A Comprehensive Windows Tutorial

Howard A. Gaberson, Oxnard, California

This article is a tutorial attempt to provide an easier analy-

sis of how windows work. I begin by looking at individual
spectrum bins as affected by off-bin-centered tones with six
different windows. I define and use the convolution theorem
to explain why the windows do what they do. I review the so-
phisticated misunderstood fat-center-peaked, picket-fence-
looking FFTs of various windows that come from adding ze-
ros to increase resolution. I then use these drawings in a
step-by-step graphical convolution to show how this or that
window will affect the FFT of your data. This is a procedure
to demonstrate how any oddly shaped, even graphically de-
fined, window affects data. These techniques are tools to al-
low you to invent a graphical window to affect data in new
ways. To demonstrate how to use graphically defined win-
dows, I use digitized drawings of a Hawaiian mountain and
the popular cartoon beagle as windows. To use these windows,
I demonstrate FFT-IFFT interpolation and resampling, which
in itself is interesting. Finally, I compare how eight windows
affect the spectrum of typical vibration data.

We need a window for vibration spectrum analysis because Figure 1. This shows a 1,024-point Hanning window applied to 1,024-
point chunk of air handler acceleration.
the beginning does not match the end of the data segment we
are analyzing and because we virtually never have an integer
number of periods of any cyclic information in the signal seg-
ment/chunk. The discrete Fourier transform (DFT) analysis is
really a Fourier series expansion of whatever segment we in-
put. It is considered periodic. Mismatched ends and noninteger
numbers of periods garbles the result. Windows somewhat al-
leviate that garbling but in a complicated way.
Windows are shapes – like a hill, a hump in the center, and
they go to zero at the ends. When you multiply these window
shapes term by term with a segment of data, you force the ends
of the data segment to zero, so at least the ends match the be-
ginning. You cannot really fix the noninteger number of peri-
ods. Figure 1 shows a window application. In Figure 1a we have
a 1,024 value plotted list of acceleration data; in Figure 1b, a
1024 value Hanning window plotted for the same time values.
Figure 1c shows the product of the two. The beginning and ends
certainly match.
I am sure there are more than 100 windows described in the
literature. Figure 2, shows six of them that I happen to use and
will illustrate here: boxcar (or uniform), Hanning, Hamming, Figure 2. Six common window shapes.
Gaussian, Kaiser Bessel, and flat-top (this is very different from
the boxcar or uniform window). Signal analyzers frequently lyzed, which almost never happens without planning. We can
offer five or six to choose from, with variably helpful discus- develop a formula for testing if a sine wave signal component
sions of when to use each. Hanning is the most widely recom- is bin centered as follows. A sine wave is given by Equation 1:
mended. Reference 1 gives 44 and Reference 2 gives eight more;
y = sin 2p ft (1)
these might be the most widely referenced windows papers, but
these authors hadn’t yet heard of the flat-top. Ron Potter 3 gives where f is the frequency in cycles/second (Hz). The period of
about five windows and is famous for an unpublished HP pub- a sine wave P (seconds per cycle) is the reciprocal of the fre-
lication. Potter did have a flat-top in there. Finally, the vener- quency (Equation 2):
able Brüel & Kjær Technical Review published the first fully 1
disclosed flat-top window and a nice version of the Kaiser P= (2)
f
Bessel window.4,5 These two publications are downloadable
from the B&K web site. I read a paper 6 from the 2004 Interna- Since we have to work with digitized data, define:
tional Modal Analysis Conference where the investigator was N = number of samples in data segment being analyzed
working on keeping the side lobes extremely low because of fs = sample rate in samples/second
all the new high-resolution, 24-bit, data-collecting hardware. The time interval between samples h is the reciprocal of the
If you really want to dig into things, Harris7 wrote a difficult sampling rate fs (Equation 3):
64-page article with excellent drawings that I recommend. 1
h= (3)
fs
Defining the Bin-Centered Concept
Non-bin centered or not, an integer number of periods or The signal duration T is given by Equation 4:
cycles in the analyzed data segment is the topic to start with. N
A sine wave component of our signal is bin centered if it has T = Nh = (4)
fs
an integer number of periods in the data segment being ana-

14 SOUND AND VIBRATION/MARCH 2006

Figure 3. Non-bin-centered tone time histories; except for the case of Figure 4. Expanded DFTs for 16 Hz; this is the bin-centered case, six
exactly 16 Hz, the beginnings and ends do not match. different windows.

