The 1-D Heat Equation

18.303 Linear Partial Dierential Equations

Matthew J. Hancock

Fall 2006

1 The 1-D Heat Equation

1.1 Physical derivation
Reference: Guenther & Lee 1.3-1.4, Myint-U & Debnath 2.1 and 2.5
[Sept. 8, 2006]
In a metal rod with non-uniform temperature, heat (thermal energy) is transferred
from regions of higher temperature to regions of lower temperature. Three physical
principles are used here.
1. Heat (or thermal) energy of a body with uniform properties:

Heat energy = cmu,

where m is the body mass, u is the temperature, c is the specic heat, units [c] =
L2 T 2 U 1 (basic units are M mass, L length, T time, U temperature). c is the energy
required to raise a unit mass of the substance 1 unit in temperature.
2. Fouriers law of heat transfer: rate of heat transfer proportional to negative
temperature gradient,
Rate of heat transfer u
= K0 (1)
area x
where K0 is the thermal conductivity, units [K0 ] = M LT 3 U 1 . In other words, heat
is transferred from areas of high temp to low temp.
3. Conservation of energy.
Consider a uniform rod of length l with non-uniform temperature lying on the
x-axis from x = 0 to x = l. By uniform rod, we mean the density , specic heat c,
thermal conductivity K0 , cross-sectional area A are ALL constant. Assume the sides

of the rod are insulated and only the ends may be exposed. Also assume there is no
heat source within the rod. Consider an arbitrary thin slice of the rod of width x
between x and x + x. The slice is so thin that the temperature throughout the slice
is u (x, t). Thus,

Heat energy of segment = c Ax u = cAxu (x, t) .

By conservation of energy,

change of
heat in from heat out from
heat energy of = .
left boundary right boundary
segment in time t

From Fouriers Law (1),

u u
cAxu (x, t + t) cAxu (x, t) = tA K0 tA K0
x x x x+x

Rearranging yields (recall , c, A, K0 are constant),

u
u (x, t + t) u (x, t) K0 x x+x
u
x x
=
t c x

Taking the limit t, x 0 gives the Heat Equation,

u 2u
= 2 (2)
t x
where
K0
= (3)
c
is called the thermal diusivity, units [] = L2 /T . Since the slice was chosen arbi
trarily, the Heat Equation (2) applies throughout the rod.

1.2 Initial condition and boundary conditions

To make use of the Heat Equation, we need more information:
1. Initial Condition (IC): in this case, the initial temperature distribution in the
rod u (x, 0).
2. Boundary Conditions (BC): in this case, the temperature of the rod is aected
by what happens at the ends, x = 0, l. What happens to the temperature at the
end of the rod must be specied. In reality, the BCs can be complicated. Here we
consider three simple cases for the boundary at x = 0.

2
 (I) Temperature prescribed at a boundary. For t > 0,

u (0, t) = u1 (t) .

(II) Insulated boundary. The heat ow can be prescribed at the boundaries,

u
K0 (0, t) = 1 (t)
x
(III) Mixed condition: an equation involving u (0, t), u/x (0, t), etc.
Example 1. Consider a rod of length l with insulated sides is given an initial
temperature distribution of f (x) degree C, for 0 < x < l. Find u (x, t) at subsequent
times t > 0 if end of rod are kept at 0o C.
The Heat Eqn and corresponding IC and BCs are thus

PDE: ut = uxx , 0 < x < l, (4)

IC: u (x, 0) = f (x) , 0 < x < l, (5)
BC: u (0, t) = u (L, t) = 0, t > 0. (6)

Physical intuition: we expect u 0 as t .

1.3 Non-dimensionalization
Dimensional (or physical) terms in the PDE (2): k, l, x, t, u. Others could be
introduced in IC and BCs. To make the solution more meaningful and simpler, we
group as many physical constants together as possible. Let the characteristic length,
time and temperature be L , T and U , respectively, with dimensions [L ] = L,
[T ] = T , [U ] = U . Introduce dimensionless variables via

x t u (x, t) f (x)
x = , t = , t =
u x, , f (
x) = . (7)
L T U U

The variables x, t, u are dimensionless (i.e. no units, [x] = 1). The sensible choice
for the characteristic length is L = l, the length of the rod. While x is in the range
0 < x < l, x is in the range 0 < x < 1.
The choice of dimensionless variables is an ART. Sometimes the statement of the
problem gives hints: e.g. the length l of the rod (1 is nicer to deal with than l, an
unspecied quantity). Often you have to solve the problem rst, look at the solution,
and try to simplify the notation.

 From the chain rule,

u u t U u

ut = = U = ,
t t t T t
u u x U u
ux = = U =
x x x L x
U 2 u
uxx =
L2 x2

Substituting these into the Heat Eqn (4) gives

u T 2 u
ut = uxx = 2
t L x2

To make the PDE simpler, we choose T = L2 / = l2 /, so that

u 2 u
= , 0 < x < 1, t > 0.
t x2
The characteristic (diusive) time scale in the problem is T = l2 /. For dierent
substances, this gives time scale over which diusion takes place in the problem. The
IC (5) and BC (6) must also be non-dimensionalized:

IC: u (x, 0) = f (x) , 0 < x < 1,

t > 0.

BC: u 0, t = u 1, t = 0,

1.4 Dimensionless problem

Dropping hats, we have the dimensionless problem

PDE: ut = uxx , 0 < x < 1, (8)

IC: u (x, 0) = f (x) , 0 < x < 1, (9)
BC: u (0, t) = u (1, t) = 0, t > 0, (10)

where x, t are dimensionless scalings of physical position and time.

2 Separation of variables
Ref: Guenther & Lee, 4.2 and 5.1, Myint-U & Debnath 6.4
[Sept 12, 2006]
We look for a solution to the dimensionless Heat Equation (8) (10) of the form

u (x, t) = X (x) T (t) (11)

4
Take the relevant partial derivatives:

uxx = X (x) T (t) , ut = X (x) T (t)

where primes denote dierentiation of a single-variable function. The PDE (8), ut =

uxx , becomes
T (t) X (x)
=
T (t) X (x)
The left hand side (l.h.s.) depends only on t and the right hand side (r.h.s.) only
depends on x. Hence if t varies and x is held xed, the r.h.s. is constant, and hence
T /T must also be constant, which we set to by convention:

T (t) X (x)
= = , = constant. (12)
T (t) X (x)

The BCs become, for t > 0,

u (0, t) = X (0) T (t) = 0

u (1, t) = X (1) T (t) = 0

Taking T (t) = 0 would give u = 0 for all time and space (called the trivial solution),
from (11), which does not satisfy the IC unless f (x) = 0. If you are lucky and
f (x) = 0, then u = 0 is the solution (this has to do with uniqueness of the solution,
which well come back to). If f (x) is not zero for all 0 < x < 1, then T (t) cannot be
zero and hence the above equations are only satised if

X (0) = X (1) = 0. (13)

2.1 Solving for X (x)

Ref: Guenther & Lee, 4.2 and 5.1 and 7.1, Myint-U & Debnath 7.1 7.3
We obtain a boundary value problem for X (x), from (12) and (13),

X (x) + X (x) = 0, 0 < x < 1, (14)

X (0) = X (1) = 0. (15)

This is an example of a Sturm-Liouville problem (from your ODEs class).
There are 3 cases: > 0, < 0 and = 0.
(i) < 0. Let = k 2 < 0. Then the solution to (14) is

X = Aekx + Bekx

5
for integration constants A, B found from imposing the BCs (15),

X (0) = A + B = 0, X (1) = Aek + Bek = 0.

The rst gives A = B, the second then gives A e2k 1 = 0, and since |k| > 0 we

have A = B = u = 0, which is the trivial solution. Thus we discard the case < 0.
(ii) = 0. Then X (x) = Ax + B and the BCs imply 0 = X (0) = B, 0 = X (1) =
A, so that A = B = u = 0. We discard this case also.
(iii) > 0. In this case, (14) is the simple harmonic equation whose solution is

X (x) = A cos x + B sin x . (16)

The BCs imply 0 = X (0) = A, and B sin = 0. We dont want B = 0, since that
would give the trivial solution u = 0, so we must have

sin = 0. (17)

Thus = n, for any nonzero integer n (n = 1, 2, 3, ...). We use subscripts
to label the particular n-value. The values of are called the eigenvalues of the
Sturm-Liouville problem (14),

n = n2 2 , n = 1, 2, 3, ...

and the corresponding solutions of (14) are called the eigenfunctions of the Sturm-Liouville
problem (14),
Xn (x) = bn sin (nx) , n = 1, 2, 3, ... (18)
We have assumed that n > 0, since n < 0 gives the same solution as n > 0.

