Chapter 11

Basics of Hermitian Geometry

11.1 Sesquilinear Forms, Hermitian Forms, Hermitian

Spaces, Pre-Hilbert Spaces

In this chapter, we generalize the basic results of Eu-

clidean geometry presented in Chapter 9 to vector spaces
over the complex numbers.

Some complications arise, due to complex conjugation.

Recall that for any complex number z 2 C, if z = x + iy

where x, y 2 R, we let <z = x, the real part of z, and
=z = y, the imaginary part of z.

We also denote the conjugate of z = x+iy as z = x iy,

and the absolute value (or length, or modulus) of z as |z|.
Recall that |z|2 = zz = x2 + y 2.

579
580 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

There are many natural situations where a map

' : E ⇥ E ! C is linear in its first argument and only
semilinear in its second argument.

For example, the natural inner product to deal with func-

tions f : R ! C, especially Fourier series, is
Z ⇡
hf, gi = f (x)g(x)dx,
⇡

which is semilinear (but not linear) in g.

Definition 11.1. Given two vector spaces E and F over

the complex field C, a function f : E ! F is semilinear
if
f (u + v) = f (u) + f (v),
f ( u) = f (u),
for all u, v 2 E and all 2 C.
11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 581

Remark : Instead of defining semilinear maps, we could

have defined the vector space E as the vector space with
the same carrier set E, whose addition is the same as that
of E, but whose multiplication by a complex number is
given by
( , u) 7! u.

Then, it is easy to check that a function f : E ! C is

semilinear i↵ f : E ! C is linear.

We can now define sesquilinear forms and Hermitian forms.

582 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Definition 11.2. Given a complex vector space E, a

function ' : E ⇥ E ! C is a sesquilinear form i↵ it is
linear in its first argument and semilinear in its second
argument, which means that
'(u1 + u2, v) = '(u1, v) + '(u2, v),
'(u, v1 + v2) = '(u, v1) + '(u, v2),
'( u, v) = '(u, v),
'(u, µv) = µ'(u, v),
for all u, v, u1, u2, v1, v2 2 E, and all , µ 2 C. A
function ' : E ⇥ E ! C is a Hermitian form i↵ it is
sesquilinear and if
'(v, u) = '(u, v)
for all all u, v 2 E.

Obviously, '(0, v) = '(u, 0) = 0.

11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 583

Also note that if ' : E ⇥ E ! C is sesquilinear, we have

'( u + µv, u + µv) = | |2'(u, u) + µ'(u, v)

+ µ'(v, u) + |µ|2'(v, v),

and if ' : E ⇥ E ! C is Hermitian, we have

'( u + µv, u + µv)

= | |2'(u, u) + 2<( µ'(u, v)) + |µ|2'(v, v).

Note that restricted to real coefficients, a sesquilinear

form is bilinear (we sometimes say R-bilinear).

The function : E ! C defined such that (u) = '(u, u)

for all u 2 E is called the quadratic form associated with
'.
584 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

The standard example of a Hermitian form on Cn is the

map ' defined such that
'((x1, . . . , xn), (y1, . . . , yn)) = x1y1 + x2y2 + · · · + xnyn.

This map is also positive definite, but before dealing with

these issues, we show the following useful proposition.

Proposition 11.1. Given a complex vector space E,

the following properties hold:
(1) A sesquilinear form ' : E ⇥ E ! C is a Hermitian
form i↵ '(u, u) 2 R for all u 2 E.
(2) If ' : E ⇥ E ! C is a sesquilinear form, then
4'(u, v) = '(u + v, u + v) '(u v, u v)
+ i'(u + iv, u + iv) i'(u iv, u iv),
and
2'(u, v) = (1 + i)('(u, u) + '(v, v))
'(u v, u v) i'(u iv, u iv).

These are called polarization identities.

11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 585

Proposition 11.1 shows that a sesquilinear form is com-

pletely determined by the quadratic form (u) = '(u, u),
even if ' is not Hermitian.

This is false for a real bilinear form, unless it is symmetric.

For example, the bilinear form ' : R2 ⇥ R2 ! R defined

such that
'((x1, y1), (x2, y2)) = x1y2 x2 y1
is not identically zero, and yet, it is null on the diagonal.

However, a real symmetric bilinear form is indeed deter-

mined by its values on the diagonal, as we saw in Chapter
9.

