Classes of Operators on Hilbert Spaces
Extended Lecture Notes
Jan Hamhalter
Department of Mathematics
Faculty od Electrical Engineering
Technicka 2, 166 27 Prague 6
Czech Republic
e-mail: [email protected]
http://math.feld.cvut.cz/hamhalte
April 10, 2008
1
1 Basic Definitions and Examples
We shall continue to denote by B(H) the set of all bounded operators acting
on a complex Hilbert space H. For T ∈ B(H) let R(T ) = T (H). I will
denote the identity operator acting on H.
Algebraic properties inside B(H) endowed with the ∗ operation lead to
striking analytic spectral properties.
numbers functions operators
complex, zz = zz complex normal, T ∗ T = T T ∗ .
real, z = z real self-adjoint, T = T ∗ .
positive, zz positive positive, T ∗ T
complex unit into unit circle unitary, T ∗ T = T T ∗ = I
{0, 1} indicator function projection, T = T 2 = T ∗
Recall that there is a bijection between bounded operators on H and
bounded conjugate linear bilinear forms on H given by
T ∈ B(H) → BT , where BT (x, y) = (T x, y) x, y ∈ H .
In fact, there is a bijection between operators acting on H and quadratic
forms on H given by the assignment
T ∈ B(H) → QT , where QT (x) = (T x, x) x∈H.
This is the content of the following proposition.
1.1. Proposition. For T1 , T2 ∈ B(H) we have that QT1 = QT2 if, and only
if, T1 = T2 .
Proof: Polarization identity for T ∈ B(H) gives us
(T x, y) = (T (x + y), x + y) − (T (x − y), x − y)
+i(T (x + iy), x + iy) − i(T (x − iy), x − iy) .
Consequently, (T1 x, x) = (T2 x, x) for all x ∈ H implies that BT1 = BT2 and
so T1 = T2 .
1
1.2. Definition. An operator T ∈ B(H) is called
(i) normal if T T ∗ = T ∗ T .
(ii) self-adjoint if T = T ∗ .
(iii) positive if (T x, x) ≥ 0 for all x ∈ H.
(iv) unitary if T ∗ T = T T ∗ = I
(v) projection if T = T 2 = T ∗ .
1.3. Proposition. Let T ∈ B(H). Then
(i) T is normal if, and only if, kT xk = kT ∗ xk for all x ∈ H.
(ii) T is self-adjoint if, and only if, (T x, x) is real for all x ∈ H.
(iii) T is unitary if, and only if, T is an inner product preserving surjection.
Proof: (i) For all x ∈ H we have
(T ∗ T x, x) − (T T ∗ x, x) = (T x, T x) − (T ∗ x, T ∗ x) = kT xk2 − kT ∗ xk2 .
This, together with Proposition 1.1, implies (i).
(ii) For all x ∈ H we have
(T x, x) − (T ∗ x, x) = (T x, x) − (x, T x) = 2i =(T x, x).
It means that T = T ∗ if, and only if, (T x, x) is real for all x ∈ H.
(iii) If T is unitary, then, for x, y ∈ H,
(T x, T y) = (x, T ∗ T x) = (x, x) .
If x ∈ R(T )⊥ , then
0 = (T T ∗ x, x) = (x, x)
and so x = 0. Hence, T is a surjection which preserves the inner product.
On the other hand, if T is an inner product preserving surjection, then T
has an inverse and
(T ∗ T x, x) = (T x, T x) = (x, x) ,
which implies that T −1 = T ∗ and so T ∗ T = T T ∗ = I.
2
1.4. Example. (Discrete diagonal) Let H be a Hilbert space with orthonor-
mal basis (en )∞
n=1 . Let g : N → C be a bounded function. Define
∞
X
Tx = g(n) (x, en )en .
n=1
(This is a diagonal operator uniquely given by T en = g(n) en .)
Let us observe that
kT k = sup |g(n)| .
n∈N
As (T en , en ) = (en , g(n)en ) we see that
∞
X
∗
T x= g(n) (x, en )en
n=1
for all x ∈ H. All diagonal operators are normal. In a certain sense the
converse holds - see below.
• T is self-adjoint ⇐⇒ g(n) = g(n) for all n ∈ N (g is real).