2.2 Solving for T (t)

When solving for X (x), we found that non-trivial solutions arose for = n2 2 for all
nonzero integers n. The equation for T (t) is thus, from (12),

T (t) = n2 2 T (t)

and, for n, the solution is

2 2 t
Tn = cn en , n = 1, 2, 3, ... (19)

where the cn s are constants of integration.

6
2.3 Full solution u (x, t)
Ref: Myint-U & Debnath 6.4, Ch 5
Putting things together, we have, from (11), (18) and (19),
2 2 t
un (x, t) = Bn sin (nx) en , n = 1, 2, 3, ... (20)

where Bn = cn bn . Each function un (x, t) is a solution to the PDE (8) and the BCs
(10). But, in general, they will not individually satisfy the IC (9),

un (x, 0) = Bn sin (nx) = f (x) .

We now apply the principle of superposition: if u1 and u2 are two solutions to the
PDE (8) and BC (10), then c1 u1 + c2 u2 is also a solution, for any constants c1 , c2 .
This relies on the linearity of the PDE and BCs. We will, of course, soon make this
more precise....
Since each un (x, 0) is a solution of the PDE, then the principle of superposition
says any nite sum is also a solution. To solve the IC, we will probably need all the
solutions un , and form the innite sum (convergence properties to be checked),


u (x, t) = un (x, t) . (21)
n=1

u (x, t) satises the BCs (10) since each un (x, t) does. Assuming term-by-term dif
ferentiation holds (to be checked) for the innite sum, then u (x, t) also satises the
PDE (8). To satisfy the IC, we need to nd Bn s such that



f (x) = u (x, 0) = un (x, 0) = Bn sin (nx) . (22)
n=1 n=1

This is the Fourier Sine Series of f (x).

To solve for the Bn s, we use the orthogonality property for the eigenfunctions
sin (nx),
1
0 m = n 1
sin (mx) sin (nx) dx = = mn (23)
0 1/2 m = n 2
where mn is the kronecker delta,

0 m = n
mn =
1 m=n

The orthogonality relation (23) is derived by substituting

2 sin (mx) sin (nx) = cos ((m n) x) cos ((m + n) x)

7
into the integral on the left hand side of (23) and noting
1
cos (mx) dx = m0 .
0

The orthogonality of the functions sin (nx) is analogous to that of the unit vectors
x and y in 2-space; the integral from 0 to 1 in (23) above is analogous to the dot
product in 2-space.
To solve for the Bn s, we multiply both sides of (22) by sin (mx) and integrate
from 0 to 1:
1
1
sin (mx) f (x) dx = Bn sin (nx) sin (mx) dx
0 n=1 0

Substituting (23) into the right hand side yields

1
1
sin (mx) f (x) dx = Bn nm
0 n=1
2
By denition of nm , the only term that is non-zero in the innite sum is the one
where n = m, thus 1
1
sin (mx) f (x) dx = Bm
0 2
Rearranging yields 1
Bm = 2 sin (mx) f (x) dx. (24)
0
The full solution is, from (20) and (21),

2 2 t
u (x, t) = Bn sin (nx) en , (25)
n=1

where Bn are given by (24).

To derive the solution (25) of the Heat Equation (8) and corresponding BCs
(10) and IC (9), we used properties of linear operators and innite series that need
justication.

3 Example : Cooling of a rod from a constant ini

tial temperature
Suppose the initial temperature distribution f (x) in the rod is constant, i.e. f (x) =
u0 . The solution for the temperature in the rod is (25),

2 2 t
u (x, t) = Bn sin (nx) en ,
n=1

where, from (24), the Fourier coecients are given by

1 1
Bn = 2 sin (nx) f (x) dx = 2u0 sin (nx) dx.
0 0

Calculating the integrals gives

1
cos (n) 1 2u0 0 n even

Bn = 2u0 sin (nx) dx = 2u0 = ((1)n 1) = 4u0
0 n n n
n odd

In other words,
4u0
B2n = 0, B2n1 =
(2n 1)
and the solution becomes

4u0 sin ((2n 1) x)
exp (2n 1)2 2 t .

u (x, t) = (26)
n=1 (2n 1)

3.1 Approximate form of solution

The number of terms of the series (26) needed to get a good approximation for u (x, t)
depends on how close t is to 0. The series is

4u0 2 t sin (3x) 92 t
u (x, t) = sin (x) e + e + .
3
The ratio of the rst and second terms is
2
|second term| e8 t |sin 3x|

=
|rst term| 3 |sin x|

2t
e8 using |sin nx| n |sin x|
1
e8 for t 2

< 0.00034

It can also be shown that the rst term dominates the sum of the rest of the terms,
and hence
4u0 2 1
u (x, t) sin (x) e t , for t 2 . (27)

What does t = 1/ 2 correspond to in physical time? In physical time t = l2 t/
(recall our scaling - here we use t as the dimensionless time and t as dimensional
time), t = 1/ 2 corresponds to:

t 15 minutes, for a 1 m rod of copper ( 1.1 cm2 sec1 )

t 169 minutes, for a 1 m rod of steel ( 0.1 cm2 sec1 )
t 47 hours, for a 1 m rod of glass ( 0.006 cm2 sec1 )

9
At t = 1/ 2 , the temperature at the center of the rod (x = 1/2) is, from (27),
4
u (x, t) u0 = 0.47u0 .
e
Thus, after a (scaled) time t = 1/ 2 , the temperature has decreased by a factor of
0.47 from the initial temperature u0 .

3.2 Geometrical visualization of the solution

[Sept 14, 2006]
To analyze qualitative features of the solution, we draw various types of curves in
2D:
1. Spatial temperature prole, given by

u = u (x, t0 )

where t0 is a xed value of x. These proles are curves in the ux-plane.

2. Temperature proles in time,

u = u (x0 , t)

where x0 is a xed value of x. These proles are curves in the ut-plane.

3. Curves of constant temperature in the xt-plane (level curves),

u (x, t) = C

where C is a constant.
Note that the solution u = u (x, t) is a 2D surface in the 3D uxt-space. The above
families of curves are the dierent cross-sections of this solution surface. Drawing the
2D cross-sections is much simpler than drawing the 3D solution surface.
Sketch typical curves: when sketching the curves in 1-3 above, we draw a few
typical curves and any special cases. While math packages such as Matlab can be
used to compute the curves from, say, 20 terms in the full power series solution (26),
the emphasis in this course is to use simple considerations to get a rough idea of what
the solution looks like. For example, one can use the rst term approximation (27),
simple physical considerations on heat transfer, and the fact that the solution u (x, t)
is continuous in x and t, so that if t1 is close to t1 , u (x, t1 ) is close to u (x, t2 ).

3.2.1 Spatial temperature proles

For xed t = t0 , the rst term approximate solution (27) is

4u0 2 t0
u (x, t) e sin (x) , t 1/ 2 . (28)

10
 Figure 1: Spatial temperature proles u(x, t0 ).

This suggests the center of the rod, x = 1/2, is a line of symmetry for u (x, t), i.e.

1 1
u + s, t = u s, t
2 2
and, for each xed time, the location of the maximum (minimum) temperature if
u0 > 0 (u0 < 0). We can prove the symmetry property by noting that the original
PDE/BC/IC problem is invariant under the transformation x 1 x. Note also
that, from (26),


cos ((2n 1) x) exp (2n 1)2 2 t

ux (x, t) = 4u0
n=1

Thus ux (1/2, t) = 0 and uxx (1/2, t) < 0, so the 2nd derivative test implies that
x = 1/2 is a local max.
In Figure 1, we have plotted two typical proles, one at early times t = t0 0 and
the other at late times t = t0 0, and two special proles, the initial temperature
at t = 0 (u = u0 ) and the temperature as t (u = 0). The prole for t = t0 0
is found from the rst term approximation (28). The line of symmetry x = 1/2 is
plotted as a dashed line for reference in Figure 1.

11
 Figure 2: Temperature proles in time u (x0 , t).

3.2.2 Temperature proles in time

Setting x = x0 in the approximate solution (27),

4u0 2
u (x0 , t) sin (x0 ) e t , t 1/ 2 .

Two typical proles are sketched in Figure 2, one near the center of the rod (x0 1/2)
and one near the edges (x0 0 or 1). To draw these we noted that the center of
the rod cools more slowly than points near the ends. One special prole is plotted,
namely the temperature at the rod ends (x = 0, 1).

