You are on page 1of 39

# OPERATOR THEORY

Solution I
1. Prove that B(F, Y ) is not a Banach space if Y is not complete. [Hint:
take a Cauchy sequence (y
n
) in Y which does not converge and consider the
sequence of operators (B
n
),
B
n
:= y
n
, F.]
Solution
Take a Cauchy sequence (y
n
) in Y which does not converge and consider the
sequence of operators (B
n
),
B
n
:= y
n
, F.
It is easy to see that B
n
B(F, Y ) and B
n
= y
n
, n N. Since
(B
n
B
m
) = (y
n
y
m
), we have B
n
B
m
= y
n
y
m
. Therefore
(B
n
) is a Cauchy sequence in B(F, Y ). Suppose there exists B B(F, Y )
such that B
n
B 0 as n +. Let y := B1 Y . Then y
n
y =
B
n
1B1 B
n
B 0 as n +, i.e. the sequence (y
n
) converges to
y. This contradiction proves that (B
n
) cannot be convergent. Hence B(F, Y )
is not a Banach space.
2. Give an example of a bounded linear operator A such that Ran(A) is
not closed. [Hint: consider the embedding X Y , where X is the space
C([0, 1]) equipped with the norm

and Y = L
p
([0, 1]) is the completion
of the normed space (C([0, 1]),
p
), 1 p < .]
Solution
Let X be the space C([0, 1]) equipped with the norm

and Y = L
p
([0, 1])
be the completion of the normed space (C([0, 1]),
p
), 1 p < . Let
A : X Y be the embedding: Af = f, f X. Then Ran(A) = C([0, 1])
is dense in Y = L
p
([0, 1]) but does not coincide with it. Therefore Ran(A)
cannot be closed.
1
3. Give an example of a normed space and an absolutely convergent series
in it, which is not convergent.
Solution
Let X be an arbitrary normed non-Banach space (e.g., X = (C([0, 1]),
p
),
1 p < ). Then there exists a Cauchy sequence (x
n
) in X which does
not converge. Since (x
n
) is Cauchy, for any k N there exists n
k
N such
that x
n
x
m
2
k
, n, m n
k
. Consider the series

k=1
(x
n
k+1
x
n
k
) . (1)
This series is absolutely convergent because

k=1
x
n
k+1
x
n
k

k=1
2
k
= 1 < +.
On the other hand, (1) is not convergent in X. Indeed, if the sequence of
partial sums

j
k=1
(x
n
k+1
x
n
k
) = x
n
j+1
x
n
1
is convergent as j +, then
so is the sequence (x
n
j+1
) and also (x
n
), since (x
n
4. Let X be the Banach space C([0, 1]) and Y be the space of all continuously
dierentiable functions on [0, 1] which equal 0 at 0. Both of the spaces are
equipped with the norm

## . Show that the linear operator B : X Y ,

(Bf)(t) :=

t
0
f()d,
is bounded, onetoone and onto, but the inverse operator B
1
: Y X
is not bounded. Compare this with the Banach theorem (bounded inverse
theorem).
Solution
2
It is clear that Bf

## , f C([0, 1]). So, B is bounded.

Suppose Bf
1
= Bf
2
, i.e.

t
0
f
1
()d =

t
0
f
2
()d, t [0, 1].
Then dierentiation gives f
1
(t) = f
2
(t), t [0, 1], i.e. f
1
= f
2
. So, B is
onetoone.
Let be continuously dierentiable and (0) = 0. Then

X and
(t) =

t
0

## ()d, t [0, 1],

i.e. = B

. So, B is onto.
It follows from the above that B is invertible and
B
1
(t) =

## (t), t [0, 1].

It is easy to see that B
1
: Y X is not bounded. Indeed,
B
1
sin(nt)

= ncos(nt)

= n n sin(nt)

, n N.
This does not contradict the Banach theorem (bounded inverse theorem) be-
cause Y equipped with the norm

## is not a Banach space.

5. Denote by C
2
([0, 1]) the space of twice continuously dierentiable func-
tions on the interval [0, 1] equipped with the norm
u := max
0s1
|u(s)| + max
0s1
|u

(s)| + max
0s1
|u

(s)|.
Let X denote the subspace of C
2
([0, 1]) containing functions satisfying bound-
ary conditions u(0) = u(1) = 0. Prove that the operator A =
d
2
ds
2
is a
bounded operator acting from X to C([0, 1]).
Solution
The boundedness of A : X C([0, 1]) follows directly from the denition
of the norm in X. (Note that C([0, 1]) is assumed to be equipped with the
norm

.)
3
6. Show that the operator A from the previous question has a bounded in-
verse A
1
: C([0, 1]) X and construct it eectively.
Solution
Consider the equation Au = f, where f C([0, 1]) is given and u X
is unknown. It is equivalent to
d
2
u(s)
ds
2
= f(s), s [0, 1]. Successive
integration gives
du(s)
ds
=

s
0
f()d + c
1
,
u(s) =

s
0

t
0
f()ddt + c
1
s + c
0
,
where c
1
and c
0
are some constants. Now the boundary conditions u(0) =
u(1) = 0 imply c
0
= 0 and
c
1
=

1
0

t
0
f()ddt.
Hence u = Rf, where
(Rf)(s) :=

s
0

t
0
f()ddt + s

1
0

t
0
f()ddt =
(s 1)

s
0

t
0
f()ddt + s

1
s

t
0
f()ddt.
It is easy to see that Rf X, f C([0, 1]) and
(ARf)(s) =
d
2
ds
2

s
0

t
0
f()ddt + s

1
0

t
0
f()ddt

=
d
ds

s
0
f()d

1
0

t
0
f()ddt

(RAu)(s) =

s
0

t
0
(u

())ddt + s

1
0

t
0
(u

())ddt =

s
0
(u

(t) u

(0))dt s

1
0
(u

(t) u

## (0))dt = u(s) u(0) u

(0)s
su(1) + su(0) + su

## (0) = u(s), s [0, 1], u X,

4
since u(0) = u(1) = 0. Hence, AR = I and RA = I, i.e. A is invertible and
A
1
= R.
The boundedness of A
1
= R : C([0, 1]) X can be obtained directly from
the denition of the norm in X. It also follows immediately from the Banach
theorem on the inverse operator. (Note that the completeness of X follows
from the well known results on uniform convergence.)
5
OPERATOR THEORY
Solution II
1. Let B B(C([0, 1])) be dened by the formula
Bf(t) = tf(t), t [0, 1].
Find (B) and the set of all eigenvalues of B.
Solution
(B) = [0, 1] and B does not have eigenvalues. This is a special case of
Question 2.
2. Let g C([0, 1]) be a xed function and let A B(C([0, 1])) be dened
by the formula
Af(t) = g(t)f(t), t [0, 1].
Find (A) and construct eectively the resolvent R(A; ). Find the eigen-
values and eigenvectors of A.
Solution
Let C, g([0, 1]) := {g(t)| t [0, 1]}. Then since g C([0, 1]),
1/(g ) C([0, 1]) and A I has an inverse R(A; ) = (A I)
1

