Tom, Dick, and Mary

Discouer the DFT

The story of three students’ discovery of the elationship
between continuous and discrete transforms

~~

J. R. DELLER, JR.

iscrete Fourier techniques are increasingly be-

D ing taught as material detached from fundamen-

tal continuous-time Fourier analysis. The
student is left with an unclear understanding, if any, of
the very significant relationships between magnitude
and phase spectra generated digitally, and the continu-
ous-time signal which is being analyzed. This article
tells the story of three undergraduate students who
discover the DFT, armed only with a knowledge of
analog methods.

Smaller and Smaller Lies

A professor once said to me “Education is the process

of telling smaller and smaller lies.” I wish I could
remember who it was so I could give proper attribution
for that wonderful quote which I have used so many
times in my own teaching. Educators (whether they be
professional or otherwise) do not ordinarily tell ma-
levolent ‘‘lies.’’More accurately put, the state of under-
standing and knowledge of the learner does not always
permit the whole “truth” to be told. By simplifying and
omitting details, the teacher lays a foundation for the
next levels (and there may be several) of under-
standing, the more complex details of which can be
presented when the student is ready to grasp them. The
simplified, alleviated-detail, versions of the subject
sometimes contain ‘‘lies’’of omission or inaccuracy for
the sake of simplicity. The process is recursive; at each
level the “lies” get smaller.
At some point, all of us stop learning about any
given subject. Inevitably, the learning stops before we
achieve the consummate “truth’ about most subjects.
In engineering, this can have unfortunate consequences
for design and innovation. It is particularly unfortunate,
however, when the incomplete version has unnecessarily “This is too much homework,” said Tom. “We’ll never
supplanted the complete understanding; that is, when the get this all done by hand.”
student could have just as easily grasped the complete picture, Dick suggested using one of the personal computers in the
but is left with an incomplete understanding. next room to “at least do the plots.” Mary pointed out that the
Such is the case with the discrete-time Fourier transform plots were continuous curves and that they could at best plot
(DTFT) and the discrete Fourier transform (DFT). These samples of the spectra. She suggested that if they could find
transforms are at the heart of modem spectral analysis, and closely spaced samples, then maybe they could “connect the
since they are almost invariably used to analyze problems dots” to get at least an approximation to the desired spectra.
arising in continuous time, it is of paramount importance that Soon they were staring at the first signal xl(t) and its Fourier
the analyst recognize the relationships between discrete and integral’
continuous spectral techniques. While the historical develop-
ment of discrete Fourier techniques is rich and detailed [1,2,
31 the fundamental roots of the methods lie in the Fourier --m
series, as is clearly brought out in the often-cited paper by
Cooley and Tukey 141. In this sense, the essence of the DFT
was discovered by Fourier in 1807 15, p. 4311, long before the wondering how they would be able to compute samples of Xi
conception of the digital computer. Earlier in this century, using discrete computations.
Fourier’s ideas were brought to bear on the contemporary “We can’t do an integral on the computer even if we just
sampled-data problem by Shannon in the United States [6,7] want values of XI(^, at samples off,” Tom correctly noted.
and Kotel’nikov in the Russian literature [6, 81. However, They thought for a while and almost abandoned the com-
much of the original motivation for the development of dis- puter when Mary recalled a point which ultimately saved the
crete Fourier techniques has been lost in recent tutorial litera- evening. She remarked that the Fourier series (FS) resulted in
ture. a sort of sampled frequency domain. Dick thought that was
The January, 1992 issue oflEEE Signal Processing Maga- irrelevant because “we are not working on FS problems-
zine celebrates the 25th anniversary of the FFT. The year 1967 xl(t) is not a periodic signal,” Dick observed, “so I don’t see
coincides roughly with the genesis of the modem discipline how we can apply the FS.”
we know as “digital signal processing.” Somewhere along the Mary pointed out that only one period of a periodic func-
way, we have lost the early understanding of the DFT. Re- tion is ever used in constructing a FS. That prompted Tom to
spected teachers and textbook writers have found it expedient suggest, “Let’s make xl(t) periodic, and see what we can do
to implicitly treat the DTFT and DFT as entities apart from with it.”
the more fundamental principles upon which they are built. Dick was still skeptical, but he agreed to go along with the
At best, some obtuse attempts may be made to tie the subjects plan. “What period should we give it?” he asked. “I don’t
together using “impulse sampling,” a “lie” from which many know,” admitted Mary, “but let’s try a general period, say Ty,
students never recover. It is assumed that the student, at the and call the periodic signal y(t). Then we’ll have this,” Mary
next level of understanding, will make the fruitful ties back said, writing on the blackboard
to the classical theory which enriches the utility and meaning-
fulness of the discrete transforms. Regrettably, this next level
of understanding may not be realized for many practitioners,
resulting in a career of spectral analysis based on an incom-
plete understanding. In my own teaching, I feel uncomfort-
able leaving the students with the “lie” that the DTFT and
D l T are something different from what they already know She then sketched the signalxl(t) (Fig. la), noting that it has
about continuous transforms. Fortunately, the following se- a “time width” Ti = 8. She also sketched two cases of y(t):
ries of events occurred in a class I was recently teaching. If Case 1 in which Ty > T I (Fig lb); and Case 2 in which
you doubt the veracity of this story, please reread the opening Ty I T I (Fig. IC).
sentence of this article. “I’m not sure where this is going,” said Tom, “but let’s
write the FS for y(t).” He wrote

Too Much Homework m

(3)

Tom, Dick, and Mary were juniors taking a course in signal m=-m
and linear system analysis. They were just completing the ’ By using this definition of the Fourier transform, rather than the similar one
part of the course covering Fourier analysis of continuous m

signals, and were working together on an assignment with based on radian frequency, XI(O) = fe-’“ dr ,we avoid some scale factors
a seemingly countless number of problems involving the -
in the following discussion. The developments remain essentially unchanged
plotting of magnitude and phase spectra based on the
if the radian-frequency definition is used.
Fourier transform.