3.2.3 Curves of constant temperature (level curves of u(x, t))

In Figure 3, we have drawn three typical level curves and two special ones, u = 0
(rod ends) and u = u0 (the initial condition). For a xed x = x0 , the temperature
in the rod decreases as t increases (motivated by the rst term approximation (28)),
as indicated by the points of intersection on the dashed line. The center of the rod
(x = 1/2) is a line of symmetry, and at any time, the maximum temperature is at
the center. Note that at t = 0, the temperature is discontinuous at x = 0, 1.
To draw the level curves, it is easiest to already have drawn the spatial temperature
proles. Draw a few horizontal broken lines across your u vs. x plot. Suppose you
draw a horizontal line u = u1 . Suppose this line u = u1 crosses one of your proles
u (x, t0 ) at position x = x1 . Then (x1 , t0 ) is a point on the level curve u (x, t) = u1 .
Now plot this point in your level curve plot. By observing where the line u = u1
crosses your various spatial proles, you ll in the level curve u (x, t) = u1 . Repeat

12

Figure 3: Curves of constant temperature u(x, t) = c, i.e., the level curves of u(x, t).

this process for a few values of u1 to obtain a few representative level curves. Plot
also the special cases: u (x, t) = u0 /w, u = 0, etc.
When drawing visualization curves, the following result is also helpful.

3.2.4 Maximum Principle for the basic Heat Problem

Ref: Guenther & Lee 5.2, Myint-U & Debnath 8.2

This result is useful when plotting solutions: the extrema of the solution of the
heat equation occurs on the space-time boundary, i.e. the maximum of the initial
condition and of the time-varying boundary conditions. More precisely, given the
heat equation with some initial condition f (x) and BCs u (0, t), u (1, t), then on a
given time interval [0, T ], the solution u (x, t) is bounded by

umin u (x, t) umax

where

umax = max max f (x) , max u (0, t) , max u (1, t)
0<x<1 0<t<T 0<t<T

umin = min min f (x) , min u (0, t) , min u (1, t)
0<x<1 0<t<T 0<t<T

13

Thus, in the example above, umin = 0 and umax = u0 , hence for all x [0, 1] and
t [0, T ], 0 u (x, t) u0 .

4 Equilibrium temperature prole (steady-state)

Intuition tells us that if the ends of the rod are held at 0o C and there are no heat
sources or sinks in the rod, the temperature in the rod will eventually reach 0. The
solution above conrms this. However, we do not have to solve the full problem to
determine the asymptotic or long-time behavior of the solution.
Instead, the equilibrium or steady-state solution u = uE (x) must be independent
of time, and will thus satisfy the PDE and BCs with ut = 0,

uE = 0, 0 < x < 1; uE (0) = uE (1) = 0.

The solution is uE = c1 x + c2 and imposing the BCs implies uE (x) = 0. In other

words, regardless of the initial temperature distribution u (x, 0) = f (x) in the rod,
the temperature eventually goes to zero.

4.1 Rate of decay of u (x, t)

How fast does u approach uE = 0? From our estimate (82) above,
2 2 t
|un (x, t)| Ben , n = 1, 2, 3, ...
n
2 2 n 2 t 2 t
where B is a constant. Noting that en t e = e , we have

Br 2
un (x, t) |un (x, t)| B rn = , r = e t < 1 (t > 0).

1r

n=1 n=1 n=1

The last step is the geometric series result n r

n=1 r = 1r , for |r| < 1. Thus
2
Be t
u (x, t) , t > 0. (29)

n=1 n 1 e2 t

Therefore

u (x, t) approaches the steady-state uE (x) = 0 exponentially fast (i.e. the rod
cools quickly)
2
the rst term in the series, sin (x) e t (term with smallest eigenvalue = )
determines the rate of decay of u (x, t)

the Bn s may also aect the rate of approach to the steady-state, for other
problems

14
4.2 Error of rst-term approximation
Using the method in the previous section, we can compute the error between the
rst-term and the full solution (26),


|u (x, t) u1 (x, t)| = un (x, t) u1 (x, t)

n=1



Br2
= un (x, t) |un (x, t)| B rn =

n=2 n=2 n=2
1r
2
where r = e t . Hence the solution u (x, t) approaches the rst term u1 (x, t) expo
nentially fast,
2
Be2 t
|u (x, t) u1 (x, t)| , t > 0. (30)
1 e2 t
With a little more work we can get a much tighter (i.e. better) upper bound. We
consider the sum

|un (x, t)|
n=N

Substituting for un (x, t) gives

e(n )2 t
2 N 2

n2 2 t N 2 2 t
|un (x, t)| B e = Be (31)
n=N n=N n=N

Heres the trick: for n N ,

n2 N 2 = (n + N ) (n N ) 2N (n N ) 0.

Since ex is a decreasing function, then

e(n )2 t e2N (nN )2 t

2 N 2
(32)

for t > 0. Using the inequality (32) in Eq. (31) gives

N 2 2 t 2t
|un (x, t)| Be e2N (nN )
n=N n=N
N
2
2 2 t

2t
n 2 2 t
e2N t
= BeN e2N = BeN .
n=N
1 e2N 2 t

Simplifying the expression on the right hand side yields

2 2
BeN t

|un (x, t)| (33)

n=N
1 e2N 2 t

15
For N = 1, (33) gives the temperature decay
2
Be t
|u (x, t)| = un (x, t) |un (x, t)| , (34)

1 e2 2 t
n=N n=N

which is slightly tighter than (29). For N = 2, (33) gives the error between u (x, t)
and the rst term approximation,
2
Be4 t
|u (x, t) u1 (x, t)| = u (x, t) |u (x, t)|

n=2 n n=2 n 1 e42 t

which is a much better upper bound than (30).

Note that for an initial condition u (x, 0) = f (x) that is is symmetric with respect
to x = 1/2 (e.g. f (x) = u0 ), then u2n = 0 for all n, and hence error between u (x, t)
and the rst term u1 (x, t) is even smaller,


|u (x, t) u1 (x, t)| = un (x, t) |un (x, t)|

n=3 n=3

Applying result (33) with N = 3 gives

2
Be9 t
|u (x, t) u1 (x, t)| .
1 e62 t
Thus, in this case, the error between the solution u (x, t) and the rst term u1 (x, t)
2
decays as e9 t - very quickly!

5 Review of Fourier Series

Ref: Guenther & Lee, 3.1, Myint-U & Debnath 5.15.3, 5.55.6
[Sept 19, 2006]
Motivation: Recall that the initial temperature distribution satises


f (x) = u (x, 0) = Bn sin (nx) .
n=1

In the example above with a constant initial temperature distribution, f (x) = u0 , we

have
4u0 sin ((2n 1) x)
u0 = . (35)
n=1 2n 1
Note that at x = 0 and x = 1, the r.h.s. does NOT converge to u0 = 0, but rather to
0 (the BCs). Note that the Fourier Sine Series of f (x) is odd and 2-periodic in space
and converges to the odd periodic extension of f (x) = u0 .

16
 Odd periodic extension The odd periodic extension of a function f (x) dened
for x [0, 1], is the Fourier Sine Series of f (x) evaluated at any x R,

f (x) =

Bn sin (nx) .
n=1

Note that since sin (nx) = sin (nx) and sin (nx) = sin (n (x + 2)), then
f (x) = f (x) and f (x + 2) = f (x). Thus f (x) is odd, 2-periodic and f (x)
equals f (x) on the open interval (0, 1). What conditions are necessary for f (x) to
equal f (x) on the closed interval [0, 1]? This is covered next.
Aside: cancelling u0 from both sides of (35) gives a really complicated way of
writing 1,

4 sin ((2n 1) x)
1= .
n=1 2n 1

5.1 Fourier Sine Series

Given an integrable function f (x) on [0, 1], the Fourier Sine Series of f (x) is


Bn sin (nx) (36)
n=1

where 1
Bn = 2 f (x) sin (nx) dx. (37)
0
The associated orthogonality properties are

1
1/2, m = n = 0,

sin (mx) sin (nx) dx =
0 0, m= n.

5.2 Fourier Cosine Series

Given an integrable function f (x) on [0, 1], the Fourier Cosine Series of f (x) is


A0 + An cos (nx) (38)
n=1

where
1
A0 = f (x) dx, (average of f (x)) (39)
0
1
An = 2 f (x) cos (nx) dx, n 1. (40)
0

17

The associated orthogonality properties are

1
1/2, m = n = 0,

cos (mx) cos (nx) dx =
0 0, m= n.

e.g. The cosine series of f (x) = u0 for x [0, 1] is just u0 . In other words, A0 = u0
and An = 0 for n 1.