## B(C([0, 1])) dened by

R(A; )f(t) = (g(t) )
1
f(t), t [0, 1].
Hence (A) g([0, 1]).
Suppose now g([0, 1]), i.e. = g(t
0
) for some t
0
[0, 1]. Then (A
I)f(t
0
) = (g(t
0
) )f(t
0
) = 0, i.e. Ran(A I) consist of functions
vanishing at t
0
. Consequently Ran(A I) = C([0, 1]) and A I is not
invertible. Therefore g([0, 1]) (A). Finally, (A) = g([0, 1]).
Take an arbitrary g([0, 1]). Let g
1
() := { [0, 1] : g() = }.
The equation Af = f, i.e. (g(t) )f(t) = 0 is equivalent to f(t) = 0,
t [0, 1] \ g
1
(). If g
1
() contains an interval of positive length, then it
is easy to see that the set {f C([0, 1]) \{0} : f(t) = 0, t [0, 1] \g
1
()}
is non-empty and coincides with the set of all eigenvectors corresponding to
1
the eigenvalue . If g
1
() does not contain an interval of positive length,
then [0, 1] \ g
1
() is dense in [0, 1] and f(t) = 0, t [0, 1] \ g
1
() implies
by continuity that f 0. In this case is not an eigenvalue.
3. Let K C be an arbitrary nonempty compact set. Construct an operator
B B(l
p
), 1 p , such that (B) = K.
Solution
Let {
k
}
kN
be a dense subset of K. Consider the operator B : l
p
l
p
dened by
Bx = (
1
x
1
,
2
x
2
, . . . ,
k
x
k
, . . . ), x = (x
1
, x
2
, . . . ) l
p
.
Then B is a bounded linear operator and
k
s are its eigenvalues. (Why?)
Consequently {
k
}
kN
(B). Since (B) is closed and {
k
}
kN
is dense in
K,
K (B).
On the other hand, let C\ K. Then d := inf
kN
|
k
| > 0 and B I
has a bounded inverse (B I)
1
: l
p
l
p
dened by
(B I)
1
x =

1

x
1
,
1

2

x
2
, . . . ,
1

k

x
k
, . . .

,
x = (x
1
, x
2
, . . . ) l
p
.
Hence (B). Therefore
(B) K.
Finally,
(B) = K
4. Let k C([0, 1]) be a given function. Consider the operator B
B(C([0, 1])) dened by the formula
(Bu)(s) =

s
0
k(t)u(t)dt.
2
Construct eectively (not as a power series!) the resolvent of A. How does
this resolvent R(B; ) behave when 0?
Solution
It is clear that Bu is continuously dierentiable for any u C([0, 1]). Hence,
Ran(B) = C([0, 1]) and B is not invertible, i.e. 0 (B).
Suppose now = 0. Consider the equation (BI)u = f, f C([0, 1]), i.e.

s
0
k(t)u(t)dt u(s) = f(s), s [0, 1]. (1)
Suppose this equation has a solution u C([0, 1]). Then

s
0
k(t)u(t)dt = f(s) + u(s), s [0, 1].
Therefore f + u is continuously dierentiable and
k(s)u(s) = (f(s) + u(s))

, s [0, 1].
If f is continuously dierentiable, then
k(s)u(s) = f

(s) + u

i.e.
u

(s)
1

k(s)u(s) =
1

## (s), s [0, 1]. (2)

Taking s = 0 in (1) gives
u(0) =
1

f(0).
Solving (2) with this initial condition we obtain
u(s) = e
1

s
0
k()d

f(0)

s
0
1

(t)e

t
0
k()d
dt

=
e
1

s
0
k()d

f(0)
1

f(t)e

t
0
k()d

s
0

s
0
f(t)k(t)e

t
0
k()d
dt

=
1

f(s)
1

s
0
f(t)k(t)e
1

s
t
k()d
dt.
Let
A

f(s) :=
1

f(s)
1

s
0
f(t)k(t)e
1

s
t
k()d
dt. (3)
3
It is easy to see that A

## : C([0, 1]) C([0, 1]) is a bounded linear operator.

The above argument shows that if f is continuously dierentiable and (1)
has a solution u C([0, 1]), then u = A

## f. In particular, (1) with f = 0 has

only a trivial solution u = 0, i.e. Ker(B I) = {0}.
For any f C([0, 1]) and u = A

f the function
f(s) + u(s) =
1

s
0
f(t)k(t)e
1

s
t
k()d
dt
is continuously dierentiable and
(f(s) + u(s))

=
1

f(s)k(s)
k(s)

s
0
f(t)k(t)e
1

s
t
k()d
dt =
k(s)A

f(s) = k(s)u(s)
(see (3)). Hence
f(s) + u(s) =

s
0
k(t)u(t)dt + const.
It follows from (3) that f(0) + u(0) = f(0) + A

f(0) = 0. Therefore
f(s) + u(s) =

s
0
k(t)u(t)dt,
i.e. f = (BI)A

## f, f C([0, 1]), i.e. (BI)A

= I. Consequently A

## is a right inverse of BI and Ran(BI) = C([0, 1]). Since Ker(BI) =

{0}, the operator B I is invertible for any = 0 and (B I)
1
= A

.
Thus
(B) = {0} and R(B; ) = A

, = 0,
where A

is given by (3).
It follows from the well known property of the resolvent that R(B; )
1/||, = 0. So R(B; ) as 0. If k 0, the above inequality
becomes an equality. Let us show that if k 0, then R(B; ) grows much
faster than 1/|| as 0 in a certain direction. Indeed,
(R(B; )1)(s) = (A

1)(s) =
1

s
0
k(t)e
1

s
t
k()d
dt =

+
1

e
1

s
t
k()d

s
0
=
1

e
1

s
0
k()d
.
Further, k 0 implies

s
0
k()d 0. Let
C := max
[0,1]

s
0
k()d

s
0
0
k()d

> 0.
4
Then for such that
1

s
0
0
k()d > 0 we have
R(B; ) R(B; )1
1
||
e
C/||
as 0.
5. Let A, B B(X). Show that for any (A)