As in the Euclidean case, Hermitian forms for which

'(u, u) 0 play an important role.
586 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Definition 11.3. Given a complex vector space E, a

Hermitian form ' : E ⇥E ! C is positive i↵ '(u, u) 0
for all u 2 E, and positive definite i↵ '(u, u) > 0 for all
u 6= 0. A pair hE, 'i where E is a complex vector space
and ' is a Hermitian form on E is called a pre-Hilbert
space if ' is positive, and a Hermitian (or unitary)
space if ' is positive definite.

We warn our readers that some authors, such as Lang

[24], define a pre-Hilbert space as what we define to be a
Hermitian space.

We prefer following the terminology used in Schwartz [28]

and Bourbaki [7].

The quantity '(u, v) is usually called the Hermitian prod-

uct of u and v. We will occasionally call it the inner
product of u and v.
11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 587

Given a pre-Hilbert space hE, 'i, as in the case of a Eu-

clidean space, we also denote '(u, v) as
u · v, or hu, vi, or (u|v),
p
and (u) as kuk.

Example 1. The complex vector space Cn under the

Hermitian form
'((x1, . . . , xn), (y1, . . . , yn)) = x1y1 + x2y2 + · · · + xnyn
is a Hermitian space.

Example 2. Let l2 denote the set of all countably in-

sequences x = (xi)i2N of complex numbers
finite P P such
that 1 i=0 |x i | 2
is defined (i.e. the sequence n
i=0 |xi |
2

converges as n ! 1).

It can be shown that the map ' : l2 ⇥ l2 ! C defined

such that
1
X
' ((xi)i2N, (yi)i2N) = xi yi
i=0

is well defined, and l2 is a Hermitian space under '. Ac-

tually, l2 is even a Hilbert space.
588 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Example 3. Consider the set Cpiece[a, b] of piecewise bounded

continuous functions f : [a, b] ! C under the Hermitian
form Z b
hf, gi = f (x)g(x)dx.
a

It is easy to check that this Hermitian form is positive,

but it is not definite. Thus, under this Hermitian form,
Cpiece[a, b] is only a pre-Hilbert space.

Example 4. Let C[a, b] be the set of complex-valued

continuous functions f : [a, b] ! C under the Hermitian
form Z b
hf, gi = f (x)g(x)dx.
a

It is easy to check that this Hermitian form is positive

definite. Thus, C[a, b] is a Hermitian space.
11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 589

Example 5. Let E = Mn(C) be the vector space of

complex n⇥n matrices. We define the Hermitian product
of two matrices A, B 2 Mn(C) as
n
X
hA, Bi = aij bij ,
i,j=1

which can be conveniently written as

hA, Bi = tr(A>B) = tr(B ⇤A).

Since this can be viewed as the standard Hermitian prod-

2
uct on Cn , it is a Hermitian product on Mn(C). The
corresponding norm
p
kAkF = tr(A⇤A)

is the Frobenius norm (see Section 6.2).

590 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

If E is finite-dimensional and if ' : E ⇥ E ! R is a

sequilinear form onPE, given any basisP(e1, . . . , en) of E,
we can write x = ni=1 xiei and y = nj=1 yj ej , and we
have
n
X
'(x, y) = xiy j '(ei, ej ).
i,j=1

If we let G = (gij ) be the matrix given by gij = '(ej , ei),

and if x and y are the column vectors associated with
(x1, . . . , xn) and (y1, . . . , yn), then we can write

'(x, y) = x>G> y = y ⇤Gx,

where y corresponds to (y 1, . . . , y n).

 11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 591

As in Section 9.1, we are committing the slightP abuse of

notation of letting x denote both the vector x = ni=1 xiei
and the column vector associated with (x1, . . . , xn) (and
similarly for y). The “correct” expression for '(x, y) is

'(x, y) = y⇤Gx = x>G>y.

 Observe that in '(x, y) = y ⇤Gx, the matrix involved

is the transpose of the matrix ('(ei, ej )). The reason
for this is that we want G to be positive definite when '
is positive definite, not G>.

Furthermore, observe that ' is Hermitian i↵ G = G⇤, and

' is positive definite i↵ the matrix G is positive definite,
that is,

(Gx)>x = x⇤Gx > 0 for all x 2 Cn, x 6= 0.

