b
i
= ̂ Z [ J - p V _/2 J C a y /(x -y ) - /S w /(x -y )*
■ - Jo-dFu) + j + j W „ , + J - ^ , But
-* -II -* -«
_ rS & i/^ -y )* r C<»/(x-y)
I (( *x--y/ )) 0 J F(xT-y))
= /r -0
= (̂o+Ol “ ^u-0) + I f a ) ~ V . ) + ^ [̂ (o*0) — ̂ *(o-0) ■ - S " /> V " '* r A j .
“ ^*) “ (̂U) rim
II' (i and /) arc points o f continuity of F .
= iz
/ > V '
*a\J,
Example:
/> V "
If CO
/ / ■ ^ ( f - x ) f i r - *
Given ^J'1 = ( + pe")’ J ? ( f - x) /w/s | t “#/#”
, J O ? ( C i s L f i a ; -
•#> > > (/■ -x j /* w li
1 7 3
172
UNIVERSITY OF IBADAN LIBRARY
C H A P T E R 10
IN T R O D U C T IO N T O M E A SU R E T H E O R Y
10.1 Introduction
Probability theory is a part of mathematics which is useful in discovering the regular
features of random events or phenomenon. In probability theory, the sigma algebra
(which we shall define later) often represents the set of available information about a
phenomenon. A function (or a function of a random variable) is measurable if and
only if it represents an outcome that is knowable based on the available information
about the experiment, the event to which it belongs and the probability function.
For us to understand how a probability measure can be obtained, let us develop an
abstract model for the probability of an event particularly for infinite sample space fl
from a specified experiment.
10.2 Abstract Model for Probability of an Event
I.etfl be the sample space such that H = {w* i = 1,2, 3 ,........ }
w, are called indecomposable outcome or simple events.
The is a decomposable or compound events, that is Ex = {wj i = 1,2,3 }
The elementary definition of probability is
PART TWO r , ( r \ _ No o f fa v o u r a b le cases^ ' T o ta l n u m b e r o f ca ffles ...................... ' '
Since events are subset of H , it follows that the union and intersection of a finite
number of events and the compliments are also events.
(1) For the model of mirror reality, the operation above can be represented by
A, B, A U B,A n B, A , B . That is all statements about events can be written in terms
of u,.n.
(2) A random for defining probability in term of weights is to allow for the fact,
that some events are more likely to occur than others. The weight of a set is just the
sum of the weights associated to each point in the set.
174 175
UNIVERSITY OF IBADAN LIBR
RY
Let ft be sure event, the impossible event will be (p. Let A be a non-empty class of
subset of ft called events. Let P(be the probability) be a real-valued function defined Any collection of events is a class of events. Classes will be denoted by A, B, etc.
on A. Such that P(E)denote the probability of event E.
The pair A, P is called the probability field and the triplet (ft, A, P) is called the Example
probability space. Let ft be the real lineR containing all the real points w. i. e. ft = {w: — oo < w <
oojalso let
10.4 Axiom for Finite Probability Space A - {w : we(—co, a)} and
(i) If Ej 6 A for i = 1,2,..., n then B = (w:we(c,d)}
Define:
n n (0 A n B ; (ii) A u B\ (iii)Ac and Bc and give your assumptions
| J e< G A and f~ | Et G A (iv) Show that the compliment of an interval need not be an interval.
<=i i«i
0 0 If E G A .then E' 6 A Solution
(ill) If E e A . then P(E) > 0, also P(ft) = 1 A r\B =
d)
The number of possible outcomes of an experiment (E) may be finite ot infinite.
Ac n B = B i f a < c < d
Let w denote a sample point (an outcome) from the experiment. Ac U B = Ac i f a < c < d
Let ft denote the totality of outcomes of E i.e. ft = {w1( w2, ...} BCAc i f a < c < d
Let event A={w: w< eft}be a subset of ft. e.g
(i) B={Wj - oo < w < co); all values on IRL On your own. define the above if c < a < d ■ or i f c < d < a
(ii) C={wi: a < w < b}-, all values in the range (a,b)
Sequences and Limits
(iii) D={w,: w0); a singleton. A sequence of sets is an ordered arrangement of sets in order of magnitude
(iv) E={w,: Wj, w2, }; a doubleton. .Monotone increasing sequence: A sequence of { sets {/ln} is said to be monotone
(v) F={w: iv. = 0); an empty set.
( increasing if An Q An+, for each An.
The class of all subsets of f t is called the power set of ft such that if f t contains n If the sequence {/!„} n = 1,2,... is monotone increasing (non-decreasing) if for every
points, there are 2n subset of ft.
Thus, if f t is finite, the number of all possible subset is also finite. n, wchave An+1 3 An
Then the limit of (/ln) is the 3mm of the sequence i.e.
The power set of f t when ft = {w,. w2, w2. w4) => 24 = 16
176
177
UNIVERSITY OF I ADAN LIBRARY
(ii) The limit of {An} is said to exist if limAn = lim An = A,
A = Y An = lim An (iii) If {A^} is not monotone and A exists then An —» A i.e. An converges to A.
nZ_ilj n—oo (iv) Even if 1im An does not exist, limAn and lim An will always exist.
or
■ n a OO
= Example:( J1' 4* = An; U a ‘ = a i. e. An T Ak k = l Consider the sequence {4n} where
= {w: iv belonging to all Ak except Av ...
A„ = w: 0 < iv < b + ̂ ^ " /n ; (b > 1)
CO
Sup
= i > Does the series {An} converge?.
k = T
For any arbitrary monotone increasing sequence {An), the limit is
OO OO
C = linAn = li sup Ak = |~= l| Ak k[=Jn Ak Solutionk fiv: 0 < w < b + —\ ,‘i f n is even,
Monotone decreasing sequence: A sequence of sets {An) is said to be monotone Let Cn = * nJ
[w: 0 < w < b + ( 7 (n + x)) j ; i f n i s odd
decreasing if An+1 Q An for each An.
If {An) n: 1, 2 ,...) of events, is monotone decreasing (non - increasing) and for every limAn = {w: 0 < iv < b)
n we have An rj An/+,',1 , then the limit is the product of event [An) i.e. Similarly,A — Final An — limn_=oo ‘A4n or [w: 0 < w < b - (Vn)]l t f n s = {w: iv belonging to at least one o f An, An_ i ...)n oo { [iv: 0 < w < b - (V (n + X) ) ] ; i f n is evenr- 1i i e A n l A limA„ = {0 < w < b]k= \ k Therefore, lim An * limA„
For any arbitrary monotone decreasing sequence {An},the limit is
Hence. {/!„} does not converge
OO
In f Exercise:
k=n If An = A: n = 1,3,5,...
= B:n = 2,4, 6, ...
Limits: B - UmAn = lim inf Ak =
k - l k = i Show that lim An = A u B, limAn = A n B
Note that When docs lim An exist.
(0 linAn £ linAn
178 179
UNIVERSITY OF IBADAN LIBRARY
Exercise: Corollary:
Examine the following for convergence, if convergent, derive the limit;
p | /l, = i4x + A\A 2 + ACXAC2A^ + -
W ^ - = (0 ,V 2 n ) .^ " « = [ - l . V(2ntI)] £=1
If
tb) An = | the s e to f rational in ( l - 1/^n + ^ 1 +- */n)j co
(c) An = 2-1/n, 2 + 2/n), n is odd. W 6 P j Ait then w belongs to some /lf
i=i
Thus w may belong to Ax or Acx or Acx or A2 or A \A \ i.e. w G Ak for some k.
10.2 Obtaining Countable Class of Disjoint =» iv £ U ?.i establishes equivalent of both sides of (*)
Lemma 1.1: Given a class = 1,2..... n}of n sets there exists a class
{/?,-, i = 1, 2.....n) of disjoint sets such that U”=i At = Ef=1 Bt
10.4.1 Definition: Additive Set Function
Proof: By induction A set function /u is said to be additive if V A,B,sJ. =
2
Note
Then = ( U ^ / l i ) U /lm+1 •
• Once the value + oc,-oojs not allowed i.e. *-co
• If all the values of «?’ 'Sl> vv . a -ADAN LIBRARY
For every decreasing sequence {En} 1
Theorem
s.t.
n0.
The probability function Plmi is a set function that has r - additive property and hence
A set function is said to be continuous if it is continuous from above and below.
is a measurement space.
Theorem
Let cp be finitely additive and continuous from below, then f i is - additive. Krample: Let (O..F) be a measurable space on which a sequence of probability t
measure Pi,Pl ,...Pu... defines a set function.
Proof < j
Show that P [ E )d . / ^ J— /^,(£)is an additive set function.
Given a sequence of disjoint sets{£„}, then 2" .,
Solution
It is required to show that
Let N be a finite number, since (p is finite additive, then (0 0 < /? .,< l
(ii) Plm, is counrably additive and is a measure
V.ns| J » “ l (iii) Prove that P(f2)=l
»N» NSi i -VX sn
-fl
I“ iX . . (y) t̂£) = ~ ̂ (E) + ^ r P:(£) + JT PJ(M + ”
Let Sv = En be an increasing sequence but -l P ,m 2 0. - L p 2lE)> 0 ...
l»«l
(p\?im S y J= Cim tp(Sn) and +
2-
y - i - . s - 2 _ = _ l » i
n r2 n “lfl * i - r i _ y 2
■ # ■ ) 0 >.,< I
(ii) 1.x! 1/i, i be a sequence of disjoint set, it is required to prove
By finite additively
•/ «* iv .y| , » ._. \_ > 'i» )1
nsI ns| from i .1’ c
(pis r - additive.
I S3
1 8 2
UNIVERSITY OF IBADAN LIBRARY
* 1 ® Proof: (Using mathematical induction on n)
4
1=1 / = nm| Uu-i * / for n = l: P {E,) = )
for n = 2: P (E , [ J E 2) = P f c ) + P (E 2 ) - P ( E xV \E2)
= t ± t r M
nm | ^ 1=1 The result is true for n = 2
Since each of /*„ is a measure and 0 < P(t) < 1 » = 3: />(£, U £ 2 U E, )= />(£,) + P(E2) + P(E,) - P(E, R E2) - P(E, R £ , ) - P(£2 R £ 3)
* +p(E ln £ , n E J)
■■■-I
1 1=1 - Z*-i ' W L, 1«-i ?z !J
= 2 > U ) * i
i-i Assuming it is true for n and also tme for n = m -1, we have
- Z1-1 ^ ) p (£ ,U £! u ...U £.,.i) = / :( u £. 1 = Z />(£ . ) - i > f e n £ y)+ I f e n ^ n s y )V i - i / <■ ! i< /< y< o«+ i i s i < y < i & n + i
^.)is countably additive. t ( - i r !p (£ ,n £ 3n ...£„ .,)
(Sii) = 1 ^ . ( 9 )
« = m : i { y £ , l = / ( [ j £ , U £ . ] - i p ( £ , n £ y n £ „ ) + x ^ n ^ n ^ n * . , )
V i - i y V i - i ) i s < s y s * s o » - i
=*±=l z± (1) 1 r £(£, n e 2 n ... n )
1 1 1 . Assuming it is true for it = m, we need to prove that the theorem is true for n = m +1
2 4 . 8 -
>1 • S in 6 S K =.— = - ^ _ = l
• -1- ^ : m
Le tE = [ jE n then
/-i.
10.5 The Halley-De-Moivre Theorem ^ Q e ,.J = P ( E U E .J
Theorem: Let { f jb e a class of events each of which belongs to a r - field 91, and
each of which may or may not occur. Then
= ! > ( £ , ) - Z p ( E ,n £ , ) + ( - i r + z ^ M ^ n £,)+ ...
P\at least one o f the event E\ occur} «■ I ISi<*> IS/< j ( £ . ) - Z ^ . n £ j + Z r a . n £ / n £ ‘ +( - T l« ’(n £ ,) This implies that the result is true for all positive integers n.
V i = | / » « l IS li j& n IS 1< 7< *S » i
1 8 4 1 8 5
UNIVERSITY OF IBADAN LIBRARY
Erwuupie: Sotuttrn:
Lc-i 11 be events which belongs to a r - field % shew that^e probability (j) Let l:\ denote the event that ihe i h letter and envelope •march
that exactly K events occurred out of n is given by
ir./!
Where S, * fl Ea f | ... D £» ) r y~ r3 X4-■ l- r 1 Since e 'x = 1 - x + :----— + ----- ...
.. 2! 31 • 4!
From Halley-De Moivie theorem
)E.] = Si - S 2-+Si - . . . + ( - l} - 'S ll {i.e.Pmi?(q/'l or 2 o r3 or...or N match) envelope ■
<«i J
If k = 0, no event occurred: .v..p(|Ji:, j - i -er'1 = 0.63212
V<»i .>
-“(A*, U£, U...U £ j - 1 ■- / ’(ft'U E, U -U £n) . = 0.6
= 1- S’. r V o . t f - S . , ( n ) Takirfe limit as N —>x>
, . - P{nOn o f the events occurred) _ l r 1 .
______ j i”
L ___ 4I-
Example 2: . . . 2! 3! 4!
Suppose /r letter and corresponding envelopes are typed by a typist. Suppose 1 + 1 1 1 1_____ .1.
further that the messenger, who is in a hurry to leave for the post office, randomly 2! 3! 4!■ft
insert letters into envelopes, thinking erroneously that all the.letters were identical.
Finch envelope contains one letter, which i.e. equally likely to be any one of the p 10.6.2 Countable Probability Space
letters. : Sometimes it is impossible for all the sample points in a fi to be equally likely.
Hence, each P, is viewed as unit probability mass among the sample points following
li) • Calculate the probability that at least one of the letters is inserted into
a certain rule or law. This law is sometimes referred to as probability distribution.
its correspondence envelope.
(it) Find the limit of this probability as N -» oo
Example: Fora geometric distribution
Suppose Q = {0.1. 2. ...}and
/ » , , * ( \ - 6 ) 0 \ x = 0. 1,2...... (0 < ^ < 1)
m 187
U '—NIVERSITY OF IBADAN LIBRARY
Then Pt - P[x) > 0, £ / | v) = J ,2) From (1) let \ = {{a}, fb. c}, fd}, fi, 0} and P{a} = P{d} = V 4 c) =
Pffi} = 1, P {0} = 0.
If ACO^then P(a] = Y j P[a ' The triplet (fi, L P) is not a probability space since ̂do not form a field.
xO.4
Poisson Distribution Exercise
C-et 0\ x = 0, 1, 2,... (2a) Is f = {/Ft, £2» ■*•.£*} afield.
(b) Hence or otherwise obtain all the elements of the a - field of t.
T^en P(s\ is a Poisson distribution and X is a Poisson random variable
(3) Consider the sample space
F = {0{W!, w2}, {w3, w4}, fi}
Definition 3:
\ f A = = {w3#w4}
A class of sets A is called a field or a - field if and only if the following conditions Show that F is a field; •
hold true.
1. If E, 6 A, then U"=i Ei 6 A
Exercise:
2. If E 6 A, then E' G A Let Er,Ei, '..,En. denote an infinite sequence of events in a — field A.
From the above, it follows that Define
3. If Ej 6 A implies U"=i Ei G A , " W . .
Example h m=n . •
OO
A = {fi, 0 }is a field.C B„=IB == [A, A }is a field m - n{A, H, 0} is no t a field, since A g C (a) Prove that BnCEnAn V-n .
G= (A, B . A B . A U B.A U B , a U B ,A U B , A B ,A B , A B , A B , fi,0 } is a field. (b) Show that {/4n} is monotone decreasing.
The class of all subset of a given set fi is a field. (c.):show that {fln} is monotone increasing.
Example 2: 10.7 Sigma Field (o - Field)
(1) Let fi = fa, b, c, d} and 5 = {{a}, fb, c, d}, fi, 0} A non-empty class of sets which is closed under complementation and countable
i-e. ? a field? Yes P(a) = ^ , P(b, c, d) = ^ unions (or countable intersection) is called a field.
//ps-CQ, l;, P) is a probability space. Note:
Yes, since ̂forms a field. • A field containing an infinite number of sets may not be a 0 - field.
1 8 9
1 8 8
UNIVERSITY OF IBADAN LIBRARY
M I
Into section oi an arbitrary number of a — fields is a o - field.
10.8.1 Borcl Set
1 0 .8 B o r c l H e l d Borel field and Borel sets play a very important role in the study of probability.
Hus is a subset of the real line. Let C be a class of all intervals of the term Monotone field: A field A is said to be a monotone field if it is closed under
(--oo, .v).* G IR as subset of the real line !RL Also let ( 0 = Tl be the minimal field monotone operations, i.e. if lim An e IF whenever {IF} is a monotone sequence of set F.
generated by d. i.c. Ane F./t,, T A => A e F
f hen 'ft contains the intervals of the form [x, oo) (i.e. compliments of (—«>, a), it also A„e ¥.An l A => A e F
contains the intervals. Theorem: A a - field is a monotone field and conversely.
Proof: Let A be a a — field and Ane A. If An T A, then A = U„ An\s a countable
(-<»,a | = n ( - 00, a + “ ). by coutable intersection
union sets of A\. Hence A e A. Similarly, if An l A.A = C\nAn is a countable
, ' (a.oo) = (-oo, a |“ by complimentation
(a.b) = (—co,/;) n (a, oo),a < b intersection of sets of A.
(a,b\,[a.b).etc fo r a.b G K. ••• A e A\, hence, A is a monotone field.
Conversely, let A be
I enuna n n a monotone field and let Ax, A2 .... be sets belonging to A.
1 ct be the class of ail intervals of the term ({,b),(a > b)a,b e IK but arbitrary. Then ( J Ak and j "~j Ak belong to A\ since A is a field.
k=l k=l
Then a ( t \ = V).
Proof: By (*) (overleaf) a.b.cty for all a, b. Hence, These are monotone sequences whose limits ( J Ak and p |A k must belong to A.
By definition of minimal field. a ( t x) c i k= 1 k=1
fo prove inclusion Thus A is a a - field.
Let x e (a.b) then.
U“ i( - n . x ) e o ( e x),V 10.9Kandom Variable in Measure Space
l et ft be the sample space with sample points w. Interest is usually in the value n*)
=> (-oo.Jf) a (et) ^ x associated w ith w.
l' c rr(/',) as defined in the last example. (a) Point function: function on the space ft to a space ft assigns to each point
If is also possible to prove that die Bore! field is the minimal field containing any one w e ft a unique point in ft denoted by X(W). Thus X(VV) is the image of the argument w
of the following-
under A' i.e. value of X at w f t —— » Q'
e, = {(—oo, x |.x 6 IR} i/rnuw n r u n g r
f ;i = ((a.ftl.fl < b.a.b e r- ’ The set Q* = |X(lv): we ft| which is a subset of Q’ is called the strict range of X.
f , - ([a. b I, a < b . a . b < ■'} If i f £1" => X is a mapping from ft to ft.
C,, - {|a.b),a < b.a.b e IK) The symbol X(vv), etc will be used to denote functions even though they denote
— 11 a c o ) , v t- IK) . t c . values of functions.
: mi
191
UNIVERSITY OF IBADAN LIBRARY
Kxample 1: X_1((w}) = {{w c“ft}: X(vv) = w1)
Let n = |0, ±1, ±2,... |; ft' = (0,1,2,... |
Note that for a point w' e ft12 one or more than one points in ft whose image under
ft = 10,1,4.9...,]; i f X {w) = w2
X is w l. Let /!' c ft1. The set of all point for which X(W)e /?1 is called the inverse of
Thus .Vis0 a mappin0g of ft into and onto Q*...i - 1 i /Sunder X denoted by X-1 (H1)- % ,i- 2 2 With every point function X, we associate a set function A”-1 whose domain is a class ± 3 3 s(j of subsets of ft and whose range is a class'^ (say) of subset ft. Then, X-1 is called the ‘inverse function' (or mapping) of X.
X ( B ) = |X(w):w E B \ . B C a
iv, = w2 => X(wl) = X ^jone - to — one X ~ l ( y ) = \ B ( B ■): B e y I
In this case X(wl) * X(w2) A w, = w2 X -I( f t) = [w:X(w)c f t| = f t
X(w) = w2is not 1 — 1 function Lemma:
Inverse mapping preserves all set relations.
