗ ) = V f (A)V ∗ (6.2.5)

for all A ∈ M+
n and all isometries V ∈ Mm,n . By Lemma 6.1.1 we have a decomposition
α
X
H= Vs As,s Vs∗
s=1

for some isometries Vs ∈ Mαn,n . Inequality (6.2.4) then yields

α
X
Tr f (H) ≤ Tr f (Vs As,s Vs∗ )
s=1

and using (6.2.5) establishes the second inequality. To prove the first inequality, we use
Theorem 6.2.1. By completing H with some zero blocks (we still suppose f (0) = 0) we
may assume that α = β is dyadic. Thus we have a decomposition
β
1X
⊕m H = Vk (⊕m ∆) Vk∗
β
k=1

where {Vk }βk=1 is a family of isometries in Mmβn,mn and m = 2β . Inequality (6.2.3) then gives
β
m 1X
Tr f (⊕ H) ≥ Tr f (Vk (⊕m ∆) Vk∗ )
β
k=1
6.3. PROOF OF THEOREM 2.1 101

and using (6.2.5) we obtain

Tr f (⊕m H) ≥ Tr f (⊕m ∆) .
The proof is completed by dividing both sides by m.

Remark 6.2.8. The first inequality of Corollary 6.2.7 is actually equivalent to Corollary 6.2.2
by a well-known majorisation principle for convex/concave functions. The above proof does
not require this principle. The simplest case of Corollary 6.2.7 is the double inequality

f (a1 + · · · + an ) ≤ Tr f (A) ≤ f (a1 ) + · · · + f (an )

for all A ∈ M+
n with diagonal entries a1 , . . . , an .

Remark 6.2.9. A special case of Corollary 6.2.7 refines a well-known determinantal inequality.
Taking as a concave function on R+ , f (t) = log(1 + t), we obtain: Let A, B ∈ M+
n . Then, for
any Hermitian X ∈ Mn such that  
A X
H=
X B
is positive semi-definite, we have

det(I + A + B) ≤ det(I + H) ≤ det(I + A) det(I + B).

This was noted in [46].

6.3 Proof of Theorem 2.1

A Clifford algebra Cβ is the associative real algebra generated by β elements q1 , . . . , qβ satisfying
the canonical anticommutation relations qi2 = 1 and

qi qj + qj qi = 0

for i 6= j. This structure was introduced by Clifford in [53]. It turned out to be of great
importance in quantum theory and operator algebras, for instance see the survey [57]. From
the relation      
0 1 1 0 1 0 0 1
+ =0
1 0 0 −1 0 −1 1 0
we infer a representation of Cβ as a a real subalgebra of M2β = ⊗β M2 by mapping the
generators qj 7→ Qj , 1 ≤ j ≤ β, where
       
j−1 1 0 0 1 β−j 1 0
Qj = ⊗ ⊗ ⊗ ⊗ . (6.3.1)
0 −1 1 0 0 1

We use these matrices in the following proof of Theorem 6.2.1.

Proof. First, replace the positive block matrix H = [As,t ] where 1 ≤ s, t, ≤ β and all blocks
are Hermitian by a bigger one in which each block in counted 2β times :
h β i
G = [Gs,t ] := [I2β ⊗ As,t ] = ⊕2 As,t
102 CHAPTER 6. PARTIAL TRACE

where Ir stands for the identity of Mr . Thus G ∈ Mβ2β n is written in β-by-β blocks in M2β n .
Then perform a unitary congruence with the unitary W ∈ Mβ2β n defined as

β
M
W = {Qj ⊗ In } (6.3.2)
j=1

where Qj is given by (6.3.1), 1 ≤ j ≤ β. Thanks to the anticommutation relation for each pair
of summands in (6.3.2),

{Qj ⊗ In } {Ql ⊗ In } + {Ql ⊗ In } {Qj ⊗ In } = 0, j 6= l,

the block matrix (with W = W ∗ )

Ω := W GW ∗ = [Ωs,t ] (6.3.3)

satisfies the following : For 1 ≤ s < t ≤ β,

Ωs,t = −Ωt,s . (6.3.4)

Next, consider the reflexion matrix

 
1 1 1
J1 = √
2 1 −1

and define inductively for all integers p > 1, a reflexion

 
1 Jp−1 Jp−1
Jp = √ ,
2 Jp−1 −Jp−1

that is Jp = ⊗p J1 . Observe, that given any matrix S ∈ M2p , S = [si,j ], such that si,j = −si,j
for all i 6= j, the matrix
T = Jp SJp∗
has its diagonal entries tj,j all equal to the normalized trace 2−p Tr S. Indeed, letting Jp = [zi,j ],
!
X X
tj,j = zj,k sk,l zl,j
k l
X
= zj,k sk,l zl,j
k,l
X X
= zj,k sk,k zk,j + zj,k sk,l zl,j
k k6=l
X
−p
=2 Tr S + (sk,l zj,k zl,j + sl,k zl,j zk,j )
k<l
= 2−p Tr S.

Now, since we assume that β = 2p for some integer p, we may perform a unitary congruence
to the matrix Ω in (6.3.3) with the unitary matrix

Rp = Jp ⊗ I2β ⊗ In
6.4. COMMENTS 103

and, making use of (6.3.4) and the above property of Jp , we note that Rp ΩRp∗ has its β diagonal
blocks (Rp ΩRp∗ )j,j , 1 ≤ j ≤ β, all equal to the matrix D ∈ M2β n ,

β
1 X n 2β o
D= ⊕ As,s .
β s=1

Thanks to the decomposition of Lemma 6.1.1 and its obvious extension for β × β partitions,
there exist some isometries Uk ∈ Mβ2β n,2β n , 1 ≤ k ≤ β, such that

β
X
Ω= Uk DUk∗ .
k=1

β β
Since Ω is unitarily equivalent to ⊕2 H, that is Ω = V ∗ (⊕2 H)V for some unitary V ∈
Mβ2β n,2β n , we get
β
2β
X
⊕ H= V Uk DUk∗ V ∗
k=1

wich is the claim of Theorem 6.2.1 by setting V Uk =: Vk , 1 ≤ k ≤ β, as 2β = m, and

D = β1 ⊕m ∆.

6.4 Comments
6.4.1 Complex matrices and small partitions
If one uses isometries with complex entries, then, in case of partitions into a small number of
β × β blocks, the number m of copies in the direct sum ⊕m H and ⊕m ∆ can be reduced. For
β = 2, Theorem 6.1.2 shows that it suffices to take m = 1. For β = 3 or β = 4 the following
result holds [46].

Theorem 6.4.1. Let H = [As,t ] ∈ M+ βn (C) be written in Hermitian blocks in Mn (C) with
Pβ
β ∈ {3, 4} and let ∆ = s=1 As,s be its partial trace. Then,
4
1X
H ⊕H = Vk (∆ ⊕ ∆) Vk∗
4
k=1

for some isometries Vk ∈ M2βn,2n (C), k = 1, 2, 3, 4.

Likewise for Theorem 6.1.2, we must consider isometries with complex entries, even for a
full matrix H with real entries. The proof makes use of quaternions and thus confines to β ≤ 4.

6.4.2 Separability criterion

Let H and F be two finite dimensional Hilbert spaces that may be either real spaces, identified
to Rn and Rm , or complex spaces, identified to Cn and Cm . The space of operators on H,
denoted by B(H), is identified with the matrix algebra Mn (with real or complex entries
according the nature of H). A positive (semi-definite) operator Z on the tensor product space
104 CHAPTER 6. PARTIAL TRACE

H ⊗ F is said to be separable if it can be decomposed as a sum of tensor products of positive

operators,
Xk
Z= Aj ⊗ Bj (6.4.1)
j=1

where Aj ’s are positive operators on H and so Bj ’s are on F (the positivity assumption on

the Aj ’s and Bj ’s is essential, otherwise 6.4.1 is always possible for any Z). It is difficult in
general to determine if a given positive operator in the matrix algebra Mn ⊗ Mm is separable
or not, though some theoretical criteria do exist [70], [51]. The partial trace of Z with respect
to H is the operator acting on F,
k
X
TrH Z = (TrAj )Bj .
j=1

These notions have their own mathematical interest and moreover play a fundamental role in
the description of bipartite systems in quantum theory, see [86, Chapter 10], where the positive
operators act on complex spaces and are usually normalized with trace one and called states.
Thus a separable state is an operator of the type (6.4.1) with Tr Z = 1. The richness of the
mathematical theory of separable operators/states and their application in quantum physics
is apparent in many places in the literature, for instance in [70] and [2]. Nielsen and Kempe in
2001 proved a majorisation separability criterion [84]. It can be stated as the following norm
comparison.

Theorem 6.4.2. Let Z be a separable state on the tensor product of two finite dimensional
Hilbert spaces H and F. Then, for all symmetric norms,
kZk ≤ kTrH Zk .

Regarding B(H ⊗ F) as Mn (Mm ), an operator Z ∈ B(H ⊗ F) is written as a block-matrix

Z = [Zi,j ] with Zi,j ∈ Mm , 1 ≤ i, j ≤ n. The partial trace of Z with respect to H is then the
sum of the diagonal blocks,
Xn
TrH Z = Zj,j .
j=1

This observation makes obvious that Theorem 6.4.2 is a straightforward consequence of Corol-
lary 6.2.2 whenever the factor H is a real Hilbert space. Indeed, we then have
k
X
Z= Aj ⊗ Bj
j=1

where, for each index j, Aj ∈ Mn (R) and Bj is Hermitian in Mm , so that Aj ⊗ Bj can be

regarded as an element of Mn (Mm ) formed of Hermitian blocks.
From Corollary 6.2.3, we may complete the majorisation of Theorem 6.4.2, when a factor
is a real space with a few more eigenvalue estimate as stated in the next corollary.

Corollary 6.4.3. Let Z be a separable positive operator on the tensor product of two finite
dimensional Hilbert space H ⊗ F with a real factor H. Then,
λ1+βk (Z) ≤ λ1+k (TrH Z)
for all k = 0, . . . , dim F − 1, where β is the smallest dyadic number such that dim H ≤ β.
6.5. AROUND THIS ARTICLE 105

Similarly, from Corollary 6.2.4, we actually have a larger set of eigenvalue inequalities.

Corollary 6.4.4. Let Z be a separable positive operator on the tensor product of two finite
dimensional Hilbert space H ⊗ F with a real factor H. Then,
1
λ1+βk (Z) ≤ λ1+k1 (TrH Z) + · · · + λ1+kβ (TrH Z)
β
for all k = 0, . . . , dim F − 1, where k1 + · · · + kβ = βk and β is the smallest dyadic number
such that dim H ≤ β.

Of course, confining to real spaces is a severe restriction. It would be desirable to obtain

similar estimates for usual complex spaces. A related problem would be to obtain a decompo-
sition like in Theorem 6.2.1 for the class of partitioned matrices considered in the full form of
Hiroshima’s theorem:

If A = [As,t ] is a positive matrix partitioned in α × α blocks such that B = [Bs,t ] := [ATs,t ] is

positive too, then the majorisation of Corollary 6.2.2 holds.

Here X T means the transposed matrix. This statement extends Corollary 6.2.2 and implies
Theorem 6.4.2.

6.5 Around this article

A companion result of Lemma 6.1.1 is given in [45] :
 
A X
Lemma 6.5.1. Let be a positive matrix partitioned into four blocks in Mn . Then,
X∗ B
for some unitary U, V ∈ M2n ,
   A+B   
A X + Re X 0 ∗ 0 0
=U 2 U +V V∗
X∗ B 0 0 0 A+B
2 − Re X
.

This lemma is the main tool to obtain original estimates between the full matrix and its
partial trace A + B. These estimates involve the geometry of the numerical range W (X). Here
we state two theorems, the next chapter will develop this topic in more details. The main
theorem of [47] reads as follows.
 
