Solutions to Partial Differential

Equations by Lawrence Evans

Matthew Kehoe

May 22, 2021

Abstract. These are my solutions to selected problems from chapters 5–9 of

Partial Differential Equations by Lawrence Evans. Any mistakes in these
solutions are my own. I plan to write more solutions in the future. If you
would like to speak with me about these solutions (or about anything related
to PDEs) then I can be contacted at [email protected].

Contents

Chapter 5 Solutions 3
5.10.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

5.10.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
5.10.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5.10.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
5.10.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

5.10.11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.10.13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.10.15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

5.10.17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.10.18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Exercise (Hardy’s Inequality on R+ ) . . . . . . . . . . . . . . . . 16

1
 Evans Chapters 5 - 9

Chapter 6 Solutions 17
6.6.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.6.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6.6.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

6.6.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.6.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.6.13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

6.6.15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Elasticity Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

Chapter 7 Solutions 33
7.5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

7.5.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
7.5.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.5.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

7.5.13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.5.14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.5.15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.5.16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

Strong Maximum Principle Exercise . . . . . . . . . . . . . . . . 45

Chapter 8 Solutions 47
8.7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.7.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
8.7.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

8.7.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.7.15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

2
 Evans Chapters 5 - 9

Chapter 9 Solutions 54
9.7.11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
9.7.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
9.7.13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

Chapter 5 Solutions

5.10.4 Assume n = 1 and u ∈ W 1,p (0, 1) for some 1 ≤ p < ∞.

(a) Show that u is equal a.e. to an absolutely continuous function u0 (which

exists a.e.) belongs to Lp (0, 1).
(b) Prove that if 1 < p < ∞, then
Z 1 1/p
1
|u(x) − u(y)| ≤ |x − y|1− p |u0 |p dt
0

for a.e. x, y ∈ [0, 1].

We first state a lemma summarizing the relationship between absolutely

continuous functions and the fundamental theorem of calculus.
Lemma 1. A function U on [a, b] is absolutely continuous if and only if
Z x
U (x) = U (a) + u(t) dt
a

for some integrable function u on [a, b].

Proof. The sufficiency part of the lemma follows directly from the fundamental
theorem of calculus. That is, if u is integrable on [a, b], and if U is defined by
Z x
U (x) := u(t) dt, a ≤ x ≤ b,
a

then U 0 (x) = u(x) for almost every x in [a, b]. To prove the necessity part, we
let U be an absolutely continuous function on [a, b]. Then U is differentiable
almost everywhere and U 0 is integrable on [a, b]. Let

Z x
G(x) := U (a) + U 0 (t) dt, x ∈ [a, b].
a

3
 Evans Chapters 5 - 9

By the fundamental theorem of calculus, G0 (x) = U 0 (x) for almost every

x ∈ [a, b]. It then follows that (U − G)0 (x) = 0 for almost every x ∈ [a, b].
Therefore, U − G is a constant. But U (a) = G(a). Therefore, U (x) = G(x) for
almost every x ∈ [a, b].
Rx
For (a), let [a, b] = [0, 1] and v(x) = 0 u0 (s) ds. Then v is an absolutely
continuous function by Lemma 1 (also see Rudin [7] for a proof). Therefore for
any test function φ ∈ Cc∞ (0, 1) :
Z 1 Z 1 Z x Z 1
v(x) − u(x) φ0 (x) dx = u0 (s) ds φ0 (x) dx − u(x)φ0 (x) dx


0 0 0 0
Z 1 Z 1 Z 1
= φ0 (x) dx u0 (s) ds − u(x)φ0 (x) dx
0 s 0
Z 1 Z 1
0
=− φ(s)u (s) ds + u0 (x)φ(x) dx
0 0
Z 1 Z 1
0
=− φ(x)u (x) dx + u0 (x)φ(x) dx = 0.
0 0

As φ was chosen arbitrarily, we see that u is equal a.e. to an absolutely

continuous function u0 as required. For (b), since u0 is in Lp (0, 1) we apply
Holder’s inequality with q = p/(p − 1):
Z x
|u(x) − u(y)| ≤ |u0 (t)|dt
y
Z x (p−1)/p Z x 1/p
≤ dt |u0 |p dt
y y
Z 1 1/p
1−1/p 0 p
≤ |x − y| |u | dt .
0

SN
5.10.6 Assume U is bounded and U ⊂⊂ i=1 Vi . Show there exist C ∞
functions ζi (i = 1, 2, . . . , N ) such that
(
0 ≤ ζi ≤ 1, spt ζi ⊂ Vi (i = 1, 2, . . . , N )
PN
i=1 ζi = 1 on U .

The functions {ζi }N

i=1 form a partition of unity.

Proof. We first complete problem 5.10.5. Let U and V be open sets with
V ⊂⊂ U . We need to show that there exists a smooth function ζ such that
ζ ≡ 1 on V and ζ = 0 near ∂U .

4
 Evans Chapters 5 - 9

As suggested in the hint by Evans, take V ⊂⊂ W ⊂⊂ U and let ε > 0 be the

distance between V and ∂U . Then define
n εo
W := x ∈ U : d(x, V ) < .
2
By making this distance small enough, we have constructed an open set W
which is contained between U and V . Let ε̃ = ε/8. Then
1 x
ηε̃ (x) = η
ε̃n ε̃
is the required mollifier as suggested in Appendix C of Evans. Define

ψ(x) := ηε̃ ∗ χW (x),

where spt(ψ) ⊆ spt(ηε̃ ) + spt(χW ) ⊂ U and is therefore a smooth function.

Hence
Z
ψ(x) = ηε̃ (x − y)χW (y) dy
n
ZR
= ηε̃ (x − y)χW (y) dy
B(x,ε̃)∩W
Z
= ηε̃ (x − y) dy.
B(x,ε̃)∩W

So if B(x, ε̃) ⊂ W , we see

Z Z
ηε̃ (y) dy = ηε̃ (y) dy = 1,
Rn B(0,ε̃)

which implies that the support is in W ∩ B(0, ε̃). As W̄ is compact, we will

cover it and its boundary by open balls. Let
N
[
W̄ ⊂ Wi
i=1

where Wi denotes an open ball which covers a portion of W and possibly the
boundary. Then we may observe that we can use the mollifier ηε̃i for every
open ball Wi where
ψi (x) := ηε̃i ∗ χW (x).
By defining
N
X ψi (x)
ζ(x) := PN ,
i=1 i=1 ψi (x)

we observe that for any fixed x ∈ U , only three terms in the sum will be
nonzero. As V ⊂⊂ W ⊂⊂ U , it is clear that ζ ≡ 1 on V and ζ = 0 near ∂U .

5
 Evans Chapters 5 - 9

SN SN
To complete 5.10.6, we assume U is bounded and U ⊂⊂ i=1 Wi ⊂⊂ i=1 Vi .
So, U has a finite cover {V1 , . . . , Vn } where for every Vi , we have a ψi as
constructed above. The support of ψi is contained entirely in Vi , ψi ≡ 1 on
Wi , and every ψi is smooth by definition. Therefore, we define
ψi (x)
ζi (x) := PN
i=1 ψi (x)
P
where i ζi ≡ 1 by construction and for all x ∈ U , the support of ζi is
contained in Vi . Also, ζi is smooth because the ψi are smooth. Thus, the
collection {ζi }N
i=1 fulfills all of the requirements and is a partition of unity
subordinate to the cover {V1 , . . . , Vn }.

5.10.7 Assume that U is bounded and there exists a smooth vector field α
such that α · ν ≥ 1 along ∂U , where ν as usual denotes the outward unit
normal. Assume 1 ≤ p < ∞.
Apply the Gauss-Green Theorem to ∂U |u|p α · ν dS, to derive a new proof of
R

the trace inequality

Z Z
|u|p dS ≤ C |Du|p + |u|p dx
∂U U
1
for all u ∈ C (Ū ).

Proof. As α · ν ≥ 1 along ∂U , we have |u|p ≤ |u|p α · ν. Then

Z Z
p
|u| dS ≤ |u|p α · ν dS
∂U ∂U
Z
= ∇ · (|u|p α) dx (Gauss−Green)
U
Z
= |u|p (∇ · α) + α · ∇|u|p dx
U
Z
≤C |u|p + |∇|u|p | dx.
U

Therefore, since
∇|u|p = p|u|p−1 (sgn u)Du,
we have for p = 1 Z Z
|u| dS ≤ C |u| + |Du| dx.
∂U U
On the other hand, if p > 1 then we apply Young’s inequality with
a = |Du|, b = |u|p−1 , q = p/(p − 1) to form
Z Z Z
p p−1
|∇|u| | dx ≤ C p|u| |Du| dx ≤ C |Du|p + (p − 1)|u|p dx.
U U U

6
 Evans Chapters 5 - 9

The constants above are different at every inequality. We may now observe
that
Z Z
|u|p dS ≤ C |u|p + |Du|p + (p − 1)|u|p dx
∂U
ZU
≤C |u|p + |Du|p dx,
U

as required.

5.10.8 Let U be bounded, with a C 1 boundary. Show that a “typical”

function u ∈ Lp (U ) (1 ≤ p < ∞) does not have a trace on ∂U . More precisely,
prove that there does not exist a bounded linear operator

T : Lp (U ) → Lp (∂U )

such that T u = u|∂U whenever u ∈ C(Ū ) ∩ Lp (U ).

Proof. We will construct a counterexample in L2 (U ). We need to show that

there does not exist a constant C > 0 such that kT ukL2 (∂U ) ≤ CkukL2 (U ) for
every u ∈ L2 (U ). Let’s consider the following sequence of continuous functions:
1
un (x) = , x ∈ U.
1 + nd(x, ∂U )

Then
0 ≤ un (x) ≤ 1,
un (x) = 1, x ∈ ∂U.
Therefore, for every x ∈ U , we see that un (x) → 0 pointwise, so by the
dominated convergence theorem

kun k2L2 (U ) → 0.

However, for every n we have

kT un k2L2 (∂U ) ≤ C 2 kun k2L2 (U ) → 0,

which implies that the area of the boundary is equal to zero. As T un = 1 for
every n we see that we have arrived at a contradiction. The same analysis
works for Lp (U ) when 1 ≤ p < ∞.

7
 Evans Chapters 5 - 9

5.10.9 Integrate by parts to prove the interpolation inequality:

1/2 1/2
kDukL2 ≤ CkukL2 kD2 ukL2
for all u ∈ Cc∞ (U ). Assume U is bounded, ∂U is smooth, and prove this
inequality if u ∈ H 2 (U ) ∩ H01 (U ).
(Hint: Take the sequences {vk }∞ ∞ 1
k=1 ⊂ Cc (U ) converging to u in H0 (U ) and
∞ ∞ 2
{wk }k=1 ⊂ C (Ū ) converging to u in H (U ).)

Proof. Let u ∈ Cc∞ (U ). Integrating by parts and applying Cauchy–Schwarz in

the last inequality forms
Z Z
Du · Du dx ≤ C |u||D2 u| dx
U U
Z 1/2 Z 1/2
2 2 2
≤C |u| dx |D u| dx ,
U U
where the boundary term disappears since u has compact support. Taking the
square root yields
1/2 1/2
kDukL2 (U ) ≤ CkukL2 (U ) kD2 ukL2 (U ) .
Following the hint provided by Evans, since H01 (U ) is the closure of C0∞ (U )
with the norm of H 1 (U ), we can find a sequence {vk }∞ 1 ∞
k=1 in H (U ) ∩ Cc (U )
1
converging to u in H0 (U ). Also, since ∂U is smooth, we can extend U to a set
V such that U ⊂⊂ V . Then, by the density of Cc∞ (V ), we can find a sequence
{wk }∞ ∞ 2
k=1 in C (Ū ) converging to u in H (U ). So, we apply the Gauss−Green
Theorem and evaluate Dvk · Dwk which after one application of
Cauchy–Schwarz yields
Z Z
Dvk · Dwk dx ≤ C |vk ||D2 wk | dx
U U
Z 1/2 Z 1/2
≤C |vk |2 dx |D2 wk |2 dx ,
U U
where the boundary term once again vanishes because vk has compact
support. As k → ∞,
Z 1/2 Z 1/2 Z 1/2 Z 1/2
C |vk |2 dx |D2 wk |2 dx →C |u|2 dx |D2 u|2 dx ,
U U U U

which is equivalent to CkukL2 (U ) kD2 ukL2 (U ) . To show that the left-hand side
converges to kDuk2L2 (U ) , we evaluate the difference and once again apply
Cauchy–Schwarz
Z Z  
(Dvk · Dwk − Du · Du) dx = Dvk · (Dwk − Du) + Du · (Dvk − Du) dx
U
ZU
≤ |Dvk | · |Dwk − Du| + |Du| · |Dvk − Du| dx
U
≤ kDvk kL2 (U ) kDwk − DukL2 (U ) + kDukL2 (U ) kDvk − DukL2 (U ) ,

8
 Evans Chapters 5 - 9

where as k → ∞, the right-hand side goes to zero since both vk , wk → u where

u ∈ H 2 (U ) ∩ H01 (U ). This implies that
Z Z
Dvk · Dwk dx → |Du|2 dx,
U U

and we may conclude

kDuk2L2 (U ) ≤ CkukL2 (U ) kD2 ukL2 (U )

which after taking the square root forms

1/2 1/2
kDukL2 (U ) ≤ CkukL2 (U ) kD2 ukL2 (U ) .

5.10.11 Suppose U is connected and u ∈ W 1,p (U ) satisfies

Du = 0 a.e. in U .

Prove u is constant a.e. in U .

Proof. First Solution: Consider the domainR Uε = {x ∈ U : dist(x, ∂U ) > ε}.

For x ∈ Uε consider the function uε (x) = ηε (x − y)u(y) dy where ηε is a
standard mollifier. Then uε ∈ C ∞ (Uε ) and
Z
Duε (x) = ηε (x − y)Du(y) dy

by the definition of the weak derivative. Since Du = 0 a.e., we have that

Duε (x) = 0 for all x ∈ Uε and hence uε is constant in Uε . Since
kuε − ukLp (Uε ) → 0 as ε → 0, we have that u is constant a.e..
Second Solution: Let ε > 0. Then consider the smooth functions

uε = ηε ∗ u ∈ C ∞ (Uε ),

where Uε = {x ∈ U : d(x, ∂U ) > ε}. By Theorem 5.3.1 in Evans, we have

Dxi (uε ) = ηε ∗ Dxi u.

Therefore, by assumption, Dxi uε = 0 a.e. in Uε . So uε is constant on each

connected subset of Uε .

Next, let x, y ∈ U . Since U is connected, there exists a polygonal path Γ ⊆ U

which connects x and y.
Let δ = min d(z, ∂U ) be the minimum distance of points in Γ to the boundary
z∈Γ
of U . Then for every ε < δ the whole polygonal curve Γ is in Uε . So x, y lie in
the same connected component of Uε . Hence uε (x) = uε (y) ≡ const.

9
 Evans Chapters 5 - 9

Figure 1: If U is connected then its subdomain Uε may not be connected.

However, any two points x, y ∈ U can be connected by a polygonal path Γ
remaining inside U . So, if ε > 0 is sufficiently small, then x and y belong to the
same connected component of Uε .

As u ∈ W 1,p (U ), Theorem 7 in Appendix C of Evans tells us that

ε→0
uε −→ u a.e. in U .

Thus, u is a constant a.e. in U .

5.10.13 Give an example of an open set U ⊂ Rn and a function

u ∈ W 1,∞ (U ), such that u is not Lipschitz continuous on U . (Hint: Take U to
be the open unit disk in R2 , with a slit removed).

