ria

Tapan K. Sarkar and C. Su Department of Electrical and Computer Engineering Syracuse University 121 Link Hall Syracuse, New York 13244-1240 USA Tel: +1 (315) 443-3775 Fax: +1 (315) 443-2583 E-mail: tksarkar@mailbox.syr.edu

Keywords: Wavelets, wavelet transforms, Fourier transform, Gabor transform

1. Abstract
The wavelet transform is described from the perspective of a Fourier transform. The relationships among the Fourier transform, the Gabor transform (windowed Fourier transform), and the wavelet transform are described. The differences are also outlined, to bring out the characteristics of the wavelet transform. The limitations of the wavelets in localizing responses in various domains are also delineated. Finally, an adaptive window is presented that may be optimally tailored to suit ones needs, and hence, possibly, the scaling functions and the wavelets.

utilized for characterizing waveforms the frequency content of which changes with time. One of the main features that distinguishes the three transforms is the choice of the window function. The purpose of the window function is to localize the information about the signal, in both space and time. To this end, the T-pulse technique provides a very flexible window function, which can be generated a priori, utilizing computerized optimization techniques. Once such methodologies are known, one can then utilize special pulse shapes to enhance detection performance. Hence, waveform shaping and analysis of wave shapes is the main theme of this paper.
3. Mathematical background

2. Introduction

n nature, one does not encounter pure single-tone signals, but signals the frequency content of which varies with time. Generally, frequency is defined as a phenomenon where the period of zero crossings of a signal has a fixed duration for all times, and frequency is related to the period. However, one may define the term instantaneous frequency by defining signals the frequency content of which changes as a function of the time or duration of the signal. A good example of this is speech, where the instantaneous frequency may change from 20 Hz to 20 kHz,depending on the system. Hence, it is interesting to develop methodologies that can be introduced to analyze such signals.
In modern times, such concepts have proven to be useful in modem radar-system analysis. Conventionally, radar has dealt with pure CW signals, which are either frequency modulated or tumed off and on to generate pulses. However, there are other radar systems that deal with wideband pulses. Hence, in understanding how such radar systems work, it is necessary to understand the timefrequency representation of waveforms, and how they are characterized and analyzed.

This section summarizes the various analytical tools that perform time-frequency analysis. The three techniques that are most popular are the Fourier transform, the Gabor transform, and the wavelet transform. These techniques are presented both from a continuous and a discrete-time signal point of view. The basic difference that distinguishes the three transfonn techniques as time-frequency localization tools is the choice OP a window function. The window function is essentially an artifact through which we observe the data. For each of the three transforms, the window function has certain fixed properties. As a result, each of these analysis techniques has preferred domains over which their application provides good results. One of the objectives of this paper is to generalize these analysis techniques, and to describe a window function that is very flexible and can deal with a very broad class of practical target-identification problems.
4. Continuous transforms

4.1 Fourier transform

In this study, we look at the classical Fourier transforms, the Gabor transforms, and the wavelet transforms, which have been

The classical Fourier-transform technique is utilized to find the frequency content of a particular wave shape, p ( t ) ,which has occurred just once, has existed for a time interval [O,T],and is nonexistent for any other times. It is well known that the frequency content of such a signal p ( t ) is given by the Fourier transform P ( w ):

This is part 2 of two-part article. Part 1, which treated discrete wavelet techniques, appeared in the October issue.

36

1045-9243/98/$10.0001998 IEEE

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

Equation (1) can be interpreted as modulating the function p ( t ) by e-wf and then integrating it. The Fourier transform is defined for all values of the variable t [Le., from --co < t < a ] , whereas the Fourier series of the function p ( t ) is defined to be periodic. That Hence, p F ( t ) , is, p ( t ) repeats itself after every period (e.g., 27~). the Fourier series of p ( t ),is defined as
m

By observing Equation (5), it is clear that the short-time Fourier transform is the convolution of the signal p ( t ) with a filter having an impulse response of the form h(t) = w(-t)e@, so that

p F (t ) =
n=-m

c, e+jnf,

where

c, = L T p ( t ) e - j dt . 27T 0

(3)

The problem here is that as the window function gets narrower, the localization information in the frequency domain is compromised. In other words, as the frequency-domain localization gets narrower, the window gets wider, so that localization information in the time domain gets compromised due to the uncertainty principle. [The uncertainty in time, A t , and the uncertainty in angular frequency, A w , are related by AwAt 2 0.5 1. Since the requirements in the time localization and frequency resolution are conflicting, one has to make some judicious choices. The best window w(t) depends on what is the meaning of the term best. The above problem has solutions only under certain conditions. For example, if we require the function w(t) to be symmetric, and we require that its energy be confined to a certain band IwI 5 B , then we know the optimum function w(t) is given by the prolate-spheroidal functions. Other criteria, such as the signal-tonoise ratio, can also be utilized in designing the function w(t), or utilizing other criteria for best [4-71. To obtain the original signal back from the short-time Fourier transform, we observe that [4]

So, for a Fourier series, the function p ( t ) is decomposed into a

sum of orthogonal functions c, elnf. Observe that the orthogonal functions into which p ( t ) is decomposed in Equation (2) is generated by integer dilations of a single function eJt. Also, note that the function p F ( t ) is periodic with period 2 n . In contrast, in the Fourier Transform, the spectrum is decomposed into non-integer dilations of the function eJf ,as we know:
. m

(4)

The problem in analyzing signals that are limited in time-Le., they exist over a finite time window and are zero elsewhere-by Fourier transform techniques is that such functions can not simultaneously be bandlimited. Hence, to represent signals p ( t ) , which exist for 0 5 t 2 T , the Fourier transform is generally not the best way to characterize such signals. However, a short-time Fourier transform has been utilized to analyze such signals in the time-frequency plane. By its very definition, the Fourier transformation uses the entire signal, and permits analysis of only the frequency distribution of energy of the signal as a whole [l]. To solve this problem, many who do spectral analysis have taken a piecewise approach. The signal is broken up into contiguous pieces, and each piece is separately Fourier transformed. The resulting family of Fourier transformations is then treated as if it were the basis for a joint time-frequency energy distribution [ 1-41. In the short-time Fourier transform, the function p ( t ) is multiplied by a window function w ( t ) , and the Fourier transform is calculated. The window function is then shifted in time, and the Fourier transform of the product is computed again. So, for a fixed shift P of the window w ( t ) , the window captures the features of the signal p ( t ) around P . The window helps to localize the timedomain data, before obtaining the frequency-domain information. Hence, the short-time Fourier transform is given by
m

If we set

p = t ,then

So, we can recover the original signal p ( t ) for all t as long as w(0)# 0 . If w(0)= 0 , then choose some other value of p that will guarantee w(0)+ 0 . Observe that it is not necessary to know w(t) for all t in order to recover p ( t ) from its short-time Fourier transform.

