Professional Documents
Culture Documents
August 1, 2010
2010
c by Christopher Heil
Detailed Solutions to Exercises
so {xn } is bounded.
(c) Given x, y ∈ H, we have
This implies that {ck (xn )}n∈N is a Cauchy sequence of scalars and hence
Pd
converges to some scalar ck . Define x = k=1 ck xk . The fact that xn → x
p
then follows just as in the proof that ℓ is complete.
1.5 We are given that kxn+1 − xn k < 2−n for every n. Choose any ε > 0,
and let N be large enough that 2−N +1 < ε. If n > m > N, then we have
n−1
X ∞
X 1 1 1
kxn − xm k ≤ kxk+1 − xk k ≤ k
= m−1 < N −1 < ε.
2 2 2
k=m k=m
Then we can find a subsequence {xnk } such that kx − xnk k > ε for every k.
But then no subsequence of {xnk } can converge to x, which is a contradiction.
1.7 It is clear that k · kR is a norm on XR , so the issue is to show that XR is
complete. Suppose that {xn } is Cauchy in XR . Since kxm −xn k = kxm −xn kR ,
we have that {xn } is Cauchy in X and therefore converges to some x ∈ X.
But then kx − xn kR = kx − xn k → 0, so xn converges to x in XR . Hence XR
is complete.
1.8 (a) The Triangle Inequality follows from d(f, h) = k(f − g) + (g − h)k ≤
kf − gk + kg − hk = d(f, g) + d(g, h).
(b) A sequence {fn }n∈N converges to f ∈ X if limn→∞ d(fn , f ) = 0, i.e.,
if
∀ ε > 0, ∃ N > 0, ∀ n ≥ N, d(fn , f ) < ε.
A sequence {fn }n∈N is Cauchy if
(c) For each n, let xn be a rational number such that π < xn < π + 1/n.
Then {xn } is Cauchy, but it does not converge in the space Q. It does converge
in the larger space R, but since the limit does not belong to Q, it is not
convergent in Q.
1.9 (a) Let δ1 , δ2 denote the first two standard basis vectors. These belong
to ℓp , but we have
kx + ykp = (1 + 1)1/p = 21/p
while
kxkp + kykp = 1 + 1 = 2.
Since p < 1 we have 21/p > 2, so the Triangle Inequality is not satisfied by
k · kp .
(b) Suppose that 0 < p < 1. Let f (t) = (1 + t)p and g(t) = 1 + tp for t > 0.
Then f (0) = 1 = g(0). Also,
1 1
f ′ (t) = p (1 + t)p−1 = p and g ′ (t) = ptp−1 = p .
(1 + t)1−p t1−p
Since 0 < 1 − p < 1, we have t1−p < (1 + t)1−p , and therefore f ′ (t) ≤ g ′ (t)
for t > 0. Hence g is increasing faster than f, and therefore f (t) ≤ g(t) for all
t ≥ 0. Next, given any a, b ≥ 0, we have
p p
b b
(a + b)p = ap 1 + ≤ ap 1 + = ap + b p .
a a
but
δ + δ
p 1 p 1 p
1 2
2
= + = p = 21−p > 1,
2 p 2 2 2
so (δ1 + δ2 )/2 does not belong to the closed unit ball. Hence this set is not
convex in ℓp . This also shows that if ε > 0 is small, then the open unit ball
B1+ε (0) is not convex. By rescaling, the unit ball B1 (0) is not convex either.
1.10 (a) Set f (t) = tθ −θt−(1−θ). Then f ′ (t) = θtθ−1 −θ. We have f ′ (t) = 0
if and only if t = 1. Also, f is increasing for 0 < t < 1 and decreasing for
t > 1, and f (1) = 0, so f (t) ≤ 0 for all t > 0, with equality only for t = 1.
(b) Note that
1 1 p p′ 1 p′
+ ′ = 1, p′ = , = , p′ − = 1.
p p p−1 p p−1 p
′
With t = ap b−p and θ = 1/p, we have by part (a) that
1 1
′
−p′ /p
p −p′ 1/p p −p′ ap b−p 1
ab = a b ≤ a b + 1− = + ′.
p p p p
′
Multiplying through by bp and using the fact that p′ − (p′ /p) = 1, we obtain
′
′ ′ ap bp
ab = a bp −p /p ≤ + ′.
p p
′ ′
Equality holds if and only if ap b−p = 1. This is equivalent to bp = ap , or
′
b = ap/p = ap−1 .
p p′
1.11 Case 1 < p < ∞. By Exercise 1.10, equality holds in ab ≤ ap + bp′ if
and only if b = ap−1 . For the normalized case kxkp = kykp′ = 1, equality in
Hölder’s Inequality requires that we have equality in equation (1.5), and this
will happen if and only if |yk | = |xk |p−1 for each k. This is equivalent to
′
|yk |p = |yk |p/(p−1) = |xk |p .
′ ′
Hence α |xk |p = β |yk |p with α = kykpp′ and β = kxkpp . On the other hand, if
either x = 0 or y = 0, then we have equality in Hölder’s Inequality, and we
′
also have α |xk |p = β |yk |p with α, β not both zero.
′
For the converse direction, suppose that α |xk |p = β |yk |p for each k ∈ I,
where α, β ∈ F are not both zero. If α = 0, then yk = 0 for every k, and hence
we trivially have kxyk1 = 0 = kxkp kykp′ . Likewise, equality holds trivially if
β = 0. Therefore, we can assume both α, β 6= 0, and by dividing both sides
′
by β, we may assume that β = 1 and α > 0. Then we have |yk |p = α |xk |p , so
′ X ′ X
kykpp′ = |yk |p = α |xk |p = α kxkpp .
k∈I k∈I
By the work above, this implies that equality holds in Hölder’s Inequality.
Case p = 1, p′ = ∞. Set M = supk |yk |. Suppose equality holds in Hölder’s
Inequality, i.e., X
X
|xk yk | = |xk | sup |yk | .
k
k∈I k∈I
Then X X
|xk yk | = M |xk |.
k∈I k∈I
Hence X
(M − |yk |) |xk | = 0,
k∈I
but 0 ≤ M − |yk | for every k, so we must have (M − |yk |) |xk | = 0 for every k.
Thus whenever xk 6= 0, we must have |yk | = M.
Conversely, if |yk | = M for all k such that xk 6= 0, equality holds in
Hölder’s Inequality.
1.12 We have ℓp ⊆ ℓ∞ for every p. Further, the constant sequence x =
(1, 1, 1, . . . ) belongs to ℓ∞ but not to any ℓp with p finite, so the inclusion is
proper.
Suppose 0 < p ≤ q < ∞ and x ∈ ℓp . If kxk∞ = 1, then
6 Detailed Solutions
X
∞ 1/q X
∞ 1/q
q p q−p
kxkq = |xk | = |xk | |xk |
k=1 k=1
X∞ 1/q
≤ |xk |p = kxkp/q
p ≤ kxkp ,
k=1
the last inequality following from the fact that p/q ≤ 1 and kxkp ≥ kxk∞ = 1.
For the general case, apply this inequality to x/kxk∞ .
To show that the inclusion is strict, set xk = k −1/p . Then since q/p > 1,
we have
X∞
q 1
kxkq = q/p
< ∞,
k=1
k
while
∞
X 1
kxkpp = = ∞.
k
k=1
2
Another example is xk = (k log k)−1/q for k ≥ 2. The Integral Test shows
that
X∞
1
kxkqq = < ∞,
k=2
k log2 k
while ∞
X 1
kxkpp = 2 = ∞.
k=2
(k log k)p/q
1.13 We need the following lemma.
Now we return to the proof of the exercise. Assume that 1 ≤ p < q < ∞.
Taking E0 = E, by applying the Lemma, we can find a set E1 ⊆ E such
that |E1 | = |E|/2. Noting that |E\E1 | = |E|/2, we apply the lemma to
find E2 ⊆ E\E1 with |E2 | = |E|/4, and note that E2 is disjoint from E1 .
Continuing in this way we construct disjoint En ⊆ E such that |En | = 2−n |E|.
Consider
Detailed Solutions 7
X
f = 2n/q χEn .
n
We have
Z X X
kf kqLq = |f |q = 2n |En | = 2n 2−n |E| = ∞,
E n n
since p/q < 1. Hence f ∈ Lp (E). Note that this f is unbounded, so this is
also an example of a function in Lp (E) that does not belong to L∞ (E).
1.14 Suppose that x ∈ ℓq (I) for some finite q. Since only countably many
components of x can be nonzero, it suffices to consider I = N.
If x = 0 then kxkp = 0 for every p, so we are done. Therefore, we may
assume x 6= 0, which implies kxk∞ 6= 0. By dividing through by kxk∞ , we
may assume that kxk∞ = 1. Then for every p we have 1 = kxk∞ ≤ kxkp . In
particular, |xk | ≤ 1 for every k. Therefore, for p ≥ q we have |xk |p ≤ |xk |q .
Hence x ∈ ℓp . Further, for p ≥ q,
X
∞ 1/p
p
kxk∞ = 1 ≤ kxkp = |xk |
k=1
X
∞ 1/p
q
≤ |xk |
k=1
= kxkq/p
q
→ 1 = kxk∞ as p → ∞,
|yM −yN | = lim |xM (k)−xN (k)| ≤ sup |xM (k)−xN (k)| = kxM −xN kℓ∞ ,
k→∞ k
≤ ε + 0 + ε = 2ε.
xN = (x(1), . . . , x(N ), 0, 0, . . . ).
Hence c00 is dense in c0 . However, c00 is not closed, since any x ∈ c0 with
infinitely many nonzero components is an accumulation point of c00 but does
not belong to c00 .
(c) Choose any x ∈ c0 . Write x = (x(1), x(2), . . . ), and set
N
X
xN = (x(1), . . . , x(N ), 0, 0, . . . ) = x(k) δk .
k=1
2
so gN → g uniformly. Hence Cc (R) is dense in C0 (R). However, if g(x) = e−x ,
then g belongs to C0 (R) but does not belong to Cc (R), so Cc (R) is not closed.
(d) Suppose that fn ∈ C(T) and fn → f uniformly. By part (a) we have
f ∈ Cb (R). Since uniform convergence implies pointwise convergence, for each
x ∈ R we have
so {fn′ }n∈N is Cauchy with respect to the uniform norm. That is, {fn′ }n∈N is a
Cauchy sequence in Cb (R). Since Cb (R) is complete, there exists a g ∈ Cb (R)
such that fn′ → g uniformly. So, the remaining point is to show that g = f ′ ,
for then we will have that fn → f in the norm of Cb1 (R).
′
To see this, fix ε > 0. Then there exists an N > 0 such that kfm −fn′ k∞ < ε
whenever m, n > N. Fix x, y ∈ R and m, n > N. Applying the Mean-Value
Theorem to the function fm − fn , there exists a point c (depending on m, n,
x, and y) between x and y such that
′
(fm − fn )(y) − (fm (x) − fn )(x) = (y − x) (fm − fn′ )(c).
Consequently,
fm (y) − fm (x) fn (y) − fn (x)
− = |f ′ (c) − f ′ (c)| ≤ kf ′ − f ′ k∞ < ε.
y−x y−x m n m n
Further, since fn′ → g uniformly, there exists an M such that kfn′ − gk∞ < ε
whenever n > M. Fix x, and suppose that |x − y| < δ. Then for n > M, N
we have
g(x) − f (y) − f (x) < |g(x) − f ′ (x)| + f ′ (x) − fn (y) − fn (x)
y−x n n
y−x
fn (y) − fn (x) f (y) − f (x)
+ −
y−x y−x
< ε + ε + ε = 3ε.
Hence
f (y) − f (x)
g(x) = lim ,
y→x y−x
so f is differentiable at x, and f ′ (x) = g(x). Thus fn → f in the norm of
Cb1 (R), so this space is complete.
A proof by induction shows that Cbm (R) is complete for each m.
(b) If we replace the norm on Cb1 (R) by the uniform norm, then it is no
longer complete. Let w(x) = max{1 − |x|, 0} be the hat function on [−1, 1].
Then we can find differentiable functions fn ∈ Cb1 (R) such that kw − fn k∞ →
0. For example, we just need to “smooth out” the corners of the graph of w
to find fn . Therefore {fn } is a Cauchy sequence in the uniform norm, but it
does not converge within Cb1 (R) because w ∈ / Cb1 (R).
1.23 (a) If f is Hölder continuous with α > 0 then
f (x) − f (y) α
lim ≤ lim C |x − y| = lim C |x − y|1−α = 0.
y→x x−y y→x |x − y| y→x
Define ∞ ∞
X X
f (0) = fn (0) and C = Cn .
n=1 n=1
Given x 6= 0, we have
∞
X ∞
X ∞
X
|fn (x)| ≤ |fn (x) − fn (0)| + |fn (0)|
n=1 n=1 n=1
∞
X
≤ Cn |x − 0|α + S0 < ∞.
n=1
P
Therefore, the series f (x) = |fn (x)| converges pointwise absolutely. Then
∞
X ∞
X
|f (x) − f (y)| ≤ |fn (x) − fn (y)| ≤ Cn |x − y|α = C |x − y|α ,
n=1 n=1
so f ∈ Cα (R). Also,
N N
X X
f (x) − fn (x) − f (y) − fn (y)
n=1 n=1
X
∞ ∞
X
= fn (x) − fn (y)
n=N +1 n=N +1
∞
X
≤ |fn (x) − fn (y)|
n=N +1
∞
X
≤ Cn |x − y|α
n=N +1
Therefore,
14 Detailed Solutions
N
X
f − fn
n=1 Cα
N
X N
X
f (x) − fn (x) − f (y) − fn (y)
N
X
= f (0) − fn (0) + sup n=1 n=1
→ 0 as N → ∞.
P
Therefore the series f = fn converges in C α -norm, so we conclude that
C α (R) is complete.
1.24 (b) Since kϕk+1 − ϕk k∞ < 2−k , we have that ϕk is Cauchy in C[0, 1] by
Exercise 1.5. Consequently ϕk converges uniformly on [0, 1], and since each
ϕk is continuous it follows that the limit ϕ is continuous.
Since
f (3−k ) − f (0) = 2−k − 0 = 2−k = (3−k )log3 2 ,
we have
f (3−k ) − f (0)
= (3−k )−1+log3 2 = 3 k(1−log3 2) .
3−k
However, 1 − log3 2 > 0, so this is not bounded independently of k. Therefore
f is not Lipschitz.
(c) Let ϕ be the Cantor–Lebesgue function. If |x − y| ≤ 3−k , then we have
|ϕ(x) − ϕ(y)| ≤ 2−k . Let k ≥ 0 be the integer such that
1 1
≤ |x − y| ≤ .
3k+1 3k
Then we have
2 1 log3 2
|ϕ(x) − ϕ(y)| ≤ = 2 ≤ 2 |x − y|log3 2 .
2k+1 3k+1
Hence ϕ is Hölder continuous for α = log3 2.
On the other hand, we have
|ϕ(3−k ) − ϕ(0)| = |2−k − 0| = 2−k = (3−k )log3 2 .
Therefore ϕ is not Hölder continuous for any exponent α > − log3 2.
Since the sum of the lengths of the intervals on which ϕ is constant is 1,
we have ϕ′ = 0 a.e.
1.25 We are given a normed linear space X that is not complete, and a a
relation ∼ on the set C of all Cauchy sequences, defined by {xn } ∼ {yn } if
limn→∞ kxn − yn k = 0.
Detailed Solutions 15
e is a normed space.
Therefore X
(c) Given x ∈ X, let [x] denote the equivalence class of the Cauchy se-
quence {x, x, x, . . . }. Then the map T : x → [x] is clearly linear, and
T (x)
e = lim kxk = kxk,
X n→∞
so T is an isometry.
We must show that the range of T is a dense subspace of X. e Given A =
e
[xn ] ∈ X, for each m ∈ N let Ym = T (xm ) = [(xm , xm , xm , . . . )]. Then since
{xn } is Cauchy in X, we have
e so T (X) is dense in X.
Hence Ym → A in X, e
(d) For this part, it will be important to distinguish between the equiv-
alence class of a sequence {xn }n∈N = {x1 , x2 , . . . } and the equivalence class
determined by a constant sequence {xn , xn , . . . }. Therefore, for this part we
will write [{xn }n∈N ] instead of [xn ].
Suppose that {XM }M∈N is a Cauchy sequence in X. e Each XM is an
equivalence class of a sequence, say
XM = [{xM (n)}n∈N ].
For each M, by part (c) there exists some yM ∈ X such that the constant
sequence determined by yM is close to XM :
1
kXM − T (yM )kXe < .
M
Call this constant sequence
Fix ε > 0. Since {yM }M∈N is Cauchy, there exists an N0 > 0 such that
M, N ≥ N0 =⇒ kyM − yN k < ε.
1
kY − XN kXe ≤ kY − YN kXe + kYN − XN kXe ≤ ε + ≤ 2ε.
N
e so X
Therefore XN → Y in X, e is complete.
Thus V is an isometry.
Finally, to show that V is onto, suppose that g is any element of Y. Then
since range(U ) is dense, there exist vectors xn ∈ X such that U (xn ) → g.
e and
Since T U −1 is an isometry, it follows that {T (xn )}n∈N is Cauchy in X,
e
hence T (xn ) → f for some f ∈ X. But then, by definition, V (f ) = g. Hence
V is surjective.
Thus V : X e → Y is an isometric isomorphism.
P P P
1.26 Assume that s = xn and t = yn converge. Let sN = N n=1 xn and
PN
tN = n=1 yn . Then
(s + t) − (sN + tN )
≤ ks − sN k + kt − tN k → 0 as N → ∞.
P
Hence (xn + yn ) exists and equals s + t.
1.27 Let ek (x) = xk . Let θM be 1 on [−M, M ], zero outside [−M − 1, M + 1],
and linear on [−M − 1, −M ] and [M, M + 1]. The set
X
N
S = rk ek θM : M, N ∈ N, rk rational
k=0
18 Detailed Solutions
P∞ There
1.29
k
is nothing to prove if y = 0, so we may assume y 6= 0. Since
k=0 kc y converges, we must have ck y k → 0 as k → ∞. Therefore M =
k
sup |cy y | < ∞. If |x| < |y| then |x/y| < 1, so
∞ ∞ k ∞ k
X X x X x
k
|ck x | = k
|ck y | ≤ M < ∞.
y y
k=0 k=0 k=0
P∞
Thus the series f (x) = k=0 ck xk converges absolutely for all |x| <
P|y|. k
Now suppose that |x| < |y|, and fix r with |x| < r < |y|. Then |ck |r <
∞ by our previous work. Let
X k
x
M = k < ∞.
r
k∈N
Then
∞
X ∞
X k ∞
|ck | rk x M X
k |ck xk−1
| ≤ k ≤ |ck | rk < ∞.
r r r
k=1 k=1 k=1
(x + h)k − xk k−1
= k ξh,k .
h
Assuming that the discrete Fubini’s theorem is justified (this is demonstrated
below), we have for those h with |x + h| ≤ r < |y| that
∞
f (x + h) − f (x) X
lim − k ck xk−1
h→0 h
k=1
X
∞ (x + h)k − xk
∞
X
= lim ck − k ck xk−1
h→0 h
k=1 k=1
∞
X
≤ lim kck (ξ k−1 − xk−1 )
h→0 h,k
k=1
∞
X k−1
= lim k |ck | ξh,k − xk−1
h→0
k=1
∞
X
= 0 = 0.
k=1
where the convergence follows since r, |x| < |y|. Since the summands in the
final series are independent of h, Fubini is justified.
1.30 Suppose that F = C (the same proofs work for F = R).
(b) Let A = [hei , ej i]dj,i=1 . Then for any x, y ∈ Cd we have
he1 , e1 i · · · hed , e1 i x1 y1
.. .. .. .. ..
Ax · y = . . . . · .
he1 , ed i · · · hed , ed i xd yd
x1 he1 , e1 i + · · · + xd hed , e1 i y1
.
. ..
= . · .
x1 he1 , ed i + · · · + xd hed , ed i yd
20 Detailed Solutions
hx, e1 i y1
= ... · ...
hx, ed i yd
2
Similarly |||x||| ≥ λ1 |x|2 , so ||| · ||| is equivalent to the Euclidean norm on Cn .
If F = R then we regard A as a complex matrix and follow the same proof.
1.31 (a) Given f, g ∈ H, we have
kf + gk2 = hf + g, f + gi
= hf, f i + hf, gi + hg, f i + hg, gi
kf + gk2 + kf − gk2
= kf k2 + hf, gi + hg, f i + kgk2 + kf k2 − hf, gi − hg, f i + kgk2
= 2kf k2 + 2kgk2 .
1.32 If p 6= 2 and δ1 , δ2 are the first two standard basis vectors in ℓp , then
while
Detailed Solutions 21
2 kδ1 k2p + kδ2 kp
2
= 2 (12 + 12 ) = 4.
These are not equal when p 6= 2, so the Parallelogram Law does not hold in
ℓp .
1.33 ⇒. The proof of the Cauchy–Bunyakovski–Schwarz Inequality shows
that for every real t we have
which yields the CBS inequality. Suppose now that equality holds in CBS,
i.e.,
|hf, gi| = kf k kgk
for some f and g. Then we have
which means that the discriminant above is zero. This implies that the poly-
nomial above does have a real root, i.e., there is be a t such that
⇐. Suppose that kaf + bgk = 0 for some scalars a, b ∈ F, not both zero.
Suppose that a = 0. Then b 6= 0 but |b| kgk = kaf + bgk = 0, so kgk = 0.
Hence 0 ≤ |hf, gi| ≤ kf k kgk = 0, so equality holds in CBS. Similarly, equality
holds if b = 0, so we can assume a, b 6= 0.
By the Polar Identity,
Hence
or
kaf k kbgk = |Rehaf, bgi|.
|ab| kf k kgk = kaf k kbgk ≤ |haf, bgi| = |ab| |hf, gi| ≤ |ab| kf k kgk.
Since ab 6= 0, equality in CBS holds.
1.34 Let h·, ·i be a semi-inner product on H. Then we proceed just as in
the proof of Cauchy–Bunyakovski–Schwarz. If x = 0 or y = 0 then there is
nothing to prove, so suppose that both are nonzero. Write hx, yi = α |hx, yi|
where α ∈ F and |α| = 1. The Polar Identity holds for semi-inner products,
so for t ∈ R we have
But then kp − qk = 0, so p = q.
1.36 If y, z ∈ A⊥ then for any x ∈ A we have
so ay + bz ∈ A⊥ . Therefore A⊥ is a subspace of H.
Suppose that yn ∈ A⊥ and yn → y in H. Then for every x ∈ A we have
by the continuity of the inner product that
Detailed Solutions 23
Therefore y ∈ A⊥ , so A⊥ is closed.
1.37 (a) ⇒ (b). Fix F = C in this part (the proof for F = R is similar).
Let p be the (unique) point in M closest to h, and let e = p − h. Choose any
f ∈ M. We must show that hf, ei = 0. Since M is a subspace, p + λf ∈ M for
any scalar λ ∈ C. Hence,
Therefore,
∀ λ ∈ C, 2 Re(λhf, ei) ≤ |λ|2 kf k2 .
If we consider λ = t > 0, then we can divide through by t to get
Hence f = 0 for p = q ∈ M.
1.38 (a) If f ∈ M then hf, yi = 0 for every y ∈ M ⊥ , so f ⊥ M ⊥ and therefore
f ∈ (M ⊥ )⊥ .
24 Detailed Solutions
Hence {xn }n∈N and {yn }n∈N are each Cauchy, so must converge in H. There-
fore xn → x and yn → y for some x, y ∈ H, and since M and N are closed we
have x ∈ M and y ∈ N. Thus, we have both xn + yn → x+ y and xn + yn → h,
so z = x + y ∈ M ⊕ N. Therefore M ⊕ N is closed.
(b) This follows from Theorem 1.43.
1.40 Let H, K be Hilbert spaces. It is clear that (h1 , k1 ), (h2 , k2 ) =
hh1 , h2 iH + hk1 , k2 iK is an inner product on H × K. The induced norm is
(h, k)
2 = khk2H + kkk2K ,
and it follows from Exercise 1.16 that H × K is a Banach space with respect
to this norm. Therefore H × K is a Hilbert space.
1.41 Let H be an inner product space. Then by Exercise 1.25, its completion
e is a Banach space.
H
e there exist xn , yn ∈ H such that xn → x and yn → y. Let
Given x, y ∈ H,
C = max{kxn k, kyn k}n∈N . Given ε > 0, there exists some N > 0 such that
|hxm , ym i − hxn , yn i|
≤ |hxm , ym i − hxm , yi| + |hxm , yi − hxn , yi| + |hxn , yi − hxn , yn i|
≤ kxm k kym − yk + kxm − xn k kyk + kxn k ky − yn k
≤ C kym − yk + C kyk + C ky − yn k ≤ 3Cε.
Hence hxn , yn i is a Cauchy sequence in F, so we can define
Detailed Solutions 25
P∞
(c) ⇐. Suppose that n=1 |cn |2 < ∞. Set
N
X N
X
sN = cn xn , tN = |cn |2 .
n=1 n=1
P∞(e) By Bessel’s Inequality and part (c), we know that the series p =
n=1 hf, xn i xn converges, so we just
have to show that it is the orthogonal
projection of f onto span {xn }n∈N . Given any k we have by the linearity
and continuity of the inner product that
∞
X
hf − p, xk i = hf, xk i − hf, xn i hxn , xk i = hf, xk i − hf, xk i = 0.
n=1
By linearity of the inner product, this implies that f −p ⊥ span {xn }n∈N . By
continuity of the inner product, this extends to f − p ⊥ span {xn } . Hence p
is indeed the orthogonal projection of f onto span {xn }n∈N .
P
(e) ⇐. Suppose x = hx, xn i xn . Then, by part (e), x = p, the orthogonal
projection of x onto span{xn }, so x ∈ span{xn }.
Now we prove the remaining implications in Theorem 1.50.
(a) ⇒ (b). If {xn } is complete,
P∞ then its closed span is all of H, so by
Exercise 1.42(e) we have f = n=1 hf, xn i xn for every f ∈ H.
(b) ⇒ (c). If statement (b) holds, then we must have cn = hx, xn i by
Theorem 1.49.
(c) ⇒ (e). Suppose that (c) holds, and choose any f, g ∈ H. Then
DX
∞ E ∞
X
∞
X
hf, gi = hf, xn i xn , g = hf, xn i xn , g = hf, xn i hxn , gi,
n=1 n=1 n=1
Therefore dist(y, M ) = 1.
(b) Let H be an infinite-dimensional Hilbert space. Choose any nonzero
vector in H, and normalize it to obtain a vector e1 with norm 1. Since H is
infinite dimensional, H 6= span{e1 }. Therefore by part (a) we can find a vector
Detailed Solutions 27
a
Fix δ > 0 small enough that a+δ > 1 − ε. By the definition of infimum, there
exists some v ∈ M such that a ≤ ku − vk < a + δ. Set
u−v
g = ,
ku − vk
= (c0 + c1 α, c1 β + c2 α, c2 β + c3 α, . . . , cN −1 β + cN α, cN β, 0, 0, . . . ),
28 Detailed Solutions
N
X
sN = ck xk
k=0
= (c0 + c1 α, c1 β + c2 α, c2 β + c3 α, . . . , cN −1 β + cN α, cN β, 0, 0, . . . ),
= (0, 0, . . . , 0, cN β, 0, 0, . . . ).
β
Since | α | < 1,
2N −2
β
ksN k2 = |cN β|2 = β 2 → 0 as N → ∞.
α
P∞
Thus sN → 0, which says that k=0 ck xk = 0.
Since {xk }k≥0 is not ω-independent, part (a) therefore implies that it
cannot be a Schauder basis.
1.47 (a) ⇒ (b). For each m, let Em be the closed subspace Em =
span{xn }n6=m . Then we can write xm uniquely as xm = pm + qm where
⊥
pm ∈ Em and qm ∈ Em . In particular, we have qm ⊥ xn for every n 6= m.
If we had hxm , qm i = 0, then we would have
For n 6= 0 we have
Z 1 Z 1/2 Z 1
fb(n) = f (x) e −2πinx
dx = e −2πinx
dx − e−2πinx dx
0 0 1/2
1/2 1
e−2πinx e−2πinx
= −
−2πin 0 −2πin 1/2
e−πin − 1 1 − e−πin
= −
−2πin −2πin
(
2(−1)n − 1 0, n even,
= = 2i
−2πin − πn , n odd.
Therefore,
2
X X 2i X 4
1 = kf k22 = |fb(n)|2 = = 2
πn .
π 2 n2
n∈Z n∈Z, n odd n∈N, n odd
< kxk2 ,
Hence kx − pk = 0, so x = p ∈ span{xn }.
1.53 Let {en } be any orthonormal basis for H. Let {fn } be an orthonormal
basis for H that contains an orthonormal basis for M, i.e., {fn }n∈I is an
orthonormal basis for M and {fn }n∈J is an orthonormal basis for M ⊥ , where
I ∪ J is a partition of N. Then
X XX
kP en k2 = |hP en , fm i|2
n n m
XX
= |hen , P fm i|2
m n
X
= kP fm k2
m
X
= 12 = |I| = dim(M ).
m∈I
⇐. Suppose that
∞ Z
X x 2
fn (t) dt = x − a, x ∈ [a, b].
n=1 a
Then,
∞
X
|hχ[a,x] , fn i|2 = x − a = kχ[a,x] k22 .
n=1
Thus, the Plancherel Equality holds for χ[a,x] , so by Exercise 1.52 we have
that χ[a,x] ∈ span{fn } for every x ∈ [a, b]. Hence χ[x,y] = χ[a,y] − χ[a,x] ∈
span{fn } for every x < y. But span{χ[x,y] : a ≤ x < y ≤ b} is the set of
“really simple functions,” which is dense in L2 [a, b] by Lemma A.28. Hence
span{fn } = L2 [a, b], as desired.
32 Detailed Solutions
1.55 We apply Vitali’s criterion with [a, b] = [0, 1]. Note that
Therefore
2 Z 2
X Z x x X e2πinx − 1 2
e 2πint
dt =
1 dt +
2πin
n∈Z 0 0 n6=0
X 2 − 2 cos 2πnx
= x2 +
4π 2 n2
n6=0
X∞
1 − cos 2πnx
= x2 + . (A)
n=1
π 2 n2
Vitali’s criterion tells us that {e2πinx }n∈Z is complete in L2 [0, 1] if and only
if 2
X Z x
e 2πint
dt = x, x ∈ [0, 1]. (B)
n∈Z 0
X∞
1 − cos 2πnx
x2 + = x.
n=1
π 2 n2
⇐. Suppose that
∞ Z ∞ Z Z 2
X b X b x (b − a)2
|hχ[a,x] , fn i|2 dx = fn (t) dt = .
2
n=1 a n=1 a a
Therefore
Detailed Solutions 33
Z b ∞
X
χ 2 χ 2
0 ≤ k [a,x] kL2 − |h [a,x] , fn i| dx
a n=1
Z b Z b ∞
X
= kχ[a,x] k2L2 − |hχ[a,x , fn i|2 dx
a a n=1
Z b
(b − a)2
= (x − a) dx − = 0.
a 2
Consequently,
∞
X
x − a = kχ[a,x] k2L2 = |hχ[a,x] , fn i|2 a.e. x.
n=1
has positive measure. Therefore, there must exist some n such that the set
E = {|f1 | > 1/n} has positive measure. Write E = ∪Fk disjointly where each
Fk has positive measure. Define
(
|Fk |1/2
m(t) = |f1 (t)| , t ∈ Fk ,
1, otherwise.
|Fk |1/2 n
|m(t)| ≤ ≤ n (b − a)1/2 ,
k
34 Detailed Solutions
so f1 /m ∈/ L2 [a, b].
Since m ∈ L∞ [a, b], we have mfn ∈ L2 [a, b] for every n. Now suppose
that g ∈ L2 [a, b] satisfies hg, mfn i = 0 for every n ≥ 2. Then (since m is
real-valued)
Z b
0 = hg, mfn i = g(t) m(t) fn (t) dt = hgm, fn i, n ≥ 2.
a
Since gm ∈ L [a, b] and {fn }n∈N is an orthonormal basis for L2 [a, b], we
2
conclude that
∞
X
gm = hgm, fn i fn = hgm, f1 i f1 .
n=1
e2πi(ξ−η)T − e−2πi(ξ−η)T
= lim = 0,
T →∞ 4T πi(ξ − η)
and Z T
1
heξ , eη i = lim e2πiξt e−2πiξt dt = 1.