The number of periods N p has to be given by the duration over

the period (Equation 5):
T
Np = (5)
P
By substituting Equations 2 and 4 into Equation 5, we find:
Nf T N / fs
Np = from = (6)
fs P 1/ f
Now, as long as Np is a whole number, the sine wave in the
data segment is bin centered or has an integer number of peri-
ods. Such test segments of a sine wave signal are needed for
calibrating a window coefficient in our spectrum calculation
programs, so this is a handy formula to keep in mind. In the
case of a sampling rate of 1,024/second and if our number of
samples is also 1,024, the number of periods will be equal to
the frequency. I will set things up that way here. To illustrate
how the end and beginning not matching affects things, I will
draw an expanded view of the end of a 16-Hz, 1,024-point sine
wave chunk connected to the beginning of the identical 16-Hz,
1,024-point sine wave chunk. This is shown for 16 Hz and sev-
eral frequencies close to 16 Hz in Figure 3. I have expanded Figure 5. Expanded DFTs, 16.1 Hz 10% off bin center, six different win-
the region where the end of the first sine meets the beginning dows; all windows seem reasonable in this situation, but the boxcar
of the second sine to make it clear that a discontinuity occurs window shows side lobe leakage.
when the frequency is not a whole number. Notice that the two
curves match perfectly (end to beginning) only for the case of
16 Hz, which makes the signal bin centered.

Non-Bin-Centered Effects
Now we are in an excellent position to consider the effects
that those six windows I mentioned help us alleviate (but can-
not really cure) the picket fence and leakage effects of a non-
bin-centered signal on the spectra we compute. In case you are
rusty, I want to review some spectrum fundamentals so you will
know what to expect. I use the acronyms DFT and FFT syn-
onymously. DFT means the discrete Fourier transform, and FFT
means the fast Fourier transform, the calculation procedure
used to calculate the DFT. All signal analyzers and data col-
lectors use an FFT to initially compute and then develope the
spectrum from that result.
The DFT of an N-point signal, sampled at fs gives us a spec-
trum of (N/2 + 1) amplitudes at frequencies from 0 to fs/2. We
frequently call these (N/2 + 1) amplitudes the bin values. The
Df or bin frequency spacing, is 1/T = fs/N. We get a frequency
value starting at f = 0, and values spaced at ∆ f all the way up
to fs/2. There will be N/2 intervals between 0 and fs/2; there-
fore, each interval or space is ∆ f = (fs/2)/(N/2) = fs/N. This is Figure 6. Expanded DFTs, 16.2 Hz, 20% off bin center, six different win-
1/T. The time between samples is 1/fs, and (N)(1/f s) = T. These dows; thin line shows actual tone positioned at 16.2 Hz.

INSTRUMENTATION REFERENCE ISSUE 15

 facts are important; you have to remember them. It is trouble
to prove, but if you want to see it proved, ask me for a copy of
Reference 8.
Now, let us check out all six windows on bin-centered and
the other Figure 3 cases with varying degrees of not being bin
centered. We start with a 1,024-sample signal of a 16-Hz sine
sampled at 1,024/sec. From Equation 6, this will have exactly
16 periods. With 1,024 samples sampled at 1,024/sec, the sig-
nal duration T equals 1 sec. On the spectrum, Df = 1/T = 1/(1
sec) = 1 Hz. If we stay with 1,024 samples and the same 1,024
sampling rate, the Df = 1 Hz stays the same no matter what we
do to the frequency to change the number of periods and ex-
amine non-bin-centered effects.
Figures 4 through 9 show an expanded view of the spectra
(around 16 Hz), with the six windows applied to the degrees
of non-bin centeredness of Figure 3. In these spectra, I have
used M ATLAB ’s ® “stem-plot” feature to emphasize the indi-
vidual bin values. Figure 4, shows the six windows applied to
the spectrum calculation when the frequency is exactly 16 Hz,
the bin-centered case. The boxcar or “no window” shows that
we have a single value of one in the spectrum in the 16-Hz bin.
All of the other five windows show content in the bins next to
Figure 7. Expanded DFTs, 16.3 Hz, 30% off bin center, six different win- 16 Hz. In this case, they are indicating content where we hap-
dows; boxcar is close to 20% in error now.
pen to know there actually is none. There is some stuff to dis-
like about windows, but as we move on, you will come to agree
that we have to use them. Both the Hanning and the Hamming
windows indicate content at 15 and 17 Hz as well. The Kaiser
Bessel window shows content in five bins, from 14 to 18 Hz.
The Gauss window shows content from 13 to 19 Hz, and the
flat-top window shows content from 12 to 20 Hz. Recall that
the signal only contains an amplitude of 1 at 16 Hz.
Figure 5 shows what happens if we raise the frequency to
16.1 Hz and thus have 16.1 periods, 10% off being bin centered.
The bin content shifts; the boxcar windowed spectrum now
shows content from 13 to 17 Hz and shows the amplitude at
16 Hz down a little. All the other windows seem to say we still
have an amplitude of 1 and a sine wave at 16 Hz. In Figure 6,
I have shown the spectra for a 16.2 Hz sine wave with 16.2
periods, 20% off bin centered. I have shown the 16.2-Hz am-
plitude as a black line. The boxcar spectrum is indicating 0.92
amplitude content at 16 Hz, with additional content spread
from probably 10 to 23 Hz. This indication of content where
there is none is called leakage. Notice that all the other win-
dows except the flat-top are indicating a content less than 1
still at 16 Hz. The flat-top is still hitting 1 right on the money
at 16 Hz, and even the 17 Hz bin is pretty close to 1.
Figure 8. Expanded DFTs, 16.5 Hz, 50% off bin center, six different win- Figure 7, shows the situation with all six windows for 16.3
dows; this is the worst case, but note the flat-top window still gets the Hz, 30% off bin centered, 16.3 periods. Now the poor boxcar
amplitude correct.
spectrum thinks we have a tone of 0.85 at 16 Hz, with content
in all the adjacent bins shown from 8 to 24 Hz. The fact that
the boxcar spectrum shows a 15% error is called the picket-
fence effect. Notice that all the other windows do much better
than the boxcar, with the flat-top showing we have content
between 16 and 17 Hz with an amplitude of 1. The Hanning
seems to beat the Hamming, and the Kaiser and the Gauss are
considerably more accurate.
The worst case is Figure 8, the spectra for 16.5 Hz or 16.5
periods. The Hanning spectrum indicates equal content at 16
and 17 Hz of about 0.84, a 16% error. That is how this picket-
fence error goes; the boxcar spectrum indicates roughly equal
content at 16 and 17 Hz of about 0.63, or a 37% picket-fence
error. Except for the flat-top, which indicates an amplitude of
1 content at 16 and 17 Hz, the Gauss window shows maybe
0.96, and the Kaiser Bessel is close behind at about 0.95. No-
tice also that the boxcar spectrum shows content in every bin,
and the other windows restrict this spreading of content. This
effect is called reducing leakage; this is a leakage of content
into adjacent bins where we know there is no content in the
true signal. Picket fence and leakage reduction: that is why we
need windows. We cannot eliminate the effects, but windows
Figure 9. Expanded DFTs, 16.7 Hz, 70% off bin center, six different win- reduce them. Figure 9 shows the same information about 16.7
dows; this is really 30% off 17 Hz. Hz, or 30% off 17 Hz.

