Lecture Notes on PDEs, part II:

Laplace’s equation, the wave equation and more

Fall 2018

Contents
1 The wave equation (introduction) 2
1.1 Solution (separation of variables) . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Standing waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Solution (eigenfunction expansion) . . . . . . . . . . . . . . . . . . . . . . . 6

2 Laplace’s equation 8
2.1 Solution in a rectangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Rectangle, with more boundary conditions . . . . . . . . . . . . . . . . . . . 11
2.2.1 Remarks on separation of variables for Laplace’s equation . . . . . . . 13

3 More solution techniques 14

3.1 Steady states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Moving inhomogeneous BCs into a source . . . . . . . . . . . . . . . . . . . 17

4 Inhomogeneous boundary conditions 20

4.1 A useful lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 PDE with Inhomogeneous BCs: example . . . . . . . . . . . . . . . . . . . . 21
4.2.1 The new part: equations for the coefficients . . . . . . . . . . . . . . 21
4.2.2 The rest of the solution . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Theoretical note: What does it mean to be a solution? . . . . . . . . . . . . 23

5 Robin boundary conditions 24

5.1 Eigenvalues in nasty cases, graphically . . . . . . . . . . . . . . . . . . . . . 24
5.2 Solving the PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

6 Equations in other geometries 29

6.1 Laplace’s equation in a disk . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

7 Appendix 31
7.1 PDEs with source terms; example . . . . . . . . . . . . . . . . . . . . . . . . 31
7.2 Good basis functions for Laplace’s equation . . . . . . . . . . . . . . . . . . 33
7.3 Compatibility conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

1
1 The wave equation (introduction)
The wave equation is the third of the essential linear PDEs in applied mathematics. In one
dimension, it has the form
utt = c2 uxx
for u(x, t). As the name suggests, the wave equation describes the propagation of waves, so
it is of fundamental importance to many fields. It describes electromagnetic waves, some
surface waves in water, vibrating strings, sound waves and much more.

Consider, as an illustrative example, a string that is fixed at ends x = 0 and x = L. It

has a constant tension T and linear mass density (i.e. mass per length) of λm . Assuming
gravity is negligible, the vertical displacement u(x, t), if it is not too large, can be described
by the wave equation:
utt = c2 uxx , x ∈ (0, L)
p
with c = T /λm . The derivation is standard (see e.g. the book). Suppose that the string
has, at t = 0, an initial displacement f (x) and initial speed of g(x). We’ll leave f and g
arbitrary for now. The initial boundary value problem for u(x, t) is

utt = c2 uxx , x ∈ (0, L), t inR

u(0, t) = 0, u(L, t) = 0, (1)
u(x, 0) = f (x), ut (x, 0) = g(x).

Note that we have two initial conditions because there are two time derivatives (unlike the
heat equation). A sketch and the domain (in the (x, t) plane) is shown below. Note that we
do not restrict t > 0 as in the heat equation.

How many boundary conditions are needed? Typically, for a PDE, to get a unique
solution we need one condition (boundary or initial) for each derivative in each variable. For
instance:
ut = uxx =⇒ one t-deriv, two x derivs =⇒ one IC, two BCs

2
and
utt = uxx =⇒ two t-derivs, two x derivs =⇒ two ICs, two BCs
In the next section, we consider Laplace’s equation uxx + uyy = 0:

uxx + uyy = 0 =⇒ two x and y derivs =⇒ four BCs.

1.1 Solution (separation of variables)

We can easily solve this equation using separation of variables. We look for a separated
solution
u = h(t)φ(x).
Substitute into the PDE and rearrange terms to get

1 h00 (t) φ00 (x)

= = −λ.
c2 h(t) φ(x)

This, along with the boundary conditions at the ends, yields the BVP for φ:

φ00 + λφ = 0, φ(0) = φ(L) = 0,

which has solutions

nπx
φn = sin , λn = n2 π 2 /L2 .
L
Now, for each λn , we solve for the solution gn (t) to the other equation:

φ00n + c2 λn φn = 0.

There are no initial conditions here because neither initial condition is separable; the initial
conditions will be applied after constructing the full series (note that if, say, g = 0 or f = 0
then we would have a condition to apply to each φn ).

The solution is then

nπct nπct
φn = an cos + bn sin .
L L
Thus the separated solutions are
nπct nπct nπx
un (x, t) = (an cos + bn sin ) sin .
L L L

The full solution to the PDE with the boundary conditions u = 0 at x = 0, L is a superpo-
sition of these waves:
∞
X nπct nπct nπx
u(x, t) = (an cos + bn sin ) sin . (2)
n=1
L L L

3
To find the coefficients, the first initial condition u(x, 0) = f (x) gives
∞
X nπx
f (x) = an sin .
n=1
L

and the second initial condition ut (x, 0) = g(x) gives

∞
cπ X nπx
g(x) = nbn sin .
L n=1 L

Both are Fourier sine series, so we easily solve for the coefficients and find

2 L
Z Z L
nπx 2 nπx
an = f (x) sin dx, bn = g(x) sin . dx. (3)
L 0 L ncπ 0 L

1.2 Standing waves

Consider a string of length π (for simplicity) and fixed ends. The separated solutions have
the form
un (t) = (an cos nct + bn sin nct) sin nx.
These solutions are called standing waves, because they have points at which the string
’stands’ still. Plotting at some time t look like

The solutions oscillate over time, but observe that

kπ
un (t) = 0 where x = , k = 0, 1, 2 · · · , n.
n
The points kπ/n are called the nodes. Note also that the amplitude is largest at the points
where sin nx = ±1, namely at
nx = π/2 + kπ.
These are called the anti-nodes.
In fact, with some effort we can show that the solution to the wave equation is really a
superposition of two waves travelling in opposite directions, reflecting off the boundaries and

4
interfering with each other. At the nodes, they cancel out exactly (destructive interference)
and at the anti-nodes, they add together exactly (constructive interference).
For the separated solutions, this is easy to show:
1 1 1
cos nct sin nx = (sin n(x + ct) + sin n(x − ct)) = hn (x + ct) + hn (x − ct).
2 2 2
It is straightforward to check that both parts of the sum are solutions to the wave equation
(’travelling waves’) although they do not individually satisfy the boundary conditions.
We can be much more general about this (it is not just true for standing waves); see
later.

Standing waves: Defining the fundamental frequency in radians per time,

ω0 = πc/L,

we can also write hn in terms of this frequency and its multiples (the harmonics):

hn = an cos ωn t + bn sin ωn t, ωn = nω0 .

