MATH 3120 - INTRO TO PARTIAL DIFFERENTIAL EQUATIONS

MARCELO DISCONZI∗

Contents
1. Abbreviations 2
2. What are partial differential equations and why do we study them? 2
3. Examples and notation 3
3.1. Laplace’s equation: 3
3.2. Heat equation or diffusion equation 4
3.3. Wave equation 4
3.4. Schödinger equation 4
3.5. Burger’s equation 4
3.6. Maxwell’s equations 4
3.7. Euler and Navier-Stokes equations 6
3.8. Other examples 6
3.9. Theory and Example 6
4. Schrödinger equation 6
4.1. Introduction 6
4.2. Separation of variables for a time-independent potential 7
4.2.1. The time-independent Schrödinger equation 7
4.2.2. The angular equation 8
4.2.3. The radial equation 12
4.3. Final comments 16
5. Separation of variables for the one-dimensional wave equation 18
6. Fourier Series 20
6.1. Piecewise Functions 21
6.2. Convergence of Fourier Series 22
6.3. Some Intuition Behind Fourier Series 24
6.4. The Fourier series of series of periodic functions, and the Fourier series of functions on
[0, L] 25
6.5. Back to the wave equation 26
7. The 1d wave equation in R 27
7.1. Regions of influence for the 1d wave equation 29
7.2. Generalized solutions 31
8. Some general tools, definitions, and conventions for the study of PDEs 32
8.1. Domains and boundaries 32
8.2. The Kronecker delta 33
8.3. Raising and lowering indices with δ 34
8.4. Calculus facts 35
9. Formal aspects of PDEs 36
10. Laplace’s Equation in Rn 40
10.1. Harmonic functions 45
∗
Vanderbilt University, Nashville, TN, USA. [email protected].

1
2 Disconzi

10.2. Further results for harmonic functions and Poisson’s equation 47

11. The wave equation in Rn 47
11.1. The Reflection Method 51
11.2. Solution for n = 3 : Kirchhoff’s formula 52
11.3. Solution for n = 2 : Poisson’s formula 53
References 54

1. Abbreviations
• ODE = ordinary differential equation
• PDE = partial differential equation
• LHS = left hand side
• RHS = right hand side
• w.r.t = with respect to
• ⇒ = implies
• LHS := RHS means that the LHS is defined by the RHS.
• nd (e.g. 1d, 2d, ...) = n dimensional

2. What are partial differential equations and why do we study them?

Recall that an ordinary differential equation (ODE) is an equation involving an unknown function
of a single variable and some of its derivatives. For example,
dy
+ y 2 = 0, (unknown y, non-linear 1st order) (2.1)
dx
y 00 + y 0 + y = 0, (unknown y, linear 2nd order) (2.2)
d2 u
(x2 − 1) + u = 0, (unknown u, linear 2nd order) (2.3)
d2 x
(2.4)
are ODEs. We can also have systems of ODEs, i.e., a system of equations involving two or more
unknown functions of a single variable and their derivatives. For example,

dt + x = 0 (unknowns: y and x, linear, 1st order)

(
dy

dx (2.5)
dt − y = 0

 u + v = 0 (unknowns: u, v, w, non linear)


 00 2

v 00 + v 2 = 0 (2.6)
 00
 0
w + w + wv = 0
are systems of ODEs. As we learn in ODE courses, one typically studies ODEs because many
phenomena in science and engineering are modeled with ODEs. A limitation of ODEs, however,
is that they are restricted to functions of a single variable, whereas many important phenomena
are described by functions of several variables . For instance, suppose we want to describe the
temperature T in a room. It will in general be different at different positions in the room, so T is
a function of (x, y, z). T can also change over time, thus T = T (t, x, y, z). An equation involving T
and its derivatives can then have derivatives with respect to any of the variables t, x, y, or z, which
will be partial derivatives, ∂t
∂
, ∂x
∂
, ∂y
∂
, ∂z
∂
. This will be a partial differential equation. Formally:
Definition 2.1. A partial differential equation is an equation involving an unknown function of
two or more variables and some of its (partial) derivatives. A system of PDEs is a system of
 MATH 3120 - Intro to PDEs 3

equations involving two or more unknown functions of two or more variables and some of their
(partial) derivatives. A solution to a PDE (or system) is a function that verifies the PDE.
Notation 2.2. Since most of the time we sill be dealing with functions of several variables, the
derivatives will be partial derivatives, but we will often omit the word “partial”, referring simply
to “derivatives.” We will also often omit “system”, and use PDE to refer to both a single equation
and systems of PDEs.
Besides applications to science and engineering PDEs are also used in many branches of math-
ematics, such as in complex analsis or geometry (see in particular Ricci flow and the Poincaré
conjecture). PDEs are also studied in mathematics for their own sake, i.e., from a “pure” point of
view.

3. Examples and notation

We will now give examples of PDEs. Along the way, we will introduce some notation that will
be used throughout.
Remark 3.1. As it was the case for ODEs, when we introduce a PDE, strictly speaking we have
to specify where the equation is defined. We will often ignore this for the time being until we get
to some more formal aspects of PDE theory.
3.1. Laplace’s equation:
4u = 0,
where 4 is the Laplacian operator defined by
∂2 ∂2 ∂2
4 := + + ,
∂x2 ∂y 2 ∂z 2
so explicitly Laplace’s equation reads:
∂2u ∂2u ∂2u
+ 2 + 2 = 0.
∂x2 ∂y ∂z

We will often denote coordinates in R3 by (x1 , x2 , x3 ), in which case we write 4 as 4 = ∂2

∂(x1 )2
+
∂2
+
∂(x2 )2
∂2
∂(x3 )2
.
We write expressions of the form u = u(x1 , x2 , x3 )to indicate the variables that a
function depends on, e.g., in this case that u is a function of x , x , and x3 . We can also consider
1 2

Laplace’s equations for a function of x1 , x2 , · · · , xn , for some arbitrary n, u = u(x1 , x2 , · · · , xn ), in

which case
∂2 ∂2 ∂2
4 := + + · · · + ,
∂(x1 )2 ∂(x2 )2 ∂(xn )2
so Laplace’s equation reads
n
∂2u ∂2u ∂2u X ∂2u
4u = + + · · · + = =0
∂(x1 )2 ∂(x2 )2 ∂(xn )2 ∂(xi )2
i=1

Laplace’s equation has many applications. Typically, u represents the density of some quantity
(e.g., a chemical concentration). Closely related to Laplace’s equation is the Poisson equation:
4u = f,
where f is a given function.
4 Disconzi

3.2. Heat equation or diffusion equation.

∂t u − 4u = 0.
This equation has many applications. For example, u can represent the temperature so that
u(t, x1 , x2 , x3 ) is the temperature at the point (x1 , x2 , x3 ) at instant t. More generally u can
represent the concentration of some quantity that diffuses over time.
Notation 3.2. Throughout these notes, we will use t to denote a time variable, unless otherwise
specified.
Remark 3.3. The heat equation is also written as ∂t u − k4u = 0, where k is a constant known as
diffusivity. In most of these notes, we will ignore physical constants in the equations, setting them
equal to 1.
3.3. Wave equation.
utt − 4u = 0
(Here we recall the notation ut = ∂t u = ∂t ,
2 u = ∂ u etc.). This equation describes a wave
∂u 2
utt = ∂tt ∂t2
propagating in a medium (e.g., a radio wave propagating in space); u is the amplitude of the wave.
Sometimes one writes utt − c2 4u = 0 where the constant c is the speed of propagation of the
wave (we will see later on why c is indeed the speed of propagation).
3.4. Schödinger equation.
∂Ψ
i + 4Ψ + V Ψ = 0,
∂t
where i is the complex unit i2 = −1, V = V (t, x1 , x2 , x3 ) is a known function called the potential
(whose specific form depends on the problem we are studying), and the unknown function Ψ, called
the wave-function, is a complex function, i.e.,
Ψ = u + iv,
where u and v are real valued functions.
The Schrödinger equation is the fundamental equation of quantum mechanics.
3.5. Burger’s equation.
ut + uux = 0.
Burger’s equation has applications in the study of shock waves.
3.6. Maxwell’s equations.
∂t E − curlB = −J,


∂t B + curlE = 0,



(3.1)

 div E = ρ,

div B = 0,


where the E and B are vector fields that are the unknown functions (or vector valued functions),
so they have three components each:
E = (E 1 , E 2 , E 3 ), (3.2)
1 2
B = (B , B , B ), 3
(3.3)
div and curl are the divergence and curl operators, sometimes written as ∇· and ∇×, respectively
(curl is also called the rotational). Let us recall the definition of these operators: for any vector
field X = (X 1 , X 2 , X 3 ), we have
div X := ∂1 X 1 + ∂2 X 2 + ∂3 X 3 ,
 MATH 3120 - Intro to PDEs 5

and
curlX := (∂2 X 3 − ∂3 X 2 , −∂1 X 3 + ∂3 X 1 , ∂2 X 3 − ∂3 X 2
where we have introduced the following notation:
∂
∂i := .
∂xi
E and B represent the electric and magnetic fields respectively. ρ represents the charge density
and J the current density, which are given.
Maxwell’s equations are the fundamental equations of electromagnetism.
−
→ −
→
Notation 3.4. Note that the above, we did not denote vectors with an “arrow” i.e., E and B , as
usually done in calculus. We will avoid using arrows for vectors - it will always be clear from the
context if a quantity is a scalar, a vector field, etc. We also denote the components or entries of
a vector with superscripts and not with subscripts as usually in calculus (i.e., X i and not Xi , but
see below for exceptions).
Similarly, we will denote points in space by a single letter without an arrow, e.g., x = (x1 , x2 , x3 )
in R3 , or more generally x = (x1 , x2 , x3 , · · · , x4 ) in R4 . So, sometimes we write expressions like
u = u(t, x) instead of u = u(t, x1 , x2 , x3 ).
Notation 3.5. The curl can be written in a compact form as
(curlX)i = ijk ∂j Xk .
meaning the ith component of the vectorcurl X
| {z }

In this expression, the following convention is adopted.  is the totally anti-symmetric symbol,
defined as
 +1 if ijk is an even permutation of 123


 := −1 if ijk is an odd permutation of 123
ijk
(3.4)
otherwise.

0


E.g., 123 = 1, 231 = 1, 213 = −1, 112 = 0. Xk means X k , but we write it here with a subscript
because of the following summation convention which will be used throughout:
Notation 3.6. When an index (such as i, j, etc.) appears repeated in an expression, once upstairs
and once downstairs, it is summed over its range.
E.g., we can write the divergence as
3
X
div X = ∂i X i = ∂i X i = ∂1 X 1 + ∂2 X 2 + ∂3 X 3 .
i=1

Remark 3.7. We will give another interpretation to Xk (i.e., X k but with the index downstairs)
which will make our conventions more systematic, later on.
In the expression for curl, for example:
(curlX)2 = 2jk ∂j Xk (3.5)
= 213 ∂1 X3 + 231 ∂3 X1 (3.6)
= ∂1 X3 + ∂3 X1 . (3.7)
We also sometimes use the notation
curl i X = (curlX)i .
6 Disconzi

3.7. Euler and Navier-Stokes equations.

(
∂t ρ + (u · ∇)ρ + ρ div u = 0
(3.8)
ρ(∂t u + (u · ∇)u) + ∇p = µ4u
These equations describe the motion of a fluid. The first equation is sometimes called the conti-
nuity equations (conservation of mass) and the second one the momentum equation (conservation
of momentum)
ρ = ρ(t, x) is a scalar function representing the fluid’s density and u = u(t, x) is a vector field
representing the fluid’s velocity. ρ and u are the unknowns. p is a given function of ρ, i.e., p = p(ρ)
(e.g., p(ρ) = ρ2 ). p represents the pressure of the fluid. µ ≥ 0 is a constant known as the viscosity of
the fluid. ∇ is the gradient operator; recall that ∇f := (∂1 f, ∂2 f, ∂3 f ), where f is a scalar function,
so the ith component reads (∇f )i = ∂i f ; we also write ∇i f for (∇f )i .
u · ∇ is the operator
u · ∇ = ui ∂i (3.9)
1 2 3
= u ∂1 + u ∂2 + u ∂3 . (3.10)
When u · ∇ acts on a vector field it does so componentwise. 4 also acts on a vector field compo-
nentwise.
These equations are known as the Navier-Stokes equations if µ > 0 and Euler equations if µ = 0.
They are the fundamental equations of hydrodynamics.
In models where the density is assumed to be constant, we take ρ = 1, we have the incompressible
Euler or Navier-Stokes equations:
(
div u = 0
(3.11)
∂t u + (u · ∇)u + ∇p = µ4u
In this case, however, it is no longer assumed that p = p(ρ), and p is given by some other expression
(we will see this later).
3.8. Other examples. There are many other important PDEs that we will not have time to
discuss. We mention a few more of them, without writing them explicitly:
Einstein’s equations: fundamental equations of general relativity
Yang-Mills equations: fundamental equations of quantum field theory
Black-Scholes equation: models the price of European options.
Remark 3.8. The concepts of the order of a PDE and of homogeneous vs. non-homogeneous
PDEs are defined similarly to their analogs in ODEs. We will define linear and non-linear PDEs
later on, but this definition is also similar to ODEs and readers should be able to identify which of
the above examples are linear or non-linear PDEs.
3.9. Theory and Example. Before investigating more general and theoretical aspects of PDEs, it
is useful to first consider a few specific equations that can be solved explicitly. Thus, the beginning
will be more computational and equation specific. Later on we will consider more robust aspects
of the general theory of PDEs.

4. Schrödinger equation
4.1. Introduction. Our goal is to investigate solutions to the Schrödinger equation,
∂Ψ ~2
i~ = − ∆Ψ + V Ψ, (4.1)
∂t 2µ
where i is the imaginary number i2 = −1; ~ = 1.51 × 10−27 erg s is Planck’s constant; µ is a
positive constant called the mass; V = V (t, x) : R × R3 → R is called the potential function; and
 MATH 3120 - Intro to PDEs 7

the unknown is the complex-valued function Ψ = Ψ(t, x) : R × R3 → C called the wave-function.

The variables t and x represent, respectively, the time and space variables.
The Schrödinger equation describes the dynamics of a particle of mass µ interacting with a
potential V , according to the laws of Quantum Mechanics. The physical interpretation of Ψ is as
follows. If U ⊆ R3 , then
Z
|Ψ(t, x)|2 dx
U
represents the probability of finding the particle in the region U at a time t. In particular, one
must have
Z
|Ψ(t, x)|2 dx = 1. (4.2)
R3
Notice that, upon multiplying Ψ by a suitable constant, condition (4.2) can always be fulfilled as
long as
Z
|Ψ(t, x)|2 dx < ∞. (4.3)
R3
Our treatment will be based on [2–4], to which the student is referred for more details.

4.2. Separation of variables for a time-independent potential. We shall assume that V

does not depend on time, i.e., V (t, x) = V (x). We will have to divide several expressions by Ψ. In
order to make this sensible, it will be assumed that Ψ does not vanish (or, at least, does not vanish
on an open set). Look for solutions of the form
Ψ(t, x) = T (t)ψ(x), (4.4)
Plugging (4.4) into (4.1) gives
T0 ~2 1
i~ =− ∆ψ + V,
T 2µ ψ
The left-hand side depends only on t, whereas the right-hand side depends only on x. Thus, both
sides have to be equal to a constant, which we denote be E. Therefore
i~T 0 = ET, (4.5)
and
~2
− ∆ψ + V ψ = Eψ. (4.6)
2µ
Equation (4.5) is easily solved. Its solution is

(4.7)
iE
T (t) = e− ~
t
,
where we ignored an arbitrary constant of integration (such constants will be neglected throughout,
as an overall constant of integration can be fixed at the very end via condition (4.2)).