Since w, = 4-2, w2 = — 2 have the same image
Proof: Let W c C c ft’, then
X(w i) = 4 = X(iv2)
X '(« ) = \w:X{w)e B \ c \w.X{w)eC\ = X ~ \C )
If ft is the real line (-co < w < co) and ft = (0 < w < co)then X(w) = exp(w) is a
(d) Indicator Function
1-1 onto function from ft to ft and 1-1 from ft to ft. If the range space is & or its A real valued function lA defined on ft as
subset, the function is said to he a .numerical’ or ‘real-valued’ function.
xa _ (= 1 if w e A
(b) Set Function ' " A (= 0 if w e.Ac
II the arguments of a function are sets of a certain class, then we have a set is called an indicator function (characteristic function by some authors). The strict
function. Suppose Zl- E A\, we associate a value p(/l), (say) then n is a set function. f.i range /,,is /„(ft) = (/„(w): weft} = {0,1}. If B is a set function and B c R, the range
may represent entity such as weight, length, measure, etc. space then lA l(B).= . ifB does not contain '0' or T
The interval (a,b) may be associated with b - a\ f(a , b) U f (c .d ) = (b - a) + = A, if U contain 'T but not '0'
(d - c), etc. - A(\ if 13 contain '0' but no t'T
I wo real valued function X and Ton ft are said to be equal iff X(iv) = Y(w) V- w e ft. = ft, if 13 contain both '0' but not'T
i. c. X = Y Thus IZl(B) = { lA < lB
denoted byX~‘({w}). The A r = 13 <--» l,\ = 1 .m
UNIVERSITY OF IBADAN LIBRARY
/„ = /2(/t) = l A(.A).!n = 1
Proof:
(ii) l(A(:) = 1 - [(A ); l(B—A) = 1(B) - >A
o n »i 00 l*i.B ...........................................cA
m u . u = f^ j^1 = I >
j=i t=i
(=i
= min (Iai*-—Un)
0 V)*(AuR) ~ U +n *B ~ IA- ' b “ maX Oa- ' b)*(A+B) = *A + 1 [i=Ji A i = ^i= l l A i ^ l A i A Ai + ^i= i 1A i ^ * A i A Aj + A k
Let Bk c fl ,then w eX_1(n Bk) <=> X(W)€ n Bk => P | B.elB
<=> X(W)€Bk Ft 1Thus, B is closed under countable intersection. Hence, IBis a o - field.
<=> w €X_1(Bk) f1t
« w e n X_1(Bk) 10.9.1 /(A) as a Measurable Function
Hence, X"'1 Since lA '(B) = { weX-'(B)
.■.X~'(BC) = (X -‘( S ))c 10.9.3 Function of Function
Clearly ( f t ) = [w:X(Ml)cft') ^ f t If X is a function from ftto fl and X is a function from fl to fl , then the function
X~'( X~l (a (f)) c A
10.9.5 Random Variable (Economic Definition) => A
Suppose ft be a sample space. Let A\ao — fieldof events associated with a certain
fixed experiment. Any real value A\ - measurable function defined on ft is called a 10.9.6 Vector Random Variable
random variable. Thus,’X is a random variable iffB~x, the a - field induced by X is Suppose w eft, the associate X(w) = (X\wy, Kfvv)) a point in the 2-dimensional
contained in A\.
huclidian R2. The Z define a function from ftto R2. Consider the class of 6 of all
Suppose we define two non-negative functions rectangles bounded by the lines xx = a.x = b,y = c,y = d,a < b,c < d arbitrary,
X(w) = *(w)» '^X(w) ^ 0 flic minimal o - field containing f in Borel field (332)in R2.
= 0, it X(W) 0 /. is called a 2-dimcntional random variable ifZ "’(332) c A\.Z~l(®2) «s a a - field
and induced by X.
X(W) ~ (̂w)> < 0 Illustration
= 0 if X(wj > 0 QO •
The above are respectively called the positive and negative parts of X. Then S„ = £ X t, E(Sn) = nA.a(Sn) = VTil
A' ’ and X are Borel function of X and will be random variable if X is a random i- I
variable
Note: flic moment generating function of Znis given as
1 11 These functions play an important role in the theory of integration of M _ „CO , (VJ
probability function.
(2) To show whether a function is a random variable, it is not necessary to
determine whether X~l (B)e A\ for every B in 33. It is sufficient to verity
X ~ l(f) c A where C is any class of subsets of R given in sub interval on log/W/(t) = —t'fnA - nA - e
page 8. = - t V 5 * - .U ( i - U + j s + s 5 + 5 ^ + - 1 )
1 % 197
UNIVERSITY OF IBADAN LIBRARY
lini log » 1 , *,2 CHAPTER 11
l|-«CX» ' * /.(£) = —2 => My{t) - e 2
LIMIT T HEOREMS AND LAW OK LARGE NUMBERS
= m g f o f A/(0,1)
Problem 11.1 Introduction
Suppose that S„ has the binomial distribution b(n,p)- show that
distribution The law of large numbers is concerned with the conditions under which the average of
%n----------- » N (0. 1) a sequence or random variable converges (in some sense) to the expected average as
the sample si/.e increases.
Theorem: Let Yn, n > 1 be a sequence of real converging to Y0 Then the sequence
r,+y2 y,+rz+r:, y1»y2+-.y,i
x ' 2 ' i n 11.2 Concept of Limit s
Also converges to Y0 However, the inverse is not true. Let .v„ be a point in some intervals oflhe real line '.H. Let / be a function which is
Proof: delined at every point of / except possibly at .v„. The limit of the function as x
Let > 0, we find n, s.t.n > n => (V̂ + ••• /„) — K„| < £ approaches v0 is /, written as
Since Yn -» K0 3 no s. t. |Tn - K0| < e/ 2 K > 1 irn /,\, = L ov /( f| > L as x —> x
l ind > n0 s. t. — K0| < e/ 2 for convinence If for any positive number X (no matter how small) there is some 8 greater than zero
We claim that n > n, => - '"0| < £ such that
Then iriii ,/ I _ |(yt+yo)+-t(yno+»b) |/ ( i Z.| < £■„ for all 0 < |.v~.v(1| < S1 it K°l “ I n
|(yl+yll)+- + (y,,o+yo) + (y„oM+y0)+--My,t + y0) I rom the above definition, the number e > 0 is first given, then we try to find a
it! n number d > 0 which satisfy the definition.
Example 1: Prove that < in i (3.r — 4) = 14
n 0 n
V \Yi- Y 0\
n i—i n0 + 1 Z_i Solution: Given A > 0 , find 8 > 0 [depending on 1 .1 s.t. 0 < jv - 6) < d. we have 1=1 i=n0+l
\ f - 14| < e
* * / 2 + '-T *el 2 -> |3.\ 4 14| |3 ( .v - 6 | 3j.v 6| < 38
S £/ 2 + e/ 2
N o te th at | \ 6 j < A
0} < -â
Example 2:
x + 1 ProofProve that ( Suppose X is continuous with density function
'-*2 3*+ 4 10
Solution: Given £0 > 0, we went to final
0 < |x - 2| < £, we have < = [ V w * + « £ V ( x) a
> £ x f ( x)dx
/ ( x) ^ _ * + l 3 x + 2 x - 2 • 8
J ' 10 3*+ 4 10 ~ 10(3* + 4) 10(3x + 4) ' 10(3x + 4) > j~af(x)dx
If x is sufficiently near to 2 so that = o ^ f ( x)dx
3.r + 4 > 10, thus 1 <1 > aP(X £ a)
10(3*+ 4) 10
/. aP[X > a) < E (X ) => P(X £ a) < -̂ a
Thus l / M — I 10(3.r-4) 100 The above is for a single variable X. Suppose we have a sequence of variable
8 = 100 £ {X„}, n = l,2,...n, then we have the Markov’s inequality for a sequence of {A',,}as
Theorem: Let / be the constant function defined by = C where C is a constant
P X .
tint f ( x ) - C
Proof: Given s>0, find 8 > 0 such that 0< |* - t7|< 5 =>|y|T)-Cj<£- 11.4 Bienayme-Chebyshev’s Inequality
Theorem: If A is a random variable with mean n and variance a 2, then for any value
The distribution of certain statistics of interest are too complicated to derive for e> 0 :
differing sample sizes. In many cases, limiting distributions can be obtained as an
approximation to the exact distribution, when the number of observation N is large. O"
Thus, most important theoretical results in probability theory are limit theorems. Proof:
Let consider some useful limit theorem. Since ( x - r t is a non-negative random variable applying the Markov’s inequality
11.3 Markov’s Inequality with a = k \ we have P{{X - /j )> K }< —-— &,then the above (*) is equivalent to
Theorem: If X is a random variable that takes only non-negative values then for any
value u > 0
2 0 1
200
UNIVERSITY OF IBADAN LIBRARY
lim P{jXM| >oo, it does
The above inequalities are important in that they enable us:
(i) derive bounds on probability when only the mean, or mean and variance of not follow that for every e: > 0. we can find a finite n0 such that for all n > n.'; the
the probability distribution are known. relations \X,\ " ( i - / > r
(i) Suppose it is known that the number of eggs sold in a poultry farm in a
month is a random variable with mean 75 crates. To show that lim pfx„|>e}= 0Hf* * ’ '
(ii) What is the probability that the sales for next month is greater than 100
crates. Solution:
(iii) Ilf the variance of the sales for the month is >5, determine the bounds on By Chebyshev’s inequality we have
the probability that sales in the coming month will be between 50 and 100
crates. E(Xh)- n.p\ Var (x)= or
v n
Jnpq
Solution: Let X be the number of eggs sold in a month But Chebyshev’s inequality states a = r
(i) by Markov’s inequality ■Jn
a 2 /
P(X>Vto)< — = - P \X - p \ > e } < ^ - fo re > 0.
v 7 100 4
(ii) by Chebyshev’s inequality a V ,
or P \X \> K a < ^ r - 4 - — 7
/>).*'-7 5 |> 25 = — 1 n| 1 k2a 2 nK:
' 1 1 252 25
p \X n\ > k o } < \
p \ x - 75| < 25} > 1 — — = — A
1 1 J 25 25
Lotting 11 — - wc have
So the probability of sales of eggs for this month is at least —24
Definition: The sequence {Xn} of a random variable is said to be stochastically P\X |> e}< -^ - = n e
convergent to zero if for every e> 0 the relation
2 0 2 203
UNIVERSITY OF IBADAN LIBRARY
PrjX . - n p \ > e } < & 11.6.1 Weak Law of Large Number (WLLN)
Let Xx,X2, ... be a sequence of iid random variable’s each having finite mean E(Xi) =
Chebyshev’s Inequality H. Then for any G > 0
This theorem is often used as a statistical tool in proving important results in statistics. > 6j _ 0
as n > co
For example: — p
lfVar(x) = 0 prove that 1'his implies that Xn -* fi
PX = B(x)=\ Proof: suppose the random variable has a finite variance a 2
Proof by Chebyshev’s inequality, for any 0 > 1. £ ^ y ar f Xi+X2..Xn\ _ — o It follows from the Chebyshev’s inequality that
as n —» oo and using the continuity property of probability and Chebyshev inequality.
Thus, as as n -» oo
n-*oo (. X\ "h X2 ...xnlim P ( -A * > € } =
” P (E2. {| 1 - " i > ^ } = 0
x n -* n
=> p[x * n\ = o
This implies strong convergence (Strong convergence) Convergence Almost Surely
X n is said to converge to X almost surely, almost certainly or almost strongly
11.5 Convergence of Random Variables
Convergence in law denoted by denoted by X „ —— if Xn(w) —> ^ (M.,for al w except for those belonging to a
L
Xn -» X if at every continually of X through distribution function F of null set N.
Limn_ 0o /y,(x) ^{x) Thus X„ -^ ± -> X iff X„(w) Ar„,.) < oo
Where Pn(x) denotes the distribution function of Xn Thus, the set of convergence of {Xn) has probability unity.
Lemma:
11.6 Laws of Large Number X n———>X iff as n ->co
This refers to the weak or strong convergence of sample mean
X # = %i t0 a corresponding population mean (/j ). -> 0 V an integer
2 0 4
205
UNIVERSITY OF IBADAN LIBRARY
Proof:
Theorem 2:IfX„ —-~>Cimplies that Fn(x) -» 0 fo rx 1 forx>Cand
Now AyOv)-* .^(w), if for arbitrary r > i , there exist some
conversely.
/?„(»•. r)s.l.V K >n„ (w,r), \X „ (w ) -X (w ] < / r Proof: If X n ——» C, Fn (x) -> F(x) where F(x) is the d.f. of the degenerate random
Moreover X m — > X imp lies that P[Xn - X] = 0 variable which takes a constant value C since
Using de-Morgan rules, fO, x < C •
,-'1 {l, x > X
[ » : \ X . ( w ) - X M \ > y $ \ = 0
r n Conversely, let Fn(x) —> F(x) as defined above.
i.e. for each r Then PrjjT ,-C |S;e] = P[A', £ C+s]+Pr[X„ S C]
[ w . \ x A » ) - x ( .w ] z . y r\\± co.
& t f u k - * l * X l ] = 0 Hence A',,--
Suppose X n's are discrete random variables taking • values
Replacing the above by the complimenting condition we have
o ,i ,2,...s.t.p (x „= ;)= /> ,.if p„->p, a s n - » 0and S takes value
/{ n k - * i < x ] ) - >i / with probability Pt (i = 0, 1, 2 ,...) and hence = 1 ■ then
Note:
Lemma 1:A sequence of random variable’s converges a.s. to a random variable iff the .ICT
sequence converges mutually almost surely. So that X a converges to X in distribution.
Lemma 2:If X n — — >A\then there exist a subsequence{Xnk}of {Xn}which Examplc:Let A^be a binomial random variable with index hand parameter
converges a.s. to X. P„ s.t.as n -> co, "Pn -> X > 0 and finite. Then we can verify
Convergence in Distribution (K = 0, 1, 2,...)
If F„[x). is the d.f. of a .random variableX,, and F(x) the d.f of random variable
The binomial random variable leads to Poisson random variable with parameter X in
X . then }Xn} is said to converge in distribution or in law or weakly. It is denoted as distribution as n-*co.
X" ——-> X, Fn -> F weakly or Fn(x) -> F(x).
Theorem 1: If X„ —— >X , then F„ F(x), x e C(f)
206 207
UNIVERSITY OF IBADAN LIBRARY
, , k - n p ......
Convergence in r,h Mean Now let y p = " r - ■ ...... .(**)
A sequence of random variables is said to converge to X in the rlh mean, denoted by Jnpq
Since n and k are large, we can expressed the factorials in (*) above by means f the
Xu ~±~>Xif H\XU - X\r -> 0 as n -> « .
Stirlings formula/approximation as
For r = 2, it is called the convergence quadratic mean or mean square.
e*"1
For r = I. it is called convergence in the first mean.
Lemma: If X „ ~ r~> X => E\Xn\ -> E\Xf b(k, m, p )— --------------
Proof: For (r < l)put(Xn - X ) and X for X in tyheincqulaity
p k q " ' k e e
4 ' f. r s £ K - * r + 4 * r (2^ (2n Y n k*y> (n-k )r k*Yi
Interchanging X n and X in inequality and combining, we have
» ( nq
E \ X l - E \ x \ < E \ X a - X \ (2n Y ^ k ( n - k ) U A n - k
Thus X n — -r-* X => E\X ,\ => E\X\ Where 0 -O lH) - 0 (k)- 0 ( n - k)
. I UsingM< l[ I +i +_l_
Then /*{.¥„ - X\ >e} < from ( . . ) the above can be rewritten in the form
Substituting for k
: . X m- ~ * X as n .-**>.
1 1
Lemma 1ltf|<---- + —+ P
The binominal distribution b (k ; n. p) approaches the normal distribution as n <- oc 1 1 12n 1 + X k K q q ‘ XtV / " ‘IJ
i.e. /?(k; n ,p )~ - = T ‘---- e ^ If we assume t,h a.t Vr« ->v nu aass nn -> °°, then 8 —>0 and e* —> 1 ■
J i n npq
Proof
Let A(k;n, p) = n\
K l(n-K ) p
_ k qn-K for large values of n.
The above represent P[SK - K ) where S t is a random variable which denote the
number of successes in n . Bernoulli trials with probability success for success in each which can be approximated by
trial.
11’we let n >xandkeep /’ fixed then p \S n - np\> np) ->0 ¥ e > 0 by the law of J — for large n Jnpq
large number. Accordingly \K -np\jn -» 0.
209
208
UNIVERSITY OF IBADAN LIBRARY
nq 11.6.2 Criterion for Convergence in ProbabilityTo estimate the quantity ̂̂ J |̂ — U- • « The following lemma gives the necessary and sufficient condition for convergence in
T'aking logarithm of the above gives probability.
Lemma
« ' o e[ y ' y ( n - K ) iog { y n _ k)
Which can be rewritten in the form x - W E ( £ i ) ^ 0 a s n *
|x„|1
X >0 iff E 0 as n -> co.
- t i p \ + xk ~ log
\ n p 1 + * J — U W JPJ
f Proof
nq 1 +■**,/— log IXI ’ • |x Inp I+X* J ~ For any X , the r.v. is bounded by unity. Taking g(x) = rj-77 1 for € > 0i nP J \V»\
Upon substitution for K.
Since x /̂7 ^ is small, we cam expand the Logarithmic function in power series. J _ W _ 1 — 1- < e [ - 1 ^ ) /li+KlJ 1+e * i ,+w j/ r+e
Using the Taylor expansion then a reminder
From RHS E M
log(] + x) = x - y + ^ - ; (o< |e3|< x ) b + w j
(* *) above becomes From LHS p \x ,\ >e]-» 0 => j 0
- | * ; + C X V i '
But K is a non -negptive r.v. so
Where C is a constant ■K
If we assume y -» 0 then can be approximated by E\ >o
— x: Theorem: Iff [x) is a continuous real valued function and
X, — X, then f(x)
Hence (*ii)is asymptotic to e ^ x‘
Gathering the estimates (i), (ii) and (iii) above, we have
b( k; n, p)~ 1 11.6.3 De Moivrc-Laplace Limit Theorem
J i n npq
If S is the number of occurrence of an event in n independent Bernoulli trial, with
The normal approximation to the binominal distribution.
probability E for success in each trial, then
2 1 0 211
UNIVERSITY OF IBADAN LIBRARY
Now using the Chebyshev’s inequality
:.p\x,\>e) 0 as n -»oo.
11.6.4 The Weak Law of Large Number = ><=}< ^ ■■, " • It follows that lim P(jXn| > e)= 0 as n-»co.
n s ' n~'*
Let A'l,X ,,...X nbe a sequence of independent and identically distributed random
variables each having mean E[X,)=/u and finite variance a 1. Then for any e> 0 11.6.6 Strong Law of Large Number (SLLN)
P^X - p| > £ j —> 0, as n —»oo This refers to the strong convergence of the sample mean to the population mean,
Proof i.e. Xn => E{Xi) = p
It follows that
i. e. Jirn P{sup \Xn - p | > e ) = 0
e [x ] - // and Var (x)= —
n Or
From Chebyshev’s inequality we have P lim [Xn = ix] = 1
n —*on
’ 1 n s"
Note that SLLN holds iff the population mean exist.
/.lim p jx - / / |>e } -0 Theorem: Let /V„X?, .- .X n be a sequence of independent and identically distributed
This theorem was first proved by Jacob Bemowlli random variables each having a finite mean // = E{X ,). Then with probability 1
X—x- -*- -X- ,-- -+- -..-.- + A,,11.6.5 Bernoulli's Law of Large Number ---- - /j as n -» co
Let fc} be a sequence at random variable with pdf Or Prjim (X, + X, +... 4- Xn )/n = p j= I
\
P 'O - P ^ f o r 0 < P < 1 and r = 0, 1, 2,...
Theorem: Let [Xk),k = 1, 2 , ... be an arbitrary sequence of random variables with
Further let X n = Yn - P sequence of random variables {A',,} is stochastically various ok and first moment Mk. If the Markov’s condition (i.e.lim n_co ak = 0)
convergent to 0 for any e> 0. i.c. lim P(jXn|> e) = 0 is satisfied then the sequence [Xk - Mk) is stochastically convergent to zero.
