Spectral Theory for Unbounded Operators

October 1, 2008

Throughout H, K denote Hilbert spaces and A, B denote operators on H, K.

1 Unbounded operators
Definition 1. An unbounded operator acting on H is a pair (A, D(A)) where
D(A) ⊂ H is a linear subspace and A : D(A) → H is linear. D(A) is called
the domain of A.

A is said to be densely defined if D(A) is dense. We consider mostly densely

defined operators. By the Graph of A we mean the following subset of H ⊕H:

G(A) := {(φ, Aφ) : π ∈ D(A)}

Definition 2. We say B (an unbounded operator) to be an extension of

A, written A ⊂ B, if G(A) ⊂ G(B) or equivalently D(A) ⊂ D(B) and
B|D(A) = A.

Definition 3. A is said to be closed if G(A) is closed. A is said to be closable

if A has a closed extension.

Proposition 1. Let A be closable. Then there exists a unique Ā such that

¯
G(Ā) = (G(A)).

Proof. Uniqueness is clear. Let π1 : H ⊕ H → H be the first projection

and Let π2 be the second projection. Let D(¯(A)) = π1 (G(A)).
¯ Let B be a
¯ ¯
closed extension of A. Hence G(A) ⊂ G(B). Hence π1 : G(A) → D(Ā) is a
¯
bijection. Let Ā = π2 ◦ π1− . Then, clearly G(Ā) = G(A). This completes the

1
proof. .

Remark. It is clear that if A is closable then A has a smallest closed ex-

tension called the closure of A which is what is constructed in the above
Proposition.

Definition 4. Let A be a densely defined operator acting on H. Let

D(A∗ ) := {φ : ψ → hAψ, φi is bounded}

Clearly D(A∗ ) is a linear subspace of H. Let φ ∈ D(A∗ ). Then ψ →

hAψ, phii admits a continuous extension to the whole of H. Hence ∃ a unique
A∗ φ such that hAψ, φi = hψ, A∗ φi. Note that φ → A∗ φ is linear. Note the
equation hAψ, φi = hψ, A∗ φi for ψ ∈ D(A), φ ∈ D(A∗ ). (A∗ , D(A∗ )) is called
the adjoint of A∗ .

Proposition 2. Let A be densely defined. Then

(1) A∗ is closed.

(2) if A is closed then D(A∗ ) is dense.

(3) If A is closed then A = A∗∗ .

Proof. Let U : H ⊕ H → H ⊕ H be defned by U (φ, ψ) = (−ψ, φ). Then U

is unitary.

Assertion: U (G(A)⊥ ) = G(A∗ ).

Let φ ∈ D(A∗ ) be given. Then

h(φ, A∗ φ), (−Aψ, ψ)i = −hφ, Aψi + hA∗ φ, ψi

= −hφ, Aψi + hφ, Aψ
=0

0
Hence G(A∗ ) ⊂ U (G(A)⊥ ). Now Let (φ, φ ) ∈ U (G(A)⊥ ) be given. Then
0 0
hAψ, φi = hψ, φ i. Hence (φ, φ ) ∈ G(A∗ ). Hence U (G(A)⊥ ⊂ G(A∗ ). This
proves the assertion. This proves (1).

2
Let A be closed. Let ξ ∈ D(A∗ )⊥ be given. Then (ξ, 0) ∈ G(A∗ )⊥ =
U (G(A)⊥⊥ ) = U (G(A)). Hence (ξ, 0) = (−Aψ, ψ) for some ψ ∈ D(A).
Hence ξ = 0. This proves (2).

Now Let A be closed. Then G(A∗∗ ) = U (G(A∗)⊥ ) = U ((U (G(A))⊥ )⊥ ) =

G(A). Hence A = A∗∗ . This proves (3). .

Remark. Note that if A is closed and densley defined then A∗ is closed and
densely defined and A = A∗∗ .

