Sampling and Aliasing

Lecture #6
Chapter 4

BME 310 Biomedical Computing - 152

J.Schesser
What Is this Course All About ?
• To Gain an Appreciation of the
Various Types of Signals and Systems
• To Analyze The Various Types of
Systems
• To Learn the Skills and Tools needed
to Perform These Analyses.
• To Understand How Computers
Process Signals and Systems
BME 310 Biomedical Computing - 153
J.Schesser
 Discrete-time Signals and Computers
• Up to now we have been studying continuous-time signals (also
called analog signals) such as

x(t )  A cos(ot   )
• However, digital computers and computer programs can not
process analog signals.
• Instead they store discrete-time versions of analog signals
x[n]  x(nTs )
• This is because digital computers can only store discrete
numbers.
– There are computers called analog computers which do process
continuous-time signals
• Since the computer only stores numbers, how does one know
what continuous-time signal it represents?
BME 310 Biomedical Computing - 154
J.Schesser
 Sampling
• We can obtain a discrete-time signal by sampling a
continuous-time signal at equally spaced time
instants, tn = nTs
x[n] = x(nTs) -∞ < n < ∞
• The individual values x[n] are called the samples of
the continuous time signal, x(t).
• The fixed time interval between samples, Ts, is also
expressed in terms of a sampling rate fs (in samples
per second) such that:
fs = 1/ Ts samples/sec.

BME 310 Biomedical Computing - 155

J.Schesser
Continuous-to-Discrete Conversion
• By using a Continuous-to-Discrete (C-to-D)
converter, we can take continuous-time signals and
form a discrete-time signal.
• There are devices called Analog-to-Digital converters
(A-to-D)
• The books chooses to distinguish an C-to-D converter
from an A-to-D converter by defining a C-to-D as an
ideal device while A-to-D converters are practical
devices where real world problems are evident.
– Problems in sampling the amplitudes accurately
– Problems in sampling at the proper times
BME 310 Biomedical Computing - 156
J.Schesser
 Discrete-Time Signals
• A discrete-time signal is
a sequence of numbers
and carry no 1

information about the 0.8

0.6

time-sequence.
0.4
0.2
0
-0.2 0 2 4 6 8 10

• Looking at the -0.4

-0.6

following diagram,
-0.8
-1

which (gray or solid)

waveform are these
(red) samples associated
with?
BME 310 Biomedical Computing - 157
J.Schesser
 Discrete-Time Sinusoidal Signals
• Since a Fourier series can be written for any continuous-time
signal, let’s concentrate on sinusoids
• We define a normalized frequency for the discrete sinusoidal
signal. x[n]  x(nTs )  A cos( nTs   )
 A cos(ˆ n   )

ˆ  Ts 
fs
• ̂ is the normalized or discrete-time frequency
• Since we can have different signals with the same ̂ , then
there can be an infinite number of continuous-time signal
which yield the same discrete-time sinusoid!

BME 310 Biomedical Computing - 158

J.Schesser
 Two Problems with Sampling
• Problem 1: How many samples are enough to
have to represent a continuous time signal?
1

0
0 2 4 6 8 10

-1

• In this figure, we have a continuous-time signal

sampled every .4 seconds (red samples) and
every 1 second (black samples).
BME 310 Biomedical Computing - 159
J.Schesser
 Discrete-Time Sinusoidal Signals
• Problem 2: Can a set of samples be represent
more that one continuous-time signal
• The discrete-time sinusoid
shown in the figure has 1

which can be obtain from, for

example, either a 1 second
sampled continuous-time 0

sinusoid with f = 0.2 Hz or 1.2 0 2 4 6 8 10

Hz.
• In the first case, where f = 0.2 -1

Hz, we have:
ˆ  2 (0.2)(1)  .4

BME 310 Biomedical Computing - 160

J.Schesser
 Discrete-Time Sinusoidal Signals

• In the first case, where f = 0.2

Hz, we have: 1

ˆ  2 (0.2)(1)  .4
• Since a sinusoid is periodic in 0
0 2 4 6 8 10

2, then for the case where

f=1.2 Hz
-1

x[n]  A cos(ˆ n   )
ˆ  2 (1.2)(1)  2.4  2.4  2  .4
x[n]  A cos(2.4 n   )  A cos(2 n  0.4 n   )  A cos(0.4 n   )
BME 310 Biomedical Computing - 161
J.Schesser
 Aliasing
• This example
illustrates that two
1

sampled sinusoids can

produce the same 0
0 2 4 6 8 10

discrete-time signal.
1. cos [2π(0.2) t]
-1

2. cos [2π(1.2) t]
• When this occurs we say that that these signals
are aliases of each other.

