Solutions To Exercises:, Y) ) Is Dense in X y /2 /2

P1: FNT
CB341/Robinson CB341-Sol February 5, 2001 10:52 Char Count= 0
Solutions to Exercises
Chapter 1
1.1 Let {x j } be a countable dense subset of X , and let {y j } be a countable
dense subset of Y . Then the countable collection {(x j , yk )} is dense in
X × Y , since for any (x, y) ∈ X × Y and any > 0 there exist x j and
yk with
x − x j X < /2 and y − yk Y < /2,
and so
(x j , yk ) − (x, y) X ×Y ≤ .
It follows that X × Y is separable, and by induction it follows that any

finite product of separable spaces is separable.
If M is a linear subspace of X then let {x j } be a countable subset
of X such that for each x ∈ X there is an x j such that |x − x j | < .
Now discard any element x j of this collection for which B(x j , ) does
not intersect M. For each remaining x j , it follows that there exists an
element m j ∈ M such that B(m j , 2) ⊃ B(x j , ). Thus this collection
{m j } has the property that for each element m ∈ M there exists an m j
such that |m − m j | < 2. Applying this construction for the sequence
n = 2−n gives a countable dense subset of M, as required.
1.2 Cover X with the collection of open balls

B(x, ).
x∈X
Since X is compact it follows that there exists a finite covering by such
1
P1: FNT
2 Solutions to Exercises
balls:

N
X⊂ B(x j , ).
j=1
It follows that for each x ∈ X there exists an x j with |x − x j | < as

required.
1.3 We first consider the case of
bounded. If u ∈ Cc0 (
) then clearly u = 0
on ∂
; it follows that if u n ∈ Cc0 (
) converges to u uniformly on
then
u = 0 on ∂
too. We now show that any function in
C00 (
) = {u ∈ C 0 (
) : u = 0 on ∂
}
can be arrived at in this way and hence that this space is the completion
of Cc0 (
) in the sup norm. Let θ be the continuous function

 x, x ≥ 1,
θ (x) = 2x − 1, 1 > x > 12 ,

0, x ≤ 12 ,
and define
u (x) = θ(|u(x)|/)u(x).
Clearly u is continuous on
, and since u is uniformly continuous on

there exists a δ such that
dist(x, ∂
) < δ ⇒ |u(x)| < /2,
that is, such that u (x) = 0 when dist(x, ∂
) < δ. It follows that
u ∈ Cc0 (
), and since,
|u(x) − u (x)| ≤ ,
u converges uniformly to u on
.
It follows that C00 (
) =
Cc0 (
) is the completion of Cc0 (
) in the sup
norm, and Cc (
) is therefore not complete.
0
When
= Rm the limit of any convergent sequence of functions in
Cc0 (Rm ) must tend to zero as |x| → ∞. This is clear, since given > 0
there exists an N such that |u n − u| ≤ for all n ≥ N . In particular, u N
is zero for all x > R N , say, and so |u| ≤ for all x > R N . The space of
all such u,

C00 (Rm ) = u ∈ Cb0 (
) : u(x) → 0 as |x| → ∞ ,
is the appropriate completion of Cc0 (Rm ). For any u ∈ C00 (Rm ), we

P1: FNT
Chapter 1 3
can use the argument above to find an approximating sequence of u ∈
Cc0 (Rm ).
1.4 If { f j } is Cauchy in the · c norm then it is Cauchy in each C n (
)
norm. Since each C n (
) is complete, f j → f in each of these spaces,
so that f ∈ C n (
) for every n and thus f ∈ C ∞ (
). It remains to show
that in fact
f c < ∞
and that
f j − f c → 0
as j → ∞. Since { f j } is Cauchy it certainly follows that for j, k ≥ N

we have

l
cn f j − f k C n (
) <
n=1
for each l < ∞, and taking the limit as k → ∞ gives

l
cn f j − f C n (
) < . (S1.1)
n=1
Using the triangle inequality in each C n (

), 0 ≤ n ≤ l, shows that

l
l
cn f C n (
) < + cn f j C n (
) ,
n=1 n=1
and so
f c ≤ + f j c .
Since (S1.1) holds for all l, we can let l → ∞ to show that

∞

cn f j − f C n (
) < ,
n=1
and so f j → f in the · c norm.

1.5 We show that C 0,γ (
) is a Banach space; the case C r,γ then follows
easily. If the sequence { f j } is Cauchy in C 0,γ (
) then given > 0 there
P1: FNT
exists an N such that for j, k ≥ N we have
|[ f j (x) − f k (x)] − [ f j (y) − f k (y)]|

f j − f k ∞ + sup ≤ .
x,y∈
|x − y|γ
Because C 0 (
) is complete we know that f j converges to some f ∈
C 0 (
). We just need to show that f is Hölder. However, since f k → f
uniformly we have
|[ f j (x) − f (x)] − [ f j (y) − f (y)]| ≤ |x − y|γ
and so
| f (x) − f (y)| ≤ | f j (x) − f j (y)| + |[ f j (x) − f (x)] + [ f j (y) − f (y)]|

≤ C j |x − y|γ + |x − y|γ ,
which shows that f ∈ C 0,γ (

).
1.6 If f ∈ C 1 (
) then |D f (x)| is uniformly bounded on
, by L, say. Since

is convex, given any two points x, y ∈
the line segment joining x
and y lies entirely in
. It follows that

1
| f (x) − f (y)| = D f (y + ξ(x − y)) · (x − y) dξ
0
≤ L|x − y|,
so f is Lipschitz.
1.7 We have

x−z y−z
|u h (x) − u h (y)| = h −m ρ −ρ u(z) dz

h h

x−z
≤ h −m ρ |u(z) − u(z + y − x)| dz

h
≤ C|y − x|γ
by using (1.7) so that u h is also Hölder.

1.8 We prove the result by induction, supposing that it is true for n = k.
Then for n = k + 1 we take p such that
k−1
1 1
+ = 1,
j=1
p j p
P1: FNT
Chapter 1 5
to obtain

| f 1 (x) · · · f k+1 (x)|d x ≤ f 1 L p1 · · · f k−1 L pk−1 f k f k+1 L p .

(S1.2)
Now, we use the standard Hölder inequality, noting that
p p
1= + ;
pk pk+1
thus
p/ pk p/ pk+1
p pk+1
( f k f k+1 ) d x ≤
p
fk k dx f k+1 dx ,

and so
f k f k+1 L p ≤ f k L pk f k+1 L pk+1 ,
which combined with (S1.2) gives (1.31) for n = k + 1. Since the

standard Hölder inequality is (1.31) for n = 2 the result follows.
1.9 Write

|u(x)| d x =
p
|u(x)|q(r − p)/(r −q) |u(x)|r ( p−q)/(r −q) d x.

Now note that

r−p p−q
+ = 1,
r −q r −q
and so using Hölder’s inequality we have

(r − p)/(r −q) ( p−q)/(r −q)
|u(x)| p d x ≤ |u(x)|q d x |u(x)|r d x ,

which becomes
q(r − p)/ p(r −q) r ( p−q)/ p(r −q)
u L p ≤ u L q u L r ,
as required.
1.10 If s ∈ S(
) then it is of the form of (1.10),

n
s(x) = c j χ [I j ](x),
j=1
P1: FNT
where the I j are m-dimensional cuboids, each of the form

m
I = [ak , bk ].
k=1
It clearly suffices to approximate χ [I ] to within in the L p norm using

an element of Cc0 (
). To do this, consider the function

m
χη = φη (xk ; ak , bk ),
k=1
where

 (x − a)/η, a ≤ x ≤ a + η,
φη (x; b, a) = 1, a + η < x < b − η,

(b − x)/η, b − η ≤ x ≤ b.
Clearly χη ∈ Cc0 (
) and converges to χ [I ] in L p (
) as η → 0.
1.11 Since |g(x)| ≤ g∞ almost everywhere, it follows that
| f (x)g(x)| ≤ | f (x)|g∞
almost everywhere, and so

| f (x)g(x)| d x ≤ | f (x)|g∞ d x ≤ f L 1 g∞ ,

as claimed.
1.12 Since {x (n) } is Cauchy, given > 0 there exists an N such that
(n)
x − x (m) ∞ ≤ for all n, m ≥ N .
l
This implies that

(n)
x − x (m) ≤ for all n, m ≥ N . (S1.3)
j j
In particular, we have x (n) (n)

j is Cauchy for each j. So x j → x j as n → ∞.
∞
It is then clear that x = {x j } ∈ l , and taking the limit m → ∞ in (S1.3)
shows that
(n)
x − x j ≤ for all n ≥ N, for all j.
j
It follows that x (n) → x in l ∞ , and so l ∞ is complete.

P1: FNT
Chapter 2 7
1.13 We know that the norm is positive definite, and so
x + λy2 = (x + λy, x + λy) = x2 + 2λ(x, y) + λ2 y2 ≥ 0.
In particular, the quadratic equation for λ,
λ2 y2 + 2λ(x, y) + x2 = 0,
can have only one distinct real root. Therefore the discriminant “b2 −
4ac” cannot be positive (which would give two real roots). In other
words,
4(x, y)2 − 4y2 x2 ≤ 0
or
|(x, y)| ≤ xy,
which is the Cauchy–Schwarz inequality. We can now write
x + y2 = x2 + 2(x, y) + y2

≤ x2 + 2xy + y2
= (x + y)2 ,
giving the triangle inequality.

1.14 We simply expand the left-hand side,
u + v2 + u − v2 = u2 + 2(u, v) + v2 + u2 − 2(u, v) + v2

= 2u2 + 2v2 ,
as required.
1.15 If {u j } is a dense subset of l 2 () then for each element γ ∈ there must
exist a u j that is within of 1 at γ and within of 0 for all other elements
of . Each such u j is distinct. It follows that if is uncountable then so
are the {u j }, and so l 2 () cannot be separable.
Chapter 2
2.1 We can apply the contraction mapping theorem to h n to deduce that h n
has a unique fixed point x ∗ ,
h n (x ∗ ) = x ∗ .
P1: FNT
If we apply h to both sides then
h(h n (x ∗ )) = h n+1 (x ∗ ) = h n (h(x ∗ )) = h(x ∗ ),
showing that h(x ∗ ) is also a fixed point of h n . Since the contraction

mapping theorem guarantees that the fixed point of h n is unique, we
must have h(x ∗ ) = h ∗ , and so h ∗ is also a fixed point of h.
2.2 The interval [1, ∞) is closed but not compact, and the map h : [1, ∞) →
[1, ∞) given by x → x + 1/x satisfies
|h(x) − h(y)| = |x − y|(1 − (x y)−1 )

< |x − y|
but clearly has no fixed point.

However, if X is compact and h : X → X satisfies
h(x) − h(y) < x − y, (S2.1)
suppose that h has no fixed point. Then
h(x) − x > 0 for all x ∈ X,
and since h(x) − x is continuous from X into R it obtains its lower

bound, so that
h(x) − x ≥ for all x ∈ X,
and there exists some y ∈ X such that h(y) − y = . However, if we

take z = h(y) then from (S2.1) we have
h(z) − z < ,
a contradiction. So h has at least one fixed point. Uniqueness follows as

in the proof of the standard contraction mapping theorem.
2.3 Take n = 2−n and apply the result of Exercise 1.2 so that there exists
finite set {x (k) (k) −k
j }, 1 ≤ j ≤ Mk , such that |x − x j | ≤ 2 . Set Nk =
k
j=1 M j , and let {x j } be the sequence
x1(1) , . . . , x M
(1)
1
, x1(2) , . . . , x M
(2)
2
, x1(3) , . . . .
2.4 Suppose that there are solutions xn (t) of
d x/dt = f (x) with x(0) = x0 (S2.2)

P1: FNT
Chapter 2 9
such that xn (τ ) → x ∗ . We need to show that there is a solution of (S2.2)
with x(τ ) = x ∗ . Now, if f is bounded then the sequence xn (t) satisfies
sup |xn (t)| ≤ |x0 | + τ f ∞ and |xn (t) − xn (s)| ≤ f ∞ |t − s|,

t∈[0,τ ]
the conditions of the Arzelà–Ascoli theorem (Theorem 2.5). It follows

that there is a subsequence that converges uniformly on [0, τ ], and as in
the proof of Theorem 2.6 the limit x(t) satisfies (S2.2). Since xn → x
uniformly on [0, τ ], in particular we have x(τ ) = x ∗ as required.
2.5 When |x| = 0 then it follows that
d d
|x|2 = 2|x| |x|,
dt dt
and (2.27) follows immediately. When |x(t0 )| = 0, since C(t) is contin-
uous, for any > 0 we have
d
1
2
|x|2 ≤ [C(t0 ) + ]|x|
dt
for t − t0 small enough, and so it follows from Lemma 2.7 that
2
|x(t)|2 ≤ [C(t0 ) + ](t − t0 ) .
Therefore
|x(t + h)| ≤ [C(t0 ) + ]t,
and so
d
|x| ≤ C(t0 ) + .
dt +
Since this holds for any > 0 we have (2.27).
2.6 If
t
y(t) = b(s)x(s) ds
0
then
dy
= b(t)x(t) ≤ a(t)b(t) + b(t)y(t),
dt
and so
t t
dy
− b(t)y(t) exp − b(s) ds ≤ a(t)b(t) exp − b(s) ds .
dt 0 0
P1: FNT
If a(t) is increasing then we can replace a(t) on [0, T ] with a(T ), and
so

t t
d
y(t) exp − b(s) ds ≤ a(T )b(t) exp − b(s) ds .
dt 0 0
Integrating both sides between 0 and T gives us

T T t
y(T ) exp − b(s) ds ≤ a(T ) b(t) exp − b(s) ds dt
0 0 0
and so
T T
y(T ) ≤ a(T ) b(t) exp b(s) ds dt.
0 t
We can integrate the right-hand side to obtain

T
y(T ) ≤ a(T ) exp b(s) ds −1 ,
0
and so, using (2.28), we have

T
x(T ) ≤ a(T ) exp b(s) ds
0
as claimed.
2.7 As in the proof of Proposition 2.10 we consider the difference of two
solutions, z(t) = x(t) − y(t), which satisfies
dz
= f (x) − g(y)
dt
= f (x) − f (y) + f (y) − g(y).
We now use Lemma 2.9 to deduce that

d
|z| ≤ | f (x) − f (y)| + | f (y) − g(y)|
dt +
≤ L|z| + f − g∞ .
An application of Gronwall’s inequality [(2.21) in Lemma 2.8] now

yields (2.29).
P1: FNT
Chapter 3 11
Chapter 3
3.1 We denote
A1 = {smallest M such that AxY ≤ Mx X for all x ∈ X }
and
A2 = sup AxY .
x X =1
First we take x =
0 and put y = x/x X ; then we have
AyY ≤ A2 ⇒ AxY ≤ A2 x X
for all x ∈ X , and so A1 ≤ A2 . Furthermore, it is clear that, for
any M,
AxY ≤ Mx X for all x∈X ⇒ A2 ≤ M,
and so A2 ≤ A1 . Thus A1 = A2 .
3.2 I is clearly bounded from C 0 ([O, L]) into itself, since
I ( f )∞ ≤ L f ∞ .
For the L 2 bound, first observe, by using the Cauchy–Schwarz inequality,
that I ( f )(x) is defined for all x if f ∈ L 2 . Then
L
|I ( f )|2 = |I ( f )(x)|2 d x
0
L x 2
= f (s) ds dx
0 0
L x x
= ds | f (s)| ds d x
2
0 0 0
L
≤L 2
| f (s)|2 ds
0
≤ L 2 | f |2 .
Thus I is a bounded operator on both spaces.