• T is positive ⇐⇒ (T x, x) ≥ 0 for all x ⇐⇒ n g(n)|(x, en )|2 ≥ 0 for all
P
x ∈ H. But this is equivalent to g(n) ≥ 0 for all n ∈ N.
• T is unitary ⇐⇒ T ∗ T = T T ∗ = I. P
But T ∗ T x = n=1 g(n)g(n)(x, en ) en = ∞
∞ 2
P
n=1 |g(n)| (x, en )en .
In other words, T is unitary ⇐⇒ |g(n)| = 1 for all n.
Now we shall deal with a generalization of this example to cover also
operators with continuous spectrum.
1.5. Example. Let (X, µ) be a σ-finite measure space. For a measurable
function, f , on X define
kf k∞ = inf{K ≥ 0 | |f (x)| ≤ K for a.a. x ∈ X}
= sup{L ≥ 0 | |f | > L on some set of nonzero measure} .
L∞ (X, µ) ... space of all measurable function with finite k · k∞ . Put H =
L2 (X, µ) and fix f ∈ L∞ (X, µ).
Define multiplication operator Mf acting on H by
Mf (g) = f g , g ∈ L2 (X, µ) .
3
Observe that
Z Z
2
|Mf g(x)| dµ(x) = |f (x)|2 |g(x)|2 dµ(x) ≤ kf k2∞ kgk2 .
X X
It implies kMf k ≤ kf k∞ .
On the other hand, if 0 < α < kf k∞ , then µ{x | |f (x)| > α} > 0 and so
there is a set Y ⊂ {x | |f (x)| > α} of finite nonzero measure. Then, for the
characteristic function χY , of the set Y
Z
2
kMf (χY )k = |f (x)|2 dµ(x) > α2 µ(Y ) = α2 kχY k2 .
Y
Therefore, kMf k ≥ α. In summary,
kMf k = kf∞ k .
We have Mf∗ = Mf and so Mf∗ Mf = Mf Mf∗ = M|f |2 . Notably, Mf is
normal.
Deep spectral theorem says that all normal operators arise in
this way !
• Mf is self-adjoint ⇐⇒ f is real:
kMf − Mf∗ k = kMf −f¯k = kf − f¯k∞ .
• Mf is positive ⇐⇒ f ≥ 0 a.e. :
Z
(Mf g, g) = f (x) |g(x)|2 dµ(x) ,
X
R
and so Mf is positive if, and only if, Y f (x)dµ(x) ≥ 0 for each measurable
set Y , which is equivalent to f ≥ 0 a.e.
• Mf is unitary ⇐⇒ |f | = 1 a.e. :
kI − Mf Mf∗ k = kM(1−|f |2 ) k = k1 − |f |2 k∞ .
4
• Mf is a projection ⇐⇒ f = χY for some measurable Y :
f is real and
kMf − Mf2 k = kMf −f 2 k = kf − f 2 k∞ .
Consequently, f is 0 or 1 a.e.
Note that Mf may have no eigenvalue: Consider L∞ [0, 1] and f (x) = x.
Let λ ∈ C.
xg(x) = λg(x) for almost all x implies (x − λ)g(x) = 0 and so g = 0 a.e. .
Matrix point of view: If H is a Hilbert space with an orthonormal basis
(en ), then T ∈ B(H) is determined by an infinite matrix
((T em , en ))m,n .
T ∗ corresponds to adjoint matrix. In Example 1.4 we have seen T whose
matrix is diagonal.
1.6. Example. Let x, y ∈ H be distinct. Define
Tx,y (z) = (z, x) y z ∈H.
For u, v ∈ H we obtain
(Tx,y u, v) = (u, x)(y, v) = (u, (v, y)x) = (u, Ty,x (v)) ,
implying
∗
Tx,y = Ty,x .
Now
∗
Tx,y Tx,y (z) = (x, x)(z, y)y = kxk2 Ty,y (z) .
By exchanging the roles of x and y, we obtain
∗
Tx,y Tx,y = kyk2 Tx,x .
(If x and y are unit vectors, then the map Tx,y is called a partial isometry
exchanging one dimensional projections onto span of x and span of y, respec-
tively. This is important for the structure theory of projections – we have to
deal with non-normal operators!)
5
1.7. Proposition. Any projection P ∈ B(H) is an orthogonal projection of
H onto P (H).
