Lecture 04

IIT Kanpur July 201
20 followers
Oct 22, 2023
Welcome to this lecture, this is the fourth lecture of week 1.
In the previous lectures, we have introduced the concept of LP spaces on measure spaces
and on top of that we have discussed properties of the P-norm.
We have justified calling it a norm and by showing the impertinent properties of the
norm.
Moreover, since LP space forms now a normed linear space, we have a natural distance function
given by the norm.
This distance function or the metric yields a topology on this LP spaces.
In this lecture, we are going to be talking about convergences in terms of this distance
or the metric.
So, without further delay let us move on to the slides.
So, let us set up the notation once more.
We are going to be talking about this P-norm and as we have already seen this P-norms gives
us a natural distance which is that given any two functions in LP, there is this natural
distance given by P-norm of f minus g.
Now, this allows us to consider convergences for measurable functions and in the special
cases of probability measures, we end up with random variables and therefore, we can talk
about convergences of random variables in this LP space sense.
So, we are going to be continuing with the notations that we have been so far using.
So, let us set up the main terminology.
First let us consider the notion of convergence in LP.
We say that a sequence of functions fn in this LP space converges in LP to another element
in LP, if f minus fn the P-norm of that goes to 0 as n goes to infinity.
Now, this concept can be restated as we have seen earlier for random variables also and
let us restate it.
So, this is now convergence in LP with the probability measures.
We say that a sequence of random variables xn in LP converges in LP to a random variable
x in LP, if the P-norm of x minus xn which is nothing but the P-th root of the P-th absolute
moment of x minus xn that should go to 0 as n goes to infinity.
Now, as you can see that this is equivalent to the fact that just the P-th absolute moment
of x minus fn should go to 0 as n goes to infinity and therefore, we can visualize it
as a convergence in the P-th mean or as it is said in the P-th moment.
Now, we will typically call it in terms of the P-th moment when we are dealing with random
variables, whenever we are dealing with generally measurable functions we are going to be calling
it a convergence in LP, but both notions are the same.
Now we introduce notations to talk about this type of convergences.
We shall write fn converges to f, but with a superscript like on the arrow with LP here.
For random variables the similar notation we will have LP on top of the arrow to denote
this convergences.
Now, in what follows we are going to be focusing on results specifically either on LP of mu
or LP of P, but not both.
Note that results which we have stated for LP of mu that is for any general measure mu
also applies to any probability measure P and therefore, it will hold for LP of P, but
some of the results may not be true the other way.
Meaning if we prove some results for LP of P considering it is a probability measure
then some of the results might not hold for general measures.
We have already seen example of such a result when we have shown that LP for P greater than
1 contained in L1 when you are dealing with a probability measure.
In case of finite measures the results continues to hold, but may not be true if you are dealing
with an infinite measure.
So, whenever you see some result proved for LP of P be careful before assuming that it
also hold for a general measure.
We now move on and discuss convergence properties.
First of all we have already seen that if a random variable is in LQ a higher Q than
P then X also is in the lower integrability space LP.
Now, that result also implies this next convergence statement that if you choose any P and Q in
this range that between 1 and infinity and P is less than Q then convergence in LQ the
higher moment implies convergence in the lower moment.
To see this you just recall that there is this interesting inequality that we have proved
in the previous lecture that was in remark 3 that the P norm of X minus Xn is dominated
by the Q norm of X minus Xn.
This is true whenever P is less than Q, but then the higher moment that is in terms of
the Q norm of X minus Xn that goes to 0 by the hypothesis and therefore, X minus Xn the
P norm of that must also go to 0 and hence we will get the required convergence as required.
Now as an exercise you can try to show that the other direction is not true meaning that
you choose P and Q like this, but assume the convergence in LP only.
Try to show by an example that it may not be converging in LQ construct explicitly the
sequence of random variables such that this statement is true in LP, but not in LQ.
We are now going to be restating the dominated convergence theorem that we have stated earlier
in proposition 2.
So, this is now proposition 8.
So, take a sequence of functions f1, f2, fn and so on and a limit function f.
Suppose this convergence holds mu almost everywhere.
If there exists a function g in LP such that modulus of fn is dominated by g, then all
these functions f, f1, f2 and so on must also be in LP.
Moreover, the convergence should hold in LP.
So, the DCT is now being restated and the integrability condition is in now in terms
of the LP integrability.
So, of course, you can restate this condition in terms of random variables, but we are not
going to be doing that.
We are now familiar with the notion that if you prove something for general measure spaces
this also applies to probability measures.
So, any result that is proved in general for measurable functions also applies to random
variables.
