Welcome to the first lecture of the second
week of this course. Today we are going to discuss something called conditional expectation.
Conditional expectation is the first most important thing, that I need to discuss with
you. The remaining part, that I was discussing with you is something I was, expecting that
you had almost some idea about. But here we start with something very new usually not
done in standard courses of probability. Standard courses mean courses which engineers usually
take or economists take, etc. So, the exposition that I found most interesting
was from a book called Probability for Finance by Sean Dineen published by American Math
Society in 2005 under the series Graduate Texts in Mathematics. So, what does conditional
expectation mean? We will see, a very simple example. Suppose, I take this S is a random
variable or S is a random variable, which tells me, how many times head appear, when
I toss a coin three times. So, how many heads in three tosses of a fair coin, repeated of
course each, first toss, second toss, third toss, three repeated tosses of a fair coin. Of course, you know this is a binomial random
variable. We have already know, we know that, if N be the number of success and Np is the
expectation, given that p is the probability of success, sorry, how many heads, so just
make a little correction, how many heads in, three repeated tosses of a fair coin. So,
it tells you that how many successes you have. So, by success, I am meaning the appearance
of a head. So, I toss any of the three can come; head, head, head, head, tail, head and
all these things. So, you know that expectation of this because,
this is a binomial random variable, is nothing but n into p where probability of success
is half and n is 3 so 3 into half which is 1.5. What happens if I now ask you this question,
find the expectation of S3? find the expected number of heads that you expect if you make
three repeated tosses of a coin, but with the knowledge now, that the first toss is
a head? So, now we are going to see how many tosses,
you have three repeated tosses, so how many heads would appear if you already know, how
many heads you are expecting rather, if you already know that the first toss is a head.
What is the answer. This is interesting. You will see now I know that the first toss
is a head. I really have to look into the second two positions, this is given. So, here
I can have head, head; head, tail; tail, head; and tail, tail. So, means, now I can have
only 1 head appearing remaining two are tail. I can have only two heads appearing there
is no tail at all so this one comes all here head comes and tail comes then tail comes
and head comes. So, if S3 is 1, if S3 can take value 1 it
can take value 2 it can take value 3. It can never take value 0 because, I know the first
one is 0 right. So, when I am talking about 1 means these two tail has come. So here among
this four, now I do not bother about it, it is already a known fact so I just have to
look into this diagram. So, when tail comes both are tail so among these four choices
I have only one choice, so there my probability is one fourth. So, 1 into one fourth, right. Now if, I have the situation, where I have
2 heads which has come in three so among these two, one has been head; so, either head, tail;
tail, head. So, any one of the two should appear with probability of the 4, this will
appear two times, so the probability would be 1/2. You see the whole working changes.
Now 3 means I have got this. So, among all the 4 I have one chance. So, it is one fourth
plus one plus three fourth. So, it gives me one, so does this answer sorry 2, 1 plus 1
2, 1 plus 1 2, sorry this answer is 2. So, I leave you with the to show that, leave
you with this question, that if the first toss is tail then what is my expected occurrence
of heads? So, that would decrease. So, here my expectation increases. When the first one
is head, I am expecting that among the 2, 3 at least 2 would be head. Here because the
first one is tail I am expecting at least among the 1 there will be 1 which will be
head. So, you see once more information is available my expectation changes. This is
called conditional expectation essentially. Now, how can I put this idea into a more rigorous
form. See, this idea is being developed on a scenario, where your omega is finite, your
sample space has finite elements. So, your Sigma-algebra or the set of all events, is
the power set of that finite set. So, let us look into this scenario. So, we have, instead of u I am writing F because,
that is what is usually written in finance books. So, let this be a probability space.