Proof: Proof
Suppose Xk arc pairwise untouched (i.e. independent). Consider the r‘h variable
W c have E (X J = 0 *, + *2+••+*„
Ym =
n Wc have
212 213
UNIVERSITY OF IBADAN LIBRARY
l A
C H A P T E R 12
E(Y^ = n k=Il^
Such Xk are pairwise uncorrelated, we have P R IN C IP L E S O F C O N V E R G E N C E AND C E N T R A L L IM IT
^ 20 'n ) = T H E O R E Mp4 LV "k=i 12.1 Introduction
The Central limit theorem is concerned with determining conditions under which the
If ̂ n sum of a large number of random variables has a probability distribution that is
Ifl im — V at - 0 approximately normal.n -00 n 2Z_i
k<=l
Then by Chebyehev’s inequality (theorem) it follows that 12.2 Convergence of Random Variable
A sequence of random variables {Xn} is said to be converge to a random variable .A" if
nl-i*moo />[|rn - E ( ) 'n) |> e ] = 0 W w)} converges to X(w)co for all w e f l Thus {Xn} is said to
Thus, the sequence — Mk} is stochastically convergent to zero converge to X everywhere.
If X n(w) converges to X(w) only for w'EQf* w e A, then C is called the set of
convergence of X,. If Ce A, then tim X n is a random variable clearly, C is the set of
all w e Cl , at which whatever be £ > 0, \Xn (w)~ A'(w)! < e for all n greater than
n = N0(w) sufficiently large symbolically for = n + m, m > 1
C = [w :X m( w ) ^ X ( w ) ]
= *•n>0 u« nni [ ~ | * ~ w - * w H
Equivalently, replacing “for every £->0 by for every — k = \, 2, ...e
k
Since C is obtained from countable operation on measurable set, C is measure from
C e A.
Now |/ ( .v ) -C | = |C -C | = 10]<£
Hence \ f (x ) — C\ < s V x
2M 215
UNIVERSITY OF IBADAN LIB
ARY
This theorem tells us that the limit of a constant 8 that constant.
This lemma provides us with sufficient evidence/condition for the convergence in
Remark: probability.
The proof of the above follows from Markov’s inequality.
If f a has the limit Las x -> x then f a is said to converge to /.
OR i.e. P[
If C is the limit of f a as x a then f a is said to converge to a constant C written as n —̂ oo
as Ax) ~*C as x-> a Theorem: Let X be a k-dimensional r. vector and g > 0 be a valued (measurable)
Note that the constant Lor C can also be a random variable. function defined on 9?*, so that g(x) is a vector random and let C > 0, then
Convergence in Probability
A sequence of random variables {XK} is said to converge to X in probability, denoted Poof: Assume X \s continuous with pdf Then
by — ~->X, If for every e > 0, as n ->oo
equivalently, if for ¥■ e > 0, as n <- oo -- fg{xltXi.... Xt ) f ( x , ..... x t ) d x „ . . . , dxk + J (gx,......,xt ) / ( * ....., x t ) d x ...... d x k
A
P[)Xm- X < e ] - ¥ l .
Note:
This concept plays an important role in statistics, i.e. consistency of estimations, weak Where A =
laws of large numbers.
( x ) ] - } & ( x i ” — >x k ) / ( * p »* )A » ~ A
Equivalcntrandom variables: Two random variables X and X ' are said to be A
equivalent if X -> X 1 a.s [almost surely] Using the result from Markov inequality
Lemma: X n— and X n——>X' =>X and X ' are equivalent.
A
This lemma shows that a sequence of random variables cannot converge in probability
= CPfeW e A] = CP\g{x) > C]
to two essentially different random variables.
Lemma: X n -»0, i f c\Xn\ ~>0 , ^ f W 2 C ] S £ M
Replacing Xn by (.Xn - X ) we have:
Note:
X n -+X<*> iff X n- X — ^ 0 If A' is of discrete type, the proof is initial analogous.
1\X" - X \ 0 implies X n — 1 Special Case I: (»)
Let X be a random variable and take g [x ) - \X ~ L \ ,/j - J i ( X \ r > 0. Then
216
2I7
UNIVERSITY OF IBADAN LIBRARY
p\x-^\>c]=A Proof:
Cr •
The above is known as Markov’s Inequality
Then X, amd Yt are standardized variables hence
Special Case II:
If r in (*) above is replaced by 2(i.e. r = 2) we have i f f - \ < E ( X , >',) K a \ i - L Note:
A A more familiar term of Canchy-Schwarz inequality is
e -[x r ) < £ ( * 2)£ ( r : )
Remark
12.4 Borel-Cantelli Lemma
Let X be a random variable with mean ^ and variance cr2 = 0. Then the above gives
In the study of sequences of events A,,A}... with Pk = P[Ak ); a significant role is
p \ x - //| £ c ]= 0 for every C > 0
played by Borel-Cantelli Lemma.
This implies that p ( X = /j ) = 1
i.e. (i) If the series converges, then a finite number at events
12.3 Cauchy-Schwarz Inequality Ak occurs with probability 1.
Let X and Y be two random variables with means u, ,/i2 and positive variance (ii) If the events are (completely) independent, the series diverges,
a' and a 2 respectively. Then then an infinite number of event At occur with probability 1.
Or equivalently, Theorem:
Let {An} n = 1,2,... be a sequence of events and P{A„ )denote the probability of the
- o fo f < E[[X - n )(Y- AA)] £ ofo}
event An where 0 < P(An)< 1. Then if
and E [ { X - ^ ) { Y - ^ ) ] = g -g \
JC
(i) If £ p (Aii)< oo with probability one only a finite number of event
iff ( * = * ) = i
An occur.
2 1 8 2 1 9
UNIVERSITY OF IBADAN LIBRARY
i.e. The infinite product of the r.h.s of the above is divergent. Hence, P ( a )= 1 which
x
30
(ii) II the events {An}n = 1,2,...are independent ^ P { A n)= with prooaouuy shows that ^ P(An) = oo
/=i A C ( J An as ( > co
n^r In simple language, the theorem states that a large number of independent random
variables has a distribution that is approximate normal. It provides a simple method
Meaning that P{a ) < A ( J a ,, < for computing approximate probability for sum of independent random variable’s and
explain the fact that many natural populations are normally distributed.
as n oo; P(An) 0
/!=/'
12.6 The Central Limit Theorem
hence, P(a ) = 0 ^ P(A„)< °° Let {Xn,n > 1} is a sequence at random variable Define
fi=l Sn = Xi + X2 + " X ni a(Sn) as the standard deviation of Sn and Zn
(ii) If An are independent and _ Sn-g(Sn)
*(Sn)
£ > ( 4 , ) = c° Then Zn converges in distribution to /V(0,1). This is an example of SLLN.
nor
then A - l - Ax xi f f at most finite number of events An occur. Example
hence, A = \ J f ) A n Suppose Xs above are i.i.d each with the Poisson distribution with parameter X. Show
<•■1 «i-l that the SLLN holds.
In view of the independence of An, we have 12.6.1 Central Limit Theorem for Independent Random Variables
Let Xi,X2, - be a sequence of independent random variable’s having means p, =
\ - f (a ) = p {a ) ^ p(\j ( ) a„ E(Xf) and variance a? = Var (XL). If (a) the Xv are uniformly bounded, that is for
k. r n | «=r
some M',P$\X‘i\. < M) = 1 for aHTaad
= L ^ f m ) = x
r * l \.«n| /
220
UNIVERSITY OF IBADAN LIBRARY
i\
We will show that with probability I,
uu CO
( h ) ^ of = co, then
1=1
Z U C X i- f i ) Let X1,X2l ...Xn be independent random variable’s with E(X[) = 0, Var (X,) = a 2
P< < a • -* 0 (a) as 7i —* co we have for some a > 0 by Kolmogoro’s inequality
ZT=i°?
Kroncker’s Lemma (Proposition)
If a1( a2, ... are real number such that
co n
Z-r < oo, converges, then lim Z)l n-*oo — i —n = 0 1 = 1 n-*oo
12.7 Strong Law of Large Numbers for Independent Random Variables
S B . X % = °
Let Xi,X2, ... be independent random variable’s with E (A",) = 0, Var (Xs) = a? < °o. l=j
By Knonecker’s proposition, we have that
If i < 0, then with probability i.
X1+Xa+-+Xn
0 a s n —» co = 1
t=i
Note: It can be observed that Kolmogorov’s inequality is a generalization of
Chebyshev’s inequality. If X has a mean n and variance a 2, then by letting n = 1 in Pr < Max > a } = 0
Kolmogorov’s inequality we obtain Ijsksn 1 %l l=J
P{\X - n\ > a) < -jj{which is Chebyshev's inequlaity) This implies that
n
Where Xl ,X2, — Xn are independent random variable’s with E{Xf) = 0, Var (Xs) = im ) * l/ n = 0 or equivalently that
of; then Chebyshev’s inequality yields -«oo Z1 j = 1
n 2
f{|Jfi + - + Xn | > a } s 2 ] ^ lim ) --------- =
n-cc Z_j n
1 = 1 ‘ 1=1
Kolmogorov’s inequality gives the same bound for the probability of larger set of and that
variables. The theorem (Kolmogorov’s) is used as a basis for the proof of the strong p f lim Xk = o] =
law of large numbers in the case where the random variable’s are assumed to by V|| •• Li )
independent but not necessarily identically distributed.
Proof: (Of strong law of large numbers for independent random variable’s)
2 2 3
2 2 2
UNIVERSITY OF
■*T8 7# . . v,..:;. . .---- — c IBA .............. —DAN LIBRARY
Definition 1: -■> A„(w) -v A'(w) as n ̂co
Two sequences h{Xn(w)}, {/„(vv)} of random variables are said to be "Trial Proof
Equivalent" a finite number of terms, Given = X„ Pr{X„\ > a„} < co
i.e. for almost all w Q, X n (vv) = Y„ (vv) => PrlEn occur infinitely often} = 0
for all but a finite number at n Pr jlim sup E n j= 0
Lemma P'1 k fw-!l Umw E- 1 ■<>
if ' Z Pr^ - x , M * y , ( w ) } < oo i.e. trials x X >
ml Pr n u ^ = 1 by de - M orgasloa
Then Pr{w: X n (vv) * Yn (vv) inf initely after }= 0
Proof /.P r = 1
Let En = [Xn(w) * ^ ( mO) => Pr [En occurs inf initely after) = 0
Since X P\En} —v oo in converges
=>Prjlim infE„j = l
PrjPirn sup En J = 0 or 1
Pr{A\(w) = y.(vv) V except a finite number of n}
Definition 2 . . . ; =>(jT.}and{Yn} are Tail Equivalent.
A sequence {y„}is said to be a truncation at the sequence {A'Jat [an)where {a,,} is a
sequence of positive real number if Examples
Let En = {vv: Yn(w) -> A„(w) as n ->»}
=*P(E) = 1
We know: \X,\ < an => -a n < X n < an i.e. A - [A*. - Y„ Vn except finitly inany nj
P(/l) = I
Y / / / X Y / / / -\ cut off the {an} at {Aj in order to obtain {T„}
’ 0 »n If we E amd w e A
we E fl A = B = {w: Y„(w) —> A„(w) as n -> co amd X„ = Y„ V except finitrly many n)
Lemma:
Let the sequence {/„} be a truncation of sequence {Xn} at sequence {a„} be finite i.e. r=.weA 'f|A=e>weB
Yn - XPrlA^I > an }< co (Efl A)< B
then thus P(E) = 1. P( A) = 1 P(B) = 1
Y„ (w) —> X n (w) as n —v oo where V. and A arc defined on the same ample space.
225
224
UNIVERSITY OF IBADAN LIBRARY
Hence, they are tail equivalent.
12.8 Bolzano-Cauchy Criterion for Convergence
Lemma: Let C be a fixed real number. If |C| < AT efor some K > 0 and every€> 0, it
follows that C = 0. By Bolzano-Cauchy criterion for convergence
Proof: suppose not. Then C * 0 £ /> (£ „ )< °o
0* I
Since € is chosen arbitrary, put e into let e= > 0, since K is given.
2k given any e> 0,3 an N0(e)
Then |C| J = Such that V-n > N 0.K e which is clearly a contradiction except for |C| = 0
hhK
^ / >(E „l< e and lettingK -» co
12.9 First Borel-Cantelli Lemma
Theorem: Let {£„} be a sequence of events each of which is a subset of ft such that
£ F where F is a a - field of sub-events of ft defined on the probability space Now Z
(ft, F.P), then ^ P ( E En)< co => pjlim supEn J= 0.
P(E)< Z p (£. 0
[iffE .} is a sequence o f events we are often interested in how many of the event occured] Where ecan be taken arbitrarily close to zero,
OR i.e. only finitely many En occurred.
If f c ) is a sequence of events in a a - field F, w here ft, F, P is a probability Is1 BC-Lemma does not require independence of the event En.
space. Then P({Eon})< 30=5 /4im supE„ }= 0 12.10 Second Borel-Cantelli Lemma
Let {£„) be a sequence of independent events on the same probability space
(fl, F, P)then if £ = mli mMt sup En;
Proof:
Let £ = »li-m«, supr En !0->1 ( £ . ) = «°
Then £ - f ) | j £ „ ; clearly E c (J E„ V-m e =>£(£„ occur infmtely often) = 1N
i.e. £jiim supEn j= 1 .
226 227
UNIVERSITY OF IBADAN LIBRARY
Proof:
Corollary to 2ml B.C. Lemma
Recall that hm sup En = p | \ J e „ s If A7 .fareindependentandX„ -> 0 (a.s)
, m=l m m
Then 2 >|jr,,|>c]<
,im *UP (En)] i = P{hm inf Enc}= pjlim (fl )} .rr»l ; . • * . '
Whatever be C> 0, finite
For any N > 0 and every K > N
Proof:
If Xn's are independent random variables /!„ = [jx„|]> C are independent. Since
Since E, s are independents E f 's are independent too. X„ 0 a.s. iff.
/ > C]< oo as n -» 0 and for any C > 0am-N m N
Since J~[(l - P{En)) < J"J e ,,[K'1 = e "" by exponential property We have / ’(lim SUP A.n J= 0
mN mN
Since PE(A„)< c o .
as K oo; £ P(En) -» oo i.e. £ P(En) = co Note:
" » N n=N
The converse of Borel-Cantelli lemma is not true if An's are not independent.
- I P ( E . )
=> lim e "" -> 0
Af-w:
=> 1 -P {E ) = 0 ^ P(e )= 1. 12.11 The Zcro-Onc-Law
Theorem: Let A„ A-,,..., be events and let A be the smallest c r - field containing each of these
< co events. Suppose E is an event in A with the property that, for any integer j\, j 2, jk pr{lim sup En }= according as J P ( E n) =
ICO that events.
Whenever Et,E 2>..., En,... are independent E and Aj, f) Aj2, f l ... fl Ajfc are independent.
r
| Then PIE) is either 0 or I.I e- (|) J^2iib]E (A n) /'(lim sup A3) = 1
Bui ±P(.4„ J~rc=> P^lim inf An )■--- 0
228
229
UNIVERSITY OF IBADAN LIBRARY
Thus lim sup An * lim inf An
n-Kt, n-*xj
Hence {An} dos not converge. By independence of E and Aj, f | Aj, D ... fl Ajk
= jA jl:'Aj,...AjkdPP(E )
n
Exercise 2: Let X have the uniform distribution (X ~ p (0 , l)) consider the sequence But (fl, A, P)is a complete probability space.
o f events {An} .
.■ .p (/fn £ )= .p M ^ p [e )
Where An = jw: X(w) < — | . Are {An}independent. For all A e A, in particular A = E.
:.P{E)={P(E)}2
P ro o f : f ( t ) = l Then P(e ) = 0 or 10 < x < 1
Completeness
A measure space (H, 0 , P) is said to be complete if Acontains all subsets of sets of
measure zero.
ThCn 5 P Â" ̂ ~ 5 n ~ +°° Harmonic Series Diverges Note
(i) A non-empty event with zero probability is negligible
But En ^ A n+lz>An+l=>... (ii) Every subset of a negligible event have-zero probability;
/ ’Jim sup A )=
I M 7 ' " ) = ^ W =0
,,,BI n**ni / Lemma
(1) Given a probability space (H.A, P)and a sequence {£„. u - 1.2,...} of event
T in )- where EnCH and E T V-n
Clearly the above violates the 2nd B-C lemma as the sequence {/*„} of events is
overlapping and the therefore not independent. Prove
(1) lim inf En C lim sup En
n- v n n-f«>
Proof: (ii) lim inf P(En)< lim inf P(En)n -* r n-Kc
on Q -A ( (2) Let (Q, F) be a measure space, on which a sequence of probability measure isLet
on AJ » 1defined. The set function P(e ) dA Pn(£)
Then J IAj{, fAl2 .....//!,* dP = P(Aj, f | Aj, f l . - f l Ajk f |£ )
(i) Show that 0 < /} ,< 1 .
= p(Aj1nA j; n . . . .n A j j / ’(£) (ii. i i ) is countably additive and is therefore a measure
(iii) Prove that Fj(f2) = 1.
230
231
UNIVERSITY OF IBADAN LIBRARY
Solution Now let Au {w: \X,(w)| ̂ e zTn}, then An l 1 .Since Pn (e )^ 1 and 0 lim (En) = 0 and iim P(An) = 0
2 n 2 " n-*w I N - X
(ii) Show l im j ; P„ (£ ) = P. (£) and d/IMmC. Jf X]dP = 0
n=l *• n»l ^ Am
This verifies the 4lh L.T. so by Lindeberg’s theorem
(Mi) P ( f i ) = Z j r P . ( n ) = i ; ^ ( i )
n = l ^ ^ X , + X 2 + - + X n_ _̂ in distribution
tTn
+ . . .
~ 2 + 4
Lindeberg’s Theorem (The Conditions of Lindcberg Theorem)
5W =
T ~ 2 Let 1 X itX 2, . :X k t
i ->^n >•••^2*3
Be a rectangular array of random variable satisfying the following condition.
12.12 Limit Theorems for Sums of Independent R.V’s
1. V-n > 1 X n ,X n ,...Xnt_ are independent
Lirideberg-Levy Theorem
Lcl X r .\\... be a sequence of ij.d.r.v each with mean 0 and variance 2. e (* j = 0;
r J (o < r < co) Then + * 7 + — + X* 71/(0, l) 3- B] = r ; + r ; : + ... + r ; 4. wrlA 5 ; > 0
zTn
Proof:
*«0
Consider an array Xi K e > 0 .
Let S„ = X„ + X„ +...+ X . and N a random variable with standard normal
* „ * i . * ,
distribution
Condition 1,2 and 3 of Lindeberg’s Theorem are satisfied. Then ——— —» -j=L= T e ^4//
5, 42n ■“
We only need to verify condition 4.
The above statement is basic to the central limit theorem
Let e> 0, Bl = n r :, then
However, i f E(XiiL ) = 0
n
—n t Z J -m p 2, \x.dpn r i* dt
Since X s are i.i.d.
and
232 233
UNIVERSITY OF IBADAN LIBRARY
Lyapunov’s Theorem 1
LetXi be a sequence of independent random variables. If a positive number 8 can be
found such that as n —» co; / > 1.
eoW-.i
u„ *=i
k = I V2 jr By hypothesis (ii) above
1 i? .+rf = 0
Proof: • /*.•; • • : . - . -, ' «<-” 6d V*
The random variables, define above satisfies 1, 2, and 3 of Lindeberg’s theorem. It
. jSl q 5* i i _ » #(o, l)
also satisfies the following: "5 ..
(i) for some fixed & > 0 > E \X » k \2+S < 00
(>o (im ■° n i Et.i £ i^-*r+' = ° then
We now need to show that condition 4 of Lindeberg’s theorem is satisfied.