Altemately, one can also recover p ( t ) from Equation (7) by multiplying both sides by G(t - p) , and integrating with respect to p . The over-bar denotes the complex conjugate. Then,

PSTFT(WP) =

jp(t)w(t -P)e-
-m

dt >

(5)

m
-m

This, of course, assumes that J\w(t-p)I d p is finite. However, if the window fkction is infinite, then one can multiply both sides of Equation (7) by another sequence ~ ( tsuch ) that

and when w ( t )= 1 for all t, one obtains the classic Fourier Transform.

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

37

can ) Thus, ,B is a shift parameter, and now the functions ~ , , ~ ( t

-m

An attempt was made by Gabor [ 8 ] ,in 1946, to develop a methodology where a function can be simultaneously localized in time and frequency. If that is possible, then the frequency content of any signal can easily be obtained, by observing the response in certain narrow frequency bands. Hence, it is possible to track the instantaneous frequency of a signal. In order to simplify the presentation, we introduce the generalized form of Parsevals theorem for two functions, p ( t ) and q(t), and their Fourier transforms, P ( w ) and Thus,

span any function p ( t ) for all possible choices of the parameter a and ,B. Now, if we look at the Fourier transform of the product of the two functions p ( t ) ~ ~ , ~ which ( t ) ,is called the Gabor transform

G,,p(w)

1 3

then

Note that (here, ~ , , ~ ( is t )assumed to be a real function)

e(@).

(P; 4)

l m 1 p(t)g(t)dt= - P ( w ) e ( o ) d o=-(Pi 2 n -m 2n -m

Q) , (11)
Hence, integrating Equation (16) with respect to O , and substituting Equation (17), one obtains

where (*;a) denotes the inner product between two functions, and the over-bar denotes the complex conjugate. Note that if [l]

then Equation (1 1) along with Equation (12) defines the inverse transform. If

e(@) = 2n6(w - w o ) and q(t) = elmot,

(13)

So, the Fourier transform of the function p ( t ), given by P(w), results from integrating the Gabor transform with respect to all possible delay parameters ,B.

then Equation (13), when substituted into Equation (1 I), defines the forward Fourier transform.

Since ( p ; q ) = -1 ( P ; Q ) , andif q ( t ) = ~ , , ~ ( t ) e @ then , 2n

co
m

In the next section, we see how Gabor, through the choice of certain window functions, made it possible not only to localize a signal in the time domain, but was also able to localize its frequency content in a narrow hequency band.
4.2 Gabor transform [l,8,9]

J p ( t ) q ( t ) d t= jp(t)w,,p(t)e-jox dt
-m
-m

= G,,p(w) =

- 1 -P(rrl)Q(Q)dQ. 2z -m

The objective of the Gabor transform is to expand p ( t ) into a set of functions that are simultaneously limited in both time and frequency. This is in contrast to the Fourier transform, where the expansion is done by functions e-Jwt that are not time limited, but highly localized in frequency. Even though, hom a strictly mathematical point of view, it is not possible to localize a function simultaneously both in the time and frequency domain, this can, however, be achieved from a practical standpoint. Let us illustrate this by considering a family of functions q ( t ) such that a member qa(t) is defined as
1 q,(t)=-e 2&
-t2

And since ~ , , ~ ( tis ) real, we have

= e-( w-n)CrejP(0-n)

4a

Also, since

Then the product p(t)q,(t) can be localized in time from a practical standpoint if a>O. This is because, beyond a certain value t = q , the function qa(t) decays down to practically zero, and so will the product p(t)q,(t). Next, we introduce another parameter, ,L?, so that the function ~ ~ , (which ~ ( t is ) real) is defined by

we have

38

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

4.3 Wavelet transform

A property not possessed by the window functions of the Gabor transform is the additional condition that the windows have no "dc" component,

It is clear from Equation (22) that the Gabor transform also localizes the Fourier transform, P(w) of p ( t ) ( i t in fact, the window function wa,B(t) is highly localized), to give its local spectral information. Hence, not only is the function p(t) localized in time by the function ~ ~ , ~ but ( talso ) , its transform P(w) is localized in frequency by the function
W ~ / ( ~ ~ ) , ~ The ( Wwidth ) .

of the window

function in the time domain, A t , is then obtained as

We have

namely, the average value of the window is zero. This property gives an extra degree of freedom for introducing a dilation (or scale) parameter in the window, in order to make the time-frequency window flexible. With this dilation parameter, the integral wavelet transform provides a flexible time-frequency window, which automatically narrows when observing high-frequency components, and widens when studying low-frequency components. Hence, it is in tune with our auditory and visual sense perceptions, which, at least in the first step, process signals in this fashion. This explains the evolution of the musical scale in the West. For example, Figure 1 shows the location of the notes C and G in the major diatonic scale, for several octaves [4]. On a logarithmic scale, they would appear to be nearly equally spaced. Thus, the notes C and G become sparser and sparser as the frequency increases. The window function in the wavelet transform takes the form of [for both positive and negative scale values of a ]

and wa,$(t)