A X
Theorem 6.5.2. Let be a positive matrix partitioned into four blocks in Mn . Sup-
X∗ B
pose that W (X) has the width ω. Then, for all symmetric norms,
 
A X
≤ kA + B + ωIk.
X∗ B

Here the width of W (X) is the smallest distance between two parallel straight lines such
that the strip between these two lines contains W (X). Hence the partial trace A + B may be
used to give an upper bound for the norms of the full block-matrix.
A lower bound is given in [41] in which the distance from 0 to W (X) contributes. We state
the main result of [41].
106 CHAPTER 6. PARTIAL TRACE
 
A X
Theorem 6.5.3. Let be a positive matrix partitioned into four blocks in Mn and let
X∗ B
d = dist(0, W (X)). Then, for all symmetric norms,
     
A X A+B A+B
≥ + dI ⊕ − dI .
X∗ B 2 2

Several consequences follow. We mention two corollaries

Corollary 6.5.4. For every positive matrix partitioned into four blocks of same size,
   
A X A+B
diam W − diam W ≥ 2d,
X∗ B 2

where d is the distance from 0 to W (X).

 
A X
Corollary 6.5.5. Let be a positive matrix partitioned into four blocks in Mn and
X∗ B
let d = dist(0, W (X)). Then,
( )
A+B 2
  
2 A X
det − d I ≥ det .
2 X∗ B

Letting X = 0, we recapture a basic property: the determinant is a log-concave map on

the positive cone of Mn . Hence Corollary 6.5.5 refines this property.
We will see in the next, short chapter several eigenvalue inequalities associated to Theorem
7.1.1.

6.6 References of Chapter 6

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York,
1996.

[28] J.-C. Bourin and F. Hiai, Norm and anti-norm inequalities for positive semi-definite
matrices, Internat. J. Math. 63 (2011), 1121-1138.

[31] J.-C. Bourin and E.-Y. Lee, Unitary orbits of Hermitian operators with convex or
concave functions, Bull. Lond. Math. Soc. 44 (2012), no. 6, 1085–1102.

[32] J.-C. Bourin and E.-Y. Lee, Decomposition and partial trace of positive matrices with
Hermitian blocks, Internat. J. Math. 24 (2013), no. 1, 1350010, 13 pp.

[41] J.-C. Bourin and E.-Y. Lee, Numerical range and positive block matrices, Bull. Aust.
Math. Soc. 103 (2021), no. 1, 69–77.

[45] J.-C. Bourin, E.-Y. Lee and M. Lin, On a decomposition lemma for positive semi-
definite block-matrices, Linear Algebra Appl. 437 (2012), 1906–1912.

[46] J.-C. Bourin, E.-Y. Lee and M. Lin, Positive matrices partitioned into a small number
of Hermitian blocks, Linear Algebra Appl. 438 (2013), no. 5, 2591–2598.
6.6. REFERENCES OF CHAPTER 6 107

[47] J.-C. Bourin, A. Mhanna, Positive block matrices and numerical ranges, C. R. Math.
Acad. Sci. Paris 355 (2017), no. 10, 1077–1081.

[51] K. Chen and L.-A. Wu, A matrix realignment method for recognizing entanglement,
Quantum Inf. Comput. 3 (2003), 193-202

[53] Clifford, Applications of Grassmann’ extensive algebra, Amer. Journ. Math. 1 (1878),
350-358.

[57] J. Dereziński, Introduction to representations of the canonical commutation and anti-

commutation relations, Lect. Note Phys. 695, 65-145 (2006), Springer.

[70] M. Horodecki, P. Horodecki, R. Horodecki, Separability of mixed states: necessary and

sufficient conditions, Phys. Lett. A 223 (1996) l-8.

[84] M. A. Nielsen and J. Kempe, Separable states are more disordered globally than locally,
Phys. Rev. Lett. 86 (2001) 5184-5187.

[86] D. Petz, Matrix Analysis with some applications, <http://www.math.hu/petz>.

[88] S. Ju. Rotfel’d, The singular values of a sum of completely continuous operators,
Topics in Mathematical Physics, Consultants Bureau, Vol. 3 (1969) 73-78.
108 CHAPTER 6. PARTIAL TRACE
Chapter 7

Block matrices and numerical range

For sake of completeness the proof of Theorem 7.1.1 is given in Section 7.4 “Around this
article”.

Eigenvalue inequalities for positive block matrices with

the inradius of the numerical range [44]
Abstract. We prove the operator norm inequality, for a positive matrix partitioned into four
blocks in Mn ,  
A X
≤ kA + Bk∞ + δ(X),
X∗ B ∞
where δ(X) is the diameter of the largest possible disc in the numerical range of X. This
shows that the inradius ε(X) := δ(X)/2 satisfies ε(X) ≥ kXk∞ − k(|X ∗ | + |X|)/2k∞ . Several
eigenvalue inequalities are derived. In particular, if X is a normal matrix whose spectrum lies
in a disc of radius r, the third eigenvalue of the full matrix is bounded by the second eigenvalue
of the sum of the diagonal block,
 
A X
λ3 ≤ λ2 (A + B) + r.
X∗ B

We think that r is optimal and we propose a conjecture related to a norm inequality of Hayashi.
Keywords. Numerical range, Partitioned matrices, eigenvalue inequalities.
2010 mathematics subject classification. 15A60, 47A12, 15A42, 47A30.

7.1 Introduction
Positive matrices partitioned into four blocks play a central role in Matrix Analysis, and in
applications, for instance quantum information theory. A lot of important theorems deal
with these matrices. Some of these results give comparison between the full matrix and its
diagonal blocks, in particular the sum of the diagonal blocks (the partial trace in the quantum
terminology). This note focuses on a recent result of Bourin and Mhana [47], involving the
numerical range of the offdiagonal block. Recall that a symmetric norm k · k on M2n means a
unitarily invariant norm. It induces a symmetric norm on Mn in an obvious way. The Schatten
p-norms k · kp , 1 ≤ p ≤ ∞, and the operator norm (p = ∞) are classical examples of symmetric
norms. The main result of [47] reads as follows.

109
110 CHAPTER 7. BLOCK MATRICES AND NUMERICAL RANGE
 
A X
Theorem 7.1.1. Let be a positive matrix partitioned into four blocks in Mn . Sup-
X∗ B
pose that W (X) has the width ω. Then, for all symmetric norms,
 
A X
≤ kA + B + ωIk.
X∗ B

Here I stands for the identity matrix, W (X) denotes the numerical range of X, and the
width of W (X) is the smallest distance between two parallel straight lines such that the strip
between these two lines contains W (X). If ω = 0, that is W (X) is a line segment, Theorem
1.1 was first proved by Mhanna [80]. Recently [41], Theorem 7.1.1 has been completed with
the reversed inequality
   A+B 
A X 2 + dI 0
≥ A+B .
X∗ B 0 2 − dI

where d := min{|z| : z ∈ W (X)} is the distance from 0 to W (X). Several applications were
derived.
Some equality cases in Theorem 7.1.1 occur for the operator norm k · k∞ with the following
block matrices, where a, b are two arbitrary nonnegative real numbers.
   
a 0 0 a
 0 b
  b 0  .


 0 b b 0 
a 0 0 a
 
0 b
This follows from the fact that W has the width 2 ||a| − |b||.
a 0
Though Theorem 7.1.1 is sharp for the operator norm, a subtle improvement is possible.
This is our concern in the next section. Once again, a geometric feature of W (X) will con-
tribute: its inradius. Our approach leads to a remarkable list of eigenvalue that cannot be
derived from the norm inequalities of Theorem 7.1.1. The last section is devoted to some
related operator norm inequalities, in particular we will discuss a property due to Hayashi
(2019) and propose a conjecture.

7.2 Eigenvalue inequalities

We define the indiameter δ(Λ) of a compact convex set Λ ⊂ C as the diameter of the largest
possible disc in Λ. For matrices X ∈ Mn , we shorten δ(W (X)) =: δ(X). Recall that the
numerical range of a two-by-two matrix is an elliptical disc (or a line segment, or a single
point).
A matrix X ∈ Mn is identified as an operator on Cn . If S is a subspace of Cn , we denote
by XS the compression of X onto S. We then define the elliptical width of X as

δ2 (X) := sup δ(XS ).

dim S=2

Of course δ2 (X) ≤ δ(X) ≤ ω where ω still denotes the width of W (X). If X is a contraction,
then δ2 (X) ≤ 1, while δ(X) may be arbitrarily close to 2 (letting n be large enough). We state
our main result.
7.2. EIGENVALUE INEQUALITIES 111
 
A X
Theorem 7.2.1. Let be a positive matrix partitioned into four blocks in Mn . Then,
X∗ B
for all j ∈ {0, 1, . . . , n − 1},
 
A X
λ1+2j ≤ λ1+j (A + B) + δ2 (X).
X∗ B

Here λ1 (S) ≥ · · · ≥ λd (S) stand for the eigenvalue of any Hermitian matrix S ∈ Md . If
we denote by λ↑1 (S) ≤ · · · ≤ λ↑d (S) these eigenvalues arranged in the increasing order, then
Theorem 7.2.1 reads as
 
↑ A X
λ2k ≤ λ↑k (A + B) + δ2 (X).
X∗ B

for all k ∈ {1, 2, . . . , n}.

The case j = 0 in Theorem 7.2.1 improves Theorem 7.1.1 for the operator norm. We may
consider that Theorem 7.2.1 is trivial for j = n − 1. Indeed, using the decomposition [31,
Lemma 3.4],
     
A X A 0 ∗ 0 0
=U U +V V ∗,
X∗ B 0 0 0 B
for some unitary matrices U, V ∈ M2n , we obtain from Weyl’s inequality [13, p. 62],
     
A X A 0 0 0
λ2n−1 ≤ λn + λn
X∗ B 0 0 0 B
= λn (A) + λn (B)
≤ λn (A + B).

We turn to the proof of the theorem.

Proof. We first consider the case j = 0. We may assume that the norm of the block matrix is
strictly greater than the norms of its two diagonal blocks A and B, otherwise the statement is
trivial. Hence we have two nonzero (column) vectors h1 , h2 ∈ Cn such that kh1 k2 + kh2 k2 = 1
and       
A X A X ∗ ∗

A X h1
λ1 = = h1 h2 .
X∗ B X∗ B ∞ X∗ B h2
Therefore, denoting by E1 and E2 the rank one projections corresponding to the one dimen-
sional subspaces spanned by h1 and by h2 , we have
     
A X E1 0 A X E1 0
=
X∗ B ∞ 0 E2 X ∗ B 0 E2 ∞

Hence, denoting by F a rank two projection such that E1 ≤ F and E2 ≤ F , we have

     
A X F 0 A X F 0
=
X∗ B ∞ 0 F X∗ B 0 F ∞
 
F AF F XF
= .
F X ∗ F F BF ∞
112 CHAPTER 7. BLOCK MATRICES AND NUMERICAL RANGE

So, letting S denote the range of F , we have

   
A X AS XS
= .
X∗ B ∞ XS∗ BS ∞

Hence applying Theorem 7.1.1 for the operator norm, we obtain

 
A X
≤ kAS + BS k∞ + ε
X∗ B ∞

where ε is the width of W (XS ). Since W (XS ) is an elliptical disc (as XS acts on a two-
dimensional space), its width equals to its indiameter, hence ε ≤ δ2 (X), and since

kAS + BS k∞ = k(A + B)S k∞ ≤ kA + Bk∞ = λ1 (A + B),

the proof for j = 0 is complete.

We turn to the general case, j = 1, . . . , n − 1. By the min-max principle,
    !
A X A X
λ1+2j ≤ inf λ1
X∗ B dim S=n−j X ∗ B S⊕S
 
AS XS
= inf λ1 ,
dim S=n−j XS∗ BS

hence, from the first part of the proof,

 
A X
λ1+2j ≤ inf λ1 (AS + BS ) + δ2 (X)
X∗ B dim S=n−j

= λ1+j (A + B) + δ2 (X)

which is the desired claim.

If X ∈ Mn , we denote by dist(X, CI) the k · k∞ -distance from X to CI. Thus, for a scalar
perturbation of a contraction, X = λI + C for some contraction C ∈ Mn and some λ ∈ C, we
have dist(X, CI) ≤ 1.
 
