Math222anotes PDF

Math 222a (Partial Differential Equations 1) Lecture
Notes
Izak Oltman
last updated: December 11, 2019
These are the lecture notes for a first semester graduate course in partial differential
equations taught at UC Berkeley by professor Sung-Jin Oh during the fall of 2019.
Contents
1 Introduction to PDE’s 3
1.1 Overview of PDE’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Basic Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 First Order (scalar) PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Method of Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Existence of Characteristic Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.7 noncharactersitic condition for kth order system . . . . . . . . . . . . . . . . . . 20
2 Distributions 25
2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 convergence of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.1 Approximation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Fundamental Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.1 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Laplace Equation 40
3.1 Cauchy Riemman Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Green’s Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4 Wave Equation 55
4.1 1 dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1
Contents
5 Fourier Transform 65
5.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.1 Duhamel’s Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6 Energy Method and Sobolev Spaces 75

6.1 Energy Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.2 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.4 Sobolev Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.5 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.6 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Appendices 92
A Midterm Review 92
A.1 First order scalar characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
A.1.1 Existence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
A.1.2 General Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
A.1.3 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
A.2 Noncharacteristic Condition for Kth order systems . . . . . . . . . . . . . . . . . 93
A.2.1 General Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.2.2 Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.3 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.3.1 Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
A.3.2 Convergence of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 95
A.3.3 Operations between distributions . . . . . . . . . . . . . . . . . . . . . . . 96
A.4 Fundamental Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
A.5 Laplace equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
B Final Review 97
B.1 Green’s Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
B.2 Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
B.3 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Index 101
– 2–
1. Introduction to PDE’s
Lecture 1 (8/29)
1 Introduction to PDE’s
The course outline will be:
1. basic issues (existence, uniqueness, regularity, ... ) will be demonstrated by means of
examples (Evans chapters 2,3,4)
2. theory of distributions and the Fourier transform (Evans chapter 4 and notes)
3. Sobolev spaces (Evans Chapter 6)
1.1 Overview of PDE’s

At a rudimentary level, a partial differential equation (PDE) is a functional equation, that
∂
involves partial differentiation ∂α . The unknown depends on several variables. PDEs
∂xα
are ubiquitous, they arise from many fields in science and mathematics.
1.2 Examples
Here are some fundamental examples of PDEs:
Example 1.2.1 (Laplace Equation). u Rd R or C, with ∆u 0, where ∆ is the
Laplace operator defined as ∆ Pdj 1 ∂j2
Example 1.2.2 (Wave Equation). ju 0 where j ∂02 ∆ (the d’Alembertian) and
u R1 d R or C. Here we have define x0

t as the time coordinate.
Both of these are scalar equations. The Laplace equation could model temperature distri-
bution at equilibrium or electrostatics which can measure the electric potential where there
is no charge. The wave equation models wave propagation (at finite speed). This could be
used to model elastic waves or acoustic waves or electromagnetic waves.
The Laplace equation is the prototypical example of elliptic PDEs while the wave is the
prototypical example of hyperbolic PDEs
Example 1.2.3 (Heat Equation). u R1 d R, C has the form ∂t ∆u 0

Example 1.2.4 (Schrodinger equation). u R1 d C with i∂t ∆u 0
The heat equation is a parabolic PDE while the Schrodinger equation is a dispersive
PDE.
– 3–
Example 1.2.5 (Cauchy-Riemann). u, v R2 R, we have:
∂x u ∂y v 0
∂y u ∂x v 0
The Cauchy-Riemman equations give necessary and sufficient conditions for a pair of
functions u, v to give us that uz iv z is holomorphic with z x iy. Also note that u
and v also satisfy the Laplace equation.
Example 1.2.6 (Maxwell Equations (vacuum)). Given E R1 3 R3 and B R1

3
R3 :
∂t E © B 0
∂t B © E 0
© E 0
© B 0
It turns out that jE j B 0.
Note that all these PDE’s can be written as:
F u 0
Note that F is linear for all these examples. So all equations stated so far are linear PDEs.
As we can write the operator of the thing equal to zero, we have a linear homogeneous
PDE. If instead we had F u f then we have a linear inhomogeneous. In the case we
have ∆u f we then have what is called the Poisson equation.
There are many nonlinear PDE examples:
Example 1.2.7 (Minimal Surface Equation). u Rd R with:
d
∂j u
Q ∂j » 0
j 1 1 SDuS2
with the notation D © ∂
This is to find a surface with minimal area that satisfies something. This is an example
of an elliptic PDE.
Example 1.2.8 (Korteng-de Vries (KDV) Equation). u R1 1 R with:
pt u ∂x3 u 6u∂x 0
This is an example of a dispersive nonlinear PDE.
Here are nonlinear systems of PDEs
– 4–
Example 1.2.9 (Ricci Flow). g a Riemannian metric (positive definite symmetric n n

matrix) with:
∂t g 2Ric g
Where Ric g 1 ∂ 2 g g 1 ∂g g 1 ∂g
Example 1.2.10 (Einstein Equation). g a Lorentzian metric which is a 1 3 1 3

symmetric metric with signature , , , (similar to diag 1, 1, 1, 1):
Ric g 0
1.3 Basic Problems

Sometimes there is no solution to a PDE, sometimes there is no unique solution, so we may
have to impose more information. One way to do this is to impose boundary values, which
is the Boundary Value Problem (BVP)
Example 1.3.1 (Dirichlet Problem). Given U òpen,bdd Rd , u Ū R with
¢̈
¨∆u f in U
¦
ü g on ∂U
¤̈
Example 1.3.2 (Neumann Problem). Given U and u as above, but:

¢̈
¨∆u f in U
¦
¨ν Du g on ∂U
¤̈
where ν is the outer normal to ∂U , and we require RU f dx R∂U gdA

Note that translated solutions are still solutions, therefore we say that solutions are
equivalent if u v c
Example 1.3.3 (Initial Value Problem (IVP)). For the heat equation:
¢̈
¨∂t u ∆u f x > 0, ª Rd
¦
ü g
¤̈
x > t 0 Rd
For Schrodinger
¢̈
ï∂t u ∆u f x > R Rd
¦
ü g
¤̈
x > t 0 Rd
For the wave equation:

¢̈ju f x > R Rd
¨
¨
¨
¦u g x > t 0 Rd
¨
¨
¨∂ u h x > t 0 Rd
¤̈ t
– 5–
Definition 1.3.1. We say a boundary value problem or initial value problem is wellposed
(Hadaward) if it satisfies
1. existence
2. uniqueness
3. continuous dependence on dataa
a PDE that fails this is called illposed
The regularity of a solution (if not we have a singularity). We also have asymptotics and
dynamics
Let’s introduce notation. First the multi-index notation: given α α 1 , . . . , α d then we

define:
∂α Dα ∂1α1 ∂2α2 ∂dαd
We call SαS P αi the order of α. The order of a PDE is the order of the highest order ∂ α
that appears in the equation.
Nonlinear PDEs are classified into three categories:
1. semilinear: 0 F u P α d aα∂ αu nonlinear stuff

S S
2. quasilinear: 0 F u P α k aαu, Du, . . . , Dk 1ux∂ αu

S S

rest
3. fully nonlinear.
Lecture 2 (9/3)
1.4 First Order (scalar) PDEs

Reference: Evans section 3.2
Here are three examples of first order PDEs (some of which are not in the text)
1. Positive example: ∂t at, x∂x u f where u R1 1 R, f R1 1 R, a R1 1 R

such that supx>R Sat, xS supx S∂x aS B M t @ ª, and f, a > C ª
2. Inviscid Burger equation ∂t u u ∂x u 0 with u R1 1 R.
This is an example of a nonlinear first order scalar PDE. Something goes wrong with
this equation.
a
that is the map from the data to the solution is continuous
– 6–
3. Lewy–Nirenberg example: ∂t it∂x u f with u, f R1 1 C.
It turns out that there exists a function f > C ª

such that there is no C 1 solution u
exists in any neighborhood of 0, 0
If we identify C by R2 , this is also an example of a linear first-order system of PDEs.
Let’s begin with a linear first order scalar PDE.
First suppose a 0 (this is known as a transport equation), then we have ∂t u f , so we

can integrate both sides to get:
ut, x S f dt C x
We could also have initial value problem:

¢̈
¨ ∂t u f
¦
ü0, x u0 x
¤̈
u0 x R0 f t , xdt
t
Then we have ut, x
Now let’s supose a is a general function which obeys the boundedness condition.
Theorem 1.4.1. Given u0 R R smooth, and f R1 1 R smooth. Then there exists a

unique solution u R1 1 R to the IVPa

∂t u a∂x u f
u0, x u0 x
Proof. (method of characteristics)
The idea is to try to make a change of variables to reduce this to the simpler case. In
other words, we want a change of variables t, x s, y such that: (
∂s ∂t a∂x (1)
We have:
∂t ∂x
∂s ∂t ∂x
∂s ∂s
∂t ∂x
∂y ∂t ∂x
∂y ∂y
a
note that this u is a global solution
– 7–
∂t ∂x
To get (1), we want: 1 and s, y ats, y , xs, y . The first equality gives us
∂s ∂s
that s t. So we would like to find:
∂x
s, y as, xs, y
∂s
This is a first order ODE. Now since SaS is bounded for each t, this allows us to conclude
that xs, y is always global (in s)a . Let’s allow x0, y y.
u0 y R0 f s , y ds
s
In our equation we had ∂s us, y f s, y therefore us, y
The last step is to show that x ( y is invertible. We have:

∂s ∂y x ∂x at, xs, y ∂y x @ M t∂y x
by comparison of ODEs, we have:
C e R0 M t dt ∂y x0, y A 0
t
S∂y xS

Therefore x is strictly increasing or strictly decreasing as we increase y, therefore we have

(
that x y is invertible.
Remark 1.4.1. This proof is an instance of the methods of characteristics. Because the
xs, y are called the characteristics and the ode involving x is called the characteristic ode.
Now let’s turn the the Inviscid Burger equation:
¢̈
¨∂t u u∂x u 0 t, x > R2
¦
ü0, x u0 x t, x > 0 R
¤̈
Theorem 1.4.2. For any smooth and compactly supported u0 R R, then there does not
exist a smooth global solution u R1 1 R with this data.

It turns out that there always exists a local solution, that is there exists δ A 0 such that
on 0, δ R, there exists a unique smooth solution u.
Proof. Suppose that u R1 1 R exists. Now let’s apply the method of characteristics to
the operator ∂t u∂x .
So we have ∂s x us, xs, y and us, y u0 y which is just constant. Visualy things
are pretty clear.
Or we could use:
0 ∂x ∂t u u∂x u ∂t ∂x u u∂x ∂x u ∂x u2
After a little, we get ∂s x x2 0, which is known as the Ricatti equation, which blows up in
finite time. This can be solved via separation of variables.
a
this is an ODE result, and requires the additional assumption that M is locally bounded. Then we apply
Picard’s theorem and prove by contradiction on a saturated solution
– 8–
Now let’s turn to Lewy-Nirenberg’s example:

∂t u it∂x u f
Theorem 1.4.3. There exists a smooth function f R1 1 C R2 with the property that
there does not exist any C 1 solution u to this PDE on any neighborhood of 0, 0
Proof. (requires complex analysis). Notation: let Br t, x > R1 1 t2 r2 @ r2 with ∂Br
the boundary. Take a smooth function f with the properties:
1. f t, x f t, x
2. for some sequence rn 0 (ex. 2 n) we want f SB

rn δn Brn δn
0 with RB rn
f dtdx x 0
Assume that such a C 1 solution to the PDE, u, exists on some neighborhood of 0, 0. Pick
1
n large enough so that Brn ` U . Then let ũt, x ut, x ut, x, then observe that ũ
2
is again a C 1 solution on Brn δ . But observe ũ is odd.

From now on u ũ. On the one hand, we have:
SBrn
∂t u it∂x u SB rn
f x0
By the divergence theorem, the left side is just:

t
º
t2 x2 u dσ
SB ∂t u it∂x u S∂B x itu
º 2
rn rn
t 2 x
Therefore uS∂Brn x 0. On the other hand, it turns out uS∂Brn 0.
By continuity, and the fact that f is zero on a bunch of thin annuli. On each annulus,
we have that f 0 ∂t u it∂x u. Let’s call the right half of the annulus U , that is:
U
rn δ 2 @ t2 x2 @ rn δ 2 , t A 0
On U , change the variables to t, x s, y

(
1 2
2
t , x so 0
1
t
∂t u i∂x u ∂s i∂y u
which is the Cauchy - Riemann operator. So now:
U
s, y ` R1 1
rn δn 2 @ 2s y 2 @ rn δn 2
We can then apply the Schwartz reflection to get a contradiction.
Lecture 3 (9/5)
Today we will cover section 3.2 from Evan’s book. Some notation things: we will use d
instead of n for the spacial dimension. We will also us xj as the jth coordinate (instead of
xj ).
– 9–
1.5 Method of Characteristics

We will be interested in the general, fully nonlinear, first order, scalar, PDE: F x, u, Du 0.
So F U R Rd R and u U R. And we will subject this to a boundary condition u g
on Γ ` ∂U . And we will assume that F and g are smooth. The coordinates in the domain
of F will be written as x, z, p1 , . . . , pd x, z, p.
Our first goal is to derive the characteristics of ODEs. To do this, we want to find the
solution ux by finding a suitable curve γ that connects x to x0 > Γ on which the evolution
of u can be computed.
Recall that the simplest example is ∂t u 0 with Γ 0 R, then the curves are just the
vertical straight lines.
Let’s parameterize our curve γ ` U with xs such that s > I. Then let the function value
along this curve be z s uxs. We would also like to keep track of the partials, by the
symbols pi s ∂i uxs
Assume that a smooth solution u exists. Then we have:

∂s pi s ṗi s Q ∂j ∂iuxsẋj s
j
Furthermore, we have:
0 ∂xi F x, ux, Dux
∂xi F x, ux, Dux ∂z F x, ux, Dux∂i u
Q∂p F x, ux, Dux∂i ∂j u (2)
j
j
So now we have second order terms in terms of first order terms. Now we will set:
ẋj s ∂pj F xs, uxs, Duxs ∂pj F xs, z s, ps (3)
Therefore (10) becomes:

0 ∂xi F x, ux, Dux ∂z F x, ux, Dux∂i u Q ẋj s∂i∂j u
j
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
ṗi s
So we get:
ṗi s ∂z F xs, z s, pspi s ∂xi F xs, z s, ps (4)
Lastly, we have:
ż s Q ∂iuxsẋis Q pis∂p F xs, z s, ps

i
(5)
i i
Then the collection of the three ODEs involving z, x, p are called the characteristic ODEs.
– 10 –
Theorem 1.5.1. If u is a smooth solution to F x, u, Du 0. Then on a curve x satisfying

(3):
z s uxs p i s ∂i uxs
will obey (4) and (5)

Example 1.5.1. Consider when F is linear. Suppose
0 F x, u, Du Q bj ∂ j u cu
j
where bj bj x and c c x .
In this case F x, z, p P j bj p j cz. Then if we plug in our characteristic equations, we

get that
ẋj s bj xs

ż s Q bj pj cxsz s
j
Remark 1.5.1.
1. The p equation is not needed
2. there is a hierarchy of the ODEs
Example 1.5.2 (Linear F).
¢̈ 1
¨x ux2 x2 ux1 u in U
¦
ü g on Γ
¤̈
With U x1 , x2 > R2 , x1 A 0, x2 A 0 and Γ x1 , 0 > R2 ` ∂U .
Therefore F x1 p2 x2 p1 z, so b x2 , x1 , and c 1. If we use the characteristic

equations, we get that ẋ1 x2 and ẋ2 x1 , and ż z.
The solution is x1 s x0 coss and x2 s x0 sins with x0 the initial condition. We
also have z s es z0 g x0 es .
Based on our domain, 0 @ s @ π ~2. So our path that we know the solution for is a quarter
circle. And we know that solution on this quarter circle is g x0 es .
Now given x1 , x2 > U , we want to find the quarter circle that hits this point.
»
That is we
1 2
want to find x0 , s such that x , x x0 cos s, x0 sin s. Then we get x0 x 2 x 2 2 ,
1
x2
and s arctan 1 . So the result is:
x
» x2
ux1 , x2 g x1 2 x2 2 exp arctan
x1
– 11 –
Remark 1.5.2. We recover the method of characteristics for ∂t at, x∂x u 0 by the
linear case.
Example 1.5.3. Consider when F is quasilinear, that is:
F x, u, Du Q bj x, u∂j u cx, u

j
In this case we have:
F Q bj x, z pj cx, z

j
Plug in our expressions, to get:

¢̈ j
¨ẋ bj xs, z s
¦
¨ż
¤̈
Pj pj ∂pj F cxs, z s
Remark 1.5.3.
1. There is no need to introduce p
2. there is no hierarchy of ODEs.

Example 1.5.4 (Burger’s Equation). Recall this equation was: ∂x1 u u∂x2 u 0. Then
we have F p1 zp2 , then ẋ1 1 and ẋ2 z and ż 0
Example 1.5.5. Consider
¢̈
¨ ∂1 u ∂2 u u2 in U
¦
ü g in Γ
¤̈
with U x1 , x2 > R2 , x2 A 0 and Γ x1 , 0 > R2 . So we have b1 1, b2 1 and c z2.

Therefore: ẋ1 ẋ2 1 and ż z 2 .
So we have x1 , x2 x 0 s, s.
dz 1 1 z 0
For the other one, we have: ds therefore s therefore z s .
z2 z s z 0 1 sz 0
g x 0
Then since z 0 g x0 , then z s .
1 g x0 s
Finally, given x1 , x2 > U , we need to find x0 and s so that x1 , x2 x 0 s, s. But
g x 1 x 2
then x0 x1 x2 , s x2 . Therefore ux
1 g x1 x2 x2
Example 1.5.6. Here will have a fully nonlinear F given as:
¢̈
¨∂x1 u∂x2 u u in U
¦
ü g on Γ
¤̈
– 12 –
Now let U x1 A 0 and Γ 0, x2 > R2 . Now we have:
ẋj ∂pj F x, z, p

ṗj ∂z F x, z, ppj ∂xj F x, z, p
ż Q pj ∂pj F x, z, p
j
Then in this case we have: F p1 p2 z, ẋ1 p2 , ẋ2 p1 , ṗ1 p1 , ṗ2 p2 and ż 2p1 p2 .
Then we have:
¢̈
¨ p 1 s p0 1 es
¦
¨ p 2 s p0 2 es
¤̈
Therefore: ż 2p0 1 p0 2 e2s , so that z z0 p0 1 p0 2 e2s 1. Lastly, ẋ1 p0 2 es 1
so that x1 p0 2 es 1 and x2 x0 p0 1 es 1.
g x 0
And z0 g x 0 , p 0 2 g x 0 , p 1

.
g x0

1
In the case where g x0 x20 , then we have that z0 x20 , p0 1 x0 , p0 2 2x0 . Now
2
given x1 , x2 > U we want to find x0 , s such that x1 , x2 x1x0 s, x2x0 s. Plugging in
what we know about these already, we get:
1
x0 x2 x1
4
s 1 2 1 1 x2 14 x1 4x2 x1
e x x
x 4 x2 14 x1 4x2 x1
x1 4x2 2
Then we get u and we simplify, to get: ux
16
Lecture 4 (9/10)
Today we will continue the material in Evans section 3.2.
Recall what we did last time with precise logical assumptions. We were looking at the
scalar first order PDE: F x, u, Dx 0 with F U R Rd R, with the boundary condition
u g on Γ ` ∂U with U a domaina
Then we derived the characteristic equations. For this we supposed that there exist a C 2
solution u to our PDE. From this we looked for a trajectory xs > U, z s > R, ps > Rd such
that uxs z s, pi s ∂i uxs and xs is chosen so that we have a closed system of
ODEs for x, z, ps.
a
open and connected set in Rd
– 13 –
The worst term is ṗi s ∂j ∂i uxsẋj s. To simplify this, we differentiate F in xi to
get:
∂xi F ∂xi F x, ux, Dux ∂z F x, ux, Dux∂i u ∂pj F x, ux, Dux∂i ∂j ux
To simplify everything, we choose ẋj ∂j ∂i uxsẋj s. Then we get: ṗi s ∂z F x, z, ppi
∂xi F x, z, p. And lastly, we get that ż Pi pi ∂pi F x, z, p.a .
1.6 Existence of Characteristic Solutions

Goal: We would like to develop local existence theory for smooth solutions to our PDE.
That is, given data near a point x0 > Γ, then under certain conditions, show that there exists
a smooth solution to our PDE in a small neighborhood V ? x0 .b
Let’s start by considering a simpler case. Let U Rd

xd A 0 and Γ ` xd 0
Let’s say we have the data x, z, p0 at x0 .