Proof. Consider the slit plane

A = {(r, θ) : 0 < r < ∞, − π < θ < π} ⊂ R2 ,

in polar coordinates. The function

 
θ
u(r, θ) = r sin
2
is continuous and locally Lipschitz. However, if we let ε > 0 and define a
compact subset of A as

K = {(r, θ) : 1/2 ≤ r ≤ 3/2, − π + ε ≤ θ ≤ π − ε} ⊂ R2 ,

then we will see that the Lipschitz constant will be unbounded. Define r = 1
and θ = ±(π − ε) and consider two points in K where x1 (r, θ) = (1, π − ε) and
x2 (r, θ) = (1, ε − π). Then when ε → 0+

sin π−ε − sin ε−π 



|u(x1 ) − u(x2 )| 2 2
sup ≥ lim+
x1 6=x2 |x1 − x2 | ε→0 |cos(π − ε) − sin(ε − π)|

2 cos (ε/2)
= lim
ε→0+ sin(π − ε)
→ ∞.

10
 Evans Chapters 5 - 9

Thus when ε → 0+ we see that the Lipschitz constant is unbounded.

Therefore u ∈ W 1,∞ (U ) is not Lipschitz continuous.

5.10.15 Fix α > 0 and let U = B 0 (0, 1). Show that there exists a constant C,
depending only on n and α, such that
Z Z
u2 dx ≤ C |Du|2 dx,
U U

provided
|{x ∈ U | u(x) = 0}| ≥ α, u ∈ H 1 (U ).

My analysis closely follows the ideas presented by Giovanni Leoni in Chapter

12 of [6]. I will show that a variant of the Poincaré inequality is true in
W 1,2 (U ).
Theorem 2 (Poincaré inequality in W 1,2 (U )). Let p = 2 and let U ⊂ R2 be a
connected extension domain for W 1,2 (U ) with finite measure. Let E ⊂ U be a
Lebesgue measurable set with positive measure. Then there exists a constant
C = C(2, U, E) > 0 such that for all u ∈ W 1,2 (U ),
Z Z
|u(x) − uE |2 dx ≤ C |Du(x)|2 dx, (1)
U U

where Z Z
1 1
uE := u(x) dx = N
u(x) dx.
|E| E L (E) E

Proof. Let E := {x ∈ U | u(x) = 0}. Then |E| > 0 and uE = 0 by definition.

Assume by contradiction that the result is false. Then, there exists a sequence
(un ) ⊂ H 1 (U ) = W 1,2 (U ) such that
Z Z
2
|un (x) − (un )E | dx > n |Dun (x)|2 dx > 0.
U U

Define
un − (un )E
vn := , (2)
kun − (un )E kL2 (U )
then vn ∈ W 1,2 (U ) and
Z
1
kvn kL2 (U ) = 1, (vn )E = 0, |Dvn |2 dx < ,
U n
where the last statement follows from
1
kDvn k2L2 (U ) < kvn k2L2 (U ) .
n

11
 Evans Chapters 5 - 9

By the Rellich–Kondrachov Compactness theorem, there exists a subsequence

(vnk ) such that vnk → v in L2 (U ) for some function v ∈ L2 (U ). So (2) and the
definition of uE imply
kvkL2 (U ) = 1, vE = 0.
Therefore for every φ ∈ Cc1 (U ) and i = 1, 2, . . . , N we find by Hölder’s
inequality
Z  Z  Z 
 v ∂φ dx = lim  vn ∂φ dx = lim  ∂vnk
     
k
φ dx
U ∂xi k→∞ ∂xi k→∞ U ∂xi
     
U
Z 1/2 Z 1/2
2 2
≤ lim |Dvnk | dx |φ| dx
k→∞ U U
→ 0,

Consequently v ∈ W 1,2 (U ) where Dv = 0 a.e. As U is connected, v must be a

constant (see 5.10.11). However, as vE = 0 and v is a constant, we see that
v = 0. This is a direct contradiction to the fact that kvkL2 (U ) = 1 and
completes the proof. As uE = 0, (1) implies that result is true for 5.10.15.

5.10.17 (Chain Rule) Assume F : R → R is C 1 , with F 0 bounded. Suppose U

is bounded and u ∈ W 1,p (U ) for some 1 ≤ p ≤ ∞. Show

v := F (u) ∈ W 1,p (U ) and vxi = F 0 (u)uxi (i = 1, . . . , n).

(Hint: Use that any sequence that converges in Lp has a subsequence that
converges pointwise a.e.)

Proof. First Solution: Let u ∈ W 1,p (U ). There exists

uk ∈ C ∞ (U ) ∩ W 1,p (U ) such that

ku − uk kW 1,p (U ) → 0

as k → ∞. Now

|F (uk )(x) − F (u` )(x)| ≤ C|uk (x) − u` (x)|

since F 0 is bounded. Moreover

DF (uk )(x) = F 0 (uk )(x)Duk (x).

From this it follows that for any test function φ ∈ Cc∞ (U )

Z Z Z
F (u)Dφ dx = lim F (uk )Dφ dx = − lim F 0 (uk )Duk φ dx.
k→∞ k→∞

12
 Evans Chapters 5 - 9

We may choose a subsequence so that uk tends a.e. to u and hence F 0 (uk ) a.e.
to F 0 (u). Now
Z 
0 0



 F (uk )Du k − F (u)Du φ dx

Z  


F 0 (uk ) − F 0 (u) Duk + F 0 (u) Duk − Du φ dx


= 
Z  
≤C kDuk k∞ |F 0 (uk ) − F (u)| + kF 0 (u)k∞ kuk − ukW 1,p (U ) |φ| dx

which tends to zero. Hence

Z Z
F (u)Dφ dx = − F 0 (u)Duφ dx.

Since U is bounded F (u) ∈ Lp (U ) and since F 0 (U ) is bounded

F 0 (U )uxi ∈ Lp (U ) and hence F (u) ∈ W 1,p (U ).
Second Solution: We assume 1 ≤ p < ∞. Let φ ∈ Cc∞ (U ). By the density
theorem, there is a sequence (un ) ⊂ C ∞ (U ) such that un → u in W 1,p (U ).
Therefore, un → u and Dun → Du a.e. in U. Defining vn = F (un ), we see that
both F, un ∈ C 1 , so vn ∈ C 1 . This implies that we may use the chain rule for
smooth functions to form
Z Z
∂φ ∂un
− F (un ) dx = F 0 (un ) φ dx. (3)
U ∂xi U ∂xi

As F 0 is bounded, we let M = sup|F 0 (t)|. Then

t∈R
Z  Z
 (F (un ) − F (u)) ∂φ dx ≤ k ∂φ kL∞ M
 
|un − u| dx → 0 as n → ∞,

U ∂xi  ∂xi U

because un → u a.e. in U . Also

Z    Z 
 0 ∂un 0 ∂u   0
  ∂u
  n ∂u 
 F (u n ) − F (u) φ dx ≤ kφkL ∞ (u n ) · −  dx
∂xi ∂xi ∂xi ∂xi
  F  
U
ZU    ∂u 
 0
+ kφkL∞ F (un ) − F 0 (u) ·   dx,
  
U ∂xi

where
Z    ∂u Z 
 0   n ∂u   ∂un ∂u 
(u )
F n   · − dx ≤ M −  dx → 0 as n → ∞.
∂xi ∂xi U ∂xi ∂xi
 
U

Similarly, since F 0 (un ) → F 0 (u) pointwise a.e. and

   ∂u   ∂u 
 0
F (un ) − F 0 (u) ·   ≤ 2M 
    
∂xi ∂xi
,

13
 Evans Chapters 5 - 9

we apply the dominated convergence theorem to conclude

Z    ∂u 
 0
F (un ) − F 0 (u) ·   dx → 0 as n → ∞.
  
U ∂xi

Together, these imply

Z   
 0 ∂un 0 ∂u 
 F (un ) − F (u) φ dx → 0 as n → ∞.

U ∂xi ∂xi 

We now take n → ∞ in (3) to obtain

Z Z
∂φ ∂u
− F (u) dx = F 0 (u) φ dx, (4)
U ∂xi U ∂xi

which is the desired form of vxi = F 0 (u)uxi . As both F and u are sufficiently
smooth, it follows that the right-hand side of (4) is in Lp (U ), which implies
that vxi ∈ Lp (U ). To see that v ∈ W 1,p (U ), we may observe
Z Z Z
p p p
|v| dx = |F (u) − F (0)| dx ≤ M |u|p < ∞.
U U U

For p = ∞, we review Chapter 5.8 of Evans. Let U be open and bounded,

with ∂U of class C 1 . Then u : U → R is Lipschitz continuous ⇐⇒
u ∈ W 1,∞ (U ). Therefore, W 1,∞ (U ) is the space of Lipschitz continuous
functions in U . So for every x, y ∈ U
Z 
 u(x) 
0
|F (u(x)) − F (u(y))| =  F (t) dt ≤ M |u(x) − u(y)|.
 
 u(y) 

Then, since u is a Lipschitz function it must have Lipschitz constant, say

|u(x) − u(y)| ≤ L|x − y| for every x, y ∈ U . This implies

|F (u(x)) − F (u(y))| ≤ M L|x − y|,

which shows that v = F (u) ∈ W 1,∞ (U ). Theorem 6 in Section 5.8 of Evans is

known as Rademacher’s Theorem. It states that if u is locally Lipschitz
continuous in U then u is differentiable almost everywhere in U . As we have
shown that v = F (u) is locally Lipschitz continuous, it follows that
vxi = F 0 (u)uxi as in the case when 1 ≤ p < ∞.

5.10.18 Assume 1 ≤ p ≤ ∞ and U is bounded.

(a) Prove that if u ∈ W 1,p (U ), then |u| ∈ W 1,p (U ).

14
 Evans Chapters 5 - 9

(b) Prove u ∈ W 1,p (U ) implies u+ , u− ∈ W 1,p (U ), and

(
Du a.e. on {u > 0}
Du+ =
0 a.e. on {u ≤ 0},
(
0 a.e. on {u ≥ 0}
Du− =
−Du a.e. on {u < 0}.
(Hint: u+ = limε→0 Fε (u), for
(
(z 2 + ε2 )1/2 − ε if z ≥ 0
Fε (z) :=
0 if z < 0.)

(c) Prove that if u ∈ W 1,p (U ), then

Du = 0 a.e. on the set {u = 0}.

Proof. We need to show that if u ∈ W 1,p (U ), then u+ , u− , |u| ∈ W 1,p (U ). By

Appendix A.3 of Evans,

u+ = max(u, 0), u− = − min(u, 0).

Following the hint for (b), we let

(√
z 2 + ε2 − ε, if z ≥ 0,
Fε (z) =
0, if z < 0,

therefore Fε (z) ∈ C 1 (R) and

√ z

, if z > 0,
(Fε )0 (z) = z 2 + ε2
0, if z ≤ 0,

which implies that k(Fε )0 kL∞ (R) ≤ 1 for every ε > 0. Also, F (z) =
limε→0 Fε (z), where (
z, if z ≥ 0,
F (z) =
0, if z < 0.
The conditions of the chain rule for W 1,p are satisfied and we have
Z Z
∂φ ∂u
Fε (u) dx = − (Fε )0 (u) φ dx,
U ∂xj U ∂x j

for every φ ∈ C0∞ (U ). Observing that

(
0 1, on {u > 0},
lim (Fε ) (u) =
ε→0 0, on {u ≤ 0},

15
 Evans Chapters 5 - 9

and
u+ = lim Fε (u) in U ,
ε→0

we apply the dominated convergence theorem k(Fε )0 kL∞ (R) < ∞ as in

5.10.17 to find
Z Z
∂φ ∂φ
u+ dx = lim Fε (u) dx
U ∂xj U ε→0 ∂x j
Z
∂φ
= lim Fε (u) dx
ε→0 U ∂xj
Z
∂u
= − lim (Fε )0 (u) φ dx (Chain Rule)
ε→0 U ∂xj
Z
∂u
=− lim (Fε )0 (u) φ dx (DCT)
ε→0 ∂xj
ZU
∂u
=− φ dx.
u>0 ∂xj

This shows that (

+ Du, a.e. on {u > 0},
Du =
0, a.e. on {u ≤ 0}.
As u− = (−u)+ , an analogous argument will complete (b). Then, (a) follows
from the fact that |u| = u+ + u− .

For (c), we have

∂u ∂u+ ∂u−
u = u+ − u− =⇒ = − ,
∂xi ∂xi ∂xi
therefore ∂u/∂xi = 0 a.e. on {u = 0} and


Du,
 a.e. on {u > 0},
D|u| = 0, a.e. on {u = 0},

−Du, a.e. on {u < 0}.


Exercise (Hardy’s Inequality on R+ ) Let p ∈ (1, ∞). Then there exists a

constant C = C(p) < ∞ such that for u ∈ W 1,p (U ) with T u = 0,
Z ∞ Z t p  p Z ∞
1 p
|f (s)|p ds.

 f (s) ds dt ≤
0
t
0
 p − 1 0

Proof. We will show

Z ∞ Z t p !1/p Z ∞ 1/p
1  p p
 f (s) ds dt
 ≤ |f (s)| ds . (5)
0
t
0 p−1 0

16
 Evans Chapters 5 - 9

Observe that Z t Z 1
1
f (s) ds = f (ts) ds.
t 0 0
Therefore by applying Minkowski’s integral inequality to the left hand side of
(5), we see

Z ∞ Z 1 p !1/p Z 1 Z ∞ 1/p
  p

 f (ts) ds dt ≤ |f (ts)| dt ds
0 0 0 0
Z 1 Z ∞ 1/p
−1/p p
= s ds |f (t)| dt
0 0
Z ∞ 1/p
p p
= |f (s)| ds ,
p−1 0

as required.

Chapter 6 Solutions

6.6.3 A function u ∈ H02 (U ) is a weak solution of this boundary-value

problem for the biharmonic equation
(
∆2 u = f in U
(∗)
u = ∂u
∂ν = 0 on ∂U

provided Z Z
∆u∆v dx = f v dx
U U

for all v ∈ H02 (U ). Given f ∈ L2 (U ), prove that there exists a unique weak
solution of (∗).

Proof. We first derive a variation of the Poincaré inequality. We have that

k·k2,2
H02 (U ) := Cc∞ (U ) . Therefore, if u ∈ H02 (U ) there exists a sequence
∞
(un ) ⊆ Cc (U ) such that un → u ∈ W 2,2 (U ). Thus H02 (U ) consist of the
functions W 2,2 (U ) such that u = 0 and ∇u = 0 on ∂U . As u, ∇u ∈ W01,2 (U )
we see that both of the following Poincaré inequalities are true:

kukL2 (U ) ≤ C1 k∇ukL2 (U ) ,

k∇ukL2 (U ) ≤ C2 k∇2 ukL2 (U ) ,

17
 Evans Chapters 5 - 9

where C1 , C2 > 0. These two inequalities allow us to obtain a Poincaré

inequality for the H02 (U )-norm

kukH02 (U ) = kuk2L2 (U ) + k∇uk2L2 (U ) + k∇2 uk2L2 (U )

≤ (1 + C1 + C2 )k∇2 uk2L2 (U )
= C̃k∇2 uk2L2 (U ) .

