Lecture : 11 Measurable functions

IIT Kanpur July 201
20 followers
Dec 2, 2023
Welcome to this lecture.
This is the first lecture of week 3.
Before we proceed, let us first quickly recall what we have done in week 1 and 2.
In week 1, we looked at the collection of events corresponding to a random experiment
or more generally certain collections of special subsets of a general non-empty set.
And we configured or we looked at certain structures like fields, sigma fields and monotone
classes.
We also studied the properties of these classes or collections of sets.
In week 2, we entered this area of measures.
We introduced this topic and the idea was that we wanted to understand which events
were more likely to occur or which sets are larger in size.
To understand this, we assigned these numerical values to these sets.
We obtained set functions.
But since our motivation is to look at these sizes of sets or likeness or appearance of
certain events, we looked at these special structures which we called as measures.
We studied these algebraic properties, continuity properties of measures and in particular,
we restricted our attention to probability measures because this is what we are interested
in.
So these are what we have covered under the materials in week 1 and 2.
So as before, let us switch over to the lecture slides and start the discussion for the materials
in week 3.
So this is the motivation to look at certain type of functions on sample spaces.
So in many random experiments, the outcomes are numerical and as we are interested in
these numerical outcomes, numerical values, numerical features, we need to consider these
functions defined on the sample space omega.
So we are looking at functions for each point or each sample point in the sample space.
So we are looking at these values assigned to each sample point and we are looking at
these collections of functions.
On the other hand, there are certain random experiments where the outcomes are non-numerical.
For example, in tossing a coin, the outcomes are either a head or a tail.
So now you will ask what is the value, but in practice, you associate certain values
to these heads or tail, you assign certain meaning.
For example, when you toss a coin, you might be calling heads and if head appears, you
win the toss.
So there is this notion of winning a toss and losing a toss.
So therefore, there is a physical meaning associated to these outcomes which are heads
and tails.
So just to understand this better, it is sometimes useful to associate certain values to these
non-numerical outcomes like head or tails.
So what you can do?
You can associate the values plus one and minus one to head and tail and you can think
of this plus one and minus one as winning ones or losing ones or winning one rupee and
losing one rupee.
So you can interpret it in this fashion and therefore, now the interpretation is much
more quantitative, it is in terms of certain values.
So this is what we are going to restrict our attention to.
So no matter what kind of outcomes we have, we will typically associate certain meaning
to them which will be numerical and we are going to study those values and get some more
intricate features or intricate analysis of the corresponding random experiment.
So in regards to this discussion that we are having, what we are going to consider are
functions of the form f with domain sample space omega and the range space real line.
So we are looking at real valued functions on the sample space and to do this analysis
that we really want to do, we require more structure on this function.
So you cannot just start looking at this arbitrary type of functions which are real valued and
defined on the sample space.
You need certain nice regularity structures on these functions.
So this is what we are going to discuss in this lecture.
So this is in terms of certain measurability structures associated to the domain space
and the range space.
So what we are going to do, so since omega is the sample space of a random experiment,
it comes with a collection of events which we have already associated the structure of
sigma field.
So on the domain side, you have the structure of a measurable space, so omega script f.
So this you can think of as the sample space together with the collection of events.
On the other side, the range side, you have the real values and on the real line, we have
discussed extensively about this sigma field which is the Borel sigma field.
So what we are going to focus on are certain structures with this information that you
have already collected like these sigma fields scripted f on the domain side and the Borel
sigma field on the range side.
So with this in hand, we are going to look at certain specific structures of these functions
f.
We are not going to look at arbitrary functions f.
So recall from note 6 of week 1 that for a function which we can consider now a general
function mapping from domain space omega 1 to a range space omega 2.
So omega 1 and omega 2 are some non-empty sets and you look at an arbitrary function
defined on top of that.
Now if you fix a subset A of omega 2, the range side, then you can define the preimage
of this set under the function small f.
This we have defined in note 6 of week 1.
So what was this?
Let's quickly recall that.
So we looked at this f inverse of A defined by all points in the domain.
So now the domain is capital omega 1.
