WEBVTT
00:00:00.650 --> 00:00:03.990
In this segment, we introduce
the concept of continuous
00:00:03.990 --> 00:00:07.230
random variables and their
characterization in terms of
00:00:07.230 --> 00:00:11.100
probability density functions,
or PDFs for short.
00:00:11.100 --> 00:00:13.810
Let us first go back to discrete
random variables.
00:00:13.810 --> 00:00:15.800
A discrete random variable
takes values
00:00:15.800 --> 00:00:17.130
in a discrete set.
00:00:17.130 --> 00:00:20.430
There is a total of one unit of
probability assigned to the
00:00:20.430 --> 00:00:21.780
possible values.
00:00:21.780 --> 00:00:26.080
And the PMF tells us exactly how
much of this probability
00:00:26.080 --> 00:00:28.100
is assigned to each value.
00:00:28.100 --> 00:00:33.590
So we can think of the bars in
the PMF as point masses with
00:00:33.590 --> 00:00:37.000
positive weight that
sit on top of each
00:00:37.000 --> 00:00:38.870
possible numerical value.
00:00:38.870 --> 00:00:41.440
And we can calculate the
probability that the random
00:00:41.440 --> 00:00:45.720
variable falls inside an
interval by adding all the
00:00:45.720 --> 00:00:49.270
masses that sit on top
of that interval.
00:00:49.270 --> 00:00:52.890
So for example, if we're looking
at the interval from a
00:00:52.890 --> 00:00:57.730
to b, the probability of this
interval is equal to the sum
00:00:57.730 --> 00:01:01.530
of the probabilities of these
three masses that fall inside
00:01:01.530 --> 00:01:02.870
this interval.
00:01:02.870 --> 00:01:05.690
On the other hand, a continuous
random variable
00:01:05.690 --> 00:01:08.860
will be taking values over
a continuous range--
00:01:08.860 --> 00:01:12.800
for example, the real line or an
interval on the real line.
00:01:12.800 --> 00:01:18.000
In this case, we still have one
total unit of probability
00:01:18.000 --> 00:01:21.350
mass that is assigned to the
possible values of the random
00:01:21.350 --> 00:01:25.390
variable, except that this unit
of mass is spread all
00:01:25.390 --> 00:01:26.860
over the real line.
00:01:26.860 --> 00:01:29.990
But it is not spread in
a uniform manner.
00:01:29.990 --> 00:01:32.210
Some parts of the real
line have more
00:01:32.210 --> 00:01:33.880
mass per unit length.
00:01:33.880 --> 00:01:35.670
Some have less.
00:01:35.670 --> 00:01:39.900
How much mass exactly is sitting
on top of each part of
00:01:39.900 --> 00:01:43.090
the real line is described by
the probability density
00:01:43.090 --> 00:01:47.380
function, this function plotted
here, which we denote
00:01:47.380 --> 00:01:48.840
with this notation.
00:01:48.840 --> 00:01:52.090
The letter f will always
indicate that we are dealing
00:01:52.090 --> 00:01:53.440
with a PDF.
00:01:53.440 --> 00:01:57.310
And the subscript will indicate
which random variable
00:01:57.310 --> 00:02:00.190
we're talking about.
00:02:00.190 --> 00:02:03.210
We use the probability density
function to calculate the
00:02:03.210 --> 00:02:07.020
probability that X lies in
a certain interval--
00:02:07.020 --> 00:02:10.560
let's say the interval
from a to b.
00:02:10.560 --> 00:02:16.079
And we calculate it by finding
the area under the PDF that
00:02:16.079 --> 00:02:19.900
sits on top of that interval.
00:02:19.900 --> 00:02:23.300
So this area here, the shaded
area, is the probability that
00:02:23.300 --> 00:02:26.100
X stakes values in
this interval.
00:02:26.100 --> 00:02:28.990
Think of probability
as snow fall.
00:02:28.990 --> 00:02:31.770
There is one pound of snow
that has fallen on
00:02:31.770 --> 00:02:34.500
top of the real line.
00:02:34.500 --> 00:02:39.160
The PDF tells us the height of
the snow accumulated over a
00:02:39.160 --> 00:02:41.329
particular point.
00:02:41.329 --> 00:02:47.190
We then find the weight of the
overall amount of snow sitting
00:02:47.190 --> 00:02:51.079
on top of an interval by
calculating the area under
00:02:51.079 --> 00:02:52.400
this curve.
00:02:52.400 --> 00:02:56.750
Of course, mathematically, area
under the curve is just
00:02:56.750 --> 00:02:57.720
an integral.
00:02:57.720 --> 00:03:01.320
So the probability that X takes
values in this interval
00:03:01.320 --> 00:03:07.790
is the integral of the PDF over
this particular interval.
00:03:07.790 --> 00:03:10.320
What properties should
the PDF have?
00:03:10.320 --> 00:03:14.780
By analogy with the discrete
case, a PDF must be
00:03:14.780 --> 00:03:18.060
non-negative, because we do
not want to get negative
00:03:18.060 --> 00:03:20.290
probabilities.