(B),
R(B; ) R(A; ) = R(B; )(A B)R(A; ).
Solution
R(B; )(A B)R(A; ) = R(B; )((A I) (B I))R(A; ) =
R(B; ) R(A; ).
5
OPERATOR THEORY
Solution III
1. Let X be a Banach space and A, B B(X).
(a) Show that if I AB is invertible, then I BA is also invertible. [Hint:
consider B(I AB)
1
A + I.]
(b) Prove that if (AB) and = 0, then (BA).
(c) Give an example of operators A and B such that 0 (AB) but 0
(BA).
(d) Show that (AB)

{0} = (BA)

{0}.
(e) Prove that r(AB) = r(BA).
Solution
(a) Suppose I AB is invertible. Then since
(I BA)

B(I AB)
1
A + I

= (I BA)B(I AB)
1
A + (I BA) =
B(I AB)(I AB)
1
A + (I BA) = BA + (I BA) = I,
and

B(I AB)
1
A + I

## (I BA) = B(I AB)

1
A(I BA) + (I BA) =
B(I AB)
1
(I AB)A + (I BA) = BA + (I BA) = I,
the operator I BA is also invertible and (I BA)
1
= B(I AB)
1
A+I.
(b) Take an arbitrary (AB) \{0}. Suppose BAI = (I
1

BA) is
invertible. Then I
1

## BA is invertible and (a) implies that I

1

AB is also
invertible. Therefore (I
1

## AB) = ABI is invertible, i.e. (AB).

This contradiction shows that BAI cannot be invertible, i.e. (BA).
(c) Consider the right and left shift operators R, L : l
p
l
p
, 1 p ,
Rx = (0, x
1
, x
2
, . . . ), Lx = (x
2
, x
3
, . . . ), x = (x
1
, x
2
, . . . ) l
p
.
It is easy to see that
LRx = L(0, x
1
, x
2
, . . . ) = x, RLx = R(x
2
, x
3
, . . . ) = (0, x
2
, x
3
, . . . ), x l
p
.
Hence LR = I, while Ran(RL) = l
p
. Therefore LR is invertible while RL is
not, i.e. 0 (RL) but 0 (LR).
1
(d) Follows from (b).
(e) Follows from (d) and the denition of the spectral radius.
2. Let X be a Banach space and let operators A, B B(X) commute:
AB = BA. Prove that r(A + B) r(A) + r(B).
Solution
Take an arbitrary > 0. The spectral radius formula implies that A
n

(r(A) +)
n
, B
n
(r(B) +)
n
for suciently large n N. Therefore there
exists M 1 such that
A
n
M(r(A) + )
n
, B
n
M(r(B) + )
n
, n N.
Since A and B commute, we have
(A + B)
n
=
n

k=0
n!
k!(n k)!
A
nk
B
k
.
Hence,
(A + B)
n

n

k=0
n!
k!(n k)!

A
nk

B
k

M
2
n

k=0
n!
k!(n k)!
(r(A) + )
nk
(r(B) + )
k
= M
2
(r(A) + r(B) + 2)
n
for any n N. Consequently
r(A + B) = lim
n+
(A + B)
n

1/n
r(A) + r(B) + 2, > 0,
i.e. r(A + B) r(A) + r(B).
3. Let k C([0, 1] [0, 1]) be a given function. Consider the operator
B B(C([0, 1])) dened by the formula
(Bu)(s) =

s
0
k(s, t)u(t)dt.
2
Find the spectral radius of B. What is the spectrum of B? [Hint: prove by
induction that
|(B
n
u)(s)|
M
n
n!
s
n
u

, n N,
for some constant M > 0.]
Solution
Let us prove by induction that
|(B
n
u)(s)|
M
n
n!
s
n
u

## , s [0, 1], n N, (1)

where
M := max
(s,t)[0,1]
2
|k(s, t)|.
For n = 0 inequality (1) is trivial. Suppose (1) holds for n = k. Then for
n = k + 1 we have

B
k+1
u

(s)

s
0
k(s, t)(B
k
u)(t)dt

s
0
|k(s, t)|

(B
k
u)(t)

dt
M

s
0

(B
k
u)(t)

dt M

s
0
M
k
k!
t
k
u

dt =
M
k+1
k!
u

s
0
t
k
dt =
M
k+1
(k + 1)!
s
k+1
u

, s [0, 1].
Hence, (1) is proved by induction.
It follows from (1) that
B
n
u

M
n
n!
u

, u C([0, 1]),
i.e.
B
n

M
n
n!
, n N.
Therefore
r(B) = lim
n+
B
n

1/n
lim
n+
M
(n!)
1/n
= 0.
Since r(B) = 0, (B) cannot contain nonzero elements. Taking into account
that (B) is nonempty we conclude (B) = {0}.
3
OPERATOR THEORY
Solution IV
1. Suppose X is a Banach space and let A B(X) be nilpotent, i.e. A
n
= 0
for some n N. Find (A).
Solution
The Spectral Mapping Theorem implies
{
n
: (A)} = (A
n
) = (0) = {0}.
Therefore, (A)
n
= 0 = 0, i.e. (A) = {0}.
2. Show that the range of the operator B : l
p
l
p
, 1 p < ,
(Bx)
n
:=
1
1 +n
2
x
n
, n N, x = (x
1
, x
2
, . . . ),
is not closed.
Solution
Take an arbitrary y = (y
1
, y
2
, . . . ) l
p
. Let y
(N)
= (y
(N)
1
, y
(N)
2
, . . . ),
y
(N)
n
:=

y
n
if n N,
0 if n > N,
N N.
Then y y
(N)
0 as N + and the vectors x
(N)
dened by x
(N)
n
:=
(1 + n
2
)y
(N)
n
belong to l
p
. (Why?) It is clear that y
(N)
= Bx
(N)
Ran(B).
So, Ran(B) is dense in l
p
. On the other hand Ran(B) does not coincide with
l
p
. Indeed, consider y l
p
, y
n
=
1
1+n
2
, n N. Suppose y Ran(B). Then
there exists x l
p
such that Bx = y, i.e.
1
1 +n
2
x
n
=
1
1 +n
2
, n N,
i.e. x = (1, 1, . . . ) l
p
. This contradiction shows that y Ran(B). Hence,
Ran(B) = l
p
. Since Ran(B) is dense in l
p
, it cannot be closed.
3. Let P B(X) be a projection, i.e. P
2
= P. Construct R(P; ).
1
Solution
If P = 0, then obviously P I = I, (P) = {0} and R(P; ) =
1
I.
If P = I, then P I = (1 )I, (P) = {1} and R(P; ) = (1 )
1
I.
Suppose now P is non-trivial, i.e. P = 0, I. Take any = 0, 1. Then using
the equalities P
2
= P, Q
2
= Q and QP = PQ = 0, where Q = I P, we
obtain