3.3 Suppose that A−1 y1 = x1 and that A−1 y2 = x2 . Then it is clear that
A(x1 + x2 ) = y1 + y2 .
Since the inverse is unique it follows that
A−1 (y1 + y2 ) = A−1 y1 + A−1 y2 .
P1: FNT
3.4 For each x ∈ X, Pn x converges to x, and so it follows that the sequence
{Pn x}∞
n=1 is bounded:
sup Pn x X < ∞

n∈Z+
for each x ∈ X . From the principle of uniform boundedness (Theo-

rem 3.7) we immediately obtain
sup Pn op < ∞,
n∈Z+
as claimed.
3.5 It is clear that φi (x)φ j (y) is an element of L 2 (
×
) and that

[φi (x)φ j (y)][φk (x)φl (y)] d x d y = δik δ jl ,

×
and so they certainly form an orthonormal set. If k ∈ L 2 (

×
) then
k(·, y) ∈ L 2 (
), and we can write
∞

k(x, y) = u i (y)φi (x),
i=1
where

u i (y) = k(x, y)φi (x) d x.

Since
2

|u i (y)| dy =
k(x, y)φi (x) d x dy
2

≤ |k(x, y)|2 d x |φi (x)|2 d x dy

≤ |k(x, y)|2 d x d y,

×
we have u i ∈ L 2 (
). So we can write
∞
u i (y) = u i (y)φ j (y) dy φ j (y),
j=1
which yields the expression

∞
k(x, y) = k(x, y)φi (x)φ j (y) d x d y φi (x)φ j (y),
i, j=1
×
as claimed.
P1: FNT
Chapter 3 13
3.6 We consider the approximations to A given by the truncated sums,

n
An u = λ j (u, w j )w j .
j=1
Using Lemma 3.12 we see that each operator An is compact. We now

want to show that
A − An op → 0,
and it then follows from Theorem 3.10 that A is compact. However, this
convergence is clear, since

∞

(A − An )u = λ j (u, w j )w j

j=n+1
∞

≤ λn+1 (u, w j )w j

j=n+1
∞
1/2

≤ λn+1 |(u, w j )| 2
j=n+1
≤ λn+1 u,
and λn+1 → 0 as n → ∞. Thus A is compact. That A is symmetric

follows by taking the inner product of Au with v to give
∞

(Au, v) = λ j (u, w j )(v, w j ) = (u, Av).
j=1
3.7 We know from Lemma 3.4 that A−1 exists iff Ker(A) = 0. So we show
that if Ax = 0 then x = 0. Because A is bounded below we have
0 = AxY ≥ kx X ,
so that x X = 0. For y ∈ R(A) we can use the lower bound on A to

deduce that
1 1
A−1 y X ≤ A A−1 yY = yY ,
k k
so that A−1 is bounded.

P1: FNT
3.8 Since G is a solution of the homogeneous equation on both sides of
x = y, we must have

C1 (y)u 1 (x), a ≤ x < y,
G(x, y) =
C2 (y)u 2 (x), y ≤ x ≤ b.
The conditions at y require
C1 (y)u 1 (y) = C2 (y)u 2 (y),

C1 (y)u 1 (y) + p(y)−1 = C2 (y)u 2 (y).
Solving these simultaneous equations for C1 and C2 gives
C1 (y) = u 2 (y)/W p (y) and C2 (y) = u 1 (y)/W p (y),
where
W p (y) = p(y)[u 1 (y)u 2 (y) − u 2 (y)u 1 (y)].
Differentiating W p with respect to y and cancelling the pu 1 u 2 terms

gives
W p = p (u 1 u 2 − u 2 u 1 ) + p[u 1 u 2 − u 2 u 1 ].
If we use the differential equation L[u 1 ] = L[u 2 ] = 0 to substitute for

the terms pu 1 and pu 2 we see that in fact W p = 0, so that W p is a
constant. We therefore obtain (3.28), and G(x, y) is symmetric.
3.9 Proposition 3.13 and Lemma 3.16 show that the integral operator K
defined by

[K u](x) = k(x, y)u(y) dy

is a compact symmetric mapping from L 2 (

) into L 2 (
). It follows
from Theorem 3.18 that K has a set of eigenfunctions u n (x) with corre-
sponding eigenvalues λn , so that K u n = λn u n :

k(x, y)u n (y) dy = λn u n (x).

Since λ j =
0 for all j there is no nonzero u such that K u = 0. In this
case KerK = {0}, and so we can expand any f ∈ L 2 (
) in terms of the
P1: FNT
Chapter 3 15
eigenfunctions of K ,
∞

f = ( f, u j )u j .
j=1
It is now easy to see that the solution of (3.29) is given by

∞
( f, u j )
u(x) = u j (x),
j=1
λj
as claimed.
3.10 We have
∞
1
A−α w j = t α−1 e−λ j t dt w j . (S3.1)
(α) 0
Now,
∞
(x) = t x−1 e−t dt,
0
and so, substituting u = λ j t in (S3.1), we have

∞
du
λ1−α u α−1 e−u = λ−α
j (α),
0
j
λj
which gives
A−α w j = λ−α
j wj
as required. Since A−α is characterised by its action on the eigenfunctions

the two expressions are equivalent.
3.11 We have
∞

As u2 = λ2s
j |c j |
2
j=1
∞

= λ2(s−α)
j |c j |2ϕ λ2α
j |c j |
2(1−ϕ)
j=1
∞
ϕ ∞
1−ϕ
2(s−α)/ϕ
2α/(1−ϕ)
≤ λj |c j |2 λj |c j |2
j=1 j=1
2ϕ 2(1−ϕ)
≤ A(s−α)/ϕ u Aα/(1−ϕ) u ,
which gives the result on setting ϕ = (k − s)/(k − l) and α =

k(s − l)/(k − l).
P1: FNT
∞
3.12 Take x = j=1 x j w j and consider the series expansion
∞
−λ j h
(e−Ah − I ) (e − 1)
x + Ax = + λj xjwj. (S3.2)
h j=1
h
Observe that
∞
−λ j h ∞
−λ j h
(e − 1) (e − 1)
+ λj xjwj = + 1 λjxjwj.
j=n+1
h j=n+1
λjh
The mean-value theorem tells us that (e−z − 1)/z ≤ 1, and so

2
∞
−λ j h
∞
(e − 1)
+ λj xjwj ≤ 4 λ2j |x j |2 , (S3.3)
h
j=n+1 j=n+1
which tends to zero as n → ∞.

It follows that given > 0 we can choose an n such that the infinite
sum in (S3.3) is bounded above by /2. It is then clear that the finite sum
n
−λ j h
(e − 1)
+ λj xjwj
j=1
h
converges to zero as h → 0, and so for small enough h the whole

expression in (S3.2) is bounded by , as required.
Chapter 4
4.1 Let P = {orthonormal subsets of H }, and define an order on P such that
a ≤ b if a ⊆ b. If {Ci } is a chain (i ∈ I) then C = ∪i Ci is an upper
bound. Zorn’s lemma implies that there is a maximal orthonormal set
{ei }i∈I . The argument of the second part of Proposition 1.23 now shows
that the {ei } form a basis.
4.2 Take z ∈ / Y . Then if w is contained in the linear span of z and Y it has a
unique decomposition of the form
w = y + αz with y ∈ Y,
as in the proof of the Hahn–Banach theorem. We can therefore define a

nonzero linear functional on the linear span of z and Y via
f (y + αz) = α.
The functional f is zero on Y , and we can extend it to a nonzero linear

functional on X by using the Hahn–Banach theorem.
P1: FNT
Chapter 4 17
4.3 It is immediate from Hölder’s inequality that
|L f (g)| ≤ f L ∞ g L 1
and so
L f (L 1 )∗ ≤ f L ∞ . (S4.1)
To show equality consider the sequence of functions
g p (x) = | f (x)| p−2 f (x).
Since f ∈ L ∞ (
) and
is bounded we have g p (x) ∈ L 1 (
) for every
p, with
p−1
g p L 1 = f L p−1 .
It follows from
p
|L f (g p )| = f L p
that
p
f L p
L f (L 1 )∗ ≥ p−1
.
f L p−1
Since f ∈ L ∞ we can use the result of Proposition 1.16,
f L ∞ = lim f L p ,
p→∞
to deduce that
L f (L 1 )∗ ≥ f L ∞ ,
which combined with (S4.1) gives the required equality.

4.4 Since M is a linear subspace of H it is also a Hilbert space. The Riesz
theorem then shows that given a linear functional f on M there exists
an m ∈ M such that
f (x) = (m, x) for all x ∈ M.
Now define F on H by
F(u) = (m, u);
it is clear that F is an extension of F and that F = f .

4.5 If x ∈ / M then the argument of Solution 4.2 shows that there exists an
element f ∈ X ∗ with f | M = 0 but f (x) = 0. So if f (x) = 0 for all
P1: FNT
such f then we must have x ∈ M. Now if xn x then for each f ∈ X ∗
with f | M = 0 we have
f (x) = lim f (xn ) = 0,

n→∞
and so it follows that x ∈ M.

The linear span of the {xn } forms a linear subspace M of X , and clearly
xn ∈ M for each n. It follows that x is contained in the linear span of the
{xn } and so can be written in the form
∞

x= cjxj. (S4.2)
j=1
[In fact x can be written as a convex combination of the {x j }, that is,

(S4.2) with c j ≥ 0 and j c j = 1; see Yosida (1980, p. 120).]
4.6 For any t ∈ [a, b],
δt : x → x(t)
is a bounded linear functional on C 0 ([a, b]). Since xn x, we have
δt (xn ) → δt (x),
and so xn (t) → x(t) for each t ∈ [a, b].

4.7 Write
xn − x2 = xn 2 + x2 − 2(x, xn ),
and then take limits on the right-hand side, using norm convergence on
xn 2 and weak convergence on (x, xn ), to show that
lim xn − x2 = 0,

n→∞
which is xn → x.
Chapter 5
5.1 Simply write
D α u, φn = (−1)|α| u, D α φn ,
and then using the definition of convergence in D(

) (Definition 5.2)
P1: FNT
Chapter 5 19
we have D α φn → D α φ in D(
), and so
D α u, φn → (−1)|α| u, D α φ
= D α u, φ,
so that D α u is indeed a distribution.

5.2 If φn ∈ D(
) with φn → φ in D(
) then ψφn → ψφ in D(
). It
follows that
ψu, φn = u, ψφn → u, ψφn = ψu, φ,
and so ψu ∈ D (
).
Given φ ∈ D(
) we have
D(ψu), φ = −ψu, φ
= −u, ψφ
= −u, ψφ + φψ + u, φψ
= Du, ψφ + u Dψ, φ
= ψ Du + u Dψ, φ,
as claimed.
5.3 Assume that | f n | ≤ M for every n. For every φ ∈ Cc∞ (
) we know that

fn φ d x (S5.1)

is a Cauchy sequence. Since Cc∞ (

)
is dense in L 2 (
) (Corollary 1.14),
for each u ∈ L (
) we can find a sequence of φn ∈ Cc∞ (
) with φn → u
2
in L 2 (
). Then, given > 0, choose K such that
|φk − u| ≤ /4M for all k≥K
and then choose N such that

( f n − f m )φk d x ≤ /2 for all n, m ≥ N .

It follows that for all n, m ≥ N

( f n − f m )u d x ≤ | f n − f m ||u − φk | + /2

≤ (2M)(/4M) + /2 = ,
and so (S5.1) is a Cauchy sequence for every u ∈ L 2 (

), showing that
f n f in L 2 (
).
P1: FNT
5.4 Suppose that the result is true for k = n. We show that it holds for
k = n + 1, which then gives a proof by induction since the statement of
Proposition 5.8 gives (5.45) for k = 1. We know that

u2H n+1 = u2H n + |D α u|2 ,
|α|=n+1
which along with the induction hypothesis becomes

u2H n+1 ≤ C(n) |D α u|2 + |D α u|2 . (S5.2)
|α|=n |α|=n+1
We therefore consider |D α u| for |α| = n. Since u ∈ H0n+1 (

) we must
have D α u ∈ H01 (
), and so
|D α u| ≤ C|D1 D α u| = C|D β u|
with |β| = n + 1, by using (5.11) from the proof of Proposition 5.8. It
follows from (S5.2) that

u2H n+1 ≤ C(n + 1) |D α u|2 ,
|α|=n+1
which is the result for k = n + 1.

