Evans PDE Solutions for Ch2 and Ch3

Osman Akar
July 2016

This document is written for the book ”Partial Differential Equations” by Lawrence C. Evans (Second
Edition). The document prepared under UCLA 2016 Pure REU Program.

SOLUTIONS OF CHAPTER 2

1. Consider the function z : < → < for fixed x ∈ <n and t ∈ (0, ∞)

z(s) = u(x + bs, t + s)ecs

Then
∂z
ż(s) := = ecs (b · Dx u(x + sb, t + s) + ut (x + sb, t + s) + cu(x + sb, t + s)) = 0
∂s
by the condition given by the problem. Therefore, z is a constant function with respect to s. Finally, by
using the fact that u = g on <n × {t = 0}, we conclude that

u(x, t) = z(0) = z(−t) = u(x − tb, 0)e−ct = g(x − tb)e−ct

2. Let O = (aij ) and φ(x) = O · x. Then it’s clear that Dφ(x) = O. Since v is defined to be
v(x) = u(O · x) = (u ◦ φ)(x), we calculate

Dv(x) = Du(φ(x)) · Dφ(x) = Du(O · x) · O

Pn
Then vxi (x) = j=1 uxj (φ(x))aji . Note that by the same calculations, we have

D(uxj ◦ φ)(x) = Duxj (φ(x)) · O

Therefore
n n n n
∂ X ∂ X X X
vxi xi (x) = vxi (x) = uxj (φ(x))aji = aji ( uxj xk (φ(x))aki ) = aji aki uxj xk (φ(x))
∂xi j=1
∂xi j=1 k=1 j,k=1

The laplacian of v is
n
X n
X n
X n
X n
X
∆v(x) = vxi xi (x) = ( aji aki uxj xk (φ(x))) = (uxj xk (φ(x)) aji aki )
i=1 i=1 j,k=1 j,k=1 i=1

Since O is an orthogonal matrix, (

n
X 1 if j = k
aji aki =
i=1
0otherwise
Lastly, we conclude our proof with
n
X
∆v(x) = (uxj xj (φ(x)) = ∆u(φ(x)) = 0
j=1

1
since u is harmonic.

3. We’ll modify the mean value property for harmonic functions. It is assumed that the reader know
the proof of the mean value property(Check Evans PDE 2.2.2, Mean Value Formulas). As in the proof
in the book, define Z
1
φ(s) = u(y)dS(y)
nα(n)sn−1 ∂B(0,s)
Then the book proves Z
1
φ0 (s) = ∆u(y)dy
nα(n)sn−1 B(0,s)

Since ∆u = −f in B(0, r)
−1
Z
φ0 (s) = f (y)dy
nα(n)sn−1 B(0,s)

By Fundamental Theorem of Calculus,

Z r
φ(r) − φ() = φ0 (s)ds


for any  > 0. Note that

Z
1
lim φ() = lim u(y)dS(y)
→0 →0 nα(n)n−1 ∂B(0,)
Z
1
= u(y)dS(y)
m(∂B(0, )) ∂B(0,)
= u(0) (3.1)
Rr
Let’s calculate 
φ0 (s)ds
r r
−1
Z Z Z
0
φ (s)ds = n−1
f (y)dyds
  nα(n)s B(0,s)
Z r
−1
Z
1
= f (y) n−1
dsdy
nα(n) B(0,r) max(|y|,) s

In the last equality, we changed the order of integration. Note that

s=r
r
−1
Z
1 1 1 1
ds = = ( − n−2 ) (3.2)
ζ sn−1 (n − 2)sn−2 n − 2 ζ n−2 r
ζ

Therefore
Z r Z Z r
1 1
φ0 (s)ds = f (y) dsdy
 nα(n) B(0,r) max(|y|,) sn−1
Z r Z r
−1
Z Z
1 1
= ( f (y) n−1
dsdy + f (y) n−1
dsdy)
nα(n) B(0,r)−B(0,) |y| s B(0,)  s
−1
Z Z
1 1 1 1 1 1
= ( f (y) ( n−2 − n−2 )dy + f (y) ( n−2 − n−2 )dy) by (3.2)
nα(n) B(0,r)−B(0,) n − 2 |y| r B(0,) n − 2  r
−1
Z Z
1 1 1 1
= ( f (y)( n−2 − n−2 )dy + f (y)( n−2 − n−2 )dy)
n(n − 2)α(n) B(0,r) |y| r B(0,)  |y|

Suppose that we proved Z

1 1
lim f (y)( − n−2 )dy = 0 (3.3)
→0 B(0,) n−2 |y|

2
Then
φ(r) − u(0) = lim φ(r) − φ() by (3.1)
→0
Z r
= lim φ0 (s)ds
→0 
−1
Z Z
1 1 1 1
= lim ( f (y)( n−2 − n−2 )dy + f (y)( n−2 − n−2 )dy)
→0 n(n − 2)α(n) B(0,r) |y| r B(0,)  |y|
−1
Z
1 1
= f (y)( n−2 − n−2 )dy
n(n − 2)α(n) B(0,r) |y| r
Therefore
Z
1 1 1
u(0) = φ(r) + f (y)( n−2 − n−2 )dy
n(n − 2)α(n) B(0,r) |y| r
Z Z
1 1 1 1
= u(y)dS(y) + f (y)( n−2 − n−2 )dy
nα(n)rn−1 ∂B(0,r) n(n − 2)α(n) B(0,r) |y| r

But u = g on ∂B(0, r), so the equality holds. So we only need to prove (3.3)
Z Z Z
1 1 1 f (y)
lim f (y)( n−2 − n−2 )dy = lim n−2 f (y)dy − lim n−2
→0 B(0,)  |y| →0  B(0,) →0 B(0,) |y|

But
Z Z
1 1
lim f (y)dy = lim nα(n)2 f (y)dy
→0 n−2 B(0,) →0 m(B(0, )) B(0,)

= lim nα(n)2 f (0)

→0
= 0
and
Z Z Z
f (y) f (y)
lim = lim dS(y)dt
→0 B(0,) |y|n−2 →0 0 ∂B(0,t) |y|n−2
Z  Z
1
= lim f (y)dS(y)dt
→0 0 tn−2 ∂B(0,t)
Z  Z
1
= lim nα(n)t f (y)dS(y)dt
→0 0 m(∂B(0, t)) ∂B(0,t)
Z 
= lim nα(n)tf (0)dt
→0 0
2
= lim nα(n)f (0)
→0 2
= 0
Therefore Z
1 1
lim f (y)( − n−2 )dy = 0
→0 B(0,) n−2 |y|
so we are done.

4. First note that U and ∂U is compact, since U is bounded and open. So max u = sup u on
both U and ∂U . Define u := u + |x|2 for each  > 0. Then for x ∈ U , ∂u ∂xi = uxi + 2xi , so


∆u = ∆u + 2n = 2n > 0 since u is harmonic and  > 0. But tr(Hess(u )) = ∆u (x) > 0,
so Hess(u )(x) is cannot be negative definite matrix. Therefore, x cannot be local maximum of the
function u , so u cannot attain its max within U . Therefore, we conclude that
max u = max u
U ∂U

3
Since we know that U is bounded, we can assume U ⊂ B(0, R) for some R > 0. Then

max u ≤ max u = max u ≤ max u + max |x|2 ≤ max u + R2

U U ∂U ∂U ∂U ∂U

Let  → 0, then max u ≤ max u. Since ∂U ⊂ U , we conclude max u = max u

U ∂U U ∂U

5. (a) As in the proof in the mean value property (or in the problem 3), define a function
Z
1
Φ(s) = v(y)dS(y)
nα(n)sn−1 ∂B(x,s)

then the book proves Z

1
Φ0 (s) = ∆v(y)dy
nα(n)sn−1 B(x,s)

so Φ0 (s) ≥ 0 since ∆v ≥ 0. Therefore Φ(r) ≥ Φ() for all r >  > 0. But lim v() = v(x) since Φ is an
→0
average integral of the function v.

(b) Assume that U is connected. Let M := max u and S = {x ∈ U v(x) = M }. If S is an empty

U
set, then clearly max = max u, so we are done. Assume that S is not an empty set. We’ll prove that S
U ∂U
is both open and relatively closed, which will led us to say S = U since U is connected. Let a ∈ S and
δ > 0 such that B(a, δ) ⊂ U . Then for all δ ≥  > 0
Z
1
v(a) = v(y)dS(y) ≤ v(a)
nα(n)sn−1 ∂B(a,δ)

therefore v = v(a) within B(a, δ), ie B(a, δ) ⊂ S, so S is open. Moreover S is relatively closed since u is a
continuous function. Thus S = U , i.e. u is a constant function in U , and also in U since u is continuous.
Therefore max = max u. Since this is true for all connected parts of the nonconnected open set U , we
U ∂U
can say max = max u.
U ∂U

Second way
If U were bounded, we could use problem 4. Define v (x) = v(x) + |x|2 . Then ∆v = ∆v + 2n > 0, the
rest are the same.