But we also have that

k∇2 uk2L2 (U ) ≤ kuk2H 2 (U ) .
0

2
Therefore k∇ uk2L2 (U ) is an equivalent norm in H02 (U ) and we think of

kuk22,2 = k∇2 uk2L2 (U ) (6)

as our norm on H02 (U ). We now claim that

k∆ukL2 (U ) = k∇2 ukL2 (U ) , (7)

for every u ∈ H02 (U ). To prove this, we consider u ∈ Cc∞ (U ). Integrating by

parts and commuting partial derivatives gives
Z Z
uxi xi uxj xj dx = uxi xj uxi xj dx
U U

for every 1 ≤ i, j ≤ N . We then sum over all i and j to arrive at

k∆ukL2 (U ) = k∇2 ukL2 (U )

for every u ∈ Cc∞ (U ). As Cc∞ (U ) is dense in H02 (U ), we pass to the limit to

find
k∆ukL2 (U ) = k∇2 ukL2 (U ) , ∀u ∈ H02 (U ).
This proves (7). Then (6) implies that

kukH02 (U ) ≤ Ck∆ukL2 (U ) . (8)

A similar analysis is shown in Stein: Singular integrals and Differentiability

Properties of Functions[8].

For the biharmonic equation, we let f ∈ L2 (U ). We consider the bilinear form

B : H02 (U ) × H02 (U ) → R defined by
Z
B[u, v] := ∆u∆v dx ∀u, v ∈ H02 (U ),
U

and the linear functional a : H02 (U ) → R defined by

Z
a(v) := f v dx ∀v ∈ H02 (U ).
U

18
 Evans Chapters 5 - 9

We need to show that the hypotheses of the Lax-Milgram theorem are

satisfied. First we observe that B is continuous because
Z 
 
|B[u, v]| =  ∆u∆v dx

U
Z 1/2 Z 1/2
≤ |∆u|2 dx |∆v|2 dx
U U
≤ CkukH02 (U ) kvkH02 (U ) ,

by the Cauchy-Schwarz inequality and the definition of the H02 (U )-norm.

Moreover, B is coercive as
Z
B[u, u] = |∆u|2 dx = k∆uk2L2 (U ) ≥ Ckuk2H 2 (U ) ,
0
U

because of (8). Also, the functional a is continuous since

Z
|a(v)| ≤ |f ||v| dx
U
Z 1/2 Z 1/2
≤ |f |2 dx |v|2 dx
U U
= kf kL2 (U ) kvkL2 (U )
≤ γkvkH01 (U ) ,

where γ := kf kL2 (U ) . Therefore, by the Lax-Milgram theorem we see that

there exists a unique u ∈ H02 (U ) such that

B[u, v] = a(v) ∀v ∈ H02 (U ).

Hence Z Z
∆u∆v dx = f v dx ∀v ∈ H02 (U ),
U U

and the Lax-Milgram theorem provides a unique weak solution of (∗).

6.6.5 Explain how to define u ∈ H 1 (U ) to be a weak solution of Poisson’s

equation with Robin boundary conditions :
(
−∆u = f in U
u + ∂u
∂ν = 0 on ∂U .

Discuss the existence and uniqueness of a weak solution for a given f ∈ L2 (U ).

19
 Evans Chapters 5 - 9

Proof. To define a weak solution of Poisson’s equation we multiply the

equation by v ∈ H 1 (U ) and integrate by parts to get
Z Z Z
∂u
− ∆uv dx = Du · Dv dx − v dS
U ∂ν
ZU Z∂U
= Du · Dv dx + uv dS
ZU ∂U

= f v dx.
U

The boundary conditions imply that u is a weak solution if

Z Z Z
Du · Dv dx + (T u)(T v) dS = f v dx, (9)
U ∂U U

where u, v ∈ H 1 (U ) and T u = u|∂U . We therefore define a weak solution to be

a function u ∈ H 1 (U ) which satisfies (9) for all v ∈ H 1 (U ). The Trace theorem
tells us that the boundary conditions can be taken in the sense of traces.

Let f ∈ L2 (U ). We will show that existence and uniqueness of a weak solution

follow from
R the Lax-Milgram theorem. Analogously to 6.6.3, the functional
a(v) := U f v dv is continuous by the Cauchy-Schwarz inequality
Z
|a(v)| ≤ |f ||v| dx ≤ kf kL2 (U ) kvkL2 (U )
U
≤ kf kL2 (U ) kvkH 1 (U )
= γkvkH 1 (U ) .

We define the bilinear form B : H 1 (U ) × H 1 (U ) → R by

Z Z
B[u, v] := Du · Dv dx + (T u)(T v) dS ∀u, v ∈ H 1 (U ).
U ∂U

We need to prove that the bilinear form is continuous and coercive. First, it is
continuous by the Cauchy-Schwarz inequality and the Trace inequality
between H 1 (U ) and L2 (∂U )
Z Z
|B[u, v]| ≤ |Du · Dv| dx + |(T u)(T v)| dS
U ∂U
≤ kDukL2 (U ) kDvkL2 (U ) + kT ukL2 (∂U ) kT vkL2 (∂U )
≤ (C + 1)kukH 1 (U ) kvkH 1 (U ) .

To show that B is coercive, assume by contradiction that it is not. Then there

is a sequence (un ) ⊂ H 1 (U ) such that kun kH 1 (U ) = 1 and B[un , un ] → 0. As
un is bounded in H 1 (U ) and H 1 (U ) is a Hilbert space, it contains a
subsequence which converges weakly to u ∈ H 1 (U ). Let’s assume that un * u.

20
 Evans Chapters 5 - 9

As ∂U is Lipschitz, we have that H 1 (U ) ⊂ L2 (U ) and therefore un → u in

L2 (U ) and Dun * Du ∈ L2 (U ).

As B[un , un ] → 0, we see that

Z Z
2
B[un , un ] = |Dun | dx + T u2n dS → 0, (10)
U ∂U

this implies that

Z Z
|Du|2 dx ≤ lim inf |Dun |2 dx → 0.
U n→∞ U

So by 5.10.11 we know that u is a constant. Now,

Z Z
lim u2n dx = u2 dx = 1.
n→∞ U U

2
However, since T un → T u in L (∂U ), we see from (10)
Z Z
2
T u dS = lim T u2n dS → 0.
∂U n→∞ ∂U

Therefore, as u is a constant it cannot be both 0 at the boundary and 1 inside

U . We have reached a contradiction and conclude that B is coercive. The
Lax-Milgram theorem then gives a unique weak solution to Poisson’s
equation.

6.6.7 Let u ∈ H 1 (Rn ) have compact support and be a weak solution of the
semilinear PDE
−∆u + c(u) = f in Rn ,
where f ∈ L2 (Rn ) and c : R → R is smooth, with c(0) = 0 and c0 ≥ 0. Assume
also c(u) ∈ L2 (Rn ). Derive the estimate

kD2 ukL2 ≤ Ckf kL2 .

(Hint: Mimic the proof of Theorem 1 in §6.3.1, but without the cutoff function
ζ.)

Proof. Define u ∈ H 1 (Rn ) as a weak solution of the semilinear PDE

Z Z Z
Du · Dv dx = f v dx − c(u)v dx, (11)
Rn Rn Rn

provided that v ∈ H 1 (Rn ). As we have compact support, we are integrating

over a large ball. Now let h > 0 be small and choose k ∈ {1, . . . , n}, and then
substitute
v := −Dk−h (Dkh u)

21
 Evans Chapters 5 - 9

into (11). This forms

Z Z Z
− Du · D(Dk−h (Dkh u)) dx = − f Dk−h (Dkh u) dx + c(u)Dk−h (Dkh u) dx
Rn Rn Rn
| {z } | {z } | {z }
A B1 B2

The “integration-by-parts” formula for difference quotients is given in §5.8.2 in

Evans by Z Z
u(Dkh v) dx = − (Dk−h u)v dx.
U U
Applying this formula on A gives
Z
A=− Du · (Dk−h (Dkh (Du)) dx
Rn
Z
= Dkh (Du) · Dkh (Du) dx
n
ZR
= |Dkh (Du)|2 dx.
Rn

Then by Cauchy’s inequality with ε and Theorem 3 in §5.8.2 of Evans

Z
|B1 | ≤ |f | · |Dk−h (Dkh u)| dx
R n
Z Z
C
≤ε |Dk−h (Dkh u)|2 dx + |f |2 dx
Rn ε Rn
Z Z
h 2 C
≤ C1 ε |Dk (Du)| dx + |f |2 dx.
Rn ε Rn

For B2 we first observe

Z 
 u(x) 
|c(u)(x)| =  c (t) dt ≤ |u(x)| · kc0 kL∞ ,
0
 
 0 

therefore
Z
|B2 | ≤ |c(u)| · |Dk−h (Dkh u)| dx
Rn
Z
C
≤ C2 ε |Dkh (Du)|2 dx + kc0 k2L∞ kuk2L2 .
R n ε

Combining the bounds for A, B1 , B2 gives

Z Z
h 2
|Dk (Du)| dx ≤ (C1 + C2 )ε |Dkh (Du)|2 dx
Rn Rn
Z
C C
+ |f |2 dx + kc0 k2L∞ kuk2L2 .
ε Rn ε

22
 Evans Chapters 5 - 9

Taking ε small enough so that it satisfies (C1 + C2 )ε = 12 yields

Z
1 C
|Dkh (Du)|2 dx ≤ kf k2L2 + kc0 k2L∞ kuk2L2 .


2 Rn ε

The above analysis holds for every k ∈ {1, . . . , n} and h > 0 small. Therefore
we apply Theorem 3 in §5.8.2 of Evans to conclude
Z
|D2 u|2 dx ≤ C̃ kf k2L2 + kuk2L2 .


Rn

Hence Du ∈ H 1 (Rn ), so u ∈ H 2 (Rn ).

6.6.8 Let
Pnu be a smooth solution of the uniformly elliptic equation
Lu = − i,j=1 aij (x)uxi uxj = 0 in U . Assume that the coefficients have
bounded derivatives. Set v := |Du|2 + λu2 and show that

Lv ≤ 0 in U

if λ is large enough. Deduce

kDukL∞ (U ) ≤ C kDukL∞ (∂U ) + kukL∞ (∂U ) .

Proof. We first show the inequality Lv ≤ 0 in U. Observe that

n
X
Lu = − aij uxi uxj = 0
i,j=1

implies
n
X n
X
D(Lu) = − Daij uxi uxj − aij Duxi uxj = 0.
i,j=1 i,j=1

Therefore
n
X n
X
− Daij uxi uxj = aij Duxi uxj .
i,j=1 i,j=1

23
 Evans Chapters 5 - 9

We then compute
n
X  
Lv = − aij Du · Du + λu2
xi xj
i,j=1
Xn  
= −2 aij Duxi xj · Du + Duxi · Duxj + λuuxi xj + λuxi uxj
i,j=1
 
n
X n
X n
X n
X
= −2  aij Duxi xj · Du + aij Duxi · Duxj + λu aij uxi xj + λ aij uxi uxj 
i,j=1 i,j=1 i,j=1 i,j=1
n
X n
X n
X
= −2 aij Duxi xj · Du − 2 aij Duxi · Duxj − 0 − 2λ aij uxi uxj
i,j=1 i,j=1 i,j=1
n
X X n n
X
=2 Daij uxi uxj · Du − 2 aij Duxi · Duxj − 2λ aij uxi uxj
i,j=1 i,j=1 i,j=1

≤ 2kDaij kL∞ (U ) |D2 u||Du| − 2θ|D2 u|2 − 2λθ|Du|2

≤ kDaij kL∞ (U ) |D2 u|2 + |Du|2 ) − 2θ|D2 u|2 − 2λθ|Du|2 (Cauchy inequality)
2 2 2
= (C1 − 2θ)|D u| + (C2 − 2λθ)|Du|
≤ 0.
Where the last inequality true provided
(C1 − 2θ)|D2 u|2 C2
λ≥ 2
+ ,
2θ|Du| 2θ
and is obtained for large λ. We now apply the weak maximum principle to
deduce that for large λ > 0
k|Du|2 kL∞ (U ) ≤ k|Du|2 + λu2 kL∞ (U )
≤ k|Du|2 + λu2 kL∞ (∂U )
≤ k|Du|2 kL∞ (∂U ) + λku2 kL∞ (∂U )

2
≤ C k|Du|kL∞ (∂U ) + kukL∞ (∂U ) ,
which implies the desired inequality.

6.6.10 Assume U is connected. Use (a) energy methods and (b) the
maximum principle to show that the only smooth solutions of the Neumann
boundary-value problem

(
−∆u = 0 in U
∂u
∂ν = 0 on ∂U

are u ≡ C, for some constant C.

24
 Evans Chapters 5 - 9

Proof. For (a), we multiply the Neumann problem by a test function v in U to

form
Z Z Z
∂u
0=− ∆uv dx = − v dS + Dv · Du dx (Green’s formula)
U ∂ν
Z ∂U U

= Dv · Du dx.
U

Letting u = v yields Z
|Du|2 dx = 0,
U
which implies that Du = 0 a.e. in U . As U is connected, we know from 5.10.11
that u ≡ C a.e. in U for some constant C.

For (b), assume by contradiction that u is not a constant. Then there must be
some x0 where u attains its maximum in Ū . If x0 ∈ U , then the strong
maximum principle implies that u is constant in contradiction to our
assumption. Therefore, the maximum can only be obtained at the boundary of
U and we conclude that u(x0 ) > u(x) for all x ∈ U . However, Hopf’s lemma
then implies that
∂u 0
(x ) > 0
∂ν
in contradiction to ∂u
∂ν = 0 on ∂U . So u must obtain its maximum inside U
and the strong maximum principle implies that u ≡ C.

Pn
6.6.13 (Courant minmax principle) Let L = − i,j=1 (aij uxi )xj , where
((aij )) is symmetric. Assume the operator L, with zero boundary conditions,
has eigenvalues 0 < λ1 < λ2 ≤ · · · . Show
λk = max
P min B[u, u] (k=1,2,. . . ).
S∈ k−1 u∈S ⊥
kukL2 =1

denotes the collection of (k − 1)-dimensional subspaces of H01 (U ).

P
Here k−1

Proof. Let λk denote an eigenvalue and φk an eigenfunction. We will apply

the theorem of compact, self-adjoint operators as in Appendix D of Evans to
select an orthonormal basis. By defining
νk = max
P min B[u, u] (k=1,2,. . . ),
S∈ k−1 u∈S ⊥
kukL2 =1

we need to show that νk = λk .

Choose any subspace S where dim S = k − 1. Then select

k
X
u= cj φj
j=1

25
 Evans Chapters 5 - 9

where φ1 , . . . , φk are orthonormal and every φj is an eigenfunction of λj for

j = 1, . . . , k. We will choose the numbers cj such that u 6= 0 and u ∈ S ⊥ . We
then select the basis {e1 , . . . , ek−1 } in S and consider the system of equations
where u is orthogonal to e1 , . . . , ek−1 in L2 (U )
k
X
cj hφj , ei i = 0, i = 1, . . . , k − 1.
j=1

By construction, the c1 , . . . , ck form k unknowns and the coefficients hφj , ei i

form k − 1 equations. Therefore there exists a nontrivial solution d1 , . . . , dk . So
k
X
u= dj φj
j=1

is in S ⊥ and is nonzero and hence we may assume that it is normalized. Then

k
X
B[u, u] = λj |cj |2
j=1

because the eigenfunctions are orthonormal. As the λj are ordered in an

increasing order we see that
k
X
B[u, u] ≤ λk |cj |2 = λk kuk2L2 = λk .
j=1

Therefore
min B[u, u] ≤ λk ,
u∈S ⊥
kukL2 =1

As this holds for any subspace S,

max
P min B[u, u] ≤ λk .
S∈ k−1 u∈S ⊥
kukL2 =1

So νk ≤ λk . The converse inequality follows from choosing

S = span(u1 , . . . , uk−1 ). Since the eigenfunctions are orthonormal and
normalized, the eigenvalues may be written as

min B[u, u] = λk ,
u∈S ⊥
kukL2 =1

which implies that

νk = max
P min B[u, u] ≥ λk ,
S∈ k−1 u∈S ⊥
kukL2 =1

hence νk ≥ λk .

