ANALYSIS TOOLS W ITH APPLICATIONS

579

35. Compact and Fredholm Operators and the Spectral Theorem

In this section H and B will be Hilbert spaces. Typically H and B will be
separable, but we will not assume this until it is needed later.
35.1. Compact Operators.
Proposition 35.1. Let M be a finite dimensional subspace of a Hilbert space H
then
(1) M is complete (hence closed).
(2) Closed bounded subsets of M are compact.
Proof. Using the Gram-Schmidt procedure, we may choose an orthonormal
basis {1 , . . . , n } of M. Define U : M Cn to be the unique unitary map such
that U i = ei where ei is the ith standard basis vector in Cn . It now follows that
M is complete and that closed bounded subsets of M are compact since the same
is true for Cn .
Definition 35.2. A bounded operator K : H B is compact if K maps bounded
sets into precompact sets, i.e. K(U ) is compact in B, where U := {x H : kxk < 1}
is the unit ball in H. Equivalently, for all bounded sequences {xn }
n=1 H, the
sequence {Kxn }
n=1 has a convergent subsequence in B.
Notice that if dim(H) = and T : H B is invertible, then T is not compact.
Definition 35.3. K : H B is said to have finite rank if Ran(K) B is finite
dimensional.
Corollary 35.4. If K : H B is a finite rank operator, then K is compact.
In particular if either dim(H) < or dim(B) < then any bounded operator
K : H B is finite rank and hence compact.
Example 35.5. Let (X, ) be a measure space, H = L2 (X, ) and
k(x, y)
where

n
X

fi (x)gi (y)

i=1

fi , gi L2 (X, ) for i = 1, . . . , n.
Define (Kf )(x) = X k(x, y)f (y)d(y), then K : L2 (X, ) L2 (X, ) is a finite
rank operator and hence compact.
R

Lemma 35.6. Let K := K(H, B) denote the compact operators from H to B.

Then K(H, B) is a norm closed subspace of L(H, B).
Proof. The fact that K is a vector subspace of L(H, B) will be left to the reader.
Now let Kn : H B be compact operators and K : H B be a bounded operator
such that limn kKn Kkop = 0. We will now show K is compact.
First Proof. Given > 0, choose N = N ( ) such that kKN Kk < .
Using the fact that KN U is precompact, choose a finite subset U such that
minx ky KN xk < for all y KN (U ) . Then for z = Kx0 K(U ) and x ,
kz Kxk = k(K KN )x0 + KN (x0 x) + (KN K)xk
2 + kKN x0 KN xk.

BRUCE K. DRIVER

580

Therefore minx kz KN xk < 3 , which shows K(U ) is 3 bounded for all > 0,
K(U ) is totally bounded and hence precompact.

Second Proof. Suppose

{xn }n=1 is a bounded sequence
in H. By compactness,
1

1
there is a subsequence xn n=1 of {xn }
such
that
K
x
1 n n=1 is convergent in
n=1
B. Working inductively, we may construct subsequences

{xn }n=1 x1n n=1 x2n n=1 {xm

n }n=1 . . .

such that {Km xm

n }n=1 is convergent in B for each m. By the usual Cantors diago
nalization procedure, let yn := xnn , then {yn }
n=1 is a subsequence of {xn }n=1 such

that {Km yn }n=1 is convergent for all m. Since

kKyn Kyl k k(K Km ) yn k + kKm (yn yl )k + k(Km K) yl )k

2 kK Km k + kKm (yn yl )k ,

lim sup kKyn Kyl k 2 kK Km k 0 as m ,

which shows

n,l

{Kyn }n=1

is Cauchy and hence convergent.

Proposition 35.7. A bounded operator K : H B is compact i there exists

finite rank operators, Kn : H B, such that kK Kn k 0 as n .
Proof. Since K(U ) is compact it contains a countable dense subset and from
this it follows that K (H) is a separable subspace of B. Let {n } be an orthonormal
N
P
(y, n )n be the orthogonal projection of y
basis for K (H) B and PN y =
n=1

onto span{n }N
n=1 . Then limN kPN y yk = 0 for all y K(H).
Define Kn Pn K a finite rank operator on H. For sake of contradiction
suppose that lim supn kK Kn k = > 0, in which case there exists xnk U
such that k(K Knk )xnk k for all nk . Since K is compact, by passing to a
subsequence if necessary, we may assume {Kxnk }
nk =1 is convergent in B. Letting
y limk Kxnk ,
k(K Knk )xnk k = k(1 Pnk )Kxnk k k(1 Pnk )(Kxnk y)k + k(1 Pnk )yk
kKxnk yk + k(1 Pnk )yk 0 as k .

But this contradicts the assumption that is positive and hence we must have
limn kK Kn k = 0, i.e. K is an operator norm limit of finite rank operators.
The converse direction follows from Corollary 35.4 and Lemma 35.6.
Corollary 35.8. If K is compact then so is K .
Proof. Let Kn = Pn K be as in the proof of Proposition 35.7, then Kn = K Pn
is still finite rank. Furthermore, using Proposition 12.16,
kK Kn k = kK Kn k 0 as n

showing K is a limit of finite rank operators and hence compact.

35.2. Hilbert Schmidt Operators.
Proposition 35.9. Let H and B be a separable Hilbert spaces, K : H B be a

bounded linear operator, {en }

n=1 and {um }m=1 be orthonormal basis for H and B
respectively. Then:

ANALYSIS TOOLS W ITH APPLICATIONS

(1)

581

P
2

kKen k2 =
m=1 kK um k allowing for the possibility that the sums
are infinite. In particular the Hilbert Schmidt norm of K,
n=1

kKkHS :=

n=1

kKen k ,

is well defined independent of the choice of orthonormal basis {en }

n=1 . We
say K : H B is a Hilbert Schmidt operator if kKkHS < and let
HS(H, B) denote the space of Hilbert Schmidt operators from H to B.
(2) For all K L(H, B), kKkHS = kK kHS and
kKkHS kKkop := sup {kKhk : h H 3 khk = 1} .
(3) The set HS(H, B) is a subspace of K(H, B) and kkHS is a norm on
HS(H, B) for which (HS(H, B), kkHS ) is a Hilbert space. The inner product on HS(H, B) is given by
(K1 , K2 )HS =

