WEBVTT
00:00:00.950 --> 00:00:05.080
Suppose I have a fair coin which
I toss multiple times.
00:00:05.080 --> 00:00:08.690
I want to model a situation
where the results of previous
00:00:08.690 --> 00:00:13.320
flips do not affect my beliefs
about the likelihood of heads
00:00:13.320 --> 00:00:15.540
in the next flip.
00:00:15.540 --> 00:00:19.280
And I would like to describe
this situation by saying that
00:00:19.280 --> 00:00:22.700
the coin tosses are
independent.
00:00:22.700 --> 00:00:25.480
You may say, we already
defined the notion of
00:00:25.480 --> 00:00:27.060
independent events.
00:00:27.060 --> 00:00:29.210
Doesn't this notion apply?
00:00:29.210 --> 00:00:30.980
Well not quite.
00:00:30.980 --> 00:00:33.660
We defined independence
of two events.
00:00:33.660 --> 00:00:36.590
But here, we want to talk
about independence of a
00:00:36.590 --> 00:00:38.330
collection of events.
00:00:38.330 --> 00:00:41.930
For example, we would like to
say that the events, heads in
00:00:41.930 --> 00:00:45.220
the first toss, heads in the
second toss, heads in the
00:00:45.220 --> 00:00:49.720
third toss, and so on,
are all independent.
00:00:49.720 --> 00:00:52.130
What is the right definition?
00:00:52.130 --> 00:00:54.970
Let us start with intuition.
00:00:54.970 --> 00:01:00.260
We will say that a family of
events are independent if
00:01:00.260 --> 00:01:05.160
knowledge about some of the
events doesn't change my
00:01:05.160 --> 00:01:10.580
beliefs, my probability model,
for the remaining events.
00:01:10.580 --> 00:01:18.770
For example, if I want to say
that events A1, A2 and so on
00:01:18.770 --> 00:01:25.330
are independent, I would like
relations such as the
00:01:25.330 --> 00:01:29.250
following to be true.
00:01:29.250 --> 00:01:36.930
The probability that event A3
happened and A4 does not
00:01:36.930 --> 00:01:42.800
happen remains the same even
if I condition on some
00:01:42.800 --> 00:01:46.220
information about some
other events.
00:01:46.220 --> 00:01:54.020
Let's say if I tell you that
A1 happens or that both A2
00:01:54.020 --> 00:01:58.880
happened and A5 did
not happen.
00:01:58.880 --> 00:02:03.000
The important thing to notice
here is that the indices
00:02:03.000 --> 00:02:07.370
involved in the event of
interest are distinct from the
00:02:07.370 --> 00:02:10.889
indices associated with the
events on which I'm given some
00:02:10.889 --> 00:02:12.040
information.
00:02:12.040 --> 00:02:15.410
I'm given some information about
the events A1, A2, and
00:02:15.410 --> 00:02:18.020
A5, what happened to them.
00:02:18.020 --> 00:02:22.130
And this information does not
affect my beliefs about
00:02:22.130 --> 00:02:25.905
something that has to do
with events A3 and A4.
00:02:28.590 --> 00:02:32.860
I would like all relations
of this kind to be true.
00:02:35.920 --> 00:02:39.650
This could be one possible
definition, just saying that
00:02:39.650 --> 00:02:43.820
the family of events are
independent if and only if any
00:02:43.820 --> 00:02:46.820
relation of this type is true.
00:02:46.820 --> 00:02:50.079
But such a definition would not
be aesthetically pleasing.
00:02:50.079 --> 00:02:53.810
Instead, we introduce the
following definition, which
00:02:53.810 --> 00:02:57.340
mimics or parallels our
earlier definition of
00:02:57.340 --> 00:02:59.320
independence of two events.
00:02:59.320 --> 00:03:03.330
We will say that a collection
of events are independent if
00:03:03.330 --> 00:03:07.280
you can calculate probabilities
of intersections
00:03:07.280 --> 00:03:13.660
of these events by multiplying
individual probabilities.
