WEBVTT

00:00:00.500 --> 00:00:03.580
We will now go through a
beautiful example, in which we

00:00:03.580 --> 00:00:08.150
approach the same question
in a number of different ways

00:00:08.150 --> 00:00:11.300
and see that by reasoning
based on the intuitive

00:00:11.300 --> 00:00:13.630
properties of a
Poisson process, we

00:00:13.630 --> 00:00:16.750
can arrive quickly
to the right answer.

00:00:16.750 --> 00:00:18.130
The problem is as follows.

00:00:18.130 --> 00:00:21.080
We have three lightbulbs,
and each light bulb

00:00:21.080 --> 00:00:24.920
is being lit at time
zero, it starts working,

00:00:24.920 --> 00:00:28.010
and the light bulb lasts
for a certain amount

00:00:28.010 --> 00:00:29.630
of time, which is random.

00:00:29.630 --> 00:00:33.090
So this light bulb lasts so
long, this one lasts so long,

00:00:33.090 --> 00:00:36.080
this one lasts that long.

00:00:36.080 --> 00:00:40.090
The lifetime of a light bulb,
the time until it burns out,

00:00:40.090 --> 00:00:42.380
will be a random
variable, and we

00:00:42.380 --> 00:00:44.220
make the following assumptions.

00:00:44.220 --> 00:00:46.670
The lifetimes of the
three light bulbs,

00:00:46.670 --> 00:00:50.100
which we denote by
X, Y, and Z, will

00:00:50.100 --> 00:00:52.880
be independent random
variables, each of which

00:00:52.880 --> 00:00:55.760
is an exponential random
variable with the parameter

00:00:55.760 --> 00:00:56.680
lambda.

00:00:56.680 --> 00:00:58.880
We're interested in the
question of calculating

00:00:58.880 --> 00:01:02.430
the expected time until
a light bulb burns out

00:01:02.430 --> 00:01:03.680
for the first time.

00:01:03.680 --> 00:01:06.870
So in this picture, the light
bulb that burned out first

00:01:06.870 --> 00:01:09.920
is the second light
bulb, and this quantity

00:01:09.920 --> 00:01:12.740
is the time until
the first burnout.

00:01:12.740 --> 00:01:15.270
How do we calculate
this quantity?

00:01:15.270 --> 00:01:18.200
The time until a first
light bulb burns out

00:01:18.200 --> 00:01:22.800
is the minimum of the times
at which each one of them

00:01:22.800 --> 00:01:24.370
burns out.

00:01:24.370 --> 00:01:26.490
So we're interested
in the expected

00:01:26.490 --> 00:01:28.560
value of this quantity.

00:01:28.560 --> 00:01:31.030
It's a random variable,
which is a function of three

00:01:31.030 --> 00:01:32.440
random variables.

00:01:32.440 --> 00:01:34.740
How do we calculate
the expected value

00:01:34.740 --> 00:01:36.740
of a function of
random variables?

00:01:36.740 --> 00:01:39.370
We can use the
expected value rule.

00:01:39.370 --> 00:01:42.410
Let us take the
function of interest,

00:01:42.410 --> 00:01:45.210
which is the minimum
of three numbers.

00:01:45.210 --> 00:01:50.370
Then we need to multiply by
the joint PDF or these three

00:01:50.370 --> 00:01:52.970
random variables
X, Y and Z. Now,

00:01:52.970 --> 00:01:55.570
because these three random
variables are independent,

00:01:55.570 --> 00:02:00.130
the joint PDF is the product
of their individual PDFs,

00:02:00.130 --> 00:02:02.230
which are all exponential.

00:02:02.230 --> 00:02:06.440
And so we obtain this
expression for the joint PDF.

00:02:06.440 --> 00:02:12.370
And we need to integrate this
over all x's, y's, and z's.

00:02:12.370 --> 00:02:14.290
So it's going to
be an integral that

00:02:14.290 --> 00:02:19.730
goes for each one of the three
variables from 0 to infinity.

00:02:19.730 --> 00:02:24.650
And here we have dx, dy, dz.

00:02:24.650 --> 00:02:27.270
So this is one approach
that can give us

00:02:27.270 --> 00:02:31.980
the answer if you're able
to manipulate and keep

00:02:31.980 --> 00:02:36.370
track of everything that happens
in this three-dimensional

00:02:36.370 --> 00:02:37.310
integral.

00:02:37.310 --> 00:02:41.770
But this is too tedious and
this is not a good way to go.

