Spectral decomposition of bounded self-adjoint operators

Anders L. Mikkelsen Marco Piazzi Sigurδur Örn Stefánsson Andreas Vigand P.

March 30, 2006

Abstract
We list some properties of self-adjoint, bounded operators and their spectra and define an ordering
for them. Then we state and prove the Hilbert theorem which allows us to decompose such operators
in terms of their unique spectral family. We use this decomposition to define how to apply functions
to the operators.

1 Introduction
When dealing with operators it can be convenient to look at them from different angles in order to simplify
their structure. An example is from the finite dimensional case where one tries to diagonalize matrices,
essentially turning a matrix into a number which can then be acted on with functions in a sensible way.
Similarly in the infinite dimensional case, it is desirable to have some kind of a procedure to understand
operators in terms of a decomposition into simpler things. The Hilbert theorem gives such a procedure
for self-adjoint bounded operators. It proves the existence of a unique spectral family corresponding to
the operator and instructs how it is possible to decompose the operator in terms of this family.
In this paper we use in most cases the notation in [1] . We will need the following definitions and
properties.
An operator A : H → H where H is a Hilbert space is called bounded if there exists a C such that
kAxk ≤ Ckxk for all x ∈ H and the smallest such number is called the norm of A i.e.

kAxk
kAk = sup . (1.0.1)
x6=0 kxk

We denote the space of bounded operators on H with L(H).

If A is bounded it is said to be self-adjoint or symmetric (in the unbounded case these are not
equivalent) if A = A∗ i.e. hAx, yi = hx, Ayi for every x, y ∈ H. The norm of a bounded, self adjoint
operator A can be written in another way

|hAx, xi|
kAk = sup = sup |hAx, xi| (1.0.2)
x6=0 kxk2 kxk=1

which follows directly from the Cauchy-Schwarz inequality and the original definition of the norm. We
will use this form of the norm all the time.
The spectrum of an operator A ∈ L(H), denoted σ(A), is defined to be the set of all points which are
not regular, where a point λ ∈ C is said to be regular if the inverse of (A − λI) exists and is bounded.
We begin the paper by describing a few useful properties of self-adjoint, bounded operators and their
spectra. Then we define how to order such operators which provides us some important results for arriving
at the Hilbert theorem.

1
2 Spectra of self-adjoint operators March 30, 2006

2 Spectra of self-adjoint operators

Below we look at some properties of self-adjoint operators and their spectra. First we state and prove
some basic results which are frequently used later on.
Property 2.1. Let A be a self-adjoint operator on H. If L ⊆ H is an invariant subspace of H with
respect to A (i.e. A(L) ⊆ L) then so is L⊥ = {x ∈ H|∀y ∈ L, hx, yi = 0}.

Proof. Take y ∈ L⊥ . Then hx, Ayi = hAx, yi = 0 for all x ∈ L since A(L) ⊆ L. Thus Ay ∈ L⊥ .
Property 2.2. hAx, xi ∈ R for all x ∈ H if and only if A is self-adjoint.

Proof. Assume first that A is self-adjoint. Then hAx, xi = hx, Axi = hAx, xi so hAx, xi ∈ R for all x ∈ H.
For the other direction note that the following holds
4hAx, yi = hA(x + y), x + yi − hA(x − y), x − yi + ihA(x + iy), x + iyi − ihA(x − iy), x − iyi.
Using this and assuming that hAx, xi ∈ R for all x ∈ H we get
4hx, Ayi = 4hAy, xi = hA(y+x), y+xi−hA(y−x), y−xi−ihA(y+ix), y+ixi+ihA(y−ix), y−ixi = 4hAx, yi
and so A is self-adjoint.

Next we look at the spectrum of a self-adjoint operator. First, let us remember the general classification
of the spectrum σ(A) ⊆ C of an operator A ∈ L(H):

1. The point spectrum of A, denoted σp (A), is the set of eigenvalues of A.

2. The continuous spectrum of A, denoted σc (A), is the set of numbers λ which are not eigenvalues
and such that Im(A − λI) is dense in H.
3. The residual spectrum of A is then the set σr (A) = σ(A)\(σp (A) ∪ σc (A)).

However, if A is self-adjoint, this classification becomes simpler. It is easy to verify that the point
spectrum is then real and that eigenvectors corresponding to different eigenvalues are orthogonal. We
get the following properties.
Property 2.3. The residual spectrum of a self-adjoint operator A ∈ L(H) is empty and so σ(A) =
σp (A) ∪ σc (A).

Proof. First note that for any operator T on H we have ker(T ∗ ) = (Im(T ))⊥ since x ∈ ker(T ∗ ) iff T ∗ x = 0
iff hy, T ∗ xi = 0 for all y ∈ H iff hT y, xi = 0 for all y ∈ H iff x ∈ Im(T )⊥ . From this we see that for any
λ ∈ C we have ker(A − λI) = (Im(A − λI))⊥ . Now, assume λ ∈ / σp (A). Since the eigenvalues are real we
have that λ ∈ / σp (A) and therefore ker(A−λI) = {0}. Then ((Im(A−λI))⊥ )⊥ = Im(A − λI) = {0}⊥ = H
which means that λ ∈ σc (A). We thus conclude that the residual spectrum is empty.

In fact the whole spectrum is real as the following property tells us

Property 2.4. The spectrum of a self-adjoint operator A ∈ L(H) is real.