The matrix G associated with a Hermitian product is

called the Gram matrix of the Hermitian product with
respect to the basis (e1, . . . , en).
592 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Proposition 11.2. Let E be a finite-dimensional vec-

tor space, and let (e1, . . . , en) be a basis of E.
1. For any Hermitian inner product h , i on E, if
G = (gij ) with gij = hej , eii is the Gram matrix
of the Hermitian product h , i w.r.t. the basis
(e1, . . . , en), then G is Hermitian positive definite.
2. For any change of basis matrix P , the Gram ma-
trix of h , i with respect to the new basis is P ⇤GP .
3. If A is any n ⇥ n Hermitian positive definite ma-
trix, then
hx, yi = y ⇤Ax
is a Hermitian product on E.

The following result reminiscent of the first polarization

identity of Proposition 11.1 can be used to prove that two
linear maps are identical.
11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 593

Proposition 11.3. Given any Hermitian space E with

Hermitian product h , i, for any linear map f : E !
E, if hf (x), xi = 0 for all x 2 E, then f = 0.

One should be careful not to apply Proposition 11.3 to a

linear map on a real Euclidean space, because it is false!
The reader should find a counterexample.

The Cauchy-Schwarz inequality and the Minkowski in-

equalities extend to pre-Hilbert spaces and to Hermitian
spaces.
594 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Proposition 11.4. Let hE, 'i be a pre-Hilbert space

with associated quadratic form . For all u, v 2 E,
we have the Cauchy-Schwarz inequality:
p p
|'(u, v)|  (u) (v).
Furthermore, if hE, 'i is a Hermitian space, the equal-
ity holds i↵ u and v are linearly dependent.

We also have the Minkovski inequality:

p p p
(u + v)  (u) + (v).
Furthermore, if hE, 'i is a Hermitian space, the equal-
ity holds i↵ u and v are linearly dependent, where in
addition, if u 6= 0 and v 6= 0, then u = v for some
real such that > 0.

As in the Euclidean case, if hE, 'i is a Hermitian space,

the Minkovski inequality
p p p
(u + v)  (u) + (v)
p
shows that the map u 7! (u) is a norm on E.
11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 595

The norm induced by ' is called the Hermitian norm

induced by '.
p
We usually denote (u) as kuk, and the Cauchy-Schwarz
inequality is written as
|u · v|  kuk kvk .

Since a Hermitian space is a normed vector space, it is

a topological space under the topology induced by the
norm (a basis for this topology is given by the open balls
B0(u, ⇢) of center u and radius ⇢ > 0, where
B0(u, ⇢) = {v 2 E | kv uk < ⇢}.

If E has finite dimension, every linear map is continuous.

596 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Remark: As in the case of real vector spaces, a norm

on a complex vector space is induced by some positive
definite Hermitian product h , i i↵ it satisfies the par-
allelogram law :

ku + vk2 + ku vk2 = 2(kuk2 + kvk2).

This time, the Hermitian product is recovered using the

polarization identity from Proposition 11.1:

4hu, vi = ku + vk2 ku vk2+i ku + ivk2 i ku ivk2 .

11.1. SESQUILINEAR FORMS, HERMITIAN FORMS 597

The Cauchy-Schwarz inequality

|u · v|  kuk kvk
shows that ' : E ⇥ E ! C is continuous, and thus, that
k k is continuous.

If hE, 'i is only pre-Hilbertian, kuk is called a semi-

norm.

In this case, the condition

kuk = 0 implies u = 0
is not necessarily true.

However, the Cauchy-Schwarz inequality shows that if

kuk = 0, then u · v = 0 for all v 2 E.

We will now basically mirror the presentation of Euclidean

geometry given in Chapter 9 rather quickly, leaving out
most proofs, except when they need to be seriously amended.
598 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

11.2 Orthogonality, Duality, Adjoint of A Linear Map

In this section, we assume that we are dealing with Her-

mitian spaces. We denote the Hermitian inner product
as u · v or hu, vi.

The concepts of orthogonality, orthogonal family of vec-

tors, orthonormal family of vectors, and orthogonal com-
plement of a set of vectors, are unchanged from the Eu-
clidean case (Definition 9.2).

