06.08.2013 Views

12. The spectral theorem - Borel calculus for bounded normal ...

12. The spectral theorem - Borel calculus for bounded normal ...

12. The spectral theorem - Borel calculus for bounded normal ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong><br />

<strong>bounded</strong> <strong>normal</strong> operators<br />

Klaus Thomsen matkt@imf.au.dk<br />

Institut <strong>for</strong> Matematiske Fag<br />

Det Naturvidenskabelige Fakultet<br />

Aarhus Universitet<br />

December 2005<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


We read Sections <strong>12.</strong>23 and <strong>12.</strong>24 in Rudins book - with some<br />

changes.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


<strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - the statement<br />

<strong>The</strong>orem<br />

Let H be a Hilbert space and T ∈ B(H) a <strong>normal</strong> operator.<br />

<strong>The</strong>re is a unique resolution of the identity E on the spectrum<br />

σ(T) of T such that<br />

<br />

<br />

T = λ dE(λ) = idσ(T) dE .<br />

σ(T)<br />

σ(T)<br />

<strong>The</strong> reader should compare this to the following result from linear<br />

algebra:<br />

Let T ∈ Mn(C) be a <strong>normal</strong> matrix. <strong>The</strong>re are then orthogonal<br />

projections E1,E2,... ,En and complex numbers λ1,λ2,... ,λn such<br />

that 1 = E1 + E2 + · · · + En and<br />

n<br />

T = λiEi.<br />

i=1<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


On the notion of spectrum<br />

Lemma<br />

Let D be a unital C ∗ -algebra and E ⊆ D a unital C ∗ -subalgebra.<br />

Let x ∈ E. <strong>The</strong>n x is invertible in E if and only if x is invertible in<br />

D.<br />

Proof.<br />

Assume that x is invertible in D. <strong>The</strong>n x ∗ is invertible in D and<br />

x −1 = x ∗ (xx ∗ ) −1 . We must show that x −1 ∈ E.<br />

Set a = xx ∗ and observe that it suffices to prove that a −1 ∈ E. Set<br />

b = a −1 and observe that a,b and 1 generate an abelian<br />

C ∗ -subalgebra A of D.<br />

By <strong>The</strong>orem 11.18 in Rudin’s book there is a compact Hausdorff<br />

space ∆ and a ∗-isomorphism Φ : A → C(∆) such that Φ(1) = 1.<br />

Set f = Φ(a),g = Φ(b) and note that g = 1<br />

f . We claim that<br />

1<br />

f ∈ C ∗ (1,f ). (1)<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


On the notion of spectrum<br />

Proof.<br />

To prove this, note first that there is a interval [ǫ,N] in the positive<br />

half-axis such that |f (x)| 2 ⊆ [ǫ,N] <strong>for</strong> all x ∈ ∆.<br />

By the classical Weierstrass <strong>theorem</strong> there is a sequence {Pn} of<br />

polynomials such that<br />

<br />

lim<br />

n→∞ sup<br />

t∈[ǫ,N]<br />

<br />

<br />

Pn(t) − 1<br />

<br />

t = 0.<br />

<strong>The</strong>n limn→∞ Pn ◦ |f | 2 = 1<br />

|f | 2 in C(∆). Since<br />

Pn ◦ |f | 2 = Pn ◦ (ff ∗ ) ∈ C ∗ (1,f ) <strong>for</strong> all n, it follows that<br />

1<br />

|f | 2 ∈ C ∗ (1,f ).<br />

Since 1<br />

f<br />

= f 1<br />

|f | 2, we see that (1) holds. It follows from (1) that<br />

b = Φ −1 (g) ∈ A. Since A ⊆ E we conclude that b ∈ E, as<br />

desired.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


On the notion of spectrum<br />

Lemma<br />

Let D be a unital C ∗ -algebra and E ⊆ D a unital C ∗ -subalgebra.<br />

Let x ∈ E. <strong>The</strong>n x has the same spectrum in E as it has in D.<br />

Proof.<br />

By Lemma 2, (λ − x) is invertible in E if and only if it is invertible<br />

in D.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(existence)<br />

Let A be the C ∗ -subalgebra of B(H) generated by T and 1. This is<br />

an abelian C ∗ -subalgebra of B(H) and <strong>The</strong>orem <strong>12.</strong>22 applies to<br />

give us a resolution of the identity E ′ on the <strong>Borel</strong> σ-algebra of the<br />

Gelfand-spectrum ∆ of A such that<br />

<br />

T = Td E.<br />

∆<br />

On the other hand, <strong>The</strong>orem 11.19 (suitably intepretated) gives us<br />

a homeomorphism κ : ∆ → σ(T) defined such that<br />

κ(ω) = ω(T).<br />

Observe that in <strong>The</strong>orem 11.19 the relevant spectrum σ(T) is<br />

taken relative to A - but by Lemma 3 this is the same as the<br />

spectrum relative to B(H).<br />

Now transfer the resolution E ′ to a resolution on σ(T) via κ. More<br />

precisely, when B ⊆ σ(T) is a <strong>Borel</strong> set, κ −1 (B) is a <strong>Borel</strong> set in ∆<br />

and we set<br />

E (B) = E ′ κ −1 (B) .<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(existence)<br />