Proof. Take λ ∈ C\R and write λ = a + ib where b 6= 0. Then

k(A − λI)xk2 = k(A − aI)x − ibxk2 = k(A − aI)xk2 − h(A − aI)x, ibxi − hibx, (A − aI)xi + b2 kxk2
= k(A − aI)xk2 + b2 kxk2 ≥ b2 kxk2
Now, take a convergent sequence (A − λI)xn → y in Im(A − λI). The above inequality then shows that
(xn ) is a Cauchy sequence and since H is complete (xn ) has a limit which we call x. Finally since A − λI
is bounded it is clear that (A − λI)xn → (A − λI)x. So we see that Im(A − λI) is closed. Since b 6= 0 we
have that λ ∈/ σp (A) . Therefore ker(A − λI) = {0}. Then λ ∈/ σc (A) since if λ ∈ σc (A) then A − λI were
bijective by the above and then also bounded by the Banach open map theorem [ [1], Theorem 4.7.1,
pg. 69]. This would mean that λ were a regular point, a contradiction. By property 2.3 the residual
spectrum is empty and so λ must be a regular point.

2
3 Order on the space of self-adjoint operators March 30, 2006

Next property concerns the existence of a spectrum point of a self-adjoint operator.

Property 2.5. Let A ∈ L(H) be a self-adjoint operator. Then either kAk or −kAk are in the spectrum
of A.

Proof. By equation (1.0.1) we can choose a sequence (xn ) with kxn k = 1 for all n such that |hAxn , xn i| →
kAk. By property 2.2 it is possible to choose a subsequence (xni ) such that hAxni , xni i → λ where
λ = ±kAk. Then

0 ≤ k(A − λI)xni k2 = kAxni k2 − 2λhAxni , xni i + λ2 kxni k2 ≤ 2λ2 − 2λhAxni , xni i → 0.

If λ is a regular point we can write

1 = k(A − λI)−1 (A − λI)xni k ≤ k(A − λI)−1 kk(A − λI)xni k → 0

which is a contradiction, so λ is a spectrum point.

3 Order on the space of self-adjoint operators

Consider a self-adjoint operator, A = A∗ , on a Hilbert space H. A is called non-negative if and only if
hAx, xi ≥ 0 for all x ∈ H, i.e. if and only if hAx, xi ≥ 0 for all x with norm 1. We express this by A ≥ 0.
Since this only makes sense if hAx, xi is real, we have that A indeed must be self-adjoint. We now define
a relation ≥ on the space of self-adjoint operators.
Definition (Order on the space of self-adjoint operators) 3.1. Let A = A∗ , B = B ∗ ∈ L(H) be
two self-adjoint operators. We say that A ≤ B if B − A ≥ 0.

This relation is well defined and it is in fact a (partial) order on our space (this is almost trivial to
verify). The ordering 3.1 has the following properties
Property (Properties of the ordering) 3.2.

i) −I ≤ A ≤ I implies kAk ≤ 1. The first inequalities in fact mean that supkxk=1 |hAx, xi| ≤ 1.
Since we have for a symmetric operator that kAk = supkxk=1 |hAx, xi| we get kAk ≤ 1 and −IkAk ≤
A ≤ IkAk.
ii) If A ≥ 0 and A ≤ 0 this implies that A = 0, where we again use kAk = supkxk=1 |hAx, xi|.

p p
iii) Let A ≥ 0. Then |hAx, yi| ≤ hAx, xi · hAy, yi. This is just a slightly more general version
of Cauchy-Schwartz inequality, see for example [1] exercise 2.1.2.

iv) If C is self-adjoint and A ≤ B, then A + C ≤ B + C.

v) If A is symmetric, then A2n ≥ 0 ∀ n ∈ N, since hA2n x, xi = hAn x, An xi = kAn xk2 ≥ 0. If A ≥ 0

then we have that A2n+1 ≥ 0, since hA2n+1 x, xi = hAAn x, An xi ≥ 0. If A ≥ 0 then we can now see
that, since all polynomials in A will have the form P (A) = kn An + kn−1 An−1 + ... + k1 A + k0 , using
the rules of the inner product we have that hkn An + kn−1 An−1 + ... + k1 A + k0 x, xi = kn hAn x, xi +
kn−1 hAn−1 x, xi+...+k1 hAx, xi+k0 hx, xi and then all polynomials in A with non-negative coefficients
will satisfy P (A) ≥ 0.
Theorem 3.3. Let (An ) be a sequence of self-adjoint, bounded operators such that A0 ≤ A1 ≤ .... ≤
An ≤ ... ≤ A where A is self-adjoint and bounded. Then there exists a strong limit of (An )n (i.e., there
exists a bounded self-adjoint operator B and An x → Bx ∀ x ∈ H).

3
3 Order on the space of self-adjoint operators March 30, 2006

Proof. For every symmetric operator A there is a real number C so A ≤ C · I. So, if we choose our
sequence as (AnC−A
1
0)
where we choose C1 so that A − A0 ≤ C1 · I, we would get 0 ≤ (AnC−A
1
0)
≤ I and we
can therefore, without loss of generality, assume that our original sequence satisfies

0 ≤ An ≤ I.