So, what is more interesting to know here, is that I take that this is finite. The cardinality,
when total number of elements of omega is finite, that is so my F is nothing but the
power set. So, F is the power set of omega. Now, assume that strictly greater than 0 for
all omega in the sample space. So, every sample point, every outcome of random experiment
the probability is strictly greater than 0, just like a head or tail scenario. Let
A be an event, such that probability of Aa is strictly bigger than and we can take it
less than 1 or strictly less than or equal to 1 does not matter much. So, this is what
I have assumed and under this assumption I have to define what is called a conditional
expectation. So, I am looking for the conditional expectation
of a random variable X, given that the event A has occurred. This is in some sense conditioning
using a Sigma-algebra because A belongs to a Sigma-algebra and soon we will come to that,
how we actually condition over a Sigma-algebra. This is nothing but the very definition, it
will be X of omega and probability of omega given the conditional probability of omega
given that the event A has occurred. This is exactly what will happen. This is exactly
the definition. So, what is probability omega intersection A. That
is probability of omega by probability of A,
if omega is in A and is 0 if, omega is not in A. This is quite simple to understand because,
you see what happens. When omega is in A, omega intersection A is omega, so you have
just P of omega by P of A, and when omega is not in A, omega intersection A is the null
event, empty, impossible event, so that time this will become 0 because, this will be the
null event. So now coming to this board again. I am rubbing
off what is in here. That is why sliding boards are so important and this place has space
for sliding boards actually. So, now I can write down the definition by putting in this
in this place. So, what I finally get is expectation X given that the event A has occurred is summation
omega element of A, X omega, P of omega divided by P of A, and this I can write as, 1 by P
of A and symbolically, you know that this sum is actually an integral if you were not
considering, the discrete distribution or discrete setup, then this would be an integral. So, essentially a sum, summation X omega,
P omega omega element of A but this we will always symbolically even if we have sum we
would symbolically write this as. So, this is just a symbolical writing for the expression
sorry is just a symbolical writing for the expression X Omega, P of omega, symbolical
writing. Of course, you can move to non-finite spaces,
which we will soon do, but this is just a useful trick to understand it. So, if you
want to write about, any other set, what about the compliment, if I know the compliment of
A has occurred. Then also, this is nothing but, the same you can actually prove that,
they are you can just replace A with. Now let us look at the Sigma-algebra which
the event A can generate, that is the smallest Sigma-algebra which contains the set A, it
should contain A, it should contain the empty set, it should contain the whole set, it should
contain its compliment. So, Sigma-algebra generated by A the empty set, the whole set
A and A compliment. So, if B, so we are just writing it B is element
of A so some sigma element B from this Sigma-algebra with element P B greater than 0, so we can
write. So now observe that, whenever B occurs, what does this means B occurs, some omega
belonged to B has occurred. When you have random experiment, what is the meaning of
B has occurred. When I say, I throw a dice, odd number has occurred. If 1 occurs odd comes
over the face, then the event that odd number has occurred has occurred. So, I can actually view this as a function.
So, for any omega that I take in B, I can define a function like this, rather a random
variable like this. So, for every omega in B, that is over the whole set B, whole event
B, this is the value. So, it remains constant over the whole event B, so if I look at it
like that, that I am looking at a function which is constant over various events, then
the conditional expectation itself can be viewed as a random variable. So, you see because it does not matter whatever,
whenever any omega occurs, B has occurred actually. So, for any omega in B this should
be the story. So, whatever once you know that, B has occurred
so for whatever omega you are taking this should be the story, this should be the answer.
So, this idea, that you can actually view, the conditional expectation itself as a random
variable is a very-very fundamental idea and is a very helpful idea. So, this is itself
a random variable r.v, where r.v. is just a short shorthand for random variables. So,
this will allow us to make shift from the standard probability space where we do not
bother about the finiteness of omegas and all those sort of things. So, first we will talk about a countable partition.
Now, what is the meaning of a countable partition of omega. So, consider a sequence of sets,
consider Gi, i is equal to 1 to infinity right, 1 2 3 by G1 G2 Gn so consider Gi where G i
is a subset of omega for all i in N set of natural numbers. Then this Gi is called a
countable partition, if number 1, Gi intersection Gj is equal to phi for all i not equal to
j, and number 2, union of Gi, i is equal to 1 to infinity, should give me back omega the
whole sample space. So, this is called a countable partition and
what we would now require, is a Sigma-algebra generated by that countable partition that
you take elements of that countable partition, when I basically, take, do not take intersections,
because intersections would ultimately generate either the, if you take intersection of the
same set it will generate the same set itself or it will generate the empty set. So basically, if you want to take any non-empty
set of a Sigma-algebra generated by the set, you just have to construct unions of this
set using some subset of n could be finite could be infinite whatever right. So, what we will be interested is, in a Sigma-algebra
generated by a countable partition, which will be a sub Sigma-algebra of F, naturally
because, we are taking a subset, when taking a subset of A and generating a Sigma-algebra. So, consider a Sigma-algebra G, subset of
f and G, G being generated by a countable partition. Then, we define it like this, and
the conditional expectation of the random variable X, conditioned on the Sigma-algebra
G, is defined as like this, at any omega right, say omega n or whatever, any omega is P Gn
integral Gn X dP because this is a countable because this is a generated by a countable
partition say G i. So, given any omega it must lie in one of
the Gi s because omega is the element of omega and this is a countable partition of omega
so it must lie in one of the Gi s. So, suppose take any omega then this will happen. This
is the way I define how you compute this random variable at the point omega if omega is element
of G n, good. I will now, been a random variable, it is
important to know that because I am claiming it to be a random variable, you should be
able to prove that, I leave this as an exercise to you, in your assignments, so of course
and this is true. You are going to prove this fact that this is measurable.