26
 Evans Chapters 5 - 9

6.6.15 (Eigenvalues and domain variations) Consider a family of smooth,

bounded domains U (τ ) ⊂ Rn that depend smoothly upon the parameter
τ ∈ R. As τ changes, each point on ∂U (τ ) moves with velocity v. For each τ ,
we consider eigenvalues λ = λ(τ ) and the corresponding eigenfunctions
w = w(x, τ ) :

(
−∆w = λw in U (τ )
w=0 on ∂U (τ ),
normalized so that kwkL2 (U (τ )) = 1. Suppose that λ and w are smooth
functions of τ and x. Prove Hadamard’s variational formula

 ∂w 2
Z  
λ̇ = −   v · ν dS,
∂U (τ ) ∂ν
 

d
where · = dτ and v · ν is the normal velocity of ∂U (τ ). (Hint: Use the calculus
formula from §C.4.)

Proof. Let w = w(x, τ ) and kwkL2 (U (τ )) = 1. Then

Z
λ(τ ) = λ(τ ) · kwkL2 (U (τ )) = λ(τ ) |w|2 dx
U (τ )
Z
= (−∆w)w dx
U (τ )
Z
= |∇w|2 dx.
U (τ )

By §C.4 in Evans,
Z
d
λ̇(τ ) = |∇w|2 dx
dτ U (τ )
Z Z
|∇w|2 v · ν dS + |∇w|2 τ dx


=
∂U (τ ) U (τ )
Z Z
= |∇w|2 v · ν dS + 2∇w∇wτ dx
∂U (τ ) U (τ )
Z Z
= |∇w|2 v · ν dS + 2w(−∆wτ ) dx (Integrate by parts).
∂U (τ ) U (τ )

To simplify the second integral, we see that

−∆wτ = (λ(τ )w)τ = λ(τ )wτ + λ̇(τ )w

27
 Evans Chapters 5 - 9

and because w = 0 on ∂U (τ )
Z
d d
0= kwkL2 (U (τ )) = |w|2 dx
dτ dτ U (τ )
Z Z
|w|2 v · ν dS + |w|2 τ dx


=
∂U (τ ) U (τ )
Z
|w|2 τ dx


=
U (τ )
Z
=2 wwτ dx.
U (τ )

Therefore
Z Z  
2
λ̇(τ ) = |∇w| v · ν dS + 2w λ(τ )wτ + λ̇(τ )w dx
∂U (τ ) U (τ )
Z Z Z
2
= |∇w| v · ν dS + 2λ(τ ) wwτ dx + 2λ̇(τ ) w2 dx
∂U (τ ) U (τ ) U (τ )
Z
= |∇w|2 v · ν dS + 2λ̇(τ ).
∂U (τ )

 2
The boundary condition w = 0 on ∂U (τ ) implies |∇w|2 is equivalent to  ∂w
∂ν

because the gradient is perpendicular to the boundary. This yields

 ∂w 2
Z  
−λ̇(τ ) =   v · ν dS,
∂U (τ ) ∂ν
 

as required.

Elasticity Exercise The homogeneous Dirichlet boundary conditions, reads

(
−divAe(u) = f in U
. (12)
u = 0 at ∂U

The unknown is a map u : U → Rn which represents the displacement of an

elastic body to which a force f : U → Rn is applied. Here, as usual, U ⊂ Rn
is an open and bounded domain. The notation e(u) = sym Du stands for the
symmetric part of the matrix of first partial derivatives of u. In components,
1
eij (u) = (∂i uj + ∂j ui ), i, j = 1, . . . , n.
2
The quantity A = (aijkl ) encodes the elastic properties of the material and can
depend on the spatial coordinate x = (x1 , . . . , xn ). It is a fourth order tensor

28
 Evans Chapters 5 - 9

defined using the indices i, j, k, l ∈ {1, . . . , n} so that, in the end, the PDE in
(7) denotes the system of equations
n
X
− ∂j (aijkl ekl (u)) = fi , i = 1, . . . , n.
j,k,l=1

Due to physical reasons, one typically has the symmetries

aijkl = ajikl , aijkl = aijlk , aijkl = aklij

for all choices of the indices, as well as the Legendre–Hadamard conditions

n
X
aijkl ξi ξj pk pl ≥ λ|ξ|2 |p|2 ∀ ξ, p ∈ Rn
i,j,k,l=1

for some λ ∈ (0, ∞). These ensure the system is uniformly elliptic. Assume
throughout that aijkl ∈ L∞ (U ).

(a) Let f ∈ L2 (U ; Rn ). Define an appropriate notion of weak solution for (12),

using the vectorial Sobolev space

H01 (U ; Rn ) = Cc∞ (U ; Rn )

where the closure is taken with respect to the norm

n
X
kukH01 (U ;Rn ) = kui kH01 (U ) .
i=1

Proof. We first multiply (12) by a test function v and integrate over U

Z Z
− divAe(u) · v dx = f v dx, for all v ∈ H01 (U ; Rn ).
U U

By the generalized form of Green’s formula the LHS becomes

Z Z Z
divAe(u) · v dx = (Ae(u))ν · v dS − Ae(u) : ∇v dx,
U ∂U U

where ν is the unit normal vector and

n
X
A:B= Aij Bij .
i,j=1

Therefore
Z Z Z
Ae(u) : ∇v dx − (Ae(u))ν · v dS = f v dx.
U ∂U U

29
 Evans Chapters 5 - 9

The boundary condition of u = 0 on ∂U implies that the second integral on

the LHS vanishes. Decomposing A into its symmetric and anti-symmetric parts
gives A = (A + AT )/2 + (A − AT )/2. This implies
1 1
∇v + ∇v T + Ae(u) : ∇v − ∇v T



Ae(u) : ∇v = Ae(u) :
2 2
= Ae(u) : e(v).

Therefore we have Z Z
Ae(u) : e(v) dx = f v dx
U U

as the weak form of (12) where v ∈ H01 (U ; Rn ).

(b) Prove that weak solutions always exist for f ∈ L2 (U ; Rn ) and are unique.
Hint: establish Korn’s inequality, which says that
Z Z
C |e(u)|2 ≥ |Du|2 ∀ u ∈ H01 (U ; Rn )
U U

for some constant C ∈ (0, ∞).

Proof. Consider B : H 1 (U ; Rn ) × H 1 (U ; Rn ) → R where we define

Z
B[u, v] = Ae(u) : e(v) dx,
U

and F : H 1 (U ; Rn ) → R where
Z
F (v) = f v dx.
U

Our problem is to find u ∈ H 1 (U ; Rn ) such that

B[u, v] = F (v), for all v ∈ H 1 (U ; Rn ).

We will apply the Lax-Milgram theorem to show that weak solutions are unique
and exist for f ∈ L2 (U ; Rn ). We first apply Hooke’s law to write

Ae(u) = 2µe(u) + λ(∇ · u)I,

where where µ > 0 and λ ≥ 0 are called the Lamé constants. Then since
I : e(v) = ∇ · v our bilinear form B[u, v] becomes
Z
B[u, v] = Ae(u) : e(v) dx = 2µ(e(u) : e(v)) + λ(∇ · u, ∇ · v).
U

30
 Evans Chapters 5 - 9

We will show that B is bounded. By the Cauchy-Schwarz inequality

|B[u, v]| = |2µ(e(u) : e(v)) + λ(∇ · u, ∇ · v)|

≤ 2µke(u)kke(v)k + λk∇ · ukk∇ · vk
≤ Ck∇ukk∇vk
≤ C̃kukH01 (U ;Rn ) kvkH01 (U ;Rn ) .

Also, F (v) is continuous through the Cauchy-Schwarz inequality

|F (v)| ≤ kf kL2 kvkL2

≤ kf kL2 kvkH01 (U ;Rn )
≤ CkvkH01 (U ;Rn ) .

To show the coercivity for B[u, v], we rely on the following

Lemma 3. (Korn’s Inequality) There is a constant C such that
Z n
X
Ck∇vk2 ≤ ke(v)k2 = eij (v)eij (v) dx.
U i,j=1

Proof of Lemma: As u = 0 on ∂U , we calculate

Z n Z n    
X X 1 ∂vi ∂vj 1 ∂vi ∂vj
eij (v)eij (v) dx = + + dx
U i,j=1 U i,j=1 2 ∂xj ∂xi 2 ∂xj ∂xi
Z X n  2  2
1 ∂vi ∂vi ∂vj ∂vj
= +2 + dx
4 U i,j=1 ∂xj ∂xj ∂xi ∂xi
n Z
1 2 1 X ∂vi ∂vj
= k∇vk + dx.
2 2 i,j=1 U ∂xj ∂xi

We then need to show that the last term on the RHS is positive. Integrating by
parts and recognizing that v = 0 on ∂U gives
n Z n Z
∂ 2 vj
Z
X ∂vi ∂vj X ∂vj
dx = − vi dx + νj vi dS
i,j=1 U ∂xj ∂xi i,j=1 U
∂xi ∂xj ∂U ∂xi
n Z Z
X ∂vi ∂vj ∂vj
= dx − νi vi dS
i,j=1 U
∂x i ∂x j ∂U ∂x j

n
! n 
X ∂vi X ∂vj 
= dx
i=1
∂xi j=1
∂xj
Z
= (∇ · v)2 dx ≥ 0,
U

31
 Evans Chapters 5 - 9

as desired. Hence ke(v)k2 ≥ Ck∇vk2 .

Coercivity of B[u, v] now follows from

B[u, u] = 2µke(u)k2 + λk∇ · uk2
≥ 2µke(u)k2
≥ Ck∇uk2
≥ C̃kuk2H 1 (U ;Rn ) .
0

Therefore weak solutions for f ∈ L (U ; Rn ) exist and are unique by the Lax-
2

Milgram theorem.
(c) Prove the regularity theorem that if A and f are infinitely differentiable
throughout U , then so must be the unique weak solution u.

Proof. By Hooke’s law, we follow Ciarlet [3] and rewrite the elasticity tensor as
Ae(u) = λ(tr e(u))I + 2µe(u) (13)
where once again µ > 0 and λ ≥ 0 are called the Lamé constants. We then
sketch a proof of the following
Theorem 4. (Regularity of weak solutions to the linear elasticity problem) Let
U be a domain in Rn with a boundary ∂U of class C 2 . Let f ∈ Lp (U ) and
p ≥ 6/5. Then the weak solution u ∈ H01 (U ) of the linear elasticity problem is
in the space W 2,p (U ) and satisfies
−div (λ(tr e(u))I + 2µe(u)) = f in Lp (U ).
Furthermore, let m ≥ 1 be an integer. If the boundary ∂U is of class C m+2 and
if f ∈ W m,p , then the weak solution u ∈ H01 (U ) is in the space W m+2,p (U ).

Sketch of proof: (1) We will proceed similar to Theorem 2 in §6.3 of Evans.

The operator of linear elasticity is strongly elliptic because of the Legendre–Hadamard
conditions. Therefore
f ∈ L2 (U ) =⇒ u ∈ H 2 (U ) ∩ H01 (U )
holds when the boundary ∂U is of class C 2 . This handles the regularity where
m = 0 and p = 2.

(2) Following the results of Agmon, Douglis, Nirenberg, and Geymonat, the
uniform ellipticity condition gives the regularity result for m = 0 and p ≥ 6/5.
I refer to Ciarlet [3] as the details of this analysis are technical.

(3) By (13), the weak solution u ∈ W 2,p (U ) ∩ H01 (U ) satisfies

Z Z
{λ(tr e(u))I + 2µe(u)} : e(v) dx = f v dx, v ∈ H01 (U ; Rn ). (14)
U U

32
 Evans Chapters 5 - 9

We will apply Green’s formula for Sobolev Spaces. Let ν = (νi ) be the outward
unit normal vector for ∂U . Then for i = 1, 2, . . . , n Green’s formula gives
Z Z Z
(∂i u)v dx = − u∂i v dx + uvνi dS.
U U ∂U

Applying the formula to the LHS of (14) with Dirichlet boundary conditions
yields
Z Z
{λ(tr e(u))I + 2µe(u)} : e(v) dx = − div {λ(tr e(u))I + 2µe(u)} · v dx,
U U

m,p
by which we then conclude that f ∈ W (U ).

(4) After establishing the regularity result

f ∈ W m,p (U ) =⇒ u ∈ W m+2,p (U )

for m = 0, we then apply the bootstrap argument to obtain higher regularity

provided that the boundary ∂U is of class C m+2 . Analagously to Theorem 6.3.2
in Evans, we repeatedly apply Theorem 4 for m = 0, 1, 2, . . . to deduce infinite
differentiability of u.

Chapter 7 Solutions

7.5.1 Prove that there is at most one smooth solution of this initial/boundary-
value problem for the heat equation with Neumann boundary conditions

ut − ∆u = f in UT

∂u
∂ν = 0 on ∂U × [0, T ]

u = g on U × {t = 0}.


Proof. Let u1 , u2 be solutions and define w := u1 −u2 . Then the initial/boundary-

value problem becomes

wt − ∆w = 0 in UT

∂w
∂ν = 0 on ∂U × [0, T ]

w = 0 on U × {t = 0}.


We need to show that w = 0. Define the energy to be

Z
2
E(t) := (w(x, t)) dx.
U

33
 Evans Chapters 5 - 9

Differentiating under the integral sign yields

Z
d
E(t) = 2 w(x, t) · wt (x, t) dx.
dt U

By the equation this expression equals

Z
d
E(t) = 2 w(x, t) · ∆w(x, t) dx.
dt U

Through integration by parts in the x variable, we deduce

Z Z
d 2 ∂w
E(t) = −2 (Dw(x, t)) dx + 2 w dS
dt ∂U ∂ν
ZU
2
= −2 (Dw(x, t)) dx
U
≤ 0.

Hence, E(t) is decreasing in t. In particular, since E(0) = 0 and since E(t) ≥ 0,

it follows that E(t) = 0 for all t ≥ 0. Hence, w = 0, as was claimed.

7.5.2 Assume u is a smooth solution of


ut − ∆u = 0
 in UT × (0, ∞)
u=0 on ∂U × [0, T ]

u=g on U × {t = 0}.


Prove the exponential decay estimate

ku(·, t)kL2 (U ) ≤ e−λ1 t kgkL2 (U ) (t ≥ 0),

where λ1 is the principal eigenvalue of −∆ (with zero boundary conditions) on

U.

