LECTURE NOTES FOR MATH 157

ALEKSEI BELTUKOV

1. Orthogonal bases
In this section we discuss certain aspects of solving the heat equation
with symmetric boundary conditions; these aspects are known collec-
tively as Fourier theory. Recall from introduction that the following
IBVP
∂u ∂ 2u
= (Heat Equation in 1 dimension)
∂t ∂x2
u(x, 0) = u0 (x) (Initial condition)
¯
¯
u(x, t)¯¯ = 0 (Dirichlet boundary conditions)
x=0,L

was found to possess a unique Fourier series solution:

∞ ·Z L ³ π n x ´ ¸ π2 n2 ³π n x´
2 X − 2 t
u(x, t) = u0 (x) sin dx e L sin .
L n=1 0 L L
We were led to that series through the following steps:
• Separation of variables led us to special solutions of the Heat
Equation in product form
2
(1) u(x, t) = e−λ t
(A cos(λ x) + B sin(λ x)) ,
with λ, A, and B being arbitrary parameters.
• Imposing Dirichlet boundary conditions, led us to conclude that
the special solutions had to be restricted to a discrete set of
Fourier modes
π 2 n2
³π n ´
un (x, t) = Bn e− L2 t sin x , n = 1, 2, 3, . . .
L
In other words, in order to satisfy the boundary conditions, we
had to set
πn
A = 0, λ = λn = , n = 1, 2, 3, . . .
L
Date: today.
1
2 ALEKSEI BELTUKOV

• Using linearity of the Heat Equation, we combined the various

Fourier modes into the general solution:
∞
X π 2 n2
³π n x´
u(x, t) = Bn e− L2
t
sin .
n=1
L

The resulting Fourier series satisfies both the PDE and the ini-
tial conditions. Further it has infinitely many Fourier coeffi-
cients Bn which can be adjusted based on the initial condition.
• We noticed that the different sine modes were orthogonal under
the functional dot product
Z L
f (x) · g(x) = f (x) g(x) dx.
0

Indeed,
Z L ³π n x´ ³π n x´
sin sin dx = 0, m 6= n.
0 L L
Dotting both sides of the initial condition
∞
X ³π n x´
u0 (x) = u(x, 0) = Bn sin .
n=1
L

with individual harmonics, led us to the integral formula for the

n-th Fourier coefficient:
¡ ¢ Z L ³π n x´
u0 (x) · sin π Ln x 2
Bn = ¡π nx¢ ¡π nx¢ = u0 (x) sin dx.
sin L · sin L L 0 L

The purpose of this section is to explain the last development in separa-

tion of variables—namely, the orthogonality of Fourier modes. At the
present stage it may seem a coincidence. However, some experimenta-
tion with boundary conditions suggests that orthogonality is intrinsic
to Fourier theory. For instance, let us change Dirichlet boundary con-
ditions to Neumann boundary conditions:
∂u ∂ 2u
= (Heat Equation in 1 dimension)
∂t ∂x2
u(x, 0) = u0 (x) (Initial condition)
¯
∂u ¯
(x, t)¯¯ = 0 (Dirichlet boundary conditions)
∂x x=0,L
 LECTURE NOTES FOR MATH 157 3

Separation of variables still gives us the old product form (1) but the
new boundary conditions lead to a series of cosine terms:
∞
X 2 n2 ³π n x´
−π t
u(x, t) = Bn e L2 cos .
n=0
L

Remarkably, the cosine modes are orthogonal under the same functional
dot product:
³π n x´ ³π n x´ Z L ³π n x´ ³π n x´
cos ·cos = cos cos dx = 0, m 6= n.
L L 0 L L
Therefore, we are led to the solution of the Neumann problem which
appears to be very similar to the previous solution of the Dirichlet’s
problem except that we now have cosines in place of sines:
Z
1 L
u(x, t) = u0 (x) dx
L 0
∞ ·Z L ³ π n x ´ ¸ π2 n2 ³π n x´
2 X − 2 t
+ u0 (x) cos dx e L cos .
L n=1 0 L L

In the solution of the Neumann problem we were forced to separate the

constant term of the series from higher harmonics because
1 · 1 = L,
whereas
³π n x´ ³π n x´ L
cos · cos = .
L L 2
Observe that the constant term
Z
1 L
u0 (x) dx
L 0
has the physical meaning of average initial temperature and it is also
the only term that survives as t → ∞. That makes sense because with
Neumann boundary conditions the rod becomes a closed system and
eventually the temperature at every point in the rod will tend to the
average temperature.
As an exercise, the reader can show that Robin’s problem leads to
the same development as before: the Fourier modes are orthogonal.
Thus orthogonality is not a fluke but is a consequence of the structure
of our IBVP as Theorem 1 demonstrates.
4 ALEKSEI BELTUKOV