(35.1)

(K1 en , K2 en ).

n=1

PN
(4) Let PN x := n=1 (x, en )en be orthogonal projection onto span {ei : i N }
H and for K HS(H, B), let Kn := KPn . Then
kK KN k2op kK KN k2HS 0 as N ,

which shows that finite rank operators are dense in (HS(H, B), kkHS ) .
(5) If L is another Hilbert space and A : L H and C : B L are bounded
operators, then
kKAkHS kKkHS kAkop and kCKkHS kKkHS kCkop .
Proof. Items 1. and 2. By Parsavals equality and Fubinis theorem for sums,

n=1

kKen k =

X
X

n=1 m=1

|(Ken , um )| =

X
X

m=1 n=1

|(en , K um )| =

m=1

kK um k .

This proves kKkHS is well defined independent of basis and that kKkHS =
kK kHS . For x H \ {0} , x/ kxk may be taken to be the first element in an
orthonormal basis for H and hence

K x kKk .
HS
kxk
Multiplying this inequality by kxk shows kKxk kKkHS kxk and hence kKkop
kKkHS .
Item 3. For K1 , K2 L(H, B),
v
u
uX
kK1 + K2 kHS = t
kK1 en + K2 en k2
n=1

v
u
uX
t
[kK1 en k + kK2 en k]2 = k{kK1 en k + kK2 en k}
n=1 k

n=1

k{kK1 en k}
n=1 k 2 + k{kK2 en k}n=1 k

= kK1 kHS + kK2 kHS .

BRUCE K. DRIVER

582

From this triangle inequality and the homogeneity properties of kkHS , we now easily see that HS(H, B) is a subspace of K(H, B) and kkHS is a norm on HS(H, B).
Since

X
X
|(K1 en , K2 en )|
kK1 en k kK2 en k
n=1

n=1

v
v
u
u
uX
uX
kK1 en k2 t
kK2 en k2 = kK1 kHS kK2 kHS ,
t
n=1

n=1

the sum in Eq. (35.1) is well defined and is easily checked to define an inner product
2
on HS(H, B) such that kKkHS = (K1 , K2 )HS . To see that HS(H, B) is complete

in this inner product suppose {Km }m=1 is a kkHS Cauchy sequence in HS(H, B).
Because L(H, B) is complete, there exists K L(H, B) such that kKm Kkop 0
as m . Since
N
X

n=1

k(K Km ) en k2 = lim

kKm Kk2HS =

n=1

N
X

n=1

k(Kl Km ) en k2 lim sup kKl Km kHS ,

l

k(K Km ) en k2 = lim

N
X

n=1

k(K Km ) en k2

lim sup kKl Km kHS 0 as m .

l

Item 4. Simply observe,

kK KN k2op kK KN k2HS =

n>N

kKen k2 0 as N .

Item 5. For C L(B, L) and K L(H, B) then

2

kCKkHS =

n=1

kCKen k kCkop

n=1

kKen k = kCkop kKkHS

and for A L (L, H) ,

kKAkHS = kA K kHS kA kop kK kHS = kAkop kKkHS .
Remark 35.10. The separability assumptions made in Proposition 35.9 are unnecessary. In general, we define
X
kKek2
kKk2HS =
e

where H is an orthonormal basis. The same proof of Item 1. of Proposition 35.9

2
shows kKkHS is well defined and kKkHS = kK kHS . If kKkHS < , then there
exists a countable subset 0 such that Ke = 0 if e \0 . Let H0 := span(0 )
and B0 := K(H0 ). Then K (H) B0 , K|H0 = 0 and hence by applying the results
of Proposition 35.9 to K|H0 : H0 B0 one easily sees that the separability of H
and B are unnecessary in Proposition 35.9.

ANALYSIS TOOLS W ITH APPLICATIONS

583

Exercise 35.1. Suppose that (X, ) is a finite measure space such that H =
L2 (X, ) is separable and k : X X R is a measurable function, such that
Z
|k(x, y)|2 d(x)d(y) < .
kkk2L2 (XX,)
XX

Define, for f H,

Kf (x) =

k(x, y)f (y)d(y),

when the integral makes sense. Show:

(1) Kf (x) is defined for a.e. x in X.
(2) The resulting function Kf is in H and K : H H is linear.
(3) kKkHS = kkkL2 (XX,) < . (This implies K HS(H, H).)
Z

35.1. Since
Z
2 Z
Z
Z

2
2
d(x)
|k(x, y)f (y)| d(y)
d(x)
|k(x, y)| d(y)
|f (y)| d(y)

kkk2 kf k2 < ,

(35.2)

we learn Kf is almost everywhere defined and that Kf H. The linearity of K is

a consequence of the linearity of the Lebesgue integral. Now suppose {n }
n=1 is
an orthonormal basis for H. From the estimate in Eq. (35.2), k(x, ) H for
a.e. x X and therefore
Z
2
Z
X

kKk2HS =
d(x) k(x, y)n (y)d(y)
=

n=1 X
Z
X
n=1

))2 =
d(x) (n , k(x,

2
d(x)
k(x, )H =

d(x)

d(x)

))2
(n , k(x,

n=1

d(y) |k(x, y)|2 = kkk22 .

Example 35.11. Suppose that Rn is a bounded set, < n, then the operator
K : L2 (, m) L2 (, m) defined by
Z
1
Kf (x) :=
f (y)dy
|x y|
is compact.
Proof. For

0, let
K f (x) :=

1
f (y)dy = [g (1 f )] (x)
|x y| +

where g (x) = |x| + 1C (x) with C Rn a suciently large ball such that
C. Since < n, it follows that

g g0 = ||

1C L1 (Rn , m).