4.2.1. The time-independent Schrödinger equation. We now focus on (4.6), known as the time-
p equation. To solve it, we assume further that V is radially symmetric,
independent Schrödinger
i.e., that V (x) = V ( x21 + x22 + x23 ) or, in spherical coordinates, that V = V (r). This assumption
suffices to treat many physical systems of interest.
Recall the expression for the Laplacian in spherical coordinates,
2 1
∆ = ∂r2 + ∂r + 2 ∆S 2 , (4.8)
r r
8 Disconzi

where
cos φ 1
∆S 2 = ∂φ2 + ∂φ + ∂2 (4.9)
sin φ sin2 φ θ
is the Laplacian on the unit sphere, r ∈ [0, ∞), φ ∈ [0, π], and θ ∈ [0, 2π). From now on we shall
work in spherical coordinates.
We suppose that
ψ(r, φ, θ) = R(r)Y (φ, θ). (4.10)
Plugging (4.10) into (4.6), and using (4.8),
~2 r2 ~2 1
 
00 2 0
− R + R + (V − E)r2 = ∆ 2 Y.
2µ R r 2µ Y S
Since the left-hand side depends only on r and the right-hand side only on (φ, θ), both sides must
be equal to a constant, which we denote by −a. Thus,
~2
  
2 0 a
− 00
R + R + V + 2 R = ER, (4.11)
2µ r r
and
~2
∆ 2 Y = −aY. (4.12)
2µ S
4.2.2. The angular equation. We first investigate (4.12), which, in light of (4.9), becomes
cos φ 1 2aµ
∂φ2 Y + ∂φ Y + 2 ∂θ2 Y = − 2 Y.
sin φ sin φ ~
Supposing
Y (φ, θ) = Φ(φ)Θ(θ), (4.13)
one gets
Θ00 sin2 φ 00 sin φ cos φ 0 2aµ sin2 φ
− = Φ + Φ + . (4.14)
Θ Φ Φ ~2
Once more, both sides ought to be equal to a constant, which we denote by b. One equation
becomes
Θ00 = −bΘ. (4.15)
To solve (4.15), we need to analyze the cases b > 0, b = 0, and b < 0. Notice the following boundary
condition: the points with coordinates θ and θ + 2π must be identified as they correspond to the
same point in R3 . Thus,
Θ(θ + 2π) = Θ(θ). (4.16)
We immediately see that the case b < 0 does not yield a solution satisfying (4.16); b = 0 and
√ (4.16)
give√Θ =constant; and b > 0 along with (4.16) give
√ that Θ is a linear combination of cos( bθ) and
sin( bθ). Moreover, 2π-periodicity requires that b = integer. All these cases can be summarized
by setting
b = m2 , m ∈ Z, (4.17)
and writing
Θ(θ) = eimθ . (4.18)
 MATH 3120 - Intro to PDEs 9

Next, we move to the Φ-equation. From (4.14) and (4.17), one has
 
sin φ d dΦ
sin φ − m2 = −λ sin2 φ, (4.19)
Φ dφ dφ
where
2µ
λ=
a, (4.20)
~2
and we used the product rule to rewrite the terms involving derivatives. In order to solve (4.19),
let us make the following change of variables,
x = cos φ, 0 ≤ φ ≤ π.
Notice that this change of variables is well-defined since cos is one-to-one for 0 ≤ φ ≤ π. The chain
rule gives
d dx d d d d
sin φ = sin φ = − sin2 φ = (cos2 φ − 1) = (x2 − 1) ,
dφ dφ dx dx dx dx
so that (4.19) becomes
m2
   
d 2 dΦ
(1 − x ) + λ− Φ = 0. (4.21)
dx dx 1 − x2
To solve (4.21), we seek for a solution of the form
|m| d|m| P (x)
Φ(x) = (1 − x2 ) 2 , (4.22)
dx|m|
where P solves
d2 P dP
(1 − x2 )
2
− 2x + λP = 0. (4.23)
dx dx
To see that this works, differentiate (4.23) |m| times, obtaining
d|m|+2 P d|m|+1 P d|m| P
(1 − x2 ) − 2(|m| + 1)x + (λ − |m|(|m| + 1)) = 0. (4.24)
dx|m|+2 dx|m|+1 dx|m|
Students are encouraged to verify (4.24) (compute the first few derivatives to see that a pattern as
(4.24) emerges). On the other hand, let Φe be defined by
|m|
Φ(x) = (1 − x2 ) 2 Φ(x)
e (4.25)
and plug this into (4.21). Computing the derivative terms,
 !
d d  |m|
 d |m| |m| |m| dΦe
(1 − x2 ) (1 − x2 ) 2 Φe = (−2x)(1 − x2 ) 2 Φ e + (1 − x2 ) 2 +1
dx dx dx 2 dx
d2 Φ
  
|m| e |m| dΦ |m| |m|
= (1 − x2 ) 2 +1 2 + (1 − x2 ) 2 + 1 (−2x) + (−2x)
dx dx 2 2
 
|m| |m| |m| |m|
+ (−2x) (1 − x2 ) 2 −1 (−2x) − 2(1 − x2 ) 2 Φ e
2 2
d2 Φ 2|m|x2
 
|m| e |m| dΦ
e |m| |m|
= (1 − x2 ) 2 +1 2 − 2x(1 − x2 ) 2 (|m| + 1) + (1 − x2 ) 2 − 2 Φ
e
dx dx 2 1 − x2
2Φ
 2
 !
|m| d e d Φ
e |m|x
= (1 − x2 ) 2 (1 − x2 ) 2 − 2x (|m| + 1) + |m| −1 Φ e .
dx dx 1 − x2
10 Disconzi

By (4.21), this has to equal

m2 m2
   
|m|
− λ− Φ=− λ− (1 − x2 ) 2 Φ(x),
e
1 − x2 1 − x2
|m|
what gives, after canceling (1 − x2 ) 2 ,
d2 Φ |m|2 x2 |m|2
 
2
e dΦ
e
(1 − x ) 2 − 2x (|m| + 1) + λΦ +
e − |m| − Φ
e = 0.
dx dx 1 − x2 1 − x2
But
|m|2 x2 |m|2 |m|(|m| + 1)x2 − |m|(|m| + 1)
− |m| − = = −|m|(|m| + 1),
1 − x2 1 − x2 1 − x2
and therefore
d2 Φ dΦ
(4.26)
e e
(1 − x2 ) 2 − 2x (|m| + 1) + (λ − |m|(|m| + 1)) Φ
e = 0.
dx dx
Comparing (4.26) with (4.24), we see that if P solves (4.23), then (4.25) solves (4.21), as claimed.
Therefore, it suffices to solve (4.23). We seek a power series solution of the form
∞
(4.27)
X
P (x) = ak xk .
k=0

Plugging (4.27) into (4.23) gives

∞
X ∞
X ∞
X
(1 − x2 ) k(k − 1)ak xk−2 − 2x kak xk−1 + λ ak xk = 0,
k=0 k=0 k=0
or yet, after rearranging some terms,
∞
X
((k + 2)(k + 1)ak+2 − (k(k + 1) − λ)ak ) xk = 0,
k=0
which implies the following recurrence relation,
k(k + 1) − λ
ak+2 = ak , k = 0, 1, 2, . . . . (4.28)
(k + 1)(k + 2)
Relation (4.28) determines all coefficients ak except for a0 and a1 , which remain arbitrary (this
is consistent with the fact that we are solving a second order ODE). Furthermore, a0 determines
all even coefficients, giving rise to an even power series, while while a1 determines all odd coeffi-
cients, giving rise to an odd power series. These two power series, even and odd, are two linearly
independent solutions of (4.23).
Next, we investigate the convergence of (4.27). Since it suffices to investigate the convergence of
the even and odd expansions separately, as these are two linearly independent solutions, the ratio
between two consecutive terms in the expansion is obtained from (4.28), yielding
ak+2 xk+2
lim = |x|2 ,
k→∞ ak xk
and thus (4.27) converges for |x| < 1 by the ratio test. We need to investigate the case |x| = 1
(which corresponds to φ = 0 or φ = π). Plugging x = ±1 into (4.27) gives
∞
(4.29)
X
P (±1) = ± ak .
k=0
 MATH 3120 - Intro to PDEs 11

From (4.28) we have

 k+2
k2 + O(k) k2
+ O(k) (k − + O(k) 2)2  k +O(kk+1 ) a0 , k even,
k+2 k+1 )
ak+2 = ak = 2 ak−2 = · · · = kkk+1 +O(k k
k 2 + O(k) k + O(k) (k − 2)2 + O(k)  k+1 +O(kk ) a1 ,
k +O(k )
k odd.
It follows that
lim ak 6= 0,
k→∞
and therefore (4.29) diverges by the divergence test, unless (4.27) is in fact a finite sum; i.e., unless
ak = 0 for all k greater than a certain `. Hence, we must have, form some non-negative integer `,
`(` + 1) − λ
a`+2 = 0 = a` ,
(` + 1)(` + 2)
which implies
λ = `(` + 1), (4.30)
provided that a` 6= 0. Relation (4.30) determines λ, and hence the separation constant a in view
of (4.20). The conclusion is that there is a family {P` } of solutions to (4.23) parametrized by
` = 0, 1, 2, . . . . After conveniently choosing a0 and a1 to obtain integer coefficients, the first few
P ’s are
P0 (x) = 1, P1 (x) = x, P2 (x) = 1 − 3x2 , P3 (x) = 3x − 5x3 .
Since P` is a polynomial of degree `, from (4.22) it follows that Φ = 0 for |m| > `. Thus, the values
of m are restricted to |m| ≤ `, i.e., the allowed m-values depend on ` and satisfy
(4.31)

m ∈ − `, −` + 1, . . . , −1, 0, 1, . . . , ` − 1, ` .
We write m = m` when we want to stress this dependence of the allowed values of m on `. One
obtains a family of solutions {Φ`m` } to (4.21) parametrized by ` and m` , where ` = 0, 1, 2, . . . and
m` satisfies (4.31). The first few Φ’s are
Φ00 (x) = 1,
1
Φ10 (x) = x, Φ1,±1 (x) = (1 − x2 ) 2 ,
1
Φ20 (x) = 1 − 3x2 , Φ2±1 (x) = (1 − x2 ) 2 x, Φ2±2 (x) = 1 − x2 ,
3 3
Φ30 (x) = 3x − 5x3 , Φ3±1 (x) = (1 − x2 ) 2 (1 − 5x2 ), Φ3±2 (x) = (1 − x2 )x, Φ3±3 (1 − x2 ) 2 .
Finally, it is necessary to rewrite our solutions in terms of the φ variable. Denoting
F`m = Φ e `m ,
` `

and using 1 − x2 2
= sin φ,
Φ`m` (φ) = sin|m` | F`m` (cos φ), ` = 0, 1, 2, . . . , |m` | ≤ `. (4.32)
Combining (4.13), (4.18), and (4.32) gives
Y`m` (φ, θ) = eim` θ sin|m` | φF`m` (cos φ), ` = 0, 1, 2, . . . , |m` | ≤ `. (4.33)
Notice that, in view of (4.12) and (4.20), Y`,m` solves
∆S 2 Y`m` = −`(` + 1)Y`m` ,
which is an eigenvalue problem for ∆S 2 ; thus, the spherical harmonics are the eigenfunctions of ∆S 2 .
We finish this section with some terminology. Equation (4.23) is known as Legendre equation, and
its solutions P` are known as Legendre polynomials. The functions F`m` are known as associated
Legendre functions. The functions Y`m` are called spherical harmonics. Legendre functions and
12 Disconzi

spherical harmonics have many important applications in Physics. The interested reader is referred
to [1] for details.
4.2.3. The radial equation. We now turn our attention to equation (4.11). Using (4.20) and (4.30),
equation (4.11) can be written as
 
1 d 2 dR 2µ R
2
r + 2 (E − V (r)) R = `(` + 1) 2 . (4.34)
r dr dr ~ r
It is important to stress that the results of section 4.2.2 are general, i.e., they apply to separation
of variables to any radially symmetric potential V = V (r). To solve (4.34), on the other hand,
we need to specify the function V (r). We shall assume that V is the potential describing the
electromagnetic interaction of an electron with a nucleus. This covers the important case when one
is solving the Schrödinger equation describing the evolution of an electron on a hydrogen atom. In
this situation, V takes the form
Ze2
V (r) = −
, (4.35)
4πε0 r
where Z is the nuclear charge (for example, Z = 1 for the hydrogen and Z = 2 for an ionized
helium atom), −e is the electron charge, where e = 1.6 × 10−19 C, and ε0 is the vacuum permitivity
whose values is ε0 = 8.85 × 10−12 F/m (farads per meters). In order to investigate solutions to
(4.34) with V given by (4.35), one needs more information about the separation constant E.
We claim that E must be real and negative. To see this, multiply equation (4.34) by r2 R∗ , where
R is the complex conjugate of R, and integrate from 0 to ∞:
∗
Z ∞
2µ ∞
  Z Z ∞ Z ∞
∗ d 2 dR 2µ
R r dr − 2 2 2
V |R| r dr − `(` + 1) 2
|R| dr = − 2 E |R|2 r2 dr, (4.36)
0 dr dr ~ 0 0 ~ 0
where we used that |R|2 = R∗ R. Integrating by parts the first term,
Z ∞ Z ∞ Z ∞
dR∗ dR 2 dR∗ dR 2
 
∗ d 2 dR dR ∞
R r dr = − r dr + R∗ r2 =− r dr (4.37)
0 dr dr 0 dr dr dr 0 0 dr dr
where it has been assumed that R∗ and dR
dr vanish sufficiently fast at ∞. Writing
R = RR + iRC ,
where RR and RC are real-valued, it comes
2 2
dR∗ dR
 
dRR dRC dRR dRC dRR dRC
=( −i )( +i )= + ,
dr dr dr dr dr dr dr dr
and we conclude that dr dr is real-valued. But from (4.36) and (4.37) we have
dR∗ dR
R ∞ dR∗ dR 2 2µ R ∞ 2 2
R∞ 2
dr dr r dr + ~2 0 V |R| r dr + `(` + 1) 0 |R| dr
E = 0
2µ ∞
R . (4.38)
~2 0
|R|2 r2 dr
Therefore, since all terms on the right-hand side are real, we conclude that E is real as well.
Students should notice that (4.38) gives an explicit expression for E in terms of (the integral of)
R and other data of the problem (although we shall derive a much more explicit expression for E,
see below).
Now that we know that E is real, let us show that it is negative1. Let us investigate the behavior
of (4.34) for large values of r, i.e., r  1. Then we can neglect the terms that contain 1r and (4.34)
gives, after expanding the terms in dr d
,
d2 R 2µE
2
≈ − 2 R. (4.39)
dr ~
1It is possible to obtain E < 0 by a more delicate analysis of (4.38), but here we employ a simpler argument.
 MATH 3120 - Intro to PDEs 13

But for r  1 we also have the approximation

d2 R dR d2 R
r + ≈ r ,
dr2 dr dr2
so that
d2 (rR) d2 R dR d2 R
= r + 2 ≈ r . (4.40)
dr2 dr2 dr dr2
Hence, multiplying (4.39) by r and using (4.40),
d2 (rR) 2µE
≈ − 2 (rR).
dr2 ~
This approximate equation can be easily solved, producing
√
−2µE
rR ≈ e± ~
r
.
If E ≥ 0, then R is a complex function which satisfies
|rR| ≈ 1 for r  1.
Then the integral
Z Z 2π Z π  Z ∞ 
|Ψ(t, x)|2 dx = |Y (φ, θ)|2 sin φ dφdθ |R(r)|2 r2 dr
R3 0 0 0

diverges since |R(r)|2 r2 ≈ 1 for large r. Consequently, condition (4.3) fails, and this does not
produce a physically sensible solution.
In light of the above arguments, we assume, once and for all, that E < 0. In this case, we can
define the real constants
2µE
β2 = − 2 , (4.41)
~
and
µZe2
γ= , (4.42)
4πε0 ~2 β
and make the real change of variables
% = 2βr.
With these definitions, equation (4.34), with V given by (4.35), becomes
   
1 d 2 dR 1 `(` + 1) γ
% + − − + R = 0. (4.43)
%2 d% d% 4 %2 %
Equation (4.43) will be solved using a power series expansion, but direct application of the method
does not work. To see this, try plugging
∞
X
R(%) = ak %k
k=0

into (4.43), obtaining

∞ ∞ ∞ ∞
X 1X X X
k(k + 1)ak %k−2 − ak %k − `(` + 1) ak %k−2 + γ ak %k−1 = 0.
4
k=0 k=0 k=0 k=0
This can be rewritten as
−`(` + 1)a0 %−2 + ((2 − `(` + 1))a1 + γa0 ) %−1
14 Disconzi

∞  
1
(4.44)
X
+ ((k + 3)(k + 2) − `(` + 1)) ak+2 + γak+1 − ak %k = 0.
4
k=0
Vanishing of each term order by order implies that a0 = 0, then a1 = 0, and subsequently ak = 0
for any k, so R = 0. We need, therefore, to try a different approach.
We shall focus on the behavior of (4.43) when %  1, in which case the equation simplifies to
 
