Windows, Harmonic Analysis, Discrete Fourier Transform: and The
Windows, Harmonic Analysis, Discrete Fourier Transform: and The
Windows, Harmonic Analysis, Discrete Fourier Transform: and The
NUC TP 532
fredric j. harris
UNDERSEA SURVEILLANCE DEPARTMENT
September 1976
CA
ADMINISTRATIVE INFORMATION
This work was performed as independent research by the author during the period,
May to August, 1976. It was supported by NUC IED computer time for generation of the
experimental test cases and plots.
The author wishes to acknowledge the influence of Don Gingras of NUC, whose
earlier related work was the motivation for this report.
J
y .i,
UNCLASSIFIED
SECURITY C1ISSIFICATION OF THIIS PAGE e"-~ Date. Fn-,d)
READ INSTRUCTIONS
REPORT D)OCUMENTATION PAGE BEFORE COMPLETING FORM
NvUnderea-Cnte
I.DISCRESTEIO (o7
RISTATEMEN R FR 1A Myporlj 7
frmnic anals~istasorm
ou
Fas"t
b
h~aa em eaprmayn oe ssn lfrhrmoi detetio and harm "' nica lyi.- Tr-R
San ieo
R 2132
EDITIONE SOEEUC OF ASPNV1I
SC RIT
LSSF C OTISP
LSF1CATION G
'DW eKrrd
ADING
UNCLASSIFIED
rCURITY CLASSIFICATION OF THIS PAGErflihei Da.a Entoed)
resolution, confidence, and bias of the estimates. We have observed that the trade-offs available
through the use of windows are not well understood nor popularly appreciated in the literature
or by the practitioner.
In addition, the requirement to apply windows to sampled data often leads to subtle
misapplications of the windows. This paper, tutorial and informational, will identify the major
considerations, effects, and pitfalls of which the signal processor should be aware. We will also
identify and clarify points of common misunderstanding concerning sampled windows.
UNCLASSIFIED
SECURITY CLASSIFICATION OF THIS PAGOE'R7en Da.te Ftrned)
iv
I
SUMMARY
OBJECTIVE
To make available a concise review of data windows and their affect on the detection
of harmonic signals in the presence of broadband noise and in the presence of nearby strong
harmonic interference. Also to call attention to a number of common errors in the applica-
tion of windows when used with the Fast Fourier Transform.
RESULTS
RECOMMENDATIONS
*1
CONTENTS
INTRODUCTION 3
V. HARMONIC RESOLUTION 53
VI. CONCLUSIONS 65
APPENDIX 67
BIBLIOGRAPHY 69
2
1. INTRODUCTION
By necessity, every observed signal we process must be of finite extent. The extent
may be adjustable and selectable, but it must be finite. Processing a finite-duration observa-
tion imposes interesting and interacting limitations on harmonic analysis. These include,
detectability of tones in the presence of broadband noise, detectability of weak tones in the
presence of nearby strong tones, resolvability of similar-strength nearby tones, resolvability
of shifting tones. and biases in estimating the parameters of any of the aforementioned sig-
na!s. Similar interactions and limitations apply t,) the analysis of broadband noise signals
such as those used in linear system identification.
For practicality, the data we process are N uniformly spaced samples of the observed
signal. For convenience. N is highly composite, and we will assume N is even. The harmonic
estimates we obtain through the discrete Fourier transform (DFT) are N uniformly spaced
samples of the associated periodic spectra. This approach is elegant and attractive when the
processing scheme is cast as a spectral decomposition in an N-dimensional orthogonal vector
space. The problem is that this elegance must often be massaged to obtain meaningful
results. We accomplish this massaging by the application of windows to the sampled data
set or, equivalently, by smoothing the spectral samples.
The two operations to which we subject the data are sampling and windowing.
These can be performed in either order. Sampling is well understood, windowing is less so,
and sampled windows for DFTs significantly less so! We will address the interacting con-
siderations of window selection in harmonic analysis and examine the special considerations
related to sampled windows for DFTs.
3
Il. HARMONIC ANALYSIS OF FINITE-EXTENT DATA
AND THE DFT
We observe that by defining a basis set over an ordered index k, we are defining the spectrum
over a line (called the frequency axis) from which we draw the concepts of bandwidth and
of frequencies close to and far from a given frequency (which is related to resolution).
For sampled signals, the basis set spanning the interval of NT seconds is identical
with the sequences obtained by uniform samples of the corresponding continuous spanning
set up to the index N/2. See Eq. (2).
We note here that uniformly spaced samples of a continuous orthogonal basis do not, in
general, form orthogonal sequences; the trigonometric functions (sampled over an integer
number of periods) are a convenient exception (rather than the rule). We also note that an
interval of length NT seconds is not the same as the interval covered by N samples separated
by intervals of T seconds. Figure I demonstrates this by sampling a function which is even
about its midpoint and of duration NT seconds. The missing end point is the beginning of
the next period of the periodic sequence and is, in fact, indistinguishable from the zero
point. This lack of symmetry due to the missing (but implied) end point is a source of con-
fusion in sampled window design. This can be traced to the early work related to con-
vergence factors for the partial sums of Fourier Series. The partial sums (or the finite
Fourier transform) always include an odd number of points and exhibit even symmetry
about the origin. Hence the literature and software libraries abound with windows designed
with true even symmetry rather than the implied symmetry with its missing end point!
5
_amna -.-
4, 011, iI , I I '
No Sample
t/T
0 1 2 3 121 4 15 16
NT Seconds
Nth T-sec. Sample
We must remember for DFT processing of sampled data that even symmetry means
that the projection upon the sampled sine sequences is identically zero; it does not mean a
matching left and right data point about the midpoint. To distinguish this symmetry from
conventional evenness we will refer to it as DFT-even (ie., a conventional even sequence with
the right end point removed). Another example of DFT-even symmetry is presented in
Fig. 2 as samples of a periodically extended triangle wave.
If we evaluate a DFT-even sequence via a finite Fourier transform (by treating the
+N/2 point as a zero-value point), the resultant continuous periodic function exhibits a non-
zero imaginary component. The DFT of the same sequence is a set of samples of the finite
Fourier transform, yet these samples exhibit an imaginary component equal to zero. Why
the disparity? We must remember that the missing end point under the DFT symmetry
contributes an imaginary sinusoidal component of period 27r/(N/2) to the finite transform
Periodic extension of
Periodic extension of continuous wave
sampled sequence
-9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9
of next
sequence
-7 -6 -5 -4 -3 -2 -10 12345 6
Figure 2. Even sequence under DFT, and periodic extension of sequence under DFT.
N6
, s
. . 9 . . . .. . . . . . . .. " ". . ." - .. . . . .. .. . . B t . . . . . . . . . i1 "
(corresponding to the odd component at sequence position N/2). The sampling positions of
the DFT are at the multiples of 2wr/N, which, of course, correspond to the zeros of the
imaginary sinusoidal component. An example of this fortuitous sampling is shown in Fig. 3.
Notice the sequence fin) is decomposed into its even and odd parts, with the odd part
supplying the imaginary sine component in the finite transform.
i p..n
-4 -3 -2 -1 0 1 2 3
I In
-4 -3 -2 -1 0 1 2 3 4
-4 -3 -2 -1 0 1 2 3
RL
IM
RL
IM
7
The selection of a finite time interval of NT seconds and of the orthogonal trigono-
metric basis (continuous or sampled) over this interval leads to an interesting peculiarity of
the spectral expansion. From the continuum of possible frequencies, only those which
coincide with the basis will project onto a single basis vector; all other frequencies will
exhibit non-zero projections on the entire basis set. This is often referred to as spectral
leakage! Notice this is a manifestation of processing finite-duration records and is not re-
lated in any way to the periodic sampling.