00:02:41.770 --> 00:02:44.329
Let us try to think
of an alternative way.

00:02:44.329 --> 00:02:46.990
Can we figure out
the distribution

00:02:46.990 --> 00:02:48.891
of this random variable?

00:02:48.891 --> 00:02:51.390
It's a function of three random
variables, so in some sense,

00:02:51.390 --> 00:02:53.530
it's a derived
distribution problem,

00:02:53.530 --> 00:02:57.320
so we can try to calculate
the cumulative distribution

00:02:57.320 --> 00:03:00.650
function of the
minimum of the three.

00:03:00.650 --> 00:03:05.350
So for the cumulative, we would
be looking at the probability

00:03:05.350 --> 00:03:10.160
that the minimum of these three
random variables is less than

00:03:10.160 --> 00:03:12.220
or equal to a certain number.

00:03:12.220 --> 00:03:14.400
It turns out that
the calculations

00:03:14.400 --> 00:03:17.570
are a little faster if we
look at the probability

00:03:17.570 --> 00:03:21.560
that this is larger than or
equal to a certain number.

00:03:21.560 --> 00:03:23.130
What is this event?

00:03:23.130 --> 00:03:26.060
The minimum is larger
than or equal to t,

00:03:26.060 --> 00:03:29.660
if, and only if,
all three of them

00:03:29.660 --> 00:03:32.000
are larger than or equal to t.

00:03:32.000 --> 00:03:34.930
So the probability of this event
is the same as the probability

00:03:34.930 --> 00:03:37.210
that X is larger
than or equal to t,

00:03:37.210 --> 00:03:41.910
Y larger than or equal to t,
and Z larger than or equal to t.

00:03:41.910 --> 00:03:45.110
But now X, Y, and
Z are independent,

00:03:45.110 --> 00:03:48.230
so we need to multiply
the probability that X

00:03:48.230 --> 00:03:50.570
is larger than or
equal to t, which

00:03:50.570 --> 00:03:52.170
for an exponential
random variable

00:03:52.170 --> 00:03:55.250
is e to the minus lambda
t, then the probability

00:03:55.250 --> 00:03:59.060
of the second event which is
again e to the minus lambda t,

00:03:59.060 --> 00:04:01.630
and, finally, the probability
of the third event, which

00:04:01.630 --> 00:04:04.010
is, again, e to
the minus lambda t,

00:04:04.010 --> 00:04:06.620
which is e to the
minus 3 lambda t.

00:04:10.730 --> 00:04:13.630
Now, we look at this
expression for the probability

00:04:13.630 --> 00:04:15.740
that the random variable
is larger than or equal

00:04:15.740 --> 00:04:19.790
to t and recognize that this
is the expression that we have

00:04:19.790 --> 00:04:25.050
when we have an exponential
random variable with parameter

00:04:25.050 --> 00:04:27.190
equal to 3 lambda.

00:04:27.190 --> 00:04:30.680
If you want to do it formally,
subtract this quantity from 1

00:04:30.680 --> 00:04:33.210
to find the CDF,
take the derivative,

00:04:33.210 --> 00:04:35.820
and you will find
an exponential PDF.

00:04:35.820 --> 00:04:39.700
So the conclusion is
that this random variable

00:04:39.700 --> 00:04:43.260
is exponential with
parameter 3 lambda.

00:04:46.000 --> 00:04:49.250
And since it's an exponential
with parameter 3 lambda,

00:04:49.250 --> 00:04:52.070
then we know what the
expected value is.

00:04:52.070 --> 00:04:55.270
It is going to be
1 over 3 lambda.

00:04:55.270 --> 00:04:57.540
And this is the
answer to the question

00:04:57.540 --> 00:05:00.350
that we were interested in.

00:05:00.350 --> 00:05:03.630
Now, the fact that this is an
exponential random variable,

00:05:03.630 --> 00:05:08.380
but with a different parameter,
is a pretty clean fact,

00:05:08.380 --> 00:05:12.190
and so it should have
a good explanation.

00:05:12.190 --> 00:05:18.930
Let us now try to think about a
good explanation for this fact.

00:05:18.930 --> 00:05:22.660
Whenever we deal with an
exponential random variable,

00:05:22.660 --> 00:05:25.120
one way of thinking
about it is that

00:05:25.120 --> 00:05:29.370
this exponential random variable
is the first time in a Poisson

00:05:29.370 --> 00:05:30.300
process.

