4. markov chains (9/23/12, cf. ross) 1. introduction 2...
TRANSCRIPT
4. Markov Chains
4. Markov Chains (9/23/12, cf. Ross)
1. Introduction
2. Chapman-Kolmogorov Equations
3. Types of States
4. Limiting Probabilities
5. Gambler’s Ruin
6. First Passage Times
7. Branching Processes
8. Time-Reversibility
1
4. Markov Chains
4.1. Introduction
Definition: A stochastic process (SP) {X(t) : t ∈ T}
is a collection of RV’s. Each X(t) is a RV; t is usually
regarded as “time.”
Example: X(t) = the number of customers in line at
the post office at time t.
Example: X(t) = the price of IBM stock at time t.
2
4. Markov Chains
T is the index set of the process. If T is countable,
then {X(t) : t ∈ T} is a discrete-time SP. If T is some
continuum, then {X(t) : t ∈ T} is a continuous-time
SP.
Example: {Xn : n = 0,1,2, . . .} (index set of non-
negative integers)
Example: {X(t) : t ≥ 0} (index set is <+)
3
4. Markov Chains
The state space of the SP is the set of all possible
values that the RV’s X(t) can take.
Example: If Xn = j, then the process is in state j at
time n.
Any realization of {X(t)} is a sample path.
4
4. Markov Chains
Definition: A Markov chain (MC) is a SP such that
whenever the process is in state i, there is a fixed
transition probability Pij that its next state will be j.
Denote the “current” state (at time n) by Xn = i.
Let the event A = {X0 = i0, X1 = i1, . . . Xn−1 = in−1}
be the previous history of the MC (before time n).
5
4. Markov Chains
{Xn} has the Markov property if it forgets about its
past, i.e.,
Pr(Xn+1 = j|A ∩Xn = i) = Pr(Xn+1 = j|Xn = i).
{Xn} is time homogeneous if
Pr(Xn+1 = j|Xn = i) = Pr(X1 = j|X0 = i) = Pij,
i.e., if the transition probabilities are independent of
n.
6
4. Markov Chains
Recap: A Markov chain is a SP such that
Pr(Xn+1 = j|A ∩Xn = i) = Pij,
i.e., the next state depends only on the current state
(and is indep of the time).
7
4. Markov Chains
Since Pij is a probability, 0 ≤ Pij ≤ 1 for all i, j.
Since the process has to go from i to some state, we
must have∑∞j=0Pij = 1, for all i. Note that it may be
possible to go from i to i (i.e., “stay” at i).
Definition: The one-step transition matrix is
P =
P00 P01 P02 · · ·P10 P11 P12 · · ·
... ... ...
8
4. Markov Chains
Example: A frog lives in a pond with three lily pads
(1,2,3). He sits on one of the pads and periodically
rolls a die. If he rolls a 1, he jumps to the lower
numbered of the two unoccupied pads. Otherwise,
he jumps to the higher numbered pad. Let X0 be the
initial pad and let Xn be his location just after the nth
jump. This is a MC since his position only depends
on the current position, and the Pij’s are independent
of n.
P =
0 1/6 5/6
1/6 0 5/61/6 5/6 0
. ♦
9
4. Markov Chains
Example: Let Xi denote the weather (rain or sun) on
day i. We’ll think of Xi−1 as yesterday, Xi as today,
and Xi+1 as tomorrow. Suppose that
Pr(Xi+1 = R | Xi−1 = R,Xi = R) = 0.7
Pr(Xi+1 = R | Xi−1 = S,Xi = R) = 0.5
Pr(Xi+1 = R | Xi−1 = R,Xi = S) = 0.4
Pr(Xi+1 = R | Xi−1 = S,Xi = S) = 0.2
10
4. Markov Chains
X0, X1, . . . isn’t quite a MC, since the probability that
it’ll rain tomorrow depends on Xi and Xi−1.
We’ll transform the process into a MC by defining the
following states in terms of today and yesterday.