16 SOUND AND VIBRATION/MARCH 2006

More DFT Background:
We have to look a little closer at the DFT and how it is used
to compute our spectrum. We have to look at the DFT of a win-
dow, and I have to explain this shifted idea. The DFT equations
are Equations 7a and 7b:
n = N -1
1
Xk = Â x ne - i 2p kn / N (7a)
N n =0
N -1
xn = Â X k e i2p nk / N (7b)
k =0

Equation 7a is the DFT analysis equation; it analyzes N

samples of digitized data, the xn into N DFT values, the Xk . The
NX k values are complex sine wave amplitude values; we say
they are complex because they contain an amplitude and a
phase or the same thing is real and imaginary parts. Equation
7b is the synthesis equation; it exactly transforms the complex
X’s back to the original x’s. However, I have the 1/N where I
think it belongs, not where it is usually placed.
To use Equations 7a and 7b for vibration analysis, there are Figure 10. The DFT concept: X (thick) is a transform value; it has a
some ground rules. The data list is sampled at sampling rate length and is at an angle, its phase. It is the starting point of the com-
fs. The xn of Equation 1 are samples from a signal that was ac- plex sine wave spiral. This represents one of the discrete complex si-
curately sampled and band limited to fs/2; this means its Fou- nusoids in 3-D. The straight-line vector from the origin represents, Xk.
The little circles on the spiral beginning at the tip of the straight line
rier transform is zero for all frequencies greater than half the are the values for the sequence of n values.
sampling rate. There is an even number of N samples in the
list. In Equations 7a and 7b, the complex exponential is also real value; it is special. It is the content at the Nyquist fre-
exactly Equation 8: quency or half the sampling rate. It turns out to be the sum of
2p kn 2p kn the sequence with alternate signs reversed. In between these
e - i 2p kn / N = cos - i sin (8) two are (N/2 – 1) unique complex values, each containing two
N N
values. The X values from X N/2+1 to X N – 1 are not unique but
We can see the frequency in that 2πkn/N, if we multiply it by are complex conjugates of the values from X N/2 – 1 down to X 1.
hfs which equals 1, by Equation 3, and rearrange as in Expres- The values symmetrical about X N/2 are complex conjugate
sion 8a: pairs. Thus for the N values of the sequence, we get N unique
values from the transform.
Ê kf ˆ
cos Á 2p s nh˜ is analagous to cos 2p ft (8a)
Ë N ¯
Relation of DFT X’s to Harmonic Content
Thus, in digitized terms, nh is the discrete time, and kf s/N, is The spectrum of a machinery vibration signal is a plot of
the frequency. The time index is n and the frequency index is content vs frequency. If you find time to look up Fourier se-
k. ries theory from any old book (25 years or so) you will find that
Let us repeat: the DFT as an exact transformation of a digi- they use a’s and b’s to do the analysis of a segment of a signal,
tized vibration signal. It transforms the data into discrete (or assumed periodic. (However, the analysis will work almost no
sampled) complex exponentials (or equivalent sinusoids). The matter what you assume; it will make a periodic function of
transform is the list of their amplitudes as a function of fre- that segment.) Then they reconstruct the signal in terms of these
quency. Equation 7a transforms the signal into N sine waves; a’s and b’s as follows. They say if x(t) is periodic with period
each X k is the amplitude and phase of a complex sine wave. T, it can be exactly synthesized from its Fourier series coeffi-
Figure 10 attempts to show one of the k discrete complex si- cients, (ak , b k ), as Equation 9:
nusoids in 3-D. X k is its complex amplitude or a vector from •
the origin to the beginning point of the discrete spiral. The little ao Ê 2p kt 2p kt ˆ
x (t ) =
2
 a cos
ËÁ k T
+ bk sin
T ¯
˜ (9)
circles on the spiral represent the values of the discrete sinu- k =1
soid for the sequence of n values. The smooth spiral on which for k = 0, 1, 2, . . . ,
the data lie is the underlying curve, the curve with time taken
to be continuous or with nh replaced by t in Equation 8a. Each The Fourier series coefficients are given by Equations 9a and
X k is the complex amplitude of one of the discrete spirals. 9b:
Equation 7b says to add them all up and you have your origi- T
ak 1 2p kt
(9a)
2 T oÚ
nal signal. The transform is the list of complex X’s. = x (t )cos dt
T
Since it is an exact transformation, we are able to exactly
inverse transform the DFT back to the original data. The trans- T
bk 1 2p kt
form is a set of amplitudes and phases of complex sine waves (9b)
2 T oÚ
= x (t )sin dt
that form a continuous curve when added together. When the T
curve is evaluated at the signal sampling instants, it exactly The ak and bk are the harmonic content. The kth harmonic is
reproduces the signal. The continuous curve, which is the sum given by Equation 10:
of the N sine waves, is the continuous periodic band-limited
2p kt 2p kt
curve the original x’s were sampled from. The inverse trans- x k (t ) = ak cos + bk sin (10)
form evaluates all these sine waves at the signal sample instants T T
and adds them up. I want to also use the word reconstruction Its frequency is k/T. We can also write the kth harmonic in
for the inverse transform operation, the IFFT. terms of amplitude and phase as follows:
The DFT is most economically computed using the FFT. I Ê 2p kt ˆ (10a)
believe M ATLAB’s algorithm is able to deliver efficient results x k (t ) = Ak cos Á - fk ˜
Ë T ¯
for all N values, not just powers of 2. The FFT will compute N
Fourier coefficients from a sequence of N numbers. The X’s are Here A k is the amplitude and φ k is the phase. The phase is the
numbered from 0 to N – 1. If N is even, X0 is the DC or average angle in radians to the first positive peak. The amplitude and
value; it is real and contains one value. X N/2 also contains one phase are given by Equations 10b and 10c:

INSTRUMENTATION REFERENCE ISSUE 17

 Ak = a2k + bk2 (10b)
Êb ˆ
f k = tan -1 Á k ˜
Ë ak ¯ (10c)

A k is the content or the amplitude of the kth harmonic or tone;

φ k is its phase. This is the quantity shown on signal analyzers
and data collectors. More manipulating leads us to the follow-
ing. The Xk ’s from k = 1 . . . (N – 1)/2 that we get from the DFT
are related to the content as Equations 11, 11a, and 11b:
ak b
Xk = -i k (11)
2 2

Ak = 2 X k (11a)

Ê Re( X k ) ˆ
j k = tan -1 Á (11b)
Ë - Im( X k ) ˜¯

The amplitude and phase of the kth harmonic defined in Equa-

tion 10a are summarized in Equations 12 and 12a:
Figure 11. DFT of a 1,024-point boxcar window; only one line, DC. F s =
1024, so T = 1 sec; Df = 1 Hz.
Ak = 2 X k , Ao = X 0 , AN / 2 = X N / 2, (12)

Ê Re( X k ) ˆ
j k = tan -1 Á (12a)
Ë - Im( X k ) ˜¯
The DFT, which is most economically computed using the
FFT, will compute N Fourier coefficients from a sequence of N
numbers. X0 is the DC or average value; it is real and contains
one value. X N/2 also contains one real value. In between these
two are (N/2 – 1) unique complex values, each containing two
values. The X values from XN/2 + 1 to XN – 1 are not unique but
are complex conjugates of the values from XN/2 – 1 down to X1 .
The values symmetrical about X N/2 are complex conjugate
pairs. Thus for the N values of the sequence, we get N unique
values from the transform.
That is dry and tough. One more DFT concept and then we
proceed, and it is ‘shifted.’ The values past XN/2 can be picked
up as a group and lifted to the other side of zero so that now
instead of having values on either side of the Nyquist value,
X N/2 , be complex conjugate pairs, the values on either side of
zero frequency or X 0 are complex conjugate pairs. In this
method of plotting a DFT we will have positive and negative Figure 12. DFT 1,024 point Hanning window; only three lines; Fs = 1024,
frequencies. This is a common method of plotting a DFT. We So T = 1 sec; Df = 1 Hz.
are used to plotting a spectrum, which is content versus fre-
quency from 0 to the Nyquist frequency, or fs/2. Many people 1Ê 2p n ˆ
w= Á 1 - cos ˜ (13)
not in this business plot the X k’s and use positive and nega- 2Ë N ¯
tive frequencies. That is what I mean by shifted. This is a raised cosine with an average or DC value of 1/2. It
has a period of our window length, 1 second, thus its Df is 1
DFT of a Window Hz. The DFT is going to be a digitized Fourier series of this 1
Now let us look at the DFT magnitude of some windows; the Hz raised cosine with an average value of 1/2 considered peri-
DFT or the X k’s, not the spectrum. I am going to make all these odic and going on forever. Figure 12a shows its time history.
windows 1,024 points long, and I am going to sample them at In Figure 12b, we see its 1,024 value DFT. We see what appears
1,024 samples per second. Therefore, the time duration of each to be a 1 at 0 frequency and a 1/2 at the 1024th value. In Fig-
window will be one second; the Df, which is one over the du- ure 12c, we see the results of shifting the 513th to the 1,024th
ration, will be 1 Hz. We will start with a boxcar window. The value over on the other side of 0. In Figure 12d we expand the
DFT is going to be digitized Fourier series of a boxcar window plot to only slow values from –10 to 10 Hz, which shows a tri-
considered periodic and going on forever. It is a straight line angle with a base from –2 to +2 Hz and an apex of 1 at 0 fre-
with an amplitude of 1. It only has a DC component; it should quency. We get this triangle, because the plotter draws straight
have one line in its DFT with amplitude 1 at frequency equal lines between the values. In Figure 12e, I use the stem plot
to 0. Figure 11 shows this. In Figure 11a, we see the 1,024 ones feature that makes it clear we only have three values; one at 0
as a horizontal line. In Figure 11b, we have the DFT magnitude frequency with an amplitude of 1, and two values of 1/2 at –1
and we see a 1 at 0 frequency. In Figure 11c, I show the shifted and +1 Hz. By Equation 12, these two values of 1/2 would add
DFT magnitude with positive and negative frequencies. We still to 1 when we convert this to a spectrum so the spectrum will