(c) We do straight calculation

∂v ∂φ(u)
= = φ0 (u)uxi
∂xi ∂xi
So
∂2v
= φ00 (u)u2xi + φ0 (u)uxi xi
∂x2i
Therefore
n
X
∆v = φ00 (u) u2xi + φ0 (u)∆u ≥ 0


i=1

(d) Follow the calculations

n
X
2
v = |Du| = u2xi
i=1
n
∂v X
= 2uxi uxi xk
∂xk i=1

4
 n
∂2v X
=2 u2xi xk + uxi xk xk
∂x2k i=1
n
X n
X n
X
∆v = 2 u2xi xk + ∆uxi ≥ ∆uxi
i,k=1 i=1 i=1
Pn
Since u is harmonic, uxi is harmonic for all i = 1, 2, ..., n. Thus ∆v ≥ i=1 ∆uxi = 0

6. Since U is bounded, we can assume that U ⊂ B(0, R) for some fixed R > 0. Choose C such that
2
C > max(1, R
2n ). Let’s define λ := max |f | (??? can we, can f be extended continuously over U ) Since
U
2 2
∆(u + λ |x| 2λ |x|
2n ) = ∆u + n 2n ≥ 0 u + λ 2n is subharmonic. By the problem 5

|x|2 |x|2
max (u + λ ) = max (u + λ )
U 2n ∂U 2n
R2
≤ max |u| + λ
∂U 2n
R2
≤ max |g| + max |f |
∂U 2n U
≤ C(max |g| + max |f |)
∂U U

7. We are assuming that u is harmonic in an open set U satisfying B(0, r) ⊂ U ⊆ <n . Let’s write
the Poisson’s formula x ∈ B(0, r),
r2 − |x|2
Z
u(y)
u(x) = dy
nα(n) ∂B(0,r) |x − y|n
Note that by the triangle inequality
r + |x| = |y| + |x| ≥ |x − y| ≥ ||y| − |x|| = r − |x|
for y ∈ ∂B(0, r). Moreover, Z
u(y)dy = nα(n)rn−1 u(0)
∂B(0,r)

by the mean value property of the harmonic functions. Then

r − |x| r2 − |x|2 r2 − |x|2
Z Z
u(y) u(y)
rn−2 u(0) = dy ≤ dy = u(x)
(r + |x|)n−1 nα(n) ∂B(0,r) (r + |x|)n nα(n) ∂B(0,r) |x − y|n
Similarly
r + |x| r2 − |x|2 r2 − |x|2
Z Z
u(y) u(y)
rn−2 n−1
u(0) = n
dy ≥ dy = u(x)
(r − |x|) nα(n) ∂B(0,r) (r − |x|) nα(n) ∂B(0,r) |x − y|n

8. (i) We’ll prove the Theorem 15 in the chapter 2.2. Take u = 1 in B(0, r) and consequently g = 1
on ∂B(0, r), then by theorem 12 (Representation Formula using Green’s function), for any x ∈ B(0, r)
Z Z
1 = u(x) = g(y)K(x, y)dS(y) = K(x, y)dS(y) (8.1)
∂B(0,r) ∂B(0,r)
R
Let’s back to the problem. First, we shall show that uxi exist and it equals to ∂B(0,r) g(y)Kxi (x, y)dS(y).
It’s enough to show that
u(x + hei ) − u(x)
Z
lim = g(y)Kxi (x, y)dS(y)
h→0 h ∂B(0,r)

5
Let’s denote
u(x + hei ) − u(x) K(x + hei , y) − K(x, y)

Z
v(x, h) := = g(y) dS(y)
h ∂B(0,r) h
By Mean Value Theorem, K(x + hei , y) − K(x, y) = hKxi (x + γ(h), y) for some 0 < γ(h) < h. Therefore
Z
v(x, h) = g(y)Kxi (x + γ(h), y)dS(y)
∂B(0,r)
Z Z


v(x, h) − g(y)Kxi (x, y)dS(y) = g(y) Kxi (x + γ(h), y) − Kxi (x, y) dS(y)
∂B(0,r) ∂B(0,r)

Let µ := dist(x, ∂B(0, r)) = inf {|x − y| y ∈ ∂B(0, r)} and consider the continuous function Kxi (x, y)
on the compact set U (0, r − µ2 ) × ∂B(0, r). Thus Kxi (x, y) is uniformly continuous. Also, since g is
continuous in the compact set ∂B(0, r), g is also bounded. Assume |g| < M . Choose  > 0. Since
Kxi (x, y) is uniformly continuous, there exist δ > 0 such that |(x1 , y1 ) − (x2 , y2 )|<2n < δ implies
|Kxi (x1 , y1 ) − Kxi (x2 , y2 )| <  Thus, for h < min {µ, δ},
Z Z


v(x, h) − g(y)Kxi (x, y)dS(y) = g(y) Kxi (x + γ(h), y) − Kxi (x, y) dS(y)
∂B(0,r) ∂B(0,r)
Z
≤ |g(y)| Kxi (x + γ(h), y) − Kxi (x, y) dS(y)
∂B(0,r)
Z
≤ |g(y)| Kxi (x + γ(h), y) − Kxi (x, y) dS(y)
∂B(0,r)
Z
≤ |g(y)|dS(y)
∂B(0,r)

≤ M nα(n)rn−1
u(x+hei )−u(x) R
which goes 0 as  goes 0. Therefore lim h exist and it equals to ∂B(0,r)
g(y)Kxi (x, y)dS(y)
h→0
In the same way, we can prove for all α = (α1 , α2 , ..., αn ),
Z
α
D u(x) = g(y)Dα K(x, y)dS(y)
∂B(0,r)
∞
since K is smooth, so is u, i.e u ∈ C (B(0, r))
R
(ii) By (i), we can say ∆u = ∂B(0,r)
g(y)∆x K(x, y)dS(y). But ∆x K(x, y) = 0 within B(0, r) (I
may add calculation) so ∆u = 0

(iii) We’ll follow the similar path as in the proof of the theorem 14. Let  > 0, since g is continuous,
choose δ > 0 such that |y − x0 | < δ implies |g(y) − g(x0 )| < . Moreover, since g is defined in the compact
set ∂B(0, r), it is bounded, assume |g| < M . By (8.1),
Z Z
0
|u(x) − g(x )| = g(y)K(x, y)dS(y) − g(x0 )K(x, y)dS(y)
∂B(0,r) ∂B(0,r)
Z
= (g(y) − g(x0 ))K(x, y)dS(y)
∂B(0,r)
Z
≤ |g(y) − g(x0 )|K(x, y)dS(y)
∂B(0,r)
Z
= |g(y) − g(x0 )|K(x, y)dS(y) +
∂B(0,r)∩B(x0 ,δ)
Z
|g(y) − g(x0 )|K(x, y)dS(y)
∂B(0,r)−B(x0 ,δ)
= I +J

6
 R R
Clearly I < ∂B(0,r)∩B(x0 ,δ)
K(x, y)dS(y) <  ∂B(0,r)
K(x, y)dS(y) =  also by (8.1). Furthermore,
0 δ 0
if |x − x | < 2 and |y − x | > δ, we have

δ 1
|y − x0 | < |y − x| + |x − x0 | < |y − x| + < |y − x| + |y − x|
2 2
therefore |y − x| > 12 |y − x0 |, so

r2 − |x|2 2n (r2 − |x|2 ) 1 2n (r2 − |x|2 )

K(x, y) = n
< 0 n
<
nα(n)|x − y| nα(n) |y − x | nα(n)δ n
δ
for |x − x0 | < 2 and y ∈ ∂B(0, r) − B(x0 , δ). By the triangle inequality, we have |g(y) − g(x0 )| < 2M .
So finally

2n (r2 − |x|2 )
Z
J < 2M dS(y)
∂B(0,r)−B(x0 ,δ) nα(n)δ n
n 2 2
2 (r − |x| )
≤ 2M nα(n)rn−1
nα(n)δ n
2n+1 M rn−1
= (r2 − |x|2 )
δn
which goes 0 as x goes x0 since x0 ∈ ∂B(0, r) so |x| → |x0 | = r

9. By Poisson’s formula for the half space, for x = (x1 , x2 , ..., xn ) ∈ <n+
Z Z
2xn g(y) 2xn g̃(ỹ)
u(x) = dS(y) = dỹ
nα(n) ∂<n+ |x − y|n nα(n) <n−1 |x − y|n

where ỹ = (y1 , y2 , ..., yn−1 ) ∈ <n−1 for y = (y1 , y2 , ..., yn−1 , 0) ∈ ∂<n+ . Let’s put x = λen = (0, 0, ..., λ),
then |x − y|2 = |ỹ|2 + λ2 , therefore the equality becomes
Z
2 λ
u(λen ) = g̃(ỹ) n dỹ
nα(n) <n−1 (|ỹ| + λ2 ) 2 +1
2

We’ll compute the partial derivative of u WRT xn at x = λen .

∂
uxn (λen ) =
u(λen )
∂λ Z
∂ 2 λ
= g̃(ỹ) n dỹ (9.0)
∂λ nα(n) <n−1 (|ỹ|2 + λ2 ) 2 +1
Z
2 ∂ λ
= g̃(ỹ) n dỹ
nα(n) <n−1 ∂λ (|ỹ|2 + λ2 ) 2 +1
|ỹ|2 − (n − 1)λ2
Z
2
= g(ỹ) n
nα(n) <n−1 |ỹ 2 + λ2 | 2 +1
Z ∞Z
2 |ỹ|2 − (n − 1)λ2
= g(ỹ) n dS(ỹ)dr (9.1)
nα(n) 0 ∂B n−1 (0,r) (|ỹ|2 + λ2 ) 2 +1
Z ∞Z
2 r2 − (n − 1)λ2
= g(ỹ) 2 n dS(ỹ)dr
nα(n) 0 ∂B n−1 (0,r) (r + λ2 ) 2 +1
R
At (9.0), we used the fact in the proof of the theorem 14, which is uxi (x) = ∂<n g(y)Kxi (x, y)dy
0
At 9.1, we used polar coordinates. B n−1 (0, r) denotes the open ball centered at origin with radious r in
<n−1 . Define functions
r2
Z
f1 (r, λ) = g(ỹ) 2 n dS(ỹ)
∂B n−1 (0,r) (r + λ2 ) 2 +1

7
and
−(n − 1)λ2
Z
f2 (r, λ) = g(ỹ) n dS(ỹ)
∂B n−1 (0,r) (r2 + λ2 ) 2 +1
Note that
Z ∞ Z 1 Z ∞ Z 1
uxn (λen ) = f1 (r, λ)dr + f1 (r, λ)dr + f2 (r, λ)dr + f2 (r, λ)dr
1 0 1 0
R∞ R∞ R1 R1
Assume that we know 1
f1 (r, λ)dr, 1
f2 (r, λ)dr and 0
f2 (r, λ)dr are bounded and lim 0
f1 (r, λ)dr =
λ→0
∞. Then we can conclude that uxn (λen ) → ∞ as λ → 0; therefore Du(λen ) is not bounded near 0. Let’s
prove the assumptions. Since we know g is bounded, we assume |g| < M

