WEBVTT

00:00:00.530 --> 00:00:03.500
This segment is probably
the most critical one

00:00:03.500 --> 00:00:07.040
for the purpose of understanding
what the Poisson process really

00:00:07.040 --> 00:00:09.260
is and how it behaves.

00:00:09.260 --> 00:00:12.050
There will be almost no
mathematical formulas.

00:00:12.050 --> 00:00:14.560
But the segment
will be quite dense

00:00:14.560 --> 00:00:16.960
in terms of
conceptual reasoning.

00:00:16.960 --> 00:00:20.100
So pay a lot of attention.

00:00:20.100 --> 00:00:24.360
In a nutshell, we will argue
that the Poisson process has

00:00:24.360 --> 00:00:28.470
memorylessness properties that
are entirely similar to those

00:00:28.470 --> 00:00:30.862
that we have seen for
the Bernoulli process.

00:00:30.862 --> 00:00:34.190
This should not be surprising,
since the Poisson process can

00:00:34.190 --> 00:00:37.710
be thought of as a limiting
case of the Bernoulli process.

00:00:37.710 --> 00:00:39.660
We will reason through
these properties,

00:00:39.660 --> 00:00:42.610
not in the style of a
formal mathematical proof,

00:00:42.610 --> 00:00:44.960
but with an intuitive argument.

00:00:44.960 --> 00:00:48.240
But I would like to assure you
that the intuitive argument can

00:00:48.240 --> 00:00:52.690
be translated into
a rigorous proof.

00:00:52.690 --> 00:00:55.050
The first property
is the following.

00:00:55.050 --> 00:00:57.970
The process starts at time 0.

00:00:57.970 --> 00:01:03.130
You come in and start
watching at let's say time 7.

00:01:03.130 --> 00:01:05.519
Or more generally,
instead of time 7,

00:01:05.519 --> 00:01:07.500
suppose that you
come in and start

00:01:07.500 --> 00:01:10.250
watching at some time, little t.

00:01:10.250 --> 00:01:13.900
The important thing here is
that little t is a constant.

00:01:13.900 --> 00:01:16.500
It's a deterministic number.

00:01:16.500 --> 00:01:22.590
Starting at that time,
what will you see?

00:01:22.590 --> 00:01:26.210
Well, the original
process was Poisson.

00:01:26.210 --> 00:01:29.140
This means that
disjoint intervals

00:01:29.140 --> 00:01:33.229
in the original process
are independent.

00:01:33.229 --> 00:01:36.450
Therefore, disjoint
intervals in the process

00:01:36.450 --> 00:01:41.360
that you will be seeing
will also be independent.

00:01:41.360 --> 00:01:45.100
Furthermore, during any
little interval of length

00:01:45.100 --> 00:01:47.509
delta in the
process that you see

00:01:47.509 --> 00:01:49.880
will still have
probability lambda times

00:01:49.880 --> 00:01:53.330
delta, approximately,
of seeing and arrival.

00:01:53.330 --> 00:01:56.650
Therefore, what you see also
satisfies the properties

00:01:56.650 --> 00:02:01.220
of a Poisson process, and
is itself a Poisson process.

00:02:01.220 --> 00:02:05.860
Second, the original
process was Poisson.

00:02:05.860 --> 00:02:08.509
So different intervals
are independent.

00:02:08.509 --> 00:02:10.520
So whatever happens
in this interval

00:02:10.520 --> 00:02:13.180
is independent from whatever
happens in that interval.

00:02:13.180 --> 00:02:16.550
But that interval corresponds
to the future of the process,

00:02:16.550 --> 00:02:21.079
and therefore, the future of the
process, what you get to see,

00:02:21.079 --> 00:02:23.990
is independent from
the past history.

00:02:23.990 --> 00:02:27.900
And so the conclusion is that
the process that you get to see

00:02:27.900 --> 00:02:32.180
is a Poisson process, which
is independent of the history

00:02:32.180 --> 00:02:35.150
until the time that
you started watching.

