Chapter

Sampling and
Aliasing 4
With this chapter we move the focus from signal modeling and
analysis, to converting signals back and forth between the analog
(continuous-time) and digital (discrete-time) domains. Back in
Chapter 2 the systems blocks C-to-D and D-to-C were intro-
duced for this purpose. The question is, how must we choose the
sampling rate in the C-to-D and D-to-C boxes so that the analog
signal can be reconstructed from its samples.
The lowpass sampling theorem states that we must sample
at a rate, f s , at least twice that of the highest frequency of interest
in analog signal x  t  . Specifically, for x  t  having spectral con-
tent extending up to B Hz, we choose f s = 1  T s  2B in form-
ing the sequence of samples
x  n  = x  nT s  –   n   . (4.1)

Sampling
• We have spent considerable time thus far, with the continu-
ous-time sinusoidal signal
x  t  = A cos  t +   , (4.2)
where t is a continuous variable
• To manipulate such signals in MATLAB or any other com-
puter too, we must actually deal with samples of x  t 

ECE 2610 Signal and Systems 4–1

 Sampling

• Recall from the course introduction, that a discrete-time sig-

nal can be obtained by uniformly sampling a continuous-time
signal at t n = nT s , i.e., x  n  = x  nT s 
– The values x  n  are samples of x  t 
– The time interval between samples is T s
– The sampling rate is f s = 1  T s
– Note, we could write x  n  = x  n  f s 
• A system which performs the sampling operation is called a
continuous-to-discrete (C-to-D) converter
Ideal
xt C-to-D x  n  = x  nT s 
Converter

Ts

Simple Sampler Ts
Switch Model xt x  n  = x  nT s 
Momentarily close to take a sample

Electronic
Subsystem: Analog B-bits, e.g., 16, 18
ADC or to
A-to-D
xt Digital
x  n  = x  nT s 
Converter

SAR Others

Flash 

SAR = successive approximation register

 = delta-sigma modulator (oversampling)

ECE 2610 Signals and Systems 4–2

 Sampling

• A real C-to-D has imperfections, with careful design they can

be minimized, or at least have negligible impact on overall
system performance
• For testing and simulation only environments we can easily
generate discrete-time signals on the computer, with no need
to actually capture and C-to-D process a live analog signal
• In this course we depict discrete-time signals as a sequence,
and plot the corresponding waveform using MATLAB’s stem
function, sometimes referred to as a lollypop plot
>> n = 0:20;
>> x = 0.8.^n;
>> stem(n,x,'filled','b','LineWidth',2)
>> grid
>> xlabel('Time Index (n)')
>> ylabel('x[n]')
1

0.9
 0, n0
0.8 xn = 
 0.8 n , n  0
0.7

0.6
x[n]

0.5
Waveform values only at
0.4 integer sample values
0.3

0.2

0.1

0
0 2 4 6 8 10 12 14 16 18 20
Time Index (n)

ECE 2610 Signals and Systems 4–3

 Sampling

Sampling Sinusoidal Signals

• We will continue to find sinusoidal signals to be useful when
operating in the discrete-time domain
• When we sample (4.2) we obtain a sinusoidal sequence
x  n  = x  nT s 
= A cos  nT s +   (4.3)
= A cos  ̂n +  
• Notice that we have defined a new frequency variable

̂  T s = 
---- rad, (4.4)
fs
known as the discrete-time frequency or normalized continu-
ous-time frequency
– Note that ̂ has units of radians, but could also be called
radians/sample, to emphasize the fact that sampling is
involved
– Note also that many values of  map to the same ̂ value
by virtue of the fact that T s is a system parameter that is
not unique either
– Since  = 2f , we could also define ˆf  fT as the dis-
s
crete-time frequency in cycles/sample
Example: Sampling Rate Comparisons
• Consider x  t  = cos  2  60  t  at sampling rates of 240
and 1000 samples per second