¢̈x0 x > Γ
¨
¨ 0
¨
¨
¨
¨z 0 g x0
¦ (6)
¨
¨ pi 0 ∂i ux0 ∂i g x0 f or i 1, . . . , d 1
¨
¨
¨
¨F x0, z 0, p0 0
¤̈
So we have 2d 1 equations for 2d 1 for x, z, p (2d 1 variables). These are known as the
compatibility conditions.
Assume that we start with x0, z 0, p0 that satisfy (6) (such a triple is called admis-
sible). Let x0 > W ` Γ. Parametrize the points in W by y y 1 , y 2 , . . . , y d 1 , 0. Now the
goal is to find the characteristics:
xy, s
z y, s
py, s
With
¢̈xy, 0 y
¨
¨
¨
¨
¨
¨z y, 0 g y
¦ (7)
¨
¨ pi y, 0 ∂i g y f or i 1, . . . , d 1
¨
¨
¨
¨p y, 0
¤̈ d F xy, 0, z y, 0, py, 0 0
To do this, we will use the implicit function theorem.

a
This can fail in many ways, one to keep in mind is if different trajectories collide
b
we mean local in space and time. And we were from now on refer to this as local existence.
– 14 –
Theorem 1.6.1 (Implicit Function Theorem). Given F Rm Rn Rm which is C 1 and

suppose that F y0 , p0 0 is satisfied for some y0 , p0 . If ∂y F is invertible (as a matrix) then
there exist neighborhoods U ? y0 and V ? p0 then there exists a C 1 function p U V such
that py0 p0 and F y, py 0 for all y > U .
Moreover, if F y, p 0 for some y > U and p > V then p p y

Therefore (7) is solvable if ∂pd F x0 , z0 , p0 x 0. In this case, this is known as nonchar-
acteristic.
In summary: we began with x0 , z0 , p0 with z0 g x0 and p0 ∂1 g, . . . , ∂d 1 g x0 , p0 d
which is admissible. And ∂pd F x0 , z0 , p0 x 0 noncharactersitic. Then we have:

¢̈xy, 0 y
¨
¨
¨
¦z y, 0 g y
¨
¨
¨py, 0
¤̈
With xy s, z y, s, py, s are given by the charactersitic equations.
Now, given x > U near x0 , we want to find y, s such that x xy s.
Lemma 1.6.1. Under the noncharactersitic assumption, the map y, s ( xy, s is invert-
(
ible near x0 . Denote the inverse by x y x, sx
Proof. Inverse function theorem. Compute using the characteristic equation that x satisfies:
Fp1 x0 , y0 , p0

k Id1
∂y jx ∂s xk Uy xn
Fpn1 x0 , z0 , p0
s 0
0 F x , z , p
pn 0 0 0
Compute the determinant which is just ẋd x0 , 0 ∂pn F x0 , z0 , p0 x 0

Theorem 1.6.2. Under the above assumption, ux z y x, sx where y, sx are
defined with x > V . Then this gives us a solution to the PDE
¢̈
¨F x, u, Du 0 in V
¦
ü g
¤̈
on Γ 9 V ` W
Consider the quasilinear 1st order PDE:
Q bj x, u∂j u cx, u 0

j
F Q bj x, zpj cx, z

j
Then the noncharacteristic condition requires ∂pd F x0 , z0 , p0 bd x0 , z0 x 0. So in the

quasilinear case, the noncharacteristic assumption is equivalent to the existence of a unique
– 15 –
choice p0 d from everything else.
In the fully nonlinear case, take F ∂x u2 ∂y u2 1. Then F p21 p12 1. Then given
1
p0 1 , then we have two choices of p0 2
2
Now let’s prove Theorem 1.6.2.
Proof. step 1 veryify that F xy, s, z y, s, py, s 0. Compute:
∂s F x, z, p Q ∂x F ẋi i ∂z F ż Q ∂j F ṗj

i j
Then apply the characterisitc equations to get that the whole thing is zero. Then since
F xy, 0, z y, 0, py, 0 0 by preparation, we are good.
step 2 We know that xy x, sx x and z y x, sx ux. So it remains to show
that ∂i ux pi y x, sx.
pi z y x, sx Q ∂y z∂iyj j

∂s z∂i s
j
And recall that ∂s z Pi pi ∂pi F . So we need to compute ∂yj z. From z y, s uxy, s ,
then we need to prove:
∂y j z Q piy, s∂y xi j
i
Define:
r j s ∂yj z Q pi y, s∂yj xi y, s

i
rj 0 0 by preparation. Then by a long computation, we get that:
∂s r j ∂z F r j
From this we get that some equation is:
Q Q pk ∂y xk ∂x yj Q pk ∂p F ∂x s Q pk Q ∂y xk ∂x yj
j k
k
i j i ∂s xk ∂xi s
j k k ´¹¹ ¹ ¹ ¸¹¹ ¹ ¹ ¶ k j
∂s ẋk
Q pk δik pi
k
Lecture 5 (9/12)
For problem number 4 of the homework (Evans section 3.3.5 problem 6(b)). Take
ux, t g x0, x, tJ 0, x, t (take this as the definition of u and verify that it is a solution).
– 16 –
Review: in the case Γ ` xd 0 U xd A 0. We wanted to solve:

¢̈
¨F x, u, Du 0 in U
¦
ü g on Γ
¤̈
Then we used the method of characteristics, picking x0 > Γ

¢̈∂ xi ∂pi F x, z, p
¨
¨
s
¨
¦ ∂s z
¨
Pi pi∂pi F x, z, p
¨
¨∂ p ∂z F x, z, ppi ∂i F x, z, p
¤̈ s i
We are interested in solving for this on an open set U . To solve this we needed initial values:
¢̈x0 x0
¨
¨
¨
¦z 0 g x0
¨
¨
¨p 0 p 0 i
¤̈ i
For p0 we need to make a choice. For p0 i ∂i g x0 with i 1, . . . , d 1. But for the normal
direction, we need to figure this out such that it solves F x0 , z0 , p0 0. Such a triple is
called admissible.
Noncharacteristic x0 , z0 , p0 means that ∂pd F x0 , z0 , p0 x 0. This allows us to solve

for pd y, 0 for y > W ? x0 and W ` Γ (unique continuous choice with pd x0 , 0 p0 d .
We also could invert y, s ( xy, s for s very small and y close to x0.
Now define ux z y x, sx. Then the claim is that this u is a solution to the bound-
ary valued problem in V . And this is smooth.
Now let’s extend this to general U and Γ (Ref: Evans Appendix C)

Definition 1.6.1 (C k Boundary). Given U open bounded in Rd . We say that ∂U is C k at
some x0 > ∂U if there exists some r A 0 and γ Rd 1 R in C k so that:
U9 B x 0 , r xd A γ x 1 , . . . , x d 1

´¹¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
ball of radius r centered at x0
And ∂U 9 B x0 , r xd γ x 1 , . . . , x d 1
So now we would like to flatten the boundary. Define the coordinates y 1 , . . . , y d so that:
¢̈ i
ÿ xi i 1, . . . , d 1
¦
ÿ d xd γ x1 , . . . , x d 1
¤̈
Then U 9 B x0 , r ` y d A 0 and ∂U 9 B x0 , r ` y d 0.
– 17 –
Note this is invertible by:

¢̈ i
¨x yi i 1, . . . , d 1
¦
¨ xd y d γ y 1 , . . . , y d
1
¤̈
If U or ∂U is C k for all k, we say it is C ª
Suppose we have the general problem:

¢̈
¨F x, u, Du 0 in U
¦
ü g on Γ
¤̈
(
With U a C domain. Then there exists x y x, smooth, invertible in a neighborhood
ª
V ? x0 . So if v y uxy , then v satisfies a PDE of the same form.
∂y
F xy , v y , Dv y xy 0
∂x
Then we can write the above operator as Gy, v y , Dv y 0. Then applying our theory of
what we have done so far to our equation Gy, v, Dv 0 then we solve the original problem.
In this case, the noncharacteristic condition now becomes:
Q ν∂U
j
∂p F x0 , z0 , p0 x 0
j
j
Where ν∂U is the unit outer normal vector to ∂U
Theorem 1.6.3. Local existence near a noncharacteristic boundary value holds for general
C domains.
ª
Let’s now talk about uniqueness. Suppose that we already have two solutions u1 , u2
that are C 2 solving:
¢̈
¨F x, u, Du 0 in U
¦
ü g on Γ
¤̈
does u1 u2 ?
The idea is the method of characterics from Γ to x in U . But we should be careful about
j
the normal derivative ν∂U ∂j u 0
Example 1.6.1. Consider ∂1 u2 1 on RC0 and u 1 on u. We have two solutions
u 1 x and u 1 x. This has to do with the choice of p0 1 Pj uj∂U p0 j .
So for uniqueness, we require that the normal derivatives also agree: P ∂∂U
j
∂j u 1
P ∂U
∂ j
∂j u2 .
– 18 –
Then for all x > U such that there exists a characteristic xs, z s, ps with x0 > Γ, z 0
ux0 , p0 Dux0 such that xs x and xs > U for all s > 0, s. Then we have
u1 x u2 x by uniqueness of ODEs.
What about continuous dependence? Continuous dependence does hold, but it is pretty
subtle.
Example 1.6.2. Consider Burger’s equation
¢̈
¨∂t u u∂x u 0 in t A 0
¦
ü g on t 0
¤̈
Suppose that g > C 2 and Yg YC 2 B A a then there exists a neighborhood V 0, T R with

T only dependent on A (in particular independent of g) such that there exists a unique C 2
solution u on V .
(
Now let’s write g u g is continuous in C 2 , in the sense that if gn g in C 2 Γ then
u gn u g in C 2 V .
1 2
However, there exist counterexamples hn , hn such that:
1 2
Yu hn u hn YC 2 V C nYhn1 hn2 YC 2 Γ
i.e. g ( u g cannot be Lipschitz C 2 C 2.
Terry Tao wrote a good blog on quasilinear well-posedness.

Here we will be looking at Evans section 4.6. We would like to study the extension of
the notion of noncharacteristic bounded data to higher order systes of ODEs.
Let’s begin with the noncharaceristic condition for a quasilinear, 1st order, scalar PDE.
Recall, we were looking at (with Γ ` xd 0):
F x, u, Du Q aj x, u∂j u a0 x, u 0

j
In this case x0 , p0 , z0 became our characteristic if and only if ad x0 , z0 x 0.
Note that given this, we can uniquely write:
F x, z, p Q aj x, z pj a0 x, z

j
x0 , z 0 is noncharacteristic, therefore this is equivalent to:

d 1
1
Q aj x0 , z0 p0 j a0 x0 , z0

p0 d
ad x 0 , z 0 j 1
a
where Yf YC k U sup α Bk supx>U S∂ α uS
S S
– 19 –
If we are interested, we can determine ∂ α ux0 under he noncharacteristic assumption.

Because from the boundary data g we got z0 , p0 1 , . . . , p0 d 1 and from nonchar we got

p0 d ∂d ux0 . Now suppose we want ∂i ∂j ux0 . This is easy if i j 1, 2, . . . , d 1. But

also, for y close to x0 , we get ∂i ∂d ux0 from i 1, . . . , d 1. To compute the second normal
derivative, we differentiate the equation in ∂d to get:
0 Q aj x, u∂j ∂du bx, u, ∂u

j
but if we evaluate at x0 , then we know everything but aj xp , z0 ∂d2 u. In other words we
can write ad x0 , z0 ∂02 x0 things that only depend on x0 , ux0 , ∂j ux0 j 1, . . . , d, other
things. But the thing on the left is nonzero, so we can solve for the thing on the left.
So the noncharacteristic assumption of x0 , z0 is equivalent to determining all derivatives

of u at x0 .
z
Lecture 6 (9/17)
1.7 noncharactersitic condition for kth order system

Reference: Evans section 4.6. The things that are important
4.6.1: noncharacteristic cauchy data for quasilinear k th order systems
4.6.3: Cauchy - Kovalewski theorem
Consider the quasilinear K-th order system of PDEs. We have u U RN , with ux
®
`Rd
1
u
. And:
uN
N
0 Q Q bαAB x, u, . . . , Dk 1uDαuB
cA x, u, . . . , Dk 1 u
Sα S kB 1
k1 2
With: bα ; U RN Rd N RN , with bα A B with A the row index and B the column index.
And C defined similarly. (I don’t really understand the indicies).
The Cauchy data is data for u, ∂ν u, . . . , ∂νkk11 u on Γ ` ∂U .

More simply, we could have Γ ` xd 0, with Cauchy data for u:
2 k 1
u, ∂d u, ∂d u, . . . , ∂d u g0 , g1 , . . . , gk1
on Γ.
– 20 –
If we assume that we have a smooth solution u, one question we will ask is whether we
can determine all derivatives Dn ux0 with x0 > Γ from the Cauchy data and the equation?
The answer is to come up with a noncharacteristic condition for the Cauchy data near x0 .

Given the Cauchy data: g0 , . . . , gk 1 near x0 > Γ. Then we can compute any Dk ux0 for

k B k except ∂dk ux0 . We can get this from the PDE. Moving things around we get:

b0,...,0,k ∂dk ux0 Q bα D

α
u cx 0
S αS k
αx0,...,0,d
We need that b0,...,0,d is invertible, in which case we have that:

∂dk ux0 1
b0,...,0,d

Q α
b α D u c x 0

Sα S k
αx0,...,0,d
We want to this again and again. Note that b0,...,0,d is invertible around x0 , therefore we can
compute ∂dk ux in a neighborhood of x0 (in Γ). Therefore we can compute all of Dk 1 ux
except ∂dk 1 ux0 .

Now take ∂d of the equation, keep ∂dk 1 u on the left hand side and the rest on the right

hand side. Then we get:
b0,...,0,d ∂dk 1 ux0

known
but we can invert and get the something known. The same condition (that b0,...,0,d is in-
vertible) allows us determine ∂dk 1 ux0 and ∂dk 1 x in a neighborhood of x0 .

Definition 1.7.1 (noncharacterisitic condition). g0 , . . . , gk 1 is noncharacteristic at
x0 if b0,...,0,d evaluated at x0 , ux0 , . . . , Dk 1 ux0 is invertiblea .

Theorem 1.7.1. If g0 , . . . , gk 1 is noncharacterisitic at x0 , then all derivatives of u at x0

can be determined.
We can generalize this to general domains. Given U a smooth domain, which means that
for all x0 > ∂U , after renaming and reorienting the coordinate axis, there exists r A 0 such
that:
∂U 9 B x, r xd γ x 1 , . . . , x d 1

with γ a smooth function. Then we can take y i xi for i 1, . . . , d 1 and y d xd
γ x1 , . . . , x d 1 .

a
the imput to b can be computed from the Cauchy data
– 21 –
Recall we can also define the unit normal to the boundary ∂U as ν∂U .
∂k
Let’s define the k-th normal derivative ux0 for x0 > ∂U which we define as:
∂ν k
d
∂k
∂ν k
ux0 Q ν i1 ν ik ∂i1 ∂ik uc0
i1 ,...,ik 1
∂k ∂
Note that k
ux0 agrees with k ux0 ν © K ux0 at the top order.
∂ν ∂ν
After playing with indicies, we get that:
∂k k
∂ν k
ux0 Q αν αDαux0
SαS k
With:
k k!

α α1 !αd !
Now the Cauchy data is:
k 1
u, ∂ν u, . . . , ∂ν k1 u g0 , g1 , . . . , gk1 in Γ
The noncharacterisitc condition comes out from the following. Idea: work iwth v y
uxy . The equation for u will translate to some equation for v in the variable y. After a
long computation, it will turn out that we get:
Q b̃αy, v, . . . , Dk 1vDαv

c̃y, . . . , Dk 1 v

Then we require that b̃0,...,0,d is invertible.
“oh I see, I had to be much more clever than this”

– Oh (2019)
To do this, we need to compute (for SαS k)
Du Dv y x Dy i ∂yi v y x

D2 u DDy i ∂yi v

So we have:
Dα u Dα v y x k
∂y d v Dy
d α
terms not involving ∂ykd v
Then we see that y d is a boundary defining function. By this we mean that ∂U ` y d 0.
And we have that from multivariable calculus that Dy d is parallel to the nornal ν.
– 22 –
This means that the original equation
bα D α u bα Dy d α ∂yKd v terms not involving ∂ykd v Q b̃α∂αk v

d c̃
α
Therefore:
b̃0,...,0,d Q bαDydα ck Q bα ν α
Sα S k S αS k
Therefore b̃ is invertible if and only if Pα S S k bα ν

α is invertible.
Example 1.7.1. Consider constant coefficient, 2nd order, scalar PDEs.

d
Q aij ∂i∂j u 0
i,j 1
Here the noncharacterisitic condition only depends on ν (or ∂U ). This requires that:
d
Q aij ν iν j x 0
i,j 1
If ∂U is given by (locally) w 0, then w is parallel to ν and the noncharacterisitc con-

©
dition becomes:
d
Q aij ∂iw∂j w x 0
i,j 1
One case of this is if aij is a positive definite matrixa . For example if aij is the identity
matrix, then our PDE becomes ∆u 0 and all boundaries are noncharacteristic (this is the
property of ellipticity).
Another example would be:
1 0
aij
0 I
Then we get the wave equation ju 0. This is characteristic if and only if ∂0 w 2

P d
i 1 ∂i w
20 (these are characteristic hypersurfaces).
Lecture 7 (9/19)
Recall our discussion on the noncharacterisitic condition for

¢̈
¦
P
¨ SαS k bα x, u, . . . , D k1 uD α u cx, u, . . . , D k1 u 0 in U
(8)
ü, ∂ν u, . . . , ∂ kk11 g0 , . . . , gk1
¤̈ ν
on Γ ` ∂U
a
P aij ξi ξj A 0
– 23 –
We say that the boundary value is noncharacteristic at some point x0 > Γ implies that
we can compute (or determine) all of the derivatives of u at x0 (Dβ ux0 ) if and only if the
matrix PSαS k bα x0 , ux0 , . . . , Dk 1 ux0 ν α is invertible.

This motivates the following theorem. Can we construct a solution of u near x0 by the
knowledge of all Dα ux0 ? The answer is yes, provided that everything can be written as
power series. Equivalently, if everything is real analytic.
Definition 1.7.2 (real-analytic). We say that f U RN is real analytic if for all x0 > U ,
there exists r A 0 and coefficients aα > RN where α are multi-indicies such that:
f x Q aα x x0 α
α
for Sx x0 S @ r
Remark 1.7.1. If f is real analytic, then we have that:
D α f x0 α!aα α1 !α2 !αd !
Definition 1.7.3 (real-analytic boundary). We say ∂U (or U ) is real analytic if for all
x0 > ∂U there exists r A 0 such that ∂U 9 B x0 , r xd γ x1 , . . . , xd 1 (after a suitable

relabeling and reoriented of the coordinates), with γ real analytic
Theorem 1.7.2 (Cauchy–Kovaleskaya Theorem). Consider the boundary valued prob-

lem (8) such that bα , c, g0 , . . . , gk 1 , ∂U are real-analytic. And suppose the boundary values

are noncharacterisitc at x0 > Γ. Then there exists a neighborhood V ? x0 such that there
exists a unique, real-analytic solution u to the boundary valued problem on V 9 U .
Moreover, this solution:

1
ux Q α! Dαux0x x 0 α
near x0
Proof. Really clever proof, see the text.
Remark 1.7.2 (about Cauchy–Kovaleskaya theorem). At first, you might think that this
theorem looks like a very general existence and uniqueness statement for PDEs. However,
real-analytic functions are too restrictive to be useful.
Suppose that f was real-analytic on Rd , and suppose that f 0 on B 0, 1, then f 0 on
Rd .
The dependence of u on the boundary data in the Cauchy–Kovaleskaya theorem is very

weak. This would make small disturbances of g’s lead to large disturbances for u.
This leads to a classical example by J. Hadamard.
– 24 –
2. Distributions
Example 1.7.2 (J.Hadamard). Consider

¢̈ 2 2
¨ ∂ 1 ∂ 2 u 0 in R2
x2 A 0
¦
ü1 , ∂2 ux0 , 0 0, k sin kx0
¤̈
This can be solved as:
ux1 , x2 0 sinkx1 f x2
Then we have:
2
0 ∂ 1 ∂22 u k 2 sinkx1 f x2 sinkx1 f 2
x sin kx1 f 2
x k f x
2 2
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
0
Once we do this, to solve for f , we get:
ux1 , x2 sinkx1 sinhkx2
is a solution to the boundary value problem. Now everything is analytic, so by the theorem,
this is the unique solution.
Now let’s perturb the initial data. Let’s have the initial data now 0, e k 1 ~ 2
k sin kx0 , then
we get:
k1 ~ 2
u e
sinkx1 sinhkx2
Now as we send k ª , then our data goes to 0, 0 in any C n -norm.
However, let’s look at the solution with x2 A 0, then we have:
e k1
~ 2
sin kx1 sinhkx2 ÐÐ
k ª 1 kx2
2
e e k1 ~ 2
sinkx1
but that thing blows up in the uniform norm.
2 Distributions
We now are turning to the second part of the course. We have two goals
1. We would like to study several fundamental linear 2nd order scalar PDEs (Laplace,
Wave, Heat, Schrodinger). (chpt 2.2, 2.4, 2.3)
2. Understand the theory of distributions. This is the “completion of differential calcu-

lus”. We will also like to study Fourier analysis (chpt 4ish). Check the bcourse site for
lecture notes on these topics.
– 25 –
2. Distributions
2.1 Motivation
Schwartz was the founder of the theory of distributions, but it was independently constructed
earlier. Let’s consider a motivating example from electrostatics.
Example 2.1.1. Given E R3 R3 the electric field and ρ R3 R the charge density
¢̈
¨© E ρ
¦
¨© E 0
¤̈
Recall from vector calculus, that © ©ϕ 0. The converse is also true, therefore there
exists some ϕ such that E ©ϕ.
A basic problem in electrostatics is to determine E (or ϕ) from ρ. Let’s solve:
∆ϕ ρ
First consider the case when ρ is the sum of point charges at xk with charges qk . By linearity,
it is enough to find:
ϕ Q q1 ϕ 0 x xk
where ϕ0 is the potential of the point charge with charge q 1 at the point 0. Then we can
say that ρ could be written as a continuous sum of point charges:
ρ S ρy dy
Then we can say that ϕ is the continuous sum of point-charge electric potentials. So:
ϕ S ρy ϕ0 x y dy
So all that remains to find is the electric potential of a point charge, ϕ0 . Note by rotational
symmetry of the problem, ϕ0 must be radial: ϕ0 ϕ0 r. By the Gauss equation, we have:
1 SB 0,r
© E S∂B 0,r

ν © ϕ0 S∂B 0,r
∂r ϕ0 ∂r ϕ0 r4πr2
11
Therefore: ∂r ϕ0 r . So we know normalize so that ϕ0 0 as r ª and then we see
4π r2
that:
1 1 1
ϕ0
4π r 4π SxS
An alternative derivation of this comes from Evans chapter 2.2.
Lecture 8 (9/24)
– 26 –
2. Distributions
2.2 Basic Definitions

I would like to point out that the word distribution to define these objects doesn’t really
make a lot of sense. I prefer the word ‘generalized functions’, I think that makes more sense.
Today will the real beginning of our discussion on distribution theory. Let’s motivate the
definition of a distribution.
Let’s try to make sense of the notation of point charges qi at points xi > R3 . Let u be the
charge distribution function of these charges. We would like:
1. u 0 outside xi
2. RU u Px >U qi
i
However, u cannot be a function that satisfies this. Rather than viewing u as a function on
R3 , it is more natural to interpret u as a functional on the set of sets. By this we mean:
u U Q qi
xi > U
Given two open disjoint subsets U, V , then uU 8 V uU uV . If we interpret these
has integrals, we have (not rigoruously):
S uχU χV S uχU S uχV
where χ is the characteristic function. It is therefore natural to interpret u as a linear func-

tional on the space of characteristic functions.
So we want to view u as a linear functional on the space of nice functions. The nicest we
can think of is the space of smooth, compactly supported functions: C0 R3 . ª
Here is a rough definition of a distribution:
Definition 2.2.1. A distribution u on a set U ` Rd is such that u C0 U ª

R a linear
functional which is also “continuous”
Notation: Given ϕ > C0 U , we will write:

ª
uϕ ù, ϕe
which mimics the L2 inner product.
We call the space C0 U the space of test functions.