But as a practice you can try to write down this statement for random variables on your
own.
Now, we establish a very important approximation result on LP.
This is approximation by simple functions.
Remember that for a measurable function we have seen such approximation results, but
in the almost everywhere situation.
The approximation require the use of the LP norm or the metric that is associated to
the LP norm.
So, we show this approximation results using the LP distances.
Start with a function f in LP, then for any epsilon positive the statement says that we
can construct a simple function g such that f minus g the P norm of that must be less
than epsilon.
Moreover, this g that we are stating here can be chosen with further additional properties
such as modulus of g should be less equals to modulus of f and g must be finite valued.
Here let me clarify the term finite valued, whenever we are talking about simple functions
it is a finite sum of indicator functions.
Now, whenever you are talking about finite valued we mean that these functions do not
take the values plus or minus infinity.
This approximation by simple functions this statement can also be restated in following
alternative variants.
So, before proving this we restate these versions first.
Consider the class of simple functions which are already in LP.
So, there might be some simple functions which are not in LP be careful with those.
So, continue with this simple functions which are already inside LP, then the closure of
such a class by closure we mean that we are using the topology in LP.
So, that closure in LP turns out to be the whole space itself.
So, this is the approximation by simple functions.
So, therefore, by limiting operations we can approximate any LP functions.
So, therefore, this closure is the whole space itself.
Another way to state the same result is that the simple functions in LP are dense in LP.
We can also state it in terms of approximations by sequences.
Start with a function f in LP, then there exist a sequence of finite valued simple functions
given by g n such that g n converges to f in the LP sense moreover modulus of g n must
be less or equals to modulus of f for all n.
Since we have restated the original result in alternative forms, let us now move on and
look at the proof of this.
Since f is measurable, first of all we can construct a sequence of finite valued simple
functions such that g n converges to f in the point wise sense moreover we can also
ensure that modulus of g should be less than modulus of f.
Here there is a minor typo this will be corrected in the final lecture notes supplied to you.
Since the limit function f is in LP, by this domination we will claim that g n must also
be in LP.
So, therefore, the sequence of simple functions that we are dealing with here are already
in LP.
Now observe that if you look at modulus of f minus g n that can be dominated by modulus
of f plus modulus of g n, but since g n's are less than modulus of f we end up with
two times modulus of f as an upper bound for modulus of f minus g n, but f is already
in LP and it immediately follows that two times modulus of f must also be in LP.
And hence by applying DCT we will claim that f minus g n must now converge in the LP sense
to 0.
Another way to state this is that g n converges to f in the LP to the function f.
Now, in particular for any epsilon positive we can find a natural number n such that f
minus g n the p norm of that should be less than epsilon.
This is following from the fact that we have shown that the sequence of values f minus
g n p norm of that that sequence of values converges to 0.
Therefore, given any epsilon positive we can choose a large enough integer n such that
nth term onwards the terms are smaller than epsilon.
And hence this g subscript capital N is our choice for a finite valued simple function
which is also in LP.
This completes the proof.
We now move forward and introduce another related concept to convergences.
This is the notion of Cauchy sequences in LP.
And as we shall see this Cauchy sequences in LP plays a very very important role.
Let us recall the definition.
We say that a sequence f n in LP is Cauchy if given any epsilon positive there exists
a natural number n such that all terms after that n onwards are closed by epsilon.
All n and m or small n and small m the corresponding functions f n and f m must be closed in LP.
LP distances closure by this epsilon distance.
This is the notion of Cauchy sequences.
We have already seen such similar definitions in the real line.
And in fact, in any matrix space such as the real line or the LP space that we have been
discussing all convergent sequences are Cauchy.
This is a important property that convergent sequences carry.
But conversely we need to recall that Cauchy sequences in R are convergent.
This is a very very important equivalence that is established on the real line.
Remember again in any matrix space all convergent sequences are Cauchy.
But for the converse direction starting with Cauchy sequences and going to convergent sequences
works in the real line.
What we are going to be discussing is a similar property in the case of LP spaces.
And remember from Cauchy sequences to convergent sequences if we can show such a implication
we get the completeness of the space.
For example, on the real line if you manage to show that Cauchy sequences are convergent
then we end up having the completeness of R.
What we are going to discuss is a similar property for the LP spaces.
Before that we recall an important result from basic measure theory.
This is the Borel-Cantelli lemma first half of it.
In a later lecture we are going to be seeing the second half of it.
So, we are going to be working with this measure space and choosing this sequence of sets coming
from the sigma field.
Suppose this sequence of sets satisfy this property that some of their sizes under the
measure mu should be finite then it follows that lim sup of a n must have zero measure
under the measure mu.
