INDEX

S. No Topic Page No.

Week 1
1 Fourier Series 1

2 Fourier Series (Contd.) 10

3 Fourier Series (Contd.) 21

4 Fourier Series (Contd.) 30

5 Fourier Series (Contd.) 49

6 Fourier Series (Contd.) 68

7 Fourier Series (Contd.) 84

Week 2
8 Fourier Transform 98

9 Fourier Transform (Contd.) 116

10 Fourier Transform (Contd.) 134

11 Fourier Transform (Contd.) 159

Week 3
12 Energy Spectral Density 175

13 Power Spectral Density 194

14 PSD of Random Signal 207

Week 4
15 Amplitude Modulation 219

16 Amplitude Modulation (Contd.) 233

17 Amplitude Modulation (Contd.) 245

18 Amplitude Modulation (Contd.) 259

Week 5
19 SSB - SC 275

20 SSB - SC (Contd.) 290

21 LVSB-SC 301

22 VSB-SC (Contd.) 315

23 Effect of Carrier Synchronization 334

24 Comparison of Different Modulation Technique 346

 Week 6
25 PLL 357

26 PLL (Contd.) 369

27 PLL (Contd.) 382

28 PLL (Contd.) and LTI 394

29 Dispersion 407

30 Channel Nonlinearities and Multipath Effects 425

Week 7
31 Probability Theory 439

32 Probability Theory (Contd.) 452

33 Probability Theory (Contd.) 467

34 Probability Theory (Contd.) 479

35 Probability Theory (Contd.) 492

36 Probability Theory (Contd.) 505

37 Probability Theory (Contd.) 520

Week 8
38 Random Process 532

39 Random Process (Contd.) 545

40 Random Process (Contd.) 564

41 Random Process (Contd.) 580

42 Random Process (Contd.) 593

43 Random Process (Contd.) 610

Week 9
44 Noise Analysis - DSB-SC 625

45 Noise Analysis - AM 637

46 Noise Analysis - SSB-SC 651

Week 10
47 Frequency Modulation 663

48 Frequency Modulation (Contd.) 676

49 Frequency Modulation (Contd.) 688

50 Frequency Modulation (Contd.) 699

51 Frequency Modulation (Contd.) 712

52 Frequency Modulation (Contd.) 726

Week 11
53 FM Noise Analysis 741

54 FM Noise Analysis (Contd.) 756

55 FM Noise Analysis (Contd.) 775

Week 12
56 Sampling Theorem 789

57 Sampling Theorem (Contd.) 803

58 FDM Vs TDM 816

59 Flat Top Vs Natural Sampling 828

60 Pulse Coded Modulation 842

 NPTEL

NPTEL ONLINE CERTIFICTAION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 01: Fourier Series

Okay, so this is the course of analog communication. So what we will do today this, for the first
class? So we will briefly discuss about what should be the outline of this particular course? What
are the things that should be generally conveyed through this course? And, what are the concepts
that we will be builded up in this courses? So that should be our first target. So, initial one hour
probably we will be spending on that. And basically, we will try to capture what we mean by
communication as such.

And where it is required, what are the basic terminology that is being used in communication.
Why those things are required? And then, we will go into the depth of communication, and
especially the analog communication. Okay.
(Refer Slide Time: 01:07)

So, whenever we talk about analog communication, before even talking about analog
communication, let us try to see what we mean by communication? So communication generally
means that we have some source of information. So if I just say some source of information, that
might be generated from any kind of a source. Like, it might be just voice! Like now, I am talking,
so this can be a source of information.

Because my voice actually contains some amount of information. It might be just the video you
are watching, it contains information. It might be some pictures, it might be some text! Whatever it
is, it is some source of information. It might be from any particular kind of source, but first we
have to take the source of information. This source of information might have a originator, or it
might be originated from some particular place. In communication what we mean, that somebody

1
else, we need to transmit or we need to actually transfer this information as fast as possible. So
that is actually communication.

So basically there should be a recipient, or I should call a receiver. It might be far away from the
source itself and source should be associated with a particular transmitter which will take those
information, convert it into a particular form of signal. We will discuss about that later on. And
then transmit it through a media, which is very important in communication, it is called channel.
And through the channel, at a distant place to the receiver, it will be transmitted. So basically if we
just define that is where the message signal comes into picture.

So form the source of information we get the message signal. It might be... we will discuss about
what kind of signal can be generated. We had just briefly told that it might be voice, or picture and
video. It might be just computer generated data. Whatever it is, so that has to be first transmitted,
and this is, here it is called transmitted signal. We will see how the message signal and
transmitted signal are different. There must be, this transmitter must do some processing on the
signal to make this communication sustainable.

And make the communication successful and then it will be launched into a channel. so channel
is a very important part of communication that is the thing which actually either a physically or
virtually connects the source to the destination or transmitted to the receiver. So it might be a
wireless channel where we just transmit it or radiate it in the wireless. In the forms of antenna or
other means and then it will be just propagated through electromagnetic wave through the
channel, and it will be received at the desired location or receiver. So that is one form of channel,
there might be channels which are not like this air. So it is like wired channel like our twisted pair
earlier being used for telephony or it might be coaxial cable that is being used for television
transmission. Or it might be, in the recent advancements which is coming up, it is called fiber
optic cable. So it might be any kind of channel that effectively transferred this transmitted signal
though it.

Okay so that is the functionality of the channel and once this is been transmitted, so after the
channel. You might be asking should this transmitted signal and receive signal be same?
effectively not! so channel does something to the transmitted signal what are the things that
channel does that is also another very important part of communication so we need to know
exactly what channel can do to my signal and we need to know how to really combat.

Any kind of impairment it puts on the transmitted signal because transmitted signal is my
information or it is carrying information or it is carrying my informations so we I need to exactly
know how it is contaminating my transmitted signal and then in the received signal whatever we
get after this contamination or distortion we get it to the receiver and then receiver does
something to get back to my information so it is that information, so we have to think about as
much faithfully we can actually put from source to receiver better the transmission quality better
the communication in that we are putting for so it is very important that whatever we are whatever
information where transmitting suppose we are sending an E-mail and that is faithfully transmitted
over this entire chain then only we will say that is a valid communication because if I write
something and that writing is being distorted, which some grammatical mistakes spelling
mistakes some words missing that will not be faithful representation of the input source
information.

So that is very important that whatever we do in between in this three blocks of transmission
channel of course we do not have any hand we do not have anything in hand so in the transmitted
and receiver block knowing what the channel is so we need to very clearly know what kind of
channel we have or we need to use or choose channel what kind of channel we should use for a
particular transmission, so knowing all this things how do we really do my transmission and how
do I do or employ something at the receiver side.

To get back my information faithfully so this is what communication actually does. so if you just
we have to say some examples. There are plenty of examples you know are telephoning is one of
the probably oldest example of communication, so where whatever voice signal will generally
generate, so what happens. There are two steps whenever you do communication so the steps
up first.

2
(Refer Slide Time: 07:54)

Step 1 generation of message signal okay so that is probably the first step.that is like for I have
taken that example of voice communication or you can talk about video communication okay so
board casting videos all those things so whatever you do first you generate the message signals
so like I am communicating and talking so that is the message signal contain some amount of
information inside.

So first generating that message single then what we have to do is second step that we need to
describe this message signal in terms of some set of symbols okay so this set of symbols will be
mostly electrical. so it might be like I am creating I means just delivering a speech. now this
speech single there will be transducer most probably which will actually covert it into a equivalent
electrical signal so that we can say that is some symbol set. so it might be just voltage level of
that electrical signal.

So over time whenever I am means varying my speech or my teach and all those things all those
quality of my speech accordingly it will be translated to a equivalent electrical signal which might
in times so if we plot that in time voltage, so it might look like something like this, according to my
variation. so this particular signal so we can call that as set of symbols. so it might be just time
varying voltage level or varying current level whichever way you want to represent it.
So it should be that massage signal through a help of transducer whatever form that message
signal input signal has, it must be converted mostly we are talking about the electrical
communication, so it must be converted to equivalent electrical signal, so that is the first stage.
okay so, or we should say second stage. The third stage of this communication is then we need
to encode these symbols into a suitable form that can be transmitted over this media.

Whatever channel we have. okay so what do you mean by suitable forms so already we have got
this electrical signal is that good enough for transmission, so this is where transmitter we have
talked about those 3 stages so this is where after the transducer when my equivalent signal may
be voice or video whatever it is that is being coveted to electrical signal now it is my task to
actually convert this into a suitable form that will be another signal of course.
But that should be suitable form so that it helps in transmitting okay so this is for the transmission
purpose, we need to convert it into suitable form so there are multiple suitable form will discuss
about them but right now I am just giving one example, so one suitable form is first of all we need
to modulate this single y modulation that probably right now it will not be very clear but we will
come back to that why that signal needs to be modulated.

But right now we just take that it needs to be modulated. what do I mean by modulation so
basically this signal as to be rided by a modulator okay so what is our modulator so modulator
generally is a sinusoidal carrier so it is a high frequency sinusoidal carrier okay so we can call that
as either sin some A sin(ωct) of frequency ωc or 2π fc or it can be even represented as

3
A cos(ωct) any form of sinusoidal which has a frequency which is represented by ωc or 2π fc, fc
is a frequency at sinusoidal okay.
A is the amplitude of that sinusoidal now what we need what we mean by converting into a
suitable form is this at this particular sinusoidal we take that as carrier like whenever we are
sending suppose letter so what do we do we put in some envelope it is almost means this carrier!
particularly is almost serving a purpose of an envelope, so it carries the information and at the
end it delivers that information and after that it is not required.
So basically we put it inside that inside the envelope, like we whenever we mail anything we put
inside a envelope, the envelope is carried and then after that envelop will be opened up, the letter
is delivered, then the envelope has no value. okay the similar thing this is a carrier and on this
carrier we try to put our information, so suppose let us say our information is let us take an
example that our information is something like this.

(Refer Slide Time: 13:38)

This is one typical signal which is like our information let us say it is, we all are familiar with binary
or digital form of data so it is 101 and this is actually time, so this is the signal we want to transmit
so what we do is eventually like this we have the carrier we have already seen that the amplitude
of that carrier should now vary according to amplitude of this particular message single okay.
So this is our message single converted to electrical and now we have a carrier which is like this a
sinusoidal carrier and the amplitude of this carrier which is this path must be varying over time
which almost mimics this message signal. so how this will look like if I modulate it should be
looking like this so basically the amplitude should be having this envelope inside the carrier
should reside, something like this.

So what is happening as now you can see the carrier almost remains the same, it is frequency it is
oscillation period and everything that remains in what is only happening is the amplitude of the
carrier is instantaneously varying as the signal, message signal is varying. So it means that this
particular carrier is carrying that message single on top of it like an envelope and it is being
carried over the amplitude okay.

So there can be another way of doing it any sinusoidal is characterized by two or three things.
right now we will take two examples, one is it is amplitude so whenever we write sinusoidal it is
A cos(ωct) right so the amplitude is 1 part of it which characterizes a sinusoidal or it is a
parameter of the sinusoidal another part is this frequency.
So now I have the option of varying any of them okay, so either I can vary this amplitude which I
have done over here or what I can do I can keep the amplitude fixed and I can start varying this
frequency, okay. So if I wish to vary the frequency how it will look like so the same thing if I
modulate with this particular signal it will look like this.

So suppose I have this particular portion 1 this particular portion 0 then 1 then 0 followed by so in
1 I will be giving a frequency which is higher so this will have higher frequency and then while 0

4
this frequency will be slightly lower that means the oscillation period will be it means the time
period will be higher and then again 1 so it will high frequency and so on. So what will be
happening if anybody sees this it is actually there is a center carrier and as we are transmitting
different kind of symbol let us say for this case 1 and 0, 2 voltage level actually.

So basically the carrier frequency is varying it is becoming suppose I have carrier frequency ωc so
it is becoming ωc + Δω at this moment and it is becoming ωc − Δω at this moment, okay so that
is another way of varying things so basically what will happen now you can see if I transmit either
this signal or that signal at the receiver end I will be able to actually decode my symbol because I
know either if I know this is means modulated at the amplitude then I know I have to just track the
amplitude of the signal how it varies with time forget about the carrier this internal variation I can
take out I can just track the amplitude and I can roughly get the signal back.

Or otherwise I can just try to see what is the frequency and how it is varying so if I just means
track the frequency I can immediately see this is the portion where high amplitude is being
transmitted because high frequency is being transmitted this is the portion where lower frequency
is being transmit so therefore automatically the message symbol should be having low amplitude,
so basically either this or this will have the same representation at a receiver, okay. So this is
where the third step we talking about this third step.

(Refer Slide Time: 18:43)

Encode this symbol suitable into a suitable form of signals which is good for transmission so this
is what we are trying to do over here. Now you might be asking why we are actually we could
have directly transmitted this signal why we are doing these things, so that is the first question
probably we will come to any communication engineers mind why we are doing these things,
there are few reasons of course the reason will not be directly clear right now but probably little bit
a head in the course it will be much more clearer.

But right now we can just say some reason, one of the reason is see look at the carrier
deliberately I have chosen very high frequency. The reason is generally whenever you are
transmitting suppose you are transmitting through an antenna okay, and you are receiving through
an antenna let us say it is wireless transmission so we are actually putting an antenna. Now the
antenna designing guideline is if you have read already antenna designing guidelines and little bit
of electromagnetic so you know that if whatever the electromagnetic wave we are putting so
whatever that is sinusoidal carrier will be putting that will be converted to the equivalent
electromagnetic wave of the same frequency.

Now whatever wavelength it has the antenna dimension should be equivalent to that wavelength
or okay, so now if I just transmit it, at a very low frequency like this one this might be having lower
frequency it is varying slowly like the wire signal it varies very slowly it has the highest frequency

5
component which is just 4kHz okay, so whatever voice tonal quality we have it has the highest
frequency which is 4kHz, okay.

So and many of them, many of the components wires components are even lower frequency
starting from 300Hz to almost like, it does not go up to 4 maybe 3.3kHz or something like that,
okay. So if you take those frequency component it is very low frequency correspondingly the
wavelength will be very high and what will happen the antenna size because that has to be
equivalent to the wavelength size that will be very big, so make a transmission I have to build a
huge antenna very big antenna maybe size of this building and all those things.

So that is not really desirable that I want to make a point to point communication between one
person to another person suppose he is carrying a mobile phone through he wish to
communicate and his antenna is very big that is not feasible or practical so that is why we need to
whenever we transmitting things the frequency component of that has to be really high, because
we are radiating through antenna and antenna size if you wish to faithfully transmit and receive
that signal antenna size has to be comparable, okay.

So that is one thing which is required so immediately it comes to our mind that maybe the
frequency has to be translated or the frequency has to be very high so how I have a voice which
is having highest frequency component probably 3.3kHz can I take this to somewhere higher
frequency so one way of doing it later on you will see mathematically right now we are not saying
we are just stating that it is some way of doing is this modulation or putting it inside the carrier it
will actually take it translate it into a very high frequency and effectively what happens their voice
modulated with that carrier will have a very high frequency component and correspondingly
antenna size requirement will be very low, very small size tiny size antenna will survive.
So that is one very practical reason why we need to do modulation, the second reason is which
probably will not be very clear right now but I am just telling it. See on the air, if you wish to
communicate that is always true that if I wish to communicate what I need this particular air
media is a shared media okay, if I wish to communicate so everybody else needs to communicate
and they want to communicate whenever they wish if two of us are trying to or multiple of us are
trying to communicate simultaneously to multiple other fellows.

And we are using the same media what will happen they will actually come into the same media
and they will collide whatever information we have they will get mixed and then it will be a very
difficult at the receiver end to actually segregate them or separate them. Now what I am saying
without any prove later on we will prove that this modulation basically segregate these things
before hand.

So whenever we are transmitting what we do I am just telling you the technique later on we will
prove that the technique helps so we actually choose different, different carrier frequency so for
my transmitting suppose station 1 is trying to transit to station 2.

(Refer Slide Time: 23:46)

6
Okay, while is the an station 3 simultaneously wish to transmit to station 4 in the same media now
what he can do, he can take a carrier frequency of f1 and he can take a carrier frequency of f2
where f1 is not equal to f2. two different frequencies and there are some criteria but it should be
the separation between them and all those things but right now you are saying but if take you
separate frequencies that there is a possibility that both the signals can coexist in the same
media.

But they will not superimpose. force basically they will have different frequencies location in the
frequency domain and you can later on you are all probably aware of it flitering. At the frequency
domain you can always put bandpass filtering to choose your own carrier and that we can
modulate it or you can get your signal back.

So this is one way of multiplexing multiple simultaneous transmissions into the same media
without disturbing each other oaky or without interfering each other so that another reason why
we should actually put a signal into a carrier okay so these things will be forward clear when we
defined them more clearly and mathematically okay.

So now going back to our steps so we have discussed about third step right so what should be
the fourth step, fourth step is so we have already encoded into a suitable form so whatever might
be the requirement we have now discussed about that and we will be actually doing that at the
transmitted side.
So this is being done at the transmitter so and this part is being done, this is the source where
message is generated. this is what the transducer is converting into electrical signal and then
from that electrical signal we actually generate suitable form which can be transmitted so that is
the duty of the transmitter and then after the transmitter we actually transmit it over the channel.
So that might be done by the help of media or any other form that is required the transmit okay so
that is has to be by an antenna, probably I am talking about. so it can put antenna, suppose it is
wireless media through attained convert it into electrical, the electro-magnetic wave and radiate it
over the channel okay.

So this is the transmission part after the transmission part, in the receiver part, we need some
more things. So if I just go to the receiver part, there, whatever encoding we have done. It is like
an envelope we are putting our letter inside the envelope at the other side if I wish to get the
message we have to open the envelope we have to tear up the envelope and then read the letter
we get the message.

(Refer Slide Time: 27:01)

7
So we have to do the same thing the reverse process so here we are putting inside the envelope
that means we are modulating so at the other side we have to accordingly demodulate the signal
so that is the fifth step, fifth step is decoding or demodulating so we are just writing it decoding,
we will see why that is written as decoding. There is multiple other steps also other then
demodulation.

The demodulation is one of the steps so decoding is the fifth step and sixth step is from after
decoding its again it goes to the transducer. because finally I want to hear it so it should be in my
speaker okay so that reverse form of transducer, or another type of transducer which converts
electrical or equivalent electrical signal into voice or speech signal. so it is recreating the original
message signal. so that is the four steps which are involved over here.

And in this four step as we have discussed our focus will be transmitter and the receiver what
transmitter does what receiver does and we have also choose the media accordingly or the
channel accordingly and we need to be aware of what actually channel can do so we will next see
what are the different characteristics of channels of course we won't go into the details right now I
have just touching up on all the topics.

Which needs to be really discussed to actually deal with communication systems so what we are
so far understood that for a communication system there must be a transmitter okay before that
there must be a source, transducer, it should n't be a part of the transmitter and it should be
anyway there, because otherwise you cannot generate the signal once we get the signal after that
we need a transmitter which encode the signal. We have seen one form of encoding, which is
called modulation. we will see other form of encoding also, later on. Then it goes into the channel,
the channel we need to choose, accordingly should we choose wireless channel or fiber optics or
coaxial cable some transmission line, twisted pair whatever that media is we have to choose and
we have to also know the characteristics of that channel so that is very important that we need to
exactly the understand the characteristics so that we know how the transmitted signal will be
actually interpreted at the other end.

So received signal will not be transmitted signal will see that in the next particular section will be
discussing about that how channel actually degrade the performance and how the channel
actually change the signal quality and then from the received signal we need to know again
decode signal to get the message back then again a set of transducer followed by means so who
ever is actually getting the information.

So it can be done okay so this is the chain of the communication now you might be asking what
are the different forms of communication that exist we have already discussed about point to
point communication so which is like from one particular point to another particular point one
particular person or one particular source to a particular destination. that is one form of
communication we all aware of starting from telephony to computer communication and all those
things.

8
But there is also another form of communication which is called broadcasting so like radio or
television where the source is unique, and he transmits that transmitted signal must go to multiple
receiver simultaneously so basically the information contain remains the same whatever in
broadcasting suppose radio whatever is being transmitted it is not unique for a particular receiver
it is same for all the receiver and it has to be simultaneously transmitted to all the receiver and
decoded simultaneously by all the receiver.

So that is one form of communication that we already are aware of broad casting so the other
form is point to point which is we call it uni-casting one particular point to another particular point
there is in between some other things which is called selective broad casting or which is also
termed as multicasting.

So there if I have multiple receiver not all of them I intent to transmit I might choose some of them
but I will still be transmitting same content to all of them simultaneously okay whomever I choose
so these are the different forms of communication that we are aware of we have seen that there
are examples of those communication, so what will do in the next section we will start talking
about our channel what do we mean by channel and how it actually effects our transmission and
why it is very important to know the channel very well before we actually device a communication
mechanism.

9
 NPTEL

NPTEL ONLINE CERTIFICTAION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 02: Fourier Series (Contd.)

Okay, So let us come back to the concept of Channel. Okay, so what do you mean by Channel? let
something we want to discuss we have already discussed about different forms of channels like, let us
say one is wireless channel which is their medium that we know and other is twisted pair okay which is
being used for our telecommunica?on transmission.

(Refer Slide Time: 00:24)

So third is Co-axial Cable this is being mostly used for the broadcas?ng television signal. this also
something we have seen already, and fourth part may be a Op?cal Fiber this is the latest one which
supports huge amounts of bandwidth, you are aware, so the fiLh might be satellite channel so on and

10
so forth so there are mul?ple channels that have communica?on channels especially that are available
through all of them actually the mode of transmission is electromagne?c wave okay.

So for wireless it is not guided so it just radiated electromagne?c wave. where as for all other cases it
actually guided electromagne?c wave. op?cal fiber it s?ll electromagne?c wave but at visible op?cal
region okay let us say that frequency such that it is called op?cal electromagne?c wave okay, and so all
of them are mostly electromagne?c wave, through that only signal propagates through a channel, which
our channel we pick.

(Refer Slide Time: 02:41)

Now what Channel can do? Let us start discussing about these things. the first thing which channel can
do. let me list out first what are the things that channel does, and then we will discuss about them. so
the first thing is, channel can introduce noise, this is probably the most important part of
communica?on. What is noise? And how noise affects? so a communica?on engineer must know very
well about noise. and that is probably the biggest enemy of our, means.. channel okay.

So second part is Distor?on, and the third part is Interference. let us talk about them one by one. So
what is noise? Noise is a again what we are doing? we are trying to transmiVed electrical signal which is
being converted into electromagne?c wave and in the channel what might happen, due to different
reason okay, so there are might be this random electromagne?c waves which can be added to this
par?cular transmiVed signal.

So what is generally say, channel is if you say. If we characterize the channel as a addi?ve channels okay,
what do you mean by addi?ve channel? it’s like a simple adder of signal. whichever signal gets means
we put as input to this channel, if there are mul?ple such signals it will just add up those signals.

So suppose I have a signal if one f1(t) okay so I am just wri?ng down a voltage of that signal and this is
the varia?on with respec?ve ?me that, this is called the func?on t for suppose this is my signal that has

11
been put into the channel and then if there is another signal that is also being kept or put inside the
channel let us say f2(t) okay so what the channel will do.

The channel will simply add these things. so at every ?me instance it will take this voltage and this
voltage it will add and put this, okay. so every ?me instance if you point wise add these two signals you
will get a composite signals which is the typical criteria of addi?ve channel. if the channel is addi?ve we
generally say that channel is linear. okay but there are non linear channel also were simple addi?on by it
does not work.

It might be may be ( f1 + f2 )2. so it might be quite start crea?ng all those square cube terms so those are
non linear channels. but here we are just saying our channel is let say that one characteris?cs of the
channel that it is a simple addi?ve channel. but it means number of signals present in the channel they
will be just means amplitude added or in terms of electromagne?c wave.

We will see that similar effect so that whenever you convert that electromagne?c wave into composite
electromagne?c waves into our signal it is just be addi?on of those to signals. Now what is Noise? Noise
is inside the channel there are electromagne?c waves which are being generated it might be due to
lightening it might be due to curious radia?on coming from outer space.

It might be the radia?on other form of radia?on that has generated on earth and being reflected back
from the ionosphere whatever it is. there might be mul?ple random signals which we have no idea
because it is generated from elsewhere we have no idea and this might be generated without our control
and this might means be an input in the channel.

So we cannot help it because we are using this channel so there will be some curious signals always
being generated or created inside that par?cular channel if it is wireless I have talked about all those
examples like electromagne?c radia?on for outside and all those things even for wired also it cannot be
fully shielded so all those from outside some electromagne?c radia?on can leak in into the channel.

And it can co-propagate along with my signal and then definitely the channel is addi?ve or linear we
should say then that signal will be added to my signal, this random signal which I have no hand in
genera?ng I have no control in it, it will be any way generated and added to my signal that is called noise.
So, if you see therefore the noise is just another random signal and calling it random because I have no
control on it.

It can be generated without my knowledge and without my influence so that is why it is actually random
source of informa?on for me so this par?cular is always inevitable, it will be present in the channel and
it will be added. So, when we say noise you might be saying okay, is noise just being added at the
channel precisely not, noise also can be added at the transmiVer.

Because transmiVer also would be having some hardware, in the hardware you will have let say all
components transistor some resistor and all of the components juts take a simplest for component let
say resistor whenever my signal it has to be transmiVed so it has be processed through this hardware
whenever hardware so it must be passing through a par?cular resistor and inside the resistor what will
happen.

12
Even if I suppose I transmit this signal or I pass this signal through the resistor it will not be completely
suppose I have a resistor through which I am passing this signal this f1(t), whatever it is okay the output
what will happen see through the resistor I will be actually crea?ng this voltage ?me varying voltage this
f1t and according to current will be flowing now what we happen inside the resistor there are mul?ple
carrier electrons there also doing brownian mo?on due to the brownian mo?on they will random
movement.

So what will happen on top of this there will be some random signal which will be added and my output
signal will look like this so it is almost followed the paVern of my f1(t) original f1(t) but it will also have
this small things added which is just being generated due to brownian mo?on of the electrons okay.

So this will be true for any hardware any electronic device we put so in transmission or transmiVed they
will be electronics which will generate this noise in receiver also there will be hardware which will also
generate this noise. so noise is not just, it is one of the means my channel is one of the source as we are
discuss but noise will also be generated at the receiver noise can also be generated and the source
where from I am actually star?ng my transmission.

So everywhere there can be noise but channel is one of the source of noise that is why whenever we are
characterizing channel we need to understand also, what should be the noise, or what should be how
much noise will be added what is the characteris?cs of that noise and all those things so, we have talked
about it is a random signal I have no knowledge about it how it will vary with ?me I have no knowledge
about it.

It is not my generated message so I do not know exactly how it will be generated okay so the paVern at
which it will be varying that will be completely random, it is true. Okay, but I have to s?ll get some
characteris?cs and that will be our major focus of this course how do you characterize noise? and in
presence of noise how do you combat noise, in the transmiVed side as well as the receiver side.

So that noise does not actually change my voice signal so you have already seen suppose my voice signal
looks like this f1(t) and on that f1(t) start pu\ng noise so what will happen, the voice signal will be liVle
bit varied and how much it will be varied how much I can tolerate all those things has to be in
considera?on wherever I am transmi\ng.

And possibly as much I can get rid of that and as much pure f1(t) can generated at the receiver that
much faithful the communica?on will be or the communica?on will have that much clarity. okay, next is
distor?on what do you mean by distor?on let us try to think about that, whenever we talked about the
channel generally the channel any form of channel you see the channel will have some low pass effect.

(Refer Slide Time: 11:58)

13
That means it can carry frequency up to some value beyond that it will not be able to carry those
frequency, those frequency will be suppressed, so that is the low pass characteris?cs though channel
transfer func?on if you see, it is almost look like this so up to some cut of frequency probably it will be
almost carrying means that signal, but aLer that it will be hugely aVenuated inverted okay so this
happens.

Suppose our signal has mul?ple components different frequency component then what will happen,
some of the frequency component will be actually equivalent carried but some of the frequency
component might get, it will suppressed. that is the effect term that we all know about low pass filtering,
who ever has read liVle bit on signals and systems, so let say I am transmi\ng this kind of signal okay.

So this is voltage level let say 0 which is voltage level let say high volt okay, so I am trying to transmit this
kind of signal now, this is where, will be later on seeing also but this where at this edge or this edge or
this edge all those high frequency components are there because at the negligible ?me the voltage level
is varying hugely.

So that actually calls for very high higher frequency very small amount of ?me voltage change or
happening, so high voltage high frequency components are involved in that part, is the part where signal
range almost constant over ?mes required not much of high frequency components again, so what
happens in a low pass filter high frequency component will be suppressed so basically has an effect this
par?cular part will be smoothen it cannot really sharply jumped.

Because the high frequency component will be suppressed so the output of a channel, if we take the
channel as low pass filter, now what will happen it will look like this almost like charging discharging of
the capacitor okay because it works as a low pass filter, so this is what will be happening. Basically you
can see I am trying to transmit something because of the low pass nature of the channel.

The signal is ge\ng liVle bit distorted okay, not only that as you go along a huge distance there will be a
high loss associated with the channel because channel might have absorp?on loss, so it liVle take away

14
of the some of the energy of the electromagne?c wave, it might be due to radia?on in the air it might
be going different direc?on so the antenna point only a por?on of that radiated energy will be linked.

So what will happen? If I transmit some amount of the energy. at the end this will really become very
small, something like this so my one now will come close to 0 because the energy level is ge\ng
depleted and also the signal form is ge\ng liVle bit distorted, so this is actually the part of the distor?on
that we generally see in the channel. It is low pass effect due to the absorp?on and aVenua?on of the
channel okay.

There are other distor?on we are talking about but we are just assuming we have been assuming that,
the probably channel is linear, okay in a linear channel, forget about the noise if no other signals are
present my signal only is being carried out, if the noise is sufficiently small then noise signal will be
par?cularly only that signal will be carried out through the channel. But what might happen if the
channel is no longer linear, then suppose that channel output if we say o(t) suppose I give input as f1(t)
and the channel is not a linear one.

So if it was linear may be there is constant term C into f1t let C is that aVenua?on, so that ?ll aVenuate
from here to here the signal level okay so let say C is .1 or .01, depending on how far your receiver he is
what kind of absorp?on or loss or aVenua?on of the par?cular channel has, so depending on that those
signal strength will be liVle lesser.

But single quality will remain the same but instead of that if the channel has not linearity, then what will
happen o(t) will be some C1 let say another constant into f1(t)2 okay if might be it is juts your quadra?c
or it might be in quadra?c with all other terms let say C1 f1(t)2+C2 f1(t) +C3 so that the channel func?on
let say and C3.

If it is like this then what is happening the o(t) is not propor?on to the ideal, when it was linear it was
propor?on all to f1(t) no longer it is propor?onal to f1(t), so there will be added distor?on okay do to
this squaring or if you have even higher order non linearity in the channels let say cube or three, power
four all those things.

Then there will be added distor?on so the signal value is like different because there will be square term
added by some other linear term and some constant terms so there will be some distor?on due to that
non linearity of the channel. so there are mul?ple ways that channel can be this distor?on one is low
pass filtering effect the second is the distor?on itself which can also keep the channel non linear okay.

Due to that linear distor?on is coming up and then you have some aVenua?on which significant reduces
the energy level of the signal so that is the second part of our discussion what channel can do the third
part is which is also inevitable which is called interference. see this interference comes when mul?ple
signals are being transmiVed simultaneously over the same channel even if you take care, very nice care
that we have discussed about modula?on there will be s?ll some spurious por?on of that other signal
which comes inside the band of our desired signal.

And that will create again some amount of means impurity into your signal, because that is the different
signal which is now ge\ng super imposed with your signal that will contaminated the signal so paVern
of your signal which carries the informa?on actually will not be sustained if that happens so these are

15
the three source that can contaminate within your channel so you have to be very careful about noise
of course you have to very careful about this distor?on.

And that is why we have discussed about the channel so you need to characterize the channel you need
to know exactly what kind of nose it gives what are the, if it is random what are the stochas?c nature of
those noise then you need to all know what kind of distor?on that channel gives if it is non linearity, how
what kind of non linearity what are the co-efficient’s of those non linearity, so all those things has to be
known.

If it has a low pass filtering effect what is the characteris?cs of that low pass filter due to that what kind
of distor?on that signal will get, and also if you as a aVenua?on how much aVenua?on it gives so all
those things has to be known and then you have to also know if there are other signal which are present
what the effects of the interference these three things should be very clear whenever you are
transmi\ng because eventually what will happen as we have discussed that many signal you transmit.

(Refer Slide Time: 19:57)

So we had this transmiVer they have this channel and we had this receiver so this is actually the
transmiVed signal so transmiVed signal may be very nice but what might happen due to the channel. I
am just taking out the interference and as well as non linearity of the channel so even if you take out
that what will happen aLer passing it through a channel this par?cular signal will become like this.

This would something like this and on top of that there will be noise added at randomly, so this is
actually going to be your received signal and that is why we said transmiVed signal and received signal
might not look alike, it might be completely different so this is one fact that we have to deal with and
this is fact that we need to build with mostly in communica?on that would probably channel will give
some distor?on some amount of the noise interference to the signal.

And my received signal will not be completely faithful representa?on of the transmiVed signal. so
therefore what I need to do. first of all we need to characterize this how much distor?on how much

16
noise will be added and we need to know how to combat this, and to get back my original signal because
that is the whole purpose of communica?on at the receiver.

So this is something will be trying to do in our communica?on okay. so next what will start means will
probably that next sec?on of our next part of next half of our course will talk about modula?on liVle bit.
so I am just here it is going through all the models of the communica?on that are required, why they are
required and what kind of things very briefly without giving any details of it very briefly making it
familiarize with these concepts.

Once we get some hold of this and we know actually what are the things we need to concentrate, then
will actually go into the details of each other. so our next target is that modula?on we have already
discussed very lightly what modula?on is, now will try to characterize those modula?on techniques what
are the different modula?on techniques that we have, how do we actually characterize them, those
things we will discuss okay.

(Refer Slide Time: 22:39)

So in modula?on especially in analog communica?on we will be talking about two forms of modula?on,
one is called con?nuous wave or CW modula?on and other one is pulse modula?on. okay these are the
two most common version of modula?on what is called con?nuous wave modula?on the other one is
pulse modula?on.

So basically whatever example we have given that modula?ng the amplitude of a carrier or modula?ng
the frequency of the carrier, that’s actually called the con?nuous wave modulate modula?on. that
means the carrier wave is con?nuous okay, that means con?nuous means it is actually in ?me if you just
go everywhere in ?me any ?me instance you defines there is some it has a con?nuity over the ?me.

And there is some amount of amplitude you will be ge\ng always. okay so that is actually con?nuous
wave modula?on in pulse modula?on what we do we actually take a signal and the signal looks like this

17
it periodic pulse so it is no longer con?nuous kind of signal. it is just defined what this small dura?on the
rest of dura?on it has nothing again for a small dura?on it is defined.

And that is called as pulse and this is actually a pulse train, a periodic pulse that goes around. so if
somehow I can put by informa?on inside this pulse okay now what are the criteria that has each of these
pulses have few things one is the width of the pulse okay like for the sinusoidal we are trying to see what
are the characteris?cs or what are the basic features that sinusoidal have, what are the main things that
sinusoidal have which we can modulate.

Where we can introduce our signal okay here are also we are almost looking like that, so we are trying to
see what the characteris?cs are basic features of this pulse. so one is this pulse width another one is this
pulse amplitude okay, like the sinusoidal there was a amplitude the pulse also has some amplitude pulse
have some width okay.

So this will be in ?me how much milli second or micro second is pulse width that is another thing. and
then if the pulse is periodic I can also vary the posi?on or loca?on of the pulse so I can slightly deviate
the loca?on of the pulse so that is called the posi?on of the pulse. That is also another variability that I
can get., so posi?on, width and amplitude. and for modula?on I can use all three of them.

(Refer Slide Time: 25:52)

Suppose I need to transmit this signal let say this signal okay and I have pulse train which are coming like
this, so in pulse amplitude modula?on what will do will actually borrow the amplitude and put it and
every ?me instance to this, so basically the amplitude of this pulse will be modulated or will be
mimicking the shape of this signal so here the pulse amplitude is with this here it will be liVle bit higher
here it will be something like this.

So that if you just connect the ?p of those pulses will see this signal being formed okay, so that is actually
called the pulse amplitude modula?on so the pulse amplitude is now being modulated according to your

18
message signal which is let say f1t okay the second part is that I have again this signal right I have this
pulse what I do is I actually modulate the width according.

Almost like frequency, so as the. my mechanism of modula?ng this width is, as I increase the amplitude
or as I see the increase in the signal amplitude I actually either increase or decrease the width of this
course okay so basically here they will be so there should be a minimal width which refers to the
minimal part of the signal okay.

And accordingly I will be varying the width from here to here there will be some scaling or varying the
width so let say this is the minimal part so here the pulse should be of minimal width and then whenever
it increases here it increases slightly so the pulse which should be the more, here it is increased slightly
more.

So here the pulse will be more the period or the center of this pulse should have the same frequency has
this, whatever ?me it has is ?me the center of this pulse will have the same ?me but width of the pulse
is now ge\ng modulated according to the message signal so basically what is happening the pulse is
being carried now and there is a characteris?cs of the pulse which gives us indica?on about the signal.

(Refer Slide Time: 29:00)

If you just look to the width of the pulse train you will get review of the signal how it should vary over
?me, same thing can be done over the pulse posi?on. so for pulse posi?on what we do we have a again
a signal and let say these are the pulses, periodic pulse width same repe??on intervals and then when
we modulate the posi?on now again what we do we actually vary the posi?on of the pulse so from here
center I can actually take the pulse here and there and how much I push the pulse the right or leL side
that depends on what is the signal strength.

So let say I always if the signal is posi?ve I go into this direc?on and let say my rule is if the signal is any
?me ge\ng nega?ve and it goes this direc?on and we say that maximum signal level let say this and
suppose that signal 0 is over here okay, and the minimum signal level which is this that is the maximum

19
right shiL or leL shiL of the pulse so let say that we have decided up to this, so here the pulse will be
maximum let say nega?ve it will be leL shiLed so it should go maximum leL shiLed.

So that is the center of pulse that should go. this pulse because it is liVle bit less nega?ve so it should be
less shiLed so this should be more closer to the center of the pulse, and so on which one at this 0 level
there is nothing this is again say this part is posi?ve, so it should be posi?ve shiLed so this should be on
this side and this is more posi?ve and let say this one again should be posi?ve shiLed on this side so if
you now see the pulses they are now distorted posi?on so we remain the same.

Amplitude remain the same just the posi?on of the pulse is ge\ng driLed, and the posi?on and loca?on
of the pulse now gets a informa?on about what the corresponding signals strength, so this is another
form of modula?on so we are now dealt with a con?nuous wave modula?on in the con?nuous wave we
have seen that there are two parameters that can be modulated.

We have the carrier we can modulate the amplitude we have already demonstrate the how the
amplitude can be modulated we can modulate the frequency that also we have demonstrated that how
the frequency can be modulated we have taken their digital or binary signal for that you can as well take
this kind up signal again but will see the same result only their there was a abrupt varia?on of amplitude.

Here probably gradual varia?on of amplitude will be happening or gradual varia?on of the frequency will
be happening. so we have amplitude modula?on and we have then frequency modula?on but let say
whenever we talked about frequency there is also another term called phase so generally any sinusoidal
may be represented as A cos(ωc t + θ ) okay.

Basically we had earlier said just ωc t but there should be some ini?al phase also so which is θ so we
actually have three parameters and accordingly we must have three modula?ons, one is called
amplitude modula?on, one is called frequency modula?on and Fm which you are familiar with, an
another one is called phase modula?on PM.

And similarly for a pulse we have three informa?on carrying things, which are pulse amplitude
modula?on that is called PAM then pulse width modula?on PWM and then the last one is pulse posi?on
modula?on is called PPM so there are three pulse modula?on there are three con?nuous wave
modula?on which will be dealing with later on right now I have just given very brief overview of this
modula?on, will try see that rela?ve merit of them how do you mathema?cally realize them.

How do you generate them? how do you demodulate them? so all those things should be part of this
course say this is just introduc?on to this part and will later on see how this things can be covered later.

20
 NPTEL ONLINE CERTIFICATION COURSE

Course

On

Analog Communication

By

Prof. Goutam Das

G S Sanyal of Telecommunications

Indian Institute of Technology Kharagpur

Lecture 03: Fourier series (Contd.)

So we have so far discussed about basic communication okay, and the course name is analog communi-
cation of course will be slowly going into analog communication but we have so far discussed
the modules of communication that the transmitter side receiver side and how a transmitter side
receiver side communicate to each other. so today what we will see that what a particular
communication system requires. So it generally requires two things or if we try to analyze
it we really require two things one is called signal.
(Refer Slide Time: 00:53)

So a little bit of understanding of signal and the other part is system so basically
we were talking about a transmitter followed by a channel and then a receiver right, so
this transmitter and receiver that is actually part of system and whatever we put as input

21
to the transmitter side that is actually signal and what that transmitter does that system
actually operates on signal and produce another signal this is how it and then the signal
propagates through the channel goes to the receiver again receiver is a system it operates
on. That signal and generates something which
is desirable okay, so this is how it the communication goes through but let us try to understand
system probably we will discuss later on it is probably the hardware that has to be designed
which will do some desired operation but initially we will be more concerned about signal so
what is signal is something where a particular parameter let us say it might be if it is
electrical signal probably the voltage or current okay which varies with time and
this is the stress with respect to time so how it varies with respect to time so let
us say if I just put it as a function of time so let us say g(t) and it varies with respect
to time the Signal is actually the description entire description over the time of our concern if
the time is truncated let us say starts from t1 and ends at t2.
So it is within the time limit you want to t2 whatever the signal at rest or let say
we are talking about voltage or we are talking about current so it is the amplitude of that
thing that is actually a signal description now signal can be of multiple pattern okay
and it can be means we are here probably in communication system or in communication analysis
we are talking about time varying or the independent variable is time it is not it is not necessarily
a time always it might be a special variable. So even a signal might be just like let us
say in television the pixel values okay, it has a special variation so in a two dimensional
space like in a television screen at every location what is the signal strength so there
the signal is actually a means independent variable is space whatever it is to coordinate
if the space is taken as 3 coordinate it is a volume then it should be 3 coordinate so
with that space how the whatever parameter that we are discussing about how that varies
okay. So that is the signal can be off means it
can vary with respect to anything but for our case or for our purpose and communication
will be mostly discussing about signal varying with time okay.
So now let us first try to see how do we means anything we define suppose signal we have
defined the way we have defined right now so that must be parameterized that means there
should be some measurable quantity in that signal so how do I suppose.

(Refer Slide Time: 04:13)

I give just a signal let us say this is g(t) okay, now if I just say that let us say it
is defined from t1 to t2 rest of the part we don't define it we can take it as 0(Zero) okay
let us say this is the signal definition now how do we actually measure some parameter
of this signal or how do I characterize this signal one way is this graphical or pictorial
depiction that is alright can we give one parameter to measure some property of this
signal. So that one important property is probably the energy of the signal we will
see that has a big consequence later on. We will see that okay, so how do we define

22
energy this is something very easy if you take g(t) as voltage and if you pass it through
a unit resistance so we will be able to calculate that it should be just g 2 (t) and you
t2

∫t
integrate it right from t1 to t2 okay. g 2(t) dt. So that is actually Eg the energy of this signal g, okay.
1
so as long as it is time bounded probably we can we can integrate it from t1 to t2 that
is fine we will get some value and that is actually characterizing that is a specific
property of a signal, different signal will have different things does not
mean that every signal will have a unique energy multiple signal might have same energy
it's just that integration value has to be same for different different signals okay, if it is
not time bounded then the integration goes from minus(-) infinity to plus(+) infinity it might
happen that this might go to infinity so signal might also have in finite energy that is possible
okay so this is one way of characterizing a signal ,the other way of characterizing a
signal is like this suppose the signal. We are talking about it is a sinusoidal okay,
so it goes like this something like this it stretches from minus infinity to plus infinity
okay so of course if we try to measure that energy of the signal you can immediately see
that we will go to infinity okay so if I just put this sine square (Sin2) suppose the frequency
is ω t dt and suppose it has a amplitude of A it will be A 2 Sin2 and then
if you just integrate it from minus infinity to plus infinity this is going to be infinity.
∞

∫−∞
A 2 sin 2(ωt) dt = ∞
so this is one example of infinite energy okay, then actually I have failed to characterize
this because it is going to infinite energy any sinusoidal I will be drawing suppose I
draw it with some other omega so it will have different frequency and again.
I try to characterize it will just have be infinite energy again so basically what has
happened I have failed to characterize the signal with respect to that particular parameter
that I have defined that is energy can we do something else can we put another measurement
parameters which will characterize the signal so that is called power so whenever we are
talking about power what we generally say that will take the overall signal energy and
will also measure the time and will divide it by time okay, so it is like this suppose
I have signal g so g 2 ( t) I integrate I integrate it from- (t) +(t) okay.
And then I divide by 1 by 2 T and then what I do I put a limit T tends to infinity so
that means by this from - (t) +(t) I calculate the energy of the signal I divided by the
time period which is 2T so I get the power of this signal and then I stretch time to
infinity that means I capture the whole signal right so this is the power of the signal g.
T

t−>∞ 2T [ ∫−T ]
1
Pg = lim g 2(t) dt

so this is another way of characterizing a signal wherever you can see that some of the
signal which has infinite energy that might have finite power like this one sinusoidal
if you just try to do it will be just means whether you do it for the entire t or you
do it for one period. And whatever calculation you get it will be
the same because that just keeps on repeating so more number of g 2( t) you get that
many number of t also you will be getting so it will get balanced and you will get similar
power okay, so accordingly we will classify signals later on so that that part will see
but these are just two parameter with which we can actually characterize the signal
this is something I wanted to tell you okay so the next part is how do we classify signal
so the classification okay.

(Refer Slide Time: 09:17)

23
So one way to classify is, c see any signal we were describing we are having a independent
variables which is time okay for our case it might be other things also so for our case
at least it is time so whenever we classify the signal with respect to time we can say
there are two things one is if I take it in discrete time that means the signal is defined
in discrete time and the other one is the signal is defined always, that is continuous
time okay so I can have a continuous time signal versus discrete time signal.
So the typical example is the g(t) we have drawn or even sinusoidal we can put that is a continuous
time signal okay instead of that if I just sample this signal at different location so
let us say it is then this and this in between suppose this is time t1 t2 t3 and so on so
these are this particular signal is only defined at time t1 at time t2 and t3 in between there
is no definition of that signal so this is called discrete time signal okay, will been
countering this particular kind of signal because we will be doing sampling of signal
that you will see later on this is a typical way of representing a signal so that will
be dealing with this later on and the other part is continuous time signal.
Where it signal is defined at every time instants between the limit we are considering ok. so
the second way of classifying a signal now we had this independent variable time we also
have this dependent variable which might be voltage for our case ok so the signal can
be also classified with respect to that the voltage can be discrete level or it can take
any value okay so accordingly we will be talking about analog signal and digital signal so analog
signal that is very important for our analog communication we are mostly will be dealing
with analog signal so do not get confused with respect to analog signal and discrete
time or continuous time. It does not mean that analog signal has to
be always continuous-time or discrete-time it can be anything only thing is that it will
be analog or it will be classified as analog signal if the voltage level or whatever that
dependent variable we are defining with a voltage level current level whatever it is
what whatever we are measuring that can take any value in a particular range okay so if
it can take any value there is no restriction within that or it is not taking some discrete
values then it is called analog signal if that is not the case it only takes few discrete
levels voltage levels then it is a digital signal typical example of digital signal is
binary signal. Where we are probably transmitting 0 volt
and 5 volt okay, so it only has two levels so that is why it is called binary you can
have other ways of representing signal multiple voltage level you can put so accordingly those
are also characterized as digital signal okay, only thing is that the levels number of levels
are more okay, so that is another way of classifying signals and do not get confused with the analog
signal digital signal or continuous-time and discrete-time the other part is which will
encounter quite a lot.

(Refer Slide Time: 13:09)

24
A lot this is periodic signal and the counterpart aperiodic. so what do we mean by periodic
signal, it's like this, suppose we have a signal g(t)which characterizes this g(t) = g(t + T0 ) or which has
this quality that if I translate the signal by a particular value T0 it looks like that
means in time if I shift the signal overall signal if I shift by T0 okay so whichever
direction it might be +T0 it might be -T0 okay so which ever direction I shift the signal
the signal just looks similar. the typical example is sinusoidal if it is defined from
- infinity to infinity. okay, so if you just shift by one period of a sinusoidal it is
just look same so if that this particular criteria is satisfied then we call that particular
signal periodic signal you will see this periodic signal also will play.
an important role in terms of signal characterization will today only we will discuss about Fourier
series you will see periodic signal has a big role to play in Fourier series the way
we represent signals so that is one thing of course if a signal is time bounded it cannot
be periodic because if a signal is time bounded it will definitely start from somewhere after
that suppose it has aperiodic nature but whatever it is if I just translate.
It will either start from somewhere it will end somewhere if I just shift it then at least
either the last or beginning will not match so for a signal to be periodic always the
criteria is signals must be stretched to - infinity and on the other side + infinity this has
to happen now whichever is not fulfilling this criteria that is called aperiodic signal
all other signals are repeated if I just put a pulse like this is nothing else is defined
from- T0 /2 to + T0 /2 it is like this the signal is defined as this so this is a
aperiodic signal because whatever shift you give you will never get back the same signal
That is not possible if the same pulse is repeated after every maybe let us say T0 then probably
if it is repeated and if it is stretched to – infinity + infinity then it is again a
periodic signal that is the difference between periodic signals and aperiodic signal next
I am just giving all the definition you will see that when we will be discussing the properties
of signal these things will come back so next is energy signal and power signal.

(Refer Slide Time: 15:57)

25
Eg < ∞ Pg = 0
Eg = ∞ Pg < ∞
So what we mean by energy signal we have already seen that signals can be characterized by
two parameters we have defined one is energy of the signal and the other one is all of
the signal defined by if signal is g(t) defined by Eg and Pg we have already seen that okay,
we have characterized that. so, if for particular signal energy is finite okay so that means
it is less than infinity and power is 0(zero) if this criteria happens most probably you can
see already the signal must be truncated it should be limited in time it should not stretch
to minus infinity plus infinity, and somehow the energy is also finite if it is truncated probably
we can always calculate the energy if it is not going to infinity okay, in between.
The signal level is not going to infinity if that is not happening all the signal levels
are finite then if time is also truncated then probably we will be able to evaluate
the energy it should have a finite energy and power automatically because time can be
stretched to infinity and if I divide by that time of course power will become 0. So those
signals are called energy signal mostly energy example of energy signals are time truncated
signals so whichever has a finite duration those are generally the energy signal and
the other criteria is the signal level be it voltage or current it should not go to
infinity okay. Then it is energy signal the power signal
is all those periodic signal we are talking about where the energy goes to infinity because
for all those periodic signal if you calculate energy that is always going to infinity but
the power is finite okay so that is actually the example of power signal this is energy
signal will see the difference between these two it is actually if you see almost periodic
a periodicity has something to do with energy signal power signal how do you characterize
them okay, and the fourth part or sorry fifth part.
Which is also our important part that’s where we will be actually our communication will
be mostly concerned with that that is called deterministic signal and random signal so
deterministic signal means that the entire time duration I know exactly how the signal
behaves okay so I know the levels or the level it attains at every instance of time if that
something I do not know whoever is characterizing the signal whoever is analyzing the shape
signal if that person does not know about what will be the value of the signal at a
particular time instants then we call that as random signal or we characterize that as
random signal. So if do not know the value of that signal
then what we are doing here and I will also give you a typical example noise is a typical
example of this random signal so what happens even though we do not know exactly what is
the value of that signal at different time instants but what we know is the statistical
property of that signal so that means we actually know that if the signal level is defined by

26
a random variable what's the PDF of that random very so these are the things.
And some more things of course right now probably we have not discussed about that so when we'll
be dealing with this in detail form we will be knowing that what exactly is required to
define a random signal and deterministic is what we know like sinusoidal or pulse train
or any kind of just single pulse so the other deterministic signals where we know the definition
of the signal for the entire duration okay. So this is typically means what we talk about
signal and a classification of signal different kinds of signals how do we classify them so
these other typical example of that so right now what we will try to do which has been
done many years before by Fourier that can we have a different representation of
signal so we know the time representation of the signal can we now try to characterize
the signal in different way a different representational all together of the signal.
So this is what we will to appreciate whatever Fourier has done what came into Fourier’s
mind and how he means logically develop that methodology we will try to explain that okay
so that that will be our next target so what Fourier have seen at that time signal is
almost similar to vector we all are aware of vectors so what we will try to do first
we will try to characterize how do we define vectors and then from that vector analogy
we will try to build up the theory of signals which Fourier has developed so this is something
we will try to do now so in vector theory what we know we know that suppose we have a vector.

(Refer Slide Time: 21:23)

Let us say this is g okay and then we have another vector that is in another direction
let us say this is x okay so these are two vectors in different direction okay. In vector
few parameters are defined so we just for recapitulation most of you all aware of this
so I will just define so whenever we say a vector space it has a dot product defined
along with that vector space. so that what we mean by dot product the definition of dot
product you already know. So if I say <g,x> this dot product I want to
get so generally it is this ||g|| that's actually the amplitude of g okay so it's actually this
value into amplitude of x into in-between whatever angle they have vector means it has
a direction so they must be having some angle between the Cos of that angle θ so this is typical
concept of vectors right the dot product concept .
⟨g, x⟩ = ∥g∥∥x∥cosθ
⟨x, x⟩ = ∥x∥2 cosθ
Okay if we try to do dot product of the vector along with itself so that must be distance
multiplied by his own distance so distance Square and Cos θ is 0 because it is the same vector
so θ is 0 because Cos 0 is 1. so generally distance square is characterized and distance Square
comes from dot product. so a vector dot product with itself gives the distance square
right so this is something I know already. now what I will try to do now is something

27
new so basically what I am trying to do.
(Refer Slide Time: 23:08)

g ≃ cx
g = cx + e
c∥x∥ = ∥g∥cosθ
c∥x∥2 = ∥g∥∥x∥cosθ
⟨g, x⟩
c=
⟨x, x⟩
Suppose I have a vector x which is a known vector let us say this vector I know already
now I have another vector g I want to represent with respect to vector x it is just a representation
so I have a arbitrary vector I am trying to represent best representation that I can do
I want to represent this vector g with respect to this vector x what will be my representation
so basically if I have to represent with respect to this vector as always in this direction
I cannot go any other direction because I have to represent with that vector only so
I can only put a scalar multiplication to that vector right so let us say that scalar
multiplication is c so I approximate g with c into x that c is a scalar okay.
So now it depends on the value of c what value of c I put so if I put c this much then it
will be let us say c is defined in such a way that c into x is up to this okay or I
can I can also define it in such a way that c is up to this okay so this is c x and here
this upto this is c x okay so I wish to represent g with respect to x so I have come up with
this value c x right so how good my representation was I need to know the error of this representation
so error must be how much I should add with c x to get g that is should be my error.
So if I just put this vector which I call e that must be the error vector so I can immediately
write in vectorial term g = c x +e where e is the error right so if I just put it this
way then my error will be this one if I just put it this way my error will be this one
now it is very trivial to see how do I minimize the error when it will be minimized when the
vector gets its projection on x it will be exactly perpendicular so c is chosen in such
a way that I get the projection of that vector immediately on that point okay ,so that should
be my best choice I cannot do anything better than this right.
So this is what I wish to do and this is how I minimize the error the and I’m just trying my
best to represent that vector given this other vector I am not doing anything else okay but

28
you will see with this a beautiful representation will come and all okay so this is what I can
do and of course this becomes my error so that is the best error I can do now whenever
this error is the best one then what I can say about this is how do I evaluate this particular
c where the error is minimized right. So this is something I wish to now means derive
right so whenever I am trying to derive that what I can say that let us say this c||x|| this
particular value when the thing is minimize the error is minimized what is that that must
be that value of g into cos θ because it is exactly the projection so g into cos θ must
be the c x okay the modulus I am talking about of course now I am not talking about the direction.
Because I have already taken the θ so c x must be g into cos θ so this is true this
should be the case now what I will do both sides I will multiply by the modulus of x
or the length of x so c mod x square that must be mod g or norm g norm x cos θ right now go
back to our earlier definition so we get c this is nothing but dot product of g and x
so that is g and x dot product and this is nothing but dot product of x with itself.
So this is how the optimal c can be represented and when we say a vector g is orthogonal that’s
the definition of orthogonality when we say a vector g is orthogonal to x if g cannot
be represented by x that means the projection will be the value of c will be 0 so I cannot
represent that vector anyway by the other vector if this is happening then only I can
say they are basically orthogonal. So immediately what should happen c should
be 0 so that means the g vector should be directly perpendicular to this x vector and
I will not be able to represent anyway ok I cannot characterize even the error because
I do not have any representation c value is 0 so if c is 0 if c has to be 0 immediately
I can see the condition of orthogonality that means g dot x must be 0 so that is the famous
orthogonality condition in vector. All these things you already know I just have
demonstrated this to mean’s so that you can appreciate that it's nothing but a representation
technique that we are trying to employ where we are trying to all we are trying is trying
to minimize the error of representation nothing else now what we'll do on this paper which
is a two dimensional thing I will just take two vector and you will see very nicely the
entire representation will be done this is how the coordinate system has been designed
so in the next section of this class we will try to do that and then we'll go to signal
with this analogy.

29
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das

GS Sanyal School of Telecommunication

Indian Institute of Technology Kharagpur

Lecture 04: Fourier Series (Contd.)

Okay so, now what we will do we have already discussed about how to
represent a particular vector with respect to another vector now what we
will do on this paper which is a two dimension we already know we will
just take a two vectors one is x.
(Refer Slide Time: 00:39)

30
g = cx x + cy y + 0
The other one is y now we just do the same exercise we will try to
represent a particular vector let say any vector let say this is g okay
again what we need to do first g needs to be represented by x so we will
again putting a perpendicular an accordingly will you calculate c or let
us called that cx okay, g also can be represented with respect to y so I
will again put a perpendicular an I will get another value c2 or cy let say
so two constant we have got, now I am trying to represent it with respect
to these two vectors.
So what I do I actually take both the vectors so basically, I am taking a
liner combination of these two vectorial representation okay so g earlier
I was just representing with respect to some c into x now I will do a
linear combination so some cx into x + some cy into y and if we can
properly minimize so right now we have just shown how to minimize
that but if you just put it this way and try to minimize it you will see that
as long as x and y or orthogonal to each other, will always be able to

31
completely represent this without any error with zero error because what
will happen if you represent by this, this will be your error now this can
be directly with represented by this one so error will be further
represented by the other vector and then the residual error that is left that
become 0.
Okay so the lesson that we have learnt that if you can so two things one
is if you can whatever dimension you have the suppose vectorial space
so this is a two dimensional vectorial space, in that vectorial space
whatever dimension you have, so it is a two dimension so two
orthogonal vector to each other. If you can find out those any two so it
does not have to be this x and y I can represent g even by this x and this
y as long as these two are orthogonal to each other okay so if you can
take any two mutually orthogonal vector because the dimension is two
with those two vectors you can always perfectly represent any vector in
that particular space, if I wish to represent it in three dimension, I need
to have three vectors so that is why they actually represent it with
respect to x y and z.
So three co-ordinates you need or three vectors you need, okay as long
as this always you know that xy and z are orthogonal to each other all
you have to do is we have to mean’s worry about the dimension of the
space you are talking about vector space you are talking about and you
have to worry about or you have to find out, if the dimension is n, n
number of mutually orthogonal vector as long as you can do that then
you get a complete we call that basis set, and with those basis set any
other vector. In that particular vector space you can represent them
without the error.

So, and eventually what happens it is just a linear combination of all

those vectors.okay, so this is the fantastic thing which has been already
understood and which also you are also familiar about this okay so you
already know that this is how it should be, what will now do will try to
see if a signal any general signal can be represented by some constituent

32
basis, this is what will be targeting and that is the basis of Fourier series
so let us try to do that let say I have a signal g (t).
(Refer Slide Time: 04:36)

g(t) ≃ cx(t)
e(t) = g(t) − cx(t)
I wish to represent it with respect to some non signal x(t) so x(t) is the
non signal okay so this is what I wish to do okay, so I am saying that if
only x(t) is there like vector I will be just multiplying it with the scalar
quantity okay so I will say it is almost represented by this c *x(t) now all
we have to do similar thing if I do this representation of course there is a
error which is called that e(t), so therefore I can write g(t) must be equal
to cx(t) +e(t) the error signal e(t) so e(t) must be all we have to do is now
minimize the error. That is what we have learnt earlier that in the vector

33
we have minimizing it there it was easier because we are just seeing
where it becomes projection they are the error vector will be of
minimum length okay that means a measurable quantity vector we are
talking about length and length is nothing but plus second norm that is
the length square of vector so that is how we where minimizing the error
basically we are doing error length minimizing that is what are doing
here also we will try to or target to minimize similar things so what we
are trying to do is let say this g(t) and x(t) they are defined over a time
let say t1 to t2 okay so it is initially for a simplicity we are saying that
this signals are time bounded it is only defined from t1 to t2 both g(t)
and x(t) so we want to now minimize the overall error, so what do you
minimize we have already characterized a signal that is actually the
energy of the signal so we will target to minimize the energy of this error
signal.

(Refer Slide Time: 06:50)

34
 t2

∫
Ee = e 2(t) dt
t1
t2

∫
= [g(t) − cx(t)] dt
2

t1
d d
[g (t) + c x (t) − 2g(t)cx(t)] dt
2 2 2
Ee =
dx dx
t2

∫
= 2cx 2(t) − 2g(t)x(t) = 0
t1

35
 t
∫t 2 g(t)x(t) dt
c= 1
t
∫t 2 x 2(t)dt
1
So our target should be let us calculate the energy of error signal that
should be integration e square t dt from t1 to t2 right now e(t) we can
replace t1 to t2 g(t) – c x(t) whole square dt okay, so this is what we are
targeting to minimize, g(t) is a known parameter or known signal or this
is the target signal that we are trying to minimize x(t) is a known signal
with which we want to approximate the other signal here with respect to
what we are trying to minimize it is to see think about that vector also
there also we are doing the same thing.
With we are trying to get a optimal c for which the error will be
minimized same thing we will be doing over here so here we should
minimize with respect c that means we will differentiate it with respect
to c and make it 0 find out that value of c that should minimize so we
just divide this. Now force integration is over t and differentiation is
over c they have no dependency so I can interchange these two right, so
I can take this thing inside differentiate g suppose I open up this square.

So it should be g2(t) differentiating with respect to c that will be 0 then –

2c g txt so if I differentiate it with just – g(t) x(t) right and differentiating
with respect to c remember, so it was c g(t) and x(t) the dc of c because
g(t) and x(t) does not depends on c, so that will be coming out it is just 1
okay the next is +c2 x2t right so c2 x2t should give me 2C x2t dt is must
be 0, this is alright so I have just taken dt c inside and I started
differentiating it I know that t g txt they do not depend on C.
So those termsI have juts open the square and they started differentiating
it okay so I have got this values, now if this should be 0 then
immediately I can find out C because C does not depend on the

36
integration variable, that is t so I can write this as 2c integration x2 t
from t1 to t2 dt = integration t1 to t2 gt xt dt I can write this or c
becomes this is fine oh so there should be a 2 because it is 2gt xt so there
should be a two cancelled so c becomes t1 to t2 d(t) x(t) dt / g txt dt
integration x2dt,so basically this is what I do get okay that is a very
important criteria. If now think about that vector an logic when I was
putting the vector thing what was my c.

(Refer Slide Time: 10:49)

t
∫t 2 g(t)x(t) dt
c= 1
t
∫t 2 x 2(t)dt
1
⟨g, x⟩
c=
⟨x, x⟩
t2

∫
⟨g(t), x(t)⟩ = g(t)x(t) dx
t1

37
 t2

∫
⟨x(t), x(t)⟩ = x 2(t) dx
t1

C was dot product g x / dot product x x right now I can define my dot
product over this signal space so that dot product now g (t) and x(t) if I
defined my dot product as this one wherever the signal is defined from t1
to t2 and then g x(t) dt okay so if defined this as dot product what
happens immediately you can see, see all the things are getting
connected now if I now put x(t) the dot product between x and x that
happens to be the energy of x there also characterizes by ideal proportion
that the way I was putting dot product.
Dot product is eventually in a vectorial space giving the, the
measurement that is the distance of that vector here also the dot product
is giving the, the measurement that I have characterized so far, that is the
energy of the signal so x(t), x(t) dot product will be t1 to t2 x2t dt this is
nothing but the energy that is very good the dot product in vector also
gives the measurable quantity that is the length of the vector dot product
defined by me over here also gives the energy which is the measurable
quantity of a particular signal.
okay.From here again we can talk about orthogonal so when a signal will
be orthogonal to another signal when that similarly in vector if I cannot
represent a vector by another vector then we say that these two are
orthogonal that means this corresponding c must be 0 here also same
thing I have defined my c as this okay, or if I just write it one more time
the c is.

(Refer Slide Time: 12:42)

38
 t
∫t 2 g(t)x(t) dt ⟨g(t), x(t)⟩
c= 1
t
=
∫t 2 x 2(t)dt ⟨x 2(t)⟩
1

Integration t1 to t2 g(t) x(t) dt divide by t1 to t2 x2t dt which is nothing

but dot product of g(t)x(t) the way I have defined dot product and dot
product of x(t), x(t) right so if c has to be 0 that means x g(t) cannot be
represented by x(t) and g(t)and x(t) are orthogonal and c must be 0
immediately the condition comes at this must be 0, so therefore g(t) x(t)
must be 0 or we can write g(t)x(t) dt = 0 so that is the famous
orthogonality condition for a signal right, so this whenever we talk about
the signal two signals x(t) and g(t) will be orthogonal to each other.
As long as this condition is satisfied if the signal is defined from t1 to t2
then the multiplication and integration within that limit must be giving
the 0, okay so this is the criteria for orthogonality and similarly we have
seen the representation okay now we will do some more things, so
whatever we have talked about right now this was actually good or a real
signal we have talked about all real signal and that is why energy
calculation was easier because energy is a measurable quantity right. So
in a measurable quantity I want something real and I was a doing
squaring and integrating it now if I have supposed x(t).

39
(Refer Slide Time: 14:32)

t2

∫t
Es = x 2(t) dt
1
t2

∫t
Es = | x(t) |2 dt
1
Okay I wish to do this integration that is t1 want t2 dt this is my energy
now if the x(t) happen to be a complex signal which has a real path and
imaginary path if it is a complex signal then I do not have this evaluation
because then immediately I can see that Eg will not be a real number, so
to do that there is a way to do that so Eg I can define in this fashion, so
basically the definition of Eg then happens to be that if it is a complex.
If it is real it will be just x2t if it is complex then I actually take that
complex and take the complex conjugate of that and multiply okay, and
then I integrate okay so that is just a slight difference in definition the
Energy definition for a real signal in slightly different from the energy
definition of the complex signal but the thing is that because complex
signal also captures the essence of, I mean a real signal so t is happening

40
is here, if it is a real then it will just become a complex conjugate is the
signal itself and if I just multiply that will be just square right so if it just
captures the other definition.
Also it is not away from that other definition, so this is my energy imme-
diately what will happen if suppose I have a now the representation
changes slightly.

(Refer Slide Time: 16:09)

g(t) ≃ cx(t)
e(t) = g(t) − cx(t)
t2

∫t
Ee = | e(t) |2 dt
1

minEe w.r.t. c
I have signal g(t) and I wish to represent it with respect to a singal x(t)
but the only difference is now that this g(t) and x(t) are no longer real
nobody has said me that these two has to be real this can be now
complex okay, so now I will try to do the same thing same derivation

41
little bit complex it will be but I wish to again derive the similar
definition for orthogonality for the representation of those c value,
optimal c value okay so that is something we wish to do okay so now let
us try to see. What will be error that is actually same thing again g(t) –
cx(t) now we have to minimize the energy of this error signal so the
energy of the error signal now because it is a complex signal, and here
even c can be complex the coefficients also can be complex that we will
see later on.
Okay, so if that is the case what we need to is we need to take that
complex conjugate multiplied with error signal and then integrate over
the limit whatever limit it has let say the limit is or just t1 t2 to whatever
okay. So immediately happens to be g(t) – cx(t) and now all I have to do
is I have to actually evaluate these things and I have to this is my energy
I have to minimize this with respect to c but this is not going to be as
easy as we have done, okay so direct differentiation will not give me
because there will be c and it is complex conjugate so I cannot
differentiate with respect to c so I have to device another tool to
minimizing okay, so this is something I will discuss now this Ee.

(Refer Slide Time: 18:18)

42
 2 2
t2 t2 t2
1 1
∫t Ex ∫t1 Ex ∫t1
Ee = | g(t) |2 dt − + c Ex − g(t)x*(t) dt
1

| u + v |2 = | u |2 + | v |2 + uv* + u*v

from the 3rd term

t2 t2 t2
1 1 1
Ex ∫t1 Ex ∫t1 Ex ∫t1
2 2
c Ex + g(t)x*(t) dt − c Ex g*(t)x(t) dt − c* Ex g(t)x*(t) dt

The way we have defined I just write it in this fashion let see whether
this is correct so t1 to t2 will prove that this is the correct representation
– Ex is the energy of signal X(t) see why I have represented this in this
fashion it is because I want to minimize it with respect c right, these two
part if you see these two part are no longer depended on C okay, so all I
have to do is somehow choose c so that this become 0 okay then only I
can say that this will be minimized with respect c so I have just device a
separate rule.
For evaluating it because direct differentiation was not possible I was
getting c,c complex conjugate and then differentiating with respect c that
was difficult okay but we have to first prove that this is the actually the
error square or error modular square integration we just keep calculating
that you will see that this is the actually the same, so basically if you just
evaluate this part okay so this you keep it as it is if we just calculate this
part so you have a this – this modular square.
Now I will just write how to evaluate suppose I have a variable u and v
which are complex number I need to evaluate this generally we can
write it in this fashion modules square mod u square plus mod v square
instead of 2uv it is written generally u is * + u* okay so this is how the
complex numbers mean square of mod of complex numbers are
represented okay, so if I just evaluate this portion with respect this what

43
do we get ? so this is actually u + v so this is becoming my u and this is
becoming my v.
So just evaluate it what do we get here we get c mod c2 Ex right because
root Ex is not a complex let say energy it is must be a real number okay
then mod of this okay so mod this square so that should be 1 / Ex and
then and modulus and integration that has nothing to do, does not have
anything to do with the integrations so it will go inside so I can write t1
to t2 g(t) x*(t) mod square t right so here I think we have forgetting that
and then it is just uv* + vu means u*v so uv* if you put so what we get.
We get C root of Ex 1/ root Ex of course there should be a – okay and
we have that same integration t1 to t2 but we have to put a * of that
right, uv* so * of that so that must be g*(t) x(t) x*(t) and it is complex
conjugate a and will give you that x(t) of dt okay so we get this – we get
then c* root Ex into 1/rootEx and then we have this integration t1 to t2
g(t) x*(t) dt right, Fine, now let us try to see if this is exactly equivalent
to our g(t) – cx(t) mod square it is eventually because what is happening
we can see this Ex, Ex gets cancelled, this Ex, Ex gets cancelled okay so
what do we get so we had let see we had t1 to t2.

(Refer Slide Time: 23:24)

44
Integration g(t) – cx(t) mod so t1 to t2 right so we have if we start
evaluating it so again we can this square we can actually evaluate it with
respect to uv mod square so this will be mod g(t)2 and then +cx(t) mod
square these two things will be coming and then g(t) c* x*(t) right and
there should be a – and then – g*(t) cx(t) right and if we just see the
previous one it is just giving me the same thing okay so if you just see
everywhere I have this C okay so if just match it so what do we get we
have got this square so this square term.
If you have evaluated we have got this c c2 Ex right which is actually x
mod x2(t) right so this term and this term so this particular term and this
term exactly matches then our g(t) exactly matches g(t)2 and that term is
also present over right, so which is this term that is already there so this
term is matching the next term is matching this particular terms gets
cancelled and the other two terms are already there right so this
representation exactly id the representation if we just do it yourself.
We will see that it is exactly the representation of this one so therefore
all we are trying to say that the representation we have put over here for
Ee that is exactly the representation of that error mod square dt, okay so
now harm in doing that representation but by doing this representation
immediately we can see that these two part are free of C so therefore all
we have to do is we have to minimize this or we have to make it 0, so to
make it 0 this inside term if I just equate it to 0 I get my C and that
should minimize the whole E because whatever else I do this is mod
square. Term so any other thing other than 0 will actually increase the
value of Ee because this term is free of c this particular term to both the
terms are free of c is this terms is anything other than 0 because there is
mod square and addition so it will always increase the value, so the
minimum value that I can get where this particular part is 0 so whenever
I put this particular part 0 immediately, what do I get for my c
calculation so c happens to be.

45
(Refer Slide Time: 26:42)

t2
1
Ex ∫ti
C Ex = g (t) x* (t) dt

1 t2
Ex ∫t1
C= g (t) x* (t) dt

t
∫ 2 g (t) x* (t) dt
= t
∫t 2 x (t) x* (t) dt
1

46
So c root Ex must be equal to 1/root Ex t1 to t2 g(t) x*(t) dt right so
that must be the case immediately I can evaluate c that is actually 1/ root
Ex integration t1 to t2 g(t) x*(t) dt so for a complex signal now you can
see we have t1 to t2 only difference is earlier it was g(t) into x(t) now I
have a x complex conjugate t dt this integration and Ex can be written
again as mod x(t)2 or x(t) into x*(t) dt, so now I can define just redefine
my for a complex signal by dot product as this that suppose I have a
signal.

(Refer Slide Time: 27:43)

t2

∫t
< g (t), x (t) > = g (t) x* (t) dt
1

47
g(t) and x(t) and I wish to do the dot product so dot product can be
defined as this g(t) x*(t) dt so this should be my new definition of dot
product in a complex signal and immediately if I take the dot product
with itself I get back the energy no problem in that and the orthogonality
condition is the c must be 0 that means g(t) must not be represent it with
respect to x(t) so if c is 0 them immediately this dot product must be 0 or
this particular integration must be 0 so this is what we can see from our
derivation next what we will do will apply this to actually get back the
Fourier Series representation.

48
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
on
Analog Communication

by
Prof. Goutam Das

GS Sanyal School of Telecommunication

Indian Institute of Technology Kharagpur
Lecture 05: Fourier Series (Contd.)

Okay so far what we have done is, we have seen a vector analogy and
try to put that into signal space okay. so we are now targeting that a par-
ticular signal must be represented by another signal and then from the
vector analogy we also know that in the signal space, if we can now, we
also now know what is the characterization of orthogonality .. two sig-
nals. So in a particular signal space if we can find out all the set of or-
thogonal signal, then we know that exactly a particular signal can be rep-
resented in linear combination of those orthogonal signal.
This is something that we have already understood and their linear coef-
ficients also how to calculate that also characterized now okay. So we
know that exactly how to evaluate those optimal c which minimizes the
whole error and if I have the complete definition, that is called the com-
pleteness of basis set. So if I have complete definition of all the basis set

49
that are possible in that particular signal space, I will be able to represent
a signal. So this what Fourier, this was the background that Fourier start-
ed, means analyzing signal and this is how he started things.
So now, our target should be that how do we represent, or how do we
find out all the orthogonal signal in a particular signal space, so initial
Fourier started with trigonometry signal.
(Refer Slide Time: 01:59)

So initially he started with let say we take this Cos m2πf0t. okay, so he
started with this and correspondingly I can also define Sin m2πf0t okay,
so these two signal we are targeting so whenever m = 1, so it is just Co-
sine signal with frequency f0, where f0 is = 1/T0 okay, so T0 is the time

50
period of that signal. So whenever we are talking about Cos signal or Sin
signal it is periodic signal we know that and the thing is that, if the con-
stitute signals are periodic and then what it will represent must also be a
periodic signal.
So what initially Fourier devised, okay I can represent signals but all
those signals must be also periodic signals that I am targeting to repre-
sent, that can be any signal it might be a triangular wave something like
this, let say this kind of thing but it must be periodic with period let say
T0, it can be a pulse and of-course it can stretch to minus infinity. So it
can be any periodic signal and hidden as started saying can we represent
this?
In a particular fashion, then he started enquiring the basic signal that we
know. Even the electromagnetic waves that propagates with cosine si-
nusoidal or sinusoidal , so can we represent all these signals with respect
to this cosine sinusoidal or sinusoidal but before doing that we need to
know the orthogonal signals in that signals space, so what he started do-
ing.
(Refer Slide Time: 04:04)

51
1 T0
2 ∫0
2cos(m2π f0t)cos(n2π f0t)dtm ≠ n

1 ∫T0 cos[(m + n)2πf t] + cos[(m − n)2πf t]dt = 0

2 0 0 0
He started enquiring whether 2 Cosine sinusoidal, cos m2πf0t for differ-
ent values of m okay, so let us say 1 is cos m2πf0t and another is cos
n2πf0t, whether these two signals are orthogonal or not. So if they need
to be orthogonal what we have what prove, it is a real signal cos, is a
real signal, so all we need to prove is and it is the periodic signal, so
within one period if the property is true, it will be true for all other peri-
od. It will be just the making the similar pattern.

52
So orthogonal means the integration must be 0, there multiplication and
integration must be 0, that is what we have understood g and x will be
orthogonal if g(t) x x(t) dt over a particular time period must be 0. So if
within 1 period it is 0, the other period will also will be behaving simi-
larly. So entire period or entire signal duration it should be 0. So this is
what we wished to prove, is it true.
So we will probably do it from let us say-T0/2 to +T0/2 we can even do it
from 0 to T0 any periodical, we can do it from 0 to T0, so over a single
period we are trying to evaluate what would be happening if we do this
integration because we know that, that is the prove of orthogonal, if this
is 0 for two different signals of-course, if it is same signal then we
should not, we should get the energy back right. so if it is two different
signals that means two different values of m and n, or we are saying m!
=n for this condition we are trying to see what will be the value okay.
So let see let us try to evaluate this, take out ½ it should be, now just put
a trigonometry form to cosine into cosine, so that should be cos a + b+
cos a-b, so it should be cos( m +n)2πf0t + cos( m –n) we are assuming
that of-course cos negative that will be again become positive so, here
either m is bigger than n or n is bigger than m it does not matter, it all
will be same. Now let see the first integration 0 to T0 cos okay. So if m is
!= n then both these are non zero.
So what is happening this is another cosine and whatever the f0 is m + n
or m – n it will be either 1 or bigger than 1, so over that period T0 if I in-
tegrate a cosine sinusoidal which is, see already 2π is there that means
either it will just make the full cycle or it will make multiple full cycle
and across the signal, if it completes a full cycle or multiples a full cycle
always the integration what will happen, sorry okay this, always the in-
tegration will be 0

53
As long as it is cosine sinusoidal signal and we know that it is either
completing a full cycle, if the difference is, let say this addition of-
course will be more than full cycle, it will be 2, 3, 4 or some integral
number. So multiples of full cycle it is completing, so if I integrate that
should be 0 definitely, so this part will be 0, even this part will also be 0
because, it is either it will be 1 full cycle because m! = n, so m-n at least
it will 1 and it might be even bigger than that.
So whatever it is, it is either a full cycle or multiples of full cycle, if I in-
tegrate again it will be 0, so the overall thing will be 0 if m! = n. so
therefore we know that all the cosine sinusoidal signal, with all the val-
ues of m going from 1 to infinite all orthogonal to each other. That is
very important findings, so he could find in that signal space in the vec-
tor space if you remember when we had this page two dimensional, we
had only 2 orthogonal vector.
Now we actually have infinite orthogonal vector right, any value of m
you put, so m starting from integer value of course, starting from 1 and
then going up to infinity, any value m or n you put, as long as m!= n, all
of them are actually orthogonal to each other. And these are actually
called the harmonics. So you have a cosine sinusoidal, you have twice
frequency cosine sinusoidal, thrice frequency cosine sinusoidal and so
on. All those frequencies are all those harmonics 1st, 2nd, 3rd, 4th, are
included and they are actually orthogonal to each other that is one find-
ing. The second part is, if I have a dc value which is just 1.
(Refer Slide Time: 10:08)

54
 T
∫0 0 1 ⋅ cos(m2πf0t)dt

∫ sin(m2πf0t)cos(n2πf0t)dt

So between that 0 to T0 whatever I have talked about, so it just takes the

value 1 okay. If this is the case will it be orthogonal to all those cos, of
course because what will happen, this as to be multiplied with all those
cos and again it will be integrated, so if I just do the integration it will be
1 into cos m2πfot dt and integrated over 0 to T0 right, so again it is co-
sine sinusoidal with m taking value 1 to infinite. So therefore this is ac-
tually because the 2π is already there and fo = 1/t0, so basically this is
either a full period or some multiples of full period.

55
So it will obviously be again integrated to 0, so even this particular sig-
nal where it is just taking a constant dc value that is also orthogonal to
all those cosine sinusoidal, so the dc value which for a single period it
takes same value so I consider all the periods it will just take a dc value,
so the dc value even is orthogonal to all cosine sinusoidal signals right.
So what we have already demonstrated.
And the next part is, this is also orthogonal to of course the dc as well as
all those cos values, will now again show that if I multiply with some
cos let say n2πfot for any value of m and n. now we do not restriction
because even if m and n are equal these two are two different signals
okay. So if I just do this integration with the same logic it will again be-
come some cos or sin and then within again either full period or multi-
ples of full period, the integration as to be done and it will always be 0.
So therefore what we have done is we could identify now few mutually
orthogonal signal and I should not say few it is infinite okay. so all the
cos m2πfot for all values of m positive integrals, all the sign to n2πfot
and dc value. These are all, infinite numbers of signals, are all mutually
orthogonal to each other. The other part which I am not covering over
here which is little bit more involved, that if I take this whole set that is
the complete set of this basis.
Like 2 dimensional vector we were saying, x and y that makes it com-
plete because only with those two vectors I can represent any vectors. In
this signal space for periodic signal we can actually prove, which will be
little bit more involved, so we are just giving the outcome that if I just
take consider these basis set, this actually make a complete basis set.
These are all the basis set that can be defined in this particular represen-
tation okay, so if I consider all of them I could actually finish represent-

56
ing all of them and any periodic signal therefore should be,mean we
should be able to represent any periodic signal with respect to these ba-
sis functions therefore okay. So what should happen then, I know that
now my.
(Refer Slide Time: 14:05)

∞ ∞
g(t) = a0 + ∑n=1 ancos2πnf0t + ∑n=1 bnsin2πnf0t
T
∫0 0 g(t) ⋅ 1dt
1 T0
∫
a0 = = g(t)dt
T0
∫0 1 ⋅ 1dt T0 0

57
 T
∫0 0 g(t)sin2π nf0t)dt
2 T0
∫
bn = = g(t)sin(2πnf0t)dt
∫0 sin (2π nf0t)dt
T0 2 T0 0

∫0 0 g(t)cos(2π nf0t)dt
T
2 ∫T0
an = = g(t)cos(2πnf0t)dt
T0 0
∫0 0 cos2(2π nf0t)dt
T

Any periodic signal g(t) I should be able to represent them linear combi-
nation with the linear combination of all these basis function right. so
therefore I can write 1st is dc, so a0 1+ anx cos 2πnfot where f0 is actu-
ally 1/T0 and this T0 is the period of this signal g(t) okay and I have to
take because all the basis has to taken, so I have to take a summation
over all possible cos or in thing or all possible values of n + the sinu-
soidal is there, so bn sin 2πnfot, summation that is the famous Fourier
series representation.
So because we have proven, that these are basis set which are mutually
orthogonal to each other and because we are also stated that these are all
that I can get in that signal space, so this is actually a complete definition
of the basis set. So therefore without error I should be able to represent
g(t) as the linear combination of all these basis set, where the coefficient
of linear combination are this a0, an and bn okay and how do you evaluate
these values? Let say a0 how should we evaluate that.
This is something that we have already seen; the evaluation of a0 should
be with those dot products right. so all we have to do is g(t) with respect
to or whatever signal is, therefore 1 dt must be integrated over a period 0
to T0 and this 1 should be also multiplied with itself, so 1 x 1 dt from 0
to t0. This has to be done; as long as we are doing this we will be able to

58
represent the whole thing right. So a0 is like this similarly an must be
whatever the g(t) is we multiply that with cos 2πnfot, integration should
be 0 to T0 and should be cos 2 2πnfot, dt that is the Fourier coefficient if
you go back and try to see whatever you have learn in Fourier series.
This is actually that is something we have already proven, this is the op-
timal an or representing that signal okay. So similarly bn also will be we
can get the formula 0 to T0 it can be - T0/2 to +T0/2 whatever it is it
should be that single period okay, so the result will be same. so again
g(t) this time it should be sin2πnfot okay, so in your Fourier transform
you might have seen that this happens to be if we do this integration it is
just 1 or dt integrated from 0 to T0, so that should be 1/T0 a0 becomes 1/
T0 that is the whole formula right g(t)dt over the period 0 to T0
This if we do the integration it should be T0/2 if you just do that integra-
tion there will be a ½ coming out and 2 going in and if you do that inte-
gration it will just find 1. So this should be 2/T0, so therefore this thing
goes away you get 2/ T0 same things happens over here so this becomes
2/T0 that is the famous Fourier series and Fourier coefficient calculation
which we could direct now prove without any ambiguity. This is very
clear to everybody, so once you have done this there is another represen-
tation, probably whenever we.
This was all fine, we know that any signal now we can actually represent
with respect to corresponding sinusoidal. So what is actually happening
all the sinusoidal harmonics are coming into the picture, so any sinu-
soidal or periodic signal signal, it is nothing but a infinite summation of
different harmonics nothing else. Any signal can completely be repre-
sented if you can evaluate through this, the appropriate coefficients will
know exactly how to combine the coefficient appropriate will know ex-

59
actly how linearly you have to combine those harmonics to get this par-
ticular signal.
So that is a very nice representation you actually now that, which sinu-
soidal with what strength you have to take and you have to add them to
get this particular signal. In a way it is also giving us a strong to tool,
you get a signal you immediately know what are the constituent sinu-
soidal with what kind of strength should be present in that particular sig-
nal okay. So whenever start talking about what are the sinusoidal con-
stituents that it means; now we are talking in terms of what frequency
components are actually there in this particular segment.
This is where the signal gets represented in the frequency domain okay,
so if we just give one example probably, we just represent this particular
signal, so that the signal is something like this
(Refer Slide Time: 20:34)

60
Let say it is w(t) it defines from –π/2 +π/2 it is like this and then again it
is π, π/2 starts again this is 2π and so on. If I represent a signal like this
which is the periodic signal and then if I have the way we have learnt to
evaluate the Fourier coefficient. If we just calculate that we can see all
the Fourier coefficients and then we can start representing this signal
with respect to frequency component, so right now I will not give the
example fully for this because we still have not equipped our self fully
with the frequency component.
But what will do we will take this example and come back after little bit
of more analysis and come back and draw actually the frequency spec-
trum. We will see that what are the harmonics which are present and
how they are present and accordingly we can actually draw the spectrum

61
of this signal, an another representation not in time domain but in fre-
quency domain. So basically we will be saying the representation, the
equivalent representation will have like this, so f0, 2 f0, 3 f0 what are the
components it has.
So it will just give another representation in frequency domain, so here it
is just f0 to f0 the frequency is increasing and whatever are the compo-
nents of those frequencies which are represent that we will be able to
characterize. So this how Fourier series actually gives us the frequency
representation or frequency domain representation of the segment. It is
nothing it is just, it says for a periodic signal what are the whenever we
say the frequency component we are actually saying what are the basis
that it has.
And basis components means its sinusoidal
So all those frequency component whenever we talk about, we plot the
frequency domain, we are actually saying that at that point there is a si-
nusoidal i.e the corresponding sinusoidal which every frequency compo-
nent we take at that point there is a sinusoidal with that strength. So fre-
quency domain representation is nothing but every frequency component
we pick actually there is a equivalent sinusoidal that actually makes the
signal. So you take of that strength sinusoidal you add all of them you
actually get the signal back. So therefore there is equivalent representa-
tion you should say , in time you see…also in frequency domain you
start seeing the similar representation of that signal.
It is just thanks to the Fourier, he could actually give the extra represen-
tation of the signal which you will see that it will helps us in many ways
for the processing of the signal. So before going into that, before charac-
terizing that let us try to see another representation which is also Fourier

62
representation, we have already seen the representation with respect to
real signal, so right now whatever basis set we have constructed these
are all are cosine sinusoidal or sinusoidal or dc value. So those all are
real signal, now what we will do, we take a signal which is real signal
g(t).
(Refer Slide Time: 24:07)

x(t) = e j2π nf0t = cos(2πf0nt) + jsin(2πf0nt)n ∈ ( − ∞, …, ∞)

x1(t) = e j2π f0nt

x2(t) = e j2π f0mt

63
 T0

∫0
< x1(t) ⋅ x2(t) > = [e j2πf0 nt e −j2πf0 mt ]dt

T
= ∫0 0 e j2π f0(n − m)t dt
T
∫0 0 e j2π f0(n − m)t dt
= | T+0
0 =0
j2π f0(n − m)

We will try now to represent it with respect some complex signal okay,
so first of all we will try to define the family of complex signal which are
mutually orthogonal to each other, then we will say what is the again
technique is the same, then say what is the overall complete basis set,
definition of complete basis set once we get that we get the complete
representation. So that particular thing is represented has exponential
complex signals. So it is like ej2πf0nt so this is test x(t) okay, so any x(t) is
taking this value where n now can take integer but it can positive and
negative.
So n now can take any value from – infinity or + infinity okay. Now let
us first try to test whether these infinite numbers of signals we have got
are they orthogonal to each other or not, that is something that we have
to prove. The signal is a complex signals because we have this complex
part already, so 1st we have proved that this is orthogonal or not, now
we have employ the complex part or complex equivalent orthogonal cri-
teria . So that means we take x(t) or x1(t) ej2πf0nt and we take another x2(t)
ej2πf0mt. Now all we have to prove is if m !=n then should we get the or-
thogonal criteria should be equal to 0.
What’s the orthogonal criteria that is actually x1(t) x2*(t) start, therefore
it should be ej2πf0nt into the complex conjugate e-j2πf0nt and if you careful-
ly see this is the complex signal that is also a periodic signal because this

64
signal can be represented as sinusoidal or cosine sinusoidal it be cos
(2πf0nt)+j sin(2πf0nt) this particular factor, so therefore again the fun-
damental frequency is related to this f0 which is = 1/T0, so the integra-
tion needs to be just done over the fundamental typing, we do not have
to do that whole integration because we know that in one integration if
we prove this gets repeated, so same technique will employ 0 to T0 the
same signal.
So now this is let us try to evaluate this, so it should be 0 to T0 ej2πf0(n-m)t
right, so now this integration how much it should be ej2πf0(n-m)t / j2pf0(n-
m)t and we have to put the value 0 to T0. So now if we just T0 is 1/f0
right so this cancelled n !=m right, so this cannot be 0 right, so this must
be some positive integer okay. It is just ej2π into some integer that we can
represent as cos and sin. Now cos 2π+j sin 2π integer, cos2π integer that
should always be having some value ,zero so what do we get if I just
represent this it should be cos2π that should be 1 right, so we get 1 for t0
and for sin it is 0, so the j part is cancelled – I have to put this limit, so
again I put 0, it should be 1 and sin it will be 0. So 1-1 gets cancelled so
this becomes 0.
So whenever m !=n it is always 0, id n =m this becomes already 1, be-
cause e0 so we know that this particular things are all orthogonal as long
as m!=n, that of-course with himself it cannot be orthogonal, so with all
other signals they are mutually orthogonal to each other. So this is some-
thing we now have proven, any value of m and n you take as long as m !
= n we could from the basic principle of the orthogonality of the com-
plex signal we could prove that they are all mutually orthogonal to each
other.
And again we are saying we have proved we have sating that these are
the all the signals that we required to represent all the entire basis set .

65
So as long as we are taking that to be true we know that this xt can be
represented with these things.
(Refer Slide Time: 30:28)

∞
cne j2πt0 nt
∑
g(t) =
n=−∞

T
∫0 0 g(t)e−j2π f0nt dt
cn =
T
∫0 0 1dt

1 T0
T0 ∫0
Cn = g(t)e −j2πf0 nt dt

66
So now what we have to do is this g(t) must be a summation of some co-
efficient let say Cn ej2πf0nt and n goes from – infinity to + infinity, that is
famous Fourier series representation, where tht Cn must be evaluated
with the same criteria of complex, so that must be integrated over that
period g(t) into complex conjugate of x, so that is why it becomes e-j2πf0nt
dt divide by this into the complex conjugate of that so they will cancel
each other it will be just 1dt this becomes 1/T0 integral from 0 to T0
g(t)e-j2πf0nt this is the reason why the inverse 1 or the coefficient one
whenever you calculate you have put e-j2πf0nt, and
whenever you are actually representing g(t) you get e+j2πf0nt.
Later on from this series will go to Fourier transform and that is why
Fourier transform always get e+j2pf0nt, and inverse transform always
get – or vice versa okay. so this happens due to signal, so no what we
have seen that we can represent any signal again with respect to another
basis set which are complex. In next what we will try to do is, we will
try to get a relationship between these two representations. What the ex-
ponential Fourier series it tells us and what the trigonometry Fourier se-
ries tells us and what is the relationship between these two representa-
tions.

67
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
on
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 06: Fourier Series (Contd.)

Okay, so what we have done so far we have try to discuss about Fourier
series right, and especially the origin of Fourier’s series it means how
Fourier’s series has come into picture. So I will just quickly recapitulate
what we have achieved so far, so given a vector analogy we have said
that any vector if you know the dimension or whatever vector space you
have if you know the dimension and if you know means as many
orthogonal vector can exist in that particular dimension.

Suppose you have a two dimension thing and you have two orthogonal
vector, so what we have said that as long as you know this two
orthogonal vector you can represent any vector in that space in that two
dimension space with respect to these two vectors especially the linear
combination of these two vector and these two orthogonal vector are
actually termed as the basis vector, right. We have taken that same

68
analogy into signal so our target there was of course we have started
targeting a periodic signal.

So what we have said given any periodic signal that can have any nature
can we really find out some known signal a linear combination of them
can give me back that same periodic signal, so any signal possible in this
world which are periodic with some period let us say T0 okay.
(Refer Slide Time: 01:52)

T/2

∫−T/2
⟨x1, x2⟩ = x1(t) x*
2
(t)dt
So can we get set of basis signal okay, or set of signals which are already
known like this orthogonal signal so already known and can be represent
and can we guarantee that any signal in that particular signal space can
be represented as linear combination of all those known signals. So what
in that aspect we have defined few things I am not going it to the details

69
of those things so I have defined what we mean by orthogonality
condition in signal space, okay.

So if two signals x1(t) and x2(t) if these two signals are there both are
periodic signal with of period T0 let us say and then how do we actually
prove that these two signals or how do we test that these signals are
orthogonal, so this is something we have already evaluated, okay. We
have evaluated two conditions one is if both signals are real then we
have evaluated one condition and then we have said if both signals are
complex then also we have evaluated one condition.

So the ultimate condition which we have also proven is that x1(t), x2(t)
over that period let us say –T0/2 to +T0/2 this must be 0 if both the
signals are real, and at the same time if signals are complex then you
have to take x1(t) and a complex conjugate of the other signal x2(t) and
you have to do the same integration over that same period and this must
give me 0, if x1(t) and x2(t) are orthogonal to each other.

So what we have to now do we have after that we have said we should

have a quest of this known signal all the orthogonal signal in that signal
space we must be able to find out as many are there, okay. So in that
process we have defined two things one is trigonometric Fourier series
that means all the signals must be decomposed into means series of
sinusoidal or cosinusoidal so that is one another one was exponential
Fourier series again all the signals must be decomposed as linear
combination of exponential functions, right.

Remember if the signal targeted signal has a period T0 all those

exponential signal or it is a complex exponential signal all the
exponential signal also must have period of T0. So in that process
trigonometric one we have already done exponential we have said that.

(Refer Slide Time: 04:40)

70
 1
xn = e j2πf0 nt f0 =
T0
∞
1 T0
T0 ∫0
cne j2πf0 nt g(t)e −j2πf0 nt dt
∑
g(t) = cn =
n=−∞
ej2πf0nt
okay, so that is the signal we are targeting and we are calling this
as all those xi(t) or xn(t) okay, so basically it is a as you can see it is
exponential and you can put Euler’s theorem and then you can get
1cos+jsin and then the period will definitely be 1/f0 right, so f0 is the
frequency so for any value of n there will be a period which is defined
by 1/f0 which is called as T0, okay so this period will always be there.

So all the signals are periodic signal effectively with period this and
remember these are all complex signal so we need to orthogonality
condition we need to actually prove it from that complex perspective or
complex signal perspective that is the first thing, okay and with this we
have actually shown that these signals are orthogonal with respect to

71
different values of n, right so we take different values of n you can
immediately show that it is orthogonal so whenever suppose we take two
values m and n and then you put that orthogonality condition that we
have just discussed you put that you will be seeing that if m is not equal
to n it will be always 0, okay.

So these signals are all for different values of n up n starting from I

mean 0 to infinity and -infinity so for all this values of n they are all
orthogonal to each other. There is another aspect of it in vector also we
have told that eventually what we were targeting any vector we were
targeting to represent by another known vector right, and in that process
we are targeting the error of this representation so we have said that this
is that error so if suppose this is my g and this is my x and if g has to be
represented by x where given some linear coefficients Cx and then this
was the error and our target was to minimize the error same thing we
have also done with respect to the signals and then we have evaluated all
the values of Cn is required for different signals right, so that is
something we have already evaluated.

So it is the error that we are targeting to minimize right, and eventually

we also said that in the vector space suppose in this one we are just
trying to represent g which is a two dimension vector by one vector so
that is why there will be always a error we can minimize that but there
will always be a error, but if we have another vector which is orthogonal
to this x let us say y we can always completely represent g with respect
to linear combination of x and y we know that already in a two
dimensional space.

We have two orthogonal vector we can always represent that, so that

means if you have a completeness of the basis definition whichever
space you are defining if all the orthogonal vectors in that or all the
orthogonal basis vector are already defined then you can always have a
linear combination and you can make your error 0 same thing happens to
the signal if I have a complete definition of basis set then I can always

72
represent a signal with respect to linear combination of them and we also
know how do we calculate the individual coefficients.

So there is just a complete recapitulation of what we have achieved so

far, okay so with that aspect we have talked about completeness and we
are just saying without prove that for all values of m let us say all values
of n, n going from -infinity touching 0 to +infinity okay, every integer
value it takes if we take all this then this makes a complete basis set this
particular ej2π. So we are just stating it without prove and then we can
say any particular signal g(t) which has a period T0 it might have inertia
that can be always represented as a linear combination so with
coefficient Cn ej2πf0nt n going for - infinity to + infinity, right.
So this is something we can always talk about where this Cn’s are
evaluated just by that error minimization process which was evaluated as
1/T0 integration whatever that g(t) is there e-j2πf0nt dt okay, and it should
be integrated over that period let us say –T0/2 to +T0 so that is the
Fourier series and corresponding coefficient we already know about,
right this is something we have already know. Now let us try to see if
this is the representation in this representation what do we get, we have
already seen that in trigonometry representation we actually get different
sinusoidal right, so at different frequency this is also exponential
representation so it must be a sinusoidal it is a complex sinusoidal, so let
us try to see what do we get…

(Refer Slide Time: 10:25)

73
 T0
*
1
T0 ∫− T0
2
[c−n]* = g(t)e −j2πf0(−n)t dt
2

π
1 2
T0 ∫− T
= g(t)e −j2πf0 nt dt = cn
2

cn = cn e −jθn

c−n = cn e jθn
So first let us try to evaluate what do we mean by this okay, so C that
coefficient evaluated at –n and I am trying to take the complex conjugate
of that, so let us first put what is C-n that should be whatever my signal
is let us say my g(t) e-j2πf0 now here that coefficient should come at that
coefficient it has to be evaluated so that particular coefficient –nt dt it

74
must be integrated from -T0/2 to +T0/2 and there is a complex conjugate
of this so the whole thing has to be conjugated right, so it should be –n I
forgotten that right.

Because we are evaluating the coefficient at –n so corresponding to the

coefficient should be –n and if this is –n what happens this -,- it becomes
+ right and then we are taking complex conjugate so integration has
nothing to do with it so it will come inside the integration g(t) is a real
signal we have already talked about that, that we are evaluating the
Fourier series of a real periodic signal g(t) okay.

So g(t) is real so a complex conjugate of that will be just g(t) and the
complex conjugate of this one so what that will be it is already -,- + and
you take the conjugate of that, that should be again – so I should
eventually get 1/T0 –T0/2 to +T0/2 g(t) e to the power again get back -,
so there are two negation -,- becomes + and then you take complex
conjugate so it again becomes -.

You can identity this, this is exactly Cn right, so what a very nice thing
we have observed out now, what happens whenever we evaluate this
coefficients C-n is actually a complex conjugate of Cn okay, so always
C-n for any value of n they are always a complex conjugate of Cn what
does that means. Let us say my these are complex number so it must
have a amplitude and a phase so I can represent that as and then there
should be a phase which is let us say e-jθn okay, I can represent it this
way.

Because is a complex number so it must have a amplitude and must have

a phase right, so that is all we are doing we are just putting this
representation as amplitude and phase, okay. So this is the case what
must be the complex conjugate of that the C-n must have amplitude
remains the same because it is complex conjugate so it must be and this
will be just conjugated so e-jθn complex conjugate will be +θn.

75
So what is happening if because now every coefficient have two
parameter if I can see it has a amplitude it has a phase, so I can have
eventually two plots one is with respect to this n I can plot the amplitude
so what happens to the amplitude if at n positive I get some value at n
negative the amplitude should remain the same right, because for Cn
also it is |Cn| for C-n also it is |Cn| right, so the amplitude should remain
the same therefore what do we get for different values of n the values
might be different.

But whatever it is at –n the value should be same right, so we get a even

symmetry in terms of the plot right, whatever we get at positive half we
get same thing at the negative half, okay so that is why it is call even
symmetry and for phase just the opposite whatever we get at positive we
will getting just the negative of that right, because you can see if it is –
θn it should be +θn if it is +θn it should be –θn so the angle should be
same but it is just negated right, so we odd symmetry for the phase plot.
So basically what we are getting now we have a signal g(t) right, it is
continuous in time what we can see now that g(t) as long as it is periodic
it can be decomposed into multiple complex sinusoidal this is something
we can immediately observe and those sinusoidal for different values of
n it is actually representing different frequency component if n=1 it is f0,
n=2 it is 2f0 if n=-1 it is –f0, n=-2 it is -2f0 so it is actually representing
the front frequency component or you should say different sinusoidal at
different frequency value, okay.

So all we are doing is now because we have got if you see this we have
got two equivalent representation of the same signal g(t) as a summation
infinite sum of course as a summation of sinusoidal at different because
this is just a complex sinusoidal which is known okay, so it is just that
complex sinusoidal all we need to know is at every frequency
component what kind of amplitude it has and what kind of frequency it
has.

76
So the information about the signal is completely carried over here
because this is a known thing okay, all the signals are known they are
just representing different complex frequencies nothing else, we just
have to know the corresponding coefficient because as the coefficient
changes it will represent different, different signal that is all we are
targeting, okay. So therefore times to prove Fourier what we get is a
separate representation of a signal. Now we do not see the signal in time
domain whenever we have a signal we know that it has a equivalent
Fourier series representation as long as it is periodic and immediately we
can plot.

(Refer Slide Time: 17:20)

And effectively we have to give two plots one for amplitude, one for
phase and we also know that the amplitude plot should be symmetric or
even symmetric, so therefore positive half or whatever it is it should be

77
mirror image of negative half and the phase plot should be odd
symmetric that means whatever we get in the positive half that should be
just negative of that in the negative half, right so that should be the case
so we should plot |Cn| and θn, so all we have to do is this and we know
that for which point in this independent axis we have to plot this are just
those n value or more precisely n*f0 values.

So those frequency those complex frequency component how much of

that component is present in phase as well as in amplitude, so how much
of those components if I mix them all together my desired signal will be
generated. So this representation is called a frequency domain
representation because we are actually decomposing our whole signal
into different frequency components that is all, all that we are doing is
this only.

There are two strikingly different representation when we do it in

trigonometric series and when we do it in exponential series, but the
function must have unique representation in frequency. So let us now try
to appreciate what exactly is the relationship between these two
representation, when we started representing it trigonometric series what
we were getting we are also getting two plots actually because we had
that sinusoidal as well as cosinusoidal at every frequency we were
having both the things, right barring 0, 0 was only having a DC values
so it does not have, it does not recognize whether it is cos or sin because
it is not having any frequency component, right it is Dc.

So at that point it was just a0 but rest of the case it was always having a
pair an and bn corresponding an is due to the basis function cos2πf0nt
and bn is due to sin okay, so these two frequency component represent. If
you see that representation there was nothing called negative frequency
okay, which we are getting over here in the exponential representation,
so exponential representation have the spectrum starting from -infinity to
+infinty so it has some component at the negative frequency as well as
positive frequency.

78
Whereas when we are representing it in trigonometric that also has and
this also have two plots one is amplitude one is phase. Whereas when we
are representing in trigonometric form it was always having two plot
corresponding to sinusoidal and cosinusoidal but it does not contain any
negative frequency, right so this is something we have observe so we are
always getting two plots and in this two plot what was coming it was
just those |Cn| at every value of n |Cn| and θn that is what we are getting,
okay and then we are getting this for the other case we are getting an and
bn. Now let us try to see if we can correlate these two things that is first
task.

(Refer Slide Time: 20:54)

∞
cne j2πf0 nt
∑
g(t) =
n=−∞

79
= c0 + { } + … + {cne j2πf0 nt + c−ne −j2πf0 nt} + …

cn = cn e −jθn

c−n = cn e jθn

And then we will come back and try to explain what do you mean by
this particular thing call negative frequency okay, so first let us be
concerned about this relationship so what happens we have seen that
there is a already we have explored that there is a relationship between
C+n and C-n so what we will try to do is we try to pair up these two
things, so we have a representation of g(t) as summation ej2πfont n going
for –infinity to + infinity okay.

So let us separate out the C0 term okay, so at C0 n is 0 so this becomes

e1 right, so that is alright and then we start pairing each of those n and –n
right, so here we will do a pair of C1 and C-1 so on let us try to do it for
Cnej2pf0nt+C-n e to the power, definitely it will be –n so –j2πf0nt okay,
let us try to see what this term together gives me, see if I can evaluate all
this terms I am almost coming to a conclusion that it is almost like this
C0 is nothing but a0 and this is actually almost like an cos and +bn sin if
I can finally represent them in that format, then I can give a relationship
between this Cn and an, bn right.

So let us try to see, now all we know is Cn is a complex conjugate of C-

n therefore I can immediately put our representation so I can say that
suppose Cn is represented as |Cn|ejθn then C-n must be represented as |
Cn|e-jθn right, this is something I can always do as long as I know this θn
and |Cn| and that I should be knowing because if I know the Cn I can
immediately calculate it is modulus and its phase right, so this is
something I can always do.

80
Now let us try to see if I put this replace this over here what do I get so
for that nth term I can get |Cn|ejθn ej2πfont+|Cn|e-jθn e-j2πf0nt okay, so this is
the nth term I am getting which is nothing but |Cn| gets common and I
get ejθn+2πf0nt+ the negative of that right, e-j same thing okay.

(Refer Slide Time: 24:22)

nth term = cn e jθne j2πf0 nt + cn e −jθne −j2πf0 nt

= cn 2 cos [θn + 2π f0nt]

an cos (2π f0nt) + bn sin (2π f0nt)

= r cos ϕ cos (2π f0nt) + r sin ϕ sin (2π f0t)

= r cos (2π f0nt − ϕ)

81
 r = 2 cn
⟹
ϕ = − θn

( an )
bn
tan−1 = − θn

an2 + bn2 = 2 cn
So if I just say it is ej some θ or let us π, +e-jØ put Euler’s theorem what
do we get, 2xcos of that cos of inside whatever Ø is there so I can write
this easily as 2cos[θn+2πfont] right, what I have eventually got is a
single cosinusoidal term right, I have got a single cosinusoidal term
nothing else this is something I am getting, okay. Similarly, for the
trigonometric series for every nth value I will be getting 1an
cos2πf0nt+bncos2πf0nt sorry sin, right I will be getting this.

Now this one what I can do an I can write as some rcosØ and bn I can
write as rsinØ this is something I can always write, right because
immediately I can see I can calculate r, r should be what root over
an2+bn2 and Ø should be if I just divide these two tan-1 bn/an such long
as I know this relationship I will be always able to put this representation
put this over here so I get rcosØ cos(2pif0nt)+rsinØ sin(2pif0nt) why I
am doing this I just want this representation again, right.

Can I see a single cos representation it is cosa cosb+sina sinb so that

must be cos(a-b) right, so I can write as cos[2pif0nt-Ø] fine. So now I
can see I have got two series one was exponential series another one was
trigonometric series from both the series I could actually get I was trying
to see the nth term and I could see that by unifying this cosiusoidal and
sinusoidal and I unifying the Cn and C-n I could get similar
representation, only thing is that those coefficient has to be now matched
right, because I will be getting exactly similar term for every n.

82
So therefore the coefficient because it is a unique signal so it must have
unique representation unique addition of sinusoidal okay, so therefore
coefficient must match so therefore r must be 2|Cn| right, and this Ø
must be minus of this θn right, so this is what we are getting. Now Ø is
tan-1bn/an that must be –θn so this is one relationship I get and r is
basically root over an2+bn2 that must be 2|Cn|.

So basically what is happening I have got a trigonometric series I have

got a exponential series, exponential series whatever amplitudes plot I
get I immediately can get a relationship between the corresponding
trigonometric series, okay. So if these two are equal the representation
should be equal and what has happen fundamentally as you can see that
because of this complex conjugate you see finally when I added those
two terms Cn and corresponding to C-n I have actually got a real signal.
So that means my signal was real people might be thinking that when I
do full exponential Fourier series I am represented them in terms of
complex signals, so what is happening to my real signal complex signals
are now creating real signal that is not the case the Cn’s are arranged in
such a way that if you take two nth terms they will those two complex
signal will actually give you back the real signal, okay. So eventually it
is nothing but both the representation tells you that it is a particular
representation where the, for a particular frequency there is an amplitude
and there is a phase nothing else both the representation gives you
similar value, okay.

So with this I will end this class the next class what we will try to do we
will try to evaluate what do we mean by this negative frequency.

83
Transcribers Name: Mayflower
Analog Communication
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur
Module-20
Lecture-07
Fourier Series (Contd.)

Okay, Now we will try to give some interpretation on this negative

frequency the concept of negative frequency. So let us first try to see,
whenever we are saying the concept of negative frequency, what exactly
we are getting? We are getting a means, as you might have seen, so we
are having this at positive frequency we are having this mod Cn.
(Refer Slide Time: 00:46)

84
 cn e j2πf0nt ⋅ e jθn

cn e j2πf0nte −jθn

e j(ω0nt + θn) = cos (ω0nt + θn) + j sin (ω0nt + θn)

And then some ej2πf0nt x ejθn right. And at the negative I have same Cn e-
j2πf0nt e-jθn this is something we have right, if you carefully see these

two things, those two are nothing but 2 phases okay. So if just ignore
this mod Cn, what do you have we have in this, this is the unit length
phasor, in the complex plane. So if have this is the real axes this is the
imaginary axes right, so this just a unit length, single length phasor
which is actually rotating as I means, see here the variable is t with
respect to t if I start increasing t, what is happening this overall angle, so
suppose of this one it is actually J 2πfont +θn.

So θn must be some constant okay which does not depend on t and after
that it is just with respect to this phasor this particular unit thing is
rotating, okay and then only we will be getting this phasor representation
because ej this one has a real part and a imaginary part whenever I put
ej2πf0 I can put as wO , w0nt + θn what does this means actually it is cos
w0nt + θn + j sin w0 nt+ θn okay. We all know that if with uniform
angular velocity w0 if we rotate a phasor, the projection to a axes along
which it has a pivot, along that axes that is what is happening the
projection actually follows a sinusoidal or co sinusoidal right, this is
something we already know.

So basically, what is happening whenever I am actually rotating this

with angular velocity w0, if I trace this projection with respect to time I,
will get at the real axes or co sinusoidal and at the imaginary axes or
sinusoidal this is actually the representation in complex to mine right, so
because, we have representing that as complex signal now what we also

85
know that eventually by our Fourier series representation, there was a
positive frequency and there was a negative frequency, how that
negative frequency looks like?

Negative frequency also is having just this representation minus some or

means minus of j 2πf0nt + θn okay. So it also has that concept of θn but
whenever we are giving the rotation it is actually if we have to trace
back it is rotating in the other direction. So this is rotating in the positive
direction whereas as we increase this θ, basically it will be if I see in the
co sinusoidal it is actually initially first it will be increasing, as I
basically increase this t whereas here if I increase t I will see that there
will be a reduction.

So basically, if it rotates on this side the other one is rotating on this side
okay. So these two phasor, just the positive one is representing a rotating
phasor which is rotating in anti-clock wise direction right, and the
negative phasor is something which is rotating in the clockwise direction
right. So the negative frequency actually means these two phasor and
means clock wise rotating phasor gives me the negative frequency.

And what is eventually happening if you see why the signal is becoming
real. I have this representation now if you are interested what you can
do, you can start putting this positive phasor and trace the signal
imaginary signal. And if you represent this rotating one which is clock
wise rotating and if you also trace this two signal, add these two up with
respect to time every time instants, you take those values and add them.
What eventually you have see that this thing will get cancelled out so in
the imaginary axes if you add these two you will get nothing it will be
always 0 ,for every value of t it will be 0, and in the real axes for every
value of t will be getting occurred, different co sinusoidal okay, the co
sinusoidal we were producing, this is exactly what is happening. So that
means my co sinusoidal I can actually have a equivalent representation,
it just mathematically equivalents a equivalent representation where I
can say that it is just a addition of two phasor, rotating clock wise and

86
counter clock wise and that counter clock wise is actually the positive
frequency and clock wise one is the negative frequency.

If I just take these two things and then plot the frequency domain
representation, that gives the due to exponential Fourier series whatever
representation we get so, and in that the frequency is just this, which is
in terms of this complex rotating phasor frequency or that phasor
rotation frequency, angular rotation okay that the phasor has okay and,
that is why because there are two counter rotating phasor, so I have
automatically on getting a concept of plus and minus.

A general sinusoidal does not have this direction, okay because general
sinusoidal is just represented with respect to w0, right we just write sin
some 2πf0nt so it is just having a frequency I do not have a concept of
negative frequency, let us whenever we start representing it in the
complex domain then we a representation because there is a direction of
rotation. The phasor generates sinusoidal but it actually generates due to
its rotation in which direction different, direction generated different
kind of complex, means components okay.

So that’s ,in a nutshell that is the means that is the realization of negative
frequency, why negative frequency comes and we have already seen that
these two are equivalent representation, here in phasor also you could
immediately realize that it is nothing but a real one if I take this counter
rotating phasors, the imaginary part gets cancelled out you can just trace
them in time, add them you will see that they get vanished.
They will just negate each other and you will see that it is getting
cancelled out whereas the co sinusoidal one, there will be a real part and
that is actually the signal that you are getting. So any real signal it must
have a equivalent phasor, rotating phasor representation and because the
rotating phasor representation has direction. So that is why we are
getting this positive and negative frequency concept so it is just a
concept as long as we are representing it in exponential.

87
So you might be asking why do I need then the exponential
representation, that might be a natural portion that we have a equivalent
trigonometry representation already the exponential one is actually
giving me that trigonometric representation on this whereas, the
exponential one is actually a complex one which is originally not
existent. However my, this one this one, the actual co sinusoidal
representation that is the real one those are the signal which exist, so
why we are actually looking for this exponential representation.

The reason is that in the trigonometric representation, you have a

sinusoidal as well as the co sinusoidal representation. okay which is
double representation and they have different, different coefficients.
okay, Instead of that if I do just exponential representation yes I get
complex number Cn but I get a unique thing, okay of course you know
that at every Cn has two components, the amplitude as well as spectrum
but sorry amplitude as well as the phase, but we know that if I just
interpret it, its just complex domain, it is just a unique representation
Cn, I do not need two An and Bn it just Cn okay.

So that is the utility of representing it in exponential Fourier series, and

you will later on see that exponential Fourier series also at zero, towards
realizing another thing where any signal can be represented that is call
the Fourier transform. Well next probably not in this class, we will try to
appreciate what happens when we try to represent any signal, not just
periodic signal with Fourier series we have already presented periodic
signal.
Later on, we will try to see that if we have any non periodic signal also
that can be represented and they are you will see, that exponential
Fourier series becomes extremely helpful, because that gets extended
towards the Fourier transform okay. So with this may be, we can give
two very important examples, so let us see the first example.

(Refer Slide Time: 11:56)

88
89
 1 π
T0 ∫−π
Dn = g(t)e −j2πf0nt dt

1 π/2 −j2πf nt
T0 ∫−π/2
= e 0 dt

1
[e − e j2πf0nπ/2]
−j2πf0 nπ/2
=
−T0 j2πf0
1
[e − e jnπ/2]
−jnπ/2
=
−T0 jn
−2j sin(nπ/2)
=
−T0 jn
sin(nπ/2) 1
= πn ⋅
2
2
1
= sinc(n /2) for n∈ℤ
2
So in the first example, what we are trying to do, we are actually just
whatever we have learned, we are trying to apply them okay, so we have
a signal this is something I have already means, given you now I just try
to solve that so this signal is something like this it has a period of 2π and
it is represented as this, this is π this is 2π ,so after every 2π it gets
repeated, same in the negative also. So at -2π, so basically this is
eventually –π/ 2, this is π/2 this amplitude value is one, so this is my g(t)
if you just take okay.

So all I want to do is, I wish to see how this signal can be decomposed
into known signals okay. So what we have to do, we have to evaluate

90
first let us say D0 okay, so D0 is nothing but 1/T0 integration now I have
to do it over a period that should be – π to + π, that can be one period it
is 2π, because this thing gets repeated at every 2π okay, so –π to + π I
have to put the signal itself okay, so which is wt x e-j2πf0nt.

So what is f0? f0 is 1/T0 and t0 is just 2π so 1/2π right, so this is what

we get. Now evaluate this integration, so all we have to do is evaluate
this, means actually Fourier series we are try to evaluate. So if you
evaluate this integration what we get see –π to π ,– π to –π/2 it is not, its
0 actually, so whatever happen is integration in that period will be 0 and
+ π/2 to π it is 0, so it is only define between – π/2 to + π/2 and the value
is 1 right, so it is just e-j. So just do the integration it should be t0 –
j2πf0n and e-j2πf0n.

So we have to put the limit so one will be π/2 – right, again – π/2 so – π/
2 become + okay, right so this is all that we get. Now you know that f0 is
1/2 π so this is actually becoming 2πf0 that becomes 1 so this entire part
you can put as one. So 1/T0 okay, Iam evaluating Dn actually right,
because at d0 it should be already n =0 so it becomes just 1 so if I just
evaluate d0, d0 it should be again 1/T0 integration from – π/2 to + π/2
only defined and wt becomes one that’s alright but this becomes
one ,already so it should be one dt.

So this becomes π divided by T0,T0 is 2 π so, this become ½ so, D0 is ½

already okay, this one so how much we get, it is e-jn π/2 - ejn π/2 right,
this is fine okay. So now all we have to, there is a minus, so this is just
again put Euler’s theorem, will be immediately seeing that the j part get
cancelled and this becomes -2cos right, sorry -2 ,no no the j part is
actually, because it is minus so the j part will be, there so it will be -2jsin
inside part right, so it becomes -2j sin n π/2 okay and then again you
have this becomes one t0 is 2π and so basically you have –j is already
there which gets cancelled t0 is 2π and n right, so 2 gets cancelled, so we
have sin n π/2/ n π so this our answer right.

91
Eventually if you plot it is actually almost like a sin x/ x kind of thing, so
I can put this as by 2 and I can multiply into ½, it is just half sin, sink
function or sin c function right. So what is happening?
(Refer Slide Time: 18:58)

If I just plot it, it is same function at D0 it is ½ and then it just goes to

again 0 at 2, goes to again 0 at 4 and so on right 6, so I just comes down
like this goes up something happens over here, that is the envelope of
course and at every value, this is 0, so this is the plot we get D0 and
corresponding Dn do you have a phase over here? You do not right,
because you have seen that all Dn it is no longer complex, so they are all
real values so basically the phase plot will be all 0, it does not have any
phase plot just amplitude plot, what eventually is happening these, these
sinusoidals are getting mixed.

92
The sinusoidal first at Dc term of ½ so basically, if you just take a signal
of strength half okay. next a sinusoidal of period one, sorry period that
2π okay and the amplitude is at this point, probably the amplitude will
be 1/ π, so with 1/ π and period that 2 π if you just put a sinusoidal, sorry
it should not be cos it should be see the phase part is not there it is real
part so it must be the cos part right so it is the cos part and the amplitude
is 1/ π.

So basically what is happening like this you keep the second one, which
is 2 x 2 π okay? That is not there that is not present, so that coefficient is
0, so keep putting all those sinusoidal you add them together you will
see that the same signal is being generated, as long as the coefficient are
right, if the coefficients are not right they will probably not reproduce
the same signal. So this is what exactly is happening so any signal if you
do Fourier analysis it will give corresponding sinusoidal okay nothing
else.

The next sinusoidal probably the next one will not be there, but the next
to next one that is the thrice the frequency, so it should be repeated here
the sinusoidal is repeating, within this period only once okay so it should
be repeating thrice, within this period with a amplitude of this much, if
just put that and you keep adding those all those sinusoidal up to infinity
you will see your signal is getting represented.
So this is what exactly happens in Fourier’s why we have done that? We
will see that later this is a typical example which will be required. But
next will be doing a particular means, analysis for a particular signal
which is immensely important you will see the sampling theorem is
based on this particular Fourier series representation, we will come back
to that but before doing that we need to understand something called the
Dirac Delta function.

(Refer Slide Time: 22:45)

93
δ(t) = 0 t ≠0
∞

∫−∞
δ(t)dt = 1
So we will try to first defined, what do we mean by Dirac delta function
so Dirac delta function by definition it is like this, is defined it was of
course, defined by Dirac and it is one its name, so delta d t is have a
value 0, as long as t is not equal to 0, that is the definition of a Dirac
delta function okay, it has some more definition we will come back to
that later okay. So the definition is that rather than 0, it is 0 everywhere it
is just at 0, it almost undefined it can take very huge value okay.

A equivalent representation of delta d t is something like this, it is the

box function okay defined from –€ / 2 to + €/ 2 and the strength of the
box is 1/€ okay, so what does this means? This will become delta or

94
this will approach to d Dirac delta, the way we have defined it if € tends
to 0, so what will happen? As € tends to 0, the separation will be, this is
the box function so it is 0 everywhere as e becomes 0 and 0 so all other
values it will have eventually 0 only at value 0 it has value 1/€, € tends
to 0 so this goes towards infinity right.

So it happens to take the position of delta dt okay. So this is one typical

example of delta function and whenever you give this definition it has a
added valuation, think about this thing, from this definition, if I integrate
d t from – infinity to + infinity according to our definition is a box
function so it should be area under this box all of the plus it is 0, so this
is this area is one so this must be one so that is the added thing added
definition of delta dt.

So delta dt must be defined with respect to this, and this one should be
added okay, which takes, almost takes delta dt towards that box function,
if these two definitions are the definitions of delta dt then, I can give a
definition of delta dt, in this manner, where I have to put additional
condition that € tends to 0. So this is in away our definition of delta dt.
Now let us try to see some property of this delta dt okay, suppose I have
function let us say xt.

(Refer Slide Time: 25:49)

95
 ∞ ϵ/2
1
∫−∞ ∫−ϵ/2 ϵ
x(t)δ(t) dt = x0 dt = x0

I multiply this function with delta dt okay, what do I get, so what is

happening? xt might be having any value think about that box definition
of delta dt, so at t = 0 ,it has around t = 0 it has this definition right –€/2
to €/2 and it is 1/€,right so now what will be the value of this, it will
actually this value can only be realize if I start first integrating it, okay
let us say I can if I can integrate this thing, why I am doing that because
delta dt integrated fashion it has some definition right. So if I integrate it,
now what is happening on top of this I am putting this definition that €
tens to 0, okay so if € is tending towards 0 so what do I get?
When I integrate as € become 0 I have a this, function is there this
function at x0, as some value okay the € is very small and infinite signal
is small, then x0 around that will almost not be varying, as long as it is a
band limited signal, that means it is not change arbitrary that this point

96
immediately, next point it can go to anywhere, as then it is a continuous,
differentiable, analytic function with all those nice property then I know
that very small € it cannot change, it cannot deviate.

So it will be as long as my € is very small within this –€2/2 to + €2/2 it

will be always taking value x0 okay. So what I can do? I can write that it
will be taking x0 value, so it should be so everywhere it is taking x0
value and this integration has to be done from –€/ 2 to +€ /2 okay, and
within this what is the value of value of delta t, that is 1/ €, so this
should be the integration x0 is now a constant because I have already
said that if e is sufficiently small x0 is not means xt, is not varying
around that value okay so x0 goes out and again I get one.
So what is eventually happening? Whenever I multiply a particular
function with delta dt and if I integrate it I will be getting that, particular
value back at that point. So the functional value at that point I will be
getting back this is called the sampling property of delta dt so, basically
whenever I multiply a function with delta dt and integrate it, it will just
sample out the value where delta d t is defined okay. So whenever I say
delta d t it is defined at 0 whenever is say delta dt – T it is defined at t =
because this has to be 0, it is defined at t = T.

So at T whatever the value of that function, it will just take out that
value, so this is why delta dt as a very nice property mathematically but
it picks up the sample value at that point wherever that delta dt is
defined. Rest of the values it will just make it 0 okay so this is very
important property of delta dt and this is some property which will be
heavily using, especially you will see the entire sampling theorem is
based on this, this particular understanding of delta dt okay.
So what will do in the next class that will try to exploit this particular
sampling condition and some other property of d to get a compulsive
understanding of a particular signal, so this is something we will do next
class.

97
Transcribers Name: Mayflower
Analog Communication
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur
Module-20
Lecture 08
Fourier Transform

Okay so welcome to analog communication course. so far we have done

some discussion about Fourier series and the representation of any signal
any periodic signal let us say with respect to some known basis periodic
signal like cosine sinusoidal ,sinusoidal or exponential sinusoidal so this
is something we have done we have also done a simple example in the
last class I just recapitulate that.

(Refer Slide Time: 0:54)

98
 ∞
1
Dne j2πf0 nt
∑
w(t) = f0 = T0 = 2π
n=−∞
2π

( 2 )
1 nπ
Dn = sin
n≠0 nπ
D0 = 1/2
So what we have done is we had a signal like this which has a period of
2π so define from –π to π and it is basically centered around 2π again
and so on so it just repeats after every 2π and the strength is 1so this is
our general signal which is w(t) okay so while representing this w(t) we
could represent w(t)as exponential Fourier series so n going on from
-infinity I am just recapitulate it the same result so it is Dn eJn2πf0t where
of course f0=1/2π which is meaning that T0 is 2π so that the period.

And we have also evaluated our DN so that happens to be whenever n is

not equal to 0 so Dn becomes 1/nπ sin nπ/2 so this was our Dn and D0
was just ½ so this is something we have already done so eventually what
we have got is something like this a spectra which was defined in
frequency domain so it was having value of 1/2 or 0.2 at 0 then at every
F0 2F0 3F0 and there are values so at F0 it is if you just put sorry at
means frequency a component 1 so that is n =1.

So there if you just put this you get 1/π and so on so at 2 so this was at 1
at 2 we got 0 and so on it was going into negative again 0 and so on and
symmetric so that is -1 -2 something like this so this is what we have got
if you see carefully that our this Dn term is now a real number all Dn
including D0 these are real numbers so initially what we have told that
Dn can be a complex number and the amplitude of that is plotted in
amplitude spectrum and the phase of that is plotted in phase spectrum.

Because the Dn are all real and so the Dn and D-n so basically what
happens the phase part is always 0 so the phase spectrum will nothing

99
but will be just 0 everywhere so it does not have any phase part right for
all the frequency component so for every other frequency component
starting from means the DC value which is at frequency 0 so it has a ½
so that means 0 you if you wish to plot this it should be something like
this all the basis function.

(Refer Slide Time: 04:13)

Dn ∠Dn
So D0 so that is in time if we just plot so that is actually half strength
DC that the strength is ½ times 0 –infinity to +infinity defined as this;
next is at frequency one Hertz so the it is a sinusoid or we should say co
sinusoidal with amplitude 1 /π so all we have to do is we have to draw a
cosine sinusoidal of frequency 1 that means the period is 1 so basically
so this part if this is 0 this must be 1 and so on it repeats after the radius
1.

100
So the period is 1 and what happens basically the strength should be 1/π
as we have got and remember because the phase is 0 so the cost actually
starts from here on the 0 phase co sinusoidal so always whenever you
are actually constructing these signals you have individual constituent of
this same signal so basically signal has been decomposed into multiple
either DC value and all other orthogonal co sinusoidal.

So we just try to plot those co sinusoidal now what is happening you

should also remember that this particular co sinusoidal if you see it from
here because it is a exponential Fourier series so this part is actually
making the eJ2π that f is now 1 okay so basically it is at 1 this is defined
so f will be 1over here so it is this right and you have a corresponding co
sinusoidal at the negative frequency so that also you will have to put so
and because the Dn are for both the cases 1 /π.

So it becomes 1 /π e-J 2πt and then if you just add these two what you get
is actually a cost with strength 1 /π okay so that means the frequency of
it is 1 and the strength is 1 /π so basically whenever I am plotting this co
sinusoidal you need to understand that we are taking a positive
frequency and the corresponding negative frequency taking both the
amplitude as well as the phase information.

So if you see over here carefully we were actually taking the amplitude
information and putting it over here because the amplitude is even
symmetric amplitude spectrum so we will have same value over here so
modulus of Dn will be put over here and the phase of Dn will be since
put over here + and - so that is the phase part because here the phase is 0
so we do not see the effect of phase.

But otherwise the phase effect should be there and the corresponding co
sinusoidal that will be happening, that will give you exactly the
amplitude as well as the phase information so cos will be either lag or
leading by that amount of phase so this is what you have to do whenever
you are trying to drew means draw the corresponding basis signals you

101
have to take from the amplitude signal the amplitude and you have to
take phase from the phase spectrum okay.

So these two information you have to take and you have to draw the
basis signal similarly the second one will also come over here if you see
the second one the second one has amplitude 0 at frequency 2 right so of
course the time period will be at1 / 2 so basically what will happen this
has a higher frequency so it will actually repeat within this okay so the
thing will be at a higher frequency but because the amplitude is 0 you do
not see this.

Because it has it’s a cosine sinusoidal where the amplitude is 0 so the

third constant is non-existent the 4th one if you see is that the amplitude
which is negative so accordingly cosine sinusoidal will have a negative
sin in front of that so the cosine sinusoidal will start at a πphase shift so
it will be looking like this frequency of that will be three times so you
start putting the three times frequency within this and it keep doing that
as you go along the frequency line.

Your frequencies will be doubled thrice or 4 times and accordingly the

repetition period will be synced and you will have to draw those things
with exact phase and amplitude once you draw this and if you add all of
them up to infinity you will see that exactly the signal that we have
targeted will be generated so this is this is the significance of it so
basically what we have done is taken a periodic signal we could
represent them with respect to corresponding co sinusoidal or DC value
and the specification that is required is completely contained in that Dn.

So Dn basically gives you two information one is the amplitude

information and other one is the phase information for every frequency
whenever you specify n that means it is n x f0 that fundamental
frequency so n x f0 at that frequency what is the corresponding
amplitude that will be the modulus of Dn and the phase of Dn will be the

102
corresponding phase so you get these two and then you plot it and
whenever you are plotting you are actually putting a co sinusoidal.

So you are taking two values from the spectrum as we have already
explained that it should be always because it is a real signal so it should
always take a exponential Fourier series one from positive and a
corresponding negative so these two together will generate your co
sinusoidal signal that is what is happening so and you will be getting as
long as you are having that complete amplitude spectrum and complete
phase spectrum for every frequency term you will be getting
corresponding amplitude.

And phase and accordingly you draw the sinusoidal and you add them
up if your phase and frequency sorry phase and amplitude information
are correct that means DN you have correctly evaluated you will always
if you sum all of them you will always get your signal back this is what
is happening that every single now is decomposed into all the harmonics
of it and every harmonic is defined for this particular targeted signal is
defined by its specific amplitude and specific phase okay.