For n > m we define Amn = An − Am ≥ 0. Since 0 ≤ An ≤ I we have kAmn k ≤ 1. Then using

the Cauchy-Schwartz inequality (first the generalized then the ordinary) it follows that for any x and
y = Amn x we have
p p
k(Am − An )xk2 = hAmn x, Amn xi = hAmn x, yi ≤ hAmn x, xi · hAmn y, yi
p p p p
= hAmn x, xi · hA2mn x, Amn xi ≤ hAmn x, xi · kA2mn xk · kAmn xk
p
≤ hAmn x, xi · kxk
since kAmn xk ≤ kxk and kA2mn xk ≤ kxk. This means that
p
kAm x − An xk2 ≤ |hAn x, xi − hAm x, xi| · kxk → 0

as n > m → ∞ for all x ∈ H, since the sequence hAn x, xi is monotone, increasing and bounded, and
therefore it converges. Hence An x is a Cauchy sequence and the limit limn→∞ An x ≡ Bx exists since
H is a Hilbert space. B depends linearly on x and 0 ≤ hAn x, xi ≤ hx, xi, so 0 ≤ hBx, xi ≤ kxk2 which
implies that B is bounded and self-adjoint by the continuity of the inner product.
Theorem (The main proposition) 3.4. Let A be such that

m·I ≤A≤M ·I (3.0.3)

for some m,M ∈ R and let P be a polynomial satisfying P (z) ≥ 0 for all z ∈ [m,M ]. Then P (A) ≥ 0.

In order to prove this we need the following lemma

Lemma 3.5. If A ≥ 0, B ≥ 0 and AB = BA, then AB ≥ 0. This lemma follows directly from the
existence of a unique square root of a self-adjoint operator [ [1], Lemma 6.3.5 pg. 98 ].

Now using lemma 3.5 we will prove the main proposition.

Proof. We assume that m < M (since in the case m=M we have A=0 and P (A)=0). Then P (z) ≥ 0 for
z ∈]m, M [ implies that P (z ) has real coefficients and it can be written in the form:
Y Y Y
P (z) = c (z − αi ) (βi − z) · ((z − γi )2 + δi2 ), γi , δi ∈ R
αi ≤m βi ≥M

where αi are all the real roots of the polynomial for z ≤ m and βi is the real roots for z ≥ M and c > 0.
The real roots on the interval ]m, M [ (which must have even multiplicity because P (z) ≥ 0) and all the
complex roots are described by ((z − γi )2 + δi2 ). Since m · I ≤ A ≤ M · I we know that A − αi I ≥ 0,
βi I − A ≥ 0 and (A − γi I)2 + δi2 I ≥ 0. Since all these operators are pairwise commutative, we know from
lemma 3.5 that their products are also non-negative.
Corollary 3.6. If m · I ≤ A ≤ M · I and P1 (t), P2 (t) are real polynomials with P1 (t) ≤ P2 (t) for all
t ∈ [m, M ], then P1 (A) ≤ P2 (A), which follows from the main proposition, since we can write P21 (t) =
P2 (t) − P1 (t) and since P21 (z) ≥ 0 for all z ∈ [m, M ]. Then P21 (A) = P2 (A) − P1 (A) ≥ 0.

4
5 The Spectral Family March 30, 2006

4 Projection operators
Let E be a linear space. A linear operator P : E 7→ E is called a projection if and only if P 2 = P . We
define E1 = ImP and E2 = ker P . We will now look at some of the properties of such operators:
Property 4.1. For all x ∈ E1 we have P x = x. Indeed if x ∈ E1 there exists a y ∈ E such that P y = x
and since P 2 = P y = x so P x = x.

Property 4.2. The operator Q = I − P is a projection (Q2 = Q) and ImP = ker Q, ker P = ImQ.
Indeed P (I - P ) = 0 implies Im(I − P ) ⊆ ker P . Furthermore if x ∈ ker P , we have (I - P )x = x - 0 =
x and this means ker P ⊆ Im(I − P ).
T
Property 4.3. Let E1 = ImP and E2 = ker P . Then E1 + E2 = E and E1 E2 = 0. Indeed PE + (I -
P )E = E and if Px = 0 and (I - P )x = 0 then 0 = (I - P )x + Px = x.

Property 4.4. Let T : E 7→ E be a linear operator, E1 + E2 = E and let P be the projection onto E1
parallel to E2 . Then P T = T P if and only if E1 and E2 are invariant subspaces of T . Indeed T E1 =
T P E1 = P T E1 ,→ E1 . Thus T : E1 → E1 . Similarly we get T E2 = T (I − P )E2 = (I − P )T E2 ,→ E2 .
Thus T : E2 → E2 . If T : E1 → E1 we have that P T E1 = T P E1 since T E1 ⊆ E1 and P E1 = E1 .
Similarly if T : E2 → E2 we have (I − P )T E2 = T (I − P )E2 since T E2 ⊆ E2 and (I − P )E2 = E2 .

A projection P is called an orthoprojection if P = P ∗ , i.e., if hP x, yi = hx, P yi for all x and y in H.

Here are two properties of such projections.
Property 4.5. If P is an orthoprojection, then ImP ⊥ ker P . For all x and y in H, hP x, (I − P )xi =
hx, (P − P 2 )yi = 0 and this means that ImP ⊥ Im(I − P ).
Property 4.6. Every orthoprojection P satisfies 0 ≤ P ≤ I. In fact hP x, xi = hP 2 x, xi = kP xk2 ≥ 0.
We now use properties 4.3 and 4.5 and get kxk2 = kP xk2 + k(I − P )xk2 and thus we get kP xk2 ≤ kxk2 .
So 0 ≤ P ≤ I.

