5 Discrete processing of analog signals

5.1 Introduction

The signals we have studied so far in the previous chapters are classified into two cate-
gories: continuous-time signals and discrete-time signals. As mentioned before, most
of the signals generated from physical phenomena are of continuous-time and they
used to be processed by analog/continuous-time systems before the digital ages came.
As a consequence of the dramatic development of digital technology over the past
few decades, which results in the availability of low-cost, lightweight, programmable,
and reproducible discrete-time systems, processing of discrete-time signals has taken
place of processing of continuous-time signals in many contexts, leading to the so-
called discrete processing of continuous-time signals.
Processing a continuous-time signal with a discrete-time system has several mer-
its that result from the power, flexibility, and reliability of discrete-time computing de-
vices such as micro-processors and specialized digital devices (e.g. digital controllers
and digital signal processors):
– a signal manipulation is much more easily carried out with arithmetic operations
of a digital device than the use of an analog system;
– implementation of a discrete-time system only involves writing/modifying a set of
program codes;
– digital signals are more robust against disturbances as there are a limited number
of signal levels involved in the systems.

The first thing of first to do for discrete-time processing of a continuous-time signal x(t)
is to convert the latter into discrete-time one, say x[n], and this is done by an operation
named sampling that leads to the following basic relation: x[n] = x(nTs ), where Ts
is the sampling period. The 2nd step for the processing is to manipulate x[n] using a
discrete-time system and the last step is to convert the output of the system back to
the continuous-time form.
It should be noted that processing of x[n] will not enhance the amount of infor-
mation on x(t). So, a number of very essential questions we ask on such a processing
include the following.
– How much information of the original signal x(t) is lost by sampling?
– Is it possible to recover x(t) from its samples x[n]?
– If the answer to the 2nd one is positive, how to reconstruct x(t) using x[n]?

An outline of this chapter is given as follows. Section 5.2 is devoted to discussing the
concept of sampling of continuous-time signals. The spectral relation is derived in
Section 5.3, which represents the DTFT of the discrete-time signal (i.e. the sampled
version of a continuous-time signal) in terms of the FT of this continuous-time signal.
160 | 5 Discrete processing of analog signals

Based on this relation, the famous sampling theorem is embodied in terms of sampling
frequency and the maximum frequency that the spectrum of the continuous-time sig-
nals is limited to. The issue on reconstruction of a continuous-time signal is studied
in Section 5.4. A hybrid system for discrete processing of continuous-time signals is
presented in Section 5.5, where the effects of under-sampling and nonideal reconstruc-
tion functions are discussed. In Section 5.6, the issue of frequency-domain sampling is
raised and leads to the important topic—discrete Fourier transform. The curtain comes
down on this chapter with the chapter summary that is given in Section 5.8.

5.2 Sampling of a continuous-time signal

To take the advantages of powerful digital computing devices, signals of continuous

variables have to be discretized. This procedure is usually referred to as sampling. Take
a time-domain signal x(t) as an example. Sampling is executed with a system whose
function is mathematically described with
x(t) → x[n] = x(nTs ). (5.1)
Physically, a sampling can be implemented with an electronic switch (see
Figure 5.1).

x(t) x[n] = x(nTs )

k
Q
Q 
 Q
t = nTs

Fig. 5.1: Implementation of sampling operation in time domain

Suppose x[n] is obtained from a continuous-time signal x(t) with sampling. In order to
keep all the information of x(t), the mapping between x(t) and its samples x[n] should
be one-to-one. Figure 5.2 show the situation where a given sequence x[n] corresponds
to three continuous-time signals.
This is definitely a unwanted situation in many contexts as the original signal x(t)
cannot be uniquely recovered from its samples x[n]. Under what conditions x(t) can
be uniquely determined by its samples x[n]? And how serious the situation is if these
conditions are not satisfied? In what follows, we will answer these two questions.

5.3 Spectral relationship and sampling theorem

Let x[n] be the discrete-time signal obtained by sampling the continuous-time signal
x(t). As the FT of x(t) and the DTFT of x[n] are discussed at the same time, in order
 5.3 Spectral relationship and sampling theorem | 161

2
x(t)

−2
0 0.2 0.4 0.6 0.8 1
(a) t

2
x1(t)

−2
0 0.2 0.4 0.6 0.8 1
(b) t

2
x2(t)

−2
0 0.2 0.4 0.6 0.8 1
(c) t

Fig. 5.2: Three continuous-time signals yielding the same x[n] when sampled with the same pe-
riod Ts .