For example, the set C[ ⇡, ⇡] of continuous functions

f : [ ⇡, ⇡] ! C is a Hermitian space under the product
Z ⇡
hf, gi = f (x)g(x)dx,
⇡

and the family (eikx)k2Z is orthogonal.

11.2. ORTHOGONALITY, DUALITY, ADJOINT OF A LINEAR MAP 599

Proposition 9.3 and 9.4 hold without any changes.

It is easy to show that

n 2 n
X X 2
X
ui = kuik + 2<(ui · uj ).
i=1 i=1 1i<jn

Analogously to the case of Euclidean spaces of finite di-

mension, the Hermitian product induces a canonical bi-
jection (i.e., independent of the choice of bases) between
the vector space E and the space E ⇤.

This is one of the places where conjugation shows up, but

in this case, troubles are minor.
600 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Given a Hermitian space E, for any vector u 2 E, let

'lu : E ! C be the map defined such that

'lu(v) = u · v, for all v 2 E.

Similarly, for any vector v 2 E, let 'rv : E ! C be the

map defined such that

'rv (u) = u · v, for all u 2 E.

Since the Hermitian product is linear in its first argument

u, the map 'rv is a linear form in E ⇤, and since it is
semilinear in its second argument v, the map 'lu is also a
linear form in E ⇤.
11.2. ORTHOGONALITY, DUALITY, ADJOINT OF A LINEAR MAP 601

Thus, we have two maps [l : E ! E ⇤ and [r : E ! E ⇤,

defined such that

[l (u) = 'lu, and [r (v) = 'rv .

Actually, it is easy to show that 'lu = 'ru and [l = [r .

Therefore, we use the notation 'u for both 'lu and 'ru,
and [ for both [l and [r .

Theorem 11.5. let E be a Hermitian space E. The

map [ : E ! E ⇤ defined such that
[(u) = 'lu = 'ru for all u 2 E
is semilinear and injective. When E is also of finite
dimension, the map [ : E ! E ⇤ is a canonical iso-
morphism.
602 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

The inverse of the isomorphism [ : E ! E ⇤ is denoted

by ] : E ⇤ ! E.

As a corollary of the isomorphism [ : E ! E ⇤, if E is

a Hermitian space of finite dimension, then every linear
form f 2 E ⇤ corresponds to a unique v 2 E, such that
f (u) = u · v, for every u 2 E.

In particular, if f is not the null form, the kernel of f ,

which is a hyperplane H, is precisely the set of vectors
that are orthogonal to v.

Remark . The “musical map” [ : E ! E ⇤ is not surjec-

tive when E has infinite dimension.

This result will be salvaged by restricting our attention to

continuous linear maps, and by assuming that the vector
space E is a Hilbert space.
11.2. ORTHOGONALITY, DUALITY, ADJOINT OF A LINEAR MAP 603

The existence of the isomorphism [ : E ! E ⇤ is crucial

to the existence of adjoint maps.

Indeed, Theorem 11.5 allows us to define the adjoint of a

linear map on a Hermitian space.

Let E be a Hermitian space of finite dimension n, and let

f : E ! E be a linear map.

For every u 2 E, the map

v 7! u · f (v)
is clearly a linear form in E ⇤, and by Theorem 11.5, there
is a unique vector in E denoted by f ⇤(u), such that
f ⇤(u) · v = u · f (v),
that is,
f ⇤(u) · v = u · f (v), for every v 2 E.
604 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Proposition 11.6. Given a Hermitian space E of fi-

nite dimension, for every linear map f : E ! E, there
is a unique linear map f ⇤ : E ! E, such that
f ⇤(u) · v = u · f (v),
for all u, v 2 E. The map f ⇤ is called the adjoint of
f (w.r.t. to the Hermitian product).

The fact that

v·u=u·v
implies that the adjoint f ⇤ of f is also characterized by
f (u) · v = u · f ⇤(v),
for all u, v 2 E. It is also obvious that f ⇤⇤ = f .
11.2. ORTHOGONALITY, DUALITY, ADJOINT OF A LINEAR MAP 605

Given two Hermitian spaces E and F , where the Hermi-

tian product on E is denoted as h , i1 and the Hermi-
tian product on F is denoted as h , i2, given any linear
map f : E ! F , it is immediately verified that the proof
of Proposition 11.6 can be adapted to show that there is
a unique linear map f ⇤ : F ! E such that
hf (u), vi2 = hu, f ⇤(v)i1
for all u 2 E and all v 2 F . The linear map f ⇤ is also
called the adjoint of f .