Definition 5. Let A be closed and densely defined. A subspace D0 is called

a core for A if A|¯D0 = A. Equivalently G(A|¯ D ) = G(A). In other words, D0
0
is a core for A if given φ ∈ D(A) there exists a sequence (φn ) ∈ D0 such that
φn → φ and Aφn → Aφ.

The following propsition identifies a core for A which we need later.

Proposition 3. Let A be closed and densely defined acting on H. Let

P : G(A) → H denote the first projection. Then the range of P P ∗ is a core
for A.

Proof. First note that P is 1-1. Hence Ker(P ) = 0 implies R(P ∗ ) is

dense in G(A). Let (φ, Aφ) ∈ G(A) be given. Then ∃ (ξn ) such that
P ∗ (ξn ) → (φ, Aφ). Let P ∗ (ξn ) = (ψn , Aψn ). Then ψn → φ and Aψn → Aφ.
But ψn = P P ∗ ξn ∈ R(P P ∗ ). Hence the range of P P ∗ is a core for A. 

Definition 6. Let A be densely defined. A is said to be symmetric if A ⊂ A∗ .

Equivalently, hAφ, ψi, h= φ, Aψi ∀φ, ψ ∈ D(A).

A is said to be selfadjoint if A = A∗ .

Proposition 4. Let A be a symmetric operator. Then the following are

equivalent:

3
 (1) A is self adjoint.

(2) ker(A∗ ± i) = {0} and A is closed.

(3) Ran(A ± i) = H.

Proof.
(1) =⇒ (2)
A is closed since A = A∗ .

We will prove that ker(A∗ + i) = {0} since the other one is similar.

Let φ ∈ Ker(A∗ + i) = Ker(A + i) be given. Then

h(A + i)φ, φi = 0
⇒ hAφ, φi + ihφ, φi = 0
⇒ hφ, Aφi + ihφ, φi = 0
⇒ ihφ, φi + ihφ, φi = 0

Hence hφ, φi = 0. Hence Ker(A∗ + i) = 0.

(2) =⇒ (3)
First We show that Ran(A+i) is dense. Let ξ ∈ Ran(A+i)⊥ be given. Then
for φ ∈ D(A), h(A + i)φ, ξi = 0. Hence hAφ, ξi = −ihφ, ξi. Hence ξ ∈ D(A∗ )
and A∗ ξ = iξ, which is a contradiction. Hence Ran(A + i) is dense.

Let φ ∈ H be given. Then ∃(φn ) ∈ D(A) such that (A + i)φn converges to φ.

Now h(A + i)(φn − φm ), (A + i)(φn − φm )i = ||A(φn − φm )||2 + ||φn − φm ||2 .
Hence (φn ) and (Aφn ) are cauchy. Since G(A) is closed, it follows that
∃ ψ ∈ D(A) such that φn → ψ and Aφn → ψ. Hence φ = (A + i)ψ. Hence
Ran(A + i) = H. The other one is similar.

(3) =⇒ (1)
Note that (Ker(A∗ + i)) ⊂ (Ran(A − i))⊥ . Hence ker(A∗ + i) = 0. Let
φ ∈ D(A∗ ) be given. Then there exists ψ ∈ D(A) such that (A∗ + i)(φ) =
(A + i)(ψ). Since A ⊂ A∗ , it follows that (A∗ + i)(φ − ψ) = 0. Hence φ = ψ.
Hence D(A∗ ) ⊂ D(A). Hence D(A) = D(A∗ ). 

4
2 Cayley Transform
In this section, we describe von Neumannn’s trick of passing from a selfad-
joint operator to a unitary operator via the cayley transform. We will use
this trick to deduce the spectral theory for selfadjoint opertors.

Lemma 1. Let A be an unbounded operator acting on H. Let A be

closed and densely defined. Suppose A : D(A) → H is a bijection. Then
A− : H → D(A) is a bounded linear operator.

Proof. Use closed graph theorem. 

Proposition 5. Let A be a selfadjoint operator. Let U = (A + i)(A − i)−1 .