An interesting thing is that if x is integrable random variable, that is it has a finite expectation
then this random variable also has a finite expectation that is the, conditional expectation,
conditioned on a sub Sigma-algebra of F which is generated by a countable partition, has
also got to be integrable. That can be proved, but I would not rather prove this fact now. So, this would be a part of the exercise.
If X is integrable, I am expecting that, you know what is integrability though we have
given some basic definitions. So, if X is integrable then, integrable means it has a
finite expectation because expectation is expressed in terms of an integral, that is
integral X dP this is finite, then E is integrable. That is this random variable also has a finite
expectation. This is important. Now, once we know this, we are now going to
state a very important property, of this random variable. This property essentially characterizes
and this idea actually allows us to move beyond Sigma-algebras, which are Sigma-algebras of
this type, that is you will now, once you know this you can, once we know this result
what we are going to state, using this idea, we can give a very general definition of a
random, conditional expectation as a random variable that is we can define the conditional
expectation, conditioned not just by a Sigma-algebra generated by a countable partition, but by
any Sigma-algebra which is a sub Sigma-algebra of F. So, you are writing this important property,
give a very small proof of this and then we will wind up today’s discussion. So, this
section is the first part of conditional expectation and tomorrow we are going to discuss the general
case and the properties of conditional expectation that is it. So, what does this proposition say? So again,
G is a sub Sigma-algebra of F, generated by a countable partition, Gi 1 to infinity. Assume,
of course this definition when you are writing, you are assuming that P Gn is strictly bigger
than 0 for all N because without that you cannot write this. Assume that P Gn that is probability of each
of these pieces of the partition is greater than equal to 0 for all N. Then E X by G this
conditional expectation is the unique G measurable, of course when I am talking
about measurability here, I am essentially talking about G measurability, is the unique
G measurable random variable such that, random variable on the probability
space of course, such that for all A in G. So, on G, integrating over a set of G, subset
of G, integrating over any element in G, any set which belongs to G, over X is same as,
integrating over the conditional expectation. So, it is essentially it says that, over G
conditional expectation and X are almost the same thing. Over such as Sigma-algebra, X
and conditional expectations, they behave in a almost similar fashion. The integrals
are same and this idea is actually used to extend, to a higher basically, when you go
to a case when you do not take this countable partition you take this as a definition of
condition expectation. You see this is how mathematics develops because,
you do it for some simple case and then you know that, it is not so easy to talk about
the general case, so you pull it up by taking the, what you know for the simple case as
a general definition and that definition will always work if the case is simple. So, we
will come to that later on. It has lot of links with certain things called Radon-Nikodym
theorem. So, let us give a proof of why this is happening. I am not going to prove, the case for uniqueness
because, that is very simple. So, to prove the uniqueness, we really have to use take
any other Y, which for this the same thing happens, which will tell me that integral
of this variable minus Y dP would be 0. So, which means that if this function is 0, what
does that mean, function is 0 almost, that is exactly. So, that is the how you receive
your uniqueness. So, we will just do the existence of this that okay if I define the random variable
in this way the way which we have defined in this particular case then this must be
equal to this. So, proof and with that we will end today’s discussion. So, I am just taking with this first with
the Gn, let us see why we are doing so. First, you have to understand why I am just taking
first with Gn because, if A is any non-empty set in G, then A is always written as, union
M subset of N Gn where M is a subset of N, could be finite could be infinite whatever. So, take some Gn from there and you are writing
this Gn, M is capital M is a set. It contains the indexes which is subset could be just
1 2, 3, 4, 5 so any subset of M. So, every A can be expressed like this right because
the intersection you could take intersection of the empty right that is why you cannot.
So, every non-empty set has to be expressed like this okay. Let us see what happens on this Gn. Now you
observe, that on this Gn for every part, if you look at the definition this is a fixed
quantity. You take any omega n element of Gn, whatever omega you take it will give you
the same value. It is a constant on that Gn part. So, I can take to be some omega integral
Gn dP, that omega is element of G n. Now what is this. This is nothing but probability
of Gn but I can again use this, I will write down the definition here, probability of Gn
integral Gn X dP probability of Gn because this is nothing but probability of Gn integral
d P Gn is probability of Gn. So, what I will get here is integral Gn X dP. Now if I am talking about integral A, so I
can write this as integral union of Gn M subset of N because, these are all disjoint, I can
actually sum them up sum each of the integrals. So, it is summation M subset of N and summation
of all the indexes of N integral Gn and this summation, what is this, for each individual
piece this is nothing but this. So, this is nothing but summation M subset of N integral
Gn X dP, which is nothing but, integral union of Gn, union of Gn is nothing but A which
is integral A X dP and that is the answer that is the proof. So, with this proof we end our discussion
here today. Thank you very much.