5 The Spectral Family

Let us consider a bounded self-adjoint operator A satisfying mI ≤ A ≤ MI, with m,M ∈ R and m ≤ M .
Let’s take a,b ∈ R with a < m and b > M and let’s K [a,b] be the set of all piecewise continuous and
bounded functions defined on [a,b] which are limits of monotone non-increasing sequences of continuous
functions defined on the same interval. This means that if f ∈ K [a,b] then there exists a sequence of
functions fn ∈ C[a,b] such that limn→∞ fn (t) = f (t), for all t ∈ [a,b], and also fn (t) ≥ fn+1 (t) ≥ f (t), for
all t ∈ [a,b] and for all n ∈ N. For all this kind of sequences we will write for short fn & f.
The following lemma will allow us to define an operator ϕ(A) for all ϕ ∈ K and for all A of the kind set
at the beginning

Lemma 5.1. Let ϕ(t) ∈ K [a,b]. Then there exists a sequence Pn (t) of polynomials such that Pn (t) &
ϕ(t) ∀t ∈[a,b].

Proof. By definition of K [a,b] there exist ϕn (t) ∈ C[a,b] such that ϕn & ϕ. Let’s consider the continuous
functions ϕn (t)+3/2n+2 and let’s apply the Weierstraß theorem: We have that for all n there exists a
polynomial P n (t) such that
  3  1
Pn (t) − ϕn (t) + n+2  ≤ n+2 .

2 2
Then, clearly, P n+1 (t) ≤ ϕn+1 (t) + 1/2n+1 ≤ ϕn (t) + 1/2n+1 ≤ P n (t). So P n (t) is a non-increasing
sequence of polynomials and it converges to ϕ(t) ∀t ∈ [a,b] since ϕn (t) does.

We are now ready to define the operator ϕ(A):

5
5 The Spectral Family March 30, 2006

Definition of ϕ(A) 5.2. Let ϕ(t) ∈ K [m,M ] and let Pn (t) & ϕ(t) ∀t ∈ [m,M ]. Then the non-increasing
sequence Pn+1 (A) ≥ Pn (A) ≥ ... is bounded. So the strong limit of Pn (A) for n → ∞ exists and we call
it ϕ(A).

We must clarify some fact here:

1. The sequence of polynomials is bounded since ϕ(t) ≥ -C (since ϕ(t) is bounded) and so P n (A) ≥
-C ·I for all n by 3.6.

2. The strong limit ϕ(A) exists by 3.3.

3. The strong limit ϕ(A) depends only on the operator A and on the function ϕ. In particular it
doesn’t depend on the sequence of polynomials P n (t) chosen to approximate ϕ(t). To prove this
we will use the following lemma.
Lemma 5.3. Let Qn (t) and Pn (t) sequences of polynomials and let ϕ(t) and ψ(t) ∈ K [m,M ] functions
in K with ψ(t) ≤ ϕ(t) ∀t ∈ [m,M ]. Let’s assume that Pn (t) & ϕ(t) and Qn (t) & ψ(t) ∀t ∈ [m,M ]. Then
lim Qn (A) := B1 = ψ(A) ≤ ϕ(A) = B2 := lim Pn (A).
n→∞ n→∞

It follows immediately from lemma 5.3 that if ϕ(t) = ψ(t) ∀t ∈ [m,M ] then B1 ≤ B2 and also B2 ≤ B1 ,
so B1 = B2 = ϕ(A) (from a basic property of the ordering of self-adjoint operators). It’s in this way clear
that the limit operator ϕ(A) doesn’t depend on the sequence of polynomials chosen to approximate ϕ(t).
Let’s now prove lemma 5.3:

Proof. since Qn (t) & ψ(t), we have that for all t ∈ [m,M ] and for all n∈N there exists an N0 (t) such
that for all N ≥ N0 (t), QN (t) < ψ(t) + 1/n. Since ψ(t) + 1/n ≤ ϕ(t) + 1/n ≤ P n (t) we can conclude
that
1
QN (t) < Pn (t) + (5.0.4)
n
(again for all t and for all n). From the continuity of the polynomials this implies in particular that for
every fixed t ∈ [m,M ] there exists an open interval I (t) where ∀t ∈ I (t) the inequality (5.0.4) holds for
N = N0 (t) and so also ∀N > N0 (t). We can build in this way a covering of [m,M ] by open intervals
and since [m,M ] (⊆ R) is compact we can extract a finite subcovering from it: let’s call this subcovering
I (ti )M
i=1 . If we now put N0 = max{N0 (ti )|1 ≤ i ≤ M } we clearly have that, ∀N > N0 , QN (t) < Pn (t)
+ n1 (and again this inequality holds for all t and for all n). This means, by 3.6, that QN (A) ≤ Pn (A)
+ n1 I and since this holds for all n we can take the limit N → ∞ obtaining B1 ≤ Pn (A) + n1 I. Now we
take the limit n → ∞ and we finally obtain B1 ≤ B2 .

Putting together the results of lemma 5.1 and lemma 5.3 we have the following:

Lemma 5.4. Let ϕ(t) ∈ K [m,M ] and let ϕn (t) ∈ C[m,M ]. Assume that ϕn (t) & ϕ(t), then ϕn (A)
→ ϕ(A) strongly.

Proof. first we note that ϕn (t) ∈ K [m,M ] for all n. In fact, for all n, ϕn is continuous (and so piecewise
continuous), bounded (since [m,M ]) is compact) and strong limit of a monotone non-increasing sequence
of continuous functions (it’s enough to consider, for example, the sequence of functions that for each fixed
n is defined by ψN (t) = ϕn (t) + 1/N ). Now we can apply lemma 5.1 to each ϕn finding, for all n, a
(n)
sequence of polynomials QN (t) & ϕn (t). By lemma 5.3 follows that since ϕn (t) ≥ ϕn+1 (t) then
(n) (n+1)
lim QN (A) = ϕn (A) ≥ ϕn+1 (A) = lim QN (A).
N →∞ N →∞

Hence ϕn (A) is a non-increasing sequence of self-adjoint bounded operators (because each ϕn (A) is strong
limit of a sequence of polynomials of A - see also property (1) immediately below the proof) and it’s
bounded (because ϕ(t) is): by theorem 3.3 we have the statement.