5.5 Consider a sequence of u n ∈ C ∞ (
) that approximates u in H k (
).
Then the derivatives of ψu n are given by the Leibniz formula (1.6)
α
α
D (ψu n ) = D β ψ D α−β u n ,
β
β≤α
and so
α
|D α (ψu n )| ≤ |D β ψ||D α−β u n |
β
β≤α

α
≤ |D β ψ| u n H k .
β
β≤α
In this way the derivatives up to and including order k are bounded in

L 2 by a constant (depending on ψ) times the H k norm of u n , and so
ψu n H k ≤ C(ψ)u n H k .
It follows that ψu n is Cauchy in H k (
), and so in the limit as n → ∞
we have ψu ∈ H k (
) with
ψu H k ≤ C(ψ)u H k
as required.
P1: FNT
Chapter 5 21
5.6 First we show that u ∈ L 2 (B(0, 1)):

2 2π 1
1
log log 1+ dx dy = r log log(1+1/r ) dr dθ,
B(0,1) |x| 0 0
which is finite since the integrand is bounded. Now, since
∂u 1 x
=
∂x log(1 + 1/|x|) |x| (1 + |x|)
2
we have

|∇u(x)|2 d x d y
B(0,1)

1 1
= dx dy
B(0,1) log(1 + 1/|x|)2 |x|2 (1 + |x|)2
2π 1
1 1
= dr. (S5.3)
0 0 log(1 + 1/r )2 r (1 + r )2
If we make the substitution u = 1/r this becomes

∞
1 1
du.
1 u + (1/u) log(1 + u)2
This integral is bounded by

∞
1
du,
1 u(log u)2
and since the integrand is the derivative of −1/ log u it follows that the
integral in (S5.3) is finite. Therefore u ∈ H 1 (B(0, 1)), even though it is
unbounded.
5.7 First integrate (5.46) with respect to x1 , so that
∞ ∞ 1/2 ∞ ∞ 1/2
|u(x)| d x1 ≤ 6
3
u D1 u dy1 u D2 u dy2
−∞ −∞ −∞ −∞
∞ 1/2
× u D3 u dy3 d x1
−∞
∞ 1/2 ∞ 1/2
≤6 u D1 u dy1 u D2 u d x1 dy2
−∞ −∞
∞ 1/2
× u D3 u d x1 dy3 .
−∞
P1: FNT
Now integrate with respect to x2 to obtain
∞ ∞ 1/2
|u(x)|3 d x1 d x2 ≤ 6 u D2 u d x1 dy2
−∞ −∞
∞ 1/2 ∞ 1/2
× u D1 u dy1 d x2 u D3 u d x1 d x2 dy3 .
−∞ −∞
Finally, integrating with respect to x3 gives

3
1/2
|u(x)| d x ≤ 6
3
uDju dx ,

j=1
and so
u3L 3 ≤ C|u|3/2 |Du|3/2 ,
which gives
1/2
u L 3 ≤ C|u|1/2 u H 1 ,
as required.
5.8 We simply apply the argument of Theorem 5.29 to the functions v =
D α u for each α with |α| ≤ j. It follows that v ∈ H k− j (
), and since
k − j > m/2 we can use Theorem 5.29 to deduce that v ∈ C 0 (
) with
v∞ ≤ Cu H k− j ≤ Cu H k .
Combining the estimates for each |α| ≤ j shows that u ∈ C j (

) with
uC j (
) ≤ Cu H k (
)
as claimed.
5.9 Suppose that the inequality does not hold. Then for each k ∈ Z+ there
must exist u k ∈ V such that
|u k | ≥ k|∇u k |. (S5.4)
If we set vk = u k /|u k | so that |vk | = 1, (S5.4) becomes
|∇vk | ≤ k −1 . (S5.5)
It follows that vk is a bounded sequence in H 1 (

), and so using
Theorem 5.32 it has a subsequence that converges in L 2 (
) to some
P1: FNT
Chapter 5 23
v ∈ V with

v(x) d x = 0 and |v| = 1. (S5.6)

However, we can also use (S5.5) along with the L 2 convergence of vk to

v to show that for any φ ∈ D(
) and any j

v D α φ d x = lim vk D j φ d x = − lim D j vk φ d x = 0.

k→∞
k→∞
It follows that Dv = 0, and so, using the hint, v is constant almost

everywhere. This contradicts (S5.6), and so we have the inequality (5.47).
5.10 Suppose that {u n } is a bounded sequence in L 2 (
). Then, since L 2 is
reflexive, there is a subsequence that converges weakly in L 2 (
), i.e. for
every φ ∈ L 2 (
) we have
(u n , φ) → (u, φ)
for some u ∈ L 2 (
). Now, suppose that u n does not converge to u in
H −1 (
), so that there exists an > 0 such that, for some subsequence
{u n },
sup |(u n − u, φ)| ≥ .

{φ∈H01 (
):φ H 1 =1}
0
Then there exist φn with φn H01 = 1 such that
|(u n − u, φn )| ≥ /2.
Since {φn } is a bounded sequence in H01 (

) and H01 (
) is compactly
embedded in L 2 (
), there exists a subsequence that is convergent in
L 2 (
) to some φ. It follows (on relabelling) that
|(u n − u, φ)| ≥ /4
for n large enough. But this contradicts the weak convergence of u n to

u in L 2 , and so we must have u n → u in H −1 (
).
5.11 Since

u= ck e2πik·x/L
k∈Zm
P1: FNT
we have
2πik
Du = ck e2πik·x/L .
k∈Zm
L
It follows that

|u|2 = L m |ck |2 and |Du|2 = L m (4π/L)2 |k|2 |ck |2 ,
k∈Zm k∈Zm
and so

L
|u| ≤ |Du|
2π
as claimed.
Chapter 6
6.1 Start with

∇u · ∇v d x = f (x)v(x) d x,

and integrate the left-hand side by parts to give

(u − f )v d x = 0.

Since u ∈ C 2 (
) and f ∈ C 0 (
), we have
ϕ ≡ u − f ∈ C 0 (
).
It therefore suffices to show that if

ϕv d x = 0 for all v ∈ Cc1 (
)

then ϕ = 0. Suppose that ϕ(x) = 0 for some x ∈

. Then since ϕ is
continuous there is a neighbourhood N of x on which ϕ(x) is of constant
sign. Taking a function v that is positive and has compact support within
N implies that

ϕv d x = ϕv d x =
0,

N
a contradiction. That u satisfies u|∂

= 0 follows from u ∈ H01 (
) ∩
C 0 (
), using Theorems 5.35 and 5.36.
P1: FNT
Chapter 6 25
6.2 Take the inner product of Lu = f with a v ∈ Cc1 (

),
m m
∂ ∂u ∂u
− ai j (x) v(x)+ bi (x) v(x)+c(x)u(x)v(x) d x

i, j=1 ∂ x i ∂x j i=1
∂ xi

= f (x)v(x) d x,

and integrate the first term by parts,

m m
∂u ∂v ∂u
ai j (x) + bi (x) v(x) + c(x)u(x)v(x) d x

i, j=1 ∂ xi ∂ x j i=1
∂ xi

= f (x)v(x) d x.

We can now introduce a bilinear form

∂u ∂v
m m
∂u
a(u, v) = ai j (x) + bi (x) v(x) + c(x)u(x)v(x) d x,

i, j=1 ∂ xi ∂ x j i=1 ∂ xi
and write the equation as
a(u, v) = ( f, v) for all v ∈ Cc1 (

).
As before we use the density of Cc1 (

) in H01 (
) to generalise to f ∈
H −1 (
) and the weak form of the problem is thus to find u ∈ H01 (
)
such that
a(u, v) = f, v for all v ∈ H01 (

).
6.3 By definition
m
m

a(u, u) = ai j (x)D j u Di u d x + bi (x)Di u u d x
i, j=1
i=1

+ c(x)u 2 d x

≥θ |∇u| d x − max bi L
2
∞ |∇u| |u| d x

i

− c L ∞ |u|2 d x.

P1: FNT
Now we use Young’s inequality with ,
a2 b2
ab ≤ + ,
2 2
to split the second term,
−1
a(u, u) ≥ 12 θ |∇u|2 d x − θ max bi L ∞ |u|2 d x

i

− c L ∞ |u|2 d x,

and so
a(u, u) ≥ Cu2H1 − λ|u|2 ,
as required.
6.4 Consider the bilinear form b(u, v) corresponding to the operator L + α.
Then
b(u, v) = a(u, v) + α (u, v)

L2
is a continuous bilinear form on H01 : clearly (u, v) is, and a(u, v) is since
m
m

|a(u, v)| ≤ |ai j ||D j u||Di v| d x + |bi ||Di u||v| d x
i, j=1
i=1

+ |c||u||v| d x

≤ Cu H 1 v H 1 .
Furthermore, b satisfies the coercivity condition, since
b(u, u) = a(u, u) + α(u, u)

≥ Cu2H1 − λ|u|2 + α|u|2
≥ Cu2H 1 .
0
We can now apply the Lax–Milgram lemma to obtain the conclusion.

6.5 First, it is easy to see that if (6.31) holds for all v ∈ H 1 (
) then choosing
v = 1 we have

f (x) d x = 0.

P1: FNT
Chapter 6 27
We cannot immediately apply the Lax–Milgram lemma to the equation
a(u, v) = ( f, v),
since
|a(u, u)| = |∇u|2 = u2H 1 − |u|2 ,
and so a(u, v) is not coercive.To deal with the L 2 part we need a Poincaré-
type inequality. Note that if
f (x) d x = 0, then

( f, v) = f, v − v(x) d x ,

since subtracting the constant from v does not make any difference, and
similarly

a(u, v) = a u, v − v(x) d x .

The weak form of the equation in this case [
f (x) d x = 0] is therefore
equivalent to
a(u, v) = ( f, v) for all v ∈ V,
where

V = u ∈ H 1 (
) : u(x) d x = 0 .

It is was shown in Exercise 5.9 that in this space
|u| ≤ C|∇u|,
and so we have
1
|a(u, u)| = |∇u|2 ≥ |u|2 + 12 |∇u|2 ≥ ku2H 1 .
2C
We can therefore apply the Lax–Milgram lemma to deduce the existence
of a weak solution of the Neumann problem.

6.6 Without the imposition of the condition Q u(x) d x = 0 Laplace’s equa-
tion on Q with periodic boundary conditions does not have a unique
P1: FNT
weak solution. In terms of the Lax–Milgram lemma this translates into
the weak problem

a(u, v) = ∇u · ∇v d x = f, v with f ∈ H −1 (Q),
Q
where we seek u ∈ L 2 (Q). But then a is not coercive on L 2 (Q), since

a(c, c) = 0 for any constant c.
6.7 First,
u(x + hei )v(x + hei ) − u(x)v(x)

Dih (uv)(x) =
h

v(x + hei ) − v(x) u(x + hei ) − u(x)
= u(x) + v(x + hei )
h h
= u(x)Dih v(x) + v(x + hei )Dih u(x).
Next, we write

u(x + hei ) − u(x)
v(x) d x

h

u(x + hei ) u(x)
= v(x) d x − v(x) d x

h
h
and change variables in the first integral, putting y = x + hei , to obtain

u(y) u(x)
v(y − hei ) dy − v(x) d x

h
h

v(x − hei ) − v(x)
= − u(x) dx

−h

= − u(x)Di−h v(x) d x.

Finally, both expressions are equal to
Di u(x + he j ) − Di u(x)
.
h
6.8 The inverse of is just the map y → x, given by

yi + z i , i = 1, . . . , m − 1,
xi =
ym + ψ(y1 + z 1 , . . . , ym−1 + z m−1 ), i = m.
P1: FNT
Chapter 7 29
Therefore
 
1 0 0 ... D1 ψ
0 1 0 ... D2 ψ 
 
 D3 ψ 
∇ψ =  0 0 1 . . . .
. . . . .. 
 .. .. .. .. . 
0 0 0 ... 1
It follows immediately that det ∇! = 1, as required.

6.9 The result of Lemma 3.26 shows that, for a general positive symmetric
linear operator whose inverse is compact,
|As u| ≤ C|Al u|(k−s)/(k−l) |Ak u|(s−l)/(k−l)
for 0 ≤ l < s < k. A = − on

with Dirichlet boundary conditions
certainly satisfies these conditions.
Taking u ∈ H0k (
), Proposition 6.19 shows that u ∈ D(A j/2 ) for all
j = 0, 1, . . . , k, and so
u H j (
) ≤ |A j/2 u| ≤ C j u H j (
) .
Therefore we have
(k−s)/(k−l) (s−l)/(k−l)
u H s (
) ≤ Cu H l (
) u H k (
) (S6.1)
for all such u.

Now take u ∈ H k (
), and use Theorem 5.20 to extend u to a function
Eu ∈ H0k (
) for some
⊃
. Then (S6.1) holds for Eu, and since E
is bounded from H j (
) into H0 (
) for each 0 ≤ j ≤ k, we have
j
(k−s)/(k−l) (s−l)/(k−l)
u H s (
) ≤ Eu H s (
) ≤ CEu H l (
) Eu H k (
)
(k−s)/(k−l) (s−l)/(k−l)
≤ Cu H l (
) u H k (
) ,
which is (6.32) for u ∈ H k (

), as required.
Chapter 7
∗∗
7.1 Define an element I ∈ X by
T
I, L = L , f (t) dt for all L ∈ X ∗. (S7.1)
0
P1: FNT
This map I is clearly linear, and it is bounded since
T
|I, L| ≤ L X ∗ f (t) X dt
0
T
≤ f (t) X dt L X ∗ ,
0
and
T
f (t) X dt < ∞
0
from (7.31). Since X is reflexive, it follows that there exists an element

y ∈ X such that
I, L = L , y for all L ∈ X ∗.
Therefore, using (S7.1), we have (7.29).

That the integral is well defined follows from Lemma 4.4, which shows
that if
L , y1 = L , y2 for all L ∈ X∗
then y1 = y2 .
7.2 Corollary 4.5 shows that there exists an element L ∈ X ∗ such that
Lop = 1 and L y = y X . Then, using (7.29), we have
T T

f (t) dt ≤ |L , f (t)| dt

0 X 0
T
≤ f (t) X dt,
0
as required.
7.3 An element v of L p (0, T ; V ) is the limit in the L p norm of a sequence
of functions vn in C 0 ([0, T ]; V ). Since such functions are uniformly
continuous on [0, T ], given > 0 we can find an integer N such that
δ = T /N satisfies
|t − s| < δ ⇒ vn (t) − vn (s)V ≤ /T 1/ p .
We can approximate vn to within in L p (0, T ; V ) by

N
vn ( jδ)χ [( jδ, ( j + 1)δ)],
j=1
P1: FNT
Chapter 7 31
an expression of the form (7.32). It follows that such elements are dense
in L p (0, T ; V ). Since C 1 ([0, T ]) is dense in L p (0, T ) we could also use
elements of the form of (7.32) with α j ∈ C 1 ([0, T ]); similarly, Cc∞ (
)
is dense in L p (
), so we could take v j ∈ Cc∞ (
).
7.4 Taking the inner product of (7.33) with Ak u n yields
d k/2 2 k+1 2 k−1 k+1
1
2
|A u n | + A 2 u n ≤ A 2 f A 2 u n ,
dt
and so, using Young’s inequality, we obtain
d k/2 2 k+1 2 k−1 2
|A u n | + A 2 u n ≤ A 2 f ,
dt
which shows that
t k+1 2
|Ak/2 u n (t)|2 + A 2 u n (s)2 ds ≤ |Ak/2 u(0)|2 + A k−1
2 f ,
which yields (7.34), and then (7.35) follows from (7.33). Therefore, using
Proposition 6.18, we get
u n ∈ L ∞ (0, T ; H k ) ∩ L 2 (0, T ; H k+1 )
and
du n /dt ∈ L 2 (0, T ; H k−1 ).
Extracting a subsequence shows that the solution u satisfies
u ∈ L 2 (0, T ; H k+1 ) and du/dt ∈ L 2 (0, T ; H k−1 ).
It follows from Corollary 7.3 that u ∈ C 0 ([0, T ]; H k ).