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 37

 -24 -20 -16 -12 8 4 0 4 8 12 16 20 24
nlw t (-1 Then he realized they were onto
(a) something. “These coefficients are
just the samples of the FT of the origi-
nal signal (with a scale factor),” he
1
said as he wrote “-Xl(mL,)” at the
T,
end of the last line.
“It looks like we’ve gotten the
samples in the frequency domain that
we want,” observed Mary. “Let’s
summarize. If we take samples of the
Fourier transform Xl(mf,) with fy =
l/TJ, T! > T I ,then scale them by& =
1 /T,.,
-24 -20 -16 -12 8 4 0 4 6 12 16 20 24
nm,t (-1
(b)

we have the FS coefficients for a pe-

riodic version of xl(t) whose “copies”
do not overlap. We have the samples
of the FT that we want, and appar-
ently haven’t lost any information
since we can recover xi(t) from the
samples by constructing the periodic
waveform using Eq. 3 , then taking the
‘primary’ period.”
“What would happen if we tried
the same trick with frequency sam-
ples taken at f , = 1/T, with T,, 5 TI?”
-24 -20 -16 -12 -8 4 0 4 6 12 16 20 24 Tom wondered. “I’ll bet those sam-
Tlm, t (we)
ples are still FS coefficients, but for
(4 the ‘overlapped’ version of y ( t ) in
Case 2.” After playing with the prob-
I
1 . ( a )The signal x l ( t ) . The “time width” is Ti = 8 s. (6) The signal y ( t ) when the period Ty = lem for a while, the group confirmed
9 is chosen greater than T I .(c) The signal y(r) when TJ,= 6 I Ti. that this was the case. They knew that
a periodic function did not have a
valid FT, but that one could be con-
Ty12 (4) structed using impulse functions. They had been wamed
a,= y(t) e-J2” mh dt ‘ about the mathematical hazards of impulse functions, but, for
-TJ2 lack of anything better to do, they finally resorted to taking
the FT of the signal they had constructed using the FS. Since
“Hey, since we’re only using one period,” Dick noticed, they weren’t sure that this manufactured signal were truly y(t),
“in Case 1 we can replace y ( t ) by x,(t) in the coefficient they called it y’(t), and wrote
computation,” and he wrote

38 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

 Now using the fact that
m

c
6 (f- mfq’) is the FT of the time
m = -m
m

signal T y c
6 (t- mT,) [9, p. 7231,
m = --m
and the property that multiplying two
l T s corresponds to convolution of
their time functions, they wrote,

-f* 0 f,

“OK!” exclaimed Mary, “Now we

know that we can always obtain at
least a periodic version of xl(t) using
scaled samples of X / ( f ) as FS coeffi-
cients. If we want one period of y(t)
to be exactlyx,(t), we must make sure
that the frequencies of the samples of
X d j ) at mf4 = m/T, are close enough
together, T, > TI.” 41, -31. -21, 4, 0 I.. 1.5f. 21, 31. 41,

Dick remarked that they were still

a long way from being able to plot
spectra on the computer. “To work
with these spectra on the computer,
we must get rid of the continuous
2.(a)A hypothetical FT, X i ( f l . for discussion purposes. Note that XI@)is generally a com-
signals in both time and frequency plex function off, which can therefore only be correctly drawn in three dimensions. How-
domains,” he said. “At least we know ever, since xl(t) is real and even, the FT XI@) happens to be real and even in this case. (b)
what we can do with frequency sam- The periodic FT Y@)when f s > 2f b. (c) The periodic FT Y@) whenf, I 2fb.
ples,” he conceded.
“We talked a lot about time-fre-
quency duality in class,” recalled Tom, “and it seems that we ‘‘Yeah,” agreed Mary, “1’11 bet there’s a general principle
should be able to ‘reverse’ the process we’vejust used in order here: If we’re willing to make one domain periodic, all the
to get samples in the time domain.” information will be contained in samples of the other. But, the

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 39

samples in one domain have to be
close enough to keep the periods from
overlapping in the other (if that’s pos-
sible). Let’s try Tom’s idea.”
Since they didn’t know what X,y)
looked like, Mary sketched a hypo-
thetical version of X , ( f ) on the board
(Fig. 2), carefully labeling its one-
sided bandwidth asb. She then drew
two periodic functions, both with pe-
riod&, and both labeled Yo.In Case 0 0.125 025 0.75
1 , the period f r > 2fh, and in Case 2 , F W W W f W)
fc=<2fh.
I
The group paused for a moment to I . Magnitude spectrum /Xi(f)/ for- the signal .xl(ti, plotted at the frequencies f = kl32 on
consider an important detail. Tom ob- o <f< 1/2. Since the signal is real and even, Xl(f) = /Xi(f)/. The numerical label at f =
jected to the sketches on the grounds 0.3125 will be sign8cant in a future discussion.
that X l ( f ) is a complex function which
could not be drawn in two dimensions. However, Mary cor-
rectly pointed out that the signal x,(t) is real and even in this
case, implying a real and even FT.“We’re just lucky here,”
Mary noted, “but 1guess in general we have to be careful with
such pictures because we are really adding complex num-
They noted that this “FT” is very close to the time signal,
bers.”
“I guess we could still get the idea of frequency overlap ca