T →∞ 2T −T
Hence, once we show that h·, ·i is an inner product, we can conclude that
{eξ }ξ∈R is an orthonormal system.
Suppose that f, g, h ∈ H and the limits defining hf, hi and hg, hi exist.
Then
Z T
1
hf + g, hi = lim (f (t) + g(t)) h(t) dt = hf, hi + hg, hi
T →∞ 2T −T
Detailed Solutions 35
exists as well, and similarly hcf, hi = chf, hi and hh, f i = hf, hi.
PM
In particular, given f, g ∈ H we can write f = m=1 cm eξm and g =
PN
n=1 dn eξn , so we have that
M X
X N M X
X N
hf, gi = cm dn heξ , eη i = cm dn δξη
m=1 n=1 m=1 n=1
Rearranging, we obtain
kLyk ≤ K kyk = K.
Hence K is an upper bound for {kLyk : kyk = 1}. But, by definition, kLk is
the least upper bound for that set, so we must have kLk ≤ K.
(e) We certainly have
kLxk
sup kLxk = sup .
kxk=1 x6=0 kxk
Continuous: Since
∞
X ∞
X
kLxk22 = k(x2 , x3 , . . . )k22 = |xk |2 ≤ |xk |2 = kxk22 ,
k=2 k=1
Detailed Solutions 37
(x2 , x3 , . . . ) = Lx = λ x = (λ x1 , λ x2 , . . . ).
Therefore:
x2 = λ x1 , x3 = λ x2 = λ2 x1 , ....
n−1
Thus xn = λ x1 for every n. But x must be an element of ℓ2 , so kxk2 must
be finite. Therefore
∞
X ∞
X ∞
X
kxk22 = |xn |2 = |λn−1 x1 |2 = |x1 |2 |λ|2n−2 < ∞.
n=1 n=1 n=1
Lx = (λ, λ2 , λ3 , . . . ) = λ (1, λ, λ2 , . . . ) = λ x.
(0, x1 , x2 , x3 , . . . ) = Rx = λ x = (λ x1 , λ x2 , . . . ).
Hence,
λ x1 = 0, λ x2 = x1 , λ x3 = x2 , etc.
That is, λ x1 = 0, and λ xn+1 = xn for every n. If λ 6= 0 then this implies
that xn = 0 for every n, a contradiction. Thus no λ 6= 0 can be an eigenvalue.
On the other hand, if λ = 0, then xn = λ xn+1 = 0 xn+1 = 0 for every n,
also a contradiction. Thus λ = 0 can’t be an eigenvalue either. Thus R has no
eigenvalues.
(c) We have
so hx, enk i = 1 for every k. This is impossible, so no such x can exist. There-
fore Mλ is not surjective. On the other hand, Mλ en = λn en and λn 6= 0, so
range(Mλ ) contains span{en }n∈N , which is dense.
1.67 If f ∈ Lp (E) with p finite, then
Z 1/p
kTm f kLp = |f (t) m(t)|p dt
E
Z 1/p
p
≤ kmk L∞ |f (t)| dt = kmkL∞ kf kLp .
E
40 Detailed Solutions
Therefore Tm is bounded and kTm k ≤ kmkL∞ . A similar proof shows that the
same inequality holds if p = ∞.
If kmkL∞ = 0 then m = 0 a.e. and Tm is the zero operator, so there is
nothing to prove in this case. Suppose M = kmkL∞ > 0 and fix ε > 0. Then
A = {x ∈ E : |m(x)| > M − ε} has positive measure. Since Lebesgue measure
is σ-finite, some set Ak = A∩[k, k+1] must have positive (and finite) measure.
If p < ∞, consider f = |Ak |−1/p χAk . We have
Z
p
kf kp = |Ak |−1 dx = 1,
Ak
but Z
kTm f kpp = |Ek |−1 |m(x)|p dx ≥ (kmk∞ − ε)p .
Ak
Hence kTm k ≥ kmk∞ −ε, and ε is arbitrary, so we conclude that kTm k−kmk∞.
On the other hand, if p = ∞, then consider f = χAk . We have kf k∞ = 1,
and
kTm f k∞ = kmχAk k∞ ≥ kmk∞ − ε,
and the inequality again follows.
1.68 Let B = {e1 , . . . , ed } be a basis for X, and let k · k1 be the ℓ1 -norm on X
Pd
with respect to that basis. Given x ∈ X, write x = k=1 xk ek . Then
X
n
n
X
kLxk =
L(xk ek )
≤ |xk | kLek k
k=1 k=1
Xn
≤ |xk | max kLek k = C kxk1 ,
k
k=1
where C = maxk kLek k < ∞. Hence L is bounded. Since all norms on X are
equivalent, it follows that L is bounded with respect to any norm on X.
1.69 (a) The norm properties follow directly from the definition. For example,
if kLk = 0, then kLf k = 0 for every unit vector f. Hence Lf = 0 for every unit
vector f. Since every vector is a scalar multiple of a unit vector, we conclude
that Lf = 0 for every f ∈ X.
Given any f ∈ X we have
k(L + K)f k = kLf + Kf k ≤ kLf k + kKf k = kLk + kKk kf k.
Taking the supremum over all unit vectors, we have kL + Kk ≤ kLk + kKk.
(b) Let X be a normed space and Y a Banach space. Assume that {An }n∈N
is a sequence of operators in B(X, Y ) that is Cauchy in operator norm. For
any given f ∈ X, we have
kAm f − An f k ≤ kAm − An k kf k,
Detailed Solutions 41
kAxk1 = kx1 v1 + · · · + xn vn k1
Hence n o
kAkℓ1 →ℓ1 ≤ max kvj k1 .
j
42 Detailed Solutions
Hence X
n
kAkℓ∞ →ℓ∞ ≤ max |aij | .
i
j=1
On the other hand, fix any particular i, and set x = (x1 , . . . , xn ) where xj aij =
|aij |. Then kxk∞ = 1 and
n n
X X
kAxk∞ ≥ (Ax)i = aij xi = |aij |.
j=1 j=1
1.71 Choose any orthonormal basis {en }n∈N for P H any orthonormal basis
{fn }n∈N for K. Define L : H → K by Lg = hg, fn i fn . The Plancherel
Equality implies that L is an isometry, and it also follows that the range of L
is closed. Since fn ∈ range(L) for each n, we conclude that range(L) = K,
so L is surjective.
1.72 Fix any f ∈ X. Since Y is dense in X, there exist gn ∈ Y such that
gn → f. Since L is bounded, we have kLgm − Lgn k ≤ kLk kgm − gn k. But
{gn }n∈N is Cauchy in X, so this implies that {Lgn }n∈N is Cauchy in Z.
Since Z is a Banach space, we conclude that there exists an h ∈ Z such that
e = h.
Lgn → h. Define Lf
e
To see that L is well defined, suppose that we also had gn′ → f for some
gn ∈ Y. Then kLgn′ − Lgn k ≤ kLk kgn′ − gn k → 0. Since Lgn → h, it follows
′
Hence kLke ≤ kLk. Combining this with the opposite inequality derived above,
we conclude that kLke = kLk.
Finally, we must show that Le is unique. Suppose that A ∈ B(X, Y ) also
satisfied A|Y = L. Then Af = Lfe for all f ∈ Y. Since Y is dense, this extends
e
by continuity to all f ∈ X, which implies that A = L.
′
1.73 (a) Consider p = 1. We know that T maps ℓp = ℓ∞ into (ℓ1 )∗ with
kµy k ≤ kykℓ∞ for each y ∈ ℓ∞ , so it remains to show that T is isometric and
surjective.
Choose y ∈ ℓ∞ . Then kδn k1 = 1, and we have
|hδn , µy i| = |yn |,
so kµy k ≥ |yn |. Since this is true for every n, we conclude that kµy k ≥ kykℓ∞ .
Hence T is an isometry.
To see that T is surjective, fix any µ ∈ (ℓ1 )∗ . Let yn = hδn , µi and set
y = (yn ). Since
This shows that kµy k ≥ kykℓ1 . Hence we have kµy k = kykℓ1 , which establishes
that T is an isometry.
1.74 By hypothesis, λ = hz, µi =6 0.
Given x ∈ X, let c = hx, µi, and set y = x − (c/λ) z. Then
c
hy, µi = hx, µi − hz, µi = c − c = 0,
λ
so y ∈ ker(µ).
To show this representation is unique, suppose that we also had x = y ′ +c′ z
where y ′ ∈ ker(µ) and c′ ∈ F. Then y − y ′ = (c′ − c)z, so since y − y ′ ∈ ker(µ)
we have
(c′ − c) λ = h(c′ − c) z, µi = hy − y ′ , µi = 0.
Therefore c = c′ , and hence y − y ′ = 0.
1.75 Given y ∈ ℓ1 , define µy : c0 → F by
∞
X
hx, µy i = xk yk , x ∈ c0 .
k=1
Hence µy is bounded and kµy k ≤ kyk1 . Let ck be the scalars of unit modulus
such that ck yk = |yk |. Set xn = (c1 , . . . , cn , 0, 0, . . . ). Then xn ∈ c0 , kxn k∞ =
1, and
Xn X n
hxn , µy i = ck y k = |yk | → kyk1 as n → ∞.
k=1 k=1
= hxn , µi
1.77 (a) We have that µh is linear, and |µh (f )| = |hf, hi| ≤ kf k khk, so µh is
bounded and kµh k ≤ khk. If h = 0, then µh = 0 and so kµh k = 0 = khk. On
the other hand, if h 6= 0 then g = h/khk is a unit vector, and
hh, hi
|µh (g)| = |hg, hi| = = khk,
khk
g = p + e, p ∈ ker(µ), e ∈ ker(µ)⊥ .
Note that µ(p) = 0 by definition, and therefore we must have µ(e) 6= 0 (since
µ(g) 6= 0). Set u = e/µ(e), and note that u ∈ ker(µ)⊥ and µ(u) = 1.
Given f ∈ H, since µ is linear and µ(u) = 1 we have
µ f − µ(f ) u = µ(f ) − µ(f ) µ(u) = 0.
Hence,
46 Detailed Solutions
1 D u E
µ(f ) = hf, ui = f, = hf, hi = µh (f )
kuk2 kuk2
where
u
h = .
kuk2
Thus L = Lh , and from part (a), we have that kLk = kLh k = khk.
It remains only to show that h is unique. Suppose that we also had µ = µh′ .
Then for every f ∈ H we have
Consequently, h − h′ = 0.
(c) Parts (a) and (b) show that T is surjective and T is norm-preserving.
Finally, the part of the proof of Theorem 1.75 given in the text shows that T
is antilinear.
1.78 ⇐. This follows from Hölder’s Inequality.
⇒. Case 1 ≤ p < ∞. The hypotheses imply that T x = hx, yi is a bounded
linear map of (c00 , k·kℓp ) into F. By Exercise 1.72, T has a unique extension to
′
a bounded linear functional on ℓp , which we also call T. Hence T ∈ (ℓp )∗ = ℓp ,
′
so there exists some z ∈ ℓp such that T x = hx, zi for all x ∈ ℓp . Therefore,
′
by considering the standard basis vectors, we see that y = z ∈ ℓp .
Case p = ∞. Let αn be scalars of unit modulus such that αn yn = |yn |.
Fix N and set α = (α1 , . . . , αN , 0, 0, . . . ). Then
N
X N
X
|yn | = αn yn = hα, yi ≤ C kαk∞ = C < ∞.
n=1 n=1
|hx, µi| = |hy, µi + c hz, µi| = |c| |hz, µi| = λ |c| ≤ 2λ kxk.
Therefore µ is bounded.
Detailed Solutions 47
1.80 Define T : X ∗ × Y ∗ → (X × Y )∗ by
Since
Hence kT (µ, ν)k ≥ k(µ, ν)k1 − 2ε, and ε > 0 is arbitrary. Combined with our
calculation above, this implies kT (µ, ν)k = k(µ, ν)k1 . Hence T is an isometry.
It remains to show that T is surjective. Suppose that λ ∈ (X × Y )∗ . Define
µ : X → F and ν : Y → F by
hx, µi = (x, 0), λ and hy, νi = (0, y), λ .
Then
|hx, µi| ≤ k(x, 0)k∞ kλk = kxk kλk,
so µ is bounded and therefore belongs to X ∗ . Similarly, ν ∈ Y ∗ . Given (x, y) ∈
X × Y we have
T (µ, ν)(x, y) = hx, µi + hy, νi = (x, 0), λ + (0, y), λ = (x, y), λ .
where we have used the fact that the norm of λ as an element of H coincides
with the operator norm of the functional on H that it induces.
2.2 Since x0 ∈
/ M, we can write x0 uniquely as
x0 = y0 + z0 , y0 ∈ M, z0 ∈ M ⊥ .
Note that
ky0 k = kx0 − z0 k = dist(x0 , M ) = d.
Set S = span{y0 }. Then S is one-dimensional, and hence is a closed sub-
space of H. Therefore we have H = S ⊕S ⊥ . Hence every x ∈ H can be written
uniquely as
x = cx y 0 + z x , cx ∈ F, zx ∈ S ⊥ .
Define µ : H → F by
hx, µi = cx , x ∈ H.
Then µ is linear and hx0 , µi = 1. Given x ∈ H, we have by the Pythagorean
Theorem that
Rearranging,
1
|hx, µi| ≤ kxk, x ∈ H,
d
so kµk ≤ 1/d. Also, y0 /d is a unit vector and
D y E 1
0
, µ = ,
d d
so kµk ≥ 1/d.
Thus µ ∈ H ∗ and kµk = 1/d. By the Riesz Representation Theorem, µ is
identified with an element of H that has the same norm. This µ has all of the
required properties.
2.3 (a) Suppose that X ∗ is separable. Then X ∗ has a countable dense sub-
set, say {µn }n∈N . Setting λn = µn /kµn k, it follows that S = {λn }n∈N is a
countable dense subset of the closed unit sphere
Detailed Solutions 49
D∗ = µ ∈ X ∗ : kµk = 1 .
kλ − λn k = sup |hx, λ − λn i|
kxk=1
≥ sup |0 − hxn , λn i|
n∈N
1
≥ .
2
But this contradicts the fact that {λn }n∈N is dense in D∗ .
Therefore we must have M = X, and therefore X is separable since M is
separable.
(b) The space ℓ1 is separable, but its dual space is ℓ∞ , which is not sepa-
rable.
2.4 Suppose that µ ∈ C[0, 1]∗ is such that hx2k , µi = 0 for every k ≥ 0.
Extend µ symmetrically to [−1, 1]. That is, given f ∈ C[−1, 1], define g,
h ∈ C[0, 1] by g(x) = f (x) and h(x) = f (−x), and set
Thus hxk , νiC[−1,1] = 0 for every k ≥ 0. Since {xk }k≥0 is complete in C[−1, 1],
we conclude that ν = 0.
Now, given f ∈ C[0, 1], extend f symmetrically to [−1, 1], i.e., set f (−x) =
f (x) for x ∈ [−1, 0). Then
T : ℓp → (ℓp )∗∗
y 7→ ye
be the canonical embedding of ℓp into (ℓp )∗∗ . That is, ye : (ℓp )∗ → F is given
by
hµ, ye i = hy, µi, µ ∈ (ℓp )∗ .
′
Given y ∈ ℓp , we know that
X
hx, µy i = xk yk x ∈ ℓp ,
k
= hz, Tp′ yi
X
= zk yk dx
k
= hy, Tp zi
= hy, νi
= hν, ye i.
hf, A∗ µ − Bµi = 0, f ∈ X.
hx, L∗ µi = hLx, µi = 0.
Also,
hx, ǫ∗ µi = hǫx, µi = hx, µi = hx, µ|M i,
so ǫ∗ µ = µ|M .
2.13 Let hx, yi = x · y = xT y be the usual sesquilinear inner product on Cn .
Then given x ∈ Cn and y ∈ Cm , we have
T
hAx, yi = (Ax)T y = xT AT y = xT (A y) = xT (AT y) = hx, AT yi.
Hence L = R∗ .
2.15 Given x, y ∈ H we have
X
∞
hMλ x, yi = λn hx, en i en , y
n=1
∞
X
= λn hx, en i hen , yi
n=1
∞
X
= λn hx, en i hen , yi
n=1
∞
X
= x, λ̄n hy, en i en = hx, Mλ̄ yi,
n=1
Detailed Solutions 53
so Mλ∗ = Mλ̄ .
Thus, Mλ is self-adjoint if and only if λ = λ̄, i.e., λ has all real components.
Finally, Mλ is positive if and only if λn ≥ 0 for all n, and is positive definite
if and only if λn > 0 for all n.
2.16 (a), (b) For any x ∈ H and y ∈ L we have
∗ ∗
(A ) x, y = hx, A∗ yi = hAx, yi
and
hBAx, yi = hAx, B ∗ yi = hx, A∗ B ∗ yi,
so (A∗ )∗ = A and (BA)∗ = A∗ B ∗ .
(c) Assume that x ∈ ker(A) and let z ∈ range(A∗ ), i.e., z = A∗ y for some
y ∈ K. Then since Ax = 0, we have hx, zi = hx, A∗ yi = hAx, yi = 0. Thus
x ∈ range(A∗ )⊥ , so ker(A) ⊆ range(A∗ )⊥ .
Now assume that x ∈ range(A∗ )⊥ . Then for any z ∈ H we have hAx, zi =
hx, A∗ zi = 0. But this implies Ax = 0, so x ∈ ker(A). Thus range(A∗ )⊥ ⊆
ker(A).
(d) Using part (c), we have
⊥
ker(A)⊥ = range(A∗ )⊥ = range(A∗ ),
⇐⇒ range(A∗ ) = H.
(f) The fact that kAk = kA∗ k follows from the corresponding result for
Banach adjoints, but we can also give a direct proof. Given any y ∈ K,
Also,
kAxk2 = hAx, Axi = hA∗ Ax, xi ≤ kA∗ Axk kxk.
Taking the supremum over all unit vectors, we obtain
Consequently kAk2 = kA∗ Ak. The final equality, kAk2 = kAA∗ k, follows by
interchanging the roles of A and A∗ .
2.17 ⇒. If P is the orthogonal projection onto M, then it is clear that P 2 = P
and range(P ) = M. If x, y ∈ H, let e = x − P x ∈ M ⊥ and f = y − P y ∈ M ⊥ .
Then
hP x, yi − hx, P yi = hP x, P y + f i − hP x + e, P yi
= hP x, P yi + hP x, f i − hP x, P yi − he, P yi
= hP x, P yi + 0 − hP x, P yi − 0 = 0.
hx − p, yi = hx, P zi − hP x, P zi
= hx, P zi − hx, P ∗ P zi
= hx, P zi − hx, P 2 zi = 0.
(ABA)∗ = A∗ B ∗ A∗ = ABA,
so ABA is self-adjoint.
(b) ⇒. Suppose A, B are self-adjoint and AB is self-adjoint. Then AB =
(AB)∗ = B ∗ A∗ = BA.
⇐. If A, B are self-adjoint and AB = BA, then (AB)∗ = B ∗ A∗ = BA =
AB, so AB is self-adjoint.
(c) Let P, Q be orthogonal projections onto one-dimensional nonorthogo-
nal lines in H. That is, P f = hf, gi g and Qf = hf, hi h for some nonorthogonal
unit vectors g, h. Then range(P Q) ⊆ span{g} while range(QP ) ⊆ span{h}.
But P Q and QP are not the zero operators, so these ranges are these 1-
dimensional lines. To see this directly, we compute that
QP f = hP f, hi h = hf, gi hg, hi h,
and
hAA∗ g, gi = hA∗ g, A∗ gi = kA∗ gk2 ≥ 0,
so A∗ A, AA∗ ≥ 0.
2.21 (a) We certainly have that Af = 0 implies A∗ Af = 0. On the other
hand, if A∗ Af = 0, then
(b) Suppose that A ≥ 0. If we fix 0 < t < kAk−1 , then for each x ∈ H we
have
htAx, xi ≤ ktAxk kxk = t kAk kxk2 < kxk2 = hIx, xi.
Hence tA ≤ I. Further, since t is a positive real scalar, the operator A has a
square root if and only if the operator tA has a square root.
(c) We show by induction that Tn+1 − Tn is a polynomial in B with all co-
efficients nonnegative. Since T1 − T0 = T1 = 21 B, this is true for the base step.
Assume that Tn − Tn−1 is a polynomial in B with all nonnegative coefficients.
Then, using the fact that B, Tn , and Tn−1 commute, we have
1 1
Tn+1 − Tn = (B + Tn2 ) − (B + Tn−1 )
2 2
1
= (Tn2 − Tn−1
2
)
2
1
= (Tn − Tn−1 ) (Tn + Tn+1 ).
2
Now, by the inductive hypothesis, Tn − Tn−1 is a polynomial in B with all
coefficients nonnegative. The same is true for Tn and Tn−1 individually, and
therefore Tn + Tn−1 is also a polynomial in B with all coefficients nonnegative.
Therefore the product (Tn − Tn−1 ) (Tn − Tn−1 ) is a polynomial in B with all
coefficients nonnegative.
(d) We have kT0 k = 0 and
1 1 1 1
kT1 k = kBk = kI − Ak ≤ (kIk + kAk) ≤ (1 + 1) = 1.
2 2 2 2
Detailed Solutions 57
so T is bounded.
By continuity of the inner product,
hT x, xi = lim hTn x, xi ≥ 0,
n→∞
so T is a positive operator.
We will show that T commutes with B. Note that
Therefore
1
T x = lim Tn+1 x = lim (B + Tn2 )x
n→∞ n→∞ 2
1 1
= Bx + lim Tn2 x
2 2 n→∞
1 1
= Bx + T 2 x.
2 2
58 Detailed Solutions
1 1 A S2
I −S = I − A + (I − S)2 = I − A + I − 2S + S 2 = I − − S + .
2 2 2 2
Rearranging, we find that A = S 2 .
Since kT k ≤ 1, we therefore have that
Hence S ≥ 0.
2.24 (a) Suppose that A ≥ 0 and hAx, xi = 0. By Theorem 2.18, there exists
a positive operator S such that S 2 = A. Hence
Since B, S ≥ 0, this implies hSy, yi = 0 = hBy, yi. Part (a) therefore implies
that Sy = 0 = By. Hence
kSx − Bxk2 = (S − B)x, (S − B)x = (S − B)(S − B)x, x
= S(S − B)x, x − hB(S − B)x, xi
= hSy, xi − hBy, xi = 0.
Therefore Sx = Bx.
2.25 (a) ⇒ (b). Suppose that E contained a nonempty open subset. Then
there is some ball Br (x) contained in E. Hence X \ E cannot be dense since
every element of X \ E lies at least a distance r from x. Thus E is not nowhere
dense.
(b) ⇒ (a). Suppose that E contains no nonempty open subsets. Choose
any x ∈ X and r ≥ 0. Then Br (x) is not a subset of E, so there must exist
some y ∈ X \ E that belongs to Br (x), i.e., d(x, y) < r. Thus X \ E is dense
in X, so E is nowhere dense.
Detailed Solutions 59
This is a closed subspace of C0 (R) with respect to the uniform norm. Given
ε > 0, let g be the hat function supported on [N, N + 1] with height ε/2.
Given any f ∈ C[−N, N ] we have h = f + g ∈ / C[−N, N ] yet
Suppose that fk ∈ Fn and fk → f in C[0, 1]. Then for each k ∈ N there exists
some xk such that
k ≥ N1 =⇒ |xnk − x0 | < δ.
k ≥ N2 =⇒ kf − fnk k∞ < ε.
60 Detailed Solutions
Suppose x0 < x < 1. Then there exists some k (in fact, infinitely many) so
that x0 ≤ xnk < x < 1. Hence we can choose k so that we have in addition
that n = nk ≥ N1 , N2 . Then we have |xn − x0 | < δ, so
|f (x) − f (x0 )| ≤ |f (x) − fn (x)| + |fn (x) − fn (xn )| + |fn (xn ) − f (xn )|
+ |f (xk ) − f (x0 )|
≤ kf − fn k∞ + n (x − xn ) + kf − fn k∞ + ε
≤ n (x − xn ) + 3ε.
|f (x) − f (x0 )| 2 kf k∞
≤ = C.
x − x0 δ
Let n ∈ N be larger than C. Then we have |f (x) − f (x0 )| ≤ n (x − x0 ) for
all x0 ≤ x < 1, so f ∈ Fn . Thus D is contained in ∪Fn , and therefore D is
a proper subset of C[0, 1]. Consequently, there must exist functions in C[0, 1]
that have no right derivative at any point.
2.29 By hypothesis, Ax = limn→∞ An x exists for each x ∈ X. This implies
that A is linear, and also
≤ kAx − An xk + M kxk
→ 0 + M kxk as n → ∞.
Hence An x → Ax as n → ∞.
(b) Let {ek } be an orthonormal basis for a separable Hilbert space H, and
let S = span{ek }, which is a dense subspace of H. Define
n
X
An x = k hx, ek i ek .
k=1
Hence, if x ∈ X then
sup |hx, x∗ i| : x∗ ∈ S ≤ sup kxk kx∗ k| : x∗ ∈ S ≤ R kxk < ∞.
Conversely, suppose that sup |hx, x∗ i| : x∗ ∈ S < ∞ for each x ∈ X.
Then the family
S = x∗ x∗ ∈S ⊆ X ∗ = B(X, F)
satisfies the hypotheses of the Uniform Boundedness Principle, so we conclude
that R = supx∗ ∈X ∗ kx∗ k < ∞. Hence S is a bounded set.
(b) Let π : X → X ∗∗ be the natural map, and note that kxk = kπ(x)k for
every x.
Suppose that S ⊆ X is bounded. Then S ∗∗ = π(x) : x ∈ S is a bounded
subset of X ∗∗ . Since X ∗ is a Banach space, we have by part (a) that for each
x∗ ∈ X ∗ ,
sup |hx∗ , π(x)i| : x ∈ S = sup |hx∗ , x∗∗ i| : x∗∗ ∈ S ∗∗ < ∞.
Conversely, suppose that sup |hx, x∗ i| : x ∈ S < ∞ for each x∗ ∈ X ∗ .
Then for each x∗ ∈ X ∗ we have
sup |hx∗ , x∗∗ i| : x∗∗ ∈ S ∗∗ = sup |hx, x∗ i| : x ∈ S < ∞.
Detailed Solutions 63
= CN kxkℓp ,
so A ∈ B(ℓp , ℓ∞ ).
2.35 ⇒. If A is surjective, then range(A) = Y, which is not meager.
⇐. Suppose that range(A) is not meager in Y. Then
S
∞
range(A) = A(BkX (0)).
k=1
By definition of nonmeager, some set A(BkX (0)) must contain an open ball
in Y. Lemma 2.26 therefore implies that A(BkX (0)) contains an open ball. But
then range(A) contains an open ball. Since range(A) is a subspace of Y, it
follows that range(A) = Y. Hence A is surjective.
Remark: The same reasoning shows that if range(A) is not closed, then it
is a meager subset of range(A).
2.36 Suppose that T : X → Y is a topological isomorphism and {xn } is com-
plete in X. Suppose µ ∈ Y ∗ and hT xn , µi = 0 for every n. Then hxn , T ∗ µi = 0
for every n. Since {xn } is complete, it follows that T ∗ µ = 0. But then for
every y ∈ Y we have
kxk = kT −1 T xk ≤ kT −1 k kT xk,
Since {yn }n∈N is Cauchy in Y, this implies that {xn }n∈N is Cauchy in X.
Since X is complete, there must be some x ∈ X such that xn → x. Since L is
continuous, this implies that yn = Lxn → Lx. But we also have yn → y, so we
conclude that y = Lx ∈ range(L). Hence range(L) is closed. The hypotheses
also imply that L is injective, so L is a linear bijection of X onto the Banach
space range(L), and therefore is a topological isomorphism by the Inverse
Mapping Theorem.
⇐. Assume that range(L) is closed and L : X → range(L) is a topological
isomorphism. Then by Exercise 2.37 we have kT −1 k−1 kxk ≤ kT xk ≤ kT k kxk
for all x ∈ X.
Detailed Solutions 65
2.39 (a) Suppose that λ ∈ ℓ∞ , i.e., λ is a bounded sequence. Then for any f
we have
∞
X ∞
X
|λn hf, en i|2 ≤ kλk2∞ hf, en i|2 = kλk2∞ kf k2 < ∞,
n=1 n=1
and
∞
X ∞
X
S (I − T ) = S − ST = Tn − T n = I.
n=0 n=1
66 Detailed Solutions
hg, νi = hT −1 g, µi, g ∈ Y.
Thus T ∗ ν = µ, so T ∗ is surjective.
Hence, we have shown that T ∗ : Y ∗ → X ∗ is a bounded bijection. The
Inverse Mapping Theorem therefore implies that T is a topological iso-
morphism. Alternatively, to prove this directly, note that if we apply the
above reasoning to the topological isomorphism T −1 : Y → X then we have
that (T −1 )∗ : X ∗ → Y ∗ is a bounded bijection. Therefore, if we have that
(T −1 )∗ = (T ∗ )−1 , then we can conclude that T ∗ is a topological isomorphism.
Therefore, fix µ ∈ X ∗ and g ∈ Y. Then g = T f for some f ∈ X, so we
have
Detailed Solutions 67
g, (T −1 )∗ µ = T −1 g, µ = hf, µi
= f, T ∗ (T ∗ )−1 µ
= T f, (T ∗ )−1 µ = g, (T ∗ )−1 µ .
Hence kT ∗ µk ≤ kµk.
Also, since T −1 : Y → X is an isometric isomorphism,
AA† y = AB −1 P y = AB −1 p = BB −1 p = p = P y.
Thus AA† = P.
(c) By Theorem 2.32, we have that range(A∗ ) is closed, and there-
fore range(A∗ ) = ker(A)⊥ . Let Q be the orthogonal projection of H onto
range(A∗ ). Given x ∈ H, write x = p + e where p = Qx ∈ range(A∗ ) and
e ∈ ker(A). Then since Ap ∈ range(A) we have P Ap = Ap, so
A† Ax = A† Ap = B −1 P Ap = B −1 Ap = B −1 Bp = p = Qx.
Thus A† A = Q.
2.44 (a) If x ∈ ker(A† ) then B −1 P x = A† x = 0. Since P x ∈ range(A) and
B −1 : range(A) → ker(A)⊥ is a topological isomorphism, this implies that
P x = 0. Since P is the orthogonal projection onto range(A), we conclude that
x ∈ range(A)⊥ .
Conversely, if x ∈ range(A)⊥ then P x = 0 and therefore A† x = B −1 P x =
0, so x ∈ ker(A† ).
(b) Note that P : K → range(A) is surjective and B −1 → range(A) →
ker(A)⊥ is a topological isomorphism, so
AA† y = y = ACy.
AA† y = 0 = ACy.
= sup kA−1/2 xk
kA1/2 xk=1
Hence
2.46 Suppose that {fn }n∈N is a Cauchy sequence with respect to ||| · |||. Since
kfm − fn kX ≤ |||fm − fn |||, it follows that {fn }n∈N is Cauchy with respect to
the original norm in X. Since X is complete, it follows that there exists some
f ∈ X such that fn → f.
Similarly, kAfm − Afn kY ≤ |||fm − fn |||, so {Afn }n∈N is Cauchy in Y.
Since Y is complete, there exists some g ∈ Y such that Afn → g.
The hypotheses therefore imply that g = Af. Hence
Hence Z x
f (x) = g(t) dt + f (0),
0
Detailed Solutions 71
= Kε + 0.
w
Since ε is arbitrary, we conclude that hxn , zi → 0. Hence xn → 0. The general
case follows by replacing xn with xn − y.
w
Case p = 1. If p = 1 then the “⇒” argument remains valid, i.e., xn → y in
1
ℓ implies that xn converges componentwisePto y and sup kxn k < ∞.
n
However, the converse fails. Set xn = n1 k=1 δk . Then kxn k1 = 1 for all n
and xn converges componentwise to 0. However, xn does not converge weakly
to 0, for if we take z = (1, 1, 1, . . . ) ∈ ℓ∞ then we have hxn , zi = 1 →
/ h0, zi.