00:03:20.290 --> 00:03:24.870
In the discrete case, the sum
of the PMF entries has to be
00:03:24.870 --> 00:03:26.680
equal to 1.
00:03:26.680 --> 00:03:31.240
In the continuous case, X is
certain to lie in the interval
00:03:31.240 --> 00:03:34.200
between minus infinity
and plus infinity.
00:03:34.200 --> 00:03:39.260
So letting a be minus infinity
and b plus infinity, we should
00:03:39.260 --> 00:03:41.680
get a probability of 1.
00:03:41.680 --> 00:03:46.180
So the total area under the PDF,
when we integrate over
00:03:46.180 --> 00:03:49.780
the entire real line, should
be equal to 1.
00:03:49.780 --> 00:03:55.140
These two conditions are all
that we need in order to have
00:03:55.140 --> 00:03:58.300
a legitimate PDF.
00:03:58.300 --> 00:04:01.280
We can now give a formal
definition of what a
00:04:01.280 --> 00:04:03.770
continuous random variable is.
00:04:03.770 --> 00:04:08.270
A continuous random variable
is a random variable whose
00:04:08.270 --> 00:04:14.380
probabilities can be described
by a PDF according to a
00:04:14.380 --> 00:04:17.209
formula of this type.
00:04:17.209 --> 00:04:19.899
An important point--
00:04:19.899 --> 00:04:22.550
the fact that a random variable
takes values on a
00:04:22.550 --> 00:04:26.980
continuous set is not enough
to make it what we call a
00:04:26.980 --> 00:04:29.180
continuous random variable.
00:04:29.180 --> 00:04:30.910
For a continuous random
variable, we're
00:04:30.910 --> 00:04:32.730
asking for a bit more--
00:04:32.730 --> 00:04:36.630
that it can be described by a
PDF, that a formula of this
00:04:36.630 --> 00:04:37.880
type is valid.
00:04:40.760 --> 00:04:44.240
Now, once we have the
probabilities of intervals as
00:04:44.240 --> 00:04:48.380
given by a PDF, we can use of
additivity to calculate the
00:04:48.380 --> 00:04:51.340
probabilities of more
complicated sets.
00:04:51.340 --> 00:04:54.520
For example, if you're
interested in the probability
00:04:54.520 --> 00:05:06.050
that X lies between 1 and 3 or
that X lies between 4 and 5--
00:05:06.050 --> 00:05:11.250
so this is the probability that
X falls in a region that
00:05:11.250 --> 00:05:14.940
consists of two disjoint
intervals.
00:05:14.940 --> 00:05:19.790
We find the probability of the
union of these two intervals,
00:05:19.790 --> 00:05:24.020
by additivity, by adding the
probabilities of the two
00:05:24.020 --> 00:05:29.480
intervals, since these intervals
are disjoint.
00:05:29.480 --> 00:05:35.840
And then we can use the PDF to
calculate the probabilities of
00:05:35.840 --> 00:05:39.380
each one of these intervals
according to this formula.
00:05:42.570 --> 00:05:45.780
At this point, you may be
wondering what happened to the
00:05:45.780 --> 00:05:49.120
sample space in all
this discussion.
00:05:49.120 --> 00:05:53.070
Well, there is still an
underlying sample space
00:05:53.070 --> 00:05:54.320
lurking in the background.
00:05:57.800 --> 00:06:01.860
And different outcomes in the
sample space result in
00:06:01.860 --> 00:06:05.670
different numerical
values for the
00:06:05.670 --> 00:06:07.135
random variable of interest.
00:06:09.790 --> 00:06:13.650
And when we talk about the
probability that X takes
00:06:13.650 --> 00:06:18.050
values between some numbers a
and b, what we really mean is
00:06:18.050 --> 00:06:23.780
the probability of those
outcomes for which the
00:06:23.780 --> 00:06:27.430
resulting value of
X lies inside
00:06:27.430 --> 00:06:29.800
this particular interval.
00:06:29.800 --> 00:06:31.720
So that's what probability
means.
00:06:31.720 --> 00:06:35.659
On the other hand, once we have
a PDF in our hands, we
00:06:35.659 --> 00:06:39.290
can completely forget about the
underlying sample space.
00:06:39.290 --> 00:06:42.890
And we can carry out any
calculations we may be
00:06:42.890 --> 00:06:46.680
interested in by just working
with the PDF.
00:06:46.680 --> 00:06:50.600
So as we move on in this course,
the sample space will
00:06:50.600 --> 00:06:52.400
be moved offstage.
00:06:52.400 --> 00:06:54.470
There will be less and
less mention of it.
00:06:54.470 --> 00:06:58.159
And we will be working just
directly with PDFs or with
00:06:58.159 --> 00:07:03.270
PMFs if we are dealing with
discrete random variables.
00:07:03.270 --> 00:07:06.470
Let us now build a little bit
on our understanding of what
00:07:06.470 --> 00:07:08.580
PDFs really are by looking at
00:07:08.580 --> 00:07:11.110
probabilities of small intervals.