Hence it follows by Proposition 11.12 ?? that

k(K K ) f kL2 () k(g0 g ) (1 f )kL2 (Rn )

k(g0 g )kL1 (Rn ) k1 f kL2 (Rn ) = k(g0 g )kL1 (Rn ) kf kL2 ( )

BRUCE K. DRIVER

584

which implies
(35.3)
kK K kB(L2 ()) kg0 g kL1 (Rn ) =

1
1

dx 0 as

|x|
C |x| +

by the dominated convergence theorem. For any > 0,

2
Z
1
dxdy < ,

|x y| +

and hence K is Hilbert Schmidt and hence compact. By Eq. (35.3), K K as

0 and hence it follows that K is compact as well.
35.3. The Spectral Theorem for Self Adjoint Compact Operators.
Lemma 35.12. Suppose T : H B is a bounded operator, then Nul(T ) =
Ran(T ) and Ran(T ) = Nul(T ) .
Proof. An element y B is in Nul(T ) i 0 = (T y, x) = (y, Ax) for all
x H which happens i y Ran(T ) . Because Ran(T ) = Ran(T ) , Ran(T ) =
Nul(T ) .
For the rest of this section, T K(H) := K(H, H) will be a self-adjoint compact
operator or S.A.C.O. for short.
Example 35.13 (Model S.A.C.O.). Let H =

1 0
0
0 2 0

T = 0
0 3

..
.. . .
.
.
.

and T be the diagonal matrix

..
.

where limn |n | = 0 and n R. Then T is a self-adjoint compact operator.

(Prove!)
The main theorem of this subsection states that up to unitary equivalence, Example 35.13 is essentially the most general example of an S.A.C.O.
Theorem 35.14. Suppose T L(H) := L(H, H) is a bounded self-adjoint operator, then
|(f, T f )|
.
kT k = sup
kf k2
f 6=0
Moreover if there exists a non-zero element g H such that
|(T g, g)|
= kT k,
kgk2

then g is an eigenvector of T with T g = g and {kT k}.

Proof. Let
M sup
f 6=0

|(f, T f )|
.
kf k2

We wish to show M = kT k. Since |(f, T f )| kf kkT f k kT kkf k2 , we see M

kT k.

ANALYSIS TOOLS W ITH APPLICATIONS

585

Conversely let f, g H and compute

(f + g, T (f + g)) (f g, T (f g))

= (f, T g) + (g, T f ) + (f, T g) + (g, T f )

= 2[(f, T g) + (T g, f )] = 2[(f, T g) + (f, T g)]
= 4Re(f, T g).

Therefore, if kf k = kgk = 1, it follows that

M

M
kf + gk2 + kf gk2 =
2kf k2 + 2kgk2 = M.
|Re(f, T g)|
4
4
By replacing f be ei f where is chosen so that ei (f, T g) is real, we find
Hence

|(f, T g)| M for all kf k = kgk = 1.

kT k =

sup
kf k=kgk=1

|(f, T g)| M.

If g H \{0} and kT k = |(T g, g)|/kgk2 then, using the Cauchy Schwarz inequality,
(35.4)

kT k =

kT gk
|(T g, g)|

kT k.
kgk2
kgk

This implies |(T g, g)| = kT gkkgk and forces equality in the Cauchy Schwarz inequality. So by Theorem 12.2, T g and g are linearly dependent, i.e. T g = g
for some C. Substituting this into (35.4) shows that || = kT k. Since T is
self-adjoint,
g),
kgk2 = (g, g) = (T g, g) = (g, T g) = (g, g) = (g,
which implies that R and therefore, {kT k}.
Theorem 35.15. Let T be a S.A.C.O., then either = kT k or = kT k is an
eigenvalue of T.
Proof. Without loss of generality we may assume that T is non-zero since
otherwise the result is trivial. By Theorem 35.14, there exists fn H such that
kfn k = 1 and
(35.5)

|(fn , T fn )|
= |(fn , T fn )| kT k as n .
kfn k2

By passing to a subsequence if necessary, we may assume that := limn (fn , T fn )

exists and {kT k}. By passing to a further subsequence if necessary, we may
assume, using the compactness of T, that T fn is convergent as well. We now compute:
0 kT fn fn k2 = kT fn k2 2(T fn , fn ) + 2
Hence
(35.6)
and therefore

2 2(T fn , fn ) + 2 2 22 + 2 = 0 as n .
T fn fn 0 as n
f lim fn =
n

1
lim T fn
n

BRUCE K. DRIVER

586

exists. By the continuity of the inner product, kf k = 1 6= 0. By passing to the limit

in Eq. (35.6) we find that T f = f.
Lemma 35.16. Let T : H H be a self-adjoint operator and M be a T invariant
subspace of H, i.e. T (M ) M. Then M is also a T invariant subspace, i.e.
T (M ) M .
Proof. Let x M and y M , then T x M and hence
0 = (T x, y) = (x, T y) for all x M.
Thus T y M .
Theorem 35.17 (Spectral Theorem). Suppose that T : H H is a non-zero
S.A.C.O., then
(1) there exists at least one eigenvalue {kT k}.
(2) There are at most countable many non-zero eigenvalues, {n }N
n=1 , where
N = is allowed. (Unless T is finite rank, N will be infinite.)
(3) The n s (including multiplicities) may be arranged so that |n | |n+1 |
for all n. If N = then limn |n | = 0. (In particular any eigenspace
for T with non-zero eigenvalue is finite dimensional.)
(4) The eigenvectors {n }N
n=1 can be chosen to be an O.N. set such that H =
span{n } Nul(T ).
(5) Using the {n }N
n=1 above,
T =

N
X

n=1

n (, n )n for all H.