Then
(i) U is a unitary operator.

(ii) 1 is not an eigenvalue of U .

Proof. First note that U is one one and onto by Proposition 4. Hence it is
enough to show that U is an isometry. Since Ran(A − i) = H, it is enough to
show that hU (A − i)ξ, U (A − i)ξi = h(A − i)ξ, (A − i)ξi for ξ ∈ D(A) which
is a simple verification. Hence U is unitary.

Suppose 1 were an eigenvalue of U . Then there exists ξ 6= 0 ∈ D(A) such

that U (A − i)ξ = (A − i)ξ. Hence (A + i)ξ = (A − i)ξ, which is a contradic-
tion. 

3 Spectral Integration
In this section, we will show how to integrate functions with respect to a pro-
jection valued measure to get operators. Fix a Hilbert space H throughout.
Let (Ω, B) be a measurable space.

Definition 7. A H valued spectral measure on is a function P : B → B(H)

such the following holds.

5
 1. P (φ) = 0 and P (Ω) = 1.

2. P (E) is a projection for every E ∈ B.

`
3. P (E F ) = P (E) + P (F ).

4. If En ↑ E then P (En ) ↑ P (E) strongly.

We start with the basic proposition which lists the properties of the spectral
measure.

Proposition 6. Let P be a spectral measure on (Ω, B). Then

1. If E ⊂ F then P (E) ≤ P (F ).

2. P (E ∩ F ) = P (E)P (F ).

Proof. First observe that (2) of definition implies P (E) ⊥ P (F ) f E ∩ F = φ.

(i)

E ⊂ F ⇒ P (F ) = P (E) + P (F \E)
⇒ P (F )P (E) = P (E)
⇒ P (E) ≤ P (F )

(ii) P (E) = P (E ∩ F ) + P (E\F ). Hence P (E)P (F ) = P (E ∩ F ). This

completes the proof. .

Definition 8. Let P be a spectral measure on (Ω, B). For ξ, η ∈ H, define

dPξ,η (E) = hP (E)ξ, ηi. Then dPξ,η is a finite measure.

The next proposition estimates the total variation of dPξ,η .

Proposition 7. |dPξ,η | ≤ ||ξ, η||.

6
 `
Proof. Let Ω := n En . Now
X X
|dPξ,η (En )| = |hP (En )ξ, ηi|
n n
1 1
X
≤ |hP (En )ξ, ξi| 2 |hP (En )η, ηi| 2
n
X
≤ ||P (En ξ)||||P (En )η||
n
1 1
X X
≤( ||P (En )ξ||2 ) 2 ( ||P (En )η||2 ) 2
n n
≤ ||ξ||||η||

Hence |dPξ,η | ≤ ||ξ||||η||. .

Let B(Ω) denote the bounded complex valued measurable R functions. Let
f ∈ B(Ω). Consider the sesquilinear form (ξ, η) → f dPξ,η .R Then the
sesquilinear
R form is bounded.
R Hence there Rexists an operator f dP such
that h( f dP )ξ, ηi = f dPξ,η . Clearly f → f dP is linear. But it is more.
It is a ∗ algebra homomorphism.

Theorem 1. Let P be a spectral measure on (Ω, B). Then there exists a

unique π : B(Ω) → B(H) such that
(i) π is a ∗ algebra homomorphsm which is unital.
(ii) π(1E ) = P (E).
(iii) π respects dominated convergence. That is if fn ∈ B(Ω) and fn → f
pointwise and sup|fn | < ∞ then π(fn ) → π(f ) strongly.
R
Proof. Uniqueness is clear. For existence, set π(f ) = f dP . It is clear that
π(1E ) = P (E). Now we prove that π(f¯) = π(f )? .
Now
Z
hπ(f )ξ, ξi = f¯dPξ,ξ
¯
Z ¯
= f dPξ,ξ

= hπ(f¯)ξ, ξi
= hξ, π(f )ξi
= hπ(f )? ξ, ξi

7
Hence π(f¯) = π(f )? .

To prove that π is an algebra homomorphism, we need the following lemma.

