1.

Introduction

We already know that if a function is ”very nice”, then one can express it as a power
series; for example

∞ ∞
xk (−x)n
ex =
X X
∀x; log(1 + x) = − ∀|x| < 1.
k=0 k! n=1 n

Indeed, in general, if f is analytic at a, then

∞
(x − a)k
f (k) (a)
X
f (x) = for all x sufficiently closed to a.
k=0 k!

Unfortunately, most functions are not that nice (indeed we will be talking about piece-
wise smooth or piecewise continuous functions.) The basic idea of Joseph Fourier
is to represent such function by an infinite sum of trigonometric functions.
For simplicity, first suppose f is a function on [0, 2π]. We hope to express it as

∞
X ∞
X
ck sin(kx − τk ) = a0 + (ak cos kx + bk sin kx).
k=0 k=1

The above is called a trigonometric series.

Basic questions:
1. What for?
2. When is it possible and in what sense?
3. How?
4. Basic properties of such expression (for the moment, it can be call a ’formal trigono-
metric series’.

Heat equation (cf. § 36/§ 31)

Let us start with a (initial) boundary value problem:

 2
 ∂u = ut = α2 uxx = α2 ∂∂xu2 , 0 < x < 1, t > 0;
∂t
(1)
 u(0, t) = u(1, t) = 0 for all t ≥ 0,

u(x, 0) = f (x), 0 < x < 1. (2)

1
2

How do we solve the equation? It seems very complicated. The following is probably
due to Joseph Fourier.
Let’s start by guessing: what if u(x, t) = T (t)X(x) (where T is only a function of t and
X is a function of x)? Of course this is usually not possible even for simple function such
as (x + t)2 , but it will be a linear combination of such forms. Anyway, do we know how
to solve this simple case?
Note that then we have a very simple form as follows:

ut = T 0 (t)X(x) and uxx = T (t)X 00 (x)

and hence
T 0 (t) X 00 (x)
= .
α2 T (t) X(x)

Notice that LHS depends only on t while RHS depends only on x. For that to hap-
pen, they must both equal to a constant. Thus, there exists a real number λ such that

T 0 (t)
=λ (3)
α2 T (t)

X 00 (x)
= λ. (4)
X(x)
We will now solve for (3) and (4). Note that they are just ordinary linear differential
equations. We now recall from (1) that we must have X(0) = X(1) = 0.

First note that for λ > 0, (4) has no nontrivial solution (i.e., a solution that is not
identically zero). This is because its general solution is of the form (we will come back to
this for more details later)
√ √
X(x) = C1 e λx
+ C2 e− λx

and hence X(0) = X(1) = 0 =⇒ C1 = C2 = 0.

Similarly, if λ = 0, then the solution of (1) is X(x) = 0.

 3

Finally, if λ = −β 2 , β > 0, then

X(x) = C1 cos βx + C2 sin βx

and X(0) = X(1) = 0 =⇒ C1 = C2 = 0 whenever β 6= kπ for all k ∈ N. However, if

β = kπ, we still have C1 = 0, but we have infinitely many solutions.

X(x) = C2 Xk (x), C2 ∈ R where Xk (x) = sin kπx.

2
For each k ∈ N, we have T 0 = −(kπ)2 α2 T and hence T (t) = ce−(kαπ) t .
2 2
Let us write Tk (t) = e−(kαπ) t . Then for each k ∈ N, Tk (t)Xk (x) = e−(kαπ) t sin kπx is a
solution to (1).

Let us note that if u1 (x, t) and u2 (x, t) are solutions of (1) then c1 u1 + c2 u2 is also a
solution of (1).

Thus if we can find {ck } such that

∞
X
ck sin kπx = f (x)
k=1

then the function

∞
2t
ck sin kπxe−(kαπ)
X

k=1
satisfies (1) and (2). We did not claim that this is the only solution that satisfies (1) and
(2). But we did find one such solution provided we are able to write
∞
X
f (x) = ck sin kπx
k=1

such that

(a)

(b)
4

Note that they hold if for example when RHS converges uniformly (but they could
hold under much weaker conditions). However, as we did not assume that you have taken
MA3110/MA3210 and indeed Fourier did not care about these issues, we may just assume
both (a) and (b) hold most of the time in this module. Of course, if there are enough
interest among students, I will go through it in the module. Anyway, we will come back
here with more detail.