So far so good we know now what is the utility of Fourier series and
how Fourier series actually represent a signal in a different domain so it
is actually we are representing the signal in frequency domain so we are
almost seeing the frequency component or the sinusoidal component of
that signal which constituent this particular signal okay so this is all
good now we have talked about a measurement of a signal right initially
in the first few classes so we have talked about energy of a signal.

(Refer Slide Time: 11:50)

103
 b

∫a
Eg = g 2(t)dt

∫a
Eg = | g(b) |2 dt

∫0
Eg = g(t)g*(t)dt
So if the signal was real we have told that the signal has to be squared
and integrated over the time that it is defined so let us say it is defined
from A to B so as long as signal g(t)is a real signal this is what we will
have to do and we get the corresponding energy of the signal this is
something we have defined that squaring the signal and then integrating

104
over the time that it is defined now if the signal is not real that it is a
complex signal.

Then we have to specify another term that means you have to take the
modulus of that signal and square it so basically you have to do this mod
g(t) square dt and you have to integrate so same thing you have to do for
calculating evaluating the energy because it is a complex signal and
energy is a real number so we need a real number so we have to do or
we have to take the signal we have to do a complex conjugate of that.
This is all something we have defined by definition this must be energy
right.

So this is something we are told already that is a measurement of signal

like we have done that vector analogies ,in vector it is the distance of the
vector or the strength of the vector which is also means measured by
distance square right so this is how it is being measured now if we just
say for any signal we want to basically get the energy value of that
signal right it is very easy take that signal integrate it over the time that it
is defined.

So suppose we have that square pulse right so remember it is a periodic

signal so definition of energy we have already talked about periodic
signal where it becomes a power signal right so the definition of energy
is not there but if we just take the definition from -π toπthen it is a
energy signal and then within this we can still define the energy of this
signal so if you wish to calculate the energy what you can do you can
just integrate from - π to π G 2 right.

If this is my signal you can get it. Is there any other way ? or in the
frequency domain can we also talk about this energy representation or
can we just say that this particular signal we have already identified that
it is means defined by multiple basis signals and their linear combination
and the coefficient of them are already known via Fourier series right so

105
if we just take this signal and take those constituent is there a possibility
of defining the energy from that perspective.

So in the frequency perspective every component how much energy it

actually provides to add up to give this energy is it possible so that
something will be now exploring so let us try to see what happens to
this.

(Refer Slide Time: 14:53)

106
 x(t), y(t)
z(t) = x(t) + y(t)
b

∫a
2
Ex = x(t) dt

∫0
Ey = | y(t) |2 dt

∫a
Ez = | z(t) |2 dt

∫a
= | x(t) + y(t) |2 dt

∫a { + x(t)y*(t) + x*(t)y(t)} dt
2 2
= | x(t) | + | y(t) |

= Ex + Ey

So before that let us say I have two signal x(t) and y(t) okay so my
definition these two signals are orthogonal to each other so I am just
trying to exploit that orthogonal property and energy calculation okay so
we are just assuming that two signals very simple two signals which are
orthogonal to each other it might be just like coswct and sinwct we have
already proven these two signal are orthogonal within a particular period
right.

So if it is defined from- π to π okay so take these to signal and let us say

I wish to generate another signal z(t) which is just an addition of these
two signals now my target is so I am almost going towards linear
combination of orthogonal signals so right now I am just giving a very
simple example that I have got two orthogonal signal and just adding
them it is a linear combination with coefficient 1 +1 so I am just adding

107
these two signal now we wish to see suppose for x(t) and y(t) I have
already evaluated the energy of them.

So this is suppose Ex and this is Ey I have evaluated the energy of that

so how I have evaluated x & y if means I do not care if they are complex
or if they are real so I just have done this x(t) modular square integration
dt over the period that it is defined okay so a to b so for our case if x(t) is
coswct so -π to +π okay so this is something I have evaluated and this
gives me EX similarly for y also I have done the same thing so I have
done a to b y(t) modular square dt that gives me Ey.

Suppose I know these two parameter Ex and Ey both of them can I now
calculate the energy of z(t) without actually means adding them
integrating and all those things so does orthogonality gives me some
idea of these things so let us try to see so energy of z(t) what does that
mean a to b modular z(t)2dt right so that actually means AB now it
should be x(t)+y(t) modulus whole square dt right now expand this so
that should be mod x(t) whole square plus mod y(t) whole square plus
x(t)y*(t)+x*(t)y(t) right.

This whole thing integration over dt right so now you can see this
particular part mod x(t) whole square dt that is actually Ex for us this
part is the Ey for us now the orthogonality plays a big role what is the
integration under t over that period as long as x(t) and y(t) are
orthogonal signal we know that that term should be 0 that is the property
of orthogonal any orthogonal signal if you take that signal and take the
complex conjugate of that multiply that and integrate over the period
that you are targeting or where the signal is defined.

It will be always as long as x(t) y(t) are orthogonal that should be always
0 so this term happens to be 0 this term happens to be 0 so basically
what we get is the energy of z(Ez) is just the summation of energy of x
and y is a very fundamental results which has big implications okay so
we will try to use this and try to prove a very fundamental theorem

108
which is called Parseval’s theorem. So let us try to see if now from this
if you just go to a linear combination of orthogonal signal.

(Refer Slide Time: 18:59)

∑
x(t) = ci xi(t)
i

∫
c1x1(t)c* x*(t)dt
1 1

∫
2
= c1 x(t)x*
2
(t)dt
2
= c1 E1
2 2
Ex = c1 E1 + c2 E2 + ⋯

109
So suppose I have a signal let us say XT that is defined as some linear
combination with coefficients Ci Xi(t) where all the xi are orthogonal to
each other okay so that means that if you take cross product so x1 with
x2T star and you integrate over the particular targeted period you will get
0 okay so these are all xi(t) are all individual orthogonal signal and as I
goes from in to its limit as many I ‘s are there okay.

So over i you have to solve ok so if this happens what we can get if you
see so the energy of x(t) if you just try to calculate this that must be we
know that it is a just take the previous example if you have two
orthogonal signals to each other it will be individual energies square or
individual energy addition of that right so if I try to calculate the
individual energy the individual constituent is suppose let us say I start
from 1 so or 0 so let us say i start from 1.

So c1x1(t) what is the energy if we know that x1(t) has an energy of E1

okay already therefore that c1X1t will have an energy of you have to
integrate modulus of this so C1 into C1* so that should be mod C1
square that will go out of the integration so if I just do that C1 X1(t) C
1* X1*(t) dt integration so C1C1* will be mod C1 square this
integration X1(t)X1*(t) dt that is the energy of x1 which is E1 so it
should be mod C1 square E1 so that is the first term.

So it should be the first term should give me the first energy component
and similarly all other energy component so there will be C2 mod square
E2+… up to as many i’ s are there all cross components because the
signals are orthogonal to each other will be just vanished so I can always
evaluate as long as I can represent a signal with respect to the constituent
orthogonal signal and they are in linear combination.
If I know all those coefficients I can always evaluate the energy of the
original signal. This is what was targeted and I can do that now let us
apply this to Fourier series let us see what happens.

(Refer Slide Time: 21:48)

110
 ∞
Dne j2πnf0t
∑
x(t) =
n=−∞
1 π j2πnf t −j2πnf t
2π ∫−π
e 0 e 0 dt = 1

2 2 2
Ex = D0 + D1 + D−1 +…
∞
2 2
∑
= D0 +2 Dn
n=1
So in Fourier series what has happened any signal x(t)a periodic signal
of course that has been represented as n=-infinity to +infinity we are
taking the exponential Fourier series formula so that should be Dn e- it
was sorry+ J2π and f0t right so that is the representation we have already

111
seen that and Dn must be evaluated as done by Fourier series right so
this is something we have already seen now what we are saying is this
that here as long as we know Dn we know that for every n these are
those orthogonal components right.

So therefore the energy of X must be the individual energy where the

coefficients are it is a linear combination of these orthogonal signal right
each of those orthogonal signal what is their corresponding energy so if I
just take e J2πnf0t and I wish to evaluate its energy so what I have to do I
will take this signal I have to do a complex conjugate 2π n f0t dt
integrate over a period let us say -π to π if I do not do this will give me 1
so I get 2π right so this is what I get immediately.

So I know that if this is a particular signal and I wish to evaluate its

energy so I can always evaluate it like this and of course whenever we
are evaluating this energy it also has 1/t0 if you remember the signal
means the way the signal has been evaluated so the corresponding Dn
term will always have 1/t0 so that 1/ t0 will come so 1/ 2π will come this
will give me 1 so that will be always giving me 1 because this 1 /t0 or 1/
t0 means 1/2π that will be there okay.

So I have to evaluate this way the energy 1 / 2π -π to π you have to do

this right and I get individual energy of those signals as one so this is
something I get so once I have got this then I know that the overall
energy should be just individuals suppose the 0th term is t0 so mod D0
square + then I have to get D1D -1 everybodies mod’s square, so D1
square + D -1 square +… now I also know that of course remember that
this coefficient square into the energy of the first one okay.

Now the energy of the first one is for any value of n that is 1 so that is
why all those values are coming to be 1 right so what we get what all
energy is just the summation of this coefficients so I can either write
them as n =-infinity to +infinity mod Dn square or I can because I know
these two are symmetric so I can write it as mod D0 square+ 2n going

112
from 1 to infinity mod n square I can also write this way this is the
famous Parseval’s theorem of computation of energy of a particular
signal.

So basically what has happened if I had a signal x(t)I was aware of its
energy so that I can do in a very simple way that okay take the signal
integrate it from within a period so let us say –π to π I can do that that is
one way of doing it but because I know the Parseval’s theorem what I
can do if I know the Fourier series representation of this signal that just
means that all that Dns I know I have evaluated all those Dn and so once
I know all those Dn the energy of the signal is also can be written or can
also be represented in a different way.

That means all constituent signals all constituent means orthogonal

signal you take their coefficient and take a modulus of that next square
add all of them you will get the energy of the signal so basically it is just
very simply understandable that a signal if it is already there you have
decomposed into multiple basis signals and each of those basis signal
has their own energy okay so the energy is dependent on the
corresponding coefficients that is the n.
So as long as you know that Dn you can take modular square you get
your energy representation of that particular constituent signal and we
call the signal is linear combination you can just take linear combination
of this energy or addition of this energy I should not say linear
combination addition of these energies that will create the overall energy
what exactly is this if you just take it from vector analogy so we have
already done a vector analogy to signal so what is happening if you have
in vector we already know.

(Refer Slide Time: 27:41)

113
That if a particular vector is represented by linear combination of two
vectors then what should be the strength of this vector that must be
through Pythagoras theorem we know because these two are orthogonal
so that must be this square plus this square right and this square is the
actually strength or measurement of that vector so measurement of this
vector plus measurement of this vector must be the measurement of this
vector as long as the measurement is distance square okay.

114
So if that is the case and if we are talking about Euclidean space we can
always see vector measurement which is distance square is just the
vector measurement of the corresponding component in means when
represented in orthogonal vector space right same thing is happening
over here if you see what is happening this Dn’s so whatever D and we
are putting over here this D n’s are actually the measurement or the
coefficients of the corresponding orthogonal signals that constituent the
signal.

So as long as we know this D n’s we can get corresponding energy

which is actually again the measurement of that particular signal or the
energy of that particular signal so measurement means energy over here
so energy of that particular signal so we get energy of all constituent
signals and it is just the addition of those energy which gives me the
overall energy of the signal that's the strength of Parseval’s Theorem.

So basically Parseval’s theorem state almost similar thing as we have got

in vector so all the cross terms are vanished because they are orthogonal
similar thing happening like vectors and only the corresponding
measurement add up to give me the measurement of the signal and our
measurement is energy so in the next class what we'll try to do after
getting this sense of signal we'll try to think about a signal which is
continuous in time that true, means in a way but most importantly if it is
a non periodic signal or energy signals. what to do what kind of
treatment we should take for those kind of signals will concern means
will concern ourselves about that in the next class.

115
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

By
Prof. Goutam Das
G S Sanyal School Of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 09: Fourier Transform (Contd).

Okay so as, we have discussed in the previous class that will be now
concerned about a signal, which is no periodic in nature.

(Refer Slide Time: 00:28)

116
 lim gT0(t) = g(t)
T0→∞

Most of the signals if you see we have defined periodic signals and the
fundamental property of a periodic signal is that it must be defined from
-∞ to ∞, that means that periodic signal must repeat at some interval, let
us say T0 which is the fundamental interval and it should go from -∞
+∞, otherwise that periodicity will not hold good. Okay so but most of
our signal actually start at a particular instance whatever practical signal
we can see suppose voice signal generated by me.

It will start at a finer particular time instance and end at the particular
time instance, they are finite in time they cannot start actually at +∞ and
-∞ and cannot go extend up to +∞ so most of our real life signals are
time bounded, okay so therefore, they have a finite energy because if
you go from -∞ to +∞ you will probably means if you wish to evaluate
power you will see probably 0 okay.

Because the signals are not stretched up to -∞ and +∞ so .our signals are
generally energy signal that means it is time bounded it starts at a
particular time and ends at a particular time. So we should have a strong
analysis of those kind of signal otherwise our analysis is not true so
whatever Fourier series gives us that is a very strong tool, we can agree
on that but still that particular thing is not good for real signal.

So we need to have good understanding about the real signal which are
finite in time okay, so let us try and take an example that is a very simple
example signal which is define as GT it starts at some value of t, let us
say a it ends at some value of t(b) so this is the time, So if this is my
present let us say 0, time zero. So this was my past at which time some
value a which is negative in time it has started and it will end at b. so
that is the future of the signal.

So this is the signal so its starts at a it ends at b, if this is a signal and I

wish to still do similar analysis as Fourier has done in his Fourier series

117
analysis, right so I still want to see what are the frequency component it
has how to define them can we get some spectrum out of it all those
things, okay but the problem is whatever Fourier has done that was
good for periodic signals, so what we will try to do or what Fourier has
also done that for a periodic signal only, we have some strong result.

So can we apply those results into this, so for that we have to forcefully
make the signal periodic so let us try to do that, let us say beyond this we
define some amount of time that is our period okay and the same signal
is repeated at every t0 so basically what happens if this is T0 my signal
was like this after another T0 there is another period of T0 where the
same signal is repeated and they are also repeats and it gets stretched to
-∞ to +∞.

So basically from our signal which was defined from a to b we have

created a period which actually takes into account this a and b, inside
that period and then we start repeating this signal whereas in the original
signal there was nothing beyond a and b or beyond even this T0 but we
have deliberately made this signal look like a periodic signal, very nice
so this is a newly constructed signal from that original signal we call this
as g T0 (t).

So that means from this signal we have transformed into a periodic

signal, right the good part is by this by means of this clever construction
we can always say if I take this limit T0-->∞ g T0 t that exactly goes to
g(t), this is what we were targeting so we were actually trying to see a
particular signal, can that be represented as a special case of periodic
signal okay, so we have constructed that particular signal and then we
are told okay.

Now if we stretch this period to ∞ okay or this t0 stretching from -∞ to

+∞ then definitely these parts will not appear so they will disappear,
because the period gets stretched up to -∞ to +∞ so that will exactly look
like our g(t) so this is what we are now targeting so what you are
targeting is deliberately we are making a periodic signal out of non
periodic or a energy signal which is time bounded so this something you

118
have done now for gT0 (t) which is a speedy periodic signal I can do
Fourier series analysis.

(Refer Slide Time: 05:54)

119
 ∞
1
Dne j2πfnt
∑
gT0(t) = f0 =
n=−∞
T0
1 T0 /2
T0 ∫−T0 /2
Dn = g(t)e −j2πf0 nt dt

∫−∞
G( f ) = g(t)e −j2πft dt

1 T0 /2
T0 ∫−T0 /2
Dn = g0(t)e −j2πf0 nt dt T0 → ∞

1 ∞
T0 ∫−∞
Dn = g(t)e −j2πnf0t dt

1
Dn = G(nf0)
T0

So for GT0 T which is a periodic signal with period T0, I can always do a
Fourier series analysis for n-∞ to +∞Dn ejn2π F0T I can write this way
where this F0=1 / T0 the period I have constructed the frequency
fundamental frequency is 1 over that, ok so this representation is true I
can always do a representation like this.

And after that we have evaluate Dn, so Dn is something we already

know it should be 1/T0 and integrate it from -t0/2 to + t0/2 over the
period gT0(t) which is the main signal and it will power -jn2π and
f0TDT so this is something we already know right, very good okay, now
let us see what happens to our representation okay so in our
representation let us try to do it this way, let us define it just arbitrarily
defining for our you will see why we are defining this.

So let us define a G(f) which is nothing but integration from -∞ to +∞

g(t) original signal it is the power -j2π F0Tdt is just defining this is our

120
by definition we are saying that because this, sorry it should be on the F
because this is our integration over t so that t will vanish it will be a
finally a function of F we are defining that as a G F function of small f
right so this definition is correct no problem in that as long as this
integration can be done.

So if we can if you say that this integration is possible we can always

say there will be this will be a function of f, because T goes away right
so we can write that function as some capital G(f) as some function,
okay so this destination is correct when this definition is correct what I
can write about DN so let us see what is DN, so now this DN if you
carefully see this DN has1 /t0 that is alright and in place of G(F) if I just
write f0, okay so if I just write f0 then what will happen I will be getting
this GT integration - t0/2 to + t0 /2 and I will have e- jn2π similar things
are there just in place of f, I have a f0, right so basically what I can say
this particular part if I put t0 tends to infinity okay.

Immediately this limit goes from -∞ to +∞ and as long as I say t0 tends

to infinity my Gnt if you see the previous definition for GT0 t if I put
T0 🡪∞ GT0t I can write as Gt so I can write Gt and replace that as
GT0t, okay so I can if I say that whenever I will put this particular limit I
will be always getting this representation, as we have G in place of f. I
will be putting f0 so this is quite obvious as you can see what is G of 0,
G of 0 will be just replace over here GF 0 so that should be GT ej2πF0T dt
okay.

Only problem is we have a GT over here so GT as long as T0 is tending

towards ∞ which is this case GT and GT0t are similar so I can GT I can
replace by GT0 T now if you go over here now this to limit T0🡪 ∞
means -T0/2 and + T t0/2 goes to - ∞ and + ∞, so I can replace these two
limit as if going to - ∞ to +∞ so therefore Dn exactly mimics these
things okay, so that is the duty I could do by this definition and by our
definition of GT and GT0T or the relationship between these two okay.

121
(Refer Slide Time: 11:04)

122
 ∞
1
G (nf0) e j2πf0 nt
∑ T
gT0(t) =
n=−∞ 0

( T0 )
1
T0 → ∞ ⇒ f0 → 0 ∵ f0 =

Δf → 0
∞
ΔfG (nΔf0) e j(nΔ f )2πt
∑
gT0(t) =
n=−∞

So once I could do this I have got a very nice construction so what is

that construction now I can see that what happens to my Dn, Dn
becomes 1/T0which I have already written this is my Dn and
immediately I can represent my GT0(t) which is the Fourier series
expansion so which is ∑n= -∞+ ∞ Dn ejn2πf0T and in place of Dn, if I can
replace this so it happens to be n=-∞ + ∞ 1/T0GF0 ejn2πF0t right, this is
what I get for this particular part okay, so far it is all good.

Now let us try to see what that this particular thing means, so what is
happening in this case we have already assumed otherwise you could not
have constructed this DN=T0🡪 ∞ so, whenever T0 goes to ∞ I have a
relationship F0 =1/T0 so what happens to F0, F0🡪 0 okay so this 1/T0 or
F0, I can that becomes infinitesimally small okay so that becomes
infinitesimally small instead of writing it as 1/T0 I can now write as ΔF
where this so, I am just replacing F0=ΔF which tends to 0 it is just for
my convenience.

Because we are used to see ΔF and this is GF0, or sorry let us see there is
things that we have missed here so it is n f0 which makes the f so it
should be written as nf0 right so that that was a mistake, I did okay so
because it is if you see the construction GF is written as Gt e j2πFT now

123
here in the construction we have e -j2πT all of this and n f0 is there so
therefore it must be nf0 right that's all right so this should be all replaced
by nf0 so ΔF nf0 must be replaced as nΔF right because f0 is replaced
now by ΔF.

(Refer Slide Time: 14:09)

∫−∞
g(t) = gT0(t) = G( f )e j2πft df nΔf → f

So GT0t newly constructed GT0 t becomes ∑n =-∞ to + ∞ 1/T0 becomes

Δ F G n f0 becomes ΔF x ejnf0 becomes ΔF2πT right so this is the new
construction that we have got for GT0t and we know here T0🡪 ∞ and ΔF
which is 1/T0 goes to 0 okay, so this is something we already know now

124
let us try to see what exactly is happening what is this part, this was
some function GF okay, now whenever we start varying this n=- ∞ + ∞
so what is happening whenever n = 0.

So it gives G0, n =1 it gives at ΔF Whatever value of GF so this must be

G on the ΔF so this is G0, at n = 2 so this is a n=1, n=2 it gives me G2 Δ s
and soon and also so on in this direction so basically it is almost like
there is a function G F which is a continuous function and whenever we
are putting this different values of n we are actually, sampling this
function at Δ F interval now as long as this Δ F is becoming
infinitesimally small.

So what is happening all these samples are coming closer and closer
together okay and in the same method what is happening to this
particular value so this is nothing but suppose I have a GF let us say this
GF looks like this. I take suppose n =0 I take this value and I take this
box so what will be the area under this box that should be Δ F x G 0 or 0
x Δ F, the next part will be again Δ F and what is this value this value is
G1x Δ F x Δ F.

So it is almost like this small boxes, I am actually taking the area under
that okay, and I am adding all this one, once I do ∑ I am actually adding
all these things right, so as long as my Δ F 🡪0 what is happening this ∑
almost becomes the integration because integration also gives me area
under a particular function so if I just now think that my overall function
is this particular thing that is gf x e j2πFT where f is replaced by n Δ F.

So basically what is happening this particular function, I am adding with

or I am multiplying with ΔF and I am summing it up, whenever we put F
as discrete values so it is nothing but this ∑ becomes the integration and
my G t0 T just happens to be our integration and because it goes from- ∞
+ ∞ whenever we put n = - ∞ what happens this n x F 0 goes to -∞ so, if
my variable is f that goes to - ∞ and whenever I put n = + ∞ it goes to+
∞ this Δ F becomes DF and I have this gf it will be e to the power j2π
this is replaced by FT.

125
So basically n Δ F is almost becoming a continuous variable F, this is
very true we have already stated that if Δ F is small, then all the values I
am taking almost becoming continuous in time, sorry in frequency so
basically I was taking initially I had this function which is gf x this, I
was sampling that function and I was trying to calculate the area under it
because I was multiplying by Δ F and taking that box and I was adding
all those things.

So this is nothing but integration that we have already understood, and

what is also happening I am making this Δ F smaller and smaller, so
those samples are coming closer together and getting very closely
packed and actually those samples are almost indicating now a
continuous variable on F, so whenever I write n ΔF that goes to F, okay
whether f varies that becomes the integration variable, so immediately I
get this and because my T0 🡪 ∞ already for this construction.

So this can be written as Gt because GT and Gt0 T becomes equivalent

as long as t0 goes to ∞, that was our definition of the signal, so we made
this a periodic signal to be periodic with period t0 and then we have told
that this t0 if we stretch to∞, it becomes a similar a periodic signal this is
something we have already stated so as long as T0 goes to∞ we get this
very relationship so what has happened.

(Refer Slide Time: 20:05)

126
 ∞

∫−∞
G( f ) = g(t)e −j2πft dt
∞

∫−∞
g(t) = G( f )e j2πft df

1
Dn = G(nΔf )
T0
Now we have a two relationship that we have said, one was our original
statement that gf must be related as this was, by definition GT it e to the
power -j2pi FC DT and now from the construction we have got another
relationship where we say GT is nothing but - ∞to + ∞ gf e to the power
+ j2πft DF, so this is the famous Fourier transform and Fourier inverse
transform theorem, you can immediately identify okay.

127
So if I have a signal GT I can always get a corresponding Fourier
transform which is Gf and if I have a signal G means if I have Fourier
transform I can always do a inverse transform to get my signal back so
this is the Fourier transform and inverse transform in this whole process
the most important part is this Gf, which is intermediately constructed so
what is this Gf, if you carefully see the construction this Gf okay is
actually almost similar to as you have seen it is almost similar to our
DN.

We have derived this earlier, here so basically we have already stated

that DN must be 1/T0 x G and LF, what was our DN, DN was for every
frequency component the corresponding amplitude and phase, DN was
giving me that now, G if as long as Δ F tends to 0 this becomes a
continuous variable ok, so this n Δ F as long as Δ F is very small, this
almost becomes a continuous variable of F so this is nothing but our T0
x DN or I can write this DN equal to this.

So basically what is happening the corresponding DN can be calculated

if I know that gf whatever that GF is and I multiply that with that very
infinitesimally small frequency component Δ F, so what is happening
suppose, I have a particular let us say of Fourier transform
representation gf okay, so each one of this for a particular value of n are
actually representing suppose this is my gf right so that sample value, if
I multiply that with Δ F so this area okay or let us say this area.

So GF x that Δ F this area as if giving me this DN, at that frequency

what is the overall coefficient of that frequency component, so what is
eventually happening as I put this Δ F tends to zero very small and small
I am actually getting all frequency component so it is almost becoming
continuous, earlier in my periodic signal I was only getting those
harmonics of f0, 2f0 and all those things now for a periodic signal what
we have observed that because for that aperiodic signal we have
represented it with respect to a periodic signal.

So if we faithfully represent back the periodic signal t0 has to be ∞, once

t0 is going towards ∞ this Δ F is becoming very small, so that means it is

128
now almost those frequency components are getting closely packed so
we are getting almost all the components all the frequency component in
the continuous domain right and everywhere if I multiply by this Δ F,
that gf whatever I have if I multiply by this Δ F, I get the strength of the
frequency component similar like this Dn right.

So what might happen suppose in a hypothetical experiment, I keep on

because I have to go means stretch my t0 towards ∞, I keep on reducing
my sorry increasing by t0 so earlier suppose this was my representation
okay, so this was my Δ s now what I do I actually increase my t0 so
immediately, what will happen if I increase my t0 so immediately this
will suppose I make t0 double so if I make t0 double immediately, what
will happen this DN value because t0 is doubled DN value will be
reduced by half okay.

So what happens the Dn value will be reduced by half but the samples
will be now more, so basically instead of this it will be all reduced by
half but the samples will be closely packed, double close the pack
because Δ F is now becoming half of the previous this one, so there will
be double closely packed and the sample values which is represented as
DN which is the integration or we should say which is the area
multiplied by Δ F G n Δ F, ok so that area is represented by these values.

So they will be becoming 1/2 because I have increased my T 0 to double

corresponding by f0 has become half, so basically sample will be there
will be double samples now more closely packed but this strength will
be reduced, and if I keep on doing this experiment what will happen I
will still get this samples or their values as long as T0 goes to ∞, their
relate means individual values will go to 0 so basically what I am
getting, I am getting at every point in continuous frequency domain.

I have some value but the corresponding value is 0, because T 0 goes to

∞, ok but the relative values are still being constructed by this G(f), so
however small you make them they have a relative nature so if this is the
strongest one and this is little weaker they will still remain the same
their individual value will tends to 0, ok but their relative nature which is

129
captured by this GF will still be the same, and that characterizes the
corresponding signals, the relative nature but individual values at a
particular frequency is no longer existent is a very typical understand.

There is a typical understanding behind this, so probably we need to

have some clarification one very simple example which is often used in
probability theory, probably will have better understanding.

(Refer Slide Time: 27:36)

90∘ 1
p= =
360∘ 4
So suppose I have a pivoted rod and I give a random force to this rod
and this rod can freely rotate around its axis, and depending on that
random force it will go and stop somewhere okay let us say where it

130
 stops so it says it stops over here, that angle is called θ now this theta
becomes the random variable, okay now if I just ask you that can you
tell me the probability that it will stop at angle 30 º, now how many
possible angles are there suppose the force that I will be exerting are
almost, means the way it is being exerted that whichever angle it will
stop.

They are all equally likely so it has an equal probability of stopping, of

stopping to any angle right so if that is the case every angle is equal
probable so they are equally favorable so now I am targeting a particular
angle30 degree, okay so how many angles in between are there are
infinite number of angles because 30.1 is the valid angle, 30.01 is a valid
annual 001 is validated so there are infinite number of angles so the
favorable case for my that30º is only one that it should stop at exactly
30º 30.0000 up to ∞ okay.

But there are infinite number of possible cases so my probability that

frequency definition of probability will give me just one upon infinite,
that is zero so stopping at a particular angle is 0okay so it doesn't have
any probability of stopping at a particular angle but if I now specify a
angle range let us say between 0 to 90º it will stop, because things are
equally likely now let us try to see, it has four quadrant so it has almost
360 degree out of that I want to cover this 90degree right, so that must
be my probability which is one by four, so whenever now I am
specifying a range, I have a valid probability value whereas if I specify a
single value I do not have anything no probability values are recorded
same thing happens to this gf.

(Refer Slide Time: 30:05)

131
 f2

∑ ∫f
ΔfG(nΔf ) ⇒
1

Whenever we are talking about Gf,I plot a Gf individual value, I do not

get anything but if I specify a range from frequency f1 to f2, it has some
value because it will be the integration over that and I will get some
value, any function if DN is the way DN is defined if you see DN is the
strength at a particular frequency okay, so if I say from this frequency to
that frequency in Fourier series term I will be just summing all DN okay.

So from that particular frequency to another frequency, I will be

summing all this DN whereas in this particular representation, if I have
to sum DN what I have to do the ∑ becomes integration we have already
seen, that DN was nothing but that Δ F x so its area actually Δ F x G
whatever that n Δ F right, so basically n and then summing over that n
value whichever in value we are targeting so f1 to f2 will have some n
value to some n value, if I am doing this it is just area and we are
summing that it just becomes in a continuous variable, that becomes the
integration from f1 to f2 I am integrating it.

132
So basically what is happening whenever we are saying that an
individual value that does not exist because it was Δ F 🡪0 this area tends
to zero so I do not get a value but if I start integrating from a particular
frequency range to another frequency range, I get some value okay so
that is the beauty of it, in the next class we will again what is the
implication of this thing.

133
 NPTEL

NPTEL ONLINECERTIFICATION COURSE

Course
On Analog communication

by
Prof.Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology kharagpur

Lecture 10: Fourier Transform (Contd.-2)

Okay so far we have discussed about Fourier transform so you got some
idea about the Fourier transform pair how they are being generated and
we have started giving some physical interpretation into this particular
after doing Fourier transform.

(Refer Slide Time: 00:39)

134
 G( f )
Dn = ΔfG(nΔf ) Δf → 0

Whatever we get the Gf) so what we have seen that by analogy because
the derivation was done through Fourier series so we could see that there
was a relationship that Dn was nothing but that Δ f into G and Δ f right
so this is something we have got that. D1 was a very important
parameter in Fourier series so it was actually giving the spectral
amplitude as well as phase of every frequency component now we have

135
seen because the Δf are becoming very small so we actually start getting
frequency component at almost every frequency.

So it is becoming continuous in frequency domain earlier it was discrete

it was at first harmonics and second harmonics and third harmonics so
what is now happening we are almost getting every harmonics but
because the Δ f is very small, so we are almost getting every frequency
component which we have already, so if I just see so what was
happening as Δ f was tending towards 0 so we were getting very closely
packed values okay, so almost continuous values.

So earlier, whenever we have plotted spectrum it was like this, so there

was some strength at different frequency f 0, 2fo and soon and the
corresponding values was even at the end here what is happening that
Dn is just not this G it is actually G into Δ f so it is almost like this box
integration okay, and on top of that as Δ l f goes to 0 these are
becoming closely packed so once they are becoming closely packed
basically I am getting a continuous spectrum and the corresponding Dn
are getting vanished okay because it is multiplied by Δf s as Δ f becomes
smaller and smaller corresponding things are becoming vanished but the
Dn might vanish, that the G still has a relative strength.

136
So what it says almost like if G multiplied by the frequency I get the
corresponding Dn okay, and we have already seen that integration gives
me some value so basically this G by itself is not the spectral component
it is defined as per unit spectral because the Δ f has to be taken out of it
so it is the overall integration that gives me the spectral component let us
say the amplitude that is given by if I just multiplied by Δ f okay so if I
just divide this Dn by Δ f, then I get this G value right.

So basically it is this G is nothing but the spectrum component par unit

frequency okay so whatever that Δ f is, per unit frequency in a very
closely packed frequency domain representation I get that at
instantaneous frequency what is the per unit frequency spectrum
component okay or the spectrum strength. so that is why whenever we
talk about G it is not the spectra itself it is actually the spectral density.

Earlier, when we are plotting Dn we are saying that frequency

component is having this much amplitude spectrum or this is that
particular frequency component is having whatever modulus Dn is this
much amplitude. now we are not saying that, we are saying per unit
spectrum but because it has to be divided by Δ f this is the amplitude
strength so basically what is happening suppose I take a G f value at a
particular frequency and I take another G f dash at another frequency so
it is actually their relative spectral strength. actual spectral strength if Δ f
is tending towards 0 is 0 for both of them but the relative nature is still

137
being intact. G f and G Δ f will have a relative strength to each other but
their actual strength.

If you wish to calculate that must be multiplied by Δ f then only you

will be getting the equivalent value with respect to Dn okay, so once I do
that and Δ f tends to 0 I do not get anything so that is why whenever we
talk about this G f it is actually a spectral density not a spectrum itself
whereas in Fourier series whatever we are getting that is actually a
spectrum because they are the individual value that we are getting that is
a real value and that value just specifies in that frequency component
how many how much strength it has. whereas here I have all the
frequency component in such as continuous.

So because I have every frequency component each individual

component like that example I have derived that each in visual
component does not have a probability value so if I say 30-degree angle
when that pivotal rod will stop and what is the Associated probability
that probability value will vanish that will be 0 here also same thing
what is the spectral strength of 5 kilo hertz if I say that will be always 0.

The answer will be that that 5 kilo hertz be 10 kilo hertz be five point
five kilo hertz or five point six kilo hertz, everywhere the spectral actual

138
spectral strength 0 but there is a relative strength between them and the
spectral density is well defined which is G. G is defined and that will
have relative nature so G at five might not be same as the at ten so they
might be different but if you just ask what's the spectral strength at five
and ten it will be all zero because by definition the G which we have
evaluated.

Now you can see that that is giving me spectral density not spectrum so
this is something why I am repeating again and again because this is
very important you need to understand that whenever we are plotting
this G(f) it is not giving me the actual spectrum it is spectral density we
are plotting actually spectral density and the difference between our
energy signal and periodic power signal. In Periodic power signal we
take Fourier series and each frequency component that we represent over
there gives me the exact value of the amplitude and corresponding phase
of that frequency component. here in the spectra so that is why that
spectra is discrete spectra it has some values where it is defined and all
other values it is not defined. and at every value you have a definite a
finite spectrum strength or I should say is the amplitude of that particular
frequency .

Whereas in a Fourier transform where we are representing energy signal

or time bounded signal we have every frequency component so we
cannot discriminate that this frequency is there, it is a continuous thing

139
with frequency whenever we plot it is a continuous thing and the
corresponding G(f) that we plot it is not spectra by itself because at a
particular frequency the height I will be getting that is actually not the
spectral strength because spectral strength is that at that particular
frequency the value G(f) into some Δ f which tends to 0. because it is a
continuous so the Δ f goes to 0 so basically the spectral strength will
still become 0 whereas what we get is that spectral strength divided by Δ
f that means spectral density per unit frequency,

how much it has. that is something that we get okay so G(f) defines that
so that is where these two particular spectrum haves light differences
and you have to be very careful about when you are generally talking
about these things often we do that mistake and in many books also
you'll see that they just represent it as spectrum so they said for a
particular time bounded signal I do Fourier transform and then I plot
G(f) that is actually my amplitude spectrum it is not or modulus g (f) I
draw I get amplitude spectrum it is not it is spectral density it is no way
spectrum because the spectrum is g (f) into Δ f which has no value.

It will be 0 everywhere if I try to plot that it is just a density which has

some meaning over there in a way it's meaningful because the signal that
we are seeing that is a finite energy signal right we are now representing
it into frequency domain if in the frequency domain every frequency

140
component has some finite value and then there are frequency
component which are infinite in number because it takes continuous
values so every frequency value it takes so then overall energy will be
infinite.

So it should not be that way it should be each individual component

should be still 0 all it is that spectral density has some values so that I
integrate from one particular frequency to another frequency, that means
I take the area which is actually the energy that will have finite value so
the integration will still have a finite value but individual one will not
have any value so this is a very important concept which you should be
able to always remember okay so whatever it is now what we have done
is we have understood what is Fourier series and what is Fourier
transform how we get Fourier transform out of Fourier series and what is
it is significance so for any signal g(t).

(Refer Slide Time: 11:12)

141
 ∞

∫−∞
G( f ) = g(t)e −j2πft dt
∞

∫−∞
g(t) = G( f )e j2πft df

We have also got these two relationship that one is this g (f) which is
Fourier transform it is G t e to the power - j 2 pi f t D t integration
-infinity +infinity and we get inverse Fourier transform which - infinity
+ infinity Gf) e to the power j 2 pi f df so this is in frequency
integration we get back our time domain signal this is the time

142
integration from the time signal we get the frequency representation of it
okay.

So this is something we have now identified now what we will talk

about why we are doing all these things because this is this has a huge
implication in communication you will see that every part of
communication techniques that has been derived that has some route in
Fourier transform or Fourier series you see that and we will one by one
derive all of them and you will see that this understanding is pretty much
essential for our overall understanding of communication technique.

So what we need to now do is try to get some very important property

which are particularly in use for communication technique so that
something will be now dealing with so before that let us try to see
whatever happens in Fourier series that it is a event symmetric spectrum
and all those things let us try to derive that part also.

(Refer Slide Time: 12:42)

143
 *
∞

[ ∫−∞ ]
G*( f ) = g(t)e −j2πft dt

∫−∞
= g(t)e j2πft dt
∞

∫−∞
= g(t)e −j2π(−f )t dt = G( f )

144
So we are trying to derive this complex conjugate of g ( f) what we get
so G (f) the expression of g (f) and we will take complex conjugates so
that is g (t) e to the power -j 2 π ft dt integration - infinity we take a
complex conjugate of this now as long as my g(t) that signal that we are
targeting to get Fourier transform is real, if this is real and if I take
complex conjugate because integration has nothing to do conjugation.

So conjugation will come inside then it will be distributed over these

two g (t) is complex conjugate because it is real so that remains g(t) so
we get minus infinity plus infinity g(t) and the complex conjugate of this
so that should be e to the power plus j 2 π ft dt okay, so what is this let
us rewrite this in a different fashion so we write g(t) e to the power - j
2π(-f) t d t okay, now if you see this is almost like a Fourier transform g
(t) into e to the power - j 2 π there is something into t dt.

So this is actually if it was f it could have been gf because it is minus f,

this must be G (-f) so if you now see it is almost giving similar result as
Fourier series so d-n if you do complex conjugate of d+n you get g-n d-
m same thing is happening for a particular frequency F if I do complex
conjugate of gf I get the negative of that so basically it says that
spectrum also if I take amplitude spectrum it will be when symmetric
and if I take same spectrum if Gf) is complex then I will get because the

145
complex conjugate is just negative so I will always get a odd symmetry
in phase.

So this will also have if you clearly see Gf) if it is complex it will have
amplitude as well as it will of phase so it will always be having two
spectrum one is for amplitude, one is for the Phase. okay so we are
saying there is a spectrum this is again you have to remember that it is a
colloquial term what you use generally but it is no longer as long as we
are talking about energy signal we are not talking about spectrum its
spectrum density so now let us try to derive few Fourier transform pairs
okay.

(Refer Slide Time: 15:31)

146
 t
g(t) = e −a u(t), a>0
∞

∫−∞
G( f ) = g(t)e −j2πft dt
∞

∫0
= e −at e −j2πft dt
∞

∫0
= e −(a+j2πf )t dt
∞
−(a+j2πf )t
e 1
= =
−(a + j2πf ) a + j2πf
0

147
So let us say if I have a signal e-at u (t) we have already defined u(t) it
sits up to from - ∞ to 0 it is 0 and then from 1 onwards sorry 0 onwards
it is actually it is value is 1 okay, so this signal we want to get the
Fourier transform of that that a> 0 let us do it in a very simple manner
so we want suppose this is g(t) because we want to evaluate G(f) which
is nothing but -∞ to + ∞ or g (t) into e –j 2π ft dt okay now g t because
of u(t) it is defined from 0 to ∞ only so this integration - ∞ it will be all
0, - ∞ 0 it will be 0 so it should be evaluated from 0 to ∞ now g t
becomes e - at 80 because u T>0 it is always 1 so it will just give me this
value e –(a+j2πf) dt now let us evaluate this so it is just exponential e to
the power - a +j2πf into t DT. So that should be -1 by a+j2πf.

So right putting value 0 and ∞ at ∞ if I put f = ∞ this goes to means

0 so we don't have a value at 0 I will be getting see for t of course it
should be, wait a second that should be there 0 and ∞. so if I put ∞
this becomes 0 right so this is the integration one by that coefficient e to
the power - j of course that a should be there so if I put t = ∞ what
happens is goes to 0 because e to the power- ∞ so that goes to 0 so that
whole value will be 0. well when I put t =0 so this becomes 1 so it
remains as this so because it is - so that - will be canceled so this
becomes a+j2πf so as you can see for a signal g(t) like this our G(f)
becomes…

148
(Refer Slide Time: 18:16)

sgn(t) = {e −atu(t) − e atu(−t)} a→0

149
G(f) =1by a+j2πf it is a complex signal okay, so we have evaluated this.
good the next part is we will try to evaluate some other things so let us
try to see if we can evaluate another very important signal so we will use
this part to evaluate that that is called a signum function. so what is
actually a signum function. A signum function is something like this it is
almost like u(t) but it has something in the negative so basically it is 1 in
the positive half, it is -1 in the negative half so it is defined as this says if
I say signum function of t that is actually+ 1 whenever t>0 at this point
it gets a value 0 so this is open so the discontinuous function of course at
t=0 and this is- 1 t> 0so if you wish to do a Fourier transform of

this signum function you will see that we will have a technique called
single sideband modulation where signum function is a very important
function you will see and that is why we are doing this Fourier transform
so it is very important that we understand the Fourier transform of this
thing how do we evaluate this Fourier transform so let us try to see
directly you won't be able to evaluate this if you just put it this way and
you will see all kinds of ∞ coming up directly we won't be able to
evaluate this. But we can evaluate in a clever way we can represent this
signum function as this it is actually e –at u(t)-e at u(-t) into u sorry. so if
you see what is happening it is actually combination of these two first is
e –a t u(-t ) what does that means e-at looks like this so u(t) means it is
defined only in the positive half right and e-at means at t = 0 it is 1 and
from there it is exponentially decaying okay, now in this one if I put a

150
tends to and a is just tending towards 0 if I just put that what will happen
this particular value will almost become one.

And this will almost mimic the positive half because his will be the
slope of this exponential decay will remain as it is so it will for every
value because a tends to almost 0 so this will always become one so this
will be very slowly decaying and it will almost mimic these things ok
and this part it is u-t that means it is defined over here this range because
u t is defined from 0 to infinity therefore u-t is defined from -∞ to ..
right. so u- t means it is defined over here and I have put e a t ok so as a
tends to 0 this will again become a value 1 for every value of t okay, but
what is happening because there is a - so it will always attain a - and it
will be something like this okay, that is slowly decaying so at if I just put
some value of a some finite value of a okay so then e to the power a t
will be as t increases this will be decreasing almost similar like this here
from - 1 it will be actually increasing and at t=∞ probably it will go to 0
so it will at the negative this one it will start from- 1 and it will go to 0
but if a tends to 0 this will almost become flat so this particular function
as a tends to 0 almost mimics the signum function so what we will do
now to get Fourier transform of signum we will take this representation
and use the previous result to evaluate the Fourier transform of signum
function.

151
(Refer Slide Time: 22:54)

152
 F[sgn(t)] = F [e −atu(t) − e atu(−t)] lim a → 0

{ }
1 1 1
= lim − =
a→0 a + j2πf a − j2πf jπf

So let us try to see my signum function has been already represented so

Fourier transform of signum t must be Fourier transform of e-at u(t) –e
at u(-t) right so this is our Fourier transform now Fourier transform of e
to the -at-ut that we have already derived. e to the -a - u t is 1 by a +j 2 π
F so this and Fourier transform if a function is having a linear
combination Fourier transforms as it is distributed that is very easy to
prove just put the Fourier transform formula and you can see that if g(t)
is represented as C 1 X t + C 2 XT.

The integration gets distributed so basically corresponding Fourier

transform you can individually take so I can do a Fourier transform of
this one which is nothing but 1a+j2πf similarly you can evaluate this one
also e to the power a-t, -ut so that is very easy if you just take this
integration and replace – t/t okay, just take a dummy variable called j = -
t it will just become this and then if you just try to evaluate that what
you will get is this one take that as homework you will be able to
evaluate this from the fundamental principle as we have derived this one

153
similarly you will be able to derive for e to the power a t u – t so that is
pretty easy.

So if you do that you will be getting this okay, and we have to also put
now that particular part comes limit a tends to 0 so here we'll have to put
limit a tends to 0 whenever we put this we immediately get one by j pi f
because a goes to 0 so we have this and we get one by j 2, 1 by because
2 is there so 1/2 + 1/2 it will be 1 ,1 by j pi f we get this so signum
function that's a very important result even though it is a discontinuous
function it has a Fourier representation and Fourier transform
correspondingly so signum function has a Fourier transform of 1 by j 2 π
f okay.

So this is again another important property now the one we will be

describing probably that's the most important property of Fourier
transform which has been used heavily for communication this is called
frequency shifting property.

(Refer Slide Time: 25:41)

154
 g(t) ⇔ G( f )
g(t)e j2πf0t ⇔ G (f − f0)
∞

∫−∞
G′( f ) = g(t)e j2πf0te −j2πft dt

What it is let us try to see suppose I have a signal g(t). I already know
the Fourier transform this so I have a known signal let us say that is g(t)
and I also know its Fourier transform let us say this is G(f) here this is
something we know already so this pair we have already evaluated for
some means by means of Fourier transform we have evaluated, so this is

155
given suppose now we are trying to prove that if this signal g(t) is
multiplied by a exponential it will ej2πf 0t if that happens
corresponding.

Fourier transform will be just that is why it is called frequency shifting

property so it will be just the spectrum will remain the same it will just
in the frequency domain translated by a amount f 0 so this is something
we wish to prove that is a very simple way of proving this so suppose
we know G(f) =- ∞ +- ∞ g (t) e –j2πftdt so this is something we know
now we want to take Fourier transform of this one so therefore we have
to do let us denote that as G dash if that should be - ∞ to + - ∞ g(t)
into e j2πft e-j2πft dt so that should be the Fourier transform.

(Refer Slide Time: 27:32)

156
 ∞

∫−∞
G′( f ) = g(t)e −j2π( f−f0)t dt

G′( f ) = G (f − f0)

Let us try to evaluate this so - ∞ +∞ g(t) e-j2πt(f-f_0) dt so if I take just

this -j2πt (f-f 0)common so we get that is my G dash f. now if we
assume see what was happening my Fourier transform if g(t) has a
Fourier transform is the ej2πt(f) into t dt then I get G(f) so if I just
replace f by f - f 0 or a new variable so therefore from that analogy I can
get this is = to G of f - f 0 which is nothing but g(t) e-j2πt(f-f)dt so this

157
is something I get so immediately I can write that G dash f is nothing but
this so I can see that if I know C g t into the e-j2πt(f)dt that is Gf)
therefore g(t) e to the power j2πt(f-f) be Gf) - f 0 this is very obvious.

Because it is a just a function of whatever is there over here so it is

becoming this back function so what we are seeing that if we multiply a
signal with an exponential ej2π 0t that corresponding frequency f 0 is
just translate the spectrum to that particular part okay, so this is pretty
understandable now what will we will try to see why exactly this is a
very fundamental property for communication in the next class.

158
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
on
Analog Communication
by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 11: Fourier Transform (Contd.-3)

Okay so we have discussed about this frequency shifting property right.

(Refer Slide Time: 00:30)

159
 ∞

∫−∞
G′( f ) = g(t)e −j2π( f−f0 )t dt

G′( f ) = G ( f − f0)

So that was the end result that we have got in the last class so basically what we have seen that if
I have a particular signal GT.

(Refer Slide Time: 00:39)

160
 g(t) ⇔ G ( f )
g(t)e j2π f0 t ⇔ G ( f − f0)
∞

∫−∞
G′( f ) = g(t)e j2π f0 t e −j2π ft dt
∞

∫−∞
G′( f ) = g(t)e −j2π( f−f0 )t dt

G′( f ) = G ( f − f0)

And I know the Fourier pair of that if I multiply that signal with e to the power J2piF0 T I get a
frequency shifting or frequency translation by an amount F0T whichever exponential I am
multiplying by okay so basically what happens the shape remains the same it just gets translated
so now let us try to see what is the implication of this thing okay the biggest implication is
suppose I have a signal G (t).

(Refer Slide Time: 01:13)

161
 g(t) ⇔ G (t)
F T [g(t)cos (2π f0t)]

[ ]
1
= F T g(t) (e j2π ft t + e −j2π f0 t)
2
1
2{
= F T [g(t)e j2π f0 t + g(t)e −j2π f0 t ]}
1
= [G ( f − f0) + G ( f + f0)]
2

Now let us say I will be multiplying this signal with cos this is the very typical actually this is
called modulation okay so let us say I will be multiplying by a cos, co sinusoidal signal 2 pi F0 t
okay,I'm doing this and this G(t) is my actual signal probably and this has a Fourier pair already
known which is gf so suppose I have a signal GT in time domain and I have a corresponding
Fourier pair which looks like this is the Gf, you see that because it is a real signal, I have
constructed the spectrum in such a way that it is even symmetric. So this is the amplitude
spectrum I have drawn correspondingly there will be a phase spectrum probably which is odd
symmetric something like that.

162
Okay, so I have got this G(f) now I know that this g(T) goes to GF okay, in frequency domain
now if I multiply by G(T) with cos w CT what will happen to this g(f)how that will look like how
the spectrum of this composite signal will look like so I have to just take a Fourier transform of
this okay.

Now what I will do, this cos I can write as ½ e to the power J2piF0T+ e to the power-J2piF0T
right, I can write this now because Fourier transform means this will get multiplied GT into this,
GT into this and then addition so linear combination Fourier transform gets distributed so I can
always write half it should be a Fourier transform of GT into e to the power J2piF0T plus Fourier
transform of GTi nbe to the power -J2p9F0T okay now go back to our previous results, so half if
GT has a Fourier transform of G F then if I multiply with this it should be G F- F0 and similarly
if I multiply by minus it should be – (F0), So that should be plus, so G (f+ F0) this is what I get
so basically what is happening first of all this gf strength is becoming half and it is getting
translated same G pattern is just getting translated to plus f0 so this is centered at plus f0 this is
centered at -f0 so suppose my f0 is something like this so what will happen this whole thing will
be little bit reduced and it will be centered around+ f0 and -f0 remember the shape of this
spectrum which actually specifies the signal quality that remains the same.

The shape does not change, the spectrum shape does not change I can again bring it back to 0 and
I will get the same thing. so this is in particular a very interesting property of Fourier transform so
what is happening, any signal if I know the Fourier transform of that and as long as the signal is
somewhat I should say band limited that means now the concept of band limited things will be
coming so what I know about the signal is if I see the spectrum of the signal the spectrum or
spectral density I should call that gf beyond some value it either completely vanishes beyond
some value B suppose either it completely vanishes or diminishes or beyond a certain level which
is insignificant okay.

So then we say that this particular signal because in time domain we cannot see this particular
signal does not contain any higher frequency component beyond some value B, so beyond B it
does not have those component, spectral component okay and then we call this signal as band
limited signals that means it is band limited up to B. so whatever spectral component it has or
frequency component it has it is limited up to bandwidth B or value, spectral value B so beyond

163
B does not have any spectral component or I should say any corresponding sinusoidal
component. So if I represent the signal like Fourier series or even in Fourier transform, so I do
not get any significant spectral component beyond this particular sinusoidal or exponential
whichever way I am representing okay, so once the signal is band limited if I try to multiply that,
that signal with a real Co-sinusoidal signal of frequency F0 where I have deliberately put this f0
to be greater than this B or sometimes much greater than this B then what will happen this signal
will get translated to that f0 okay.

So if F0 is not greater than B then what might happen so suppose f0 is somewhere over here I
mean the signal is or somewhere over here so this signal will be something like this and it will be
centered again at -f0 it will look like this so these two things will overlap so that is something I
do not want, okay so why I do not want that will be clear later on but right now we are saying that
this is what if happens this condition is true, there is a signal that particular spectrum part is still
as it is,it might be a little bit reduced but whenever they are reduce their relative part is
equivalently reduced.

So that means the spectral feature or characteristics which define the signal that remains the
same, so that relative strength at every point are relatively equivalent compared to this one okay
so the signal pattern remains almost the same so what is happening I am just actually translating
them to a higher frequency it remains the relative shape remains the same I am just translating
them to a higher frequency and this particular part is called modulation why now we can define
what is modulation and what is the advantage of doing modulation.

Suppose, I have a channel we have talked about channel in the first few process and it is a shared
channel and everybody wish to use that same channel so suppose I want to transmit something on
the channel so I put some voice signal to my transducer it converts it into a electrical domain
signal it becomes a voice signal and then it is being through antenna it is being radiated into the
air and by corresponding recipient wants to receive this as long as I am the only one talking to
one particular guy this is all good.

(Refer Slide Time: 08:52)

164
But suppose two fellows wish to communicate now after suppose this is by speech signals it has a
random variation and the corresponding spectrum looks like this okay it will, it will be a band
emitted spectrum in speech we say that beyond three point four kilo Hertz there is nothing all are
insignificant so it is band limited up to 3.4 kilo Hz so whatever spectrum component we will
have, it will be up to 3.4 kilo Hertz.okay, so another guys so this is my g1 (T) which is my signal,
another guy wish to also communicate he also has generated a separate speech signal.

A different kind of speed signal so this we call as g2T which also looks like similar spectrum
because it is speed so it will still have same kind of frequency component but it might look little
bit the spectrum might look little different but it will be almost up to 3.4 KHZ. now if I just super
impose this G 1 and G2 and put through antenna at the same time what will happen these two
spectrum will coexist and then neither in time domain because time domain the signal will be just
added. If it is additive channel and in frequency domain also the frequency component will be
just added then it is very hard to separate them out because neither they are separated in time
domain nor they are separated infrequency domain so I will have no device or no mechanism to

165
separate them out so if my receiver wish to now listen he will get a jumbled signal or added
signal g1 with g2 and that will completely deliver him a completely different speech.

It will be, it will be just a distorted speech so he will not, neither he will be able to listen neither
the other guy wish to receive g2T will be able to receive, so that is where if I wish to multiplex
multiple users data into the same common channel I need something else the device that was
designed by the understanding of frequency shifting property is something like this let us
multiply this G1T with some cos 2pi F0T or F1 T let us say.

Then what will happen this particular frequency component will go around f1and -f1 and it will
sit nicely over there and when we are trying transmit G to T we do a separate multiplication, with
again a cos term but this time the cos frequency will be different > let us do it at F2 T well F2 is
predominantly different from F1 and this particular signal will nicely be sitting over here and at
-F2 the good part is now in the frequency domain they have separate location and at the receiver
side what I can do is I can just filter them out. So there is a device called filter which most of you
are familiar that just specifically takes some of the frequency or it just passes some of the
frequency component and it suppresses other frequency components so if this is the composite
signal coming to the filter and my filter is tuned at f1 so what will happen is a band pass filter so
it will just take this amount and it will reject this whereas the other guy can tune his filter at f2
and he can reject this so both of them will get their original signal as if it is transmitted on the air
single.

As long as we choose this f1, f2 carefully and as many we wish we can actually multiplex many
we can have corresponding f3 corresponding F4and so on we can actually multiplex multiple
users simultaneously transmitting over the channels and their separate existence is still remaining
because of this frequency translation property what is happening, whenever we multiply with cos
2 pi F1 T or F0 T or F2 T or F3 T we know that the spectral component or the relative spectrum
which is the characteristics of the signal remains the same it just gets translated to a different
frequency band and all We have to do at the receiver we have to carefully choose the frequency
band put a filter and take extract our own signal and reject all of the signal this is facilitated by
this communication technique or that is why this particular part is called modulation so what you
are trying to do whenever you are trying to use the media which is being used by all others you
translate or you multiply by co-sinusoidal signal to translate the frequency spectrum into a higher
frequency and then actually transmit it in the common media.

166
Everybody will be doing that as long as they are using separate bands to transmit they will have
separate identity, in time domain it might not have separate identity but because we have done
this, frequency domain they will always have their own separate identity that is fantastic! so this
particular thing is called frequency division multiplexing so what we are trying to do is because
now you can see all these things are so important.

Because we have understanding of Fourier series and transform and because of that we could get
another representation of signal and from the Fourier transform only we could derive that there is
a property called frequency shifting property by which we can actually keep the signal intact but
translate it into a higher frequency and because we have some device called filters which can
separate out some portion of frequency.

So we can actually give separate entity for separate signal in the frequency domain so the entire
manipulation is done in the frequency domain if we would not have understood Fourier transform
would not have understood the frequency representation of a signal we could not have produce
these things okay if we were not able to understand the mathematics behind that frequency
shifting property we could not have produced this thing.

So it is very important means now you can appreciate it is very important for communication any
simple technique that has been applied in communication it is deeply rooted in Fourier series or
Fourier transform, it is absolutely necessary that you understand those concepts very clearly okay
so now we have understood some portion of it we have talked about this modulation why this
modulation is required we have told that one particular thing is this multiplexing.

So this is now probably clear you have stated earlier, but now it is all mathematically clear that
what we are trying to do and why that is important there is another aspect of this frequency
shifting property or this modulation that is we have already talked about that that whenever you
put multiply with cos this one you can see the frequency component that it is having whenever
we are transmitting it is around f0 some +/- so earlier it was having, if you just see this one it was
having frequency from 0 to some 3.4 kilo Hertz.

167
Now suppose I put them at it is a very high frequency let us have some 400 megahertz so what
will happen it will just go center around 400 megahertz and it will just be around that 400
megahertz 400+ 3.4 and 400 -3.4 okay so it will be around that 400 megahertz only so the
frequency component it predominantly had that has been transmitted in the air will be around that
400 megahertz whenever it is 400 megahertz the frequency component.

The corresponding wavelength will be very short okay 3.4 kilo Hertz one by that you will see the
corresponding wavelength is quite big whereas 400 megahertz you should do one by that you will
see because of that 10 to the power 6 it will be very small and we have also discussed that
whenever you are putting a transmitting antenna or receiving antenna to actually transfer the
energy properly we need to have the antenna size or to capture the energy properly through the
antenna we need to have an antenna size which is comparable to the wavelength of the
correspond frequency that may translate or transferring okay.

So it is very much essential that the frequency is higher to facilitate the antenna size to be
smaller, so the with modulation we are also achieving that particular part because what is
happening here if you see there are very small frequency component which will have huge
wavelength because it is inversely proportional and the antenna size will become very huge
whereas by this, by just this simple technique by multiplying the signal with a cos it serves two
purposes one is it translate it to a very high frequency immediately corresponding wavelength
will be very smaller.

So I can devise very small antenna for my transmitter and receiver and on top of that we are now
creating in the frequency domain multiple such places where we can start multiplexing multiple
signals simultaneously and transmit them through the air media simultaneously and just use a
filter to take my own signal and reject other signal so that is actually the story behind modulation
you will see that later on okay.

But right now we are happy with this frequency shifting property which will help us to do
modulation okay the next property that we are trying to provide is.

168
(Refer Slide Time: 19:37)

∫−∞
g(t) * w(t) = g(τ)ω(t − τ)dτ

F [g1(t) * g2(t)] ⇔ G1( f )G2( f )

Something called convolution I am not going to prove that but I will just state what do we mean
by convolution so this convolution is given by this star symbol so that means the signal is
convoluted with another signal GT is convoluted with another signal WT what does that means
convolution actually means this, so what is happening so you take this signal GT Tau is just do
not worry about tau, that the dummy variable of, in this integration okay.

So take the signal GT and for w, you can see it is already -tau so basically you reverse this so you
take W- tau so like u -T, I was taking earlier so take w- tau and then time shifted by t amount
okay so do this and then keep varying this tau multiply this and integrate so that is called the
convolution so basically any signal you take and the other signal you invert it first in time domain

169
and then time shift and keep varying this time shift and you integrate it multiplication and
integration .

So basically one particular signal you will take that will be suppose this is my GT let us say this
is like this and suppose the other one w T is something like this, now what you do first you flip
this w T so from there it would be W- T or- tau so this is w tau this is G tau okay so time I am
defining respect to tau and what you do for a particular so then you time shift it at tau = T you put
it okay so this will be just at t okay. So whatever that value of T wherever that T will be it will
look like that so at, so this will be wt-tau so what different value of 7 this will this signal will be
I mean going to 0 at different location and then you multiply these two things and integrate it, so
basically you will be keep on shifting this for different value of T and you will be getting a
different value of those integrations so this is called convolution and there is a well known result
in Fourier transform so I will give this as homework you have to prove that if I have two signal G
1 T & G 2 T if I convolute them and try to take a Fourier transform of them if I individually know
the Fourier transform of G 1 T which is suppose G1F and G 2 T which is G 2 F we can always
prove that this should be G 1 F into G Z F.

So in time if we convolute in frequency it becomes multiplication and if in frequency if we

convolute in time it will be multiplication that comes from another property of Fourier transform
which is called a duality so the duality property if you wish to know that is also very simple you
can just very quickly give them.

(Refer Slide Time: 23:02)

170
 g(t) ⇔ G ( f )
∞

∫−∞
G( f ) = g(t)e −j2π ft dt
∞

∫−∞
G (t) = g( f )e −j2π ft d f, −f = x
∞

∫−∞
= g(−x)e +j2πxt d x

So we know that GT I have a Fourier transform which is called G F okay so in that case what we
can write is this G F is nothing but -infinity to plus infinity GT e to the power - j2 pi Ft dt right,
now so this is something we already know we also know because it is a Fourier transform so we
can also represent this G T as -infinity plus infinity gf e to the power +J2pi F T d T sorry d F right
we know this.

171
Now what we can do, instead of F if we just put -F so then from this equation we can get GT I
can write -infinity plus infinity so f is replaced by-f. So what will happen this is G -F and here we
will just get e power -J 2pi FT right, so this is something we will get okay, now all you have to do
is, okay so what we have done we have just replaced this okay so we can do one more thing
instead of doing it over here I can,I can also so do it over here so basically what we are trying to
do I have got this right now you just replace T by F so T = F if I just put that so what will happen
this GF will become GT and I get -infinity to plus infinity this becomes G F e to the power -J2piF
becomes replaced by T and T becomes replaced by F. So I get F T and T becomes G f right, so
this is something I get okay so I get Gt is equal to -infinity to + infinity G F e to the power -J2pi F
T d F okay so now you can see that, if I have this as
frequency domain okay so this becomes T right so this is almost becoming inverse Fourier
transform but I only have one problem this there is a minus over here inverse Fourier transform is
having plus so what I can do is instead of this-f I can put -F right.

So immediately what will happen if F is replaced by minus F or I can write minus F means,
minus F as some X so immediately what will happen so this will become X so this is becoming
G- X because itsF, F is - X, G- X e to the power minus this would become plus because - F
becomes X so J2pi T X and d this would be dx, - and + so that will just change the limit so this
will still remain the same okay so G T becomes this now you can replace X by F immediately.

(Refer Slide Time: 27:00)

172
 ∞

∫−∞
G (t) = g(−f )e j2π ft d f

G (t) ⇔ g(−f )
g(t) ⇔ G (t)

What you get our G T is becoming just -infinity to plus infinity G -F e to the power J2piFTd f
right so what is happening now I can see this G T and G -F are Fourier transform to each other so
G -F if I take inverse Fourier transform I get G T so that becomes a Fourier pair so basically we
have started with a Fourier pair GT which has a Fourier transform of G F if I just, this whatever
time domain function I have got that if I represent in frequency domain if I just take negative of
that, that will have whatever frequency domain function I have, if I just represent it in time
domain I will be getting that so that is called the duality of Fourier transform so that means
suppose I have already got a Fourier transform let us say I do a Fourier transform of some, this
kind of pulse okay which is defined from - T/ 2 to + T /2 take that as homework if you do a
Fourier transform easily you can just put the fundamental Fourier transform this things you will
see that it will be a sin function it is defined by this it is Fourier transform will be GF should be
tau sin C pi F tau, sin C means sine of this argument divided by this argument.

173
So sin pi F tau divided by pi F tau so this gf becomes this, which will be
defined a,s at t equal to 0 it looks like this at means tau T = 0, it is having a value of tau
at 1/tau, it goes to 0 again at 2 / tau it goes to 0 and so on, at- 1 /tau, it goes to 0 and so on, okay
so if I already have got a GT which is looking like this and if I already have got a GF which looks
like this, now if I give a sinc function in time domain that means this becomes GT and if I wish to
actually get a Fourier transform this would be just similar like this this duality formula tells me
that.

So if the frequency domain 1 okay frequency domain 1 is now becoming my time domain
function so now the sinc function which was the frequency domain representation that becomes
my time function this will be just same time domain function will be the Fourier transform now
this will be frequency domain okay so this will be now this is in time domain this is in frequency
domain if I take this in time domain it will just be frequency domain the time domain1 becomes
the frequency domain 1 only thing is that it must be having this -F but because this function is
symmetric over T or F, So it will remain the same.

so they become ups Fourier conjugate of each others that means a square box function if I take, if
I do a Fourier transform I get a sinc function, if I take a sinc function in time domain if I do a
Fourier transform I get a box function in frequency domain, so if I know one I will be always
knowing the conjugate of that so that is the duality property of Fourier transform.

so with all these basics what we will try to do is we will try to now go into the measurement part
which we have ignored means we have done that for Fourier series but for Fourier transform the
measurement part that means the energy or power we have still not devised that so now we will
try to see for a signal the very important property is the measurement of the signal which is
energy so we will try to see how to evaluate the energy of a signal so that will be our next target.

thank you.

174
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
on
Analog Communication
by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 12: Energy Spectral Density

Ok, so in the last class what we have discussed is how from last few classes I should say,
how from Fourier series towards Fourier transform. so that means we have a energy signal means
we have a power signal from there, we get Fourier series, that we have already explored and then
if instead of energy signal means power signal we have a non periodic signal., that is energy sig-
nal which is time bounded for that how do we actually evaluate the frequency components. so
what we have seen that whenever, we have put a trick over there, so instead of doing the analyses
completely in the new fashion we have actually borrowed the idea from series-what we did we
have taken a time period bigger than the existence of the signal.

(Refer Slide Time: 01:21)

175
That means if the signal is defined from let say a to b we have taken the time period as the time
period which includes the entire thing and then we started repeating the entire signal. So what
was happening the targeted signal was repeated over the period T if I say this is T, and then we
said that immediately if I start repeating it up to time plus infinity and minus infinity, then imme-
diately it becomes a again a power signal and it is a periodic signal. so we get a Fourier series ac-
cordingly.

And then we started stretching this T to infinity. so immediately you could see that you are get-
ting a Fourier transform case. so we have given interpretation of Fourier transform, what exactly
it means for a signal g t we get a transform GF. what is the meaning of that 2 he have said that at
any frequency f, the value of GF is not actually the spectrum component as it was for a discrete
Fourier series case so we have told that generally gives me 0.

(Refer Slide Time: 02:44)

176
g(t)

But if I just multiple GF with some ∇f, so within this ∇f how much frequency component is there
can be actually evaluated. so that is why GF is called as spectral density. because GF ∇F is al-
most similar to that Dn, and then G (F) ∇ F/∇ F is actually giving me G(f) . So this is actually the
spectrum divided by the frequency part so it becomes the spectrum density so this is something
we have already discussed. okay so we have given some interpretation of GF what that mean.
now what we wish to do is now from here to measurement.

We have already define that energy or power is kind of measurement for us. it is almost like vec-
tor also, we get the distance of the vector which is measurement. we have shown that similarly if
we just integrate either g2t or mod gt squared we get energy and corresponding if it’s the power
signal then we can calculate evaluate power just dividing by time stretching the time up to infini-
ty.

177
So okay, so will now try to get into the energy or power of a particular signal. so our target is
similar to what we have done we also have to evaluated the power of the periodic signal right!
that something we have already talked about, and we have given the equivalence, means power
calculation theorem which is called Parseval's Power theorem. Similar theorem will be trying to
drive for a time limited signal or a energy signal. okay, let's try to see.

(Refer Slide Time: 04:28)

g(t)

So if I have a signal g t which is time bounded so the signal can be anything and I’m just taking
this signal, it might be of any nature. so it might be of this nature, define between some finite
amount of time. okay so if I just take that g t, I wish to now evaluate the energy of this signal so
you might be saying okay, it is very easy, we already have defined this particular thing.

(Refer Slide Time: 04:59)

178
 ∞ ∞

∫−∞ ∫−∞
2
| g(t) | dt = g 2(t)dt

so it is all about integrating mod gt2 at eight. or if it s real signal then it is just integrating g
square t d t right. so this is our means this happens the signal is real this happens when this signal
is complex. right or we can write mod gt2 as g t, g star t d e right. so what we will do we probably
take the more generic definition which is the complex signal definition because complex signal
already takes the real part eight. of it is real then g start t will be gt then immediately we get g2
right.

So for that case I wish to evaluate this you might be asking this, this is my energy. you might be
asking this, by definition already he have defined this. we have by definition of Fourier series and
Fourier transform we have already defined that this should be the energy right. so this is some-

179
thing that was already derived so what we are targeting now. like what we have done in Parseval's
theorem for Fourier series, we wish to also define energy from the frequency spectrum.

This is what we are seeing from the time okay. So g t has equivalent Fourier transform, let's say
that in G f because g t is time limited signal. so it must have a corresponding Fourier transform
which is we'll defined okay so as long as GT as some property that we discussed already. So GF
is already defined so I want to see in the frequency domain can we evaluate the power of all en-
ergy similarly. so for this signal because it’s a time limited so it should be a energy signal so we
will be interested in energy. so let’s try to see if we can define that thing. so what we can do is
something like this.

(Refer Slide Time: 07:13)

∫−∞
Eg = g(t)g*(t)dt g(t) ⇔ G ( f )

180
We have already talked about the inverse Fourier transform okay. so in that inverse Fourier series
you have said g t is represented as long as GF is known. it is -∞ to ∞ G (f) ej2πft df so that is the
theorem we have proven already. That is the inverse Fourier transform so this G start will be re-
placing by this. so it will be one integration over f G f, of course because we have to take com-
plex conjugate because this is g t we need to get g*t, g*t will be g*f so there will be minus sign
over here. So I can replace it by that. so it will be a star f e to the power-j two pi f t d f so no prob-
lem in that whatever we have derived so far we are just using those things. because we want to
take it to frequency domain that why we are applying to get inverse Fourier transform. So the
time domain signal I’m actually representing it through inverse Fourier transform to frequency
domain signal. but so far I have already succeeded in replacing one of them. Of course g t should
be still there. So this is where we have replaced g start. okay we have to see what happens to this
sheet so lets see this integration limit okay.

(Refer Slide Time: 08:46)

181
 ∞

∫−∞
Eg = g(t)g*(t)dt g(t) ⇔ G ( f )

∞ ∞

∫−∞ [ ∫−∞ ]
= g(t) G*(t)e −j2π ft d f dt

∞ ∞

∫−∞ [ ∫−∞ ]
= G*( f ) g(t)e −j2π ft dt d f

∞ ∞

∫−∞ ∫−∞
= G*( f )G ( f )d f = | G ( f ) |2 d f

So they are depending on each other I can always spilt these two integration. I can take the fre-
quency integration outside so whatever terms which involve all the frequency I can take that out-
side. so g2f, sorry G start will be this. inside the time integration should be there where GT is al-
ready there e to the power -J two ti f t. 1 this is also a thing which is depend on time. So I cannot
take that outside the integration. so dt integration is done first and then df. Right, no problem I
have just rearranged it and I am allowed to do that rearranging because the limits are not depend-
ed on each other okay. and whichever function are depend on both f and t, I have not taken them
out. so I have done all things correctly.

Now let us see what happens over here, can you identify this thing. it is actually the Fourier trans-
form of g t we have already defined. that and okay and we have derived that. So this is Fourier
transform of g t therefore this must be GF so that is GF, d f so that is basically what has happened
we have started the definition of energy from time domain and eventually we could get into fre-
quency domain with same energy. so what happens in frequency domain. It is nothing but mod
gf2df. right !

So basically, in frequency domain. now you can see also what is happening my GF might be
complex term. we have seen that already it might have a phase component, it might have a ampli-
tude component. What here we are doing because we are measuring the real quantity, Eg so basi-
cally we taking the formula says that I must take the modulus of GF which is real again and
square of that, integrate over the entire frequency domain. whatever I get back, it is actually ener-
gy.
so in the frequency domain whatever pattern I’m getting.

(Refer Slide Time: 10:54)

182
| G ( f ) |2

This mod GF2 if I plot. okay so, suppose this frequency and I’m plotting this I have a corre-
sponding g t say g t looks like this. and I’m plotting mod GF2. Eventually looking like this. So
you will later on see if g t is this box function it becomes okay Sinc square so sinx/x square of
that okay. so it look like this. so if my GT is this, I plot this is actually mod GF2. this I call as en-
ergy spectral density’ or ESD. why it is called so because first of all if you integrate it to minus
infinity to plus infinity, that means all the frequency components, you are actually trying to eval-
uate something which is mod Gf square. I still do not know what that is. so if I integrate that I get
the overall energy. good and if I suppose, try to see what are the frequency component it has or
the measurable quantity of some frequency component it has. If I try to see that, what I will be
doing I will be passing this through a bandpass filter, a very narrow band pass filter of, let’s say a
band pass filter centered around f c and with, this one as delta f, so the filter if you correctly plot
that, the transfer function will look like this. So it will have, a paste, there will be flat band, where
it will pass every frequency. And-fo aho, because of the symmetric nature in Fourier Transforms.
so it will have some things. okay,

So if I have pass this what will happen, I will have the multiplication of these two. right! so im-
mediately what I will get, only at that frequency, this mod GF square will exist. because it will

183
just pass, and in all other frequency it will be 0, and then if I just try to see that suppose the band
is over here. lets' say so what I will get is just the same almost similar pattern over here. and if I
just now integrate what should I except now, that in those frequency band how much power it
has. okay so in that del F, because of overall power is if I integrate it fully. now I am passing it
through
filter. Filk will be passing only on that band whatever it has exactly as it is.

so if I now passing, after panning through filter now I wish to calculate the energy that must be
those frequency component energy. Because those are passed through this filter so that frequency
component energy is becoming just the integration of this mod G f square over that frequency
band. so this is the reason why this is called energy spectral
density because what is happening for a particular frequency band energy is there that is being
characterized by this mod Gf squared.

(Refer Slide Time: 14:28)

| G ( f ) |2

184
Okay. If you just target the particular frequency at that frequency just if you take only that fre-
quency, you will not get any energy. because if you integrate, it will be 0. right. because I don’t
have any space to integrate. it’s just a on a point I am trying to do integration and that will be o.
Because there del F is o. If I allow some del f, I will be immediately getting some amount of en-
ergy, and that energy characterises, at that frequency it del F is sufficiently small, at that frequen-
cy around that how much energy part unit band it has. because it is integration of that, that means,
it this del f is sufficiently small,

(Refer Slide Time: 15:07)

2 | G ( f ) |2 Δ f
Δf

That means, it is almost flat right. And then the value of mod G f squared into del f. Because it is
almost flat. it can be approximated on flat.

If this is drawn like this and I have chosen a very small this one you can see that top is almost flat
if I make by Δf tending to 0. So, then the overall energy will be, this mod Gf square almost simi-

185
lar concept into delf. Of course because it is two side band, 2 sided it should be 2 into that. Be-
cause d will getting the same thing in both the sides. okay. if I just ignore this 2, it will be always
there which ever band I choose it will always there. so what is happening If I divide this by def f.
okay. I get to see per unit band, because I’m dividing it by frequency. This mod GF2 it is actually
giving me the energy per unit band. so that is why it’s called energy spectral density, because
mode GF2 gives me idea that at any frequency I choose, permit band how much energy is being
there, in that signal.

So it actually characterises the over all signal, became you have told that this energy is a mea-
surement so it characterises the overall signal. it just says every frequency turn, how much energy
or spectral density, I should not say spectrum should be spectral density how much spectral densi-
ty it has in that target ed frequency.

okay because it characterises that, so that is another way of characterizing the whole signal okay.
so that is the importance or significance of this. so we have now got some measurement or mea-
surable quantity in the frequency domain as well. Because now we have defined two separate
domain for representing a signal one is the time domain, another one is the frequency domain.
now I have got the frequency domain representation of a measurable which is energy okay. or
energy spectral density. That is something we have got. now we will introduce another thing.

(Refer Slide Time: 17:30)

186
Okay, so that is, if I have a signal gt how I do l get energy spectral density. the technique is very
simple.first do the Fourier transform. okay, so whenever I do a Fourier transform,I will be getting
corresponding mod G f square. first I will be getting GF and then from that Gt I will be calculat-
ing mod GF square. that becomes my energy spectrum density. this is one way of evaluating en-
ergy spectral density. but I would say most of the time the helpful one is not this one. We Will try
to see the what is the other method to actually evaluate energy spectral density. so for that we try
to define a particular term which is called time auto correlation function. What is time auto
fruition. so, suppose I have a time bounded signal GT might be like this might be anything, but
bounded in time. okay so this GT if I multiple with either a advanced or the delayed version of
the same signal. t plus τ let us say it's advance by τ amount of time. and if I integrate this over the
entire time. so I take a signal.

So it might be this signal if I give this signal and then probably this means actually it will be de-
layed. so it will be shifted backwards so if I just shift it by the amount. so this is actually τ and
then I will multiple there two, GT and this is GT plus τ. so i t I multiply these two signal and start
integrating it. we are saying we are defining it. Tell right now you will not appreciate why we are

187
doing this. Later on you will see a big implication in communication or signal processing. We are
defining this phi g τ. Of course we are integrating it over to you but τ Will remain. it is a function
of τ, for different values of τ the integration will be different. you can see already if I put a τ
which is sufficient in large than this particular duration of this signal; then it becomes o. whereas
here it is not 0. It has a overlapping part. Of course this will be depending on τ I choose and it is a
function of τ. okay. so, this is our time auto correlation function. first of all let us see some prop-
erty of it. will later on link this two something of our interest, and then we will appreciate why we
are defining this.but let us first try to characterize this particular signal. so first thing we wish to
proof is. That it is an even symmetric function.

(Refer Slide Time: 20:47)

188
 ∞

∫−∞
ψg(τ) = g(t)g(t + τ)dt t+τ =z
∞

∫−∞
ψg(τ) = g(z − τ)g(z)d z
∞

∫−∞
= g(t)g(t − τ)dt = ψg(−τ)

ψg(z) = ψg(−τ)

So let us say my definition was this okay. so now let us substitute t plus τ as z okay. so what will
happen to t, that should be z-τ, right, and DT will be τ is constant. so for this integration τ is a
constant. d t will be d z. So this psi g e should be-infinity to t infinity g t g ok, now we have to do
the substitution. so t will become gz-τ, and t +τ will become dz. so again z is just a dummy vari-
able. I can replace that with t. So it will become -infinity to + infinity g t g t-tau at. just think
about this. This definition. what is happening if this is est g tau this is actually psi g-tau by this
definition. Because if tau is replaced by -tau, this automatically becomes. If I replace tau by-tau,
this becomes psi g -fan which is exactly equal to this particular thing. But this equal to this.
Therefore this must be equal to this. so I can see my psi g tau function as an even symmetric
function. Became the negate of tan. If I replace by-tau, I get the same value. so it is an even
symmetric function. right

okay next the most important thing I will be doing. so now what I wish to do is I wish to Fourier
transform of this one. It will be very clear after sometime why i am doing this. okay, so right now
just doing it.

(Refer Slide Time: 23:10)

189
190
F T ⋅ [ψg(τ)]
∞ ∞

∫−∞ [ ∫−∞ ]
= g(t)g(t + τ)dt e −j2π fτ dτ

∞ ∞

∫−∞ [ −∞
∫ ]
= g(t) g(t + τ)e −j2π fτ dτ dt

t+τ=y
dz = dy
∞ ∞

∫−∞ [ ∫−∞ ]
= g(t) g(y)e −j2π f y e j2π ft d y dt

∞ ∞

∫−∞ [ ∫−∞ ]
= g(t)e j2π ft g(y)e −j2π f y d y dt

∫−∞
= g(t)e j2π ft G (t)dt

= G ( f )G (−f )
= G ( f )G*( f ) ( ∵ G (−f ) = G*( f ))
F T{ψg(τ)} = | G ( f ) |2

I wish to take,Fourier transform of this particular signal.right. Psi g τ of course when I’m saying I
am taking a Fourier transform, my Fourier transform is actually transforming from tau Domain to
some other domain. Let us call that as f domain. no longer that t domain, here the independent
variable is τ. so Fourier transform should be on τ. Let us put that in Fourier Transform. Fourier
transform means I have to put the function.

So which is nothing but, -infinity to + infinity g t, g t + tau d t. This is the whole signal that has to
be Fourier transformed. Into e to the power-j z pi f t, sorry f tau, Became the Variable is now tau.
d tau integration-infinity to + infinity. So that is the Fourier Transform right. Let's now try to see
if we can evaluate this. Again I will do a change of this integration. became they are not indepen-
dent of each other. The limits are not independent. so I can keep the t integration out and the tau
integration inside. So which ever is free of tau, that will be going out. So g t goes out, I have-in-
finity to + nifty g t + tau e to the power-J a pit tau d tau.

So this is something which is there inside. Now again I will do the same trick IT + tau I will re-
place with some y let’s say. So immediately what do you get, minus infinity 2 + infinity GT, now
minus infinity 2 + infinity, it is becoming my y, g y. okay. Now tau is a variable. So tau must be,

191
and for this integration t is a constant. Because this integration is over τ. so t is a constant for that
integration. So we must not bother about T. Tau becomes y minus T, this becomes e to the power-
j s pi f y into e to the power + j a pi ft. right. Just replaced tau by y-t So y minus is there -t be-
comes + d t. Lets see Joe this integrations inside, this is a constant term. Because it has nothing
dependent on tau. I can take that out. right. what I have is ,I should also say D tan, became t is
constant that should be replaced by d y. And the integration because it is minus infinity 2+ infini-
ty that is not changing. now I have got g y e to the power j 2 p i fy dy. minus infinity to pan infin-
ity, d t. Whats thisThat is actually Fourier transform of G. If I known already Fourier transform
suppose GT over any variable GY or whatever it is because the transform takes it to another do-
main. So this is G F suppose , that must be, y is just a dummy variable. I can put, treat y as t. twill
be getting Q F over here. So I can immediately write this-in infinity to +Infinity g t e to the pow-
er-j s pi f t.

This becomes - G F d t. There should be, this was plus, this time was plus, fine. Now GF is no
longer a function of T I can take that out GF goes out what I’m left with GT E to the power of J 2
pi ft d t. If this was minus I could have got the Fourier transform right I can forcibly put a minus
over here then I have to put minus F over here. Right immediately what happens this is the Fouri-
er transform of water work is there inside so that should be G minus F so I can write G F Into G
minus F right we have also proven that Fourier transformer is even symmetric. So G mins F is
Nothing but the complex conjugate of GF so ji minus F is G*F. I can put that as GF into G * F
now you can see the beauty of it no what has happened I have started with autocorrelation func-
tion. This is why I have defined the autocorrelation function because I knew that autocorrelation
function if I take Fourier transform I will be getting back my energy spectral density. so this is
that is why, we are telling there is another way of defining energy spectral density.

it is just' you take first, instead of taking directly Fourier transform we will see why this is re-
quired.you take first take the signal, take the auto correlation of that-okay and then you do Fouri-
er transform. And immediately you will get back your auto correlation function later on we will
see this is actually the famous Wiener-Khinchin relationship. of course we are done that for only
time correlation. This has Wiener Khinchin theorem has higher implication. when we study about
random process we will see that it is also related that. so this is/right now we should be satisfied
with if you just take a signal, single signal don’t think about randomness in it a single defined
signal if I take a time auto correlation okay I will do a Fourier transform on’ that I..And I get back
my energy spectral density, the way we have defined what is energy spectral density. Earlier we
had already defined what is energy spectral density, so now with time auto correlation function.

192
We are getting back energy spectrum density. what we will do next is this was done for a energy
signal. we will do the same thing for a power signal and then we will try to interrupt why this is
so much important. okay, why we have to take this D2 apart. where we have take evaluate this
time auto correlation function and then try to take Fourier transform. we will try to give a simple
example where you will see that Fourier transform is not good enough just directly taking Fourier
transform might not be good enough..

193
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
on
Analog Communication
by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 13: Power Spectral Density

Okay so we have already discussed about for energy signal time auto correlation function and we
have also proven that whenever we take the Fourier transform of time autocorrelation function
that becomes the energy spectral density we will just try to see we have also discussed in one of
the previous class the modulation so let us try to see what happens to means whenever we modu-
late a signal what happens to the energy of it and the corresponding energy spectral density okay.

(Refer Slide Time: 00:56)

194
 g(t) ⇔ G ( f )
ϕ(t) = g(t)cos (2π f0t)
1
2[
ϕ( f ) = G ( f + f0) + G ( f − f )]

F T {ψq(t)} = | ϕ( f ) |2
1
[ G ( f + f0) + G ( f − f0 )2]
2
=
4

So let us see we have a signal gt which is again a time limited and we want to generate a modu-
lated signal or a frequency shifted signal which is nothing but gt multiplied by cos 2p F0T so we
have seen already that if we multiply by co-sinusoidal it gets frequency shifted in plus and minus
as well as and the corresponding signal is called the modulated signal okay so I want to now
evaluate the power spectral density of this pi t right.

195
So basically what I wish to do is si phi t which is the spectral density of this one okay so first of
all if you wish to do that there is one technique we have already told that if I can get the Fourier
transform of this one that is something we have already evaluated if gt we know the Fourier
transform is capital gf then immediately we know that pi F should be half GF +f0 +GF-F0 this is
something we have proven in one of the previous class right so that is the frequency shifting
property actually of Fourier transform so if I know this so basically I know the Fourier transform
of this one all I have to do is take mod and square so that should be modulus of half Gf+ F0+ GF-
F02:

So that means ¼ right so if I just try to put suppose gf was something like this of course. What we
are saying that the signal is you later on see that if a signal is time-limited then it will not be band
width-limited okay so it will have frequency component up to infinity whenever a signal is time
limited but what we are saying that probably it has an effective bandwidth that means the other
component that it has earlier on what whatever we have discussed if gt was something like this a
time limited signal.

Then gf we have seen it has all the other lobes side lobes okay so we are saying that those are in
significant okay so the signal is having effective bandwidth up to this as long as the signal looks
like this then if I modulate so that pi F will look like at F0 there will be this one at minus F0 there
will be this one okay if F0 is sufficiently large compared to its bandwidth effective bandwidth
then what will happen.

These two things will not have any overlap in frequency domain okay as long as that is happen-
ing so this is actually characterized by this one and this is characterized by this one okay so as
long as these two are non-overlapping their square the cross product should be 0 because they are
non overlapping so I can always write if that condition is true that F0 is much, much bigger than
the bandwidth of the signal okay, effective bandwidth of the signal.

Then I can write this squares must be just summation of 2 square right so that is what we get now
if I just talk a. bout suppose gt was having a measurable quantity which is Eg let us say that is the
measurable quantity it might be interpreted as minus infinity plus infinity integrated G2t in time
or mod G2F or mod GF2 integrated from minus infinity plus infinity okay.

196
So anyway that is the Eg right so if that is Eg this if I integrate over frequency band so suppose I
want to I wish to now calculate the energy of this signal the modulated signal so e pi that must be
integration of this energy spectral density over the entire frequency okay so that integration
should be one-fourth integration of this one plus integration of right now this integration is just
equivalent Eg because it is nothing but the same gf shifted right the same gf shifted if I just inte-
grate mod gf2 that must give me same thing.

(Refer Slide Time: 06:41)

1 Eg
4[ g g]
Eϕ = E + E =
2

cos(ωmt)cos(ωct)

So this must be Eg this must be also Eg so what do I get effectively 1/4 Eg + Eg so that means
my energy becomes Eg/2 so that must be the E pi so always remember this is going to be very
essential for our analysis of any modulation always remember whenever we modulate that means

197
our signal is multiplied by a cost of unit amplitude always the energy gets half due to the process
of modulation.

That is a very peculiar thing you might be asking why this is happening I am having a signal gt I
am multiplying that signal gt by another co sinusoidal and by the process of this multiplication I
am losing energy you know what actually is happening it is very easy to understand so where the
energy is getting lost it is because suppose I have a signal let us say that signal is just tone okay
so it is just a cos w Mt okay.

So whenever I multiply this signal with another higher frequency signal okay let us say that is cos
w 0T once I multiply what happens basically it will be looking like this cos w MT into cos w0T
so this particular term becomes actually the amplitude of this cos w 0 T because cos w 0 T has
higher frequency so means if you see this one varying this will be varying at a very higher rate
while this one is varying this will vary at a very low rate.

So therefore actually you can interpret it as if this is varying and the amplitude is slowly varying
with respect to this one okay so eventually what will happen it is a cos w 0T only the amplitude
will have this envelope so it will look like this let us try it one more time it will look like some-
thing like this okay so what is happening now you can see earlier it was a cos w MT whatever
energy it was having or it might be just a square wave whatever it is whatever energy it was hav-
ing.

Now what is happening that particular part suppose if it was just a pulse train then after modula-
tion it will look like this inside that there will be this sinusoidal again this inside that there will be
cosine. So what is happening earlier it was almost keeping that same level of amplitude so the
energy if you integrate okay it will be much more whereas here there is oscillation so the energy
is getting degraded due to this modulation due to this oscillation okay which is also true. If you
just see the RMS value of cos that is always 1 by root 2 okay so the power if you wish to calcu-
late that is 1/2 so basically whenever you multiply its getting multiplied by 1/2 so whatever it is
the reason means there might be multiple reason that you can state but what happens always re-
member that whenever you multiply with a co sinusoidal or a sinusoidal your energy will be de-
graded it becomes half of the original signal energy.

198
That always happens okay so this is just to demonstrate the utility of or how to utilize our energy
spectral density whenever we do some modulation or some modification of original signal what
happens and how do you evaluate again the energy spectral density so it is just to give an exam-
ple there might be multiple such example and we can take all of them as homework okay so what
now we wish to do is we wish to do the similar thing for a power signal okay.

(Refer Slide Time: 11:40)

199
 1 T/2 2
T→∞ T ∫−T/2
Pg = lim g (t)dt

{0 | t | > T /2
g(t) ∣ t ∣ ≤ T /2
gT (t) =

lim gT (t) = g(t)

T→∞
1 T/2 2
T→∞ T ∫−T/2
Pg = lim gT (t)dt

1
= lim EgT
T→∞ T

1 ∞ ∞

T→∞ T ∫−∞ ( ∫−∞ )

2 2
= lim GT ( f ) df ∵ EgT = GT ( f ) df

2
∞ GT ( f )
∫−∞
= lim df
T→∞ T

So now let us say we have a power signal gt it might be periodic it might not be periodic whatev-
er it is power signal means first of all it is stretching from minus infinity to plus infinity and it has
energy that is infinite because it is stretching from minus infinity to plus infinity but it has a finite
power that means over a means if you just calculate the energy for the entire time duration from
minus infinity to plus infinity and divide by T you will always get a finite
value.

So it has a finite power but infinite energy so that characterizes for power signal or periodic sig-
nal or power signal okay so if I have a power signal okay now we have already stated that the
power must be calculated like this limit T tends to infinity 1/ T integration minus infinity sorry
minus T/2 +T /2 e2Tdt that means I am almost trying to evaluate the energy but divided by time
so what I am doing I am taking a valid T or finite T calculating the power So calculating the en-
ergy dividing by T and then all we are doing is stretching that limit to infinity okay so that was
our definition of power earlier we have done that ok if this is real signal if this is a complex sig-
nal then this should be mod gt square or g t into g star t so that same thing so this is all good but
now we should do suppose I have a power signal first thing I will do I will truncate that signal so
I have suppose a power signal which is stretching from minus infinity to plus infinity.

200
I first define a T and I define a signal g Tt which is defined as GT if T is or I should say mod T is
less than or equal to T/2 so if this is suppose 0 that s minus T /2 and thats plus T/ 2 so truncated
with centering at 0 and we are saying that G TT will be exactly equal to GT so it just carries the,
the property of GT from minus T /2 to plus T/ 2 rest of the places it is 0 so that is truncation so
that means the whole signal up to T I am just taking, rest of the things I am making 0 okay.

So that is my new definition of signal right so what we are trying to do again now you can see
means in Fourier transform we keep on doing that earlier we have derived Fourier series from
Fourier series we have actually taken it to Fourier transform because Fourier series was defined
for power signal or I should say periodic signal then from there we wanted to derive something
for time limited signal or the energy signal so we have again use the truncation property here also
we are almost doing that earlier we are defined energy spectral density so energy spectral density
was defined for a energy signal which is time-limited right.

So now we have got a time means unbounded signal so it is defined from minus infinity to plus
infinity so what we are trying to do we wish to employ the same thing again so we are now first
truncating the signal and then will put the same trick T tends to infinity right so this truncation if I
put T tends to infinity GTT goes to GT so if I just say GTT limit T tends to infinity by definition
it should be equal to g T so this particular thing I know but eventually what has happened the
GTT that I have defined that is actually now a energy signal because it is truncated so it must
have a finite energy as long as some of the portion are not going to infinity or some of the portion
are not blowing up ok, so as long as the signal is characterized by that which was also the charac-
teristics of GT so therefore I must always get a energy signal by doing that truncation okay so by
definition now my GTT is an energy signal right.

So now this p g I can define as limit T tends to infinity 1 by t minus T/ 2 to plus T/ 2 see this in-
tegration is defined from minus divided to plus T/2 so instead of G2t I can write also GTT square
that will give me same thing these two are equivalent according to my definition because GTT is
exactly equal to GT from minus T by 2 to plus T /2 my integration was always being done from
minus T /2 to plus T/ 2.

So therefore GTT if I integrate or GT if I integrate both will give the same result so that must be
Pg okay so if this is the case now you can see for this GTT this particular part because it is a
time-limited signal that must be the energy right this is the energy of that signal so I can immedi-

201
ately write limit T tends to infinity 1 by T this is I can write this as EGT that means energy of that
truncated signal, I can write this.

So Pg is nothing but this thing okay only thing is that earlier Pg was defined for GT now Pg is
defined for a truncated signal where I have the energy now the good part is this energy because it
is a truncated signal so it is time limited I can actually now relate this to energy spectral density
okay so I can immediately write this as limit T tends to infinity okay 1/T I can put this Eg T as if I
have this GTT that signal if I take a Fourier transform of that that should be GTF and if I take
mod of that and if I do a integration over full frequency minus infinity plus infinity that must be
my EGT right.

This is well-defined because I have now evaluating it for a energy signal so I should be able to
write this so immediately I can write instead of this minus infinity plus infinity modulus of GTF2
DF right so this is something I can always do fine if this is the case now this T has nothing to do
with this F right I can take all these things inside so I can write minus infinity plus infinity limit T
tends to infinity mod GTF whole square divided by T this whole thing into DF right.

So what I have got eventually the power is now related to this where I have this particular part
this I can call as the power spectral density because what is happening now similarly like the en-
ergy spectral density I can this is defined as of course is the energy spectral density so it should
be a spectral density so this if I integrate over entire f I actually get back my power so that must
be defined as power spectral density of my signal.

(Refer Slide Time: 20:02)

202
 2
GT ( f )
Sg( f ) = lim
T→∞ T
1 T/2
T→∞ T ∫−T
Rg(τ) = lim g(t)g(t + τ)dt
2
1 ∞
T→∞ T ∫−∞
= lim gT (t)gT (t + τ)dt

2
ψgT (τ) ⇔ Gτ ( f )
2
GT ( f )
Rg(τ) ⇔ lim
T→∞ T
Rg(τ) ⇔ Sg( f )

So I can now write that, suppose power spectral density is defined as SGF so, that I can write as
limit T tends to infinity mod GTF whole square divided by T right good so this is actually the de-

203
finition of a power spectral density so any signal GT if you are willing to define the power spec-
tral density what you do you truncate that over at T evaluate the Fourier transform get modular
square that means the energy spectral density of that and then you put that T tends to infinity that
energy spectral density divided by T. So that must be my energy spectral density by definition we
have seen that.

okay, and we have also proven that if Iintegrate this over F the entire f domain I will be getting
my power back right and that is why it is called power spectral density okay again it should be
related to autocorrelation function so what I wish to do now is I wish to define the autocorrelation
function for a power signal okay so that is r Gtau.

So what will happen because it is a power signal so just multiplying the signal and its delayed
version and then integrating will give me always infinity, okay. because that is very obvious if I
just put tau equal to zero immediately I get G2T integrated over minus infinity to plus infinity
that is already I know already that energy is infinity so there is no point in evaluating that I
should be evaluating it in terms of almost similar to power so what I will be doing again I will be
doing similar thing I will be putting limit T tends to infinity I will take 1 by T and then I will in-
tegrate minus T/2 to plus T/2 GT okay.

That is now that definition of time autocorrelation function for a power signal okay because for
power signal if I don't do this things I will always get infinite value okay so I have avoided this
but remember this is just a definition okay why we have defined this way because we will now
see that if I do a Fourier transform of this again similar to that previous one, We will get back our
power spectral density so this is what we will be trying to prove now okay so let us try to see how
do we define that okay so this is my RG tau okay I can put this as limit t tends to infinity 1/T so
this integration what I can do now is instead of minus T/ 2 to plus T/2 I can put this integration
from minus infinity to plus infinity provided I put the truncated signal over here ok instead of GT
because it is defined from minus infinity okay.

This integration allows me to do from minus T/2 to plus T/2 so if I just truncate that then I am
still limited within that t only okay so therefore I can always take this stretch this integration from
minus infinity plus infinity as long as I am applying this okay so this is good no problem in that

204
now let us see what is this, this is actually for the truncated signal time autocorrelation function
which is the energy signals we already know that okay.

So if this is the energy signal then Fourier transform of this one must be the energy spectral den-
sity so this entire thing has a Fourier transform pair or I can say they are paired by only this part
not limit T tends to infinity and one by T not that part only this part has a Fourier transform pair
which is GTF whole square I already know that therefore this RG tau must have a Fourier trans-
form because this T has nothing to do okay with it,. That with Fourier transform,Fourier trans-
form will not have that capital T okay so r G ta must have a Fourier transform pair which is noth-
ing but this will be as it is there only this part will be added so I will say it is a Fourier transform
pair remember it is not equal it is just limit T tends to infinity 1 by T GT f mod square so this part
is the Fourier transform of this one now what is this we have already proven that, that is sa F so
therefore r G tau and our defined s Gf are Fourier transform pair.

So we know that for a power signal now if we the way we have defined time autocorrelation
function if we do a Fourier transform of that, the time autocorrelation function will be always get-
ting that our power spectral density right so this is another very important result that we were try-
ing to get so what we have eventually done in these two classes, we could actually earlier we
have done Fourier transform but Fourier transform was little vague because there was no measur-
able quantity.

Now from that same part we could actually get a measurable quantity and we could relate that in
the frequency domain, time domain we had that measurable quantity we knew what is the energy
or power for corresponding energy signal and power signal but in the frequency domain we never
had that relationship. what we have done first we have proven that okay in the frequency domain
also we can get corresponding energy spectral density and power spectral density, which is noth-
ing but if it is a energy signal then all you have to do is take the Fourier transform of that signal
and then mod that square okay and we could also prove that if you integrate over frequency you
get the overall energy.

For the power signal we just have to do little bit extra, we need to truncate that signal do calculate
the energy spectrum and then whatever truncation value we have taken we have to divide by that

205
truncation and take limit that truncation parameter T tends to infinity that gives me the power
spectral density, We could again prove that if you integrate that we will be getting overall our PG
of that signal but then we have proven two detours for calculating energy spectral density and
power spectral density one was if we for an energy signal if we do auto correlation function the
way we have defined it, if we do autocorrelation function we know that once we calculate auto-
correlation function immediately if you take Fourier transform we will be getting your energy
spectral density. That is another way of doing it without doing the Fourier transform of the signal
directly you do the Fourier transform of the autocorrelation function of course you will be asking
why I am exerting myself to first calculate autocorrelation function then take that autocorrelation
function anyway again you have to do Fourier transform will clarify that in the next class but that
is another way of doing it okay.

So you first evaluate that autocorrelation function the way it has been defined for energy and
power signal both similar thing only the autocorrelation function definition is little bit different so
if you do autocorrelation function of the energy signal you will get corresponding energy spectral
density if you do Fourier transform if you do autocorrelation of the power signal, Then you will
get automatically if you do Fourier transform will be getting corresponding power spectral densi-
ty now the big question is why I need that second path but remember probably that is the most
essential part and most important part to evaluate any kind of realistic signal next we will give
one example without talking about random process you will probably be able to appreciate why
this is so much required. okay.

206
 NPTEL

NPTEL Online Certification Course

Course
On
Analog Communication

By
Prof.Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-14: PSD of Random Signal

Okay, so in the last class we have probably defined what is energy spectral density what is power
spectral density and we have devised way to evaluate energy spectral density and power spectral
density, so there was simple direct way you directly do the Fourier transform of the signal and
then take mod square of that you definitely get spectral density power spectral density just you
have to do a truncation and then do it okay.

Almost similar method but the problem with this let us try to see where we will have problems so
the first part is which is the most important part most of our signals are random in nature that
means the signals will not be deterministic so whenever we are doing this Fourier transform so
far whatever we have talked about it is all deterministic signal we have exactly defined it over
time that it should be at this particular time it should transit from 0 to 1 And remain 1 the come
to zero and so on.

So everything was completely deterministic the entire time means definition of the signal was
completely explained what will generally happen see the signals are important mostly for a
receiver, because the transmitter when they he is transmitting suppose I am trying to transit a
voice, my own voice to myself as no information okay it gives me nothing because I already
know I have created that voice so that will not give me any extra information.

207
Generally whenever we transmit signal the signals are important for the receiver because
receiver does not what I am going to transmit so for him that carries information and means of
course in this course we will not be doing but if you read little bit of information theory it will
know means more uncertainty you have in the signal more amount of information it is carried
which is very simply we can state suppose in the newspaper everyday you see that it is written
that sun rises in the east okay.

So that is almost known information we do not want that to be newspaper because that is
something you know so will that be a news that will be nonsense for you so any information
which is certainty or which will eventually occur and you are aware of that will never give you
information, but if you say suppose that means IS has bombed in let us say London okay, that
might be very important information for you okay whether important not important it is good bad
that is a different thing.

But that is a information because you do not see that happening everyday it is very uncertain to
you and that is suddenly happen and that particular thing we will give some information so for a
receiver part whichever it is we form communication prospective from general communication
like I have given that example of newspaper so always if you have some amount of uncertainty
or I should say some amount of randomness associated to the signal then only it carries
information.

So therefore all real signal if you see receiver perspective it should be random to him it should
not before and now there is no point in transmitting a sinusoidal to him because what does he get
extra from that it is just a sinusoidal if we gets a first period he known’s that it will be just keep
on repeating, so exact pattern of sinusoidal we knowing to him, so there will be no information
in first carrying out a sinusoidal if I just say a pulse strain okay so let us say it is like this.

(Refer Slide Time: 04:31)

208
 1 T/2
T→∞ T ∫−T/2
Rg(τ) = lim g(t)g(t − τ)dt
Tb
1 N 2
N→∞ N Tb ∫−N Tb
= lim g(t)g(t − τ)dt
2

(2 )
1 Tb
= lim N −τ
N→∞ N Tb

It becomes one then 0 and it is a pulse strain only so it is always 1 then 0, then 0 to 1 is that a
information for him it will never be information for him because he knows form the first pattern
if he observes this pattern he knows this pattern will be keep on repeating in time afterwards so
there is no information he will not be even interested in looking into that.

209
So where ever the receiver will be interested that means the signal that receiver will be interested
will be the most realistic signal that we wish to communicate and those signals from the receiver
perspective will be random so if receiver wish to do some processing on that it must be a random
signal which is not expected to him does not it should not have any pattern on it should not be a
deterministic signal for him.

So mostly whenever we say we are doing all this signal analysis so far we happen doing for
deterministic signal only for some purpose of course because we have to first understand that
deterministic signal analysis then only we will be able to appreciate the random signal analysis
so as long as the signal is deterministic so far whatever we are doing that is not very interesting
and we know that whatever tools so far we have devised like let us say Fourier transform so for a
deterministic signal.

There is a existence of Fourier transmit we can do Fourier because if you wish to do Fourier
transform that first thing you have to do is we have to take g(t)it is for –j to πft but you have to
integrate it from - ∞ to + ∞ that means if you wish to get the g(f) you need to know the exact
behavior g(t) over entire duration of time from -∞ to + ∞ this is absolutely required so if you
wish to do Fourier transform you need to know then signal completely complete definition of
signal is required to be there with you.

But now we are seeing that most of the realistic signals suppose receiver which you know what
kind of signal i have what are the frequency component it has but unfortunately the receiver
does not know about the signal pattern it does not if he knows the signal pattern he will not be
interested in receiving as we have already stated so this is the case then we have a dilemma, can
we now in show that or a realistic random signal okay which carry some information.

Where I dont know exactly what the pattern will be so the signal is not no long the deterministic
will can I take that other path of evaluating energy spectral density or power spectral density and
I still be able to get the spectrum component or spectrum density can I told okay so that will be
on next start, so we will just take one example, so very simple example is generally we transmit
seen in from here we are just now going towards the just to demonstrate it.

210
In a easier way we just going towards from analog communication to digital communication
okay, so in digital we will be knowing that what we do generally if we have even any digital
signal also comes from analog signal how it come. I will tell you so if we have a signal suppose
let say this is g(t) that just let say this is voice okay so what we generally do we first sample this
signal.

And then we note the amplitude of those signals and then we note what should be the maximum
swing of this signal .okay, so we know the maximum gmax or amplitude max that it can go up to
if I run a signal for the entire duration and the g mean of the signal if we know this range then we
know that the signal whatever sample will be take those samples will be within this.

Okay so once I know this range what I will do I will divide this entire range into smaller, smaller
ranges okay as when we have wish so that is how many I will be defining that depends on what
kind of things are we want design and then what I do, for each of this samples I want to see
which that smaller range it corresponds to and then for every range I give of equivalent digital
representation I give a equivalent digital representation okay so what happens suppose I have
this length I sub divide this.

Into let say 256 small, small ranges now this number 256 it is actually 28 so basically I know that
if I take 8 bits and if my signals are only binary can take only one or 0 then with that 8 bits I can
actually represent 256 different patterns okay so this is already known to you, so basically 0 0 all
8 zero’s will be 1 pattern 7 zero’s last 111 will be this next pattern and so on, we can represent
upto 256 such patterns okay so then what will happen where ever this sample stands I can
immediately say okay it is in this particular one.

That has a code book in the binary representation I will use that code book and I represent that
sample with code to like this every sample I will be able to represent in some binary
representation and if I just put them serially one after the other I will be getting a binary signal
stream, which in a way represent this one and later on when will be talking about sampling
theorem , PCM we will see that it actually faithfully represent the signal so you can actually
come there form that binary representation you can actually take out extract that entire signal the

211
quality signal of course there will be some noise associated with it will also discuss about those
things which is called Quantization noise.

Right now we should not be worried about that part but what we know is that we can for any
signal we can have equivalent binary representation or these days computer communication is
very famous so in computer communication what happens every key you press let say a it has
ASCI key representation and finally a binary representation again, so whatever we are
generating over there that also is directly generating some binary streams so let say those binary
streams that is the digital communication actually.

I wish to just transmitted those binary that means I will be transmitting either a high signal or
whichever way I represent that a high signal for let say if it is 1 and a low signal is be 0 and I do
not know have to represent any other it has some other advantage will discussed that later on
why digital communication is better than analog communication so right now we are not
bothered about that I just wish to say that okay, this is the valid communication which can also
transmit some analog data that is generated like this voice signal that we have generated so it can
represent those data also.

So let us say we represent this 1 and 0 in this fashion, so this is my 0 voltage level 1 is
represented by a pulse like this of duration Tb where for the first +Tb/2 or 1/2 duration it will be
having some positive voltage it can be 5 volt or something like that, okay and rest of the time it
returns back to 0, okay so this is called a return to 0 pulse we will later on if we have to do line
coding or in digital communication probably we will between line coding then you will be
knowing all these things.

But right now it is just the representation so 1 is represented by this and our corresponding 0 is
represented let us say so this is 1 and binary 0 is represented by this so it will for first Tb/2 or 1/2
cycle it will be negative of that same voltage so if this was +v this will be –v and the rest of the
part it will be 0 so this is by representation of 1 and 0.

Now let us think of our random sequence which actually carries information at the receiver so
that has this binary I mean random binary schemes that means 1s and 0s are all mixed it can be 1

212
it be 0 I do not have prior information, so it can be of any way so it can be a suppose first one is
1 then the next one is 1, next one is 0, next one is 1 again another 1 then one 0 so it can be of any
pattern you can see that the patterns are broken randomly okay.

So it is not a usual pattern it can be anything there might be a stream of five 1s and then followed
by seven 0s whatever it is all there, only thing is that any time you transfer this there is a
statistical property associated with it okay, so whenever we talk about random signal there
should be a statistical property. So the only statistical property we are just means exacting over
here is that it does not have all this bits does not have any correlation that means if this is 1 does
not have any correlation that the next one should be 1 or 0 okay.

So it can be with equal probability 1 or 0, and also we have equal distribution of 1s and 0 so that
means if I see observe this pulse it is almost like tossing a coin if we toss it for enough number of
time you will see probably head and tail number of that frequency relation number of heads and
number of tails will be almost equivalent here also same thing if I just stretch this pulse up to -8
to +8 we will see that as many numbers of 1 will be coming almost similar so there is no
biasness towards 1 or 0, okay.

So they are equal probability with probability half right, so these two statistical property is given
for this pulse so that means the pulse has a statistical definition or statistical property associated
with it when the pulse is no longer deterministic it can have any pattern okay, so this is
something I know. Now if I just say can you evaluate the Fourier transform of this not possible
because I do not know it is a ransom pulse I only know about the statistical characteristics that if
this one is 1 I know that the next with equal probability might be 1 might be 0.

But for sure I do not know, and also when this how this 1s and 0s are arranged I have no
information so if I just say okay, I need to have in the receiver side I need to know what are the
frequency component it has for that I need to evaluate the power spectral to density of this one
but if I say okay give me the power spectral density will not be able to evaluate because the first
stumbling block will be you would not be able to do Fourier transform of this particular to see.

213
So if you cannot Fourier transform you cannot even evaluate mod g(f)2 if this is g(t) mod g(f)2 is
not evaluated, can I now take the other way I have defined to actually get a Fourier it means
power spectral density of this one. So this will be our next target, so for auto correlation function
what I need to do of course I can see this is a power signal because it is stretches for - ∞ to + ∞
okay. I will do truncation and all those things I will do that but it is a power signal.

Because I am not saying that this 1s and 0s are up starting from here and ends over there I am
just saying okay, it is a continuous stream of bits which are going on and see the time infinity
also you must always argue that why we are doing this, why we are calculating I mean treating
this as a power signal and why we are evaluating power spectral density you might be always
asking because any signal actually starts at a finite time and it will end at a finite time no signals
are actually if it is not that the signals ready sorry, radio waves that are transmitted from big band
and it keeps on coming.

If that is not the signal we know all are the signals man generate a signals if you speak you are
not speaking or talking from minus infinity and you will not be you cannot even survive up to
plus infinity so it will never happen it will always be time bounded then it should be all energy
signal and then we should do energy spectral density only but the counter argument is that the
receiver will be seen the signal for a very small duration and within that you wish to know okay
what is the power and all those things.

The time duration he is evaluating it compare to that the signal the time duration for which the
signal lost is much bigger so in that perspective or in that relative sense you can always say that
this is almost like a plus infinity and minus infinity to me because when I am trying the evaluate
the signal probably compare to that the signal existed much bigger time so I can always treat that
as if it stretches to minus infinity and plus infinity and I can treat that as a power signal.

So that is why most of the time you will be actually going towards power spectral density rather
than in a spectral density okay. So anyway those are the motivation why we are doing this so we
wish to so this is a power signal for to us so all we wish to do is we wish to calculate the time
auto co relation function right so that is our target. So that means what we have to do we have to

214
shift this signal little bit slightly and then we have to multiply so let us say we have shifted by a
part T(Tau) okay.

So what will happen? If as long as this t is less than this tb /2 tb / 2 is the duration or tb is the
overall bit duration so as long as it is less than all is than equal to this tb/2 I know that there will
be some over lapping and each pulse will have same overlap. Because it is not going beyond that
pulse duration so it should be if I just shift by only that t which is less that tb/ 2 it still have
similar overlap at every pulse so this pulse will also have similar overlap so that is the overlap
region this is the overlap region and same thing will be happening over here.

That is the overlap region and so on so all shifted by t right. Now let us say because it is power
signal I need to Truncate it get it so let us say I have Truncate it up to t, and I have chosen by for
these of calculation I have chosen by t in such a way that some n number of pulses exactly will
sit within this okay so I will get my t is defined in such a way that I will get exactly suppose N
number of pulse okay.

So I am free to select any t there is no problem in that and we also know as t tens to infinity this
n will also go to infinity infinite number of bits will be coming so infinite number of such pulses
will be coming okay. So now my auto co relation function the way we have defined it, it should
be because it is a power signal so it should be limit takes in to infinity 1/t okay and then we have
to integrate right from –t/2 to + t/2 and then we will be doing this g(t)g(t-T) dt right this is what
we will be doing.

Now what is my t now t is n x tb because there are exactly n number of pulse so I can just re
write this n tens to infinity 1/ because t tens to infinity means N tens to infinity instead t I can
write N tb integration this goes –NTb /2 to + NTb/ 2 and then g(t) g(t-T) dt right I can write it
this way, now let us say what is the overlap that we are getting so I have already assume that t is
less than tb/2 right.

215
So by overlapping area is this was tb / 2 – T if you take so it must be if the amplitude level is +1
and -1 taken as that so then my overlapping area should be tb/2 – T that is the overlapping area
right because the amplitude is 1 so it will be just multiplied by 1 will get this value all the right.
How many such pulses I will be getting N number of so this whole integration I can actually
write as n x this area right so I can write limit intends to infinity 1/NTb nx tb /2 – T I can writer it
this way.

(Refer Slide Time: 23:48)

216
 Tb ( 2 )
1 Tb
= −τ

2( Tb )
1 2τ
Rg(τ) = 1−

(1 − )
1 2|τ| Tb
2 Tb
|τ | ≤ 2
Rg(τ) =
Tb
0 |τ | > 2

( 2 )
Tb π f Tb
Sg( f ) = sinc2
4

So now what happens n gets cancelled so my limit is no long that required because nothing else
is there which is putting limit on N. So therefore I get this = Tb/ 2 – T and there is 1/ Tb or I can
just rearrange it, ½ 1- 2T/ Tb right, I can write this. So this becomes my RG T which is nothing
but a linear function, starts from 0.5 and goes to 0 okay, so goes to 0 at Tb/ 2 and because it is a
RG T we have already proven that it is a even symmetric function, so it should be exactly like
this upto - Tb/ 2 that something we know, so we can write RG T should be ½ 1- 2|T|/ Tb as long
as |T| is less than Tb/2 good upto this we have the value.

But what about T can take any value, if now I wish to take T > Tb /2, so if I just see this will
happen, if T > Tb /2 this will shifted beyond this pulse. So his corresponding pulse will be
shifted beyond the corresponding pulse. So it will now have overlapping with the next pulse or
next to next pulse whatever it is. Whatever happens it is not his own pulse, it will be next pulse.
Now there is exactly half probability that will be same.

That means 1 will overlapping with 1 or 0 will overlapping with 0 because the 1 and 0 first of all
not correlated and equally probably, so there is exactly half probability or I should say ¼
probability that both will be one, so if I just add them half probability then both will be either 1
or both 0. In both cases over lapping gives me positive value and there is another half probability
that they will be actually, one will be overlapping with 0, so that they will create a negative
value, or 0 will be overlapping with 1.

There will be creating a negative value, so because there will be infinite number of such pulses,
so if I just all those things has to be added okay. So what is happening? Half of them will give

217
me positive values because they are equally likely half of them will give negative value. It will
get cancelled and I will get 0 all the time. So this becomes 0 all the time as long as my T > Tb /2.
This is clear because the statically definition of the signal was earlier given. Though the signal
was not deterministic but this time auto correlation I could evaluate because of the signal already
that has been given.

So from there I immediately get a time auto correlation function which is RG T, now this is the
deterministic signal, now I can do Fourier transform it is very easy, it becomes because it is a
triangular one it is sinc square okay, so it happens to be this where it is actually g(f) is Tb/ 4, so
you take this and transform will see that, that becomes sinc 2 of (πf Tb)/2 and that is the pattern.
So immediately you can see that I had random pulse which is going to the receiver.

Now receiver has to know the frequency components because accordingly he can designed these
things like he use to put some fit there. You might be asking why the filtering is required there
might be signal also, so he as to separate them out not only that, even if there are no other signals
you will later on see whenever we will be talking about noise, noise remains everywhere in
entire band. If I just do not filter out my signal then if I put all pass filter I will get my signal.

But I also get my entire noise from the entire band, should be the signal, so I wish to reduce the
noise then I should put up the filter so that out of noise are surpassed. So for that reason also I
need to know where the signal is existed or which frequency band my signal is existed, so for
that I need to first know even if the signal is random I need to know what is the shape of that
signal that for a random signal, I cannot imply the step method of energy or prospective.

So I have to imply the second method that I have proven like through, so that chest tells us
probably the 2nd method is much more important than the first method because we have already
explained that a signal which is important to the receiver must be a random signal, and for the
random signal it is very difficult direct Fourier transform. Rather auto correlation if you know
some statistical property that is pretty much possible okay. So we will end today class over here,
next class we will probably we will start discussing about modulation.

218
 NPTEL

NPTEL Online Certification Course

Course
On
Analog Communication

By
Prof.Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-15: Amplitude Modulation

Okay so far we have already discussed about some aspect of signals like Fourier series, Fourier
transform and then energy spectral density, power spectral density, what is the significance of
them? Given one example why autocorrelation function is important to evaluate energy or power
spectral density. So this is something we have already covered, now let us see means so far there
were some touches where we could talk about communication but we have not gone inside the
communication aspect. So today what we will try to do actually try to see how this understanding
of signal can make us means to communicate okay or can aid us to communicate. So this is
something which we will be targeting today.

so if we start seeing communication as we have discussed earlier there are two aspects of
communication, so one is of course the signal, which is the source means that is the thing that has
to be communicated and the other part is the system, which somehow is a transmitter receiver in
between a channel, so either the transmitter or receiver processes the signal. So this system
receiver or transmitter are actually a system which processes the signal towards a desired output,
so that the signal can be transmitted faithfully at the receiver side and all the information can be
faithfully transmitted.

So this is the purpose of the system designed right, so system design has some purpose because it
operates on a signal, so any system that we are targeting must serve those purpose. So for a
communication system the most important part as we have so far explored is probably the

219
modulation and demodulation, so this is the first thing we will be exploring that how to devise a
system does modulation and demodulation add transmitter and receiver respectively okay.

So if you just start seeing communication there are in a way two types of communication.

(Refer Slide Time: 02:31)

One is called baseband communication and the second is carrier communication, okay so what do
we mean by these two? So baseband communication means we actually transmit it as a low-pass
signal, okay so whenever we are transmitting if in the frequency domain we try to see it will
occupy some low pass band. So that is generally called baseband communication, so why it is
called baseband communication generally the signals that we generate the raw form signals, let us
say we generate voice signal through a transducer, so basically we talk the transducer converts it
into an electrical signal.

220
If we just try to see how that electrical signal looks like we will see for voice signal it occupies a
band starting from 300 Hz towards 3.4 kHz, so it occupies some band like this and of course
because it is a spectrum and it's represented by Fourier transform form so it should be having
equivalent negative part, so - 300 to - 3.4, so this is typical voice signal that looks like if you if
you have the most two most popular signals probably where one was voice signal and the other
one was which was popularized by radio transmission okay. So and the other one was probably
the video signals okay.

So if I just try to see the video signal that has a higher band that means again there should be a
video transducer which converts the video signal into equivalent electrical signal, so what it does
actually if you have a screen so every second there are few frames, so frame represent the entire
screen and then you have pixels on the screen all those pixels will have their own grayscale value
if it is a black and white if it is colored then there will be three composite colors and each color
will have their own value gray scale value again okay. So those signals has to be means if I just
put them on time, so for each frame we just one after another pixel value we put we get a signal
and that particular signal if I just represent it in electrical domain the way I have told. So
basically if you have a screen you scan them pixel by pixel from this direction left to right again
you do for the next line and so on you do all lines every pixel value you put a voltage next pixel
value you put another voltage and you put it on time so this is overall frame if you do that will be
one frame followed by another frame and another frame so on, so you get a signal equivalent
electrical signal, if you just again do our Fourier transform of that and you represent it in
frequency domain you will see that it occupies up to 4.5 Mh, okay that's quite big of course the
video signal has a higher bandwidth it occupies the higher bandwidth.

So this is the raw video signal as you can see these are all low-pass equivalent signals okay, so it
is centered around zero and it through a low-pass filter you can always extract that signal okay, so
low-pass filter bandwidth has to be accordingly adjusted. So these are the signals which are
means if raw signal I wish to transmit that will be termed as baseband signal okay, so if I wish to
transmit that should be a baseband signal, that is about baseband signal. Even telephony which is
probably we have already talked about telephony, so what happens? There also we transmit voice,
voice but the voice is encoded.

So basically if I in the last class we were demonstrating that the voice I will be sampling each
sample will be represented by binary bits and if we just transmit that binary bits we have already

221
seen that a random sequence of binary bits if I just see the spectrum through autocorrelation we
have drawn the spectrum autocorrelation was something like this and the corresponding spectrum
was sin2 right.

(Refer Slide Time: 07:18)

So it was like this again a if you just see it is a low-pass equivalent signal, so these are all
baseband signal if I wish to transmit them RAW this would be baseband signals, so and the
corresponding communication will be termed as baseband communication. So I don't do any
modulation over here that means I do not put it in any higher frequency, so as long as I am not
doing that that particular communication is termed as baseband communication.

If instead we have already talked about frequency division multiplexing that means in frequency
domain I wish to put those signals which are band limited in different different frequency
location okay. So that I can multiplex multiple signals and simultaneously transmit them they are

222
in time domain they will be overlapping okay but in frequency domain there will be non
overlapping and I can always employ a band pass filter to extract my own desired signal.

So basically a meaningful transmission can be or simultaneously multiple meaningful

transmission can go through the channel, it can utilize the same channel and I can transmit
multiple signals. So thats, that is a basis or the basic requirements for doing a modulation okay,
we have also discussed that there is another reason why we should do modulation?

That was typically for radio transmission, so if we wish to broadcast the signal that has to be
received by a antenna and the antenna size has to be comparable to the wavelength of a signal, so
this is something we have already discussed that if I wish to do that the frequency has to be
higher, so I should instead of if I wish to transmit baseband signal antenna size will be very big,
instead of doing that same signal I can actually modulate, it will serve two purpose first of all
because it goes to higher frequency my actual antenna which catches the signal will be much
smaller and then after that I can employ a low pass filter to extract that signal and then do a
demodulation to get that signal again back to baseband.

Remember the signal is by itself is defined in baseband, so if I wish to again either listen to that
voice or watch that television I need to convert that to the baseband, so my demodulation actually
means if I do a modulation then de- modulation actually means again from that pass band I have
to bring that signal to the base band. So this is all that we mean by demodulation so definitely
now you can see that there is a purpose, two purpose probably, one is for multiplexing, another
one is for reducing the antenna size, we need to do modulation and that is when the first hurdle of
communication comes into picture or first design of.system comes into picture. That I have to
now devise a system which modulates the signal that means does the frequency shifting and I
have to design a signal which demodulates our system which demodulates the signal that is
already modulated signal, I can catch through antenna after doing low band pass filtering I need
to demodulate it to bring it back to baseband, so that I can put it to the transducer and again either
listen or watch whatever I wish to do okay. So let us try to see what this modulation means and
why we are saying that to be a carrier modulation okay or carrier communication.

(Refer Slide Time: 10:47)

223
S(t) = A(t)cos (ωc(t)t + φ(t))
S(t) = m(t)cos (ωct)

So let us say whenever we are talking about carrier I have a signal I am just representing a
sinusoidal signal, so sinusoidal or co sinusoidal it must have a cause and it has actually three
parameters. One is the amplitude of this co sinusoidal which I can write as A, the other one is the
frequency of this co sinusoidal which I can write as WC or 2pi FC whichever way I wish to write
into t plus there should be a phase right, let us call that pi. So this is a sinusoidal at frequency WC
and generally what we say because this is at some frequency we call that a carrier.

Why it is carrier? That will be clear because it does not by itself carry any message or any
information, it somehow modulates the message that it carries, so that is why it is a carrier it is
almost like a envelope that carries your letter okay, so it just helps you to carry the letter it has no
importance just it has to carry the, carry the thing that you are trying to transmit just delivers it to

224
the destination and after that it has no use usage. So you will be again demodulating so if we get
getting rid of this carrier okay. So right now we can see there are three terms in the carrier, so
while modulating I can use one of these three to carry my signal okay.

So suppose I have a message okay what I can do I can either make this amplitude time varying
which is proportional to my message signal okay, so that way also I can carry my message signal
on top of this one or I can vary the frequency with respect to time which is proportional to my
message signal again over a frequency I am carrying the signal or I can vary the phase,
proportional to my message signal and accordingly there will be different kinds of modulation, so
if I carry it in my amplitude that is called amplitude modulation AM, if I carry it through my
frequency so that is called frequency modulation or popularly termed as FM and if I carry it
through my phase and that is called phase modulation or PM.

So accordingly there will be three kinds of different kinds of modulation which has their own
advantage disadvantage, they will have their own types of system designing will see those things
so we will try to explore these things, amplitude modulation, frequency modulation and phase
modulation, so these are the three predominant analog form of modulation that exist. Of course it
is going over a sinusoidal carrier okay.

So the carrier is there and I wish to do some modulation so initially for the beginning of our
course probably will be more concerned about amplitude modulation later on will explore the
other things, which are these two are again termed as angle modulation because overall this can
be termed as a phase of a co-sinusoidal okay it has a frequency part, it has a phase part. So
overall I can talk about with the T in it we can talk about that's the overall time varying phase or
the angle okay, so that is why it is called angle modulation or I should not write AM then it will
be confused with amplitude modulation. So it is actually angle modulation okay, so initially we'll
be concerned about the amplitude modulation. So let us try to see the most basic form of
amplitude modulation, so we have told that in amplitude modulation my message let us say I
have a message signal which looks like this okay, so some message signal this looks like this it is
m t of course it might have random variation as we have already termed about it that if it is a
voice signal for the receiver it must be a random signal.

So anyway it is some kind of signal okay which is time varying and the time variation might be
random and on top of that we also have another requirements that this must be band-limited that
means if I take the Fourier transform, how we will be taking Fourier transform? Again will be
probably knowing the statistical property of this and we will do our time autocorrelation function
and we will do a Fourier transform so we will get some Fourier transform, let us say this is the

225
Fourier transform of this one our requirement is this Fourier transform should be essentially band
limited okay.

So that means the frequency terms that are existent for this signal are essentially within certain
band or otherwise we can say if it is not band limited the out of the band or whichever band we
are targeting beyond this there might be some frequency component but those are insignificant
okay, so we do not have to be bothered about that whether that is transmitted or not that will not
distort the signal beyond something beyond our perception.

So we will say that this is band limited so this goes from - B to + B okay right. So I have a signal
now which is a random signal probably which has a equivalent Fourier transform representation
or I should say a power spectral density I have okay. So these two things are there or I can always
say it has a means MF correspondingly it will have mod MF2 which is either energy spectral
density or power spectral density depending on how you define it okay, so we are just right now
thinking about that we have a signal. And we have corresponding amplitude spectrum that is the
simplest way to represent it just other things can be done later on okay. So these two things are
known suppose for a signal, now we have to design a system which does modulation so what we
need to do the first way of doing modulation is frequency using frequency shifting property of
Fourier transform.

So if I have mt we have already told that it should be carried by the carrier so mt must come over
here, so my modulation must look like mt cos w CT + some pi constant pi , which must not be
time varying because we are just putting the message in the amplitude not in frequency or phase
that should be remaining constant. So even for simplicity we can take that as zero even if you
take a constant value that will not disturb you, so we can say this is my modulated signal.

So here the terminology will be this is my carrier this is my modulating signal or the message
signal and this is the modulated signal, as you can now see we are familiar with this already.

(Refer Slide Time: 18:06)

226
 S(t) = m(t)cos (ωct)
1
2[
S( f ) = M ( f + fc) + M ( f − fc)]

So if St becomes mt cos Wct immediately if I do Fourier transform of this so Sf should be we

have already derived this that the famous frequency shifting property, so that becomes M f +
suppose this w c is 2 pi fc and should be f + fc + M f - fc right, we are saying this is our
amplitude modulation. So how this the spectrum will look like so if I was having M f like this
wherever I modulate that goes around fc and - fc and the corresponding strength becomes 1/2 so
if this is like this will be half of that amplitude will be half of that, but the shape will remain the
same so that is the mf- fc and this is the mf+ fc and you can see after modulation because it is a
real signal.So it is a even symmetric representation okay,

so the conditions are, that this f – e whenever I modulate there are two things which are
happening, what was the bandwidth of this signal? The bandwidth was, it is up to B right, so the
overall bandwidth is zero to B and 0 to-B negative half I can ignore to define bandwidth, so it is

227
actually bandwidth is B so the baseband signal has bandwidth B whenever I do modulation what
happens because the shape remains the same this should be fc - B and this should be fc + B. So
overall bandwidth will be doubled that is called 2B okay and what happens if you see this signal
was even symmetric.

So this particular half and this particular half, this half does not carry any extra information as
long as I know the pattern of this half that will be repeated in the just in as a mirror image in the
other half also okay, so basically the information is contained in one half or within if I say just a
baseband then within that B bandwidth what is happening after modulation my band width is
getting increased. now the overall band width if you see it is defined from fc - fB to fc + fB, just
if I see the positive half, negative half we do not have to see for bandwidth calculation so that
becomes 2 B and what is happening, this is still even symmetric.

So this is symmetric and plus around this because the shape remains the same around this point
fc, it is still even symmetric or this is symmetric so still the information is carried by this
particular band only which we term as upper sideband and this particular thing we call as lower
sideband and upper sideband and lower side band are just mirror image to each other, so they do
not carry extra information whatever upper side band carries same information about that signal
is being carried by lower side band.

But unfortunately whenever I do modulation,I will be requiring this entire 2B band to transmit
my signal. So what is happening eventually my signal bandwidth is B but after doing modulation
I am occupying at least 2B band and that's being occupied by me, my own signal. So in a way
there is a band width wise if we start asking there is a wastage or I should say double wastage
because for every B band width signal I am occupying 2B band width, later on we will see if we
can have some band width efficient modulation technique. But right now we should be content
with that, that is the simplest form of modulation I just do this I get this particular thing I know
that band width wise it is not the most efficient one and of course, we will be requiring this
because band width is a commodity you can see that if you have some sudden band to operate
you will be putting your signals one after another and they should be non overlapping if they are
band limited.

228
So basically, how many signal you can put within that band will be restricted by what is the band
it is occupying, so therefore if you can reduce the band of course, it will be occupying more
number of signals that means more number of simultaneous transmission you can do. So here
what will be happening if you just do this modulation as many for us for same band suppose
where you can operate if you can operate with n number of means you can transmit n number of
simultaneous communication, what will happen? If I could have reduced this to half I could have
transmitted twice that number that's a big number suppose I can transmit 6 radio channel over a
band now I can say I can transmit 12 radio channel within that same band.