Let Var(X,)=b for / = 1, 2 then Bn =bTn
Setting E(X,) = ar, condition 4 becomes
/ \ x a*hVn
Vr,b2\ |,r-a|!>*^fl, - a^ dFM = o+eJ/r+fp -»
which approaches zero since the Var(Xt)< oo anc/ 6 = 0
Now wc need to show that condition (2) implies condition (4)
This follows from the inequality
234 235
UNIVERSITY OF IBADAN LIBRARY
Where /?(ll|is the initial distribution
C H A P T E R 13 //, is the drift vector
IN T R O D U C T IO N T O B RO W N IA N M O T IO N ^ i s the diffusion matrix
13.1 Brownian Motion (Weiner Process)
Brownian motion describes the macroscopic picture of a particle emerging in random 13.2 Brownian Process
system defined be a host of microscopic random effects in d-dimensional space, Peter If the drift vector is zero and the diffusion matrix is the identity, then is
& Yuval (2008). At any step on the microscopic level, the particle receives a termed/referred to as the Standard Brownian Motion. Hence, the macroscopic picture
displacement caused by other particles hitting if or by an external forces so that it’s emerging from a random walk can be fully described by a Standard Brownian
•• ** H Motion.
posterior at time-zero is So, its posterior at time n is given by S„ = S0 + ^ x , where
■̂1
the displacements X t,X 3,... are assumed to be independent, identically distributed 13.3 Multinomial Distribution and Gaussian Process
The most important joint distribution is the multivariate normal (or the multinomial)
random variables with value in TRd. The process {S’,,:/*£ 0}is a random walk, the distribution. It arises in many applications and has some properties that makes its
displacements represent the microscopic inputs. Thus Brownian motion is a kind of manipulation very simple.
stochastic process. If A is any (/jx»)symmctric matrix, consider the quadratic form
Any continuous time stochastic process {#(/):/£0} describing the
= A X
macroscopic feature of a random walk should have the following properties: n n
(i) For all time 0 ,< /, <... * x 0} has almost surely continuous paths.
Let V = A'1, then V is also positive-definite and symmetric.
(iv) It follows from the CI.T that these feature implies the existence of and
a matrix le '.R ^ su ch that for every t > Q and h > 0 , the increment is Definition 1:
A collection (X„ Kjwhich has the joint density.
multivariate normally distributed with mean h/.i and covariance matrix /?XXr
Any process {.V,} with the above feature seem be represented by (2x)% m(def.V)T
& i = A .i + M + £ / W J'” t>0 is said to have the multinomial distribution A'fO.K)
237
236
UNIVERSITY OF IBADAN LIBRARY
13.4 Properties of a Brownian motion (B. M)
Definition 2: If //,, are finite real numbers then X = The following are the properties of a Brownian motion.
C^i+M ^2 + /^2» -> ^ n + A. )joint p.d.f 1. The Brownian motion is a Gaussian process with autocovariance function.
V { s , t ) = E ( x , . X , )
(27r)/^-(det. P ) ^ exp ~ (±J Y~'(x~m)} a°d l *s sa>d to have the multinomial
= min (5,/)
distribution n (̂ j, v ) 2. The autocovariance function
P(.v,r) = min (i-,/)
Definition 3: Letr be any set (usually a subset of the real axis). For every t fe r re t i.c. symmetric for r = (0,co)
A ^ b e a random variable defined on a probability space (Q, A,P). Then the family 3. Let A ^be a B.M. process and define A'(.s,f)= Xlt) - X {s), the increment
,w): re r} at random variables is called a Stochastic process.
process on the interval' (s,t\ Then A(j ,/) ~ A^(0,/-i)
Definition 4: Let V(s,t) = e \ X ^ - ' / u, \ x ^ - J} be the autocovariance function at 4. Given the Brownian motion process
M M for all relevant values of t and s and pt = E ] x ^ \ p s = ZsjAQ,)] £ ( / ) = 4 K /+A M d f )
Definition 5: A stochastic process Af(r,w) with the property that all its. finite = 3/r
5. The Brownian motion process is continuous everywhere but is nowhere
dimensional distribution are multinomial and E(X,}=0,
differentiable.
E ( X „ X , ) = V { s , t )
Where K(v ) is a positive-definite function on r , is called a .Gaussian Process with Definition 7: Let T be any set (usually infinite) and possibly uncountable) and let
autocovariancc function P{-, •): TXT -> 9? be a function with the two properties.
Remark: (ii) for any finite subset }er and any real numbers Z2,...,Zn not
• Two Gaussian processes with the same autocovariance function have all zero
the same finite-dimensional distribution
• The most important example of a Gaussian process is the Weiner (or I.-1 2/-iM ' , < > , .* ,> 0
Brownian motion) process. then P(v ) is called a positive-definite function on T
Lemma:
Definition 6: A Gaussian process is said to be a wiener (Brownian) process
P(/,./,) - min (r,,r,) is a positive definite function on r
i f (/') r = (0, ao)
(H) -T1(1) =- 0 and
(iii) = m in(j,/)
238 1I
UNIVERSITY OF IBADAN LIBRARY
Proof:
(/) Clearly V(/,, /,) = V (/,, /,)
(/'/') I f 0 < I, < /, then z * . + ('z -O + (/> -':)[ Z * , z * > +(/«/■I
Z<-i /E1 ^ i > t e *, = Z<-i Z,-»i min 0 / . t e • Z f c ' . - i J /E-I* /
„ i ‘-i ,-i Where /„ - ()./, >
(1 <./.) Clearly the last expression in the “curly bracket] is a positive number.
•S/rar min(/i./()-/. /dr / = j and J(.v,/) = min (.y./) is positive definite.
Since b> symmetry, we may interchange / and7 to cover cases in which f
Theorem:
- I ' . T + 22 . L * , Let V(/)be a Wiener process and let X(.s.t)- ^ lit * A',.,denote the increment of the
••I /"ni
process on an interval (.v,/), then
Expanding the square bracket gives
(/) A (.v,/) - iV(0,/- .v)
= * .-+ 22 , 2 ;* , (//) // (.v,, /,) anil (a,, /.) arc disjouit intervals.then X ( s t. tx)andX(s,, t2)are
stochastically independent.
= *,’ + 22,(2 ,.,)+ 2* .!* ,.,)+ ... + 22,(2 . J + 2(2 ,.2 .)
Proof
(i) A',#,is Gaussian, therefore the joint distribution of A',f, and .Y(#, is
■ f t * . /"I*!
multinomial and so .Yj(J - Aj,, - A',, N(o.r’)
I * , I - I I * , where r - I a r ( X ( s j ))(-/♦I
-4 * 1 .1 - - O '
Writing the expression in full, we have
= 4 ^ i , - 2avv, + -v,;,]
- / , k IZ . (3, + * + . . . + 2 . f} ~ r ( t . / ) - 2 l ( s . t ) + V (.v..v)
1 C |(^; + - H )' - (£; *- £ , + ... + } - / - 2.V + .v
• - • M . , ' * J -?„■’ } - / • .V
Hut / w ,. by hypothesis, rewriting mi) 'A c ;*- , K ^ ( / ; I + ^ ( a , 1 [0 , t = 0
“ V[U. ^ (', ^ 1 •?; )+ ^ * *2)
l /
— o Show that is a Brownian process.
Since Cfn>{s,l) = 0, there exist stochastic independence.
Exercise 1: For any real set of number C,,C,.... C,(and real values random variable Hint for Solution
•j.V, J". show that ]T ]T (?,(?,£■(A', - / / J iA ' , -p ) is positive scmi-dcllnite for Calculate !'(•. ) o/
,.i /-i Note that for a B.M. process // = 0
//)(); //)=£[();,,)(>;„))
Hint for Solution: Let >'■ = A', - / / , then £ ()') = 0 and V u r \^ C iY, ] = Zij ^ C ,) '
= .v/»n. n{I i .-i •,
W s j
i-i i- i #-ri -- min (.v./)/
Exercise 4:
M 1 cl A'Ufbc a Brownian motion process and let rr,,, - A„, — / A*(ll;
0 < / < I
Example 2: Calculate the autocovariancc function of the Gauss-Markov Process Find the or /',(.•>./)
Hint for Solution
Hint for Solution: Assuming /T()7) = 0; Nolo that /:(-/*„,)- 0 nm/
! > ./) • - A-[(c"A>:- )(e - Ye-'')]
, )(«•*)]
Cf e:> for l > s
i.' <‘*'1 t* for l < s
* 1 for 1 = s
- f, i/ l-s.f
Example 3:
242 24?
UNIVERSITY OF IBADAN LIBRARY
^ E [ ^ r s X j X {l)- t ( x J ]
= £[a'\t)X U) - sX{t)X {,) - X[t>f X [x) + tsX
= P M - . sI'(1,0 - /P ( 1,j )+&K(U)
= Min( s j ) - S min ( i t ) - t min(l,.?) + / s min (l,$)
= .v - st - is + st; for s < l
- t - t s - s t + st; for t < s
= M in ( s j ) - s t ; V-sJ
PART THREE
244 245
UNIVERSITY OF IBADAN LIBRARY
C H A P T E R 14
Definition
IN T R O D U C T IO N T O S T O C H A S T IC P R O C E S S E S A stochastic process is any process that evolves with time. A few examples are data
on weather, stock market indices, air-pollution data, demographic data, and political
14.1 Basic Concepts tracking polls. These also have in common that successive observations are typically
Researchers in science, engineering, computing, business studies and economics quite not independent, such collection of observations is called a stochastic process.
often need to model real-world situations using stochastic models in order to Therefore, a stochastic process is a collection of random variables that take values in a
understand, analyze, and make inferences about real-world random phenomena. set S, the state space. The collection is indexed by another setT, the index set.
Finding a model usually begins with fitting some existing simple stochastic process to
the observed data to see if this process is an adequate approximation to the real-world The two most common index sets are the natural numbers T = {0,1, 2,...}, and the
situation. nonnegative real numbers which usually represent discrete time and continuous time,
*> '• "• '■ i‘ respectively. The first index set thus gives a sequence of random variables
Stochastic models are used in several fields of research. Some models used in the (X0,XVX2, — )and the second, a collection of random variables {AT,,,, t > 0 j, one
engineering sciences are models of traffic flow, queuing models, and reliability
random variable for each time t. In general, the index set does not have to describe
models, spatial and spatial-temporal models. In the computer sciences, the queuing
theory issued in performance models to compare the performance of different time but is also commonly used to describe spatial location.
computer systems. The state space can be finite countable infinite, or uncountable, depending on the
application.
Learning stochastic processes requires a good knowledge o f the probability theory,
advanced calculus, matrix algebra and a general level o f mathematical maturity. 14.1.1 Applications of Stochastic Processes
Nowadays, however, less probability theory, calculus, matrix algebra and differential The followings are some areas of Stochastic Processes:
equations arc taught in the undergraduate courses. This makes it a little bit difficult to (i) Marketing: To study customers or consumer buying behaviour and forecast.
teach stochastic processes to undergraduate students. (ii) Finance: To study the customer’s account recordable behaviour and forecast.
(iii) Personnel: To study and determine the manpower requirement of an
The mathematical techniques and the numerical computation used in stochastic organization.
models are not very simple. In an introductory course, the hope is to teach students a (iv) Production: To study and evaluate alternative maintenance policies,
small number of stochastic models effectively to enable them to start thinking about inventory, and so on, in industries.
the applications of stochastic processes in their area of research. These small numbers (v) Transport: To effectively control flow and congestion in the transport
of stochastic models are the core topics to be taught in an introductory course on industry.
stochastic processes directed to researchers in the physical sciences, engineering,
operational research and computing science. These researchers have a stronger 14.2 Discrete-Time Markov Chains
background in mathematics and probability than researchers in the biological You arc playing a lotto, in each round betting N13 on odd. You start with N30 and
sciences. after each round record your new fortune. Suppose that the first five rounds give the
247
UNIVERSITY OF IBADAN LIBRARY
sequence loss. loss, win, win, win, which gives the sequence of fortunes, 9, 8, 9. 10, For a transformation matrix, a 2-level change of state will produce 2 by a matrix, a 3- lcvel change produces 3 by 3 matrix and so on
1 1, and that you wish to find the distribution of your fortune after the next round,
18
given this information. Your fortune will be 12 if you win which has probability — 14.3 Classification of General Stochastic Processes
The main elements of distinguishing stochastic process are in the nature of the state
and 10 if you lose, with probability —20 . One thing we realize is that this depends
3 8 space, the index parameter T, and the dependence relations among the random variables XL.
only on the fact that the current fortune is N il and not the values prior to that. In
general, if your fortunes in the first of rounds are the random variables^, ...,Xn. the
conditional distribution of ^n+l given Xv ...,Xn depends only on Xn. This is a
14.3.1 State Space 5
fundamental property and we state the following general condition.
This is the space in which the possible values of each 0 1/7) 1 ^ = 1 , for/;eS
14.4 Classical Type of Stochastic Processes
for all ( e S
We now describe (first brielly) then in details some of the classical types of stochastic
processes characterized by different dependence relationships among At . Unless
14.2.1 The Transition Matrix
In changing from one stale to another in any Markov system, a measure of probability random .staled, we lake T - [(), -x-] and assume the random variables A", are real
valued
is always attached. Ii is the collection of all such probabilistic measures which are
arrange din rows and columns that ids called the transition matrix.
UNIVERSITY OF IBADAN LIBRARY
Then for any t and s we have
14.4.1 Process with Stationary Independent Increment / ( / + *) = £ [X ,„ -X „]
If the random variables X l2>- X,,. X ,j-X l2,...,Xtn — X ln_{ are independent for all = £ [X,.s-A'.v + X5. -X „ ]
= £ [X ,.S- X ,] + E[XJ.-X „]
choices of £1( t2, .... ^satisfying £, < /, < ...< /„ then we say that Xt is a process with
independent increments. = £ [ X , - X „ ] + £ [ X , - X „ ]
If the index set contains a smallest index t0, it is also assumed
X c - X li,...,Xtn- X ln_l are independent. If the index set is divided,
Using the property of stationary increments
what is 7 = (0 ,1 ,...), then a process with independent were reduces to a sequence of
= / M - / W
independent random variables Z0 = Xq,Z{ = X, - X ^ . i = 1,2,3, ...in the sense that
The only solution to the functional equation / ( / + s = / ( / ) + f ( s ) = /( /) / .
knowing the individual probabilities/distributions of ZQ,ZV ... enable us to determine
the joint distributions o f any finite set of Xt, in fact that of differentiating with respect to t and independently with respect to s we have
X, = Z„ + Z, + ... + Z,, 1 =0,1,2,... f ( r + s) = f ' ( r ) = f \ s ) .
Therefore for 5 = l, we find f ( t) = constant = f(i) = c. Integrating this elementary
Remarks'Definition differential equation yields f [ t ) = cl + d.
1. 1! the distribution of the increments X(t, + h ) -X ( t t) depends only on the But / ( 0 ) = 2, / ( 0 ) implies / ( 0 ) = 0 and therefore d = 0.
length h of the interval and not on the time t, the process is said to have
TTh. eref.o re expressi- on / ( 'x)r = / ( ! > «s/ f n %
sia/iunun • increment.
2. For a process with stationary increments, the distribution of X(/2 + h ) - X ( t 2), =>E[X,] = M n +M, 1 as requires.
no matter what the values of h, t2 and h.
3. We now state a theorem;
14.5 Markov Processes
If a process {Xr t eT}, whereT = [0, oo] or T = (0,1,2,...) has stationary
A Markov process is a process with the property that, given the value ofXt , the values
independent increments and has a finite mean, then it true that: of Xs, S > t, do not depend on the value if X u < t; that is, the probability of any
£ (X ,)= M n + M, where M0 = £ ( x J and M ,= £ (X ,) -M 0 particular future behaviour of the process, when the present state is known exactly, is
07 = 07 + 07 where not altered by additional knowledge concerning the past behaviour, (provided our
£7,; - £ [(X„ - M„)] and |.v: s;«.A[ Pr {x, s and 2. Explain the concept of a simple Markov Chain.
is basic to the study of the structure of Markov process. We may express the condition
( 1) as follows: 3. Define the following:
p r{« < x , ^ tyx ,i x,>Ks = x^ - x ,„ = "„) = H x„> tn, t A) where ) £ |a < £ < b} (a) Slate Space (5)
(b) Index Set (7 )
14.5.1 Martingales (c) Renewal Process
Let (.V,) be a real-valued stochastic process with discrete or count parameter set. We
say that (A',) is a Martingale if. for all t, and if for any
< /, e (X1i1.,|X 1I r/,.... Xln =o„) = c for all values of ai, a2, ... a„.
14.5.2 Renewal Process
A renewal process is a sequence Tk of independent and identically distributed (i . i .d )
positive random variables, repressing the lifetimes of some “units”. The first unit is
placed at time zero; it falls at lime /', and is immediately replaced a new unit which
then fails at time 7', + 7'2and so on. the motivating the name “renewal process”. The
time of the nth renewal is S„ - 7] + 7', t-... + Tn.
A renewal counting process N, counts the number of renewals in the interval [o.tj.
formally .V, = n for Sn < ( < Sn,t, n = 0 ,1.2 ....
Remarks: I lie Poisson process with parameter A is a renewal counting process for
which the unit lifetimes have exponential distribution with common parameter A
Other examples such as Poisson process, birth and death processes and Branching
Process v\ ill he considered in small details.
UNIVERSITY OF IBADAN LIBRARY
C H A P T E R 15 P {x < i) = 1 - q ,
G E N E R A T IN G FU N C TIO N S A N D M A R K O V C H A IN S So that the probability generating function follows
p(x) = Zi=oPi * ' = I: (X1)
15.1 Introduction
Also for the joint probability, we have the generating function as
Generating function is of central importance in the handling of stochastic processes
involving integral-valued random variables not only in theoretical analysis that also in Q CO = £«=0 Qi
practical appreciations. Stochastic process involves all process dealing with
We can see that (?(*) is not the same as P(x)
individuals’ populations, which may be biological organisms, radioactive atoms, or
telephone calls.
Q(x) do not in general constitute probability distribution despite the fact the
15.2 Basic Definitions and Tail Probabilities coefficients are probabilities.
Suppose we have a sequence of real numbers a 0, a a...... Involving the doming Note that
variable x, we may define a formula sothatP(i) = 1,
A (x ) = Go*0 + a ^x1 + a2x 2 + ••• = £?= Qaix i and /P (x ) /< ^T /p .xV
If the series converges in some real inference - x 0 < x < x0, then the function A (x) is
known as the generating functions of the sequence { a j. We may also see this as a < ^ Pj. if / x / < 1
transformation that carries the sequence unit the function A(x). If the sequence {a,} is < 1
bounded, then a comparison with the geometric series shows that A(x) converge at This means that P(x) is absolutely convergent at least for /x /< 1. But for Q(x), all
least for f x f x j . coefficients are less than unity, this making Q(x) to converge absolutely at least in the
II the following restriction is introduced open interval / x /< 1 .
n
Converting P(x) andQ(x), we have
t'=0
( l -x )Q O O = l - P O )
Then the corresponding function A(x) is viewed as a probability-generating function.
Specifically, consider the probability distribution given by which is easily seen when the coefficient of both sides are compared,
H x = i) = Pi for the mean and variance of p,-. we have
Where X is an integral valued random variable assuming the values 0,1,2 .... U = /•(*) = £ ip, =p<( 1)
Consequently, we define the tail probabilities as i = 0
P{x > i} = q,
= q‘ =(D
Bui the usual distribution function is i = 0
then E\x (x - 1)] = £ t(t - l)p< = p " (l) = 2 Q \l)
254 255
UNIVERSITY OF IBADAN LIBRARY
So that (T2 = var(x ) = p"( 1) + p '( l) - (F 1 ( l ))2
= 2g '( i) + q ( i ) - ((D)2 , . V N u*(it)r!
In the same vein rth factorial moment n[r) about the origin to be 0 (0 = l + ) —*->r=i r!
function and continues
fcWx-DO- 2) .... (x - r + 1)] = £ (i - lXi - 2).... (i - r + l)Pi the characteristics function exist always both for discrete
function.