As in the Euclidean case, a linear map f : E ! E (where

E is a finite-dimensional Hermitian space) is self-adjoint
if f = f ⇤. The map f is positive semidefinite i↵

hf (x), xi 0 all x 2 E;

positive definite i↵
hf (x), xi > 0 all x 2 E, x 6= 0.
606 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

An interesting corollary of Proposition 11.3 is that a pos-

itive semidefinite linear map must be self-adjoint. In fact,
we can prove a slightly more general result.

Proposition 11.7. Given any finite-dimensional Her-

mitian space E with Hermitian product h , i, for
any linear map f : E ! E, if hf (x), xi 2 R for all
x 2 E, then f is self-adjoint. In particular, any posi-
tive semidefinite linear map f : E ! E is self-adjoint.

As in the Euclidean case, Theorem 11.5 can be used to

show that any Hermitian space of finite dimension has an
orthonormal basis. The proof is unchanged.
11.2. ORTHOGONALITY, DUALITY, ADJOINT OF A LINEAR MAP 607

Proposition 11.8. Given any nontrivial Hermitian

space E of finite dimension n 1, there is an or-
thonormal basis (u1, . . . , un) for E.

The Gram–Schmidt orthonormalization procedure also

applies to Hermitian spaces of finite dimension, without
any changes from the Euclidean case!
Proposition 11.9. Given any nontrivial Hermitian
space E of finite dimension n 1, from any basis
(e1, . . . , en) for E, we can construct an orthonormal
basis
(u1, . . . , un) for E, with the property that for every k,
1  k  n, the families (e1, . . . , ek ) and (u1, . . . , uk )
generate the same subspace.

Remarks: The remarks made after Proposition 9.8 also

apply here, except that in the QR-decomposition, Q is a
unitary matrix.
608 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

As a consequence of Proposition 9.7 (or Proposition 11.9),

given any Hermitian space of finite dimension n, if (e1, . . . ,
en) is an orthonormal basis for E, then for any two vec-
tors u = u1e1 + · · · + unen and v = v1e1 + · · · + vnen, the
Hermitian product u · v is expressed as
n
X
u · v = (u1e1 + · · · + unen) · (v1e1 + · · · + vnen) = ui v i ,
i=1

and the norm kuk as

v
u n
uX
kuk = ku1e1 + · · · + unenk = t |ui|2.
i=1

Proposition 9.9 also holds unchanged.

Proposition 11.10. Given any nontrivial Hermitian

space E of finite dimension n 1, for any subspace F
of dimension k, the orthogonal complement F ? of F
has dimension n k, and E = F F ?. Furthermore,
we have F ?? = F .
11.3. LINEAR ISOMETRIES (ALSO CALLED UNITARY TRANSFORMATIONS) 609

11.3 Linear Isometries (also called Unitary Transfor-

mations)

In this section, we consider linear maps between Hermi-

tian spaces that preserve the Hermitian norm.

All definitions given for Euclidean spaces in Section 9.3

extend to Hermitian spaces, except that orthogonal trans-
formations are called unitary transformation, but Propo-
sition 9.10 only extends with a modified condition (2).

Indeed, the old proof that (2) implies (3) does not work,
and the implication is in fact false! It can be repaired by
strengthening condition (2). For the sake of completeness,
we state the Hermitian version of Definition 9.3.

Definition 11.4. Given any two nontrivial Hermitian

spaces E and F of the same finite dimension n, a function
f : E ! F is a unitary transformation, or a linear
isometry i↵ it is linear and
kf (u)k = kuk ,
for all u 2 E.
610 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Proposition 9.10 can be salvaged by strengthening condi-

tion (2).

Proposition 11.11. Given any two nontrivial Her-

mitian space E and F of the same finite dimension
n, for every function f : E ! F , the following prop-
erties are equivalent:
(1) f is a linear map and kf (u)k = kuk, for all u 2 E;
(2) kf (v) f (u)k = kv uk and f (iu) = if (u), for
all u, v 2 E;
(3) f (u) · f (v) = u · v, for all u, v 2 E.
Furthermore, such a map is bijective.