= -w

&

__

Hence,

The above function is admissible as a window function in the wavelet transform (or succinctly, a wavelet) provided Equation (30) is satisfied. The integral wavelet transform of p ( t ) is defined by WTp:

and the width of the spectrum of the window function in the frequency domain is given by A. :

with a 8 0 . The shortcoming of the short-time Fourier techniques is taken care of by utilizing the scaled window ~ , , ~ ( t For ) . a fixed time

So the product of the functions p(t)wa,s(l) is localized in time at

& , and their p9 2

spectrum [it is the spectrum of the windowed

1 . In addition, version of p ( t ) ]is localized in frequency at R 4& please note that

1 AtAW= -. 2
(28)

width of the window ~ ( t in ) Equation (7), when p ( t ) is a highfrequency signal, many cycles are captured by the window, whereas if p ( t ) is a low-frequency signal, only very few cycles are within the same window. Thus, the accuracy of the estimate of the Fourier transform is poor at low frequencies, and improves as the frequency increases. One way to correct this problem is to replace the window w(t) with a function of both frequency and time, so that the time-domain plot of the window gets wider as frequencies

111 1 1
p
C Q C

'
c
:,.cu.nn

So, the width of the time-frequency window is unchanged for

observing the spectrum of p(t) at all frequencies. In fact, it is seen that the Gabor transform is essentially a short-time Fourier transform with the smallest time-frequency window. It is the smallest time-frequency window because the principle of uncertainty is exactly equal to 1/2, as given by Equation (28).

le!

aw H I
628 H t

wz

i
1 2 2 1 HI

B 116 I

Figure 1. The pitch frequencies corresponding to the keys of C and G on a piano. These correspond to the major diatonic scale of Western music. The spacing is very nonuniform, and will appear to be almost uniform on a logarithmic scale.

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6,December 1998

39

decrease: so that the window adjusts its width according to frequency. This is accomplished by the wavelet transform, as outlined above. For example, if

From the wavelet transform given by Equations (31) or (35), the original function can be recovered, utilizing

then in the frequency domain,

n

where a > 0 . The constant C, is given by

(33) Thus, all the possible states of the window are obtained by frequency scaling the response of the window function. This is in contrast to the STFT (short-time Fourier transform), where the various window functions were obtained as W(w)eJpW, Le., by modulating the spectrum of the window function, which is fixed. Hence, if the center of the window function w(t) is given by

(39) Thus, Equation (39) implies that only certain classes of window functions can be utilized in the wavelet transform, namely, those responses that decay at least as fast as
1

The convergence of the integral in Equation (38) is defined in a weak sense [2], Le., taking the inner product of both sides of Equation (38) with any function g ( x ) L 2 and commuting the inner product with the integral over a,P on the right-hand side leads to the true formula. Since for any absolutely integrable function W ( o )-the Fourier transform of w(t)-is continuous, Equation (39) can only be evaluated provided

t * , and the width of the window in the time domain is given by

A ,:

then the function ~ , , ~ ( t is) a window function with its center

at p + at*,and its width is aA7. In addition, its spectrum is given by

w(0)= 0 ,
m

(40)

where W ( o ) is the Fourier transform of w(t) From Parseval's relation in Equation (19), and utilizing Equation (34), the wavelet transform WTp o f p is given by

or, equivalently, j w ( t ) d t = 0 , i.e., the wavelet w(t) has no dc

-m

value, as mentioned in Equation (29). Implementation of the wavelet transform is equivalent to filtering the signal p ( t ) by a bank of bandpass filters that are of constant Q. The center hequencies of the filters are offset from one another by an octave and, in the limit of Equation (39), this guarantees that the center frequencies of the bandpass filters will never be zero. In the continuous domain, all three techniques (Fourier, Gabor, and wavelet) have good theoretical properties. The question is, what happens in the discrete domain? Do all these properties cany over to the discrete domain, or do certain additional constraints need to be imposed?!

(35)

The window function in the frequency domain is centered at w *,

2A : and has a width -,

factor

with the exception of the multiplicative

a
~

, and the phase factor eJpW, The wavelet transform

274&/ provides local information in the frequency window

5. Discrete transforms

Note that in the wavelet type of analysis, if w * of W ( a w ) is assumed to be positive, then the ratio center frequency bandwidth 2A:ia

w./a

w*

]
.

2Az

(37)

The problem of dealing with the discrete transforms has some advantages and disadvantages. For example, the Fourier transform for the discrete case changes over to the Fourier-series representation. The Gabor transform, on the other hand, provides a nonphysical representation of the spectrum for a certain class of window functions. The discrete-wavelet transform provides a fast algorithm for computation of the wavelet coefficients from a multirate digital-filtering point of view, without even introducing the concept of wavelets.

is independent of the scaling factor, a . The class of bandpass filters represented by Equation (36) as a function of a has the property given in Equation (37), and these are called constant-Q filters. This type of processing is done by the human ear, at least in the first stage of signal detection [4].

5.1 A discrete short-time Fourier transform (DSTFT)

The discrete representation of Equation ( 5 ) is given by [4]

40

/E& Antennas and Propagation Magazine, Vol. 40,No. 6,December 1998

Gabor presented a heuristic argument for iteratively estimating the coefficients ake so as to obtain an approximation j ( t ) . However, no justification of this process was given [8,9]. where Martin Bastiaans [lo] used a different set of expansions for p ( t ) , given by and the p ( n ) are the sampled versions of p ( t ) , with p as an integer. So, the window function is represented by values that are sampled at w(n - p) . The inverse DSTFT of PDLr~FT is given by He demonstrated that when gke(t) replaces Y k t ( t ) in Equation (48), then, indeed, the representation of Equations (48) and (SO) form a complete set, and the representation error, p ( t ) - j ( t ) goes to zero for a sufficiently large number of basis functions gke(t).The important difference between the representations of Equations (49) and (SO) are the choices

(43)