00:05:30.300 --> 00:05:33.040
So imagine that there's
a Poisson process that

00:05:33.040 --> 00:05:38.400
runs forever, and X is
the first arrival time.

00:05:38.400 --> 00:05:41.180
For this light bulb, we
can think the same way,

00:05:41.180 --> 00:05:44.650
and since X and Y are
assumed to be independent,

00:05:44.650 --> 00:05:48.010
we can assume that here we
have an independent Poisson

00:05:48.010 --> 00:05:50.360
process, independent
from the first one,

00:05:50.360 --> 00:05:53.690
it has its own arrival times,
and Y is the first arrival

00:05:53.690 --> 00:05:55.640
time in this Poisson process.

00:05:55.640 --> 00:05:59.040
And finally, we have a third
independent Poisson process,

00:05:59.040 --> 00:06:02.180
and the random variable
Z is the first arrival

00:06:02.180 --> 00:06:04.290
time in that Poisson process.

00:06:04.290 --> 00:06:08.010
So X, Y and Z are
interpreted as first arrivals

00:06:08.010 --> 00:06:11.040
in three independent
Poisson processes.

00:06:11.040 --> 00:06:14.200
Now, let us take these
three Poisson processes

00:06:14.200 --> 00:06:15.540
and merge them.

00:06:15.540 --> 00:06:18.750
If we merge these three
processes, what we obtain

00:06:18.750 --> 00:06:22.890
is a merged process, which
is Poisson with parameter

00:06:22.890 --> 00:06:27.500
equal to the sum of the rates
or parameters of each one

00:06:27.500 --> 00:06:29.820
of the processes, so
it's a Poisson process

00:06:29.820 --> 00:06:32.450
with parameter 3 lambda.

00:06:32.450 --> 00:06:37.280
Now, how can we interpret the
random variable of interest,

00:06:37.280 --> 00:06:41.420
the first burnout time, in
terms of the merged process?

00:06:41.420 --> 00:06:44.200
So the merged process
has an arrival

00:06:44.200 --> 00:06:47.680
whenever one of those three
processes has an arrival.

00:06:47.680 --> 00:06:50.570
The first arrival in
the merged process

00:06:50.570 --> 00:06:55.390
will happen the first time that
one of these three processes

00:06:55.390 --> 00:06:57.570
is going to have an arrival.

00:06:57.570 --> 00:07:00.990
Therefore, we can interpret
the random variable

00:07:00.990 --> 00:07:03.180
of interest, the
first burnout time,

00:07:03.180 --> 00:07:06.650
as the first arrival
time in a merged process.

00:07:06.650 --> 00:07:10.045
But now the merged process
is Poisson with parameter 3

00:07:10.045 --> 00:07:13.740
lambda, therefore,
this random variable

00:07:13.740 --> 00:07:18.100
is going to be an exponential
random variable with parameter

00:07:18.100 --> 00:07:19.830
3 lambda.

00:07:19.830 --> 00:07:23.010
And from this, we also
obtain the expected value

00:07:23.010 --> 00:07:25.890
of that random variable.

00:07:25.890 --> 00:07:30.270
The beauty of this last approach
for coming up with the answer

00:07:30.270 --> 00:07:33.390
by reasoning in terms of
merged Poisson processes

00:07:33.390 --> 00:07:37.540
is that we didn't have to do
any calculations at all, just

00:07:37.540 --> 00:07:40.960
use the intuitive understanding
of Poisson processes

00:07:40.960 --> 00:07:43.010
and, especially,
the interpretation

00:07:43.010 --> 00:07:46.830
of an exponential random
variable as the first arrival

00:07:46.830 --> 00:07:50.400
time in a Poisson process.

00:07:50.400 --> 00:07:54.190
Let us now try to solve a
somewhat harder problem.

00:07:54.190 --> 00:07:56.750
Let us try to calculate
the expected time

00:07:56.750 --> 00:08:01.430
until all the light
bulbs burn out.

00:08:01.430 --> 00:08:05.540
So one light bulb will burn out,
then another one will burn out,

00:08:05.540 --> 00:08:08.110
and, finally, the third
one will burn out.

00:08:08.110 --> 00:08:11.700
We want to find the expected
time until this happens.

00:08:11.700 --> 00:08:15.720
Once more, we will be thinking
of these burnout times

00:08:15.720 --> 00:08:19.690
as the first arrival times
in Poisson processes.