∞ ∞
r2
Z Z Z
f1 (r, λ)dr ≤ g(ỹ) n dS(ỹ)dr
1 1 ∂B n−1 (0,r) (r2 + λ2 ) 2 +1
∞Z
r2
Z
≤ M n dS(ỹ)dr
1 ∂B n−1 (0,r) (r2 + λ2 ) 2 +1
∞
r2
Z
= M (n − 1)α(n − 1)rn−2 2 n dS(ỹ)dr
1 (r + λ2 ) 2 +1
Z ∞
rn
= M (n − 1)α(n − 1) n dr (9.2)
1 (r2 + λ2 ) 2 +1
Z ∞
r
≤ M (n − 1)α(n − 1) 3 dr
1 (r + λ2 ) 2
2
 ∞
−1
= M (n − 1)α(n − 1) 1
(r2 + λ2 ) 2 1
1
= M (n − 1)α(n − 1) 2
λ +1
≤ M (n − 1)α(n − 1)
Note that at (9.2), we used the inequality
n−1
rn

r r r
2 2 n
+1
= 3 1 ≤ 3
(r + λ ) 2 (r + λ2 ) 2
2 (r2 + λ2 ) 2 (r2 + λ2 ) 2
R∞
We find that 1 f1 (r, λ)dr is bounded. For the second assumption
Z ∞ Z ∞Z
−(n − 1)λ2
f2 (r, λ)dr ≤ g(ỹ) 2 n dS(ỹ)dr
1 1 ∂B n−1 (0,r) (r + λ2 ) 2 +1
Z ∞Z
1
≤ M (n − 1)λ2 2 + λ2 ) n
dS(ỹ)dr
n−1 (r 2 +1
1 ∂B (0,r)
Z ∞
rn−2
= M (n − 1)2 α(n − 1)λ2 n dr (9.3)
1 (r2 + λ2 ) 2 +1
Z ∞
r
≤ M (n − 1)2 α(n − 1)λ2 3 dr
1 (r + λ2 ) 2
2
 ∞
−1
= M (n − 1)2 α(n − 1)λ2 1
(r2 + λ2 ) 2 1
λ2
= M (n − 1)2 α(n − 1) 2
λ +1
2
≤ M (n − 1) α(n − 1)
At (9.3) we used the inequality
n−1
rn−2

r r 1 r
2 2 n
+1
= 3 1 2
≤ 3
(r + λ ) 2 (r + λ2 ) 2
2 (r2 + λ2 ) 2 r (r + λ2 ) 2
2

8
The inequality is true for r > 1, which is enough since we integrate from 1 to ∞. For third assumption,
realize that for r ≤ 1
−(n − 1)λ2 rn−1
Z
2 2
f2 (r) = |ỹ| 2 n dS(ỹ) = −λ (n − 1) α(n − 1) n
∂B n−1 (0,r) (r + λ2 ) 2 +1 (r2 + λ2 ) 2 +1

Since g(ỹ) = |ỹ| for |ỹ| ≤ 1. Therefore

Z ∞ 1
rn−1
Z
f2 (r, λ)dr ≤ λ2 (n − 1)2 α(n − 1) n
1 0 (r2 + λ2 ) 2 +1
Z 1
r
≤ λ2 (n − 1)2 α(n − 1) (9.4)
0 (r2 + λ2 )2
 r=1 
2 2 −1
≤ λ (n − 1) α(n − 1)
2(r2 + λ2 ) 0
1
= (n − 1)2 α(n − 1)
2(λ2 + 1)
1
≤ (n − 1)2 α(n − 1)
2
At (9.4), we used the inequality
n−2
rn−1

r r r
n = 2 1 ≤
(r2 + λ2 ) 2 +1 (r + λ2 )2 (r2 + λ2 ) 2 (r2 + λ2 )2

since n > 1.
Let’s prove the last assumption. For r < 1

r2
Z
f1 (r, λ) = g(ỹ) n dS(ỹ)
∂B n−1 (0,r) (r2 + λ2 ) 2 +1
r2
Z
= |ỹ| n dS(ỹ)
∂B n−1 (0,r) (r2 + λ2 ) 2 +1
r3
Z
= n dS(ỹ)
∂B n−1 (0,r) (r2 + λ2 ) 2 +1
r3
= (n − 1)α(n − 1)rn−2 n
(r2 + λ2 ) 2 +1
rn+1
= (n − 1)α(n − 1) 2 n
(r + λ2 ) 2 +1

since g(ỹ) = |ỹ| for |ỹ| ≤ 1. Let (n − 1)α(n − 1) = µ. Then

Z 1 Z 1
rn+1
f1 (r, λ)dr = µ 2 2 n +1
dr
0 0 (r + λ ) 2
Z 1
rn+1
≥ µ 2 2 n +1
dr (9.5)
λ (r + λ ) 2
Z 1
1 1
≥ µ n
+1 r
dr
λ 2 2
Z 1
1 1
= µ n +1 dr
2 2
λ r
 1
µ
= n+2 log(r)
2 2 λ
µ
= n+2 (−log(λ))
2 2

9
which goes infinity as λ → 0. Note that at (9.5), we used inequality

rn+1 1 1
2 2 n
+1
≥ n +1
(r + λ ) 2 22 r

which is equivalent
n n
(2r2 ) 2 +1 ≥ (r2 + λ2 ) 2 +1
which is true for r ≥ λ, so we are done.

x̃
11. First, let’s focus on the function x̃ : <n → <n which takes x to |x|2 .

x̃ = x̃(x) = (x1 |x|−2 , x2 |x|−2 , ..., xn |x|−2 )

and
(|x|2 − 2x21 )|x|−4 −2x1 x2 |x|−4 −2x1 x3 |x|−4 −2x1 xn |x|−4
 
···
 −2x2 x1 |x|−4 (|x|2 − 2x22 )|x|−4 −2x2 x3 |x|−4 ··· −2x2 xn |x|−4 
−4
−2x3 x2 |x|−4 (|x|2 − 2x23 )|x|−4 −2x3 xn |x|−4
 
D(x̃) = Dx (x̃) =  −2x3 x1 |x| ···
 

 .. .. .. .. .. 
 . . . . . 
−2xn x1 |x|−4 −2xn x2 |x|−4 −2xn x3 |x|−4 ··· (|x|2 − 2x2n )|x|−4

Let’s denote D(x̃) = (αji (x)) where

(
−2xi xj |x|−4
 
∂ xi if j 6= i
αji (x)


= D(x̃) ij = =
∂xj |x|2 (|x|2 − 2x2i )|x|−4 if j = i

Clearly αji = αij , ie the matrix D(x̃) is symmetric. Let ri (x) = (αi1 (x), αi2 (x), ..., αin (x)) = (α1i (x), α2i (x), ..., αni (x)
xi


denote the ith row and column. Note that ri (x) = D |x| 2 .

Lemma 1. D(x̃) · D(x̃) = |x|−4 · I

Proof We need to prove that (
|x|−4 if j = i
ri (x) · rj (x) =
0 otherwise
For j = i
n
X
ri (x) · ri (x) = (αki (x))2
k=1
n
X
= ((|x|2 − 2x2i )|x|−4 )2 + (−2xi xk |x|−4 )2
k=1,k6=i
n
X
= (|x|4 − 4|x|2 x2i + 4x4i )|x|−8 + 4x2i x2k |x|−8
k=1,k6=i
 n
X 
−8 4 2 2 2 2
= |x| |x| − 4|x| xi + 4xi xk
k=1
= |x|−4

10
 For j 6= i
n
X
ri (x) · rj (x) = αki (x)αkj (x)
k=1
n
X
= −2xj xi |x|−4 (|x|2 − 2x2i )|x|−4 − 2xj xi |x|−4 (|x|2 − 2x2j )|x|−4 + 4xi xj x2k |x|−8
k=1,k6=i,j
 n
X 
= 2xi xj |x|−8 2x2i + 2x2j − 2|x|2 + 2x2k
k=1,k6=i,j
 n
X 
= 2xi xj |x|−8 − 2|x|2 + 2x2k
k=1
= 0

So the lemma is proved.

Corollary (1) (
|x|−4 if j = i
ri (x) · rj (x) =
0 otherwise
Let’s continue the problem. By product rule

D(ũ(x)) = D(u(x̃)|x|2−n ) = Du(x̃) · D(x̃)|x|2−n + u(x̃)D(|x|2−n )

So
∂
ũ(x) = Du(x̃) · ri (x)|x|2−n + u(x̃)(2 − n)xi |x|−n
∂xi
∂2
 
∂ ∂
ũ(x) = (Du(x̃)) · ri (x)|x|2−n
+ Du(x̃) · ri (x) |x|2−n + (2 − n)Du(x̃) · ri (x)xi |x|−n +
∂x2i ∂xi ∂xi
 
∂
+(2 − n) u(x̃) xi |x|−n + (2 − n)u(x̃)|x|−n + n(n − 2)u(x̃)x2i |x|−n−2
∂xi
We shall prove that
n
X ∂
(1) (Du(x̃)) · ri (x)|x|2−n = 0
i=1
∂x i

n  
X ∂
(2) Du(x̃) · ri (x) |x|2−n = −2(n − 2)|x|−2−n Du(x̃) · x
i=1
∂xi
n
X
(3) (2 − n)Du(x̃) · ri (x)xi |x|−n = (n − 2)|x|−2−n Du(x̃) · x
i=1
n  
X ∂
(4) (2 − n) u(x̃) xi |x|−n = (n − 2)|x|−2−n Du(x̃) · x
i=1
∂x i

n
X
(5) n(n − 2)u(x̃)x2i |x|−n−2 = n(n − 2)u(x̃)|x|−n
i=1
Pn
Note that (2)+(3)+(4)=0 and (5)=n(n − 2)u(x̃)|x|−n = − i=1 (2 − n)u(x̃)|x|−n , therefore if (1)-
(5) are all true, we can conclude that ∆ũ = 0. Let’s start to prove claims
Pn ∂
Proof (1) We shall prove that i=1 ∂xi (Du(x̃)) · ri (x) = 0, then we can immediately conclude that