This is the Borel-Cantelli lemma first half.
This is also an important result as you see in probability theory where a measure mu is
replaced by a probability measure.
As a consequence of this we are going to be obtaining a important technical lemma.
Suppose I have a sequence of functions fn in LP such that the n plus 1th function and
the nth functions are closure in P-norm by this distance 4 to the power minus n for all
n.
Then the claim is that the sequence of function must converge in the almost everywhere sense.
To see this we start by looking at the sets a n, n equal to 1, 2 and so on defined by
this set of points where these functions differ by more than 2 power minus n distance.
This is point wise distances that we are looking at.
Now remember according to the notation that we adopt we are going to be removing this
small omegas and simply rewriting this inequality here.
So, this is again another shorthand description of the sets a n that we are going to be using.
Now the Markov inequality that we have just talked about gives us that the size of the
sets under the measure mu can be dominated by the P-th power of this norm and with this
appropriate scaling.
This is the application of Markov's inequality, but since we have chosen these sets in this
familiar manner and these sequence of functions satisfy this 4 power minus n distance.
So, raise to the power P you get minus n to the power P and finally the size of the sets
a n must be dominated by 2 power minus n P.
This is a very standard manipulation that we have performed and as a consequence of
that some of the sizes of the sets turns out to be finite since we are working with P between
1 and infinity.
Now it is the application of Borel-Cantelli lemma first half and we will immediately claim
that lim sup of these sets that we are considering must have 0 measure.
Now while dealing with integrations we have stated earlier also that we are going to be
ignoring these sets of measure 0.
So, on the complement of this lim sup set that we see here we are going to be doing
our analysis.
Now on the complement by the description of the sets a n we conclude that for large enough
n we will now continue to have the upper bound given by this.
Remember the sets a n were given by the other side of the inequality and lim sup a n complement
essentially says that after a stage onwards you are going to be having the other side
of the inequality.
Therefore, for all large n you end up with this upper bound which is pretty short which
is 2 power minus n.
And hence as a consequence of this we get the point wise convergence of the sequence
f n omega.
This f n omega is a real sequence and this bound actually implies the Cauchy property.
By the completeness you immediately have the convergence.
And hence what we have established is that on the complement of this null set which is
the complement of this lim sup this on this complement we have the point wise convergence
of the function values.
Now on the set of measure 0 we do not really care about the limit function values.
So, therefore, we essentially conclude that outside this null set anyway we have shown
the convergence.
So, therefore, this sequence of functions converges almost everywhere.
So, therefore, this technical lemma is proved that as soon as this sequence of functions
f n satisfy the relation that f n plus 1 minus f n in p norm is smaller than 4 to the power
minus n then this will converge in the almost everywhere sense also.
Now we are going to be commenting on a specific part of this proof that we have not really
explicitly identified the limit function here.
So, what we have done is that on the complement of the null set we have essentially identified
the limit, but have not really identified the whole limit function on the whole domain,
but we can do it very easily as done in the following expressions.
So, what we do?
We state this function on the domain omega with real numbers as its values given by this
that on the complement of this lim sup set we continue to take the limit values as already
justified and on the set of measure 0 which is the lim sup we will just set it to be 0
that will completely describe this limit function and as we can see it is a very standard easy
verification that this limit function f that we have been claiming this is actually a measurable
function and this almost everywhere convergence will hold.
Now here of course, we need to comment that instead of the value 0 we could also use any
other known measurable functions and that will also work as a limit function.
But we have to keep on the complement of the null set the value given by the limit function
only on the null set we can change the values according to our wish coming from any measurable
function for example, we have used the constant function 0 here.
Now combining all these results that we have seen so far we now come to a very important
structural property of LP spaces.
So far what we have already observed is that LP spaces together with the norm forms a normed
linear space, but with this property of completeness that we have been discussing so far we claim
that all the Cauchy sequences in LP are convergent and as a consequence of that what we end up
having is that this normed linear space LP are complete.
So, let us establish this result.
So, we have to start off with Cauchy sequences in LP and show that they are convergent, but
before we prove this let us make some comments about this result.
So, what we have is that this LP spaces are normed linear spaces and this previous theorem
is saying that this spaces must be complete.
So, LP spaces are complete normed linear spaces which are nothing, but Banach spaces.
So, therefore, we have very special structure of this LP spaces which turn out to be Banach
spaces.
Now, we go on into the proof and here we shall see the application of the technical lemma
that we have established earlier.
We start with a Cauchy sequence fn.