ECE 2610 Signals and Systems 4–4

 Sampling

– The corresponding sample spacing values are

1 1
T s = --------- = 4.1666 ms T s = ------------ = 1ms
240 1000
>> t = 0:1/2000:.02;
>> x = cos(2*pi*60*t); % approx. to continuous-time
>> t240 = 0:1/240:.02;
>> n240 = 0:length(t240)-1;
>> x240 = cos(2*pi*60/240*n240); % fs = 240 Hz
>> axis([0 4.8 -1 1]) % axis scale since .02*240 = 4.8
>> t1000 = 0:1/1000:.02;
>> n1000 = 0:length(t1000)-1;
>> x1000 = cos(2*pi*60/1000*n1000); % fs = 1000 Hz
1

0.5
x(t)

0
Continuous
−0.5

−1
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Time (s)
1
4 samples/period
0.5
[n]
240

0
fs = 240 Hz
x

−0.5

−1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

1
16.67 samples/period
0.5
[n]
1000

0
x

−0.5 fs = 1000 Hz
−1
0 2 4 6 8 10 12 14 16 18 20
Sample Index (n)

ECE 2610 Signals and Systems 4–5

 Sampling

• The analog frequency is 2  60 rad/s or 60 Hz

• When sampling with f s = 240 and 1000 Hz
̂ 240 = 2  60  240 = 2  0.25  rad
̂ 1000 = 2  60  1000 = 2  0.06  rad
respectively
• The sinusoidal sequences are
x 240  n  = cos  0.5n 
x 1000  n  = cos  0.12n 
respectively
• It turns out that we can reconstruct the original x  t  from
either sequence
• Are there other continuous-time sinusoids that when sam-
pled, result in the same sequence values as x 240 and x 1000 ?
• Are there other sinusoid sequences of different frequency ̂
that result in the same sequence values?

The Concept of Aliasing

• In this section we begin a discussion of the very important
signal processing topic known as aliasing
• Alias as found in the Oxford American dictionary: noun
– A false or assumed identity: a spy operating under an alias.
– Computing: an alternative name or label that refers to a file, com-
mand, address, or other item, and can be used to locate or access it.

ECE 2610 Signals and Systems 4–6

 Sampling

– Telecommunications: each of a set of signal frequencies that, when

sampled at a given uniform rate, would give the same set of sampled
values, and thus might be incorrectly substituted for one another
when reconstructing the original signal.
• Consider the sinusoidal sequence
x 1  n  = cos  0.4n  (4.5)
– Clearly, ̂ = 0.4
• We know that cosine is a mod 2 function, so
x 2  n  = cos  2.4n 
= cos   2 + 0.4 n  = cos  0.4n + 2n  (4.6)
= cos  0.4n  = x 1  n 
– We see that ̂ = 2.4 gives the same sequence values as
̂ = 0.4 , so 2.4 and 0.4 are aliases of each other
• We can generalize the above to any 2 multiple, i.e.,
̂ l = ̂ 0 + 2l l = 0 1 2 3  (4.7)
result in identical frequency samples for cos  ̂ l n  due to the
mod 2 property of sine and cosine
• We can take this one step further by noting that since
cos    = cos  –   , we can write
x 3  n  = cos  1.6n 
= cos   2 – 0.4 n  = cos  2n – 0.4n  (4.8)
= cos  – 0.4n  = cos  0.4n 

ECE 2610 Signals and Systems 4–7

 Sampling

– We see that ̂ = 1.6 gives the same sequence values as

̂ = 0.4 , so 1.6 and 0.4 are aliases of each other
• We can generalize this result to saying
̂ l = 2l – ̂ 0 l = 0 1 2 3  (4.9)
result in identical frequency samples for cos  ̂ l n  due to the
mod 2 property and evenness property of cosine
– This result also holds for sine, expect the amplitude is
inverted since sin  –   = – sin   
• In summary, for any integer l, and discrete-time frequency
̂ 0 , the frequencies
̂ 0 ̂ 0 + 2l 2l – ̂ 0 l = 1 2 3  (4.10)
all produce the same sequence values with cosine, and with
sine may differ by the numeric sign
– A generalization to handle both cosine and sine is to con-
sider the inclusion of an arbitrary phase  ,
A cos   ̂ + 2l n +   = A cos  ̂n + 2l  n +  
= A cos  ̂n +  
A cos   2l – ̂ n –   = A cos  2l  n – ̂n –   (4.11)
= A cos  – ̂n –  
= A cos  ̂n +  
– Note in the second grouping the sign change in the phase
• The frequencies of (4.10) are aliases of each other, in terms
of discrete-time frequencies

