Markov Processes

Download Report

Transcript Markov Processes

Markov Processes
and
Birth-Death Processes
J. M. Akinpelu
Exponential Distribution
Definition. A continuous random variable X has an
exponential distribution with parameter  > 0 if its
probability density function is given by
e  x ,
f ( x)  
 0,
x0
x  0.
Its distribution function is given by
1  e  x ,
F ( x)  
 0,
x0
x  0.
Exponential Distribution
Theorem 1. A continuous R.V. X is exponentially
distributed if and only if for s, t  0,
P{X  s  t | X  t}  P{X  s}
(1)
or equivalently,
P{X  s  t}  P{X  s}P{X  t}.
A random variable with this property is said to be
memoryless.
Exponential Distribution
Proof: If X is exponentially distributed, (1) follows readily.
Now assume (1). Define F(x) = P{X ≤ x}, f (x) = F(x),
and, G(x) = P{X > x}. It follows that G(x) = ‒ f (x). Now
fix x. For h  0,
G( x  h)  G( x)G(h) .
This implies that, taking the derivative wrt x,
dG ( x  h )
dx
 G ( x) G (h)
dG ( x  h )
dx
  f ( x)G (h)
dG ( x  h )
G (h)
  f ( x)dx.
Exponential Distribution
Letting x = 0 and integrating both sides from 0 to t
gives
dG ( h )
G (h)
t

  f (0)dx
t
dG ( h )
G (h)
0
   f (0)dx
0
log(G (h))
t
0
  f (0) x 0
log(G (t ))   f (0)t
G (t )  e  f ( 0)t .
t
Exponential Distribution
Theorem 2. A R.V. X is exponentially distributed
if and only if for h  0,
P{X  h}  h  o(h)
P{X  h}  1  h  o(h) .
Exponential Distribution
Proof: Let X be exponentially distributed, then
for h  0,
 h
P{ X  h}  1  e


( h) n
 1  1  
 n 1 n!
n

( h)
 h  
n!
n2
 h  o( h).
The converse is left as an exercise.



Exponential Distribution
1.2
1
F (x )
0.8
0.6
0.4
0.2
slope (rate) ≈ 
0
0
1
2
3
x
4
5
6
Markov Process
A continuous time stochastic process {Xt, t  0}
with state space E is called a Markov process provided
that
P{ X s t  j | X s  i, X u  xu , 0  u  s}
 P( X s t  j | X s  i}
for all states i, j  E and all s, t  0.
known
0
s
s+t
Markov Process
We restrict ourselves to Markov processes for which the
state space E = {0, 1, 2, …}, and such that the
conditional probabilities
Pij (t )  P{X st  j | X s  i}
are independent of s. Such a Markov process is called
time-homogeneous.
Pij(t) is called the transition function of the Markov
process X.
Markov Process - Example
Let X be a Markov process with
 r0 (t ) r1 (t ) r2 (t )


r0 (t ) r1 (t )

P(t )  
r0 (t )



0







where
e t (t ) j i
Pij (t )  r j i (t ) 
, 0i j
( j  i)!
for some  > 0. X is a Poisson process.
Chapman-Kolmogorov Equations
Theorem 3. For i, j  E, t, s  0,
P ij (s  t )   Pik (s) Pkj (t ).
kE
Realization of a Markov Process
Xt()
7
S4
6
S2
5
4
S3
3
S1
S5
2
S0
1
0
T0
T1
T2
T3
T4
T5
t
Time Spent in a State
Theorem 4. Let t  0, and n satisfy Tn ≤ t < Tn+1, and let Wt =
Tn+1 – t. Let i  E, u  0, and define
G(u)  P{Wt  u | X t  i} .
Then
G(u  v)  G(u)G(v).
Note: This implies that the distribution of time remaining in a
state is exponentially distributed, regardless of the time
already spent in that state.
Wt
Tn
t
t+u
Tn+1
Time Spent in a State
Proof: We first note that due to the time homogeneity of X, G(u)
is independent of t. If we fix i, then we have
G (u  v)  P{Wt  u  v | X t  i}
 P{Wt  u, W t u  v | X t  i}
 P{Wt  u | X t  i}P{W t u  v | Wt  u, X t  i}
 P{Wt  u | X t  i}P{Wt u  v | X t u  i}
 G (u ) G (v).
An Alternative Characterization of a
Markov Process
Theorem 5. Let X ={Xt, t  0} be a Markov process. Let T0, T1,
…, be the successive state transition times and let S0, S1, …, be
the successive states visited by X. There exists some number i
such that for any non-negative integer n, for any j  E, and t > 0,
P{S n1  j, Tn1  Tn  t | S 0 ,, S n1 , S n  i ; T0 ,, Tn }
 Q(i, j ) e  it
where
Qij  0, Qii  0,
Q
jE
ij
 1.
An Alternative Characterization of a
Markov Process
This implies that the successive states visited by
a Markov process form a Markov chain with
transition matrix Q.
A Markov process is irreducible recurrent if its
underlying Markov chain is irreducible recurrent.
Kolmogorov Equations
Theorem 6.
Pij (t )   i Qik Pkj (t )  iPij (t )
k i
and, under suitable regularity conditions,
Pij (t )   k Qkj Pik (t )  jP ij (t ) .
k j
These are Kolmogorov’s Backward and Forward
Equations.
Kolmogorov Equations
Proof (Forward Equation): For t, h  0,
Pij (t  h)   Pik (t )  k h Qkj  P ij (t ) 1  v j h  o(h) .
k j
Hence
Pij (t  h)  Pij (t )
h
o(h)
  k Qkj Pik (t )  jP ij (t ) 
.
h
k j
Taking the limit as h  0, we get our result.
Limiting Probabilities
Theorem 7. If a Markov process is irreducible recurrent, then
limiting probabilities
P j  lim Pij (t )
t 
exist independent of i, and satisfy
 j P j   k Qkj P k
k j
for all j. These are referred to as “balance equations”. Together
with the condition
 Pj  1 ,
j
they uniquely determine the limiting distribution.
Birth-Death Processes
Definition. A birth-death process {X(t), t  0} is a Markov
process such that, if the process is in state j, then the only
transitions allowed are to state j + 1 or to state j – 1 (if j > 0).
It follows that there exist non-negative values j and j,
j = 0, 1, 2, …, (called the birth rates and death rates) so that,
P{ X t  h  j | X t  j  1}   j 1h  o(h)
P{ X t  h  j | X t  j  1}   j 1h  o(h)
P{ X t  h  j | X t  j}  1   j h   j h  o(h)
P{ X t  h  j | X t  i}  o(h)
if | j  i | 1.
Birth and Death Rates
j-1
j
j
j-1
j
j+1
j+1
Note:
1. The expected time in state j before entering state j+1 is 1/j;
the expected time in state j before entering state j‒1 is 1/j.
2. The rate corresponding to state j is vj = j + j.
Differential-Difference Equations for
a Birth-Death Process
It follows that, if Pj (t )  P{X (t )  j} , then
d
Pj (t )  j 1P j 1 (t )   j 1 Pj 1 (t )  ( j   j ) Pj (t ),
dt
j0
d
P0 (t )  1 P1 (t )  0 P0 (t ).
dt
Together with the state distribution at time 0, this
completely describes the behavior of the birthdeath process.
Birth-Death Processes - Example
Pure birth process with constant birth rate
j =  > 0, j = 0 for all j. Assume that