0 : Xi−1 = R, Xi = R
1 : Xi−1 = S, Xi = R
2 : Xi−1 = R, Xi = S
3 : Xi−1 = S, Xi = S
11
4. Markov Chains
Thus, we have, e.g.,
Pr(Xi+1 = R | Xi−1 = R,Xi = R) = P00 = 0.7
Pr(Xi+1 = S | Xi−1 = R,Xi = R) = P02 = 0.3
Using similar reasoning, we get
P =
0.7 0 0.3 00.5 0 0.5 00 0.4 0 0.60 0.2 0 0.8
. ♦
12
4. Markov Chains
Example: A MC whose state space is given by the
integers is called a random walk if Pi,i+1 = p and
Pi,i−1 = 1− p.
P =
... ... ... ... ...· · · 1− p 0 p 0 0 · · ·· · · 0 1− p 0 p 0 · · ·· · · 0 0 1− p 0 p · · ·· · · 0 0 0 1− p 0 · · ·
... ... ... ... ...
. ♦
13
4. Markov Chains
Example (Gambler’s Ruin): Every time a gambler
plays a game, he wins $1 w.p. p, and he loses $1
w.p. 1− p. He stops playing as soon as his fortune is
either $0 or $N. The gambler’s fortune is a MC with
the following Pij’s:
Pi,i+1 = p, i = 1,2, . . . , N − 1
Pi,i−1 = 1− p, i = 1,2, . . . , N − 1
P0,0 = PN,N = 1
0 and N are absorbing states — once the process
enters one of these states, it can’t leave. ♦14
4. Markov Chains
Example (Ehrenfest Model): A random walk on a fi-
nite set of states with “reflecting” boundaries. Set of
states is {1,2, . . . , a}.
Pij =
a−ia if j = i+ 1
ia if j = i− 1
0 otherwise
Idea: Suppose A has i marbles, B has a− i. Select a
marble at random, and put it in the other container.
♦15
4. Markov Chains
4.2 Chapman-Kolmogorov Equations
Definition: The n-step transition probability that a
process currently in state i will be in state j after n
additional transitions is
P(n)ij ≡ Pr(Xn = j|X0 = i), n, i, j ≥ 0.
Note that P (1)ij = Pij, and
P(0)ij =
1 if i = j0 otherwise
.
16
4. Markov Chains
Theorem (C-K Equations):
P(n+m)ij =
∞∑k=0
P(n)ik P
(m)kj .
Think of going from i to j in n + m steps with an
intermediate stop in state k after n steps; then sum
over all possible k values.
17
4. Markov Chains
Proof: By definition,
P(n+m)ij
= Pr(Xn+m = j|X0 = i)
=∞∑k=0
Pr(Xn+m = j ∩Xn = k|X0 = i) (total prob)
=∞∑k=0
Pr(Xn+m = j|X0 = i ∩Xn = k)Pr(Xn = k|X0 = i)
(since Pr(A ∩ C|B) = Pr(A|B ∩ C)Pr(C|B))
=∞∑k=0
Pr(Xn+m = j|Xn = k)Pr(Xn = k|X0 = i)
(Markov property). ♦
18
4. Markov Chains
Definition: The n-step transition matrix is
P(n) =
P(n)00 P
(n)01 P
(n)02 · · ·
P(n)10 P
(n)11 P
(n)12 · · ·
... ... ...
The C-K equations imply P(n+m) = P(n)P(m).
In particular, P(2) = P(1)P(1) = PP = P2.
By induction, P(n) = Pn.19
4. Markov Chains
Example: Let Xi = 0 if it rains on day i; otherwise,
Xi = 1. Suppose P00 = 0.7 and P10 = 0.4. Then
P =
0.7 0.30.4 0.6
.Suppose it rains on Monday. Then the prob that it
rains on Friday is P (4)00 . Note that
P(4) = P4 =
0.7 0.30.4 0.6
4
=
0.5749 0.42510.5668 0.4332
,so that P (4)
00 = 0.5749. ♦
20
4. Markov Chains
Unconditional Probabilities
Suppose we know the “initial” probabilities,
αi ≡ Pr(X0 = i), i = 0,1, . . . .
(Note that∑iαi = 1.) Then by total probability,
Pr(Xn = j) =∞∑i=0
Pr(Xn = j ∩X0 = i)
=∞∑i=0
Pr(Xn = j|X0 = i)Pr(X0 = i)
=∞∑i=0
P(n)ij αi.
21
4. Markov Chains
Example: In the above example, suppose α0 = 0.4
and α1 = 0.6. Find the prob that it will not rain on
the 4th day after we start keeping records (assuming
nothing about the first day).