ECE 2610 Signals and Systems 4–8

 Sampling

• The smallest value, ̂  [0 ) , is called the principal alias

• These aliased frequencies extend to sampling a continuous-
time sinusoid using the fact that ̂ = T s or  = ̂  T s
= ̂f s , thus we may rewrite (4.10) in terms of the continu-
ous-time frequency  0
 0  0 + 2lf s 2lf s –  0 l = 1 2 3 (4.12)
– In terms of frequency in Hz we also have
f 0 f 0 + lf s lf s – f 0 l = 1 2 3 (4.13)
• When viewed in the continuous-time domain, this means that
sampling A cos  2f 0 t +   with t  nT s results in
A cos  2f 0  nT s  +   = A cos  2  f 0 + lf s   nT s  +  
(4.14)
= A cos  2  lf s – f 0   nT s  –  
being equivalent sequences for any n and any l
Example: Input a 60 Hz, 340 Hz, or 460 Hz Sinusoid with
f s = 400 Hz
• The analog signal is
x 1  t  = cos  260t +   3 
x 2  t  = cos  2340t –   3 
x 3  t  = cos  2460t +   3 
• We sample x i  t  , i = 1, 2, 3 at rate f s = 400 Hz
>> ta = 0:1/4000:2/60; % analog time axis
>> xa1 = cos(2*pi*60*ta+pi/3);
>> xa2 = cos(2*pi*340*ta-pi/3);

ECE 2610 Signals and Systems 4–9

 Sampling

>> xa3 = cos(2*pi*460*ta+pi/3);

>> tn = 0:1/400:2/60; % discrete-time axis as n*Ts
>> xn1 = cos(2*pi*60*tn+pi/3);
>> xn2 = cos(2*pi*340*tn-pi/3);
>> xn3 = cos(2*pi*460*tn+pi/3);
f 1 = 60 Hz  f s = 400 Hz
1
Analog
0.5
Discrete
x1(t), x1[n]

0
All the sample
−0.5 values match
−1
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035
f 2 = 340 Hz  f s = 400 Hz
1

0.5
x2(t), x2[n]

−0.5

−1
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035
f 3 = 460 Hz  f s = 400 Hz
1

0.5
x3(t), x3[n]

−0.5

−1
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035
Sample Index × Ts = nTs (s)

• We have used (4.14) to set this example up, so we expected

the sample values for all three signals to be identical
• This shows that 60, 340, and 460, are aliased frequencies,
when the sampling rate is 400 Hz

ECE 2610 Signals and Systems 4–10

 Sampling

– Note: 400 - 340 = 60 Hz and 460 - 400 = 60 Hz

• The discrete-time frequencies are  i = 0.3 1.7 2.3
– Note: 2 – 1.7 = 0.3 rad and 2.3 – 2 = 0.3 rad

Example: 5 cos  7.3n +   4  versus 5 cos  0.7n +   4 

• To start with we need to see if either
7.3 = 0.7 + 2l
or 7.3 = 2l – 0.7
for l a positive integer
• Solving the first equation, we see that l = 3.3 , which is not
an integer
• Solving the second equation, we see that l = 4 , which is an
integer
0.7 7.3

̂
0  2 3 4 5 6 7 8
What are some other valid alias frequencies?

• The phase does not agree with (4.11), so we will use MAT-
LAB to see if 5 cos  0.7n +   4   5 cos  0.7n –   4  to
make the samples agree in a time alignment sense
>> n = 0:10; % discrete time axis
>> x1 = 5*cos(7.3*pi*n+pi/4);
>> x2 = 5*cos(0.7*pi*n+pi/4);
>> x3 = 5*cos(0.7*pi*n-pi/4);
>> na = 0:1/200:10; % continuous time axis
>> x1a = 5*cos(7.3*pi*na+pi/4);

ECE 2610 Signals and Systems 4–11

 Sampling

>> x2a = 5*cos(0.7*pi*na+pi/4);

>> x3a = 5*cos(0.7*pi*na-pi/4);
5 cos  7.3n +   4 
5
x [n]

Agree
0
1
but wrong phase
Samples match,

−5
0 1 2 3 4 5 6 7 8 9 10
5 cos  0.7n +   4 
5
x [n]