7.5 Since the {w j } are orthogonal in H01 and orthonormal in L 2 the equation
for u n becomes
λ j un j = f j ,
where λ j ≡ w j 2 and f j = ( f, w j ). It follows that
u n j = f j /λ j ,
independent of n. In particular we have

n
fj
un = wj,
λ
j=1 j
P1: FNT
and so, for m > n,

m
f j2
u m − u n 2 ≤ .
j = n+1
λj
Since we have the Poincaré inequality we must have λ j ≥ λ, for some

λ, and so
1 2
m
u m − u n ≤
2
f .
λ j = n+1 j
Since f ∈ L 2 (
) it follows that u n converges in H01 (
) to u =
∞
j=1 f j w j /λ j .
Now, we know that
((u n , v)) = (Pn f, v) for all v ∈ Pn H01 (

).
Since
((u n , v)) = ((u n , Pn v)) and (Pn f, v) = (Pn f, Pn v)
for all v ∈ H01 (

), we in fact have
((u n , v)) = (Pn f, v) for all v ∈ H01 (

).
Since u n → u in H01 (
) we know that
((u n , v)) → ((u, v)),
and since Pn f → f in L 2 (
) we must have
((u, v)) = ( f, v) for all v ∈ H01 (

),
and u is a weak solution of (7.36) as required.
Chapter 8
8.1 We show in general that if Z = X ∩ Y , with norm
u Z = u X + uY ,

P1: FNT
Chapter 8 33
then Z ∗ = X ∗ + Y ∗ . First, it is clear that if f = f 1 + f 2 , with f 1 ∈ X ∗
and f 2 ∈ Y ∗ , then for u ∈ X ∩ Y
| f 1 + f 2 , u| ≤ | f 1 , u| + | f 2 , u|

≤ f 1 X ∗ u X + f 2 Y ∗ uY
≤ ( f 1 X ∗ + f 2 Y ∗ )u X ∩Y .
Thus X ∗ + Y ∗ ⊂ (X ∩ Y )∗ . Now, if f ∈ (X ∩ Y )∗ then, since it is a linear

functional on a linear subspace of X , application of the Hahn–Banach
theorem (Theorem 4.3) tells us it has an extension f 1 that is a linear
functional on the whole of X (we could use Y rather than X here if we
wished). Thus (X ∩ Y )∗ ⊂ X ∗ ⊂ X ∗ + Y ∗ , and so we have the required
equality.
8.2 Follow the argument of Theorem 7.2, except approximate u by a se-
quence u n ∈ C 1 ([0, T ]; H 1 ) such that
un → in L 2 (0, T ; H 1 ) ∩ L p (
T )
and
du n /dt → du/dt in L 2 (0, T ; H −1 ) + L q (

T ).
We will denote by X (t1 , t2 ) the space
L 2 (t1 , t2 ; H 1 ) ∩ L p (
× (t1 , t2 )),
and by X ∗ (t1 , t2 ) the space
L 2 (t1 , t2 ; H −1 ) + L q (
× (t1 , t2 )).
We now estimate
T t
1
|u n (t)|2 d x = |u n (t)|2 dt d x + 2 u̇ n (s)u n (s) ds

T
0
t∗
T
1
≤ |u n (t)|2 dt d x + 2u̇ n X ∗ (t ∗ ,t) u n X (t ∗ ,t)
T
0
T
1
≤ |u n (t)|2 dt d x + 2u̇ n X ∗ (0,T ) u n X (0,T ) ,
T
0
showing once again that u n is also a Cauchy sequence in C 0 ([0, T ]; L 2 )

and hence that u ∈ C 0 ([0, T ]; L 2 ) as claimed.
P1: FNT
8.3 Integrating by parts gives

∂ 2un
− f (u n ) 2 d x

j ∂x j

∂u n 2
= f (u n ) dx + f (u n )∇u n · n d S.

j ∂x j ∂
We can estimate the extra term by

f (u n )∇u n · n d S ≤ | f (0)| |∇u n | d S
∂
∂
≤ | f (0)||∂
|1/2 ∇u n L 2 (∂
)
≤ Cu n H 1 (∂
)
≤ Cu n H 2 (
) ,
using the trace theorem (Theorem 5.35). Since we have
u H 2 (
) ≤ C|Au|
from Theorem 6.16, we can write

d
1
2
u n 2 + |Au n |2 ≤ lu n 2 + C|Au n |.
dt
Using Young’s inequality on the last term and rearranging finally gives
d
u n 2 + |Au n |2 ≤ 2lu n 2 + C,
dt
which integrates to give the bound
T T
u n (T ) +
2
|Au n (s)| ds ≤ 2l
2
u n (t)2 dt + u 0 2 + C T.
0 0
Thus u n is uniformly bounded in L 2 (0, T ; D(A)) [and L ∞ (0, T ; V )],

where (8.19) is used as before to guarantee that u n ∈ L 2 (0, T ; V ).
8.4 In this case we can follow the proof of Proposition 8.6 until the line

|F(u) − F(v)|2 ≤ C|u − v|2L 2 p 1 + |u|2L 2qγ + |v|2L 2qγ .
We now have to be more careful with our use of the Sobolev embedding
theorem, since the highest we can go is H 1 ⊂ L 6 . We therefore need
2p ≤ 6 and 2qγ ≤ 6, where ( p, q) are conjugate.
The first conditions forces us to take p ≤ 3, and hence we must have

P1: FNT
Chapter 8 35
q ≥ 3/2, which shows that the largest possible value for γ is 2, as
claimed. Provided that γ ≤ 2 we can write
|F(u) − F(v)|2 ≤ Cu − v2H 1 (1 + u H 1 + v H 1 ),
as in Proposition 8.6.
8.5 Since u, v ∈ L 2 (0, T ; D(A) we can use Corollary 7.3 to take the inner
product of
dw
+ Aw = F(u) − F(v)
dt
with Aw to obtain, using (8.31),
d
1
2
w2 + |Aw|2 = (F(u) − F(v), Aw)
dt
≤ C(1 + |Au| + |Av|)1/2 w1/2 |Aw|3/2 .
We now use Young’s inequality to split the right-hand side,

d 3
1
2
w2 + |Aw|2 ≤ |Aw|2 + C(1 + |Au| + |Av|)2 w2 ,
dt 4
and so
d
1
2
w2 ≤ C(1 + |Au| + |Av|)2 w2 .
dt
This yields
t
w(t)2 ≤ w(0)2 exp C(1 + |Au(s)|2 + |Av(s)|2 ) ds ,
0
which gives continuous dependence on initial conditions since we know

that both u and v are elements of L 2 (0, T ; D(A)).
8.6 Setting g(s) = e−A(t−s) u(s), we have
∂g du
= Ae−A(t−s) u(s) + e−A(t−s)
∂s ds
= Ae−A(t−s) u(s) + e−A(t−s) [−Au + f (u(s))]
= e−A(t−s) f (u(s)),
so that integrating with respect to s between 0 and t gives

t
g(t) − g(0) = e−A(t−s) f (u(s)) ds.
0
This rearranges to give (8.33).

P1: FNT
Chapter 9
9.1 Taking the divergence of the governing equation yields
u = ∇ · f,
since all the other terms are divergence free. A solution of this equation in
the periodic case when f ∈ L̇ 2 (Q) has been obtained as Equation (9.10).
Note that if f ∈ H then this implies that p = 0 (or, equivalently, a
constant).
9.2 We have
1/2 1/2
|u|4L 4 = |u|4 d x ≤ |u|6 |u|2
Q Q Q
= u3L 6 |u|
≤ ku3 |u|,
since H 1 (Q) ⊂ L 6 (Q) (see Theorem 5.31).

9.3 Applying the Cauchy–Schwarz inequality first in the variable j and then
in the variable i, we get
m 1/2 m 1/2
m

ai bi, j c j ≤ |ai bi, j |2 |c j |2

i, j=1 i, j=1 j=1
m 1/2
m

m

= ai bi, j |c j |2

i=1 j=1 j=1
1/2 1/2 1/2

m
m
m
≤ |ai |2 |bi, j |2 |c j |2 ,
i=1 i, j=1 j=1
as claimed.
9.4 If m = 2, we have
|b(u, v, w)| ≤ k|u|1/2 u1/2 |v|1/2 v1/2 w
[using b(u, v, w) = −b(u, w, v)], so that
B(u, u), w ≤ k|u|uw,
and therefore
B(u, u)V ∗ ≤ k|u|u.

P1: FNT
Chapter 9 37
If m = 3, we have
|b(u, v, w)| ≤ k|u|1/4 u3/4 |v|1/4 v3/4 w
[using b(u, v, w) = −b(u, w, v) again], and so
B(u, u), w ≤ k|u|1/2 u3/2 w,
giving
B(u, u)V ∗ ≤ k|u|1/2 u3/2 .
9.5 Take ( p, q) = (2, 2) if m = 2 and ( p, q) = (4/3, 4) if m = 3. We know

∗
that Bn B in L p (0, T ; V ∗ ), where Bn = B(u n , u n ) and B = B(u, u).
∗
We need to show that Pn Bn B in the same sense. For ψ ∈ L q (0, T ; V )
we have
T T T
Pn Bn (t)− B, ψ dt = Pn Bn − Bn , ψ dt + Bn − B, ψ dt.
0 0 0
∗
The second term converges since Bn B, so we have to treat only the
first term. We rewrite this as
T
Bn (t), Q n ψ dt.
0
Since functions of the form

k
ψ= ψ j α j (t), ψ j ∈ V, α j ∈ C 1 ([0, T ], R) (S9.1)
j=1
are dense in L q (0, T ; V ) (see Exercise 7.3) we can consider

$ %
T
k
Bn , Q n ψ j α j (t) dt.
0 j=1
Since Bn is uniformly bounded in L p (0, T ; V ∗ ) when m = 2, we can

use the fact that Q n ψ j → ψ j in V to show the required convergence
for all ψ of the form (S9.1). The density of such ψ in L q (0, T ; V ) then
gives the full result.
P1: FNT
9.6 If u ∈ L 4 (0, T ; V ) then we can estimate b(w, u, w) in (9.41) differently,

writing
d
1
2
|w|2 + νw2 ≤ k|w|1/2 w3/2 u
dt
ν c
≤ w2 + 3 |w|2 u4 ,
2 ν
which becomes, dropping the terms in w2 ,
d
|w|2 ≤ C|w|2 u4 .
dt
Integrating gives
t
|w(t)| ≤ |w(0)| exp
2 2
u(s) ds ,
4
0
which implies uniqueness provided that u ∈ L 4 (0, T ; V ).
Chapter 10
10.1 If not, then there exist an > 0 and sequences δn → 0, xn ∈ K ,
yn ∈ H , such that
|xn − yn | ≤ δn and | f (xn ) − f (yn )| > .
Since K is compact there is a subsequence of the {xn } (relabel this xn )

such that xn → x ∗ ∈ X . Now,
|x ∗ − yn | ≤ |x ∗ − xn | + |xn − yn | → 0 as n → ∞, (S10.1)
and
| f (x ∗ )− f (yn )| ≥ | f (xn )− f (yn )|−| f (xn )− f (x ∗ )|≥/2 (S10.2)
if n is sufficiently large, since f is continuous at x ∗ . But then (S10.1)

and (S10.2) say precisely that f is not continuous at x ∗ , which is a
contradiction.
10.2 The set in (10.23) is bounded since

S(t)B, (S10.3)
t≥t0 (B)
P1: FNT
Chapter 10 39
with t0 (B) from Definition 10.2, is a subset of B, and

S(t)B (S10.4)
0≤t≤t0 (B)
is bounded since B is bounded and S(t) is continuous. Similarly, if

B is compact then (S10.3) is a closed subset of B, and (S10.4) is the
continuous image of the compact set B × [0, t0 (B)]: both parts are
compact, and therefore so is (10.23). That (10.23) is positively invariant
is clear by definition.
10.3 In this example ω(0) = 0 and ω(x) = {|x| = 1} if x = 0. So
$(B) = {(0, 0)} ∪ {|x| = 1}
(which is clearly not connected). Since ω(x) = (1, 0) for all x with
|x| = 1,
$[$(B)] = {(0, 0), (1, 0)},
so that $[$(B)] = $(B) as claimed.

10.4 We show that, for a bounded set X ,
ω1 (X ) = {y : S(tn )xn → y},
where tn → ∞ and xn ∈ X , is equal to

&
ω2 (X ) = S(s)X .
t≥0 s≥t
If y ∈ ω1 (X ) then clearly

y∈ S(s)X
s≥t
for all t ≥ 0 and hence in y ∈ ω2 (X ). So ω1 (X ) ⊂ ω2 (X ).

Conversely, if y ∈ ω2 (X ) then for any t ≥ 0

y∈ S(s)X ,
s≥t
and so there are sequences {τm(t) }, with τm(t) ≥ t, and {xm(t) } ∈ X with
S(τm(t) )xm(t) → y. Now consider t = 1, 2, . . . and pick tn from τm(n) and
P1: FNT
xn from xm(n) such that

(n) (n)
S τ x − y ≤ 1/n.
m m
Then S(tn )xn → y with tn → ∞, since tn ≥ n, showing that y ∈

ω1 (X ). This gives ω2 (X ) ⊂ ω1 (X ), and so ω1 (X ) = ω2 (X ).
10.5 If y ∈ S(t)B for all t ≥ 0 then for any tn there is an xn ∈ B with
y = S(tn )xn , so clearly y ∈ ω2 (B) (as defined in the previous solution).
Conversely, if y ∈ ω2 (B) then we must have

y∈ S(s)B.
s≥t
Now, if τ ≥ t0 (B), then
S(t)B ⊃ S(t + τ )B,
and so then

S(t)B ⊃ S(t + τ )B.
τ ≥t0 (B)
Since S(t)B is closed

S(t)B ⊃ S(t + τ )B y,
τ ≥t0 (B)
that is, y ∈ S(t)B for all t ≥ 0.

