You are on page 1of 9

Lecture 10

Weston Barger

July 27, 2016

1 Heat Equation on the Infinite Domain


We begin by considering the heat equation in 1 dimension without boundaries. The problem is thus
(
x ∈ (–∞, ∞)
ut = σuxx , .
u(x , 0) = f (x )

When we solved the boundary value problem, we asserted boundary values at both ends of the domain.
There are usually physical conditions imposed at x → ±∞. In the simplest case, we assume that f (x ) → 0
as x → ±∞. This means that the initial heat profile decays towards 0. Physically, this means that for all
time t > 0, u(x , t ) → 0 as x → ±∞. In this way, our problem as homogeneous “boundary” conditions. Are
problem is thus

 x ∈ (–∞, ∞)


q:homoHeatOnR} ut = σuxx , u(x , 0) = f (x ) . (1.1)


 u(±∞, t ) = 0.

Let us make the simplifying assumption that f is continuous.


We proceed by generalizing the method of separation of variables. Suppose that

{eq:heatSep} u(x , t ) = w (t )φ(x ). (1.2)

Inserting (1.2) into (1.1), we get

w 0 (t ) φ00 (x ) φ00 (x ) = λφ(x )


{eq:ODEs} = =λ ⇒ (1.3)
σw (t ) φ(x ) w 0 (t ) = λσw (t )

Recall that the solution to the ODE in φ is

r 2 = λ > 0, φ(x ) = c1 e rx + c2 e –rx


λ=0 φ(x ) = c1 x + c2
–r 2 =λ<0 φ(x ) = c1 cos(rx ) + c2 sin(rx )

If we impose the condition that φ(x ) → 0 as x → ±∞, we are left with no non-trival solutions. That is,
imposing the same “boundary” condition on φ that we do on u gives us only boring solutions. So, we relax

1
that condition to |φ(x )| < ∞ as x → ±∞. Hopefully, even with this relaxation, we will still be able to get
a function u(x , t ) which satisfies the “boundary” conditions in (1.1).
With this relaxation, we are left with

φ(x ) = c1 cos(rx ) + c2 sin(rx ), r > 0.

It should be noted that there are no conditions imposed on r other than non-negativity. When we were
solving the problem on [–L, L] we achieved a discrete spectrum i.e. λ took values in a countable set. Here,
λ takes on all negative numbers. In this case, we say that the spectrum is continuous.
Solving the ODE (1.3) in w , we get
2
w (t ) = w0 e –σr t ,

so by superposition we have
Z ∞
2
u(x , t ) = [c1 (r ) cos(rx ) + c2 (r ) sin(rx )] e –σr t dr .
0

Note that this is the continuous analog of the Fourier series. In the continuous case, it is conventional to
denote r = ω. So, we rewrite
Z ∞
2
u(x , t ) = [c1 (ω) cos(ωx ) + c2 (ω) sin(ωx )] e –σω t dω.
0

Proceeding similarly to the complex Fourier series case covered last lecture, we can set c(ω) = (c1 (ω) +
ic2 (ω))/2 and write
Z ∞
2
{eq:usorta} u(x , t ) = c(ω)e –i ωx e –σω t dω. (1.4)
–∞

Applying our initial condition yields


Z ∞
q:initApplied} f (x ) = c(ω)e –i ωx dx . (1.5)
–∞

We are still left to determine the “coefficients” c(ω). We will give a heuristic argument to compute cn .
For a continuous function f (x ), we saw that

" Z L #
X 1
f (x ) = f (y)e inπy/L dy e –inπx /L .
n=–∞
2L –L

We saw that on the finite interval [–L, L], the allowable wave number ω(= r ) were

(r =) ω = .
L
We see that
(n + 1)π nπ π
∆ω = – = ,
L L L
which implies that
1 π/L 1 π/L ∆ω
= · = = .
2L π/L 2L 2π 2π

2
Thus,

" Z L #
X 1 i ωy
{eq:riemann} f (x ) = f (y)e dy e –i ωx ∆ω. (1.6)
n=–∞
2π –L

The sum (1.6) looks like a Riemann sum in n whose “integration variable” is ω. Letting L → ∞, we see that
∆ω → 0, our partition gets smaller and smaller, and in the limit we get
Z ∞ Z ∞ 
1 i ωy
:shakeyFTrans} f (x ) = f (y)e dy e –i ωx dω. (1.7)
–∞ 2π –∞

Let us take (1.7) as fact and make the definition

Definition 1.1. Let f (x ) be a function on (–∞, ∞). We call the function


Z ∞
1
F (ω) := f (x )e i ωx dx .
2π –∞
the Fourier Transform of f (x ). We often write

F (ω) = F{f (x )}.

The Fourier transform F (ω) is analogous to the Fourier coefficient cn . F (ω) represents the amplitude of
the wave component of f (x ) with wave number ω. We make the following definition

Definition 1.2. Let F (ω) be a function on (–∞, ∞). We call the function
Z ∞
f (x ) := F (ω)e –i ωx dω.
–∞

the Inverse Fourier Transform of f (x ). We often write

f (x ) = F–1 {F (ω)}.