So cost-wise, that will be much more effective we will be definitely targeting that, but right now
we should be contained with this modulation, so this is this particular modulation is called double
sideband with suppressed carrier we will talk about why it is suppressed area, right now it is
double sideband that is really understandable because we are transmitting both the side band,
upper as well as lower sideband. When we are see after doing modulation we will be putting it in
the antenna now it is actually serving me two purpose one is it has been translated to a higher
frequency, so the antenna size will be smaller and it is occupying a particular band so I can
actually multiplex, if I was trying to transmit raw base band signal to Y signal will be occupying
the same baseband and we won't be able to multiplex them. So two purpose are being served.

Another thing that has to be seen is, what should be my fc? very carefully see, if I have chosen
my fc over here okay, so by signal whenever I modulate g it will be left shifted and right shifted,
so what will happen this two things after left shifting and right shifting they will overlap with
each other and in the frequency domain you cannot actually separate them, now you cannot
employ any filtering technique to separate them out okay.

So what will happen? Your signal will have some distortion because it will, this gets overlapped,
so it will look like something like this which is a completely distorted version if I just, so it still
remains a low-pass signal which has a frequency. Suppose I am centered at fc so it goes up to fc +
B and it is of course - fc -B right, so now,I cannot employ up after doing this modulation I cannot
employ any kind of band pass filtering, so I will have to employ a low-pass filtering of bandwidth
fc + B okay and that will distort the signal.

229
if I now put it in the means of course it is already based band, so if I just try to see the time
domain signal it will be distorted because the frequency for components has been distorted, so
signal will be distorted. So therefore if I don't wish to have Distortion what I need to do is, at
least I should ensure, that whenever I put them they are separated.

So that means my fc, this fc, must be bigger than the B and generally, it should be much much
bigger than the B, so that is some criteria which has to be satisfied, whatever modulation you do
wherever you wish to put, but you have to always ensure that whatever frequency you are
choosing for your carrier that must be greater than the bandwidth of the signal, and that is why it
is very essential that you know the Fourier or energy spectral density of your signal you are
modulating voice whichever way you are modulating because it has a much lower bandwidth, if
you employ the same thing for video probably it will not work because video is occupying a
much bigger baseband frequency. So you have to accordingly see where I can put which
particular frequency I can choose and put my video signals that is very essential.

Whenever you are doing modulation that is one part which is essential and also you have to see
what frequency you are choosing because according to your antenna size will be decided, so both
the things will actually determine what frequency you should choose okay. Now let us see what
happens to this DSPSC.

(Refer Slide Time: 27:12)

230
m(t)cos (ωct)

So basically what has happened,I have a signal mt for modulating what I have to do? I need to
multiply it with cos w Ct so basically from signal I am now trying to devise a system okay, so
what I need is a multiplier circuit somehow I have to realize this multiplication. A circuit which
where I will be giving two input one is my message signal itself or modulating signal Mt and I
will be giving a carrier signal and somehow my system should multiply this, will see how this
multiplication can be done and then the output will be mt cos w Ct, for my purpose this is my
modulation.

So if I can realize a multiplier circuit that makes my modulator system ready and then I just put it
in the antenna and it will be transmitting this particular signal because this characterizes what I
wish to achieve in modulation okay. In time domain what will be happening, so suppose I had a
signal which I have shown earlier here if I wish to modulate it, so basically what will be
happening I am modulating with the carrier frequency which is cos w Ct as you can see the time

231
varying of amplitude which is much slower than this time variation, because it is at very high
frequency okay.

So that will be happening now if I just multiply this the amplitude of this one, this carrier
frequency must carry this signal, so this must be reflected in the amplitude so what I will see is
something like this, this should make my envelope of the carrier so if I just envelope means it
should be from both sides. So that draws the envelope within that the carrier will be oscillating,
so it should if this is 0 sorry this is 0 let me draw it again with something happening over here we
will discuss that in the next class but this is what will be in the time domain okay, where in the
envelope the message will be carried over.

232
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 16: Amplitude Modulation (Contd.)

Okay so in the last class whatever we were describing is we are trying to show that if we do a
simplest form of modulation which we are calling as DSBSC that DSB part is clear probably
SC is still not clear that will be clear after some discussion some amount of discussion but
that particular modulation which is the simplest of amplitude modulation where just we
multiply a signal m(t) with cos ωct so we have shown the frequency domain response of that.

(Refer Slide Time: 00:55)

233
So basically it will be just centered around fc and that same criteria will be coming that this
will be M(f – fc) this will be M(f + fc) and of course with half, so it will be centered around
fc, so this is the frequency domain representation and the corresponding time domain
representation we have understood from this particular thing that it is carriers whose
amplitude is instantaneously varying with respect to the message signal which is a slow
varying signal according to our assumption.

Here also that assumption comes into picture because fc has to be much much bigger than B
which is the highest frequency component that the message signal has, so therefore that will
be much slower varying signal and amplitude of that carrier will be slowly varying
accordingly, so that means that m(t) should be observable in the envelope of the signal that is
being transmitted so it is a co sinusoidal signal only the envelope will be tracing that m(t)
okay.

So what is happening there is a possibility that m(t) might be positive or m(t) might be
negative like over here, so this is the crossover point right beyond this m(t) is positive just
after this m(t) will be negative so what will happen in the carrier, so the carrier that cos is
going up to this point it is positive immediately after this point will be negative so in the
frequency means in the phase of the carrier what will happen from positive to negative
immediately jumps.

So there should be a 180 phase shift immediately right because the amplitude is yet just
getting means from positive to negative, so therefore in the carrier phase there should be 180
phase shift so wherever there will be 0 crossing in this DSBSC in time domain if you just
observe the signal you will always see 180 phase reversal whichever way whichever phase it
cuts over there it will just be 180 phase reversal.

So every point you will be getting that that is a huge implication we will see that that makes
the demodulation very costly will come across that and that is where the suppressed carrier
and non suppressed carrier comes into picture we will come to those things but we have from
the time domain we just try to analyze whatever is happening okay so there is we know as
many times there will be this zero crossover that many 180 phase the word cell will be
happening if this was also having a pattern like this there also we would have observed
similar things.

So it will be like this the envelop will be just like this and again if things are coming out like
this there will be a 180 phase reversal immediately okay, so as many 0 crossover will be
happening that many phase reversal will be happening, so we have understood what is the
modulation right what will be corresponding demodulation that means I will be getting
receiving this m(t) cos ωct signals from the air through my antenna. Now I have to get back

234
my m(t) okay so how do I get that very simple operation another multiplication will give me
that.

(Refer Slide Time: 04:20)

ϕ(t) = m(t)cos (ωct) cos (ωct)

1
= m(t)[2 cos2 (ωct)]
2
1
= m(t)[cos (2ωct) + 1]
2
1 1
= m(t) + m(t)cos (2ωct)
2 2
1 1
Φ( f ) = M( f ) + [M ( f + 2fc) + M ( f − 2fc)]
2 4

If you just see m(t)cos ωct I am getting if I just multiply it with another cos ωct locally
generated what will happen, so this is if I just take ½ m(t) this is 2 cos2 ωct cos2 ωct can be
written as cos 2ωct right we can just write this immediately what we get we separate this 2
out I get ½ m(t) + ½ m(t) cos2 ωct fine this is what I get now just see the signal now you will

235
appreciate why this Fourier transform and all those things were so important why time
domain and frequency domain we have to observe the signal.

Let us try to do a Fourier transform of this signal of this composite signal let us call this as ?
(t) and then I wish to evaluate ? (f) what is ? (f) so I have ½ over here so that should be ½
m(t) the Fourier transform should be M(f) and this is multiplied by cos 2 ωct so therefore
there should be a frequency shifting property so I can write ½ and then this particular part
should take me to cos f + fc sorry of course m f+ fc + m f- fc right this is fine okay.

So I will get that particular thing but there should be a ½ also coming out of this one because
this whenever I multiply by cos there should be another ½. So this must become ¼ right sorry
there is 2 ωc so it should be 2fc okay fine, so if I now just plot this frequency response.

(Refer Slide Time: 06:38)

1 1
M( f ) ⇔ m(t)
2 2
How it will look like so this is suppose my M(f) was something like this so this is M(f) ½
M(f) so this is that ½ M(f) and at 2fc there will be ¼ M(f) – fc and at – 2fc there should be ¼
of M(f + fc) fine to exit a fine. Now it is very simple you can see that just by putting a filter I
can extract my signal back I will be putting a low-pass filter over here which has a cutoff

236
frequency which is much lower than this 2 fc but bigger than this B so I just employ a filter
like that I will be getting my ½ M(f) back which has a Fourier inverse transform which is ½
m(t) so I will be getting this signal back so immediately you can see my demodulated circuit
is almost ready what I need to do in demodulators.

Whatever modulated signal m(t) cos ωct I will be getting I will do another multiplication with
cos ωct whatever I get I pass it through a low-pass filter even the filter design is also known
the cutoff frequency must be suppose the cutoff frequency fc cutoff okay so that must be
bigger than B we should not say much bigger than because B but that must be lesser than this
2fc at least as long as I am doing that I will be getting my signal ½ m(t) over here because
this part will be rejected.

So very nice we could realize see this is the importance of signal and from signal to system
that is why probably we have devoted so much time in doing frequency response Fourier
transform how to see a signal infrequency domain so those are the things just giving us tools
enough tools to actually manipulate the signal and then understand what kind of systems I
should put so that whatever I wish I will be able to achieve that so basically your modulator
and demodulator is now ready for this DSB- SC of course we still have not characterized
what this multiplier circuit will be how do I achieve a multiplication.

Okay but we have now understood that if I know how multiplication has to be done I need a
multiplier circuit followed by a low-pass filter that will make my demodulator I need in the
modulator I need just a multiplier circuit nothing else okay, so now let us try to see what are
the difficulties over here the first difficulty that comes out is generating this one that is a big
challenge because if you see very carefully the frequency and of course the phase also we are
not writing the phase these two has to completely match over here.

Then only that cos 2 ωc will be coming out and then only the Fourier transform will give me
very nicely m(t) later on will prove that if there is a phase drift or frequency drift and if we
multiply these two I have a chance of not getting anything over there okay so it is we have to
be very cautious about generating this local sinusoidal that has to be completely in sync with
the carriers by which the signal has originally modulated the problem is that nobody will give
me that signal right.

Because if the modulator and demodulator are sitting at the same place what is the point in
doing communication because when we wish to do communication we wish to transmit it
over a longer distance if I know that same carrier I can give in both places from the same
circuit then probably my modulator and demodulator are already sitting in the same place or
otherwise I have to separately again communicate the carrier as well right to a long distance
so just to send my message over a carrier I have to again transmit my carrier along with it.

237
That is one difficulty okay the second difficulty is even if I try to send my carrier there is a
possibility that this modulated signal and the carrier because they are going through the
channel they might go through some frequency and phase drift it might happen due to
Doppler effect due to other effects means that are there due to the channel effect mostly, so
there will be a drift in phase as well as frequency of the carrier signal and that will be random
as well as the modulating signal.

So if I wish to really means even if I have that carrier and I am transmitting along with that
there is a possibility that these two are going through different phase and frequency drift and
at the end they are not in sync in terms of frequency and phase again if I multiply I will not
get the proper representation, so what I have to generally do that this particular signal that I
am getting already I know that it already has of course are contaminated sinusoidal can I
extract that carrier out of this.

If I can do that from there if I can generate this cos ωc then I am fine so that is called a carrier
recovery there is a means there will be in this course only there will be a few classes devoted
towards that that how do we do that carrier recovery that is a big circuit again it must be
locked with the incoming frequency and phase that is termed as phase lock loop will see
those circuitry but that is the part which is required otherwise your demodulation will not be
good.

So that is the difficulty we are having in this particular modulation scheme that we need to
have another carrier which is completely in synchronism with the incoming carrier frequency
and phase this is the one difficulty that we have that means the receiver design becomes little
more complicated because we have to do this carrier recovery on top of this whole thing right
so if I now ask when we are designing a system is this desirable for let us say a broadcasting
kind of system okay.

Broadcasting means like the radio transmission we had those big authority and all other
things where we used to just transmit something from a big antenna okay and that was
broadcasted to everybody was listening to same voice okay, so this was broadcasted and
everybody must have their own receiver and they must detect it okay in that kind of thing I
can actually make my transmit a little costly because that is common to everybody that cost
will be shared among all the users whereas there are multiple users who are try to willing to
receive this signal their receiver must not be very costly.

Because if the receiver is costly that cost directly will come on the user so in a broadcasting
system generally my target should be that the receiver is little bit simplified and the
transmitter probably is little more complicated okay why I am saying all these things this will
give me another design direction where I will probably take out this difficulty of getting this
carrier recovery circuit into it and then multiplying it.

238
So this entire stuff I will take out and I will employ another modulation scheme which will be
just a simplified modified version of this where the receiver will be will be becoming very
simple but the transmitter will be little more complicated we will show you what kind of
complication we will be having in the transmitter probably transmitter will be little less
efficient okay but that is pretty obvious whenever we have one-to-one communication
probably this is better.

Because then I cannot make the transmitter very costly is one-to-one communication again if
I increase the transmitter cost that will come to the user so there I need to have a balance that
transmitter receiver must be almost similarly equivalent complex, so there will probably we
can employ this particular technique so that is why that short-range radio communication
people have used SSB sorry DSB double sideband suppress carrier this particular modulation
techniques that we are discussing about.

So whenever we have one-to-one that short-range radio communication which is not

broadcasting in nature there we can employ this kind of technique okay now let us try to see
that we have talked about this demodulation let us first for DSB - SC let us try to see what are
the different kind of modulator that we can generate, so we have talked about that multiplier
let us say there are if you go into market you will see that there are multiplier chips with
differential amplifier and all those things it is little complicated okay.

So we can always employ a direct modulation by that technique but there are other
techniques and which will be discussing other very simplified techniques to do modulation
we will discuss those things.

(Refer Slide Time: 16:46)

239
y(t) = a x(t) + bx 2(t)
z(t) = y1(t) − y2(t) = a x1(t) + bx12(t) − a x2(t) − bx22(t)
= a [x1(t) − x2(t)] + b [x12(t) − x22(t)]
= 2am(t) + b (x1(t) − x2(t)) (x1(t) + x2(t))
= 2am(t) + 4bm(t)cos (ωct)

The first one is called the nonlinear modulator so in that case we will be using a nonlinear
device okay what do we mean by nonlinear device that if I give our input x(t) the output will
just not be a linear scale factor of this okay what will happen it will produce some square
term as well it is just a nonlinear device, so output will be means showing non-linearity with
respect to the input so suppose my x(t) is this and output is y(t) and if I have y(t) following
this relationship a x(t) + bx 2 (t).

That is the simplest non-linearity we can get that is the quadratic one okay, so we can also
have other non-linearity or other higher order like cubic so it will be having some another
constant c x3 or so on higher polynomial also but we can easily get this quadratic non-

240
linearity and realize this by some devices which we all know like transistor or diode okay so
they if you see their characteristic function that has a non-linearity because the characteristic
function generally goes like this right.

And if you bias it in certain region you will see that it will follow quadratic nature okay, so
this kind of nature so if you give input output will be just in a quadratic form with some A
and B that will depend on the diode characteristics okay but if we have a nonlinear device
which is let us say a diode properly biased so that we get a quadratic non-linearity into it and
that is this device and now if we can connect this in this fashion so my two input if you know
are m(t) and cos ωct I need to produce the multiplication term.

So what I do is something like this I have two adders so these are just you can put them as op
amp adder okay to just add the signal so this goes over here this comes over here actually this
is adder directed sorry and this is a subtractor or so it is + 0r - I get x1(t) over here x2(t) over
here and then I pass it through a nonlinear device of this nature again I pass it through a
nonlinear device of this nature so I get y1(t) and y2(t) after passing it through this I pass it
through another adder or I should call it subtract and then whatever I get I pass it through a
band pass filter centered around +- ωc or fc okay and the output I will be getting will be
proving that it is actually 4 bm(t) cos ωc if I adjust my be to be ¼ then it is actually m(t) cos
ωc whichever is our target okay.

So how this works it is very simple you just refuse to those algebraic manipulations so if I
have this x1 and x2(t) after nonlinear device suppose this is z(t) what is z(t) is y1(t) - y2(t)
whereas y1 (t) is actually a x1(t) + bx1(t)2 - y2(t)is a x2 (t) – bx2(t)2 right or I can write as a
x1(t) – x2(t) + bx12 (t) – x22 (t) right now what is x1(t) x1(t) is m(t) + cos ωct and what is
x2(t) that is m(t)- cos ωt ct okay so x1 – x2(t) just m(t)will remain right so that should be
2m(t) and this is (A + B)2 - (A-B)2 because x1(t) is if this is or let us say a I should not say A
let us say C and D.

So (C +D) 2- (C – D)2that should be 4 right C x D so therefore it should be 4 x B C x D is

m(t) into cos ωct so m(t) x cos ωct right now what we are doing now the frequency domain
term will come into picture so this m(t) if I take it into frequency domain that should be m(f)
and this if I take into frequency domain that should be m +m(f) sorry and m(f+ fc) and m(f-
fc) if I put my band pass filter around fc then this term should pass through and this will be
not going through it so therefore at the end only this term will survive which is this right.

So you can see we can actually devise a multiplier circuit by two nonlinear device and three
adder adders are very easy to device just take an op-amp and you can you can make a adder
right so 3 op-amp and two nonlinear device properly biased let us say diode properly bias so
that we get this quadratic relationship media tube and a band pass filter again band pass filter
can be designed using op-amp and some active filtering okay.

241
So that is now you can see that the multiplication circuit so this is one way of doing
multiplication there is another way that is called the switching modulation.

(Refer Slide Time: 23:12)

ϕ(t) = m(t)ω(t)

2 π( )
1 2 1 1
ω(t) = + cos (ωct) − cos (3ωct) + cos (5ωct)⋯
3 5
1 2 1
ϕ(t) = m(t) + m(t)cos (ωct) − m(t)cos (3ωct) ⋅ ⋅
2 π 3

We are just trying to show you which other devices that can be employed to do this
modulation right so for switching modulation what we will be doing we know that a
particular transistor or CMOS circuit can work as a switch so in the gate of a particular this
switching transistor if we just put a signal like this which is having let us say +5volt for ½ the
duration and it is being 0 for half the duration.

So what will happen whenever this is being put in the gate it will be on so it will pass the
signal so if I just put in the emitter to collector if I just put my message signal so whenever at
the gate it is getting + 5 volt it will be on then that signal will be passing through it and

242
whenever it is off that will not pass through it so if I just put a resistor across that if I take the
voltage I will see that it is gets switched.

So basically if my signal is something like this and if I just switch it through this, so
whenever this part is on the signals will follow rest of the part it will be 0 again the signal
will follow rest of the part it will be 0 so basically what is happening my message signal is
getting switched through this pulse okay so almost what we are doing message signal
multiplied by this pulse okay so if I represent this as w(t) and this is m(t) I can actually
connect this w(t) in the gate of that switch or transistor.

And in the emitter to collector I can put a resistor across which I will be taking the voltage
and I across the bias this one an emitter I can put my m(t) signal okay so then the output of
the resistor will be this modulated signal right or whether it is modulated or not I do not know
it is a switched signal right now basically that output will be just multiplication of these two
as I can see if this is I put this as one then immediately it will be just a multiplication okay.

So I get my output which we say ϕ (t) is just m(t) ω(t) but the ω (t) if you carefully see that is
a periodic signal so I can do a Fourier series analysis this is where you can see all those
techniques that we have used will actually be used over here so ω (t) I can just expand it in
Fourier series so you can just do it represent this one as this which we have already done
probably something like this okay.

So it is this is overall T this is sorry this is this is T and this is- T / 2 this is T / 2 and this is-
T / 2 and with this period it is period it gets repeated okay, so this one if you just do Fourier
series analysis you can represent it as this ½ so that means the DC part will be ½ that
coefficient next coefficient will be 2/ π ( cos ω c(t) the next coefficient 2 ω ct will not be
there it is the 3 ω ct part which will come 1/3 cos 3 ω ct + so it gets alternative + and – so 1/5
cos 5 ω ct and so on okay.

So basically it has all the odd frequency harmonics and the coefficients will be alternatively+
- and the corresponding coefficient becomes 2 ϕ x 1 by whatever the frequency whatever the
harmonics okay so this is what ω ct is therefore my ϕ (t) will be immediately I can see I
multiply this so ½ m(t) +2 / π m(t) ? cos ω ct - 2 / π x 1 / 3 m(t) cos 3 ω ct and so on okay
again do a Fourier transform of this because ?f wish to see so if I do a Fourier transform I can
see there will be some part at baseband next m(t)cos ω ct so that should be around fc next
will be around 3 fc.

Now if I just employ a band pass filter around fc properly design then I will be just getting
this signal all other terms will be neglected so immediately I get my modulator because this is
m(t) cos ω ct okay so if I just.

(Refer Slide Time: 28:53)

243
2
m(t)cos (ωct)
π
So basically what I have to do I have to put a switch which has two input one is that w(t) and
the other one is that m(t)after that whatever I get that is this ϕ (t) must be passed through a
band pass filter centered around +- ω c whatever I get that is actually 2 / π m(t) cos ω c t
right so that is another way of doing multiplication this is called the switching modulator so
in the next class probably we will be discussing more about the relative advantage and
disadvantage of all this circuitry.

244
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-17: Amplitude Modulation (Contd.)

Okay, so we have seen few types of modulator to generate just multiplication of two signals,
now what we have promised that we will try to see the relative advantage and disadvantage
of these things. Now we are actually coming into the engineering part of it, so let us say the
previous one that we have done there.

(Refer Slide Time: 00:44)

245
4bm(t)cos (ωct)
2
m(t)cos (ωct)
π
m(t)cos (ωct)

We have got the output as 4bm(t) cos(2ωct) if we just multiply we get m(t) cos(ω ct) right, so
these are the three things we are getting, so what is happening if you just carefully see
especially this one when I am doing multiplication think about my power most of the power
because it was having the infinite Fourier series representation many part of power has gone
to 3FC, 5FC, 7FC and all those things so those power because I just employ up a band pass
filtering those things are all wasted right, so that is why I am getting a pretty low power 2/p
which is getting linked over here right.

So that is one disadvantage of doing this, so if I do with the switching circuit I will be
actually leaking power into higher harmonics. So effectively I will be losing power and that is
very important because whatever or I put I want to actually employ that for my modulation
purpose on it so they are all getting wasted so it is very important that I carefully design my
system and that is why probably people have tried to employ direct multiplication instead of
doing this circuit okay.

So that is one aspect of it that we should think of directly multiplying it if we can okay, and
that is why there are devised multiplier chips which will be generally using for multiplying
the signals okay. Now linear one also is quite alright as long as my B is not very small if the
nonlinear device I means where the biasing is very sensitive if the coefficient becomes too
small especially that quadratic related to that coefficient becomes too small then probably
again this will be very low and I will be most of my power will go into the linear part which
is anyway getting cancelled out, okay so that is also very important okay.

So after this let us try to see another form of modulation okay, so the other form of
modulation is called that is actually termed as AM.

(Refer Slide Time: 03:31)

246
A cos (ωct) + m(t)cos (ωct)
= [A + m(t)]cos (ωct)

Because historically that came before amplitude modulation so you can call that as DSB with
carrier okay, so that will also have double sideband but the carrier will be present what does
that means so that means you are at earlier generating m(t) cos(ωct) I will add a carrier to it
which is pure carrier Acos(ωct) where A is amplitude of the carrier, so this is what is
amplitude modulation so it is a very simple one but you will see that lot of advantage will get
out of this and this is being employed for broadcasting generally.

Because the ct corresponding we see what becomes very simple, so we just add a carrier to it
and you will see that the entire complicacy of recovering carrier and then multiplying with
the carrier just vanishes. How we will see now, so if I just do this I can write it as Am(t)
cos(ωct) right, so what is happening it is almost like a DSBSC it is just my m(t) is DC shifted
so whatever message signal I have suppose I have a message signal something like this okay,
so this is my m(t)corresponding sum m(f) I have.

247
Now if I directly do modulation the way I was doing DSBSC what will happen this will
create envelope and then there will be a carrier okay, and fd 0 crossing there will be 180
phase shift this is something we have seen. Now what we are doing this m(t) we are adding a
DC to it so basically this will be that is let us say that is A so m(t) will be over this and
suppose this is the minimum which is called my -mp okay, the minimum voltage of my m(t)
message signal and somehow my A is greater than this mp then the entire signal will be above
0 level.

So basically what I am trying to do is I am trying to avoid any 0 crossing that is all I am

trying to employ over here, so by doing this that means adding carrier actually means that in
the message signal I am adding a DC term and why I am doing that because I knew that
whenever I will be multiplying by cause if there is a 0 crossing there will be a phase reversal
I want to avoid that phase reversal.

For good reason we will come to that, but once we avoid that phase reversal what will happen
now if I multiply this with cos(ωct) my envelope will look like this and inside the carrier will
be flowing and we know that we have avoided 0 crossing so there will be no 0 crossing inside
okay, and there will be no phase reversal inside fine, why this is so much advantageous or
what is the advantage of this and what is the disadvantage of this let us try to discuss that.

First the advantage part, the corresponding demodulator will be much more simplified, so let
us try to see what kind of demodulator I can put for this.

(Refer Slide Time: 07:12)

248
[A + m(t)]cos (ωct)

So what I will do, I will actually for the demodulators I will pass it first through a diode I
completely changing the demodulator now I am no longer multiplying it with cos(ωct) I
know that multiplying with cos(ωct) will still give me result but I do not want to do that I am
just simplifying it, so I will pass it through a diode, what diode will do this it will act as a half
wave rectifier right, so only in the positive half it will pass the signal negative half it will
make it a 0 okay.

So it will just almost take out this half where it is less than 0 so that part it will take out after
passing it through a diode what I will do is I will pass it through a low pass filter and then I
will pass it through a capacitor followed by a resistor which is just nothing but a DC blocker
we will see why that is required and that will make give me my m(t) or very precisely I will
get half sorry 1/p m(t) as long as I am putting this A+m(t)cos(ωct) over here that means
amplitude modulated signal if I put it over here if I just do this circuitry I will get this so let
us try to prove that, okay. So let see whenever I pass it through a diode what is exactly
happening.

(Refer Slide Time: 08:58)

249
A cos (ωct) + m(t)cos (ωct)
= [A + m(t)]cos (ωct)

So I have this signal once I pass it through a diode it is just taking the positive half and the
negative half it is taking out it is again taking the positive half negative how it is taking out
again taking the positive half what does this means this almost means this signal multiplied
by a pulse strength which is exactly synchronized with this co sinusoidal is not it, it can be
represented as this because I am sending this signal only the negative of is going off that
means in the negative half if I have 0 and multiply this it will become 0 and in the positive
half if I keep 1 it will just represent that signal.

So basically this after diode or the diode can be represented as a multiplier which will just
multiply the signal whatever is coming A+m(t) cos(ωct) multiplied by a w(t), where w(t) is
represented by that pulse strain.

(Refer Slide Time: 10:01)

250
 [2 π ( )]
1 2 1 1
= [A + m(t)]cos (ωct) + cos ωct − cos 3ωct + cos 5ωct …
3 5
1
= (m(t) + A) + ⋯
π

which looks like this, remember for the pulse strength it should be having 50% duty cycle
because for a cosinusoidal I have to take the full positive half and I have to negate the full
negative house so it must be 50% duty cycle pulse that is the first thing. Second thing is this
time period must coincide with the cosinusoidal time period, so I had this 1/fc time period
that must be this T so that should be exactly equivalent and that is automatically happening
because I was is doing that I do not have to really bring a signal which is of this nature it is
actually the diet characteristics is almost like this ωt t as if completely synchronized with the
signal getting multiplied, right.

So I can write my output as A+m(t) cos(ωct) which is my original signal multiplied by w(t)
that is output of the diode I can write this now this w(t) I can again expand in Fourier series

251
this signal I can expand in Fourier series so I can write that w(t) instead of w(t) I can write
again that expansion Fourier series expansion which we have seen just assume means in the
previous class probably so 1/2 2/p cos(ωct)-or I can keep 2/p out so 1/3cos3(ωct)
+1/5cos5(ωct) and so on.

Now just to the multiplication what I will get this A+m(t)cos(ωct) into this one that is the first
term, second term will have A+m(t)cos(ωct)2p into this cos(ωct) that will give me actually
one base band part and the other higher frequency part and all other multiplication if I see
that will be all either at ?ct or some high harmonics of ωct, so if I just pass it through a low
pass filter only that part will come out which is nothing but it is 2/p into half will be there.

So 1/2[2/p m(t)] will be there plus there will be all other harmonics if you just multiply all
these things we will just see there will be all other harmonics some terms into m(t) cos(ωct)
and cos(2 ωct) 3 ωct and all those things okay, so I will only have this part now if I put the
pass filter which will take out all these harmonics which will give me just this which is 1/p
m(t) right sorry, m(t)+A there should be this term remains the same.

So m(t)+A1/p so at this point after low pass filter I will be getting 1/p m(t)+ a DC term, now
this capacitor resistor that will just act as a DC blocker so it will just block the DC that A part
I will get 1/pm(t) so what has happened the one that has happened over here my receiver
circuit has become very simple I just have to put a diode no complicated carrier tracking I do
not have to do that carrier track because automatically the diode is doing the carrier tracking
for me.

You see where that cos(ωct) multiplication is happening over here I had this term earlier I
was deliberately multiplying it with cos(ωct) now diode is actually helping me to multiply it
with w(t) which has a cos(ωct) okay, so basically my signal itself is carrying the carrier which
I have told earlier also so it has the carrier information but in such a way it is now carrying
that with a diode I can actually do the multiplication and I get everything out I do not have to
do any synchronization.

252
You know why this is happening because this is a periodic signal w(t) I can write as a
periodic signal. If there was a random phase reversal which was happening in DSBSC then
what will happen let us see I have a random phase reversal.

(Refer Slide Time: 14:42)

1 1
wc
< < RC < 2πB

So my signal was like this I do not add any DC so what will happen within this, this will
sorry this will become my envelope and within this there will be a co-sinusoidal which will
be just having a carrier flip over here. Now I pass it through a diode what will happen, diode
will be that w(t) corresponding to that diode it will be positive here negative here positive
here or I should say positive here 0 here okay, because it is taking out all these negative
cycles but here there is 180 phase reversal so there will be positive followed by positive that
is breaking my periodic nature, right.

So the w(t) I will be multiplying that is no longer periodic that is a random signal, now
because where that crossover will be happening I do not know that depends on the signal, so

253
anywhere it can happen so that completely breaks out my periodic nature of w(t) so at that
point I cannot write my w(t) as this one very nicely.

So therefore that multiplication I was getting I am not sure now and that is the reason why I
will not be getting my signal back if I just put a diode so that diode detection that followed by
low pass filtering detection can only be possible if I put a DC term so that my entire message
signal goes into the positive half.

And whenever I do modulation no 0 crossing is there in the envelope and correspondingly no

phase reversal is happening as long as that is happening I know my diode detector will be all
good multiplier. You can also think about this see it this thing is carrying the carrier
information but let us try to see how I can extract the carrier, so basically whatever is coming
over here I can excite up model means this things a sinusoidal generated by this but what will
happen after some time the excitement will be given by this which is a phase reversal.

So that will just counteract on the exciter whatever excitation it was creating it was creating a
sinusoidal now it will just be a phase reversal version of that so it will just counteract and it
will actually negate that thing. So basically if I on a long term basis I wish to track the carrier
I will not get anything because it will keep on reversing and I will keep on means denying my
carrier generation process so that is what happens whenever we do DSBSC.

So DSBSC carrier recovery has to be done completely separately we cannot really employ
these things, whereas if we just add that A term into it wonderful thing happens and my
receiver becomes very simplified because in the receiver I just get a diode followed by low
pass filtering that is it nothing else is required. The same receiver we can also employ or we
can also think that receiver in a different manner.

Suppose I have my AM signal which is mostly this is called the envelope detector but which
is mostly employed so I will be putting a diode on this okay, and then followed by a kind of
charging and discharging path and here I will be taking the signal out that is my receiver, so
what is happening over here let us say I have this signal of course it has to be AM so let us
say I have this signal and inside I have this carrier.

254
Okay, now whenever it passes through diode I will be just getting rid of this so I will be
getting this positive half and then see whenever the diode is on this thing will pass it through
it and probably what will happen this capacitor will get charged through this okay, so if diode
internal resistance I assume to be very small so this RC value will be sufficiently small to
make the charging time constant very quick it will be charged very quickly okay, so this is the
charging path.

Whenever diode is in reverse bias condition then this capacitor is not getting any supply so it
must be discharged but it does not get any passed through this diode because now it is in
reverse bias condition so it should be discharged through this path now if I put design by R
and corresponding C in such a way that while discharging it just tracks the envelope okay, so
basically what is happening.

I have this so while charging it quickly charges whenever diode is on it follows the carrier
very quickly gets charged but while discharging if the discharging time constant also is fast
that means this RC value is in such a way design that it is also faster then it will discharge
full wave. But basically to receive my message signal I want to detect this envelope because
that envelope only carries my message signal, so what I wish to do is while discharging it
does not get discharged very quickly.

Again next cycle comes back then the diode gets charged so it again gets charged quickly, but
while discharging again it follows a slower path it is not discharging very quickly then what
will happen it will just keep tracking the envelope okay, if the discharging time constant is
too slow then what will happen if you see this negative slope then the discharging time
constant will be too slow will be too slow that is also something which I have to take care of,
the discharging time consent should be such that I will be always tracking the envelope to do
that I have to design my RC.

So basically what should happen this RC must be greater than or much, much greater than 1/
wc okay, so that it does not follow the carrier but it must be less than my 1/2 π B, where B is
the highest frequency that I can have in my message signal so this envelope, so this RC must
be less than that so that whenever it is getting discharged it is not even slower than my slope
of the message signal so as long as I am actually considering this particular condition because
B and wc are known whenever I am doing modulation I know with what frequency I am

255
modulating and I also know what is the, if the signal is message signal is band limited what is
the maximum frequency that it contains.

So if I know these two parameter and if I know this condition I will be accordingly designing
by RC and immediately I know that it will be tracking my envelope, so finally my signal will
look like some which is almost tracking the signal right, after that there will be there are
some aberration just a low pass filtering will do this will make it smoother okay, so
immediately I get this so this followed by another low pass filtering probably will smoothen
the envelope and of course there should be a DC blocker that has to be there because this help
this will whenever I track the envelope that will have a DC I should have a DC blocker so
immediately I get that minus such.

So there are two circuit that we have discussed one was diode followed by a low-pass filter
almost employing similar technology or I can think it from time domain so that is to
representation one I was thinking in frequency domain that my diode generates a signal
which has all components all harmonics as well as the baseband why I was employing the
low-pass filtering so that all other harmonics are rejected.

And after that I knew that it is A+m(t) so as I was putting a DC blocker here I am actually
seeing it from time domain so that is why both the representations are good so I am seeing it
from time domain what I am trying to see is in the time domain whenever I pass it through a
diode it is actually just giving me the positive half next it just has to track the envelope so for
that I need a very faster charging and very slow discharging time constant that is what I have
devised and immediately we could get the envelope tracker.

After that it is some low-pass filtering DC blocking to smoothen the signal according to my
requirement okay, so this gives a very simplified way of the demodulating and this particular
modulation technique for that reason is as we have seen that it has a carrier. So just if we now
try to see the frequency domain representation of the modulated signal.

(Refer Slide Time: 24:16)

256
A cos (ωct) + m(t)cos (ωct)
A A 1 1
⇔ δ ( f − fc) + δ ( f + fc) + M ( f − fc) + M( f + fc)
2 2 2 2
Which is a cos(wct)+m(t) cos(wct) if I do a frequency domain representation this m(t)
cos(wct) of course that will be 1/2m(f+fc)+ 1/2(f-fc) this is all right but this cos(wct) that will
be 2 δ function right, so this will be A/2 a δ function at +fc δ function at -fc right, so
basically if I just try to draw the spectrum what will happen at fc there will be these two term
m(f) term and there will be a δ function which is actually eventually representing the carrier,
so that means it has two side band like earlier one but it also has this carrier term so that is
why it is called just DSP or AM I am not suppressing carrier.

Whereas in the other one I was actually deliberately suppressing the carrier I was just sending
the multiplication term I was not putting the carrier into it but I know now the difficulty for
the receiver if I do not transmit the carrier, but is it all rosy in the next class will probably
prove that is not the case whenever we put a carrier we are again sending unnecessary signals
due to that there will be extra amount of power I will be sending so the overall power
efficiency will be lesser in this system.

257
And that is why probably you will see the transmitter has to release more amount of power to
get similar kind of effect at the receiver, so that means transmitter has to be has to employ
some power amplifier and all those things okay, so because it has to transmit huge amount of
power transmitter becomes more means engineering design wise more complex and more
costlier. So we will try to in the next class we will try to see that how four AM the transmitter
is costlier which was our design target that I want to for a broadcasting system I want to make
the transmitter little costlier whereas the receivers will be much more cheaper.

And for point to point I want to make the transmitter little bit cost effective and receivers
similarly costly okay, so we will see that part also that this has one advantage but there will
be another disadvantage in to it. So once we have done that in the next class we will go
towards more bandwidth efficient modulation amplitude modulation those are like single side
band we are now transmitting double sideband we have already discussed that both the side
bands are carrying equal amount of information.

So we do not need both side band, so we will try to see some more bandwidth efficient
modulation like single sideband which is called SSBSC single side band suppressed carrier
again we would not be transmitting carrier or there is something called vestigial side band we
will see the application of these two some cases like for voice you can employ single
sideband whereas for video you cannot employ that you have to go for vestigial side band so
we will see that and it is single side band probably will be the most bandwidth efficient
whereas the serial side band will be less bandwidth efficient.

There is also another modulation scheme which is also bandwidth efficient which is called
quadrature amplitude modulation that also will discuss and we will see that relative
advantage and disadvantage of them.

258
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-18: Amplitude Modulation (Contd.)

So far we have discussed about DSB-SC and amplitude modulation, so these are two version
of amplitude modulation that we have discussed to the previous few classes. So DSB-SC was
something where it is called double sideband but also suppress carrier that means you
modulate it the way we model it is just frequency shifting actually so you multiply the
message signal by a cos wct where wc is the frequency of the carrier okay. So if you just do
that then you get double sideband but if you on top of that if you add carrier to it then you get
amplitude modulation.

(Refer Slide Time: 01:08)

259
 QDSB−SC(t) = m(t)cos(ωct)
ϕAM (t) = m(t)cos ωct + A cos ωct
So the DSB-SC if I just write f DSB-SC(t) that should be if my message signal for
modulating signal is empty and carrier is cos wct so this is DSB-SC and amplitude
modulation is nothing but the same thing which is m(t) cos wct plus A cos wct so that is the
carrier part we add to the signal and we have seen the benefit limited benefit of it.

So here in AM we have seen the receiver becomes very simple it is just envelope detector
whichever way you do that we have already discussed two methods of envelope detection
and for DSB-SC we need in we need to generate local carrier, so for that we need to extract
carrier from the incoming signal so that requires carrier synchronization and all those things.

So the modulation of DSB-SC was very simple you have to again be multiplied with cos wct
followed by a map of sorry low-pass filter, so this is something we have already discussed
now let us try to see it is just some more discussion about this you will be mostly seeing that
there is a importance of tone modulation because through tone modulation we get means try
to get information about what happens whenever we model it so tone modulation means it is
just it is not a message actually this is the sinusoidal I want to modulate it.

260
(Refer Slide Time: 02:39)

m(t) = cos(ωmt)
ϕDSB−SC (t) = m(t)cos (ωct)
1
=
2 [2 cos (ωmt) cos (ωct)]
1
=
2 [cos (ωm + ωc) t + cos (ωc − ωm) t]

So my modulating signal that MT I have a specific representation so far we are not

representing it okay, we are just saying it is some kind of signal it maybe voice signal or
something. So here in tone modulation, so we call it tone modulation it is just a indicating
modulation actual modulation will never be that because modulating signal we have already
discussed that must have some randomness in it and if you just say it is just cos wmt it is not
carrying any information but it is just to say how the modulating modulated signal will look
like what should be the frequency spectrum of it and all those things.

261
So if I have m(t) equal to cos wmt so therefore f DSB-SC(t) which is the modulated one must
be empty into cos wct which is nothing but cos wmt and E cos wct so I take 1/2 x 2 so this I
can put as cos A so it cos A cos B formula, so I can just put that as of course our assumption
is carrier frequency is much higher than for modulation that has to be there it should be wc-
wm, so if you just try to see in the spectrum what is happening our modulating signal that
was sinusoidal or a cos we have already said that cos in the frequency domain can be
represented as 2 δ function right.

So if I just put this as M(f) so this is at wm this is at – wm so whenever we modulate what

happens there are two cos now one is cos wc + wm so if this is wc there should be one cos
function which is w at wc + wm and there should be another co sinusoidal which is at wc-wm
and of course the negative half should have the symmetric part, so if this is - wc so there
should be two part over here with amplitude half right.

So all this must have half amplitude so that is ΦDSB-SC(f) so whenever we model it actually
if a tone we modulate with a carrier we just get to go sinusoidal and the corresponding time
domain representation will be something like this.

(Refer Slide Time: 05:37)

262
ϕAM (t) = cos (ωn) cos (ωct) + A cos (ωct)

So this is your modulating signal with frequency wm and let us say that expect area which is
having a frequency much higher so whenever you modulate what we see that becomes the
since that should represent the envelope so it should be something like this and of a similar
presentation and a positive and negative up and inside with wc frequency it should and there
should be a phase reversal we have discussed about that also everywhere wherever there is a
voltage means whenever it crosses 0 there should be a phase reversal.

So that should be the time domain representation so if you just try to see what this is this is
actually two co sinusoidal beating with each other the addition of two co sinusoidal signal
with frequency wc+wt and wc-wm once you do that if you just take these two co sinusoidal
signal and then add them together you will see that this pattern will be generated. So that is
what is happening over here so this is how the DS-SC should look like whenever we add
carrier to it what happens.

So in the frequency domain if we try to employ so ΦAM(t) that of course cause O MT

because that is the modulating one that must be multiplied by cos wmt plus there should be a
cos wct right okay carrier term should be there so therefore this part we have already been
said what should be the frequency domain corresponding frequency domain representation.
So that should be two co sinusoidal this should be wc so this must be wc+wm and this must
be wc-wm and due to the presence of KD s there must be a by two strength carrier at O C.

So it is just we can see that whenever we have which carrier there should be a d function
which is just representing this carrier at O C and – O C of course because it must be given
symmetric okay.

(Refer Slide Time: 08:21)

263
So if I have just general modulated signal of this M(f) so that is the M(t) has a Fourier
transform of M(f) let us say so if I modulate that with wc that should be my representation if I
do DSB-SC if I add Carrier there should be just a delta function at wc and –wc A/2. So this is
what we have been told and this is just a tone modulation version of that so now what we
wish to do is we wish to see that the same modulation technique can we employ for other
things one is called the frequency mixer or converter let us first tries to understand the utility
of that.

So generally what will be happening suppose I wish to put some signal at a higher frequency
okay, so the higher frequency at that higher frequency let us say the available frequency
because I need to get a frequency slot which is in the chair media nobody is occupied okay so
that might be very high pretty high. So what I need to do while putting I need to modulate
with that wc but in between I might have to do some signal processing like I might have to
amplify the signal and so on some other signal processing I might have to employ okay.

So if I wish to do that what I can do is something like this I can first modulate the signal or
even demodulation also similar thing will be happening so I can put it into a intermediate
frequency let's call it wif similarly in the negative half also same thing will be happening and
then after that from wif will be actually translating it towards wc or sometimes I might do

264
the reverse thing that after receiving okay whatever signal I receive so suppose my modulated
signal is coming.

So this is centered around wc - wc I put a band pass filter centered around wc and the
bandwidth should be greater than the 2 B value right we have discussed about the bandwidth
for a moderated simple okay. So I will get just this one if there are some spurious thing in
other frequencies that will be all means rejected I will just get my signal now once I get this
signal I might have to amplify the signal okay.

So the receive signal has to be amplified but at that high frequency if I wish to put amplifier
circuit at high frequencies any circuit design will be very difficult because they'll all start
radiating things okay so there any circuit component you start putting they will be already
getting so I cannot really do employ any signal processing at that high frequency. So what I
wish to do I still need to do amplification on any other signal processing that I wish to do.

So I need to first translate it to a frequency band where I can do my signal processing so the
good advantage is while modulating and transmitting the signal I am still using wc which is a
requirement in the carrier or sorry requirement in the channel because I need to put a signal in
that particular frequency event that is the only free frequency band, but I also know at the
same time that processing I would not be able to do at that frequency.

So I have to first receive that so receiving is still alright I can put a band pass filter at that
frequency centered around wc with to V bandwidth I can take that signal now with a very
simple circuit if I can just down convert it to a lower frequency then I can employ all my
signal pulsing things on that, so all I have to do is if I get this signal.

(Refer Slide Time: 12:30)

265
 m(t)cos (ωct) ⟹ m(t)cos (ωIF t)
m(t)cos (ωct) 2 cos (ωc ∓ ωIF) t

= m(t)[cos (2ωc + ωIF) t + cos (ωIF t)]

That is a center around wc-wc that is my receive signal I wish to at that frequency level. Now
I have rejected after band pass filtering I have rejected all other things now I want to I cannot
do s passing over there I want to actually translate it back to wif which is pretty low
frequency where I can do signal processing, so all I have to do is from this to this I have to do
a conversion okay which means see the spectrum of the signal frequency components
remains the same it is just getting translated into one frequency to another frequency.

So all we have to do is if the modulated signal was empty cos wct from there I will have to
create M(t) cos wif is all that I will have to create okay, so that is something we will have to
create. Now this is something which can be done almost equivalent with a similar modulation
technique. So what I can do suppose this is something I am receiving okay. So this whatever I
am receiving cos wct I will actually multiply this with two I am just taking it we will see that
why I am taking two, but I just have to multiply it with something called cos wc + wif(t)
okay.

266
So this is all that I will have to do so I have to locally generate a particular co sinusoidal
which is centered at wc if+ w or I can do it even at - both will give me same results okay, so I
just have to do that and I have to multiply it. So my circuit will look like this is m(t) cos wct
which is coming to me I receive it through a antenna then I put a multiplier circuit whichever
where we have already talked about how to design a multiplier circuit and then I multiply this
with cos wc - or + wif(t) okay.

Whatever I do get after that I put a band pass filter okay, which is centered around that O IF
you will say that this will give me this particular signal back why let us try to manipulate this
particular part, so I am just doing multiplication so after multiplication I will get because it is
2 cos x cos A cos B so I can I can put cos (A + B) + cos (A- B) so cos A + B suppose I am
just taking the positive sign, so that should be 2 wc + wif x (t) + cos wif and this final this.

So that should be just wif(t) right which is my target. Now this is that wif this is at very high
frequency twice of wc plus wif right. So if I just put a band pass filter around wif' okay which
is having a bandwidth of to be so what will happen this whenever I see this will be creating
something at wif' and it will be creating something at very far which is 2wc +wif is right so
that is what I will be getting.

Now if I just put a band pass filter I am just showing the positive spectrum similar things will
be there in the negative half also if I just put a band pass filter this will be rejected so
immediately what do I get I get this empty so this term will be cancelled the left over will be
cos wif which is my desired signal right. So this particular methodology is called as either
frequency mixer or frequency converter because what we are doing we are not changing the
message signal we just translating the carrier from a higher frequency to a desired lower
frequency which can be done.

Here also you might be asking that I have to generate a local carrier does that need to be
synchronized with the input carrier because this is at a separate frequency we do not need any
synchronization okay, so that is the advantage of it whereas when you are doing coherent
demodulation in that or we should say multiplied demodulation that means you again
multiply with cos wct at that point also we are putting you are multiplying with it and we are
putting a low-pass filter but they are the frequencies has to be completely synchronized we

267
have talked about that that if this is cos wct that has to be cos wt here neither the frequencies
are synchronized so it must be some other frequency.

So I can always generate it and we will later on prove also the phase also need not to be
synchronized with it okay so we will prove that when we talk about that carrier
synchronization and carrier recovery okay but right now we can just give this information
that we or this frequency converter we do not need additional requirement of synchronization
no synchronization is required I can just multiply with any local carrier that I generate I need
to just roughly know what is the wc okay.

So it needs to be closer to that and then I need to generate something at wc-wif as long as I
am doing that that should be fine. So I will be able to do this so this is a very important tool
in communication again because the frequency converters as I have already pointed out it
might be required in many cases where you might have to modulate at a very high frequency
and you might have to translate that down to a desired frequency band where you can do
signal processing which might be a big requirement.

So that is one technique you can already see that which can be employed for communication
but it is actually not affecting what you are doing in the channel you are still choosing that wc
right so that is one thing we have talked about the other thing we wanted to talk about is
actually the relative advantage or disadvantage due to this amplitude modulation okay. So
advantage we have already talked about that the receiver side is very simple right we just
need to do a envelope tracker or envelope detection which is a very simple circuit we have
already talked about that that you need a just a diode and then followed by a low-pass filter
okay either low-pass filter or a RC charging and discharging circuitry which properly design
R and C we have already talked about that.

So it might become very simple you do not need any carrier synchronization you do not have
to generate carrier yourself, so all those things are not required whereas if you wish to do
without carrier that means not amplitude modulation as DSB-SC sequence carrier a subset
you might have to locally generate a carrier which is synchronized with the incoming carrier
or you might have to actually take the carrier I mean extract the carrier from the incoming
signal and then demodulate it that is you have to multiply.

268
So let us measure strategies it might be advantageous okay in that aspect let us try to see if
there is any disadvantage because there are no free lunches if we try to get some advantage
there must be some disadvantage. So let us try to see what the disadvantage we get so
whenever we add a carrier what was our purpose this is also something we have discussed,
that whenever we add a carrier the message signal whatever way it was.

(Refer Slide Time: 20:26)

ϕAM (t) = m(t)cos ωct + A cos ωct

= [A + m(t)]cos (ωct)

The lowest amplitude whenever we add carrier actually almost like adding DC value to the
method signal and then doing the DSB_SC right, so the DC value which is actually the
amplitude of the carrier that must be greater than the lowest amplitude of the message signal
this is something we have already proven that my m(t)coswt add A cos wct to it which
actually this is ΦAM(t) so this is we can just say m(t) + A x cos wct. So this is my new
message signal where with empty I am adding a and if I wish to do envelope detection still
we have seen that it must not have any zero crossing because whenever I have zero crossing I
will have that phase 180 phase shifting and then I would not be my envelope tracker will not
give me the signal back.

269
So even if I add a DC to it and it still has some spurious part which is below means zero level
so if I just try to see the envelope it will have this thing and then the carrier will run on top of
this and there will be a phase reversal and if I wish to do envelope detection so envelope will
detect this and this so my detected signal will look like this which is not the original replica
of this one right. So whatever I need to do I need to take this up so carrier must be added in
such a way that the amplitude of the carrier is always is whenever I add that amplitude of the
carrier is always greater than even the lowest value of my message signal okay so let us say
that lowest value is - MP okay.

So therefore I have a condition that a must be greater than this MP okay this must happen and
in that aspect we define a term called modulation index which is defined as mu which is
actually MP /A where MP is the lowest value of my message signal okay or you can write
that as M in whichever is good for you but anyway this is the modulation index that has been
defined. So if you just see this definition immediately you can see that µ must be of course
greater than zero because if it is less than zero then A will be negative with respect to MP
right MP is already a positive because - of MP is the negative values so MP will be always
positive.

So modulus value of that and then if mu is negative then a will be negative which is not good
okay, so all we will have to do is that mu must be that is a trivial one it must be greater than
zero all this but the other one is very important because A must be greater than this so
immediately from here I can write MP divided by a that must be less than one I just took a
down say MP by a less than 1 so I can also write this at max I can have equality but anything
other than equality I put always I will have the danger of this kind of scenario okay. So I
should be this modulation index I should be always ensuring that this is happening okay

(Refer Slide Time: 24:08)

270
 A2
Pc = Psignal = Ps /2
2
Ps /2 Ps
η= =
A2
+
Ps A 2 + Ps
2 2

Now let us try to see what is the effect of this okay, so let us say I have some let us say I put a
ΦAM(t) which is nothing but some A cos wct that is the carrier part and I have some message
signals into cos wct right. Now this is the carrier part and this is that PSBs see what okay, so
what is the power of this one so I am trying to calculate the power of this one so I know it is a
co sinusoidal so it should be A2 / 2 if you just evaluate the power of that so this is P carrier
PC I call it and what is the power of P signal suppose MT has a power PS that is something I
know already the message signal has a power of PS.

Now if I multiply with cos wct we know that the power becomes 1/2 we have already
discussed that in a previous class so this must be PS by 2 that is the power of this part okay
fine. S so this is my 2 power so I call this P signals okay. Now let us say talk about some

271
efficiency okay ultimately what I am trying to do I am trying to actually transmit empty
through the channel okay.

So the significant amount of power which is which is actually required or which is the
valuable power that is actually this PS right the signal power so that is the power I wish
means I wish I could have transmitted if I was not employing any carrier modulation I could
have been in the baseband I could have transmitted this the empty was getting transmitted so
corresponding signal power was PS.

So this is the one I will be actually I will be wishing to transmit whereas due to this particular
modulation I am transmitting a little bit more so what is the thing I am transmitting so I
should have transmitted PS but right now I am transmitting what I am transmitting this A2 / 2
+ PS/2 okay if I compare it with DSB-SC in DSB-SB I will be transmitting only this so how
much I will be transmitting PS / 2 so that I would have eventually transmitted if I was
transmitting DHBSC.

Now because of this particular thing I have to transmit power for this as well as power for
this so this is the overall power I am transmitting right so immediately I get so that is the
efficiency to actually which I could have done with PS now I am transmitting A2 + PS so that
is the overall power efficiency that I get doing this right so I am transmitting extra amount of
power which is a square. So this is always sitting inside okay.

Now let us try to see with this efficiency factor so better the efficiency I am better off if I just
transmit DHBSC then below also I will be I will not be transferred in this A2 so it will be
100% efficient that means whatever power I have to transmit I am just transmitting that much
but if I do AM then probably I am transmitting little extra power that means I am wasting that
power because for transmitting carrier I am getting no advantage that is not carrying my
signal right so what I can write suppose if I have a tone modulation right.

(Refer Slide Time: 28:07)

272
m(t)cos (ωct) = μ A cos (ωmt) cos (ωct)
μ2 A2
PS =
2
μ 2 A 2 /2 μ2
η= =
A2 + 2
μ2 A2 2 + μ2

So my MP is something called a cos wmt and I have a modulation index of MU that is added
to it okay, because if I just write it as some a s cos wmt okay, so what is the lowest value of
that it is - As okay what is the TM for this which I was talking about the lowest value of this
that must be a s okay. So I should have my carrier signal AC so that means AC AS / AC must
be my modulation index okay.

So therefore I can write a s as my µ whatever modulation index I am employing into AC so

or a if I just put this as a so we will do a so I can write this right so that is my MT okay or MT
cos wct if I just write okay then this should be multiplied by cos wct right. So my overall
signal is a cos wct + µ a cos wmt x cos wct so this is what I am again we are demonstrating a
tone modulation right.

273
So in this case for the message signal which is this part what is the power of it? It is µ2 A2
right so that is the power divided by two that is the power of the message signal so this I can
call as PS and immediately I can evaluate the efficiency which is sorry µ2 A2 / 2 right
divided by A2 + µ2 A2 by two according to that formula that PS divided by PA or sorry a
square plus again PS right. So this is what I get a I can cancel out so I get µ2 / 2 + µ2 right.

If I just put µ = 1 that is the best I can do that is the highest modulation if µ = 1 I will be
seeing almost this envelop just touches it does not cross over anything bigger than that will
be below this so if I just put µ = 1 immediately what kind of efficiency I get to go to 1it will
be 1/3, so 33% efficiency, so that is the best I can get so immediately you can see that is
where I pay DSBSC whatever power efficiency I was getting compared to that in AM I am
getting33% efficiency on D so that is 1/3 power is the useful power rest of the things are just
wasted okay. So we will end it over here for this class next class we'll start defining other
bandwidth efficient modulation.

274
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-19: SSB - SC

Okay so the last class we have already discussed about the advantage disadvantage of AM
DSB-SC but both AM DSB-SC if you see how it occupies the spectrum.

(Refer Slide Time: 00:36)

275
M( f ) = M+(F ) + M−( f )
It is always either suppressed carrier it will be occupying this entire to be banned if I put the
carrier on top of this nothing else changes at least in the frequency domain nothing else
changes still it occupies to be but I can see already if you see the message signal which is
getting translated over here and creating this to be banned it is even symmetric right this M(f)
must be a Fourier transform of a real signal, so if it must be even symmetric so whatever
spectrum component you have the negative of also that there will be a symmetric spectrum
component.

So I know already that only this half carries the whole information I do not need that other
half this other half when it was actually a baseband signal it was occupying only B band it is
just because of this modulation I have suddenly increased it to be right and this is
just repeating the negative half which is a mirror image of the positive half so it is not
required actually to transmit that I can always get back my signal we will discuss that later
on through the positive half only or just one half.

Either positive or negative whichever half you require from that portion only I can get extract
the whole signal that is possible will show that that is possible so if we can employ something
where we do not have to actually transmit both the bands then, I can save some bandwidth
because this is the band he will be occupying in the channel so to be is lost from me if I can
reduce it to be other B I can actually transmit another signal and if all the signals are off B
bandwidth then I can actually transmit double the number of signals with same band utilizing
the same time right.

So that is why we can now think of employing some bandwidth efficient so this will be all
processing in the looking at the frequency domain, so bandwidth efficient transmission or
modulation technique so that should be our first target so to do that as you can see that I need
to get one half of it right, so what will say that if this is M(f) I represent this as one half this
and another half this I call as M + (f) and this I call as M – (f) that is the negative part of the
spectrum and the positive part of the spectrum.

And if you just see if I just add these two I will get my M(f) so M(f) is nothing but M + (f)
M – (f) right so that is the representation okay, so this representation will be trying to utilize
for generating our signal you will see that the frequency spectrum how it generates our proper
frequency efficient modulation techniques, so will you see from our discussion later all but to
do this we need to employ another technique or tool that is required to be studied first then
only we will understand how to employ this technique so this is called Hilbert transform.

(Refer Slide Time: 03:59)

276
 1 ∞ x(α)
π ∫−∞ t − α
xh(t) = H{x(t)} = dα

1
= x(t) *
πt
xh(t) = x(t) − j sgn( f )
1
⇔ − j sgn(t)
πt

Okay let us try to understand what is Hilbert transform actually I am just moving definition
now we will understand why this is really required so we are saying Hilbert transform is
nothing but H{x(t)} which is nothing but Hilbert transform of xt and it is represented as this
1/π it is just the definition I am giving okay so it is of course a function of T and I have
taken that dummy variables or integration a and it is actually x(α)dα divided by t – a.

If you try to see what this is this is almost like a convolution we have already defined what is
convolution it is a convolution of this x(t) convoluted way with 1 / πt okay, the convolution
means its suppose I want to convolute mo m1 (t) with m2 (t) we have already defined that so
you need to get m1 (T) and m2( t-T) dT right like that you have to integrate from –infinity t

277
+infinity similar thing is happening 1 by πt is one of the function let us say m2 (t) and x(t) is
the if you just do convolution you will be getting this okay.

And that instead of t we are saying this is that a right or if the convolution integration
variable is t then this should be that alpha okay, so I can I identify immediately that x h(t) is
if I do Hilbert transform it is nothing but the convolution of x(t) and 1 / πt why I am doing
that because we also know we have means we have already discussed about that that if some
signal is convoluted in time domain if I do a Fourier transform of that it should be
multiplication of those two signals so if I represent x h(f) as the Fourier transform of xh(t).

Then that must be equal to the Fourier transforms suppose x(t)has a Fourier transform of x(f)
and 1/ πt we have already defined the Fourier transform of that okay so 1/ ?t is actually the
Fourier transform is – j signum (f) is something we have defined we have actually done this
thing for signum (f) t okay, and then we know also that anything in time domain if I have a
Fourier transform frequency domain also the duality property I can use and if you just utilize
those two things you will see that that signal function the Fourier transform signum (f) is
actually a Fourier transform of - 1 /J πt .

So from there we get dis conservation right so J Signum( f) actually has a Fourier transform
– j signum( f) has a Fourier transform of 1 / πt so this is something I know so 1 / πt the
Fourier transform is - J Signum( F) so which is nothing but J x(f) Signum (f) so this is
something I know about the Hilbert transform okay, so if I pass a signal through our Hilbert
transform circuitry.

(Refer Slide Time: 08:08)

278
xh( f ) = x( f )∥( f )
H( f ) = − j sgn( f )
−j = 1e −jπ/2f > 0
{j = 1e jπ/2
=
f<0

Suppose this is my Hilbert transform circuitry so if I pass x(t) through it I will be getting
xh(t) now I want to get the characteristics of this that is the h(f) of this one but that transfer
function of this one so I know that xh(f) if this is a the transfer function is HF that should be
accept into h(f) right so if I just go back to my previous relationship then what is my h(f) if
this is the Hilbert transformer the h(f) must be j Signum (f) right we have already proven that
that x h(f) is accept into – jSignum (f) therefore h(f) must be j signal s so this is actually the
transfer function of a Hilbert transform.

Okay which I can write as -J when F is greater than 0 and + J when F is less than 0 right
because Signum function we have already defined Signum function is nothing but this okay,
so whenever F is greater than 0 it must be + 1 so this is + 1 and this is - 1 we have already
defined that so this is f so whenever F is positive it is + 1 so that must be - J and whenever F
is negative its -1 so that should be - J into - 1 that should be + J.

279
So this is what we get h(f) definition okay which I can further write as 1 into e to the power
-J π/ 2 and this we can write as 1 ej π/2 right J can be written as this okay just put euler's
theorem you will see that okay because it will be Cos π/ 2 + J Sin π/2 Cos π/ 2 will be 0
means if it is this then it should be J π/ 2 J Sin π/2 that should be J okay and this would be - e
so if you just see the transfer function looks like this any transfer function this is the transfer
function any transfer function must t have or amplitude.

This one plot and a frequency plot the amplitude if mod h(f) I wish to plot that should be
constant that is 1 if you take mod of this that is always 1 so h (f) has mod of H(f) that is
always 1 and now the phase one let us say Θ h(f) I want to plot okay so what that should be
that should be whenever f is positive it should be Θ is - π by 2 so that should be - π / 2 and
this is + π / 2 this π by 2 this is – π /2 or right so that means whenever I have a positive
frequency component it is actually shifting the phase see the amplitude remains the same.

So amplitude I do not touch whenever I pass it through this whichever frequency component
suppose all tones I am passing so whichever frequency component is there for the positive
frequency component it is giving π / 2 phase shift and for the negative frequency component
because the negative frequency component will be exactly opposite it is again giving a π /2
phase shift so it is always giving a – π /2 phase shift to every component every frequency
component so whichever it is the positive it is – ?/2.

In the negative because it is already either π or - π so it is actually + π /2 if you just take it

out it is again shifting by - π /2 so it is actually always every frequency component that is
give it is giving ?/2 phase shift or I should say rather more concisely that – π /2 phase shift, so
this is a typical circuitry which does not touches the amplitude but it just take every
frequency and give π /2 phase shift so any frequency component if you it will just give you π/
2 phase shift here I would like to point it out there are two different things we suppose we
have a signal we give a delay okay.

(Refer Slide Time: 12:47)

280
x(t) = c1 cos (ωct) + c1 cos (2ωct)

= c1 cos (ωc (t − t0)) + c2 cos (2ωc (t − t0))

So let us say I have a signal x(t) which is nothing but let us say it has two frequency
components Cos wCT + Cos 2 wCT okay with some coefficient C1 C2 yes I am providing a
delay to it okay so this is C 1 Cos wc delay means it will be delayed by that much time so t -
t0 + C let us say this is C2 C2 Cos 2 wCt - t0 okay, so let us say I have a circuit through
which if it passes it gets a delay okay so it just appears little bit later because the circuitry
provides some delay and that delay is T0.

So if I have that then what is happening this particular frequency component Cos wct how
much overall delay it is getting w0 T 0 sorry wc t0 and this frequency component how much
delay it is getting it is getting 2 wC T0 right different frequency.

Component will get different overall delay okay so that is the overall delay that they are
getting right so sinusoidal will be delayed by this much amount, so as you can see different
frequency component sorry I should not say delay was already T0 so that is the phase shift
we are getting we should say we should now talk about phase shift so basically this particular

281
frequency component is getting a phase shift of wC T0 and this particular frequency
component is getting a phase shift.

Which is 2 wC T0 right these other two phase shift you are giving if I was passing it through
a delay element whereas when we are passing it to a Hilbert transform what will happen both
of them will get equivalent phase shift all π /2 so both of them both the frequency component
will get equivalent phase shift so immediately what will happen they will have equal phase
shift but the delay will be different for different frequency element so that means if a
constituent signal is composed by multiple frequency element.

They will be separate we delayed so basically after Hilbert transform will not get back similar
signal whereas here the signal means composition will remain the same so there is a
difference between delay and this one it works almost similarly because it does not patch the
amplitude it just works on the phase okay for every frequency component but here in Hilbert
transform you are giving constant phase shift to every frequency component where in delay
element you are giving constant delay shift to every frequency element that means not
equivalent phase it will be actually linearly going growing with the frequency term if you just
see the phase element is 2 wC T0 for 2 wC.

Frequency it is actually wCT 0 that means it gets doubled over here and whatever amount it
was having over here, so whenever your frequency gets doubled the phase shift will be
doubled so if you just plot a delay element that look like this whereas if you plot the Hilbert
transform that will also have amplitude spectrum as this mod HF will be this the delay
element will be linear so whatever is the frequency it will be just scaling with the frequency
term okay if the frequency doubles the delayed out double so it will be just linear with 45
degree angle okay.

Whereas for Hilbert transform you are constantly giving constant phase shift delay will be
different okay, so there the phase shifter is this so that is the difference between these two
both of them are delay element or a Hilbert transform both of them actually shifts different
frequency with either different phase or this one gives same phase shifting and they will not
touch the amplitude but there is a certain difference and that is why whenever you do Hilbert
transform you will have a deformation to the signal.

Whereas whenever you do a means you provide a delay element you will never see any
deformation to the signal this is the basic between these two okay and as you can see this is
very easy to realize okay, that is the circuit which is very easy to realize just keep on putting
register whenever it register means it will be just a propagation delay through it, so if you just
keep on putting a register or a transmission line your signal will come it will take some order
time to come out from this output.

282
So it will just be delayed version of the input signal so this is very easy to realize any real
circuitry will give you that as long as it is not touching the amplitude whereas this one is
harder to realize because it has a frequency selective delay different element gets different
frequency and even gets different amount of delay okay phase shift is same, so this is some
circuitry which is very difficult to realize okay so fine therefore.

(Refer Slide Time: 18:30)

x(t) → x1(t)
If you wish to employ a circuitry which does this transformation, so all we have to do is we
have to understand that to do this I have to take every frequency component and give a phase
shift which is π /2 okay, so original delay element will not do that you have to selectively take
maybe with a very narrow band pass filter, you have to take every frequency element and for
everybody you have to give π /2phase shift okay so that is what you will have to do it means
generally it is a difficult circuit to realize there are techniques.

But it is a difficult circuit to realize but we will see that there is a way to actually bypass this
and when first try to appreciate that where this is since why we are discussing this of course
we started talking about the bandwidth efficient modulation, so why this is required so will

283
now will probably try to appreciate that part okay , so first let us go back to my original
proposition.

(Refer Slide Time: 19:32)

M( f ) = M+( f ) + M−( f )

{M−( f ) = M( f )U(−f )
M+( f ) = M( f )U( f )

That a signal M(f) which is the frequency domain representation of some M(t) okay so let us
say this is my empty this signal can be represented as this Plus this is my M + (f ) and this is
my M – (F) right where I represent M (F) as M + (F) plus M –( F) right, now if I wish to do a
from D s be a bandwidth efficient one which is generally originally called as SSPs see that
means signal sideband suppressed carrier so what I try to do is because I know one of them is
actually giving me the information so I wish to keep only one of them.

So when I do modulation at FC I just put one of them in the negative ass what should be
remaining it should be the other one because it needs to be symmetric otherwise the signal
will not be real and I cannot transmit a means complex signal in the air right that is not

284
possible I cannot even generate a complex signal this is just representation complex signals a
representation okay.

So if I wish to do that I should either represent it this way or I can represent it in another way
okay, so this is actually termed as upper sideband and this is termed as lower sideband
because at the positive half you can see this is actually the lower portion of it okay that is the
negative part of it so that is why this is called the lower sideband and this is where you are at
the positive half you are actually transmitting the upper sideband because there are two
sideband.

So this I call as upper sideband and this I call as lower sideband so here at the positive half of
the frequency I am transmitting either upper sideband or lower sideband, so any of these two
are valid modulation if I can do this modulation first of all I have to be convinced that this
modulation is possible first let us try to see the mathematical treatment of it so what is this is
actually let us first try to represent this M + (F) okay and then we will try to represent this
one so what is M +( F).

If M (f) is x U(F) what is U(f) is nothing but this right so less than F less than 0 its 0 almost
like U(t) okay it is a unit step function and greater than F it is 1right so if I multiply this 1
with this immediately what do I get I get only the positive of the M +( F) must be M(F) into
us similarly M –(f) if you –f is defined as this then it should be M(F) into u – f, now this you
F the good part is can be represented with respect to Signum function because Signum which
is nothing but.

(Refer Slide Time: 22:55)

285
 1
u( f ) = [sgn( f ) + 1]
2
1
M+( f ) = M( f ) [sgn( f ) + 1]
2
1
= [M( f ) + M( f )sgn( f )]
2

2[ ]
1 1
= M( f ) + (−j)M( f )sgn( f )
−j
1
= [M( f ) + jMh( f )]
2

This thing okay if I add 1 to it then what will happen this is 1 this is - 1 if I add 1 this will go
to 2 and this will be 0 and then put 1/2 so if this is my Signum f if I write Signum um f +1
into half that must be my you f right so you f I can write this way why I am doing this
because I want to deliberately bring that Signum f because I already know that J s f is the

286
representation of Hilbert transform or - J Signum f so I am deliberately bringing it okay, so
therefore now I can write M + F which we have written as M( F) into u( F).

So that I can write now as MF into 1/2 Signum f +1 right this I can write as half just M(F) +
M(F) * signum(f) right I can write this way this is MF and what is this so if I just introduce a
J over here - 1 by J means 1/ J should be taken over here, so I can write 1 by j into j
M(F )Signum s right if I introduce a minus term over here that should be minus 1 by j
multiplies above and down by j we will get this is actually –J2 that is plus 1so you get just J
over here right.

So I can write half this is M(F) plus or minus J this is m h(f) is not it there is a sign problem
M(F) s mess I need a - J or that should equal so this is what we get ,so immediately I can see
that this M + f can be represented as half of MF + J into the Hilbert transform of the message
C so this can be represented if you just do similar things.

(Refer Slide Time: 25:53)

M( f ) = M+( f ) + M−( f )

{M−( f ) = M( f )U(−f )
M+( f ) = M( f )U( f )

287
You can see also my M – (F) can be represented as ½ {m(f)-mh(f)}you can also do this okay
just similar signal we just have to represent it okay this way, so now knowing these two
things let us try to represent this t SSB with USB okay, so what is this part this part is just M
+ this is the plus half f translated at FC right so it must be M (f - FC) and this part what is
this is M -the minus half translated asked minus FC so that must be F plus FC therefore.

(Refer Slide Time: 26:48)

1
M( f ) = [M( f ) − jMh( f )]
2
u
ϕSSB ( f ) = M+ ( f − fc) + M− ( f + fc)
1 1
=
2 [M ( f − f0) + jMh ( f − fc)] + 2[
M ( f + fc) − jMh ( f + fc)]
1 j
=
2[ M ( f − fc) + M ( f + fc) ] 2 [Mh ( f − fC) − Mh ( f + fC)]
+
1
⇔ m(t)cos ωct − [Mh(t)e j2πfc t − Mh(t)e −j2πfc t ]
2j
⇔ m(t)cos ωct − mh(t)sin ωct

288
I can write phi suppose SSB for the upper sideband so u(f) this must be equal to M +(F – fc)
M – (f + fc) right this is the representation if it has to be a lower on accordingly, we will have
to do that okay, so this is pretty much here how I can represent my upper sideband SSB
modulated signal right now all I have to do this M + I already have expression, so I will have
to just replace that so M + I already know that M+ is nothing but half into MF+ j m h(f) right
so I will just write that so I can write half M + can be written as ½ [m(f-fc)+jmh(f-fc)]this
part again.

I will replace half M F plus s c- j MH s - fc right that is the representation okay, whichever
has j thermal will take them out, so basically what do I get I get ½[m(f-fc)+m(f+fc)]so this is
the term which is free of J and then if I just take - one by J common okay, so 1 by 2 J
immediately inside I will get MH f - FC- m hm s + XC it this is just algebraic manipulation
what is this is the modulation of empty or MF I should say, so if I just wish to now do a
Fourier transform or inverse Fourier transform of this one.

What do I get this must be represent as MT Cos wCT because that half time is it on okay
because with half it will be this okay so empty Cos w CT this will be represented and this
particular term that should be1/ 2 J is there and this is just for the sinusoidal one if I multiply
M HT if this m HT is the Hilbert transform equivalent time domain signal if I just multiply
my m HT into sin wCT I will be getting the corresponding Fourier transforms this one
because Sin? C will have 1 by 2 J and then you have those modulation term so it will just be
translated at + FC and –system.

So basically what I can say now by looking at this spectrum so here I was doing all the
manipulation in the spectrum because I know that SSB means it should be half of the
spectrum right, so I can do the manipulation in the spectrum and I have the understanding of
Hilbert transform so this is what I of having so initially that spectrum I have represented with
respect to us from there I went to Signum F I knew that with Signum F I can represent it with
respect to the Hilbert transform and.

Then from there I could do some algebraic manipulation and I could see that actual SSB
signal is nothing but this so that means you take the message signal multiplied by cost and
you take the Hilbert transform the message signal multiplied by sign means subtract these,
two you get your SSB signal so generation of SSB just comes from this we have seen the
frequency domain representation now in time domain we can immediately get the
representation okay, but then in the next class what we will try to do try to appreciate.

That this MHT we have already talked about that circuitry to do just 90 degree phase shift for
each component is a very difficult one we will try to devise a circuit which is known as wave
a circuit which does it very nicely, okay so we will try to see that and try to see the
corresponding mathematical representation of that thank you.

289
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-20: SSB - SC (contd.)

Okay so we have already talked about single sideband modulation right in the last class so
what we wish to do now we have already told that the single side band modulation we have
already seen the frequency domain representation and through Hilbert transform we could
finally get that the SSB signal if it is upper side band.

(Refer Slide Time: 00:44)

290
m(t)cos ωct − mh(t)sin ωct
It looks like this M(t) cos wCT –mH(t) sin wCT take that as a homework if we just do the
lower side band this will just become plus nothing else okay so this can be proven so either
upper sideband or lower side band it is just this separation of this plus sign or minus sign
okay so let us say we are just taking this so that means the generation is pretty easy you take
this empty okay take into two arms okay.

And then here it gives this Hilbert transform which is minus p/2 shift okay remember this
minus p/2 shifter for each frequency component we have already have one understanding of
that so you take that and then here you generate a cos wCT that if we put it over here multiply
it so that will generate empty into cos wCT that you take in one arm here you give some p/2
phase shift okay so that will generate cos wCT sorry from cos wCT sin wCT and then you
multiply so this will generate MHT into sin wC so here you will get MHT sin wCT and here
you will get MT cos wCT.

So if you wish to do upper side band you just add with this Plus this minus that will be your
SSB it okay so that is the easiest way we have understood so basically the modulator has
been just defined by knowing that mathematical understanding so you wanted to generate that
immediately understood that okay so it is nothing but this so I need to do this MHT so you
devise a Hilbert transform circuit once you have that it is just nothing but two modulators one
Hilbert transform right.

291
You can even tell that this is just the DSP SC modulator multiplication is nothing but the SJC
modulated so basically this two DSP modulator and you only have 1 π /2 phase shifter which
is the Hilbert transform and there is some things another adder is required that is all but we
have already talked about that this particular thing that π /2 phase sheet for each frequency
that is a very hard circuit to realize can we get some other circuitry which can do this.

(Refer Slide Time: 03:29)

So I will just give the practical circuitry that was proposed by waiver that particular circuitry
so the practical circuitry is looks like this so you have m(t) so this is one arm where you have
a multiplier you multiply this by cos w0T or 2π f0 we will talk about this w0 that is a or 2pF0
what is that we will talk about that and this one in the other arm you multiply by sin w0t okay
after that you put a low pass filter I will also talk about the bandwidth of the loop means cut
off frequency of that low pass filter once you have done that then you multiply with actual
carrier so this is cos wCT and this is where you multiply with find wCT and you subtract
them what I am saying I will be getting SSB signal.

As long as I turn this frequency properly and I put this w0 properly see the technique we are
applying that has a big historical significance and that is why probably SSB you cannot you

292
can generate for voice but you cannot generate for video signal we will talk about that where
you can generate SSB signal it is like this suppose I have a message signal which has no DC
part or close to DC component.

I have a typical message signal like that so let us say it looks like this any voice looks like
this voice has a minimum frequency so less than 300 hertz there is nothing okay that will not
significantly affect my voice signal okay so it has a minimum frequency let us call that as FA
and it has a higher frequency FB beyond which I do not need that so for Y is typically three
point four kilo Hertz right so 300 to 3.4 so FM is 300 FB is 3.14 kilo watt if the signal looks
like this what I can do the basic functionality of this is something like this that I will take this
signal to a small intermediate frequency okay.

And after taking it to small intermediate frequency probably I will put my filtering so it is
something like this why I am doing this because I have to somehow employ a filtering
technique to create SSB what I can do is something like this that I modulate it first let us say I
get a modulation something like this and then I employ a low-pass filter which has a cutoff
frequency exactly at this and I roll off like this then in reject this it will reject this or I employ
a high-pass filter which other transfer frequency transfer function like this.

Then I will reject this part and I will get this very nice for suppose a signal even like this I
can do that I do a modulation so it takes it over there and then I put our ideal high-pass filter
so this is where the problem comes do you have something called ideal high-pass filter an
ideal low-pass filter which has a sharp cut off you will never find that always any practical
filter which are stabilizable will have a roll-off right and as you go to higher frequencies the
amount of frequency it will take for roll-off will become higher and higher right so what is
happening it will have our own of once it has a roll of fit is actually disturbing the SSB
signal.

So I cannot allow that so basically I need to have a signal where I can still run this roll-off so
I was talking about this free zone if I have that free zone probably I will be able to allow this
roll off the float off of the filter I can allow this and I also know that at a very higher
frequency I can do this that I immediately modulate to wC whichever frequency I need let us
say 900 mega or something like that and at that part I only have a separation of 300+, 300,
300 this side and 300 this side.

So 600 hard within 600 hard if I wish to put my roll-off that will be very high order filter and
the corresponding filter transfer function will be almost non stabilizer they will not be able to
and you will not be able to realize this whereas what I can do is something like this at a lower
frequency I generate this filtering effect and then I translate it to a higher frequency I can do
that this is what has been done over here so if you see now the circuit it is clear you are doing
at a lower frequency modulation.

293
Then you are employing the low-pass filtering effect so that the roll-off is still realizable
after that you are modulating with a carrier right that is all you are trying to do so that is the
basic circuitry of wave a circuit okay but we will see he has done it in a little better way so
we will try to employ that part so what he has done is something like this yes so this is the
signal representation this is my FA this is my FB -FA and this is -FB.

(Refer Slide Time: 09:24)

fa + fb
f0 =
2
fb − fa
2

So the modulation he has employed is at a frequency this F0 is nothing but FA plus FB

divided by 2 it will be very clear why he has chosen this frequency okay so immediately if
you do that what will happen if I do modulate with this frequency so it will be translated by
f0 on this side and F0 on the other side right there will be a translation f0 on this side and f0
or the other side so if I just give pulse wave 0 translation let us say it shifts at the positive

294
frequency side then what will happen his F minus a plus F0 will be the frequency where it
gets translated.

So -FA+FB, FA+ FB/ 2which is nothing but FB -FA right so this frequency goes over here
this FA- FB goes to -FB+ FA+FB/2 so that is actually FA- FB/2 or nothing but minus of FB-
FA because they should FB is always greater than FA right so it just gets centered at zero and
this will be your new to frequency the spectrum looks similar so only sorry I have done it
wrong so this becomes FB- FA/2 this becomes FB –FA/2-1, 1 right and the separation
immediately.

You can see if FP minus is which is the separation of this one so whenever I do give a
translation this part shifts to the middle this part will go into F0 right so there will be one part
centered around 2F0 and when I give negative shaped what will happen similar thing will be
happening this will be centered around this and the other part will get shifted over here at
centered around minus 2 at 0 right.

Now all I will be doing the filter he has designed is actually the filter cutoff frequency is this
FB-FA /2 so that means he is trying to take this portion okay so that is the only meaningful
information he will be required so he is just taking this portion after the filter once he takes
that then he does rest of the things okay so this is what he is trying to apply so immediately
we can see that F 0 is designed as this and the filter cut off frequency is FB-FA/2 that what
has been employed so immediately.

You know for wise what should be our overall things right so that's something we already
know okay fine and remember the big advantage is now it is no longer a even a band pass
filter it is just a low-pass filter that I will have to put okay and that no pass filter is means it is
the lowest frequency where you can roll off the low pass filter because any other low-pass
filter if you just modulate it to a higher frequency the low-pass filter will still go to some
higher frequency and then you have to put the roll-off you wish to put the role of because the
role of region.

You have only 600 for voice right you want to keep it a slow as possible so that the roll-off
can still be easily understandable or realizable so this is probably the best one you can do any
other things you do your filter overall cutoff frequency will be little higher and the runoff
of600 hard might not be sufficient okay so probably the waiver circuit is an intelligent design
where is the best you can do you bring both the sideband overlapping in the middle but
because he has those two arms you will see that he has manipulated it very well after doing
the subtraction.

It will just generate with these two bands just generate the USB or LSB okay so we will see
that part now so if I just go back to his circuit that was his circuit so basically what we do we

295
take the MT in the upper arm we multiply it with cos w0 T or 2 π F0 T right as zero value we
already know now fA + SBp/2 so MT is or equivalent MF can be written as M plus F plus M
minus s that is the upper sideband and lower side better right.

(Refer Slide Time: 14:50)

M( f ) = M+( f ) + M−( f )
ϕ(t) = m(t)cos(2π f t)
1
= m(t) [e j2πfc t + e j2πfc t ]
2
1
ϕ( f ) ⇔ {M+ ( f − f0) + M− ( f − f0) + M+ ( f + f0) + M−( f + f0)}
2
1
ϕF (t) = {M− ( f − f0) + M+ ( f + f0)}
2

So what is happening I am if I represent this π T as my MT into cos 2π f0T just after the first
multiplication in the upper right so I can write this as M T into half e power j 2 π F 0 T plus
e-j 2 π F 0 T so I can write it this way now if I just do a Fourier transform which is p FT π F
right if I just take a Fourier transform of this so what that should be half it is a Fourier

296
transform of MP multiplied by this MP is already this one this one multiplied by this, this
gives the frequency translation right.

It will give a frequency translation of F my it will take it to it is like F- F0 right and -J2 π F0
will give produce F+F0 right so this will be if I just write M+ F minus F0 plus M-F- F0
because that is the whole MF right inverse multiplied by this that will produce these two
terms and M F multiplied by this that will produce another two terms which will be at plus F0
so M plus 0 M minus 0 right.

So we get this p F alright now what we have employed after pF we have put a filter okay so
in the filters which are the terms which will be vanished so this is the M plus one if you just
go back to our filter representation so this is so this gets translated to this side that is actually
multiplication by it is power J 2π F0T okay.

So there the negative 1 is getting survived see the negative 1/2 is I give a translation negative
1/2 is coming over here positive 1/2 is going over there this is the M plus part okay so I can
write this M + F- F0 and this is actually M+F+F0 and this I can write as M+ F plus so since
on the other side and this is actually M minus s plus s 0 right so these two terms are getting
rejected after the finishing so what I will be left with if I represent that as f F that should be
half into so two terms which are M -F+F 0 that gets cancelled and this other term which is
this, this gets cancelled okay.

(Refer Slide Time: 18:55)

297
 1
M+ (t + f0) − M− ( f − f0)]
2j [
ϕ′F (t) =

ϕF (t)cos(ωct)

1 1
⟹ ϕF (t) [e j2π fc t + e j2π fc t ] ⇔ [M+ ( f + f0 − fc) + +M− ( f − f0 − fC) + M+ ( f + f0 + fc) M− ( f − f0 + fc)]
2 4
1
ϕ′F(t)sin ωct ⇔ − {M+ ( f + f0 − f0) − M ( f − f0 − tc) + M+ ( f + f0 + f0) + M−( f − f0 + fc )}
4

So we are left with i/2j[mf(f+f0)-m(f-f0)] so you have understood this part similarly we will
have to do in the lower half also we have to multiply by sin means it should be whenever we
do multiplication by sin it will be 1by 2 J and here it should be this plus minus this right and
we will have to do the translation so if I do further with the time I will be getting if I just
write that as ? - after filtering it should be just 1/ 2 J just do it yourself algebraic manipulation
F+ M 0- M - f - f zero so we get this okay.

So now these two terms has to be multiplied with cos wCT + sin wCT so if I do cos wCT
multiplication so immediately what do we get so suppose the upper arm on trying to do so
that is corresponding time domain signals let us say ? FT which is having a frequency domain

298
representation we have already done that which is ? FF so this into cos wCT right this is what
we will have to do which is nothing but or I should say equal to ? F P into again.

We can write half it before J 2 π FCT + e- J2 π FCT okay so if I just employ same thing same
technique of frequency translation will get again this half and there is already one half so that
should be 1/ 4 and I will get a term of M +F+ f0-FC this is one term then I will get M- F- f0
-FC just another term so this is multiplication by FC and then -FC will give mean other two
terms which is M+F+f0+ FC and then I get M –F- F0 +FC.

So this four terms I get okay after multiplication with cost of course the Fourier transform of
that remember it is not equal we are doing the Fourier transform of that in the Fourier
transform because it is multiplication bye to the power J 2 π F 0 T it will be just translated at
frequency FC and -FC both the terms okay we had two terms M s and a minus and 10 T will
be happening when we multiply by minus J 2 π F 0 T so we will translate it to plus F 0 so M
plus that plus FC plus, plus X.

So it will have both the terms translated FC this is what we get for the first time in the second
down we have that ? F-T already that needs to be multiplied by sin wCT again do the same
thing okay sign it can put as 1by 2 J e to the power plus J 2 π FCT minus e to the power
minus J 2 π FCT so if you just try to write that and again do a Fourier transform so what you
will get is minus 1 /4 and then the terms will be if you just do that same translation so it
should be M plus you already had a component F plus F 0 that will be translated at -FC –M-
F -F0 –FC.

And then another minus M plus F plus F 0 plus FC and then you have M minus F minus F
0plus FC right so this is what we get okay now all you have to do is subtraction this to this
right because at the end after doing these two multiplication you are just subtracting from this
to this now you see some of the terms are getting cancelled because this minus this minus,
minus will become plus so you are actually adding these two okay so basically if you see this
M minus F minus F 0 minus FC that is plus so this term will be canceled.

(Refer Slide Time: 22:57)

And then this term will be canceled so what you are left with is to this terms and to this term
so 2 by 4 it will be which is becoming 1 by 2 and you get M plus F plus s 0 minus FC and
you have another term which is M minus F minus F 0 plus FC very nicely this is created by
SSB signals if you just see what it is it is the M plus part which is centered at frequency or
which is translated at frequency if I just because FC is big soma or FC is the greater one so I
can write it FC minus f0 right and this m- has been translated to f plus FC minus f0.

So the M minus part okay which is that lower part has been centered at C so M minus part n
minus part is centered at this FC plus f0 right so that should be M minus FC plus f0 okay so

299
this is where it comes this becomes the FC plus f0 the center part okay and this M plus part
comes to this again sorry FC minus l0 u minus and this gets centered at the M plus part
centered at FC minus f0 okay.

So you get up clear SSB representation but you could see that we have never employed any
Hilbert transform over here it is just filtering and that one filter that we have employed where
we can actually put roll-off provided there is that 600 hard free zone if there is no free zone I
cannot put that filter because I will not be getting even in low pass filter at a very low
frequency domain also I cannot get a sharp cut off right that is not possible there is no filter in
this world which can have a sharp cutoff so due to that only voice signal.

Where you have that free frequency band at a large or around DC you can actually employ
SSB whereas for video which has the DC term and all the frequency bands in the lower
frequency zone are occupied you cannot really employ any SSP so SSB modulation is not
possible for video signal whereas SSB modulation for voice signals that is possible okay
whereas if you see the Y's band is pretty small it is just 300 hard to 3.42 so they are probably
we will not be benefitting by reducing this whereas in video band it is almost like four point
five megahertz.

If I just modulate it with DSPs C I will be getting almost 9mega but I know that SSB I will
not be able to employ it over there because it occupies the voice band sorry video band looks
like this so no way I can put up filters like this whatever I do I will never be able to employ
this filtering it so there is a big difficulty in this so I will never be able to work out that in
video so can we get any other modulation that might not give me as efficient as SSB but
some efficiency where I can still have a filter roll-off.

That is where the modulation technique called vestigial sideband gets popular okay and we
have just seen the end of probably analog video transmission so, so far it was analog video
transmission even in many places I think it is still analog video transmission if it is analog
video transmission it is employed with vestigial sideband okay so the vestigial sideband will
see that there is a filter roll-off till we can employ and it might not be as efficient.

As the SSB that means it just occupies half the frequency but it will still be good enough
okay so what we will do next is we will try to see what is this vestigial sideband modulation
okay that we can employ for a video signal well after doing that we'll also try to see how do
we do modulate see signal as well as the SSB signal because we still have to know this is all
about modulation right we have not talked about demodulation so we will talk about
demodulation of these signals later on okay thank you.

300
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-21: VSB-SC

Okay so in the last class probably we have discussed about SSB of course suppressed carrier, so
we started already exploring modulation technique which are bandwidth efficient. So you have
already shown that whenever we do a modulation in the DSB SC we have shown whenever we
do modulation basically the bandwidth gets doubled.

And because in the baseband of the spectrum symmetry in the positive and negative half the
double band width will have the upper sideband and lower side band which are just mirror image
to each other. So means immediately it is clear that we don't have to transmit that whole
information the information is already contained in one of the side band, so either upper sideband
or lower side band.

So it will be good if we can just transmit one of the side band and somehow devise a technique
where we can modulate in such a way and also we can demodulate it so this was our target right
and in that process we have shown that yes it is possible, it is possible through a transform called
Hilbert transform, so if you do Hilbert transform.

Then you can represent this particular signal as some mt cos wct plus mht sin wct of course it Is
plus or minus so accordingly will be getting either lower sideband or upper sideband. So that's
actually we have proven from the spectrum that if we have SSB Sc signal that can be represented
as this where this mht is nothing but the Hilbert transform that means mt passed through a
transformer which is the Hilbert transform and we get mht correspondingly okay.

So that is what we have already shown that this is possible so all we have to do is take the signal
do a Hilbert transform multiply the original signal by cos Wct and the Hilbert transform signal by
sine wct through our adder either add or subtract accordingly we will get your means overall

301
representation of upper sideband or lower side. So this is what we have already shown that is one
way of defining it but we have seen the Hilbert transform is not that easy to realize, what we have
seen that actually the Hilbert transform is almost like it keeps the amplitude same, so whenever it
passes through that transform it keeps the amplitude same for phase it gives for every frequency
component it gives pi/ 2 phase shift.

That's something we have discussed that which is difficult to achieve okay, so that is why we
have devised another method which is called wavers method, which is means selective filtering
method that means we just do the filtering have means to do a modulation in such a way that in
the low means we put a low pass filter at the low frequency domain we can do the filtering and
then we can translate it to the frequency where we wish to put it okay. So that way we can
generate this kind of signal but for that we require that the signal has zero component or zero
power around the frequency zero, so that means no DC value and around that nothing is there
because that is where the filter roll-off can be put over okay.

So that is required so there are two methods to actually modulate or generate SSB signals this is
something we have seen and then for demodulation it's pretty easy we have seen that also that
this particular signal that phi SSB SCT that we have got you just multiply this with a cos wct
immediately you will be getting a term which is related to baseband of that signal.

And then there will be terms of higher frequency cos 2 Wct and sin 2 Wct you put a low-pass
filter those things will be rejected all those higher frequency term you can get extract your empty
but for that you have to remember that I need to have a local carrier okay, which is completely in
synchronous in phase and frequency with respect to the incoming signal means carrier frequency
and phase.

That is absolutely required we will try to see whether that Is possible or not okay, so this is
something we have discussed so far. What now would like to discuss that is there any other
demodulation technique that is possible in SSB, SC or I should not say SSB SC is there any other
demodulation technique if I add carrier to it okay so all we are trying to do now.

(Refer Slide Time: 05:13)

302
 ϕSSB(t) = A cos ωct + m(t)cos ωct ∓ mh(t)sin ωct
= [A + m(t)]cos (ωct) + mh(t)sin ωct
E(t)cos θ(t) E(t)sin θ(t)
E 2(t) = [A + m(t)]2 + [mh(t)]
2

We are calling this as SSB signal that means no suppressed carrier so we will be adding carrier to
it a cos wct and the original SSB signal which is if my message signal or modulating signal is mt
so mt cos wct minus or plus depending on whether upper side band or lower side band I am
taking mht sin Wct this is the representation of SSB, I am just adding carrier to it at the
transmitter end I can always add the carrier because that is where I have the local carrier which,
with which we are modulating it so I can always add the carrier to it okay.

So this is what I will be now transmitting through the air instead of just transmitting SSB so that
is why we are calling it SSB only SSB no suppressed carrier, so carrier is already over there, so
that this particular signal I can re represent as A plus mt cos Wct plus or minus I am just taking
one of them so mht sin Wct okay so it is a cos term and a sin term I can represent these two terms

303
as another some amplitude, let's say Et into let us say cos another theta t whatever phase that that
is that might have some frequency term phase term whatever it is overall phase if I just say theta t
I can always write that this is my cos means Et cos theta t and this is my Et sin theta t.

I can write that way and then immediately I will be getting another co sinusoidal term that you
can already see that the way I represent it accordingly I will be getting another co sinusoidal term
of which the amplitude will be this square plus this square right means that Et 2 okay so whatever
that Et so this I can write as Et cos thetaT and this let us say I write ET sine thetaT I can always
write that where as what will be the Et so if this is equivalent to this is equivalent to this if I just
square this and add that should be ET 2 so therefore Et 2 should be A plus mt2 plus mht 2 okay.

So what does that mean, that means that same signal I can actually represent as a Co sinusoidal
signal and this Et becomes the envelope of that co sinusoidal signal that means it is a co
sinusoidal signal whose amplitude is varying with time, now it is no longer a constant term it is a
time varying amplitude similar like DSB or amplitude modulation we have done so there also
there was a means carrier and the amplitude was modulated right, so same thing is happening
over here so Et is the envelope which is carrying that message signal in some way it has a term
which is equivalent to this mathematically and of course there will be a phase associated with it
theta t but I am not bothered about theta t right now.

Because what we are trying to do because we have added the carrier we want to actually do
envelope detector okay, so carrier detection we have already seen in amplitude modulation we
could add carrier in such a way that with some condition that the carrier strength should be at
least bigger than the lowest amplitude okay, modulus of that.

So we have tried to do employ the same thing we are adding carrier hoping that we can do same
thing there it was very simple it was just something into co sinusoidal here we have constructed
that co sinusoidal because we want to just track the envelope of that co sinusoidal that is what we
are trying to do. So now let us see if the envelope detector gives me the signal back that is what
we are trying to do.

(Refer Slide Time: 09:39)

304
 [A + m(t)]2 + (mh(t))
2
E(t) =

= {A 2 + 2Am(t) + m 2(t) + mh2(t)}

Y2

1/2
2m(t) m 2(t) mh2(t)
{ A2 }
=A 1+ + +
A A2
m(t) < < A
mh(t) < < A
1

{ A }
2m(t) 2
≃A 1+

{ A }
m(t)
≃A 1+ = A + m(t)

305
So if that is e square then what is Et because this is the envelope I will be detecting through a
diode detector if I just put a pass it through a diode detector and then charging discharging the
way I have designed it so if I just do that, that should be vA + mt 2 plus mht 2 so this I can further
write this to the power half A square plus 2Amt+ m 2t plus mh2 t okay,I take A out so A2 out
because there is a square root so it should be A so I get 1plus 2 mt /A plus m 2 t /A square plus
mh2 t /A 2 to the power half now I will try to put some approximation okay means where I can
actually detect mt.

So I know that the square terms are bothering me so I need to somehow take this out when I can
neglect this if MT is much less than A and also you can see Hilbert transform what it does every
frequency component it actually gives the same amplitude so if MT is much less than T I can also
write MHT that should be also much less than A so this is something I can write because in the
amplitude of Hilbert transform it's in the transfer function.

The amplitude actually gets translated with unity gain every frequency component so I can write
this safely if this is the case then what will happen this quadratic term I can neglect if this is
happening MT divided by a and you can also see because I want to put this that is why I have
taken out a okay so approximately this I can write as a 1+ 2 mt so the linear term remains and all
other quadratic comes are ignored now what I can do again I can do a another approximation
because MT by a again I know that all higher order terms can be neglected.

So this to the power half I can put in Taylor series okay and all higher order term starting from
quadratic to cubic and all those things I can neglect and this will be approximated as a 1 plus
there will be a half factor coming because the second term will have half so this 2 will be
cancelled I will get 1+MT divided by a ok so that's again approximately this immediately you can
see this is actually a+Mt I just have to put up DC blocker And I get my mt back so the condition
is in SSB also like AM, like DSB also had a version AM, where I was adding carrier right and
with carrier I was able to do my receiver very simple or do my receiving very simple in a very
simple way so I was able to do that what I was doing I was just putting envelope detector
followed by a DC blocker same thing I can do over here but there is a condition

For A M also there was a condition right that modulation index has to be just less than 1 or I need
to say that this mod of Mt ok at least the negative part of it and modulus of that must be less than
a this condition has to be satisfied always ok I was having this condition less than equal to now I
have a condition which is much more stronger what does that mean we have already seen that this
was when it was taking equality with the tone modulation we have proven that at that time I had
the energy efficiency which is 33% anything more or mu lesser than 1 the n energy efficiency
will be lesser and lesser

306
Here what is happening I need to choose a which is much bigger than mt so therefore the
modulation index corresponding to this or I should say energy efficiency which will be much
more Watts because I have to really put very strong carrier so I'm just wasting power by putting a
very strong carrier because otherwise I won't be able to detect this okay because for detection I
have already seen mathematically I need to have that approximation if that is not there all
quadratic term and cubic term will be coming into picture and that will actually distort my signals
okay.

So on top of mt there will be M square T and M cube t and all those quadratic all higher
harmonics terms will be coming and mixing with the signal that will distort my signal so I cannot
really afford to do that if I wish to detect I need to have this condition and immediately the power
efficiency goes away ok so basically as we have stated that there are no free lunches SSB the
bandwidth efficient scheme that's very nice but in SSB if you wish to do the receiver technology
little bit simplified and you are a carrier and that carrier strength or carrier power has to be very
high so you pay in power as penalty right

And later on when the carrier recovery circuit will be designing you'll also see for SSB and the
other version VSD that will be discussing probably later on you will see that for them carrier
recovery is almost impossible you cannot really do carrier recovery for this bandwidth efficient
schemes it is very hard to do that will prove that so if that is not possible then definitely we have
to take this particular thing where you do not suppress the carrier and immediately the power
efficiency goes for a toss so we have two options either you make it bandwidth efficient but then
power efficiency will not be that good or you make it power efficient then bandwidth efficiency
will not be that good.

So this is what we could see so far we have discussed something about single sideband now in
single side band we have also talked about that it is only good if you wish to really do this
receiver circuit sorry transmitter circuit will be simplified like the wave or circuit okay where you
can just put filtering approach for that you need to have a spectrum of the modulating signal
which is devoid of any component around DC okay.

So that's pretty much required but we have seen that for voice that's true this is this fortunately
this was happening so people could employ SSB in voice but for video the spectrum looks like
this so there are strong term around DC value and you cannot ignore them so that is not possible
to ignore this if you will ignore that probably signal will be distorted so you, you are not means
for modulation you cannot really distort the signal that's something you is not allowed for
modulation.

307
So that this is the case SSB is almost impossible we have said if we wish to do it through filtering
that filter should run like this sharp cut off which makes it non realizable in reality so it will not
be stabilizer right so what's the technique that we should employ so here will probably do a
compromisation will do will take some amount of bandwidth efficiency means we'll compromise
on bandwidth from SSB but also will gain little bit on power will show you that demonstrative
that okay.

So that is the vs BSC that means whenever we do VSB so if you just demonstrate suppose this is
my MT and corresponding MF which has a bandwidth B if I do DSB then at FC and minus FC
this will be modulated and it will occupy band from minus B to means FC minus B to FC plus B
so almost 2B bandwidth right if I do SSB then I will be taking if I take upper sideband I will be
just taking one of the side band the upper one okay so this is at FC minus FC again this is B.

So it's bandwidth efficient instead of 2B it's this and because I cannot create this if something
around zero is already there what I will be doing I will try to create something which is like this
okay so basically what I am trying to do I am trying to suppress some portion in my valid band of
USB.

(Refer Slide Time: 18:53)

308
And I am including some portion in the LSB okay and with that I am not including that whole
band I am NOT going running up to B it will be somewhere it will be ended so my overall
bandwidth will be this much let's say 70% 75% of the overall 2B so I still I am benefitting by
saving 25% of the overall band okay where as from SSB compared to SSB I am taking extra let
us say band okay so this is what I will be trying.

(Refer Slide Time: 19:29)

309
And if I wish to do that the corresponding filtering you can see that might have a nice roll off
because I do not have to do a sharp cutoff at that frequency so this is called the VSP or we call it
vestigial sideband okay so it's a little bit distorted sideband we are taking and we are actually
taking a mix of both side band both way distorted but you will see that will in such a way we will
design the filter so that we can actually get back the original signal right.

So basically what we are trying to do is something like this again will employ the same SSB
technique we will try to appreciate it because it's being appreciated at the frequency domain we
will try to appreciate it from the frequency domain. And then go back to the time domain
characteristics okay so it's actually nothing but the modulated signal means or we can say the
DSBSC that is the modulated one followed by a band pass or sorry a high-pass filtering okay
either high pass or it can be a band pass also you can you can end it later on so the upper one it's
not a problem anywhere you can end only the lower part the filter has to be properly designed
okay so we can say it's modulated signal let's say which is MF+FC and MF- FC this is just the
spectrum of the DSBSC followed by a filter transfer function.

310
The filter transfer function we know that it has a either high-pass or we should say a band pass
characteristics okay it might be high-pass we can just run it start from here and run it for the
entire these things but later on we will also understand if we try to do that probably that's not
really required we do not have to really include that entire band so what we can do we can just
end it somewhere so we can just put a band pass filter with the typical characteristics.

So this band pass filter we are defining as hif so basically to generate my VSB signal all I have to
do I have to modulate it followed by a band pass filter right now let's say I need to so this is my
modulated signal after modulating I wish to because VSB and SSB almost looks like similar
thing so what I wish to do my demodulator should be almost similar if I don't add carrier so it
should be something like this, this modulated signal will be transmitted over the channel.

And then at the receiver I will first through antenna take that I will multiply this with a cos wct
okay followed by some low-pass filtering that is what we do because whenever you multiply
there will be a means baseband term as well as some high frequency term at 2 FC or minus 2 FC
you want to reject them so followed by a low-pass filters the same technique I wish to employ but
while employing them I need to understand what kind of low pass filtering I will be putting so
that it is consistent with this h I f which I am putting that band pass filtering I am putting at the
transmitter and I get my signal back that is what I am trying to do now.

(Refer Slide Time: 23:33)

311
 ϕVSB( f ) = [M ( f + fc) + M ( f − fc)] Hi( f )
ϕVSB(t)
e(t) = ϕVSB(t)2 cos (ωct)
E( f ) = ϕVSB ( f + fC) + ϕVSB ( f − fc)
M( f ) = E′( f ) = [ϕVSB ( f + fC) + ϕVSB ( f − fc)]H0( f )

So what I will do suppose corresponding to this signal I have VSB T so it's just the Fourier
transform or inverse Fourier transform of this okay so whatever this is right now I am not
bothered about it so by detection or demodulation will be suppose it's after multiplication I will
get this et so that should be Phi VSBt multiplied by a co sinusoidal I am just deliberately putting
a 2 amplitude 2 so that is just a constant it does not matter you will see that this 2 is required right
this is what I wish to do so immediately the frequency domain how it will look like so if I just try
to plot EF how this should look like so that should be Phi VSB now 2 cos Wct if I have given So

312
it should be plus FC minus FC so this should be if F plus FC right so this is what we get after
modulating this now what will do we have to follow means pass this EF through a low-pass filter
so let's say that has a transfer function of HOF so then whatever E'f I will be getting that should
be this Phi vSB signal or this multiplied by co sinusoidal that must be pass through this so that
should be vsb f plus FC + Phi VSB f - FC multiplied HOF right and what do we expect and if I
design my HOF correctly this should be my M F right so I should get my signal back because I
am doing the demodulation.

So in the demodulation I will be getting my Phi VSB like SSB or DSB I will multiply by cos and
then put a low pass filter I should get my signal that is what I have done for all other modulation
so same thing I should expect over here okay so that should be MF now Phi VSB I already know
the characteristics right so that I will put over here and also I am aware that this is a low-pass
filter so any higher frequency term that will be created low-pass filter will be able to reject them
right. So let us try to put that so what do we get I will get so let us just in this Phi VSB f plus FC,

(Refer Slide Time: 26:06)

313
 {[ ( }
M f + 2fc) + M( f )] Hi ( f + fc)+ [M( f ) + M ( f − 2f2)] Hi ( f − ft) H0( f )

M( f )[Hi ( f + fc) + Hi ( f − fc)] H0( f ) = M( f )

1
H0( f ) =
Hi( f + fc ) + Hi ( f − fc)

Let us replace by f plus FC and for F minus FC let us replace F minus FC and put it over here that
is all we will be doing so that should create MF plus 2 FC if I am considering F plus FC and then
the next term which is MF minus FC and F minus FC plus FC so that should be MF followed by
H I it's F plus FC because it was F.

So it should be F plus FC and the other term should be M F minus, F plus FC minus FC so that
should be MF + Mf-2FC and h if minus FC right this is the, this is this EF part ok just by
replacing this over here ok so this is the EF part in EF I have just substituted this and wherever
required in place of f i put s plus FC or F minus SC and I have substituted that multiplied by HOF
must give me MF this multiplied by HOF but now we have to see whenever I multiply by HOF
that's how the higher frequency term will be canceled look at this term where it is going it's
actually MF shifted by 2fc in the negative half.

So this must be cancelled because this is followed by a low-pass filter so low pass filter must
cancel this same thing will be happening over here this must be canceled so what we'll be left
with is M F Hi f + FC plus again M F so I can take common MF + H I F - fc *HOF= M (f)
because that must give me after passing through low pass filter I must get back my signal so that
is the condition and M f gets cancelled I get a relationship between my input filter and the output
filter or I should say filter at modulator. And filter at demodulator I get a relationship among them

so I can just write my output filter should look like this right so that should be the output filter
given the input filter okay so this is the relationship which will be happening okay now what
we'll do we have just created a relationship between the filters we have still not seen how the vsb
will look like means how do we characterize this filter so all those things we'll do in the next
class thank you.

314
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-22: VSB-SC (Contd.)

Okay all right so we have for in case of VSB- SC we have already seen the input-output filtering
characteristics and their relationships right, so let us just assume one particular filter.

(Refer Slide Time: 00:33)

315
 1
H0( f ) = |f | ≤ B
Hi ( f + f0) + Hi ( f − f0)
H0( f ) = 1 |f | ≤ B
Hi ( f + fc)+Hi ( f − tc) = 1

So we have already seen this relationship H0(f) is 1/ Hi(f+ fc) + Hi (f – fc) right this is something
we have seen okay but also remember this particular criteria is required only when I am within
my band because what I was doing,I was trying to detect my signal which is m(f) which is valid
up to B so this filter characteristics must be valid up to B after that whatever happens I am not
bothered as long as it is a low-pass and it is rejecting that higher frequency term so as long as this
is happening I know that things will be all fine okay.

So because I have to only concentrate on that band, rest of the portion are not that important okay
because anyway nothing will be there okay so that is what we are trying to see. now if we just
employ this method that within this band this H0 (f) is a very simple filter it is a simple low-pass
filter okay ideal low-pass filter let us think about that so I can say that H0(f) must be one within
this band for low pass filter it must be flat otherwise the signal will be distorted my m(f) will be
distorted right.

So H0(f) must have unity gain within that band of interest outside that band it might be anything
so I am not that much bothered only thing is that it must be having a low-pass characteristic so it
might be something like this or it might be also something where the roll-off is there so within
the band as long as it is unity I have no problem so that must be ensured rest of things are not that
much important to me okay because it is all defined within that band only my actual signal was
within that band. So I all I have to do is within that band it should be alright okay.

All the other places what is happening it is not very important but one thing you need to keep in
mind that it should not be in means spreaded in such a way that that 2fc term is coming back, so
it must be having that low-pass characteristics and it is not taking that very high frequency term
which we have rejected while doing this calculation right.

(Refer Slide Time: 03:09)

316
 {[ ( }
M f + 2fc) + M( f )] Hi ( f + fc)+ [M( f ) + M ( f − 2f2)] Hi ( f − ft) H0( f )

M( f )[Hi ( f + fc) + Hi ( f − fc)] H0( f ) = M( f )

1
H0( f ) =
Hi( f + fc ) + Hi ( f − fc)

That relationship we have got by saying that these two terms we can actually cancel so it should
not be taking these two terms, so whatever it is it must be means satisfying that criteria as long as
it is doing that I have no problem okay.

(Refer Slide Time: 03:27)

317
 1
H0( f ) = |f | ≤ B
Hi ( f + f0) + Hi ( f − f0)
H0( f ) = 1 |f | ≤ B
Hi ( f + fc)+Hi ( f − tc) = 1

So if I assume this so within the band I can write that immediately Hi(f + f c ) + Hi (f – fc) that
must be 1 so this happens to be one of the desired criteria of designing my input filter okay output
filter now is becoming very simple it is just a low-pass filter as long as it is flat within that band
of interest it is alright so if I just take that then I can see that input filter has to be properly
designed whatever it is it must satisfy this thing.so let us try to see first what is this input filter
this is how that should look like, okay.

(Refer Slide Time: 04:16)

318
So what you are saying that this input filter is a band pass filter right so suppose I have a DSB –
SC modulated signal okay now I want to put a input filter so what do I do I put an input filter
which has a characteristic something like this okay and again it should be symmetric because it is
a real circuit suppose I have a characteristic like this now here I can see the slope is not very
sharp and it will of course distort the signals but what will happen let us say this is fc and this is -
fc.

(Refer Slide Time: 05:09)

319
Now you shift, you put this thing, this criteria right.

(Refer Slide Time: 05:12)

320
Let us say this is 1/2 and this is also 1/2 or I should not say 1/2, let us say, this is 1 okay so what
will happen now I have to shift this two, so what all I have to do is this H whatever H I have I
have to shift it to + fc and shift it to- fc and add these two and then I have to see whether it is
giving me 1 then only I can say that this is a valid filter for me okay the way I have adjusted it so
this frequency and this frequency will be all similar right.

So if this is fc + some delta this must be also delta if this is also delta this must be also delta okay
so this thing has to be all properly fixed okay so if this is the case now shifted by fc so what will
happen this will be completely shifted by fc and you shift on the other side by fc so they will
actually, sorry, slope should be all same I will actually over lap like this that is the 0 this, these
are deltas now because it is a constant slope if you add this two they will all give 1.

So if you just add these two things will just give one so it becomes flat and you can now see from
– B to +B it remains flat, so that means this criteria is being fulfilled if I design my filter this way,
right, so there now you can see that whatever relationship that we have got over here I can
actually have a realistic filter okay where there might be some roll-off okay of course it needs to
be linear or maybe some mere linear part I can just select and that will still suffice okay.

So I will be suffering probably the filter will have some this thing so that will create a little bit of
distortion but as long as the slope is almost tracking whatever I wish to do it will almost give me
similar result so if I can do that I can immediately see that if I put that Hi(f) + fc and Hi (f) – fc, I

321
can always get back what whatever relation I have means to get my actual SSB signal back from
the DSBSC signal actual signal back whatever criteria I put I can actually apply that so now I
have seen that there is a realistic filter realization that can be employed over here and which can
still give me due to the filter designing still give me back my signal back okay.

So once we have done that let us try to see whether VSB -SC signal also has similar
representation as SSB signal that will be our next target okay.

(Refer Slide Time: 08:43)

ϕVSB(t) = m(t)cos (ωct) + mv(t)sin(ωct)

Mv( f ) = M( f )F( f )
M ( f + fc) + M ( f − fc)
1
2j { v
ϕVSB(t) = M ( f − fc) − Mv( f + fc)}
+
2
1 1
= M ( f − fc) {1 − jF ( f − fc)} + M ( f + f0) {1 + jF ( f + fc)}
2 2

322
So for that what will try to see that will assume that maybe that is the case okay so we will say
VSB( t) is nothing but some m(t) cos wct which is almost similar to a SSB signal and some mv(t)
sin wct okay so we just try to prove that there is a representation like this okay so we are just it is
almost like we are assuming this and then try to see whether we get something which are not
contradictory okay.

So that is what we are targeting so we are assuming that okay there is a relationship,
representation which is like this where this mv(t) is nothing but my message signal pass through a
filter okay or I can write mv(f) this is actually my M (f) followed by a filter called F(f) okay thats
the filter transfer function so some filter I will be requiring through which if I pass this M(f) and
of course we will later on see that this F (f) must have a low pass characteristics okay.

So it is a filter which I am aiming I still do not know I will be proving what this filter is so which
gives me this mv so far we are just assuming okay that this mv (t) can be represented as this
almost similar to our SSB, in SSB what was happening the signal was passed through a Hilbert
transform, so inverse transom has a transfer function like this filter and we were just passing it
through it same thing we are doing.

Because we have said that we want our representation almost similar like SSB that is what we are
doing okay so immediately what we can write this Phi VSB (f) we can now represent it in
frequency domain right so if MV(f) is written as this I can now write this as M(f) + fc+ M f - fc
of course because it is cos w ct so divided by 2 plus because it is sin so it should be 1 / 2j and I
will have this MV (f – fc) minus Mv (f + fc) right I will have this now Mv I can write as M(f) and
F(f) I can write that I can take these two in the same part.

So I will get finally mf – fc if I just take f- fc there will be a 1/2 term and I will have 1 –j F (f-fc)
if you just do a algebraic ordering and similarly if I just do M(f+ fc) I will get again another 1/2 1
+j F(f+ fc) I will get this right, so this is what we are getting so what is happening now this M (f-
fc) and M(f + fc) see this is my VSB(f) what was my earlier assumption it must be M (f –fc)
+M(f +fc) into Hi(f) that was my earlier assumption okay or I should not say assumption that was
the actual generation.

So this must be similar to that if that has to be the case then this and this both should be Hi(f)
right so this should be Hi (f) and this should be also Hi (f) so let us try to put that, okay, that okay
if this hasto be equal then this should be both of them should be equal to Hi(f) if I put that do I
get a contradictory result or do I get something which is what we have proven already.

(Refer Slide Time: 13:00)

323
 1
2{
Hi( f ) = 1 − jF ( f − fc)}
1
Hi ( f + fc) = {1 − jF( f )}
2
1
Hi( f ) = {1 + jF ( f + f0)}
2
1
Hi ( f − fc) = {1 + jF( f )}
2
=1

So if I just say Hi(f) is 1/2 {1 - j F(f-fc)} right so from here I can calculate Hif + fc right which is
nothing but 1/2 {1 – j that should be F(f) right and from the other relationship I can get Hi (f)
should be1/2 {1+j F (f+fc)} from here I can get Hi(f – fc) okay so immediately that will be 1/2 {1
+ j F (f)}now what was my relationship Hi f + fc and Hif - fc if I add these two that must be one,
add this two you can immediately see this will get cancelled and I will get 1/2 + 1/2 that is 1.

324
So it is not giving any contradictory result okay so the filter that I have designed it is actually
being consistent with that so therefore I can say my filter F is having this relationship okay, so it
is just that filter if I just pass it, pass my signal through this filter I will be getting the
representation which is already being proven for SSB- SC I get almost similar representation for
VSB- SC.

(Refer Slide Time: 14:40)

ϕVSB(t) = m(t)cos (ωct) + m v(t)sin(ωct)

Mv( f ) = M( f )F( f )
M ( f + fc) + M ( f − fc)1
2j { v
ϕVSB(t) = +M ( f − fc) − Mv( f + fc )}
2
1 1
= M ( f − fc) {1 − jF ( f − fc)} + M ( f + f0) {1 + jF ( f + fc)}
2 2

So that is why I can also see now things are all consistent I can also see if I just multiply this by a
2 cos wct immediately I get my signal back because I am multiplied by 2 cos wct put a low pass

325
filter I will get my mt back that is consistent earlier also we have proposed the same thing so we
can see that with this particular thing will be always getting back this thing right.

(Refer Slide Time: 15:10)

m(t) 2 cos(ωct)
So basically I can now put my modulator as this so I will be having m(t) sorry means for the
modulation what I can do that m(t) I can multiply by a cos first so let us say 2 cos wct followed
by Hi(f) the filter I have designed so the demodulator will be take this sorry take this one
multiply with because we have now seen that it is represented as this right it is represented as this
so if I just multiply by cos and followed by a low-pass filter I should be getting so multiplied by 2
cos wct followed by whatever that Hof which is a low-pass filter.I get my m(t) back okay.

so this is actually the VSB- SC modulation and demodulation it is pretty simple that only thing is
that this filter has to be carefully designed the way we have demonstrated in the earlier case so
that has to be very carefully designed so that at Hif + fc and Hi f -fc gives me 1 in all the places
only thing is that beyond that it is not that much required.

(Refer Slide Time: 16:38)

326
So whenever you are designing that filter you can always say that I need to have this after that it
can have any roll-off so let us say this is my fc this is B up to B I mean fc+B and let us say this is
my fc – fc - fc - B after that it might have any roll-off that is not a problem because if you just do
F + so this is the spectrum if I just do Hi (f + fc) and f –fc what will happen these two will add up
to 1 so up to B and - B it will be flat after that it will have this and this roll-off so no problem in
that, that can be anything only thing is that I have to ensure that whatever that filter that is flat up
to B beyond that what I put it is not a problem okay.

So that is how generally as VSD signals are being generated so basically you have to, all you
have to do is you have to apply this filter and you have to characterize this roll off properly
whatever roll off you wish to do.now regarding the roll off it depends on what kind of circuit you
can choose and according there will be a trade-off of how much bandwidth you take suppose you
want to have a little lesser roll-off okay.

Then the filter characterization accordingly will be simpler because you do not need to go for
very high, higher order filters, so what will happen it will take some extra band if you start going
towards higher and higher order filter the roll-off will be sharper and we will be saving more and
more bandwidth and you can see that at the end when you can make this roll off almost sudden

327
then it goes back to SSB so DSB is nothing but from DSB slowly you can move up to SSB and it
all depends on what kind of roll off you wish to give to your filter okay.

So depending on that your filter designing will be more challenging or less challenging but
whenever you have more challenging filter design you will be saving on bandwidth whenever
you have a less challenging filter design probably you will be, will not be saving bandwidth that
much okay so in TV spectrum probably they go up to the 50% of this band so they take over all
band if it is, 2B they take 75% of that, so that is what they do they save 25% of spectrum okay.

And also they put carrier you will see that why that carrier has to be put that we will discuss later
on whenever we talk about carrier recovery will be discussing more about that carrier recovery
means why carrier has to be put in SSB signal okay.

(Refer Slide Time: 21:32)

328
ϕQAM = m1(t)cos (ωct) + m2(t)sin (ωct)
x1(t) = ϕQAM(t) 2 cos (ωct)
= m1(t) + m1(t)cos 2ωct + m2(t) sin (2ωct)
x2(t) = ϕQAM(t) 2 sin (ωct)

So after this now we have discussed about two band-width efficient scheme one is SSB one is
DSB – SC sorry one is SSB- SC right now we will discuss another bandwidth efficient scheme
which is called clam QAM-quadrature amplitude modulation, so what is this is something like
that, let us say I wish to generate QAM so what I do actually instead of modulating one signal
now I take two message signals one is m1(t) and the other one is m2 (t) it is almost like
multiplexing I am taking this two message signal and within the band of 2B I wish to transmit
both of them as long as both of them are actually demonstrating similar band so they are similar
signals they might be having different spectra different time domain representation but the overall
band that is being considered for them are similar as long as that is happening I can actually
represent them as this composite signal.

So it should be m1(t) cos wct + m2(t) sin wct so that is exactly what we are trying to achieve
okay so through this motivation so in QAM what we are targeting is, we wish to basically
transcript two signals simultaneously okay if you just try to see the spectrum that it will be
occupying so if I multiply by cos Wct it will still have similar spectrum shift okay it will still be
occupying that fc +B 2 fc – d and the negative house and if m2(t) is also having B bandwidth if I
multiply by sin it will also be occupying same band okay.

So both of them will occupy the similar band and they will superimpose but just we have to see
because the way we are producing them we are multiplexing them in the same band so they are
getting superimposed is there a possibility that we can separate them out that is what we will have
to appreciate can we separate them out after this so modulation was fine I know that they occupy
if they are having similar baseband representation there will be occupying similar band but can I
really demodulate them and separate them out.

That something which will be will be discussing now so what we have to do is suppose I want to
modulate demodulate back m1(t) that is my target so what I do, I take this QAM sorry QAM
signal okay quadrature amplitude modulated signal Phi QAM put a band pass filter around fc of
2B So I will be getting the entire this composite signal okay I will now multiply this with
2coswct okay.

329
So I will just multiply with cos, co-sinusoidal so what do I get so this is I am calling this as x1 (t)
okay so what do I get now this will be multiplied by cos so that should be cos2 so I will have 1
m1(t) + cos 2 wct so m1(t) into cos 2 wct and this will be 2 sin wct cos wct so that should be m2
(t) sin 2 wct right.

Now a very nice thing has happened both of these things has gone to a higher frequency only
thing that is left which is at the base band is m1(t) I put up band pass filter I get this signal back
so basically I have option of demodulating only m1t from this similarly if I just multiply this by
sin you will see that I have an option of de-modulating m2 (t) back so if I produce x2 (t) which is
Phi QAM of course that should be T into 2 sin wct.

(Refer Slide Time: 24:16)

= m1(t)sin (2ωct) + m2(t)[cos (2ωct) + 1]

= m2(t)

330
Now what will happen this is the first term is now becoming m1(t) sin 2wct and then it should be
m2 (t) into 2 sin2 that's actually cos 2 wct - 1 or 1 minus right so 1 minus this right, so what I get
again this term that is a modulated signal again going at 2 wct so I can neglect this, this into this,
that can be neglected as well so I get back m2(t) after I do a band pass filtering, so basically I
have now option of demodulating both the signals okay simultaneously but the good part is both
of them are occupying B band so it is actually sorry 2B band so within 2B band I am able to
transmit two such signals it is almost similar means bandwidth efficient as SSB signal because in
SSB also if I had this 2B band I was putting 2 SSB signal whereas here I am probably occupying
a composite signal and it is having both of them right so m1 (t) as well as m2t and I know now
know that I can actually demodulate back both of them.

So equivalent I am able to transmit almost similar and similar bandwidth efficient modulated
signal as SSB so what will be the corresponding modulator and demodulator.

(Refer Slide Time: 26:06)

So I just have to generate a local cos wct right multiply that with our message signal which is m1
(t) and then from cos I have to produce sin so put a pi/2 phase shift multiplied with another signal

331
m2 (t) right so whatever you get you add them together and put it in the channel right so that is
the Phi QAM (t) once you receive that you put in two arm and then again locally you generate
two cos wct you directly multiply and then followed by a low-pass filter will be getting m1(t)
back if you just give a pi/2 phase shift.

So this will become sin multiply and then followed by a low-pass filter will be getting m2 (t) so
that is actually the x1 sorry that is the x1(t) and this is the x2(t) just after or before the modulation
right so this is what which will be happening if I do or the modulation right so modulation
demodulation is pretty simple right now, can we now say that this is probably the best modulation
scheme where right now we cannot give that answer but bandwidth efficient wise probably yes it
is equivalent to DSB.

And we can say that the transmitter part is not as hard as sorry bandwidth efficient wise it is
equivalent to SSP but transmission part is not as hard as SSP because we do not have to employ
any filtering like our SSP okay so either be it wave a circuit or any other way do not have to do
that okay no Hilbert transform is informed means involved in this right so that is also a good part
and we are able to modulate similar things but later on you will see that if this there is a carrier
recovery process because I am doing this demodulation coherently right.

We are means there is non-coherent and coherent demodulation probably I have not talked about
this terms but I have been keep on mentioning about it so non-coherent is like that envelope
detector means I do not need a local carrier I do not have to generate local carrier which is in
synchronous in phase and frequency to the incoming carrier I do not have to do that that is called
non-coherent that means the receiver part is pretty simple and I am not having any influence on
carrier generation okay whereas coherent is means I am subject to my receiver performance will
be subject to how good I am in generating the local carrier at the receiver something is coming
already with carriers because it is modulated so that same carrier I have to somehow extract so
that the phase and frequency are completely in synchronous with the incoming carrier.

And then with that local oscillator which is generating synchronous phase and frequency carrier I
have to demodulate it okay so whatever we have seen so far amplitude modulation was probably
the non-coherent means there I can employ non-coherent demodulation DSB- SC definitely
requires a coherent demodulation because I have to multiply by carrier SSB also if I do not put
carrier it requires a coherent demodulation and for VSB also we have seen that it requires a
coherent demodulation.

This one definitely I can see already that it requires a coherent demodulation okay so this
coherent demodulation if I have there are all possibility that whatever carrier I will be generating

332
that might have some phase drift or frequency drift compared to the incoming carrier and now
what we will try to do we will try to look into that part if there are drift what happens to all this
coherent demodulation part okay.

So if there are some drift in frequency or phase how I will be suffering what kind of suffering I
will get for DSP what kind of suffering I will get for quadrature amplitude modulation what kind
of suffering I will get SSB and DSB . similar things would be happening but specifically I will be
targeting these two and there you will see probably this is not as good as DSP- SC both of them
will have detrimental effect but probably this one will have more detrimental effect.

And will also be able to prove that whoever is bandwidth efficient like SSB or VSB will be also
able to prove that carrier recovery is almost impossible over there okay similar thing over here
also so carrier recovery will be impossible but the first thing, second thing is even if we somehow
get the carrier if you wish to demodulate if you have some phase or frequency offset you will be
having a detrimental effect over it so in the next class probably we will be discussing that part in
detail thank you.

333
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-23: Effect of Carrier Synchronization

Okay we have talked about this coherent and non coherent demodulation it is, it's part of the
receiver circuitry okay. We have already talked about that, that if we have to do coherent
demodulation we need to extract the carrier and what will be trying to prove today if that carrier
recovery means how important that carrier recovery is. If there is some error in the carrier
recovery what kind of detrimental we will have in the recovery of the signal. So that is something
will try to exploit today. So let us have see for a DSB signal.

(Refer Slide Time: 01:02)

334
 ϕDSB−SC = m(t)cos (ωct)
ϕDSB−SC(t)2 cos(ωct + θ)
= m(t)2 cos (ωct) cos (ωct + θ)

= m(t)[cos (2ωct + θ) + cos(θ)]

⟹ m(t)cos(θ)

So let say my DSB –SC which is nothing but m(t) cos Wct right. So that is a typical DSB signal
where mt is the modulating signal band limited of course it occupies in baseband up to band
width D and then we have chosen carrier which is wc much bigger than D okay. So that is the
DSB signal. Now let say what I will be doing I will be locally generating a carrier okay. To
demodulate it first I will take it from the channel through antenna and then locally generate a
carrier whichever way I get that carrier.

335
Let say I get that carrier I will multiply with this carrier right so generally my demodulation
process is something like this. DSB – SC (t) multiplied by 2 cos Wct okay so multiply by this and
then pass it through a low pass filter. Is what we do, low pass filter which as a bandwidth roughly
or correct frequency at D. so this is what we do, but let's say this carrier that I will be multiply
with it is not completely in synchronize. Let say that as a phase off set of Theta with a incoming
carrier okay.

So let say that that is happening that I have probably detect the frequency very nicely the
frequency is completely in synchronize but the phase that is off set of Theta okay. But that is the
carrier I am generating so I will multiply with that carrier so do I get okay so I can replace this by
this so I will get m(t) 2 comes over here cos Wct into cos (Wct+Theta) so that is 2 cos a cos b it
should be cos a+b into cos a-b okay. So I can immediately write m(t) this is cos A+B means cos
Wc t sorry 2 Wct+ Wct, 2Wct+ theta and + cos is Wc t +theta-Wc t so that should be theta right.

Now I will be putting a low pass filter so this term will be canceled because that's at frequency
2wct that 2wc so I will be cancelling out this term that will just of the term after demodulation I
will be getting m(t) cos theta. earlier when it was fully synchronized there is no phase off set in
the local carrier, I was getting only M (t) that was good. Now I am gettingm(t) into cos Theta
okay. So whatever off set we have, this is not a time varying term as you can see. So basically if
there is a phase off set I get almost mt multiplied by a constant term. So the message signal will
have similar means and duration in time okay. So the pattern of the message signal remains the
same. It is just get modulated sorry it' 8 just get multiplied by a constant term which is called cos
theta, the theta is the offset. The problem with this is whats the maximum value is of cos theta
that is 1 but I have a chance getting 0 over here.

Suppose that theta off set is pi/2 then I get 0 so what do I get, I get nothing so there is a danger of
huge attenuation, if I have a phase off set I do not know that phase off set will be random right if
somehow I choose a phase a I get a phase offset of pi/2 or around pi/2 I will have a huge off set
means huge attenuation of the message signal. The message signal pattern is not getting changed
because it is getting multiplied by a constant term.

So there is no problem in that message signal pattern is still remaining the same only thing is that
unnecessarily due to my demodulation there is an attenuation term which is coming up if my
phase off set is near to 0 I know that not much attenuation will be happening. But if it is near to
pi/2 I have a big problem okay. My message signal will the strength will go down heavily so this
one problem if there is a phase off set, now let see if there is a frequency drift okay.

(Refer Slide Time: 05:52)

336
ϕDSB−SC(t) 2 cos ((ωc + δ)t)
= m(t)2 cos (ωct) cos (ωct + δt)

= m(t)[cos (2ωct + δt) + cos(δt)]

= m(t)cos(δt)

So I will again have Phi DSB -SC (T) okay which is now will be modulated sorry this is already
modulated signals I get that okay. Now what should be means what will be doing will be now
demodulating it okay, so while demodulating I will be again multiplying by the local carrier, let
us now say this is modulated with wc but this local carrier has a frequency drift, so it is means I
could not track that frequency so I get a delta drift to the frequency.

So let us say wc is the targeted frequency but I get w c+delta or – delta whichever way you take
and there is no phase shift let us say if one term of that there is phase shift there will be additional
effect we know already what will happen if there is a phase shift, and let us say phase there is no

337
drift but frequency I could not track completely so there is a drift in frequency. So now I can
write this as m (t)cos(wct) cos(wct+delta t) right, and the two term is there again I will put cos
A.B so it should be m (t)cosA +B so that should be (2 wct+ delta t) + cos (delta t) right.

Let us say now what will happen there is a frequency term at m(t) modulated by some very small
frequency drift because whenever I am tracking the carrier I will not have huge drift I will have a
small drift, so there will be a small shift little bit of shift in the frequency so there will be a small
drift of that in frequency domain if I wish to track that so what will happen there will be a I
means it will not come as m (t) there will be a small drift of that right, so it should have
something like this okay, so there will be a delta drift and the other part will actually go into 2
wc+ delta and -2 Wc it means -2 wc-delta okay.

So that is why it will go if I put low pass filtering now it will just take this thing rest of the things
will be gone so I will be not getting this I will still get this which is nothing but m(t) cos delta t I
should not say delta t it is delta into t right. Now this term is no longer a constant term okay, what
it is actually the m(t) is now getting a modulated term with cos some small frequency delta into t,
so this will distort the signal.