Proof. As u is a smooth solution of the initial/boundary-value problem we in-

tegrate by parts in the x variable to obtain
Z
d 1
ku(·, t)k2L2 (U ) = uut dx
dt 2
ZU
= u∆u dx
U
Z Z
∂u
=− |Du|2 dx + u dS
U ∂U ∂ν
Z
=− |Du|2 dx.
U

34
 Evans Chapters 5 - 9

From Theorem 2 in §6.5, we know that Rayleigh’s formula is expressed as

|Du|2 dx
R
B[u, u] U
λ1 = min 2 = min 2 ,
u∈H01 (U ) kukL2 (U ) u∈H01 (U ) kukL2 (U )
u6=0 u6=0

d 1 2
where we may observe that dt 2 ku(·, t)kL2 (U ) = −B[u, u] by the weak formation.
This implies
d
ku(·, t)k2L2 (U ) ≤ −2λ1 kuk2L2 (U ) .
dt
We now apply Gronwall’s inequality to obtain the exponential decay estimate.
Letting η(t) = ku(·, t)k2L2 (U ) we see that

η 0 (t) ≤ φ(t)η(t) + ψ(t)

where φ(t) is a constant and ψ(t) is zero. This gives

Rt
η(t) ≤ e 0
φ(s) ds
η(0) = e−2λ1 t η(0).

By the initial condition u = g at t = 0, we see that

η(0) = ku(·, 0)k2L2 (U ) = kgk2L2 (U ) .

This forms
ku(·, t)k2L2 (U ) ≤ e−2λ1 t kgk2L2 (U ) ,
as desired.

7.5.5 Assume (
uk * u in L2 (0, T ; H01 (U ))
u0k * v in L2 (0, T ; H −1 (U )).

Prove that v = u0 . (Hint: Let φ ∈ Cc1 (0, T ), w ∈ H01 (U ). Then

Z T Z T
hu0k , φwi dt =− huk , φ0 wi dt.)
0 0

Proof. As in §D.4, uk * u in L2 (0, T ; H01 (U )) and u0k * v in L2 (0, T ; H −1 (U ))

means
Z T Z T
huk , hi dt → hu, hi dt for all h ∈ L2 (0, T ; H −1 (U )),
0 0
Z T Z T
hu0k , gi dt → hv, gi dt for all g ∈ L2 (0, T ; H01 (U )),
0 0

35
 Evans Chapters 5 - 9

where h , i denotes the pairing between H −1 (U ) and H01 (U ). Following the

hint, we let φ ∈ Cc1 (0, T ), w ∈ H01 (U ), and g = φ(t)w. This implies that
g ∈ L2 (0, T, H01 (U )). We then rewrite the above expressions as
Z T Z T
huk , φ0 wi dt → hu, φ0 wi dt, (15)
0 0
Z T Z T
hu0k , φwi dt → hv, φwi dt. (16)
0 0

Theorem 8 (Bochner) in §E.5 states that if f : [0, T ] → X is strongly measurable

then * Z + Z
T T
u∗ , f (t) dt = hu∗ , f (t)i dt,
0 0

for every u∗ ∈ X ∗ . We first deduce

*Z +
Z T Z T T
0 0 0
hu, φ wi dt = huφ , wi dt = u(t)φ (t) dt, w , (17)
0 0 0

and then observe

Z T *Z +
Z T T
hv, φwi dt = hvφ, wi dt = v(t)φ(t) dt, w . (18)
0 0 0

Combining (17) and (18) with the hint

Z T Z T
hu0k , φwi dt =− huk , φ0 wi dt
0 0

and taking the limit gives

*Z + * Z +
T T
0
u(t)φ (t) dt, w = − v(t)φ(t) dt, w
0 0

for every w ∈ H01 (U ). If we didn’t want to use the hint, then we could observe
by (15) and (16) that

36
 Evans Chapters 5 - 9

*Z +
T Z T
0
u(t)φ (t) dt, w = huφ0 , wi dt
0 0
Z T
= hu, φ0 wi dt
0
Z T
= lim huk , φ0 wi dt
k→∞ 0
Z T
= lim huk φ0 , wi dt
k→∞ 0
*Z +
T
0
= lim uk (t)φ (t) dt, w
k→∞ 0
* Z +
T
= lim − u0k (t)φ(t) dt, w
k→∞ 0
Z T
= lim −hu0k φ, wi dt
k→∞ 0
Z T
= lim −hu0k , φwi dt
k→∞ 0
Z T
= −hv, φwi dt
0
*Z +
T
= −v(t)φ(t) dt, w .
0

Both methods imply that

Z T Z T
0
u(t)φ (t) dt = − v(t)φ(t) dt
0 0

for every φ ∈ Cc1 (0, T ). Hence, v = u0 .

7.5.6 Suppose H is a Hilbert space and uk * u in L2 (0, T ; H). Assume further

we have the uniform bounds

ess sup kuk (t)k ≤ C (k = 1 . . .)

0≤t≤T

for some constant C. Prove

ess sup ku(t)k ≤ C.

0≤t≤T

Rb
(Hint: We have a
(v, uk (t)) dt ≤ Ckvk|b − a| for 0 ≤ a ≤ b ≤ T and v ∈ H.)

37
 Evans Chapters 5 - 9

Proof. Since uk * u in L2 (0, T ; H) we know from §D.4 that

Z b Z b
(v, uk (t)) dt → (v, u(t)) dt, for all v ∈ L2 (0, T ; H).
a a

Following the hint

Z b
(v, uk (t)) dt ≤ Ckvk|b − a|, 0 ≤ a ≤ b ≤ T, v ∈ H,
a

we let k → ∞ to deduce
Z b
(v, u(t)) dt ≤ Ckvk|b − a|.
a

Then since |(v, u(t))| ≤ kvkku(t)k by Cauchy–Schwarz we observe that taking

the supremum where kvk ≤ 1 gives
Z b Z b
sup |(v, u(t))| dt = ku(t)k dt ≤ C|b − a|.
a ||v||≤1 a

Hence ku(t)k ≤ C for all 0 ≤ a ≤ b ≤ T . This implies

ess sup ku(t)k ≤ C.

0≤t≤T

7.5.12 Prove the resolvent identities (12) and (13) in §7.4.1.

Proof. The requirement for λ to be in the resolvent set is that the operator
λI − A maps D(A) bijectively onto X. As λ, µ ∈ p(A) we see that

λI − A : D(A) → X, µI − A : D(A) → X

are both bijective operators onto X and are invertible. This implies that the
operator times its inverse is the identity. By spectral theory,

Rλ (λI − A) = (λI − A)Rλ = I,

Rµ (µI − A) = (µI − A)Rµ = I.

Hence

Rλ − Rµ = Rλ (µI − A)Rµ − Rλ (λI − A)Rµ

= Rλ ((µI − A) − (λI − A))Rµ
= Rλ (µ − λ)Rµ
= (µ − λ)Rλ Rµ ,

38
 Evans Chapters 5 - 9

which proves the first resolvent identity. The second identity follows from re-
peating the analysis and switching µ and λ. Then

Rµ − Rλ = Rµ (λI − A)Rλ − Rµ (µI − A)Rλ

= Rµ ((λI − A) − (µI − A))Rλ
= Rµ (λ − µ)Rλ
= (λ − µ)Rµ Rλ .

Therefore
Rλ − Rµ
Rλ Rµ =
µ−λ
and we have that
Rµ − Rλ Rλ − Rµ
Rµ Rλ = = = Rλ Rµ .
λ−µ µ−λ

7.5.13 Justify the equality

Z ∞ Z ∞
−λt
A e S(t)u dt = e−λt AS(t)u dt
0 0

used in (16) of §7.4.1. (Hint: Approximate the integral by a Riemann sum and
recall that A is a closed operator.)

Proof. A is a closed means that whenever uk ∈ D(A) (k = 1, . . .) and uk →

u, Auk → v as k → ∞, then

u ∈ D(A), v = Au.

So we need to make a suitable choice for uk . Following the hint in Evans, we

look at the Laplace transform of the semigroup
Z ∞
eλt S(t)u dt
0

and approximate it by Riemann sums by letting

k
X ∗
uk = e−λti S(t∗i )u∆t,
i=1

where ti−1 < t∗i

< ti represent subintervals. By Theorem 1 in §7.4 of Evans,
u ∈ D(A) implies that S(t)u ∈ D(A) for every t ≥ 0. As {S(t)}t≥0 is a one-
parameter family of linear operators, we see that uk ∈ D(A) because of the

39
 Evans Chapters 5 - 9

linearity of the operator S(t). Furthermore, A : D(A) ⊂ X → X is also a linear

operator. This implies that we can approximate the RHS of the equality
Z ∞
e−λt AS(t)u dt
0

by Riemann sums where

k
X ∗
Auk = e−λti AS(t∗i )u∆t.
i=1

Taking the limit as k → ∞ yields

k Z ∞
X ∗
lim e−λti S(t∗i )u∆t = eλt S(t)u dt,
k→∞ 0
i=1

k Z ∞
X ∗
lim e−λti AS(t∗i )u∆t = eλt AS(t)u dt.
k→∞ 0
i=1

Therefore we see that as k → ∞,

Z ∞ Z ∞
uk → eλt S(t)u dt, Auk → eλt AS(t)u dt .
0 0
| {z } | {z }
u v

Since A is closed, we know that u ∈ D(A) and v = Au. Furthermore, since A is

linear we may conclude
Z ∞ Z ∞
eλt AS(t)u dt = A eλt S(t)u dt,
0 0

as required.

7.5.14 Define for t > 0

Z
[S(t)g](x) = Φ(x − y, t)g(y) dy (x ∈ Rn ),
Rn

where g : Rn → R and Φ is the fundamental solution of the heat equation. Also

set S(0)g = g.
(a) Prove {S(t)}t≥0 is a contraction semigroup on L2 (Rn ).
(b) Show {S(t)}t≥0 is not a contraction semigroup on L∞ (Rn ).

Proof. For (a) we know from (9) in §2.3 of Evans that the fundamental solution
of the heat equation in Rn may be written as the convolution
Z
1 |x−y|2
− 4t
[S(t)g](x) = e g(y) dy
(4πt)n/2 Rn

40
 Evans Chapters 5 - 9

where x ∈ Rn , t > 0, g ∈ X for X := Lp (Rn ), 1 ≤ p < ∞. By defining

2
Jt (x) := (4πt)−n/2 e−|x| /4t
,

we see that we can write

[S(t)g](x) = Jt ∗ g(x).

A function g : Rn → R is said to be rapidly decreasing if it is infinitely many

times differentiable (g ∈ C ∞ (Rn )) and

lim |x|k Dα g(x) = 0 for all k ∈ N and α ∈ Nn .

|x|→∞

The space

S (Rn ) := {g ∈ C ∞ (Rn ) : f is rapidly decreasing}

is called the Schwartz space. The function Jt : Rn → R given by

2
Jt (x) := (4πt)−n/2 e−|x| /4t
, x ∈ Rn ,

belongs to S (Rn ) for every t > 0 ([4], §2.13).

To show that S(t) is a contraction semigroup on L2 (Rn ), we first let 1 ≤ p < ∞

and observe that the integral defining [S(t)g](x) for every Lp (Rn ) exists because
Jt ∈ S (Rn ). Furthermore,

kS(t)gkLp ≤ kJt kL1 · kgkLp ≤ kgkLp

by Young’s inequality (for p = 2, we have 11 + 21 = 12 + 1). Therefore, every S(t)

is a contraction on Lp (Rn ). The remaining properties (4) − (6) in §7.4 of Evans
follow from the Gaussian kernel Jt defined above. These are

1. Jt > 0
2. Js ∗ Jt = Js+t
R
3. Rn Jt (x) dx = 1

The first property is clear from the definition of Jt . The second property follows
from
Z Z
1 |x−y|2 1 |x−z|2
− 4(s+t) − 4t
n/2
e g(y) dy = n/2
e ×
Rn (4π(s + t)) n (4πt)
Z R
1 |z−y|2

n/2
e− 4s g(y) dydz
Rn (4πs)

where
Z
1 |x−y|2 1 1 |x−z|2 |z−y|2
− 4(s+t)
e = e− 4t e− 4s dz.
(4π(s + t))n/2 (4πt)n/2 (4πs)n/2 Rn

41
 Evans Chapters 5 - 9

The third property is shown in Evans in the Lemma on page 46.

The first property implies that if g ≥ 0 then S(t)g = Jt ∗ g > 0 and (4) from
Evans is verified because S(0)g(x) = g(x). The second property implies that
Ss St = Ss+t and therefore (5) from Evans is verified. The third property implies
that if g ∈ L2 ∩ Lp , where Lp is the space of functions for which p ∈ [1, ∞] then

kS(t)gkLp = kJt ∗ gkLp ≤ kJt kL1 · kgkLp ≤ kgkLp

analogously as before. These properties imply that the operators {S(t)}t≥0 form
a semigroup where Ss St = Ss+t . The first property implies that the semigroup
is positive and S(t) maps positive functions into positive functions. In fact, it
maps positive functions into strictly positive functions. To verify (6) in Evans we
show that S(t)g → g in X = Lp (Rn ) as t → 0+ if g ∈ Lp (Rn ). For g ∈ Lp (Rn )
we find that
Z Z p
kS(t)g − gkpLp =
 

 Jt (y)g(x − y) dy − g(x) dx

Rn Rn
Z Z
 h i p
=  n t
 J (y) g(x − y) − g(x) dy  dx
Rn R
Z Z
 h √ i p
=  n J1 (v) g(x − tv) − g(x) dv  dx
 
Rn R
Z Z  √ p
≤ J1 (v)g(x − tv) − g(x) dv dx
 
n n
ZR R Z  √ p
= J1 (v) g(x − tv) − g(x) dx dv
 
Rn Rn

because the integral of Jt is one. This in combination with a change of variable

allows us to put g(x) under the integral sign and deduce
Z
 h √ i p Z  √ p
J (v) g(x − tv) − g(x) dv  ≤ J (v) − tv) − g(x) dv
 
1 1
 g(x
 
Rn Rn
R  √ p
by Hölder’s inequality. The function ψ(v, t) := Rn g(x − tv) − g(x) dx goes
to zero as t → 0+ for every v (see Rudin [7]) by a well known property of Lp
functions. The function is also bounded above by 2p kgkpLp therefore we can
apply the dominated convergence theorem to conclude that kS(t)g − gkpLp → 0
as t → 0+ .
The third property implies that the semigroup of operators S can be extended
from L2 ∩ Lp to Lp and that the extensions are contractive on every Lp -space,

kS(t)kLp →Lp = sup{kS(t)gkLp : kgkLp ≤ 1} ≤ 1

for every p ∈ [1, ∞]. Therefore {S(t)}t≥0 is a contraction semigroup on L2 (Rn ).

It is interesting to observe that property (6) in §7.4 of Evans is satisfied for
1 ≤ p < ∞ but fails when p = ∞ because the map isn’t continuous at t = 0.

42
 Evans Chapters 5 - 9

For (b), we may observe that {S(t)}t≥0 is not a contraction semigroup on

L∞ (Rn ) because it doesn’t satisfy property (6) in §7.4 of Evans which states
that for every u ∈ X, the mapping t → S(t)u is continuous from [0, ∞) into X.
For n = 1, we let u(x) be the characteristic function that is 0 when x ≥ 0 and
1 when x < 0. The integral becomes for every t > 0
Z x
1 2
[S(t)u](x) = √ e−y /4t dy,
2 πt −∞
√
which is continuous. Substituting s = y/ t and letting x approach zero gives
Z 0
1 2 1
[S(t)u](0) = √ e−s /4 ds = .
2 π −∞ 2
However, the distance between S(t)u and u remains positive for all t > 0
kS(t)u − ukL∞ > 0,
so the map isn’t continuous at t = 0. Therefore {S(t)}t≥0 is not a contraction
semigroup on L∞ (Rn ).

7.5.15 Let {S(t)}t≥0 be a contraction semigroup on X, with generator A.

Inductively define D(Ak ) := {u ∈ D(Ak−1 ) | Ak−1 u ∈ D(A)} (k = 2, . . .).
Show that if u ∈ D(Ak ) for some k, then S(t)u ∈ D(Ak ) for each t ≥ 0.