5.3 The (full) Fourier Series

The Full Fourier Series of an integrable function f (x), now dened on [1, 1], is

f (x) = a0 +

(an cos (nx) + bn sin (nx)) (41)
n=1

where
1 1

a0 = f (x) dx
2 1
1
an = f (x) cos (nx) dx
1
1
bn = f (x) sin (nx) dx
1

The associated orthogonality properties of sin and cos on [1, 1] are, for any m, n =
1, 2, 3, ... 1
sin (mx) cos (nx) dx = 0, all m, n,
1

1
1, m = n = 0,

sin (mx) sin (nx) dx =
1 0, m = n,

1
1, m = n = 0,

cos (mx) cos (nx) dx =
1 0, m = n.

5.4 Piecewise Smooth

Ref: Guenther & Lee p. 50,
Provided a function f (x) is integrable, its Fourier coecients can be calculated.
It does not follow, however, that the corresponding Fourier Series (Sine, Cosine or
Full) converges or has the sum f (x). In order to ensure this, f (x) must satisfy some
stronger conditions.

18
 Denition Piecewise Smooth function: A function f (x) dened on a closed
interval [a, b] is said to be piecewise smooth on [a, b] if there is a partition of [a, b],

a = x0 < x1 < x2 < < xn = b

such that f has a continuous derivative (i.e. C 1 ) on each closed subinterval [xm , xm+1 ].
E.g. any function that is C 1 on an interval [a, b] is, of course, piecewise smooth on
[a, b].
E.g. the function
2x, 0 x 1/2
f (x) =
1/2, 1/2 < x 1
is piecewise smooth on [0, 1], but is not continuous on [0, 1].
E.g. the function f (x) = |x| is both continuous and piecewise smooth on [1, 1],
despite f (x) not being dened at x = 0. This is because we partition [1, 1] into
two subintervals [1, 0] and [0, 1]. When worrying about f (x) near x = 0, note that
on [0, 1] we only care about the left limit f (0 ) and for [1, 0], we only care about
the right limit f (0+ ).
E.g. the function f (x) = |x|1/2 is continuous on [1, 1] but not piecewise smooth
on [1, 1], since f (0 ) and f (0+ ) do not exist.

5.5 Convergence of Fourier Series

Ref: Guenther & Lee p. 49 and (optional) 3.3, Myint-U & Debnath 5.10
Theorem [Convergence of the Fourier Sine and Cosine Series]: If f (x) is piecewise
smooth on the closed interval [0, 1] and continuous on the open interval (0, 1), then
the Fourier Sine and Cosine Series converge for all x [0, 1] and have the sum f (x)
for all x (0, 1).
Note: Suppose f (x) is piecewise smooth on [0, 1] and is continuous on (0, 1) except
at a jump discontinuity at x = a. Then the Fourier Sine and Cosine Series converge
to f (x) on (0, 1) and converge to the average of the left and right limits at x = a,
i.e. (f (a) + f (a+)) /2. At the endpoints x = 0, 1, the Sine series converges to zero,
since sin (nx) = 0 at x = 0, 1 for all n.
Theorem [Convergence of the Full Fourier Series]: If f (x) is piecewise smooth on
the closed interval [1, 1] and continuous on the open interval (1, 1), then the Full
Fourier Series converges for all x [1, 1] and has the sum f (x) for all x (1, 1).

5.6 Comments
Ref: see problems Guenther & Lee p. 53

19
 Given a function f (x) that is piecewise smooth on [0, 1] and continuous on (0, 1),
the Fourier Sine and Cosine Series of f (x) converge on [0, 1] and equal f (x) on the
open interval (0, 1) (i.e. perhaps excluding the endpoints). Thus, for any x (0, 1),



A0 + An cos (nx) = f (x) = Bn sin (nx) .
n=1 n=1

In other words, the Fourier Cosine Series (left hand side) and the Fourier Sine Series
(right hand side) are two dierent representations for the same function f (x), on the
open interval (0, 1). The values at the endpoints x = 0, 1 may not be the same. The
choice of Sine or Cosine series is determined from the type of eigenfunctions that give
solutions to the Heat Equation and BCs.

6 Well-Posed Problems
Ref: Guenther & Lee 1.8 and 5.2 (in particular p. 160 and preceding pages, and p.
150)
Ref: Myint-U & Debnath 1.2, 6.5
We call a mathematical model or equation or problem well-posed if it satises the
following 3 conditions:

1. [Existence of a solution]: The mathematical model has at least 1 solution.

Physical interpretation: the system exists over at least some nite time interval.

2. [Uniqueness of solution]: The mathematical model has at most 1 solution. Phys

ical interpretation: identical initial states of the system lead to the same out
come.

3. [Continuous dependence on parameters]: The solution of the mathematical

model depends continuously on initial conditions and parameters. Physical
interpretation: small changes in initial states (or parameters) of the system
produce small changes in the outcome.

If an IVP (initial value problem) or BIVP (boundary initial value problem - e.g.
Heat Problem) satises 1, 2, 3 then it is well-posed.
Example: for the basic Heat Problem, we showed 1 by construction a solution
using the method of separation of variables. Continuous dependence is more dicult
to show (need to know about norms), but it is true, and we will use this fact when
sketching solutions. Also, when drawing level curves u (x, t) = const, small changes
in parameters (x, t) leads to a small change in u. We now prove the 2nd part of
well-posedness, uniqueness of solution, for the basic heat problem.

20
6.1 Uniqueness of solution to the Heat Problem
Denition We dene the two space-time sets

D = {(x, t) : 0 x 1, t > 0}
D = {(x, t) : 0 x 1, t 0}

and the space of functions

C 2 D = u (x, t) : uxx continuous in D and u continuous in D .

In other words, the space of functions that are twice continuously dierentiable on
[0, 1] for t > 0 and continuous on [0, 1] for t 0.
Theorem The basic Heat Problem, i.e. the Heat Equation (8) with BC (10)
and IC (9),

PDE: ut = uxx , 0 < x < 1,

IC: u (x, 0) = f (x) , 0 < x < 1,
BC: u (0, t) = u (1, t) = 0, t > 0,

has at most one solution in the space of functions C 2 D .

Proof : Consider two solutions u1 , u2 C 2 D to the Heat Problem. Let v =

u1 u2 . We aim to show that v = 0 on [0, 1], which would prove that u1 = u2 and
then solution to the Heat Equation (8) with BC (10) and IC (9) is unique. Since each
of u1 , u2 satises (8), (9), and (10), the function v satises

vt = (u1 u2 )t (42)
= u1t u2t
= u1xx u2xx
= (u1 u2 )xx
= vxx , 0 < x < 1,

and similarly,

IC: v (x, 0) = u1 (x, 0) u2 (x, 0) = f (x) f (x) = 0, 0 < x < 1, (43)

BC: v (0, t) = u1 (0, t) u2 (0, t) = 0, v (1, t) = 0, t > 0. (44)

Dene the function

1
V (t) = v 2 (x, t) dx 0, t 0.
0

21
Showing that v (x, t) = 0 reduces to showing that V (t) = 0, since v (x, t) is continuous
on [0, 1] for all t 0 and if there was a point x such that v (x, t) = 0, then V (t)
would be strictly greater than 0. To show V (t) = 0, we dierentiate V (t) in time
and substitute for vt from the PDE (42),
1 1
dV
=2 vvt dx = 2 vvxx dx
dt 0 0

Integrating by parts (note: vvxx = (vvx )x vx2 ) gives

1 1
dV 1 2
=2 vvxx dx = 2 vvx |x=0 vx dx
dt 0 0

Using the BCs (44) gives

1
dV

= 2 vx2 dx 0.
dt 0
The IC (43) implies that
1
V (0) = v 2 (x, 0) dx = 0.
0

Thus, V (t) 0, dV /dt 0, and V (0) = 0, i.e. V (t) is a non-negative, non-

increasing function of time whose initial value is zero. Thus, for all time, V (t) = 0
and v (x, t) = 0 for all x [0, 1], implying that u1 = u2 . This proves that the solution
to the Heat Equation (8), its IC (9), and BCs (10) is unique, i.e. there is at most one
solution. Uniqueness proofs for other types of BCs follows in a similar manner.

7 Variations on the basic Heat Problem

[Sept 21, 2006]
We now consider variations to the basic Heat Problem, including dierent types
of boundary conditions and the presence of sources and sinks.

7.1 Boundary conditions

7.1.1 Type I BCs (Dirichlet conditions)

Ref: Guenther & Lee p. 149

Type I, or Dirichlet, BCs specify the temperature u (x, t) at the end points of the
rod, for t > 0,

u (0, t) = g1 (t) ,
u (1, t) = g2 (t) .

22
Type I Homogeneous BCs are

u (0, t) = 0,
u (1, t) = 0.