= p « ( 1) =
0 ,( 0 = £ > * / ■ «
t=i
From these result, several other generating function could be obtain such as the
and
moment generating function, characteristics function, cumulative generating function. 30
x = j e ltx f ( x )d x
15.3 Moment-Generating Function — 0 3
This is define as where the Fourier transform o f / (x ) is
A1x(t) = E(eCx) 30
for X discrete witth probability p,-, we have / m = T j W o d w
- 0 0
A range simpler generating function is that of the cumulants. When the natural
Mx(t) = 'Yj e tipi = P (e f) logarithm of either the mgfo r the c f is generated, it results into the cumulant-
generating function, which is simpler to handle than the former two.
for X continues with frequency function f (a>u ) , we have This is given by Kx(t) = logMx(t)
Mx(t) = J f{u )d u
— 00
obtaining the Taylor series expansion of My(t) r!
we have whore /fr is the rth cumulant.
M(t) = 1 + Zr=i V } tv
In handling discrete variables, the functional moment generating-function is also
r! useful, which is defined as
where is the rth moment assume the original.
Because of the limitation of the moment generation function ( in that it does not Q(a) = P ( l + y ) = e[Cl + y)i]
always exist) the characteristics function become appropriate which is define by = 1 ! Ir=lUr!(r)yr 0 (t) = E{eitx)
flic Taylor expansion is similar where uir) is the rth factorial moment about the origin.
256
257
UNIVERSITY OF IBADAN LIBRARY
15.4 Convolutions Jusl as the case of two sequences, several sequences can also be combining together.
Let there be two non-negative independent integral-valued random variables X, Ywith The generating function of the convolution is simply the product of the individual
generating functions. That is. if we have the sequence {a;) * {£, ) * {c,} * {d,) * .... the
p d f
P(x = 0 = a, generating function becomes /l(x) B(x) C(x) D (x )....
and Given the sum of several independent random variables,
P(y = /) = bj the probability of the joint event (x = y, - j ) is given as aibj.
Syi = Xj + X i + x ? + ••• + X n
Where Xk have a common probability distribution given by p,-, with pgfP(x), then the
Let there be a new random variable S = x 4- y the event (s = k) is made up of the
pg/ol'5,, is {(P(x)}71. Further, the distribution of 5„ is given by a sequence of
mutually exclusive events (X = 0, Y = k), ( X = 1 ,Y = k - 1 ) ,. . , (X = k,Y = 0)
probabilities which is the n-fold c o(pn.v)o •lu.t.i..o..n* {opf .){ p=*} with r if its written as {pi) * ipi)
Given the distribution of 5 as
Pis = k ) = ck 15.5 Compound Distributions
Suppose the number of random variables contributing to the sum is itself a random
then it can be shown that
Ck = a0bk + a-i bk. 1 + — + arb0 variable. Thai is
When two sequence of numbers which may not be probabilities are compounded, then SN = + x2 + — + *n
it is called a convolution which ca{nC kb}e =re p{reks}e nted generally as wherea * [bk] P{xk = i} = f i '
Given the following general functions p{N = n} = g lx
> » (* ) -2 5 o « i* <’|
P{Sn = /) = /i,.
and the corresponding p d f be given as
C(x) = l i .o Q x 'J F W - £ f i * ‘ ^
we can then write
C(x) = A (x)B(x) Q ( * ) = l 9 n * "
this is because, multiplying the two series A{x) and 5(x), and given the coefficients n (x ) = Z /ijX '.
ol'x* as ck. Simple probability consideration show that we can write the probability distribution
When considering probability distribution functions, the probability-function of the ol'S„ as
sum.5, of two independent non-negative integrated-valued random variable X and Kis ^ = p{s„ = /}
simply the product of the letters probability-generating functions. = £ p {/ V = r,}P(Sn = l/N = n)
259
258
UNIVERSITY OF IBADAN LIBRARY
Set L = 0, (i, = i. and 1 = 0
Tor llxcd n. the distribution of Sn is the n-fold convolution of {F,} with itself, that is 15.6.2 Transition Diagram
(/•;}." Thus A transition diagram is a graphical representation o f the process with arrows from
E<-uF{5(I = l /N = njx* = {F(x)}n each stale to indicate the possible direction of movement together with the
Thus the probability generating function//(x) can be expressed as corresponding transition probabilities against the arrow s.
1 = 0 Kxample 15.1
*' ^ gnp{Sn = l /N = n) Consider a process w ith three possible slates av a 2, cmcla2. I .el p,,-: i = 1, 2,3, j =-
n= 0 1 , 2 , 3 , denote the transition from one state to the other.
S n ^ p t f n = l /N = n}*' The corresponding transition diagram is as follows:
n=0 i=0
I h i = ' A >
= ny=0 s „ { f ^ ) } n
= G ( /M )
Thus gives a functionally simple form for thepg/'of the compound distribution {A;} of
the sum SN
15.6 Markov Chain The diagram above represents a square matrix
It would be o f interest to define the joint probability o f the entire experiment. This P = (p „) i = 1 .2 .--------n, j = 1.2.----------
will be a very complicated or intricate problem.Early in the 20lh century, a Russian
Mathematician A.A Markov, provide a simplification of the problem by making the 15.6.3 Transition M atrix
assumption that the outcome of a trial XL depends on the outcome of the immediate To even transition diagram, there exist a transition matrix and vice versa. F or the
proceeding trial Xt_, (and on if only) and effects Xc+1 (next trial) only. The resulting example 16.1. the transition matrix is as given below :
process is known as Markov Chain.
15.6.1 Transition Problem Pn Pl2 P13'
P21 P22 P-a
If a,- denote the state of the process X, and a,, i not equal toj denotes the state of the p = .P31 P32 P n
process X, f ,. then there is a problem of going from a, to a, denoted by pi;- define as,
Pn ~ W o , = ci, / X ( = a,)
2 0 1
260
UNIVERSI Y OF IBADAN LIBRARY
This is a one-step transition matrix for every given i.{p,y} indicate the branch problem
in a tree diagram. In general. Pou PU, Poi
rPu P12 ...... - Pin Pn, P>. Pn.
P21 P22 .....
IPnl Pn 2 .....
Pn
and.
□ « / = i Example 15.2 (Forecasting the Weather)
y=i Suppose that the chance of rain tomorrow depends on the previous weather conditions
For any given t. p,y is the probability of transition to a. given that the process was in only through whether or not it is raining today and not on past weather conditions.
slate a,-. Suppose also that if it rains today, then it will rain tomorrow with probability o ; and
if it docs not rain today, then it will rain tomorrow with probability /?.
In this section, we consider a stochastic process {Xn n= 0,1,2,...} that takes in a II we say that the process is in state 0 when it rains and state 1 when it does not rain,
finite or countable number of possible values unless otherwise mentioned, his set of then the preceding is a two state Markov chain whose transition probabilities are
possible values of the process will be tested by the set of non-negative integers (0, I, given by
2. If X„ = 1. the process is said to be in state / at time n. We suppose that a l-c r „ ( a 1 - o '
whenever the process is in state /, there is a fixed probability piy-that it will set be in fi l - /? j { P | - / ? J
state j. That is we suppose that
=- ]\XU -■ ..... X, = l,,X„ = /„ }= P„ or all statesi„, il t ..., in- j .
i./and Vn > 0. Such a stochastic process is known as a Markov chain. The value p,y Example 15.3
represents the probability that process will, when in state i, next make a transition into Suppose that company XYZ has three departments a}, a2 and a3. The employees lean
stale j. Since probabilities arc non-negative and since the process must make a to be transferred to another department at the end of the year as follows:
i) A man who is in a: . must be transferred only to a2
transition into some state, we have that Ptj > 0, i, j > 0 ; P0 = 1, i = 0 ,1,...
h-o ii) A man who is in a2 cannot be transferred to a lt but can be transferred to a2 or a3
with equal probability.
P denote the matrix of one-step transition probabilities p,y. so that iii) A man who is in a 3 cannot be transferred to a2 but can be transferred either to a3
with probability “/g or a I with probability Draw a one state transition diagram,
and matrix.
UNIVERSITY OF IBADAN LI " ^BRARY
First problem: Suppose the process state in other 1, what is the probability that after
n-steps it will be in state j? Consider a process with only three states al, a2 and a3.
What is the probability that after two steps the process will be in state j . f o r j = 1,2,3
given that the initial state of the process is i . fo r i = 1,2,3. P{X2 = a J X , = fl,J = PU .PU = PPu m
By assuring that i = 1, we obtain a probability tree for the process as follows: P{X2 = a2\X0 = a ,) = P12.P2, = P2, m P»m
P[X2 = a 3|X, = a,) = P.j .Pj, = P3iw P»m
PllPll + P12P21 + P l3P3lP llP l2 + Pl2P22 + Pl3P32PllPl3 + P12P23 + P13P33
P2lPl l +■ P22P21 + P23P3lP2lPl2 + P22P22 + P23P32P2lPl3 + P22P23 + P23P33
P3lPll + P32P21 + P33P3lP3lP l2 + P32P22 + P33P32P3lPl3 + P32P23 + P33P33
Assume tliai i = 2. then
P{-f:; — “ l / — a2) ~ P21 ■ Pi 2 — P2I
. , . , (2)
/ *0 - a 2) ~ P22 P22 ~ P22
/■'{a'x = a< / = a2} = P23 P32 = P^J P23
265
264
UNIVERSITY OF IBADAN LIBRARY
P r/3) = V 4
Pn(3) = V 6
Assume that i = 3, then
31 In the same veinP { x 2 = a \ / x 0 = a 3 ) = p 3 i . P i 3 = P
P { X 2 = a 2 / X 0 = a 3 ) = P32-P23 = P 32
P { x 2 = a 2 / x 0 = a 3} = P 3 3 . P 33 = V r s
It could be seen that p (n ̂ = p"
Example 15.4
Use a probability tree to fmdp(3) in example 16.3
P{x3 = a 2 / x0 = a2} = X/ 2 l f 2 . V 2 = V s
P{*3 = as / xn = a2} = V 2 • V 2 • V2 = V s
P{^3 = a l / xo = a 2) = V 2 ' V 2 • V 3 = V l2
= aj) = 1. V 2. V = V P{x3 = a 3 / x 0 = a2} = V 2 • V 2 ■ 2/3 = * 4 {x3 a 2 / x a = 2 4 P{x3 = a 2 / x 0 = a2} = 1/ 2 . 1/ 3 l = V 6
^ { * 3 = aa ,j //X o == aa ,j)} == 11 -. V1 / 2 • aV/ 2 == VV4 P{*3 = az / *0 = ai) = V 2 • V2 • V2 = V 6 PP[x3 == x„ a,} = 2 3 = V 6 P{^3 = / *0 = Qz) = V 2 • 7 3 • V 3 = V 9{*3 a2/ x 0 = 1-a/ 2 2/3 3 p{x3 = a3/x0 = a2) = V2 • 2/ 3 • 2/3 = 2/g
Pl3<3)=
267
266
UNIVERSITY OF IBADAN LIBRARY
^ ] l !' — V 12 + V69 -7 /72/346 Therefore.- % + v =
r-(3)
^23" ' = V i 3 + V6 + % = 3?/ 72 Pu Pu pS [V s v 4 7/ l 2
,j(3) ...
P21 P22 P23 = 36 7h.A 37/ 72
*3^ P S P23 2S/S4
Note that = p (” l5p = p (0)p"
at n = 1 , pll) = p “»p
at 11 = 2, p(2) = pCilp = pC0)p 2
at n = 3, p(:̂ = pt'z)p = pWp*
This implies that
_ V „(»- ’ ’
Pij ~ / _ P I k Pkj
k
Definition
Let {9fn , n = 0,1, 2 ....} denote a square of real valued variable index by n. The value
of x for given n is the state of the process at the n th step.
= a , / x0 = a,} = V 3 • 1. V 2 = V 6 P{xn = j / *„_•! = i} is a one-step transition probability matrix. The index n denote
l>{xi = a3 / x„ = a3) = V 3 • 1. V 2 = V 6 something close to time and therefore depend on xn. x, xn_2 *0 anc* not on
P{x3 = a 2 / *„ = a ,j = 2/ 3 V 3 • 1 = 2/ g -V7l + l'*71 + 2>
The Markov assumption is that
' ( * 3 = « . / ^ = «3} = 2/ 3.2/ 3 1/3 = 4/ 27 P^xv ~ in /xv- l — jn-l> xn-2 ~ jn-2>—->x 0 = jo) = P{xn = j r / xn- 1 = /n + l<)
^ * 3 = * 3 / Vo =• a3} 2/ 3.2/3.2/3 = fl/2? The conditional distribution of xn given the whole past history of the process must
equal to the conditional distribution of xn given x ,,^ .
*1* « 4/27
=• V 6 + % = ?/18 Example 15.5: (A Communication System)
Pn,nJ , = l //6 ++ 8//2 7 _- 2 5 //54 Consider a communications system which transmits in digits 0 and I. Each digit
transmitted must pass through several stages, at each of which more is a probability p
2<»K 269
UNIVERSITY OF IBADAN LIBRA
a R
-J Y
^vj
Ci-Oi "
. In order words, we can-say. that the.proccss is in
that the digit entered will be unchanged when it leaves. Letting d'ndenote the digit
State 0 if it rained both today and yesterday;
entering the nth stage, then {Af„,n = 0,l...} is a two-state Markov chain having a State 1 if it rained today but not yesterday;
transition probability matrix. State 2 if it rained yesterday but not today;
P 1 - P (P l - P } State 3 if it did not rain either yesterday or today.P = P =
1 - P P 1 - P P
The preceding would then represent a fair-state Markov chain having a transition on
probability matrix.
Example 15.6 0.7 0 0.3 0
On any given day Gary is either cheerful (C), so-so (5), or glum (G). If he is cheerful 0.5 0 0.5 0
today, then he will be C, S, or G tomorrow with respective probabilities 0.5, 0.4, 0.1. 0 0.4 0 0.6
If he is feeling so-so today, then he will be C, S, or G tomorrow with probabilities 0.3, 0 0.2 0 0.8
0.4, 0.3. If he is glum today, then he will be C, S, G tomorrow with probabilities 0.2, You should carefully check the matrix P, and make sure you understand how it was
0.3. 05. obtained.
Letting Xn denote Gary's mood on the nth day. then {Xn, n > 0} is a three states
Markov chain (State 0 = C. state 1 = 5., State 2 = G) with transition probability 15.7 Stationarity Assumption
matrix. A Markov chain is stationary if for m =£ n
0.5 0.4
P[xn = j n / xn-\ = jn - 1>} = P{xm = jm /xm-l ~ jm- 3'}
0.1
0.3 0.4 0.3 or simply.
0.2 P{xn = j / x r - 1 = i.) = P {X m = j/xm- 1 = 00.3 0.5
In this case the one-step transition probability does not depend on the step number. It
Example 15.7: (Transforming a process into a Markov chain) is therefore sufficient
Suppose that whether or not it rains today depends on previous weather conditions For us to state only the one-step transition probabilities.
through the last two days. Specifically, suppose that if it has rained for the past two We therefore set n = 1 and obtain
days. Even it will rain tomorrow with probability 0.7. if it rained today but. not
yesterday, then it will rain tomorrow with probability 0.5; if it rained yesterday but pjy° = ^ {* 1 = j / x o = 0
not today, then it will rain tomorrow with probability 0.4; if it has not rained in the P,("J = P{.xn = j/x o =
past two days, then it will rain tomorrow with probability 0.2 .
II wo lei lhi.> state at time n depend only on whether or not it is raining at time n. then For, n = 0, this leads to
>he preceding mode! is not a Maikov chain (why not?). However, we can transform p!j) = 1 for j = i
ilus model into a Markov chain by saying that the state at any time is determined by
the weather conditions during both that day and the previous day.
271
UNIVERSITY OF IBADAN LIBRARY
P i j is the probability of the first event, and that of the second is
P i( i ' = 0 for J f t
Pi"1 = P{xn = j / x o = 0 by definition
T,kPikakj
= Y,kP{xn = j , x n. i = k / x 0 = i} marginal from joint
= k ,x0 = i}P{xn. j = /c/x0 = i) Consider a process with the following three states; a1( a2 a3l where afis an absorbing
state, and others are transient.
^ P A n =7 An -1 = W A n - l = kAo = 0
k ,
— P22a2l
= Yk j v* ~ ' )pKi
15.8 Absorbing Markov Chain
A stale in a Markov chain is absorbing if it is impossible to move out of that state.
That is, the process stays there. A Markov chain is absorbing if it can’t least one
absorbing state. That is,
Pjj = 1.0 — P 23a 3 l
A state in a Markov chain is transient or non-absorbing if it is possible to get out of
that stale. That is
Pjj =£ 1.0 for state j.
15.8.1 Probability of a Markov Process ending in a Given Absorbing State
This depend on the given in that state. Let atj denote the probability that an absorbing
chain will be absorbed in state if it states in the non-absorbing state a,.
Method 1 Then
There arc two possibilities, either the first transition is to state ay (in which case the a U = p i x < = a/Ao = )
chain is immediately absorbed) or the first transition is to some transient or non
absorbing state ak ,k * j, and then the process immediately enters states a, f r o m a k.
These arc two mutual exclusive events.
272 273
UNIVERSITY OF IBADAN LIBRARY
Bv substitution
As an example consider the following transition matrix lor absorbing Markov chain
a2t ~ P21 + P a i + P 3O31 with four states. Note that an absorbing state is indicated by probability l2 2 2 2
v 4 v 4 V 2 0
p = V 3 v 3 0 v 3
ciij is a one-linear equation in several unknowns. Construct a corresponding linear 0 0 1 0
equation by using each o f the other transit state as initial state. 0 0 0 1
In the given example, a2l is a linear equation in two unknowns. Note that Ptj is
Note that the absorbing state are a3 and a4.
obtained from the given one step transition matrix. The onlyunknown are akJ-, all k =£ Suppose that we want a13, that is the probability starting from a. will get absorbed in
j slate a3 . In other word, we want the probability that the chain will enter a3 from a1.
The corresponding a21is given by l
Then aiywill give us
a33 = Pl.3 + P n a13 + Pl2a23
a23 = P23 + P23a13 + P22a23
Substitute forpiy. noting that akj is unknown.
~ P?.2a21
ars = V 2 + V 4 a i3 + V 4 fl23
a23 — 0 + 1/g «13 + V 3 a23
~ P33°31 Solving the simultaneous equation, we obtain
« i3 = 4/ 5 and a23 = 2/ 5
The matrix becomes
| fl13 a l4|
I °23
Alternatively.
l hen Naive all values of’equation ii for all ^ j t sue ii as a2 1 and u?1) simullnneouslv
! :>> rr Vi, + 2.H Piltakj
274
27?
UNIVERSITY OF IBADAN LIBRARY
Wo can write this matrix form. Let A denotes the matrix of aiy R denotes the matrix
oI'Pij. Q denotes «3 fl2
The matrix o f pik. That is
A = (af/) = ia kj} -s x r
r
R = (Pa) s x r 1 0 0
0 = i P i k ) S X 5
Then aijcan be written as
0 1 0
A = R + QA fl4
Where
r = number o f absorbing states v 2 0 V *
s = number o f transit states
Step I: Arrange the rows and columns of the one-step transition matrix in which a 0 v 3 V s
way that the absorbing states appear first in the rows and first in the columns.
V
Step 2: partition the new one-step transition matrix as follows
r
- r ~ \ Step 4: Find I-Q and hence ( / - ( ? ) 1
absorbing states 0
transient states 5 ■< R 3/ 4 " 'A( /-< ? ) =
J - v 3 2/3
hrx r )> 0 ( rx s )> N(sxr)> Q (sxs ) u - «?i = (3A ) (2/ 3) - (V 4) (V 3) = ( 5/ i 2)
Step 3: Solve for A. the matrix of the unknown, as follow
[2/ 3 V 3
(/ -Q )A = R C0f(/- c ) = k 3/ J
A = ( / -< ? ) - '/* cofT(/ - Q) - Adj(l - (?) = 2/3 v 3
Since (l-Q) is non-singular and so has an inverse. (/ - (?)_1 is known as the V3 3A
fundamental matrix. , Ad](I - Q) 2/ 3 v 4
l or example, the above 4 x 4 matrix incan be rearranged as follows: since V - Q r = Je t( , _ Q) 12AV3 V4
277
UNIVERSITY OF IBADAN LIBRARY
r8/ s 3/ s i
i 4/s 9/5j
Therefore /! = ( / - Q)_1R =
L4/s 9/ 5J ° V 3.