(6) The spectrum of T is (T ) = {0}

n=1 {n }.
Proof. We will find n s and n s recursively. Let 1 {kT k} and 1 H
such that T 1 = 1 1 as in Theorem 35.15. Take M1 = span(1 ) so T (M1 ) M1 .
By Lemma 35.16, T M1 M1 . Define T1 : M1 M1 via T1 = T |M1 . Then T1
is again a compact operator. If T1 = 0, we are done.
If T1 6= 0, by Theorem 35.15 there exists 2 {kT k1 } and 2 M1 such that
k2 k = 1 and T1 2 = T 2 = 2 2 . Let M2 span(1 , 2 ). Again T (M2 ) M2
and hence T2 T |M2 : M2 M2 is compact. Again if T2 = 0 we are done.
If T2 6= 0. Then by Theorem 35.15 there exists 3 {kT k2 } and 3 M2 such
that k3 k = 1 and T2 3 = T 3 = 3 3 . Continuing this way indefinitely or until we
reach a point where Tn = 0, we construct a sequence {n }N
n=1 of eigenvalues and
orthonormal eigenvectors {n }N
such
that
|
|

|
|
with
the further property
i
i+1
n=1
that
(35.7)

|i | =

kT k
{1 ,2 ,...i1 } kk
sup

If N = then limi |i | = 0 for if not there would exist > 0 such that
1
|i | > 0 for all i. In this case {i /i }
.
i=1 is sequence in H bounded by
By compactness of T, there exists a subsequence ik such that ik = T ik /ik is
convergent. But this is impossible since {ik } is an orthonormal set. Hence we
must have that = 0.

ANALYSIS TOOLS W ITH APPLICATIONS

587

Let M span{i }N
i=1 with N = possible. Then T (M ) M and hence
T (M ) M . Using Eq. (35.7),
kT |M k kT |Mn k = |n | 0 as n
showing T |M 0.
Define P0 to be orthogonal projection onto M . Then for H,
= P0 + (1 P0 ) = P0 +
and

N
X
(, i )i
i=1

N
N
X
X
(, i )i =
i (, i )i .
T = T P0 + T
i=1

i=1

Since {n } (T ) and (T ) is closed, it follows that 0 (T ) and hence {n }

n=1
{0} (T ). Suppose that z
/ {n }
n=1 {0} and let d be the distance between
z and {n }
n=1 {0}. Notice that d > 0 because limn n = 0. A few simple
computations show that:
(T zI) =
(T z)

N
X
(, i )(i z)i zP0 ,
i=1

exists,
(T zI)1 =

and
k(T zI)1 k2 =

N
X
i=1

N
X
(, i )(i z)1 i z 1 P0 ,
i=1

|(, i )|2

1
1
+ 2 kP0 k2
2
|i z|
|z|
!

2 X
N
1

|(, i )|2 + kP0 k2

d
i=1

1
kk2 .
d2

We have thus shown that (T zI)1 exists, k(T zI)1 k d1 < and hence
z
/ (T ).
35.4. Structure of Compact Operators.
Theorem 35.18. Let K : H B be a compact operator. Then there exists N
N
N {} , orthonormal subsets {n }N
n=1 H and {n }n=1 B and a sequences
{n }N
n=1 C such that limn n = 0 if N = and
Kf =

N
X

n=1

n (f, n )n for all f H.

Proof. The operator K K K(H) is self-adjoint and hence by Theorem 35.17,

there exists an orthonormal set {n }N

n=1 H and {n }n=1 (0, ) such that
K Kf =

N
X

n=1

n (f, n )n for all f H.

BRUCE K. DRIVER

588

Let n :=

n and

K K K(H) be defined by

N
X

K Kf =
n (f, n )n for all f H.
n=1

Define U L(H, B) so that U =K (K K)

(35.8)

Uf =

N
X

1/2

, or more precisely by

1
n (f, n )Kn .

n=1

The operator U is well defined because

1
1 1

1 1 2
(1
n Kn , m Km ) = n m (n , K Km ) = n m m m,n = m,n

which shows 1
n Kn n=1 is an orthonormal subset of B. Moreover this also shows
2

kU f k =

N
X

n=1

|(f, n )| = kP f k

where P = PNul(K) . Replacing f by (K K)1/2 f in Eq. (35.8) shows

(35.9)

1/2

U (K K)

f=

N
X

1
n ((K K)

1/2

f, n )Kn =

n=1

(f, n )Kn = Kf,

n=1

PN
since f = n=1 (f, n )n + P f.
From Eq. (35.9) it follows that
Kf =

N
X

N
X

n (f, n )U n =

n=1

N
X

n (f, n )n

n=1

where {n }N
n=1 is the orthonormal sequence in B defined by
n := U n = 1
n Kn .

35.4.1. Trace Class Operators. We will say K K(H) is trace class if

N
X

n <
tr( K K) :=
n=1

in which case we define

tr(K) =

N
X

n (n , n ).

n=1

Notice that if {em }

m=1 is any orthonormal basis in H (or for the Ran(K) if H is
not separable) then
M
X

(Kem , em ) =

m=1

M X
N
N
M
X
X
X
(
n (em , n )n , em ) =
n
(em , n )(n , em )

m=1 n=1

N
X

n=1

n (PM n , n )

n=1

m=1

ANALYSIS TOOLS W ITH APPLICATIONS

589

where PM is orthogonal projection onto Span(e1 , . . . , eM ). Therefore by dominated

convergence theorem ,

(Kem , em ) = lim

m=1

N
X

N
X

n (PM n , n ) =

n=1

N
X

n lim (PM n , n )

n=1

n (n , n ) = tr(K).

n=1

35.5. Fredholm Operators.

Lemma 35.19. Let M H be a closed subspace and V H be a finite dimensional
subspace. Then M +V is closed as well. In particular if codim(M ) dim(H/M ) <
and W H is a subspace such that M W, then W is closed and codim(W ) <
.
Proof. Let P : H M be orthogonal projection and let V0 := (I P ) V. Since

dim(V0 ) dim(V ) < , V0 is still closed. Also it is easily seen that M +V = M V0

from which it follows that M + V is closed because {zn = mn + vn } M V0 is

convergent i {mn } M and {vn } V0 are convergent.
If codim(M ) < and M W, there is a finite dimensional subspace V H
such that W = M + V and so by what we have just proved, W is closed as well. It
should also be clear that codim(W ) codim(M ) < .
Lemma 35.20. If K : H B is a finite rank operator, then there exists
{n }kn=1 H and {n }kn=1 B such that
P
(1) Kx = kn=1 (x, n )n for all x H.
P
(2) K y = kn=1 (y, n )n for all y B, in particular K is still finite rank.
For the next two items, further assume B = H.
(3) dim Nul(I + K) < .
(4) dim coker(I + K) < , Ran(I + K) is closed and
Ran(I + K) = Nul(I + K ) .