72 Detailed Solutions
(b) Since ℓp is reflexive for 1 < p < ∞, we only have to consider the cases
p = 1 and p = ∞.
w*
Case p = 1. ⇒. Suppose that xn −→ y in ℓ1 = c0 ∗ . Since δk ∈ c0 , we
therefore have that
N
X ∞
X
≤ |z(k)| |xn (k)| + ε |xn (k)|
k=1 k=N +1
N
X
≤ |z(k)| |xn (k)| + ε kxn k1
k=1
N
X
≤ |z(k)| |xn (k)| + Kε.
k=1
Therefore
X
N
lim |hz, xn i| ≤ lim |z(k)| |xn (k)| + Kε = Kε.
n→∞ n→∞
k=1
w*
Since ε is arbitrary, we conclude that hz, xn i → 0. Hence xn −→ 0. The general
case follows by replacing xn with xn − y.
w* ∗
Case p = ∞. ⇒. Suppose that xn −→ y in ℓ∞ = ℓ1 . Since ek ∈ ℓ1 , we
therefore have that
N
X ∞
X
≤ |z(k)| |xn (k)| + K |z(k)|
k=1 k=N +1
N
X
≤ |z(k)| |xn (k)| + Kε.
k=1
Therefore
X
N
lim |hz, xn i| ≤ lim |z(k)| |xn (k)| + Kε = Kε.
n→∞ n→∞
k=1
w*
Since ε is arbitrary, we conclude that hz, xn i → 0. Hence xn −→ 0. The general
case follows by replacing xn with xn − y.
2.52 Let {xn } be an orthonormal sequence in a Hilbert space H, and choose
any y ∈ H. Then by Bessel’s Inequality, we have
∞
X
|hxn , yi|2 ≤ kyk2 < ∞,
n=1
w
so hxn , yi → 0. Hence xn → 0.
Detailed Solutions to Exercises from Chapter 3
P P
3.1 Theorem 1.49 tells us that cn xn converges if and only if |cn |2 < ∞.
However, the latter sum is a series of nonnegative scalars, so if is converges
then it converges absolutely
P and hence unconditionally.
P Thus, if σ is P any bi-
jection of N, then |2 < ∞ and only if
|cnP |cσ(n) |2 < ∞. Hence cn xn
converges if and only if cσ(n) xσ(n) converges for every bijection σ.
P
3.2 We are given a conditionally convergent series cn of real scalars.
(a) Let (pn ) be the sequence of nonnegative terms of (cn ) in order,Pand
let (qP
n ) be the sequence of negativeP terms of P(cn ) in order. If bothP pn
and qn converge, then If both pn and (−qn ) converge, so (pn −
qn ) converges. But this is a series containing only nonnegative terms, so it
converges absolutely. Therefore P every rearrangement of this P sum converges,
One of these rearrangements is |cn |, so we conclude that cn converges
absolutely and hence unconditionally, which is a contradiction.
P
(b) Fix x ∈ R. Since cn converges, we must have cn → 0. Let (pn ) be
the positive terms of (cn ) in decreasing order, and let (qn ) be the negative
terms of (cn ) in increasing order. Since (pn ) diverges, there is a first integer
M1 such that
S1 = p1 + · · · + pM1 ≥ x.
Then since (qn ) diverges, there is a first integer N1 such that
Continuing in this way (and interleaving any cn that are zero), we obtain a
bijection σ of N, and the corresponding partial sums converge to x because
the terms pn , qn decrease to zero.
(c),
P (d) The proof of Lemma 3.3 constructs a permutation σ of N such
that cσ(n) diverges to ∞. Modifying that proof slightly, we can find an M1
such that
p1 + · · · + pM1 > 1,
then an N1 such that
converge to f, so we have
fnM +1 − fn1 → f as M → ∞.
Hence
fnM → g = f + fn1 as M → ∞.
Thus {fn }n∈N has a subsequence {fnk }k∈N that converges to g ∈ X.
But {fn }n∈N is also Cauchy sequence, so given ε > 0, there exists a K > 0
such that kg − fnk k < ε for all nk > K. Also, there exists an N such that
kfm − fn k < ε for all m, n > N. Suppose that n > N. Then since the nk
76 Detailed Solutions
are strictly increasing, there exists some nk greater than both K and N. For
this nk we have
Hence fn → g, so X is complete.
P P
3.4 Assume that m n kxmn k < ∞. Choosing any ordering σ of N × N,
P P
we have that k kxσ(k) k < ∞. Thus x = k xσ(k) converges absolutely in X,
hence converges regardless of the ordering. For each n, the series
X
ym = xmn
n
so the series X
y = ym
m
→ 0 as N → ∞.
in turn. Then
P the preceding calculation says that a subsequence of the partial
sums of k xσ(k) converge to y. However, we know that this series converges
and equals x, so we conclude that y = x and therefore
XX X
xmn = y = x = xσ(k) .
m n k
Further, the final series on the line above is independent of the choice of σ.
Detailed Solutions 77
P∞ Ak
so the series eA = k=0 k! converges absolutely in operator norm. Since
B(X) is a Banach space under the operator norm, we conclude that the series
converges in B(X), hence eA ∈ B(X).
(c) Let sN , tN be the N th partial sums for the series defining eA and eB ,
respectively:
XN XN
Ak Bk
sN = and tN = .
k! k!
k=0 k=0
A B
Then since sN → e , tN → e , and {tN }N ∈N is a bounded sequence (since
it converges), we have that
∞ X
X k
k! kAj B k−j k
=
j! (k − j)! k!
k=0 j=0
78 Detailed Solutions
X∞ X∞
kAkj kBkk−j
≤
j=0
j! (k − j)!
k=j
∞
X ∞
kAkj X kBkk−j
=
j=0
j! (k − j)!
k=j
∞
X ∞
kAkj X kBkk
=
j=0
j! k!
k=0
∞
X kAkj
= ekBk
j=0
j!
∞ X
X k
k! Aj B k−j
=
j! (k − j)! k!
k=0 j=0
X∞ X∞
Aj B k−j
=
j=0
j! (k − j)!
k=j
∞
X ∞
Aj X B k−j
=
j=0
j! (k − j)!
k=j
∞
X ∞
Aj X B k
=
j=0
j! k!
k=0
∞
X j
A B
= e
j=0
j!
= eA eB .
and similarly (eiA )∗ eiA = I. In particular, eiA is bijection. Hence eiA is uni-
tary by Exercise 2.22.
3.6 We are given an orthonormal set {xi }i∈I in a Hilbert space H.
(a) Fix x ∈ H. Suppose that hx, xi i =
6 0 for uncountably many i. Let
(c) ⇒ (a). If hx, xi i = 0 for every i then statement (c) implies that x = 0,
and therefore {xi }i∈I is complete.
′
3.8 (a) We are given that T (µ) = hxn , µi ∈ ℓp for every µ ∈ X ∗ . Consider
the case 1 < p ≤ ∞, so 1 ≤ p′ < ∞. For each N > 0, define
TN (µ) = hx1 , µi, . . . , hxN , µi, 0, 0, . . . .
′
Then TN : X ∗ → ℓp is linear, and
N
X N
X
′ ′ ′ ′ ′
kTN (µ)kpℓp′ = |hxn , µi|p ≤ kxn kp kµkp = CN kµkp ,
n=1 n=1
where
N
X ′
CN = kxn kp < ∞.
n=1
Hence,
∞
X N
X
′ ′ ′
kT (µ)kpℓp′ = |hxn , µi|p = lim |hxn , µi|p
N →∞
n=1 n=1
′
= lim kTN (µ)kpℓp′
N →∞
′ ′ ′ ′
≤ sup kTN kp kµkp = B p kµkp .
N
′ ′
Therefore T (µ) ∈ ℓp , and T : X ∗ → ℓp is bounded, with kT k ≤ B.
For the case p = 1, we have p′ = ∞ and
where
CN = sup kxn k.
1≤n≤N
Hence,
Detailed Solutions 81
kT (µ)kℓ∞ = sup |hxn , µi| = sup kTN (µ)k ≤ sup kTN k kµk = B kµk.
n N N
N
X
≤ sup |λn cn | |hxn , µi|
kµk=1 n=M+1
X
N 1/p X 1/p′
p p′
≤ sup |cn | |hxn , µi|
kµk=1 n=M+1 n
X
N 1/p
p
≤ sup B kµk |cn |
kµk=1 n=M+1
X
N 1/p
p
= B |cn | ,
n=M+1
X ′
Uc = cn xn ∈ X and T µ = hxn , µi ∈ ℓp .
n
Therefore
X
hc, U ∗ µi = hU c, µi = cn xn , µ
n
X
= cn hxn , µi
n
D E
= (cn ), hxn , µi = hc, T µi.
Therefore T = U ∗ .
(e) Now assume that 1 < p < ∞ and X is reflexive. Then 1 < p′ < ∞, so
p′ ∗ ′
(ℓ ) = ℓp . Therefore, since T : X ∗ → ℓp , we have that T ∗ : ℓp → X ∗∗ = X.
Given c ∈ ℓp and x ∈ X, we have
X D X E
hx, T ∗ ci = hT x, ci = hx, xn i cn = x, cn xn = hx, U ci,
n n
P
where we have used the fact that the series U c = cn xn converges in norm.
Hence U = T ∗ .
∞
3.9 (a) Let {δn } be the sequence
P of standard basis vectors in ℓ . Then for
every finite set F ⊆ N we have k n∈F δn kℓ∞ = 1.PThus R = 1 (and similarly
RE = RΛ = 1). However, by the same reasoning, δn is not a Cauchy series,
hence does not converge in ℓ∞ .
Detailed Solutions 83
(b) The same argument as in part (a) is valid in the space c, which is
separable.
3.10 Suppose thatP0 < A ≤ kxn k ≤ B < ∞ for some sequence {xn } in a
Hilbert space H. If cn xn converges unconditionally, then we have by Orlicz’s
Theorem that X X
A |cn |2 ≤ kcn xn k2 < ∞,
n
2
so (cn )ℓ .
2
On
P the other hand, if we fix any nonzero vector x ∈ H, then (1/n) ∈ ℓ
but x/n does not converge.
P
3.11 We are given complex-valued functions in Lp (E) such that fn con-
verges
P unconditionally. Write f n = g n + ih n where g n , h n are real-valued. Let
f = fn and similarly write f = g + ih. Given any permutation σ of N, we
have
XN XN
g −
gσ(n)
≤
f − fσ(n)
p,
n=1 Lp n=1 L
P P
so g = gn converges unconditionally, and similarly h = hn converges
unconditionally. Using the fact that the result has been proved for real-valued
functions, we therefore have that
Z X ∞ p/2
|fn (x)|2 dx
E n=1
Z X
∞ p/2
= |gn (x)|2 + |hn (x)|2 dx
E n=1
Z X
∞ ∞
X p/2
2 2
= 2 max |g n (x)| , |h n (x)| dx
E n=1 n=1
Z X
∞ p/2 X
∞ p/2
p/2 2 2
= 2 max |gn (x)| , |hn (x)| dx
E n=1 n=1
Z X
∞ p/2 Z X
∞ p/2
p/2 2 p/2 2
≤ 2 |gn (x)| dx + 2 |hn (x)| dx
E n=1 E n=1
< ∞.
Alternatively, we could simplify the proof by appealing to the Triangle In-
equality in ℓp/2 , noting that 21 ≤ p2 ≤ 1.
P
3.12 Suppose that f = cn fn converges unconditionally in Lp (E). Applying
Orlicz’s Theorem, we have
X X
A2 |cn |2 ≤ kcn fn k2Lp < ∞.
n n
84 Detailed Solutions
P
Hence |cn |2 < ∞.
P PN
Since f = cn fn converges in P Lp (E), the partial sums SN =P n=1 cn fn
converge in Lp -norm to f. Since |cn |2 < ∞, we know that g = cn fn con-
verges in L (E). Hence SN converges in L2 -norm to g. Since Lp convergence
2
Z X p/p
= |fn (x)|p dx
E n
Z X p/2
≤ |fn (x)|2 dx
E n
< ∞,
N X
X ∞
= |hxn , ak i| < ∞.
k=1 n=1
P
Therefore xn converges absolutely.
Detailed Solutions to Exercises from Chapter 4
4.1 Let
S = A ⊆ X : A is finitely linearly independent .
Note that S is nonempty since any singleton {x} with x 6= 0 is linearly inde-
pendent. The inclusion relation ⊆ is a partial order on S.
Suppose that C is a chain in S, say C = {Ai }i∈I where I is some index set.
By definition, each set Ai is finitely independent, and we claim that A = ∪Ai
is also finitely independent. To see this, suppose that x1 , . . . , xn ∈ A. Then
for each k = 1, . . . , n we have xk ∈ Aik for some ik ∈ I. Since C is a chain, it
is linearly ordered by inclusion. Therefore there is a largest set Aik , i.e., there
is a j such that Aik ⊆ Aij for k = 1, . . . , n. Hence x1 , . . . , xn belong to Aij
and therefore are independent. Therefore A ∈ S, and since we have Ai ⊆ A
for each i ∈ I, the set A is an upper bound for the chain C.
Therefore, Zorn’s Lemma implies that S contains a maximal element B.
By definition, B is finitely independent, so if its finite span is X then it is a
Hamel basis for X. Suppose that there exists some x ∈ X \ span(B). Then
B ′ = B∪{x} is finitely independent and hence belongs to S. However, B ( B ′ ,
which implies that B is not a maximal element in S. This is a contradiction,
so we must have X = span(B).
4.2 (a) First proof. Let X be an infinite-dimensional Banach space. By defini-
tion, any Hamel basis Hamel basis for X is infinite. Suppose that {fn }n∈N was
a countable Hamel basis for X. Since {fn }n∈N is finitely linearly independent,
we must have fn 6= 0 for every n, so kfn k 6= 0 for every n.
For each N ∈ N, define
FN = span{f1 , . . . , fN }.
But then we have am (x) 6= 0 for infinitely many m, contradicting the fact
that only finitely many am (x) can be nonzero.
(b) Suppose that S is an infinite-dimensional subspace of X that has a
countable Hamel basis {fn }n∈N . Then just as in part (a) we can write S =
∪Fn where Fn = span{f1 , . . . , fN } is closed because it is finite dimensional.
Hence S is a meager subset of X.
(c) We can write S
Cc (R) = C[−N, N ],
N ∈N
(a) Let {δn } be the standard basis for ℓ1 (any Banach space with a basis
can be substituted in this example). Then {δ1 , . . . , δN } is a Hamel basis for
E = span{δ1 , . . . , δN }. Let M = span{δn }n>N , and let {yi }i∈I be a Hamel
basis for M. Since every vector in X can be written uniquely as x = e + m
where e ∈ E and m ∈ M, it follows that {δ1 , . . . , δN } ∪ {yi }i∈I is a Hamel
basis for X. In fact, every vector x ∈ X can be written uniquely as
X
x = x1 δ1 + · · · + xN δN + bi (x) yi .
i∈I
kt (x, µ(x))kY = k(tx, µ(tx))kY = ktxk = |t| kxk = |t| k(x, µ(x)kY .
Therefore Y is a normed
space.
Suppose that (xn , µ(xn )) is a Cauchy sequence in Y. Then {xn } is a
Cauchy sequence in X, so there exists some x ∈ X such that xn → x. But
then
k(x, µ(x)) − (xn , µ(xn ))kY = kx − xn kX → 0,
so Y is complete.
(b) Since µ is unbounded, we can find xn → x in X such that µ(xn ) does
not converge. Hence, although (xn , µ(xn )) converges to (x, µ(x)) in Y, it does
not converge in the norm of X1 . Therefore the embedding I(x) = x of Y into
X1 is not continuous.
4.5 Let {xi }i∈I be a basis for R over the field Q. Choose any countable
subsequence J = {j1 , j2 , . . . } of I. Define f (xjn ) = n for n ∈ N and f (xi ) = 0
for i ∈ I\J0 . Given a nonzero x ∈ R, by definition of Hamel basis we can
PN
write x = k=1 ck xik for some unique P i1 , . . . , iN ∈ I and unique nonzero
N
rational scalars c1 , . . . , cN . Define f (x) = k=1 ck f (xik ). This f is a Q-linear
function on R, i.e., f (x + ry) = f (x) + rf (y) for all x, y ∈ R and r ∈ Q.
If f was R-linear, then we would have f (x) = ax for some a ∈ R. Hence
m = f (xjm ) = axjm and n = f (xjn ) = axjn , so a 6= 0 and
a a
xj − xj = 0.
m m n n
Therefore
1 1
xj − xj = 0,
m m n n
so {xi }i∈I is not Q-independent, which is a contradiction. Hence f cannot be
R-linear.
P
4.6 (a) Given xP∈ X, there exist unique scalars (cn ) such that x = cn xn .
Therefore x = (cn /λn ) (λn xn ),Pso we need only show P that this represen-
tation is unique. If we have x = dn (λn xn ), then x = (dn λn ) xn , so by
uniqueness dn λn = cn for every n, and therefore dn = cn /λn . Hence {λn xn }
is a basis.
(b) Note that the partial sums associated with {λn xn } are
N
X N
X
SN x = (cn /λn ) (λn xn ) = cn xn ,
n=1 n=1
which are exactly equal to the partial sums associated with {xn }. Similarly,
if we consider a different ordering of the index set, the partial sums do not
change. Hence if {xn } is unconditional, then so is {λn xn }.
90 Detailed Solutions
Consequently,
X∞ X∞
hx, en i hx, en i
c1 = hx, e1 i + n−1
= .
n=2
2 n=1
2n−1
N
X N
X
hzN , µi = cn hδn , µi = |hδn , µi|.
n=1 n=1
Therefore
∞
X ∞
X
|xn − x0 | |hδn , µi| ≤ 2 kxk ℓ∞ |hδn , µi| ≤ 2 kxkℓ∞ kµk < ∞,
n=1 n=1
where ∞
X
y0µ = hδ0 , µi − hδn , µi and ynµ = hδn , µi.
n=1
Since
|hx, µi| ≤ kxk∞ kykℓ1 ,
we have µ ∈ c∗ . If m ∈ N then
∞
X
hδm , µi = 0 · y0 + δmn yn = ym ,
n=1
so ∞
X
y0 = hδ0 , µi − hδm , µi.
n=1
Therefore T µ = y, so T is surjective.
4.9 We are given yn = (1, . . . , 1, 0, 0, . . . ). Suppose that x ∈ c0 , and set
cn = xn − xn+1 , n ∈ N.
PN
Set SN x = n=1 cn yn . Then
N
X X
N N
X
SN x = cn y n = cn , cn , . . . , cN , 0, 0, . . . ,
n=1 n=1 n=2
and therefore
Detailed Solutions 95
N
X N
X
x − SN x = x1 − cn , x2 − cn , . . . , xN − cN , xN +1 , xN +1 , . . .
n=1 n=2
= (xN +1 , xN +1 , . . . ),
so
kx − SN xkℓ∞ = |xN +1 | → 0.
Consequently,
∞
X
x = cn y n ,
n=1
Hence
∞
X ∞
X ∞
X ∞
X
c1 y 1 = cn y n − cn yn = x1 − x2 = dn yn − dn yn = d1 y1 .
n=1 n=2 n=1 n=2
For m > N,
X
N N
X
(SN x)m = cn z n = x1 + (xn − xn−1 ) = xN .
n=1 m n=2
Hence
SN x = (x1 , . . . , xN , xN , xN , . . . ),
so
x − SN x = (0, . . . , 0, xN +1 − xN , xN +2 − xN , . . . ),
Now, xN → x0 , so given ε > 0 there exists an N0 such that
n ≥ N0 =⇒ |xn − x0 | < ε.
1 1
= z1 − z2 + z3 − · · · .
2 3
The vector x belongs to c since
n
X (−1)k+1
lim = ln 2.
n→∞ k
k=1
However,
Detailed Solutions 97
1 1 1 1
z1 + z2 + · · · + zn = 1, 1 + , . . . , 1 + · · · + , 0, 0, . . . ,
2 n 2 n
which does not converge in c. Hence the basis representation of this x does
not converge unconditionally, so {zn } is not an unconditional basis for c0 .
4.11 (a) Fix 1 ≤ p < ∞, and let {δn } be the standard basis for ℓp . Given
x ∈ ℓp and N ∈ N, we have
N
X N
X
kSN xkpp = |xn |p ≤ |xn |p = kxkpp .
n=1 n=1
= (x0 , x0 , . . . ) + (x1 − x0 , . . . , xN − x0 , 0, 0, . . . )
= (x1 , . . . , xN , x0 , x0 , . . . ).
Hence T is well defined, and furthermore the calculation above shows that
kT (y)k ≤ kyk1 , so T is bounded.
Suppose that T (y) = 0. Then, by the uniqueness of the coefficients in a
basis expansion, we must have yn /kxn k = 0 for every n, which implies yn = 0
for every n and hence y = 0. Thus T is injective.
Let {an } denote the coefficient functionals
P associated with the basis {xn }.
Suppose that x ∈ X, so we have x = an (x) xn . Set yn = an (x) kxn k. Then,
∞
X ∞
X
|yn | = |an (x)| kxn k < ∞,
n=1 n=1
(b) Suppose that {yn }n∈N is Cauchy with respect to ||| · |||. For each N we
have
kSN ym − SN yn k = kSN (ym − yn )k ≤ |||ym − yn |||.
Since {yn }n∈N is Cauchy with respect to ||| · |||, we conclude that {SN yn }n∈N
is Cauchy with respect to k · k. However, X is complete with respect to k · k,
so this implies that {SN yn }n∈N converges with respect to k · k, i.e., there
exists some zN ∈ X such that limn→∞ kzN − SN yn k = 0. Since each SN yn
belongs to range(SN ) = span{x1 , . . . , xN }, which is closed since it is finite
dimensional, we must have zN ∈ span{x1 , . . . , xN } as well.
(c) Fix ε > 0. Then since {yn }n∈N is Cauchy with respect to ||| · |||, there
exists an n0 such that
M, N > N0 =⇒ kSM yn − SN yn k ≤ ε.
Hence,
N
X N
X
ak (zN ) xk = zN = SN zN +1 = ak (zN +1 ) xk .
k=1 k=1
= fNk (xNk
= kxNk − SNk xNk k
S
M
K ⊆ Bε (yk ).
k=1
∀ N ≥ N0 , ∀ k = 1, . . . , M, kyk − SN yk k ≤ ε.
Given x ∈ K, there exists some k such that kx−yk k < ε. Hence for all N ≥ N0
we have
∀ N ≥ N0 , sup kx − SN xk ≤ (C + 2) ε.
x∈K
Thus SN → I uniformly on K.
(b) Suppose that Y has the approximation property and X is an arbi-
trary Banach space. Let T : X → Y be a compact operator. Then there exist
finite-rank operators TN ∈ B(Y ) that converge uniformly to the identity I on
compact subsets of Y. Let B = {x ∈ X : kxk ≤ 1} be the closed unit sphere
in X. Since T is compact, D = T (B) is a compact subset of Y. Therefore
kT − TN T k = sup kT x − TN T xk
kxk≤1
= sup ky − TN yk
kxk≤1,y=T x
≤ sup ky − TN yk
y∈D
→ 0 as N → ∞.
4.19 Exercise 2.41 implies that Y is a Banach space and Lemma 4.18 implies
that {yn } is a basis for Y. Since we are given that T xn = yn for every n, the
two bases {xn } and {yn } are equivalent by definition.
4.20 (c) Suppose {xn } is an absolutely convergent basis for a Banach space X
that is equivalent
P to a basis {yn } for a Banach space Y. Given y ∈ Y, we can
write y = cn yn . Let T : X → Y be a topological isomorphismP such that
T xn = yn for every n. By Theorem 4.20, it follows that x = P cn xn , where
x = T −1 y. Since {xn } is absolutely convergent, we must have |cn | < ∞.
Therefore
X X X X
kcn yn k = |cn | kT xn k ≤ kT k |cn | kxn k = kT k kcn xn k < ∞.
T x = (x0 , x1 − x0 , x2 − x0 , . . . ).
so T is bounded.
Suppose that T x = 0. Then by definition of T we have x0 = 0, and hence
0 = T x = (0, x1 , x2 , . . . ). Therefore x = 0, so T is injective.
Let y ∈ c0 be given. Then y0 = 0, so if we set x = (0, y1 , y2 , . . . ) then we
have x ∈ c and T x = y. Hence T is injective.
104 Detailed Solutions
T δ0 = (1, 1 − 1, 1 − 1, . . . ) = (1, 0, 0, . . . ) = δ1 ,
and
T δn = (0, . . . , 0, 1 − 0, 0, 0, . . . ) = δn+1 .
Hence T maps the basis {δn }n≥0 onto the basis {δn }n∈N , so these two bases
are equivalent.
(b) Suppose that x ∈ c0 and kxk∞ = 1. Then |xn | ≤ 1 for every n and
limn→∞ xn = 0. Hence there is some k such that |xk | = 1 and some n such
that |xn | < 1. Then we can find scalars yn , zn not equal to xn and with |yn |,
|zn | < 1 such that xn = yn +z
2
n
. Let ym = xm and zm = xm for all m 6= n. For
m = k in particular we have |yk | = |zk | = |xk | = 1, so kyk∞ = kzk∞ = 1.
Now consider the unit ball in c, and consider the vector x = (1, 1, . . . ).
Suppose that y, z ∈ c satisfy kyk∞ = kzk∞ = 1 and y+z 2 = x. Then yn + zn =
2xn = 2 for every n. But |yn |, |zn | ≤ 1 for every n, so this is only possible if
yn = zn = 1. To see this explicitly, write yn = an + ibn and zn = cn + idn .
Then we have an + cn = 2 and bn + dn = 0, but −1 ≤ an , cn ≤ 1, so this
implies an = cn = 1 and hence bn = dn = 0.
(c) Suppose that T : c0 → c was an isometric isomorphism. Fix x ∈ c0 with
kxk∞ = 1, and let y 6= z ∈ c0 be as constructed in part (b). Then kT xk∞ =
kT yk∞ = kT zk∞ = 1 since T is an isometry. Further, T x = (T y + T z)/2
since T is linear. But then part (b) implies that T y = T x = T z. Since T is
invertible, it follows that y = x = z, which is a contradiction. Therefore no
such isometric isomorphism can exist.
4.23 We have seen that f ∈ C[0, 1] can be written
n
∞ 2X
X −1
f = aχ + bℓ + cn,k sn,k ,
n=0 k=0
with uniform convergence of this series. Every element of the Schauder system
except χ vanishes at t = 0, so f (0) = aχ(0) = a. Hence a is unique. Now
consider n
X∞ 2X −1
f − aχ = bℓ + cn,k sn,k .
n=0 k=0
Each element of the right-hand side vanishes at t = 1 except for the function ℓ.
Therefore f (1) − aχ(1) = bℓ(1), so b = f (1) − a = f (1) − f (0). Hence b is
unique. Now consider
n
∞ 2X
X −1
f − aχ − bℓ = cn,k sn,k .
n=0 k=0
Detailed Solutions 105
Each element of the right-hand side vanishes at t = 1/2 except for s0,0 . There-
fore
f (1/2) − aχ(1/2) − bℓ(1/2) = c0,0 s0,0 (1/2),
so
c0,0 = f (1/2) − a − b = f (1/2) − f (0) − f (1).
Hence c0,0 is unique. Continuing in this way we obtain the uniqueness of the
representation of f.
4.24 (a) Suppose that fn ∈ C(T) and fn → g uniformly. Then g is contin-
uous. Since we have fn (t + 1) = fn (t) for every t, it follows from pointwise
convergence that g is 1-periodic. Hence g ∈ C(T), so C(T) is a closed sub-
space of C(R) and therefore is a Banach space with respect to the uniform
norm.
(b) Let S = {f ∈ C[0, 1] : f (0) = f (1)}. Define T : C(T) → S by T f =
f |[0,1] . Clearly T is linear and isometric. Given g ∈ S we have g(1) = g(0).
Therefore if we set f (t+k) = g(t) for t ∈ [0, 1) and k ∈ Z, then g is a 1-periodic
function and T f = g. Hence T is surjective, so is an isometric isomorphism.
(c) The proof that Lp (T) and Lp [0, 1] are isometrically isomorphic is sim-
ilar.
(d) Consider fα (t) = tα where α ∈ R. If αp + 1 > 0 then
Z 1 Z 1 1
p αp tαp+1 1
|fα (t)| dt = t dt = = < ∞.
0 0 αp + 1 0 αp + 1
On the other hand, if αp + 1 = 0 then
Z 1 Z 1
|fα (t)|p dt = t−1 dt = ∞,
0 0
DX
N E D N
X E
lim hx, an i xn , x∗ = lim hx, an i xn , x∗
N →∞ N →∞
n=1 n=1
DX
∞ E
= hx, an i xn , x∗ = hx, x∗ i.
n=1
P
It therefore remains only to show that the Prepresentation x = hx, an i xn
is unique. Suppose that we also had x = cn xn , with weak convergence of
m. Then since am ∈ X ∗ , we have by the weak
this series. Fix any particularP
convergence of the series x = cn xn that
DX
N E
hx, am i = lim cn xn , am
N →∞
n=1
N
X
= lim cn hxn , am i
N →∞
n=1
N
X
= lim cn δnm = cm .
N →∞
n=1
Hence the representation is unique, and therefore {xn } is a weak basis for X.
4.26 (a) Recall that weakly convergent sequences are bounded
P (Theo-
rem 2.38). Therefore, if (cn ) ∈ Y then k(cn )kY < ∞ since cn xn =
PN
limN →∞ n=1 cn xn converges weakly. The remainder of the proof is now
identical to the proof of Theorem 4.12(a).
(b) Suppose
P that {xn } is a weak basis for X. Define the map T : Y → X
by T (cn ) = cn xn , where this series converges weakly. This mapping is well
defined by the definition of Y. It is clearly linear, and it is bijective because
{xn } is a weak basis. Finally, if (cn ) ∈ Y then
∞
N
X
X
kT (cn )k =
cn xn
≤ sup
cn xn
N
= k(cn )kY ,
n=1 n=1
hx1 , y1 i = hδ1 , y1 i = 1,
hxn , y1 i = (−1)n δ1 + δn , y1 = 0,
hxm , yn i = (−1)m δ1 + δm , δn = δmn ,
Therefore
∞ ∞
X X
kz − SN zkℓ1 =
z n δn − (−1)n+1 zn δ1
n=N +1 n=N +1 ℓ1
X
∞
∞
X
≤
z n δn
+ |zn |
n=N +1 ℓ1 n=N +1
→ 0 as N → ∞.
N
X −1 X
∞
= z 1 δ1 + z n δn − z k δN
n=2 k=N +1
N
X −1 X
∞
= z n δn − z k δN .
n=1 k=N +1
Therefore
∞ X
∞
X
kz − SN zkℓ1
=
z n δn + z k δN
n=N k=N +1 ℓ1
∞
∞
X
X
≤
z δ
n n
+ |zk |
n=N ℓ1 k=N +1
→ 0 as N → ∞.
= yN → 0 as N → ∞.
Hence weak* representations of vectors with respect to {xn } are not unique,
so {xn } is not a weak* basis for ℓ1 .
Detailed Solutions to Exercises from Chapter 5
5.1 ⇒. Suppose there is a unique sequence {an } ⊆ X ∗ that is biorthogonal
to {xn }. We know that {xn } is minimal by part (a), so it remains to show
that {xn } is complete. Suppose that x∗ ∈ X ∗ is a continuous linear functional
such that hxn , x∗ i = 0 for every n. Then
kx − qn kLp ≤ kx − qn k∞ → 0 as n → ∞,
where the series converges in norm. Applying the Pythagorean Theorem, the
square of the norms of the partial sums of this series are
N
X
2
tN =
c0 e1 + (cn en + cn en+1 )
n=1
= kc0 e1 + c1 e1 + c1 e2 + c2 e2 + c2 e3 + · · · + cN eN + cN eN +1 k2
en = n (fn − e1 ) = n (fn − f1 ).
Therefore
X∞ XN
|cn |2 |cn |2
0 ≤ = lim ≤ lim ksN k2 = 0.
n=2
n2 N →∞
n=2
n2 N →∞
This is only possible if cn = 0 for all n ≥ 2. But then we are left with c1 e1 = 0,
so we must also have c1 = 0 since e1 6= 0. Thus F is ω-independent.
5.4 Define
yn = nen − (n + 1)en+1 .
Then for m, n ∈ N we have
m D
X E
ek
hxm , yn i = , nen − (n + 1)en+1
k
k=1
m
X m
X
nδk,n (n + 1)δk,n+1
= −
k k
k=1 k=1
0 − 0, n > m,
= 1 − 0, n = m,
1 − 1, n < m,
= δmn .
Thus {xn } is bounded above and below in norm. The same is not true of {yn }
since the Plancherel Equality tells us that
kyn k2 = n2 + (n + 1)2 .
Therefore {xn } cannot be a Schauder basis since if it was then its dual se-
quence would have to be bounded above and below P as well.
