CHAPTER 6

The Heat Equation

We introduce several PDE techniques in the context of the heat equation:
The Fundamental Solution is the heart of the theory of innite domain prob-
lems. The fundamental solution also has to do with bounded domains, when
we introduce Greens functions later.
The Maximum Principle applies to the heat equation in domains bounded
in space and time. It is an important property of parabolic equations used
to deduce a variety of results such as uniqueness of solutions, comparison
principles. (E.g., if boundary conditions are changed in a way that suggests
intuitively the resulting temperature should be smaller, this can be proved
using the maximum principle.)
The Energy Method works analogously to the wave equation, except that
the physical (heat) energy is less interesting than a mathematical energy,
which typically decays. As for the wave equation, this leads to straightfor-
ward uniqueness results. It is also useful for obtaining estimates on solutions
that are part of the existence and regularity theory for parabolic equations.
Initial Boundary Value Problems. We will spend some time describing ex-
plicit solutions, expressed as innite series of functions, of the heat equation
plus initial and boundary conditions. There is a general technique frequently
referred to as separation of variables, or as eigenfunction expansions. The
development of this technique leads us to an analysis of eigenvalue problems
for ordinary and partial dierential equations, and to the analysis of Fourier
series.
6.1. The Fundamental Solution
To start with, we consider the heat equation in one space variable, plus time.
In this section, we derive the fundamental solution and show how it is used
to solve the Cauchy problem:
u
t
= ku
xx
, |x| < , t > 0
u(x, 0) = g(x), |x| <
59
60 6. THE HEAT EQUATION
The heat equation has a scale invariance property that is analogous to scale
invariance of the wave equation or scalar conservation laws, but the scaling
is dierent.
Let a > 0 be a constant. Under the scaling x ax, t a
2
t the
heat equation is unchanged. More precisely, if we introduce the change of
variables: t = a
2
t, x = ax, then the heat equation becomes
u
t
= ku
xx
This scale invariance suggests that we seek solutions v depending on the
similarity variable
x
2
t
, or on
x

t
. However, there is a property of the heat
equation we would like to preserve in our similarity solution, that of con-
servation of energy. Suppose u is a solution of the heat equation with the
property that |
_

u(x, 0) dx| < , and u

x
(x.t) 0 as x . Then,
integrating the PDE, we nd
d
dt
_

u(x, t) dx = 0,
so that the total heat energy is conserved:
(1.1)
_

u(x, t) dx = constant.
However,
_

v
_
x

t
_
dx = t
1
2
_

v(y) dy.
This suggests we should scale the function v by t

1
2
:
(1.2) u(x, t) =
1

t
v
_
x

t
_
With this scaling, heat is conserved in the sense of (1.1).
Substituting (1.2) into the PDE leads to an ODE for v = v(y), with non-
constant coecients:
(1.3) kv

(y) +
1
2
yv

(y) +
1
2
v(y) = 0
Since this is a second order equation, we should have two independent solu-
tions. First rewrite the ODE as
kv

(y) +
1
2
(yv(y))

= 0.
Thus,
kv

(y) +
1
2
yv(y) = constant.
6.1. THE FUNDAMENTAL SOLUTION 61
Since we are really only seeking one solution, it is convenient to set the
constant to zero, and write the solution of the homogenous equation:
v(y) = Ae

y
2
4k
.
Converting back to (x, t), with y =
x

t
, we obtain the similarity solution
(1.4) u(x, t) = A
1

t
e

x
2
4kt
Usually, we choose a particular value of A so that the constant in (1.1) is
unity; for this choice of constant, we have the fundamental solution of the
heat equation:
(x, t) =
1

4kt
e

x
2
4kt
.
In higher dimensions, x R
n
, the fundamental solution takes a similar form:
(x, t) =
1
(4kt)
n/2
e
|x|
2
/4kt
= (r, t), r = |x|,
where (r, t) =
1
(4kt)
n/2
e
r
2
/4kt
. Then u(x, t) = (x, t) satises
u
t
= ku, = div grad (the Laplacian with respect to x).
Exercise. Show that (r, t) =
rr
+
n1
r

r
. Hence,

t
= k
_

rr
+
n 1
r

r
_
.
Properties of the Fundamental Solution (x, t)
1. (x, t) > 0 for all x R, t > 0
2. is C

in (x, t), t > 0

3.
_
R
n
(x, t)dx = 1 for all t > 0.
Properties 1 and 2 are obvious; here is the proof of property 3 for n = 1.
_