% V 5
7s 3/sJ
15.8.2 The Expected Number of Times a Markov Process will be in each Possible
Starting Transient (Absorbing) State
Lei N = (liij) where Uij is the number of times the chain is in transient state a;-
given the initial state is at.
Lot n,7 denote the mean number of time that the chain is in transient state a, .
Let N denote the matrix of n (y, which is a square matrix since i and j range over the
transient stateds. Consider the state at time 1. That is, the first time interval is spent in state a, (a( is
Consider a chain with the three states in (a), a a, a2, a3 where aj is the absorbing transient state). If i =£ j and the transition probability pik given the probability that the
state. Assume that the initial state process is a2. process will be in aK from at . Then
nij = T.kPiknkj
nii = Piknki = 1
= dii + 'LkPiknki
Which is combined into
n i i = d U + Y j P ikU ki ' = l> f° r ‘ “ j
l<
= 0, for i j
279
278
UNIVERSITY OF IBADAN LIBRARY
i----
o
r—i
LD
co
L̂T)
<»
In matrix form this can be written as
Recall
djj = Vij + I,kPikakj
d 1 d2 • dn
dx d2 . dn
r " N
d, 1 0 0 r ^ \
dx i 0 0d2 0 1 0
d2 0 1 0
dn 0 0 1
dn 0 0 1
J J
15.8.3 The Length of Time (Expected Number of Steps) Required before
Absorbtion Occurs
For any given initial transient state a, the expected number of step required before {p*} = Q
absorption is given by the elements of the rector M = N
N = 1 + QN
t = Z ” "
i n = (/ - o r 1 1 = (/ -
\)'- i
Thus the element of the fundamental matrix give the expected number of times the
Let c be a column rector with the same number of elements as the columns of N and
every element of C is unity. process will spend in given transient states for any given initial transient state.
Then 8/s 3/s = rl.6 0.61
n = (/ - Q r l =
t = NC a2 4 / c 9 /c l0.8 1 .8J
Interpretation:
T Starting in state au the expected number of times in state al before absorption occurs
where c = 1 is 1.6. Similarly, starting from aa the expected number of times in state a2 before
1 .
By using the above fundamental matrix we find absorption is 1 .8.
The expected number of transition before absorption is
aij™ij = Pijmij + Sfc Pikakj mij
summing over/
15.8.4 The Number of Transitions that will occur before a particular Absorbing
State is reached M = Y j a 0.
or
ZjP j P i 0.5 0.0 0.75 P i
Pi = 0.5 0.5 0.25 Pi
0.5 0.0. Pi.
Pk = I , PiPuc Pi. 0.0
A probability distribution which satisfies pk is called invariant or stationary
= 0.5//! + 0.75//3
distribution (lor a given Markov Chain). In this case row ofP(n) is the probability
//2 = 0.5//, + 0.5//2 + 0.25//3
vector// = (/tl t //2, •••)• Hence, given
Pi = 0 .5 //2
pin) _ p ln-l/p Thus
nl i—m on P(n) = Um P^ -^ P
Pi ~ -SP\
71 —• co
P\Pi PlP2 and
PlP2 — Pi Pi \P] P2 = 2a<3
Substituting we have
4
This can be written as Pi = 3^1
P = pP IJv imposing the normalizing condition on the sum ut we obtain
or P, + P2 ■+ " 1
p T = Pr pT 4 2
Pi "F -j « i + jr/M
I lien:Ibr.:
2X8
2 6 9
UNIVERSITY OF IBADAN LIBRARY
1 D H R
This means therefore that
4 2 D S 2 v 2 0
Pi = and n3 = -
P = H v 4 v 2 V
Thus n = (jiiPiPz) = (V 3 4/ g 2/ 9) R 0 v 2 V
This gives a sample method of obtainingP(7l) than raising Pto power n. V J
Findlimn_co P(n) and give all possible interpretation of the result.
Interpretation:
can be interpreted as follows: 16.4 First-Passage and First-Return Probabilities
1. Probability of a distant state: if a point in time Is fixed in the distant future , //y We shall approach this topic by way o f asking certain questions.
is the probability that the process will be as state j. Q l : What is the probability that in a process stating from a(. the first entry to a;
2. As a time average: if the process is operated for a long time, /iy is the fraction occurs at the nthstep?
of time that the process we be at state j. Q2: What is the number of stepsn, required to reach state ay for the first time?
3. As a fraction of process: if many identical processes are operated For Ql. consider the function/?^ which is the probability that the process will enter
simultaneously, /iy is a fraction of the process that can be found in state j after slate j at the nth step given that it is in state i of the initial step. That is,
a long time.
4. Reciprocal of mean number of transition: iij is the reciprocal of the mean P p> = P[Xn-‘ j\Xo = i)
number of transition between recurrence of the state, that is, average (a) In this case, the process would enter state ay, after onlyk, 1 < k < n — 1 ,
number of transition before a steps.
man inay will come back to a*. (b) After that is called either stay three is ay or change to another state and then
return to ay. For Q2, the probability/^ that the process will reach state ay
Example 16.2 for the first time at the nth step given that it stated from a, is called first-
An individual of unknown genetic character is crossed with a hybrid. The offspring is passage probability and is define as
again crossed with a hybrid, and so on. The states are dominant(D), hybrid (H) and fij * = P{Xn = j-X71—1 ^ J'Xn-2 * j > —>X\ * j\Xo — 0
recessive (/?). The transition probabilities are
Definition: First-Passage Probability
This is the probability that the process is in state ay at time nand not before, given that
it was in state a, at time 0.
290 291
UNIVERSITY OF IBADAN LIBRARY
Tlu> implies that the probability that n steps are required to reach state aj for the first 0.5 0 0
0 0.5 0
time given that the process siartsfroin slate a,. p ! ? * 0 0 0
Clearly
- 0. the process is still at a,- 0.5 0.25 0 0.5 0 O' 0.25 0.25 O'
0 0.5 0.5 0 0.5 0 = 0 0.25 0
/jy11 = Pi,, the one-step transition probability, t =£ j 0.75 0.25 0 . 0 0 0. .0.375 0.125 0.
A Iso. 0.250 0.500 0.250'
P ? JpW = 0.375 0.375 0.250
.0.375 0.500 0.125.
Then, 0.250 0.500 0.250' 0.25 0.25
O'
0.375 0.375 0.250 - 0 0.25 0n
v u H)i - h i I' ll
/ ,>ij P>i
/<=1
or.
_ JJ) r;"’} ore known.
number of steps to get from i to /(the first passage). That is, the number of steps
required to reach a ;- for the first time.
Kxample 16.2
(ii) The number of steps required to get from i to j is therefore a random
Consider (lie problem of departmental transfer in chapter 17.
variableA/,,. with
0.5 0.25 0 />{*.) = "} = / «
r .. 0 0.5 0 .
0.75 0.25 0
2 9 3
UNIVERSITY OF IBADAN LIBRARY
16.6 First Return (Recurrence) m i l = X p t j + ^ ( 1 + m k j ) p lk
( i) If; = t, f£ n) gives the probability of the first return to state a f. For example, k * j
the probability that the person transferred from department a £ will return to a, for = Pi j + ^ Pik + P i k ™k j
the first time at time n. k * j k * j
Corresponding to = ^ Pik + ^ Pi k™k j
all k k * j
fii ̂ = P[Xn = ^ 2 it —,Xi ^tlA'o = i)
= 1 + Y j VikTrLki
k*J
( iii) The equation relating / J n) to would also be the same.
But since £ kpik = 1, This expressesm^- as a linear function of m kj as the unknowns.
t i n) = n * i = zi*0 = 0 = 1 (ii) By using the same relation for other m £y’s a complete set of linear equation
(ii) Then Nu is a random variable whose value is the recurrence of state a£. (equation to the number of unknowns) can be expressed.
(iii) Since {/it(n)} for fixed i , j gives the distribution of fyy. the mean first passage (iii) A solution of the linear equation gives the mean first passage time from any
time from a,- to ay denoted by m,yis given by state into state j .
oo (iv) Mean first recurrence times are obtained in the same way.
m a = E(Nu ) = Y j n^ /n>
n = ] Example 16.3
Consider the three-department job assignment. How many assignments will occur, in
(iv) where t = mu is the mean first recurrence time. the average, before a man who is first assignment to ax (engineering) will be assigned
to a 3 (sales)? That is, what is m13?
16.6.1 Calculation ofm iy-
( I ) The formula in (6.6) for would required the complete first passage time Solution to Example 16.3
distribution for solution to be obtained. Using the formula for m,y
( 2) A simplification of the problem is obtained by conditioning the formula for 7 n 13 = 1 + P n m 13 +- p 12m 23
m(; on the state at step 1. That is. on one value of i at a time. There are two unknowns. Hence we form a similar equation for m23 as follows.
(3) Given that the process is in state at at time 0, either the next state is a; in m23 = 1 + Pzim l3 + P22m 23
which case /V(/ = 1 , or it is in some other state ak afier which it enter state Now. recall that
ar in which case the passage time will be m kj = 1 + NkJ, the passage time 0.5 0.5 0 '
p = 0 0.5 0.5
from ak to fly. .0.75 0.25 0 .
By substitution we obtain
(i) Thus m13 = 1 + 0.5m)3 + 0.5m23
2 9 4 2 9 5
UNIVERSITY OF IBADAN LIBRARY
m23 = 1 + 0.5m23
Solving the simultaneous equation, we found that C H A P T E R 19
mi3 = 4- and m 23 = 2 CH APM AN-K O LM O G O R O V EQUATIONS AND
C LASSIFIC ATIO N OF STATES
Practice Questions
1. Define the term, steady state probability
17.1 Introduction
2. Write an expression for a limiting distribution.
The n,h-step transition probabilities P" is the probability that a process in state / will
3. Solve completely the problem in example 5.1, and draw all the graphs.
4. Use matrix multiplication and limiting probabilities to solve Problem 5.2. be in state j after n additional transitions that is,
5. In the post test in lecture four, obtain the stationary probabilities. /7= H * ,,-„ -^ |X „= ;} ,n 20 ,i ,j ;> 0 .
6. Define and write an expression for The Chapman-Kolmogorov equations provide a method for computing these n-step
(a) First-passage probability. transition probabilities. These equations are:
(b) First-return probability. p;p?kp;' for all n, m > 0, all i, j
7. Using the post test of lecture four, find the mean first passage time from state
5 to state 4 by making state 4 absorbing. (This has nothing to do with states
and are established by observing that
{1,2}.) (The University o fS ydney, 2009)
8. The transition matrix P of a Markov chain X = (Xn: n > 0) is fT " = |x.. = 'l
= Z ^ f r — “ -/ .X. = K |X ,= /}
. 1 2 3 4. 5 • K
r - 1 p \x - . - J - !x - - *• x » = ' M x „ = * 1* . = -I
I •. o 0 0 0: 1 1 K-1'
2 . 0 ■ 0 1/3 1/2 0 - i c c
3 0 0 1 " 0 0 (.=ii
4 .: 0 1/3 0 1/6 1 /2 If wc let P"” denote the matrix of n-step transition probabilities, P”, then it can be
5 1 /2 0 ■ .0 0 1 / 2 . asserted that
p i n —m) __ p i n I p ir n I
V ■ ■
(a) Specify the classes of this chain and determine vyhether they are transient, null where the dot represents matrix multiplication.
recurrent or positive recurrent. Hence,
(b) Find all stationary distributions for this chain. p m ) _ p p i i i - i ) _ p p p i " - - i _ _ p "
(c) Find the mean recurrence time m.jj for all positive recurrent states. and thus P,nl- may be calculated by multiplying the matrix P by itself n times.
(The University o f Sydney. 20 JO) pniis said to be Accessible from slate / if for some P >0. Two states / and j
accessible to each other is said to communicate and wc write / j.
296
297
UNIVERSITY OF IBADAN LIBRARY
Let denotes the one-step transition probabilities, and />! = Pt Solution:
Observe that P"k P" represents the probability that starting in / the process will go to The one-slep transition on probability matrix is given by
'0.7 0.3'|
stated in n + ^transitions through a path which takes it into K at the nth transition. p =
.0.4 0.6 J
17.1.1 Proof of C - K Equations 07 fo.i 0.3Hence. Pm = P: °-3)
Using remark (3) above, summing over all intermediate states /(yields the probability 0.4 0.6 1,0.4 0.5
that the process will be in state j after n + m transitions. We have
r r = e k « = j \x , =;} 0.61 0.39'
v 0.52 0.48,
T " = L k . . , = ^ x . = * k = ' ' }
f 0.61 0.39'j 0.61 0.39'j
= 5> k,., = V ,K .t .X 0 = /}p {x „ = K\X„ = l} 0.52 0.48 J ,0-52 0.48 J
*«0
p;
*=0 f 0.579 0.4251 \
[ 0.5668 0.4332 J
Matrix ofn-slep transition probabilities: P(nl
Let Plnl denote the matrix of n -step transition probabilities P'j then the C-K Equation Hence, the required probability P*n equal 0.5749.
asserts that
pin-rm) _ p lm ) p(m) Example 17.2
Consider Example 2.4. Given that it rained on Monday and Tuesday, what is the
By induction
p‘ probability that it will rain on Thursday?p\.»\ _ p i" - \* k ) _ p i i-l _ p n
That is the n -step transition matrix may be obtained by multiplying the matrix P by
Solution
itself n times.
The two-step transition matrix is given by
Example 17.1
Consider example in which the weather is considered as a two-state Markov chain. If
a = 0.7 and p = 0.4, the calculate the probability that it will rain four days from today
given that it is raining today.
298 299
UNIVERSITY OF IBADAN LIBRARY
Proof: the Is' two parts follow trivially from the definition of communication. To
0.7 0 0.3 0 '0.7 0 0.3 0 prove (iii) suppose that /<-> /., and j k then there exists m,- n such that
p{2) = p'- = 0.5 0 0.5 0 0.5 0 0.5 0 P"' > 0, P" > 0. Hence,
0 0.4 0 0.6 0 0.4 0 0.6
0 0.2 0 0.2 0 P',r = t P"' P* - K p"k > 0
/•=U
Similarly, we may show there exists an S for which Pks. > 0 . Two states that
'0.49 0.12 0.21 0.1 8n
0.35 communicate are said to be in the same class and by the proposition any two classes 0.20 0.15 0.30
0.20 .arc either disjoint or identical. We say that the Markov chain is Irreducible. If there is 0.12 0.20 0.48
, 0.10 0.16 oniy one class- that is, if all states communicate with each other.0.10 0.64,
Stale is said to have period d i f P ” = 0, whenever n is not divisible by d and d is the
Since rain on Thursday is equivalent to the process being in either state 0 or state 1 on greatest integer with this property. (If P* = 0, for all n > 0; then define the period of i
Thursday, the required probability is given by />2„ + = 0.49 + 0.12 = 0.61 to be infinite). A state with period 1 is said to be A periodic. Let d (i) denote the
period of/, we can show that periodicity is a class property.
17.2 Classification of States
In order to analyze precisely the asymptotic behaviour of the Markov chain process, I7.2.2Recurrcnt (or Persistent) State
we need to introduce some principles of classifying state of a Markov chain. A state f e S is said to be Recurrent if Pr(7] i; yesterday, then it will be sunny tomorrow with probability 0.6; if it was cloudy today
(iii) If i j, then j <-» i; then i k. but sunny yesterday, then it will be sunny tomorrow with probability 4; if it was
cloudy for the last two days, then it will be sunny tomorrow with probability 1 . '
301
300
UNIVER ©ooSITY OF IBADAN
LIBRARY
oo->
o
Definitely, the model above is not a Markov chain. However, such a model can be (b) Obtain the transition matrix
transformed into a Markov chain. (c) Find the stationary distribution in terms of p and q where p + q = 1
(a) Transform this into a Markov chain
(b) Obtain the transition probability matrix Solution
(e) Find the stationary distribution of this Markov chain. The state are (2, 0), (1,0), (1, 1), (0, 1).
The transition matrix is
Solution
To state-* (2,0) (l,0) (l,l) (O.l)
(a) Suppose we say that the state at any time is determined by the weather
From state
conditions during both that day and the previous day. We say the process is in: 9 P 0 . 0
State (S. S) if it was sunny both today and yesterday; P = (2,0) 0 0 <7 p
State (S. C) if it was sunny both yesterday but cloudy today; (1 . 0 ) P 0 0
Slate (C, S) if it was cloudy yesterday but sunny today; 0.0 0 I 0 0
State (C, C) if it was cloudy both today and yesterday (o.i)
(b) The transition probability matrix is 17.3 Discrete Time Process
Today’s state 1. Consider a series of events E resulting from the repetition of the same
(S. S) (S, C) (C,S) (C,C) experiment and occurring consecutively. The common examples are telephone Yesterday's staie(S, S)
.8 .2 0 0 calls, average customers at a service point, chromosomes breakages and (S,C)
0 0 .4 .6 radiation, and so on(C,S)
.6 .4 0 0 2. The occurrence are assumed to be of the same kind y.The number n of events (C,C)
0 0 .1 .9 in a given interval t is a random variable.
3. Letz ( t ) denote the total number o f occurrences within an arbitrary time
2. An airline reservation system has two computers only one of which is in interval t.
4 I t lP .U) = P(*tO = n).
operation at any given time. A computer may brake down on any given day which
probability p. there is a single repair facility which takes at least 2 days to restore a Assumptions
computer to normal. The facilities are such that only one computer and a time can be i. Pn(t) depends only on the time interval of duration t, and does not depend on
dealt with. the initial instant.
ii. The probability that E will occur more than once is the time interval that is
(a) Fonn a Markov chain by taking as states the pairs (x, y) where x is the number infinitesimally small (that is, negligible).
of machines in operating condition at the end of a day and y is 1 if a day’s labour has iii. The probability that E occur once in the interval dt is proportionaly to that
been expended on a machine not yet repaired and 0 otherwise. interval and is written
As Xdt.
302 303
UNIVERSITY OF IBA
AN LIBRARY
ii. it is independent of the values of z(t)prior to t0.
Assumptions on z ( t ) Thus if z ( t0) is known, the value z(t) (which is determined by the probability of
i. the initial z ( t) is 0. occurrence in the interval z) depends solely on the law of probability that governs the
ii. z(t)) increases by I when E occurs.
increment nafter t„.
iii. z(t) remains constant when E does not occur.
The random variablez(t) follows a defined poisson process and constitute an example
of a Markov chain. Jt is defined completely by the probability
(Z7t Y e - O '
Pn(0 = n = 0,1...n!
17.4 The Poisson Process
Under the assumption of the continuity of time we can expand py (At) in Maclaurin
series
Pi At = P i(0) + pi(0)A t + ^pi'(0)(A t)2 + -
= Pi(0) + p i(0 )A t + o(At)2
But pjCO) = 0
The probability that Ecan occur 0 time or once is
iv. z(t) = 0, 1 , 2,..., n , ... .At random instants ty, t2, ..., t*, ...it jump abruptly from
0 to 1 , 1 to 2, and 2 to 3, ... . The increment of z(t) at time interval t is equal to the p{z(At) = 0 orz(At) = 1} = p0At + pyAt
number n of events that have occurred. Now consider pnAt
v. If we know the value of z (t0)att0(the initial instant) we can find the value at p0At + pi At + ••• = 1
t = t„ + At.
z(t) = z(t0 + At) which is necessary.
= z (t0) + z(At)
= z(t0) + p0At = 1 - px At - p2At - •••*
The increment z(At) = n is characterized by the properties of the probability of = l - p ,A t
occurrence in the time interval if This is so because of the assumption 1 - py At - o(At)2. But our only interest is in
i. its probability is Pn(At) Pn(t)-
304 305
UNIVERSITY OF IBADAN LIBRARY
Event E can occur precisely n times during the interval t + At if the following This equation does not hold for n = 0. We can use the forward chapman-kolmogorov
mutually exclusive events are time. equation.
i. E occurs n times in the interval t, 0 times in the interval At. p0( t+ At) = Po(t)p0( At)
ii. E occurs n - 1 times in interval t, once in the interval At. = p0(t)[l-A A t]
iii. E occurs n - 2 times in the interval £, twice in the interval At.