̃
to avoid any confusion we use X(jω ), rather than X(jω ), to denote the FT of x(t) and
jΩ
X(e ) for the DTFT of x[n].
Note that the mapping between x(t) and x[n] is one-to-one if and only if so is that
̃ and X(.). So, let us look at the relationship between the
between the two spectra X(.)
̃
two functions X(.) and X(.):
+∞

x(t) ↔
FT
̃
X(jω ) = ∫ x(t)e−jω t dt
−∞
↓ t = nT ↓ ??
+∞
DTFT
x[n] ↔ X(ejΩ ) = ∑ x[n]e−jΩ n
n=−∞
162 | 5 Discrete processing of analog signals

̃
It follows from the IFT of X(jω ) and (5.1) that
+∞
1 ̃
x[n] = ∫ X(jω )ejω nTs dω .
2π
−∞

Now, divide the region (−∞, + ∞) for the variable ω into a set of small intervals
with the points
2πk + π
ωk ≜ , k = ⋅ ⋅ ⋅ , −2, −1, 0, 1, 2, . . . .
Ts
1
∑+∞
(2πk+π)/T
̃ jω nTs
k=−∞ ∫(2πk−π)/T X(jω )e
s
Then x[n] = 2π
dω and hence with ξ = Ts ω
s

(2πk+π)
1 +∞ 1 ̃ /T )ejξ n dξ .
x[n] = ∑ ∫ X(jξ s
2π k=−∞ Ts
(2πk−π)

With the intermediate variable Ω = ξ − 2πk defined, we finally have

π
1 1 +∞ ̃ Ω + 2πk jΩ n
x[n] = ∫ ∑ X (j ) e dΩ . (5.2)
2π Ts k=−∞ Ts
−π

Denote
1 +∞ ̃ Ω + 2πk
Φ (Ω ) ≜ ∑ X (j ).
Ts k=−∞ Ts
Note that Φ (Ω ) is periodic: Φ (Ω + 2π) = Φ (Ω ). It is clear that if c[m] are the FS
coefficients of Φ (Ω ), (5.2) suggests that c[m] = x[−m] and hence (noting Ω0 = 2π/T0 =
2π/2π = 1)
Φ (Ω ) = ∑ c[m]ejmΩ0 Ω = ∑ x[n]e−jnΩ = X(ejΩ ),
m n

which yields the following two equivalent relationships:

{
{ 1 +∞ ̃ Ω + 2kπ
{
{ X(ejΩ ) = ∑ X (j )
{
{ Ts k=−∞ Ts
{
{
{
{ Ω 2π (5.3)
{
{ (ω = , ωs ≜ )
{
{ Ts Ts
{
{
{
{ +∞
{
{ X(ejTs ω ) = 1 ∑ X(j(ω̃ + kωs )).
{ Ts k=−∞

2
ω sin(ω t/2)
As seen in Example 4.10, x(t) = 2πM ( ω Mt/2 ) ↔ X(jω ̃ ) = (1 − ω|ω | ) w2ωM (ω )
M M
with the latter shown in Figure 5.3(a). Figures 5.3(b) and (c) show the spectra of x[n]
obtained with x(t) sampled using different Ts .
Let x(t) be a signal of bandwidth limited to ωM like the one shown above. It is easy
to see that if ωM < ωs − ωM , i.e. ωs > 2ωM , then
̃
Ts X(ejTs ω ) = X(jω ), |ω | < ωs /2. (5.4)
 5.3 Spectral relationship and sampling theorem | 163
|X̃( jω)|

–ωM 0 ωM
(a) ω
|X(ejωTs)|

–1
Ts

–ωs –ωM 0 ωM ωs
(b) ω
|X(ejωTs)|

–1
Ts

–ωs –ωM 0 ωM ωs
(c) ω

̃ )|; (b) |X(ejω Ts )| with Ts =

Fig. 5.3: (a) |X(jω 4π
; (c) |X(ejω Ts )| with Ts = 4π
.
3ωM 5ωM

̃
This means that the analog spectrum X(jω ) can be extracted from the corresponding
jΩ
digital spectrum X(e ). We therefore have the following important theorem.

Theorem 5.1: Suppose x(t) is bandlimited:

̃
X(jω ) = 0, ∀ |ω | ⩾ ωM .

Then x(t) can be recovered from its samples x[n] = x(nTs ) if

2π
ωs ≜ > 2ωM ≜ ωN , (5.5)
Ts

where ωN is usually referred to as the Nyquist rate.

This is also called Nyquist theorem or sampling theorem.

Example 5.1: Let x(t) be a signal with a band limited to ωM and pτ (t) is a periodic
signal:
+∞
1
pτ (t) = ∑ wτ (t − kTs ).
k=−∞
τ
̂ = x(t)pτ (t).
Analyze the spectrum of x(t)
164 | 5 Discrete processing of analog signals

Solution: First of all,

pτ (t) = ∑ c[k]ejωs kt ,
k
where the FS coefficients are given by
1 sin(kωs τ /2) 1 sin(kωs τ /2)
c[k] = τ −1 τ = , ∀ k.
Ts kωs τ /2 Ts kωs τ /2
Therefore,

x(t) ̂
̂ = x(t)pτ (t) ↔ X(jω ) = ∑ c[k]X(j(ω − kωs )).
k

2
x(t)

−2
0 0.2 0.4 0.6 0.8 1
(a)
(t)

2
τpτ(t)

−1
0 0.2 0.4 0.6 0.8 1
(b)
(t)

4
τx͂(t)

−2
0 0.2 0.4 0.6 0.8 1
(c)
(t)

̃
Fig. 5.4: A mixer equivalent to a sampler. (a) x(t); (b) τ pτ (t); (c) τ x(t).