So you are choosing all points or some all points omega 1 coming from the domain side
such that the value associated to those domain points, the f of omega 1, the associated value
lands up in the set capital A. So f inverse of A that is nothing but the preimage of A
which is defined as all points in the domain side such that the values of the function
on those points lands up in the set capital A.
So remember f inverse here does not represent the inverse of f because a priori whenever
you are taking a function, it's not really invertible.
So you don't know whether the function is 1 to 1 on 2.
You don't know anything like that.
So here you should treat f inverse of A as one symbol instead of treating it as two separate
symbols like f inverse and A. So you should think of it as one big symbol signifying this
preimage for A.
So now here f inverse you can now think of it as a function that associates any subset
on the range side to its preimage or this is now connecting or associating sets from
the range side to the domain side.
So this you can think of as a function in that sense.
But this is not the inverse function of the function f.
Now if the function f has an inverse, then of course you can connect the notion of this
inverse with the actual inverse of a function.
That you can try to work it out.
But now what we are interested in are properties of these preimages.
So the properties of preimages comes in this form.
So take this general function which starts in the domain omega 1, lands up in the range
omega 2 and choose a collection of subsets from the range side.
So ai, i belongs to lambda.
So where lambda is some index set.
So here it's some arbitrary set.
It could be finite, countable, uncountable, anything.
So it's just some arbitrary index set.
So you are looking at a collection of subsets of omega 2.
You are not associating any structure of any type.
So function is general, a collection of sets is general.
Now what we are discussing now are properties of the preimages of the sets ai under the
function f.
So what are the properties?
So the first property says no matter which set you take, if you look at the complement
of the preimage, so f inverse of ai that's a set on the domain side.
If you look at the complement, that can be again identified as a preimage of another
set and that turns out to be the preimage of ai complement.
So this is a nice fact.
You can try to prove this.
Similarly, there are these nice properties that connects the unions and intersections
over these indexing sets.
So here we are looking at that same index set lambda.
This is some arbitrary index set and it simply says that whether you take unions or intersections
of the preimages, these unions or intersections can simply go inside.
So by that I mean, if you are looking at the unions of all preimages, it's nothing
but the preimage of the unions of ai and intersections you have a similar result.
So try to prove this.
So these are some very nice useful observations about preimage of a set under the function
f.
Now, with this in hand, we are now focused at the preimages of sets coming from the range
side and we are going to look at certain properties which restrict the collections
of functions on the sample spaces and with this we now arrive at the first definition.
So we are going to talk about measurable functions.
Let omega 1 f1 and omega 2 f2 be two measurable spaces.
So I take two measurable spaces on domains omega 1 and omega 2 and I already need to
have sigma fields on top of them.
So I have script f1 and script f2, the appropriate sigma fields on these domains.
Now I look at a general function f from omega 1 to omega 2.
So with domain omega 1, range omega 2.
So we say that this function is f1 over f2, script f1 over script f2 measurable if for
all preimages we land up in the domain side.
So you take all subsets that are on the range side sigma field which is the script f2.
So you take all such subsets which are in the range side sigma field.
Look at the preimages, those preimages I want them to be on the domain side sigma field.
So this is the restriction we impose here and if it so happens for a function f defined
on a domain omega 1 to the range omega 2 with this associated sigma field script f1,
script f2, if it so happens for such a function then I am going to call it script f1 over
script f2 measurable.
So this is the definition, this is the restriction that we impose on a general function.
So if it happens then we introduce this concept of measurability.
Now in many situations there are standard choices of sigma fields.
So for example you can start with a finite sample space and then you see that the power
set is a natural example of a sigma field there.
On the real line we have extensively discussed this Borel sigma field, these are certain
natural choices.
So once you have this natural or standard choices you typically like to work on only
this type of sigma fields.
So you will put this natural or standard choices on the domain side and the range side as appropriate.
If in any discussion the sigma fields are kept fixed then for simplicity we may simply
say that f is measurable instead of saying f is script f1 over script f2 measurable.
So two points to note, first thing is that there may be standard choices for sigma fields
so we will use that but once you have this standard choices the sigma fields are now
understood from the context.