And so on. p0( t + A t)-p o ( t) n ^
---------- Tt---------- —
These lead to the following:
pn(t + At) = pn(t)p 0At + pn_1(t)PiAt + pn_2(t)p2At + - In the limit asAt -» 0
Pi At = 1 — p[ (0)At — o(At)2 Po(0 = -^Po(t)
sincepi(O) = 0
Letp] (0) = A, then At the beginning of the interval t, we have
Po At = 1 - AAt - o(At)2
po(0) = 1 andp^CO) = 0 , 7i * 0
PiAt can be written as
P i At = AAt + 0(At)2 Divide the limit result by p0(t), we have
So that £o(£) _
pn(t + At) = pn(t) [ l - AAt] + Prj-i(£)AAt + o(At)2 Po(0 Po(0PO
= Pn(0 - P„(OAAt + p„_i(t)AAt + o(At)2
Pn(t d + At) - pn(t) = [pn_t ( 0 - p„(t)]AAt + o(At)2 or— l° 9ePo^ = _A
Divide both side by At
p„(t + A t ) - p n(Q J^ lO g e P o ( t) = ~ x j dt
At = ^ b n - l ( 0 - P n ( 0 ] + 0(At) 2
As At -> 0 logep0(O= - X t
^ P n ( t) = pn'( t ) = APn-l(0 “ Apn(t), 71 = 0, 1,2,...
PoCO =
3 0 6 307
UNIVERSITY OF IBADAN LIBRARY
Ane -Xt An t j e~Xttr
^p„(t) can still be written as N o w p M = T 7 T = 1 F ^ W ’ i ~ 0,1,2'
0pn(O = *P n-l(0 “ Apn(0 » n > 0 so that
Ae ' Xl \ t Je~Xtt r _
Atn = 1 P , ( 0 ~Z) + A _ ( r + ; ) ! ! !
0Pi(O = *Po(O -A pi(0
putr = 1 , j = 0
At n = 2 At ° e -xtt1
Pi ( 0 = = Xte~Xt = (At)e_Xt
Dp2(t) = Apj(t) - Ap2(t) (1 + 0) ! 1 !
Atn = 3 Then
Z)p3(t) = Ap2(t) -A p 3(t) (D + A )p 2( t ) = X2te~xt
Dpx(t) can be rewritten as A2 te-Xc _ X2t j e~xit r
p2( 0 =
(D +A)p1(t) = Ap0(t) D + A “ (r + / ) ! ! !
So that we have put r = 1 , j = 1
2^-Xt
(D + A)Pl(t) = Xe-Xt X2t2e~xi (*Q 'e
P2( 0 = 2 ! 2 !
Divide through by (£) + A)
Consequently
Xe~xt Xe~Xct r ^ _ X3te~xt X3t j e~Xct r
Pi (0 = D + A ( r + 1 ) ! 1 ! = D + A = (r + » ! ! !
This is a general solution. put r = 1 , j = 2
^3j3g-At _ O T _ f ‘
p3( 0 =
Xe~Xlt r \ e~ Xtt r 3!
3!
may also be written as
Noticetliat c T T i j r n r! 1 !
309
308
UNIVERSITY OF IBADAN LIBRARY
In general, Suppose that a continuous-lime Markov chain enters state / at some time, say time 0,
C O = (At)ne~XtPn ,n = 0, 1 , 2, and suppose that the process does not leave state / (that is, a transition does not occur) n\ during the next s time units. What is the probability that the process will not have
state / during the following t time units?
If we fix t, At is a fixed parameter for the distribution and the set P i(t),p2(t).... then To answer this, note that as the process is in state /' at time s, it follows, by the
gives a probability distribution of the process at the fixed time interval which is a Markovian property, that the probability it remains in that state during the interval
Poisson distribution. In terms of a counts of events the above results shows that the [s,s + t] is just the (unconditional) probability that it stays in state i for at least t time
member of events occurring in a fixed time interval t is distributed as a Poisson with units. That is, if we test t; denote the amount of time that the process stays in state i
parameter At. before making a transition into a different state, then
Also since the mean of the Poisson distribution is equal to the parameter At, At can be
P = {ri > 5’ + r|7j > 5 }= /3{7; > t)
interpreted as the expected number of events that can occur in time t. the quantityA is
the average or mean rate of occurrence of E.
for all s ,t > 0. Hence, the random variable 7) is memoryless and must thus be
exponentially distributed.
17.5 Continuous Time Process The above gives us a way of construction a continuous-time Markov chain, namely, it
A continuous-time Markov chain is a stochastic process having the Markovian
is a stochastic process having the properties that each time it enters state /:
property that the conditional distribution of the future state at timet + s, given the
(i) the amount of time it spends in that state before making a transition into a
present state at t and all past states depends only on the present state and is
different state is exponentially distributed with rate say vL \ and
independent of the past. Thus, this lecture establishes the fact that a continuous time
(ii) when the process leaves state /'. it will next enter state j with same
process is also distributed as an exponential probability
probability, call it p,-y, where * i pfj = 1.
17.5.1 Definition and Properties
Consider a continuous-time stochastic process [x{l), t > o} taking on values in the set A state i lor which u,- = 00 is called an instantaneous state since when entered it is
of non-negative integers. In analogy with the definition of a discrete-time Markov instantaneously left. Whereas such states are theoretically possible, we shall; assume
chain, given earlier, we say that the process [x{l), t > o} is a continuous-time Markov throughout that 0 < v>. < co for all /. (If v, = 0, then state /' is called absorbing since
chain if for all s , t > 0 and non-negative integers i , j , X 0 < u < s, once entered it is never life).
Hence, lor our purposes or continuous-time Markov chain is a stochastic process that
moves from state to slate in accordance with a (discrete-time) Markov chain, but is
such that the amount of time it spends in each state, before proceeding to the next
If, in addition P^\X(,t i )= j |.Y(j) = l} is independent of s, then the continuous-time stale is exponentially distributed. In addition, the amount of time one process spends
in slate t and the next state visited, must be independent random variables. For if the
Markov chain is said to have stationary or homogeneous transition probabilities. All
next slate visited were dependent on 7). then information as to how long the process
Markov chains we consider will be assumed to have stationary transition probabilities.
310 311
UNIVERSITY OF IBADAN LIBRARY
has already been in state i would be relevant to the prediction of the next state-and
this would contradict the Markovian assumption. . A continuous time stochastic process is said to have Markov property and is called a
A continuous-time Markov chain is said to be regular if, with probability 1, the continuous time Markov process if for all t* > > -- .ty > t0 satisfying the
number of transitions in any finite length of time is finite. An example of a non- condition tn > tn > t0.
regular Markov chain is the one having.
Pi, i+ 1 = 1, Vi = i2 1. P(X(tn) = jn/X (tn- i ) = jn-l>X(tn- 2) ~ jn-Z' •••••»^(^o) = jo)
It can be shown that this Markov chain-which always goes from state i to i + 1, = P( X(tn) = jn/X ( tn_i) = j n- i)
spending an exponentially distributed amount of time with mean ' /2 in state i - will,
with positive probability, make an infinite number of transitions in any time interval This is the independent probability and it state that all that is needed to predict the
of length t ,t > 0. We shall assume from now on that all Markov chains considered state of the process at time n is the state of the process at the immediately preceding
are regular. time.
Let qtj be defined by 2. A Markov process is said to be time-homogeneous or stationary if
Qij = VtPij, V i * j
P{X(t2) = y /^ ( t1) = i) = P iX iti -h ) = j /X ( 0) = i)V i and j, tj < t 2
Since v, is the rate at which the process leaves state i and p,y is the probability that it
In words, the process is stationary or time homogeneous if the conditional probability
then goes to j , it follows that qtj is the rate when in state i that the process makes a in (2) depends only on the time interval between the events considered, rather on the
transition into state j \ and in fact we call qtj the transition rate from / to j. absolute time. Note that ‘time-homogeneous’ and ‘stationary’ denote sameness in
Let us denote by Pij(t) the probability that a Markov chain, presently in state i, will be lime. We can also know that a stationary Markov process is defined completely by the
in state j after an additional time t transitional probability function which we defined as
P j M = = M m = '}
PijCO = p M O = ;'/* ( 0) = i }
17.6 The Exponential Process The fundamental equation for stationary Markov process is Chapman-Kolmogorov
Let us consider a finite state but continuous time process. Let X (t) denote a random equation for p,y(t + r). By definition,
variable. The value of X (t) at fixed t is the state of the process at time t.
A time dependent process is the set (/(t)fo r given t > 0. X fo)depends on ^ > t0, fjtl(l + r) = P[XU + t) = j/X (0 ) = 0
and not on t2* ><£2- The process is continuos if t can take value on the t-axis. = ^ P[X(t + r) = j ,X ( t) = k /X {0) = 0 Marginal from joint
Definition k
Using Markov assumption
- ^ P{X(l + t) = j /X { t) = l(,X(0) = i) P{X(t) = k,X (0) = 0
k
312
313
UNIVERSITY OF IBADAN LIBRARY
But P{X{t) = k ,X( 0) = /} = P{X(t) = k / X (0 ) = flP fvff)) = i}
Under the assumption that py(t) is a continuous function oft, we can express
Therefore, Py(At)by the use ofMaclaurin series.
Pij(t + r) = P{X(t + t ) = j /X( t ) = k,X( 0) = 0 W O = k / X (0) = i}
k P u m = p u ( o ) + + i p ^ ' t o x ^ o 3 + -
This is because PfA'(0) = t} = 1
Thus = PiyCO) + Pi'yC0)^t + OĈ Jt)2
^ P « t + r) = j /X ( t ) = = k /X (0) = /)
k LetpijiP) = Ay
By the stationary assumption in (2) Py(At) = Pi/ (o) + Ay At -I- o(At)2 for i * j
Plj(t + r) = £ P{X(t) = y/X(0) = 4} P{X(t) = k/XQS) = £} Py(At) = Ay A t + o(At)2 for i = ;
k
= Pkj(j)Ptktf) (By definition) •
k Also, let p'y(0) = Xjj
This is the general form of Chapman-Kolmogorov equation.
Py(At) = 1 + Py y At + o(At)2
A specified form of this is:
= 1 + Xjj At + o(At)2
P i j ( t + A t ) = Y P i k ( t ) p k j ( A t )
k
The above is forward Chapman-Kolmogrov equation. Sincep-y(O) = 0 /o r i * j is a minimum, Ay is positive. Also, since Py(0)
For i = j is a maximum, Ay is non-positive.
The forward Chapman-Kolmogrov equation is given as We can unite the forward Chapman-Kolmogorov.
Pij(At + t) = ]T p ijt(d t)p k;(t) Py(t + At) = ^ P i k (t)pky(4t)
k k
= Py (t)pyy(At) + ^ p ifc (t)pky(At)
We expect the following to hold k * j
i ) 0 < P i j ( t ) < 1 f o r a l l t
ii) Pij{ 0) = W ( 0) = j/X { 0) = t) = l , i = j
= 0 , i * j
And for any given i
iii) PikCO = I , W O = j / X (0) = 0 = 1
314 315
UNIVERSITY OF IBADAN LIBRARY
Winch wo can write as
dpjj(t)
P ,-/(t + A t) = PiyCOfl + V c + uC4t)2] + ^ pik ( 0 [ 4 ; 4 t + 0(At)2] dt
k * j i
= Pij(t) + Vij^jjAt + py (t)0 (d t)2 + Y j [ P i M l kjA t + Pik(t)OOdt)2]
k*i £ p « (0 ) = °
Pij(.t + At) - Pij(t) i
At
Z f t = °
v -1 j
Pij(t)Xjj + 2 ^ Pik ( 0 Xkj
k*j X A«v = xn + 2 a ‘7 = 0
y y**
+ M £C ft) I t « jP a (t)O(^Q2
4 t d t
Thus since every of A (diagonal element) is non-negative, the diagonal element Ay
dt must be equal in magnitude and opposite in signal to the sum of the other element in k
the same row. Ay is called the transition rate from i to j for iit j . Ay can be
'I'he limit as d t -» 0
interpreted as the parameter of negative exponential distribution. For each Ay, the
dPi/(0 V exponential distribution gives the distribution of time spent in a state i, given that j is
~ ~ d t~ = Z j Pik the next step. Thus if Ty is the random variable with Ayk
In matrix form.
(a= diagonal element) i ’(Ty) = f'HJ
d m dp a it) So that Ay can be estimated as the inverse of a sample mean.
dt dt A= f t ; )
/ (* ) = <* c.-te
P M = (pyM ) or f ( t ) = Ay-c " V . x > 0
B u t ,^ p „ ( t ) = 1 with mean
CO CO
d E(Ty)= J t / ( ( ) d t = J a ljer x‘l‘ dt
dt ^ P./(0 = 0 0 0
317
UNIVERSITY OF IBADAN LIBRARY
C H A P T E R 20
= j - [ IN T R O D U C T IO N T O T H E T H E O R Y O F G A M E S AND O U E U IN G
M O D E L S
=-j—sinccr(2) = 1
A ij A ij 1 Games Tkeary
(mines theory is a branch of Stochastic Processes that can be applied to a situation
Suppose thill we have the likelihood such as business, stock trading, politics, and so on. where the person involved can be
n
referred to as a player ox simply a gambler
«=1
18.2 Gambler’s Ruin
Consider a gambler who plays a game of chance against an adversary. Suppose that at
-n lo g X tj- 'Y ^ X ij t the start of the game, the gambler deposits an amount in nairaZ. The adversary deposit
dlogL _ n s r 1
0 N ■ /. in naira where N is the cumulated initial capital.
a i < 1. 'The role of the game is that if the gambler wins a game he takes N1 from
- = - = t the adversary and loses same to the adversary otherwise.
2. The game terminates. If dither player loses all his deposits. When the
gambler loses all his deposit, he is said to be ruined:
This implies that, 3. No game is jumped.
We cun pul the money on a number scale.
C = Vf
Practice Questions
I. Considered a two-state process such as the operation of a loom for weaving I 2 .......... Z ............................................N
cloth. The two-state for the looms are 0, the loom is shut off and the operator is
repairing it. And 1, the loom is operating and the operator is idle. Consider the 1 lie uaiii in loss is represented by movement along the scale. Gambler’s gain is
operating and repair time as continuous. Assume that the constant proportionality is 3 represented by movement to the right observed and its loss represented by movement
lor repair transition and 2 for breakdown transition. Find the probability distribution m the lelty observed.
of the repair and the operation time. No point .ii. tin. scale is jumped. Movement in either direction on the scale is by pure
Obtain the general form of the Chapman-Kolmogorov (C — K) equation. chance. The movement along the scale can be seen as that of a particle that moves at
3. Show that Aiy transition from i to j , V i & j . is 1/p
319
318
UNIVERSITY OF IBADAN LIBRARY
random forward and backward. Because of that the process is. known as random
tvalk. The points on the scale represent the state of the process. Since P{R\Zj] — qZj, P[Zk) = p and P{Z,} = q
Movement- to a point on the scale depends on the point the gambler (or adversary) is
at currently. It is therefore a Markov Chain. Generally (10.1) can be written as
q-/, = VQxk + qq/(. 1 -q/2) = q(q/2-qz3)
probability of winning a game.
Let q = 1 - p, the probability of moving to the left of Z, that is losing a game (by the This implies that
gambler). Let the points on the scale be denoted by Z0lZv ...,ZN and qZj, the
{R\Z2) = P{Zi,R} + P{Zl,R) - q/< = r(q/A - q*s)
= P(R\Z,)P{Z.3) + P{R\7.JP{Z,) To unify these equations we define
q-,N=q^= o
We can write this as These arc boundary conditions on qy . This becomes
q/2 = pqx3 + *,, i < z, < n - i q-/3 - q/., = Tq/.,
320
UNIVERSITY OF IBADAN LIBRARY
This extends to other equation in the system 1 - r ‘
~ <1*, = r ^ z , 1 - r 9z4
1 - r
- ̂
1 - r ‘
<7z0 ~ i +/*/*•-■ = 1 i f p*q
Then by substitution, equation 10.12 becomes
I his is the same as
The general solution can be written as
I Tie particular solution o f 10.1 1 can be written as
I his becomes The boundary conditions
. V 1 ' \c /X ' q0 = 1 and q N = 0
That is, when Z= 0, qz — 1
and when Z= /V, qz — 0
I )i vide by XY‘
This implies that
A + B = 1, Z = 0
/4 + 5( % ) = °’ Z = yV
325
UNIVERSITY OF IBADAN LIBRARY
Solving the system of equations, we obtain In general1' ’
qz = A + BZ
Under the boundary conditions
q0 = I,- qN ==0 . at Z = 0 and N respectively.
Substitute for A and B, we have
Thus,
>4 = 1 at Z - 0
and
A + B N = 0 ' at Z = N
Thus,
Substitute for A and B
18.2.2 Gambler’s Expected Gain(G)
Possible Values q2 = \ - — ZN
Gain N -Z with probability l - q z
Loss Z with probability q 7 Substituting for qz
e ( G ) = 4 - ( i - Z / n I - z
The expected gain is
£■((7) = [Combined Capital) (Probability of gain) - (initial papital) = n {z/ n ) - z
= N ( \ - q , ) - Z = 0
That is
18.2,3 Expected Duration of the Game
C(C/)= M (l-q z )- Z
Assume that the expected duration of the game has a known value Dz . If the first trial
If /;== - ()r q + p * 0 results in a success, the game continues as if the initial position wasZ + 1 .
Now, the initial position is Z, so that
We can write qz as a function of Z
Dz = PDZ+, + q D 7,
/., = /* //,: + /r-4pq The complete solution is
- 2p Dz = UZ+ Vz
Multiplying the solutions results in
x = l
P
328 329
UNIVERSITY OF IBADAN LIBRARY
The required boundary conditions are- - So that
A + B — 0, Z = 0, Dy — 1 Nq-p
M S w % A = —= —— ; Z = N, D n =0
q - p 1
For Z — 0, Dz — 0 Substituting for A and B in (10.39) we have
A + B = 0 J L J L (1 V
n _ Q - P ,. Q - P ^\Pp J'
L>Z ~ 7 ~ 7 7 T + 7------------- ---- + '
For Z = N,D n = 0
q - p N 'W
q - p q - p
So that ' { %
183 Queuing Theory
The principal pioneer o f queuing system was A.R. Erlang, who began in 1908 to
This results in study problems of telephone congestion for the Copenhagen Telephone Company. He
was concerned with problems such as the following: A manually operated telephone
exchange has a limited number (one or more) of operations when a subscriber
attempts to make a call, the subscriber must wait if all the operations are already busy
making connections for other subscribers. It is of interest to study the waiting time of
subscribers e.g. the average waiting time and the chance that a subscriber will obtain
service immediately without waiting and to examine how much the waiting times will
So that we have be affected if the number of operations is affected or conditions are changed in any
other way. If there are more or if service can be speeded up, subscribers will be
^ r = 0 pleased because waiting will be reduced, but the improved facility will become
- o expensive to maintain, therefore, a reasonable balance must be stntck.
Solving for A, 183.1 Applications of Queuing Theory
- N When persons or things needing the services of a facility or persons arrive at a service
q - p channel or counter on the account that the facility or persons cannot serve all at a
time, a queue or waiting line is formed. Examples of this include:
330 331
UNIVERSITY OF IBADAN LIBRARY
(i) cars arriving at a fuel station waiting to be served.
(ii) persons waiting at a bus station waiting to be checked in.
(iii) books arriving at a librarians desk.
(iv) patients waiting to see a doctor or community health dispenser.
(v) customers arriving at a departmental store (supermarket).
(vi) clients waiting to see the Customer Service Executive or Officer.
Arriving customers Served customers
Queuing theory is applied into every field of human endeavour. This is because there leaving
is no perfect service or treatment that can be meted out. Below are some of the fields
of application: ' 1
(i) Business - banks, supermarket, booking offices, and so on. discouraged n
(ii) Industries - servicing of automatic machines, production lines, storage, and so customers
on. leaving
(iii) Engineering - telephony, communication networks, electronic computers, and
so on. A QUEUING SYSTEM
(iv) Transportation - airports, harbours, railways, traffic operations in cities, postal (OR by Swarup et al. 1978, p505)
services, and so on.