A X
Corollary 7.2.2. Let be a positive matrix partitioned into four blocks in Mn . Then,
X∗ B
for all j ∈ {0, 1, . . . , n − 1},
 
A X
λ1+2j ≤ λ1+j (A + B) + dist(X, CI).
X∗ B

Proof. For any subspace S ⊂ C n , we have

dist(X, CI) ≥ dist(XS , CIS ).

If S has dimension 2, then

dist(XS , CIS ) ≥ δ (W (XS )) .
Therefore dist(X, CI) ≥ δ2 (X) and Theorem 7.2.1 completes the proof.
7.3. NORM INEQUALITIES 113

Corollary 7.2.3. Let A, B ∈ Mn . Then, for every j ≥ 0 such that 1 + 2j ≤ n,

λ1+2j (A∗ A + B ∗ B) ≤ λ1+j (AA∗ + BB ∗ ) + δ2 (AB ∗ )

Proof. Note that

λ1+2j (A∗ A + B ∗ B) = λ1+2j (T ∗ T ) = λ1+2j (T T ∗ )
AA∗ AB ∗
   
A
with T = and T T ∗ = so that Theorem 7.2.1 yields the desired claim.
B BA∗ BB ∗
 
A N
Corollary 7.2.4. Let be a positive matrix partitioned into four blocks in Mn . If N
N∗ B
is normal and its spectrum is contained in a disc of radius r, then,
 
A N
λ1+2j ≤ λ1+j (A + B) + r.
N∗ B
for all j = 0, 1, . . . , n − 1.

Proof. Corollary 7.2.4 is a special case of corollary 7.2.2, as N = λI + R, where λ is the center
of the disc of radius r containing the spectrum of N , and kN k∞ ≤ r.

Question 7.2.5. Fix r > 0 and ε > 0. Can we find (with n largeenough) a normal matrix
A N
N with spectrum in a disc of radius r and a positive block matrix such that
N∗ B
 
A N
λ1+2j ≥ λ1+j (A + B) + r − ε
N∗ B
for some j ∈ {0, . . . , n − 1} ? Is it true for for j = 0 ?

7.3 Norm inequalities

Corollary 7.2.4 with j = 0 reads as follows.
 
A N
Corollary 7.3.1. Let be a positive matrix partitioned into four blocks in Mn . If N
N∗ B
is normal and its spectrum is contained in a disc of radius r, then,
 
A N
≤ kA + Bk∞ + r.
N∗ B ∞

We do not know wether the constant r is sharp or not (Question 7.2.5). If n = 2, we can
replace r by 0 as the numerical range of N is then a line segment. If n = 3 there are some
simple examples with N = U unitary such that
 
A U
> kA + Bk∞ .
U∗ B ∞
See Hayashi’s example in the discussion of [H, Problem 3] and the interesting study and
examples in [G] where we further have A + B = kI for some scalars k. The next result is due
to Hayashi [H, Theorem 2.5].
114 CHAPTER 7. BLOCK MATRICES AND NUMERICAL RANGE

Theorem 7.3.2. Suppose that X ∈ Mn is invertible with n distinct singular values. If the
inequality  
A X
≤ kA + Bk∞ .
X∗ B ∞
holds for all positive block-matrix with X as off-diagonal block, then X is normal.

Theorem 7.3.2 and Theorem 7.1.1 suggest a natural conjecture. If W (T ) is line segment,
then T is a so-called essentially Hermitian matrix.
Conjecture 7.3.3. Let X ∈ Mn . If the inequality
 
A X
≤ kA + Bk∞
X∗ B ∞
holds for all positive block-matrix with X as off-diagonal block, then X is essentially Hermitian.

If we replace the operator norm by the Frobenius (or Hilbert-Schmidt) norm k · k2 then the
following characterization holds.
Proposition 7.3.4. Let X ∈ Mn . Then, the inequality
 
A X
≤ kA + Bk2
X∗ B 2
holds for all positive block-matrix with X as off-diagonal block if and only if X is normal.

Proof. Suppose that X is normal. To prove the inequality, squaring both side, it suffices to
establish the trace inequality
Tr X ∗ X ≤ Tr AB. (7.3.1)
Note that X = A1/2 KB 1/2 , for some contraction K. Recall that, for all symmetric norms on
Mn , and any normal matrix N Mn , decomposed as N = ST , we have kN k ≤ kT Sk. Therefore
kXk = kA1/2 KB 1/2 k ≤ kKB 1/2 A1/2 k.
Squaring this inequality with the Frobenius norm yields thedesired inequality (7.3.1).
|X ∗ | X
Suppose that X is nonnormal, and note that is positive semidefinite and sat-
X ∗ |X|
isfies  ∗  2
|X | X
= 4k|X|k22
X ∗ |X| 2
while
k|X ∗ | + |X|k22 = 2k|X|k22 + 2Tr |X||X ∗ |
In the Hilbert space (Mn , k · k2 ), the assumption
k|X|k2 = k|X ∗ |k2 , 6 |X ∗ |
|X| =
ensures strict inequality in the Cauchy-Schwarz inequality
Tr |X||X ∗ | < kXk22 .
Therefore
2
|X ∗ | X
 
∗
k|X | + |X|k22 <
X ∗ |X| 2
and this completes the proof.
7.3. NORM INEQUALITIES 115

Proposition 7.3.4 suggests a question: for which p ∈ [1, ∞], the schatten p-norm inequality
 
A N
≤ kA + Bkp
N∗ B p

holds for any positive partitioned matrices with a normal off-diagonal block N ?

Corollary 7.3.5. Let H, K, X ∈ Mn be Hermitian. If X is invertible and HK is a scalar

perturbation of a contraction, then,

XH 2 X + X −1 K 2 X −1 ∞
≤ HX 2 H + KX −2 K ∞
+ 1.

Proof. We apply Corollary 7.2.3 with j = 0 and A = HX, B = KX −1 , to get

XH 2 X + X −1 K 2 X −1 ∞
≤ HX 2 H + KX −2 K ∞
+ δ2 (HK)

Since (HK)S is a scalar perturbation of a contraction acting on a space of dimension 2,

necessarily δ2 (HK) ≤ 1.

For a normal operator, the numerical range is the convex hull of the spectrum. For a non
normal operator X, several lower bounds for the indiameter of W (X) can be obtained from
the left and right modulus |X ∗ | and |X|.

Corollary 7.3.6. Let X ∈ Mn and let f (t) and g(t) are two nonnegative functions defined on
[0, ∞) such that f (t)g(t) = t2 . Then,

δ2 (X) ≥ kf (|X|) + g(|X|)k∞ − kf (|X ∗ |) + g(|X|)k∞ .

Proof. First, observe that we have a function h(t) defined on [0, ∞) such that

f (t) = th2 (t1/2 ), g(t) = th−2 (t1/2 ), (7.3.2)

and h(t) > 0 for all t ≥ 0 (we may, for instance, set h(0) = 1). Hence h(T ) is invertible for
any positive T , and from the polar decomposition

X = |X ∗ |1/2 U |X|1/2

with a unitary factor U , we infer the factorization

X = |X ∗ |1/2 h(|X ∗ |1/2 )U |X|1/2 h−1 (|X|1/2 ).

Thus X = AB ∗ where A = |X ∗ |1/2 h(|X ∗ |1/2 ) and B ∗ = U |X|1/2 h−1 (|X|1/2 ). Therefore
Corollary 7.2.3 yields

|X ∗ |h2 (|X ∗ |1/2 ) + U |X|h−2 (|X|1/2 )U ∗ ≤ |X|h2 (|X|1/2 ) + |X|h−2 (|X|1/2 )U ∗ + δ2 (X)
∞ ∞

Using (7.3.2) and the fact that ϕ(|X ∗ |) = U ϕ(|X|)U ∗ for any function ϕ(t) defined on [0, ∞),
the proof is complete.

The following special case shows that Corollary 7.3.6 is rather optimal.
116 CHAPTER 7. BLOCK MATRICES AND NUMERICAL RANGE

Corollary 7.3.7. If X ∈ Mn has a numerical range of inradius ε(X), then, for all a ∈ C,

|X − aI| + |X ∗ − aI|
ε(X) ≥ kX − aIk∞ − .
2 ∞

If X ∈ M2 and a = τ is the normalized trace of X, then this inequality is an equality.

Proof. Applying Corollary 7.3.6 with X − aI and f (t) = g(t) = t yields the inequality. If
X ∈ M2 , then X is unitarily equivalent to
 
τ y
.
x τ

So
|X − τ I| + |X ∗ − τ I|
   
|x| 0 1 |x| + |y| 0
kX − τ Ik∞ − = −
2 ∞
0 |y| ∞ 2 0 |x| + |y| ∞
= ||x| − |y||
= ε(X)

establishing the desired equality.

The special case f (t) = g(t) = t in Corollary 7.3.6 seems important, we record it as a
proposition:

Proposition 7.3.8. The elliptical width of the numerical range of X ∈ Mn satisfies

δ2 (X) ≥ 2kXk∞ − k|X| + |X ∗ |k∞ .

In particular, the inradius ε(X) of the numerical range of X satisfies

ε(X) ≥ kXk∞ − k(|X| + |X ∗ |)/2k∞ .

Remark 7.3.9. Our results still hold for operators on infinite dimensional separable Hilbert
space (assuming in Corollary 7.3.6 that f (t) and g(t) are Borel functions).

7.4 Around this article

We give the proof of Theorem 7.1.1.

Proof. (Theorem 7.1.1) By using the unitary congruence implemented by

 iθ 
e 0
0 I

we see that our block matrix is unitarily equivalent to

eiθ X
 
A
.
e−iθ X ∗ B
7.5. REFERENCES OF CHAPTER 7 117

As W (eiθ X) = eiθ W (X), by choosing the adequate θ and replacing X by eiθ X, we may and
do assume that W (X) lies in a strip S of width ω and parallel to the imaginary axis,
S = { x + iy : y ∈ R, r ≤ x ≤ r + ω}.
The projection property for the real part Re W (X) = W (Re X), then ensures that
rI ≤ Re X ≤ (r + ω)I. (7.4.1)
Now we use the decomposition [45, Corollary 2.1] derived from (8.2.1),
   A+B   
A X + Re X 0 ∗ 0 0
=U 2 U +V V∗ (7.4.2)
X∗ B 0 0 0 A+B2 − Re X
for some unitaries U, V ∈ M2n . Note that the two matrices in the right hand side of (7.4.2)
are positive since so are
     
  A X I   A X I
I I and I −I .
X∗ B I X ∗ B −I
Combining (7.4.1) and (7.4.2) yields
   A+B   
A X + (r + ω)I 0 ∗ 0 0
≤U 2 U +V A+B V∗
X∗ B 0 0 0 2 − rI
where the two matrices of the right hand side are positive. From each Ky Fan k-norm,
k = 1, 2, . . . , 2n, we then have
 
A X A+B A+B
∗ ≤ + (r + ω)I + − rI
X B (k) 2 (k) 2 (k)

= kA + B + ωIk(k) .
The Ky Fan principle then guarantees that this inequality hold for all symmetric norms.

