PHYS 332: Junior Physics Laboratory II

Notes on Fourier Transforms

The Fourier transform is a generalization of the Fourier series representation of
functions. The Fourier series is limited to periodic functions, while the Fourier transform can
be used for a larger class of functions which are not necessarily periodic. Since the transform
is essential to the understanding of several exercises, we briefly explain some basic Fourier
transform concepts.
A. Fourier series
Recall the definition of the Fourier series expansion for a periodic function u(t). If
u(t + T) = u(t), then
#

u(t) =

$ c(!

) exp(i! m t)

(1)

1 T /2
# u(t) exp("i! m t) dt
T "T / 2

(2)

m="#

where !m = 2"m/T and

c(! m ) =

If u(t) is a function of time it is natural to call !m an angular frequency and to refer to c(!m)
as the amplitude of u(t) at frequency !m. Together, the c(!m) comprise the frequency
spectrum of u(t).
Computing the frequency spectrum for a sine or square wave of period T is fairly
simple. For the sine wave, there are only two non-zero coefficients
c(! m ) =

1
2

m = 1

(3)

as one might expect. Note, though, the odd fact that the formalism requires both positive and
negative frequencies. The spectrum of the square wave is slightly more complicated
c(! m ) =

2
"m

m odd

(4)

and again contains both positive and negative frequencies. The interesting feature here is that
the rapid rise and fall of the square wave leads to spectral components at high frequency, as
demonstrated in Fig. 1. This is a general effect, and can help one understand the qualitative
connection between time and frequency descriptions.
S3
S1
S4
S2

Fig. 1 Building up a square wave from Fourier series components. Each graph is a plot of Sn,
the sum of the first n terms in the series expansion, for increasing values of n.
B. Fourier transforms
For many purposes a knowledge of the c(!m) may be more useful than knowing u(t)
itself. It is therefore desirable to extend the idea of a frequency spectrum to non-periodic
functions, for example single pulses or finite wave trains. This can be done by replacing the
sum of Eq. 1 by an integral, giving

u(t) =

1
2!

+$

% c(" ) exp(i "t) d"

#$

(5)

and
+#

c(! ) =

$ u(t) exp("i!t) dt

(6)

"#

where ! is now a continuous variable. The complex function c(!) is called the Fourier
transform of u(t) and is again the amplitude of u(t) at frequency !.
Some examples may make the ideas clearer. The transform of exp(i!0t) is a delta
function, !(! - !0). This says that the spectrum of exp(i!0t) has a large amplitude only at !
= !0, which seems reasonable since the function is periodic with angular frequency !0.
The Fourier transform of a pulse centered at t = 0,
#% = 1
u(t) $
%& = 0

t ! " /2

(7)

t > " /2

is simply

c(! ) =

sin(!" /2)
(! /2)

(8)

Note that a short pulse (small !) leads to a significant c(") over a wide frequency range,
while a long pulse (large !) implies c(") decreases quickly with frequency. As with the
Fourier series results, we might have expected this outcome because a rapidly changing
signal must be made up of high frequencies.
The Fourier transform has many other amusing mathematical properties, but only one
more is of concern here: The Fourier transform of c(") is u(-t) (except for a normalization
factor). Therefore, if we transform a function and then transform the result we will almost
recover the original function.
The function c(") contains both amplitude and phase information needed to reconstruct
the original signal. We are usually interested only in intensity information, not phase, so we can
employ the magnitude-squared of the transform

P(! ) = c(! )

(9)

where P(") is called the power spectral density of the signal, with units of volts2/Hz. For a realvalued signal, the only case of physical interest, P(") is an even function of ", so all the physical
information is contained in the positive frequency domain. Integrating P(") over all frequencies,
we recover the mean-square voltage
"
2

(10)

V = 2 P(! ) d!
0

further reinforcing the connection to intensity.

More physically, the power part of the name comes from the proportionality to voltage
squared, and it is a spectral density because it is the voltage squared per unit frequency interval.
Effectively, P(") tells us how much power our signal has at a specified frequency.
So far we have considered functions of one variable, which we thought of as a time,
but there is nothing to prevent a generalization to functions of several variables. It turns out
to be useful to consider the case of two spatial variables, say (x,y). Formally,

u(x, y) =

1
2!