The physical signicance of these BCs for the rod is that the ends are kept at 0o C.
The solution to the Heat Equation with Type I BCs was considered in class. After
separation of variables u (x, t) = X (x) T (t), the associated Sturm-Liouville Boundary
Value Problem for X (x) is

X + X = 0; X (0) = X (1) = 0.

The eigenfunctions are Xn (x) = Bn sin (nx).

7.1.2 Type II BCs (Newmann conditions)

Ref: Guenther & Lee p. 152 problem 1

Type II, or Newmann, BCs specify the rate of change of temperature u/x (or
heat ux) at the ends of the rod, for t > 0,
u
(0, t) = g1 (t) ,
x
u
(1, t) = g2 (t) .
x
Type II Homogeneous BCs are
u
(0, t) = 0,
x
u
(1, t) = 0.
x
The physical signicance of these BCs for the rod is that the ends are insulated.
These lead to another relatively simple solution involving a cosine series (see problem
6 on PS 1). After separation of variables u (x, t) = X (x) T (t), the associated Sturm-
Liouville Boundary Value Problem for X (x) is

X + X = 0; X (0) = X (1) = 0.

The eigenfunctions are X0 (x) = A0 = const and Xn (x) = An cos (nx).

7.1.3 Type III BCs (Mixed)

The general Type III BCs are a mixture of Type I and II, for t > 0,
u
1 (0, t) + 2 u (0, t) = g1 (t) ,
x
u
3 (1, t) + 4 u (1, t) = g2 (t) .
x

23
After separation of variables u (x, t) = X (x) T (t), the associated Sturm-Liouville
Boundary Value Problem for X (x) is

X + X = 0,

1 X (0) + 2 X (0) = 0
3 X (0) + 4 X (0) = 0

The associated eigenfunctions depend on the values of the constants 1,2,3,4 .

Example 1. 1 = 4 = 1, 2 = 3 = 0. Then

X + X = 0; X (0) = X (1) = 0

and the eigenfunctions are Xn = An cos 2n21 x . Note: the constant X0 = A0 is not

an eigenfunction here.
Example 2. 1 = 4 = 0, 2 = 3 = 1. Then

X + X = 0; X (0) = X (1) = 0

and the eigenfunctions are Xn = Bn sin 2n21 x . Note: the constant X0 = A0 is not

an eigenfunction here.
Note: Starting with the BCs in Example 1 and rotating the rod about x = 1/2,
youd get the BCs in Example 2. It is not surprising then that under the change of
variables x = 1x, Example 1 becomes Example 2, and vice versa. The eigenfunctions
also possess this symmetry, since

2n 1 n 2n 1
sin (1 x) = (1) cos x .
2 2

Since we can absorb the (1)n into the constant Bn , the eigenfunctions of Example
1 become those of Example 2 under the transformation x = 1 x, and vice versa.

7.2 Solving the Heat Problem with Inhomogeneous (time

independent) BCs
Ref: Guenther & Lee p. 149
Consider the Heat Problem with inhomogeneous Type I BCs,

ut = uxx , 0<x<1
u (0, t) = 0, u (1, t) = u1 = const, t > 0, (45)
u (x, 0) = 0, 0 < x < 1.

24
Directly applying separation of variables u (x, t) = X (x) T (t) is not useful, because
wed obtain X (1) T (t) = u1 for t > 0. The strategy is to rewrite the solution u (x, t)
in terms of a new variable v (x, t) such that the new problem for v has homogeneous
BCs!
Step 1. Find the steady-state, or equilibrium solution uE (x), since this by deni
tion must satisfy the PDE and the BCs,
uE = 0, 0<x<1
uE (0) = 0, uE (1) = u1 = const.
Solving for uE gives uE (x) = u1 x.
Step 2. Transform variables by introducing a new variable v (x, t),
v (x, t) = u (x, t) uE (x) = u (x, t) u1 x. (46)
Substituting this into the Heat Problem (45) gives a new Heat Problem,
vt = vxx , 0<x<1
v (0, t) = 0, v (1, t) = 0, t > 0, (47)
v (x, 0) = u1 x, 0 < x < 1.
Notice that the BCs are now homogeneous, and the IC is now inhomogeneous. Notice
also that we know how to solve this - since its the basic Heat Problem! Based on our
work, we know that the solution to (47) is

1
n2 2 t
v (x, t) = Bn sin (nx) e , Bn = 2 f (x) sin (nx) dx (48)
n=1 0

where f (x) = u1 x. Substituting for f (x) and integrating by parts, we nd

1
Bn = 2u1 x sin (nx) dx
0
1 1
x cos (nx) 1
= 2u1 + cos (nx) dx
n
x=0 n 0
2u1 (1)n
= (49)
n

Step 3. Transform back to u (x, t), from (46), (48) and (49),

2u1 (1)n 2 2
u (x, t) = uE (x) + v (x, t) = u1 x + sin (nx) en t . (50)
n=1 n
The term uE (x) is the steady state and the term v (x, t) is called the transient, since
it exists initially to satisfy the initial condition but vanishes as t . You can check
for yourself by direct substitution that Eq. (50) is the solution to the inhomogeneous
Heat Problem (45), i.e. the PDE, BCs and IC.

25
7.3 Heat sources
7.3.1 Derivation

Ref: Guenther & Lee p. 6 (Eq. 3-3)

To add a heat source to the derivation of the Heat Equation, we modify the energy
balance equation to read,

change of
heat in from heat out from heat generated
heat energy of = + .
left boundary right boundary in segment
segment in time t

Let Q (x, t) be the heat generated per unit time per unit volume at position x in the
rod. Then the last term in the energy balance equation is just QAxt. Applying
Fouriers Law (1) gives

u u
cAxu (x, t + t) cAxu (x, t) = tA K0 tA K0
x x x x+x
+QAxt

The last term is new; the others we had for the rod without sources. Dividing by
Axt and rearranging yields
u u
u (x, t + t) u (x, t) K0 x x+x
x x Q
= + .
t c x c

Taking the limit t, x 0 gives the Heat Equation with a heat source,

u 2u Q
= 2 + (51)
t x c

Introducing non-dimensional variables x = x/l, t = t/l2 gives

u 2 u l2 Q
= + (52)
t x2 c

Dening the dimensionless source term q = l2 Q/ (c) and dropping tildes gives the
dimensionless Heat Problem with a source,

u 2u
= +q (53)
t x2

26

7.3.2 Solution method

Ref: Guenther & Lee p. 147 149, Myint-U & Debnath 6.7 (exercises)
The simplest case is that of a constant source q = q (x) in the rod. The Heat
Problem becomes,

ut = uxx + q (x) , 0 < x < 1, (54)

u (0, t) = b1 , u (1, t) = b2 , t > 0,
u (x, 0) = f (x) , 0 < x < 1.

If q (x) > 0, heat is generated at x in the rod; if q (x) < 0, heat is absorbed.
The solution method is the same as that for inhomogeneous BCs: nd the equi
librium solution uE (x) that satises the PDE and the BCs,

0 = uE + q (x) , 0 < x < 1,

uE (0) = b1 , uE (1) = b2 .

Then let
v (x, t) = u (x, t) uE (x) .
Substituting u (x, t) = v (x, t) + uE (x) into (54) gives a problem for v (x, t),

vt = vxx , 0 < x < 1,

v (0, t) = 0, v (1, t) = 0, t > 0,
v (x, 0) = f (x) uE (x) , 0 < x < 1.

Note that v (x, t) satises the homogeneous Heat Equation (PDE) and homogeneous
BCs, i.e. the basic Heat Problem. Solve the Heat Problem for v (x, t) and then obtain
u (x, t) = v (x, t) + uE (x).
Note that things get complicated if the source is time-dependent - we wont see
that in this course.

7.4 Periodic boundary conditions

Ref: Guenther & Lee p. 189-190; for alternate method, see Guether & Lee p. 149
and then p. 147
[Sept 26, 2006]
Above, we solved the heat problem with inhomogeneous, but time-independent,
BCs by using the steady-state. We now show how to solve the heat problem with inho
mogeneous, but time-varying, BCs. We consider the heat problem with an oscillatory

27

(and periodic) BC,

ut = uxx , 0<x<1
u (0, t) = A cos t, u (1, t) = 0, t > 0, (55)
u (x, 0) = f (x) , 0 < x < 1.