Proof. We first recall the proof of (i) and (ii) in Theorem 1 of §7.4 of Evans.
Let u ∈ D(A). Then
S(s)S(t)u − S(t)u S(t)S(s)u − S(t)u
lim = lim +
s−>0+ s s−>0 s
S(s)u − u
= S(t) lim +
s−>0 s
= S(t)Au.
Thus S(t)u ∈ D(A) and AS(t)u = S(t)Au. We will now assume that u ∈ D(Ak )
and show that S(t)u ∈ D(Ak ). By the induction hypothesis, we have D(Ak ) :=
{u ∈ D(Ak−1 ) | Ak−1 u ∈ D(A)} where k ≥ 2. The base case for k = 1 is
handled through Theorem 1. Let n = k − 1 so that the difference quotient
becomes
S(s)S(t)An u − S(t)An u S(t)S(s)An u − S(t)An u
lim = lim
s−>0+ s s−>0+ s
S(s)An u − An u
= S(t) lim + .
s−>0 s
This shows that S(t)An u ∈ D(A). However, in order to prove that S(t)u ∈
D(Ak ), we need to show that An S(t)u ∈ D(A). Hence we apply part (ii) of
Theorem 1 to see that S(t)An u = An S(t)u. Thus S(t)u ∈ D(Ak ).

43
 Evans Chapters 5 - 9

7.5.16 Use Problem 15 to prove that if u is the semigroup solution in X =

L2 (U ) of 
ut − ∆u = 0
 in UT
u=0 on ∂U × [0, T ]

u=g on U × {t = 0},


with g ∈ Cc∞ (U ), then u(·, t) ∈ C ∞ (U ), for each 0 ≤ t ≤ T .

Proof. To get a representation formula for the solution, we apply the Fourier
transform. We denote û(ξ, t) as the Fourier transform of u with respect to the
space variable x. This gives

2
 ût = −|ξ| û in ÛT

û = 0 on ∂ Û × [0, T ]

û = ĝ on Û × {t = 0},


2
whose solution is û(ξ, t) = ĝ(ξ)e−|ξ| t . Taking the inverse Fourier transform, we
get u = S(·)g, where the heat semigroup {S(t)}t≥0 is defined by
Z
1 |x−y|2
− 4t
(S(t)g)(x) = e g(y) dy, t > 0, x ∈ Rn .
(4πt)n/2 Rn

Here, (S(0)g)(x) = g(x) and the remainder of the semigroup properties for
(S(t))t≥0 follow from the analysis in 7.5.14. As before, S(t)g = Jt ∗ f where
Z
1 −|x|2 /4t
Jt (x) = e , Jt (x) dx = 1, t > 0,
(4πt)n/2 Rn

and ∗ denotes convolution. By Young’s inequality,

kS(t)gkL2 ≤ kgkL2 , t > 0, p = 2.

Therefore (S(t))t≥0 is a contraction semigroup. By the previous exercise, S(t)u ∈

X = L2 (U ). Furthermore, since ut = ∆u we see that the generator A is ∆. Ex-
ercise 15 then implies that if u ∈ D(∆k ) where k ∈ N then S(t)u ∈ D(∆k ). We
first show that g ∈ D(Ak ) = D(∆k ).

Following Brezis ([2], §10.1) we let u ∈ D(A) = H01 (U ) ∩ H 2 (U ). Then we see

that D(A2 ) = {u ∈ H01 (U ) ∩ H 2 (U ) | ∆u ∈ H01 (U ) ∩ H 2 (U )} and D(Ak ) =
{u ∈ D(Ak−1 ) | Ak−1 u ∈ D(A)} = {u ∈ H01 (U ) ∩ H 2 (U ) | ∆k−1 u ∈ H01 (U ) ∩
H 2 (U )} by the previous exercise. We will show that g ∈ D(∆k ) by applying the
induction defined in problem 15. For the base step, we know that g ∈ Cc∞ (U ) ⊆
Cc∞ (U ) = H01 (U ). Furthermore, the compact support implies that the L2 -norm
of g and its derivatives are finite and therefore g ∈ H 2 (U ). This shows that
g ∈ D(A) = D(∆). We then see that at order k, g ∈ D(Ak ) = D(∆k ) because
if g ∈ D(∆k−1 ) then Ak−1 g = ∆k−1 g ∈ Cc∞ (U ) ⊆ H01 (U ) and once again the
L2 -norm of g and its derivatives are finite hence g ∈ H 2 (U ).

44
 Evans Chapters 5 - 9

Since g ∈ D(∆k ), we know that u(·, t) = S(t)g ∈ D(∆k ) for all finite time
intervals. It remains that show that D(∆k ) ⊆ H 2k (U ) for all k ∈ N. This
follows from the observation that

u ∈ D(Ak ) =⇒ ∆k−1 u ∈ H01 (U ) ∩ H 2 (U ) =⇒ u ∈ H01 (U ) ∩ H 2k (U ).

We then apply Theorem 4 (Morrey’s inequality) and Theorem 6 in §5.6.2 and

§5.6.3 of Evans to deduce the embedding into C ∞ (U ). This gives the desired
embedding into C ∞ (U ) through the use of Sobolev embedding.

Strong Maximum Principle Exercise Prove the following strong maximum

principle for weak sub-solutions of the given elliptic PDE:
Theorem 5. Let U ⊆ Rn be open, bounded, and connected, and let A : U →
Symn×n have
λI ≤ A(x) ≤ ΛI a.e. in U
where λ, Λ ∈ (0, ∞). Let u ∈ H 1 (u) satisfy
Z
hA(x)Du, Dvi ≤ 0
U

for all non-negative v ∈ H01 (U ). If for some compactly contained ball BR (x0 ) ⊂⊂
U we have that
ess sup u = ess sup u,
BR (x0 ) U

then u must be constant a.e. in U .

Hint: Apply Moser’s weak Harnack inequality. We rely on the following two
lemmas.
Lemma 6. Suppose u ∈ H 1 (U ) is a weak subsolution of L, i.e., Lu ≥ 0 and
Z
hA(x)Du, Dvi ≤ 0,
U

where
λI ≤ A(x) ≤ ΛI a.e. in U
for λ, Λ ∈ (0, ∞). Then u is bounded above on any U0 ⊂⊂ U . Therefore, if u is
a weak solution of Lu = 0, it is bounded above and below on any U0 satisfying
these conditions.
Lemma 7. Let u be a positive supersolution in B4r0 (x0 ) ⊂ Rn . For 0 < p <
n
n−2 , and if n ≥ 3 then

Z !1/p
p C
− u dx ≤ 2 ess inf u,
n
B2r0 (x0 )
n−2 − p Br0 (x0 )

45
 Evans Chapters 5 - 9

where C = C(n, Λλ ). In the case n = 2, the estimate is true provided 0 < p < ∞
and the constant C depends on p and Λ C
λ in place of ( n −p)2 .
n−2

Proof. As there is some compactly contained ball BR (x0 ) ⊂⊂ U such that

ess sup u = ess sup u

BR (x0 ) U

it must also be true that there is another ball Br0 (y0 ) with B4r0 (y0 ) ⊂ U such
that
ess sup u = ess sup u. (19)
Br0 (y0 ) U

Moreover, by Lemma 6 we see that we can take ess sup u to be finite since
U
we know that ess sup u < ∞. Define M to be a positive number such that
BR (x0 )
M > ess sup u. Then M − u is a positive supersolution and we can apply our
U
second lemma due to Moser. As M − u is a positive supersolution, we can apply
Lemma 7 to it. Passing to the limit, we see that the inequality holds for

M = ess sup u. (20)

U

Therefore, we set p = 1 in Lemma 7 and recall ess inf(−u) = −ess sup(u) to

deduce Z
− (M − u) dx ≤ C̃ ess inf (M − u) = 0,
B2r0 (y0 ) Br0 (y0 )

because of (19) and (20). As M is equal to the supremum of u over the domain,
we also know that u ≤ M . Therefore we have that

u=M (21)

a.e. on B2r0 (y0 ). Given that we have found that u is constant on a ball of
radius 2r0 , we now extend this result to the entire domain. Let y ∈ U be
arbitrary. Then there is a sequence of balls Bi := Bri (yi ) for i = 0, . . . , n such
that B4ri (yi ) ⊂ U and Bi−1 ∩ Bi 6= ∅ for i = 1, . . . , n. Furthermore, y ∈ Bn .
Since B0 ∩ B1 6= ∅ and we previously showed that u = M a.e. on B2r0 (y0 ), we
have
ess sup u = M.
B1

By the same reasoning as before, we find that u = M a.e. on the ball B2r1 (y1 ).
We then iterate this process over every ball to obtain

u=M

a.e. on B2rn (yn ). As y ∈ Brn (yn ), we see that u(y) = M . Furthermore, since
y ∈ U was arbitrary, we have that u ≡ M a.e. on all of U .

46
 Evans Chapters 5 - 9

Chapter 8 Solutions

8.7.1 This problem illustrates that a weakly convergent subsequence can be

rather badly behaved.
(a) Prove uk (x) = sin(kx) * 0 as k → ∞ in L2 (0, 1).
(b) Fix a, b ∈ R, 0 < λ < 1. Define
(
a if j/k ≤ x < (j + λ)/k
uk (x) := (j = 0, . . . , k − 1).
b if (j + λ)/k ≤ x < (j + 1)/k

Prove uk * u := λa + (1 − λ)b in L2 (0, 1).

Proof. For (a), we apply the Riemann–Lebesgue lemma which states that the
Fourier transform of an L1 function vanishes at infinity. As sin(kx) = (eikx −
e−ikx )/(2i), we observe that for any f ∈ L2 (0, 1), f χ[0,1] ∈ L1 (R), and therefore
1
eikx − e−ikx
Z Z
lim sin(kx)f dx = lim f χ[0,1] dx = 0.
k→∞ 0 k→∞ R 2i

Hence, uk (x) = sin(kx) * 0 in L2 (0, 1).

For (b), we apply the definition of weak convergence. We need to prove that
Z 1 Z 1
lim uk (x)g(x) dx = u(x)g(x) dx
k→∞ 0 0

for every g ∈ L2 (0, 1). It suffices to prove this in g ∈ C0∞ (0, 1) because these
functions are dense in L2 (0, 1). Then
k−1 j+1
Z 1 XZ k
uk (x)g(x) dx = uk (x)g(x) dx
j
0 j=0 k

j+λ j+1
k−1 Z Z !
X k k
= ag(x) dx + bg(x) dx .
j j+λ
j=0 k k

As g(x) is continuous, we have that for every j ∈ {0, . . . , k − 1}, there exist µj
and νj in [j/k, (j + 1)/k] such that
Z j+λ
k λ
ag(x) dx = a g(µj ),
j
k
k
j+1
1−λ
Z k
bg(x) dx = b g(νj ),
j+λ
k
k

47
 Evans Chapters 5 - 9

as a consequence of the mean value theorem and the continuity of g(x). There-
fore
Z 1 k−1 k−1
1X 1X
uk (x)g(x) dx = aλ g(µj ) + b(1 − λ) g(νj ).
0 k j=0 k j=0
R1
However, both of these are Riemann sums which converge to 0 g(x) dx. We
pass to the limit to deduce
Z 1 Z 1 Z 1
lim uk (x)g(x) dx = aλ g(x) dx + b(1 − λ) g(x) dx
k→∞ 0 0 0
 Z 1
= aλ + b(1 − λ) g(x) dx
0

and by the density of g ∈ C0∞ (0, 1) in L2 (0, 1) we have uk * u := λa + (1 − λ)b.

Remark: An alternative way to analyze 8.7.1(b) is to define
(
(1 − λ)(a − b) if j/k ≤ x < (j + λ)/k
vk (x) := uk (x)−[λa + (1 − λ)b] =
λ(b − a) if (j + λ)/k ≤ x < (j + 1)/k.

Then Z (j+1)/k
vk (x) dx = 0.
j/k
R
So by defining the characteristic function g = χ[a,b] , we see that gvk dx → 0
(because the rationals are dense). The same is true for a step function, as
these are a linear combination of characteristic functions of an interval. As step
functions are dense in L2 , we see that this would also imply that uk * u :=
λa + (1 − λ)b.

8.7.2 Find L = L(p, z, x) so that the PDE

−∆u + Dφ · Du = f in U
R
is the Euler-Lagrange equation corresponding to the fuctional I[w] := U
L(Dw, w, x) dx.
(Hint: Look for a Lagrangian with an exponential term.)

Proof. We are given that L(p, z, x) = L(Du, u, x) and by the hint we consider
the exponential term e−φ(x) . Let
 
1 2
L(p, z, x) = e−φ(x) |p| − zf (x) .
2

48
 Evans Chapters 5 - 9

Then  
1
L(Du, u, x) = e−φ(x) |Du|2 − uf (x)
2
and the functional is
Z  
1
I[u] = e−φ(x) |Du|2 − uf (x) dx.
U 2
Therefore we can minimize the function by evaluating the derivative at time τ
Z
0 d
I [u] = I[u] = e−φ(x) (Du · Dũ − ũf (x)) dx
dτ U
d
where ũ = dτ u. As in the derivation shown in §8.1.2, evaluating the derivative
at τ = 0 will minimize the function. We compute
Z
I 0 [u] = e−φ(x) (Du · Dũ − ũf (x)) dx
ZU
= e−φ(x) Du · Dũ − ũe−φ(x) f (x) dx
U
Z
= ũe−φ(x) Dφ · Du − ũe−φ(x) ∆u − ũe−φ(x) f (x) dx
U
Z
= ũe−φ(x) (Dφ · Du − ∆u − f (x)) dx
U
= 0.
Hence,
−∆u + Dφ · Du = f
as desired.

8.7.4 Assume η : Rn → R is C 1 .
(a) Show L(P, z, x) = η(z) det P (P ∈ Mn×m , z ∈ Rn ) is a null Lagrangian.

(b) Deduce that if u : Rn → Rn is C 2 , then

Z
η(u) det Du dx
U

depends only on u|∂U .

Proof. For (a), we apply two identities from the Lemma (Divergence-free rows)
on page 464 of Evans:
(det P )I = P T (cof P ) (22)
and
n
X
(det P )δij = pki (cof P )kj , (i, j = 1, . . . , n). (23)
k=1

49
 Evans Chapters 5 - 9

Then since L(P, z, x) = η(z) det P , we see that

k
Lpki (Du, u, x) = η(u)(cof(Du))i (i, k = 1, . . . , n)

and then compute the first term in the Euler-Lagrange equations. (22) and (23)
forms
n 
X  n X
X n
− Lpki (Du, u, x) =− (ηzj (u)) ujxi (cof Du)ki
xi
i=1 i=1 j=1
Xn
+ η(u) (cof Du)ki,xi
i=1
n X
X n
=− (ηzj (u)) ujxi (cof Du)ki
i=1 j=1
n
X
=− ηzj (u)δjk det(Du)
j=1
Pn Pn
because det(Du)δjk = i=1 pji (cof Du)ki and j=1 (cof Du)kj,xj = 0. The sec-
ond term in the Euler-Lagrange equations becomes
Lzk (Du, u, x) = ηzk (u) det(Du).
Therefore since determinants are null Lagrangians,
n 
X 
− Lpki (Du, u, x) + Lzk (Du, u, x)
xi
i=1
n
X
=− ηzj (u)δjk det(Du) + ηzk (u) det(Du)
j=1

= 0.
Hence, L(P, z, x) = η(z) det P is a null Lagrangian.