First we show that f dPξ,η = dPπ(f )ξ,η for f simple.

Since both sides are linear, it is enough to show that 1E dPξ,η = dPπ(1E )ξ,η .
Now as 1E dPξ,η (F ) = hP (E ∩ F )ξ, ηi = hP (E)P (F )ξ, ηi = dPπ(1E )ξ,η (F ) it
follows that 1E dPξ,η = dPπ(1E )ξ,η .
HenceRf dPξ,η = dPRπ(f )ξ,η for f simple.
Thus gf dPξ,η = gdPπ(f )ξ,η for g ∈ B(Ω) and f simple.
Hence π(gf ) = π(g)π(f ) where g ∈ B(Ω) and f simple. Since π preserves ∗
one has π(f¯ḡ) = π(f¯)π(ḡ) for f simple and g ∈ B(Ω). So π(f g) = π(f )π(g)
for f Rsimple and g ∈
R B(Ω).
Now f gdPξ,η = f dPRπ(g)ξ,η for f Rsimple. Hence gdPξ,η = dPπ(g)ξ,η for
g ∈ B(Ω). Thus again f gdPξ,η = f dPπ(g)ξ,η . Hence π(f g) = π(f )π(g).
This shows that π is an algebra homomorphism.

Now Let us prove the dominated convergence. Let fn ∈ B(Ω) be such that
fn → f pointwise and sup|fn | < ∞. Now, by the usual lebesque dominated
convergence theorem, it follows that π(fn ) converges weakly to π(f ). Since
π is an algebra homomorphism, it follows that π(fn ) converges strongly to
π(f ). .

The following proposition is easy to prove and the proof is left to the reader.
Proposition 8. The following are equivalent. Let f ∈ B(Ω).
R
1. f dP = 0.
2. f 1E dP = 0 for E ∈ B.
R

3. f = 0 a.e.
Let L∞ (Ω, P ) denote the equivalence class under the relation f ∼ g if and
only if f = q a.e.

Suppose
P 1 H is separable. Let ξn be an orthonormal basis for H. Let µ(E) =
n 2n hP (E)ξn , ξn i. Then µ is a finite measure. µ is absolutely continous
w.r.t P i.e P (E) = 0 ⇔ µ(E) = 0. Hence L∞ (Ω, µ) is isomorphic to
L∞ (Ω, P ). Hence L∞R(Ω, P ) is a von Neumann algebra. By the above prop-
sition, the map f → f dP factors through L∞ (Ω, P ) which is one-one. It is
also normal.

8
4 Spectral Theorem for unitary operators
We need the following spectral theorem for unitary operators to deduce the
spectral theorem for selfadjoint operators via the cayley transform. The proof
of the following proposition can be found in [Arv].

Proposition 9. Let U ∈ B(H) be a unitary operator. Then there exists a

unique spectral
R measure on the unit circle (with the Borel σ algebra) such
that U = z dP (z).

5 Spectral integration for unbounded func-

tions
We have already seen that by integrating bounded functions with respect to
a spectral measure we get bounded operators. In this section, we show how
to get unbounded operators by integrating unbounded functions.

Let (Ω, B) be a measurable space. Let P be a H valued spectral measure on

(Ω, B). Let f : Ω →R C be a measurable function.
Let Df = {ξ ∈ H : |f |2 dPξ,ξ < ∞}. We assert that Df is a linear subspace
of H.