Clearly we have
& &
S(t)B ⊂ S(nT )B.
t≥0 n∈Z+
If u ∈ S(nT )B for all n ∈ Z+ then in particular u ∈ S(n 0 T )B, provided

that n 0 is large enough that n 0 T ≥ t0 (B), where
S(t)B ⊂ B for all t ≥ t0 (B).
Since u ∈ S(nT )B we have u = S(nT )y with y ∈ B, and it follows

that for all t ≥ 0
S(t)u = S(t + nT )y = S(τ )y ∈ B,
since τ ≥ t0 (B). Therefore u ∈ S(t)B for all t ≥ n 0 T . It follows that

u ∈ ω2 (B), giving the required equality.
P1: FNT
Chapter 10 41
10.6 y ∈ ω(Y ) if S(tn )yn → y with tn → ∞ and yn ∈ Y . Then yn ∈ X also
and so ω(X ) ⊃ ω(Y ). If Y absorbs X in a time t0 (assuming X to be
bounded) and if S(tn )xn → x, then
S(tn − t0 )[S(t0 )xn ] → x,
and since tn − t0 → ∞ and S(t0 )xn ∈ Y , ω(X ) ⊂ ω(Y ), so then

ω(X ) = ω(Y ).
10.7 First, the set
∞

Ki (S10.5)
i= j
is clearly closed, and since all sets K i lie within 1/j of K j if i ≥ j it is

also bounded, and hence compact. It follows that K ∞ , the intersection
of a decreasing sequence of compact sets, is itself compact.
Now, it is clear by a similar argument that
dist(K ∞ , K j ) ≤ j −1 .
Conversely, if u ∈ K j then dist(u, K i ) ≤ j −1 for all i ≥ j. So certainly

∞

dist u, K i ≤ j −1 .
i= j
In particular, there exist points u i ∈ K i , i ≥ j, such that
|u i − u| ≤ j −1 .
Since each u i is contained in the compact set (S10.5) (with j = 1)

then there exists a subsequence of the u i that converges to some u ∗ . It
follows that u ∗ ∈ K ∞ , and by construction |u − u ∗ | ≤ j −1 . Therefore
dist(K j , K ∞ ) ≤ j −1 ,
and so
distH (K j , K ∞ ) ≤ j −1 :
K j converges to K ∞ in the Hausdorff metric.

10.8 To show that the inverse is continuous, suppose not. Then there exist
an > 0 and a sequence {xn } ∈ f (X ) with xn → y ∈ f (X ) but
| f −1 (xn ) − f −1 (y)| ≥ . However, f −1 (xn ) ∈ X , and since X is
P1: FNT
compact there exists a subsequence xn j such that f −1 (xn j ) → z. Since

f is continuous, it follows that xn j → f (z). Since f is injective, it
follows from f (z) = y that z = f −1 (y), which is a contradiction. So
f −1 is continuous on f (X ).
10.9 Proposition 10.14 says that, given 1 and T > 0, there exists a time τ1
such that, for all t ≥ τ1 ,
dist(u(t), A) ≤ δ(1 , T ).
So we can track the trajectory u(t) within a distance 1 for a time T

starting at any time t ≥ τ1 .
We can replace T with 2T and apply the same argument for 2 =
1 /2, that is, there exists a time τ2 such that, for all t ≥ τ2 ,
dist(u(t), A) ≤ δ(2 , 2T ),
and then the trajectory u(t) can be tracked for a time 2T starting at any
time t ≥ τ2 .
Thus u(t) can be followed from τ1 to τ2 by a distance 1 with a finite
number of trajectories on A of time length T , and when we reach τ2 ,
we can start to track u(t) within a distance 2 with trajectories on A of
time length 2T , until we reach a τ3 after which we can track within a
distance 3 for a time length 3T , etc.
The “jumps” are bounded by k + k+1 , since
|vk+1 − S(tk+1 − tk )vk |

≤ |vk+1 − u(tk+1 )| + |u(tk + (tk+1 − tk )) − S(tk+1 − tk )vk |
≤ k+1 + k .
10.10 Take > 0. Then there is a T > 0 such that
dist(S(t)B1 , A) + dist(S(t)B2 , A) < for all t ≥ T.
Also, by the uniform continuity of the semigroup, there is a δ > 0 such

that
dist(S(t)B1 , S(t)B2 ) ≤ for all t ∈ [0, T ]
provided that dist(B1 , B2 ) ≤ δ. The argument is symmetric, which

gives the result.
P1: FNT
Chapter 11 43
Chapter 11
11.1 Using Young’s inequality on (11.30) we can deduce that

p p
|u| ≤
2
|u| p d x + |
|.
2
p−2
So we can write (11.6) as

d 2α2 2 2
1
|u|2 + u2 + |u| ≤ + k |
|.
2
dt p p−2
Neglecting the u2 term we can write

d 2α2 2 2
1
|u|2 + |u| ≤ + k |
|.
2
dt p p−2
We can now apply the Gronwall inequality to deduce an asymptotic

bound on |u(t)|, as in Proposition 11.1. (The expression for the bound
will be a more complicated expression than before.)
11.2 Proceeding as advised, we obtain
s s
d
y(s) exp − g(τ ) dτ ≤ h(s) exp − g(τ ) dτ ≤ h(s),
ds t t
and integrating both sides between s and t + r gives

t+r
y(t + r ) ≤ y(s) exp g(τ ) dτ
s
t+r t+r
+ h(τ ) dτ exp g(τ ) dτ
s s
≤ (y(s) + a2 ) exp(a1 ).
Integrating both sides for t ≤ s ≤ t + r gives the result as stated.

11.3 Taking the inner product of
du n
+ Au n = Pn f (u n )
dt
with t 2 Au n we obtain

du n 2
, t Au n + t 2 |Au n |2 = (Pn f (u n ), t 2 Au n ),
dt
P1: FNT
which, using the methods leading to (8.27) for the right-hand side,
becomes
d
1
2
tu n 2 − 2tu n 2 + t 2 |Au n |2 ≤ lt 2 u n 2 .
dt
Integrating from 0 to T gives
T T
T u n 2 + t 2 |Au n |2 dt ≤ (2t + lt 2 )u n 2 dt.
0 0
Since we already know that u n ∈ L 2 (0, T ; V ), it follows that u n ∈

L 2 (t, T ; D(A)) for any t > 0. Since H 2 (
) ⊂ C 0 (
) if m ≤ 3 we
also have Pn f (u n ) ∈ L 2 (t, T ; L 2 ), and so it follows that du n /dt ∈
L 2 (t, T ; H ). Taking limits shows that the solution u satisfies
u ∈ L 2 (t, T ; D(A)) and du/dt ∈ L 2 (t, T ; L 2 ).
Application of Corollary 7.3 then makes the “formal” calculations at

the beginning of Section 11.1.2 rigorous.
11.4 Observe that for s < 0 we have
f (s)|s| ≥ α2 |s| p − k,
and so in particular
f (s) ≥ 0 for all s < (k/α2 )1/ p . (S11.1)
Now set M = (k/α2 )1/ p , multiply Equation (11.1) by (u(x) + M)− ,

and integrate to obtain

1 d
2
(u(x) + M)− +2
|∇(u + M)− | =
2
f (u)(u + M)− d x
dt

≤ 0,
using (S11.1). It follows, using the Poincaré inequality, that

1 d
2
(u(x) + M) 2
− d x ≤ −C (u(x) + M)2− d x,
dt

and so as in the last part of the argument given in Theorem 11.6, we

must have

(u(x) + M)2− d x = 0

for all u ∈ A.
P1: FNT
Chapter 12 45
Chapter 12
12.1 If u is smooth then Au = −u, and we have
2
b(u, u, A u) =
2
u i (Di u j )Dk2 Dl2 u j d x
i, j,k,l=1

2

= Dk2 (u i (Di u j )) Dl2 u j d x
i, j,k,l=1
'
= Dk2 u i (Di u j ) + 2(Dk u i )(Dk Di u j )
i, j,k,l
(
+ u i Di Dk2 u j
Dl2 u j

= b(Au, u, Au) + 2 (Dk u i )(Dk Di u j ) Dl2 u j d x
i, j,k,l
+ b(u, Au, Au)

2
= b(Au, u, Au) + 2 b(Dk u, Dk u, Au),
k=1
as claimed. The result follows for general u by taking limits.

To obtain inequality (12.23), use (9.26) to give

2
|b(u, u, A2 u)| ≤ k|A3/2 u||Au|u + 2 |b(D j u, Au, D j u)|
j=1

2
≤ k|A 3/2
u||Au|u + 2k |D j u|D j uAu
j=1
≤ k|A3/2 u||Au|u
2 1/2 2 1/2

+ 2k |D j u|2 D j u2 Au.
j=1 j=1
Since
u2 = a(u, u) = Au, u = (A1/2 u, A1/2 u) = |A1/2 u|2 ,
this becomes
|b(u, u, A2 u)| ≤ 3k|A3/2 u||Au|u
ν 3/2 2 9k 2
≤ |A u| + u2 |Au|2 ,
4 ν
as required.
P1: FNT
12.2 Take the inner product of
du/dt + ν Au + B(u, u) = f
with A2 u to obtain
d
1
2
|Au|2 + ν|A3/2 u|2 = −b(u, u, A2 u) + ( f, A2 u),
dt
and use the estimate (12.23) from the previous exercise to write
d
1
2
|Au|2 + ν|A3/2 u|2
dt
≤ |b(u, u, A2 u)| + f |A3/2 u|
ν C f 2 ν
≤ |A3/2 u|2 + u2 |Au|2 + + |A3/2 u|2 ,
4 ν ν 4
so that
d 2 f 2 C
|Au|2 + ν|A3/2 u|2 ≤ + u2 |Au|2 .
dt ν ν
Using a similar trick as we did for the absorbing set in V , we integrate
this equation between s and t, with t < s < t + 1, so that

2M C t+1
|Au(t + 1)|2 ≤ |Au(s)|2 + + u(s)2 |Au(s)|2 ds,
ν ν t
where we have used (12.24). Integrating again with respect to s between
t and t + 1 gives
t+1
2M C t+1
|Au(t + 1)|2 ≤ |Au(s)|2 ds + + u(s)2 |Au(s)|2 ds.
t ν ν t
(S12.1)
Now, if t ≥ t1 (|u 0 |) then we know that
t+1
u(s) ≤ ρV and |Au(s)|2 ds ≤ I A ,
t
and so if it follows that then

2M C
|Au(t + 1)|2 ≤ ρ A ≡ I A + + ρV2 I A ,
ν ν
an absorbing set in D(A).
12.3 Suppose that u n ∈ V with u n ≤ k and that u n → u in H . Then there
exists a subsequence u n j such that u n j v in V , so that v ≤ k.
Since V ⊂⊂ H , it follows that u n j → v in H , and so in particular we
must have u = v, which implies that u ≤ k.
P1: FNT
Chapter 12 47
12.4 If u ∈ D(A) with

u= u k e2πik·x/L
k∈Z2
then we can estimate u∞ by

u∞ ≤ |u k |.
k∈Z2
Split the sum into two parts,

u∞ ≤ |u k | + |u k |.
|k|≤κ |k|>κ
We now use the Cauchy–Schwarz inequality on each piece,

u∞ ≤ (|u k | × 1) + (|u k ||k|2 × |k|−2 )
|k|≤κ |k|>κ
1/2 1/2

≤ |u k | 2
1
|k|≤κ |k|≤κ
1/2 1/2

−4
+ |u k | |k| 2 4
|k| .
|k|>κ |k|>κ
Since

1 ≤ Cκ 2 and |k|4 ≤ Cκ −2 ,
|k|≤κ |k|>κ
this becomes
u∞ ≤ C(κ|u| + κ −1 |Au|).
To make both terms on the right-hand side the same, we choose κ =

|Au|1/2 |u|−1/2 , obtaining
u∞ ≤ C|u|1/2 |Au|1/2 .
12.5 We have already derived in (12.20) the inequality
d 2| f |2
u2 + ν|Au|2 ≤ + Cu6 ,
dt ν
and since we have a uniform bound on u for t large enough, we obtain
a uniform bound on the integral of |Au(s)|2 ,
t0 +1
|Au(s)|2 ds ≤ C1 . (S12.2)
t0
P1: FNT
Following the analysis in Proposition 12.4, we estimate
|u t | ≤ ν|Au| + |B(u, u)| + | f |,
and using (12.25) this becomes
|u t | ≤ ν|Au| + ku3/2 |Au|1/2 + | f |.
An application of Young’s inequality yields
|u t | ≤ c|Au| + Cu2 + | f |,
and so for t large enough,
|u t | ≤ c|Au| + CρV2 + | f |.