0
2

−5
0 1 2 3 4 5 6 7 8 9 10
5 cos  0.7n –   4 
5
x [n]

0
3

−5
0 1 2 3 4 5 6 7 8 9 10
Sample Index (n)

The Spectrum of a Discrete-Time Signal

• As alluded to in the previous example, a spectrum plot can be
helpful in understanding aliasing
• From the earlier discussion of line spectra, we know that for
each real cosine at ̂ 0 , the result is spectral lines at  ̂ 0
• When we consider the aliased frequency possibilities for a
single real cosine signal, we have spectral lines not only at
 ̂ 0 , but at all  2l frequency offsets, that is

ECE 2610 Signals and Systems 4–12

 Sampling

 ̂ 0  2l l = 0 1 2 3  (4.15)
• The principal aliases occur when l = 0 , as these are the only
frequencies on the interval [ –  )
Example: x  n  = cos  0.4n 
• The line spectra plot of this discrete-time sinusoid is shown
below
Principal
-2 Offset 1 Aliases 2 Offset
---
2

... ...

̂
– 3 – 2 –  – ̂ 0 0 ̂ 0  2 3

• A particularly useful view of the alias frequencies is to con-

sider a folded strip of paper, with folds at integer multiples of
 , with the strip representing frequencies along the  -axis
Folded Strip is the  Axis
̂
Principle
alias range
0 2 4

–  3
– 2  + ̂ 0 – ̂ 0 ̂ 0 2 – ̂ 0 ̂ 0 + 2 4 – ̂ 0

All of the alias frequencies are on a the same line when the paper
is folded like an accordian, hence the term folded frequencies.

ECE 2610 Signals and Systems 4–13

 Sampling

The Sampling Theorem

• The lowpass sampling theorem states that we must sample at
a rate, f s , at least twice that of the highest frequency in the
analog signal x  t  . Specifically, for x  t  having spectral con-
tent extending up to B Hz, we must choose f s = 1  T s  2B
Example: Sampling with f s = 2000 Hz
• When we sample an analog signal at 2000 Hz the maximum
usable frequency range (positive frequencies) is 0 to f s  2 Hz
• This is a result of the sampling theorem, which says that we
must sample at a rate that is twice the highest frequency to
avoid aliasing; in this case 1000 Hz is that maximum
• If the signal being sampled violates the sampling theorem,
aliasing will occur (see the figure below)
Principle Alias Frequency Range [0, fs/2]

f (Hz)


̂ = ----
0  2 fs
(rad)

ECE 2610 Signals and Systems 4–14

 Sampling

• An input frequency of 1500 Hz aliases to 500 Hz, as does an

input frequency of 2500 Hz
• The behavior of input frequencies being converted to princi-
ple value alias frequencies, continues as f increases
• Notice also that the discrete-time frequency axis can be dis-
played just below the continuous-time frequency axis, using
the fact that ̂ = 2f  f s rad
• We can just as easily map from the  –axis back to the con-
tinuous-time frequency axis via f = ̂f s   2 
• Working this in MATLAB we start by writing a support func-
tion
function f_out = prin_alias(f_in,fs)
% f_out = prin_alias(f_in,fs)
%
% Mark Wickert, October 2006

f_out = f_in;

for n=1:length(f_in)
while f_out(n) > fs/2
f_out(n) = abs(f_out(n) - fs);
end
end
• We now create a frquency vector that sweeps from 0 to 2500
and assume that f s = 2000 Hz
>> f = 0:5:2500;
>> f_alias = prin_alias(f,2000);
>> plot(f,f)
>> hold

ECE 2610 Signals and Systems 4–15

 Sampling

Current plot held

>> plot(f,f_alias,'r')
>> grid
>> xlabel('Frequency (Hz)')
>> ylabel('Apparent Frequency (Hz)')
>> print -tiff -depsc f_alias.eps
2500

Before sampling
2000
Apparent Frequency (Hz)

1500

1000

After sampling
with fs = 2000 Hz
500

0
0 500 1000 1500 2000 2500
Frequency (Hz)