Pr(X4 = 1) =∞∑i=0
P(4)i1 αi
= P(4)01 α0 + P
(4)11 α1
= (0.4251)(0.4) + (0.4332)(0.6)
= 0.4300. ♦
22
4. Markov Chains
4.3 Types of States
Definition: If P(n)ij > 0 for some n ≥ 0, state j is
accessible from i.
Notation: i→ j.
Definition: If i→ j and j → i, then i and j communi-
cate.
Notation: i↔ j.23
4. Markov Chains
Theorem: Communication is an equivalence relation:
(i) i↔ i for all i (reflexive).
(ii) i↔ j implies j ↔ i (symmetric).
(iii) i↔ j and j ↔ k imply i↔ k (transitive).
Proof: (i) and (ii) are trivial, so we’ll only do (iii). To
do so, suppose i ↔ j and j ↔ k. Then there are n,m
such that P (n)ij > 0 and P
(m)jk > 0. So by C-K,
P(n+m)ik =
∞∑r=0
P(n)ir P
(m)rk ≥ P
(n)ij P
(m)jk > 0.
Thus, i→ k. Similarly, k → i. ♦24
4. Markov Chains
Definition: An equivalence class consists of all states
that communicate with each other.
Remark: Easy to see that two equiv classes are dis-
joint.
Example: The following P has equiv classes {0,1} and
{2,3}.
P =
1/2 1/2 0 01/2 1/2 0 0
0 0 3/4 1/40 0 1/4 3/4
. ♦
25
4. Markov Chains
Example: P again has equiv classes {0,1} and {2,3}— note that 1 isn’t accessible from 2.
P =
1/2 1/2 0 01/2 1/4 1/4 0
0 0 3/4 1/40 0 1/4 3/4
. ♦
Definition: A MC is irreducible if there is only one
equiv class (i.e., if all states communicate).
Example: The previous two examples are not irre-
ducible. ♦26
4. Markov Chains
Example: The following P is irreducible since all states
communicate (“loop” technique: 0→ 1→ 0).
P =
1/2 1/21/4 3/4
. ♦
Example: P is irreducible since 0→ 2→ 1→ 0.
P =
1/4 0 3/4
1 0 00 1/2 1/2
. ♦
27
4. Markov Chains
Definition: The probability that the MC eventually
returns to state i is
fi ≡ Pr(Xn = i for some n ≥ 1|X0 = i).
Example: The following MC has equiv classes {0,1},
{2}, and {3}, the latter of which is absorbing.
P =
1/2 1/2 0 01/2 1/2 0 01/4 1/4 1/4 1/4
0 0 0 1
.
We have f0 = f1 = 1, f2 = 1/4, f3 = 1. ♦28
4. Markov Chains
Remark: The fi’s are usually hard to compute.
Definition: If fi = 1, state i is recurrent. If fi < 1,
state i is transient.
Theorem: Suppose X0 = i. Let N denote the number
of times that the MC is in state i (before leaving i
forever). Note that N ≥ 1 since X0 = i. Then i is
recurrent iff E[N ] = ∞ (and i is transient iff E[N ] <
∞).
29
4. Markov Chains
Proof: If i is recurrent, it’s easy to see that the MC
returns to i an infinite number of times; so E[N ] =∞.
Otherwise, suppose i is transient. Then
Pr(N = 1) = 1− fi (never returns)
Pr(N = 2) = fi(1− fi) (returns exactly once)
...
Pr(N = k) = fk−1i (1− fi) (returns k − 1 times)
So N ∼ Geom(1 − fi). Finally, since fi < 1, we have
E[N ] = 11−fi <∞. ♦
30
4. Markov Chains
Theorem: i is recurrent iff∑∞n=1P
(n)ii = ∞. (So i is
transient iff∑∞n=1P
(n)ii <∞.)
Proof: Define the event
An ≡ 1 if Xn = i
0 if Xn 6= i.
Note that N ≡ ∑∞n=1An is the number of returns to i.
