Chapter 6

Sampling Theorem

Sampling theorem plays a crucial role in modern digital signal processing. The the-
orem concerns about the minimum sampling rate required to convert a continuous
time signal to a digital signal, without loss of information.

6.1 Analog to Digital Conversion

Consider the following system shown in Fig. 6.1. This system is called an analog-
to-digital (A/D) conversion system. The basic idea of A/D conversion is to take a
continuous-time signal, and convert it to a discrete-time signal.

Figure 6.1: An analog to digital (A/D) conversion system.

Mathematically, if the continuous-time signal is x(t), we can collect a set of samples

by multiplying x(t) with an impulse train p(t):
1
X
p(t) = (t nT ),
n= 1

83
84 CHAPTER 6. SAMPLING THEOREM

where T is the period of the impulse train. Multiplying x(t) with p(t) yields

xp (t) = x(t)p(t)
X1
= x(t) (t nT )
n= 1
1
X
= x(t) (t nT )
n= 1
X1
= x(nT ) (t nT ).
n= 1

Pictorially, xp (t) is a set of impulses bounded by the envelop x(t) as shown in Fig.
6.2.

Figure 6.2: An example of A/D conversion. The output signal xp (t) represents a set
of samples of the signal x(t).

We may regard xp (t) as the samples of x(t). Note that xp (t) is still a continuous-time
signal! (We can view xp (t) as a discrete-time signal if we define xp [n] = x(nT ). But
this is not an important issue here.)

6.2 Frequency Analysis of A/D Conversion

Having an explanation of the A/D conversion in time domain, we now want to study
the A/D conversion in the frequency domain. (Why? We need it for the develop-
ment of Sampling Theorem!) So, how do the frequency responses X(j!), P (j!) and
Xp (j!) look like?
6.2. FREQUENCY ANALYSIS OF A/D CONVERSION 85

6.2.1 How does P (j!) look like?

Let’s start with P (j!). From Table 4.2 of the textbook, we know that
1
X 1
2⇡ X
F.T. 2⇡k
p(t) = (t nT ) ! (! ) = P (j!) (6.1)
n= 1
T k= 1 T

This means that the frequency response of the impulse train p(t) is another impulse
train. The only di↵erence is that the period of p(t) is T , whereas the period of P (j!)
is 2⇡
T
.

Figure 6.3: Illustration of X(j!) and P (j!).

6.2.2 How does Xp (j!) look like?

Next, suppose that the signal x(t) has a frequency response X(j!). We want to know
the frequency response of the output xp (t). From the definition of xp (t), we know
know that
xp (t) = x(t)p(t).
Therefore, by the multiplication property of Fourier Transform, we have
1
Xp (j!) = X(j!) ⇤ P (j!).
2⇡
86 CHAPTER 6. SAMPLING THEOREM

Shown in Fig. 6.3 are the frequency response of X(j!) and P (j!) respectively. To
perform the convolution in frequency domain, we first note that P (j!) is an impulse
train. Therefore, convolving X(j!) with P (j!) is basically producing replicates at
every 2⇡
T
. The result is shown in Fig. 6.4.

Figure 6.4: Convolution between X(j!) and P (j!) yields periodic replicates of
X(j!).

Mathematically, the output Xp (j!) is given by

Z 1
1 1
Xp (j!) = X(j!) ⇤ P (j!) = X(j✓)P (j(! ✓))d✓
2⇡ 2⇡ 1
Z 1 " 1
#
1 2⇡ X 2⇡k
= X(j✓) (! ✓ ) d✓
2⇡ 1 T k= 1 T
1 Z 1
1 X 2⇡k
= X(j✓) (! ✓ )d✓
T k= 1 1 T
1 ✓ ◆
1 X 2⇡k
= X j(! ) .
T k= 1 T
The result is illustrated in Fig. 6.5.

6.2.3 What happens if T becomes larger and larger ?

If T becomes larger and larger (i.e., we take fewer and fewer samples), we know
from the definition of p(t) that the period (in time domain) between two consecutive
6.2. FREQUENCY ANALYSIS OF A/D CONVERSION 87

Figure 6.5: Illustration of xp (t) and Xp (j!).

impulses increases (i.e., farther apart). In frequency domain, since

1
2⇡ X 2⇡k
P (j!) = (! ),
T k= 1 T

the period 2⇡
T
reduces! In other words, the impulses are more packed in frequency
domain when T increases. Fig. 6.6 illustrates this idea.

Figure 6.6: When T increases, the period in frequency domain reduces.

If we consider Xp (j!), which is a periodic replicate of X(j!) at the impulses given

by P (j!), we see that the separation between replicates reduces. When T hits cer-
tain limit, the separation becomes zero; and beyond that limit, the replicates start to
overlap! When the frequency replicates overlap, we say that there is aliasing.
88 CHAPTER 6. SAMPLING THEOREM

Figure 6.7: When T is sufficiently large, there will be overlap between consecutive
replicates.