Now since it is given to be Cauchy with any choice such as 1 by 4 we can get a number
capital N1 such that all the terms after this number N1 must be closer to each other by
this distance.
Now as soon as we obtain this N1 we set the first function g1 to be f of N1.
Likewise, we continue and what we do is that we now choose a smaller number which is 4
to the power minus 2 and accordingly we will be obtaining a number N2 such that all the
subsequent terms are closer to each other by this rate and without loss of generality
what we can do we can choose this N2 to be greater than N1.
With those choices we now set g2 to be f of N2.
Then likewise once we have obtained the first k many functions g1, g2, gk we have to explain
how to get the next function.
To do that first we find Nk plus 1 greater than Nk such that all the subsequent functions
after that stage onwards must be closer to each other by this distance and as soon as
you obtain such a number Nk you set gk plus 1 to be that function.
Therefore this way we have managed to construct a subsequence of the original functions itself.
Remember gks are special choices of the functions from the original sequence and therefore gks
are actually a subsequence of the original sequence.
Now this subsequence has been constructed with the property that the subsequent distances
between all the successive functions is less than 4 to the power minus N.
That is how we have constructed this subsequence.
By this earlier lemma what we have is that this sequence of functions converges mu almost
everywhere to a limit function and therefore, we just give it a name call it g.
Therefore, we have a subsequence which converges almost everywhere to a measurable function
g.
We first claim that this limit function must be in Lp and finally, we are going to be showing
that the original sequence itself must converge to this limit function in the Lp sense.
So, we have managed to identify the limit function of the original sequence.
To establish this first fix an epsilon positive then again by the Cauchy property of the original
sequence we can get a number capital N such that successive distances between functions
after the Nth stage will be less than epsilon to the power 1 by p.
In particular if we choose our Nks which are larger than capital N then replacing the fm
by f of Nk we end up with this relation and here as we let k going to infinity fix the
capital N let k go to infinity we will end up with that the lim inf of these quantities
is upper bounded by this epsilon.
The only difference from the previous relation to the this relation is that we have raised
it to the power p and therefore, we have gotten rid of the epsilon to the power 1 by p and
this appears as a epsilon on our left hand side.
Now, remember the norm raised to the power p actually is an integral and therefore, here
we can apply the Fatou's lemma push the lim inf inside and therefore, you will get this
inequality here this is an application of the Fatou's lemma.
Once you have that observe that we have already established the point wise convergence of
this subsequences here and that is nothing, but what we denoted by g.
So, that next expression is an equality because by identifying the limit for this is a point
wise limit, but then this again can be expressed in terms of the pth power of the p norm between
fn minus g which we now happen to have to be less than epsilon and in particular fn
minus g p norm of that has to be finite and therefore, it is in lp.
Now, since this can be made smaller than epsilon for all large capital N, then fn minus g converges
to 0 in the lp sense therefore, as a whole what we end up having is that the function
g that we have obtained as a limit of subsequences now can be rewritten in this algebraic form
as a difference between fn and fn minus g both of which are lp functions.
Since lp is a linear space g must also be in lp and finally, once we have established
fn minus g converges to 0 in the lp we just rewrite that relation and end up having that
the original sequence fn itself converges to g in the lp sense and therefore, this completes
the proof.
The final statement that we are going to be talking about in this lecture is this important
result that what is the relation between lp convergences and almost everywhere convergences.
In the DCT as we have seen almost everywhere convergences with some integrability condition
implies the lp in convergences.
On the other hand we are going to be seeing this important relation from lp convergences
to almost everywhere convergences, but only through a subsequence.
So, the result states like this we take a lp convergent sequence then we can get a subsequence
which converges to the limit function in the almost everywhere sense.
To see this start with the convergent sequence now since this is convergent this is Cauchy
and with the type of arguments that we have been doing earlier we can construct a subsequence
such that the successive functions are closer to each other in the p norm by 4 to the power
minus k distance for every k.
And therefore, by our earlier technical lemma this specific subsequence that we construct
converges almost everywhere and let us call this limit function g.
As we have seen earlier this limit function this convergence actually holds also in lp,
but observe here that we have started off with the original sequence converging to the
limit function f in the lp sense and therefore, as a consequence of that by considering the
uniqueness of the limit we can check that f and g must match almost everywhere.
And therefore, as a consequence of this we will immediately get that f of nk the subsequence
must converge to the limit function f in the almost everywhere sense.
Please complete the argument this is left as an exercise for you this completes the
proof ok.
So, up to this we have discussed some important properties of lp spaces and prove that lp
spaces forms Banach spaces.
So, in the next lecture we are going to be talking about spaces related to be lp spaces
we stop here.
Thank you.