Let ξ, η ∈ H. Then,

dPξ+η,ξ+η (E) = hP (E)(ξ + η), P (E)(ξ + η)i

= ||P (E)ξ||2 + 2hP (E)ξ, P (E)ηi + ||P (E)η||2
≤ 2(dPξ,ξ (E) + dPη,η (E))

Hence ξ, η ∈ Df implies ξ + η ∈ Df . It is clear that if ξ ∈ Df and α ∈ C

implies αξ ∈ Df . Hence it follows that Df is a subspace of H. We now claim
that Df is a dense subspace. Let En := {x :| f (x) |≤ n}. For any ξ ∈ H,
π(1En )ξ ∈ Df and π(1En )ξR converges to ξ. This proves that Df is dense.
Now we will make sense of f dP as an unbounded operator defined on Df .
To that end, let sn be a sequence of bounded functions such that | sn |≤| f |
and sn converges pointwise to fR. For ξ ∈ Df , note that π(sn )ξ is convergent.
For (k π(sn )ξ − π(sm )ξ k)2 = | sn − sm |2 dPξ,ξ and sn converges to f in
L2 (dPξ,ξ ) and hence π(sn )ξ is cauchy. Also if tn were another sequence of
bounded functions for which | tn |≤| f | then k π(sn )ξ −π(tn )ξ k=k sn −tn k2

9
where k k2 is the norm in L2 (dPξ,ξ and hence π(tn )ξ converges to the same
limit
R as that of the limit Rof the sequence π(sn )ξ. We define this value as
( f dP )ξ. It is clear that f dP is a linear operator with domain Df .

Proposition 10. Let f : Ω → C be measurable. Then

R R R
1. RFor ξ ∈ Df , k ( f dP )ξ k2 = | f |2 dPξ,ξ and h( f dP )ξ, xii =
f dPξ,ξ .
R
2. Let g R: Ω → C be bounded.
R R For ξ ∈ D f
R , ( R )ξ ∈ Df and ξ ∈ Df g
gdP
and ( f gdP )ξ = ( f dP )( gdP )ξ = ( gdP )( f dP )ξ

The proof is left to the reader as it follows straight from the definitions and
theorem 1

Theorem 2. Let f be a measurable complex valued function on Ω. Then

R
1. The unbounded operator f dP is closed.

2. The adjoint of f dP is f¯dP .

R R

R
3. If f is real then the operator f dP is selfadjoint.
R
Proof. Let us denote the operator f dP by T (f ). Let ξn → ξ and T (f )ξn →
η. Let Ek := {| f |≤ k}. Now by (2) of the proposition 10 one has
hT (f )P (Ek )ξ, T (f )P (Ek )ξi = hP (Ek )T (f )ξn , P (Ek )T (f )ξn i ≤ hT (f )ξn , T (f )ξn i.
Hence supk,n hT (f )P (Ek )ξn , T (f )P (Ek )ξn i is finite which we denote by say
M . Since ξn converges to ξ, oneRhas the inequality hT (f 1EkR)ξ, iT (f 1Ek ξ ≤ M
for every k which implies that | f |2 dPξ,ξ ≤ M . Hence | f |2 dPξ,ξ ≤ M .
This proves that ξ ∈ Df .
Also note that for every k, n, One has T (f )P (Ek )ξn = P (Ek )T (f )ξn . Hence
as n → ∞, one has T (f 1Ek )ξ = P (Ek )η for every k. Now by the definition
of T (f ) as k → ∞, one has T (f )ξ = η. This proves that T (f ) is closed. Now
we will prove that T (f )∗ = T (f¯). It is clear that Df¯ = Df . Let ξ ∈ Df¯. For
η ∈ Df , hT (f )η, ξi = limn→∞ hT (s¯n )η, ξi=limn→∞ hη, T (s¯n ξ)i =hη, T (f¯)ξi.
Hence T (f¯) ⊂ T (f )∗ .
Let η ∈ Dom(T (f )∗ ). Then there exists M > 0 such that for every ξ ∈ Df
one has | hT (f )ξ, ηi |≤ M k ξ k. Let Ek := {| f |≤ k}. Now one has
the
R following inequality | hT (f )P (Ek )T (f¯)P (Ek )η, ηi ≤ M . Thus one has
2
| f | 1Ek dPη,η ≤ M for every k. Hence η ∈ Df = Df¯. This proves that
T (f¯) = T (f )∗ . This completes the proof. 