7.5 References of Chapter 7

[G] M. Gumus, J. Liu, S. Raouafi, T-Y. Tam, Positive semi-definite 2 × 2 block matrices
and norm inequalities, Linear Algebra Appl. 551 (2018), 83–91.
[H] T. Hayashi, On a norm inequality for a positive block-matrix, Linear Algebra Appl.
566 (2019), 86–97.
[41] J.-C. Bourin and E.-Y. Lee, Numerical range and positive block matrices, Bull. Aust.
Math. Soc. 103 (2021), no. 1, 69–77.
[44] J.-C. Bourin and E.-Y. Lee, Eigenvalue inequalities for positive block matrices with
the inradius of the numerical range, preprint
[45] J.-C. Bourin, E.-Y. Lee and M. Lin, On a decomposition lemma for positive semi-
definite block-matrices, Linear Algebra Appl. 437 (2012), 1906–1912.
[47] J.-C. Bourin, A. Mhanna, Positive block matrices and numerical ranges, C. R. Math.
Acad. Sci. Paris 355 (2017), no. 10, 1077–1081.
[80] A. Mhanna, On symmetric norm inequalities and positive definite block-matrices, Math.
Inequal. Appl. 21 (2018), no. 1, 133–138.
118 CHAPTER 7. BLOCK MATRICES AND NUMERICAL RANGE
Chapter 8

Block matrices and Pythagoras

A Pythagorean theorem for partitioned matrices [42]

Abstract. We establish a Pythagorean theorem for the absolute values of the blocks of a
partitioned matrix. This leads to a series of remarkable operator inequalities. For instance, if
the matrix A is partitioned into three blocks A, B, C, then

|A|3 ≥ U |A|3 U ∗ + V |B|3 V ∗ + W |C|3 W ∗ ,

√
3|A| ≥ U |A|U ∗ + V |B|V ∗ + W |C|W ∗ ,
for some isometries U, V, W , and

µ24 (A) ≤ µ23 (A) + µ22 (B) + µ21 (C)

where µj stands for the j-th singular value. Our theorem may be used to extend a result by
Bhatia and Kittaneh for the Schatten p-noms and to give a singular value version of Cauchy’s
Interlacing Theorem.
Keywords. Partitioned matrices, functional calculus, matrix inequalities.
2010 mathematics subject classification. 15A18, 15A60, 47A30.

8.1 Introduction
Let Md denote the space of d-by-d matrices. If A ∈ Md , the polar decomposition holds,

A = U |A| (8.1.1)

where |A| ∈ Md is positive semi-definite and U ∈ Md is a unitary matrix. The matrix |A| is
called the absolute value of A, and its eigenvalues are the singular values of A. The absolute
value can be defined for d × d′ matrices A ∈ Md,d′ as a positive matrix |A| ∈ Md′ , and the
factor U in (8.1.1) is an isometry (d ≥ d′ ) or a coisometry (d < d′ ).
If A is partitioned in some number of rectangular blocks, say four blocks A, B, C, D, it is
of interest to have a relation between the absolute value |A| and the absolute values of the
blocks. By using the standard inner product of Md,d′ , we immediately have the trace relation

Tr |A|2 = Tr |A|2 + Tr |B|2 + Tr |C|2 + Tr |D|2 .

119
120 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

This note aims to point out a much stronger Pythagorean theorem, Theorem 8.2.1, and several
consequences. This result holds for many partitionings of A, not only when A is a block matrix
partitioned into p × q blocks. For instance, one may consider the matrix
 
a1 a2 b1 b2 b3
a3 a4 b4 b5 b6 
 
a5 a6
A= c1 c2 d1 

a7 a8 c3 c4 d2 
a9 a10 c5 c6 d3

partitioned into four obvious blocks A, B, C, D.

If A is partitioned into r blocks Ak ∈ Mnk ,mk , we write
r
[
A= Ak = A1 ∪ · · · ∪ Ar (8.1.2)
k=1

where we can use the = sign if one considers Ak not only as an element of Mnk ,mk but also as
a submatrix of A with its position in A.
We say that the partitioning (8.1.2) is column compatible, or that A is partitioned into
column compatible blocks if for all pairs of indexes k, l, either Ak and Al lie on the same set
of columns of A, or Ak and Al lie on two disjoint sets of columns of A. Similarly, (8.1.2) is
row compatible, if for all pairs of indexes k, l, either Ak and Al lie on the same set of rows of
A, orAk and Al lie on two disjoint sets of rows of A.
Our Pythagorean Theorem 8.2.1 will be stated for row or column compatible blocks. An
application is a Theorem of Bhatia and Kittaneh for the Schatten p-norms (Corollary 8.2.5).
Another application is an inequality for the singular values of compression onto hyperplanes.
A matrix A ∈ Md is an operator on Cd . Given a hyperplane S of Cd , we have a unit vector
h such that S = h⊥ , that is x ∈ S ⇐⇒ hh, xi = h∗ x = 0. The compression AS of A onto
S is the operator acting on S defined as the restriction of EA to S where E stands for the
(orthogonal) projection onto S. Theorem 8.2.1 entails a bound for the singular values of AS
in terms of those of A. These results are given in Section 3; we state a special case in the
following corollary. Let µj denote the j-th singular value arranged in nonincreasing order.

Corollary 8.1.1. Let A ∈ Md be a normal matrix and let S be a hyperplane of Cd orthogonal

to the unit vector h. Set β = kAhk2 − |hh, Ahi|2 . Then, for j = 1, . . . , d − 1,

µ2j (A) ≥ µ2j (AS ) ≥ µ2j+1 (A) − β.

We discuss the case of four and five blocks in Section 4. For four blocks, our Pythagorean
theorem entails an interesting inequality stated in the next corollary.

Corollary 8.1.2. Let A ∈ Md,d′ be partitioned into four blocks A, B, C, D. Then, there exist
some isometries U, V, W, X of suitable sizes such that

2|A| ≥ U |A|U ∗ + V |B|V ∗ + W |C|W ∗ + X|D|X ∗ .

The last section is devoted to several other operator inequalities such as the first inequality
in the abstract.
8.2. A PYTHAGOREAN THEOREM 121

8.2 A Pythagorean theorem

Theorem 8.2.1. Let A ∈ Md,d′ be partitioned into r row or column compatible blocks Ak ∈
Mnk ,mk . Then, there exist some isometries Uk ∈ Md′ ,mk such that
r
X
|A|2 = Uk |Ak |2 Uk∗ .
k=1

Recall that U ∈ Md′ ,m , m ≤ d′ , is an isometry if U ∗ U = 1m , the identity on Cm . If

A ∈ Md,1 , then the theorem reads as Pythagoras’ Theorem.

Proof. Consider a positive matrix in Mn+m partitioned as

 
A X
X∗ B

with diagonal blocks A ∈ Mn and B ∈ Mm . By [31, Lemma 3.4] we have two unitary matrices
U, V ∈ Mn+m such that
     
A X A 0 ∗ 0 0
=U U +V V ∗, (8.2.1)
X∗ B 0 0 0 B

equivalently,  
A X
= U1 AU1∗ + U2 BU2∗
X∗ B
for two isometry matrices U1 ∈ Mn+m,n and U2 ∈ Mn+m,m . An obvious iteration of (8.2.1)
shows that, given a positive block matrix in Mm partitioned into p × p blocks,

B = (Bi,j )1≤i,j≤p ,

with square diagonal blocks Bi,i ∈ Mni and n1 + · · · + np = m, we have the decomposition
p
X
B= Ui Bi,i Ui∗ (8.2.2)
i=1

for some isometries Ui ∈ Mm,ni .

We use (8.2.2) to prove the theorem. Consider first the column compatible case. Thus we
have a partitioning into p block columns,

A = C1 ∪ · · · ∪ Cp , (8.2.3)

and each block Ak belongs to one block column Cq . By relabelling the Ak ’s if necessary, we
may assume that we have p integers 1 = α1 < α2 < · · · < αp < r such that

Cq = Aαq ∪ · · · ∪ Aαq+1 −1 , 1 ≤ q < p, and Cp = Aαp ∪ · · · ∪ Aαr .

We also have a partitioning into p block rows,

A∗ = C∗1 ∪ · · · ∪ C∗p , (8.2.4)

122 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

and multiplying (8.2.4) and (8.2.3) we then obtain a block matrix for A∗ A = |A|2 ∈ Md′ ,

|A|2 = (C∗i Cj )1≤i,j≤p .

By the decomposition (8.2.2) we have

p
X
2
|A| = Ui C∗i Ci Ui∗
i=1

for some isometries Ui ∈ Md′ ,ni , where ni is the number of columns of Ci . Hence, with the
convention αp+1 := r + 1,
Xp αi+1
X−1
|A|2 = Ui A∗k Ak Ui∗
i=1 k=αi

establishing the theorem for a column compatible partitioning.

Now, we turn to the row compatible case. Thus we have a partitioning into p block rows,

A = R1 ∪ · · · ∪ Rp , (8.2.5)

and each block Ak belongs to one block row Rq and, as in the column compatible case, we
may assume that we have p integers 1 = α1 < α2 < · · · < αp < r such that

Rq = Aαq ∪ · · · ∪ Aαq+1 −1 , 1 ≤ q < p, and Rp = Aαp ∪ · · · ∪ Aαr .

We also have a partitioning into p block columns,

A∗ = R∗1 ∪ · · · ∪ R∗p (8.2.6)

Mutiply (8.2.6) and (8.2.5) and note that

p
X
2
|A| = R∗l Rl . (8.2.7)
l=1

with p block matrices in Md′ , (l = 1, . . . , p),

R∗l Rl = (A∗i Aj )αl ≤i,j<αl+1 (8.2.8)

where we still use αp+1 := r + 1. Applying the decomposition (8.2.2) to the block matrices
(8.2.8) yields
αl+1 −1
X
∗
Rl Rl = Ui |Ai |2 Ui∗
i=αl

for some isometries Ui of suitable sizes, and combining with (8.2.7) completes the proof.

Denote by µ1 (S) ≥ µ2 (S) ≥ · · · the singular values of a matrix S ∈ Mn,m . This list is often
limited to min{n, m} elements, however we can naturally define µk (S) = 0 for any index k
larger than min{n, m}. Given two matrices of same size, a classical inequality of Weyl asserts
that
µj+k+1 (S + T ) ≤ µj+1 (S) + µk+1 (T )
for all nonnegative integers j and k. This inequality and Theorem 8.2.1 entail the next corol-
lary.
8.2. A PYTHAGOREAN THEOREM 123

Corollary 8.2.2. Let A ∈ Md,d′ be partitioned into r row or column compatible blocks Ak ∈
Mnk ,mk . Then, for all nonnegative integers j1 , j2 , . . . , jr ,
r
X
µ2j1 +j2 +···+jr +1 (A) ≤ µ2jk +1 (Ak ).
k=1

A special case of this inequality is given in the abstract for three blocks and j1 = 2, j2 = 1,
j3 = 0.
Since any partitioning into three blocks is row or column compatible we have the next
corollary.
Corollary 8.2.3. Let A ∈ Md,d′ be partitioned into three blocks A, B, C. Then, there exist
some isometries U, V, W of suitable sizes such that

|A|2 = U |A|2 U ∗ + V |B|2 V ∗ + W |C|2 W ∗ .

By using the triangle inequality for the Schatten p-norms we have the trace inequality
1/p 1/p 1/p 1/p
Tr |A|2p ≤ Tr |A|2p + Tr |B|2p + Tr |C|2p
   
, p ≥ 1, (8.2.9)

equivalently
kAk2q ≤ kAk2q + kBk2q + kCk2q (8.2.10)
for all Schatten q-norms, q ≥ 2.
Theorem 8.2.1 entails another interesting relation between the blocks of a partitioned matrix
and the full matrix.
Corollary 8.2.4. Let A ∈ Md,d′ be partitioned into r row P
or column compatible blocks Ak ∈
Mnk ,mk . Then, for some isometries Vj ∈ Mm,d′ , with m = rk=1 mk ,
r r
M 1X
|Ak |2 = Vj |A|2 Vj∗ .
r
k=1 j=1

Proof. From Theorem 8.2.1 and the main result of [35] we have
r r
M 1X
Uk |Ak |2 Uk = Wj |A|2 Wj∗
r
k=1 j=1

for some isometries Uk ∈ Md′ ,mk and some isometries Wj ∈ Mrd′ ,d′ . Since
r
( r )
M M
|Ak |2 = C Uk |Ak |2 Uk C ∗
k=1 k=1

for some contraction C ∈ Mm,rd′ , we infer

r r
M
2 1X
|Ak | = CWj |A|2 Wj∗ C ∗ .
r
k=1 j=1

If |A| is invertible, then, taking trace, the above equality ensures that contractions CWj satisfy
Wj∗ C ∗ CWj = 1d′ for all j. Hence the result is proved with Vj = CWj . The general case follows
by a limit argument.
124 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

We are in a position to estimate the Schatten norms of the blocks with the full matrix.
The following corollary was first obtained by Bhatia and Kittaneh [17] in case of a matrix
partitioned into n × n blocks.
Corollary 8.2.5. Let A ∈ Md,d′ be partitioned into r row or column compatible blocks Ak ∈
Mnk ,mk . Then, for all Schatten q-norms, q ≥ 2,
r r
2
−1
X X
r q kAk k2q ≤ kAk2q ≤ kAk k2q
k=1 k=1
These two inequalities are reversed for 2 > q > 0.