(x, t)dx =
1

4kt
_

x
2
4kt
dx
62 6. THE HEAT EQUATION
=
1

e
y
2
dy (let y =
x

4kt
, dy =
dx

4kt
)
= 1.
x
y
1
4 kt
0
y = (x,t)
Figure 6.1. Graph of the fundamental solution for the heat
equation, with t > 0. .
Therefore, (x, t) is a probability distribution for each t > 0, with interesting
dependence on t in the limits t and t 0 :
The area under the graph is 1 for all t > 0, yet as t , max
x
(x, t) 0;
the tail spreads out to maintain
_
= 1
As t 0 the maximum (at x = 0) blows up like
1

t
, but the integral remains
constant. We also observe (x, t) 0 for x = 0, as t 0 + .
6.2. The Cauchy Problem for the Heat Equation
As for the wave equation, the Cauchy problem is the pure initial value
problem, here stated in the one-dimensional case, n = 1:
(2.5)
u
t
= ku
xx
, |x| < , t > 0 (a)
u(x, 0) = g(x), |x| < (b)
6.2. THE CAUCHY PROBLEM FOR THE HEAT EQUATION 63
Recall that the fundamental solution
(x, t) =
1

4kt
e

x
2
4kt
satises (2.5(a)) for t > 0.
Now (x y, t) is a solution of (2.5(a)) for all y, by translation invariance:
x x y does not change the heat equation. Thus,
(x y, t)g(y)
is also a solution of (2.5(a)). For later reference, we note that the heat
equation is invariant under time translation also.
By linearity and homogeneity of the PDE, we can also take linear combina-
tions of solutions. This suggests that
(2.6) u(x, t) =
_

(x y, t)g(y)dy
should also be a solution. Moreover, properties of suggest that as t 0+,
u(x, t) g(x), since (x y, t) collapses to zero away from y = x, and
blows up at y = x in such a way (i.e., preserving
_
= 1) that the initial
condition is satised in the sense u(x, t) g(x) as t 0 + .
It is straightforward to check that the integrals for u, u
t
, u
xx
all converge
provided g C(R) is bounded. Then
u
t
=
_

t
(x y, t)g(y)dy; u
xx
=
_

x
2
(x y, t)g(y)dy,
so that u satises the PDE for t > 0.
It is more complicated to check that the initial condition is satised. We
need to show u(x, 0) = g(x). But t = 0 is a singular point for : (x, t) is
not dened at t = 0. To get an idea of why lim
t0+
u(x, t) = g(x), lets x
x.
Then, for > 0,
_

(x y, t)g(y)dy
=
_
(xy)<
(x y, t)g(y)dy +
_
|xy|>
(x y, t)g(y)dy

_
|xy|<
(x y, t)g(x)dy
By continuity, g(y) g(x) for y near x, which explains how the rst integral
is approximately the nal line. The second integral approaches zero as
t 0+, because 0 uniformly, and exponentially, away from y = x as
t 0+.
64 6. THE HEAT EQUATION
Theorem: Let g C(R) be bounded, and let u(x, t) be given be the formula
(2.6). Then
1. u is C

in (x, t) for t > 0;

2. u satises the heat equation
u
t
= ku
xx
, x R, t > 0;
3. lim
(x, t) (x
0
, 0)
t > 0
u(x, t) = g(x
0
) for all x
0
R.
Proof:
Property 1 follows because is C

for t > 0, and derivatives of all decay

exponentially as |x| , so the integrals converge.
Property 2 follows from
t
= k
xx
, t > 0.
To prove property 3, we look at the dierence |u(x, t) g(x
0
)|, estimate
and show that the pieces we get behave as we expect, i.e., as in the rough
argument preceeding the proof.
Let > 0. (This measures |u(x, t)g(x
0
)|.) Then, since
_
dx = 1, we have
(2.7) |u(x, t) g(x
0
)| =

(x y, t)(g(y) g(x
0
)) dy

,
Let > 0 (we choose below), and break up the integrals in (2.7):
(2.8)
|u(x, t)g(x
0
)|