Also, {yn } is incomplete, because if we set x = en /n, then for each m
we have
X∞
1
x, yn = hem , nen − (n + 1)en+1 i = 1 − 1 = 0.
m=1
m
x1 , . . . , xN ∈ range(PN ).
Pn xN = Pn PN −1 xN = Pn 0 = 0.
Therefore
xN ∈ ker(P1 ) ∩ · · · ∩ ker(PN −1 ).
Combining these statements, we see that
114 Detailed Solutions
(
xm , m ≤ n,
Pn xm =
0, m > n.
P
We claim that {xn } is ω-independent. Suppose that cn xn = 0. Then
since P1 is continuous,
X X
0 = P1 cn xn = cn P1 xn = c1 x1 .
n n
Let
X
M1
rk
=
cm,k xm
m=1
N
X X
M1 X
M2
≤
dn,k yn
pk ⊗ qk − cm,k xm ⊗
k=1 m=1 n=1 HS
N
X M1
X
N
X M2
X
M1
X
≤
pk − c x
kq k +
q − d y
c x
m,k m
k
k n,k n
m,k m
k=1 m=1 k=1 n=1 k=1
N
X N
X
ε ε
≤ kqk k + rk
N kqk k N rk
k=1 k=1
= 2ε.
Z
By the uniqueness statement in Exercise B.9(b), to show that SN 2 =
X Y
SN ⊗ SN it suffices to check that equality holds when these operators are
applied to simple tensors x ⊗ y. We compute that
2
N
X
Z
SN 2 (x ⊗ y) = x ⊗ y, ck zk
k=1
N X
X N
= x ⊗ y, am ⊗ bn (xm ⊗ yn )
m=1 n=1
N X
X N
= hx, am i hy, bn i (xm ⊗ yn )
m=1 n=1
X
N X
N
= hx, am i xm ⊗ hy, bn i yn
m=1 n=1
X Y
= SN x ⊗ SN y
X Y
= SN ⊗ SN (x ⊗ y).
Therefore equality of operators holds, and Exercise B.9(b) also tells us that
Z
X
Y
S 2
≤
SN
kSN k ≤ CX CY .
N
X Y X X Y
= SN ⊗ SN (x ⊗ y) + SN +1 − S N ⊗ S ℓ (x ⊗ y).
Therefore
Z
S 2
≤
S X ⊗ S Y
+
(S X − S X ) ⊗ S Y
N +ℓ N N N +1 N ℓ
X
Y
X
≤
SN
S
+
S
N
X
Y
N +1 − SN Sℓ
≤ CX CY + 2CX CY = 3 CX CY .
X Y X X Y
= SN ⊗ SN + (SN +1 − SN ) ⊗ SN +1 (x ⊗ y)
ℓ
X
+ hx, aj i hy, bN +1 i (xj ⊗ yN +1 )
j=1
X Y X X Y
= SN ⊗ SN + (SN +1 − SN ) ⊗ SN +1 (x ⊗ y)
ℓ
X
+ hx, aj i xj ⊗ hy, bN +1 i yN +1
j=1
X Y X X Y
= SN ⊗ SN + (SN +1 − SN ) ⊗ SN +1 (x ⊗ y)
X ℓ
+ hx, aj i xj ⊗ hy, bN +1 i yN +1
j=1
X Y X X Y
= SN ⊗ SN + (SN +1 − SN ) ⊗ SN +1 (x ⊗ y)
+ (SℓX x) ⊗ (SN
Y Y
+1 y − SN y)
X Y X X Y
= SN ⊗ SN + (SN +1 − SN ) ⊗ SN +1 (x ⊗ y)
+ SℓX ⊗ (SNY Y
+1 y − SN y) (x ⊗ y).
Therefore
118 Detailed Solutions
Z
X
S 2
≤ 3 CX CY +
(SN X Y
+1 − SN ) ⊗ Sℓ
N +N +1+ℓ
X
≤ 3 CX CY +
SN X
Y
+1 − SN Sℓ
≤ 3 CX CY + 2CX CY = 5 CX CY .
5.9 Without loss of generality, we consider g(t) = |t|α and e
g (t) = |t|−α and
the corresponding systems of weighted exponentials in L2 (T) = L2 [− 21 , 12 ].
(a) Since 0 < α < 1/2, we have 0 < 2α < 1 and 1 − 2α > 0. Therefore
Z 1/2 Z 1/2 1/2
2 2t2α+1 2 2−2α−1
|t|α dt = 2 t2α dt = = < ∞,
−1/2 0 2α + 1 0 2α + 1
and
Z 1/2 Z 1/2 1/2
2 2t1−2α 2 22α−1
|t|−α dt = 2 t−2α dt = = < ∞.
−1/2 0 1 − 2α 0 1 − 2α
α −α
Hence |t| and |t| both belong to L2 (T).
Further,
α 2πimt
|t| e , |t|−α e2πint = e2πimt , e2πint = δmn ,
so {|t|α e2πint }n∈Z and {|t|−α e2πint }n∈Z are biorthogonal.
Suppose that f ∈ L2 (T) and f, |t|α e2πint = 0 for every n. Then we have
f (t) |t|α , e2πint = 0 for every n. The function f (t) |t|α belongs to L1 (T),
and Fourier coefficients of functions in L1 (T) are unique, so f (t) |t|α = 0 for
every n. Therefore f (t) |t|α = 0 a.e., which implies that f = 0 a.e. hence
{|t|α e2πint }n∈Z is complete. A similar argument shows that {|t|−α e2πint }n∈Z
is complete.
Note: In fact, since |t|α is bounded, the proof for the completeness of
α 2πint
{|t| e }n∈Z only needs to appeal to the fact that the exponentials are
complete in L2 (T). However, |t|−α is unbounded, so this case requires more
care.
(b) It suffices to consider intervals I that contain the origin, since that is
the only zero or singularity of g or eg . Since functions in L2 (T) are 1-periodic,
it also suffices to consider intervals of length at most 1. So, fix I = [−a, b]
where 0 < a, b < 1.
Because 0 < 1 − 2α < 1 + 2α < 2, we have that
Z b Z a Z b
1 1
|t|−2α dt = t−2α dt + t−2α dt
a + b −a a+b 0 0
1−2α a
1−2α b
1 t
= + t
a + b 1 − 2α 0 1 − 2α 0
1 a1−2α + b1−2α
= .
a+b 1 − 2α
Detailed Solutions 119
Combining with a similar calculation for the reciprocal exponent, we find that
Z Z
1 2 1 1
|g(t)| dt dt
|I| I |I| I |g(t)|2
Z b Z b
1 2α 1 −2α
= |t| dt |t| dt
a + b −a a + b −a
1 1
= 2
(a1+2α + b1+2α ) (a1−2α + b1−2α )
(a + b) (1 + 2α) (1 − 2α)
Now, since 0 < a < 1 and 1 < 1 + 2α < 2, we have a1+2α ≤ a1 = a, and also
a1−2α ≤ 1. Hence we can continue equation (A) as follows:
1 a2 + a + b + b 2 1
(A) ≤ 2 2 2
≤ .
1 − 4α a + 2ab + b 1 − 4α2
Since this quantity is independent of a and b, we conclude that |g(t)|2 = |t|2α
is an A2 weight.
5.10 Let BC = {xn } and BR = {x1 , ix1 , x2 , ix2 , . . . }. Let us write R-span
and C-span to denote spans with respect to the real or complex field, etc.
⇒. Suppose that BC is a basis for X. Then BC is C-complete, so it follows
that BR is R-complete. Also, we have that xn 6= 0 for every n, so to prove
that BR is a basis we can use Theorem 5.17.
Given real scalars an , bn , set cn = an + ibn . Let C be the basis constant
for BC . Then for M ≤ N we have
M M
M
X X
X
an xn +
bn ixn
=
cn xn
n=1 n=1 n=1
N
X
≤ C
c x
n n
n=1
X
N
= C
(an + ibn ) xn
n=1
X
N N
X
= C
a x
n n + b ix
n n
.
n=1 n=1
meaning that the partial sums of this series converge to x. Considering the
“even” partial sums, we see that
N −1
N
X
X
kf k∞ =
c g
n n
≤
c g
n n
= kf + cN gN k∞ (A).
n=1 ∞ n=1 ∞
(c) We have
2 |c|p = 21−p |2c|p = 21−p |(c + cN ) + (c − cN )|p ≤ |c + cN |p + |c − cN |p .
′
5.15 (a) Since hn is bounded, it belongs to Lp [0, 1] = Lp [0, 1]∗ for each
1 ≤ p < ∞. Since hhm , hn i = δmn , the Haar system is its own biorthogonal
system.
122 Detailed Solutions
(b) By definition, {hn } is complete in its closed span in L∞ [0, 1], and
hn 6= 0 for every n. We will appeal to Theorem 5.17 to show that {hn } is a
basis for its closed span.
Given N ∈ N and scalars c1 , . . . , cN , just as in the discussion in Section 5.5,
if we set
N
X −1 XN
gN −1 = cn h n and gN = cn h n
n=1 n=1
then gN −1 and gN agree except possibly on the dyadic interval I where hN is
nonzero. Let I1 , I2 denote the left and right halves of I. Then gN −1 takes a
constant value c on I, while gN = c + dN on I1 and gN = c − dN on I2 for an
appropriate constant dN . Hence
sup |gN −1 (t)| = |c| and sup |gN (t)| = max{|c + dN |, |c − dN |}.
t∈I t∈I
Since
|2c| = |c + dN + c − dN | ≤ |c + dN | + |c − dN |,
we must have at least one of |c + dN | ≥ |c| or |c − dN | ≥ |c|. Hence |gN −1 (t)| ≤
/ I. Consequently, kgN −1 kL∞ ≤
|gN (t)| for each t ∈ I, and they are equal for t ∈
kgN kL∞ , so Theorem 5.17 implies that {hn } is a basis for its closed span.
(c) The characteristic function of every dyadic interval [ 2kn , k+1
2n ) belongs to
span{hn }. If f ∈ C[0, 1] then f is uniformly continuous, so we can approximate
it as closely as we like in the uniform norm by a step function of the form
n
2X −1
g = c k χ[ k
, k+1 .
2n 2n )
k=0
= 2−1−α K 2−nα−n/2 ,
In essence, the conclusion holds for all n large enough, and those k such that
ψn,k is “localized at x.”
5.17 (a) Assume that {xn } is a bounded basis for X. Then, by defini-
tion, 0 < inf kxn k ≤ sup kxn k < ∞. Further, by Theorem 4.13 we have
1 ≤ kan k kxn k ≤ 2C, where C is the basis constant for {xn }. Therefore
0 < inf kan k ≤ sup kan k < ∞. Combined with Theorem 5.21, this implies
that {an } is a bounded basis for its closed span in X ∗ .
(b) Consider the Babenko example {|t|1/4 e2πint }n∈Z , which is a Schauder
basis for L2 [− 21 , 12 ]. Note that
Z 1/2 Z 1/2 1/2
1/4 2πint 2 2t3/2 4 1 21/2
|t| e dt = 2 t1/2 dt = 2 = = .
3 0 3 2 3/2 3
−1/2 0
Since these two quantities are not reciprocals, if we normalize {|t|1/4 e2πint }n∈Z
then the dual system will not be normalized.
5.18 Let {xn } be a minimal system in a Banach space X with a biorthogonal
system {an } that is a basis for X ∗ . Then {π(xn )} ⊆ X ∗∗ is biorthogonal to
{an } ⊆ X ∗ , so Theorem 5.21 implies that {π(xn )} is a basis for its closed
span in X ∗∗ .
Suppose that µ ∈ X ∗ and hxn , µi = 0 for every n. Since µ ∈ X ∗ and {an }
is a basis for X ∗ , we have that
X
X
µ = µ, π(xn ) an = hxn , µi an = 0.
n n
N
N
X
X
cn (xn − yn )
=
hx, an i (xn − yn )
n=1 n=1
N
X
≤ |hx, an i| kxn − yn k
n=1
N
X
≤ kxk kan k kxn − yn k
n=1
X
N
≤ λ kxk =
c x
n n
.
n=1
[x] = {y ∈ V : x ∼ y} = {y ∈ V : x − y ∈ M }
= {x − m : m ∈ M }
= {x + m : m ∈ M } = x + M.
so ψ is linear.
If ψ(x + M ) = 0 then T x = 0 so x ∈ M and hence x + M = M, the zero
element of X/M. Hence ψ is injective.
By definition, range(ψ) = X/M, so ψ is surjective.
(d) Let M be a subspace of a vector space V, and let N ⊆ V be a sub-
space such that M ∩ N = {0} and M + N = V. Then each x ∈ V can be
written uniquely as x = m(x) + n(x) where m(x) ∈ M and n(x) ∈ N. The
mapping n : V → N is linear and surjective, and n(x) = 0 if and only if
x = m(x) ∈ M, so ker(n) = M. By the Isomorphism Theorem, there exists
a linear bijection ψ : V /M → N. Since linear bijections are vector space iso-
morphisms, they preserve dimension. Hence dim(N ) = dim(V /M ), and this
number is independent of N.
Detailed Solutions to Exercises from Chapter 6
6.1 (a) ⇒ (b). Assume that {xn } is an unconditional basis for X,P and let σ be
any permutation of N. Choose any xP ∈ X. Then the series x = hx, an i xn
converges unconditionally, so x = hx, aσ(n) i xσ(n) converges by Corol-
lary 3.11. We must show that this is the unique
P representation of x in terms
of {xσ(n) }. Suppose that we also had x = cn xσ(n) for some scalars (cn ).
Then, since {xn } and {an } are biorthogonal and aσ(m) is continuous, we have
X X X
hx, aσ(m) i = cn hxσ(n) , aσ(m) i = cn δσ(n),σ(m) = cn δnm = cm .
n n n
6.5 (a) Fix 1 ≤ p < 2 and set en (t) = e2πint . Suppose that {en }n∈Z was
an unconditional basis for Lp (T). Since the exponentials are biothogonal to
P
themselves, it follows that f = n∈Z fb(n) en converges unconditionally for
each f ∈ Lp (T). However, since ken kLp = 1 for every n, we have by Theo-
P b
rem 3.28, which is a corollary to Orlicz’s Theorem, that |f (n)|2 < ∞ and
2 2 p
f ∈ L (T). Since L (T) is a proper subset of L (T) when p < 2, this is a
contradiction.
(b) Fix 2 < p < ∞. If {en }n∈Z is an unconditional basis for Lp (T), then
since Lp (T) is reflexive we would have that its biorthogonal system, which is
′
{en }n∈Z , would be an unconditional basis for Lp (T) by Exercise 6.4. This
contradicts part (a).
6.6 It is not hard to see that ≤ is a partial order on X and that requirements
(a) and (b) in Definition 3.35 are satisfied.
P P
(c) Given
P x = anP xn and y = bn xn in X, set cn = max{an , bn }. Since
the series
P an xn and
P bn yn converge unconditionally, Theorem P 3.10 implies
that |an |xn and |bn |xn converge unconditionally. Hence (|an |+ |bn |)xn
converges unconditionally.
P Since |cn | ≤ |an |+|bn |, it follows from Theorem 3.10
that z = cn xn converges unconditionally. We will show that z is a least
upper bound for x, y.
Since an , bn ≤ cn , by definition of the ordering we have x, y ≤ z. P Suppose
that x, y ≤ w. Since {xn } is an unconditional basis, we can write w = dn xn .
By definition of the order, an , bn ≤ dn for every n. Hence cn = max{an , bn } ≤
dn , so z ≤ w. Therefore z is a least upper bound. The construction of a greatest
lower bound is similar.
P P
(d)
PFix x = an xn and y = bn xn in X. As above, we know that
p = |an |xn converges unconditionally. Further, since scalars are real we
have |aPn | = max{an , −an }, so p = |x| in the sense of Definition 3.35. Similarly,
|y| = |bn |xn .
Suppose that |x| ≤ |y|, which means that |an | ≤ |bn | for every n. Then by
criterion (c) of Theorem 6.7, for each N we have
N
N
X
X
an xn
≤ KΛ
bn xn
.
n=1 n=1
6.7 No. Let h1 = χ[0,1] and h2 = χ[0,1/2) − χ[1/2,1) . Using the ordering from
Exercise 6.6 we have
128 Detailed Solutions
However, the functions χ[0,1] and 2χ[0,1/2) are not comparable using the point-
wise a.e. comparison.
6.8 (a) We have a1 = 1, a2 = 3/2, and
aN −1 + aN −2
aN = 1 + .
2
We claim that aN ≥ N/2 for every N. This is true for N = 1 and N = 2, and
if it is true for all integers between 1 and N for some N ≥ 2 then we have
aN + aN −1 1 N N − 1
aN +1 = 1 + ≥ 1+ +
2 2 2 2
N 1 N 3 N +1
= 1+ − = + ≥ .
2 4 2 4 2
Hence the claim follows by induction, and therefore aN → ∞.
(b) If we set bN = hN (µN ), then we have
bN −1 + bN −2
+ 1, N odd,
2
bN =
bN −1 + bN −2 − 1, N even.
2
We claim that 1 ≤ bN ≤ 2 for N odd and −1 ≤ bN ≤ 0 for N even.
We have b1 = 1 and b2 = −1/2. Suppose the claim holds for all integers
between 1 and N for some N ≥ 2. If N is odd then N + 1 and N − 1 are even,
so
bN + bN −1 2+0
bN +1 = −1 ≤ −1 = 1−1 = 0
2 2
and
bN + bN −1 1−1
bN +1 = −1 ≥ − 1 = −1.
2 2
If N is even then N + 1 and N − 1 are odd, so
bN + bN −1 2+0
bN +1 = +1 ≤ +1 ≤ 1+1 = 2
2 2
and
bN + bN −1 1−1
bN +1 = −1 ≥ + 1 = 1.
2 2
In particular, |bN | ≤ 2 for every N.
Since hN is piecewise linear, its global max and min must occur at one of
its knots, which are all located at t = µn for some 1 ≤ n ≤ N. Hence
khN k∞ = max{|bn | : n = 1, . . . , N } ≤ 2.
Detailed Solutions to Exercises from Chapter 7
7.1 Let {en } and {fn } be orthonormal bases for a Hilbert space H.
(a) {e1 , e2 , e2 , e3 , e3 , e3 , . . . } is norm-bounded above but is not Bessel.
(b) {e1 + e2 , e2 , e3 , . . . } is a Bessel sequence that is a nonorthogonal basis
for H.
(c) {2−1/2 en , 2−1/2 fn } is a Bessel sequence since
X 1X 1X
|hx, 2−1/2 en i|2 + |hx, 2−1/2 fn i|2 = |hx, en i|2 + |hx, fn i|2
n
2 n 2 n
= kxk2 ,
On the other hand, if finitely many elements are removed from {2−n em }m,n∈N ,
then it still contains a multiple of every vector en , and hence is complete.
Moreover, a similar calculation shows that the frame operator is S = I.
(e) {nen } is anP
unconditional basis
P for H, but it is not a Bessel sequence.
For example, x = en /n ∈ H, but |hx, nen i|2 = ∞.
(f) Consider the Babenko example, {|t|α e2πint }n∈Z where 0 < α < 1/2,
which is a conditional basis for H = L2 (T) = L2 [− 21 , 21 ]. Each element of this
basis has the same norm, so except for a scaling factor it is a normalized basis.
Set en (t) = e2πint and fn (t) = |t|α en (t). Given f ∈ H, we have g(t) =
α
|t| f (t) ∈ H, and
Z Z
hf, fn i = f (t) |t|α en (t) dt = g(t) en (t) dt = hg, en i.
However, Z r
kgr k2L2 = 1 dt = r.
0
Therefore
1 X
|hfr , hn i|2 = 2r−1/2 → ∞ as r → 0+ .
kgr k2L2
n∈Z
≤ B lim kyk k2
k→∞
= B kxk2 .
PN
Since this is true for every N, we conclude that n=1 |hx, xn i|2 ≤ B kxk2 , so
{xn } is a Bessel sequence.
(a) ⇒ (c). Suppose that {xn } is a Bessel sequence with Bessel bound B.
Then by Theorem 7.2, the synthesis operator R for {xn } is bounded, with
Detailed Solutions 131
P
2
kRk ≤ B 1/2 . We therefore have
cn xn
≤ B kck2ℓ2 for all c ∈ ℓ2 , which
implies statement (c).
(c) ⇒ (d). Suppose that statement (c) holds. Fix c = (cn ) ∈ ℓ2 . Given
M < N, set d = (0, . . . , 0, cM+1 , . . . , cN , 0, 0, . . . ). Then
X
2
N
N
X
c x
= kRdk22 ≤ kRk2 kdk22 = kRk2 |cn |2 .
n n
ℓ ℓ
n=M+1 n=M+1
P
Therefore cn xn is Cauchy, and therefore converges.
P
(d) ⇒ (e). Suppose statement (d) holds. Then T c = cn xn converges for
PN
each c ∈ ℓ2 , so T : H → ℓ2 and T is clearly linear. Define TN c = n=1 cn xn .
Then
N
X X
N 1/2 X
N 1/2
2 2
kTN ck ≤ |cn | kxn k ≤ |cn | kxn k ≤ CN kckℓ2 ,
n=1 n=1 n=1
7.5 Suppose {xn } is Bessel with Bessel bound B, and let {en } be the standard
basis for ℓ2 . The synthesis operator for {xn } is bounded, so kxn k2 = kRen k2 ≤
B ken k2ℓ2 = B. Hence all Bessel sequences are bounded. Alternatively, we note
that for any m we have
X
kxm k4 = |hxm , xm i|2 ≤ |hxm , xn i|2 ≤ B kxm k2 .
n
(b) No. Let {en } be an orthonormal basis for H and set xn = nen . The
sequence {xn } is not a Bessel sequence. However, if we set yn = en /n then
{yn } is complete, and for each m ∈ N we have
X X
|hym , xn i|2 = |hem /m, nen i|2 = |hem /m, memi|2 = 1.
n n
P P
7.9 (a) We are given that K = m n |hxn , xn i|2 < ∞. Let G = [hxn , xm i]
be the Gram matrix for {xn }. Given c ∈ c00 , applying Cauchy–Bunyakovski–
Schwarz we have
X X 2
2
kGckℓ2 = hxn , xm i cn
m n
X X X
≤ |hxn , xm i|2 |cn |2
m n n
= K kck2ℓ2 .
X X 2
1/2 1/2
≤ |hxn , xm i| |hxn , xm i| |cn |
m n
X X X
≤ |hxn , xm i| |hxn , xm i| |cn |2
m n n
X X
≤ K |hxn , xm i| |cn |2
m n
X X
= K |hxn , xm i| |cn |2
n m
X
≤ K2 |cn |2
n
= K kck2ℓ2 .
2
7.11 By Theorem 7.4 there exist orthonormal bases {em }, {fn } for H, K and
bounded maps U ∈ B(H), V ∈ B(K) such that U em = xm and V fn = yn for
all m, n. By Theorem B.10, {em ⊗fn }m,n∈N is an orthonormal basis for H ⊗K.
By Exercise B.9 there exists a unique bounded operator U ⊗ V ∈ B(H ⊗ K)
that satisfies (U ⊗ V )(x ⊗ y) = U x ⊗ V y for all x ∈ H, y ∈ K. By Exercise 7.6
we know that bounded operators map Bessel sequences to Bessel sequences,
so
{xm ⊗ fn }m,n∈N = {U em ⊗ V fn }m,n∈N = (U ⊗ V )(em ⊗ fn ) m,n∈N
is Bessel.
7.12 (a) The (m, n)-entry of L is
(
1/m, m ≥ n,
Lmn =
1/n, m ≤ n.
Hence CC ∗ = L.
Detailed Solutions 135
(b) We have
I − (I − C)(I − C)∗ = I − I + C + C ∗ − CC ∗ = C + C ∗ − CC ∗ .
By inspection, this is the diagonal matrix with entries 1, 1/2, . . . , and therefore
it is a positive matrix.
(c) Let M = (I −C)(I −C)∗ . Then I −M ≥ 0, so we have h(I −M )x, xi ≥ 0
for every x. Therefore
hM x, xi ≤ kIxkx = kxk2 .
kM k = sup hM x, xi ≤ 1.
kxk=1
(d) We have
Hence
kLk = kCC ∗ k = kCk2 ≤ 22 = 4.
7.13 This follows from Exercise 3.1 and the definition of Riesz bases.
7.14 Let {x1 , . . . , xd } be a basis for a finite-dimensional Hilbert space V. Then
it has a dual basis {y1 , . . . , yd }. Both of these are complete Bessel sequences,
so it follows from Theorem 7.13 that they are Riesz bases.
Another approach is to note that {x1 , . . . , xd } is a bounded unconditional
basis, and hence is a Riesz basis.
Yet another is to observe that
d
X
hGc, ci =
cn xn
.
n=1
∞
N
X
X
≤ kTk k
c a e
n nk n
k=1 n=1
∞
X X
N 1/2
2
= kTk k |cn ank |
k=1 n=1
Detailed Solutions 137
∞
X X
N 1/2
≤ kTk k sup |ank | |cn |2
n
k=1 n=1
X
N 1/2
= λ |cn |2 .
n=1
while Z 1 2k
2 1
kTk fε k2L2 = x2k dx ≥ −ε · ε,
1
2 −ε
2
so 2k
2 kTk fε k2L2 1
kTk k ≥ ≥ −ε .
kfε k2L2 2
Since ε > 0 is arbitrary, this implies that kTk k2 ≥ 2−2k . Taking square roots
and combining with the converse estimate, we see that kTk k = 2−k .
(b) Expanding e2πinx in a Taylor series, we see that pointwise we have
en (x) − eλn (x) = e2πinx − e2πiλn x
= e2πinx 1 − e2πi(λn −λ)x
∞ k
X 2πi(λn − n)
2πinx
= e 1− xk
k!
k=0
X
∞ k
2πinx 2πi(λn − n)
= −e xk
k!
k=1
∞ k
X 2πi(λn − n)
= − xk e2πinx
k!
k=1
∞
X
= ank Tk en (x).
k=1
138 Detailed Solutions
= eπδ < ∞.
In fact, the same calculations show that the series converges absolutely in
L∞ [− 21 , 12 ]. Therefore the pointwise and L2 -limits are the same a.e., so
∞
X
en − eλn = ank Tk en ,
k=1
Applying Exercise 7.19, if λ < 1 then {eλn }n∈Z is a Riesz basis for L2 [− 21 , 12 ].
This happens if eπδ − 1 < 1, or
eπδ < 2.
X3 √ 2 √ 2
3 1 3 1
2 2
|hv, xn i| = |v2 | + − v1 − v2 + v1 − v2
n=1
2 2 2 2
3 1
= |v2 |2 + 2 |v1 |2 + 2 |v2 |2
4 4
3 3 3
= |v1 |2 + |v2 |2 = kvk2 .
2 2 2
8.2 Let A, B be frame bounds for {xn } and let C be a Bessel bound for {yn }.
Then for each x ∈ H we have
X
A kxk2 ≤ |hx, xn i|2
n
X X
≤ |hx, xn i|2 + |hx, yn i|2 ≤ B kxk2 + C kxk2 .
n n
≥ A lim kyk k2
k→∞
= A kxk2 .
P∞
Since this is true for every N, we conclude that n=1 |hx, xn i|2 ≥ A kxk2 , so
{xn } is a frame.
8.5 (a) We know that the frame operator is bounded and self-adjoint. We
will show that kS − AIk = 0. Since {xn }n∈N is a tight frame,
X
hSx, xi = |hx, xn i|2 = A kxk2 .
n
140 Detailed Solutions
Therefore,
(S − AI)x, x = hSx, xi − A hx, xi = A kxk2 − A kxk2 = 0,
so
kS − AIk = sup (S − AI)x, x = 0.
kxk=1
Consequently, X
Ax = Sx = hx, xn i xn .
n
Since
|hx, xm i|2 ≤ kxk2 kxm k2 ,
we therefore have that
X
|hx, xn i|2 = A kxk2 − |hx, xm i|2
n6=m
≥ A kxk2 − kxk2 kxm k2 = A − kxm k2 kxk2 .
1
We must show that the scalars cn = A hx, xn i are unique. Assume that we
P
also had x = dn xn , and fix any m ∈ N. Then by orthogonality,
DX
∞ E
hx, xm i = dn xn , xm
n=1
∞
X
= dn hxn , xm i = dm hxm , xm i = dm kxm k2 = A dm .
n=1
1
Therefore dm = A hx, xm i = cm . Thus we have uniqueness, so {xn }n∈N is a
Schauder basis for H.
142 Detailed Solutions
(c) ⇒ (d). This follows immediately from the fact that every basis is ω-
independent.
(d) ⇒ (b). Assume that {xn }n∈N is ω-independent. This implies in par-
ticular that xn 6= 0 for every n. Fix m ∈ N. Then since S = AI,
∞ ∞
1 1 X X
1
xm = Sxm = hxm , xn i xn = A hxm , xn i xn .
A A n=1 n=1
f = χ[0,0+ε] − χ[ 1b , 1b +ε] .
Then Z Z 1
ε b +ε
2πibnt −2πibnt
hf, e i = e dt − e−2πibnt dt = 0.
1
0 b
it follows from Exercise 8.7 that {e2πibnt }n∈Z is not a basis, and it also follows
from Exercise 8.6 that {e2πibnt }n6=m is a frame for any m.
Second, {e2πibnt }n∈Z is not an orthogonal sequence, because
144 Detailed Solutions
Z 1
he2πimbt , e2πibnt i = e2πi(m−n)bt dt
0
1
e2πi(m−n)bt
=
2πi(m − n)b 0
e2πi(m−n)b 1
= − 6= 0.
2πi(m − n)b 2πi(m − n)b
More precisely, this quantity is not zero for every choice of m and n, e.g.,
consider m = 1 and n = 0. Exercise 8.7 again implies we do not have a basis.
Third, we show directly that the constant function f = 1 does not have
a unique representation in terms of the exponentials {e2πibnt }. Since f (t) =
e2πi0bt , one expansion is
X
f (t) = δmn e2πibnt .
n∈Z
Another expansion is provided by the tight frame property (see Exercise 8.5):
1 X
f (t) = hf, e2πibnt i e2πibnt .
A
n∈Z
1/2 1/2
≤ B kf k2 + K kf k2
= (B 1/2 + K 1/2 ) kf k.
Detailed Solutions 145
Then,
X
∞ 1/2 X
∞ 1/2 X
∞ 1/2
2 2 2
|hf, yn i| ≥ |hf, xn i| − |hf, xn − yn i|
n=1 n=1 n=1
1/2 1/2
≥ A kf k2 − K kf k2
= (A1/2 − K 1/2 ) kf k.
Since A > K, this gives us a lower frame bound for {yn } of (A1/2 − K 1/2 )2 .
P
(b) We are given a sequence {zn } such that K = khn k2 < ∞. Using
Cauchy–Bunyakovski–Schwarz, we compute
X X X
|hx, zn i|2 ≤ kxk2 kzn k2 = kxk2 kzn k2 = K kxk2 .
n n n
We can estimate these norms using that fact that |eit − 1| ≤ |t| for every t:
Z 1
ke2πibnt − e2πiλn t k22 = |e2πibnt − e2πiλn t |2 dt
0
Z 1
= |e2πi(bn−λn )t − 1|2 dt
0
Z 1
≤ |2π(bn − λn )t|2 dt
0
4π 2
= |bn − λn |2 .
3
Therefore, {e2πiλn t }n∈Z will be a frame if
146 Detailed Solutions
X 3
|bn − λn |2 < .
n
4π 2 b
Surprisingly (perhaps), this result is not even close to being optimal. This
result requires the that λn become closer and closer to nb as |n| increases. The
optimal result does not require this, only that the λn have a certain maximum
distance from nb. Here is the optimal result for b = 1, the case where you get
both a frame and a basis.
Kadec’s 41 -Theorem. If supn |n − λn | ≤ L < 41 , then {e2πiλn t }n∈Z is an
exact frame for L2 [0, 1].
This result is optimal in the sense that the theorem is false if L = 1/4. Nat-
urally, the proof is more difficult than the proof of our result, but it isn’t ter-
ribly complicated. If you’re interested, look in the nicely written book [You01]
by Young.
8.11 Fix vectors v1 , . . . , vN ∈ Fd . Let h·, ·i denote the dot product on Fd ,
and let AH denote the Hermitian of a matrix A.
(a) If x ∈ Fd then
N
X N
X
|hx, xn i|2 ≤ kxk2 kxn k2 = B kxk2 ,
n=1 n=1
d
X
≤ λd |hx, wj i|2 = λd kxk2 ,
j=1
so it is a quasibasis.