The physical meaning of the BC u (0, t) = A cos t is that we keep changing, in a

periodic fashion, the temperature at the end x = 0 of the rod.
We dont expect the solution to be independent of time as t , since were
changing the temperature periodically at one end. However, we do expect that after
an initial transient time, the solution will become periodic with angular frequency ,
i.e.
u (x, t) = v (x, t) + A (x) cos (t + (x))
where v (x, t) 0 as t is the transient, A (x) cos (t + (x)) is what we call
the quasi-steady state, A (x) and (x) are the amplitude and phase of the quasi
steady state. To solve the problem, the goal is to rst nd A (x) and (x), and then
v (x, t), if necessary. Often we might not care about the transient state, if we are
more interested in the solution after long times.

7.4.1 Complexify the problem

We use the notation Re {z}, Im {z} to denote the real and imaginary parts of a
complex number z. Note that
1
(z + z ) , cos = Re ei

Re {z} =
2
1
(z z ) , sin = Im ei

Im {z} =
2i
where asterisks denote the complex conjugate ((x + iy) = x iy). Thus, we can
write our quasi-steady solution in terms of complex exponentials,

A (x) cos (t + (x)) = Re A (x) ei(x) eit = Re U (x) eit

where, for convenience, we have replaced A (x) ei(x) with the complex function U (x).
We do this because complex exponentials are much easier to work with than cos (t)
and sin(t). Note that U (x) has magnitude A (x) = |U (x)| and phase (x) =
Im {U (x)}
arctan Re{U (x)} . The phase (x) delays the eects of what is happening at the end
of the rod: if the end is heated at time t = t1 , the eect is not felt at the center until
a later time t = (1/2) / + t1 . The following result will be useful.

28
 Lemma [Zero sum of complex exponentials] If, for two complex constants a, b,
we have
aeit + beit = 0 (56)
for all times t in some open interval, then a = b = 0.
Proof: Dierentiate (56) in time t,

i aeit beit = 0.

(57)

Adding (56) to 1/i(57) gives

2aeit = 0.
Since eit is never zero (|eit | = 1), then a = 0. From (56), beit = 0 and hence
b = 0.
Note that we could also use the Wronskian to show this:

it it eit eit
W e ,e = det = 2i = 0
ieit ieit

and hence eit and eit are linearly independent, meaning that a = b = 0.

7.4.2 ODE and ICs for quasi-steady state

Step 1. Find the quasi-steady state solution to the PDE and BCs of the Heat Problem
(55) of the form
1
uSS (x, t) = Re U (x) eit = U (x) eit + U (x) eit = A (x) cos (t + (x))

2
(58)
where U (x) is a complex valued function. Substituting (58) for u (x, t) into the PDE
in (55) gives
1 1
iU (x) eit iU (x) eit = U (x) eit + U (x) eit ,

2 2
for 0 < x < 1 and t > 0. Multiplying both sides by 2 and re-grouping terms yields

(iU (x) U (x)) eit + (iU (x) U (x)) eit = 0, 0 < x < 1, t > 0.
(59)
Applying the Lemma to (59) gives

iU (x) U (x) = 0 = iU (x) U (x) (60)

29
Note that the left and right hand sides are the complex conjugates of one another,
and hence they both say the same thing (so from now on well write one or the other).
Substituting (58) into the BCs in (55) gives

1 A it 1
U (0) eit + U (0) eit = e + eit , U (1) eit + U (1) eit = 0,

2 2 2
(61)
for t > 0. Grouping the coecients of eit and applying the Lemma yields

U (0) = A, U (1) = 0. (62)

To summarize, the problem for the complex amplitude U (x) of the quasi-steady
state uSS (x, t) is, from (60) and (62),

U (x) iU (x) = 0; U (0) = A, U (1) = 0. (63)

Note that (1 + i)2 = 2i, and hence

2
1
i = (1 + i)2 = (1 + i) .
2 2

Therefore, (63) can be rewritten as

2

U (1 + i) U = 0; U (0) = A, U (1) = 0. (64)
2

7.4.3 Solving for quasi-steady state

Solving the ODE (64) gives

U = c1 exp (1 + i) x + c2 exp (1 + i) x (65)
2 2

where c1 , c2 are integration constants. Imposing the BCs gives

A = U (0) = c1 + c2 , 0 = U (1) = c1 exp (1 + i) + c2 exp (1 + i) .
2 2

Solving this set of linear equations for the unknowns (c1 , c2 ) and substituting these
back into gives

A exp 2
(1 + i)
c1 =
exp 2
(1 + i) exp 2
(1 + i)
A exp 2 (1 + i)

c2 = A c1 =
exp 2
(1 + i) exp 2
(1 + i)

30
 1 0

0.8
0.1

phase(U(x))
0.6
|U(x)|/A

0.2
0.4

0.3
0.2

0 0.4
0 0.5 1 0 0.5 1
x x

Figure 4: At left, the magnitude of U (x) (solid) and |U (x)| (dash). At right, the
phase of U (x).

Substituting these into (65) gives

(1 + i) (1 x) exp 2 (1 + i) (1 x)

exp 2
U = A
.

exp
2
(1 + i)
exp
2
(1 + i)

Therefore, the quasi-steady-state solution to the heat problem is

exp 2
(1 + i) (1 x) exp (1 + i) (1 x)
uSS (x, t) = Re 2 Aeit .

exp
2
(1 + i)
exp
2 (1 + i)

It is easy to check that uSS (x, t) satises the PDE and BCs in (55). Also, note that
the square of the magnitude of the denominator is
2

exp
(1 + i) exp (1 + i)
= 2 cosh 2 cos 2
> 0

2 2

which is greater than zero since > 0 and hence cosh 2 > 1 cos 2. In Figure
4, the magnitude and phase of U (x) are plotted as solid lines. The straight dashed
line is drawn with |U (x)| for comparison, illustrating that |U (x)| is nearly linear in x.
The phase of U (x) is negative, indicating a delay between what happens at a point
x on the rod and what happens at the end x = 0.

7.4.4 Solving for the transient

Step 2. Solve for the transient, dened as before,

v (x, t) = u (x, t) uSS (x, t) . (66)

31

Substituting (66) into the heat problem (55), given that uSS (x, t) satises the PDE
and BCs in (55), gives the following problem for v (x, t),

vt = vxx , 0<x<1
v (0, t) = 0, v (1, t) = 0, t > 0, (67)
v (x, 0) = f2 (x) , 0 < x < 1,

where the initial condition f2 (x) is given by

f2 (x) = u (x, 0) uSS (x, 0)

exp 2
(1 + i) (1 x) exp (1 + i) (1 x)
= f (x) Re A 2 .
exp 2
(1 + i) exp 2 (1 + i)

The problem for v (x, t) is the familiar basic Heat Problem whose solution is given by
(25), (24) with f (x) replaced by f2 (x),

1
2 2
v (x, t) = Bn sin (nx) en t , Bn = 2 f2 (x) sin (nx) dx.
n=1 0

7.4.5 Full solution

The full solution to the problem is

(1 + i) (1 x) exp 2 (1 + i) (1 x)

exp 2 it
u (x, t) = Re Ae
(1 + i) exp 2 (1 + i)

exp 2

2 2 t
+ Bn sin (nx) en .

n=1

The rst term is the quasi-steady state, whose amplitude at each x is constant, plus
a transient part v (x, t) that decays exponentially as t . If the IC f (x) were
given, then we could compute the Bn s.

7.4.6 Similar problem: heating/cooling of earths surface

Consider a vertical column in the earths crust that is cooled in the winter and heated
in the summer, at the surface. We take the x-coordinate to be pointing vertically
downward with x = 0 corresponding to the earths surface. For simplicity, we model
the column of earth by the semi-innite line 0 x < . We crudely model the
heating and cooling at the surface as u (0, t) = A cos t, where = 2/ and the
(scaled) period corresponds to 1 year. Under our scaling, = (1 year)/l2 . The
boundary condition as x is that the temperature u is bounded ( is at the

32

bottom of the earths crust, still far away from the core, whose eects are neglected).
What is the quasi-steady state?
The quasi-steady state satises the Heat Equation and the BCs,

(uSS )t = (uSS )xx , 0<x< (68)

uSS (0, t) = T0 + T1 cos t, uSS bounded as x , t > 0.

We use superposition: u (x, t) = u0 + u1 , where

(u0 )t = (u0 )xx , (u1 )t = (u1 )xx , 0 < x < , t > 0.

u0 (0, t) = T0 , u1 (0, t) = T1 cos t, u0 , u1 bounded as x ,

Obviously, u0 (x, t) = T0 works, and by uniqueness, we know this is the only solution
for u0 (x, t). To solve for u1 , we proceed as before and let u1 (x, t) = Re {U (x) eit }
to obtain

U (x) iU (x) = 0, 0<x< (69)

U (0) = T1 , U bonded as x , t > 0.