Physically, these correspond to standing waves, which oscillate at the frequency ωn and have
fixed nodes (with u = 0) at equally spaced points x = Lm/nπ for m = 0, · · · n. Note that
the ’fundamental frequency’ is typically written in cycles per time, so
c
f0 = , fn = nf0 .
2L

Example: plucking a string

Suppose a string from a guitar or harp is plucked. The initial displacement will be something
like a triangular shape, such as
(
2Ax/L 0 ≤ x < L/2
f=
2A(L − x)/L L/2 < x < L

where A is the initial displacement at x = L/2. The initial speed is g = 0. In that case, it is
straightforward to show that bn = 0 and
8A nπ
an = 2 2
sin .
π n 2
In terms of the frequencies ωn , the response of the string is
∞
X nπx
u(x, t) = an cos(2πωn t) sin .
n=1
L

5
Since sin nπ/2 = 0 for n even, the string, when plucked exactly at the center, vibrates with
only the odd harmonics, and the amplitude of the harmonics decay quadratically with n.
Note that because the initial displacement is not an eigenfunction, there are an infinite num-
ber of harmonics present. For a musical instrument, this is ideal, since the sound is much
better when it is a mix of frequencies (a pure tone of one frequency is not pleasant).

Superposition of initial conditions: The principle of superposition gives some insight

into the parts of the solution (2). We can split the IBVP (1) into one part with zero initial
velocity (ut = 0) and one with zero initial displacement (u = 0). To be precise, let v solve

vtt = c2 vxx , t > 0, x ∈ (0, L),

v(0, t) = 0, v(L, t) = 0, t > 0 (4)
v(x, 0) = f (x), vt (x, 0) = 0

and let w solve

wtt = c2 wxx , t > 0, x ∈ (0, L),
w(0, t) = 0, w(L, t) = 0, t > 0 (5)
w(x, 0) = 0, wt (x, 0) = g(x).

Then the solution u(x, t) to (1) is

u = v + w.
Notice that the two pieces correspond to the sine/cosine terms in the full solution:
∞   X ∞  
X nπct nπx nπct nπx
u(x, t) = an cos sin + bn sin sin .
n=1
L L n=1
L L
| {z } | {z }
v(x,t) w(x,t)

To see this, note that from (3), if g(x) = 0 then the bn ’s are zero; if f (x) = 0 then the an ’s
are zero. It follows that the two terms above are really v and w.

1.3 Solution (eigenfunction expansion)

The derivation using an eigenfunction expansion follows the same pattern as the heat equa-
tion. Again, let us consider the Dirichlet problem (1). Write the PDE as

utt = −L[u], Lu = uxx . (6)

The eigenfunction problem for φ(x) is

Lφ = λφ, φ satisfies the BCs

6
which is just the familiar eigenvalue problem from the heat equation,

−φ00 = λφ, φ(0) = φ(L) = 0.

The eigenvalues/functions are

nπx
λn = n2 π 2 /L2 , φn = sin , n ≥ 1.
L
The solution therefore has an eigenfunction expansion
∞
X
u(x, t) = hn (t)φn (x).
n=1

Plug into the PDE:

∞
X ∞
X
h00n (t)φn (x) = hn (t)φ00n (x)
n=1 n=1
X∞
=− λn hn (t)φn (x).
n=1

Rearrange to get
∞
X
(h00n (t) + λn hn (t))φn (x) = 0
n=1
so
h00n (t) + λn hn (t) = 0 for n ≥ 1.
From here, the solution is the same as for separation of variables - solve for φn , then apply
the initial conditions.

7
2 Laplace’s equation
In two dimensions the heat equation1 is

ut = α(uxx + uyy ) = α∆u

where ∆u = uxx + uyy is the Laplacian of u (the operator ∆ is the ’Laplacian’). If the
solution reaches an equilibrium, the resulting steady state will satisfy

uxx + uyy = 0. (7)

This equation is Laplace’s equation in two dimensions, one of the essential equations in
applied mathematics (and the most important for time-independent problems). Note that
in general, the Laplacian for a function u(x1 , · · · , xn ) in Rn → R is defined to be the sum of
the second partial derivatives:
n
X ∂ 2u
∆u = .
j=1
∂x2j

Laplace’s equation is then compactly written as

∆u = 0.

The inhomogeneous case, i.e.

∆u = f
the equation is called Poisson’s equation. Innumerable physical systems are described by
Laplace’s equation or Poisson’s equation, beyond steady states for the heat equation: invis-
cid fluid flow (e.g. flow past an airfoil), stress in a solid, electric fields, wavefunctions (time
independence) in quantum mechanics, and more.

The two differences with the wave equation

utt = c2 uxx

are:

• We specify boundary conditions in both directions, not initial conditions in t.

• There is an opposite sign; we have uxx = −uyy rather than utt = c2 uxx .

The first point changes the way the problem is solved slightly; the second point changes the
answer. Note that there is also no coefficient, but this is not really important (we can just
as easily solve uxx + k 2 uyy = 0).
1
The derivation follows the same argument as what we did in one dimension.

8
2.1 Solution in a rectangle
We can solve Laplace’s equation in a bounded domain by the same techniques used for the
heat and wave equation.

Consider the following boundary value problem in a square of side length 1:

0 = uxx + uyy , x ∈ (0, 1), y ∈ (0, 1)
u(x, 0) = 0, u(x, 1) = 0, x ∈ (0, 1)
u(0, y) = 0, u(1, y) = f (y), y ∈ (0, 1).

The boundary conditions are all homogeneous (shown in blue above) except on the right
edge (y = 0)). Motivated by this, we will try to get eigenfunctions φ(y), since the eigenvalue
problem requires us to impose homogeneous boundary conditions.

Look for a separated solution

u = g(x)φ(y).
Substitute into the PDE to get
0 = g 00 (x)φ(y) + g(x)h00 (y)
and then separate:
h00 (y) g 00 (x)
− = = λ.
φ(y) g(x)
This leads to the pair of ODEs
φ00 (y) + λφ(y) = 0, g 00 (x) = λg(x).
Applying the boundary conditions on the sides x = 0 and x = 1, we get the BVP
φ00 (y) + λφ(y) = 0, φ(0) = φ(1) = 0.
We know the solutions to the above; they are
φn (y) = sin nπy, λn = n2 π 2 , n ≥ 1.