Windows are applied to data to reduce the spectral leakage associated with finite
observation intervals. From one viewpoint, the window is applied to the data to reduce the
order of the discontinuity at the boundary of the periodic extension. This is accomplished
by matching as many orders of derivative as possible at the boundary. The easiest match, of
course, is zero. Thus windowed data are smoothly brought to zero at the end points so that
the periodic extension is continuous in many orders of derivative.
From another viewpoint, the window is applied to the basis set so that a signal of
arbitrary frequency will exhibit a significant projection only on those basis vectors having a
frequency close to the signal frequency. Of course both viewpoints lead to identical results.
We can gain insight into window design by occasionally switching between these two
viewpoints.
NT
Observed Signal Periodic
I J I J extension
8
Ill. WINDOWS AND FIGURES OF MERIT
Windows are used in harmonic analysis to 1educe the undesirable effects related to
spectral leakage. Windows impact on many attributes of a harmonic processor; these include
detectability, resolution, dynamic range. confidence, and ease of implementation. We would
like to identify the major parameters that will allow performance comparisons between
different windows. We can best identify these parameters by examining the effects on
harmonic analysis of a window.
An essentially bandlimited signal f (t) with Fourier transform F (w) can be described
by the uniformly sampled data set f (nT). This data set defines the periodically extended
spectrum FT (w.) by its Fourier series expansion as identified in Eqs. (3).
+00
+00
+7r/T
Where
IF (w) 1 =0; 1I 27r
and where
F T (w) --F (w ); < 27rl
For (real world) machine processing, the data must be of finite extent, and the summation
of Eq. (3b) can only be performed as a finite sum as indicated in Eqs. (4).
9
(N/2)-l
F (w) = f (nT) e -ij nT N even (4b)
n=-N/2
(N/2)-!
F (wk) f(nT)e-jWkkn T ;Neven (4c)
n=-N/2
where
wk - k
We recognize Eq. (4a) as the finite Fourier transform, a summation addressed for the
convenience of its even symmetry. Equation (4b) is the finite Fourier transform with the
right end point removed, and Eq. (4c) is the DFT sampling of Eq. (4b). Of course for actual
processing, we require realizability, and the summation will shift N/2 positions. This will
affect only the phase angles of the transforms, so for the convenience of symmetry we will
address the windows as being centered at the origin. We also identify this convenience as a
major source of window misapplication. The shift of N/2 points is often overlooked or is
improperly handled in the definition of the window. This is particularly so when the
windowing is performed as a spectral convolution. See the discussion on the Hanning
window under the cosal(x) windows.
The question now posed is, to what extent is the finite summation of Eq. (4a) a
meaningful approximation of the infinite summation of Eq. (3b)? In fact, we address the
question for a more general case of an arbitrary window applied to the time function (or
series) as presented in Eq. (5).
where
w (nT) = 0; In I> (N even)
2
and
: an (nT) = w (-nT); n -,w ( !T) =0
2 2
Let us now examine the effects of the window on our spectral estimates. Equation
(5) shows that the transform Fw (w) is the transform of a product. As indicated in Eq. (6),
this is equivalent to the convolution of the two corresponding transforms (see Appendix).
10
F (o)= F (x)W (w-x) dx/27r (6)
or
Fw (w) = F (w)*W (o)
Equation (6) is the key to the effects of processing finite-extent data. The equation can be
interpreted in two equivalent and enlightening ways, which will be more easily visualized
with the aid of an example. The example we choose is the sampled rectangle window;
w (nT) = 1.0. We know W (w) is the Dirichiet kernel shown in Eq. (7).
7.sin[IjoT]
-j_
W = e(7) (¢o)
sin [ woT]
Except for the linear phase shift term (which will change due to the N/2 point shift
for realizability), the transform has the form indicated in Fig. 5.
W(w)
N
The first observation concerning Eq. (6) is that the value of Fw (w)at a particular o,say
Wo= o,is the sum of all of the spectral contributions at each o weighted by the window
centered at woo and measured at co. See Fig. 6.
II
W(w 0 - W)
Noise Spectra
WOW
00 /
WO1W
W jr
_ _ 41
0 W
12
Ill. A. EQUIVALENT NOISE BANDWIDTH
From Fig. 6 we observe that the amplitude of the harmonic estimate at a given fre-
quency is biased by the accumulated broadband noise included in the bandwidth of the
window. In this sense, the window behaves as a filter, gathering contributions for its
estimate over its bandwidth. For the harmonic detection problem, we desire to minimize
this accumulated noise signal, and we accomplish this with small-bandwidth windows. A
convenient measure of this bandwidth is the Equivalent Noise Bandwidth (ENBW) of the
window. This is the width of a rectangle filter of the same peak power gain that would
accumulate the same noise power. See Fig. 7.
F(w) 2
_F1 - peak power gain = IF(O)12
r-
I I
I I
.equivalent noise bandwidth
I
i
The peak power gain of the window occurs at w = 0, the zero frequency power gain, and is
defined in Eq. (10).
Thus the equivalent noise bandwidth (normalized by the 1/T bandwidth) is defined in
Eq. (I ) and is tabulated for the windows of this report in Table 1.
.w2 (nT)
ENBW= n (1i)
ww(nT)]2
A concept closely allied to ENBW is Processing Gain and Processing loss of a win-
dowed transform. We can think of the discrete Fourier transform as a bank of filters
matched to the set of basis tones. From this perspective we can examine the processing gain
(sometimes called the coherent gain) of the filter, and we can examine the processing loss
due to the window reducing the data to zero near the boundaries. Let the input sampled
sequence be defined by Eq. (12):
14
The incoherent component of the windowed transform is given in Eq. (14a), and the
incoherent power, (the mean square value of this component) is given in Eq. (14b).
-j o o
w (nT) w (mT) E jq (nT) q*(mT)l e G knT e+jw kmT
Finally, processing gain (PG), defined as the ratio of output signal-to-noise ratio to
input signal-to-noise ratio, is given in Eq. (15).
where
WS_ 7r
S 2 N NT
Scalloping loss represents the maximum reduction in processing gain due to signal frequency.
This loss has been computed for the windows of this report and has been included in Table I.
Returning to Eq. (6) and to Fig. 6, we observe the spectral measurement is affected
not only by the broadband noise spectrum, but also by the narrowband spectrum which
falls
=
within the bandwidth of the window. In fact, a given spectral component say at
Wo W. will contribute output (or will be observed) at another frequency, say at W0= Woa
according to the gain of the window centered at o and measured at wa. This is the effect
normally referred to as spectral leakage and is demonstrated in Fig. 8 with the transform of
a finite-duration tone of frequency w o
This leakage causes a bias in the amplitude and the position of a harmonic estimate.
Even for the case of a single real harmonic line (not at a DFT sample point), the leakage
from the kernel on the negative frequency axis biases the kernel on the positive frequency
line. This bias is most severe and most bothersome for the detection of small signals in the
16
F(w)
Signal at w ° being 0
monitored as a
0signal at w
presence of nearby large signals. To reduce the effects of this bias, the window should
exhibit low-amplitude sidelobes far from the central main lobe and the transition to the low
sidelobes should be very rapid. One indicator of how well a window suppresses leakage is
the peak sidelobe level (relative to the main lPbe): another is the asymptotic rate of falloff
of these sidelobes. These indicators are listed in Table 1.