Example: Compact Disk Digital Audio

• Compact disk (CD) digital audio uses a sampling rate of
f s = 44.1 kHz
• From the sampling theorem, this means that signals having
frequency content up to 22.05 kHz can be represented
• High quality audio signal processing equipment generally
has an upper frequency limit of 20 kHz
– Musical instruments can easily produce harmonics above
20 kHz, but human’s cannot hear these signals

ECE 2610 Signals and Systems 4–16

 Sampling

• The fact that aliasing occurs when the sampling theorem is

violated leads us to the topic of reconstructing a signal from
its samples
• In the previous example with f s = 2000 Hz, we see that tak-
ing into account the principle alias frequency range, the
usable frequency band is only [0, 1000] Hz

Ideal Reconstruction
• Reconstruction refers using just the samples x  n  = x  nT s 
to return to the original continuous-time signal x  t 
• Ideal reconstruction refers to exact reconstruction of x  t 
from its samples so long as the sampling theorem is satisfied
• In the extreme case example, this means that a sinusoid hav-
ing frequency just less than f s  2 , can be reconstructed from
samples taken at rate f s
• The block diagram of an ideal discrete-to-continuous (D-to-
C) converter is shown below

Ideal
yn D-to-C yt
Converter

1-
f s = ----
Ts

• In very simple terms the D-to-C performs interpolation on the

sample values y  n  as they are placed on the time axis at
spacing T s s

ECE 2610 Signals and Systems 4–17

 Sampling

– There is an ideal interpolation function that is discussed in

detail in Chapter 12 of the text
• Consider placing the sample values directly on the time axis
yt
y1 pulse width << Ts
y2
y  –2  y0
y3
–Ts 4T s 5T s 6T s
t
– 2T s Ts 2T s 3T s
y  –1  y6
y4 y5

• The D-to-C places the y  n  values on the time axis and then
must interpolate signal waveform values in between the
sequence (sample) values
• Two very simple interpolation functions are zero-order hold
and linear interpolation
• With zero-order hold each sample value is represented as a
rectangular pulse of width T s and height y  n 
– Real world digital-to-analog converters (DACs) perform
this type of interpolation
• With linear interpolation the continuous waveform values
between each sample value are formed by connecting a line

ECE 2610 Signals and Systems 4–18

 Sampling

between the y  n  values

y  –2  yt Similar to actual
y1
y2 y3 DAC output
Zero-Order y0
Hold Interp. 4T s
–Ts 5T s 6T s
t
– 2T s Ts 2T s 3T s
y  –1 
y4 y5 y6

yt
y1
y2
Linear y  –2  y0
Interp.
y3
–Ts 4T s 5T s 6T s
t
– 2T s Ts 2T s 3T s
y  –1  y6
y4 y5

• Both cases introduce errors, so it is clear that something bet-

ter must exist
• For D-to-C conversion using pulses, we can write

yt =  y  n p  t – nT s  (4.16)
n = –
where p  t  is a rectangular pulse of duration T s
• A complete sampling and reconstruction system requires
both a C-to-D and a D-to-C

xt Ideal xn Direct yn Ideal yt

C-to-D Connection D-to-C
Converter DSP System Converter
yn = xn
fs fs

ECE 2610 Signals and Systems 4–19

 Sampling

• With this system we can sample analog signal x  t  to pro-

duce x  n  , and at the very least we may pass x  n  directly to
y  n  , then reconstruct the samples y  n  into y  t 
– The DSP system that sits between the C-to-D and D-to-C,
should do something useful, but as a starting point we con-
sider how well a direct connection system does at returning
y t  x t
– As long as the sampling theorem is satisfied, we expect
that y  t  will be close to x  t  for frequency content in x  t 
that is less than f s  2 Hz
– What if some of the signals contained in x  t  do not satisfy
the sampling theorem?
– Typically the C-to-D is designed to block signals above
f s  2 from entering the C-to-D (antialiasing filter)
– A practical D-to-C is designed to reconstruct the principle
alias frequencies that span
̂   –     f   – f s  2 f s  2  (4.17)