The bound in (S12.2) therefore implies a bound on |u t |2 ,
t0 +1
|u t (s)|2 ds ≤ C2 . (S12.3)
t0
Now differentiate
u t + ν Au + B(u, u) = f
with respect to t to obtain
u tt + ν Au t + B(u t , u) + B(u, u t ) = 0
and take the inner product with u t so that

d
1
2
|u t |2 + νu t 2 ≤ |b(u t , u, u t )|
dt
≤ ku|u t |1/2 u t 3/2
3ν k 4 u4 |u t |2
≤ u t 2 + .
4 4ν 3
Using once again the asymptotic bound on u, we have for t ≥ t0 that
d
|u t |2 ≤ C3 |u t |2 .
dt
We use the usual trick, integrating between s and t + 1, with t < s <
t + 1,
t+1
|u t (t + 1)| ≤ |u t (s)| + C4
2 2
|u t (s)|2 ds,
t
P1: FNT
Chapter 13 49
and then between t and t + 1 (with respect to s) so that
t+1
|u t (t + 1)|2 ≤ (1 + C4 ) |u t |2 ds
t
≤ (1 + C4 )C3 , (S12.4)
by (S12.3).
To end, we show that |u t | bounds |Au|. From the equation we have
ν|Au| ≤ |u t | + |B(u, u)| + | f |,
or with (12.25)
ν|Au| ≤ |u t | + k|Au|1/2 u3/2 + | f |,
and so after using Young’s inequality and rearranging we have
|Au| ≤ C(|u t | + u3 + | f |).
Together with (S12.4) we obtain
|Au(t)| ≤ ρ D
for all t ≥ 1 + t0 (u 0 ). So we have an absorbing set in D(A) and

hence a global attractor for the 3D equations.
Chapter 13
13.1 Let G(X, ) be the number of boxes in a fixed cubic lattice, with sides
, that are necessary to cover X . Since each cube with side sits inside
a ball of radius , N (X, ) ≤ G(X, ), and so
d f (X ) ≤ dbox (X ).
Also, since any ball with side is contained within at most 2m different
boxes in the grid, we have G(X, ) ≤ 2m N (X, ). Therefore
log G(X, )
dbox (X ) = lim sup
→0 − log
m log 2 + log N (X, )
≤ lim sup
→0 − log
log N (X, )
= lim sup
→0 − log
= d f (X ),
P1: FNT
giving equality between box-counting dimension and fractal dimension
in Rm .
13.2 If n+1 ≤ < n then we have
log N (X, ) log N (X, n+1 )
≤
− log − log n
log N (X, n+1 )
≤
− log n+1 + log(n+1 /n )
log N (X, n+1 )
≤ ,
− log n+1 + log α
and so
log N (X, ) log N (X, n )
lim sup ≤ lim sup .
→0 − log n→∞ − log n
That this inequality holds in the opposite sense is straightforward, and
hence we obtain the desired
√ equality.
13.3 The sequence m = ( 2 log m)−1 , m ≥ 2, satisfies
m+1 log m log 2
= ≥ ,
m log(m + 1) log 3
and so we can use the result of the previous exercise. Note that we have

en ek 2 1 1 2

log n − log k = (log n)2 + (log k)2 ≤ (log n)2
for n > k, and so the first m − 1 elements from Hlog will belong to
distinct balls of radius m . It follows that
N (Hlog ) ≥ m − 1,
and so
log N (Hlog , m )
d f (Hlog ) ≥ lim sup
m→∞ log m
log(m − 1)
≥ lim sup √ = ∞,
m→∞ log( 2 log m)
which implies that d f (Hlog ) = ∞, as claimed.
13.4 At the jth stage of construction the middle-α set Cα consists of 2 j
intervals of length β j , where β = (1 − α)/2. It follows that
N (Cα , β j ) = 2 j .
P1: FNT
Chapter 13 51
Therefore, using the result of Exercise 13.2 we can calculate
log 2 j log 2
d f (Cα ) = lim sup = .
j→∞ log β j log β
13.5 Clearly
∞
∞

µ X k , d, ≤ µ(X k , d, ).
k=1 k=1
Since µ(X k , d, ) is nondecreasing in we have
µ(X k , d, ) ≤ Hd (X k )
for each k, and so for every > 0 we have

∞ ∞

µ X k , d, ≤ Hd (X k ).
k=1 k=1
We can now take the limit as → 0 on the left-hand side to obtain

∞ ∞

H d
Xk ≤ Hd (X k )
k=1 k=1
as claimed.
13.6 The map L taking e(i) into v (i) (1 ≤ i ≤ n) is given by

n
L= v (k) (e(k) )T ,
k=1
( j)
and since ei(k) = δik , the components of L are L i j = vi :
( j)
(L T L)i j = vk(i) vk = v (i) · v ( j) = Mi j .
13.7 M is real and symmetric since
Mi j = δx (i) · δx ( j) .
It follows that its eigenvalues λ j are real, and one can find an orthonor-
mal set of eigenvectors e(k) with
Me(k) = λk e(k) .
P1: FNT
To show that λk > 0, consider
λk = e(k) Me(k) = ei(k) δxs(i) δxs( j) e(k)

T (k) 2
j = |v | ≥ 0,
where v (k) is the vector given by its components
vs(k) = ei(k) δxs(i) .
If v (k) = 0 then the two different initial conditions

n
δx(0) = 0 and δx(0) = ei(k) δx (i)
i=1
have the same solution at time t, contradicting uniqueness. So all the

eigenvalues are strictly positive.
13.8 Writing M as

n
M= λ j e j e Tj ,
j=1
we have

n
log M = log λ j e j e Tj .
j=1
Clearly

n
Tr[log M] = log λ j , (S13.1)
j=1
and since

n
detM = λj
j=1
the required result follows immediately.

Since Tr[log M] is given by (S13.1), we have
d λ̇i n
Tr[log M] = .
dt λ
i=1 i
P1: FNT
Chapter 13 53
The right-hand side of (13.34) is
n
−1 d M
ei , M ei
i=1
dt
) n *
n
n

= ei , λ−1 T
j ejej λ̇k ek ekT + λk ėk ekT + λk ek ėkT ei
i=1 j=1 k=1
) *

n
n

= λi−1 eiT λ̇k ek ekT + λk ėk ekT + λk ek ėkT ei
i=1 k=1
) *

n
n
= λi−1 λ̇i eiT + λi ėiT + λk (ei , ėk )ekT ei
i=1 k=1

n
' −1 (
= λi λ̇i + (ėi , ei ) + (ei , ėi )
i=1

n
= λi−1 λ̇i ,
i=1
since dtd (ei , ei ) = 0.

13.9 Since the eigenvalues are proportional to the sums of squares of m
integers, we will have reached the eigenvalue mk 2 once we have taken
k m combinations of integers. Thus
λk m = Cmk 2 ,
and so if k m < n < (k + 1)m we obtain
Cmk 2 ≤ λn ≤ Cm(k + 1)m .
We now have
k < n 1/m < (k + 1)
and so
1 1/m
2
n < k < k + 1 < 2n 1/m .
This gives
cn 2/m ≤ λn ≤ Cn 2/m ,
as required.
P1: FNT
13.10 Taking the inner product of (13.27) with U we obtain
d
1
2
|U |2 + νU 2 = −b(U, u, U ),
dt
and so
d
1
2
|U |2 + νU 2 ≤ k|U |U |u|.
dt
Using Young’s inequality and rearranging we get
d
|U |2 + νU 2 ≤ C|U |2 . (S13.2)
dt
That bounded sets in L 2 are mapped into bounded sets in L 2 follows
by neglecting the term in U 2 and applying Gronwall’s inequality
(Lemma 2.8),
|U (t)|2 ≤ eCt |U (0)|2 = eCt |ξ |2 . (S13.3)
To show that we in fact obtain a bounded set in H 1 , we first return

to (S13.2) and integrate between t/2 and t to obtain
t t
ν U (s) ds ≤ C
2
|U (s)|2 ds + |U (t/2)|2 ≤ C(t)|U (t/2)|2 ,
t/2 t/2
(S13.4)
using (S13.3). Now we take the inner product of (13.27) with AU ,
which gives
d
1
2
U 2 + ν|AU |2 = −b(u, U, AU ) − b(U, u, AU ).
dt
Using (9.27) we obtain
d
1
2
U 2 + ν|AU |2 ≤ k |u|1/2 u1/2 U 1/2 |AU |3/2
dt

+ |U |1/2 U 1/2 u1/2 |Au|1/2 |AU | ,
and after using Young’s inequality and rearranging we have

d
U 2 + ν|AU |2 ≤ CU 2 .
dt
Expression (S13.4) allows us to use the “uniform Gronwall” trick and
find a bound on U valid for all t > 0. Thus $(t; u 0 ) is compact for
all t > 0.
P1: FNT
Chapter 14 55
13.11 If we integrate (12.6) between 0 and T we obtain
T
T | f |2
ν |Au(s)|2 ds ≤ + u(0)2 .
0 ν
Dividing by T and taking the limit as T → ∞ yields

1 T | f |2
lim sup |Au(s)|2 ds ≤ 2 ,
T →∞ T 0 ν
since there is an absorbing set in V . Therefore
| f |2
χ≤ = ν 3 L −6 G 2 .
L 2ν
The only length that can be formed from χ and ν is
3 1/6
ν
Lχ = ,
χ
and this implies (13.35).
Chapter 14
14.1 Since A is compact, it is bounded and certainly contained in B(0, r )
for some r > 0. So Nr (A) = 1. We consider
S(B(0, r ) ∩ A),
which by our assumption can be covered by K 0 balls, centred in A, and

of radius r/2. So
N (A, r/2) = K 0 .
Now consider each one of the balls in this covering, and apply our
assumption again to show that
S(B(ai , r/2) ∩ A)
can be covered by K 0 balls of radius r/4, so that
N (A, r/4) = K 02 .
Iterating this argument, we can see that
N (A, 2−k r ) = K 0k .
P1: FNT
So therefore, using the result of Exercise 13.2, we have
log(N (A, 2−k r ))

d f (A) = lim
k→∞ log(2k )
k log K 0
≤
k log 2
log α
≤ n0 ,
log 2
precisely (14.33).
14.2 We have, for any u ∈ D(A1/2 ),
u2 = a(u, u) = (A1/2 u, A1/2 u) = |A1/2 u|.
Expanding p in terms of the eigenfunctions of A gives

n
p= ( p, w j )w j ,
j=1
and so

n
p2 = λ j |( p, w j )|2 ≤ λn | p|2 .
j=1
Similarly,

n
q2 = λ j |(q, w j )|2 ≥ λn+1 |q|2 .
j=1
The other two inequalities in the exercise follow easily from these.
14.3 Differentiating gives

d da λb db λa
= exp(λa/C(a +b)) 1− + 1+ .
dt dt C(a+b) dt C(a+b)
Since we have (14.34), the coefficient of da/dt is negative, whereas

the coefficient of db/dt is positive. It follows that we can substitute in
the inequalities for da/dt and db/dt, which gives d /dt ≤ 0.
14.4 Write

|u(x) − u(y)| ≤ |e2πik·x/L − e2πik·y/L ||ck |,
k∈Z2
P1: FNT
Chapter 14 57
and use (14.35) to deduce that

|u(x) − u(y)| ≤ C|x − y|1/2 |ck ||k|1/2
k∈Z2
1/2 1/2
|k|
≤ C|x − y| 1/2
(1 + |k| )|ck |
4 2
(1 + |k|4 )
k∈Z2 k∈Z2
≤ Cu H 2 |x − y| 1/2
.

[ k∈Z2 |k|/(1 + |k|4 ) is finite.]
14.5 Since (6.14) shows that u H 2 = C|Au| for u ∈ D(A), we can use the
result of the previous exercise to deduce that
|u(x) − u(y)| ≤ c|Au||x − y|1/2 .
Expression (14.36) follows immediately from this and the definitions

of d(N ) and η(u).
14.6 Choose > 0. Then there exists a T such that b(t) ≤ /2 for all t ≥ T .
Hence for t ≥ T ,
dX
+ a X ≤ /2.
dt
By Gronwall’s inequality (Lemma 2.8),
X (T + t) ≤ X (T )e−at + /2,
and so choosing τ large enough that
ke−aτ < /2,
we have
X (t) ≤ for all t ≥ T + τ,
so that X (t) → 0.
14.7 Using the bound on b given in (9.25), we can write
d
1
2
w2 + ν|Aw|2 ≤ w∞ w|Au|
dt
≤ [η(w) + cd(N )1/2 |Aw|]w|Au|
−1/2
≤ η(w)w|Au| + cd(N )1/2 λ1 |Aw|2 |Au|,
P1: FNT
using (14.36), and therefore
d ' −1/2 (
1
2
w2 + ν − cλ1 d(N )1/2 |Au| λ1 w2 ≤ η(w)w|Au|.
dt
Now, we know that A is bounded in V and D(A), so that
d
1
2
w2 + [ν − cλ1/2 ρ A d(N )1/2 ]λ1 w2 ≤ 2ρV ρ A η(w).
dt
Now, choose δ such that
1/2
µ = ν − cλ1 ρ A δ 1/2 > 0.
Then we have, for d(N ) < δ,
d
1
2
w2 + µw2 ≤ 2ρV ρ A η(w). (S14.1)
dt
By assumption, we know that η(w) → 0, and since the attractor is
bounded in V we have w(t)2 ≤ 4ρV2 . The result of the previous
exercise applied to (S14.1) now shows (14.37).
14.8 (i) Take the inner product of (14.38) with qn = Q n u to obtain
d
1
2
|qn |2 + (Au, qn ) = (F(u), qn ).
dt
Now, notice that
|(Au, qn )| = |(Aqn , qn )| ≥ λn+1 |qn |2 ,
and so
d
1
2
|qn |2 + λn+1 |qn |2 ≤ C0 |qn |,
dt
from which, using the result of Exercise 2.5, we see that
d
|qn | ≤ −λn+1 |qn | + C0 ,
dt +
which gives
C0
|Q n u(t)| ≤ + |Q n u(0)|, (S14.2)
λn+1
using the Gronwall lemma (Lemma 2.8).

P1: FNT
Chapter 15 59
(ii) Writing p(t) = Pn u(t) and q(t) = Q n u(t), p solves the equation
d p/dt + Ap = Pn F( p + q).
Thus the equation for w = p − pn is
dw/dt + Aw = Pn F( pn ) − Pn F( p + q).
Taking the inner product with w and using the Lipschitz property
of F gives
d
1
2
|w|2 + w2 ≤ C1 |w|2 + C1 |q||w|.
dt
Hence
d
|w| ≤ C1 |w| + C1 |q|,
dt +
and so, using the bound in (S14.2) and the Gronwall lemma as
above we obtain

−1 C0
|Pn u(t) − pn (t)| ≤ C1 + |Q n u(0)| eC1 t .
λn+1
Combining this with (S14.2) yields

C0
|u(t) − pn (t)| ≤ C1−1 + |Q n u(0)| (C1 + eC1 t ),
λn+1
and since we know that λn+1 → ∞ and |Q n u(0)| → 0 as n →
∞, it follows that pn (t) converges to u(t) as claimed.
Chapter 15
15.1 For any point v ∈ H ,

dist(v, M)2 = inf |Pv − p|2 + |Qv − φ( p)|2
p∈P H
and
|Qv − φ(Pv)|2 = |Qv − φ( p) + φ( p) − φ(Pv)|2
≤ 2|Qv − φ( p)|2 + 2|φ( p) − φ(Pv)|2
≤ 2|Qv − φ( p)|2 + 2l 2 |Pv − p|2

≤ c2 |Qv − φ( p)|2 + |Pv − p|2
P1: FNT
for all p ∈ P H , where c2 = 2 max(l 2 , 1). Therefore
|Qv − φ(Pv)| ≤ c dist(v, M).
The other implication is obvious.