$ $

% % c (" ," ) exp[i2! ( x"

x

+ y" y ) d" x d" y

#$ #$

# #

c (! x ,! y ) =

$ $ u(x, y) exp["i2% ( x!

"# "#

+ y! y ) dxdy

(11)

(12)

where !x and !y are called spatial frequencies, with units of (length)-1. Just as ordinary
frequency measures how fast a signal changes in time, a spatial frequency measures how
rapidly a signal varies in space. Since the formalism is the same, all of our previous
observations about the properties of the transforms hold for two-variable functions.
C. Sampled data considerations
So far we have tacitly assumed that the function u(t) is known at all values of t, but it
is more common to sample the input with an analog to digital converter at regular intervals !t
so that the required computations can be done digitally. Also, the signal can only be sampled
for a finite time period, say from -T/2 to T/2. When transforming such sampled data, we
compute a discrete approximation to the ideal Fourier transform. The discrete transform of an
N-point array un is defined to be
N !1

$
nm '
c m = " un exp&!i2#
)
%
N (
n= 0

(13)

for m = 0,1,...,N-1. For a real input function un, c*N-m = cm, so we obtain N/2 distinct points
in the frequency spectrum, uniformly spaced between f = 0 and the Nyquist frequency fN =
1/(2!t). The spacing between points, or the frequency resolution, is just 1/N!t = 1/T, where T
is the total length of the input record. Except as noted below, the discrete transform is, for our
purposes, equivalent to the continuous Fourier transform. A very efficient algorithm, called
the Fast Fourier Transform or FFT, exists to compute the sum in Eq. 13, provided that N is
exactly a power of two.

Fig. 2 Sampling of sinusoidal waves. Note that there is no way to distinguish between the
two sine waves on the basis of the regularly-spaced samples.
The first limitation of the sampling process is the possibility of aliasing, as illustrated in
Fig. 2. Evidently input frequencies from zero to 1/(2!t) are all distinguishable, but if higher
frequencies are present they will be confused with some lower frequency. An example is shown
in Fig. 3, where the higher harmonics of the square wave signal are aliased to lower frequencies,
producing a misleading spectrum. The Nyquist frequency, fN = 1/(2!t), is the highest frequency
that can be measured with a sample spacing of !t. Experimentally, it is important to avoid
aliasing by inserting an analog low-pass filter in the input before sampling and then choosing !t
small enough that the Nyquist frequency is well above the pass limit of the analog filter.
A more subtle problem arises from the finite length of the sampled record. In general, the
signal will not be exactly zero at the start and end of the record, so it appears to change very

10

10

10

Power

Power

10

10

-5

100

200
300
400
Frequency (Hz)

500

10

-5

100

200
300
400
Frequency (Hz)

500

Fig. 3 Power spectra of square wave signals with a fundamental of 157 Hz. (Left) Without
filtering, there are numerous spurious peaks due to aliasing. (Right) With proper low-pass
filtering, only the first and third harmonic appear below the Nyquist frequency.
5

rapidly when cut off at the ends. This is equivalent to introducing spurious high frequencies into
the signal, which may then be aliased to frequencies below fN and appear in the spectrum.
Depending on the information that is desired, it may be possible to improve the measurement by
smoothly forcing the data to zero at the ends of the time interval, rather than simply truncating.
This is done by multiplying the input data by a window function before performing the
transform. For example, the Hann window for an N point sample

w(n) = 0.5 ! 0.5cos(2"n /N)

(14)

is often used for general purpose analysis. The effect of a Hann window on the filtered square
wave signal is shown in Fig. 4. Other window functions can be designed to optimize the
resolution for closely spaced frequencies or to obtain accurate amplitude estimates.

10

10

Power

Power

10

10

-5

100

200
300
Frequency (Hz)

400

500

10

10

-5

100

200
300
Frequency (Hz)

400

500

Fig. 4 Power spectrum of a square-wave signal (Left) without windowing and (Right) with a
Hann window applied.
6