ESS 522

2014
7. Tapering and Other Practical Considerations

In this lecture we discuss some practical aspects of using the discrete Fourier transform (DFT).
Amplitude and Power Spectra
We can write the Fourier transform in terms of its amplitude and phase
X( f ) = X( f ) exp i ( f )
(7-1)
where
|X (f)| is the amplitude spectrum
(f) is the phase spectrum
In some geophysical applications (e.g., electromagnetics) the phase is very important but in
many it is not and we just consider the amplitude spectrum
In engineering signal processing the power spectrum, |X(f)|2 ), is normally used.
Very often amplitude and power spectra are plotted with logarithmic scale. The decibel scale is
a logarithmic scale in which an order of magnitude change in power is 10 dB and an order of
magnitude change in amplitude is 20 dB. Amplitudes or powers in decibels are always
expressed relative to a reference.
Spectral Leakage and Windowing.
Note this section will be covered in Exercise 4, question 1. If you are confused by these notes
after the lecture, I would suggest doing the exercise and then trying to read them again.
Imagine that we obtain a discrete time series with a finite number of samples from a process that
is continuous in time (e.g., a year of temperature data or 10 seconds of data from a permanent
seismic installation). If we take a discrete time series with a finite number of samples and obtain
its DFT we are making the tacit assumption that the time series is periodic even though it is not.
Any discontinuity between the 1st and last time sample (i.e., a jump in the value or a change in
the gradient) will contribute to the frequency spectrum in a manner that is generally undesirable.
This effect is known as spectral leakage.
One can understand this effect by envisioning the discrete series as the result of truncating a
longer waveform by multiplying with a boxcar. This is equivalent to convolving the spectrum of
the untruncated time series with spectrum of a boxcar (i.e. a sinc function). The side lobes in the
sinc function leak frequencies from one frequency to another during the convolution in the
frequency domain (see Exercise 2, question 2v).
The solution to this problem is to multiply the finite sampled time series by a windowing
function or a taper that decreases smoothly to zero or near zero at its ends. There are literally
dozens of different choices of tapers (see the Matlab window function), many of which look
very similar, and some sophisticated tapering techniques (see multi-taper spectral analysis
method in Chapter 10 of Time Series Analysis and Inverse Theory by David Gubbins, Cambridge
University Press, 2004).
7-1

ESS 522

2014

In the exercise, we will consider one taper, the Hanning window or raised cosine window (the
word window is used synonymously with taper), which is defined for N points by
i

1 cos 2 N +1
wi =
(7-2)
, i = 1, 2,...N
2

(Note that the indexing works so as to exclude the zero points that would lie at either end of the
raised cosine)
The effects of applying a taper are:
1. To decrease the time series to zero or near zero at its start and end so that there is no sharp
discontinuity between the 1st and last point in the periodic time series.
2. To change the weighting of samples in the time series so that those near the middle contribute
more to the Fourier transform. Much of the theory of spectral analysis was developed by
engineers to look at stationary signals such as noise on an electrical circuit. For such signals one
can remove the bias resulting from applying the taper by normalizing the window so its rootmean squared amplitude is unity. That is
wi
(7-3)
winormalized =
1 N 2
w
N i =1 i
For non-stationary signals (the characteristics of many signals of interest in the geosciences such
as seismic waveforms change with time) tapering may bias the spectral amplitudes even if the
taper is normalized
3. To reduce the resolution of the spectrum by averaging adjacent samples. We can understand
this effect by looking at the amplitude spectrum of the windowing function and remembering
that multiplication in the time domain is convolution in the frequency domain. Ideally our
windowing function would have a very narrow central lobe (minimal averaging and loss or
resolution) but one of the tradeoffs in window design is that the narrower the central lobe, the
less well the window suppresses the effects of truncating the waveform. This is manifested by
side-lobes in the frequency domain that leak signal from other frequencies into each spectral
sample (spectral leakage). You will look at this effect in Exercise 4, question 1
Padding Time Series with Zeros
Because of the computational savings of the Fast Fourier Transform, it is most efficient to
compute the Fourier transform of a time series that has 2M samples. To accomplish this one can
simply pad the end of the time series with zeros (after applying a taper to the original time
series). Sometimes it is desirable for display purposes to add more zeros than is necessary to
increase the number of samples to the nearest power of 2. For example one might pad a 100
point time series with 412 zeros rather than 28.
7-2