BME 310 Biomedical Computing - 162

J.Schesser
 Aliasing
• There are more alias signals for this example:
1. x(t) = cos (2π(0.2) t) => x[n] = cos (2π(0.2) 1n) = cos (0.4π n)
2. x(t) = cos (2π(1.2) t) => x[n] = cos (2π(1.2) 1n) = cos (2.4π n) =
cos (0.4 π n + 2 π n) = cos (0.4π n)
ˆ l  0.4  2l for l  0,1,2,3,
Since cos(2 -  )  cos( ), ˆ l  0.4  2l for l  0,1,2,3,
3. x(t) = cos (2π(.8) t) => x[n] = cos (2π(.8) 1n) = cos (1.6π n) =
cos (2 π n - 0.4 π n) = cos (0.4π n)
• In summary, (for l = positive 1

or negative integer):
0

̂ o, ̂ o  2 l, 2 l  ̂ o
0 2 4 6 8 10

where ̂ o is called the principal alias

-1

BME 310 Biomedical Computing - 163

J.Schesser
 Aliasing
• Let’s look at signals of the form: cos(ωlt)
cos(l t )  cos(ˆ l n)
sampled

ˆ l
where l   ˆ l f s and ˆ l  ˆ o  2l , ˆ o is the principal alias, and l is an integer.
Ts
(ˆ o  2l ) f s
Therefore, l  ˆ l f s  (ˆ o  2l ) f s and f l 
2
(2l  ˆ o ) f s
since cos( )  cos( )  cos(2   ), then we can have ˆ l  2l  ˆ o and f l 
2
(2l  ˆ o ) f s
or l  2l  ˆ o and f l 
2

In our example, ˆ o is 0.4 and f s  1. Then,

l  2l  ˆ o  2l  0.4  0.4 , 1.6 , 2.4 , 3.6 , ....rad/sec
(2l  ˆ o ) f s 2l  0.4
fl    l  0.2  0.2 , 0.8 1.2 , 1.8, ..... Hz
2 2
BME 310 Biomedical Computing - 164
J.Schesser
 Shannon’s Sampling Theorem
• How frequently do we need to sample?
• The solution: Shannon’s Sampling Theorem:
A continuous-time signal x(t) with frequencies
no higher than fmax can be reconstructed
exactly from its samples x[n] = x(nTs), if the
samples are taken a rate fs = 1 / Ts that is
greater than 2 fmax.
• Note that the minimum sampling rate, 2 fmax ,
is called the Nyquist rate.
BME 310 Biomedical Computing - 165
J.Schesser
 Spectrum of the Discrete-time
Signal
0.5

-2.4 -1.6 -0.4 0.4 1.6 2.4

• There are an infinite number of frequency

components of discrete-time signal
• They consists of the principal along with the
other aliases (an infinite number of them).
BME 310 Biomedical Computing - 166
J.Schesser
 Nyquist Rate
• Shannon’s theorem tell us that if we have at
least 2 samples per period of a sinusoid, we
have enough information to reconstruct the
sinusoid.
• What happens if we sample at a rate which is
less than the Nyquist Rate?
– Aliasing will occur!!!!

BME 310 Biomedical Computing - 167

J.Schesser
 Ideal Reconstruction
• The sampling theorem suggests that a process exists
for reconstructing a continuous-time signal from its
samples.
• If we know the sampling rate and know its spectrum
then we can reconstruct the continuous-time signal by
scaling the principal alias of the discrete-time signal
to the frequency of the continuous signal.
• The normalized frequency will always be in the range
between 0 ~ π and be the principal alias if the
sampling rate is greater than the Nyquist rate.