Observe that from f (iu) = if (u), for u = 0, we get

f (0) = if (0), which implies that f (0) = 0.
11.3. LINEAR ISOMETRIES (ALSO CALLED UNITARY TRANSFORMATIONS) 611

Remarks: (i) In the Euclidean case, we proved that the

assumption

(2’) kf (v) f (u)k = kv uk, for all u, v 2 E, and

f (0) = 0;

implies (3). For this, we used the polarization identity

2u · v = kuk2 + kvk2 ku vk2 .

In the Hermitian case, the polarization identity involves

the complex number i.

In fact, the implication (2’) implies (3) is false in the

Hermitian case! Conjugation z 7! z satifies (2’) since
|z2 z1| = |z2 z1| = |z2 z1|,
and yet, it is not linear!
612 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

(ii) If we modify (2) by changing the second condition by

now requiring that there is some ⌧ 2 E such that

f (⌧ + iu) = f (⌧ ) + i(f (⌧ + u) f (⌧ ))

for all u 2 E, then the function g : E ! E defined such

that

g(u) = f (⌧ + u) f (⌧ )

satisfies the old conditions of (2), and the implications

(2) ! (3) and (3) ! (1) prove that g is linear, and thus
that f is affine.

In view of the first remark, some condition involving i is

needed on f , in addition to the fact that f is distance-
preserving.

We are now going to take a closer look at the isometries

f : E ! E of a Hermitian space of finite dimension.
11.4. THE UNITARY GROUP, UNITARY MATRICES 613

11.4 The Unitary Group, Unitary Matrices

In this section, as a mirror image of our treatment of

the isometries of a Euclidean space, we explore some of
the fundamental properties of the unitary group and of
unitary matrices.

Definition 11.5. Given a complex m ⇥ n matrix A,

the transpose A> of A is the n ⇥ m matrix A> = (a>
i j)
defined such that
a>i j = aj i
and the conjugate A of A is the m ⇥ n matrix A = (bi j )
defined such that
b i j = ai j
for all i, j, 1  i  m, 1  j  n. The adjoint A⇤ of A
is the matrix defined such that
A⇤ = (A>) = (A)>.
614 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Proposition 11.12. Let E be any Hermitian space of

finite dimension n, and let f : E ! E be any linear
map. The following properties hold:
(1) The linear map f : E ! E is an isometry i↵
f f ⇤ = f ⇤ f = id.

(2) For every orthonormal basis (e1, . . . , en) of E, if

the matrix of f is A, then the matrix of f ⇤ is the
adjoint A⇤ of A, and f is an isometry i↵ A satisfies
the idendities
A A⇤ = A⇤A = In,
where In denotes the identity matrix of order n,
i↵ the columns of A form an orthonormal basis of
Cn, i↵ the rows of A form an orthonormal basis of
Cn .

Proposition 9.12 also motivates the following definition.

11.4. THE UNITARY GROUP, UNITARY MATRICES 615

Definition 11.6. A complex n ⇥ n matrix is a unitary

matrix i↵
A A⇤ = A⇤A = In.

Remarks: The conditions A A⇤ = In, A⇤A = In, and

A 1 = A⇤, are equivalent.

Given any two orthonormal bases (u1, . . . , un) and

(v1, . . . , vn), if P is the change of basis matrix from
(u1, . . . , un) to (v1, . . . , vn), it is easy to show that the
matrix P is unitary.

The proof of Proposition 11.11 (3) also shows that if f is

an isometry, then the image of an orthonormal basis
(u1, . . . , un) is an orthonormal basis.

If f is a unitary transformation and A is its matrix with

respect to any orthonormal basis, we have | det(A)| = 1.
616 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Definition 11.7. Given a Hermitian space E of dimen-

sion n, the set of isometries f : E ! E forms a subgroup
of GL(E, C) denoted as U(E), or U(n) when E = Cn,
called the unitary group (of E). For every isometry,
f , we have | det(f )| = 1, where det(f ) denotes the de-
terminant of f . The isometries such that det(f ) = 1
are called rotations, or proper isometries, or proper
unitary transformations, and they form a subgroup of
the special linear group SL(E, C) (and of U(E)), de-
noted as SU(E), or SU(n) when E = Cn, called the
special unitary group (of E). The isometries such that
det(f ) 6= 1 are called improper isometries, or improper
unitary transformations, or flip transformations.