Proof: Put M = P (H). M is closed because P 2 = P implies that
P (H) = {x ∈ H | P x = x}. We can write H = M ⊕ M ⊥ .
If y ∈ M ⊥ , then (x, P y) = (P x, y) = 0 for each x ∈ H and so P y = 0.
Therefore, if z = x + y, where x ∈ M and y ∈ M ⊥ , then
P (x + y) = x .
2 Spectral Theory of Normal Operators
Recall that for T ∈ B(H) the spectrum, Sp T , is a subset of C defined by
λ ∈ Sp T ⇐⇒ (T − λI) has not an inverse in B(H) .
Point spectrum
Spp T = σp (T ) = {λ ∈ C | (T − λI) is not one-to one .}
In other words, for each λ ∈ σp (T ) there is a nonzero y in H such that
Ty = λy .
Vector y is called an eigenvector.
Spectral radius r(T ) = sup{|λ| | λ ∈ Sp(T )}.
Spectrum is always a compact subset of C.
2.1. Proposition. Let T ∈ B(H) be normal. Then the following statements
hold:
(i) If T x = λ x for some λ ∈ C and x ∈ H, then T ∗ x = λx.
(ii) If λ1 6= λ2 are complex numbers, then
Ker(T − λ1 I) ⊥ Ker(T − λ2 I) .
6
Proof (i) By normality of T , for each x ∈ H,
k(T − λI)xk = k(T − λI)∗ xk = k(T ∗ − λ̄ I)xk .
It implies (i).
(ii) Suppose that x, y ∈ H and λ1 6= λ2 are in C such that T x = λ1 x,
T y = λ2 y. Then
λ1 (x, y) = (T x, y) = (x, T ∗ y) = (x, λ2 y) = λ2 (x, y) .
Since λ1 6= λ2 , (x, y) = 0.
Spectrum of a normal operator has a simpler structure than in general
case.
2.2. Proposition. Let T ∈ B(H) be a normal operator. Then
λ 6∈ Sp T ⇐⇒ there is c > 0 such that
k(T − λI)xk > ckxk for all x ∈ H . (1)
Proof: Without loss of generality assume that λ = 0. Suppose that there
is c > 0 satisfying the condition (1). Then T is one-to-one. It follows from
(1) that range T (H) is complete and thereby closed in H. It remains to prove
that R(T ) = H. Choose x ∈ R(T )⊥ . Then
0 = (x, T T ∗ x) = (x, T ∗ T x) = (T x, T x) = kT xk2 ≥ c2 kxk2 .
In other words, x = 0 and R(T ) = H. So (1) implies 0 6∈ Sp T . The reverse
implication is clear.
2.3. Corollary. If T ∈ B(H) is normal and λ ∈ Sp T \ σp (T ),
then (T − λI)(H) is not closed.
Proof: If T − λI is one-to-one and (T − λI)(H) is closed, then, by the
Inverse Mapping Theorem, there is a continuous linear map
S : (T − λI)(H) → H such that S(T − λI)x = x for all x ∈ H. It means
that kxk ≤ kSkk(T − λI)xk. As kSk = 6 0, we see that
1
k(T − λI)xk ≥ kxk .
kSk
In view of Proposition 2.2, λ 6∈ Sp T.
7
2.4. Corollary. (Approximate Spectrum) If T ∈ B(H) is normal, then λ ∈
Sp T if, and only if, there is a sequence (xn ) of unit vectors such that
k(T − λI)xn k → 0 as n → ∞.
Proof: By Proposition 2.2 λ ∈ Sp T ⇐⇒ inf kxk=1 k(T − λI)xk = 0.
Spectrum of a normal operator is equal to approximate point spectrum.
2.5. Corollary. If T ∈ B(H) is normal, then
Sp T ⊂ {(T x, x) | kxk = 1} .
Proof: If λ ∈ Sp T then there is a sequence (xn ) of unit vectors such that
kT xn − λxn k → 0 .
It implies
(T xn − λxn , xn ) → 0
(T xn , xn ) → λ .
2.6. Theorem. If T ∈ B(H) is a normal operator, then the following state-
ments hold:
(i) T is self-adjoint if, and only if, Sp T ⊂ R.
(ii) T is positive if, and only if, Sp T ⊂ R+ .
(iii) T is unitary if, and only if, Sp T ⊂ {z ∈ C | |z| = 1}.
