WEBVTT

00:00:00.580 --> 00:00:04.710
In this problem, we're given a
collection of 10 variables, x1

00:00:04.710 --> 00:00:10.580
through x10, where each i, xi,
is a uniform random variable

00:00:10.580 --> 00:00:12.345
between 0 and 1.

00:00:12.345 --> 00:00:18.710
So each i is uniform between 0
and 1, and all 10 variables

00:00:18.710 --> 00:00:19.950
are independent.

00:00:19.950 --> 00:00:26.640
And we'd like to develop a bound
on the probability that

00:00:26.640 --> 00:00:34.860
some of the 10 variables, 1 to
10, being greater than 7 using

00:00:34.860 --> 00:00:36.250
different methods.

00:00:36.250 --> 00:00:39.960
So in part A we'll be using
the Markov's inequality

00:00:39.960 --> 00:00:41.240
written here.

00:00:41.240 --> 00:00:44.060
That is, if we have a random
variable, positive random

00:00:44.060 --> 00:00:48.350
variable x, the probability x is
greater than a, where a is

00:00:48.350 --> 00:00:52.480
again some positive number, is
bounded above by the expected

00:00:52.480 --> 00:00:55.450
value of x divided by a.

00:00:55.450 --> 00:00:58.140
And let's see how that works
out in our situation.

00:00:58.140 --> 00:01:02.910
In our situation, we will call
x the summation of i equal to

00:01:02.910 --> 00:01:13.010
1 to 10xi, and therefore, E of
x is simply 10 times E of x1,

00:01:13.010 --> 00:01:16.470
the individual ones, and
this gives us 5.

00:01:16.470 --> 00:01:21.970
Here we use used the linearity
of expectation such that the

00:01:21.970 --> 00:01:24.530
expectation of the sum of the
random variable is simply the

00:01:24.530 --> 00:01:27.020
sum of the expectations.

00:01:27.020 --> 00:01:29.600
Now, we can invoke Markov's
Inequality.

00:01:29.600 --> 00:01:32.630
It says x greater
or equal to 7.

00:01:32.630 --> 00:01:42.750
This is less than E of x over 7,
and this gives us 5 over 7.

00:01:42.750 --> 00:01:45.780
For part B, let's see if we can
improve the bound we got

00:01:45.780 --> 00:01:49.860
in part A using the Chebyshev
inequality, which takes into

00:01:49.860 --> 00:01:52.950
account the variance of
random variable x.

00:01:52.950 --> 00:01:56.700
Again, to refresh you on this,
the Chebyshev Inequality says

00:01:56.700 --> 00:02:01.010
the probability that x deviates
from its mean E of x,

00:02:01.010 --> 00:02:05.850
by more than a is bound above
by the variance of x divided

00:02:05.850 --> 00:02:08.090
by a squared.

00:02:08.090 --> 00:02:11.960
So we have to actually do some
work to transform the

00:02:11.960 --> 00:02:15.680
probability we're interested
in, which is x greater or

00:02:15.680 --> 00:02:20.140
equal to 7, into the form that's
convenient to use using

00:02:20.140 --> 00:02:22.430
the Chebyshev Inequality.

00:02:22.430 --> 00:02:26.510
To do so, we'll rewrite this
probability as the probability

00:02:26.510 --> 00:02:32.410
of x minus 5 greater or equal
to 2 simply by moving 5 from

00:02:32.410 --> 00:02:33.920
the right to the left.

00:02:33.920 --> 00:02:37.470
The reason we chose 5 is because
5 is equal to the

00:02:37.470 --> 00:02:42.660
expected value of x from part
A as we know before.

00:02:42.660 --> 00:02:45.880
And in fact, this quantity is
also equal to the probability

00:02:45.880 --> 00:02:49.430
that x minus 5 less or
equal to negative 2.

00:02:52.490 --> 00:02:55.730
To see why this is true, recall
that x is simply the

00:02:55.730 --> 00:03:01.670
summation of the xi's, the 10
xi's, and each xi is a uniform

00:03:01.670 --> 00:03:04.370
random variable between
0 and 1.

00:03:04.370 --> 00:03:08.100
And therefore, each xi, the
distribution of which is

00:03:08.100 --> 00:03:11.470
symmetric around its mean 1/2.

00:03:11.470 --> 00:03:14.290
So we can see that after we
add up all the xi's, the

00:03:14.290 --> 00:03:16.960
resulting distribution
x is also symmetric

00:03:16.960 --> 00:03:19.050
around its mean 5.

00:03:19.050 --> 00:03:21.840
And as a result, the probability
of x minus 5

00:03:21.840 --> 00:03:26.380
greater than 2 is now equal to
the probability that x minus 5

00:03:26.380 --> 00:03:28.670
less than negative 2.

00:03:28.670 --> 00:03:32.120
And knowing these two, we can
then say they're both equal to

00:03:32.120 --> 00:03:38.440
1/2 the probability x minus 5
absolute value greater or

00:03:38.440 --> 00:03:42.270
equal to 2, because this term
right here is simply the sum

00:03:42.270 --> 00:03:44.850
of both terms here and here.