have a 1 at 0 frequency. In Figure 11d, I expand the plot to only have a 0 frequency or DC value of 1 and a 1 Hz value of 1. They
slow values from –10 to 10 Hz, which shows a triangle with a both should be 1/2, but I normalized the DFT plot to make the
base from –1 to 1 Hz and an apex of 1 at 0 frequency; we get largest value 1. The Hanning window has a conveniently
this triangle because the plotter draws straight lines between simple DFT; three numbers. We will soon discuss that the con-
the values. In Figure 11e, I use the stem plot feature that makes volution can be used to apply the window, and a convolution
it clear we only have one value at zero frequency with an am- with just these three numbers made significant economic sense
plitude of 1. in the late 1960s and early ‘70s when the new signal analyzers
Now let us look at the DFT magnitude of a Hanning window. became digital and started using the FFT. Most of the popular
The equation of the Hanning window is Equation 13: windows have simple DFTs for this reason.

18 SOUND AND VIBRATION/MARCH 2006

Effect of Adding Lots of Zeros
This may seem silly for a minute, but just for a minute. The
DFT of a digitized time history returns N + 1 frequency values
when it receives N time history values equally spaced between
zero and fs. I will append nine times 1,024 zeros to the end of
a 1,024 point window time history. Now this forces the DFT to
analyze a periodic time history with the window, a nine times
as long string of zeros, and the window again, and so on; the
DFT will now calculate 10,241 frequency values. For the box-
car or uniform window, this is very different from our periodic
boxcar that was just a horizontal line. Figure 13a, shows a time
history for a 10,240-point boxcar window in which the first
1,024 values are ones. (You have to append a lot of zeros to see
the window shape.) The time history is 10 seconds long. The
sample rate is 1024 sample/second, so the first 1,024 points
have a duration of 1 second. Figure 13b is a view of just the
first 1,050 points of the time history to show that we have the
same boxcar window. Figure 13c shows the magnitude of the
DFT of the zero-padded boxcar 10-second time history. Now
the plot appears to show the value 1 at 0 frequency and a 1 at
the highest frequency. The shifted DFT magnitude in Figure
13d shows content near 0 frequency. To see what is going on Figure 13. 1,024 point boxcar window, 9¥ zero-padded, expanded, DFT,
near 0, I expanded the plot to just show the region from -5 to shifted, expanded, semi-log plot.
+5 Hz in Figure 13e. This is the amazing result of adding the
zeros. Instead of getting the single line of Figure 11e, we now
have a continuous curve with a main lobe 2 Hz wide between
an array of side lobes 1 Hz wide. In Figure 11f, I have plotted
a semi-log plot showing the log of the DFT values. Quite a com-
plicated result. This is the kind of DFT I will use to show what
the windows do.
I will pursue adding zeros in examples with the Hanning
window shown in Figure 14. Our Hanning window in Figure
12 was a cosine with a amplitude of 1/2 and a mean value of
1/2 oscillating forever. In Figure 14a, we see the Hanning win-
dow with nine times as many zero values appended. Figure 14c
shows the DFT magnitude and Figure 14d the shifted DFT
magnitude, again with all of the content apparently clustered
near 0 frequency. But reducing the frequency range from –5 to
+5 Hz shows a main lobe 4 Hz wide with a slight ripple appar-
ent. This ripple becomes apparent on the semi-log plot of Fig-
ure 14f. The main lobe is fatter, 4 Hz wide, and the ripple
magnitude is considerably reduced; the highest lobe is about
a 0.02. These side lobes are again 1 Hz wide.
Finally, in Figure 15, I did the same thing for the flat-top
window to show the extremely wide main lobe, which we shall Figure 14. 1,024 point Hanning window, 9¥ zero-padded DFT, shifted,
see enables the flat-top window to provide extreme amplitude expanded, and semi-log plot.
accuracy. The reason it is called the flat-top is because the main
lobe has a flat top. The side lobes are so small they do not show
in this plot. Now we move on to the convolution theorem where
we will use these ideas.