1 d 2 dR R
% ≈ . (4.45)
%2 d% d% 4
This (approximate) equation can be solved as follows. Look for a solution of the form eA% . Plugging
into the equation we find A = − 12 , i.e., e− 2 is a (approximate) solution of (4.45). This suggests2
%

looking for solutions of (4.43) in the form

%
R(%) = e− 2 G(%). (4.46)
Plugging (4.46) into (4.43) gives an equation for G,
d2 G
   
2 ∂G γ − 1 `(` + 1)
+ − 1 + − G = 0. (4.47)
d%2 % ∂% % %2
We seek a solution of the form
∞ ∞
(4.48)
X X
G(%) = %s ak %k = ak %k+s ,
k=0 k=0

where s is to be determined. The term has been included due to the %1 terms in the equation, as
%s
these may lead to singular terms that do not fit into a general recurrence relation, as it occurred in
(4.44). Notice that the traditional procedure is included in this approach by simply setting s = 0.
Plugging (4.48) into (4.47) gives, after some algebra,
(s(s + 1) − `(` + 1)) a0 %s−2
∞ 
X  (4.49)
+ ((s + k + 1)(s + k + 2) − `(` + 1))ak+1 − (s + k + 1 − γ)ak %s+k−1 = 0.
k=0

The vanishing of the term in %s−2 requires

s(s + 1) − `(` + 1) = 0,
which has roots s = ` and s = −(` + 1). This latter root is rejected on the basis that it does not
yield a finite solution when % → 0+ , i.e., G(%) blows up at the origin when s = −(` + 1) (recall that
` is non-negative).
Using s = `, one finds from (4.49) the following recurrence relation,
k+`+1−γ
ak+1 = ak . (4.50)
(k + ` + 1)(k + ` + 2) − `(` + 1)
From (4.50) and the ratio test, we see at once that (4.48), with s = `, converges for all values of %.
In order for (4.48) to be an acceptable solution, we also must verify (4.3). From (4.50), it follows
that
k + ··· 1 + ···
ak+1 = 2 ak = ak ,
k + ··· k + ···
2The reader may remember that when one solves second order ODEs with constant coefficients, sometimes we
have to multiply a solution by a suitable power of the variable in order to produce a particular solution or a second
linearly independent solution. What it is being done here resembles that: we have some information about solutions,
% %
i.e., that e− 2 solves the equation (in an approximate sense) for large values of %. Thus, we try multiplying by e− 2
to construct the full, exact solution.
 MATH 3120 - Intro to PDEs 15

and
k − 1 + ··· 1 + ···
ak = 2
ak−1 = ak−1 ,
(k − 1) + · · · (k − 1) + · · ·
so that
1 + ··· 1 + ··· 1 + ···
ak+1 = ak = ak−1
k + ··· k + · · · (k − 1) + · · ·
1 + ···
= ak−1 .
k(k − 1) + · · ·
Continuing this way,
1 + ···
ak+1 = ak−j .
k(k − 1)(k − 2) · · · (k − j) + · · ·
Remembering that
∞
X 1 k
e% = % ,
k!
k=0
we see that G(%) is asymptotic to %s e% , i.e., its series expansion behaves very much like the series
of %` e% (recall that s = `):
G(%) ∼ %` e% ,
which implies, upon recalling (4.46),
% % %
R(%) = e− 2 G(%) ∼ e− 2 %` e% = %` e 2 ,
which diverges when % → ∞. As a consequence, (4.3) is not satisfied. This will be the case, unless
the series (4.48) terminates, i.e., unless ak = 0 for all k greater than a certain n. From (4.50), this
means
k + ` + 1 − γ = 0,
i.e.,
γ = k + ` + 1.
In particular, γ has to be an integer,
γ = n, n = ` + 1, ` + 2, . . . .
With this, the series terminates at the (n − (` + 1))th term, and G is a polynomial of degree n − 1.
Recalling (4.41) and (4.42), we have found the possible values for the separation constant E = En ,
namely,
µZ 2 e4
En = − , n = 1, 2, 3, . . . (4.51)
2(4πε0 )2 ~2 n2
We write Rn` to indicate that R is parametrized by the integers n and `, with n = `, ` + 1, . . . . We
can now write, for each n, the corresponding Rn` by using (4.50) to find the polynomial G = Gn` ,
and then Rn` via (4.46). Unwrapping all our definitions,
Zr `
   
Zr
− nα Zr
Rn` (r) = e 0 Gn` ,
nα0 nα0
where
4πε0 ~2
α0 = .
µe2
16 Disconzi

In light of (4.10), we see that ψ is also parametrized by n, `, and m` . Instead of thinking of n varying
according to n = `, ` + 1, . . . , we can equivalently think of ` as constrained by ` = 0, 1, . . . , n − 1,
for each given n = 1, 2, . . . , what is more convenient in order to organize the parameters n, `, m` .
We obtain, therefore, a family of solutions to (4.6),
ψn`m` = Rn` Y`m` , (4.52)
where
n = 1, 2, 3, . . . ,
` = 0, 1, 2, . . . , n − 1, (4.53)
m` = −`, −` + 1, . . . , 0, . . . , ` − 1, `.
Our final solution is then given, in view of (4.4) and (4.7), by
iEn
Ψ(t, x) = An`m` e− ~
t
ψn`m` (x),
where n, `, and m` satisfy (4.53), En and ψn`m` are given by (4.51) and (4.52), respectively, and
An`m` is a constant (depending on n, `, and m` ) that ensures (4.2), i.e., An`m` is given by
Z − 1
2
2
An`m` = |ψn`m` | .
R3

The term e− ~ t does not contribute to |Ψ|2 (since (e− ~ t )∗ (e− ~ t ) = (e+ ) = 1). It
iEn iEn iEn iEn iEn
~
t
)(e− ~
t

is customary to absorb the constant An`m` into ψn`m` , in which case

Z
|ψn`m` |2 = 1.
R3

Or course, (4.2) is automatically satisfied in this case.

4.3. Final comments. We close with some remarks about the physical meaning of the problem
we just described. Readers are referred to [4] for a more thorough physical discussion. Below, we
list some the first few ψn`m` .
n ` m` ψn`m`
 3
− Zr
1 0 0
2
ψ100 = √1π aZ0 e a0
 3   Zr
2 0 0
2 − 2a
ψ200 = 4√12π aZ0 2 − Zra0 e
0

 3 Zr
2 1 0
2 Zr −
ψ210 = 4√12π aZ0 a0 e
2a0 cos φ

 3 Zr
2 1 ±1
2 Zr −
ψ21±1 = 8√12π aZ0 2a0 sin φe±iθ
a0 e
 3   Zr
3 0 0
2 Z 2 r2 −
ψ300 = 81√1 3π aZ0 27 − 18 Zr a0 + 2 z02
e 3a0
√  3   Zr
3 1 0 Zr − 3a0
2
ψ310 = 81√2π aZ0 6 − Zr a0 a0 e cos φ
√   3   Zr
3 1 ±1 Zr − 3a0
2
ψ31±1 = 81√2π aZ0 6 − Zr a0 a0 e cos φe±iθ
  3 2 2 Zr
3 2 0
2 Z r −
ψ320 = 81√1 6π aZ0 a20
e 3a0 (3 cos2 φ − 1)
  3 2 2 Zr
3 2 ±1
2 Z r −
ψ32±1 = 811√π aZ0 a20
e 3a0 sin φ cos φe±iθ
  3 2 2 Zr
3 2 ±2
2 Z r −
ψ32±2 = 2621√π aZ0 a2
e 3a0 sin2 φe±2iθ
0
 MATH 3120 - Intro to PDEs 17

It is possible to show that the constants En , `(`+1), and m` have important physical interpretation:
En corresponds to the electron energy, `(` + 1) to the magnitude of its orbital angular momentum,
and m` to the projection of the orbital angular momentum onto the z-axis. The reader should notice
that these quantities cannot be arbitrary, being allowed to take values only on a countable set of
multiples of integers. This is a distinctive feature of Quantum Mechanics (we say that the energy
and orbital angular momentum are “quantized”). The indices n, `, and m` are called quantum
numbers.
One-electron atoms with ` = 0, 1, 2, 3 are labeled s, p, d, f . In hydrogen and hydrogen-like atoms,
this letter is preceded by a number giving the energy level n. Thus, the lowest energy state of the
hydrogen atom is 1s; the next to the lowest are 2s and 2p; the next 3s, 3p, 3d and so on. These
are the so-called “atomic orbitals” that the student is likely to have learned in Chemistry. Remem-
bering that |Ψ|2 is a probability density, what these orbitals represent are “clouds of probability,”
highlighting the regions of three-dimensional space where it is more likely to find the electron. A
few illustrations of the atomic orbitals are given in Figures 1, 2, and 3. These figures were generated
with the Mathematica package Visualizing Atomic Orbitals that can be found at
http://demonstrations.wolfram.com/VisualizingAtomicOrbitals/

Figure 1. An illustration of the Figure 2. An illustration of the

orbital 1s (n = 1, ` = 0, m` = 0). orbital 2p (n = 2, ` = 1, m` = 0).

Figure 3. An illustration of the orbital 3d (n = 3, ` = 2, m` = 0).

We finish mentioning that in a more detailed treatment of the problem, µ is not exactly the
mass of the particle being described, but rather the reduced mass of the system. This is because,
strictly speaking, the electron does not orbit the nucleus, but both orbit the center of mass of the
system electron–nucleus. This is very much like the situation of the Earth orbiting the Sun: both
bodies move due to their reciprocal gravitational attraction, although the Sun, begin much more
massive, barely feels the pull caused by Earth’s gravitational field, and that is why one usually
thinks of the Earth orbiting an standing-still Sun. A similar situation occurs for the nucleus and
the electron. We remark, however, that the calculations we presented apply, with no change, to
this more accurate situation: we only have to change the value of µ to be the reduced mass.
18 Disconzi

5. Separation of variables for the one-dimensional wave equation

Consider the wave equation in one dimension:
utt − c2 uxx = 0. (c 6= 0)
Notation 5.1. Whenever a PDE involves the time variable, by the dimension we always mean the
spatial dimension. E.g., the one-dimensional wave equation (abbreviated 1d wave equation) is the
wave equation for u = u(t, x) with x ∈ R.
We are interested in the case when the spatial variable belongs to a compact interval, e.g., 0 ≤
x ≤ L, for some L ≥ 0, and u vanishes at the extremities of the interval, i.e., u(t, 0) = 0 = u(t, L).
This is the situation describing a string that can vibrate in the vertical direction with its ends fixed,
with u(t, x) representing the string amplitude at x at time t:

u(t, x)

x=0 x=L

The conditions u(t, 0) = 0 and u(t, L) = 0 are called boundary conditions because they are
conditions imposed on the solution on the boundary of the domain where it is defined. Thus, the
problem can be stated as
= 0 in (0, ∞) × (0, L)

 utt − c2 uxx
u(t, 0) = 0 (5.1)
u(t, L) = 0


where the wave equation is defined for all t ∈ (0, ∞) and x ∈ (0, L). This is called a boundary
value problem (BVP) because it consist of a PDE plus boundary conditions. Sometimes we refer
to a boundary value problem simply as PDE. In the HW, you will be asked to show that applying
separation of variables we obtain the following family of solutions:
  nπc   nπc   nπ 
un (t, x) = an cos t + bn sin t sin x
L L L
where n = 1, 2, 3, ... and an and bn are arbitrary constants. Since the equations is linear, sums fo
the above function are solutions, i.e.,
XN XN   nπc   nπc   nπ 
un (t, x) = an cos t + bn sin t sin x
L L L
n=1 n=1
is also a solution.
Because of this hold for any N , we should be able to sum all the way to infinity and still get a
solution. In other word, the most general solution to the above boundary value problem is
∞   nπc   nπc   nπ 
(5.2)
X
u(t, x) = an cos t + bn sin t sin x
L L L
n=1
provided that this expression makes sense, i.e., the series converges.
Terminology. It often happens in PDEs that we have situations as the above, i.e., we have a
formula for a would-be solution, but we do not know if the formula is in fact well-defined (e.g.,
we have a series that might not converge, or a function that might not be differentiable, etc.).
“Solutions” of this type are called formal solutions. In other words, a formal solution is a candidate
for a solution, but extra work must be done or further assumptions made in order to show that
they are in fact solutions.
 MATH 3120 - Intro to PDEs 19

The convergence of the above series cannot be decided without further information about the
problem. This is because, as stated, the coefficients an and bn in the formal solution are arbitrary,
and it is not difficult to see that we can make different choices of these coefficients in order to make
the series converge or diverge.
Therefore, we consider the above BVP (5.1) supplemented by initial conditions, i.e., we assumed
given functions g and h defined on [0, L] and look for a solution u such that

u(0, x) = g(x), ∂t u(0, x) = h(x), 0 ≤ x ≤ L.

Similarly to what happens in ODEs, we expect that once initial conditions are given, we will no
longer obtain a general solution but rather the unique solutions that satisfies the initial conditions.

Remark 5.2. Note that any multiple of the (formal) solution u will also be a (formal) solution.
This is encoded in the arbitrariness of an and bn , since if we multiple u by a constant A, we can
simply redefine new coefficients as ãn = Aan , b̃n = Abn . This freedom, however, is not present
once we consider initial conditions, since if u(0, x) = g(x), ∂t u(0, x) = h(x) then Au(0, x) 6= g(x),
A∂t u(0, x) 6= h(x) (unless A = 1).

This pervious remark suggests that the coefficients an and bn should be determined from the
initial condition. Before investigating this, let us state the full problem. We want to find u such
that
utt − c2 uxx = 0 in (0, ∞) × (0, L), c > 0


u(t, 0) = u(t, L) = 0,

t≥0

(5.3)
 u(0, x) = g(x), 0≤x≤L
∂t u(0, x) = h(x), 0 ≤ x ≤ L,



The above problem is called an initial-boundary value problem (IBVP) since it is a PDE with
boundary conditions and initial conditions provided, although we sometimes call it simply a PDE.
Two initial conditions are prescribed, i.e., u(0, x) and ∂t u(0, x), because the wave equation is
second order in time. Note that g and h have to satisfy the following compatibility conditions:

g(0) = g(L) = h(0) = h(L) = 0. (5.4)

We have already derived a formal solution to the wave equation satisfying the boundary condi-
tions (5.2). It remains to investigate the initial conditions. Plugging t = 0:
∞
X  nπ 
u(0, x) = g(x) = an sin x .
L
n=1

Differentiating u with respect to t and plugging t = 0:

∞
X nπc  nπ 
∂t u(0, x) = h(x) = bn sin x .
L L
n=1

Since g and h are in principle arbitrary, the above is essential asking whether it is possible to write
an arbitrary function on [0, L] as a series of sine functions with suitable coefficients. Or, rephrasing
the equation in a more appropriate form, we are asking: what are the functions on [0, L] that can be
written as a convergent series of sine functions with suitable coefficients? The functions for which
this is true will provide us with a class of functions for which the above initial-boundary problem
(IBVP) admits a solution.
The subject that investigates questions of this type is known as Fourier series. We will now make
a digression to study Fourier series. After that, we will return to the wave equation.
20 Disconzi

6. Fourier Series
We begin with the definition of Fourier series.
Definition 6.1. Let I = (−L, L) or [−L, L], L > 0, and f : I → R be integrable on I. The Fourier
series of f , denoted F.S.{f }, is the series
∞
a0 X   nπx   nπx 
F.S.{f }(x) := + an cos + bn sin , (6.1)
2 L L
n=1
where the coefficients an and bn are given by
1 L
Z  nπx 
an = f (x) cos dx, n = 0, 1, 2, ... (6.2)
L −L L
1 L
Z  nπx 
bn = f (x) sin dx, n = 1, 2, 3, ... (6.3)
L −L L
The coefficients an and bn are called Fourier coefficients.
Remark 6.2.
(1) F.S.{f } is a series constructed out of f . We are not claiming that F.S.{f } = f . In fact, at
this point we are not evening claiming that F.S.{f } converges (although we want to find
conditions for which it converges, and for which F.S.{f } = f ).
(2) The Fourier coefficients are well defined in view of the integrability of f .
(3) We introduced Fourier series for functions defined on an interval [−L, L]. This set-up is
slightly different than what we encountered above for the wave equation, where we worked
on the interval [0, L]. We will relate Fourier series on [−L, L] with functions defined on
[0, L] later on.
(4) The Fourier series is a series of sine and cosine. The situation discussed above in the wave
equation is a particular case where only sine is present (i.e., an = 0).
Example 6.3. Find the Fourier series of