00:03:48.810 --> 00:03:51.270
At this point, we have
transformed the probability of

00:03:51.270 --> 00:03:54.970
x greater or equal to 7 into the
form right here, such that

00:03:54.970 --> 00:03:59.090
we can apply the Chebyshev's
Inequality basically directly.

00:03:59.090 --> 00:04:02.050
And we'll write the probably
here being less than or equal

00:04:02.050 --> 00:04:08.560
to 1/2 times, applying the
Chebyshev Inequality, variance

00:04:08.560 --> 00:04:11.510
of x divided by 2 squared.

00:04:11.510 --> 00:04:16.589
Now, 2 is the same as a
right here, and this

00:04:16.589 --> 00:04:20.230
gives us 1/8 times--

00:04:20.230 --> 00:04:24.230
now, the variance of x, we know
is 10 times the variance

00:04:24.230 --> 00:04:29.220
of a uniform random variable
between 0 and 1, which is

00:04:29.220 --> 00:04:33.230
1/12, and that gives us 5/48.

00:04:36.500 --> 00:04:38.900
Now, let's compare this with
the number we got earlier

00:04:38.900 --> 00:04:42.650
using the Markov Inequality,
which was 5/7.

00:04:42.650 --> 00:04:46.770
We see that 5/48 is much
smaller, and this tells us

00:04:46.770 --> 00:04:49.540
that, at least for this example,
using the Chebyshev

00:04:49.540 --> 00:04:52.720
Inequality combined with the
information of the variance of

00:04:52.720 --> 00:04:56.030
x, we're able to get a stronger
upper bound on the

00:04:56.030 --> 00:04:59.150
probability of the event that
we're interested in.

00:04:59.150 --> 00:05:01.890
Now, in part B, we saw that
by using the additional

00:05:01.890 --> 00:05:04.340
information of the variance
combined with the Chebyshev

00:05:04.340 --> 00:05:08.190
Inequality, we can improve upon
bound given by Markov's

00:05:08.190 --> 00:05:09.650
Inequality.

00:05:09.650 --> 00:05:12.420
Now, in part C, we'll use
a somewhat more powerful

00:05:12.420 --> 00:05:15.230
approach in addition to the
Chebyshev Inequality, the

00:05:15.230 --> 00:05:16.910
so-called central
limit theorem.

00:05:16.910 --> 00:05:19.450
Let's see if we can even
get a better bound.

00:05:19.450 --> 00:05:22.600
To remind you what a central
limit theorem is, let's say we

00:05:22.600 --> 00:05:28.740
have a summation of i equal
to 1 to some number n of

00:05:28.740 --> 00:05:30.610
independent and identically
distributed

00:05:30.610 --> 00:05:32.420
random variables xi.

00:05:32.420 --> 00:05:34.720
Now, the central limit theorem
says the following.

00:05:34.720 --> 00:05:38.770
We take the sum right here, and
subtract out its means,

00:05:38.770 --> 00:05:44.890
which is E of the same
summation, and further, we'll

00:05:44.890 --> 00:05:48.500
divide out, what we call
normalize, by the standard

00:05:48.500 --> 00:05:50.275
deviation of the summation.

00:05:50.275 --> 00:05:55.580
In other words, the square
root of the variance

00:05:55.580 --> 00:05:57.000
of the sum of xi.

00:05:59.500 --> 00:06:05.230
So if we perform this procedure
right here, then as

00:06:05.230 --> 00:06:09.090
the number of terms in the sums
going to infinity, here

00:06:09.090 --> 00:06:13.270
as in goes to infinity, we will
actually see that this

00:06:13.270 --> 00:06:17.490
random variable will converge
in distribution in some way

00:06:17.490 --> 00:06:20.960
that will eventually look like
a standard normal random

00:06:20.960 --> 00:06:23.880
variable with means 0 and 1.

00:06:23.880 --> 00:06:26.380
And since we know how the
distribution of a standard

00:06:26.380 --> 00:06:29.790
normal looks like, we can go to
table and look up certain

00:06:29.790 --> 00:06:32.580
properties of the resulting
distribution.

00:06:32.580 --> 00:06:34.430
So that is a plan to do.

00:06:34.430 --> 00:06:36.970
So right now, we have
about 10 variables.

00:06:36.970 --> 00:06:40.680
It's not that many compared to
a huge numbering, but again,

00:06:40.680 --> 00:06:43.310
if we believe it's a good
approximation, we can get some

00:06:43.310 --> 00:06:46.600
information out of it by using
the central limit theorem.

00:06:46.600 --> 00:06:49.840
So we are interesting knowing
that probability summation of

00:06:49.840 --> 00:06:55.250
i equal to 1 to 10 x1 greater
or equal to 7.

00:06:55.250 --> 00:06:59.410
We'll rewrite this as 1 minus
the probability the summation

00:06:59.410 --> 00:07:06.270
i equal to 1 to 10, and
xi less equal to 7.

00:07:06.270 --> 00:07:08.310
Now, we're going to apply
the scaling to the

00:07:08.310 --> 00:07:10.570
summation right here.