BME 310 Biomedical Computing - 168

J.Schesser
Ideal Reconstruction Continued
• If continuous-time signal has a frequency of ω, then the discrete-time
signal will have a principal alias of

ˆ  Ts 
fs
• So we can use this equation to determine the frequency of the
continuous-time signal from the principal alias:
ˆ
  ˆ f s 
Ts
• Note that the normalized frequency must be less than  if the Nyquist
rate is used
2f MAX 2f MAX
ˆ   MAX Ts  2f MAX Ts   
fs f s ( 2 f MAX )
• And the reconstructed continuous-time frequency must be
ˆ f s  f s f s
  2 f  ˆ f s  f     f max
2 2 2
BME 310 Biomedical Computing - 169
J.Schesser
 Oversampling
• When we sample at a rate which is greater than the Nyquist
rate, we say we are oversampling.
• If we are sampling a 100 Hz signal, the Nyquist rate is 200
samples/second => x(t)=cos(2π(100)t+π/3)
• If we sample at 2.5 times the Nyquist rate, then fs = 500
samples/sec
• This will yield a normalized frequency at 2π(100/500) = 0.4π
1

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

-1

BME 310 Biomedical Computing - 170

J.Schesser
 Oversampling
• Since we are greater than the Nyquist rate, the normalized
frequency will be < π which means it is the principal alias.
• And we get back the original continuous frequency when we
do the reconstruction
• f = 0.4πfs / 2π = 0.4π500 / 2π = 0.2 (500) = 100 Hz

0.5

-2.4π -1.6π -0.4π 0.4π 1.6π 2.4π

BME 310 Biomedical Computing - 171

J.Schesser
 Undersampling and Aliasing
• When we sample at a rate which is less than the Nyquist rate,
we say we are undersampling and aliasing will yield
misleading results.
• If we are sampling a 100 Hz signal, the Nyquist rate is 200
samples/second => x(t)=cos(2π(100)t+π/3)
• If we sample at .4 times the Nyquist rate, then fs = 80 s/sec
• This will yield a normalized frequency at 2π(100/80) = 2.5π
1

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

-1
BME 310 Biomedical Computing - 172
J.Schesser
 Undersampling
• Since it is > π, 2.5π is NOT the principal alias
• The principle alias is 2.5π - 2π = 0.5π

• Using 0.5π as the principal alias and performing a reconstruction we then

have:
f = 0.5π fs / 2π = 0.5π 80 / 2π = 0.5 (40) = 20 Hz and we have
reconstructed the wrong signal!!!

0.5

-2.5π -1.5π -0.5π 0.5π 1.5π 2.5π

BME 310 Biomedical Computing - 173

J.Schesser
 The Alias Problem due to
Undersampling

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

-1

BME 310 Biomedical Computing - 174

J.Schesser
 Aliasing and Folding
• Your book treats undersampling in terms of
aliasing and folding
• During reconstruction, both of these
phenomenon will produce erroneous results.
• The difference between aliasing and folding
has to do with which part of the spectrum
created the alias.

BME 310 Biomedical Computing - 175

J.Schesser
Discrete-to-Continuous Conversion
• An D-to-C converter uses the samples to
reconstruct the continuous-time signal by
interpolation.
• There are various interpolation algorithms
which may be used:
– Zero-Order Hold
– Linear
– Cubic Spline

BME 310 Biomedical Computing - 176

J.Schesser
 Interpolation
1 1

0 0
0 2 4 6 8 10 0 2 4 6 8 10

-1 -1

Zero-Order Hold Linear

• Oversampling always improves the reconstruction

• Best reconstruction is Low Pass Filter or what the text
calls: Ideal Bandlimited Interpolation
BME 310 Biomedical Computing - 177
J.Schesser
 Non-sinusoidal Signals
• Since a Fourier series can be written for any
continuous-time signal, the sampling and
reconstruction processes for any continuous-
time signal is the same
– Shannon’s Sampling theorem
– Nyquist Rate fs ≥ 2fmax to eliminate aliasing
– Oversampling to improve interpolation
– Ideal (low pass filter) Bandlimited interpolation

BME 310 Biomedical Computing - 178

J.Schesser
 Homework
• Exercises:
– 4.1 – 4.5
• Problems:
– 4.1, 4.2, 4.3,
– 4.4, Use Matlab to plot the signal in part a.
– 4.5, 4.8, 4.11, 4.19

BME 310 Biomedical Computing - 179

J.Schesser