Proposition 1.3. We have that


f (x+ ) + f (x– )
= F–1 {F{f (x )}}.
2
When f (x ) is continuous, we have

f (x ) = F–1 {F{f (x )}}.

Note: This is analogous to the “averaging” result for Fourier series. Recall that for piecewise smooth
functions f , their Fourier series converged to (f (x+ ) + f (x– ))/2.

1.1 Solution
First, in order to complete the problem (1.1), we need to know the inverse Fourier Transform of the “bell-
shaped” curve
2
G(ω) = e –σω t .

3
This is a well-known result and we will not derive it here,
r
π – x2 2
g(x ) = e 4σt = F–1 {e –σω t }.
σt
We are now ready to find c(ω). We saw from (1.5) that
Z ∞
f (x ) = c(ω)e –i ωx dω = F–1 {c(ω)}.
–∞
Thus,

c(ω) = F{f (x )}.

Thus, by (1.4), the full solution of (1.1) is


Z ∞
2
u(x , t ) = c(ω)e –i ωx e –σω t dω
Z–∞∞ 1 Z ∞ 
2
= f (y)e i ωy dy e –i ωx e –σω t dω
–∞ 2π –∞
Z ∞Z ∞
1 2
= f (y)e –i ω(x –y) e –σω t dω dy
2π –∞ –∞
Z ∞ Z ∞ 
1 2
= f (y) e –i ω(x –y) e –σω t dω dy
2π –∞ –∞
Z ∞ r
π – (x –y)2

1
= f (y) e 4σt dy
2π –∞ σt
Z ∞
1 (x –y)2
{eq:heatsoln} = f (y) √ e – 4σt dy. (1.8)
–∞ 4πσt
This form clearly shows the solution’s dependence on the entire initial temperature distribution. The
function
1 (x –y)2
G(x , t ; y, 0) = √ e – 4σt
4πσt
is a very important function. It describes the influence that the initial heat profile f (y) has at a point (x , t ).
Since G(t , x ; y, 0) is never 0, we see that the information encoded by f travels “infinitely fast.” What I mean
is, for each t > 0, f (y) contributes to u(x , t ) for all y ∈ R. It can be shown that
1 (x –y)2
lim √ e – 4σt = δ(x – y).
t →0 4πσt
Therefore, (1.8) satisfies the initial condition.

2 Transforms of Derivatives
Using separation of variables, we motivated and defined the Fourier transform. However, if we know that
we should be using Fourier transforms, we can start in the beginning by transforming the entire PDE. For
the heat equation

ut = σuxx ⇒ F{ut } = σF{uxx }.

4
Note that since F is defined as an integral, F is linear i.e. constants can be pulled out.
We now need to compute the Fourier transform of derivatives of u. We define
Z ∞
1
U (ω, t ) = u(x , t )e i ωx dx = F{u(x , t )}.
2π –∞
Note that u(x , t ) is also a function of time. We simply keep t fixed when performing the Fourier transform.
Spacial Fourier transforms of time derivatives are not difficult. Since we are not integrating with respect to
time,
  Z ∞
∂u 1 ∂
F = u(x , t )e i ωx dx
∂t 2π –∞ ∂t
 Z ∞ 
∂ 1 i ωx
= u(x , t )e dx
∂t 2π –∞

= F{u(x , t )}
∂t
{eq:FTtime} = U t (x , t ). (2.1)

Thus, the Fourier transform of time derivatives equals the time derivative of the Fourier trans-
form.
Let us consider how Fourier transforms act on spacial derivatives. We have
  Z ∞
∂u 1 ∂u i ωx
F = e dx
∂x 2π –∞ ∂x
ue i ωx ∞ iω ∞
Z
= – u(x , t )e i ωx dx
2π –∞ 2π –∞

= –i ωF{u(x , t )}.

Therefore, we have
 
∂u
{eq:FTfirst} F = –i ωF{u} = –i ωU (ω, t ). (2.2)
∂x
In a similar manner, higher derivatives may be obtained:
 2   
∂ u ∂u
{eq:FTsecond} F = –i ωF = (–i ω)2 U (ω, t ). (2.3)
∂x 2 ∂x
In general, the Fourier transform of the k th derivatives with respect to x of a function is
(–i ω)k times the Fourier transform of the function. In mathematical notation,
( )
∂k u
{eq:FTkth} F = (–i ω)k U (ω, t ). (2.4)
∂x k

Remark 2.1. We have to be a little careful here. We obtained the relation (2.2) using integration by
parts. This argument relied on the fact that u(x , t ) → 0 as x → ±∞. We have implicitly assumed that
ux (x , t ) → 0 as x → ±∞ when deriving (2.3). One can derive (2.3) using integration by parts twice to see
j
this. Continuing, we assume that ∂∂xuj → 0 as x → ±∞ for 1 ≤ j ≤ k – 1 when we derived (2.4).
When we use Fourier transforms to solve a PDE that is k th order in x , we will find solutions
whose derivatives up to k – 1th order decay at towards infinity.