00:02:35.150 --> 00:02:37.970
And we say, therefore,
that what you see

00:02:37.970 --> 00:02:42.520
is a fresh starting process.

00:02:42.520 --> 00:02:46.280
The Poisson process
starts fresh at time t.

00:02:46.280 --> 00:02:48.640
We have the fresh
start property.

00:02:48.640 --> 00:02:51.360
And similar to the language we
use for the Bernoulli process,

00:02:51.360 --> 00:02:55.340
the fresh start property means
that you see a process that's

00:02:55.340 --> 00:02:57.630
independent of
the past and which

00:02:57.630 --> 00:03:01.890
has the same statistical
properties as if this was time

00:03:01.890 --> 00:03:06.930
0, as if the process was
just starting right now.

00:03:06.930 --> 00:03:09.810
One consequence of this
fresh start property

00:03:09.810 --> 00:03:11.230
is the following.

00:03:11.230 --> 00:03:13.600
You start watching at time t.

00:03:13.600 --> 00:03:15.720
And you're interested
in the time

00:03:15.720 --> 00:03:19.820
it takes until the next arrival.

00:03:19.820 --> 00:03:23.640
What are the properties
of this random variable?

00:03:23.640 --> 00:03:26.490
Well, since you have a fresh
starting Poisson process

00:03:26.490 --> 00:03:30.410
at this time, this is the
time until the first arrival

00:03:30.410 --> 00:03:34.270
in this fresh starting
Poisson process.

00:03:34.270 --> 00:03:36.940
And the time until
the first arrival

00:03:36.940 --> 00:03:39.079
in a process that
is just starting,

00:03:39.079 --> 00:03:42.760
we know that it has an
exponential distribution.

00:03:42.760 --> 00:03:46.090
So this is going to be an
exponential random variable

00:03:46.090 --> 00:03:48.579
with the same parameter, lambda.

00:03:48.579 --> 00:03:51.329
Furthermore, because the
process starts fresh,

00:03:51.329 --> 00:03:55.120
whatever happens in the future
is independent from the past.

00:03:55.120 --> 00:03:58.310
And so this random variable,
the remaining time,

00:03:58.310 --> 00:04:03.010
is independent of whatever
happened in the past until time

00:04:03.010 --> 00:04:03.510
t.

00:04:09.810 --> 00:04:13.480
Now let us look at a
somewhat different situation.

00:04:13.480 --> 00:04:19.000
You start watching the
process at time T1.

00:04:19.000 --> 00:04:23.600
Time T1 is the time
of the first arrival.

00:04:23.600 --> 00:04:27.130
And you start
watching from here on.

00:04:27.130 --> 00:04:29.530
What is it that
you're going to see?

00:04:29.530 --> 00:04:31.860
Suppose that the
first arrival happens,

00:04:31.860 --> 00:04:34.760
let's say, at time equal to 3.

00:04:34.760 --> 00:04:37.659
So we're conditioning
on this event.

00:04:37.659 --> 00:04:41.290
In that case, you start
watching the process at time 3.

00:04:41.290 --> 00:04:45.540
And you also know that the first
arrival happened at time 3.

00:04:45.540 --> 00:04:49.220
But this fact about the first
arrival happening at time 3

00:04:49.220 --> 00:04:53.710
belongs to the history of
the process until time 3.

00:04:53.710 --> 00:04:55.850
This is information
about the past,

00:04:55.850 --> 00:05:00.400
and does not affect what is
going to happen after time 3.

00:05:00.400 --> 00:05:04.710
The process after time 3 will
be independent from the history

00:05:04.710 --> 00:05:08.120
until time 3 and whatever
happened until that time.

00:05:08.120 --> 00:05:11.840
So starting at that particular
time 3, what you see

00:05:11.840 --> 00:05:15.860
is a Poisson process that is
independent from the past.