−1, −π ≤ x < 0
f (x) =
1, 0 ≤ x ≤ π.
We compute:
1 π
Z
an = f (x) cos(nx)dx = 0 (even-odd functions)
π −π
Z π
2 π
Z
1
bn = f (x) sin(nx)dx = f (x) sin(nx)dx
π −π π 0 |{z}
=1
π
2 1 (−1)n
  
2 cos(nx)
= − = −
π n 0 π n n
0 n even

=
nπ n odd.
4

Thus:
∞ 
2 X 1 − (−1)n

F.S.{f }(x) = sin(nx)
π n
n=1
 
4 1 1
= sin(x) + sin(3x) + sin(5x) + ... .
π 3 5
Note that f (0) = 1 but F.S.{f }(0) = 0, so F.S.{f } 6= f .
 MATH 3120 - Intro to PDEs 21

Example 6.4. Find the Fourier series of f (x) = |x|, −1 ≤ x ≤ 1. Compute:

Z 1 Z 1
a0 = f (x)dx = 2 xdx = 1,
−1 0
Z 1 Z 1
2
an = f (x) cos(nπx)dx = 2 x cos(nπx)dx = ((−1)n − 1), n∈N
−1 0 π 2 n2
Z 1
bn = f (x) sin(nπx)dx = 0 (even-odd).
−1
Thus
∞
1 X 2
F.S.{f }(x) = + ((−1)n − 1) cos(nπx)
2 n2 π 2
n=1
 
1 4 1 1
= − 2 cos(πx) + cos(3πx) + cos(5πx) + ... .
2 π 9 25
6.1. Piecewise Functions. We begin with some definitions.
Definition 6.5. Let I ⊂ R be an interval. A function f : I → R is called k-times continuously
differentiable if all its derivatives up to order k exists and are continuous. We denote by C k (I)
the space of all k-times continuously differentiable functions on I. Note that C 0 (I) is the space of
continuous functions on I. We denote by C ∞ (I) the space of infinitely many times differentiable
functions on I. Sometimes we say simply that “f is C k ” to mean that f ∈ C k (I). We write simply
C k for C k (I) if I is implicitly understood. C ∞ functions are also called smooth functions.
Example 6.6. ex ∈ C ∞ (R), |x| ∈ C 0 (R). The function f : R → R defined by
x sin x1 , x =
 2

6 0
f (x) =
0, x=0
is C 0 , it is differentiable, but it is not C 1 . This is because f 0 (x) exists for every x (including x = 0)
but f 0 is not continuous at x = 0.
Remark 6.7. Note that C k (I) ⊂ C ` (I) if k > ` and C ∞ (I) = ∞ k
T
k=0 C (I).

Definition 6.8. Let I ⊂ R be an interval. We say that f : I → R is a piecewise C k function if f

is C k except possibly at a countable number of isolated points.
Example 6.9.
(1) The functions |x| and

1, x ≥ 0
f (x) =
−1, x < 0
are piecewise smooth (C ∞ ) functions.
(2) Below is piecewise C ∞ function.

−2 −1 0 1 2

(3) The function f : [0, 1] → R given by

22 Disconzi

...

1 1 1
8 4 2 1

is not piecewise C k because the set of points where it fails to be C k are not isolated.
6.2. Convergence of Fourier Series.
Notation 6.10. We denote by f (x+ ) and f (x− ) the right and left values of f at x, defined by
f (x+ ) = lim f (x + h), f (x− ) = lim f (x + h).
h→0+ h→0−

If f is continuous at x, then f (x+ ) = f (x− ) = f (x), but otherwise these values might differ.
Example 6.11. Consider the unit step function,

1, x ≥ 0
H(t) =
−1, x < 0.
Then H(0+ ) = 1 and H(0− ) = −1.
Theorem 6.12. Let f be a piecewise C 1 function on [−L, L]. Then, for any x ∈ (−L, L):
1
F.S.{f }(x) = (f (x+ ) + f (x− )), (6.4)
2
and
1
F.S.{f }(±L) = (f (−L+ ) + f (L− )). (6.5)
2
In particular, F.S.{f } converges.
From the above theorem, we see that F.S.{f }(x) = f (x) when f is continuous at x. Thus, if f
is piecewise C 1 and C 0 , we have:
∞
a0 X   nπx   nπx 
f (x) = + an cos + bn sin . (6.6)
2 L L
n=1

Example 6.13. We graph


−1, −π ≤ x < 0
f (x) =
1, 0 ≤ x ≤ π

and F.S.{f }(x) below (note that f is piecewise C 1 )

Example 6.14. Since |x| is continuous and piecewise C 1 :
∞
1 X 2
|x| = + ((−1)n − 1) cos(nπx).
2 n2 π 2
n=1

Next, we consider the differentiation and integration of Fourier series term by term.
 MATH 3120 - Intro to PDEs 23

graph of f graph of F.S.{f }

f (0+ ) = 1
1 f (π − ) = 1 1

−π π −π π

f (−π + ) = −1 f (0− ) = −1
−1 −1

Theorem 6.15. Let f be a piecewise C 2 and continuous function on [−L, L], and assume that
f (−L) = f (L). Then, the Fourier series of f 0 can be obtained from that of f by differentiation
term-by-term. More precisely, writing
∞
a0 X   nπx   nπx 
f (x) = + an cos + bn sin ,
2 L L
n=1

we have
∞ 
X   nπx 0   nπx 0 
0
F.S.{f }(x) = an cos + bn sin
L L
n=1
∞   nπx 
X −an nπ  nπx  b nπx
n
= sin + cos .
L L L L
n=1

In particular, if f0 is continuous at x, we have

∞
X nπ   nπx   nπx 
f 0 (x) = −an sin + bn cos .
L L L
n=1

Example 6.16. To see that we cannot always differentiate a Fourier series term by term, consider
f (x) = x, −π ≤ x ≤ π. Its Fourier series is
∞
X (−1)n+1
F.S.{f }(x) = 2 sin(nx),
n
n=1

which converges for any x, but the term-by-term differentiated series, which is
∞
X
2 (−1)n+1 cos(nx)
n=1

diverges for every x.

Theorem 6.17. Let f be piecewise continuous on [−L, L] with Fourier series
∞ 
1  nπx   nπx 
(6.7)
X
F.S.{f }(x) = a0 + an cos + bn sin .
2 L L
n=1

Then, for any x ∈ [−L, L]:

Z x Z x ∞ Z x     
1 nπt nπt
(6.8)
X
f (t)dt = a0 dt + an cos + bn sin dt.
−L −L 2 L L
n=1 −L
24 Disconzi

6.3. Some Intuition Behind Fourier Series. Let us make some comments about the way the
Fourier series is defined. Given f defined on [−L, L], our goal is to write:
∞
a0 X   nπx   nπx 
f (x) = + an cos + bn sin
2 L L
n=1
Let us make an analogy with the following problem: given a vector v ∈ Rn , we want to write
Xn
v= ci ei ,
i=1
where {ei }ni=1 is an orthogonal basis of Rn . In other words, we have to find the coefficients ci . Since
the vectors ei are orthogonal
ei · ej = 0 if i 6= j,
where · is the dot product, a.k.a. inner product of vectors. Thus, for each j = 1, ..., n:
n
X v · ej
ej · v = ci ej · ei = cj ej · ej =⇒ cj = .
ej · ej
i=1
We want to do something similar to find the Fourier coefficients an and bn . Consider the functions
1  nπx   nπx 
E0 (x) = , En (x) = cos , Ẽn (x) = sin n = 1, 2, ....
2 L L
Then:
∞
X
f = a0 E0 + (an En + bn Ẽn ).
n=1

This is very similar to the case in Rn .

In fact, the space of piecewise C k is a vector space, so (6.3)
is an equality between vectors, although C k is an infinite dimensional vector space so we need a
basis with infinitely many vectors.
To find the Fourier coefficients the series same way we found the coefficients cj above, we need
the analogue of the dot product for functions. It cannot be the usual product of functions, since
the product of two functions is another function, whereas the dot product of two vectors is not
another vector but a number. We also want our “dot product” for functions to have all the standard
properties of the dot product of vectors. The relevant product for functions is defined below.
Definition 6.18. Let I ⊂ R be an interval. The L2 inner product, or simply inner product, of two
functions f, g : I → R is defined as
Z
hf, giL2 := f (x)g(x)dx
I
whenever the integral on the RHS is well-defined. We often write h·, ·i for h·, ·iL2 . The L2 norm, or
simply norm, of f : I → R is defined as
p
kf kL2 := hf, f i.
We sometimes write k·k for k·kL2 . We also write h·, ·iL2 (I) and k·kL2 (I) if we want to emphasize the
interval I.
It is a simple exercise to show that h·, ·iL2 has all the following properties, which are similar to
the properties of the dot product:
(1) hf, gi ∈ R when defined
(2) hf, gi = hg, f i
(3) hf, ag + bhi = ahf, gi + bhf, hi, a, b ∈ R, f, g, h functions
(4) hf, 0i = 0
 MATH 3120 - Intro to PDEs 25

(5) hf, f i ≥ 0. In particular, k·kL2 is a real number if hf, f i < ∞.

Remark 6.19. The dot product has the property v · v = 0 ⇒ v = 0. This is not true for h·, ·iL2 ,
as the example

1, x = 0
f (x) =
0, otherwise
shows. However, if f is continuous, then it is true that hf, f iL2 = 0 ⇒ f = 0.
Consider now I = [−L, L] and let us go back to (6.3). A simple computation shows that
hEn , Em i = 0 if n 6= m, hẼn , Ẽm i = 0 if n 6= m, hEn , Ẽm i = 0, hẼn , Ẽn i = L,
2, n = 0
 L
hEn , En i =
L, n > 0.
Taking the inner product of (6.3) with Em , Ẽm , m ≥ 1, and E0 , gives:
∞
X
hf, Em i = a0 hE0 , Em i + (an hEn , Em i + bn hẼn , Em i)
n=1
hf, Em i
= am hEm , Em i = am L ⇒ am =
L
∞
X
hf, E0 i = a0 hE0 , E0 i + (an hEn , E0 i + bn hẼn , E0 i)
n=1
L 2
= a0 hE0 , E0 i = a0 ⇒ a0 = hf, E0 i
2 L
∞
X
hf, Ẽm i = a0 hE0 , Ẽm i + (an hEn , Ẽm i + bn hẼn , Ẽm i)
n=1
hf, Ẽm i
= bm hẼm , Ẽm i = bm L ⇒ bm =
L
Writing explicitly h·, ·i in terms of an integral and using the definitions of En , Ẽn , we see that the
expressions we found for an , bn are exactly the Fourier coefficients.
6.4. The Fourier series of series of periodic functions, and the Fourier series of functions
on [0, L]. Suppose that f is defined on R and has period 2L, i.e., f (x) = f (x + 2L) for all x. Thus,
all information about f is determined by its values on [−L, L]. We can defined the Fourier series
for f as a function on [−L, L], and all the previous results are immediately adapted to this case.
Moreover, given a function (−L, L), we can extend it to a periodic function on R and consider its
Fourier series (note, however, that this extension is not unique). This is illustrated in the picture
below:

−L L −2L −L L 2L

Consider now a function f defined on [0, L]. We define its cosine Fourier series by
∞  
a0 X nπL
cos
F.S. {f }(x) = + an cos , x ∈ [0, L], (6.9)
2 x
n=1
26 Disconzi

where
2 L
Z  nπx 
an = f (x) cos dx.
L 0 L
Extend f to an even function on [−L, L] by

f (x), 0 ≤ x ≤ L,
˜
f (x) = (6.10)
f (−x), −L ≤ x < 0.
The Fourier coefficients of f˜ are
1 L ˜ 2 L
Z  nπx  Z  nπx 
ãn = f (x) cos dx = f (x) cos dx = an , (6.11)
L −L L L 0 L
1 L ˜
Z  nπx 
b̃n = f (x) sin dx = 0, (6.12)
L −L L
where we use that f˜ is even. Thus, for x ∈ [0, L]
F.S.{f˜}(x) = F.S.cos {f }(x). (6.13)
In other words, the cosine Fourier series of f : [0, L] → R equals the restriction to [0, L] of the
Fourier series of the even extension of f .
Similarly, we define the sine Fourier series of f : [0, L] → R by
∞  nπx 
(6.14)
X
sin
F.S. {f }(x) = bn sin
L
n=1
where Z L
2  nπx 
bn = f (x) sin dx.
L 0 L
Letting f˜ be an odd extension of f ,

f (x), 0 ≤ x ≤ L,
f˜(x) = (6.15)
−f (−x), −L ≤ x < 0.
We find the Fourier coefficients of f˜ to be
1 L ˜
Z  nπx 
ãn = f (x) cos dx = 0, (6.16)
L −L L
1 L ˜ 2 L
Z  nπx  Z  nπx 
b̃n = f (x) sin dx = f (x) sin dx = bn , (6.17)
L −L L L 0 L
thus F.S.{f˜}(x) = F.S.sin {f }(x), x ∈ [0, L]. In other words, the sine Fourier series of f : [0, L] → R
equals the restriction to [0, L] of the Fourier series of the odd extension of f .
We conclude that the theorems on convergence, differentiation, and integration of Fourier series
are immediately applicable to the sine and cosine Fourier series.

6.5. Back to the wave equation. We are now ready to discuss the problem
= 0 in (0, ∞) × (0, L), c > 0

 utt − c2 uxx
= 0,

u(t, 0) = u(t, L) t≥0

(6.18)
 u(0, x) = g(x), 0≤x≤L
=

∂t u(0, x) h(x), 0 ≤ x ≤ L,


where g and h are given functions satisfying the compatibility conditions

g(0) = g(L) = 0 = h(0) = h(L).
 MATH 3120 - Intro to PDEs 27

We saw that a formed solution to this problem is given by:

∞     
nπct nπct  nπx 
(6.19)
X
u(t, x) = an cos + bn sin sin
L L L
n=1

where an and bn are to be determined by

∞
X  nπx 
g(x) = an sin ,
L
n=1

and
∞
X nπc  nπx 
h(x) = bn sin .
L L
n=1

The last two expressions mean that g and h equal their sine Fourier series, with Fourier coefficients
given by an and nπcL bn , respectively. These equalities will in fact be true if we make suitable
assumptions on g and h. Let us assume that g and h are C 2 functions. Then, from the previous
theorems for Fourier series, we know that g and h equal their sine Fourier series, and the coefficients
an and bn are given by
2 L
Z  nπx  Z L
2  nπx 
an = g(x) sin dx, bn = h(x) sin dx. (6.20)
L 0 L nπc 0 L
Our assumptions on g and h allow us to compute the coefficients an and bn . We will have to
develop a few more tools before we are able to show that (6.19) is in fact a solutions. However, we
summarize the result here; its proof will be postponed (in fact, it will be assigned as the HW after
more background is developed).
Theorem 6.20. Consider the problem (5.3) and assume that g and h are C 2 functions such that
g(0) = g(L) = 0 = h(0) = h(L),
g (0) = g 00 (L) = 0 = h00 (0) = h00 (L).
00

Then a solution to (5.3) is given by (6.19), where an and bn are given by (6.20).
Remark 6.21. We will explain the assumptions involving second derivatives of g and h when we
prove this theorem.