Proof. For p := q/2 ≥ 1, the second inequality contains (8.2.10) and immediately follows from
Theorem 8.2.1 and the triangle inequality for the Schatten p-norms. Corollary 8.2.4 gives
k|A|2 kp ≥ |A1 |2 ⊕ · · · ⊕ |Ar |2 p

and since the concavity of t 7→ t1/p entails

1/p 1
−1
|A1 |2 ⊕ · · · ⊕ |Ar |2 = k|A1 |2 kpp + · · · k|Ar |2 kpp k|A1 |2 kp + · · · + k|Ar |2 kp


p
≥ rp
we get the first inequality. These inequalities are reversed for 0 < p < 1.

Corollary 8.2.4 is relevant to Majorisation Theory. We take this opportunity to point out
an interesting fact about majorisation in the next proposition. Though this result might be
well-known to some experts, it does not seem to be in the literature. Let A, B ∈ M+ n , the
positive semi-definite cone of Mn . The majorisation A ≺ B means that
k
X k
X
µj (A) ≤ µj (B)
j=1 j=1

for all k = 1, 2, . . . n, with equality for k = n. The majorisation A ≺ B is equivalent to

n
X
A= αi Ui BUi∗
i=1

for some unitary matrices Ui ∈ Mn and weights αi ≥ 0 with ni=1 αi = 1. This can be easily
P
derived from Caratheodory’s theorem [94]. A more accurate statement holds.
Proposition 8.2.6. Let A, B ∈ M+
n , A ≺ B.Then, for some unitary matrices Ui ∈ Mn ,
n
1X
A= Ui BUi∗ .
n
i=1

Proof. By the Schur-Horn Theorem, we may assume that A is the diagonal part of B. Then
we use the simple idea of Equation (2) in the nice paper of Bhatia [14].

Note that Corollary 8.2.4 may be restated as

r r
M
2 1X
|Ak | = Uj (|A|2 ⊕ O)Uj∗ .
r
k=1 j=1

for some unitary matrices Uj and some fixed zero matrix O. Hence we have an average of r
matrices in the unitary orbit of |A|2 ⊕ O, this number r being (much) smaller than the one
given by Proposition 8.2.6, d = m1 + · · · + mr .
8.3. COMPRESSION ONTO A HYPERPLANE 125

8.3 Compression onto a hyperplane

By a hyperplane of Cd we mean a vector subspace of dimension d − 1. The next corollary is a
singular value version of Cauchy’s Interlacing Theorem [13, p. 59].

Corollary 8.3.1. Let A ∈ Md and let S be a hyperplane of Cd orthogonal to the unit vector
h. Set β = min{kAhk2 , kA∗ hk2 } − |hh, Ahi|2 . Then, for all j = 1, . . . , d − 1,

µ2j (A) ≥ µ2j (AS ) ≥ µ2j+1 (A) − β.

√
This double inequality is stronger than µj (A) ≥ µj (AS ) ≥ µj+1 (A) − β. If A is a normal
matrix, then kAhk = kA∗ hk and we have Corollary 8.1.1. If A = V is a unitary matrix,
µj (V ) = 1 for all j and kV hk = kV ∗ hk = 1 for all unit vectors, so we deduce from Corollary
8.3.1 that µj (VS ) ≥ |hh, V hi|. In fact one can easily check that µj (VS ) = 1 for j ≤ d − 2 and
µd−1 (VS ) = |hh, V hi|. Hence Corollary 8.3.1 is sharp.

Proof. (Corollary 8.3.1) The inequality µj (A) ≥ µj (AS ) is trivial. To deal with the other
inequality we may assume that h is the last vector of the canonical basis and that AS is the
submatrix of A obtained by deleting the last column and the last line. We partition A as

A = AS ∪ B ∪ C

where B contains the d − 1 entries below AS and C is the last column of A. We then apply to
this partitioning Corollary 8.2.2 with j1 = j − 1, j2 = 0, and j3 = 1 to get

µ2j+1 (A) ≤ µ2j (AS ) + µ21 (B) + µ22 (C).

Since µ2 (C) = 0, we have

µ2j+1 (A) − µ21 (B) ≤ µ2j (AS )
Observe that µ21 (B) = kA∗ hk2 − |hh, Ahi|2 , hence

µ2j (AS ) ≥ µ2j+1 (A) − kA∗ hk2 + |hh, Ahi|2 . (8.3.1)

We may also partition A as

A = AS ∪ R ∪ L
where R contains the d − 1 entries on the right of AS and L stands for the last line of A.
Arguing as above with R and L in place of B and C yields

µ2j+1 (A) − µ2j (AS ) ≤ µ21 (R) = kAhk2 − |hh, Ahi|2 (8.3.2)

Combining (8.3.1) and (8.3.2) completes the proof.

Corollary 8.3.2. Let A ∈ Md,d′ be partitioned into r row or column compatible blocks Ak ∈
Mnk ,mk . Then, for each block Ak and all j ≥ 1,
X
µ2j (A) − µ2j (Ak ) ≤ µ21 (Al ).
l6=k

Proof. Apply Corollary 8.2.2 with jk = j − 1 and jl = 0 for all l 6= k.

126 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

Corollary 8.3.3. Let A ∈ Md and let S be a hyperplane of Cd orthogonal to the unit vector
h. Then for all j = 1, . . . , d − 1,

µ2j (A) − µ2j (AS ) ≤ kAhk2 + kA∗ hk2 − |hh, Ahi|2 .

Proof. We may suppose that h is the last vector of the canonical basis and we partition A into
three blocks : AS , the last column of A, and the d − 1 entries below AS . We then apply the
previous corollary.

8.4 Four and five blocks

Partitionings into four blocks are not necessarily row or column compatible. However, for such
partitionings, Theorem 8.2.1 still holds.

Corollary 8.4.1. Let A ∈ Md,d′ be partitioned into four blocks A, B, C, D. Then, there exist
some isometries U, V, W, X of suitable sizes such that

|A|2 = U |A|2 U ∗ + V |B|2 V ∗ + W |C|2 W ∗ + X|D|2 X ∗ .

Proof. We assume that A is the block in the upper left corner and we distinguish three cases.
(1) A has the same number d of lines as A. In such a case, letting A′ = B ∪ C ∪ D, the
partitioning A = A ∪ A′ is column compatible, and we have two isometry matrices U, U ′ such
that
|A|2 = U |A|2 U ∗ + U ′ |A′ |2 U ′∗ . (8.4.1)
Since A′ is partitioned into three blocks, necessarily a row or column partitioning, we can
apply the theorem to obtain the decomposition

|A′ |2 = V ′ |B|2 V ′∗ + W ′ |B|2 W ′∗ + X ′ |B|2 X ′∗ (8.4.2)

for some isometry matrices V ′ , W ′ , X ′ of suitable sizes. Combining (8.4.1) and (8.4.2) we
get the conclusion of the corollary with the isometry matrices V = U ′ V ′ , W = U ′ W ′ , and
X = U ′X ′.
(2) A has the same number d′ of columns as A. Letting again A′ = B ∪ C ∪ D, the
partitioning A = A ∪ A′ is row compatible, and we may argue as in case (1).
(3) A has l < d lines and c < d′ columns. There exist then a block, say B, on the top
position, and just on the right of A, and another block, say C just below A and on the left
side. We consider three subcases (a), (b), (c).
(a) B has fewer than l lines. Then, the last block D is necessarily below B with the same
number of columns as B, and so C has either the same number of columns as A or C has d′
columns as A. In the first case, A = A ∪ C ∪ B ∪ D is a column compatible partitioning and
we can apply the theorem. In the second case, the situation is the same as in (2).
(b) B has exactly l lines, like A. We denote by γ the number of columns of B and we
consider three situations.
(i) C has more than c + γ columns. Then necessarily C has d′ columns and D is the upper
right block with l lines, hence A = A ∪ B ∪ D ∪ C is a line compatible partitioning and we
may apply the theorem.
8.4. FOUR AND FIVE BLOCKS 127

(ii) C has exactly c + γ columns. Then, letting A′′ = A ∪ B ∪ C with have a partitioning
into three blocks, and A = A′′ ∪ D. Thus applying the theorem twice as in case (1) yields the
conclusion.
(iii) C has fewer than c + γ columns. Then D is the lower right block, with the same
number of lines as B, and A is partitioned into line compatible blocks. Thus the theorem can
be applied.
(c) B has more lines than A. Let λ be the number of line of B. Hence λ > l. There exist
two situations
(iv) λ < d. Then D is the lower right block, with the same number of columns as B, and
A is partitioned into line compatible blocks. Thus we may apply the theorem.
(v) λ = d. Then A′′′ = A ∪ C ∪ D is a partitioning into three blocks and A = A′′′ ∪ B, thus
applying twice the theorem completes the proof.
We do not know whether Corollary 8.4.1 can be extended or not to any partitioning in five
blocks. For instance we are not able to prove or disprove a version of Corollary 8.4.1 for the
matrices  
  a1 a2 a3 b1 b2
a 1 a 2 b1 a4 a5 a6 b3 b4 
 
A = d1 x b2
  or A =   d1 d 2 x b5 b 6


d2 c1 c2  d3 d4 c1 c2 c3 
d5 d6 c4 c5 c6
partitioned into five obvious blocks A, B, C, D, X. Hence, that Theorem 8.2.1 holds or not
for any partitioning into five blocks is an open problem. More generally, we may consider the
following two questions.
Question 8.4.2. For which partitionings does Theorem 8.2.1 hold ? For which partitionings
does Corollary 8.2.2 hold ?

Matrices partitioned into four blocks (usually of same size) are comon examples of parti-
tionings. A nontrivial inequality follows from the previous corollary.
Corollary 8.4.3. Let A ∈ Md,d′ be partitioned into four blocks A, B, C, D, and let p > 2.
Then, there exist some isometries U, V, W, X of suitable sizes such that
22−p |A|p ≤ U |A|p U ∗ + V |B|p V ∗ + W |C|p W ∗ + X|D|p X ∗ .
The inequality reverses for 2 > p > 0.

Letting p = 1 we have Corollary 8.1.2 with the constant 2 which is sharp, even for a positive
block matrix, as shown by the simple example
 
A A
A= .
A A

Proof. For any monotone convex function f (t) on the nonnegative axis, we have thanks to [31,
Corollary 2.4] and Corollary 8.4.1,
 2
U |A|2 U ∗ + V |B|2 V ∗ + W |C|2 W ∗ + X|D|2 X ∗
 
|A|
f =f
4 4
f (U |A| U ) + V (f |B| V ∗ ) + f (W |C|2 W ∗ ) + f (X|D|2 X ∗ ) ∗
2 ∗ 2
≤Λ Λ
4
128 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

for some unitary matrix Λ ∈ M′d . Picking f (t) = tp/2 with p > 2 yields the result. The reverse
inequalities hold for monotone concave functions f (t) and 0 < p < 2.

Remark 8.4.4. The version of Corollary 8.4.3 for three blocks A, B, C, and p = 1 reads as the
inequality of the abstract,
√
3|A| ≥ U |A|U ∗ + V |B|V ∗ + W |C|W ∗ .
√
The constant 3 is the best one: we cannot take a smaller constant for
 
x y z
A= x y
 z
x y z

partitioned
√ into its three lines. For two blocks, a similar sharp inequality holds with the
constant 2.

8.5 Concave or convex functions

For sake of simplicity we state our results for a square matrix A partitioned into blocks. By
adding some zero rows or zero columns to a rectangular matrix, we could obtain statements
for rectangular matrices (Remark 8.5.9).
Suppose that A ∈ Md is partitioned into blocks Ak ∈ Mnk ,mk , k = 1, . . . , r. From Thomp-
son’s triangle inequality ([92] or [13, p. 74] we have
r
X
|A| ≤ Uk |Ak |Uk∗ (8.5.1)
k=1

for some isometry matrices Uk ∈ Md,mk . The equality of Theorem 8.2.1 and (8.5.1) suggest
several other inequalities, in particular, if A is partioned in row or column compatible blocks,
r
X
3
|A| ≥ Vk |Ak |3 Vk∗ (8.5.2)
k=1

for some isometries Vk ∈ Md,mk . This is indeed true as shown in the following theorem. We
do not know if (8.5.2) can be extended to any partitioning. Corollary 8.4.1 and the proof of
Theorem 8.5.1 show that (8.5.2) holds for four blocks. The case of five blocks is open.