11
Pn ∂ ∂ ∂
i=1 ∂xi (Du(x̃)) · ri (x)|x|2−n = 0. First, let’s calculate ∂x i
(Du(x̃)) = ∂x i
(ux1 (x̃), ux2 (x̃), ..., uxn (x̃)).
For j = 1, 2, 3, ..., n,
  n
∂ X
uxj (x̃) = Duxj (x̃) · ri (x) = (uxj x1 (x̃), uxj x2 (x̃), ..., uxj xn (x̃)) · ri (x) = uxj xk αki (x)
∂xi
k=1

So
n n
∂ XX
(Du(x̃)) · ri (x) = uxj xk αki (x)αji (x)
∂xi j=1 k=1

Therefore

n n X
n X
n
X ∂ X
(Du(x̃)) · ri (x) = uxj xk (x̃)αki (x)αji (x)
i=1
∂x i i=1 j=1 k=1
Xn n
X
= uxj xk (x̃) αki (x)αji (x)
j,k=1 i=1
Xn
= uxj xk (x̃)rj (x)ri (x) (11.1)
j,k=1
X n
= uxj xj (x̃)|x|−4
j=1

= |x|−4 ∆u(x̃)
= 0
since u is harmonic. At (11.1), we used corollary 1

∂
Proof (2) Let’s start by calculating ∂xi ri (x). For j 6= i
∂ i ∂
α (x) = (−2xi xj |x|−4 ) = −2xj |x|−4 + 8xj x2i |x|−6
∂xi j ∂xi
and for j = i
∂ i ∂
α = (|x|−2 − 2x2i |x|−4 ) = −6xi |x|−4 + 8x3i |x|−6
∂xi i ∂xi
So
n   n X
n
X ∂ X ∂ i
Du(x̃) · ri (x) = uxj (x̃) α (x)
i=1
∂xi i=1 j=1
∂xi j
n n
X X ∂ i
= uxj (x̃) αj (x)
j=1 i=1
∂xi

Xn  n
X 
−4 3 −6 −4 2 −6
= uxj (x̃) − 6xj |x| + 8xj |x| + −2xj |x| + 8xj xi |x|
j=1 i=1,i6=j
n
X  n
X 
−4 −6 2
= uxj (x̃) (−2n − 4)xj |x| + 8xj |x| xi
j=1 i=1
Xn
= uxj (x̃) (−2n − 4)xj |x|−4 + 8xj |x|−4 )
j=1
n
X
= −2(n − 2)|x|−4 uxj (x̃) · xj
j=1

= −2(n − 2)|x|−4 Du(x̃) · x

12
Finally
n  
X ∂
Du(x̃) · ri (x) |x|2−n = −2(n − 2)|x|−4 Du(x̃) · x|x|2−n = −2(n − 2)|x|−n−2 Du(x̃) · x
i=1
∂xi

so we proved (2).

Proof (3) We do straight calculation

n
X n X
X n
Du(x̃) · ri (x)xi = uxj (x̃)αji (x)xi
i=1 i=1 j=1
n 
X n
X 
= uxj (x̃) αji (x)xi
j=1 i=1
Xn   n
X 
−2 3 −4 2 −4
= uxj (x̃) xj |x| − 2xj |x| + −2xi xj |x|
j=1 i=1,i6=j
n 
X  n
X 
−2 2 −4
= uxj (x̃) xj |x| − 2xj xi |x|
j=1 i=1
n
X
= − uxj (x̃)xj |x|−2
j=1

= −|x|−2 Du(x̃) · x

Finally
n
X
(2 − n)Du(x̃) · ri (x)xi |x|−n = (2 − n)|x|−n (−|x|−2 Du(x̃) · x
i=1
= (n − 2)|x|−2−n Du(x̃) · x

∂
Proof (4) Since ∂xi ux̃ = Du(x̃) = Du(x̃)ri (x), we can immediately see (4)=(3), so we are done.

Proof (5) This is also trivial, as follows

n
X n
X
n(n − 2)u(x̃)x2i |x|−n−2 = n(n − 2)u(x̃)|x|−n−2 x2i
i=1 i=1
= n(n − 2)u(x̃)|x|−n−2 |x|2
= n(n − 2)u(x̃)|x|−n

12. Let’s use the notation u(x, t; λ) := uλ (x, t) = u(λx, λ2 t).

(a) We do straight calculation
∂ ∂
u(x, t; λ) = u(λx, λ2 t) = λ2 ut (λx, λ2 t)
∂t ∂t
∂ ∂
u(x, t; λ) = u(λx, λ2 t = λuxi (λx, λ2 t)
∂xi ∂xi
∂2 ∂
u(x, t; λ) = λuxi (λx, λ2 t) = λ2 uxi xi u(λx, λ2 t)
∂x2i ∂xi

13
Therefore
ut (x, t; λ) − ∆x u(x, t; λ) = λ2 ut (λx, λ2 t) − λ2 ∆x u(λx, λ2 t) = 0
so uλ (x, t) = u(x, t; λ) solves heat equation.

(b)
∂ ∂
u(x, t; λ) = u(λx, λ2 t) = x · Dx u(λx, λ2 t) + 2λtut (λx, λ2 t)
∂λ ∂λ
Since u(x, t) is smooth, so is u(x, t; λ). Therefore the derivatives commutes. More precisely,
 
∂ ∂
∂ ∂
( − ∆x ) u(x, t; λ) = ( − ∆x )u(x, t; λ) = 0
∂t ∂λ ∂λ ∂t
∂
by (a). So the function ∂λ u(x, t; λ) solves the heat equation for fixed λ. Take λ = 1, then

∂ ∂
u(x, t; λ) = u(x, t; 1) = x · Dx u(x, t) + 2tut (x, t)
∂λ ∂λ
solves the heat equation.

x
13. (a) Let z = z(x, t) = √
t
. Then

∂ x 1 x
ut (x, t) = v( √ ) = − 3 v 0 (z(x, t))
∂t t 2 t2
∂ x 1
ux (x, t) = v( √ ) = √ v 0 (z(x, t))
∂x t t
∂ 1 0 x 1
ux x(x, t) = √ v ( √ ) = v 00 (z(x, t))
∂x t t t
Therefore
ut = ux x
⇐⇒
1 x 0 1
− 3 v (z(x, t)) = v 00 (z(x, t))
2 t2 t
⇐⇒
1
− z(x, t)v 0 (z(x, t)) = v 00 (z(x, t))
2
⇐⇒
z
(*) v (z) + v 0 (z) = 0
00
2
v 00
By (*), v 0 satisfies v0 = − z2 , so

v 00 −z z2
Z Z
log(v 0 ) = + C1 = + C1 = − + C1
v0 2 4
2 2
v 0 (z) = e−z /4+C1
= ce−z /4

where c = eC1 is a constant. Thus

Z Z
2
v(z) = v 0 (s)ds + d = c e−s /4
ds + d

14
 14. Define function v(x, t) = u(x, t)ect . Then

vt − ∆v = ut ect + ucect − ect ∆u = ect (ut − ∆u + cu) = ect f (14.1)

and clearly v = g on <n × {t = 0}. Using THEOREM 2 in section 2.3.1, the function
Z Z tZ
v(x, t) = Φ(x − y, t)g(y)dy + Φ(x − y, t − s)f (y, s)ecs dyds
<n 0 <n

solves the equation (14.1). Thus the function u(x, t) = v(x, t)e−ct solves the system of equation.

16. Define u (x, t) = u(x, t) − t for each  > 0. We will prove first that u attains it’s maximum
on ΓT . For sake of contradiction, assume that u has maximum at (x0 , t0 ) ∈ UT = U × (0, T ]. Define
v : U → < with v(x) = u (x, t0 ). Then v has maximum at x0 ∈ U . Since u ∈ C12 (UT ) ∩ C(UT ), we have
v ∈ C12 (U ) ∩ C(U ). Thus we must have Hess(v)(x0 ) is a negative definite matrix. So we must have

∆v(x0 ) = tr(Hess(v)(x0 )) < 0 (16.1)

We’ll consider two cases.

Case 1: t0 < T
∂
Since (x0 , t0 ) ∈ U × (0, T ) is global maximum, we have 0 0
∂t u (x , t ) = 0. So

∂
0= u (x0 , t0 ) = ut (x0 , t0 ) − 
∂t
Therefore we have
ut (x0 , t0 ) =  (16.2)
But

∆v(x0 ) = ∆x u (x0 , t0 )
= ∆u(x0 , t0 )
= ut (x0 , t0 ) by (16.2)
= 
> 0

using (16.1), we get a contradiction.

Case 2: t0 = T
Since u (x0 , t0 ) is global maximum, there exist 0 < T 0 < T such that ∂ 0
∂t u (x , t) ≥ 0 for all t ∈ (T 0 , T ).
Since u solve heat equation
∂
0 ≤ u (x0 , t)
∂t
= ut (x0 , t) − 
= ∆u(x0 , t) − 

so
∆u(x0 , t) ≥  (16.3)
for all t ∈ (T 0 , T ). Also we know that ∆v(x0 ) = ∆x u (x0 , T ) = ∆x u(x0 , T ). Since u ∈ C12 (UT ) ∩ C(UT ),
∆x u is continuous. Using (16.3), we conclude that ∆v(x0 ) ≥  > 0, which is a contradiction with (16.1).
Corollary: For all  > 0
max u = max u
UT ΓT

15
 By using the corollary, let’s prove that u also attains its maximum on ΓT . Since u = u + t

max u ≤ max (u + t)

UT UT
≤ max u + max t (use corollary)
UT UT
≤ max u + T
ΓT
≤ max u + T
ΓT

Let  → 0, then max u ≤ max u. Since ΓT ⊂ U T , we conclude max u = max u

UT ΓT UT ΓT

17. NOTE : The reader should read the proof of THEOREM 3(A mean-value property for the
heat equation) at 2.3.2 before start reading the solution.