If we set

f l = n ,then
(44)

and Hence, one can recover p ( n ) as long as w(0) # 0 . If w(0)= 0 , then one chooses a value of /3 for which w(m) = w(n - p) z 0 , and the procedure continues.
An altemative representation can be made, provided we have

where Tis the sampling interval. One of the objectives of this representation is to look at the ! ) in the time-frequency plane, and cell centered at the point ( t k ,f how its surrounding regions are distributed. The main objectives of the representation of Equation (48) is that if the signal within the T T window NT - - 5 t 5 NT + - contains a pure sinusoid of fie2 2 quency fo, then we would hope that the expression for the timelimited spectrum would be

For Equation (44), p ( n ) is recovered by

It is interesting to note that the inversion formula is not unique. For example, if zo is a zero of the z transform of the conjugate of the window function, i.e., zo is a zero of the polynomial
m

&(z) = C"(k)z-k,
k=-m

(47) in Equation (46) by

and that IP(NT,f,T)I2 would have a significant magnitude, primarily near f0 . However, if one looks at the representation of p ( t ) given by (48) and (49), then it is not at all clear that there would not be any leakage from nearby cells, where k = N + 1 , N + 2 , or T)I . Indeed, the leakage N - 1, N - 2 , and so on, to IP(NT,f, from neighboring cells can produce erroneous interpretations, as described in [9]. This is because neither Equation (49) nor (SO) has a finite support in the time domain. They extend for all times. One would expect the signals centered around the cell NT to contribute to P(NT,t , f) if and only if the time-domain support in the representation in Equation (48) is finite. Otherwise, the results are not relevant. This is why Lemer [ 111 extended the representation in Equation (SO) to have the basis signals of the general form
2

then if we

replace

P'~TFT(z,~)

P'sTFT(z,m) + z: , then Equation (46) is still satisfied. This is in contrast to the conventional, discrete Fourier transform (or, equivalently, the Fourier series), which provides a unique inverse.

5.2 Discrete Gabor transform

In the discrete Gabor transform, the objective is to represent the signal, p ( t ),by a series,

where (49)

v ( t ) is considered to be a "convenient" finite-energy function, the energy of which is concentrated near t = 0 , and the energy spectrum of which, IV(w)l , is concentrated near w = 0 . This was later 41
2

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 199%

modified by Roach [SI, who demonstrated that unless the function v ( t ) is of finite support in the time domain, representations of the form in Equation (48) produce an energy spectrum IP(NT,f,T)I that has no clear relationship to the energy of the signal concentrated in the corresponding interval in time. Note that this precludes the Gaussian fimction or the prolate-spheroidal functions as possible window functions, since they are of infinite duration! It has been shown that for a proper time-hequency representation, the expansion must be of the form [9]
8 2

Equation (62) guarantees that the coefficient ckf for a given segment centered at t = tk is completely independent of the signal outside the Mh segment. To confirm this, observe that

Now note that, always, k = m : otherwise, the window functions w(t - t k ) and w(t - tm) are disjoint. We have

-m

tk -Ti2

where

and finally The window function ~ ( t in ) Equation (56) is chosen in such a way that
Ti2
WT(W)=

~ ( t )
-Ti2

dt ,

(57)

and

with (59) The class of window functions that possesses this property is formed by convolving a rectangular pulse of duration (T - p) with
a symmetric positive pulse a ( t ) having duration 1 __

Utilizing the property of Equations (58) and (59) of the window function, it is seen that Equation (65) becomes unity for e = n and zero otherwise, satisfying Equation (62), and hence Equation (61) holds. The interesting point in Equation (61) is that the coefficient for a given segment in time is completely independent of the signal outside the kth cell, since the coefficients in different time segments are orthonormal, Le., the total energy in the signal is then the sum of all the individual coefficients squared. It is interesting to observe that Equations (56) and (61) have close resemblance to the windowed Fourier transform, presented in the earlier sections, when the window function becomes a rectangular window. However, any window function that results from the convolution of any rectangular pulse with a symmetric window function will provide the mathematical requirements for ~ ( t in ) Equations (56) to (59).
5.3 Discrete wavelet transform

p and area

T-p'
Here,

If, in the continuous wavelet transform, one uses integer values for some integers k,n in Equation (32), and usually assumes

p = 2knT (one can assume T = 1 without loss of generality for the

discrete case) and a = 2 k , then the discrete wavelet transform of p ( t ) is given by [4] where Q is the integer frequency-spacing factor. The coefficients eke in Equation (55) can be given by

The inverse transform is given by provided

p ( t )=

WT,(k,n)2-k'2w(2-kt-n~),

(67)

under the condition that the window functions, given by where the over-bar denotes the conjugate.

42

/Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

are orthonormal, i.e.,

W ( X )=

dm4(2x- m) .
m=-m

The + ( x ) are called the scaling functions, and are generated by the dilation equation The shift integers n are chosen in such a way that ~ ( 2 - ~n tT ) covers the whole line for all values of t. The wavelet transform thus separates the object into different components in its transform domain, and studies each component with a resolution matched to its scale. Another interesting property to observe is that the original signal in Equation (67) is recovered from the discretewavelet-transform coefficients by filtering with the window ~,,~(t of ) Equation (68). Equivalently, the discrete wavelet trans(73) where, in each case, only a finite number of coefficients cm and dm are different from zero.

Here, 4 does not have integral zero, but w does, and q5 is form is obtained by filtering the signal p ( t ) by ~ ~ , ~ ( ort ) , normalized such that 2-kZw(-2-kt).
m

J #(x)dx = 1.
The wavelet series of Equation (67) amounts to expanding the function p ( t ) in terms of wavelets wk,n(t) or, equivalently, by the shifted-dilated versions of the same window function, so that
-m

(74)

We define (bk,n even though

q5k,n
= 2-k24(2-%

4 is not a wavelet, i.e.,

(75)

-n).