Figure 9 suggests another criterion with which we should be concerned in the win-
dow selection process. Since the window imposes an effective bandwidth on the spectral
line, we would be interested in the minimum separation between two equal-strength lines
such that for arbitrary spectral locations their respective main lobes can be resolved. The
classic criterion for this resolution is the width of the window at the half-power points (the
3.0-dB bandwidth). This criterion reflects the fact that two equal-strength main lobes
A-
-~ ig\
17
separated in frequency by less than their 3.0-dB bandwidths will exhibit a single spectral
peak and will not be resolved as two distinct lines. The problem with this criterion is that it
does not work for the coherent addition we find in the DFT. The DFT output points are
the coherent addition of the spectral components weighted through the window at a given
frequency.
If two kernels are contributing to the coherent summation, the sum at the crossover
point (nominally halfway between them) must be smaller than the individual peaks if the
two peaks are to be resolved. Thus, at the crossover points of the kernels, the gain from
each kernel must be less than 0.5, or the crossover points must occur beyond the 6.0 dB-
points of the windows. Table I lists the 6.0-dB bandwidths of the various windows
examined in this report. From the table, we see that the 6.0-dB bandwidth varies from 1.2
bins to 2.6 bins, where a bin is the fundamental frequency resolution Ws/N. The 3.0-dB
bandwidth does have utility as a performance indicator as shown in the next paragraph.
Remember however, it is the 6.0-dB bandwidth which defines the resolution of the windowed
DFT.
From Table I we see that the noise bandwidth always exceeds the 3.0-dB bandwidth.
The difference between the two, referenced to the 3.0-dB bandwidth, appears to be a sensi-
tive indicator of overall window performance. We have observed that for all the good win-
dows on the table, this indicator was found to be in the range of 4 to 5.5 percent.
Those windows for which this ratio is outside that range either have a wide main lobe or a
high sidelobe structure and, hence, are characterized by high processing loss or by poor two-
tone detection capabilities. Those windows for which this ratio is inside the 4- to S-percent
range are found in the lower left corner of the performance comparison chart, Fig. 10,
which is described next.
While Table I does list the common performance parameters of the windows ex-
amined in this report, the mass of numbers is not enlightening. We do realize that the side-
lobe level (to reduce bias) and the worst-case processing loss (to maximize detectability) are
probably the most important parameters on the table. Figure 10 shows the relative position
of the windows as a function of these parameters. Windows residing in the lower left corner
of the figure are the good performing windows. They exhibit low sidelobe levels and low
worst-case processing loss. Of course, the proof of the pudding is in the eating; see the
conclusion section and Section V.
18
TABLIE I. WINDOWS AND FIGURES O1: MERIT.
COSa x) a - 1.0 -23 -12 0.64 1.23 1.20 2.10 3.01 1.65
HANNING a = 2.0 -32 -18 0.50 1.50 1.44 1.42 3.18 2.00
a - 3.0 -39 -24 0.42 1.73 1.66 1.08 3.47 2.32
a = 4.0 -47 -30 0.38 1.94 1.86 0.86 3.75 2.59
TUKEY a = 0.25 -14 -18 0.88 1.10 1.01 2.96 3.39 1.38
a-0.50 -15 -18 0.75 1.22 1.15 2.24 3.11 1.57
a = 0.75 -19 -18 0.63 1.36 1.31 1.73 3.07 1.80
HANNING- a = 0.5 -35 -18 0.43 1.61 1.54 1.26 3.33 2,14
POISSON a - 1.0 -39 -18 0.38 1.73 1.64 1.11 3.50 2.30
a - 2.0 NONE -18 0.29 2.02 1.87 0.87 3.94 2.65
19
Ck .5
A*So
H inningng
-40 2.40 a.30a-r.
-J 30
anma 0 4.0
C',e
0
zjHimn
-;-- k903,0
:20
IV. CLASSIC WINDOWS
We will now catalogue some well-known (and some not well-known) windows. For
each window we will comment on the justification for its use and identify its significant
parameters. All the windows will be presented as even (about the origin) sequences with an
odd number of points. To convert the window to DFT-even, the right end point will be dis-
carded and the sequence will be shifted so that the left end point coincides with the origin.
We will also use normalized coordinates with sample period T = 1.0, so that W is periodic in
27r and, hence, will be identified as 0.
The rectangle window is unity over the observation interval and can be thought of as
a gating function applied to the data so that they are of finite extent. The window for a
finite Fourier transform is defined in Eq. (I 7a) and is shown in Fig. I I. The same window
for a discrete Fourier transform is defined in Eq. (1 7b).
=
w (n) 1.0; n = -N _ ,,N (1 7a)
2.2
=
w(n) 1.0;n = 0,1 .... N-1 (17b)
The transform of this window is seen to be the Dirichlet Kernel, which exhibits a DFT main-
lobe width of 2(2ir/N) and a first sidelobe level approximately 13 dB down from the main-
lobe peak. The sidelobe fall off at 6.0 dB per octave, which of course is expected for a func-
tion with a discontinuity. The parameters of the DFT window are listed in Table 1.
With the rectangle now defined, we can answer the question posed earlier: in what
sense does the finite sum of Eq. (I 8a) approximate the infinite sum of Eq. (1 8b)?
+N/2
+_0
which the f ()'sare the Fourier series coefficients. We lso recognize that the finite sum is
simply the partial sum of the series. From this viewpoint we can cast the question in terms
of the convergence properties of the partial sums of Fourier series. From this work we know
the partial sum is the least mean square error approximation to the infinite sum.
21
We observe that mean square convergence is a convenient analytic concept, but it is
not attractive for finite estimates or for numerical approximations. Mean square estimates
tend to oscillate about their mean and do not exhibit uniform convergence. (The approxi-
mation at a point of continuity may get worse if more terms are added to the partial sum.)
We normally observe this behavior near points of discontinuity as the ringing we call Gibbs
phenomenon. It is this oscillatory behavior we are trying to suppress by the use of other
windows.
The triangle window for a finite Fourier transform is defined in Eq. (1 9a) and is
shown in Fig. 12. The same window for a DFT is defined in Eq. (1 9b).
=
N
= W(N-n); n ...... N-1
2
The transform of this window is seen to be the squared Dirichlet kernel. Its main-lobe
width is twice that of the rectangle's and the first sidelobe level is approximately 26 dB
down from the main-lobe peak, again, twice that of the rectangle's. The sidelobes fall off
at -1 2 dB per octave, reflecting the discontinuity of the window residing in the first deriva-
tive (rather than in the function itself). The triangle is the simplest window which exhibits
a non-negative transform. This property can be realized by convolving any window (of half
extent) with itself. The resultant window's transform is the square of the original window's
transform!
22
however, and he questioned if some massaging of coefficients might lead to a new set with
more desirable convergence properties. The oscillation of the partial sum and the contrac-
tion of those oscillations as the order of the partial sum increased suggested that an average
of the partial sums would be a smoother function. Figure 13 presents an expansion of two
partial sums near a discontinuity. Notice the average of the two expansions is smoother
than either.
Fm() F(O
W2
100
F0 (0) fo
F4() 0 1 2 3t. I 4f 3 2f 1 Of
We see that the Fejer convergence factors applied to the Fourier series coefficients is, in
fact, a Triangle window.
23
IV. C. COSa (X) WINDOWS
This is actually a family of windows dependent upon the parameter a, with a nor-
mally being an integer. Attractions of this family include the ease with which the terms can
be generated and the easily identified properties of the transform of the cosine function.