ª
How “flexible” is this space? There exists one nontrivial element:
ψR R
– 27 –
2. Distributions
with the property ψ > C ª

such that suppψ x ψ x x 0 ` 0, ª. We can set:
¢̈ 1~x
ë xA0
ψ ¦
¨0
¤̈
xB0
Then ψ works. Now consider ϕ Rd R by writing ϕx ψ 1 x1 2 xd 2 , then
supp ϕ ` B 0, 1.
Definition 2.2.2 (Convolution). Given f a locally integrable function, ϕ > C0 Rd , then ª
we write:
f ϕ x S f y ϕx y dy
Remark 2.2.1. Note that even if f is nondifferentiable, then f ϕ is C , because:

ª
∂ xj f ϕ ∂xj S f y ϕx y dy S f y ∂xj ϕx y dy
Where the second equality relies on f being locally integrable and ϕ smooth.
Remark 2.2.2. If f > C Rd and ϕ > C0 Rd . Then suppf ϕ ` supp f supp ϕ. Where
ª
supp f supp ϕ x y > Rd x > supp f, y > supp ϕ

In particular, if f > C0 Rd and ϕ > C0 Rd , then f ª
ϕ > C0 Rd
ª
Lemma 2.2.1 (Approximation by mollification). Let f > C k Rd and ϕ > C0 Rd with ª
R ϕ 1. Then given δ A 0, define:

1
ϕδ x ϕ x ~ δ
δd
Therefore supp ϕδ ` B 0, δ and R ϕδ 1. Then f ϕδ ÐÐÐÐÐ

unif ormly
f as δ 0. Moreover, up
to k derivatives uniformly converge:
sup SDα f ϕδ f S 0
Rd
for all α, SαS B k.

Proof. When SαS 0 and k 0. Then we have:
f ϕ δ x f x S f y ϕδ x y dy f x S f x y ϕδ y dy S f xϕδ y dy
S ϕδ y f x y f xdy S f x δz f xϕz dz
Then as δ 0, we use continuity of f to get that this integral goes to zero.
For higher k with SαS A 0, you let the derivatives fall on f , then proceed with the same
proof.
– 28 –
2. Distributions
Now what is the topology that we should take?

Definition 2.2.3 (Convergence of Test Functions). We say that a sequence ϕj > C0 U ª
converges to ϕ > C0 U (uj u) if there exists a compact set K ` U such that supp ϕj ` K
ª
(for all but finitely many) and supp ϕ ` K and supx>K SDα ϕj x Dα ϕxS 0 as j ª for
all multi-index α
Remark 2.2.3 (Optional). This isn’t a Frechet space or a Banach space
Remark 2.2.4 (Optional). This is the strongest topology such that every C0 K embeds ª
into this space for every compact subset K. Where C0 K is the space of functions in
ª
supp ϕ ` K endowed with ϕj ϕ if and only if supx>K SDα ϕj x Dα ϕxS 0 for all α.
This is a complete and metrizable set (Frechet).
Definition 2.2.4 (Distribution). u C0 U U is a distribution if it is linear and
ª
continuous in the sense that if ϕj is a sequence of test functions which converge to ϕ in the
above sense, then uϕj converges to uϕ.
Lemma 2.2.2. Let u C0 U ª
R be a linear functional. Then u is a distribution (i.e.
continuous) if and only if (“boundedness”) for all compact K ` U there exists N A 0 and
C CK,N A 0 such that for all ϕ > C0 U , supp ϕ ` K:
ª
S ù, ϕe S B CK,N Q sup SDα ϕxS

x> K
SαSBN
Proof. If you have boundedness, then you have continuity. We can ask (by linearity) if
ù, ϕj ϕe 0. But ϕj ϕ 0 in C0 U there exist K compact such that supx>K SDα ϕj x
ª
ϕxS 0. By boundedness, there exists a N and C such that
S ù, ϕj ϕe S B C Q sup SDα ϕj x ϕxS

x> K
0
SαSBN
For the other direction, suppose that u doesn’t have boundedness. Then we get a compact
K ` U such that for all N , there exists ϕN with S ù, ϕN e S C N PSαSBN supx SDα ϕxS. Then
set ψN x ϕN xN PSαSBN supx SDα ϕxS 1 . Then SDα ψN S B 1~N forall SαS B N . Then

ψN 0 in C0 U but not because the above.

ª
Lecture 9 (9/26)
Recall that we defined the notation of a distribution.

Definition 2.2.5 (distribution). A distribution u is a linear functional ϕ ( ù, ϕe on
C0 U that is continuous.
ª
i.e. if ϕj ϕ in the sense that there exists a compact set K ` U such that supp ϕj , supp ϕ `
K, we have:
sup SDα ϕj x Dα ϕS 0

x> K
as j ª for all α. Then ù, ϕj e ù, ϕe.
– 29 –
2. Distributions
Lemma 2.2.3. A linear functional u C0 U ª

R is a distribution if and only if it is
“bounded” in the sense that for all compact subsets K ` U there exists an N and CK,N A 0
such that for all supp ϕ ` K, we have that S ù, ϕe S B CK,N supx>K supSαSBN SDα ϕxS
We will call the space of all distributions D U

Definition 2.2.6 (Order of a distribution). We say that a distribution u > D U is said
to be of order B N if there exists N such that for all compact sets K there exists CK,N A 0
such that S ù, ϕe S B CK,N supx>K supSαSBN SDα ϕxS
The order of a distribution is the minimum among such N .
Definition 2.2.7 (Support of a Distribution). Given an open subset V ` U we say that

u vanishes on V if for all ϕ > C0 U , supp ϕ ` V implies that ù, ϕe 0.
ª
Then set Vmax V u vanishes on V . Then set supp u c

Vmax
Example 2.2.1. Any continuous function u defines a distribution by ù, ϕe R uϕ for all
ϕ > C 0 U .
ª
The order of such a distribution is zero and the support definitions agree
Example 2.2.2. Locally integrable functions are distributions. Recall u a function on U is

locally integrable if for all compact subsets K ` U we have that RK SuS @ ª.
This is a distribution, the order is 0 and supp u suppudu
Example 2.2.3. The signed Borel measure on U , B U , defines a distribution (just integrate
against it). The order is 0.
Example 2.2.4. The simplest example of order N is ù, ϕe Dα ϕ
2.3 Operations
Now let’s discuss the basic operations for distributions.
The basic principal (the adjoint method): take an operation P on a function u, we will
try to compute P so that R P uϕ R uP ϕ for all test functions ϕ. Then given a distribution

u > D U , we define P u by `P u, ϕe ù, P ϕe

Example 2.3.1. Consider multiplication of u > D U by f > C ª

U . When u > C0 U
ª
S f uϕ S uf ϕ
Following this we say that `f u, ϕe ù, f ϕe
– 30 –
2. Distributions
Example 2.3.2 (Differentiation). Start with u > C0 U and ϕ > C0 U then ª ª
S Dα uϕ 1
S αS
S uDα ϕ
So define: `Dα u, ϕe 1SαS ù, D α ϕe
Note that all distributions are differentiable. So distribution theory extends differential
calclus. Moreover, all distributions are locally Dα f for f continuous functions.
Example 2.3.3 (convolution). Let’s first consider u > D U and v > C0 U . What is u v? ª
If v, ϕ > C0 U , then recall:

ª
u v x S uy v x y dy
So naturally:
ù v, ϕe S u v xϕx U uy v x y dyϕxdx S uy S v x y ϕxdx dy

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
v ϕy
So ù v, ϕe ù, v
ϕe
Because v > C0 U , then u v is not just a distribution, it is a smooth function.
ª
Proof. Consider u v x ù, v x e. By continuity of u and the fact that the map
(
x v x is continuous in the space of test functions.
So if xn x, then v xn y v x y in C0 U . This means that u v xn

ª
u v x
ie. as a function, u v x is continuous. Note that:
S u v xϕx ù v, ϕe
it is easy to show smoothness.

Let’s discuss more examples of distributions.
Example 2.3.4 (Delta Function). Delta distribution δx0 > D U where `δx0 , ϕe
ϕx0 .
This is equivalently given by an atomic measure of mass 1 supported x0
Example 2.3.5 (Heaviside Function). Heaviside function H R R with:
¢̈
¨1 xA0
H x ¦
¨0
¤̈
xB0
Then H
δ0 . Because
ª
`H

, ϕe `H, ϕ e

S0 ϕ xdx

ϕ0
– 31 –
2. Distributions
Remark 2.3.1. We can generalize this computation to give a distribution theoretic proof of
the divergence theorem.
Remark 2.3.2. Trying to make sense of 1~x as a distribution. This isn’t a L1loc function.
However, does there exists a distribution u such that u 1~x in R 0. Take log SxS and
differentiate it as a distribution (this is known as taking the principal value of 1~x.
ª
ϕ ª L 0
`1~x, ϕe S ª x
dx S ª
log SxSϕ xdx

S0 log SxSϕ ϕ0
S L log SxSϕ

ϕ0
L ϕx ϕ0 ε ϕx ϕ0 ε ϕx ϕ0 L ϕx ϕ0

SL x
dx SL x
dx S
ε x
dx S
ε x
dx
(Yeah I know, a bunch of those bounds should be negative) So we then send this to infinity.
The middle term ends up going to zero so we define:
ε ϕx ϕ0 L ϕx ϕ0
`P V 1~x, e ϕ lim S S
ε 0 L x ε x
To show the middle term actually goes to zero we note that:
ϕx ϕ0 ε ε 1
x Sε
dx S S ϕ xσ dσdx
ε 0

Then we can bound ϕ and send ε to zero to get that this integral goes to zero.

Lecture 10 (10/1)
Today we will continue on our discussion of distribution theory. Recall that distributions
(D U ) are defined as linear functionals on C0 U that are continuous in the sense of con-
ª
vergence on C0 U .ª
We discussed basic operations on distributions: multiplication by a test function, differ-

entiation, convolution with a test function.
2.4 convergence of distributions

Let’s now discuss sequences of distributions.
Definition 2.4.1 (Convergence in the sense of distributions). Given a sequence un >
D U , we say that un u > D U in the sense of distributions if for all ϕ > C0 U ,
ª
ùn , ϕe ù, ϕe
Remark 2.4.1. This is the same of weak convergence of linear functionals our space.
Theorem 2.4.1 (Sequential Compactness). Let un > D U and suppose that for all
ϕ > C0 U , ùn , ϕe is convergent. Then the claim is that the linear functional u defined by
ª
ù, ϕe lim ùn , ϕe (9)

is a distribution (i.e. it is continuous on C0 U ).
ª
– 32 –
2. Distributions
Proof. (optional) It’s clear that u as defined above is linear. So it suffices to check continuity.
Let K be any compact subset of U . Then u obeys (9) for ϕ > C0 K . Note that this
ª
space is nice (for a fixed K). Convergence is just ϕn ϕ when for all N , SαS B N , we have
supx>K SDα un Dα uxS 0. So define the seminorm ρN ϕ supSαSBN supx>K SDα ϕxS.
Then let:
PN ϕ ψ
dϕ, ψ Q2 N 1

PN ϕ ψ
N
then we have convergence of ϕn if and only if dϕn , ϕ 0. So C0 K , d defines the

ª
same topology as above and d is invariant in the sense that dϕ v, ψ v (.ϕ, ψ for all
ϕ, ψ, v > C0 K , and d is complete.
ª
These must be proven. But once these are shown, then un is the family of linear
functionals on C0 K which is a complete metric space (compatible with the vector space
ª
structure). Then we can apply the boundedness principal (or the Banach-Steinhaus theo-
rem). If ùn , ϕe is bounded for each ϕ, then un is equicontinuous.
By equicontinuity, we have that uSC0ª K is continuous and K was arbitrary. Then since
C0 U is the strongest topology such that C0 K C0 U imbedds continuously. It follows
ª ª ª
that u is continuous on C0 U
ª
Example 2.4.1. Consider a sequence of functions un > L1loc U . In this case you can connect
the notion of pointwise convergence and the convergence in the sense of distributions via the
dominated convergence theorem.
That is, if un u in the pointwise sense, and there exists some function g > L1loc such
that Sun S B g almost everywhere on U , then un u in the sense of distributions
Proof. We just need to show that R un ϕ R uϕ for all ϕ > C0 U . But this follows imme-
ª
diately from the dominated convergence theorem.
Example 2.4.2. Let h > C with h 1 on 1, ª and supp h ` 0, ª. Then if we take
ª
limδ 0 hx~δ this converges to the heaviside function pointwise and in the sense of distri-
butions.
Example 2.4.3. Consider the sequence:

¢̈ inx
ë xA0
un x ¦
¨0
¤̈
xB0
Then un x does not have a pointwise limit for 0 @ x x 2πZ. But this sequence does have a
limit in the sense of distributions, and it is zero.
– 33 –
2. Distributions
To see this, pick phi > C0 U , then we have: ª
ª
1 ª
1 inx ª
1 ª
ùn , ϕe S einx ϕxdx S

in S0
∂x einx ϕxdx e ϕxU einx ∂x ϕxdx
0 0 in in 0
1 1 ª
ϕ0 S einx ∂x ϕxdx

in in 0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
bang
1
But we can control (bang) as its absolute value is bounded by R0 S∂x ϕxSdx which goes to
ª
n
zero.
1
Example 2.4.4. Take the same un but normalize it by n. Then the limit of limn nun δ0 ª
i
in the sense of distributions (check).
Example 2.4.5 (approximation of the identity). Let ϕδ be the standard mollifer with
support contained in a ball of radius δ. Then we saw that ϕδ u u as δ 0.
Now if we take the same ϕδ , then ϕδ δ0 as δ 0 in the sense of distributions.

Proof. Fix ψ > C0 Rd , then:
ª
`ϕδ , ψ e S δ d ϕδ 1 xψ xdx

S ϕz ψ δz dz
then apply dominated convergence theorem.
Remark 2.4.2. δ0 u u
Proposition 2.4.1. u > D Rd take ϕ >
C0 Rd with R ϕ
ª
1, ϕδ x δ d ϕx~δ , then

ϕδ u u in the sense of distributions.

Proof.
`ϕδ u, ψ e ù, ϕδ ψe

with:
S S ϕδ z ϕz ydz ÐÐ ψy
δ 0
ϕδ ψ

ϕδ x y ψ xdx
repeat for Dα ϕδ ψ , then ϕδ

ψ ψ in C0 U . Then by continuity
ª
of u, we get
ù, ϕ0 ψ e

ù, ψ e
Remark 2.4.3. ϕδ u is a smooth function. Therefore every distribution can be approximated

by smooth functions.
Take u > D U , then there exist a sequence uj > C0 U such that uj
ª
u in the sense of
distributions.
Here is a construction. We would like to construct a sequence Kj of compact subset of

U which is nicely nested such that:
1. Kj ` Kj0 1
2. j Kj U
now construct χj > C0 U such that χj 1 on Kj but supp χj ` Kj 1 . Then we can compute
ª

ϕδ χj u then set uj be a sequence of these then we are done.
– 34 –
2. Distributions
2.4.1 Approximation Method

An operation P on smooth (smooth and compactly supported) functions is generalized to
u > D U by taking a sequence of continuous functions uj > C U such that uj u in the
ª
sense of distributions and define P u lim P uj
Example 2.4.6. Prove that if u > D U such that ∂j u

0 then u is a constant. This is
trivial with the approximation method:
Let uδ ϕδ u, this converges to u distributionally. But we also know that ∂j uδ ϕδ ∂i u 0,

therefore uδ 0, therefore u must be zero.
Example 2.4.7 (Differentiation of the characteristic function). Recall last time we

computed that H δ0 .
Now let U ` Rd be an open domain. Define 1U χU x as the characteristic function.

Then the claim is that if U has a C 1 boundary then we have:
∂j χU ν∂U j dS
with dS the Euclidean surface element. So `ϕ, dS e R∂U ϕS∂U dS.

This actually implies the divergence theorem. Let bj > C ª
Rd , then we have:
d d
SU © b d1U , Q ∂j bj i Q a∂j 1U , bj f S∂U ν bdS
j 1 j 1
Proof. Note that ∂j 1U is supported in ∂U . It suffices to compute this on balls that cover
∂U . Then B 9 ∂U xd γ x1 , . . . , xd 1 and B 9 U xd @ γ 1 , . . . , xd 1 . Then we have:

1 1
1U x lim hδ
γ x , . . . , xd 1
x
d
δ 0
set ṽ ∂1 γ, . . . , ∂d 1 γ, 1. Then we can differentiate 1U by differentiating all the hδ ’s. By

computaion we have:
1 1
∂j h h δ

γ x , . . . , xd 1
x δ
d 1
ṽj
now for each x1 , . . . , xd 1 as δ goes to zero, this converges to the distribution δγ x1 ,...,xd1 xd ,

therefore for each ϕ > C0 U with supp ϕ > Bα we have:

ª
`∂j 1U , ϕe SB α
ṽj ϕx1 , . . . , xd 1 , γ x1 , . . . , xd
1
dx
1
dxd 1
»
S∂B α
ν 1 S©γ S2 ϕx1 , . . . , xd 1 , γ x1 , . . . , xd 1
dx
1
dxd
1
S∂U νj ϕxdS
Where I dropped a negative sign somewhere.
– 35 –
2. Distributions
Lecture 11 (10/3)
Today will be the last class where we will discuss distribution theory. We will:
introduce the notion of fundamental solutions
discuss operations between distributions
Let’s consider multiplication of distributions.
º
Take
º
u, v > D U . What does u v look like?

This doesn’t always work, for example 1~ x and 1~ x are locally integrable, therefore they
define distributions, however, their multiplication gives 1~x which isn’t locally integrable.
So in general, this is impossible, but if for all x > U , at least one of u or v is smooth, then
u v is well defined.
Definition 2.4.2 (Singular Support). Let u > D U . We say that u is smooth on an

open subset V ` U if ù, ϕe R ũϕ for some ũ > C and for all ϕ > C0 U with supp ϕ ` V .
ª ª
We define the singular support as:

singsupp u U V ` U an open subset in U on which u is smooth
Intuitively, this is the collection of points where the distribution fails to be a smooth
function
Proposition 2.4.2. If the singular supports of u and v are disjoint, then their product is
well-defined.
Proof. Construct a smooth function χ such that χ 1 on singsupp U with supp χ ` U
singsupp u. Then for all ϕ > C0 U we have ϕ χϕ 1 χϕ. So supp χϕ ` U singsupp u
ª
and supp1 χϕ ` U singsupp v.
Then ùv, ϕe ùv, χϕe ùv, 1 χϕe `v, uχϕe ù, v 1 χϕe
Remark 2.4.4. The approximation method is more flexible than the adjoint method.
The converse of Proposition 2.4.2 is false as the following example shows:
Example 2.4.8. In R2 , let u dSx 0 and v dSy 0 . So now supp u singsupp u x 0
and supp v singsupp v y 0. Now singsupp u 9 singsupp v 0 but uv is well-defined.
We have u limδ 0 δ 1 ψ δ 1 x, ψ > C0 R with

ª
Rψ 1. and v defined similarly. Then
we have uv“ ” limδ δ 2 ψ x~δ ψ y ~δ δ0

Now let’s discuss the convolution of two distributions. In general you cannot define this.
Proposition 2.4.3. If u and/or v (distributions) have compact support, then u v is well
defined and moreover u v v u and supp u v ` supp u supp v
Proof. welldefined: recall when v is a function ù v, ϕe ù, v
ϕe, where:
v ϕy

S v x y ϕxdx `v, ϕx e
without loss of generality, suppose that supp v is compact. Then since ϕ > C0 Rd then ª
v ϕ > C Rd even when v > D U and suppv ϕ ` supp v supp ϕ. Therefore
ª
v ϕ > C0 Rd therefore ù v, ϕe ù, v ϕe is well-defined.

ª
– 36 –
2. Distributions
2.5 Fundamental Solutions

By the previous proposition, for any u > D Rd we can always make sense of u δ0 δ0 u

which is by definition just u. If u > C0 Rd , then ux `δx , ue `δ0 x , ue δ0 ux.
ª
We can write:
`δ0 x , ue S δ0 x y uy dy“ ” ux
Example 2.5.1. If V is a finite dimensional vector space, then we can write v Pi viei if
ei span V .
Now suppose that P u v where P is a linear operator from U to V , then if we are

interested in the existence of solutions to this equation follows if we know P ui ei . Then if
v > V with v P vi ei then u P vi ui solves P u v
So the above is like saying δ0 x span the whole space C0 Rd
ª
Let’s consider P as a linear scalar partial differential operator of the form:

P u Q aα x D α u
SαSBk
And aα is a real-valued smooth function. Then P u makes sense for u > D Rd
Definition 2.5.1 (Fundamental Solution). Given y > Rd , we say Ey is a fundamental

solution for P if it solves the equation P Ey δy δ y
The application is that if we would like to solve P u v we expect that u“ ” R v y Ey x
to be a solution (in analogy to the finite dimensional vector space example)
v x S v y δ0 x y dy
So we have (hopefully)
PS v y Ey xdy S v y δ0 x y dy v x
but these equalities need to be checked in a case-by-case basis.
2.5.1 Uniqueness
Example 2.5.2. Let U and V be finite dimensional vector spaces and P be a liner operator.
Let P u v. Let ϕ > V then:

ù, P ϕe `P u, ϕe

`v, ϕe (10)
With P V

U the adjoint. Suppose there exist a set of vectors θi that span U . If for

all θi we can find E i > V such that P E i θi , then by (10) (replacing ϕ with E i )

au, θ i f av, E i f
for all i. If our dual space could seperate points, then this would give uniqueness. Because if
u1 , u2 both solve P u v, then we must have that ù1 , θi e ù2 , θi e for all θi > U .
– 37 –
2. Distributions
Now given P define the formal adjoint as:

P v Q 1 α Dαaαu

S S
S α S BK
note that if one of u or v is smooth an compactly supported, then `P u, v e ù, P v e (this
follows by integrating by parts).