7. The 1d wave equation in R

We now consider the problem for u = u(t, x):
= 0 in (0, ∞) × (∞, ∞), c > 0

 utt − c2 uxx
u(0, x) = u0 (x), −∞ ≤ x ≤ ∞ (7.1)
∂t u(0, x) = u1 (x), −∞ ≤ x ≤ ∞,


This is an initial-value problem for the wave equation (IVP). Compared to the initial-boundary
value problem we studied earlier, we see that now x ∈ R, so there are no boundary conditions. This
initial-value problem is also known as the Cauchy problem for the wave equation, a terminology
that we will explain in more detail later on. We refer to the functions u0 and u1 as (initial) data
for the Cauchy problem. A solution to this Cauchy problem is a function that satisfies the wave
equation and the initial conditions.
We had defined the spaces C k (I) for an interval I ⊂ R. For functions of two variables, we can
similarly define C k (R2 ), which we will use here. We will define general C k spaces for functions
several variables later on.
28 Disconzi

Proposition 7.1. Let u ∈ C 2 (R2 ) be a solution to the 1d wave equation. Then, there exists
functions F, G ∈ C 2 (R) such that
u(t, x) = F (x + ct) + G(x − ct).
Proof. Set α := x + ct, β := x − ct, so that t = 1
2c (α − β), x = 12 (α + β), and
v(α, β) := u(t(α, β), x(α, β)).
Then, from u(t, x) = v(α(t, x), β(t, x)) we find
ut = vα αt + vβ βt = cvα − cvβ ,
utt = cvαα αt + cvαβ βt − cvβα αt − cvββ βt
= c2 vαα − c2 vαβ − c2 vβα + c2 vββ ,
ux = vα αx + vβ βx = vα + vβ ,
uxx = vαα αx + vαβ βx + vβα αx + vββ βx
= vαα + vαβ + vβα + vββ .
Thus, 0 = utt − c2 uxx = −4c2 vαβ , where used that vαβ = vβα since v is C 2 (because u is C 2 and
the change of coordinates (t, x) 7→ (α, β) is C ∞ ). Thus, in (α, β) coordinates the wave equation
reads: vαβ = 0.
Therefore, (vα )β = 0 implies that vα is a function of α only: vα (α, β) = f (α) for some C 1
function f . Integrating w.r.t. α gives
Z
v(α, β) = f (α)dα + G(β),

for some function G. Note that F := f (α)dα is C 2 , thus so is G. Therefore, v(α, β) = F (α)+G(β),
R
and in (t, x) coordinates: u(t, x) = F (x + ct) + G(x − ct). 
The above formula has a clear physical interpretation. At t = 0, u(0, x) = F (x) + G(x). For
each t > 0, the graph of G(x − ct) is the graph of G(x) moved of units to the right, so the graph
of G(x) is moving to the right with speed c. G(x − ct) is called a forward wave. Similarly, the
graph of F (x) is moving to the left and F (x + ct) is called a backward wave. The general solution
is thus a sum (or a superposition) of a forward and backward wave, and we see that the constant
c is indeed the speed of propagation of the wave.

G(x) G(x − ct)

Notation 7.2. Having found the interpretation of the constant c, we will often set c = 1.
Proposition 7.3. Let u ∈ C 2 ([0, ∞) × R) be a solution to the Cauchy problem for the 1d wave
equation with data u0 , u1 . Then
u0 (t + x) + u0 (x − t) 1 x+t
Z
u(t, x) = + u1 (y)dy. (7.2)
2 2 x−t
This formula is known as D’Alembert’s formula.
 MATH 3120 - Intro to PDEs 29

Proof. Note that u0 ∈ C 2 , u1 ∈ C 1 . From

u(t, x) = F (x + t) + G(x − t)

in the pervious result, we have

u(0, x) = F (x) + G(x) = u0 (x),

ut (0, x) = F 0 (x) − G0 (x) = u1 (x).

Integrating this last equality:

Z x
F (x) − G(x) = u1 (y)dy + C
|{z} ,
0
F (0)−G(0)

adding to u(0, x) :
Z x
1 1 C
F (x) = u0 (x) + u1 (y)dy + .
2 2 0 2

Plugging back into u(0, x) :

Z x
1 1 C
G(x) = u0 (x) − u1 (y)dy − .
2 2 0 2

Replacing x 7→ x + t in F and x 7→ x − t in G and adding gives the formula. 

The last two propositions derived formulas for C 2 solutions of the wave equation given such a
solution. The next result shows that solutions actually exist.

Theorem 7.4 (∃! of 1d wave). Let u0 ∈ C 2 (R) and u1 ∈ C 1 (R). Then there exists a unique
u ∈ C 2 ([0, ∞) × R) that solves the Cauchy problem for the 1d wave equation with data u0 , u1 .
Moreover, u is given by D’Alembert’s formula.

Proof. Given two C 2 ([0, ∞) × R) solutions, both satisfy D’Alembert’s formula (with the same
u0 , u1 ) thus they are equal, establishing uniqueness. To prove existence, defined u by D’Alembert’s
formula. Then u ∈ C 2 ([0, ∞) × R) since u0 ∈ C 2 and u1 ∈ C 1 , and by construction u satisfies the
wave equation and the initial conditions. 

Definition 7.5. The lines x + t = constant and x − t = constant in the (t, x) plane (or x + ct =
constant, x − ct = constant for c 6= 1) are called the characteristics (or characteristic curves) of
the wave equation. They (and their generalizations to higher dimensions) are very important to
understand solutions to the wave equation, as we will see.

7.1. Regions of influence for the 1d wave equation. Suppose u1 = 0 and u0 (x) = 0 for
x∈/ [a, b]. Since u0 (x + t) and u0 (x − t) are constant along the lines x + t = constant and x − t =
constant, respectively, we see that u(t, x) 6= 0 only possibly for points (t, x) that lie int he region
determined by the region lying between the characteristics emanating from a and b as indicated in
the figure:
30 Disconzi

x+t=a x+t=b x−t=a x−t=b

u(t, x) = 0
b<x+t
x−t<a

a≤x+t≤b a≤x−t≤b

x−t≤x+t<a b<x−t≤x+t
u(t, x) = 0 u(t, x) = 0
x
a b

Notation 7.6. Although we ordered the coordinates as (t, x), we will often draw the (t, x) plane
with the x-axis on the horizontal.

Suppose now that u0 = 0 and that u1 (x) = 0 for x ∈ / [a, b]. Then x−t u1 (y)dy = 0 whenever we
R x+t

have [x − t, x + t] ∩ [a, b] =, i.e., if x + t < a or x − t > b. Therefore, u(t, x) 6= 0 possibly only in the
region {x + t ≥ a} ∩ {x − t ≤ b}, as depicted in the figure

x+t=a x−t=b
x+t≥a
and
x−t≤b

x−t≤x+t<a b<x−t≤x+t
u(t, x) = 0 u(t, x) = 0

a b x

For general u0 and u1 , we can therefore precisely track how the values of u(t, x) are influenced
by the values of the initial conditions. It follows that the values of the data on an interval [a, b] can
only affect the values of u(t, x) for (t, x) ∈ {x + t ≥ a} ∩ {x − t ≤ b}. This reflects the fact that
waves travel at a finite speed. The regions (t, x) ∈ {x + t ≥ a} ∩ {x − t ≤ b} is called the domain
of influence of [a, b]. Consider now a point (t0 , x0 ) and u(t0 , x0 ). Let D be the triangle with vertex
(t0 , x0 ) determined by x + t = x0 + t0 , x − t = x0 − t0 , and t = 0:
 MATH 3120 - Intro to PDEs 31

(t0 , x0 )

x − t = x0 − t 0 x + t = x0 + t 0

x
x = x0 − t 0 x = x0 + t 0

Then
x0 +t0
u0 (x0 + t0 ) + u0 (x0 − t0 ) 1
Z
u(t0 , x0 ) = + u1 (y)dy.
2 2 x0 −t0
and we see that u(t0 , x0 ) is completely determined by the values of the initial data on the interval
[x0 − t0 , x0 + t0 ]. The region D is called the (past) domain of dependence of (t0 , x0 ).

7.2. Generalized solutions. Note that the RHS of D’Alembert’s formula (7.2) makes sense when
u0 and u1 are piecewise functions. This motivates the following definition.
Definition 7.7. Let u0 be a piecewise C 2 function and u1 a piecewise C 1 function. Then u given
by D’Alembert’s formula is called a generalized solution to the wave equation. If u0 and u1 are
C 2 and C 1 functions, respectively, then u is called a classical solution. When u is a generalized
solution, the points where u fails to be C 2 are called singularities of the solution (sometimes we
abuse language and say singularities of the wave equation).
To understand what is going on, consider the case when for fixed t0 . u is C 2 except at the point
(t0 , x0 ). Writing u(t, x) = F (x + t) + G(x − t), we see that F is not C 2 at x0 + t0 and/or G is not C 2
at x0 − t0 . The two characteristics passing through (t0 , x0 ) are x + t = x0 + t0 and x − t = x0 − t0 .
Thus, for any fixed t1 , u(t1 , x) fails to be C 2 except at one or two points, namely, x± such that
x+ + t1 = x0 + t0 , x− − t1 x = x0 − t0 .

(t1 , x− ) (t1 , x+ ) t = t0

(t0 , x0 )

x − t = x0 − t 0 x + t = x0 + t 0

x
32 Disconzi

This shows that the singularities of the wave equation remain localized in space and travel along
the characteristics.
We will see that the results we obtained for the 1d wave equation (existence and uniqueness for
the Cauchy problem, existence of domains of influence/dependence, propagation of singularities
along characteristics) hold for the wave equation in higher dimensions and, in fact, for a class of
equations called hyperbolic, of which the wave equation is the prototypical example.

8. Some general tools, definitions, and conventions for the study of PDEs
In order to advance further our study of PDEs, in particular to study PDEs in Rn , we will recall
a few tools from multivariable calculus and introduce some convenient notation/terminology.

8.1. Domains and boundaries.

Definition 8.1. A domain in Rn is an open connected subset of Rn . If Ω ⊂ Rn is a domain, we

denote by Ω̄ its closure in Rn . The boundary of a domain Ω, denoted ∂Ω, is the set ∂Ω := Ω̄ \ Ω.
We say that a boundary ∂Ω has regularity C k or is a C k boundary if it can be written locally as
the graph of a C k function.

Notation 8.2. We note by |x| the Euclidean norm of an element x ∈ Rn . Ω and ∂Ω will always
denote a domain and its boundary, unless stated otherwise.

Example 8.3. B n := {x ∈ Rn | |x| < 1} is a domain in Rn . Its boundary is the n − 1 dimensional

sphere: S n−1 := ∂B n = {x ∈ Rn | |x| = 1}. It is not difficult to see that S n−1 is C ∞ , i.e., B n has a
C ∞ boundary. For example, the upper cap of S n−1 , given by S n−1 ∩ {xn > 0}, is the graph of the
function f : B n−1 ⊂ Rn−1 → R given by

p
f (x1 , ..., xn−1 ) = 1 − (x1 )2 − ... − (xn−1 )2 ,

which is C ∞ .

Notation 8.4. When talking about maps between subsets of Rn and Rm , we will often write
f : U ⊂ Rn → Rm , where it is implicitly understood that the domain U of f is an open set (unless
said otherwise).

Recall that if f : U ⊂ Rn → R is C 1 , for each x ∈ U the graph of f at (x, f (x)) admits a

tangent plane. Thus, if ∂Ω is C 1 , for each x ∈ ∂Ω there exists a tangent plan to ∂Ω at x, denoted
Tx ∂Ω. The unit outer normal to ∂Ω at x is by definition the unit normal to Tx ∂Ω that points to
the exterior of Ω. The collection of the unit outer normals Nx as x varies over ∂Ω forms a vector
field over ∂Ω which is called the unit outer normal vector field. We sometimes refer simply to “the
unit outer normal” when the context makes it clear whether we are talking about the vector field
as a specific vector field.
 MATH 3120 - Intro to PDEs 33

unit outer normal

x
Tx ∂Ω

Remark 8.5. Above, we took for granted that students recall (or have seen) the definition of a
connected set in Rn . Intuitively, a set is connected if is not “split into separate parts:”

Ω
Ω

Connected Not connected

For the time being, this intuitive notion will suffice for students who have not see the precise
definition. The mathematical definition of connectedness will be given later on.
8.2. The Kronecker delta.
Definition 8.6. The Kronecker delta symbol in n dimensions or simply the Kronecker delta when
the dimension is implicitly understood, is defined as the collection of numbers {δij }ni,j=1 such that
δi,j = 1 if i = j and δij = 0 if i 6= j. We identify the Kronecker delta with the entries of the
n × n identity matrix in the standard coordinates. We also define δ ij := δij , which we also call the
Kronecker delta and identity with the entries of the identity matrix.
Recall that the Euclidean inner product, a.k.a. the dot product, of vectors in Rn is the map:
h·, ·i : Rn × Rn → R
given in standard coordinates by:
n
X
hX, Y i = X iY i,
i=1
which is also denoted by X · Y . We can write hX, Y i as (recall over sum convention):
hX, Y i = δij X i Y j
In view of this last formula, we also identify the Kronecker delta with the Euclidean inner product.
34 Disconzi

8.3. Raising and lowering indices with δ. Given a vector X = (X 1 , ..., X n ), we define
Xi := δij X j , i = 1, ..., n.
We say that we are lowering the indices of X and identify the n-tuple (X1 , X2 , ..., Xn ) with the
vector X itself.
The point of introducing Xi is to achieve consistency with our convention of summing indices
that appear once up and once down. For example, if we want the inner product
n
X
hX, Y i = X iY i
i=1

using our sum convention (thus avoiding to write i=1 ), one of the indices i needs to be downstairs:
Pn

hX, Y i = X i Yi ,
so that we had to break with our convention that vectors have indices upstairs. However, if we now
interpret Yi as lowering the indices of Y , then
hX, Y i = δij X i Y j = X i δij Y j = X i Yi .
Similarly, recall that we wrote
curl i X = ijk ∂j Xk ,
where we had artificially written Xk with an index downstairs, thus breaking with our convention
that vectors had an index upstairs. But now we have a proper way of thinking of Xk as δkj X j .
Note that using δij we could completely avoid writing vectors with indices downstairs, i.e., every
time that Xi appears in a forumla we can replace it with δij X j . E.g.,
curl i X = ijk δk` ∂j X ` .
But the point is precisely to have a compact notation, so δk` ∂X ` = ∂j δk` X ` = ∂j Xk .
Remark 8.7. In the above computations, note that we can move δk` pass the derivative because
δk` is constant for each fixed k and `, i.e., δk` is not a function of the coordinates.
We extend the lowering of indices to any object indexed by i1 , ..., i` , ij ∈ {1, ..., n}, j = 1, ..., n.
E.g.:
ijk := δi` `jk ,
ij k := δj` i`k , etc.
Note that it is important to keep the order of the indices on the LHS due to the anti-symmetry
of , so that ij k 6= ijk . In fact, the order of indices always matter unless one is dealing with objects
that are symmetric in the respective indices. E.g., if aij are the entires of a matrix, then
aij := δi` a`j
and in general aij 6= aj i . However, if the matrix is symmetric, aij = aji , then aij = aj i , and we
write aji for aij .
The same way we lowered indices using δij , we can raise indices using δ ij . For instance, given
an object indexed by downstairs indices ij, i.e., Aij , we set
Aij := δ i` A`j .
Again, the order of the indices on the LHS matters unless the object is symmetric. It follows that
we can define the Kronecker delta with one index up and one down:
δji = δ i` δ`j .
 MATH 3120 - Intro to PDEs 35

It follows that 
1, i = j,
δji =
0, i 6= j.
Note that raising and then lowering (or vice-versa) an index gives the same object back. E.g.:
Xi = δij X j ⇒ X i = δ ij Xj = δ ij δj` X ` = X i ,
| {z }
δ`i

where we used δ`i = 0 for i 6= `.