Once you have this standard choices the sigma fields are understood from the context then
for simplicity of the notation you will simply say that f is measurable instead of explicitly
mentioning this sigma field script f1 script f2.
So for example on any of these sides on the domain side or the range side if the real
set of real numbers appear and you are going to use the Borel sigma field then you can
ignore that and do not mention it just say that f is measurable this will be understood
from the context that you are using the Borel sigma field.
So this is the simplicity of the notation.
Now we focus our attention to the specific case when the range side is the real line
so as discussed above.
So when you are looking at certain random experiments with numerical outcomes you exactly
get this type of a function so each sample point gets associated a real number otherwise
if the random experiment appears with non-numerical outcomes you will associate certain meaning
or values to this numerical values and therefore obtain these type of functions such that it
is defined on the domain omega and the range side is the real line.
So as motivated like that we are going to focus our attention to the case of functions
which are defined on domain omega the sample space and the range side is the real line.
So in the next definition this is explained in more concrete detail.
So we are looking at specific type of measurable functions now.
So what is this?
So now the domain side is fixed that is omega script f so that is some measurable space
that is given to you it may be coming from a random experiment or maybe you can look
at some general non-empty set together with some appropriate choice of a sigma field.
So that gives you a measurable space and that you are going to put on the domain side.
On the range side as discussed we are going to look at the real line and we will be considering
the Borel sigma field on top of the real line.
So that will give you the appropriate measurable space on the range side.
Now look at the measurable functions which comes out from these conventions from these
choices of measurable spaces.
We say that the function f which is defined on the domain omega with values in the real
line to be script f over Br the Borel sigma field measurable or simply Borel measurable
if the same condition that you take now all Borel sets coming from the real line.
So you are looking at all Borel subsets of the real line which is on the range side you
are looking at all subsets looking at the pre-images now.
The pre-images are subsets on the domain side and you want those subsets to be in your collection
script f.
If it so happens you are going to call this function small f as script f over Br measurable
or for simplicity of notation you will simply say its Borel measurable.
So you are going to talk about Borel measurable functions on domains omega and the idea is
this that once the range is real line and you are talking about the Borel sigma field
you don't have to mention it explicitly once more its understood from the context.
So you are looking at for all sets on the range side all Borel subsets on the range
side you are looking at their pre-images these pre-images must fall in the domain side collection
the sigma field script f.
So if this happens you are going to call it a Borel measurable function.
So notice that before defining this measurability concepts we are fixing the sigma fields.
So this you have mentioned earlier that we have also talked about this standard choices
of sigma fields and other things.
Now we have always fixed the sigma fields on the domain and the range first before defining
the measurability.
What is the reason for this?
The reason is this if you change the sigma fields on either the range side or the domain
side you will be getting different types of Borel measurable functions.
So if you make a specific choice of sigma fields on the domain side and the range side
you will get some type of functions which will turn out to be measurable with respect
to those sigma fields.
But then if you change the sigma fields either on the domain side or the range side you might
get other functions which will now turn out to be Borel measurable or measurable and you
might get certain functions which were measurable in the earlier sense but no more for the new
sigma fields.
So you have to be careful with these choices of sigma fields.
Once you make these choices before your discussion you always make these choices keep it fixed
and then continue on the discussion.
So we are going to do this.
We are also going to use this shorthand notation or the statement.
What does it read?
It reads as follows small f maps from omega scripted f RBR.
What does it mean?
You are explicitly stating the domain and the range which is that the domain is omega
range is R but you are also putting in the sigma fields appropriate sigma fields on the
domain side and the range side.
So just to be clear the function the given function is purely a function defined on the
domain.
So for each point in the domain omega you are associating a real number.
So that is how the function is defined.
But whenever you are talking about the measurability you have to mention the appropriate sigma
fields.
So what you do whenever you are talking about this measurability you explicitly state that.
So that this notation is now explaining these choices.
You simply write f from omega f it is as if it is coming from that appropriate measurable
structure on the domain side and lands up on the measurable space on the range side.
So that is RBR.
So that is how the measurability should be explained.
So if you write this then it will mean the Borel measurability of small f is exactly
as per the above definition meaning you explicitly specify the sigma fields.