(iv) T is a projection if, and only if, Sp T ⊂ {0, 1}.
Proof: We shall prove the implications =⇒ (the reverse implications are
more complicated). In case of (i) and (ii), these implications follow from
Corollary 2.5.
Suppose that T is unitary. Then kT k = 1 and so |(T x, x)| ≤ kxk2 = 1
for all unit vectors x. By Proposition 2.5 Sp T is a subset of the unit disc. If
λ ∈ Sp T , then λ is nonzero and λ1 ∈ Sp T −1 = Sp T ∗ . As T ∗ is unitary we
have that | λ1 | ≤ 1. Hence, |λ| = 1.
(iv) is a consequence of proposition 1.7
8
2.7. Proposition. (C ∗ -property) If T ∈ B(H), then
kT ∗ T k = kT k2 .
Proof: First observe that kT k = kT ∗ k (consider e.g. corresponding bilin-
ear forms). For x ∈ H we have
kT xk2 = (T x, T x) = (T ∗ T x, x) ≤ kT ∗ T k · kxk2 .
Hence,
kT k2 ≤ kT ∗ T k ≤ kT ∗ k · kT k = kT k2 .
2.8. Proposition. If T ∈ B(H) is normal, then
r(T ) = kT k .
Proof: First suppose that T is self-adjoint. Then by the C ∗ -property
kT k2 = kT ∗ T k = kT 2 k .
If T is normal, then
kT 2 k2 = k(T 2 )∗ T 2 k = k(T ∗ T )2 k = kT ∗ T k2 = kT k4 .
(We have used the fact that T ∗ T is self-adjoint.) Consequently,
kT 2 k = kT k2 ,
n n
and in turn kT 2 k = kT k2 for all n. By the spectral radius formula
n n
r(T ) = lim kT n k1/n = lim kT 2 k1/2 = lim kT k = kT k .
n n n
Very useful concept in operator theory is that of numerical range of an
operator T ∈ B(H):
N (T ) = {(T x, x) | kxk = 1} .
By a numerical radius of T we mean
n(T ) = sup |(T x, x)| .
kxk=1
It is clear that, in general, n(T ) ≤ kT k.
9
2.9. Proposition. Let T ∈ B(H). Then the following statements hold
(i) If T is normal, then
kT k = r(T ) = n(T ) .
(ii) If T is self-adjoint, then kT k or −kT k is in Sp T .
Proof: (i) If T is normal, then Sp T ⊂ N (T ) by Proposition 2.5. Obvi-
ously,
r(T ) ≤ n(T ) ≤ kT k = r(T )
and so r(T ) = n(T ).
(ii) By working with kT k−1 T in place of T , we can assume that kT k = 1.
Then there is a sequence of unit vectors (xn ) such that kT xn k → 1. Thanks
to this
k(I − T 2 ) xn k2 = kxn k2 + kT 2 xn k2 − 2<(T 2 xn , xn ) ≤ 2 − 2kT xn k2 → 0
as n → ∞. We see that 1 ∈ Sp T 2 . It means that T + I or T − I has no
inverse, for otherwise,
T 2 − I = (T + I)(T − I)
would have an inverse, which is not possible.
In view of the previous result we can say that the norm of a normal
operator is given by the extreme of the corresponding quadratic form.
3 Algebraic Aspects and Applications
The facts mentioned below follow from Exercises. The set of normal operators
is stable under forming powers and scalar multiples. If T is normal, then the
smallest ∗-subalgebra of B(H) containing T is commutative. Any operator
T ∈ B(H) can be written as T = T1 + iT2 , where T1 and T2 are self-adjoint.
Moreover, any self adjoint operator is a difference of two positive operators.
If T is self-adjoint, then T 2 is always positive. The converse also holds. (The
proof is more complicated and will be omitted.)
3.1. Proposition. For T ∈ B(H) the following conditions are equivalent:
10
(i) T is positive
(ii) T = A∗ A for some A ∈ B(H).
(iii) T = S 2 for some self-adjoint S ∈ B(H). (S is denoted by T 1/2 and
called the square root of T ).
If T is self-adjoint, then eiT is unitary. The converse also holds:
3.2. Proposition. For any unitary operator U ∈ B(H) there is a self-
adjoint operator T ∈ B(H) with kT k ≤ 2π such that U = eiT .