(1 )
1
P
1
Q

(P I) =

(1 )
1
P
1
Q

((1 )P Q) =
P +Q = I
and similarly
(P I)

(1 )
1
P
1
Q

= I.
Thus
R(P; ) = (1 )
1
P
1
Q.
4. Let A
1
l
B(Y, X) be a left inverse of A B(X, Y ), i.e. A
1
l
A = I
X
.
Find (AA
1
l
).
Let B
1
r
B(Y, X) be a right inverse of B B(X, Y ), i.e. BB
1
r
= I
Y
. Find
(B
1
r
B).
Solution
It is easy to see that AA
1
l
is a projection. Indeed,
(AA
1
l
)
2
= AA
1
l
AA
1
l
= AI
X
A
1
l
= AA
1
l
.
It is also clear that AA
1
l
= 0. Indeed, if AA
1
l
= 0 then A
1
l
(AA
1
l
)A =
A
1
l
0A = 0, i.e. I
X
= 0. Contradiction! (We assume of course that
X = {0}.) So, AA
1
l
is a nonzero projection. Hence (AA
1
l
) = {0, 1}
if AA
1
l
= I, i.e. if A is not invertible. If A is invertible, i.e. AA
1
l
= I, then
obviously (AA
1
l
) = {1}.
Similarly
(B
1
r
B) =

## {0, 1} if B is not invertible,

{1} if B is invertible.
2
OPERATOR THEORY
Solution V
1. Let B B(X) and let T B(Y, X) be invertible: T
1
B(X, Y ). Prove
that
(B) = (T
1
BT).
Solution
Since T is invertible,
(B) B I is invertible T
1
(B I)T is invertible
T
1
BT I is invertible (T
1
BT),
i.e. (B) = (T
1
BT).
2. Consider the right-shift operator R : l

dened by
Rx = (0, x
1
, x
2
, . . . ), x = (x
1
, x
2
, . . . ) l

.
Find the eigenvalues and the spectrum of this operator. Is this operator
compact?
Solution
It is clear that Ker(R) = {0}, i.e. 0 is not an eigenvalue of R. Suppose = 0
is an eigenvalue of R. Then Rx = x for some non-zero x. So,
0 = x
1
x
1
= x
2
x
2
= x
3

Solving the last system we obtain x = 0. Contradiction! Thus R does not
have eigenvalues.
1
It is clear that Rx = x for any x l

## . Therefore R = 1 and (R)

{ C : || 1}. Let us take an arbitrary C such that 0 < || < 1.
Suppose (R). Then the equation
(R I)x = y (1)
has a unique solution x l

for any y l

## . For y = (1, 0, 0, . . . ), (1) takes

the form
x
1
= 1
x
1
x
2
= 0
x
2
x
3
= 0

Solving the last system we obtain x
k
=
k
. Since || < 1, the element
x = (x
1
, x
2
, . . . ) does not belong to l

## . The obtained contradiction shows

that (R). Hence, { C : 0 < || < 1} (R). Taking into account
that (R) is closed we obtain
(R) = { C : || 1}.
The operator R is not compact. This follows from the fact that R is an isom-
etry or from the fact that its spectrum cannot be the spectrum of a compact
operator.
3. Consider the set
M = {x l

: |x
n
| n

, n N} l

,
where > 0 is a xed number. Prove that M is compact.
Solution
It is clear that M is a closed set. Take an arbitrary > 0. There exists N N
such that n

## < , n > N. Let x

(l)
= (x
(l)
1
, . . . , x
(l)
N
) C
N
, l = 1, . . . , L be
a nite net of the set
M
N
:= { x = (x
1
, . . . , x
N
) C
N
: |x
n
| n

, n = 1, . . . , N}.
Such an net can be easily constructed with the help of nets of the disks
{x
n
C : |x
n
| n

## }. (The existence of such an net also follows

2
from the fact that M
N
is relatively compact as a bounded subset of the
nitedimensional space C
N
.) Now it is easy to see that
x
(l)
:= (x
(l)
1
, . . . , x
(l)
N
, 0, 0, . . . ), l = 1, . . . , L
is an net of M. Hence M is a closed relatively compact set, i.e. a compact
set.
(An alternative proof: it is clear that M = T(S

), where T B(l

),
Tx := (x
1
, 2

x
2
, 3

x
3
, . . . , n

x
n
, . . . ), x = (x
1
, x
2
, . . . ) l

and S

## is the unit ball of l

:
S

= {x l

: |x
n
| 1, n N}.
It is easy to see that T T
N
0 as N +, where T
N
is a nite rank
operator dened by
T
N
x := (x
1
, 2