(a)We modify the proof of the THEOREM 3. Put v instead of u in the proof. We know that

φ0 (r) = A+B
Z Z n
1 2n X
= −4nv s ψ − vy yi dyds
rn+1 E(r) s i=1 i

Since ψ defined to be

n |y|2
ψ = − log(−4πs) + + nlog r = log(Φ(y, −s)rn )
2 4s
ψ ≥ 0 in E(r) because Φ(y, −s)rn ≥ 1 in E(r). Thus 4nψ(vs − ∆v) ≤ 0, −4nψvs ≥ −4n∆v. Then we
have inequality
Z Z n
1 2n X
φ0 (r) = −4nvs ψ − vy yi dyds
rn+1 E(r) s i=1 i
Z Z n
1 2n X
≥ −4n∆vψ − vy yi dyds
rn+1 E(r) s i=1 i
= 0

according to the proof of the Theorem 3. So we have

φ(r) ≥ φ()

for all r >  > 0. But we know

lim φ() = 4v(0, 0)
→0

So we have inequality
|y|2
Z Z
1
v(y, s) dyds = φ(r) ≥ 4v(0, 0)
rn E(r) s2
At first, the book choses x = t = 0 without losing generality, so we

|x − y|2
Z Z
1
v(y, s) dyds ≥ v(x, t)
4rn E(x,t;r) (t − s)2

16
 (b) We modify the proof of the THEOREM 4 at 2.3.2 (Strong Maximum Principle for the Heat
Equation). In the proof, put v instead of u. Assume that there exist a point (x0 , t0 ) ∈ UT with
v(x0 , t0 ) = M := max u. Then for sufficiently small r > 0, E(x0 , t0 ; r) ⊂ UT , thus
UT

M = v(x0 , t0 )
|x − y|2
Z Z
1
≤ v(y, s) dyds
4rn E(x0 ,t0 ;r) (t − s)2
|x − y|2
Z Z
1
≤ M n 2
dyds
4r E(x0 ,t0 ;r) (t − s)
= M
so we must have v(y, s) = M for (y, s) ∈ E(x0 , t0 ; r). The rest of the proof is the same.

(c) Follow the equations

vt = φ0 (u)ut
vxi = φ0 (u)uxi
vxi xi = φ00 (u)u2xi + φ0 (u)uxi xi
Thus we have
n
X
vt − ∆v = φ0 (u)ut − φ0 (u)∆u − φ00 (u) u2xi ≤ 0
i=1
2
Pn
(d) v = |Du| + u2t = u2t + 2
j=1 uxj , so we
n
X
vt = 2ut utt + 2 uxj uxj t (17.1)
j=1

n
X
vxi = 2ut utxi + 2 uxj uxj xi
j=1
n
X
vxi xi = 2 ut utxi xi + u2txi + uxj uxj xi xi + u2xj xi


j=1

so we have
n
X
∆v = vx i x i
i=1
n 
X n
X 
= 2 ut utxi xi + u2txi + uxj uxj xi xi + u2xj xi
i=1 j=1
X n n
X n
X 
= 2 u2txi + u2xi xj + ut ∆ut + uxj ∆uxj (17.2)
i=1 i,j=1 j=1

since u solves heat equation, so does ut and uxi for i = 1, 2, ..., n. This is true since differentiation
commutes. Thus (17.2) becomes
X n n
X Xn 
2 2
∆v = 2 utxi + uxi xj + ut ∆ut + uxj ∆uxj
i=1 i,j=1 j=1
n
X Xn  n
X
= 2 u2txi + u2xi xj + 2ut utt + 2 uxj uxj t
i=1 i,j=1 j=1
n
X
≥ 2ut utt + 2 uxj uxj t
j=1
= vt (by 17.1)

17
thus we have ∆v ≥ vt , so v is a subsolution.

18. Since u is smooth, differentiation commutes, so

vtt − ∆v = uttt − ∆ut

∂
= (utt − ∆u)
∂t
= 0

Moreover, v = ut = h on <n × {t = 0} and

∂ ∂
vt (x, 0) = ut (x, 0) = h(x) = 0
∂t ∂t
so vt = 0 on <n × {t = 0}.

∂
19. (a) We have 0 = uxy = ∂x uy , so uy is constant as x changes, thus uy only depends on y. Say
uy (x, y) = c(y). By Fundamental Theorem of Calculus, we have
Z y Z y
u(x, y) − u(x, 0) = uy (x, z)dz = c(z)dz (19.1)
0 0
Ry
Let F (x) := u(x, 0) and G(y) := 0
c(z)dz. By 19.1, we have u(x, y) = F (x) + G(y).

(b) ξ and η are defined to be ξ = ξ(x, t) = x + t and η = η(x, t) = x − t. We have an equality

ξ+η ξ−η
u(x, t) = u( , )
2 2
Thus we have
1 ξ+η ξ−η 1 ξ+η ξ−η
uξ = uξ (x, t) = ux ( , ) + ut ( , )
2 2 2 2 2 2
1 1 1 1 1 1 1
uξη = uξη (x, t) = ( uxx − uxt ) + ( utx − utt ) = (uxx − utt )
2 2 2 2 2 2 4
Now it is clear to see
utt − uxx = 0 ⇐⇒ uξη = 0

18
 SOLUTIONS OF CHAPTER 3

1. Trivially ut (x, t, a, b) = −H(a) and Du(x, t, a, b) = a so ut + H(Du) = −H(a) + H(a) = 0.

Denote y = (x, t) ∈ <n+1 , c = (a, b) ∈ <n+1 and u(y, c) = u(x, t, a, b). Only thing left to prove is
2
rank(Dc u, Dyc u) = n + 1
 
x1 1 0 0 ··· 0 0
 x2
 0 1 0 ··· 0 0 
2  .. .
.. .
.. .
.. . .. .
.. .. 
Dc u, Dyc u =  . .
 
xn 0 0 0 ··· 1 0
1 −Ha1 (a) −Ha2 (a) −Ha3 (a) · · · −Han (a) 0
2 2
The matrix Dc u, Dya u (i.e. the left (n + 1) × (n + 1) part of the matrix Dc u, Dyc u) has determinant
(−1)n , so it has rank n + 1, so we are done.

2. (i) for u(x; a) = x1 + a2 x2 − 2a

1
Da u(x; a) = 2ax2 − 2 so for φ(x) = φ(x1 , x2 ) = x2 , the function solves the equation Da u(x; φ(x)) = 0.
The envelope is
1 1
v(x) = u(x; φ(x)) = u(x1 , x2 ; ) = x1 −
x2 x2

(ii) for u(x; a) = 2a1 x1 + 2a2 x2 − x3 + a21 + a22

Da u(x; a) = (2x1 + 2a1 , 2x2 + 2a2 ) so the function φ(x) = φ(x1 , x2 , x3 ) = (−x1 , −x2 ) solves the equation
Da u(x; φ(x)) = 0. The envelope is

v(x) = u(x; φ(x)) = −x21 − x22 + x3

3. (a) Let’s write the equation

G(x, u, a) = G(x, u(x, a), a) = 0

Gxi (x, u, a) + Gz (x, u, a)uxi = 0

Pn
for the function G(x, z, a) = j=1 aj x2j + z 3 The equation is

2ai xi + 3u2 uxi = 0

Therefore
3 2
u uxi xi = −ai x2i
2
if we add up for all i’s, we find that
n n
3 2 X3 X
u Du · x = u2 uxi xi = −ai x2i = u3
2 i=1
2 i=1

Therefore u(x, a) solves the PDE

3
F (Du, u, x) = Du · x − u = 0
2
(b) to be continued...

4. (a) Let’s write y = (x, t) and consider characteristic equations for y(s) = (x(s), t(s)), q(s) =
(Dux (y(s)), ut (y(s))) and z(s) = u(y(s)) Then the PDE becomes

F (q, z, y) = q · (b, 1) − f (y) = 0

19
Therefore
Dq F (q, z, y) = (b, 1)
Dy F (q, z, y) = Dy f (y)
Dz F (q, z, y) = 0
Then the characteristic equations are
.
y(s) = (b, 1)
.
z 0 (s) = z(s) = (b, 1) · q(s) = b · Du(y(s)) + ut (y(s)) = f (y(s))
.
q(s) = −Dy f (y(s))
(b) Let’s choose y(s) = (bs + c, s) for some constant c ∈ <n . We need to find proper s and s to have
equality y(s) = (x, t). Clearly t = s and c = x − bt. Therefore

y(s) = (x + b(s − t), s)

By Fundamental theorem of calculus,

Z t
z(t) − z(0) = z 0 (s)ds
0
Z t
= f (z(s))ds
0
Z t
= f (x + b(s − t), s)ds
0

But z(0) = u(y(0)) = u(x − bt, 0) = g(x − bt), so

Z t
u(x, t) = z(t) = g(x − bt) + f (x + b(s − t), s)ds
0

which agrees with the formula in 2.1.2

5. (a) Let’s read the equation as

F (Du, u, x) = x · Du − 2u = 0

so the function F (p, z, x) satisfies

F (p, z, x) = p · x − 2z = 0
Dp F (p, z, x) = x
Dx F (p, z, x) = p
Dz F (p, z, x) = −2
Therefore the characteristic equations are
.
x(s) = x(s)
.
z(s) = x(s) · p(s) = 2z(s)
.
p(s) = −p(s) − (−2)p(s) = p(s)
Let’s choose
x(s) = (x1 (s), x2 (s)) = (C1 es , es )
for some constant C1 . Moreover, z(s) must be

z(s) = C2 e2s

20
but
g(C1 ) = u(C1 , 1) = u(x(0)) = z(0) = C2
so
z(s) = g(C1 )e2s
For given (a, b), if we choose s = log(b) and C1 = ab , we find that

x(s) = x(log(b)) = (C1 b, b) = (a, b)

Thus
a
u(a, b) = u(x(log(b))) = z(log(b)) = g(C1 )e2log(b) = g( )b2
b
x1 2
Finally, let’s prove that the function u(x1 , x2 ) = g( x2 )x2 indeed solves the equation system

x · Du − 2u = 0

u(x1 , 1) = g(x1 )
The second equation is trivial, for the first equation
x1 1 2 x1 x1 −x1 x1
x1 ux1 + x2 ux2 − 2u = x1 g 0 ( ) x + x2 (g( )2x2 + g 0 ( ) 2 x22 ) − 2g( )x22 = 0
x2 x2 2 x2 x2 x2 x2