Theorem 8.5.1. Let A ∈ Md be partitioned into r row or column √ compatible blocks Ak ∈

Mnk ,mk , and let ψ(t) be a monotone function on [0, ∞) such that ψ( t) is convex and ψ(0) = 0.
Then there exist some isometries Vk ∈ Md,mk such that
r
X
ψ(|A|) ≥ Vk ψ(|Ak |)Vk∗ .
k=1

Theorem 8.5.1 considerably improves (8.2.9). A special case with ψ(t) = t3 is given in the
abstract.
8.5. CONCAVE OR CONVEX FUNCTIONS 129

Proof. Let g(t) be a monotone convex function on [0, ∞) such that g(0) ≤ 0, and let A, B ∈ Mn
be positive (semidefinite). By [12] or [31, Corollary 3.2] we have

g(A + B) ≥ U g(A)U ∗ + V g(B)V ∗

for some unitary matrices U, V ∈ Mn . Using this inequality and Theorem 8.2.1 we infer
r
X
g(|A|2 ) ≥ Wk g(Uk |Ak |2 Uk∗ )Wk∗
k=1

for some unitary matrices Wk and some isometry matrices Uk ∈ Md,mk . If g(0) = 0, we have
g(Uk |Ak |2 Uk∗ ) = Uk g(|Ak |2 )Uk∗ . Hence
r
X
2
g(|A| ) ≥ Vk g(|Ak |2 )Vk∗
k=1
√
with the isometry matrices Vk = Wk Uk . Applying this to g(t) = ψ( t) completes the proof.

Corollary 8.5.2. Let A ∈ Md be partitioned into r row or column compatible

√ blocks Ak ∈
Mnk ,mk , and let ϕ(t) be a nonnegative function on [0, ∞) such that ϕ( t) is concave. Then
there exist some isometries Uk ∈ Md,nk such that
r
X
ϕ(|A|) ≤ Uk ϕ(|Ak |)Uk∗ .
k=1

√
Proof. Since ϕ( t) is nonnegative and concave, it is necessarily a monotone function (non-
decreasing), hence continuous on (0, ∞). Since we are dealing with matrices we may further
suppose that ϕ(t) is also continuous at t = 0.
(1) Assume that ϕ(0) = 0. Theorem 8.5.1 applied to ψ(t) = −ϕ(t) proves the corollary.
(2) Assume that ϕ(0) > 0. Since the continuous functional calculus is continuous on the
positive semidefinite cone of any Mm , by a limit argument, we may assume that |A| is invertible.
So, suppose√that the √ spectrum√of |A|2 lies in an interval [r 2 , s2 ] with r > 0.√Define a convex
function φ( t) by φ( t) = ϕ( t) for t ≥ r 2 , φ(0) = 0, and the graph of φ( t) on [0, r 2 ] is a
line segment. Hence φ(t) ≤ ϕ(t) and φ(|A|) = ϕ(|A|). Applying case (1) to φ yields
r
X r
X
ϕ(|A|) = φ(|A|) ≤ Uk φ(|Ak |)Uk∗ ≤ Uk ϕ(|Ak |)Uk∗
k=1 k=1

for some isometry matrices Uk .

The next three corollaries follow from Corollary 8.5.2.

Corollary 8.5.3. Let A ∈ Mmn be partitioned into an m × m family of blocks Ai,j ∈ Mn , and
let 0 < q ≤ 2. Then there exist some isometries Ui,j ∈ Mmn,n such that
m
X
q
|A| ≤ Ui,j |Ai,j |q Ui,j
∗
.
i,j=1
130 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS

Corollary 8.5.4. Let A ∈ Mmn be partitioned into an m × m family of blocks Ai,j ∈ Mn , let
s ≥ 1 and 0 < q ≤ 2. Then,
m
{Tr |A|qs }1/s ≤ {Tr |Ai,j |qs }1/s .
X

i,j=1

Corollary 8.5.5. Let A ∈ Mn , let ck be the norm of the k-th column of A and let 0 < q ≤ 2.
Then there exist some rank one projections Ek ∈ Mn such that
n
cqk Ek
X
|A|q ≤
k=1

The last corollaries follow from Theorem 8.5.1.

Corollary 8.5.6. Let A ∈ Mmn be partitioned into an m × m family of blocks Ai,j ∈ Mn , and
let p ≥ 2. Then there exist some isometries Ui,j ∈ Mmn,n such that
m
X
p
|A| ≥ Ui,j |Ai,j |p Ui,j
∗
.
i,j=1

Corollary 8.5.7. Let A ∈ Mmn be partitioned into an m × m family of blocks Ai,j ∈ Mn , let
0 ≤ s ≤ 1 and p ≥ 2. Then,
m
ps 1/s
{Tr |Ai,j |ps }1/s .
X
{Tr |A| } ≥
i,j=1

Corollary 8.5.8. Let A ∈ Mn , let rk be the norm of the k-th row of A and let p ≥ 2. Then
there exist some rank one projections Ek ∈ Mn such that
n
cpk Ek .
X
p
|A| ≥
k=1

Remark 8.5.9. The proof of Theorem 8.5.1 is the same for a d × d′ matrix A. So Corollary
8.5.2 also holds for A ∈ Md,d′ if ϕ(0) = 0. In case of d ≥ d′ , we may again use a limit argument
and assume that |A| is invertible. In case of d′ > d we may argue as follows. Add some zero
lines to A in order to obtain a square matrix A0 ∈ M′d . Let B1 . . . , Bp be the blocks at the
bottom of A, and R1 , . . . , Rq be the remaining blocks of A. Add some zeros to the blocks Bi
in order to obtain blocks Bi0 of A0 in such a way that
!  
[ [
A0 = Bi0 ∪  Rj 
i j

is a row or column compatible partitioning of A0 . Since it is a square matrix, we may apply

Corollary 8.5.2 and since |A0 | = |A| and |Bi0 | = |B i |, we see that Corollary 8.5.2 holds for
d × d′ matrices.
8.6. AROUND THIS ARTICLE 131

8.6 Around this article

For a positive block-diagonal matrix, Theorem 8.2.1 is trivial. The following statement [35] is
more interesting.

Theorem 8.6.1. Let Ai ∈ M+

n , i ∈ Im . Then, for some isometries Vk ∈ Mmn,n , k ∈ Im ,

m m
(m )
M 1 X X
Ai = Vk Ai Vk∗ .
m
i=1 k=1 i=1

This says that direct sums are averages of usual sums, up to isometric congruences. It is a
genuine non-commutative fact with no analogous statement for positive vectors and permuta-
tions of their components.

8.7 References of Chapter 8

[12] J.S. Aujla and J.-C. Bourin, Eigenvalue inequalities for convex and log-convex func-
tions, Linear Algebra Appl. 424 (2007), 25–35.

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York,

1996.

[14] R. Bhatia, Pinching, trimming, truncating, and averaging of matrices. Amer. Math.
Monthly 107 (2000),

[17] R. Bhatia and F. Kittaneh, Norm inequalities for partitioned operators and an appli-
cation. Math. Ann. 287 (1990), no. 4, 719–726. no. 7, 602–608.

[31] J.-C. Bourin and E.-Y. Lee, Unitary orbits of Hermitian operators with convex or
concave functions, Bull. Lond. Math. Soc. 44 (2012), no. 6, 1085–1102.

[35] J.-C. Bourin and E.-Y. Lee, Direct sums of positive semi-definite matrices. Linear
Algebra Appl. 463 (2014), 273–281.

[42] J.-C. Bourin and E.-Y. Lee, A Pythagorean Theorem for partitioned matrices, Proc.
Amer. Math. Soc., in press.

[92] R.-C. Thompson, Convex and concave functions of singular values of matrix sums,
Pacific J. Math. 66 (1976), 285–290.

[94] X. Zhan, The sharp Rado theorem for majorizations, Amer. Math. Monthly 110 (2003)
152–153.
132 CHAPTER 8. BLOCK MATRICES AND PYTHAGORAS
Chapter 9

The spectral Theorem

9.1 Introduction
Hermitian (or symmetric) matrices acting on Hn = Cn (or Rn ) form a nice finite dimensional
real vector space, let us denote it by S(Hn ), with a order structure, A ≤ B, whenever B − A
is positive semidefinte.
The fundamental property of S(Hn ) is the matrix spectral theorem asserting that a
Hermitian matrix can be diagonalized in a suitable orthonormal basis. It is a very important
theorem and, important too, it is an easy result well understood by students. Combined
with the order structure, this theorem shows that Hermitian matrices can be regarded as a
generalization of Rn and opens the way of a full theory of Matrix Analysis, where facts for
real finite sequences of numbers search for counterparts in the Hermitian matrix world.
But, the extension of this diagonalization theorem for operators on a infinite dimensional
separable Hilbert space H - a key stone result of Functional Analysis - is more delicate: the
standard literature invoke abstract constructions based on the Gelfand isomorphism between
some abstract algebras. It might be a too abstract and unnatural approach; the link with
matrices unexpectedly disappears though operators obviously have matrix representations.
However, and fortunately as it is conceptually desirable, it is possible, to give a simple proof
deriving from the matrix case. Hence operators are not more complicated than matrices any
longer ! We do this rather easy and definitely pleasant job in this note.

9.2 The matrix case, and two natural definitions for

operators
Let us first recall a sketch of the proof of the spectral theorem for Hermitian matrices.

• if a subspace S ⊂ Hn is invariant by A, then so is its orthocomplement S ⊥ .

• λ↑1 (A) := minhh, Ahi must be an eigenvalue,

so that repeating the process with the restriction of A to the orthocomplement .... This is the
variational method. and we arrive to the following spectral decomposition of A:
k
λ↑j (A)Pj
X
A= (9.2.1)
j=1

133
134 CHAPTER 9. THE SPECTRAL THEOREM

A very important feature of this decompostion is to allow a matrix functional calculus, given
f (t) ∈ C 0 , the matrix f (A) makes sense. This is the starting point of a branch of Matrix
analysis. Note that the matrix functional calculus behave well with the order structure of C 0 ,:
f (t) ≥ g(t)) implies f (A) ≥ g(A) for all Hermitian matrices A. this will be crucial in our
approach.
Now, let us consider the typical situation in the infinite dimensional case, the multiplication
operator on L2 ([0, 1],
Z : f (t) 7→ tf (t), f (t) ∈ L2 ([0, 1], dt)
This basic example shows that it is problematic to use eigenvalues and projection Pj as in
(9.2.1). However it seems possible to have a corresponding notion for the sum

Ej = P1 + P2 + · · · Pj , j = 1, 2 · · · (9.2.2)

of the first j spectral projections of A by using the operators

Ew : f (t) 7→ I[0,w] (t)f (t), w ∈ [0, 1].

This is a very good point, because we may write the matrix spectral theorem (9.2.1) as:

A = λ↑1 (A)E1 + λ↑2 (A)(E2 − E1 ) + λ↑3 (A)(E3 − E2 ) + .... + λ↑n (A)(En − En−1 ) (9.2.3)

We will obtain a similar statement for Hilbert space operators. To adapt the matrix proof
to the infinite dimensional setting, we need two simple definitions. Let S(H) be the set of
Hermitian operators on a infinite dimensional, separable Hilbert space H.

Definition 9.2.1. (sot-convergence, or strong convergence) We say that the sequence {Xn } in
S(H) is sot convergent, or strongly converges, if it is norm bounded and, for all vectors h ∈ H
(equivalently for all elements of an orthonormal basis), {Xn h} converges in H). If {Xn } is
sot convergent, then there exists (a unic) X ∈ S(H) such that Xh = limn Xn h for all vectors
h ∈ H and we say that X is the strong limit of Xn , written as X = sot limn→∞ Xn .