The general solution to the ODE (69) is

U = c1 exp (1 + i) x + c2 exp (1 + i) x .
2 2

The boundedness criterion gives c2 = 0, since that term blows up as x . The

BC at the surface (x = 0) gives c1 = T1 . Hence


U = T1 exp (1 + i) x .
2

Putting things together, we have

it
uSS (x, t) = T0 + Re T1 exp (1 + i) x e
2



x
= T0 + T1 e 2 Re exp i x + it

2



x
= T0 + T1 e 2 cos
x + t . (70)

uSS (x, t) is plotted at various dimensionless times

t/ = 0, 1/4, 1/2, 3/4, 1 in Figure
x
5. Dashed lines give the amplitude T0 T1 e 2 of the quasi-steady-state uSS (x, t).
Physical questions: What is the ideal depth for a wine cellar? We want the wine
to be relatively cool compared to the summer temperature and relatively warm to the

33
 0
0.75 1.75 0
1 1.25 1.5 0.5 0.25

5 0 5 10 15 20 25
uSS(x,t)

Figure 5: Plot of uSS (x, t) at various times. Numbers in gure indicate t/.

34

winter temperature, and yet we want the cellar close to the surface (to avoid climbing
too many stairs). Then we must nd the smallest depth x such that the temperature
uSS (x, t) will be opposite in phase to the surface temperature uSS (0, t). We take
= 2 103 cm2 /s and l = 1 m. Recall that the period is 1 year, = (/l2 ) (1 year)
and 1 year is 3.15 107 s. From the solution (70), the phase of uSS (x, t) is reversed
when

x=
2
Solving for x gives

x = 2/
Returning to dimensional coordinates, we have

2
x = lx = l = (1 year) = (2 103 cm2 ) 3.15 107 = 4.45 m.
2
At this depth, the amplitude of temperature variation is

T1 e 2 x = T1 e 0.04T1

Thus, the temperature variations are only 4% of what they are at the surface. And
being out of phase with the surface temperature, the temperature at x = 4.45 m is
cold in the summer and warm in the winter. This is the ideal depth of a wine cellar.
Note: for a dierent solution to this problem using Laplace transforms (not covered
in this course), see Myint-U & Debnath Example 11.10.5.

8 Linearity, Homogeneity, and Superposition

Ref: Myint-U & Debnath 1.1, 1.3, 1.4
[Sept 28, 2006]
Denition Linear space: A set V is a linear space if, for any two elements1 v1 ,
v2 V and any scalar (i.e. number) k R, the terms v1 +v2 and kv1 are also elements
of V.
E.g. Let S denote the set of functions that are C 2 (twice-continuously dieren
tiable) in x and C 1 (continuously dierentiable) in t, for x [0, 1], t 0. We write S
as
S = {f (x, t) | fxx , ft continuous for x [0, 1] , t 0} (71)
If f1 , f2 S, i.e. f1 , f2 are functions that have continuous second derivatives in space
and continuous derivatives in time, and k R, then f1 + f2 and kf1 also have the
1
Note that the symbol means an element of. So x [0, l] means x is in the interval [0, l].

35

same, and hence are both elements of S. Thus, by denition, S forms a linear space
2
over the real numbers R. For instance, f1 (x, t) = sin (x) e t and f2 = x2 + t3 .
Denition Linear operator: An operator L : V W is linear if

L (v1 + v2 ) = L (v1 ) + L (v2 ) , (72)

L (kv1 ) = kL (v1 ) , (73)

for all v1 , v2 V, k R. The rst property is the summation property; the second is
the scalar multiplication property.
The term operator is very general. It could be a linear transformation of vectors
in a vector space, or the derivative operation /x acting on functions. Dierential
operators are just operators that contain derivatives. So /x and /x + /y are
dierential operators. These operators are functions that act on other functions. For
example, the x-derivative operation is linear on a space of functions, even though the
functions the derivative acts on might not be linear in x:
2
cos x + x2 =

(cos x) + x .
x x x
E.g. the identity operator I, which maps each element to itself, i.e. for all elements
v V we have I (v) = v. Check for yourself that I satises (72) and (73) for an
arbitrary linear space V.
E.g. consider the partial derivative operator acting on S (dened in (71)),

u
D (u) = , u S.
x
From the well-known properties of the derivative, namely
u v
(u + v) = + , u, v S,
x x x
u
(ku) = k , u S, k R
x x
it follows that the operator D satises properties (72) and (73) and is therefore a
linear operator.
E.g. Consider the operator that denes the Heat Equation

2
L= 2
t x
and acts over the linear space S of functions C 2 in x, C 1 in t. The Heat Equation can
be written as
u 2 u
L (u) = 2 = 0.
t x

36

For any two functions u, v S and a real k R,

2


L (u + v) = (u + v)
t x2
2 u 2 u v 2 v
= (u + v) 2 (u + v) = 2+ 2 = L (u) + L (v)
t x t x t x
2 2 u 2 u


L (ku) = (ku) = (ku) 2 (ku) = k 2 = kL (u)
t x2 t x t x
Thus, L satises properties (72) and (73) and is therefore a linear operator.

8.1 Linear and homogeneous PDE, BC, IC

Consider a dierential operator L and operators B1 , B2 , I that dene the following
problem:

PDE: L (u) = h (x, t) , with

BC: B1 (u (0, t)) = g1 (t) , B2 (u (1, t)) = g2 (t)
IC: I (u (x, 0)) = f (x) .

Denition The PDE is linear if the operator L is linear.

Denition The PDE is homogeneous if h (x, t) = 0.
Denition The BCs are linear if B1 , B2 are linear. The BCs are homogeneous if
g1 (t) = g2 (t) = 0.
Denition The IC is linear if I is a linear. The IC is homogeneous if f (x) = 0.

8.2 The Principle of Superposition

The Principle of Superposition has three parts, all which follow from the denition
of a linear operator:
(1) If L (u) = 0 is a linear, homogeneous PDE and u1 , u2 are solutions, then
c1 u1 + c2 u2 is also a solution, for all c1 , c2 R.
(2) If u1 is a solution to L (u1 ) = f1 , u2 is a solution to L (u2 ) = f2 , and L is
linear, then c1 u1 + c2 u2 is a solution to

L (u) = c1 f1 + c2 f2 ,

for all c1 , c2 R.
(3) If a PDE, its BCs, and its IC are all linear, then solutions can be superposed
in a manner similar to (1) and (2).

37
 The proofs of (1)-(3) follow from the denitions above. For example, the proof of
(1) is

L (c1 u1 + c2 u2 ) = L (c1 u1 ) + L (c2 u2 )

= c1 L (u1 ) + c2 L (u2 ) = c1 0 + c2 0 = 0

The rst step follows from the addition rule of linear operators; the second from the
scalar multiplication rule, and the third from the fact that u1 and u2 are solutions of
L (u) = 0. The proofs of parts (2) and (3) are similar.
E.g. Suppose u1 is a solution to

L (u) = ut uxx = 0
B1 (u (0, t)) = u (0, t) = 0
B2 (u (1, t)) = u (1, t) = 0
I (u (x, 0)) = u (x, 0) = 100

and u2 is a solution to

L (u) = ut uxx = 0
B1 (u (0, t)) = u (0, t) = 100
B2 (u (1, t)) = u (1, t) = 0
I (u (x, 0)) = u (x, 0) = 0

Then 2u1 u2 would solve

L (u) = ut uxx = 0
B1 (u (0, t)) = u (0, t) = 100
B2 (u (1, t)) = u (1, t) = 0
I (u (x, 0)) = u (x, 0) = 200

To solve L (u) = g with inhomogeneous BCs using superposition, we can solve

two simpler problems: L (u) = g with homogeneous BCs and L (u) = 0 with inhomo
geneous BCs.

8.3 Application to the solution of the Heat Equation

Recall that we showed that un (x, t) satised the PDE and BCs. Since the PDE (8)
and BCs (10) are linear and homogeneous, then we apply the Principle of Superposi
tion (1) repeatedly to nd that the innite sum u (x, t) given in (80) is also a solution,
provided of course that it can be dierentiated term-by-term.