In Physics: t ∈ R → eitH , where H is Hamiltonian (Energy), describes
time development of the system (solution of the Schrödinger equation).
Another important example of a unitary map is the Fourier-Plancherel
transform: Z
ˆ 1
2
f ∈ L (R) → f (ω) = √ f (t)e−itω dt .
2π R
4 Compact Operators
Notation:
B – Banach space
For x ∈ B and ε > 0 denote Bε (x) = {y | kx − yk ≤ ε}
B1 = B1 (0).
4.1. Definition. A set X in a Banach space B is said to be compact if for
each system U of open subsets of B with X ⊂ ∪O∈U O there is a finite subset
U 0 ⊂ U with X ⊂ ∪O∈U 0 O. A set X ⊂ B is said to be relatively compact if
its closure, X, is compact.
Related concept to compactness is total boundedness.
4.2. Definition. A set X in a Banach space B is said to be totally bounded
if for each ε > 0 there exist x1 , . . . , xn ∈ X such that
X ⊂ ∪ni=1 Bε (xi ) .
4.3. Theorem. X ⊂ B is compact if, and only if, X is closed and totally
bounded. X ⊂ B is relatively compact if, and only if, X is totally bounded.
11
Basic facts about compact sets:
• If X ⊂ B is relatively compact, then for each sequence (xn ) ⊂ X there
is a cauchy subsequence (xnk ).
• Any relatively compact set is bounded.
• Any bounded set in a finite-dimensional space is relatively compact.
• Unit ball B1 is compact if, and only if, dim B < ∞.
• Let f : B1 → B2 be a continuous map between Banach spaces. If
X ⊂ B1 is (relatively) compact, then the image f (X) is (relatively)
compact in B2 .
4.4. Definition. A linear operator T : F → G between Banach spaces F
and G is called compact if
T (F1 ) is relatively compact .
Basic facts about compact operators:
T : F → G is a linear map between Banach spaces.
• T is compact if it maps bounded sets to relatively compact sets. In
particular, compact maps are continuous.
• If T is compact, then for each bounded sequence (xn ) ⊂ F there is a
subsequence (xnk ) such that (T xnk ) is convergent.
• The identity map on a Banach space B is compact if, and only if,
dim B < ∞.
• Any bounded operator with finite-dimensional range is compact.
4.5. Corollary. If T : B → B is a compact map, then for each nonzero
λ∈C
dim Ker(T − λI) < ∞
12
Proof: T restricted to Ker(T − λI) is a nonzero multiple of I. Therefore
T : Ker(T − λI) → Ker(T − λI) is compact if, and only if, dim Ker(T − λI) <
∞.
Compact operators on Banach spaces have special spectral properties.
4.6. Theorem. Let T : B → B be a compact operator. Then Sp T is count-
able, and each nonzero point of Sp T is an eigenvalue and an isolated point of
Sp T . For each nonzero λ ∈ Sp T , the space Ker(T −λI) has finite dimension.
We shall prove this theorem for normal compact operators on Hilbert
spaces later.
Notation:
K(H) ... compact operators acting on a Hilbert space H
F (H) ... finite rank operators acting on a Hilbert space H.
(T ∈ F (H) if, and only if, dim T (H) < ∞.)
These classes of operators form a special structure in B(H)
4.7. Definition. An ideal J ⊂ B(H) is a linear subspace of B(H) such that
S T, TS∈J
whenever T ∈ J and S ∈ B(H).
4.8. Proposition. (i) F (H) ⊂ K(H) and each T ∈ F (H) is a linear com-
bination of the operators of the form
Tx,y (z) = (z, x) y ,
where x, y ∈ H.
(ii) F (H) and K(H) are ideals in B(H). Moreover, K(H) is a closed
ideal in B(H).
13
Proof: (i) F (H) ⊂ K(H) because any finite-dimensional operator is com-
pact.
Take T ∈ F (H) and let P be the projection of H onto R(T ). Then
P = P1 + P2 + · · · + Pn ,
where each Pi is a one-dimensional projection. As
T = PT
the problem of description of T reduces to a rank one operator: dim R(T ) = 1.
Suppose R(T ) = span{y}, where kyk = 1. Then, for each z ∈ H,
T z = (T z, y)y = (z, T ∗ y)y
and so T = TT ∗ y,y .