8.14 ⇒. This is proved in Theorem 8.13.
(b) ⇒ (a). Assume that statementP (b) holds. By continuity of the inner
product, it follows that hSx, xi = |hx, xn i|2 . Since we have AI ≤ S ≤ BI,
we therefore have Then hAIx, xi ≤ hSx, xi ≤ hBIx, xi for every x ∈ H, which
implies that {xn } is a frame for H.
8.15 We are given positive operators U, V ∈ B(H) with U ≤ V. Since U
and V are self-adjoint, we have by Theorem 2.15 that
kU k = sup hU x, xi ≤ sup hV x, xi = kV k.
kxk=1 kxk=1
0 ≤ BI − S ≤ B− AI = (B − A) I.
Therefore
1 B−A
0 ≤ I− S ≤ I.
B B
Also,
(B + A) I − 2S = BI + AI − 2S ≤ BI + AI − 2AI = BI − AI = (B − A) I,
so
2 B−A
I− S ≤ I.
A+B B+A
Since 0 < A ≤ B < ∞, using part (a), we therefore have that
1
B−A B−A
I − S
≤ kIk = < 1
B B B
Detailed Solutions 149
and
2
B−A B−A
I − S
≤ kIk = < 1.
A+B B+A B+A
By Neumann series, we therefore have that S is a topological isomorphism.
8.16 (a) We are given a frame {xn } with an alternative dual {yn } that is a
Bessel sequence. 2
Given y ∈ H, we have P hy, xn i ∈ ℓ since {xn } is a frame. Since {yn }
is Bessel, the sequence hy, xn i yn therefore converges, and hence for every
x ∈ H we have
X X X
hx, yi = hx, yn i xn , y = hx, yn i hxn , yi = x, hy, xn i yn .
n n n
P
Therefore y = hy, xn i yn .
Let A, B be frame
P bounds for {yn }. Given y ∈ H, since {yn } is a dual of
{xn } we have y = hy, yn i xn . Therefore
X 2
2 2
kyk = sup |hy, xi| = sup hy, yn i xn x
kxk=1 kxk=1 n
X X 2
2
≤ sup |hy, yn i| |xn x|
kxk=1 n n
X
≤ sup |hy, yn i|2 B kxk2
kxk=1 n
X
= B |hy, yn i|2 .
n
Thus {yn } has a lower frame bound of B −1 and hence is a frame. By the
above work, {xn } is a dual for {yn }, so {xn } is an alternative dual frame for
{yn }.
(b) ⇒. Suppose that {yn } = {e xn } is the canonical dual for {xn }. Then it
is certainly an alternative dual, and given x ∈ H we have
e = hx, x
Cx en i = hx, S −1 xn i = hS −1 x, xn i = CS −1 x.
e
Since S is a topological isomorphism, it follows that range(C) = range(C).
⇐. Suppose that {yn } is a Bessel sequence that is an alternative dual to
e Since {yn } is Bessel, by definition
{xn } that satisfies range(C) = range(C).
of alternative dual we have
X
x = e = C ∗ Cx,
hx, yn i xn = RCx x ∈ H.
n
Hence
e = I = C
C∗C e ∗ C.
Therefore
e ∗C
CC e = C
e
e ∗ z = z for z ∈ range(C)
and hence CC e = range(C). Therefore
e ∗ Cx = Cx,
CC x ∈ H.
e ∗ C = C. Taking adjoints,
That is, CC
e ∗ C)∗ = C ∗ C C
C ∗ = (CC e∗ ,
so
e∗ δn = C ∗ Cyn = Syn ,
xn = C ∗ δn = C ∗ C C
where S is the frame operator for {xn }. Hence yn = S −1 xn = x
en , so {yn } is
the canonical dual frame for {xn }.
8.17 (a) Let A, B be frame bounds for {xn }. Then given x ∈ M we have
X X X
|hx, P xn i|2 = |hP x, xn i|2 = |hx, xn i|2 ≤ B kxk2 ,
n n n
and similarly for the lower frame bound. Hence {P xn } is a frame for M with
frame bounds A, B.
(b) Suppose that P S = SP. Since S is a topological isomorphism, it follows
that
S −1 P = S −1 P SS −1 = S −1 SP S −1 = P S −1 ,
so S −1 commutes with P as well. Let T be the frame operator for {P xn } as
a frame for M. Using the self-adjointness of P, given x ∈ H we compute that
X X
TPx = hP x, P xn i P xn = P hP 2 x, xn i xn
n n
= P SP x = SP 2 x = SP x.
T −1 (P xn ) = (S|M )−1 P xn = S −1 P xn = P S −1 xn = P x
en ,
en }.
so the canonical dual frame is {P x
Detailed Solutions 151
Given c ∈ ℓ2 , we have
X X X
e =
Rc en =
cn x cn S −1 xn = S −1 cn xn = S −1 Rc.
n n n
(b) Let A, B be frame bounds for {xn } and let C, P R be its analysis and
synthesis operators. Since {e xn } is a frame, the series en converges for all
cn x
e be the synthesis operator for {e
c = (cn ) ∈ ℓ2 . Let R xn }. Since B −1 , A−1 are
frame bounds for {e xn }, we have kRk e 2 ≤ A−1 . The operator
X
Pc = en , xk
cn x = hRc,e xk i = C Rc e
n k∈N
e 2 kck22 ≤ B kck22 .
kP ck2ℓ2 ≤ kCk2 kRk ℓ ℓ
A
e = C RCx
P c = C Rc e = Cx = c.
e = hRc, Rdi
hP c, di = hc, P di = hc, C Rdi e = h0, Rdi
e = 0,
Therefore {yn } is an alternative dual to {xn }. Since {yn } is also Bessel, The-
orem 8.18(a) implies that {yn } is a frame.
152 Detailed Solutions
(a) ⇒ (b). Suppose that {yn } is an alternative dual frame for {xn }, and
let V : ℓ2 → H be the synthesis operator for {yn }. Then V is bounded and
V δn = yn by definition. Further, Theorem 8.18(a) implies that {xn } is a dual
frame for {yn }, so if x ∈ H then
X
V Cx = V hx, xn i = hx, xn i yn = x.
n
Letting {txn } denote the canonical dual frame, biorthogonality implies that
X
em i =
hx, x em i = hx, ym i.
hx, yn i hxn , x
n
kyk kyk
= ≤ kT ∗ yk ≤ kT ∗k kyk = kT k kyk.
−1
kT k kT ∗ −1 k
Combining these equations we find that
154 Detailed Solutions
A kyk2
≤ hT ST ∗y, yi ≤ B kT k2 kyk2 ,
kT −1 k2
e
Since S is a topological isomorphism, it follows that range(C) = range(C).
(b) The synthesis operator R maps ℓ2 onto H. Let T be the operator R
restricted to ker(R)⊥ = range(C) = range(C).e Then T : range(C) → H is a
topological isomorphism. Since the range of T is all of H, the pseudoinverse
of R is T −1 : H → range(C). We must show that T −1 = C. e Since {e
xn } is the
e = I and RC
canonical dual, we have that RC e = I.
Note that T Ce : H → H. If x ∈ H, then
X
e = RCx
T −1 Cx e = en i xn = x.
hx, x
n
e c = CR
CT e Cx
e = Cx
e = c.
e = A−1 , as desired.
Hence C
Detailed Solutions 155
(c) The fact that kCk2 is the optimal upper frame bound for {xn } follows
from the definition:
X
|hx, xn i|2 = kCxk2ℓ2 ≤ kCk2 kxk2 ,
n
kSk = sup |hSx, xi| = sup |hC ∗ Cx, xi| = sup kCxk2 = kCk2 ,
kxk=1 kxk=1 kxk=1
Hence
em i m∈N 6∈ ℓ2 ,
(cm )m∈N = −c0 hx0 , x
which is a contradiction. Therefore we must have c0 = 0. But then
∞
X
cn xn = 0,
n=1
which implies cn = 0 for all n > 0 since {xn } is a basis. Hence {xn }n≥0 is
ℓ2 -independent.
On the other hand, {xn }n≥0 is not ω-independent because we have
∞
X
−x0 + en i xn = −x0 + x0 = 0.
hx, x
n=1
Thus T is isometric on M.
(b) ⇐ . We have that {PM δn } is a Parseval frame for M, and since T
is an isometric isomorphism it follows that the image of {PM δn } under T is
Parseval as well.
8.30 (a) ⇒. Let M and T be as in Corollary 8.29. Define K = H × M ⊥ . By
Exercise 1.40, this is a Hilbert space with respect to the natural inner product.
Define U : K → ℓ2 by U (x, c) = T −1 x + c. Since T −1 x ∈ M and c ∈ M ⊥ , we
have
kU (x, c)k2ℓ2 = kT −1 x + ck2ℓ2 = kT −1 xk2ℓ2 + kck2ℓ2 .
Since T : M → H is a topological isomorphism, it follows that U is a topolog-
ical isomorphism as well. Let V = U −1 . Given a ∈ ℓ2 , write a = b + c where
b ∈ M and c ∈ M ⊥ . Then U (T b, c) = b + c = a, so V a = (T b, c).
Let δn = bn + cn where bn = PM δn ∈ M and cn = δn − bn ∈ M ⊥ . Let
en = V δn . Then {en } is a Riesz basis for K. Further,
PH en = PH V δn = PH (T bn , cn ) = T bn = T PM δn = xn .
so {2−m/2 em d
n }m∈N,n=1,...,d is a Parseval frame for F .
(b) If {xn }n∈I is a frame that is bounded below then its associated canon-
ical Parseval frame {S 1/2 xn }n∈I is also a frame that is bounded below, so it
suffices to assume that {xn }n∈I is Parseval.
By Corollary 8.29 there exists some Hilbert space K ⊇ Fd and some or-
thonormal basis {en }n∈I for K such that xn = P en where P is the orthogonal
projection of K onto Fd . But then Exercise 1.53 implies that
158 Detailed Solutions
X X
d = dim(Fd ) = kP en k2 = kxn k2 = |I|,
n∈I n∈I
so {xn } is Parseval.
8.34 Since M and H are separable and infinite dimensional, there exists an
isometric isomorphism T : H → M. Since {PM δn } is a Parseval frame for M,
{xn } = {T −1 PM δn } is a Parseval frame for H. The analysis operator for {xn }
is
Cx = hx, xn i = x, T −1 PM δn
∗ −1
= (T ) x, PM δn
= hPM (T ∗ )−1 x, δn i
= PM (T ∗ )−1 x.
≥ AX kxk2 + AY kyk2
≥ A kxk2 + kyk2
2
= A
(x, y)
H×K ,
Hence AX = 1 is a lower frame bound for {xn }, and similarly for the other
frame bounds. Therefore {xn } and {yn } are Parseval.
(b) Assume that range(CX ) ⊕ range(CY ) = ℓ2 . Then by the calculation in
part (a) we have range(C)
= range(CX ) + range(CY ) = ℓ2 , so it follows from
Theorem 8.32 that (xn , yn ) is a Riesz basis for H × K.
A Riesz basis is an orthonormal basis if and only
if it is a Parseval frame,
so by combining with part (a) we conclude that (xn , yn ) is orthonormal if
and only if both {xn } and {yn } are Parseval.
(c) Let M be a closed subspace of ℓ2 such that both M and M ⊥ are
infinite dimensional. Then {PM δn } is a Parseval frame for M, but since M is
proper subspace this frame cannot be a Riesz basis (Exercise 8.25). Likewise
{PM ⊥ δn } is a Parseval frame for M ⊥ that is not a Riesz basis. The analysis
operator for {PM δn } is
CM x = hx, PM δn i = hPM x, δn i = PM x,
(SX ⊗ SY )(x ⊗ y) = SX x ⊗ SY y
X X
= hx, xm i xm ⊗ hy, yn i yn
m n
XX
= hxm , xi hy, yn i (xm ⊗ yn )
m n
X
= x ⊗ y, xm ⊗ yn (xm ⊗ yn )
m,n
= T (x ⊗ y).
By the uniqueness statements in Exercise B.9, we conclude that T = SX ⊗ SY
is a topological isomorphism.
This implies that {xm ⊗ yn }m,n∈N is a frame. For example, if we let R
denote the analysis operator for {xm ⊗ yn }m,n∈N , then we have T = RR∗ and
therefore R must be surjective since T is a topological isomorphism. It follows
from Theorem 8.29 that {xm ⊗ yn }m,n∈N is a frame.
(b) This follows from the identification of L2 (E × F ) with L2 (E) ⊗ L2 (F )
given in Theorem B.13.
8.37 (a) ⇒ (b). Suppose Pthat {xn }n∈F/ is a Riesz basis forP H, where F is a
2
finite subset of N. Then n∈F / c n xn converges if and only if / |cn | < ∞.
n∈F
Since F is finite, it follows that the same equivalence holds for the summation
over all n.
w
(b) ⇒ (c). We verify the next step in the construction. Since φn → 0, there
exists an m3 > m2 such that
X
N2
hφ , δ i δ
< 1.
m 3 n n
2 8
n=1 ℓ
so we can choose N3 large enough that this quantity exceeds 1/2. Similarly,
N3
N3
X
X
1 1
hφm3 , δn i P δn
hφm3 , δn i δn
2 ≤
2 + 8 → 0 + 8,
n=N2 +1 ℓ n=1 ℓ
By hypothesis, we have
X
N 1/2
= |hφmj , δn i|2
n=M
≤ kφmj kℓ2 = 1.
Fix any N1 + 1 ≤ M < N < ∞. Let j be the smallest integer such that
Nj ≤ M < Nj+1 and let ℓ be the largest integer such that Nℓ < N ≤ Nℓ+1 .
Then
N
Nj+1
X
X −1/2
c P δ
≤
j hφ , δ i P δ
n n
2
mj n n
2
n=M ℓ n=M ℓ
ℓ Nk+1
X X
+
k −1/2 hφmk , δn i P δn
k=j+1 n=Nk +1 ℓ2
X
N
+
ℓ −1/2
hφmℓ , δn i P δn
n=Nℓ +1 ℓ2
ℓ
X
≤ j −1/2 + kψk kℓ2 + ℓ−1/2
k=j+1
ℓ
X
≤ j −1/2 + k −1/2 2−k + ℓ−1/2 .
k=j+1
∞
Nk+1
2
X
X
= k −1
hφ , δ i
mk n n
δ
k=1 n=Nk +1 ℓ2
∞
X 1
≥ k −1 = ∞.
4
k=1
Detailed Solutions 163
lim hym1 , yk i = 0,
k→∞
1
we can choose vectors ym k
such that
1 ε
|hym1 , ym i| < , k > 1.
k
2k+1
Then we have X
1 ε
|hym1 , ym i| < .
k
2
k>1
1
Set m2 = ym 1
. Since
lim hym2 , yk i = 0,
k→∞
2 1
we can choose a subsequence {ym k
} of {ym k
} such that
2 ε
|hym2 , ym i| < , k > 2.
k
2k+2
Then we have
164 Detailed Solutions
X ε
2
|hym2 , ym i| < .
k
4
k>2
Therefore XX X ε
|hymk , ymj i| ≤ = ε.
2k
k j>k k
Hence XX XX
|hymk , ymj i| ≤ 2 |hymk , ymj i| ≤ 2ε.
k j6=k k j>k
8.40 (a) ⇒ (b). Assume that statement (a) holds. We are given that {xn }n6=nk
has a lower frame bound of L. Since S −1/2 is a topological isomorphsm and
kS 1/2 k2 ≤ B, Exercise 8.24 implies that {S −1/2 xn }n6=nk is a frame with
lower frame bound L/B. However, since {S −1/2 xn } is a Parseval frame, Ex-
ercise 8.6 implies that the optimal lower frame bound for {S −1/2 xn }n6=nk is
1 − kS −1/2 xnk k2 . Therefore we must have L/B ≤ 1 − kS −1/2 xnk k2 , so
2 L
enk i =
S −1/2 xnk
≤ 1 − .
hxnk , x
B
(b) ⇒ (a). Assume that D = supk hxnk , x enk i < 1. Fix any particular n.
Then 1 − kS 1/2 xnk k2 ≥ 1 − D > 0. Since 1 − kS 1/2 xnk k2 is the optimal lower
bound for {S −1/2 xn }n6=nk , we conclude that 1 − D is a lower frame bound.
Since S 1/2 is a topological isomorphism with kS −1/2 k2 ≤ 1/A, it follows
by Exercise 8.24 that {xn }n6=nk is a frame for H with lower frame bound
L = A(1 − D).
Conclusion. Suppose that statements (a), (b) hold. Then by the argument
2
above, we have that
S −1/2 xnk
≤ 1 − L/B for every k. Since {S −1/2 xn }
is a Parseval frame, Theorem 8.44 implies that there exists an infinite set
J ⊆ {nk } such that {S −1/2 xn }n∈J
/ is a frame with frame bounds L/B − ε, 1.
Since S 1/2 is a topological isomorphism, kS −1/2 k2 ≤ A, and kS 1/2 k2 ≤ B, it
follows that {xn }n∈J
/ is a frame with frame bounds LA/B − Aε, B.
RV ∗ = C ∗ V ∗ = (V C)∗ = (CU )∗ = U ∗ C ∗ = U ∗ R.
is invariant under U ∗ .
(d) Similar to part (c).
8.42 (i) Suppose that U f = λf for some nontrivial f ∈ L2 (T) and some
λ ∈ C. Then e2πibt f (t) = λf (t) a.e. Since f is not the zero function, there
must be a set E with positive measure such that f (t) 6= 0 for t ∈ E. But then
e2πibt = λ for t ∈ E. Since the set of t such that e2πibt takes a common value
is countable, this is a contradiction. Hence U has no eigenvalues.
(ii) Suppose that V c = λc for some nonzero c ∈ ℓ2 (Z) and λ ∈ C. If λ = 0
then (cn−1 )n∈Z = V c = 0, which implies c = 0, so λ = 0 is not an eigenvalue.
But then we have (cn−1 )n∈Z = V c = (λcn )n∈Z , so cn = λn c0 for n ∈ Z. We
must have c0 6= 0 since otherwise cn = 0 for every n. But if c0 6= 0 then
we cannot have c ∈ ℓ2 (Z), which again is a contradiction. Hence V has no
eigenvalues. Note that it is important here that we are doing a shift-operator
on ℓ2 (Z), not ℓ2 (N) (compare Exercise 2.19).
(iii) The adjoint of U is multiplication by e−2πibt , so the same proof as
part (a) shows that U ∗ has no eigenvalues.
(iv) The adjoint of V is the left-shift operator on ℓ2 (Z), so the same proof
as part (b) shows that V ∗ has no eigenvalues.
(b) Exercise 8.9 tells us that {e2πibnt }n∈Z is a b−1 -tight frame. Renormal-
izing, {be2πibnt }n∈Z is a Parseval frame. However, each element has norm
Theorem 8.44 therefore implies that an infinite subset can be removed yet still
leave a frame.
Detailed Solutions to Exercises from Chapter 9
9.2 (a) Suppose that f ∈ C0 (R), and choose ε > 0. Then there exists an
R > 0 such that |f (x)| < ε/2 for all |x| > R − 1. Define
f (x), |x| ≤ R + 1,
g(x) = linear, R + 1 ≤ |x| ≤ R + 2,
0, |x| ≥ R + 2.
Since translation is isometric on Lp (R), we therefore have for |a| < δ that
≤ ε + ε + ε = 3ε.
Hence Ta f → f in Lp (R) as a → 0.
9.3 If g ∈ L1 (R) then
Detailed Solutions 167
Z Z
g(x − a) e−2πiξx dx = g(x) e−2πiξ(x+a) dx
∧
(Ta g) (ξ) =
Z
= e−2πiaξ g(x) e−2πiξx dx
and
Z
e2πibx g(x) e−2πiξx dx
∧
(Mb g) (ξ) =
Z
= g(x) e−2πi(ξ−b)x dx = b
g(ξ − b) = Tb b
g(ξ).
9.4 Consider the case m = 1. Assume that f (x), xf (x) ∈ L1 (R). We must
show that the limit
′ fb(ξ + η) − fb(ξ)
fb (ξ) = lim
η→0 η
Z
e−2πi(ξ+η)x − e−2πiξx
= lim f (x) dx (A)
η→0 η
exists. We do this by applying the Lebesgue Dominated Convergence Theo-
rem. Define ex (ξ) = e−2πiξx . Then the integrand in equation (A) converges
pointwise for almost every x, because
ex (ξ + η) − ex (ξ)
lim f (x) = f (x) e′x (ξ) = f (x) (−2πix) e−2πiξx .
η→0 η
Hence the Lebesgue Dominated Convergence Theorem implies that the limit
in equation (A) exists, and that we have
Z
′ e−2πi(ξ+η)x − e−2πiξx
fb (ξ) = lim f (x) dx
η→0 η
Z
= f (x) (−2πix) e−2πiξx dx
∧
= (−2πix)f (x) (ξ).
= kf k1 kgk1 .
R
Fubini’s Theorem therefore implies that (f ∗ g)(x) = f (y) g(x − y) dy exists
for almost every x and is an integrable function of x.
(c) Fubini’s Theorem allows us to interchange integrals in the following
calculation:
Z
(f ∗ g)(x) e−2πiξx dx
∧
(f ∗ g) (ξ) =
ZZ
= f (y) g(x − y) e−2πiξx dy dx
Detailed Solutions 169
Z Z
= f (y) e−2πiξy g(x − y) e−2πiξ(x−y) dx dy
Z Z
= f (y) e−2πiξy g(x) e−2πiξx dx dy
Z
= f (y) e−2πiξy gb(ξ) dy
= fb(ξ) gb(ξ).
(d) By definition, supp(f ) is the smallest closed set such that f = 0 a.e.
off supp(f ), and by hypothesis supp(f ) is compact. Therefore, we can fix a
representative of f such that f (x) = 0 for every x ∈/ supp(f ), and likewise we
fix similar representatives of g and f ∗ g. Let
K = supp(f ) + supp(g).
R
Suppose that (f ∗ g)(x) = f (y) g(x − y) dy was nonzero. Then there must be
a y such that y ∈ supp(f ) and x − y ∈ supp(g) simultaneously, for otherwise
f (y) g(x − y) would be zero for every y and therefore f ∗ g(x) would be zero.
Hence, x = y + (x − y) ∈ supp(f ) + supp(g). This shows that
As a function of y,
f (y) g(x + h − y) − g(x − y) ≤ |f (y)| kg ′ k∞ ∈ L1 (R).
h
(f ∗ g)(x + h) − (f ∗ g)(x)
(f ∗ g)′ (x) = lim
h→0 h
Z
g(x + h − y) − g(x − y)
= lim f (y) dy
h→0 h
Z
= f (y) g ′ (x − y) dy = (f ∗ g ′ )(x).
Thus f ∗ g is differentiable.
9.6 (b) Because L1 (R) ∩ L2 (R) is dense in L2 (R), the relations from part (a)
continue to hold when we extend the Fourier transform to L2 (R). For example,
if we fix g ∈ L2 (R) then there exist functions fn ∈ L1 (R) ∩ L2 (R) such that
fn → g in L2 -norm. As the Fourier transform is unitary, we also have fbn → b g
in L2 -norm. By passing to subsequences if necessary, we can assume that these
convergences hold pointwise almost everywhere as well (see Theorem A.23).
Since (Ta fn ) = M−a fbn , it follows that (Ta g) = M−a g. This is an equality
∧ ∧
(b) Since Cc2 (R) is dense in L2 (R), given g ∈ L2 (R) we can find
fn ∈ Cc2 (R) such that fn → g in L2 -norm. Then there is a subsequence that
converges pointwise almost everywhere, so by applying part (a) and taking
∧∧
pointwise limits we obtain g (ξ) = g(−ξ) for almost every ξ.
9.9 Assume that f, g are absolutely continuous on [a, b]. Then f ′ , g ′ ∈ L1 [a, b],
so f ′ (x) g ′ (y) ∈ L1 [a, b]2 . Letting E = {(x, y) ∈ [a, b]2 : x ≤ y}, we compute
that
ZZ Z b Z y
f ′ (x) g ′ (y) dx dy = f ′ (x) dx g ′ (y) dy
E a a
Z b
= f (y) − f (a) g ′ (y) dy
a
Z b Z b
= f (y) g ′ (y) dy − f (a) g ′ (y) dy
a a
Z b
= f (y) g ′ (y) dy − f (a) g(b) − g(a) .
a
Fubini’s Theorem implies that these two integrals are equal, so by rearranging
we obtain
Z b Z b
′
f (x) g (x) dx = f (b)g(b) − f (a)g(a) − f ′ (x) g(x) dx.
a a
Detailed Solutions to Exercises from Chapter 10
∨
10.1 Fix f ∈ L2[−Ω,Ω] (R). Then f, f ∈ L2 (R) and supp(f ) ⊆ [−Ω, Ω].
∨
Consequently f is bandlimited to [−Ω, Ω] because
∨ ∧
supp( f ) = supp(f ) ⊆ [−Ω, Ω].
∨
Hence f ∈ FL2[−Ω,Ω] (R), so the inverse Fourier transform maps L2[−Ω,Ω] (R)
into FL2[−Ω,Ω] (R).
To see that this mapping is surjective, fix g ∈ FL2[−Ω,Ω](R). Then g,
g ∈ L2 (R) and supp(b
b g ) ⊆ [−Ω, Ω]. Hence gb ∈ L2[−Ω,Ω](R), and it is mapped
to g by the inverse Fourier transform.
The argument for the Fourier transform is similar, based on the facts that
∨
fb(ξ) = f (−ξ) a.e., and the interval [−Ω, Ω] is symmetric about the origin.
10.2 Suppose f ∈ FL2[−Ω,Ω] (R), so in particular we have supp(fb) ⊆ [−Ω, Ω].
Since
(Dr f ) (ξ) = D1/r fb(ξ) = r−1/2 fb(ξ/r),
∧
∧ ∧
the function (Dr f ) (ξ) can only be nonzero if −Ω ≤ ξ/r ≤ Ω. Hence (Dr f )
is supported in [−rΩ, rΩ], and therefore Dr f ∈ FL2[−rΩ,rΩ](R). Thus Dr
maps FL2[−Ω,Ω] (R) into FL2[−rΩ,rΩ](R), and a similar argument shows that
Dr is surjective.
10.3 (a) Suppose that fn ∈ PW(R), f ∈ L2 (R), and fn → f in L2 -norm.
The Fourier transform is unitary, so fbn → fb in L2 -norm. Consequently, there
is subsequence that converges pointwise a.e. But fbn (ξ) = 0 for almost every
ξ∈ / [−1/2, 1/2], so it follows that fb(ξ) = 0 for almost every ξ ∈
/ [−1/2, 1/2].
∨
(b) First we use that fact that f = fb to give an explicit proof that
f ∈ PW(R) is continuous. Given x, y ∈ R we have
∨ ∨
|f (x) − f (y)| = | fb (x) − fb (y)|
Z Z
= fb(ξ) e2πiξx dξ − fb(ξ) e2πiξy dξ
Z 1/2
= fb(ξ) e2πiξx − e2πiξy dξ
−1/2
Z 1/2
≤ |fb(ξ)| |e2πiξy | |e2πiξ(x−y) − 1| dξ
−1/2
Z 1/2
≤ |fb(ξ)| 2π|ξ| |x − y| dξ
−1/2
Detailed Solutions 173
Z 1/2 1/2 Z 1/2 1/2
≤ 2π |x − y| |fb(ξ)|2 dξ |ξ|2 dξ
−1/2 −1/2
1
= 2π |x − y| kfbkL2
2 · 31/2
= 3−1/2 π kf kL2 |x − y|.
X∞ X∞
|f (n) (a)| π n |x − a|n
|x − a|n ≤ kfbkL1 = kfbkL1 eπ|x−a| < ∞,
n=0
n! n=0
n!
the Taylor series for f converges absolutely at each x, and in fact it converges
uniformly on every compact set.
However, we still have to show that the Taylor series converges to f (x).
The remainder term Rn is defined by
∞
X f (k) (a)
f (x) = (x − a)k + Rn (x),
k!
k=0
i.e., (Ta dπ ) = ea χ[− 12 , 12 ] . Since {eλn }n∈N is a frame for L2 (T), it follows that
∧
{eλn χ[− 1 , 1 ] }n∈N is a frame for L2 1 1 (R), and therefore {Tλn dπ }n∈N is a
2 2 [− 2 , 2 ]
frame for PW(R). The canonical dual frame is {e
sn }n∈N . Therefore, for each
f ∈ PW(R) we have
X X
f = hf, Tλn dπ i sen = f (λn ) sen ,
n n
Detailed Solutions 175
hf ϕ, en i = 0, n ∈ Z.
= kf ϕk2L2
Z 1
= |f (x)|2 |ϕ(x)|2 dx
0
Z
= |f (x)|2 |ϕ(x)|2 dx
Zϕ
Z
≥ A |f (x)|2 dx
Zϕ
176 Detailed Solutions
Z 1
= A |f (x)|2 dx
0
= A kf k2L2 .
We compute that
X
ck−n e−2πikx
∧
(Tn c) (ξ) =
k∈Z
X
= ck e−2πi(k+n)x
k∈Z
X
= e−2πinx ck e−2πikx = e−2πinx b
c (ξ).
k∈Z
10.12 Fix g ∈ L2 (R) and a > 0, and let h(x) = g(ax). If {Tak g}k∈Z is a frame
sequence, then since the dilation operator Da f (x) = a1/2 f (ax) is unitary, it
follows that If {Da (Tak g)}k∈Z is a frame sequence as well. Now,
Da Tak g(x) = a1/2 (Tak g)(ax) = a1/2 g(ax − ak) = a1/2 h(x − k).
XZ a
= f (x + an) dx
n∈Z 0
XZ an+a
= f (x) dx
n∈Z an
Z ∞
= f (x) dx.
−∞
PN
10.14 Suppose that k=1 Tak g = 0 a.e. Taking the Fourier transform of both
sides, we have that
N
X
M−ak b
g = 0 a.e.
k=1
But then,
X
N
−2πiak ξ
e gb(ξ) = 0 a.e.
k=1
10.15 Set
N
X
mN (ξ) = ck e−2πikξ .
k=−N
P −2πikξ
Since b
c (ξ) = k∈Z ck e converges in L2 (T) we have mN → b c in L2 -norm
as N → ∞. Hence there exists some subsequence that converges pointwise a.e.,
say mNk (ξ) → b c (ξ) a.e. as k → ∞.
By hypothesis we also have that mN gb → F in L2 (R) as N → ∞, so
g → F in L2 -norm as well, and therefore some subsequence of this con-
mN k b
verges to F pointwise a.e. But mNk (ξ) gb(ξ) → b c (ξ) gb(ξ) a.e., so we must have
F (ξ) = bc (ξ) gb(ξ) a.e.
10.16 (b) ⇐. Suppose that A ≤ Φg ≤ B a.e. off Zg . Then by part (a) we know
that T (g) is a Bessel sequence with Bessel bound B. Note that T (g) ⊆ Hg , so
if we establish that the lower frame bound holds on Hg then T (g) is a frame
for Hg .
So, fix any F ∈ Hg . Then since {e−k }k∈Z is an orthonormal basis for
L2 (T), and F and Φg both vanish on Zg , we can compute that
X
X
2
F, Gk 2 = F Φ1/2
g , e−k
k∈Z k∈Z
2
=
F Φ1/2
g
2
L (T)
Z 1
= |F (ξ)|2 Φg (ξ) dξ
0
Z 1
≥ A |F (ξ)|2 dξ
0
2
= A
F
L2 (T) .
= δjk .
Hence {Fk }k∈Z is biorthogonal to {Gk }k∈Z , and therefore these sequences are
minimal in L2 (T). Since U maps V0 (g) unitarily onto L2 (T), we conclude that
Detailed Solutions 179
again with unconditional convergence of the series. Exercise 10.15 then implies
that fb = b
cbg.
⇐. Suppose f ∈ L2 (T) satisfies fb = m gb for some function m ∈ L2 (T).
c for some sequence c = ℓ2 (Z). Set
Then m = b
cN = (. . . , 0, 0, c−N , . . . , cN , 0, 0, . . . ).
Then
N
X
cc
N (ξ) = ck e−2πikξ → b
c (ξ) = m(ξ),
k=−N
XZ 1
= c (ξ + j) b
|b g(ξ + j) − cc g(ξ + j)|2 dξ
N (ξ + j) b
j∈Z 0
Z 1 X
2
= c (ξ) − cc
|b N (ξ)| gb(ξ + j)|2 dξ
0 j∈Z
Z 1
2
= c (ξ) − cc
|b N (ξ)| Φg (ξ) dξ
0
Z 1
2
≤ B c (ξ) − cc
|b N (ξ)| dξ
0
→ 0 as N → ∞.