̂ One observes that when τ /Ts is very

Figure 5.4 shows graphically x(t), τ pτ (t) and τ x(t).
small, c[k] ≈ 1/Ts , ∀ k and hence X(jω̂ ) is very close to X(ejω Ts ), where X(ejΩ ) is the
DTFT of x[n] = x(nTs ). See (5.3). This means that the signal x(t) ̂ with small τ /Ts is
almost equivalent to the discrete-time signal x[n], while the former is obtained with
an analog multiplier. Furthermore, note that pτ (t) becomes an impulse train when
τ → 0. This is why the sample sequence x[n] is usually modelled as the product of
the analog sinal x(t) and the impulse train.
 5.4 Reconstruction of continuous-time signals | 165

5.4 Reconstruction of continuous-time signals

Let hd (t) be the unit impulse response of an LTI system Hd (jω ). We can construct a
continuous-time signal with the discrete-time signal x[n] in the following way

+∞
̂ = ∑ x[m]hd (t − mTs ),
x[n] → x(t) (5.6)
m=−∞

which can be implemented using the system shown in Figure 5.5, where Mix is a mix-
ture such that xp (t) = ∑+∞
m=−∞ x[m]δ (t−mTs ) and as to be discussed late, can be replaced
by a zero—or first order hold circuit in practice.

x[n] xp (t) ̂
x(t) ̄
x(t)
- Mix - hd (t) - H (jω ) -
i

Fig. 5.5: Structure of the discrete-analog-convertor (DAC).

̂ and the original x(t)?

What is the relationship between the reconstructed signal x(t)
Let us look at the problem from frequency-domain as it seems impossible to get an
answer from time-domain.
Applying the FT to both sides of the above equation yields
+∞ +∞
̂
X(jω ) = ∑ x[m]Hd (jω )e−jmω Ts = Hd (jω ) ∑ x[m]e−jmω Ts
m=−∞ m=−∞
jω Ts
= Hd (jω )X(e ). (5.7)

Take
sin(ωl t)
Hd (jω ) = Ts w2ωl (ω ) ≜ H0 (jω ) ⇔ h0 (t) = Ts , (5.8)
πt
where w2ωl (ω ) is the window function. It then turns out from (5.6) that
+∞ +∞
sin(ωl (t − mTs ))
x̂0 (t) = ∑ x[m]h0 (t − mTs ) = Ts ∑ x[m] . (5.9)
m=−∞ m=−∞ π(t − mTs )

̂
It follows from (5.3) that under the recovery condition (5.5), X(jω ̃
) = X(jω ) and
hence
x̂0 (t) ≡ x(t),

as long as
ωs
wM < ω l ⩽ . (5.10)
2
166 | 5 Discrete processing of analog signals

It is interesting to note that x̂0 (nTs ) = x[n] is always true if Hd (jω ) = k−1 Ts w2ωl (ω ),
kω
where ωl = 2 s for any integer k > 1. This provides a way to generate different analog
signals that yield the same discrete sequence x[n] when sampled with Ts .
Note that (5.6) can be rewritten into
n +∞
̂ = ∑ x[m]hd (t − mTs ) + ∑ x[m]hd (t − mTs ).
x(t)
m=−∞ m=n+1

In an on-line (i.e. real-time) signal reconstruction system, for t < (n + 1)Ts the
samples of discrete-time sequence x[m] are available just for m ⩽ n. In that case, the
2nd term of the above equation should be nil for all t < (n + 1)Ts in order to avoid
̂
using x[n + 1], x[n + 2], ⋅ ⋅ ⋅ in evaluating x(t). To achieve that, it suffices to ensure
hd (t) ≡ 0, ∀ t < 0. This means that the system Hd (jω ) should be causal !
sin(ω t)
Look at (5.9) in which the ideal reconstruction system h0 (t) = Ts πt l is non-
causal. It is of theoretical importance, but in practice it has to be replaced with an
implementable reconstruction system.

Zero Order Hold (ZOH): It is defined as h1 (t) ≜ wTs (t − Ts /2), leading to x̂1 (t) =
∑+∞
m=−∞ x[m]h1 (t − mTs ). See Figure 5.6 as an example.
sin(ω T /2)
As Hd (jω ) = H1 (jω ) = 2 ω
s
e−jω T2 /2 , we can see that the reconstructed signal
x̂1 (t) is different from x(t) due to the following three factors: i) there is a time delay
of Ts /2; ii) the magnitude response H1 (jω ) is not flat within [−ωM , ωM ], and iii) the
magnitude response H1 (jω ) is not nil constantly outside [−ωM , ωM ]. The first two fac-
tors may not be serious when Ts is very small, while the third one can introduce high
frequency noises. This is demonstrated with Figure 5.7.