The inner products of Equation (71) can be evaluated using Ifwe further assume that the wavelets that Equation (69) holds), then
wk,n(X)

are orthogonal (i.e.,

So, the problem of finding the wavelet coefficients is reduced to

By comparing Equations (67) and (71), it is apparent that the (k, n)th wavelet coefficient of p is given by the integral wavelet transform ofp, if the same orthogonal wavelets are used in both the integral wavelet transform and in the wavelet series. So the wavelet series provides an approximation for p that is not necessarily the
least-squares (l?) orthogonal projection that a Fourier Series provides. Also, the wavelets provide an unconditional basis for that of computing ( p ;$ j , k ) . Note that

(77)

f for

1 < i < co . Since 2 has no unconditional bases, wavelets cannot do the impossible. However, they can still do a better job than the Fourier expansion, by displaying no Gibbs phenomenon for approximating functions that are discontinuous, by utilizing the Haar wavelets, for example. However, if the wavelets used in the approximation are not discontinuous, like the Haar wavelets (or Walsh functions), but utilize continuous functions, instead, then the Gibbs phenomenon is visible in the wavelet approximation. The problem now at hand is to determine if are there any numerically stable algorithms to compute the wavelet coefficients Ck,* in Equation (71). Specifically, in real life, p is not a given function, but is a sampled function. Computing the integrals of then requires a quadrature formula. For the smallest value of koften referred to by the scale parameter, Le., most negative kEquation (71) will not involve many samples ofp, and one can do the computation quickly. For large scales, however, one faces large integrals, which might considerably slow down the computation of the wavelet transform o f any given function. Especially for on-line implementations, one should avoid having to compute these long integrals. One way out is the technique used in multirate/multiresolution analysis, by introducing an auxiliary function, @(x), so that [2]

so that ( P ; $ ~ , can ~ ) be computed recursively, starting from the smallest scale (most-negative k ) to the largest. The advantage of this procedure is that it is numerically robust, namely, even though the wavelet coefficients ck,nin Equation (71) are computed with low precision-say, with a couple of bits-one can still reproduce p ( t ) with comparatively much higher precision [2].
In summary, what we have done is as follows. Consider p ( t ) to be a function of time. We have taken the spectrum of p ( t ) , and have separated the spectrum into octaves of widths A u k , such that ~ 2 k + z ] ,for its frequency band w has been divided into [ 2 k to all values of k. We now define wavelets in each frequency bin Amk , and approximate p ( t ) by them. If we choose

4(t) =
then

-3

sin nt nt

and the wavelet expansion of p ( t ) is given by

~ ~ , = ~ 2k2y1(2kt ( t ) - n) (with T = 1),

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6,December 1998

The functions ~ ~ , ~ are ( orthonormal t ) because their bandwidths are non-overlapping, namely, for a fixed k, P k ( 0 ) has the bandwidth A m k , which is [ 2 k n , 2k+n]. So, the wavelet expansion of a function is complete in the sense that it makes an approximation by orthogonal functions that have non-overlapping bandwidth.

possibly, sine functions. Even though they generate orthogonal bases with some localization, there is a problem, as with the Gabor basis. Unless the windows have the specific properties as outlined by Equations (56) to (59) the windowed localized bases of Equation (61), which are similar to the Malvar wavelets, do not always provide the correct localization picture in the transformed domain!

6. A discussion of possible choices of the window function

As concluded by Vaidyanathan [ 4 ] , even though the continuous wavelet transform has a wider scope with deeper mathematical issues, the discrete wavelet transform is quite simple, and can be explained in terms of basic filter theory. Even before the development of wavelets, nonuniform filter banks had been used in speech processing by [12, 131. The motivation was that the nonuniform bandwidths could be used to exploit the nonuniform frequency resolution of the human ear [14]. So if the wavelet application is already in the digital domain, it is really not necessary to understand the deeper results of the scaling function 4(t) and wavelets
~ ( t . In ) this case, all we need to focus on is the dilation equation and the shift principle to generate a complete basis.

The claim that a particular transform provides simultaneous localization, strictly in time and frequency, is preposterous. What we have seen is that if a function is limited in time, it cannot simultaneously be limited in frequency, Le., bandlimited. Hence, resolution in time is accomplished at the expense of the frequency, such that for the best window, we have the situation that the Heisenberg principle of uncertainty, A t A o 5 112 is satisfied with the equality. The equality is achieved only for a Gaussian window function. However, as we have seen for the Gabor transform, the Gaussian window provides nonphysical power spectral density for a region of the function localized in time. The wavelet transform cannot do any better! For all other choices of the window function, A t A , is > 112. Other possible choices for the window may be the prolatespheroidal functions. Unfortunately the prolate-spheroidal functions are not limited in time, but they are strictly bandlimited. An altemate choice is to deal with the truncated prolate-spheroidal functions, which are strictly time limited, but one does not have any control over the spectrum of such functions [15-171. That is why a flexible methodology has been developed, where one can design a window shape that it is finite in time and, in addition, can focus the energy into a pre-specified band [18, 191. However, what one may try to do is to have a finite time window for localizing a process in time. However, its shape can be manipulated in such a way that it is approximately bandlimited for most practical purposes, Le., 99.9% of its energy is within a very narrow prescribed band. Such a construct is called a T pulse. A T pulse is a pulse limited in time T. Moreover, its spectral content, Le., 99.9% of its energy, is focussed within a few band times (i.e.,

5.3.1 First word of caution. Often, over-zealous researchers, in order to push the advantages of this newly developed tool (which, of course, is quite useful for selective problems), overlook two important factors.