9
Now we solve for g for each λn . Note that there is only one boundary condition (at x = 1);
we leave the f (x) condition for later (it will require using the full series). We solve

g 00 − n2 π 2 g = 0, g(0) = 0

to get
gn (x) = an sinh nπx.
The solution
un = gn (x)φn (y)
satisfies the PDE and all the boundary conditions except u(x, 0) = f (x). To satisfy this, we
need to write u as a sum all of the separated solutions gn φn :
∞
X
u(x, y) = an sinh nπx sin nπy.
n=1

Now apply u(1, y) = f (y) (the boundary condition at x = 1) to get

∞
X
f (x) = an sinh nπ sin nπy.
n=1

Note that the functions φn = sin πy are orthogonal in L2 [0, 1] (we have shown this several
times at this point!). As always, take inner products of both sides with hm = sin mπy to get
the coefficients:
hf, hm i = (am sinh mπ) hhm , hm i
so
1 hf, φn i
an = , n ≥ 1.
sinh nπ hφn , φn i

Remark (convergence): In terms of the eigenfunction basis (set φn = sin nπy; the basis
is {φn (y)}), we have
∞
X
u(x, y) = an sinh nπxφn (y).
n=1

Unlike the heat equation, the coefficients have an sinh πx, which contains a positive expo-
nential (sinh nπx ∼ enπ|x| /2 as n → ∞). However, the an ’s have a sinh nπ in the denominator,
which makes sure the coefficients are small enough that the series convergences. it can be
shown that
sinh nπx
an = decays exponentially as n → ∞
sinh nπ
for x in the domain. This is true since as n → ∞, the numerator grows like enπ|x| /2 while
the denominator grows like enπ /2, so if |x| < 1 the rate is faster for the denominator.

10
2.2 Rectangle, with more boundary conditions
Let’s return to the rectangle example and consider how to solve the problem when there
are inhomogeneous boundary conditions applied at all the sides for Laplace’s equation in a
rectangle of width A and height B:
0 = uxx + uyy , x ∈ (0, a), y ∈ (0, b)
u(x, 0) = f1 (x), u(x, 1) = f2 (x), x ∈ (0, A) (8)
u(0, y) = g1 (y), u(1, y) = g2 (y), y ∈ (0, B).

Both pairs of opposite sides (in blue and red above) could have non-homogeneous BCs. Our
method only works if one of those pairs is homogeneous.
To solve (8), we use superposition and break the problem up into parts. Each part will
take care of one (or two) of the boundaries and leave all the others zero. When added to-
gether, the sum of the parts will satisfy all the boundary conditions.

We find v, w solving
0 = vxx + vyy , x ∈ (0, A), y ∈ (0, B)
v(x, 0) = 0, v(x, B) = 0, x ∈ (0, A) (9)
v(0, y) = g1 (y), v(A, y) = g2 (y), y ∈ (0, B).

0 = wxx + wyy , x ∈ (0, A), y ∈ (0, B)

w(x, 0) = f1 (x), w(x, B) = f2 (x), x ∈ (0, A) (10)
w(0, y) = 0, w(A, y) = 0, y ∈ (0, B).
The sum u = v + w is then the solution to (8). The solutions u along with v, w for a specific
choice of initial condition are shown in Figure 1.

Solving for v: To solve (9), look for a separated solution

v = h(x)φ(y).

This leads to the boundary value problem

φ00 + λφ = 0, φ(0) = φ(b) = 0.

11
The solutions are
nπy
φn (y) = sin , λn = n2 π 2 /B 2 .
B
There are no boundary conditions we can apply for h (both boundaries have inhomogeneous
terms), which satisfies
h00n − λn hn = 0,
so we take the general solution. Set µn = nπ/a. The right choice of basis for solutions to the
ODE has one basis function vanish at x = 0 and the other at x = A:

hn (x) = an sinh(µn (A − x)) + bn sinh µn x.

See subsection 7.2 for details. Adding up all the separated solutions, the solution for v is
∞ h
X i
v(x, y) = an sinh(µn (A − x)) + bn sinh(µn x) φn (y).
n=1

Now the choice of basis becomes useful because it makes only one set of coeffiients appear
at each boundary. At x = 0:
∞
X
g1 (y) = v(0, y) = an sinh(µn )φn (y)
n=1

and so
B
hg1 , φn i
Z
2 nπy
an sinh µn = = g1 (y) sin dy.
hφn , φn i B 0 B
2
where hf, gi is the inner product on L [0, B]. At the x = A boundary:
∞
X
g2 (y) = v(A, y) = bn sinh(µn A)φn (y)
n=1

and so
B
hg2 , φn i
Z
2 nπy
bn sinh(µn A) = = g2 (y) sin dy.
hφn , φn i B 0 B
Finding w that solves (10) is the same process, and one gets a similar expression (left as an
exercise). Finally, the solution to the original problem (8) is

u = v + w.

12
Figure 1: Solution u to (8) and the solutions v and w to (9) and (10) for f1 = f2 = x(1 − x)
and g1 = g2 = y(1 − y).

2.2.1 Remarks on separation of variables for Laplace’s equation

Just as with the heat equation, if there are more complicated inhomogeneous terms, e.g.

0 = uxx + uyy + f (x, y)

then the eigenfunction method is required unless you are lucky and there is a ‘particular’
solution you can subtract out to remove the inhomogeneous terms.

When applying the eigenfunction method, one must pick a direction for the eigenfunctions,
either X X
u= cn (x)φn (y) or u = cn (y)φn (x).
The correct choice is one where the boundary conditions are homogeneous (if both work, then
it does not matter which you choose). The details are somewhat involved but straightforward
in concept.

13
Non-separable boundary conditions: There is a more fundamental concern: the
geometry of the problem. Notice that in order to ‘separate’ the problem, we needed each
boundary condition to involve only one coordinate. This was true, for instance, for the
rectangle (one edge is x = const, the other is y = const) and for a circle, annulus or wedge
(which have edges r = const or θ = const).

Dealing with boundary conditions that are not separable - for instance, flow of water
past a fish (at least a non-spherical fish) - is challenging and requires more sophisticated
techniques.

3 More solution techniques

The following techniques are all ways of reducing more complicated problems to simpler
ones. The ’simple’ problems are

a) a homogeneous PDE and homogeneous BCs

b) an inhomogeneous PDE (i.e. with a source term) and homogeneous BCs

Both (a) and (b) can be solved using eigenfunction expansions; (a) is even simpler and can
be solved using separation of variables.2 Note that ’homogeneous BCs’ here means that there
are enough homogeneous BCs to get eigenfunctions; the other boundaries may be allowed to
stay inhomogeneous.

3.1 Steady states

We convert an inhomogeneous heat equation to a homogeneous problem when the inho-
mogeneous terms are all time-independent. In doing so, we obtain an easy method for
determining the limit of the solution as t → ∞.