6
5 The Spectral Family March 30, 2006

Let’s study now some basic properties of the operator ϕ(A):

(1) ϕ(A) is self-adjoint (ϕ∈K ). This follows from the fact that for every real polynomial Pn (t) and for
every self-adjoint operator A the operator Pn (A) is also clearly self-adjoint. Since ∀ϕ∈K ϕ(A) is strong
limit of Pn (A), the continuity of the inner product guarantees that ϕ(A) is self-adjoint.
(i)
(2) (ϕ1 + ϕ2)(A) = ϕ1 (A)  + ϕ2 (A) (ϕi ∈K for i = 1, 2). In fact, taken Pn (t) & ϕi (t) for i = 1, 2, we
(1) (2) (1) (2)
have clearly Pn + Pn (t) ≡ Pn (t) + Pn (t) := (ϕ1 + ϕ2 )(t) & ϕ1 (t) + ϕ2 (t) (so the same limits
(i)
hold with the operator A) and we know from lemma 5.3 that the choice of the polynomials Pn doesn’t
influence the limiting operator ϕ1 + ϕ2 .
(3) (cϕ)(A) = cϕ(A) (for all c > 0 and for all ϕ∈K ). The proof is the same as in (2).
(4) (ϕ1 ϕ2 )(A) = ϕ1 (A) ϕ2 (A) (ϕi ∈K and ϕi ≥ 0 for i = 1, 2; in case ϕ1 or ϕ2 is not ≥ 0 the product
operator ϕ1 ϕ2 may not belong to K ). The proof is again the same as in (2).
(5) ϕ1 (t) ≥ ϕ2 (t) implies ϕ1 (A) ≥ ϕ2 (A) (ϕi ∈K for i = 1, 2). We proved this in the previous lemma
5.3.
We want now to extend the correspondence between the class K and the space L(H) (of bounded opera-
tors on the Hilbert space H) stated by definition 5.2 and lemma 5.1 : to this goal we introduce the class
of functions K of the form ψ = f - ϕ with f, ϕ ∈ K. First of all we define ψ(A) = (f - ϕ)(A) ≡ f (A) - ϕ(A)
and we check that the operator ψ is well-defined (this meaning that it doesn’t depend on the particular
choice of the difference f - ϕ). Let’s suppose that ψ = f1 - ϕ1 and ψ = f2 - ϕ2 (fi , ϕi ∈K for i = 1, 2):
then clearly f1 + ϕ2 = f2 + ϕ1 and so (by property (2) of the operator ϕ(A)) f1 (A) + ϕ2 (A) = f2 (A)
+ ϕ1 (A). Hence f1 (A) - ϕ1 (A) = f2 (A) - ϕ2 (A) and the operator ψ is well defined. The operators ψ(A)
have the same properties of the operators ϕ(A) with two important differences: properties (3) and (4)
above are valid, when ψ ∈ K, for every constant c (and not only c > 0) and ∀ψi ∈ K (and not only ψi ≥
0): in fact, taken constants ci ∈ R and functions ψi ∈ K (i = 1, 2) such that ψ1 + c1 ≥ 0 and ψ2 + c2 ≥
0, we can define ψ1 ψ2 ≡ (ψ1 + c1 )(ψ2 + c2 ) - c1 ψ2 - c2 ψ1 - c1 c2 and we see that the function ψ1 ψ2 ∈ K.
So the operator (ψ1 ψ2 )(A) is defined and it’s equal to ψ1 (A)ψ2 (A).
In this way we have constructed a correspondence between the class of functions K and the space of
bounded self-adjoint operators. This correspondence turns out to be a functional calculus for a bounded
self-adjoint operator A such that mI ≤ A ≤ MI. In fact, K is a unital algebra of functions (the unit being
the constant function ψ(t) = 1) defined on the interval [a,b] ⊇ σ(A) and containing (clearly) the identical
function Id : t 7→ t, L(H) is also a unital algebra (this is a general fact) and the correspondence Θ:
K 7→ L(H) is a unital homomorphism between algebras such that Θ(Id ) = A. We leave to the reader the
detailed proof of the above statements. This functional calculus allows us to define functions of bounded
self-adjoint operators and it’s very important in the study of spectral theory: in this way it should be
clear why we have expanded the class of functions K to the class K.

With the apparatus we have developed we can now derive the spectral decomposition of an (arbitrary)
bounded self-adjoint operator A and define in particular what is a spectral family.
Let’s take the following function, defined ∀λ ∈ [a,b]:

 1 for t ≤ λ
eλ (t) =
0 for t > λ


Clearly this function is in K [a,b], so we are allowed to define the operator eλ (A) = Eλ . The family
of operators {Eλ }λ∈R now defined is an example of a spectral family we were searching for and it’s of
fundamental importance for the Hilbert theorem on the spectral decomposition of a bounded self-adjoint
operator. This family has 3 basic properties that makes it a so-called spectral family:
(i) Eλ are orthoprojections for all λ. In fact Eλ2 = Eλ for all λ, since eλ (t)·eλ (t) = eλ (t) (and then
the result follows from property (5) above), and moreover Eλ = (Eλ )∗ for all λ (as stated before this is
a general fact that holds every time we have a self-adjoint bounded operator A and a function ϕ ∈ K ).
But these are exactly the properties that define an orthoprojections (see section 4). We have also the
following:
(a)Eλ = 0 for a < λ < m
(b)Eλ = I for M ≤ λ < b
These two properties follow from the fact that mI ≤ A ≤ MI.
(ii) Eλ ·Eµ = Eλ for λ < µ. This follows from the fact that obviously eλ (t)·eµ (t) = eλ (t) for λ < µ and
then from property (5) above. This property it’s clearly equivalent to say that Eλ ≤ Eµ for λ < µ.