x
2
, . . . , N

x
N
, 0, 0 . . . ), x = (x
1
, x
2
, . . . ) l

.
Hence T is a compact operator and M = T(S

## ) is a closed relatively com-

pact set, i.e. a compact set.)
4. Let g C([0, 1]) be a xed function. Consider the operator A
B(C([0, 1])) dened by the formula
(Au)(s) := g(s)u(s),
i.e. the operator of multiplication by g. Is this operator compact?
Solution
It is clear that if g 0 then A is compact. Let us prove that t if g 0 then
A is not compact. Indeed, since g 0, there exists a subinterval [a, b] [0, 1]
such that m := min
s[a,b]
|g(s)| > 0. Consider the sequence u
n
C([0, 1]),
n N, u
n
(s) := sin(2
n sa
ba
), s [0, 1]. It is clear that (u
n
) is a bounded
sequence. On the other hand (Au
n
) does not have Cauchy subsequences.
Indeed, take arbitrary k, n N. Assume for deniteness that k > n, i.e.
k n + 1. Let s
n
:= a + 2
(n+1)
(b a). Then s
n
[a, b] and
Au
k
Au
n
= max
s[0,1]
|g(s)(u
k
(s) u
n
(s))| m max
s[a,b]
|u
k
(s) u
n
(s)|
m|u
k
(s
n
) u
n
(s
n
)| = m| sin(2
kn1
) sin(/2)| = m|0 1| = m > 0.
3
Since (Au
n
) does not have Cauchy subsequences, A is not compact.
(An alternative proof: according to the solution of Exercise 2, Sheet II,
(A) = g([0, 1]). If g 0 is a constant, then 0 g([0, 1]) and (A) cannot
be the spectrum of a compact operator. If g is nonconstant, then g([0, 1]) is
a connected subset of C consisting of more than one point and (A) cannot
be the spectrum of a compact operator.)
5. Let X be an innite-dimensional Banach space and B, T B(X). Which
of the following statements are true?
(i) If BT is compact then either B or T is compact.
(ii) If T
2
= 0 then T is compact.
(iii) If T
n
= I for some n N then T is not compact.
Solution
(i) is false. This follows from the fact that (ii) is false.
(ii) is false. Indeed, let X = l
p
, 1 p + and
Tx = (0, x
1
, 0, x
3
, 0, x
5
, 0, . . . ), x = (x
1
, x
2
, x
3
, . . . ) l
p
.
Then T
2
= 0 and it is easy to see that T is not compact (why?).
(iii) is true. Indeed, suppose T is compact. Then I = T
n
is also compact,
which is impossible, since X is innite-dimensional. Contradiction!
4
OPERATOR THEORY
Solution VI
1. Prove that the norm
p
on l
p
, p = 2, is not induced by an inner product.
(Hint: Prove that for x = (1, 1, 0, . . . ) l
p
and y = (1, 1, 0, . . . ) l
p
the
parallelogram law fails.)
Solution
It is easy to see that
x + y
p
= 2 = x y
p
, x
p
= 2
1/p
= y
p
.
Hence the parallelogram law
x + y
2
+x y
2
= 2(x
2
+y
2
)
is equivalent to 8 = 4 2
2/p
, i.e. to p = 2. So, if p = 2, the parallelogram law
fails, i.e. the norm
p
is not induced by an inner product.
2. Prove that the norm
p
, p = 2 on C([0, 1]) is not induced by an inner
product. (Hint: Prove that for functions f(t) = 1/2 t and
g(t) =
_
1/2 t if 0 t 1/2 ,
t 1/2 if 1/2 < t 1 ,
the parallelogram law fails).
Solution
We have
f(t) + g(t) =
_
1 2t if 0 t 1/2 ,
0 if 1/2 < t 1 ,
f(t) g(t) =
_
0 if 0 t 1/2 ,
1 2t if 1/2 < t 1 ,
1
f + g
p
=
_
_
1/2
0
(1 2t)
p
dt
_
1/p
=
1
(2(p + 1))
1/p
,
f g
p
=
__
1
1/2
|1 2t|
p
dt
_
1/p
=
__
1
1/2
(2t 1)
p
dt
_
1/p
=
1
(2(p + 1))
1/p
,
f
p
=
__
1
0
|1/2 t|
p
dt
_
1/p
=
_
_
1/2
0
(1/2 t)
p
dt +
_
1
1/2
(t 1/2)
p
dt
_
1/p
=
_
2
1
(p + 1)2
p+1
_
1/p
=
1
2(p + 1)
1/p
= g
p
.
Hence the parallelogram law is equivalent to 2 2
2/p
= 1, i.e. to p = 2. So,
if p = 2, the parallelogram law fails, i.e. the norm
p
is not induced by an
inner product.
3. Let {e
n
}
nN
be an orthonormal set in an inner product space H. Prove
that

n=1
|(x, e
n
)(y, e
n
)| xy, x, y H.
Solution
Using the CauchySchwarz inequality for l
2
and Bessels inequality for H we
obtain

n=1
|(x, e
n
)(y, e
n
)|
_

n=1
|(x, e
n
)|
2
_
1/2
_

n=1
|(y, e
n
)|
2
_
1/2
xy.
2
4. Prove that in a complex inner product space H the following equalities
hold:
(x, y) =
1
N
N

k=1
x + e
2ik/N
y
2
e
2ik/N
for N 3,
(x, y) =
1
2
_
2
0
x + e
i
y
2
e
i
d, x, y H.
Solution
1
N
N

k=1
x + e
2ik/N
y
2
e
2ik/N
=
1
N
N

k=1
(x + e
2ik/N
y, x + e
2ik/N
y)e
2ik/N
=
1
N
N

k=1
_
x
2
e
2ik/N
+ (y, x)e
4ik/N
+ (x, y) +y
2
e
2ik/N
_
=
1
N
x
2
e
2i(N+1)/N
e
2i/N
e
2i/N
1
+
1
N
(y, x)
e
4i(N+1)/N
e
4i/N
e
4i/N
1
+ (x, y) +
1
N
y
2
e
2i(N+1)/N
e
2i/N
e
2i/N
1
= (x, y),
since e
2i/N
= 1 and e
4i/N
= 1 for N 3.
Similarly
1
2
_
2
0
x + e
i
y
2
e
i
d =
1
2
_
2
0
(x + e
i
y, x + e
i
y)e
i
d =
1
2
x
2
_
2
0
e
i
d +
1
2
(y, x)
_
2
0
e
2i
d +
1
2
(x, y)
_
2
0
1d +
1
2
y
2
_
2
0
e
i
d = (x, y).
5. Show that A

## = spanA for any subset of a Hilbert space.

Solution
3
Claim I: For any subset M of a Hilbert space H the orthogonal complement
M

## is a closed linear subspace of H (see Proposition 3.15(i)). Indeed, if

z
1
, z
2
M

, then
(z
1
+ z
2
, y) = (z
1
, y) + (z
2
, y) = 0 + 0 = 0, y M, , F
and hence, z
1
+ z
2
M

. So, M

is a linear subspace of H.
Suppose z Cl(M

## ). Then there exist z

n
M

, n N such that z
n
z
as n +. Therefore
(z, y) = lim
n+
(z
n
, y) = lim
n+
0 = 0, y M,
i.e. z M

. Thus M

Claim II: A A

## (see Proposition 3.17(ii)). Indeed, for any x A we have

(x, y) = (y, x) = 0, y A

,
i.e. x A

.
Since A

= (A

## is a closed linear subspace of H, we obtain

spanA A

.
Now, take any x A

## . Since H = spanA (spanA)

, we have
x = z + y, z spanA, y (spanA)

,
and therefore, (x, y) = (z, y) + y
2
. Since y A

, we obtain 0 = y
2
, i.e.
y = 0, i.e. x = z, i.e. x spanA. Consequently A

spanA. Finally,
A

= spanA.
6. Let M and N be closed subspaces of a Hilbert space. Show that
(M + N)

= M

, (M N)

= Cl(M

+ N

).
Solution
4
Let z (M + N)

, i.e.
(z, x + y) = 0, x M, y N. (1)
Taking x = 0 or y = 0 we obtain
(z, x) = 0, x M, (z, y) = 0, y N, (2)
i.e. z M

. Hence (M + N)

.
Suppose now z M

## , i.e. (2) holds. Then obviously, (1) holds, i.e

z (M + N)

. Therefore M

(M + N)

. Thus
(M + N)

= M

. (3)
Writing (3) for M

and N

## instead of M and N and using Exercise 5, we

obtain
(M

+ N

= M

= M N,
since M and N are closed linear subspaces. Taking the orthogonal com-
plements of the LHS and the RHS and using Exercise 5 again, we arrive
at
Cl(M

