WEBVTT

00:00:01.260 --> 00:00:05.320
By now, we have introduced
all sorts of PMFs for

00:00:05.320 --> 00:00:06.710
the discrete case.

00:00:06.710 --> 00:00:09.990
The joint PMF, the
conditional PMF--

00:00:09.990 --> 00:00:11.670
given an event--

00:00:11.670 --> 00:00:15.450
and the conditional PMF of one
random variable given another.

00:00:15.450 --> 00:00:18.060
And we're moving along with
the program of defining

00:00:18.060 --> 00:00:20.720
analogous concepts for
the continuous case.

00:00:20.720 --> 00:00:24.630
We have already discussed the
joint PDF and the conditional

00:00:24.630 --> 00:00:28.030
PDF, given an event.

00:00:28.030 --> 00:00:32.560
The next item in our menu is to
define a conditional PDF of

00:00:32.560 --> 00:00:36.150
one random variable, given
another random variable.

00:00:36.150 --> 00:00:38.800
We proceed by first looking
at the definition for the

00:00:38.800 --> 00:00:40.120
discrete case.

00:00:40.120 --> 00:00:43.720
A typical entry of the
conditional PMF is just a

00:00:43.720 --> 00:00:47.090
conditional probability, but
in different notation.

00:00:47.090 --> 00:00:49.790
And using the definition of
conditional probabilities,

00:00:49.790 --> 00:00:53.540
this is equal to the ratio of
the joint divided by the

00:00:53.540 --> 00:00:56.610
probability of the conditioning
event.

00:00:56.610 --> 00:00:59.560
Unfortunately, in the continuous
case, a definition

00:00:59.560 --> 00:01:03.140
of this form would be
problematic, because the event

00:01:03.140 --> 00:01:07.150
that Y takes on a specific value
is an event that has 0

00:01:07.150 --> 00:01:07.930
probability.

00:01:07.930 --> 00:01:10.700
And we know that we cannot
condition on a

00:01:10.700 --> 00:01:12.930
0 probability event.

00:01:12.930 --> 00:01:16.830
However, we can take this
expression as a guide on how

00:01:16.830 --> 00:01:20.110
to define a conditional PDF
in the continuous case.

00:01:20.110 --> 00:01:23.530
And this is the definition,
which just mimics the formula

00:01:23.530 --> 00:01:25.430
that we have up here.

00:01:25.430 --> 00:01:29.050
Notice that this conditional
PDF-- defined this way-- is

00:01:29.050 --> 00:01:32.229
well defined, as long
as the denominator

00:01:32.229 --> 00:01:34.670
is a positive quantity.

00:01:34.670 --> 00:01:37.370
Let us now try to make sense
of this definition.

00:01:37.370 --> 00:01:39.789
Let us first recall the
interpretation of the

00:01:39.789 --> 00:01:43.820
conditional PDF, given an event,
A, that has positive

00:01:43.820 --> 00:01:45.320
probability.

00:01:45.320 --> 00:01:48.280
We know that the PDF is used to
determine the probability

00:01:48.280 --> 00:01:49.740
of a small interval.

00:01:49.740 --> 00:01:54.780
And similarly, the conditional
PDF is used to calculate the

00:01:54.780 --> 00:01:58.620
conditional probability of a
small interval given the

00:01:58.620 --> 00:02:00.880
conditioning event.

00:02:00.880 --> 00:02:03.610
We would like to do something
similar for the conditional

00:02:03.610 --> 00:02:07.780
PDF, where we would like to
take the event A to be

00:02:07.780 --> 00:02:12.070
something like the event that Y
is equal to some particular

00:02:12.070 --> 00:02:13.450
value-- little y.

00:02:13.450 --> 00:02:15.780
But as we said, this is
problematic, because this

00:02:15.780 --> 00:02:19.110
event does not have positive
probability.

00:02:19.110 --> 00:02:23.480
So instead, we can take A to
be the event that Y is

00:02:23.480 --> 00:02:26.670
approximately equal to
a certain value.

00:02:26.670 --> 00:02:29.910
So we're dealing with a little
interval around this value,

00:02:29.910 --> 00:02:33.500
little y, which in general would
be an event of positive

00:02:33.500 --> 00:02:34.540
probability.

00:02:34.540 --> 00:02:38.530
And we can try to have a
similar interpretation.

00:02:38.530 --> 00:02:41.140
Let us see how this works out.

00:02:41.140 --> 00:02:45.510
So what does it mean that Y is
approximately equal to some

00:02:45.510 --> 00:02:47.000
particular value, little y?

00:02:47.000 --> 00:02:49.420
We interpret that as follows.

00:02:49.420 --> 00:02:53.050
We're told that the random
variable, Y, takes a value

00:02:53.050 --> 00:02:54.660
that is within epsilon--

00:02:54.660 --> 00:02:56.590
where epsilon is a
small number--

00:02:56.590 --> 00:02:58.780
of a given value, little y.

00:02:58.780 --> 00:03:01.290
And given this conditioning
information, we want to

00:03:01.290 --> 00:03:04.160
calculate the probability
of a small interval.

00:03:04.160 --> 00:03:05.600
How do we do that?

00:03:05.600 --> 00:03:06.370
Well, here--

00:03:06.370 --> 00:03:08.485
because this, in general,
will be a

00:03:08.485 --> 00:03:10.110
positive probability event--

00:03:10.110 --> 00:03:13.920
we can use the definition of
conditional probabilities.