For (b), we apply Theorem 1 (Null Lagrangians and boundary conditions) on

page 463 of Evans. We know that null Lagrangians are significant because their
corresponding energy depends only on the boundary conditions. As u is C 2 and
we showed that L(P, z, x) = η(z) det P or L(Du, u, x) = η(u) det(Du) is a null
Lagragian in (a), all of the conditions of Theorem 1 are satisfied.

This implies that if there is another C 2 function ũ such that u ≡ ũ on ∂U then

I[u] = I[Ũ ]
thus Z Z
L(Du, u, x) dx = L(Dũ, ũ, x) dx.
U U
R
So we can find U
η(u) det Du dx by the value of u as its boundary, u|∂U .

50
 Evans Chapters 5 - 9

8.7.7 Let m = n. Prove

L(P ) = tr(P 2 ) − tr(P )2 (P ∈ Mm×n )

is a null Lagrangian.

Proof. We will match the notation on page 462 of Evans and represent P ∈
Mm×n by  1
p1 . . . p1n


P =
 .. 
. 
m m
p1 . . . pn m×n
Pn
where P = (pki ) for i rows, k columns, and tr(A) = i=1 aii . Expanding L(P )
gives
n n
!2 n
X X X
2 2 k i i
tr(P ) − tr(P ) = pi pk − pi = pki pik − pii pkk .
i,k=1 i=1 i,k=1

We then see that the RHS consists of a sum of subdeterminants of the matrix
 i
pi pik


pki pkk

obtained from evaluating the larger matrix P ∈ Mm×n . We need to show that
these subdeterminants are null Lagrangian. To do this, we first observe that if
i = k then we are done. On the other hand, if i 6= k then it is easier to apply
the notation shown in lecture. Observing that p is represented by Du and that
we are taking the derivative with respect to subscript gives
 
∂i ui ∂k ui
.
∂i uk ∂k uk
The determinant of the above matrix is
 
∂i ui ∂k ui
det = ∂i ui ∂k uk − ∂i uk ∂k ui .
∂i uk ∂k uk
Therefore
W (F ) = Fii Fkk − Fik Fki
where
∂W ∂W
= Fkk , = −Fik ,
∂Fii ∂Fki
∂W ∂W
= −Fki , = Fii .
∂Fik ∂Fkk
The Euler-Lagrange equations are
 
∂ ∂W
(Du) = ∂i (∂k uk ) − ∂k (∂i uk ) = (∂ik − ∂ki )uk = 0
∂xj ∂F ij

51
 Evans Chapters 5 - 9

and  
∂ ∂W
(Du) = ∂k (∂i ui ) − ∂i (∂k ui ) = (∂ki − ∂ik )ui = 0.
∂xj ∂F kj
clear that all of the 2 × 2 subdeterminants are null Lagrangian
Therefore, it isP
n
and −L(P ) = i,k=1 ∂i ui ∂k uk − ∂i uk ∂k ui as a sum of subdeterminants is null
Lagrangian.
Alternatively we can consider

L̃ = Lki = pki pik − pii pkk

and then compute the Euler-Lagrange equations abstractly through the simpli-
fied notation on page 456 and 463. Defining

Dpi L(Du) = The gradient of L(Du) with respect to the index row i

we find that the Euler-Lagrange equations (17) on page 463 may be written as
the divergence

−∇ · Dpi L̃(Du) + Dzi L̃(Du) = −(uk )xk xi + (uk )xk xi = 0

−∇ · Dpk L̃(Du) + Dzk L̃(Du) = −(ui )xk xi + (ui )xk xi = 0.

Either method shows that all of the 2 × 2 subdeterminants are null Lagrangian
and therefore L(P ) is null Lagrangian.

8.7.15 (Pointwise gradient constraint)

(a) Show that there exists a unique minimizer u ∈ A of
Z
1
I[w] := |Dw|2 − f w dx,
U 2

where f ∈ L2 (U ) and

A := {w ∈ H01 (U ) | |Dw| ≤ 1 a.e.}.

(b) Prove Z Z
Du · D(w − u) dx ≥ f (w − u) dx
U U
for all w ∈ A.

Proof. We follow the proofs of Theorem 3 and Theorem 4 in §8.4.2 of Evans.

The difference in this problem is that the admissible set now has the obstacle
of |Dw| ≤ 1 a.e. in U instead of w ≥ h a.e. in U .

52
 Evans Chapters 5 - 9

For (a), the existence of a minimizer follows from choosing a minimizing se-
quence {uk }∞
k=1 ⊂ A with

I[uk ] → m = inf I[w].

w∈A

We then extract a subsequence

ukj * u weakly in H01 (U )

with I[u] ≤ m. We will be done if we can show

|Du| ≤ 1 a.e.,

so that u ∈ A. Here, it is clear that A is convex and closed. By Mazur’s

Theorem, A is weakly closed. So by compactness ukj → u in L2 (U ). Since
Dukj ≤ 1 a.e., it follows that Du ≤ 1 a.e. and therefore u ∈ A.
Remark: It isn’t necessary to apply Mazur’s Theorem here. Since we know that
I(w) is weakly lower semi-continuous and coercive and that A is convex and
closed, we can apply §5.7 (the Rellich-Kondrachov Theorem) to show that there
exists a minimizer u ∈ A.
To show uniqueness, we assume that u and ũ ∈ A are two minimizers where
u 6= ũ. Then w := u+ũ
2 ∈ A and
Z  2  
1  Du + Dũ  u + ũ
I[w] =  −f dx
U 2 2 2

Z  
1 2 2
 u + ũ
= |Du| + 2Du · Dũ + |Dũ| − f dx.
U 8 2

Since 2a · b = |a|2 + |b|2 − |a − b|2 , we find that

Z  
1  u + ũ
I[w] = 2|Du|2 + 2|Dũ|2 − |Du − Dũ|2 − f dx
U 8 2
Z Z
1 1 1 1
< |Du|2 − f u dx + |Dũ|2 − f ũ dx
2 U 2 2 U 2
1 1
= I[u] + I[ũ],
2 2
the strict inequality following from u 6≡ ũ. This is a direct contradiction because
we assumed that u and ũ are minimizers. Therefore, there exists a unique
function u ∈ A satisfying I[u] = min I[w].
w∈A

For (b), we fix an arbitrary w ∈ A. Since A is convex we know that for every
0 ≤ τ ≤ 1,
u + τ (w − u) = (1 − τ )u + τ w ∈ A.
Hence by defining
i(τ ) := I[u + τ (w − u)],

53
 Evans Chapters 5 - 9

we have that i(0) ≤ i(τ ) for every 0 ≤ τ ≤ 1. Therefore

i0 (0) ≥ 0. (24)

Then if 0 < τ ≤ 1,

i(τ ) − i(0) |Du + τ D(w − u)|2 − |Du|2

Z
1

= − f u + τ (w − u) − u dx
τ τ U 2
τ |D(w − u)|2
Z
= Du · D(w − u) + − f (w − u) dx.
U 2

Hence (24) implies

Z
0
0 ≤ i (0) = Du · D(w − u) − f (w − u) dx,
U

which gives the desired inequality.

We may also define v := w − u so that the inequality becomes
Z Z
Du · Dv dx ≥ f v dx
U U

which is an analogue of the Euler-Lagrange equation in terms of an equality.

This inequality is associated to I(w) where w ∈ A and v = w − u.

Chapter 9 Solutions

9.7.11 Give a simple example showing that the flow

u0 ∈ −∂I[u] (t > 0) (*)

may be irreversible. (That is, find a Hilbert space H and a convex, proper, lower
semicontinuous function I : H → (−∞, ∞] such that the semigroup solution of
(∗) satisfies
S(t)u = S(t)û
for some t > 0 and u 6= û.)

Proof. We consider a piecewise defined lower semi-continuous function defined

by breaking up the real line into two convex functions. Let

I : R → (−∞, ∞]

be defined by (
|u| if |u| ≤ 1
I(u) :=
u2 if |u| > 1.

54
 Evans Chapters 5 - 9

Figure 2: The piecewise defined lower semi-continuous function I(u). The slope
of I(u) is ±1 when |u| ≤ 1 and ±2 when |u| > 1.

where I is convex, proper, continuous, and lower semicontinuous. We then

compute ∂I[u]. By construction, I is smooth away from u = 0 and u = ±1, so
the subdifferential is the derivative I 0 (u). At u = 0, lines with slopes of 1 and
−1 lie below the graph of I. At u = 1, lines with slopes between 1 and 2 lie
below the graph of I. Also, at u = −1, lines with slopes between −2 and −1 lie
below the graph of I. Therefore


 {2u} if u < −1
[−2, −1] u = −1




 if
{−1} if −1 < u < 0



∂I(u) := [−1, 1] if u=0

{1} if 0<u<1








 [1, 2] if u=1
{2u}

if u > 1.

By the theory of gradient flows in Evans, we know that the gradient flow
(
u0 ∈ −∂I[u] (t > 0)
u(0) = u0

has a unique continuous solution u(t) = S(t)u0 . We investigate the solution

when u0 > 1.
When u > 1, ∂I[u] = {2u}, so we have that u0 = −2u and therefore u(t) =
u0 e−2t . This is true until u(t) reaches u = 1, which happens when 1 = u0 e−2t
or t = (1/2) ln(u0 ). We now move to u ∈ (0, 1). Here, u0 = −1 and therefore
u(t) = −t + u1 . At u = 1, we have that 1 = −t + u1 = −(1/2) ln(u0 ) so that
u1 = 1 + (1/2) ln(u0 ). Hence u(t) = 1 − (t − (1/2) ln(u0 )). This is true until u(t)
reaches 0, where t = 1 + (1/2) ln(u0 ). For u ≤ 0, we have that u(t) ≡ constant

55
 Evans Chapters 5 - 9

by construction. Combining all of these cases


−2t
u0 e
 if 0 ≤ t ≤ (1/2) ln(u0 )
u(t) = 1 − (t − (1/2) ln(u0 ) if (1/2) ln(u0 ) ≤ t ≤ (1/2) ln(u0 ) + 1

0 if t ≥ (1/2) ln(u0 ) + 1.


Therefore we can choose two initial conditions u0 and û0 and then go forward
in time t. Eventually there will be a time t > 0 such that S(t)u0 = S(t)û0 = 0.
Thus, the flow is irreversible.

9.7.12 Let u = u(x, t) denote the height at x ∈ R2 of a sandpile that grows as

sand is added at a rate f = f (x, t) ≥ 0. We assume the stability condition

|Du| ≤ 1,

meaning that nowhere can the sandpile have slope greater than 1. As usual,
Dx u = Du. We propose the dynamics

ut − div(ADu) = f in R2 × (0, ∞),

where a = a(x, t) ≥ 0 describes the flow rate of sand rolling downhill, that is,
in the direction −Du. Suppose that

spt a ⊆ {|Du| = 1},

so that the sand flows downhill only if the slope is one.

Show that the foregoing implies

f − ut ∈ ∂I[u],

for the convex function

(
0 if u ∈ L2 (R2 ), |Du| ≤ 1 a.e.
I(u) :=
∞ otherwise.

Proof. We follow the analysis done by Aronsson, Evans, and Wu in [1] and [5].
We first proceed informally and analyze the physical aspects of the model.
Consider the non-linear p-Laplacian

∆p u = div(|Du|p−2 Du)

which for large values of p is known as a prototype of a slow/fast diffusion

operator. We are interested in understanding what happens in the fast and
slow diffusion regions in the limit as p → ∞. This problem is related to existing

56
 Evans Chapters 5 - 9

techniques known for Monge–Kantorovich type optimal mass transfer problems.

We introduce some standard notation from convex analysis. First, we define
K := {u ∈ L2 (Rn ) : |Du| ≤ 1 a.e}
and (
0 if u ∈ K
I(u) :=
∞ otherwise.
Here u∗ minimizes K[·] over K. Its corresponding Euler-Lagrange equation is
f + − f − ∈ I[u∗ ] (25)
which is defined as
I[v] ≥ I[u∗ ] + (f + − f − , v − u∗ )L2
for every v ∈ L2 (Rn ). We first rewrite (25) in a more familiar form.

Lemma 8. Suppose that f + and f − are Lipschitz continuous.

(i) Then there exists a nonnegative L∞ function a such that
− div(aDu∗ ) = f + − f − in Rn . (26)
(ii) In addition
|Du| = 1 a.e. on the set {a > 0}.
Here, a is known as the transport density. The PDE (26) is nonlinear and a is
a Lagrange multiplier corresponding to the constraint |Du∗ | ≤ 1.

Sketch of Proof: Let n + 1 ≤ p < ∞. We then approximate by the quasilinear

PDE defined for the p-Laplacian
− div(|Dup |p−2 Dup ) = f + − f − , (27)
which corresponds to maximizing
Z

1
Kp [u] := u f + − f − − |Du|p dz.
Rn p
By the maximum principle, we find that
sup |up |, |Dup |, |Dup |p ≤ C < ∞
p

for some constant C ∈ (0, ∞). It then follows that there exists a sequence
pk → ∞ such that

∗
upk → u
 locally uniformly
Dupk → Du∗ boundedly, a.e.
* a weakly ∗ in L∞ .
 p−2
|Dupk |


57
 Evans Chapters 5 - 9

Therefore, we can pass to the limit in (27) to obtain (26).

For this problem, we have a sandpile model that evolves over time. In a physical
context, it is a Monge–Kantorovich mass transfer mechanism which interacts on
a fast time scale. As u is the height of the sandpile, we impose the constraint

|Du| ≤ 1 (28)

everywhere. The constraint has the physical interpretation that the the slope
will not remain in equilibrium and that the sand will flow downhill when the
slope is one. Suppose f ≥ 0 is a source term which represents the rate at
which sand is added to the sandpile whose initial height is zero. This would
initially suggest that the model would be ut = f in any region where the con-
straint (28) is active. However, adding more sand in a particular location would
break the constraint and it would be perfectly reasonable to expect that these
newly added sand particles would instantaneously roll downhill. The particles
would continue rolling downhill until stopping at a rest state where adding new
particles maintains the constraint.

This gives the following evolutionary model for growing sandpiles

(
f − ut = ∂I[u] (t > 0)
u=0 (t = 0),

where I[·] is defined by

I[v] ≥ I[u∗ ] + (f + − f − , v − u∗ )L2

for every v ∈ L2 (Rn ). Applying Lemma 8 we see that f + = f and f − = ut

so that ut − div(ADu) = f . The physical interpretation is that for every
moment of time, the mass dµ+ = f + (·, t)dx is instantly and optimally trans-
ported downhill by the potential u(·, t) into the mass dµ− = ut (·, t)dy. There-
fore, the height u(·, t) of the sandpile is also the potential which generates the
Monge–Kantorovich reallocation of f + dx to ut dy. This relation forces the dy-
namics of the growing sandpile model.
Now, we analyze why the dynamics give the evolution equation. Let us consider
evolutions governed by the p-Laplacian:
(
up,t − ∆p up = fp in Rn × (0, ∞)
up = g on Rn × {t = 0}

and study the behavior of the solutions up as p → ∞. Here f ≥ 0 represents

a given source term, which we interpret physically as adding material to an
evolving system. As p → ∞, we will see that up → p and that the limit u solves
the evolution equation
f − ut = ∂I[u], t > 0.