+ N

) = (M N)

,
since M

+ N

is a linear subspace.
7. Show that M := {x = (x
n
) l
2
: x
2n
= 0, n N} is a closed subspace
of l
2
. Find M

.
Solution
Take any x, y M. It is clear that for any , F
(x + y)
2n
= x
2n
+ y
2n
= 0,
i.e. x + y M. Hence M is a linear subspace of l
2
. Let us prove that it
is closed.
Take any x Cl(M). There exist x
(k)
M such that x
(k)
x as k +.
Since x
(k)
2n
= 0, we obtain
x
2n
= lim
k+
x
(k)
2n
= 0,
5
i.e. x M. Hence M is closed.
Further,
z M

(z, x) = 0, x M

n=0
z
2n+1
x
2n+1
= 0 for all x
2n+1
F, n N
such that

n=0
|x
2n+1
|
2
< +
z
2n+1
= 0, n = 0, 1, . . .
Therefore
M

= {z = (z
n
) l
2
: z
2n+1
= 0, n = 0, 1, . . . }.
8. Show that vectors x
1
, . . . , x
N
in an inner product space H are linearly
independent i their Gram matrix (a
jk
)
N
j,k=1
= ((x
k
, x
j
))
N
j,k=1
is nonsingular,
i.e. i the corresponding Gram determinant det((x
k
, x
j
)) does not equal zero.
Take an arbitrary x H and set b
j
= (x, x
j
). Show that, whether or not x
j
are linearly independent, the system of equations
N

k=1
a
jk
c
k
= b
j
, j = 1, . . . , N,
is solvable and that for any solution (c
1
, . . . , c
N
) the vector

N
j=1
c
j
x
j
is the
nearest to x point of lin{x
1
, . . . , x
N
}.
Solution
Let c
1
, . . . , c
N
F. It is clear that
N

k=1
c
k
x
k
= 0
_
N

k=1
c
k
x
k
, y
_
= 0, y lin{x
1
, . . . , x
N
}
_
N

k=1
c
k
x
k
, x
j
_
= 0, j = 1, . . . , N
N

k=1
a
jk
c
k
= 0, j = 1, . . . , N.
6
Therefore
the vectors x
1
, . . . , x
N
are linearly independent
N

k=1
c
k
x
k
= 0 i c
1
= = c
N
= 0
the system
N

k=1
a
jk
c
k
= 0, j = 1, . . . , N has only
the trivial solution c
1
= = c
N
= 0
det((x
k
, x
j
)) = det(a
jk
) = 0.
For any x H and b
j
= (x, x
j
), j = 1, . . . , N we have
the vector
N

k=1
c
k
x
k
is the nearest to x point of
lin{x
1
, . . . , x
N
}
x
N

k=1
c
k
x
k
lin{x
1
, . . . , x
N
}

_
x
N

k=1
c
k
x
k
, x
j
_
= 0, j = 1, . . . , N
N

k=1
a
jk
c
k
= b
j
, j = 1, . . . , N.
7
CM414Z/CMMS08 OPERATOR THEORY
Solutions to 2007 Exam
Each question will be marked out of 25. The tentative marks for parts
are indicated in square brackets.
1. Let X be a Banach space and B B(X), i.e. let B be a bounded
linear operator acting on X. Dene the kernel Ker(B) and the range
Ran(B) of B.
Suppose there exists a constant c > 0 such that
Bx cx, x X. (1)
Show that Ker(B) = {0} and Ran(B) is closed.
Suppose A B(X) is left invertible and C B(X) is right invertible,
i.e. there exist A
1
l
, C
1
r
B(X) such that A
1
l
A = I and CC
1
r
= I,
where I is the identity operator. Prove that Ran(A) and Ran(C) are
closed. Show that Ker(A) = {0}. Give an example of a right invertible
C B(X) such that Ker(C) = {0}.
Explain what is meant by saying that P B(X) is a projection.
Let A and C be as above. Prove that AA
1
l
and C
1
r
C are projections.
Solution
Ker(B) := {x X| Bx = 0}, Ran(B) := {Bx| x X}.
Bookwork. [2]
Suppose (1) holds with a positive constant c. If x Ker(B), then it
follows from (1) that x = 0, i.e. x = 0. Hence Ker(B) = {0}.
Let us now prove that Ran(B) is closed. For any y Cl(Ran(B)) there
exist x
n
X such that Bx
n
y as n . Then
x
n
x
m
c
1
Bx
n
Bx
m
0 as n, m .
Hence (x
n
) is a Cauchy sequence in the Banach space X. Let us denote
its limit by x. We have
Bx = B
_
lim
n
x
n
_
= lim
n
Bx
n
= y.
1
Thus y Ran(B), i.e. Ran(B) is closed.
Bookwork. [7]
Since
x = A
1
l
Ax A
1
l
Ax, x X,
A satises (1) with c = 1/A
1
l
. Hence Ran(A) is closed.
Unseen. [3]
Since
x = CC
1
r
x Ran(C), x X,
Ran(C) = X. Hence Ran(C) is clearly closed.
Unseen. [3]
Suppose x Ker(A). Then
x = A
1
l
Ax = A
1
l
(Ax) = A
1
l
0 = 0.
So, Ker(A) = {0}.
Unseen. [3]
Let X = l
2
and C be the left shift operator:
Cx = (x
2
, x
3
, . . . ), x = (x
1
, x
2
, . . . ) l
2
.
It is easy to see that C is right invertible with a right inverse
C
1
r
x = (0, x
1
, x
2
, x
3
, . . . ), x = (x
1
, x
2
, . . . ) l
2
,
and
Ker(C) = {(, 0, 0, 0, . . . )| C} = {0}.
Unseen in this form. [3]
P is called a projection if P
2
= P.
Bookwork. [1]
(AA
1
l
)
2
= AA
1
l
AA
1
l
= AIA
1
l
= AA
1
l
(C
1
r
C)
2
= C
1
r
CC
1
r
C = C
1
r
IC = C
1
r
C.
Homework. [3]
2
2. Let B be a bounded linear operator acting in a Banach space X. For
the operator B dene the following notions: norm, eigenvalue, eigenvec-
tor, resolvent set, spectrum, resolvent.
Let C([, ]) be the Banach space of all continuous complex valued
functions on [, ] equipped with the standard norm
f = max
t[,]
|f(t)|.
Consider the operator A : C([, ]) C([, ]) dened by
(Af)(t) = (sin t)f(t), t , f C([, ]).
Find A.
Find the spectrum (A) and eigenvalues of A.
Consider the bounded linear operator V : C([, ]) C([, ]) de-
ned by
(V f)(t) = f(t), t [, ], f C([, ]).
Find the spectrum of AV + V A.
Solution
If B is bounded we dene its norm by the equality
B := inf{C : Bx Cx, x X}.
(Any of the following equivalent denitions will be accepted as a correct one:
B = inf{C : Bx C, all x s.t. x 1}
= inf{C : Bx C, all x s.t. x = 1} = sup
_
Bx
x
: x = 0
_
= sup{Bx : x 1} = sup{Bx : x = 1}. )
A number C is called an eigenvalue of B if there exists an x X,
x = 0, called an eigenvector of B, such that Bx = x. The resolvent
set (B) is the set of all C for which there exists a bounded lin-
ear operator R(B; ), called the resolvent, such that R(B; )(B I) =
(B I)R(B; ) = I, i.e. R(B; ) = (B I)
1
. The spectrum (B) is
dened as (B) = C \ (B), i.e. (B) is the set of all C such that
B I is not invertible on X.
Bookwork. [4]
It is clear that
Af = max
t[,]
|(sin t)f(t)| max
t[,]
|f(t)| = f, f C([, ]).
3
So, A 1. On the other hand, we have for f 1
Af = max
t[,]
|sin t| = 1 = f.
Hence A 1. Finally, A = 1.
Unseen. [4]
Let C, [1, 1] = {sin t| t [, ]}. Then 1/(sin t )
C([, ]) and AI has an inverse R(A; ) = (AI)
1
B(C([, ]))
dened by
R(A; )g(t) = (sin t )
1
g(t), t [, ].
Hence (A) [1, 1].
Unseen but similar to problems from homework and from 2006 paper. [4]
Suppose now [1, 1], i.e. = sin t