Simple Examples
(1) f (x) = sin 2πx + 7 sin 20πx

(2) f (x) = cos(2πx) sin(6πx)

(3) f (x) = sin(2πx) sin(3πx)

sin((N + 1/2)πx) − sin(πx/2)

(4) f (x) = .
sin(πx/2)

(5) x(x − 1)?

 5

(6) Consider the following (initial) boundary value problem (wave equation §37, 45, )

utt = uxx , 0 < x < π, t > 0; u(0, t) = u(π, t) = 0 for all t ≥ 0, (5)

u(x, 0) = f (x), ut (x, 0) = g(x), 0 < x < π. (6)

Again, we will do it quickly here for you to have a quick overview and come back
later with more detail.

(7) A simple boundary value problem in ODE:

y 00 + y = sin3 x, 0 < x < π, y(0) = y(π) = 0.

6

2. Fourier Series (cf. Chapter 1)

Similar to power series expansion, we would like to express a function in terms of sum
of trigonometric functions. Here is the most common example:
∞
X
f (x) = a0 + (ak cos kx + bk sin kx)
k=1

which is a (formal) trigonometric series or also called Fourier series. Obviously, we cannot
hope that the representation is pointwise (infinite sum!) for all functions. In particular
as RHS is periodic of period 2π (if it converges), we may conclude that f must also be
periodic. However, in general, we do not need pointwise equality in practice. Indeed, we
only need a representation that is able to do jobs that we need to do.
Let us now try to find out some types of functions that have such representation.

Recall that for a function to have a power series expansion, the function has to be an-
alytic (at the reference point 0) and hence infinitely differentiable. However, necessary
conditions for a function (indeed possible for a generalized function) to have a Fourier
series is much weaker. We are not going to discuss the collection of all functions that
have a Fourier series, instead, we will just concentrate on functions that are not “too
badly behaved”. The following (piecewise defined) functions are the basic element in this
module.

Piecewise continuous functions (cf. §1)

First of all recall that a function f is said to be continuous at x0 if

lim f (x) = lim− f (x) = f (x0 ).

x→x+
0 x→x0

Also, recall that f (x− +

0 ) = lim− f (x) and f (x0 ) = lim+ f (x) are the left hand limit and
x→x0 x→x0
right hand limit at x0 respectively.
Examples
 7

Definition 2.1. A function f is said to be piecewise continuous on [a, b] if it has a

partition a = x0 < x1 < x2 < · · · < xn = b such that f is uniformly continuous on each
interval (xi−1 , xi ) for i = 1, · · · , n.

Remark 2.2. This is not the definition used in your textbook but the two definitions
are equivalent. Indeed, it is a consequence of uniform extension theorem (for continuous
functions, refer to your note on continuous functions and uniform continuous function in
MA2108). In particular, we have
(1) both f (a+ ), f (b− ) exist, and
−
(2) f (x+
i ), f (xi ) exist for all i = 1, · · · , n − 1.

Moreover, f needs not be defined at any xi , consequently the domain of f may be a

subset of [a, b] that contains [a, b] \ {x0 , · · · , xn }.

Remark:
(1) Continuous functions on [a,b] are piecewise continuous on [a,b]. However, contin-
uous functions on (a, b) need not be piecewise continuous on [a,b].
(2) Similar to continuous functions, it is easy to verify that product and linear sum
of piecewise continuous functions are again piecewise continuous.
(3) For those who know the definition of Riemann integral may want to verity that
piecewise continuous functions on [a, b] are Riemann integrable.

More examples:

[Exercise] Are they piecewise continuous on [−π, π]?

√
1. x sin(1/x)
1
2. cos
x
8

3. the greatest integer function of 1/x

√
4. the greatest integer function of 20 x
5. tan(x/2)

Similarly, we can define piecewise smooth functions (to be exact, it should be called
piecewise continuously differentiable functions):
Definition A function f is said to be piecewise smooth (cf. p 27) on [a, b] if both f 0
and f are piecewise continuous on [a, b].