Hence the strong convergence is merely the pointwise convergence in S(H). Of course, it is
a classical notion and ”sot” refers to strong operator topology.
Given A ∈ S(H), and fixing a (orthonormal) basis, we may write A as an infinite matrix
A = (ai,j ), and so extracting the n-by-n left upper corners, n = 1, 2, 3, .... we obtain a basic
sequence of Hermitian matrices for which the matrix spectral theorem and the functional
calculus are available. Since we wish to extend it to A, it seems natural to use this sequence
of matrices, and the following terminology is convenient.

Definition 9.2.2. The sequence {An } in S(H) is a basic sequence for A if there exists an
increasing sequence of finite rank projections {En }, sot-convergent to the identity I, such that
An = En AEn .

9.3 Continuous functional calculus

In the introduction we have noted that the matrix spectral theorem allows to define the
functional calculus of Hermitian matrices. Our strategy in the Hilbert space operator case
is opposite: we first built up a continuous functional calculus (thanks to the the two above
definition, and -of course- to the matrix spectral theorem)
9.3. CONTINUOUS FUNCTIONAL CALCULUS 135

Lemma 9.3.1. If {An }n∈N is a basic sequence for A ∈ S(H) and p(t) is a polynomial, then

sot lim p(An ) = p(A).

n→∞

Proof.

Lemma 9.3.2. Let f (t) ∈ C 0 and let A ∈ S(H). Then, there exists an operator in S(H), that
we denote f (A), such that
f (A) = sot lim f (An ) (9.3.1)
n→∞

for any basic sequence {An }n∈N for A. Moreover f (A)A = Af (A).

Thus, we have natural way to define operators f (A) from continuous function f (t) and
Hermitian arguments A. This is called the (operator) continous functional calculus. Lemma
2.1 shows that for a polynomial, the continous functional calculus yields, of course, the same
operator as its algebraic definition.

Proof. Fix a unit vector h ∈ H and a scalar ε, and pick a polynomial p(t) such that |f (t) −
p(t)| < ε whenever |t| ≤ kAk. Since kAn k ≤ kAk, for any n, we infer

kf (An ) − p(An )k ≤ ε. (9.3.2)

Let n0 be an integer such that

kp(An )h − p(A)hk ≤ ε (9.3.3)
for all n ≥ n0 . Combining (9.3.2) and (9.3.3) we get kf (An )h − p(A)hk ≤ 2ε for all n ≥ n0 .
As ε can be arbitrarily small, we obtain that {f (An )h} is a Cauchy sequence in H and thus
has a limit that we call f (A)h. It obviously defines a linear operator as f (An ) are linear
operators. And since kf (An )k ≤ max{|f (t)| : −kAk ≤ t ≤ kAk}, f (A) is a bounded linear
operator. Thus we have the strong limit in S, f (A) = sot limn f (An ). To check that f (A)
does not depend on the basic sequence of A, suppose that {A′n } is another basic sequence of
A and note that we also have kf (A′n )h − p(A)hk ≤ 2ε for n large enough and thus {f ′ (An )}
also converges to f (A)h. Finally, as f (An )An = An f (An ), taking the strong limits, we infer
f (A)A = Af (A).

For a finite rank operator Z ∈ S(H), rank Z = m, with spectral decomposition

m
zj↑ Pj ,
X
Z=
j=1

where zj are the nonzero eigenvalues of Z counted with their multiplicities and Pj are corre-
sponding rank one spectral projections, the functional calculus with f ∈ C 0 is obtained by the
simple formulae
m
f (zj↑ )Pj ,
X
f (Z) = f (0)P0 + (9.3.4)
j=1

P0 standing for the projection onto the nullspace of Z. Therefore, given two continuous
functions such that f (t) ≥ g(t), we have f (Z) ≥ g(Z). Applying this to a basic sequence we
obtain the following remark.
136 CHAPTER 9. THE SPECTRAL THEOREM

Remark 9.3.3. If f (t), g(t) ∈ C 0 satysfy f (t) ≥ g(t), then f (A) ≥ g(A) for all A ∈ S(H).

The functional calculus for a finite rank operator (9.3.4) also shows that

hh, f (Z)f i ≤ sup f (t)

t∈R

for all unit vectors h. Applying this to a basic sequence provides the next simple observation.
Remark 9.3.4. If f (t) ∈ C 0 and A ∈ S(H), then, for all unit vectors h,

hh, f (A)hi ≤ sup f (t).

t∈R

9.4 Spectral projections

The aim of this section is to define the operator version of the matrix spectral projections
(9.2.2). We first give a useful fact, known as Vigier’s theorem.

Lemma 9.4.1. Let {Zn }n∈N be a decreasing sequence of positive operators on H, that is
Zn ≥ Zn+1 for all n. Then there exists a positive operator Z such that

Z = sot lim Zn
n→∞

Proof. The wot convergence follows from polarization. Then, we may infer the sot convergence.
Indeed, given h ∈ H, one has

kZn+p h − Zn hk2 = hh, (Zn − Zn+p )2 hi

= hh, (Zn − Zn+p )1/2 (Zn − Zn+p )(Zn − Zn+p )1/2 hi
≤ kZn − Zn+p khh, (Zn − Zn+p )hi.

Lemma 9.4.2. Let A ∈ S(H). For λ ∈ (−∞, ∞) and n ∈ N∗ , define the three piece affine
continuous function fλ,n (t) such that

1
 if t ≤ λ,
fλ,n (t) := 1 + n(λ − t) if λ ≤ t ≤ λ + n−1 ,

0 if λ + n−1 ≤ t.


Then, there exists a projection E(λ) ∈ S(H) commuting with A such that

E(λ) = sot lim fλ,n (A)

n→∞

and hh, AE(λ)hi ≤ λ for all unit vectors h.

Proof. Note that fλ,n (t) ≥ fλ,n+1 (t) ≥ 0, n ∈ N. Remark 9.3.3 shows that {fλ,n (A)}n∈N is
a decreasing sequence of positive operators. Thanks to Lemma 9.3.2 we thus have a positive
operator E(λ) as the strong limit

E(λ) = sot lim fλ,n (A).

n→∞
9.5. THE SPECTRAL THEOREM 137

Since fλ,n (A)A = Afλ,n (A), we also have the commutativity of A with this strong limit,
E(λ)A = AE(λ). Further, by Remark 9.3.4, given a unit vectors h,

hh, Afλ,n (A)hi ≤ sup tfλ,n (t)

t∈R

As the supremum tends to λ as n → ∞, we get hh, AE(λ)hi ≤ λ for all unit vectors h.
Since 1 ≥ fλ,n (t), Remark 9.3.3 shows that I ≥ E(λ) and hence E(λ) ≥ E 2 (λ). We also
have E 2 (λ) ≥ E(λ), indeed,
 2
E 2 (λ) = sot lim fλ,n (A)
n→∞
2
= sot lim fλ,n (A)
n→∞
≥ sot lim fλ,m (A) = E(λ)
m→∞

2 (t) ≥ f
as for each integer n there exists an integer m such that fλ,n 2
λ,m (t). Thus E (λ) = E(λ),
i.e., E(λ) is a projection.

Remark 9.4.3. The projection valued map λ → E(λ) is increasing and sot right continuous,
we call this map (as well as the family {E(λ)} with λ ∈ R) the spectral measure of A.
The previous lemma shows that given two self-adjoint operators A and B with respectives
spectral measures {E(λ)} and {F (ν)}, then the commutativity assumption AB = BA ensures
the commutativity of the spectral measures: E(λ)F (ν) = F (ν)E(λ).

9.5 The spectral theorem

In case of A ∈ S(H), the matrix decomposition (9.2.3) (involving the matrix spectral mea-
sure....) remains valid in a continuous form:

Theorem 9.5.1. Let A ∈ S(H) with its spectral measure E(λ). Then
Z
A = λ dE(λ)

Now we explain the Stieltjes-integral notation employed, and next give the proof which is a
straightforward consequence of the existence of the spectral measure built up in the previous
Lemma.
We may detail a little bit more the theorem, The spectral measure vanishes on an open set
ω(A) of R:
x ∈ ω(A) ⇐⇒ ∃ε > 0 s.t. E(x + ε) − E(x − ε) = 0.
The complementary set of ω is the compact set σ(A), the spectrum of A. Then we may write
the spectral theorem as

Theorem 9.5.2. Let A ∈ S(H) with its spectral measure E(λ). Then
Z
A= λ dE(λ).
σ(A)

Denote by C ∗ (A) the unital C ∗ -algebra spanned by A. We have the following concrete
version of the Gelfand isomorphism mentionned in the introduction.
138 CHAPTER 9. THE SPECTRAL THEOREM

Corollary 9.5.3. Let A ∈ S(H). Then the functional calculus from C 0 (σ(A)) to C ∗ (A),
Z
f (t) 7→ f (A) := f (λ) dE(λ).
σ(A)

is an isometric ∗-isomorphism.

From Theorem 9.5.1 and Remark 9.4.3 we get:

Corollary 9.5.4. Let A, B ∈ S(H) be commuting. Then AB ∈ S(H).

Taking advantage that a normal operator N has a decomposition N = A + iB in which A

and B are two commuting self-adjoint operators, the above corollary and Remark 9.4.3 show
that A and B have two commuting respective spectral measures, say E(x) and F (y), x, y ∈ R.
Thus we obtain a spectral measure G(z), z ∈ C, for N and get the following spectral theorem
for normal operators:

Theorem 9.5.5. Let N be a normal operator with its spectral measure G(λ). Then
Z Z
N= λ dG(λ) = λ dG(λ)
C σ(N )
Bibliography

[1] C.A. Akemann and J. Anderson. Lyapunov theorems for operator algebras. Mem. Amer.
Math. Soc., 94 no 458, 1991.

[2] E. Alfsen and F. Shultz, Unique decompositions, faces, and automorphisms of separable
states, J. Math. Phys. 51, 052201 (2010).

[3] J.H. Anderson and J.G. Stampfli, Commutators and compressions, Israel J. Math. 10
(1971), 433–441.

[4] T. Ando, Comparison of norms |||f (A) − f (B)||| and kf (|A − B|)|||. Math. Z. 197 (1988),
no. 3, 403-409.

[5] T. Ando, Majorization, doubly stochastic matrices, and comparison of eigenvalues, Linear
Algebra Appl. 118 (1989), 163-248.

[6] T. Ando and X. Zhan, Norm inequalities related to operator monotone functions, Math.
Ann. 315 (1999) 771-780.

[7] H. Araki, On an inequality of Lieb and Thirring, Let. Math. Phys. 19 (1990 )167-170.

[8] K. Audenaert, On the Araki-Lieb-Thirring Inequality, Int. J. Inf. Syst. Sci. 4 (2008),
78-83.

[9] K. Audenaert and F. Hiai, Reciprocal Lie-Trotter formula, Linear Mult. Algebra, in press

[10] K. Audenaert and J. S. Aujla, On Ando’s inequalities for convex and concave functions,
arXiv:0704.0099v1.

[11] J. S. Aujla and F. C. Silva, Weak majorization inequalities and convex functions, Linear
Algebra Appl. 369 (2003), 217-233.

[12] J.S. Aujla and J.-C. Bourin, Eigenvalue inequalities for convex and log-convex functions,
Linear Algebra Appl. 424 (2007), 25–35.

[13] R. Bhatia, Matrix Analysis, Gradutate Texts in Mathematics, Springer, New-York, 1996.

[14] R. Bhatia, Pinching, trimming, truncating, and averaging of matrices. Amer. Math.
Monthly 107 (2000), no. 7, 602–608.

[15] R. Bhatia, Positive Definite Matrices, Princeton University press, Princeton 2007.

[16] R. Bhatia and J. Holbrook, On the Clarkson-McCarthy inequalities, Math. Ann. 281
(1988), no. 1, 7–12.

139
140 BIBLIOGRAPHY

[17] R. Bhatia and F. Kittaneh, Norm inequalities for partitioned operators and an application.
Math. Ann. 287 (1990), no. 4, 719–726.