Proof.
(1) Choose {n }k1 to be an orthonormal basis for Ran(K). Then for x H,
Kx =

k
X

(Kx, n )n =

n=1

k
X

n=1

(x, K n )n =

k
X

(x, n )n

n=1

where n K n .
(2) Item 2. is a simple computation left to the reader.
(3) Since Nul(I + K) = {x H | x = Kx} Ran(K) it is finite dimensional.
(4) Since x = (I + K)x Ran(I + K)for x Nul(K), Nul(K) Ran(I + K).
Since {1 , 2 , . . . , k } Nul(K), H = Nul(K) + span ({1 , 2 , . . . , k })
and thus codim (Nul(K)) < . From these comments and Lemma 35.19,
Ran(I + K) is closed and codim (Ran(I + K)) codim (Nul(K)) < .
The assertion that Ran(I + K) = Nul(I + K ) is a consequence of Lemma
35.12 below.

590

BRUCE K. DRIVER

Definition 35.21. A bounded operator F : H B is Fredholm i the

dim Nul(F ) < , dim coker(F ) < and Ran(F ) is closed in B. (Recall:
coker(F ) B/Ran(F ).) The index of F is the integer,
(35.10)

(35.11)

index(F ) = dim Nul(F ) dim coker(F )

= dim Nul(F ) dim Nul(F )

Notice that equations (35.10) and (35.11) are the same since, (using Ran(F ) is
closed)
B = Ran(F ) Ran(F ) = Ran(F ) Nul(F )
so that coker(F ) = B/Ran(F )
= Nul(F ).
Lemma 35.22. The requirement that Ran(F ) is closed in Defintion 35.21 is redundant.
Proof. By restricting F to Nul(F ) , we may assume without loss of generality
that Nul(F ) = {0}. Assuming dim coker(F ) < , there exists a finite dimensional
subspace V B such that B = Ran(F ) V. Since V is finite dimensional, V is
closed and hence B = V V . Let : B V be the orthogonal projection
operator onto V and let G F : H V which is continuous, being the
composition of two bounded transformations. Since G is a linear isomorphism, as
the reader should check, the open mapping theorem implies the inverse operator
G1 : V H is bounded.
Suppose that hn H is a sequence such that limn F (hn ) =: b exists in B.
Then by composing this last equation with , we find that limn G(hn ) = (b)
exists in V . Composing this equation with G1 shows that h := limn hn =
G1 (b) exists in H. Therefore, F (hn ) F (h) Ran(F ), which shows that
Ran(F ) is closed.
Remark 35.23. It is essential that the subspace M Ran(F ) in Lemma 35.22 is
the image of a bounded operator, for it is not true that every finite codimensional
subspace M of a Banach space B is necessarily closed. To see this suppose that B
is a separable infinite dimensional Banach space and let A B be an algebraic
basis for B, which exists by a Zorns lemma argument. Since dim(B) = and B
is complete, A must be uncountable. Indeed, if A were countable we could write
B =
n=1 Bn where Bn are finite dimensional (necessarily closed) subspaces of B.
This shows that B is the countable union of nowhere dense closed subsets which
violates the Baire Category theorem.
By separability of B, there exists a countable subset A0 A such that the closure
of M0 span(A0 ) is equal to B. Choose x0 A \ A0 , and let M span(A \ {x0 }).
0 = M
, while codim(M ) = 1. Clearly this M can
Then M0 M so that B = M
not be closed.
Example 35.24. Suppose that H and B are finite dimensional Hilbert spaces and
F : H B is Fredholm. Then
(35.12)

index(F ) = dim(B) dim(H).

The formula in Eq. (35.12) may be verified using the rank nullity theorem,
dim(H) = dim Nul(F ) + dim Ran(F ),
and the fact that
dim(B/Ran(F )) = dim(B) dim Ran(F ).

ANALYSIS TOOLS W ITH APPLICATIONS

591

Theorem 35.25. A bounded operator F : H B is Fredholm i there exists

a bounded operator A : B H such that AF I and F A I are both compact
operators. (In fact we may choose A so that AF I and F A I are both finite
rank operators.)
Proof. () Suppose F is Fredholm, then F : Nul(F ) Ran(F ) is a bijective bounded linear map between Hilbert spaces. (Recall that Ran(F ) is a closed
subspace of B and hence a Hilbert space.) Let F be the inverse of this mapa
bounded map by the open mapping theorem. Let P : H Ran(F ) be orthogonal
projection and set A F P . Then AF I = F P F I = F F I = Q where Q is
the orthogonal projection onto Nul(F ). Similarly, F A I = F F P I = (I P ).
Because I P and Q are finite rank projections and hence compact, both AF I
and F A I are compact.
() We first show that the operator A : B H may be modified so that
AF I and F A I are both finite rank operators. To this end let G AF I (G
is compact) and choose a finite rank approximation G1 to G such that G = G1 + E
where kEk < 1. Define AL : B H to be the operator AL (I + E)1 A. Since
AF = (I + E) + G1 ,
AL F = (I + E)1 AF = I + (I + E)1 G1 = I + KL
where KL is a finite rank operator. Similarly there exists a bounded operator
AR : B H and a finite rank operator KR such that F AR = I + KR . Notice that
AL F AR = AR + KL AR on one hand and AL F AR = AL + AL KR on the other.
Therefore, AL AR = AL KR KL AR =: S is a finite rank operator. Therefore
F AL = F (AR + S) = I + KR + F S, so that F AL I = KR F S is still a finite rank
operator. Thus we have shown that there exists a bounded operator A : B H
I and F A I are both finite rank operators.
such that AF
We now assume that A is chosen such that AF I = G1 , F A I = G2 are
finite rank. Clearly Nul(F ) Nul(AF ) = Nul(I + G1 ) and Ran(F ) Ran(F A) =
Ran(I + G2 ). The theorem now follows from Lemma 35.19 and Lemma 35.20.
Corollary 35.26. If F : H
B is Fredholm then F is Fredholm and index(F ) =

index(F ).
Proof. Choose A : B H such that both AF I and F A I are compact.
Then F A I and A F I are compact which implies that F is Fredholm. The
assertion, index(F ) = index(F ), follows directly from Eq. (35.11).
Lemma 35.27. A bounded operator F : H B is Fredholm if and only if there
exists orthogonal decompositions H = H1 H2 and B = B1 B2 such that
(1) H1 and B1 are closed subspaces,
(2) H2 and B2 are finite dimensional subspaces, and
(3) F has the block diagonal form

(35.13)

F =

F11
F21

F12
F22

H1

H2

B1

B2

with F11 : H1 B1 being a bounded invertible operator.

