EE338 - Digital Signal Processing and its Applications

Vikram M. Gadre

1 Introduction 2.1 “What is a satisfactory way in which

we can discretize a signal?”
• Def- Continuum: A set where one has a ‘distance’
notion and no matter how small the distance between • By satisfactory, we mean, not to lose critical informa-
two elements, there is another with a smaller distance tion
from both. • By discretize, we mean, use only the values of the
signal at chosen values of the independent variable.
• Continuum could be more than one dimension.

• In real life situations, mostly, both independent and 2.2 Choosing the value of independent
dependent variable are continuous. variable for discretization
• Start with a single sine wave, if that is conquered,
• Here, we aim to deal with them by and equivalent Fourier Series will help to do the rest.
signal where independent variable is DISCRETE and
dependent is CONTINUOUS. x(t) = A0 cos(ω0 + φ0 )

• Sampling at t=nT,
• Equivalent implies we do not lose any information in
the process of discretization. x(nT ) = x[n] = A0 cos(ω0 nT + φ0 )
= A0 cos(ω0 nT + φ0 + 2πnk)
• So, when a variable is discrete, we have a notion of
= A0 cos(2πnk − ω0 nT − φ0 )
nearest sample.
n, k ∈ Z
• Def- Signal: Relation between an independent and
• Here, although ambiguity is introduced, but no infor-
a dependent variable.
mation is lost.
• Def- Signal Processing: Modification and separa- x[n] = A0 cos(ω0 nT + φ0 )
tion of signals. Example, in an audio, it is a 2π
mixture of male and female voices, enhance a sin- = A0 cos(( k + ω0 )nT + φ0 )
T
gle component(Modification), get a single compo- 2π
nent(Separation). = A0 cos(( k − ω0 )nT − φ0 )
T
• The above equations show that x[n] could’ve come
from multiple possible frequencies.
2 Discretization of continuous inde-
• A numerical example, let,
pendent variables
T = 1ms → 1/T = 1KHz
• Sine wave is a common function to break a signal into ω0 = 2π(0.1)KHz
components as it is sort of the smoothest wave that
can be found in nature. It is persistent and its fre- Here, the original sinusoid has a frequency of 0.1KHz
quency remains same for a large no. of operations. and rate of sampling is 1KHz. So, the samples
obtained could’ve come from (A0 , φ0 , 0.1KHz) or
• For example, adding 2 triangular or square waves with (A0 , −φ0 , 0.9KHz) or (A0 , φ0 , 1.1KHz)...
same frequency, the result is not a wave with the same • A little visual example,
frequency.

• Taking an example of the sine waves,

A1 cos(ω0 t + φ1 ) + A2 cos(ω0 t + φ2 )
= (A1 cos φ1 + A2 sin φ2 ) cos ω0 t− • Adding all the possible sinusoids that could give the
(A1 sin φ1 + A2 sin φ2 ) sin ω0 t same samples, we see that it is constructive at points
= A3 cos(ω0 t + φ) of sampling and destructive at others. (TRY TO
PROVE THIS)
A3 =??, φ =??
• Let original signal be x(t), the sampled signal be xs (t).
• We use Fourier Transform to regard a continuous sig- • On frequency axis, sampling creates ”copies” of
nal x(t) as (possibly infinite) sum or linear combina- the original signal at a collection of frequen-
tion of sine waves of frequency varying, in principle, cies. Consider 3 sine waves sampled at 1KHz
from zero to infinity. (A1 , φ1 , 0.1KHz)(A2 , φ2 , 0.13KHz), (A3 , φ3 , 0.15KHz)
then, the following copies will be generated, the orig-
• Shannon’s sampling theorem: Only band-limited sig- inal ones, (A3 , −φ3 , 0.85KHz)(A2 , −φ2 , 0.87KHz),
nals can be discretized without any loss of informa- (A1 , −φ1 , 0.9KHz), (A1 , φ1 , 1.1KHz)(A2 , φ2 , 1.13KHz),
tion. (Pretty evident from the sampling theorem) (A3 , φ3 , 1.15KHz) and so on.
• In the spectral domain, reconstructing x(t) 3 Linear and Shift Invariant (LSI)
from xs (t): retain the original signal, and re-
move the all the copies and mirror images that
Systems
appear after sampling. • Linearity: Principle of superposition
• We can do this if neither the copies nor the i.e., if x1 [n] → y1 [n] and x2 [n] → y2 [n], then
mirror images ”pollute” or ”overlap” the orig- αx1 [n] + βx2 [n] → αy1 [n] + βy2 [n]
inal signal spectrum. ∀x1 , y1 , x2 , y2 , α, β ∈ C