These properties are particularly attractive under the DFT. The window for a finite Fourier
transform is defined in Eq. (2 1a) and for a DFT in Eq. (21b). Notice the effect due to the
change of the origin.
The most common values of t are the integers I through 4, with 2 being the most well
known (as the Hanning window). This window is identified for values of a equal to I and 2
in Eqs. (22a), (22b), (23a) and (23b), (the "a" for the finite transform, the "b" for the
DFT). The windows are shown for a integer values of I through 4 in Figs. 14 through 17.
N
=0.5 l0 +c os1[-L ]; -l,0,1,
Notice as a becomes larger, the windows become smoother and the transform reflects this
increased smoothness in decreased sidelobe level and faster falloff of the sidelobes, but with
an increased width of the main lobe.
24
Of particular interest in this family, is the Iann window (after the Austrian meteor-
ologist. Julius Von Hann).* Not only is this window continuous, but so is its first derivative.
Since the discontinuity of this window resides in the second derivative, the transform falls
off as I /w3 or as 18 dB per octave. Let us closely examine the transform of this window.
We will gain some interesting insight and learn of a clever application of the window under
the DFT.
The sampled Hanning window can be written as the sum of the sequence as, indicated
in Eq. (24a). Each sequence has the easily recognized transform indicated in Eq. (24b).
=
w (n) 0.5 + 0.5 cos 2 :n - .- 1,0,1 ...... (24a)
where
-j- sin
D (0)
sin [10]
We recognize the Dirichlet kernel at the origin as the transform of the constant 0.5 samples
and the pair of translated kernels as the transform of the single cycle of cosine samples.
Note that the translated kernels are located on the first zeros of the center kernel and are
half the size of the center kernel. Also the sidelobes of the translated kernel are about half
the size and are of opposite phase of the sidelobes of the central kernel. The summation of
the three kernels' sidelobes, being in phase opposition, tends to cancel the sidelobe struc-
ture. This cancelling summation is demonstrated in Fig. 18.
For the special case of the DFT, the Hanning window is sampled at multiples of
21r/N, which of course are the locations of the zeros of the central Dirichlet kernel. Thus
only three non-zero samples are taken in the sampling process. The positions of these sam-
ples are at -27r/N, 0, and +21r/N. The value of the samples obtained from Eq. (24b)
(including the phase factor e-j(N/ 2 )0 to account for the N/2 shift) are -1/4. +1/2. -1/4.
respectively. Note the minus signs. These result from the shift in the origin for the window.
Without the shift, the phase term is missing and the coefficients are all positive: 1/4, 1/2.
1/4. These are incorrect for DFT processing, but they find their way into much of the
literature and practice.
Rather than apply the window as a product in the time domain, we always have the
option to apply it as a convolution in the frequency domain. The attraction of the Harming
window for this application is twofold; first, the window spectra is non-zero at only three
*The correct name of this window is "'Hann." The term "Hanning"is used in this report to reflect conventional usage.
The derived term "Hann 'd"isalso widely used.
25
W(O)
0.5 DO8)
020821r 21r
02DO+N 0.2 WOe-Ni
I
- '1
data points, and second, the sample values are binary fractions, which can be implemented
as right shifts. Thus the Hanning-windowed spectral points obtained from the rectangle-
windowed spectral points are obtained as indicated in Eq. (25) as two real adds and two
4, binary shifts (to multiply by 1/2).
Thus a Hanning window applied to a real transrorm of length N can be performed as N real
multiplies on the time sequence or as 2N real adds and 2N binary shifts on the spectral data.
One other mildly important consideration, if the window is to be applied to the time data, is
that the samples of the window must be stored somewhere, which normally means addi-
tional memory or hardware. It so happens that the samples of the cosine for the Hanning
window are already stored in the machine as the trig-table for the FFT; thus the window
requires no additional storage.
The Hamming window can be thought of as a modified Hanning window. (Note the
potential source of confusion in the similarities of the two names.) Referring back to Figs.
15 and 18. we note the inexact cancellation of the sidelobes from the summation of the
three kernels. We can construct a window by adjusting the relative size of the kernels as
indicated in Eq. (26a) to achieve a more desirable form of cancellation. Perfect cancellation
of the first sidelobe (at 0 = 2.5 12r/Ni ) corresponds to the Hamming window as indicated
in Eq. (26b).
26
w (n) =a + (l-) cos L (26a)
or
The Hamming window is shown in Fig. 19. Notice the deep attenuation at the missing side-
lobe position. Note also that the small discontinuity at the boundary of the window has
resulted in a I/w (6.0 dB per octave) rate of falloff. The better sidelobe cancellation does
result in a much lower initial sidelobe level of-42 dB. Table I lists the parameters of this
window. Also note the loss of binary weighting: hence the loss of ease of a spectral convolu-
tion implementation.
I. Riesz Window
The Riesz window, identified in Eq. (27), is the simplest continuous polynomial
window. It exhibits a discontinuous first derivative at the boundaries: hence its transform
fall off like I/c 2 . The window is shown in Fig. 20. The first sidelobe is -22 dB from the
main lobe. This window is similar to the cosine lobe [Eq. (22)] as can be demonstrated by
examining its Taylor series expansion.
27
2. Riemann Window
The Riemann window, defined in Eq. (28), is the central lobe of the SINC kernel.
This window is continuous, with a discontinuous first derivative at the boundary. It is
similar to the Riesz and Cosine Lobe windows. The Riemann window is shown in Fig. 21.
R 17r] 2
3. de la Valle'-Poussin Window
=
w (n) 1.0 - 6 [.20-
--- i 0 In (29)
LNI-2J N/2J 4
4. Tukey Window
The Tukey window, often called the cosine-tapered window, is best imagined as a
cosine lobe of width (ci/2)N convolved with a rectangle window of width (1.0 - o/2)N. Of
course the resultant transform is the product of the two corresponding transforms. The
window represents an attempt to smoothly set the data to zero at the boundaries while not
significantly reducing the processing gain of the windowed transform. The window evolves
from the rectangle to the Hanning window as the parameter 0i varies from zero to unity.
The family of windows exhibits a confusing array of sidelobe levels arising from the product
of the two component transforms. The window is defined in Eq. (30).
n N
=0.5 .0+cos wr N] L_ IntI<- 2
The window is shown in Figs. 23, 24, and 25 for values of a equal to 0.25, 0.50, and 0.75,
respectively.
28
5. Bohman Window
0 < InI<N
6. Poisson Window
_ InI
i
w(n)=e N/2 0< InI<-2 (32)
This is actually a family of windows parameterized on the variable Of. Since it exhibits a
discontinuity at the boundaries, the transform can falloff no faster than lI/0. The window
is shown in Figs. 27, 28, and 29 for values of at equal to 2.0, 3.0, and 4.0. respectively.
Notice as the discontinuity at the boundaries becomes smaller, the sidelobe structure merges
into the asymptote. Also note the very wide main lobe: this will be observed in Table I as a
large equivalent noise bandwidth and as a large worst-case processing loss.
7. Hanning-Poisson Window
The Hanning-Poisson window is constructed as the product of the Harming and the
Poisson windows. The family is defined in Eq. (33).
r InI
w (n) = 0.5 [10+ Cos [,, n]e aN/2 (33)
29
8. Cauchy Window
The window is shown in Figs. 33, 34, and 35 for values of a equal to 3.0, 4.0, and 5.0,
respectively. Note the transform of the Cauchy window is a two-sided exponential (see
Poisson windows), which when presented on a log magnitude scale is essentially an isosceles
triangle. This causes the window to exhibit a very wide main lobe and to have a large
equivalent noise bandwidth.