Consider the IBVP

ut = uxx + h(x), x ∈ [0, `], t > 0
u(0, t) = A, u(`, t) = B, t > 0 (11)
u(x, 0) = f (x)

which represents heat flow with a time-independent source and/or ends fixed at some temper-
ature. The expectation is that over time, the heat will approach a steady state (equilibrium):

u(x) = lim u(x, t).

t→∞

2
There is a method for reducing (b) to (a) that we will not cover here since we can solve (a); this method
is called Duhamel’s principle.

14
 Formally, we can obtain this equilibrium shape as follows: if u(x) is a steady-state, then
it solves the PDE but does not depend on time. Thus it must satisfy

0 = uxx + h(x), u(0) = A, u(`) = B.

This we can then solve. The important point is that the difference betwee the PDE solution
and steady state,
v =u−u
solves the homogeneous IBVP

vt = vxx , x ∈ [0, `], t > 0

v(0, t) = 0, v(`, t) = 0, t > 0 (12)
v(x, 0) = f (x) − u(x)

So to solve (11) we can find the steady state (formally), subtract it out and then solve (12)
for the ‘homogeneous’ part.

Note that the inhomogeneous term could appear in the source or in the boundaries.

Example: Consider the inhomogeneous problem

ut = uxx + 1, x ∈ (0, 1), t > 0

u(0, t) = 0, u(1, t) = 0, t > 0
u(x, 0) = f (x).

There is a constant input source of heat, the heat is kept fixed at both ends and there some
initial distribution f (x) of heat. We want to show that the heat distribution converges to
some equilibrium shape as t → ∞. This is done in two steps:

1. Compute the steady state, assuming it exists: From knowledge of the heat equa-
tion, we expect a steady state to exist (diffusion wants to spread things out). Let us assume
that it exists, i.e. there is a function u(x) such that

u(x) = lim u(x, t). (13)

t→∞

Plug u(x) into the PDE and BCs to get

uxx = −1, u(0) = u(1) = 0

which is easily solved (directly) to obtain

1
u(x) = x(1 − x). (14)
2

15
Key point: At this point, we have shown that if a steady state exists, it must be (14). It
does not yet follow that the limit (13) exists!

2. Show directly that it is really a steady state: Now let

v(x, t) = u(x, t) − u(x).

The function v is the difference between the solution and the steady state, so we want to
show that v → 0 as t → ∞. Note that u is the solution to the IBVP

ut = uxx + 1, x ∈ (0, 1), t > 0

u(0, t) = 0, u(1, t) = 0, t > 0
u(x, 0) = u(x)

By linearity/superposition v satisfies

vt = vxx , x ∈ (0, 1), t > 0

v(0, t) = 0, v(1, t) = 0, t > 0
v(x, 0) = f (x) − u(x)

which is the difference of the IBVP satisfied by u and the one satisfied by u.

But this equation is just the heat equation (homogeneous) with Dirichlet boundary con-
ditions; the solution is
∞
X
v(x, t) = bn e−λn t sin λn x (15)
n=1
2 2
where λn = n π . and
Z 1 Z 1
bn = 2 v(x, 0) sin nπx dx = 2 (f (x) − u(x)) sin πx dx.
0 0

The solution to the inhomogeneous problem (14) is then

∞
1 X
u(x, t) = u + v = x(1 − x) + bn e−λn t sin λn x.
2 n=1

To verify that u is actually the steady state, note that all the eigenvalues are positive, so

lim u(x, t) = u(x) + lim v(x, t) = u(x).

t→∞ t→∞

16
When does the method fail? This trick works when there is a steady state but only
when the source term and boundary conditions do not depend on time. For instance,

ut = tuxx + sin x

cannot be solved using this method. Assuming ut = 0 is not enough since we also need to
take t → ∞ and we cannot find a u = u(x) that solves

tuxx + sin x = 0.

We can, however, solve the full problem using the eigenfunction method and find the steady
state directly.

3.2 Moving inhomogeneous BCs into a source

Suggesiton: We will need to use the eigenfunction expansion method here; a review example
is included in the appendix (subsection 7.1). We can use a trick similar to the steady state to
remove inhomgeneous boundary conditions. It is too much to try to find a simple function
that satisfies the PDE and the BCs simultaneously. However, we can instead compromise
and find a function w that satisfies the boundary conditions but not the PDE. Then

v =u−w

will satisfy the homogeneous boundary conditions but will have both a different IC and
an extra source term. For example, consider the heat equation with time-dependent
boundary conditions:

ut = uxx + f (x, t), x ∈ [0, 1], t > 0,

u(0, t) = g(t), u(1, t) = h(t), t > 0,
u(x, 0) = u0 (x).

There are many choices for a function w that satisfies the boundary conditions. For Dirichlet
BCs, the easiest is just to construct a line that goes from (0, g(t)) to (1, h(t)):

w(x, t) = (1 − x)g(t) + xh(t).

Now define v(x, t) = u(x, t) − w(x, t). Then

v(0, t) = u(0, t) − w(0, t) = g(t) − g(t) = 0,

and similarly for v(1, t), so

v(0, t) = v(1, t) = 0.
Now we find the PDE for v by linearity/superposition. First, we need to find the ’source’
term for w. To do so, plug w into the PDE wt = wxx ; what is left-over is the source term.
Explicitly, we have
wt = wxx + g(x, t)

17
where
g = wt − wxx = (1 − x)g 0 (t) + xh0 (t).
It follows that
vt = ut − wt = uxx + f − (wxx + g) = vxx + f˜
where the new sourcce term for v is
f˜(x, t) = f − g = f − (1 − x)g 0 (t) − xh0 (t).
Finally, for the initial condition,
v(x, 0) = u(x, 0) − w(x, 0) = u0 (x) − (1 − x)g(0) − xh(0).
The problem to solve for v is then
vt = vxx + f˜(x, t), x ∈ [0, 1], t > 0,
v(0, t) = 0, v(1, t) = 0, t > 0,
v(x, 0) = u0 (x) − (1 − x)g(0) − xh(0).
We obtain v using the eigenfunction expansion method and then add w back in to get
u = v + w.

Illustrative example
We solve
ut = uxx , x ∈ (0, π), t > 0,
u(0, t) = 0, u(π, t) = At, t > 0,
u(x, 0) = 0.
Set
Axt
w= (16)
π
and
v = u − w.
Then wt = wxx + Ax/π, so the PDE for v is
vt = ut − wt
= uxx − (wxx + Ax/π)
Ax
= vxx − .
π
Since w(x, 0) = 0, the initial condition for v is the same for u (i.e. v(x, 0) = 0). Thus the
IBVP for v we need to solve is
Ax
vt = vxx − , x ∈ (0, π), t > 0,
π
v(0, t) = 0, v(π, t) = 0, t > 0,
v(x, 0) = 0.