The first one is termed a wavelet crime by Strang and Nguyen. The problem has to do with the selection of sample values of the function as the wavelet coefficients. As Strang and Nguyen point out, ...start with a function ~ ( t )Its . samples x ( n ) are often the input to the filter bank (to generate the wavelet decomposition). Is this legal? NO. IT IS A WAVELET CRIME. Some cant imagine doing it, others cant image not doing it. Is this crime convenient? YES. We may not know the whole function n(t), it may not be a combination of 4(t - k ) , and computing the true coefficients in x a ( k ) 4 ( t - k ) may take too long. But the crime cannot go unnoticed-we have to discuss it [20,p. 2321. The problem is in Equations (71) and (76),where the samples of x ( n ) are used at the zeroth scale to initiate the pyramid algorithm for doing a discrete wavelet decomposition. As Strang and Nguyen point out, [20, p. 2321 ...the pyramid algorithm acts on the numbers x ( n ) as if they were expansion coefficients of its

z), where n is about 2 or 3. T

The design of the T pulse is carried out in the discrete domain. Let us assume a discrete signal sequence f(m), which is defined for m = 0, 1, 2, ..., N , - 1, and is identically zero outside these N m values. Let us assume there are N, samples in one baud time (in an approximate way, the baud time is the time duration between zero crossings of a signal). Then, the total number of baud times N, i s

t n ) , i.e., underlying function x s ( t ) = C x ( n ) @ (-

Does xs(t) have the correct sample values x ( n ) ? This seems a minimum requirement. They further point out the correct way to solve the problem. There is a perennial problem of relating the sampled values of the function to the continuous function, as the mapping from the sampled z domain to the continuous s domain is not unique [21].

The discrete Fourier transform (DFT) of the signal f(m) is given by

for k 5.3.2 Second word of caution. It is possible to generate a local orthogonal basis, by using the Malvar wavelets, for example [22]. The Malvar wavelets are basically windowed cosine and,
44

= 0,

1,..., Nk -1

So, in the frequency domain, N k is the total number of samples of the DFT sequence F ( k ). In the frequency domain, if we

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6,December 1998

assume there are N,. samples per baud rate (the inverse of baud time), then

and increasing N,. increases the resolution in the frequency domain. In the T pulse construction, the objective is to maximize the in-band energy within the set 4 = {-Nb I k I Nb}. Or, equivalently, it minimizes the energy outside the Nb samples. In addition, the wave shape has to be orthogonal to its shifted version, and this will minimize the inter-symbol interference. This implies that In summary, the following observations are of importance.
I. Note that the objective is to minimize the cost function, J, such that E,,, is minimum with em = 0 , and ep = 0 for p = O , l , ..., N , - l .

m=O

2. The weights w e , wm, and w p ( p = 0, 1,..., N , - 1 ) should be adjusted in a search procedure to achieve the above goal. The minimization process can be outlined as follows:

for k = 0 , 1 , ..., N c - I , where 6 ( k ) is the impulse function. This guarantees that if the waveform is shifted by a baud time or its multiples, then the wave shape is orthogonal to itself. This is termed zero inter-symbol interference. Note that when k = 0, it is the square of the function itself, and no constraint needs to be put on that. In addition, we need to put in a dc constraint, i.e., the wave shape should have no dc component. Hence,

Step 1. Choose an initial guess for f (m)and an initial guess for the weights we, wm, and w p , p = 0, 1, ..., Ne - 1 . Step 2. Compute the gradient of the functional Jwith respect to f ( m) . This is given by

So, the cost function, J , that will be minimized is

N -1

p=o

where w e , wm, and wp are various weights to the errors E,,,

, em, ep . The weights should be adjusted in a search procedure that has been designed to minimize J.
In addition,
E,,,
=

assuming f ( m ) is real. Next, an optimum step length to update the signal sequence f ( m) is chosen, through one-dimensional searches.

out-of-band energy

and

m=O
-0.4

N,-1

e,

= C f ( m ) f ( m + p N s ) - 6 ( p )for p = l , 2, ..., N c - l .
m=O

(90)

-0.61

-0.8
1

Equivalently,

1 0

20

30

40

50

60

70

BO

90

100

Time Index
m=O

Figure 2. A baseband T pulse with zero mean and uon intersymbol interference.
45

IEEE Antennas and Propagation Magazine, Vol. 40,No. 6 , December 1998

20

Note that throughout the process, we is a fixed nonzero value, since the absolute value of the total energy is not important. However, we can be increased during the process, if the in-band energy to generate the T pulse is unexpectedly low. This may happen since the cost function may have more than one local minimum, and increasing w e can make one jump out of an undesired local minimum.

Frequency Index Figure 3. The spectrum of the zero-mean baseband Tpulse.

An example is presented, showing the synthesized T pulse in the time domain, in Figure 2. Its spectrum is in Figure 3. The T pulse has been designed for Nbaud= 4 , Nband= 130, and for 100 samples. Note that the T pulse has zero mean, and is orthogonal to its shifted versions. The second figure shows the spectrum of the T pulse. It is seen that the bandwidth of the Tpulse is 13 units, as seen from the FFT. As seen, one can design a window function as shown in Figure 2 that has zero mean and, in addition, has no intersymbol interference, i.e., the waveform is orthogonal to its shifted version. This window function is called the T pulse or the timelimited pulse. The spectrum of the Tpulse is illustrated in Figure 3, and is shown to be very highly localized. To further illustrate the application of time limiting and, practically, band-limiting a signal, we consider the electromagnetic scattering from a thin wire. If a Gaussian-shaped pulse excites a thin-wire antenna 2.0 m in length and with a radius of 0.01 m, then the current at the center point of the wire is given in Figure 4. If we

-1.5
0

10

20

30

50

Time (light-meters) Figure 4. The current at the center point of a thin-wire antenna 2.0 m in length and with a radius of 0.01 m, excited by a Gaussian-s~aped pulse.

10

20

30

40

50

60

,o

80

90

100

,io

120

130

Time Index Figure 6. A T pulse with a dc value that is not zero.

-30
-40

Frequency Index Figure 5. The Fourier transform of the current in Figure 4.

Step 3. If the norm of the previous gradient vector is not small enough, go to Step 2. Otherwise, see whether the
orthogonality errors lepl, for p = 1, 2, ..., N,
- 1,

-'to -120 0

are small enough.

-.....- ,........
IO
20

\ , , , , .--..,..I.....

If the errors are small enough, stop the process. If the errors are still considered to be large, increase each w p ( p = 1, 2, ..., N, -1) by a factor, and then go to Step 2.
46

30

50

SO

Frequency Index Figure 7. The spectral content of the Tpulse shown in Figure 6.

/Antennas and Propagation Magazine, Vol. 40,No. 6,December 1998

1.0

0.9

0.8.
0.7

0.6

0.1
0.4.