Recall that ∂i = ∂x
∂
i . We define the derivative with an index upstairs by:

∂ i := δ ij ∂j .
Using this notation, we can write the Laplacian as:
∆ = ∂ i ∂i = δ ij ∂i ∂j .
We sometimes abbreviate ∂ij 2 = ∂ ∂ , ∂ 3 = ∂ ∂ ∂ , etc.
i j ijk i j k
Important remark. The use of the Kronecker delta and the raising and lowering of indices
provide us with a convenient and compact notation. But the overall discussion and definitions
probably seem a bit ad hoc. It turns out that these ideas can be given a more satisfactory content
within the language of differential geometry. For example, the Kronecker delta can be introduced
not as a “collection of symbols” but rather as a tensor satisfying certain properties. The raising
the lowering of indices can be interpreted as a map, given by the inner product, that identifies
elements of a vector space and its dual, on vector fields and one forms; or yet more generally as
the identification of covariant and contravariant tensors. Since we will not be discussing differential
geometry (except for some elementary aspects tied to PDEs), here we will take a purely instrumental
point of view, using the above machinery mostly as a matter of convenient notation.
8.4. Calculus facts. We collect a few calculus facts that we will use alter on.
Definition 8.8. We say that a map f is k-times continuously differentiable if all its partial deriva-
tives up to order k exist and are continuous in the domain of f . We denote the space of k-times
continuously differentiable functions in U ⊂ Rn by C k (U ). Sometimes we write simply C k if U is im-
plicitly understood, and sometimes we say simply “f is C k ” to mean that f is k-times continuously
differentiable.
Integration by parts. If u, v ∈ C 1 (Ω̄), then
Z Z Z
∂i uvdx = − u∂i vdx + uvν i dS,
Ω Ω ∂Ω
i = 1, ...n, where ν = is the unit outer normal to ∂Ω and dS is the volume element
(ν , ..., ν n )
1

induced on ∂Ω.
Students who have not seen the above integration by parts in Rn can view it as a generalization
of the divergence theorem in R3 . The latter can be written (using Stewart’s Calculus notation):
y −
→ x− → − →
div F dV = F · dS .
E S
−
→
Take F = uv − →e i , where −
→
e i has 1 in the ith component and zero in the remaining components.
Then,
−
→
div F = ∂i uv + u∂i v.
−
→
For example, if −
→e = e = (1, 0, 0), and writing F = (F , F , F ), so that
i 1 x y z
−
→
div F = ∂x Fx + ∂y Fy + ∂z Fz ,
36 Disconzi

we find
−
→
div F = div(uv, 0, 0) = ∂x (uv) = ∂x uv + u∂x v,
−
→ →
and similarly for −
→
e 2 and −
→
e 3 . Recalling also that d S = −
n dS, where −→n is the unit outer normal,
−
→ − → −
→ −
→ −
→ −
→
F · d S = (uv e ) · n dS = uv e · n dS.
i i

But −
→
ei·−
→
n = ith component of −
→
n = ni , thus
−
→ − →
F · d S = uvni .
Plugging the above into the divergence theorem:
y x
(∂i uv + u∂i v)dV = uvni dS
E S
which is the formula we stated in a different notation.
Definition 8.9. Let u ∈ C 1 (Ω̄). The normal derivative of u, denoted ∂ν ,
∂u
is a function defined on
∂Ωby
∂u
:= ∇u · ν,
∂ν
where ν is the unit outer normal to ∂Ω and ∇ is the gradient.
From the integration by parts formula we can derive the following formulas (sometimes called
Green’s identities): for u ∈ C 1 (Ω̄):
Z Z
∂i udx = uν i dS.
Ω ∂Ω
For u, v ∈ C 2 (Ω̄):
Z Z
∂u
∆udx = dS,
Ω ∂Ω ∂ν
Z Z Z
∂v
∇u · ∇vdx = − u∆vdx + u dS,
Ω Ω ∂Ω ∂ν
Z Z  
∂v ∂u
(u∆v − v∆u)dx = u −v dS.
Ω ∂Ω ∂ν ∂ν
9. Formal aspects of PDEs
Definition 9.1. A vector of the form
α = (α1 , ..., αn ),
where each entry is a non-negative integer is called a multiindex of order |α| = α1 + ... + αn . Given
a multiindex, we define:
∂ |α| u
Dα u := ,
∂(x1 )α1 ...∂(xn )αn
where u = u(x1 , ..., xn ). If k is a non-negative integer,
Dk u := {Dα u | |α| = k}
is the set of all k-th order partial derivatives of u. When k = 1 we identify Du with the gradient
of u. When k = 2 we identify D2 u with the Hessian matrix of u:
∂2u ∂2u
 
1 2 . . . 1 n
 ∂(x. ) ..
∂x ∂x
..
D2 u =  . .

. .

 
2
∂ u 2
∂ u
∂xn ∂x1
. . . ∂(xn )2
 MATH 3120 - Intro to PDEs 37

We can regard Dk u(x) as a point in Rn . Its norm is

k

sX
|Dk u(x)| = |Dα u(x)|2 ,
|α|=k

where |α|≤k means the sum is over all multiindices of order k. If u = (u1 , ..., um ) is vector valued,
P

we define
Dα u := (Dα u1 , ..., Dα um )
and set
Dk u := {Dα u | |α| = k},
and sX
k
|D u| = |Dα u(x)|2
|α|=k

as before.
We will now restate the definition of PDEs using the above notation. This new definition agrees
with the one previously given.
Definition 9.2. Let Ω ⊂ Rn be a domain and k ≥ 1 be a non-negative integer. An expression of
the form
F (Dk u(x), Dk−1 u(x), ..., Du(x), u(x), x) = 0,
x ∈ Ω, is called a k-th order partial differential equation (PDE), where:
k k−1
F : Rn × Rn × ... × Rn × R × Ω → R
is given and u : Ω → R is unknown. A solution to the PDE is a function u that verifies the PDE.
Sometimes we drop x from the notation and state the PDE as
F (Dk u, Dk−1 u, ..., Du, u, x) = 0 in Ω.
Ω is sometimes called the domain of definition of the PDE.
Example 9.3. ∆u = 0 in R3 can be written as
F (D2 u, Du, u, x) = 0 in R3
with F : R9 × R3 × R × |{z}
R3 → R given by the following expression. First, we label the coordinates
Ω
in R9 × R3 × R × R3 according to the order of the derivatives and x, i.e.,
∂2u ∂2u ∂ 2 u ∂u ∂u
1 2
, 1 2
, ..., 3 2
, 1 , ..., 3 , u, x1 , x2 , x3 ,
∂(x ) ∂x ∂x ∂(x ) ∂x ∂x
so
F = F (p11 , p12 , p13 , p21 , ..., p23 , p1 , p2 , p3 , p, x1 , x2 , x3 ).
| {z } | {z }
9 entries 3 entries
Then F is given by
F (p11 , ..., x3 ) = p11 + p22 + p33 .
Example 9.4. ∆u = f in R3 , where f (x) = (x1 )2 + (x2 )2 + (x3 )2 , can be written, using the
notation of the previous example, as in the definition with F given by
F (p11 , ..., x3 ) = p11 + p22 + p33 − ((x1 )2 + (x2 )2 + (x3 )2 ).
38 Disconzi

Definition 9.5. A PDE

F (Dk u, Dk−1 u, ..., Du, u, x) = 0
is called linear if F is linear in all its entries except possibly in x. Otherwise it is called non-linear.
More precisely, denoting F : Rn × Rn × ... × Rn × Rn × Ω → R, by F = F (− →
k k−1
p , x)
−
→p = ( pk,1 , ..., pk,nk , pk−1,1 , ..., pk−1,nk−1 , ..., p),
| {z } | {z }
nk entries for Rnk nk−1 entries for Rn
k−1

we can write F (−→p , x) = FH (−

→
p , x) + FI (x), where FI contains all terms that do not depend on − →
p
−
→
(i.e., terms that do not depend on u or its derivatives). The PDE is linear if FH ( p , x) is a linear
function of −→
p for fixed x. FH is called the homogeneous part of F and FI the inhomogeneous part.
The PDE is called homogenous if FI = 0 and inhomogeneous otherwise.
We clarify that when we say that FH is linear in, say, the entry Dk u, we mean that it is linear
in each component of Dk u separately. For instance, FH (Du, u, x) = 0 is linear if it is linear in Du.
Since Du = (∂1 u, ..., ∂n u) we mean that F is linear in each entry of (∂1 u, ..., ∂n u) plus in the entry
u. A linear PDE F (Dk u, ..., u, x) = 0 can always be written as
X
aα Dα u = f,
|α|≤k

where the aα and f are known functions defined on Ω. If the PDE is also homogenous then f = 0.
A PDE as defined above, where the unknown is a single function on Ω, is also called a scalar PDE.
Definition 9.6.
(1) A k th order PDE is called semi-linear if it has the form
X
aα Dα u + a0 (Dk−1 u, .., Du, u, x) = 0,
|α|=k

where the aα : Ω → R and a0 : Rn × ... × Rn × R × Ω → R are given functions.

k−1

(2) A k th order PDE is called quasi-linear if it has the form

X
aα (Dk−1 u, .., Du, u, x)Dα u + a0 (Dk−1 u, .., Du, u, x) = 0,
|α|=k

where aα , a0 : Rn × ... × Rn × R × Ω → R are known.

k−1

(3) A PDE is called fully nonlinear if it depends nonlinearly on its highest order derivative.
Definition 9.7. An expression of the form
F (Dk u(x), Dk−1 u(x), ..., Du(x), u(x), x) = 0,
is called a k th order system of PDEs, where
k k−1
F = (F 1 , ..., F ` ) : Rmn × Rmn × ... × Rmn × Rm × Ω → R`
is given and u = (u1 , ..., um ) : Ω → Rm is the unknown. A solution to the system of PDEs is a
function u : Ω → Rm that satisfies the system of PDEs. We sometimes drop the x-dependence and
write
F (Dk u, ..., Du, u, x) = 0
in Ω. We sometimes refer to a system of PDEs simply as a PDE.
The definitions of (non)linear, (non)homogenous, semi-linear, quasi-linear generalize in a straight-
forward fashion to systems. In particular, a linear system can be written as
X
Aα Dα u = f,
|α|≤k
 MATH 3120 - Intro to PDEs 39

where Aα : Ω → R`m are known ` × m matrices (depending on x ∈ Ω) and f : Ω → R` is a known

function (f = 0 if the system is homogeneous).
Having introduced the basic definitions and terminology for PDEs, let us discuss the case of
evolution equations, i.e., when one of the variables represents time.
When we study a PDE where one of the variables is the time variable, it is convenient to separate
time and space and denote the spatial variables by (x1 , ..., xn ) and the time variables by x0 . In this
case we have n + 1 variables and extend the multi-index notation to
∂ |α| u
α = (α0 , ..., αn ), |α| = α0 + ... + αn , Dα u = .
∂(x0 )α0 ∂(x1 )α1 ...∂(xn )αn
The domain of definition of the PDE in this case is Ω ⊂ Rn+1 , but it is convenient to take it to
be (TI , TF ) × Ω ⊂ Rn+1 , for some interval (TI , TF ) ⊂ R and some domain Ω ⊂ Rn . Typically
(TI , TF ) = (0, T ) for some T > 0. We also write Rn+1 = R × Rn wen we want to emphasize the
the first coordinate, x0 , correspond to time. We also write t := x0 for the time variable. Thus,
∂t = ∂x0 .
∂ ∂

Notation 9.8. We extend our indices convention by adopting the convention that Latin Lower-
case indices range from 1 to n (as we have used so far) and Greek lower-case indices range from 0
to n. For instance,
aα ∂α u = a0 ∂0 u + ai ∂i u
= a0 ∂t u + ai ∂i u
= a0 ∂t u + a1 ∂1 u + ... + an ∂n u.
Note that we use Greek letters to denote both indices varying from 0 to n and multi-indices.
The context will make the distinction clear. In particular, note that for multi-indices we never
use the convention that repeated indices are summed. Thus, for example, in aα ∂α , α is an index
summed from 0 to n, whereas in |α|≤k aα Dα , α is a multi-index summed over all multi-indices
P

with |α| ≤ k. Finally, if α = (α0 , ..., αn ) is a multi-index, we write − →α for its “spatial part,” i.e.,
−
→α = (α1 , ..., αn ).
We next state some useful calculus facts using multi-index notation. The formulas below
involve functions u = u(x1 , ..., xn ) and α = (α1 , ..., αn ), but clearly similar formulas hold for
u = u(x0 , x1 , ..., xn ) and α = (α0 , α1 , ..., αn ). For multi-indices α and β define
n
α ≤ β ⇔ αi ≤ βi for each i = 1, ..., n,
Y
α! = α1 !α2 !...αn !, xα = xαi i .
i=1

(1) Multinomial theorem:

X |α|  
|α| |α|!
k
(x1 + ... + xn ) = x , where
α
=
α α α!
|α|=k

(2) Leibniz’s formula or product rule:

X α  
α α!
Dα (uv) = Dβ uDα−β v, where = .
β β β!(α − β)!
β≤α

(3) Taylor’s forumla:

X 1
u(x) = Dα u(0)xα + O(|x|k+1 ) as x → 0.
α!
|α|≤k

Above, u, v : Rn → R are sufficiently regular as to make the formulas valid.

40 Disconzi

Remark 9.9. When we introduce a PDE, we indicate the domain Ω where it is defined, which
says that we are looking for a solution that is defined in Ω. It may happen, however (and it is
often the case for non-linear PDEs) that we are able to find a solution u but u is defined only on a
smaller domain Ω0 ⊂ Ω. I.e., u satisfies the PDE only for x ∈ Ω0 , where Ω0 is strictly smaller than
Ω. In fact, we a priori do not know whether it is possible to satisfy the PDE for all x ∈ Ω. We
still call such a u that is defined only on Ω0 a solution, and sometimes call it a local solution if we
want to emphasize that the solution we found is defined on a domain smaller than where the PDE
was originally stated. In other words, the domain of definition of the PDE is a guide that helps us
define the problem, but it can happen that solution are only defined in a subset of Ω.
Example 9.10. Let us illustrate this situation with a simple ODE example. Consider the Riccati
equation y 0 = y 2 in Ω = (0, ∞) with initial condition y(0) = 1. The solution is y(t) = 1−t 1
.
This solution, however, is not defined for t = 1. Thus we in fact have a local solution defined on
Ω = (0, 1) (we do not take Ω0 = (0, 1) ∪ (1, ∞) because this set is not connected; and we take the
portion (0,1)because we need to approached zero to satisfy the initial condition)
We can also defined boundary value problems, initial value problems, and initial-boundary value
problems as we had done for the 1d wave equation. We will not give these general definitions here,
but will introduce them as needed to study specific problems. We note that in such cases we will in
general seek a solution defined on a larger domain than Ω. For example, we may want u : Ω̄ → R
in a boundary value problem. What exactly is required is usually a case-by-case analysis.
Important notation on constants In what follows we are going to derive estimates and
computations that involve numerical constants whose specific value will not be important. Thus,
we will denote by C > 0 a generic positive constant that can vary from line to line. C will
generally depend on a fixed data of the problem (e.g., the dimension n). Sometimes we indicate
the dependence of C using subscripts or function notation, e.g. Cn or C(n).

10. Laplace’s Equation in Rn

We are going to study Laplace’s equation in Rn :
4u = 0 in Rn ,
and its inhomogeneous version known as Poisson’s equation:
4u = f in Rn ,
where f : Rn → R is given.
We begin looking for a solution of the form u(x) = v(r) where r = |x| = ((x1 )2 + ... + (xn )2 )1/2
is the distance to the origin. The motivation to look for such a solution is that Laplace’s equation
is rotationally invariant (this will be a HW). Direct computation gives:
xi i i 2 (xi )2
 
0x 2 00 (x ) 0 1
∂i r = , x 6= 0, ∂i u = v , ∂i u = v +v − 3 .
r r r2 r r
Summing from 1 to n:
n−1 0
4u = v 00 +
v.
r
Hence 4u = 0 iff v 00 + n−1
r v = 0, which is an ODE for v. If n 6= 0 we cam write it as
0 0

v 00 1−n
(ln(v 0 ))0 = 0
= ,
v r
which gives v 0 (r) = A
rn−1
for some constant A. If r > 0, integrating again we find

a ln r + b, n = 2
v(r) = a
rn−2
+ b, n ≥ 3,
 MATH 3120 - Intro to PDEs 41

where a and b are arbitrary constants. This calculation motivates the following definition.

Definition 10.1. The function

 1
2π ln|x|, n=2
Γ(x) :=
n(2−n)ωn |x|n−2 ,
1 1
n ≥ 3,

is called the fundamental solution of Laplace’s equation.

Notation 10.2. We denote by Br (x) the (open) ball of radius r centered at x in Rn , i.e.,

Br (x) := {y ∈ Rn | |x − y| < r}.

Sometimes we write Brn (x) to emphasize the dimension. We denote

ωn := vol(B1n (0)).

In particular, ω3 = 43 π.

Note that 4Γ(x) = 0 for x 6= 0 by construction. Sometimes we write Γ(|x|) to emphasize the
radial dependence on r = |x|. Before solving Laplace’s equation, we need one more definition.

Definition 10.3. The support of a map f : U → R is the set

supp(f ) := {x ∈ U | f (x) 6= 0},

where the bar is the closure. Recall that a set U ⊂ Rn is called compact if it is closed and bounded.
We say that f has compact support if supp(f ) is compact. We denote by Cck (U ) the space of C k
functions in U with compact support.

Theorem 10.4. Let f ∈ Cc2 (Rn ). Set

Z
u(x) = Γ(x − y)f (y)dy = (Γ ∗ f )(x).
Rn

Then:
(1) u is well-defined
(2) u ∈ C 2 (Rn )
(3) 4u = f in Rn .