Therefore, in order to avoid aliasing, T cannot be too large. If we define the sampling
rate to be
2⇡
!s = ,
T

then smaller T implies higher !s . In other words, there is a minimum sampling rate
such that no aliasing occurs.

Figure 6.8: Meanings of high sampling rate v.s. low sampling rate.

6.2.4 What is the minimum sampling rate such that there is

no aliasing?

Here, let us assume that the signal x(t) is band-limited. That is, we assume X(j!) = 0
for all |!| > W , where W is known as the band-width.

To answer this question, we need the Sampling Theorem.

6.3. SAMPLING THEOREM 89

Figure 6.9: Left: A band limited signal (since X(j!) = 0 for all ! > |W |.) Right: A
band non-limited signal.

6.3 Sampling Theorem

Theorem 11 (Sampling Theorem). Let x(t) be a band limited signal with X(j!) = 0
for all |!| > W . Then the minimum sampling rate such that no aliasing occurs in
Xp (j!) is

!s > 2W,
2⇡
where !s = T
.

6.3.1 Explanation
Suppose x(t) has bandwidth W . The tightest arrangement that no aliasing occurs is
shown in Fig. 6.10

Figure 6.10: Minimum sampling rate that there is no aliasing.

2⇡
In this case, we see that the sampling rate !s (= T
) is

!s = 2W.
90 CHAPTER 6. SAMPLING THEOREM

If T is larger (or !s is smaller), then 2⇡

T
becomes less than 2W , and aliasing occurs.
Therefore, the minimum sampling rate to ensure no aliasing is

!s > 2W.

6.3.2 Example
Suppose there is a signal with maximum frequency 40kHz. What is the minimum
sampling rate ?

Figure 6.11: Example: Minimum sampling frequency.

Answer :
Since ! = 2⇡f , we know that the max frequency (in rad) is ! = 2⇡(40 ⇥ 103 ) =
80 ⇥ 103 ⇡ (rad). Therefore, the minimum Sampling rate is: 2 ⇥ (80 ⇥ 103 ⇡), which is
160 ⇥ 103 ⇡ (rad) = 80kHz.

6.4 Digital to Analog Conversion

In the previous sections, we studied A/D conversion. Now, given a discrete-time
signal (assume no aliasing), we would like to construct the continuous time signal.

6.4.1 Given Xp (t) (no aliasing), how do I recover x(t)?

If no aliasing occurs during the sampling processing (i.e., multiply x(t) with p(t)),
then we can apply a lowpass filter H(j!) to extract the x(t) from xp (t). Fig. 6.12
shows a schematic diagram of how this is performed.
To see how an ideal lowpass filter can extract x(t) from xp (t), we first look at the
frequency response of Xp (j!). Suppose that p(t) has a period of T (so that !s = 2⇡
T
).
6.4. DIGITAL TO ANALOG CONVERSION 91

Figure 6.12: Schematic diagram of recovering x(t) from xp (t). The filter H(j!) is
assumed to be an ideal lowpass filter.

Then 1
1X
Xp (j!) = X(j(! k!s )).
T 1
As shown in the top left of Fig. 6.13, Xp (j!) is a periodic replicate of X(j!). Since
we assume that there is no aliasing, the replicate covering the y-axis is identical to
X(j!). That is, for |!| < !2s ,

Xp (j!) = X(j!).

Now, if we apply an ideal lowpass filter (shown in bottom left of Fig. 6.13):
(
1, |!| < !2s ,
H(j!) =
0, otherwise,

then
Xp (j!)H(j!) = X(j!),
for all !. Taking the inverse continuous-time Fourier transform, we can obtain x(t).

6.4.2 If Xp (t) has aliasing, can I still recover x(t) from xp (t) ?
The answer is NO. If aliasing occurs, then the condition

Xp (j!) = X(j!)
92 CHAPTER 6. SAMPLING THEOREM

Figure 6.13: Left: Multiplication between Xp (j!) and the lowpass filter H(j!). The
extracted output X̂(j!) is identical to X(j!) if no aliasing occurs. By applying
inverse Fourier transform to X̂(j!) we can obtain x(t).

does not hold for all |!| < !2s . Consequently, even if we apply the lowpass filter H(j!)
to Xp (j!), the result is not X(j!). This can be seen in Fig. 6.14.

Figure 6.14: If aliasing occurs, we are unable to recover x(t) from xp (t) by using an
ideal lowpass filter.
6.4. DIGITAL TO ANALOG CONVERSION 93

6.4.3 What can I do if my sampling device does not support

a very high sampling rate ?
• Method 1: Buy a better sampling device !

• Method 2: Send signals with narrower bandwidth or limit the bandwidth be-
fore sending :

• Method 3: Go to grad school and learn more cool methods !!