1 Fourier Series: 1.1 The Complex Exponential Form
1 Fourier Series: 1.1 The Complex Exponential Form
ENERGY 281
Spring Quarter 2007-08
Lecture 5 Notes
These notes are based on Rosalind Archer’s PE281 lecture notes, with some
revisions by Jim Lambers.
1 Fourier Series
Recall that in Lecture 2, when we considered PDE on bounded spatial do-
mains, we expressed solutions in terms of a Fourier sine series, in the case of
Dirichlet boundary conditions, or Fourier cosine series, in the case of Neu-
mann boundary conditions. Now, suppose that a solution u(x, t) of a PDE
on the interval [0, L] is intended to satisfy periodic boundary conditions
u(0, t) = u(L, t), t > 0. (1)
Then, since sin x and cos x are both 2π-periodic, the solution can be
expressed in terms of both sines and cosines, as follows:
∞
X 2πjx 2πjx
u(x, t) = a0 (t) + aj (t) cos + bj (t) sin , (2)
j=1
L L
1
where
1 L Z
û(ω, t) = √ e−2πiωx/L u(x, t) dx. (7)
L 0
Clearly, this series resembles the formula for the inverse Fourier transform,
while the formula for the coefficients resembles that of the Fourier transform.
1.2 Orthogonality
Like the sines and cosines in (2), the functions e2πiωx/L are orthogonal with
respect to the inner product
Z L
(u, v) = u(x)v(x) dx, (8)
0
where the complex conjugation of u(x) is necessary to ensure that the norm
k · k defined by q
kuk = (u, u) (9)
satisfies one of the essential properties of norms, that the norm of a function
must be nonnegative.
Specifically, we have
(
2πiωx/L 2πiηx/L
L ω=η
e ,e = . (10)
0 ω=
6 η
√
This explains the presence of the scaling constant 1/ L in (6). It normalizes
the functions e2πiωx/L so that they form an orthonormal set, meaning that
they are orthogonal to one another, and have unit norm.
One consequence of this orthonormality is an analogue of Parsevals iden-
tity for Fourier series. We say that f (x) is square-integrable on [0, L] if
Z L
|f (x)|2 dx < ∞. (11)
0
That is, the above integral must be finite. If such a function is also piecewise
continuous, the following identity is satisfied:
∞
|fˆ(ω)|2 = kf k2 ,
X
(12)
ω=−∞
2
1.3 Convergence
The Fourier series for an L-periodic function f (x) will converge to f (x) at
any point in [0, L] at which f is continuously differentiable. If f has a jump
discontinuity at a point c, then the series will converge to 12 [f (c+ ) + f (c− )],
where
f (c+ ) = lim f (x), f (c− ) = lim f (x). (13)
x→c+ x→c−
where each f˜(ω) approximates the corresponding coefficient fˆ(ω) of the true
Fourier series. Ideally, this approximate series should satisfy
3
are obtained by approximating the integrals that defined the coefficients of
the the Fourier series:
1 NX−1
f˜(ω) = √ e−2πiωxj /L f (xj )∆x, ω = −N/2 + 1, . . . , N/2. (18)
L j=0
Because the functions e2πiωx/L are orthogonal with respect to the discrete
inner product
N
X −1
(u, v)N = ∆x u(xj )v(xj ), (19)
j=0
4
3 De-Noising and Aliasing
Suppose we have 128 data points sampled from the following function over
[0, 2π]:
f (x) = sin(10x) + noise. (24)
The function f (x), shown in Figure 1(a), is quite noisy. However, by taking
the discrete Fourier transform, (Figure 1(b)) we can extract the original sine
wave quite easily. The Fourier transform shows two distinct spikes, one at
n = 11 and one at n = 119. These correspond to frequencies of ±10 i.e. the
frequency of the original sine wave. The first N/2 + 1 values of the Fourier
transform correspond to frequencies of 0 ≤ f ≤ fmax , where fmax = N/2.
The remaining N/2 − 1 values of the Fourier transform correspond to the
frequencies −fmax < f < 0.
1.2
1
1
0.5
0.8
|Tf(ω)|
y
0
0.6
−0.5 0.4
−1 0.2
0 2 4 6 −50 0 50
x n
Figure 1: (a) Left plot: noisy signal (b) Right plot: discrete Fourier trans-
form
Note that the DFT only considers a finite range of frequencies. If there
are frequencies beyond this present in the true Fourier transform, an effect
known as aliasing occurs. The effect of aliasing is shown in Figure 2: it
“folds” these frequencies back into the computed Fourier transform. Specif-
ically,
∞
f˜(ω) = fˆ(ω + ℓN ),
X
−N/2 + 1 ≤ ω ≤ N/2. (25)
ℓ=−∞
5
Aliasing can be avoided by filtering the function before the DFT is applied,
to prevent high-frequency components from “contaminating” the coefficients
of the DFT.
− 1 0 1 f
2∆ 2∆
∆x NX−1
f˜(ω) = √ e−2πijω/N f (xj )
L j=0
N/2−1 N/2−1
∆x X −2πiω(2j)/N ∆x X −2πiω(2j+1)/N
= √ e f (x2j ) + √ e f (x2j+1 )
L j=0 L j=0
N/2−1
∆x X −2πiωj/(N/2)
= √ e f (x2j ) + (26)
L j=0
6
N/2−1
∆x
√ Wω
X
e−2πiωj/(N/2) f (x2j+1 ) (27)
L j=0
where
W = e−2πi/N . (28)
It follows that
1 1
f˜(ω) = f˜e (ω) + W ω f˜o (ω), ω = −N/2 + 1, . . . , N/2, (29)
2 2
where f˜e (ω) is the DFT of f obtained from its values at the even-numbered
points of the N -point grid on which f is defined, and f˜o (ω) is the DFT of
f obtained from its values at the the odd-numbered points. Because the
coefficients of a DFT of length N are N -periodic, in view of the identity
e2πi = 1, evaluation of f˜e and f˜o at ω between −N/2 + 1 and N/2 is valid,
even though they are transforms of length N/2 instead of N .
This reduction to half-size transforms can be performed recursively; i.e.
a transform of length N/2 can be written as the sum of two transforms of
length N/4, etc. Because only O(N ) operations are needed to construct a
transform of length N from two transforms of length N/2, the entire process
requires only O(N log2 N ) operations.