18
For the source term, expand x in terms of the eigenfunctions:
∞
X
x= bn φn , bn = −2 cos(nπ)/n.
n=1
P
Now solve for v = cn (t)φn (x) using an eigenfunction expansion to get

Abn
c0n (t) + λn cn (t) = − , cn (0) = 0.
π
The solution for this ODE is
Abn −λn t
cn (t) = (e − 1), (17)
λn π
so the solution to the IBVP is
∞
At X
u(x, t) = x+ cn (t)φn (x)
π n=1

with cn ’s given by (17) and bn ’s by (16). Note that the Atx/π term is not a ’particular
solution’ to the PDE, since it only satisfies the boundary conditions and not the PDE.
In the next section, we will use another method that can be used to solve the original
problem directly, which gives us u as an actual sum of a homogeneous and particular solution.

19
4 Inhomogeneous boundary conditions
When boundary conditions are inhomogeneous, the method above will not work. The issue
is that the eigenfunctions must satisfy the homogeneous boundary conditions, so it seems
X
u(x, t) = cn (t)φn (x) (18)
n

should also satisfy the homogeneous BCs. Thus we cannot just substitute u into the
PDE/BCs and solve for cn . For instance, if

u(0, t) = 0, u(π, t) = At

and φn = sin nx then X

u(π, t) = cn (t)φn (π) = 0
n

which gets us nowhere. However, the series (18) is still a good starting point.

4.1 A useful lemma

R`
Recall that the inner product on L2 [0, `] is hf, gi = 0
f (x)g(x) dx.

Lagrange’s identity [simple version]: Consider the operator

Lf = −f 00 .

If f and g are functions in L2 [0, `] then

`
0 0
hLf, gi = (f g − f g) + hf, Lgi. (19)

0

Moreover, if we have standard homogeneous boundary conditions then

hLf, gi = hf, Lgi for all f, g satisfying the BCs.

In this case we say the the operator L with the boundary conditions3 is self-adjoint.

The identity generalizes the fact that for a real symmetric matrix A and vectors x, y ∈ Rn ,

(Ax) · y = (Ax)T y = xT At y = xT (Ay) = x · (Ay).

The proof of (19) is simple; just integrate by parts twice (left as an exercise).
3
Dirichlet, Neumann or Robin, i.e. α1 f (0) + α2 f 0 (0) = 0 and β1 f (`) + β2 f 0 (`) = 0.

20
4.2 PDE with Inhomogeneous BCs: example
An example will serve to illustrate the idea. We solve the heat equation in [0, π] with a
time-dependent boundary condition:
ut = uxx , x ∈ (0, π), t > 0,
u(0, t) = 0, u(π, t) = At, t > 0, (20)
u(x, 0) = f (x).
The eigenfunctions/values are
φn = sin nx, λn = n2 , n ≥ 1.
The solution has an eigenfunction expansion
∞
X
u(x, t) = cn (t)φn (x).
n=1

The idea is to use the formula for the coefficients, the PDE and some manipulations
(including the identity (19)) to get equations for cn (t).

4.2.1 The new part: equations for the coefficients

Since the φn ’s form an orthogonal basis, we have that
hu, φn i 2
cn (t) = = hu, φn i. (21)
hφn , φn i π
Of course, u is unknown, so we need to get ODEs for cn (t). To do so, take the inner product
of the PDE with φn :
hut , φn i = huxx , φn i.
Our goal is to write everything in the form h· · · , φn i, since this will give back cn (t). The
first term is c0n (t) since the ∂t
∂
can be factored out:
∂
hut , φn i = hu, φn i = c0n (t).
∂t
Letting Lu = −uxx , the second term is −hLu, φn i. It follows that
π 0
c (t) = hut , φn i
2 n
= −hLu, φn i
π
= (uφ0n − ux φn ) − hu, Lφn i (using the identity (19))

0
= −At cos nπ − hu, Lφn i (using the BCs)
= −At cos nπ − hu, λn φn i (since φn is an eigenfunction)
This gives
π 0 π
cn (t) = −At cos nπ − λn hu, φn i = −At cos nπ − cn (t). (22)
2 2
using the formula for the coefficients (21) again.

21
Boundary term details: We know that u has boundary conditions

u(0, t) = 0, u(π, t) = At

and φn (by definition) has

φn (0) = φn (π) = 0, φ0n (0) = n, φ0n (π) = cos nπ

which lets us simplify the boundary term:

π
(uφ0n − ux φn ) = u(π)φ0n (π) − u(0)φ0n (0) = At cos nπ.

0

Because of the inhomogeneous BC, the boundary term is not zero!

4.2.2 The rest of the solution

Now that we have equations for the cn ’s, the rest of the solution work the same way as
before. For brevity, set
2An cos(nπ)
γn = − . (23)
π
The ODE for cn is then
c0n (t) + λn cn (t) = γn t
As before, write the initial condition in terms of the eigenfunction basis:
∞
2 π
X Z
f (x) = an φn (x), an = f (x) sin nx dx. (24)
n=1
π 0

Then u(x, 0) = f (x) gives the initial condition for cn :

cn (0) = an .

To solve, use an integrating factor:

(eλn t cn )0 = γn eλn t t

to obtain Z t
−λn t −λn t
c n = an e + γn e eλn s s ds.
0
Evaluating the integral we get
γn
cn (t) = an e−λn t + λn t − 1 + e−λn t .


2
λn
While not required, we can plug in λ2n and γn from (23) we get
2 2A cos(nπ)  2 −n2 t

cn (t) = an e−n t − n t − 1 + e . (25)
πn3
22
The solution is then given by
∞
X
u(x, t) = cn (t)φn (x)
n=1

with cn (t) given by (25) and the an ’s by (24). Note that the first term in the expression (25)
for cn (t) gives the homogeneous solution; the second term is the response to the inhomoge-
neous boundary conditions:
∞ ∞
X
−n2 t 2A X cos nπ  2 −n2 t

u(x, t) = an e φn (x) − n t − 1 + e φn (x).
n=1
π n=1 n3

In fact, this solution is the same as the one we derived using the superposition trick in
section 3.2.

4.3 Theoretical note: What does it mean to be a solution?

Consider the solution we just found for the IBVP (20) with inhomogeneous boundary con-
ditions. It has the form ∞
X
u(x, t) = cn (t)φn (x). (26)
n=1

However, the eigenfunctions satisfy the homogeneous boundary conditions:

φn (0) = φn (π) = 0.