Windows are smooth positive functions with tall, thin (ie., concentrated) Fourier
transforms. From the generalized uncertainty principle, we know we cannot simultaneously
concentrate both a signal and its Fourier transform. If our measure of concentration is the
mean square time duration T and the mean square bandwidth W, we know all functions
satisfy the inequality of Eq. (35), with equality being achieved only for the Gaussian Pulse.
T W -> 1 (35)
47r
Thus the Gaussian pulse, characterized by minimum time-bandwidth product, is a reasonable
candidate for a window. When we use the Gaussian pulse as a window we have to truncate
or discard the tails. By restricting the pulse to finite length, the window no longer is mini-
mum time-bandwidth. If the truncation point is beyond the three-sigma point, the error
should be small and the window should be a good approximation to minimum time-
bandwidth.
2
w (n) = e[ ] n 0 NI (36)
This window is parameterized on a, which acts as the reciprocal of the standard deviation, a
measure of the width of its Fourier transform. Increased t will decrease the width of the
window and reduce the severity of the discontinuity at the boundaries. This will result in an
increased width transform main lobe and decreased sidelobe levels. The window is presented
in Figs. 36, 37, and 38 for values of a equal to 2.5, 3.0. and 3.5. respectively. Note the
rapid drop-off rate of sidelobe level in the exchange of sidelobe level for main-lobe width.
The figures of merit for this window are listed in Table I.
Following the reasoning of the previous section. we seek a window which, for a
known finite duration, in some sense exhibits a narrow bandwidth. We now take a lead
30
from the antenna design people who have faced and solved a similar problem. The problem
is to illuminate an antenna of finite aperture to achieve a narrow main-lobe beam pattern
while simultaneously restricting sidelobe response. (The antenna designer calls his weighting
procedure shading.) The closed-form solution to the minimum mainlobe width for a given
sidelobe level is the Dolph-Tchebyshev window (shading). The continuous solution to the
problem exhibits impulses at the boundaries, which restricts continuous realizations to
approximation, (the Taylor approximation). The discrete or sampled window is not so
restricted and the solution can be implemented exactly.
The relation T n (X) = COS (nO) describes a mapping between the nth order
Tchebyshev (algebraic) polynomial and the nth order trigonometric polynomial. The Dolph-
Tchebyshev window is defined with this mapping in Eq. (37) in terms of uniformly spaced
samples of the window's Fourier transform. To obtain the corresponding window time
samples w (n), we simply perform a DFT on the samples W (k) and then scale for unity peak
amplitude. The parameter o represents the log of the ratio of main-lobe level to sidelobe
level. Thus a value of o equal to 3.0 represents sidelobes 3.0 decades down from the main
lobe, or sidelobes 60.0 dB below the main lobe. The (-I )k alternates the sign of successive
transform samples to reflect the shifted origin in the time domain.
and
c-lI(x) =,-tan- [X V- :
The window is presented in Figs. 39, 40, 41, and 42 for values of af equal to 2.5. 3.0, 3.5,
and 4.0, respectively. Note the uniformity of the sidelobe structure; almost sinusoidal! It
is this uniform oscillation which is responsible for the impulses in the window.
be the Gaussian. In Section G we sought the function with restricted time duration, which
minimized the main-lobe width for a given sidelobe level. We now consider a similar prob-
lem. For a restricted energy, determine the function of restricted time duration T which
maximizes the energy in the band of frequencies, W. Slepian, Pollak, and Landau have deter-
mined this function as a family parameterized over the time-bandwidth product TW, the
31
prolate-spheroidal wave functions of order zero. Kaiser has discovered a simple approxima-
tion to these functions in terms of the zero-order modified Bessel function of the first kind.
The Kaiser-Bessel window is defined in Eq. (38). The parameter ira is half of the time-
bandwidth product.
7ro~ .°
0 N
w (n) = 0< Ini < (38)
where
k=0
This window is presknted in Figs. 43, 44, 45, and 46 for values of a equal to 2.0, 2.5, 3.0,
and 3.5, respectively. Note the trade-off between sidelobe level and main-lobe width.
We now examine the last criterion of optimality for a window, We have already
described the Slepian, Pollak, and Landau criterion. Subject to the constraints of fixed
energy and fixed duration, determine the function which maximizes the energy in the band
of frequencies W. A related criterion, subject to the constraints of fixed area and fixed
duration, is to determine the function which minimizes the energy (or the weighted energy)
outside the band of frequencies W. This is a reasonable criterion since we recognize that the
transform of a good window should minimize the energy it gathers from frequencies re-
moved from its center frequency. Till now, we have been responding to this goal by maxi-
mizing the concentration of the transform at its main lobe.
H ) 1p0
2 n (ira,O) (rat
72 (39)
n I - A2n
Here the ?2n is the eigenvalue corresponding to the associated prolate-spheroidal wave func-
tion I 02n (x,y)| and the ira is the selected half time-bandwidth product. The summation
converges quite rapidly and is often approximated by the first term or by the first two
terms. The first term happens to be the solution of the Slepian, Pollak, and Landau prob-
lem, which we have already examined as th, Kaiser-Bessel window.
32
Minimize fI(W) 2 _W dw (40)
This criterion is one which is a compromise between the Dolph-Tchebyshev and the Kaiser-
Bessel window criteria.
Like the Dolph-Tchebyshev window, the Fourier transform is more easily defined,
and the window time-samples are obtained by an inverse DFT and an appropriate scale fac-
tor. The transform samples are defined in Eq. (41).' See also Eq. (37).
[+ABI[vjk] 2 + 1.0]
where
-
A = sinh (C)= 1 0 2a 1
= cosh ~c
1
y(k)= Ncos cos 7(
This window is presented in Figs. 47, 48, and 49 for values of Ceequal to 3.0, 3.5, and 4.0,
respectively. The main-lobe structure is practically indistinguishable from the Kaiser-
Bessel main lobe. The figures of merit listed on Table I suggest that for the same sidelobe
level, this window does indeed reside between the Kaiser-Bessel and the Dolph-Tchebyshev
windows. It is interesting to examine Fig. 10 and note where this window is located with
respect to the Kaiser-Bessel window; striking similarity in performance!
33
-25 -20 -15 10 -5 0 10 to 15 20 25 2-a-20-15 -10 - 0 a 10 i15 2 25
1.0 l.0
-1 -i 1 _T
0 v00
-- Ia
*10 , a
Figure 11. Rectangle window, Fourier transform, Figure 12. Triangle window, Fourier transform,
log-magnitude of transform. and log-magnitude of transform.
34
1.21 1.21
1.00 1.M
1.0 1.0
0 0
Ode Ode
-20 -20
-40 -40
II
S1 -- 0
2
Figure 14. Cos (n v/N) window, Fourier trans- Figure 15. Cos (n n/N) window, Fourier trans-
form, and log-magnitude of transform. form, and log-magnitude of transform.
35
1.25 125
2-IS
25 1O
S 0 5 10 15 I0 21 -25 -20 -15 -10 -5 0 5 10 15 20 25
-. 0 -0
11
-00 -60
0 r-1 e
-10 i t0
Figure 16. Cos3 (n if/N) window, Fourier trans- Figure 17. Cos 4 (n nr/N) window, Fourier trans-
form, and log-magnitude of transform. form, and log-magnitude of transform.