Then by the trick that allows us to treat the ex-
pected value of an indicator function as a probability,
we have. . .31
4. Markov Chains
∞∑n=1
P(n)ii =
∞∑n=1
Pr(Xn = i|X0 = i)
=∞∑n=1
E[An|X0 = i] (trick)
= E[ ∞∑n=1
An
∣∣∣∣∣X0 = i
]= E[N |X0 = i] (N = number of returns)
= ∞
⇔ i is recur (by previous theorem). ♦
32
4. Markov Chains
Corollary 1: If i is recur and i↔ j, then j is recur.
Proof: See Ross. ♦
Corollary 2: In a MC with a finite number of states,
not all of the states can be transient.
Proof: Suppose not. Then the MC will run out of
states not to go to an infinite number of times. This
is a contradiction. ♦33
4. Markov Chains
Corollary 3: If one state in an equiv class is transient,
then all states are trans.
Proof: Suppose not, i.e., suppose there’s a recur
state. Since all states in the equiv class communi-
cate, Corollary 1 implies all states are recur. This is
a contradiction. ♦
Corollary 4: All states in a finite irreducible MC are
recurrent.
34
4. Markov Chains
Proof: Suppose not, i.e., suppose there’s a trans
state. Then Corollary 3 implies all states are trans.
But this contradicts Corollary 1. ♦
Definition: By Corollary 1, all states in an equiv class
are recur if one state in that class is recur. Such a
class is a recurrent equiv class.
By Corollary 3, all states in an equiv class are trans
if one state in that class is trans. Such a class is a
transient equiv class.35
4. Markov Chains
Example: Consider the prob transition matrix
P =
1/2 1/21/4 3/4
.Clearly, all states communicate. So this is a finite,
irreducible MC. So Corollary 4 implies all states are
recurrent. ♦
36
4. Markov Chains
Example: Consider
P =
1/4 0 0 3/4
1 0 0 00 1 0 00 0 1 0
.
Loop: 0 → 3 → 2 → 1 → 0. Thus, all states commu-
nicate; so they’re all recurrent. ♦
37
4. Markov Chains
Example: Consider
P =
1/4 0 3/4 0 00 1/2 0 1/2 0
1/2 0 1/2 0 00 1/2 0 1/2 0
1/5 1/5 0 0 3/5
.
The equiv classes are {0,2} (recur), {1,3} (recur),
and {4} (trans). ♦
38
4. Markov Chains
Example: Random Walk: A drunk walks on the inte-
gers 0,±1,±2, . . . with transition probabilities
Pi,i+1 = p
Pi,i−1 = q = 1− p
(i.e., he steps to the right w.p. p and to the left w.p.
1− p).
39
4. Markov Chains
The prob transition matrix is
P =
...q 0 p 0 00 q 0 p 0
. . . 0 0 q 0 p . . .0 0 0 q 0
...
.
Are the states recurrent or transient?
Clearly, all states communicate. So Corollary 1 implies
that if one of the states are recur, then they all are.
Otherwise, all states will be transient.40
4. Markov Chains
Consider a typical state 0. If 0 is recurrent [transient],
then all states will be recurrent [transient]. We’ll find
out which is the case by calculating∑∞n=1P
(n)00 .
Suppose the drunk starts at 0. Since it’s impossible
for him to return to 0 in an odd number of steps, we
see that P (2n+1)00 = 0 for all n ≥ 0.
41
4. Markov Chains
So the only chance he has of returning to 0 is if he’s
taken an even number of steps, say 2n. Of these
steps, n must be taken to the left, and n to the right.
So, thinking binomial, we have
P(2n)00 =
2nn
pnqn =(2n)!
n!n!pnqn, n ≥ 1.
Aside: For large n, Stirling’s approximation says that
n! ≈√
2π nn+12 e−n.
42
4. Markov Chains
After the smoke clears,
P(2n)00 ≈
[4p(1− p)]n√πn
,
so that
∞∑n=1
P(n)00 =
∞∑n=1
P(2n)00
=∞∑n=1
[4p(1− p)]n√πn
=∞ if p = 1/2<∞ if p 6= 1/2
.
So the MC is recur if p = 1/2 and trans otherwise.
♦
43
4. Markov Chains
Definition: If p = 1/2, the random walk is symmetric.
Remark: A 2-dimensional r.w. with probability 1/4 of
going each way yields a recurrent MC.
A 3-dimensional r.w. with probability 1/6 of going
each way (N, S, E, W, up, down) yields a transient
MC.