00:05:15.860 --> 00:05:18.830
Now, this argument
is valid even if I

00:05:18.830 --> 00:05:24.260
were to use here a
3.5 or 3.4 or 3.7.

00:05:24.260 --> 00:05:28.130
No matter when this
first arrival occurred,

00:05:28.130 --> 00:05:30.820
what I see starting
from this time

00:05:30.820 --> 00:05:36.860
is a Poisson process which
is independent from the past.

00:05:36.860 --> 00:05:38.850
At the time of
the first arrival,

00:05:38.850 --> 00:05:43.230
the process just starts fresh.

00:05:43.230 --> 00:05:47.070
As a consequence of this, and by
repeating the argument that we

00:05:47.070 --> 00:05:50.260
carried out for the remaining
time until the next arrival

00:05:50.260 --> 00:05:52.900
up here, we can
repeat this argument

00:05:52.900 --> 00:05:56.590
and argue that the time
until the next arrival

00:05:56.590 --> 00:06:00.440
in this fresh
starting process, this

00:06:00.440 --> 00:06:05.490
will also be an exponential
random variable.

00:06:05.490 --> 00:06:07.780
Now, this time until
the next arrival

00:06:07.780 --> 00:06:10.890
is the difference between
the second arrival

00:06:10.890 --> 00:06:13.380
time and the first arrival time.

00:06:13.380 --> 00:06:16.310
And we denote it by T2.

00:06:16.310 --> 00:06:21.280
What we just argued is that
this time until the next arrival

00:06:21.280 --> 00:06:25.230
is going to be an
exponential random variable.

00:06:25.230 --> 00:06:29.620
And also, it is
independent from the past.

00:06:29.620 --> 00:06:35.620
And in particular, it
is independent from T1.

00:06:35.620 --> 00:06:38.270
So the time until
the second arrival,

00:06:38.270 --> 00:06:42.500
starting from the first arrival,
the second inter-arrival time

00:06:42.500 --> 00:06:46.350
is a random variable that has
an exponential distribution that

00:06:46.350 --> 00:06:49.190
is the same distribution
as that of T1,

00:06:49.190 --> 00:06:52.030
and is independent from T1.

00:06:52.030 --> 00:06:54.870
Now we can extend this
argument and look at the kth

00:06:54.870 --> 00:06:56.550
inter-arrival time.

00:06:56.550 --> 00:07:00.410
For example, if the
arrival numbered k minus 1

00:07:00.410 --> 00:07:03.960
occurred here, and the
k arrival occurs here,

00:07:03.960 --> 00:07:08.650
this difference, here
we denote it by Tk,

00:07:08.650 --> 00:07:12.270
and by arguing in a similar
way that the process starts

00:07:12.270 --> 00:07:16.490
fresh at this particular time,
the time until the next arrival

00:07:16.490 --> 00:07:19.400
will also be an
exponential random variable

00:07:19.400 --> 00:07:21.510
with the same distribution.

00:07:21.510 --> 00:07:23.660
And furthermore,
will be independent

00:07:23.660 --> 00:07:26.540
from the past history,
and therefore,

00:07:26.540 --> 00:07:31.000
independent from the
earlier inter-arrival times.

00:07:31.000 --> 00:07:34.730
And this has lots of
important implications.

00:07:34.730 --> 00:07:38.840
For example, the time
until the kth arrival,

00:07:38.840 --> 00:07:43.260
which is the sum of the
first k inter-arrival times,

00:07:43.260 --> 00:07:47.850
is the sum of independent,
identically distributed,

00:07:47.850 --> 00:07:50.920
exponential random variables.

00:07:50.920 --> 00:07:55.170
In particular, this means
that we can find the PDF of Yk

00:07:55.170 --> 00:07:58.370
by convolving the
exponential PDF

00:07:58.370 --> 00:08:03.350
of these inter-arrival times,
convolving this exponential PDF

00:08:03.350 --> 00:08:05.570
with itself k times.