00:03:13.660 --> 00:03:17.820
And this should be possible
for all choices of indices
00:03:17.820 --> 00:03:23.870
involved and for any number
or events involved.
00:03:23.870 --> 00:03:26.920
Let us translate this into
something concrete.
00:03:26.920 --> 00:03:32.610
Consider the case of three
events, A1, A2, and A3.
00:03:32.610 --> 00:03:38.180
Our definition requires that
we can calculate the
00:03:38.180 --> 00:03:41.690
probability of the intersection
of two events by
00:03:41.690 --> 00:03:45.770
multiplying individual
probabilities.
00:03:45.770 --> 00:03:49.320
And we would like all of these
three relations to be true,
00:03:49.320 --> 00:03:52.900
because this property should
be true for any
00:03:52.900 --> 00:03:55.970
choice of the indices.
00:03:55.970 --> 00:03:57.860
What do we have here?
00:03:57.860 --> 00:04:02.740
This relation tells us that
A1 and A2 are independent.
00:04:02.740 --> 00:04:06.440
This relation tells us that
A1 and A3 are independent.
00:04:06.440 --> 00:04:10.780
This relation tells us that
A2 and A3 are independent.
00:04:10.780 --> 00:04:16.110
We call this situation pairwise
independence.
00:04:16.110 --> 00:04:18.589
But the definition requires
something more.
00:04:18.589 --> 00:04:21.420
It requires that the probability
of three-way
00:04:21.420 --> 00:04:25.890
intersections can also be
calculated the same way by
00:04:25.890 --> 00:04:29.040
multiplying individual
probabilities.
00:04:29.040 --> 00:04:33.590
And this additional condition
does make a difference, as
00:04:33.590 --> 00:04:37.470
we're going to see in
a later example.
00:04:37.470 --> 00:04:41.210
Is this the right definition?
00:04:41.210 --> 00:04:42.210
Yes.
00:04:42.210 --> 00:04:46.170
One can prove formally that
if the conditions in this
00:04:46.170 --> 00:04:51.370
definition are satisfied,
then any formula of
00:04:51.370 --> 00:04:53.700
this kind is true.
00:04:53.700 --> 00:04:55.670
In particular, we also have
00:04:55.670 --> 00:04:58.230
relations such as the following.
00:04:58.230 --> 00:05:05.010
The probability of event A3 is
the same as the probability of
00:05:05.010 --> 00:05:11.180
event A3, given that
A1 and A2 occurred.
00:05:11.180 --> 00:05:16.270
Or the probability of
A3, given that A1
00:05:16.270 --> 00:05:19.530
occurred but A2 didn't.
00:05:19.530 --> 00:05:23.500
Or we can continue this
similarly, the probability of
00:05:23.500 --> 00:05:27.790
A3 given that A1 did
not occur, and A2
00:05:27.790 --> 00:05:29.790
occurred, and so on.
00:05:29.790 --> 00:05:32.930
So any kind of information that
I might give you about
00:05:32.930 --> 00:05:34.530
events A1 and A2--
00:05:34.530 --> 00:05:37.390
which one of them occurred
and which one didn't--
00:05:37.390 --> 00:05:42.090
is not going to affect my
beliefs about the event A3.
00:05:42.090 --> 00:05:44.600
The conditional probabilities
are going to be the same as
00:05:44.600 --> 00:05:47.850
the unconditional
probabilities.
00:05:47.850 --> 00:05:50.940
I told you that this definition
implies that all
00:05:50.940 --> 00:05:52.500
relations of this kind [are]
00:05:52.500 --> 00:05:53.200
true.
00:05:53.200 --> 00:05:54.550
This can be proved.
00:05:54.550 --> 00:05:56.970
The proof is a bit tedious.
00:05:56.970 --> 00:05:58.740
And we will not go through it.