_
|x
0
y|<
(xy, t)(g(y)g(x
0
))dy

_
|x
0
y|
(xy, t)(g(y)g(y
0
))dy

Now we use two ways to show the two integrals are small:
1st: g(y) g(x
0
). For y near x, (x y, t) blows up as t 0, but
_
(x y, t) dy is bounded uniformly by 1.
2nd:
_
|x
0
y|
(x y, t) dy 0 as t 0, provided |x x
0
| <

2
, while
g(y) is bounded.
We write the right hand side of (2.8) as
I

+ J

Choose > 0 so that |g(y) g(x

0
)| < for |y x
0
| < (by continuity of g
at x
0
).
6.2. THE CAUCHY PROBLEM FOR THE HEAT EQUATION 65
Then
I

_
|x
0
y|<
(x y, t)|g(y) g(x
0
)|dy

_
|x
0
y|<
(x y, t)dy .
The second integral is somewhat trickier: Since g is bounded, there is K > 0
such that |g(y)| K, for all y. Thus,
J

_
|x
0
y|>
(x y, t)|g(y) g(x
0
)|dy
2K
_
|x
0
y|>
1

4kt
e

(xy)
2
4kt
dy
Here is the second use of : Consider x satisfying |x x
0
| <

2
. Then
|x y| >

2
in the range of integration |x
0
y| > . But this is not a good
enough estimate of the exponential, because we would still be left with an
integral over an innite interval of a small but positive quantity. So, we
need to observe that in the region of integration, |x y|
1
2
|y x
0
|. Then
J

k
_
|x
0
y|>
1

t
e

(x
0
y)
2
16kt
dy
C
_
|z|

4

kt
e
z
2
dz < ,
for t > 0 suciently small.
Thus, |u(x, t) g(x
0
)| < 2 for |x x
0
| <

2
, t > 0 suciently small.
This proves property 3.
Remarks
1. lim
t0+
(x, t) is not a function in the usual sense, but is a distribution
or generalized function called the Dirac delta function (x) :
_
R
(x y)g(y) dy = g(x),
where the integral is a notational convenience. One way to interpret the delta
function and the integral is to recognize the delta function as a measure that
places unit mass at x = 0 and zero mass at each x = 0.
2.
_

(x y, t) g(y) dy is the convolution of the function (, t) with g

More generally, for integrable functions , on R the convolution product
of and is also a function dened by
(x) =
_

(x y)(y) dy.
66 6. THE HEAT EQUATION
6.2.1. Using the Fundamental Solution to Solve Quarter-plane
Problems. :
Consider the so-called quarter-plane problem:
1) u
t
= ku
xx
, x > 0, t > 0
2) u(0, t) = 0, t > 0. (Homogeneous Dirichlet boundary condition.)
3) u(x, 0) = g(x), x > 0
As for the wave equation, we reect the initial data so the solution satises
the boundary conditions:
Let g(x) be the odd extension of g(x).
g(x) =
_
g(x), x > 0
g(x), x < 0
and dene
(2.9) u(x, t) =
_

(x y, t) g(y)dy
Then
u
t
= ku
xx
,
u(x, 0) = g(x) = g(x), x > 0
u(0, t) =
_

(y, t) g(y)dy = 0, t > 0,

since (y, t) is an even function of y, and g is an odd function.
It is straightforward to check directly that u(x, t) is an odd function of
x R. That is, the symmetry in the initial data is carried through to the
same symmetry in the solution.
Now replace g(y) with g(y) using g(y) = g(y), y < 0. Then
u(x, t) =
_

0
(x y, t)g(y)dy +
_
0

(x y, t)(g(y))dy
=
_

0
(x y, t)g(y)dy
_

0
(x + y, t)g(y)dy
=
_

0
((x y, t) (x + y, t))g(y)dy
6.3. THE ENERGY METHOD. 67
With a homogeneous Neumann boundary condition instead, we extend the
initial data to be even:
u
t
= ku
xx
, x > 0, t > 0
u
x
(0, t) = 0 (Homogeneous Neumann Condition)
u(x, 0) = g(x), x > 0
Solution: Extend g using the even extension, so that the rst derivative is
zero at x = 0.
Then u(x, t) =
_