for t

Then (A I)f(t

) = (sin t

)f(t

## ) = 0, i.e. Ran(A I) consists

of functions vanishing at t

## . Consequently Ran(A I) = C([, ])

and A I is not invertible. Therefore [1, 1] (A). Finally, (A) =
[1, 1].
Unseen but similar to problems from homework and from 2006 paper. [4]
Let [1, 1]. The equation Af = f, i.e. (sin t )f(t) 0 is equiv-
alent to f(t) = 0, t [, ] \ {t

}, where

t

## is the second solution

of sin t = , t [, ]. So, f 0 by continuity. Hence cannot be an
eigenvalue of A, i.e. A has no eigenvalues.
Unseen but similar to problems from homework and from 2006 paper. [4]
Since
(AV + V A)f(t) = (sin t)f(t) + (sin(t))f(t)
= (sin t)f(t) (sin t)f(t) = 0,
t [, ], f C([, ]),
AV + V A = 0 and (AV + V A) = {0}.
Unseen. [5]
4
3. Let X and Y be normed spaces. Explain what is meant by saying
that T : X Y is a compact operator.
Let X be a Banach space and K : X X be a compact operator.
Suppose that = 0 belongs to the spectrum of K and , C belong
to the resolvent set of K.
Show that the operator
A := (K I)
1
(K I)
1
is compact. (Here and below, you may use without justication any
textbook material on compact operators.)
Show that ( )
1
is an eigenvalue of (K I)
1
.
Show that
:=
( )
( )( )
is an eigenvalue of nite multiplicity of the operator A if = .
Solution
Let X and Y be normed spaces. A linear operator T : X Y is called
compact if it maps bounded sets of X into relatively compact sets of Y ,
i.e. if for any bounded sequence z
n
X, n N, the sequence (Tz
n
)
nN
has a Cauchy subsequence.
Bookwork. [2]
Let X be a Banach space and K : X X be compact.
Suppose that , C belong to the resolvent set of K. Then
(K I)
1
(K I)
1
= (K I)
1
((K I) (K I))(K I)
1
= ( )(K I)
1
K(K I)
1
and this operator is compact as a product of the compact operator T and
bounded operators (K I)
1
and (K I)
1
.
Unseen. [7]
Since = 0 belongs to the spectrum of K and K is compact, is an eigen-
value K. Let x X \ {0} be a corresponding eigenvector: Kx = x.
Then (K I)x = ( )x. Hence ( )
1
x = (K I)
1
x, i.e.
( )
1
is an eigenvalue of (K I)
1
.
Unseen. [7]
Similarly ( )
1
x = (K I)
1
x. Therefore
Ax = (K I)
1
x (K I)
1
x = ( )
1
x ( )
1
x
=
_

_
x =
( )
( )( )
x = x.
5
Hence is an eigenvalue of A.
Unseen. [7]
Since = 0, this is a nonzero eigenvalue if = . Since A is a compact
operator, has a nite multiplicity.
Unseen. [2]
6
4. Let H be a Hilbert space.
Show that vectors x
1
, . . . , x
N
H are linearly independent if and only if
their Gram matrix (a
jk
)
N
j,k=1
= ((x
k
, x
j
))
N
j,k=1
is nonsingular, i.e. if and
only if the corresponding Gram determinant det((x
k
, x
j
)) does not equal
zero.
Dene the orthogonal complement of a set M H.
Let H = l
2
. Show that L := {x = (x
n
)
nN
l
2
| x
2k1
= x
2k
, k N} is a
closed subspace of l
2
. Find the orthogonal complement of L.
Solution
Let c
1
, . . . , c
N
F. It is clear that
N

k=1
c
k
x
k
= 0
_
N

k=1
c
k
x
k
, y
_
= 0, y lin{x
1
, . . . , x
N
}
_
N

k=1
c
k
x
k
, x
j
_
= 0, j = 1, . . . , N
N

k=1
a
jk
c
k
= 0, j = 1, . . . , N.
Therefore
the vectors x
1
, . . . , x
N
are linearly independent
N

k=1
c
k
x
k
= 0 i c
1
= = c
N
= 0
the system
N

k=1
a
jk
c
k
= 0, j = 1, . . . , N has only
the trivial solution c
1
= = c
N
= 0
det((x
k
, x
j
)) = det(a
jk
) = 0.
Homework. [10]
The orthogonal complement M

## of a set M H is the set

M

:= {z H : (z, x) = 0, x M}.
Bookwork. [1]
Take any x, y L l
2
. It is clear that for any scalars and ,
(x + y)
2k1
= x
2k1
+ y
2k1
= x
2k
+ y
2k
= (x + y)
2k
, k N,
i.e. x + y L. Hence L is a linear subspace of l
2
. Let us prove that
it is closed.
7
Take any x Cl(L). There exist x
(j)
L such that x
(j)
x as j +.
Since x
(j)
2k1
= x
(j)
2k
, we obtain
x
2k1
= lim
j+
x
(j)
2k1
= lim
j+
x
(j)
2k
= x
2k
, k N,
i.e. x L. Hence L is closed.
Unseen. [7]
Further,
z L