00:03:13.920 --> 00:03:17.900
And it would be equal to the
probability of both events

00:03:17.900 --> 00:03:20.600
happening, divided by the
probability of the

00:03:20.600 --> 00:03:22.820
conditioning event.

00:03:22.820 --> 00:03:25.010
What is the probability of
both events happening?

00:03:25.010 --> 00:03:30.700
This is a probability of a small
rectangle in xy space.

00:03:30.700 --> 00:03:38.570
At that rectangle, the joint
PDF, has a certain value.

00:03:38.570 --> 00:03:41.670
And because we're integrating
over that rectangle--

00:03:41.670 --> 00:03:45.000
and that rectangle has
dimensions delta and epsilon--

00:03:45.000 --> 00:03:48.150
of that probability, that
small rectangle, is

00:03:48.150 --> 00:03:50.690
approximately equal to this.

00:03:50.690 --> 00:03:54.190
Then we need the denominator,
which is the probability of

00:03:54.190 --> 00:03:55.829
the conditioning event.

00:03:55.829 --> 00:04:00.750
And this is approximately equal
to the density of Y

00:04:00.750 --> 00:04:03.350
evaluated at that point,
times the length

00:04:03.350 --> 00:04:05.370
of the small interval.

00:04:05.370 --> 00:04:08.940
We cancel the epsilons.

00:04:08.940 --> 00:04:12.250
And then we notice that the
ratio we have here is what we

00:04:12.250 --> 00:04:16.290
defined as the conditional
PDF.

00:04:16.290 --> 00:04:22.390
So we get this relation
times delta.

00:04:22.390 --> 00:04:25.010
So what do we see?

00:04:25.010 --> 00:04:32.760
We see that the probability of a
small interval is equal to a

00:04:32.760 --> 00:04:36.060
PDF times the length of
the small interval.

00:04:36.060 --> 00:04:40.530
However, because we are
conditioning on Y being

00:04:40.530 --> 00:04:44.520
approximately equal to a certain
value, we end up using

00:04:44.520 --> 00:04:48.060
a corresponding conditional PDF,
where the conditional PDF

00:04:48.060 --> 00:04:49.680
is defined this way.

00:04:49.680 --> 00:04:52.020
So we now have an interpretation
of the

00:04:52.020 --> 00:04:54.470
conditional PDF in terms of

00:04:54.470 --> 00:04:56.575
probabilities of small intervals.

00:04:59.890 --> 00:05:02.100
Now that we have an intuitive
interpretation of the

00:05:02.100 --> 00:05:06.940
conditional PDF, we can also
use it to calculate

00:05:06.940 --> 00:05:09.330
conditional probabilities
of more general

00:05:09.330 --> 00:05:11.860
events, not just intervals.

00:05:11.860 --> 00:05:14.920
And we do this as follows.

00:05:14.920 --> 00:05:17.660
In general, for continuous
random variables, we can find

00:05:17.660 --> 00:05:20.240
the probability that X belongs
to a certain set by

00:05:20.240 --> 00:05:23.340
integrating a PDF
over that set.

00:05:23.340 --> 00:05:26.470
Because here we're dealing with
a conditional situation

00:05:26.470 --> 00:05:30.570
where we're given the value of
Y, we use the conditional PDF

00:05:30.570 --> 00:05:32.340
instead of the true PDF.

00:05:32.340 --> 00:05:35.710
And this way, we calculate the
conditional probability.

00:05:35.710 --> 00:05:39.120
Now, the difficulty is that this
conditional probability

00:05:39.120 --> 00:05:43.409
is not a well-defined quantity
according to what we did early

00:05:43.409 --> 00:05:44.340
on in this class.

00:05:44.340 --> 00:05:46.960
We cannot condition on zero
probability events.

00:05:46.960 --> 00:05:50.270
But we can get the around this
difficulty as follows.

00:05:50.270 --> 00:05:53.580
This quantity is well-defined.

00:05:53.580 --> 00:05:58.640
And we can use this quantity
as the definition of this

00:05:58.640 --> 00:06:00.630
conditional probability.

00:06:00.630 --> 00:06:03.800
And so we have managed to
provide definition of

00:06:03.800 --> 00:06:07.810
conditional probabilities, given
a 0 probability event of

00:06:07.810 --> 00:06:10.250
a certain type.

00:06:10.250 --> 00:06:13.140
It turns out that this
definition is sound and

00:06:13.140 --> 00:06:16.780
consistent with everything
else that we are doing.

00:06:16.780 --> 00:06:20.770
But when we're dealing with
particular problems and

00:06:20.770 --> 00:06:24.420
applications, we can generally
forget about all of these

00:06:24.420 --> 00:06:27.800
subtleties that we have
been discussing here.

00:06:27.800 --> 00:06:29.540
The bottom line is
that we will be

00:06:29.540 --> 00:06:32.190
treating conditional PDFs--

00:06:32.190 --> 00:06:35.480
given the value of a random
variable, Y--

00:06:35.480 --> 00:06:41.980
just as ordinary PDFs, but given
the information that

00:06:41.980 --> 00:06:44.940
this random variable took
on a specific value.

00:06:44.940 --> 00:06:47.640
And in that conditional
universe, we will calculate

00:06:47.640 --> 00:06:51.770
probabilities the usual way,
by using conditional PDFs

00:06:51.770 --> 00:06:53.400
instead of ordinary PDFs.