0
((x y, t) + (x + y, t)) g(y) dy.
To check the boundary condition:
u
x
(0, t) =
_

0
(
x
(y, t) +
x
(y, t) g(y) dy.
(y, t) is even in y so
x
(y, t) is odd. Thus, u
x
(0, t) = 0.
6.3. The Energy Method.
Consider
u
t
= k u
xx
, a x b, t > 0.
multiplying by u and integrating over [a, b],
_
b
a
uu
t
dx = k
_
b
a
uu
xx
dx.
Therefore,
d
dt
_
b
a
1
2
u
2
dx = kuu
x
|
b
a
k
_
b
a
u
2
x
dx.
Thus, the energy integral (not the heat energy) E(t) =
_
b
a
1
2
u
2
(x, t)dx is
decreasing in time if uu
x
|
b
a
0.
For example, if either u or u
x
is zero at each end point x = a, x = b, then
the energy integral decreases in t.
In this case, we have the important comparison to the initial data:
E(t) E(0), t > 0.
I.e.,
(3.10)
_
b
a
u
2
(x, t)dx
_
b
a
u
2
(x, 0)dx, t > 0.
68 6. THE HEAT EQUATION
6.3.1. Using the Energy Method to prove Uniqueness. Consider
the initial boundary value problem
(IBVP)
_

_
u
t
= ku
xx
+ f(x, t), 0 < x < l, t > 0
u(0, t) = g(t), t > 0
u(L, t) = h(t), t > 0
u(x, 0) = (x), 0 < x < L
where f, g, h, are given functions.
The energy inequality works in any number of dimensions (will involve di-
vergence theorem and Greens identities instead of integration by parts).
Theorem. If u
1
, u
2
solve (IBVP) and are C
2
functions, then u
1
= u
2
everywhere.
Proof: Let u = u
1
u
2
Then u satises (IBVP) with zero data: f g h 0. But we know
that E(t) is decreasing, from the energy inequality
0
_
L
0
(u(x, t))
2
dx
_
L
0
u(x, 0)
2
dx = 0
Therefore, u 0, so that u
1
u
2
.
Remark. In general, the existence of solutions is harder to establish than
uniqueness.
6.3.2. The Energy Principle in Higher Dimensions: Consider a
bounded open subset U of R
n
, and the initial boundary value problem
u
t
= ku, x U t > 0
u(x, t) = 0, x U, t > 0
u(x, 0) = (x), x U
We begin by dening the energy integral as in one dimension:
(3.11) E(t) =
1
2
_
U
u(x, t)
2
dx.
Then
6.4. THE MAXIMUM PRINCIPLE 69
E

(t) =
_
U
uu
t
dx = k
_
U
uu dx
= k
_
u
uu n dS k
_
U
u u dx
= k
_
U
|u|
2
dx 0.
Exercise: Prove the energy inequality for u
t
= .(k(x, u)u) where
k(x, u) R
n
is a given positive function. The only dierence is that k gets
moved under the integral sign. Does this enable us to prove uniqueness of
solutions for the quasilinear heat equation?
6.4. The Maximum Principle
For T > 0, and U R
n
, we use the notation U
T
= U (0, T]. Note that
U
T
includes the top U {t = T}. The following theorem, the Maximum
Principle, states that the maximum of any (smooth) solution of the heat
equation occurs either initially (at t = 0), or on the boundary of the domain.
These parts of the boundary of U
T
are known as the parabolic boundary
T
:

T
= U
T
U
T
= (U [0, T]) (U {t = 0}).
Solutions of the heat equation should have two spatial derivatives and one
time derivative, so we dene the appropriate space of functions on U
T
:
C
2
1
(U
T
) = {u = u(x, t) : u, u
t
, D
x
u, D
2
x
u C(U
T
)}.
In order to compare values of u in U
T
with values on the boundary, we
require in the theorem that u should be continuous on U
T
.
Theorem. The Maximum Principle. Let u C(U
T
) C
2
1
(U
T
) satisfy
u
t
= ku, (x, t) U
T
.
Then
max
U
T
u(x, t) = max