7
5 The Spectral Family March 30, 2006

(iii) Eλ is continuous from the right with respect to λ in the strong sense. The proof of this
fact is the following: let’s suppose that ϕn (t) is a sequence of continuous functions such that ϕn (t) &
eλ (t) ∀t∈[a,b] and such that ϕn (t) ≥ eλ+ n1 (t) for all n and for all t. The existence of such a sequence is
ensured by the fact that eλ (t)∈K [a,b] and by the fact that clearly we can build our sequence choosing
in particular ϕ1 (t) ≥ 1 ∀t ∈ [λ,λ+1]
h (fori this it’s enough define ϕ1 in this way)h and then,
i for all n ≥
1
1, 1 ≤ ϕn+1 (t) ≤ ϕn (t) ∀t ∈ λ,λ+ n+1 ; clearly for all t outside the interval λ,λ+ n1 we choose the
sequence in such a way that ϕn (t) ≥ ϕn+1 (t) ≥ eλ (t) for all n.
Now, by lemma 5.4 and property (ii) we can conclude that ϕn (A) ≥ Eλ+ n1 ≥ Eλ+αn ≥ Eλ for every αn
such that 0 ≤ αn ≤ 1/n. Since for n → ∞, ϕn (A) & Eλ and αn → 0, it’s clear that Eλ+αn → Eλ for
αn → 0, that is what we wanted to prove.

In general, an arbitrary family of operators {Eλ }λ∈R that fulfil the properties (i)-(iii) is
called a spectral family of operators. We can note that once we have a bounded self-adjoint oper-
ator A such that mI ≤ A ≤ MI there exists such a family: By all our previous discussion it should be
clear that it’s enough to consider the function eλ (t) as defined before to build up such a family.

The spectral family has an additional important property:

(iv) Eλ ·A = A·Eλ . This follows from the fact that by definition Eλ = eλ (A) is the strong limit of a
sequence of polynomials of A (so clearly each polynomial commutes with A and by continuity of the
product of operators also their strong limit does). In particular this means, by property 4.4, that ImEλ
= Hλ is an invariant subspace (of the Hilbert space H) with respect to the operator A.

Now that we have defined a spectral family, the first central ingredient of the Hilbert theorem, we can
move on to the description of the second corner-stone of the theorem, the construction of the so-called
spectral integral.
For this construction we need a very important inequality:
Let λ1 < λ2 (λi ∈ R for i = 1, 2). Then
     
λ1 · eλ2 (t) − eλ1 (t) ≤ t · eλ2 (t) − eλ1 (t) ≤ λ2 · eλ2 (t) − eλ1 (t) .

The inequality holds because eλ2 (t) - eλ1 (t) = 0 for t < λ1 and for t > λ2 , while for λ1 ≤ t ≤ λ2 we have
eλ2 (t) - eλ1 (t) = 1: it’s then easy to see that for t < λ1 and for t > λ2 the inequality is simply stating 0
≤ 0, while for λ1 ≤ t ≤ λ2 the inequality states that λ1 ≤ t ≤ λ2 .
Now, using property (5), we can insert the operator A in place of t in the above inequality obtaining the
fundamental      
λ1 · Eλ2 − Eλ1 ≤ A · Eλ2 − Eλ1 ≤ λ2 · Eλ2 − Eλ1 (5.0.5)

(we recall that Eλ2 = eλ2 (A) and Eλ1 = eλ1 (A)).
This is the main inequality we will soon use to construct the spectral integral.
Let us now investigate some particular cases of special interest and some properties of the main inequality:
1) clearly if we take λ1 < m and λ2 = λ then, by property (i) of the spectral family, we have AEλ ≤ λEλ .
2) if we take λ1 = λ and λ2 >M then,  again by property (i) of the spectral family, we obtain λ(I - Eλ )
≤ A(I - Eλ ) i.e. (A - λI )Eλ ≤ A - λI .
3) for this property we first observe that Eλ1 λ2 ≡ Eλ2 - Eλ1 is an orthoprojection. For this we must verify
that (Eλ1 λ2 )2 = Eλ1 λ2 and that (Eλ1 λ2 )∗ = Eλ1 λ2 . We have (Eλ1 λ2 )2 = (Eλ2 )2 + (Eλ2 )2 - 2Eλ1 ·Eλ2 =
Eλ2 + Eλ1 - 2Eλ1 = Eλ2 - Eλ1 . In the first equality we have used the fact that Eλ1 and Eλ2 commute
(by property (ii) of the spectral family); in the second equality we have used the fact that Eλ1 and Eλ2
are orthoprojections, within the fact that, since λ1 < λ2 , Eλ1 ·Eλ2 = Eλ1 (again by property (ii)). Then
we have (Eλ1 λ2 )∗ = (Eλ2 - Eλ1 )∗ = (Eλ2 )∗ - (Eλ1 )∗ = Eλ2 - Eλ1 . In the last equality we have used the
fact that Eλ1 and Eλ2 are orthoprojections.
Now that we have proved that Eλ1 λ2 is an orthoprojection, we can note that clearly it commutes with A
(because Eλ1 and Eλ2 does) and this two facts together ensure us that the subspace ImEλ1 λ2 = Hλ1 λ2
is invariant with respect to the operator A (by 4.4). Moreover, recalling that Eλ1 λ2 |Hλ1 λ2 = I |Hλ1 λ2 (by
4.1), it follows directly from the main inequality that λ1 I |Hλ1 λ2 ≤ A|Hλ1 λ2 ≤ λ2 I |Hλ1 λ2 ∀x ∈ Hλ1 λ2 .
Therefore, setting ε = λ2 − λ1 , we have that -εI |Hλ1 λ2 ≤ (A - λI )|Hλ1 λ2 ≤ εI |Hλ1 λ2 , ∀λ ∈ [λ1 ,λ2 ]. Now