So frequency drift is even more detrimental than the phase drift, phase drift my message signal
was almost similar okay, and I was getting something over there as long as the phase was not
going near to pi/2 because then the attenuation, it was just the attenuated signal and the
attenuation was not very big okay, so that way I was safe but here what is happening my message
signal is now getting modulated means whenever I demodulate I expect that I should be getting
message signal but it is no longer the message signal it has some modulation on top of this which
is modulated by that delta okay, the problem with this is as you can see in the spectrum because
this delta is small so there will be a small drift only so you will never be so this is centered at
delta and this will be centered at – delta.

So the positive shift and negative shift will super impose and they will actually generate a
different kind of signal or different kind of spectrum which will completely change the overall
signal quality and I will never be able to separate that out because a higher modulation will
always keep the signal shape intact or spectrum shape intact but this is a lower modulation so
they get overlap the negative half and positive half gets overlapped and I get something which is
completely different okay.

So I will never be able to recover my signal back from this if this happens, because I know that
the delta will be smaller and however small it is it will start distorting the signal okay. So that has
a huge detrimental effect in demodulation so the, I am just telling this because now we can
appreciate why carrier recover is so important and why carrier phase and frequency

338
synchronization is so important without that any coherent detection is off no meaning okay. Now
let us try to see what happens if I have this quadrature amplitude modulation QAM.

(Refer Slide Time: 11:34)

ϕQAM = m1(t)cos (ωct) + m2(t)sin (ωct)

x1(t) = ϕQAM(t) 2 cos (ωct + θ)

= m1(t)[2 cos (ωct) cos (ωct + θ)] + m2(t)[2 cos (ωct + θ) sin ωct]

= m1(t)cos θ + m1(t)cos (2ωc + m2(t)sin (2ωct + θ) − m2(t)sin θt + θ)

= m1(t)cos θ − m2(t)sin θ

So let us say, I have Phi QAM t which is m1 t cos wCT + M2T sinwCt that is what we have said
that will be this, now I will demodulating I have to demodulate it. let say I want to just
demodulate this M1 T so what I have to do? I have to multiply with a local co sinusoidal let us

339
say that has a phase drift. So I will be multiplying this Phi QAM t with 2 cos wCT + a drift of
theta okay. Let us say this is my X1 T what do I get? So it will be M1t 2 cosw CT x cos wCt +
theta right I get this + M2t 2 cos wCT + theta into sin wCT.

So all we have to do is cos A cos B and cos sin B so we put the formula of that so we will be
getting M1T cos Theta, A- B is this is A this is B + M1T cos 2 wCT + theta and will have M2T
sin 2 w CT + theta – M2T sin theta fine. I will be putting passing it through a low-pass filter these
two term will be gone what do I left with m1t cos theta – m2t sin theta earlier for DSB SC I was
getting this right, which was okay because cos theta is constant term now what I am getting?

I am getting this signal better Fourier signal coming from here also so now because I have
modulated it in such a way that they are actually co occupying the same band so I will
demodulating if I have a phase drift I will interference coming from the other signal. This is more
detrimental, earlier I was just getting a attenuated signal we have told that there are no free
lunches, you got of very nice frequency or band width response by QAM but if you have a drift
in carrier for DSB SC it just add some attenuation. Whereas for this one you will have a
interference term coming from other signal that you cannot do anything right now because they
will all mixed in the base band and you have no way to actually demodulate them any further
okay. So it will just interfere your signal and it will corrupt your signal so this is the problem that
QAM will have which again you will see that our SSB will not have that SSB will not have
probably similar problem.

But QAM will have a bigger problem coming some amount of interference from other signal
okay. so after knowing this that the importance of carrier recovery and importance of not having
any phase or frequency drift and which particular modulations scheme suffers from a phase or
frequency drift in what extent. So after seeing that let us try to see if for VSB or SSB signal if we
wish to actually demodulate it sorry not demodulate we wish to get carrier back what will be the
associated problem okay. So before demonstrating that let us try to see how do I actually get a
carrier back from a DSB signal double side band modulated.

(Refer Slide Time: 16:22)

340
ϕDSB−SC (t) = m1(t)cos (ωct)
So let us say Phi DSB SC signal which is nothing but MT cos WCT right this is by DSB signal.
So what I can do this MT cos wCt I wish to take the carrier so this will be suppose this is my MF
I do a modulation it goes over there okay so it has all the frequency term if I can put a very
narrow band pass filter centered around that fc whatever I will be getting that will give me means
the frequency term almost okay.

The problem is with this particular method if you just say that anyway this has been modulated
with carrier If I put a very narrow band I will be getting my carrier back that something I should
be getting but that has some danger first of all to put a narrow band, band pass filter you need to
know the frequency okay roughly probably you will be knowing the frequency but you might not
know it exactly so you will have to put at least some amount of bigger band over there.

But that not has detrimental as the next one I will talking about suppose this is a voice signal
what will happen? Will it have any DC value?

No, the spectrum will look like this once I modulate there will be nothing over there. So around
the carrier I have nothing so I put a band pass filter I get 0, I would not be able to detect any

341
signal. So even though I know there is a carrier involved in it but somehow the carrier is evading
me. I am not able to get that carrier right. So what I can employ over here is something like this.

(Refer Slide Time: 18:24)

m 2(t)cos2 ωct = 12 m 2(t) + 12 m 2(t)cos(2ωct)

I have this signal right this MT cos WCT even if this has no value around WC can I somehow
generates something around WC or WC what is the chance of doing that let us put the circuit
which is the squaring of it okay, so let us have MT cos WCT I square it so what do I get? Here I
will be getting M2T cos 2 WCT right which I can write as ½ m2 T + ½ m2 T cos 2 WCT okay if
it has some phase that also will be tract over there, so I will get this now if I put a band pass filter
this will be gone.

But for this one I have now a better guarante of getting something because what is happening I
am now actually suppose it was not having any DC term, it was having some spectra which was
devoid of any DC term or around that DC term. Now this is by MF what I am getting over here is
m2 T in frequency domain what will happen I will get convoluted right, so if I convolute this

342
signal with itself what will happen?It will have the highest peak around 0, okay so whenever I
convolute this signal and then plot it so around 0 it will be having very high value and after that it
will just get reduced. So, basically if I say M2t multiplied by this, this particular thing not the
actual spectrum will actually go over to wct right, so even if I never had anything for mf. I'll now
be having something not only that by squaring I ensure that, that gets convoluted so it will be
having peak over here which is expected. I want most of the power around the centre because
that's where I'll be extracting my carrier. This is just for carrier extraction. Remember, this is not
the signal demodulation. I take that part I'll just tap some power from there and I’ll try to actually
extract carrier from here. right so thats what I'm trying to do. So of course I can always do that, I
can multiply with this means, I can actually square this squaring is possible, any non-linear
circuit you can employ you can get square of that. so, and after that you just put a band-pass filter
around that.Now you don't have to for sure know the fc what is happening? because it'll have a
sharper peak so if you give little bit of extra band it'll actually thick that frequency that is the
major frequency component so it will thick that frequency not only that because this carrier is
already there 2Wct whatever phase it has whatever frequency it has it will take that so all you
have to do is you have to put a band pass filter and then average it okay, the amplitude has to be
averaged. So, from there carrier is already, carrier information is already coming over there so
m2t has nothing to do with it okay, so you will be getting the extract in carrier information if you
do a squaring right so that's pretty nice for DSB-SC I can see that even if it does not have any DC
component, even if I modulate it I don't get anything around wc but if I square it I'll be getting
something which is exactly mimicing the input phase and frequency because this 2 wct if there is
a drift in wc that drift will be reflected over there so I'll be getting the same drift over there. There
will be drift in phase that will be reflected over this 2 wct. whatever is coming I'm actually
squaring that only so that will have exact same frequency and phase and if I put a band-pass filter
it will just extract that thing okay, it's not become easier because m2 t is no longer 0 over there so
I will, I will get some amount of signal at 2Wc. all I have to do is if I need to get wc I have to half
the frequency. There are frequency means this divider and all those circuitry are already
available. o if I get 2 wc, I can always create wc from that. That is not a big problem there are
circuits where frequency halfing is possible. OK you can immediately with a comparator convert
into digital Signal with one and zero. And then you can do a frequency halfing by any counter
and all those things. OK that’s that is pretty much possible that is not a big problem so that I can
do then whatever I get that is exactly synchronous to incoming phase and frequency. OKs o this
is a technique whenever I do a squaring I will get my frequency back we will see also there are
some more things which is which will be the discussion of our next class probably that is called
phase locked loop we will come back to that but right ow we should be pretty happy with this that
we canot least extract the carrier. now let us discuss if it was a SSB signal. Can we extract the
carrier. DSB we can see. now let’s see we have a SSB Signal. so let’s say, we call that VSS b OK,

343
if I have the carrier VSSBT that I can write as m t cos omega c t minus or plus I can take one of
them for upper side band or lower sideband MH t sin omega c t. That is the representation of
SSB. VSB will also have similar representation. So we will see the similar things will be
happening in VSB. Ok

So now I can represent that in a similar fashion has we have done for envelope detection of VSB
with carrier so I can do represent m (t) and mh(t) has some et into cos some theta t and et x some
sin theta t right I can write that so immediately I will be writing E(t) this as cos wct depending on
+ or – I can write that as theta(t) right if this is-, it will be the fat. if this is + , it will be-theta t,
right, so I get this VSSB can be represent this now if I just square it, it's all about squaring earlier
we have seen with square we are getting things.

So now you square it [VssB (t)]2 what do we get ,we get E2 (t) cos 2 [Wct + or-theta(t)] take half
E2t I get [1 + cos [2 Wct + Theta(t]) I do get this, Now this is around 2wc so if I just put a narrow
bandpass filter I will be getting this thing so what we are getting again narrow band pass filter the
carrier that we are getting that's not to 2 Wct, it has some modulated part which is not the pure
carrier okay, earlier I was getting the carrier back

For DSB-S c I was getting the carrier back fortunately I was getting just 2wct if it has a drifted
2wct + theta but now I am getting some thetat which is actually determine by this m(t) and mh(t)
which is coming from the message and that will actually change, this is not pure co-sinusoidal
now this is something else so that I cannot take as a carrier signal because it is a corrupted carrier
signal which Is not pure co-sinusoidal so whatever I do even after squaring I do not get my
carrier back okay.

I can put up employ band pass filter around 2WCt but whatever I will be getting that is not
actually a carrier so with that I won't be able to employ any demodulation so that is why you can
see for VSSB also same thing will be happening because this will be instead of Mh it will be mv,
after passing through that filter right. So again there will be some theta term in the carrier if you
means try to square it, so you will never be able to through a narrow pass filter get extract that
carrier from there so that is why if for SSB and VSB if you wish to de- modulate it without
carrier that is not possible. first of all we will not carrier back so all you will have to do is ,you
will have to send some carrier along with it.

So that is where it becomes power inefficient because we have already seen that if we just try to
put some power for VSB if we wish to do envelop detection then we will have to actually put
huge amount of power to employ that envelop detection of course it can be proven thatfor VSB
the power will not be that high as compared to SSB but you will have to put power which is very
high compared to our amplitude modulation that we have employed.

344
So that is pretty much the discussion from engineering point of view from modulation,
demodulation point of viewthat what happens in different kind of modulation they have some
advantage some disadvantage we have explored that and now we are in a position to actually
explore more on carrier recovery because we have already seen anywhere we wish to employ
coherent demodulation we have to have some nice carrier recovery circuit. we have already for
DSB SC, we have already told that may be squaring might give me a carrier but we'll try to
means make that even more accurate so that's where the PLL- circuitary will come into the
picture which is called phase locked loop that has a big importance in communication system so
we will try to now explore now onwards explore the PLL circuitry from next class on thank you.

345
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-24: Comparison of Different Modulation Techniques

So far we have discussed different analog amplitude modulation schemes, so let us try to list all
that we have discussed so far.

(Refer Slide Time: 00:35)

346
So one is basically, DSB-SC that is the first one probably we have discussed and then we have
seen a simple variation of that so that is also a DSB-SC but we add means some carrier to it so we
can call that as amplitude modulation or just simple DSP with carrier okay, so this is just a
derived version of that and then we went towards more bandwidth efficient modulation schemes
which are single sideband maybe suppress carriers maybe with carriers so we will see that also so
every modulation schemes that we can think of will have suppressed carrier that means you do
not add any carrier so in the frequency domain there will be no delta term at wc which is the
carrier or it can be with carrier so there will be a delta term.

And then we have talked about another one which is called VSB vestigial sideband may be with
suppressed carrier again maybe with carrier, so it might be just SSB VSB with carrier okay, and
then finally we have talked about quadrature amplitude modulation. So these three actually
comes under or I should say comes under band width efficient modulation and the corresponding
version where we add carrier so those are with carrier and wherever we are suppressing carrier
that we write with this term SC okay.

So let us try to see the means just a comprehensive summary of all this the relative advantage
disadvantage what can be done what cannot be done let us just try to summarize all these things
because we have already covered all of them. So for DSB-SC in time domain.

(Refer Slide Time: 02:52)

347
DSB-SC ⟹ m(t)cos(ωct)
A M ⟹ A cos (ωct) + m(t)cos (ωct)
= [A + m(t)]cos(ωct)
SSB-SC ⟹ m(t)cos ωct ∓ mh(t)sin ωct

Suppose I have a modulating signal or you can call that as a message signal which is m(t) always
remember that whenever we are trying to modulate it, it has a corresponding Fourier transform in
frequency domain and that must be band limited, so frequency components are defined up to it is
mostly base band signal whenever we are talking about that is base band limited up to B
bandwidth the significant frequency components are there beyond that those are all insignificant
or nothing is present okay.

So that is our assumption so modulating signal m(t) and if I wish to represent DSB-SC so we
have already seen that in time domain the representation should be m(t) cos(wct) right, it is just a
multiplication with respect to a cosinusoidal term correspondingly the modulated frequency

348
domain response will go around + fc and -fc and the shape will remain the same if I just talk
about another version AM then I add a carrier to it so it is A cos (wct)+m(t) cos(Wct).

So we just try to now summarize everything and try to show what we have achieved so far, so it
is nothing but A +m (t) cos(wct) so therefore it is just like a DSB modulated signal but the
modulating signal you add with a DC level A or actually in circuitry you add a carrier to it with
strength A okay, or power A2/2 so what will happen the corresponding spectra will just look like
DSB-SC at fc at –fc but there will be a carrier added to it that is the delta function which is being
added.

So this is something we have seen probably in that course of our discussion we have not talked
about something which will try to reveal but before that let me just characterize the other things
also and then will reveal that okay, so for SSB-SC what we do is of course the m(t) will be there
and that needs to be multiplied by cos(wct) but then we actually that same message signal should
be passed through a Hilbert transform and or go through Hilbert transform.

And correspondingly we get mh(t) after getting a Hilbert transform the message signal where
Hilbert transform we have already characterized it is nothing but introducing or adding a phase p/
2 or each frequency component irrespective of the frequency component it is always adding a
constant phase, okay.

So this will be + or - depending on whether we are targeting USB or LSB means upper sideband
or low, lower sideband and this should be multiplied by sin(wct) so the corresponding frequency
domain response should be either this that means only one side band at area fc or -fc will be there
or the other side bands so this is the USB, upper sideband and this is the LSB okay, so either this
one or this one okay.

So whatever I was mentioning just a few minutes back that there is something we have not talked
about so the thing is that here whenever we represent this you can see that it has a predominant
representation that a particular signal multiplied by a co-sinusoidal term of the carrier and some
modified version of that signal multiplied by a sinusoidal term right, so this is what is happening
so this whenever we multiply with the cosinusoidal of the carrier whatever we put over here that
we call it as in phase term.

And whenever we multiply with sinusoidal we call that a quadrature term or Q term okay, so this
is in phase and this is quadrature or we simply call that I or Q okay, so for SSB-SC you can see
the in phase term is m(t) and the quadrature term is the Hilbert transform of m(t) whereas for
DSB-SC the in phase term is m (t) quadrature term is 0 nothing is there okay, for amplitude
modulation the in phase term is A+ m(t) so a DC added to the method signal and there is no
quadrature term that's all is happening so this is just characterizing the signal.

349
(Refer Slide Time: 08:23)

VSB-SC m(t)cos ωct + mv(t)

QAM m1(t)cos (ωct) + m2(t)sin (ωct)

For VSB-SC again we have seen so we just try to summarize again so it should be m(t) cos(wct)
okay, + some mv(t) this is actually that filter we were talking about right, so for VSB-SC if you
remember what we have said that we will be generating a DSB signal then pass it through a filter
band pass filter which is characterized as or the transfer function is Hi(f) okay, so we are talking
about this filter if we just pass it through that particular correspondence and corresponding to that
there is a filter and we can get a representation of mv(t) which was we have represented that filter
as F(f) right.

And if I pass my message signal through that filter whatever I will be getting that we term as
mv(t) okay, so we have already done these things so this can be next represented the modulated
signal can be represented as so again we can see there is a in phase term and there is a quadrature
term for this for both SSB and VSB-SC we can add carrier signal at any time if we wish so the

350
corresponding signal should look like so if this was SSB it should be something like this okay, so
that should be VSB-SC and then we have talked about QAM which actually takes the entire band
but it represents two signals.

So basically what we have said if we have a message signal m1(t) and m2(t) and we can represent
this at m1(t) we can put in in phase term and m2(t) in quadrature term, so I can just write m1 (t)
cos(Wct)+ m2(t) sin(wct) so that in a nutshell of course the bandwidth will look like same but it
will be just added part the frequency response will be added with respect to m1 frequency
response and m2's frequency response right.

So that should be the case but we have seen already we have demonstrated that this can be very
clearly separated out okay, so that can be done. So this is something we have already seen so far
and we have means these are the versions or variation of amplitude modulation schemes that we
have, now if you just compare them let us try to compare these things.

(Refer Slide Time: 11:07)

351
A ≥ | m(t) |
A ≫ | m(t) |
So if you just write Am DSB-SC, SSB-SC we can write SC general SSB also we can write VSB-
SC and let us say QAM right, now let us try to see where this things can be used of course we can
also have a just VSB okay, so AM generally used for broadcasting okay, because we have seen
that whenever we use AM and if we have this condition that my A is just greater than equal to for
all values of t as long as this is happening I know that I will not have any phase reversal so just
by envelope detection with a diode followed by a charging and discharging circuit.

I can actually demodulate it so the detector circuit becomes very simple and that is why probably
I will be targeting this for a broadcasting system because everybody will have the receiver they
are not transmitting anything it is just receiving things like television broadcasting or radio
broadcasting that used to happen earlier so it is in that case I will probably employ this because a
receiver which is at the user premises will be more cost effective so that is something we have
already seen.

But of course because we have to add that carrier so this will not be that power efficient, so we
have also seen that if we talk about so this is scenario where this can be used right, so next if I
talk about power efficiency this is probably not that power efficient of course the corresponding
DSB-SC will be much power efficient because I don't have to add that carrier and unnecessarily
means we have already demonstrated that right, with the modulation index and then we have also
talked about the power efficiency.

So this will be power efficient SSB-SC definitely will be power efficient because there also I'm
not adding carrier but here we have already shown that if we wish to add carrier and then we
want to demodulate probably a huge amount of carrier has to be added. So here if I wish to add
carrier and we want to demodulate through that envelope detection A must be much, much
greater than this m(t) so the amount of power that will be wasted for transmitting carrier which is
not useful means signal so there will be wasting huge amount of energy right.

Again VSB original VBS-SC will be power efficient or energy efficient I should say this will not
be okay, so here also that same condition as we have stated probably not proven for SSB you
have proven similar argument will come for the VSB also there also A must be much, much
bigger than this m(t) will be able to show that because it is almost similar structure, so this can be
again proven so of course this will not be energy efficient and even what will happen this SSB
and VSB if we wish to do for means use them for broadcasting then definitely we will have to
means operate at very low power efficient scheme, okay so that will be always happening so if
we wish to do that.

352
We have also proven another thing which is called the carrier recovery can we do carrier recovery
or cannot this one actually does not require carrier recovery it does not require because I can
whatever happens because I am doing amplitude modulation so carrier recovery is not required
carrier is already inbuilt I can just put a diode and we have demonstrated tha,t that acts as if the
carrier is already there and we are de-modulating.

So this does not require so I can say not required whereas for DSB-SC yes, carrier recovery is
required that is a big circuitry we have already started talking about the carrier recovery we have
promised that next we will be discussing phase lock loop which is integral part of carrier
recovery. So the DSB-SC definitely require carrier recovery SSB-SC this is where there is
something that we have proven in the last class that this for SSB-SC the carrier recovery is very
difficult.

But the DSB-SC we have seen that we square it we always get back the carrier but for, but for
SSB-SC and VSB-SC even if we square you won't get the carrier term, pure carrier term it will
not be at 2 Wct so there will be a phase variation with respect to time so you would not be getting
the pure carrier, so carrier recovery is not possible here also it is not possible that is a big blow in
modulation technique.

Because these two are really efficient in terms of bandwidth SSB particularly it gives us twice the
benefit right, so two signals I can put in place of one signal compared to a.m. and DSB-SC
whereas VSB-SC not that much efficient as SSB but okay it gives some benefit but the problem
is I cannot really and they are also energy efficient because I am not putting carrier to it but I
cannot really do carrier recovery from them the original signal that has been transmitted from
there I cannot extract the carrier so all I will have to do probably I will have to separately transmit
pilot carrier along with them if I wish to do demodulation properly.

So that is the additional headache that will have to take if we wish to do that these two okay,
carrier recovery means I am putting carrier so it is possible that I don't have to do carrier recovery
I can just do same envelope detection over here but energy efficiency is very low that is what
happens okay. So these are the few things that I wanted to discuss means this is something we
have already covered we have proven all of them but I wanted to summarize them.

Next another one thing probably we have discussed that is actually related to the carrier mismatch
okay, so there are two mismatch we have seen these are the things also we have demonstrated
that there might be a phase mismatch between the incoming carrier and the local carrier that has
been generated once will be discussing this PLL you'll see that that, that is possible that mightbe
happening.

353
So if there is a constant phase mismatch between the incoming carrier and my locally generated
carrier whichever way I can I can square that pass it through a PLL whatever carrier I get if that
has a constant phase mismatch then what will happen, so if I have phase mismatch of delta and
then we have also discussed a frequency mismatch of some delta so this probably we have given
as some delta and something like this right.

So if I have a phase mismatch what we have seen for this one DSB-SC we have demonstrated
already what we have seen that probably will get the signal intact but only thing is that there will
be a attenuation of this term cos delta so modulated signal after we demodulate I will get back my
m(t) but m(t) will be multiplied by this cos delta. Now the problem is if the phase mismatch goes
around pi/2 then this cos delta will be 0 so attenuation will be very high I would not get my signal
back so that is the detrimental effect that we will have.

But if there is a frequency mismatch what we have seen that this will be multiplied my m(t) will
be multiplied by cos delta(t) so there will be a modulation term which will be and which will vary
the essential nature of the message signal itself so that little more detrimental so if there is a
frequency mismatch I have to be very careful about that so I cannot really allow frequency
mismatch in the incoming carrier and the local carrier I am generating.

All these things will be very clear in PLL we will see the strategy that we have to take to ensure
these things knowing that this will be happening then for QAM we have discussed some more
things that if there is a phase mismatch so we have already seen that okay, probably it will not
means the message signal will have something like that m1(t) suppose I want to demodulate
m1(t) so it will have this cos delta term but there will be also additional term with respect to
m2(t) so that will actually give me interference to my signal, so that is even more detrimental for
QAM.

So QAM in terms of phase mismatch and frequency mismatch it is more vulnerable because I am
actually modulating two signals and putting in the same band so I will always expect if there is a
phase mismatch immediately there will be our interference term coming from the other signals
which will power up my signals so that will happen whereas for phase mismatch in DSB nothing
will happen it will just have attenuation term same thing will happen if just you carry on the same
technique you will see that it will be just repeated over here also it will have similar effect as
DSB-SC okay.

So VSB or SSB if you carry out the same thing you will see that similar effect will be happening
for them but QAM that is more detrimental so if you just modulate with respect to QAM or if you
are transmitting with by employing QAM you have to be very careful about the frequency and
phase mismatch in your local carrier okay, frequency mismatch of course there will be

354
modulation sorry, interference term plus there will be additional modulation term which will be
coming out similar to this okay.

So these are probably what we have discussed so far okay, so just if we should summarize that
where, what can be used and what are the things that I need to take care of so whether its energy
efficient whether the carrier recovery is required not required accordingly the receiver circuit how
it will be can we use it for broadcasting, can you use it for one-to-one communication or point-to-
point communication and if there are in the carrier recovery if it is required at all that means I
cannot employ a means envelope detection.

Then what kind of effect I will get if there is phase mismatch or frequency mismatch so this is
something we have so far discussed. There are, there were some more things which has to be
means which we have also discussed in the process that was related to two of the in a sorry,
bandwidth efficient schemes.

(Refer Slide Time: 23:34)

355
So one is SSB let us say SSB- SC, and the other one is VSB-SC so we have seen that this is good
for if we employ a Waiver circuitry this is good if we have a signal which has nothing around 0 in
the frequency band so that is voice signals so this is good for voice modulation but this is not
good for video modulation because that has something so video spectrum looks like this whereas
voice spectrum looks like this, so nothing around 0 okay.

So for this SSB-SC is good we know that SBS-SC can be generated by filtering approach okay,
so for that filtering approach only we are talking about this for the other approach where you
have to take Hilbert transform that is very difficult because every frequency component has to be
given pi/2 phase shift which is very critical and that is very difficult to realise in realistic circuit
we have given alternate circuit which is called Waiver circuit, which is with respect to filtering so
you do first filtering then you modulate then you do another filtering and you get your signals
right, modulated signal.

So for that we need we can only modulate voice signal we cannot really employ this technique
for video signal whereas VSB was actually designed keeping in mind that if a signal has
something around 0 what should we do so the this particular thing can still be realized using a
realistic filtering technique, because it allows the filters to have some role of some realistic roll
on, so that is something which we have discussed so far okay.

So now we can see that the only part which is still not very clear probably is the carrier recovery
so we have to now discuss in big details about the carrier recovery and what is the consequence
when we get some phase delay when we get frequency delay sorry, when we get some phase
delay among the carrier and the locally generated carriers or some frequency deviation among the
locally generated carrier and incoming carrier.

So in the next class what we will try to do will try to means get into inside the PLL circuit and try
to see what happens over there and do some analysis okay, thank you.

356
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-25: PLL

Okay so has promised in the previous class let us start discussing about PLL which is called
phase lock loop.

(Refer Slide Time: 00:35)

357
A cos [ωct + θi(t)]
A cos [ωct + θ]

Okay this is a typically circuit which will be a mostly in many places will be used in
communication so it has utilization in a digital communication has well because there are also
you need to do clock recovery for clock recovery PLL’s one of the integral part or the topic we
are discuss in this PLL that is where it is very important that carrier recovery we have already
seen that if I wished to do coherent deduction or coherent demodulation where I need to generate
a local carrier we have already seen the effect of that local carrier drift.

So we need to have something where the carrier is properly synchronized and locked to the
incoming carrier okay in with respect to frequency and phase okay and we will later on we will
also see that this particular circuitry also being employed heavily for FM demodulation we will
once we understand the mathematics of or the basic principle of this circuit will then we will
discuss about frequency modulation and we will see that how this can very easily used for FM
demodulation.

So this is one of the circuit for communication probably one of the most important circuit let us
try to discuss about it is essentially a means we are saying phase lock that means we are trying to
lock the phase okay so whether it is frequency or phase suppose I have cos sinusoidal let us say A
cos Wc t + some input phase it is having right it might be time varying might not be even Wc
might be time varying might not be time varying whatever it is okay.

So we are saying that this is the phase and this is the frequency or over all we can say this is
actually the time varying over all phase of that co-sinusoidal, so this is the phase and we want to
lock this overall phase that is our target okay within this there are two constituent part one is the
frequency and another one is the phase okay so pure phase so and generally what happens if I
have the representation like this where wc and theta are no time varying.

That means they are constant with respective time then that separation of frequency and phase
comes into picture, okay otherwise it is the overall phase so once both of them are varying with
time I do not really I cannot really discriminate what is frequency and what is phase where as if
this is the way it is happening then I know that there is one part which is multiplied by t so that is
my frequency + there is some constant part that is the phase okay so that is how we give a
definition of frequency and phase is separate them extract them out.

Otherwise in a completely time varying scenario where wc and theta are also variables of time I
cannot really discriminate between them it is just overall phase of that co-sinusoidal okay which

358
is time varying, so and that is why this particular circuitry is called phase lock loop that means
whatever this entire phase is I want to lock it to that okay it might vary with time but my things
also should have equivalent variation whatever I will be generating at my local oscillator.

That must have equivalent variation so this is what I want to do, so definitely if I wish to lock
something immediately one thing comes into your anybody’s mind that is feedback loop, so that
is why it is called as a loop circuit okay so we will have a feedback and with that feedback we
will try to actually minimize the phase error between the incoming and the outgoing.

So that is the overall principle the principle is very simple I will have a phase which is incoming
so let us say I will have co-sinusoidal which is coming with a incoming phase of theta it and
locally I will generating some phase theta o(t) and my target should be that this difference theta 0
to theta i or theta i to theta0 must go to 0 so I need to create a feedback loop, so that error which
is the phase error must go to 0 that can only happen with the feedback. If I try to do that so my
circuitry also will look like that it is a feedback circuit, so the circuit is something like this.

(Refer Slide Time: 05:21)

359
 ω(t) = ωc + ce0(t)
VCO output = B cos [ωct + θ0(t)]
ωc + θ′0(t)
θ′0(t) = ce0(t)

It means before drawing the circuit we should talk about what are things it consists of 1) it
consists of a VCO this is probably the first time you are hearing that is called voltage controlled
oscillator this is the typical circuitry where you actually give a out input voltage and proportional
to that voltage it will this is the oscillator that is why it is called voltage control oscillator.

So it will create a conciliation, but the frequency of oscillation so it will actually create a
sinusoidal wave but the frequency of oscillation which is that w c must be proportional to the
input voltage that we are giving okay so that is called VCO voltage control oscillator that means
it is a typical circuitry where you give a input voltage and you can expect that outside or output
will be getting a sinusoidal which has a frequency which is proportional to the input voltage okay.

So typically we write it like this so the Wct okay which is the frequency of oscillation of this
voltage control oscillator, so this is the output okay output is actually I should say it is a cos wct
or wt okay so it is a co-sinusoidal which as a instantaneous frequency of this one, and this must
be some wc constant I will talk about that + some constant into e0t where e0t is the input of the
VCO it gives me some I can take A cos wt right.

Where wt is this okay so you can immediately see that wt that is the frequency of that cos
sinusoidal which is coming out of VCO that is proportional to e0t but there are proportional
means it must be linear so this is a linear curve, with two constant which are specific parameter to
the VCO one is the wc what is Wc that means if I actually do not give any voltage as input and if
I keep the VCO to run freely.

That means I give power VCO is operating so I keep him to run freely then this is the frequency
at which he will be running okay so this is the output frequency we will generating that is why it
is called free running frequency okay so that is the typical parameter to VCO different can have a
different standard and different free running frequency so accordingly we will be buying a VCO
with your targeted carrier range right it should be around that your free running frequency should
be around that.

So that is the Wc and c is just a constant which is a by which the et is multiplied so that is
probably that slope of that linear variation so that is also typical to a particular VCO okay so
these two parameter will be given for a chosen VCO so what will happen if I have some C sorry

360
some e0T input to it will keep on running at different different frequency now if I put the VCO in
a loop circuit remember my target is this e0t.

If that is the error voltage let's say or that is equivalent to the error phase difference then I will be
targeting that error must be zero okay in the loop that is how I will be designing the loop we will
see that later on but remember if I put itin a feedback loop then this error will be 0 because the
error will be generated by a phase difference between the input phase and output phase so that
can only be 0 if these two frequency are locked that means the it will not be we the error is 0
VCO will not be a in a loop configuration VCO will not be operating at wc because the input
error is 0 in a loop it will actually run at the frequency Which is equivalent to the incoming
frequency un-phase so thing you have to always consider once I write this equation you will be
always thinking that as if I put e0t = 0 it will be always running at wc but that is not the case it is
the case if I am not putting it inside the loop it will be it will not be the case if I put inside the
loop because inside the loop it will actually run at the frequency which is completely in
synchronized with the incoming phase and frequency because then only the difference will be 0
and the VCO will not change further.

So it will be located to that so we will see that so I can, I can see this now let's say the VCO out
put so the VCO if it is running at this frequency so what should we get let us say VCO output is
according to my target is some B cos Wct + theta output okay so this is my target I want a VCO
which must run with this frequency and phase okay.

But what is happening if the VCO input is e0t then that must be the wt okay fine so from here this
is a phase so what should be the instantaneous frequency I should differentiate it because when
this phase term was constant it was Wct + some theta if I differentiate I get the frequency so for
instantaneous frequency from the instantaneous phase it is always the differentiation of it so that
must be Wc + first derivative of theta ot so this must be the VCO instantaneous frequency where
as I have already said at this must be the VCO instantaneous frequency that wt okay.

If I give e0t as input signal so this two must be same if I put that VCO inside so therefore I can
always write that theta output t differentiation must be equal to this inside the loop okay when I
put the VCO inside VCO output will be this that is my output phase and I assume that it is
running at wc okay so if that is case this output correspondingly I will be having a instantaneous
frequency so I calculate that and I have also told that VCO instantaneous frequency must be this
so therefore there must be a relationship between the output phase as well as the input voltage. So
that is the relationship we were targeting and we have got this. Okay now let us see how do we
use the VCO for PLL.

(Refer Slide Time: 13:05)

361
AB sin [ωct + θi(t)] cos [ωct + θ0(t)]
1
= AB [sin (θi(t) − θ0(t)) + sin [2ωct + θi(t) + θ0(x)]]
2
So it is like this I will have a signal which is A sin wct + theta i t okay remember this, this is the
carrier I am getting you might be asking okay what where do I get this carrier so that is the first
question because the PLL if this is the PLL we are designing so PLL will be at the receiver circuit
do not have the carrier, so we have already talked about that squaring and narrow band pass filter
passing it through that we get the carrier with some phase Whatever carrier phase as carrier
frequency or carrier phase this is that so basically before PLL always you will employing that
squaring circuit and for SBB and VSB we have told that PLL cannot even work because that
squaring does not give me carrier so if I do not get a carrier over here pure carrier over here Vco
cannot help okay.

362
So it is very important that after squaring and putting a narrow band pass filter I must get a carrier
if I am getting that then I can employee receiver technique so VCO will take this carrier whatever
it is with a phase and frequency.

And he will have a multiplier circuit okay so in that multiplier what he is doing is giving the
output from VCO okay so VCO will generate as we have already discussed that it should be
generating these B cos (Wct + theta) output t right so that is the signal VCO is generating we
have already told that okay.

If the output phase is theta we will have multiplier circuit followed by a low pass nature of loop
filter with the transfer function of H(s) whatever it is, it is a low pass filter we will see why low
pass filter is requiring and that should produce the error signal which should be fade back
because it is a feedback circuit to with VCO okay.

So this is all we are doing, let us try to see if this is correct okay this is what we are targeting so
what is happening whenever we multiply what do we get we will be getting a multiplication of
these two so that should be so I have this and I have this so if I multiply I will be getting after
multiplication, so it should be A x B sin Wct + theta i t right that is the incoming signal with
incoming phase and then the VCO generated wall.

very carefully check I have already assumed that probably frequencies are same okay that is not a
requirement will just after sometimes we will prove that this is not a requirement okay we can
always have any frequency that VCO is running at or any input frequency which is different from
the VCO free running frequency we can still track it we will prove that but right now we are
probably assuming that probably the frequency is already synchronized it is the phase I want to
track okay.

So if that is a case so I will multiply these two okay so after multiplication it is just half AB so we
will have 2sin A cos B so I can just means put sin A + B and A – B that formula okay so we will
be getting sin this minus this so that should be wct will be cancelled so I will get theta i t – theta
ot right and then + sin I will have this plus this so that should be 2 Wct + theta i t + theta ot right I
will have these two term now you can see after the multiplier this where I will be getting these
two.

I want at this point just the phase difference term, so I do not want this what is the procedure to
reject that just a low pass filter so that is why this must be a low pass filter this will reject this
part if it is designed probably so that 2 Wct term it cancels out okay so immediately at this point I
will be getting just the sin this difference, base difference pass through the filter transfer function,
okay.

363
(Refer Slide Time: 18:16)

H(s) ⇔ h(t)
1
e0(t) = h(t) * AB sin [θi(t) − θ0(t)]
2
CAB t
2 ∫−2
θ′0(t) = h(t − x)sin [θi(x) − θ0(x)] d x

1
CB = K
2

So let us say the filter transfer function is Hs and corresponding if I do inverse transform the
corresponding impulse response is h (t) so what should I expect at the output of the filter that
must be convolution with this one of the sin signal so it must be my e0(t) which is generated as
error that must be equals to that h(t) that is the filter trans impulse response convolution with

364
whatever signal I have got ½ AB sin theta it – theta ot right because I know the filter is already a
high pass sorry low pass filter. I have neglected the other term because that is not required so this
must be my overall output e0t, e0t I can write that as means I can multiply this by c so I multiply
by c what was c e0t that was actually theta ot differentiation so I get this equation c of course AB
½ comes out integration -infinity to + infinity so I am just writing that h(t) convolution this so
that should be h(t – x) sin theta ix- theta ox so I have just put the convolution.

Okay where I will be now putting this ½ CB that is a term completely or defined by the VCO
because C is also a term of VCO and it generates amplitude of B so these are terminal means
these are all parameter of VCO so therefore ½ CB I can just represent as a parameter of VCO so I
can write that as k so I can have then Ak and integration this term right where this I can write
now as theta et or ex right.

So far I have done this now let us try to see if I can get a equivalence circuit of this one okay so
the equivalence circuit now I wish to define is see earlier what I was having I was having a cos
sinusoidal if I just put that circuit this is was the actual circuit of VCO I was putting two cos
sinusoidal I was putting a multiplier the circuit immediately becomes a nonlinear circuit right I
have no way to actually analyse it through linear circuitry okay or linear circuit theory measure.

So from there I wish to now transform this to a linear circuit right so what I do instead of taking
this sinusoidal and co sinusoidal I take it from the phase perceptive as if my input is phase output
is phase only okay so what I do is I transform this circuit.

(Refer Slide Time: 21:38)

365
 θe(t) → 0
sin θe(t) ≃ θe(t)

As If I am giving a input of theta it okay instead of giving that sinusoidal I can just think about
that I am the sinusoidal all other parameters are fixed on so that as nothing to do with the circuit
A and wc those are not going to change it is just this phase which will be actually tracked and
controlled okay, so this phase as to be properly designed so what I can say is this is the phase
which is coming in and then after that I can put a adder circuit.

So this is where the multiplier becomes adder because if I see it from phase terminology it is just
phase addition or subtraction okay so what has happen inside if you see I have a sinusoidal and I
have a phase subtraction so this subtraction part only I will pick. So what I will say that I have a
output phase which is being generated from the VCO okay so VCO I will put equivalent circuitry
okay Later on but this is the phase that is being generated I will put + here – here, so that should
be theta et but what I have got inside before the loop filter is sin of that right, so I pass through a

366
sin converter so this is just take that phase and convert it into sin then I get a exactly what should
be the input of that loop filter then I put this H(S) but in that H(s) now I should also have
something which is if I have just done sine this then sorry I have just convoluted sine with this
one this particular term should be part of it.

So what I can do now in H(s) I can include that that is constant term so I can just put instead of
H(s) I can put A k H(s) by then this sin will just convert it to sin theta i-theta o then Ak will be
multiplied because either it is the signal or filter it does not matter because this is a constant term
H (S) is automatically there whatever I get over here that I can term as my c e0t right so that I can
term as c0t so if this is ce0t because that is what we have told that is the error signal so this after
convolution I get the error signal ce0t so if this is ce0t then what should be the output that is just a
differentiation of the so this must be theta ’

So what should be this circuit then this is a integrator that is the beauty of it so it can immediately
see very nicely that non linear circuit if I just visualize it from the phase perspective I can
actually transform it into a linear circuit I have adder I have integrator in S domain this will be
also a treated as linear part H(s) is a realistic filter so no problem in all these things only
problematic thing is this sinusoidal so we will now we will see in the next class probably we will
try to see how to get rid of this what we will say a very simple assumption will get rid of this.

We will say we are doing a small error analysis that means the VCO as already all most tracked it
or generally if I have a carrier signal I will actually give a input which is very close to the carrier
because it cannot be that I have a carrier coming at five megahertz and I am just starting form 1
megahertz I will never do that because I know roughly it is 5 megahertz so VCO that I will be
putting free running frequency will be already around 5 megahertz just that small error that needs
to tracked.

So if that is the case the phase error should be very small so I can say this theta et is very close to
0 if that is the case what can I write about sin theta et that must be almost equal to theta et that is
it I do not need this and immediately I can say at entire circuit becomes just a linear circuit and I
can just do the equivalent transfer function calculation and the whole analysis whether I will be
getting a means whether this particular thing should give me tracking and what kind of tracking it
will be any track phase any track frequency and all those things so in the next class.

What we will try to do we will do this small signal analysis or small error analysis at try to see
what kind of tracking is required only thing that is in my hand VCO it is integrator I cannot touch
that all those constant I can change but that will not change the overall transfer function of the
overall circuitry now only thing that is in my hand is the filter transfer function H (s) so we will

367
see what kind of filter to the employed and correspondingly what kind of tracking I get in terms
of frequency and phase so in the next class we discuss that thank you.

368
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-26: PLL (Contd.)

Okay so in the last class we have started discussing about PLL and some analysis of PLL right so
how good it is tracking we have already discussed that its very important that we track the
incoming frequency and phase well and we have seen that detrimental effect if we cannot track
either the phase of frequency specially the frequency is very important because that gives
modulation term if you just think about DSBSC or any other thing and of course phase is also
important because it gives attenuation if we do not track it properly so we have already started
doing the analysis of PLL we have already given the basics circuit diagram of PLL means it is we
have already told that its probably of operates in a feedback loop with a special component that is
called VCO right voltage control oscillator.

(Refer Slide Time: 01:36)

369
A sin[ωct + ψ (t)]

= A sin ωct + (ω0 − ωc)t + ψ (t)

θi (t)

So whatever input voltage you give accordingly it will oscillate okay and the frequency of the
oscillation will be a linear means it will have a linear response with respect to the input voltage
right so that is what we have done and what we have done the basic PLL circuit that we have
given was having a multiplier followed by loop filter And this the error signal that was getting
generated and then there is a feedback in the feedback loop you have that VCO okay so whatever
VCO produces that is Bcos WCt+theta output T oaky and the input signal which is A cos or sin
wCt + theta it so this is how we represent it so sinusoidal is coming we generate co- sinusoidal
signal over here from VCO and all we wish to do is this two phase we want to get proper
synchronism in this two phase input phase as well as output phase okay.

370
So one thing we have not probably covered that here if you see we are already assuming that if
the incoming frequency is Wc then the phase of the sorry frequency of the VCO also should be w
c let say we have already talked about free running frequency of VCO let say that is Wc okay or
Fc, so that is already given and my incoming frequency is not matching to that it might be closed
but it is not matching then what happen okay.

Then will this VCO technique has still operate at that point let us try to see that okay so let say I
have an incoming signal which is A sin instead of Wct that has a different frequency. Let's say
w0t+ some input phase let say si t is coming over okay so this is just a modified one of this one.
We have just assuming over here that there is no even no matching in frequency earlier whenever
we started analyzing it you are saying that in analysis you will assume that the frequencies are
matched.

And then there is a time varying phase and we want to match that phase okay so that was our
target now we are saying that can be really extend that same analysis for a frequency mismatch
case okay so if this is case I can actually rewrite this I can write it this way wct + (w0-wc) t+si t
no problem in doing that right I can always do that and what I can do I can actually take this part
and define that as the input phase theta it.

So I can always do that because any way this was not a constant thing this was the variable of
time so I can add another time varying part to it and I can consider that the whole thing composite
thing has theta i t and then what happens immediately whatever frequency comes I can actually
represent that in terms of the free running frequency of VCO and then I can accordingly
manipulate the input phase okay.

And the tracking will be on that phase only right so I can always frequency mismatch I can
actually put that in phase mismatched out and in then do the analysis similarly the way we have
been doing it okay so alright after this we as we have now a good understanding that phase and
frequency both we can treat similarly what we have done.

We have actually converted this particular circuit that we have shown with the multiplier because
that was actually a non linear circuit equivalent linear circuit where you are treating phase as the
input and phase as the VCO generated things okay so immediately we can see that it will be a
linear circuit.

Because whenever we multiply there will be a phase+ term and phase – term and then there is a
loop filter which is predominantly assumed to be low pass filter so the + term will be cancelled it
just – term which will be there so if phase wise we think about that circuitry it just a input phase-
the output phase so that is the linear circuitry right.

371
(Refer Slide Time: 06:02)

θe(t) → 0
θe(t) ≪ π /2
sin θe ≃ θe

So that is what we have already employed so what we have said we will see it from the phase
perspective not an actual signal perspective okay so immediately what happens so this is theta i t
so there is theta O t which is coming in and that creates my theta e t after subtraction okay after
the difference of course because the original signal was some cos and sin so you multiply there
will be cos sin a+b and sin a-b and a + b terms get cancelled so the rest will be sin a-b so if this
the phase difference, There should be of circuitry which converts it into sin right that should be
pass to the low pass filters according to our low pass into equation that we have derived in the
previous class right so that should pass through the low pass filter and equivalently we have
written that should AK and Hs right.

372
That Hs is a transfer function of the low pass filter and AK are system defined parameter either it
comes from VCO or it comes from the means it generally comes from the either signal or VCO
right so that is the system generated parameter okay so this should generate the error signal which
we are calling as Ce0t okay and then that gets feeded to the VCO we have also seen if we see
from the phase perspective that becomes integrator.

These are the things we have already proven right so it either – infinity to zero or if the signal
starts from 0 you can take it from 0 to t so as – infinity to t so this is just integrated circuit where
as the input is actually the output phase derivative of that to which is equal to Ce0t right so that
gets feed into this because it is integrated immediately convert it to theta o t and then gets feeded
over this circuit right.

So this was the equivalent circuit that we have already means transform so this circuit
transformation we have already done in the previous class right and that was easier now we
wanted to do the analysis of this circuit whether it is stabilizable whether we get this error to be
means in a time if we try to see what kind of filter designing will give me this error to 0 okay.

Because that was we want if you want the phase matching then theta i should be equivalent to
theta o that has to happen then theta e must be 0 right so if we wish to get that this must be 0 okay
so phase must be matched and then this error signal must be 0 this is what we want actually so
that is something we wanted to achieve right.

So if that happens then immediately there will be a locking in the input and the output phase okay
so for that as you can see this K parameter that was typical to this VCO okay it was having the
parameter B which is the actually if you just see over here that B is the strength of the VCO
output signal so that is the typical to VCO.

And then it was also having a parameter called C which is also part of that linear variation with
respect to VCO input voltage and frequency that it generates so this is also parameter typical to
the VCO and K is actually half b*c we have proven that okay so that is the typical VCO related
parameter okay.

So I cannot really touch that once the VCO is given I cannot touch that A is also something which
is coming along with the input signal right I cannot really touch that so what I can touch is this
Hs so that loop filter designing has to be done properly so that I can prove that this particular
thing actually track my input frequency and phase. That should be an effect now on wards okay

so let us try to see how do you do that this entire circuitry if you see if I just convert it into S
domain this is fine that should be integration so it should be 1/S right we know that integration
equivalently in Laplace domain and it should be 1/S that is all fine.

373
This is alright, typical function of S and no problem in that addition no problem in that only
problem or the non linear part that is where in the sin of this signal co- sinusoidal of something if
you wish to calculate the transform function that will be non linear thing right so the transfer
function calculation of transfer function is going to be very difficult.

So that is the only disturbing element now what will do is will try to think that whatever happens
the incoming frequency and the corresponding phase is already very close to the one that is
generated by the VCO okay so this small assumption I will take and immediately we can say
theta e t is actually very close to 0 okay or it is very small I can just write that.

You can write that theta et is much, much less than basically I can say pi/2 okay if I write that so
basically the error in phase is much smaller in that case the sinusoidal can be approximated as
angle itself so I can write sin theta e as almost equivalent to theta or if it is theta e t must be
almost equivalent to theta et this is something I can write okay, once I write that immediately this
sinusoidal goes away for small error whenever the error is within a very small amount I know
that I can have some analysis will characterize how much small it should be so all those things
will be characterized.

(Refer Slide Time: 12:29)

374
 AKH(s)
θ0(s) = [θi(s) − θ0(s)]
s

[ ]
AKH(s) AKH(s)
θ0(s) 1 + = θi(s)
s s
θ0(s) AKH(s)
=
θi(s) s + AKH(s)

But right now you can see that if it is stage error is much smaller than pi/2 we can actually
approximate this sin as this signal itself immediately my transfer function of this particular thing
becomes very simplified so I have a 1/S over here due to the integration and I have AK Hs over
here right fine.

So this is actually my theta it or in S domain I can write it has theta is this is actually theta ot or I
can write it has theta os right so at this point I get theta i s - theta os right this is + this is – so
basically what happens theta i s - theta os right this if I multiply so this is that part if I multiply by

375
this and this then I get my theta os so I can write theta os is nothing but this into AK Hs/S this is
overall equation I can write immediately I get a relationship between my output phase and input
phase right.

(Refer Slide Time: 14:32)

θe(s) = θ0(s) − θi(s)

A sin [ωct + ψ0]

= A sin ωct + (ω0 − ωc)t + ψ0

θi (t)
ω0 − ωc ψ0
θi(s) = +
s2 s

s + AKH(s) [ s 2 s ]
s ω0 − ωc ψ0
θe(s) = +

376
So I can write theta os should be theta os if I put it this side AK Hs/S is equal to theta is and AK
Hs/S so I can write theta os/ theta is = AK Hs/S +AK Hs right. That something I can write what is
theta es that is equal to theta is-theta os okay and immediately from there I can get a also has I got
a relationship between theta 0 and theta i I can get a relationship between theta e and theta i so
immediately I can write this theta es is equal to S/ S+AK Hs that just get manipulation of theta is
if just put that over here I will be able to get this relationship also right so far it is all good what I
have to now see is I give a input phase and i want to see that my theta es in equivalent time
domain if I put t tends to infinity it should go to 0 okay

So that should be my target or it should go to something wherever it goes that my actually phase
error finally I will be getting that the steady state phase error because there might be oscillation
and all other things but as t tends to infinity whatever I get that should be the phase error I will be
getting right so I.. because theta es is actually phase error between the input and the output okay.

So I will try to characterize that one what I do let us say my incoming signal has a frequency
error okay let say it is A sin wot+phi o okay so basically what I am trying to do is I have a input
signal now which is a sinusoidal okay this I might have created by squaring it and then filtering it
properly okay.

So which we have already discussed that double side band surpassed carrier if we get, we actually
square it and do a filtering whatever frequency we get we actually make it half then from Wc we
get wc or whatever it is that might be little bit deviated so whatever we do get we are that is
actually at means we do not say Wc so w 0 let say so we get this right.

So whatever it is we will be getting that now at the VOC, VOC generates that w c okay so
immediately I have to calculate the effective theta it with respect to that wc so I can write this as
usual or as previous case I can write this is wct + w o - Wct + Phi 0 right so that becomes my
theta it right so Wct , Wc is already that free running frequency of the VCO.

So therefore this must to be my input phase okay so if this is the input phase what is the
corresponding S domain representation okay so Wo-Wct, t in S domain is 1/S2 right so it should
be Wo-Wc/S2 so just go through the list of Laplace transform will be seeing that means the linear
in time that gets 1/S2 and a constant will be 1/S right so I can right that has I do we get this
rightso theta is, is evaluated.

Now let see theta es from this relationship so immediately I can write theta es as S/S+AK Hs into
theta is right so that was what we have already written so theta is I can write over here so that
should be wo-wc/S2Phi0/S right so we get this now here all other things are known only thing
that is not known is Hs so let try to put some Hs over here.

377
(Refer Slide Time: 19:15)

H(s) = 1

s + AK [ s 2 s ]
s ω0 − ωc ψ1
θe(s) = +
ω − ωc ψ0
= 0 +
s(s + Ak) (s + Ak)
(ω0 − ωC) /AK (ω0 − ωC) /AK ψ0
= − +
s s + AK (s + Ak)

t→0 { AK }
ω0 − ωc ω0 − ωc −AKt
lim θe(t) = lim − e + φ 0e −AKt
t→0 AK

So let say first thing I do is a low pass filter okay so Hs must be one within the band of interest
then it should be 0 so I can put Hs within the band of interest should be one okay so if i just

378
replies that so my theta es must be S/S+Hs becomes 1 so it should be AK and I do have Wo-Wc/
S2 +Phi 0/S right this I can write so this should be Wo-Wc/S*S+AK right + shi o S gets cancelled
so it is S +AK All I wish to do is theta es I have got I want to go to time domain representation.

So inverse transform has to be done so for this I know how to evaluate inverse Laplace transform
okay because it is in form of 1/S + some constant this is still not I can do partial fraction
evaluation and I can again represent it this way and it turns out to be w0-Wc/AK/S-W0-Wc/AK/S
+AK right+Phi 0/S +AK this is fine if you just add these two you can see that things will be
cancelled and finally get w0-Wc okay.

So this is the valid representation and immediately if I just wished to do from the list of Laplace
inverse transform I can evaluate this as whenever it is juts 1/S that becomes that thing only
whatever I have so it should be w0-Wc/AK whenever it is S+AK it should have a exponential
term okay e to the power minus that into T so it should be w0-Wc/AK that should be there and
there should be term called e to the power -AKT same thing will be happening over here also so
that should be + Phi o e to the power -AKT.

So I have got my Phi e t representation or theta et representation now all I have to do is I have to
see the final value okay whenever I put time tends to be infinity that should be the steady state
response of this particular means circuit right so for steady state response all I have to do is I have
to now put limit t tends to infinity immediately I can see this term will survive.

(Refer Slide Time: 22:26)

379
 ω0 − ωc
θe(α) = AK

Because that is the constant term this as t goes to infinity this goes to 0 so this so these two terms
will be cancelled out so what I get is the steady state phase error lets called that Theta e at infinity
that happens to be w0-Wc/AK so what has happened I have put a first order low pass filters
because Hs was 1 within the band of interest after that it must be a low pass filter so whenever I
put that first order characteristics or within the band all pass characteristics.

What I can see that this should be the final phase error so there was a input frequency error right
we have W0 and the free running frequency was wc so it can actually track that frequency there
was no problem in that it has tracked that frequency where as only there will be a phase error
which si constant phase error which is related to these parameters okay.

The input frequency the free running frequency A which is the input amplitude and K which is
typical to our phase lock loop okay or this VCO so that should be the overall phase error
fortunately this is constant right constant in time so will be getting the constant phase error if I
just do this what does this says for our DSBSC modulation.

380
It says very simple thing that there will be phase error and that phase error is constant over time
constant phase error so that means I will be getting attenuation in my signal we have seen that so
if there is a phase error in the input VCO generated out or VCO whatever VCO generates if there
is a phase error then there should be a amplitude degradation or I should say attenuation.

So attenuation should be cos of this term so whatever that phase error cos of that term only there
should be a amplitude attenuation so as long as we can set Wo Wc means we now know how to
design all this okay and we know how much phase shift will be there if I have that kind of loop
filter okay so next what will try to do will try to see that my changing filter characteristics.

We had first order filter which was Hs equal to 1 if I just change the characteristics of the filter
can we really do little bit later can we really take the phase error also out of the picture if this is
something we can achieve then we know that even that attenuation will not come into pictures so
in the next class will try to evaluate that part. If we just give some different kind of filter
characteristics can we get something better than this okay so that should be our point of
discussion thank you.

381
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-27: PLL (Contd.)-2

Okay so in the last class we have discussed about the means characteristics of a PLL and how it
can track these and frequency errors, so we have seen that just by giving a first order filter which
is h = 1 within the band of interest and it is a low pass filter of course because it has to reject that
higher frequency term so if that we give we could see that it tracks frequency very nicely, so
basically exactly tracks the frequency in phase it gives a constant phase error okay so this is what
we have already observed right. So now what we wish to do is can be better this so our next
target should be will take H (S).

(Refer Slide Time: 01:06)

382
 s+a
H(s) =
s

[ s2 s ]
s ω0 − ωc φ0
θe(s) = +
s + AK ( s )
s+a

s + Ak(s + a) [ s 2 s ]
s2 ω0 − ωc φ0
= 2 +

θe(α) = lim θe(t) = lim sθe(s) = 0

t→α s→0

Which is of second order so lets say the transfer function is something like this S + A / S so this is
just a realistic filter we can design it with any RC circuitry so that is possible so let us say we will
take this so immediately what will happen by theta es which is nothing but this s / s + Ak hs so in
place of hs we can write s + A / S that is my hs into theta is which is W0 – Wc / s2 + Psi 0
divided by s this is something we have already proved, so we have said that it is the similar kind
of input which is coming at a different frequency of w0.

383
Not exactly matching with the free running frequency and it has a constant phase if that is the
type of input of course if it has something other than this characteristic will have to do something
else, okay s as long as this is something which that happening that where exact frequency I do not
know there is a deviation and there is a phase that phase is whatever that phase is that remains
same that is not the random phase okay.

So as long as this is happening we have the error signal as this okay so let us simplify this little
bit so we get it should S2 and then S2 A + Ak S + A and we have W0 – WC S2 + psi 0 / S okay
and most you have learned Laplace transform in network theory you might have learn the final
value theorem so instead of doing this suppose I want to evaluate that theta e at + infinity right,
so that I can do by evaluating theta et and putting limit t tends to infinity same thing is equivalent
to if I just have s domain representation. theta s multiplied by s limit s tends to 0 so that is the
result of final value theorem so will be utilizing this we our target is to evaluate at time t tends to
infinity what remains what is the phase error that remains that should be my final phase error, so
let us try to evaluate this part so if I just do that you can see that if S tends to 0 I multiplying by S
so for this term S2 gets cancelled S turns to 0 this will have some value but I will still have a if I
just put that S tends to 0, this particular part will go towards 0 same thing will be happening over
here because here only one S will be cancelled so there is already 1s another S will be coming
from here if I put s tends to 0 that part also will go to 0 so this must be = 0 if I put s tends to 0
okay

so my final value is 0 so now you can see very nicely if I put a second order loop filter
immediately frequency has been tracked I know that because the way I have represented already
let us in the input for this one the means or the VCO generated one will be exactly tracking that
w0 okay no problem in that even in phase also there is no error so it tracks the phase exactly
accurately you might be arguing there is a certain difference the way we have if I just give you
that one of the earlier example, where we were putting the VCO circuit so if you see, whenever
my phase error is 0 that means this is exactly tracked okay no problem in that but the problem
that will be occurring is this is sin and this is Cos so there will be no matter how well you track
the phase and frequency.

So if you exactly track the phase and frequency there will be a Pi/ 2 phase shift so if you wish to
use that VCO output as a means demodulator, you have to give another Pi/2 phase shift to that
particular thing so this is absolutely mandatory okay that you will have to again from co-
sinusoidal, you have to translate into sinusoidal or sinusoidal to Co-Sinusoidal, you will have to
do that because, the input whatever phase it has there will be after complete tracking of the loop
there will be a Pi/2 phase shift that something we can see.

384
If this is Sin this is Cos okay so that Pi/2 phase shift has to be somehow eradicated and that can
only be done by employing a Pi/2 phase shifter again after the Vco okay so whatever Vco output
we get that is completely in synchronous with the incoming this one barring a Pi/2 phase shift so
now I can employ that Pi/2 so earlier whenever we said there was a phase difference of that w0 –
WC / AK when we were putting a first order loop filter that means that was already there plus
there will be another Pi/2 okay.

So that was the overall phase shift whereas here the phase difference has been track completely
so you have exactly Pi/ 2 phase shift that has to be somehow taken out from the circuitry
whenever you design your VCO right so this is what we have now design now you can just play
around with it you can actually show that if suppose whatever that we have written

(Refer Slide Time: 06:58)

A sin [ωct + ψ (t)]

it is this A sin W0t + Phi 0 we have written this right now if here this w0 also varies with time let
say it is linearly varies with time, so of course in modulation that will not be happening but for
some reason for a particular amount of time if that is the case your first order will not be able to

385
even track it, even frequency tracking will not be happening, that will fail you can immediately
prove that you won't get to a final value, that will still oscillate so you will not be able to track
that where as your second order will track that but there will be a constant phase shift, will be
able to prove that similar method you just apply employ just put w0 as a linear function of time
okay so if you just do that will be able to show that second order will actually give you a constant
phase shift, whereas if you now go to third order filter so higher the order

Probably the tracking becomes better and better so if you put a third order filter you will, you can
prove that it will track the phase and frequency completely so that something which will be
happening but remember the entire analysis has been done for a small error signal or small phase
error signals so if that is not happening you probably cannot argue about this outcomes okay so
this will be always happening as long as the phase and frequency are almost synchronized that
means they are very close to each other.

Juts there is little bit of deviation and we know that even a little bit of deviation can be
detrimental for our demodulation so we wish to make it completely synchronized that is why this
PLL is being employed okay, so now we will try to see that part if there is a deviation okay so let
say.

(Refer Slide Time: 08:59)

386
H(s) = 1
h(t) = δ(t)
·
θ0(x) = AK sin (θe(t))
θi(t) = (ω0 − ωc )t + ψ0
θe(t) = θi(t) − θ0(t)
· · ·
θe(t) = θi(t) − θ0(t) = (ω0 − ωc) − AK sin (θe(t))

I have a loop filter okay which is equals to 1 okay so loop filter is just that first order filter we
were trying to employ okay so this is something we are assuming of course it will have a low pas
characteristics but what we can also assume that 2 WC is very high so I can almost assume it to
be a band pass within the band of my interest and then accordingly I can write this ht to be a delta
function right because that is just a inverse Fourier transform okay hs is 1 so therefore ht will be
just inverse the Laplace transform of this one.

So that should be delta t okay so this is what we get now we do not actually put that Sin theta=
theta because right now whatever analysis will be doing that is not good for means hold good for

387
this small signal analysis, let us assume this part okay so if this is the case what we can write is
we can write that our equation okay so the equation is if we just go back to our filter or is in this
previous one okay.

So we just go back to this so there is sinusoidal after that I am just putting Ak hs is becoming one
right because that is what I have employed okay if there is a sinusoidal then I cannot really call
that as theta e I can just write that as if that is a sinusoidal of theta e multiplied by Ak so that
should be filtered over here, so my theta output derivative should be that one okay so I can write
immediately that theta output t the derivative of that must be Ak Sin of theta et can I write this
okay so that directly comes from that.

Phase module where no approximation is being taken now earlier we were taking an
approximation we are saying Sin theta et is equivalent to theta et that was the approximation we
were taking now we are not taking any approximation, okay so we just saying according to this
particular phase lock loop whatever means from the phase perspective whatever linear circuitry
we have or non linear circuitry we have put this should be the case okay so that is fine.

So this is the case now I can also write my let say incoming phase like I was writing previously
so that should be w0– WC into t plus Phi 0 so this is equivalent earlier whatever I was writing
that it is coming at W0t + Phi 0 so if I just compare it with the Vco free running then the phase
input phase which is actually theta it that must be equal to this we have already done that, okay so
this is that phase now let us try to write what is the theta e so theta e is basically theta it – theta ot
right.

Now I can derive mean take a derivative over here so immediately theta et derivative that
becomes theta it derivative – theta ot derivative, theta it I already know the derivative of that
should be w0 – WC because it is just linear of time and theta ot derivative I already know this so
that must be AK Sin theta et fine, so upto this there is no problem I can write that, write this one
in this fashion okay so once I have got this now what I can do because we do not want to now the
solving part will be difficult.

Because there is sinusoidal and the derivative of that solving will be little bit difficult but we do
not need to solve it we can actually say something about it, so let us try to actually Plot theta et
and means theta e derivative with respect to theta et let us plot that okay so if I just plot that it is
just a shifted sinusoidal.

(Refer Slide Time: 13:51)

388
 ·
θe(t) = 0
ω0 − ωc
θ3 = sin−1
AK
ω − ωc
θ3 ≃ 0
AK
ω0 − ωc
≪ π /2
AK
·
θe(t) = (ω0 − ωc) − AK sin θe(t)
0 = (ω0 − ωc) − AK sin θ3

Okay so it will look like this whereas it will be touching at 0 at w0 – WC because see this is
where theta et derivative I am plotting and this where I am plotting theta et whenever you put
theta et = 0 we get theta e derivative t is w0 – WC so this must be cutting because at that point
theta et is 0 so this must be cutting at this point okay, and this particular value how much is this

389
that should be w0 – WC – AK the maximum of the amplitude right okay now let us try to see
where exactly I will get that phase lock.

Okay or equilibrium so equilibrium means at the input to the VCO I have constant thing it is not
changing okay so that means this theta et derivative that is equal to 0 because that was the input
to the if you see that is the input to the VCO right this was the input to the VOC if that is the
input to the VCO if that input is 0, that means there is no point in changing the VCO now VCO is
locked so whenever that inputs becomes 0 that is where it will be in equilibrium okay so
whatever that is

So whenever theta e derivative is 0 I know that those are the equilibrium now let us try to analyze
where those equilibrium points are theta e derivative is on the y axis so wherever it crosses this 0
in y those are the equilibrium points so this must be let us mark them let say this is actually theta
1 this is theta 2 this is theta 3 this is theta 4 right like this there will be if I just keep on going
there should be a theta5 and so on and theta – 1 – 2 like that there will be many okay.

So there are infinite equilibrium point in this case that something we can see already now we
have to see whether this equilibriums are stable, or unstable okay stable means what is stable
equilibrium that means you give perturbation means you deviate little bit it will try to come back
to that same equilibrium okay so that is a equilibrium and even if you perturb it you just deviate it
and it will again try to track it back

and unstable is, It is in a equilibrium in perturb it will go away from that okay so it will go further
away again and it will again keep tracking and go away from this so those are the stable and
unstable equilibrium so let us try to see that okay so let say around this there was a initial phase
error okay, so see whatever happens because I could see that the overall phase trajectory should
be on this so it should be moving or along this whatever happens the loop equation actually gets
bounded over there so it must be actually moving around that.Sinusoidal this is what should be
have happening so let say initially I have a phase error right, so that phase error let us called that
as theta 0 okay so this was theta 0 at that point what is the derivative that is negative right so the
derivative of that is negative so immediately what will try to do because the derivative is negative
it must be trying to lift that phase up okay so that particular part so because the derivative is
negative and it will try to negative and it will try to go towards the 0 okay so if that happens.

Then immediately what will happen to the phase it can go up only in this direction okay that
derivative can go up only in this direction it is going towards 0 okay so if that happens the phase
is actually getting reduced and it is whenever you put a phase over here, it will always try to go
towards this same thing will be happening over here if you just put a phase error over there it will

390
actually try to track it back to this equilibrium so that is why I can see that event I am at that
equilibrium I give a perturbation

It will always be coming back to that particular equilibrium so I can immediately call that
equilibrium to their stable equilibrium okay so this is happen ing to be a right whereas you can
also test this or this equilibrium we will see that they just go away once you put over there in
those equilibrium if you put a perturbation okay immediately you will see it will just take it on
the other side so it will just take it away from the equilibrium so you can immediately say that
probably

In this particular equilibrium its locked so that will be happening so this for me theta 3 is a stable
equilibrium okay at theta 3 what will be happening so I had a guiding equation right so the
guiding equation was theta e’ =w0 – Wc – AK Sin theta et right, at theta 3 this must be 0 that is
what we are seeing that theta3 it is a equilibrium that means the derivative must be 0 so this I can
put a 0 so 0 is = w0 – WC – AK Sin theta 3 or I can write Sin or theta 3 = Sin inverse w0 – WC /
AK a very nice thing has happened if you now see what has happened if this is pretty small that
means w 0 – WC / AK is much smaller than Pi/2

I can write this theta3 Almost approximately equal to w0 – WC / AK remember what was a filter
I have put, it was H (S) first order filter, in first order filter when we did for a small signal
analysis the final value means we put the final theorem or we put the at t tends to infinity what
will be the phase error that was this exactly we are again when we start putting this particular
approximation again that this is much lesser than or phase error is much lesser than Pi/2.

Again see that we are getting back to same the so it is all consistent but what we also see that here
probably you can tell about more on equilibrium, so what kind of equilibrium we can get earlier
we were just solving for one equilibrium because we have already saying that it is very closer
okay that is much lesser then Pi/2 so we are already in this equilibrium and we are trying to solve
that okay whereas here if you see we have multiple almost infinite number of equilibrium and we
can immediately see that half of them are stable equilibrium and half of them are unstable
equilibrium or non stable equilibrium so this is something which is happening over here, so this is
the extra thing you can say and there also because for any this approximated analysis if you do
the general analysis from there if you again could back the approximation probably you should be
getting back the same result so that is what we are getting, okay so this characterizes the PLL
means almost in little bit of greater details.

Of course there are other means non linear analysis of this one but that probably will not be
required for this course but here at least we have understood what can happen to this particular
PLL when it can track the frequency or phase what are the condition of doing that all those things

391
here also we can also see that see when if you just try to see this curve where it crosses it crosses
means what is the minimum point that is actually w 0 – WC – AK okay.

(Refer Slide Time: 23:56)

(ω0 − ωc) − AK < 0

ω0 − ωc ≤ AK

So it is actually w 0 – WC – AK right so this is what we are getting so for some reason if this
particular sinusoidal goes above the 0 at x axis okay above the x axis will you have any
equilibrium point will not have any equilibrium point so that means there is a condition on this w
0, wc and AK which are the parameter of incoming whatever phase you are getting or frequency
you getting and the PLL free running frequency as well as some of the parameters of the PLL
okay like K

392
so as long as these things It is this must be less than 0 okay so if this happening then only will
have a tracking of PLL right, otherwise not because otherwise whatever curve we have put it is
just go above x axis and there will be no actually cutting point where your PLL that theta e dot
become 0 if that is not 0 then the VCO input will not be 0 so it will keep on changing the output,
right as long as there is something at the input of VCO, VCO will keep on changing the
frequencies so there will be no lock that will be achieved through the PLL so to achieve lock in
PLL I need to have this sinusoidal coming below 0.

And for that I need this condition or I should say w0 – WC must be less than = AK okay or we
should actually write a modulus of that because it might happen that WC > w 0 so that might also
happen whatever happens it should be this okay so that condition should be always satisfied so
and that relationship, must be there to get a lock in the PLL so which also says that your
frequency should not be too much deviating from the targeted phase lock loop free running
frequency that should be the case.

You cannot really hope to track back a particular input sinusoidal phase and frequency when your
PLL is deviating sorry the PLL has VCO and VCO free running frequency is deviating too much
so that is a ill design that you know your carrier is at 500MHz probably and you come up with
the PLL which has a free running frequency of 5MHz it does not make sense, it must be around
that 500MHz you know that exactly I would not be able to track but it must be around that so I
say that I will be creating a500MHz.

But of course there will be a deviation because of manufacturing and all those things so if this is
500.1 this might be 499.9 something like that and then the rest of the things it can track and that
is very clear because we need this, otherwise you do not have a tracking okay so this probably
ends our discussion on PLL so we have seen the effectiveness of PLL we have seen how to what
are the design criteria of PLL.

What are the things we need to do and what are the consequence if we cannot track it properly
that is also something we have already seen it is a very important integrated circuit so which
should be studied in depth okay, so what will do in the next class is a we will try to see in a
DSBSC demodulation how this PLL and the squaring of the signal which is the overall thing for
carrier recovery can be a very nice way implemented so that there is a famous circuit called
Costas loop will try to explain that part.

That actually gives you a practicalDSBSC demodulator not just multiplier because for multiplier
we need to assume that there is a complete tracking of phase and frequency so what will be
giving is our realistic practical circuitry. So we will try to see how the concept of PLL is almost
being employed over there okay thank you.

393
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-28: PLL (Contd.) and LTI

Okay, so probably last two or three classes we have been discussing heavily on PLL so we have
almost finished the analysis part of PLL, now what we wish to do is try to see a practical
application of it so that is actually DSB-SC demodulation where the PLL has been heavily used
so we will see that we will later on when we will study FM we will see also this particular thing
PLL can be also employed for FM demodulation we will explore that part and that is why
probably PLL is one of the most important circuitry in communication system. So let us try to
draw that circuit.

(Refer Slide Time: 01:07)

394
R sin(2θe) = 0

So it is nothing but your let us say m(t) cos (wct+theta i) that is what is coming okay, so this is
actually just m(t) is your message signal and we have done a DSB-SC modulation, so we have
multiplied with a cos Wct, but what might happen if I have a phase okay, theta i which might be
included in it okay. So this is coming we take it into two arms and then we use to multiplier
circuit over here that multiplier circuit actually gets input from a VCO so that is where tr hat
almost the PLL circuitry is being created.

So VCO will definitely create another cosinusoidal let us say okay, or it might create a it means
sinusoidal whatever it is let us say we create a cosinusoidal okay, so VCO will create a
cosinusoidal which is we termed at as 2cos(wct)+ let us say that is theta o right, same thing is
coming over here I give a –pi/2 phase shift okay and then multiply so this will be sum 2sin (wct
+theta o) right, this is what we get when both the arms are passed through a low pass filter okay.

ASo what will happen that multiplication term will be creating a this plus this and this minus this
right, let us say the minus because wc we have told already there is a matching even if there is no

395
matching we know that how to deal with that w 0 I can represent as w c-W 0 and + Wc, so we
can always do that okay. So frequency we do not have to really think about it is all about phase
okay, so there will be a 2wc +theta i+ theta o will be there whatever it is and there should be a
theta i-theta o low pass filter will take 2 Wct term will be only having this m (t) cos theta e which
is just the difference of these two.

Of course I am not writing that it is a function of t it must be a function of t okay, so this theta e
will be a function of t. so this must be my output and what I do I need to have a feedback so I
take it over here fine, I put another multiplier circuit this is almost working like a squaring circuit
so here also there should be a low pass filter what do I get because it is sin so cosinusoidal
multiplied by a sin it will be sin a+b sin a-b then sin a-b will be surviving so m (t) sin theta e will
be surviving.

I multiply these two right, and then I pass it through a low pass filter again which is a very
narrow band low pass filter okay, it just around DC it is try to pick that DC term nothing else,
okay. So that is this thing so this is a narrow band low pass filter so what do I get over here is 1/2
or I should not say 1/2 it is just m2 yes, there should be a 1/2 m2(t) it is cos theta e into sin theta e
right, so that should be sin 2theta e right, that must be get into this.

And now I am passing through a narrow band low pass filter so what will happen whatever that
m(t) variation around DC it will almost take a constant value, okay so m(t) might have some
variation right, m2 (t), because I am just taking a very narrow band low pass filter so it will just
take a constant term so at the input of the VCO it should be a constant sin (2 theta e) which be
coming over here, okay then from the VCO the output goes, fine.

See what will happen, VCO will be keep on running till what time till it gets some input which is
0 right, if it does not get that he will keep on running right, so when it is 0 then what will happen
to this sin R is a constant let us say sin(2 theta e) that must be 0, so immediately what happens to
the value of theta e that must be 0 that means it gets a I means the VCO will stop deviating its
phase that means this theta o will stop deviating when this theta e is becoming 0 that means we
have a phase lock in theta i and theta e.

And once that is happening what will be the output, let us see that it is the modulated output as
you can sorry, demodulated output. Because m(t) cos theta e if theta e is 0 this should be 1 so I
get m(t) back, so this is where I get my output back, so that is the beauty of the costas loop. So
costas loop actually is doing everything it is doing this m2 (t) term, as you can see this 2 gets
multiplied and I get m2(t) so that part is happening which was the original proposal that you have
to square it.

396
After squaring it is also putting a PLL circuitry because the main part of PLL which is a VCO
followed by a loop in a, with a low pass filter with a loop filter that is all happening over here
okay. So it has in a way both the things but they are combined in nice way so that it is even
producing the modulated output sorry, the demodulated output that m(t) is getting produced
whenever there is a locking okay, of course here also you can do the same PLL analysis, you can
show that if the phase errors are very large you would not be able to track it so all those things
you can do.

But of course we have already done that for PLL this will be just a duplication so we do not want
to go into that direction, but what we can now see that a very useful practical circuit can be
designed employing both the things the squaring, filtering then PLL okay. And our original
proposal just to remind you was something like this.

(Refer Slide Time: 08:32)

We have our m(t) cos(wct) you do pass it through a squaring circuit and then put a band pass
filter right, this band pass filter must be narrow band like that low pass filter and whatever you

397
get okay, after this you will be getting at 2 wct right, you put a PLL which has a free running
frequency around that 2 wct then you will get a pure, purely tracked sinusoidal once you get that
because this PLL will help you to track even if there are some phase error, so it will also track
that once you get that then you do a 2:1 frequency divider okay, you get kcos Wct.

After this you need to put a demodulator so this was the original proposal which has been very
nicely taken inside the PLL means inside sorry, inside the costas loop so this was proposed by
Costas and generally all the DSB-SC demodulator follows the costas loop because that is
probably integrate the PLL squaring circuit and all this frequency divider and everything those
are not required they integrated very nicely and give you the demodulated output finally, okay.

So that will almost end our I means discussion of circuit as well as in short we should say system
as well as signal for the amplitude modulated things, we have already I think few class back we
have already summarized all those modulation technique and we have also talked about I means
different possibility of imperfection that come within this we have characterized them in terms of
their advantages, relative advantages, disadvantages in terms of power efficiency in terms of what
kind of system we can use it for can it be useful voice or video we have also talked about can it
be used for broadcasting or point to point transmission if there are if it is bandwidth efficient that
is something we have characterized all those things we have already done.

Now the thing which is left and which is probably the most important part of communication is
characterization of channel and the effect say it good or bad mostly bad, bad effect of channel and
how you come back that channel that is something which will be means in next few class we will
dealing with, okay today we will probably start little bit but it is a channel characterization which
is come next and in the channel characterization one of the most important part is noise
characterization that in presence of noise, What happens to the modulation system means
whenever I put it in the channel there will be we have talked about in our some of the previous
classes that noise will be present even in transmitter as well as in channel and in the receiver, so
that will be there plus channel will have some other imperfections will, which also will be
characterizing.

In presence of all this what happens to our means all the modulation technique that we have
discussed so far, so we will try to means we have done a comparison but now probably the most
important comparison which is coming up is in presence of all this imperfection or impurities of
channel who survives best okay, so for that two things will have to do one is we need to
understand a system called linear time in variant system we need to understand the characteristics
of that, that is one part.

398
And the second part will we need to understand random process because without random process
or without the basic understanding of random process will not be able to characterize noise or
even interference okay, so these two things if you wish to characterize we need to have a good
understanding of random process. So we will first we will try to talk about the other than noise
other imperfections which are little bit easier to deal with and not many will be in detailed dealt
with in this particular course but we will touch them at least.

And for the noise part we will do our very detailed discussion on random process before going
into the analysis of all this systems okay, amplitude modulation to DSB-SC to SSB, VSB and all
other systems, so before going into that we will do a very detail regress understanding or rigorous
teaching of random process, so this is something which we do. So today what we will try to do is
we will try to discuss about this LTI.

(Refer Slide Time: 13:47)

x1 ⟶ y1
x2 ⟶ y2
(c1x1 + c2 x2) ⟶ c1y1 + c2 y2

399
Which is called linear time in variant system, so there are two terms actually one is this many of
you are already familiar with this but let us just give a brief over view there are two terms very
important terms one is called the linear and the other one is this time in variant TI okay, so we
will have to characterize these two things separately what do you mean by linear, linearity in a
circuit so a circuit is linear if we say, suppose I give a input to a circuit and I get a output y(t)
okay.

Let us say if I give input as x1(t) I get a output corresponding output as y1(t) and I give another
input x2(t) and I get a corresponding output of y2 (t). Now what we can say if we give a linear
combination of these two input so let us say c1 into x1 (t)+ c2 into x2(t) a circuit will be linear
probably you might have done that in your network theory course. If the output is also similarly
means with similar scaling factor linear combination of the corresponding output, so if x1(t)
produce y1(t) and x2(t) produce y2(t) then output of this must be c1y1 (t)+ c2y2(t).

Whichever circuit actually provides this functionality we call them linear circuit okay, all those
circuitry which are multiplier means or those quadratics sinusoidal will not provide this
characteristics okay, it must have a linear relationship that is something which has to be there. So
and you can do it for any number of signal okay, so as many linear combination you will take
always you will be getting at the output the linear combination of the corresponding output.

So this is something which has to be characteristics of a linear circuit, so whenever the circuitry
is like that we call that as a linear circuitry and what is time invariant.

(Refer Slide Time: 16:03)

400
That means if I give a signal x(t) the output y(t) should be just a replica of the input only with the
delay okay, so if this particular signal whatever component it has okay, all those signals if I am
talking about a time invariant circuit that means it might at most provide a delay to the input the
output must be a delayed version of that nothing else, there should not be any distortion of the
overall signal, okay.

So the output characteristics remain the same it just gets delayed as long as the circuit is time
invariant. So basically what we can say if I give a input as suppose this x(t) so y(t) must be
something which is the delayed version of that, so I can say x(t-T) okay , and there must be some
other coefficient that can be there but time in variant means the signal quality should not be
distorted okay, that is the most important part or I can say that every component should be
equivalently delayed whichever corresponding component it has, so let us say I havex1(t)+x2(t)
what will happen if I pass it through it.

So it be if there is a delay it will be equivalent delay to both those signals, okay so there will be
equivalent delay nothing else. If there is a composite signal and at the output they are differently

401
delayed then the composition will actually be distorted it will not keep the similar structure of the
input signal. So the circuitry which provides this particular functionality we call them as time
invariant as well, okay. So all the circuitry that we will be looking for now onwards are those
linear time invariant system, okay.

So whichever is linear the way we have discussed and time invariant and then from these two
property it can be proven that suppose I have a system okay, now system is characterized by the
impulse response of it so we just give a delta over here and of course there must be a output I will
be observing, and this is called the impulse response of a system. So if I provide a delta function
at the input that Dirac delta function actually and then whatever output I get that is actually the
means the impulse response of that system.

And we also know that if it is linear time invariant then it or we should say it can be proven that
if I now for this system let us say I put a signal x(t) then what should be the output, output should
be the convolution of this impulse response and the input signal this can be just proven from
those two property that if the circuit is linear and if the circuit is time invariant then I will be
always able to prove that means I am not doing that prove over here because that is this is not the
forum for that.

But we will be always able to prove that it must be this the output should be this, whatever signal
you put always the output will be this and the entire circuitry is completely characterized by the
impulse response which is this h(t) I do not need any other description. The circuitry is uniquely
identified by it is impulse response, and corresponding transfer function which is just the Fourier
transform of it.

(Refer Slide Time: 20:08)

402
 y(t) = x(t) * h(t)
Y( f ) = X( f ) ⋅ H( f )
H( f ) = ke −j2πftd

So basically if I say this is h(t) if I give x(t) I get a output y(t) which is nothing but x (t)
convolution of this sorry, h(t) okay, so if I just take a Fourier transform of this part Y(f) is the
corresponding Fourier transform of the output that must be convolution in time domain must be
multiplication in Fourier domain that is something we have already proven so that must be
X(f)H(f) right, so this H(f) is called as the as you all know called the transfer function of this
particular thing whatever circuitry you are talking about or whatever system you are talking
about.

So this either the h(t) which is the impulse response or the Fourier transform of this which is
called the transfer function of that particular system, this particular relationship should always
hold for a linear time invariant system or in time domain I should say this particular relationship
should always hold and the corresponding impulse response characterizes the whole system this
is something for a linear time-invariant system I will be always able to tell that, okay.

403
So now what I will talk about is about a system which is distortion-less, so now my target will be
to talk about a system which is distortionless, okay. So if I have to have a system which is
distortionless so you can now see that this X(f) and this Y (f) must have similar characteristics
okay, X(f) sorry Y (f) should exactly equivalent to the X(f) okay, it should almost look like
similar thing.

So if this has to happen then what do I first need, that the amplitude of this or amplitude spectrum
of this must be 1 because otherwise there is no possibility that X(f) and Y(f) will have identical
characteristics, so any other things will create distortion see any other things which are which
might not, might be still linear time invariant like a filter let us say okay, so if I just employ a
filter, filter will have some characteristics H(f) will be suppose let us say first order low pass
filter.

So it will have some characteristics where some of the frequency element will be means having
similar amplitude they will pass some of the frequency element will be suppressed and if our
signal targeted signal X(f) is not band limited the his high frequency term will be attenuated more
than the low frequency term, so it is getting distorted at the output, okay so filter always creates a
distortion right.

So if I need to have distortionless signal at the output so therefore the distortion less things
whatever I am characterizing that must be a all pass filter, it should have H(f) which is constant 1
in all frequencies so it just passes all the frequency with equal or I should say no attenuation,
right. So that is the amplitude part what should be happening to the phase let us try to see that
part, okay. Earlier we have done a exercise, suppose we have at the input we have two sinusoidal
okay, for those two sinusoidal if I need to have that signal not getting distorted that means both
the sinusoidal should be equivalently delayed okay.

So what should be equivalent phase difference that should be created while passing through or
what should be the equivalence phase difference in each of those sinusoidal that will be created
by passing through this particular system. We have already seen that it actually linearly scales
with the frequency, this is something we have already proven in one of the, so that means we can
immediately see if we now start taking more number of sinusoidal so it must be linear whatever
the frequency it must be proportional to that frequency, okay, the phase that will be created.

So therefore the phase spectrum should look like a linear thing okay, so it should be always a
linear characteristics stretching from-infinity to +infinity okay. What does this means this means
my H(f) must have a amplitude spectra which is 1 okay, and there should be a phase which is e to
the power -j2pif and there should be a constant thing right, because it should be linear with

404
respect to this is the overall phase so phase has already f so it this must be constant so let us say
some td I put. So this must be that H(f) okay.

So whatever I do that must be the H(f) if this is the H(f) let us try to see the corresponding
impulse response okay, or instead of 1 I can also write some k okay, so what should be the
corresponding impulse response okay, or let us also forget about that let us say I have a X (f).

(Refer Slide Time: 25:56)

Y( f ) = X( f )ke −j2πftd
y(t) = k x(t − td )

Forget about the impulse response that we will come back later on, so let us say I have a X (f)
which is the input that should be passing through a system which is H (f)so that must be k e to the
power -j2piftd this must be my Y(f) so whenever we multiply in a Fourier transform for a k is a
constant term, so we do not have to worry about that whenever suppose this X(f) has a
corresponding x(t), whenever we multiply with e to the power -j2piftd what happens in the time
domain.

405
I will be getting a delay okay, of value td so this must give me y(t) as k x(t-td), if k is 1 this is
actually the distortionless signal transmission so if I have a system like that I can immediately
corresponding time domain signal I can see that produces a distortionless signal, because it just
the whole signal it delays it that means the signal characteristics is not changing and whole signal
either it will not attenuate anything or if I have a constant attenuation it gives constant attenuation
to every time or every frequency component, so it will again not change the overall signal quality.

So what do I, why I am talking about this so that means in the channel whatever I am transferring
let us say m(t) cos(Wct) that is what I am transmitting, at the receiver what I expect, I expect the
same thing but now in between there is channel if I characterize this channel as a system I need to
ensure that channel is a distortionless channel, it should be linear time invariant of course more
than that I should be saying that channel must give me this characteristics, so it must be a
distortionless that means channel must be all pass filter which is ideal, it has linear phase
spectrum.

If the channel deviates from there, now we will be talking about the impairment of channel, if
channel deviates either in amplitude spectra or the phase spectra I will have corresponding
distortion that I have to deal with at the receiver end. So this is why we started with this thing,
once we understand this we know that what kind of imperfection I should be getting and means
what should I say for the misfortune of all the communication engineers the channel is not a all
pass filter, it is a low pass filter whichever way you see it.

It is a low pass filter it has different amplitude attenuation at different frequency component and
the phase is also not always linear. So there will be additional distortion which will be
characterizing in the next class that will be happening whenever you pass signal through any
channel, so we have to be ready to compared those things at the other end, okay or we have be
choose our channel carefully, okay where we can almost see this characteristics. So we will try to
characterize that in the next class, thank you.

406
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-29: Dispersion

So I think we have started characterizing the channel the last class we discussed linear time
invariant system and what kind of system we need for realizing a distortionless system this is
something we have done so for linear time invariant system we have already told that it has to be
linear that means input whatever you give corresponding output if you know if you give the
linear combination of the input so you should be expecting the linear combination of the output
and time invariant means suppose a particular delay XT within a system gives you YT if I delay
the input so if I pass t-t0 so the output will be just similar just delayed by same amount of time.

(Refer Slide Time: 01:14)

407
x(t − t0) ⟶ y(t − t0)

So it will be just a same output so basically if I give a delay it remains time invariant now output
also will be just delayed by that signal that signal strength and the output signal the way it looks
there will be no change in that okay so that was the linear time in variant system we have told
that it can be characterized by only one thing which is the transfer function or the impulses of it
and always we can prove that the if I put a input as XT the output will be always if I know the
impulse response of that particular system so it should be convulsion of those impulse response
and input function.

(Refer Slide Time: 02:00)

408
 y(t) = x(t) * h(t)
Y(F ) = X( f )H( f )

So always yt must be xt*ht okay so this is always true and the corresponding Fourier transform
should be transfer function multiplied by the input signal so this two things are always true for a
linear time invariant or system right so this is something we have told and then we went to a
special class of linear time invariant system so where we were talking about dispersion sorry
distortion less system so for that we have told that distortionless means the amplitude should
remain the same okay so what does that means amplitude remains the same means that HF should
give me the same characteristics means after passing it through HF

(Refer Slide Time: 02:52)

409
H( f ) = 1

The YF should have the same character tics HF sorry XF so therefore HF must be 1 so I can write
HF must be 1 if it is any signal I have to say that for all frequency value HF must be 1 so it
should look like amplitude spectra or the trans function the amplitude representation of that must
look like this all frequency component if XF is band limited so that means that XF you are
talking about is something like this then if I wish to get YF similar to XF I only need to ensure
that within the band HF is 1 beyond that whatever happens I do not care because any way I do
not have any frequency component of that signal beyond that band okay, so therefore I need to at
least ensure that if it is a band limited signal that within that band so –b to + b HF must be at least
1 beyond that it can be anything and if it not band limited signal then I have to say that HF must
be 1 at all frequency or I should say mode HF must be 1 for all frequency okay.

(Refer Slide Time: 04:01)

410
Y( f ) = X( f )e −j2πft0
H( f ) = 1.x(t − t0)

So, more precisely because phase we have not talked about and then distortion less means there is
also we have to talk about the phase and the effect of phase so what will happen if we have
already told that distortion less means I put a input to a particular system okay which does not
create any distortion so therefore the input should look same and the only thing that can happen it
might get translated, in time nothing else okay so there might be a certain delay to the signal okay
but the exact signal should be replicated at the output the whole signal might get translated not
some portion of it get translated to other portion does not get translated then probably the signal
will be different.

So that as to happen therefore my output should be whatever XT I give I must get something like
Xt-t0 so immediately I know for the corresponding HF we can now do the Fourier transform of
this one so output should be YF which is we already know that should be XFe to the power
-j2pift0 so this is something we know so this is the XF so therefore this system HF must be this in
so I can write my HF as 1.e-j2pift0 so this 1 is the amplitude part of that and this is the phase part

411
of this so I should say 2pift0 is minus thats actually the phase so if I plot that phase it should look
like this so it should be linear okay with the slope of this to okay.

So that is all that is happening for a means now we have characterized distortion less system that
means if you give a input to the distortion less system I should be expecting just the replica of
that maximum I can means take it into a account of it so this is what I have done XT I told output
must be XT-t0 if you say okay I can still take an attenuation, because that might be inevitable if I
transmit it, transmit it the signal over a channel this is the channel characteristics we are talking
about so if we transmitted there must be a attenuation but the attenuation should be with
respective of frequency that means all the frequency must have similar attenuation so I should be
able to see

(Refer Slide Time: 06:38)

Y( f ) = X( f )ke −j2πft0
f
H( f ) = k . x(t − t0)

412
If XT is there I will be able to see K into this where K is less than 1 okay so immediately k will
come out then this will become k so basically what happens my HF is mod HF is still the same
instead of 1 it is k okay and the phase is linear with respect to frequency so this is the character
tics of the distortionless system or I should now start talking about channel so because we started
characterizing the channel what I want from the channel whatever I transmit I need to receive
similar thing that is why the channel is characterize it should be ideally distortionless that means
it is neither creating any deformities or distortion in the overall signal character tics it might just
give you some delay of course that will be there because if I transmit over a long distance I know
that even the fastest of carrier which is light in free space probably that will still have a velocity
okay a finite velocity c.

And of course, for a distance there will always a delay so anything I will be transmitting there
will be a delay okay so that delay I can take no problem with that but signal there will be no
distortion okay so if I wish that my channel should have this characteristics it is almost like all
past characteristics and the phase should be linear so Ideally I want that and if I say it is already
we have talked about that if I say that if it is band limited then I just need to ensure that it is flat.

So the channel characteristics or channel transfer function is flat over that band of interest okay
now what will try to see is we will try to see if this is possible in the channel and if somehow that
is not the characteristics of the channel what happens ideally we will be expecting this but
probably we would not get that so let us try to see if this is not happen so I will just go through
two examples

(Refer Slide Time: 08:49)

413
 {0
(1 + k cos 2π f T )e −j2πftd | f | ≤ B
H( f ) =
|f | > B

One is that my amplitude is something like this so I’m expecting a signal of band width b so okay
so this is not the signal this is actually HF suppose I have a signal which is anything let us say
okay this is actually G or let us say GF corresponding there is a signal which is band limited
within –b to +b now I pass this signal through this, which is my channel characteristics so the
channel characteristics is the phase is linear but the amplitude has a curvature okay it isnot fixed
which is idea if this happens what will be my overall reception that is what I want to see okay so
let us say I will be transmitting particular pulse okay so let us say this the pulse so this is that
pulse or may be some other pulse which looks like something like this okay.

So which ever pulse it is I will be transmitting this pulse and assume that it means almost band
limited that means a whatever is outside its band okay that is suppressed or I might even
represent the pulse with the sins okay so I represent the pulse with the sink then it is band limited
but of course the pulse will not the time limited that is something I know but the sink will
probably will die down after sometime so this is, suppose if this is my pulse okay So that pulse
will be band limited that is something I know already because I have been represented it as sink

414
pulse now suppose that sink pulse is been transmitted over this channel characteristics which is
something like this.

So it is actually 1+kcos2pifT. e-2jpift. okay, when f <or equal to b and zero when f>b okay so this
is the characteristics I have taken as sinusoidal its just truncated sinusoidal so it was like this and
I have just truncated it up to – b and +b and this t is the period of that sinusoidal or cosinusoidal I
should say okay so that's my overall transfer characteristics or I should actually write it like this,
it is let us say 1 the sinusoidal actually goes like this okay so or I should say it goes like this, so it
is above one I have something called 1 above 1 there is a swing, swing of strength k right so this
is what is happened so that is the sinusoidal I have a sinusoidal and above 1

(Refer Slide Time: 12:28)

There is a swing of strength k lets say this swing is 2k so at this point it is 1+k so this is what it is
at f= 0 it actually 1+k so this much be 1+k and b whatever value it will get and this kind of I am
just creating a shape okay and so that it is not flat and I wish to actually see what happens to my
input pulse so this is my HF

(Refer Slide Time: 13:11)

415
Y( f ) = H( f )G( f )

( 2B )
t
= G( f )π ( 1 + k cos 2π f T ) e −j2πftd

So therefore what should be my YF I will be passing GF through that so it should be HF .GH

right so within the band of my interest okay what I will be getting I will be getting this so this
particular thing is HF is actually if I multiple this it should be multiplied with this one over a box
of –b to +b so I can write this as GF into a box function which is of Size 2b okay so –b to +b is
multiplied by that because rest of the part it is 0 so if I just multiply this particular function with a
box function then I will be getting that particular part rest of the part will be 0 so this I will be
getting into that whole function so 1+ kcos2pift. e-2j2piftd I can write in this way so now I have
two terms I can have GF into this now what I can write because this multiplied by one will
actually give me GF only because outside nothing else is there so that will be all zero

(Refer Slide Time: 14:34)

416
Y( f ) = H( f )G( f )

( 2B ) (
t
= G( f )π 1 + k cos 2π f T ) e −j2πftd

= G( f )e −j2πftd + (G( f )K cos 2π f T )e −j2πftd

k k
y(t) = g(t − td ) + g(t − td − T ) + g(t − td + T )
2 2

So I can whenever I multiply GF with the box function I will be getting just GF so GF into 1 into
this e -j2piftd that's one term + GF again multiplied by the box function that should be, as G (f) is
band limited that should return me G (f) only as long as this box functions beyond –b + b or this
bandwidth so I will be getting kcos2pift. e-j2pift fine so this is what I am getting now I wish to
see the output signal that means after transferring through this transfer function which is my
channel now what do I expect at the other end so YT must be Fourier inverse transform of this
Fourier inverse transform of GF is GT if I multiply by this we know the Fourier transform of that,
that should be GT right now this is the other thing where this cos function I can write as
summation of 2 exponential.

417
So it will be ej2pift and plus e-j2pft/2 so I will get a k/2 outside and then Gf into e to the power
plus that term and e to the power this will be there so, therefore I should get something which is
again shifted that shifted t-td-T and there should be another term K/2 g shifted as t/td+ T that is
all I will be getting if I just put the Euler's theorem and put ej2pifT + e-j2pift and there's a half
term coming out so that’s why I am getting K/2 so what do I get now I had a pulse ideally I
should have expected this right my target was that if I want distortionless channel then I should
be getting just the delayed version of that pulse that was all good but now I have two Fourier
terms which are further advanced delayed or came little bit earlier okay so if I just have a pulse.

(Refer Slide Time: 17:09)

= 2T + Δ

Let us say that is this thing which is this pulse so what will happen so first part will be just the
delayed version of that pulse so this will be delayed lets say my tdso it will be delayed by that
amount the other part will be it will be further delayed by td +t right and attenuation there will be
some attenuation which is K/2 so there should be some attenuated pulse of same duration where
this point is td + t and there should be another point where this is probably td-t right so the

418
summation of these two things whichever way you represent it, it will look like this right the
pulse will look like its summation of all these three

So what has happened now it actually starts from this td -t and it goes up to td+t plus the pulse
duration let us say that duration is delta okay so overall pulse duration is now happening to be td
+t + delta that is this end point – td + t right so it happens to be td gets cancel so your pulse rate is
2t+ delta which should be ideally just delta but now I am getting a distorted pulse this is
definitely a distorted pulse and the pulse width has been extra stretched by this amount of 2t.

And you can see thats just coming from this cosinusoidal, okay so make this flatter this co
sinusoidal will go away and you will get your pulse back exactly this particular part that we have
described is called dispersion whenever you will be discussing means later on in other courses
when in we will be discussing about digital communication will see that, that has a profound
effect in pulse communication on digital communication so whatever pulse will be putting what
happens whenever you have non ideal channel. The way we have described that you have some
amplitude variation of the transduction of a particular filter or particular channel then you will get
dispersion that means it actually broaden your pulse so if you have a continuous stream of pulse
which are carrying information okay the pulses are carrying either is it 0 or 1 accordingly the
pulses are encoded if you have that and if the pulses are getting broaden what will happen it will
actually spill over the next pulse okay because you have a dispersion or the pulse broadening

So it will spill over the next pulse the problem is more it spills then the next pulse detection will
become difficult so this is a particular thing which is called inter symbol interface ISI here in this
term in digital communication more often than" analog communication of course the same effect
will be there in analog communication so in analog communication you do not have pulse you
will see whatever is transmitted that is getting actually stretched so every in time whatever is
there that is getting stretched .

And they are getting distorting the other time signal in a means which are in advance in time
okay they will keep distorting it okay so that distortion probably in analog communication will be
there you would not be able to take them out in digital what happens because it is just encoded
signal either 1 or 0 so you still have possibility of detecting it even thou there is a distortion okay
so suppose I’m transmitting 1 or 0 its just is it 1 or 0 that’s all I have to detect if there is a spill
over if I can still say that okay by the pattern of the signal I can detect that it is 1 or it is 0.

So if I can do then it is good enough okay but there will be a point when I will not be able to
detect that and that is where the inter symbol inferences actually becomes very severe so we will
see the kind of channel impairments we are talking about here where the channel transform
function is little bit non ideal there will be creating this is kind of interference the kind of

419
interference that we are creating that is within the same channel that means the frequency band
we are there we are continuously transmitting signal the suppose it is a pulse range so single
pulse is giving interfaces to the next pulse that is what called within inter channel interferences so
the interferences is created by the same signal.

So the signal itself is the creating interferences to himself so that’s the problem whenever you
have a non ideal characteristics of a particular channel okay and this particular phenomena is
called as dispersion so what will generally happen if I can just given example so if you see I have
suppose a pulse stream of this transmitted so what will happen once if it is a non ideal channel
then what you will see.

(Refer Slide Time: 23:15)

Once the pulse is transmitted it will have this kind of distortion because most of the time we will
have low pass effect in the channel and low pass effect will start creating this thing it will start
smoothing the sharper edges okay so it will start creating this kind of thing okay now this is 0 if
this sharpening sorry this de sharpening of the edges is too much then what will happen it will go

420
like this and it will go like this so 1's & 0's will be its almost like capacitor charging and capacitor
discharge because we will be anyway low pass filter you can always realise with Rc circuit right.

So if the channel is low pass filter it will be almost like passing a pulse through a RC circuit and
this is what will be happened so what is eventually happening you can see now the effect of inter
symbol interferes it is just happening because of the non ideal character tics of the channel
because you are passing this pulse through the channel and this is what we are getting and then in
this portion can you now detect whether its if the pulse was transmitted like this you can easily
detect it is 0 can you now detect whether this is 0 or 1 you cannot probably because for detecting
whether this is 1 or 0 you will probably put a threshold.

And you will try to see whether the signal level is beyond that threshold okay its below threshold
you say it is 0 if its above threshold you say it is 1 but that is getting mark here over there you
won't be able to detect that so this is where means you previous signal its actually because the
spurious the spreaded part of the previous signal has not really vanished and that is creating
interferences over here okay so this is what happens.

So within the same channel it is just the previous part of your signal is creating interferences to
you which will be happening if you have non idea channel, okay so let us try to see if the same
thing is happening in phase so I have a modified phase if that is happening okay so let us say I
have a channel which looks like this.

(Refer Slide Time: 25:30)

421
| H( f ) | = 1
θh(t) = − 2π f0 − k sin(2π f T )
Y( f ) = G( f )e −j2πft0e −k sin 2πf T k≪1
Y( f ) = G( f )e −j2πft0

I have channel which is ideal in amplitude so this is 1 but the phase part how it should be, it
should be linear with respect to f but suppose it is not so itis -2pift0 and there is some modulation
-ksin2pift and we are assuming that this k <<1 the most of the time we will see this non ideality
of the channel will not be too heavy in the previous example also there was in the amplitude there
was k cos so that k value will be pretty low okay once the k value is pretty low there will not be
still some spreading but it will not be that severe okay so the other two pulse delayed pulse we
have generated if I can just give those bring that example back.

(Refer Slide Time: 26:31)

422
| H( f ) | = 1
θh(t) = − 2π f0 − k sin(2π f T )
Y( f ) = G( f )e −j2πft0e −k sin 2πf T k≪1
Y( f ) = G( f )e −j2πft0

So if this particular k value is very low they will have a lower strength and the corresponding this
one will be of very low strength so the pulse spreading that is being happening will be pretty low
in strength right so that is what will be happening but so this is something where the non ideality
will not be very severe okay so that is why I wanted to tell this but here also probably the non
ideality is coming due to this portion that will probably not be very severe so whatever that is if
you just consider this lets try to see what will be the effect of it okay.

So again let us try to compute the YF that should be let us say I have a GF multiplied by the filter
transfer function okay which is one into e to the power minus this or e to the power this so it
should be just e-j2pift0 .e-ksin2pifT now we will be putting some approximation over here so GF
this is fine e-j2pift0 now this you expand in tailor series so will be we are trying to put this

423
approximation we expand in Taylor series and anything because k is in the argument right so
there should be terms of k,k2,k3 and all those things higher order terms because k is very small
higher order terms we can neglect so I can just take the first term and the second means just
varying with this one k that terms so first and second term.

So I can write this as 1-jksin2pifT right this is something I can write now just expand it so this is
as it is e-j2pifto so the first term should be as it is the second term should be –jk again I have this
sin so I can write that like the previous one again e to the power j-? divided by 2J there should be
a 2j coming out and I should be getting this gfe-j2pift0. e-j2pifT and then minus minus plus so
that should be jk/2j and then j gets cancelled everywhere and gf e-j2pift0. e-j2pifT right this is
what we get again I want to see the pulse that exactly the same if you see just there will be minus
sign so this becomes gt-t0 this if you do Fourier transform or inverse Fourier transform this one
will be –k/2gt-t0+t and then this should be +k/2gt-t0-T. so I’m getting three pulses again similarly
there will have they will be creating some distortion.

And there will be a pulse ready so whether we get a pulse transmitted through a transfer function
which is non ideal either in the amplitude or in the phase I almost get the similar effect that there
will be a distortion which getting created in the pulse and the distortion is nothing but it broadens
the pulse and it start creating symbol interface okay so this is something we have seen and we are
calling this as dispersion next will see if the channel is little bit differently non ideal that means if
it is showing non linear characteristics then we will probably see that it will not only create
interference to his own channel it will start creating interferences to other channels okay.

So will see that will be more detrimental in our system where frequency division multiplexing is
being used so if you have multiple signals the transmitted over multiple frequency band 1
particular thing if the channel is non ideal it will start creating Fourier frequency term into the
other channels okay so that’s the other effect of channel we will try to examine in the next class.
Okay. Thank you.

424
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-30: Channel Nonlinearities and Multi-path Effects

Okay so let us now try to see the nonlinear effect of channel so when we say a channel is
nonlinear we can say if suppose I have a signal called GT.

(Refer Slide Time: 00:35)

425
 y(t) = a0 + a1g(t) + a2 g 2(t)…
g(t) ⇔ G( f )
G( f ) * G( f )

Which is being transferred over the channel and after passing through the channel if I expect the
output Y T which is just not the linear means part of it or linear means this transformation of it.it
also gives some square cube and all other terms so if this looks like this some a zero which is the
constant term then a 1 into G T + some a 2 into G a 2 into G Square t + order dot has many things
are there.

So up to this the channel is linear because whatever you give you get the linear means all the
linearity property of that LTI system we have talked about that will be satisfied but if you just go
to this whenever it start taking the square term or any other higher term you will see that it is
basically creating means that LTI the property of that is not no longer holding and it will start
creating some spurious frequency term that is the something we are interested now so that's what
will be happening we means it's little bit clearer because you can see if as long as it is just up to
this part the linear term or the first-order term then what happens.

If I do a Fourier transform so the frequency component that will be involved in it will be still the
frequency component that is already involved in GT ok so it will just be equivalent frequency
component in the output whereas if I start taking this term what does this means G Square t in
frequency domain what that will be GF means if the GT has a Fourier transform of Gf then it
should be convolution of g f with itself whenever we convolute very simple example is if I take
that box function okay.

If I convolute this box function with itself this will create a triangular function so it will look like
this and if this is band limited to B that box function the way it is this will go up to -2b +2b the
Band width gets doubled and so on if you start creating taking G cube it will actually occupy 3
times the bandwidth and keep on increasing one after another okay so I can see already if I take
the nonlinear part nonlinear portion first of all just think about this signal this particular signal
okay.

This is the Box signal coming from sync function so if my signal was sync okay Sin c,
corresponding G F should be this. if I just take that second order term what will happen I will get
this so within the band what is happening there is a distortion already the box function has been
transformed to a triangular function so there is a distortion within the channel but the good means
bad part is it is now starting to create some distortion out-of-band okay.

426
So suppose I had expectation that my particular thing like voice it is band limited it is up to -B
to+ B I modulate it and put it in the channel right and the neighboring channel I also put another
voice so those two voice are being carried over the channel if the channel has non-linearity then
for this signal as well as the neighboring signal which is also suppose a box signal okay in
frequency domain so for both of them, we'll start creating this nonlinear effect so they will spread
in frequency domain.

The earlier one we have seen when it was having means dispersion it was getting spreaded in
time domain here what we are seeing due to nonlinear effect probably we are getting spread in
frequency domain and that is now creating a inter channel interference so earlier one was intra
channel interference this is called inter that is something which will be happening whenever we
pass it through a nonlinear channel but of course what will happen the channel is it might be non-
ideal but these coefficients will be very small because generally channel will not do too much of
means will not create too much of non-linearity it will create some amount of non-linearity but
the coefficients will be pretty small and accordingly you will see the effect of it okay but
whatever it is you have to always keep this in mind that channel might show some non-linearity
and accordingly you have to come back okay so how we will come back will later on will see but
this something you have to be sure that the channel might be the ideal channel distortion less
channel we have talked about it might not be this.

So it might have that dispersion which we have discussed already if the phase is not non-ideal or
the amplitude is non-ideal in the transfer function or the channel might be just a nonlinear thing
and which will create this kind of interference so let us just give one example so if we just have a
signal YT.

(Refer Slide Time: 06:19)

427
 y(t) = x(t) + 0.000158x 2(t)
x(t) = 2000 sinc(200πt)

( 2000 )
f
X( f ) = Π

y(t) = 2000 sinc(200πt) + 0.316 × sin2(200πt)

( 2000 ) ( 4000 )
f f
Y( f ) = = Π + 0.316Δ

Which is something like this which is let us say XT + some this is just an example and I have
taken from the books let's say this is this is my YT as you can see it's already the nonlinear
coefficient is pretty small okay so that X square T the corresponding coefficient this is small that
is what will be happen generally okay now let us see let us examine what kind of effect will be
getting with this.

428
So let us say I put a XT which is something like this 2000 sin c 2000 piT okay so it's a sinc
function in frequency domain why we have taken a sinc function because we want to represent it
as a box function okay in frequency domain so this is our X T so definitely X F will be a box
function of duration 2000 so that, that is the case if I pass this thing function over the channel
then my Y T should be 2000 sinc 2000 Pi T + this will be there and then 2 000 square so 2 000
square into this will give you this number 0.316okay or let us say 2000 square not 2000 square
you multiply this by 2 000 you get this and there will be another 2 000 which should be left and
then you get a sinc square okay.

So I can write this as this is just a means if I take YF this is just a box function okay of strength
2000 or sorry the width 2000 and this is just 316 if I take and this is actually a triangular function
so I can take a triangular function of this one 4000 because 2000 convoluted with 2000 box
function so that creates a 4000 band triangular function so it I just slot them they look like this the
first part will be thats this part that remains the same function.

(Refer Slide Time: 08:57)

Which is a box function running from this 1000 to + 1000 okay and the next one is if I see this
one it has a strength of 0.316 so at this tip it should be 0.316 and it should be going as a triangular

429
function going from -2000 to +2000 okay what will be happening I will have addition of these
two so if I just see it should look like this and there should be some part right so this should be
my overall output function Yf.

Now you can see there is a distortion because ideally I should have expected this thing this is the
distortion part which is coming in band so non linearity actually has two effect it corrupts your
signal and it corrupts other signal as well so it has a distortion and not only that out-of-band it has
created some spurious thing which will create distortion to others okay so you have to be very
careful about a non linear channel.

Because nonlinear channel creates not only problem for you it will also additionally create
problem for all your neighboring FDM channels or it starts creating inter channel interference
that is something we should be very careful since the channel is nonlinear we should be very
careful about it okay and accordingly we should combat so there are methods of combating
nonlinear channel but right now we are just stating what should be the ideal channel and if there
is deviation what should be the effect on the signal.

okay because whenever we have said we wish to do communication we have almost assumed as
if the channel was ideal so you have characterized what do we mean by ideal that means
whatever we transmit we are almost expecting all the time that at the receiver I will be receiving
almost similar thing this is probably not true as you can now see okay so it can have dispersion
we have already talked about that it can have non-linearity this is something we are now talking
about next we will talk about something else.

Okay which is called especially being observed in a wireless channel it is called multipath effect
of fading okay so what is this whenever you transmit a signal with your antenna most probably
you would not be able to restrict means if the antenna is Omni directional then you would not be
able to restrict the signal to a particular direction and generally in many cases that is not the
intention because you do not know exactly where your receiver you target your antenna to a
particular direction with a very pencil precise pencil beam probably you miss the target you do
not get you do not transmit a signal where the user is okay.

So that is not very good so generally I would not bother myself to see where my recipient is I will
just broadcast the signal on the air so that means I need only directional antenna and which will
broadcast in every direction the signal and what will happen when you are broadcasting you are
not sitting in a free space right means and then transmitting it you are sitting in some locality
where there are buildings there are trees there are many other things which can be act which can
act as a reflector to your signal.

430
So what might happen suppose I have reflector over here this is my receiver antenna this is my
transmitter antenna so whenever I transmit it goes in every direction so the one that goes over this
direction will directly be received one that goes in this direction okay this should be reflected that
or some portion of that will be reflected back it might absorb something it might scatter
something another direction but some portion of that will be reflected back and that will still be
received over here This particular model is called multipath effect okay

even in other scenario where we have a guided communication like let us say fiber-optic
communication okay so if we have a fiber-optic communication there also it will be it will have
some diameter of your code where you are launching the light and then due to total internal
reflection.

So basically the code looks like this and there is a cladding of that fiber and you launch the optic
means light and then it gets totally internally reflected and gets guided but what will happen there
also there might be different rays launched at different angles and they will go through different
path length okay.

So there also the same phenomena will be happening so whatever happens this might be a very
low dimension that is why there will be not much difference between those rays whereas here
there will be huge amount of difference today that goes directly to a receiver and the Ray that
takes a d2 and there might be multiple reflectors multiple such things can be happening okay so
all those things will get inside this effectively what is happening effectively suppose let us say I
again.

Let us take a means example of pulse I transmit a pulse some portion of its power is getting
directly linked so the delay on this will be much lesser whereas some portion of that power is
getting on this link so there the delay will be higher so basically a multipath means it's different
delayed signal replica of that same signal is coming and sitting at the receiver okay this is what
happens whenever we talk about multipath fading okay.

(Refer Slide Time: 15:36)

431
H( f ) = e −j2πftd + αe −j2πf(td +Δ)
= e −j2πftd (1 + α cos(2π fΔ) + jα sin(2π fΔ))

= [1 + α cos(2π fΔ)]2 + α 2 sin2(2π fΔ)

α sin(2π fΔ)
= 1 + α 2 + +2α cos(2π fΔ) exp(−j{2π f td + tan−1{ } })
(1 + α cos(2π fΔ)

So let us just try to see a channel model where only two rays we can think about okay so that two
Ray model one is straightly going to the receiver and the other one is getting reflected from
another reflector and coming back okay so the first one the Ray, which is going directly that'll
have some delay because transmitter receiver are not co-located so there should become delay of
TD after which you will be receiving okay and there can be further signal which gets a little bit
more attenuation let's say alpha why that is happening because it is taking a longer path then
getting reflected that might absorb some amount of signal.

432
So that will have higher attenuation probably and after that it will also go through a delay which
is definitely greater than TD because it is taking a longer path so that should be TD +some delta
let us say and these two signal will be added simply at the receiver because receiver cannot
separate them they are just coming from different direction receiver also has omni directional
antenna because he also does not know from which direction his signal will be coming.

So he has to keep Omni directional antenna so all the things which should be coming from that
particular transmitter will all be linked to him they are all in same frequency so you have no way
to separate them so they will all come into your receiver and will be received so if I now see the
channel even though it was usual channel that has a different characteristics now so this is now
my channel characteristics due to this propagation model okay.

If just two rays our there if there are multiple rays I have to put that many arms with all different
values and different Delta values I have to put all those multiple arms and all of them should be
summed over here and that is why we probably do not call that as two path it is multi path so
multiple paths can be there all of them should be accounted for my signal versus for simplicity
and to get some insight were just taking two ray model or two path model okay.

So if this is the case what the transfer function of this filter let us try to first identify this is just a
delay element so that should give me if I try to characterize HF that should be just e to the power
-J2 piFTD it is just a delay element of fixed delays TD so it should give means I have not taken
any attenuation over here so it should give me ideally 1 as the amplitude and the delay should be
this + because there is a adder so + there should be a attenuation of alpha so this should be alpha
into e to the power-J2piF this delay should be TD + T right so this is what we get. This is my H
(f), now lets try to see the amplitude part of H (f) and phase part of H (f) okay, so I can take out e
to the power -j 2 Pif td what do I get, I get 1+this I can write with the Euler's theorem again as
cos + j something okay,so I can write cos 2 Pif so wait a second 2 pi f td is already gone so 2 pi ft
+ j sin 2 pifT right I can write is it t or delta t oh sorry delta I have just written t so that should be
delta.

So I can write it this way right now I have to get the phase part and the amplitude part so this is
already a complex one I need to first try to evaluate the amplitude part of that so that should be
square root of this square so 1 + cos 2PiF delta square + this square right so sin square 2PiF delta
right and there should be a alpha right somewhere we have missed that alpha by no alpha okay so
that should be alpha square is we missed that alpha.

So if I just simplify this what we get root over 1+ this should be alpha square cos square + alpha
square sin square that should be giving me alpha square + 2 alpha this should be means that is 2
alpha cos this one right that is my amplitude and also similarly I can evaluate the phase that

433
should be just an inverse this-this right so I get this phase is already there + that tan inverse part
so I should be getting exponential -J I get 2 pi FTD which is this phase + tan inverse that this
thing divided by this whole thing right.

So I can just write that part it is just this part comes in the numerator and this whole part goes in
the denominator ok so I have now evaluated for that HF the amplitude and phase vector or phase
part right now let us see what kind of amplitude and phase it has so this amplitude if I carefully
observe it what is happening it is actually with respect to Delta it has a modulation or it is a
sinusoidal okay basically if I wish to plot that amplitude it's a periodic signal I should say well
this will be happening in the amplitude.

(Refer Slide Time: 21:52)

434
 1 + α 2 + +2α cos(2π fΔ) …
n
f=
2(Δ)
α=1
| H( f ) | = 0

So this is HF there should be some with some period it's actually oscillating so that depends on
this and then phase if I wish to plot so this is suppose the phase so phase has a linear part this part
+ tan inverse something which is also oscillated because we have a cos term and sin term so that
is also oscillatory so phase will also be with some other oscillation will have oscillatory part
because it is also a periodic signal whatever oscillation that will be that might not be exactly co-
sinusoidal or sinusoidal but that will have some oscillatory part so whenever we do a multipath or
two path fading so this is the impairment or this is the impurities that we see in the channel
characteristics and if you just try to characterize it little bit more what can we see let us say I had
that amplitude spectra right that is one + alpha square + 2 alpha cos 2pif delta, delta right so this
was the case this is something we were having ok so in this if we just put this F equals to N by 2
delta okay if you just put this okay so what do we expect over here if I am putting this.

So immediately what do we get so if we have this n now I can put at a start putting n as odd and
even right so let us say n I put as odd immediately what will be the value of this, this must be -1
right that whole thing right and then if you just calculate this how much this will be1-alpha whole
square, square root root so that is 1-alpha now let us say alpha equals to 1I am just approximating
that both the paths are having similar strength if this is happening now what do I get I get nothing
because my HF modulus of that gets 0 so all the frequency term which has this value where n is
odd will give me nothing at the output and if I just take it as even.

And if I take again alpha equal to 1 this will be just instead of becoming 1-1 this will be 1 + 1
okay so that will be a huge strength that will be coming to my receiver so what I can now see
depending on this f I have a different characteristics basically the HF is, this undulation is there
that means some of the frequencies are really getting heavily attenuated okay if alpha is one this
will almost go like this okay so this will be zero in some of the frequency in some of the
frequency is very high so that is exactly what is happening which is it which is termed as you
might have heard about this term that's called frequency selective fading.

So what my channel the way I have characterized it I have taken two path model and with that
two paths I have characterized the overall characteristics of the channels okay so that HF I have

435
evaluated then I could see that, that HF phase and frequency response both have some
undulations or some modulation or I should say it's periodic in nature and then going deep into
the details of it what I could do is I could set this alpha equal to one and then try to see if I put
this F as n by two delta So two delta gets cancelled and this is PI into n right and now n if I put
odd or even I get different result for this cos it might be + 1 or-1 whenever I put odd it is -1
whenever I put even it is + 1 and accordingly I can see a destructive interference and constructive
interference is being created so different frequency term because different n I put different
frequency I will be getting for different frequency I will be getting at different characteristics and
this exactly is termed as frequency selective fading that means your channel acts differently over
different frequency band okay and there are methods to actually combat these things.

S o what you do if you really wish to transmit a very broadband signal over a channel which has
this kind of characteristics What will happen some of the selectively it will actually attenuate
some of the some portion of the signal and selectively it will enhance some portion of the signal
or it will transmit as it is some portion of the signal that's not very good instead of that if you can
segregate your entire frequency band into smaller bands and in each band you transmit something
then it will have less effect of frequency selective fading and that is what is being done in
orthogonal frequency-division multiplexing, which probably is not the topic of our discussion
over here.

But this is what people do instead of utilizing the whole band they subdivide the band into
smaller bands and they try to transmit something on that those smaller bands independently so
that they can be detected independently in a of course in smarter way which is part of that OFDM
but this is the basis for doing that and that is why in a typical wireless channel people do this
because you can see that multipath fading comes from a wireless channel and we could also
understand the frequency selective fading part of a channel this will always come in multipath
channel how do we combat immediately this comes to our mind.

because in this particular course We won't be actually dealing with multipath fading in most of
the time so maybe we can just give some hint how we can combat this so whatever has happened
to the channel can be reverse that so that means a particular portion if we just go back to this
same channel suppose let us say some part of the signal has been delayed but less delayed some
part of the signal has been higher delayed and some part of the signal has not been attenuated
some part of the signal has been higher attenuated okay. If I just do the reverse thing okay

so I also almost the reverse because whatever has happened in the channel I want to really negate
that or nullify that what I can do if I can just do the reverse thing that means whoever has got
lower delay can I give higher delay to him and whoever has got higher delay can I give the lower
delay to him whoever has got higher attenuation I give lower attenuation so this is actually called

436
channel equalization this is something which you will be seeing in your digital communication
course heavily channel equalization is a very important factor for your receiver designing.

But that is what is channel equalization that means you assume that channel will have this
multipath effect and you try to equalize or that means you try to reverse that and that is being
generally realized with a tapped delay line so that means you actually take the signal you tap at
different power levels and then you actually adjust with a particular attenuation factor as well as a
delay factor and try to again add them together to negate the channel now what will be the
coefficient optimal coefficient of this delay.

And as well as this alpha that you have to actually understand or you have to know that is why
what they do they train the channel that means you first transmit a known signal over the channel
try to adjust this parameters try to see whether that known signal can be better realized after
doing all these things adjust all those parameters of alpha that means the attenuation factor that
you are putting in the delay factor you are putting you equalize them this is actually called the
equalization you equalize them.

So that you get a better response then once you have characterized the inverse of the channel and
assuming that the channel is not time varying that means the channel is not varying over time this
reflection and all those things are almost fixed it is almost coming from building not from a car or
something which is moving up object if this is the assumption underlying assumption then the
channel will remain the same and then the equalization that you have done that you can use for
unknown signal detection okay.

So this is what people do for doing channel equalization and of course in this course will not be
dealing with that but this is a very important phenomena that happens in the channel and you
have to means I have just given this example to let you know that how you can actually combat it
physically okay so this is something which almost means tell us what are the things that can be
there in the channel and these are the things which you have to combat we'll probably we still
have not touched something which is another impairment which will detrimental affect our
transmission that is called the noise.

So next class onward what we'll try to do is we will try to characterize noise first because that
means even if you do not have wireless channels because noise can be even be generated at
receiver so even if you do not have a channel still noise can be there so it is that detrimental so if
you are transmitting means the transmitter and receiver area the same location you do not have
channel no impairments coming from the channel still you can have noise because your
transmitter receiver can generate noise and we will try to characterize noise.

437
So and that is why probably noise is the most important thing which has to be combated and we
will devote quite number of classes towards understanding the noise better and then towards
understanding how we combat noise or in presence of noise what kind of things you should
expect so those things will try to analytically devise those things and we have to really to get a
good understanding of those things we need to have a good understanding of random process so
our next few classes will be devoted towards understanding random process okay thank you.

438
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-31: Probability Theory

Okay, so far we have discussed about signals little bit of system and then try to describe about the
modulation scheme which are employed for amplitude modulation mostly and then we started
characterizing different kind of amplitude modulation that are there I have done a comparison
and finally we discussed about little bit about the channel okay. And then we said that one of the
most detrimental part of channel which will be probably either channel or receiver or I should say
rather impairment is noise this is something we have already said.

But to understand noise probably we have to really understand random process, so what we will
try to do probably give you a brief overview of random process but I do not assume any prerequi-
site. So what I will do I will give a first a brief overview of probability theory basic probability
theory and then from there we will go to random variable and then from there we will develop the
theory of random process probably which will be all targeted towards communication or charac-
terization of noise.

So let us start with the concept of probability, so we all know there are events which are means
eventually looks to be random to us, so let us just discuss one event which is almost known by
everybody like either tossing a coin or rolling a die right so whenever we do that we say that it is
random but most probably it is not random as we state that means whenever we toss a coin if we
can calculate all the forces that that is being exerted on that coin lets say which angle I put it on
my nail and then when I toss how much force I give in which direction and the air friction and all
other things if I just gravitational force how it change changes with time so everything if I calcu-
late and then probably elasticity of the ground where it is hitting.

So if I can calculate all these things probably there will be no randomness associated with it I can
definitely calculate and tell you what will be the outcome of this event okay but as I have started
describing you might have seen that there are lot of process a complicated process involved and
associated physics involved in it. So if I really wish to means take a journey into that and try to

439
calculate this will be a tedious job which as in any engineering or scientific study we do not want
to do that we really wish to get some extract or some meaningful information from there and then
try to see whether that can be used further.

So this is something where probability theory came into picture what people started doing they
started looking into that particular random event I have talked about that tossing a coin instead of
going through the entire physics of means tossing the coin means exerting the force and then coin
flipping in the air and dropping in the ground, so instead of going through all this physics they
started thinking that okay given all those things let us try to see what is the outcome of that okay
that experiment.

So if it is a coin and we assume that the coin whenever it is landing it is not landing on its side or
edges so it will be probably landing on either of its face so there are only two outcome of this en -
tire process.

(Refer Slide Time: 04:23)

440
 n1
pH ¿
N
N −n1
pT ¿
N
lim n1
1
pH ¿ N →∞ =
N 2
1
pT ¿
2

So one is we call it head or tail okay so these are the only two outcome of this whole experiment
right then people started saying okay if this is the only two outcome why do I have to go through
all these details let us try to see intuitively can we give some understanding of this events okay.
So because finally I need to see whatever happens what is the overall outcome and what is the as-
sociated means mathematical understanding of or physical understanding that we can derive from
this process instead of going through all these details people started saying okay if this is the case
let us try to do it this way that I will take either that single coin and I will keep flipping it okay.

So I will do this I will keep repeating this experiment okay let us say n number of times and I af -
ter every time I do this experiment I tend to record what is outcome okay and I start counting it so
out of this n how many times let us say n1 times I have observed that head has occurred and defi-
nitely n - n1 times because we have already assumed there is no other outcome. So it will not
land on a edge of the coin and all those things so n - n times there will be tail okay and then we
people started questioning that what is the relative occurrence of these things okay.

So or let us say frequency of each of those events which will be happening so the frequency is if I
just take this experiment this simple experiment so I say from that experiment allowed outcome I
can say n1 by n that is the relative occurrence of I say head, so let us call that as P H okay and P
T is some n - n 1 / n okay and if I now start doing this experiment for a very large number of time
okay.

So what does that means mathematically that I will take limit n tends to infinity okay and people
started doing that and then they could observe that if the coin is means we should say unbiased
coin or the coin we see that it has no preference for any one of the surface head or tail whenever
it is landing see it might happen sometimes that the edge is designed in such a way that it has a
preference towards a particular service it might happen that it is little tilted over there so if that is
not the case people have seen that this n1 almost tends to become n /2 okay and finally what we
get so we get this Pa which is limit n tends to infinity n 1/ n that goes towards 1/2 and same thing
happens for sorry pH and something happens for PT that also goes to half.

So this was probably the initial definition of probability okay, so people started saying that proba-
bility of this particular event head happening is ½ that means if I just do this experiment for n
number of times I will almost see n /2 times head is occurring and n by 2times tail will be occur-

441
ring so this is the frequency definition of probability. So I associate with every event that are the
outcome of that experiment whatever random experiment I am doing. So those events I associate
with a value which is called probability.

So this is just percentage I should say means or I should not say percentage it is just a fraction
that defines how much means if I just multiply that with 100 probably that will give me how
much percentage of time it will favor a particular outcome okay. So that is actually the definition
very definition of probability this is called the frequency definition of probability and all these
number this half that is a magic number which has come from basically experimentation pure ex-
perimentation like any other scientific experimentation okay.

So these are the axiom we should say axioms of probability and these are just basically observed
over time and people have understood that probably this should be the case there is no basis that
this should be happening people have not solved the behind means underlying physics to try to
come up with this they have seen they have tried testing with different number of coins in differ-
ent location with different number of times with different kind of people who were tossing it.

So they have seen all these observation and everywhere they could see that it is close to this and
that has given that axiom that okay that should be the probability from there something else
comes into picture okay, so people started defining instead of this definition they started defining
that whenever I do a experimentation how many equally likely events are possible in a particular
experimentation like in this experiment I have head and tail which are equally likely okay. So
whenever I start calculating the overall probability so I count those equally likely or equal possi-
ble events okay I count those things and then create a space okay. So like here I have a space of
head and tail only two elements are there okay.

(Refer Slide Time: 10:40)

442
 1
pT = pH =
2

And we say the probability of head will be out of them for that particular event we are calling
means this will be more clearer with a bigger example probably right now we just defining it, so
for the event we are targeting how many equally likely elements are favorable to that event okay
here for head there is only one event which is favorable tail is not favorable towards head because
if it is tail then it is not head right.

So only one event are favorable towards this divided by as many events are there equally likely
events so two I get the same answer same thing happens for PT that should be 1 by 2 right there
are two events out of them favorable to head is one okay. So let us just give one little bit bigger
example and then probably you will understand a little bit more. So let us try to say suppose I
have a six phase die and I roll it and we say that the die is fair that means any surface appearing
at the top is equally likely okay.

So that means how many events equally likely events are there already by my definition I get six
equally likely events right.

(Refer Slide Time: 11:59)

443
So I can define them as a 1 a 2 a 3 a 4 a 5 and a 6 that means in the top surface one comes top
surface 2 comes top surface 3 comes and so on okay. So there are 6 such events and that define
my entire space whenever I define this collection of events I call this a sample space that is a def-
inition by definition I call this a sample space and these are the individual element of that sample
space, so these are the elements it is just the abstraction we are trying to do okay all of them we
have already told by definition are equally likely okay.

Now let us say that I define an event now we just diverting from this I could have defined an
event which is that I will be means what is the probability that I will be receiving one in the top
surface. So that is actually there are how many are favorable to this particular thing it is only one
this one only will give me one all others will be giving me something else. So one divided by
how many total events are there six so it should be 1 by 6 similarly if I talk about 2 that should be
also 1 by 6 and so on.

But if I just now change the definition or change the means the yes definition of event let us say
my event is that whenever I roll a die what is the probability that I will be getting a even number
okay that is the event definition now. So out of them now you can see the favorable elements, so
one can appear two can appear three can appear four can appear five can appear six can appear
out of them this will give me a even number, this will give me a even number, and this will give

444
me a even number so how many are favorable to my experimentation or my definition of event
that is three divided by total number of elements that six so that probability of getting even num-
ber is half similarly probability of getting odd number is also half right. So that is how we define
event over here okay, so let us say suppose the same thing we take it.

(Refer Slide Time: 14:30)

S ¿ {a1 , a2 , … ,a6 }
A0 ¿ {a1 , a3 , a5 }
Ae ¿ { a2 , a4 ,a6 }
B ¿ { a1 , a2 , a3 , a4 }
A0∩ B ¿ {a 1 , a0 }
Ae ∩ B ¿ { a 2 , a4 }
A0 ∪ Ae ¿Φ
A0 ∪ Ae ¿S

So let us say this is a 1 a 2 a 3 a 4 a 5 and a 6 that is my sample space I call it s okay, so the sam -
ple space now I will give us set theoretic definition of this particular thing so sample space actu-
ally is a means group of elements so I can call that as a set so this particular set is defined by

445
these elements a 1 a 2 a 6 six elements are there okay. Out of that I define an event which is
called a or I should say a o that means the event that has element which are actually showing me
odd number in the top of the surface.