Consider the fundamental solutions Ey x to P Ey δ y , then you expect a charac-

terization of u in P u v:
S S P uxEy xdx S uxP Ey xdx S

?
v xEy xdx

uxδ0 x y dx u y
Example 2.5.3. Let P ∂x on R. We know that

d d
H δ0 so H x y δ0 x y δy
dx dx
therefore H is a fundamental solution for ∂x .
So let’s suppose we would like to solve ∂x u u. Then by our formula we have:

x
ux S v y H x y dy S ª
v y dy
The adjoint operator is just P

∂x therefore the solutions we are looking for with
P
Ey x δ0 x y . Just let Ey x

H y x. Then compute:
y
S S
?
uy v xH y xdx v xdx
ª
when v > C0 R, we have:

ª
S ∂x uH y xdx S ux∂x H y xdx u y
Suppose we are interested in solving ∂x u v on a, b.
S 1a,b x∂x uxH y xdx S ux∂x 1a,b xH y xdx
Then since ∂x 1a,b δa δb we get that the above becomes:
uy uaH y a ubH y b
because y @ b. And so we get that uy ua Ra vdx and we have the fundamental theorem
y
of calculus.
Lecture 12 (10/8)
Today we will cover the Laplace equation (Evans 2.2).
Recall we were discussing the fundamental solution. If P is a linear operator of the form:
P Q aαDα
SαSBk
– 38 –
2. Distributions
we say that Ey is a fundamental solution for P at y > Rd if P Ey δy δ0 x y .
Once we have this, we can prove existence. Given a nice function f , then if we formally
write:
u f S f y Ey xdy
Then we expect that Pu f f.
We also get uniqueness, we look at the adjoint of P which is characterized by:

` P u, ve ù, P ve
after integrating by parts, we get that:

P v Q

1α Dα aα v
S α S Bk
Then the fundamental solution can be defined as P E x

δx . So now given a nice u, we
have:
ux ù, δx e ù, P E xe `P u, E xe

Then we can recover P u (which gives us uniqueness). This formula above we will call the
representation formula for u.
We also have uniqueness for a boundary valued problem. This is the same idea where we
write for x > U :
ux ù, δx e `1U u, δx e `1U u, P E x e

`1U P u, E xe
terms involving © 1U
But these extra terms give us the contributions of the boundary values of u.
Remark 2.5.1. If you consider constant coefficient P. This is the case where aα are con-
stants. There are two simplifications that occur.
1. if P E0 δ0 then Ey x E0 x y is a fundamental solution for P at P E0 y
δ0 y δy . So we get translational invariance.
2. P v P α Bk 1 α aαDαv P v x. Therefore if we set E xy

S S
S S

E0 x y then
Py E x P E0x y δ0x y δx

So we will now only consider the constant coefficient P . Then Ey x E0 x y and
E
x y E0 x y . Then we have that:
u f S f y E0 x y dy f E 0 x
So for uniqueness, E
x y E 0 x y :
` P u, E xe S P uyE0x

y dy Pu E0 x
– 39 –
3. Laplace Equation
3 Laplace Equation
Now we will consider P ∆. Let’s begin by deriving fundamental solutions to the Laplacian.
That is solve ∆E0 δ0 in Rd for all d C 2. We already did this. We will use rotational
invariance of ∆. That is given a rotation matrix R on Rd , then we have:
∆U Rx ∆uRx
Then we can look for E0 E0 r. So we would like a∆E0 , 1B 0,r f aδ0 , 1B 0,r f 1.
Integrating by parts the thing on the left, we get:
x
S ©E0 © 1B 0,r S © E0 νdS∂B 0,r S SxS
©E0 dS∂B 0,r S∂B 0, r S∂r E0 dαdrd 1 ∂r E0 r

we say αd is the volume of the unit ball in Rd . Therefore:

1
∂r E 0 r
dαdrd 1
So we have the fundamental solution

¢̈ 1
¨
¨ logr d 2
¨ 2π
E0 r ¦ 1 1
¨
¨ dC3
¨ d2
¤̈ dd 2αd r
Remark 3.0.1. There is no universal way to determine fundamental solutions.

Existence: of © u f in Rd . We have u f f E0 x which is:
1
cd S f y dy
Sx y Sd 2
as long as f > D Rd with compact support, ∆u f f

Uniqueness: Take u > D Rd with compact support. Then u

∆u E0 (what?)
Theorem 3.0.1 (Analytic regularity to solutions to ∆u 0). If u > D R d

is a
solution to ∆u 0 (i.e. u is harmonic). Then u is smooth and analytic.
Proof. We have (since ∆u 0):
∆χu ∆χu 2©χ © u
Now if x > Rd , choose χ > C0 Rd such that χ
ª
1 in a neighborhood of x, then © χ x 0
near x.
Now:
χu ∆χu E0 S ∆χu 2©χ © uy E0 x y dy
Now E0 is not smooth only when x y. By smoothness of E0 x y outside of x y, then

χu is smooth.
– 40 –
3. Laplace Equation
Theorem 3.0.2 (Derivative estimate for a harmonic function). If u is harmonic then

there exists a constant C C SαS such that
1
SD
α
uxS B C
rd Sα S SB x,y

SuSdy
Proof. Take χ > C , χ ª

1 for x > B x, r~2, supp χ ` B x, r, S©χS B c~r. Then:
Dα ux Dα χux S ∆χuy 2©χ © uy Dα E0 x y dy
Then ∆χ is supported in the annulus of radii r~2 and r. Then we get control of the thing
by:
C
rdSαS SB x,r

SuSdy something
where something comes from integration by parts.

Theorem 3.0.3 (Liouville). If u is a bounded harmonic function defined on Rd then u is
constant.
Proof. Use the above theorem.
Theorem 3.0.4. Given f > D Rd with compact support (d C 3). Any bounded solution to

∆u f is of the form:
u u f c
for some constant c.

Proof. If u, u are solutions, then u u is bounded and harmonic and therefore equal to a

constant.
Theorem 3.0.5 (Mean Value Property). If u is harmonic then:
1 1
ux S
S∂B x, r S ∂B x,r
udS y S
SB x, r B x,r
u y
Proof. Note that the two are equivalent (omitted)
We will just prove the fist equality.
ux S 1B x,r y uy ∆E0 x y dy S ©1B x,r y uy © E0 x y dy
S ν y u y © E0 x y dS∂B x,r S 1B x,r y ©u © E 0 x y y
S uy ν © E0 x y dS∂B x,r S © 1B x,r y ©uE0 x y dy
the second term is just:
S ν © uE0 x y dS∂B x,r
– 41 –
3. Laplace Equation
So we ultimately get:
ux S ν © uE0 x y dS∂B x,r S uy ν © E0 x y dS∂B x,r
note that the second term is:

1
S
udS
And the first term can be shown to be zero by the divergence theorem(possibly?).
Lecture 13 (10/15)
Day of the midterm Lecture 14 (10/17)
Today we will spend 20-30 minutes discussing the midterm and review what we have done
so far in the course.
General Theory
1. For first-order scalar PDE’s we studied the method of characteristics. The idea is that
we can always reduce PDE’s to first order ODE’s which then we can use ODE results
to get existence and uniqueness. This required noncharacteristic boundary data. In
the quasilinear case, the noncharacteristic boundary data is equivalent to the condition
that we are able to compute all derivatives of the solution at the boundary point. We
extended this to general quasilinear k-th order systems (see Holmgien’s theorem about
an application). We then discussed Cauchy Koweleskaya, but it isn’t the end due to
Hadamard’s example. He emphasized well-posedness (which asks not only existence
and uniqueness, but also continuous dependence).
2. Then we turned to constant coefficient 2nd order scalar PDE’s. We are studying the
four equations:
¢̈∆u 0
¨
¨
¨
¨
¨
¨ju 0
¦
¨
¨ ∂t ∆u 0
¨
¨
¨
ï∂ ∆u
¤̈ t 0
this can be studied on a case by case basis (like in Evans), but we can study them at
the same time using distribution theory. This allows us to make use of the fundamental
solution: P u δy and P u δx . We gen existence for P u f . And we get uniqueness:

if u is nice (eg compact support) then ux E0 P u, u is only defined on Ω so

ux E0 P u R∂Ω
What’s to come: for the next 2-3 weeks we will focus on these equations, then we will
discuss Sobolev spaces. We wil be studying A-priori estimates. We will be studying:
¢̈
¦
P
¨ u f in Ω
ü g on Γ
¤̈
– 42 –
3. Laplace Equation
we first assume a solution exists, then we prove estimates on them. For existence we would
like ù, P ϕe `f, ϕe. This is something we can define if P is injective. Then the linear

functional ù, P ϕe is well-defined in span P ϕ ϕ > C0 U . This strategy favors L2 type
ª
spaces, in which we can find solutions u > L2 , and all derivatives will stay in L2 . A Sobolev
spcae is H k u Dα u > L2 , SαS B k .
3.1 Cauchy Riemman Equation

Let’s apply what we have learned to the Cauchy Riemman equation for f u iv:
∂x u ∂y v 0
∂y u ∂x v 0
This is equivalent to ∂x i∂y f 0. Note that if this is true, then each component of f is
harmonic as ∂x i∂y ∂x i∂y ∂x2 ∂y2 ∆. Now in 2 d the fundamental solution is:
1
∆ log r δ0
2π
1 1
Decomposing ∆, we get the fundamental solution ∂x i∂y log r x~r 2 iy ~r 2
2π 2π
1
2πz
1 1
Theorem 3.1.1. E0 is a fundamental solution for ∂x i∂y
2π z
Theorem 3.1.2. If ∂x i∂y f 0 for f > D U then f > C
ª
U
Proof. Let z0 > U , let χ be 1 near z0 and zero outside a small ball centered at z0 . Then:
∂ x i∂y χf f ∂x i∂y χ
Then since χf is compactly supported, we can use the representation formula to get that:
χf E0 ∂x i∂y χf E0 ∂x χ i∂y χf z

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
F
But this last term vanishes near z0 , if we let that term be F , then we have (if F is a function):
S E0 z wF wdwx dwy
which is smooth. If F isn’t smooth, then we use approximation
Corollary 3.1.1 (Morera’s Theorem). If f is C 1 on U and RΓ f z dz for all closed curves
Γ, then f is holomorphic (i.e. f is a C solution to CR equation)
ª
– 43 –
3. Laplace Equation
Proof. We say that f is a distribution solution if and only if `f, ∂x i∂y ϕe 0 for all ϕ >
C0 U complex valued. Where the inner product is now defined as ù, ve R uzvzdxdy
ª
(which isn’t real valued anymore). We have that if:
0 S f z ∂x i∂y ϕz dxdy
for all ϕ then f is smooth.

Lemma 3.1.1.
1
S hz ∂x i∂y χΩ dxdy
i S∂Ω
hz dx idy
By hypothesis, we have that `f, ∂x i∂y χΩ e 0. Let ϕ χΩ be an approximation,

then we can approximate ϕ be step functions in L1 . Then `∂x i∂y f, χΩ e 0 implies that
`∂x i∂y f, ϕe 0 for all ϕ > C0 U ª
Theorem 3.1.3 (Cauchy Integral formula).
1 f z
f z0 S
2π ∂Ω z z0
dz
Proof. (uniqueness in the case of u on Ω). We have:
f z0 f z0 χΩ z0 S δ z z0 χΩ z0 f z dz S ∂x i∂y E0 z0 z χz f z dz
then we integrate by parts, when the differential operator falls on f we get zero, when it
falls on χ we use the lemma to get:
1 1 1 dz
2π S z0 z
∂X i∂y χΩ z f z dxdy S
2πi ∂Ω z z0
f z
where there was a sign error.

Lecture 15 (10/22)
Today we will continue the discussion of the Laplace equation. We were looking at:
∆u f
what we did last time was we computed the (radial) fundamental solutions for ∆. That is
∆E0 δ0 which had the form:
¢̈ 1
¨
¨ log SxS d 2
¨ 2π
E0 x ¦ 1
¨
¨ SxSd2 dC3
¨
¤̈ dα d d 2
with αd the volume of the d-dimensional unit ball.
Recall that we proved:
– 44 –
3. Laplace Equation
existence of a solution u f to ∆u f for f a compactly supported continuous function

on Rd where:
u f x E0 f x
uniqueness due to the representation formula. When u is compactly supported:
E0 ∆u u
combined with the regularity (smooth, analytic) of E0 outside x 0 leads to smooth-

ness (and analycity) of u near x > U , ∆u 0 in U
derivative estimate: given ∆u 0 in U and B x, r ` U then:
Cα
r1 SαS SB x,r
SD
α
uxS B
SuSdy
Liouvile theorem: if u is a harmonic bounded function on Rd , then u is constant.
in d C 3, any bounded solution to ∆u f in Rd is of the form u u f c where c is

some constant.
see the notes for d 2
We said that u has to be a compactly supported distribution for: u E0 ∆u to hold.
If u is a distribution solution to ∆u 0 in U . Then we truncate and approximate u by
distributions with compact support.
3.2 Boundary Value Problem

Now will try to apply the representation formula for a boundary valued problem. Suppose
we are solving for u in U . So we have for x > U :
ux 1U ux δ0 1U ux ∆E0 1U ux Q ∂j ∂j E0 1U ux
j
Q ∂j E0 ∂j 1U ux Q ∂j E0 ∂j u1U x Q∂j E0 u∂j 1U x

j j j
E 0 ∆u1U x Q E0 ∂j u∂j 1U x Q∂j E0 u∂j 1U x

j j
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
B C
where we used the fact that 1U u has compact support. Recall that ∂j 1U νj dS∂U . So C
has a term that becomes:
Q ∂j E0 u∂j 1U x Q S∂U ∂y E0x y uy νj dS y S∂U ν DE0 x y uy dS y
j j
And B becomes:
Q E0 ∂j u∂j 1U x S∂U ν Duy E0 x y dS y

j
– 45 –
3. Laplace Equation
Theorem 3.2.1. If u is a smooth function on Ū then:
ux SU E 0 x y ∆uy dy S

∂U
E0 x y ν Duy dy S
∂U
ν Dy E0 x y uy dy
Corollary 3.2.1 (Mean Value Property). If u is a harmonic function in U such that

B x, r ` U , then:
1
u x
dαdrd 1 SB x,r
uy dS y
Proof. Apply the above theorem with:
Ẽ0 x E0 x E0 r
where E0 r is the value of E0 y for Sy S r. Then we get:
u x S∂B x,r
ν Dy Ẽ x y uy dS y
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∂r Ẽ0 r
1
but we know that ∂r E
dαdrd 1
Remark 3.2.1. The same proof applies to ∆u x 0. This leaves to a nice proof of Jensen’s
formula in complex analysis.
Theorem 3.2.2 (Maximal Principal). Let U be a bounded C 1 -domain (open and con-
nected). Let u > C 2 U 9 C 0 Ū with ∆u 0 in U then:
1. (weak maximal principal)
max u max u
Ū ∂U
2. (strong maximal principal): if x0 > U and ux0 maxŪ u, then u is constant.
Proof. It suffices to prove the strong maximal principal. Let M maxŪ u. Let B x, r ` U .
Then:
1
M ux0 S
udS y B M
this means that u M on ∂B x, r (because we have R∂B M u dS y ).

´¹¹ ¹ ¸¹¹ ¹ ¶
We can do this
C0
for all balls in U , but this implies that u M in B x, r. Then by connectedness, we are
donea
a
let Z x > U ux M . This is open by the aboe computation and closed as it is the preimage of a
closed set, therefore it is U
– 46 –
3. Laplace Equation
Corollary 3.2.2 (Uniqueness of Dirichlet Problem). Given U a bounded C 1 domain.

The Dirichlet problem is:
¢̈
¨∆u f in U
¦
ÿ g on ∂U
¤̈
note that ν Du is not prescribed. If f and g are continuous and u > C 2 U 9 C Ū then u
is unique.
Proof. If u and u solve this, then u u is harmonic with boundary value 0. Then by the

maximal principal we get that u u 0 (show B and C).

Remark 3.2.2. This uniqueness theorem is a global phenomenon!

Remark 3.2.3. The uniqueness that comes from the maximal principal implies the existence
of a representation formula.
Proof. (sketch) Given x > U with ∆u 0. If we take uS∂U > C ∂U and map it to ux > R.
Then this map is bounded by the maximal principal. Then we look at the linear subspace
V uS∂U of ∆u 0 in U, u > C 2 U 9 C Ū . Then we use Hahn-Banach.
Note that this is almost useless, as we do not get what the formula is, just that it exists.
Theorem 3.2.3 (Harnack’s inequality). Given u > C 2 U 9 C Ū , ∆u 0, u C 0. Then
on any bounded domain V ` V̄ ` U :
sup u B C inf u
V̄ V̄
Where C depends only on U and V

Proof. We will apply the mena value property for balls:
1
ux S
SB x, r S B x,r
uy dy
1
this follows by averaging R udS for 0 @ r @ r.

Now take r P distV̄ , ∂U . Then take x, y > V such that Sx y S @ r. Then we can write:
1 1 2d
ux S
SB x, r S B x,r
udz B S
SB x, r S B y,2r
udy S
SB y, 2r S B y,2r
udy B 2d uy
using that B x, r ` B y, 2r. This tells us:
2 d uy B ux B 2d uy

then we cover U with balls of radius r. So if we jump N balls, we get:
2 dN
uy B ux B 2dN uy
since V̄ is compact, we know that N is bounded and we are done.
– 47 –
3. Laplace Equation
Lecture 16 (10/24)
Recall our discussion of the Laplace equation. We got uniqueness of a solution to the Dirchlet
problem for a bounded domain U
¢̈
¨∆u f in U
¦
ü g on ∂U
¤̈
If f, g are continuous and u > C 2 U 9 C Ū , then we have uniqueness.
3.3 Green’s Function

Definition 3.3.1 (Green’s function). Given a domain U and some y > U , we say that
G , y > D U is a Green’s function for U at y if

¢̈
¨∆x G , y δ0 x y in U
¦
¨Gx, y 0
¤̈
x > ∂U
Remark 3.3.1. This definition is opposite to Evan’s definition.

Note that Gx, y E0 x y is a harmonic function. So one way to find Gx, y is to take
Gx, y E0 x y hy x such that h solves:
¢̈
¨∆hy x 0 in U
¦
¨hy x E0 x y
¤̈
f or x > ∂U
So, existence theory for the Dirichlet problem implies the existence of a Green’s function.
Now, if you have a Green’s function, you have existence of a solution u f to:
¢̈
¨∆u f f in U
¦
ü f 0 on ∂U
¤̈
where we set:
u f x S f y Gx, y dy
This has to be justified, but morally speaking, one can solve:

¢̈
¨∆u 0 in U
¦
ü g on ∂U
¤̈
by working with v u g̃. We find a C ª

extension g̃ > C ª
Ū such that g̃ S∂U g. Then:
¢̈
¨∆v ∆u ∆g̃ ∆g̃ in U
¦
¨v 0 on ∂U
¤̈
– 48 –
3. Laplace Equation
So if we solve v and put u v g̃, then u solves the original problem.
So we have that “homogeneous and nontrivial boundary value” can be solved by solving
“inhomogeneous and trivial boundary value”
It turns out that existence to Dirichlet problem is equivalent to the existence of a Green’s
function.
Remark 3.3.2. It turns out that G is a C ª

function in U U y x
Theorem 3.3.1 (uniqueness and symmetry of Green’s functions). Let U be a bounded

C 1 domain and suppose there exists a Green’s function Gx, y . Then (for x x y, x, y > U :
1. If G x, y is also a Green’s function, then Gx, y

G x, y

2. Gx, y Gy, x
Proof.
G x, y

S δ0 x z G z, y dz
S 1U z δ0 x z G z, y dz

S ∆z Gz, xG z, y 1U z dz

d d
QS ∂j Gz, x∂j G z, y 1U z dz Q S ∂j Gz, xG z, y ∂y 1U z dz
j 1 j 1
d d
QS Gz, x∂j2 G z, y 1U z dz Q S Gz, x∂j G z, y ∂j 1U z dz
j 1 j 1
d
QS ∂j Gz, xG z, y ∂j 1U z dz
j 1
S Gz, xδ0 z, y dz 0 0

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Gy,x
Remark 3.3.3. We required that ∂G to be continuous up to ∂U

So now if we take G G , then we see the symmetry property for any Greens’ function.
Then we use this to see that G x, y Gy, x
Remark 3.3.4. Gx, is a solution to the adjoint problem.

