WEBVTT

00:00:00.660 --> 00:00:04.340
In this segment, we will first
discuss and compare different

00:00:04.340 --> 00:00:06.890
views of the sum of independent
identically

00:00:06.890 --> 00:00:08.870
distributed random variables.

00:00:08.870 --> 00:00:11.150
And then, we will conclude
with a statement of the

00:00:11.150 --> 00:00:12.930
central limit theorem.

00:00:12.930 --> 00:00:16.070
So let X1 up to Xn be
independent identically

00:00:16.070 --> 00:00:19.440
distributed random variables
that have a certain finite

00:00:19.440 --> 00:00:21.520
mean and finite variance
that we'll denote by

00:00:21.520 --> 00:00:23.440
mu and sigma squared.

00:00:23.440 --> 00:00:27.140
In order to have a concrete
example in our hands, let us

00:00:27.140 --> 00:00:30.490
assume that this random variable
has a distribution,

00:00:30.490 --> 00:00:35.220
let's say, a PDF, that ranges
from minus 1 to plus 1 and has

00:00:35.220 --> 00:00:38.632
a mean of 0.

00:00:38.632 --> 00:00:42.080
Let us look at the sum of
these random variables.

00:00:42.080 --> 00:00:46.040
The sum has a variance of n
times sigma squared, which

00:00:46.040 --> 00:00:47.420
goes to infinity.

00:00:47.420 --> 00:00:50.490
And correspondingly, the
standard deviation of this

00:00:50.490 --> 00:00:53.740
random variable also
grows to infinity.

00:00:53.740 --> 00:00:56.140
This random variable takes
values between

00:00:56.140 --> 00:00:59.820
minus n and plus n.

00:00:59.820 --> 00:01:02.770
And because the variance and
the standard deviation

00:01:02.770 --> 00:01:07.460
increase, this means that for
larger and larger n, the width

00:01:07.460 --> 00:01:11.460
of this distribution is going
to be larger and larger.

00:01:14.080 --> 00:01:18.940
We can obtain a different view
of this sum if we divided by n

00:01:18.940 --> 00:01:22.470
in, which case, we obtain
the sample mean.

00:01:22.470 --> 00:01:26.039
In this case, the variance
goes to 0

00:01:26.039 --> 00:01:28.060
as n goes to infinity.

00:01:28.060 --> 00:01:32.280
And as a consequence, the
distribution is highly

00:01:32.280 --> 00:01:35.360
concentrated around 0.

00:01:35.360 --> 00:01:40.729
This is also what the weak law
of large numbers tells us.

00:01:40.729 --> 00:01:44.350
The bulk of the distribution
is concentrated in an

00:01:44.350 --> 00:01:47.320
arbitrarily small interval
around 0.

00:01:47.320 --> 00:01:50.690
So this width becomes
smaller and smaller

00:01:50.690 --> 00:01:53.710
as n goes to infinity.

00:01:53.710 --> 00:01:57.770
So in this case, we obtain
a limiting distribution.

00:01:57.770 --> 00:02:00.560
But this limiting distribution
is trivial.

00:02:00.560 --> 00:02:01.530
It's degenerate.

00:02:01.530 --> 00:02:04.900
It's all concentrated
on a single point.

00:02:04.900 --> 00:02:08.250
How can we make it so that we
obtain a limiting distribution

00:02:08.250 --> 00:02:10.038
that is more interesting?

00:02:10.038 --> 00:02:14.900
The key is to divide not by
n, but to divide by the

00:02:14.900 --> 00:02:16.570
square root of n.

00:02:16.570 --> 00:02:18.410
This has the following effect.

00:02:18.410 --> 00:02:22.390
The variance of this ratio
is calculated as follows.

00:02:22.390 --> 00:02:25.930
We take the variance of the
numerator, which is n times

00:02:25.930 --> 00:02:27.350
sigma squared.

00:02:27.350 --> 00:02:29.800
And then we divide by
the square of this

00:02:29.800 --> 00:02:31.760
number, which is n.

00:02:31.760 --> 00:02:36.440
And therefore, the variance
is equal to sigma squared.

00:02:36.440 --> 00:02:42.180
What's important here is that
the variance stays constant.

00:02:42.180 --> 00:02:47.110
No matter what n is, the width
of this distribution is going

00:02:47.110 --> 00:02:50.820
to be more or less the same.

00:02:50.820 --> 00:02:54.860
The distribution itself might
change as n changes.

00:02:54.860 --> 00:02:55.990
But the distribution--

00:02:55.990 --> 00:02:59.240
at least in this case, where
we assume 0 mean, the

00:02:59.240 --> 00:03:00.920
distribution stays in place.

00:03:00.920 --> 00:03:03.640
It doesn't move to the
right or to the left.

00:03:03.640 --> 00:03:06.000
And its width stays the same.

00:03:06.000 --> 00:03:09.930
So one can wonder, in the limit,
as n goes to infinity,

00:03:09.930 --> 00:03:12.430
does this shape start
to approach a

00:03:12.430 --> 00:03:14.400
certain limiting shape?

00:03:14.400 --> 00:03:17.720
And if it does, what is the
limiting shape that it

00:03:17.720 --> 00:03:18.990
approaches?