using these diagrams,’‘ Dick observed. Dick was correct, but

the magnitude spectra do not simply add directly as we might
-m
infer from these pictures, I told them the next day as we
discussed this point in my office. Textbooks do not always
make this point clear, and we must be careful with such but agreed to ignore this fact temporarily in order to preserve the
sketches. analogy to the frequency-sampling case. They carefully wrote
The other detail that they neglected is the fact that a signal down the “FS” for the periodic function (Eq. lo), lettingfplay
cannot be both time-limited, as in the case with xl(t), and the customary role of “time,” and fs play the role of the funda-
frequency bandlimited, as implied by Mary’s sketch of X,y) mental “period.” Accordingly, l#s = Ts is the analogous quantity
[9, p. 1541. This negligence tumed out to be beneficial, for, to fundamental “frequency.” In these terms, the “FS” is
had they remembered this fact, they might have been discour-
aged from making X , ( f ) periodic. We discussed this point at
some length the following day.
The students began to scrutinize the cases of Figs. 2b and
2c. In either case, they speculated that the periodic function

could be represented by a “FS” whose coefficients were The “FS coefficients are scaled samples of the (reversed time)
related to samples of time. It was simply a matter of reversing waveform x’i(t) = ~ i ( - - t ) so
the roles of time and frequency in the analysis. Dick guessed
correctly that the time samples would be xl(nTs), with Ts = m

I & by making an analogy to what happened in the dual

problem. n=-m
Following the analysis in the frequency-sampling prob-
lem, the group quickly discovered that their intuition was
correct. The periodic (and complex) frequency function Y ( f ) The group decided they didn’t like using the samples in
could be represented by a “FS.” They first computed the “FT” reversed time, so they made the simple change of replacing
for the “signal” Xl(f), calling it x’l(~), -n by t i , to obtain

40 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

 m m complete continuous-time signal like x i ( t ) is recoverable
from the time samples.xi(nTs). In theory, the method for doing
so is obvious from your work: Form the “FS” for Y(f) as in
Eq. 16. The first period of Y ( f ) is the FT X i ( f ) , so we can
“We’ve got it!” Mary said excitedly. “As long as we take inverse FT this initial period to recover the signal xl(t). In
fs = UT, > 2fh, then the first period of Y ( f ) will be exactly practice, we apply a lowpass filter to recover the signal
XIV).’’ occupying the primary frequency band.
“But we need samples of Xl(f),” Dick reminder her, “in “This lowpass filtering operation can be shown to be a
order to plot them and work with them on the computer.” discrete convolution operation involving the samples of the
“No problem,” said Tom, “we simply calculate Yo
=X,(f), signal as inputs to the filter. The process is usually called
at any frequency we want. For example, take the frequency interpolation, since it allows us to interpolate between sample
fo,” and he wrote on the board values in the signal.
“As long as you sample xl(t) fast enough, the FS Yo is an
m (17) excellent candidate for spectral analysis of the original signal,
since it represents an exactly periodic version of X / ( f ) . But
n =-m even when samples are not taken fast enough (or cannot be
because of infinite bandwidth), Yy> is closely related to X , ( f )
“This is valid for any frequency sample -fh Sf0 <fh since in that it is a periodic, but aliased, version of X , ( f ) .The term
that’s the range over which Yo= X/(f).” aliasing refers to the overlap in the frequency spectrum which
Using sampled values of the given signals, and the expres- causes frequencies abovefsI2 in the signal to be confused with
sion they had derived in Eq. 17, Tom, Dick, and Mary were those in the band 0 If <f,i2.”
able to computer generate the plots for
each of their homework problems.
Not knowing the bandwidth, fb, for
any of the signals, they had to guess

\
the sample period T, to use in each
case. They arbitrarily chose T, = 16=
1 s for all signals. In some cases, this
was satisfactory; in others, quite un-
satisfactory. The magnitude spectrum
1
: for result for signal x/(t) is shown in
Fig. 3. (Note that the phase is either
zero or pi radians at every sample,
:
0 0 0.3125 0.625 0.9375 1.25 1.5825 1.875 2.1875 2.5

since the signal x/(t) is real and even, Fnqumcy, f (Hz)

resulting in a real, even lT. ) 3
Figure
was obtained using Eq. 17 at the fre-
quency samples VJ32 = k/32 for k =
0, 1, ...16. Only the positive frequen-
cies are plotted, since the spectrum is
0 0.3125 0.025 0.8375 125 1.5625 1.875 2.1875 2.5
an even function of frequency. la0
When we met in my office the next
day, Tom, Dick, and Mary were justi- to
fiably excited about their discovery.

Gold Stars: Part I F 1

I commended them for their innova-
tion, and took the opportunity to ex-
plain just how profound their work
was. “First of all,” I told them, “in the
i :o:

0.Wl

time sampling case, what you have

discovered is the essence of a funda- 0.m1

mental result in information theory

known as Shannon’s Sampling Theo-
rem [6,7]. This principle says that if
we sample a time waveform at a high- . “True”magnitude spectrum for the signal x l ( t ) . ( a )Magnitude plotted on a linear scale.
( h )Log scale.
enough rate,fs = l/Ts > 2fbr then a

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 41

 FT). We studied two signals in their
homework in some detail to under-
stand this point. First, I generated a
plot representing a very good ap-