It follows that u(0, t) = 0 and u(π, t) = 0. But we solved for u using the inhomogeneous
boundary conditions! The apparent contradiction is due to the fact that the equality in the
representation (26) is not pointwise; it is equal in the same sense we had for Fourier series.
Thus when we say that (26) is a solution to the IBVP, it does not mean that the series
satisfies the boundary conditions at the endpoints. However, the right/left limits satisfy
the left/right BCs For example,, for (26) solving (20) we have

lim u(x, t) = 0, lim u(x, t) = At

x→0+ x→π −

even though u(π, t) = 0. This means that the series for u, close to the boundary, is correct;
it just might be incorrect at the boundary exactly). Thus the fact that the series is wrong
at the boundaries is not much of a worry (if handled carefully).

23
5 Robin boundary conditions
Returning to homogeneous problems, we now solve the heat equation with a boundary
condition involving u and ux (Robin). Two new issues arise:
• We will be unable to get explicit solutions for the eigenvalues λn , so a graphical method
must be used to find and estimate them.
• There can be eigenvalues in all of the cases (negative and positive!).
Consider a metal bar of length 1 with temperature fixed at one end, with the other end being
heated. Suppose the temperature u(x, t) satisfies

ut = uxx , for x ∈ (0, 1), t > 0 (27)

with boundary conditions

u(0, t) = 0, ux (1, t) − βu(1, t) = 0 (28)

and initial condition
u(x, 0) = f (x).
Here β > 0 describes the strength of the heating; the flux at the boundary is

−ux = −βu.

Heat is added at a rate βu if u > 0 (and lost if u < 0). The value of β determines the
balance between inflow and outflow of heat, so the solution behavior as t → ∞ will depend
on β; we will be able to determine it by looking at the eigenvalues.

5.1 Eigenvalues in nasty cases, graphically

The eigenvalue problem is

− φ00 = λφ, φ(0) = 0, φ0 (1) − βφ(1) = 0 (29)

For this problems, the eigenvalues cannot be obtained exactly. Instead, we need to use a
graphical argument to locate them.

Negative eigenvalues: First, suppose λ < 0. Then the general solution is

φ = c1 eµx + c2 e−µx
√
where µ = −λ. Since φ(0) = 0, we need c1 = −c2 , so

φ = eµx − e−µx .

Now impose the Robin boundary condition:

0 = µ(eµ + e−µ ) − β(eµ − e−µ ).

24
We cannot solve for µ here. Instead, rearrange to get
e2µ + 1
β=µ := g(µ).
e2µ − 1
A solution exists for each µ > 0 such that g(µ) = β (draw a horizontal line at β across the
graph of g(µ) for µ > 0. From a plot of g(µ) (check that g(µ) > 1 for µ > 0), it is clear that
there are no positive solutions for β ≤ 1 and exactly one when β > 1.

No zero eigenvalue: If λ = 0 the general solution is

φ = c1 x + c2 .
The boundary condition φ(0) = 0 forces c2 = 0 so φ = c1 x. Imposing the other boundary
condition,
c1 − βc1 = 0
which has no solutions (except zero) when β 6= 1 but has a solution when β = 1; the eigen-
function is then φ = x.
√
Positive eigenvalues: Finally, If λ > 0 then (with µ = λ)
φ = c1 sin µx + c2 cos µx.
The boundary condition φ(0) = 0 requires c2 = 0. The other condition gives
µ cos µ − β sin µ = 0.
Rearranging, we need µ that satisfies
µ
= tan µ. (30)
β
There is no exact solution, but the solutions are the intersections of the line µ/β and tan µ
at positive values µ.

Note that tan µ has asymptotes at π/2 + nπ. Let I0 = (0, π/2) and let In = (π/2 +
(n − 1)π, π/2 + nπ be the intervals for each of the branches of tan µ (for µ > 0.

First note that in I0 , tan µ starts at 0 and increases to ∞. Since (tan µ)0 = 1 at µ = 0
and (µ/β)0 = 1/β > 1, the line starts above tan µ then intersects it once in I0 (see plot).

For the other intervals, note that tan µ is one-to-one and goes from −∞ to ∞ in each
In , so the line µ/β clearly intersects the branch at a unique point in In .

It follows that there is a sequence of positive solutions µn to (30) with

0 < µ0 < π/2, π/2 + (n − 1)π < µn < π/2 + nπ for n ≥ 1.
The corresponding eigenvalues and eigenfunctions are
λn = µ2n , φn = sin µn x for n = 1, 2, · · · .

25
5.2 Solving the PDE
In all cases, there are eigenvalues λn > 0 for n ≥ 1 with eigenfunctions
p
φn = sin λn x, n = 1, 2, · · · .
However, the first eigenvalue is different for each of the three cases, which determines the
behavior of the solution. The details of the eigenvalue problem are in subsection 5.1.

Summary: solution to the eigenvalue problem for (27)-(28) The eigenvalues depend
on β. For all β 6= 0, there is a sequence

0 < λ1 < λ2 < · · · → ∞

of eigenvalues with eigenfunctions

p
φn = sin λn x, quad n = 1, 2, · · · .

• If β < 1 there are no other eigenvalues.

• If β = 1 there is a zero eigenvalue:

λ0 = 0, φ0 = x.

• If β > 1 there is no zero eigenvalue but there is a single negative eigenvalue:

p
λ0 < 0, ψ0 = sinh( −λ0 x).

Case 1 (β < 1; decay):

The solution has the form ∞
X
u(x, t) = an (t)φn (x).
n=1
Following the same steps as ?? (note that the eigenfunctions are different, but the steps are
exactly the same here), we get
a0n (t) + λn an (t) = 0 =⇒ an (t) = bn e−λn t .
Thus the solution to the PDE (27) with the BCs (28) is
∞
X p
u(x, t) = an e−λn t sin λn x. (31)
n=1

Finally, to satisfy the initial condition, evaluate the series at t = 0 and set it equal to f (x):
∞
X p
f (x) = an φn , with φn = sin λn x. (32)
n=1

26
The φn ’s are an orthogonal basis (a result of the theorem stated in our original discussion of
eigenvalue problems), so we simply take the inner product with φm on both sides to get
∞
X
hf, φm i = an hφn , φm i = am hφm , φm i.
n=1

Thus the constants an are given by

R1
hf, φn i 0
f (x)φn (x) dx
an = = R1 . (33)
hφn , φn i φ2 dx
n
0

For completeness, we could evaluate the denominator by direction integration. Define

Z 1 Z 1 Z 1 √
2 2
p 1 − cos λn x 1 1 p
κn = φn dx = sin λn x dx = dx = − √ sin λn .
0 0 0 2 2 2 λn
Then, explicitly, the coefficients an are given by
Z 1
1 p
an = f (x) sin λn x dx.
κn 0

Case 2 (β = 1; steady state):

There is an extra eigenvalue λ0 = 0 and eigenfunction φ0 = x. The solution has the form
∞
X
u(x, t) = a0 (t)φ0 (x) + an (t)φn (x).
n=1

Solving for the coefficients an (t) is exactly the same; we get

an (t) = bn e−λn t for n = 1, 2, · · · .