The Gram–Schmidt orthonormalization procedure imme-

diately yields the QR-decomposition for matrices.

Proposition 11.13. Given any n ⇥ n complex matrix

A, if A is invertible then there is a unitary matrix Q
and an upper triangular matrix R with positive diag-
onal entries such that A = QR.

The proof is absolutely the same as in the real case!

11.4. THE UNITARY GROUP, UNITARY MATRICES 617

We have the following version of the Hadamard inequality

for complex matrices.

Proposition 11.14. (Hadamard) For any complex

n ⇥ n matrix A = (aij ), we have
n ✓X
Y n ◆1/2
| det(A)|  |aij |2
i=1 j=1
n ✓X
Y n ◆1/2
and | det(A)|  |aij |2 .
j=1 i=1

Moreover, equality holds i↵ either A has a zero col-

umn in the left inequality or a zero row in the right
inequality, or A is unitary.
618 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

We also have the following version of Proposition 9.15 for

Hermitian matrices.

Proposition 11.15. (Hadamard) For any complex

n ⇥ n matrix A = (aij ), if A is Hermitian positive
semidefinite, then we have
n
Y
det(A)  aii.
i=1

Moreover, if A is positive definite, then equality holds

i↵ A is a diagonal matrix.
11.5. ORTHOGONAL PROJECTIONS AND INVOLUTIONS 619

11.5 Orthogonal Projections and Involutions

In this section, we assume that the field K is not a field

of characteristic 2.

Recall that a linear map f : E ! E is an involution i↵

f 2 = id, and is idempotent i↵ f 2 = f . We know from
Proposition 3.5 that if f is idempotent, then

E = Im(f ) Ker (f ),

and that the restriction of f to its image is the identity.

For this reason, a linear involution is called a projection.

Proposition 11.16. For any linear map f : E ! E,

we have f 2 = id i↵ 12 (id f ) is a projection i↵ 12 (id+f )
is a projection; in this case, f is equal to the di↵erence
of the two projections 12 (id + f ) and 12 (id f ).
620 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Let U + = Ker ( 12 (id f )) and let U = Im( 12 (id f )).

If f 2 = id, then
(id + f ) (id f ) = id f 2 = id id = 0,
which implies that
✓ ◆ ✓ ◆
1 1
Im (id + f ) ✓ Ker (id f) .
2 2

1
Conversely, if u 2 Ker 2 (id f ) , then f (u) = u, so
1 1
(id + f )(u) = (u + u) = u,
2 2
and thus
✓ ◆ ✓ ◆
1 1
Ker (id f) ✓ Im (id + f ) .
2 2
Therefore,
✓ ◆ ✓ ◆
1 1
U + = Ker (id f) = Im (id + f ) ,
2 2

and so, f (u) = u on U + and f (u) = u on U .

11.5. ORTHOGONAL PROJECTIONS AND INVOLUTIONS 621

The involutions of E that are unitary transformations are

characterized as follows.

Proposition 11.17. Let f 2 GL(E) be an involu-

tion. The following properties are equivalent:
(a) The map f is unitary; that is, f 2 U(E).
(b) The subspaces U = Im( 12 (id f )) and
U + = Im( 12 (id + f )) are orthogonal.
Furthermore, if E is finite-dimensional, then (a)
and (b) are equivalent to
(c) The map is self-adjoint; that is, f = f ⇤.

A unitary involution is the identity on

U + = Im( 12 (id + f )), and
f (v) = v for all v 2 U = Im( 12 (id f )).

Furthermore, E is an orthogonal direct sum

E = U+ U .

We say that f is an orthogonal reflection about U +.

622 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

In the special case where U + is a hyperplane, we say

that f is a hyperplane reflection. We already studied
hyperplane reflections in the Euclidean case; see Chapter
10.

If f : E ! E is a projection (f 2 = f ), then
(id 2f )2 = id 4f + 4f 2 = id 4f + 4f = id,
so id 2f is an involution. As a consequence, we get the
following result.

Proposition 11.18. If f : E ! E is a projection

(f 2 = f ), then Ker (f ) and Im(f ) are orthogonal i↵
f⇤ = f.