8
6 The spectral integral March 30, 2006

it follows immediately from 3.2 (i) that

kA - λI kHλ1 λ2 ≤ ε ∀λ ∈ [λ1 , λ2 ], ∀ε.

We have in this way arrived at a very important relation telling us that on the subspace Hλ1 λ2 our
bounded self-adjoint operator A can get infinitely close to a constant operator λI. It’s important to note
that we don’t know a priori if Hλ1 λ2 is the trivial subspace 0. If it isn’t, then the above relation tells
us also that λ is ”almost” an eigenvalue of A (so λ ∈ σp (A)) and every x 6=0 ∈ Hλ1 λ2 is ”almost” an
eigenvector of A.
The construction of the spectral integral will enable us to improve this first approximation of a bounded
self-adjoint operator A.

6 The spectral integral

In this section we are considering a self-adjoint bounded operator A = A∗ with mI ≤ A ≤ M I and an
interval ]a, b[ containing [m, M ]. Consider a partition, a < λ0 < m ≤ λ1 < λ2 < ... < λn−1 ≤ M < λn < b
of the open interval ]a, b[. Assume that the norm of the partition (i.e. the number maxi |λi+1 − λi |) is
less than ε > 0. Take some µi ∈ [λi , λi+1 ] for all i = 0, 1, ..., n − 1. Using the main operator inequality
5.0.5 for all 0 ≤ i ≤ n − 1

λi (Eλi+1 − Eλi ) ≤ A(Eλi+1 − Eλi ) ≤ λi+1 (Eλi+1 − Eλi )

and adding all the terms we get

n−1
X  n−1
X  n−1
X
λk (Eλk+1 − Eλk ) ≤ A (Eλk+1 − Eλk ) = A ≤ λk+1 (Eλk+1 − Eλk )
k=0 k=0 k=0

Pn−1
where we have used that k=0 (Eλk+1 − Eλk ) = Eλn − Eλ0 = I.
Pn−1
Now consider the operator k=0 µk (Eλk+1 − Eλk ). By choice −ε ≤ λk − µk and λk+1 − µk ≤ ε, this gives
n−1
X
n−1
X
−εI = −ε (Eλk+1 − Eλk ) ≤ (λk − µk )(Eλk+1 − Eλk )
k=0 k=0
n−1
X n−1
X
≤A− µk (Eλk+1 − Eλk ) ≤ (λk+1 − µk )(Eλk+1 − Eλk ) ≤ εI
k=0 k=0
(6.0.6)
Pn−1
By the property 3.2 (i) we therefore see that ||A− k=0 µk (Eλk+1 −Eλk )|| ≤ ε for any partition with norm
Pn−1
less than ε > 0 and any choice of points, µi , in the partition-intervals. So the operator n=0 µk (Eλk+1 −
Eλk ) converges towards A as ε → 0 (in the operator norm). We write the limit of the sum as an integral,
and we then have Z b
A= λdEλ (6.0.7)
a
This is a meaningful notation since the convergence of the sum only depends on the size of the partition
norm. Note that a is any number so that a < m and b is any number so that b > M . Because of
R M +δ
the continuity from the right noted in (iii) in the previous section, we see that A = a λdEλ , δ > 0,
RM RM
converges (in the strong sense) towards a λdEλ . Therefore A = a λdEλ . From the property (i) in
R∞
the previous section we may also (formally) write A as A = −∞ λdEλ .
If A is a self-adjoint operator on a finite vector space we know that A can be diagonalized, i.e. it can
be decomposed in ”smaller” operators that act just as multiplication by a real number. Therefore we see
that the spectral integral formula for A (6.0.7) looks very much like the finite case. We would therefore
also expect that ||Ax||2 = λ2 dhEλ x, xi, let’s prove it;
R

First of all, how should the integral be understood? By (ii) in the previous section and the definition
of the ordering we see that the function λ 7→ hEλ x, xi is monotone increasing for all x. This means

9
6 The spectral integral March 30, 2006

that the integral can be understood as the Riemann-Stiltjies integral and it exists (since λ 7−→ λ2 is
continuous). Let us return to the construction of the spectral integral from before. Consider the partition
and define ∆Eλi = Eλi+1 − Eλi . The ∆Eλi ’s are pairwise orthogonal (meaning ∆Eλi ∆Eλj = 0 for j 6= i)
orthoprojections as the reader
P may easily verify for himself. We saw in the construction of the spectral
integral
P that the operator i µi ∆Eλi is convergent towards A in the operator norm. This implies that
µ
i i ∆E λi x → Ax in H for all x. Collecting all these facts we get
X X X X X
|| µi ∆Eλi x||2 = h µi ∆Eλi x, µj ∆Eλj xi = µi µj h∆Eλi x, ∆Eλj xi = µ2i h∆Eλi x, xi → ||Ax||2
i i j ij i
(6.0.8)
which proves the claim.
We collect all the above results in the following theorem
Theorem (Hilbert theorem) 6.1. For every self-adjoint bounded operator A = A∗ on a Hilbert space
H such that mI ≤ A ≤ M I, there exists a unique spectral family {Eλ }λ∈R such that
R∞
i) A = −∞
λdEλ with the integral convergent in the operator norm.
ii) For all λ ∈ R the operator Eλ commutes with any operator that commutes with A.
iii) ||Ax||2 = λ2 dhEλ x, xi (understood as the Riemann-Stiljies integral ).
R