’wl proximation to the true continuous-

time magnitude spectrum of signal
x , ( t ) (Fig. 4a) for the sake of discus-
sion. “You were lucky in this case,” I
told them, “since there is energy in
this signal beyond your assumed
bandwidth of&/2 = 1/2. This is more
evident on the log scale (Fig. 4b).
However, the amount of energy is
negligible compared to that in the
range 0 S f < 1/2, so there is little dis-
tortion of the spectrum on the range
you plotted. (Imagine what would
happen if Fig. 4a or Fig. 4b were made
periodic with periodf, = 1 as above.)
However, there is some aliasing as we
I see, for example, by comparing your
sampleatthefrequencyf=0.3125,with
the same frequency on the ‘true’ spec-
trum. We see that, on your plot,
K1(0.3125)l = 0.723, whereas on the
‘true’ spectrum, lX1(0.3125)I=0.520.”
So we concluded that we will not
always be able to “guess” a reason-
able sample rate. The most flagrant
example of aliasing occurred in the
next problem which involved the sig-
nal

loo 1 x2(0 =
e4.03‘ sin(%) [u(t- 1) - u(t - 93.1)]

(18)
with u(t), the unit step function, de-
fined to be unity for t 2 0 and zero
otherwise. Based on our previous dis-
0.01 4 cussion, the group now understood
0 0.5 1 1.5 2 25
that a sample rate offs = 1 was inap-
propriate since the energy in this
damped sinusoid is concentrated
. ( a )Alleged magnitude spectrum for signal x2(t) on the range 0 2 f I 1t2 obtained by using a r o u n d the frequency
Eq. 16 with T,$= I at frequency samples f = kl128, k = 0, I,..., 64. (h)Extending the compu- fo = 5/2n = 0.796Hz.“A s a m p l i n g ‘
tations of part (a)to include the range -1 2 f < 1 =fs. (c) The true spectrum for signal x2(t). rate of f s = 2 (0.796) = 1.6 Hz is
needed to prevent ‘overlap’ in the pe-
riodic FT,”Tom noted correctly.
We digressed momentarily to discuss the phenomenon of
“What happened in this case is very revealing,” I told
aliasing. The students pointed out that they didn’t know what
them. “Look at your spectral magnitude plot (Fig. 5a). It
sample rate to use for the various signals in their homework.
to indicate energy atf=o.2 Hz, but to show
Even if they had been aware of the sampling theorem, they
had not known the bandwidths of the signals, since the task energy at the frequency of the sinusoid (which is not even in
of the problems was to plot the spectra. By using an arbitrary the range of Your Plot). What happened?’’
sample rate of& = 1 Hz, they had sometimes gotten a poor I suggested that they extend their plot to include one or
result (in terms of the approximation to the continuous-time more periods of the “overlapped” FT. Using the program

42 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

 1
0.9
0.8
0.7

B :::0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30
Sampt., n (Actualtlma In sac,t 8 n X 0.001)
6. The N = 15-length discrete-timesignalx3(nTS),n = 0 ,I ,...,N - 1 , with Ts = I tns, used as an example in the development of the DFT.

they had written the night before, they did so (Fig. 5b), “So the DTFT is very important because it allows us to
knowing that the result would be periodic with periodf, = 1. compute a periodic replica of the continuous FT using just
While they were doing that, I plotted the true spectrum of time samples,” Dick noted.
x z ( t ) (Fig. 5c). “As long as we sample fast enough in time,” I added.
“I see what happened,” said Tom. “The bogus peak we got “Now what is the effect of the scale factor omission in the
in the range 0 Ifs 112 is the result of one of the copies of the definition of the DTFT?” I asked.
true spectrum that got shifted into that range.” They identified “No big deal,” said Tom, “it just means that the DTFT is
the “correct” peak in the periodic spectrum and convinced a scaled periodic replica of the original FT.” Looking back at
themselves that this problem would not have occurred if a his notes from the night before (Eq. lo), he wrote
proper sampling rate had been used.
“Now let me tell you a further significance of your result,”
I continued. “The ‘FS’ Yoyou’ve discovered is very close to
what is known as the discrete-time Fourier transform (DTFT)
for the discrete time signal consisting of the samples
xl(n7‘), n = ...- 1, 0, 1, 2, . ... Usually, books will tell you that
“So is the DTFT something new?” I asked them. “Or is it
the DTFT is the something like the following.” I wrote on the a natural extension of classical theory?” The question was
board rhetorical, and they were very proud of their accomplishment.
“When you see the DTFT in your work with discrete-time
signals, you will forever remember its origins since you
worked so hard to discover them.”
“Now I have a challenge for you,” I said after they had
congratulated themselves. Tom, Dick, and Mary looked
“That’s just our ‘FS’ except for a scale factor 16,” Mary astonished since they had been awake most of the night facing
said. “Yes,” I agreed, “which happened to be unity in your what they considered to be a more than sufficient challenge.
numerical work anyway. So can you tell me the significance “In the first part of your work, you show how to use a set of
of the DTFT? Think about what we’ve just worked through.” frequency samples to represent a continuous-time waveform.
“Well, we wanted to be able to plot spectra using the In the second part, you show how to use time samples to
computer, so we had to have discrete samples in both do- represent a continuous-frequency function. (Then you use
mains,” Tom said. “Yes,” I interjected, “and more generally, that result to compute frequency samples. To use a computer
computers are being used to replace all kinds of analog for spectral analysis, it would be useful to have a transform
systems and filters so that the computations, and in this case, which goes back and forth between sets of samples. You are
spectral analysis, must be done with discrete samples. Usu- close. See if you can find such a transform.”
ally, however, the samples relate to continuous phenomena “Let me give you a hint,” I added. “You notice that your
since most signals are continua. So we’d like to be able to DTFT computationuses a doubly-infinite number of time
perform ‘meaningful’ spectral analysis using the computer.” samples.”
“You mean we want to be able to relate the results to the “Yeah,” Tom agreed, “we had to chop off the signal after
continuous world, since that’s where they really come from,” a large number of samples if the original signal was not
Mary remarked. ‘time-limited’. I mean if T I = -. So I guess we only got an
“Precisely.” approximate spectrum.”