(ii) K(H) is a subspace of B(H) because the sum of finitely many totally
bounded sets is totally bounded and scalar multiple of a totally bounded set
is totally bounded as well. If S ∈ B(H), then
T ∈ F (H) =⇒ ST, T S ∈ F (H) (linear algebra)
T ∈ K(H) =⇒ ST, T S ∈ K(H).
(Last implication is due to the fact that bounded operators map rela-
tively compact sets to relatively compact sets and that continuous image of
a relatively compact set is relatively compact.)
Closedness of K(H):
Suppose that (Tn ) ⊂ K(H), Tn → T ∈ B(H). Given ε > 0 there is n0
such that kT − Tn k < ε/3 whenever n ≥ n0 . There are x1 , . . . , xk ∈ H1 such
that
k
[ ε
Tn0 (H1 ) ⊂ B(Tn0 xi , ) .
i=1
3
Now for any x ∈ H1 , T x ∈ T (H1 ) and
kT x − Tn0 xk ≤ ε/3 .
14
There is 1 ≤ j ≤ k such that
kTn0 x − Tn0 xj k ≤ ε/3
and so
kT x − T xj k ≤ kT x − Tn0 xk + kTn0 x − Tn0 xj k + kTn0 xj − T xj k
≤ ε/3 + ε/3 + ε/3 = ε.
Hence, T (H1 ) is totally bounded.
4.9. Example. Suppose that (en )∞ n=1 is an orthonormal basis of H and T ∈
1
B(H) is defined by T en = n en . (T is a diagonal operator.) Set
N
X 1
TN x = (x, en )en .
n=1
n
Then
∞ ∞
2
X 1 2
X 1 1
k(T − TN )xk = k (x, en )en k = 2
|(x, en )|2 ≤ 2
kxk2 .
n=N +1
n n=N +1
n (N + 1)
Therefore kT − TN k → 0 as N → ∞. By the previous result T is compact.
(Observe that the same is true whenever T en = λn en , where λn → 0.)
c0 ,→ K(H) – noncommutative c0 .
We shall now develop theory of self-adjoint and normal compact opera-
tors.
4.10. Proposition. If T ∈ K(H) is self-adjoint, then kT k or −kT k must
be an eigenvalue of T .
Proof: (We know that kT k or −kT k is in Sp T .)
Without loss of generality assume that kT k = 1. Then
1 = kT k = sup |(T x, x)| .
kxk=1
15
As (T x, x) is real for all x, there exists a sequence (xn ) of unit vectors such
that
(T xn , xn ) → 1( or → − 1 which is the same) .
Using compactness we can pass to a subsequence of (xn ), denoted by the
same symbol, such that
T xn → x ∈ H1 .
Then
(x, xn ) → 1 and so xn → x .
Now xn → x, T xn → x implies T x = x.
4.11. Proposition. Let T ∈ K(H) and (en )∞
n=1 be an orthonormal sequence
in H. Then
T en → 0 as n → ∞ .
Proof: Without loss of generality we can assume that
lim T en = x .
n
Suppose that x 6= 0 and try to reach a contradiction. Given n we can find
k(n) such that
1
kT em − xk ≤ √
n
for all m ≥ k(n). Now set
1
un = √ (ek(n) + ek(n)+1 + · · · + ek(n)+n−1 ).
n
Then kun k = 1. As
k(n)+n−1 k(n)+n−1
X X 1
k T ei k ≥ knxk − kT ei − xk ≥ nkxk − n √
n
i=k(n) i=k(n)
we obtain
1 h 1 i
kT un k ≥ √ n · kxk − √
n n
√
= nkxk − 1 → ∞ as n → ∞ .
So T is unbounded - a contradiction.
16
4.12. Theorem. (Spectral theorem for normal compact operators) Let H
be a separable Hilbert space of infinite dimension and T a normal compact
operator acting on H. Then there is an orthonormal basis (en )∞
n=1 of H and
a sequence of complex numbers λn → 0 such that
∞
X
Tx = λn (x, en )en (2)
n=1
for all x ∈ H.
Proof: We show first that T is diagonalizable. By Zorn’s lemma there is
a maximal orthonormal set E of eigenvectors of T . If L is the closed linear
span of E, then H = L ⊕ L⊥ . Observe that L⊥ is T -invariant. For this
fix x ∈ L⊥ and take arbitrary y ∈ E. Then there is a scalar λ such that
T y = λy. It gives
(y, T x) = (T ∗ y, x) = (λy, x) = 0 .