44
4. Markov Chains
4.4 Limiting Probabilities
Example: Note that the following matrices appear to
be converging. . . .
P =
0.7 0.30.4 0.6
, P(2) =
0.61 0.390.52 0.48
,
P(4) =
0.575 0.4250.567 0.433
, P(8) =
0.572 0.4280.570 0.430
, . . .
45
4. Markov Chains
Definition: Suppose that P (n)ii = 0 whenever n is not
divisible by d, and suppose that d is the largest integer
with this property. Then state i has period d. Think
of d as the greatest common divisor of all n values for
which P(n)ii > 0.
Example: All states have period 3.
P =
0 1 00 0 11 0 0
. ♦
46
4. Markov Chains
Definition: A state with period 1 is aperiodic.
Example:
P =
0 1 0 01 0 0 0
1/4 1/4 1/4 1/40 0 1/2 1/2
.
Here, states 0 and 1 have period 2, while states 2 and
3 are aperiodic. ♦
47
4. Markov Chains
Definition: Suppose state i is recurrent and X0 = i.
If the expected time until the process returns to i is
finite, then i is positive recurrent.
Remark: It turns out that. . .
(1) In a finite MC, all recur states are positive recur.
(2) In an∞-state MC, there may be some recur states
that are not positive recur. Such states are null recur.
Definition: Pos recur, aperiodic states are ergodic.
48
4. Markov Chains
Theorem: For an irreducible, ergodic MC,
(1) πj ≡ limn→∞P(n)ij exists and is independent of i.
(The πj’s are called limiting probabilities.)
(2) πj is the unique, nonnegative solution ofπj =
∑∞i=0 πiPij, j ≥ 0
1 =∑∞j=0 πj
.
In vector notation, this can be written as π = πP.
“Heuristic “proof”: see Ross. ♦49
4. Markov Chains
Remarks: (1) πj is also the long-run proportion of
time that the MC will be in state j. The πj’s are often
called stationary probs — since if Pr(X0 = j) = πj,
then Pr(Xn = j) = πj for all n.
(2) In the irred, pos recur, periodic case, πj can only
be interpreted as the long-run proportion of time in j.
(3) Let mjj ≡ expected number of transitions needed
to go from j to j. Since, on average, the MC spends
1 time unit in state j for every mjj time units, we have
mjj = 1/πj.
50
4. Markov Chains
Example: Find the limiting probabilities of
P =
0.5 0.4 0.10.3 0.4 0.30.2 0.3 0.5
.Solve πj =
∑∞i=0 πiPij (π = πP), i.e.,
π0 = π0P00 + π1P10 + π2P20 = 0.5π0 + 0.3π1 + 0.2π2,
π1 = π0P01 + π1P11 + π2P21 = 0.4π0 + 0.4π1 + 0.3π2,
π2 = π0P02 + π1P12 + π2P22 = 0.1π0 + 0.3π1 + 0.5π2,
and π0 + π1 + π2 = 1. Get π ={
2162,
2362,
1862
}. ♦
51
4. Markov Chains
Definition: A transition matrix P is doubly stochastic
if each column (and row) sums to 1.
Theorem: If, in addition to the conditions of the pre-
vious theorem, P is a doubly stochastic n× n matrix,
then πj = 1/n for all j.
Proof: Just plug in πj = 1/n for all j into π = πP
to verify that it works. Since this solution must be
unique, we’re done. ♦52
4. Markov Chains
Example: Find the limiting probabilities of
P =
0.5 0.4 0.10.3 0.3 0.40.2 0.3 0.5
.
This is a doubly stochastic matrix, so we immediately
have that π0 = π1 = π2 = 1/3. ♦
53
4. Markov Chains
4.5 Gambler’s Ruin Problem
Each time a gambler plays, he wins $1 w.p. p and loses
$1 w.p. 1− p = q. Each play is independent. Suppose
he starts with $i. Find the probability that his fortune
will hit $N (i.e., he breaks the bank) before it hits $0
(i.e., he is ruined).
54
4. Markov Chains
Let Xn denote his fortune at time n. Clearly, {Xn} is
a MC.
Note Pi,i+1 = p and Pi,i−1 = q for i = 1,2, . . . N − 1.
Further, P00 = 1 = PNN .