Furthermore, given this decomposition, index(F ) = dim(H2 ) dim(B2 ).

BRUCE K. DRIVER

592

Proof. If F is Fredholm, set H1 = Nul(F

) , H2 = Nul(F ), B1 = Ran(F ),
F11 0
and B2 = Ran(F ) . Then F =
, where F11 F |H1 : H1 B 1 is
0
0
invertible.

1
F11 0
For the converse, assume that F is given as in Eq. (35.13). Let A
0
0
then

1
1
I 0
I F11
F12
F12
0 F11
AF =
=
+
,
0 I
0
0
0
I

so that AF I is finite rank. Similarly one shows that F A I is finite rank, which
shows that F is Fredholm.

x1
Now to compute the index of F, notice that
Nul(F ) i
x2
F11 x1 + F12 x2 = 0
F21 x1 + F22 x2 = 0
1
1
F12 x2 and (F21 F11
F12 + F22 )x2 = 0. Let D
which happens i x1 = F11
1
(F22 F21 F11 F12 ) : H2 B2 , then the mapping

1
F12 x2
F11
Nul(F )
x2 Nul(D)
x2

is a linear isomorphism of vector spaces so that Nul(F )

= Nul(D). Since
F =

F11

F12

F21

F22

B1

B2

H1
,
H2

similar reasoning implies Nul(F )

= Nul(D ). This shows that index(F ) =
index(D). But we have already seen in Example 35.24 that index(D) = dim H2
dim B2 .
Proposition 35.28. Let F be a Fredholm operator and K be a compact operator
from H B. Further assume T : B X (where X is another Hilbert space) is
also Fredholm. Then
(1) the Fredholm operators form an open subset of the bounded operators. Moreover if E : H B is a bounded operator with kEk suciently small we have
index(F ) =index(F + E).
(2) F + K is Fredholm and index(F ) = index(F + K).
(3) T F is Fredholm and index(T F ) = index(T ) + index(F )
Proof.
(1) We know F may be written in the block form given in Eq. (35.13) with
F11 : H1 B1 being a bounded invertible operator. Decompose E into the
block form as

E11 E12
E=
E21 E22

and choose kEk suciently small such that kE11 k is suciently small to
guarantee that F11 + E11 is still invertible. (Recall that the
invertible opF11 + E11
erators form an open set.) Thus F + E =
has the block

ANALYSIS TOOLS W ITH APPLICATIONS

593

form of a Fredholm operator and the index may be computed as:

index(F + E) = dim H2 dim B2 = index(F ).
(2) Given K : H B compact, it is easily seen that F + K is still Fredholm.
Indeed if A : B H is a bounded operator such that G1 AF I and
G2 F A I are both compact, then A(F + K) I = G1 + AK and
(F + K)A I = G2 + KA are both compact. Hence F + K is Fredholm
by Theorem 35.25. By item 1., the function f (t) index(F + tK) is a
continuous locally constant function of t R and hence is constant. In
particular, index(F + K) = f (1) = f (0) = index(F ).
(3) It is easily seen, using Theorem 35.25 that the product of two Fredholm
operators is again Fredholm. So it only remains to verify the index formula
in item 3.
For this let H1 Nul(F ) , H2 Nul(F ), B1 Ran(T ) = T (H1 ), and
B2 Ran(T ) = Nul(T ). Then F decomposes into the block form:
F =

F
0

0
0

H1

H2

B1
,
B2

where F = F |H1 : H1 B1 is an invertible operator. Let Y1 T (B1 )

and Y2 Y1 = T (B1 ) . Notice that Y1 = T (B1 ) = T Q(B1 ), where
Q : B B1 B is orthogonal projection onto B1 . Since B1 is closed
and B2 is finite dimensional, Q is Fredholm. Hence T Q is Fredholm and
Y1 = T Q(B1 ) is closed in Y and is of finite codimension. Using the above
decompositions, we may write T in the block form:
T =

T11
T21

T12
T22

B1

B2

Y1
.
Y2

0 T12
: B Y is a finite rank operator and hence
T21 T22
RF : H Y is finite rank, index(T R) = index(T ) and index(T F RF ) =
index(T F ).
loss of generality we may assume that T has the
Hence without

T 0
form T =
, (T = T11 ) and hence
0 0
Since R =

TF =

TF
0

0
0

H1

H2

Y1
.
Y2

We now compute the index(T ). Notice that Nul(T ) = Nul(T) B2 and

Ran(T ) = T(B1 ) = Y1 . So
index(T ) = index(T) + dim(B2 ) dim(Y2 ).
Similarly,
index(T F ) = index(TF ) + dim(H2 ) dim(Y2 ),
and as we have already seen
index(F ) = dim(H2 ) dim(B2 ).

BRUCE K. DRIVER

594

Therefore,
index(T F ) index(T ) index(F ) = index(TF ) index(T).

Since F is invertible, Ran(T) = Ran(TF ) and Nul(T)

= Nul(TF ). Thus
index(TF ) index(T) = 0 and the theorem is proved.
35.6. Tensor Product Spaces . References for this section are Reed and Simon
[?] (Volume 1, Chapter VI.5), Simon [?], and Schatten [?]. See also Reed and Simon
[?] (Volume 2 IX.4 and XIII.17).
Let H and K be separable Hilbert spaces and H K will denote the usual
Hilbert completion of the algebraic tensors H f K. Recall that the inner product
on H K is determined by (hk, h0 k0 ) = (h, h0 )(k, k0 ). The following proposition
is well known.
Proposition 35.29 (Structure of H K). There is a bounded linear map T :
H K B(K, H) determined by
T (h k)k 0 (k, k 0 )h for all k, k 0 K and h H.