Proof. We will carry out the proof for n ≥ 3. The case n = 2 is done with similar arguments.
To begin, recall that a continuous function over a compact set is always has a maximum and
minimum. Therefore, since f has compact support, there exists a constant C > 0 such that
|f (x)| ≤ C for every x. Moreover, again by the compact support of f , there exists a R > 0 such
that
Z Z
Γ(x − y)f (y)dy = Γ(x − y)f (y)dy.
Rn BR (x)

Thus,
Z Z Z
1
Γ(x − y)f (y)dy ≤ C |Γ(x − y)|dy ≤ C dy.
Rn BR (x) BR (x) |x − y|n−2

We now take polar coordinates (r, ω) centered at x, where r = distance to x and ω ∈ S n−1 (n − 1
dimensional unit sphere), sothat y − x = rω, |x − y| = r.
42 Disconzi

y−x
z = rω

y
ω

x
∂B1 (x) ' S n−1

S n−1

In these coordinates dy = rn−1 dω, where dω is the volume element on S n−1 ( for n = 3, dω =
sin(φ)dφdθ). Then
Z Z RZ Z R Z
1 1 n−1
n−2
dy = n−2
r drdω = rdr dω = C,
BR (x) |x − y| 0 S n−1 r 0 S n−1

showing that u is well defined, i.e., (1).

To prove (2), first make a change of variables z = x − y, so
Z Z
u(x) = Γ(x − y)f (y)dy = Γ(z)f (x − z)dz.
Rn Rn

Note that ∂i f and ∂ij

2 f also have compact support, thus an argument to the above shows that

Z Z
|Γ(y)∂i f (x − y)|dy and 2
|Γ(y)∂ij f (x − y)|dy
Rn Rn

are well defined. Let ei = (0, ..., 1, ..., 0) be the canonical basis vectors in Rn and let h > 0. Then,
for any x:
 
u(x + hei ) − u(x) f (x + hei − y) − f (x − y)
Z
= Γ(y) dy
h Rn h
 
f (x + hei − y) − f (x − y)
Z
= Γ(y) dy,
BR (0) h

where the second equality holds for a sufficiently large R in view of the compact support of f . Since
f (x + ei h − y) − f (x − y)
lim = ∂i f (x − y)
h→0 h
and the integral of Γ(y)∂i f (x − y) is well defined,
 
u(x + hei ) − u(x) f (x + hei − y) − f (x − y)
Z
lim = lim Γ(y) dy
h→0 h h→0 Rn h
 
f (x + hei − y) − f (x − y)
Z
= Γ(y) lim dy
Rn h→0 h
Z
= Γ(y)∂i f (x − y)dy,
Rn
 MATH 3120 - Intro to PDEs 43

showing that the limit limh→0 u(x+hehi )−u(x) exists, i.e., ∂i u(x) exists. Repeating this argument with
f (x − y) replaced by ∂i f (x − y) we conclude that ∂ij 2 u(x) exists and

Z
2 2
∂ij u(x) = Γ(y)∂ij f (x − y)dy.
Rn

To show that u ∈ C 2 (Rn ), it remains to show that ∂ij 2 u is continuous. Fix x ∈ Rn and  > 0, and
0
consider:
Z
2 2 2 2
|∂ij u(x0 ) − ∂ij u(x)| = Γ(y)(∂ij f (x0 − y) − ∂ij f (x − y))dy
Rn
Z
2 2
≤ |Γ(y)||∂ij f (x0 − y) − ∂ij f (x − y)|dy.
Rn

Since ∂ij
2 f is continuous and has compact support it is uniformly continuous, i.e., given 0 , there

exists a δ > 0 such that |∂ij 2 f (z) − ∂ 2 f (y)| < 0 whenever |z − y| < δ. Putting 0 =  , with
ij C
C = BR (0) |Γ(y)|dy ( which we already know to be finite), we find that if |x0 − x| < δ, so that
R

|(x0 − y) − (x − y)| < δ, we obtain that

Z
2 2 2 2
|∂ij u(x0 ) − ∂ij u(x)| ≤ |Γ(y)| |∂ij f (x0 − y) − ∂ij f (x − y)| dy < ,
BR (0) | {z }
<0

showing that u ∈ C 2 (Rn ).

To show (3), from the expression for ∂ij u we obtain
Z
ij 2
4u(x) = δ ∂ij u(x) = Γ(y)4x f (x − y)dy,
Rn
Z Z
= Γ(y)4x f (x − y)dy + Γ(y)4x f (x − y)dy =: I1 + I2 ,
Rn \B (0) B (0)

where  > 0 and we write 4x to emphasize that in 4x f (x − y) the Laplacian is with respect to the
x variable. Noticing that 4x f (x − y) = 4y f (x − y), Green’s identities give:
Z
I1 = Γ(y)4y f (x − y)dy
Rn \B (0)
Z Z
∂f
=0 ∇Γ(y) · ∇y f (x − y)dy + Γ(y) (x − y)dS(y)
Rn \B (0) ∂B (0) ∂ν
 
=: I11 + I12 ,
where we write ∇y and dS(y) to emphasize that the gradient and integration over ∂B (0) are on
the y variable. We also notice that in the integration by parts there is no term to be “evaluated at
∞” since f has compact support.
Let’s now analyze the integrals I2 , I11 , and I  . Observe that:
12
Z Z

|I2 | ≤ |Γ(y)| |4x f (x − y)| ≤ C |Γ(y)|dy
B (0) | {z } B (0)
C
Z 
1 n−1
≤C n−2
r dr = C2 .
0 r
Since dS(y) = n−1 dω and |Γ(y)| ≤ C/n−2 on ∂B (0):
Z
 ∂f
|I12 |≤ |Γ(y)| (x − y) dS(y) ≤ C.
∂B (0) ∂ν
44 Disconzi

For I11
 , we integrate by parts again:
Z

I11 = − ∇Γ(y) · ∇y f (x − y)dy
Rn \B (0)
Z Z
∂Γ
= 4Γ(y)f (x − y)dy − (y)f (x − y)dS(y)
Rn \B (0) ∂B (0) ∂ν
Z
∂Γ
=0− (y)f (x − y)dS(y),
∂B (0) ∂ν
where we used that 4Γ(y) = 0 for y 6= 0.
From the explicit expression for Γ(y), compute:
1 y
∇Γ(y) = , y 6= 0.
nωn |y|n
The unit outer normal in the integral is given by ν = −y
|y| , thus

1 |y|2
Z Z
 1
I11 = n+1
f (x − y)dS(y) = f (x − y)dS(y),
∂B (0) nωn |y| nωn n−1 ∂B (0)
since |y| =  on ∂B (0).

 ν

B (x)

B (0) x

z
y

Making a change of variables x − y = z, we find

Z
 1
I11 = f (z)dS(z).
nωn n−1 ∂B (0)
Note that nωn n−1 is the surface are, or volume, of ∂B (0) (e.g., for n = 3, nωn n−1 = 4π2 ), so we
write Z
 1
I11 = f (y)dS(y).
vol(∂B (0)) ∂B (x)
 MATH 3120 - Intro to PDEs 45

Since we have Z Z
4u(x) = (. . . ) + (. . . ) = I1 + I2
Rn \B (0) B (0)
which is valid for any  > 0, we conclude that
4u(x) = lim I1 + lim I2
→0+ →0+
if the limits exist. From the foregoing:
lim I2 = 0,
→0+
lim I1 = lim I11
 
+ lim I12
→0+ →0+ →0+
| {z }
=0
Z
1
= lim f (y)dS(y).
→0+ vol(∂B (x)) ∂B (x)

The result (3) now follows from the lemma stated right below, whose proof will be a HW. 
Lemma 10.5. For any continuous function h:
Z
1
lim h(y)dS(y) = h(x)
→0+ vol(∂B (x)) ∂B (x)
Z
1
lim h(y)dy = h(x)
→0+ vol(B (x)) B (x)

Proof. HW 
Remark 10.6. From the expression for Γ(x) we obtain the following useful estimates:
C C
|DΓ(x)| ≤ n−1
, |D2 Γ(x)| ≤ , x 6= 0.
|x| |x|n
10.1. Harmonic functions.
Definition 10.7. A solution to Laplace’s equation is called a harmonic function. We say that u
is a harmonic function (or simply harmonic) in Ω if we want to emphasize that it solves Laplace’s
equation in Ω.
Theorem 10.8 (mean value formula for Laplace’s equation). Let u ∈ C 2 (Ω) be harmonic in Ω.
Then Z Z
1 1
u(x) = udS = udy,
vol(∂Br (x)) ∂Br (x) vol(Br (x)) Br (x)
for each Br (x) ⊂ Ω.
Remark 10.9. This theorem says that harmonic functions are “non-local” since their value at x
depends on their values on ∂Br (x); in particular r can be arbitrarily large for Ω = Rn .
Proof. Define Z
1
f (r) := u(y)dS(y).
vol(∂Br (x)) ∂Br (x)
Changing variables z = r ,
y−x
recalling that dS = rn−1 dω, vol(∂B
r (x)) = nωn r
n−1 :
Z
1
f (r) = u(x + rz)dS(z).
nωn ∂B1 (0)
Taking the derivative and noticing that we can differentiate under the integral:
Z
1
f 0 (r) = ∇u(x + rz) · zdS(z).
nωn ∂B1 (0)
46 Disconzi

Changing variables back to y:

 
y−x
Z
0 1
f (r) = ∇u(y) · dS(y).
nωn rn−1 ∂Br (x) r
Since y−x
r = ν =unit outer normal to ∂Br (x):
Z
1
f 0 (r) = ∇u(y) · νdS(y)
nωn rn−1 ∂Br (x)
Z
1 ∂u
= n−1
(y)dS(y)
nωn r ∂Br (x) ∂ν
Z
1
= 4u(y)dy = 0
nωn rn−1 Br (x)
where we used Green’s identities. Thus, f (r) is constant so
Z
1
u(y)dS(y) = f (r) = lim f (r)
vol(∂Br (x)) ∂Br (x) r→0+
Z
1
= lim u(y)dS(y) = u(x),
r→0 vol(∂Br (x)) ∂Br (x)
+

showing the first equality. For the second, integrate in polar coordinates to find
Z Z r Z !
1 1
u(y)dy = udS ds = u(x).
vol(Br (x)) Br (x) ωn rn 0 ∂Bs (x)
| {z }
=nωn sn−1 u(x)


Theorem 10.10 (converse of the mean value property). If u ∈ C 2 (Ω) is such that
Z
1
u(x) = udS
vol(∂Br (x)) ∂Br (x)
for each Br (x) ⊂ Ω. Then u is harmonic.
Proof. This will be a HW. 
Definition 10.11. Let U ⊂ Rn .
We say that a subset V ⊂ U is relatively open, or open in U ,
if V = U ∩ W for some open set W ⊂ Rn . V ⊂ U is said to relatively closed, or closed in U , if
V = U ∩ W for some closed set W ⊂ Rn . A set Ω ⊂ Rn is called connected if the only non-empty
subset of Ω that is both open and closed in Ω is Ω itself.
Remark 10.12. Sometimes we say simply that V ⊂ U is open/closed to mean that is open/closed
in U , i.e., U is implicitly understood.
Students who have no seen the definition of connected sets are encouraged to think about how
the above definition corresponds to the intuition that Ω cannot be “split into separate pieces. ”
Theorem 10.13 (maximum principle). Suppose that u ∈ C 2 (Ω) ∩ C 0 (Ω̄) is harmonic where Ω is
bounded. Then
max u = max u.
Ω̄ ∂Ω
Moreover, if u(x0 ) = maxΩ̄ u for some x0 ∈ Ω, then u is constant.
Remark 10.14. Replacing u by −u we obtain similar statements with min. Thus, we can summa-
rize the maximum principle by saying that a harmonic function achieves its maximum and minimum
on the boundary.
 MATH 3120 - Intro to PDEs 47

Proof. Suppose that for some x0 ∈ Ω, we have u(x0 ) = M = maxΩ̄ u. For 0 < r < dist(x0 , ∂Ω), the
mean value property gives:
Z
1
M = u(x0 ) = udy ≤ M.
vol(Br (x0 )) Br (x0 )
Equality in ≤ happens only if u(y) = M for all y ∈ Br (x0 ). Therefore the set A := {x ∈ Ω | u(x) =
M } is both open and closed in Ω, thus, A = Ω, showing the second statement. The first statement
follows form the second. 
10.2. Further results for harmonic functions and Poisson’s equation. Here we list a few
important results concerning 4u = f that we will not prove.
Theorem 10.15 (Liouville’s Theorem). Suppose that u : Rn → R is harmonic and bounded (i.e.,
there exists a constant M ≥ 0 such that |u(x)| ≤ M for all x ∈ Rn ). Then u is constant.
Definition 10.16. Let f : Ω → R and g : ∂Ω → R be given. The following boundary-value
problem
4u = f in Ω

u = g on ∂Ω
is called the (inhomogeneous) Dirichlet problem for the Laplacian.
Theorem 10.17. Let Ω ⊂ Rn be a bounded domain with a C 3 boundary. Let f ∈ C 1 (Ω̄) and
g ∈ C 3 (Ω̄). Then, there exists a unique solution u ∈ C 2 (Ω̄) to the Dirichlet problem
4u = f in Ω

u = g on ∂Ω
Remark 10.18. To solve Poisson’s equation in Rn we introduced the fundamental solution. One
approach to solve the Dirichlet problem is to introduce an analogue of the fundamental solution
which takes the boundary into account, known as the Green function.

11. The wave equation in Rn

Here we will study the Cauchy problem for the wave equation in Rn , i.e.,
 u = 0 in [0, ∞) × Rn


u = u0 on {t = 0} × Rn
∂t u = u1 on {t = 0} × Rn


where  := −∂t2 + 4 is called the D’Alembertian (or the wave operator) and u0 , u1 : Rn → R are
given. The initial conditions can also be stated as u(0, x) = u0 (x), ∂t u(0, x) = u1 (x), x ∈ Rn .
Definition 11.1. The sets
Ct0 ,x0 := {(t, x) ∈ (−∞, ∞) × Rn | |x − x0 | ≤ |t − t0 |},
Ct+0 ,x0 := {(t, x) ∈ (−∞, ∞) × Rn | |x − x0 | ≤ t − t0 },
Ct−0 ,x0 := {(t, x) ∈ (−∞, ∞) × Rn | |x − x0 | ≤ t0 − t},
are called, respectively, the light-cone, future light-cone, and past light-cone with vertex at (t0 , x0 ).
The sets
Kt0 ,x0 := Ct0 ,x0 ∩ {t ≥ 0}
Kt+0 ,x0 := Ct+0 ,x0 ∩ {t ≥ 0},
Kt−0 ,x0 := Ct−0 ,x0 ∩ {t ≥ 0},
are called, respectively, the light-cone, future light-cone, and past light-cone for positive time with
vertex at (t0 , x0 ).
48 Disconzi

We often omit “for positive time” and refer to the sets K as light-cones. We also refer to a part
of a cone, e.g., for 0 ≤ t ≤ T , as the truncated (future, past) light-cone.
(t0 , x0 )
Ct+0 ,x0

Ct−0 ,x0

(t0 , x0 ) (t0 , x0 )
(t0 , x0 )

Kt−0 ,x0

{t = 0}
Ct0 ,x0

Lemma 11.2 (differentiation of moving regions). Let Ω(τ ) ⊂ Rn be a family of bounded domains
with smooth boundary depending smoothly on the parameter τ . Let v be the velocity of the moving
boundary ∂Ω(τ ) and ν the unit outer normal to ∂Ω(τ ). If f = t(τ, x) is smooth then
Z Z Z
d
f dx = ∂τ f dx + f v · νdS.
dτ Ω(τ ) Ω(τ ) ∂Ω(τ )

Proof. HW. (Compare this with the fundamental theorem of calculus). 

Theorem 11.3 (finite propagation speed). Let u ∈ C 2 ([0, ∞) × Rn ) be a solution to the Cauchy
problem for the wave equation. If u0 = u1 = 0 on {t = 0} × Bt0 (x0 ), then u = 0 within Kt−0 ,x0 .
(Thus, the solution at (t0 , x0 ) depends only on the data on Bt0 (x0 ) and the cone Kt−0 ,x0 is also called
a domain of dependence).
Proof. Define the “energy” as
Z
1
E(t) = ((∂t u)2 + |∇u|2 )dx, 0 ≤ t ≤ t0 .
2 Bt0 −t (x0 )

Then:
Z Z
dE 1
= (∂t u∂t2 u + ∇u · ∇∂t u)dx + ((∂t u)2 + |∇u|2 )v · νdS.
dt Bt0 −t (x0 ) 2 ∂Bt0 −t (x0 )

The points on the boundary move inward orthogonally to the spheres ∂Bt0 −t (x0 ) and with finite
speed linear in t, thus v = −ν.