Therefore, T restricts to a compact normal operator acting on L⊥ . We are
going to show that L⊥ = {0}.
First we show that any nonzero point λ in the spectrum of T is an ein-
genvalue. By Corollary 2.4 there is a sequence (xn ) in H1 such that
T xn − λxn → 0
as n → ∞. As T is compact we can, by passing to a subsequence, assume
that
lim T xn = y .
n
y
Then λxn → y and so xn → λ
. In turn, y 6= 0,
Ty
y = lim T xn = ,
n λ
saying that
T y = λy .
Now, if L were nonzero, then T would have a nonzero eigenvector in L⊥ ,
⊥
which is excluded by maximality of E. Hence L⊥ = {0}.
Summing it up, E is an orthonormal basis of H and so that T is diago-
nalizable. In other words, T is of the form (2) for some sequence (λn ). That
λn → 0 follows from Proposition 4.11.
17
5 Trace Class and Hilbert-Schmidt Opera-
tors
• Applications to integral equations, Gaussian stochastic processes, uni-
tary representations of locally compact groups, ...
• quantization of `1 , `2 .
5.1. Definition. Let T ∈ B(H) be a positive operator and (en )∞
n=1 an or-
thonormal basis of H. Define
∞
X
trace T = (T en , en ) .
n=1
(It may happen that trace T = ∞.)
Remarks: In the matrix representation trace T is a sum of diagonal ele-
ments.
5.2. Proposition. (i) For a given positive T ∈ B(H), trace T does not
depend on the choice of an orthonormal basis (en ).
(ii)
trace(T1 + T2 ) = trace T1 + trace T2
trace(λT1 ) = λ trace T1 ,
whenever T1 , T2 ≥ 0 and λ ≥ 0.
Proof: (i) Fix two orthonormal bases (ek ) and (fk ) and T ∈ B(H). Then
∞
X ∞
X ∞
X
(T ek , ek ) = (T ek , fl )(ek , fl ) = (ek , T fl )(ek , fl )
k=1 l,k=1 l,k=1
X∞ X∞
= (T fl , ek )(fl , ek ) = (fl , T fl )
l,k=1 l=1
X ∞
= (T fl , fl ) .
l=1
(ii) obvious
18
5.3. Corollary. trace(U ∗ T U ) = trace T whenever U is unitary and T ≥ 0.
5.4. Example. If T is a positive operator acting on an n-dimensional Hilbert
space, then
trace T = λ1 + λ2 + · · · + λn ,
where λi ’s are eigenvalues of T (counted with multiplicity).
5.5. Definition. (i) A positive T ∈ B(H) is a trace class operator if
trace T < ∞.
(ii)
L1 (H) = span{T ≥ 0 | trace T < ∞}
is the set of trace class operators.
If T ∈ L1 (H), then
T = P1 − P2 + i(P3 − P4 ) ,
where Pi ≥ 0 and trace Pi < ∞. The decomposition is not unique, but the
basic properties of the trace imply that there is a unique linear functional,
denoted by trace, on L1 (H) defined by
trace T = trace P1 − trace P2 + i(trace P3 − trace P4 ) .
Obviously, for every T ∈ L1 (H) and every orthonormal basis e1 , e2 , . . . we
have ∞
X
trace T = (T en , en ) ,
n=1
where the series on the right hand side is absolutely convergent.
5.6. Definition. An operator T ∈ B(H) is called a Hilbert-Schmidt operator
if
trace(T ∗ T ) < ∞ .
L2 (H) ..... set of all Hilbert-Schmidt operators acting on H.
Observe that ∞
X
2
T ∈ L (H) ⇐⇒ kT en k2 < ∞
n=1
for any orthonormal basis (en )∞
n=1 of H.
19
5.7. Proposition. L2 (H) is a self-adjoint ideal in B(H).
Proof: Let A, B ∈ L2 (H).
parallelogram law:
(A + B)∗ (A + B) + (A − B)∗ (A − B) = 2 A∗ A + 2 B ∗ B .
It implies that
0 ≤ (A + B)∗ (A + B) ≤ 2A∗ A + 2B ∗ B .