Ts 3Ts
First Order Hold (FOH): It is defined as h2 (t) ≜ Tt [wTs (t − 2
) − wTs (t − 2
)]. See
s
Figure 5.8(a). This leads to x̂2 (t) = ∑+∞
m=−∞ x[m]h2 (t − mTs ).
It can be shown that
+∞
x[m + 1] − x[m]
x̂2 (t) = ∑ [x[m] + (t − mTs )]h1 (t − mTs ).
m=−∞ Ts

Clearly, the first order hold yields a better result. Why is that?
To remove these undesired high frequency components, the continuous-time sig-
̂ from the hold is usually smoothed by a smoothing analog low-pass filter –
nal x(t)
anti-imaging filter Hi (jω ). See Figure 5.5.
Based on (5.7), we have

̄
X(jω ) = Hi (jω )Hd (jω )X(ejω Ts ). (5.11)

What is the best Hi (jω )?

Before turning to the next topic, we should point out that the condition specified
by (5.5) is just a sufficient condition to ensure a continuous-time signal to be restored
 5.4 Reconstruction of continuous-time signals | 167

1.5

1
h1(t)

0.5

0
−0.3 −0.2 −0.1 0 0.1 0.2 0.3
(a)
t

3
x1̂ (t)

0
0 0.2 0.4 0.6 0.8 1
(b)
t

Fig. 5.6: Reconstruction of x(t) (dotted-line) using ZOH, where Ts = 0.1.

|X̃( jω)|

|H1(jω)|

1
Ts

–ωm 0 ωm –2ωs –ωs 0 ωs 2ωs

ω ω
|X(ejTsω)|

1
|X̃1(jω)|

Ts 1

–ωs –ωm 0 ωm ωs –ωs –ωm 0 ωm ωs

ω ω

Fig. 5.7: Demonstration of the imperfection of the ZOH reconstruction.

168 | 5 Discrete processing of analog signals

1.5

1
h2(t)

0.5

0
−0.2 −0.1 0 0.1 0.2 0.3 0.4
(a) t

3
x2̂ (t)

0
0 0.2 0.4 0.6 0.8 1
(b) t

Fig. 5.8: Reconstruction of x(t) (dotted-line) using FOH, where Ts = 0.1.

from its discrete-time counterpart. There are signals which do not satisfy this condi-
tion but still can be recovered from their samples.
Look at the spectrum of a bandpass signal depicted in Figure 5.9. In communica-
tions, it is quite often that the band width of such a signal is relatively much smaller
than ω1 , i.e. Bw ≜ ω2 − ω1 << ω1 .
As indicated by Figure 5.9, where

ω̃ 1 ≜ mωs − ω1 , ω̃ 2 ≜ (m + 1)ωs − ω2 ,

once there exists a positive integer m such that

−ω1 + mωs < ω1 mωs < 2ω1

{ ⇒ {
ω2 < −ω2 + (m + 1)ωs 2ω2 < (m + 1)ωs

̃
there will be no spectral overlapping between the shifted versions of X(jω ) in the DTFT
 5.4 Reconstruction of continuous-time signals | 169

̃ )
X(jω

-
−ω2 −ω1 0 ω1 ω2

̃
X(j(ω − mωs ))

∙ -
0 ω̃ 1 mωs

̃
X(j(ω − (m + 1)ωs ))

∙ -
0 ω̃ 2 (m + 1)ωs
Fig. 5.9: The spectrum of x(t) and its shifted versions.

spectrum X(ejω Ts ) given by (5.3). The above is equivalent to

2ω1 − (ωs − 2Bw ) < mωs < 2ω1 (5.12)

and holds for some positive integer m.¹

For example, assume ω1 = 5 Bw . It is easy to verify that the choice: ωs = 3.2Bw ,
m = 3 is one of the solutions to (5.12). The corresponding ωs is smaller than the Nyquist
frequency 2ωM = 2ω2 = 12Bw .
Denote
ω − ω1 ω + ω1
ωl ≜ 2 , ω0 ≜ 2 ,
2 2
and h0 (t) ↔ H0 (jω ) = Ts w2ωl (ω )—the ideal low-pass filter as defined in (5.8). If the
following (band-pass) reconstruction system

Hd (jω ) = H0 (j(ω + ω0 )) + H0 (j(ω − ω0 )) ↔ hd (t) = 2h0 (t) cos(ω0 t)

is used, the original signal x(t) can be recovered with

+∞
x(t) = ∑ x[m]hd (t − mTs ).
m=−∞

In this case, the sampling frequency ωs satisfies (5.12) and is not necessarily larger
than the Nyquist rate 2ω2 .

ω2 +ω1
1 It can be shown that there is no spectrum overlapping once ωs = 2m
, where m is the biggest
ω +ω
integer satisfying m ⩽ 2(ω2 −ω1 ) . See Problem 8.21.
2 1
170 | 5 Discrete processing of analog signals

5.5 Hybrid systems for discrete processing

We have realized from the discussions above that in order to recover the continuous-
time signal the sampling frequency has to be chosen carefully such that there is a
̃
shifted version of X(jω ) in the digital spectrum Ts X(ejω Ts ).
Figures 5.10(a)-(b) show a speech signal x0 (t) and its spectrum. This signal is de-
filed by a noise of higher frequencies than 20 kHz. Figures 5.10(c)–(d) yield the time
representation and the frequency representation of the defiled signal x1 (t). It is well
known that our human being’s ears are not very sensitive to frequencies higher than
20 kHz, so such a noise is not very annoying if one listens to it directly with an analog
instrument.