ECE 2610 Signals and Systems 4–20

 Spectrum View of Sampling and Reconstruction

Spectrum View of Sampling and Reconstruc-

tion
• We now view the spectra associated with a cosine signal
passing through a C-to-D/D-to-C system
• Assume that x  t  = cos  2f 0 t 
• The sampling rate will be fixed at f s = 2000 Hz
f s = 2000 Hz Spectrum of 500 Hz Cosine x(t)
1
---
2

f
– 2000 – 1000 -500 0 500 1000 2000 (Hz)
Spectrum of Sampled 500 Hz Cosine x[n] = y[n]
Alias Alias 1 Alias Alias
---
2

(rad)
̂
–2  –  -0.5 0 0.5  2
f
– 2000 – 1000 0 1000 2000 (Hz)
Spectrum of Reconstructed 500 Hz Cosine y(t)
1
--- Reconstruction
2
Band

f
– 2000 – 1000 -500 0 500 1000 2000 (Hz)

ECE 2610 Signals and Systems 4–21

 Spectrum View of Sampling and Reconstruction

f s = 2000 Hz Spectrum of 1500 Hz Cosine x(t)

1
---
2

f
– 2000 – 1000 -500 0 500 1000 2000 (Hz)
Spectrum of Sampled 1500 Hz Cosine x[n] = y[n]
Alias Alias 1
--- Alias Alias
2

(rad)
̂
– 2  -1.5 –  -0.5 0 0.5  1.5 2
f
– 2000 – 1000 0 1000 2000 (Hz)
Spectrum of Reconstructed 1500 Hz Cosine y(t)
1
---
2 Reconstruction
Band

f
– 2000 – 1000 -500 0 500 1000 2000 (Hz)
• We see that the 1500 Hz sinusoid is aliased to 500 Hz, and
when it is output as y  t  , we have no idea that it arrived at the
input as 1500 Hz
• What are some other inputs that will produce a 500 Hz out-
put?

The Ideal Bandlimited Interpolation

• In Chapter 12 of the text it is shown that ideal D-to-C conver-
sion utilizes an interpolating pulse shape of the form
sin  t  T s 
pt = - –   t  
-------------------------- (4.18)
 t  T s 

ECE 2610 Signals and Systems 4–22

 Spectrum View of Sampling and Reconstruction

• The function sin  x    x  is known as the sinc function

• Note that interpolation with this function means that all sam-
ples are required to reconstruct y  t  , since the extent of p  t 
is doubly infinite
– In practice this form of reconstruction is not possible

pt 1

0.8
Zero crossings
0.6 at integer multiples
of Ts
0.4

0.2
t-
----
-7.5 -5 -2.5 2.5 5 7.5 T s
-0.2

• A Mathematica animation showing that when the sinc()

pulses are weighted by the sample values, delayed, and then
summed, high quality reconstruction (interpolation) is possi-
ble
– The code used to create the animation
Manipulate@
Show@Plot@Cos@2 p f t + fD, 8t, 0, 10<, PlotStyle Ø [email protected], RGBColor@1, 0, 0D<D,
Plot@Sum@Cos@2 p f n + fD Sinc@p Ht - nLD, 8n, 0, 10<D, 8t, 0, 10<,
PlotStyle Ø 8Thick, RGBColor@0, 1, 0D<D, DiscretePlot@Cos@2 p f n + fD,
8n, 0, 10<, Filling Ø Axis, PlotStyle Ø [email protected], RGBColor@1, 0, 0D<D,
Plot@Table@Cos@2 p f n + fD Sinc@p Ht - nLD, 8n, 0, 10<D, 8t, 0, 10<, PlotRange Ø AllD,
PlotRange Ø AllD,
88f, 1 ê 10<, 81 ê 2, 1 ê 3, 1 ê 4, 1 ê 10, 1 ê 20<<, 88f, 0<, 80, p ê 4, p ê 2<<D

ECE 2610 Signals and Systems 4–23

 Spectrum View of Sampling and Reconstruction

• The final display showing an interpolated output for a single

sinusoid
– The input signal is x  t  = cos  2ft +   and we assume
that T s = 1 , so in sampling we let t  n
– With f = 1  4 (4 samples per period) and  =   4 we
have the following display:

1 1 1 1 1
f Green thick = intepolated output
2 3 4 10 20
Red dashed = input
Blue thin = sinc interpolated samples
p p
f 0 Red points = sample value
4 2

ECE 2610 Signals and Systems 4–24