(z, x) = 0, x L

k=1
(z
2k1
+ z
2k
)x
2k
= 0 for all x
2k
F, k N
such that

k=1
|x
2k
|
2
< +
z
2k1
= z
2k
, k N.
Therefore
M

= {z = (z
n
)
nN
l
2
| z
2k1
= z
2k
, k N}.
Unseen. [7]
8
5. Let H be a complex Hilbert space.
Let B : H H be a bounded linear operator. Explain what is meant
by saying that
i) B is self-adjoint, ii) B is unitary, iii) B is normal.
Show that for any bounded linear operator A : H H there exist unique
self-adjoint operators B and C such that A = B + iC. Prove that A is
normal if and only if B and C commute: BC = CB.
Dene the numerical range Num(A) of A.
Consider the operator A : l
2
l
2
dened by
Ax = (x
1
,
x
2
2
,
x
3
3
, . . . , (1)
k
x
k
k
, . . . ), x = (x
1
, x
2
, . . . ) l
2
.
Find Num(A).
Solution
The operator B is called

= B, i.e. if
(Bx, y) = (x, By), x, y H;
(ii) unitary if B

B = BB

= I, i.e. if B
1
= B

;
(iii) normal if BB

= B

B.
Bookwork. [2]
Suppose there exist self-adjoint operators B and C such that A = B+iC.
Then A

= B iC and we have
B =
1
2
(A + A

), C =
1
2i
(A A

). (2)
It is clear that the above operators are self-adjoint and A = B + iC.
Homework and 2005 paper. [4]
Suppose B and C commute. Then
AA

= (B + iC)(B iC) = B
2
+ iCB iBC + C
2
= B
2
+ iBC iCB + C
2
= (B iC)(B + iC) = A

A,
i.e. A is normal.
Unseen. [4]
Suppose now A is normal. Then (B +iC)(B iC) = (B iC)(B +iC),
i.e.
B
2
+ iCB iBC + C
2
= B
2
+ iBC iCB + C
2
,
i.e. CB BC = BC CB, i.e. CB = BC.
(An alternative way of proving this is to use (2).)
Unseen. [4]
9
Num(A) := {(Ax, x) : x = 1, x H}.
Bookwork. [1]
Take an arbitrary x l
2
such that x
2
= 1. Then
(Ax, x) =

k=1
(1)
k
k
|x
k
|
2

j=1
1
2j
|x
2j
|
2

1
2

j=1
|x
2j
|
2

1
2

k=1
|x
k
|
2
=
1
2
,
(Ax, x) =

k=1
(1)
k
k
|x
k
|
2

j=1
1
2j 1
|x
2j1
|
2

j=1
|x
2j1
|
2

k=1
|x
k
|
2
= 1.
Hence Num(A) [1,
1
2
].
Unseen. [5]
Take an arbitrary [1,
1
2
]. Let t =
1
3
(1 2). Then 0 t 1
and t +
1
2
(1 t) = . Let x = (

t,

1 t, 0, 0, . . . ) l
2
. Then
x
2
= t + 1 t = 1 and
(Ax, x) = t +
1
2
(1 t) = .
Therefore Num(A), [1,
1
2
], i.e. [1,
1
2
] Num(A). Finally,
Num(A) = [1,
1
2
].
Unseen. [5]
10
6. Let H be a complex Hilbert space. State the Hilbert-Schmidt theorem
for a compact self-adjoint linear operator on H.
Let T : H H be a compact self-adjoint linear operator.
Use the Hilbert-Schmidt theorem to prove that there exists a compact
linear operator S : H H such that S
2
= T.
Suppose additionally that T is not a nite rank operator. Which of the
following statements is correct?
(i) There is a unique S : H H such that S
2
= T.
(ii) There are exactly two operators S : H H such that S
2
= T.
(iii) There are innitely many operators S : H H such that S
2
= T.
Suppose H is separable. Explain what it means to say that K : H H
is a Hilbert-Schmidt operator.
Give an example of a non-Hilbert-Schmidt operator S : H H such
that T = S
2
is Hilbert-Schmidt.
Solution
The Hilbert-Schmidt theorem: Let T : H H be a compact self-adjoint
operator. Then there exists a nite or a countable orthonormal set
{e
n
}
N
n=1
, N N {}, of eigenvectors of T such that any x H
has a unique representation of the form
x =
N

n=1
c
n
e
n
+ y, y Ker(T), c
n
C. (3)
One then has
Tx =
N

n=1

n
c
n
e
n
, (4)
where
n
= 0 is the eigenvalue of T corresponding to the eigenvector e
n
.
Moreover,
(T) \ {0} = {
n
}
N
n=1
R, |
1
| |
2
|
and
lim
n

n
= 0 if N = .
Bookwork. [4]
Consider the operator
Sx :=
N

n=1
_

n
c
n
e
n
11
(see (3)). It is clear that S
2
= T (see (4)). If N < +then S is obviously
a bounded nite rank operator. Hence S is compact. If N = +,
consider the bounded nite rank operators
S
j
x :=
j

n=1
_

n
c
n
e
n
, j N.
We have
(S S
j
)x
2
=
_
_
_
_
_

n=j+1
_

n
c
n
e
n
_
_
_
_
_
2
=

n=j+1
|
n
||c
n
|
2
sup
nj+1
|
n
|

n=j+1
|c
n
|
2
= |
j+1
|

n=j+1
|c
n
|
2
|
j+1
| x
2
.
Hence
S S
j

_
|
j+1
| 0 as j +,
and S is compact as the norm limit of a sequence of compact (nite rank)
operators.
Unseen. [7]
If T is not a nite rank operator, then N = + in (3), (4). It is clear
that for any choice of + or for each n, the operator

Sx :=

n=1

n
c
n
e
n
satises the equality

S
2
= T. Since the set of such operators is innite
(in fact, has cardinality of the continuum), there are innitely many op-
erators S : H H such that S
2
= T.
Unseen. [6]
Let Hbe separable and let {e
n
}
nN
be an arbitrary complete orthonormal
set in H. K is called a Hilbert-Schmidt operator if

n=1
Ke
n

2
< .
Bookwork. [2]
Let
Sx :=
N

n=1
1

n
(x, e
n
) e
n
.
Then
T = S
2
=
N

n=1
1
n
(x, e
n
) e
n
,
12
and

n=1
Se
n

2
=

n=1
1
n
= ,

n=1
Te
n

2
=

n=1
1
n
2
< .
So, T is Hilbert-Schmidt and S is not.
Unseen. [6]
13