So a o which is this so I can write a o as within the set it is just a subset which has the element a1
a3 and a5 right similarly I can write ae that is the even number which happens to be this a 2 a 4 a
6 right up to this it is all fine right now I start defining something else because now I will actually
go inside the probability theory okay. So let us try to define another event which is telling me that
whenever I roll a die my event is that I should observe a number which is less than equal to 4
okay so who should be the part of this particular event that should be these things so either 1 2
3or 4 those are all less than or equal to4 so I call that event B fine so I have defined now three
events within the same sample space okay.

I have 6 elements and with that I started defining events so I have a definition of events and ac-
cordingly I have defined some three events so this is a 1 a 2 a 3 a 4 now in terms of set theory let
us start calling the union of event and intersection of events because I have two events now let us
try to see if I can define in set theory we know about Union that means taking the elements of
means elements either that exist over here or there that is Union and intersection is it should be
and that means it should be both existent on both the set or both the I should say collection or
subset.

So if I just try to write ao inter section B what do I get or a o intersection B so this is my ao this is
my ae so this has element this and this a1 and a3 right similarly I can define ae intersection B and
this has element a2 and a4 okay. So far have still not gone into probability okay I am just trying
to see the events and their intersection or Union if I try to write the union of this sorry intersec -
tion of this what do I get that is a null set because these two has no common event so I can write
that as a null set or I can write it this way no events are included in that okay or no elements are
included in that. And of course ao unionae must be giving me s the overall set so these are some
of the things we have observed now let us try to see if I wish to from there from that definition.

(Refer Slide Time: 18:59)

446
 lim (n1 +n2 )
N →∞
p [ A0 ∩ B ] ¿ p [ A0 ∪ B ] ¿ p [ A 0 ∪ B ] ¿
N
¿ ¿ p [ A0 ∪ B ] ¿ p [ A 0 ] + p [ B ]− p [ A 0 ∩ B ]

If I wish to calculate the probability that ao intersection B is happening or probability that ao

union B is happening so this I should say joint event simultaneously ao and B happening so that
means I would say that I roll a dice what is the probability that the number I will be seeing that is
both even and less than 4 that is this event and the Associated probability and if I say that what is
the probability that both are means the event that are happening those are both odd as well as less
than 4 so that should be this and associated probability value that I wish to assign okay.

Probability means again that same definition that if I have all outcomes now it is becoming a little
bit more complicated earlier it was much more simpler now it is becoming more complicated be-
cause I have joint event defined so both the things should be satisfied what is the relative fre-
quency of that how many times if I start doing this experiment means that means n number of
times I roll the die and if n tends to infinity what should be the number of times that I will be see-
ing this particular joint event occurring okay so that is what we wish to evaluate okay.

447
So if I wish to evaluate that how do I evaluate this particular things okay, so first of all let us try
to see if I wish to do it for this one let us say probability of this okay probability of ao Union ae
okay this is something I wish to do now ao and ae has no common element right what I can do I
can basically individually calculate the probability axiom says that I can individually calculate
the events and their associated probability and I can add them it is very simple it can be easily ob-
servable because if I just do the experiment for n number of times whichever number of times
this will be happening this will not be happening okay.

So if I observe them as N 1 number of times and if I observe this one as n2 number of times they
should be completely if this is happening this will never be happening that is not the case I cannot
say that same thing for a o and B I cannot say these things okay. So whenever I will be calculat-
ing suppose ao Union B I won't be able to say this thing that a particular event that is favorable
towards ao that is not favorable towards B or it is probably that proper term is mutually exclusive
that means if ao happens B will not be happening I cannot say that there are elements where both
the things will be happening right and there are elements also where one will be happening the
other one will not be happening.

So this kind of complicated definitions are already observed over here but in this case whenever I
have this definition that means they are we call this as two sets are mutually exclusive they do not
have any common elements if that is the case then whenever this will be happening suppose I
keep counting this n1 is the count that you know has happened whenever a o has happened I am
sure that the other one has not happened and whenever the other one has happened I am sure that
this one has not happened.

So therefore if I just count this one and this one over all this union has happened can be just sum-
mation of them I am sure okay there will be no extra term counted if I just add these two wher-
ever in this case there are some common commonality that means there might be one experimen-
tation where I get suppose a o and B I am talking about so I get suppose it is odd number so I get
one that is one valid experiment so that particular one I count for which one should I count it for
this one or this one basically it will be counted for both of them but this is just one experiment I
am counting two times if I just add those counting it will just take extra count.

And obviously that means derived probability value will be wrong of this joint event so that is the
problem over here if they are mutually exclusive I can do this and then I can write limit n tends to
infinity that is the probability of this event ao Union ae and then I can separate them limit n tends
to infinity n 1 by n plus limit n tends to infinity n 2 by n this is exactly the probability of a o, so I
can write this as probability of ao plus this is probability of ae so this is actually where probabil-
ity gets distributed over the Union.

If the events are mutually exclusive okay so that is probably the first axiom of probability theory
that we can get if we can define two events which are mutually exclusive then the probability gets
added okay if they are not mutually exclusive then what happens so basically what we can see

448
just see it from set theory theoretic point of view so I have a set A I have a set B right I wish to do
A union B so what will happen if I just take as many times A happens plus as many times B hap-
pens then this particular section will be counted twice. So I have to take that out so therefore this
must be P of Ao + P of B minus P of Ao intersection B I can write this so thats the probably the
as long as they are not mutually not exclusive that's probably the second axiom of probability the-
ory that if two events are not mutually exclusive I have to do it always this way okay this particu-
lar case right. so now let us try to see that union if I wish to calculate if I know individual event
occurrence those are good but I also have that intersection so I have to now try to evaluate the in -
tersection of a occurrence

(Refer Slide Time: 26:10)

lim n2 lim n2
N→∞ N→∞ n1
p [ A0 ∩ B ] ¿ = .
N n1 N
¿ ¿ p( A 0 / B) p(B)

So let us say this I wish to calculate this right so this is something I want to calculate okay so how
do I do that? This actually means that both the things has happened right, so let us say things
event ao has also occurred and event B also has occurred let us try to again do the frequency defi-

449
nition let us say I have rolled the die for n number of times okay out of them I have seen that n1
number of times B is occurring okay.

So basically I am just trying to see that n times I have rolled the dice I am just counting where B
has occurred I am not bothered about a o now okay so I am trying to see where B has occurred
because I wish to evaluate where a B are both occurring so if a B are both occurring B has to oc-
cur so I am just counting those numbers where B has occurred that is N 1 okay.

Out of them how many times a has occurred now there is a restriction given B has occurred be-
cause this n 1 times I am filtering out those are the event means those are the outcomes or those
are the experiment experimental outcome where I have seen B event B has occurred that means I
have seen some suppose B is that less than 4 so that has occurred okay out of them I am now try-
ing to evaluate how many times a has occurred let us say that is n 2 so n2 is actually a number
where it is basically this is counted from that set of N 1 those events only I have observed out of
them wherever a also has been satisfied I am just counting out them.

So it should be I always know that N2 must be less than equal to N1 that should be the case be-
cause out of them only I am trying to see okay. So this is where a has occurred given B also has
occurred okay so that is the definition or that is how it is being written that event a or I should say
event a O has occurred given that B has also occurred okay.

So I can just now get this a intersection B what is the definition of that limit n tends to infinity fi -
nally a intersection B where both a and B has occurred that is actually n 2, so it should be n 2/
capital N this n 2 I can write so this I can write as n 2 / N 1 n 2 n1 divided by n that is simple just
both sides I have multiplied and divided by n1 now let us divide that limit n tends to infinity n 2 /
N 1 and limit n tends to infinity n 1 / n.

Now what is this and what is this? This is very clear this was the first experiment right where out
of that n number N 1 number of time B has occurred so that must be the probability of B right
what is this, this is actually probability that given B has occurred what is the frequency of a oc-
curring so this must be the probability that ao has occurred given B has occurred this particular
part because it is n2 by n1 so that n1 number is already telling that B has occurred out of them or
out of those experiment where B has already occurred I am trying to see a has occurred so this
must be the probability that a o has occurred given B has already occurred.

Because whenever I am trying to calculate this that means I have already ensured that B has oc-
curred okay, so that is where the conditional probability comes into picture this is called the con-
ditional probability this actually means that a particular event conditioned on that a particular
event has already occurred what is the associated probability that another particular event will be
occurring okay.

450
So we could now evaluate by this definition frequency definition that what is the intersection
probability what will do in the next class we will try to see the implication of these things okay,
so we will discuss these things in details in the next class, thank you.

451
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-32: Probability Theory (Contd.)

Okay so in the previous class we have already defined means we have given one example of
rolling a die we have said that there are two events which is called the event where the means top
surface of the die will be, will be showing me a even number and that is called or odd number
okay so that AO is called that it will be showing odd number and then another event we defined
that is B that the number it will be showing that is <= 4 okay. so we started evaluating the joint
probability that P(A0) or AO and B both are… okay.

(Refer Slide Time: 01:03)

452
 lim n2 lim n2
N→∞ N→∞ n1
p [ A0 ∩ B ] ¿ = .
N n1 N = p( A 0 )¿
¿ ¿ p( A 0 / B) p(B)
¿ ¿
And then from the frequency definition we could see that it should be this where we have defined
a conditional probability that a particular event has already occurred means given that what is the
probability of the other event happening and P B okay multiplied by PB is something we could
see this is another axiom of probability theory p[A o intersection B] so that is that is pretty much
what mathematical theory does the initial part you take it from experimentation okay.

You build up the first few forms of axioms and from there in a deductive logic which we are do -
ing now okay so you have given that definition frequency definition to probability and after that
it is just deductive logic with which we are actually going in forward okay so this is one axiom
which is being derived from that frequency definition of probability theory okay so this is some-
thing we have already understood.

Now let us say that this ao the event is independent of B what do I mean by that that is a very im-
portant definition in probability theory that two events are independent okay so what do I mean
by that, that means t-hat occurrence of ao does not really depend on whether B has occurred or

453
not okay so whenever we are trying to evaluate this n2 how many number of times Ao will occur
given B has occurred even if I do not consider B I will probably get the same number.

If this is happening that means I have no dependency for the probability calculation of this partic-
ular event on the previous event I will give some example so basically in that case what I can say
is p AO given B is independent of B whether B has occurred or not I will get the same number so
I can write this as P [Ao] and immediately this formula turns out to be that is a very important
theorem of probability if two of the events for which you are calculating the joint probability or
the intersection.

If both of them are independent their probability just gets multiplied probability of individual
event gets multiplied we will just give you some example and it will be very clear.

(Refer Slide Time: 03:47)

2
p B=
3
1
pW =
3

454
So let us take this let us first give one example one very simple example or this dependency so let
us say I have a particular arm with two red ball and two black balls so I have this thing where I
have two black ball and two let us say two white wall now if I cannot see what I am picking up
and randomly I pick a thing there is a probability that will be either red or it will be black okay if
me picking a ball is any of them is equally likely then what should be the probability that I will
be picking up means let say white ball that should be half because there are two favorable and
there are four overall things okay.

What is the probability that I will be so this is probability of picking white ball probability of
picking black ball should be also half okay so now I say I define two events one is first time I will
take out a ball okay and I see what it is and then I put the ball back inside the? and next time
again I takeout about okay so the first time I am taking out that is a random event and second
time I am taking out that is also a random event let us call A1 and a2 the outcome of a2 does it
depend on the outcome of a1. Absolutely no because I have replaced the ball so the experimental
scenario goes back to the previous case okay so whether in a1 I have taken white or red I have re-
placed the same ball the overall means sample space remains the same so whatever again I will
be taking out if things remains equally probable I will be getting same probability so this a1 and
a2 these two events are independent of each other but suppose I do not do that in the first experi-
ment I take out something and I do not replace it okay.

so immediately what will happen the probability will be now dependent on the outcome of the
first event because if I in the first event I have picked a white ball then this balance will be little
bit changed for the second experiment there will be two black and one white so automatically p
black will become1 sorry 2/3 and p white will become 1/3 and if I, in the first experiment I pick a
black ball then there verse will be happening so now I can see my probability calculation or the
overall event is depended on the first event okay so this is where I say two events are dependent
or independent

(Refer Slide Time: 07:15)

455
 1
p [ A0 ∩ B ] ¿
3

So let us just for fun we go back to the same example we have taken sorry now lets define this as
my B this, this as Ao this as Ae lets try to see if I wish to calculate P A o intersection B I first as-
sume that they are probably independent I still do not know whether Ao and B are independent or
not we will try to derive this okay or try to see that and you will see a very interesting case so
suppose I assume that so immediately what will happen because they are independent the proba-
bility theory according to our derivation it says it should be this .

So what is probability of ao that is actually three events / 6 events so that should be 3/6 or 1 / 2 x
What is probability of B that is actually 4/6 so that is 2/3 so it is 1/3 right if I take them to be in-
dependent I might be wrong I will try to verify that now I will do the right calculation if I do not
have knowledge of independency I will do a right calculation so which says probability of Ao
given it is B x Probability of B this was the actual definition if they are not independent.

So I still do not know whether they are independent or not so this should be the right calculation
this is more generic so let us now try to see what is the probability that if given B has happened
what is the probability that Ao will be happening so given B has happened how many cases are

456
there I will be restricted already okay so given B has happened I have only four events out of
them what is the probability that Ao that is my favorable case will be happening.

That is actually two it is 2 x 2/4 x Probability of B how much is that that’s again same thing prob-
ability of B remains the same 2 / 3 that is 1/ 2 so I get the same result which says yes these two
events are independent I just change this definition I change the definition of B I say my B is this
you will see a fantastic thing happening if my B is this that means B that event B is actually less
than 5 <= 5 happening okay.

So let us try to do that same calculation if they are independent what should be this PAo that is
actually 3 / 6 so that is again 1/2 and P B that is actually 5/6 right so if they are independent this
is the case right now let us see if they are not independent so what do I get I do get over here B
has already occurred okay out of them how many are A o that should be 3 /5 x PB which is 2/3
see the probabilities are not same so the event whether it will be independent or not it all depends
on whether this is equal to this probability A o given B is equal to probability Ao that I can evalu-
ate.

So if that is not happening that means the events are not independent see whenever I was taking 4
because what was happening actually because it was taking same number of odd and even so
eventually or half of them were even so eventually my independency was coming out whenever
that has not happened because I have taken now 5 number so there is a bias towards odd event
once that has happened immediately my independence goes away.

So why I have given this example a slight definition of the event or slight variation in the defini -
tion of the event can change your dependency so you have to be very careful whenever you are
writing or joint event to be the multiplication of two event so whenever you do that that will be
very careful you have to ensure that is the case otherwise you will be doing a mistake this is more
generic definition with a conditional probability so always you should take that into account.

(Refer Slide Time: 12:53)

457
p( A ∩ B) ¿ p( A / B) p( B)= P (B/ A) P( A) ¿ p( A ∩ A ) ¿= p( A )+ p( A )¿ p( A ∩ B)¿=p (B/ A) p(B)¿ p( A ∪ B
o e e o
¿ ¿ P(B)

So now let us see we have already defined P let us just go back from that example let us just take
two things two events a and B so a intersection B we have said it is P A given B x P B okay these
two event I have no preference so I can even write this as P B given A x P A there is no problem
in that because I can just replace A / B and B /A I get the same formula right so these two are
equal and immediately I can see that P a given B can be calculated because these two are equal I
can always calculate B given A PA /PB that is the famous Baye's theorem.

So what we are doing we are just building up the probability theory entire probability theory just
from that initial fre quency postulate with that only we defined all these things and now we are
getting one after another theorem so these are all linked and they have some importance also you
later on see why this is very important okay so we will talk about that later on when we give
more generalized thing so now I would like to touch on two things okay we have now talked
about two particular kind of event one is independent event one is one is mutually exclusive
events are they same thing or are they not okay.

Basically it is not completely different do not confuse independent event with mutual indepen-
dence or sorry mutual exclusiveness in mutual exclusiveness what we are saying just take that ex-

458
ample of Ao and Ae odd and even in that case if Ao happens that means the odd number I get im-
mediately I can say that event will not be happening so basically I am almost saying there is a
strong dependent instead of becoming independent mutually exclusive events are strongly depen-
dent if I say this event has occurred that actually means the other event has not occurred.

So I have some already some understanding of the other event by defining this event okay so they
must be strongly dependent so mutually exclusive events are never independent mutually exclu-
siveness has nothing to do with independence people often get confused with these two definition
they start thinking that mutually exclusiveness is probably also defining independence no it is not
that is the mistake people do often when they are just starting to learn probability theory.

So that is not the case in mutually exclusiveness remember the probability axiom that if two
events are mutually exclusive then I can write this and if two events are mutually independent
suppose A and B then I can write this so always remember this and if two events are not mutually
exclusive then I can write suppose a union B I wish to write so that must be set theory we know
PA + PB we have already talked about that it will actually count two times so I have to take this
out.

If they are independent but not mutually exclusive then I get this theorem this happens so the two
events A and B are independent but they are not mutually exclusive then this should be happen-
ing if they are mutually exclusive this should be 0 okay so whether they are independent they are
not independent means actually mutually exclusive means they are not independent so and this
immediately happens to be 0 okay. So that is something which we should be keeping in mind
whenever we are defining probability now let us do something more okay.

(Refer Slide Time: 17:35)

459
 ∪ Ai ¿S
i
Ai ∩ A j ¿ Φ∀i≠ j
B ¿ B ∩S
¿ ¿( B∩ A 1)∪ (B ∩ A 2 )…∪(B ∩ A n )
n n
p(B) ¿ ∑ p ( B ∩ A i ) = ∑ p ( B/ A i ) p( A i)
i=1 i=1

Just take the Bayes theorem towards more generic one so lets say I have a event B okay so this is
the sample space of B lets say in between all those events are actually or all those elements are
included and this sample space I now actually include it within a particular sample space called A
and I divide that a into some mutually exclusive partition okay I call that A1 A2 A3 A4 and so on
so what I have done is I have taken this B I have taken a superset of B which is A and then I have
actually partitioned a in mutually exclusive parts okay.

And I am calling them a I and what I am saying that this a actually happens to cover the entire
sample space that I am experimenting with so that means all the events that can happen that is in-
cluded in the entire A okay so a include every everything in that sample space B might not be-
cause B is a subset of a so B might not include that but A include everything so I can always
write union AI because they are mutually exclusive so they will be just added one after another or

460
their elements will be just added over I if I add them so that is that must give me the overall some
sample space that is according to definition okay.

And I also have this definition AI intersection AJ is null set for all I not equal to J so this is the
definition of my a that is how I have created my A know what I can write about B see B , because
B is a means proper subset of A or that sample space S I can always write B equal to this because
it will have C when I take if it is a proper subset the elements of that will be actually elements of
that and intersection of the bigger set because it will have the same elements right so I can write
this now yes I can substitute in this one so it should be B intersection [Union of AI] okay.

So now because they are mutually exclusive I can actually just flip these two so I can write this as
which is very obvious from the Venn diagram also the diagram I have drawn so B I can write as
this Plus this where this is actually B intersection AI or A1 this is actually B intersection A 2 if I
just add all of them I must be getting B back so this is what is happening so U (B n A) suppose it
goes up to N and so I have made n partition here it is n=7 right.

So I can write this now because AI are all mutually exclusive so the B n AI are mutually exclu-
sive that is very obvious from here also they are all mutually exclusive because AI are mutually
exclusive so if I take B intersection AI that must be a subset of that now B intersection A1 and B
n A2 if A1 A2 are already mutually exclusive they have no common element therefore that small
subset must not have any common element so I can always write these are all mutually exclusive.

So now probability of B what can I write that should be probability of this whole thing because B
is this now because they are mutually exclusive now I put that probability theorem of mutually
exclusiveness so that should be sum of all those probabilities so I can just write summation over I
going from 1 to N it is P B n AI, I can write this no problem in that now I can put Bayes theorem
the way I have learned it so I =1 to n this intersection is I can write it in terms of conditional
probability so I can write P B given AI x P AI that is something which I can write now okay. PB
that event we were talking about which is a subset of that whole thing must be this.

(Refer Slide Time: 22:59)

461
 p ( B/ A i ) p ( A j )
p( A j /B) ¿
p(B)
¿ ¿
So eventually what happens I have earlier proven that P some AI given B or lets say lets just put J
because I will be putting a dummy variable so this is something I have already proven that that
should be P B given AJ PA J / P B now just now I have proven PB should be a summation term
so I can just write this as given AJ PAJ /summation I =1 to n P B given a I that is the most
generic form of Bayes theorem okay you might be asking why we are doing this why Bayes theo-
rem is so important okay.

So what we should do is we should give some example of these things okay that whatever we are
doing that is that is very important but before giving that example probably we will have to now
go into another definition once we get means our hands very clear on that definition probably it
will be easier to give that example will give well we will pick on every important communication
example to say why this is very important okay why we are actually learning Bayes theorem as a
communication expert why this is very important for us okay.

462
But before that let us try to do something let us try to give another definition which is very impor-
tant to probability theory which is called random variable so what do I mean by random variable
or RV.

(Refer Slide Time: 24:50)

x 1 , x2 ∈ R
1
p( H)= p(T )= .
2

See so far we have been talking about set elements events are like if I roll a die it is just a descrip-
tion right it is the top surface getting one or top surface getting two and all those things mathe-
matically those are not that description is not very sound right I cannot represent them mathemat-
ically I cannot write anything over there okay in terms of mathematical equation and all those
things so it is very important that we define those things into a sound mathematical background
okay.

And that is where random variable comes into picture so basically random variable is nothing but
mapping those abstract things of definition of events into the real axis so suppose I have head and
tail these were two outcomes of tossing a coin okay so this was a abstract thing it is just a word

463
that head or tail it has no mathematical notion it does not go into some number and all those
things so I cannot really manipulate them.

So now what I do I do a mapping I say that head gets mapped to -1 in the real axis and tail gets
mapped to + 1 or vice versa whichever way you wish to put it okay so if I just do that then I get a
definition of random variable so because this head and tail we are actually random outcome the
event is defined but I do not know if I do the experiment whether I will be getting head or tail that
is why I call it random because before hand before doing it I have no notion of or no understand -
ing what will be the outcome of that experiment.

So I cannot say that but what I can actually define is probability of them that means what is the
chance of getting head I can I can define that that is what we have defined so far so probability of
that event I have already defined so I have said probability of head is actually 1/2 and I have said
probability of tail is actually 1/2 this is something I have told now what I will do instead of say-
ing probability of a event which is abstractly defined I will say probability of this +1 and -1 and I
will put some number.

Immediately what do I get I get a functional representation of this probability so what I do this
head I put it to some discrete value which I have done X sorry this -1 to the real axis so I call it
X1 so X1 =-1 and tell I put it into the real axis I call this X 2 okay which is a real number so X1
and X2 are taken from the real number okay set of real number so X 2 + 1 it can be anything ei-
ther you give -1 + 1 I can also give 0 and 2, 0 and 5 whatever you wish you can give that but as
long as your definition is clear you know that it will be mapped as a function okay.

So at minus 1 this will have a value of 1/2 and at +1 it will have a value of 1/2 and the Associated
function is defined as PMF or probability mass function it is actually defining that if I map
uniquely this is this must be one-to-one mapping remember every event should be uniquely
mapped to the real axis there should not be any ambiguity if head and tail I map to the same point
I will just get ½ + ½ one and I will not have any distinct means distinguishing feature between
head and tail whereas actually they are two different events ok ay.

So I need to have a one-to-one mapping to the real axis and then I plot the probability with re-
spect to them whatever I get that I call as probability mass function that means in the real axis as
if that is representing the random variable that I am describing by doing this mapping okay and
here because I have countable number of events so that is why the overall mapping will be dis-
crete mapping remember if a random experiment gives me uncountable number of events then I
can actually map it to a continuous variable.

Whereas if I have countable number of events I will only be mapping them to a finite number of
or if not finite at least countable events and there will be all discrete so I get discrete probability
values and the Associated function I call it probability mass function okay we will also see the
other definition of it whenever I will have continuous one will have probability density function

464
so this is called probability mass function okay fine so this definition is there is no problem in
that okay so how do I define this.

(Refer Slide Time: 30:14)

X ={ x 1 , x2 } ¿ p X ( x1 ) =1/2 ¿ pX ( x2 ) =1/2 ¿ X ¿ { x1 , x2 ,… x6 }
p X ( x1 ) ¿ ⋮ ¿ ¿

So I say The Associated random variable is X and this X might take value now we are defining
earlier we are saying just head or tail now we can say this takes value X1 or X2 right if it is head
tail if it is a die I can write it is X1 X2 up to X6 right if it is any other experiment suppose its card
so there will be 52 such elements right so it depends on what kind of experiment you are doing
accordingly how many events are there or elements are there will be all mapped to a real axis and
those becomes the random elements.

So I define a random variable X and those have elements X 1 X 2 X 3 and then the PMF I define
it this way P probability of this random variable taking a value X1 okay or you can put first
bracket as well this actually says that your random variable which is this X defined by this X this

465
is a random variable it takes a value x1 what is the Associated probability so if I am doing this
experiment this must be half if x1 is defined as head and px X 2 that becomes half that is all.

If this is the random experiment then I call P X x1 that should be 1 / 6 and so on all of them will
be 1 / 6 okay so that is how I define it and that is how I get the probability mass function simi-
larly I can start defining the joint event okay.

(Refer Slide Time: 32:14)

So the joint event will be it is like this so let us say I put a value X I and Y I so this is the defini -
tion of joint event that means basically now I am trying to see a probability that two random
events or two random variables taking corresponding the value of X 1 or X I from the random
variable X and y1 from the random variable Y this is the Associated joint probability of x and y
simultaneously happening which are taking simultaneously value X I and Y I.

So this is the definition of the joint probability what we will try to do next is probably we will try
to give the axioms of these things now in terms of random variables so far we are defined in
terms of set theory and events only but we are not mapped it properly to the random variable now
we will actually properly map it to the random variable which we have just started we will do all
those axioms we will actually map it to proper random variable and then try to see what is the
outcome of them ok and then from there we will go towards the definition of continuous random
variable okay thank you.

466
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-33: Probability Theory (Contd.)

Okay so I think we have in the last class we have discussed already about basic theory of proba-
bility right so that is something we started discussing we have already proven Bayes theorem we
have given the definition of random variable that is just like a mapping from events to value in
radial axis so that is something we have started and then we define something where the random
events are discrete that means it is countable, so whatever events that will be happening that is
countable right you toss a coin.

So you have only two outcomes that you can you can map to a real axis maybe put to 0 or 1 or
means head and head or minus 1 or 1 it is just countable so for that we started defining corre-
sponding probabilities of each events okay.

(Refer Slide Time: 01:22)

467
 p X ( x1 ) ¿Prob( X=x 1 )
1
p( X=−1) ¿
2
1 ¿
p( X=1) ¿
2
n
∑ p ( x=x i ) =1 ¿
i=1

So we have given a definition that is called where X is the random variable it is actually saying
the probability that, that random variable X is equal to a particular event X I or the real axis rep-
resentation of that event X I what is the probability associated probability of that and we have
also told because if we just consider all the events and they are probability if we just sum them
that is a certain event because we must be seeing something is happening okay.

So like we say probability of head that is let us say x = - 1 that is probably tail probability of tail
is 1/2 and then we say probability of head that is +1 that is equal to1/2 so if we just add these two
that means something will be happening so that is a certain event and that addition of probability
must be 1 we have also given a fundamental theorem of probability theory that is called this sum-
mation when it is taken over all the events that can be occurring in that particular experimentation
that must be always one.

So probability actually sums up to one the way we have defined probability of a random variable
okay or associated probability for a particular random variable for a particular value when it takes

468
a particular value. So this is always true that is something we have already discussed okay we
have also discussed something called means same representation for joint event where you have
two random variable for two events okay given one example of such thing.

So if we have joint event so we can represent it as this x and y okay these two are joint event that
means X represent the random variable associated with the experimentation where we are consid-
ering that event X or collection of event X and then collection of event Y if you have that so to
separate this one and then we say probability of X Y XI Y I that actually means that probability
that random variable X takes value X I from its sample space and simultaneously random vari-
able Y takes value Y I from its sample space okay.

So this is something which can be defined this might not be just same I this might be also Y J be-
cause that might have a different variability altogether okay so this is called that joint event and
associated joint probability okay and similarly if we just sum this over all the possible combina-
tion of X and Y's we must get one, so if you just sum it will be a joint summation because I have
to do combination for all x and y so this is for all I going from 1 to N and all J going from what-
ever value it goes so let us say this is n X and this is n Y if we just take all of this that must be
one okay.

So this is for joint event and accordingly we can keep on augmenting that we can have as many
events jointly happening as we wish and accordingly the joint distribution will be formed okay so
after defining this with this same definition of notational aspect we started defining the condi-
tional event right.

(Refer Slide Time: 05:15)

469
p X Y ( x i , y j ) ¿ p X ∣ Y ( x i ∣ y j ) pY ( y j )
=∑ p X ∣ Y ( x i ∣ y j ) pY ( y j ) ¿= pY ( y j ) ∑ pX ∣Y ( xi ∣ y j ) ¿=pY ( y j ) ¿
¿ ¿ i i

So we have also told that the conditional event which is represented as this, so this X given Y it is
actually a conditional thing that means there are two random experimentation simultaneously go-
ing on which have some random outcomes one is represented by a random variable X which is al-
ready mapped to a rail axis and the other one is represented by random experimentation which is
a random variable Y which is mapped to again real axis okay.

So X given Y means that I have already observed the event Y which is taking value equal to YJ
given that what is the probability that I will be observing that X is happening to be X I that asso-
ciated probability this is that conditional probability we talk about okay. So whenever we say this
is conditional probability so if you just sum this over I okay so what do you expect this is a con-
ditional probability that means y YJ that is already given okay given that all the possible proba-
bility values for seeing the event x equals to x i okay.

So if I take all of them that must be again giving me probability of one because given that y J has
happened if I am just trying to test what's the Associated probability of each of the possible out-
comes of X I and their associated probability if I add all those that must be a certain things be-

470
cause something of those X I will be happening anyway, so that probability becomes one very
simple example that we have taken probably we had two joint observation you remember that we
were just trying to roll a die and then we were saying whether it is even or odd that is one event
and then there was another event that it is less than 4 or greater than 4 right.

So if suppose I say it is already even that is known that is J for me that it can take two outcomes
even or odd so I am just saying that given that it is even okay that means only two four and six
are possible outcomes so only those cases I will be taking, so I will be rolling the die it is a ran -
dom experimentation of course but I will be only marking those events where either 2, 4 or 6 is
happening all other events are nonentity for me because they are not part of this experimentation
okay.

So I will just take those things and start counting that frequency definition of probability theory
and then what I will say I will try to see a particular XI is happening whether it is less than 4 or
greater than 4 okay, these two things are there so here if you see less than equal to 4 are 2 so there
are 2 favorable outcomes towards this and 1 favorable outcomes which is not towards this so this
X I, I can say or this X I can say it is less than this or greater than this okay. So if this is my defi -
nition of two events and accordingly X is defined so X I will take two values only let us say 0 and
1 again I am mapping.

So 0 means actually it is less than 4 okay less than or equal to 4, 1 means it is actually greater
than 4 and then if I just evaluate the calculation or probability calculation what is this suppose let
us say 0 given it is even okay or let us say even I represent as one so even and odd these are two
events which are mapped to the real axis I say this is 1 this is 0 okay as I have told both the
events x and y or both the experimentation x and y and associated random variable must be going
to the real axis so this has gone to the real axis I have just denoted 1 & 0 similarly less than equal
to 4 and greater than 4 these are the two outcomes this must be mapped to a real axis so let us say
I put it as 0 and let us say I put it as 1 okay.

So whenever I write this that actually means that 1 has occurred in Y so that means this has oc-
curred so it is already even-numbered given that it is even-numbered what's the probability that
this has occurred okay that means it is less than 4 this probability I wish to calculate I have al-
ready seen that there are 1,3,5 also out of them these are only my favorable condition towards this
Y event that y equal to 1 out of them what's the probability that this X will be 0 so there are 2 fa-
vorable outcome towards that and 1 non favorable they are equally likely because the faces are
equally likely to come.

So I can very fairly say that should be 2 by 3 that probability because 2 favorable and overall
means possibility are 3 similarly P X Y if I wish to calculate 1 1 that must be 1 / 3 because 1
means this greater than 4 there is only one favorable case out of 3 so that must be 1 by 3 if you
sum them that should give you one okay. So with that simple experiment we can actually again
map it towards a conditional probability calculation and we can get back to this theory that is

471
very important that conditional event if the condition upon which it depends on that is already
fixed then if you just take a summation over the one which is dependent on if you just take a
summation of all the probabilities on that random variable you must be getting one.

Similarly it should be also true if I put it like this P Y given X suppose it is the same experiment I
just now turn the condition okay so given that it is less than equal to four what is the probability
that this is even or odd something like that okay so then YJ must come over here and X I so now
it is given X I I am summing over all J this must be also giving me one that should be always
happening you can again test it with that same experiment we have done okay. So you will see
that this is happening for all possible cases okay so these two are very important relationship and
then we have also understood the bayes theorem okay.

So what is base theorem Bayes theorem we have called this is the joint event we have already
talked about X I and let us say YJ according to base theorem that must be the conditional one x
which is the independent thing means on which the whole experiment depends so P X given Y
suppose X is conditioned means over Y so that should be X i given y J Py YJ be careful about the
notation whenever it is conditional it should be like this that random variable x given random
variable y whenever it is not conditional and it is a single random variable so it should be like this
whenever it is joined it should be like this okay so that is the notational means representation
okay.

So this must be happening and it should be also true for Y given X so bayes theorem tells me that
this is true right so now let us try to evaluate something like this P XY that is the joint one I am
taking and I sum it over one of them ok let us say I sum it over I so I can write over I I will be
summing P let us give take this definition this X given Y okay, so X given Y so I am just writing
the expanding the joint distribution okay or joint probability with the condition 1 so X i given y j
py YJ I can write it this way okay.

Now this particular part that does not depend on I so I can take that out so I can really take this py
y J out summation i p x given Y X i given YJ now this already we have proven that this should be
one just now proven that okay, so if this is 1 this must be py YJ that is a fantastic thing what is
happening if you take the joint distribution and you we call this as marginalization if you just sum
it over one of the random variable okay all possible cases of that will be getting the marginalized
distribution recall that.

So this is actually from the joint distribution we can get the individual distribution similarly we
can do it for X also so we just have to sum it over J, I will be getting px X I you can you can try
this just have to use the other one probably okay. So this is how you can always any joint distri-
bution given you can get the individual distribution that is possible if we call this distribution
okay basically this is called actually our probability mass distribution probably we have not men-
tioned that so because this is discrete event.

472
(Refer Slide Time: 15:36)

0 ≤ p X ( xi )≤1

So whenever we get for a particular random variable P X X I we know all the values of this for
every possible values of X I we call this as a PMF or probability mass function okay so and if
you just plot it should be like this all those X I they are now we have mapped it to a real axis so
there will be some value some discrete value over this particular thing and there will be a Associ-
ated probability value which will be plotted over here and we know that the sum of all this proba-
bility must be one that has to be and they must be always positive.

Because it is probability value probability cannot be negative because you will be see the fre-
quency definition of probability says that I will be doing this experiment for let us say infinite
number of time out of them how many will be in favor of my X equal to X I okay that can never
be negative neither that N can be negative nor that how many means with what frequency I will
be getting a favorable outcome that can at least be 0 it cannot be less than that I cannot have a
negative event happening the negative count of event happening.

So I will be always getting a positive number so definitely the probability value associated proba-
bility value whether it is small or big it should be some positive thing and sum of positive should

473
be giving me 1 so therefore I can very fairly say from this that P of X XI must be lying between 0
and 1 because if their sum is 1 individually they must be fraction okay so this is always true for
any probability mass function or any probability value okay so these are something which are al-
ready known to you okay.

So what we have now learned is we know how to define in a for a discrete random variable how
to define Joint Distribution and from there how to marginalize to get individual distribution of
each of the random variable that is something we know and The Associated conditional distribu-
tion how it is related to joint distribution this is something we have already derived you might be
now asking all this Bayes theorem conditional rule what that has to do with communication.

So let me just give you a simple example which might not be having anything to do with analog
communication in particular it is mostly the application is in digital communication but you will
see later on when we will be proving something on analog communication there are some re-
quirement of that also but I will just give this example because visually this is very satisfying
okay.

(Refer Slide Time: 18:22)

474
 p ( x i=0 ) ¿Q= pX (0)
p ( x i=1 ) ¿1−Q= p X (1)
pY ∣ X (0 ∣ 0) ¿ pY ∣ X (1∣ 1)=1− pe
pY ∣ X (1 ∣ 0) ¿ pY ∣ X (0 ∣ 1)= pe

So let us try to take this example that this is my transmitter and this is my receiver so from the
transmitter side why I am giving this digital example because you will see that there are only very
possible number of symbols that can be happening, so definitely we will have to calculate proba-
bility values for a less number of events okay. So that why that is why it is little bit easier to
tackle, so let us say from the transmitter side I can only transmit two possible kind of signal okay
one is high voltage or let us call it one and one is a low voltage or almost zero voltage which let
us call that zero okay.

So it is almost like a binary bit stream which you are trying to transmit and that probably carries
information as long as it is properly encoded okay, so it is streams of ones and zeros I am trying
to transmit with some symbol okay so let us say at the transmitter there is a because we are trans -
mitting events so there is a random occurrence of these events with let us call that a variable X X
can take 1 or X can take 0 so these are the two possible outcome and associated there are some
probability okay.

So let us say with Q probability at the transmitter side it is 1 and with 1 minus Q probability it is
0 because only 2 are there and that probability sum of the probability must be 1 we have already
seen that so this is the only possible scenario okay because we only have two events or two sam-
ple values okay. So this probability that XI equals to one okay, so this is Q and probability that X
this X I = 0, 1 -Q okay these are called the prior probability what does that means that while
transmitting this is the probability with which I generate either 0 or 1.

So that is why these are called means without seeing what is happening at the receiver this is
what is actually happening at the transmitter, so this is called the prior probability that means be-
fore transmission this is this is the statistical balance between the events okay which we are try-
ing to transmit and decode on the other side.

Now we have already talked about channel so will be this 1 or 0 whatever it is it is a voltage level
or some specific type of symbol whichever way we represent it when we put it in the channel
okay a channel will do its own corruption it will probably add noise it will distort things and all
those things will be happening and at the end that the receiver will employ something to detect it
okay.

But what will happen because the channel has already corrupted it I might have a wrong detec-
tion at the receiver side okay, so there might be erroneous detection at the receiver side so what
might happen? Suppose I transmit and let us say this I still say this is y okay random event Y

475
which is the reception okay now because this is already random in between whatever things are
happening that is also random so what I will be receiving that must be a random things okay.

So that might be one and that might be also zero so I can only I have only two possibility I al-
ready know that whatever he is transmitting that is limited to one or zero so whatever I will be re-
ceiving somehow I will be encoding this and I will be detecting only one and zero okay, so this is
two things I will be detecting now what might happen because of the channel there might be er-
ror.

So that I transmit one and he also receives one this has certain probability so let us call this one -
PE where PE is the error probability so immediately I can see that I transmit one and zero is re-
ceived that must have a probability of P because again these two are the only possible that once I
transmit one there are only two possibility that either I will be receiving one or I will be receiving
zero.

So there are two possibility and we have already told that sum of probabilities must be one so if
error probability that one gets flipped to 0 if this is PE somehow I will have to calculate that you
will see in your digital communication how to calculate that so if this is P that must be 1- P let us
say that same thing happens for the zero also. So the probability that I will be correctly receiving
it that zero I transmit and I receive 0 that is 1 - PE and I will be it is wrongly receiving it is PE
okay.

So that is what happens in the whole scenario now of course we will be asking how do I evaluate
PE that is a separate domain okay we will have to see how do we evaluate those things but that
will not be part of our discussion in analog communication that should be the part of digital com-
munication discussion okay. But whatever that is this is the scenario okay. So now let us try to
define whatever we have got so far so if you see over here there are two random events okay, one
is this X which is one of the random variable and another one is the Y what is transmitted and
what is received okay both of them I have now mapped into real axes it can take values 0 and one
other one also can take zero and one okay.

So I know the prior probability which is px 1 I already know okay px that that random variable
will be taking a value 1 that I already know that I have told Q and P X0 that is 1- Q right now
over here whatever I know those are the conditional probability if you see very carefully so I am
just mapping whatever we have learned to a physical example okay. So this is the conditional
probability that given 1 has been transmitted what is the probability that I will be getting 1 or 0
okay.

So I can immediately write that P Y given X 0 0 that means that X that is the given one that if
that was 0, 0 was transmitted what is the probability that I will be getting 0 that must be accord-
ing to our definition 1 - P same thing will be happening according to our definition it is symmet -

476
ric actually 1 and 0 are both behaving in a symmetric manner due to the channel. So P that Y
given X and I will be transmitting 1.

And I will be receiving also 1 these are all 1 - P and just the other thing that P Y given X I will be
transmitting 0 and I will be getting 1 or p y given X I will be transmitting 1 and I will be getting 0
this is P so what I can see that I already know all the possible conditional probabilities right. So
these two things I know I know the prior probability now I will apply Bayes theorem so I know Y
given X I want to now know what is the probability that X given Y okay.

So this is something I wish to know and also I wish to know what is the probability that I will be
receiving 1 or I will be receiving 0 that is actually the receiver probability okay so that is the mar-
gin probability coming from the Joint Distribution so if I can define now P XY and this X I Y I or
Y J if I can define this I have already talked about marginalizing. So I can marginalize it over X
and then I can get p y okay.

So there is a possibility that I can now get like prior probability this probability of earlier I would
I was defining probability of transmitting something now due to the channel there will be some-
thing else which will be happening now I wish to know what is the probability that I will be re-
ceiving one and what is the probability that I will be receiving zero so that is something now I
have enough tool to calculate that okay and you will be asking why I need the reverse thing that
means P now I will be evaluating X given Y with some values, why do I need this?

So in detection what mostly you are interested that at the receiver I detect something okay after
doing all those things I just see okay this is 1 or 0 finally I will be taking that now whether it is if
I detect that as 1 or 0 I need to understand that okay I have received one but I have a doubt be -
cause the channel was creating errors I have a doubt whether one was transmitted or not now I
wish to know if I have received one was it one and with what probability was it one actually
transmitted and with what probability that should be given by this px given Y.

So now given thing is known to the receiver because receiver will be receiving that receiver does
not know what was transmitted he is trying to guess what could have been transmitted, so from
this he will be able to backtrack or back calculate this particular information okay and then he
will be able to guess probably what kind of error he is expecting right.

So after detecting was the overall error that he is making because once he has this probability
value he will be knowing that how many of those things means if he knows the probability of er-
ror he will be also knowing if I receive this many out of the definition of frequency definition of
probability we can always get to know how many of them will be in error okay.

So this is something he will be knowing and then he can actually try to manipulate that but if he
does not know this relationship he will not be able to manipulate that manipulation means proba-
bly he has to design a better detector he has to probably use a different channel where the error is

477
less or error is more or some other things he has to put over there okay. So all those things you
can do only if he has or the receiver has this understanding okay.

So that is why it is very important that we have this mapping all the conditional probability, mar-
ginal probability and the overall distribution that is well known and if one is known how to calcu-
late the other one that is pretty much required in any communication. So what we will do now we
have so far done even the example done for discrete case we will just now try to see what hap -
pens for a continuous case which will be more interesting for us in analog communication spe-
cially, thank you.

478
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-34: Probability Theory (Contd.)

Okay so in this particular example we have already seen that how do we calculate means if I have
the transmitted that prior probability known and then channel what it is doing that is also known
to me how do I calculate what is happening at the other end and then whatever is happening at the
other end from that we can also start guessing what might have happened in the transmitter side.

So you remember in communication it is, it is the communication actually becomes successful if

transmitter and receiver does not have any other form of communication because the very com-
munication you are building up just to give information from transmitter to receiver if you al-
ready assume okay for that you need another communication then it just goes on right.

So basically you have to always understand that whenever we are doing communication probably
transmitter does not know what will be happening at the receiver and receiver will also have no
information about what is happening in the transmitter okay so they will not have this linked in-
formation only they will see something which is being transmitted over the channel they will re-
ceive that if it is contaminated by the channel. They have no way to know what exactly has been
done by the channel okay

so they can only have why we are doing all these things because they can only have a statistical
guess nothing else they cannot exactly know in a deterministic way what exactly has happened to
my signal okay.

(Refer Slide Time: 01:56)

479
pX Y ( xi , y j) ¿ p X ∣ Y ( x i ∣ y j ) pY ( y j )
pY ( 1 ) ¿∑ p X Y ( xi , 1 )
i
¿ ¿ Q (1− pe )+(1−Q) pe
pY (0) ¿(1−Q)(1− pe)+Q pe

So let us try to see with that example so we have we could define this right PXY Xi Yj this is
something we can define from the information that we have already got so we have already told
that this, this and these four we know okay so this PXYj Xi Yj that is actually the joint so from
Bayes theorem we can easily evaluate now this must be Py given X, Yj given Xi, Px Xi okay so
this is something I can write this is equal to this why I have written with the condition on X.

That is because I already know that okay I have not written it with a condition on Y because that
is something I do not know okay so this is something I know we have also said how to marginal -
ize it now let us try to marginalize and get what do I want to get I do not want to get PX that is al-
ready known okay.

I want to get Py so I wish to get Py let us say 1 okay so what do I have to do I have to marginal -
ize this that means I have to sum over X so I will be doing it over i PXY Xi and YJ, I have al -
ready told that takes a value 1 so that should be 1 if I put 0 then I will get this done for 0 okay so
this is all that I will have to do so I can immediately write it in this format.

480
So I can write PY given X, okay so now it should be YI okay so I will just write Yi given X
where right so YI and now I have to put this Xi right so Xi I can put anything so Yi is for me it is
actually 1 so I can put Yi as 1 X can be taken a value of Xi so if I just do a sum I have two possi -
bilities right so I can do it for 1 and then multiply Px1 and I can do it for 0 Py given X.

So this should be 1 always given 0 Px 0 so I can write this way now you can see I know all this
already okay so I can actually calculate this so this is the beauty of it once I know all these things
I know how to calculate the individual probability also which is unknown so I had the prior prob-
ability I know all those conditional probability given X that means given that what is transmitted.

So I can now calculate what is the associated probability of Py okay so I can write this as P x1
that is probably px1 is Q and this is 1-PE and then this is 1-Q and this is probably PE so that is
how I calculate Py1 similarly you can also see Py0 can be evaluated in a similar manner so that
should be 1- Q *1 -P you can just test this and Q* Pe this should be the case okay so I get these

two fine now I wish to calculate which I was telling that given Y I want to see what was X okay.

(Refer Slide Time: 05:41)

pY ∣ X (0 ∣ 1) p X (1)
p X ∣Y (1∣ 0) ¿
pY (1)
¿ ¿

481
So my evaluation should be PX given Y okay let us say sorry I wish to calculate this, this specifi-
cally means that I have received at the receiver 0 what is the probability that 1 has been transmit-
ted so that means I have received an erroneous thing okay certainly whatever was transmitted was
1 and I have now received 0 so I am trying to evaluate this probability after receiving okay.

So we can just put Bayes theorem again so Px given Y can be calculated as Py given X, 0 1 P x1
given Py0 so this I can do okay and immediately this is already known which is actually PE this
is known this is Q and P y0 I have already evaluated which is 1- Q *1-P+ Q *P so I get this so
this is the way we calculate this particular part so whenever we have something received we are
trying to guess what is the probability that something else was transmitted okay for our case
something else is just another one so if we just calculate that that must be our error probability
okay.

So that is the probability that given I have received 1 what is a probability that I have probably
received an erroneous thing that something else was transmitted similarly I can do it for given I
have received 0 what is the probability that I have received probably an erroneous thing that one
was transmitted I have received this so this must be your error at the receiver side and we will try
to minimize these things.

So all you will be doing is once you get this now you have to see what has to be minimized okay
so something which can be minimized is PE that is the channel error so you will have to see how
to manipulate that so that we get overall error probability minimize so this is something people
do often okay so given these things now let us try to define another thing which will actually lead
us towards a continuous random variable definition.

(Refer Slide Time: 08:16)

482
 F X (x) ¿ p( X ≤ x)
F X (x) ≥0
F X (−∞ ) ¿0
F X (∞) ¿1
F X ( x1 ) ≤ F X ( x2 ) x 1 ≤ x2

So we will define something called cumulative distribution function so that is called CDF so what
is CDF our CDF definition is something like this we call it capital F so capital FX X so X is a
random variable this X is something we will talk about that so this is nothing but probability that
this random variable takes a value less than equal to that particular thing I have defined okay.

So in the real axis suppose for our that rolling of die we have events which are 1 2 3 4 5 6 when -
ever we say that X is less than equal to 4 that happens to be a CDF up to 4 when we say X is less
than equal to 5 that happens to be a CDF of parameter 5 so you keep doing that so that means we
will have to add all this probability till that point in the real axis so we will start from probability
minus infinity And we will go up to that point X so this is that definition okay

so immediately what we can say there are few things which we can immediately say about this
FXX can we say this that is always true because whatever happens it is just a probability and
probability must be we have already discussed that probability must be greater than equal to 0 so
this must be the case that is the first thing at minus infinity.

483
So that means this sum goes up to minus infinity okay and it starts at minus infinity so it could
not take anything any probability value so this must be 0 at infinity what it should be at infinity
means I have taken all the probability that means all possible probability values for all the ran-
dom outcomes I have already sum them and we know that summation of all those probability
must be 1 so this is a function which starts at 0 at a minus infinity level in the real axis and goes
up to plus 1 okay.

(Refer Slide Time: 11:13)

F X (x 2) ¿ p( X < x 2)
¿ ¿ p ( ( X ≤ x 1 ) ) + p ( ( x1 < X ≤ x 2 ) ) ≥ F X ( x 1 ) x 2 ≥ x 1 ¿
¿ ¿
Now let us see the nature of the function so we will now next prove this part that this is always
less than as long as this happens okay so we will be trying to prove this so this is will be with this
we will be able to say that it is a monotonically increasing function okay so let us see how do we
prove this so let us write FX X2 as probability by definition X must be less than equal to x2 this I
can write as PX less than equal to x1 so this Union X from this I can always write that X less than
equal to x2 means that I take those X which goes up to because I know x2 is greater than X1.

484
So I go up to X 1 and then the rest of the set I take so these two set I can always write it's the
probability that this Union this now these two set by definition they are disjoint because whatever
is covered within this is not covered within this okay so they are disjoint so we have already
proven that probability of union of two disjoint events must be summation of probability.

So this must be probability X less than equal to X 1 plus probability that your X remains between
these two now what is this by definition this is f X x1 this is a probability value whatever it is that
must be always positive so this is always positive so what is happening this particular value is
some value plus some positive term.

So this must be greater than equal to this at least this can be 0 so I can from this definition imme -
diately I can write FX x2 is always greater than equal to FX x1 for all x2 by definition we have
taken X 2 is greater than otherwise I could not have written this like x2 greater than equal to x1
so for all these things this must be happening so I can always write that this is just monotonically
increasing function okay.

It will never have a depth so it will never go down it will just increase 0 it will start it will go up
to one so all these things we have proven so this particular function some characterization we
have got now okay now let us try to define something which will take us towards the continuous
part okay so let us say now I have a random variable which is continuous what do I mean by that
let us give an example so we have already given an example in one of our earlier class.

We have told suppose I have a Pivoted rod and just I give exact a random force it starts freely ro-
tating around the point where it is pivoted okay and then it will stop somewhere it completely
where it will stop it completely depends on what kind of force I insert and the amount of force
that I can give that is a continuous any force I can give okay so accordingly he can stop between
that 0 and 2pi so it is a pivoted rot which is getting rotated and stopping somewhere at theta
sothis theta which is a random variable.

(Refer Slide Time: 14:30)

p (x< X ≤ x+ Δ x) ¿ F x (x + Δ x)−F x (x)

lim p( x< X ≤ x+ Δ x)
p X (x) ¿ Δx → 0
Δx
lim F x (x+ Δ x)−F x ( x)
Δx → 0 d F x ( x)
¿
Δx dx
d F x ( x)
p X (x) ¿
dx

485
Now where it will stop so that can be anywhere between 0 to 2pi and he can take any value so
this random variable theta now it is not taking a discrete value it can take any value 0 to 2pi any
value you can think of or any real number you can think of between 0 to 2pi it can take that so
this is a typical example of a random variable which is continuously okay so that means if I just
try to similarly map it to real axis between 0 to 2pi.

Now it can take any value okay so the definition of PMF we have defined that is not valid over
here let us try to see can we give some other definition over here so let us say that I have this ran-
dom variable which is taking a value between X and X + delta X so from that previous prove let
us take X as X 1 and this X+ deltaX as X 2 I can immediately write this as just do not give this
equality probably so this I can write as FX X +deltaX- FX X this is alright because we have
started defining it with FX x2 this is my x2 is equal to we have already proven.

That should be FX x1 plus this probability that it lies between x1 and x2 okay so this particular
probability value that it lies excluding of X from X to X+deltaX that is lying on that interval that
probability is defined by this okay so now what I do we actually divide this by del X okay so I
can I can or this side you can divide this, this FX x+ del x- FX X divide this by deI X and take a
limit deI X tends to 0 what do I get I will be getting this is differentiation of this FX.

According to the basic definition of differentiation so eventually what I have done is I have di-
vided this part by del X and then try to evaluate this so if this is that probability this was just the
separation okay so if I just try to relate this to this I can always write that probably this part multi-
plied by del X and taking del X tends to 0 right so this particular part whatever it is okay.

486
So that must be this part right now we all we have done is just divided by del X and taken limit
del X tends to 0 define this as some px X which is defining it so this divided by del X which is
defining this as px X so I immediately get PX to be differentiation of this one right that definition
we get and in that process I was more interested towards this, this probability so what that is I can
immediately see this px X if I multiply with del X and take the limit that del X tends to 0.

That must be my this Px that is the probability so what is eventually happening suppose I have a
particular probability mapping on the real axis and I have said this is my X and this is my X+
deIX okay and I am trying to find out this particular part which is a probability that it will be ly-
ing between this and this so that overall probability I am trying to report as del X goes towards
zero what will happen the variation of that okay will become vanishing there will be no variation
okay.

So at that point I can say safely that it is actually whatever that Px X if I just plot it over X and it
is the area under that part which defines my probability that it will lie between X and X+ delX so
this probability is nothing but whatever this particular thing I have got by differentiation of this
CDF cumulative distribution function if I multiply that with a del X that almost means del X
tends to 0.

So this PxX will not be varying this would be almost constant so that x del X is almost getting the
area under that curve okay so which is the definition now we still have not defined what this Px X
is but we can now see that there is a particular function px X whenever we take it to the continu-
ous domain this is just a differentiation of the CDF if CDF is defined the way we have defined I
that it is less than something okay.

(Refer Slide Time: 20:52)

487
lim P x ( x) Δ x
Δ x →0

So that can be defined so that probability that CDF if I differentiate it I get whatever value I get if
I plot it with respect to X so if I just plot it with respect to X and then choose a particular X and if
I get the area under it for a very small del X that must give me the probability that it is lying be-
tween X and X + del X okay right so that is the probability that it will be now what do I get from
there suppose let us say del X goes towards zero.

So this will be almost coming to this point so almost I am going back to a probability of a value
of a random value like our means discrete random variable what will be that value irrespective of
pxX if pxX is bounded we know that P xx must be bounded as long as that is happening what do
we get pxX into del X, deI X towards tends towards 0 so I get an individual probability 0 so in a
continuous variable like that example that I give a random force and it just keeps rotating and
fixed at a particular theta okay.

So from the very definition frequency definition or favorable cases let us try to think about this
that it will be stopping exactly at theta equal to 60 degree what do I get about the associated prob-
ability how many favorable outcomes are there only one that is the 60 degree okay how many
overall possible outcomes are there how many angles I can get infinite so what is my probability
it should be 1 by infinity.

So that must be 0 so individual angles I must get a probability value of 0 so remember that is why
we take an example of this particular thing knowing that many of you might be knowing proba-
bility theory and given the similar explanation for Fourier series to Fourier transform the same
thing we are doing or dealing with again its PMF or discrete random variable.

488
And we will be calling that as PDF and remember we will not be calling that deliberately add as
distribution function that should be a density function it is, it is whenever we say PDF it is actu-
ally probability density function not distribution it is it has nothing to do with distribution, distri-
bution is not that px X distribution is multiplied by this and then limit the del X tends to 0.

So this is actually the probability that gives me probability what PX X does not give me probabil-
ity it just gives me almost like a density that as if at that point if I now multiply with del x suffi-
ciently del X small so that this does not get changed and then I will be getting the area will be
defining the probability okay.

So as if the probability divided by this del X gives me this okay as long as del X is sufficiently
small so that is why it is becoming density that per unit on that real axis per unit that random
variable change what is the probability so that is why it is density it is not probability so individu-
ally at a particular point the probability will be always 0 but there will be a relative nature of this
and that tells you where exactly the probability values are higher and where exactly the probabil-
ity values are lower.

So suppose I have just drawn a PDF like this then I will be able to say that around zero it has a
higher probability I would not be able to say specifically that at 0 this is the probability that will
be 0 at one this is the probability that will be all zero but I can say around zero this is the proba-
bility and this is having higher probability around zero whereas as I go away from zero it will
have lower probability so from the nature of this curve.

(Refer Slide Time: 25:20)

489
p ( ( x1 < X ≤ x2 ) ) ¿ F X (x 2 )−F X (x 1 )
x2
¿ ¿ ∫ x p X ( x)d x
1

d F X (x)
¿ p X ( x)
dx
x1
F X ( x) ¿ ∫ −∞ p X ( x)d x

I will be always getting that hint so that is one thing and the second thing that I wanted to discuss
is something like this I can always write P of suppose so this particular part we have already seen
that can be written as FX x2 minus FX x1 this is something we have already proven so what is
this, this I can write from minus infinity to plus x2 this FX x2 we have already shown that deriva-
tive of FX X is equal to the PDF.

Therefore from the relationship of derivative and integration we can say FXs is just the integra-
tion of Px X so I can write that probability distribution function going up to X2 must be interre-
lated this way minus similarly I can write right so very nicely because I have this definition so
FX X DX was Px X.

Therefore from the conjugate definition of differentiation and integration I can always write this
okay so I will be able to write FX X must be integration minus infinity to X px suppose T DT so
this is something I could write so here also I can write that way immediately from this I can write
it from x1 to x2 again from the definition of integration with respect to limit.

490
I can write this so what it says that if I target instead of a single value a particular zone on that red
random axis or real axis where the random variable is defined I can always integrate that PDF
from that particular value to whatever value I wish to evaluate it so I will be getting the corre-
sponding probability so if you just think about that pivoted rod if I just say instead of specifying a
60 degree angle I say between 0 to 60 degree it will be lying what's the probability associated
probability.

So then there is a ratio between 0 to 60 degree is a solid 60 degree angle and what is the overall
angle that is 360 degree so 60 degree/360 degree should be my probability which is 1/6 okay so
whenever we talk about a range because it is density I integrate it and I get a probability associ-
ated with a range but individual value will be all 0 and that is why we were talking about this
thing right so in this range if you integrate the value will be higher associated probability will be
higher.

And if you integrate in this range if even if the ranges are same probability the integrated value
will be smaller and the associated probability will be smaller so you can always say I have more
likelihood to get this random variable whereas I have less likelihood towards getting this random
variable that has lower probability so PDF actually tells you that information okay or gives you
that information and this is how they are linked so will be always able to evaluate whenever we
need to evaluate That it is the probability between this to this we will be able to calculate that
okay

so with this definition probably we have now gone into the definition of continuous random vari-
able and its associated functions like CDF remember CDF is still called cumulative distribution
function because that is till the distribution up to x1 what is the overall probability that is still a
probability where as PDF that is not probability that is a density.

So often people do this mistake even in Fourier series and transform also same thing happens
whenever you go to transform its spectral density remember that wherever whenever you are
talking about Fourier series in each frequency will be getting a particular amplitude value or
overall energy value or power value I should say okay so there is a definite power value whereas
for Fourier transform you do not have a power value.

It is spectral density actually and there also you have seen that it is almost similar concept you
have if you integrate from some particular value to some particular value you get instead of prob-
ability there you get some amount of power or energy okay so with this definition will stop today
and then what we will do next is try to see some more things which will actually get us toward
random process so some more things on probability theory thank you.

491
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-35: Probability Theory (Contd.)

So we have so far discussed about discreet random variable as well as continuous random vari-
able, so in the last class probably we have started discussing already about the continuous random
variable and it is associated PDF and CDF.

(Refer Slide Time: 00:43)

492
 ∞
∫ −∞ pX (x)d x ¿1
x1
F X ( x) ¿∫ −∞ pX (x)d x
p X ( x) ≥0
x2
p ( ( x1 < X ≤ x2 ) ) ¿ ∫ x p X ( x)d x
1

So some of the property that we have already mentioned is something like this 1 is which is a uni-
versal theory of probability that if PXX is the PDF of a random variable X and this must be 1 and
the CDF and PDF are related like this. Of course its probability so this must be greater than equal
to 0. So these 3 things we have already discussed about them we also told what do we mean by
this PDF it is an density function remember even it is called PDF its probability density function
not distribution. So Basically This PXX does not give you probability its value is always or value
at a particular X is always 0 it just provide the density and that is why the other formula that p
suppose X not that equality so that always comes up to be This is always true so from a particular
value to the other value If we wish to calculate that our experiment that random variable will lie
between this 2 value that has a probability term and that probability term we get through integra-
tion of this PDF so because it is an density function so if you do integration then you get the
probability but individual at particular X you do not get anything we have also given an example
of these things right.

493
So what will be doing now in communication one of most important PDF or most well known
PDF that is being applied you just keep an property of that it just a physical thing we need to just
understand some of the property and some the function associated with that so that is called
Gaussian random variable which is defined as this defined by the PDF.

(Refer Slide Time: 02:54)

1 −x / 2
2

p X (x ) ¿ e
√2 π
¿ ¿
That is one definition or it can be another definition where you have this so of course as you can
see if I just put sigma= to 1 and m = 0 I get back this one so that's more generalised gaussian dis-
tribution where this m is called the if you see thats called the mean we will see how the mean can
be calculated for the random variable so then we will understand what do we mean by mean and
this sigma is means the standard deviation or sigma square is the variance.So as long as the vari-
ance is 1 and your mean is 0 we get back to this, that definition so this is the typical gaussian dis-
tribution it if the mean is 0 it look like this.

The famous bell curve or if the mean is m so it will centered around m it look like this as we in-
crease the sigma what will happen this will become more and more flatter and the top will come

494
down because the area under this must be always be 1. And as you decrease the sigma this the top
will go up and up and the this will become more sharper. So that is typical Gaussian distribution.
So what will try to see this Gaussian Distribution has been related to some other function, like
they call it either Complementary function or Error function or the Cue function.

So we will see what are those functions and how it is related so that is something we will try to
define because most of the times may be not much in analog communication but when we will be
going into digital communication we will see that this cue function error function those are the
main functions which keeps coming back. For the noise analysis or communication analysis with
noise.

(Refer Slide Time: 05:30)

495
 1 a −x / 2
2

Q( y) ¿ ∫ e dx
√2 π y
1 −x / 2
2

p X ( x) ¿ e
√2 π
y 1 −x / 2
2 1 x
F X ( y) ¿ ∫ −∞ e d x = erfc ( ) ¿
√2 π 2 √2
F X ( x) ¿ 1−Q(x)
2 ∞ −y2

erfc (x) ¿ ∫ e dy
√π x
¿ ¿
Okay so let us, try to define a function called Q (Y). This is by definition this given as one by two
pi some Y to infinity e to the power minus X square by two dx. This can be easily related to our
that PDF or CDF as such so if Gaussian distribution we have already written as this PXX is 1 by
root 2 pi e to the power –x square by 2. So of course, Fx(x) will become – infinity to x and the
same thing right. So this thing with the dummy variable let us say 1 by root 2pi e to the power or
if we just put it with variable Y so this will become this one –x square by 2dx.

So immediately we can see that, this and this are complementary okay! So I can write Fx (x) = 1-
Qx so this is always true. Remember it is a distribution so the integration of that from – infinity +
infinity must be one. So that is what we have applied over here that – infinity + infinity is 1- in-
finity to Y if I just take that out this will automatically become Y to infinity Right! So that’s the
way they are related

There is another function which is Error Function or ERFC that is defined as this 2 by root Pi so
these are just by definition, these functions are defined so this is called error complementary
function okay. That's by definition is this, this complementary function is of x of course so the in-
tegration it is 2 by root Pi so it goes from x to infinity e to the power –y square immediately you
can see what I need is to define it with respect to the Gaussian or Q function

So if I just go back to the q Function this has the relationship with the Q Function but this is x
square by 2 so I have to replace it, y must be replaced as x by root 2 right! So if I just do that im-
mediately I will be getting that particular function and correspondingly what will happen this will
become 2q so that just by doing that replacement and then changing the integration so Y if you
replace it with respect to another dummy variable take U so put it as U by 2, u by root 2

Then this will be immediately U sq/2 and here, that here that 2 will be coming into the limit and
then you will see that it will become x root 2 and there will be a two factor which will be can-
celled out okay. So this is how it is and we can also write this Qx can be as well as written as half
just the reverse thing okay. So here the Error FC or ERFC was written as a function of Qx and we
are just doing the reverse thing same thing

496
Just a manipulation of should I put in the this one x/root2 or x root2 ok! so if you just put that au-
tomatically you can manipulate and you can get this relationship ok! This is just to know that
how this Error function or Error Complementary Function or Complementary Error Function or
this Q function how they are related to the Gaussian PDF Ok!

Because most of the time we will see that these functions are means numerically evaluated for
different values of X so therefore Gaussian can be immediately mapped to this functions ok! So
that is why it is often required. So it is good to know these things okay. This was just for general
information so that whenever you are handling Gaussian Function probably you will know these
things.

Now what we will be doing what we have already done for discrete random variable that joint
distribution, conditional distribution conditional probability all those things how do you define
and characterize them in the continuous version so that is something we will start. So first we will
start with the joint distribution.

(Refer Slide Time: 10:58)

497
 Δ
FXY ¿ P( X ≤ x ,Y ≤ y)
∂2
P X Y ( x , y) ¿ ❑ F X Y (x , y)
∂x ∂y
PXY ( x , y ) Δ x Δ y ¿ P( x< X ≤ x + Δ x , y <Y ≤ y+ Δ y)
∞ ∞
∫ −∞ ∫ −∞ pX Y (x , y) d x d y ¿1

So that means we have two random variable which takes continuous value one is X and another
one is Y okay. So what I need to know is this Fx(x,y) = which is by definition written as proba-
bility that my random variable x is less than equal to some x this x specified x and the random
variable Y less than equal to that specified y okay. So this is called the joint CDF Ok! So when-
ever we talk about joint CDF that means it is actually means there are underlined two random
variable

And whenever we specify that joint CDF with two functional input asX and Y then we are actu-
ally saying that my, it is the probability that one of the variable X capital x that is always less
than equal to some x specified over here and the other variable Y capital Y less than equal to this
particular y okay. So whenever this is happening if you just calculate the overall probability that
with what probability this particular thing phenomena is occurring then we get this one okay.

And with the same definition we can define because it is joined so there should be a double dif-
ferentiation with respect to CDF to get the PDF right. So Pxy with the similar mathematical thing
which we have already proven CDF and PDF are related to differentiation and integration.

So we can just write it as double derivative with respect to x and y. so this is the corresponding
density joint density function.

So this is actually the joint density function that we were talking about okay. So this actually tells
you what it tells I will list that so this like the single variable one Pxy (x,y)its density so I have to
multiply with the corresponding delta means change of random variable so this define a probabil-
ity that my x or the random variable x is greater than x and less than equal to x plus del x so it lies
between x and x plus del x and the other one jointly both the things are happening so other ran-
dom variable y that lying between y and y+ del y.

So that is the definition of the PDF right. So PDF is the density function as long as we multiply
by the change okay. So we get the area under that and our assumption is this del x and del y are
sufficiently small so that things are not changing over there ok! So this actually characterizes the
probability that x lies at the vicinity of x that means x to x+ del x and y lies in the vicinity of y.
that means y+, y to y+ del y right! So this is the definition which we have seen already for the
single variable right. It is just the extension.

So immediately we also know because this is the probability. So if we just integrate over all pos -
sible values of x and y I must be getting 1 because that is the sudden event because I am search -

498
ing through all values of x all values of y. So I will be getting something definitely. So if I just in-
tegrate it from –infinity to + infinity this Pxy (x,y) d x d y that must be giving me one. So this is
something again the fundamental postulate of probability theory.

Because it is probability so if we just integrate it over the whole region of interest I must be get-
ting one okay. So this will be always happening so let us try to see from joint distribution in our
other case the discrete random variable we have defined the marginal distribution. So let us try to
see whether that is happening over here.

(Refer Slide Time: 14:37)

lim pX (x) Δ x ¿ lim p( x< x ≤ x+ Δ x ,−α < y< x)

Δ x →0 Δx → 0
∞
d u ∫ −∞ p X Y (u , x)d w = ∫−∞ p X Y (u , w)d w ¿
x+ Δ x ∞
¿ ¿ lim ∫ x
Δ x →0
¿ ¿
Similar definition of marginal distribution, so let us try to characterize this thing that this is my
marginal distribution okay. So that means now I have made it free of y I knew that there was a
experiment which was going on where two variables were randomly varying x and y, so but I m
interested in for any value for y I do not care what is the probability density function of x that is
what I am interested in.

499
So if I do this that tells me the probability that x lies between x and x+ del x so that is the proba-
bility that I am targeting! So I can write this as limit del x tends to 0 that is the random variable
but remember I also had y in the definition but now I do not care about y so I am saying that y
must be lying anywhere. So from – infinity to +infinity okay, by definition I can write this I am
only interested in x that I want to restrict my x between x and x+ del x so that is my px(X) x del x
where del x tend to 0.

But I have only the joint distribution so that also specifies y so y I do not want to keep any re-
strictions so I can take any value from –infinity and +infinity okay. So this is the probability I am
targeting if I wish to only get x as marginal distribution right. Let us try to evaluate this thing, so
I can now write it to be limit del x tends to 0. So now let us write it from the definition of PDF so
y goes from –infinity to + infinity.

Therefore the integration this pxy this must be integrated overall y that is the first task! – infinity
to +infinity for x I should be integrating it from x to x+ del x right. So this integration should be x
to x + del x right. So I am integrating for y from – infinity to + infinity and for x I am integrating
from x to x+ del x of course you can always argue that okay, it is excluding that value of x but by
the definition of integration I do not have to take that.

As long as I am integrating if you just divide it into small small area the last one you can take it
out. So this is what happens right. So this was according to the definition of probability theory
this is a probability and we always know that any probability is given by the integration of it is
PDF from particular value that we have specified over here for x this is the value we have speci-
fied so that is where we are integrating okay. Y that is the value over which we are integrating

But we have to do another thing because this is in the limit so maybe we can just define it as U
and W so two dummy variable. We should not mix the limit with the internal integration variable.
That is all right, so Pxy u v we are integrating from for y this and x this okay, the assumption is
this del x tends to 0 so within that interval this is not changing as it is varying it is not changing
so what I can do , I can actually separate out this integration

I can still write del x tends to 0 I can take x to x+ del x this where ever x variability is there that
means du and y –infinity + infinity Pxy(u,w) dw right! I can write it this way. Why I could sepa-
rate this because I know that this particular thing is constant with respect to x. So therefore, x in-
tegration and y integration can be separated out. I will have nothing within this. So this happens
to be del x because integrating du from x to x+ del x so that should be del x

So I can write limit del x tend to 0. This is del x and this – infinity to + infinity Pxy (u,w)dw.
Now match these two, I get a definition of this, so immediately I can write my Px (x) is nothing
but the integration with respect to the other variable. So I take the same this one same joint distri-
bution I integrate it with respect to y from – infinity to + infinity so that is called the marginaliza-

500
tion. So a particular if it is a joint distribution I want to marginalize it so marginalize it with re-
spect to the other variable. I get the marginal distribution with respect to whatever I am targeting

Here we are targeting for x and the we have to marginalize it respect to y. same thing will be hap-
pening if we wish to target Py(Y) then we have to marginalize with respect to x. so that is how
the joint distribution and the marginal distribution are related.

(Refer Slide Time: 20:51)

pX ∣ Y ( x ∣ y ) pY ( y ) ¿ pX Y ( x , y )
∞ ∞ pX Y ( x , y )
∫ −∞ p X ∣ Y ( x ∣ y ) d x ¿ ∫ −∞ dx
pY
p (y)
¿ ¿ Y =1
pY ( y )

Now let us try to see the conditional distribution, so the Bayes theorem still prevails I can always
write the way we have written Px/y (x/y)Py (y)=Pxy(x,y) so this is always true we have already
proven this. So the joint distribution is related to sorry, conditional distribution is somehow re-
lated to the joint distribution. Here we should not talk about any distribution we are just saying

501
that conditional probability is just related to the joint probability for a particular targeted value of
x y. that is how we are discussing about these things.

Now what we can write, we can prove that this is also forming a PDF. What does that mean?
That means, if I integrate it with respect to suppose x over – infinity + infinity I must be getting
one. Because it is a finally this is nothing but given random variable y = y this is the probability
of x=x. this is a density function but if I just take it multiply with del x and limit del x tends to 0
we can say that x lies between x+ del x. I will be able to talk about that probability

So if that is the case if I integrate that function with respect to x because x is the variable over
here y is the condition so that condition I am not bothered about if I integrate it with respect to x,
I must be getting 1 , if this is a this should be a PDF or probability density form. So I should be
proving that. Let us see – infinity + infinity Px/y(x/y)dx so that is what I want to evaluate so I can
write this as from this formula Pxy(X,Y)/Py(y) dx I can write it this way.

Just coming from this formula, its Px given y is nothing but pxy divided by py. This is the inte-
gration over x so y has nothing to do with it so I will take that out. What is this? This is just we
have proven now that’s the marginal distribution. A marginalizing with respect to x. therefore I
will be left with y so that must be Py(y) divided by Py(y) = 1. So you could see that this is actu -
ally a PDF. Whatever we get because we have just from the Bayes theorem directly correlated it.

Now we can see it is actually a joint PDF sorry conditional PDF. So we can evaluate this PDF.
Similarly, for the other things also, y/x similar things will be getting. Now from here conditional
we for the discrete random variable we came to the independency. So same thing we will do over
here.

(Refer Slide Time: 24:33)

502
p X ∣Y ( x ∣ y ) ¿ p X ( x)
x y x y
p X Y ( x , y ) ¿ pX ∣ Y ( x ∣ y ) pY ( y )= ∫ −∞ ∫−∞ p X Y (u ,w)d u d w ¿= ∫ −∞ ∫ −∞ p X (u) pY (w)d u d w ¿=F X ( x) F Y ( y)¿
¿ ¿
Px given y, x will be independent of y if this is just this does not depend on y. So whatever the
value of y given y, I don’t care! I will still get the xx if this is happening then I will know that it is
a independent event means x doesn’t depend on the occurrence of y. y can happen, might not
happen still my x will be following the same thing. It is like the tossing the coin twice. If it is a
fairly unencoded unbiased coin then what will happen? The second toss the outcome does not de-
pend on first toss what has happened. Whether that was head or tail I do not care, I still have sim-
ilar probability value of the second toss. that's same thing independent.

And if that is the case now go to this pxy definition what it was according to the Bayes theorem
this was px given y pyy this was the definition. Now this is according to this, so we immediately
see if they are independent then the PDF also gets multiplied or separated out in the multiplica-
tion form. That is a very interesting theorem of probability theory. So if two random events are
independent then I can always write this otherwise not because otherwise you would not be able
to write this thing.

503
And immediately we can also get into this, this is actually the CDF of the joint distribution so this
I can write as from – infinity to x and – infinity to Y, Px y (u,w) two dummy variable I have
taken du dw. So I can write this now the good part is as they are independent they can be sepa-
rated out so I can write this. Because they are separated I can now separate these two integrations.
The integration with respect to variable u and w can be separated. So I get – infinity + infinity
PXu du sorry -infinity to x and – infinity toy Py w dw which is nothing but the fxx and fxy so this
will be, if I write this will be equal to fxx and fyy even their CDF gets multiplied form.

So so far these are the things which we could means it is almost the similar thing that we have de-
rived for discrete random variable we have derived all those things for continuous time random
variable. So in the next one we will try to see is what is the statistical property or measurement
property that we can extract and how we relate them to probability okay. That is something we
will try to see in the next class. Thank you.

504
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-36: Probability Theory (Contd.)

Okay so for Gaussian random variable we have already started discussing about mean, Variance,
standard deviation, so let us try to characterize those things now okay. What do you mean? And
how there are related to corresponding PDF and CDF, so that is something that which we will
target in this class. So let us try to see the basic or classic definition of mean.

(Refer Slide Time: 00:49)

505
 N1x1 + N2 x2 + ⋯ + Nn xn
x̄ =
∑ Ni

( ∑ Ni ) ( ΣNi )
N1 N2
= x1 + x2 + …

= pX (x1) x1 + pX (x2) x2 + …
N
xi pX (xi)
∑
x̄ =
i=1

It is just the average, how do you do averaging, we actually suppose I go to the class and I say
that I want to get average height of all the students, so what I do? You actually ask every student
their height, you note them down, and you add them together and divided by the number of
students!"right. That is how we calculate average, so this well known thing okay. Now suppose
we want to do this in a little different way, we say we want average height but we want to just do
this with just the integer number.

506
Okay so it is 5 feet, 6 feet, 7 feet, or he feet something like that okay, we will have no other
things. So basically what we will say that if somebody is having height from 4 to 5 will report
that as 4 okay, if somebody whoever are falling from 5 to 6 will report that as 5 and so on. If I
just do that then it will be not a single observation for each one of them okay. so will have to then
see, how many of them are falling under that category and then accordingly we have to take.

So what will happen? It is like this will probably after doing that we wish to calculate that
average, so then we are saying, okay how many of you are between 4 and 5, we will take that
number. Let say that number is N1 okay, so and for each of them how much I'll be evaluating that
is to me is 4 or I can take that as 4.5 also whichever I take that N1 into whatever the value, so let
say that is x1 I will do that. And then I will also calculate how many are from 5 to 6, so I will take
the number I will multiply by the number that it represents okay x2

It might be 5.5 okay, something like this and I will do for N number of such things, so let say Nn
number last one I get xn and then I will divide the whole thing by the overall number of students,
so it is nothing sum of this Ni. So in h category they were N1, N2 or N3, N I number so I will
divide by this. So that gives me a sense of average, if I do it this way okay. Otherwise it is also
true, this was a specially created experimentation but if I wish to calculate suppose I take a coin
and I say if it is head then it should be +1 or if it is tail it is -1.

We have already told random variable means it has to be map whatever the events are there that
has to be marked to a real axis, so we have marked it – 1 and +1 and now I do this experiment for
let say 5 lakhs time okay and then I count, how many heads are occurred? I multiply all of them
with +1, and how many tails are occurred? I multiply all of them with -1 okay. So that as many
numbers into +1, as many numbers into -1 and divide by the total number. That is how evaluate
things, so it is always true.

So for the dice also it will similar thing there will be 6 such part right. Now let us try to see little
bit more carefully, so this N1 divided by this summation Ni that is factor which coming into x1.
We will always be getting this kind of factor,N2 /summationNi okay. So this summation of Ni is
actually the amount of time I have done that experiment, repeated that experiment, so that is N.
my frequency definition of probability was that Ni take it towards infinity. So once this is going
towards to infinity, what happens to this? This actually tells the probability that this x1 will be
happening.

That is what we have defined That's how we have defined it okay, so for head and tail if we just
take that from 5 lakhs to may be 5 core or even more then almost half the time it will be head,
almost half the time it will be tail. So how many times it has occurred, favourable things has
occurred divided by as many times the experiment has be run, so this ratio is always in frequency

507
term gives me the probability. So this must be as summationNi tends towards infinity this must be
the probability of associated random value x1

So I can write this as px of x1, that x1 will be happening into x1 + this I can write px of x2 x x2
similarly it goes on okay. So immediately I can see the x bar is nothing but in a discrete case as
many times I will be doing this experiment right, so that is the famous definition of mean. So
mean is nothing but your random variables are there, all possible things you take and you
multiply with associated probability, you add it up for all the possible outcomes whatever you get
that is mean.

So from classic definition of mean we could see that, this is the definition of mean, w. r. t its
associated probability okay. if we just extent it from summation to integration that is how we go
to continuous one, so that immediately that summation will become that integration and inside it
will be associated PDF.

(Refer Slide Time: 07:17)

508
 ∞

∫−∞
x̄ = xpX (x)d x

ȳ = g(x)
N

∑ ( i) X ( i)
= g x p x
i=1
∞

∫−∞
g(x) = g(x)pX (x)d x

So I can write x bar if it is a now the continuous variable, so that should be now that summation
will become integration from - infinity to + infinity in the entire real axis where this can happen,
x into px x dx right, so that is for the continuous variable , so we get a associated mean, so this
how mean been calculated and that is why probably thatPDF is so important because mean is
something which is called the average. So whenever there is a randomness happening you tend to
get the average, try to guess the average what is the average of it, so on a average what should I
expect from this random experimentation.

So whenever you try to guess the average you need to have this PDF associated probability
density function of that particular or for the case of discrete variable you need to have that PMF,
so either PDF or PMF will be searching because you know that through which we will be able
calculate the associated mean, one of measurable parameter okay. So now let say if I have a
function of a random variable okay, so what do I mean by that?

Suppose this x is the random variable and what I get is the function which maps that random
variable to some other real axis. So it is just the functional mapping, which ever you wish to do
that okay. so suppose there is something called temperature variation okay, so something varies
with the temperature square okay. Now temperature in a day is a random variable, you take that
and your that targeted variable is varying with temperature square.

So it is mapping that random variable to that particular temperature square variable what you
have to find oaky, so that kind of thing. So that will be our definitely another random variable
which completely depends on input random variable okay, so that if I just say y, I might have to
finally, I might not observe the temperature but I might observe that outcome which is the
temperature square probably okay, so that is how I can measure things.

So I might have to get my y bar, whatever I am getting that random variable, so this y is also
random variable which is just eventually function of random variable. So I need to evaluate the
mean of that, so what happens, if I wish to calculate this y bar that should be just this g x bar, so

509
whenever there is a x happening right, to calculate the average I was multiplying x with it
associated probability. Now this x will change to g x, the associated PDF is still on x, for each the
value of x that is the, if I just correlate this probability to frequency term that should be the case.

So I can always write this as Summation if it goes up to n, so g xi into it's associated probability,
so this is something that I can write or in a continuous one it will just be the integration. So g x
must be-infinity to + infinity right. so similar definition it is instead of x it has become gx, so this
is more generic, if gx = x then I get the earlier definition that is the because it is a direct mapping,
it maps to itself, 1 to 1 mapping and maps to itself only if the gx = x.

If any other thing is still have the some kind of mapping and I will take that associated
probability and I will still get the average of that okay. So this is the functional average we talked
about, now let say suppose I have.

(Refer Slide Time: 11:43)

510
 ∞ ∞

∫−∞ ∫−∞
g(x, y) = g(x, y)pXY (x, y)d x dy

g1(x, y) + g2(x, y) = g1(x, y) + g2(x, y)

x + y = x + y if g1(x, y) = x, g2(x, y) = y

A joint thing okay, where there are two random variable x and y and they have a joint PDF which
is defined as Pxy x, y and I also have a joint function; I want to evaluate that associated mean
okay. That is nothing but same thing I have to multiply for each x,y with it is associated PDF and
integrate over the entire means domain of x and y, so if I wish to calculate g (x,y) bar that must be
- infinity to + infinity,-infinity to + infinity this Pxy and this is should be multiplied with gxy
right. So that is the average of g xy that we will be able to calculate, if we just do it that way
okay.

So if this is the case I can also now start writing if I have g1 x,y that is 1 mapping and another g2
x,y another mapping, random variable remains the same. I wish to evaluate the average of this
one, so this is very clearly it can be seen that, this g1 (x,y) g 2 (x, y) will be separated out then
two integration will be separated out and I can immediately that it is nothing but g1 x, y bar + g2
x,y bar right this is very true because once you replace this whole thing over here then two
integration will be separated out and I will get this same result.

So any functional mapping if two addition functional mapping I wish to get the mean of that, I
can separate out as if I calculate the functional mean of the 1st function and functional mean of
the 2nd function and I add them, so this is always true that is pretty obvious from the integration
definition or through PDF the way we define mean right. So now I can just say that this g1 x, y is
nothing but x and this g2 x,y is nothing but y. I can have this functional mapping right.

So it is always possible, it is a function of x and y but the probably y = 0 always, this is also
function of x and y but x = 0. So this can happen this is a valid function, if I just put it over there,
so what will happen x + y bar is always xbar + ybar, so this is always true. For this is to happen
you do not need anything remember no independence among x y and all other things are required.
Mean is always additive, if we take 2 random variable whether they are independent co related
whatever it is, you don!t care you can separately calculate the mean and you can always add that.

So this something which is very important, you should always remember that for mean you do
not have to think whether those two random variable are at all dependent on or independent or
any other relation, you do not have to care about that. You can always say okay if I can
individually get the mean of these two I can always add them up that should be the overall mean,
so this is always true okay. Now let see if this g x, y is nothing but

511
(Refer Slide Time: 15:42)

g(x, y) = g1(x)g2(y)
∞ ∞

∫−∞ ∫−∞
g(x, y) = g(x, y)pXY (x, y)d x d y
∞ ∞

∫−∞ ∫−∞
g(x, y) = g1(x)g2(y)pXY (x, y)d x d y
∞ ∞

∫−∞ ∫−∞
= g1(x)pX (x)d x g2(y)pY (y)d y

= g1(x) g2(y)
XY = x y

a product of 2 functions where they are separately just function of x or y okay, if this is the
functional mapping and this is the nature of these two function then what how do we evaluate the
overall mean value. So let say I want to still get g x, y for this particular function the mean of that

512
right, so I can again write-infinity to + infinity,- infinity to + infinity so this is g (x, y), P (xy), x, y
dx, dy. Now I know that this can be separated right. Now the question is if this can be separated?

That is the condition where we will be able to separate these two if x and y are independent,
immediately I have already proven that it can be written as Px x and Py y, now the integration can
be separated out, so I can write this as gx into P x x dx integration - infinity and + infinity gy x Py
y, so this is nothing but gx and gy . So this will always happen, remember not necessarily this
joint function has to be same mapping g this can be g1, g2 whatever it is that might be different
also, so in that case it will be just g1, g2 whatever it is, so this would be g1 g2, so this would be
g1 g 2 the mapping can be different.

The one we have taken probably that is the specific function we all know that for Gaussian that is
what happens but you do not have to really have that similar kind of mapping. Any way whatever
the mapping is you need to be or it should be possible that you separate them out okay, so and
you can put them as multiplication where 1 part is just the function of one of the variable other
part is just the function of other variable. As long as this is possible and if random variable that is
x and y, those are independent then you know that overall this functional mean is just the
multiplication of their mean. So if now that function I take as x and y x multiplication, that's
separable exactly.

(Refer Slide Time: 18:31)

513
g(x, y) = x y
g(x) = x
g(y) = y
xy = x y

Because you have just this multiplication this, this is just the function of x, this is just the
function of y that has happened, so my g(xy) is nothing but xy and gx this also the typical
example where both are remaining as same function. This is x gy just y okay, as long as this is
happening I can write this when if x y are independent. So this is the strong condition if two of
the random variable are independent then they are mean of multiplication it is nothing but
multiplication of mean. This is a very strong condition for xy, so this always happens okay xy bar
is equal to x bar into y bar

Now let us talk about some of the higher moments, so mean is we called that, most of you will be
knowing already that is the 1st moment actually, the higher moments are defined as this.

(Refer Slide Time: 19:54)

514
 ∞

∫−∞
n
X = x n pX (x)d x

variance = (X − X̄ )2
= (x 2 − 2x x̄ + (x̄)2)
= x¯2 − 2 x̄ x̄ + (x̄)2
= x¯2 − (x̄)2
2
∞ ∞

∫−∞ [ ∫−∞ ]
= x 2 pX (x)d x − xpX (x)d x

This is by definition xn bar okay, so higher moment why mean is called the 1st moment, it is just
because if you , just put n= 1 you get the equation of mean, so that is why mean or average value
is the 1st moment and then if you just put it has 2 it has the specific mean, so that is the 2nd
moment and from 2nd moment you can actually get the variance or standard deviation that we
will talk about now.

515
So when we talk about variance, where actually mean by this, this as a big significance in any
statistical process, what it is saying? It is actually whenever we write this that means we are
taking an average okay, so bar means average. This x is the random variable right we are picking
different values what we do we first calculate the mean and here what we are trying to do for
every value of x we are actually taking how far it is from the mean right.

And square that because if they are on both sides of mean then they will cancel each other, if I
just add them together, so we square and add them add all of them, divide by as many random
samples are there divide by that N I will get this average here. So this is actually saying, it has the
physical significance, the saying how far it is deviating from it mean value. The square of those
deviations so basically you are taking all the samples if this value is very low that means
immediately you can say most of the values are lying very close to the mean value.

If this is very high that means it is flatter we are talking about the Gaussian with sigma 2 low or
high that is the case. If it is having a higher value that means the most of the values are actually
lying away from mean. so overall mean can be seen. I might I suppose the height of a particular
class or section if that is reported from it is all over the place, suppose let say from 4 to 7 feet,
everything you observe okay. then you can say that may be the mean value is still 5.5 or 5.6
something like that.

But there is a huge deviation, this might happen where as if I just do it for the age in a particular
class what will happen they will all running very close to the average age because generally in a
class we just give admission if their ages are almost similar right. so if you do that so for age
what will happen, you will immediately see that their ages are not deviating from the mean value.
Whereas height or any other weight if you start observing that you will see a lot of deviation
might vary quiet randomly there will be a huge deviation.

So in that case just mean does not characterize anything, for a particular random experiment if the
mean suppose for the 2 random experimentation the mean might still remain same but it might
deviate quite small from the mean value or it might deviate heavily from the mean value. So
these two samples are quite different, if you just observe those two samples or those two random
outcomes we will see there, they have a huge difference, probably mean will be still the same.

So a typical example if I give, if you take the age of a particular class 6 or lets say 6th standard
okay, if you take the random samples from 6th standard or if you go and take the age value of
competitive exam and there is a coaching center over there and you are trying to take that age
value, you will see a lot of deviation okay. where as here 6th standard if you just take those age
probably it will be almost similar okay. So that is what happens, you can have similar mean value
reported but you can have a large deviation.

516
So basically that characterizes the randomness of the associated process, so that is why just mean
will not characterize anything, you have to go to this variance, which says something about it but
variance also it is not enough, we will see that all those higher moments that as to be
characterized, then only the complete random process can be or complete variability of that
random variable you will be able to characterize it. The best thing to characterize a random
variable is this one. This provides all the information because once you know the distribution its
all, all this average thing, standard deviation those came first before even understanding this
means PDF or PMF associated PMF.

People started seeing okay if something is random can we measure some or guess something
random variability, so they started with mean initially then they could see that okay if there is
wider variability and smaller variability still the mean is same, they went to the second order
statistics or that variance or standard deviation will define them, okay So then they could see
okay there is some more things that we can extract but they could see that you go to the higher
order probably something else will be happening so every higher older is required.

But finally people could understand it is just that PDF or PMF which characterizes everything
because once you know this you can actually generate all the moments. So if some how you can
get the PDF you have everything. So all the statistics related to is already in your grip, so that is
why PDF is probably and thats the relationship between PDF and all those moments right but any
way so what we were discussing that if I have the PDF, I can go to the 1st older ,2nd order and all
those things.

So this was my definition of variance let see how it is related to those 1st and 2nd order moments.
So if just expand this I can write this is just the average value, so this can be x2 – 2x x bar + x
bar2 remember here # bar is the mean value that is not the random variable, so taking bar of that
will return you the same value right because its same, every time it will be same thing. So
whether you take the average or any other thing, you will be getting the same value.

So now this averaging suppose it has the associated PDF it is just an integration right, so
integration with addition and substraction that will be distributed, so I can write this as well as x2
bar – 2 this 2 is constant # bar is constant so that will come out from integration and it will be just
integration of # into P xx so that should be another x bar + this is constant that should be x bar2.
So finally what do I get -this is 2 x bar 2 + x bar2 so this is nothing but #2 bar - # bar 2 . So once
you have evaluated your mean and then through this process you have evaluated your second
moment, so just the difference of these two are your variance. Calculation of variance which was
of measurable quantity people used to do, they use to always take every value from mean how far
it deviates takes the square of that add all of them for all the samples divided by sample number.

517
This thing can now easily be calculated if you know the probability of them okay so that was our
target that we wanted to all these measurable parameter that people will used to do how do we
correlate that with our this moment definition right. so we could achieve that so this we can write
as x2 Pxx dx integration-infinity to + infinity minus # P xx dx integration-infinity whole square
of that so that is all immediately we get our variance and standard deviation is nothing but the sq.
root of that variance because just you do not want them to cancel it out, you have squared it. And
then you have taken the average of square but if I want to see the actual deviation so I need to
again do a square root of that. So that square root of variance is just a standard deviation thats by
definition okay so that is characterised by Sigma

(Refer Slide Time: 30:03)

σ= variance

= (X − X̄ )2

Often that is square root of variance or sq root of x2 bar – x bar whole square right so this is how
you get the sigma, this was the sigma which we are characterizing for Gaussian. So you

518
remember there were two parameters for the Gaussian distribution, one was the mean which was
x bar, so that m was actually x bar and you can verify that take that Gaussian distribution you
plug in x into PXX integrate it you will always be getting this m back where Gaussian was
defined by X-m whole square divided by 2 sigma square

And again do that variance calculation you will get back this sigma. So this is what happens, in
the next class what we will do, so far we have defined all these things in the next class we will try
to do is we will try to see some more relationship between these two particular part that variance
means, how they are related for a particular random experiment when they are independent, they
are not independent how do means, mean we have already seen whether they are independent, not
independent it is always addition of mean.

But for standard deviation what happens or variance what happens can they be evaluated
separately so that something we will try to characterize and then we will probably go into slowly
towards our new definition Which is correlation okay so that is something that we will try to do
thank you.

519
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-37: Probability theory (contd.)

Okay so we have so far characterized what we mean by mean and what we mean by variance or
standard deviation for mean we have already said that its if we have two random variable whether
they are dependent independent correlated uncorrelated we don't care it's mean is always addition
okay so separately you can calculate mean and you just add them you get the overall mean value
of that but suppose

(Refer Slide Time: 00:56)

520
z = x +y z̄ = x̄ + ȳ x, y independent
σ22 = (z − z̄ )2
= (x + y − x̄ − ȳ)2
= ((x − x̄) + (y − ȳ))2
= (x − x̄)2 + (y − ȳ)2 + 2(x − x̄)(y − ȳ)
= (x − x̄)2 + (y − ȳ)2 + 2(x y − x̄ y − x ȳ + x̄ ȳ)
= (x − x̄)2 + (y − ȳ)2 + 2(x¯y − x̄ ȳ)
= σx2 + σy2 + 0

I have some Z which is the summation of two random variables right X + y so for that we have
said that mean of Z should be this so that is something we have already proved now let us try to
see what happens to the variance here will insert or will means say something that x and y are
independent let us try to see what is the implication of that so I would like to calculate this Sigma
Z square which is the variance of Z, Z is a new random variable which is constructed by making
X + y where x and y are two independent random variable okay.

521
So Sigma Z square is nothing but we have already characterized that Z-Z bar whole square bar
we have already said that so this is the average of that random variable - mean square so now
replace z by this thing z by X + y and Z bar by X bar + y bar I can do that so this is X + y-X bar-y
bar whole square I can Club this as X-X bar + y-y bar right this can be written as X-X bar whole
square + y-y bar whole square + 2 into X-X bar into y-y bar now the overall averaging. We know
that it is just multiplication with PDF joint or marginal whatever it is and then integration so
integration gets distributed with respect to addition therefore this must be this bar this bar and this
overall bar right now let us see what is this is just Sigma Square by definition this is just Sigma Y
square by definition okay by definition of variance let us talk about these things so first of all we
can say this let us try to evaluate this X-X bar.

So this will be XY-X bar Y-XY bar + X bar Y bar again it will be distributed okay so this should
be X Y bar minus now let us see X bar is a constant so it will be just X bar Y bar this is also same
thing Y bar is constant so this should be X bar Y bar + X bar Y bar, for independence we have
already proven that if X Y are independent then X Y bar is equal to X bar into y bar so this must
be 0 once this is 0.

So if you just go back to our previous this one that Sigma Z square is Sigma X square + Sigma Y
Square and this is 0 that is a very fundamental result that if two events are independent remember
for mean independence was not required but for standard deviation or sorry variance the
independence is required and then only the variance of the addition of two random variable will
be addition of their variance right this only happens when these two random variables are
independent if not this factor will always be there and this will not be canceled out okay.

We have seen that this factor is 2 * X y bar-X bar Y bar so as long as they are not 0 this factor
will always be there okay so that is called means that is how you evaluate whenever two random
variables are independent okay now from there a new thing comes into picture what we do is
something like this that particular thing we have got X-X bar into y-y bar we define that as a
correlation between two variable x and y.

So we can immediately see if they are independent the Sigma X Y will be 0 we call that then X Y
are uncorrelated so independence means that they will be uncorrelated but not vice versa okay
because actually independence means little bit stronger condition it's just not independence
means X Y bar will be X bar Y bar that's alright but if that happens immediately Sigma X Y will
be 0okay whereas for independence you need something more that the PDF has to be distributed
the joint PDF P X Y X Y that must be distributed.

It should be p xx Into py y this must be happening which is a stronger condition if that happens
we know that there becomes uncorrelated but that does not mean un correlation is just this, this

522
must be equal to 0 it does not mean that XY must be means having a PDF which gets separated
out okay. So always do not get confused with the definition of Independence and
uncorrelatedness generally what we should say independence is a stronger condition as long as
two variables are independent I know that they are uncorrelated but the reverse is not always true
okay they might be uncorrelated still it might not be independent okay.

So this might happen because independence requires stronger condition so let us try to see what
this correlation captures it is almost if instead of X I put sorry instead of Y I put X it is actually
going back to variance definition because that happens to be X-X bar whole square bar okay so
basically your correlation takes you towards the variance if x and y are same okay so let us try to
understand what does this means what is actually correlation so let us try to do some means try to
list out some random experiment so let's say suppose the experiments are like this I in a day or in
means in a day what is the temperature that is one random variable for me okay so that I record
and also I take another random events which is the selling of soft drinks okay.

(Refer Slide Time: 06:49)

523
(x − x̄)(y − ȳ) = σxy
σxy ⪋ 0

So how much quantity of soft drinks has-been sold in a day let us say these are two random
events so temperature can be anything different day I see different values I will be getting same
thing will be happening with the soft drink means sold that number now let us try to plot them
what does that means we choose multiple number of days okay and this is the x axis and y axis in
x axis we will be plotting what is the temperature in a day and in y axis will be plotting what is
the amount of soft drink that has been sold.

So every this is called scatter diagram where every day will be put somewhere okay so what does
that means a particular point corresponds to what was the temperature on that day anyhow many
means what is the amount of soft drinks that has been sold so every point represents a particular
day it is a sample value of that day okay and sample value of this joint event I am actually
measuring two random number and try to plot it over there okay now plot these things it is all
fine now if these two events which we could see that as the temperature increases probably more
soft drinks or cold drinks will be sold.

So these two events are highly correlate one actually influences the others so what will happen in
the scatter diagram also you will see similar things whenever this is high this also must be high
whenever this is low this must be low so it will all be scattered in that region okay never it will
happen that this is higher and this is lower or this is lower and this is higher so it will not be
scattered over the entire thing this will never happen okay so if I now so it will all be scattered in
this fashion if you just write or draw y equal to X it will be around this okay and then if I just take
out from each of the samples the X bar so what will happen this will get shifted if I now plot X-X
bar and y-y bar.

So it will be plotted like this it will be clubbed around this once this kind of plot you are
observing then you can say in the scatter diagram that they are highly correlated similar thing
might happen when they are just negatively correlated so you could see a plot like this what will
happen whenever particular part is high the other part will be low it is it is just negatively
correlated that means if this is high then that should be low okay.

And if you have something which is completely uncorrelated the scatter diagram will show you
all over the place something everywhere it might be high with low, low with high both are same
so all kinds of things will be getting if they are uncorrelated if you take out this means bar X-X
bar and Y-y bar they will be all populated around zero everywhere and then immediately if you
take these numbers X-X bar into y-y bar they will actually tend to cancel out each other.

524
Because it scattered everywhere right they will not produce similar things here they will always
produce positive number here they will always produce negative number okay so if you add them
all probably they will produce a huge negative number here they will produce a huge positive
number that means they are hugely correlated either positively correlated or negatively correlated
okay whereas this one will give you almost 0.

So that is where the scatter diagram can help you to see that if two events are correlated or
uncorrelated that is pretty obvious are pretty clear which will be happening so basically this
Sigma X Y will tend towards zero if two events are completely unfolded that means this kind of
scatter diagram you observe okay and if that is not the case probably you'll see that they diverge
and they'll have a huge value okay and for positive correlation it should be Sigma X Y should be
generally greater than 0 and negative correlation must be less than zero and there is a technique to
normalize it so what generally people will do this Sigma X Y is divided by Sigma X into Sigma Y
that is called the Rho X Y just to normalize it what will happen whatever the value of Sigma X Y
this cannot be bigger than this multiplication right because that is the biggest that it can get so if
that is the case this will be always less than equal to 1 and greater than equal to -1 so at -1 it will
be highest negatively co means correlated fully correlated and at + 1 it is positively fully
correlated.

And at 0 they are completely uncorrelated ok so this is how it goes so Rho X Y if you plot at 0 its
uncorrelated this side the positive correlation increases this side the negative correlation increases
it ends at-1 ends at + 1 because you have already normalized it so you cannot get a value bigger
than 1 okay if it does not normalize it can be even infinity right so that's the concept of
correlation and therefore I can relate if suppose I have a random variable z which is X + y then I
can write now which was earlier done Sigma Z square that is nothing but Sigma X square +
Sigma Y square + 2Sigma X Y so that's the correlation okay.

So as long as this correlation is 0 this would be just Sigma X square + Sigma Y square otherwise
this correlation term has to be kept in okay so this is something you will see later on that if they
are not independent and not uncorrelated then what is the effect of that what happens in a random
variable where you add two random variables so we see most of the time why we are so much
bothered means so far in the previous class as well as this class we have been evaluating the
addition of random variable.

Why we are doing that let us first try to ask ourselves why this is so much required in a course
like analog communication that we are so much concerned about addition of random variable so
you know in signal what happens we have already talked about channel being linear right so if
there is noise and the signal being transmitted most of the time what will happen noise and signal
will be added with each other okay.

525
Now most of the time noise and signal are independent both independent and uncorrelated this is
the fact most of the time that happens sometimes there are correlation there are noise which
depends on what kind of signal has been transmitted like in optical that is what happens if you
transmit one the amount of noise will be higher if you transmit 0 the amount of noise will be
lower so if that happens there is a correlation also.

(Refer Slide Time: 13:40)

σxy
−1 ≤ ρxy = ≤1
∂x σy
Z=X+Y
σZ2 = σx2 + σy2 + 2σxy

526
Z =X+Y Y=Z −X
Fz(z) = P(Z ≤ z)
= P(X ≤ ∞, Y ≤ z − x)
∞ z−x

∫−∞ ∫−∞
= pXY (u, w)du dw

∞ z−x

∫−∞ [ ∫−∞ ]
= pXY (u, w)dw du

So but whatever it is when signal level you are adding because the channel is already assumed to
be linear so there will be just addition of these two signal so your outcome which is the output its
just addition of your regular signal + the noise if the regular signal for the receiver it can still be a
random thing right if the receiver already knows what is coming in a deterministic manner he
doesn't have to he doesn’t bother to receive that okay.

He wants to receive that because he does not know that there is uncertainty over that signal and
that is why he knows that there is some information in that so he wants to receive it so that is
unknown signal which is also a random thing for him and the noise which is being added at the
channel or at the receiver wherever it is it is also unknown to him it is random so these two
signals will often be added and that is why we are so much interested in addition of two random
variable.

So we are characterizing see you might be thinking that we are Just in the background we are just
doing mathematics but not necessarily we are choosing that part of mathematics which will be
interesting for us while characterizing a communication channel or means signal processing right
so that is why we are saying so much things about addition of two random variable right now
they we have already talked about when they are independents, so if noise and signals are
independent then of course whatever theory we have talked about independence that will be
applicable if they are not independent like we have told optical noise.

That is not independent of the signal so then of course the correlation has to be taken into account
while characterizing them so let us now try to see that suppose x and y these are two let's say
independent random variable what I knows something like this I know this for ✗ Whats the
Associated PDF now we are going to PDF so far we were just doing things with respect to
measurable quantity that's the variance and mean now let's go towards more concrete things
where everything can be characterized.

527
So the PDF. so suppose this x and y for both of them I know the individual PDF okay let us also
say that these two are independent so for independence only we are now proving this whatever
theorem will be proving okay now let us say my z which is a new random variable which is the
addition of these two random variable, knowing these two PDF is it possible to get the PDF
object so that will be our next target how do you get that okay.

So PDF generally comes from CDF we can always differentiate any CDF to get PDF so let's
characterize the CDF of z which is nothing but probability that this random variable z will be less
than equal to some defined Z over here okay so from here what do we know, we know that I can
say Y should be if this is the relationship I know that always Z is x + y then I can always say Y
must be Z-X I can characterize this.

So this means actually probability Z has to be less than Z means probability that my X which is a
see once I characterize X I can get something on Y right with respect to this z so Z less than Z
means actually X for any value of x which is less than infinity I must have a Y which is less than
equal to Z-X then z will be less than z this z right always I will be able to ensure this so this and
this these two are equivalent definition very carefully see this because what I am saying that I
need a probability that my Z random variable must be less than equal to z but I also know that my
x+ y is equal to z.

So now what I am trying to see because I know the joint means I know individual distribution
right so that is something I know so let us say this particular event is nothing but that I take any
event from X, x can take any value but I need to ensure that my X + y must be still less than X +
y will be that Z that Z must be still less than Z that means Y must have this condition whatever
the value of X it must be less than equal to Z-X then I know that whatever that associated things I
will be getting there the overall Z will be always less than this z if I in these two things because Z
is always being created by the addition of these two as long as I ensure this I will be having this
right so if this is the case I can always write this so this is again joint PDF from there I can
calculate this so what do I have to do for X I have no restriction I can go from -infinity to +
infinity but for y I have a restriction I can go from -infinity to this value.

so y should be going from -infinity to 2z - X into P X Y okay. •Inside let us say u W du dw right
so this is the joint distribution of X and y I still do not know but probably I do not have to we will
see that ok but according to my definition this should be equivalent right now let us try to do
something so I can write this from -infinity to + infinity ok so this is du and within that my
integration -infinity to z-X P XY u, w d w I have-not separated this integration remember its still
this integration inside this integration I am just writing it in this way okay.

528
So you can even write D u over here or you can just write D u over here fine now what I will try
to do is I will try to differentiate this because I wanted PDF of this right I will try to differentiate
this with respect to Z that must be giving me the PDF of Z so P Zz I know that is nothing but
DDZ of FZz which is nothing but d dZ of integration -infinity + infinity I have D u and inside I
have integration-infinity to Z-X PXY u w DW right this

now this integration limit has no Z. so I can take this into differentiation inside no problem in that
only inside of this I won't be able to take because limit already has Z so I can write it -infinity to
+ infinity D DZ now because the limit is Z-X I will put that as DD Z-X of course I can put a
chain rule so I can write, I can write this no problem in that and I will have this integration
infinity to Z-X PXY u w DW right this is just 1 ok what is this you have this conjugate of
integration and differentiation okay.

So this differentiation and this integration they will cancel each other okay because that's just
complementary so they will cancel each other and what do we finally get -infinity to + infinity I
will have this P XY or u w right this is fine seconds that should be with its limit sorry I should not
write this I must be hiding -infinity + infinity see whenever I do this cancellation what will
happen this will just be up to that function up to this value this integration and differentiation gets
cancelled but after cancelling the value that remains is that value right so I can write this as P XY
this u now becomes X and W Becomes up to this z- X and I will be having if it is X I will be
having X okay.

(Refer Slide Time: 24:05)

529
 d
pZ (z) = FZ (z)
dz
d ∞ z−x

dz ∫−∞ [ ∫−α ]
= PXY (u, x)d w du

∞ z−x

∫−∞ d(z − x) [ ∫−α ] dz

d d(z − x)
= PXY (u, x)d w du

∫−∞
= PXY (x, z − x)d x
∞

∫−∞
= PX (x)PY (x, z − x)d x

= PX (x) * PY (y)

So I did nothing instead of X I have taken a dummy variable DX okay and this automatically
goes up to Z-X so P XY that goes up to the z-X this gets cancelled out and it is at that functional

530
value right so that is what I get and I have after all these things I have this thing right now if this
two are independent I have already told that p x this x and y are independent so immediately I can
write -infinity + infinity this must be p xx into Py z-X DX right as long as they are independent
otherwise no what do I get, I get a convolution you can you can identify this function this is just
convolution of p x and trance because this variable is Z.

So it takes you to Z so basically that is a very important result that whenever you have addition of
two random variable if I know individual PDF and if I know these two random variable are
independent see in the proof steps all those things were used if they are independent I know that
they are that new random variable created through the addition is having a PDF which is just the
convolution of the Constituent PDF there is a very strong result as long as they are independents
in our case when signals and noise those are characterized by some PDF if I get an additional
signal which is the addition of these two as long as I know signal and noise are independent I can
always say it will be just a convolution of these two PDF right so that is something will be with
this with the help of this we will be able to characterize a PDF of the joint signal that so that is a
very strong tool which will be required later on so probably what has happened with all these
things will probably end my discussion of random variables.

So we have now got enough tool to characterize our signals and characterize our transmitted
signal as well as noise when they get added and all those things we will be able to do that but
now we will go to another step where this random variable will become a function of time which
is a very means obvious thing that happens in signal transmission.

So and that particular thing is called random process so now with the help of this random
variables and associated understanding we will try to build up our own understanding of random
process which will help us actually characterize noise as well as signals okay so then only we will
be able to talk about a communication which is contaminated by noise how do we characterize
that so from next class onwards we will start discussing about this random process and its
characteristics okay thank you.

531
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-38: Random Process

Okay, so we have already discussed about random variables right. so that is something we have
already done in a previous class. so we have told that will be probably starting discussing about
random process. and that will be in next few class probably will be occupying ourselves with the
discussion of random process.how it is different from random variable, why it is important? so all
those things and then how to manipulate a random process so that in our case where either the
signal is random or the noise is random we can actually analyze things, okay.

So will be in the process we will also see if a random process passes through a system, because
most of the time we will be passing signals through our system so how it interact with the system
also that will be something we will be discussing in next few classes, okay. so let us try to
understand the basic fundamental difference between random variable at random process. so I
will start with an example which is probably a widely accepted example to understand random
process, so what we do is something like this.

(Refer Slide Time: 01:35)

532
We take a resistor okay, and we put a ammeter against it okay , so no voltage source nothing, just
ammeter against a resistor? so what will happen it will try to measure the current through the
resistor, okay. so initially we will be wondering why there should not be any, I have not put any
voltage so there should not be any current.but what happens inside resistor this is just apiece of
wire probably, and inside that there are free electrons and all those free electrons or we should
say charged particle which are free to move, so what will happen they will be because they
mobile, so they will be also doing Brownian motion.

Okay so they will they will go here and there they will get collided, again deflected back in a
random fashion due to that Brownian motion there will be some net current which will be
recorded. okay and that net current is completely random, because you never knowhow that net
flow of overall electron density will be accounted, because it is all random motion. so what will
happen if you just see against that the current against that resistance.

533
So what we will see is something like this very low amplitude of current which is randomly
varying and this is this variation is over time. so you are recording over time and you just through
the ammeter whatever you are getting you are recording those values of current through that
resistor in time. so this is the nature you will be probably seeing and it will just keep continuing.
of course, the amplitude will be very low because it's just due to the Brownian motion-okay, so
this random amount of current if you. If you now start giving a voltage source across this, so
what will happen due to that voltage source there will be a current constant current okay, across
the resistance.so immediately you will see that there is a DC SH shift to this particular random
current and on top of that DC this DC shift will be due to this voltage source, and whatever the
values are resistance due to that whatever current you can calculate Ohm's law you put ammeter
also if you have something on that you take that resistance also> and you calculate overall
current.

You will see that current will be there but on top of that there will be a random variation almost
similar like that, okay. so that is due to that Brownian motion. okay, however small that
amplitude is of that current, there will be some amount of current which will be randomly
varying. so what we do now, suppose we prepare this same kind of resistance.so we try to, this is
a hypothetical experiment probably, we try to prepare say resistance multiple number and at the
same time instance.

We say in the experimental this one laboratory, so we say multiple students to actually measure
with a similar ammeter across that resistance, and at every time instance they should record this
for currents. right, so what we will see, we will see every student recording different kind of
current pattern. so if suppose this is for student 1) this is for student 2, and so on, and you will see
that similar kind of current pattern, but which will have different random variation which will be
recorded by different students-okay, So this is what will be happening, because it is though they
are identically prepared at the same time they are measuring because these are two different
resistance they are Brownian motion they are independent, so there will be, means differently
creating net current through the resistor, and you will be seeing something. and it is guaranteed
that because net means Brownian motion over a time if you see that should be 0. so overall
average current if you wish to see that we all zero. here also it will be 0, here also it will be 0, and
so on as many you can take everywhere the average current will be 0. but there will be a random
variation and that random variation reported by student 1 will be completely different from
student 2, and so on that is what will be happening.

534
now let us say we take this different signals, For a sufficient amount of time they have measured
it, they recorded it they have given it to you. and now I fix a time. right, let us say time T equal to
t1. okay, so at that time instants every student will be reporting some amount of current okay, so
take that value now these values which are reported if I have enough number of students let us
say almost infinite number of students if I have that, then these things at a particular time instants
that means I have almost taken a snap shot of all these signals that has been produced by the
resistor, identically prepared resistors by an experimented or measured current through it by
different students.

If I just take a snapshot at a particular time, we get different voltage reported. and then whatever
we get those will be random values at a particular instance so this is actually a random variable •
at a particular time instants if I take from different students different reading whatever they are
they are just like random variable and over there we can construct whatever we have constructed
so far that means we take those random variables we can actually try to produce what is a PDF
Associated PDF of that, we can also try to see what is the mean, what is the standard deviation,
variance, and all those things higher moments. We can produce all those things. okay, so we will
be getting that suppose I decide, okay no'T 1 I am NOT happy with, I will go for another time let
us say that is T 2, and again I do that, okay so I take all those samples again it is a random
variable at a different time instants, right! so there also I can again do all those things. okay so I
can I can get mean, standard deviation, variance, PDF, everything that I wish. okay, what might
surprise you is something like this you go from this time to this time you try to see the overall
PDF of the random variable that you have recorded, You will see that that will be same. that is
very surprising, because this has been experimentally demonstrated that happens okay.

why that happens, and what kind of class of signal those are will classify them later on, but this is
something which happens that means there is something underlying which tells us that a
particular signal you take, okay, which is random in nature. So that means, I can take different
kind of those signals which are identically prepared, and then I take samples at a particular time
instants, from all the sample function we should call, okay. so remember samples value and
sample function there is a difference between these two. so this particular function that entire
function in time domain that is called the sample function. whereas this at a particular time
instance whatever value I am recording that is actually sample value, and those values we are
saying that forms a random variable.

okay right, and then I can get a PDF, but what we are now seeing is something more interesting,
that the sample if you take, sample function if you take samples from all those sample functions,
and then try to evaluate the statistical property they remains almost the same. so that tells us that
may be something more has to be said about these things okay.

535
Sol in that/means with that observation only people started defining something else which is
called random process.

(Refer Slide Time: 10:04)

Which is I should say nothing but a random variable, but it is a function of time also. so
whenever we say XT, at a particular time instant this X, is just a random variable. okay, but this
random variable keeps on taking values at different instants of time.okay, and so whenever it
takes if I just go through the time and takes different values and I construct a function that
becomes one sample function of a random process from which it is drawn. so the associated
random process is just that experiment of putting identical resistor,And then measuring current
through it which is completely generated due to our inside randomness happening which is called
Brownian motion. okay, due to that whatever current is happening, for the entire time duration I if
I record the overall sample function, those sample functions are now random, because as you
have seen, one particular sample function is quite different from the other sample function. okay,
so we can even say that as if, like random variable, suppose head and tail so you have two
outcomes and whenever you do experiment, one of them are chosen. right, here also I can say
that as if there are different kind of sample signals and when you do a experiment one of them are
chosen. okay, which one will be chosen it completely depends.

536
okay, so I will give you one very simple example probably that will take out the doubt from you.
so let us say we have a 4-bit signal which is being transmitted. now I know that it's 4-bit each bit
of duration T. okay, so T might be anything.

It might be 1 second or it might be 1millisecond 1 microsecond whatever it is not bothered about

that. but it is just a 4-bit representation, that means four bit pattern which is completely binary it
can only take 1 or 0 value right but whether it will be a particular bit will be 0 or 1 that is
something I do not know. beforehand, means when we are at the receiver beforehand, I do not
know whether this is 1 or 0 it can be anything.

okay, now suppose the first bit I receive. I see that that is probably 1, so I can construct the first
bit that is 1. then I receive the second bit I see that that is also 1. then I receive the third bit it is 0,
and the fourth bit is 1. so immediately I can see, after receiving the whole sequence, I have a
pattern which is being received. but this is actually one of the random pattern that could have
been generated at the other side this is just one of them. this is the sample function that I have
received. but I could have received equally likely other sample function also like it might be 0 ,0
followed by 1 and then followed by 1, it could have been this sample function.

or there might be multiple such things how many of them will be there because each bit has two
options so it is 2 to the power 4, so there will be 16 options. just like you are rolling a die or
tossing a coin right. it has two options or this one, rolling a die has six options. but there we do
not have the notion of time.

It is just single thing that you observe. right, you do not have that notion of time. here we have
the notion of time, it is random, what is coming I am still not knowing. it is just coming to me
and after I detect the whole sequence probably I will know which sample function among those
16 possible things, has received by me. but whatever it is it has a randomness associated at every
bit, so that is what we are trying to characterize that every time instance probably here, we just
have four distinct time instants. okay, so there are many other but we do not have to specify them
because it remains the same. if it is 1it will be remaining same, as one for the entire bit duration
T. right.

so there are 4 time instants where it might change. okay, so it is a function of that time instance
where things might change, and that randomness is already there, if I would have talked about
just a random variable then I could have sampled it over here-and it is just 2 values 1 or 0.
whereas because I am talking about the whole signal that is why we have 4 sample value and
each of them characterizes a particular sample function okay.

So you have to see them in conjugate, and then only we will be getting an idea about what will be
the overall sample function.In digital it is little bit easier because we have accountable numbers

537
of options. right, so 4 bit we already have 16 options not more than that not less than that. this
many options are there. but if I just say it is a continuous signal like that signal generated by the
resistor. ok so it can take infinite options. so there might be associated with it there might be
infinite number of sample function which can take any value at any time instants. right.

So I can this way I can construct any other sample signal and all of them are basically a particular
sample function. any of them might happen. so this is something which is actually characterized
by a random process so whenever we talk about random process, associated with it a signal
comes into picture and that signal we talk about is a random signal. ok that will be any of those
possible signal that can be constructed so if you start talking about possibility we know that at
every time instance there is upper level and lower level that can be created by this particular
resistor. okay so of upper level and lower level of current.

If I know that among this any voltage level at this time instants. again next time instants which is
infinitely small distance away from that time instance you have considered. there also any
random value it can take. so you can now see there will be infinitely large number of possibilities
of sample signal that you can produce. because immediately over here only it can take infinite
amount of values and then the next one depending on what value it has taken it can take again
another infinite number of values and so on.

So like that, there will be some means there are infinite number of sample signal and probably
one of them will be coming to you-right, and it is not guaranteed that always that signal will be
coming to you so you experiment it today, probably one particular signal next day probably it will
be different-right, so like that. whatever it is we are saying that particular thing that signals are
represented by XT, because it is a it has a variation of time whereas at every time instant it is a
random variable.

So that is what random process is. now let us talk about characterization of a random process
right. so whenever we say characterization of a random process, we have already talked about
two things, one is you take a sample time. okay, so or you take a snapshot and then try to see the
statistical property of it. you will be immediately getting over all those sample functions some
values at that time, and you try to construct a random variable out of that and then try to see the
PDF of it. okay, and all other moments right.

So this is one way of characterizing it that is all good, but probably this is not all that we can do.
okay’ there is something more.

(Refer Slide Time: 18:19)

538
So I have a signal like this,I have another signal like this. now what I wish to also see, that at a
particular time instants. if I have this value in another time instants, whatever value I will be
getting, or whatever statistical property I will be getting, if I take the snap shot among all the
signals will that have dependency on the statistical property that I have observed earlier or later
time instants. so this is also very important that is there a statistical we have talked about
dependency correlation and all those things.

So now we are trying to see that, that in time I am taking snapshot and immediately I am
generating random variable. another time I take a snapshot immediately I generate another set of
random variable, means another random variable with different set of values. these two things are
they dependent, or are they not dependent.because that also comes out from the signal quality of
signal specification. and that characterizes the signal as well. because if there is a dependency I
need to know what kind of dependency.

okay, so it might happen that the signal even though it is random, at a particular time it is,
suppose it has picked one random number, but the next time instants, because of the low past
nature that means the signal does not vary too much. okay, low pass nature means what? within
some amount offtime it has only low frequency values high frequency values are not there, so

539
within a certain amount of time it cannot really vary beyond something.that is called low pass
nature, because the slope will be lesser than, means it has to be less than something right.

So if that is the case, immediately see.I have already spoken that within this it can take from this
value to this value, then I can specify that whenever I from this to this I go. and if I know what is
the maximum frequency it can take, immediately from here if it has already taken a value this
much over here, there will be a bound up to which the value can vary. this bound can only be
calculated If I know the dependency.

so that is why it is very important. that I also have some understanding about the dependency of a
sample signal with respect to the previous or previous to previous all those sample signal. so what
do I have to do that we have already talked about random variable and this is why we have talked
so much about random variable dependency, joint PDF, and all those things so right now you can
only immediately see that two things which might be dependent. okay.

So if I wish to characterize what do I have to do I need to go towards the joint PDF of these two
random variable so a random process I know that it is an infinite collection of random
variables.and there should be associated dependency among them. so if I need to characterize the
entire random process I need to actually capture all those dependency. or rather I should say'The
Associated joint PDF' of all of them. so that is where the next part comes where I will be
characterizing a random process.

(Refer Slide Time: 22:02)

540
FX (x1, x2, …, xn; t1, t2, …, tn) = P [X (t1) ≤ x1, X (t2) ≤ x2, …, X (tn) ≤ xn]
∂n
pX (x1, x2, …, xn; t1, t2, …, tn) = [FX (⋯)]
∂x1∂x2⋯∂xn

So what do I do, I wish to suppose X (t) is that random process so I wish to do a characterization.
that means CDF I will be calculating.so in CDF I will talk about X 1 X 2 X n what are these?
bracket, semicolon again T, t 2.. I will just explain what are this.T n. what does this means
actually it is actually the sample we were having we define a time T1. T 2 and so on up to TN, at
every instance we say the associated random variable is here X 1, here it is X 2, and so on up to
X n okay. so these are the associated random variable and what.'

We are trying to do is at those time instants we wish to see the joint CDF of those random
variables. I have told you that they might have dependency so I want to capture that full
dependency. so only way to capture that full dependency is get a joint PDF, or joint CDF among
them so this is what you will have to do and remember it has to be done overall sample time that
means this T will go towards infinity so every sample time you take try to capture all those
samples which will be the random variable.

And then take the joint CDF of all of them right so this will just specify that my probability that
X at T1 is less than equal to X1, comma X at T2 is less than equal to X 2. and X at T 2 3 is less
than equal to X 3 and so on.X at TN less than equal to xn so if I can characterize this entire

541
probability, then only I know that that entire random process I have captured it well.which is a
daunting task you can already see that if I have a signal, it is a continuous time means and then
immediately you have infinite number of time instants and at every time instants.

You need to actually get joint PDF that means or joint CDF that means that random variable at T
1must be less than equal to some specified value X 1, and so on that will give you the joint CDF
x1 x2 and xn right and if you just do a n th order differentiation you will get the CDF. so px this
x1 x2 up to xn that should be just an n th order differentiation with respect to all these variables.
right, DX 1 DX 2 and the CDF fx. right, so once I have this probably I know that I have
characterized the whole signal okay.

Which is already we have we could see that it is a daunting task. okay so any random process
characterization of that fully is really beyond our capability. most of the time. so therefore what
we will try to do is we will try to restrict ourselves, okay try to see something and try to give
some definition to it. so what we will start doing is something like this we will go to first order,
okay. so what do we mean by first order so we will talk about mean.

(Refer Slide Time: 25:54)

542
 ∞

∫−∞
X(t) = xpX (x; t) d x

Rx (t1, t2) = X (t1) X (t2)

∞ ∞

∫−∞ ∫−∞
= x1x2 px (x1, x2; t1, t2) d x1d x2

What do you mean by mean that is defined as this so what we are doing over here is we just take
one sample time that is T and then try to get the mean of it. so a particular sample time and just
calculate the mean so this is called the first order. immediately what will happen so I have two
mean means it is the value multiplied by the PDF. so X and in PDF because it is a one sample so I
do not need that Joint Distribution I only need one of them. right, so at time T so I will be
calculating this px X atT.X so if I can do that I will get mean. right, so this is the first order we'll
talk about okay.

543
So this just we are trying to get the mean or we are trying to characterize the PDF at a particular
point, so in time. so in time just take a snapshot, try to evaluate the mean or you try to evaluate
the PDF. once you get the PDF you will be getting the mean as well. okay next we will try talk
about the second order that means at least two time instants we will take> okay so that is called
the autocorrelation function. you might have already seen out of correlation function.

But that was a different kind that was in time we are just shifting it and multiplying the same
signal so it was the same signal.here it is not say you have multiple signals. okay, but we will try
to evaluate the autocorrelation function. what is that? we call that our Xt1 t2 so it's two time, so
therefore it must be at t1 I get the X.at T 2 I get the same, from that same random means sample
signal.I take that and I take the average, overall average.this is, remember there must be a joint
PDF and it is average so it should be a double integral. okay, so it must be - infinity to plus
infinity minus infinity - infinity I take suppose I pick value x1 at T 1 and X 2 at t2 and then I need
to get this joint distribution, which is x1 x2 at t1 and t2. so now I have to get at two time instants
what is the Joint Distribution among them. okay, and DX 1 DX 2. this is all that I will have to do.

so basically I take 2 or infinite number of sample signal at t1 at t2, I just try to get the
multiplication of these two. ok and get the joint PDF of them and then try to evaluate the overall
average value. okay, so which is the autocorrelation function. now from this autocorrelation
function you can see also, if I just make t1 equals T 2 it will just go back to the same time, and
then I will get the second order so that is why over here with the single this one. I have not
defined the second order or third order, because, that will be just a outcome of this thing. where
you just put t1 equals t 2. you will get back that. ok, so we have defined two, so on. we can just
keep on defining things, so what we will do, we will just get an idea of how to define them at
higher order, and then we will try to talk about the minimum requirement that we need to talk
about, so that the signal is almost characterized.so in the next class we will do that.

544
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-39: Random Process (Contd.)

Okay, so we have already started characterizing random process, right. We have said how to
evaluate mean and autocorrelation function, means how to characterize random process first of
all, and then why we need at different time instance and there means correlation that is
something we have already talked about. And then we have started talking about evaluation of
mean and autocorrelation function, right. So these two things we have already done.

(Refer Slide Time: 00:49)

545
pX (x1, x2, …, xn; t1, t2, …, tn)
∞

∫−∞
X(t) = xpX (x; t) d x

Now for the higher-order what we will have to do is definitely we will have to evaluate this ‘Px’
at let us say nth one so we have to do this joint distribution at different time instance. And then
multiply this with x1, x2, x3 up to xn and do nth order integration, and you will get the nth order
autocorrelation; Okay, or correlation I should say. Okay. So this is something we can keep on
doing for any value of n.

But now what will be interesting for us, let us try to see those signals; the signals that will be
interesting for us for which we can do some analysis. The signal that will be interesting, so
2

546
suppose a signal is going on, okay, if the statistical property of the signal basically depends on
which time I observe, then I have a signal which is very hard to characterize, because the signal
statistics is keep on varying it is not just the signal, signals keep on varying that is all true.

But now we are talking about the statistical property of the signal. If that also keeps on varying
and I have almost nothing to, means no hope to at least analyze that signal. So now we will be
talking about a class of signal where the signal might vary in time, but the statistical property
probably does not vary with time, which is, means defined as stationary process. So, we will now
try to define the stationary process. Okay.

So, this is something we will try to define. So, I have told, a stationary process will be or a
stationary random process will be, whose statistical property does not change over time. Okay.
What does that mean? That means all those PDF and joint PDF we are defining; they are not
really dependent on the time where I start. Okay. So, let us talk about that mean first, so mean
was; my mean XT bar was something like –∞ to +∞ x px xT dx, right.

So, this was my mean. Now this PDF if irrespective of any time T it is same; that means anytime
you sample, I have talked about that noise signal, I have told that probably the PDF at every
sample instants gives me the similar PDF. That was the first insight of a stationary process. Okay,
this is what happens generally in most of the signals that we study. Most of the time it will
probably be stationary. Okay; If there are no other influences, time varying influences in the
signal. Okay.

We will see that again we will differ from this and will say most of the realistic signal are not
stationary, because it comes through the channel and the channel keeps on varying over time. But
for the time being let us say that most of the signals that we observe are stationary. What does
that mean; that every instance you take associated PDF; that means like that experiment we are
saying that multiple students will be there will sample it and they will be recording the current,
will sample it at particular time, let us say 12 ‘O’ clock in the noon and we will say okay give me
whatever sample you have got; we plot the histogram get the PDF, normalize it get the PDF.
That’s fine one part. Then we say, okay 1 ‘O’ clock whatever you have observed give me that;
again, we do the histogram plot and normalize it get the PDF.

547
We will see that those two PDF will be identical; and any time we take that, probably all of them
will be identical. Okay. That is the time when we would say that mean is irrespective of time. It
will just X bar. Because if the PDF is same and the signal value, because the PDF is same, so
therefore it is mean, because PDF characterizes mean, right.

So, it’s mean also will be same, irrespective of whatever time you take. If the PDF is stationary
over time, the mean value also will be stationary overtime. It is not going to change. So that is
the insight we will get; that first order it is stationary, if at every time instance the PDF is that
single PDF of course we are just taking that at that time instance a sample and talking about this.

So, if this is happening then we can say first order it is stationary. Okay. Next the second order
comes. Now what we do, we actually take the joint PDF with two samples or two time instances.
Okay. And then again do the histogram plotting with normalization, try to see what is the joint
PDF. If the joint PDF again remains stationary for any two time instances we take, but remember
the time separation has to be same. Okay.

So, because I am taking a joint PDF and I have just said the statistical property should not be
dependent on the first time instance I take or the origin of my calculation. Okay. But that does
not mean any time I take because any time I take, the nature of dependency will be varying. Just
a simple low-pass filtering will tell you that. If you take it closer enough, the variation that it can
take, means suppose I have a signal it is a low pass signal.

(Refer Slide Time: 06:46)

548
pX (x1, x2, …, xn; t1, t2, …, tn) = pX (x1, x2, …, xn; t2 − t1, t3 − t1, …, tn − t1)

So, from here whatever value it has taken, if I take a very closer time instance, it will have a
lower variation, but if I go further away it will have a higher variation. Okay. So therefore, the
time separation between two points I am considering, that is very important, that has to be there,
that is the fundamental characteristics of the signal.

But where from I start, whether I take this and this or this and this, it should be independent of
that. So, if I can now say that the autocorrelation function which we have defined which is the
second-order thing, right. So, which is that Rx(t1, t2), it was earlier a function of t1 and t2. Now if
I can say that this is only dependent on the separation between them, not individual values okay.