Convolution Theorem
The only way I can analyze or explain window effects (that
is, why the window affects leakage, the content in the adjacent
bin, and reduces picket fence errors) is via the convolution
theorem,9 which is difficult. The convolution theorem for us
is: the DFT of the term-by-term product of two digitized time
signals (a digitized time signal is a long list of digitized accel-
eration or velocity values) is equal to the convolution of the
DFTs of each of the signals. If we want the DFT of the product
of two time signals, we can get it by a convolution of the two
individual DFTs. And that is our situation exactly. We want to
understand the DFT of a windowed signal, and we will do it
by looking at the convolution of the DFT of unwindowed sig-
nal and the DFT of the window. So we have to use the convo-
lution.
What is a convolution? It is horribly confusing, important,
Figure 15. 1,024 point flat-top window, DFT, shifted, expanded, semi-
and widely used. It drives most students nuts and still confuses log plot.
me every now and then. The equation is:
q = N -1 For our situation, Y(q) represents the DFT of our un-
X (k ) = Â Y (q )W ( k - q ) (14) windowed data as collected, and W is the DFT of the window.
q =0 X is the DFT of the windowed data that we can convert to the

INSTRUMENTATION REFERENCE ISSUE 19

spectrum we see in the signal analyzer. The equation says (and
realize that X, Y and W are lists of N numbers, like 1,024 num-
bers) that the value of the windowed DFT at frequency k is the
sum of the products of the flipped, shifted, window DFT val-
ues centered at frequency k and the DFT of the data. We are
going to do this in the explanation and I will draw it shortly,
so I think this will become clear. But remember: sum of prod-
ucts of flipped, shifted, window DFT, and data DFT.

Applying Convolution Theorem to Windows Analysis

How does the convolution theorem help us understand win-
dows? We will view the window as a long digitized time sig-
nal, a string of mostly zeros with a non-zero digitized window
shape in the center. Thus, the process of term-by-term multi-
plying the long window with the long signal conceptually
forms our windowed segment that our analyzer/data collector
FFTs. I need this concept to explain what we saw in Figures 4
through 9. Figures 16 and 17 illustrate this for the boxcar and
Hanning windows.
In Figure 16, by speaking of a widowed segment, I mean
considering the term-by-term product of the upper plot (pre-
tend it is a long vibration signal, say 50,000 points) and the
middle plot is an equally long list of zeros with a set of ones Figure 16. Boxcar window to select a segment of data from long time
history.
(perhaps 1,024) in the region where we want to select a seg-
ment of the signal for analysis. By multiplying these two graphs
or data lists term by term, we end of with the third plot, a seg-
ment of the long signal. The third plot represents a chunk of
data our signal analyzer would analyze. By speaking of a wid-
owed segment, for now, I mean the product of the upper plot
(the signal), and the middle plot of the long list of zeros with a
set of ones in the middle. Similarly, Figure 17 shows the same
concept, except this time we have a centered Hanning window
flanked by a huge number of zeros. Now you can see why I
needed the DFT of the window with the huge amount of zeros.
We know that the DFT of a signal returns spectrum values at
a set of frequencies ∆f apart. The ∆ f is 1/T, where T is the time
duration of the signal segment. If we have a signal that is stable
and infinitely long (let fe mean forever) Tfe is infinitely long,
N fe is infinite, ∆ ffe is infinitesimally small; its DFT has infi-
nitely many lines and is exact for every tone or harmonic con-
tent the signal contains. This is an idea of the true spectrum
concept: no picket fence or frequency error and no leakage. It
has lines everywhere and is exact for every tone in that stable
signal. That is the true spectrum we want, but we cannot have
it because we cannot collect or analyze such a long signal.
We use the infinitely long zeros-window-zeros of Figures 16b
and 17b to select a chunk T long of that stable, very long time Figure 17. Hanning window to select a segment of data from long time
history. The DFT of this windowed chunk will still have lines history.
everywhere, because it had infinitely many time values. So it
will also have lines at the frequencies that our signal analyzer
can compute. To explain what the window will do to various
kinds of signals analyzed on any real signal analyzer, we will
convolve the highly detailed zero-padded window DFT with
the DFT of a sine wave signal. We know the DFT of our T length
signal analyzer spectrum is going to be a discrete vector with
values for frequencies starting at 0, and at every 1/T, on up to
N/2.
And what does this ‘convolved’ mean? This means the value
of the convolution at frequency k. Or: flip the window DFT left
right, center it on one k value, term by term multiply, and add
up the products. Summarizing: take the DFT of the window,
which is symmetric if plotted with positive and negative fre-
quencies – center it at some frequency f1 on the DFT of the sig-
nal – term-by-term multiply the two DFTs together; add up the
products; and that is value of the convolution for f1. Continue
this for every frequency of interest.
Let us illustrate this with the examples of Figure 7, the case
of a 30% of bin-centered signal. Figure 18 illustrates this for
the boxcar window. The true signal is a tone at 16.3 Hz shown
as a black line. In Figure 18a, the window DFT (blue) is cen-
tered at 13 Hz. The product of the DFT of the window with the
DFT of the 16.3 Hz tone is short red line at 13 Hz. In Figure Figure 18. Convolution steps of boxcar window with 16.3 Hz tone.