¢̈
¨∆Gx, δx
¦
¨Gx, 0
¤̈
Revised definition of Greens’ function:
– 49 –
3. Laplace Equation
Definition 3.3.2. A Greens function is a solution to:

¢̈
¨∆G , y δy in U
¦
¨G , y 0 on ∂U
¤̈
and G , y > C 1 Ū y

Now let’s turn to construction of Green’s functions on some domains. Here we will use a
technique called the method of image charges.
The idea is that the fundamental solution E0 x, y models the electric potential of a unit
point charge which is situated at y. To ensure that on the boundary that the potential is
zero, we simply pretend there is an opposite charge on the other side of the boundary. So
we would get Gx, y E0 x y E0 x ȳ , where ȳ is y reflected across the boundary
Example 3.3.1. Let U B 0, 1.
If we look for Sx y S Sx ȳ S then we get hyperplanes. Recall that for d C 3 we have:

1 1
E0 X
dαdd 2 SxSd 2
so we get that
1
E 0 x E0 λx
λd2
for λ A 0
So now we can consider the hypersurface for R A 1 P x > Rd Sx y S RSx ȳ S.
On the plane, this is the circle.
In this case for y > U :

Gx, y E0 x y E0 Sy Sx ȳ
y
where ȳ
Sy S2
Proof. (that this is the Green’s function) Let y > B 0, 1. It suffices to check that Sx y S
Sy SSx ȳ S for x > ∂B 0, 1 if and only if SxS2 1.
But this is just:

2
SxS 2x y Sy S2 2
Sy S SxS
2
2x ȳ Sȳ S2
but the right side is:
2 2 y Sy 2 S 2 2
Sy S SxS 2x Sy S SxS 2x y 1
Sy S2 Sy 4 S
– 50 –
3. Laplace Equation
In Summary we have the following Greens functions:

1. For Rd
x > Rd xd A 0, then if y > Rd let ȳ

y 1 , . . . , y d1 , y d . Then we set:
Gx, y E0 x y E0 x ȳ
y
2. For B 0, 1 we have y > B 0, 1 ȳ then:
Sy S2
Gx, y E0 x y E0 Sy Sx ȳ
Theorem 3.3.2 (Representation formula for the Dirichlet problem). Let u > C 2 U 9
C Ū , then we have that:
ux SU Gx, y ∆uy dy S

∂U
ν y Dy Gx, y uy dS y
Proof.
ux S δ0 x y uy 1U y dy S ∆y Gx, y uy 1U y dy
d d
QS ∂yj Gx, y ∂yj uy 1U y dy Q S ∂yj Gx, y uy ∂yj 1U y dy
j 1 j 1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
A
d
A S Gx, y ∆uu1U y dy Q S Gy, x∂j u∂j 1U dy
j 1
First assume that u > C Ū . Then the first and third term is zero by Gx, y
ª
0 for y > ∂U .
For general u, you can approximate u in C 2 U 9 C Ū .
Theorem 3.3.3 (Poisson integral formula). Given g > C ∂U (for today U is the half
plane or the unit ball). If we define:
ux S ν y Dy Gx, y g y dy
then u > C 2 U 9 C Ū and u g on ∂U and ∆u 0

Lecture 17 (10/29)
Recall that we were discussing the Green’s function for the Laplace equation. It solved:
¢̈
¨∆G , y δ y in U
¦
¨G , y 0 on ∂U
¤̈
For a domain U ` Rd with y > U . We found that G is unique and that Gx, y Gy, x. So
we also have:
¢̈
¨∆Gx, δ x in U
¦
¨Gx, 0 in ∂U
¤̈
which is the definition in Evan’s.
– 51 –
3. Laplace Equation
Theorem 3.3.4 (Poisson Integral Formula). Given U a C 1 -domain and u > C ª

Ū ,
then:

∂U ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Poisson integral kernel
Remark 3.3.5. Note that this representation formula does not have a term with the normal
of u.
Remark 3.3.6. ∆u 0 is what is usually called Poisson integral formula.
Remark 3.3.7. If we want to solve:

¢̈
¨∆u g in U
¦ (11)
ü h in ∂U
¤̈
If a solution exists, then:
u x SU Gx, y ∆uy dy S

∂U
ν Dy Gx, y udS
In fact, we can define:
ũx SU Gx, ygydy S∂U ν gDy Gx, yhydS y

it turns out that ũx is indeed the solution to the Dirichlet problem (11), but we will not
justify this here.
It isn’t hard to show that ∆ũ g, it is difficult to show that ũx ÐxÐÐÐÐ
x >∂U 0
hx0
We had examples of Green’s functions using the method of image charges.
Example 3.3.2. When U Rd

xd A 0. In this case, we had:
Gx, y E0 x y E0 x ȳ
with y y 1 , . . . , y d1 , y d
Example 3.3.3. When U B 0, 1, we used the fact that if y > B 0, 1 and let ȳ Sy S2 y,
then we have that:
∂B 0, 1 x > Rd Sx y S Sy SSx ȳ S
for any y > B 0, 1. This suggested that we take:
Gx, y E0 x, y E0 Sy Sx ȳ
if y 0, then we set Gx, y E0 x, y 0 E0 x, y
– 52 –
3. Laplace Equation
Given an integral operator:
T f x S K x, y f y dy
then K is the kernel of the integral operator. By the Schwartz kernel theorem, any linear
T C0 X D Y is of the form T f `K x, , f e
ª
Now we would like to compute the Poisson kernel: ν y Dy Gx, y for the two above
examples.
Example 3.3.4. Swap x and y, set:
Gx, y E0 y x E0 y x̄
ν y Dy on ∂U xd 0 is just ∂yd , so we have:
1 y d xd 1
∂yd E0 y x ∂yd Sy xSE0 Sx y S

dαd Sy xS Sy xSd1
where we used that:

1
1
E 0 x

dαd r 1
d
If we restrict to y y , 0, then we get:
1 xd
αdd Sy xSd
Then we compute:
1 x̄d
∂yd E0 y x̄
dαd Sy x̄Sd
1 xd
then restrict to y y 1 , 0, to get . So our Poisson kernel becomes:
dαd Sy xSd
2xd 1
ν y Dy Gx, y
dαd Sy xSd
there was a sign error, the final answer is right. So the Poisson integral formula for harmonic
u in Rd is:

2xd 1
ux S dαd Sy xSd
hy dy
Remark 3.3.8. ux hx0 as x x0 if h > C ª

∂U
– 53 –
3. Laplace Equation
Example 3.3.5. For the ball, we have:
Gx, y E0 y x E0 SxSSy x̄S
and again we will use that:

1 1
E 0 r

dαd rd 1
then we have that on ∂B 0, 1:
ν y Dy Q y α ∂α
α
Then we have:
1 1
Q y α ∂y α E0 y x Q yα∂y αS y xS
dαd Sy xSd 1

α α
Sy S2y x 1

dαd Sy xSd
And we have:
1 1
Q yα∂y α E0 SxSSy x̄S Q y α ∂y α S xSSy x̄S
dαd SxS Sy x̄Sd
d 1 1

α α
1 1 Pα y α y α x̄α 1
dαd SxSd 2 Sy x̄S

Sy x̄Sd 1
1 1 Sy S2 yx̄
dαd SxSd 2 Sy x̄Sd
then if we restrict both to ∂B 0, 1, then we get:

1y x 1

dαd Sy xSd
for the first term, and the second term becomes something else, and after we add them up,
we get:
SxS2 1 1
dαd Sy xSd
Note that if u is harmonic in B 0, 1, we get:
1 SxS2 hy
u x S∂B 0,1

ν y Dy Gx, y hy dS y S∂B 0,1
dαd Sy xS
d
dS y
– 54 –
4. Wave Equation
4 Wave Equation
The wave equation is an equation on ϕ R1 d R where the coordinates of the domain are

x0 , x1 , . . . , xd with x0 t and our operator is:
1 2
j ϕ ∂ ∆ϕ 0
c2 t
this is called the d’Alembertian operator. This is the prototypical hyperbolic equation.
We always take c 1 in math courses. We have three goals:
1. find an explicit fundamental solution for j. We will figure out that there are a lot of
symmetries of j.
2. derive an (explicit) representation formula for:

¢̈jϕ f, in R1 d
t A 0
¨
¨
¨
¦ϕ g, on ∂R1 d (12)
¨
¨

¨∂ ϕ h on ∂R1 d
¤̈ t
3. Prove existence of a solution to (12)
4.1 1 dimension
Let’s consider the domain R1 1 . Then we have:

j ∂t2 ∂x2 ∂t ∂x ∂t ∂x
So let’s take advantage of this factoring by finding the fundamental solution. Let’s make a
change of coordinates u t x, v t x. Then we get that du dt dx, dv dt dx threfore
1 1 ∂t ∂x 1 1
dt du dv and dx du dv this means that ∂u ∂t ∂x ∂t ∂x and
2 2 ∂u ∂u 2 2
1 1
∂v ∂t ∂x .
2 2
Then if E is our fundamental solution, we have:
j E0 4∂u ∂v E0 u, v 2δ0 u, v
where we had to use the approximation method to compute δ with a change of variables. So
we just need to solve:
1
∂u ∂v E0 δ0 uδ0 v
2
So that:
1
E0 H u c1 H v v2
2
– 55 –
4. Wave Equation
The forward fundamental solution E is defined so that E is a fundamental solution and

supp E ` t C 0. This forces c1 c2 0. Therefore the fowards fundamental solution has

the form:
1 1
E H u H v H t x H t x
2 2
Lecture 18 (10/31)
Let’s rectify what was messed up last time.
Theorem 4.1.1 (Change of Variables for Distributions). Suppose that Φ is a diffeo-

morphism (smooth change of coordinates) from X1 to X2 (open subsets in Rd . And suppose
that u > D X2 , then we can define u X Φ, by for all ϕ > C0 X1 :
ª
1 1
ù X Φ, ϕe du, ϕXΦ i

S det DΦS
Remark 4.1.1. u (u X Φ is continuous and linear in u.
Proof. First assume that u > C0 X2 ,then let y 1 , . . . , y d

ª
Φx1 , . . . , xd , then we have:
∂ x 1 , . . . , x d
S uΦxϕxdx S u y ϕ Φ 1
y W
∂ y 1 , . . . , y d
W dy
1 1
du, ϕΦ y i

S det DΦy S
Now since Φ is a diffeomorphism, we have that we still get a test function for the input of
u
1
Corollary 4.1.1. δ0 X Φ δΦ1 0
S det DΦS
Let’s now return to the wave equation:

2
j ϕ ∂t ∆ϕ 0
this is the sign choice that people usually use in general relativity. For R1 1 , we got:
1
E0 H t x c1 H t x c2
2
Then we introduced something called the forward fundamental solution, which required
that supp E ` t C 0. This is the relevant choice for solving the wave equation because we

would like to know what happens in the future. This requires jE δ0 and E 0 before
seeing δ0 . Then we get c1 c2 0, and get the unique forward fundamental solution:
1
E H t x H t x
2
– 56 –
4. Wave Equation
Our goal today is to derive a representation formula using E . In 1 1 dimension, we get
that for all ϕ > C Rd , for all t, x > Rd then, we get d’Alembert’s formula
ª

1 t x t s
1 1 x 6
2 S0 Sx t s 2 Sx t

ϕt, x jϕ s, y dyds ϕ 0, x t ϕ 0, x t ∂t ϕ0, y dy
2
Note that there is a sign change in Evans due to the choice of sign of j.
Let’s turn to the derivation of the reresentation formula for any dimension (given that
we know the forward fundamental solution). Then we will go towards a way of deriving
fundamental solutions for all dimensions.
Let E be a forward fundamental solution to

j in R1 d . In other words we get that:

¢̈
¨jE δ0
¦
¨supp E ` t C 0
¤̈
Assume in addition that for all bounded interval
supp E 9 t > I (13)
is compact. It turns out (13) is required for the following

Lemma 4.1.1. Let u > D R1 d , and supp u ` t > L, ª, then u E is well-defined.
Proof. (sketch) Suppose that u is a C ª

function and E is a function. Let’s make sense of
u E t, x, this is just:

u E t, x U us, y E t s, x y dsdy
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
compactly supported
Proof. (real) Given an interval I a, b with a, b > R 8 ª, let χI be a smooth function
that is 1 on I and 0 in s B a1 and s C b1. Then we will consider χI uE for all bounded I.
Now for ϕ > C0 Rd , we have:

ª
?
`χI u E , ϕe ù E , χI ϕe du, E χI ϕ i
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
need to show is a test function
Actually we will prove that χ L, ª E χI ϕ . For functions, we had:
f

g y S f x y g xdx
then we had supp f

g ` supp f supp g.
– 57 –
4. Wave Equation
Now suppϕχI ` t > I , then we would like to show that supp f

g ` L, ª. Now
E χ M, E 1 χ M, E . Now:
ª ª
suppχ M, E ª

χI ϕ ` t C M t > I
Yeah...I got a little lost, this is one of those things that is easier to just work out by yourself.
Theorem 4.1.2 (Uniqueness of forward fundamental solutions). Let E be a forward
fundamental solution with the property (13) and let E be any forward fundamental solution

to:
¢̈
¨jE δ0
¦
¨supp E ` t C 0
¤̈
Then E

E

Proof. We have:
E δ0 E jE E0 E E E E E δ0 E

j j

Theorem 4.1.3. Let ϕ > C ª

Rd , t, x > Rd . Then:

ϕt, x E δt ϕδt 0 ∂t Et ϕ∂t 0 Et jϕ1tC0
Proof. We have:
ϕ1tC0 δ0 ϕ1tC0 jE ϕ1tC0
2
∂ E ϕ1tC0 ∆E ϕ1tC0
t
∂t Et ∂t ϕ1tC0 ∂t E ϕδt 0 E ∆ϕ1tC0

2
E ∂t ϕ1tC0 E ∂t ϕδt 0 E ∆ϕ1tC0 ∂t E
ϕδt 0
Looking at the representation formula and pretend everything is a function, we have:

ª
E jϕ1tC0 t, x S0 S j ϕs, y E t s, x y dyds

Then if you believe that supp E ` SxS B t, then you are integrating over the triangle below

the thing.
1
Let’s specialize to d 1, then E H txH tx, then let’s look at the representation
2
formula. Then we have:
1 1
2 S S 2 U

E ψ t, x ψ s, y H t s x y H t s x y dsdy ψ s, y dsdy
– 58 –
4. Wave Equation
where we are integrating over s, y t s C x y, t s C x y , s C 0. This is telling us

that x t s C y C x t s. So we get:
1 t x t s
2 S0 Sx t s

ψ s, y dyds

Then if we apply this expression to the last term we get:
1 t x t s
2 S0 Sx t s
E j ϕ jϕdyds

and we get that:

1 x t
2 Sx t
E
∂t ϕδt 0 ∂t ϕ0, y dy

and:
1 x t
1
2 Sx t
∂t E
ϕδt 0 ∂t ϕ0, y dy ϕ0, x t ϕ0, x t
2
Lecture 19 (11/5)
We were discussion the wave equation:
j ϕ 0
1
In R1 1 we were able to compute the fundamental solution E t, xH t xH t x. In
2
R1 d , given a forward fundamental solution E with the condition that for all bounded inter-

vals supp E 9 t, x t > I is compact. Then we were able to construct the representation

formula for the Cuachy problem:

¢̈
¨j ϕ f in R1 d

¦ (14)
¨ϕ, ∂t ϕ g, h in 0 Rd
¤̈
then:
ϕt, x E ∂t ϕδt 0 ∂t E ϕδt 0 E jϕ1R1d

If everything is a function, then we have:
E f S E t s, x y f s, y dsdy

then we are just integrating over the inverted cone in R1 d . This is known as “finite speed
of propagation”.
– 59 –
4. Wave Equation
Remark 4.1.2. In the case of wave equations, we can take the representation formula to
write down the solution to the initial value problem. For t, x > R1 d , define

ϕt, x E hδt 0 ∂ t E gδt 0 E f 1R1d
Then the claim is that for nice enough f, g, h (for instance all smooth) then ϕt, x does solve
the initial value problem (14).
The goal today is to find the foward fundamental solution in R1 d for d C 1. To do this,
we will exploit the symmetries of j to narrow down the possible candidates for E .
Symmetries of j, we would like to find L R1 d R1 d such that:
j ϕ X L jϕ X L
One example is would be translations. If x0 > R1 d , then we have:
ϕt, x x0
j j ϕt, x x0
This is not helpful for E because jE

δ is not invariant under translations.
We are instead interested in symmetries that fix 0, 0. These are called the Lorentz
transformations. These are:
1. (rotations), R > SOda . Then jϕt, Rx j ϕt, Rx.
2. (reflections) L t, x1 , . . . , xd t, x1 , . . . , xd or we would reflect a spacial coordinate.

Then since all derivatives are second order, then this is invariant.
3. (Lorentz boosts) Lorentz boost with speed v and direction x1 is the formula:
1
t, x ( ºt vx1 x1 vt
,º
1 v2 1 v2
Any other Lorentz transform is given by a composition of the above. Another way to
characterize these is L is a Lorentz transformation if and only if it preserves:
t2 x1 2 xd 2
Another property of the Lorentz transformation is the behavior of j under scaling. For all
(
λ A 0, given the scaling transformation t, x t~λ, x~λ, then we have:
1
j ϕt~λ, x~λ jϕt~λ, x~λ
λ2
Definition 4.1.1 (Homogeneity). We say a function ϕ Rn 0 R is homogeneous of
1
degree a if ϕx~λ ϕx
λa
a
RRt Rt R I and det R 1
– 60 –
4. Wave Equation
Example 4.1.1. If n 1 and ϕ xa for x A 0. Then this is homogeneous of degree a.
Remark 4.1.3. If ϕ is homogeneous of degree a, then jϕ is homogeneous of degree a 2
For distributions consider:
`ϕ ~λ, ψ e S ϕx~λψ xdx λd S ϕy ψ λy dy aϕ, λd ψ λ f
Definition 4.1.2. For ϕ > D Rn or

D R n

0 is homogeneous of degree a if for all
ψ > C0 Rn or C0 Rn 0 we have:
ª ª
n a
`ϕ, λ ψ λ e λ `ϕ, ψ e

Example 4.1.2. δ0 on Rn , then the degree is n.
So if we are looking for jE δ0 and j lowers the degree by 2 and δ0 has degree 1 d,

it is reasonable to look for a homogeneous E of degree d 1
Example 4.1.3. Consider ϕx xa for x A 0 and 0 for x @ 0. Then ϕ is a homogeneous

function of degree a on R 0. When a A 1, then ϕ can be extended as a homogeneous
distribution on the whole R.
ª
`ϕ, ψ e S0 x a ψ xdx

when a 1, no homogeneous distribution agrees on R.

Lemma 4.1.2. If ϕ is a homogeneous distribution of degree a A . n in Rn 0, then there

exists a unique distribution ϕ̃ > D Rn that is homogeneous of degree a and agrees with ϕ in

Rn 0.
Now we are looking for E .
Motivated by the homogeity of j and δ0 we may assume that E is a homogeneous
distribution on R1 d of degree d 1. By the previous lemma, it is enough to construct

E in R1 d 0, 0

s2 t, x t2 SxS2 is invariant under all Lorentz transformations, we will assume that
kinda d1 2
E χ 2 t SxS2
We wanted the forward fundamental solution. So we rather define E as:
d1 2 2
E 1t> 0, ª χ 2 t SxS
d1
to make sure that this is well-defined, it is good to require that χ 2 vanishes for s @ 0.
Claim 4.1.1.
j E 0 in R1 d
0, 0
– 61 –
4. Wave Equation
Theorem 4.1.4. If ϕ > D Rn and supp ϕ ` 0, then ϕ Pα cαDαδ0 for at most finitely

many cα x 0
Using this theorem, we have that jE Pα cα Dα δ0 and jE is homogeneous of degree d 1.

This implies that jE c0 δ0 . So the last thing you have to check is that c0 x 0 One way to

do this is to compute, which is possible, but difficult). But this fact is not difficult, you just
need the uniqueness of jϕ 0 in R1 d and supp ϕ ` R1 d then ϕ 0 (this can be proved by

Fourier transform or integration by parts).
d1
Remark 4.1.4. χ 2 is a distribution so the composition of it with t2 SxS2 should be
understood by approximations.
Let’s now prove the claim.
d1
Proof. letting χ χ 2
d1 d1
j1tC0 χ 2
2
t SxS
2
1tC0 j χ 2 t
2

2
SxS 1tA0 ∂t 2tx1 t2 SxS2 Q ∂j 2xj x1 t2 SxS2
j
lots of computation. We have to use Euler’s identity:

d1
sχ s
χ s
2
d1
d1
this is because χ is a homogeneosus distribution of degree

1
2 2
Lecture 20 (11/7)
Recall that we were aiming to derive the fundamental solution for j in R1 d . what we
were doing was looking at symmetries of j, that is finding L R1 d R1 d linear such that
s2 t, x t2 SxS2 s2 Lt, Lx. And we were looking at homogenity of j and δ0 . And
looking at the forward condition supp E ` R1 d . Then our guess was the E is of the form:

d1 2 2
Ẽ 1R1d χ 2 t SxS
d1 d1
with χ
2
homogeneous of degree
d1
2
in D R with supp χ

2
` 0, ª.
We computed that jẼ 0 outside the origin, i.e. jE ` 0, 0 and is homogeneous of

degree d 1 which tells us that jẼ cδ0 . Then it is not too difficult to show that c x 0,

but refer to the class notes to compute it.

d1
How does χ 2 look like? Let’s construct a family of distributions, χa for a > R such

that it is a distribution on R, it is homogeneous of degree a and supp χa ` 0, ª.
– 62 –
4. Wave Equation
For a A 1, then there is a obvious candidate:
χ̃a
H x x a
To extend this to a B 1, we will differentiate our distribution. Observe that by computation
(for a A 1) we have (you have to check this):
d a
χ̃
aH xxa 1
aχ̃a
1
dx
Then let’s just define: χ̃ a by using:

1 d a
χ̃a
χ̃
1
a 1 dx
this is good, and will give nice distributions χ̃a which will happen if a

1, 2, 3, . . . . That’s
why we need to be a little smarter, let:
χa
ca H x x a
with ca canceling the possibly zero. If we set:

1
χa
H x x a
Γa 1
for a A 1 where:
x dx
ª
Γa S0 xa e
x
then we wouldn’t have this problem. To see this:
d a 1 Γaa a
χ
aH xxa 1
χ

1
dx Γa 1 Γa 1
Then since Γa 1 aΓa, so the above is just χa 1 . So for a B 1, we define:

d a dk a
χa

χ
1
χ

k
dx dxk
for k > N large enough so that a k A 1.
Example 4.1.4. χ0
H x. And:
k k 1
χ δ0 x

Also we have:
1~2 1 1~2
χ º H x x

π
º 1~2k 1
º H xx1~2 k

this uses the computation that Γ1~2 π. Therefore we have χ
π
– 63 –
4. Wave Equation
Remark 4.1.5. Up to a constant, this is the unique distribution with the three desired
properties.
d1
δ
Remark 4.1.6. j Ẽ
2π 2 0 . This means that:
1 d1
2 2
E 1 d χ
d1 R1
2
t SxS
2π 2
Proposition 4.1.1 (Properties of E ).
1. Finite speed of propagation (weak Huyghen’s principal) This follows from the
fact that:
supp E ` t, x > R

0 B SxS B t
With the representation formula that we just derived:
ϕt, x E ∂t ϕδt 0 ∂ t E ϕδt 0 E j ϕ
with ϕ > C0 R1 d and all these convolutions involve integration on inverted cones
ª

which have finite area at time equal to zero.

2. Strong Huyghens principle when d is odd and C 3. Note that:
d1 2 2
E c1R1d χ 2 t SxS
which is only supported it t, x t SxS which is the boundary of the light cone.
1 1
Remark 4.1.7. When d 1, we get E 1 11 H t2 SxS2 H t x H t x
2 R

2
Remark 4.1.8. IF d 2 then we have:
1 1
E 1R12 1~2
2π t SxS2
2

If you plug this into the representation formula, you will get Poisson formula for the 2
dimensional wave equation.
Remark 4.1.9. For d 3, then we get:
1
E 1 13 δ0 t2 SxS2
2π R

1
the easiest way to compute this is by the approximation method, with ϕs~λ δ0 as λ 0,
λ
so we get:
1
º
4
dSC0
2πt
with C0 t, x t SxS. Plugging this into the representation formula gives you the Kir-

choff formula in R1 3 .

Remark 4.1.10. In Evan’s book, for chapter 2.4, there is an alternative derivation of the
representation formula for j using a very different method (the method of spherical means).
– 64 –
5. Fourier Transform
5 Fourier Transform
Let’s motivate why the Fourier transform takes the form that it does. This will be a brief
two lecture crash course on the Fourier transform.
The Fourier transform can be thought of as a change of basis in the space of functions.
One way to understand f x is a a description of f in the basis that consists of δ0 x y for
all y > Rd . We can do this because we can write:
f x S f y δ0 x y dy
This is not good for understanding ∂x or any Dα . It is reasonable to ask for a basis that di-
agonalizes out basis with respect to our linear transformation of differentiation. The Fourier
transform selects a basis eiξ x ξ>Rd .
An important property of eiξ x is that:
∂j eiξ x iξj eiξ x
The Fourier transform, F f ξ is the change of basis formula such that:
f S aξ eiξ x dξ
with:
F f ξ aξ
We need to compute the transitional matrix m so that:
δ0 x y S my, ξ eix ξ dξ
Note first that:
S my, ξ eix ξ dξ δ0 x y S m0, ξ eiξx y

dξ S m0, ξ e iξy iξ x
e
therefore my, ξ m0, ξ e iξy .
Note that:
0 x j δ0 x S xj m0, ξ eix ξ dξ i S m0, ξ ∂ξj eix ξ dξ i S ∂ξ m0, ξ eix ξ dξ
We then conclude that m0, ξ c. And we have already dedcued δ0 x y R ce iξ y eiξ xdξ.