00:07:10.570 --> 00:07:20.050
So this is equal to 1 minus the
probability summation i

00:07:20.050 --> 00:07:25.880
equal to 1 to 10xi minus 5.

00:07:25.880 --> 00:07:29.020
Because we know from previous
parts that 5 is the expected

00:07:29.020 --> 00:07:32.090
value of the sum right
here, and divided by

00:07:32.090 --> 00:07:36.420
square root of 10/12.

00:07:36.420 --> 00:07:40.390
Again, earlier we know that
10/12 is the variance of the

00:07:40.390 --> 00:07:42.340
sum of xi's.

00:07:42.340 --> 00:07:45.890
And we'll do the same on the
other side, writing it 7 minus

00:07:45.890 --> 00:07:48.930
5 divided by square
root of 10/12.

00:07:52.280 --> 00:07:55.710
Now, if we compute out the
quantity right here, we know

00:07:55.710 --> 00:08:02.550
that this quantity is roughly
2.19, and by the central limit

00:08:02.550 --> 00:08:07.030
theorem, if we believe 10 is a
large enough number, then this

00:08:07.030 --> 00:08:13.390
will be roughly equal to 1 minus
the CDF of a standard

00:08:13.390 --> 00:08:17.200
normal evaluated at 2.19.

00:08:17.200 --> 00:08:20.460
And we could look up the number
in the table, and this

00:08:20.460 --> 00:08:27.910
gives us number roughly,
0.014.

00:08:27.910 --> 00:08:30.580
Now let's do a quick summary of
what this problem is about.

00:08:30.580 --> 00:08:33.720
We're asked to compute the
probability of x greater or

00:08:33.720 --> 00:08:38.909
equal to 7, where x is the sum
of 10 uniform random variables

00:08:38.909 --> 00:08:41.850
between 0 and 1, so
we'll call it xi.

00:08:41.850 --> 00:08:44.560
We know that because each
random variable has

00:08:44.560 --> 00:08:48.390
expectation 1/2, adding 10
of them up, gives us

00:08:48.390 --> 00:08:50.820
expectation of 5.

00:08:50.820 --> 00:08:54.320
So this is essentially asking,
what is the chance that x is

00:08:54.320 --> 00:08:57.600
more than two away from
its expectation?

00:08:57.600 --> 00:09:01.240
So if this is a real line, and
5 is here, maybe x has some

00:09:01.240 --> 00:09:05.180
distribution around 5, so the
center what the expected value

00:09:05.180 --> 00:09:09.650
is at 5, we wonder how likely is
it for us to see something

00:09:09.650 --> 00:09:11.930
greater than 7?

00:09:11.930 --> 00:09:16.430
Now, let's see where do we land
on the probably spectrum

00:09:16.430 --> 00:09:17.890
from 0 to 1.

00:09:17.890 --> 00:09:20.110
Well, without using any
information, we know the

00:09:20.110 --> 00:09:23.710
probability cannot be greater
than 1, so a trivial upper

00:09:23.710 --> 00:09:27.990
bound for the probability
right here will be 1.

00:09:27.990 --> 00:09:31.570
Well, for the first part we use
Markov's Inequality and

00:09:31.570 --> 00:09:35.750
that gives us some number, which
is roughly equal to 0.7.

00:09:35.750 --> 00:09:40.320
In fact, we got number 5/7,
and this is from Markov's

00:09:40.320 --> 00:09:41.420
Inequality.

00:09:41.420 --> 00:09:44.210
Oh, it's better than 1, already
telling us it cannot

00:09:44.210 --> 00:09:48.500
be between 0.7 and 1, but
can we do better?

00:09:48.500 --> 00:09:51.980
Well, the part B, we see that
all the way, using the

00:09:51.980 --> 00:09:54.430
additional information variance,
we can get this

00:09:54.430 --> 00:09:59.510
number down to 5/48, which
is roughly 0.1.

00:09:59.510 --> 00:10:01.910
Already, that's much
better than 0.7.

00:10:01.910 --> 00:10:03.230
Can we even do better?

00:10:03.230 --> 00:10:05.450
And this is the Chebyshev,
and it turns out we

00:10:05.450 --> 00:10:06.960
can indeed do better.

00:10:06.960 --> 00:10:09.390
Using the central limit theorem,
we can squeeze this

00:10:09.390 --> 00:10:14.370
number all the way down to
0.014, almost a 10 times

00:10:14.370 --> 00:10:16.910
improvement over the
previous number.

00:10:16.910 --> 00:10:18.630
This is from central
limit theorem.

00:10:18.630 --> 00:10:22.010
As we can see, by using
different bounding techniques,

00:10:22.010 --> 00:10:25.250
we can progressively improve the
bound on the probability

00:10:25.250 --> 00:10:28.810
of x exceeding 7, and from this
problem we learned that

00:10:28.810 --> 00:10:33.480
even with 10 variables, the
truth is more like this, which

00:10:33.480 --> 00:10:36.850
says that the distribution of
x concentrates very heavily

00:10:36.850 --> 00:10:40.080
around 5, and hence, the
probability of x being greater

00:10:40.080 --> 00:10:43.400
or equal to 7 could be much
smaller than one might expect.