[Example]

Remark 2.3. Again, a piecewise smooth function needs not be differentiable at all points.
Moreover, the following consequence of L’Hopital rule may be useful to us.
f (x) − f (a+ )
Fact: if lim+ f 0 (x) = l, then lim+ = l. Similar fact holds for left limit.
x→a x→a x−a

[Exercise] Are they piecewise smooth?

(1) x2 cos(1/x) on [0,1]
(2) the greatest integer function of 4x2 on [0,3].
(3) the greatest integer function of 1/ cos x on [0, π].
(4) product and linear sum of piecewise smooth functions.
 9

Integration and Integration by parts

1. When f is a piecewise continuous function on [a, b], we often evaluate the integral
of f as follows:

Z b n Z xj
X n Z xj
X
f (x)dx = f (x)dx = gj (x)dx
a j=1 xj−1 j=1 xj−1
where x0 = a < x1 < · · · < xn = b is such that f is uniformly continuous on each
(xj−1 , xj ), and gj (x) on [xj−1 , xj ] is the continuous extension of f on (xj−1 , xj ); this is
always possible by uniform extension theorem of continuous function (cf. MA2108).

2. When P is a polynomial of degree less than m and f is continuous, the following

computation is often useful and standard: (cf. p11)

Z
P f dx = P F1 − P 0 F2 + P 00 F3 − · · · + · · · + (−1)m P (m) Fm+1 + C
where P (j) is the j derivative of P and F1 is an antiderivative of f and in general Fj+1 is
an antiderivative of Fj .

3. If P is a polynomial (or any other nice function) and f is only piecewise continuous,
then we will do step 1 first before step 2. Note that in general, it may not be true that
Z b Z b
P 0 f dx = f (b)P (b) − f (a)P (a) − P f 0 dx (∗)
a a

if f is only piecewise smooth.

Example
Z π
(x2 + 2x) sin xdx =
0

Exercise: Can you find the correct formula for (*)? Assuming f is uniformly contin-
uous on (a, c) and (c, b).
10

Periodic functions We know that trigonometric functions are periodic. What does
that mean?
A function f (x) on R is said to be periodic if there exists a > 0 such that

f (x + a) = f (x) for all x ∈ R.

We said f is periodic with period α, if α is (usually) the smallest positive number (if that
exists) such that the above mentioned identity holds with a = α.

Examples cos x, sin x are periodic functions with period 2π. sin 2πx + cos πx is a
periodic function with period 2.

Remark 2.4. (1) If f is a periodic function with period τ , then for any k ∈ Z, we
have f (x + kτ ) = f (x) for all x ∈ R.
(2) If f is a function on (0, a], a > 0, then it can be extended to a periodic function
on R with period a or a/k, k ∈ N.

Computation of Fourier series Z π

First let us recall that if m, n ∈ Z, then eimx e−inx dx = 2πδ(m − n) where δ(n) = 0
−π
for n 6= 0 and δ(0) = 1 (δ is sometimes called the ’Delta function’).
 11

Hence, we have

Z π
eimx + e−imx inx
e dx = πδ(m − n), if m, n ∈ N ∪ {0}.
−π 2
Since eimx = cos mx + i sin mx (De Moivre formula), we have the following orthogonal
property (m, n ∈ N ∪ {0}):

Z π Z π
cos(nx) cos(mx)dx = real part of eimx cos nxdx = πδ(m − n).
−π −π
The above is equal to 2π if m = n = 0. Of course, one could also use trigonometric
identity for the above computation. Similarly,
Z π
cos(nx) sin(mx)dx = 0 (even if n = m)
−π
Z π
sin(nx) sin(mx)dx = πδ(m − n).
−π
The idea of finding Fourier series expansion is indeed very simple. For illustration, let
us first assume it is possible to write
∞
X
f (x) = bk sin kx (which is usually known as Fourier sine series on [0, π]).
k=1