Bt0 −t0 (x0 ) v at t0

v at t

Bt0 −t (x0 )

Kt−0 ,x0

Integrating by parts:
Z Z Z
∂u
∇u · ∇∂t udx = − 4u∂t udx + ∂t udS.
Bt0 −t (x0 ) Bt0 −t (x0 ) ∂Bt0 −t (x0 ) ∂ν
 MATH 3120 - Intro to PDEs 49

Thus,
Z Z Z
dE 2 ∂u 1
= (∂t u − 4u) ∂t udS + ∂t u − ((∂t u)2 + |∇u|2 )dS
dt Bt0 −t (x0 ) | {z } ∂Bt0 −t (x0 ) ∂ν 2 ∂Bt0 −t (x0 )
=0
Z  
∂u 1 2 1 2
= ∂t u − (∂t u) − |∇u| dS
∂Bt0 −t (x0 ) ∂ν 2 2
Z  
1 2 1 2
≤ |∇u||∂t u| − (∂t u) − |∇u| dS,
∂Bt0 −t (x0 ) 2 2
where we used that ∂u
∂ν ∂t u ≤ ∂u
∂ν ∂t u = ∂u
∂ν |∂t u| and
∂u
= |∇ · ν| ≤ |∇u| |ν| = |∇u|.
∂ν |{z}
=1

Now apply the Cauchy-Schwarz inequality ab ≤ 2 + 2 with a = |∇u|, b = |∂t u|, to get
a2 b2
Z  
dE 1 2 1 2 1 2 1 2
≤ |∇u| + (∂t u) − (∂t u) − |∇u| = 0,
dt ∂Bt0 −t (x0 ) 2 2 2 2
thus E(t) is decreasing. Since E(t) ≥ 0 and
Z
1
E(0) = ((∂t u(0, x))2 + |∇u(0, x)|2 )dx = 0.
2 ∂Bt0 (x0 ) | {z } | {z }
=u1 (x)=0 |∇u0 (0,x)|=0

We conclude that E(t) = 0 for all 0 ≤ t ≤ t0 .

Since E(t) is the integral of a positive continuous function over Bt−t0 (x0 ), E(t) = 0 implies that,
for each t, the integrand must vanish, i.e.,
(∂t u(t, x))2 + |∇u(t, x)|2 = 0
for all (t, x) ∈ Kt−0 ,x0 , which then implies ∂t u(t, x) = 0 and ∇u(t, x) = 0 for all (t, x) ∈ Kt−0 ,x0 .
Since Kt−0 ,x0 is connected, we conclude that u is constant in time and space within Kt−0 ,x0 , i.e.,
u(t, x) = C =constant in Kt−0 ,x0 . Since u(0, x) = u0 (x) = 0, C must be zero. 
Notation 11.4. Henceforth, we assume that n ≥ 2. Set
Z
1
U (t, x; r) := u(t, y)dS(y),
vol(∂Br (x)) ∂Br (x)
Z
1
U0 (x; r) := u0 (t, y)dS(y),
vol(∂Br (x)) ∂Br (x)
Z
1
U1 (x; r) := u1 (t, y)dS(y)
vol(∂Br (x)) ∂Br (x)
which are spherical average over ∂Br (x).
Proposition 11.5 (Euler-Poisson-Darboux equation). Let u ∈ C m ([0, ∞) × Rn ), m ≥ 2 be a
solution to the Cauchy problem for the wave equation. For fixed x ∈ Rn , consider U = U (t, x; r) as
a function of t and r. Then U ∈ C m ([0, ∞) × [0, ∞)) and U satisfies the Euler-Poisson-Darboux
equation:
n−1

∂t U − ∂r U − r ∂r U = 0 on (0, ∞) × (0, ∞),
 2 2


 U = U0 on {t = 0} × (0, ∞),
on {t = 0} × (0, ∞).

∂t U = U1

50 Disconzi

Proof. Differentiability with respect to t is immediate, as is differentiability with respect to r for

r > 0.
Arguing as in the proof of the mean value formula for Laplace’s equation:
Z
r 1
∂r U (t, x; r) = 4u(t, y)dy.
n vol(Br (x)) Br (x)

This implies limr→0+ ∂r U (t, x; r) = 0. Next,

Z  Z
1 1 r 1
∂r2 U (t, x; r) = 4u(t, y)dy + ∂r 4u(t, y)dy
n vol(Br (x)) Br (x) n vol(Br (x)) Br (x)
Z
r 1
+ ∂r 4u(t, y)dy.
n vol(Br (x)) Br (x)

But ∂r y)dS(y), and recall that vol(Br (x)) = ωn rn , so

R R
Br (x) 4u(t, y)dy = ∂Br (x) 4u(t,  
r 1
n vol(Br (x)) = 1
nωn r n−1 = 1
,
vol(∂Br (x)) n
r
∂r
1 r 1
vol(Br (x)) = n ∂r ωn rn = − ωn1rn = − vol(B1r (x)) , so
  Z Z
1 1 1
∂r2 U (t, x; r) = −1 4u(t, y)dy + 4u(t, y)dS(y).
n vol(Br (x)) Br (x) vol(∂Br (x)) ∂Br (x)

This implies that lim ∂r2 U (t, x; r) = n1 4u(t, x).

r→0+
Proceeding this way we compute all derivates of U with respect to r and conclude that U ∈
C m ([0, ∞) × [0, ∞)).
Returning to the expression for ∂r (U ):
Z Z
r 1 r
∂r U = 4u = ∂r2 u,
n vol(Br (x)) Br (x) n Br (x)

thus
rn
 Z
n−1
∂r (r ∂r U ) = ∂r ∂t2 u
n vol(Br (x)) Br (x)
Z !
1
= ∂r ∂2u
nωn Br (x) t
Z
1
= ∂2
nωn ∂Br (x) t
rn−1
Z
= ∂2u
vol(∂Br (x)) ∂Br (x) t
Z !
1
= rn−1 ∂t2 u
vol(∂Br (x)) ∂Br (x)
= rn−1 ∂t2 U.

On the other hand:

∂r (rn−1 ∂r U ) = (n − 1)rn−2 ∂r U + rn−1 ∂r2 U,

and equating the two right hand sides gives the result. 
 MATH 3120 - Intro to PDEs 51

11.1. The Reflection Method. We will use the function U (t, x; r) to reduce the higher dimen-
sional wave equation to the 1d wave equation, for which D’Alembert’s formula is available, in the
variable t and r. However, U (t, x; r) is defined only for r ≥ 0, whereas D’Alembert’s formula is for
−∞ < r < ∞. Thus, we first consider the system:
utt − uxx = 0 in (0, ∞) × (0 × ∞),


u = u0 on {t = 0} × (0, ×),




 ∂t u = u1 on {t = 0} × (0, ∞),
u = 0 on (0, ∞) × {x = 0},



where u0 (0) = u1 (0) = 0. Consider odd extensions, where t ≥ 0 :

( ( (
u(t, x) x≥0 u0 (x) x≥0 u1 (x) x≥0
u
e(t, x) = , uf0 (x) = , u
f1 (x) =
−u(t, −x) x ≤ 0 −u0 (−x) x ≤ 0 −u1 (−x) x ≤ 0.

A solution to the problem on (0, ∞) × (0, ∞) is obtained by solving

exx = 0 in (0, ∞) × R,

u
 ett − u
u
e=uf0 on {t = 0} × R,
f1 on {t = 0} × R,

∂t u
e=u


and restricting to (0, ∞) × (0, ∞) where u e = u. D’Alembert’s formula gives

Z x+t
1 1
u
e(t, x) = 2 (f f0 (x − t)) + 2
u0 (x + t) + u u
f1 (y)dy.
x−t

Consider now t ≥ 0 and x ≥ 0, so that u e(t, x) = u(t, x). Then x + t ≥ 0 so that u

f0 (x + t) = u0 (x + t).
If x ≥ t, then the variable of integration, y, satisfies y ≥ R0, since y ∈[x − t, x + t]. In the case
f1 (y) = u1 (y). Thus u(t, x) = (u0 (x + t) + u0 (x − t))/2 + x−t u1 (y)dy /2 for x ≥ t. If 0 ≤ x ≤ t,
x+t
u
then u u0 (−(x − t)) and
f0 (x − t) = −f
Z x+t Z 0 Z x+t
u
f1 (y)dy = u
f1 (y)dy + u
f1 (y)dy
x−t x−t 0
Z 0 Z x+t
=− u1 (−y)dy + u1 (y)dy
x−t 0
Z 0 Z x+t
= u1 (y)dy + u1 (y)dy
−x+t 0
Z x+t
= u1 (y)dy.
−x+t
R 
Thus, u(t, x) = (u0 (x + t) − u0 (−x + t))/2 + −x+t u1 (y)dy /2 for 0 ≤ x ≤ t. Summarizing:
x+t

 Z x+t
1 1
 2 (u0 (x + t) + u0 (x − t)) + 2 u1 (y)dy x≥t≥0



x−t
u(t, x) = Z x+t
1 1

 2 (u0 (x + t) − u0 (−x + t)) + 2 u1 (y)dy 0 ≤ x ≤ t.


−x+t

Note that u is not C 2 except if u000 (0) = 0. Note also that u(t, 0) = 0.
This solution can be understood as follows: for x ≥ t ≥ 0, finite propagation speed implies that
the solution “does not see” the boundary. For 0 ≤ x ≤ t, the waves traveling to the left are reflected
52 Disconzi

on the boundary where u = 0.

t

x≤t

x≥t domain of dependence

−x0 x0 x

11.2. Solution for n = 3 : Kirchhoff’s formula. Set U e = rU, U f1 = rU1 , where

f0 = rU0 , U
U, U0 , U1 are as in the Euler-Poisson-Darboux equation (see Notation 11.4). Then,
∂t2 U
e = r∂ 2 U
t
 
2 3−1
= r ∂r U + ∂r U
r
= r∂r2 U + 2∂r U
= ∂r2 (rU ) = ∂r2 (U
e ),

so Ue solves the 1d wave equation on (0, ∞) × (0, ∞) with initial conditions U e (0, r) = U
f0 (r),
∂t U (0, r) = U1 (r). By the reflection method discussed above, we have
e f
  Z r+t
1 f 1
U (t, x; r) = 2 U0 (r + t) − U0 (−r + t) + 2
e f U
f1 (y)dy
−r+t

for 0 ≤ r ≤ t, where we used the notation U f0 (r + t) and Uf1 for U

f0 (x; r + t), U
f1 (x; y).
For the definition of Ue and U and the above formula:
Z
1
u(t, x) = lim u(t, y)dS(y)
r→0+ vol(∂Br (x)) ∂Br (x)

= lim U (t, x; r)
r→0+

U
e (t, x; r)
= lim
r→0+ r
f0 (t + r) − U
f0 (t − r) Z t+r
U 1
= lim + lim U
f1 (y)dy.
r→0+ 2r r→0+ 2r t−r
Note that
f0 (t + r) − U
U f0 (t − r) f0 (t + 2r) − U
U f0 (t)
lim = lim f0 0 (t),
=U
r→0+ 2r r→0 + 2r
and Z t+r
1
lim Uf1 (y)dy = Uf1 (t)
r→0+ 2r t−r

(this equality is simply limr→0+ vol(Br (x)) Br (x) f (y)dy = f (x) for n = 1). So,
1
R

0
u(t, x) = U
f0 (t) + U
f1 (t).
 MATH 3120 - Intro to PDEs 53

Invoking the definition of U

f0 and U
f1 :
Z ! Z
∂ t t
u(t, x) = u0 (y)dS(y) + u1 (y)dS(y). (11.1)
∂t vol(∂Bt (x) ∂Bt (x) vol(Bt (x)) ∂Bt (x)
Making the change of variable z :− (y − x)/t (recall that we are treating the n = 3 case, so in the
calculations that follow n = 3, but we write n for the sake of clearer notation):
Z Z
1 1
u0 (y)dS(y) = u0 (y)dS(y)
vol(∂Bt (x)) ∂Bt (x) nωn tn−1 ∂Bt (x)
Z
1
= u0 (x + tz)tn−1 dS(z)
nωn tn−1 ∂B1 (x)
Z
1
= u0 (x + tz)dS(z).
nωn ∂B1 (x)
Then,
Z ! Z
∂ 1 1 ∂
u0 (y)dS(y) = u0 (x + tz)dS(z)
∂t vol(∂Bt (x)) ∂Bt (x) nωn ∂t ∂B1 (x)
Z
1
= (∇u0 (x + tz) · z)dS(z).
nωn ∂B1 (x)
Changing variables back to y, that is y = x + tz and recalling that dS(y) = tn−1 dS(z) :
!  
y−x
Z Z
∂ 1 1
u0 (y)dS(y) = ∇u0 (y) · dS(y).
∂t vol(∂Bt (x)) ∂Bt (x) vol(∂Bt (x)) ∂Bt (x) t
Using this in Equation 11.1’s expression for u(t, x), yields Kirchhoff’s formula:
Z
1
u(t, x) = (u0 (y) + tu1 (y))dS(y)
vol(∂Bt (x)) ∂Bt (x)
Z (11.2)
1
+ (∇u0 (y) · (y − x))dS(y).
vol(∂Bt (x)) ∂Bt (x)

Theorem 11.6. Let u0 ∈ C 3 (R3 ) and u1 ∈ C 2 (R3 ). Then, there exists a unique u ∈ C 2 ([0, ∞)×R3 )
that is a solution to the Cauchy problem for the wave equation in three spatial dimensions. Moreover,
u is given by Kirchhoff’s formula 11.2.
Proof. Define u by Kirchhoff’s formula. By construction it is a solution with the stated regularity.
Uniqueness follows from the finite speed of propagation property. 
11.3. Solution for n = 2 : Poisson’s formula. We now consider u ∈ C 2 ([0, ∞) × R2 ) a solution
to the wave equation for n = 2. Then
v(t, x1 , x2 , x3 ) := u(t, x1 , x2 )
is a solution for the wave equation in n = 3 dimensions with data v0 (x1 , x2 , x3 ) := u0 (x1 , x2 ) and
v1 (x1 , x2 , x3 ) := u1 (x1 , x2 ). Let us write x = (x1 , x2 ) and x = (x1 , x2 , 0). Thus, from the n = 3 case:
Z ! Z
∂ t t
u(t, x) = v(t, x) = v0 dS + v1 dS,
∂t vol(∂B t (x) ∂B t (x) vol(∂B t (x) ∂B t (x)
where B t (x) is the ball in R3 with center x and radius t and dS is the volume element on ∂B t (x).
We now rewrite this formula with integrals involving only variables in R2 .
54 Disconzi

The integral over ∂B t (x) can be written as

Z Z Z
= +
+ −
,
∂B t (x) ∂B t (x) ∂B t (x)
−
where ∂B t (x) and ∂B t (x) are the upper and lower hemispheres of ∂B t (x), respectively. The upper
+

cap ∂B t (x) is parametrized by f (y) = t2 − (y − x)2 , y = (y 1 , y 2 ) ∈ Bt (x), x = (x1 , x2 ), where

+ p

Bt (x) is the ball of radius t and center x in R2 . Recalling the formula for integrals along a surface
given by a graph:
Z Z
1 1 p
v 0 dS = 2
u0 (y) 1 + |∇f (y)|2 dy,
vol(∂B t (x)) ∂B +
t (x) 4πt Bt (x)

where we used that v0 (x1 , x2 , x3 ) = u0 (x1 , x2 ). This last fact also implies that ∂B + (x) v0 dS =
R
R t
−
∂B (x)
v 0 dS,
t

Thus,
Z Z
1 2 p
v0 dS = u0 (y) 1 + |∇f (y)|2 dy
vol(∂B t (x)) ∂B t (x) 4πt2 Bt (x)
Z
1 u0 (y)
= p dy.
2πt Bt (x) t2 − (y − x)2

References
[1] E. Butkov. Mathematical Physics. Addison-Wesley Educational Publishers Inc; New edition, 1973.
[2] L. D. Faddeev and O. A. Yakubovskiĭ. Lectures on quantum mechanics for mathematics students, volume 47 of
Student Mathematical Library. American Mathematical Society, Providence, RI, 2009. Translated from the 1980
Russian original by Harold McFaden, With an appendix by Leon Takhtajan.
[3] Leon A. Takhtajan. Quantum mechanics for mathematicians, volume 95 of Graduate Studies in Mathematics.
American Mathematical Society, Providence, RI, 2008.
[4] Steven Weinberg. Lectures on quantum mechanics. Cambridge University Press, Cambridge, second edition, 2015.