× 10−3
0.5 5
4
X0(jω)
x0(t)

3
0
2
1
−0.5 0
0 0.02 0.04 0.06 0.08 0.1 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
× 104
(a) t (b) ω

× 10−3
0.5 5
4
X1(jω)

3
x1(t)

0
2
1
−0.5 0
0 0.02 0.04 0.06 0.08 0.1 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
× 104
(c) t (d) ω

× 10−3
0.5 5
4
3
Xˆ0(jω)
x0̂ (t)

0
2
1
−0.5 0
0 0.02 0.04 0.06 0.08 0.1 −1 −0.5 0 0.5 1
× 104
(e) t (f) ω

Fig. 5.10: (a)—(b): for x0 (t); (c)–(d): for x1 (t); (e)–(f): for x0̂ (t).

If we discretize this signal with a sampling frequency 8.00kHz, the high frequency
spectrum than 20kHz will be folded around the frequency ωs . This aliasing effect can
damage badly the reconstructed signal. Figures 5.10(e)–(f) depict the reconstructed
signal x̂0 (t) from samples of x1 (t) using the ideal construction function h0 (t) and the
corresponding spectrum.
 5.6 Discrete Fourier transform | 171

In many applications, the analog signals have a very wide bandwidth of spectrum.
In order to reduce the effect of the aliasing phenomenon, the signals are pre-filtered
before sampled. Such a filter, denoted as Ha (jω ), is usually called aliasing filter and is
intended to limit the spectrum of the output xa (t) to ωs /2 without damaging too much
the information of x(t). See Figure 5.11.

x(t) xa (t) xa [n] y[n] ̂

y(t) ̄
y(t)

- Ha (jω ) - ADC - H(ejΩ ) - DAC - Hi (jω ) -

Fig. 5.11: Block-diagram of discrete processing of continuous-time signals.

Suppose that we have a continuous-time system g(t) to process x(t): y(t) = g(t) ∗ x(t)
such that y(t) is of a spectrum limited to ωs /2. The block-diagram of discrete process-
ing of the same signal is shown in Figure 5.11. Based on the relations established be-
fore, we have

{
{
{ X̃ a (jω ) = Ha (jω )X(jω )
{
{
{
{
{
{ Xa (ejω Ts ) = Ts−1 ∑k X̃ a (j(ω − kωs ))
{
{
{
{ Y(ejω Ts ) = H(ejω Ts )Xa (ejω Ts )
{
{
{
{
{ ̄
Y(jω ) = Hi (jω )Hd (jω )Y(ejω Ts ).
{
Therefore, we have

̄
Y(jω ) = Ts−1 ∑ Hd (jω )Hi (jω )H(ejω Ts )Ha (j(ω − kωs ))X(j(ω − kωs )). (5.13)
k

Clearly, under the assumption that Ha (jω ) = 0, ∀ |ω | > ωs /2, we will have

̄ = y(t) = g(t) ∗ x(t),

y(t)

as long as
Hd (jω )Hi (jω )H(ejω Ts )Ha (jω ) = G(jω ), ∀ |ω | < ωs /2.

5.6 Discrete Fourier transform

Figure 5.12 depicts a discrete-time signal x[n] and its magnitude spectrum |X(ejΩ )|.
As is known, x[n] can be represented using X(ejΩ )—a function of the continuous
variable Ω . It is therefore not an efficient way to represent the sequence x[n] by us-
ing X(ejΩ ) since to do so, one would need a memory device of infinite bits, which is
practically impossible.
172 | 5 Discrete processing of analog signals

2
1.5
1
x[n]

0.5
0
−0.5
0 5 10 15 20 25 30 35 40 45 50
(a) n

20

15
|X(ejΩ)|

10

0
0 1 2 3 4 5 6
(b) Ω

Fig. 5.12: (a) a finite duration x[n] of 24 samples; (b) |X(ejΩ )|.

To overcome this problem, like what we have just discussed above on discrete pro-
cessing of continuous-time signals, we can evaluate the DTFT of x[n] just for a set of
frequency points only. As most of our readers have realized, this is something like a
sampling procedure.
The discrete Fourier transform (DFT) of a discrete-time signal x[n] is the DTFT of
the same signal but computed at the frequency points

2πk
Ωk = ≜ kΩs , k ∈ Z, (5.14)
Ns

where Ωs = 2π/Ns is the sampling period with Ns a positive integer indicating the
number of samples taken within 2π.
The Ns -point DFT of x[n] is then defined as
+∞ 2πkn
X[k] ≜ ∑ x[n]e−j Ns = X(ejΩs k ). (5.15)
n=−∞

Note X[k] defined by (5.15) is periodic with period Ns . So, the DTFS suggests
Ns −1
−jΩs mk
̃
X[k] = ∑ x[m]e , (5.16)
m=0
 5.6 Discrete Fourier transform | 173

̃
where² the sequence x[m] (i.e. the DTFS coefficients of X[k]) are given by
N −1
1 s
̃
x[m] = ∑ X[k]ejΩs km , m = 0, 1, . . . , Ns − 1, (5.17)
Ns k=0

called the Ns -point inverse discrete Fourier transform (IDTF) of X[k].