A projection such that f = f ⇤ is called an orthogonal

projection.
11.5. ORTHOGONAL PROJECTIONS AND INVOLUTIONS 623

If (a1 . . . , ak ) are k linearly independent vectors in Rn, let

us determine the matrix P of the orthogonal projection
onto the subspace of Rn spanned by (a1, . . . , ak ).

Let A be the n ⇥ k matrix whose jth column consists of

the coordinates of the vector aj over the canonical basis
(e1, . . . , en).

The matrix P of the projection onto the subspace spanned

by (a1 . . . , ak ) is given by

P = A(A>A) 1A>.

The reader should check that P 2 = P and P > = P .

624 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

11.6 Dual Norms

If (E, k k) and (F, k k) are two normed vector spaces and

if we let L(E; F ) denote the set of all continuous (equiv-
alently, bounded) linear maps from E to F , then, we can
define the operator norm (or subordinate norm) k k on
L(E; F ) as follows: for every f 2 L(E; F ),

kf (x)k
kf k = sup = sup kf (x)k .
x2E kxk x2E
x6=0 kxk=1

In particular, if F = C, then L(E; F ) = E 0 is the dual

space of E, and we get the operator norm denoted by
k k⇤ given by

kf k⇤ = sup |f (x)|.
x2E
kxk=1

The norm k k⇤ is called the dual norm of k k on E 0.

11.6. DUAL NORMS 625

Let us now assume that E is a finite-dimensional Hermi-

tian space, in which case E 0 = E ⇤.

Theorem 11.5 implies that for every linear form f 2 E ⇤,

there is a unique vector y 2 E so that

f (x) = hx, yi, for all x 2 E,

and so we can write

kf k⇤ = sup |hx, yi|.

x2E
kxk=1

The above suggests defining a norm k kD on E.

626 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Definition 11.8. If E is a finite-dimensional Hermitian

space and k k is any norm on E, for any y 2 E we let

kykD = sup |hx, yi|,

x2E
kxk=1

be the dual norm of k k (on E). If E is a real Euclidean

space, then the dual norm is defined by

kykD = sup hx, yi

x2E
kxk=1
for all y 2 E.

Beware that k k is generally not the Hermitian norm as-

sociated with the Hermitian innner product.

The dual norm shows up in convex programming; see

Boyd and Vandenberghe [8], Chapters 2, 3, 6, 9.
11.6. DUAL NORMS 627

The fact that k kD is a norm follows from the fact that

k k⇤ is a norm and can also be checked directly.

It is worth noting that the triangle inequality for k kD

comes “for free,” in the sense that it holds for any function
p : E ! R.

If p : E ! R is a function such that

(1) p(x) 0 for all x 2 E, and p(x) = 0 i↵ x = 0;
(2) p( x) = | |p(x), for all x 2 E and all 2 C;
(3) p is continuous, in the sense that for some basis
(e1, . . . , en) of E, the function
(x1, . . . , xn) 7! p(x1e1 + · · · + xnen)
from Cn to R is continuous;
then we say that p is a pre-norm.
628 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Obviously, every norm is a pre-norm, but a pre-norm may

not satisfy the triangle inequality.

However, we just showed that the dual norm of any pre-

norm is actually a norm.

It is not hard to show that

kykD = sup |hx, yi| = sup <hx, yi.

x2E x2E
kxk=1 kxk=1

Proposition 11.19. For all x, y 2 E, we have

|hx, yi|  kxk kykD
|hx, yi|  kxkD kyk .
11.6. DUAL NORMS 629

We can show that

kykD
1 = kyk1
kykD
1 = kyk1
kykD
2 = kyk2 .

Thus, the Euclidean norm is autodual. More generally, if

p, q 1 and 1/p + 1/q = 1, we have

kykD
p = kykq .

It can also be shown that the dual of the spectral norm

is the trace norm (or nuclear norm) from Section 14.3.
630 CHAPTER 11. BASICS OF HERMITIAN GEOMETRY

Theorem 11.20. If E is a finite-dimensional Hermi-

tian space, then for any norm k k on E, we have
kykDD = kyk
for all y 2 E.

The proof makes use of the fact that a nonempty, closed,

convex set has a supporting hyperplane through each
of its boundary points, a result known as Minkowski’s
lemma.

This result is a consequence of the Hahn–Banach theo-

rem; see Gallier [14].

More details on dual norms and unitarily invariant norms

can be found in Horn and Johnson [19] (Chapters 5 and
7).