So if the sigma fields are understood then you can remove it you don't have to mention
it but it is always good to keep track of the sigma fields on the domain side and the
range side.
And to do that we are going to use this term f from omega f omega script f to RBR.
That notation that function is measurable.
So we are going to write this statement.
With this introduction of measurable functions we are now ready to talk about examples of
Borel measurable functions.
So let's start with the simplest example that you can think of.
So you take a measurable space omega script f and fix a constant.
So the simplest function that you can think of is a constant function.
So what you do?
So for all points in the domain omega you associate the value that constant value small
c.
So that gives you a constant function on the domain.
Now you are going to ask is this constant function a measurable function or a Borel
measurable function.
So on the domain side you have already been given some arbitrary sigma field arbitrary
but fixed script f.
So that's a sigma field that is on the domain side.
On the range side you have the real line which is of course having the Borel sigma field
and with these two specific choices of the sigma fields you are going to talk about the
measurability of the function.
But now what you need to do?
You are going to choose subsets Borel subsets from the range side.
So here you are going to choose this sets A and you are going to look at the pre-images
of those sets.
So it turns out that when the function f is just a constant function taking value the
constant value c then the pre-images f inverse of a that can be only two possible things.
What is this?
So it depends on the choice whether the constant value is in the set A or not.
If the constant value is not in the set A then the pre-image turns out to be an empty
set.
If the constant value is in A then the pre-image turns out to be the whole set.
Why is this?
So again go back to the definition.
What is f inverse of A?
You are going to look at all possible points in the domain such that the f of omega that
value lands up in the set A.
But for this specific function the constant function f of omega no matter what omega is
just a constant c.
So therefore if your constant value c is in the set A then you are going to get all points
in the domain side as its pre-image because all those points are landing up exactly in
c so therefore f of omega is in A.
So therefore you get the whole set.
But if the point c is not in the set A then the point is that f of omega is never in A
for all possible omega in the domain side.
So therefore you end up with the empty set.
So now you get these two possible cases for the pre-images.
So no matter what A you are choosing you always end up with these two choices either the empty
set or the whole set.
But then if you take any arbitrary sigma field it will always contain this empty set or the
whole set that's by construction by definition and hence the pre-images are always in the
sigma field no matter which Borel subset you choose from the range side.
So for all possible Borel subsets if you look at the pre-images they always fall in the
domain side sigma field and therefore it turns out to be Borel measurable.
So this is therefore Borel measurable or F script F over BR measurable for any choice
of sigma field.
So this is happening because the empty set and the whole set is always in the sigma field.
So that is what I have already explained.
Please try to write it down I have left this as an exercise.
So please check that no matter which sigma field you choose constant functions always
remain Borel measurable.
So now you are interested in some more general functions.
So we have already introduced this indicator function subsets and that's what we are going
to look at next.
So take a measurable space and choose some subset on the domain side.
So B is a subset on the domain side.
So let's take some arbitrary subset it need not be in the sigma field script F.
So now look at the indicator function.
So how is it defined?
So for all point in the domain all you need to check whether the point is in the set B
or not.
If it is there you associate the value 1 otherwise you associate the value 0.
It's simply a 0 1 valued function.
That's the function.
Now you want to check that given this choice of script F sigma field on the domain side
you are going to check whether this function this indicator function is measurable or not.
So again what you need to do you are going to choose these Borel subsets from the range
side.
So take this sets A and you are going to look at the pre-images and you are going to check
whether the pre-images fall in your sigma field on the domain side script F.
So now what are the possible cases?
So you can now check as done for the constant function case that you can split it into four
possible cases.
So for the constant function case you had two choices but now you have to look at these
two values 0s and 1s and with these possible choices if 0s and 1s belongs to A or 1 does
not belong to A and so on.
So you have these four possible choices you will get these different different sets.
If neither of 0 and 1 is in A you are going to get empty set as the pre-image and so on.
Please check this.
This is left as an exercise for you.
But then what do you want?
You wanted this indicator B to be measurable.
Now remember these four possible options are all these things that will appear as the pre-images.