̃ and x[n] are the inverse of X[k] and X(ejΩ ), respectively, and X[k] = X(ejΩs k ),
As x[n]
a fundamental question to be asked is: what is the relationship between x[n] ̃ and x[n]?
Inserting X[k] by (5.15) into (5.17) yields
N −1 N −1
1 s +∞ +∞ s
̃
x[m] = ∑ ∑ x[n]e−jΩs kn ejΩs km = ∑ x[n] ∑ ejΩs k(m−n) .
Ns k=0 n=−∞ n=−∞ k=0

It follows from Ωs = 2π/Ns that

Ns −1
∑ ejΩs k(m−n) = Ns δ [m − n − lNs ]
k=0

holds for any integer −∞ < l <+ ∞ and hence

+∞
̃
x[m] = ∑ x[m − lNs ]. (5.18)
l=−∞

It is easy understand that (5.18) is in general not invertible, this is to say that x[n]
cannot be uniquely determined from x[n] ̃ for any Ns .
However, if x[n] is of finite duration L, say x[n] = 0, ∀n < N1 , n > N1 + L − 1. Then

̃ = x[n],
x[n] n = N1 , N1 + 1, . . . , N1 + L − 1,

as long as
Ns ⩾ L. (5.19)
The above is actually consistent with the sampling theorem specified with (5.5),
while (5.18) is the discrete counterpart of (5.3). It should be noted that (5.19) is just a
sufficient condition for the sampling number Ns to ensure the reconstruction of x[n]
from its DFT.

Remark: For a finite duration signal, say x[n] = x[n](u[n]−u[n−N]), xp [n] = ∑+∞
k=−∞ x[n−
kN] is periodic with a period of N. As known before (see (3.50)), the DTFS coefficients
of such a periodic signal is given by
1 2πk
Xp [k] = X(ej N ), ∀ k.
N

2 Note that the periodic signal X[k] is decomposed using bases {e−jΩs m , m = 0, 1, ⋅ ⋅ ⋅ , Ns − 1} rather
than {ejΩs m , m = 0, 1, ⋅ ⋅ ⋅ , Ns − 1} as they are identical.
174 | 5 Discrete processing of analog signals

Clearly, the N-DFT of x[n] is X[k] = NXp [k]. In that sense, DFT and DTFS are equiva-
lent though they are derived from the concepts of sampling and transforming, respec-
tively.

Example 5.2: Consider the signal x[n] shown in Figure 5.12, which is of a duration of
L = 24 samples, starting from n = 0 and ending at n = 23. We computed its Ns -DFT
and then corresponding IDFT x[n] ̃ for different Ns = 12, 24, 28.
Case 1: Ns = 12 < L, corresponding a under-sampling situation. Figure 5.13(a)
2πk
shows the 24 samples |X[k]| = |X(ej 24 )|, while Figure 5.13(b) yields the corresponding
̃
IDFT x[n].
̃ is totally different from x[n] for n = 0, 1, 2, . . . , 23 due to the
As seen, the IDFT x[n]
aliasing effect.
Case 2: Ns = 24 = L, corresponding to the critical sampling. The results are rep-
resented in Figure 5.14.

20

15

10
|X[k]|

0
0 1 2 3 4 5 6
Ωk = 2πk
(a) 12

3
2.5
2
1.5
1
x͂[n]

0.5
0
−0.5
0 5 10 15 20 25 30 35 40 45 50
(b) n

̃
Fig. 5.13: Top:12-point DFT of x[n] shown in Figure 5.12; Bottom:the IDFT x[n].

̃
For this case, as observed, x[n] = x[n], n = 0, 1, 2, ⋅ ⋅ ⋅ , 23, which means that x[n] can
be recovered from X[k] as there is no overlapping between x[n] and x[n + 24m] for any
nonzero integer.
Case 3: Ns = 28 > L, corresponding to an over-sampling. The results are repre-
sented in Figure 5.15.
 5.7 Compressed sensing | 175

20

15
|X[k]|

10

0
0 1 2 3 4 5 6
Ωk = 2πk
(a) 24

2
1.5
1
x͂[n]

0.5
0
−0.5
0 5 10 15 20 25 30 35 40 45 50
(b)
n

̃
Fig. 5.14: Top:24-point DFT of x[n] shown in Figure 5.12; bottom:the IDFT x[n].

̃
As expected, x[n] = x[n], n = 0, 1, 2, ⋅ ⋅ ⋅ , 23, which is confirmed in Figure 5.15(a).
Though over-sampling provides a better resolution than the critical-sampling, both
have exactly the same information on x[n].