We have 3 equiv classes: {0} (recur), {1,2, . . . , N −1}
(trans), and {N} (recur).
55
4. Markov Chains
By a standard one-step conditioning argument,
Pi ≡ Pr(Eventually hit $N |X0 = i)
= Pr(Event. hit N |X1 = i+ 1 and X0 = i)
×Pr(X1 = i+ 1|X0 = i)
+ Pr(Event. hit N |X1 = i− 1 and X0 = i)
×Pr(X1 = i− 1|X0 = i)
= Pr(Event. hit N |X1 = i+ 1)p
+ Pr(Event. hit N |X1 = i− 1)q
= pPi+1 + qPi−1, i = 1,2, . . . , N − 1.
56
4. Markov Chains
Since p+ q = 1, we have
pPi + qPi = pPi+1 + qPi−1
iff
p(Pi+1 − Pi) = q(Pi − Pi−1)
iff
Pi+1 − Pi =q
p(Pi − Pi−1), i = 1,2, . . . , N − 1.
57
4. Markov Chains
Since P0 = 0, we have
P2 − P1 =q
pP1
P3 − P2 =q
p(P2 − P1) =
qp
2
P1
...
Pi − Pi−1 =q
p(Pi−1 − Pi−2) =
qp
i−1
P1.
Summing up the LHS terms and the RHS terms,
i∑j=2
(Pj − Pj−1) = Pi − P1 =i−1∑j=1
qp
j P1.
58
4. Markov Chains
This implies that
Pi = P1
i−1∑j=0
qp
j =
1−(q/p)i
1−(q/p) P1 if q 6= p (p 6= 1/2)
iP1 if q = p (p = 1/2)
.
In particular, note that
1 = PN =
1−(q/p)N
1−(q/p) P1 if p 6= 1/2
NP1 if p = 1/2
.
59
4. Markov Chains
Thus,
P1 =
1−(q/p)
1−(q/p)Nif p 6= 1/2
1/N if p = 1/2
,
so that
Pi =
1−(q/p)i
1−(q/p)Nif p 6= 1/2
i/N if p = 1/2
. ♦
By the way, as N →∞,
Pi →
1− (q/p)i if p > 1/2
0 if p ≤ 1/2. ♦
60
4. Markov Chains
Example: A guy can somehow win any blackjack hand
w.p. 0.6. If he wins, he fortune increases by $100;
a loss costs him $100. Suppose he starts out with
$500, and that he’ll quit playing as soon as his fortune
hits $0 or $1500. What’s the probability that he’ll
eventually hit $1500?
P5 =1− (0.4/0.6)5
1− (0.4/0.6)15= 0.870. ♦
61
4. Markov Chains
4.6 First Passage Time from State 0 to State N
P(n)ij ≡ P (Xn = j|X0 = i)
Definition: The probability that the first passage time
from i to j is n is
f(n)ij ≡ P (Xn = j|X0 = i,Xk 6= j, k = 0,1, . . . , n− 1).
This is the probability that the MC goes from i to j
in exactly n steps (without passing thru j along the
way).
62
4. Markov Chains
Remarks:
(1) By definition, f(1)ij = P
(1)ij = Pij
(2) f(n)ij = P
(n)ij −
∑n−1k=1 f
(k)ij P
(n−k)jj
P(n)ij = Prob. of going from i to j in n steps
f(k)ij = Prob. of i to j for first time in k steps
P(n−k)jj = Prob. of j to j in remaining n− k steps
63
4. Markov Chains
Special Case: Start in state 0 and state N is an ab-
sorbing (“trapping”) state.
f(1)0N = P
(1)0N = P0N
f(2)0N = P
(2)0N − f
(1)0N P
(1)NN = P
(2)0N − P
(1)0N
f(3)0N = P
(3)0N − f
(1)0N − f
(2)0N
= P(3)0N − P
(1)0N − (P (2)
0N − P(1)0N ) = P
(3)0N − P
(2)0N
...
f(n)0N = P
(n)0N − P
(n−1)0N
f(n)0N ’s can be calculated iteratively starting at f(1)
0N .
64
4. Markov Chains
Define T ≡ first passage time from 0 to N
E(T k) =∞∑n=1
nkPr(T = n) =∞∑n=1
nkf(n)0N
=∞∑n=1
nk(P (n)0N − P
(n−1)0N )
Usually use a computer to calculate this.