In fact, all of the series above converge unconditionally, but the important
fact is that by applying the inverse Fourier transform we obtain
X
f = ck Tk g ∈ V0 (g).
k∈Z
10.18 (a) We are given that b g(ξ) 6= 0 for almost every ξ. Suppose that
f ∈ V0 (g). Then there exist functions fn ∈ V0 (g) such that fn → f in L2 -
norm. By passing to a subsequence, we may assume that fn → f pointwise a.e.
Now, each function fn can be written as
M
X
fn = ck T k g
k=−M
fbn (ξ)
p(ξ) = lim pn (ξ) = lim .
n→∞ n→∞ gb(ξ)
Therefore fb(ξ) = p(ξ) gb(ξ) a.e. Finally, as each function pn is 1-periodic, the
same is true of p.
Detailed Solutions 181
so |m(ξ)| = 1 a.e.
⇐. If gb = m fb where m is 1-periodic and |m| = 1 a.e., then, since T (f )
is a frame, Exercise 10.17 implies that g ∈ span(T (f )). Since we also have
fb = m gb, it follows symmetrically that f ∈ span(T (g)). Hence span(T (f )) =
span(T (g)).
(b) ⇐. If g = α Tk f where α is a scalar then
g(ξ)| = |α| |fb(ξ)|. By part (a), if T (f ) and T (g) have the same spans then
so |b
we must have |α| = 1.
⇒. Since g belongs to span(T (f )) and the integer translates of f are
orthonormal, we have
X X
g = hg, Tk f i Tk f = αk Tk f.
k∈Z k∈Z
However, if f and g are each compactly supported, then only finitely many of
the inner products hg, Tk f i can be nonzero. This implies that
X
gb(ξ) = hg, Tk f i M−k fb(ξ) = α(ξ) fb(ξ)
k∈Z
where X X
α(ξ) = hg, Tk f i e−2πikξ = αk e−2πikξ ,
k∈Z k∈Z
X X
= ek−j
αj α e−2πikξ
j∈Z k∈Z
X
= e)j e−2πikξ ,
(α ∗ α
j∈Z
e)j = δ0j ,
(α ∗ α j ∈ Z.
X Z 1 1/2
2
= |g(x + k)| dx
k∈Z 0
N Z
X 1 1/2
= |g(x + k)|2 dx
k=−N 0
< ∞.
X N
X
g · χ[k,k+1] k2∞ =
kb g · χ[k,k+1] k2∞ < ∞.
kb
k∈Z k=−N
X |b
g (ξ + j)|2
=
Φg (ξ + j)
j∈Z
X |b
g (ξ + j)|2
= = 1,
Φg (ξ)
j∈Z
while for ξ ∈ Z we have Φg♯ (ξ) = 0. Theorem 10.19 therefore implies that
T (g ♯ ) is a Parseval frame for V0 (g ♯ ) = V0 (g).
If T (g) is a Riesz basis for V0 (g), then |Z| = 0, so Φg♯ = 1 a.e., and
therefore T (g ♯ ) is an orthonormal basis for V0 (g).
10.22 We compute that
Z 1
2
hw, wi = 2 (1 − x)2 dx =
0 3
and Z 1
1
hw, T−1 wi = hw, T1 wi = x (1 − x) = .
0 6
Since supp(w) = [−1, 1], we have hw, Tk wi for all |k| > 1. By equation (10.20),
we therefore have
X 2 1 2πiξ 2 + cos 2πξ
Φw = hw, Tk wi e−2πikξ = + e + e−2πiξ = .
3 6 3
k∈Z
Therefore
184 Detailed Solutions
1
≤ Φw (ξ) ≤ 1, all ξ ∈ R,
3
so T (w) is a Riesz basis for its closed span.
We compute that
X
Φwe (ξ) = be + k)|2
|w(ξ
k∈Z
X |w(ξb + k)|2
=
Φw (ξ + k)2
k∈Z
P
b + k)|2
k∈Z |w(ξ
=
Φw (ξ)2
Φw (ξ)
=
Φw (ξ)2
1
=
Φw (ξ)
3
= .
2 + cos 2πξ
Similarly,
h i X
wc♯ , w
be (ξ) = c♯ (ξ + k) w(ξ
w be + k)
k∈Z
X b + k) w(ξ + k)
w(ξ
=
k∈Z
Φw (ξ + k)1/2 Φw (ξ + k)
P
k∈Z b + k)|2
|w(ξ
=
Φw (ξ + k)3/2
Φw (ξ)
=
Φw (ξ)3/2
1
=
Φw (ξ)1/2
1/2
3
= .
2 + cos 2πξ
If (ak ) is not uniformly separated then there must exist some j 6= k such
that |aj − ak | < δ, and without loss of generality we may assume that j < k.
Let fjk = Taj g −Tak g , and note that the basis hypothesis implies that fjk 6= 0.
Further, we have kfjk kL2 < ε by equation (∗). Considering the basis expan-
sions, since j < k we have
Therefore
kSj (fjk )kL2 kTaj gkL2 kgkL2
C ≥ kSj k ≥ > = .
kfjk kL2 ε ε
we have that {Mbn Tak g}k,n∈Z is a frame if and only if {Tak Mbn g}k,n∈Z is a
frame. In this case, we have for f ∈ L2 (R) that
X X
hf, Mbn Tak gi Mbn Tak g = hf, e2πiakbn Tak Mbn gi e2πiakbn Tak Mbn g
k,n k,n
X
= e−2πiakbn e2πiakbn hf, Tak Mbn gi Tak Mbn g
k,n
X
= hf, Tak Mbn gi Tak Mbn g.
k,n
Therefore the frame operators for these two frames are identical.
11.2 (b) The general linear group of degree 3 is the set of all 3 × 3 invertible
real matrices. It is denoted by GL3 (R), and it is a group under matrix mul-
tiplication. Since H3 ⊆ GL3 (R), we just have to show that H3 is a subgroup.
We have
1 b t 1 d u 1 b + d t + u + bc
0 1 a 0 1 c = 0 1 b + d ∈ H3 ,
0 0 1 0 0 1 0 0 1
(ak, bn, abi) ∗ (aℓ, bm, abj) = (ak + aℓ, bn + bm, abi + abj + bnaℓ)
= a(k + ℓ), b(n + m), ab(i + j + nℓ) ∈ G,
= Ta Sf.
Also,
X
S(Mb f ) = hMb f, Mbn Tak gi Mbn Tak g
k,n
X
= hf, M−b Mbn Tak gi Mbn Tak g
k,n
X
= hf, Mb(n−1) Tak gi Mbn Tak g
k,n
X
= hf, Mbn Tak gi Mb(n+1) Tak g
k,n
X
= hf, Mbn Tak gi Mb Mbn Tak g
k,n
X
= Mb hf, Mbn Tak gi Mbn Tak g
k,n
= Mb Sf.
Therefore S commutes with Ta and Mb , and hence commutes with Tak = (Ta )k
and Mbn = (Mb )n . Consequently,
and therefore
(c) Suppose that G(g, a, b) is a Gabor frame. By part (b), its canonical
dual is G(e
g , a, b). By Corollary 8.23, these frames are exact if and only if
Detailed Solutions 189
hMbn Tak g, Mbn Tak ge i = 1 for every k, n. However, all of these inner products
are equal:
hMbn Tak g, Mbn Tak ge i = hTak g, M−bn Mbn Tak e
gi
= hTak g, Tak ge i = hg, T−ak Tak ge i = hg, e
g i.
Hence G(g, a, b) is a Riesz basis if and only if hg, e
g i = 1.
(d) We have seen that the frame operator commutes with Tak and Mbn ,
and that the same is true for S −1 . Since S −1 is a positive operator, it has a
square root S −1/2 . By Theorem 2.18, S −1/2 commutes with every operator
that commutes with S −1 . Therefore S −1/2 commutes with Tak and Mbn , and
hence
S −1/2 (Mbn Tak f ) = Mbn Tak (S −1/2 f ), f ∈ L2 (R).
Corollary 8.28 therefore implies that G(S −1/2 g, a, b) is a Parseval frame.
11.4 (a) The Fourier transform is a unitary operator on L2 (R), so the image
of a frame under the Fourier transform is still a frame. Applying equations
(9.2) and (9.3), we have
∧
F G(g, a, b) = (Mbn Tak g) k,n∈Z
= Tbn M−ak bg k,n∈Z
= Tbn Mak gb k,n∈Z .
Hence
G(g, a, b) is a frame ⇐⇒ Tbn Mak b
g k,n∈Z is a frame
⇐⇒ Mak Tbn b
g k,n∈Z is a frame
⇐⇒ G(b
g , b, a) is a frame.
Note that since the Fourier transform is unitary, the frame bounds of G(g, a, b)
are the same as those of G(bg , b, a).
11.5 (a) The fact that G(χ[0,1] , 1, 1) is orthonormal follows from the fact
that {e2πinx }n∈Z is an orthonormal sequence in L2 [a, a + 1] for any a ∈ R. It
is a frame by Theorem 11.4, or we can prove this directly as follows. Given
f ∈ L2 (R), let fk = f χ[k,k+1] . Note that
Z k+1
hfk , en iL2 [k,k+1] = fk (x) e−2πinx dx
k
Z ∞
= f (x) χ[k,k+1] (x) e−2πinx dx
−∞
Z ∞
= f (x) Mn Tk χ[0,1] (x) dx
−∞
190 Detailed Solutions
= f, Mn Tk χ[0,1] L2 (R) .
Since {e2πinx }n∈Z is an orthonormal basis for L2 [k, k + 1], we can use the
Planchel Equality to compute that
Z ∞ X Z k+1
2 2
kf kL2 = |f (x)| dx = |fk (x)|2 dx
−∞ k∈Z k
X
= kfk k2L2 [k,k+1]
k∈Z
X
= kfbk k2ℓ2 (Z)
k∈Z
XX
= |fbk (n)|2
k∈Z n∈Z
XX
= |hfk , en iL2 [k,k+1] |2
k∈Z n∈Z
X X
= f, Mn Tk χ[0,1] 2 2
L (R)
k∈Z n∈Z
where e
g = bg/G0 . Hence the dual frame is
−1
S (Mbn Tak g) k,n∈Z = Mbn Tak ge k,n∈Z = G(g, a, b).
hTak g, Tℓa gi = 0, k 6= ℓ ∈ Z.
If k = ℓ and m 6= n, then
Z ak+a
g(x − ak)
Mbm Tak g, Mbn Tak e
g = e2πibmx g(x − ak) e−2πibnx dx
ak G0 (x − ak)
Z ak+a
|g(x − ak|2
= e2πi(bm−bn)x dx
ak G0 (x − ak)
Z ak+a
= e2πi(bm−bn)x dx = 0.
ak
Hence G(g, a, b) has a biorthogonal system. Since we are given that G(g, a, b)
is a frame, it therefore is an exact frame and hence is a Riesz basis.
(c), (d) Suppose 0 < ab ≤ 1, and g satisfies the hypotheses of the
Painless Nonorthogonal Expansions. Since the canonical dual of G(g, a, b) is
G(e
g , a, b), Corollary 8.23 implies that these frames are exact if and only if
hMbn Tak g, Mbn Tak ge i = 1 for every k, n. Now, all of these inner products are
equal to hg, ge i, and since ge = bg/G0 we have
Z b−1 Z b−1
|g(x)|2
hg, e
gi = g(x) ge(x) dx = b P dx.
0 0 k |g(x − ak)|2
192 Detailed Solutions
P
Since |g(x)|2 ≤ k |g(x − ak)|2 , the integrand on the line above lies between 0
and 1 almost everywhere. Hence
Z b−1
|g(x)|2
b 1− P dx ≥ 0,
0 k |g(x − ak)|2
and the integrand is nonnegative a.e. Therefore, this integral is zero if and
only if
|g(x)|2
1− P 2
= 0 a.e. on [0, b−1 ].
k |g(x − ak)|
That is,
In case these hold, by examining the supports of g(x − ak) we see that G0 = 0
a.e. on [0, b−1 − a] and [a, b−1 ]. However, to be a frame we must at least
have that G0 is nonzero a.e. Therefore, if G(g, a, b) is a Riesz basis then these
intervals must have zero measure, which requires a = b−1 . Thus, we can only
have a Riesz basis when ab = 1.
11.7 Suppose g ∈ Cc (R) is not the zero function. Let b be small enough that
supp(g) is contained in some interval I of length 1/b. Since g is continuous
and not everywhere zero, there exists some interval J such that g(x) 6= 0
for all x ∈ J. Let a > 0 be smaller than the length of J. Then G0 (x) =
P 2
k |g(x − ak)| is continuous, a-periodic, and everywhere nonzero. Therefore
0 < inf G0 ≤ sup G0 < ∞, so G(g, a, b) is a frame by Theorem 11.4.
11.8 Suppose 0 < a < b−1 0 . The support condition on g combined with the
fact that g(x) > 0 on (0, b−1
0 ) implies that
X
0 < A ≤ |g(x − ak)|2 ≤ B < ∞
k∈Z
where
pn+1 (t) = 2t3 pn (t) − t2 p′n (t).
Since 2t3 pn (t) has degree 3n + 3 while t2 p′n (t) has degree 3n + 1, we conclude
that their difference has degree 3n + 3.
Finally, since pn is a polynomial,
−2 2
lim f (n) (x) = lim+ pn (x−1 ) e−x = lim pn (t) e−t = 0.
x→0+ x→0 t→∞
is infinitely differentiable because for any given x there are at most a fixed
finite number of terms in the sum. Because a < b−1 it is also strictly positive
everywhere, so
f
g = 1/2
F0
is infinitely differentiable and supported in [0, b−1 ]. Further,
X
G0 (x) = |g(x − ak)|2
k∈Z
X |f (x − ak)|2
= P 2
k∈Z j∈Z |f (x − aj)|
P
|f (x − ak)|2
= Pk∈Z 2
= 1.
j∈Z |f (x − aj)|
Theorem 11.4 therefore implies that G(g, a, b) is a b−1 -tight frame. By rescal-
ing, we can make it a Parseval frame.
11.11 (a) If G(g, a, b) is a Riesz frame, then it is a frame, and ab = 1 by
Corollary 11.7(e).
Conversely, if ab = 1 and G(g, a, b) is a frame then Corollary 11.7(e) implies
that it is a Riesz basis.
194 Detailed Solutions
X X X 1/q X
kfn · χ[j,j+1] kLp ≤ kfn · χ[k,k+1] kqLp = kfn kW < ∞.
n n k∈Z n
P
Therefore the series gj = n fn · χ[j,j+1] converges absolutely in Lp [j, j + 1].
Define g on R by setting g(x) = gj (x) for x ∈ [j, j + 1]. Then g · χ[j,j+1] =
gj ∈ Lp [j, j + 1], and by applying the Triangle Inequality in ℓq we obtain
X 1/q
kgkW = kg · χ[k,k+1] kqLp
k∈Z
X
X
q 1/q
=
fn · [k,k+1]
χ
p
k∈Z n L
X X q 1/q
= kfn · χ[k,k+1] kLp
k∈Z n
X X 1/q
≤ kfn · χ[k,k+1] kqLp
n k∈Z
X
= kfn kW < ∞.
n
E = [0, 12 ) ∪ [1 + 21 , 1 + 43 ) ∪ [2 + 34 , 2 + 87 ) ∪ + · · · .
and the “pieces” have no overlap when they are translated back to the inter-
val [0, 1). This implies that {Mn g}n∈Z is orthonormal. Together, G(g, 1, 1) is
orthonormal.
Suppose that f is orthogonal to Mn Tk g for every k, n ∈ Z. Fix k = 0.
Take f · χE and translate each “piece” back into the interval [0, 1). Because
the exponentials e2πinx are 1-periodic, the resulting function h on [0, 1) is
196 Detailed Solutions
orthogonal to e2πinx for each n ∈ Z, and therefore is zero a.e. Repeating for
n 6= 0 shows that f = 0 a.e. everywhere. Hence G(g, 1, 1, ) is complete.
11.17 (a) If f ∈ W (L∞ , ℓ1 ) then
Z ∞ X Z k+1 X
|f (x)| dx = |f (x)| dx ≤ sup |f (x)| = kf kW (L∞ ,ℓ1 ) ,
−∞ k∈Z k k∈Z x∈[k,k+1]
so f ∈ L1 (R). Also
sup |f (x)| = sup sup |f (x)| ≤ kf kW (L∞ ,ℓ1 ) ,
x∈R k∈Z x∈[k,k+1]
≤ sup |f (y)|
y∈[k−j−1,k−j+1]
= kf · χ[k−j−1,k+j+1] kL∞
(d) The argument for the lower inequality is entirely symmetrical to the
one given for the upper inequality.
11.18 There are several points in the computation that should be justified.
(a) Since g belongs to W (L∞ , ℓ1 ), it is integrable. As f is bounded, the
function f (x) e−2πibnx g(x − ak) is integrable on R, and therefore the equality
Z ∞
f (x) e−2πibnx g(x − ak) dx
−∞
Z b−1 X
j
= f x− j
b e−2πibn x− b
g x − ak − j
b dx
0 j∈Z
is justified.
(b) Because f and Fk are bounded and g is integrable, the function
f (x) g(x − ak) Fk (x)
is integrable. This justifies the equality
Z b−1 X
f x − jb g x − ak − jb Fk x − jb dx
0 j∈Z
Z ∞
= f (x) g(x − ak) Fk (x) dx
−∞
XXZ R
j
= |f (x)| |g(x − ak)| |f x − b | |g x − ak − jb | dx
k∈Z j∈Z −R
XXZ R
≤ |f (x)| kTak g · χ[−R,R] kL∞
k∈Z j∈Z −R
× |f x − jb | kTak+ j g · χ[−R,R] kL∞ dx
b
X X
2
≤ 2R kf kL∞ kTak g · χ[−R,R] kL∞ kT j g · χ[−R,R] kL∞
b
k∈Z j∈Z
11.19 Suppose
G(g, a, b) = Mbn Tak h k,n∈Z
is a frame.
11.20 Since G(χ[0,1] , 1, 1) and {Enk }k,n∈Z are each orthonormal basis, it
follows from Theorem 4.20 that they are equivalent bases. Hence there is a
topological isomorphism Z such that Z(Mn Tk χ[0,1] ) = Enk . The fact that
both systems are orthonormal bases implies that this mapping is unitary, and
that Z is unique.
11.21 Fix f ∈ L2 (R) or f ∈ W (Lp , ℓ1 ). Given x, ξ ∈ R and m, n ∈ Z, we
have
X
Zf (x + m, ξ + n) = f (x + m − j) e2πij(ξ+n)
j∈Z
X
= f (x + m − j) e2πijξ
j∈Z
X
= f (x − j) e2πi(j+m)ξ
j∈Z
X
= e2πimξ f (x + j) e2πijξ
j∈Z
2πimξ
= e Zf (x, ξ).
Detailed Solutions 199
= 0.
11.22 (a) Because f ∈ L1 (R), we have
XZ 1 XZ 1
2πijξ
|f (x − j) e | dξ = |f (x − j)| dξ = kf kL1 < ∞.
j∈Z 0 j∈Z 0
Z a+1 X Z 1
= f (x) − f (x − j) e2πijξ dξ dx
a j∈Z 0
Z a+1 X
= f (x) − f (x − j) δ0,j dx
a j∈Z
= 0.
200 Detailed Solutions
→ 0 as a → 0,
XZ 1 Z 1
= f (x − j) e2πijξ e−2πinx e2πikξ dx dξ
j∈Z 0 0
XZ 1 Z 1
= f (x − j) e−2πinx dx e2πi(j+k)ξ dξ
j∈Z 0 0
XZ 1
= f (x − j) e−2πinx dx (δ−j,k )
j∈Z 0
Detailed Solutions 201
Z 1
= f (x + k) e−2πinx dx
0
= f χ[k,k+1] , e2πinx L2 [k,k+1] .
∧
= f χ[k,k+1] (n).
Thus, the (n, −k)th Fourier coefficient of Zf ∈ L2 (Q) is the nth Fourier
coefficient of f χ[k,k+1] ∈ L2 [k, k + 1] ⊆ L1 [k, k + 1].
On the other hand, the series defining Zg converges absolutely in the
c
norm of L1 (Q), so an entirely similar calculation shows that Zg(n, −k) =
∧
χ
g [k,k+1] (n). Since Zf = Zg, we therefore have
∧ ∧
f χ[k,k+1] (n) = g χ[k,k+1] (n), k, n ∈ Z.
Now, f χ[k,k+1] and g χ[k,k+1] both belong to L1 [k, k + 1], and the Uniqueness
Theorem for Fourier coefficients (Theorem 13.26) tells us that functions in
L1 [k, k+1] are uniquely determined by their Fourier coefficients, so this implies
that f χ[k,k+1] = g χ[k,k+1] a.e. for each k ∈ Z. Hence f = g a.e.
But this tells us that every function f ∈ L2 (R) equals a function
g ∈ L1 (R), and hence L2 (R) ⊆ L1 (R). This is false, so we have obtained
a contradiction.
(d) Suppose Z −1 : range(Z) → L1 (R) was continuous. By definition of in-
verse, we have Z −1 Zf = f for f ∈ L1 (R) and ZZ −1 F = F for F ∈ range(Z).
However, range(Z) is dense, so Z −1 has a unique extension to a continuous
mapping Z −1 : L1 (Q) → L1 (R). By continuity, we obtain ZZ −1 F = F for all
F ∈ L1 (Q). But then Z is surjective, which contradicts part (c).
11.23 We are given f ∈ L2 (R) such that Zf is continuous.
(a) If f is even then
X 1
Zf (1/2, 1/2) = f ( 21 − j) e2πi 2 j
j∈Z
X 1
= f (− 21 − j) e2πi 2 j
j∈Z
X 1
= f (− 21 − (j − 1)) e2πi 2 (j−1)
j∈Z
X 1
= f ( 21 − j) e2πi 2 j e−πi
j∈Z
X 1
= − f ( 21 − j) e2πi 2 j = −Zf (1/2, 1/2).
j∈Z
Therefore Zf (0, 0) = 0.
Since Zf is 1-periodic in the second variable,
X 1
Zf (0, 1/2) = f (−j) e2πi 2 j
j∈Z
X 1
= − f (j) e2πi 2 j
j∈Z
X 1
= − f (−j) e2πi 2 (−j)
j∈Z
Therefore Zf (1/2, 0) = 0.
(c) If f is real-valued then
X 1
Zf (x, 1/2) = f (x − j) e2πi 2 j
j∈Z
X 1
= f (x − j) e−2πi 2 j = Zf (x, −1/2) = Zf (x, 1/2).
j∈Z
Therefore Zf (x, 1/2) is real for each x ∈ [0, 1]. However, by quasiperiodicity,
Detailed Solutions 203
Thus, Zf (0, 1/2 and Zf (1, 1/2) have opposite signs. Since Zf (x, 1/2) is con-
tinuous and real-valued, the Intermediate Value Theorem implies that there
must be some x ∈ [0, 1] such that Zf (x, 1/2) = 0.
11.24 The proof is very similar to the proof of Exercise 10.8 so we will only
expand on the following point.
(d) ⇒. Suppose that G(g, 1, 1) is a frame. Then it is a Bessel sequence,
so by part (b) there exists some B > 0 such that |Zg|2 ≤ B a.e., and B is
a Bessel bound. An almost identical argument based on the lower the frame
bound shows that there exists some A > 0 such that |Zg|2 ≥ A a.e., and A is
a lower frame bound.
Since Q has finite measure and we have A ≤ |Zg|2 ≤ B a.e., it follows that
1/Zg ∈ L2 (Q). Therefore G(g, 1, 1) is exact by part (b). Finally, exact frames
coincide with Riesz bases.
PN
11.25 (a) Suppose p(x) = k=−N ck e2πikx satisfies |p(x)| = 1 a.e. Set ck = 0
for |k| > N. Then for almost every x we have
N
X N
X
1 = |p(x)|2 = cj ck e2πi(j−k)x
j=−N k=−N
XX
= cj ck+j e2πi(j−(k+j))x
j∈Z k∈Z
XX
= cj ck+j e2πikx .
k∈Z j∈Z
cn e2πinx a.e. (in fact, equality must hold everywhere since both sides are
continuous). Finally, |cn | = |p(x)| = 1.
(b) ⇒. Suppose that g ∈ L2 (R) is compactly supported and G(g, 1, 1) is
an
P orthonormal basis for L2 (R). Then |Zg|2 = 1 a.e. by Theorem 11.31 and
|g(x − j)|2 = 1 a.e. by Theorem 11.6.
Fix any particular representative of g. If supp(g) ⊆ [−N, N + 1] then
N
X
Zg(x, ξ) = g(x − j) e2πijξ .
j=−N
Since {Enk }k,n∈Z is an orthonormal basis for L2 (Q) and c ∈ ℓ2 (Z2 ), the series
X
F = cnk Enk
k,n∈Z
However, Zg(x, ξ) 6= 0 for almost every (x, ξ), so this implies that F = 0 a.e.
Finally, the fact that {Enk }k,n∈Z is an orthonormal basis therefore implies
that c = 0.
11.27 Suppose g ∈ L2 (R) and x |g(x)|2 → c as x → ∞. If c > 0, then
for x large enough we have |g(x)|2 ≥ c/x, which contradicts the fact that g is
square integrable.
11.28 Let fn = χ[n,n+1] . Then kfn kL2 = 1, and
Z n+1 n+1
x3 (n + 1)3 − n3
kP fn k2L2 = 2
x dx = = .
n 2 n 2
kgkL2 for every n, but gn′ (x) = n3/2 g ′ (nx), so kgn′ kL2 = nkg ′ kL2 . Hence
kM gn kL2 → ∞ as n → ∞, so M is unbounded.
11.29 The position and momentum operators map S(R) into itself simply
by definition of the Schwartz class. Given f, g ∈ S(R), we have
Z Z
hP f, gi = xf (x) g(x) dx = f (x) xg(x) dx = hf, P gi,
1
= − f (x).
2πi
To relate this to equation (11.32), note that if f is smooth enough and has
sufficient decay then the following integrals are well defined, and by rearrang-
ing terms and using integration by parts we have
Z Z
1 1
M f, P f − P f, M f = f ′ (x) xf (x) dx − xf (x) f ′ (x) dx
2πi 2πi
Z Z
1 1
= xf ′ (x) f (x) dx − (P f )′ (x) f (x) dx
2πi 2πi
Z Z
= P M f (x) f (x) dx − M P f (x) f (x) dx
= P M f − M P f, f
= [P, M ]f, f
1
= − hf, f i
2πi
1
= − kf k2L2 .
2πi
This is equation (11.32), noting that the function g in that equation has unit
norm.
11.31 (a) For f, g ∈ S we have that
[A, B]f, f = ABf, f − BAf, f
= Bf, Af − Af, Bf
= 2i Im Bf, Af .
Hence
[A, B]f, f = 2 Im Bf, Af
≤ 2 Bf, Af
≤ 2 kBf k kAf k.
(b) Equality holds in the first inequality above if and only if hBf, Af i is
purely imaginary. Equality holds in the second inequality above if and only if
Bf is a scalar multiple of Af. So, we must have Bf = aAf for some a ∈ C,
and therefore hBf, Af i = akAf k2 . In order for this quantity to be purely
imaginary, we must have a = ci for some real c.
11.32 Given f ∈ S(R) we have
Detailed Solutions 207
A0 + z = 0 + λ1 = λ1 ,
Ae1 + z = u + λ1 = λ2 ,
Ae2 + z = v + λ1 = λ3 .
where c′k = ck e2πi(qk +b)a . Since e2πibx is everywhere nonzero and G(g, Λ) is
independent, this implies that c′k = 0 for each k, and hence ck = 0 for each k.
Therefore G(g, Λ + z) is linearly independent. The converse implication is
similar.
208 Detailed Solutions
2 2
where c′k = ck eπirpk . Since eπirx is everywhere nonzero, we conclude that
PN ′
k=1 ck Mqk Tpk g = 0 a.e. Since G(g, Λ) is independent, this implies that
c′k = 0 for every k, and hence ck = 0 for every k. Therefore G(h, Sr (Λ)) is
independent. The converse implication is similar.
(c) Suppose that G(g, Λ) is independent. Note that
p 0 −1 pk −qk
R k = = ,
qk 1 0 qk pk
∨
∨ N PN ∨
so G( g , R(Λ)) = Mpk T−qk g k=1 . Suppose that k=1 ck M−pk Tqk g = 0 a.e.
Then for almost every x we have
N
X ∨
0 = ck Mpk T−qk g (ξ)
k=1
N
X ∨
= ck (Tpk Mqk g) (ξ)
k=1
Detailed Solutions 209
N
X ∨
= c′k (Mqk Tpk g) (ξ),
k=1
where c′k
=e −2πipk qk
ck . Since the Fourier transform is unitary and G(g, Λ) is
independent, we conclude that c′k = 0 for every k, and therefore ck = 0 for
∨
each k. Hence G( g , R(Λ)) is independent.
(d) Suppose that G(g, Λ) is independent. Note that
pk a 0 pk apk
Da = = ,
qk 0 1/a qk qk /a
N PN
so G(h, Da (Λ)) = Mqk /a Tapk h k=1 . If k=1 ck Mqk /a Tapk h = 0 a.e., then for
almost every y = ax we have
N
X
0 = ck Mqk /a Tapk h(y)
k=1
N
X
= ck e2πi(qk /a)(ax) h(ax − apk )
k=1
N
X
= ck e2πiqk ax g(x − pk )
k=1
N
X
= ck Mqk Tpk g(x).
k=1
Since G(g, Λ) is independent, this implies that ck = 0 for each k, and therefore
G(h, Da (Λ)) is independent.
(e) Suppose that
x y
A = , xw − yz = 1.
zw
If y 6= 0, then
−y 0 1 0 0 −1 1 0
D−y Syw RSx/y =
0 −1/y yw 1 1 0 x/y 1
" #
x y
= xyw−y yw
y2 y
" #
x y
= xw−1
y w
x y
= = A.
zw
210 Detailed Solutions
By part (a) this implies that gb = 0 a.e., and therefore g = 0 a.e. since the
Fourier transform is unitary.
(c) We are given N distinct but collinear points Λ = {(pk , qk )}N k=1 . By
Exercise 11.35(a), if z ∈ R2 then G(g, Λ) is independent if and only if G(g, Λ +
z) is independent. Therefore we can assume that Λ is contained in a line that
passes through the origin.
N
If this line is vertical, then pk = 0 for every k and G(g, Λ) = Mqk g k=1
is independent by part (a).
If this line is not vertical then it has some finite slope r. Define h(x) =
2
e−πirx g(x). By Exercise 11.35(b), G(g, Λ) is linearly independent if and only
if G(h, S−r (Λ)) is linearly independent. However,
pk 1 0 pk pk
S−r = = ,
qk −r 1 rpk 0
Detailed Solutions 211
N
so G(h, S−r (Λ)) = Tpk h k=1 . This set is independent by part (b) of this
exercise, so we conclude that G(g, Λ) is independent as well.
11.38 Fix any nonzero g ∈ L2 (R) with g supported on a half-line of the
form [a, ∞). By taking a as large as possible we may assume that g is not
zero almost everywhere on [a, a + ε] for all ε > 0. Given finitely many distinct
points in R2 , we can write these points as
i.e., for each distinct translate we group the corresponding modulates together.
We can assume that p1 < · · · < pN . Given scalars ck,j , suppose that
X Mk
N X N
X
0 = ck,j Mqk,j Tpk g(x) = mk (x) g(x − pk ) a.e., (∗)
k=1 j=1 k=1
where
Mk
X
mk (x) = ck,j e2πiqk,j x .
j=1
where
pn+1 (x) = p′n (x) − 4πxpn (x).
Since p′n has degree n−1 and 4πxpn (x) has degree n+1, we conclude that pn+1
has degree n + 1, and the leading coefficient is (−4π) · (−4π)n = (−4π)n+1 .