(b) We can read the equation as F (Du, u, x) = 0 where

F (p, z, x) = p · (x1 , 2x2 , 1) − 3z = 0

the derivatives are

Dp F (p, z, x) = (x1 , 2x2 , 1)
Dx F (p, z, x) = (p1 , 2p2 , 0)
Dz F (p, z, x) = −3
so the characteristic equations are
.
x(s) = (x1 , 2x2 , 1)
.
z(s) = Dp F (p, z, x) · p = 3z(s)
.
p(s) = −(p1 , 2p2 , 0) − (−3)p = (2p1 , p2 , 3p3 )
Considering equations, we can choose x(s) = (C1 es , C2 e2s , s) and z(s) = C3 e3s . Putting s = 0

C3 = z(0) = u(x(0)) = u(C1 , C2 , 0) = g(C1 , C2 )

Let (x1 , x2 , x3 ) be given. Only thing left is to choose proper C1 , C2 and s such that x(s) = (x1 , x2 , x3 ).
Clearly s = x3 , C1 = x1 e−x3 and C2 = x2 e−2x3 . Finally

u(x1 , x2 , x3 ) = u(x(x3 )) = g(C1 , C2 )e3x3 = g(x1 e−x3 , x2 e−2x3 )e3x3

Let’s prove u indeed solves the PDE.

x1 ux1 = x1 gx1 (x1 e−x3 , x2 e−2x3 )e2x3

2x2 ux2 = 2x2 gx2 (x1 e−x3 , x2 e−2x3 )ex3

ux3 = −gx1 x1 e−x1 e3x3 − gx2 2x2 e−2x3 )e3x3 + 3ge3x3 = −x1 gx1 e2x2 − 2x2 ex3 + 3ge3x3
Therefore F (Du, u, x) = x1 ux1 + 2x2 ux2 + ux3 − 3u = 0

(c) We can read the equation as F (Du, u, x) = 0 where

F (p, z, x) = p · (z, 1) − 1

21
the derivatives are
Dp F (p, z, x) = (z, 1)
Dx F (p, z, x) = 0
Dz F (p, z, x) = p1
thus the characteristic equations are
.
x(s) = (z(s), 1)
.
z(s) = 1
Choose z(s) = s + c1 and x2 (s) = s. Since x01 (s) = z(s) = s + c1 , choose x1 (s) = 21 (x + c1 )2 + c2 . Assume
that for some s, x1 (s) = s = x2 (s), then
1
s + c1 = z(s) = u(x(s)) = u(s, s) = s
2
so s must be −2c1 . Let’s put s = −2c1 in the equation x1 (s) = s to find c2 .
1
−2c1 = x(−2c1 ) = c21 + c2

. Thus c2 = −2c1 − 12 c21 . Let’s write c = c1 to simplify the notation. In summation, we find that

1 1 1
x(s) = ( (s + c)2 − c2 − 2c, s) = ( s2 + sc − 2c, s)
2 2 2
and
1
u( s2 + sc − 2c, s) = u(x(s)) = z(s) = s + c
2
We need to find proper s and c for given (x1 , x2 ) such that x(s) = (x1 , x2 ).
1
(x1 , x2 ) = x(s) = ( s2 + sc − 2c, s)
2
so x2 = s and c(s − 2) = x1 − 21 s2 = x1 − 12 x22 . Finally, for x2 6= 2,

2x1 − x22
c=
2(x2 − 2)

and
2x1 − x22 x2 − 4x2 + 2x1
u(x1 , x2 ) = u(x(x2 ))z(x2 ) = x2 + c = x2 + = 2
2(x2 − 2) 2(x2 − 2)
x22 −4x2 +2x1 x1
Let’s show that u = 2(x2 −2) is indeed a solution. For x1 = x2 6= 2, clearly u(x1 , x1 ) = 2 . Also

1
ux1 (x1 , x2 ) =
x2 − 2

(2x2 − 4)2(x2 − 2) − 2(x22 − 4x2 + 2x1 )

ux2 (x1 , x2 ) = = 1 − uux1
(2(x2 − 2))2
Thus uux1 + ux2 = 1, so we are done.

6. (a) Let x = (x1 , x2 , ..., xn ) and b = (b1 , b2 , ..., bn )

n
∂xi X Y
J(s, x, t) = detDx x(s, x, t) = det( )= (−1)sgn(n) xjxσ (j)
∂xj j=1
σ∈Sn

22
 Qn
therefore J is a linear combination of functions fσ (s, x, t) = (−1)sgn(σ) j=1 xjxσ (j) (s, x, t) where σ ∈ Sn .
We need to prove that
X ∂fσ X
Js = = div(b(x)) fσ
∂s
σ∈Sn σ∈Sn

Let’s start to compute RHS

n n n
∂fσ ∂ Y j X ∂ i Y
= (−1)sgn(σ) xxσ (j) = (−1)sgn(σ) xxσ (i) xjxσ (j)
∂s ∂s j=1 i=1
∂s
j=1, j6=i

Since x is smooth, we know that differentiation commutes, so

∂ i ∂
x = xi
∂s xσ (i) ∂xσ(i) s

By the given condition ∂x

∂s = ẋ = b(x) we know that xis = bi (x) for i = 1, 2, ..., n. Thus
n
∂ i ∂ ∂ X
xxσ (i) = xis = bi (x) = bixk xkxσ(i)
∂s ∂xσ(i) ∂xσ(i)
k=1

Therefore we find that

n
n  X  n 
∂fσ X Y
= (−1)sgn(σ) bixk xkxσ(i) xjxσ (j)
∂s i=1 k=1 j=1, j6=i

Thus
X ∂fσ
Js =
∂s
σ∈Sn

X X n
n  X  n
Y !
= (−1) sgn(σ)
bixk xkxσ(i) xjxσ (j)
σ∈Sn i=1 k=1 j=1, j6=i
Xn X n  X n
Y 
= bixk (−1)sgn(σ) xkxσ(i) xjxσ (j)
i=1 k=1 σ∈Sn j=1, j6=i

but
X n
Y
(−1)sgn(σ) xkxσ(i) xjxσ (j) = detDx (x1 , x, ..., xi−1 , xk , xi+1 , ..., xn )
σ∈Sn j=1, j6=i

which is equal to J for i = k and 0 for i 6= k. Therefore

n
X
Js = bixi J = div(b(x))J
i=1

(b) Note that the problem is written wrong, errata! We’ll use characteristic equations.

ut + div(ub) = ut + Du · b + udiv(b)

We can read the equation as

F (Du, ut , u, x, t) = 0
where for q, y ∈ <n+1 and z ∈ <

F (q, z, y) = q · (b, 1) + zdiv(b) = 0

The derivatives are

Dq F (q, z, y) = (b, 1)

23
 Dz F (q, z, y) = div(b)
Dy F (q, z, y) = 0
Thus the characteristic equations are
ẏ(s) = (b, 1)
zs (s) = ż(s) = (b, 1) · q = −z(s)div(b)
Since we are given the fact that ẋ = b(x), we can choose y(s) = (x(s), s)

claim: z(s)J(s) is a constant function.

proof: We use Euler Formula, which we proved in (a)

∂
(zJ) = zs J + zJs = −zdiv(b)J + zdiv(b)J = 0
∂s
For given (x, t) ∈ <n × [0, ∞), choose x(s) = x(s, x, t). Since zJ is constant
z(t)J(t) = z(0)J(0)
but
z(t) = u(x(t), t) = u(x(t, x, t), t) = u(x, t)
J(t, x, t) = detDx x(t, x, t) = detDx (x) = 1
so
u(x, t) = z(t)J(t)
= z(0)J(0)
= u(x(0, x, t), 0)J(0, x, t)
= g(x(0, x, t))J(0, x, t)
since u = g on <n × {t = 0}

7. NOTE: Before start reading the solution, the reader should read appendix C.1, and section
3.2.3, especially Lemma 1.

Define functions Φ and Ψ as in appendix C.1. Define v(y) = u(Ψ(y)) : V → < as in 3.1.3.a. Then by
(26), we have an equality
F (Dv(y)DΦ(Ψ(y)), v(y), Ψ(y)) = F (Du(Ψ(y)), u(Ψ(y)), Ψ(y)) = 0 (7.1)
Define function G : <2n+1 → < such that
G(q, z, y) = F (qDΦ(Ψ(y)), z, Ψ(y)) (7.2)
By 7.1, we have
G(Dv(y), v(y), y) = 0
0
Moreover V := Φ(U ) is flat near x ∈ Γ. Thus by LEMMA 1, we have noncharacteristic boundary
condition
Gqn (q 0 , z 0 , y 0 ) 6= 0
where y 0 = Φ(x0 ) and q 0 = p0 DΨ(y 0 ). By the definition of Φ
 
1 0 0 ··· 0 0
 0
 1 0 ··· 0 0 
 .. .
.. .
.. .. .. .. 
DΦ(x) =  . . . .
 
 0 0 0 ··· 1 0
−γx1 −γx2 −γx3 · · · −γxn−1 1

24
Let ci (x) define the ith column of the matrix. Then we have
(
∂ −γxj (x1 , x2 , ..., xn−1 ) for j = 1, 2, ..., n
(q · cj (x)T ) =
∂qn 1 for j = n
By equality in 7.2
∂
Gqn (q, z, y) = F (qDΦ(Ψ(y)), z, Ψ(y))
∂qn
n
X ∂
= Fqj (qDΦ(Ψ(y)), z, Ψ(y)) (q · ci (Ψ(y)T ) (7.3)
j=1
∂q n

In particular, when x is near x0 , the equation xn − γ(x1 , x2 , ..., xn−1 ) = 0 defines the boundary for U .
Thus v(x0 ) = (−γx1 , −γx2 , ..., −γxn−1 , 1) is the normal vector at the point x0 . Using 7.3
n
X ∂ 0
Gqn (q 0 , z 0 , y 0 ) = Fqj (q 0 DΦ(Ψ(y 0 )), z 0 , Ψ(y 0 )) (q DΦ(Ψ(y 0 )))
j=1
∂qj

But by definitions of y 0 and q 0

Ψ(y 0 ) = Ψ(Φ(x0 ))x0
q 0 DΦ(Ψ(y 0 )) = p0 DΨ(x0 )DΦ(x0 ) = p0
Thus
Gqn (q 0 , z 0 , y 0 ) = Fp (p0 , z 0 , x0 ) · v(x0 )
Therefore the noncharacteristic condition becomes

Fp (p0 , z 0 , x0 ) · v(x0 ) 6= 0

when Γ is not flat near x0 .

