Lecture 2: Non-uniqueness Due to Sampling

Prof. Vikram M Gadre

EE210.1X
Indian Institute of Technology Bombay
April 2, 2015

IIT Bombay, 2015

This work is licensed under a Creative Commons
Attribution-ShareAlike 4.0 International License.
IITBombayX EE210.1X

1 Introduction
In previous lecture, we used a-priori information for filling information between
different samples. For example take a exponential signal

x(t) = Aet

We can do a unique reconstruction by just doing two measurements and filling

everything in between by using a priori information of x(t). Here we will look
at what happens when we dont have a-priori information for reconstruction.

2 Non-uniqueness due to lack of a-priori Infor-

mation
Lets sketch the reconstruction of x(t) for two measurements

x(t1 ) = Aet1

x(t2 ) = Aet2
and t1 6= t2 . If we have, as a-priori information, the form of the function (expo-
nential in this case) we can fill everything in between t1 and t2 and elsewhere
using that information (black line in Figure 1). Now if we know the same sam-
ples x(t1 ) and x(t2 ) and the a-priori form is not given, how will we proceed with
the reconstruction? The problem is not that there are no possibilities but that
there are too many of them. Some of them are shown in Figure 1 along with
the reconstruction when the given form is exponential.

Figure 1: Non-unique reconstruction due to lack of a-priori Information

We can say the problem of sampling and reconstruction is similar to solving

a crossword. We can think of clues as a-priori information and letters provided
as samples. Without any a-priori information we can fill the crossword any way
we want but with clues we are restricted in our choice of filling up the crossword.

2
IITBombayX EE210.1X

Crosswords are generally designed such that there is a unique solution. Similarly
in case of sampling and reconstruction a strong a-priori information restricts us
to a unique reconstruction.

2.1 Degree of non-uniqueness in reconstruction

When we know nothing about the form of signal we see there is an infinity
of possibilities. To answer how big this infinity is, we must first see different
classes of infinity. There is an infinity of natural numbers, infinity of integers,
positive even/odd numbers etc. which are all same. Then there is an infinity
of real numbers which is bigger or higher than the infinity of integers in the
sense that we cant put real numbers in one to one correspondence to integers
without missing some real numbers. In case of reconstruction we need to define
a value at every point. Some finite number of points have fixed value which are
basically the samples but there is an infinity of points which dont have fixed
value and can take any value in the absence of a-priori information. This infinity
of possible reconstructions is thus even higher than infinity of real numbers. Now
we can understand the degree of non-uniqueness of this problem.
How uniquely we can solve this problem of reconstruction depends on whether
the given a-priori information completely complements the samples we have. We
have previously seen that using Fourier transform we can represent a signal in
terms of sinusoid. Let us take a sine wave and sample it and see what ambigu-
ities we create.

3 Sampling Sinusoid
Let us take a single sine wave and sample it uniformly i.e sampling with equal
spacing between adjacent points. Now we need to find what are all sine waves
which have these samples. The problem of ambiguity arises because we can
miss an integer number of cycles between two samples. We can also arrive at
a wrong edge. So with each possibility of losing a integer number of cycles one
can also arrive at the right or the wrong edge. Lets analyse this mathematically
by taking
Ao cos(o t + o )
as the form of sinusoid where o is the angular frequency and To is the time
period
2
o =
To
Now the process of sampling creates many ghost or monster frequencies
which have the same samples at the same points. Lets denote these frequencies
by kj where j (j = 1, 2) corresponds to two possibilities of correct or wrong
edge.
x(t) = Ao cos(kj t + kj )
k = 1, 2, 3, ...
When x(t) is sampled at time nTs where n belongs to set of integers, we get

x[nTs ] = Ao cos(o nTs + o )

3
IITBombayX EE210.1X

The non-uniqueness comes from being able to add a phase to this. Lets add
2kn to the phase where n is the sampling instant and k belongs to set of
positive integers.
Ao cos(o nTs 2kn + o )
2
= Ao cos(nTs (o k) + o )
Ts
1 k
= Ao cos(2nTs ( ) + o )
To Ts
We need to have more than one sample in a cycle. We can see importance of
this by taking a extreme example where we take just one sample per cycle and
at the same location in every cycle. We will get same sample in every cycle. We
cant even tell whether it is a DC signal (a constant signal) or a sinusoid. Thus
we can say that time interval between the samples must be less than cycle time.
( T1o Tks ) represents all possible Hertz frequencies. Lets take the possibility
of k = 1
1 1
( )<0
To Ts
as Ts < To but we dont want negative frequencies. Lets go back to original
expression and see how we can correct this.
1 k
Ao cos(2nTs ( ) + o )
To Ts
1 k
= Ao cos((2nTs ( ) + o ))
To Ts
If we take + we get the same expression and if we take we get
k 1
= Ao cos(2nTs ( + ) o )
Ts To
we see that the phase is reversed.

3.1 Ghost Frequencies: An Example

Lets take the case of k = 1 again. Take + for ( T1s + 1
To )

1 1
Ao cos(2nTs ( + ) + o )
To Ts
Take for ( T1s + 1
To )

1 1
Ao cos(2nTs ( ) o )
Ts To
To
Take Ts = 4
1 1 5
+ =
To Ts To
1 1 3
=
Ts To To
Lets analyse this situation graphically. Take To = 2 . In Figure 2, the black
reconstruction is the true sinusoid and red is the sinusoid with ghost frequency
= T5o

4
IITBombayX EE210.1X

Figure 2: True sinusoid and Ghost frequencies

Exercise : Draw the sinusoid when you almost miss the cycle and arrive on
the wrong edge and verify the ghost frequency = T3o and the spacing to be 3
2