2
11.40 We are given h(x) = p(x) e−πx where p is a nontrivial polynomial.
Given a finite set of distinct points in R2 we can write these points as
where
Mk
X 2
mk (x) = ck,j e−πpk e2πiqk,j x .
j=1
Now suppose that |pk − p′k | < ε and |qk − qk′ | < ε for k = 1, . . . , N. Then
for any scalars c1 , . . . , cN we have
XN
XN
ck Mqk Tpk g
≤
ck (Mqk − Mqk′ )Tpk g
2 2
k=1 k=1
XN
XN
+
ck Mqk′ (Tpk − Tp′k )g
+
ck Mqk′ Tp′k g
2 2
k=1 k=1
N
X
≤ |ck | kMqk −qk′ Tpk g − Tpk gk2
k=1
N
X
XN
+ |ck | kTpk −p′k g − gk2 +
ck Mqk′ Tp′k g
2
k=1 k=1
N
X
XN
≤ 2δ |ck | +
ck Mqk′ Tp′k g
2
k=1 k=1
X
N 1/2
XN
≤ 2δN 1/2 |ck |2 +
ck Mqk′ Tp′k g
.
2
k=1 k=1
X
N 1/2
XN
1/2 1/2 2
A − 2δN |ck | ≤
ck Mqk′ Tp′k g
.
2
k=1 k=1
PN
Since A1/2 − 2δN 1/2 > 0, it follows that if k=1 ck Mqk′ Tp′k g = 0 a.e., then
c1 = · · · = cN = 0.
11.42 If m is differentiable at x = 0 and m(0) = 0, then we have
|m(x)|
C1 ≤ ≤ C2 , |x| < δ.
|x|
Consequently
C1 |x| ≤ |m(x)| ≤ C2 |x|, |x| < δ,
and therefore
Now,
Z δ Z δ
ln |x| dx = 2 ln x dx = 2δ (ln δ − 1).
δ 0
Rδ
Since ln C1 + ln |x| ≤ p(x) ≤ ln C2 + ln |x|, we conclude that −δ
p(x) dx exists
and is finite.
Detailed Solutions to Exercises from Chapter 12
12.1 If W(ψ, a, b) is a wavelet frame then
XX
S(Da f ) = hDa f, Dan Tbk ψi Dan Tbk ψ
n∈Z k∈Z
XX
= hf, D1/a Dan Tbk ψi Dan Tbk ψ
n∈Z k∈Z
XX
= hf, Dan−1 Tbk ψi Dan Tbk ψ
n∈Z k∈Z
XX
= hf, Dan Tbk ψi Dan+1 Tbk ψ
n∈Z k∈Z
XX
= hf, Dan Tbk ψi Da Dan Tbk ψ
n∈Z k∈Z
XX
= Da hf, Dan Tbk ψi Dan Tbk ψ
n∈Z k∈Z
= Da Sf.
Thus, if we set ψek = S −1 (Tbk ψ) for k ∈ Z, then the canonical dual simply
consists of dilations of these functions ψek , i.e., the canonical dual is
Dan ψek k,n∈Z .
12.2 (a) The general linear group of degree 2 is the set of all 2 × 2 invertible
real matrices. It is denoted by GL2 (R), and it is a group under matrix mul-
tiplication. Since A2 ⊆ GL2 (R), we just have to show that A2 is a subgroup.
We have
a0 c0 ac 0
= ∈ A2 ,
b1 d1 bc + d 1
so A2 is closed under composition. Also,
−1
a0 1/a 0
= ∈ A2 ,
b1 −b/a 1
Dan T(am j+k)b = Dan−m Dam Tam bj Tbk = Dan−m Tbj Dam Tbk ∈ G.
(d) Fix (u, v) ∈ A3 , and note that u > 0. Given a function F that is
integrable with respect to da a db, by making the change of variable c = au
followed by d = b + (cv)/u, we compute that
Z ∞Z ∞ Z ∞Z ∞
da da
F (u, v) ∗ (a, b) db = F (au, av + b) db
−∞ 0 a −∞ 0 a
Z ∞Z ∞ dc/u
cv
= F c, +b db
−∞ 0 u c/u
Z ∞Z ∞ dc
cv
= F c, + b db
0 −∞ u c
Z ∞Z ∞
dc
= F (c, d) dd
0 −∞ c
Z ∞Z ∞
da
= F (a, b) db.
−∞ 0 a
12.3 (a) By equation (12.3), the image of W(ψ, a, b) under the Fourier trans-
form is
c a, b) = a−n/2 e2πibka−n ξ ψ(a
W(ψ, b −n ξ)
k,n∈Z
n/2 2πibkan ξ
= a e b n ξ)
ψ(a .
k,n∈Z
Since the Fourier transform is unitary, our goal is to show that this collection
is a Parseval frame for L2 [0, ∞). Note that, with n fixed,
n/2 1/2 2πibkan ξ
a b e k∈Z
is an orthonormal basis for L2 (I) where I is any interval of length a−n b−1 .
Let f be any function in Cc (R). For simplicity let us take c = 1, but the
same argument applies to any positive c. Let d = 1 + b−1 , so ψb ∈ L2 (R)
is supported within [1, d], which is an interval of length b−1 . The dilated
b n ξ) belongs to L2 (In ) where In = [a−n , a−n c], an interval of
function ψ(a
b n ξ)
length a−n (d − 1) = a−n b−1 . Since f is bounded, the product f (ξ) · ψ(a
2
also belongs to L (In ). Applying the Plancherel Equality, we therefore have
Z ∞
b n ξ)|2 dξ =
|f (ξ) ψ(a b n ξ)
f (ξ) ψ(a
2
0 L (In )
X D E 2
= b n ξ), an/2 b1/2 e2πibkan ξ
f (ξ) ψ(a
L2 (In )
k∈Z
2
X Z
= an b b n ξ) e−2πibkan ξ
f (ξ) ψ(a
k∈Z In
2
X Z ∞
= an b b n ξ e−2πibkan ξ .
f (ξ) ψ(a
k∈Z 0
X D E 2
= b b n ξ) e2πibkan ξ
f (ξ), an/2 ψ(a .
L2 [0,∞)
k∈Z
for all f ∈ Cc (R). Since Cc (R) is dense in L2 (R), it follows that G(g, a, b) is
a tight frame with frame bound Ab−1 .
(b) By applying the same techniques as in Example 11.5 and Exer-
cise 11.10, we can choose a, b, and ψ so that ψb is as smooth as we like,
even infinitely differentiable. We just need to make sure that the dilations of
the interval [c, c + b−1 ] have nontrivial overlaps, i.e., we need to choose a, b so
that ac < c + b−1 .
(c) We take a = 2 and b = 1. Define ψ1 ∈ L2 (R) by
0, x < 1/2,
linear, 1/2 ≤ x ≤ 3/4,
ψb1 (ξ)2 = 1, 3/4 ≤ x ≤ 1,
linear, 1 ≤ x ≤ 3/2,
0, x > 3/2.
P
If we set c = 1/2 then we have c + b−1 = 3/2. Also, n∈Z |ψb1 (ξ)|2 = 1 on
2
(0, ∞), so W(ψ1 ) is a Parseval frame for H+ (R) by part (a).
Now let ψ2 ∈ L (R) be the function such that ψb2 (ξ) = ψ(−ξ).
2 b Then
2
W(ψ1 ) ∪ W(ψ2 ) is a Parseval frame for L (R). Explicitly, we just have to set
ψ2 (x) = ψ(−x), for then
Z Z
ψb2 (ξ) = ψ(−x) e−2πiξx dx = b
ψ(x) e2πiξx dx = ψ(−ξ).
12.4 (a) Let Ek = (E ∩ [k, k + 1]). We claim that, modulo sets of measure
zero, we have S
(Ek − k) = [0, 1],
k∈Z
XZ
= |U f (x)|2 dx
k∈Z Ek
Z
= |U f (x)|2 dx = kU f k2L2 (E) .
E
U en (x) = en (x − k) = en (x).
Since {en }n∈Z is an orthonormal basis for L2 [0, 1] and U is unitary, it follows
that {en }n∈Z is an orthonormal basis for L2 (E).
(b) By part (a), {ek }k∈Z is an orthonormal basis for L2 (E). Since the
dilation Da−n is unitary, {Da−n ek }k∈Z is an orthonormal basis for L2 (an E).
Since ∪(an E) = R with overlaps of measure zero, we conclude that
Da−n (ek χE ) k,n∈Z = Da−n Mk χE k,n∈Z
is admissible, and ψε → ψ as ε → 0.
12.9 If ψb is continuous and ψ(0)
b b
6= 0, then |ψ(ξ)| > δ > 0 on some neighbor-
b
hood of the origin, which implies that ψ is not admissible.
12.10 We fill in the details of the claims made in Example 12.12.
Exercise 10.3 shows that V0 = PW(R) is a closed subspace of L2 (R).
Since Vn consists of functions whose Fourier transforms are supported
within [−2n−1 , 2n−1 ], the union of the Vn contains all functions in L2 (R)
with compactly supported Fourier transforms. Since the Fourier transform is
unitary and
L2c (R) = {f ∈ L2 (R) : supp(f ) is compact}
is dense in L2 (R), it follows that ∪n∈Z Vn is a dense subspace of L2 (R) as
well.
Detailed Solutions 221
Hence
n−1
X X
Pn f = P−n f + bk,m D2m Tk ψ,
m=−n k∈Z
where
bk,m = hf, D2m Tk ψi.
Since Pn f → f and P−n f → 0 in L2 -norm, we conclude that
XX
f = bk,m D2m Tk ψ.
m∈Z k∈Z
Hence Pn → f in L2 -norm as n → ∞.
(e) The space of bounded, compactly supported functions is dense in
L2 (R), so suppose first that f is bounded and compactly supported, say
supp(f ) ⊆ [−R, R]. Then
X
kPn f k2L2 = f, D2n Tk ϕ 2
k∈Z
X
= D2−n f, Tk ϕ 2
k∈Z
222 Detailed Solutions
Z 2
X ∞
= −n/2 −n
x) ϕ(x − k) dx
2 f (2
k∈Z −∞
Z 2n R 2
X
−n
= 2 f (2 −n
x) ϕ(x − k) dx
k∈Z −2n R
X Z 2n R Z 2n R
−n −n 2
≤ 2 |f (2 x)| dx |ϕ(x − k)|2 dx
k∈Z −2n R −2n R
X Z 2n R Z 2n R+k
≤ 2−n kf k2L∞ dx |ϕ(x)|2 dx
k∈Z −2n R −2n R+k
X Z 2n R+k
≤ 2R kf k2L∞ |ϕ(x)|2 dx
k∈Z −2n R+k
Z
= 2R kf k2L∞ |ϕ(x)|2 dx,
En
where S
En = −2n R + k, 2n R + k .
k∈Z
2 sin(πk/2)
=
πk
0,
k even,
= 2/(πk), k = . . . , −7, −3, 1, 5, . . . ,
−2/(πk), k = . . . , −5, −1, 3, 7, . . . .
We conclude that m0 (ξ/2) = χ[− 12 , 21 ] , at least on the interval [−1, 1]. Hence
m0 (ξ) = χ[− 41 , 14 ] , extended 1-periodically to a function on T.
12.16 It is the Cantor–Lebesgue function, reflected about the point x = 1 to
create a continuous function supported on [0, 2].
12.17 By hypothesis, there some sequence of partial sums such that
X
ϕ = lim 2−1/2 ck D2 Tk ϕ,
n→∞
k∈Fn
1 X
= lim ck e−2πik(ξ/2) ϕ(ξ/2).
b
2 n→∞
k∈Fn
1 X
m0 (ξ) = lim ck e−2πikξ ,
2 n→∞
k∈Fn
and by passing to a subsequence yet again we may assume that this conver-
gence is pointwise a.e. Hence for almost every ξ, since Fn is a finite set we can
compute that
X
1
b
ϕ(ξ) = lim ck e−2πik(ξ/2) ϕ(ξ/2)
b
2 n→∞
k∈Fn
X
1
= lim ck e−2πik(ξ/2) ϕ(ξ/2)
b
2 n→∞
k∈Fn
b
= m0 (ξ/2) ϕ(ξ/2).
b
ϕ(ξ) b
= m0 (ξ/2) ϕ(ξ/2)
X
1
= ck e−2πik(ξ/2) ϕ(ξ/2)
b
2
k∈Z
X
= 2−1/2 ck 2−1/2 e−2πik(ξ/2) ϕ(ξ/2)
b
k∈Z
X
= 2−1/2 ck D1/2 M−k ϕ(ξ)
b
k∈Z
X
2−1/2 ck (D2 Tk ϕ) (ξ).
∧
=
k∈Z
Because all of the series converge absolutely, the justification of the factoring
that relates the second to third lines follows from an argument that is similar
to (but even more simple than) the ones used in Exercise 10.15 or 12.17. Since
the Fourier transform is a unitary mapping, it follows that
X
ϕ = 2−1/2 ck D2 Tk ϕ.
k∈Z
(a) Since (ck )k∈Z ∈ ℓ2 (Z), the series defining m0 converges uncondition-
P
ally in L2 (T). If {Tk ϕ}k∈Z is a Bessel sequence, then the series ck D 2 T k ϕ
converges unconditionally in L2 (R). Therefore we can compute just as in the
argument above, appealing this time to Exercise 10.15 for justification of the
factoring.
12.19 (a) We compute that
226 Detailed Solutions
XX X X
|cj dk−j | = |cj | |dk−j |
k∈Z j∈Z j∈Z k∈Z
X X
= |cj | |dk |
j∈Z k∈Z
1X X
= dk cj−k (ϕ ∗ ψ)(2x − j)
2
k∈Z j∈Z
1X X
= dk cj−k (ϕ ∗ ψ)(2x − j)
2
j∈Z k∈Z
1X
= (c ∗ d)j (ϕ ∗ ψ)(2x − j).
2
j∈Z
Hence ϕ ∗ ψ is refinable with refinement coefficients 12 (c ∗ d)j j∈Z .
The use of Fubini’s Theorem is justified by performing a similar caculation
to show that
Detailed Solutions 227
Z ∞ Z ∞ XX
|cj | |dk | |ϕ(2y − j)| |ψ(2x − 2y − k)| dy
−∞ −∞ j∈Z k∈Z
1
≤ kc ∗ dkℓ1 kϕ ∗ ψkL1 < ∞.
2
The symbol for this refinement equation is
∞
1 X 1
mc∗d (ξ) = (c ∗ d)k e−2πikξ
2 2
k=−∞
∞ ∞
1 X X
= cj dk−j e−2πi(j+k−j)ξ
4
k=−∞ j=−∞
∞ ∞
1 X X
= cj e−2πijξ dk−j e−2πi(k−j)ξ
4 j=−∞
k=−∞
∞
X ∞
X
1
= cj e−2πijξ dk e−2πikξ
4 j=−∞ k=−∞
∞
X ∞
1 −2πijξ 1 X −2πikξ
= cj e dk e
2 j=−∞
2
k=−∞
= mc (ξ) md (ξ).
B0 1 1
B1 1/2 1 1/2
Fix 0 ≤ j ≤ n. If 1 ≤ ℓ ≤ j then
0 ≤ j − ℓ ≤ n − 1,
(j−ℓ)
so Mn−1 (1/2) = 0. Therefore
j
X
(j) j (j−ℓ)
Mn(j) (1/2) = M0 (1/2) Mn−1(1/2) + (−πi)ℓ · 0 · Mn−1 (1/2)
ℓ
ℓ=1
= 0.
Detailed Solutions 229
(j)
That is, Mn (1/2) = 0 for j = 0, . . . , n.
(c) By definition, B0 = χ. Exercise 9.5 shows that L1 (R) is closed under
convolution, so we have by induction that Bn = Bn−1 ∗ χ ∈ L1 (R) for every n,
and also Bn is compactly supported. Let
sin πξ
dπ (ξ) = ,
πξ
sin πξ
b (ξ) = M−1/2 dπ (ξ) = e−πiξ
χ .
πξ
Therefore
n+1 n+1
cn (ξ) = sin πξ
B b
χ (ξ) = e−πi(n+1)ξ ,
πξ
cn (ξ) (1 − e−2πiξ )
= B
cn (ξ) e−πiξ (eπiξ − e−πiξ )
= B
cn (ξ) e−πiξ 2i sin(πξ)
= B
[
= −B ′
n+1 (ξ).
∧
[
Therefore (T1 Bn − Bn ) = B ′
n+1 , so by the Uniqueness Theorem we have
′
T1 Bn − Bn = Bn+1 .
(e) We know that B1 = χ ∗ χ ∈ Cc (R). Also, the fact that Bcn has the form
c
given in part (a) implies that Bn , ξ n−1 c 1
Bn (ξ) ∈ L (R), so Theorem 9.15 and
the Inversion Formula imply that
230 Detailed Solutions
cn )
Bn = (B
∨
∈ C0n−1 (R).
= [ϕ(j)]j=1,...,N .
T
Thus ϕ(1), . . . , ϕ(N − 1) is a 1-eigenvector of the matrix M.
(c) If ϕ is differentiable, then we have
N
X N
X
d
ϕ′ (x) = ck ϕ(2x − k) = 2ck ϕ′ (2x − k),
dx
k=0 k=0
Detailed Solutions 231
so ϕ′ is refinable.
′ ′
(d) If ϕ is differentiable and ϕ is continuous, then applying part (b) to ϕ
we see that ϕ′ (1), . . . , ϕ′ (N −1) is an eigenvalue for 2M for the eigenvalue 1.
Hence 1/2 is an eigenvalue for M.
(e) Since M is an (N −1)×(N −1) matrix, it can have at most N −1 distinct
eigenvalues. In order for ϕ to have N − 2 continuous derivatives, each of 1,
1/2, . . . , 2−(N −1) must be eigenvalues of M. This is the maximum possible
number of continuous derivatives that ϕ can have.
12.22 Suppose ϕ ∈ L2 (R) is refinable and compactly supported and
{Tk ϕ}k∈Z is orthonormal. The refinement equation is
X
ϕ(x) = 2−1/2 ck 21/2 ϕ(2x − k).
k∈Z
1XX 1XX
= cn ck e−2πi(n−k)ξ + cn ck e−2πi(n−k)(ξ+1/2)
4 4
k∈Z n∈Z k∈Z n∈Z
1XX 1XX
= cn+k ck e−2πinξ + cn+k ck e−2πin(ξ+1/2)
4 4
k∈Z n∈Z k∈Z n∈Z
1XX 1XX
= cn+k ck e−2πinξ + cn+k ck e−πin e−2πinξ
4 4
k∈Z n∈Z k∈Z n∈Z
1XX 1XX
= cn+k ck e−2πinξ + cn+k ck (−1)n e−2πinξ
4 4
k∈Z n∈Z k∈Z n∈Z
1XX
= (cn+k + (−1)n cn+k ) ck e−2πinξ
4
k∈Z n∈Z
1XX
= c2n+k ck e−2πinξ
2
k∈Z n∈Z
1X X
= c2n+k ck e−2πinξ
2
n∈Z k∈Z
1X
= a−n e−2πinξ
2
n∈Z
1X
= an e2πinξ .
2
n∈Z
c(n) = an
Since a ∈ ℓ1 (Z), it follows that the Fourier coefficients of M are M
for n ∈ Z. Consequently, M = 1 a.e. if and only if a/2 is the delta sequence,
i.e., an = 2δ0n for n ∈ Z.
12.24 Without loss of generality, we can assume that 0 < δ ≤ 1. Since
|eiθ − 1| ≤ min{|θ|, 2}, we compute that
1 X 1 X
|m0 (ξ) − 1| = ck e 2πikξ
− ck
2 2
k∈Z k∈Z
1 X 2πikξ
≤ |e − 1|δ |e2πikξ − 1|1−δ |ck |
2
k∈Z
X
≤ |πξ|δ |k|δ 21−δ |ck | = C|ξ|δ .
k∈Z
The remainder of the proof is similar to the proof of Theorem 12.25. We have
Detailed Solutions 233
n
Y
|Pn (ξ)| = m0 (2−j ξ) − 1 + 1
j=1
n
Y
≤ C 2−jδ |ξ|δ + 1
j=1
n
Y −jδ
|ξ|δ
≤ eC 2
j=1
Pn
2−jδ |ξ|δ
= eC j=1
′
|ξ|δ
≤ eC .
sup |Pn (ξ) − Pn−1 (ξ)| = sup |m0 (2−n ξ) − 1| |Pn−1 (ξ)|
ξ∈[−R,R] ξ∈[−R,R]
′
Rδ
≤ C 2−nδ Rδ eC .
P∞
Since n=1 2−nδ < ∞, this implies that {Pn }n∈Z is Cauchy in the uniform
norm on [−R, R].
12.25 (a) Since c = (. . . , 0, 0, c0 , . . . , cN , 0, 0, . . . ) is a finite sequence, equa-
tions (12.42) and (12.43) are equivalent. Suppose that N = 2m is even.
If ck and ck+2m are both nonzero then we must have 0 ≤ k ≤ 2m and
0 ≤ k + 2m ≤ 2m. This implies k = 0. Hence
X
0 = δ0,−2m = ck ck+2m = c0 c2m = c0 cN 6= 0,
k∈Z
sin πξ
b (ξ) = M1/2 (χ[− 1 , 1 ] )∧ (ξ) = eπiξ
χ .
2 2 πξ
(b) The box function satisfies the refinement equation χ(x) = χ(2x) +
χ(2x − 1). This is the refinement equation with c0 = c1 = 1 and all other
P
ck = 0. Therefore, we have k∈Z ck ck−2n = 2δ0n , which is equivalent to
|m0 (ξ)|2 + |m0 (ξ + 12 )|2 = 1.
(c) Fix ξ ∈ R. Using the double-angle formula sin 2x = 2 sin x cos x, we
have
234 Detailed Solutions
n
Y
cos(2−j πξ) = cos(πξ) · · · cos(2−n πξ)
j=1
sin(πξ)
=
2n sin(2−n πξ)
sin(πξ)
→ as n → ∞,
πξ
sin x
where at the last step we have used the fact that x → 1 as x → 0. This
yields Viète’s formula.
(d) Note that
1 eπiξ −πiξ
m0 (ξ) = 1 + e2πiξ ) = e + eπiξ
2 2
e πiξ
= 2 cos(πξ = eπiξ cos(πξ).
2
By parts (a), (b) and Theorem 12.29, we have
Y∞
sin πξ
eπiξ b (ξ) =
= χ m0 (2−j ξ)
πξ j=1
∞
Y −j
= eπi2 ξ
cos(2−j πξ)
j=1
P∞ ∞
Y
πi2−j ξ
= e j=1 cos(2−j πξ)
j=1
∞
Y
= eπiξ cos(2−j πξ),
j=1
(j) (−2πi)j X j
m0 (ξ) = k ck e−2πikx ,
2
k∈Z
Detailed Solutions 235
we have
(j) (−2πi)j X j (−2πi)j X
m0 (1/2) = k ck e−πik = (−1)k k j ck .
2 2
k∈Z k∈Z
1
belongs to L (T). In particular, p is 1-periodic. Also, by refinability we have
X
p(x) = ϕ(x + j)
j∈Z
XX
= ck ϕ(2x + 2j − k)
j∈Z k∈Z
X X
= ck ϕ(2x + 2j − k)
k∈Z j∈Z
X X X X
= c2k ϕ(2x + 2j − 2k) + c2k+1 ϕ(2x + 2j − 2k − 1)
k∈Z j∈Z k∈Z j∈Z
X X X X
= c2k ϕ(2x + 2j) + c2k+1 ϕ(2x + 2j − 1)
k∈Z j∈Z k∈Z j∈Z
X X
= ϕ(2x + 2j) + ϕ(2x + 2j − 1)
j∈Z j∈Z
X
= ϕ(2x + j)
j∈Z
= p(2x).
Hence p satisfies both p(x + 1) = p(x) and p(2x) = p(x) for a.e. x ∈ T. The
mapping τ x = 2x mod 1 from [0, 1) onto itself is ergodic. The Birkhoff Ergodic
Theorem [Wal82, Thm. 1.14] therefore implies that there is a constant C such
that
n−1
1X
p(x) = p(τ k x) → C a.e. as n → ∞.
n
k=0
Hence p(x) = C a.e.
To evaluate this constant, we integrate the periodization of ϕ over a period:
Z 1 Z 1X Z ∞
C = C dx = ϕ(x + k) dx = b
ϕ(x) dx = ϕ(0).
0 0 k∈Z −∞
236 Detailed Solutions
12.30 Set X
q(x) = (j + a) ϕ(x + j).
j∈Z
Since ϕ is compactly supported, for any given x only finitely many terms in
this series can be nonzero. Similarly, all sums in this solution will have finitely
many nonzero terms, so there are no convergence issues to consider.
We have
X
q(x + 1) = (j − a) ϕ(x + 1 + j)
j∈Z
X
= (j − a − 1) ϕ(x + j)
j∈Z
X X
= (j − a) ϕ(x + j) − ϕ(x + j)
j∈Z j∈Z
= q(x) − 1.
1X X
= 2j ϕ(2x + 2j) − a ϕ(2x + 2j)
2
j∈Z j∈Z
X
1 X
+ 2k c2k ϕ(2x + 2j)
2
k∈Z j∈Z
1X
+ (2j − 1 + 1) ϕ(2x + 2j − 1)
2
j∈Z
X
− a ϕ(2x + 2j − 1)
j∈Z
X
1 X
+ (2k + 1 − 1) c2k+1 ϕ(2x + 2j − 1)
2
k∈Z j∈Z
1X X
= 2j ϕ(2x + 2j) − a ϕ(2x + 2j)
2
j∈Z j∈Z
X
1 X
+ 2k c2k ϕ(2x + 2j)
2
k∈Z j∈Z
1X
+ (2j − 1) ϕ(2x + 2j − 1)
2
j∈Z
238 Detailed Solutions
1X
+ ϕ(2x + 2j − 1)
2
j∈Z
X
− a ϕ(2x + 2j − 1)
j∈Z
X
1 X
+ (2k + 1) c2k+1 ϕ(2x + 2j − 1)
2
k∈Z j∈Z
X
1 X
− c2k+1 ϕ(2x + 2j − 1)
2
k∈Z j∈Z
1X X
= j ϕ(2x + j) − a ϕ(2x + j)
2
j∈Z j∈Z
aX aX
+ ϕ(2x + 2j) + ϕ(2x + 2j − 1)
2 2
j∈Z j∈Z
1X aX
= j ϕ(2x + j) − ϕ(2x + j)
2 2
j∈Z j∈Z
q(2x)
= .
2
Therefore
q(2x) 2x h(2x)
h(x) = q(x) − x = − = .
2 2 2
Hence h satisfies both h(x + 1) = h(x) and h(2x) = 2h(x) for a.e. x ∈ T.
The mapping τ x = 2x mod 1 from [0, 1) onto itself is ergodic. The Birkhoff
Ergodic Theorem [Wal82, Thm. 1.14] therefore implies that there is a constant
C such that
n−1 Z 1
1X
h(τ k x) → C = h(x) dx a.e. as n → ∞.
n 0
k=0
But
n−1 n−1 n−1
1X 1X 1X k 2n − 1
h(τ k x) = h(2k x) = 2 h(x) = h(x).
n n n n
k=0 k=0 k=0
Since this converges to the finite number C we must have h(x) = 0 a.e.
12.31 (a) The fact that the Hilbert transform is unitary follows from the
Plancherel Equality and the fact that | − i sign(ξ)| = 1 for every ξ.
If f ∈ L1 (R)∩L2 (R), then fb is a continuous function. However, if fb(0) 6= 0,
∧
then (Hf ) is not continuous, so we cannot have Hf ∈ L1 (R).
(b) By Lemma 12.15, we have
Detailed Solutions 239
b
ϕ(ξ) b
= m0 (ξ/2) ϕ(ξ/2) a.e.
Let θ(ξ) = −i sign(ξ), and note that θ(2ξ) = θ(ξ) for every ξ. Then for almost
every ξ we have
∧ ∧
b
(Hϕ) (ξ) = θ(ξ) ϕ(ξ) b
= θ(ξ/2) m0 (ξ/2) ϕ(ξ/2) = m0 (ξ/2) (Hϕ) (ξ/2).
Since (ck )k∈Z ∈ ℓ1 (Z), Exercise 12.18 implies that Hϕ satisfies the refinement
equation.
12.32 Since the hypotheses of Theorem 12.25 are satisfied, we know that the
infinite product defining P converges uniformly on compact sets. Also,
C = kP · χ[−1,1] k∞ < ∞.
Suppose 2n−1 ≤ |ξ| < 2n , where n > 1. Then 2−n ξ ∈ [−1, 1], so
n
Y
|P (ξ)| = |m0 (2−j ξ)| |P (2−n ξ)| ≤ C km0 kn∞ ≤ C Rn .
j=1
Set α = log2 R, and let M > α be an integer. Then since α > 0 and |ξ| ≥ 1,
we have
n−1
Rn = R 2log2 R = R 2(n−1)α ≤ R |ξ|α ≤ R |ξ|M .
|P (ξ)| ≤ CR |ξ|M .
b
ϕ(ξ) b
= m0 (ξ/2) ϕ(ξ/2).
d −1/2
(D1/2 f )′ (x) = 2 f (x/2) = 2−3/2 f ′ (x/2),
dx
and by induction,
Hence δ (j) = 2j+1 2−1/2 D2 δ (j) . In distributional shorthand notation, this says
that δ (j) (x) = 2j+1 δ (j) (2j x).
12.35 (a) Theorem 12.25 implies that P is a continuous function, hence is
locally integrable. By Exercise 12.32, P has polynomial growth at infinity, and
∨
therefore defines a tempered distribution. Its inverse Fourier transform µ = P
is therefore a tempered distribution as well. Given f ∈ S(R) we have
X X
f, 2−1/2 ck D2 Tk µ = 2−1/2 ck T−k D1/2 f, µ
k∈Z k∈Z
X
2−1/2 ck (T−k D1/2 f ) , µ
∧
= b
k∈Z
Detailed Solutions 241
X
= 2−1/2 ck Mk D2 fb, µ
b
k∈Z
X Z ∞
= 2 −1/2
ck e2πikξ 21/2 fb(2ξ) P (ξ) dξ
k∈Z −∞
X Z ∞
dξ
= ck e2πikξ/2 fb(ξ) P (ξ/2)
−∞ 2
k∈Z
Z ∞
1X dξ
= fb(ξ) ck e2πikξ/2 P (ξ/2)
−∞ 2 2
k∈Z
Z ∞
dξ
= fb(ξ) m0 (ξ)/2) P (ξ/2)
−∞ 2
Z ∞
dξ
= fb(ξ) P (ξ)
−∞ 2
= fb, µ
b = hf, µi.
P
Therefore k∈Z 2−1/2 ck D2 Tk µ = µ.
(b) Rearranging the computations in part (a), if ν ∈ S ′ (R) satisfies the
refinement equation and νb is continuous, then we have νb(ξ) = νb(ξ/2) m0 (ξ/2).
Since the hypotheses of Theorem 12.25 are satisfied, iterating this equation
leads to
νb(ξ) = νb(0) P (ξ) = νb(0) µ
b(ξ).
Hence ν is a scalar multiple of µ.
(c) Since P is locally integrable with polynomial growth and θ is bounded,
the product θP is locally integrable with polynomial growth, and therefore
is a tempered distribution. If we set νb = θP, then νb(ξ) = νb(ξ/2) m0 (ξ/2).
Then a computation similiar to the one from part (a) shows that ν satisfies
the refinement equation in a distributional sense.
(d) Given any ξ, we have θ(ξ) = θ(2−n ξ) for every n, so
Suppose that 0 < m ≤ n and 2m−1 ≤ |ξ| < 2m . Then 2−m ξ ∈ [−1, 1], so
m
Y
|Pn (ξ)| = |m0 (2−j ξ)| |Pn−m (2−m ξ)| ≤ C km0 km m
∞ ≤ CR .
j=1
Detailed Solutions 243
Suppose on the other hand that m > n and 2m−1 ≤ |ξ| < 2m . Since Pn
is 2n -periodic, there is some integer ℓ such that |ξ − 2n ℓ| ∈ [0, 2n ]. Hence
2i ≤ |ξ − 2n ℓ| < 2i−1 for some i ≤ n, so
Just as in the proof of Exercise 12.32, if we set α = log2 R and let M > α be
an integer then we have
so
D
|fb(ξ)| ≤ , |ξ| ≥ 1.
|ξ|M+2
Choose T > 1 large enough that
Z
1
dξ ≤ ε.
|ξ|>T |ξ|2
Then
|hf, µ − µn i|
= fb, µ
b−µ cn
Z T Z
≤ |fb(ξ)| |P (ξ) − Pn (ξ)| dξ + |fb(ξ)| |P (ξ) − Pn (ξ)| dξ
−T |ξ|>T
Z
≤ kfbkL1
P − Pn · χ[−T,T ]
∞ + |fb(ξ)| 2C |ξ|M dξ
|ξ|>T
Z
12
≤ kfbkL1
P − Pn · χ[−T,T ]
∞ + 2CD dξ
|ξ|>T |ξ|
≤ kfbkL1
P − Pn · χ[−T,T ]
∞ + 2CDε.