38
9 Uniform convergence, dierentiation and inte
gration of an innite series
Ref: Guenther & Lee Ch 3 talk about these issue for the Fourier Series, in particular
uniform convergence p. 50, dierentiation and integration of an innite series on p.
58 and 59, and the Weirstrauss M-Test on p. 59.
Ref: Myint-U & Debnath 5.10, 5.13 (problem 12 p. 136)
The innite series solution (25) to the Heat Equation (8) only makes sense if it
converges uniformly2 on the interval [0, 1]. The reason is that to satisfy the PDE,
we must be able to integrate and dierentiate the innite series term-by-term. This
can only be done if the innite series AND its derivatives converge uniformly. The
following results3 dictate when we can dierentiate and integrate an innite series
term-by-term.
Denition [Uniform convergence of a series] The series


fn (x)
n=1

of functions fn (x) dened on some interval [a, b] converges if for every > 0, there
exits an N0 () 1 such that


fn (x) < , for all N N0 .



n=N

Theorem [Term-by-term dierentiation] If, on an interval x [a, b],

1. f (x) = fn (x) converges uniformly,
n=1

2. fn (x) converges uniformly,
n=1
3. fn (x) are continuous,

then the series may be dierentiated term-by-term,

f (x) = fn (x) .
n=1

2
The precise denitions are outlined in 9.4 (optional reading).
3
Proofs of these theorems [optional reading] can be found in any good text on real analysis [e.g.
Rudin] - start by reading 9.4.

39
 Theorem [Term-by-term integration] If, on an interval x [a, b],


1. f (x) = fn (x) converges uniformly,
n=1

2. fn (x) are integrable,

then the series may be integrated term-by-term,
b
b
f (x) dx = fn (x) dx.
a n=1 a

Thus, uniform convergence of an innite series of functions is an important prop

erty. To check that an innite series has this property, we use the following tests in
succession (examples to follow):
Theorem [Weirstrass M-Test for uniform convergence of a series of functions]
Suppose {fn (x)}n=1 is a sequence of functions dened on an interval [a, b], and sup
pose
|fn (x)| Mn (x [a, b] , n = 1, 2, 3, ...).

Then the series of functions n=1 fn (x) converges uniformly on [a, b] if the series of
numbers

n=1 Mn converges absolutely.
Theorem [Ratio Test for convergence of a series of numbers] The series

n=1 an
converges absolutely if the ratio of successive terms is less than a constant r < 1, i.e.

|an+1 |
r < 1, (74)
|an |

for all n N 1.
Note: N is present to allow the rst N 1 terms in the series not to obey the
ratio rule (74).
Theorem [Convergence of an alternating series] Suppose

1. |a0 | |a1 | |a2 |

2. a2n1 0, a2n 0 (n = 1, 2, 3 . . .)
3. limn an = 0

then the sum

an
n=1
converges.
Note: This is not an absolute form of convergence, since the series n

n=1 (1) /n

converges, but n=1 1/n does not.

40

9.1 Examples for the Ratio Test

E.g. Consider the innite series of numbers
n
1
(75)
n=1
2
n
Writing this as a series n=1 an , we identify an = (1/2) . We form the ratio of
successive elements in the series,

|an+1 | (1/2)n+1 1
= n = <1
|an | (1/2) 2

Thus, the innite series (75) satises the requirements of the Ratio Test with r = 1/2,
and hence (75) converges absolutely.
E.g. Consider the innite series

1
(76)
n=1
n

Writing this as a series n=1 an , we identify an = 1/n. We form the ratio of successive
elements in the series,
|an+1 | 1/ (n + 1) n
= =
|an | 1/n n+1
n
Note that limn n+1 = 1, and hence there is no upper bound r < 1 that is greater
than |an+1 | / |an | for ALL n. So the Ratio Test fails, i.e. it gives no information. It
turns out that this series diverges, i.e. the sum is innite.
E.g. Consider the innite series

1
(77)
n=1
n2

Again, writing this as a series 2

n=1 an , we identify an = 1/n . We form the ratio of
successive elements in the series,

|an+1 | 1/ (n + 1)2 n2
= =
|an | 1/n2 (n + 1)2
n 2
Note again that limn (n+1) 2 = 1, and hence there is no r < 1 that is greater than

|an+1 | / |an | for ALL n. So the Ratio Test gives no information again. However, it
turns out that this series converges:

1 2
=
n=1
n2 6

41

So the fact that the Ratio Test fails does not imply anything about the convergence
of the series!
Note that the innite series
1
(78)
n=1
np
converges for p > 1 and diverges (is innite) for p 1.

9.2 Examples of series of functions

Note,


sin (nx)
n=1

does not converge at certain points in [0, 1], and hence it cannot converge uniformly
on the interval. In particular, at x = 1/2, we have


n
sin (nx) = sin = 1 + 0 1 + 0 + 1 + 0 1 +
n=1 n=1
2

The partial sums (i.e. sums of the rst n terms) change from 1 to 0 forever. Thus the
sum does not converge, since otherwise the more terms we add, the closer the sum
must get to a single number.
Consider
sin (nx)
2
(79)
n=1
n
Since the nth term is bounded in absolute value by 1/n2 , and since 2

n=1 1/n con
verges (absolutely), then the Weirstrass M-Test says the sum (79) converges uniformly
on [0, 1].

9.3 Application to the solution of the Heat Equation

Ref: Guenther & Lee p. 145 147
We now use the Weirstrass M-Test and the Ratio Test to show that the innite
series solution (25) to the Heat Equation converges uniformly,


2 2 t
u (x, t) = un (x, t) = Bn sin (nx) en , (80)
n=1 n=1

for all positive time, i.e. t t0 > 0, and space x [0, 1], provided the initial condition
f (x) is piecewise continuous.

42
 To apply the M-Test, we need bounds on |un (x, t)|,

2 2 2 2
|un (x, t)| = Bn sin (nx) en t |Bn | en t0 , for all x [0, 1] . (81)

We now need a bound on the Fourier coecients |Bn |. Note that from Eq. (24),
1 1 1

|Bm | = 2
sin (mx) f (x) dx 2
|sin (mx) f (x)| dx 2 |f (x)| dx, (82)
0 0 0

for all x [0, 1]. To obtain the inequality (82), we used the fact that |sin (mx)| 1
and, for any integrable function h (x),
b b


h (x) dx
|h (x)| dx.
a a

Youve seen this integral inequality, I hope, in past Calculus classes. We combine (81)
and (82), to obtain
|un (x, t)| Mn (83)
where 1
2 2 t
Mn = 2 |f (x)| dx en 0
.
0

To apply the Weirstrass M-Test, we rst need to show that the innite series of
numbers

n=1 Mn converges absolutely (we will use the Ratio Test). Forming the
ratio of successive terms yields
2 2
Mn+1 e(n+1) t0 (n2 (n+1)2 )2 t0 = e(2n+1)2 t0 e2 t0 < 1,
= n2 2 t = e n = 1, 2, 3, ...
Mn e 0

2t
Thus, by the Ratio Test with r = e 0
< 1, the sum


Mn
n=1

converges absolutely, and hence by Eq. (81) and the Weirstrass M-Test,

n=1 un (x, t)
converges uniformly for x [0, 1] and t t0 > 0.
A similar argument holds for the convergence of the derivatives ut and uxx . Thus,
for all t t0 > 0, the innite series (80) for u may be dierentiated term-by-term
and since each un (x, t) satises the PDE and BCs, then so does u (x, t). Later, after
considering properties of Fourier Series, we will show that u converges even at t = 0
(given conditions on the initial condition f (x)).

43

9.4 Background [optional]

[Note: You are not responsible for the material in this subsection 9.4 - it is only added
for completeness]
Ref: Chapters 3 & 7 of Principles of Mathematical Analysis, W. Rudin, McGraw-
Hill, 1976.
Denition Convergence of a series of numbers: the series of real numbers

n=1 an
converges if for every > 0, there is an integer N such that for any n N ,


am < .



m=n
Denition Absolute convergence of a series of numbers: the series of real numbers

n=1 a n is said to converge absolutely if the series n=1 |an | converges.

Denition Uniform convergence of a sequence of functions: A sequence of func

tions {fn (x)} n=1 dened on a subset E R converges uniformly on E to a function

f (x) if for every > 0 there is an integer N such that for any n N ,
|fn (x) f (x)| <
for all x E.
Note: pointwise convergence does not imply uniform convergence, i.e. in the
denition, for each , one N works for all x in E. For example, consider the sequence
of functions {xn }
n=1 on the interval [0, 1]. These converge pointwise to the function

0, 0 x < 1,
f (x) =
1, x=1
on [0, 1], but do not converge uniformly.

Denition Uniform convergence of a series of functions: A series of functions

n=1 fn (x) dened on a subset E R converges uniformly on E to a function g (x)

if the partial sums
m

sm (x) = fn (x)
n=1
converge uniformly to g (x) on E.
Consider the following sum,


cos (nx)

n=1
n
does converge, but this is a weaker form of convergence (pointwise, not uniform or
absolute), by the Alternating Series Test (terms alternate in sign, absolute value of
the terms goes to zero as n ). Consult Rudin for the Alternating Series Test, if
desired.

44