5.7 Compressed sensing

We have pointed out that Nyquist Theorem yields a sufficient condition that guaran-
tees analog signals can be recovered exactly from their samples, in another words,
xa (t) can be totally represented by x[n] = xa (nTs ) as long as (5.5) is satisfied. Now, let
ψ1 (t) = ej2π 310t and ψ2 (t) = ej2π 499 and

xa (t) = α1 ψ1 (t) + α2 ψ2 (t). (5.20)

Clearly, xa (t) is bandlimited to fM = 499 Hz. If we sample this signal with fs =

1000 Hz, we have 1000 samples x[n] per second. Can we use much fewer samples
of x[n] (than 1000) if we know a priori that xa (t) has a structure given by (5.20) with
ψ1 (t), ψ2 (t) known, say just two samples xa (t1 ), xa (t2 )? This is possible! Surprised?
In fact, it follows from (5.20) that

xa (t1 ) ψ (t ) ψ2 (t1 ) α α
[ ]=[ 1 1 ][ 1 ] ≜ Ψ [ 1 ]. (5.21)
xa (t2 ) ψ1 (t2 ) ψ2 (t2 ) α2 α2
176 | 5 Discrete processing of analog signals

20

15
|X[k]|

10

0
0 1 2 3 4 5 6
Ωk = 2πk
(a) 28

2
1.5
1
x͂[n]

0.5
0
−0.5
0 5 10 15 20 25 30 35 40 45 50
(b)
n

̃
Fig. 5.15: Top:28-point DFT of x[n] shown in Figure 5.12; Bottom:the IDFT x[n].

As we have seen, we can obtain the signal parameters α1 , α2 from just two samples
of xa (t) as long as the matrix Ψ is nonsingular, and hence the whole signa xa (t) with
(5.20).
Consider a more complicated situation, where
L
xa (t) = ∑ αk ψk (t)
k=1
T
= [ ψ1 (t) ⋅ ⋅ ⋅ ψL (t) ] [ α1 ⋅ ⋅ ⋅ αL ] ≜ ψ (t) α , (5.22)

with the set {ψk (t)} given. Suppose we know that there are at most K (K << L) elements
in the vector α are nonzero but the corresponding positions are unknown. How many
samples of xa (t) do we need in order to represent xa (t) completely? Obviously,

xa (t1 ) ψ (t1 ) α1
[ x (t ) ] [ ψ (t ) ][ ]
[ a 2 ] [ 2 ][ α2 ]
[ ]=[ ][ ] ⇐⇒ x = Ψ α . (5.23)
[ .. ] [ .. ][ .. ]
[ . ] [ . ][ . ]
x
[ a M)
(t ] [ ψ (t M) α
][ M ]

As the above is a under-determined equation (due to M < L), for a given x, con-
sisting of M samples of xa (t), the solutions α of the above equation are not unique and
to find the one that has at most K nonzero elements, we have to carry out an exhaus-
tive searching, which is very time consumable when L is very big. This is one of key
 5.8 Summary | 177

problems dealt with in compressed sensing (CS) – an emerging research area in signal
processing community.
CS has attracted a lot of attention since its introduction at the early of this century.
A “holy grail” of the compressed sensing is to build acquisition devices that exploit
signal structure in order to reduce the sample rate, and consequently demands on stor-
age and bandwidth required for transmission. As its core, CS is a mathematical frame-
work that studies accurate recovery of a signal represented by a vector of length L from
M << L measurements, effectively performing compression during signal acquisition.
The measurement paradigm consists of linear projections, or inner product, of the sig-
nal vector into a set of carefully chosen projection vectors that act as a multitude of
probes on the information contained in the signal. For more details on CS, we refer
to [4].

5.8 Summary

The main focus of this chapter is on discrete processing of continuous-time signals.

The development flows from the techniques derived in the previous chapters, partic-
ularly Chapters 3 and 4.
Starting from the basic relationship between a continuous-time signal x(t) and
its discrete-time counterpart x[n], we have derived the important spectral relationship
between the DTFT of x[n] and the FT of x(t). Based on such a relation one sufficient
condition for completely recovering x(t) from its samples x[n] is obtained, embodied
in the sampling theorem. Reconstruction of a continuous-time signal, the effects of
aliasing phenomena and nonideal reconstruction functions in the hybrid system of
ADC and DAC have been discussed.
Sampling the DTFT of a discrete-time signal has also been studied in this chapter,
leading to the important topic— DFT. It should be pointed out that among the Fourier
analysis-based transforms we have discussed so far, the DFT is the one that has been
most popularly used in many applications of digital signal processing, including lin-
ear filtering, correlation analysis, and spectral analysis. The key reason for its impor-
tance is the existence of efficient algorithms for computing the Ns -point DFT, among
which there is one called fast Fourier transform (FFT) when Ns is a power of two. More
profound discussions on the DFT are beyond the scope of this book as the relevant
topics are parts of core contents for the subjects on digital signal processing.
The signal transforms discussed so far provide very powerful tools for us to an-
alyze signals and systems. In terms of signal analysis, these classical transforms are
still popularly used now-days in many areas of research and applications. Recall the
FT/DTFT-based concept of frequency response of LTI systems. One of the requirements
by this concept is that the systems should be stable. This is because the FT-based trans-
forms cannot handle the class of unstable signals and systems. To overcome this prob-
178 | 5 Discrete processing of analog signals

lem, more general transforms are definitely needed and this is the topic we will have
in the next chapter.