Hence
lim sup |hf, µ − µn i| ≤ 0 + 2CDε.
n→∞
244 Detailed Solutions
w*
Since ε is arbitrary, we conclude that hf, µn i → hf, µi, and therefore µn −→ µ.
Suppose that f ∈ Cc∞ (R) satisfies supp(f ) ⊆ R\[0, N ]. Then since
supp(µn ) ⊆ [0, N ], we have hf, µn i = 0 by definition of support. Consequently
where
1X
m1 (ξ) = dk e−2πikξ .
2
k∈Z
1X
= ck eπik e2πi(k−1)ξ
2
k∈Z
Detailed Solutions 245
1X 1
= e−2πiξ ck e2πik(ξ+ 2 )
2
k∈Z
−2πiξ
= e m0 (ξ + 21 ).
ξ
ξ
ξ 1
ξ 1
= m1 2 m0 2 + m1 2 + 2 m0 2 + 2
= 0,
246 Detailed Solutions
b
|ϕ(ξ)|2 b 2 = |m0 (ξ/2)|2 |ϕ(ξ/2)|
+ |ψ(ξ)| b 2
+ |m1 (ξ/2)|2 |ϕ(ξ/2)|
b 2
2
b
= |ϕ(ξ/2)| a.e.
(b) Applying part (a), we turn the series into a telescoping sum:
N
X N
X
b n ξ)|2 =
|ψ(2 b n−1 ξ)|2 − |ϕ(2
|ϕ(2 b n ξ)|2
n=1 n=1
2
b
= |ϕ(ξ)| b N ξ)|2 .
− |ϕ(2 (A)
If we know in addition that ϕ ∈ L1 (R) then we have by the the Riemann–
Lebesgue Lemma that ϕ b is continuous and ϕ(ω)
b → 0 as |ω| → ∞. Hence, in
this case we have
∞
X
b n ξ)|2 =
|ψ(2 lim b 2 − |ϕ(2
|ϕ(ξ)| b N ξ)|2 b
≤ |ϕ(ξ)|2
a.e.
N →∞
n=1
However, if we only know that ϕ ∈ L2 (R), then we can reach the same
conclusion by applying the Lebesgue Dominated Convergence Theorem. Note
that by equation (A), for every N we have
N
X
b n ξ)|2 = |ϕ(ξ)|
|ψ(2 b 2 − |ϕ(2
b N ξ)|2 ≤ |ϕ(ξ)|
b 2
.
n=1
P∞ b n ξ)|2 converges pointwise almost everywhere
Now, the series ρ(ξ) = n=1 |ψ(2
in the extended real sense since each term is nonnegative. For almost every ξ
we have that
N
X ∞
X
lim b n ξ)|2 =
|ψ(2 b n ξ)|2 ≤ |ϕ(ξ)|
|ψ(2 b 2 ∈ L1 (R).
N →∞
n=1 n=1
PN b n ξ)|2
The Dominated Convergence Theorem therefore implies that n=1 |ψ(2
P∞ b n ξ)|2 in L1 -norm. Hence
converges to n=1 |ψ(2
Z ∞ X∞
0 ≤ b
|ϕ(ξ)|2
− b n ξ)|2 dξ
|ψ(2
−∞ n=1
Z ∞ N
X
= b
lim |ϕ(ξ)|2
− b n 2
|ψ(2 ξ)| dξ
−∞ N →∞ n=1
Z ∞ N
X
= lim b
|ϕ(ξ)|2
− b n ξ)|2 dξ
|ψ(2
N →∞ −∞ n=1
Detailed Solutions 247
Z ∞
= lim b N ξ)|2 dξ
|ϕ(2
N →∞ −∞
Z ∞
−N 2
= lim 2 b
|ϕ(ξ)| dξ
N →∞ −∞
Hence ∞
X
b n ξ)|2 = |ϕ(ξ)|
|ψ(2 b 2 a.e.
n=1
(c) Theorem 12.6 implies that if ψ is any function that generates a dyadic
P b n 2
wavelet orthonormal basis, then we must have n∈Z |ψ(2 ξ)| = 1 a.e.,
whether ψ is associated with an MRA or not. In this problem we are given the
additional information that ψ is associated with an MRA and asked to try to
prove that the same conclusion holds without appealing to Theorem 12.6. I’m
not sure how to do this if we assume only that the scaling function ϕ belongs
to L2 (R). However, if we assume ϕ ∈ L1 (R) ∩ L2 (R) then ϕ b is continuous
b
and we have ϕ(0) = 1. In this case we can use use part (b) and compute that
X ∞
X
b n ξ)|2 =
|ψ(2 lim b n ξ)|2
|ψ(2
N →∞
n∈Z n=−N
∞
X
= lim b n−N −1 ξ)|2
|ψ(2
N →∞
n=1
∞
X
= lim b n (ξ/2N +1 ))|2
|ψ(2
N →∞
n=1
N +1 2
= b
lim |ϕ(ξ/2 )|
N →∞
2
b
= |ϕ(0)| = 1.
12.39 Since the boundary points of intervals have measure zero, we can ignore
them throughout.
(a) The set E is congruent mod 1 to
5 1 5 2 1 2
7 , 1 ∪ 2 , 7 ∪ 7 , 2 ∪ 0, 7 = [0, 1].
Therefore E tiles the real line by integer translates with overlaps of measure
zero.
Also, dilates by 2 of E have overlaps of measure zero. Since
E/4 ∪ E/2 ∪ E ∪ 2E ∪ 4E
contains
248 Detailed Solutions
1 4
4 8 2 16
2, 7 ∪ 1, 87 ∪ 72 , 21 ∪ 2, 16
7 ∪ 7, 1 ∪ 7, 2 = 7, 7
and dilations by 2 of this last interval cover the real line, we see that E tiles
the positive axis by dilations. Since E is symmetric, it also tiles the negative
axis.
Exercise 12.4 therefore implies that E is a wavelet set.
(b) Define ψb = χE , and suppose that ψ was associated with an MRA and
a scaling function ϕ. Then by Exercise 12.38, we have
∞
X
b
|ϕ(ξ)|2
= b n ξ)|2 = χF
|ψ(2
n=1
where
F = − 78 , −1 ∪ − 74 , − 12 ∪ − 72 , 72 ∪ 21 , 74 ∪ 1, 87 .
(c) Since
b
ϕ(ξ) b
= m0 (ξ/2) ϕ(ξ/2),
2 2
we conclude that m0 (ξ) = 1 on − 7 , 7 . Since m0 is 1-periodic, it is therefore 1
on the interval 75 , 79 . Hence if ξ ∈ 1, 87 then we have
b
0 = ϕ(2ξ) b
= m0 (ξ) ϕ(ξ) = 1,
which is a contradiction.
b
(d) Since |ϕ(ξ)| = 1 both for ξ ∈ 0, 71 and for ξ ∈ 1, 87 , we have
X
2
Φϕ (ξ) = b
|ϕ(ξ)| = 2, ξ ∈ 0, 17 .
k∈Z
Thus ν ′′′ is continuous, so ν ∈ C 3 (R). Further, 0 < ν(x) < 1 for x ∈ (0, 1), so
ν is a C 3 sigmoid function.
12.33 (Details for the proof of Theorem 12.33).
(a) Given equation (12.58), which states that
Detailed Solutions 249
= mf (ξ)/m0 (ξ + 12 )
= mf (ξ + 21 ) + 1
2 /m0 (ξ + 12 ) = −λ(ξ + 21 ).
Since translation is isometric on Lp (T), we therefore have for |a| < δ that
kf − Ta f kLp ≤ kf − gkLp + kg − Ta gkLp + kTa g − Ta f kLp
≤ ε + ε + ε = 3ε.
Hence Ta f → f in Lp (T) as a → 0.
13.2 (a) We have
Z 1
f (x − a) e−2πinx dx
∧
(Ta f ) (n) =
0
Z 1
= f (x) e−2πin(x+a) dx
0
Z 1
= e −2πina
f (x) e−2πinx dx = M−a fb(n).
0
(a’) We have
X
c(n − m) e−2πinx
∧
(Tm c) (x) =
n∈Z
X
= c(n) e−2πi(n+m)x
n∈Z
X
= e−2πimx c(n) e−2πinx = M−m b
c (x).
n∈Z
Detailed Solutions 251
(b’) We have
∧
X
(Ma c) (x) = (Ma c)(n) e−2πinx
n∈Z
X
= e2πian c(n) e−2πinx
n∈Z
X
= c(n) e−2πin(x−a) = b
c (x − a).
n∈Z
= 0 + 2πinfb(n).
c (ξ + h) − b
b c (ξ) X e−2πin(ξ+h) − e−2πinξ
= cn .
h h
n∈Z
2π|nh|
≤ |cn | |e−2πinh |
|h|
= 2π|ncn | ∈ ℓ1 (Z).
252 Detailed Solutions
b
c (ξ + h) − b
c (ξ) X e−2πin(ξ+h) − e−2πinξ
lim = lim cn
h→0 h h→0 h
n∈Z
X e−2πin(ξ+h) − e−2πinξ
= cn lim
h→0 h
n∈Z
X
= −2πi ncn e−2πinξ .
n∈Z
b
cb ′ (ξ) = d(ξ)
(b) Given f , g ∈ ℓ1 (Z), the functions fb and gb are continuous, and the series
defining them converge absolutely. Hence we can compute pointwise that
X
(f ∗ g)(n) e−2πinx
∧
(f ∗ g) (x) =
n∈Z
X X
= f (m) g(n − m) e−2πinx
n∈Z m∈Z
X X
= f (m) e−2πimx g(n − m) e−2πi(n−m)x
m∈Z n∈Z
X X
= f (m) e−2πimx g(n) e−2πinx
m∈Z n∈Z
X
= f (m) e−2πimx gb(x)
m∈Z
= fb(x) gb(x).
13.11 Set f ∗ (x) = f (−x). Then, making the change of variables x 7→ −x,
Z 1/2
fc∗ (−n) = f (x) e−2πi(−n)x dx
−1/2
Z −1/2
= − f (−x) e−2πi(−n)(−x) dx
1/2
Z 1/2
= f (x) e−2πinx dx = fb(n).
−1/2
∨
Finally, f (n) = fb(−n) by definition.
13.12 Let δ = δ0 = δ0n n∈Z . If c ∈ ℓp (Z) then we have
X
(c ∗ δ)(n) = c(n − k) δ(k) = c(n),
k∈Z
so c ∗ δ = c.
254 Detailed Solutions
= kf kL1 kg − Ta gk∞
→ 0 as a → 0.
= kf kLp kg − Ta gkLp′
→ 0 as a → 0.
kf − f ∗ kN kpLp
Z 1 Z 1 p
= f (x) − f (x − t) k (t) dt dx
N
0 0
Z 1 Z 1 p
′
≤ |f (x) − f (x − t)| |kN (t)|1/p |kN (t)|1/p dt dx
0 0
Z 1 Z 1 p/p Z 1 p/p′
p
≤ |f (x) − f (x − t)| |kN (t)| dt |kN (t)| dt dx
0 0 0
Z 1 Z 1
p/p′
= kkN kL1 |f (x) − f (x − t)|p |kN (t)| dt dx
0 0
Z 1 Z 1
p/p′ p
≤ K |f (x) − f (x − t)| dx |kN (t)| dt
0 0
Z 1
′
= K p/p kf − Tt f kpLp |kN (t)| dt.
0
From this point onwards, the proof is identical to the proof of Theorem 13.13,
using the fact that translation is strongly continuous in Lp (T) when p < ∞.
13.17 By Exercise 13.15 and 11.9, we can create an approximate identity
{kN } such that kN ∈ C ∞ (T) for every N. Given f ∈ Lp (T), it follows from
Exercise 13.13 that f ∗ kN ∈ C ∞ (T). By Theorem 13.13 we have f ∗ kN → f
256 Detailed Solutions
in Lp -norm. Hence C ∞ (T) is dense in Lp (T) when 1 ≤ p < ∞, and the same
argument with p = ∞ establishes denseness in C(T).
13.18 If f is Hölder continuous on T, then, using equation (13.13), we have
Z Z α
1 1 1 1 1 1 α 1 1
|fb(n)| ≤ f (x) − f x − dx ≤ dx = .
2 0 2n 2 0 2n 2 2|n|
13.19 By definition and the fact that (δm ) (t) = e2πimt , we have
∨
N
X
∨
χN (x) = e2πimx = dN (x).
m=−N
∨
Also, χbN = χN since χN is even.
Let ω = e2πix , and let
N
X N
X
s = dN (x) = e2πimx = ωm.
m=−N m=−N
Then we have
sω = s − ω −N + ω N +1 ,
so
s(ω − 1) = ω N +1 − ω −N .
Multiplying both sides by ω −1/2 = e−πix , we obtain
Now,
ω 1/2 − ω −1/2 = eπix − e−πix = 2i sin πx,
and likewise
so
sin π(2N + 1)x
s = .
sin πx
13.20 We have dN ∈ L1 (T) since it is continuous on T, and also
Z 1 Z 1 N
X
dN = e2πinx dx = 1.
0 0 n=−N
However, we will show that sup kdN kL1 = ∞. Using the fact that
Z 1 1
− cos πx 2
sin πx dx = = π,
0 π 0
Detailed Solutions 257
0.4
0.2
-0.2
-0.4
2
N
X 1 1 1
≤ α + + −
π2 k 2 π
k=1
N
1 1 2 X1
= α+ − + 2 ,
2 π π k
k=1
where Z 1
sin πx
α = dx ≈ 0.58949 < 1.
0 πx
Hence
N N
2 4 X1 4 X1
kdN k1 ≤ 2α + 1 − + 2 ≈ 2 + 1.54236 . . . .
π π k π k
k=1 k=1
Hence
N
2 4 X1
kdN k1 ≤ 2α + 1 − + 2
π π k
k=1
2 4
≤ 2α + 1 − + 2 (1 + ln N )
π π
2 4 4
= 2α + 1 − + 2 + 2 ln N,
π π π
where γ ≈ 0.577 . . . is Euler’s constant. Numerically,
2 4 4 4
kdN k1 ≤ 2α + 1 − + 2 + 2 ln N ≈ 2 ln N + 1.94764 . . . .
π π π π
13.21 (a) We have
XN XN n
X
1 1
σN = sn = ak
N + 1 n=0 N + 1 n=0
k=−n
N
X N
X
1
= ak
N +1
k=−N n=|k|
N
|k|
N
X X
N − |k| + 1
= ak = 1− ak .
N +1 N +1
k=−N k=−N
lim WN (k) ak = ak
N →∞
and
WN (k) |ak | ≤ |ak |.
260 Detailed Solutions
More generally, suppose that we know that the partial sums sN converge
to L. Choose any ε > 0. Then there exists an M such that
N > M =⇒ |L − sN | < ε.
N Z
X |n| 1
−2πint
= 1− f (t) e dt e2πinx
N +1 0
n=−N
Detailed Solutions 261
Z N
1 X |n| 2πin(x−t)
= f (t) 1− e dt
0 N +1
n=−N
Z 1
= f (t) wN (x − t) dt
0
= (f ∗ wN )(x).
∨ PN
13.23 (a) Fix any N ∈ N. Note that χN (x) = n=−N e2πinx . Using Exer-
cise 13.21, we can therefore compute that
∨
N
X |n| 2πinx
WN (x) = 1− e
N +1
n=−N
∨ ∨
χ0 (x) + · · · + χN (x)
=
N +1
N
1 X sin(2N + 1)πx
=
N + 1 n=0 sin πx
N
1 X e(2n+1)πix − e−(2n+1)πix
=
N + 1 n=0 2i sin πx
XN X N
1 1
= eπix e2nπix − e−πix e−2nπix
N + 1 2i sin πx n=0 n=0
2πi(N +1)x −2πi(N +1)x
1 1 πix e −1 −πix e −1
= e − e
N + 1 2i sin πx e2πix − 1 e−2πix − 1
2πi(N +1)x
1 1 e −1 e−2πi(N +1)x − 1
= −
N + 1 2i sin πx eπix − e−πix e−πix − eπix
2πi(N +1)x
1 1 e − 2 + e−2πi(N +1)x
=
N + 1 2i sin πx eπix − e−πix
πi(N +1)x 2
1 1 (e − e−πi(N +1)x)
=
N + 1 2i sin πx 2i sin πx
1 (2i sin π(N + 1)x)2
=
N +1 (2i sin πx)2
2
1 sin π(N + 1)
= .
N +1 sin πx
N
X |n|
= 1− δ0n = 1.
N +1
n=−N
13.24 (a) Suppose that f ∈ L1 (T) and fb ∈ ℓ2 (Z). Then we have fb ∈ ℓ1 (Z),
so the Inversion Formula applies. In particular, f is continuous and bounded
on T, so f ∈ L2 (T).
(b) If f ∈ L2 (T), then we know that the Plancherel Equality holds.
So, suppose that f ∈ L1 (T) \ L2 (T). Then we must have fb ∈ / ℓ2 (Z) by
b
part (a). Hence we have both kf kL2 = ∞ and kf kL2 = ∞, so again the
Plancherel Equality holds.
13.25 Integration by parts shows that
Z 1 i ,
n 6= 0,
xe −2πinx
dx = 2πn
0
1
, n = 0,
2
and
πin + 1
Z 1 , n 6= 0,
2π 2 n2
x2 e−2πinx dx =
0 1, n = 0.
3
Hence the Fourier coefficients of f (x) = π 2 (x2 −x+ 61 ) (extended periodically)
are 1 1 1
fb(0) = π 2 − + = 0
3 2 6
and, for n 6= 0,
Detailed Solutions 263
πin + 1 i πin + 1 πin 1
fb(n) = π 2 − = π2 − = .
2π 2 n2 2πn 2
2π n 2 2
2π n 2 2n2
Hence fb ∈ ℓ1 (Z), and therefore the Fourier series for f converges uniformly
on T to the continuous function f . Combining positive and negative terms,
we therefore have that
X∞ X X
cos 2πnx 1 2πinx
= e = fb(n) e2πinx = f (x),
n=1
n2 2n 2
n∈Z,n6=0 n∈Z
where the series converges uniformly on [0, 1). Since f and each term in the
series is 1-periodic, we also have uniform convergence on any compact subset
of R, except that we must remember that f is periodic, and hence is given by
the formula f (x) = π 2 (x2 − x + 61 ) only for x ∈ [0, 1].
13.26 (a) If f ∈ C 2 (T), then applying Exercise 13.3 twice, we have
Therefore
X X |fc′′ (n)| X kfc′′ k∞ X kf ′′ kL1
kfbkL1 = |fb(n)| ≤ 2 2
≤ 2 2
≤ < ∞.
4π n 4π n 4π 2 n2
n∈Z n∈Z n∈Z n∈Z
Then we have
S
∞
A(T) = FN .
N =1
≤ kf − fk k∞ → 0 as k → ∞.
= lim inf N = N.
k→∞
Thus FN is closed.
Finally, since A(T) is a proper subspace of C(T), it cannot contain any
open subsets of C(T). Hence FN has no interior points, and therefore is
nowhere dense in C(T). Therefore A(T) is a countable union of nowhere
dense sets, and hence is meager by definition.
Detailed Solutions to Exercises from Chapter 14
= fb · χ[0,2N ]
o ∧
= (SN f ) . ⊓
⊔
14.2 (a) ⇒ (b) + (c). This follows because the basis constant for {e2πinx }n∈Z
a p
is the supremum of all kSN kLp →Lp and kSN kL →Lp .
(b) ⇐⇒ (d). If f ∈ Lp (T) then
o
kSN f kLp = kMN SN M−N f kLp
= kSN M−N f kLp
≤ kSN kLp→Lp kM−N f kLp
= kSN kLp→Lp kf kLp .
o p
Hence kSN kL →Lp ≤ kSN kLp →Lp . Conversely,
Therefore,
2N
X
t t t
SN SN f (x) = SN g(x) = −i sign(n) gb(n) e2πinx
n=−2N
2N
X
= −i sign(n) (−i) sign(n) fb(n) e2πinx
n=−2N
2N
X
= − fb(n) e2πinx + fb(0)
n=−2N
Rearranging,
t t
S2N f = −SN SN f + fb(0).
Also,
2N
X 2N
X
t
S2N f (x) + iSN f (x) = fb(n) e2πinx + sign(n) fb(n) e2πinx
n=−2N n=−2N
2N
X
= 2 fb(n) e2πinx + fb(0)
n=1
2N
X
= 2 fb(n) e2πinx − fb(0)
n=0
o
= 2 SN f (x) − 2 fb(0).
Therefore,
o t t t t
2 SN f = S2N f + iSN f + fb(0) = −SN SN f + iSN f + 2 fb(0).
Hence
o t t t
2 kSN f k ≤ kSN SN f kLp + kSN f kLp + kfb(0) · 1kLp
t t
≤ kSN k2Lp→Lp kf kLp + kSN kLp →Lp kf kLp + 2 |fb(0)| k1kLp
t t
≤ kSN k2Lp→Lp kf kLp + kSN kLp →Lp kf kLp + 2 kf kLp · 1
t t
= kSN k2Lp →Lp + kSN kLp →Lp + 2 kf kLp .
Thus,
o 1 t 1 t
kSN f k ≤ kS k2 p p + kSN kLp →Lp + 1.
2 N L →L 2
Detailed Solutions 267
g ′ (0) = f ′ (a).
Further,
Since (Hfk ) (n) = −i sign(n) fbk (n) for |n| ≤ 2N and 0 otherwise, we conclude
∧
f + iHf fb(0)
Rf = + · 1.
2 2
For n > 0 we therefore have that
For n = 0 we have
= fb · χ[0,∞) − T2N +1 (T−2N −1 fb) · χ[0,∞)
′
Therefore H is bounded on a dense subset of Lp (T), and therefore is bounded
′
on all of Lp (T) by Theorem 14.7. Moreover, the above work shows that
B.1 (c) Suppose that {xn }n∈N is a bounded sequence in X. Since T1 is com-
(1) (1)
pact, there exists a subsequence {xn }n∈N of {xn }n∈N such that {T1 xn }n∈N
(2)
converges. Then since T2 is compact, there exists a subsequence {xn }n∈N of
(1) (2) (2)
{xn }n∈N such that {T2 xn }n∈N converges (note that {T1 xn }n∈N also con-
verges). Continue to construct subsequences in this way.
(m) (k)
Note that xm is a member of the subsequence {xn }n∈N for each
(m)
k = 1, . . . , m. In particular, for every m we have that xm belongs to the
(1) (m)
subsequence {xn }n∈N . For m ≥ 2 we have that xm belongs to the sub-
(2) (m) (k)
sequence {xn }n∈N . In general, {xm }m≥k is a subsequence of {xn }n∈N .
(m) (m)
Therefore {Tk xm }m≥k converges. For simplicity, let ym = xm . Since our
original sequence is bounded, R = sup kym k < ∞.
Fix ε > 0. Then there exists some k such that kT − Tk k < ε. Since {Tk ym }
converges it is Cauchy, so there is some some M such that kTk ym − Tk yn k < ε
for all m, n ≥ M. Therefore for all m, n ≥ M we have
kT ym − T yn k ≤ kT ym − Tk ym k + kTk ym − Tk yn k + kTk yn − T yn k
≤ kT − Tk k kym k + ε + kTk − T k kynk
≤ Rε + ε + Rε.
Using this and the continuity of the inner product, we therefore have
where {en }n∈N is an orthonormal basis for the separable Hilbert space H.
Since λ = (λn ) is bounded, we know that Mλ is bounded operator on H.
(a) ⇒. Suppose that λn → 0 as n → ∞. Define
272 Detailed Solutions
N
X
TN f = λn hf, en i en .
n=1
This operator is linear, bounded, and has finite rank since range(TN ) ⊆
span{e1 , . . . , eN }. Further, TN is a good approximation to Mλ , because by
using the Plancherel Equality we can compute that
∞
2
X
2
k(Mλ − TN )f k =
λn hf, en i en
n=N +1
∞
X
= |λn |2 |hf, en i|2
n=N +1
∞
X
≤ sup |λn |2 |hf, en i|2
n>N n=N +1
≤ sup |λn |2 kf k2 .
n>N
Hence λ ∈ c0 .
(b) Suppose that λ ∈ c0 and λn 6= 0 for every n. Exercise 1.66 shows that
range(Mλ ) is a dense but proper subspace of H. Let D = {x ∈ H : kxk ≤ 1},
and suppose that Mλ (D) was closed. Choose any y ∈ H. Since range(Mλ ) is
dense, we can find xn ∈ H such that Mλ xn → y. Since {xn }n∈N is bounded,
there is some r > 0 such that
Mλ xn ∈ Mλ (r D) ∈ r Mλ (D)
for every n. Since r Mλ (D) is closed, this implies that y ∈ r Mλ (D). Hence
y = Mλ x for some x with kxk ≤ r. Thus y ∈ range(Mλ ). Hence Mλ is
surjective, which is a contradiction. Therefore Mλ (D) cannot be closed.
B.4 (a) Let E = {en } be an orthonormal basis for H and let F = {fn } be
an orthonormal basis for K. Then, by the Plancherel Equality and Tonelli’s
Theorem for series,
Detailed Solutions 273
X XX
kT emk2 = |hT em , fn i|2
m m n
XX X
= |hem , T ∗ fn i|2 = kT ∗ fn k2 .
n m n
These quantities may be infinite, but they are either all infinite, or all finite
and equal. THus, no matter what orthonormal bases we choose, we have
X X
kT em k2 = kT ∗ fn k2 .
m n
P
Fixing F, we see that kT kHS = kT emk isPindependent of the choice of
orthonormal basis for H. Likewise, kT ∗ kHS = kT fnk is independent of the
∗
choice of orthonormal basis for K, and furthermore kT P kHS = kT kHS.
Also, if x ∈ H then, since T is continuous, T x = hx, en i T en . Hence
X
kT xk ≤ |hx, en i| kT en k
n
X 1/2 X 1/2
≤ hx, en i|2 kT enk2 = kxk kT kHS,
n n
so kT k ≤ kT kHS.
(b) We have by definition that 0 ≤ kT kHS < ∞ for each T ∈ B2 (H, K).
If kT kHS = 0, then we must have kT enk = 0 for every element of an
orthonormal basis {en }n∈N . By linearity and continuity, we must then have
T f = 0 for every f ∈ H, so T = 0.
Homogeneity follows from the homogeneity of the Hilbert space and ℓ2
norms, i.e., kcT kHS = |c| kT kHS.
The Triangle inequality likewise follows from the corresponding inequali-
ties for H and for ℓ2 . Specifically, if {en }n∈N is any orthonormal basis for H,
then
X ∞ 1/2
kT + U kHS = kT en + U en k2
n=1
X
∞ 1/2
2
≤ kT enk + kU en k
n=1
X
∞ 1/2 X
∞ 1/2
2 2
≤ kT enk + kU en k
n=1 n=1
= kT kHS + kU kHS .
This shows that k · kHS is a norm. Now we must show that B2 (H, K) is
complete with respect to k·kHS. Suppose that {Tn }n∈N is a Cauchy sequence in
274 Detailed Solutions
B2 (H, K). Then by part (a), kTm −Tnk ≤ kTm −Tn kHS , so {Tn }n∈N is Cauchy
in operator norm. Since B(H, K) is complete, there exists T ∈ B(H, K) such
that Tn → T in operator norm. We must show that Tn → T in Hilbert–
Schmidt norm.
If we fix ε > 0, then there exists an N such that kTm − Tn kHS ≤ ε for all
m, n > N. Choose any orthonormal basis {en }n∈N for H. Since Tn → T in
operator norm, for any fixed k we have Tn ek → T ek as n → ∞. Therefore, if
n > N then we have for each M > 0 that
M
X M
X
kT ek − Tn ek k2 = lim kTm ek − Tn ek k2
m→∞
k=1 k=1
M
X
= lim kTm ek − Tn ek k2
m→∞
k=1
∞
X
≤ lim sup kTm ek − Tn ek k2
m→∞
k=1
≤ kT kHS kU kHS,
P
so hT, U iHS = hT en , U en i is well defined. Further, if {fn } is any orthonor-
mal basis for K then, applying Fubini’s Theorem, we have
X XX
hT em , U em i = hT em , fn i hfn , U em i
m m n
XX
= hem , T ∗ fn i hU ∗ fn , em i
m n
Detailed Solutions 275
XX
= hU ∗ fn , em i hem , T ∗ fn i
n m
X
= hU ∗ fn , T ∗ fn i.
n
≤ kT kHS kU kHS .
In any case, hT, U iHS is well defined and independent of the choice of or-
thonormal basis {en }.
Finally, h·, ·iHS satisfies all the properties of an inner product, so since we
have already proved that B2 (H, K) is complete, it is a Hilbert space.
(c) If T ∈ B2 (H, K) and B ∈ B(K) then
∞
X ∞
X
kBT k2HS = kBT en k2 ≤ kBk2 kT en k2 = kBk2 kT k2HS.
n=1 n=1
kT AkHS = k(T A)∗ kHS = kA∗ T ∗ kHS ≤ kA∗ k kT ∗kHS = kAk kT kHS.
X d
X d
X
kLen k2 = kLen k2 ≤ kLk2 ken k2 = d kLk2 < ∞.
n n=1 n=1
Hence L is Hilbert–Schmidt.
Now suppose that T ∈ B2 (H, K) is given, and let {en } be an orthonormal
basis for H. For each N ∈ N define
N
X
TN x = hx, en i T en , x ∈ H.
n=1
276 Detailed Solutions
where xk = T ∗ yk .
⇐. If T has this form, then range(T ) ⊆ span{y1 , . . . , yN }, so T has finite
rank.
(c) By part (b),
= hx2 , x1 i hy1 , y2 i.
(d) We have
(x⊗y)(w), z = hw, xi y, z = hw, xi hy, zi = w, hz, yi x = w, y⊗x .
B.8 Since
Z Z
hem ⊗ fn , ej ⊗ fk i = (em ⊗ fn )(x, y) (ej ⊗ fk )(x, y) dx dy
F E
Z Z
= em (x) fn (y) ej (x) fk (x) dx dy
F E
Z Z
= ej (x) em (x) fn (y) fk (y) dy dx
E F
Z
= ej (x) em (x) hfn , fk i dx
E
where Z
hm (y) = F (x, y) em (x) dx.
E
where Ek and Fk have finite measure. Since χEk , χFk ∈ L2 (R), we can write
X X
χE k = amk em and χFk = bnk fn
m∈N n∈N
for some scalars amk and bnk , where the series converge in L2 -norm. Hence,
N
X
H(x, y) = ck χEk (x) χFk (y)
k=1
N
X X X
= ck amk em (x) bnk fn (y)
k=1 m n
X XX
N
= ck amk bnk em (x) fn (y)
m n k=1
XX
= dmn (em ⊗ fn )(x, y),
m n
PN
where dmn = k=1 ck amk bnk are some new scalars (note that this is a finite
sum, so is well defined). The partial sums of this series converge to H and lie
in span{em ⊗ fn }m,n∈N . Therefore F can be approximated arbitrarily closely
by functions in this span. Therefore {em ⊗ fn }m,n∈N is complete.
A third alternative is to show that the Plancherel Equality holds. This
approach is similar to what is done in the proof of Exercise 8.36.
280 Detailed Solutions
B.9 (a) Fix U ∈ B(H) and V ∈ B(K), and choose g ∈ H and h ∈ K. Since
g ⊗ h ∈ B(H, K) is a bounded rank one operator, we know that V (g ⊗ h)U ∗ ∈
B(H, K) is a well-defined bounded operator, and given f ∈ H we have
V (g ⊗ h)U ∗ f = V (g ⊗ h)(U ∗ f )
= V hU ∗ f, gi h
= hf, U gi V h
= (U g ⊗ V h)(f ).
Hence U (g ⊗ h)V ∗ = U g ⊗ V h.
(b) Define T (F ) = V F U ∗ for F ∈ B2 (H, K). Then T (F ) ∈ B2 (H, K) by
Theorem B.8(c), and that theorem also implies that
(U1 ⊗ V1 )(U2 ⊗ V2 ) = U1 U2 ⊗ V1 V2 .
(d) Suppose that U and V are topological isomorphisms. Then by part (c),
(U −1 ⊗ V −1 )(U ⊗ V ) = U −1 U ⊗ V −1 V = I ⊗ I = I,