1 25
1.00X
4 .2
T
- i we t
1.0 1.0
0 0
OdD adS
-20 -20
-40-40
Figure 19. Hamming wifdtow, Fourier transform, Figure 20. Riesz window, Fourier transform,
and log-magnitude of Fourier transform, and log-magnitude of transform.
37
•I O01 0
1.0 1.0
0.5 0'5
J-*1-
0d
-20 -20
*10 -w 0 ! I
'-40 -40
-I0 10
Figure 21. Riemann window, Fourier transform, Figure 22. "dela Valle'-Poussin window, Fourier
and log-magnitude of transform. transform and log-magnitude of transform.
38
1.25 1.25
1.00 1.00
1.0 1.0
0L5 0.5
Frrr 0
vJ w i
Oda 043
-2020 -20
.40
Figure 23. 25% cosine taper (tukey) window, Figure 24. 50%cosine taper (tukey) window,
Fuirtransform and log-magnitude Fourier transform and log-magnitude
of transform. Of transform.
39P
100 1.00
1.0 1.0
0.5 0.5
VI
0 10
e -w T 1--1
OdB OdB
-20 -20
-20
9_ 0 0
I I 0 IT 0
Figure 25. 75% cosine taper (tukey) window, Figure 26. Bohman window, Fourier transform,
Fourier transform and log-magnitude and log-magnitude of transform.
of transform.
40
-25 -20 1.25
1.01.
0.5
r ,
*10
OdS aa
-20
00
Fiue2.Poisson window, Fourier transformn, Figure 28. Poisson window, Fourier transform,
and log-magnitude of transform. (a 2.0) and log-magnitude of transform. (a =3.0)
41
7]
25i
T
4-
1.25
.00.,
{ 1.25
1.00
T T
t 1.0 1.0
," I0.5
0.5
-100
.r
I I + l I III' .
I, ..
.4 -20 0
00
-
+, !o42
-40
Figure 29. Poisson window, Fourier transfonm, Figure 30. Hanning-Poisson window, Fourier
and log-magnitude of transform. (a = 4.0) transform, and log-magnitude of transform.
42 05
T 125 1.25
100 1.00
l,IIi ]:
'~~. jjiiI
T T
I t.0
T
1.0
0.505
0e
-,rr LV I Ir-t
OdB Odo
--- 20 -20
43
4.
1.25
100.0a
T '.25
".0 1.0
/ 0.5 O.S
A
0 -I I
0d 0 d
0 . - 0
Figure 33. Cauchy window, Fourier transform, Figure 34. Cauchy window, Fourier transform,
and log-magnitude of transform. (a = 3.0) and log-magnitude of transform. (a = 4.0)
44
1 .2 102S
1.00 0
T T
I -, 6 ! i'
44 1.0 1.0
0 .50
-# 0 S OS i 0I Od$
-200
40
4
O10"Wfill
Ti
-tI - 40
Figure 35. Cauchy window, Fourier transform, Figure 36. Gaussian window, Fourier transform,
and log-masnitude of transform. (a = 5.0) and log-magnitude of transform. (a = 2.5)
45
1.00 1.01
-21-20 -15 -10 -5 0 £ 10 1s 20 21 -21 -20 -IS -10 -5 0 5 10 1s 20 21
T e. A 5
-I 0 r 0 I
Ode Ode
-20 2
-40 -40
T
-60
-t 0 - 0
Figure 37. Gaussian window, Fourier transform, Figure 38. Gaussian window, Fourier transform,
and log-magnitude of transform. (a = 3.0) and log-magnitude of transform. (a = 3.5)
46
10.
00
1.0 1.0
-11
0 .5M
\ Ode 00
-20 .20
-~4-
-1 0 I' "0 2
Figure 39. Dolph-Tchebyshev window, Fourier Figure 40. Dolph-Tchebyshev window, Fourier
transform, and Jog-magnitude of transform, transform, and log-magnitude of transform.
(a= 2.5) (a 3.0)
47
1.00 1.01)
'
-25 -20 -5 -10 05 20 2 -26 -20 -15 -10 -6 0 S 10 IS 20 2
T
1
.0 1.0
0.$ 0.6
0 0
/ OdB OdSl
-20 -20
-40 . -40
Tl-
S- -II 1
trasfrm an d . iit~o tsom 48; i trnfom an logmagie of tranf.
Figure 41. Dolph-Tchebyshev window, Fourier Figure 42. Dolph-Tchebyshev window, Fourier
: transform, and log-magitude of transform. transform, and log-maguitude of transform.
(a =3.5) (az4.0)
48
1.25 25
1.00 1.00
- + ±r..ui.J....
, iTli"
+±...
Ii1'l
-5 -2 -15 -10 -5 0 5 10 15 20 25 -25 -20 -I -10 -5 0 5 10 is 20 25
7 T
7 +
00 1.0
0.5 .5~
-v0 0
Od8 OdO
(p20 ± ~-20
-40 4
, !0 I T I e I I I
Figure 43. Kaiser.Bessel window, Fourier trans- Figure 44. Kaiser-Bessel window, Fourier trans-
form, and log-magnitude of transform. form and log-magnitude of transform.
(a = 2.0) (a = 2.5)
49
i 1.25 1.25
1.00 1.00
Il
olin
i-
!l''(",l
'
il 0 1 1 1
0.$ 0.5
0 5
!" 1
! t 1
it OdL
-20
- l8 - t i i -
OdL
-20
'1we
-40
S -40
i -so -60
-0 -, - r
Figure 45. Kaiser-Bessei window, Fourier trans- Figure 46. Kaiser-Bessel window, Fourier trans-
form, and log-magnitude of transform. form, and log-magnitude of transform.
(a = 3.0) (a = 3.5) ,
50
I
T 1.25 T 1.25
100 1.00
T T
S 1.0 1.0
0.5 0.5
-#0 N -w
20 -20
I- 40 -40
0
k ~ -60
Ii '-
-# 0 W - 0
Figure 47. Barcilon-Temes window, Fourier Figure 48. Barcilon-Temes window, Fourier
transform, and log-magnitude of transform. transform, and log-magnitude of transform.
(a =3.0) (a 3.5)
5'i
1.25S
1.0
0.5
01
adsl
-20
-40
-60
52
V. HARMONIC RESOLUTION
We now modify the signal slightly so that the larger signal resides midway between
two DFT bins: in particular, at 10.5 ts/N. The smaller signal still resides in the sixteenth bin.
The power spectrum of this signal is shown in Fig. 5 I. We note that the sidelobe structure
of the larger signal has completely swamped the main lobe of the smaller signal. In fact, we
know (see Fig. 1 I) that the sidelobe amplitude of the rectangle window at 5.5 bins from
the center is only 25 dB down from the peak. Thus the second signal (5.5 bins away) could
not be detected because it was more than 26 dB down and, hence, hidden by the sidelobe.
(The 26 dB comes from the -25-dB sidelobe level minus the 3.9-dB processing loss of the
window plus 3.0 dB for positive detection.) We also note the obvious asymmetry around
the main lobe centered at 10.5 bins. This is due to the coherent addition of the sidelobe
structures of the pair of kernels located at the plus and minus 10.5 bin positions. We are
observing the self-leakage between the positive and the negative frequencies. Figure 52 is
the power spectrum of the signal pair modified so that the large-amplitude signal resides at
the 10.25-bin position. Note the change in asymmetry of the main lobe and the reduction
in the sidelobe level. We still can not observe the second signal located at bin position 16.0.
