Stochastic Processes

- Poisson and renewal processes [Link] Next Generation Communication Networks Lab. Division of Applied Mathematics KAIST

Stochastic processes

Reference S.M. Ross, Stochastic Process, 2nd ed., John Wiley & Sons, Inc., 1996, chapter 2.
A stochastic process X(t) is a family of random variables indexed by a time parameter t

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

An example of a stochastic process X(t)

X(t)

a sample path

time

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

Independent Increments : A stochastic process X(t) is said to have independent increments if X(t2) - X(t1) and X(t4)-X(t3) are independent for any t1 < t2 < t3 < t4. (c.f. disjoint intervals)

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

Stationary Increments : A stochastic process X(t) is said to have stationary increments if X(t2+s)-X(t1+s) and X(t2)-X(t1) have the same distribution for all t1 < t2, s > 0.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

The Poisson process

A stochastic process X(t) is called a Poisson process having rate (>0) if it satisfies the following: X(0) = 0 X(t) has independent and stationary increments P{X(t) = n} = e-t(t)n/n!
little o function (o(h)): A function f(x) is called o(h) if limh! o f(h)/h = 0. Then, we see P{X(h) = 1} = h + o(h) P{X(h) 2} = o(h)
Division of Applied Mathematics, KAIST

Next Generation Communication Networks Lab.

An equivalent definition

A stochastic process X(t) is called a Poisson process having rate (>0) if it satisfies the following: X(0) = 0 X(t) has independent and stationary increments P{X(h) = 1} = h + o(h) P{X(h) 2} = o(h)

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

Inter-arrival times tn for a Poisson process

From the definition, we have P{t1>t} = P{X(t) = 0} = e-t.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

and consequently,

Similarly, for n 3, we can show that Xn are exponentially distributed with parameter . In addition, the independence of Xn is trivial from the fact that the Poisson process has independent increments.
Division of Applied Mathematics, KAIST

Next Generation Communication Networks Lab.

Therefore, we have the following: For a Poisson process, inter-arrival times {tn, n 1} are independent and identically distributed (i.i.d.) exponential random variables with rate .

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

Examples of a Poisson process

Voice call arrivals at a base station

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

10

KAIST ARA BBS : Exponential inter-access times Access times during [12:00 18:00]

Empirical data

Empirical data

The Q-Q plot

The Kolmogorov-Smirnov Test

11

Internet traffic in the backbone Traffic appears Poisson at sub-second time scale T. Karagiannis et. al, A nonstationary Poisson view of internet traffic, INFOCOM 2004, 1558-1569.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

12

The complementary distribution function of the Packet interarrival times

exponential distribution

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

13

Properties of the Poisson process

Superposition of two independent Poisson processes with rate 1 and 2, respectively, is a Poisson process with rate 1+2.
Proof: Let X1(t) and X2(t) be the two Poisson processes with rate 1 and 2, respectively. Note that the PGF of a Poisson distribution with parameter t is given by e(z-1)t.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

14

Random selection: From a Poisson process with parameter if we select each arrival with probability p, then the selected arrivals form a new Poisson process with parameter p.
Proof: Let Y(t) be the selected process. Then,

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

15

A 22 switch with Poisson arrivals

Poisson(2)

1-q q

Poisson(1(1-p)+2q)
Next Generation Communication Networks Lab.
Division of Applied Mathematics, KAIST

16

Random split: A Poisson process can be split into two independent sub-processes by random selection with probability p. The two processes are Poisson with p and Poisson with (1-p).
Proof: Let X1(t) and X2(t) be two sub-processes.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

17

Conditioning on the number of arrivals: Given that we have X(t) = n arrivals in [0,t], the n arrival times 1, , n are the order statistics corresponding to n independent random variables uniformly distributed in the interval [0,t]

By letting hi ! 0 we obtain the conditional joint density function f(t1,t2,,tn) = n!/tn.

18

c.f. Order statistics generated by uniform random variables

Let X1, , Xn be independent uniform random variables in [0,t]. Consider the following ordered random variables X(1) = max{X1, ,Xn}, , X(n) = min{X1, , Xn}. Then for 0<t1 < < tn<t and sufficiently small > 0

Hence, given that X(t) = n, the arrival epochs i, 1 i n, of the Poisson process are the order statistics generated by n independent random variables uniformly distributed in the interval [0,t]
Division of Applied Mathematics, KAIST

Next Generation Communication Networks Lab.