Now we have:
f x S f y δ0 x y dy cS S f y e iξ y
ydy eiξ x dξ
– 65 –
Therefore we have:
F f ξ d S f y e iξ y
dy
so all we need to do is to determine the constant c.
We want to determine c, where c m0, ξ c. In other words we want to determine:
δ0 x c S eix ξ dξ
Note that
δ0 x δ0 x1 δ0 x2 δ0 xd
And note that the above integral can be written as:
c S ex dξ1 S ex
1ξ d ξd
1
dξ d
So if we just compute c1 for
δ0 x c1 S eixξ dξ
in R, then cd cd1
Claim 5.0.1. c1 2π 1
Proof.
Method 1:
ª 0
lim S e εSξ S ixξ
e dξ lim S e ξ εix
dξ S eξε ix
dξ
ε 0 ε 0 0 ª
1 1
lim
ε 0 ε ix ε ix
1 1 1
lim
i ε 0 x iε x iε
ε
2 lim 2 2 2πδ0
ε 0 x ε
Lecture 21 (11/12)
Recall our discussion on the Fourier transform. We motivated the Fourier transform as a
change of basis from δ0 x y y>Rd to a new basis eiξ x ξ>Rd . And we have:
S f y δ0 x y dy f x S aξ eiξ x dξ
– 66 –
wo we would like to relate f y with aξ . The heart of the matter is to write:
δ0 x c S eix ξ dξ
we had one way to compute c 2π d . Let’s compute this a different way, using an algebraic
method.
Let’s start with the knowledge:
δ0 x c S eiξ x dξ
for d 1. Given any function f , we have:
f 0 S f y δ0 y dy c̄ U f y e iξy

dydξ
Let fÂξ R e iξ y f ydy. So we have

f 0 c̄ S fÂξ dξ
Some properties:
1. xe iξx i∂ξ e ixξ
2. ∂x e iξx iξe ixξ
So if x ∂x f 0, then we get that i∂ξ ξ fÂ 0. This is easy to solve, and we get:
1 2
f x f 0e 2 x
1 2
x
Let’s normalize f 0 1 to get f x e
2 , this gives us:
1 2
fÂξ fÂ0e
2
ξ
but we have:
º
S S
1
fÂ0 f y dy e
2
Sy S 2
dy 2π
º 1
Therefore fÂξ 2πe
2
Sξ S2
. Therefore:
º
f 0c̄ S fÂξ dξ 2πc̄ S e
1
Sξ S2
1
2 dξ 2πc̄
therefore c c̄ 2π 1
Let’s move to a rigorous derivation. Let’s begin with some comments about complex
valued distributions. Let C0 U C be the set of smooth, compactly supported, complex
ª
valued functions in U . We say uj u if and only if Ruj u and Iuj u.
Now given u, v > C0 U C, then we let ù, v e

ª
R uv̄dx be the Hermitian product.
– 67 –
Definition 5.0.1 (Conjugate Linear Complex Distribution). We say a complex valued

distribution, u, is a continuous conjugate linear functional on C0 U C almost everywhere, ª
i.e.
uϕ ψ uϕ uψ , ucϕ c̄uϕ
This generalizes uϕ ù, ϕe when u > C0 U C ª
Remark 5.0.1. Equivalently, u > D U C if and only if u

v iw with v, w distributions.
Definition 5.0.2. Given f > C0 Rd C, let: ª
F f ξ fÂξ S f y e iξ y
dy
Let us introduce F
in the following way
dξ
à, be2π d dξ S ab̄
2π d
then define:
` F f, ae 2π dξ
d
`f, F
ae
We can compute this by:
`F f , ae2πd dξ U f y e iξy
a ξ
dξ
2π d
S f y S aξ eiξ y dξ 2π d dy
Then we will define:
F
af ǎx S aξ eiξ x
dξ
2π d
we will soon prove that F

F 1
Lemma 5.0.1.
1. if f > L1 , then F f is well-defined and Y F f YLª B Yf YL1
2. for all x, y > Rd ,
F f xξ e ix ξ F f ξ

F f ξ η F eiη f ξ

3. if f, ∂j f > L1 , then F ∂j f iξj F f ξ
4. if f and xj f > L1 , then F xj f i∂x iF f ξ
– 68 –
Remark 5.0.2. By the Fourier transform, regularity of f is equivalent to decay of F f

Remark 5.0.3. F
ax 2π d F ax
We would now like to build a space such that it is closed under the Fourier transform
(note that test functions and L1 functions both won’t work).
Definition 5.0.3. Schwartz class on Rd
S R d C ϕ >C ª
R
d
C ¦α, β > Nn , Yxβ ∂ α ϕYLª @ ª
we say that ϕj ϕ in S if and only if:
sup Sxβ Dα ϕj ϕS 0
x>Rd
for all α, β
By our lemma, we have that S is closed under the Fourier transform, and the adjoint of
the Fourier transform. Therefore F S S can act on the dual space of the Schwartz class.

Definition 5.0.4 (Tempered distributions).

S Rd

C u continuous, conjugate-linear functions on S Rd, C
. Where we say u is continuous ϕj S if uϕj ϕϕ
ϕ in
So we get F S

S by `F u, ϕe 2π dξ ù, F ϕe and F S

d

S
(Defined similiarly,
no time).
Remark 5.0.4. All test funcitons are Schwartz functions but the Gaussian is a Schwartz
function, but not a distrubution.
Also S ` D Rd

C, but this is a strict subset.
Remark 5.0.5. Usually, the way to compute F u when u > S , one approximates u by uj > L1
(in the topology of S ). And
F uj S uj y e iξ y
dy
Theorem 5.0.1 (Fourier Inversion, Plancherel).

1. (Fourier Inversion on S ) For f > S Rd, C, F F f FF
f f
2. (Plancherel’s theorem) for f, g > S Rd C, we have:
S F f F
dξ
S f ḡdx g
2π d
In particular:
Yf YL2
2
Y F f Y2 2
L dξ

2π d
and F extends uniquely to F L2 L2 by density.
– 69 –
3. (FI in S ) For u > S Rd

C, FF f
FF u u
Proof. It suffices to show that FF id in S or S . Because the other FF

id follows from
F 2π dF ,

(1) in S , f > S , F f > F . So we have:
FF

f S F f ξ eiξ x
dξ
2π d
lim S
ε 0
F f ξ e εSξ1 SSξd S iξ x
e
dξ
2π d
dξ
lim U f y e iξ y
e εSξ1 S Sξd S iξ x
e dy
ε 0 2π d
d
lim S
ε 0
MS e εSξj Siξj xy
2π
d
dξ f y dy
j 1
S δ0 x1 y 1 δ0 xd y d f y dy f x
Where we used the following fact that we proved last time
lim S e εSξ S iξ x
e dξ δ0 x2π
ε 0
(2) (sketch) FF

id on S 1, This key point is to justify that:
δ0 F 1
There are two steps to show this. First ∂ξj 1 0 implies that xj F 1 0 for all j, With a
little more argument, this reduces to computing the constant cδ0 F 1. This follows by
computing:
,F
1 S xS 2 1 1 SxS2
1 be 2 , δ0 g be 2 g 1~c
c
Remark 5.0.6. The Fourier inversion formula and the Plancheral theorem is all about
finding a way to justify:
dξ
δ0 S eiξ x
2π
Remark 5.0.7. The Fourier transform maps convolutions to products (and vice-versa)
F f g F f F g
And
dξ
a 2πd dξ bη S a ξ b η ξ
2π d
Then:
F 1
a 2πd dξ bη F 1
aF 1
b
– 70 –
Remark 5.0.8. Let L Rd Rd be an invertible linear map. Then:

F f X Lξ det L
1
F f L 1 t
ξ
Lemma 5.0.2. You can compute the Fourier transform of any Gaussian. If A is a symmetric
positive definite matrix, then:
F 1 t
e 2 x Ax 2π
d~2
S det AS
1~2 21 ξ t A1 ξ
e
This is the end of our discussion of the Fourier transform.
Lecture 22 (11/14)
5.1 Applications
The goal today is to give some applications of the Fourier transform to the study of evolu-
tionary PDE’s. The main ideas to get across are:
1. using the Fourier transform to solve homogeneous evolutionary PDE’s. Like PDEs of
the type ∂t f u 0
2. Duhamel’s principle: using the homogeneous solution to solve the inhomogeneous pde.
3. discuss the Fourier transform of the foward fundamental solution. The main example
to study is the heat equation
¢̈
¨ ∂ t ∆u f in R1 d t A 0

¦
ü g in t 0
¤̈
and the Schrodinger equation
¢̈
ï∂t ∆u f in R1 d t A 0

¦
ü g in t 0
¤̈
these and the wave equation are worked out in the notes.
Let’s begin by studying the homogeneous equation:
¢̈
¨ ∂ t ∆u 0 in R1
d
¦
ü g on t 0
¤̈
From now on, the Fourier transform only relies of the space fourier transform, and F t will
mean the time-spacial Fourier transform.
Ât, ξ
Letting u F ut, ξ . So we get the new problem
¢̈
¨∂t Sξ S2 Â
ut, ξ 0
¦ (15)
ü Â0, ξ Â g ξ
¤̈
so solving this first order ODE, we get:
tSξ S2
Ât, ξ
u Â
g ξ e
– 71 –
Proposition 5.1.1 (Solvability for the heat equation).
1. (existence) for g > L2 , then there exists a solution u > Ct 0, ª L2 to (15) such that
Yut, YL2 B Yg YL2
2. If u, u are solutions to (15), u, u > Ct 0, ª, L2 , then u

ua

5.1.1 Duhamel’s Principle
∂ t Atu f
where At PN α α
j 0 aα t, xD with D not containing ∂t . This is a first-order evolutionary
equation.Suppose that we know how to solve the homogeneous problem. In other words,
given a space X of distributions on Rd and there exists the solution operator S t, s g
where g > X and t C s is such that:
∂t AtS t, sg 0

S s, s g g
For fixed s, S t, s g > Ct s, ª, X .
Then we can compute:
∂t At1 s, tS t, s g

ª
We know that: At1 s, ª tu 1s, ª tAt u and we have:
∂t 1s, ª tu δ0 t su 1s, ª t∂t u
So we get:
∂ t At1 s, ª tS t, s g δ0 t sS t, s g δ0 t s g
So the affect of initial data is an instantaneous forcing term.
Now given any F t, x R F s, xδ0t sds, this suggests that we want to define
t
uF t, x S S t, s F s, x1s, ª tds S ª
S t, s F s, xds
This is called Duhamel’s Formula This computation tells us that:
∂t At u F S ∂ t At1 s, ª tS t, s F 9s, ds S F s, xδ0 t sds F t, x
a
use the formula and Plancherel, then apply ODE uniqueness
– 72 –
Usually, X will have a norm defined on it, and we will have YS t, s g YX B C Yg YX . In this
case if F > L1t 0, T , X (that is YF t, YX is absolutely integrable) then uF > Ct 0, t, X .
And moreover, if supp F ` 0 B t B T , then
t t
uF S ª
S t, sF sds S0 S t, sF sds
and UF 0, x 0.
Let’s now apply Duhamell’s principle to:

¢̈
¨∂t Sξ S2 Â
ut, ξ 10, ª tfÂt, ξ
¦
ü Â0, ξ Â g ξ
¤̈
this tells you that:

t t
ÂF t, x
u S ª
S t, s 10, ª fÂsds S0 e
tsSξ S2
fÂsds
this is supposed to give us the solution to the inhomogeneous equation

¢̈
¨∂t Sξ S2 Â
uF 10, ª tfÂ
¦
ü Â 0, ξ 0
¤̈ F
Then we get that:

t
Ât, ξ
u e tSξ S2
Â
g ξ S0 e tsSξ S2
fÂs, ξ ds
Alternatively we could have done the following: we could compute:
F 1
e tSξ S2
Â
g F 1
e tSξ S2
g º
1
d S e
S xy S2
4t g y
4πt
then we can apply Duhamnels principal directly to ∂t ∆ to get:
t
ut, x δ t, 0 g S S t, s f sds
0
which solves the inhomogeneous problem

¢̈
¨∂t ∆u f
¦
ü g
¤̈
Theorem 5.1.1.
1. (existence) for g > L2 and f > L1 0, T L2 , there exists a solution to the inhomoge-
neous equation such that Yut, YL2 B Yg YL2 R0 Yf s, YL2 a
t
a
notation: u > Lpt I; X iff t ( Y ut, YX > LPt I
– 73 –
2. (uniqueness) If u and u are in Ct 0, T ; L2 and the inhomogeneous equation then

u u
The crux of Duhamels’ formula is:
∂ t At1 s, ª tS t, s g δ0 t s g
We can get the forward fundamental solution via the Fourier transform. We have:
∂t ∆E δ0 tδ0 x

supp E ` t C 0
This formula tells us that we must have that:
E 10, ª tS t, 0 δ0 x
Recall that F δ0 1. This means that:
Â S Æ
u t, 0 δ0
should solve
¢̈
¨∂t Sξ S2 Â
ut, ξ 0
¦
ü Â0, ξ 1
¤̈
but we know that e tSξ S2 solves this, therefore we get that:
tF 1 tSξ S2 1 S xS2

E 10, e 10, t º e

4t
ª ª
d
rπt
Remark 5.1.1. E is smooth outside of the space time origin

Remark 5.1.2. E is analytic outside t

0
Remark 5.1.3. For any u > S 1 R1 d with supp u ` t C L , then u E is well-defined.
Now let’s discuss the space time Fourier transform of the forward fundamental solution.
Let’s start with:
∂t ∆E δ0
let’s take the Fourier transform to get:

2
iτS ξ S t,x F E 1
provided that iτ Sξ S2 x 0, we have
F t,x E
iτ
1
Sξ S2
– 74 –
6. Energy Method and Sobolev Spaces
This isn’t quite enough, we also need to use the support property supp E ` t C 0.
We can approximate E by e εt E as ε 0 . Now we can try to compute:

εt εt εt εt
∂t ∆e E e ∂ t ∆E εe E δ0 t, x εe E

putting these together, we get:

εt
∂t ∆ ε e E δ0 t, x

then take the Fourier transform of this to get:
iτ Sξ S
2
εFt,x e εt
Eτ 1
for all ε A 0, but note that:
F t,x e εtE

1
iτ YxiS2 ε
so now:
Ft,x E lim Ft,x e

ε 0
εt
E
ε 0
lim
1
iτ Sξ S2 ε
Remark 5.1.4. Analyticity in the lower (or upper) half-plane of the Fourier transform
cooresponds to the original distribution in t C 0 or t B 0
Remark 5.1.5.
Lemma 5.1.1. Let a > H a > C Ia C 0, then:
lim F t 1
1

i10, teiat
ε 0
ª
τ iε a
Assuming this lemma, we see that
E Ft,x1 Ft,xE

lim Ft,x1
ε 0

1
it iε Sξ S2

F 1 2
10,ª teSξ S t
Lecture 23 (11/19)
6 Energy Method and Sobolev Spaces

Now we will move on to discussion on the energy method and Sobolev spaces. This will
follow Evans chapter 5.
– 75 –
6.1 Energy Method

This is the third general way to trying to understand 2nd order PDEs. This is glorified
integration by parts.
Example 6.1.1 (Laplace equation). We will focus on proving a-priori estimates. By
which we will assume there is a regular solution to the PDE, then we will prove inequalities
about norms of the solution with respect to the data. Let’s look at the Dirichlet problem
¢̈
¨∆u f in U
¦ (16)
ü g on ∂Ω
¤̈
Proposition 6.1.1. Suppose that U is a C 1 bounded domain and u > C 2 Ū , f > C 0 Ū ,
g > C 0 ∂U . Then the solution, u, to (16) is unique.
Proof. We have already seen a proof, but here we present a different proof.
Let u1 , u2 be two solutions to (16). Then if we let v u2 u1 > C 2 Ū , then this solves
the homogeneous equation:
¢̈
¨∆v 0 in U
¦
¨v 0 on ∂U
¤̈
The key idea is: multiply the equation by v, integrate over U , then integrate by parts.
0 SU ∆vvdx SU v
© © vdx S
∂U
ν © vvdS∂U
therefore R S©v S2 dx 0, this tells us that © v 0, therefore v 0 in U by the boundary

condition.
So the energy method is a very loose term but it is:
Multiply the equation by a suitable Xu (suitable operator), then we integrate by parts.
Example 6.1.2 (Heat Equation). Consider the heat equation:
¢̈
¨∂t u ∆u f in R1
d
¦
ü g on t 0
¤̈
Proposition 6.1.2. If u > Ct 0, T , L2 , © u > L2t,x 0, t Rd and f > L1t 0, T , L2 and
g > L2 . Then u is unique and
YutYL2 Y©uYL2 0,T ,L2 B Yu0YL2 Yf YL1 0,T L2
Proof. In this case we have:

t t
S0 S f udxdt
S0 S ∂t ∆uut , xdxdt

t 1 2 t
S0 S 2
∂t u dxdt
S0 S ∆udxdt
t 1 2 t
S0 S 2
∂t u dxdt
S0 S © u © udxdt Boundary term
– 76 –
the Boundary term is zero by taking limits and using the assumptions on u and ©u. But
then we get:
1 1 t
2S 2S S0 S
2 2 2
u t, x dx u 0, x dx S©uS dxdt

Then we get the inequality via rearrangement, Cauchy Schwartz, and Minkowski’s inequality.
Another main thing we do in the energy method is: to apply operators Y to u, then look at
P Y u , then apply the process as above.
Notice that for any ∂j , if ∂t ∆u f , then ∂t ∆∂j u ∂j f , this is because ∂j commutes
with the heat operator. So if we apply the same method to this resulting equation, then we
can get estimates:
sup YDα utYL2 Y©Dα uYL2t,x 0,T Rd B sup YDα g YL2 YDα f YL1 0,T ;L2
SαSBk S α S Bk
Remark 6.1.1. For the wave equation, j u f , a “good” multiplier is ∂t u. This gives us
the physical conservation of energy.a
6.2 Sobolev Spaces

From the energy method, you expect to be able to control something like:
α
YD uYL2
One natural question is: can we say other things about the solution from the control of these
kinds of norms?
We know that if f > Cc1 R, then by the fundamental theorem of calculus
x
f x S ª
f x dx B Yf

YL1
Now if x > I (a compact interval) and supp f ` I, then we have:
sup Sf S B Yf
YL1 I B Yf
YL2 I SI S
1~2
I
by Holder’s inequality. Now it turns out we can generalize this to multiple dimensions, which
are known as Sobolev inequalities.
Definition 6.2.1 (Sobolev Space). Given U a domain in Rd , k > ZC0 , 1 B p B ª, then we

let:
W k,p U u > Lp U Dα u > Lp U ¦ SαS B k

a
this multiplier makes sense due to Lagrangian mechanics
– 77 –
where Dα are taken in the sense of distributions.
We associate a norm:
YuYW k,p U Q Y D α u YL p U
S α S Bk
So Sobolev inequalities will look like
YD
α
uYLq B YuYW k,p
Remark 6.2.1. We want a “function space” framework for converting a-prioi estimates to
existence results.
In linear algebra, given a n m matrix L and we are interested in solving Lu f , then

the existence of u such that Lu f is equivalent to `f, ϕe 0 for all ϕ such that L ϕ 0
(which is equivalent to ImL ker L Ù ). It turns out the Sobolev spaces are the ideal

space to do the same thing.
Let’s introduce a little more notation. If ϕ > C0 U then ϕ > W k,p U for all k, p. In
ª
other words C0 U ` W k,p U , and we can consider the completion of C0 U in W k,p U

ª ª
with respect to Y YW k,p U . The resulting space will be denoted W0k,p U (the completion
of test function in the Sobolev space.
In general, this is a strictly smaller subset, and you should think of these function as
elements in the Sobolev space whose boundary value on ∂U is zero.
The case p 2 is of importance (these are the bounds we get from the energy method).
We let H k U W k,2 U . And we often write H0k U W0k,2 U .
Proposition 6.2.1 (Completeness of Sobolev Space, Norm Equivalence To Fourier

Transform).
1. For all k, p, W k,p U is complete under Y YW k,p U . In other words W k,p U is a Banach
space.
2. There is a characterization of H k Rd in terms of the Fourier transform, that is:
YuYH k Rd Y1 Sξ Sk u

Âξ YL2
ξ
where A B if and only there exists C A 0 such that A B CB and B B CA
Now let’s discuss some basic properties of W k,p U .
1. approximation theorem
2. extension theorems
3. trace theorem
– 78 –
(
Lemma 6.2.1. Given y > Rd , let τy ux ux y viewed as a linear operator. Then τy
is continuous on Lp as a function of y (implicitly, this means u > Lp ). In other words:
lim Yτy u uYLp 0

y 0
Proof. See Evans appendix.

Corollary 6.2.1. For all u > L, , given a mollifer ϕ, then:
lim Yϕε u uYLp 0

ε 0
Proof. We have:
1
\ S εd
ϕy ~εux y dy ux\
Lp
Y S ϕz ux εz dz S ϕz uxdz YLp
Y S ϕz ux εz uxdz YLp
B S ϕz Yτεz u uYLp dz
take ε 0 and apply the dominated convergence theorem.

Theorem 6.2.1 (Approximation). By the same token if u > W k,p Rd , then ϕε u u in
W k,p Rd . And we know that ϕε u > C Rd ª
Proposition 6.2.2. If u > W k,p U , then there exists uε > C ª

U such that uε u in
W k,p U
Proof. Consider a sequence of open sets Vn such that Vn is bounded, V̄n is compact, Vn ` Vn 1
and n 1 Vn U . Every open covering has a smooth partition of unity χn subordinate to Vn .