Now remember that a sigma field script F always contains the empty set and the whole set.
So you don't need to worry about these two things.
But a priori you do not know whether the sets B or B complement is in the collection script
F. So to check that indicator B is measurable what you need to check the set B is in the
collection script F. So I am saying that if you start with the assumption that B is in
script F then you can show that indicator B is F over B are measurable.
Why?
Because if B is in scripted F then B complement is also in scripted F and therefore all possible
pre-images for Borel sets will land up in your domain side sigma field script F. So
therefore indicator B will become Borel measurable.
On the other hand if you start off with assuming Borel measurability of the function indicator
B then you must have these pre-images B or B complement to be in your sigma field and
that simply reduces to the condition B belongs to script F. So please try to write it down.
I have already explained the ideals.
So therefore for indicator of B to be measurable where B is a subset on the domain side thus
the set B must be in the sigma field.
So what is the interpretation now?
So we saw that the constant function is measurable with respect to any sigma field that you choose
on the domain side.
But when you went to a non-constant function this indicator B takes two values.
So whenever you shifted away from the constant function immediately there came some restrictions
on the measurability conditions.
So for this to be measurable, this to be Borel measurable you only need to choose those sigma
fields where the set B is included.
If you don't have the set B in your sigma field on the domain side then indicator B
will not be measurable.
So that's what we have discussed.
So therefore this already put some restrictions on the sigma fields or on the other side on
the measurable functions.
So therefore first you always guessed constant functions as measurable functions but whenever
you move on to specific choices of sigma fields then for non-constant functions you have to
be very very careful and it is very important to identify which are the non-constant measurable
functions given the situation.
So given the appropriate choices of the sigma fields that you start off with you always
should try to identify whether there are any non-constant Borel measurable functions.
Now with these two examples at hand we are not really satisfied.
We would like to obtain more examples.
So now in what follows we are going to discuss a large class of examples.
What do we do?
So we consider functions with range r but the domain is also taken to be r.
So we are taking r to r functions and we are going to talk about Borel measurability of
such functions meaning we are going to put Borel sigma field both on the range side as
well as the domain side because both domain and the range are the real line.
We are going to make this choice and we are going to talk about the Borel measurability
here.
Now you can check that if you take a continuous function then you can look at the pre-images
of all open subsets and it will turn out to be open.
So if you start off with a continuous function it will imply that pre-images of open sets
is open.
On the other hand if you assume this condition that the pre-images of all open sets is open
then you can show that the function is continuous.
So this is an if and only if condition.
This is a statement that comes from the real analysis please try to work this out.
Now using this exercise we get this nice collections or large class of examples for Borel measurable
functions but with the domain real line.
So this result follows using principle of good sets which we had discussed in note 22
of week 1 and combining with this exercise that we have mentioned about identification
of continuous functions with this pre-image conditions for open subsets.
So if you use the continuous functions, if you look at the continuous functions on the
real line then using the principle of good sets what you can show is that these continuous
functions turn out to be Borel measurable.
So you get a large class of examples.
So here what we are doing?
We are looking at real valued continuous functions on the real line and we are choosing the Borel
sigma fields on both sides, the range and the domain and we are saying that if you take
a continuous function then it's automatically Borel measurable.
So that's the statement of the proposition.
But since this requires this principle of good sets which we have skipped earlier, we
are not going to discuss the proof.
We are going to assume this as a fact and we are going to use this.
So for concrete examples of concrete continuous functions and hence measurable functions you
can choose to look at polynomials with real coefficients.
So if you choose any polynomial with real coefficients they will turn out to be Borel
measurable because these are continuous functions.
Moreover you can think of other nice continuous functions like trigonometric functions like
sine functions and cosine functions.
They are also continuous functions defined on the real line and therefore they will turn
out to be Borel measurable.
So therefore you have a large class of examples immediately after you get this proposition
and other than that we have discussed these two examples.
One was the constant functions, another was the indicator functions.
So we are going to continue these discussions for Borel measurable functions but we are
going to look at functions taking values in higher dimensions in Euclidean spaces, higher
dimensional Euclidean spaces.
This we are going to discuss in the next lecture.
We stop here.