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 43

 “Yes, and that will always be the case in discrete-time so x3(t) is nonzero on 0 I t < 15.1 ms.
analysis,” I pointed out. “You can never enter an infinite Suppose T, is the sampling period you choose and that you
number of time samples. There are smart ways to t”ite get N samples on the nonzero range,
long signals to get better estimates of the spectra, but that’s x3(0),x 3 ( ~ x~3)( ,2 ~ , )...,
, X ~ ( N- 1 ) ~ These
~ . samples comprise
another story [lo, 111. Let’sjust assume that we have
which starts at time t = 0 and ‘ends’ at some finite time. An
the nonzero part of the discrete-time signal . ~ ( n n.’’ 1
sketched the samples on board for T, = 1 ms (Fig. 6).
example is xz(t) in your homework, but the next one, x d t ) , is X 3 ( t ) bandlimited to finite f h , ) ) they wanted to
simpler: know. “and, if SO, is ”& > 2fh?”
-

“xj(t) can’t be bandlimited because it is of finite time

[u(t)- U ( ? - 0.015
x3(t)= (0.9)-loaot 11 (21) duration. However, it can be assumed to be practically ban-
dlimited to some finite fh, either inherently or because of a
pre-processing lowpass filtering op-
eration. For your theoretical work,
1 assume that there is practically an
0.0 fh < m but that you don’t know it’s
0.8 value,” I responded. “This means that
0.7 you have to be prepared for some
0.1
aliasing since you don’t know the

1: 03
proper value offs.”
“I appreciate what you did last
night! Now see if you can take it to
the next logical step.”
02 I’d like to tell you that Tom, Dick,
0.1 and Mary excitedly left my office,
0 eager to tackle this new challenge.
0 However, this is a true story. Some-
thing unfit for print was muttered in
the hallway as they walked away.
1
0.9
0.8 it Onward to the DFT

1:
0.7
The three students puzzled over Eqns.
7 and 9 for a while thinking maybe
they could just use a similar equation
0.3
to Eq. 7 at times t = nTs to get the time
02
0.1
samples they wanted. “But we don’t
0
really have X 3 v ) to sample,” Tom
0 5 10 15 20 25 30 36 4 0 4 5 6 0 5 5 pointed out. “What we have is a peri-
odic version of X 3 v ) that we again
call Yo. In this case, since we have
only N time samples, Eq. 16 can be
1.4 1 simplified.” He wrote
12

1
N- 1 (22)

j: “~
fl=O

0.4
“Of course, we can compute any
02 samples off that we want, but which
0 ones?” Dick wondered. They pon-
0 5 10 15 dered this for a long time. Mary reit-
erated the point that the signal xj(t)
could not be frequency bandlimited
because it is time-limited. Neverthe-
7. The signal y(nT,) of Eq. 24 when (a)fu = llNTs =fa;
(bjf? < llNTs; ( c j f , 2 I I N T S . Each
less, she suggested that they go back
waveform repeats indefinitely in both positive and negative time.
and look at what happened in their

44 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

first analysis in which X / ( f ) was as-
sumed (incorrectly) to be frequency-
limited tofh. “Look at Eqs. 7 and 9,”
she instructed her classmates. “Zf we
had samples of XjV, to work with, and
ifwe took these samples at spacingf,.
= l/NT, =&IN, then Eq. 7 would pro-
duce, according to Eq. 3:”

m (23)
y ( t )= C x3(t - kNTs)
k=- (R 0 fJ2
“fW)
“And, if we used Eq. 7 at times nTs, 8. Hypothetical FT Y3(3 defined to be identical to Y ( f ,of Fig. 2c on its primary p e r i o d .
then we would get the numbers:”
m
(24) the original signal, x~(t)had been bandlimited and sampled
y(nTs) = x3(nTs - kNT,) fast enough.
k=-m “Right on both counts,” said Tom, “but Mary is consider-
ing the general case in which Y v ) and YjV, might involve
They sketched the sequence y(nT,) on the blackboard (Fig. some aliasing. I also see what she’s trying to do by making
7a) and discovered that the first N points were exactly the y.&) a continuous-time signal. She’s taking us back to the very
sequence xj(nT,). In fact, they varied the value off, = 1/T, and first problem we studied last night- continuous-time, dis-
discovered that using a smaller value off, = 1/NT, would crete-frequency .”
cause “extra space” (zeros) between the copies of x/(t) in y ( t ) “I see where we’re going,” responded Dick, “but if we
(Fig. 7b), and “overlap” (time aliasing) of the copies would follow the approach we did last night, we’ll end up writing a
occur iff, were chosen too large,& > 1/T, (Fig. 7c). FS for a periodic version of yj(r), not y(t). Then we’ll evaluate
“But we don’t have the samples of X j ( f ) to work with,” at discrete times, and end up with samples of a periodic
noted Dick. “We have only a periodic, maybe aliased, version version of yj(t), not samples of y(t), and certainly not the
of Xjv).” samples xj(nT,).”
“I have a hunch that samples of Y v ) taken at the right “Go ahead and do it anyway,” Tom persisted. ‘‘I think it
will work.” So they began to meticulously follow the steps
spacing f , S J i N will still give us back the periodic sequence
that had become familiar by now. First they created the
(Eq. 24),” Tom said. ”Something like what happened when
periodic extension of the somewhat mysterious signal yj(r):
we sampled the aliased periodic FT last night (Eqns. 7-9),” he
added with some uncertainty in his voice. m
They knew that this case had to have something to do with
the first analysis they had done the night before concerning w(t)= y3(t - iT,])
frequency sampling. They began to pore over those results. I = -m
“Look at the pattern we followed there,” said Mary. “We had
a bandlimited signalxl(t) and we let it become periodic so that anticipating that they would sample Y3(f) at spacing fw =
we could discretize the frequency domain. I think we want to UTw. They all agreed that T w would have to be some multiple
do the same thing here. Forget the fact that Y(f) is periodic. of Ts because the period would ultimately have to be some
That’s what’s throwing us off. Just look at one period of it,” integer number of samples, say,
and she wrote on the blackboard the following definition:
1
T w = M T ~ ,or f w = -
MTs