00:08:05.570 --> 00:08:09.490
And this is indeed one
way to find the PDF of Yk.

00:08:09.490 --> 00:08:12.340
But fortunately for us,
we were able to find it

00:08:12.340 --> 00:08:13.970
with a much simpler argument.

00:08:13.970 --> 00:08:16.590
And we already know what it is.

00:08:16.590 --> 00:08:20.490
But this property here is
also useful for finding

00:08:20.490 --> 00:08:22.790
the mean and the variance of Yk.

00:08:22.790 --> 00:08:25.850
The mean of the sum is
the sum of the means.

00:08:25.850 --> 00:08:27.960
And since the random
variables are independent,

00:08:27.960 --> 00:08:30.980
the variance of the sum is
the sum of the variances.

00:08:30.980 --> 00:08:34.280
We know what is the mean and
the variance of an exponential.

00:08:34.280 --> 00:08:37.308
And so by multiplying
that by k, we

00:08:37.308 --> 00:08:40.429
obtain the mean of
the kth arrival time

00:08:40.429 --> 00:08:43.630
and the variance of
the kth arrival time.

00:08:43.630 --> 00:08:45.720
And so we now know
the mean and variance

00:08:45.720 --> 00:08:49.820
of the Erlang PDF of order k.

00:08:49.820 --> 00:08:53.290
A second implication
of this property

00:08:53.290 --> 00:08:56.150
is more theoretical,
more conceptual.

00:08:56.150 --> 00:08:58.960
Recall that we defined
the Poisson process

00:08:58.960 --> 00:09:01.320
in terms of an
independence assumption

00:09:01.320 --> 00:09:04.540
and an assumption on the
probability of arrivals

00:09:04.540 --> 00:09:06.340
during a small interval.

00:09:06.340 --> 00:09:11.780
But we could have defined the
Poisson process as follows.

00:09:11.780 --> 00:09:14.600
Consider a sequence of
independent, identically

00:09:14.600 --> 00:09:16.830
distributed exponentials.

00:09:16.830 --> 00:09:18.650
Call them Tk.

00:09:18.650 --> 00:09:23.370
And use these to define
the arrival times.

00:09:23.370 --> 00:09:27.140
This is a way of
constructing a process.

00:09:27.140 --> 00:09:29.880
What we argued in
this segment is

00:09:29.880 --> 00:09:33.500
that a Poisson process under
the original definition

00:09:33.500 --> 00:09:36.970
satisfies this new definition.

00:09:36.970 --> 00:09:38.830
One can complete
the argument to show

00:09:38.830 --> 00:09:40.770
that the two definitions
are equivalent.

00:09:40.770 --> 00:09:44.260
It is possible to argue
that if we define an arrival

00:09:44.260 --> 00:09:48.570
process in this manner, this
arrival process will also

00:09:48.570 --> 00:09:52.490
satisfy the basic properties
of the Poisson process.

00:09:52.490 --> 00:09:54.720
This argument can
indeed be carried out,

00:09:54.720 --> 00:09:57.810
but we will not go through it.

00:09:57.810 --> 00:10:01.720
A final implication, which
is a little more practical.

00:10:01.720 --> 00:10:04.440
If you want to simulate
the Poisson process,

00:10:04.440 --> 00:10:06.590
how would you do it?

00:10:06.590 --> 00:10:11.130
Given what we now know, the most
natural way is the following.

00:10:11.130 --> 00:10:15.320
We generate independent,
identically distributed,

00:10:15.320 --> 00:10:17.560
exponential random
variables, using

00:10:17.560 --> 00:10:20.070
for example a random
number generator.

00:10:20.070 --> 00:10:23.730
And then use these
exponential random variables

00:10:23.730 --> 00:10:27.530
to construct the values of
the inter arrival times.

00:10:27.530 --> 00:10:30.710
And this way, construct
a complete time history

00:10:30.710 --> 00:10:33.050
of the Poisson process.