00:07:11.110 --> 00:07:16.660
Let us look at an interval that
starts at some a and goes
00:07:16.660 --> 00:07:20.340
up to some number
a plus delta.
00:07:20.340 --> 00:07:23.410
So here, delta is a
positive number.
00:07:23.410 --> 00:07:25.270
But we're interested
in the case where
00:07:25.270 --> 00:07:28.150
delta is very small.
00:07:28.150 --> 00:07:33.750
Let us look at the probability
that X falls in this interval.
00:07:33.750 --> 00:07:40.640
The probability that X lies
inside this interval is the
00:07:40.640 --> 00:07:42.390
area of this region.
00:07:45.650 --> 00:07:50.630
On the other hand, as long as
f does not change too much
00:07:50.630 --> 00:07:53.590
over this little interval, which
will be the case if we
00:07:53.590 --> 00:07:59.520
have a continuous density f,
then we can approximate the
00:07:59.520 --> 00:08:04.670
area we have of this region by
the area of a rectangle where
00:08:04.670 --> 00:08:08.130
we keep the height constant.
00:08:08.130 --> 00:08:14.800
The area of this rectangle is
equal to the height, which is
00:08:14.800 --> 00:08:21.360
the value of the PDF at the
point a, times the base of the
00:08:21.360 --> 00:08:24.360
rectangle, which is
equal to delta.
00:08:24.360 --> 00:08:27.560
So this gives us an
interpretation of PDFs in
00:08:27.560 --> 00:08:30.940
terms of probabilities
of small intervals.
00:08:30.940 --> 00:08:34.370
If we take this factor of delta
and send it to the other
00:08:34.370 --> 00:08:37.960
side in this approximate
equality, we see that the
00:08:37.960 --> 00:08:43.308
value of the PDF can be
interpreted as probability per
00:08:43.308 --> 00:08:44.760
unit length.
00:08:44.760 --> 00:08:47.350
So PDFs are not probabilities.
00:08:47.350 --> 00:08:48.750
They are densities.
00:08:48.750 --> 00:08:52.770
Their units are probability
per unit length.
00:08:52.770 --> 00:08:57.710
Now, if the probability per unit
length is finite and the
00:08:57.710 --> 00:09:04.410
length delta is sent to 0, we
will get 0 probability.
00:09:04.410 --> 00:09:09.690
More formally, if we look at
this integral and we let b to
00:09:09.690 --> 00:09:16.200
be the same as a, then we obtain
the probability that X
00:09:16.200 --> 00:09:19.150
is equal to a.
00:09:19.150 --> 00:09:21.620
And on that side, we
get an integral
00:09:21.620 --> 00:09:24.176
over a 0 length interval.
00:09:24.176 --> 00:09:26.800
And that integral is
going to be 0.
00:09:26.800 --> 00:09:30.960
So we obtain that the
probability that X takes a
00:09:30.960 --> 00:09:35.900
value equal to a specific,
particular point--
00:09:35.900 --> 00:09:39.220
that probability is going
to be equal to 0.
00:09:39.220 --> 00:09:42.520
So for a continuous random
variable, any particular point
00:09:42.520 --> 00:09:44.710
has 0 probability.
00:09:44.710 --> 00:09:50.280
Yet somehow, collectively, the
infinitely many points in an
00:09:50.280 --> 00:09:55.070
interval together will have
positive probablility.
00:09:55.070 --> 00:09:56.650
Is this a puzzle?
00:09:56.650 --> 00:09:57.260
Not really.
00:09:57.260 --> 00:09:59.890
That's exactly what happens,
also, with the
00:09:59.890 --> 00:10:02.050
ordinary notion of length.
00:10:02.050 --> 00:10:06.590
Single points have 0 length, yet
by putting together lots
00:10:06.590 --> 00:10:14.030
of points, we can create a set
that has a positive length.
00:10:14.030 --> 00:10:17.680
And a final consequence of the
fact that individual points
00:10:17.680 --> 00:10:20.650
have 0 length.
00:10:20.650 --> 00:10:24.380
Using the additivity axiom,
the probability that our
00:10:24.380 --> 00:10:29.680
random variable takes values
inside an interval is equal to
00:10:29.680 --> 00:10:33.150
the probability that our random
variable takes a value
00:10:33.150 --> 00:10:37.780
of a plus the probability that
our random variable takes a
00:10:37.780 --> 00:10:42.470
value of b plus the probability
that our random
00:10:42.470 --> 00:10:47.390
variable is strictly
between a and b.
00:10:47.390 --> 00:10:54.030
According to our discussion,
this term is equal to 0.
00:10:54.030 --> 00:10:56.230
And this term is equal to 0.
00:10:56.230 --> 00:10:59.070
And so we conclude that the
probability of a closed
00:10:59.070 --> 00:11:00.650
interval is the same as the
00:11:00.650 --> 00:11:03.100
probability of an open interval.
00:11:03.100 --> 00:11:05.340
When calculating probabilities,
it does not
00:11:05.340 --> 00:11:08.300
matter whether we include
the endpoints or not.