Since L[φ0 ] = 0, the n = 0 term yields

a00 (t) = 0.

It follows that the solution is

∞
X p
u(x, t) = b0 x + bn e−λn t sin λn x.
n=1

To solve for b0 , use the initial condition. We require

f (x) = b0 x + (terms orthogonal to x).

Taking the inner product with x, we get

R1 Z 1
0
xf (x) dx
b0 = R 1 =3 xf (x) dx.
x 2 dx 0
0

27
Since λn > 0 except for n = 0, all the terms past n = 0 vanish (quickly) as t → ∞. This
suggests that
lim u(x, t) = a0 x
t→∞
i.e. the solution converges to the ’steady state’ a0 x Note that if it happens to be true that
Z 1
xf (x) dx = 0
0
then a0 = 0 and the solution converges to zero. That is, the zero mode is responsible for the
convergence to a non-zero steady state; without it, solutions will just go to zero.

Case 3 (β > 1; growth):

The solution is ∞
X p
−λ0 t
u(x, t) = b0 e sinh µ0 x + bn e−λn t sin λn x.
n=1
By the theorem, φ0 = sinh µ0 x is orthogonal to the other basis functions, so
R1
hf, φ0 i f (x) sinh µ0 x dx
b0 = = 0R 1 .
hφ0 , φ0 i sinh2 µ0 x dx
0
Since λ0 < 0, the first term grows exponentially. We call this term an unstable mode
of the system. Thus the solution u(x, t) will have exponential growth unless it happens to
be true that the initial condition has no component in the unstable mode, i.e.
Z 1
f (x) sinh µ0 x dx = 0
0
which makes b0 = 0.

Physical interpretation:
β < 1: More heat leaves through x = 0 than enters the system, so the heat decays to zero.
β = 1: There is a balance between heat entering/leaving the system, so there is a steady
state as t → ∞. Typically, this steady state will be non-zero, which occurs precisely
when Z 1
b0 = 3 xf (x) dx 6= 0.
0

β > 1 Enough heat enters that it collects in the metal bar and the temperature grows
exponentially due to the unstable mode. For most initial conditions, u(x, t) → ∞
as t → ∞ for x ∈ (0, 1). More precisely, the exponential growth rate is −λ0 . The
exception is if f does not have a φ0 component (b0 = 0)

So long as f has a φ0 component (b0 6= 0), no matter how small the coefficient
is to start, it will grow exponentially with time and eventually dominate the solution
(since all the other terms decay).

28
 7-

€
c})"\
t(
2
-(\
\,4

1
//

/
6 Equations in other geometries

5
o
2\\
.{
\-z-
r)
d-)

-a
(*-\
t[
)
6.1 Laplace’s equation in a disk
We now solve Laplace’s equation inside a disk of radius R. In polar coordinates, Laplace’s
equation ∆u = 0 for u(r, θ) in the disk becomes
1 1

o
urr + ur + 2 uθθ = 0, r ∈ (0, R), θ ∈ [0, 2π]

/-
r r

Oo
Z
l1
o
Let us assume that u is specified on the boundary of the circle:

u(R, θ) = f (θ).

g
e
tt
\<
\Y

-€
6
f,

(T)
a\
g
(-.
O
\

We start by looking for a separated solution

u = g(r)h(θ).

Substituting into the PDE we get

1 1
g 00 (r)h(θ) + g 0 (r)h(θ) + 2 g(r)h00 (θ) = 0.
r r
Divide by h(θ) and g(r) and move all the θ terms to the right:

r2 g 00 (r) + rg 0 (r) h00 (θ)

= −λ .
g(r) h(θ)
We arrive at the equations

r2 g 00 + rg 0 − λg = 0, h00 (θ) + λh(θ) = 0.

Since the inhomogeneous part f (r) is at a boundary r = R, we want to look in the other
direction first (θ) where the boundary conditions are homogeneous.

There are no explicit boundary conditions in θ; however, because θ is an angle there are
implied periodic boundary conditions

u(r, 0) = u(r, 2π), uθ (r, 0) = uθ (r, 2π).

29
Thus we need to solve

h00 (θ) + λh(θ) = 0, h(0) = h(2π), h0 (0) = h0 (2π).

This is an eigenvalue problem with periodic boundary conditions. The solutions are
h0 (θ) = a0 for λ = 0 and

hn (θ) = an cos nθ + bn sin nθ, λn = n2 , n≥1

where an , bn are arbitrary. To obtain gn , we need to solve

r2 gn00 (r) + rgn0 (r) − n2 gn (r) = 0.

Again, there are no explicit boundary conditions, but we want the solution in the disk to be
finite, so implicitly we have the condition

gn (r) is bounded , r ∈ [0, R].

The ODE is an Euler equation, so guess g = rp and obtain

p(p − 1) + p − n2 = 0 =⇒ p = ±n.

The solution is then

gn = cn rn + dn r−n .
By the boundedness condition, dn = 0. The separated solutions for u are then
a0
u0 = , un = rn (an cos nθ + bn sin nθ), n≥1
2
for arbitrary constants an and bn . Note that the cn ’s are not necessary (since an , bn are
already arbitrary) and that the constant for u0 was chosen to be a0 /2 for convenience.

The full solution is ∞

a0 X n
u(r, θ) = + r (an cos nθ + bn sin nθ). (34)
2 n=1

We now find the constants so that u satisfies the condition at r = R:

∞
a0 X n
f (θ) = u(R, θ) = + R (an cos nθ + bn sin nθ).
2 n=1

This is just the Fourier series (with coefficients Rn an and Rn bn ) for f (θ), so
Z 2π Z 2π
1 1
an = f (θ) cos nθ dθ, bn = f (θ) sin nθ dθ. (35)
πRn 0 πRn 0
Now we are done. In summary, the series (34) with coefficients (35) is the solution to
1 1
urr + ur + 2 uθθ = 0, r ∈ (0, R), θ ∈ [0, 2π]
r r
30
with boundary condition
u(R, θ) = f (θ)
and implied boundary conditions

u bounded for r ∈ [0, R], u periodic in θ.

Integral formula: This particular problem is a remarkable case where the infinite series
can be simplified. After some calculations, one ends up with Poisson’s integral formula
Z 2π
1 (R2 − r2 )f (φ)
u(r, θ) = dφ.
2π 0 R2 + r2 − 2rR cos(θ − φ)

The formula, however, is not of any use to us here besides looking nice.