We now apply different windows to the two-tone signal to demonstrate the differ-
ence in second-tone detectability. For some of the windows, the poorer resolution occurs
when the large signal is at 10.0 bins rather than at 10.5 bins. We will always present the
window with the large signal at the location corresponding to worst-case resolution.
The first window we apply is the triangle window (see Fig. 53). The sidelobes have
fallen by a factor of two over the rectangle windows' lobes (eg., the -35-dB level has fallen
to -70 dB). The sidelobes of the larger signal have fallen to approximately -43 dB at the
second signal so that it is barely detectable. If there were any noise in the signal, the second
tone would probably not have been detected.
The next windows we apply are the coscl (x) family. For the cosine lobe, a = 1.0.
shown in Fig. 54 we observe a phase cancellation in the sidelobe of the large signal located
at the small signal position. This cannot be considered a positive detection. We also see the
spectral leakage of the main lobe over the frequency axis. Signals below this leakage level
would not be detected. With a = 2.0 we have the Hanning window, which is presented in
Fig. 55. We detect the second signal and observe a 3.0-dB null between the two lobes. This
53
is still a marginal detection. For the cos3 (x) window presented in Fig. 56, we detect the
second signal and observe a 9.0-dB null between the lobes. We also see the improved side-
lobe response. Finally for the cos 4 (x) window presented in Fig. 57, we detect the second
signal and observe a 7.0-dB null between the lobes. Here we witness the reduced return for
the trade between sidelobe level and main-lobe width. In obtaining further reduction in
sidelobe level we have caused the increased main-lobe width to encroach upon the second
signal.
We next apply the Hamming window and present the result in Fig. 58. Here we
observe the second signal some 35 dB down, approximately 3.0 dB over the sidelobe
response of the large signal. Here, too, we observe the phase cancellation and the leakage
between the positive and the negative frequency components. Signals more than 50 dB
down would not be detected in the presence of the larger signal.
The Riesz window is the first of our constructed windows and is presented in Fig.
59. We have not detected the second signal, but we do observe its affect as a 20.O-dB null
due to phase cancellation of a sidelobe in the large signals' kernel.
The result of a Riemann window is presented in Fig. 60. Here, too, we have no
detection of the second signal. We do have a small null due to phase cancellation at the
second signal. We also have a large sidelobe response.
The result of applying the Tukey family of windows is presented in Figs. 62, 63, and
64. In Fig. 62 (the 25-percent taper) we see the lack of second-signal detection due to the
high sidelobe structure of the dominant rectangle window. In Fig. 63 (the 50-percent taper)
we observe a lack of second-signal detection, with the second signal actually filling in one of
the nulls of the first signals' kernel. In Fig. 64 (the 75-percent taper) we witness a marginal
detection in the still high sidelobes of the larger signal. This is still an unsatisfying window
because of the artifacts.
The Bohman-construction window is applied and presented in Fig. 65. The second
signal has been detected and the null between the two lobes is approximately 6.0 dB. This
isn't bad, but we can still do better. Note where the Bohman window resides in Fig. 10.
The result of applying the Poisson-window family is presented in Figs. 66, 67, and
68. The second signal is not detected for any of the selected parameter values due to the
high sidelobe levels of the larger signal. We anticipated this poor performance in Table I by
the large difference between the 3.0-dB and the equivalent noise bandwidths.
54
The Cauchy-fanily windows have been applied and the results are presented in Figs.
72, 73. and 74. Here too we have a lack of satisfactory detection of the second signal and
the poor sidelobe response. This was predicted by the large difference between the 3.0 dB
and the equivalent noise bandwidths as listed in Table 1.
We now apply the Gaussian family of windows and present the results in Figs. 75, 76,
and 77. The second signal is detected in all three figures. We note as we further depress the
sidelobe structure to enhance second-signal detection, the null deepens to approximately
16.0 dB and then becomes poorer as the main-lobe width increases and starts to overlap the
lobe of the smaller signal.
The Kaiser-Bessel family is presented in Figs. 83 through 86. Here, too, we have
positive second-signal detection. Again, we see the effect of trading increased main-lobe
width for decreased sidelobe level. The null between the two lobes reaches a maximum of
22.0 dB as the sidelobe structure falls and then becomes poorer with further sidelobe level
improvement. Note that this window can maintain a 20.0-dB null between the two signal
lobes and still hold the leakage to more than 70 dB down over the entire spectrum.
Figures 87, 88, and 89 present the performance of the Barcilon-Temes window.
Note the positive detection of the second signal. There are slight sidelobe artifacts. The
window can maintain a 20.0-dB null between the two signal lobes. The performance of this
window is slightly shy of that of the Kaiser-Bessel window, but the two are remarkably
similar.
55
rON~ FFT Bin A-0.
S 1. 10.0 1.00
S u 2. 16.0 0.01
-20
-40
-0
. 10,. 1.00
0 10 20 30 40 60 6 7060 li 0 100
-00 dIB
Sg.2
FF7
Bin
60
Anhp.
00
0dF7Bi
0.01 gnl2. 6.
Am.
t -40
-40
56
p
tI
OdeT F Ai.
Ain Ode FFTBin Anpi
Signsai1. 10.0 1.00 Signal. 1 0.0 1.00
Signl 2. 16.0 0.01 Signal 2. 16.0 001
GO -60
o 10 20 0o ,0 so 6 0o 60 s0 100 0 10 2o 0o 4o ;0 60 70 O go 100
Figure 54. Cos (n it/N) window. Figure 56. Cos3 (n ir/N) window.
-20 t -20
-40 -40
-60 _ - +
30 40 5 0 k S 10k02 0 6 0
60 70 80 90 100 0 30 40 S 60 100
Figure 55. Cos 2 (n if/N) (Hanning) window. Figure 57. Cos4 (n if/N) window.
57
. . .... ...
.. ...... ..... . .. : ...
-..i.. . - ln[ iI .. -. .. .. . - ll ill iir . . . .... . . -- " " - j
1. FFT Sin Amp.
Si10.5 1.00
0d-
T* Sp1.
FIFTBin
10.0
AnW.
1.0
-20
-20 A
Sipal 2. 16.0 1 Signal 2. 16.0 0.01
-40
.0 -60
tO
10 20 36 40 060 70 80 0 16'a 0 10 20 30 4s560 7 500 100
Figure 58. Hamming window. Figure 60. Riemann window.
0d -
FFTB m Ampl.
OdT igalI.a.0
Sina 2. 16.0
'0SiWWa
0.01
1.
Signal 2.
105
16.0
1.00
0.01
-20 - -4oT
U
*1-
-60 6
O1 2o !ko fi 70 so 90 l00
lO o 4 Wo 6o "0 0 1ol0; o 30 4o
58
I
-20
-40
60 10 2 0 0 5 6 0 80 10
-60 ~
-200
0 10 20 30 40 W0 60 70 6 96 10 0
Figure 63. Tukey (50% cosine taper) window.
-20 2
-40 4
-60 __ __ _ __0_ _
59
FPT~m
Ampi . FFT On .
A10.V
FFT- Sin Aro. 0W
S= S :I IS~o
0 0.01 u 2. 16.0 0.01
-0 120
-40 -40
-60 so
0 0 20 36 40 O 70 so 1 0 10 20 30 40 16 00 7v s w
Figure 66. Poisson window (a = 2.0). Figure 69. Hanning-Poisson window (a = 0.5).
-20 -20
-40 -40
-60 .- 6O
1k
0 1 36 0o 6 so go 3b 10 20 20 so s 100k
Figure 67. Poisson window (a = 3.0). Figure 70. Hanning-Poisson window (a = 1.0).