0.3.

0.2
0. I -0.0

-0.1.

6-0.a

-0.1.
-0.4'

.CI

VI

-0,s.
-0.6.
-0.7

-0.0

-1.0
0
10

-'O 20

+.--,
0

, . , . . . ~~.
IO

., . , , . . .
0 M

30

40

1 1 0

70

80

90

100

110

120

130

zo

30

70

Time Index Figure 8. The amplitude-modulated carrier with the T pulse.

Frequency Index Figure 11. The spectral content of the signal in Figure 11.
observe the Fourier transform of the current, it is given in Figure 5. Now, the question is, it is possible to excite the second resonance in the wire structure, which is 20 dB down, without exciting the first resonance? To this effect, we amplitude modulate a carrier frequency that is at the second resonance by the T pulse. In this example, we consider a T pulse with a dc value that is not zero, as is shown in Figure 6. The spectral content of the T pulse is given by Figure 7. The amplitude-modulated carrier with the T pulse is shown in Figure 8. Its spectral content is given in Figure 9. Observe that the spectral energy is dominant around the second resonance. Now, if we excite the dipole antenna with this incident pulse of Figure 8, the current at the center is given in Figure 10. The spectral content is given in Figure 11. From Figure 11, it is quite clear that it is possible to generate the second resonance without exciting the dominant first resonance. Hence, this multiresolution analysis can have significant potential application in target identification.

E
Y

e
Q)

r A

.IO0
0
10 20

30

50

60

Frequency Index Figure 9. The spectral content of the signal in Figure 8.

Finally, we conclude this section by pointing out the differences between a T pulse and a wavelet. In the T pulse, a window has been created that is not only limited in time, but 99% of the energy is focused in a narrow band and, in this way, an attempt is made to satisfy the equality in the Heisenberg principle of uncertainty, i.e.,
A,A,
4 0.5.

0.017 0 015
0.013.

O.OI,
0.00s.
0.007.

0.00S.

0.003.
0.001.
-0.001.

-0.003. -0.005 -0.007.

-0.009.
-0.0,1.

The expansion in Equation (70) can be done in terms of a dilated and shifted version of the T pulse in an approximate fashion. For the wavelets. the decomposition by Equation (70) is essentially exact, and perfect reconstruction can be done, even after the functions have been down-sampled [4]. Also, in Equation (70), there is a desire to limit the number of nonzero coefficients in Equation (70), to improve the efficiency of the decomposition. This is achieved by enforcing that the pth derivative and all its lower derivatives are zero at co = 0 , Le.,

-0.013'
-0.O,S,

-0.017

\-.-.-.-~,-~.-0

.,..
30

. . . ... , . . ,.. . .. .
0

(93)

IO

20

HI

Time (light-meters) Figure 10. The current at the center of the dipole antenna, when excited by the incident pulse of Figure 8.

Depending on the nature of the requirements for the problem of interest, one can use either of the constraints on the window fiinctions.

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

47

7. Conclusion

This paper provided a survey of the short-time Fourier techniques, the Gabor transform, and the wavelet transform. The strengths and the weaknesses of each method have been identified. The unity in diversity between the three transforms is the choice of the appropriate window function. In addition, it has been shown how to apply numerical-optimization techniques to design a flexible window function, such that it is strictly time limited, yet 99.9% of its energy can be focused in a narrow band.

13. R. W. Schafer, L. R. Rabiner, and 0. Herrmann, FIR Digital Filter Banks for Speech Analysis, Bell System Technical Journal, 54, March 1975, pp. 531-544. 14. J. L. Flanagan, Speech Analysis, Synthesis and Perception, New York, Springer-Verlag, 1972. 15. I. Gerst and J. Diamond, The Elimination of Intersymbol f the IRE, 19, Interference by Input Signal Shaping, Proceedings o 1963, pp. 1195-1203. 16. P. H. Halpern, Optimum Finite Duration Nyquist Signals, IEEE Transactions on Communications, COM-27, June 1979, pp. 886-888. 17. E. Panayirci and N. Tugbay, Optimum Design of Finite Duration Nyquist Signals, Signal Processing, 7,1984, pp. 57-64. 18. Y. Hua and T. K. Sarkar, Design of Optimum Discrete Finite Duration Orthogonal Nyquist Signals, IEEE Transactions on Acoustics, Speech and Signal Processing, ASSP-36, 4, April 1988, pp. 606-608. 19. T. K. Sarkar, S. M. Naryana, H. Wang, M. Wicks, and M. Salazar-Palma, Wavelets and T-pulses, in L. Carin and L. B. Felsen (eds.), Ultra- Wideband Short-Pulse Electromagnetics 2, New York, Plenum Press, 1995, pp. 475-485. 20. G. Strang and T. Nguyen, Wavelets and Filter Banks, Wellesley Cambridge Press, 1996. 21. T. K. Sarkar, N. Radhaknshna, and H. Chen, Survey of Various z-Domain to s-Domain Transformations, IEEE Transactions on Instrumentation and Measurements, IM-35, December 1986, pp. 508-520. 22. M. Vettereli and J. Kovacevic, Wavelets and Subband Coding, New York, Prentice Hall, 1995.

8. Acknowledgment The authors would like to acknowledge the suggestions of the Editor-in-Chief regarding ways to improve the readability of the article.
9. References

1. C. K. Chui, An Introduction to Wavelets, New York, Academic Press, 1992. 2. I. Daubechies, Ten Lectures on Wavelets, Philadelphia, Society of Industrial and Applied Mathematics (CBMS, 61), 1992.
3. R. Gopinath and C. S. Bums, Wavelet Transforms and Filter Banks, in C. K. Chui (ed.), Wavelets-A Tutorial in Theoly and Applications, New Nork, Academic Press, 1992, pp. 603-654.