Now note that

Z π
1Z π
sin nx sin mxdx = sin nx sin mxdx = δ(n − m)π/2.
0 2 −π
As we are going to integrate, let us assume f is piecewise continuous on [0, π]. Thus,
multiplying the above expression of sine series by sin nx and integrate on [0, π], we have
Z π ∞
Z π X
f (x) sin nxdx = ( bk sin kx) sin nxdx.
0 0 k=1

Now if we can interchange the infinite sum with the integral, we have
∞ Z π
X
RHS = bk sin kx sin nxdx = bn π/2.
k=1 0

So we conclude that if there is a Fourier expansion as above, we should have

2Zπ
bn = f (x) sin nxdx for all n.
π 0

Indeed, the above computation is valid for finite sums and thus Fourier probably think
that it should be also true for infinite sum. After Fourier, many mathematicians have
provided sufficient condition in order to interchange the infinite sum and the integral
12

(note that it is always possible to do that if it is only a finite sum since integration is
known to be ’linear’). Riemann saw that it is possible if one has uniform convergence of
the infinite sum. However, in general, it is possible (to interchange) under much weaker
condition instead of uniform convergence (unfortunately, it requires Lebesgue’s work in
measure and integral MA4262).

Our computation above only conclude that if f (x) has a Fourier sine series, then bk
must be of the above mentioned form. We have yet to prove that the Fourier sine series
is a meaningful expansion. Indeed, most of the work about whether the expansion is
meaningful were not done by Fourier. For completeness of this subject, we will justify it
later in the course.
 13

Let us make some quick observations and remarks.

1. The above Fourier sine series is an odd function and of period 2π and it is zero at
both the origin and π.
2. There is unfortunately no reason why the infinite sum converges. Indeed, there is an
example where f is continuous and periodic but its Fourier series diverges at a point (and
hence one can produce a periodic continuous functions with Fourier series with arbitrary
finite number of points of divergence). However, we know that Fourier series expansion
is meaningful when f is piecewise smooth (and we will prove it). Moreover, we will also
discuss ’Cesaro’s mean to reproduce the pointwise value for piecewise continuous function.
3. Since in general one may not get pointwise equality, to reduce confusion, we some
time replace 0 =0 by 0 ∼0 for the trigonometric representation.

Some examples
∞
X (−1)k
(1) x ∼ −2 sin kx
k=1 k
∞
4X sin(2k − 1)x
(2) 1 ∼ .
π k=1 2k − 1
∞
−(−1)k π 2 2(−1)k − 1
!
2 2X
(3) x ∼ + sin kx.
π k=1 k k3

Indeed, (to be proved later)

∞
4X sin(2k − 1)x
1= for x ∈ (0, π)
π k=1 2k − 1

and
14

Fourier Cosine series

We could also try expressing a function f on [0, π] as a Fourier cosine series:
∞
X ∞
X
f (x) = ak cos kx = a0 + ak cos kx.
k=0 k=1

This form is slightly different from your textbook (cf. §2 p4).

Carrying out the computation similar to Fourier sine series, we have

2Zπ 1Zπ
ak = f (x) cos kxdx, k ∈ N and a0 = f (x)dx.
π 0 π 0

Once again, you may want to compute the following yourselves.

∞
π 4X cos(2k − 1)x
x∼ − .
2 π k=1 (2k − 1)2
∞
π2 (−1)k
x2 ∼
X
+4 cos kx.
3 k=1 k2
1 2 X (−1)k
χ[0,π/2] ∼ − cos(2k − 1)x.
2 π k=1 2k − 1

Here are some questions for you.

P∞
Question 1. Why don’t we talk about expansion k=1 ak cos kx, that is, to start the
sum from k = 1 and not zero.

P∞
Question 2. How about b0 + k=1 bk sin kx?
 15

Question 3. While Fourier sine series gives us an odd function, what does Fourier
cosine series give us?