(WARNING! Don’t break this up into 2 separate ∞
summations!) Stop calculating when f(n)0N ≈ 0.
65
4. Markov Chains
2nd Special Case: 2 absorbing states N,N ′
Same procedure as before but divide each f(n)0N , f(n)
0N ′
by the probs. of being trapped. So probs. of first
passage times to N , N ′ in n steps are
f(n)0N∑∞
k=1 f(k)0N
andf
(n)0N ′∑∞
k=1 f(k)0N ′
.
66
4. Markov Chains
4.7 Branching Processes ← Special class of MC’s
Suppose X0 is the number of individuals in a certain
population. Suppose the probability that any individ-
ual will have exactly j offspring during its lifetime is
Pj, j ≥ 0. (Assume that the number of offspring from
one individual is independent of the number from any
other individual.)
67
4. Markov Chains
X0 ≡ size of the 0th generation
X1 ≡ size of the 1st gener’n = # kids produced by
individuals from 0th gener’n.
...
Xn ≡ size of the nth gener’n = # kids produced by
indiv.’s from (n− 1)st gener’n.
Then {Xn : n ≥ 0} is a MC with the non-negative
integers as its state space. Pij ≡ P (Xn+1 = j|Xn = i).
68
4. Markov Chains
Remarks:
(1) 0 is recurrent since P00 = 1.
(2) If P0 > 0, then all other states are transient.
(Proof: If P0 > 0, then Pi0 = P i0 > 0. If i is recurrent,
we’d eventually go to state 0. Contradiction.)
These two remarks imply that the population either
dies out or its size →∞.
69
4. Markov Chains
Denote µ ≡ ∑∞j=0 jPj, the mean number of offspring
of a particular individual.
Denote σ2 ≡ ∑∞j=0(j − µ)2Pj, the variance.
Suppose X0 = 1. In order to calculate E[Xn] and
Var(Xn), note that
Xn =Xn−1∑i=1
Zi
where Zi is the # of kids from indiv. i of gener’n
(n− 1).70
4. Markov Chains
Since Xn−1 is indep of the Zi’s,
E[Xn] = E[Xn−1∑i=1
Zi
]= E[Xn−1]E[Zi]
= µE[Xn−1].
Since X0 = 1,
E[X1] = µ
E[X2] = µE[X1] = µ2
...
E[Xn] = µn.
71
4. Markov Chains
Similarly,
Var(Xn) =
σ2µn−1
(µn−1µ−1
), if µ 6= 1
nσ2, if µ = 1
72
4. Markov Chains
Denote π0 ≡ limn→∞Pr(Xn = 0|X0 = 1) = prob that
the population will eventually die out (given X0 = 1).
Fact: If µ < 1, then π0 = 1.
Proof:
Pr(Xn ≥ 1) =∞∑j=1
Pr(Xn = j)
≤∞∑j=1
jPr(Xn = j)
= E[Xn] = µn→ 0 as n→∞.
Fact: If µ = 1, then π0 = 1.73
4. Markov Chains
What about the case when µ > 1?
Here, it turns out that π0 < 1, i.e., the prob. popula-
tion dies out is < 1.
π0 = Pr(pop’n dies out)
=∞∑j=0
Pr(pop’n dies out|X1 = j)︸ ︷︷ ︸πj0
Pr(X1 = j)︸ ︷︷ ︸Pj
,
where πj0 implies that the families started by the j
members of the first generation all die out (indep’ly).
74
4. Markov Chains
Summary:
π0 =∞∑j=0
πj0Pj (∗)
For µ > 1, π0 is the smallest positive number satisfy-
ing (*).
75
4. Markov Chains
Example: Suppose P0 = 14, P1 = 1
4, P2 = 12.
µ =∞∑j=0
jPj = 0 ·1
4+ 1 ·
1
4+ 2 ·
1
2=
5
4> 1
Furthermore, (*) implies
π0 = π00 ·
1
4+ π1
0 ·1
4+ π2
0 ·1
2=
1
4+
1
4π0 +
1
2π2
0
⇔ 2π20 − 3π0 + 1 = 0
Smallest positive sol’n is π0 = 12.
76