20 SOUND AND VIBRATION/MARCH 2006

Figure 19. Convolution steps of Hanning window with 16.3-Hz tone; this
shows why the several lines appear.
Figure 22. After editing, Diamond Head was 696 samples; it was
resampled to 1,024.

18b, the window is centered at 14 Hz. The product of the DFT

of the window with the DFT of the 16.3 Hz tone is the red line
at 14 Hz. This red line height is equal to the height of the in-
tersection of the blue window DFT with the black tone. In Fig-
ure 18c, the window is centered at 15 Hz. The six positions of
the window DFT result in the six red lines shown on Figure
18f. This is the same result I obtained by taking the boxcar
windowed FFT and shown in the stem plot of Figure 7a. The
same example is illustrated for the Hanning and the flat-top
windows in Figures 19 and 20. That is how the window works
and why we saw the results of Figures 4 through 9.

Main Lobe Significance

My next observation is also difficult. The convolution of the
window DFT with the data DFT is the same as the convolution
of the data DFT with the window DFT. By looking at Figures
19 and 20, notice that if we had a set of vertical grid lines at
12, 13, 14, 15, 16, 17, 18, 19, 20 Hz (which are the bin-centered
frequencies, k/T) and we centered the window DFT on the 16.3
Hz tone, the intersection of the window DFT with the grid lines
Figure 20. Convolution steps of flat-top window with 16.3-Hz tone;
shows line build-up and accuracy due to the wide flat-top of the win-
would give us the same result that we came up with in the
dow DFT. above paragraph by shifting the window to do the convolution.
This is illustrated in Figure 21. The shape of the main lobe of
the zero-padded window DFT sets the lines or bin values
caused by any tone. The equation of the main lobe is well
known for the standard windows. If you believe that a group
of lines is caused by a tone, you can use the formulas from the
appendix in Reference 12 to estimate the true frequency and
amplitude. Especially if there is not too much going on in the
little side lobes.

Graphically Defined Windows

Since we have discussed how to estimate window effects, I
want to present a procedure for inventing a window defined
only graphically. I chose two shapes to illustrate the idea. One
is Diamond Head, a well-known mountain crater on the Hawai-
ian island of Oahu, and the other is well known beagle comic
character sleeping on his doghouse. In both cases, I scanned
the images into the computer and then digitized them with digi-
tizing software.11 The digitizing software does not allow you
to select the sampling rate or the number of samples, so these
windows must be resampled. The Diamond Head window was
originally sampled to 696 samples; this was resampled to 1024
samples by adding zeros to the DFT and inverse (IFFT) trans-
Figure 21. The green curve is the Hanning window zero-padded DFT forming it back to the time domain. 10 Figure 22 gives a plot of
centered on the 16.3-Hz tone. Where it crosses the bin center frequen-
cies (11, 12, 13, 14, 15, 16, 17, 18, 19, 20, etc.) are the height of the
the original digitizing and the resampled version. Figure 23
lines it will induce in the DFT. These red lines are the same as indi- shows the zero-padded version, its DFT, and both linear and
cated in Figure 19. semi-log plots of its expanded central region. We are adding

INSTRUMENTATION REFERENCE ISSUE 21

Figure 23. Diamond Head window, 9¥ zero-padded, DFT, shifted, lin- Figure 24. The digitized and edited, as well as resampled beagle win-
ear, semi-log. dow.

only high-frequency zeros and an even number for economi-

cal computing. The current Nyquist or center value is the high-
est frequency value and is real. We make the new Nyquist 0.
Half of the old Nyquist is placed on each side as highest fre-
quency non-zero values. (Add a complex zero.) Then place an
odd number of complex zeros in the transform center for the
desired values in reconstruction. Because of the 1/N in M ATLAB,
the reconstructed window must be renormalized to make its
maximum value 1. Adding zeros is a sensible practical proce-
dure and certainly worked well here.
In the case of the beagle window, I digitized the drawing with
170 samples, and this was resampled by adding zeros to the
DFT, and inverse transforming to bring the time plot to 1,024
samples. Figure 24 shows a plot of both the original and
resampled versions of the window, and Figure 25 shows the
zero-padded version, its DFT, and both linear and semi log plots
of its expanded central region.