ESS 522

2014

Adding zeros can introduce a spectral leakage effect since the gradient changes abruptly to zero
in the padded region you can minimize this by applying a window to the data before padding.
Padding with zeros decreases the spacing of frequency samples but it does NOT increase the
number of independent spectral values. Instead the values at additional frequencies are
interpolated and the resulting spectrum looks smoother.
Spectra of Long Time Series
Although the FFT and modern computers allow Fourier transforms to be computed for long time
series, one will eventually reach a computational threshold. Equally importantly, there is often
no need for the very fine frequency resolution that would accompany Fourier transform of a long
series. For instance, if one was looking at the spectral amplitude characteristics of a years worth
of seismic noise sampled at 100 Hz (something that I have done), there would be nearly 232
points and even if an FFT was computationally feasible for such a long series it is hard to
imagine why one would want frequency samples spaced at 3 x 10-8 Hz.
One option would be to smooth the output spectrum with a running mean (or better still a
Gaussian). Another practical approach is to divide long time series into shorter windows and
take the mean spectral amplitudes or energies (amplitudes squared) of all the windows. A taper
will generally be applied to each window to minimize spectral leakage. Since the tapering
functions tend down weight samples near the end of each window it is common practice to allow
an overlap of 50% between adjacent windows. The method of using overlapping windows to
calculate power spectra is commonly known as Welchs and is applied by the Matlab function
pwelch (psd in earlier versions).
Spectrograms
Sometimes we are interested in how the spectrum of a signal varies with time. This can be
visualized with a spectrogram which is a contour plot (normally a color contour plot) of the
spectral amplitude or logarithm of the spectral amplitude against time on the x-axis and
frequency on the y-axis. Each column of the spectrogram is constructed from a short windowed
segment of data centered on the time shown on the x-axis. In Matlab spectrograms can be
created and plotted with the command spectrogram.m

7-3

ESS 522

2014

Figure 7.1 Example of a spectrogram for 14 minutes of data showing earthquakes (vertical
regions of high amplitude) and tremor (high amplitudes around 1 Hz). Taken from the website of
the Iceland Meteorological Office at http://hraun.vedur.is/ja/englishweb/eruption_grimsvotn.html
Scaling of the DFT
The DFT can be used to approximate the continuous transform of a non-periodic signal of finite
duration, if it is scaled by the sample interval t. To show this we consider the continuous
Fourier transform
G( f ) =

g(t)exp ( i2 ft ) dt

(7-4)

If g(t) is only non-zero between 0 and T = Nt, we can convert it to a sum of the discrete time
series gt which yields:

( )

N 1

( ) (

G f = g(t)exp i2 ft dt g jt exp i2 fjt t

0

j=0

N 1
i2 kjt
= t g jt exp
=
N t

j=0

( )

N 1
i2 kj
= t g jt exp
= t DFT[g j ] = tGk
N

j=0

( )

7-4

(7-5)

ESS 522

2014

where f = k/Nt and we have assumed the version of the discrete Fourier transform used in
Matlab with the 1/N scaling in the inverse transform. That is
N 1
2 ikj
(7-6)
Gk = g j exp

N
j=0
Scaling by the sample interval normalizes spectra to the continuous Fourier transform. Even if
we are not interested in continuous Fourier transforms, this normalization allows us to directly
compare spectral amplitudes obtained from discrete time series with different sample rates.
Parsevals Theorem
Parsvals theorem states that the energy of a signal, that is the integral (or sum) of the square of
the signal, is equal to the integral (or sum) of the square of its Fourier transform Consider a
signal of infinite duration. For the Fourier transform to be convergent the energy of the time
series, E must be bounded
E=

g (t )

dt <

(7-7)

The inverse Fourier transform for a continuous series is

g ( t ) ==

G( f )exp (i2 ft ) df

(7-8)

If we substitute this for one g(t) term in equation (7-7) we get

(7-8)
E = g (t ) 2 dt = g (t ) G ( f ) exp(i 2 ft )df dt

By reversing the order of integration, we can write

2
E = g (t ) 2 dt = G ( f ) g (t ) exp(i 2 ft )dt df = G ( f )G * ( f )df = G ( f ) df (7-9)

where G* is the complex conjugate (Remember that for a complex variable we can write
G = a + ib , G* = a ib , and GG* = a 2 + b 2 ). We can think of g(t)2 as the distribution of
energy in the time domain and |G(f)|2 as the distribution of energy in the frequency domain.
Now for a discrete time series which we assume represents a discrete version of a continuous
time series that is zero outside the limits of the discrete time series, we can write the time domain
representation of the energy as expressed in equation (7-7) as
T

N 1

j =0

E = g (t ) 2 dt g ( j t ) 2 t

(7-9)

The amplitude of time samples can be considered to be in units of energy / unit time interval .
In the frequency domain we get a discrete version of equation (7-9)
E=

N 1

N 1

G ( f ) df = G (k f ) f = t 2 GDiscrete,k f
2

k =0

k =0

where the t2 in the right hand term adjusts the scaling of the discrete Fourier transform as
required by equation (7-5).
7-5

(7-10)

ESS 522

2014

The units for the continuous Fourier transform, G(f) are

discrete Fourier transform they are
is the total spectral width.

energy /unit of frequency but for the

energy /total spectral width of the DFT since the term 1/t

Absolute Spectral Amplitudes

In many applications we have no interest in the absolute values of the spectral amplitudes - We
are either interested in relative values (which frequencies dominate a spectra) or we just want to
manipulate a signal in the frequency domain (e.g, convolve or filter it) before returning to the
time domain. However, if you are making use of the absolute spectral values you need to be
careful to:
1. Scale the tapers used to minimize spectral leakage so that their root mean squared amplitude is
unity. Otherwise the tapers will rescale the data.
2. Scale the discrete amplitude spectra with t (equation 7-3) and power spectra with t2 so that
they have the same units as their continuous spectra.
3. Understand how the apparatus that recorded the digital data responds to different frequencies
(i.e., know its impulse response) and apply the appropriate correction so that your spectra are in
the units of physical interest.

7-6