8. (Note that there is an Errata in the statement of problem). We shall prove that the function u
which satisfies the condition

u = u(x, t) = g(x − tF0 (u)) = g(x − tF0 (u(x, t)))

provides implicit solution of the conservation law

ut + divF(u) = ut + Du · F0 = 0

Let’s calculate ut (x, t)

∂
ut (x, t) = g(x − tF0 (u))
∂t
n
X ∂
= gxi (x − tF0 (u)) (xi − t(Fi )0 (u))
i=1
∂t
n
X
= gxi (x − tF0 (u))(−(Fi )0 (u) − t(Fi )00 (u)ut )
i=1
= −Dg(x − tF0 (u)) · F0 (u) − tut Dg(x − tF0 (u)) · F00 (u)

Therefore  
ut (x, t) 1 + tDg(x − tF0 (u)) · F00 (u) = −Dg(x − tF0 (u)) · F0 (u)

25
Let’s define γ(x, t) = 1 + tDg(x − tF0 (u)) · F00 (u) to make equation looks simpler. Basically,

ut (x, t)γ(x, t) = −Dg(x − tF0 (u)) · F0 (u) (8.1)

Let’s calculate divF(u) = Du · F0 (u). For i = 1, 2, ..., n

∂
uxi (x, t) = g(x − tF0 (u))
∂xi
n
X ∂
= gxj (x − tF0 (u)) (xj − t(Fj )0 (u))
j=1
∂x i
n
X
= gxi (x − tF0 (u))(1 − t(Fi )00 (u)uxi ) + gxj (x − tF0 (u))(−t(Fj )00 (u)uxi )
j=1,j6=i
n
X
= gxi (x − tF0 (u)) − tuxi gxj (x − tF0 (u))(Fj )00 (u)
j=1

= gxi (x − tF0 (u)) − tuxi Du(x − tF0 (u))F00 (u)

Therefore
γ(x, t)uxi (x, t) = gxi (x − tF0 (u))
Thus

γ(x, t)divF(u) = γ(x, t)Du · F0 (u)

Xn
= γ(x, t)uxi (x, t)(Fi )0 (u)
i=1
n
X
= gxi (x − tF0 (u))(Fi )0 (u)
i=1
= Dg(x − tF0 (u))F0 (u)
 
Using (8.1), we conclude that γ(x, t) ut + divF(u) = 0. If we are given that

γ(x, t) = 1 + tDg(x − tF0 (u)) · F00 (u) 6= 0

we can say that ut + divF(u) = 0, i.e. the function u solves the conservation law.

9. NOTE: Before start reading the solution, the reader should know the proof of the THEOREM
1 (Euler-Lagrange Equations) in section 3.3.1.

(a) Let’s define S := {y ∈ C ∞ ([0, t]; <n ) y(0) = y(t) = 0} and i : < → < such that
Z t
i(τ ) := I[x(·) + τ y(·)] = L(ẋ(s) + τ ẏ(s), x(s) + τ y(s))ds
0

Clearly for any y ∈ S and τ ∈ <, x + τ y ∈ A. Therefore i(τ ) has minimum at τ = 0 for any fixed y ∈ P,
so we know i0 (τ ) = 0. As we see in the proof of the THEOREM 1
n
Z tX
0
i (τ ) = Lvi (ẋ + τ ẏ, x + τ y)ẏi + Lxi (ẋ + τ ẏ, x + τ y)yi
0 i=1

Thus at τ = 0
n
Z tX
0 = i(0) = Lvi (ẋ, x)ẏi + Lxi (ẋ, x)yi ds (9.1)
0 i=1

26
By integration by parts

0 = Lvi (ẋ(t), x(t))yi (t) − Lvi (ẋ(0), x(0))yi (0)

Z t 
∂ i
= Lv (ẋ(s), x(s))y (s) ds
0 ∂s i
Z t   
∂
= Lvi (ẋ(s), x(s) yi (s) + Lvi (ẋ(s), x(s))ẏi (s) ds
0 ∂s

Therefore Z t Z t 
∂
Lvi (ẋ(s), x(s))ẏi (s)ds = − Lvi (ẋ(s), x(s)) yi (s)ds
0 0 ∂s
If we put this equality in (9.1), we find that
n Z t 
X ∂
0= − Lv (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) yi ds
i=1 0 ∂s i

Since this equality is true for any smooth function y ∈ S, we conclude that
∂
− Lv (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) = 0
∂s i
for all i = 1, 2, .., n. In other words, x(s) satisfies Euler-Lagrange Equations.

(b) Now consider the set P := {y ∈ C ∞ ([0, t]; <n ) y(t) = 0}. Clearly for any y ∈ P and τ ∈ <,
x + τ y ∈ A, so i(τ ) has minimum at τ = 0 for any fixed y ∈ P. Thus i0 (0) = 0. By integration by parts
Z t   
∂
−Lvi (ẋ(0), x(0))yi (0) = Lvi (ẋ(s), x(s) yi (s) + Lvi (ẋ(s), x(s))ẏi (s) ds
0 ∂s

Therefore
Z t Z t 
i ∂
Lvi (ẋ(s), x(s))ẏ (s)ds = − Lv (ẋ(s), x(s)) yi (s)ds − Lvi (ẋ(0), x(0))yi (0)
0 0 ∂s i

If we put this equality in (9.1), we find that

0 = i0 (0)
n Z t 
X ∂
= − Lvi (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) yi ds − Dv L(ẋ(0), x(0)) · y(0)
i=1 0 ∂s

∂
But − ∂s Lvi (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) = 0 by (a), so Dv L(ẋ(0), x(0)) · y(0) = 0 for any y ∈ P, thus
Dv L(ẋ(0), x(0)) = 0

(c) For any y ∈ S and τ ∈ <, x + τ y ∈ A and g(x(0)) = g(x(0) + τ y(0)). Define j : < → < such
that Z t
j(τ ) = i(τ ) + g(x(0) + τ y(0)) = L(ẋ(s) + τ ẏ(s), x(s) + τ y(s))ds + g(x(0) + τ y(0))
0

By similar reasoning as in (a),

n Z t 
X ∂
0 = j 0 (0) = − Lvi (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) yi ds
i=1 0 ∂s

notice that we got rid of g(x(0) + τ y(0)) since it is constant as τ changes. Therefore x satisfies Euler-
Lagrange equations.

27
 Now consider y ∈ P and define j(τ ) in the same way. Then

∂
j 0 (τ ) = i0 (τ ) + g(x(0) + τ y(0)) = i0 (τ ) + Dg(x(0) + τ y(0)) · y(0)
∂τ
So at τ = 0, j 0 (0) = i0 (0) + Dg(x(0)) · y(0). But from (b), we know
n Z t 
0
X ∂
i (0) = − Lv (ẋ(s), x(s)) + Lxi (ẋ(s), x(s)) yi ds − Dv L(ẋ(0), x(0)) · y(0)
i=1 0 ∂s i
= −Dv L(ẋ(0), x(0)) · y(0)

since x satisfies Euler-Lagrange Equations. Therefore

0 = j 0 (0)
= −Dv L(ẋ(0), x(0)) · y(0) + Dg(x(0)) · y(0)
 
= Dg(x(0)) − Dv L(ẋ(0), x(0)) · y(0)

Since the equality is true for any y ∈ P, we must have Dv L(ẋ(0), x(0)) = Dg(x(0)) as an initial condition
for minimizer x.

10. (a) Firstly, let’s remember Y oung 0 s Inequality: Let p, q ∈ (1, ∞) and a, b ∈ [0, ∞), then the
inequality
ap bp
+ ≥ ab
p p
holds. By the definition,

|p|r
L(v) = sup {v · p − H(p)} = sup {v · p − }
p∈<n p∈<n r

|p|r |v|s
By Y oung 0 s Inequality and Cauchy 0 s Inequality, we know that r + s ≥ |p||v| ≥ p · v, so

|p|r |v|s
L(v) = sup {v · p − }≤ (10.1)
p∈<n r s
s−r
. If we put p = v|v| r

s−r r
|p|r s−r v|v| r
v·p− = v · v|v| r −
r r
s+r |v|s
= |v| r − (10.2)
r
s |v|s
= |v| −
r
|v|s
=
s
s+r sr |v|s |v|s
At (10.2), we used r = r = s. Thus L(v) ≥ s . Combining (10.1), we conclude that L(v) = s .