Moreover T (H K) = HS(K, H) the Hilbert Schmidt operators from K to H.

The map T : H K HS(K, H) is unitary equivalence of Hilbert spaces. Finally,
any A H K may be expressed as
(35.14)

A=

n=1

n hn kn ,

where {hn } and {kP

n } are orthonormal sets in H and K respectively and {n } R
such that kAk2 = |n |2 < .
P
} are orthonormal bases for H
Proof. Let A aji hj ki , where {hi } and {kj P
2
and K respectively
and
{a
}

R
such
that
kAk
=
|aji |2 < . Then evidently,
ji
P
T (A)k aji hj (ki , k) and
X X
XX
XX
kT (A)kk2 =
|
aji (ki , k)|2
|aji |2 |(ki , k)|2
|aji |2 kkk2 .
j

Thus T : H K B(K, H) is bounded. Moreover,

X
X
kT (A)ki k2 =
|aji |2 = kAk2 ,
kT (A)k2HS
ij

which proves the T is an isometry.

We will now prove that T is surjective and at the same time prove Eq. (35.14).
To motivate the construction, suppose that Q = T (A) where A is given as in Eq.
(35.14). Then
Q Q = T (

n=1

n kn hn )T (

n=1

n hn kn ) = T (

n=1

2n kn kn ).

That is {kn } is an orthonormal basis for (nulQ Q) with Q Qkn = 2n kn . Also

Qkn = n hn , so that hn = 1
n Qkn .
We will now reverse the above argument. Let Q HS(K, H). Then Q Q is a selfadjoint compact operator on K. Therefore there is an orthonormal basis {kn }
n=1

ANALYSIS TOOLS W ITH APPLICATIONS

595

for the (nulQ Q) which consists of eigenvectors of Q Q. Let n (0, ) such

that Q Qkn = 2n kn and set hn = 1
n Qkn . Notice that
1
1
1
1
1 2
(hn , hm ) = (1
n Qkn , m Qkm ) = (n kn , m Q Qkm ) = (n kn , m m km ) = mn ,

so that {hn } is an orthonormal set in H. Define

X
n hn kn
A=
n=1

and notice that T (A)kn = n hn = Qkn for all n and T (A)k = 0 for all k nulQ =
nulQ Q. That
is T (A) = Q. P
Therefore T is surjective and Eq. (35.14)
Pholds.

i
i
Recall that 1 z = 1
c
z
for
|z|
<
1,
where
c

0
and
i=1 i
i=1 ci < .
For an operator A on H such that A 0 and kAkB(H) 1, the square root of A is
given by

A=I
ci (A I)i .
i=1

See Theorem VI.9 on p. 196 of Reed and Simon

[?]. The next proposition is

problem 14 and 15 on p. 217 of [?]. Let |A| A A.

Proposition 35.30 (Square Root). Suppose that An

and A are positive operators

on H and kAAn kB(H) 0 as n , then An A in B(H) also. Moreover,

An and A are general bounded operators on H and An A in the operator norm
then |An | |A|.
Proof. With out loss of generality, assume that kAn k 1 for all n. This implies
also that that kAk 1. Then

X
p

A An =
ci {(An I)i (A I)i }
i=1

and hence
(35.15)

X
p

k A An k
ci k(An I)i (A I)i k.
i=1

For the moment we will make the additional assumption that An I, where
(0, 1). Then 0 I An (1 )I and in particular kI An kB(H) (1 ).
Now suppose that Q, R, S, T are operators on H, then QR ST = (Q S)R +
S(R T ) and hence
kQR ST k kQ SkkRk + kSkkR T k.

Setting Q = An I, R (An I)i1 , S (A I) and T = (A I)i1 in this last

inequality gives
k(An I)i (A I)i k kAn Akk(An I)i1 k + k(A I)kk(An I)i1 (A I)i1 k
(35.16)

kAn Ak(1 )i1 + (1 )k(An I)i1 (A I)i1 k.

It now follows by induction that

k(An I)i (A I)i k i(1 )i1 kAn Ak.

Inserting this estimate into (35.15) shows that

X
p

1
1 1
1
kAAn k = kAAn k 0.
ci i(1 )i1 kAn Ak = p
k A An k
2
2
1 (1 )
i=1

BRUCE K. DRIVER

596

Therefore we have shown if An I for all n and An A in norm then

in norm.
For the general case where An 0, we find that for all > 0
p

(35.17)
lim
An + = A + .

An

54

By the spectral theorem

k A + Ak max | x + x|
x(A)

max | x + x| 0 as

0xkAk

0.

Since the above estimates are uniform in A 0 such that kAk is bounded, it is now
an easy matter to conclude that Eq. (35.17) holds even when = 0.
Now suppose that An A in B(H) and An and A are general operators. Then
An An A A in B(H). So by what we have already proved,

p
|An | An An |A| A A in B(H) as n .

Notation 35.31. In the future we will identify A H K with T (A) HS(K, H)

and drop T from the notation. So that with this notation we have (h k)k 0 =
(k, k 0 )h.
p

Let A H H, we set kAk1 tr A A tr T (A) T (A) and we let

H 1 H {A H H : kAk1 < }.

We will now
Pcompute kAk1 for A H H described as in Eq. (35.14). First notice
that A =
n=1 n kn hn and
A A =

n=1

2n kn kn .

P
P
Hence P
A A = n=1 |n |kn kn and hence kAk1 = n=1 |n |. Also notice that
2
kAk2 =
n=1 |n | and kAkop = maxn |n |. Since
kAk21 = {

n=1

|n |}2

n=1

|n |2 = kAk2 ,

we have the following relations among the various norms,

kAkop kAk kAk1 .