ª
By this we mean that χn > C Rd such that supp χn ` Vn and Pn 1 χn x
ª ª
1 and for
all x > U , only finitely many terms in the sum are nonzero. And 0 B χn B 1
Then we can split u Pn 1 χn u which can be defined on Rd . Then the idea is to approxi-
ª
mate each χn u > W k,p Rd by extending it by 0 outside Vn . Then you take an approximation:
vn ϕεn χn u
such that Yvn χn uY B δn ε. Then we let vε Pn vn where δn is chosen so that P δn 1, then

ª ª
Yvn uYW k,p Q Yvn χn uYW k,p Rd B Q δn ε

n 1 n 1
, which is bounded by ε.
Proposition 6.2.3 (Approximation of elements in C Ū ). If U is a C 1 -domain and ª
u > W k,p U , then there exists a sequence uε > C Ū such that uε u in W k,p U .
ª
Proof. (sketch) It was a picture.
– 79 –
Lecture 24 (11/21)
We will continue the discussion on Sobolev spaces today. Recall we say that u > W k,p U
if:
YuYW k,p Q Y D α u YL p @ª
SαSBk
We then approximated elements of the Sobolev space that were better behaved. If u >
W k,p Rd , with ϕ > C0 U , R ϕ 1, ϕε ε d ϕ ~ε, then:
ª
ϕε u u
in W k,p Rd . Also if u > W k,p U then there exists a sequence uj > C ª
U such that uj u
in W k,p U .
We also have that if u > W k,p U with U a C 1 domain, then there is a sequence uj > C ª
Ū
such that uj u in W k,p U .
Proposition 6.2.4. If u > W k,p Rd , k is a non-negative integer, 1 B p @ ª, then χx~Ru

u in W k,p Rd as R ª (for χ a smooth function that is 1 on B 0, 1 and 0 outside B 0, 2)
Proof. For k 1 (the others are similar). We need to show that:
YD χ ~Ru uYLp Yχ ~Ru uYLp 0
as R ª . To show this we right the first term as:
χ ~RDu Du Dχ ~Ru
which can both be shown to go to zero by simple estimates.
Corollary 6.2.2. C0 Rd is a dense subset of W k,pRd

ª
(This can’t be right...we defined W0k,p to be the completion of test functions, but by this
corollary, this means that W0k,p W k,p )
Theorem 6.2.2 (Extension). Suppose U is a bounded domain with ∂U a C k boundary.

Let V be an open set such that V a Ū . Then there exists a linear operator E W k,p U
W k,p Rd such that:
1. E uU
U
u
2. supp E u ` V for all u > W k,p U
3. Y E uYW k,p Rd B C YuYW k,p U where C depends on k, p, V, U .
– 80 –
Proof. (sketch)
Step 0: it suffices to consider u > C ª
Ū (by approximation methods).
step 1: Reduce the result to the following: there exists an extension operator E on
U B 0, 1 9 Rd and V B 0, 2, defined for u > C Ū and supp u ` B 0, 1 9 xd C 0 and
ª
YE uYW k,p Rd B C YuYW k,p U .

Here we just argue by picture, I did not draw the picture...
Step 2: proof of the claim in part 1. The key is to use reflection. Let x
x1 , . . . , xd1
then define:
¢̈
d üx, xd C 0
ũx , x
¦
üx , xd ,
¤̈
xd @ 0
This is a continuous extension. However, can we extend so that Dũx , 0 Dux , 0?
Can we do this for all derivatives? The key part is to show that the normal derivatives agree:
∂xl d ũx , 0

∂xl d ũx , 0

For l 0, . . . , k. So let’s introduce α0 , . . . , αk , β0 , . . . , βk . Then define:

¢̈
¨ u x , x d , xd C 0
ũx , xd

¦ k
P
¨ l 0 αl ux , βl xd ,
¤̈
xd @ 0
with αi > R and βi A 0. So now if we compute, for xd @ 0:

k k
∂xl d ũx , xd

∂xl d Q α m u x ,
β m xd Q αm βm l ∂xl d ux , βm xd

m 0 m 0
So now if we want the two limits as xd 0 (from both sides), we get the following linear
system of equations for β and α:
1 α0 αk
1 β0 α0 β0 αk

k
1 β0 α0 βk k αk
Which is just the matrix equation:

0 βk 0 α0
1 β0

1 β0 k βk k αk
which is just the Vandermond matrix, which has determinant:
M βi βj
0Bi@j Bk
– 81 –
so if we choose βi ’s to be pairwise distinct, then there exists α0 , . . . , αk so that this matrix
holds.
So now define:
E
u χũ
with χ equal 1 on the support of U and 0 outside a B 0, 1 9 xd C 1~2. Then you check
the estimate.
Remark 6.2.2 (Stein). There exists a universal way to extend Sobolev spaces Euniv for all
W k,p (k C 0, 1 B p @ ª which only requires U to be C 1 .
6.3 Traces
Here we will restrict u to ∂U
Theorem 6.3.1. Let k C 1 and 1 B p @ ª, U a C 1 domain
1. u > C ª
Ū and define:
tru uS∂U
Then this extends uniquely as a bounded map from W k,p U to Lp ∂U such that:
Ytr uYLp ∂U B C YuYW 1,p U
2. If u > W k,p U and tru 0, then u > W0k,p U and vice-versa.

Proof. See Evans section 5.5
Remark 6.3.1. In fact, the image of tr leads to the notion of fractional regularity Besov
spaces.
It turns out trH k U Hk 1~2 ∂U , with U is C k (this can be defined with the Fourier
transform)
s
YuYH s Rd Âξ YL2
Y1 Sξ S u
ξ
6.4 Sobolev Inequalities

This is all expressions of the form:
YuYW p,q B C YuYW k,p U

or the term on the left could be one of many types of norms.
A key element in this theory is the following inequality
– 82 –
Theorem 6.4.1 (Gagliardo – Nirenberk - Sobolev Inequality). Let u > C0 Rd , then: ª
YuYLd ~ d1 Rd B cYDuYL1 Rd

We can remember d~d 1 through dimensional analysis (or scaling analysis). If we have
ux, define uλ x uxλ 1 . Then we have:

YDuλ YL1 S SDuλ Sdx λd 1

S SDuy Sdy
We also have that Yuλ YLp λd~p YuYLp In order for the inequality to hold for all u and all λ,
d
we would like p .
d1
Proof. Given u > C0 Rd , we know that we can write:
ª
xi
ux S ª
∂ i u x 1 , . . . , yi
®
, . . . , xd dy i
ith component
This tells us that:
sup SuxS B S S∂i ux

1
, . . . , xi , . . . , xd Sdy i
xi > R ª
Lemma 6.4.1 (Looms-Whitney-Inequality). Let fi be functions on Rd such that:

Â ,...,x C 0
f i x 1 , . . . , x i d
Âi indicates that we are fixing that coordinate (so the function doesn’t depend on that
where x
coordinate) then:
d
S f1 fd dx B M Yfi YLd1 Rd1
i 1
Proof.
S f1 fd dx1 f1 S f2 fd dx1 B f1 Yf2 YLd11 Yfd YLd11

x x
then we integrate this thing and apply Holders inequality again and again and everything
will end up working out.
With this we are almost done. Define:
fi sup Sux1 , . . . , xd S B S SDux1 , . . . , y i , . . . , xd Sdy i
xi > R R
Let us apply the Looms-Whitney inequality to:
S d~d1
SuS dx1 dxd B S f1
1~d1
fd
1~d1
B
®
M 1
Y fi
~d1
YLd1
i
LW
B MS Sfi Sdx 1

i d 1~d
Â dx
dx 1
B MS SDuSdx1 dxd 1~d 1
i i
S SDuSdx
1
dxd d~d 1
– 83 –
Remark 6.4.1. Isopermetric inequality Give a C 1 bounded domain, then:
SV olU S
d1~d
B C SArea∂U S
Lecture 25 (12/3)
Recall our discussion on Sobolev inequalities. Last time we proved the Gaglaudo-Nirenberg
Sobolev inequality for u > C0 Rd 9 W 1,1 Rd , we got that:
ª
YuYLd ~ d1 Rd B C YDuYL1 Rd

We will now upgrade this to more general cases when u > C0 Rd 9 W 1,p Rd ª
Theorem 6.4.2. For u > C0 Rd 9 W 1,p Rd , 1 B p @ d, we have that:

ª
YuYp B C YDuYLp Rd (17)

dp
where after scaling analysis, it is easy to see that p
dp
d
Proof. Let v SuSγ so that γ p , then apply the W 1,1 version of GN S to v:

d1
d1
S B c S SDSuSγ Sdx B c S SuSγ

d
B cYuγ
d
1 1
Sv S d1 dx SDuSdx YLp YDuYLp

claim: p γ 1

p , this has to be true because we respected scaling.

We have only discussed nice functions, but we can extend this to non compactly supported
ones:
dp
Theorem 6.4.3 (Sobolev inequalities for W 1,p U ). Let 1 B p @ d, p
dp
1. if u > W01,p U , then (17) holds
2. If U is a bounded C 1 domain, then for any u > W 1,p U , (17) holds on the domain U .
When p d, p
ª , but :
YuYLª Rd B YDuYLd Rd

is only true when d 1. There is a nice counterexample in Evans. Let U B 0, 1 and
u log log1 1~SxS.
It turns out that often in analysis, the following substitute for L ª

is useful. The space
we have in mind is called BM O. We say u > BM O if
sup
B x,r
S~ B x,r
Su aB x,r

uSdx @ ª
– 84 –
this is known as bounded mean oscillation.
If p A d, in this case we have:
YuYLª r B cYu YW p
where r is the modulus of continuity.
Definition 6.4.1 (Holder space). for u > C 0 K with K a closed subset of Rd ,
Sux uy
uC 0,α sup
x,y >K Sx y Sα
where 0 @ α @ 1. Then define:
C 0,α K u > C 0 K uC 0,ª @ ª
which we equip with the norm:
YuYC 0,ª K YuYC 0 K uC 0,ª K

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
sup norm
Remark 6.4.2. C 0,α K is a complete space.

Remark 6.4.3. If u > C 0,α K , then u > C 0,α K with 0 @ α @ α @ 1
Remark 6.4.4. If uC 0,α @ ª for α A 1, then u is constant.
Theorem 6.4.4 (Morey’s inequality). For u > C ª

Rd 9 W 1,p Rd (for p A d)
YuYC 0,α Rd B C YuYW 1,p Rd
d
with α 1
p
Proof. The key is the following proposition:
Proposition 6.4.1. for u > C ª
Rd , with x > Rd and r A 0, then we have:
1 SDuy S
S
SB x, r S B x,r
Suy uxSdy B C S
B x,r Sx y Sd 1
dy
with C only depending on d

Proof. The idea is to compute:
S∂B x,r

Suy uxSdy rd 1
S∂B 0,1
Sux rω uxSdω
– 85 –
then note that:

r r
ux rω ux S0 ∂s ux sω ds S0 ω © ux sω ds
therefore:
r
rd 1
S∂B 0,1

S Sdω rd 1
S∂B 0,1 S S0

ω ux sω sd 1 dsSdω
computation follows, I couldn’t really read the board, I’ll try to fill this in later.
Now let’s prove the main thing. First for all x > Rd , we have:
SuxS Sux aB x,1

uy dy S a
B 0,1
SuSdy Ba
B ,x1
Sux uy S a Suy Sdy
both of these can be trivially bounded (see notes). The hard part is bounded the semi-norm.
Note that we have:
Sux uy S B Sux a uz dz S S a uz dz uy S

B x,r9B y,r B x,r9B y,r
Ba Sux uz Sdz aB x,r Suz uy Sdz

B x,r9B y,r 9B y,r
This can be controlled by our previous results, scaling estimates on regions, and Holder’s
inequality.
Definition 6.4.2 (Version). We say u is a version of u if u
u almost everywhere.

Theorem 6.4.5 (Sobolev Inequality for u > W 1,p U , p A d). given p A d and α 1 d~p
1. If u > W01,p U , then there exists a version u > C 0,α Ū and Yu
YC 0,α Ū B C YuYW 1,p U
2. Assume also that U is a bounded C 1 domain. For any u > W 1,p U there exists a
version u > C 0, Ū and the above inequality holds.
ª
Theorem 6.4.6 (General Sobolev inequality for W k,p U ). Let k be any non-negative
integer and 1 B p B ª and assue either U is a domain and u > W0k,p U or U is a bounded
C k domain and u > W k,p U .
1. let l be a nonnegative integer, 0 B l B k, 1 B q @ ª, d~q l C d~p k, then u > W l,q U
and YuYW l,q U B C YuYW k,p U
2. let l be a nonnegative integer 0 B l B k, 0 @ α @ 1 if l α C d~p k, then there exists a

version u > C l,α Ū and YuYC l,α Ū B C YuYW k,p U

Lecture 26 (12/5)
Today we will wrap up our discussion of Sobolev spaces. We will study compactness
properties with the main theorem the Relich - Konarchov theorem. We will also look at
Poincare’s inequality. And lastly we will look at the dual space of Sobolev spaces, which will
naturally lead to the notation of negative regularity Sobolev spaces.
– 86 –
6.5 Compactness
We want to look at ui ` W 1,p U given Yui YW 1,p U B C, then this turns out to be compact
in an appropriate larger space.
Definition 6.5.1 (Compact Embedding of Banach Spaces). Let’s say we have X, Y
YX and Y, Y YY two normed spaces with Y ` X. We say that Y, Y YY embeddes compactly
into X, Y YX (and say Y bb X if for all bounded sequences in Y has a convergent sub-
sequence in X.
Theorem 6.5.1 (Relich–Konoluchov). Let U be a C 1 bounded domain, 1 B p @ ª and

1 B q @ p , then W 1,p U bb Lq U . Recall we have:

YuYLp U B C YuYW 1,p U

and W 1,p U b Lp U
Proof. The key ingredients to this are:

Theorem 6.5.2 (Arezela-Ascoli). Let K be a compact metric space with un ` C K ,
then given
1. (pointwise boundedness) for all x > K, we have un x is bounded
2. (equicontinuity) for all ε A 0 there exists δ0 such that Sun x un y S @ ε for all n and
Sx y S @ δ.
then there exists a subsequence uni that is convergent in the uniform topology.
Another key ingredient is a property of mollification called accelerated convergence.
Proposition 6.5.1. Let ϕ > C0 Rd with R ϕ 1 and define ϕε x ε d ϕxε 1 , then
ª
ϕe u u as ε 0 in Lp . It turns out that if you assume that u > W 1,p U , then the rate of
convergence in ϕε u u can be quantified.
Here is a more general proposition
Proposition 6.5.2. Let k be a positive integer, 1 B p @ ª, let ϕ > C0 Rd such that ª
Rϕ 1
and R xα ϕdx 0 for all 1 B SαS B k 1. Then if u > W k,p Rd then:
Yϕε u uYLp B Cεk Q Y D α u YL p
Sα S k
Proof. When k 1:
ε
ux ϕε ux S ϕz ux ux εz dz S ϕ z S
0
∂z ux tz dz
ε
S ϕz S
0
z Dux tz dz
– 87 –
then by the Minkowski inequality, we have:

ε
Yux ϕε uxYLp B S S0 Yϕz z Dux tz YLpx dtdz
ε
BS S0 Sϕz z SYDux tz YLp dtdz
B CεYDuYLp
Now let’s prove the RK theorem.
Step 1: It suffices to prove the case when q p. We want to show that if Yun YW 1,p U B C
then there exists a subsequence uni so that Yuni unj YLp 0 as i, j ª .
Indeed, if p @ q, then we have that all the other norms are bounded. If p @ q @ p then
use Holder’s inequality which states that:
Yv YLq B Yv YθLp Yv Y1Lrθ
with q 1 p 1θ r
1 1 θ
Step 2: we would like to prove that there exists a subsequnce uni such that for all ε A 0
there exists I with Yuni uni YLp U @ ε if i, i C I.
Fix a bounded subset of Rd (call it V ) and let Ū ` V . Let’s extend un to elements in

with supp un ` V and
W 1,p Rd
Y E u n YW 1,p Rd B C Yun YW 1,p U B C
Now let un E un . Choose δ and let vn un ϕδ and supp vn ` Ṽ another bounded set with
V̄ ` Ṽ , let δbe such that:
Yun vn YLp B ε~3
(where we use the proposition). Then note that vn is smooth, then we have:
vn x S ϕδ x y vn y dy
then by Holder’s inequality, we have that:
Yvn YLª B Cδ
and also
YDvn YLª B Cδ
– 88 –
this means that vn has a convergent subsequence by Arzela-Ascoli. So there exists vni so
that Yvni vni YLp @ ε~3, for i, i C I.

So now we have:
Yuni uni YLp B Yuni vni YLp Yvni vni SLp Yuni vni YLp B ε
now we must preform a diagonalization argument.
Let’s now consider one nice application of this theorem. This is a key way to estimate u
by Du.
Proposition 6.5.3 (Poincare’s Inequality). Let U be a C 1 bounded domain, 1 B p @ ª,

then:
\u aU u\L P U
B C YDuYLp U
with C only depending on U and p.
Proof. Assume that this fails. Then for all n C 1 there exists un ~ 0 so that:
Yu n aU u n Y LP U C nYDun YLP U
Now let:
un Ù un
vn
Yun Ù un YLp
1
Therefore: Yvn YLp 1 and ` vn 0, and
n
Yvn YLP U C YDvn YLp so that YDvn YLp U 0 as
N ª.
We know that vn is bounded in W 1,p . THerefore there exists vni with vni v in Lp U .
In particular:
Yv YLp lim Yvni YLp 1

i ª
on the other hand:
aU v lim a vni
i ª U
0
and finally, YDvn YLp 0 this implies that Dv 0 in the sense of distributions, therefore v is
a constant and we get a contradiction.
– 89 –
6.6 Duality
Definition 6.6.1 (Negative Regularity Sobolev Spaces). Let k be a non-negative inte-
ger and 1 B p @ ª, let:
¢̈ £̈
> D U ¦ SαS B k, § gα > L U , u
¨ ¨
W k,p
U ¦u
¨
p
QD α
gα §
¨
¤̈ SαSBk ¥̈
with norm:
YuYW k,p U inf

g
Q YgαYpL 1 pp
~
P α Bk Dα gα
α
u SαSBk
S S
Theorem 6.6.1. For all domains U , k a positive integer, 1 @ p @ ª:

k,p k,p
W0 U W U

– 90 –
Appendices
A Midterm Review
A.1 First order scalar characteristics
We are trying to solve the first order scalar PDE:
¢̈
¨F x, u, Du 0 U R Rd
¦ (18)
ü g Γ ∂U
¤̈
With x > U ` Rd , u Rd R. Then if we assume there exists a solution, then along any path
x I Rd , we have z s uxs with:
¢̈ẋj ∂pj F
¨
¨
¨
¦ṗi ∂z F pi ∂xi F
¨
¨
¨ż
¤̈ Pi pi∂pi F
Where in F we replaced u with z and ∂i with pi .
Corollary A.1.1. If F P bj ∂ j u c is linear or quasilinear, then:
¢̈ j
¨ẋ bj
¦
¨ż c
¤̈
A.1.1 Existence
Given our PDE (18), we would like to know when this method works, whether a solution
exists. If our boundary is flat (Γ ` xd 0) then for any characteristic path, we know it
must satisfy compatability conditions:
¢̈x0 x > Γ
¨
¨ 0
¨
¨
¨
¨z 0 g x0
¦
¨
¨ pi 0 ∂i g x0 i 1, . . . , d 1
¨
¨
¨
¨F x0, z 0, p0
¤̈
To apply our method of characteristics, we require the information of pd 0 which we can
get if ∂pd F x0 , z0 , p0 x 0 (this is known as the noncharacteristic condition).
Now we would like to show that u has a solution near x0 . To do this let’s find a whole
bunch of characteristics where the first input indicates where the path starts, and the second
indicates the time on the path:
xy, s
z y, s
py, s
– 91 –
A. Midterm Review
with y y 1 , y 2 , . . . , y d 1 , 0. Then in order to find the solution ux we must figure out the

characteristic that hits x. That is we must determine y and s such that xy, s xa . It
turns out that under the noncharacterisitic assumption, this map is locally invertible, and
(
we denote the inverse: x y x, sx, which gives rise to the main theorem:
Theorem A.1.1. Given our PDE 18 and a point x0 > Γ such that ∂pd F x 0 at x0 , then there
exists a neighborhood of x0 such that ux z y x, sx solves the PDE for all x in this
neighborhood.
Remark A.1.1. The nonchraracteristic assumption of x0 , z0 is equivalent to determining
all derivatives of u (in every direction) at x0 .
A.1.2 General Boundary

For a smooth boundary Γ we can do the same thing, but the noncharacterisitc condition
becomes:
Q ν∂U
j
∂p F x0 , z0 , p0 x 0
j
j
A.1.3 Uniqueness
Given two solutions to such a PDE whose sum of their normal derivatives agree, then they
are the same.
A.2 Noncharacteristic Condition for Kth order systems

Our new quasilinear K-th order PDE is of the form:
N
F Q Q bαBA x, u, . . . , Dk 1uDαuB
cA x, u, . . . , Dk 1 u

0 (19)
S αS kB 1
k 1
RN , U ` Rd , bα U RN Rd N
2
with u U RN . So bα evaluated at each point is a
N N matrix. The A, B component of this matrix is bα A
th
B . Therefore a very compact
(but uninformative) way of writing F would be:
F Q bα D α u c
Sα S k
But just note that F is a set of N equations. Now the Cauchy data (initial conditions) for
this PDE would be to prescribe normal derivatives. If our boundary is flat, Γ ` xd 0,
then our Cauchy data is:
k 1
u, ∂d u, . . . , ∂d u g0 , g1 , . . . , gk1
Then define the noncharacteristic condition for the Cauchy data for a point x0 > Γ as when
the matrix:
b0,0,...,0,k x0 , ux0 , . . . , Dk 1 ux0
is invertible. If this is the case then we can determine all derivatives of u at x0 .

a
lol, this notation could use work, I literally just figured this out while writing it
– 92 –
A. Midterm Review
A.2.1 General Domains

For a general domain, the noncharacteristic condition requires that:
Q bα ν α
SαS k
is invertible, were ν is the outer unit normal to the boundary.
A.2.2 Power Series

Theorem A.2.1. Given the PDE (19) such that all functions and boundary are real analytic
and x > Γ is noncharacteristic, then there exists a unique, local, real-analytic solution u with:
1
ux Q α! Dαux0x x 0 α
near x0
A.3 Distributions
Definition A.3.1 (Test Function). A test function on a set U ` Rd is a smooth compactly
supported function on U . The space of test functions on U is denoted C0 R ª
Definition A.3.2 (Convergence of Test Functions). We say fn f converges in the

space of test functions (with fn , f > C0 U ) if there exists a compact set containing the
ª
support of alla the functions on which all derivatives uniformly converge.