• Linearity = Additivity + Homogeneity

• The trouble maker is essentially the first mirror im-
age. Let the maximum signal frequency be fn and • Additivity: Linearity where α, β = 1
the sampling frequency be fs . So the first copy occu-
pies fs → fs + fn and the first mirror image occupies • Homogeneity/scaling: Linearity where x2 [n] = 0 ∀ n
fs − fn → fs . Hence, fs − fn > fn or,
• Example: linear system, y[n] = αx[n] + βx[n − 1]
fs > 2fn
• Example:additive but not homogeneous, y[n]=Re{x[n]}
This is known as Nyquist Principle of Sampling • Example: homogeneous but not additive
or the Sampling Theorem y[n] = x2 [n]/x[n − 1]
• Doubt: So we prove that adding the possible signals • PROVE THAT ADDITIVITY + HOMOGENEITY
from the sample return a signal similar to the sam- = LINEARITY
ple, only with infinite value at non zero points. So
this added signal has all possible frequencies, but we • Shift Invariance:
pass the sampled signal through the low pass filter, If x[n] → y[n], then x[n − n0 ] → y[n − n0 ] ∀ x[n],
so essentially the amplitude of the original signal fre- n0 ∈ Z. Example, y[n] = 3x[n] +4x[n-1]
quencies we obtain will tend to zero right?
• Example that is NOT shift invariant, y[n]= nx[n]
• If this is not true, we have Aliasing. Alias means
• If a system is both linear and shift invariant, only one
false name or identity.
experiment is required to characterize the complete
• Food for thought: Consider the frequency range of system.
3-5KHz, can we sample at less then 10KHz? • Theorem: A LSI system is completely characterized
Sampling at 5 KHz should work fine as wee, only that by its response to a unit impulse sequence δ[n][P ] .
we get the mirror image and frequencies will have to
be modified. • Characterization means being able to determine the
output given ANY input.
• While designing systems for processing, we need them
to be ”impartial” wrt the original phases and ampli- • The output sequence is obtained by some opera-
tudes on the frequency axis, and we see that we need tion between the input and unit impulse response se-
Linearity and shift-invariance. quence, we call it the convolution.
∞ ∞
P S P
• A discrete system is a relationship between all samples x[k]δ[n − k] −→ x[k]h[n − k] = y[n]
k=−∞ k=−∞
y[n], for all integer n, and all samples x[n] in general.
So it is NOT a point to point relation. One interpretation for this summation is, we construct
many sequences and add them together to obtain y[n].
• We wish to decouple dealing with sine waves of dif-
ferent frequencies, i.e. if the input is a sine wave, we Or, we can also obtain y[n] point by point for each n.
expect to get a sine wave of the same frequency. We think of h[n-k] as a sequence in k with fixed n.
We shift h[k] backward by n and then take the mirror
• Sequences: Functions of an integer variable (ex. sam- image of the sequence. We multiply point wise with
pling index) x[k] and the summation is the value at y[n].
It can also be seen as x[k] being static(platform) and
• Complex Functions are allowed. Phasors help to keep
h[n-k] being dynamic(train) and moved.
the system as a constant ratio even after the phase as
well as amplitude is changed.

• Phasor: A complex no. rotating with a single angular

frequency ω0 . Also, the real and imaginary parts are
sinusoids with angular frequency ω0 .

• For the above mentioned properties, we need the sys-

tem to be

– Linear
– Shift Invariant
– Stable
 Appendix: Proofs, Calculations and Challenges
1. Proof that a LSI system is completely characterized
by its response to a unit impulse sequence:
∞
P
x[n]= x[k]δ[n − k]
k=−∞
Now assume that the unit impulse (δ[n]) response of
the LSI system is given by h[n].
Using the shift invariance property,
S
δ[n − k] −
→ h[n − k]
Using the homogeneity property,
S
x[k]δ[n − k] −
→ x[k]h[n − k]
Note here that h[n-k] is a sequence and x[k] is some
constant.
Using the additivity property,
∞ ∞
P S P
x[k]δ[n − k] −
→ x[k]h[n − k] = y[n]
k=−∞ k=−∞
Notice, the proof is constructive.