Consequently,
trace[(A + B)∗ (A + B)] ≤ 2 trace A∗ A + 2 trace B ∗ B < ∞ ,
and so A + B ∈ L2 (H). Hence L2 (H) is a subspace of B(H).
We shall now prove that trace A∗ A = trace AA∗ (this is of independent im-
portance). For this fix two orthonormal basis (en ) and (fk ) of H. Then
∞
X ∞
X
2
kAen k = |(Aen , fk )|2
n=1 n,k=1
X∞ ∞
X
∗
= |(en , A fk )| = 2
|(A∗ fk , en )|2
n,k=1 n,k=1
X∞
= kA∗ fk k2 .
k=1
In other words, L2 (H) is self-adjoint. If B ∈ B(H) and A ∈ L2 (H), then
∞
X ∞
X
2 2
kBAen k ≤ kBk kAen k2 < ∞ .
n=1 n=1
So L2 (H) is a left ideal and by self-adjointeness it is an ideal.
5.8. Proposition.
L2 (H) ⊂ K(H) .
20
Proof: F (H) ⊂ K(H). For any x ∈ H1 and any T ∈ B(H)
∞
X
2
kT xk ≤ kT en k2 = trace T ∗ T ,
n=1
where (en )∞
n=1 is an orthonormal basis containing x. This means that
kT k2 ≤ trace T ∗ T .
Suppose now that T ∈ L2 (H). Fix an orthonormal basis (en )∞ n=1 . Let PN be
the orthogonal projection onto span{e1 , . . . , eN }. Put FN = T PN . Then
∞
X
2 ∗
kT −FN k ≤ trace((I −PN )T T (I −PN )) = kT en k2 → ∞ for N → ∞ .
n=N +1
Therefore T ∈ K(H).
5.9. Corollary. A normal Hilbert-Schmidt operator T is diagonalizable and
for its sequence (λn ) of eigenvalues we have
∞
X
|λn |2 < ∞ .
n=1
A natural inner product can be introduced on L2 (H).
For A, B ∈ L2 (H) define
∞
X ∞
X
(A, B)2 = (B ∗ Aen , en ) = (Aen , Ben ) .
n=1 n=1
X∞ ∞
X ∞
X
2
( |(Aen , Ben )| ≤ k(Aen k k(Ben k2 < ∞ .)
n=1 n=1 n=1
Note that kAk2 = trace A∗ A.
It can be proved that (L2 (H), (·, ·)2 ) is a Hilbert space.
21
Hilbert-Schmidt integral operator:
(X, µ) .... σ-finite measure space.
k(x, y) ∈ H = L2 (X × X, µ × µ) .
Define an operator T on L2 (X, µ) by
Z
T ξ(x) = k(x, y)ξ(y)dµ(y) . (3)
X
This definition is correct because k(x, ·) ∈ L2 (X, µ) for a.a. x ∈ X.
Another application of Fubini’s theorem implies that for ξ, ν ∈ L2 (X, µ)
Z Z
|T ξ(x)||ν(x)|dµ(x) ≤ |k(x, y)||ξ(x)||ν(x)|dµ(x)dµ(y) ≤ kkkkξk kνk
X X×X
Hence T ∈ B(L2 (X, µ)) and kT k ≤ kkk.
Let us now compute the trace of T ∗ T . Choose an orthonormal basis
(en )∞ 2
n=1 of L (X, µ). Then
umn (x, y) = en (x)em (y) m, n = 1, 2, . . . .
form an orthonormal basis of L2 (X × X, µ × µ). We observe
Z Z
(T em , en ) = T em (x)en (x)dµ(x) = k(x, y)en (x)em (y)dµ(y)dµ(x)
X X×X
= (k, umn ) .
In turn,
∞
X ∞
X
∗ 2
trace T T = |(T em , en )| = |(k, umn )|2 = kkk2 .
m,n=1 m,n=1
We summarize the results of this discussion in the following proposition:
22
5.10. Proposition. Let (X, µ) be a σ-finite measure space. For every func-
tion k ∈ L2 (X × X, µ × µ) there is a unique bounded operator Tk on L2 (X, µ)
satisfying Z
T ξ(x) = k(x, y)ξ(y)dµ(y) ξ ∈ L2 (X, µ) .
X
Then Tk is a Hilbert-Schmidt operator with the Hilbert-Schmidt norm kkk.
23