15.2 Using Proposition 15.3 we see that the attractor lies in the graph of
some Lipschitz function : Pn H → Q n H . We can therefore project
the dynamics on A onto Pn H by writing
d p/dt + Ap = Pn F( p + ( p)). (S15.1)
It is easy to show that (S15.1) is a Lipschitz ODE on Pn H , since
|Pn F( p+ ( p))−Pn F( p+ ( p))| ≤ |F( p + ( p))−F( p + ( p))|

≤ C| p + ( p) − p − ( p)|

≤ C | p − p| + | ( p) − ( p)|
≤ 2C| p − p|.
We know that if u(t) is a solution in A then p(t) = Pn u(t) is a so-

lution of (S15.1) lying in Pn A. Since (S15.1) is Lipschitz its solu-
tions are unique, and so in particular Pn A is an invariant set. Thus
(S15.1) is a finite-dimensional system that reproduces the dynamics
on A. [The advantage of the inertial form over (S15.1) is that Pn A
is the attractor of the finite-dimensional system, not just an invariant
set.]
15.3 Since F = 0 outside B(0, ρ),
't0 ⊂ S(t0 ){u : u ∈ Pn H : ρ ≤ |u| ≤ ρeλn+1 t0 }.
The cone invariance part of the strong squeezing property then shows
that for any two points u 1 and u 2 in 't0 we must have
|Q n (u 1 − u 2 )| ≤ |Pn (u 1 − u 2 )|.
If we write

'= S(t)
0≤t<∞
then the function defined by
(Pn u) = Q n u for all u∈'

P1: FNT
Chapter 15 61
is Lipschitz on its domain of definition, Pn ' = Pn B(0, ρ). Clearly '

is positively invariant, and so M is invariant.
To show that A ⊂ M, suppose that u ∈ A and v ∈ M with
Pn u = Pn v but Q n u = Q n v. Then, using the invariance of ' and A,
we have u = S(t)u t with u t ∈ M, and v = S(t)vt with vt ∈ A. Thus
|Q n (u − v)| ≤ |Q n (u t − vt )|e−kt
≤ 2ρe−kt , (S15.2)
since both A and ' are subsets of B(0, ρ). Since (S15.2) holds for all
t ≥ 0, we must have Q n u = Q n v. Thus u = v and A ⊂ M as claimed.
15.4 We have
135 = 12 + 22 + 32 + 112 ,
136 = 62 + 102 ,
137 = 42 + 112 ,
138 = 12 + 32 + 82 + 82 ,
all as sums of (at most) four squares.

15.5 (i) If u(t) is a solution of (15.24), then p(t) = Pn u(t) is the solution
of the equation
d p/dt + Ap = Pn F( p(t) + q(t)).
Since F is Lipschitz, it follows that
|Pn F( p(t)+q(t))−Pn F( p(t)+ ( p(t)))| ≤ C1 |q(t)− ( p(t))|

≤ C1 Ce−kt ,
where the result of Exercise 15.1 has been used.

(ii) Let u(t) = p(t) + ( p(t)). Then u(t) ∈ M, so we just have to
show the exponential convergence in (15.26). To do this, we write
|u(t) − u(t)| ≤ | p(t) + q(t) − p(t) − ( p(t))|

+ | p(t) + ( p(t)) − p(t) − ( p(t))|
≤ |q(t) − ( p(t))| + 2| p(t) − p(t)|
−kt
≤ Ce + 2De−kt = Me−K t ,
where we have used the result of Exercise 15.1 again and the
Lipschitz property of .
P1: FNT
Chapter 16
16.1 ω(r ) is clearly well defined, since the set
{(x, y) ∈ X × X : |x − y| ≤ r }
is a compact subset of X × X . The convexity property follows easily,

since
ω(r + s) = sup | f (x) − f (z)|

|x−z|≤r +s
≤ sup | f (x) − f (y)| + | f (y) − f (z)|

|x−y|≤r, |y−z|≤s
≤ sup | f (x) − f (y)| + sup | f (x) − f (y)|

|x−y|≤r |y−z|≤s
= ω(r ) + ω(s),
where to prevent too clumsy notation we have assumed throughout that

x, y, z ∈ X .
16.2 (i) X can be covered by N (X, ) balls of radius and, in particular,
lies within of the space spanned by the centres of these balls.
Therefore d(X, ) ≤ N (X, ), and the inequality follows.
(ii) Simply choose any open subset O in Rn . Then d f (O) = n but
since O ⊂ Rn we must have τ (O) = 0.
16.3 Consider the projection Pn onto the space spanned by the first n eigen-
functions of A,

n
Pn u = (u, w j )w j ,
j=1
and its orthogonal complement Q n = I − Pn . Then
|u − Pn u| = |Q n u|
= |Q n A−s/2 As/2 u|
≤ Q n A−s/2 op |As/2 u|
−s/2
≤ λn+1 u H s
≤ Cn −2s/m
for some constant C. Clearly,

log log C
log d(X, ) ≤ + ,
−2s/m 2s/m
P1: FNT
Chapter 16 63
and so one obtain (16.23). If X is bounded in D(Ar ) for any r then it
follows from (16.23) that τ (X ) = 0, and so one can obtain any θ in the
range
2d f (X )
0<θ <1− .
k
We can now obtain any θ < 1 by choosing k large enough.
16.4 Write w = u − v for u, v ∈ A. If A is Lipschitz continuous from A
into H then
|Aw| ≤ L|w|
for some L. Now split w = Pn w + Q n w, and observe that we have

both
|Aw|2 = |A(Pn w+Q n w)|2 = |A(Pn w)|2 +|A(Q n w)|2 ≥ λ2n+1 |Q n w|2
and
|Aw|2 ≤ L 2 |w|2 ≤ L 2 |Pn w|2 + L 2 |Q n w|2 .
Since λn → ∞ as n → ∞, we can choose n large enough that λn+1 >

L, and then write
λ2n+1 − L 2 |Q n w|2 ≤ L 2 |Pn w|2 ,
that is,
1/2
L2
|Q n w| ≤ |Pn w|.
λ2n+1 − L 2
It follows that we can define (Pn u) = Q n u uniquely for each u ∈ A,

and then
1/2
L2
| ( p1 ) − ( p2 )| ≤ | p1 − p2 |,
λ2n+1 − L 2
so that (cf. Proposition 15.3) the attractor is a subset of a Lipschitz

graph over Pn H .
16.5 Since X is the attractor for ẋ = g(x), given > 0, there exists a δ > 0
such that if x(0) ∈ N (X, δ) then the solution x(t) of ẋ = g(x) remains
within N (X, ) for all t ≥ 0.
P1: FNT
Define f˜(x) on a closed subset of Rn by

f (x), dist(x, X ) ≤ δ/4,
f˜(x) =
0, dist(x, X ) ≥ δ/2.
Since f˜ is Lipschitz on its domain of definition, it can be extended

using Theorem 16.4 to a function F(x) that is Lipschitz on Rn . Now
consider
ẋ = F(x) + g(x). (S16.1)
Clearly X is an invariant subset for (S16.1), since F(x) + g(x) = F(x)

on X . To show that the attractor of (S16.1) lies within an N (X, ) it
suffices to show that N (X, ) is absorbing. This follows from the choice
of δ and the fact that F(x) + g(x) = g(x) outside N (X, δ/2).
Chapter 17
17.1 Integrating (17.3) between 0 and L and using the periodic boundary
conditions gives
L L
|Du|2 = − u D 2 u d x,
0 0
which implies (17.4) after an application of the Cauchy–Schwarz in-

equality. For u ∈ Ḣ 2p the result follows by finding a sequence {u n } ∈ Ċ 2p
that converges to u in the norm of Hp2 .
17.2 Multiplying (17.5) by a function φ in Ċ 2p and integrating by parts twice
gives
L L
(D 2 u)(D 2 φ) d x = f (x)φ(x) d x. (S17.1)
0 0
Define a bilinear form a(u, v) : Ḣ 2p × Ḣ 2p → R by
L
a(u, v) = (D 2 u)(D 2 v) d x,
0
and then, using the density of Ċ 2p in Ḣ 2p , we see that (S17.1) becomes

(17.6).
P1: FNT
Chapter 17 65
17.3 Since a(u, v) is equivalent to the inner product on Ḣ 2p (by the general
Poincaré inequality from Exercise 5.4), we can use the Riesz represen-
tation theorem to deduce the existence of a unique solution u ∈ Ḣ 2p of
(17.6) for any f ∈ H −2 .
In particular if f ∈ L̇ 2 then u ∈ Ḣ 2p , which is a compact subset of L̇ 2 ,
using the Rellich–Kondrachov compactness theorem (Theorem 5.32).
It follows that the inverse of A is compact, and A itself is clearly sym-
metric. We can therefore apply Corollary 3.26 to deduce that A has an
orthonormal set of eigenfunctions {w j } that form a basis for L̇ 2 .
17.4 The orthogonality property (17.8) follows easily, since for u ∈ Ċ 2p ,
L L
2 du d
b(u, u, u) = u(x) dx = 1
3
u(x)3 d x = 0,
0 dx 0 dx
using the periodic boundary conditions. The result follows for all u ∈
Ḣ 2p by taking limits. Similarly for the cyclic equality, after an integration
by parts, we have
L L L
uvx w d x = − (uw)x v d x = − vwx u + wu x v d x.
0 0 0
The inequalities in (17.10) follow from the estimate

uvw d x ≤ u∞ |v||w| ≤ |Du||v||w|,
since H 1 ⊂ C 0 on a one-dimensional domain (Theorem 5.31).

17.5 Taking the inner product of (17.12) with u n gives
d
1
2
|u n |2 + a(u n , u n ) + (D 2 u n , u n ) + (Pn B(u n , u n ), u n ) = 0.
dt
Since
(Pn B(u n , u n ), u n ) = (B(u n , u n ), Pn u n ) = (B(u n , u n ), u n ) = 0
by (17.8), we obtain
d
1
2
|u n |2 + |D 2 u n |2 = |Du n |2 .
dt
Using (17.4) we have
d
1
2
|u n |2 + |D 2 u n |2 ≤ |u n ||D 2 u n |
dt
≤ 12 |u n |2 + 12 |D 2 u n |2 ,
P1: FNT
and so
d
|u n |2 + |D 2 u n |2 ≤ |u n |2 . (S17.2)
dt
Dropping the term in |D 2 u n |2 and integrating we get
|u n (t)|2 ≤ et |u n (0)|2 ,
so clearly
un is uniformly bounded in L ∞ (0, T ; L̇ 2 ).
Integrating (S17.2) as it stands then gives

t t
|u n (t)| +
2
|D u n (s)| ds ≤
2 2
|u n (s)|2 ds + |u n (0)|
0 0
and in particular shows that

un is uniformly bounded in L 2 0, T ; Ḣ 2p .
It follows from these estimates, the equality
du n /dt = −Au − D 2 u − B(u, u),
and Poincaré’s inequality (17.2) that
du n /dt is uniformly bounded in L 2 (0, T ; H −2 ),
and we have obtained the bounds in (17.13).

17.6 Extracting subsequences from the {u n } and relabelling as necessary we
find a u such that

u ∈ L 2 0, T ; Ḣ 2p ∩ L ∞ (0, T ; L̇ 2 ) with du/dt ∈ L 2 (0, T ; H −2 ),
and

un u in L 2 0, T ; Ḣ 2p ,
∗
un u in L ∞ (0, T ; L̇ 2 ),
∗
du n /dt du/dt in L 2 (0, T ; H −2 ).
We can also use the compactness theorem (Theorem 8.1) to find a

subsequence with the additional strong convergence

un → u in L 2 0, T ; Ḣ 1p ,
P1: FNT
Chapter 17 67
since Ḣ 2p ⊂⊂ Ḣ 1p ⊂ H −2 . It is simple to show the weak-* convergence

in L 2 (0, T ; H −2 ) of all the terms in the equation, except for the non-
linear term. For this we need the strong convergence in L 2 (0, T ; Ḣ 1p )
and the uniform bound on u n in L ∞ (0, T ; L̇ 2 ). We need to show that

T T
b(u n , u n , v) dt → b(u, u, v) dt for all v ∈ L 2 0, T ; Ḣ 2p .
0 0
Using (17.9) we write
b(u n , u n , v) − b(u, u, v) = b(u n −u, u n , v) + b(u, u n − u, v)

= −b(u n , v, u n − u) − b(v, u n − u, u n ) + b(u, u n − u, v),
and then for the first term

T T
|b(u n , v, u n − u)| dt ≤ k |u n ||D 2 v||u n − u| dt
0 0
≤ ku n L ∞ (0,T ; L̇ 2 ) v L 2 (0,T ; Ḣ 2p ) u n − u L 2 (0,T ; L̇ 2 )
→ 0,
and for the second and third terms

T T
|b(v, u n − u, u n )| dt ≤ k |Dv||D(u n − u)||u n | dt
0 0
≤ ku n L ∞ (0,T ; L̇ 2 ) v L 2 (0,T ; Ḣ 2p ) u n − u L 2 (0,T ; Ḣ 1p )
→ 0,
giving the required convergence. That

∗
Pn B(u n , u n ) B(u, u)
follows as in Exercise 9.5.