(b) We need to determine

L(v) := H ∗ (v) = sup {v · p − H(p)}

p∈<n

28
(For the sake of the easiness of the notation, we assume that v and p are column vectors) Fix v and
define function f = fv : <n → < such that
n n
X 1 X
f (p) = v · p − H(p) = (vi − bi )pi − aij pi pj
i=1
2 i,j=1
Pn
Notice two things: f ∈ C ∞ (<n ; R), and i,j=1 aij pi pj = pT Ap. Assume that we also know f is bounded.
Then f must have maximum, i.e there exist p∗ such that f (p∗ ) = sup f (p) = L(v). Since f is smooth,
p∈<n
p∗ must be a critical point, so we have Df (p∗ ) = 0. Let ci denote ith column of the matrix A (and also
ith row since A is a symmetric matrix). Then by simple calculation
∂
f (p∗ ) = (vi − bi ) − ri · p∗
∂pi
Thus
Df (p∗ ) = v − b − Ap∗ (10.3)
Since A is a positive definite matrix, it is also invertible. So we can solve the equation (10.3) for
p∗ = A−1 (v − b). Moreover, for this particular choice of p∗ , Hess(f )(p∗ ) = −A, which is negative
definite, so p is actually local maximum. Since it is only critical point and we assumed f to be bounded,
p∗ must be the global maximum. Therefore

L(v) = f (p∗ )
1
= (v − b) · p∗ − (p∗ )T Ap∗
2
1
= (v − b) · (A−1 (v − b)) − (A−1 (v − b))T A(A−1 (v − b)) (10.4)
2
1
= (v − b)T (A−1 (v − b)) − ((v − b)T A−1 )A(A−1 (v − b))
2
1
= (v − b)T A−1 (v − b)
2
At (10.4), we used that
(A−1 (v − b)) = (v − b)T (A−1 )T
and

(A−1 )T = A−1 A(A−1 )T

= A−1 AT (A−1 )T
= A−1 (A−1 A)T
= A−1 (I)T
= A−1

(briefly we proved that if A is symmetric, so is A−1 ) We only need to prove that f is bounded. Since
the function x → xT Ax is continuous and the set B(0, 1) ⊂ <n is compact, xT Ax is bounded on the set
p
B(0, 1). Assume xT Ax > M for |x| = 1. Since A is positive definite, we can choose M > 0. Let p1 = |p|

f (p) = (v − b) · p − pT Ap
≤ |v − b||p| − |p|2 pT1 Ap1 (by Cauchy’s Inequality)
< |v − b||p| − M |p|2
1
= M |p|(|v − b| − M |p|)
M
1 |v − b|2
≤ (by Aritmetic-Geometric Mean Inequality)
M 4
Thus f is bounded, so we are done.

29
 11. Assume that v ∈ ∂H(p). Let’s prove p · v = H(p) + L(v). For any r ∈ <n

r · v − H(r) ≤ r · v − (H(p) + v · (r − p)) = v · p − H(p)

Thus
L(v) = sup {v · r − H(r)} ≤ v · p − H(p)
r∈<n

but for r = p, v · r − H(r) = v · p − H(p), so L(v) = v · p − H(p).

Now assume that p · v = H(p) + L(v), we’ll prove v ∈ ∂H(p). For any r ∈ <n

v · p − H(p) = L(v)
= sup {v · s − H(s)}
s∈<n
≥ v · r − H(r)

So
˙ − p)
H(r) ≥ H(p) + v (r
thus v ∈ ∂H(p). Since we have duality between L and H

v ∈ ∂H(p) ⇐⇒ p · v = H(p) + L(v) ⇐⇒ p ∈ ∂H(v)

12. First observe that

max {−H1 (p) − H2 (−p)} = − minn {H1 (p) + H2 (−p)}

p∈<n p∈<

so we need to prove
min {L1 (v) + L2 (v)} + minn {H1 (p) + H2 (−p)} = 0
v∈<n p∈<

By definition of H1 = L∗1 and H2 = L∗2 , for all p ∈ <n

H1 (p) + H2 (−p) = sup {v1 · p − L1 (v1 )} + sup {−v2 · p − L2 (v2 )}

v1 ∈<n v2 ∈<n
≥ sup {−L1 (v) − L2 (v)}
v∈<n
= − minn {L1 (v) + L2 (v)}
v∈<

Therefore we have
min {H1 (p) + H2 (−p)} ≥ − minn {L1 (v) + L2 (v)}
p∈<n v∈<

so we proved
min {L1 (v) + L2 (v)} + minn {H1 (p) + H2 (−p)} ≥ 0
v∈<n p∈<

Next we prove that it is also ≤ 0. First let’s remember some facts about Hamiltonian and Lagrangian.

Lemma (THEOREM 3 at 3.3.2) The three statements

p · v = L(v) + H(p)

p = DL(v)
v = DL(p)

30
are equivalent, provided that H = L∗ , L is differentiable at v and H is differentiable at p.

Assume that L1 (v) + L2 (v) obtains its min at v ∗ . In other words

min {L1 (v) + L2 (v)} = L1 (v ∗ ) + L2 (v ∗ )

v∈<n

Then we must have 0 = D(L1 + L2 )(v ∗ ) = DL1 (v ∗ ) + DL2 (v ∗ ) = 0. Choose

p∗ = DL1 (v ∗ ) = −DL2 (v ∗ )

then by the Lemma,

L1 (v ∗ ) + H1 (p∗ ) = v ∗ · p∗
L2 (v ∗ ) + H2 (−p∗ ) = −v ∗ · p∗
so

min {L1 (v) + L2 (v)} + minn {H1 (p) + H2 (−p)} ≤ L1 (v ∗ ) + L2 (v ∗ ) + H1 (p∗ ) + H2 (−p∗ )
v∈<n p∈<
= v ∗ · p∗ − v ∗ · p∗
= 0

thus we are done.

13. Define function f (y) = tL( x−y 0

t ) + g(y) and assume f takes its minimum at y . Then we
x−y −1 x−y
must have Df (y 0 ) = 0. But we know Df (y) = tDL( t ) t + Dg(y) = Dg(y) − DL( t ), so we have
0
DL( x−y 0
t ) = Dg(y ). Then by the Lemma from the problem 12

x − y0
DH(Dg(y 0 )) =
t
but we are given
|x − y 0 |
R ≥ |DH(Dg(y 0 ))| =
t
which means y 0 ∈ B(x, Rt). Thus we must have
x−y x−y
min {tL( ) + g(y)} = min {tL( ) + g(y)}
y∈<n t y∈B(x,Rt) t
so we are done.

14. We have Hamiltonian H(p) = |p|2 . Let’s determine L := H ∗ .

|v|2
Claim L(v) = 4 for all v ∈ <n

Proof By Arithmetic Mean-Geometric mean and Cauchy inequalities, we have

|v|2 |v|2
v · p − H(p) = v · p − |p|2 ≤ |v||p| − |p|2 ≤ ( + |p|2 ) − |p|2 =
4 4
|v|2 |v|2 |v|2
for all p ∈ <n . Thus we have L(v) ≤ 4 . But for p∗ = v2 , p∗ · v − |p∗ |2 = 4 , so we have L(v) = 4 .
Let’s apply Hopf-Lax formula for u.
x−y
u(x, t) = min {tL( ) + g(y)}
y∈<n t
|x − y|2
= minn { + g(y)}
y∈< 4t

31
Since g(y) = u(y, 0) is given in the problem as
(
0 if y ∈ E
g(y) =
∞ if y ∈
/E

|x − y|2
u(x, t) = minn { + g(y)}
y∈< 4t
|x − y|2
= min{ }
y∈E 4t
1
= dist(x, E)2
4t

15. We shall fill the gaps in the Lemma 4 from 3.3.3. Let’s prove (36). Define function H 0 (p) =
H(p) − θ2 |p|2 , then for all ξ ∈ <n
n
X
ξ T Hess(H 0 )(p)ξ = Hpi pj (p)ξi ξj − θ|ξ|2 ≥ 0
i,j=1

thus H 0 is a convex function. So we have

p1 + p2 p1 + p2 θ
H( ) = H 0( ) + |p1 + p2 |2
2 2 8
1 0 1 θ
≤ H (p1 ) + H 0 (p2 ) + |p1 + p2 |2 (since H 0 is convex)
2 2 8
1 θ 1 θ θ
= H(p1 ) − |p1 |2 + H(p2 ) − |p2 |2 + |p1 + p2 |2
2 4 2 4 8
1 1 θ
= H(p1 ) + H(p2 ) − (2|p1 | + 2|p2 | − |p1 + p2 |2 )
2 2
2 2 8
1 1 θ
= H(p1 ) + H(p2 ) − |p1 − p2 |2
2 2 8
so we proved (36). Let’s prove (37). Assume that

L(v1 ) = p1 v1 − H(p1 )

L(v2 ) = v2 p2 − H(p2 )
So
1 1 p1 v1 + p2 v2 1 1
L(v1 ) + L(v2 ) = − H(p1 ) − H(p2 ) (15.1)
2 2 2 2 2

v1 + v2 1 v1 + v2 p1 + p2 p1 + p2 1
L( ) + |v1 − v2 |2 ≥ − H( ) + |v1 − v2 |2
2 8θ 2 2 2 8θ
(v1 + v2 ) · (p1 + p2 ) 1 1 θ 1
≥ − H(p1 ) − H(p2 ) + |p1 − p2 |2 + |v1 − v2 |2 by (36)
4 2 2 8 8θ
1 1 (v1 − v2 )(p1 − p2 ) θ 1
= L(v1 ) + L(v2 ) − + |p1 − p2 | + |v1 − v2 |2 by (15.1)
2
2 2 4 8 8θ
1 1 (v1 − v2 ) · (p1 − p2 ) |p1 − p2 ||v1 − v2 |
≥ L(v1 ) + L(v2 ) − + (15.2)
2 2 4 4
1 1
≥ L(v1 ) + L(v2 ) by Cauchy’s inequality
2 2
Note that at (15.2), we used Arithmetic-Geometric Mean Inequality.

32
 16. Assume that
x − y1
u1 (x, t) = tL( ) + g 1 (y1 )
t
x − y2
u2 (x, t) = tL( ) + g 2 (y2 )
t
Thus
x−y x − y2
u1 (x, t) − u2 (x, t) = minn {tL( ) + g 1 (y)} − tL( ) + g 2 (y2 ) put y = y2
y∈< t t
x − y2 x − y2
≤ tL( ) + g 1 (y2 ) − tL( ) − g 2 (y2 )
t t
= g 1 (y2 ) − g 2 (y2 )

Similarly
u2 (x, t) − u1 (x, t) ≤ g 2 (y1 ) − g 1 (y1 )
So either |u2 (x, t) − u1 (x, t)| ≤ |g 2 (y1 ) − g 1 (y1 )| or |u1 (x, t) − u2 (x, t)| ≤ |g 1 (y2 ) − g 2 (y2 )|. In each case,

|u1 (x, t) − u2 (x, t)| ≤ sup|g 1 − g 2 |

<n

Bibliography

[1] Evans, Lawrence C., ”Partial Differential Equations”,

[2] Rudin, Walter, ”Principles of Mathematical Analysis”, 3rd ed., McGraw-Hill, Inc.,

33