T
u(x, t).
Remarks:
1. Suppose u has a local maximum at (x, t) U
T
. Then u
t
= 0 = u
x
= 0,
u 0. If u < 0 at (x, t), then the pde gives us a contradiction:
0 = u
t
= ku < 0
Although this is not a proof, since we have to handle the degenerate case in
which u = 0, it has the main idea; the proof merely modies u to remove
a possibly degenerate maximum.
70 6. THE HEAT EQUATION
2. U
T
is closed, so u achieves its maximum somewhere in U
T
: there is an
(x
0
, t
0
) such that u(x
0
, t
0
) = max
U
T
u(x, t).
Proof: Let M = max

T
u(x, t)
U
T

-
L
x
T
6
t
0
Goal: Prove that U(x, t) M for all (x, t) U
T
To deal with the possibility u = 0 at a maximum, we perturb u a bit:
Let v(x, t) = u(x, t) + |x|
2
, > 0. Then
v
t
kv = u
t
ku
. .
2 k n < 0
0
(4.12)
Now suppose v has a local maximum in U
T
, at P
0
= (x
0
, t
0
) with t
0
< T.
Then v
t
= v
x
= 0 at P
0
, and v 0 at P
0
. But this contradicts (4.12), so
v cannot have a max in the interior of U
T
.
Now suppose v has a maximum on the line t = T, at P
1
= (x
1
, t = T). Then
v
x
= 0, v
xx
0 and v
t
0 at P
1
.
1
Now we have v
t
kv 0, again
contradicting (4.12).
1
The nal inequality is easily proved by contradiction - vt < 0 at P1, then v(x1, t) >
v(x1, T) for t < T close to t T, contradicting the assumption that v has a maximum at
P1.
6.5. DUHAMELS PRINCIPLE FOR THE INHOMOGENEOUS HEAT EQUATION 71
Therefore maximum of v on U
T
occurs on
T
. I.e., v(x, t) max
(y,t)U
T
v(y, t)
for all (x, t) U
T
. We have proved
u(x, t) + |x|
2
max
(y,t)
T
(u(y, t) + |y|
2
) max

T
u + C for all (x, t) U
T
,
where C = max
yU
|y|
2
. Thus,
u(x, t) M + (C |x|
2
)
M + C
Since > 0 is arbitrary, we have u(x, t) M for all (x, t) U
T
.
Remarks:
1. The weak maximum principle is easy to prove. The related strong maxi-
mum principle is somewhat harder to prove. The strong maximum principle
states that, provided U is connected, then the maximum of u is achieved
only on the parabolic boundary, unless u is constant throughout U
T
.
2. By applying the maximum principle to u, which also satises the con-
ditions of the Theorem, we see that there is a corresponding minimum prin-
ciple:
min
U
T
u(x, t) = min

T
u(x, t).
6.5. Duhamels Principle for the Inhomogeneous Heat Equation
Consider the example of the heat equation on the whole real line, with a
heat source or sink represented by the density function f(x, t) :
(P)
_
u
t
= ku
xx
+ f(x, t), |x| < , t > 0
u(x, 0) = 0 |x| <
The fundamental solution (x, t) of the homogeneous heat equation satises
u
t
= ku
xx
, for t > 0. But then, for any y, s > 0 the shifted function (x
y, ts) satises the equation for t > s, and we can multiply by an amplitude
f(y, s), so that we have a collection of solutions (x y, t s)f(y, s) of the
heat equation, with y R and t > s. Summing (i.e., integrating) these
solutions over y R, with s xed, we have that
u(x, t; s) =
_

(x y, t s)f(y, s) dy is a solution for all t > s,

satisfying u(x, t = s; s) = f(x, s)
72 6. THE HEAT EQUATION
Now the idea is to integrate u(x, t; s) with respect to s from 0 to t :
Dene u(x, t) =
_
t
0
u(x, t, s)ds. Then
u
t
ku
xx
= u(x, t; t) +
_
t
0
( u
t
(x, t; s) k u
xx
(x, t; s))ds
= f(x, t)
Note that this calculation is slightly misleading, because we have dieren-
tiated under the integral sign, when (x y, t s) has a singularity at
x = y, t = s, on the boundary of the domain of integration. However, this
singularity can be handled with the appropriate limit, and the result is the
same. (See Evans [PDEs], page 50 for the details.)
Finally, we observe
u(x, 0) = 0 < x < .