Finally, the continuity of u into L̇ 2 follows from the generalisation
of Theorem 7.2 discussed after its formal statement in Chapter 7.
17.7 The equation for the difference w of two solutions, w = u − v, is
wt + wx x x x + wx x + wu x + vwx = 0.
Taking the inner product with w we obtain
d
1
2
|w|2 + |D 2 w|2 − |Dw|2 = −b(w, u, w) − b(v, w, w).
dt
P1: FNT
Estimating the terms on the right-hand side by using (17.10) we have
d
1
2
|w|2 + |D 2 w|2 ≤ |Dw|2 + |D 2 u||w|2 + |v||Dw|2 .
dt
Using (17.4) and Young’s inequality gives
d
1
2
|w|2 + |D 2 w|2 ≤ (1 + |v|)|w||D 2 w| + |D 2 u||w 2 |
dt
≤ 12 |D 2 w|2 + C(1 + |D 2 u| + |v|2 )|w|2 ,
and so
d
|w|2 + |D 2 w|2 ≤ C(1 + |D 2 u| + |v|2 )|w|2 . (S17.3)
dt
Neglecting the term in |D 2 w|2 and integrating from 0 to t shows
(17.15). Since u, v ∈ L 2 (0, T ; Ḣ 2p ), it follows that w(t) = 0 for all
t if w(0) = 0, which gives uniqueness.
17.8 Choosing α = 6 we have
d 2 1 2 2
|v| + 2 |D v| + 2|v|2 ≤ 12 |g|2 , (S17.4)
dt
and so in particular
d 2
|v| ≤ −2|v|2 + 12 |g|2 .
dt
The Gronwall inequality (Lemma 2.8) now shows that
|v(t)|2 ≤ |v(0)|2 e−2t + 14 |g|2 (1 − e−2t ). (S17.5)
Since u = φ + v and φ ∈ Ċ ∞ p is constant, it follows that there is an
absorbing set for u(t) in L 2 .
We can also obtain from (S17.4) a bound on the integral of |D 2 v|2 ,
t+1
1
2
|D 2 v(s)|2 ds ≤ 12 |g|2 + |v(t)|2 ,
t
or for |D u| the bound

2 2
t+1
|D 2 u(s)|2 ds ≤ |g|2 + |D 2 φ|2 + |v(t)|2 .
t
It follows from (S17.5) that if t is large enough then

t+1
|D 2 u(s)|2 ds ≤ M, (S17.6)
t
and we have both bounds in (17.18).

P1: FNT
Chapter 17 69
17.9 Taking the inner product of (17.11) with −D 2 u we obtain
d
1
2
|Du|2 + |D 3 u|2 = |D 2 u|2 + b(u, u, D 2 u).
dt
We now we estimate the right-hand side by using (17.10),
d
1
2
|Du|2 + |D 3 u|2 ≤ |D 2 u|2 + |D 2 u||Du|2 .
dt
Neglecting the term in |D 3 u|2 we have
d
|Du|2 ≤ |D 2 u|2 + |D 2 u||Du|2 .
dt
Note that this is in the form in which the uniform Gronwall lemma of
Exercise 11.2 is applicable, since we have a uniform estimate on the
integral of |D 2 u| provided in (S17.6) above. It follows that there is an
1
absorbing set in Ḣ p .
We have therefore obtained a compact absorbing set in L 2 and proved
the existence of a global attractor.
17.10 As in the proof of Theorem 13.20, we consider the equation for θ =
u − v − U,
θt + θx x x x + θx x + θ u x + wwx = 0,
where w = u − v. Taking the inner product with θ yields
d 2
1
2
|θ| + |D 2 θ |2 = |Dθ|2 − b(θ, u, θ ) − b(w, w, θ ).
dt
Using (17.4) and (17.10) on the right-hand side we obtain
d 2
1
2
|θ | + |D 2 θ|2 ≤ |θ ||D 2 θ | + |θ |2 |D 2 u| + |Dw|2 |θ |
dt
≤ 12 |θ|2 + 12 |D 2 θ|2 + |D 2 u||θ |2 + 12 |Dw|4 + 12 |θ |2 ,
and so
d 2
|θ| + |D 2 θ|2 ≤ 2(1 + |D 2 u|)|θ |2 + |Dw|4 .
dt
It follows from Gronwall’s inequality (Lemma 2.8), since θ(0) = 0, that
t
|θ (t)| ≤ k(t)
2
|Dw(s)|4 ds,
0
P1: FNT
and so, using (17.4), we get
t
|θ (t)|2 ≤ k |w(s)|2 |D 2 w(s)|2 ds.
0
Returning to (S17.3),
d
|w|2 + |D 2 w|2 ≤ C(1 + |D 2 u| + |v|2 )|w|2 ,
dt
multiplying both sides by |w|2 , and integrating we obtain
t t
|w(s)|2 |D 2 w(s)|2 ds ≤ C |w(s)|4 ds + 14 |w(0)|4 .
0 0
Using (17.15) we have

t
|w(s)|2 |D 2 w(s)|2 ds ≤ C(t)|w(0)|4 ,
0
and hence
|θ(t)|2 ≤ K (t)|w(0)|4 .
The uniform differentiability property now follows.

17.11 To show that $(t; u 0 ) is compact take the inner product of (17.19) with
U to obtain
d
1
2
|U |2 + |D 2 U |2 − |DU |2 + b(U, u, U ) + b(u, U, U ) = 0.
dt
Using the cyclic property (17.9) and the bound in (17.10) we have
d
1
2
|U |2 + |D 2 U |2 ≤ C|DU |2 .
dt
Using (17.4) and Young’s inequality we end up with
d
|U |2 + |D 2 U |2 ≤ C|U |2 . (S17.7)
dt
Dropping the term in |D 2 U |2 shows that
|U (t)|2 ≤ eCt |ξ |2 , (S17.8)
and integrating between t/2 and t shows that (cf. Exercise 13.10)
t
|D 2 U (s)|2 ds ≤ C(t)|U (t/2)|2 . (S17.9)
t/2
P1: FNT
Chapter 17 71
Now take the inner product of (17.19) with −D 2 U and obtain
d
1
2
|DU |2 + |D 3 U |2 = |D 2 U |2 + b(U, u, D 2 U ) + b(u, U, D 2 U )
dt
≤ |D 2 U |2 + |DU ||Du||D 2 U | + |u||D 2 U |2 ,
by using (17.10). We can use the Poincaré inequality (17.2) and drop
the term in |D 3 U |2 to give
d
|DU |2 ≤ C|D 2 U |2 .
dt
Using (S17.9) and the uniform Gronwall “trick” shows that a bounded
set in L 2 becomes a bounded set in H 1 , and so $(t; u 0 ) is compact for
all t > 0 as claimed.
17.12 We use (17.4) to estimate the second term on the right-hand side by
1/2 1/2

n
n
n
n
|Dφ j | ≤
2
|φ j ||D φ j | ≤
2
|φ j |
2
|D φ j |
2 2
.
j=1 j=1 j=1 j=1
Since the {φ j } are orthonormal, |φ j |2 = 1, giving

1/2

n
n
n
|Dφ j | ≤ n
2 1/2
|D φ j |
2 2
≤n+ 1
4
|D 2 φ j |2 .
j=1 j=1 j=1
To estimate the final term, we use the Cauchy–Schwarz inequality,

L
φ 2j Du d x ≤ |φ 2j ||Du| = φ j 2L 4 |Du|
0
≤ C|Dφ j |2 ,
since |Du| is bounded on A and H 1 ⊂ L 4 . Now, using (17.4), we have

L
φ 2j Du d x ≤ C|φ j ||D 2 φ j |
0
≤ C|φ j |2 + 14 |D 2 φ j |2 .
Combining these estimates we have

n
n
(Lφ j , φ j ) ≤ − 12 |D 2 φ j |2 + Mn.
j=1 j=1
P1: FNT
Since the eigenvalues λ j of A = D 4 are proportional to j 4 , it follows

(cf. final part of the argument in the proof of Lemma 13.17) that

n
|D 2 φ j |2 ≥ Cn 5 .
j=1
Therefore we need
−Cn 5 + Mn < 0,
which occurs provided that n > (M/C)1/4 . The KSE therefore has a
finite-dimensional attractor.
17.13 For v ∈ D(A1/2 ) we have
L
(N (u), v) = u(Du)v + (D 2 u)v d x
0
L
=− 1 2
2
u Dv − u D 2 v d x,
0
and so
|(N (u), v)| ≤ 12 |u|2 Dv L ∞ + |u||D 2 v|.
Since H 1 ⊂ L ∞ and D(A1/2 ) ⊂ H 2 then
|(N (u), v)| ≤ C(|u| + 1)|u||A1/2 v|,
as required.
17.14 For w ∈ D(A1/2 ),
L
(N (u) − N (v), w) = (u Du − v Dv)w + D 2 (u − v)w d x
0
L
= (u
1 2
2
− v 2 )Dw + (u − v)(D 2 w) d x,
0
and so
|(N (u) − N (v), w)| ≤ ( 12 (|u + v| + 1)|u|Dw L ∞

≤ c(|u + v| + 1)|u||A1/2 w|,
where the same embedding results as those given above were used.

Solutions To Exercises:, Y) ) Is Dense in X y /2 /2

Uploaded by

Copyright:

Available Formats

Solutions To Exercises:, Y) ) Is Dense in X y /2 /2

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Solutions To Exercises:, Y) ) Is Dense in X y /2 /2

Uploaded by

Copyright:

Available Formats

P1: FNT

CB341/Robinson CB341-Sol February 5, 2001 10:52 Char Count= 0

x − x j  X < /2 and y − yk Y < /2,

(x j , yk ) − (x, y) X ×Y ≤ .

It follows that X × Y is separable, and by induction it follows that any

Since X is compact it follows that there exists a finite covering by such

It follows that for each x ∈ X there exists an x j with |x − x j | <  as

is the appropriate completion of Cc0 (Rm ). For any u ∈ C00 (Rm ), we

as j → ∞. Since { f j } is Cauchy it certainly follows that for j, k ≥ N

for each l < ∞, and taking the limit as k → ∞ gives

Using the triangle inequality in each C n (

Since (S1.1) holds for all l, we can let l → ∞ to show that

and so f j → f in the  · c norm.

|[ f j (x) − f k (x)] − [ f j (y) − f k (y)]|

|[ f j (x) − f (x)] − [ f j (y) − f (y)]| ≤ |x − y|γ

| f (x) − f (y)| ≤ | f j (x) − f j (y)| + |[ f j (x) − f (x)] + [ f j (y) − f (y)]|

which shows that f ∈ C 0,γ (

by using (1.7) so that u h is also Hölder.

 f k f k+1  L p ≤  f k  L pk  f k+1  L pk+1 ,

which combined with (S1.2) gives (1.31) for n = k + 1. Since the

Now note that

and so using Hölder’s inequality we have

It clearly suffices to approximate χ [I ] to within  in the L p norm using

almost everywhere, and so

This implies that

In particular, we have x (n) (n)

It follows that x (n) → x in l ∞ , and so l ∞ is complete.

x + λy2 = (x + λy, x + λy) = x2 + 2λ(x, y) + λ2 y2 ≥ 0.

In particular, the quadratic equation for λ,

λ2 y2 + 2λ(x, y) + x2 = 0,

4(x, y)2 − 4y2 x2 ≤ 0

|(x, y)| ≤ xy,

which is the Cauchy–Schwarz inequality. We can now write

x + y2 = x2 + 2(x, y) + y2

giving the triangle inequality.

u + v2 + u − v2 = u2 + 2(u, v) + v2 + u2 − 2(u, v) + v2

h(h n (x ∗ )) = h n+1 (x ∗ ) = h n (h(x ∗ )) = h(x ∗ ),

showing that h(x ∗ ) is also a fixed point of h n . Since the contraction

|h(x) − h(y)| = |x − y|(1 − (x y)−1 )

but clearly has no fixed point.

h(x) − h(y) < x − y, (S2.1)

suppose that h has no fixed point. Then

h(x) − x > 0 for all x ∈ X,

and since h(x) − x is continuous from X into R it obtains its lower

h(x) − x ≥  for all x ∈ X,

and there exists some y ∈ X such that h(y) − y = . However, if we

a contradiction. So h has at least one fixed point. Uniqueness follows as

2.4 Suppose that there are solutions xn (t) of

d x/dt = f (x) with x(0) = x0 (S2.2)

sup |xn (t)| ≤ |x0 | + τ  f ∞ and |xn (t) − xn (s)| ≤  f ∞ |t − s|,

the conditions of the Arzelà–Ascoli theorem (Theorem 2.5). It follows

|x(t + h)| ≤ [C(t0 ) + ]t,

Integrating both sides between 0 and T gives us

We can integrate the right-hand side to obtain

and so, using (2.28), we have

We now use Lemma 2.9 to deduce that

An application of Gronwall’s inequality [(2.21) in Lemma 2.8] now

Thus I is a bounded operator on both spaces.

x − x j X < /2 and y − yk Y < /2,

(x j , yk ) − (x, y) X ×Y ≤ .

It follows that for each x ∈ X there exists an x j with |x − x j | < as

and so f j → f in the · c norm.

|[ f j (x) − f (x)] − [ f j (y) − f (y)]| ≤ |x − y|γ

f k f k+1 L p ≤ f k L pk f k+1 L pk+1 ,

It clearly suffices to approximate χ [I ] to within in the L p norm using

x + λy2 = (x + λy, x + λy) = x2 + 2λ(x, y) + λ2 y2 ≥ 0.

λ2 y2 + 2λ(x, y) + x2 = 0,

4(x, y)2 − 4y2 x2 ≤ 0

|(x, y)| ≤ xy,

x + y2 = x2 + 2(x, y) + y2

u + v2 + u − v2 = u2 + 2(u, v) + v2 + u2 − 2(u, v) + v2

h(x) − h(y) < x − y, (S2.1)

h(x) − x > 0 for all x ∈ X,

and since h(x) − x is continuous from X into R it obtains its lower

h(x) − x ≥ for all x ∈ X,

and there exists some y ∈ X such that h(y) − y = . However, if we

sup |xn (t)| ≤ |x0 | + τ f ∞ and |xn (t) − xn (s)| ≤ f ∞ |t − s|,

|x(t + h)| ≤ [C(t0 ) + ]t,

sup Pn x X < ∞

so that x X = 0. For y ∈ R(A) we can use the lower bound on A to

W p (y) = p(y)[u 1 (y)u 2 (y) − u 2 (y)u 1 (y)].

Differentiating W p with respect to y and cancelling the pu 1 u 2 terms

W p = p (u 1 u 2 − u 2 u 1 ) + p[u 1 u 2 − u 2 u 1 ].

it is clear that F is an extension of F and that F = f .

is a bounded linear functional on C 0 ([a, b]). Since xn x, we have

xn − x2 = xn 2 + x2 − 2(x, xn ),

lim xn − x2 = 0,

ψu, φn = u, ψφn → u, ψφn = ψu, φ,

|φk − u| ≤ /4M for all k≥K

u3L 3 ≤ C|u|3/2 |Du|3/2 ,

v∞ ≤ Cu H k− j ≤ Cu H k .

sup |(u n − u, φ)| ≥ .

Then there exist φn with φn H01 = 1 such that

|(u n − u, φ)| ≥ /4