Definition A.3.3 (Distribution). A distribution on a set U ` Rd is a linear functionalb
from C0 U to R which is continuous with respect to convergence of test functions.
ª
Theorem A.3.1 (Boundedness). A linear functional u C0 U R is a distribution if ª
and only if for all compact sets K ` U , there exists a N and C C K, N such that for all
ϕ > C0 U with supp ϕ ` K:
ª
S ù, ϕe S BC Q sup SDα ϕxS

x> K
SαSBN
Definition A.3.4 (Order of a Distribution). The largest such N in the above theorem
is called the order of a distribution.
Theorem A.3.2. If u is a distribution that agrees with 1~x on 0, ª then the order of u is
at least 1.
Definition A.3.5 (Convolution). If f > L1loc and ϕ > C0 Rd then: ª
f ϕx SR d
f y ϕx y dy
we can also define the starred convolution as:
f
ϕ x SR d
ϕy f y xdy
a
but finitely many of course
b
linear map from a vector space to its base field
– 93 –
A. Midterm Review
Theorem A.3.3. If f > L1loc and ϕ > C0 Rd then f ª

ϕ>C ª
Theorem A.3.4. If f > C Rd and ϕ > C0 Rd then suppf

ª
ϕ ` supp f supp ϕ
Definition A.3.6 (Mollifer). A mollifier is a function ϕ > C0 Rd with ª

Rϕ 1
Theorem A.3.5 (Approximation). If ϕ is a mollifier, let ϕδ δ d ϕxδ 1 . Then if
f > C R then f ϕδ converges uniformly to f (an up to and including k derivatives

k d
converge uniformly).
Theorem A.3.6. If f > L1loc then f defines a distribution.
A.3.1 Operations
We may do the following operations with u a distribution, f > C ª
U , ϕ, v > C 0 R d
ª
1. `f u, ϕe ù, f ϕe
2. `Dα u, ϕe 1SαS ù, D α ϕe
3. ù v, ϕe ù, v

ϕe
Theorem A.3.7. u v > C ª
Theorem A.3.8. δ0 u u
Theorem A.3.9. If U is an open, connected, bounded subset of Rd with a C 1 boundary and

1U is the characteristic function (a distribution) then:
∂j 1U ν j dS∂U
with ν the unit outer normal of ∂U and dS∂U the surface measure.
A.3.2 Convergence of Distributions

Definition A.3.7. un (distributions) converge in distribution to a distribution u if for all
ϕ > C0 Rd then ùn , ϕe ù, ϕe
ª
Theorem A.3.10. If un u as distributions, then Dα un u as distributions for all α
Theorem A.3.11 (Sequential Compactness). If un are distributions and for all ϕ >
C0 U , ùn, ϕe converges, then u defined as ù, ϕe limn ùn, ϕe is a distribution.
ª
Theorem A.3.12. If ϕ is a moliffier, then ϕδ converges in distribution to δ0
Theorem A.3.13. If ϕ is a mollifier and u a distribution then ϕδ u are smooth functions

that converge in distribution to u.
– 94 –
A. Midterm Review
A.3.3 Operations between distributions

Definition A.3.8 (Singular Support). The set of points on which a distribution fails to
a be a smooth function.
Theorem A.3.14. If two distributions have disjoint singular support, then their product is
defined.
Theorem A.3.15. If u or v have compact support, then u v v u is well-defined

(ù v, ϕe ù, v ϕe) and supp u v ` supp u supp v.

A.4 Fundamental Solutions

For this section consider the linear scalar partial differential operator:
P u Q aα x D α u
SαSBk
with aα real valued smooth.
Definition A.4.1 (Fundamental Solution). Given y > Rd , the fundamental solution Ey

is such that P Ey x δy x δ0 x y
Then formally, a solution to this PDE is ux R vyE y xdy.

Definition A.4.2. The formal adjoint to the above linear operator is:
P v Q aαxDαaαu

SαSBk
which we expect to satisfy `P u, ve ù, P ve
Remark A.4.1. Finding fundamental solutions Ey x to the adjoint of

P gives us unique-
ness, this is because if u solves P u v, then:
ux ù, δx e ù, P E xe `P u, E xe

`v, E e
x
Which expresses u in terms of v.
Theorem A.4.1. If P is constant coefficient (ie all aα are constant) and P E0 δ0 , then
1. Ey x E0 x y
2. P v x P v

x

3. E
x y E0 x y
– 95 –
B. Final Review
A.5 Laplace equation

The fundamental solution to the Laplace equation in Rd is:
¢̈ log r
¨
¨ d 2
¨ 2π
E0 r ¦ 1
¨
¨ dC3
¨ d2
¤̈ dd 2αdr
Theorem A.5.1. If u > D Rd and ∆u

0 then u is smooth and harmonic.
Theorem A.5.2. If u is harmonic, then there exists a constant C C SαS sch that:
1
SD
α
uxS B C
rd S αS SB x,r

SuSdy
Theorem A.5.3. Bounded harmonic functions are constant.

Theorem A.5.4. For f > D Rd with compact support with d C 3, then solutions to ∆u

f
are unique upto addition by a constant.
Theorem A.5.5. If u is harmonic, then:
1 1
ux S
S∂B x, r ∂B x,r
udS y S
SB x, r S B x,r
u y
B Final Review
B.1 Green’s Functions
Given E0 as the fundamental solution to the Laplace equation, U an open set and u a smooth
function on Ū , then we have:
ux SU E 0 x y ∆uy dy S

∂U
E0 x y ν Duy dy S
∂U
ν Dy E0 x y uy dy
(20)
Now if u is harmonic and we choose E0 so that it vanishes on the boundary (a Green

function) then we can get results like:
Theorem B.1.1 (Mean Value Theorem). if u is harmonic then:
1
u x
dαdrd 1 SB x,r
uy dS y
Concretly, we have that:

Definition B.1.1 (Green’s Function). Given an open set U with y > U , then G , y >
C 1 Ū y is the Green’s function at y if:
¢̈
¨∆x Gx, y δ0 x y , x>U
¦
¨Gx, y 0,
¤̈
x > ∂U
– 96 –
B. Final Review
Theorem B.1.2. On an a bounded open set with C 1 boundary, Green’s functions are unique
and symmetric (Gx, y Gy, x).
We can construct Green’s functions for certain domains by the method of image charges.
This takes advantage of the symmetries of our domain, which can only be done in specific
circumstances.
By looking at (20) we have the following representation formula for u > C 2 U 9 C Ū :

∂U
This leads to
Theorem B.1.3 (Poisson Integral Formula).
ux S∂U ν y Dy Gx, y g y dS y
solves the ODE:

¢̈
¨∆u 0 x>U
¦
ü g
¤̈
x > ∂U
We even have the stronger formula (with certain conditions):
ux SU Gx, ygydy S∂U ν y

Dy Gx, y hy dS y
solves:
¢̈
¨∆u g x>U
¦
ü f
¤̈
x > ∂U
B.2 Wave Equation

Here we define the d’Alembertian operator j ∂t2 ∆ and ask to solve:
¢̈jϕ 0 t, x > R1 d
¨
¨
¨
¦ϕ g t, x > ∂R1 d
¨
¨

¨∂ ϕ h t, x > ∂R1 d

¤̈ t
In one dimension, we use change of variables to determine that the foward fundamental
solution (fundamental solution supported in nonnegative time) is:
1
E t, x
H t x H t x
2
with H the Heaviside function.
– 97 –
B. Final Review
Definition B.2.1 (Forward Fundamental Solution). To solve the wave equation we look
for forward fundamental solutions E that satisfy
1. E
j δ0
2. supp E ` t C 0

3. supp E 9 t > I is compact for all intervals I.

Theorem B.2.1. Forward fundamental solutions are unique
Theorem B.2.2. For ϕ > C ª
Rd and t, x > Rd , then:
ϕt, x E ∂t ϕδt 0 ∂t E ϕδt 0 E j ϕ1tC0
For the one dimensional case, this leads to

Theorem B.2.3 (d’Alembert’s Formula).
1 t x t s
1 1 x t
2 S0 Sx t s 2 Sx t

ϕt, x jϕ s, y dyds ϕ 0, x t ϕ 0, x t ∂t ϕ0, y dy
2
To find the forward fundamental solution for arbitrary dimensions, we take advantage of
the fact that reflections, rotations, and Lorentz boots are conserved under j. Also by
looking at scaling properties, we know that E should to be homogeneous of degree d 1,

ie:
aE , λd1 ϕλ f λd 1

È , ϕe
B.3 Fourier Transform

The motivation is to preform a change of coordinates in the space of functions to something
that behaves well with differentiation. We simply define:
Definition B.3.1 (Fourier Transfom).
F f xξ
1
S f x e
2π d Rd
iξ x
dx fÂξ
And we have the inverse:

Definition B.3.2 (Inverse Fourier Transform).
F 1 Â
f ξ x
1
2π d
S fÂξ eiξ xdξ
Proposition B.3.1. If f > L1 , we have fÂ is well defined and YfÂYLª B Yf YL1
Proposition B.3.2. For f > L1 :
F f xξ e ix ξ
F f xξ
F f xξ η F eiηx
f xξ
– 98 –
B. Final Review
Ä
Proposition B.3.3. ∂ iξj fÂ
jf
Ä
Proposition B.3.4. x jf i∂ξ fÂ
Proposition B.3.5. fÅ
g fÂÂ
g
Proposition B.3.6. F 1 f
2π d dξ g F 1 f F 1 g
The best space to take the Fourier transform is the space of Scwartz functions, which
are like test functions, but are allowed to have non-compact support as long as the decay
quickly
Definition B.3.3 (Schwartz Class on Rd ).
S Rd C ϕ > C Rd C Yxβ DαϕYLª @

α, β > Nn
ª
ª ¦
Then F and F 1 map S to S . The space of conjugate linear functionals on S are called

tempered distributions and we are allowed to take Fourier transforms (and inverse Fourier
transforms) of these elements (which will be tempered distributions).
Theorem B.3.1 (Fourier Inversion on S ). for f > S , we have that F 1F

f FF 1 f
f . The same is true for tempered distributions.
Theorem B.3.2 (Plancherel’s Theorem). For f, g > S , we have the conservation of the
two inner products:
dξ
S fÂ
g dx S fÂÂ
g
2π d
There are important integrals to keep in mind to help compute Fourier transforms
1. R eiξ xdξ limε 0 R e ε ξ eiξ xdξ S S

δ0 x2π
2. R e 1 2 ξ dξ 2πd 2
~ S S2 ~
º
3. F e 1 2 x 2πe 1 2ξ
~ 2 ~ 2
– 99 –
Index
Index
C k boundary, 17 Harnack’s inequality, 47
heat equation, 3
admissible, 14, 15 energy, 76
approximation, 28 Heaviside function, 31
Arezela-Ascoli theorem, 87 Holder space, 85
Huyghen’s principal, 64
Burger’s equation, 12, 19
hyperbolic, 3
Cauchy Riemman equations, 4
implicit function theorem, 15
characteristic ODEs, 10
initial value problem, 5
compact embedding, 87
inviscid Burger equation, 8
compatibility conditions, 14
conjugate linear, 68 KDV equation, 4
convolution, 28
Laplace Equation
d’Alembert’s formula, 57 energy, 76
d’Alembertian operator, 55 Laplace equation, 3
dirac delta function, 31 fundamental solution, 40
Dirichlet problem, 5 maximal principal, 46
representation, 51 mean value property, 46
uniqueness, 47 lecture
dispersive, 3 01 (8/29), 3
distribution, 29 02 (9/3), 6
complex valued, 67 03 (9/5), 9
convergence, 32 04 (9/10), 13
convolution, 31 05 (9/12), 16
differentiation, 31 06 (9/17), 20
order, 30 07 (9/19), 23
singular support, 36 08 (9/24), 26
support, 30 09 (9/26), 29
distributions 10 (10/1), 32
change of variables, 56 11 (10/3), 36
Duhamel’s Formula, 72 12 (10/8), 38
13 (10/15), 42
Einstein equation, 5
14 (10/17), 42
elliptic, 3
15 (10/22), 44
finite speed of propagation, 64 16 (10/24), 48
formal adjoint, 38 17 (10/29), 51
forward fundamental solution, 56 18 (10/31), 56
Fourier inversion theorem, 69 19 (11/5), 59
fundamental solution, 37 20 (11/7), 62
21 (11/12), 66
Gagliardo inequality, 83 22 (11/14), 71
Green’s function, 48 23 (11/19), 75
– 100 –
Index
24 (11/21), 80 question, 20, 40, 42, 80

25 (12/3), 84 quote, 22
26 (12/5), 86
linear homogeneous, 4 real analytic, 24
Looms-Whitney inequality, 83 Relich-Konoluchov theorem, 87
Lorentz transformation, 60 representation formula, 39
Ricci flow, 5
Maxwell equations, 4
method of index charges, 50 Schrodinger equation, 3
minimal surface equation, 4 Schwartz class, 69
mollifier, 28 Sobolev Space
Morey’s inequality, 85 approximation, 79
noncharacteristic, 15, 24 Sobolev space, 77
Nuemman problem, 5 H k , 78
W0k,p , 78
parabolic, 3 extension, 80
Plancherel theorem, 69
Poincare inequality, 89 tempered distributions, 69
Poisson equation, 4 test functions, 27
Poisson integral formula, 51
principal value, 32 wave equation, 3
– 101 –

Math222anotes PDF

Uploaded by

Copyright:

Available Formats

Math222anotes PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Math222anotes PDF

Uploaded by

Copyright:

Available Formats

Math 222a (Partial Differential Equations 1) Lecture

6 Energy Method and Sobolev Spaces 75

3. Sobolev spaces (Evans Chapter 6)

1.1 Overview of PDE’s

Example 1.2.3 (Heat Equation). u R1  d R, C has the form ∂t  ∆u 0

Example 1.2.5 (Cauchy-Riemann). u, v R2 R, we have:

Example 1.2.6 (Maxwell Equations (vacuum)). Given E R1  3 R3 and B R1

It turns out that jE j B 0.

Note that all these PDE’s can be written as:

There are many nonlinear PDE examples:

Example 1.2.7 (Minimal Surface Equation). u Rd R with:

with the notation D © ∂

Example 1.2.8 (Korteng-de Vries (KDV) Equation). u R1  1 R with:

This is an example of a dispersive nonlinear PDE.

Here are nonlinear systems of PDEs

Example 1.2.9 (Ricci Flow). g a Riemannian metric (positive definite symmetric n  n

Where Ric g  1 ∂ 2 g  g 1 ∂g g 1 ∂g

Example 1.2.10 (Einstein Equation). g a Lorentzian metric which is a 1  3  1  3

1.3 Basic Problems

Example 1.3.2 (Neumann Problem). Given U and u as above, but:

where ν is the outer normal to ∂U , and we require RU f dx R∂U gdA

For the wave equation:

3. continuous dependence on dataa

a PDE that fails this is called illposed

Let’s introduce notation. First the multi-index notation: given α α 1 , . . . , α d  then we

∂α Dα ∂1α1 ∂2α2 ∂dαd

Nonlinear PDEs are classified into three categories:

1. semilinear: 0 F u P α d aα∂ αu nonlinear stuff

2. quasilinear: 0 F u P α k aαu, Du, . . . , Dk 1ux∂ αu

1.4 First Order (scalar) PDEs

1. Positive example: ∂t  at, x∂x u f where u R1 1 R, f R1   1 R, a R1  1 R

2. Inviscid Burger equation ∂t u  u ∂x u 0 with u R1  1 R.

3. Lewy–Nirenberg example: ∂t  it∂x u f with u, f R1  1 C.

It turns out that there exists a function f > C ª

If we identify C by R2 , this is also an example of a linear first-order system of PDEs.

Let’s begin with a linear first order scalar PDE.

First suppose a 0 (this is known as a transport equation), then we have ∂t u f , so we

We could also have initial value problem:

Theorem 1.4.1. Given u0 R R smooth, and f R1  1 R smooth. Then there exists a

Proof. (method of characteristics)

The last step is to show that x ( y is invertible. We have:

Therefore x is strictly increasing or strictly decreasing as we increase y, therefore we have

Now let’s turn to Lewy-Nirenberg’s example:

From now on u ũ. On the one hand, we have:

By the divergence theorem, the left side is just:

On U , change the variables to t, x s, y 

1.5 Method of Characteristics

Assume that a smooth solution u exists. Then we have:

Therefore (10) becomes:

ż s Q ∂iuxsẋis Q pis∂p F xs, z s, ps

Theorem 1.5.1. If u is a smooth solution to F x, u, Du 0. Then on a curve x satisfying

will obey (4) and (5)

where bj bj x and c c x .

In this case F x, z, p P j bj p j  cz. Then if we plug in our characteristic equations, we

ẋj s bj xs

With U x1 , x2  > R2 , x1 A 0, x2 A 0 and Γ x1 , 0 > R2  ` ∂U .

Therefore F x1 p2  x2 p1  z, so b x2 , x1 , and c  1. If we use the characteristic

F x, u, Du Q bj x, u∂j u  cx, u

In this case we have:

Example 1.2.3 (Heat Equation). u R1 d R, C has the form ∂t ∆u 0

Example 1.2.6 (Maxwell Equations (vacuum)). Given E R1 3 R3 and B R1

Example 1.2.8 (Korteng-de Vries (KDV) Equation). u R1 1 R with:

Example 1.2.9 (Ricci Flow). g a Riemannian metric (positive definite symmetric n n

Where Ric g 1 ∂ 2 g g 1 ∂g g 1 ∂g

Example 1.2.10 (Einstein Equation). g a Lorentzian metric which is a 1 3 1 3

Let’s introduce notation. First the multi-index notation: given α α 1 , . . . , α d then we

∂α Dα ∂1α1 ∂2α2 ∂dαd

1. semilinear: 0 F u P α d aα∂ αu nonlinear stuff

2. quasilinear: 0 F u P α k aαu, Du, . . . , Dk 1ux∂ αu

1. Positive example: ∂t at, x∂x u f where u R1 1 R, f R1 1 R, a R1 1 R

2. Inviscid Burger equation ∂t u u ∂x u 0 with u R1 1 R.

3. Lewy–Nirenberg example: ∂t it∂x u f with u, f R1 1 C.

Theorem 1.4.1. Given u0 R R smooth, and f R1 1 R smooth. Then there exists a

On U , change the variables to t, x s, y

ż s Q ∂iuxsẋis Q pis∂p F xs, z s, ps

Theorem 1.5.1. If u is a smooth solution to F x, u, Du 0. Then on a curve x satisfying

where bj bj x and c c x .

In this case F x, z, p P j bj p j cz. Then if we plug in our characteristic equations, we

ẋj s bj xs

With U x1 , x2 > R2 , x1 A 0, x2 A 0 and Γ x1 , 0 > R2 ` ∂U .

Therefore F x1 p2 x2 p1 z, so b x2 , x1 , and c 1. If we use the characteristic

F x, u, Du Q bj x, u∂j u cx, u

F Q bj x, z pj cx, z

with U x1 , x2 > R2 , x2 A 0 and Γ x1 , 0 > R2 . So we have b1 1, b2 1 and c z2.

So we have x1 , x2 x 0 s, s.

Now let U x1 A 0 and Γ 0, x2 > R2 . Now we have:

ẋj ∂pj F x, z, p

Let’s start by considering a simpler case. Let U Rd

Let’s say we have the data x, z, p0 at x0 .

Theorem 1.6.1 (Implicit Function Theorem). Given F Rm Rn Rm which is C 1 and

Moreover, if F y, p 0 for some y > U and p > V then p p y

which is admissible. And ∂pd F x0 , z0 , p0 x 0 noncharactersitic. Then we have:

Compute the determinant which is just ẋd x0 , 0 ∂pn F x0 , z0 , p0 x 0

Q bj x, u∂j u cx, u 0

F Q bj x, zpj cx, z

Then the noncharacteristic condition requires ∂pd F x0 , z0 , p0 bd x0 , z0 x 0. So in the

choice p0 d from everything else.

∂s F x, z, p Q ∂x F ẋi i ∂z F ż Q ∂j F ṗj

pi z y x, sx Q ∂y z∂iyj j

r j s ∂yj z Q pi y, s∂yj xi y, s

rj 0 0 by preparation. Then by a long computation, we get that:

Review: in the case Γ ` xd 0 U xd A 0. We wanted to solve:

Noncharacteristic x0 , z0 , p0 means that ∂pd F x0 , z0 , p0 x 0. This allows us to solve

Then U 9 B x0 , r ` y d A 0 and ∂U 9 B x0 , r ` y d 0.

V ? x0 . So if v y uxy , then v satisfies a PDE of the same form.

does u1 u2 ?

u1 x u2 x by uniqueness of ODEs.

Suppose that g > C 2 and Yg YC 2 B A a then there exists a neighborhood V 0, T R with

F x, u, Du Q aj x, u∂j u a0 x, u 0

In this case x0 , p0 , z0 became our characteristic if and only if ad x0 , z0 x 0.

F x, z, p Q aj x, z pj a0 x, z

x0 , z 0 is noncharacteristic, therefore this is equivalent to:

Q aj x0 , z0 p0 j a0 x0 , z0

If we are interested, we can determine ∂ α ux0 under he noncharacteristic assumption.

p0 d ∂d ux0 . Now suppose we want ∂i ∂j ux0 . This is easy if i j 1, 2, . . . , d 1. But

0 Q aj x, u∂j ∂du bx, u, ∂u

So the noncharacteristic assumption of x0 , z0 is equivalent to determining all derivatives

4.6.1: noncharacteristic cauchy data for quasilinear k th order systems

4.6.3: Cauchy - Kovalewski theorem

The Cauchy data is data for u, ∂ν u, . . . , ∂νkk11 u on Γ ` ∂U .

b0,...,0,k ∂dk ux0 Q bα D

We need that b0,...,0,d is invertible, in which case we have that:

except ∂dk 1 ux0 .

b0,...,0,d ∂dk 1 ux0

Definition 1.7.1 (noncharacterisitic condition). g0 , . . . , gk 1 is noncharacteristic at

x0 if b0,...,0,d evaluated at x0 , ux0 , . . . , Dk 1 ux0 is invertiblea .

Theorem 1.7.1. If g0 , . . . , gk 1 is noncharacterisitic at x0 , then all derivatives of u at x0

with γ a smooth function. Then we can take y i xi for i 1, . . . , d 1 and y d xd

Then we require that b̃0,...,0,d is invertible.

Du Dv y x Dy i ∂yi v y x