(35.18)

Proposition 35.32. There is a continuous linear map C : H 1 H R such that

C(h k) = (h, k) for all h, k H. If A H 1 H, then
X
(35.19)
CA =
(em em , A),
where {em } is any orthonormal basis for H. Moreover, if A H 1 H is positive,
i.e. T (A) is a non-negative operator, then kAk1 = CA.

54 It is possible to give a more elementary proof here. Indeed, assume further that kAk < 1,

then for

P
i
i
(0, 1 ), k A + Ak
i=1 ci k(A + ) A k. But

i
i
X
X
i k ik
i k
kA
k
kAkik = (kAk + )i kAki ,
k
k
k=1
k=1
p
p

kAk + kAk 0 as 0 uniformly in A 0 such that

so that k A + Ak
kAk < 1.

k(A + )i Ai k

ANALYSIS TOOLS W ITH APPLICATIONS

597

P
Proof. Let A P
H 1 H be given as in Eq. (35.14) with P
n=1 |n | = kAk1 < .
Then define CA

(h
,
k
)
and
notice
that
|CA|

|n | = kAk1 , which
n
n
n
n=1
shows that C is a contraction on H 1 H. (Using the universal property of H f H
it is easily seen that C is well defined.) Also notice that for M Z+ that
(35.20)

M
X

(em em , A) =

m=1

(35.21)

M
X
X

(em em , n hn kn , ),

n=1 m=1

n (PM hn , kn ),

n=1

where PMPdenotes orthogonal projection onto span{em }M

m=1 . Since |n (PM hn , kn )|
|n | and
|
|
=
kAk
<
,
we
may
let
M

in Eq. (35.21) to find that

n
1
n=1

(em em , A) =

m=1

n (hn , kn ) = CA.

n=1

This proves Eq. (35.19).

For the final assertion, suppose that A 0. Then there is an orthonormal
P basis

{kn }
of
eigenvectors
of
A.
That
is
A
=
n kn
n=1 for the (nulA) which consists
P
P
kn and n 0 for all n. Thus CA = n and kAk1 = n .
Proposition 35.33 (Noncommutative Fatou s Lemma). Let An be a sequence of
positive operators on a Hilbert space H and An A weakly as n , then
trA lim inf trAn .

(35.22)

Also if An H 1 H and An A in B(H), then

kAk1 lim inf kAn k1 .

(35.23)

Proof. Let An be a sequence of positive operators on a Hilbert space H and

An A weakly as n and {ek }
k=1 be an orthonormal basis for H. Then by
Fatous lemma for sums,
trA =

(Aek , ek ) =

k=1

lim inf
n

k=1

k=1

lim (An ek , ek )

(An ek , ek ) = lim inf trAn .

n

Now suppose that An H 1 H and An A in B(H). Then by Proposition 35.30, |An | |A| in B(H) as well. Hence by Eq. (35.22), kAk1 tr|A|
lim inf n tr|An | lim inf n kAn k1 .
Proposition 35.34. Let X be a Banach space, B : H K X be a bounded
bi-linear form, and kBk sup{|B(h, k)| : khkkkk 1}. Then there is a unique
k) = B(h, k). Moreover
: H 1 K X such that B(h
bounded linear map B

kBkop = kBk.

BRUCE K. DRIVER

598

P
is
Proof. Let A =
kn H 1 K as in Eq. (35.14). Clearly, if B
n=1 n hnP

to exist we must have B(A)

n=1 n B(hn , kn ). Notice that

n=1

|n ||B(hn , kn )|

n=1

|n |kBk = kAk1 kBk.

op kBk.
The opposite inequality
This shows that B(A)
is well defined and that kBk
follows from the trivial computation:
k)| : kh 1 kk1 = 1} kBk
op .
kBk = sup{|B(h, k)| : khkkkk = 1} = sup{|B(h
Lemma 35.35. Suppose that P B(H) and Q B(K), then P Q : H K
H K is a bounded operator. Moreover, P Q(H 1 K) H 1 K and we have
the norm equalities
kP QkB(HK) = kP kB(H) kQkB(K)
and
kP QkB(H1 K) = kP kB(H) kQkB(K) .
We will give essentially the same proof of kP QkB(HK) = kP kB(H) kQkB(K)
as the proof on p. 299 of Reed and Simon [?]. Let A H K as in Eq. (35.14).
Then

X
n P hn kn
(P I)A =
n=1

and hence

(P I)A{(P I)A} =
Therefore,

n=1

2n P hn P hn .

k(P I)Ak2 = tr(P I)A{(P I)A}

=

n=1

2n (P hn , P hn ) kP k2
2

= kP k

kAk21 ,

2n

n=1

which shows that Thus kP IkB(HK) kP k. By symmetry, kI QkB(HK)

kQk. Since P Q = (P I)(I Q), we have
kP QkB(HK) kP kB(H) kQkB(K) .
The reverse inequality is easily proved by considering P Q on elements of the
form h k H K.
Proof. Now suppose that A H 1 K as in Eq. (35.14). Then
k(P Q)Ak1
which shows that

n=1

|n |kP hn Qkn k1 kP kkQk

n=1

|n | = kP kkQkkAk,

kP QkB(H1 K) kP kB(H) kQkB(K) .

Again the reverse inequality is easily proved by considering P Q on elements of
the form h k H 1 K.

ANALYSIS TOOLS W ITH APPLICATIONS

599

Lemma 35.36. Suppose that Pm and Qm are orthogonal projections on H and K

respectively which are strongly convergent to the identity on H and K respectively.
Then Pm Qm : H 1 K
H 1 K also converges strongly to the identity in
H 1 K.
P
Proof. Let A = n=1 n hn kn H 1 K as in Eq. (35.14). Then
kPm Qm A Ak1

n=1

X
n=1

X
n=1

X
n=1

|n |kPm hn Qm kn hn kn k1

|n |k(Pm hn hn ) Qm kn + hn (Qm kn kn )k1

|n |{kPm hn hn kkQm kn k + khn kkQm kn kn k}
|n |{kPm hn hn k + kQm kn kn k} 0 as m

by the dominated convergence theorem.