4. P. P. Vaidyanathan, Multirate Systems and Filter Banks, New Jersey, Prentice Hall, 1993.

5. F. I. Tseng, T. K. Sarkar, and D. D. Weiner, A Novel Window for Harmonic Analysis, IEEE Transactions on Acoustics, Speech andsignal Processing, April 1981, pp. 177-188.
6. W. F. Walker, T. K. Sarkar, F. I. Tseng, andD. D. Weiner, Carrier Frequency Estimation Based on the Location of the Spectral Peak of a Windowed Sample of Carrier Plus Noise, IEEE Transactions on Instrumentation and Measurements, IM-31, December 1982, pp. 239-249. 7. W. F. Walker, T. K. Sarkar, F. I. Tseng, and J. Cross, Optimum Windows for Carrier Frequency Estimation, IEEE Transactions on Geoscience Electronics, GE-21, October 1983, pp. 446-454. 8. D. Gabor, Theory of Communications, Journal o f the Institute for Electrical Engineers, November 1946, pp. 429-457. 9. J. E. Roach, A Vector Space Approach to Time-Variant Energy Spectral Analysis, PhD thesis, Syracuse University, Syracuse, New York, 1982. 10. M. Bastiaans, Gabors Expansion of a Signal into Gaussian Elementary Signals, Proceedings ofthe IEEE, 68, April 1980, pp. 538-539. 11. R. Lemer, Representation of Signals, in E. Baghdady (ed.), Lectures on Communrcation Svstem Theow, New York, McGrawHill, 1961, Chapter 10. 12. G. A. Nelson, L. L. Pfeiffer, and R. C. Wood, High Speed Octave Band Digital Filtering, IEEE Transactions on Audio and Electroacoustics, AU-20, March 1972, pp. 8-65.
48

Introducing Feature Article Authors

Tapan Kumar Sarkar received the B. Tech. degree from the Indian Institute of Technology, Kharagpur, India, the MScE degree from the University of New Brunswick, Fredericton, Canada, and the MS and PhD degrees from Syracuse University, Syracuse, New York in 1969, 1971, and 1975, respectively.
From 1975 to 1976, he was with the TACO Division of the General Instruments Corporation. From 1976 to 1985, he was with the Rochester Institute of Technology, Rochester, N Y . From 1977 to 1978, he was a Research Fellow at the Gordon McKay Laboratory, Harvard University, Cambridge, MA. He is now a Professor in the Department of Electrical and Computer Engineering, Syracuse University; Syracuse, N Y . He has authored or co-authored more than 170 joumal articles, and has written chapters in ten books. His current research interests deal with numerical solutions of operator equations arising in electromagnetics and signal processing, with application to system design. Dr. Sarkar is a registered Professional Engineer in the State of New York. He was an Associate Editor for Feature Articles of the IEEE Antennas and Propagation Society Newsletter, and he

IEEE Antennas and Propagation Magazine, Vol. 40, No. 6 , December 1998

was the Technical Program Chairman for the 1988 IEEE Antennas and Propagation Society International Symposium and URSI Radio Science Meeting. He was the Chairman of the Inter-Commission Working Group of URSI on Time-Domain Metrology. He is a member of Sigma Xi and USNC/URSI Commissions A and B. He received one of the best solution awards in May, 1977, at the Rome Air Development Center (RADC) Spectral Estimation Workshop. He received the Best Paper Award of the IEEE Trunsactions on Electromagnetic Compatibility in 1979, and at the 1997 National Radar Conference. He is a Fellow of the IEEE. He received the degree of Docteur Honoris Causa from the UniversitC Blaise Pascal, Clermont-Ferrand, France, in 1998. Chaowei Su was bom in Jiangsu, Peoples Republic of China, on April 23, 1961. He received the BS and MSc degrees from Northwestem Polytechnical University (NPU), both in the Department of Applied Mathematics, in 1981 and 1986, respectively. He joined the Department of Applied Mathematics at NPU in 1982. Mr. Su was appointed as an Associate Professor and a full Professor at NPU in 1993 and 1996, respectively. He was a Visiting Scholar at S U N Y , Stony Brook, New York, USA, between December, 1990, and June, 1992, supported by a Grumman Fellowship. Since August, 1996, he has been a Visiting Professor at Syracuse University, USA. He is an author of Numerical Methods in Inverse Problems o f Partial Differential Equations and Their Applications (published by NPU Press). His main research interests are in the area of numerical methods of electromagnetic scattering and inverse scattering, and inverse problems of partial differential equations.
111111111111111111111111111111111111111111111111111111111111111111111111111

~~

N A

SENIOR DESIGN ENGINEER

South Florida

Senior level position with South Florida avionics company. Design experience with Broad Band Aircraft Antennas from 1 MHz to 5 GHz desirable. BSEE or higher Degree required. Full benefits.

DAYTON-GRANGER, INC. P.O. Box 350550 Ft. Lauderdale, FL 33335 TEL: 954-463-3451

FAX: 954-761-3172 WWW.daytongranger.com E-mail: sales@daytongranger.com

So why settle for a limited planning tool?

for indoor and outdoor microcell stu&s

EDX SignalPro includes SignalMXTM

Some planning tools are great at cellular or PCS hut useless otherwise. EDX SignalProTM is a ful-featured planning tool for Windows@95/NT that provides comprehensive, accurate designs for any system in the 30 MHz to 100 GHz frequency range. Cellular Wireless local loop PCS Microcell Pqng Indoor wireless MDS Landmobile LMDS Microwavelink AM, FM, and TV broadcast (DTV)
EDX SignalPro and its modules offer 21 different propagation models, including transnutter-specilkmeasurement-derived models, and full 3D ray-tracing models

(pat.pend.) making it the worlds only planning tool with multimedia simulations of system performance. EDX also offers an extensive collection of terrain, groundcover,demographic,map image, and vector map data for most areas of the world. For 13 years EDX has been the leader in PC-based wireless planning tools. W e provlde features and performance that are not availahle from any other tool-at any price-on any platform. Contact us today for a fully operational demo CD (not just a slideshow) Youll discover why EDX is the name to know in wreless system planning tools.

Tools for Wireless Design

!EEE Antennas and Propagation Magazine, Vol. 40,No. 6 , December 1998

49