So, if I can say this is just Rx(t2 - t1) it just depends on the separation of them, but not where the
‘t1’ is. So ‘t1’ can be anywhere, but if you keep the same separation you can put ‘t1’ at anywhere
what your statistical property will be getting, it remains the same. Then we can say at the second

549
order also, the function or whatever random process we are taking, its statistical property remains
stationary and so on.

If I go to the suppose nth order, then I should specify that Px is x1, x2 up to xn the way we have
specified in time t1, t2,…… tn must be equal to, now again the differences right, that should be
same. So therefore, I should be able to say that it should be x1, x2,…. xn at time I can specify let
us say (t2-t1), (t3-t1)…..(tn-t1) okay. So, wherever the t1 is just as long as t1 to t2, t1 to t3, t1 to t4 are
all same, I can shift the whole thing take the whole statistics, it will remain the same.

Then I can say at the nth order also it is stationary. Okay. So, first order, second order like that nth
order up to infinite order whatever order you can go, if you can say all of them are remaining
stationary, that means it does not depend on the origin, then the whole signal has the statistical
property which is stationary. And that is a very strong condition. We call those kind of signals as
stationary or those kind of associated process which is a combination of those signals associated,
that process is called a stationary random process.

So that is the definition of stationary random process, but as you have seen that stationarity of a
random process, proving that is going to be a daunting task. So, what we will do we will be
happy with just two order.

550
(Refer Slide Time: 09:52)

X(t) = x
Rx (t1, t2) = Rx (t2 − t1)

So, people are happy with to order. We just say that X bar T is just X bar does not depend on
time and Rx(t1, t2) is just dependent on the separation of them. If this happens, then in the first
order and second order they are stationary. All other higher orders we do not test, we just say it is
a wide sense stationary process.

551
Of course, it is a weaker condition or we say generally say weak stationary process. Okay, so that
means first and second order we have tested. It is just our limitation of computation. Okay, so we
have tested first and second order. We have seen that it is stationary. We do not go (means wish
to go) to higher order because that is too complicated to check. So, we just declare probably first
and second order it is already stationary; and you will see all the signal processing that will be
done will mostly dependent on this first order statistics. It does not go any higher. Okay.

They are important, but you will see for all our calculation probably, it is not going any higher
order up to any higher order statistics. So up to this if they are stationary, we can say that we can
just assume them to be stationary and we call them wide sense stationary. Most of the signals that
we will be dealing with will be (means we will assume) that they are wide sense stationary,
because later on see that if they are not wide sense stationary, there is no analysis.

Because if the statistical property also varies with time, I have no hope of any analysis if they are
stationary that means statistical properties are fixed, then I have some average sense of analysis
for these random signals as well; okay, which will be eventually used for our noise analysis and
any random reception analysis. Okay.

So far, I hope this is clear what we mean by stationary and wide sense stationary. Okay. Now we
will just try to give one example, so just an example where you will be able to see that a
particular signal means all these concepts, we have told that will be captured. Okay.

(Refer Slide Time: 12:22)

552
x(t) = A cos(ωct + θ )

So, let us say I have a signal which is actually A cos(ωct+θ). Okay, so this is a usual co-sinusoidal
signal with a phase frequency and amplitude. Now we are just adding some more things to make
it random. This ‘A’ is fixed. Amplitude does not vary with time. Okay, or it does not vary over
signals. This ωc is fixed, frequency is fixed, but this θ might vary and vary in a random fashion.

So, the θ has a PDF of, means it is uniformly distributed between 0 to 2π. So, because PDF has
to be integrated to one, so distance should be 1/2π right, so that is the PDF of θ, right. So, this is
something I know. So, what will be associated with that, there will be multiple random signals
now will be generated.

So, if I just try to see one possibility is it has a zero phase, right. So, if it is ‘0’, it will just up
your co-sinusoidal signal. The other possibility it has some phase so that with that lag the co-

553
sinusoidal will be starting, so there will be infinite such phases between 0 to 2π and all those
signals are those random signals that we were talking about.

Of course their amplitude is not changing, but it is the phase they are changing randomly. So all
those sample signals, so even when θ is probably ‘π/2’ then immediately it becomes sine, so it
will literally be something like this. So, this and this too are valid signals generated from the
same process, same random process, only when the θ varies. So θ is ‘0’ over here. θ is ‘π/2’ over
here and so on up to 2π you can.

So, it can take any of these, these are all they are actually one-one random signal. Any of them
might be chosen, because when I am transmitting, I do not know that θ, what that θ will be at the
receiver side. So, any of them will be chosen for me. Now what I wish to do is, I wish to see
whether this particular process is that stationary process or is that a wide sense stationary
process. So, this is something I would like to evaluate. So, I would like to see that θ with this
PDF of θ given will this be in time if I wish to see is this a stationary process. So, let us first
evaluate the average value of that right.

(Refer Slide Time: 15:34)

10

554
x(t) = A cos(ωC t + θ )
2π
1
∫0
= Acos (ωct + θ) dθ
2π
A 2π
2π ∫0
= Acos (ωct + θ) dθ

=0

So, let us say I wish to do this, that should be A cos(ωct+θ) average. Remember, whenever they
are talking about average, so far, we have been (in signal) we are talking about average in time,
right. This is not average in time; this is on that random variable. Okay. So, if I wish to do that
average, it is again a functional average of that θ.

11

555
We have done that already. So, what is that? That is actually this function multiplied PDF of that
random variable integrated over that range of that PDF, right. So, PDF is raised, I have seen
already it is raised over 0 to 2π right, so it should be integrated from 0 to 2π, A cos(ωct+θ) as a
functional average remember into PDF of θ which is 1/2π within this range dθ, so that should be
the evaluation.

It is not time averaged, remember that. We will come to time average later on. We will be
defining another subclass of process. So A is a constant 0 to 2π 1/2π goes out, so it is just
Acos(ωct+θ) dθ. So, if I just do this from 0 to 2π, you will see that this will be 0 because we are
integrating cos from 0 to 2π, right.

So θ varies from 0 to 2π. Any co-sinusoidal signal within a particular period if you integrate that
should be 0. So, this is 0. The good part that has happened is, I have started with any T.
Remember that T was there so it can be any T, but for any T, I could now prove that this is 0.
This will be 0 for any value of T, so that must be constant over time that is the good part.

So, it has a mean value of 0 at every time instance. So that means first order, I can say this is
stationary. This is the statistical property. Remember I am doing ‘n’ sample average, and the ‘n’
sample average is giving me a constant value right, for this signal it does not depend on time. So,
the average value of this signal does not depend on time, so that is the first step towards telling
that it might be a stationary process. Now we have to do for the second one. So for the second
one what we have to do is something like this.

(Refer Slide Time: 18:17)

12

556
Rx (t1, t2) = A cos(ωct1 + θ )A cos(ωct1 + θ )
2π
1
∫0
= A2 cos(ωct1 + θ )cos(ωct1 + θ ) dθ
2π
A 2 1 2π
2 2π ∫0
= cos(ωct1 + θ )cos(ωct1 + θ )dθ

2 2π ∫0 [ ]
A 2 1 2π
= cos ωc (t2 − t1) + cos (ωc (t1 + t2) + 2θ) dθ

A2
= cos ωc (t2 − t1) + 0
2
A2
= cos ωc (t2 − t1)
2

13

557
So, we have to evaluate Rx(t1, t2) so at value of t1 I will be taking a value of this particular signal
and t2 will be taking a value of that signal and then we have to do n sample average over the
variation of whatever random variation we have that θ. So this must be A cos(ωct1+θ) A
cos(ωct2+θ) average. This is ‘n’ sample average. What is the variation? It is just θ variation so
average again will be a functional average so over θ a random variable.

So this must be, A2 integration 0 to 2π PDF of that is 1/2π dθ and into this function which is
cos(ωct1+θ) cos(ωct2+θ), right. This is the one. I can take out, so A2/2 to 1/2π comes out 0 to 2π
so now this will be 2 cosA cosB so I can write that as cosωc this minus this θ gets cancelled, so
that will be t2-t1 and plus cosωc so this plus this so which should be t1+t2+2θ right. So, this
whole thing dθ. Alright so we are just doing the ‘n’ sample average nothing else. So A2/2 the first
term that is not dependent on θ, so therefore the first term if you integrate will be distributed of
course first time if you integrate just that thing comes out 0 to 2πdθ will be integrated so that will
be 2π this 2π, 2π gets canceled so I just get cosωc(t2-t1) right.

The second term, it will be this cos something which is not a θ variable, 2θ integrated over 0 to
2π again because it is cos and going through 1 period so it will just be cancelled whether it is
whatever it is it will be just cancelled out, cos positive and negative up will cancel each other so
this integration will always be 0. So that part will be 0 whatever happens. So now let us see what
has happened to my Rx(t1, t2). Is this separately dependent on t1, t2? No, it is just dependent on
the separation.

So immediately I can see it is actually just dependent on the separation nothing else. So, I can
now say in the second order also it is stationary, because my definition of stationarity was that in
the second order it has to be just dependent on the difference of it, nothing else. Okay. And so on
you can keep doing that, and you will be able to prove that it is stationary at every order. Okay,
but probably will be happy, content with this. You can test for higher order also.

But we will be happy with this because up to two orders we already have a strong condition. We
can say this is wide sense stationary already. So that particular function we have chosen,

14

558
fortunately even though θ has a random variability we could see us two means random process
which is stationary over time.

So, its autocorrelation function is dependent on just the separation of those two time instants
whatever you take, okay; and its mean is constant, it does not vary overtime. Okay. So
immediately we can declare, this is probably a random process which is stationary. Now we will
talk about something else which is called the Ergodic RP.

(Refer Slide Time: 22:27)

15

559
⏞ 1 T/2
T→α T ∫−T/2
x(t) = lim x(t)dt

Rx(τ) = x(t)x(t + τ)
T/2

T→∞ ∫−T/2
= lim x(t)x(t + τ)dt

So what do we mean by Ergodicity. Let’s now try to talk about that. We have already talked
about the stationarity. Now what we do, we will pick one of the sample. Instead of taking all of
them, we will pick randomly one of the sample. Any one of them is good. Okay, so there is no
preference for that. Any one of the sample we will pick, and we will try to do the same thing in
time, the averaging and second order statistics means statistics here it will not be statistics, it is
the second order autocorrelation function, the way we have defined auto correlation function,
okay, in the previous means previous few classes right.

So, let us try to do that. So first is the average, so we say that was ‘n’ sample average. This is
time average. So, time average is nothing but this, which is limit T tends to infinity because it is
a periodic signal, so if this is a periodic signal then I can just talk about time average as if I take
it for a big period do the average and make the period stretched to infinity. So, 1/T – T/2, + T/2
x(t)dt.

So this is the time average, so over the entire time duration you pick all the samples and then
means you cannot pick samples here because it is continuous, so you integrate it and divide by T
because for T time duration you have integrated. If T is stretched to infinity then you get for the
entire sample or entire signal. So that is the average. Remember what we are doing for the ‘n’
sample, we are taking a single time and we are taking multiple samples we are doing ‘n’ sample
average. Everywhere we are picking our value and we are doing average of them.

So that is called the ‘n’ sample average. Here what we are doing, among all those sample signal,
we have picked one of the signals. No preference for it. Any signal will do. Take that signal and
do overtime averaging for that entire signal. You do not consider any other signals, just that
signal do a time average. You get this, okay and the autocorrelation the way we have defined

16

560
autocorrelation so Rx(τ) will be defined as x(t) x(t + τ) and then we do time average, which is
similar.

Limit T tends to ∞ 1/T integration –T/2 to +T/2 x(t) x(t+τ) this is how we defined autocorrelation
function in time domain, right. So, if I evaluate this one, then I get the autocorrelation. Now if
for me or for a particular sample space, if this happens that any sample, I pick I evaluate these
two things and they are exactly equal as the one we have got for the stationary process, if this
happens, then the associated random process is called Ergodic random process. Okay.

So, Ergodicity means that whether you do it in ‘n’ sample or you pick a sample (any sample it
will do) and do it in time, both the things will give you same result. Okay. So as long as that is
happening, then we can say that this process underlying process is probably Ergodic process.
Okay. What I would request you means I won’t do that because it is a very simple one (simple
exercise), I would request you to just test the sample we have taken.

Take any θ, okay, Acos(ωct+θ). Here a particular sample means any θ you can pick, out of those
random 0 to 2π, any value you can pick any of them, it does not have any preference, that also
will be clear to you. So, pick any of them either 0 or π or 2π or π/2 or π/4 whichever value you
pick, we will be able to show that the time average if you calculate, because here the time
average will be just if you do it within a single period, because it will be just repeated.

So, you do not have to do it for the entire thing because it is a periodic signal. So, you just have
to do it over a period and over a period if you just do time average as well as autocorrelation
function, they will be same as what you have calculated. The same value will be coming out.
Okay, so this is something we will be able to prove and then you can show that the signal we
have picked that is why we have picked it hat signal that is Ergodic signal as well, okay or we
should say wide sense Ergodic, because we are not proven for higher order statistics. We are not
going to order 3 or order 4 and so on okay because that would be too. So, at least up to second
order you can prove that it is stationary and then finally you will be able to prove that it is also
Ergodic, that means any sample any θ you put pick, it will give you same result.

17

561
So that is the strong condition for most of the means considered signal probably we’ll see and
many things can be proven for these kind of signals. Either it has to be stationary or it has to be
even more stringent that it is Ergodic. Okay, so I will just before ending this class I will just give
you one very simple example of Ergodicity. Probably that will give you some idea of Ergodicity.
This is again an example picked from the book we are following. So, suppose we have a
Manhattan Street kind of configuration. Okay.

(Refer Slide Time: 28:21)

And you have all those southeast and northwest, sorry north-south and east-west, these streets are
exactly perpendicular to each other. Okay, so every street has crossing. Okay. Now we say that
every crossing whatever signal that are there, with 75% chance that that can be green and 25%
chance is there that can be red. Okay. Now let us try to do this experiment, that a particular car
going from this to this over a particular street going from south to north, okay it will be crossing
multiple these things, multiple signals.

18

562
And then what he has to see is and we also assume that all these signals are independent of each
other. They are not correlated that if this is green that will be also green with this probability and
all those things they are completely independent of each other. So if I just see one driver who is
going from north to south and crossing multiple such signals, okay and then try to evaluate how
many times he will be or how out of all those signals he has encountered, how many signals he
will be observing as red and how many signals he will be observing as green.

That should be exactly from our definition we know that 75% of the chance or 75% of the
signals he has encountered probably will be green and 25% of them will be red so I can
immediately see from that drivers point of view, and that driver can be any driver. Now let us say
multiple drivers are flocking around. Okay, so they are actually coming to a particular signal.
How many of them will actually report that I have seen green in that particular one and how
many of them will say that I have seen red in that? because it is 75%, and because the drivers are
coming at any time, okay, so because it is 75% times it is remaining green and 25% times it is
remaining red, so probably you will see that out of all those drivers, 75% will report that at that
particular signal point I have seen green and 25% of them will report that I have seen red. So
what is happening, out of all those drivers one of them you picked, and you actually went
through time, because whenever he is driving at different time different signal he is observing
and he is reporting.

Whatever he is reporting, that PDF is exactly true for if you just sample it over a particular signal
as if there are single in our definition single time instants, okay. So, single signal is single time
instants and multiple signals are different time instants. So a particular signal if it is a particular
driver going through it that means it is over time going through and whatever average value he is
observing, that is becoming same as if you take more number of drivers in a particular signal
point.

So this particular scenario is actually an Ergodic process because for any driver you pick, they
will report the same thing. Okay, so this is just to demonstrate which are the process that can be
Ergodic and what is the meaning of Ergodicity, right. So with that probably we will end this
particular class and then we will try to show you if a signal is stationary what can we do with
that? What is the advantage we get if a stationary signal is there? Okay, in the next class. Thank
you.
19

563
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 40: Random Process (Contd.)

Okay, so we have talked about stationarity and ergodicity, right. What we’ll now show you is
(something for means) some class of signal that is probably wide sense stationary signal, we’ll
show you (demonstrate you) that some more property of that signal is still predictable. Okay, so
that is something we will try to appreciate today. So, what we mean by that?

Let us say for any signal, all those deterministic signals so far what we have done, we are trying
to see what is happening at a frequency domain, right. So that was very interesting for us because
with that we were means employing filtering and all other things which frequency component it
has, how to enhance the signal, all those things, lots of processing and all those were possible is
we visualize the signal in the frequency domain.

And whenever we said that frequency domain we want to visualize, I need to do a Fourier
transform. So, for Fourier transform the input has to be deterministic signal. So far this was all
good because I was always dealing with deterministic signal, either a pulse or a sinusoidal or
something of that nature, but now we have already declare that most of the signals are not
deterministic because suppose receiver want to design a filter, he does not know what is coming,
he has no idea, it will be a random either a random pulse or a random signal amplitude
modulated, how the envelop is varying, he has no idea. On top of that there will be noise.

564
There is no idea about all these things. Okay, because before he receives how can he know what
is there in that signal right, so it is random to him, but he wants to design a filter for this, but
filter requires frequency domain characteristics but it is a random signal. So now what do we do
for this? How do we characterize for a signal what will be the spectrum quality of it? but we
know that whenever we wish to go to spectrum, we have already devised a strong tool which is
called Fourier transform. Right now, because a signal is random, we have no way to characterize
it, so it is very difficult we cannot do that.

So, let us try to see if we can do something. Whenever we are having a random thing, generally
how do we measure that or characterize that? It is in average sense, so whenever we have
different kinds of height in a class, if we wish to characterize the height of the class, we say
average height of the class. So similar thing we will try to do over here, so because it is random
process so that means all the signal that will be coming that is random, so we will try to
characterize it from the average perspective. Okay.

So, we will give some average sense of it. What does that means? So, we assume that these
signals are power signal, so therefore I will be interested in power spectral density right. So,
what was power spectral density? We have already derived this formula.

(Refer Slide Time: 03:28)

565
 2
xT (t)
Sx( f ) = lim xT (t) = x(t)π (t /T )
T→α T
2
xT (t)
Sx( f ) = lim
T→α T
∞

∫−∞
xT ( f ) = xT (t)e −j2πft dt

T/2

∫−T/2
= xT (t)e −j2πft dt

So we have said that power spectral density is something like this right, so this is something we
have told that if I have a corresponding x(t) which is truncated, so I create a x(t) where it is
nothing but x(t) into a box function of duration ‘T’ right, so it is multiplied by a box function
3

566
where it is defined from – t to + t and rest of the things are 0. So it is just basically truncated. So
this I can get a corresponding Fourier transform, as long as x(t) is known to me. Okay.

So, if x(t) deterministic one then xT(t) also will be also a deterministic one, and then what I can
do? I can take a corresponding XT(f), okay. Fourier transform I can do, so it is this Fourier
transform then divided by T, T tends to ∞, that was the definition of my SX(f) but now my x(t) is
a random process, so what I have to do is (I need to) if I wish to calculate this power spectral
density, that must be the average power spectral density.

So, therefore that average power spectral density should be, I should take average of this, so |
XT(f)|2/T, average of that. So, all I have to do is, I have to do the ‘n’ sample average of this whole
thing, okay. So that should characterize in average sense what will be the power spectral density
and I need to see whether this is something I can characterize. So that is something will try to do
in this class, okay.

So, let us say I have a particular signal x(t) whose Fourier transform is this means XT(f), so
therefore I can write the corresponding signal xT(t) it should be e–j2πftdt, –∞ to +∞ right. I can
write this, because this is just the Fourier transform of the truncated signal, as long as I have
picked a sample, okay. So I am ‘n’ sample or ‘n’ sample averaging I will do later, so this is just a
picked sample and then because x(t) is truncated already I can write that this is –T/2 to +T/2
x(t)e–j 2πftdt I can write this, because it is truncated as long as I take the limit from –T/2 to +T/2
over that ‘T’ duration, whether I write xT(t) or x(t) it is the same, because within that truncation it
is the same thing it is just beyond this it is 0 for xT(t) and it is having some value for x(t) because
I am not going beyond that so I can write this; and this and this are equivalent because beyond
+t/2 this is ‘0’ so there will be no value, so this and this are equivalent, I can write this way, okay.

(Refer Slide Time: 06:40)

567
 T/2

∫−T/2
XT (−f ) = x(t)e j2πft dt

| XT ( f ) |2 = XT (−f )XT (−f )

T/2 T/2

∫−T/2 ( 1) ∫−T/2 ( 2)
+j2πft1
= x t e dt1 x t e j2πft2 dt2

| XT ( f ) |2 1 T/2 T/2
T→∞ T ∫−T/2 ∫−T/2
lim = lim x(t1)x(t2)e −j2πf (t2−t1)dt1 dt2
T→∞ T
1 T/2 T/2
T→∞ T ∫−T/2 ∫−T/2
= lim x(t1)x(t2)e −j2πf (t2−t1)dt1 dt2

Similarly I can also define XT(-f), right, so which will be nothing but –T/2 to +T/2 x(t) this will
be just +e+j2πft because it is –f, so – of – will be plus dt, okay. Now what is |XT(f)|2 that I write as
XT(f)XT(-f) why? Because we know that XT(-f) is complex conjugate of XT(f), this is something

568
we have proven already, so if I just multiply these two that must be giving me |XT(f)|2 for a real
signal of course, so I can write these two.

And because there are two t integrations, I will take two dummy variables t1 and t2 right, so we
just do it –T/2 to +T/2 x(t1)ej2πft1 so I am just first writing XT(-f) dt1, –t/2 to +t/2 x(t2)e–j2πft2dt2 so
this is this one, this is this one, dt1 and dt2. This is just two dummy variables I have taken. It was
intergraded over t because both of them are t, so because I will be taking double integral, means I
will be representing them as double integral so I do not want to confuse those two random means
those two variables.

So because this and this are independent, so I take that whole integral out so from –T/2 to +T/2,
–T/2 to +T/2 x(t1)x(t2)e-j2πf(t2-t1)dt1dt2 right. I can write this. Now I wish to get that XT(f) average
right, which will be nothing but if I just take divide by T, take an average and take limit T tends
to ∞ if I just do that. So, what will happen to this side, it should be average and there should be
1/T limit T tends to ∞, so I am just modifying it, right.

So that should be the overall thing. Now this ‘n’ sample average is over these values right, these
are the random variable that is one integral okay, and that has nothing to do with this t1 and t2
integral, so therefore I can actually exchange these two integrals. ‘n’ sample will be also I
integration over this random variable. Okay, that x(t1) and x(t2) okay, which has nothing to do
with this other random variable, sorry other variable of integration, okay.

So therefore, I can take the ‘n’ sample average inside, no problem in that. It is just and the limits
are not dependent, so I can take them inside, so I can just write this similarly T tends to ∞ this 1/
T remains, this remains, so it is x(t1)x(t2) average e-j2πf(t2–t1)dt1dt2 right. Now comes the
stationarity property, what is this? That is actually the autocorrelation function in ‘n’ sample
domain right, so if the process is at least wide sense stationary, that means the second order
should be just the autocorrelation function which depends on the separation of these two, so I can
write this.

569
(Refer Slide Time: 10:58)

1 T/2 T/2
∫ ∫
Sx( f ) = lim Rx(t2 − t1)e −j2πf (t2−t1)dt1 dt2
T→∞ T −T/2 −T/2

1 T/2 T/2
T→∞ T ∫−T/2 ∫−T/2
= lim ϕ(t2 − t1)dt1 dt2

So therefore SX(f), I can write as limit T tends to ∞ 1/T –T/2 to +T/2, another integration –T/2 to
+T/2 this I can write as RX(t2–t1)e-j2πf(t2 – t1)dt1dt2 right. I can write this, so this whole thing now as
long as they are wide sense stationary this whole thing no becomes a variable of t2–t1 so I can
defined that as φ(t2–t1) okay, so therefore the whole process becomes limit T tends to ∞ 1/T –T/2
to +T/2 again –T/2 to +T/2 φ(t2–t1) dt1dt2 right, so this is what we have got so far.

Now we will do a trick, so from this double integral we want to convert it to a single integral,
that something we wish to do. Why we are doing? Because we want to actually this SX(f) we
7

570
want to relate it to the autocorrelation function okay, so we have related it somewhat but we need
a little bit more concrete nature of it, okay. So let us try to see if we can further simplify it.

(Refer Slide Time: 12:25)

( cos θ )
T−τ
area of strip = cos θΔτ

So what we wish to do, this t2–t1 that should be a variable. So, we define τ as t2–t1. Okay, let us
see how the integration variable changes so this was my t1 and this is my t2. The overall
integration if we just see this, this is happening over this box of dt1dt2 where t1 is varying from –
T/2 to +T/2 and t2 is varying from –T/2 to +T/2 so it will be a box of size T x T okay, so which is
something like this.

571
Where this is +T/2, this is +T/2, this is –T/2 and this is –T/2 right, so now we need to also define
in this plane, this particular variable which is τ=t2–t1, this is nothing but the straight line of 450
slope, okay because t2 is exactly t1+τ, where m is ‘1’, ‘1’ means tan θ should be ‘1’, so that
means it has a 450 slope and it cuts at, that is the cutting point at τ, if I take the next line and I
increase τ by this becomes τ+Δτ, so this must be the equation t2–t1=τ+Δτ so that is the equation
of this line.

And this line will have equation t2–t1=τ okay, so now we wish to convert the integration over this
τ how do I do that? So instead of integrating it over suppose at t1 and t2 I have a box right of dt1
and dt2, instead of that box if I just do this integration over this strips, okay and then what will
happen? If I just this τ if I start varying it so at τ = t it will just be crossing it that is very clear
because this is T/2, –T/2, –T/2.

So, at τ = t it will just be touching that, okay. So, at τ = T to τ = –T, if I just keep changing τ, it
will just switch through the entire thing. All I have to do is because of that whatever strips will be
happening I need to take the area of that, over that area I have to do the integration. So that is all
I have to do, and if this dτ is sufficiently small, why I am doing this because I had a inside
variable which is φ(t2–t1) which is already φ(τ), so if Δτ is sufficiently small over that strip φ(τ)
will not be changing.

So that is why I can do this integration, okay just the area dt1dt2 has to be changed to this
particular area. As long as I am doing that, it should be converted to a single integral right, so I
have to then evaluate the area, okay because I have already seen the limit of τ that goes from –T
to +T okay, so I just have to get the area, so for calculating that area I need to find out this length,
okay because the angle I know already so immediately I will be getting this value, okay so this is
something I can always do.

So basically if I just draw it little bit bigger, so let us say this is actually t2– t1=τ, so what is this
point? At this point by t2 is T/2 so t2 is T/2. What will be t1? That must be this, okay. So t1 is
actually t2–τ and t2 is T/2, so T/2–τ so that is this value. I need this, so that must be because it is
9

572
already in negative, so it must be this value – this value that will become negative so they will
give me the distance.

So this value is already t2–τ, so if I wish to get this that must be T/2–τ– this value, what is the
value over here, –T/2, so –(–T/2) that gives me (t–τ), so this particular value, sorry this particular
value is (t–τ), if this is θ and what should be this? this should be (t – τ)/cos θ, right, so I have got
one arm of this. I need to also get another arm of this, sorry this particular thing, these triangles
will be very small and it can be cancelled out okay so that is that will be in consider because that
will be Δτ2, so I can neglect that as long as Δτ is small.

So, I just need to get this box. Okay, so if now you can see basically this is Δτ right, because over
here that was the increment τ to τ+Δτ so this must be Δτ, what is this? This is already θ, that
particular small angle between this line and this line that is θ. Okay, so immediately that should
be Δτ cos θ, so this multiplied by Δτ cos θ must be giving me this whole area this strip area, cos θ
and cos θ gets cancelled so that must be (t–τ) Δτ, that is the actually the strip area and whenever
we do integration, so at every value of this τ this area has to be considered, okay.

So therefore, the integration should be; okay I should also specify another thing that this was true
when I was in the positive half of τ. If I go to the negative half of τ, you will be able to similarly
prove that this should be (t+τ) Δτ. This is something can be easily proven similar way, okay. So
basically, I can write that this should be (t–|τ|)Δτ right, I can write this area as this. Once this is
the case, now I go back to my calculation.

(Refer Slide Time: 19:48)

10

573
 1 T/2 T/2
T→∞ T ∫−T/2 ∫−T/2
Sx( f ) = lim ϕ(t2 − t1)dt1 dt2

1 T
T→∞ T ∫−T
= lim ϕ(τ)(T − | τ | )dτ

{ T }
∞
|τ |
∫−∞
= ϕ(τ) 1 − dτ
∞ ∞

∫−∞ ∫−∞
= ϕ(τ)dτ = Rx(τ)e −j2πfτ dτ

So SX(f) was actually limit I will write it one more time 1/T, –T/2 to +T/2, again –T/2 to +T/2
φ(t2–t1)dt1dt2. This integration now can be converted, limit T tends to ∞ 1/T, this becomes a
single integral now over τ. So, the area for that integration is (t–|τ|)Δτ that becomes dτ, and this is
already φ(τ) integration and the limits goes from –T to +T that is what we have seen, so –T to +T
right.
11

574
So, this is the overall integration. Limit T tends to ∞, I take the T inside because that has nothing
to do with the integration, so if I just take that what will happen? Because it goes to ∞ so I can
even write that to be ∞, so basically I can take this limit. I might write this to –∞ to +∞ φ(τ), and
now this will be (1–|τ|/T), of course here that limit T tends to ∞ should be there, okay, dτ.

Now this is something where we have to specify that this as T goes to ∞, probably this might be
canceled out, but we should be very careful about that we have to say this thing that as long as
this particular integration φ(τ)|τ|dτ is bounded. Most of the signal that we’ll be considering this
will be bounded. Once this is bounded, divided by T okay and T tends to ∞ that goes to 0. So
therefore, this term gets cancelled, I will be left with –∞ to +∞, φ(τ)dτ, which is nothing but – ∞
to + ∞.

Now replace φ(τ), it was actually RX(τ) e-j 2πf τdτ. Can you identify this form? This is nothing but
the Fourier transform of RX(τ). That is a fundamental theorem of random process that whenever
we have a random process which is characterized as either wide sense stationary or stationary, I
will be having a RX(τ) calculated or (t2–t1) calculated. If I know that, I can always relate means
do a Fourier transform of that.

(Refer Slide Time: 22:37)

12

575
SX ( f ) ⇔ RX (τ)

So that means RX(τ), if I just do Fourier transform, I get back my SX(f) which is defined in
average way. So, this is the average power spectral density, because we told for a random thing
random signal, I can never calculate the actual power spectral density, because every signal will
have its own power spectral density, but what we can characterize is that average power spectral
density. And we have got now a very strong formula where it is characterized that, if I know that
signal is stationary it would not have happened if the signal was not stationary.

Because then the whole process will not be proven. If the signal is stationary or at least wide
sense stationary, that means up to second order it is stationary, if that happens then the
autocorrelation function if I can somehow evaluate, you just take a Fourier transform you get the
average power spectral density. And once you get the average power spectral density, now you
can even though that was a random process you have the average power spectral density, with
that you can design the system. Okay, that is the advantage you get.

13

576
Otherwise you could not have done this. So, this is the famous Wiener Khintchine theorem. That
is the WK theorem we all know about. So, this is that fundamental property of a random process.
It says that as long as and that is why we were saying so much of things about our stationary or
wide sense stationary random process because as long as random process is stationary or wide
sense stationary, we know that we can always derive its average power spectral density which
will be used for characterize in signal, okay.

So that means if I have a random signal, I should not now worry. I just have to test whether it is
at least wide sense stationary, that means second order statistics is still stationary as long as I
know that I know I still have a means to go to the spectrum. Earlier I was little bit afraid that if
the signal is random, I have no way to go to the spectrum I do not know how to exploit the
spectral property of the system of the signals, okay. But now with this strong relationship I know
that as long as I can evaluate somehow the autocorrelation function, I will be able to characterize
the corresponding average spectrum, okay.

So, this is a fundamental theorem and this will be used. Now let us after knowing this, let us talk
about something which is very fundament al to noise analysis, let us try to see what this noise.
So, noise is something which is completely random to us. That means if I absorb something at
the next time, I do not know what will be happening. That is why we call that noise. Okay,
because that is completely random and therefore in the next time instance even though it is
infinitesimally small, just at ‘t’ I observe some sample value even at t+Δt let’s say over here,
even if Δt tends to 0, I have no way to say where that will be. That can take any random value.

This particular process we characterize as noise, okay, so that is to our mind that is probably the
noise, okay, we will see that as a special characteristics but if that is the noise, then what is the let
us try to ask because if this is noise this is the random process. I need to first try to understand
what will be the corresponding autocorrelation function, okay. I can immediately see because at
this time seeing something, I am not able to at the next time what will be happening I am not
able to tell that means it is completely uncorrelated.

14

577
So time we just progress little bit I must not any correlation, so the correlation value should be 0,
so what I can say but signal with itself has full correlation will be very high correlation because
that I know exactly that signal. Only thing is that I just shift the time even a little bit, I know that
there will be no correlation absolutely 0 correlation. So, if this is the case when the
autocorrelation function should ideally look like a δ function, right. So that is noise fouls.

Whenever we characterize noise, we say it does not have any correlation to next time instants,
and immediately single function comes to our mind that is the δ function which some strength of
that δ, let us say that is ε, okay. So, with that in mind immediately what will be the corresponding
Fourier transform that should be the power spectral density, we know δ Fourier transform is dc,
so this is the power strength of it, so this is actually for noise this is the power spectral density,
what does this say?

That it has equivalent component at every frequency or equivalent amplitude of power at every
frequency and that is why probably we say it is wide noise, that means all equivalently all
frequency components are present and it will be wide noise, because it is the most random noise
that we know which has no correlation what so ever even if I know some (means observe some)
sample value over here at a particular instance of time, the next time instance I cannot predict
anything about it.

So whenever the signal is like that, I have a auto correlation function which is δ, so immediately
you can see why this particular thing is so important is that I have a sense of creating
autocorrelation function from the description of that noise that I have in mind, the most random
noise that I know and then immediately from there I know the spectral property of that noise
because of Wiener Khinchine theorem, so that is the strength of Wiener Khinchine theorem.

So, once I know that random process even though the signal can be anything, see the signal can
be of any nature I do not even know, even though the signal is like this, I have a very nice
spectral property of it. That is the fundamental process over here. So immediately I know the
noise characteristics and this is actually called the white noise, okay.

15

578
 So once we have that spectral property of noise, now we can actually start dealing with this
noise that what this noise will do to my signal and how I should restrict this noise, so
immediately you know that the filtering is the big thing in noise (means combating noise)
whenever your signal is, just filter that portion out because rest of the noise will be immediately
suppressed. If you do not apply filtering see even if you do not have any other signal, still you
are actually taking lot of noise to the receiver and that we will contaminate your signal.

Whereas if you just suppress it, a lot of power signal power we’ll be relate a signal power to this
autocorrelation as well as the spectral property of it, but you will see that lot of extra power noise
power which is contaminating the signal are unnecessarily coming to your receiver, so that is
why filter has a big role whenever you are combating noise, so what we’ll do, next we will try to
appreciate all these things to see how noise analysis can be dealt with in a communication system
or particularly analog communication system okay. Thank you.

16

579
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-41: Random Process (Contd.)

Okay, so in last few classes probably we have started discussing about random process and in the
very last class I think we have proven one of the most important property of a random process
which will be heavily used for our case of noise analysis, so that is called the Wiener khinchin
theorem right. So that is something we have already proven. we have said that if a process is at
least wide-sense stationary and then what we can do is we can always evaluate the
autocorrelation function that is the ensemble autocorrelation the way we have defined and then
we know that that if it is wide-sense stationary then ensemble autocorrelation function should be
just depending on the separation of time instants where the samples are picked.

And if you do a Fourier transform of that autocorrelation function then you get the average power
spectral density, so this is this is one of the most fundamental theorem that we have proven which
will be heavily used because that is how we can actually take any random signal and then try to
guess what will be the spectral quality of that signal right which spectrum components are present
which are not present.

So although spectral analysis or visualizing the signal from the frequency perspective that will be
useful for us okay so let us try to define on that same line some more property regarding this
autocorrelation function, so let us say I have a wide sense stationary process XT okay.

(Refer Slide Time: 01:59)

580
 Rx(τ) = x(t)x(t + τ)
Rx(−τ) = x(t)x(t − τ)
= x(σ + τ)x(σ) t−τ =σ
= Rx(τ)

Rx(0) = x(t)x(t) = x 2(t) = x 2

∞

∫−∞
2π f τ
Rx(τ) = Sx( f )e j df
∞

∫−∞
Rx(0) = Sx( f )d f = x 2

So whenever we are mentioning we are always mentioning that it is at least wide sense stationary
whether it is stationary or not because we had to go for a higher order so that is something we
have seen, so at least wide-sense stationary has to be mentioned so this is the process which has a

581
autocorrelation function once it is wide-sense stationary then we know that autocorrelation
function just depends on the time difference between the samples we are picking.

So it is RX tau which we have defined as X T XT + tau and the ensemble average over that right,
so that was our definition immediately if we try to guess what will be our X- tau so that should be
xt x (t-tau) average right. Now what we can do is we can we can assume that t - tau is sigma
right, so immediately we can represent t to be sigma + tau right and I can write this as X sigma +
tau X sigma average, now sigma is just a time right so it can be any dummy variable so I can
again put T in place of sigma and it is just for evaluating that dummy variable and it does not
depend on basically we have told already this entire autocorrelation function does not depend on
that particular value right T where it starts or sigma

So it does not matter whatever I take so I can immediately say that this and these two are
equivalent so this is actually R X tau so this is a fundamental property of autocorrelation function
we have also proven this same thing for a time autocorrelation part right earlier when we were
dealing with deterministic signal at that time we have proven that autocorrelation function for a
particular sample signal which is deterministic in nature it's always even symmetric for a real
signal right same thing is happening for a random process also.

So the way we define autocorrelation function we can again see that it is even symmetric that
means - tau is same as our X tau right, so this is true and then we can also try to guess what is this
that autocorrelation function when this tau separation is zero which is if I just put it in this
definition it becomes XT multiplication XT or that is just X 2 T average and then we have said
this is wide-sense stationary signal so therefore the means first-order or second-order if I just take
no separation, so first order second order up to means any order mean should be our moment
should be independent of time, so it should be X2 bar and that is the means that is the standard
deviation or means variance of this signal okay.

So what it says that the autocorrelation function already captures the means associated whatever
underlying associated random process is there any time you actually sample it can be any time
because every time is equivalent and if you take the variance of it, it should be that okay so as
long as the mean is same this is actually becoming the power of that signal right. So that is also
very true because we have also told that X 2 t if we just try to guess how much is that.

So that should be the power and in a different way also we can correlate it to the power if
suppose the signal is also ergodic, then what will happen? The second order means this same
average time average if you wish to calculate that should give me the power so that is equivalent
because it is ergodic signal so immediately we can say this gets correlated or related to the power
of the signal okay.

582
So what we can understand from these things that autocorrelation is a very strong property of a
signal it just gives me back many information about the signal itself okay, I do a Fourier
transform I get the spectrum of the signal of course the average spectrum I put 0 in place of tau
then I get the power of the signal back ok as long as the signal holds property like it is ergodic. it
is wide sense stationary and all those things okay.

The other part is this what I know that this is r X tau is actually the Fourier inverse Fourier
transform of the power spectral density, so if I write power spectral density as s X F okay so
therefore it must be e to the power J 2 pi F tau D F right, now we know that at RX tau = 0 it
becomes power, so let us try to evaluate that r x 0 over here also within the integration tau must
be 0 so once tau is 0 this becomes 1 so it is just -infinity to + infinity s X F DF this is a well-
known thing that that is the power so power is nothing but the average sense power spectral
density that you have got for this particular random signal.

So that power spectral density like the deterministic signal you just integrate from -infinity to +
infinity over the frequency domain you will be again getting back the power okay, so it is all
sitting very nicely mathematically as you can see the way we have we have actually developed a
strong tool to evaluate the random process and then once we have developed that tool we can see
everything is now almost getting similar representation with respect to our deterministic signal
okay, whatever we have understood in deterministic signal here also we are almost getting similar
representation we know a power spectral density of course this power spectral density is a
average power spectral density.

But it behaves like for a deterministic signal whatever power spectral density you are having for
that also we are integrating from -infinity to + infinity in frequency domain we are getting back
power here also same thing is happening we are integrating it we are getting power back okay, so
that is very nice everything is sitting nicely. Now let us try to probably in the last class we have
already started characterizing noise with this random process. So we have told the most random
noise that we can think of is having probably autocorrelation function.

(Refer Slide Time: 09:11)

583
 N
Rx(τ) = δ(t)
2
Which is RX tau which must be almost like delta okay with some strength let us say that, that is
related to the noise power okay let us say that strength is n / 2 okay so therefore that strength
must be n / 2 okay, so we are just saying that this is probably the representation of noise there
were some reasoning behind it that we were saying this is the noise because it is a kind of signal
where in the current time if I have some if I observed some voltage level or some current level or
power level whatever it is next time instance however in finitely similarly small or closer it is to
this current time instant I will have no predictability that means I will not be able to say that it
should be within this range I will have no predictability it can take any value any value possible
from -infinity to + infinity.

So that is probably the most random noise that can come across whenever we are transmitting or
that we can come across whenever we are transmitting some signal right, so that is why we say
that is basic noise which is something which is completely random with respect to our signal
transmission and the autocorrelation function immediately is very high correlation whenever we

584
are not shifting it but a slight shift in fact even if it is infinitly small that should take the
autocorrelation function to zero okay.

So it should not be correlated because I cannot have any predictability with respect to the current
observation, so that should be represented as delta function immediately what we can see for
noise the power spectral density becomes because we know that if it is a delta function of e / 2 if
you just do a Fourier transform that's what we have understood that Fourier transform of
autocorrelation function should be giving me back the power spectral density, so that should be
this so it should be with e / 2 or n / 2 whichever way you represent so that should be the noise.

So this is the most random noise that you can counter and fortunately most of the noise you'll be
seeing looks almost like that within the band of our interest of course it will not be flat over the
entire band because then noise power if you start integrating it from-infinity to + infinity that
what we have understood that noise for will be from -infinity to + infinity you will have to
integrate but if you start doing that it will be infinite right.

So noise power will be infinite in that case, but within the band of our interest probably this will
be flat okay so that is a very important understanding that it will be probably flat of course at a
very higher band it will start showing some low-pass effect okay, but within the band of your
interest this will remain flat okay so that is the characteristics of a general noise. Now let us try to
see most of the time what we will be encountering in the communication means whenever there
is noise in communication what we can see if the noise is like this suppose from the channel it is
coming like this so if I just allow the entire noise then huge amount of power because you
integrate over the entire frequency domain huge amount of noise power will be coming into my
signal and that will of course contaminate the overall signal.

So I want to reduce the noise power because I want to increase enhance the signal-to-noise ratio
or signal-to-noise power signal power to noise power signal power means whatever message I am
transmitting related to that the power related to that and noise power is whatever is coming out
okay out of this process. Now I wish to suppress this noise power because otherwise my signal-
to-noise ratio will be very high which is not good because if noise is bigger than signal I will not
be able to means decode my signal okay.

So if this is the case then what I will be employing most of the time I know my signal will be
within a band of interest if it is modulated it will be in a band if or maybe a in a pass band if it is
not modulated directly transmitted in the baseband it will be a low-pass signal okay, so whatever
it is I will have to employ either a low-pass filter or a band pass filter so that is why you will see
that most of our energy now will be concentrated on characterizing this low-pass noise and band
pass noise okay.

585
So that is something we will be concentrating on so first let us try to see what happens to the low-
pass noise so I know that noise spectral density is something like this which is flat okay, now I let
us say I pass it through a ideal low-pass filter which has a characteristics of this from minus B to
B it has transfer function one and rest of the places it is just zero okay. So if I just pass this
through this process what do we get okay so this is something probably we will have to now see
that a particular random process if it is passed means passing through a particular transfer
function or a particular system how the output will look like okay.

So that something will try to characterize now and after that probably will characterize this low-
pass, so let us say I have a system.

(Refer Slide Time: 14:55)

∞ ∞

∫−∞ ∫−∞
RY (t, t + τ) = h(α)h(β )Rx(τ + α − β )dαdβ

RY (t, t + τ) = h(τ) * h(−τ) * Rx(τ)

SY ( f ) = | H( f ) |2 Sx( f )

586
Which has a transfer function of H(F) okay, and the associated impulse response which is the
inverse transform of this which is H(T) okay, so I know this and I do give our input X (T) which
is let us say a stationary or at least wide-sense stationary random process and I want to get some
output and I want to now characterize this output now it would also will be random okay it
should be a random process but I want to characterize this random process okay.

But what we know that if for a given XT and if I know this HT this should be for a linear time-
invariant system it should be convolution of these two, so I can immediately evaluate YT should
be -infinity to + infinity H alphaX t - alphaD alpha right this is something I know okay because I
have told that this particular system through which I am passing my signal that is linear time-
invariant and we have already characterized that for linear time invariant this is what happens
output and input get this relationship as long as I know the transfer function or the Associated
impulse response okay.

So this is the case now I want to characterize this output process so what I have to do how do I
characterize so far we have discussed that characterizing any random process requires calculation
of autocorrelation function, so therefore I need to evaluate also YT + tau right it is all about that
so YT + tau will be just its when I am passing my XT + tau that should be because it is linear
time-invariant so time wise it should be just YT + tau because it is time invariant so YT + tau
must be according to our understanding of linear time-in variant system that that should be X T
+tau - aIpha right this is something we have already understood that linear time invariant system
should work like this.

So therefore now all I have to do is find out the autocorrelation function, so r Y tau I will try to
find out which is nothing but Y T YT + tau and symbol average right which is nothing but these
two and symbol average so -infinity + infinity H of course there are two integration I will be
merging these two integrations so this integration variable inside variable I will take any I can
take any dummy variable so I will take alpha for 1 and ß for another one right so the first one I
will pick aIpha x t-alphaD alpha second one I will pick as ß, so that should be H ß X T + t - ß D ß
right and I need to do this.

Now these two integration are not dependent on each other so I can actually Club them together
now this ensemble averaging that is really on X that has nothing to do with this internal variable
of aIpha and ß so therefore ensemble average can be taken inside so I can write this as H alpha
does not depend on that so which is not dependent on that X random process so H alpha remains
as it is H ß remains as it is and XT alphaX T + t - ß I take an average of that D alpha D ß I can
write this right.

587
Now what is this? This is if t -alpha I take as something okay let us say sigma and this should be
sigma plus something okay I can immediately write that as sigma plus something, so this is just
nothing but autocorrelation function of X which I know already and I know X is stationary see
for this probably I have written a wrong thing I still didn't know while doing this whether the
output process will remain stationary or not I had no idea I have written this but I should not have
written this I should have written this as t T + tau because I do not know whether this will be
stationary or not okay or wide-sense stationary or not.

But X I know for sure because that is the into input process I have given and I have chosen it to
be wide sense stationary right, so if that is the case this must be function of just tau which is the
separation between these two right. So I can write this as ry so let us write it correctly t t + tau
must be -infiniy+ infinity-infinity + infinity this is H alpha this is H ß and this becomes this
minus this if I just do so that should be T will be cancelled and I have tau + alpha ß so that should
be tau + alpha-ß D alpha D ß okay if you just carefully see this is just to basically convolution
nothing other than that.

So it can be written as H tau I would not do that this is take that as homework this is just a two
convolution we already know single convolution if you just do it two times this will be a
convolution of H tau H - tau and rx tau okay this is the convolution of these three things. So
therefore it is just a variable of tau only nothing else so therefore I can now write my output
process that ry that is just a variable of tau that has nothing to do with T because T is already
getting cancelled due to the property of X where X was wide sense stationary okay.

So immediately I can see the output process which is being generated it is also a wide sense
stationary process, so if the input remains wide sense stationary I can pass it through any LTI
linear time-invariant system I will always get output process which is also a wide sense stationary
process okay. And the process is related to this now I can do a Fourier transform of this, so
immediately I will get s Y F because this is the autocorrelation function so Fourier transform of
that Wiener khinchin process tells me as long as this is wide sense stationary which is the case I
will be getting my average power spectral density so I get average power spectral density.

Now if I do Fourier transform of this should be HF and this should be H - F or H * F so that must
be mod HF2 and Fourier transform of this must be sxf because this is the autocorrelation function
of the input signal that is a very nice fundamental result we have got okay, so two things we have
derived through this process one is that I give a input to a linear time-invariant system remember
all these things has to happen it has to be linear time-invariant we have put all the formulas of YT
YT + tau knowing that it is linear time anyway otherwise I could not have written that okay.

588
So the process has to be linear time-invariant and the input should be wide sense stationary at
least if this two condition happens properly I know that the output process also will be wide sense
stationary at least okay and on top of that I can get the power spectral density which is nothing
but the transfer function square of that linear time-invariant system or mod2 of that linear time-
invariant system multiplied by the input average power spectral density okay, so this is a very
fundamental result which we will use for characterizing that low-pass system okay. So now try to
just see what we were doing for that low-pass system I was having a noise.

(Refer Slide Time: 23:23)

N
× 2B = NB
2
Which was characterized by this ideal noise of having power spectral density so this is my input
let us say SX F, so this is my input which has a power spectral density of this which is constant
over the entire frequency range having value eta/ 2 okay, now I pass it through a ideal low-pass
filter having bandwidth B. So what do I get this is one so therefore I must just get this into mod if
this is HF mod of this square because this is all the one between minus B to B it will remain 1.

589
So I will just get eta/ 2over that band and rest of the cases it should be all 0 so my output that s y
F must be like this it is eta/2 from - B to + B and everywhere else it is 0 right and then actually I
can talk about this process I cannot talk about the power of this because as I have told that
probably we are just representing it but it is actually showing infinite power because from
-infinity+ infinity you integrate it will be infinity but actually it is not like that it should have a
low pass characteristics at a very high frequency probably it will show some low pass
characteristics but I am not bothered about that it is flat in the region of interest of my frequency
okay.

And then I wish to see what will be the output noise power now I can integrate this to get my
output noise power so that should be nothing but eta/ 2 x 2 B because it is all constant, so
integration will be just multiplication of this the band and that overall power spectral density that
should be etax B so this happens to be a low-pass noise power okay. So whenever you have a
noise which is most random possibly that can be and most of the cases it is okay so and that noise
if you pass through a low-pass filter with all our derivation winner from starting from wiener
Khinechin to what happens if I pass it through a LTI system.

So after doing all these things this simple understanding we have got that the output noise power
I will be able to evaluate this is going to be a very important part of our noise analysis further you
have to be very careful about doing these things correctly okay, so once this is being done let us
try to understand some more property of this whole processes okay. So one thing is that is called
in when we were talking about random variable we are actually trying to show the joint PDF of
two random variables right.

(Refer Slide Time: 26:26)

590
Rxy(t1, t2) = x(t)y (t2) = Rxy (t2 − t1)
Rxy = x(t)y(t + τ)
= xy

Same thing will now try to do over here also, and when we will be doing noise analysis you will
see why that is important it is really required that we take two processes and then try to see their
means either correlation or cross correlation whatever you term them okay, so without proof I
will give some of the theorems which are required for us because if we start proving all of them
probably it will be means we would not be able to cover the major part of analog communication
in the duration of this course.

So I will give some property of it so the first part is cross correlation okay, so cross correlation
function between two random processes x and y both of them we assume to be wide sense
stationary at least so both x and y or I should say XT and YT they are independently means I
should not say independent over here because that is another property I should just say that they
are probably means they are actually who at least wide sense stationary okay up to second-order I
know that they are stationary then the cross correlation is defined as this okay.

591
Still I have not exerted the property of there being means wide sense stationary if they are then I
should be getting this just a difference of these two so this must be rXY just a function of t2 - t1 it
should not have the sense of origin that it should not be dependent on sense of origin it is just the
difference between these two time instant it should be dependent on that okay or I can write this
as rXY tau where t is defined as t2 - t1 that time difference all right, so this is what it should be.

Now and this will happen see both of them are probably wide sense stationary but if this has to
happen then there is probably I have not mentioned another extra property which is required
that's called they should be jointly stationary as well so independently they might be means
separately they might be stationary wide sense stationary but they should be also jointly either
fully stationary or wide sense stationary that has to happen then only their cross correlation will
be dependent on just the separation of the time instants we are taking.

So there is a now you can see that there is a concept of joint stationary okay, now next thing will
be defining which is called uncorrelated we have already seen it for random variables now we
will see for random process what do we mean by uncorrelated it is just by definition okay the
uncorrelated random process at that time this are X Y tau that we have defined if they are jointly
stationary okay.

So this is as we know it is XT and sorry okay, so what should happen this must be okay, so when
they are uncorrelated I should be able to just get the autocorrelation joint or cross correlation
function to be just the multiplication of their respective mean value okay and because it is
stationary these two processes are stationary so there is no notion of time because mean at any
time will be same okay.

So therefore if two processes are uncorrelated this must be happening okay, so what we will do?
We will try to define some more property of these things in the next class and we will see how
they are means they will be utilized in our further processing okay, thank you.

592
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 42: Random Process (Contd.)

So we have already started talking about cross correlation function right so that is something we
have already done, so what we will try to do now we will get some more property of this cross
correlation okay we have defined two things one is cross correlation and then we have given a
property of join stationary okay and then we have said what is the property when two signals or
two random process will become uncorrelated we have given that okay.

The next part of the definition is when two random process we call them again by definition they
are in coherent okay or we should say orthogonal, orthogonality deterministic signal we have
already seen now let us see for random process what is the definition of orthogonality or in
coherency both are interchangeably used so I have already told it must be if it has to be in
coherent or orthogonal it must be first uncorrelated so therefore I must have this property.

(Refer Slide Time: 01:20)

593
Rxy(τ) = x y = 0
x = 0 or y = 0
X*T ( f )YT ( f )
Sxy( f ) = lim
T→α T
RXY (τ) ⇔ SXY ( f )
RXY (τ) = RXY (−τ)

This must be there now among them if one of them is 0 mean okay either x bar is 0 or y bar is 0
okay that happens then this happens to be 0 and those signals are generally termed has incoherent
or orthogonal okay so this is just by definition so we are just defining different kind of signals
that will be encountering of course accordingly that Rxy that joint correlation will have it is own
definition okay or cross correlation will have it is won definition so we are just defining them.

So we are saying that if Rxy tau is just it is Rxy T, T + tau it is just function of tau then we say it
is jointly stationary okay if Rxy tau is equal to x bar . y bar or x bar into y bar then we say it is
uncorrelated if Rxy tau = 0 then we say those two signals are incoherent or orthogonal okay so

594
that is by definition and then because we have this we will also have corresponding power
spectral analogy okay.

So similarly we can actually put across power spectral density once again by definition it is just
the fourier transform of this one right so we call it Sxy (f) which is just nothing but limit by
dentition T tends to infinity you trunk it because of basic definition is you trunk the actual signal
if it is auto correlation or sorry auto correlation and it is own power spectral density then it is just
xtf mod square right.

But here because 2 signals are there so you first so for the first signal okay and then do for the
next signal and do for time do in simple average 1/T so that is the definition of this one and if you
take this definition again go through the same process of the way we have to win 3:38 theorem it
will be able to show that Rxy tau as long as they are jointly stationary will be able to prove that
there are Fourier these two are Fourier there okay.

So this is something we will be able to show that and similarly we will be able to also show that
are Rxy tau is a even symmetric function okay so that is will be but when we say even symmetric
it is something that is some hint into it or some trick into it becomes Ryx- tau okay so we are not
proving this results it can be done similarly okay it is not very complicated the way we are
proven that similarly if you know the Rxy tau definition and you just put the things you will see
that this has to be the case okay.

So that must be case and immediately by taking this relationship We can also show that RSxy f
must be equal to Syx – f so this directly comes from here okay so these are the property which
are for 2 different random process and they are correlated or cross correlation they can get this
properties okay cross power spectral density or cross correlation okay.

(Refer Slide Time: 05:15)

595
 z(t) = x(t) + y(t)
Rz(τ) = z(t) + z(t + τ)
= [x(t) + y(t)][x(t + τ) + y(t + τ)]
= x(t)x(t + τ) + y(t)y(t + τ) + x(t)y(t + τ) + y(t)x(t + τ)
= = Rx(τ) + Ry(τ) + Rxy(τ) + Ryx(τ)
= Rx(τ) + Ry(τ) + 2 x̄ ⋅ ȳ = Rx(τ) + Ry(τ)

Now with this let us try to see if I have a signal or a random process which is just a summation of
2 random process it is going to be very important because I might have whenever we are
receiving something I might have one random process which is associated with the signal itself
and another random process which is actually associated with the noise and most of the time my
channel is linear so as long has it is linear definitely the overall signal I will be getting that will
be just addition of these 2.

So that is why it is called additive channel okay so whenever I will be receiving something I can
always by the virtue of this concept that entire channel is linier so I will be able to say that it is
just additive the noise will be added on top of that signal so two random process often in my

596
receiver will be additive so I need to also see how do I characterize that okay so if I wish to
characterize random process can be characterized by it's auto correlation function so I need to
evacuate Rz tau okay which is nothing but z t z t+ tau and simple average we have the definition
now zt you replace by xt and yt.

So xt + yt and xt + tau + yt + tau right if first term will be xt, xt + tau and this because it is simple
average it will be distributed over the addition so it should be x t xt+ tau and symbol plus there
will be another term related to y so yt yt + tau that is where you will see that y that cross
correlation is so important here that term will becoming so you will have multiplication with
these 2 and you will have a multiplication with these 2 so xt (y) t + tau and yt x t + tau so this is
nothing but the auto correlation function of x.

So I can write that Rx tau as long as my x and y are all wide send 7:49 stationary at least so it will
be just dependent on tau this is also same thing so I can write if these two process x and y are
jointly stationary or jointly wide send stationary then I can write this has R xy tau and this will be
R because y is taken first so that should be yx tau right so that is something what we are getting
okay.

Now let us start putting all those assumption or all those definition that we have given so suppose
the process is just not jointly stationary these two process they are also uncorrected then
immediately what will happen this Rzz will be Rx tau + Ry tau + this should be x bar y bar this
also should be y bar x bar or x bar y bar that same so that should be 2 x bar y bar okay now if
they are incoherent or orthogonal now that must be 0 so I get Rx tau put tau = 0 what do I get.

(Refer Slide Time: 09:10)

597
Rz(0) = Rx(0) + Rx(0)
z2 = x2 + y2

For remember whatever we are doing that is for incoherent okay or orthogonal system what is
this we have already proven that should be the power this is what happen for signal which is
incoherent or orthogonal okay if two signals two input process are incoherent or orthogonal most
of the time this is what is happening okay your signal will be probably even if signal does not
have a zero mean you noise definitely will have 0 mean.

So we have told that and they will be actually means uncorrelated okay so that is always true
signal cannot be correlated with respect to noise and noise cannot be correlated with respect to
they have different source where they are getting created so they cannot be correlated okay so if
that is the case then they will be definitely uncorrelated and even if your signal that depends on
how your transmitting if that is not having 0 mean but your noise will definitely have 0 mean.

So we have told that it will be incoherent if one of them are having 0 mean so immediately I can
say that noise + signal which is coming to me in the receiver will always be incoherent or
orthogonal so I can always write this so therefore I can always write so that is a very fundamental

598
thing that for our analysis we can always very safely say that as long as they are stationary and
uncorrelated and all those properties holds I can always take the overall power that I am receiving
it is just addition of noise power + signal power.

This we take has assumption but this is the fundamental basis behind it happens because of these
thing if they were correlated if they were means in coherent I could not have written this so you
have to be very careful when ever your putting this formula you have to also cheek whether all
the assumptions that we have already stated are valid for the system as long they are valid you
can take this assumption otherwise no okay. So this is something you should always remember
okay.

So now we are almost in the mind set or frame work to now analyze we have already analyzed
the means low pass noise now in receiver you can always see there are two ways the receiver can
function one is I am transmitting base band data okay then probably it is around 0 the frequency
component that is there suppose voice I do not modulate I just transmit it as it is then probably it
will be around that 0 and I have to employ a low pass filtering.

So the noise that will be there in the channel or at the receiver it will just go through that low pass
filter okay so that case low pass filtering version of noise is good enough but when I do
modulation what I will be doing I will be translating my signal I will do modulation, modulation.
I have, we have already see that is just translating to a higher frequency so I will take that entire
band and put in a higher band.

And at that point if I wish to reject some amount of noise what I have to do I have to employ a
band pass filter always you will see your receiver will start with a band pass filter that serves two
purpose one is of course it gives you the rejection of all other signals you just want to get your
own signal so it just rejects all other signals that is the first thing second thing is it also rejects
most of the noise out of band which is not inside your band because inside your band you cannot
do anything noise will be added with your signal.

And you have no way to separate them out if you are not doing digital processing okay as long as
signals are analog you have no way to actually suppress them but what you can do outside the
band where you are not interested just put a band pass filter it will reject all the noise the way we
have seen for low pass filtering that happen so that way any receiver if you have modulation
always we will start with a band pass filter as a noise that comes that goes through that band pass
filter so we have to first characterize what happens to the noise if it passes through a band pass
filter. So that will be your next target okay.

(Refer Slide Time: 13:53)

599
x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)

So let us say I have a noise which is a band pass noise okay whose spectrum looks like this it is
similar so that is 0 that is fc that is –fc this is that 2B that is the band width okay and because the
noise was flat so this particular value is Theta/2 which ever was there so same noise same power
spectral density it is we are still assuming that it is flat over the band of our interest okay so it is
still remaining flat it will probably drop some other frequency some other very high frequency
but not in the band of our interest.

So as long as we are taking that assumption the noise, band pass noise will look like this okay
and associated to that this is the spectrum so associated that I will have a noise equivalent random
process this xt I will now say something and then we will prove that this what can be done at this
xt whenever it has a characteristics of this one that it is a band pass noise can be represented as
this.

Where xc and xs we'll later on prove that they are uncorrelated in coherent random process so
they are actually jointly of course jointly stationary that is the first thing so they will be
uncorrelated to each other and they are incoherent to each other so we will prove this okay so

600
these 2 are 2 separate random process Xc and Xs this is called the in phase part and this is called
the quadrature part of noise.

Whatever I have written I have not proven that okay so we will try to prove this because then
what will happen any noise that comes in once you pass through a band pass filter we will be able
to write it this way okay and that will help you because we will also get the property of these 2
things okay and that will actually help you will in the process we will characterize both of them
okay.

Xc and Xs what is the associated auto correlation function what is there cross correlation function
and then what are the relationship between them as well as the power spectral density so we will
characterise everything but we are first we are saying that Xt can be represented like this okay
where wc is related to fc 2 Pifc okay so this will be always possible so let us try to see what are
these things.

So for that I will draw one circuit okay just to prove this we will see we will appreciate after I
analyze this but initially just take the circuit as it okay so there is multiplier so Xt whatever the
input signal I have which is a band pass signal right that is a it exits in a band I multiply this with
2 cos Wct + theta where theta is a know random variable okay I get a1 over here I pass it through
a ideal low pass filter of frequency response HOF where HOF looks like this.

It is ideal filter form – B to + B transfer function it is 1 okay I pass it through it I call this b1 and
the I again multiply this with cos Wct + theta remember I am just creating a hypothetical circuit
but this circuit is realizable because these 2 are having same phase that means it is means actually
generated through same oscillator okay so that is possible in a circuit further this is c I have a
adder over here which creates d and the other arm I will have similar multiplier but now I will
multiplying with sin.

This is should be a2 again ideal similar low pass filter which has same transfer function so it will
be 2 another multiplier where I multiply again with the sin add these two and I get t okay so I can
write d as yt that is my output so what will I try to see is the transfer function of this particular
system that I am creating okay so I will first try to evaluate what is the transfer function of this
then you will understand.

Why I am doing that all those things will be very clear okay is I am just first trying to get the
transfer function of this whole system how do I get a transfer function I first need to get the
impulse response of it so basically I excite this with a delta t and try to see what will be the
corresponding output and when creating this transfer function I have to also make sure that this is
linear time invariant .

601
So therefore what I will do instead of just putting a delta t I will give a delay to that delta okay so
basically I will put a delta t – alpha okay so I will just give delay so that I want to test that output
it is just a delayed same amount of delay to the output it is just function of that entire t – alpha
that is what is happening okay so that is something I want to test then it will be linear time in
variant okay.

(Refer Slide Time: 20: 34)

a1 = 2 cos (ωc α + θ) δ(t − α)

a2 = 2 sin (ωc α + θ) δ(t − α)

So let us try to characterize this whole thing so at a1 what do I get I have a delta multiplied by
cos so any function multiplied by delta is just pick the functional value at that point so if I
multiplying deltat – alpha that means at t = alpha it w ill pick the functional value and it will put a
delta function on over there so therefore it will be if you see over here a1 is nothing but this cos is
multiplied by delta t – alpha so therefore it will just get multiplied and t will be replaced by
aIpha.

602
So I will just get a1 has 2 cos wc aIpha + Theta delta (t – alpha) right this is what I should expect
at a1 similarly a2 must be 2 sin wc alpha + theta delta(t – alpha) right so these 2 thing I get a1 a2
value now this is just a delta function pass through a if I just again come back to the same picture.

(Refer Slide Time: 21:49)

x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)

This a1 passes through a ideal low pass filter which has a corresponding transfer function as h0t
so at delta function now is the input to a h0t what do I expect at the output it should be
convoluted with h0t and h0t convoluted with a delta function must give me the same function we
know that any function convoluted with delta will give me that same function okay.

(Refer Slide Time: 22:21)

603
 a1 = 2 cos (ωc α + θ) δ(t − α)
a2 = 2 sin (ωc α + θ) δ(t − α)
b1 = 2 cos (ωc α + θ) h0(t − α)
b2 = 2 sin (ωc α + θ) ho(t − α)

d = ho(t − α)[cos (ωc α + θ) cos (ωct + θ) + sin (ωc α + θ) sin (ωct + θ)]

y(t) = 2h0(t − α)cos (ωct − α)

So therefore if I just try to calculate b 1 at must be same this is a constant for it right so 2 cos wc
alpha+ theta now delta will be convoluted so it should be h0t – alpha similarly b2 should be 2 sin
wc alpha + theta delta sorry h0t - alpha after that what we are doing so from b1 and b2 we want
to get c1 and c2 then add these 2 we will get yt right so what is c1 that is just multiplication of
this and c2 just multiplication of this so therefore overall d if I write that is nothing but this h0t -
alpha will be there in both the cases.

604
t-alpha I will have 2 everywhere so 2 I can take out and this is nothing but cos wc alpha + theta
intocos it will be multiplied by wct + theta+ sin w c aIpha + theta into sin Wct + Theta right see
why I have created that whole circuit because I wanted a cos a-b formula nothing else this whole
circuit was created because I wanted that cos a-b okay y cos a-b immediately you can see if I just
do cos a- b so it will be Wct + theta, theta get cancelled it will be just wct- aIpha h0t – alpha into
cos wc t – alpha that makes it a linear time in variant system because it is just a function of this yt
what I get because I have put a delta with a delay of alpha then what do I get it is just function of
that t – aIpha so entire gets similarly delayed right so it is time in variant so I could prove my
point by that okay.

(Refer Slide Time: 24:51)

y(t) = 2h0(t)cos (ωct)

h(t) = 2h0(t)cos (ωct)
H( f ) = H0( f + fc ) + H0( f − fc )

So over all if I just say that my yt if I just exceed with delta t it should be 2h0 t into cos wct right
I can just write this when I put delta into it so that must be my impulse response therefore

605
because I have put delta t as my x and I have got this so what is this, this is nothing but let us try
to just see this is actually this happens to be my ht now which is 2 h0t because it is exceed by
delta t and I have got yt has this.

This must be my impulse response of that whole system which is this okay so if I just take it to
the frequency domain so what will be my Hf that must be so cos I am multiplying so this is just a
frequency translation, of this h0t if that is H0F. so H0F goes to + fc and –fc, there will be a ½ due
to that cos and that ½ with this 2 gets canceled. right, so I get H0f + fc + H0 f- fc. so if my H0
was something like this – B to B. what is hf? nothing but, this gets translated to + fc and –fc. this
0 + fc – fc that is 2B how does it look?

Its just like a band pass filter. so therefore with all these things what I have created is, a ideal
band pass filter. so what was my x?that was within that band, okay, a band pass noise, we are
talking about band pass noise. if x is that, because x was actually my band pass noise, if I pass
through a band pass filter, which is having equivalent band, so this yt will be exactly same as xt
because it will just pass everything. so that is already a band limited band pass signal passing
through the same band will just give me same thing nothing extra will be coming out.

(Refer Slide Time: 27:39)

606
SY ( f ) = | H( f ) |2 SX (t)
y(t) = x(t)

So therefore, I can always write. that my Syf what that should be, that should be mod. we have
already proven, that Syf output should be Sxf into mod Hf square. right, but my Syf is exactly
equivalent to my Sxf. right, so therefore I can immediately write my yt is nothing but, xt. This is
something I know already by design.

(Refer Slide Time: 28:24)

Okay, So now, if I just take back that same thing. okay, so this is now becoming my xt. now let us
trace back, what is xt, xt is something over here, let us say that is b1 I call that as xct. whatever it
is, and I call this as xst this multiplied by cos and this multiplied by sin added should be, means
giving me back xt.

(Refer Slide Time: 28:50)

607
SY ( f ) = | H( f ) |2 SX (t)
y(t) = x(t)
x(t) = xc(t)cos (ωct + θ) + xs(t)sin (ωct + θ)
x(t) = xc(t)cos (ωct) + xs(t)sin (ωct) θ =0

So therefore I can always write xt as something, I do not want that is which is x ct multiplied by
cos omega ct, according to the design of my over all circuit and xst sin omega ct + theta. this I
true for nay value of theta, see even if put ? = 0 it must be still true. so I can always write
therefore xt is actually nothing but xct cos omg a ct + xst sin omg a ct. where xct is nothing but
my signal at b1, and xst is nothing but my signal at b2. now if yi just ask what is this signal? this
was after passing through a low pass filter.

(Refer Slide Time: 29:50)

608
So this must be a pass band signal or base band signal. so it is a low pass equivalent signal and
this is also a low pas equivalent signal. so the band pass signal, or band pass noise is nothing but
means represented by in phase component and quadrature component, which are just loss
equivalent signal. that is something we are proven right now. so what we will try to so in the next
class, is to see the characterize of those xct and xst.

What are they? because we already through the circuit we know, they also have relationship if we
just see the other half of the circuit/will be able to get some relationship between xct and xt, and
from there we will try to characterize what they are. what are the, their specific properties. we
will be able to characterize those things. so in the next class properly w will try to do that. okay,
thank you.

609
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 43: Random Process (Contd.)

Okay so what we have done so far in the last class is that we are trying to characterize the band
pass noise right and for that target we have actually devised a circuit and we are trying to evaluate
the transfer function of that circuit right so by exciting it with a delta function we first proven that
that particular circuit is a linear time-invariant circuit and then we have proven actually derived
the transfer function of that circuit and we could prove that it is just nothing but a band pass filter
okay.

And then we could show that if a band pass signal exactly matching with its band if it is passed
through this it will remain same okay so that is why Xt if that is a band pass noise exactly
matching with its band then at the output also it will remain as that and then we could say that
okay the Xt can be represented because of that circuitry can be represented as some Xc and Xs
in-phase and quadrature right. So that is something we have proven.

(Refer Slide Time: 01:31)

610
x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)

So therefore finally it came to that Xt is Xct cos ωct + X st sin ωct where we know that this Xct I
can actually trace back them in my circuit the circuit we have shown.

(Refer Slide Time: 01:50)

611
x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)

Sso Xct is nothing but equivalent to b1 and Xst is nothing but b2 okay so we already know that
what they are okay. So now let us try to see what is the characteristics of this Xct so Xct if we just
trace back through this circuit it is nothing but this Xt okay multiplied by this and then pass
through a low-pass filter ideal low-pass filter right that is actually b1.

(Refer Slide Time: 02:31)

612
 x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)
xc(t) = 2x(t)cos (ωct)
xs(t) = 2x(t)sin (ωct)
X( f ) = X( f + fc ) + X( f − fc )

So I can write that as 2xt cos ωct we are again taking θ to be 0 okay because it is valid for any θ
so I can always take θ = 0 so this is xct pass through our ideal low-pass filter as long as my Xt is
band limited okay so basically the Xt is limited to - B to +B within that fc then if I just multiply it
with cos so it will actually come down over here I will just take that low-pass portion of it okay.

So it remains this only we have to make sure that it is within that band - B to +B okay so that is
actually Xct which is low-pass equivalent of course it has to be passed through a ideal low-pass
filter and Xst similarly should be 2 x Xt sin ωct right so let us first try to see this one so what
happened my Xt suppose that looks like something like this that is so this is fc it must be
symmetric.

613
So something like this so let us say this is - B - fc or I should say - fc - B and this is- fc +B that
point okay so this is actually my 2b similarly this is fc +B and this is fc – B this is my 2b okay so
now if this is Xt what is Xct the in phase component that is nothing one this xt multiply cos and
two is there because whenever I multiply cos in the spectrum if I wish to see there will be a ½
term so that will be cancelled out so it is just whenever I multiply cos this whole thing will be
shifted to +fc and - fc right.

Once I shift this this particular thing once I shift to + this will come over here and then this
particular term will go to how very high frequency 2fc and when I shift to the other side this will
come over here and this term will go to - 2fc those things will be anyway cancelled because there
is a low-pass filter so what will remain over here at the center is something like this from - B to
+B this particular pattern will be repeated and from this to this particular pattern right.

It is the addition of these two which will be the Xc okay so or if I just take a Fourier transform
this so Fourier transform let us say it is X let us say Xc or Xcf right so what that should be that
should be actually Xf +fc f- fc so original f f + f c +Xf –fc but this must be defined for a f where
mod f ≤ B okay which is just like this all the higher frequency term will be neglected it is just
these two part will overlap and whatever I get okay so that is actually my Xct and corresponding
Fourier transform of that okay.

So I immediately get the corresponding Fourier transform this one right it is very easy to get
directly the Fourier transform as long as I know the Xt means the power spectral density so
basically what I can do instead of doing this I can relate them in terms of power spectral density.

(Refer Slide Time: 07:00)

614
SXc( f ) = [SX ( f + fc) + SX ( f − fc)] |f | < B

SXs( f ) = [SX ( f + fc) + SX ( f − fc)] |f | < B

xc2(t) = xs2(t) = x 2(t)

So I can write Sxc f which will be just this whatever Sx suppose Sxf is known that is for the Xt I
know it is average Fourier means average power spectral density so this should be multiplied by2
cos ωct so therefore it should be that Sxf must be shifted to + fc – fc so I can write Sxf sorry + fc
+ Sx f – fc and then I have to put this condition that mod f ≤ b it should be this and then for
because it is power spectral density so therefore for cos there will be a ½ term which will be
squared.

So there should be a 1 / 4 and there is a 2 term already that should be squared so it would be 4, 4,
4 gets cancelled so this happens to be my power spectral density of Sxc f okay so without going
into the autocorrelation or cross correlation function I directly define my power spectral density
as long as Xt is defined I can always get the corresponding Xc that is the in phase terms, power
spectral density similarly Sxsf also will be same because whatever sign is there it does not matter

615
when you take power spectral density it will be squared and all other things will be cancelled it
will just be + again f +fc Sx f – fc again defined for this okay.

So immediately I can see the power that we wish to evaluate so if you just try to calculate this
power that must be the integration of this one for the whole spectrum okay so this is just defined
within B so it should be this + this okay so Sx + fc within this and -fc within this so that should
be the power of this similarly Xs 2 t also have because they are having same similar power
spectral density.

So it must be same and not only that this is also same as this because think about these two power
spectral density.

(Refer Slide Time: 09:31)

x(t) = xc(t)cos (ωct) + xs(t)sin (ωct)

xc(t) = 2x(t)cos (ωct)
xs(t) = 2x(t)sin (ωct)
X( f ) = X( f + fc ) + X( f − fc )

616
So if I just go back to this example what is happening the power spectral density of this one will
be this + this same thing happening over here it is just this portion is being repeated over here this
portion is being repeated over here without having any attenuation so the power spectral density
will be just addition of these two which is same.

(Refer Slide Time: 09:48)

x 2(t) = 2NB

xc2(t) = 2NB

So therefore the overall power remains the same for the in-phase component or the quadrature
component as well as the random process that we are talking about this will be always happening
as long as the bands are properly adjusted okay so therefore what we have seen so far is we have
seen how do we actually evaluate a band pass random process through equivalent to low pass
random process which are called in-phase and quadrature component okay so that Xct and Xst
and we have also seen the interrelationship between their power spectral density okay.

So that band pass process how the Xct can be related to the corresponding Xt that is something
we have seen and we have evaluated their power also okay so this is something we have already

617
done now what we wish to do is something like this that we want to evaluate the cross-correlation
of this Xc and Xs now we have to two things so this Xct and Xst + τ which is nothing but Rxc xs
okay.

If it will be stationary, so what will be able to prove that this will be just 0. okay so for that you
just have to take this relationship Xc and Xs and follow through the means you know already, the
autocorrelation function of Rx just go through that follow the steps you will be able to prove that
this is zero. so basically this means they are uncorrelated they are incoherent. this is something
will be always able to prove okay.

So that is another thing which will be probably means heavily required for our analysis of noise.
so whenever later on you will see whenever we have a band pass noise, that is coming out
because of the band pass filtering will be always writing it in terms of in-phase and
quadrature.because once the noise comes in it is the noise + signal which goes through the whole
demodulation process.

So we have to actually take the signal as well as noise going through the demodulation process.
so if that has to be done then I need to have a representation of the noise. so that is the
representation of the noise that NT suppose it is noise which is a random process that must be Nst
sin ω CT + Nct sin a cos ωct. so always I will be able to write that way and then I also know they
are corresponding power spectral density their cross correlation all those things we know already.
because we have characterized them already. so once though all these things are known you will
be able to see that we have a strong tool set which will give us some analysis. okay.

So all we are trying to do over this exercise. that definition of random process.what stationary
how do I get power spectral density, what is auto- correlation, and then if I have a low-pass noise,
band pass noise, what are the corresponding representations, so all these things we are doing, just
to add ourselves to do noise analysis of the system that we have already discussed. so far all those
different kinds of amplitude modulation systems okay.

So we will just give two more things which will complete this entire discussion probably, so
those two things we will just discuss briefly, and after that we will give you an outline of how we
actually can do noise analysis. so let us just first try to see go back to our noise, which was
something like this, right so we want to characterize that noise so which is having a strength of η/
2 righ.

Now this noise is passed through a band pass filter. okay centered around fc and band is.to be
similarly - fc. band right. how the noise will look like, after passing it through that so it will be a
band pass process, so that should look like this. so it should be still the strength should be eta/ 2

618
because these are all one. so mod hf2 will still remain 1, pass through this we have already
proven that it should be mod hf2 the input noise, input noisy η / 2.

So we just multiply that and we should get this, right so that strength should be η/ 2. okay defined
at - fc +fc and within this band of 2b. so what is the noise power? so let us say this noise is
defined as xt, so x2t that must be integrated, noise power is just integrated over this entire band.
okay so which is nothing but η / 2 so 2 x integration from this to this, right so that is nothing 2b x
η/2. so that is η x b'2 times so that should be 2 x η x B that is my overall noise power right.

Now let us see what is my Xct. so what is Xct. Xct must be this shifted to this, and this part
shifted to this,so therefore how much I do get.it should be band limited from - B to Bibut this one
will be shifted, and this one also will be shifted so overall strength will be η/ 2 + η /2. so that
must be η. what is the power of this one.

So Xct square bar there is also 2 η x B if I integrate over this. which we have already said that
this and this power will be equivalent. similar things will be happening for Xst as well. right for
Xst also same representation will be happening and I get 2η B, so that is how we will be
evaluating the noise power and will be evaluating the corresponding power spectral density of
noise.

Whenever we have a noise. okay, so in our noise analysis will be taking the noise as this which is
a flat spectrum, okay and whichever filter will be employing be at low pass or band pass will be
always taking them to be ideal.because if you do not take ideal.then probably the analysis will
become too much complicated okay you can always do that, but just to see the performance
probably for both, means if I wish to compare performance of two systems,I will take ideal filter
for both of them and then try to analyze it.

Because it just becomes, life becomes much more simpler. okay so that is something we will be
trying to do. now let us try to talk about another thing.okay which is called that noise we are
characterizing, what is that noise you might be hearing this particular term so often that the noise
is AWG okay.

(Refer Slide Time: 17:48)

619
For AWGN or the channel is called as AWGN channel. okay what do I mean by this?the first term
is called the additive. it is just defining the noise, right.that means I have already assumed that
noise either is generated in the channel or in the receiver circuit, whichever place it is being
generated, its then passing through, the channel will be linear and the circuit also will be linear.
so as long as they are linear, so basically the noise will be just added on top of signal.so it is
additive noise not multiplicative, or any other kind of noise.

So it is just additive noise, so you will be always assuming that if I have signal the noise will be
just added to it.okay, so that is the first thing additive.then the next term called white, so the noise
that we have characterized which is probably the most random noise that we know which has
autocorrelation function of impulse, and corresponding spectral density which is flat. this is what
characterized it to be white, that means it has all the frequency component equivalently, with
equivalent strength.it just comes from that concept of white light.

So what is white light that it has all other lights or all other wavelength that characterizes
different light, whether it is violet or blue or any other things. so all those are equally added to
create white light okay so same thing the white noise is being created by equivalent noise
component at every frequency, as many frequency component you can think of up to infinity. so
all those frequency components are equivalently present in this particular noise, and because the

620
noises like this it just comes from the autocorrelation function of our understanding from there,
because it is like this that every frequency component or equivalent present we call this as a white
noise.

So most of the time will be considered white noise, the next part is G which is the Gaussian, okay
so what does this means? this actually says that see this is the power spectral density, this is the
autocorrelation, actual noise will look like something like this, right and it will because it is a
random process you can have different such samples, okay in finite number of such samples
which will look different if you take different samples, fix any instant it is a stationary process,
we have already talked about that that whatever that noise is that is already stationary okay.

So anytime you pick,I know that the statistical property does not change. okay so if I just pick
this time instant or this time instants, and try to take this samples put a PDF, and what I will
observe that this is following a Gaussian process. we have already discussed about Gaussian
process and that is why we have discussed Gaussian process.because it is so important that most
of the things probably we have not discussed one important property of a random variables, that
if you have many number of independent random variables. infinitely many and if you just all of
them are independent, if you just add them together the PDF that you will be getting that always
goes towards Gaussian okay.

That is called the central limit theorem, we have not discussed, that but in probability theory we
will be learning those things. so because of central limit theorem, we know that always a process
which is constructed of many infinite independent processes will tend towards Gaussian process.
so the noise is generally being generated by infinite number of things, which are naturally
happening inside the nature. okay so because they are all independent and they are all adding up
to create noise.

So we can all very safely say, that it follows central limit and the noise process will have PDF
which is Gaussian. so that is why this noise generally is taken to be Gaussian. okay, so what does
that means? that any sample time you take, take the ensemble, put them in a histogram, normalize
them, whatever PDF you will be getting that will look exactly like Gaussian PDF. with a
particular σ square. okay, which is actually the power of that, right and that σ2 of that Gaussian
should be equated to the power that you are calculating from the from the noise. like for this band
pass noise, we have evaluated this 2 η B is the noise power. that must be the σ2 of the
corresponding Gaussian.

Because the Gaussian process has two things one is mean and variance. these two has to be
specified. so variance is specified. and all these Gaussian noise are generally because they are
created by random process which can means either create positive voltage or negative voltage

621
equally likely. so it is often having zero mean. okay so always the signal which is characterized
by noise signal will always have zero mean, so we can always say that it is always characterized
by a Gaussian process, which has zero mean and the σ2 is just equivalent to the power that we are
exactly okay.

The noise power we are thinking of. so that noise power will depend on that power spectral
density of it, and the band we are operating at, so accordingly we can the way we have calculated
noise power, will be getting that power, and that σ2 will be equivalent to that power, because
these two are, means exactly equal-okay so that is why whenever we talk about a particular
channel, and a noise process that is being added, we always take help of additive white Gaussian
noise, there are other noise for different system, like you know optical system probably it will not
be similar things okay.

So because of photon counting, and that counting process becoming poison. so there will be
different kind of noises which can also be added. but generally speaking most of the receivers are
being analyzed with respect to additive white Gaussian noise. so that is why it will be very
important that we know this all these terms. so whenever will be because of this presence of this
noise whenever will be characterizing noise will always say the noise amplitude will be just
directly added to the signal amplitude. nothing else its additive, okay and the noise power spectral
density will be flat, with the means power spectral density specified as η,or η by 2. or whatever
you like, and the noise statistics or the PDF if you just take a particular time instants, freeze the
time instant and take the PDF will always get that to Gaussian.

These are the criteria which will be always fulfilled whenever we, means concern ourselves with
respect to the noise that we are analyzing. okay so this is one thing I wanted to characterize, the
other thing I wanted to say that this in-phase and quadrature representation that we have done that
is not unique.

(Refer Slide Time: 25:44)

622
So let us try to appreciate that part, you remember we had a means we are just trying to show that
suppose I have a band pass signal. or random process which is characterized like this. okay and
this is the band where it is defined. now I want to have a quadrature and in phase representation,
what I will do first I will choose fc.

Now carefully check I can choose the fc over here, I can choose the fc over here, once I choose
the fc over here what I have to make sure, that the entire signal means that Xc if I wish to
represent, that my bandwidth will be represented in such a way that the entire signal comes under
that band. so accordingly I will be defining my band. if my fc is over here,then I will have to
define my band accordingly. once I define a fc and corresponding band, then immediately if you
try to represent the Xc or Xs that will look different. definitely because they will have a different
kind of shape which will be translated into the central frequency.

So there is no unique representation of these two things. okay so whenever we talk about that in
phase and quadrature component it is not by definition unique. it just depends on where you put
your fc that central frequency.because you will be given a band pass equivalent signal, nobody
will tell you with that spectrum that this is the central frequency. that is according to your choice.
you can choose a central frequency, accordingly you will have to adjust the band and once you
adjust this band and central frequency.

623
There will be a definite Xc, and Xf which will be created. and this will keep on changing as you
change your Xc and Fs. as you have seen also that particular circuit we have drawn, that is
completely the transfer function of that circuit depends on what cos ω CT we give. if we keep
changing that ωc, and what kind of bandwidth we put for that middle low-pass filter, that h0f if
we change these two,the representation accordingly will change, so be very careful whenever you
are representing it that you should understand that there is no unique representation, it just
depends on your fc definition.

And the band definition. as you define them accordingly or Xc or Xs will be changed.I can give
you a very simple example, the one we were doing that band past noise, that this was my fc by
default, and this was 2B, right and this was - fc and I was having this. now suppose I do not do
that I define fc over here, so then the band has to be taken as this/that mean 4 B and then if I
translate this, what will happen the overall band will be from - 2 B to +2 B and the strength will
be just η / 2.

Earlier, what was the representation, because I have defined it at the central so the representation
was from – B to +B. and the strength was η. right, so you can see already there are two different
representation. as I define my fc differently and accordingly corresponding back. but the overall
power if you wish to calculate it will remain the same.so the noise power and the associated in-
phase and quadrature component, whatever you define. their power will remain the same. it is
just their spectral property will change. due to your definition of fc and 2 B.

So I think we have almost finished our discussion of random process, and we are now almost
ready towards jumping into noise analysis.so what we will do in the next class is we will try to
see how the means what should be our way of doing noise analysis. because whenever we do not
do noise analysis, we want to compare two system. and suppose DSBC an amplitude modulation,
how they perform in terms of noise. can I say this is better in terms of noise. okay or in terms of
noise cancellation.

So how do we actually benchmark them. what is the criteria to do that?how we evaluate our
things? so that we can actually compare two processes so first we will say that. with that we will
have a process which will be created means for evaluating this particular modulation schemes
and then we will go on analyzing those modulations okay thank you.

624
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 44: Noise Analysis – DSB- SC

Okay so I think we have gathered enough information regarding the random process and now
probably we are almost ready towards utilizing this means understanding of random process
towards doing the noise analysis of some of the systems which we have already discussed. So I
will first give a outline of the basis of our noise analysis means how we wish to do that, what is
the measurable quantity? How do we really characterize a particular modulation scheme in terms
of noise analysis?

So what we will as whenever you are trying to modulate something we will try to see if we have
not employed that modulation that means the original data which is in base band if we would
have directly transmitted it and in the receiver just with a low pass filter if we would have
received that, so what could have been the signal to noise ratio that we will try to measure 1st. so
basically our that we will call as reference signal noised ratio okay.

(Refer Slide Time: 01:38)

625
 w

∫−w
Sm( f ) d f = P

w(t) = m(t) + n(t)

PN = N0W

So that reference signal noised ratio how would we like to compute, like suppose I have the
message signal mt which might a voice signal let say. Of course it is a band limited signal so it
goes from –w to + w, so that is the message signal, frequency response of that message signal.
We also know that the power of overall signal that is actually –w to + w if the corresponding
power spectral density let say that is s(f) or smf this if we integrate we get the power P, so this
something we know.

That is the overall power, so this is already given, so this is the overall power of the message
signal and that will be means received at the receiver. So I am not saying that it will be
transmitted or we are saying in the channel there is no attenuation let say. So either it is
transmitted with this amount of power or there is some attenuation so just at the beginning of the

626
receiver or the antenna of the receiver this is the power that we will be getting okay, if we just
transmit the signal. Alright.

So either way you can take assumption, so very clear cut assumption is we are not modeling the
attenuation that is being created or power distribution, that is being created by the channel okay
so that is something we are not because here we are just interested in channel noise not the
attenuation okay. so we will take that assumption that whatever is transmitted that will be
received at the receiver. So this is the power that will be if in the base band directly I transmit the
signal without any modulation, so that should be the power which will be launched at the
transmitter and that will be received.

And this corresponding signal is band limited, so what we should employ the best thing we have
already talked about AWG and noise so we will say that in the channel probably AWG and noise
will be added okay. So it is the white noise so the spectrum is flat and it looks like this. With the
strength it has 0/2 okay so that is the power spectral density of noise PSD of the noise signal lets
say that is ntSo this is something we know already the noise characteristic and we also know that
the noise follows the Gaussian statistics that means at the particular time you sample from the
entire ensemble of noise signal that are possible and then if you plot the histogram with the
normalization it will get the Gaussian PDF okay. And this noise will be in the channel added to
the signal that is also something we know. That is why you have called it additive wide Gaussian
noise okay. So in the channel what will happen if I transmit this m(t) at the receiver because we
have no attenuation so m(t) will be intact at the receiver + noise signal n (t) will be added

That should be my signal that is received okay, so let call that as let say w(t) okay. So this is what
will be received by the receiver, now what do I want to do? It is low pass equivalent signal we
have already told that if it is a low pass signal I am only interested about the message, if I allow
the entire band to come ofcourse my message will come and nothing else we are transmitting lets
assume nothing else we are transmitting but the entire noise power will actually come to receiver
so signal noise ratio will be very high

Instead of that we can just employ a low pass filter of bandwidth w, that –w to + w, if we do that
how much noise will be coming in that will be from –w to + w if you integrate this. So that
should 2w x N0/2, so overall noise power should be N0 x w, 2 gets cancelled so this is the noise
power PN okay, so this what I will be receiving out of which the noise power is this the signal
power I already know because we said it is not getting attenuated, so that should be the signal
power.

So over all we can say signal to noise ratio should be corresponding signal power divided by
corresponding noise power, so this is how we will be characterizing the system, if we just, this is

627
the reference SINR or SNR. When we transmit the signal raw, in its raw form without any
modulation okay so that should be the reference part and then what we will do, we will try to
employ some modulation scheme. we do not know what modulation scheme.

(Refer Slide Time: 06:54)

w(t) = s(t) + n(t)

(SNR)mod
FOM =
(SNR)ref

so due to that there will be some signal s (t) which will be transmitted probably okay, so it will be
a modulated version of m (t). so if it isDSBSC then it will be m(t) cos ✗ cos 2 wct. If it is SSB
then it will be mt cos w ct +mht which is Hilbert transform of mtx sin ɷ ct, something like that
either + or - okay . so whatever the modulation scheme I will be transmitting this okay. So this is
something we will be launching and then at the receiver there will be some procedure because
what we wish to calculate is after the entire reception process, whatever signal we get what is the
power of it and some spurious noise that will be also going through that receiver process.

628
And due to that at the end, there will some noise power and we have to take the ratio of these
two. okay, so that will be the signaled noise ratio of the modulated signal okay. so we modulate
we launch s (t) and then at the receiver we receive it along with noise. So the signal s(t) + m (t)
will be my input at the receiver then the receiver will do all it is processing. If it is the envelope
detector, it will do that, if there is a bandpass filter it will do that, if there is alow pass filter, there
is a mixture so all those things will be done equally with the signal as well as the noise.

So that is something which we will have to carry through the entire experimentation. okay and
after that we will try the message signal coming out as usual but there will be also a noise part
which will be coming out, which might be now a complicated noise signal okay because it will
go through the receiver process. and then finally we will try to evaluate the overall power of that
noise which will be after the reception and the signal power that will getting after the reception.

The ratio of that should be my, after modulation and de-modulation that should be my signal to
noise ratio okay, so this is the signal to noise ratio which we will be getting for a particular
modulation scheme. and the signal tonoise ratio for reference we have got,that is we do not any
modulation just base band we transmit, and we will be trying to do is, we will try to create a
figure of merit of a modulation scheme so we call it as FOM.

That should be this SNR which is for that modulation scheme, so for a particular modulation
scheme/ SNR reference okay. So whatever value we will be getting if we are doing fair compose
for everything we will try to see, whether the SNR improves from it is base band or SNR
decrease from its baseband.So that is something we want to see, we want to see which
modulation actually enhances the SNR performance or which modulation scheme decreases the
SNR performance.

So that is why FOM is introduced but here there is something which we have not still talked
about, that is the launch power, so if we do not really equate the launch power for this reference
signal as well as the modulated signal, then we are comparing something to some other thing
okay. We are not doing a fair comparison, because I can put any power in that particular
modulated signal and if you wish to compare them probably that something might be better.

So what we have to also ensure in this process if we really wish to test it, we have to ensure that
the power s(t) that is been launched that must be = to power that we will be putting in the base
band, these two power should be equated. So then we know at the transmitter we are transmitting
equal power then it is going through the entire receiver process and then finally for both the
scheme, which is 1 is base band transmission, because now the power are equal . So whatever
noise is being added to this process, what is the figure of the merit of the modulation.

629
So probably once we do the calculation it will be very clear but this is the methodology will be
applying or adopting towards analyzing this kind of system. Any of those systems. So let us try to
see what we can do for particular system. So let say DSBSC modulation right.

(Refer Slide Time: 11:32)

s(t) = m(t)cos(2π fct)

So the 1st reference part will be I will put this m(t) right and then this particular this message
signal will be transmitted, message signal at it is base band I just transmit it. at the receiver so this
is the Tx part, at the receiver what I will do I will take this signal. Now the signal will be
contaminated with noise. So it should be this m (t) + n(t) and I know that it will additive, so it is +
always, so that should be incident on my receiver.

At the receiver I have a very simple technique we have already talked about that, it will just be a
low pass filter having a band width which is equivalent to that the signal bandwidth, so that
should be –w to + w. so that low pass filter will be, and that should be my output signal y(t),so

630
this is my output okay. So this is for the reference we will be doing. Next I will put a modulated
signal which is nothing but s (t) which is = m(t) cos 2πfct right.

So this I will transmit, so while transmitting, so this is for the modulation, this is for DSB.SC.so
we are now trying to now analyze the whole system for DSB- SC okay. so whenever we are
modulating we need to ensure this power and this power both of them are launching equivalent
power, so this is something that we have to make sure okay. I will receive somewhere so this
should bes(t)+n (t). 1st thing we will do is this is modulated so the frequency will be centered
around ɷc or fc okay.

Sp I have to employee a band pass filter because we have already told that whenever it is
modulated I want to take just my signal which is centered on ɷc with band of 2w right. so I want
also take just minimum amount of noise, so therefore I will put a band pass filter so just the
minimum amount of noise which is centered around fc and 2w bandwidth that will be taken into
the account, so I should be putting a band pass filter which is centered around fc which is exactly
the modulating frequency.

And bandwidth is 2w. okay so that is the 1st thing that we will be putting so this we call as let say
x(t). okay after passing through the band pass filter. Now there will be 1st thing for DSB – SC
what do we do? We first put through a mixture, we will be multiplying by cos 2 ɷct or cos 2πfct.
so we just multiply by this, some how we will do carrier synchronization we are not talking about
that part okay. So this is just the simplest part where the carrier synchronization is differently
done. It is not the costas receiver or some other thing, it the simplest thingWhere separately we
do the carrier synchronization. we are not bother about that right now.

So with that I will get v(t) which will be generated this we are calling as w(t) and then of-course
whenever we multiply by cos 2πfct there will be higher frequency term that will be created which
we have to filter out, so we will be putting a low pass filter, that is actually the overall DSB-SC
demodulation. low pass filter which is again centered, low pass filter which has a bandwidth of
w.and that must be my y(t) or output okay. Now what is happening? This entire process your
noise also will go through this process.

So you have to characterize it properly that whole thing, the whole thing we have to properly
characterise. The 1st process of this is equating these 2 power , so let us try to see how we do
that. So already we have said that my P is, we have already told that, this is something we are
assuming, that the overall signal, power spectral density if we integrate over the band of interest
that is the power we get. okay. Now whenever we modulate, let us try to see what kind of power
we will be getting because we have to also launch similar amount of power okay, so that is
something we will have to do.

631
So let us try to see. when we modulate what we will get? If we modulate I will be getting this, so
that is the s(t) which is m (t) cos 2πfct right. What will be corresponding power spectral density?
So if this was SMF, after modulation we already know from the power spectral density, and
Fourier transform property, that this will be just, if this was my SMF this will be going towards
+fc – fc with the same band of 2w and the strength will just half right.

So basically sorry the strength will not half it is ¼. because mf becomes ½ and square of that
because, power spectral density will be mod square, so it will be ¼, so this will be 1/4 th of that
whole thing or overall if I integrate this must be 1/4th of this thing okay. so over all power of this
one will be 1/4th of this, so if this was P, that should be P/2 this integration will give me p/2 this
integration will give me another P/4. So P/4 over P/4 over here, so if you add that so over all
power should be P/2.

So this is something we know any were you multiply with a cos sinusoidal or sinusoidal you will
be halfling the power, overall power. So immediately we know that the launch power should be
P/2 so this is something we should be aware of okay. Alright, so this is something already
happening over here.Now let say for the base band can be analyze this , so to be more specific or
generalized.

(Refer Slide Time: 19:16)

632
 s(t) = cAc m(t)cos(2π fct)
c 2 Ac2 P
Ps =
2
PN = N0W
c 2 Ac2 P
SNR ref =
2N0W

Let us try putting this instead of just putting m (t) cos ɷct let us try to put a factor into it okay. So
where there is the amplitude of this cos s and there is the modulation factor let say that is c, that is
within our hand and we will be equating the power so it does not matter right. So let us say
instead of s (t) writing as m(t) cos ɷct. We write it as c ac m (t) cos 2πfct. So immediately what
will be the power of it? The power of this signal.

That must be m(t) cos ɷct that must be P/2 and this because it power spectral density, it should be
square of this, power should be p (s) should c2 ac2/ P/2. So that is happens to be if I try to
transmit this particular signal that happens to be the power. What we said already that this power
must be also launched when we do base band that reference signal because we have equate these
two power. we need to actually put the same power. So we need to equate these two powers. so
this is Ps the same power is to be launched for base band.

We will be launching again instead of this launching this P power in the base band will be
launching this amount of power, so which is c2 A c2 P/2 this amount of power, so accordingly
will amplify the signal and whatever need to do we will do that okay, so that the overall power
will become this. So if this the power, noise remains the same right so what will be noise power?
Noise power will be whatever we have calculated so which was N0 W so that is the noise power.

So therefore what is the SNR reference? that should be this power divided by this noise power
already passing through low pass filter, this noised power as been already evaluated passing it
through the low pass filter of band width w. okay. So therefore it should be c2 Ac2 P/2 N0 W. so
that is one part of our calculation. SNR reference has been evaluated, so now let us try to do for
the actual modulated signal right.

For the actual modulated signal, We have drawn it, so this must be my overall de-modulation
process where s(t) is given by this of course c x Ac should be there. okay so let us try to write
that.

(Refer Slide Time: 22:22)

633
x(t) = cAc cos (2π fct) m(t) + nI (t)cos (ωct) + nQ(t)sin (ωct)
v(t) = x(t)cos(2π fct)
= cAc cos2 (2π f0t) m(t) + nI (t)cos2 (2π fct) + nQ(t)cos (ωct) sin (ωct)
1 1 1
= cAc(1 + cos(4π fct))m(t) + nI (t)(1 + cos(4π fct) + nQ(t)(sin(4π fct)
2 2 2

So my X(t), so we will try to write that x(t), x(t) is after passing the signal + noise through a band
pass filter. so that is x (t). What is x (t)? x (t) must have that s (t) and the noise which is pass
through the band pass filter. In the last class we have already characterized a band pass noise.
right we have told that, a band pass noise must have an in phase component, and a quadrature
component.

we have already characterized those things. Those are also random process which has. we have
also evaluated that they do not have any cross relation right, this is something that we have
already, and they are generally orthogonal to each other. okay, this is something that we have
already stated. So we know that the noise must be now represented as in-phase, quadrature
components, so I can write the noise as in-phase, quadrature components. so that is something we

634
have learnt from the random process, band pass random process analysis. Okay, So x(t) must be
my signal because signal through that band pass filter remains in intact'.

Because it is a band pass signal of that same band, which ever I am putting, so signal remains
unchanged, so that should be c Ac cos 2πfct x m(t) that must be my signal + the noise which was
n t, that was the overall noise, now it is pass through the band pass filter. So I can write as nIt
which is the in phase part, x cos ɷct which is central frequency or2πfct, central frequency of the
bandpass filter. +nq (t) which is the quadrature component x sin ɷ ct. We have already told that is
the representation we can get whenever we pass a white noise through a band limit, or band pass
filter. or a band pass noise can be represented as this. So I got this n t, Now what will happen? If I
again go back to circuit, x (t) gets multiplied cos 2πfct right, so what eventually is happening, my
signal, not only the signal is getting multiplied by cos 2πfct, the noise will also be multiplied by
that. So we are actually in the process that additive noise we are taking through the receiver and
every step, we are trying to see what is happening to the noise after doing that step.

So let say, that is v(t) which is nothing but x(t) cos ɷ ct or 2πfct that must be this multiplied by
cos 2πfct, so c Ac cos2 2πfct x m(t) + n I t. again cos multiplied by cos, that is cos 2 2πfct + nqt
cos 2πfct into sin 2πfct. right, so this is something what that we get. So we get the overall
expression after the multiplier okay. Now what we will do? We pass it through the low pass filter;
first of all let us try to see which are the higher frequency term, so this cos square ½ CAc.

So cos2 will become 2 cos2, so I can write as 1 cos 2, 1+ 4 πfct x m(t). similarly I can write for
this one also, n I t,1 + cos 4 πfct. this is again ½ , this is 2cos x sin so that should be sin, n q t,
sin4 πfct. Now let us pass it through. the last part of the receiver circuit is this low pass filter of
band with w. So which are the components of this one, so I have this component, this component,
this component, this component and this component, out of them which are higher frequency
term, which the low pass filter will reject?

That must be rejected, because fc is much bigger than already w, therefore 2fc must even bigger
than that w. so that must be rejected. So this alo must be rejected, and this must be rejected, so
finally after low pass filter.

(Refer Slide Time: 27:45)

635
 1 1
y(t) = 2
Ac m(t) + 2 nI (t)

Let’s say this, Y(t) it remains to ½ CAc x m(t), that is related to message signal + noise term
which is ½ n I t okay. so we have left this two terms, which is very good for us because the signal
and the noise term are clearly separated after DSB-SC. so all we have to now do is, try to
evaluate the power of this one, and try to evaluate the power of this one, and take the ratio of
these two, and then try to see how this is compared to the reference S I N R. So what we will do
in the next class we will just try to see this figure of merit of DSB-SC, okay. thank you.

636
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 45: Noise Analysis - AM

Okay so in the previous class what we have tried is we were trying to actually get a methodology
or propose a methodology where we can actually compared in terms of noise analysis different
modulation scheme, so we have already proposed one where we have told that okay will take
base band transmission and we will take the corresponding modulated transmission and for this
purpose we will try to equate these two power, base band whatever power we are transmitting we
should be also transmitting the same power at the modulation part, so basically our methodology
was for the modulated first calculate the power the same power you put in the base band okay
and then base band we could evaluate because base band whenever you transmit at the receiver
side it is very easy it is just a low pass filter noise will be added in the channel and low pass filter,
so we put actually evaluate the SNR signal noise Ratio and we call that means our reference SNR
okay and then for the modulated we have actually done the entire modulation process for DSB-
SC we have already analyzed it.

We have told that we have the receiver front and we have band pass filter that is of course for
rejection of noise followed by the entire demodulation okay, so after doing that entire
demodulation we just have taken 1 help of this noise representation which is the band pass noise
representation okay or band limited noise representation so that particular in phase then
quadrature we have written after doing the entire demodulation process we could see that I
ultimate output.

(Refer Slide Time: 02:11)

637
 1 1
y(t) =
CAc m(t) + nI (t)
2 2
1
PS = c 2 Ac2 P
4
1
PN = N02w
4
1 2 2
4
c Ac P
(SNR)DSB-SC = 1
Nw
2 0
c 2 Ac2 P
=
2N0 w

Of a DSB SC signal should be this is something we have already prove that okay, so now let us
try to see how do we calculate the signal power so the signal power must be coming from this see
mt gives me power of p and that is multiplied by c AC and half so power spectral density if you
just go and then integrate I should be getting one forth C2 AC2 and P right, so that must be the

638
signal power after demodulation now I have to get the noise power that should this power okay
which is nothing but 1/4 because there is a 1/2

And the integration of power spectral density of ni, this is something we have already
characterized in our band pass random process analysis, so we have told that if I have a
corresponding n which is going from that center frequency minus fc +fc and this by this 2w we
have told this ni or nq will be nothing but this shifted to this side plus shifted to this right okay, so
if this strength was eta 0/2 both of them will be shifted and the overall strength will be eta 0 or n0
so this be n0 define from – w to + w right.

So that is my ni or the power spectral density of ni s ni' (f) so what's the overall noise power then
it is just the integration of the power spectral density so that must be N0 into 2w right, so that is
my noise power there is a factor of ¼ so it should be ¼ into N02w so that is actual so this if I
write as this so that is actual noise power right, so what is my signal to noise ratio for this
modulated scheme, so if I just say SNR for this DSB SC that must be this power ¼ C2 AC2 P /
this noise power which is nothing but ½ N0 w.

So that 2 gets cancelled I get C2 AC2 P / 2 N0 w so that is the SNR for DSBSC if you just go
back and try to check the reference SNR was exactly the same okay the expression was exactly
the same so therefore the figure of merit must be 1.

(Refer Slide Time: 05:33)

639
 (SNR)DSB-SC
FOM = =1
(SNR)ref
s(t) = Ac [1 + Ka m(t)] cos (2π fct)
w(t) = s(t) + n(t)

Must be this SNR for DBSC / SNR of the reference because they are exactly the same. so it
should be 1. so for DSBSC in presence of noise it gives me no degradation or no advantage
compared to base band transmission.this is something we have understood, whatever the SNR I
get when I do base band transmission the same amount of power if they transmit whatever
advantage I will be getting,DSBSC does not give me any degradation or any advantage. okay, so
this is something we have now figured out for DSBSC that something which happens okay.

So next what we will try do is, we will try to see.if the am version of it where you do not suppress
the career, is it having similar thing. okay or is it something else then only we will be able to
compare whether noise performance wise these two modulations schemes are equivalent or one
has some advantage over the other, so let us try to do the noise analysis or amplitude modulation.

640
process will be same so the process must have a base band equivalent transmission which we
have already seen.

okay and there should be a modulation. Now the modulation our signal is something like this,
you already know this. okay, so where this KA is related to the modulation index, they define this
when we did our Am and this is the amplitude of the career signal. okay so basically what we do
nothing but we actually add a DC to the signal and then modulate, or do our DSBSC. or
otherwise in a way you can see it as the DSBSC + the career part if you just multiply this that
should be the carrier part.

so the career has been also added Over here. so this our modulation corresponding de-
modulation, the de-modulation of am was easier, what we did is something like this you have this
St + noise which let say wt coming through the channel, at the receiver what you do you first put
any receiver will have band pass filter that is guaranteed, because you want to reject the noise as
much you want if it is a low pass signal you will have a low pass filter, if it is band means signal
having some band at a particular band, then we have to put a band pass filter so you put a band
pass filter followed by just envelop detection. this was our amplitude demodulation right, so we
will not change that circuitry. so that must be our yt. okay, so let us now try to see what kind of
power we are talking. first of all, see anywhere the techniques are same we need to launch same
power for the base band transmission as well as the modulated transmission. okay, so that is why
because the modulation transmission is will be completed to calculate.

First We will calculate the modulator transmission power and that same power will be put in the
base band. Earlier also we done same thing, so here also we will do the same thing.

(Refer Slide Time: 09:40)

641
 Ac2 Ac2 Ka2 P
power = +
2 2
Ac2 [1 + Ka2 p]
SNR ref =
2N0W
v(t) = [Ac + Ac Ka m(t)] cos (2π fct) + nI cos (2π fct) + nQ(t)sin(2π fct)

[Ac + Ac Ka m(t) + nI (t)] + nQ(t)

2 2
y(t) =

so what is the power that is launched whenever we do this modulation. okay so the power that
will be launched. see this AC Cos ωCt, that is the Cos sinusoidal signal with amplitude AC. we
know that the power of that is AC^2 / 2 right this is something we have already done, so that
must be AC ^ 2 / 2 +. this AC into KA that is already there, right, Earlier also we had done this, C
into AC right, so instead of C it is KA so that must be AC^2 into KA and this mt, if that has a
power of PI then it should be' because it is multiplied by Cos, so it should be P / 2.

So I can write as AC ^ 2 K2 P / 2, this is the power if I modulate will be launched.so what I have
to do, I have to equate the power, so therefore this is the power I will be also putting in the

642
baseband. so base band power is this. where is the base band noise, that is equivalent to the
previous case because it is just N0 / 2 from – w to = w. so overall power noise power is N0 / 2
into 2w which is N0 into w. so therefore our SNR reference must be this divided by this, so it
must be AC^2 that it is common 1 + K2P / 2, N 0w right.

so that is the reference SNR that we get now we have to do the envelope detection. so we will
again employ the same thing, so I have this signal plus noise it should be pass through a band
pass filter. this band pass filter because it exactly matches with the band of this one, so ST will
remain intact, the noise will be passing through the band pass filters so it will have a equivalent
band pass representation-okay, with in phase and quadrature. so therefore that after the band pas
filter, if I call that signal as Vt.

That must be our AC + AC KA mt Cos 2πfct. right, so this is the message signal or modulated
signal plus n I t Cos 2 pi fct + nQt Sin 2πfct just the noise gets our in phase quadrature
representation. right now next what will happen, this particular Vt will go through envelope
detection, that means whatever I have, this is a sinusoidal or this can be termed as a sinusoidal, it
is envelope will be tracked. so I will all I have to do is, I have to now see what is the envelope of
this one which has a frequency of fc.

Because this as this career which is a very high frequency. so it is actually the underline
frequency term is Cos ωCt we just need to see the envelope of that. okay, there is a Cos term and
there is a Sin term. we know how to calculate the envelope that should be the coefficient of Cos ^
2 +co-efficient of Sin square, square root that must be the envelope. we have already derived this
path. okay, so let see what is the coefficient of Cos. coefficient of Cos is AC, AC KA mt + nIt.
right so that is the coefficient of Cos.

So this square plus, coefficient of Sin is Sin 2πfc is nQ. so nQt^ 2 and then the square root of that,
that must be the envelope. so this is my yt, then because the envelope detector will exactly detect
the envelope of this career. okay so that must be my overall envelope. Now we will insert
something, okay because this is the complex this one now you can see the noise square, square
root, all of kind of things are coming. okay so it is no longer or very nice linear thing. like we
have assumed or we have we could see for the DSBSC, so it going to be harder to analyze okay
but we will take a very simplified assumption over here.

(Refer Slide Time: 15:09)

643
 nI (t), nQ(t) ≪ Ac
y(t) = Ac Ka m(t) + nI (t)
Ac2 Ka2 P
SNRAM =
2N0 w

we will say that this noise term generally that should be the case this. This ni, both ni or nQ
which is the in phase or quadrature component that must be much less than this AC. okay so this
is one assumption we are taking. with that assumption, generally you will have to put a high
carrier. so if you are doing that, as long as this is valid, you can do some analysis. so let us try to
see what that would be. so immediately yt I can write, so we have this square plus this square,
this particular term as AC^2. so this will be much lesser compared to this particular part.

Okay, because this is already lesser then square of that will be much lesser, so I can neglect that
part. so immediately if I neglect what will I have, I will have a square and a square root that we
cancel out each other. so I will get this AC + AC KA mt + mIt so again we get back similar
things, where this is probably the signal part, and this is the noise part-this is the DC part we have
already seen that after envelope. generally you will be have a DC blocker which will block this
part.

644
So this could be already gone will have just this power and this power so now we have to again
evaluate the signal to noise ratio, okay the signal power is how much? it is actually AC^2 KA^2
into P / 2. so that is the signal power and the noise power is this nI whatever that is! okay, so nI
we have already evaluated that, that must be, its strength it eta 0 going from – w to + w. so that
must be 2 N0w. right, so this, if I put I will be getting AC^2 K^2, oh sorry this two, while I have
put there should not be any 2, because there is no modulation term it is just directly mt so it
should be, this P / 2, N 0w right, so that is the SNR of Am now if you wish to calculate the
overall figure of merit.

(Refer Slide Time: 17:55)

(SNR)AM
FOM =
(SNR)ref
ka2 P
=
1 + ka2 P

Okay, figure of merit of this modulation, so figure of merit will be first we have to put SNR of
Am / SNR of the reference. SNR of Am we have got AC^2 K^2 P / 2 N 0 w and SNR reference

645
we have got as what was it AC^2 1 + K^2 P / 2 N 0 w. AC^2 gets cancelled 2 now gets cancelled,
so you are that with AC^2 P / 1 + K^2. right which is for sure, less than 1. because I already have
a K^2P I have something added to it, so definitely I know the figure of merit of this one is less
than

1. okay, so what does that means that actually says.

That after doing this analysis we could understand that probably DSBSC is better noise resilient
compared to Am. because m to get similar amount of signal to noise ratio, will have to launch a
huge amount of power. okay because that career power are putting that is wasted, and that is not
really giving me any help in terms of noise cancellation right, it is not boosting the signal power
which will be the actual signal power, that will be demodulated, so because of that we have a
degraded signal to noise ration performance. okay so if we just give on example probably.

(Refer Slide Time: 20:02)

646
if m(t) = Am cos(2π fmt)
s(t) = Ac[1 + μ cos(2π fmt)]cos(2π fct)
μ = Ka Am
1
P = Am2
2
ka2 Am
2
ka2 P 2 μ2
FOM = = =
1 + ka2 P 1+
ka2 Am
2
2 + μ2
2

So if we take our mt to be some Am, just a tone modulation okay so if we just take that, so mt is
also a sinusoidal so basically we are modulating our slow varying sinusoidal with the faster
varying sinusoidal . which we have also done for demonstration purpose right, so this is usual
practice we just want to see what happens. okay and let say our modulation index in that case is
something like this 1 + µ, so this is the modulation we employ.µ is the modulation index okay, so
if this is the case now what is µ this is actually KA into Am.

I can write in this way and my power is ½ Am ^ 2 because it is a sinusoidal, so the power should
be amplitude square divided by 2 right, so this is my overall power. okay so therefore the figure
of merit that we have already calculated that is K^2P divided by 1 + K^2P. which if we just put so
KA okay we can put it in terms of µ so K^2P if I just put ½ Am ^ 2 so I get KA Am ^ 2 / 2 1 +
KA ^ 2 / 2 that is nothing but this is actually µ ^ 2 / 2 + 1 + 1 µ ^ 2 / 2 or I can write as µ ^ 2 2 + µ
^2. what is the highest value of µ that I can take ? one.

we have already seen that for proper Am modulation, so that we have enough DC shift it does not
do any 0 crossing. so to do that, so that envelope detection is proper, we need to maximum µ I
can take is 1. so if I just put that maximum µ, I can immediately get figure of merit to be one
third. so that is the maximum I can achieve. so it will be whatever happens for a tone modulation
we can already show that this is the best figure of merit we can get it can become even worse
than this.

This is the best we can get. so this is the best we can get. so this one third time worser than the
corresponding DSB-SC modulation. okay, so that something we wanted to say at, now with the
noise analyses we can actually insert that okay one of them is better and the other one is probably
not that better. what will do next is, means similarly you can start doing analysis for everything. I
will just demonstrate one of them, because it is little more critical so that is why I will
demonstrate the other one which is the SSB.

(Refer Slide Time: 23:37)

647
 s(t) = m(t)cos(2π fct) + mh(t)sin(2π fct)
Mh( f ) = jM( f )sgn( f )

{jM( f ) if f < 0
−jM( f )if f > 0
=

m(t) ⇔ M( f )
w

∫−w
Sm( f ) d f = P

mh(t) ⇔ Mh( f )

Okay, so single side band of-course, suppressed carrier. so this is something I will try to
demonstrate next and what will try to do is, this will almost give us most of the tools that are
required while doing this analysis. if you wish to do for VSB, it is almost similar, only thing is
that, for the filtering you have to take a proper care of that, because VSB has a different kind of
filtering it is no long or a sharp filtering. okay, so accordingly the noise and all those things will
be accordingly at just it we will have to do that okay.

648
So I will not demonstrate that VSB part, but probably the SSB I will demonstrate because we will
see something more critical will come up. with the noise processing. okay, so for SSB what we
need to do is again the similar thing. okay so we have to actually define a signal for SSB which is
nothing but mt Cos2πfct + mht Sin2πfct. this can be + or – depending on upper side band or
lower side band which we are transmitting. okay and mht is something we know already it is the
Hilbert transform of our signal, right.

so this is something we already know. so the characterization of that is if we write the frequency
response of mht is mft or we should not say frequency response, the frequency representation of
that or the amplitude spectrum of that so that is the mhf that must be we already know that should
be j Mf Signum f. This is something we have already defined. which is nothing but plus – jMf if f
is > 0 or + jMf if f is < 0. right, one of them you can make equal okay so this is something we
already know.

Why where telling all these things, because we will have to finally calculate the power. to
evaluate the power I already know the power of this mt. mht I have to now calculate. okay what
is the corresponding power of that. because otherwise I want to b able to evaluate the overall
power that will be transmitted. right, because whenever I see we transmit about that signal I have
to first get the power and that power I will be putting in the base band, so that something we have
to first characterize okay.

So to evaluate the power let us try to see how do you do that. we know that mt has the
corresponding Fourier transform has mf and it as a corresponding power spectral density which is
Smf. right, and you also know that –w to, it is a band limited signal. and – w to +w Smf this is
something we have we are keep on telling that this is p. that something we know.

Okay, now let us try to see what will happen to mh. so immediately we have already told that
mht, that must have a Fourier transform. which is mhf okay, which looks like this. now let us go
to the power spectral density. so whenever you will be evaluating power spectral density.it is a
modulus of this square, you do the more or less this +j or – j has no effect. so you know that the
power spectral density, whatever happens it will be equivalent to whatever we get through mf.
okay so both the power spectral density will be equivalent where.

Amplitude spectrum or I should say, means amplitude and phase spectrum will be different,
amplitude spectrum will still remain the same, only phase spectrum will be different. but if I go to
power spectral density, they have exacted no difference. okay, so therefore for this one also, if I
just write that as Smh f. so that also I can insert that is band limited, and –w to w if I integrate,
that must be, also P. this is the extra information I get. okay, so what will try do in the next class
is taking these two information.

649
We will try to first evaluate the power of this. okay, and then we will try to see what is the
corresponding overall noise analysis can be. okay thank you.

650
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-46: Noise Analysis – SSB- SC

So we have started doing the analysis of this SSB-SC right the first task is as we are seeing
through the last two lectures that the first task is evaluate the power of the modulated signal that
is the most important part. In SSB we have already seen, if I just put the previous light. So what
we have seen that modulated signal is empty cos ɷct up to that it was all fine but it also has either
+ or - mht sin ɷc t or 2 π fct so evaluation of power becomes more critical over here because for
mt we have a definition the power I know but mh (t) we do not know.

So that is why we did some derivation step to actually realize that the power of that or the power
spectral density of mh (t) is same as the power spectral density of m(t). Therefore the overall
power because they are also similarly band-limited, so the overall power will remain P. So now
for the modulated signal if this is P, the corresponding multiplication with cos that must be P / 2
and if this is also P the corresponding multiplication with cos or sin whatever you multiply the
overall power will be again 1/2 so it should be P / 2 so P / 2 + P / 2 it is the overall P power that is
being transmitted. So this is something we could know after doing all these things.

(Refer Slide Time: 01:54)

651
 Ps = P
N0
noise power = 2w
2
P
SNR ref =
wN0

So this s(t) corresponding power or Ps that is actually P, that P is, the way it is defined mt
corresponding power spectral density if you integrate from- W to + W that is something you get
okay. So now we know the power of it at the modulated level, so therefore the baseband also
must transmit that same power P fortunately here we are getting exactly equivalent power okay.
So if I just transmit mt whatever power it will be that is the same power we are getting okay.

Now this is the power will be transmitting what is the noise again same noise so the noise power
that must be it is a low-pass filter from - W to + W with strength n0 / 2, so 2 W into n0 / 2 which
is nothing but W n0 right so therefore SNR reference must be P / WN0 so far it is quite easy. Now
let us try to draw the receiver chain because now we will have to go through the receiver and then
try to see at the output what happens what the corresponding noise and what is the corresponding
signal power okay. So receiver chain if I just draw it. let us do it in a new page.

652
(Refer Slide Time: 03:31)

w(t) = s(t) + n(t)

v(t) = m(t)cos(2π fct) + mh(t)sin(2π fct) + nI (t)cos (2π ( fc + w/2)t) + nQ(t)sin (2π ( fc + w/2)t)

So that is let us say w(t) which is that signal s (t) that we have defined which is a modulated
signal + noise this will be incident on the receiver, so that goes to the receiver. okay, So in the
very beginning of the receiver you will have a multiplier, SSB also more demodulated in a
similar way as DSP is demodulated. so you will be putting a multiplier, but before that we have to
employ a band pass filter. so this should not be the first part, there should be a band pass filter.
Now the important thing, this particular band pass filter the band is no longer 2w. because in SSB
we have already restricted it to W.

Because we have taken either the upper band or the lower band. so the corresponding modulation
whichever way it looks, it will be either from fc to fc + W and- fc to- fc - W, or the other way.
okay one of the band will be taken, so therefore if my interest region is only this I will be also
employing a band pass filter which is only of this width. So instead of putting a band pass filter
of width 2w, I will be now putting a band pass filter of width W. because I want to cancel as
much noise I can.

653
If I can do that why should I waste my this thing. and why should I take some more noise, so
because it is a more spectral efficient modulation scheme, so I know that I can employ a smaller
bandwidth filter. okay. and there is also another thing what is the center frequency of this
particular filter. that is no longer fc. that is actually fc + W / 2, so this is a big change which
happens whenever you employ SSB. so the field that filters center frequency is now changed,
what will be the consequence of this?

The noise will be characterized in a different way because that is the frequency now of the band
pass, that is the central frequency of the band pass noise, so therefore if I put it as in-phase and
quadrature component the central frequency will not be ɷ c. it will be fc + W / 2 or 2 π into fc +
W / 2. okay so that is the change which will be happening in SSB. okay,So all right I pass through
bandpass filter so what do I get, I will be getting s(t) will remain as it is because s (t) is exactly
passing through this filter which is designed for him only.

So therefore, if suppose I call this to be V t so my V t should be, that should look like mt so st
will remain the same cos 2 π fct + mHt sine 2π fct, so these two things are there. okay, plus, now
the noise part okay, noise is a band Limited noise. so therefore it must have a in-phase
component, but the corresponding cos should be at this frequency. so 2 πfc + W / 2 + NQt sine 2
π, sorry, the t was missing over here.

Right, So this must be my VT. slightly modified because of the filter characteristics. so means
whenever you are doing noise analysis, it is very important that you understand all this process.
okay, and you characterize the noise properly, that is the most important part, because otherwise
the calculation will be wrong. So after that what do we do, after passing through the band pass
filter will be multiplying /cos 2 π fct.

(Refer Slide Time: 08:00)

654
 (2) 2 (2)
1 1 w 1 w
y(t) = m(t) + nI (t)cos 2π + nQ(t)cos 2π
2 2
SNQ( f ) = SNI ( f )
N0W
8

So this Vt will be now multiplied by cos 2 π fct which is nothing but this mt will be multiplied by
this cos squared 2 π fct + mh t sine 2 π fct x cos 2 π fct + n i t cos2π FC + w / 2 into t into cos 2 π
fct + n cube t, into sine 2 π fc + W / 2t into cos 2 π, right, this is what will be happening.
immediately I can simplify this because this is cos square, I can get half and then I will get two
terms, one is mt and the other one will be mt cos 4 π fct. Similarly I can also take ½ 'this will be
mht sin 4 π fct.

So these are all higher frequency terms which will be cancelled by the next low-pass filter now
these two, it cos a cos b. if I just take half a cos a + b+ cos a - b. so I can write ½ nit cos of a + b,
so that must be 2π into 2fc fc + W / 2 t + 1/2 n It. I will have cos a - b so that must be2 π W / 2 t
right I will have this is again a higher frequency term this will be canceled. Similarly for sine
also, I will have because it is sign a cos b so it should be sine a + b + sine a - B. so it should be ½

655
n q t, sin a +B means same thing. so it should be 2 π2 fc + W / 2 into t. + 1/2 n q t this will be sine
a – B.

So that should be signed 2 Π W / 2 T. this will again get cancelled by the low-pass filter. so after
the low-pass filter will be left with half mt +, I will have this half nit cos 2 π W / 2 x T + 1/2 nQ t
sine 2 π w / 2 into t. right, this is what we get at the output, so that is my yt. now clearly you can
see that has a message term and noise term. So this is the first time because of this filter central
frequency shifted we get both the in-phase and quadrate term in my output noise. okay, and now
we will have to probably get the overall ponce spectral density of this noise.

Now though all those things will be useful that they have no cross correlation. because if they are
there are cross Correlation, there will be cross term which will be generated.but that thing will be
we already know that not probably there.it is all the events symmetric spectrum and they do not
have any cross correlation term, okay and they are orthogonal to each other. So basically the
overall power spectral density will be just power spectral density of this + power spectral density
of this one.

Okay so we will have to first evaluate the power spectral density of this one, so let us try to see
what is the power spectral density. first of all what is the ni, so this was actually, let us try to
evaluate this so this was at fc this is my n or s n f. I am trying to draw okay so this was as fc or -
fc and this is at - fc - W this is at + fc this is at + fc + W right. So if I now try to evaluate the ni
that should be my central frequency. this will be shifted over here and this will be shifted over
here the strength was n 0/ 2. right so in I or n cube will be going from - W / 2 because it is the
central frequency is basically fc + W/ 2.

And that much shift we give so therefore it will be going from - W/ 2 to + W / 2, so - W / 2 to +

W / 2 unlike the previous one. and it goes up to n 0. right, there it gets added. So this is my SN I
F or S and Q F both of them. okay, Now let us just consider this one. now nIt is getting a
modulated term. so I will have to first get a modulation term on this. how much modulation W /
2. So there will be a shift of W / 2 in this side as well as this side. and because it is a
multiplication with cos, so it must be 1 / 4, so what do I get?

So this n 0 will become n 0 / 4 and will be shifted to W / 2. so this will be shifted to W / 2, so

from then it will occupy 0 to W with a strength of n0 / 4. and shifted on the left, so that will
occupy this part, it will come up to W and the strength will be again n 0 /4. so that should be the
spectrum.this is fine, n i cos 2 π w / 2t should have a this spectrum. and why I should consider
this whole spectrum because my filter that low-pass filter is having a filter cut off at W. So
therefore this entire spectrum as noise will actually enter through the low-pass filter.

656
So that should be the case. what is the corresponding power? That is actually n 0/ 4 x 2 w. so n 0
W / 2 fine. you have a half term over here, so I will get 1 / 4 so noise due to this, noise power due
to this will be this much. right same thing will happen over here, either you multiply / cos or sin,
in power spectral you will have no effect. similar thing will be produced over here because the
frequencies is also that shifting frequency or translation frequency is same. So it will get again
this is actually n 0 W / 8 right, so the other one also will be n 0 W / 8. overall it should be if I just
add these 2, n 0 W / 4.

(Refer Slide Time: 16:49)

N0 w
PN =
4
P
PS =
4
P
SNRSSB-SC =
N0 w
P
N0 w
FOM = P
=1
N0 w

657
So I do get overall noise power which is n0 W / 4. this is fine, so this is my noise power. what is
the signal power? mt therefore it should be P, and 1/2 factor is there in power spectral density it
will be 1/2^2. So it should be 1 / 4 P. so the power is P / 4 so the signal-to-noise ratio for SSB
must be P. right, now the good part, what was the reference signal to noise power ?exactly the
same. So therefore my figure of Merit must be this P / n0 w divided / p + 0 W, again we get a
figure of merit of 1.

So even though SSB is a band width efficient modulation scheme, we will be expecting that
probably it will be taking less amount of noise, so it must be more efficient than DSB-SC. but
after the calculation we could see that is not the case it is exactly equivalent efficient as the
baseband and as the DSB-SC, there is no difference in these two, and both of them are equally
efficient compared to our amplitude modulation. in terms of noise analysis we are just talking
about noise analysis, but of course SSB has other advantage because it uses the frequency in a
better way.

Because it just with the same frequency it can transmit two signals. potentially two similar kind
of signal. because it just takes the half frequency of DS B- SC. right, But we could see noise
analysis wise, there is no difference this is a very fundamental result that comes generally it is a
counter intuitive result. because we expect that because it is more spectrum efficient. So probably
it will if I put a employ a proper filtering, probably it will take less amount of noise, so it must be
more efficient people often also do another mistake while calculating the power of SSB signal.

(Refer Slide Time: 19:45)

658
So what people do you think about a filtering method. okay, so this is a mistake people often do
of course it is it should not be done. So if you just think about filtering method. what happens
your DSB will look like this. and then as if half the power will be transmitting. okay, what we
have seen that SSB exactly transmit P power. whereas if you just calculate it this way. you will
think that this is already DSB so it must be P / 2, and you are transmitting half of that power. so it
should be P /4. that is not the case.

Because in a way you have to think that whenever you employ a filtering, you are actually
rejecting a band that means you are wasting the power, so this Pby 2, basically getting
counteracted by the filtering filter is rejecting that that means you have to actually generate more
power to produce a certain power in the modulated signal. okay suppose you are targeting P out
to be transmitted through the modulation process. Then you have to actually generate 2 P power
over there. then only after filtering you will get P power.

so that's a common mistake I have seen over the years, people Just do, while calculating the
power of DSB SC, sorry SSB, they just keep doing this twice halfing of power that they say okay
it is DSB SC so what all power gets 1/2 because I am multiplying by cos.I again employ a filter
which takes out half of the power so it becomes P by 4. that is a very common mistake which
people often do so we should not do that. And if you do it from the other side, which we have

659
done you will see that actually a P power, which is to be launched.so this probably ends our
discussion of noise analysis.

Because the other thing that can be done we have also discussed about two other modulation
scheme. one is quadrate amplitude modulation QAM, and the other one is VSB.VSB it is a little
bit more trickier. because VSB will not have ideal filtering, and therefore the noise calculation
will not be like that very simple integration okay. So it will have a particular role off filtering. and
that filtering has to be considered whenever you are doing that and accordingly the overall
calculation has to be done because the input filter and output filter for VSB will be typical and
that typical filtering has to be taken into account to actually evaluate the overall noise power that
will be linked to your modulation.

so that is something you can take as homework. so analysis quadrate amplitude is just almost
equivalent to DSB. because it is just two DSB-SC simultaneously you are putting. then they will
have both the noise and then you just multiply by cos and sine you have to just see that. So it will
be almost similar toDSB SC analysis, there will be no difference in to that okay so it will it will
almost be same. So after doing this probably we have we told that we will be comparing all these
modulation schemes in terms of various aspect.