5.9 Problems

Problem 5.1: Let X(jω ) = w2B (ω ) be the FT of a signal x(t). Sketch the spectrum X(ejω Ts )
4π 4π
of x[n] = x(nTs ) for Ts = 3B , 5B , respectively.

Problem 5.2: A continuous-time signal x(t) is obtained at the output of an ideal low-
pass filter with cutoff frequency ωc = 1000π. If x(t) is sampled, which of the following
sampling periods would guarantee that x(t) can be recovered from its sampled version
using appropriate lowpass filter?
(a) T = 0.5 × 10−3 ;
(b) T = 2 × 10−3 ;
(c) T = 10−4 .

Problem 5.3: Let x(t) be a signal with Nyquist rate ω0 . Determine the Nyquist rate for
each of the following signals:
(a) x(t) + x(t − 1);
(b) dx(t)
dt
;
(c) x2 (t);
(d) x(t) cos(ω0 t).

Problem 5.4: Consider a signal x(t) = x0 (t) + w(t), where the desired signal x0 (t) is
bandlimited to 4000Hz and the noise signal w(t) has a spectrum within [4050, 5500]
(in Hz). We wish to detect x0 (t) from x(t) by discrete processing of the latter. The whole
system consists of sampling (ADC), digital filtering, and reconstruction (DAC). Specify
each part of the system as much as you can.

Problem 5.5: Let xa (t) ↔ Xa (jω ) be an FT pair and

+∞
p(t) = ∑ (−1)k δ (t − kT).
k=−∞

(a) Compute the FT of y(t) = xa (t)p(t).

(b) Denote w[n] = x[n] ∑+∞ k
k=−∞ (−1) δ [n − k],where x[n] = xa (nT). Derive the expression
for the spectrum (i.e. the DTFT) of w[n] in terms of Xa (jω ). Is it possible to recover
xa (t) from the sample sequence w[n]? Justify your answer.

Problem 5.6: Shown in Figure 5.16 is a system in which the input signal is multiplied
by a periodic square wave. The period of s(t) is T. The input signal is band limited
󵄨 󵄨
with 󵄨󵄨󵄨X(jω )󵄨󵄨󵄨 = 0 for |ω | ⩾ ωM . Determine, in terms of ωM , the maximum value of T for
 5.9 Problems | 179

which there is no aliasing among the replicas of X(jω ) in W(jω ) for different Δ values
given in the following:

(1) Δ = T/3, (2) Δ = T/4

s(t)

x(t) w(t)
1

T t
s(t) –1

Fig. 5.16: Block-diagram of the system for Problem 5.6.

Problem 5.7: In practice, the ideal sampling x[n] = x(nTs ) is replaced using
nTs
1
̂ ≜
x[n] ∫ x(τ )dτ ,
Ts
(n−1)Ts

which is the mean value of x(t) within the interval ((n − 1)Ts , nTs ).
̃ is actually equal to the ideal sampling of an analog signal y(t), i.e.
(a) Show that x[n]
̃ = y(nTs ), where y(t) = h(t) ∗ x(t). Find such a h(t).
x[n]
̃ jω Ts ) in terms of X(jω ), T and H(jω );.
(b) Derive the expression for X(e s
(c) Assume that x(t) is band limited to the range |ω | < 3π/(4Ts ). Design a discrete-
time system Q(ejΩ ) such that it yields x[n] when excited with x[n].̃

Problem 5.8: Let h1 (t), h2 (t) be the zero- and first-hold defined in the textbook. x1 (t)
is the reconstructed signal from x(nTs ) using h1 (t), i.e. x1 (t) = ∑+∞
n=−∞ x(nTs )h1 (t − nTs ).
Suppose y(t) is the output of an LTI system h(t) when the input is x1 (t). Design h(t) such
that y(t) = ∑+∞
n=−∞ x(nTs )h2 (t − nTs ). This design actually provides a way to implement
the first-order hold operation based on the zero-order hold.

Problem 5.9: Given that x[n] ↔ X(ejΩ ) and Xk is the N-point DFT of x[n].
(a) Find out the DTFT and N−point DFT of y[n] = (−1)n x[n], respectively.
(b) Denote z[n] = ∑7m=0 x[n − mN]. Compute the 8N−point DFT of z[n].

/2)
Problem 5.10: Let x[n] ↔ X(ejΩ ) = sin(11Ω
sin(Ω /2)
and Xk be the N-point DFT of x[n]. Find
̃ for N = 6, 15, 20, respectively, and compare them with x[n].
the IDFT x[n]