Lemma
R 2. Let P be a spectral measure on theR (R, BR ). Then the operator
−1 1
A := xdP (x) is selfadjoint. Also (A − i) = x−i dP (x).

10
Proof. The fact that A is selfadjoint will follow from the last theorem. Let
1
g(x) = x−i . Then g is a bounded function. Let f (x) = x − i. Then for ξ ∈
Dom(A), T (g)T (f )ξ = T (f )T (g)ξ = T (f g)ξ = ξ. Hence T (g)(A − i)ξ = ξ
for ξ ∈ Dom(A). Hence T (g) = (A − i)−1 . .

6 Spectral theorem for unbounded operators

We start with a lemma that gives the change of variable formula in the
spectral measure setup. We leave the easy proof to the reader.
0 0 0
Lemma 3. Let (Ω, B) and (Ω , B ) be measurable spaces and let φ : Ω → Ω
0
be measurable. Let P be a spectral measure on (Ω, B) and let P = P φ−1 .
Then
0 0
1. P is a spectral measure on Ω .
0 R 0 R
2. For f : Ω → C measurable, f dP = f ◦ φdP .
Theorem [von Neumann]. Let A be a selfadjoint operator on a Hilbert
spaceRH. Then there exists a unique spectral measure P on (R, BR ) such that
A = xdP (x).
0
Proof. Let U := (A + i)(A − i) be the cayley transform of A. Then U
is a unitary operator on H. Hence, by the spectral theorem for unitary
operators,
R there exists a unique spectral measure Q on T such that A =
zdQ(z)Since 1 is not an eigen value of U , Q(1) = 0. Now the map φ :
R → T − 1 given by φ(t) = t+i t−i
is a bijection and let ψ be it’s inverse.
Then ψ(z)R = if racz +R1z − 1. Let P := Q|T−1 ◦ ψ −1 . Let T denote the
operator xdP (x) = T−1 i( z+1 z−1
)dQ(z). Then T is selfadjoint and (T −
U −1
i) = T 2i dQ(z)= 2i . But since f racU − 12i = (A − i)−1 it follows that
−1
R z−1

(T − i)−1 = (A − i)−1 . Hence T = A. The uniqueness of the spectral measure

P follows from the uniqueness of the spectral measure Q. 
Remark. Let U be a unitary on H and P a spectral measure on (Ω, B). Let
P
R U be definedR by PU (E) = U P (E)U ∗ . Then PU is a spectral measure and
f dPU = U ( f dP )U ∗ for measurable functions.
Definition 9. Let M be a von Neumann algebra on H with the identity
operator being in M . Let A be an unbounded operator. We say A is affliated
0 0
to M written AηM if and only if U AU ∗ = A for every unitary U ∈ M ( M
denotes the commutant of M )
Proposition 11. Let A be a selfadjoint operator on H and Let M be a von
Neumann algebra on H. Then the following are equivalent.

11
 1. The operator A is affliated to M .

2. If P is the spectral measure for A then P (E) ∈ M for every E ∈ BR .

0
xdP (x))U ∗ =
R
Proof.
R First assume
R that AηM . Let
R U be a unitary in M . Then U (
xdP (x). Hence xdPU (x) = xdP (x). By the uniqueness of the spectral
measure it follows that PU = P . Hence for E ∈ BR , PU (E) = P (E). Hence
0
P (E) commutes with every unitary in M . By commutant theorem it follows
that P (E) ∈ M .
0
Now assume R(2). Let U ∈ M be a unitary. Then dPU = dP . Hence
xdPU (x) = xdP (x). Thus U AU ∗ = A. Hence A is affliated to M .
R


Definition 10. Let A be a selfadjoint operator on H. Denote the smallest

vonNeumann algebra which is affliated to A by W ∗ {A}.

Proposition 12. Let P be the spectral measure on (R, BR ) for the selfadjoint
operator A. Then the image of L∞ (R, P ) under the map f → f dP is
R

W ∗ {A}.

12