So now just the one part which was left when we did that comparison chart that was the noise
analysis, now we could see the noise in terms of noise how they perform. so probably DSB SC
and SSB are the best in terms of noise, whereas AM is not that good, noise wise and VSB you
will see that will be little bit in-efficient because we will have a roll off filter, so that will take
additional noise. so it will not be that as good as that quadrate amplitude will be similar to DSB
as we have told already, so it takes means the modulation and demodulation process is almost
similar.

So after doing all this one thing you have to keep in mind that probably we have said over here
that for SSB analysis, we have already told that noise analysis wise it is as good as DSB. there is
a gross assumption over here which probably we have not stated implicitly, we are told that this
SSB can be demodulated with a fresh carrier, and we have already seen in our earlier classes that
SSB carrier recovery is not that easy. So therefore generally in SSB either people add huge
amount of carrier, like amplitude modulation because the carrier recovery is not as easy as DSB
SC.

So huge amount of carrier once you add that SSB noise analysis also will be as poor, or even
poorer than your amplitude modulation. because you will have to add a huge amount of carrier to
really and then employ envelope detection, so it will be similar to that particular process, if
fortunately you have carrier/ some method either by transmitting pilot carrier or some other

660
method, because from the signal itself you cannot extract the carrier because that Hilbert
transform and the signal will complicate the carrier recovery process.

So carrier recovery cannot be done as it is like putting a Costas loop or something like that so
there you will have to probably if you wish to do a coherent demodulation, you will have to either
put a pilot carrier or your carrier somehow in a miraculous way, should be synchronized. so that
is something which has to be done. okay, So we have for our analysis assume that probably that π
lot carrier is available. and that is why SSB is probably looking to be as good as DSB . but that's
not the case, we should be always keeping that carrier recovery is a very important circuit which
we have just taken that it is recoverable probable.

Which is not true, mathematically. we have already proven that, and this completes our amplitude
modulation schemes and their analysis, complete analysis. okay, now what we will do from the
next class onwards, we will start looking at the another very important modulation which is
called angle modulation. okay, where the amplitude will remain fixed. whereas over here, you
might have seen that we are just putting all the things all the message signal in the amplitude of a
particular sinusoidal.

In the next class onwards we will try to see that will keep the amplitude fixed and we will put the
message signal in the frequency part or phase part. either will vary the frequency of the sinusoidal
with respect to our message signal. so the modulation will be done at the frequency level or will
vary the phase part of it and try to modulate it. okay, So that is overall these two things are almost
similar you will also means we will demonstrate that so you will see that corresponding
modulation is called angle modulation.

And especially the frequency modulation part is famous as FM or frequency modulation. and
phase modulation is called p.m. but p.m., FM effectively they are almost similar things. the FM
will be also able to prove that it has a much better noise immunity. okay, so that is something we
will try to prove with our noise analysis again we will try to give the full-blown analysis of FM,
means how it performs with respect to noise or in presence of noise. and how it is much better
than the a.m. noise performance.

So that is something we will discuss, and historically probably that is the reason why FM become
more important because it has a better noise cancellation technique and it survives, means it
keeps the quality of the signal much better than AM. so initially people started with AM
historically, but then people could see that FM is much better modulation technique, so people
went to FM. but FM initially was not being accepted in a glad manner by the community. that is
because there is a, there was initially means a particular doubt in people's mind that FM might
have a very huge bandwidth.

661
So we will also try to do that particular part of analysis. that is really what people have thought
that FM bandwidth is probably very huge whenever we do FM modulation because the frequency
start varying. and then, effectively people initially said that FM probably has infinite bandwidth.
we will try to prove that and we will try to see that probably effective bandwidth is still not
infinite. it is band limited still. because if you can only transmit one FM through the entire
channel, then it is not good .

Because all others transmission will be canceled out, but fortunately that has not happened and
this was mostly proven by a person whose name is almost synonymous with FM that is
Armstrong. so we will try to see what he could prove and make FM fundamentally very
important modulation scheme in the next class. thank you.

662
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-47: Frequency Modulation

Okay so we have already discussed about the noise in AM. so we have almost finished whatever
we wanted to do in AM. we have given a brief introduction in the last class about FM. so today
what we will do, we try to explore FM more. there were a few controversy around FM. so
initially people thought probably with FM there is a huge bandwidth that advantage. so they, they
thought okay probably FM will give us the smallest bandwidth that is possible.

And it was the initiative of bandwidth, means you know that DSB was there, then SSB came, or
QAM came, so people were thinking of a more bandwidth efficient modulation scheme. and then
people came up with FM and they thought okay this is probably the most bandwidth efficient. I
can actually vary my bandwidth as I wish, and I can make the bandwidth almost infinitesimally
small.

We will demonstrate that.okay, why people have thought. that but that was a mistake probably. so
initially FM become very popular, that they can actually save band width. by this! then finally
when people started doing FM they could realize that probably that is also something will prove,
that FM has infinite bandwidth. instead of means controlling the bandwidth to the extent that it
will be almost tend to 0.

And we can make really multiplexing hugely efficient by putting so many FM carriers. people
realize that one single FM is almost taking in finite bandwidth. so immediately what happens the
faith of FM was very bleak. people could understand that probably this is not the modulation
scheme that we should try for. and then the bandwidth problem was in a way solved people could
realize that okay.

663
It is neither 0 or tending towards 0 and nor tending towards infinity it is somewhere in between.
and we will try to prove that part because FM bandwidth has bothered people, researcher
demonstrator and implementer alike for long time. so we will do that, that is one of the major
thing. and then people thought okay it is, it is a bandwidth not bandwidth efficient scheme and
they could prove that it is much lesser efficient compared to DSB-sc.

So it was even worse than DSB Sc see of course it is not that worse that it takes in finite
bandwidth but it was worse than DSB SC people could realize that, that the boundary I can get is
not zero bandwidth so it will be at least DSB sC and it can be even more so immediately the idea
of FM was having a very means negative response from the entire research community and entire
implementers.

And then Armstrong, thanks to Armstrong actually FM become so popular. So he started

inquiring other thing' that okay' communication we generally try to see bandwidth and all those
things that is all fine bandwidth efficient whether it is energy efficient and all those things, but
there is also another thing that you put a signal for communication and then, then there is noise,
there are interference there are all those impairments that we have discussed which comes from
the channel. channel non-linearity and then he started understanding whether FM has a good
response in terms of those things.

And then he could realize that that is also something we'll be proving, that FM has in potentially
very good response against all these impure impairments-channel non-linearity interference and
then noise in particular. so that is the point where FM started becoming again popular but
Armstrong had to really fight hard but that was legal battle. and nobody was trying to accept his
technique and then he had to fight a very long legal battle and then finally he committed suicide
also because of that. because that took a toll from him, but anyway, because of him probably and
all his battle if FM become the winner and then people could see that all the radio actually started
becoming FM.

Because you might have also experienced if you have heard about Vividh Bharati in earlier days,
and then all FM channels FM has much better clarity so that is because it is more prone to noise
interference channels a non-linearity and all those things. so it always transmits data with much
bigger clarity it is easier to receive the data in its original form so all those things actually helped
FM to become a much, much more popular broadcasting channel.

So what we will try to do, this is just a summary of what we are targeting in FM but we will try to
see all these points that we have discussed that the bandwidth issue. Then the channel non-
linearity how FM is better than better in that aspect, all those other channel impairments like

664
noise and interference so we will try to prove all those things. but before that the first task is how
do we really realize fm.

(Refer Slide Time: 06:04)

A cos(ωct + θ )
(ωc(t)t + θ0)
θ(t) = (ωct + θ0)
t1 < t < t2
Δt = t2 − t1
θi = (ωit + θ0)

What is the basis or genesis of FM? So let us try to see that part. So we have already told that it
all started with the parameters of a sinusoidal right, it is all carrier modulation so we need to see
what are the things we can modulate. So we have said that a sinusoidal or co-sinusoidal it is
defined by three parameter. right, so one is the amplitude the other one is the frequency and other
one is the phase. okay, so whenever you do amplitude modulation that is something we have
covered already.

665
When you start doing modulating these two parameters will also prove that they are probably
similar, so if you start doing that modulation probably that is when the angle modulation comes
into picture. that means either the ωc will be linearly varying with respect to the message signal
or the theta that is there will be linearly varying with respect to the modulating signal.like the
amplitude was earlier varying with respect to the modulating signal right.

If you just think about DSB SC. so this was the target so then people started thinking that is FM
and a means p.m. and that phase modulation or frequency modulation are there almost two
similar things ? so for that we need to establish the relationship between phase and frequency. so
if we start varying frequency or phase what we will see is overall this whole thing inside will be
varying with respect to time okay so let us just call this as a overall phase, okay which is like Co-
sinusoidal what value it takes the angular value that it takes inside.

So that is the phase. it is a co sinusoidal of that particular value. for a signal which is having this
varying this parameter with respect to time. let us plot that. okay so with respect to let us say time
we start plotting this so what will happen. this particular phase will of course we can see with
respect to time it will increase. okay because there is a function T and there will be some
variation because this ωc might vary with time theta might vary with time.

So there will be some variation. so let us say probably this particular part, is right now not
varying. okay, so let us say it is like this ωct + some θ zero okay, so entire variation is over here
so we are trying to capture the FM part. okay so because this ωct that is also a function of T, this
whole thing will now become a random variation of this particular thing. okay, so therefore if
Omega C also was constant I could have expected something like this. a linear curve because
then that theta T would be ωct + θ where θ 0 is a constant ωC is a constant.

If that is the case, then θT must be a linear function of T. so it should be like this where it cuts at θ
zero, and the slope is ωC. right, this is something what we know okay now what will happen if
ωC start varying with respect to time then there will be a phase response which will be something
like this. okay so let us concentrate just for the time being a particular time T1 and a particular
time which is close enough which is called T2 and we are trying to concentrate on this interval,
that is this t, between this T 1 and T 2 okay.

And what we also do that separation we call that ∆T. that is T 2 -T 1 we make it infinitesimally
small. so almost tending towards 0 if you just do that, then add that at a particular instant let us
say at T1. what will be the instantaneous frequency? okay, so that is something we are trying to
evaluate so at that point this must be the phase, let us say that is θ I, so I can say that theta I. at
that instance should be at this point if I see from t1to t2. if this θ I does not vary too much okay.

666
So then between t1 to t2 for all these values of T, it must be remaining constant. okay so that θ I,
can just write at that point whatever ωC I will be getting into T plus θ0. suppose I can write this
way. then what will happen this particular part if I just put a tangent over here, that must be the
linear curve, so if I just say this is that ωC okay.

Because of that tangent whatever slope it will be creating. I say that is the means that is ωC. so,
so for this instant I can write that ω C into t+ θ0. wherever that tangent will cut this particular
part. I can write it this way right, so or instead of writing this as ωC I can write it as instantaneous
frequency at that instance. okay similarly at every instance, I can start putting tangent where ever
it will cut that will become my theta initial phase and then accordingly this particular thing will
be coming out right.

So what I can see over here immediately that basically the instantaneous frequency is becoming
the slope of the tangent I draw over that theta i curve, this is something very clear almost like
when you start talking about velocity and acceleration. okay, so if the velocity was not constant
then the acceleration was similarly defined. right it is the rate of change of velocity. and if the
velocity goes in any extent then you start defining it with respect to the tangent. now the tangent
from the differential calculus.We know that that is the differentiation of this particular part.

(Refer Slide Time: 12:53)

667
 dθ
ωi =
dt
dθ
ω(t) =
dt
t

∫−α
θ= ω(α)dα

so immediately what we can say this ω I which is the instantaneous frequency that must be the
differentiation of the phase, which is happening over there at time T. okay, so this relationship I
can always write.or I can write ωCT must be D θ, DT. so time varying relationship of ω. I can get
which is our differentiation of the overall phase of the co sinusoidal okay.

So this is something I can write now from here I can get the definition of θ. we know that
differentiation and integration are actually conjugate. okay so immediately I can write theta as
from minus infinity to T ω α D α so this relationship also I can easily write. right so these two
relationships fundamentally comes from our understanding of frequency and phase. okay so if I
just take overall phase and then try to define what is frequency and frequency immediately that

668
angular frequency. immediately becomes the instantaneous differentiation at that particular point
okay.

(Refer Slide Time: 14:01)

ωct + Kp m(t)

A cos[θ(t)] = A cos [ωct + Kp m(t)]

d ·
ω(t) = θ(t) = ωc + Kp m(t)
dt

So knowing this relationship let us try to do you see what will be happening if I try to modulate a
signal, that angular things we wish to modulate. okay, so let us say I have a θ T, okay so now this
θ T, I want to vary instantaneously with respect to the input message signal right so this is
something I can do so mean immediately what I can do it must, must be having a constant part
ωCT plus it should be a linear variation with respect to that. so I can put some KP mp okay.

So what I am doing the overall phase that is being created that has a constant angular this one due
to an angular frequency. constant this plus means a constant angular frequency due to that
whatever phase is being created plus there is a variation due to the message signal itself. so this I

669
can call a phase modulation because the overall phase, how we define phase it is should be ωct
plus a θ now I am varying that θ linearly with respect to this m t.

So whichever way MT is varying I have just a linear means coefficient constant which is constant
KP I multiply with that I get the phase and that phase I will be inserting in the signal. so therefore
if my θ T becomes like this. the phase modulated signal must look like a cos θ T which is nothing
but a cos ωct, right this is what happens? no problem in this now what we will try to see that what
happens in the frequency of this part.

So what do we know about frequency? so frequency must be if this is that phase θ T, my

frequency must be d DT of this. so that should be my ω T. right, so if I differentiate this that
should be ω C plus KP differentiation of that. okay, so the differentiation I represent with the dot.
so this becomes my frequency. now what was my idea of FM modulation? the idea was that the
frequency must now vary linearly with respect to a particular signal which is almost happening as
you can see over here.

The frequency instantaneous frequency which has a constant term that you can forget, and this is
where the m dot t is actually model modulating the frequency. so what I can say if I do a phase
modulation.that is nothing but also a frequency modulation but in the phase modulation, I modify
the phase with respect to M T whereas if I have a frequency modulator, similar phase modulation
I can create if first I differentiate my signal and then do a FM so correspondingly what will be
created that should be my p.m. right, so phase modulation is nothing but I first do a frequency
modulation.

But when I do frequency modulation, I also remember that if I do the frequency modulation with
respect to the differentiation of that particular signal, then automatically that depicts the phase
modulation. right, so this is what we are trying to do. so therefore I can say that suppose I have a
frequency modulator, and I want to generate a phase modulator, so what do I do, I take my MT
first I pass it through a differentiator circuit. okay, and then give it to the FM so FM will modify
the frequency accordingly and that, that is nothing but this. if the frequency gets modulated like
this automatically the phase will be this.

(Refer Slide Time: 18:29)

670
ωc(t) = ωct + Kf m(t)
t t

∫−α ∫−α
θ(t) = ωc(α)dα = [ωc α + Kf m(α)]dα
t

∫−α
= ωct + Kf ω(α)dα

So that becomes this whole module becomes the PA. right this is just by seeing the relationship
between phase and frequency we could see that. okay similarly what we can write suppose I do a
FM modulation now, so what will happen my ωct should be some constant frequency plus some
KF another constant into MT. because now I wish to do frequency modulation so frequency must
be linearly means dependent with respect to the input signal, that is what is happening if this is
the frequency.

Then what should be the phase ? phase must be integration we have already derived that
integration from minus infinity ωct or αD α,which is nothing but integration minus infinity to T.
ω C α+ KFm α D α . this happens to be just ω C into T. so this is ω C into T, and this is just K F
comes out integration minus infinity to T, M α D also right, so that is my phase so what I can do
from ωc.

671
So basically I need, suppose I have a phase modulator. now you can see it very clearly I have a
phase modulator, so what I do first I pass it through my message signal pass it through our
integrator circuit okay, so I will get this and then I actually modulate the means,I do a phase
modulation with that, I get actually the frequency modulator. so suppose I have a phase
modulator then what do I do the message signal I pass it through an integrator, then I put a phase
modulation" correspondingly whatever I get that must be a frequency modulated.

So from this relationship between phase and frequency what we can see that these two are almost
equivalent. if I have one modulator I can always construct another modulator, just by using a
linear circuit in front of that. either a differentiator or an integrator. so if I have a PM I use a
integrator to get FM. if I have a FM modulator then I just use a differentiator circuit in front of
that to get PM. so this is something we will be able to always do. and that is why probably all our
discussion,I can just do it for one the other one will follow automatically because the modulation
as well as the modulation all these things you can do very similarly.

(Refer Slide Time: 21:35)

672
 t

[ ∫−α ]
ϕGAM (t) = A cos ωct + m(α)h(t − α)dα

= A cos [ωct + Kp m(t)]

h(t) = Kpδ(t)
h(t) = Kp h(t)
t

[ ∫−α ]
ϕGAM (t) = A cos ωct + Kf m(α) dα

Okay, so with that can we just give a unified representation. so that something will just say so we
call it our generalized so this is just the more, means most generalized FM or PM modulation that
we can give or any other angular modulation that we can give. so we can see,if I just write it in
this expression.

So it is a cos ω CT plus integration from minus infinity to T, M α h t -α D α, I will just explain

what why, what I have just written. okay, so we are saying that this is probably the most
generalized representation of either PM or FM or any other variety of angle modulation. why is
that? so I will just give you two examples, let us say HT is equals to K P ∆T, then what happens
to this so if I just take this as ∆ and if I just integrate a function multiplied by a delta function.it
will just give me, return me back the function itself. right.

So immediately what I will get, if I put this, this will become a cos ωct plus I will get KP Mp.
which is nothing but F M, sorry PM. so that is the phase modulated signal, because the phase is
now that phase portion is now linearly modulated with respect to the message signal. if instead of
that,I start putting HT equal to KF UT, now let us see what will happen, so then this Phi
generalized angle modulation T, will have a form which is nothing but ωCT plus, now I
multiplied with UT or UT -α 'that means it is actually just remains as m α, because that will be
from, from T equals to α to infinity this will be one.

Okay, So wherever I take that m α, it will be one over there, so it should be just mα, rest of the
things will be vanished. so if I just do this integration, you just look like minus infinity T that KF
should be over here,Mα Dα can you again identify this form? this is actually the FM form. okay
so the frequency modulation can be realized by realizing at different HT. so it's just almost like a
means this HT you can already see it is just like a impulse, impulse response of a particular
transfer function, that I am putting in front of the modulator.

673
okay so that is what you are trying to do. so you just maybe you are having a FM modulator
where the phase modulation you are creating by that transfer function. okay, so whichever
transforms on you choose accordingly either PM or FM will be generated. and not only that we
can also see a plethora of other things which are generated which you do not know. because you
can now take H anything and this will be one version of angle modulation.

Okay, So this generalized representation just tells us that we do not have just one option or two
options, that FM or PM there are in between multiple other options we put will probably get one
kind of angle modulation. okay, whichever way that is, but it will have one angle modulation
because it is just with respect to my input signal it is just modulating the angle. okay, so it just
depends on what kind of transfer function you are putting. so we have given two examples those
two examples actually takes us to PM or FM and that is why when we are talking this as
generalized angle modulation probably that is true.

Because we can we can even get multiple other angle modulation using just this particular form.
so once this is being done, what we can now start talking about is a bandwidth of FM. okay, so
this was something which has been means which was bothering people over the year. as I have
already told in the introduction part of this particular class and that FM was initially thought that
it has, means I can, I can vary the bandwidth according to my wish so I can have even zero band
width.

So that was the initial assumption, then people started seeing probably no that is not true if I
actually have infinite band width. okay, so it went from one extreme to another extreme and why
does that has happened I probably will try to examine that portion. so what we'll try to do in the
next class is try to see this why people had so many different version of FM bandwidth. so this is
something we will try to explore in a better way, that that lead us to actual bandwidth calculation
which is known as Carson's formula.

So we will first go through all this jungle of different bandwidth proposition for FM, and then we
will probably try to hit the right thing where the actual FM bandwidth is being evaluated. it is
neither infinity not zero it is somewhere in between, and then people could realize which I have
already told that it is, initially people thought it is bandwidth efficient so it was not bandwidth
efficient then people started thinking it might have infinite bandwidth and people could realize
that it is at least as worse as DSB-S C. it is not better than that.

So if you have to consider FM, of course it is not infinite/but it will be as worse as DSB SC, you
cannot have anything lesser than a DSB SC so it cannot be any way band width efficient
compared to any other amplitude modulation schemes. so whichever we have discussed from

674
DSB SC to QAM toto VSB to SSB, it is not better than any of those things. so that something
will try to prove first and then we will see the other benefit of it, okay thank you.

675
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-48: Frequency Modulation (Contd.)

Okay, so coming back to the calculation of FM bandwidth. Let us try to do calculate that
one ,suppose because of its FM so what we have will have to do, we have already have seen
that ,we generally any co-sinusoidal whenever we represent we represent with respect to the over
all phase, so I should represent it with the integration okay.

(Refer Slide Time: 00:55)

676
 t

∫−α
a(t) = ω(α)dα

̂ (t) = Re[ϕ̂ (t)]

ϕFM FM
·
ϕ̂ = Ae c t+Kf m(t)]
FM
j[ω

= Ae [ωc t]e j[k f a(t)]

j 2 kj2 a 2(t)
[ ]
= A[cos ωct + j sin ωct] 1 + jkf a(t) + +⋯
2!
Kf2 a 2(t)cos ω0t
ϕ̂
FM (t) = A cos ωct − Kf a(t)sin (ωct) − +⋯
2!

Because whatever mt ill be giving , we have already seen that form and it should be a cos some
wct+ integration of m αt α, right into km ,so let say that integration part I represent with another
thing a(t) so integration -∞ to t m(α) dα, this is just for representation we do not want to always
write that integration form, so we have just taking that as another function of t because I am
integrating up to t right.

So then we define something we will see why we are defining this so which we are not calling as
Ø Fm but Ø Fm ^ , why because I am defining the complex signal . So I will define it like this A
ej w ct+Kfa(t) it is very easy to see, if I just take the real part of this I must be getting Fm the real
part of this will be it is cos+j sin so I will get the cos part.

So that, that is the a cos w ct+ kfa(t) where it is this one that is Fm already discuss that, so I just I
am trying to see what will happen to this, okay so I know that my Ø Fm(t) Fm modulated signal
should be the real part of Ø Fm ^ (t) right, is something I know already in the background so let
us try to see I can write this as jw ct ej kfa(t) right ,now what I do I expand this in an exponential
series okay nothing else and this I write as cos w ct+j sin w ct will see why I am writing this.

So I just representing both of them in two different forms nothing else, one is I am just taking
Fourier’s formula ,another one I am putting the exponential series ,so this should be 1st term
should be 1 next term should be this jk fa(t) right divided by factorial 1 which is 1 then + this
square it should be j2 k2 a2(t)/ factorial 2 and so on all other had terms up to infinity right.

So what you can see every odd term ,okay wherever I am doing means this is probably 0th term,
this is the 1st term, the 2nd ,3rd so every in that way if I just represent every odd term will have
means ,that will be an complex one now I should say that be imaginary term right, that j will be
multiplied here it will be just j2 so I can put that as minus okay.

677
So it be like that and I have this so therefore whenever I write this ,I can just take the real part of
it of this multiplication, so there is something I can start doing what I see ,wherever this j term is
there that should be multiplied with sin and wherever there is no basically j term that should be
multiplied with cos wct that must be this ,so therefore immediately I can write ØFm(t) from this
understanding I will just take all those multiplication where I have real term.

So the first one will be A cos wct*1, the second term should be this multiply by this okay so that
must be minus because j2 will give you minus, kf a(t) okay so that multiplied by sin wct right
and so on right there will be other terms like minus kf2 a2(t) cos wct/ factorial 2 , this multiply
by this and it will just continue okay as many term of their which are free of j all the real terms
so that must ØFm.

So what I have done is I have got an expansion of Ø Fm utilizing this okay, carefully see it has a
co-sinusoidal which is a carrier term, the next one is just like DSB-SC okay by A(t) ,that
integrated signal whatever it is that is as if you are doing with DSB SC , of-course in DSB-SC
we do it by cos or by sin okay, but the spectrum will look like the same okay so there will be a
translation around wc , which will have a band, equivalent to this band of A(t) right.

Then immediately you can see that at least Fm modulation will have a band equivalent to,
integration of the signal now what we can also say very carefully see ,suppose this m(t) that is
the band limited signal right, what do I mean by A(t) that means m (t) is lineally pass through the
integrator okay.

So if m(t) is band limited the integrated version also should be just band limited because it just it
is getting pass through linear circuit, so if I just means this signal pass through the linear
circuit ,so what will happen wherever it is 0 there it will be 0 say if it is band limited wherever it
does not have any component it will also not have any component, so if that is band limited this
should be also band limited with almost equivalent band okay.

So if that is the case then I know that this is A(t) also will be band limited with equivalent band,
because integrator is just nothing but 1/j2pf right, so that something we can just represent if I just
put that immediately I calculate band, band will remain the same ,so a(t) in the frequency domain
if I try to see that has equivalent band as m(t) or m(f).

If I do this modulation so first two term if I just inspect , I have a carrier and I have already
almost similar like dsp modulation which takes the entire band so initially whatever we have said
that fm probably might have 0 band, that is not true from this analysis we have already see, but
not only that rather stories to you just start in inspecting this one this is also shifted at wc so it
around wc but what is the band it is having the signal which is square of a, so it is square t what

678
happens in the frequency domain ,whatever is multiplied in time domain in frequency domain if
suppose a(t).

Other frequency domain or Fourier transform it is A(f) so therefore in frequency domain that
should be convoluted right, now we have already proven this that two things convoluted , the
band will be both the things are same, so if you convolute with itself the band becomes twice,
okay so this will have now 2B band and it will have all the higher order terms up to infinity.

So there will be the term were I will get A to the power nt which will have n times the band and
then n goes to infinity, so that I the effecting band of the Fm ,that goes to infinity this is when
people started getting discourage them about FM, they could see that from the analysis it is very
clear that Fm band is infinite.

So whenever I do this modulation because all this square cube terms and order terms are being
already embedded ,okay so I am getting all every frequency component getting populated by this
okay, so that was one apprehension people had for Fm, okay so this is I hope this is particularly
clear right ,so I get infinite that is alright.

(Refer Slide Time: 09:53)

679
m(t)
| a(t) |
| Kf a(t) | < ∞
(Kf a(t))n
⟹ 0 n→∞
n!

But here there is the catch, let us try to see that, if suppose this message signal the regular signal
which is bounded, let say that is the case so what is a(t), if m(t) remains bounded and it is a
regular signal when the integration of it also must be bounded ,so this is something we should be
because otherwise the signal means the integration goes to infinity.

Then doing transformer is not possible okay so that is not our regular signal or any signal that
can be handling in communication systems because that will not be possible okay, if we wish to
calculate the power and all those things it will go infinity right, so our energy will go to infinity
so that is the kind of signal we are not expecting that kind of signal okay so basically what will
happen even the differentiation due to that our integration due the reason will remain bounded
okay.

So that I can always say this is bounded, this is does not approach to infinity I can then say kf
which is constant into a(t) so that is also remain bounded okay, so this must not this must be less
than infinity I can always write that, okay now a bounded value what we are getting the nth term
if we try to inspect which as k(f) a(t) to the power n right.

So I have k(f) a(t) to the power n divided by factorial n, we know that this polynomial term does
not go as fast as this factorial term, that is true because we are polynomial term whatever we are
multiplying it is the same thing getting multiplied n times where the factorial terms goes much
faster okay.

If that is the case this is already bounded so this will goes faster than this one, okay once I
already know this ,then I can say that as n approaches to infinity this will approach to 0 okay, so
all his not lost over here what I can see, though I have infinite band at the strength of that band
whichever is created at higher order term it was going to infinity because the nth terms was
creating huge amount of band width as b right .

So as goes to infinity this band goes to infinity, but what I could also see that the nth terms as n
goes to infinity that gets vanished, so basically the importance of that will be vanishingly small
and then I can conclude that most probably all this terms are not important some of those terms
will be important with that there will be some band which might be bigger than of course b.

680
That is something which is true but it will not be infinite most probably okay, so this is just
intuition that is coming directly from the fm signal analysis, we still have not prove anything
about it expanded we just doing it just to exert something about the band okay ,so once we have
this understanding ,can we now go about analyzing the fm band okay.So that will be our next
task .

(Refer Slide Time: 13:35)

m(tK )
ωi = ωc + Kf m(tk )

So let say I modulate the signal which looks like this okay, so this is my m(t) or this is my
message signal, I modulate this signal okay, first of all probably we still have not proven that will
resolve the sampling theorem ,so we have not proven this particular thing but it just say that you
most of you, later on will also prove that most of you know that this particular signal as long as it
is band limited.

That means the bandwidth is less than B ,then I have effective Nyquist sampling which is two
times of this one okay so 2B if I just sample this signal those samples are already represented it

681
of this entire signal, I do not have to really take as long as it is band limited, I do not have to
really take all the samples infinite number of sample that every time instance.

I can just sample it at an instance of ½ B or even higher ,okay the sampling rate at least has to be
this lower than this sampling rate will not represent the signal faithfully, this is something will
prove but this is the nyquist sampling theorem okay, so if I just sample it to this I will still have, I
will just take this samples and those samples faithfully represent the whole signal.

What does that means? I can actually start constructing box of that, of course you can see
because I have taken the box little bit bigger actually it will come from the band width of it ,
actually gives all this slope so accordingly the band will defined and accordingly by sample size
or this simple interval would be defined.

So this is not properly drawn the if I just plot this into frequency domain I will see probably
much higher frequency terms of because of this variation and due to that ½ will be much smaller
but whatever it is there will be ½ B which faithfully represent it ,I can just take those samples
and I can make this sample top flatter, so I will say that particular flat sampled signal whatever, I
am this box if I just put one of the another whatever I get that actually faithfully represent this
signal.

I am not saying any wrong thing over there ,because I know those samples are true representative
of the whole signal will be proving that also sampling interferences say that so basically I can
say at this particular signal can even be represented as those boxes okay, now what happens at
this instant what is the height of this box.

That is actually at that time instant whatever sample value I get ,so that must be at ,suppose this
is t(k) value at t(k) whatever message signal I get ,whatever sample value I get and in Fm what
will be doing will be actually picking those samples and with those samples you will be
generating a frequency right.

That is exactly what Fm is so general Fm says that you take every sample and start creating the
corresponding frequencies but we are saying because the message can be only represented by
these samples faithfully so I will take that sample and for that amount of period will keep the
frequency same which is determine by this m(tk).

So at that m(tk) what will be the instant in your frequency wi, that should be because the
frequency so it should be wc+K(f)*m(tk) this is what we have said already for fm that at any
time whatever the message signal it should be proportional to instant frequency ,should be
proportional to that, that is what we are doing and that k(f) is the proportionality constant.

682
According to our fm modulation definition , so at this instant I will be just creating a frequency
which is exactly this and because I have represented this as a box signal so that means this
duration the signal will stand remains m(tk) so therefore the frequency will remain m(tk) so I can
now say that the entire signal if I do a modulation, this will just the a pulse over which duration
this frequency should prevail.

The frequency is this ,next time whenever I put I have the change frequency which will be at that
instant whatever m(tk) value will be that particular one, so accordingly I can represent the fm
modulation so I can just sub divided into those pulse and modulate that with the frequency that
comes out of it.

At that instant whatever value I get ,so this is the faithful representation of fm modulation I can
immediately say from the way we are constructing it ,that advantage you get is now the actually
have pulse which is getting modulated at a frequency which is known to me ,okay so how does
the pulse will be represented in suppose I want to see the frequency domain representation so
what will happen this particular pulse, what is that it is the unit pulse okay if I go to the
frequency domain how that will look like that will be our sinc function if 1/2B is the width of
that pulse because of that sampling theorem.

So what will happen this one by that should be the point where it will be cutting means the
frequency axes in the positive half as well as minus half one by that should be the frequency axes
where it will be cutting the frequency axes okay if I take it to w that should be 2pi into that so
1/1/2 be should be 2B*2pi that should be our that 4piB so I can represent this as sinc function at
this point is 4piB if I put it in w right.

But what is happening that particular pulse is getting amplitude modulated, it is a pulse and then
I have modulating it to the sinusoidal right of frequency this , so immediately this will be shifted
modulation means that pulse getting shifted in the frequency domain to this particular frequency,
so correspondingly if I start drawing the frequency domain chart.

(Refer Slide Time: 21:10)

683
BFM = [2Kf mp + 8πB]

I will see for this m (tk) what will be the frequency that should wc+kf*m(tk) at that centre
frequency there will be the pulse ,sorry there will be a sinc function which is having a this zero
crossing will be +4piB so wc+k fm(tk) what is the maximum that I can get over this center
frequency that ,should be the maximum voltage that the signal can get.

Let us called that as mp and the maximum of the minimum voltage that I get let us call that –mp
so it just seeing the overall signal just seeing the up to the maximum mp and minimum of –mp, if
that the case this entire frequency domain will be populated with so many of this sinc function at
every impulse I will be getting corresponding sinc function right but the sinc function the
maximum that I can get will be populated at wc+k(f)mp it will not go beyond this.

And the minimum that I will get that should be at wc-k(f)mp it cannot be anything less than
this ,and then there will be the sinc on top of this here it 4piB and there will be the sinc on top of
this, this is 4piB so overall this variation were the meaningful bandwidth will be because beyond
this sinc will get diminishing all the sinc will be diminishing right.

Beyond this point, so I can say this is probably my valid band width and what that is that is
actually this minus this so that should be 2*kf*mp + in this side 4piB from this side 4piB so it

684
should be 8piB that should be the band width Bfm band width of fm right ,so you can see that in
fm band width there is a term called this which is particularly dependent on the maximum signal
strength, as well as the kf value that you supply + this 8piB where B is the messaged signal
bandwidth okay.

(Refer Slide Time: 24:27)

kf mp − (−mp kf)
Δf =
2π
k f mp
Δf =
2π
kf mp = πΔ f
BFM = [2Kf mp + 8πB]
= 2πΔ f + 8πB
= 2π[Δ f + 4B]

685
So these two terms actually comes into the picture whenever B try to calculate the fm
value ,okay so I just try to see what this means okay so this is actually ,what is the maximum
deviation in frequency I can get in fm what do I mean by that so by maximum signal is mp
kf*mp is the maximum angular frequency that I can create.

At the minimum angular frequency that I can create around wc that should be –mp*kf right
because the minimum signal will be –mp so this is the deviation between the maximum
frequency and the minimum frequency that I get okay in w, if I wished to represented it in
frequency domain delf that must be divided by 2pi.

So this is the frequency deviation that happens due to my signal variation ,as you can see that is
directly dependent on this kf factor, so more kf I put probably there will be more frequency
deviation less or kf if I put there will be less or frequency deviation right and immediately what
do I get that should be 2kf mp/2pi right.That is my del f .

And fm band width now I can represent with respect to this del f so kf*mp that is actually pi*del
f, I can write it this way okay so immediately I can write this kf*mp that is actually p*del f ,so
therefore this Bfm which was earlier derived as 2kf*mp+8pB that I can write as 2pdel f+8pB or
2p I can take common del f+4B okay.

So this Bfm was in w domain ,if I wish to get that fm frequency band in f domain so I have to
divide by 2p so this Bfm in frequency domain let us call as F, that should be del f+4B this is the
fundamental relation that we are getting that fm frequency is due to this frequency deviation, that
is happening okay and the 4B which is particularly band width means relative to band width of
the message signal initially you know what happened people were mistaken ,that in fm the
frequency terms are getting populated just because of I am varying the frequency okay.

So that because of that frequency deviation ,so they were thinking that just frequency terms will
capture my band width because every frequency will have deviation, sorry every amplitude will
have the deviation so ,maximum amplitude I can calculate from their I can get overall deviation
which is the del f so if I make del f tends to 0.

Then my fm modulated signals also will have 0 frequency band width ,or times to 0 frequency
this was initial understanding of people they were thinking just this del f I can capture ,I can
capture the amplitude that will be the fm band so initially before understanding all this
maths ,people were thinking that it is fm means for different frequency I actually give a different
sinusoidal , so if I have if I can capture this entire del f variation that must be all those sinusoidal
will be populated over there.

686
So that must be my over all bandwidth , so if I can make my del f literally very small which is
called narrow band ,will demonstrate that part so which is called narrow band fm, so if I can
make this del f infinitesimally small by just making kf very small ,if I just do kf very small
whatever no matter whatever the message signal variation it will be bounded.If Kf is very small
then del f will be very small because del f Is kf*mp/pi.

So that is the case I can always take this very small and the corresponding band width of fm will
be infinitesimally small, so that was the initial understanding if people were thinking and
propagating this news that fm will be the savior , it gets the modulation scheme over the band
width becomes vanishing this small, this is not true that we have understood already that
whatever I do I can make del f to 0 okay but this 4B will be there .

And then people started means trying to see what will happen if we have a narrow band version
of this so in the next class what will try to do is will try to see if I do a narrow band version of
this, what do I get out of this and then will this formula faithful because this formula was derived
in adhoc manner ,you have seen that already had some intuition that we are not exactly
calculating because at every value there will be a sinc function ,then how the sinc function dies
and what will be the overall bandwidth, we are just saying that wherever the last sinc function
goes to 0 that is my bandwidth that’s very adhoc way of saying bandwidth or defining
bandwidth, so we would like to see will more inside into that and try to give some intuitive
understanding about this formula the corrected version of this formula and you will see later on.

We will improve that, that was actually thanks to Carsons formula this is actually del f +2B it is
not just 4B will see that correction that comes from tone modulation, so will try to capture that
part and try to solve the fm band width means whatever jinks they had at that time. Okay, thank
you.

687
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 49: Frequency Modulation (Contd.)

Okay so in the last class probably we have already started discussing about FM bandwidth. so let
us just explore a little bit more on FM bandwidth. but before that for our own advantage we
would like to give some intuitive understanding of narrowband FM.

(Refer Slide Time: 00:45)

688
 Kf2 a 2(t)cos ω0t
ϕFM (t) = A cos ωct − Kf a(t)sin (ωct) − +⋯
2!
Kf < < 1
ϕFM (t) = A cos ωct − Kf a(t)sin (ωct)
BFM = Δ f + 4B

You remember, We have derived this formula that ϕfm (t)that was actually A [cos ωct when we
took the real part of this, -kfa (t)sinwct-kf^2a^2 (t)/2 factorial, coswt and upto .∞]. now what we
have said that initially people started saying that we want to have a very small deviation. and that
will create a narrowband FM at that time people were thinking that narrow band FM means it is
almost zero bandwidth. okay so that Δ F will be zero. so with that understanding we say that K F
<<1, very small value of K. immediately what happens, all these higher terms actually goes away.
squared to cubed- up to ∞ goes away. and the ϕfm becomes this. which is almost like AM, okay'
though we are calling it FM.it almost means the functionalities almost looks like FM okay.

So we have a carrier term over here which is this, and then we have just 80 x sine ωct. it was for
a.m. it was cosωct and there was a plus sign. just that difference otherwise, it looks like a AM.
okay so earlier that was our amplitude modulation, right. how on earth this is becoming a
frequency modulation?this is something comes to our mind every time we ask this, but remember
here also the amplitude variation will be very restricted because the K value is very small- okay,
so a multiplied by this KF into A t.

So whatever that a t variation is, it will be very small, it will not be very significant so the
amplitude variation as we can see it will be restricted. okay and because we have got it from the
FM formula, so this will also have corresponding frequency variation, will later on, when we will
be drawing FM phaser diagram, probably you will understand this. right now we are not means
we are not trying to prove that part. but this faithfully represents that above 1 and that was
according to us FM signal. so we know that at least we could prove that this does not have too
much of amplitude variation. that is for sure okay.

So that is something, but there will be some amplitude variation once we start truncating it, that
all those higher signal actually they were balancing those amplitude variation, so once we target
truncate that however small there will be some amplitude variation we will also see when
narrowband FM will be generating how do we really control that part. so that something also we
will see. but this is what happens. At this instance, if I now talk about because we were talking
about this narrowband FM we have just demonstrated because we wanted to say talk about
bandwidth so we were talking about bandwidth.

689
So at this intense. if this is the FM earlier understanding was narrowband FM must have
frequency deviation almost zero, so the overall bandwidth should be zero. but this is coming out
to be narrowband FM so what's the bandwidth of this, this has this modulation term so it must be
having a bandwidth of two B. if B is the bandwidth of that. so immediately I can see if I just
make narrowband FM, that means ΔF tending towards zero or very low value of K F, then this
particular function tells me the bandwidth of that FM must be 2b. but what was our derived
formula.

So β FM in frequency in F domain, not ω domain, angular frequency domain. so that was ΔF +

4b right. now make ΔF tends to 0 so this ad-hoc formula tells me this is 4B but from here direct
from FM. I can get that this must be 2B. right, so there is some fallacy, in this still, but what I can
understand because this was a adhoc formula, so I should not believe in this formula directly I
know that this formula should be almost correct, but it should not completely believe on this
formula will till I derive some very accurate kind of formula. but with this ad hoc formula I can
say that probably the overall bandwidth because what we were doing, wherever the means that
sinc function goes to zero we are taking up to that.

So maybe that is over estimation, because I am almost going towards zero and then after that all
the other ripples will be smaller and the ripples that are coming from internal sinc, this is just the
boundary sink internal sinc there will be even smaller. okay, so probably we are overestimating it
so it should be intuitively we can say because this must be 2B, whenever I put ΔF tends to zero,
so in Duty B I can say it must be this is the upper bound of that, so it must be something like this
we will come back to this. okay so this must be the FM bandwidth. okay, let us try to see how we
can deal with that okay.

(Refer Slide Time: 06:50)

690
m(t) = α cos(ωmt)
t

∫−∞
a(t) = α cos(ωm x) d x
α
= sin(ωmt) a(−∞) = 0
ωm
α
̂ = Ae j[ωc t+Kf ωm sin(ωm t)]
ϕFM Δω = Kf mp = αKf
Δω Δω Δω
= Ae j[ωc t+ ωm sin(ωm t)] β= =
Bw ωm
= Ae j[ωc t+β sin(ωm t)]
= Ae jωc t e jβ sin(ωm t)

So to deal with that, let us try to define one thing which is called frequency deviation rate ratio. it
directly comes from that particular formula where we are saying that probably this FM bandwidth
should be this ΔF+B. wait a second, Where is my derivation, correct Δ F + 4b, yes, so we are
saying ΔF+4b. right, okay. so this particular deviation I am taking this sometimes you will see
this derivation means, this ΔF I have taken it from the entire deviation. so it sometimes taken

691
from zero how much it varies in the positive and negative halt. so that way they write also to Δ F
so according to the formula gets modified okay.

So if I write that way then it becomes 2Δ F + 4 B. okay if the delay definition is not the entire
deviation it's just from the center, that ωc how much in the positive half, and how much in the
negative half it goes, so if I just take it that way, so it becomes this or we can write - Δ f means 2
into Δ F + 2 B. and we have said that it maybe not UB, it may be 2B. so, with that approximation
b FM must be 2 (Δ F + B) . okay, so one of this formula will be correct.we do not know right
now. we have derived this one but we are saying that probably from narrowband FM this should
be the correct one okay.

Now whatever it is the derivation ratio is defined as this. I take this B out, then I will have this
thing ΔF divided by B plus one. this particular thing is called termed as deviation ratio. okay this
is actually tells us how much deviation I get in frequency divided by what is the bandwidth of the
original message signal. so that ratio is defined as a term called β. okay in our next derivation
probably this beta will be useful. so this is actually the deviation ratio that we are talking about so
now let us try to do this. okay suppose I have a message signal which is actually a tone
modulated signal or which is a tone or which is a single frequency.

So cos ωt, so I am not now, this was by the derivation of Carson’s since how he could derive FM
bandwidth-so we are just going through his derivation so here what we are saying, and earlier we
are taking any FM. so any message signal you could take and then we are trying to derive the
whole formula. so now we restrict yourself we are saying probably we are just test it for a tone
modulated signal and we are do the proper derivation. The proper bandwidth derivation. so we
are saying that a message signal is just a tone or a co sinusoidal at frequency ωm. so this is ωm.
and of-course omega M must be much, much smaller than ωc.

So this condition is valid. okay, so if this is the message signal, first I have to for constructing
theFM I have to first construct a t. okay, so what will be a t that will be the integration from
minus ∞ to that particular T value.now if I wish to do that, it will be minus in ∞ to t. α cos ω MT,
sorry, α D α or I should not say α that should be some other things, so let us say it is X DX. okay
because α is already there. so I have taken a dummy variable I just have to take a dummy variable
which is X. so omega M X DX okay omega m inside this. now this integration if I wish to
evaluate, I already know how to evaluate that but at minus ∞ it will create problem for me.

So if I just say that at minus infinity I already know that this particular thing will be, means this a
T will be already zero-so this is something if I know. means we are just assuming that. at minus ∞
we know this-okay so if I just assume that at a minus ∞ I know that is zero, so then immediately I
can write this as integration of that and putting as T.so this should be α divided by ω M sine

692
omega M. now in place of X I will put T because at minus ∞ I know it is zero. so this is my aT.
right, so now I wish to evaluate the bandwidth of it so what I will do I will first go to the actual
FM signal. so let us do that the way I have done that ψfm (t). let us try to evaluate this part.

So this is actually Ae ^ (iωct+ kfα/ωm sin ωmt) so KF into alpha divided by ω M sine ωM T,
right. I can write it this way. now I can also write we have already derived, this Δω is actually KF
into MP this is something we have already derived, so this I can write as α into KF because MP
comes α this is m t-okay, what is the maximum value of this that is actually α and minimum value
minus MP is just minus α. so MP becomes alpha. so I can write Δ ω in terms of α and KF. Right,
so this is something I can write.

So KF into α immediately I can write as Δω. so I can write this as e ^-Jω CT+ KF into α I can
write as Δω/ωm, sine ωmt. okay, now what is β that deviation ratio. that was actually Δ F. or if I
represent it in terms of ω that should be Δω divided by the bandwidth in ω domain-okay, now for
this particular message signal what is the band or bandwidth it is just a tone at ω M. so the
bandwidth must be that only. right, so B must be ω M.so I can write this P as ω m and del ω is
this one.

So Δω by ωM must be my β. so β Δω and B must be ω M so this becomes my beta. so I can write

this as a Ae ^Jω C T + β sine ω MT. fine, if I separate it, okay, so that is my phi hat FM. now try
to see this signal. because I have sinusoidal it is a periodic signal. right with frequency ω m or
this angular frequency ωn. so because it is a periodic signal what can I do, my Fourier series cell
it tells me that this can be re presented as a Fourier series expansion. okay so that understanding
tells me that.

(Refer Slide Time: 15:44)

693
 1 π j( β sin x−nx)
2π ∫−π
e dx

= Jn(β )
∞
ϕ̂ Jn(β )e j(ωc t + nωm t)
FM (t) ∑
=A
n=−∞
∞
Jn(β )cos (ωct + nωmt)
∑
ϕFM (t) = A
n=−∞
n >β +1

I can write this e^Jβsine ωmt as Fourier series expansion with coefficient DN.and the frequency I
know already ωm so this must be e ^ Jmt and this n varies from -∞ to +∞. okay, wait a second I
have not written it correctly. yeah that is correct. so this must be the case. so basically at every
ωm I will be getting this term, probably that M I have not written over here. okay so it must be
that every frequency term I will be getting okay with respect to this, and yeah it is not m, it
should be n.

694
So for every value of n I will be getting a corresponding N and I will be getting that frequency
term. right now this DN how do I calculate? so DN becomes the Fourier integral. so that should
be 1/2π so that is or 1/T so the period is basically the overall period is ωm that is in angular
frequency so it should be ωm/2π or the time should be 2π/ωm. so it should be -π/ωm to + π/ ω M
and 1/T means it should be ωm/2π. and the integration goes this is the signal itself it will power
e^Jβsin (ωmt)-right, multiplied by e^Jnωmt. right that is the Fourier series analysis DT right.

Now just put means pie by this T by ω M as X, so immediately what you will get is this becomes
1/2π goes from -π e^ j β sinx e^-jnx dx or, I can write π2πe^-Jβsine x dx. okay'this particular
integral is not something that can be calculated easily okay that has to be numerically evaluated
but this has a name this was by Bessel, and this is called Bessel function of means first means,
this is in n th order actually. so because this n there is there. so it is called an n th order Bessel
function. so I can write this as this integration 1 /2 integration minus π to π eJβsine X minus nxdx
this can be written as J n β.

So this particular integral I can write it as J n beta-right so immediately what happens by

representation that phi f m hat t what happens to that?this DN can be now represented as this. so
immediately I can represent this whole thing as summation n goes from minus infinity to plus
infinity, j n β , which is that coefficient, and I will have e^jωct was already there multiplied and
here it should be plus n ωmt which is coming due to the Fourier representation, the Fourier series
representation. so this is something what we see. immediately I should take the real part of it so I
will be getting this just the cost component of all of these come things. so ϕ FM T I can write as a
summation n equals to minus infinity plus infinity J n beta cosine of ωct + n ω M T.

So you can see that it is actually creating C at ω C I have something and then plus n minus n all
infinite terms it is creating. okay so this is something which is happening. so again we can say
probably the overall frequency is infinite. okay so that is something we can immediately from
here we can say because it goes from minus infinity to + infinity.all of them will be populated but
if you carefully analyze this J n β it is revealed, that as n goes bigger than β +1 the significance of
this or the value of J n β becomes insignificant. so this is something from the Bessels function
plot for different values of β and n.you can you can actually see that okay.

So for every n there is a β+1, means beyond which that n value is not really significant. okay so
what we have to do. we have to. so for every value of n the overall significant term we can
actually evaluate. okay so what we can say the significant sideband should be up to β+1 because
n this value of n for a particular β value that β value is defined already right. so β value is coming
from the FM modulation Δ f I have given and the land with of this one.

695
So once I provide a β value. I know n becoming beyond β+1 that is becoming insignificant. so
basically what happens to my bandwidth the overall bandwidth becomes from ω C, I take β+ 1on
this side and β+1 on this side.that must be the overall band-okay so I can immediately write that
my FM bandwidth for this tone modulation must be 2 β+1. okay so this is something I can
immediately write. so this happens to be my overall FM bandwidth but this is just a number.

(Refer Slide Time: 23:17)

BFM = 2(β + 1)fm

( fm )
Δf
=2 + 1 fm

= 2(Δ f + fm )

I still have not really accounted for band width. each one of those numbers I take so this many
numbers of frequency I will be taking, what's the separation between them, that is actually either
ωm in ω domain, or FM infrequency domain or F domain. so if I really wish to call this a
bandwidth I need to multiply this with FM. because that many terms I am taking.so how many
overall this thing I will be taking. that should be this multiplied by FM. okay this FM okay or I
can write that as ω M divided by 2 π. right, so if you start writing that way you will see that if I

696
just write it, so suppose let's keep it as FM so I multiplied this so this becomes 2, now this beta
earlier we have represented with respect to the ω okay.

So β is just a ratio, either I do it with respect to ω or B it will be just 2 π gets cancelled. so I can
even write β with respect to ΔF and corresponding FM-so I can write that as they left by FM + 1
into F M. right I can write this way immediately what do I get ΔF + fm. From this is the famous
Carson's theorem. so immediately you can see for tone modulation, the overall bandwidth of FM
is nothing but 2 into that Δ F which we have defined as frequency deviation due to the FM
modulation. itself plus the bandwidth of the message signal.here it was tone modulated so FM
becomes the bandwidth of it. okay what he could do is to prove this only that for a tone
modulated signal this will be happening.

And we have seen that for a generalized modulated system probably this will be ΔF Plus here it
should be means overall 4B. So it must be taken as 2B inside.okay but what we have seen also
that narrowband FM tells me that this should be plus B. so that is why this formula was more
accepted there was no formal proof of this because it is very difficult for any other things. so
there was no formal proof of FM bandwidth overall but people could understand that this must be
this. okay so that is why finally they could take the overall FM bandwidth was 2 into Δf +B and
there was the other formula where it was 2.Δ F+2B okay.

So one of them is acceptable this was more acceptable because, the it came from the tone
modulation that's first thing where we did proper bandwidth calculation. and second thing is this
actually for the narrowband FM this defines the correct bandwidth. okay where this formula does
not define it correctly it gives me 4b which is not true for narrow band FM when Δ F tends to 0.
okay so that is why historically this bandwidth definition was more widely accepted.

And then we could see all kinds of relationship with respect to this. if my Δ F becomes much,
much bigger than B the thing that people were actually saying that the frequency deviation should
be the FM bandwidth that actually happens where Δ F is already that frequency it is not the
narrowband part-it is the wider band part where you have a huge deviation that means the K
value is already big huge deviation, on top of that the bandwidth that is coming from this
particular thing the actual message signal bandwidth that will be insignificant. so there I can say
this Bfm is nothing but 2 into the deviation.

Okay as long as the deviation is taken from omega c to +1/2 or -1/2, whichever not the whole
deviation then it will be just del f. okay so that is something we have already told. so all the
understandings were correct but in some perspective. okay people thought that it is in finite
bandwidth, yes it is in finite bandwidth. but some of the part of that bandwidth becomes

697
insignificant which we have seen through the Bessel function proved also for tone modulation,
that Bessel function itself whenever you take your n greater than that β+1. that becomes
insignificant. okay that is numerically seen okay.

So that has been observed or you could also see that from the perspective of the our derivation,
that we have taken sampled values and then try to derive it. then also we could see that, means,
no not that part, actually expanding that FM and then getting all those a square, a cube, a to the
power n term and then n tends to ∞, that becomes convolution of all those a t.that also takes it to
infinity. so infinite bandwidth conjecture was correct there's no doubt about it. but the thing is
that some of the things are insignificant that is also something we have proven that this a to the
power n KF divided by factorial n actually goes to 0.

And then we could also see that it is not Δ F, because it is Δ F if the bandwidth is already that Δ F
variation is already sufficiently large. initially people thought that it should be del f, which is true,
if it is already wider band FM. that means it already covers a huge amount of bandwidth, then we
can say it is just del f. okay but initially what people thought with that logic that it is Δ F they
could actually make the Δ F very small and they were thinking that it should be just Δ F. but that
was wrong because Δ F is always supplemented with a β. and if Δ F goes to 0. that 2B at least will
be there.

So whatever happens that 2B factor will be there. so it will never be that is what we wanted to
prove that it will never be anything better than double sideband.it will at least be worser than that,
whatever it is ! so bandwidth wise what we can prove now FM is probably not as inefficient as
people who are thinking that it covers the entire band up to infinity. so it is not as inefficient as
that, but it is much inefficient in terms of spectral usage, compared to any other FM counterpart,
sorry a.m. counterpart so this is something we have clear understanding now.

So that is where after this bandwidth calculation and all those things. the reputation of FM got
dented initially people started jumping into FM because they thought this is bandwidth efficient
but that got dented because of this derivation-so now we have seen that probably FM is not as
good what we will try to do in the next few classes will try to prove why FM is still good.in terms
of noise cancellation/interference cancellation and channel non-linearity.so all the impairments
that comes from channel-okay, so thank you.

698
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-50: Frequency Modulation (Cont.d)

Okay, so in the last few classes we have I think already started discussing about the FM
bandwidth.and we have also discussed a FM modulation index, so that is something we have
already discussed which is related to the frequency spread, and in the bandwidth also we have
defined what is the relationship between the overall FM bandwidth and the frequency spread due
to FM modulation, as well as overall the baseband bandwidth of the signal that is being
modulated okay.

So those things and Carson's formula for tone modulation what happens to FM bandwidth, we
have we have seen it from different angle and has means started gathering information about
what should be the FM band, and there was some fallacy around the calculation of actual FM
bandwidth that also we have discussed in detail. now what we will do is we will try to see how
FM is being generated and means what are the different options of FM generation and then we
will try to see how you can demodulate that.

So the means next 2 or 3classes probably will be discussing about FM modulation and
demodulation. because we have seen that is the most important part, so that is where the signals
and systems actually interact with each other and we need to define a system which does the
proposed modulation or demodulation task, so we will start with FM modulation.

(Refer Slide Time: 01:53)

699
ϕFM (t) = ≃ A [cos (ω0t) − Kf a(t)sin ωct]

Okay so that is the first thing, but before we start we should also characterize two things.
probably we have already started characterizing that, we have told that there are something called
narrowband FM, which is called NBFM and the counterpart of that is wideband FM. okay, so
what is the narrowband FM we have told that if the frequency deviation is very small that means,
especially that KF value is very small much, much smaller than one. okay, so whenever this
happens for our FM modulation, so corresponding FM is termed as narrowband FM. and we have
also seen that Taylor series with that Taylor series expansion of that FM modulated signal.

We have seen that narrowband FM when this particular KF into a T is much smaller, we can
actually represent it as this. This is something we have already explored, approximately can be
represented as Cos ω CT - KF into a t Sin ωCT, where ω C is the carrier frequency of that means
with which we are doing that that modulation and what is a t, that is actually the means, we for
doing FM we have to first differentiate it that or integrate it for FM and PM, so that is that part
right. so this is what we have we have already discussed about that.

700
Okay, So this KF into a(t )which is actually for FM it should be integration, right, integration mα
dα - ∞ to T okay, so that is something which is already there taken care of. so m also this is
actually the message signal which we are trying to, so that is actually we have told that there are
if you do Taylor series expansion there are other terms which are being neglected. because that
will be, because of Taylor series expansion there will be KF2 a-square t and all other terms
because this is already very smaller than much smaller than, so we are telling that KS a (t) it is
much smaller than one, so we can actually neglect that.

Okay Sin ω CT of course is bounded by 1. so modulus of that, so we are neglecting those terms
as long as this is small so we get this particular things and then we have told how then the
narrowband FM that modulation looks very simple, it is almost like our this DSB SC or
amplitude modulation kind of thing, where you multiply this with this, so if you just now put that
so what we have to do we have this mt.

(Refer Slide Time: 05:08)

Whatever that signal is we put through a integrator, we get a (t) and then will you DSB SC
modulation where we supply actually a Sin ω CT, so basically we will have a oscillator which
will, which will be generating a Cos ω CT. okay, and will give a π by two phase shift, to get

701
Sin.and then this has to be added Plus this is minus, what you get is, narrowband FM. okay, so it
is easier enough because what we have to do we have to just generate this a Cos ω CT which is
coming from this path.

So we are getting a Cos ω CT, now this is getting π by two phase shifted, so we get a Sine ω CT
multiplied by a( t) and then we are taking minus of that, so this plus and this minus so we get this
particular thing. okay so of course you say that somewhere KF has to be there, so we will put KF
either over here, or we can include it over here. so wherever you wish to take that KF it becomes
like that. okay, so this is how FM narrowband FM as long as whatever that KF you are choosing.

That is small enough and your means, or I should say K F into a Tl that is small enough-okay if
that is happening you will only have too much of deviation, and the corresponding one it do not
really has to go and modulate the frequency with respect to the input signal-because, we know
that as long as that is small we have that expression.okay, so immediately you will get FM
modulation for demodulation also it is very simple. you just do the reverse process because it is
almost like am modulation or FM modulation okay.

So right now will not talk about FM demodulation, so we are just talking about FM modulation.
so this is the way we will be generating as long as the frequency deviation is small this is the way
we will be generating narrowband FM. okay, if you have to just do p.m. then you do not need this
integration also, we directly supply it. right, m(t) you directly supply, so m (t) you will be
supplying and that directly goes into this DSB sc. okay, let us other things are same.

so this is that DSB yes other things are just duplication of this particular thing. okay, So
narrowband FM and that will be corresponding narrowband pm of course you will have to again
multiplied with KP, okay KF for FM,KP for pm. and KP into Mt must be much smaller, otherwise
we will not have that approximation. okay, so this is the generation process now let us try to see
what exactly is happening ?

(Refer Slide Time: 08:29)

702
ϕFM (t) ≃ A [cos (ωct) − Kf a(t)sin ωct]
= A E(t)cos(ωct + θ(t))
E(t) = 1 + Kf2 a 2(t)

θ(t) = tan−1[kf a(t)]

I have this particular term. Let us say, so I FMT which is a Cos ω CT, sorry -kf at Sine ωCT.right,
so I have this. what is happening a this can be written as a sinusoidal. I can write it as t Cos ω CT
+ some θ T. what is Et? that is actually or I can write it as a Et, so then Et should be 1 +
KF^2,A^2t right, because I am just trying to represent it as a co-sinusoidal. okay, so accordingly
we can put that and this is what we get because I can put that as Sine, this as one as Cos θ T and
then you manipulate it. immediately we can also see that θ T will be just Tan inverse KF into at.

So this model, means this particular representation we already know, ok, so from here to here if
we wish to represent immediately a T and θT can be represented as this In FM what we expected
was something like A Cos some phase variation. right, so the amplitude remains the same
whereas in the narrowband FM of-course there is a variation, with respect to T of the amplitude,

703
this is what we are getting and the θT what we expected, suppose if it is FM then generally θ T
should be just this portion because θ T is ω C T plus this theta T.

If it is original FM this should be KF into integration mα dα that is actually a(t) integration of that
from - ∞ to T. so KF into a T you should be expecting inside the phase. right, because we want a
frequency which is proportional to the message signal, whereas phase will be integration of that
right, so this I expected but what I am getting is instead of getting it as Kf at I am getting tan
inverse of that. as long as you can see already as long as Kf at small enough than universe of θ
will be equal to θ.

So I have equivalent representation and as long as Kf at is much smaller than one. I can actually
make it as one. okay, but what we can also see from here means however small this is if this is
not sufficiently small, then there is a time variation of the amplitude. however small it is there
will be okay so first task when I am generating FM because I know whatever the frequency
variation is keeping that intact, can I make this amplitude constant. so that the first thing, because
that that should be the way FM is modulated okay, so FM modulation whenever I do that
amplitude must not vary with time.

Here we have a variation. So we should be somehow employ something some circuitry which
actually keeps that amplitude fixed, while the it will not touch the phase, phase will remain the
same. what is the process of doing that, so whatever amplitude variation we have, so suppose FM
modulated signal it might look like this, something like that. okay suppose it looks like this. okay,
so there is a slight amplitude variation which should be generally constant,that should have
happened well that is not happening, let us say. if this is the case what I can do I can pass it
through a special circuit which is called band pass limiter. what is this? This is first a hard limiter
followed by a band pass filter. okay, so let us try to understand what is hard limiter, so whenever
that kind of signal is there so suppose I have this.

(Refer Slide Time: 13:01)

704
vi(t) = A(t)cos(θt)

{−1if cos θ ≤ 0
1 if cos θ > 0
vo(t) =

4 1 1
vo(t) = [cos θ − cos 3θ + cos 5θ − ⋯]
π 3 5
t

[ ∫−α ]
v0 ωct + kf m(α)dα

[ ∫ ]
t

π[ [ ] 3
4 1
∫−α
= cos ωct + kf m(α)dα − cos 3ωct + 3kf ⋯

I have a V i t some at, Cos, let us say some θ T, okay I am just trying to write/that there is a
amplitude variation, there is a θ variation. okay, so what it will do it will from this input it will
generate an output VOt, which is something like this, it will be plus one or minus one if this Cos
θ, whatever that is that might be a variation of time and all those things,I do not care, as long as
this Cos θ is greater than 0, and so this is the hard limiter that means I had this signal, so

705
whenever it is above 0 I must have plus one over there, and whenever it is below 0 I must have –
1 over there.

So this should correspond. okay so it is just almost keeping the frequency variation intact, but
amplitude it is because it is just making it plus one or minus one so there is nothing in between so
it is just making the amplitude constant' okay, so that is called the hard limiter, so this is, this is
the output of that. so whenever Cos θ will be greater than 0,I have +1 whenever Cos θ is less than
0, I will have -1. so this is what will be the output okay, so after passing it through a hard limiter
with respect to now, of course there is a θ variation.

But we can ignore that, if as long as we plot, we are plotting it with respect to θ we can just try to
see, that what will be this output. okay, so this V output, if we say that is a function of θ how does
that look like ? so it is like this, it depends on that Cos θ if Cos θ greater than one, sorry greater
than 0, it is plus one.and less than 0 it is – 1, so with respect to θ if I plot. so it will be just with
respect to θ it is a square pulse. with alternate plus and minus. right, with respect to θ.

if θ has some other variation, I do not know because I am just plotting it with respect to θ. how θ
either varies with time I am not plotting it with time. I am now just plotting it with θ-so whatever
the value of θ, it will just look like a square pulse having values plus and minus. so that we have
already represented if you just think about that square pulse we have already represented in
Fourier series, of-course we are saying that it is with respect to T, now we are doing it with
respect to θ. so if I just do that if you remember that we have already derived that. so that should
be Cos θ – 1 third all our terms will be there, or all harmonics will be there,Cos 3 θ + 1/5 Cos 5 θ
- dot up to ∞.

So this representation we have already done that in our Fourier series analysis. okay, so as long as
I have θ 'V output θ must be represented as this. because that with respect to theta it looks like
this something like this. okay this is fine now I can write, what is this θ? this θT is actually the
overall FM modulated thing plus the career part.

okay so let us approximate that, this is actually ωCT + let us say KF integration minus infinity to
T m α dα. I can write it this way. okay of-course I do not have exactly this, I have some Tan
inverse of this one. okay Tan inverse of KF into at-I I do have that, but whether I have that,I do
not have that, let us say that probably this is the case, I can instead of this I can always write it is
as Tan inverse of KF into this. which is at actually - ∞ to T.M αD α. I even can replace this
particular part as this, rest of the part will be remaining same ωCT. okay, so this whole thing I can
now write as 4 by π.

Now cos of this, so which is ω CT plus either I write it KF into this or Tan inverse, whichever, let
us take this one. it is just easier to write, as long as they are similar, and then – 1 1/3Cos I will get

706
three ω CT plus 3 KF same thing, and so on. now all I have to do I have told that my hard limiter,
this is the hard limiter output I have to pass it through a band pass filter which matches with the
FM frequency and the center frequency is ω C, then what will happen all these terms will be
neglected.

(Refer Slide Time: 18:23)

[ [ ∫−∞ ]]
4A −1
= cos ωct + tan kf m(α)dα
π
t

[ ∫−∞ ]
2
x(t) = cos ωct + kf m(α)d x

[ ∫−∞ ]
2
y(t) = a2 cos ωct + kf m(α)d x

[ ]
1 1
∫−κ
= a2 + a2 cos 2ωct + 2Kf m(α)dα
2 2

707
so whatever I will be getting. That will be nothing but4 by π Cos ω CT plus, in actual I will get
tan inverse , KF integration - ∞ to TM α dα, this is what I will be getting. okay, so that will be
after passing it through a hard limiter and band pass filter, so basically what I will do. I will
generate a narrow band FM, then I wish to make the amplitude before launching it I wish to make
the amplitude constant, so that is what I have done. so here you can see that amplitude variation
is gone. okay.so whatever the amplitude variation it was having that is all gone I just have 4 π and
of course probably that a factor will be still there.

That a will be still there, because that is that was already there in that. okay, so a factor will be
there and then I have cos this one. as long as Tan inverse data is equal to theta, I have FM
modulated signal. okay, which will be happening if this is small enough. so whenever we are
generating narrowband FM-generally it is augmented by this. that band pass limiter. okay the
reason is I can very easily do that,I will always have narrowband FM whenever I generate.I know
there will be a frequency deviation, which is a little bit non-ideal of course, but as long as my KF
is adjusted.

That is correct but there will be also a amplitude deviation which should not be there and I have a
very simple circuitry the way I have demonstrated just now just now. you can just employ that
and you can make the amplitude fixed at-least and there might be any variation of this frequency
that's all right. you will also see if the amplitude variation is there what is the consequence of
that. will we will demonstrate that also why we are doing that that will be more clearer, but this is
something you can do, and that is what the modulator chain will do you will have first
narrowband FM generator followed by a band pass limiter and then you launch that signal. okay
so this is the way narrowband FM can be generated.

Now let us try to see historically what Armstrong did. he was actually taking that narrowband FM
generated signal and then trying to make it wideband FM. okay so let us try to see or appreciate
how he has done that. okay, so that will be our first exercise. okay, so what he has done so he has
employed a method called indirect method, this is of course we are targeting the FM modulation.
and doing it indirectly. that means we will first generate narrowband FM from there we will
convert it to wideband FM not directly generating it.

okay we will also see what is, what we mean by direct generation, this is the indirect method. in
the indirect method what he started doing is, started exploring some of the nonlinear devices.
okay, let us see what happens, so suppose I have a quadratic response nonlinear device, so that
means if I have input x. so the output will be this, with this diodes transistor you can you can
realize this with a particular biasing You can actually get a region where it is nonlinear.

708
okay, so you can do that so therefore, suppose I give you a narrowband FM signal, so let us say
FM modulated signal, the corresponding YT should be like this a 2. my xt is Cos ωCT plus
ideally it should be Tan inverse that, I am just writing as long as it is narrowband we can even
write KF. okay, so that should be my xt, now I put it over here. okay, so what will happen I will
have square so it should be Cos ^2 ωCT + KF. so immediately what I can do, because it is Cos ^
2. I can have two terms.

So one will be half a2 plus there will be another term which is half a 2 Cos 2 of this. so I will get
2 ωCT + 2 KF. Now he will be able to appreciate why he was doing that. so what is happening
there is a two factor which has come over here. that something probably was not desired, but we
will discuss about that but what was desired was there is a two factor which has come over here.
okay why this is desired because, now we are actually going towards from narrowband FM to
wideband FM, that means the frequency deviation has to be increased.

So what has happened, this will effectively increase the KF and instead of Kf it will become 2
KF, so that means the deviation has become double. whatever deviation I was having earlier in
the narrowband FM that is becoming double. okay, so it will keep on happening and the good part
is, if it is Tan inverse the two will be outside. remember that, so this two remains outside, and
then you will have this Tan inverse this KF into this part. okay, so inside tan inverse it is still that
narrowband FM, so you can still approximate that to be KF into minus this, sorry KF into
integration - ∞ to t the m alpha D α right, so this particular part remains intact you are just
multiplying actually were supplying terms multiplication multiplicative terms to KF.

So you are slowly increasing that KF portion without hampering what was happening or what has
been appreciated in narrowband FM. okay, so this was his method what he did instead of doing
this, he went directly to in an nth order non-linearity.

(Refer Slide Time: 25:13)

709
y(t) = a0 + a1x(t) + a2 x 2(t) + ⋯ + an x n(t)
= c0 + c1 cos [2ωct + 2 tan−1(Kf a(t))] + ⋯ + cn cos [nωct + n tan−1(Kf a(t))]

So basically he told can I, if I can get a circuitry like this suppose it is a 0 a1 xt so it has a n th
order non-linearity and a 2 X^2 T we have just taken this term as if we have assumed a 0, a1, and
all other terms are 0. okay, so immediately if we just put it as whatever modulation we had, what
we will get and then rearranging them there will be two of that whole θT, three of that θ T, four of
that θ T, rearranging all those things I can just write it as C 0 + C 1 into Cos of-course the C0 C1
will be some there will be some dependency with respect to a0 a1 a2 all those things. okay so
they can be means that can be a corresponding equations.

So that will be Cos ω CT plus of-course it should be general ideally tan inverse. so I can I can just
write it that way, it is just Tan inverse Kf, let us write a(t )on the just for the brevity of it okay, so
tan inverse this plus some c2 into Cos 2 ωCT+ 2, sorry the KF should be there, no that is fine.
okay, so this is that, then 2 into tan inverse Kf at. right, and then up to Cn Cos n ωC T plus n into
Tan inverse KF into a T. right, this is what will be happening, now what I do. I need a FM which
has n times deviation .

710
So I will put a filter around this so at n ω C, instead of putting it at Omega C, I will be putting a
filter around n ω C with a prop appropriate FM bandwidth. because, now the FM bandwidth will
be recalculated. because it has n times deviation. okay, so if earlier with narrowband FM it was 𝛿
F, now it will be n 𝛿 F. okay so earlier if this was 𝛿 F, now it should be n into 𝛿 F. so if I just put a
filter over there I will be getting just this term so that is actually a FM modulated signal but what
has happened, see this tan inverse this one can still be approximated as KF into at, because the
original KF into at that remains still small I can write that as KF at only thing is that that gets n
times multiplied.

so overall I get a wider band FM. okay this is the best part of it. okay so that is what Armstrong
did in a very ingenious way, and this particular technique is called as frequency multiplier. it is an
indirect method and you can see already it is doing nothing but multiplication of frequency, it is
generating that. so it is generating your FM modulated signal at n ω C. okay with a deviation
which is n times the deviation of the corresponding narrowband FM okay.

So with that understanding we will try to see what are the shortcomings of this thing. okay, so
that is something we will try to analyze in the next phase or next class and we will also try to see
exactly in a practical scenario how do you design a FM transceiver. all this multiplicative factor
how we take into account if we have a targeted frequency, where the FM has to be put that means
some ω C is targeted. how do we actually manipulate that, because here something is not in our
hand we will explain that in the next class. okay, thank you.

711
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 51: Frequency Modulation (Contd.)

Okay, so we have already started discussing about generation of FM, right. We have seen how to
generate a narrowband FM which is a simple way to generate (means) whatever circuitry we
have learned already for AM modulation, almost the same thing we can use (utilize), so that was
the advantage, so at that time people were happy to use the same circuitry and then we could see
that with a kind of nonlinear devices, we can actually do frequency multiplication which was
demonstrated by Armstrong and then we could actually generate even wider band FM, okay.

So, you have to remember we will also see the other way of generating FM, but at that time
historically you have to see, at that time probably those circuits of direct generation of FM were
not available, so that is why probably this indirect way of generating FM was popular at that
time, so people who are just trying to see if we can get an engineer circuit which is almost
derived from the amplitude modulation part, okay, so that was the sole reason why people started
going into (means) narrowband FM to wider band FM.

But what you should appreciate in this course again and again I'm probably exerting that; that all
this circuitry actually comes from your mathematical understanding of the signal, okay, that's
very important that everywhere all the circuitry you are actually doing, you are doing it from the
mathematical understanding of the underground or underlying equations, okay, so this is
something which is very important in (communication school) any communication course. We
1

712
will see that, okay, but anyway, now let us try to see how do we in a practical circuit generate a
wider band FM, okay.

(Refer Slide Time: 02:35)

[ ∫−∞ ]
4A
= cos ωct + tan−1 k f m(α)d α
π
t

[ ∫−∞ ]
x (t) = cos2 ωct + k f m(α)d x

y(t) = a 2 x 2(t)
t

[ ∫−∞ ]
2
y(t) = a 2 cos ωct + k f m(α)d x

[ ]
1 1
∫−κ
= a 2 + a 2 cos 2ωct + 2Kf m(α)d α
2 2

So, what will happen? In any wider band FM, we’ll specify something, so last time whatever we
have derived probably, so this was something we were doing, so whenever we made a square, we

713
could see that it's actually going up to the square of it, okay, so the frequency deviation becomes
square. If we have an nth order (this one) non-linearity then it goes up to nth order; but the
problem with that is the central frequency also goes to the nth order.

So, whenever we use an nth order non-linearity, so basically the initial frequency deviation
suppose that was Δf, so I need a Δf which is ‘n’ times this, so I get nΔf but by-product of that is
my frequency also gets shifted to nωc, so now we will try to see that if we target a particular
frequency where we wish to put the FM and we target a different whatever I wish independent of
that frequency, a deviation how do we really make a circuitry out of that? okay, so that is actually
(means) where we'll be doing this FM in multiple stages, okay, or this nonlinear non-linearity
that we are putting, we’ll actually do it in multiple stages, so let us try to appreciate that part.

So, suppose we have some message signal m(t), we integrate it and we pass it through that
DSBSC, so that is that narrowband FM modulator, so this is the overall narrowband FM
modulator, okay, so after this the narrowband FM modulator will have some specification. First
of all, it will have a center frequency, so let us say ‘fc1’. Let us say that is 200 kHz, okay, we are
just taking one example which is also given in the book of B. P. Lathi, so we're just taking that
example for simplicity.

And suppose that has a Δf which is just 25 Hz which is small enough (pretty small compared to
that ‘fc1’), right, so this is this is pretty small Δf, so basically this can be characterized as a
narrowband FM, no problem in that. Suppose I want to actually generate a FM signal at
frequency 91.2 MHz, so this is something you are used to (91.2 FM), okay, so that's where I wish
to put my signal and the Δf at that signal frequency also is specified, so let us call that as some
‘fc1’ okay and there Δf1 that's also specified, that it is having 76.8 kHz, so that's the frequency
deviation that I want, okay.

So, this is the deviation I actually wish to have. This is actually taken from actual example.
That's exactly what deviation they have at 91.2 FM, okay, so that's 76.8 kHz overall bandwidth
will be double of that, okay, because with that frequency the actual voice bandwidth will be

714
much smaller, so Carson's formula you can put and we are almost getting similar thing around
150 kHz of bandwidth, okay, so from this Δf to this.

Immediately what I will be trying, probably that means I need a multiplication. How much
multiplication? That should be this Δf divided by this Δf. That I’ll see immediately, so if I just do
that, immediately I will get a number, but the problem is once I get that number, if I directly
apply the multiplication, this will not take it to 91.2, to some other frequency.

This is where I will have problem, okay, so I have to do it the way it is being done in stages,
okay, so what I will do? I will first multiply it to take it to a higher deviation and higher this one.
We’ll of course appreciate the design of it, okay, with a simple problem; but then what I will do?
Because I want to now shift my carrier frequency, so I will just do a translation, that's always
easy. I (give a) take a local oscillator, from there a cosinusoidal signal and then I just get a
translation with a band pass filter, again I will be picking one of the particular frequency band
where I want to put that, okay.

So, this is something I will be employing, so I will be actually putting through a local oscillator
and then local oscillator will have its own ‘fLo’, let's say and then whenever I multiply these two
what will happen? Wherever I have whatever that ‘fc’ that centre frequency will be, it will
generate that fc + fLo and that fc – fLo. Now (I can) if I wish to have fc + fLo, I will just filter that
out and then do further multiplication.

In that paper what will happen? I will have an extra variability with the local oscillator frequency
and then I will be able to probably adjust both these things, that how much stages of
multiplication so that multiplication stage I will subdivide into two parts and then in between I’ll
put a local oscillator which translates the frequency, so that after doing these two multiplication
along with this local oscillator facilitated (translation) frequency translation, I will get exactly at
this frequency, so this is (what so) general FM what they do.

If this is the case then there will be a frequency multiplier, okay, so this is the frequency
multiplier we are talking about, that means nth order non-linearity which translates both the
things. It increases the deviation of FM (as well as), so this is the indirect method, so with that
4

715
we probably create ‘fc2’ which will be 12.8 MHz, so from 200 kHz to 12.8 MHz accordingly you
can see how much multiplication has to be given, so this actually multiplied into 64, so 200 x 64
you will get this and according to your Δf also will be changed.

So, if this was Δf1, this should be Δf2 which should be 25 x 64, right, so that becomes 1.6 kHz. At
this you give a frequency converter, so that is the frequency converter where you actually from a
local oscillator, you take a signal which is 10.9 MHz. This is whatever we are saying that's
actually being done, so what will happen? If I just do this, (your) this frequency the ‘fc2’ will get
translated to a higher frequency but Δf will remain the same because I am not (tinkering with I
am not) passing it through a local oscillator.

It's just a frequency translator. It will just take it to a higher frequency, nothing else. The
frequency deviation will not have (any means) any effect on that, so what will happen? I will
now go to, so this is 12.8 minus this, probably my ‘fc3’ should be 1.9 MHz, okay, so as I have
told this might create both the frequency. This plus this and this minus this, so here I am
interested in this minus this, so I will be putting accordingly a bandpass filter which will just take
out my FM signal.

So, Δf3 remains the same 1.6 MHz, right. Then in the last stage I will be putting another
frequency multiplier, that's multiplication with 48 and then you will see immediately you get a
FM which is having ‘fc4’ that is exactly 91.2 MHz and the Δf let's say for that will be exactly this
76.8 kHz, so basically you have employed a multiplication of 64 x 48, but you didn't do it
simultaneously or together.

(Refer Slide Time: 11:35)

716
 fc1 = 200 KHz Δ f1 = 25 Hz
fc2 = 12.8 MHz Δ f1 = 1.6 KHz
fc3 = 1.9 MHz Δ f3 = 1.6 KHz
fc4 = 91.2 MHz Δ f4 = 76.8 KHz

What you did? You’ve splitted this multiplication, so you have employed two circuit to multiply
it, okay do the frequency multiplication and in between you have adjusted your frequency so that
with the final multiplication you will get your desired FM signal, okay.

So, if we just try to do a simple problem probably it will be more clearer, so let's say I have a
narrowband FM where the ‘fc1’ that's the initial carrier which is 20 kHz, okay; and this Δf1 the
initial that narrowband FM initial frequency deviation that's actually 5 Hz, okay, so that's very
small so it should be classified as narrow band FM; and what I want is, I have a desired wide-
band FM where the specifications are ‘fc4’ that's 97.3 x 106 or I should say 97.3 MHz FM and Δf4
is 10 kHz, okay, 10.24 kHz, so that's the Δf4 I wish.

So, this is where I want to model it and that's the frequency deviation that I wish; and what we
have also told, that all the multiplier stages are actually (this can all means) we only have
6

717
quadratic non-linearity, ok, nothing else; so it must be 2 to the power something we whatever we
wish to realize that should be just it has quadratic, so it can only make a square, so it can
generate a multiplier which is multiplication by 2, so if I just want to have multiplication by 4 I'll
be putting 2 stages, but multiplication by 3 that's not realizable because I only have quadratic
non-linearity, ok.

So, whatever I will have, whatever multiplication stages I will be putting that should be 2 to the
power something, nothing else can be done over here, okay, so that's one restriction that is
coming from the circuit side and there is another restriction that your local oscillator ‘fLo’ must
be within this 400 kHz to 500 kHz, okay. That's another restriction, okay, so this is one restriction
and your multiplications are (all or frequency multiplier) that's actually only quadratic non-
linearity is available, okay, so this is something you already know, okay, so therefore the two
stages we were talking about that there will be two multiplication stages, in between there should
be a local oscillator induced frequency translation, okay.

So, the two stages with that whatever multiplication we can do let us say in the first stage we
have 2n1 and in the next stage we have 2n2, so overall multiplication will be 2n1 x 2n2, so this is
almost like the previous one that's actually here 64 and here 48, but 48 cannot be realized over
here, so whatever will be putting over here that must be realized at 2 to the power something, ok,
so, two stages we are putting, so overall it should be 2n1+n2, okay, so that must be the overall
multiplication factor, so I can get a value of n1 + n2.

How do I get that? So, Δf4/Δf1 that must be the overall multiplication because this is the
multiplication available to us so that should be this divided by, so that becomes 2048, okay,
which is the way it is it's 2 to the power something, it is actually 211 right, so I can write this as
211, so I immediately get a equation n1 + n2 must be 11, so this is something I know. Now how I
split that depends on me. Whether I should get five and six or six and five or seven and four or
four and seven, whichever way I wish to do it, that depends on me; but there is also another thing
I have to keep in mind, that my local oscillator must be restricted to this one.

(Refer Slide Time: 17:25)

718
 NBFC fc1 = 20KHz Δ f1 = 5 Hz
Desired WBFM ⟹ fc4 = 97.3MHz Δ f4 = 10,240Hz
400kHz ⩽ fLO ⩽ 500kHz
2n1 ⋅ 2n 2 = 2n1+n 2 = 211
Δf 10240
∵ 4 = = 2048
Δ f1 5
n1 + n2 = 11

Whatever I keep I will have to finally do a translation. That will be done by a local oscillator and
after doing that translation, I should also match the frequency, so that it goes to 97.3 from this 20
kHz. That's also something I will have to keep in mind, so let us try to see what will be
happening, so let us say what is suppose the first stage is ‘n1’ multiplication, second stage is ‘n2’
multiplication, so and the convention I have taken in the previous example that this becomes ‘fc1’

then what should be my ‘fc2’ that is this ‘2n1’ is the 64, so that multiplied by this one.

So, if I say ‘fc2’ that must be ‘fc1’ which was 20 kHz according to our example, okay, multiplied
by 2n1, so I get a value of ‘fc2’, right, so this is something I have got; and correspondingly I can
also write ‘fc4’, that must be 2n2 x fc3, right. This ‘fc3’ is being generated due to the local
oscillator, means due to the interaction of ‘fc2’ and local oscillator right, so if you just try to see
over here that's the ‘fc1’ Δf. From ‘fc1’ with multiplication I get ‘fc2’, okay and this ‘fc4’ is my final

719
target. ‘fc4’ is generated due to this multiplication, so 2n2 x fc3 should be ‘fc4’. That is what we
have written. ‘fc4’ must be 2n2 x fc3 and ‘fc2’ should be 2n1 x fc1. That's something we have written,
so these two equations are always valid.

Now what will happen? Once we get these two equations, now I have another relationship
between ‘fc2’ and ‘fc3’, so how the ‘fc2’ and ‘fc3’ is being generated? My ‘fc3’ can be either fc2±fLo
or it can be fLo±fc2, if fLo is much bigger, okay, so it depends on these four possibilities, so it can
be also fLo±fc2, so I have three options. Actually ‘+’ will be in both cases same, so I should write
this, so there are these three or options which I can play with. With that the frequency should all
match, so this equation, this equation and this equation, wherever it satisfies that ‘fLo’ is within
400 and 500 MHz; as well as all n1 and n2 are integer value. If these all these things are
satisfiable, then only I can get a solution.

So, what we'll have to do is; now (start doing) start taking all three cases one by one, so the
solution I won't be doing. It's easy you can do that, so the solution we’ll be just take one by one.
For each case try to see if we get something meaningful. You try to calculate the ‘fLo’ from this
particular equation, so suppose first case I take ‘fc3’ is equal to fc2+fLo, then immediately ‘fLo’ will
be fc3 − fc2, ok, so ‘fc3’ you can represent with respect to ‘fc4’, which you know, ok; and ‘fc2’ you
already know.

(Refer Slide Time: 20:40)

720
fc2 = 20KHz × 2n1
fc4 = 2n 2 ⋅ fc3
fc3 = fc2 ± fL 0 /fL 0 − fc2
fc3 = fc2 + fL 0 = fL0 = fc3 − fc2

So, once you start putting that, we'll be getting our relationship for ‘fLo’. Now ‘fLo’ you put that
condition it must be 400 to 500 and then try to see whether n1 or n2, I can choose as integer. Any
integer value will give me a feasible solution. If that is not the case, go for the second case and if
that is also not satisfiable, go for the third case and you'll see probably solution in one of them,
ok, so, this is how a general FM transmitter when you try to do it in indirect method is being
generated, okay.

So, of course we'll have a unit which is called the nonlinear modulator, ok, so that nonlinear
modulator generally people will be designing a unique one because they don't want to design for
anything customized that you say okay, I want 63 non-linearity. That's very difficult to design, so
generally people will say ok I will just give you quadratic one; and then you want to make
whatever you like this particular practical circuitry we have seen that I will give you a quadratic
one. Now you whatever you wish to make you take multiple such quadratic one and you just
design the circuit, okay.

But you can see already there are difficulties in FM modulation, so what's happening? Whenever
I do that, initial modulation that narrowband one was very easy, but after that it's just too many
non-linearity stages followed by a frequency translator and then again followed by some more
non-linearity stages and this non-linearity has to be very closely monitored and controlled. Little
bit of biasing factor which goes here and there your complete non-linearity will change and
accordingly there will be different frequency which will be generated.

So, probably that was a very problematic one in FM generation, so what we'll try to see is what
happens (if we means) if we can try to get a FM modulator which is a little bit simpler than this,
okay and in that quest only people started getting the circuitry for direct modulation of FM, okay,
so we will now try to talk about the direct modulation, okay, so FM direct modulation has been
actually done.

10

721
This is just there is only one device which can do that. That's called VCO or Voltage Controlled
Oscillator. So, this is the single device which can actually do FM modulation directly, okay, so
we'll try to just explore what this VCO is; so basically what happens? People realize that this is a
Voltage Controlled oscillator. That means it's actually an oscillator where you give an input
voltage, okay, so whatever input voltage we’ll be giving, depending on the amplitude of that
voltage, your oscillation frequency should be varying, okay. That is why it is named as or termed
as Voltage Controlled Oscillator, okay.

So, generally an oscillator is designed by a combination of inductor and capacitor, right, so this is
something we all know, so in any of the oscillator either you take it Hartley or Colpitts
oscillators, so our oscillation frequency is generally where L is the inductor and C is the
capacitor. What people have realized over years that if you have a semiconductor device and you
reverse bias it, that is a means we know that in reverse bias it acts as a capacitor and what people
have seen also that you can actually control that capacitor with respect to the input bias voltage,
so that was the sole initiative which has started giving some hope to generate Voltage Controlled
Oscillator.

So, basically what you can now do is that capacitor value can be (we are targeting this capacitor
value can be) now modulated with our input signal, okay, so the bias you give that reverse bias
suppose semiconductor device which has a capacitor value and that you use as part of that
semiconductor capacitor sorry sorry that oscillator whether it's be it Hartley or Colpitts
whichever oscillator you have, you take this as a capacitor and reverse bias semiconductor and
then the wires well voltage you give to that, the capacitance value will be dependent on that. So,
that's called actually vari-cap, okay, so it's known as vari-cap or variable capacitor, okay.

So, if we have that, so let us say we have that equation of that capacitor value is something like
this Co – K m(t), so these are actually constants of that vari-cap, so what happens if you give an
input voltage and you will have a capacitor, that capacitance (that) value will be dependent on
this and it has a linear variation over a small range. Of course, it's not for any range. Any input
voltage you give, it will not be following a linear relationship, but for a particular region you'll
be finding that it is following a linear region or linear variability, so we'll be targeting that one,
okay.

11

722
So, immediately what happens to my ‘ωo’ of that particular oscillator? So, that should be my L
does not vary but C is now actually Co – K m(t). I can write it this way or I can write it as I take
out LCo, so I'll be having something called [1 – (K m(t)/Co)]1/2 or I can write it as (1/√LCo) [1 –
(K m(t)/Co)]−1/2. You know why I am writing this, because now I will do a Taylor series
expansion. That's something which I will try to do.

As long as this Km(t) << Co, this particular term will be small enough and if I do a Taylor series
expansion what I will get? I will get the first term, then the second term is still valid because it's
the linear term after that it's just quadratic of this one, where this Km(t)/Co is much less than 1,
so I can actually neglect those terms. So, basically immediately I will get this ωo = (1/√LCo) [1 +
(K m(t)/2Co)] and all other terms will be neglected, so I have just done a Taylor series expansion
as long as this particular thing is valid.

(Refer Slide Time: 28:06)

12

723
 1
ω0 =
LC
c = c0 − k m(t)
1
ω0 =
L [C0 − Km(t)]
1
=
LC1 (1 − C0 )
km(t) 1/2

−1/2

LC1 ( C0 )
1 k m(t)
= 1−

LC1 ( C0 )
1 k m(t)
= 1− ∵ Km(t) ≪ C0

So, generally what will happen this m(t) it remains linear in a very small region, so this condition
is almost all the time (means) valid, so that is the case we get a ‘ωo’, ok, so I write that ‘ωo’ as
ωc[1 + Km(t)/2Co], where ωc = 1/√LCo. I can write that, ok or I can write this as ωc + Kf m(t),
where Kf is just ωc (K/2Co), fine, so this is actually the vari-caps rest capacitor probably when
m(t) is 0. This is the inductor value which is fixed for that particular oscillator and these are all
the parameters which are coming from the particular oscillator.

So, immediately I can get some Kf of that particular oscillator and I get some ωc of that
oscillator, so this ωc is actually termed as free running frequency. Whenever m(t) is 0, that's the
frequency with which it will be oscillating; and then depending on m(t) there will be a linear
variation where the slope of that variation is Kf which is also determined by the particular
capacitor, so this particular oscillator now what is happening? His frequency is a potentially
linear function of my input voltage, so that is what I wanted to design.

13

724
(Refer Slide Time: 30:00)

[ 2C0 ]
Km(t) 1
ω0 = ωc 1 + ωc =
LC0
wc k
= ωc + kf m(t) kf =
2C0

I wanted to design a particular oscillator whose oscillation frequency is directly dependent or

linearly dependent on the input voltage, so once this is being done, it's very easy you'll now see
to generate FM, okay, so in the next class what we'll try to explore is, with this how do I generate
FM and then probably after generation we will try to appreciate how do you demodulate FM,
okay, so that's something next we'll be trying to see. Thank you.

14

725
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 52: Frequency Modulation (Contd.)

(Refer Slide Time: 00:25)

726
ω0 = ωc + Kf m(t)
t

[ ∫−∞ ]
cos ωct + kf m(α)dα

Okay, so what we have started exploring is the frequency output of the oscillator we were
targeting which is; we have proven that that should be ωc+Kf m(t). We have defined what is ωc
and what is Kf, okay, so this is something we have already discussed, right. Now FM generation
once we have this circuit that means a particular oscillator with an inductor which is fixed and a
vary-cap where the it varies with the input that bias voltage, okay.

So, once we have that, we will be getting the output oscillation frequency is accordingly varying,
okay, so what I will do, if you now start giving an input voltage as your m(t) into that vary-cap,
immediately the output oscillation that we will be getting, that must be FM modulated signal
because that will be the m(t), immediately the oscillation frequency, suppose it generates A cos
some frequency, so that must be this, okay so and immediately you will get if this is the
frequency the phase will be ωct.

This is something we have already explored, that should be Kf integration -∞ to t m(α) dα, that
is actually the FM modulated signal, so that gives me a direct modulation, where this Kf is
something which is chosen by that particular vary-cap, right, so whatever I choose as my VCO, I
will be getting accordingly the FM deviation and everything, so I can choose my parameter
accordingly and I will get my FM, so I have to choose accordingly the free running frequency for
that FM as well as the Kf value and that will give me the correct output, right.

That is called the direct method, so you can see already lot of complicacy that were arising from
narrowband FM and then with an indirect method to generate wide band FM, that goes away and
of course there also there was a problem that narrowband FM, if it is not sufficiently narrower
then there will be a tan-1 (Kf AT) term right, so instead of just Kf AT terms, so that problem was
there, whereas here there is no problem like that, so this is called the direct method of FM
generation, okay.

727
So, let us try to see now how do you do a FM demodulation? Okay, so our next target should be
FM modulation is almost done, so we have learned narrowband FM how that can be generated;
then for wider band FM we have seen two methods, one is direct, one is indirect. So, in the
indirect method we have to just put frequency multiplier and we have seen also how to adjust the
center frequency as well as frequency deviation, whereas in direct method it is just choosing a
proper VCO parameter okay.

Sorry the direct method it is just choosing a proper VCO parameter and it is very easy. Just
across the vary-cap you give your input voltage and you will be getting your corresponding FM
modulated output at the output of that VCO, okay, so let us now try to discuss about FM
demodulation. Again, you will see just the mathematics tells us what should be the FM
demodulation. Suppose I have a ψFM(t).

(Refer Slide Time: 04:06)

728
 t

[ ∫−∞ ]
ψFM (t) = A cos ωct + kf m(α)dα

[ ]
= A [ωct + kf m(t)] sin ωct + kf
∫
′
ψFM (t) m(α)dα
−∞
Kf m(t) + ωc ≥ 0
Kf mp + ωc ≥ 0
ωc ≥ Kf mp

And what it is? it is A cos (ωct + Kf) right, this is my ψFM(t). I want to demodulate it. The first
thing that I will do is, if you carefully see this, I just differentiate this signal, so let us try to do
this. What do we get? So, simple differentiation you do, so whenever we differentiate, it should
have that chain rule, so differentiation of cos must be giving me sin and then inside part also has
to be differentiated, right, so inside part if I differentiate, this must be ωct must give me ωc
because I am different shading with respect to ‘t’ right.

And then Kf is a constant differentiation and integration actually cancel each other and this will
give me m(t), right, so that is the whole thing that I get after differentiation and cos must be also
differentiated, so that must give me sin (ωct + Kf), right, so what we can see, that FM modulated
signal if we just differentiate what happens, it will generate another sinusoidal but the whole
signal varying part comes into the amplitude, okay, now what we do, whatever the sinusoidal, so
it will actually look like this, suppose ideal differentiator how does that look? That is generally
j2πf, okay.

Differentiator d/dt if you take the corresponding Fourier transform, so that looks like this, so
ideal differentiator should be something like this constant means if I put the amplitude part of it,
so that should look like this it is a linear function of ‘f’, okay, so if my ωc is somewhere over
here and it remains linear over there means if it is an ideal differentiator, so I should be expecting
something like this from that. That is the output, okay and after this, how the signal will look
like? So, I had initially FM moderated signal.

729
So, which was having something some modulation, so it was varying with respect to the
amplitude of it. It is just the frequency deviation. If I just differentiate what happens? This
frequency variation also comes into amplitude, whenever there is a higher variation of frequency,
so, basically, I will have accordingly a higher amplitude and correspondingly if there is a lower
variation, I will get a lower amplitude, right.

So, this will be means once I pass it through a differentiator, it will look like this; and inside also
same pattern will be coming, something like this, okay. This is something which will be
happening, so whatever is happening, what we can see that message signal is almost in the
envelope, so all I have to now do is envelope detection, nothing else as long as I can ensure that
envelope detection gives me a signal back, as long as this is not doing a zero crossing. We have
already learned that in AM modulation.

So, what is the condition that this will not be crossing zero? If this remains always positive, so
that means I have to write Kf m(t)+ωc must be greater than 0, right, this is something I will have
to write. Now what is the minimum value of this where it might cross 0? That means the minima
of this must also; this is true for all ‘t’ that means the minima of this must be also greater than 0,
so minima let us say that is – or that is –mp let us say, okay.

So then I can write –Kf mp, where mp is a positive number – of that is the minimum, +ωc must be
≥ 0 or ωc must be ≥ Kf mp, so this is the condition I get. What is Kf mp? That is the frequency
deviation we have talked about, okay, so basically my carrier frequency must be bigger than
frequency deviation. This is something which always I will be doing because if the carrier
frequency because I know from Carson's formula that the bandwidth is more than 2 times of this
deviation. If my carrier is not even bigger than this particular thing there will be aliasing, right.

We are actually putting at the carrier ωc, now the bandwidth is definitely bigger than my Δf,
twice of Δf right, so or Δωc if this is already bigger than my ωc then what will happen? This will
come even beyond ‘0’ and there will be aliasing, so definitely whenever I do FM, I will make
sure my ωc is bigger than that Δf at least. It has to be bigger than Δf+B, according to Carson's

730
formula right, so it has to be done. If that is being always done, so this condition will always be
there that is prevalent. We know that this will be happening.

If that is the case, I know that the envelope will always remain above 0, it will not have any 0
crossing, so I will actually this is guaranteed that I will never have to cross 0, so if I just detect
the envelope, I will get my signal back, so that makes the FM demodulation pretty simple. All
you have to do; you have to pass it through an ideal differentiator followed by a simple envelope
detector, the one we have designed for AM modulation demodulation.

So, this is what happens and you know that it is guaranteed as long as you make sure that the FM
modulated signal that you generate that is not creating aliasing, that means ωc is bigger than at
least that Δω, you are pretty sure that your envelope will be above always and if you just track
the envelope, you will get your message signal back. So, tracking the envelope will give you ωc
+Kf m(t). This term will be gone because you are just tracking the envelope. That sinusoidal
variation will be gone.

(Refer Slide Time: 11:20)

731
 j2π f RC
H( f ) = ≃ j2π f RC 2π f RC ≪ 1
1 + j2π f RC
1
ωc ≪
RC

So, you will just get Aωc+AKf m(t). Now this is just a DC term, you block the DC with a
capacitor, you will get this part which is the message signal, so indirect sorry this method of
(means) this differentiator induced method of FM demodulation, that is very simple. All you’ll
have to do is, you have to find out ideal differentiator circuit. Now we will discuss about that.
That is little bit difficult (to find out the ideal differentiator), okay.

So, generally what people have done, means you cannot actually get an ideal differentiator. That
is not possible in circuitry. Every capacitor we’ll put, there will be some spurious resistor in that
capacitor. There will be some other effect and always you will see that it cannot work as an ideal
differentiator, so that is not possible. There is nothing called ideal differentiator. What we will
do? We will employ something like this RC circuit. This is actually a high-pass filter, right.
7

732
So, how the high-pass filter will look like? It will actually look like this, but there is a region,
generally high pass filter we are not bothered about this roll-off. Generally, we want to neglect
that roll-off in high pass filter. We are most more concerned about where the filter response is flat
but this is further for the FM demodulation. We are actually bothered more about this roll-off, so
we want to see where exactly it remains linear. That is where it almost behaves like an ideal
differentiator.

So, basically we have to target a particular frequency zone or we have to design our high pass
filter in such a way that in the frequency of interest which is this ωc and around that Δf+ and Δf−
right, so that region it remains linear, okay, so that is something we will have to find out, so let us
for our ideal this particular filter, let us try to see that what is the transfer function? So that
should be j2πfRC, if you just put it accordingly, so this is what we get, okay, H(f) will be just
this, okay. Now this can be approximated as j2πfRC, if this j2πfRC << 1, because then this term
will be neglected.

So, there will be only 1, so I get this. This is almost like an ideal differentiator, so this will
happen if this condition is valid, so now I can get a condition on my ωc, okay, so what I can do?
This j2πfc let us put it as ωc, so, ωc << 1/RC, okay, so this is the condition I get. If I choose it
accordingly then the ωc will be falling in this region where it looks like ideal differentiator, so all
I will have to do is, I have to choose a RC corresponding to my ωc that has been put over there,
accordingly I put my RC so maybe I can fix C and then try to find out what should be my R and
then try to find out the RC value which is much bigger than this ωc.

Then I know that around that ωc, it remains linear, because this particular transfer function
characteristics will be valid. That approximation also will be valid and it will almost behave like
an ideal (low-pass field sorry) high-pass filter sorry ideal differentiator, right and then if I just
pass my signal through this, at the output I will get a differentiator or differentiated output at that
particular frequency, so this is all that I will have to do. It is simple enough. All we will have to
do is we have to design our RC accordingly and then try to get a particular differentiation circuit.

733
Once this is being done, I know that the (means) envelope tracking will be very simple.
Whatever we have done at AM, that has to be mimicked over here, okay, this is one way of
(means) demodulating FM.

(Refer Slide Time: 16:16)

t
π
∫−∞
θi(t) = kf m(α)dα +
2
t

[ 2]
π
∫−∞
A sin ωct + kf m(α)dα +

t
π
∫−∞
θ0(t) = kf m(α)dα + − θe(t)
2
1 1
e0(t) = θ0(t) = Kf m(t) − θe(t)
c c
9

734
There is another way, that is FM demodulation using PLL, okay so if you remember that for PLL
we have also used a VCO right. That was how we have designed PLL, so for FM modulation we
have already used VCO. Now for demodulation also we will be just using PLL and PLL also has
VCO in it, so basically FM modulation demodulation both will have key component as the VCO,
so once you have VCO, you can do both modulation demodulation if we can appreciate this
particular circuitry.

So, let us try to see how FM demodulation can be done with a PLL, so this is another application
of PLL which is coming out. Earlier we have seen that for carrier tracking. Probably PLL is very
good and with that target only because we are talking about phase locked loop which was with
the target of carrier locking, carrier phase and frequency locking we have discussed about that
effectively.

So, now we will try to see the other application of PLL which is FM demodulation, so let us say
the PLL generates its input this one to which it gets locked, that is a sinusoidal, so that is
something like sin (ωct + θI(t)) where θI(t) is the input, so what we’ll do, we’ll actually give this
(FM signal) FM modulated signal to the PLL input, so what will happen to this θI(t)? That must
be whatever phase it is getting, so FM is A cosωct + the FM modulated part that Kf At. Okay, so
that means that θI(t) must have this Kf At + because it’s cos and PLL input takes sin, so there
should be a π/2 phase shift.

So, therefore θI(t) should be Kf ∫−∞ to t m(α) dα + π/2. If this becomes θI(t), so immediately I can
put my FM input should be A sin ωct, so I can put A sin (ωct + θI(t)) which is Kf ∫−∞ to t m(α) dα
+ π/2, right, that immediately becomes cos, so it is A cos ωct + this. That is actually FM signal, so
therefore if I wish to put FM signal to PLL input and if we correspond it, we get my θI(t) as this
one, right. This is fine.

Let us say the output phase error that is generated due to the locking that is θe(t). Of course this
should be small enough we will see that, so basically therefore what is θo(t) which is being
generated after the VCO (means the PLL output)? That should be this input phase − this θ, right,

10

735
so input phase is Kf ∫−∞ to t m(α) dα + π/2 − θe(t), right, so that is what it will lock to and it will
get this output phase, but what we also know that this θo(t).

If we now try to see the error signal that is being generated by PLL, so this that is eo(t) that is
nothing but the differentiation of this output phase, and with a factor 1/C which is the factor of
PLL right, so 1/C differentiation of θo(t). Now let us try to differentiate it, so it should be 1/C and
if θo(t) we have to differentiate, so this term will be gone, I will have a differentiation of this one
and I will have differentiation of this one, so differentiation of that one is Kf and if I just
differentiate, integration differentiation cancels each other so I get m(t), right.

And I will get differentiation of this. Now that is exactly the error frequency because phase
differentiation is the error frequency. PLL if it is properly locked, then frequency error must be
zero, so this I can almost say it should be (1/C) Kf m(t) which is very good, because that is
actually the message signal with some constant factor which I do not bother, so basically at the
error of PLL I get FM demodulated signal, if I give at the input of PLL FM modulated signal.

That is, whatever analysis we have done for PLL that actually directly comes from (means it
directly comes from) there. If we (means we) just take a PLL circuit, earlier we are not bothered
about the error signal of PLL, we are not bothered about that.

Now in this particular demodulation what we will try to do? We will try to suppose at that time
we were bothered about the VCO generated output, okay, now we are actually looking into the
error signal that is being generated after the PLL (after the loop filter), so what we have to do is
we’ll give as input the FM modulated signal. We (do not actually tracking we) are not tracking
anything, so we are not worried about the VCO generated output, we are not worried about now.

Now we are not worried about that. We’ll take that from the loop filter whatever is coming out,
we will take that and we could prove mathematically, again you can see why we are able to use
this, because mathematically this is getting proved, so it is the circuit almost operating the way it

11

736
is defined, its transfer function and everything operating on the signal this is what it is giving, so
once you prove that mathematically you know that with this circuit, I can do these things.

So, we can see that the error signal is actually becoming the differentiation of that particular
thing, okay so differentiation sorry differentiation of this output signal and that becomes happens
to be proportional to my message signal, so that means at the error I am getting my demodulated
signal, right, so what we have so far done is by we employed two methods of FM. One was
direct method; one was indirect method. In direct method we could be used VCO; indirect
method we had to do a lot of things, a lot of multiplier part lot of nonlinear circuit and all those
things.

And then for demodulation also we could see there are two methods. One is through
differentiation followed by envelope detector which is simple enough, so for differentiation you
will probably have to employ a linear part of a high-pass filter okay and the other part is just use
PLL for demodulating FM signal, so these are two things that can be done to demodulate FM
signal, okay.

So far we have analyzed about FM bandwidth and how we can generate or demodulate FM
modulated signal and FM demodulated signal, okay, so this is something we have analyzed, but
first of all we need to understand that why we should study FM or why one should employ FM?
So, this is something we will now be exploring, so we will try to see or try to appreciate that FM
has some good characteristics. Initially people would have thought probably it is the bandwidth
efficient protocol sorry bandwidth efficient modulation technique.

But that is not the case. We have already proven that that this is not probably as bandwidth
efficient as any of the AM modulated signal, but there are some advantages and this is something
which we will be exploring next, so the first thing is we have already discussed that there are any
modulated signal that will be transferring that has some effect when it is passed through a
channel, right.

So, what are those effects. The first effect we have discussed is, if the channel is nonlinear, okay,
so that is the first thing which happens, so we have also seen that if we have AM modulated
12

737
signal and if there is a non-linearity, there will be detrimental effect. This is something we have
already appreciated and we have seen that, okay, so what we now wish to see, that for FM is
there any effect or is it very (means very) much superior in terms of modulation that any non-
linearity in the channel can be just rejected by FM, okay. So that is something we wish to see.

(Refer Slide Time: 25:37)

y(t) = a0 + a1x(t) + a2 x 2(t) + ⋯ + an x n(t)

t t

[ ∫−∞ ] [ ∫−∞ ]
= c0 + c1 cos ωct + Kf m(α)dα + c2 cos 2ωct + 2Kf m(α)dα + ⋯

So, let us try to appreciate that, so let us say we have an arbitrary non-linearity, so y(t) is some
ao+a1x(t)+a2x2(t) (x(t) is the input), some nth order non-linearity in the channel. Now what will
happen? I will be launching FM signal, so FM signal that means my x(t) should be A cos [ωct+Kf
∫−∞ to t m(α) dα]. So, whatever discussion we are doing, that is equally valid for FM as well as
PM right.

13

738
So, both are equivalent. That is something we have proven. Instead of integration you will be
putting just m(t) over here, okay, so if I just put this x(t), what will be my y(t)? This is something
we have just done for our direct modulation, right, so what will happen? If I just put it over here,
I will be just seeing something like Co+C1 (x sorry) cos of, course this A can be taken inside C,
so do not worry about that, so it should be cos [ωct + Kf ∫−∞ to t m(α) dα] +C2 cos [2ωct +2Kf
∫−∞ to t m(α) dα] and so on up to a nth term, but what has happened now?

This FM modulated thing after it passes through this particular channel will probably get all
these extra higher terms and my receiver what generally that will have? At the means at the front
end of receiver we have already talked about that. It wants to neglect the effect of noise, so it will
have a band pass filter. Where that band pass filter will be centered at? At ωc and the band will be
just FM band, so if I just pass it through that band pass filter, these things all will be neglected,
so all those higher frequency term where this is getting contaminated because the frequency
deviation is getting twice thrice and all those things, they will be all cancelled out.

What will happen? I will only have this particular thing, so even if I have channel non-linearity, I
do not bother about it, because FM automatically due to that band pass filter will cancel out that
effect of non-linearity and it will get pure FM modulated signal even after passing through a
nonlinear channel, so that is a very big advantage which FM has or FM has that edge over
amplitude modulated signal because in amplitude modulated signal if you multiply it, that m2(t)
term will be coming out because it is in the amplitude.

So, that multiplication will create a multiplication in the signal also and then that will create
problem for you, whereas that is not happening over here, okay, so that is a big advantage which
we will have. Whenever we have a channel which is slightly nonlinear, so in that channel if you
just put FM, that is more protected compared to your any form of AM modulated signal, okay, so
this is the first thing where we could get some advantage of FM. We’ll see if there are
interference, see whatever I have told that is actually creating a source of interference also.

So, if the channel is nonlinear and I have here probably at ωc my FM signal but at 2ωc there
might be some others’ FM signal where due to these things, in the channel they will be all

14

739
created, so this will create interference to them. However small that non-linearity is, these a2, a3
coefficients are small, they might be smaller, but they will still create some effect at those
frequencies, so those are treated as interference.

So, what we will see in the next class that in presence of interference, how FM survives, okay, so
that is something we will try to appreciate in the next class; and then at the end we will also try
to do noise analysis and we will be able to prove that in any time in any channel, FM is much
better in terms of noise cancellation compared to any of its AM counterpart, okay, so that is
something we will be proving and with that we will probably end our discussion of FM, okay,
thank you.

15

740
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 53: FM Noise Analysis

Okay, so far, I think we have already seen how do we really see the benefit of FM. So, that is
what we started doing. We have already finished discussing about FM bandwidth. Initially
people thought its bandwidth benefit. We have discarded that and then we could appreciate the
circuits that has to be produced for FM generation and demodulation; and then we started
capturing the advantages of FM. Mostly all the channel impairments probably FM take care of
that in a better fashion compared to its AM counterpart.

So that is what we started discussing. We have already given one example where the channel
non-linearity we have seen that other modulation schemes (like Amplitude Modulation) they are
probably not that good to handle channel non-linearity. Whereas in FM we could see that (it can
vary means) due to the inherent quality of FM it can really handle non-linearity very well. So
that is the sole reason why for a high-power transmission where you need to really put Class C
amplifier which goes into non-linearity.

So, for those kind of transmission FM was the de-facto modulation scheme because it has a good
cancellation mechanism which is inherent to FM. Now today what we'll try to see that in
presence of interference what happens to FM modulation? Okay, or what kind of protection it has
in presence of interference? So, just to capture that because the full-blown analysis is little bit

741
involved. So, what we will try to see is that (as if FM means) we are trying to transmit a FM
which is either a constant thing or it does not have any modulation. It is just the carrier we are
transmitting but of course you will see the essence of the cancellation (is already means that) will
be understood through this analysis.

(Refer Slide Time: 02:29)

742
 A cos(ωct)
I cos((ωc + ω)t)
r (t) = A cos(ωct) + I cos((ωc + ω)t)
= [A + I cos(ωt)]cos (ωct) − I sin(ωt)sin (ωct)
= Er (t)cos [ωct + ψd (t)]

[ [A + I cos(ωt)] ]
I sin(ωt)
ψd (t) = tan−1 A ≫I

[ ]
I sin(ωt)
≃ tan−1
A
sin(ωt)
≃
θ

So, what we will say, we will say unmodulated carrier A cos (ωct) is being transmitted. So, of
course if it is FM modulated so there should be some phase term which is for FM it will be our

743
integral of M α D α x K for PM it will be just Mt into KP probably. So, there should be a phase
term we are just neglecting that. That means unmodulated just the carrier we are sending. Okay,
so there should be a phase term we are just neglecting it for our purpose of analysis and you'll
see that it is not that much required because that the demodulation will still employ FM
demodulation and we will get some insight.

So, let us say this is the desired carrier and nearby that we have some interferer which has a
strength of ‘I’ and it is nearby, so the frequency is (ωc+ω)t. Okay, so it is just this is the carrier
frequency, let us say at ωc and there is nearby some strength ‘I’ this is having strength at (ωc+ω).
So, that is the interferer actually, and we have to see what is the effect of this interference and
what FM does with this interference, so that is something we wish to check.

So basically what will happen? Suppose my received signal will be because in channel, I will
have this carrier as well as this interferer so it will be addition of these two, right. So, I will have
A cos ωct + I cos (ωc+ω)t. This is what I will be getting. I can just open this cos (ωc+ω) and then
I will be getting [A+I cos (ωt)] cos (ωct) just algebraic manipulation −I sin (ωt) sin (ωct) right.
This is what I will be getting at the FM and we are thinking that this I is close enough in the band
pass filtering zone of FM.

So basically it is an interferer which is within the band, so it will not be canceled by the filter, so
it will still be inside; and we want to see if those are the valid interferer and we want to see what
is the effect of that interference into the FM modulated signal, so this is what I we are trying to
analyze, okay. So, this can be written as, because it is cos ωc, sin ωc and this can be represented
as an overall cosinusoidal with an envelope and a composite phase. Okay, where of course this
Er(t) should be this square plus this square and square root; and we are interested in phase
because FM demodulator will actually detect this phase and differentiate it right.

So this phase will be that must be tan-1 we know this divided by this whole thing, so this divided
by this so that should be I sin ωt divided by A + I cos ωt. Now if it is an interferer within the
band, we expect that the interference probably will be smaller in strength because if interference
is almost of similar strength with the signal itself probably nobody can do anything with that. It

744
will still be present, so we will assume that because it's interference so of course ‘A’ must be
much greater than ‘I’ okay.

So, this is an interference in picture also we have almost demonstrated similar thing. So, if that is
the case then I see because it is I cos ωt, this will never be bigger than ‘A’, so this can be
neglected compared to ‘A’ and then this will be so this becomes tan-1 [I sin (ωt)/A]. Again I/A
that is a very small number, so this can still be approximated, so these are all approximations so
tan-1 θ will be just tan θ. I sin (ωt)/A right.

So that is becoming my phase (the FM over all FM phase), okay. So, if a modulated signal and
the phase part of that so when we demodulate it, what do we do? We take the phase and we
actually differentiate it, whatever we get (that should be sorry that is not theta that should be so
we differentiate it) that will be the frequency and that must be my signal. That is how FM
demodulation goes, it detects the phase and then differentiate it to get this right. So that we have
already understood because in the phase it will be integration and then you differentiate it you
get the overall phase. So here also because it is a FM demodulation change so we will be doing
employing same thing, so this particular phase we’ll be differentiating and whatever we’ll be
getting that is my receive signal.

(Refer Slide Time: 07:56)

745
 dt [ ]
d ψd (t) d I sin(ωt)
yd (t) = =
dt A
Iω
= cos(ωt)
A

So, what we can say that d/dt of ψd(t) let us call that as yd(t), that must be the demodulated
signal. So, that must be so d/dt of ψd(t) whatever we have calculated, that approximate value I sin
ωt/A, okay. So, if I differentiate it, it will be (I ω/A) cos ωt right. This is what we get. So,
basically (we have) what we have done? We have tried to see the interference, we have just said
that we are probably transmitting a FM signal but we are just considering it’s unmodulated FM
that means the just the carrier we are (just for testing purpose just a carrier we are) sending and
nearby that carrier within the band of FM, I have an interferer which is at (ωc+ω), okay.

746
So, these two signals are in composite falling my receiver. I am just trying to detect the phase,
the way FM will be doing and phase if you differentiate, that should be the frequency variation
and that we call as a detection. Okay, so this is what I will be detecting right. So, that is my phase
and now let us see what happens, okay. If I just try to plot so this is actually; see actual my FM
modulated one was having nothing, so in the detection I should get a zero thing, okay because
there was no modulation so it must be a zero, so it should be getting zero but it is getting this
thing.

So, that is my interference in this case, (interference getting transformed after doing means FM
demodulation). So, what is that? Let us try to see that. First of all, there is a factor of I/A, okay.
So, we have already said if ‘A’ is bigger than ‘I’ so there is already means the interference is
becoming ‘A’ times lesser, okay. So, when it is interfering with FM because of FM demodulation
process, it is already becoming ‘A’ times lesser. As long as ‘A’ is higher, probably it will be
already lesser, that is one thing. Second thing is, it is getting multiplied with ‘ω’. So, if you see
that ‘ω’ is the separation right.

So if I just plot it with respect to my frequency ‘ω’ and then try to plot this interference how it
will be? At ω = 0 it is 0 and then there is a linear curve, okay. So of course you can always say
that there is A cos ωt term but okay if this cos ωt and because I do not plot it with respect to ‘t’, I
will take the probably worst case scenario, so it is the highest value of this. This is one, so it will
be just Iω/A, so I am plotting the highest value of interference, that's what we are interested in
that how much interference at the peak will be coming to me.

So that at most will be this, which is (I/A) ω, so it goes linear with ‘ω’, so this is the interference.
Here remember deliberately we have taken out the signal, so we should expect after FM
demodulation, zero (nothing) but we are getting something and we are calling that as
interference, so that was just to make the overall analysis simplified. If you give FM, then there
will be more involved term, so we just wanted to simplify that. We are just capturing the
interference effect.

747
So, what we can see that as ‘ω’ is closer (this is a very interesting thing), the interference term is
smaller. So, as the interfering signal comes closer to you, you are more afraid that because the
interference is closer, probably more interference I will be getting, but FM has a cancellation
part. Okay, due to its modulation and demodulation it cancels that, so basically if the interferer is
closer to it, it overrides it in a better fashion. As the interferer goes away from it (of course
within the band of FM, if it goes beyond that you will your filter will cancel it, but within the
band of it) if it goes little further, probably a little bit more interference will be coming to you,
okay.

So, this is what will be happening. So, what we can see already that FM has a nice cancellation
property of interferer. So, any interference that is coming in the band of FM, the closer it is with
the FM carrier, the more it will be cancelled. Not only that, it also gets overridden by the carrier
strength. So, if the carrier strength is enough, the interferer (if that is much smaller than that so
overall worst-case interference) will already be overridden by this FM. So, this is particularly
termed as capture effect (is called as capture effect) of FM, so that means what it does? Actual
FM signal if you transmit, that is your desired signal, so whenever you are demodulating, it will
actually have an overriding factor over all the close by interferers. So, it captures that and
actually exert more power in terms of demodulated output, okay.

So that is a very good property because we want that, we want interference cancellation and FM
automatically due to its inherent in a demodulation technique, it cancels them out. That is a very
nice property. So, what people have seen over here, that in AM, if there is a nearby interferer,
that must be almost 35 db less, then only you get proper signal to interference ratio, okay,
whereas for FM you can go as far up to 6 db less, okay.

So that is a big advantage, so almost 29 db extra interference still you are good enough, almost
both the receivers will behave equivalently. So that is the advantage of FM we were talking
about, that FM we have are now shown two advantages. One was already with the nonlinear
channel, that FM cancels it out very nicely due to its inherent demodulation and modulation
technique and now we could show that if there are interferers even within the band, FM actually

748
overrides them, which is called (termed) as capture effect and it (means) gives huge advantage
over the interferers whenever you are transmitting FM.

So next what we will try to do is, we will try to show another big advantage of FM, that is the
noise cancellation which is almost similar to this part but probably the analysis will be little more
involved, so we will try to capture the overall (interference sorry overall) noise effect in FM,
okay. So, for that we will go back to our noise analysis in similar fashion, so what we did in our
noise analysis, it was a simple thing. We have drawn the receiver transmitter module and then we
(means) for AM analysis also we (have done) always try to capture the figure of Merit. What
does that mean? If I recollect, that means (that) if you transmit the same signal (in same) or insert
or put same amount of power as you are putting in FM in the baseband and try to see what is the
performance.

So, this is the benchmarking one, that is the reference one. Try to see if I do not do FM, I just
insert same amount of power that FM has to put, okay, FM modulated signal has to put, so I put
the same amount of power and try to see what is the performance if I just transmit it in baseband
without doing any modulation and on the other side any demodulation, only that filtering will be
there (low-pass filtering). So, with that we get a reference SNR, okay, and then what we do with
the same power because now the powers are equivalent.

So we actually transmit the FM modulated signal and then in the channel (as expected noise will
be added) after that the entire demodulation of FM will go on, then we will get probably the
signal and noise will go through this demodulation process and finally we will get some
expression for signal and noise, and that will be the actual signal-to-noise ratio. If we get the
power of these two and then after FM demodulation, we try to see what is the SNR and then we
compare these two SNRs compared to the reference SNR how better or how worse this is and
that gives us a ratio which is called the figure of Merit.

So, we have already calculated (evaluated) figure of Merit for AM modulation DS, BSC, SB SC
so we have done all those things and we could see that they are d SB SC and s SB SC they are
having figure of Merit of 1. That means it has no advantage with respect to equivalent baseband
transmission, whereas AM was even worse. It means with a tone modulation we have proven that
9

749
the highest it can get is 1/3 okay, it will be always less than that. So now we will try to see with
FM what happens? okay. So, let us first try to draw the receiver chain, that is the first task
probably so I have FM.

(Refer Slide Time: 17:41)

10

750
 t t

[ ∫−∞ ] ∫−∞
s(t) = Ac cos ωct + kf′ m(α)dα = Ac cos((2π fct + 2πkf m(α)dα)

n(t) = nI (t)cos (2π fct) − nQ(t)cos (2π fct)

= r (t)cos(2π fct + ψ (t))
r (t) = nI2(t) + nQ2 (t)

[ nQ(t) ]
nI (t)
ψ (t) = tan−1

That is the summation, this is the channel so basically FM modulated signal I will be
putting over here which is my signal which is just FM modulated signal that means I can
write it as Ac cos ωct + Kf if it is represented in ‘ω’ domain and 2πKf if it is represented in
‘f’ domain. Okay, so accordingly Kf -∞ to ‘t’ m(α)dα that is it. Okay, so this has an
equivalent representation if I do it for frequency domain, so I can write it Ac cos 2πfct +
in instead of ‘Kf’ let us call this as K_f. so here we will be writing 2πKf.

So, this is in frequency domain ‘f’ domain we should call; and this is an angular frequency
domain ‘ω’ domain, okay so that is my FM. So, this is what will be inserted over here; and from
here in the channel there will be noise which will be added over here, right; and after that, after
the channel the FM demodulation starts, so the first part of demodulation should always be a
band pass filter, so this is the band pass filter of FM. Now this band pass filter is no way related
to just the frequency of the modulating signal. It is the FM band, okay.

So, we call that as from -BT/2 to +BT/2, so BT is the overall FM band suppose which comes from
the Carson's formula. Okay, so we take that, it has a part of the ‘Δf’ it has a part of ‘B’, so it is
actually 2Δf + B, right. So, that is what we have understood, so this is that particular part. Okay,
so that band pass filter should be put so that the entire FM band comes into my receiver. After
that so we call this signal as x(t), okay. This should have noise part as well as the FM modulated
signal. After that what we’ll be putting? We’ll be putting a FM demodulation.
11

751
So, the simplest one we know is a differentiator followed by an envelope detector, right. That is
one of the FM demodulators we have seen. It is just you differentiate it; the entire frequency
variation comes into amplitude as well and then you do envelope detection. We have already
proven in the last class that with envelope detection, we’ll be always getting our signal back,
right.

So that (is the whole purpose of this) is probably the main part of FM demodulation. Okay, so
after that let us say we get a v(t). After this in FM demodulation, this will be clearly understood
from the noise analysis that we need to also employ a low-pass filtering. We will see that later on
why that is required.

So, this low-pass filter is just of the message signal band, so that goes from -W to +W and after
this we get our output, okay. This low-pass filter will just facilitate us in terms of noise
cancellation and you will see that FM only gives this facility. All other modulation scheme does
not give this facility and that along with some other things which we will be exploring gives the
FM edge over other modulation scheme.

So, we will come to that, okay. So, this is the overall receiver architecture, right. So, let us see
how the demodulation happens. So, I have a signal which is s(t) of course but I also have a noise
which is band pass noise, so I can always write it in terms of in-phase and quadrature
component, so I can write it as nI(t) cos (2πfct). Here (I can means) the representation does not
depend on whether I write + or –; so I can take any one of them, so for our advantage which will
be clear in the next phase of derivation, we take – so it is equivalent. It is just a representation.
Whether you take + or – (it) spectrum wise there is no difference, okay.

And representation wise also there is no difference, okay, so this can be written because the
central frequency of my band pass filter is this ‘fc’ therefore this amount of noise which is from
-BT to +BT over that band or -BT/2 to +BT/2 that means BT that is the FM bandwidth, over that
band the overall noise will be coming through the receiver and that is the band pass
representation of that particular signal. Okay, this can further be written as r(t) as any summation

12

752
of addition or subtraction of two sinusoidal can be represented as another sinusoidal so this can
be written as r(t) cos [2πfct + ψ(t)], right.

So, this is something I can write, whereof course r(t) you know already it should be root over
nI(t)2 + nQ(t)2. That is the r(t) and ψ(t) must be tan-1 this divided by this, so it should be ‘nQ’. So
as long as these two nI and nQ (we have probably not talked about these part and due to time
constraint probably we would not be able to prove that, but this is a very important derivation in
random process that if these two) we have already proven that they are independent, okay.

So that is something we have proven, in fact we have proven they are orthogonal also, so as long
as that is happening and they are Gaussian, okay so that is something if we assume generally that
is what happens, then we can always prove that this r(t) actually follows a Rayleigh distribution.
So, this is something which (will be) can be proven. I am not going into the details of that
because that means take us to somewhere else so due to time constraint we are just taking that.
So, it is a Rayleigh distributed one and the corresponding ψ(t) also can be proven that that is
uniformly distributed between 0 to 2π, okay.

So this is something which will be happening (you know random process), so these are also a
separate random process which follows like this nI and nQ or overall NT that follows a Gaussian
distribution. It can be proven that these are also random process which follows this follows
Rayleigh distribution, this follows a uniform distribution between 0 to 2π, okay, so that
something can be proven, so we are just leaving it over here but probably you'll see that's not that
much required for our derivation, okay.

(Refer Slide Time: 25:18)

13

753
 t

∫−∞
s(t) = Ac cos[2π fct + 2πkf m(α)dα]

x(t) = s(t) + n(t)

t

[ ∫−∞ ]
= Ac cos 2π fct + 2πkf m(α)dα + r (t)cos(2π fct + ψ (t))

So anyway, we have got this. So, immediately what we can write is our s(t) that is something we
know already it is Ac cos (2πfct) that's the transmitted one +2πKf in terms of frequency if we
represent - ∞ to t m(α) dα so that is the FM modulated signal. So overall signal that is x(t) after
the band pass filter will be this s(t)+n(t) right, so s(t) happens to be Ac cos [2πfct+2πKf
integration - ∞ to t m(α) dα] so this is the part; and the noise also has its equivalent
representation which I can write as r(t) cos [2πfct + ψ(t)], okay, that is how we have represented

14

754
at the noise that is the after the band pass filter. So then this is a band pass noise and this is the
actual FM.

Because FM the band pass filter just exactly allows the entire FM signal, so FM signal remains
undisturbed, okay, so this is what we get after the band pass filter. Now what we will have to do
is next we have to see if this is the signal plus noise, what my discriminator will do or probably I
have not given that term so this entire part that differentiation followed by envelope detection
that is called termed as discriminator in FM demodulation, so that is called overall a
discriminator circuit. So, what this entire circuit will do, that is something we have to see
because I earlier whatever we have seen that was easier because we are directly differentiating
this term but now, we have the noise which is a random process again.

So therefore, that direct differentiation will not work probably. we have to do it means borrowing
the theory of a random process. Ok, so that is something we'll have to do and equivalently we
will have to calculate the overall power spectral density of the noise that we get after the
discriminator, that will be the next task that we’ll have to do to analyze the FM in presence of
noise, okay, so in the next class probably we will discuss in details how we can analyze FM in
presence of noise. Okay, thank you.

15

755
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 54: FM Noise Analysis (Contd.)

Okay so let us come back to the same equation that we have derived in the previous class, that is
the x(t) which is in the FM demodulator circuit.

(Refer Slide Time: 00:29)

756
 t t

[ ∫−∞ ] ∫−∞
s(t) = Ac cos ωct + kf′ m(α)dα = Ac cos[2π fct + 2πkf m(α)dα]

n(t) = nI (t)cos (2π fct) − nQ(t)cos (2π fct)

= r (t)cos(2π fct + ψ (t))
r (t) = nI2(t) + nQ2 (t)

[ nQ(t) ]
nI (t)
ψ (t) = tan−1

That is where we are now, so now we want to see that effect of this thing okay, so after the band
pass filter we have already characterized the signal plus noise, so this is the noise part. To
characterize this let us try to draw a phasor diagram which is famous in FM, so what we will do,

757
we’ll try to draw the phasor diagram. Phasor means you know already we’ll be actually putting
the phase in angle and the amplitude will be put over along the radius of that, so it will be a two-
dimensional diagram. Now in any phasor the reference is very important, so here what we will
do, we’ll take the reference as this one signal reference, so that must be the reference where we
should start, okay.

So this is something we call as (let us see that is called as) ϕ(t), okay so we are calling this
particular part as ϕ(t) okay, so we can even write this as Ac cos [2πfct+ ϕ(t)], we’re just writing
that as a phase ϕ(t), then r(t) cos [2πfct + ψ(t)], okay this is the reference. So therefore I can put
Ac over here, because in phasor it’s whatever the angle, now this angle is as taken as reference,
so that must be the 0 angle, with respect to that we will have to put this, okay what is the angle of
that one okay, and these are in phasor this will to two vectors right, which has this angle and this
angle, so this is the reference angle. Therefore, what is the angle of this one of this phasor? That
must be this minus this right, that is the angle so that must be 2πfct gets cancelled.

So ψ(t)-ϕ(t) that becomes the angle of this one, okay with respect to this reference. So, remember
that we have taken the reference with respect to this, so that becomes the 0. So therefore, this
angle this entire angle because the phasor is just something cos something, so therefore that
amplitude gets into the radius of that particular part and the angle will be corresponding angle
will be this one. So, it is just the polar plot where the amplitude is the ‘r’ of the polar coordinate
and θ is the θ it has okay.

(Refer Slide Time: 03:19)

758
So, with this reference we have this amplitude as Ac and we get this one drawn, so this will be
having amplitude r(t) whatever that is and this will have a phase of ψ(t)-ϕ(t), now that should
look like something like this, okay because we are adding these two, so in vectorial we can write
from here same phasor, so this angle should be this angle, so therefore this must be ψ(t)-ϕ(t)
right, so overall this x(t) must be the resultant one (resultant phasor), so that must be by x(t)
okay, so this is just a phasor representation of that whole thing, okay. So, immediately that must
have some angle, okay.

Let us call that as θ(t), of course it is reference to this one, so this must be θ(t)-ϕ(t) right, because
the reference is already ϕ(t), so therefore if this has some angle θ(t), therefore the actual angle
between these two should be θ(t)-ϕ(t), right. So, now I can write this θ(t)-ϕ(t). How do I
represent that? So basically, I can see from the phasor diagram that tan of this must be this

759
divided by this. If I put a perpendicular from this point to this point, this entire thing divided by
this entire thing, so just this angle should be tan-1, or maybe you can write it fresh.