IO, otherwise Then they wrote the FS:

m
and drew the sketch in Fig. 8. “Of course, we’ll let y j ( t ) be
the corresponding time signal. I’m not sure what y3(t) is, but
I know we will make it periodic if we build a FS out of k-m

samples of Yj(f),and, if Tom’s hunch is right, the signal y j ( t ) T,J2

will match y(t) at the sample times nTs,” Mary deduced.
“By chopping off Y v ) to create Yjv),” Dick pointed out,
“we force yj(t) to be a continuous-time signal.” He also
observed that yj(t) would be exactly xj(t), and Y.{(f)= X j ( f ) if They knew from past experience that whether w ( t ) were

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 45

an “overlapped” version of y3(t)or not (see Eqs. 5-9),
They noted that they could replace Y 3 0 by Y O by restricting
1 (30) the summation:
yk = - Y3(kfw)
TnJ (32)
1
w(t)= -
TW
C Y(w%,)eJ2*IFfHr
Therefore, k
m
(31) k such that -2 -< kfnz < fs
-fs
-2

k=-m They now had frequency samples on the right side with
which they could compute time sam-
ples. In particular, they wanted sam-

c
*
‘i 7
ples at times nT,, n = 0, 1, . .. , N-1.
So they evaluated w(t) at the discrete
times,

ii: w(nTs)= -
TW
m

k=-w
“Great,” muttered Dick, sarcasti-
(33)
1 Z Y ( k f w )e j2*kfwnTs

cally, “we can compute samples of

some signal we don’t understand or
want.”
“Hold on,” said Mary. “These are
samples of a periodic version of y&).
We just have to figure out how y3(f) is
related to our original time samples.
“Look,” she said, as she wrote

“Hey, that just looks like a FS coeffi-

cient computation for the periodic
Yo,’’ exclaimed Tom.
Mary was busy rifling back
through her notes from the night be-
fore. “Yeah, we’ve seen that before!”
After a brief search she rediscovered
Eq. 14 and wrote, triumphantly,

P-~I= x3(nT.,)
.j A,
y3(nT.,) =fs (35)
then,
- .
0

Id
9. Samples of the (a)magnitude, and ( b )phase, spectra of the discrete-time signal x3(nTs) of
Recalling their agreement (Eq. 27),
Fig. 6 obtained using Eq. 39. (c)Magnitude spectrum of the continuous-time signal x3(t) of
Eq. 21.
she finally wrote

46 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

 m
(37) the corresponding magnitudes and phases shown in Figs. 9a
w(nTs)= x3(nTs - &ITs) and 9b, respectively. By hand, they plotted the true magnitude
2z-x
spectrum (shown in part (c) of the same figure) to see how
closely their samples matched the true spectrum.
“Voila! w(nT,)is just a periodic, possibly aliased version
of x3(nT,),” Tom said excitedly. “As long as we take M 2 N, Gold Stars: Part I1
then the samples MO), w(T,),...,w((N-1)Z’Jwill be the time
samples we are looking for!” Referring back to Eqs. 32 and With understandable pride, the students brought me Eqs. 38
27 with M = N,he wrote and 39 and their plots the next day. I applauded their efforts
and suggested we make a few minor adjustments.
“Your Eq. 39 is used to compute samples of Suppose Yo.
we wanted to compute samples of the more conventionally
used ‘DTFT’ that we discussed yesterday. What would that
look like?”
They reviewed our discussion of the DTFT and realized
that only a scale factor change was necessary. Since
Y D T d f =fsY(f),
) we wrote

n = 0, 1, ..., N-1

“That gives us the N time samples in terms of N frequency

samples,” said Dick. “Now we need to get the N frequency
samples using the N time samples. I think we can use the result The inverse relation also had to account for the scale factor.
we used for the plots in our homework.”
They dug back through their notes and found Eq. 16.
Noting that x,(nT,) is zero except for the first N points, they
wrote
(39)

“You have also done the logical thing of sampling the FT

N- 1 (or now DTFT) over the ‘primary band’. That is, you took the
=-
1
C x 3 ( n ~e-j2nn
~)
kiN
,
-fi kfs
uchthat-<-<-
fi
fs 2 - N 2 -. < f < f-.s The
appropriate samples on the frequency range -
fl=O 2 - 2
problem with this is that the values of k (when written in
They computed the 15 samples in frequency using Eq. 39 and general terms) depend on whether N is odd or even. When N

2 1 t t
1 -
0
0 1
1
2
T
3
?
4
?
5
T
6
?
7
?
8
?
9
? ? ? T
1 0 1 1 1 2 1 3 1 4 1 5
1
Frequency sample, k (Actual frequency In H q f = lo00 W16)

1
10. Magnitude spectrum of the 16-point DFT of the sequence x3(nTs)