Side Lobe Height

From our discussion of Figures 18, 19, and 20, we can see
the problems of side lobe height. The windowed spectrum level
at any of the frequencies will be the sum of the products of the
window DFT and the data DFT. I illustrated the easy problem Figure 25. Beagle window, 9¥ zero-padded, DFT, shifted DFT, linear,
of the convolution with a single tone. Windows with high DFT semi-log plots.
side lobes or skirts allow noise and other tones or content to
contaminate or increase the apparent level for the line at which
the window is centered. Figures 26 and 27 show composite
linear and semi-log plots of the side lobes of the six windows.

Effects on Machinery Vibration Data

Finally, I made two busy plots, Figures 28, and 29, showing
the eight windows applied to some pump vibration data. They
provide an opportunity to examine spectral changes expected
from different windows. The running speed is close to 20 Hz.
There is a strong 2¥ component near 40 Hz, probably due to
misalignment. The content at 7¥ is because the pump has seven
vanes. The high skirt characteristic of the boxcar window
shows well on Figure 29, the semi-log plot. The wide flat-top
window smears the peaks and obscures detail. It is interesting
that the beagle and the Diamond Head widows do as well as
they do.

Conclusion
I hope some of this adds some insight to your understand-
ing of windows. Maybe it will give you enough background to
tackle some of the difficult references. It certainly seems to me
from References 1, 2, and 9, that many bright people have de- Figure 26. DFTs of common windows, shifted, expanded, zero-padded,
voted years to developing windows. All of the figures and cal- linear plot.

22 SOUND AND VIBRATION/MARCH 2006

Figure 27. Semi-log plot of shifted, zero-padded, expanded windows
showing skirts. Figure 29. Eight semi-log spectra from the bearing of a large pump, each
analyzed with the indicated window. High skirts of the boxcar window
show well around 40 Hz. The windows have an effect to be sure.

References
1. Harris, F. J.. “On the use of Windows for Harmonic Analysis with
the Discrete Fourier Transform,” Proc. IEEE, Vol. 66, No. 1, pp. 51-
83, January 1978.
2. Nuttall, A. H., “Some Windows with Very Good Sidelobe Behavior,”
IEEE Trans. on Acoustics, Speech, and Signal Processing, Vol.
ASSP-29, No. 1, pp. 84-91, February 1981.
3. Potter, R. W., “Compilation of Time Windows and Line Shapes for
Fourier Analysis,” Handout notes from an HP seminar circa 1978
4. Gade, S., and Herlufsen, H., “Use of Weighting Functions in DFT/
FFT Analysis (Part I),” Brüel Kjær Technical Review, No. 3, pp. 1-
28, 1987.
5. Gade, S., and Herlufsen, H., “Use of Weighting Functions in DFT/
FFT Analysis (Part II),” Brüel & Kjær Technical Review, No. 4, pp.
1-35, 1987.
6. Tran, T., Dahl, M., Claesson, I., and Lago, T., “High Accuracy Win-
dows for Today’s 24 Bit ADCs,” IMAC 22, Session 29, Signal Pro-
cessing, Soc. for Experimental Mechs., www.sem.org.
7. Harris, F. J., Trigonometric Transforms, a Unique Introduction to the
FFT, Spectral Dynamics, Inc., San Marcos, CA., October 1977.
8. Gaberson, H. A., “The DFT and the FFT as a Discrete Fourier Se-
ries of Sampled Data,” MFPT Advanced Signal Analysis Course
Notes, Section 2, January 2003.
9. McConnell, K. G., Vibration Testing; Theory and Practice, John
Wiley & Sons, Inc., pp. 266-278, 1995. (Most digital signal process-
ing books cover the convolution theorem, but this one is more di-
rected to our area.)
10. Gaberson, H. A.; “Using the FFT for Filtering, Transient Details, and
Resampling,” Proc. National Technical Training Symposium and
Figure 28. Eight linear spectra from the bearing of a large pump, each 27th Annual Meeting, Vibration Institute, Willowbrook, IL, pp. 127-
analyzed with the indicated window. Note the errors in peak ampli- 136, July 2003.
tudes; the flat-top should be most accurate. 11. UN-SCAN-IT, Silk Scientific Inc., Orem, UT.
12. M ATLAB ® for Windows, Version 5.3, High-Performance Numeric
Computation and Visualization Software, The MathWorks, Inc.,
culations were done with M ATLAB.12 I kept copies of all the little Natick, MA, 1999.
programs used to draw the figures. If you would like copies of
the programs to see how I have done them, let me know. The author can be contacted at: [email protected].

INSTRUMENTATION REFERENCE ISSUE 23