(Refer Slide Time: 05:08)

[ Ac + r (t)sin[ψ (t) − ϕ(t)] ]

r (t)sin[ψ (t) − ϕ(t)]
θ(t) − ϕ(t) = tan−1

r (t)sin[ψ (t) − ϕ(t)

θ(t) ≃ ϕ(t) +
A
1 d
v(t) = [θ(t)]
2π dt

2π dt [ ]
1 dϕ(t) 1 d r (t)sin[ψ (t) − ϕ(t)
= +
2π dt A

760
So this angle should be θ(t)-ϕ(t)=tan-1 of let us see the diagram, that should be this divided by
this, this should be r(t) sin of that, sin of this angle and what is this, this should be Ac+r(t) cos of
this angle this internal angle, so I can write it as r(t)sin[ψ(t)-ϕ(t)]/Ac+r(t)cos [ψ(t)-ϕ(t)], okay, so
that’s all fine. Now whenever we are doing analysis of noise, we expect that noise has to be (that
is) lesser power when compared to my actual signal or that Ac, so basically my r(t) should be
always much lesser than Ac, that is approximation we can always take, so immediately I can
neglect this part, so I can write this θ(t)=ϕ(t) + tan-1 of this divided by Ac, again it is r(t)/Ac.

So I can again neglect because that angle will be smaller because sin only can get up to +1, so
the maximum value will be r(t)/Ac, so which is again r(t) is smaller than Ac, so tan-1 of θ must be
equal to θ so I can write this as r(t) approximately r(t)sin[ψ(t)-ϕ(t)]/Ac, okay. So, this is the
approximation which is valid as long as this condition is true, okay. So, I get my overall θ(t)
which is actually the phase of the resultant signal that means that x(t), if you see over here that is
the phase of x(t). What discriminator will do or what the FM demodulator will do? It will
actually take that phase, differentiate it, whatever that is that must be my signal, okay.

So, discriminator output should give the differentiation of this phase, right. So basically my
output signal let us call that as let us say v(t), because if we take this diagram so that is v(t), so
the output should be (v(t) should be) according to my understanding of discriminator circuit, it
should be differentiation of this phase, but remember the differentiation of the phase gives you
the ‘ω’ variance, so if I need to have frequency variance which we are doing for this calculation
so there must be 1/2π. So, (1/2π) d/dt[θ(t)], that must be what we expect after the FM
demodulation. So, let us try to see that, so it should be (1/2π) dϕ(t)/dt + there should be d/dt of
this one. Let us first try to evaluate this part. What was my ϕ(t)? So ϕ(t)=2π Kf integration - ∞ to
t m(α) dα. So, what is dϕ(t)/dt?

Integration and differentiation get cancelled out, so it should be 2π Kf, and if I differentiate it,
that should be m(t), right. So, I must expect this. So therefore, v(t) if I write (1/2π) dϕ(t)/dt, that
is this one, so that should be 2π Kf m(t). 2π gets cancelled + I have (1/2π) d/dt [r(t)/Ac] whatever
we have derived. This is exactly my message signal, that is very good, so this must be accounted
as noise overall but the problem is if I just try to put that as suppose nd(t), so what do I get as
6

761
noise nd(t), that must be (1/2π) d/dt [r(t)sin[ψ(t)-ϕ(t)]/Ac]. What I can see is, there is ψ(t) and ϕ(t).
ψ(t) is related to the noise, it is completely related to the noise no problem no issue with that, but
ϕ(t) is still related to a message signal, so in that noise there is a part of message signal already.

So, that’s where we have to be little bit worried that this noise some message signal part. Now
again we will talk about something which we’ll not prove because of time constraint, so (that
was) that derivation came from a famous paper written by Rice at 1963. If you are interested go
to that paper, it is very famous paper for FM noise analysis, so what he could prove at this
particular noise is independent of the message signal, okay. So, whatever the message signal is, it
is independent of that but not only that he could also prove that, see this ψ(t) we have already we
have shown that it is actually without proof we have said that it is actually (means it is) a random
process which is having a uniform distribution between 0 to 2π, right.

So, what he could prove that this also which is independent of message signal even though ϕ(t) is
there it will be independent, it will become independent and this is also uniformly distributed
between 0 to 2π. So, basically what I can do, instead of taking that part we can just take this as
sin ψ(t), so this was the approximation we could take and he could prove also that this
approximation is valid, so therefore I can write this without loss of any generality, Ac is a
constant, that goes out, so d/dt of r(t) sin ψ(t). Can you identify this? What is this? This is
actually the quadrature term. Go back to that filter representation of band pass signal, you will
see that r(t) sin ψ(t) is actually the quadrature term of the noise, okay so that is nQ(t).

(Refer Slide Time: 13:18)

762
 1 d
nd (t) = [nQ(t)]
2π Ac dt
d
⇔ j2π f
dt
1
SNd ( f ) = (2π f )2 SNQ( f )
(2π Ac)
2

( Ac ) SNQ( f ) | f | ⩽ 2
f 2 B
1
=
0 other wise.

So therefore, immediately we can say the FM noise which is nd(t) is nothing but 1/(2πAc) and
then it is d/dt nq(t), okay, so this is something we can already say. So, it is 1/(2πAc)
differentiation of nq(t), so now nq(t) we already know, right. So, nq(t) is something like this. If No/
8

763
2 is the noise strength, so it should be No because it has to be shifted left at right, they will
superimpose, we have already done that. So, it will be No over the desire band, so that band
becomes -BT/2 to BT/2, okay fine. So, this is the overall nq(t). Now let us try to see what this
means nd(t), so nd(t) is nothing but this nq(t) passed through an ideal differentiator because that is
the operator which is being operated over nq(t).

So it is nothing but a circuit which is an ideal differentiator and pass your nq(t) whatever you get
at the output that is nd(t) with this factor, that factor is alright, we don’t have to bother about that
factor, so that is what is my nd(t) right, so let us now try to see ideal differentiator. What is the
Fourier transform of that? That is actually j2πf, so this is the Fourier transform of that. We have
already proven that ideal differentiator also are linear time invariant circuit, that we already, okay
so if we have a random process input to a (this is something that we have proven; input to a)
linear time invariant circuit, then the output random process will be |hf2| where hf is the transfer
function of that particular linear time invariant circuit or whatever you put into the input sine or
power spectral density.

So basically, we can write power spectral density of this SNd(f) must be equal to the |hf2| which is
(2πf)2 SNq(f) and also you have this part, so therefore it should be 1/(2πAc)2. So, this must be my
SNd(f), that characterizes my noise after the discriminator circuit, okay. So, what do I get from
there? So, I get this SNq(f) is already there, here (2π)2 gets cancelled so I get (f/Ac)2 SNq(f), right,
so this is what I will be getting and this is defined. Remember the filtering was already employed
so this must be defined from -BT/2 to BT/2, so this is true if my |f|<BT/2, and this must be 0
because otherwise there is no noise, because it has already been cancelled due to the ideal nature
of band pass filter.

So that must be 0 otherwise, so can I now plot this? Okay, so if I plot this, SNq(f) is something
like this which is constant between -BT/2 to BT/2, but SNd(f) that has f2, so at frequency 0 it must
be 0 and then it should have a parabolic thing, so it must be something like this, so this is the FM
after demodulation FM noise characteristics, so it never remains similar as AM, any other AM
counterpart, because in AM counterpart always whenever we were analyzing the noise it was like
this, but this is what is happening. If you think little bit carefully, we have already proven that
9

764
interference effect. If you take all this noise as it small, small at a particular small band, small
amplitude interferers what FM does? It captures it right.

And closer to the FM, it captures it more, that means it suppresses it. Same performance is
happening over here, as you can see closer to the central frequency, it is actually suppressing
those noise and as you go away from the central frequency it gets its some significant part, okay,
so FM due to the capture effect only, here the physical meaning is again that same capture effect
we have talked about but we can see now mathematically that is what it is happening, so FM due
to capture effect actually shows this particular thing (this kind of representation). Let us now try
to see what will be the overall noise, okay.

So now this is my noise and at the output we are also getting the signal part was this part Kf m(t),
so this is the signal and the noise whatever nd(t) we have talked about, that is the power spectral
density of this. Now these two things we’ll be putting. Now already my signal has come into the
base band. Up to the modulation it is just m(t) as the signal is defined from –w to +w. I have
noise which is defined to the band pass filtering and after that whatever demodulation has
happened, my noise is still defined from –BT to +BT. now this is FM bandwidth, FM bandwidth
we now already proven through Carson's formula that it is weight narrow band or white band
whatever it is BT is always bigger than ‘w’, okay or BT/2 which is one sides FM bandwidth that
is always bigger than that ‘w’ which is the message band width.

So I for sure I know this BT/2 is greater than or equal w, so I can now employ because my
message will be somewhere here only within ‘w’, beyond that whatever is there, that is not
required for my signal and I can also see due to FM capture effect, that’s where the noise is
having higher power, so I should employ a low pass filter around that message signal, okay,
going from –w to +w so that is what I will be doing next. I will be passing it through a low pass
filter of bandwidth ‘w’, so it will actually take from -w to +w so this will definitely have lesser
noise, so if I now integrate this power spectral density, whatever I get that must be the noise
power. Let us try to calculate the noise power.

(Refer Slide Time: 21:09)

10

765
 ( Ac ) N0 | f | ⩽ 2
f 2 B1
SNd ( f ) =
0 other wise.
2
w

∫−w ( Ac ) 0
f
Pnoise = N df

N0 2W 3
= 2
Ac 3
Psignal = Kf2 P

So I can now define after doing this low pass filtering this SNd(f) must be same thing (f/
Ac)2 SNq(f), where |f|≤w because it is restricted to that, because I have already done a low
11

766
pass filtering and it will be 0 otherwise. So, what is SNq(f) in that? That is just No, so (I
can) instead of this I can write No right. So, what is the overall power? We know from
power spectral density you have to just integrate it, so here it is defined only from -w to
+w.

So, the overall power of noise should be integrated from –w to +w is (f2/Ac2) No df, that must be
my overall noise power. So, No/Ac2 goes out, this is f2 so it will be f3/3, then if you put ‘w’ so that
become actually 2w3/3. This is the noise power. What is the corresponding signal power? So, I
had Kf m(t), if the overall signal power is ‘P’ so it should be Kf2 P, that must be my signal power.
Therefore, what is the signal to noise ratio for FM?

(Refer Slide Time: 22:55)

12

767
 3Kf2 PAc2
SNRFM =
2N0 w 3
Ac2
SNRR =
2N0 w
3Ac2 kf22N0 w 3Kf2 P
FOM = =
2N0 w 3 Ac2 w2

So, SNRFM=3Kf2PAc2/2Now3. So, this is what we get as SNR of FM. Now you have to do the
figure of merit calculation. For figure of merit calculation what we have to do? In the base band
we have to transmit (whatever we have transmitted for) whatever power we have put for FM
transmission FM. FM was just Ac something is there, okay, so what will be the power of that?

13

768
That will be the Ac2/2, because whatever is inside, that is just phase, that will not contribute to
the power of this signal, so Ac2/2 will be the power that will be transmitted in the base band.

And what will be the noise? (The noise goes from my) It will have a strength of No/2 and it goes
from –w to +w, so what all noise will be Now, if we integrate, okay, so that is the noise power, so
therefore overall SNR for the reference one should be Ac2/2N0w. So, therefore the figure of merit
that must be this divided by this, right, so figure of merit must be 3Ac2Kf2P/2N0w3 divided by
this, so this particular term goes up 2Now/Ac2. So, we are left with 3Kf2P/w2, right. That is the
figure of merit. Probably from here we are not able to appreciate what has happened in FM. That
is still something we still do not know. Let us try to evaluate it in terms of tone modulation then
probably we will able to match it with respect to AM what happens, okay.

(Refer Slide Time: 25:53)

14

769
 1
m(t) = cos(ωmt) P =
2
2πkf sin (ωmt)
[ ]
S(t) = Ac )cos 2π fct +
2π fm
kf sin (ωmt)
[ ]
= Ac )cos 2π fct +
fm
3Kf2 P 3 kf
2
FOM = =
w2 2 f m2
3 2
= β
2

Let us try to do on a tone modulation, so in tone modulation what will happen, by m(t) that is
actually nothing but some cos(ωmt), okay, this is just a tone modulation, this is just a cosinusoidal
I am trying to transmit, so immediately I can calculate what is the power of that, so that means
‘P’ should be 1/2 because it is just a cos, so ‘P’ should be ½, so this is something I know, okay.
What will be my s(t)? s(t) should be Ac cos(2πfct+Kf) and then if this Kf is represented in terms
of let us say ‘ω’, so then that is Kf, otherwise there should be 2π, all those things, okay.

So if that is the case Kf and then I need to have a integration of this one, so if I just to a
integration what will happen? If I do integration that should be sin(ωmt)/ωm. Now if Kf is
represented in frequency that must have 2π, I can here also write 2πfm, 2π gets cancelled, so this
is my representation, okay, so I can write this as Ac cos[2πfct+ (Kf/fm)sin (2πfmt)], okay, so that is
my representation. Now the Figure of Merit which we wanted to calculate, what is that? It is
actually 3Kf2P/w2, okay. What is ‘w’ for me? That is actually fm because the message signal is
having just an impulse of strength ¼ at +fm and -fm, so what is the overall bandwidth? That
should be this one.

So, ‘w’ becomes fm right. P becomes ½, so I can write this as 3/2, then I have Kf2/fm2 or I can
write (3/2) (Kf/fm)2. Can you now identify what this is? This is actually the β. We have already

15

770
proven that for tone modulation that Kf/fm is the β or the we should say modulation index or
frequency deviation index, okay, so this must be the β, so I can write this as 3/2β2. What now I
see, now we get a particular relationship between the FOM (Figure Of Merit) with respect to β.
The good part is, here if I increase β, FOM increases, that is one thing. What does that mean?
That means if I increase β, I am actually going towards a wider band FM, because the frequency
deviation that means the Kf value is getting increased with respect to fm that is the frequency
deviation is getting increased which is actually termed as wider band FM.

So if I make the FM more wider, that means I increase the bandwidth of the FM, I get a benefit
in terms of noise cancellation because its Figure of Merit gets improved, so this is what we
wanted to discuss that FM is the only modulation scheme (none of other modulation scheme will
probably do that) that where I can exchange my frequency with respect to noise performances, so
if I exchange means, I can make this worse and get a better benefit, so basically I can increase
the FM bandwidth where I am guaranteed to have a better performance in terms of noise
conservation. This is what is happening in FM. Now let us try to compare this. What was the best
FOM (Figure of Merit) that we could get for AM?

(Refer Slide Time: 30:35)

16

771
3 2 1
β >
2 3
2
β> = 0.471
3

That was 1/3rd we have proven with a tone modulation, so what we can write that this, if this has
to be anything better than this 1/3rd, my 3/2β2 >1/3 okay, and then immediately I can get a β>√2/3
which comes out to be some 0.471, so any β which is greater than this, will always give me
benefit in terms of noise cancellation with respect to FM, that will be happening. I can take to
any β value where this will be much bigger, okay so basically, I have now a controllability
parameter or a controllable parameter which I can just keep controlling to get better and better
noise performance.

17

772
So, in FM we can actually do we can make the bandwidth wider off course the spectral
efficiency will be reduced but to get better signal quality. That is always possible, where as in
AM you can never do that. If you increase bandwidth more noise will be coming inside and you
will get a worse performance and you cannot decrease the bandwidth because the bandwidth you
take for AM that is the optimal one. Anything you decrease your signal will be distorted, so there
is no option over there, no variability over there, where as in FM you can do that.

And this also the reason why people have preferred wider band FM compared to narrow band
counterpart, because in wider band FM, you can a better noise cancellation and that is why we
have seen that Armstrong tried so hard to generate wider band FM with series of multiplier
circuits from a narrow band counterpart because generation of narrow band was that time easy
when VCO was not there and then they could actually reproduce wider band FM by doing this.
This also gives us another advantage if we just try to see, what was the FM noise performance?
This was the noise spectrum, whereas (the message will be so this is from –w to +w, whereas)
the message will look like this.

So, from –w to +w or it might have nothing around 0 if it is a voice signal. Whatever it is we can
see that more noise is added at the higher frequency part. To combat that what we can do? We
can do some pre processing at the signal level. We can give something where the signal higher
frequency components are boosted compared to the lower frequency component because lower
frequency components are much better off in terms of noise because they are getting added with
lesser amount of noise, whereas the higher frequency components are getting more amount
noise, so if we just boost them little bit, what will happen? We’ll get a better signal to noise
characteristics at the higher frequency.

And then after doing FM demodulation and everything, we can actually whatever boosting we
have done, we can just reverse that; so that means after noise cancellation and everything that
has been employed, we can just redo the whole thing and we will get probably a better analysis
because noise power will be already same whatever will be coming in, and after that if we do a
reverse whatever filter we apply at the beginning, if we just reverse it, we’ll probably get a better
performance in FM.
18

773
So that is what people have employed which is famously termed as pre-emphasis and de-
emphasis, so you do a preprocessing of a frame signal to give it a better noise cancellation. So,
next class probably we will be discussing this pre-emphasis and how much it gives benefit with
the practical circuit, okay. Thank you.

19

774
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 55: FM Noise Analysis (Contd.)

Okay, so what we have done so far is (means) discussion of FM and it is noise analysis, so we
have also proven that in presence of noise, FM performs better. In a way that we have also shown
that FM is probably the only modulation scheme where we can actually exchange the bandwidth
efficiency with the noise performance. What does that mean? That means if we increase the noise
band width or (sorry) if you increase the FM bandwidth (that means make a wider band FM, that
is what we can do) if we do that, immediately we can see there will be a positive effect in terms
of noise cancellation, okay.

So, (more) wider the FM bandwidth, better noise characteristics we’ll be able to see, so this is
something which has been already proven. We have also done that for a tone modulation. We
have shown how better it is in terms of noise cancellation with respect to AM and what
parameter set like the modulation index which is directly proportional to the FM bandwidth or
frequency deviation.

So, how we can set that to make it better than AM, so this is something we have already
demonstrated and at the end of last lecture. We could also show that and even before also we
could show that FM has a noise cancellation or even interference cancellation when the interferer
signal is closer to FM signal, so this is something which has happened and especially that was
proven that FM noise due to the demodulation process looks like this.

775
(Refer Slide Time: 02:05)

w < BT /2

So basically, the noise spectrum or power spectral density looks like this, if this is -BT/2 to +BT/
2, so that is something we have already shown, so what effectively we mean by this that, if this is
the FM carrier and around which the FM modulated signal is there so something closer either
noise or interference very closer to that central frequency of that FM modulated signal, that will
be canceled out, whereas as we go away from the central frequency the noise power spectral
density also gets boosted up as well as interference also we have seen that.

That is something which is to be noted for FM, so this is typical of FM. Now generally what will
happen, whenever we within this there will be the means this is the low-pass equivalent noise
characteristics right, so if I just plot the signal spectra that should look like this, okay, so this is
my M(f) power spectral density, or power spectral density let us say SM(f), okay, with respect to
frequency, okay, so we know that it is band limited probably that is the band so -W to +W and of
course ‘W’ must be less than BT/2.
2

776
So, that is a condition which generally happens, we have already shown that and this BT is
(actually) the actual FM band width whereas ‘W’ is the message signal bandwidth, so FM
bandwidth is always greater than that ‘W’. That is something we know, so what happens in this
side at the higher frequency? Generally, the message signal shows that because it is band limited
so the spectrum component will be having lesser amount of power, okay.

And whenever it is transmitted after the receiver, the discriminator circuit followed by a low-pass
filter, what we will see at this position noise is more enhanced due to FM demodulation and this
is the point where signals are (means) having less power, so (if the signal is having) at this
frequency of course overall, we have already analyzed but at this frequency we just concentrate
on the higher frequency side of the message signal or message band.

We can see the noise power is higher, whereas signal power is lower, so maybe something can be
done over here, okay so what is that something? The something is like this; see noise is getting
added at the channel right, that has nothing to do with the transmitter, so I have a transmitter
followed by a channel through which it propagates. This is where noise gets added and followed
by a receiver. After that I get my signal as well as noise, okay, the composite part of it, so this
particular noise after going through (means) added in the channel and going through the receiver,
that looks like this, okay. Now I can do something to the signal. What I can do is effectively
because I know that noise will be already higher at the higher frequency.

So, I can deliberately boost the signal at the higher frequency, okay. That boosting will not boost
the noise because that boosting will be done at the transmitter, so which is independent of noise;
before even noise gets added, I boost the signal so that that signal is already of high power then
the noise gets added. Due to the demodulation process, noise will be as it is, but the signal has
been already boosted at the transmitter side, so at the receiver side, signal will have that boosted
things.

Now what we have to do, of course the signal if I boost it, the relative amplitude at different
frequency level has been changed, so basically that will distort the signal, so if I have boosted it,
I need to put a de-booster or some filter which is just cancelling out this boosting at the receiver;
so the fun begins when I start putting that; so if I start putting that, so suppose I have a receiver,
after the receiver I put that reverse thing, at that time what will happen? The signal it will come
back to original.

But because this is actually reducing the overall amplitude or overall power of the high
frequency term, so what it will do to the noise, noise is as it is because it has not gone through
3

777
some boosting, so it will be as it is, so here if we are de-boosting it or if we are reducing the
amplitude of the high-frequency part, so noise will go through that reduction in overall power, so
at the output I will see the same signal because it was boosted at the transmitter so before
transmission, I have this boosting and here it is the reverse operation.

So, I will get the same signal but this noise power will be much more reduced; so, this is a
technique which has been employed in FM which helps means combating noise even better,
okay. We’ll probably show with the practical almost first order this boosting and de-boosting
filters which are called as pre-emphasis and de-emphasis, because we emphasize the higher
frequency before transmitting and the emphasis is just doing the reverse so pre-emphasis then de-
emphasis, so we de-emphasize it and bring it back to the original after the transmitter.

(Refer Slide Time: 08:23)

778
Hpe( f )Hde( f ) = 1
2
SN0( f ) = HDe( f ) SNd ( f )

So this pre-emphasis de-emphasis actually gives you, with the practical implementation of pre-
emphasis de-emphasis we will be showing that almost 13 dB gain in terms of signal to noise
ratio, okay, so means 13 dB, okay, so that is what we want to achieve, let us see how that can be
done, okay, so the overall thing is there should be a H-pre-emphasis or pre-emphasis just before
the transmission, so our message signal m(t) must go through this and then followed by whatever
FM transmitter we have. Any transmitter is okay.

And then it goes to the channel where the noise is being added with the modulated signal
(modulated pre-emphasized signal), okay; so, then I go to the receiver which is FM receiver,
okay followed by of course I have to cancel that pre-emphasis so I have to put a H-de-emphasis
filter; and that should be my transmitter receiver chain and what we should also realize that this
HPe(f) Hde(f) this must be ‘1’, because they should just cancel out each other, okay.

So whatever the transfer function of the pre-emphasis de-emphasis must have inverse of that, so
that they just cancel out each other and the message signal passed through this chain of pre-
emphasis and de-emphasis will actually have the same (quality) spectral quality as well as time
domain quality, okay, so this is something which has to be happening, okay, so if this is the part
then let us try to see after the receiver, okay after going through de-emphasis what should be my
signal and noise? Signal should remain as it is as we have discussed earlier but the noise part will
go through this.

So basically we have already told that this SNd(f) is the noise power spectral density at this point,
okay, so that goes through the de-emphasis filter, so it must be passing through the de-emphasis
filter and that is what the noise will be, so if I just at the output if I wish to get SNo(f), that must
be passing through a filter, so which is also pre-emphasizes de-emphasis will be (means) proving
that they are generally chosen to be linear time-invariant, so I can again put that whatever
random process theorem we have proven that any random process with a power spectral density,
if it passes through a particular filter which is linear time-invariant, then the output power
spectral density will be the input power spectral density into |hf2|.

So same thing will be happening so if this is that Hde(f) so that must be |Hde(f)|2 SNd(f), which
was the noise power spectral density over here. This is that part which we have already derived
that looks like this, right, so and that is defined from -W to +W, right, because this is after
5

779
employing the low-pass filter also this is the entire (it should be FM, so entire) FM demodulation
chain okay.

(Refer Slide Time: 12:10)

w
N0 2
∫−w
2
noise powerp/d = f HDe( f ) df
Ac2
w
N0 2
∫−w
noise powerwithout p/d = f df
Ac2
2W 3
I= 2
w
3 ∫−w f 2 HDe( f ) df

780
So that is after the demodulation and (after following means) after putting the low-pass filter of
the message bandwidth, so that should be the noise spectral density, okay, so this is what
happens. Now let us go back and try to see what is the overall noise power, okay and what we
will try to evaluate that what is the relative benefit by doing this de-emphasis, okay, so let us say
I have this particular thing which is defined. This is actually average noise power without pre-
emphasis de-emphasis, so this is without pre-emphasis or de-emphasis filter and the ratio of
average noise power with pre-emphasis de-emphasis filter.

So, I am just trying to see if I (means) put this pre-emphasis de-emphasis what will be the noise
power at the output provided that same kind of message signal and same kind of noise is being
added in the channel? If I do not put it what will be the output noise power? So, I am just trying
to take the ratio of this, this will be the improvement because the signal power remains the same
only the noise power will be reduced by this factor, okay.

So, this is something I am trying to calculate, so what will be the noise power with pre-emphasis
de-emphasis? So, this is with P/D, so that must be we know already this relationship SNd(f) |
Hde(f)|2. I have to integrate it from -W to +W, right. It is valid from -W to +W, so all I have to do
is that only, so if I just do that, we have already derived the noise power spectral density SNd(f),
so let us put that, so that was equal to (No/Ac2) f2. Now along with that there will be a
multiplication term of |Hde(f)|, right and this must be integrated.

So, that is the overall power spectral density. this was the power spectral density for SNd(f) and
multiplied by this square, then I have to integrate it from -W to +W, right, so this should be my
overall this one; and what has happened for noise power without P/D, we have already
calculated, so that should be noise power without P/D. This is something we have already
demonstrated so that is just minus -W to +W and (No/Ac2) f2 df, right, so this is something we
have already evaluated.

So, to get ‘I’, we have to take the ratio of these two, so therefore that ‘I’ should be after doing all
this calculation, what we will have, we will be having so without one on the top and with bottom
so that should be this, so this should be 2W3, so after doing this simplification, ‘3’ and I will be
only left with this integration, f |Hde(f)|2 df, that's my ‘I’, right, so and (No/Ac2) gets cancelled
because this is a ratio of noise power, so (I will be just) I have to evaluate this okay.

So, let us try to see if we have a practical filter, for this pre-emphasis and de-emphasis (how do
we means) what should be the performance, okay, so let us say we employ a pre-emphasis filter
which is almost which looks like a high-pass filter because (I want to really) pre-emphasis means
the higher frequency I should boost and lower frequency should (have means they should) be
7

781
suppressed, so if I put a practical high-pass filter which has a nice means smooth slope
something like that.

(Refer Slide Time: 15:54)

R≪r
2π fcr ≪ 1
jf
Hpe( f ) = 1 +
f0
1
HDe( f ) = jf
1+ f0
1
fc =
2πCr

782
So, it will actually boost this higher part and it will (means) reduce the lower part sufficiently,
right, so a high-pass filter, if we just give one example, so this is C r and followed by of course
we have to amplify it also, so ‘R’ and that is the amplifier, that is the input impedance of the
amplifier, okay, so this is that high-pass filter part, okay, so if I just put my message signal over
here, whatever at the output I will be getting, that is actually the pre-emphasized message signal,
okay.

What is the corresponding transfer function? So, we have to also consider that this ‘R’ is much,
much lesser than this ‘r’, okay, so that is something we will have to ensure and also the cut off
frequency this 2πfCr that must be much, much less than ‘1’, so it must be much lesser means this
particular overall thing, okay, that ‘C’ and ‘r’, so this is these two things we have to ensure just to
make the high-pass filter so that at the low lower band also it is passing something, otherwise the
message signal will be completely suppressed.

So, I do not want that, so I just want to emphasize the higher frequency part, so it will definitely
what I will try to do, the high-pass cutoff frequency I will bring it very close to ‘0’, means it
should be closer to that, so that the higher frequency is getting a little bit amplified (or means)
and the lower frequency is getting a little bit suppressed, so this is something which is
happening, not completely getting suppressed, okay, so I do not want a high-pass filter which is
like this, okay.

So, then the lower frequency which is the actual message signal will completely be gone, so I do
not want that, so this that is why this restriction I will be putting and accordingly the Hpe(f)
should look like 1+j(f/fo), where fo is 1/(2πCr), so this is just that filter characteristics we are
putting, so if Hpe(f) is this, immediately Hde(f) becomes 1 by that, so 1/(1+j(f/fo)), so of course
that should have a reverse characteristics and fortunately that can be realized, so if we just put r
over here and C over here.

So, it is the reverse thing low-pass equivalent part, okay, and if I just put that, this will have a
transfer function, if you try and look at your filter designing skills, you will be knowing that that
should be the transfer function of this particular filter, so we have already designed Hpe and Hde
and that condition that this multiplication must be ‘1’ that is being satisfied, ok, the way we have
taken the transfer function and it is as you can see from the filter these are all first order filters,
okay, so be it high pass/ low pass these are all first order filters. If we have this, then can we now
calculate the ‘I’? So that ‘I’ which was we have derived the formula is this, in that formula I
think we have missed something.
9

783
(Refer Slide Time: 19:17)

2W 3
I= 2
w
3 ∫−w f HDe( f ) df
2W 3
=
w f 2 df
3 ∫−w f
1 + ( f )2
0

(w/f0)
3

[ 0 ]
3 (ω /f0) − tan−1 ( fw )

So, earlier when we have shown that square was missing, so it should be f2|Hde(f)|2 df, right, so
that must be the overall thing and now just put that filter Hde(f) you already know, so if I just put
that filter transfer function, so I get -W to +W, f2 df divided by that filter characteristics |Hde(f)|2

10

784
so that is (f/fo)2, okay, now it is just doing this integration which is a simple integration, so (this
can be means) you can separate them out and you will get tan-1 integration, so if you just do that
it will become, finally I am just giving the final form, so that should be (W/fo) – tan-1(W/fo).

So this fo has to be manipulated, it should be written as 1+ so here ‘f’ is there, take ‘fo’ out, so
this will become (fo – f) and then this +1 –1, so this +1, that gets cancelled, so this will be just df
integration so that will be f which gives you this –1/(1+(f/fo)), so f/fo you substitute with some
variable and then do that, it will be just tan-1 this, okay, so (that is the) that we know already, that
is a simple method of doing it, okay so this is what is happening, this is becoming ‘I’ and if we
just take ‘fo’ to be 2.1 kHz and if we take ‘W’ to be 15 kHz for a typical FM transmission.

So, you must be seeing that this is not just voice, this is actually a whole band of even musical
instruments also comes under this, so 15 kHz means voice was just 3.4 kHz, so this is actually a
wider band message signal, okay, which includes voice as well as some musical instruments
which has higher frequency component, so if that is the case and we choose ‘fo’ to be 2.1 kHz, so
it depends on what ‘fo’ you choose because we can see already this ‘I’ is function of ‘fo’, so if we
choose this thing, then immediately ‘I’ becomes 22, and in dB terms that is 13 dB, that is what
we have told that in a practical FM case we can actually reduce the noise.

So, this becomes 22 times higher, that means without this thing pre-emphasis de-emphasis
whatever noise we get that is 22 times higher than things with pre-emphasis emphasis, so
basically 13 dB higher noise you will be getting at the receiver if you do not employ pre-
emphasis de-emphasis, so that is a big benefit so a SNR will be accordingly you can go almost
13 dB means you will have a 13 dB margin in the SNR.

So, this is the advantage that you get using FM because FM can also give you this facility, so it
already had we have already seen that it has a nice noise cancellation, not only that you can
employ more things to cancel further noise, okay, so and that is the reason why earlier days
people used to transmit just voice. At that time probably amplitude modulation was still okay, but
when people started transmitting all kinds of musical instruments, music and everything, then the
(means) overall message bandwidth was not restricted to just 3.4 kHz or 4 kHz.

So, it (means) went beyond that and it went up to 15 kHz or 20 kHz. At that time if you employ
AM modulation, huge amount of noise was coming in and that was really killing the quality of
the signal, so FM came up very handy because in FM, we could see more the bandwidth, better
the noise cancellation, so more clarity we’ll get. That is what happens, okay.

11

785
So, I think we have now almost proven the strength of FM in a way, so what next, we'll try to see
(that means) from this perspective it looks like (FM is all means) implementation of FM is all
rosy. It is just one of the best modulation techniques that can happen and all those things, ok but
means none of the story in this world are like that. It is not always (means) everybody or
particular part is always winner, so there are some hind side of FM. We will try to also means
just if not quantitatively at least qualitatively clarify them in this particular part of the course.

So, we will now try to see another interesting part of FM (which is called FM means), if FM gets
captured by the noise, it is very bad, so that is something we'll try to capture over here which is
called the threshold effect of FM or it is also termed as click noise in FM, ok, so when that
happens, if you remember when we were doing FM analysis in one place we were drawing that
phasor diagram and we have also (means) mentioned that probably the FM noise signal, the
amplitude of the noise or overall power of the noise is much lesser than compared to the signal
part.

(Refer Slide Time: 25:56)

12

786
 t

∫−∞
ϕ(t) = kf m(α)dα

r (t)cos[ωct − ψ (t)]
ψ (t) − ϕ(t)

So, this is when FM does very nicely or performs do perform very nicely, but when noise
becomes comparable to the signal, that is when all the problem starts coming, ok in FM, so let us
try to appreciate that, so what happens, if you remember that phasor diagram we have drawn, so
this was ‘Ac’ which is the signal and that is in the phasor what we have said that is (means) the
phasor phase is the reference, so because FM has Ac cos (ωct + Kf into integration m(α) dα), so
that entire thing is the difference. On top of that we were adding noise, so noise as we have said
that it can be represented as r(t) cos (ωct + ψ(t)), right.

That is something we have said, so this is that means after doing that quadrature and phase
representation of band pass noise then we could represent it in this fashion, okay, so and we
could see that that will be somewhere over here where this angle is this ψ(t) – φ(t) whatever FM
phase is, where that FM modulation term comes so that φ(t) is actually Kf integration −∞ to t,
m(α)dα right, we have discussed that, so this becomes that ψ(t) – φ(t) right and on top of that,
this is actually the r(t) and that is rI(t) and that is actually rQ(t), right. This is the overall noise
plus signal, okay (Phasor), and this is something we have characterized, that is the x(t) right, now
let us see what will happen to our FM, okay.

So, suppose around this I start means what will happen? This noise is uncorrelated and it is
independent it is probably one of the most random things that is happening, so this noise
amplitude as well as the angle will keep on varying. This will almost remain the same, whereas
on top of that this will be varying, so this tip wherever it is this tip will be wandering around that
tip, it can be anything depending on how the noise amplitude is, so this may wander within this
circle probably, okay.

And the amplitude also might slightly become bigger or smaller, but as long as this amplitude r(t)
is much, much lesser than ‘Ac’ whatever wherever they wander, they will be remaining very
close to this particular point only that x(t) only, okay so they might wander around but they will
remain over there, so basically the phase as you can see overall phase will not be deviating too
much, okay because they are they are small in amplitude, they are varying on that location
wandering around that location.

13

787
So overall phase we will still be joining some point over here random point over here, so the
phase does not get change hugely, so basically if I just track the phase, it will be wandering
around that value, so this is probably that θ(t), but what will happen if this amplitude becomes
bigger? Then the circle that it actually wanders around is becoming much bigger and then what
might happen (we'll see probably in the next class that what might happen) that might really
create a huge amount of effect in the phase variation.

And the phase variation is actually this that is the θ(t) which you are trying to track where the
FM is modulated. Differentiation of that as is actually the demodulated signal, so if this varies a
lot, there will be some impulses which will be generated, so these impulse noises are actually
termed as click noise. We will see that, so if noise amplitude is very high, then probably
whenever you are demodulating FM, you will be seeing those clicking sound in-between
randomly happening, so that is actually termed as FM noise, okay (FM click noise), so we will
discuss about that in the next class more on these things and then we will also talk about
threshold effect, okay, thank you.

14

788
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-56: Sampling Theorem

Okay so we were in the last class we were actually discussing about the click noise okay, so that
something one topic we have started. So will not go into the details of it, but let us give you some
basic idea of what that click noise is. so basically we have drawn that pharos diagram right.

(Refer Slide Time: 00:39)

789
So this is that AC that is that point on top of that, the noise is being added. Now the problem is as
long as this noise amplitude is smaller than A, there is no problem. but if the noise starts getting,
suppose this is my A, and this is the noise amplitude, okay. Now what might happen if this is big
enough this might even go in this direction. so what is happening to my phase? it was on this
direction okay so the phase this is the reference phase which is over here. this immediately goes
into reverse direction.

It might even go into this direction. okay so the phase becomes so basically there might become a
random variation of phase, which was not happening over here. Here if it wanders,I know that it
will still be remaining on that same, around theta okay, so it is not getting a huge variation due to
noise, okay. So if you just plot that theta, so probably I wish to have a theta variation like this on
top of that there will be small noise induced theta variation. okay, and I will differentiate it, they
will smoothen out even more, and I will get my actual message signal variation.

Whereas, what will happen over here.once there is a variation, and if I means if this goes from
one particular, suppose it is in this direction, now noise is random, so it can immediately in the
next' this one, it can go into this direction, so there will be almost a PI phase shift. okay, so there
is a possibility now because the noise amplitude is higher, so there is a possibility that randomly

790
after sometimes suddenly a pi phase shift happens. So what will that means this particular thing
will have a certain PI jump. okay.

And then again it will probably track, again another PI jump might be there, or it might have
again come back, so all these slopes if you differentiate it what will happen they will start
creating, suppose after differentiating I was getting this. but this particular point because there is
a sudden thing, there will be a impulse noise. which will be created as many pi shifts are
happening. okay, either in the positive or negative whatever happens, there will be positive
impulse or negative impulse. It depends on what kind of things are there but whatever it is there
will be a impulse noise which will be coming and they will be coming randomly because noise
characteristics you do not know, when it will do that pi shift okay.

Here even if it does PI shift, my noise due to noise this variation will not be huge. it will just be a
small amount of modulation. but because in the phase there might be a PI phase shift, so there
might be a click noise which will be coming. And this clicks are random in nature, when they will
be coming that depends on noise characteristics.

okay, So what is the threshold effect, so this is called click noise.so basically what you will see
that after FM demodulation.it will have that original message which because it has a nice noise
cancellation, so overall noise cancellation will be good. So the message signal will be smooth,
but suddenly there will be impulse, and then again smooth suddenly there will be impulse. so this
will be creating a cracking noise in the receiver. So whenever you are hearing there will be
cracking noise one after another, this might happen whoever means have listened to FM radio
you might have heard those clicking sounds. okay, very rarely happening.

But what will happen if the noise amplitude is too high, and it remains too high, and the noise
nature actually varies this means this arrival of Pi phase shift, it happens too often. so within a
second, if there are too many clicking sounds so this particular part whenever this within a second
how many clicking sounds are being heard, that increases then, that is called the threshold effect
in FM. So at that time, probably the message is still being demodulated very nicely, but there are
too many impulses on that.

So basically our means ear, whenever we hear it probably, that clicking sounds, they are too
often, so that will be only reverberating inside our ear. so we will not be able to hear whatever is
actually being transmitted. the message signal will have no consequence it will be just a
collection of clicking sounds one after another coming. So that is when we say that FM goes into
threshold. so there are analysis that with what SNR value, FM goes to, means threshold zone
what is the number of clicks in one second that will be coming'.

791
So all those things are there probably will not bother our self due to time constraint in
characterizing those things. but it is good to just know that FM has that kind of effect. okay, so it
is not all rosy, as we have thought. whenever the signal-to-noise ratio becomes smaller, so it
either goes very close to unity or even less than unity. then we have a big means) challenge in
receiving FM signal-

okay. So I think with that we have almost finished our discussion of FM. so whatever we wanted
to discuss about FM we have finished that. Now we are in a position to actually compare FM and
AM with their full scale. so what is the bandwidth efficiency, what is the corresponding noise
characteristics or noise cancellation characteristics, how it behaves in terms of nonlinear channel
characteristics, what kind of interference cancellation technique they, they can employ. so all
those things already has been evaluated.

okay, So now we will probably start with another form of analog modulation it is still being
called analog modulation, because that was probably the transition phase between analog and
digital transmission. So we will start.we will try to cover also that, that particular analog form of
means modulation technique which is called pulse modulation technique. we have already
mentioned it in the beginning of our course.

(Refer Slide Time: 07:23)

792
So today we will try to, means it is called pulse modulation. okay, so this is something we will try
to cover. but to understand pulse modulation will have to probably start appreciating some of the
fundamentals of this pulse modulation. or some of the signal aspect, or signal theoretic aspect of
this pulse modulation. So that is something which we will be doing first means dealing with
them, and then we will come to the fundamentals of pulse modulation.

okay. So pulse modulation has three types. like AM, FM, they had or AM and angle modulation,
amplitude and angle modulation.they had their own types, like for angle modulation, we had PM
and FM. like amplitude modulation we had/means amplitude modulation without means
suppressing the carrier, then DSB-SC, SSB-SC then vestigial sideband, so here also there are
three techniques. so one is called pulse amplitude modulation,PAM.the other one is called pulse
width modulation, and the other one is called pulse position modulation okay.

So it is almost analogous to, basically frequency phase and amplitude. okay, so if you see PAM,
that is actually like this. whenever we say pulse amplitude modulation, it is in terms of some
discrete pulses, so the entire modulation happens with pulses. pulse means it is on for some
duration, and rest of the duration it is off, probably. okay. So that is generally termed as pulse

793
modulation, so we will be modulating generally with in terms of this kind of pulse train, which is
just this. these are periodic signal after every, let us say TS amount of time it repeats.

And the pulse is on for some amount of time. let us say T P. and T s - T P time it will be off. so
the modulation happens with these pulses like we had carrier we have pulse train over here to do
the modulation, so it almost analogous to the carrier. okay, and when we say pulse amplitude
modulation, that means if I have a signal, the signal will be covered in the envelope of this pulse.
so that means the pulse amplitude will be modulated according to the signal. So if I have a pulse
over here, that will a higher amplitude next if I have a pulse over here, that will have a smaller
amplitude. so it will be just like this if you take the pulse tip, and connect them they will create
the envelope of the signal. so again there will be a pulse that will be negative and so on. so if
there is a pulse that will be positive, so if you just connect the tip, that will just look like the
original signal. okay. So that is pulse amplitude modulation.

whereas pulse width modulation is you, now start varying the width of the pulse. okay, according
to the amplitude of the signal message signal. so basically what will happen if the amplitude is
higher it depends on how you wish to modulate it so you can you can take a decision, that if the
amplitude is higher, I will have lower width pulse. and if the amplitude is lower I will have higher
width pulse, so what will do probably here. if you do pulse width modulation here the pulse width
will be smaller. which means now you are actually modulating that Tp okay.

So that is something we are doing. over here the amplitude is lower, so probably the pulse width
will be little higher, and over here amplitude is even lower so pulse width must be even higher.
then again it is becoming smaller as it goes up probably the pulse width becomes, means of
course, we should have same amplitude. Now it is almost like now frequency modulation okay,
so the amplitude of the pulse remains similar it is just the width you are modulating. the width of
the pulse, okay.and the other thing is the pulse position modulation which is almost like a phase
modulation.

So basically you position your pulse depending on the amplitude of this particular thing. so
basically whoever will be having higher amplitude probably they will be closer to the this pulse
starting time. okay so what will happen if we just draw that at this point, because the amplitude is
higher, so pulse will be closer to that pulse starting time. at this point because this is lower, so this
will be further away.the pulse now the width will be same amplitude will be same.it is just the
position of the pulse which is being varied here, it is more negative so it will go much more away
from the, this one/again it will come closer or something like this.

okay, so this is called pulse position modulation, this is called pulse width modulation, and this is
called pulse amplitude modulation. okay, so this is something which we will be trying to employ.

794
okay, so let us now try to see, as we have said that before we do this, we need to have a
mathematical foundation of this pulse modulation. So that we can analyze their characteristics
and everything else. so that signal processing part of that or probably the signal analysis part of
that/has to be dealt first, before we even jump into these modulation techniques.

(Refer Slide Time: 13:43)

∞
δ (t − λTs)
∑
δTs(t) =
n=−∞
g(t) = g(t)δTs(t)
∞
Dne jnωs t
∑
δTs(t) =
n=−∞
2π
ωs = = 2π fs
Ts
1 Ts /2
∫
Dn = δ(t)e −jnωs t dt
Ts −Ts /2

795
So if you just see we have probably discussed this what is actually a pulse modulation. So in
pulse modulation what you are trying to do, suppose I have a message signal. So whenever we
could have actually transmitted this message signal. okay that is without modulation we are
transmitting it. Now instead of doing that, probably I will have some theoretical foundation, that
not all the time, because this is a continuous time continuous amplitude signal okay, so this is
time, and this is amplitude, it might be voltage it might be whatever it is whichever parameter we
put for this.

So that is taking any value, and also the time every time the signal exists. so if some theory can
tell us, that probably for this signal description I do not need all the time, so I just need some
portion of that time, to define that signal. or maybe some discrete values of time where that has to
be justified, rest of the time I do not have to have a definition of this signal and I can still
reconstruct the signal as it is.

If we can do that then probably will have some modulation format which is similar to pulse
modulation. so in the pulse modulation basically what we are saying is, in discrete time, so time
actually we are making a discrete, in discrete time we are trying to get the signal and send the
signal, send some part of the signal, rest of the time we are keeping it open. where others can
modulate. So it is also a multiplexing scheme like in frequency, FDM we have told, that
infrequency domain we can multiplex things.this is actually where/in time domain we can
multiplex things.

So that is called TDM, so we will discuss more about it later on, but that is what we try to do that
time domain, can we actually discretize it, and only take discrete-time events or time samples to
represent the whole signal. so that is what we are now targeting. Later one when from here, will
be transiting towards more of a digital transmission. what we will do we will not only discretize
the time we will also try to discretize the vertical axis. so then our amplitude also will be having
discrete values, and we will take those values and represent them in a digital format.

So that is actually what is termed as digital transmission. so that is why I told, that this is one step
ahead towards digital communication. so you are just discretizing or trying to discretize the time,
or trying to free up the time. that you do not use it for whole time to represent the signal, so that
is something we will try to, and now theoretically analyze. So this was first done by Nyquest.
most of you must be knowing about Nyquist sampling theorem, so that is something we will try
to demonstrate now.

okay, so what is Nyquist sampling theorem.that actually tells that if I have some trains of impulse
function or delta function, so which looks like this. so this is a periodic signal where at a
particular interval, let us say that is TS, I put a delta function. okay so this i represent as d d TS T.

796
okay, which is nothing but a train of delta function. So we can write this as delta t - n TS where n
goes from -infinity to + infinity.

okay, so this is just at different location in time there are d function okay. And it is a periodic one
after every TS it gets repeated. so the basic signal is something like this from - TS by 2 sorry,- TS
by 2 - D s by 2 this is just a d function. and it is defined from here. this gets repeated, ok. so
basically that is the Train of d function and sampling theorem says, see we have already told that
we want discrete values of it, so sampling theorem says, that we multiply these two so whenever
we multiply, with d if you multiply what happens it just picks that value at that point wherever
that d is defined.

So it will be actually picking those samples at those d locations. so the output will be something
like, this if I just draw the envelope, so this is where1 d so at that point d of the strength which is
following this gt will be captured. and so on okay, so this I can write as d t okay, so what is this,
this is nothing but our gt multiplied by d tst right. so this original function multiplied by this train
of d function. if you just multiply these two we get a sampling. okay.

So sampling is nothing but we are discretizing in time. so basically we are picking that sample
value at that location and then ignoring all others between this TS, and again at the boundary of
TS we are picking another sample value which is having amplitude exactly equivalent to that at
that point whatever GT has sampled value or amplitude value, so it is just picking those samples.
So we have got are presentation, so this is actually a sampled version of our signal okay. now let
us try to see what happens in frequency domain, so that is what we want to understand.

So this GT what is the corresponding frequency domain representation, what we are saying is this
gt has a frequency domain representation of gf which looks like this, which is defined from - let
us say B 2 + B. so eventually we are saying the signal we are sampling, that is a band limited
signal. and a low-pass equivalent signal, so it is a low pass band limited signal having maximum
bandwidth or maximum frequency component up to B. So it has a band width of B. so that, and
that is the gf that is the Fourier transform of that particular signal okay.

Now if we wish to do Fourier transform of this one. what we know it is the Fourier transform of
this,Fourier transform of this, and because they are multiplied in time domain it should be
convoluted in the frequency domain.this is something we know. so first we already know the
Fourier transform of this gt, so we need to understand what is the Fourier transform of this del
TST that is something we will have to first do. Then only we will be able to characterize the
Fourier transform of that g hat t okay, or g bar t okay.

So this del TST' what is this, is actually this looks like this, it is a periodic signal.with this as
being repeated every time, with a period of T s, so therefore I can represent this as a Fourier

797
series, so if I can do that, so what I can say is it is actually summation n equal to -infinity to +
infinity, okay Dn e to the power; omegas t. where omega s is just 2 pi by TS. or I can write it as
2pi fs okay. where fs is 1 by t s. okay so I can from myFourier series understanding I know that
any periodic signal can be written like this. where what is DN? must be calculated as Fourier has
said for Fourier series analysis.

So that must be integration over the period which is- TS by- 2 + TS by 2, and then we have to
multiply with the function itself which is d T. so d T into e -j n omega s t DT right. that should be
my D N. what is this it is d T multiplied by this. that must be at T = 0 whatever value of this is
there, so that is Z, that is 1, so 1 now you integrate this okay. So basically what will happen if I
we have also seen the property of d T, if it is integrated from -in fig to + infinity with a function,
it will just give me as output the value at 0, whatever it is.

So d T if is if it is multiplied by a function, and then integrated over the entire duration,I will be
getting this particular function, so if I integrate it from - TS by 2 or from -infinity to +infinity,
because d T is 0 anywhere else. okay so I can always write that, as if it is integrated from -infinity
to +infinity and then I will be getting at T equal to 0, whatever value I do get-okay so that should
be, what I was telling, that at T equal to 0 whatever this value that should be the value, so e to the
power 0, that is 1, so it must be 1 by Ts right.so DN becomes 1 by TS, so it is not a function of n
anymore. So it is independent of n. so I can actually take this outside the summation.

(Refer Slide Time: 24:09)

798
 1 ∞ −jnω t
∑
δTs(t) = e s
Ts n=−∞
1 ∞ 1
∑
= δ( f − n fs) fs =
Ts n=−∞ Ts
1 ∞
∑
g( f ) = G( f ) * δ( f − n fs)
Ts n=−∞
1 ∞
∑
= G( f ) * δ( f − n fs)
Ts n=−∞
1 ∞
∑
= G( f − n fs)
Ts n=−∞

So what I can write this d TST, I can write as 1 by TS, because I can take it outside the
summation n-infinity to + infinity e to the power j omega s t right. I can write it this way. what is
the Fourier transform of this? let us say that is del Ts f.this will remain the same. it is a
summation of exponential 1 each exponential if I take Fourier transform, what it creates, it

799
creates a delta function in the frequency domain at that location wherever it is. okay, so at that
frequency location. So basically this will be just a sum of delta function. so this happens to be,
right. this is something we already know. that if I have this, means I can I can get a time period
accordingly which is 1 by means 2 pi by TS, This can be represented.

At TS it will be creating delta, this is something we have already known, this is the relationship,
Fourier conjugate relationship, to the e to the power j 2 pi fst will create a d function at FS. okay
so this I already know, so it will be just a summation of d function. that is very good. now,I need
to find out this g bar f. which is gf or maybe capital gf, convolution of this which is 1 by Ts
summation n -infinity to + infinity delta t - n TS. right,I can write it this way. now, the
convolution is integration and this is a summation these two does not depend, so I can
interchange them.

So gf I can take inside, there is no problem in that because this summation has nothing to do with
F. so I can write this as 1 by TS, summation n-infinity to + infinity, this is nothing but gf
convolution d d ta, sorry this should be, it is a frequency domain representation, so that must be f.
right, so that is f - n TS. I have done a wrong thing. this is actually if I am putting it at f, it should
be n fs. okay where fs is one by TS. so that is the Fourier representation, right. if I put it in
Omega it will be n Omega s okay.

So that is, that is the wrong thing I have done. so, that must be fs, so this is what we are getting.
now we know anything convoluted with delta must be giving me the same thing at where the
delta is. okay, so the convolution theorem with respect to d this will give me TS summation and n
-infinity to + infinity this will be just gf-n fs. that is what happens.Just that g goes into every fs.
So in frequency domain what will be the corresponding representation?

(Refer Slide Time: 27:46)

800
fs − B > B
fs > 2B

If I just try to see, so that was my gf. now I am trying to get this g bar f right. so how that will be
represented? this is this entire thing gets repeated at every fs. so this is 0. At fs also this will get
repeated. at 2 fs because n goes from -infinity to + infinity. at two fs also this gets repeated, so
this is a fs, this is two fs, and so on.what is the strength of them? so the strength should be this
one by TS. so that must be the strength, so this should be 1 by T s. and this gf gets repeated
everywhere, even in the negative side also.

This will be at-fs, I will have that same repetition. right, so I will have this reputation from, input
to + infinity.Now the thing is that, i f I wish to keep the signal intact, that means after doing this I
can still reconstruct this signal.how in the frequency domain if you see, how I can reconstruct this
signal? if whenever I am shifting this. if the fs is chosen carefully, so that this and this are
separated, they are not overlapping. So you can immediately see if my fs is too small, suppose I
put fs over here, so what will happen this is getting repeated, and this is getting repeated, now
these things are getting superimposed.

801
So if fs is small, then I will have problem, so what should be my optimal fs so that I can still
separate them out, with a low-pass filter. that is all I will have to do. right, so for that what is this
point, this is fs + B, sorry this is + B just/and what is this points this point is fs - B, so I need to
ensure that fs - B is greater than + B. so I need to just ensure this. fs sorry, fs - B is greater than +
B or fs must be greater than 2B.that is the famous Nyquist theorem. so, as long as whatever the
bandwidth of the signal, if twice that, if I means twice of that Which is 2B. I do my sampling
frequency, that means I choose my TS in such a way that my sampling frequency which is 1 by
TS, that becomes bigger than this 21. I am fine. I know that I can put those samples and then I
can transmit it those samples if I just pass through a low-pass filter which has a bandwidth of B,
it will be able to trace that signal back. so I can reconstruct that signal. I am not losing any
information so I can still recover that signal back. it is only when this condition is violated, I
know there will be a super imposition and my sampling is not giving me very nice reconstruction
of the signal.

This is quite obvious. Suppose, I have a signal something like this tokay and then if I just take the
samples like this, what will happen if I start try to reconstruct, without knowing what has
happened inside, so I will just reconstruct it this way. right, so it will take out all the variation,
whereas if I start getting samples like this, it will capture the entire variation.

So basically, I and Nyquist criteria tells me that at least what should be the/means> what should
be that TS, what should be that value. so from here I can calculate one by TS must be greater than
2 B. or I can say this TS must be less than 1 by 2 B.so that is the maximum value of TS,1 by 2B. I
cannot really go beyond this. if my TS becomes more than 1 by 2B. I will be losing some of the
information. So my sampling will not be good representative of my signal. so as long as I am
while sampling I am considering this,I know that the sample will be a true representative.

So this was the sampling theorem which has given a huge impetus for modulation. because with
this only we know that probably the entire time domain I do not need. I can just pick some
discrete values of time, where I represent the signal and that is good enough. ok, so what we will
try to do we will come back and try to see, yes they are good enough. Those samples are good
enough we will try to show that theoretically.

802
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 57: Sampling Theorem (Contd.)

Okay, so we have started pulse modulation, that something we have already started. and we have
told that we wish to actually establish the theory behind this pulse modulation. So pulse
modulation does take discrete samples of time, signal right. So if you wish to do that we have
already proven the Nyquist Theorem, that yes there is a possibility that I can take just samples
and that will still, with the low pass filter if we put some criteria on the sampling we have a
possibility of getting the signal back, original signal back from the sampled signal. this is
something we have already proven.

And next what we wish to see that yes those samples which are, means following the Nyquist
criteria are actually the true representative of the signal. this is called the reconstruction theorem
or the interpolation theorem. So we will try to see that those samples are good enough, in certain
way if we present them they actually represent the whole signal.okay. So let us try to see those
things.

(Refer Slide Time: 01:32)

803
 ∞
g (nTn) δ (t − nTn)
∑
g(t) =
n=−∞

( 2B )
f
H( f ) = Ts Π

h(t) = Ts2B sinc(2πBt)

So what we have so far said, that this gt bar right, which is the sampled version, what is that, that
was actually at that instance we had some sample, okay. So that sample multiplied by Δ at that
point okay. So let us call that sample at nth instance as g(nTs) okay. So this just represents that at
n th instance what is the sampled signal of that particular G okay.

This should be multiplied by Δ(t-nts) okay. So at nts what is the sample, this one, and we are
multiplying there with the Δ. so that must be this gbar, because gbar was nothing but that sampled
with that strength, there was a Δ function at that location. So that is what we are doing, so this
must be a representation and goes from -∞ to +∞ right. So this is our gbar right.

Now what we have said that if it is following Nyquist criteria that is happening, then I can
actually filter this, that is what we have said that gbar T, if we try to see the frequency domain

804
representation, that looks like this as long as Nyquist criteria is fulfilled. If I just put a ideal low
pass filter, which is having a band of B. okay, this must give me back my original signal okay.

And also we need to see that this low pass filter that I will be putting, if I wish to go back to
actual original G, what was the strength of this one, that was 1/Ts. So low pass filter must have a
gain of Ts, so that they cancels each other. okay. So this has already gets in terms of power gets
reduced by 1/Ts if we can amplify them in terms of, sorry, amplitude gets reduced by 1/Ts, if we
amplify them by Ts factor this will originally go to our original Gf.

So basically this pass through a low pass filter must give me GF as well as Nyquist criteria is
fulfilled, right. So what we have to do, this signal has to be passed through a ideal low pass filter
of bandwidth B, and gain Ts right. So this is a ideal low pass filter, so at this side if we put G bar
T, I must be getting GT, this is something we have understood, as long as the sampling has been
done with Nyquist criteria. okay, that means that Ts is less than 1/2B okay, so that condition has
been already satisfied okay. So what is my HF then, that is ideal low pass filter, which has a gain
of Ts, so it must have a gain and that is a box function, so like in box function if we represent by
π, so it must be a box function of let us say F1 frequency with the box width of 2B, because it
should look like this from –B to +B.

So it is a box of width 2B, so this is a box function with a gain Ts and this one right. So what is
the corresponding Ht, which is the impulse response of this filter. okay, I know the HF
corresponding Ht, we know for a box function the corresponding Fourier Ht will be Fourier
inverse transform, so that is a sinc function. we already know that. So that sinc function should be
Ts remains over here that should be 2B because it is a 2B box function, so 2B sinc sin C, 2πBt
right.

Now, let us say we have just satisfied the Nyquist criteria. okay that means our Ts what have to
happen that must be less than 2B right. Just satisfied means, I also take the equality condition.
Now there is a case that when I can take the equality condition, if the highest frequency, the
signal we are considering highest frequency of that is not a impulse okay, there is no impulse in
the highest frequency and it almost goes to 0 okay. If this criteria is fulfilled then I can always
sample it, so what we are trying to say is like this.

(Refer Slide Time: 06:35)

805
 1
Ts =
2B
h(t) = sinc(2π Bt)

Suppose I have a signal whose frequency response is this, what is this, this is a cost Omega Ct , if
I now sample it there is a problem, if I just do it with a Nyquist frequency, of means, that
sampling frequency is just matching with the frequency of the sinusoidal. If that is the case then I
will have a problem, because this is a band limited signal of course, with the bandwidth suppose
this is omega M let us say, If this is omega M, the band or if this is FM, at FM we have
something okay. So this is FM, so this is a band limited signal of highest frequency FM, but at
FM it has a impulse. There if we start doing sampling, probably there will be super imposition of
things okay.

So as long as there is no sample at the end of the band, we can always even do Nyquist sampling
at that frequency okay.So basically instead of writing this less than,I can also write less than
equal to and the best I can do is Ts=1/2B. So let us say we have done this sampling,just with the
Nyquist sampling bit. So then Ts=1/2B, so this becomes 1, so I can write this as sinc 2πBT right.

(Refer Slide Time: 08:12)

806
 ∑ ( s) (
g(t) = g kT δ t − kTs)
k
h(t) = sinc(2πBt)

[∑ ]
g(t) * h(t) = g (kTs) δ (t − kTs) * sinc(2πBt)
k

∑ ( s) (
= g kT δ t − kTs) * sinc(2πBt)
k

∑ ( s)
= g kT sinc(2πB(t − kTs))
k

So this is all fine, now what I have to do is I had a g(t), right! which is represented as this.
Summation, let us say if have k or n which every you put k gKTs delta t – kTs, you can even put
n'earlier I was writing n'so you can also do that. okay, so this I the gT and my ht is sinc 2 p i Bt.
right, so what will be the out/it is convolution we know it is a linear time invariant system
because it is ideal, this one. so the output will be convolution of ht and g hat t, so ghat t

807
convolution of h(t) what is that?this convoluted with this one, this one convoluted with this one,
but this is a delta function.

So convolution will just, keep this signal intact, put it will just go at that delta location. right, so
what will happen it will be just have this g(kTs) delta t – kTs, this whole thing, convoluted with
sinc (2ΠBt). right, because this summation has nothing to do no k factor over here so I can take
this inside, the Σ so I can write k g(kTs) so this is just a 𝛿 convolution with this so 𝛿 (t – kTs)
convolution sinc (2ΠBt). this is just a sinc function but at that 𝛿 location so this must be sinc2 Π
B. Instead of t it will be (t - kTs) right.

(Refer Slide Time: 10:36)

= ∑k g (kTs) sinc(2πBt − πk)

So what is that/this is actually Σ k g(kTs) sinc (2ΠBt - (2ΠB x kTs) B x 2Bts we already know, it
is one. so this will remain as Π k'right, so what exactly is happening we already know that if we
do Nyquist sampling, and the from the frequency domain we have a understanding that if we put

808
this filter will be at the output will be just getting out the signal. so this must be gT. according to
understanding of the frequency domain, so if this is gT what is actually happening? let us say this
is my gT, I have now taken those samples at gkTs right.

And at that sample we are putting a sinc function. okay, so at that position we are putting a sinc
function with this parameter. so basically that sinc we will putting a sinc which goes to 0 over
here and then keeps on going. and here also there will also be sinc which goes to 0 over here
because this is actually just what is the separation that is 1 by 2B, and this sinc function with
2ΠBt, this is just a shifting factor just exit to next this one okay.

So basically they goes to at 1 by 2B, they goes to 0. so it is just those sinc function, if you just put
one after another. they will reconstruct the signal. so basically what happens we have understood
now, that a signal is nothing but you take those samples and put a sinc function over there. okay,
so that is all you will have to do. so basically it is nothing but addition of all those sinc which
recreates the signal.

So that is why it is called, that this particular thing is the interpolation of the same signal. gT can
be represented as a infinite sum like this one, infinite sum of sinc function. so sinc becomes the
bases, almost like Fourier series. so Fourier series we are saying' that any signal/can be/if it is a
periodic signal can be represented as the infinite sum of those exponential things' here also, it is a
infinite sum of delayed sinc function having strength, where there also, every frequency term was
having strength.

So basically this sinc function becomes the basis function and having strength which is exactly
the sampled valued at that location. wherever that sinc is defined. okay, so this I called the
reconstruction theory of sampling. so basically, why we are able to do sampling ? because
underlying we have this understanding. we know that, if I do sampling with Nyquist criteria,I can
always all those samples I can represent with the corresponding sinc function, and I will be able
to represent it.

This passing through a ideal low pass filter is nothing but, that every sample is now getting that
sinc function. because the filter has ideal filter has impulse response of sinc. so every pulse that
comes and falls on that, at that location there will be sinc function, which will be getting created.
so basically, this passing through ideal low pass filter just recreates those sinc function for every
impulse that you give at the input and that is what happens.

Because the summation of all those sinc function are the reconstructed message signal. so it gets
reconstructed. so reconstruction of a message signal is not a big deal, and that is why, we know
that sampling signal will always give us opportunity to again reconstruct it later on. okay, so this

809
is something we should have in minds whenever we are doing it. but remember it can only be
done if we are following Nyquist criteria.

Otherwise, there will be distortion. okay, so that particular distortion is called aliasing, in the
literature. so basically whenever you have a band limited signal. okay, of let us say – B to +
Bland when your sampling, the sampling frequency is less than 2B. so what will happen the Fs
might be it will not be this is 2 B. it should be beyond that, but we do not, suppose deliberately
but it over here, so what will happen this will get repeated at every Fs.

So at this location, there is a huge amount of aliasing which will be actually creating the signal
like . this because they will added, and then after that even if you put a low pass filter. so what
will happen?the signal will not be the original one. it will more look like this, this is the aliasing'
so, basically, whenever you do sampling you have to careful that you are employing Nyquist
criteria, otherwise your sampling will, will not give you opportunity, even if you put ideal low
pass filter, will not give you opportunity, to recreate your original message signal.so that is very
important, so and that is why sampling theorem is there okay.

So now, what will try to do before even going towards the modulation, like amplitude modulation
or pulse amplitude modulation, pulse position modulation or means, the width modulation. what
will try to do, we will try to see some kind of sampling which are more practical-okay, so we will
start discussing about some practical form of sampling.

(Refer Slide Time: 17:02)

810
 ∑ ( s) (
g̃(t) = g kT p t − kTs)
k

∑ ( s) {
= g kT p(t) * δ (t − kTs)}
k

[∑ ]
= p(t) * g (kTs) δ (t − kTs)
k
= p(t) * g(t)

So, practical sampling means, see as you might have seen that I did sampling of a signal which is
the band limited signal. so once it is band limited, the power spectral density is something which
is also band limited, and we integrate it from – infinity to + infinite, we get a finite power. right,
but whenever we sample it, after sampling what is happening ? within this 1 / ts times it is overall
amplitude spectrum is getting reduced. but this is getting repeated upto ∞ right, so overall powr is
becoming infinite.

So where from this, just by doing sampling we are getting infinite power. where from we are
getting this power? so basically what is happening, if you think about the sampler which was that

811
train of 𝛿 function. we have already earlier proven then there are 𝛿 function has, means single 𝛿
function as infinite energy. if that is the case, a 𝛿 function train will have infinite power. that is the
source of power. so basically we will never be able to achieve a 𝛿 function, or 𝛿 train or train of 𝛿
function.

This is something which is not possible, because that if you wish to get that kind of signal, that
will give you means that will be actually generating infinite power. which is not possible. so this
ideal sampling which was proposed by Nyquist, is not possible in practice, so we have to now
think about some sampling which are actually valid in practical scenario. so what will happen the
pulses is that will be creating, that will be of finite width, it is not infinitesimally small/like 𝛿
function. 𝛿 function the pulse width, we cannot talk about that width, that is infinitesimally small,
we have already given one example in our earlier class, that we have said/that is the pulse width,
let say like it – ε, valid from – ε/ 2 to + ε / 2 and the pulse width will b 1/ ε'then the area under it
is remains 1. and then we make ε tends to 0. so what happens is pulse with almost vanishes.
because the pulse rate is ε that goes, then means goes towards 0. and this goes ∞ • so that is
actually impulse. so in impulse, the pulse width is infinitesimally small, it tends to 0. whereas
actual pulse will have a finite duration.

So now, whatever modulation will be discussing or whatever, means, sampling will be

discussing, will be discussing with respect to pulse of finite width, or arbitrary shaped pulse,
which are practically generatable and available. okay, so therefore, suppose we multiply the
signal gt with respect to that kind of pulse'let say we have a pulse Pt/which looks like this. okay,
it goes to 0, so this is our Pt, not an impulse. so if this is the case, then when I do, suppose I take a
train of this pulse/so how that will look like ? so there will be a pulse again after Ts, there will be
another such pulse, and so on. which will get repeated from - ∞ to ∞, so if my signal original
signal gtl I start multiplying with this particular pulse, it will look like this' n or K whichever you
put. KTs now it will be multiplied not with 𝛿. but with this pulse, right. so this is what will be
happen ing. but what we can say is any pulse, is nothing but the convolution of that pulse with the
𝛿 'so Pt – nts, I can always write, mathematically, as this, this is nothing but Pt – NTs convoluted
withior sorry Pt convoluted with 𝛿t – NTs. okay so whenever we do this convolution, and this
will be creating a pulse at, same similar shape pulse at ti and t – NTs. so this to our equivalent
representation, that is something we know from the theory of 𝛿 function.because this Pt has
nothing to do, it does not have, sorry. this will nothing to do with Kiso this Pt can go out with
convolution, so sum and convolution can be basically flipped. so this happens to be VKTS into 𝛿t
– KTS right.

812
And we term this as g 𝛿t okay, so because this is not original sampling, so what we are not
writing it at as g hat or, g sorry g bar t which was original sampling with 𝛿. so this sampled with
that pulse and then we could prove this, so I can write that it is nothing but Pt, that pulse is pulse,
convolution with, this is actually, again we are going back to the sampling with 𝛿 train. okay, for
mathematically representation. so this we can write as g bar t. so that what happens, whenever we
actually sample it with any arbitrary pulse, so we are actually theorizing it, so that we can sample
it with any arbitrary pulse, then we can take some finite with pulse, probably. okay. So now if we
just do Fourier transform of this.

(Refer Slide Time: 23:29)

G̃( f ) = P( f )G( f )
1
G ( f − n fs)
Ts ∑
= P( f )
n
1
∑ (
= P( f ) G f − n fs)
Ts n

813
So basically if we write that as let us say capital, G tild f that is the Fourier transform of g. tilde t,
'let us say' these are Fourier pair and we also know G t as a Fourier pair of gf, and pt and Fourier
pair of pf, suppose we know that, so for the pulse also we know the Fourier transform. if that is
the case, what will be g ~ t, g ~f because g ~f we have already proven that is p t convolution g hat
or g bar t, so it must be in frequency domain multiplication. so that should be pf which is the
Fourier transform of the pulse itself multiplied by this g bar f right.

Now this g bar f we have already evaluated earlier, while doing the sampling theorem, so we can
write this tf, this should be this we have already proven that, that is 1/ts Summation, if we just
write n or k let us write n, so this is gl and means f- nt or fs right. So this is something we have
already proven that, this g hat t is this, equal to this. something we have proven in our sampling
theorem, by doing the Fourier series and then followed by Fourier transform, so that is what
happens.

So basically what we get, is 1/ts, p f. I can pf inside, right. so if the something wishes, or
otherwise I can keep pf over here, so basically it is Summation n, g f – nfs. right, that is all that
happens. right, Whenever we have any arbitrary pulse 1 now for any arbitrary pulse, we have a
representation right. so if this pulse was ideally δ t. then what will be the corresponding pf? can
be constant, so if this is a constant, immediately what I get? I get my earlier representation okay.

And of course that ts can has to be given, so that this ts gets cancelled. so basically if the pulse is
something like this,I can always put our low pass filter with ts gain, then ts will be cancelled out.
with the frequency ideal low pass filter and then because this gf – nfc just look like this. okay it
gets repeated, and then ideal low pass filtering will just give me back gf. but now what is
happening, this if I see this gf – fc that remains the same, let us say we have just does the, this FS
has ben chosen/so that they are just equal to/this fs is just equal to 2b. so then this will be just
repeated from there with the strength of each of them are one / ts right.

So this is my g, means this is actually my g hat, or g dot f. this if I pass through a ideal low pass
filter, with the gain of ts, so that ts and 1/ts gets cancelled, and I get this, original gf. but now,
because I have done the sampling with the pulse, this pf should be multiplied with this function.
Now any non ideal case of pfl suppose let us say my pf look like this. or my pt looks like this ra
pulse of certain duration, or my pt looks like this okay.

So it is defined from –tp / 2 to + tp / 2 okay, strength is one. so that is my pt, immediately what
will become the pf? Pf will look like, a because this is a box function, so it should be a sinc
function. because the width is tp, this should be 1/2tp. okay so that should be my pf, now this
should be multiplied with this one, so what will happen? Once I multiply, this particular gf also

814
will get a multiplication with this pf. so basically the gf, whenever it gets multiplicities the
frequency spectrum of that gf gets little bit distorted.

So this is what happens if I give a instead of ideal pulse, if I give some other non ideal pulse,I
will always get a distortion, but this distortion I can restrict by what method, if I make this tp
smaller and smaller, what will happens?this 1/2tp will go away further away and this will almost
become flatter in the region of interest.from –b to + b. there it will not get any distortion due to
that, due to the presence of the spectrum of pf.

So that is what we will have to do whenever your pulse is not a impulse function, we can
understand that but what you can always do, that instead of having a impulse function, you can
have a finite duration pulse, but lesser the width of the pulse better you will be in terms of
distortion of the original signal. And then if this is flat enough, so this will remain almost intact. I
can now put a ideal low pass filtering and I can get my signal back without much distortion.

So this is the theorem, when we do not put impulse or let us say impulse function, or if a train of
impulse for our sampling, because pulse will means pulse represented by impulse cannot be
generated, or recreated using any general or day to day circuitry/because it required infinite
energy. so if we give a finite duration pulse, the lesser the duration is, falter its frequency
response will be, and less amount of distortion you will be getting due to the presence of that
pulse okay.

So in the next class, what we will try to do, we will try to see different kind of pulse amplitude
modulation by employing this' means practical, means pulse train.okay, so that something will try
to study and will try to see what is the effect of them. and then probably will have a brief
discussion about what is pulse amplitude modulation, how we can employ, pulse position
modulation and pulse width modulation okay, thank you.

815
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 58: FDM Vs TDM

Okay, so, so far I think we have started discussing about sampling theorem, so why we started
discussing this is because we have already done the amplitude modulation and the frequency and
phase modulation. So this is something we have already seen that how that works, and what are
the relative benefits or what are the pros and cons of these things. And then after that we have
promised to start another form of analog modulation which is probably the boundary towards
digital modulation.

So this is called the pulse modulation, so there are few types of pulse modulation that we have
talked about, one is called pulse amplitude modulation where the information actually is being
carried over a pulse, but it is on the amplitude of the pulse. So basically if you see instead of
using carrier in pulse modulation instead of using carrier like analog or all other analog version
like amplitude or frequency modulation, what we are trying to do is.

(Refer Slide Time: 01:33)

816
We have a pulse train which looks like this okay. So we have a pulse train it might have different
shape, so that is something we will explore. So we have this periodic pulse train, and then we
want to modulate our information on this pulse, so pulse is almost working like a carrier. So what
we can do is there are three types of pulse modulation, we have already discussed that.

One is pulse amplitude modulation, so in pulse amplitude modulation what we do, suppose we
have a message signal okay. So depending on and suppose this is the pulse strength which goes
like this okay, it will have uniform period of either T or Ts let us say. So basically what we do, we
modulate the amplitude of the pulses according to the input signal. So it will be let us say this
becomes the envelope of the pulses and the pulses are just the tip like this.

So basically the amplitude of the pulse or the relative amplitude of the successive pulse actually
carry out the message information. So this is one kind of pulse modulation, this is called pulse
amplitude modulation of PAM, so that is something we have seen. We have also, we have talked
about very briefly other two version of pulse modulation, one is called the pulse width
modulation.

817
So like in the carrier we had amplitude, we had frequency and phase, for pulse also there are few
things which actually can tell us about modulation or possibility of modulation. So one is of
course the pulse amplitude that can be varied. Also the pulse position can be varied according to
the amplitude of the message signal. So what we can do this, because it is a periodic one, so it has
a definite pulse position.

Now we can start varying this position of the pulse keeping the pulse width and amplitude same
according to the message signal. So let us say we can have this convention that whenever the
pulse amplitude is higher then probably we will deviate from the center point little bit more
whereas whenever the amplitude is lesser will deviate much lesser and something like this. So
with this we can have a pulse position modulation where the pulse width remains the same, pulse
amplitude remains the same.

It is the pulse position relative to where the pulse should be, original pulse should be, original
periodic pulse should be. So if we do that then the position of the pulse is actually carrying the
information of the, about the message signal. So that is one way of doing it, another way of doing
it is now because we have three parameters one is the pulse amplitude, other one is the pulse
position and the third one is the pulse width.

So we can also do pulse width modulation okay, so this one was called as pulse position
modulation, and in width the pulse remains in the same location as the input sorry it should go
from here okay as the input. But the width will be now varying according to the amplitude, so
this might have a higher width, next might have a lower width and so on, this might have a bit
higher and so on.

So basically the pulse amplitude remains the same, and the position remains the same that is
starting point of each of the pulse that is fixed at after every Ts duration the next pulse will be
starting. But the width of the pulse will be now varying according to the message signal. So these
are the three possible modulation that can happen. So we actually call these kind of modulation as
discrete time modulation.

So basically earlier whatever we are doing the signal was continuous time, because we had a
carrier and that was existing for all the time instance. Whereas, pulse that exist for some duration
of time, rest of the time there is nothing. So we can call this as discrete time modulation, whereas
the amp means the information that it carries that can be means that is not discrete that can take
any value so either it is in width or position or amplitude that can take any value so that is why it
is still a analog modulation because the information that it carries that is in analog form that is not
discredited okay.

818
So to develop the theory of this pulse modulation what we have started is something like this
initially we have talked about theorem which is well know and all of you might be already
knowing in some version so we have given a theorem which is called the Nyquist sampling
theorem okay and at that point what we have taken.

(Refer Slide Time: 06:57)

fs ≥ 2B

fs < 2B

So I am just quickly recapitulating what we have done so nyquist sampling theorem means the
pulse is where we still have pulse train but the pulse where impulse okay so individual pulse was
just a impulse function or we should say kronecker delta 𝛿 function okay so it is just a 𝛿 function
continues time 𝛿 function so which is defined has the pulse okay with that we started actually
means started doing almost like pulse amplitude modulation or we were rather telling that
because it is a impulse function so we were actually telling them as if we were talking samples of
the signal.

819
So if we had a signal we are just multiplying by this so what because if you multiply by 𝛿 it will
just pick the sample at that instance so it was just picking this instances the sample value at that
instances and these sample values where now the where represented as discrete time continuous
amplitude so it is still analog signal okay.

So what we could see from Nyquist sampling that basically the spectrum suppose this particular
signal just the signal has a spectrum like this it is a band limited signal from –B to + B if we just
sample it goes down by so suppose this is 0 goes down by actually 1/Ts the Ts is this or 1/Ts is
the frequency of those pluses so basically it goes down by 1/Ts and it gets repeated at the
sampling frequency so basically if this is 1/Ts so this is will be 1/Ts next will be 2/Ts and so on
and similarly it also gets repeated over – 1/Ts or you can call that as Fs okay.

So this is what happens and then whenever you do sample why we are sampling it we want to
reconstruct the signal so that was our target that if we sample it we should have provision to
extract the signal it is also most similar like that we modulate we must have a means whenever
we do modulation we also have some mechanism to actually demodulate it so that we get the
original message signal back here also we are trying to do same thing.

We are trying to see if we sample it or let us say with impulse train we are doing impulse
amplitude modulation the we need to have a Mechanism which takes my signal out undistorted so
if that has to happen then has you can see this spectrum of this signal is remaining in tacked has
long as we have already proven and this FS see this is B this is B that Fs is greater than 2B or I
should say greater than equal to that equal to becomes that N quest sampling ratio. Okay.

So long as that is happening there is no problem if it is not true that fs is less than 2B then what
will happen when ever I do this sampling so that means actually you are under sampling so you
take a sample over here because fs is less than 2B so what you do you take sample over here so
this is will be the second sample probably this will be 3rd sample so if you under sample it then
what will happen this things.

So because fs is now no longer so this is fs no longer greater than 2B so there will be a

interpenetration of these 2 spectrum and then what we have, we have a distortion over here okay
so this is called aliasing okay so it will the spectrum will look like because these two will super
impose so it will be smoothen out over here so it will look like something like this okay so if you
now see this is not equivalent to our message signal, so if I satisfy the Nyquist criteria that means
my sampling frequency is greater than to twice of the band width of the message signal.

Then I can always say that I can employ a ideal low pass filter okay which will exactly take out
because there is no aliasing there is no super imposition of the neighboring means that spectrum

820
that is being created side spectrum so there is no over lapping so the spectrum remains intact and
I can put a ideal low pass filter of band with B and I can take my, extract my signal out so if this,
nyquist criteria is satisfied I can get the signal out thats the whole thing, in a way we can also
explain this physically,

So if suppose I have a signal what we mean by highest band width okay so highest band is
suppose B we are saying that means that is the highest frequency so over a unit time because this
is the highest frequency so they will be means restriction of rate of change of the signal so
basically if my sampling rate is somewhere related to this particular band width so basically I will
ensure that I sample at a instance where I can still have the tracking of the signal, so what does
that means that actually mean.

Suppose I have a signal like this and because this variations are there so a particular highest
frequency component is already there if I start sampling it like this then what is happening if I try
reconstruct it this sample, this sample and this sample the easier method is interpolation and if I
interpolate them we will see probably it will be smoothen out so all this variations are not getting
captured, whereas Nyquist criteria is saying that I have to depending on the bandwidth I have
take enough number of sample, So that this particular variation within time is being captured by
those samples and it nyquist criteria actually tells what are the minimum number of samples
depending on the bandwidth of the original message signal are required to represent the signal.
okay so this is something we have already done we have proven the Nyquist theorem also and we
have done another part of it which is called the interpolation theorem, so basically what we have
shown through Interpolation theorem so I am just recapturing, captulating all these things we are
not again mathematically proving it because that part we have done.

(Refer Slide Time: 14:25)

821
g(t) = ∑k g (kTs) sinc(2πBt − πk)

So what we have said suppose I now have just the samples okay whenever we say we pass it
through a low pass filter, so low pass filter if you see the transfer function that is a basically a box
function, so what will be the corresponding Fourier inverse transform that should be a sinc
function so if this impulses are getting passing through this particular impulse response, there
should be a convolution and everywhere a sinc function will sit together means sit one on top of
one another and we have also proven that the 0 are exactly over the other samples.

So if you start putting all those sinc function and we have proven that if you just a every sample
you start putting the sinc function corresponding sinc function with this amplitude, so you will
see the whole signal those sinc function getting added will create the original signal original
message signal which will just go through those steps okay, so this is the reconstruction theorem
so we have already told that our gt a message signal can be represented as this g of KTs.

So these are those sample values multiplied by sinc function okay sinc function appropriately
delayed so depending on the value of k so that is why you are putting a sinc function over here,
over here , depending on the value of k if k is 0 you put a sinc function over here k is one you put

822
a sinc function over here and so on okay. So accordingly you fill up all the sinc function and that
should give you the original message signal back okay.

So this is called the interpolation theorem so basically Nyquist sampling theorem and
interpolation theorem together they tell us that those samples whenever you follow the criteria of
Nyquist sampling those samples are good enough to represent the signal okay, if you just pass it
through a those samples if you just pass it through a low pass filter we probably get your signal
back so this is a very important result and this also gives us some possibility of again
multiplexing signal so that something let us try to see what happens exactly.

(Refer Slide Time: 17:05)

So basically suppose the multiplexing was one of the important thing that was required for our
transmission so we have said that suppose I have multiple signals okay and I want to
simultaneously transmit that over a same media so let us say air is the media or if I have a wired
media so everybody wish to actually simultaneously communicate through that same media okay
so this is something we have done in amplitude modulation or even frequency modulation, so or
general modulation theorem.

823
So what we are doing we were actually in the same media we were modulating with the different
frequency of a carrier so what was happening in frequency domain, they were, suppose this is
corresponding to f1 this is f1 f okay this is f2 or let us say g 1f, g2 f and so on, so this is at fc this
is at fc’ so we put them in different frequency so that they co-exist in different, different
frequency band and whenever we have to demodulate we first employ a band pass filter we have
already seen that band pass filter is probably the first part of de-modulation.

So we employ a band pass filter which is exactly matching with the desired signal so basically the
desired signal will be at some frequency if I already have understanding means I know already
that frequency so I will be putting my band pass filter centered around that so that I can just filter
out the desired signal and I can reject all the other signals okay. this serves two purpose we have
already we have seen that one is that it actually means take the desired signal and also it is rejects
some of the noise which are present in the entire band.

So this is something we have done and this particular technique was named as FDM frequency
division multiplexing okay, so we were trying to do frequency division multiplexing over there so
that we can reuse the channel for means actually putting simultaneous transmission. This is
something we have already discussed we have done that through modulation. Now by this
sampling theorem we have another opportunity so what is happening in sampling theorem, we
suppose I have a signal let us say that is again g1 t okay I put few samples to represent it now if
you just see it in time domain these samples we have already proven through interpolation
theorem that they are the sole representative of the entire signal.

So I do not need all the description of this signal for the entire duration of time I can just take
those few samples that means discrete time description of the signal is good enough for me to
represent a whole signal okay. if the signal is band limited as Nyquist has already pointed it out
so if this is something which is happening then I know that these samples are good enough to
represent the signal so basically what this is doing? This sampling theorem is doing? it is freeing
up of some of the time which are not being utilize now for transmission of the signal.

So whenever we transmit the signal now after doing this sampling we will be only transmitting
those samples so only few instance of the entire time line I will be transmitting those samples rest
of the time my transmitter as well as my receiver will be ideal so I can actually employ these
things to multiplex other signals okay so that is my purpose now what I want to do suppose I
have another signal g2t which is having a different shape so I also can samples so this was the
sampling instance so these two signals simultaneously exists so time has unit definition and what
I do because I know that these particular parts are free in time domain.

824
And my sampling does not say exactly where from I start sampling it can be anywhere but only
the regular interval should be maintained so if this two have equivalent band width okay so that
means this also is having B may be the shape is little bit different so –b to +b and this is probably
another voice signal suppose let say this is Voice signal so this 3.4 kilo Hz I have another noise
signal where the shape is little bit different.

But it is still from –b to +b is that happening when the sampling rate will be same equivalent for
both these signals but what I can do, I can actually start sampling it little bit opposite. so let say
that is ∆ off set so then because the sampling rate is same so it should be second sample also
should be ∆ offset and so on for every other samples.

Now what I can do, I can actually superimpose these things or add these two signals and transmit
in the channel so how that will look like so there will be one sample from here and next sample
will be from here, then one sample from here, next sample will be from here, so if you just keep
doing that one sample from here, and next sample from here, if we start joining the corresponding
sample tip you will get the original message signal back but these samples will co exist in the
channel.

Now what I have to do is if I do this at the other end I need to know exactly the timing is very
important earlier what was happening when I was doing frequency division multiplexing the
frequency was important, the frequency information should be known that which central, which
carrier frequency I have modulated I have to put my band pass filter accordingly.

Here the multiplexing is being done in time domain so that is why it is called TDM time division
multiplexing and the time information is very important so I need to precise it know where my
first signal sample start if I know that I only take this sample and then skip all the samples which
are intermediate which are due to multiplexing other signals actually.

So I skip this side signal and I can pick this one just after that Ts duration sampling means of
duration or inter sampling deviation so I can pick this signal as long as I am properly time
synchronies then I will pick this one so I will be picking my signals and then I reconstruct that
means I just pass it through low pass filters I reconstruct the whole signal and I get back my
original signal okay.

So this is something which is possible so this has been made possible because of sampling
theorem so you have see there was probably in communication or analog communication I should
say there are two fundamental theorems one is called the modulation theorem which actually if
you multiply with the carrier it translate the signal to a different frequency which is centered
around that carrier

825
And if you sample it properly knowing the band with and then following Nyquist criteria then
you are actually freeing up some of the time between samples and where you can multiplex other
signals so these are potentially two multiplexing schemes that has been employed in
communication you will see later on that mostly your means AM modulated signals fm
modulated signals goes through this.

Pulse modulated signals are, Because it is already with the pulse so it goes towards the tDM
multiplexing scheme. so it can still be done width fdm and sometimes the mix of tdm and fdm is
being employed. so these are the things which can be done but pulse modulated signals they are
closer to time deviation multiplexing and that is why tdm is employed for multiplexing in pulse
modulated signal okay.

So this is something which we have understood. now let us try to see another every fundamental
thing in communication or this kind of pulse modulated signal. so what is happening I have a
signal,I sample it, right. let say these are ts and my sampling frequency is 1/ts. okay, which
satisfies the criteria so that means fs must be greater than equal to 2B. okay,

So this is something which is being satisfied. so minimum I can take to make it let say. because if
I take because one my samples will be closer because sampling frequency increases, so ts will be
decreasing because its samples will be closer less time I will get between the samples. so the best
I can do.is if fs is just to be and, I am sure that the highest frequency whenever the band with I
defined highest frequency does not have any means corresponding all component okay.

So that is something which is required. because otherwise there will be again aliasing. so if the
highest frequency does not have any significant power component or it does not have any
impulse, or in kind of carrier. so if I just sample it at 2b, then what will happen this will be 1/2b.
so that TS becomes because FS becomes 2b and this becomes TS becomes 1/2b. okay, so this is
something I already know. now let us try to see how much of information I am transferring. okay,
with this. so basically whenever I am transferring how much information I am transferring. see
how many samples I am transferring, each of the sample actually carriers information.

Okay,So if the frequency is, the time is 1/2b frequency is 2B. so basically 2B samples I am
transferring per second. okay, and that occupies how much bandwidth, whenever I am transfusing
it off course because of sampling it is go over here, over here, and all those things. but these
things are not required for me. when I will be at the demodulated side.I will be putting I will be
putting a filter over here. okay, so basically my signal is confirmed over this. okay, so B. what
does that means? so basically this 2B samples I can now transmit with B bandwidth.because this
2B samples I am transferring these are the information.

826
So 2B information per second I can transmit with b bandwidth. okay, or B Hertz bandwidth.if it is
represented in F domain. not in ω domain. so with B, 2B information I can transmit. so per hertz
how much information I can transform? basically 2 information per second. so this is the very
fundamental result of, later on we will see for digital communication this is probably the most
fundamental result that every hertz, if you do sampling every hertz possibly carriers two
information. whatever that information is, per second. Okay every hertz is capable of carrying
two information for second. this directly comes from the sampling theorem. okay if you do
sampling little bit means, if you do over sampling probably this will be little lesser.

But this is the best you can do every hertz at most can transmit 2 hertz per second. okay, so this is
the very important and fundamental restriction we should say whenever we employee sampling.
and then each samples are information for us. okay, so later on we will see that when the from
pulse amplitude modulation to pulse coded modulation the PCM, we go, this is actually theories
the amount of bandwidth that is required for a PCM transmission.

So that is something we will see later on. but that is for now. for the time being we should be
happy about this result. so basically in this particular class we have discuss three fundamental
result, side by side. one is the modulation techniques which is the multiplication with the carrier
frequency and there is a translation which helps in FDM because the division multiplication the
next one is the sampling theorem, which helps us in, this is actually sampling and freeing up
some of the time.

So any signal does not have to represented all the time. it just needs few samples, few discrete
time representation of the signal. so that is how we get time division multiplication or we have
opportunity to do time division multiplication. and then we also understood whenever we do this
sampling and each sample becomes information for us, so every hertz is capable of transmitting
two information per second. okay, so that is the relationship between bandwidth and the
information transferred. okay, so this relationship will keep on talking about later on and
whenever you will be doing the digital communication probe this is the most important and
fundamental postulate, that will be required.

Okay, So with this background what we will do we will start now exploring some of the more
generic practical sampling that can be happening. because this impulse sampling is something
which is impractical, because each impulse will have a infinite energy we have already proven
that. so that is not possible we cannot really generate infinite power. so we need to see if we
employees some practical which are the restriction of circuits, if employee those kind of sampling
what happens? okay, so that will be our next discussion thank you.

827
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture 59: Flat Top Vs Natural Sampling

Okay so in the previous class we have talked about few fundamental aspect of communication,
now we are into pulsed modulation, so specially pulse amplitude modulation, so let us try to see
if we try to employee a practical pulse modulation okay.

(Refer Slide Time: 00:43)

828
 ∑ ( s) (
g̃(t) = g kT p t − nTs)
n

[∑ ]
= p(t) * g (kTs) δ (t − nTs)
n
= p(t) * g(t)
1
G ( f − n fs)
Ts ∑
G̃( f ) = P( f )
n

So let say I have a signal it looks like this. okay, I want to do a pulse amplitude modulation but
remember the pulse is longer impulse like one, we have taken impulse strength for proving our
Nyquist theorem, it is not, it is unlike that. So let say the pulse might have some shape and we are
calling this as P t. correspondingly we will also have a Fourier transform of this, let say, some P
fit might be anything, it might be band limited, it might not be band limited so that depends on
the pulse, shape of the pulse okay.

So let say we have we know this there is some pulse and we know the corresponding frequency
response of that pulse. so we are trying to make this analysis generic, so that for any pulse this
will be true. So we can have pulse like this of finite duration and this will just the special case of
that Pt. we will try to see these things later but right now we are saying any pulse any pt which as
the corresponding furrier transform which is Pf. So these 2 things are known.

So basically what we are doing ,we are again sampling it but now and the sampling intervals are
again related to the Nyquist interval and all those things okay.That's there, But now what is
happening this a particular part will be multiplied by this kind of pulse. similar to Pt.then this
one, something like this and so on. So again the tip of the pulse basically constructive. okay, so
this we are trying to do.

So let us try to see whenever we do this modulation, what exactly we are getting. so let us say we
call this particular modulated pulse train we call this has g tilda (t)-okay, so what is this? we are
just taking a sample at every instance separated by Ts. so we are taking samples and multiplying
that whatever instance that what every the strength of that sample with the pulse. okay so
basically our this thing will be represented has nothing but this g (nTs), whatever the pulse is t-
nTs right so that is actually our this particular signal.

So basically what is happening every time at a particular instance suppose n = 0, so that instance
I get this amplitude that multiplied by this p, I get this pulse. bigger pulse probably. next instance
at n = 1 so I get gTs and then multiplied by pulse must be shifted by Ts amount. so n = 1 so it

829
should be shifted by Ts amount and again multiplication. so this is exactly what is happening
whenever I means represent this pulse modulation. okay so this is basically my over all pulse
modulated signal.

Now p, I can always write has pulse and the convolute it with 𝛿. because I know any signal
convoluted with 𝛿 remains the same.so I can definitely write this as Pt convoluted Σ n gnTs that
has nothing to do with the t, so therefore this is just a constant thing and this Σ and the
convolution can be means basically exchanged. so whether I write it inside the convolution or
outside the convolution it is all the same.

O basically 𝛿t – nTs where ever the 𝛿 position is p will convoluted with that and create a pulse at
that position. corresponding the gnTs will be multiplied and get the same thing. so I can have a
representation like this. what we can see what is this part we have already studied. this is
basically the impulse sample version. so earlier we have called this has g^ t. okay this is the
impulse sampled version of the signal. okay so I can write this g ~t which is actually means
modulated with practical pulse which is Pt which has a frequency response of pf so I get this.
okay.

No problem in that, now let us try to see what will be the corresponding frequency response. okay
or if I do Fourier transform so let us say I call this as G~f, the Fourier transform of this one. so
that must be Fourier transform of this one, now it is convoluted in time domain, so in frequency
domain they must be multiplied. so what was the fierier transform of this was 1/Ts, we have
already proven. and G (f-nfs) okay.

So this has been already proven, we have already proven in that impulse means if you take
impulse train and then multiply the signal with that or modulate the signal with that we get this
frequency response. where G is actually this Gf is the Fourier transform of gt. okay, so this is
something we know already, that is a band limited signal and all those things are there and our FS
is also following that Nyquist criteria okay.

So this is this part G^T that is the Fourier transform of that Pt, I know Pf. so that should be my
Fourier transform so this is the overall Fourier transform. okay, so this is something we have now
understood that if I basically try to modulate my signal with a realistic pulse then the
corresponding Fourier transform will be this Fourier transform which was the original Fourier
transform if I do it with a impulse okay.

830
Which, with which we have prove in the Nyquist sampling and then we could show the
interpolation that, just by the low pass filter we can extract that signal out. now we are seeing that
if we have, means non-ideal or non impulse, pulse train, then the overall Fourier transform has to
be multiplied by the Fourier transform of that pulse okay.

So let us try to employee these thing for a particles scenario. let us say of course impulse I cannot
generate, but I can always generate this kind of pulse which is of finite duration.

(Refer Slide Time: 08:30)

( Tp ) ( Tp )
t − 0.5T t
p(t) = π =π

p( f ) = Tp sinc (2π f Tp)

1
G̃( f ) = Tp sinc (2π f Tp) G ( f − n fs)
Ts ∑
n

831
Let us say the duration of this pulse is Tp. okay where of course Tp must be less than Ts. okay,
wherever I am constructing a pulse train, so basically this is the pulse the next pulse so this
between these two pulse the difference is Ts, which is the sampling interval land this is Tp.
okay.so then what is the corresponding Pt?' that is actually a box function.

So this is a box function of duration TP. so duration Tp and because this I take as 0, so basically
box function, this must be the center point. so there is a 0.5Tp shift. so I can write this as T –
0.5Pt okay, so this is my pt. right, so that happens to be Pt. immediately what I can do?for box
function I can always generate a pf. corresponding pf is something we can always do. so what
will be that pf ? how do I get that pf? this is just a delay.

So you can even ignore that if I have my pulse described like this. then I can be ignored that part,
so I can as well write it as πt/Tp. so that means, it is a box function which gets repeated and that
has a. sorry it is not repeated, it is the single pulse we are describing. it is a box function centered
around 0/and it has the width of the Tp, so basically it looks like this. okay so it is defined from –
Tp / 2 + Tp/2 + Tp/2 and this condition satisfied, so immediately if this is the box function,I will
get a corresponding sinc function.

Tf so that should be Tp sink 2πfTp. right ,so this is my pf so therefore, my g ~f must look like Tp
sinc 2πfTp okay into 1/Ts gf – nfs. so this what I get. so what is happening if you just try to see in
this spectrum, so I have this structure which was the earlier structure as long as I satisfied the
Nyquist criteria. so it gets 1/Ts down if this is my – B to + B, this is my G(f), so it gets 1/Ts
down, and this is till define from – B to + B and it gets repeated after every np 1/Ts or equal to
Fs.okay so that what is happening.

So this still remains so this is that part, this gets multiplied with this part. okay so this is the
disturbing part, because what will happen, how does this looks like this is the sinc function. so
that must be, if I just draw it down, so that is a sinc function. which is touching 0 and 1/2 Tp, and
– 1 / Tp. okay, now what will happen? sorry, it should be, I have done little bit wrong, so it should
be centered over him. so let me draw it one more time.

(Refer Slide Time: 12:32)

832
So this is, and this is the center. so this looks like ½ T p.okay and – 1 /2 Tp. and these two gets
multiplied, if you see what is happening? earlier what we were doing, we were just putting a band
pass filter, getting our signal back, but now my signal is a composite signal multiplication of
these two, and unfortunately for me this is getting multiplied by this, which does not have a flat
characteristics over the entire frequency band of interest, so what will happen there will be
because of this multiplication there will be a distortion to this signal which will be created. so
whenever we put a pulse like that, there is a possibility of distortion. so what will happen once we
multiply these two, probably it will look like, these things will be little bit suppressed, so we will
look like this. should of-course be symmetric. so little bit of distortion has been created. it is not
ideal for this one,I can still put a band pass filter now – B to, sorry low pass filter of – B to + B. I
get this signal back.

Which is actually a multiplication of these two. but our demand is this one. okay so that is where
we are having a problem now. now let us try to see how do I make this flatter ? so this will be
flattery if this point where it is touching 0 stretches out, so it goes away. then it almost looks like
a flatter. so if I just draw it if this touches over here, that's – 1/2 P, 2 P and this is + ½ Tp then it
becomes so basically, as it goes, now you can see within the band of interest it is almost flatter, so

833
whenever I multiply with it the original shape will be in tact. what this means? so ½ p I am
increasing that means the Tp has to be reduced. right.

So what happens to the pulse? This is my ts, and I am trying to reduce the pulse width so this will
be exactly flatter when I have the pulse which has almost infinite similarly vanishing width. okay
which is an impulse function. again so I know that if it is impulse then this will be completely flat
and I do not have an any distortion. but I cannot generate impulse, so what I need to do is, I need
to generate a pulse definitely, finite width pulse, because that is what I can do, because then the
energy will be restricted. but I need to also ensure that the pulse width should be as small as
possible. because then they will be no distortion due to this modulation technique okay.

So this is something we have to keep in mind, and this corresponding sampling is termed as flat
top sampling. so let us try to see in time domain what is happening and why it is being called flat
of sampling.

(Refer Slide Time: 16:10)

So I have this message signal, I do have a pulse which are like this. some finite width pulse. and
at every instance, I take the sample value, and I represent a pulse with that sample value. next

834
time, I take another sample value and I represent the pulse with that amplitude of that sample
value. next time again I get another pulse and I get a, so if you see the pulse with which we are
modulating the top of the pulse remains flat. because we have taken a square pulse, so a box of
function.

Because the top of the pulse is flat, and it just takes one instance of the sample value, so the
sampling looks like this. I take a sample I maintain that. okay, so this is being realized in practical
circuit by sample and hold circuit. so basically what we do we take a sample and we hold it for
the pulse duration, and that is how the top of the pulse becomes flat and corresponding pulse is
called flat top pulse.

This kind of flat top pulse is done by sample and hold circuit. okay so which probably we want to
be discussing that because that is part of the circuit if you just do and look for sample and hold
circuit you will get plenty of example of that. different realization of sample and hold circuit, so
with sample and hold circuit people do this flat top sampling, or sampling with flat of pulses or
pulse strain.

So the disadvantage of this one is, what we can see that we need to really make sure that the
pulses are very small, that is the first thing, second thing is probably there is a corresponding
distortion which is coming, and that is why probably we are making this pulse width very small.
we have already demonstrate that, that 1/tp has to be very big that means, the tp has to be very
small-okay.

So the pulse width has to be small, and with this whatever finite pulse width you take there will
be certain amount of distortion, because it is a overall spectrum is getting multiplied by pf so that
is giving distortion to the bens band signal as well/and whenever we low pass filter it will get
some amount of that distortion-okay so that is the disadvantage of this kind of sampling, okay, is
there any other sampling?so let us try to see, if flat top sampling is one of them, is there any other
counter part of this? that will be our next target. So the counter part of flat top sampling is call
natural sampling.

(Refer Slide Time: 19:28)

835
 ∞
Qne jnωs t
∑
qTs(t) =
n=−∞
Tp /2
1
Ts ∫−Tp /2
Qn = q(t)e −jnωs t
\
Tp /2
1 e jnωs t
=
Ts −jnωs
−Tp /2

What is natural sampling? So natural sampling is something like this, I have the signal I have
pulse train. okay so whether probably this pulse amplitude is unity, assumed to be unity. so it has
it is a pulse train so it has pulse from -∞ to + ∞. so what I do, in the previous one I was taking a
sample and then multiplying by that pulse. okay or I was with that sample amplitude I was
representing a pulse. okaylso just that pulse was getting multiplied by that sample amplitude.

Here what I will do, this gt I will actually multiplied with this, means pulse train . so let us say
that is qt. okay or let us say qt st because this is the ts okay, so I will be multiplying it, what will

836
happen? Wherever it is one, it will be, fallen because I am multiplying, so because it will be
multiplied by 1, so this will be sampled all right. but during this duration it will just follow
because it is multiplication one with this one, so it will just follow the way signal is, whenever
the pulse is on, so it will just follow that signal. so it happens to be this way. so basically as you
can see you again connect these things your signal will be degenerated okay.

So this is called the natural sampling, the natural sampling means within the pulse duration you
are actually allowing the signal or the pulse shape or the top of the pulse to follow the signal. that
is what we are trying to do. okay so can I now characteristics it in frequency domain and then
tried to see the way we have done for the flat top sampling. That is something will we try to do.

So for that, first of all, we have to characterize this things.so Qts (t) what is that?that is a periodic
signal. right so if it is the periodic signal,I can always write in Fourier series/so n from minus
infinity to plus infinity, sum Qn, are you can write it as dn/the way we are writing'e to the power j
n ωst, where ωs is equal to 2πfs where fs is 1/ts right.

Because that is the fundamental frequency. so we will get this. where the Qn the way Qn is
calculated, so that should be 1/ts,- the pulse is defined. so if this isthe pulse from –tp/2 to + tp/2.
So –tp/2 to + tp/2, is defined. qt e -jn ωst dt, so we can do this, so if we just evaluated this, what
you get 1/ts. okay and you have, because this qt remains 1 within this so I will have 1. from – tp/2
to + tp because this is pulse is having unit amplitude, that is what we have assumed. so it is just e
to the power this, you integrate it, so e -jn ωst divided by in omegas, putting-tp/2 and + tp/2, okay
so that is what we are getting.

(Refer Slide Time: 23:46)

837
 Tρ
2 e jnωs 2 − e −jnωsTp /2
=
nωsTs 2j

[ 2 ]
2 nωsTs
= sin
nωsTs
Tp
= sinc [nπ fsTp]
Ts
TP ∞
̂ sin (nπ fsTp) δ ( f − n fs)
∑
G( f ) = G( f ) *
TS n=−∞

And then, if we evaluate this, this happens to be, okay so if I just do simplification after putting
tp/2 and –tp/2, so I get 2/n ωs ts ,e^jn ωstp/2- e^-jn ωsTp/2/2j, I have taken this form because it
will be our sin function. so I can write 2/n ωsts this is sin n ωs tp/2, and this I can manipulate, so
basically to bring out these things so that I can get a sinc function. so finally if I just do that
manipulation I will get tp/ts sinc nπfs, so I have put ωs is equal to π fs Tp.

838
This what we get. okay so qN's are all being represented now, now what is, we have got by qts.
okay so what will be the corresponding frequency response. because Qts is nothing but this qn e
to the power this ∑ - infinity and + infinity. so this is just become impulse function at frequency
domain at those frequency and ωs or n fs okay.

So if I now, means what we are tried to do is, if you just see this sample qts is getting multiplied
with gt. okay so if it is multiplication in because we have already characterize the Fourier
transform of this one, because we have got Fourier series and then you do the transform you will
get all kinds of delta function with the strength as q n corresponding strength okay.

So now if these two signal are getting multiplied in time domain exactly to create this,so
therefore Fourier transform of this one will be convolution of gf and this one. fine, so that if I
represent that has as g tilda f. that must be Gf, which is the original missing signal, convolution
Tp/ts sinc nπfs Tp. so this n goes – infinity to + infinity.

And I have the delta that every fs, corresponding fs. right, so this is the Fourier transform of that
qt st, because it is just exponential. so corresponding they will be creating delta at every fs. so I
get and the strength will be corresponding sinc so this is what we get, now gf convoluted with
this ∑ and this convolution. I can inter exchanged, so gf can go inside then g (f)convoluted with Δ
, again it becomes just g (f)-nfs. so I can write this as tp/ts∑n sinc (nπfctp), this becomes of g(f)-
nfs. right so that happens to be my overall Fourier transform. okay so now what you can see is
this.

(Refer Slide Time: 27:46)

839
 TP ∞
sin (nπ fsTp) G ( f − n fs)
∑
=
TS n=−∞

That the Fourier transform is nothing but at every n. so basically whenever n=0 I get g(f) so n =0
means it is at the center frequency n=1 means right side-1 into left side so like this. so there will
be frequency term, which has, so it will be just exactly g(f) . so the frequency spectrum remains
the same, only thing is that, it has the strength which is sinc function. okay and tp/ts. so this sinc
function what happens this will be bigger and the next one will smaller and so on. they just keep
on going like this.

So this tip means the strength of them follows a sinc function at every location of phase. so this is
0 this is fs this is –fs and at every location of fs, the tip of them actually follows sinc function. so
what is happening I am getting two advantages over here' one is that if I now employee, so this is
still from –b to +b, I employee a low pass filter, this has no distortion because this is just exactly
equivalent to G(f) . with that sinc multiplied.okay, and the other advantage I get, if I do original
sampling what was happening my power was getting distributed everywhere. right where as here
the central lobe which is the lobe of my interest is getting more power and the side lobes, which I

840
should reject anywhere are getting lesser power, because it has a sinc function. so the overall
strength of that follows a sinc function.

So that is the good thing because most of the power are concentrated on the low pass band, and I
will be employing filtering to take this. so most of the power I will be getting while demodulating
it. That is first thing, second thing is unlike the flat top sampling, it does not go through
distortion. these are two most important advantage of this one. and it has another advantage
which we can easily see that because the strength of is proportional to tp, so what is happening as
I increase my tp.this strength in the central lobe will increase. okay, that is a very good
thing.because what will happen as you increase tpi there is also another counter effect> this tp
will increase, so this will factor will increase, because it is a sinc, this as you increase t p, this will
go to 0 very soon in the side lobe.

So basically the side lobes are getting decreased even more, so what is happening as I increase
tp,I will be getting more and more power over here, and let us try to see what happens if I
increase tp. suppose if I have a signal like this, I want to sample it, so I sample it with these
pulses. let us say 50 % duty cycle. so whatever so this is so my tp is just ½ of this ts. so say this is
ts, so what will happen, it will just fallow the signal over this ½, and then it will be 0 'and then
again over this ½ it will follow this, next will be zero. so I have seen that as I increased t p of-
course I will be getting more power concentrated over here, which is very obvious because, if I
just make it full that means tp=ts, then the entire power because it will just look like this same
signal entire power will come to my central lobe, is that something I want, probably not, because
why against sampling?

we were doing sampling to get those central, to get those sample values and in between you are
freeing the time to do time division multiplication. so more tp we take, probably we will have less
opportunity multiplex anything else. so there is a trade of between these, which we can
immediately see. so probably with this we have a fair bit of understanding about natural and flat
of sampling.

These are the two most important sampling that are available practically, so next what we will do
we will try to see what kind of circuitry we can put for doing this natural sampling. okay, so in
the next class we will discuss little bit of circuitry of these modulation techniques and then we
will probably go to means digital version of this sampling, or pulse amplitude modulation this is
called PCM. Okay, Thank you.

841
 NPTEL

NPTEL ONLINE CERTIFICATION COURSE

Course
On
Analog Communication

by
Prof. Goutam Das
G S Sanyal School of Telecommunications
Indian Institute of Technology Kharagpur

Lecture-60: Pulse Coded Modulation

Okay so as we have already discussed about different kinds of sampling which is the major tool
for doing the pulse amplitude modulation. So what will, probably this will be our last class. what
will try to see is, how this can be realized in circuit.first of all, and then from here how do we go
towards our digitized version of this. okay so that is probably the linkage of analog
communication to digital communication. so this is with this only we will end our course.

Okay let us try to see what actually we employee as a circuit for a pulse amplitude model. let us
say natural sampling you want to do, because we have seen already some advantages of natural
sampling so what we can do.

(Refer Slide Time: 01: 20)

842
Let us take any switch okay so it can be a transistor switch or it can be a Mos switch whichever
switch you have this. so basically what you try do is you put your message signal mt over here. of
course it is having wires and a resistor over here, you also have a resistor over here, and over here
what you do you actually supply the pulse so this is where you will be supplying pulse. okay so
whenever the pulse is high that means it is on and that time this will be on and there will be a
current flow which will be proportional, because the message signal you are putting over here.

So that will proportional to your message signal so the message signal varies in time like this the
current also will be varying according to this in nature, and the voltage across this or this resistor,
so we will also have similar variation. but whenever it is off nothing will flow through it because
of the resistor or that switch will be off, and you will get me. so basically what will happen as a
output of this resistor what you will see whenever it is on it will follow the signal whenever it is
off nothing is getting flowed through it.

Again whenever it is on it is follow the signal and current so that is all natural sampling. with a
very simple circuit we can actually get natural sampling so that is also another advantage of
natural sampling that natural sampling can be realized with a very simple switch either a CMOS

843
switch or a transistor switch can be done very easily just supply the pulse supply the message
signal and automatically it could be modulated with that switcher okay.

So that is the realization of a pulse amplitude modulation. okay what we can also do is pulse
width and pulse position modulation. but probably for time savings we will skip those things
because of they are not much use and not that important. so what we will try to do in this class is
we try to see from PAM how do we go to the digitized version of this. okay so let us say I have
either a natural sampling okay, but a natural sampling become very small duration okay.

So that means it is almost equivalent to flattop sampling because the duration is so small that the
variation will not be it will be overlooked so it almost happens like a flattop sampling. so if this is
the pulse modulation I get, now what I do is this particular pulse I have got a amplitude at that
particular instance okay. Let us say my message signal it has a range amplitude range from + mp
to – mp.

So that is the overall range of amplitude variation. immediately I can also say that pulse
amplitude because it is unit, let us say it is the pulse is are of unit amplitude so whenever we
multiply it just takes the message and signals amplitude so basically even the pulse amplitude
also will vary in that range. so highest pulse amplitude can be +mp and the lowest pulse
amplitude can be –mp.

Okay so this is the range over which it varies so I de-mark that range so from + mp to – mp. now
what I want to do is I want to digitize this okay, so to digitize this means I want to have a unique
representation of each of this samples which can be represented by 2 bit, by binary representation
1 and 0. which is easier to encode and easier to transmit we will see the advantages of that also
okay.

So I want to digitize it what is the technique that I should employee. whenever I digitize let us
say I digitize with let us say 8 bit. okay so that means each of the sample represent with 8 number
of bits.with l bit how many representation I can get? that is something we know 2^8
representation. okay that is 256 representation. so if I do represent every sample with 8 bit then I
can only get 256 positive representation.

But within this +mp to –mp the signal is analog.so how many representations are there infinite
representation, every voltage is a positive representation. but I only have restricted number of
representation. so what I do is a very important part where you do from analog communication to
digital communication. or do employee a method which is called quantization. so what does that
means quantization means basically this range you take.

844
And you subdivide this range into multiple sub ranges. how many such things we will be taking?
exactly 256 such ranges. because you have to push to 256 representation . so you take small sub
range and any signal happens to be within this range will be encoded as the middle value. and it
will have a corresponding binary representation. or a 8 bit representation why I have taken 256
because I have with 8 bit 256 distinct representation. so if this happens to be within this range
entire range can be represented or with a unique binary code let us say all 1. so 8 ones like this is
can be represented with all 0 and so on.

Every interval can have a unique representation and what we do.whenever we represent it that is
all fine, that I get a signal which is in between somewhere, it can be anywhere. so I represent that
as the corresponding binary code. so this is some kind of modulation again then I am actually
modifying the same signal for a better representation. okay so or a digital representation. so this
some kind of modulation, and at the other end, I have to also demodulate it,or have to de-map it
decode it.

So for decoding what do I do? now I have this unique representation/whenever I get this unique
representation I will be able to put a corresponding sample value over there.but now the problem
starts because which sample value I put all these things from this all those sample values in
between intermediate sample values are mapped to this unique representation. so when I de- map
I have no information where exactly it was, because I do not have that information.

So what I will do I will actually put this central point as the representation. so immediately you
can see there is a error. because it might be this one and I represent with the central value so
basically the amplitude was this much I have represented with this. so what might happen.

(Refer Slide Time: 09:48)

845
Suppose my message signals where like this the samples are like this, and then I encoded with
binary, so anything falls under this so basically what happen I represent it with the middle value.
okay and then I joins the middle values, whereas actually it should be this tips, it should be
joined. there should be little bit of error due to this quantization. this is called quantization error.
okay so whenever we digitize the signal we will always be having a quantization error, and this
process of digitization is called pulse coded modulation.

Because what you are doing is taking a code/sorry taking a pulse and you are encoding it into a
digital or mapping it into a digital bit stream. okay that is why it is called pcm or pulse coded
modulation. okay why signal use this means even these days also the telephone that we use the
landline phone calls we make that goes through this. so basically you hear voice signal first of all
it is being sampled with.

Because the voice signal, we assume that it has maximum band width as 4 KHZ. so what will be
the corresponding might be straight that should be 8 K samples per second. each of this samples
are represented by 8 bit. so overall bit rate will be 8 into 8. so each of those samples are
represented by 8bits. so 8 into 8 which is 64 kb/sec.so that is why the voice is always being

846
transmitted with 64 kbps is the de-facto industrial standard which has happened that is because of
this pcm encoded thing which has 256 representation for quantisation.

Okay so whatever this is we have now seen that there is a quantisation error which is happening
due to this process. so can we characterize this quantisation. so let us try to see if we can do that.

(Refer Slide Time: 12:12)

m (kTs) sinc(2πBt − πk)

∑
m(t) =
k

̂ = m̂ (kTs) sinc(2πBt − πk)

∑
m(t)
k
̂ − m(t)
q(t) = m(t)

∑[ s ) − m(kTs )] sinc(2πBt − πk)

= ̂
m(kT
k

so let us see my message how it was, We had this interpolation relationship which tells me, that
this is actually nothing but, samples taken at particular instants and represented by a
corresponding sinc function. whereas my quantisation 1 will be almost similar but only thing is

847
that instead of taking this or we taking the middle value. so let us call that as m^. because I
cannot get all these values/this can be any arbitrary number, that means any arbitrary value
between –mp to + mp.

I will be taking all those quantized middle point. okay so that let us call that as m^ kts and that
should again, because these are creating a sample at the other end or we actually representing
again with the sinc function-so that I get my message back. So due to this process there are what
is that error. so let us say that is the quantisation error I that is by this m^(t) –m(t),which is
nothing but if I just represent it'k ∑ which of those sample error.

So M ^ (kts) –m(kts) sinc (2π8t-kπ). so this is my quantisation error. so this is the errors signal I
need to see how much power so this error signal is almost like an noise sitting over there. right.so
if I just try to see.

(Refer Slide Time: 14: 11)

This was my original signal right, the above one. and this is my after quantisation and
demodulation or means decoding this is what I get. so the error is actually on noise as if being
added on the signal to get my signal right.so that can be treated as noise, and how do you

848
characterize noise so this is something which can be termed as quantisation noise.okay so this is a
particular noise if you see whenever you go from analog communication to digital
communication this is a particular noise which is been created at the source.

So for all the noise that we have talked about are either generated at the transmitter circuit, but
which was very negligible, so we have never considered that. in the channel of-course there were
noise and in the receiver circuit. so mostly channel and receiver circuit that where the noise were
there.now this is the noise which is being created at the encoder. so whenever your quantizing
you are potentially creating noise.

So this is source induced noise. so your process of transmission is generally creating this noise.
okay so this noise how do I characterize the noise, we need to get the noise power.so basically I
want to evaluate the noise power.

(Refer Slide Time: 15:36)

849
 m (kTs) sinc(2πBt − πk)
∑
m(t) =
k

̂ = m̂ (kTs) sinc(2πBt − πk)

∑
m(t)
k
̂ − m(t)
q(t) = m(t)

∑[ s ) − m(kTs )] sinc(2πBt − πk)

= ̂
m(kT
k
q(kTs) = [m(kT
̂ s ) − m(kTs )]

So How do I get the noise power? So I square it and take an average. so in time domain if I just
write this it should be this, the way we have defined power. right, this is how we have defined
power. so this how we should be evaluating that. so if I now put the formula of qt. this was ∑k q
(kts) okay, so that m ^ kts-mkts we are terming it as q(kts) okay.

(Refer Slide Time: 16:29)

850
 1 T/2 2
T→∞ T ∫−T/2
˜2
q (t) = lim q (t)dt

2
1 T/2
T→∞ T ∫−T/2 [ ∑ ]
= lim q(kTs)sinc(2πBt − πk) dt
k
∞ 0 m≠n
∫−∞ { 2B m = n
sin c(2πBt − 2nπ)sin c(2πBt − nπ)dt = 1

1 T/2
T→∞ T ∫−T/2
˜2
q 2(kTs)sinc2(2πBt − πk) dt
∑
q (t) = lim
k
1 1
q 2(kTs)
T→∞ T ∑
= lim
k
2B

So if we write just q(kts) this is sinc (2π8t-kπ )^2 dt. now this is the homework what you can take
for your own benefit. so what happens we will be able to prove that this sinc functions are
mutually orthogonal for different values of k. that means, we can we will be able to show that
from – infinity to + infinity' if we multiply sinc(2π8t-mπ) sinc (2π8t-nπ)dt, so that is two signals
with different values of m and n.they will be orthogonal if m equal n, gives me some value and m
not equal n, gives me 0. so that is what happens.

It is 0 when m is not equal to 0, this is something which we will be able to prove and this is equal
to 1/2b when m=n. okay so these are orthogonal, so what will happen this squared term. basically
whenever you square it, there will be all square terms + 2 into this cross term, with different
values of k, like m and n. so all those terms will be 0. so only the square terms will be remaining.
so what we can write from here, it is nothing but as t tends to infinity. 1/T (- 1/T to 1/T). so just ∑
square term kts.

And this whenever m=n that is the squared term that should be 1/2b. okay and of-course there
should be, this means that sinc2 will come, then you integrate, basically you get 1/2. right so this
is what we get. of-course because we have already done that integration, with that integration, we
will be getting this because t tends to α is there so –t to + t goes to α 'and then that integration we
can evaluate so you get ½. I can write immediately limit t tends to α.

And then, I can, of come the integration, sorry, I have written little bit wrong fashion. So I should
have written this as 1/2bt sinc2 (2π8t-kπ)dt. okay I should have written this. And then, the
integration I can take inside, because this has nothing to do this the summation, and this has
nothing to do with that integration variable and limit.so that they can come out, so I can always

851
get 1/Iso this is equal to limit it tend as t tends to α ,1/T will be there, ∑ this term goes out, q^ 2kt
and then I will have this integration, because t tends to infinity, so it goes for – infinity + infinity,
this integration which happens to be 1/2 B. okay this is how I can write this and then.

(Refer Slide Time: 20:03)

What do I get iI get limit t tends to α ,1/2bt ∑k q^2(kts). what is this means?so basically I'm
taking all those quantisation values the square of them. let us for our duration of t how many such
values I should be getting? I will be exactly getting 2b into t. because 2b is the sampling rate.
okay, so power unit second I will be getting 2 B such samples.in t duration I will be getting 2bt.
what does this means actually, this means the average value of this quantization.

So q^2(kts) it is the average value of that. okay so whatever those values are, at different k I need
to take the that means the mean of the square or the average of that. because I am actually
summing all of them divided by as many samples are there in t. and then t is stretched to α, means
I take enough number of samples. so that it is actual average. so basically I just have to do
averaging of that. That should be my quantization noises. So this is something we have to proven
now.

okay so now let us try to see how do I get this average. so let us at this point will be doing two
things. one is we first assume which is also a valid assumption that this q thing that is of-course
obvious, that will be a random process. so basically what is happening it is a discrete time
random process. so it has values at different time instants.so this is the time description.

But we also have a enssembled description of this one. okay so there might be infinite number of
such sample function which has different kind of variation. and we say that this is a stationary
random process as well as ergodic. okay we are true but people have tested that is true always.
okay so if that is the case the ensemble average will be the time average.

So we were try to calculate the time average that it must be the ensemble average and now you
will make further assumption, that those samples actual samples of m is whenever we take m
(kts) those samples are informally falling between –mp to + mp. so they are uniformly
distributed. so a sample which are restricted over a particular duration, let us say the duration is
∆b, because the error will be only on that.

It means he is always, any sample within this will be approximated by the central value. so any
error that can actually percolate that will be restricted to the boundary of that particular zone.
okay let us say that is ∆v. what is ∆v? then ∆v is equal to overall range is from +mp to –mp. so
that is 2mp/ as many number of levels I will be creating. so let us say that is Lifor our case our
example L was 256. so 2mp / L is ∆v.

852
And what happens within this we are now assuming that the samples are uniformly distributed.
that is what our assumption is the samples are uniformly distributed, what is the corresponding
error? because will be any sample we get we will be representing by the middle value. okay so
the corresponding duration will be from the middle value to whatever it is. so let us say if it is
uniformly distributed in the picture it should be 1/∆ v going from -∆ v/2 to +∆ v/2. because the
middle point is the central point. so what is happening, this is 0 this -∆ v/2 and this is +∆ v/2.

So this is that whole range which is being represented by this. okay within this the sample might
lie uniformly distributed so that is why 1/∆b is its Pdf and then I have to take q^2 I want to take
the average of that, whatever that value of q is the deviation is the error. so q^2dqi have to do. pdf
I have already taken out. so this it will evaluate what we will get it/will be just it means, integrate
it and then put ∆ V so we will be getting ∆ v^2/12. that is the quantization.

Because I wanted to evaluate the mean of the square of it. so that is what we are doing, mean of
the square of it. so mean of the square of it is square, integrate over its range, which is the pdf
range which is from -∆ v/+∆ v/2 and 1/∆ v is it's constant pdf. so I get this. now what is ∆ V, I can
put that should be (2mp/l)^2 divided by 12. or mp^2/3L^2. this is the noise. this is called the
quantisation noise of a signal okay.

(Refer Slide Time: 26:00)

853
 ̂ = m(t) + q(t)
m(t)
S =˜
0m 2(t)
S0 3L 2˜
m 2(t)
=
N0 mp2
L = 2n
n = log2 n

So now suppose I have transmitted a message signal mt, along with that I have a quantisation
noise which is qt and I get this noise contaminated signal m^,because you are actually receiving
those m^ and we are connecting them only. we are getting this much error. this was the error
signal. so immediately what will be my s 0, let us say that is the message signals power. so what
will be my signal to noise ratio that should be the message signals power divided by the noise.

Which is mt^2 into 3L^2. right, so this is how signal to noise to noise ratio is calculated in our
typical this kind of pcm transmitter receiver. okay now let us try to characterize it little bit so
what is exactly happening is something like this. remember we if we had because it is a binary

854
representation so if we have 8bit, how many representation you are getting or what was the L
value that was actually 256.

So if we have in general n bits for representation, we will be getting 2^n =L at many

representation. okay alright or we can write that L = 2^n or n = log2 L relationship we know
already. okay now let us also try to see what is happening to our overall band width. okay so we
already know that every hertz, I can transmit 2 bits. okay so basically I had B that is the overall
band width I could eventually transmit 2B samples okay.

And overall, how much information I am getting. so for this B band width because of multiple
sampling 2B samples I am generating, for that is sample rate. each of those samples are now
represented by n bit. so this many bits Iam representing. okay so to transmit this how much
bandwidth I required? this many bits per second so divided by 2 will be the among the bandwidth
I required. so b* n hertz band width I require. so that becomes by effective transmission
bandwidth, Bt.

(Refer Slide Time: 28:49)

855
 BT
n=
B
BT
L = 2n = 2 B
S0 ˜
m 2(t)
= 3L 2
N0 mp2
= cL 2 = c(2)(2n) = c22BT /B

Okay, so basically I can write that my this value of n is nothing but, my effective bandwidth,
transmission bandwidth divided by B. So this is the bandwidth that I have to take to transmit
these. Became I have converted them into bits and then i am transmitting thou bits. okay. so this
is my effective bandwidth which be occupying the channel and this is the base. that means actual
signal bandwidth. okay, so this what I require. now i f I go back to the relationship between L and
m so l is 2^n or I can write 2 ^ BT/B. I can wite this way.what happens to my signal to my noise
ratio, we have already represented signal to noise ratio as 3l^2 m^2t /mp^2. this 3 and these
things if you just make them constant.

Suppose signal I put a particular fixed value of power mp the variation remains the same. so
basically it happens to be just some cL^2. now l if we replace by this we get C, L means 2' n,
c(2)^2n. or I can write c(2) to the power 2 BT divided by B.okay, so something we get. that is the
signal to noise ratio. now if I represent the signal to noise ratio.in db what I have to do I will have
take 10 log signal noise ratio in db.

So or we can write 10 log 10 c 2 ^ n, okay sorry 2 ^ 2 nor I can write log because it is log is to be
distributed so log10c +c is a the constant and log10 2 into an. right, a log 10 n2 is just 3. so this
happens to be, sorry log10 2 into 10 that is this particular part I can write as α log 10 c because it
is a constant, into 10. that is α and here we can write 10 log 10 2 that is 3, * 2 that is 6 n. then I
get my signal to noise ratio in db as this.

okay, n can be further written as α into 6, n is, I have already seen that n is Bt /b. and I have got
as wonderful thing over here. probably realized, what has happened? you can see, that as I
increase my Bt, what it does that means? I am actually taking more samples. So suppose I go
increase by n from 8 to 9, immediately what will be the band with increment ? so here the band
width, if it is 8 it should be for voice channel, it should be 32kbps or 32 hertz the kilo hertz okay.

So if we just calculate, the way I have calculated. okay 64Kbps that should be because every half
hertz gives you one bit. so 32 kilo hertz. so that is the band width. if from their If I go to 9, I get
36 kilo hertz. so only four kilo hertz of increment of the band width. but signal to noise ratio what
is happening, because n is going from 8 to 9 I get a 6 dB increment.

856
So a quadruple increment of my signal noise ratio. so basically here you can see, like fm I
increase my band I get a better noise quality or better noise cancellation quality or signal to noise
ratio. but the advantage, is in fm we have proven that if you double the band width then you get a
6 dB increment or what quadrupled your signal to noise ratio.

Whereas here, you just increase by 4 kilo hertz. that is not doubling the band width but you still
get that kind of benefit. so basically this particular thing as a huge potential towards the noise
cancellation. you might be asking okay we have just taken the source noise. what is happening to
the channel noise. so of course that is true there will be channel noise also over here but that is
where the digital communication gets advantage.

(Refer Slide Time: 34:01)

857
( N0 ) ( N0 )
S0 S0
= 10 log10
dB
= 10 log10 c22n
= 10 [log10 c + 2n log10 2]

( N0 )
S0
= α + 6n
dB
BT
=α+6
B

So what happens in digital communication whenever we are communicating the digital signal. so
let say I communicate like this it is all 1 and 0. I know that might be 1, it might be 0so that is
what digital communication comes into the picture.suppose this signal goes through the channel
which are low pass characteristics, it has an attenuation, plus noise gets added. so due to low pass
characteristics it will remain like this.

And then on top of that there will be some noise added. and the signal will be look like this. so all
those nice patterns of 1 and 0 actually goes away. but the good part is suppose, you. from here,
you start integrating it, what will happen? because of noises sitting on the top of this and the
noise mean is actually 0.

Because generally, all the noises that we have seen so far these are all 0 mean noise signal with
some standard deviation. so its because you are integrating it, what you need is just this bit. all
unlike analog communication, every location what is the value, that is not important. It is just, we
have to decide within this duration, is it 1 or 0. that is good enough for digital communication. so
do you start integrating it. so these are the things which you learning in the digital
communication, I am just giving you example so that you can understand.

So once we integrate, the noise effects gets cancelled, because noise will be 0 mean. so if it is +
or – , once you integrate the noise effect will be cancelled and because of the signal even though
it is low pass filtered and attenuated, it will still integrate. and then you can put a threshold if it is
bigger than that threshold you declared it to be one, if it is less than that threshold, you declared it
be 0. so that way you will have a very cleaner discussion of 0 and 1 and what you can also do,
you can actually regenerate in between. so before the signal degrades to a limit or to a value from
you cannot really again reconstruct the signal.

858
So what you do in between you put regenerator which does these things. tell you that okay, this
was 1, this was 0. reshape the pulse. so again it will reconstruct the pulses because he knows once
it is 0 then it would look like this, once it is 1 it would look like this. and again retransmit it. so
basically he makes the pulse again very nice and transmitted so that you can come back the noise,
attenuation and all other effects of channel. so that is why digital communication is generally
prone, sorry, which it is much more effective in terms of noise cancellation and all other things.
and we can if we employ and there are other techniques also to basically if a bit is getting
erroneous, you have other techniques to basically checked whether it is erroneous or nots you can
correct it, and all those things. so employing all those things you can almost negate the noise in
the channel, or the receiver. okay.

Then the quantization noise becomes the more significant one. and in quantization noise we have
now shown at we have a huge benefit compared to all other modulation schemes. and that is the
sole reason why all pulse coded modulation started becoming very popular. okay, compared to
analog part. so I think we have almost discussed enough it analog communication, almost came
to the door step or the door way or digital communication. so now onwards probably you can
start discussing about.

How digital communication becomes effective and what are the techniques employed. how do
you to noise analysis? what are the signal analysis that you employ digital communication ? those
things can be explored further. with that I will probably end this course and means from this
onward you can take it forward towards more advance communication techniques. Okay, thank
you.

859
 THIS BOOK
IS NOT FOR
SALE
NOR COMMERCIAL USE

(044) 2257 5905/08

nptel.ac.in
swayam.gov.in