[18] J.-C. Bourin, Some inequalities for norms on matrices and operators, Linear Algebra Appl.
292 (1999), no. 1–3, 139–154.

[19] J.-C. Bourin, Singular values of compressions, restrictions and dilations, Linear Algebra
Appl. 360 (2003), 259–272.

[20] J.-C. Bourin, Total dilations, Linear Algebra Appl. 368 (2003), 159–169.

[21] J.-C. Bourin, Compressions and pinchings, J. Operator Theory 50 (2003), no. 2, 211-220.

[22] J.-C. Bourin, Convexity or concavity inequalities for Hermitian operators. Math. Inequal.
Appl. 7 (2004), no. 4, 607–620.

[23] J.-C. Bourin, Hermitian operators and convex functions, J. Inequal. Pure Appl. Math. 6
(2005), Article 139, 6 pp.

[24] J.-C. Bourin, A concavity inequality for symmetric norms, Linear Algebra Appl. 413
(2006), 212-217.

[25] J.-C. Bourin, Matrix versions of some classical inequalities. Linear Algebra Appl. 416
(2006), no. 2–3, 890–907.

[26] J.-C. Bourin, Matrix subadditivity inequalities and block-matrices, Internat. J. Math. 20
(2009), no. 6, 679–691.

[27] J.-C. Bourin, A matrix subadditivity inequality for symmetric norms, Proc. Amer. Math.
Soc. 138 (2010), no. 2, 495–504.

[28] J.-C. Bourin and F. Hiai, Norm and anti-norm inequalities for positive semi-definite ma-
trices, Internat. J. Math. 63 (2011), 1121-1138.

[29] J.-C. Bourin and F. Hiai, Jensen and Minkowski inequalities for operator means and
anti-norms. Linear Algebra Appl. 456 (2014), 22–53.

[30] J.-C. Bourin and E.-Y. Lee, Concave functions of positive operators, sums, and congru-
ences, J. Operator Theory 63 (2010), 151–157.

[31] J.-C. Bourin and E.-Y. Lee, Unitary orbits of Hermitian operators with convex or concave
functions, Bull. Lond. Math. Soc. 44 (2012), no. 6, 1085–1102.

[32] J.-C. Bourin and E.-Y. Lee, Decomposition and partial trace of positive matrices with
Hermitian blocks, Internat. J. Math. 24 (2013), no. 1, 1350010, 13 pp.

[33] J.-C. Bourin and E.-Y. Lee, Sums of Murray-von Neumann equivalent positive operators,
C. R. Math. Acad. Sci. Paris 351 (2013), no. 19-20, 761-764.

[34] J.-C. Bourin and E.-Y. Lee, Sums of unitarily equivalent positive operators, C. R. Math.
Acad. Sci. Paris 352 (2014), no. 5, 435–439.

[35] J.-C. Bourin and E.-Y. Lee, Direct sums of positive semi-definite matrices. Linear Algebra
Appl. 463 (2014), 273–281.
BIBLIOGRAPHY 141

[36] J.-C. Bourin and E.-Y. Lee, Matrix inequalities from a two variables functional, Internat.
J. Math. 27 (2016), no. 9, 1650071, 19 pp.

[37] J.-C. Bourin and E.-Y. Lee, Pinchings and positive linear maps, J. Funct. Anal. 270
(2016), no. 1, 359–374.

[38] J.-C. Bourin and E.-Y. Lee, Positive linear maps on normal matrices, Internat. J. Math.
29 (2018), no. 12, 1850088, 11 pp.

[39] J.-C. Bourin and E.-Y. Lee, On the Russo-Dye theorem for positive linear maps. Linear
Algebra Appl. 571 (2019), 92–102.

[40] J.-C. Bourin and E.-Y. Lee, Clarkson-McCarthy inequalities with unitary and isometry
orbits, Linear Algebra Appl. 601 (2020), 170–179.

[41] J.-C. Bourin and E.-Y. Lee, Numerical range and positive block matrices, Bull. Aust.
Math. Soc. 103 (2021), no. 1, 69–77.

[42] J.-C. Bourin and E.-Y. Lee, A Pythagorean Theorem for partitioned matrices, Proc.
Amer. Math. Soc., in press.

[43] J.-C. Bourin and E.-Y. Lee, Matrix inequalities and majorizations around Hermite-
Hadamard’s inequality, Canad. Bull. Math., in press.

[44] J.-C. Bourin and E.-Y. Lee, Eigenvalue inequalities for positive block matrices with the
inradius of the numerical range, Internat. J. Math. 33 (2022), no. 1, 10 pp.

[45] J.-C. Bourin, E.-Y. Lee and M. Lin, On a decomposition lemma for positive semi-definite
block-matrices, Linear Algebra Appl. 437 (2012), 1906–1912.

[46] J.-C. Bourin, E.-Y. Lee and M. Lin, Positive matrices partitioned into a small number of
Hermitian blocks, Linear Algebra Appl. 438 (2013), no. 5, 2591–2598.

[47] J.-C. Bourin, A. Mhanna, Positive block matrices and numerical ranges, C. R. Math.
Acad. Sci. Paris 355 (2017), no. 10, 1077–1081.

[48] J.-C. Bourin and E. Ricard, An asymmetric Kadison’s inequality, Linear Algebra Appl.
433 (2010) 499–510.

[49] J.-C. Bourin and M. Uchiyama, A matrix subadditivity inequality for f (A + B) and
f (A) + f (B), Linear Algebra Appl. 423 (2007), 512–518.

[50] L. G. Brown and H. Kosaki, Jensen’s inequality in semi-finite von Neuman algebras, J.
Operator theory 23 (1990), 3–19.

[51] K. Chen and L.-A. Wu, A matrix realignment method for recognizing entanglement,
Quantum Inf. Comput. 3 (2003), 193-202.

[52] M.-D. Choi, A Schwarz inequality for positive linear maps on C ∗ -algebras, Illinois J.
Math. 18 (1974), 565–574.

[53] Clifford, Applications of Grassmann’ extensive algebra, Amer. Journ. Math. 1 (1878),
350-358.
142 BIBLIOGRAPHY

[54] J.E. Cohen, Spectral inequalities for matrix exponentials, Linear Algebra Appl. 111 (1988)
25-28.

[55] J.E. Cohen, S. Friedland, T. Kate, and F. Kelly, Eigenvalue inequalities for products of
matrix exponentials, Linear Algebra Appl. 45 (1982) 55-95.

[56] C. Davis, A Schwarz inequality for convex operator functions, Proc. Amer. Math. Soc. 8
(1957), 42-44.

[57] J. Dereziński, Introduction to representations of the canonical commutation and anticom-

mutation relations, Lect. Note Phys. 695, 65-145 (2006), Springer.

[58] C. Dragan and V. Kaftal, Sums of equivalent sequences of positive operators in von
Neumann factors, preprint, arXiv:1504.03193.

[59] D.J.H. Garling, Inequalities - A journey into linear analysis. Cambridge University Press,
Cambridge, 2007.

[60] P.R. Halmos, Numerical ranges and normal dilations, Acta Sci. Math. (Szeged) 25 (1964)
1–5.

[61] F. Hansen, An operator inequality, Math. Ann. 246 (1979/80), no. 3, 249–250.

[62] F. Hansen and G. K. Pedersen, Jensen’s inequality for operators and Löwner’s theorem,
Math. Ann. 258 (1982), 229–241.

[63] F. Hansen and G. K. Pedersen, Jensen’s operator inequality, Bull. London Math. Soc. 35
(2003), no. 4, 553–564.

[64] F. Hiai, Matrix Analysis: Matrix Monotone Functions, Matrix Means, and Majorization
(GSIS selected lectures), Interdisciplinary Information Sciences 16 (2010), 139–248.

[65] F. Hiai, Log-majorizations and norm inequalities for exponential operators, Banach Cen-
ter Pub. 38 (1997) 119–181.

[66] F. Hiai, A generalization of Araki’s log-majorization. Linear Algebra Appl. 501 (2016),
1–16.

[67] F. Hiai, D. Petz, Introduction to Matrix Analysis and applications. Universitext, Springer,
New Delhi, 2014.

[68] T. Hiroshima, Majorization Criterion for Distillability of a Bipartite Quantum State,

(ArXiv: quantum-ph, 2003).

[69] O. Hirzallah and F. Kittaneh, Non-commutative Clarkson inequalities for unitarily in-
variant norms, Pacific J. Math. 202 (2002), no. 2, 363–369.

[70] M. Horodecki, P. Horodecki, R. Horodecki, Separability of mixed states: necessary and

sufficient conditions, Phys. Lett. A 223 (1996) l-8.

[71] R. Kadison, The Pythagorean theorem. II. The infinite discrete case, Proc. Natl. Acad.
Sci. USA 99 (2002), no. 8, 5217-5222.

[72] R. Kadison and I. Singer, Extensions of pure states, Amer. J. Math. (1959), 383-400.
BIBLIOGRAPHY 143

[73] M. Kennedy and P. Skoufranis, The Schur-Horn Problem for Normal Operators, Proc.
London Math. Soc., in press, arXiv:1501.06457.

[74] H. Kosaki, Arithmetic-geometric mean and related inequalities for operators. J. Funct.
Anal. 156 (1998), no. 2, 429-451.

[75] E.-Y. Lee, Extension of Rotfel’d Theorem, Linear Algebra Appl. 435 (2010), 735–741.

[76] E.-Y. Lee, How to compare the absolute values of operator sums and the sums of absolute
values ?, to appear in Operator and Matrices.

[77] E. Lieb and W. Thirring, in E. Lieb, B. Simon and A. S. Wightman (eds), Studies in
Mathematical Physics, Princeton Press, I976, pp. 301-302

[78] J. Loreaux and G. Weiss, Diagonality and idempotents with applications to problems in
operator theory and frame theory, J. Operator Theory, in press, arXiv:1410.7441.

[79] R. Mathias, Concavity of monotone matrix functions of finite order, Linear and Multilin-
ear Algebra 27 (1990), no. 2, 129-138.

[80] A. Mhanna, On symmetric norm inequalities and positive definite block-matrices, Math.
Inequal. Appl. 21 (2018), no. 1, 133–138.

[81] M. S. Moslehian, Matrix Hermite-Hadamard type inequalities, Houston J. Math. 39

(2013), no. 1, 177–189.

[82] V. Müller, Y. Tomilov, In search of convexity: diagonals and numerical ranges, Bull.
London Math soc., 53 (2021), no. 4, 1016–1029.

[83] C. P. Niculescu; L.-E. Persson, Old and new on the Hermite-Hadamard inequality, Real
Anal. Exchange 29 (2003/04), no. 2, 663–685.

[84] M. A. Nielsen and J. Kempe, Separable states are more disordered globally than locally,
Phys. Rev. Lett. 86 (2001) 5184-5187.

[85] C. Pearcy and D. Topping, Sums of small numbers of idempotents, Michigan J. Math.
14 (1967) 453–465.

[86] D. Petz, Matrix Analysis with some applications, <http://www.math.hu/petz>.

[87] M. Ravichandran, The Schur-Horn theorem in von Neumann algebras, preprint,

arXiv:1209.0909.

[88] S. Ju. Rotfel’d, The singular values of a sum of completely continuous operators, Topics
in Mathematical Physics, Consultants Bureau, Vol. 3 (1969) 73-78.

[89] B. Simon, Trace ideal and their applications, Cambridge University Press, Cambridge,
1979.

[90] W. F. Stinespring, Positive functions on C ∗ -algebras, Proc. Amer. Math. Soc. 6, (1955).
211-216.

[91] C.J. Thompson, Inequalities and partial orders in matrix spaces, Indiana Univ. Math. J.
21 (1971) 469-480.
144 BIBLIOGRAPHY

[92] R.-C. Thompson, Convex and concave functions of singular values of matrix sums, Pacific
J. Math. 66 (1976), 285–290.

[93] P.Y. Wu, Additive combination of special operators, Banach Center Publ. 30 (1994),
337-361.

[94] X. Zhan, The sharp Rado theorem for majorizations, Amer. Math. Monthly 110 (2003)
152–153.