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 47

 N- 1 N- 1 never use the DFT to study the continuous spectrum?”
is odd, then the values are k = - -, ..., 0, .. . -; 2 but when “ N ~ I, guess you can always sample fast enough to
2
N N the aliasing insignificant,” said Mary.
N is even, k = - -, ..., 0, ...,- - 1.” “Yes, or more likely, you will filter out all energy above
2 2
fs/2 and just resign yourself to studying the ‘lower’ frequen-
“That we have to have different formulas for cies,” I added. “OK, so let’s assume that one period of the
the frequency samples,” observed Tom. DTFT is a very good approximation to X3(f).”
“Right. Unless we can find a way around the problem. ‘‘Then the DFT frequency samples will let us study the FT
matsignal processing people usually do is exploit the fact of the continuous-time signal using just additions and multi-
6 plications, no integrals!” said Tom. “And we can go back and
that the numbers Ymm [z)e’2’krJ’Nare periodic with period
I .
forth between the frequency and time samples with the rela-
N. (Remember that the\DdFT has period fr, which corre- tions we derived.”
sponds to N samples.) Let’s show this: “Yes, and you did that all with FS theory, which is in the
spirit of the original developments of the transform,” I re-
minded them. “So when you see the DFT in your later work,
you will know that it is not a ‘separate’ transform, but that it
has real meaning for analyzing continuous-time signals,
which, after all, is what we are usually trying to do.”
and
“Let’s look at another important point which is inherent in
your work,” I encouraged them. “Assume that there is negli-
gible aliasing so that Fig. 9c on 0 If I f J 2 = 500Hz repre-
sents the first half period of the DTFT except for a scale factor

Y D T-
r):
What this means is that, in either case, instead of computing
~ for negative values of k, we can compute over
nonnegative k values only. In either case, N odd or even, the
values of k used are k = 0, 1,2, ..., N-1.” The students played
$,. Your samples on this range k = 0, 1, ..., 8 in either Fig. 9a
or 10 compare favorably with the continuous spectrum. That
is, they appear to be proper samples of the DTFT spectrum.”
“But we don’t see much of the detail in the spectrum,” Tom
pointed out. “Yes, that’s exactly what I wanted you to notice,”
with a few values of N to convince themselves that this was I said. “Is there a remedy for this problem? Remember how
true. “Therefore, we can use Eq. 40 for these nonnegative k’s you sampled the FT at arbitrary frequencies in doing your
and modify Eq. 41 to read:” homework plots (Eq. 17)?
“I think the same basic idea applies here,” said Mary. “If
we rewrite the D F I relation (Eq. 40) as

(45)
“The pair of relations (Eq. 40, used fork = 0, 1 , 2,..., N-1;
and Eq. U ) ,are the essence of the discrete Fourier transform k n=O
(DFT),” I elaborated. There is a group of efficient algorithms
for computing the DFT relations that take advantage of the it is obvious that the kth result is the sample of the DTFT at
symmetry properties of the quantities involved. These algo- frequency f = k fslN. I guess we could compute these samples
rithms are collectively known as the fast Fourier transform at whatever frequencies we want.”
(FFT) (e.g, see [lo-131). We took a moment to compute the “That’s right,” I responded, “and a common thing to do is
DFT of Eq. 40 and to plot the resulting magnitude spectrum to compute these samples at the frequencies kf,/M for k = 0,
(Fig. 10) and to verify that the inverse, Eq. 44, indeed pro- 1,..., M-1 with M > N . This gives us M equally (and more
T . ~ ) in Fig. 6. The picture
duced the 16 samples of X ~ ( ~ shown closely) spaced samples of the D T R . M is frequently taken
made clear that the DFT are samples of the first period of the to be a power of two, since efficient and widely-available DFT
DTFT on nonnegative frequencies, rather than the “primary (FFT) algorithms exist for this case. In this case, the DFT
period” that straddles f = 0. relation becomes
“Let’s review the significance of your DFT relations,” I
encouraged them. “We will always have to work with a
finite-duration signal on the computer, like x3(t), Dick began.
“We then take N samples. The forward transform 40) (m.
allows us to compute samples of the DTFT.”
“What is the significance of those samples?” I asked.
“Well, the DTFT will have to be an aliased version of the
continuous-time FT, Xi(f),” Mary said. “ x l ( t ) cannot be ban-
dlimited because it is time-limited.’’
“Right,” I agreed. “Do you suppose this means that we can The last equality follows because x3(nTs) is assumed to be

40 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994

zero outside the initial N points. This is equivalent to taking N- 1 (47)
the M-point DFT of the N-point signal x3(nTs) to which M - ~ 3 ( k=) x3(n) e -j2nnklN , k = 0 , 1 , 2 ,..., N-1
N zeros have been appended. You will get better resolution n=O
of the DTFT corresponding to the assumed N-length signal
x3(nTs)using this process. If x3(nTs) is, in fact, longer than N and
points in duration, the only way to get both better resolution
and a better DTFT to sample is to add more real data (not just N- 1 (48)
1
appended zeros).”
We proceeded to plot the 64-point DFT of x3(nTs) by
4 n )=-
N
C X3(k) CJ~’~’’” n=O, 1 , ..., N-1
k=O
appending 48 zeros to the original 16 points. The magnitude
spectrum result is shown in Fig. 11. Better resolution of the
continuous-time spectrum is evident in the result.
“Does Eq. 44 still retum the correct signal x3(nT,) when M
~ , or,
where X3(k) stands for what we have called Y D T -
(t]
is used in place of N?” Dick wanted to know. “Good ques-
tion,” I responded. “The answer is yes, and it is not difficult
to demonstrate by putting Eq. 46 into Eq. 44. So, again, we
since in this casefs = 1, Y D T -
~
(F] . Please note again that
this is not the same thing as simp y taking samples of the
original time waveform at frequencyfs = 1.
have a complete DFT pair with the ability to sample the DTFT The students were pleased with their discovery, and men-
for a finite-duration signal as closely as desirable.” tally and physically exhausted from their labors. They were
“Finally, there’s one more nuance that you need to be packing up their books when I concluded the session.
aware of,” I cautioned the students. “It is customary in digital “I’m very proud of you for the hard work you did,” I told
signal processing to ‘normalize’ the problem so that the them, “and since you seemed to enjoy this challenge so much,
sampling frequency is effectively unity, fs = 1. Another way ...” But before I could finish the sentence, Tom, Dick, and
to look at this is that the time samples are indexed by integers, Mary made a mad dash for the office door and were gone.
for example, x,(O), x I (l), ..., xl(N) so that the sample period,
T,, is effectively normalized to one. This reindexing has the
accompanying effect of normalizing the frequency axis so No More Lies
that samples are taken in the interval 0 S f < 1. Please notice
that this process is not equivalent to sampling the original It is becoming increasingly likely that Tom, Dick, and Mary
signal at a rate& = 1. Rather it amounts to reindexing samples will spend an engineering career designing and analyzing
that have been taken at an appropriate rate. You will explore spectra using only discrete techniques. The days of “pencil
these details more formally in later work in digital signal and paper” integral transforms as a basic design tools are
processing. However, I wanted you to be aware of this fact receding into history. However, the students’ basic notions of
because you will often see the DFT relations written simply spectra, frequency, phase, energy, resonance, and so on, are
as, for example, all based in continuous-time concepts, and the problems they
will work on are inevitably drawn from the continuous world.