5
Now, that we have (2.1) and (2.4) we can apply the Fourier transform to a PDE. The heat equation
becomes

{eq:heat.ode} F {ut } = F {σuxx } ⇒ U (ω, t ) = σ(–i ω)2 U (ω, t ) = –σω 2 U (ω, t ). (2.5)
∂t
We have reduced the heat equation to a linear ODE.

Remark 2.2. Generally, the Fourier transform of a linear PDE with constant coefficients is a linear ODE.

The ODE (2.5) has the solution


2
heat.ode.soln} U (ω, t ) = c(ω)e –σω t . (2.6)

To determine c(ω), we note that

c(ω) = U (ω, 0) = F {u(x , 0)} = F {f (x )} .

Thus, c(ω) is the Fourier transform of the initial condition. This is the same result obtained by separation
of variables. We would now like to take the inverse Fourier transform of (2.6). We need a result

Proposition 2.3. Suppose that f (x ) and g(x ) are functions and

F (ω) = F {f (x )} , G(ω) = F {g(x )} .

Define
Z ∞
1
H (ω) := F (ω)G(ω), and h(x ) := g(y)f (x – y) dy.
2π –∞
Then we have
Z ∞
:convolution} F–1 {H (ω)} = F–1 {F (ω)G(ω)} = g(y)f (x – y) dy = h(x ). (2.7)
–∞

The function h(x ) is called the convolution of g and f and is sometimes denoted g ∗ f .

Proof. Note that


Z ∞
F–1 {H (ω)} = H (ω)e –ωx dω
Z–∞

= F (ω)G(ω)e –ωx dω
Z–∞
∞  Z ∞ 
1
= F (ω) g(y)e i ωy dy e –ωx dω.
–∞ 2π –∞
Rearranging the orders of integration, we have
Z ∞  Z ∞  Z ∞ Z ∞ 
1 1
F (ω) g(y)e i ωy dy e –ωx dω = g(y) F (ω)e –i ω(x –y) dω dy
–∞ 2π –∞ 2π
Z–∞

–∞
1
{eq:hint} = g(y)f (x – y) dy. (2.8)
2π –∞

This proves (2.7).

6
Remark 2.4. Applying a change of variable to the integral (2.8) shows that
Z ∞ Z ∞
1 1
g(y)f (x – y) dy = g(x – z )f (z ) dz .
2π –∞ 2π –∞

Therefore, g ∗ f = f ∗ g, and we may choose either function to have the shifted argument.

We can apply (2.7) to solve the heat equation. Recall that u(x , 0) = f (x ), and define
2
F (ω) := F {f (x )} = c(ω), G(ω, t ) := e –σω t .

By (2.7),

u(x , t ) = F–1 U (ω, t )




= F–1 {F (ω)G(ω, t )}
n 2
o
= F–1 F (ω)e –σω t
Z ∞ r
1 π –(x –y)2 /(4σt )
= f (y) e dy
2π –∞ σt
Z ∞
1 2
=√ f (y)e –(x –y) /(4σt ) dy.
4πσt –∞
We employed the following steps to solve the heat equation:

1. Fourier transform the heat equation

2. Solve the resulting ODE

3. Apply the initial conditions

4. Use the convolution theorem

2.1 Wave equation on infinite domain


Consider the wave equation on the infinite domain

 x ∈ (–∞, ∞), t ∈ (0, ∞),


utt = c 2 uxx , u(x , 0) = f (x ),


 u (x , 0) = 0.
t

We consider the case that ut (x , 0) = g(x ) = 0 for simplicity. Taking the Fourier transform of the wave
equation gives

∂2
{eq:wave.ode} F {utt } = c 2 F {uxx } ⇒ U (ω, t ) = –c 2 ω 2 U (ω, t ), (2.9)
∂t
With the initial conditions

U (ω, 0) = F {f (x )} , U (ω, 0) = 0.
∂t

7
Solving the ODE (2.9) yields

U (ω, t ) = A(ω) cos (cωt ) + B (ω) sin (cωt ) .

Applying the initial condition

∂U (ω, 0)
= cωB (ω) = 0 ⇒ B (ω) = 0,
∂t
and

U (x , 0) = A(ω) = F {f (x )} .

Using the inverse Fourier transform, the solution of the one-dimensional wave equation is
Z ∞
u(x , t ) = U (ω, 0) cos (cωt ) e –i ωx dω.
–∞

Using Euler’s formula

e iy + e –iy
cos(y) = ,
2
we get
Z ∞
u(x , t ) = U (ω, 0) cos (cωt ) e –i ωx dω
–∞
1 ∞
Z  
= U (ω, 0) e –i ω(x –ct ) + e –i ω(x +ct ) dω.
2 –∞

Since U (ω, 0) is the Fourier transform of f (x ), we have


Z ∞
f (x ) = U (ω, 0)e –i ωx dω.
–∞

Thus,

1 ∞
Z  
u(x , t ) = U (ω, 0) e –i ω(x –ct ) + e –i ω(x +ct ) dω
2 –∞
1
= (f (x – ct ) + f (x + ct )) ,
2
which is d’Alembert’s solution.

8
Page 369 (382) section 9.7 of R. Haberman.
Can be used much like Laplace transforms to solve PDEs.

You might also like