19

PASTA (Poisson arrivals see time average) Consider a queueing system with Poisson arrivals with rate . i = the probability that the system is in state i in the steady state i(a) = the probability that the system is in state i at an arbitrary arrival instant The PASTA property tells that i = i(a).

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

20

a counterexample of the PASTA property packets arrive at every seconds the service times of packets are all 0.9 seconds the number of packets just before an arrival = 0 the time average number of packets = 0.9
packet departures

packet arrivals
Next Generation Communication Networks Lab.
Division of Applied Mathematics, KAIST

21

Renewal Process

Reference: S.M. Ross, Stochastic Process, 2nd ed., John Wiley & Sons, Inc., 1996, chapter 3.
Definition of a renewal process Let {Tn, n 1} be a sequence of nonnegative independent random variables with a common distribution F(t) with F(0)<1. Let S0 = 0 and Sn = i=1n Tn, n 1. Then the process N(t) = sup{n 0 | Sn t}, the number of renewals (or arrivals) by time t, is called a renewal process.

time Renewals (arrivals, events)

22

Properties of a renewal process

Since F(0) < 1, 0 < E[T1] 1 By strong law of large numbers, Sn = i=1n Tn goes to infinity a.s. as n ! 1 because E[T1] > 0. Therefore, N(t) < 1 a.s. for every finite t > 0. In addition,

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

23

Distribution of N(t)

Note that
Then,

Let m(t) = E[N(t)] and Fn(t) is the dist. ft of Sn.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

24

The Elementary Renewal Theorem

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

25

The Blackwells Theorem

Lattice distribution A nonnegative R.V. X is said to be lattice if there exists d 0 (called span) such that n=01 P{X=nd} = 1.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

26

Forward and Backward recurrence times

backward recurrence time (or the age of a renewal) Tb(t) = t - SN(t) forward recurrence time (or the remaining life time) Tf(t) = SN(t)+1 - t : the interval length including the present time TN(t)+1 = Tb(t) + Tf(t)

present time t
Next Generation Communication Networks Lab.
Division of Applied Mathematics, KAIST

27

Equilibrium distribution

Assume that F(t) is a non-lattice distribution with mean E[T]. Then,

Fe(t) = (1/E[T]) s0x 1-F(y) dy is called the equilibrium distribution associated with F(t).
28

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

Similarly, for a discrete R.V. T with p.m.f. (probability mass function) p(k), k 1 we have its equilibrium distribution

Applications: For a batch arrival process, consider a tagged customer. P{limn!1 Tb(n) = k} is p.m.f. of the number of customers in front of our tagged customer in the batch containing our tagged customer P{limn!1 Tf(n) = k} is p.m.f. of the number of customers behind our tagged customer in the batch containing our tagged customer P{limn!1 TN(n)+1 = k} is p.m.f. of the size of the batch containing our tagged customer
Division of Applied Mathematics, KAIST

Next Generation Communication Networks Lab.

29

Why we need Fe(x) ?

Let X(i) be the i-th batch size in a batch arrival process. We assume {X(i)} are i.i.d. random variables with P{X=j} = aj. compute P{an arbitrary customer is in the j-th position in a batch}.

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

30

Alternating renewal process

an i.i.d. sequence {Zn, n 1} for ON periods an i.i.d. sequence {Yn, n 1} for OFF periods For each n, Zn and Yn may be dependent {(Zn, Yn, n 1)} is called an alternating renewal process an alternating renewal process is a good model for an ON and OFF source.

0 (off)

1 (on)

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

31

Limiting Theorem for alternating renewal process Assume that the system is on at time 0. Let p(t) = P{The system is on at time t}. Assume that F(x) is distribution of Zn+Yn, which is non-lattice with finite mean. Then limt!1 p(t) = E[Z1]/E[Z1+Y1].

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST

32

Exponential ON and OFF source

In the steady state

For N homogeneous ON and OFF sources,

0 (off)

1 (on)
33

Next Generation Communication Networks Lab.

Division of Applied Mathematics, KAIST