t: 3 T
2
I~ 01 0
TTTlTTTTTTTT?TTT??T???T???T~~TTT~TTT?TTTTTlTT~lT?l, 1

I . Magnitude spectrum of the 64-point DFT of the sequence x3(nTs).

APRIL 1994 IEEE SIGNAL PROCESSING MAGAZINE 4!

Henceforth, Tom, Dick, and Mary will benefit from the agony ing, New York IEEE Press, 1972.
I put them through. They now understand the notion of 5. W.E. Boyce and R.C. DiPrima, Elementary Differential Equations and-
Boundary Value Problems (2nd edition), New York: Wiley, 1969.
sampling as one of constructing Fourier series for periodic
extensions of the “signal” in the other domain. This will allow 6. A.J. Jem, “The Shannon sampling theorem- Its various extensions
andapplications: A tutorial review,” Proc. IEEE, 65, pp. 1565-1596, Nov.
them to meaningfully interpret their results in light of their 1977.
understanding of continuous-time concepts. They know the
7. C.E. Shannon, “Communication in the presence of noise,” Proc. IRE,37,
truth. pp. 10-21, Jan. 1949.
8. V.A. Kotel’nikov, “On the transmission capacity of ‘ether’ and wire in
J . R . Deller, Jr. is a professor in the Department of Elec- electrocommunications,”Izd. Red. Upr. RKKA, Moscow, 1933.
trical Engineering, Michigan State University, East Lansing, 9. G. E. Carlson, Signal and Linear System Analysis, Boston: Houghton-
MI. Mifflin, 1992.
10. A.V. Oppenheim and R.W. Schafer, Discrete-Time Signal Processing,
Englewood Cliffs, New Jersey: Prentice-Hall, 1989.
References 11. J.G. Proakis and D.G. Manolakis, Digital Signal Processing: Principles,
Algorithms, and Applications, (2nd edition), New York: Macmillan, 1992.
1. J.W. Cooley, “How the gained acceptance,” IEEE Signal Processing
12. E.O. Brigham, The Fast Fourier Transform and Its Application, Engle-
Magaiine, pp. 10-13, Jan. 1992. wood Cliffs, New Jersey: Rentice-Hall, 1988.
2. J.W. Cooley, P.A.W. Lewis, and P.D. Welch, “Historical notes on the fast
13. G.D. Berglund, “A guided tour of the fast Fourier transform,” IEEE
Fourier transform,” IEEE Trans. Audio and Electroacoust., AU-15,pp.
Spect”,6,pp.41-51, July 1969.Reprintedin: L.R. RabinerandC.M.Rader
76-79, June 1967. Reprinted in: L.R. Rabiner and C.M. Rader (editors), (editors), Digital Signal Processing, New York IEEE Press, 1972.
Digital Signal Processing, New York: IEEE Press, 1972.
3. J.W. Cooley, P.A.W. Lewis, and P.D. Welch, “The finite Fourier trans-
form,” IEEE Trans. Audio and Electroacoust., AU-17, p p 77-85, June 1969.
Reprinted in: L.R. Rabiner and C.M. Rader (editors), Dlgrtal Signal Process-
ing, New York: IEEE Press, 1972.
Editor’s Note
4.J.W. Cooley and J.W. Tukey, “An algorithm for the machine calculation
of complex Fourier series,” Math. of Computation, 19, pp. 297-301, 1965. This article has been formally reviewed by four referees
Reprinted in: L.R. Rabiner and C.M. Rader (editors), Digital Signal Process- according to customary IEEE procedures.

Introducing the 1994

IEEE All-Society Periodicals Package
The core collection of quality journals and magazines in
electrical engineering, electronics and computing...
...subscuibe today and get over one third of them @e!
I n electrical engineering, a “core” is a mass of material that directs and increases
the strength of a field. The 89 journals and magazines in IEEEs All-soCiety Periodicals
Package virtually define the fields of electrical engineering, electronics, and computing.
For a free brochure or to order call: 908-982-0060 or fax: 908-982-9667

The Institute of Electrical and Electronics Engineers, Inc.

445 Hoes Lane, P.O. Box 1331, Piscataway, NJ 08855-1331,U S A .

50 IEEE SIGNAL PROCESSING MAGAZINE APRIL 1994