For any function f on [−π, π], one could always write it as a sum of an odd function
and an even function. Indeed, note that the function fE (x) = (f (x) + f (−x))/2 is even
and fO (x) = (f (x) − f (−x))/2 is odd and f (x) = fE (x) + fO (x) for all x. Note that f is
piecewise continuous on [−π, π] if and only if both fE and fO are piecewise continuous.
We could then expand fE as a Fourier cosine series and fO as a Fourier sine series and
hence we have a Fourier series:

∞
X ∞
X
f (x) = a0 + ak cos kx + bk sin kx.
k=1 k=1
(Note that the RHS of the above equality is a function of period 2π if it converges.)
We could also compute directly just as before and observe that for any n ∈ N

Z π ∞
X ∞
X
(a0 + ak cos kx + bk sin kx) sin nxdx
−π k=1 k=1
Z π ∞
X Z π ∞
X Z π
= a0 sin nxdx + ak cos kx sin nxdx + bk sin kx sin nxdx
−π k=1 −π k=1 −π
Z π
2
= bn sin nxdx = bn π
−π

(provided we areZable to interchange the above integral with infinite sum).

1 π
Hence bn = f (x) sin nxdx.
π −π Z
1 π
Similarly, an = f (x) cos nxdx, if n ∈ N.
π −π
16

Finally, we have
Z π Z π ∞
X ∞
X Z π
f (x)dx = (a0 + ak cos kx + bk sin kx)dx = a0 dx = 2πa0 .
−π −π k=1 k=1 −π

1 Zπ
Hence a0 = f (x)dx.
2π −π

Remark 2.5. (1) After checking through the above, we do not see why there is a need for
f be periodic. All we need to compute an and bn is the fact that f is piecewise continuous
(indeed, it is enough to assume that f is Riemann integrable). The question is why should
the representation be meaningful.
(2) Note that if f (x) = f (−x), then bk are zero for all k (sin kx are odd functions).
Similarly, if f (x) = −f (−x), then ak are zero for all k.

More examples

(1) f (x) = 3 sin2 x + cos x.

(2) f (x) = sin 2x sin 3x

(3) f (x) = x sin x.

 17

In general, a periodic function may have period other than 2π. What if the period of
a function is not 2π? For example the function sin 2πx or cos πx + sin 3πx. Then it is
natural for us to modify our representation accordingly. In general if f has period L, then

∞ ∞
X 2kπx X 2kπx
f (x) = a0 + ak cos + bk sin .
k=1 L k=1 L
And we can compute the Fourier series just as before:

2 Z L/2 2πnx
an = f (x) cos dx
L −L/2 L
2 Z L/2 2πnx
bn = f (x) sin dx
L −L/2 L
1 Z L/2
a0 = f (x)dx.
L −L/2

Remark (1) Once again, we see that it is sufficient to assume the integral makes sense
in order to compute an , bn . Is there any reason that f must be periodic?
(2) Instead of doing the computation similarly, it is also possible to do it by change of
variables to the Fourier series that we have already obtained (see §8). Indeed, change of
variable can be a useful tool in cutting down steps of computation.
18

More examples: Compute Fourier series for the following

(1) f (x) = x on [0, π].

(2) f (x) = x on [0, 2π].

(3) f (x) = x on [−π, π].

(4) f (x) = x(1 − x) on [0, 1].

(5) ebx on [−1, 1].

[Exercise]
1. f (x) = |x| on [-1,1].

2. f (x) = x sin x on [−π, π].

 19

3. f (x) = π − x on [0, π].

4. ebx on [0, π].

Relationship of Fourier series, Fourier sine series and Fourier cosine series
Let f be a function on [0, π] such that f (0) = f (π) = 0. Then we can do any of the
following.
(1) Extend f to a function on R of period π.
(2) Extend f to an odd function on R of period 2π.
(3) Extend f to an even function on R of period 2π.
20

If we do (1), we can find a Fourier series expansion.

If we do (2), we can still get a Fourier series expansion. However, note that it is actually
just a Fourier sine series as there will be no terms of cos kx in the expansion.
Finally, if we do (3), we will get a Fourier cosine series.

[Exercise] Find the Fourier series, Fourier cosine series and Fourier sine series for each
of the following functions:
(1) x sin x on [0, π];
(2) x on [0, π];
(3) x on [0,1];
(4) 1 + 2x on

[0,1].
 −1, if 0 < x < 1
2
(5) f (x) =  1
.
1, if 2 < x < 1