1
Pj (0)  

0
if j  0
if j  0.
Then solving the difference-differential equations for this
process gives
( t ) e
Pj (t ) 
j!
j
 t
.
Birth-Death Processes - Example
Pure death process with proportional death rate
j = 0 for all j, j = j > 0 for 1 ≤ j ≤ N, j = 0 otherwise,
and

1
Pj (0)  

0
if j  N
otherwise.
Then solving the difference-differential equations for this
process gives
 N  t j
Pj (t )   (e ) (1  et ) N  j
 j
0  j  N.
Limiting Probabilities
Now assume that limiting probabilities Pj exist.
They must satisfy:
or
0  j 1P j 1   j 1 Pj 1  ( j   j ) Pj ,
0  1 P1  0 P0
( j   j ) Pj  j 1P j 1  j 1Pj 1 ,
0 P0  1P1.
j 0
j0
(*)
Limiting Probabilities
These are the balance equations for a birth-death
process. Together with the condition

P
j 0
j
1,
they uniquely define the limiting probabilities.
Limiting Probabilities
From (*), one can prove by induction that
i
Pj  P0 
i 0  i 1
j 1
j  0, 1, 2,  .
When Do Limiting Probabilities
Exist?
Define
i
S  1  
.
j 1 i 0 i 1

j 1
It is easy to show that
Po  S
1
if S < . (This is equivalent to the condition P0 > 0.)
Furthermore, all of the states are recurrent positive, i.e.,
ergodic. If S = , then either all of the states are
recurrent null or all of the states are transient, and
limiting probabilities do not exist.
Flow Balance Method
Draw a closed boundary around state j:
j-1
j
j-1
j
Global balance
equation:
j
j+1
j+1
 j 1Pj 1   j 1Pj 1  ( j   j )Pj
“flow in = flow out”
Flow Balance Method
Draw a closed boundary between state j and state j–1:
j-1
j
j
j-1
j
Detailed balance
equation:
j+1
j+1
 j 1Pj 1   j Pj
Example
Machine repair problem. Suppose there are m machines
serviced by one repairman. Each machine runs without
failure, independent of all others, an exponential time
with mean 1/. When it fails, it waits until the
repairman can come to repair it, and the repair itself
takes an exponentially distributed amount of time with
mean 1/. Once repaired, the machine is as good as
new.
What is the probability that j machines are failed?
Example
Let Pj be the steady-state probability of j failed
machines.
 j-1=(m‒j+1)
j=(m‒j)
j
j‒1
j= 
j+1
j+1= 
 (m  j  1) Pj 1  Pj 1  [ (m  j )   ]Pj
m P0  P1
Pm1  Pm
Example
j-1=(m‒j+1)
j=(m‒j)
j
j‒1
j+1
j=
i
Pj  P0 
i  0  i 1
j 1
 (m  i )
 P0 

i 0
j+1=
j 1
P0 
 P0 m(m  1)  (m  j  1)( /  ) j
1

1   m(m  1) (m  j  1) 
j 1

m
j
Example
How would this example change if there were m
(or more) repairmen?
Homework
No homework this week due to test next week.
References
1. Erhan Cinlar, Introduction to Stochastic Processes,
Prentice-Hall, Inc., 1975.
2. Leonard Kleinrock, Queueing Systems, Volume I:
Theory, John Wiley & Sons, 1975.
3. Sheldon M. Ross, Introduction to Probability
Models, Ninth Edition, Elsevier Inc., 2007.