Ol. - . 043
=112. 16.0 0.01 1~. 1"0 '
S~nL2 Ito "I0
-20
-40 -40
-u-j L3
1102'4 k k io o 010 202046 A 1
Figure 68. Poisson window (a 4.0). Figure 71. Hanning-Poisson window (a = 2.0).
60
od-FFTBgin Ampi. 0dB FFT bin A
10.5 1.00 SpD . 1". ;r.
Signal1.
-20 2
-40 -40
-80 -60
Figure 73. Cauchy window (a = 4.0). Figure 76. Gaussian window (a =3.0).
-20
- 20
-40 -40
Figure 74. Cauchy window (a = 5.0). Figure 77. Gaussian window (a =3.5).
61
10. 1.9. 1"* am
-20 -20
-40 -40
16.0 0.01
Spz 14.0 0.01
-40 -20
-40
-m6
0
lS 06 10 20 40 06 gO 110i30
1O
o''C 20 36 -40 60 ?TO 30
Figure 79. Dolph-Tchebyshev window (a =2.5). Figure 82. Dolph-Tchebyshev window (a 4.0).
V
S.l10.5
2. IS0
SWuW 0.01
62
0 Pl FTI Art*#. adf1 FFT Bi Amw.
5w210 0.00 1.
las~ 1.00
2.wl 16.0 0.01 2.W ~
-20 -20
-,40
o+ -40
o
-604 -60
k k-l
0 10 20 30 40- 5060 70 60 90 100 0 10 20 30 40 5080 70 SWOO0 10
Figure 83. Kaiser-Bessel window (a =2.0). Figure 85. Kaiser-Bessel window (a =3.0).
-20 -20
-0; -40
Figure 84. Kaiser-Bessel window (a = 2.5). Figure 86. Kaiser-Bessel window (a 3.5).
63
T n Oe
Od- FFT sn AnO.
I .Oo O.00
Sga2. 16.0 0.01
-20
-40
-0
k
0 10 20 30 40 m0 6 0 So 9o 100
Figure 87. Barcion-Temes window (a = 3.0).
Ode- FI Ailn
mn.
. 1 1.00
2 1.0 0.01
-20
-40
-00
o 0 20 206 1 0 6 70 6 50 1
Figure 88. Barcilon-Temes window (a = 3.5).
-20
-60a
0 k
20 k 4p 16 60 70 a060 100flta
Figure 89. Barcilon-Temes window (a = 4.0).
64
VI. CONCLUSIONS
We have examined some classic "windows and some windows which satisfy some
criteria of optimality. In particular, we have described their effects on the problem of
general harmonic analysis of tones in broadband noise and of tones in the presence of other
tones. We have observed that when the DFT is used as a harmonic energy-detector, the
worst-case processing loss due to the windows appears to be lower bounded by 3.0 dB and
(for good windows) ipper bounded near 3.75 dB. This suggests that the choice of particular
windows has very little effect on worst-case performance in DFT energy detection. We have
concluded that a good performance indicator for the window is the difference between tile
equivalent noise bandwidth and the 3.0-dB bandwidth normalized by the 3.0-dB bandwidth.
The windows which perform well (as indicated in Fig. 10) exhibit values of this ratio be-
tween 4.0 and 5.5 percent. The range of this ratio for the windows listed in Table I is 3.2
to 22.9 percent.
For multiple-tone detection via the DFT, the window employed does have a con-
siderable effect. Maximum dynamic range of multi-tone detection requires the transform of
the window to exhibit a highly concentrated central lobe with very low sidelobe structure.
We have demonstrated that many classic windows satisfy this criterion with varying degrees
of success and some not at all. We have demonstrated the optimal windows perform best in
positive detection of nearby tones of significantly different amplitudes. Also for the same
dynamic range, the three optimal windows (Kaiser-Bessel, Dolph-Tchebyshev, and Barcilon-
Ternes) are roughly equivalent with the Kaiser-Bessel, demonstrating minor performance
ad ,antages over the other two. We note that while the Dolph-Tchebyshev window appears
to be the best window by virtue of its relative position in Fig. 10, the coherent addition of
its constant-level sidelobes detracts from its performance in multi-tone detection. Also the
sidelobe structure of the Dolph-Tchebyshev window exhibits extreme sensitivity to co-
efficient errors. This would affect its performance in machines operating with single-
precision fixed-point arithmetic. This suggests that the Kaiser-Bessel or the Barcilon-Temes
window be declared the top performer. My preference is the Kaiser-Bessel window. Among
other reasons, the coefficients are easy to generate and the trade-off of sidelobe level as a
function of time-bandwidth product is fairly simple.
Finally, we comment that all of the conclusions presented about the window
performance in spectral analysis are also applicable to shading for array processing of spatial
sampled data, including FFT beamforming.
65
APPENDIX
THE EQUIVALENCE OF WINDOWING IN THE TIME DOMAIN TO
CONVOLUTION IN THE FREQUENCY DOMAIN
Let
+*0
f (t) F (w) e-Jwt do/27r
and
+N/2
W (W) = w (nT) e+jwfn T
n=-N/2
Then
+00
F w (W) = I w (nT) f (nT) e+jwnT
becomes
[00 2
= F (x) w (nT) e+j(c( - x) nT dx/2w
=L
L"
+0F
n=-N/2
or
F (w)=F (w) *W (j)
67
BIBLIOGRAPHY
Barcilon, V., Temes, G., "Optimum Impulse Response and the Van Der Maas Function,"
IEEE Transactionson Circuit Theory, Vol. CT-19, No. 4, July 1972, pp. 336-342.
Blackman, R. B., Data Smoothing and Prediction, Addison Wesley, 1965.
Blackman, R. B., Tukey, J. W., The Measurement of Power Spectra, Dover Publications,
inc., 1958.
Brillinger, D. R., Time Series Data Analysis and Theory, Holt-Rinehart and Winston, 1975.
Gingras, D., Time Series Windows for Improving Discrete Spectra Estimation, Naval Under-
sea Research and Development Center, NUC TN-715, April 1972.
Harris, B., Spectral Analysis of Time Series, John Wiley and Sons, 1966.
harris, f. j., "Digital Signal Processing," Class notes, San Diego State University, 1971.
harris, f. j., High-Resolution Spectral Analysis With Arbitrary Spectral Centers and
Adjustable Spectral Resolutions, Naval Undersea Center, NUC TP-440, February 1976
(also appears in Journalof Computers & ElectricalEngineering, Vol. 3, 1976, pp.
171-191.
Helms, H. D., "Nonrecursive Digital Filters; Design Methods for Achieving Specifications on
Frequency Response," IEEE Transactionson Audio and Electroacoustics, Vol. AU-I 6,
No. 3, September 1968, pp. 336-342.
Kuo, F. F., Kaiser, J. F., System Analysis by Digital Computer, John Wiley and Sons, 1966,
pp. 232-238.
Landau, H., Pollak, H., "Prolate-Spheroidal Wave Functions, Fourier Analysis and
Uncertainty - 1I," Bell Telephone System Journal, Vol. 40, January 1961, pp. 65-84.
Parzen, E., "Mathematical Considerations in the Estimation of Spectra," Technometrics,
Vol. 3, No. 2, May 1961, pp. 167-190.
Slepian, D., Pollak, H., "Prolate-Spheroidal Wave Functions, Fourier Analysis and
Uncertainty - 1," Bell Telephone System Journal. Vol. 40, January 1961, pp. 43-64.
69