We will not prove the uniqueness of the spectral family since it requires knowledge of measure theory.
With the construction of the spectral family and the spectral integral it becomes more transparent what
happens when a (continuous) function is applied to an operator as the following theorem tells us
Theorem (Expanded spectral integral) 6.2. Let A = A∗ be a bounded self-adjoint operator (with
RM
mI ≤ A ≤ M I) on a HilbertP space H. Let ϕ ∈ C([a, M ]) with a < m. Define the integral m−0 ϕ(λ)dEλ
as the limit of the sum i ϕ(µi )∆Eλi as the norm of the partition goes to zero. The sum converges
RM
towards ϕ(A) in the operator norm; ϕ(A) = m−0 ϕ(λ)dEλ and hence ||ϕ(A)||2 = ϕ2 (λ)hEλ x, xi (the
R

last statement follows frow the proof before).

Proof. First assume that ϕ(t) = tk , k ∈ N. Then

X  X  X Z M
Ak ← µi1 ∆Eλi1 · ... · µik ∆Eλik = µki ∆Eλi → λk dEλ
i1 ik i c

as the partition norm goes to zero, where c is just a number such that a < c < m. From this we
RM
conclude that for any polynomial, P (t), we have that P (A) = c P (λ)dEλ . Let Pn be a sequence so
that Pn & ϕ and for a given ε let P be a polynomial in the sequence so that |ϕ(t) − P (t)| ≤ ε/3
for all t ∈ [a, b]. By the fundamental theorem 3.4 and the definition of ϕ(A) we see that this implies
−ε/3 I ≤ ϕ(A) − P (A) ≤ ε/3 I and hence ||ϕ(A) − P (A)|| ≤ ε/3 I. We know for a partition {λi } of [a, b]
with some norm less than ε that
X
||P (A) − P (µi )∆Eλi || ≤ ε/3.
i
P P
P the operator T = i P (µi )∆Eλi − i ϕ(µi )∆Eλi . Since −ε/3 ≤ P (µi ) − ϕ(µi ) ≤ ε/3 for all
Now define
i and i ∆Eλi = I we have that −ε/3 I ≤ T ≤ ε/3 I and therefore ||T || ≤ ε/3.
We now see that
X    X  X X 
||ϕ(A)− ϕ(µi )∆Eλi || = || ϕ(A)−P (A) + P (A)− P (µi )∆Eλi + P (µi )∆Eλi − ϕ(µi )∆Eλi || ≤ ε
i i i i

which proves the statement.

Example 1. Consider the Hilbert space L[0,1]. Let A be the linear operator Ax(t) = tx(t). We see
that A is self-adjoint since for all x(t), y(t) ∈ L[0,1] we have
Z 1 Z 1 Z 1 Z 1
hAx(t), y(t)i = Ax(t)y(t)dt = tx(t)y(t)dt = x(t)ty(t)dt = x(t)Ay(t)dt = hx(t), Ay(t)i
0 0 0 0

10
References March 30, 2006

and it is bounded since

Z 1 
sup |hAx, xi| =  t|x(t)|2 dt ≤ 1.
 
kxk=1 0

We can find the spectral family corresponding to A by first noting that Ak x(t) = tk x(t) where k is a
positive integer and hence for any polynomial P we have P (A)x(t) = P (t)x(t). Then by definition 5.2
we see that for any φ ∈ K[0, 1] we have φ(A)x(t) = φ(t)x(t) and in particular Eλ x(t) = eλ (t)x(t). To
check agreement with the Hilbert theorem we can for example verify property iii) explicitly. Since A is a
positive operator and has norm less than one we can choose the integration limit in the spetral integral
to be [0,1]. Then we get
Z 1 Z 1 Z λ  Z 1 Z 1
λ2 dhEλ x(t), x(t)i = λ2 d |x(t)|2 = λ2 |x(λ)|2 dλ = |Ax(λ)|2 dλ = kAx(t)k2
0 0 0 0 0

which agrees with the Hilbert theorem.

Example 2. If we have a positive, self-adjoint, bounded operator A, then it is enough to integrate
over the positive real numbers in the extended spectral integral. Then it makes sense to apply a square
root function to A, i.e.
Z ∞ √
SQRT (A) = λdEλ .
0

This application of the expanded spectral integral gives an alternative definition of a square root to
the one mentioned in Lemma 3.5. However it would be disappointing if these two did not agree. Let’s
conclude by showing that indeed they are the same. We have

X√  X √  X√ Xq
SQRT (A)SQRT (A) ← µi ∆Eλi µj ∆Eλj = µi µj ∆Eλi ∆Eλj = µ2i ∆Eλi → A.
i j i,j i

√ √
By the uniqueness of A we conclude that SQRT (A) = A.

References
[1] Eidelman, Y., Milman, V. and Tsolomitis, A. Functional Analysis, An Introduction. American Math-
ematical Society, 2004.

1, 2, 3, 3.5

11