Theorem 1 (Orthogonality). Consider the Heat Equation IBVP with

Robin’s boundary conditions:

∂u ∂ 2u
= (Heat Equation in 1 dimension)
∂t ∂x2
u(x, 0) = u0 (x) (Initial condition)
¯
∂u ¯¯
α u(x, t) + β = 0 (Robin boundary conditions)
∂x ¯x=0,L

Let gn (x) denote the solution of the associated ordinary BVP that arises
during separation of variables:
¯
d2 gn 2 dgn ¯¯
(2) + λn gn = 0, α gn (x) + β (x)¯ = 0.
dx2 dx x=0,L

We call the function gn an eigenfunction corresponding to the eigen-

value λn and tacitly assume that the eigenvalues λn form a discrete
sequence. Let gn and gm be two eigenfunctions corresponding to two
distinct eigenvalues. Then gn and gm are orthogonal with respect to the
usual dot product on [0, L]:
Z L
gn (x) · gm (x) = gn (x) gm (x) dx = 0.
0

Proof. Notice that the theorem asserts orthogonality of Fourier modes

without reference to any trigonometry: the eigenfunctions gn do not
need to be known explicitly. This suggests that we should work with
the BVP (??) rather than its trigonometric solution. In the light of
Equation (??), what we do know about gn and gm can be restated as
follows:

d2 gn
= −λ2n gn
dx2
d2 gm
2
= −λ2m gm
dx
Further, we know that both gn and gm satisfy symmetric Robin bound-
ary conditions. If we multiply the first ODE (for gn ) by gm and the
second by gm and subtract one from another, the results will be

d2 gm d2 gn
gn − g m = (λ2n − λ2m ) gn gm .
dx2 dx2
 LECTURE NOTES FOR MATH 157 5

Now integrate both sides from 0 to L to get

Z L· ¸
d2 gm d2 gn
gn (x) 2
(x) − gm (x) 2
(x) dx = (λ2n − λ2m )
0 dx dx
Z L
× gn (x) gm (x) dx = (λ2n − λ2m ) gn · gm .
0
Notice that we have essentially obtained an expression for the dot prod-
uct of gn · gm which we can now manipulate as follows:
Z L· ¸
d2 gm d2 gn
gn (x) (x) − gm (x) (x) dx
0 dx2 dx2
Z L · ¸
d dgm dgn
= gn (x) (x) − gm (x) (x) dx
0 dx dx dx
¯x=L
dgm dgn ¯¯
= gn (x) (x) − gm (x) (x)¯
dx dx x=0
¯ ¯ ¯x=L
¯ g (x) gm (x) ¯ ¯
= ¯¯ dgnn dgm
¯¯
¯¯
dx
(x) dx
(x) x=0
We now argue that the determinant vanishes at both endpoints due to
Robin’s boundary conditions: indeed, at the endpoints of [0, L] we can
substitute derivatives with scaled functions
¯
dgn,m α ¯
= − gn,m (x)¯¯
dx β x=0,L

and then the determinant vanishes (if β = 0, we have the already

investigated Dirichlet case.) We have thus shown that
(λ2n − λ2m ) gn · gm = 0, n 6= m.
Since λn 6= λm we must have
gn · gm = 0,
which shows that the eigenfunctions corresponding to distinct eigen-
values are orthogonal. ¤

Exercises
(1) Solve the Heat Equation with Robin’s boundary conditions:
¯
∂u ¯¯
u(x, t) + 2 ¯ = 0.
∂x x=0,L

Show every step and be sure to show orthogonality of the Fourier

modes: this will confirm Theorem 1.
6 ALEKSEI BELTUKOV

(2) Legendre polynomials form an orthogonal system on the inter-

val [−1, 1] under the dot product defined by
Z 1
f ·g = f (x) g(x) dx.
−1
The first two polynomials are given by
P0 (x) = 1, P1 (x) = x.
Find the quadratic P2 using orthogonality and the fact that
Pn (1) = 1 for each n.
(3) Consider the function y = cos(π x/2) defined on the interval
[−1, 1]. Approximate y using (i) Taylor polynomial of order 2
and (ii) Legendre polynomials of order 0, 1, 2. Plot the functions
and both approximations on the same plot. Which approxima-
tion is better and why?