This is a resolution of the identity: E(∅) = 0 and<br />

E (σ(T)) = E ′ (∆) = 1. And E (B) is obviously a self-adjoint<br />

projection because E ′ κ −1 (B) is.<br />

When x ∈ H we see that<br />

Ex(B) = 〈E(B)x,x〉 = E ′ κ −1 (B) x,x = E ′ x ◦ κ−1 (B),<br />

which is a measure because E ′ x is (!).<br />

<strong>The</strong> next step is to show that<br />

<br />

T =<br />

σ(T)<br />

Let x ∈ H and note that<br />

<br />

λ dE x,x =<br />

σ(T)<br />

<br />

σ(T)<br />

= λ dE ′ x ◦ κ −1 <br />

= κ dE ′ x.<br />

σ(T)<br />

λ dE(λ). (2)<br />

∆<br />

λ Ex<br />

(3)<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(existence)<br />

<strong>The</strong> last equation may not be too well-known so here is an<br />

argument: We claim that in general<br />

<br />

<br />

f dE<br />

σ(T)<br />

′ x ◦ κ −1 =<br />

f ◦ κ dE<br />

∆<br />

′ x<br />

when f ∈ BB (σ(T)).<br />

<strong>The</strong> last identity in (3) follows from (4) by taking f to be the<br />

identity function on σ(T).<br />

To establish (4) note that it suffices to check on simple functions.<br />

Hence it suffices in fact to establish the equality when f = 1B <strong>for</strong><br />

some <strong>Borel</strong> set B ⊆ σ(T). But 1B ◦ κ = 1κ−1 (B) and hence the (4)<br />

reads E ′ <br />

x κ−1 (B) = E ′<br />

x κ−1 (B) in this case - a statement which<br />

is hard to argue against.<br />

(4)<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(existence)<br />

To complete the calculation started in (3) observe that κ = T<br />

because κ(ω) = ω(T) = T(ω) <strong>for</strong> all ω ∈ ∆.<br />

Hence<br />

<br />

<br />

=<br />

∆<br />

λ dE<br />

σ(T)<br />

= 〈Tx,x〉 .<br />

T dE ′ x =<br />

x,x<br />

<br />

∆<br />

=<br />

T dE<br />

Since x ∈ H was arbitrary, this proves (2).<br />

κ dE<br />

∆<br />

′ x<br />

<br />

x,x<br />

(5)<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

To establish the uniqueness part in the <strong>spectral</strong> <strong>theorem</strong> we have to<br />

work a little more than Rudin does because we have not required<br />

the measures arising from a resolution of the identity to be<br />

reagular. <strong>The</strong> next lemma gives us what we need:<br />

Lemma<br />

Let X be a compact metric space, and let µ be a finite <strong>Borel</strong><br />

measure on X. <strong>The</strong>n µ is both inner- and outer regular, i.e<br />

and<br />

<strong>for</strong> every <strong>Borel</strong> set B ⊆ X<br />

µ(B) = sup {µ(K) : K ⊆ B compact}<br />

µ(B) = inf {µ(V) : E ⊆ V open}.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

(proof of lemma)<br />

Let R be the collection of subsets E in X with the following<br />

property:<br />

∀ǫ > 0, ∃F closed and U open such that F ⊂ E ⊂<br />

U and µ (U\F) < ǫ.<br />

We claim that R is a σ-algebra. To see this, observe first that<br />

X ∈ R. Let then A ∈ R.<br />

To see that A c = X \A ∈ R, let ǫ > 0. <strong>The</strong>re is an open set U and<br />

a closed set F such that F ⊆ A ⊆ U such that µ (U\F) < ǫ. Set<br />

F ′ = X \U and U ′ = X \F and note that F ′ is closed, that U ′ is<br />

open and that F ′ ⊆ A c ⊆ U ′ . Since U ′ \F ′ = U\F we find that<br />

µ (U ′ \F ′ ) = µ (U\F) < ǫ.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

(proof of lemma)<br />

This shows that A c ∈ R. Let next A1,A2,A3,... be a sequence of<br />

sets from R. To see that ∞<br />

i=1 Ai ∈ R, let ǫ > 0. For each n there<br />

are Fn closed, Un open such that Fn ⊆ An ⊆ Un and<br />

µ (Un\Fn) < ǫ<br />

3 n.<br />

Set C = <br />

n Fn. Since µ(C) < ∞ we see that 1C is µ-integrable.<br />

Note that 1 S n<br />

k=1 F k increases to 1C pointwise so that (e.g.)<br />

Lebesgue’s <strong>theorem</strong> on monotone convergence implies that<br />

µ(C) = limn→∞ µ ( n<br />

k=1 Fk). (This can also be seen by use of<br />

Proposition 2 on page 255 in Royden: ’Real Analysis’.)<br />

We can there<strong>for</strong>e choose an n so big that µ (C\ n . It<br />

follows then that F = n i=1 Ai,<br />

k=1 Fn) < ǫ<br />

2<br />

k=1 Fn is a closed subset of ∞ U = ∞ k=1 Uk is an open set containing ∞ i=1 Ai and that<br />