00:03:18.990 --> 00:03:21.210
The central limit theorem
will give us the

00:03:21.210 --> 00:03:22.600
answers to these questions.

00:03:25.110 --> 00:03:28.270
The setting for the central
limit theorem will be pretty

00:03:28.270 --> 00:03:31.770
much the setting that we
were just discussing.

00:03:31.770 --> 00:03:35.300
So we will be looking into the
case where we divide the sum

00:03:35.300 --> 00:03:38.810
of the random variables by
square root of n, except for a

00:03:38.810 --> 00:03:41.880
few additional twists.

00:03:41.880 --> 00:03:47.880
Since this ratio has a variance
of sigma squared, it

00:03:47.880 --> 00:03:52.930
would help to divide by a
further factor of sigma here

00:03:52.930 --> 00:03:56.579
so that the variance is
going to become 1.

00:03:56.579 --> 00:03:58.829
And there's another issue.

00:03:58.829 --> 00:04:03.210
If the mean of the X's
is non-zero, then the

00:04:03.210 --> 00:04:07.750
distribution is centered at
a quantity that keeps

00:04:07.750 --> 00:04:09.460
changing with n.

00:04:09.460 --> 00:04:13.540
So the distribution will be
drifting away from 0.

00:04:13.540 --> 00:04:15.060
It's not staying in place.

00:04:15.060 --> 00:04:17.050
And so it wouldn't
have any hope of

00:04:17.050 --> 00:04:19.180
converging to something.

00:04:19.180 --> 00:04:22.880
For this reason, instead of
looking at this ratio in

00:04:22.880 --> 00:04:28.440
particular, what we do is we
first subtract the mean of the

00:04:28.440 --> 00:04:31.570
sum, which is n times
the mean.

00:04:31.570 --> 00:04:37.050
And then we divide by a further
factor of sigma.

00:04:37.050 --> 00:04:41.680
This random variable that
we obtain here has nice

00:04:41.680 --> 00:04:43.000
properties.

00:04:43.000 --> 00:04:46.550
The mean of this random variable
is equal to 0,

00:04:46.550 --> 00:04:50.520
because we did subtract
the mean of the X's.

00:04:50.520 --> 00:04:55.409
And the variance of this random
variable is going to be

00:04:55.409 --> 00:04:57.690
equal to 1.

00:04:57.690 --> 00:05:01.050
The reason is that the variance
is the variance of

00:05:01.050 --> 00:05:05.600
the numerator, which is n times
sigma squared, divided

00:05:05.600 --> 00:05:08.620
by the square of the
denominator, which is also n

00:05:08.620 --> 00:05:10.510
times sigma squared.

00:05:10.510 --> 00:05:14.900
So as n changes, the
distribution of the random

00:05:14.900 --> 00:05:18.360
variable Zn stays in place.

00:05:18.360 --> 00:05:20.066
It has a mean of 0.

00:05:20.066 --> 00:05:24.230
And its width, more or less,
stays the same, because we

00:05:24.230 --> 00:05:25.560
have a constant variance.

00:05:28.620 --> 00:05:33.040
We will compare this random
variable with a standard

00:05:33.040 --> 00:05:38.460
normal random variable that has
0 mean and unit variance.

00:05:38.460 --> 00:05:43.100
The central limit theorem states
the amazing fact that

00:05:43.100 --> 00:05:47.909
as n goes to infinity, the
distribution of this random

00:05:47.909 --> 00:05:51.670
variable converges to the
standard normal distribution

00:05:51.670 --> 00:05:53.440
in the following sense--

00:05:53.440 --> 00:05:59.350
that this probability here
converges to that probability

00:05:59.350 --> 00:06:02.640
for any choice of little z.

00:06:02.640 --> 00:06:06.670
Now, what we have here is just
the CDF of this random

00:06:06.670 --> 00:06:07.930
variable, Zn.

00:06:07.930 --> 00:06:11.040
So it tells us that the CDF
of the random variable Zn

00:06:11.040 --> 00:06:14.020
converges to the CDF of
a standard normal.

00:06:14.020 --> 00:06:17.520
And fortunately, for the
standard normal, the CDF is

00:06:17.520 --> 00:06:19.360
available in tables.

00:06:19.360 --> 00:06:22.960
So if we needed to calculate the
numerical value here, we

00:06:22.960 --> 00:06:25.640
can just look up the
normal tables.

00:06:25.640 --> 00:06:28.980
And this suggests an
approximation to this

00:06:28.980 --> 00:06:33.230
probability for the case where
n is finite but large.

00:06:33.230 --> 00:06:36.900
When n is large, we can
approximate this probability

00:06:36.900 --> 00:06:40.140
by this probability on the
right, which we can find from

00:06:40.140 --> 00:06:42.190
the normal tables.

00:06:42.190 --> 00:06:45.590
The central limit theorem is
a very important result.

00:06:45.590 --> 00:06:48.909
For this reason, we will spend
some time discussing how to

00:06:48.909 --> 00:06:53.460
interpret it, what it means, how
we use it, and we will go

00:06:53.460 --> 00:06:56.790
through a few examples to see
how we actually apply it.