7 Appendix
7.1 PDEs with source terms; example
Consider the following IBVP for the heat equation with a time-dependent source:
ut = uxx + e−t , x ∈ (0, π), t > 0,
u(0, t) = 0, u(π, t) = 0, t > 0, (36)
u(x, 0) = f (x).
Refer to the previous notes for the outline of the method.

First, we find the appropriate eigenvalues/eigenfunctions for the homogeneous problem

(as if the source term were zero). For (36), the eigenfunctions/values are

φn = sin nx, λn = n2 , n ≥ 1.

The eigenfunctions form a basis for L2 [0, π]. Now write all the functions in the PDE in terms
of the eigenfunction basis. The solution has the form
∞
X
u(x, t) = cn (t)φn (x)
n=1

for unknown cn (t)0 . The source term is

∞
X
−t
e = γn (t)φn (x).
n=1

The γn ’s are easily found through the usual formula:

he−t , φn i 2 π −t
Z
2
γn (t) = = e sin nx dx = e−t (1 − cos(nπ)).
hφn , φn i π 0 nπ

31
The inner product is an integral in x, so the dependence on t causes no trouble. For conve-
nience let us write γn (t) = an e−t where
2
an = (1 − cos(nπ)). (37)
nπ
Note that in this case we also could have just written
∞
X
1= an φn
n=1

and then multiplied by e−t instead of directly computing the expansion of e−t .

Finally, for the initial condition,

∞ Z π
X 2
f (x) = bn φn (x), bn = f (x) sin nx dx. (38)
n=1
π 0

Now plug in these series into the PDE to obtain

∞
X ∞
X ∞
X
c0n (t)φn (x) = cn (t)φ00n (x) + γn (t)φn (x) (39)
n=1 n=1 n=1
X∞
= (−cn (t)λn + γn (t))φn (x) (40)
n=1
(41)

using that φ00n = −λn φn . Since {φn } is a basis, the coefficients of φn on either side must be
equal for all n, so
c0n (t) + λn cn (t) = γn (t), n ≥ 1.
From the initial condition:
∞
X ∞
X
u(x, 0) = f (x) =⇒ cn (0)φn (x) = bn φn (x)
n=1 n=1

We then solve the IVP for the coefficients,

c0n (t) + λn cn (t) = γn (t), cn (0) = bn , (42)

to obtain Z t
−λn t −λn t
cn (t) = bn e +e eλn s γn (s) ds.
0
We solved for γn before, so the term on the right is
(
t t an
(e−t − e−λn t ) n > 1
Z Z
e−λn t eλn s γn (s) ds = an e−λn t eλn s e−s ds = λn −1
.
0 0 a1 te−t n=1

32
Highlight (word of caution): Some casework may be required in finding the cn ’s, de-
pending on λn and the inhomogeneous term γn . Here, we could get the general solution
to (42) for all n at once, but then had to do some casework to compute because γ1 (t) is a
homogeneous solution, but γn for n > 1 is not.

Finally, combining everything, the solution u(x, t) is

∞  
−t −t
X
−n2 t an −t −n2 t
u(x, t) = (b1 e + a1 te ) sin x + bn e + 2 (e − e ) sin nx dx.
n=2
n −1

with an and bn from (37) and (38). Note that

"∞ # " ∞   #
X 2
X a n 2
u(x, t) = bn e−n t sin nx + a1 te−t sin x + 2−1
(e−t − e−n t ) sin nx dx
n=1 n=2
n

The first term in square brackets depends only on the initial condition; it is the homogeneous
solution. The second term depends only on the source term (a ’particular’ solution with a
zero initial condition).

7.2 Good basis functions for Laplace’s equation

It will be useful tto recall
ex − e−x ex + e−x
sinh x = , cosh x = .
2 2
In solving Laplace’s equation we end up needing to deal with

g 00 − λg = 0
√
for positive eigenvalues λ > 0. The solutions, with µ = λ, are

g = c1 eµx + c2 e−µx .

Instead, we could choose any linear combination of the two as basis functions. Given some
boundary conditions, we want to choose the basis functions so that one of them vanishes at
each boundary condition. For instance, if the boundary conditions are

u(0, y) = f1 (y), u(1, y) = f2 (y)

Then for each n we seek basis solutions g1 , g2 such that

g1 (0) = 0, g2 (0) = 0.

For g1 , we get c2 = −c1 so

g1 = c1 (eµx − e−µx ) = 2c1 sinh x =⇒ g1 = sinh x.

33
In a similar way, we find [see homework] that
g2 = sinh(µ(x − 1)).
As another example, if the boundary conditions at x = 0 and x = 1 are
u(0, y) = 0, ux (1, y) = f (y)
then we would choose
gn (x) = an sinh(µn x) + bn cosh(µn (x − 1)).
For each boundary condition, one of the basis functions vanishes:
gn (0) = bn cosh(µn ), gn0 (1) = an cosh(µn ).

7.3 Compatibility conditions

Consider the Neumann problem for Laplace’s equation in a disk of radius R,
1 1
0 = (rur )r + 2 uθθ = 0, ur (R, θ) = f (θ). (43)
r r
Recall that the general solution to the PDE in the disk (without the condition at R) is
∞
a0 X n
u= + r (an cos nθ + bn sin nθ).
2 n=1
R 2π
In order to be able to satisfy the boundary condition ur (r, θ) = 0, we must have 0 f (θ) dθ =
0 (why?). This is called the compatibility condition, which says that the Neumann
problem only has a solution for certain constraints at the boundary.
For physical context: Laplace’s equation is a steady-state for heat conduction. If the
system is in equilibrium, nothing is entering or leaving, so the total flux through the boundary
must be zero, which is exactly the compatibility condition.
The compatibility condition can be obtained without solving the PDE. Integrating both
sides of the PDE (43) over the disk and using the boundary condition, we find that
Z 2π
0= f (θ) dθ.
0
R 2π R R
Note that the integral over the disk is 0 0
· · · r drdθ.

The compatibility condition for a general Neumann problem is that the integral of the flux
through the boundary must be zero. Notice that if this condition is satisfied, the solution
is only unique up to a constant. To get a unique solution, we would need to also specify
some other constraint (that would depend on the problem), e.g. the amount of stuff in the
domain: Z
u dA = M0 .

Contrast this with other steady state problems with Dirichlet (u = A) or radiation boundary
conditions, where there is only one solution. For the heat equation with a steady state, we
always have an initial condition that was not a Neumann condition, u(x, 0) = f (x), so the
solution (with boundary conditions) is unique and there is no compatibility condition needed.

34