00:08:22.560 --> 00:08:25.780
The total time until
the third burnout

00:08:25.780 --> 00:08:29.860
can be split into
different periods.

00:08:29.860 --> 00:08:33.808
There's a time until one
light bulb burns out.

00:08:33.808 --> 00:08:37.058
And the expected
value of this period

00:08:37.058 --> 00:08:40.120
here is going to
be 1 over 3 lambda.

00:08:43.049 --> 00:08:45.540
What happens at this time?

00:08:45.540 --> 00:08:50.300
The second light bulb has burned
out, so we can forget about it,

00:08:50.300 --> 00:08:52.960
take it out of the picture.

00:08:52.960 --> 00:08:55.830
We have two lightbulbs.

00:08:55.830 --> 00:08:58.390
Let us look at the
time it will take

00:08:58.390 --> 00:09:02.610
until one of these two
light bulbs burns out.

00:09:02.610 --> 00:09:07.110
So we're interested in
this period of time.

00:09:07.110 --> 00:09:11.370
Now, the Poisson process
starts fresh at this time.

00:09:11.370 --> 00:09:13.940
After this time,
whatever happens

00:09:13.940 --> 00:09:16.640
is just an ordinary
Poisson process

00:09:16.640 --> 00:09:19.350
as if it were
starting at this time.

00:09:19.350 --> 00:09:23.630
So this is going to be an
exponential random variable

00:09:23.630 --> 00:09:24.930
starting from this time.

00:09:24.930 --> 00:09:28.680
And this is going to be another
exponential random variable.

00:09:28.680 --> 00:09:32.670
So the time until the next light
bulb burns out in this case

00:09:32.670 --> 00:09:34.410
is going to be
the minimum of two

00:09:34.410 --> 00:09:36.370
exponential random variables.

00:09:36.370 --> 00:09:39.300
We can think again about
merging these two Poisson

00:09:39.300 --> 00:09:42.475
processes to obtain a Poisson
process with total rate 2

00:09:42.475 --> 00:09:46.450
lambda, and the time until
one of these two turns out

00:09:46.450 --> 00:09:50.670
is going to be the first arrival
time in that merged process.

00:09:50.670 --> 00:09:54.775
And so the expected time
until the first arrival

00:09:54.775 --> 00:09:59.730
of the merged process is
going to be 1 over 2 lambda.

00:09:59.730 --> 00:10:03.540
And finally, once this
burnout has happened,

00:10:03.540 --> 00:10:06.160
we can forget about
this light bulb.

00:10:06.160 --> 00:10:10.200
We're left just with one light
bulb, and starting from here,

00:10:10.200 --> 00:10:13.130
we wait until that
light bulb burns out.

00:10:13.130 --> 00:10:16.480
Once more, because of the fresh
start property of the Poisson

00:10:16.480 --> 00:10:19.960
process, starting from
here until it burns out

00:10:19.960 --> 00:10:22.010
is going to be a
random variable, which

00:10:22.010 --> 00:10:24.210
is an exponential
random variable.

00:10:24.210 --> 00:10:27.150
And in this case, an exponential
random variable with rate

00:10:27.150 --> 00:10:28.620
just lambda.

00:10:28.620 --> 00:10:31.100
And by adding these
three quantities,

00:10:31.100 --> 00:10:38.440
we get the expected time until
all three have burned out.

00:10:38.440 --> 00:10:39.970
This is a problem
that would have

00:10:39.970 --> 00:10:44.960
been quite hard to solve
in a more analytical way.

00:10:44.960 --> 00:10:47.670
We're dealing with a random
variable, which is now

00:10:47.670 --> 00:10:52.450
the maximum of X, Y, and
Z. And the distribution

00:10:52.450 --> 00:10:56.770
of this random variable is
not so simple to write down.

00:10:56.770 --> 00:11:00.130
So that would not be
a very good approach

00:11:00.130 --> 00:11:02.430
for going about this problem.

00:11:02.430 --> 00:11:04.930
But we managed to find
the expected value

00:11:04.930 --> 00:11:08.030
of this random variable
without having to write down

00:11:08.030 --> 00:11:12.020
its distribution, by
breaking this random variable

00:11:12.020 --> 00:11:17.650
into a sum of three particular
random variables, each of which

00:11:17.650 --> 00:11:20.490
had a nice intuitive
interpretation.

00:11:20.490 --> 00:11:24.580
And that was the key to the
solution of this problem.