58
 Evans Chapters 5 - 9

We start by sending p → ∞. Then for every T > 0 we have that the functions
{up }p≥n+1 are bounded in L∞ (Rn × [0, T ]) and {Dup , up,t }p≥n+1 are bounded
in L2 (Rn × [0, T ]). Further, the functions {up }p≥n+1 have a uniform bound
supported in Rn × [0, T ]. Therefore we can extract a sequence pi → ∞ and a
limit u so that for every T > 0
(
upi → u a.e. and in L2 (Rn × (0, T ))
(29)
Dupi * Du, upi,t * ut weakly in L2 (Rn × (0, T )).

We then reinterpret the evolution PDE as

fp − up,t = ∂Ip [up ], t > 0,

where we will send p → ∞ to recover the original PDE. We make some explicit
assumptions regarding the source terms {fp }p≥n+1 . For this, we select m distinct
points {dk }m n
k=1 ⊂ R and m smooth, nonnegative functions of time {fk (t)}k=1 .
m

This will imply that the measure

m
X
f= fk (t)δdk (x)
k=1

will record point sources at the sites {dk }m m

k=1 with rates {fk (t)}k=1 . For every
k n
k = 1, . . . , m and n + 1 ≤ p ≤ ∞ we let dp : R → R be smooth functions
satisfying (
supp(dpk ) ⊂ B(dk , rp ), dpk ≥ 0,
dp dx = 1
R
B(dk ,rp ) k
as p → ∞
where rp → 0 as p → ∞. This is why we require spt a ⊆ {|Du| = 1}. We then
take
m
X
fp (x, t) = fk (t)dpk (x) (n + 1 ≤ p ≤ ∞)
k=1

as a smooth approximation to f . With this setup, we arrive at the following

Theorem 9. The limit function u defined by (29) satisfies

f − ut ∈ ∂I[u] a.e. t > 0.

Setup of Theorem 9: We interpret f − ut ∈ ∂I[u] as

Z
I[v] ≥ I[u] + (f (x, t) − ut (x, t))(v − u(x, t)) dx, (30)
Rn

for every v ∈ L2 (Rn ) at a.e. t > 0. We also suppose that v is Lipschitz and that
the Lipschitz constant is at most one. Then the expression
Z
f (x, t)(v − u(x, t)) dx
Rn

59
 Evans Chapters 5 - 9

may defined as
m
X
fk (t)(v(dk ) − u(dk , t)).
k=1

Proof of Theorem 9: As |Du| ≤ 1, we have that I [u(·, t)] = 0 for a.e. t > 0. We
will also assume that v has compact support. To verify (30) is true, we need to
show that Z
(f (x, t) − ut (x, t))(v − u(x, t)) dx ≤ 0 (31)
Rn
for a.e. t > 0. To do this, we fix two times 0 < t1 < t2 . Our original PDE
becomes
fp − up,t = ∂Ip [up ].
This implies
Z t2 Z t2 Z
(t2 − t1 )Ip [v] ≥ I [up (·, t)] dt + (fp (x, t) − up,t (x, t))(v − up (x, t)) dxdt
t1 t1 Rn
Z t2 Z
≥ (fp (x, t) − up.t (x, t))(v − up (x, t)) dxdt, (32)
t1 Rn

Then since |Dv| ≤ 1 a.e. and v is compactly supported, we have

lim Ip [v] = 0.
p→∞

By (29), we see that the weak convergence implies

Z t2 Z Z t2 Z
upi,t (x, t)(v − upi (x, t)) dxdt → ut (x, t)(v − u(x, t)) dxdt. (33)
t1 Rn t1 Rn

It remains to show that

Z t2 Z m Z
X t2
fpi (x, t)(v − upi (x, t)) dxdt → fk (t)(v(dk ) − u(dk , t)) dt. (34)
t1 Rn k=1 t1

Since spt(a) ⊆ {|Du| = 1} it follows that spt(dpk ) ⊆ {|Du| = 1} which implies

Z t2 Z Z m
t2 X
fpi (x, t)v dxdt → fk (t)v(dk ) dt. (35)
t1 Rn t1 k=1

So for a.e. t > 0 and p ≥ 2n we obtain

kup (·, t)kC 0,1/2 ≤ C (kDup (·, t)kL2n + kup (·, t)kL∞ ) .

The C 0,1/2 norm is defined by

|w(x) − w(y)|
kwkC 0,1/2 = kwkL∞ + sup .
x6=y |x − y|1/2

60
 Evans Chapters 5 - 9

This gives
Z t2 Z t2
kup (x, t)kC 0,1/2 dt ≤ C + C kDup (x, t)kL2n dt
t1 t1
Z t2 Z 1/2n
2n
=C +C |Dup (x, t)| dxdt
t1 Rn
Z t2 Z 1/p
≤C +C |Dup (x, t)|p dxdt
t1 Rn
≤ C,

where all of the integrations are done over finite values. Therefore we find
Z t2 Z m Z t2
X Z
fpi (x, t)upi (x, t) dxdt = fk (t) dpki (x)upi (x, t) dxdt
t1 Rn k=1 t1 B(dk ,rpi )

Xm Z t2
= fk (t)upi (dk , t) dt+
k=1 t1
m
X Z t2 Z  
fk (t) dpki (x) upi (x, t) − upi (dk , t) dxdt
k=1 t1 B(dk ,rpi )

:= A1 + A2 .

We bound both of these expressions. First

Xm Z t2
|A2 | ≤ C sup |upi (x, t) − upi (dk , t)| dt
k=1 t1 B(dk ,rpi )
Z t2
≤ Crp1/2
i
kupi (x, t)kC 0,1/2 dt
t1

≤ Crp1/2
i
,
R t2
because t1
kup (x, t)kC 0,1/2 dt ≤ C. We then fix r > 0 to find
m Z
X t2
A1 = fk (t)upi (dk , t) dt
k=1 t1
Xm Z t2 Z
= fk (t)− upi (dk , t) − upi (x, t) dxdt
k=1 t1 B(dk ,r)

Xm Z t2 Z
+ fk (t)− upi (x, t) dxdt
k=1 t1 B(dk ,r)

:= B1 + B2 .

Proceeding as before,
|B1 | ≤ Cr1/2 .

61
 Evans Chapters 5 - 9

To estimate B2 we apply (29) and pass to limits to find

Z Z 
 t2 m Z t2
X Z 
lim sup  fpi (x, t)upi (x, t) dxdt − fk (t)− u(x, t) dxdt
 
pi →∞  t1 R n t1 B(dk ,r) 
k=1
1/2
≤ Cr .
Now since |Du| ≤ 1 a.e. we have
m Z
X
− |u(x, t) − u(dk , t)| dx ≤ Cr,
k=1 B(dk ,r)

which implies
Z Z m Z t2

 t2 X 
lim sup  fpi (x, t)upi (x, t) dxdt − fk (t)u(dk , t) dt ≤ Cr1/2
 
pi →∞  t1 Rn t1 
k=1

for every r > 0. In combination with (35), this justifies (34).

Then (32) − (34) imply

m
Z t2 X Z t2 Z
0≥ fk (t)(v(dk ) − u(dk , t)) dt − ut (x, t)(v − u(x, t)) dxdt
t1 k=1 t1 Rn

for every 0 ≤ t1 < t2 and v as above. Hence

Xm Z
0≥ fk (t)(v(dk ) − u(dk , t)) − ut (x, t)(v − u(x, t)) dx
k=1 Rn

for a.e. t ≥ 0 and v. This establishes (31) and completes the proof. The exercise
then follows from the assumptions of |Du| ≤ 1 and spt a ⊆ {|Du| = 1} which
are included in the assumptions of Theorem 9. The last assumption that
ut − div(Adu) = f in R2 × (0, ∞),
is used in defining the degenerate parabolic PDE
(
up,t − div(|Dup |p−2 Dup ) = fp in Rn × (0, ∞)
up = g on Rn × {t = 0}

where n + 1 ≤ p < ∞ taking the limit as p → ∞ gives the desired evolution

equation.

9.7.13 Assume u is a smooth solution of the gradient flow system (33) in

§9.6.3, where L satisfies the uniform convexity condition (29). Show there exists
constants C, γ > 0 such that
Z
u2t (x, t) dx ≤ Ce−γt (t > 0).
U

62
 Evans Chapters 5 - 9

Proof. The idea behind this problem is that convex gradient flows converge
exponentially to a minimizer.
We first review the theory of gradient flows. In general, a partial differential
equation is a gradient flow of an energy functional E : X → (−∞, +∞] on a
metric space (X, d) if the equation may be written as

d
u(t) = −∇d E(u(t)), u(0) = u ∈ X,
dt
for a generalized notion of gradient ∇d .

To get a feel for gradient flows in L2 (U ), we review the heat equation. It is

defined by the Dirichlet energy
( R
1
|∇u(x)|2 dx, for u ∈ H01 (U )
E(u) := 2 U
+∞ otherwise.

Here E(u) isn’t continuous with respect to the L2 (U ) norm, even on the re-
stricted set G := H01 (U ) ∩ H 2 (U ). However, its directional derivatives are well
defined for every u ∈ G since

E(u + hv) − E(u)

Z Z
lim = ∇u · ∇v = − ∆uv, ∀v ∈ Cc∞ (U ).
h→0 h U U

By this, we define ∇L2 E(u) to be the vector field satisfying

E(u + hv) − E(u)

(∇L2 E(u), v)L2 (U ) = lim = (−∆u, v)L2 (U ) , ∀v ∈ L2 (U ).
h→0 h
Hence, the gradient flow of the Dirichlet energy is the heat equation
d
u(t) = −∇L2 E(u(t)) = ∆u(t).
dt
By linearity of the gradient and convexity of | · |2 , the Dirichlet energy is convex
on L2 (U ).
We now review gradient flows on Hilbert spaces. In general, the theory can be
seen as a nonlinear, infinite dimensional generalization of the theory of ordinary
differential equations. Suppose we have a functional E : H → R ∪ {+∞} on a
Hilbert space H. We then define the Hilbert space gradient analogously to the
definition of ∇L2 . We also define three other properties.
Definition 1 (Hilbert space gradient). ∇H E(u) ∈ H is a Hilbert space gradient
of E at u if
E(u + hv) − E(u)
(∇H E(u), v) = lim , ∀v ∈ H.
h→0 h

63
 Evans Chapters 5 - 9

Definition 2 (L2 gradientR and functional derivative). Consider the class of

integral functions E(u) = F (x, u(x), ∇u(x))dx on L2 (U ). If u is sufficiently
regular then the functional derivative δE 2
δu ∈ L (U ) exists and satisfies

E(u + hv) − E(u)

Z
δE
lim = v, ∀v ∈ Cc∞ (U ).
h→0 h U δu

In our case, we identify ∇L2 E(u) with δE

δu . With this definition of Hilbert space
gradient, we define the Hilbert space gradient flow.
Definition 3 (Hilbert space gradient flow). The gradient flow of E with respect
to H is
d
u(t) = −∇H E(u(t)), u(0) = u.
dt

Definition 4 (Convexity inequality). Given θ ∈ R and a functional E : H →

R ∪ {+∞} we obtain for convex E
θ
(∇E(u) − ∇E(v), v − u) ≥ |v − u|2 .
2
These give rise to the exponential decay of the gradient. That is,

|∇E(u(t))| ≤ e−θt |∇E(u(0))|. (36)

To see this, we impose the additional assumption that for all u ∈ G := H01 (U ) ∩
H 2 (U ), there exists D2 E(u) : G → H satisfying
(∇E(u + hv) − ∇E(u), w)
(D2 E(u)v, w) = lim , ∀v ∈ G.
h→0 h
The convexity inequality then implies
(∇E(u + hv) − ∇E(u), hv) θ |hv|2 θ
(D2 E(u)v, v) = lim 2
≥ lim = |v|2 .
h→0 h h→0 2 h2 2
Therefore
d 1 θ
|∇E(u(t))|2 = −(D2 E(u(t))∇E(u(t)), ∇E(u(t))) ≤ − |∇E(u(t))|2 ,
dt 2 2
and integrating yields the inequality (15). In our case

ut = −∇E(u(t))

so by defining H(t) := |∇E(u(t))|2 we have

d 1 θ
H(t) ≤ − H(t).
dt 2 2
By the inequality
d
H(t)e−θt = H 0 (t)e−θt − θH(t)e−θt ≤ 0


dt

64
 Evans Chapters 5 - 9

we see that H(t) ≤ H(0)e−θt therefore implying

Z Z
u2t (x, t) dx ≤ e−θt |∇E(u(0))|2 = Ce−γt .
U U

The same analysis follows from applying Gronwall’s inequality after integration.
Remark: Perhaps Evans intended to “take a more direct route to arrive at this
inequality.” Returning back to the energy functional E : L2 (U ) → R ∪ {+∞}
defined by ( R
1
|∇u(x)|2 dx, for u ∈ H01 (U )
E(u) := 2 U
+∞ otherwise.
We integrate by parts to find
Z Z
d
E(u(t)) = ∇u(t) · ∇ut (t) = − |∆u(t)|2 = −k∇E(u)k2L2 (U ) .
dt U U

We may also observe that

Z
d
k∇E(u)k2L2 (U ) = 2 ∆u(t)∆ut (t)
dt U
Z
=2 ∆u(t)∆(∆u(t))
U
Z
= −2 |∇(∆u(t))|2
U
≤ 0,

which shows that there is exponential convergence to the unique minimum of

E(u) over L2 (u). Applying similar analysis as above with the uniform convexity
condition gives
d
H(t) ≤ −θH(t).
dt
where H(t) is the L2 norm squared of the energy. Proceeding as before, we see
that
d
H(t)e−θt = H 0 (t)e−θt − θH(t)e−θt ≤ 0


dt
implies H(t) ≤ H(0)e−θt . Hence
Z Z
u2t (x, t) dx ≤ e−θt |∇E(u(0))|2 = Ce−γt .
U U

65
 Evans Chapters 5 - 9

References

[1] G. Aronsson, L.C. Evans, and Y. Wu. “Fast/Slow Diffusion and Growing
Sandpiles”. In: Journal of Differential Equations 131.2 (1996), pp. 304–
335. issn: 0022-0396. doi: https : / / doi . org / 10 . 1006 / jdeq . 1996 .
0166. url: http://www.sciencedirect.com/science/article/pii/
S0022039696901667.
[2] Haim Brezis. Functional Analysis, Sobolev Spaces and Partial Differential
Equations. Springer New York, 2010. doi: 10.1007/978-0-387-70914-7.
url: https://doi.org/10.1007%2F978-0-387-70914-7.
[3] P.G. Ciarlet. Three-Dimensional Elasticity. Mathematical Elasticity. Else-
vier Science, 1994. isbn: 9780444817761. url: https://books.google.
com/books?id=sQiOzyTOJXUC.
[4] Klaus-Jochen Engel and Rainer Nagel. “One-parameter semigroups for lin-
ear evolution equations”. In: Semigroup forum. Vol. 63. 2. Springer. 2001,
pp. 278–280.
[5] Lawrence C Evans. “Partial differential equations and Monge-Kantorovich
mass transfer”. In: Current developments in mathematics 1997.1 (1997),
pp. 65–126.
[6] Giovanni Leoni. “A First Course in Sobolev Spaces”. In: (Jan. 2009). doi:
10.1090/gsm/105.
[7] Walter Rudin. Real and Complex Analysis, 3rd Ed. USA: McGraw-Hill,
Inc., 1987. isbn: 0070542341.
[8] ELIAS M. STEIN. Singular Integrals and Differentiability Properties of
Functions (PMS-30). Princeton University Press, 1970. isbn: 9780691080796.
url: http://www.jstor.org/stable/j.ctt1bpmb07.

66