µ (U\F) ≤ µ (U\C) + µ (C\F)<br />

<br />

∞<br />

<br />

≤ µ Uk\Fk + ǫ<br />

2 ≤<br />

∞ ǫ ǫ<br />

+ ≤ ǫ.<br />

3k 2<br />

k=1<br />

k=1<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

(proof of lemma)<br />

We have now shown that R is a σ-algebra. Let F ⊆ X be a closed<br />

set. Let d be a metric <strong>for</strong> the topology on X. For each n,<br />

<br />

Un = x ∈ X : ∃y ∈ F : d(x,y) < 1<br />

<br />

n<br />

is an open set in X.<br />

We leave the reader to check that U1 ⊇ U2 ⊇ U3 ⊇ ..., and that<br />

∞<br />

n=1 Un = F. When this is established, it follows that<br />

limn→∞ µ (Un) = µ(F), see e.g Proposition 2 on page 255 in<br />

Royden: ’Real Analysis’.<br />

For any given ǫ > 0 we can the find an n such that µ (Un\F) < ǫ,<br />

and we conclude that F ∈ R. Thus R is a σ-algebra containing all<br />

closed sets, and must there<strong>for</strong>e contained all <strong>Borel</strong> sets.<br />

It follows that µ is both inner and outer regular.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

Now to the proof of the uniqueness part in <strong>The</strong>orem <strong>12.</strong>23:<br />

Let E and F be two resolutions of the identity on σ(T) such that<br />

<br />

<br />

λ dE(λ) = T = λ dF(λ). (6)<br />

σ(T)<br />

σ(T)<br />

By <strong>The</strong>orem <strong>12.</strong>21 there are then ∗-homomorphisms<br />

ΨE : L ∞ (E) → B(H) and ΨF : L ∞ (F) → B(H) such that<br />

and<br />

<br />

〈ΨE(f )x,x〉 =<br />

<br />

〈ΨF(g)x,x〉 =<br />

σ(T)<br />

σ(T)<br />

f dEx, f ∈ L ∞ (E),<br />

g dFx, g ∈ L ∞ (F).<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

<strong>The</strong> equation (6) means that ΨE and ΨF agree on the identity<br />

function id σ(T) on σ(T). Since ΨE and ΨF are both<br />

∗-homomorphisms it follows that they agree on any polynomial in<br />

the variables z and z.<br />

Since these polynomials are dense in C (σ(T)) by the<br />

Stone-Weierstrass <strong>theorem</strong>, it follows by continuity that ΨE and<br />

ΨF agree on C (σ(T)).<br />

Now the regularity comes in: Let x ∈ H. We need to prove that<br />

Ex(B) = Fx(B) <strong>for</strong> every <strong>Borel</strong> set B ⊆ σ(T).<br />

Let ǫ > 0. Since σ(T) is a compact metric space and both<br />

meausures, Ex and Fx, are finite we conclude from Lemma 4 that<br />

there are compact subsets KE ⊆ B and KF ⊆ B and open sets<br />

B ⊆ UE and B ⊆ UF such that<br />

and<br />

Ex (UE \KE) ≤ ǫ<br />

Fx (UF \KF) ≤ ǫ.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm


Proof of the <strong>spectral</strong> <strong>theorem</strong> - <strong>The</strong>orem <strong>12.</strong>23(uniqueness)<br />

Set K = KE ∪ KF and U = UE ∩ UF, and note that K ⊆ B ⊆ U.<br />

Since K is compact and U open, Urysohn’s lemma provides us with<br />

a function K ≺ f ≺ U.<br />

Note first that Ex (U\K) ≤ Ex (UE \KE) ≤ ǫ and<br />

Fx (U\K) ≤ Fx (UF \KF) ≤ ǫ. <br />

<br />

<strong>The</strong>n |1B − f | ≤ 2 1U\K and hence Ex (B) − <br />

<br />

σ(T) f dEx<br />

≤<br />

<br />

σ(T) |1B − f | dEx ≤ <br />

σ(T) 2 1U\K dEx = 2Ex (U\K) ≤ 2ǫ.<br />

<strong>The</strong> same estimates with E replaced by F yields also that<br />

<br />

Fx (B) − <br />

<br />

σ(T) f dFx<br />

≤ 2ǫ.<br />

Since <br />

σ(T) f dEx = <br />

σ(T) f dFx we find that<br />

|Fx (B) − Ex(B)| ≤ 4ǫ. Since ǫ > 0 was arbitrary we conclude that<br />

Fx (B) = Ex(B). Since B was an arbitrary <strong>Borel</strong> set and x an<br />

arbitrary vector in H, we conclude that E = F.<br />

Klaus Thomsen <strong>12.</strong> <strong>The</strong> <strong>spectral</strong> <strong>theorem</strong> - <strong>Borel</strong> <strong>calculus</strong> <strong>for</strong> <strong>bounded</strong> norm

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!