(v) Others - elevators, restaurants, barber shops, and so on. 18.3.3 Components of the Queue System
A queue situation can be divided into five elements. These are:
18.3.2 Concept and Definition (i) 'Arrival mode
Queuing theory is concerned with the design and planning of service facilities to meet (ii) Service mechanism
a randomly fluctuating demand for service in order to minimize congestion and (iii) Service channels
maintain economic balance between service cost and waiting cost. The cost here (iv) System capacity
refers to time. (v) Queue discipline
A queuing system is composed o f customers arriving at a service channel and is
attended to by any one or more o f the service attendants. If a customer is not served (i) Arrival Mode - this refers to the rate at which customers arrive at a service
immediately he may decide to wait. In the process, however, a few customers may centre and the statistical law which governs the pattern of arrival.
leave the line if they cannot wait. At the end of the process, served customers leave Certain definitions pertaining to the arrival of customers:
the system. bulk or batch arrival: more than one arrival allowed to enter into the system
simultaneously.
balk: customers deciding not o enter a queue because it is long or lengthy.
renege: customer leaving a queue due to impatience.
332 333
UNIVERSITY OF IBADAN LIBRARY
Symbols and Notations
jockey, customer jostling among parallel queues. We shall employ the following symbols and notations this lecture:.
stationary, arrival pattern which does-not change with time. n = . number of customers in the system, both waiting and in service,
transient: a time-dependent arrivalj>ro«ess. A = average number o f customers arriving per unit o f time
The arrival mode is always denoted by M. . average number of customers being served per unit of time
A
• r 4 traffic intensity
(ii) Service Mechanism - this refers to the number o f service points that are
available and the duration o f service. When the service points or servers are infinite, c = number o f parallel service channels (servers)
the service will be instantaneous, which will result in no queue. In case of finite E ( n ) = average number o f customers in the system, both waiting and in
points, queue is inevitable. Customers can be served according to a specific order, service
which may be in batches o f fixed size or of variable size. This system is called bulk E (m ) = average number o f customers within in the queue
service system. E (v) = average waiting time of customers in the system, both waiting and in
service.
(iii) Service Channels - where there are more than one channel of service, then £ 0 ) = average waiting time of a customer in the queue
arrangement of service may be in parallel or series, or a combination of both, P n ( 0 = probability that there are n customers in the system at any time t,
depending on the system design. both waiting and in service.
^n = time independent probability that there are n customers in the system,
(iv) System Capacity - most queuing system are limited in such a way that both waiting and in service.
waiting rooms are all accommodating. This gives limit to the number o f customers
that can be accepted to the waiting line at any given time. Such situation gives rise to 18.4 The Basic Queuing Process
f in ite source queues, and results in forced balk. The statistical pattern by which customers arrive over a period o f time must be
specified.
(v) Queue Discipline - this is a method of customer selection for service when a It is usually assumed that they are generated according to a Poisson process that is, the
queue has been formed. The different forms of discipline include: number of customer who arrives until any specific time has a Poisson distribution.
(ai) First Come, First Served (FCFS), or The Poisson distribution involves the probability of occurrence of an arrival and is
(aii) First In, First-Out (FIFO) independent o f what has occurred in the preceding observation. This Poisson
(b) First In, LasvOut (FILO) assumption indicates the number of arrivals per unit time(A) (or mean arrival rate),
(c) Last In, First Out (LIFO) while1/^ on the lengthy o f interval between two consecutive arrivals. This time
(d) First in. First Out with Priority (FIFOP) between two consecutive arrivals is referred to as "inter-arrival time.”
(e) Selection for Service In Random Order (SIRO) The mean service rate /i is the number of customers served per unit time whole
average service time (V jt) >s the l'me un‘ts Per customer service time delivered is
335
334
UNIVERSITY OF IBADAN LIBRARY
18.6 Classification of Queuing System
given by an experiment distribution where the servicing of a customer takes place Queuing systems, generally, may be completely specified in the following symbolic
between the timet andt + At.
forms:
(a|Z>|c):(d|e)
18.5 Poisson Process and Exponential Distribution
Description
In queuing theory, the arrival rate and service rate follow a Poisson distribution.
First symbol (a) - type of distribution of inter-arrival times
However, it should be noted that the number of occurrences in some time interval is a
Second symbol (b) - type of distribution of inter-service times
Poisson random variate, and the time between successive occurrences is an
Third symbol (c) - number o f servers
exponential distribution. Both are equivalent
Fourth symbol (d) - system capacity
Fifth symbol (e) - queue discipline
18.5.1 Axioms of the Poisson Process
Given an arrival process [ N ( t ) , t> 0 ] , where N (t) denotes the total number of For the first and second symbols, the following letters may be used:
arrivals up to time t, N (0) = 0. an arrival characterized by the following assumptions M = Poisson arrival or departure distributions
(axioms) can be described as a Poisson process; Ek = Erlangian or Gamma inter-arrival or service distribution
G1 = General input distribution
AXIOM 1 - the number of arrivals in non-overlapping intervals are statistically
G = General service time distribution
independent. This means there is independent increment in the process.
An example of a queue system is
AXIOMS 2 - the probability of more than one arrival between time t and time (M\Ek\Cy.(N\SIRO)
t + At is o(At); this means there is negligibility in the probability of two or more
arrivals during the small time interval At. This implies that Queuing system is classified into
p0(AO + p1 (At) + o(At) = 1 (i) Poisson Queues
(ii) Non-Poisson Queues
AXIOMS 3 - the probability that an arrival occurs between time t and time t + At
isAAt + o(At). This implies that
Definitions
Pi (At) = A At + o(At) Transient State: When a queuing system ahs its operating characteristic (e.g. input,
Where A, a constant, is independent o f /V(t), At is an incremental element, and output, mean queue length, etc) dependent upon time, then it is said to be in transient
o(A t) represents the terms such that state.
Ahtm- 0 —°(-A—t) = n0 Steady State: This is a queue system that is independent of time.At Assume Pn(t) to be the probability that there are n customers in the system at time i,
then the steady state use becomes
337
336
UNIVERSITY OF IBADAN LIBRARY
lim| -nP*n (/) = Pn (independent of t) This can be re-written asP.(i + A/) = PH( l \ 1 - AA/ + 0(At)][l - //At + 0(At]+ />„(r)[AA/][/zAt]+ P „ ^ h ^ x + ° 1
18.7 Poisson Queues
18.7.1 The M\M\1 System
Suppose n = 0, we have
This deals with the process where arrivals and departures occur randomly over time
P0{i + A/) = P0(t Jl - XAt + o(At)]+ Px (rXl - + °(A0]
generally known as birth-death process.
[//A/ + o(A/)] + o[A/]
1. Model 1: (M |A f|l):(o o |FIFO)
In this model, we have Poisson input, exponential service, single channel, infinite = P0(/J l-/lA /]+ P l(t)/^ r + o(A/)
system capacity and first in first out basis.
If P„(t), be the probability that there are ncustomers in the system at time t, then in We can record the difference equation
order to write the difference equation for P„{t), we first consider how the system can
P„ (/ + At) - P„(/) = - U + p)A tPn (/) + /iA/ Pntl (/) + XAt Pn, (/) +o(A/); n > 1
get to state En at time t + At. To be in state En of time / + A t , the system could have and
been in the state £nat time t and have no arrivals or service competitions in Ator be in P0 (/ + A/) •- P0 (l) = - XAtP0 (t)+ (iAt P, (/) + o (At)
state £n_i of time t and have, during A,, one service completion and no arrivals. If
Then,
we assume that n > 1 (having arrivals and service independent of each other), it
can be easily seen that lira P|' (' + ^ ' = - U + m ) P J ' ) + 0 ) + V>,„(/) + o(At)A t
a n d
Pn (t + At) = P„(t). P(no arrivals in At). P{no service completions in At)
lim - ^ , ( / ) + / iP ,( 0 + o(A t)
+Pn (0 - P{one arrival in At).P (one service in At) Ai—*0 A/
T^n + l (t). P(one service com pleted in At). P(_no arrivals in At)
+Pn-i( t) .P (o n e arriva l in A t).P (no service completions in At) + o(At) So that we have
n > 1 - P j n = P ( t ) = A A + p ) P jn + p P J l ) + J.P-i(t) n > I
dt
and
338 339
UNIVERSITY OF IBADAN LIBRARY
In general we have
~dt p„w = K ( ' ) = - K ( ' ) + mP M P. = 1 - 1 P, Vn
P.
The above are known as difference equations in n and t. The steady-state solutions Proof
for Pn in the system at an arbitrary point of time is obtained by taking the limit as By mathematical condition, we have
/ —► oG..
rp,, *\ n >1 r n ~ - pr n- \ »
If the steady-state exists (/l < /j, as t - » co), then P P
P„(t) —> P„ and Pn (/) —> 0 as t —»oo
A + p '
- i
P P \ P
I f A = p there exist no queue X -'+ fiX 1 A"
If — > I we have an explosive state
P ’
Using the condition of steady state,we have = - Pr
0 = -(A + ^ + / / P „ , + ^ _ , ; n > 1
9 .
Using the boundary condition; Z ^ . = 1. then 6.5 becomes
and
0 = -AP0 + fjPy >=Z P* = ^ o Z
n-o l/'J
Using iterate procedure we have
= P., Sum o f geometric series where — < 1
'pi ^ p' ii
P 1 - P
P
' A >
p2 = f x + » V - a k =
p J p = />
V - p J
f - 1 This implies that p ,=
p ; p l / ' J 1 - P
Resulting in the steady-state
P„ = p" (1 — p \ p < 1 and n >0
341
340
UNIVERSITY OF IBADAN LIBRARY
This is the probability distribution of queue length. (jii) Average queue length
Characteristics of Model 1
£ ( » . ) = I » p.;
(i) Probability of queue size greater or equal to n.
where m = n -1 (that is number of customers in queue
/>(* „ ) = ! > , . = z o - p ) p i minus customer in service)
A' *n K = - i ) P „ = i > r , , - l P ,
t=n r
= 0 - p ) p " t p ‘ ~ - 2
K - =( n-0
> p . -
i i i
= 0j ^ v = . * 7 ^ — [ i - ( i - p ) ]
1-P 1- P
(ii) Average number of customers in the system ■ r b - ’
. P 2
E ( n ) = Y j n PB = £ n (l-p )p " 1 - P
n=0 0) = ^ m- ■
V ' P(m > 0
££de 1 P
p - /1
= p ( i - p ) f S p - , Since /? < 1
P ( P - * )
d e ^
= P ( ' - p ) LO-p)-J This is because P (m > 0) = P(n > l) = ]jT P(1 - /}, - Px • L"»u
P (v) The fluctuation (variance) of queue length
1 - / 9 ^ - / l
T ( * ) = I« = o [ n - i r (n ) ] 2P„
= E^=on2P n - [ ^ ( n )]2
By algebraic transformations,
P(n) = ( l - p ) £ £ - [ £ f
_ P
( 1 -P )2
342 343
UNIVERSITY OF IBADAN LIBRARY
V'.-(O) :P(w = 0)
=
(n-W = P (No customer on the systemn upon arrival)
Example 18.1 To find y/uin for / > 0,' we suppose there be n customers in the system upon arrival,
A TV repairman finds that the time spent on his jobs has an exponential distribution l or a customer to go into service at time between 0 mid t, it means all the customers
with mean 30 minutes. If he repairs sets in the order in which they come in and if the
must have been served at time t.
arrival of sets is approximately Poisson with an average rate of 10 per day.
Therefore,
(i) What is the repairman’s expected idle time each day?
(ii) How many jobs are ahead o f the average set just brought in? t//i (,)=: p [(n -1) customers are served at timet) P [one customer being served in timedt]
Solution {idt
2 = — = — . setsperhour
8 ' 4 The waiting time w is therefore
w < t]
// = ^>V60 = 2 setsperfhour
(i) The probability o f no unit in the queue is asZ»t’i^ J o^ - ( / ) + V'-C«)
Po n 8 8
Hence the idle time for repairman in 8 hour days
= - , 0 = 3 hours „ z l ( « - i ;
8 i
= (l - p ) p - //t (l - /o)dt + (l - p)
E(n) = - V j o b s o(ii)
2 - /V 4 3 - \ - pe I> 0
18.7.2 Waiting Time Distribution for Model 1 The distribution of waiting time in queue is
Waiting time is mostly a continuous random variable and there is a non-zero I " P / = 0
probability of delay being zero. Denote time spent in queue by w. Let (/„.(/) be the ^ {,) = 1 - / » / > o
cumulative probability distribution so that from a complex randomness of the Poisson,
we have
345
344
UNIVERSITY OF IBADAN LIBRARY
Characteristics of Waiting Time Distribution for Model 1
( i) Average waiting time of a customer (in the queue) (iv) Average waiting time that a customer spends in the system including service
X
E(v) = | t.\//(wl w > 0)ir
£ ( h -) =
0 o
u>
= f tpp{\ - p)
o 0
I *
= -------[ x e's dx, for (// - X) = x
P - K
1
P _ A p - X
p ( l~ p ) p (p ~ X )
Relation between Average Queue Length and Average Waiting Time
(ii) Average waiting time of an arrival that has to want (Little’s Formula)
E (w /w > 0)= A2
p[w> 0) E(m) p(p-x)
E(w) = A £ (v )= —!—
p ( p - * ) \ / p p(p ~ x) P - X
1
It can be seen that E(n) = A E(v), E(n) = X E(w) and E(v) = E(w) + —
P ~ X P
Example 18.2
We note that P(w > 0) = 1 - P(w = 0)= 1 - (l - p ) = p Amvals at a telephone both are considered to be Poisson with an average time of 10
minutes between one arrival and the next. The length of a phone call is assumed to be
(iii) For the busy period distribution, suppose v is the random variable denoting the distributed exponentially with mean 3 minutes.
total time that a customer had to spend in the system including service. This makes (i) What is the probability that a person arriving at the booth will have to wait?
the cumulative density function to be (ii) The telephone department will initial a record booth when convinced that an
• arrival would expect waiting far at least 3 minutes for phone. By how much
v{w /w > 0) = — U ; where ^(w ) = [ipw (/)]
P[w > 0) at should the flow of arrivals increase in order to justify a record booth.
A /
l P ) / l / 'J
t .> 0
347
346
UNIVERSITY OF IBADAN LIBRARY
Solution
E(v) = \ E(n) = —
We are given A u-A.
This result applies to the FIFO SIRO and LIFO cases. These three queue discipline
^ = K) = 0, 1®Person Per minute sonly differ in the distribution of waiting time when the probabilities of along and
and short waiting times change depending upon the discipline used. When the waiting
time distribution is not required, the symbol GD(general discipline) can be used to
p = ̂ = 0.33 person per minute
represents the three queue disciplines above.
(/) P(w > o ) = l - / >u = l - f l - —
l P ) 18.7.4 Modellll (M |M |l):(A f|F /FO )
_ A _ 0.01 There is a deviation from the previous model 1 (especially 1) because the number of
~ M ~ 0.33 customers is now finite (W). As long as n < N, the difference equated o f model
= 0.33 remains valid for this model. If the system is in state Ew, then the probability of an
(ii) The installation of record booth will be justified if the arrival rate is greater arrival into the system is zero.
than the waiting time. Then the length of queue will go on increasing. Thus, the additional difference equation for n = N becomes
Now, E(w) = , ^— r = 3 p „ { t + a/ ) = p H ( i ) [i - M 'l+ ^ v - i ( 0 - M i - H + < > (a 0
MKM-A)
A1 resulting in the differential-difference equation.
0.33 (0.33-A1)
Where E(w) = 3 and A = A'(w) for record booth. On simplification this yields 4at p n( / ) = - / / P n C 0 + ^ ^ - , « )
A1 = 0.16. hence the arrival rate should become 0.16 person per minute to justifies the and gives the resultant steady state difference equation
record booth. 0 = - / i Pn + A P n, ( O
18.7.3 ModelII(Af |M |1): (oo|S //?0) Given the interval 1 < n < N -1 , the complete set of steady-state difference equations
This model is similar to model 1. The only difference is in the service discipline. The for this model is as follows.
first follow the FIFO rule, while this follows the SJRO rule. We recall that the /^ ,= A P 0
derivation of Pn for model I does not depend on any specific queue discipline, it may pP.,.i = (A + ai)P, - A P„ ,
then be concluded that for the SIRO rule case, we must have.
p„ =( ] - p) p " , n >0 P*3,. = A*\ ,
The average number of customer in the systemv£(n) remains the same irrespective of
cases, FIFO or SIRO. Provided P„ remains unchanged, £ (n ) remain the same in all
queue discipline, thus
348 349
UNIVERSITY OF IBADAN LIBRARY
\s in model I, by iterative procedure, the first two difference equations are
0 -/g )p "
P« = ( j j P „ '.n < N -l 1 „ N * I P * 1
; 0 < n < N
n (he same manner, the value of Pn holds for the last difference equation if n = N.
Thus, we have
N + 1 (p = 0
= p" P0; n < N Note that the steady-state solution exists even for p > 1. Intuitively, there is sense in
Using the boundary condition, we can obtain the value of P0. this since the process is prevented from blowing up by the maximum limit.
N Thus, given N ->■ co, the steady-state solution results in
Boundary condition is ^ P = P,
n=0 P„ = (l - p)p" n< co
Thus
Which is the same as that in model 1.
1 = ^. 2 > n
i - p > Characteristics of Model III
{ i - (i) Average number of customers in the system is given by p
/> (N + j) E(n) = Y inPil =P„YJnp"
n “II n “0
Thus,
t!>dt dp
\ - p N+l
1 - p * «
P0 = P'>P Tdp L i1 *
-P
N + \
( I - P ) :
Hence p [l-(N + l ) p N + N p H*'\
( M O V )
350 351
UNIVERSITY OF IBADAN LIBRARY
(ii) Average queue length
We know that Pn = p(> e", thus
£("0 = Z l ( ” _1) Pn = E (n) ~d=Y .P'■ n= I (fl) P, =(0.53) (0.5) = 0.27
= £ ( « ) - ( ! -P „ ) P2 = (0.53) (0:5)2 = 0.13
P3 = (0.53) (0.5)3 = 0.07
(b) E(n) = 1(0.27)+ 2(0.12)+ 3(0.07) = 0.74
_ p 2f l - A T p " - , + ( A f - l ) /) K l
V p ) ( i > ) Hence, the coverage number of trains in the queue is 0.74, and each train takes on an
(iii) Average waiting time. average 'A (0.085) hours for getting service. As the arrival of new train expects to
Using Little’s formula: find on average of 0.74 trains in the system before it.
E(w) = (0.74) (0.085) hours
E{v) = ~ ^ where A1 is the mean rate of customers entering the system and is equal
A = 0.0629 hours or 38 minutes
to a ( i -/> ,.)
18.7.5 Model IV (Birth- Death Process)
Thus, E(w) = E(y) - — =
P X Assume the system to be in date En, the probability of a birth occurring in a small
Example 18.3 time interval At is considered as AnAt + o(At); and that of the death is considered as
At a railway station, only one train is handled at a time. The railway yard is sufficient finAt + o(At),n > 1. The system being in En at time t means it will remain in En at
only for two trains to wait while the other is given signal to leave the station. Trains timet -F At provided there is no birth and no death/on birth and one death, or the
arrive at the station at an average rate of 6 per hour and the railway station can handle system might have been in E ^ a n d had a birth, or in En+1and had a death. Thus, this
them on an average of 12 per hours. Assuming Poisson arrivals and exponential result in
service distribution,
(a) Find the steady-state probabilities for the various numbers of trains in the Pn (t + At) = Pn (t) (1 - AnA t - o(At)){\ - iinAt - o(At) ) 4- Pn+1(t) (Mn+i^t
system. + o (A t) ) ( l - An+1At - o{At) + Pn- i ( t ) (An_aAt + o (A t))(l
(b) Also, find the average waiting time of a new train coming into the yard. - lin_xA t - o { A t ) + o(At), n > 1
Solution P0 (t + At) = P0( t ) ( l — A0A t- o(At) + P^O C /ijA t -H\o(At)) + o(At),
n = 0 \
2 = 6 / / = 12, p = — = 0.5
12 Dividing by At, and taking limit as At -*0, the diffential - difference equations results
Probability of no train in the system (both waiting and in service is
d
P„. = ——-^-7 = — = 0.53 “JT^nCO — ~ (A n + P)i)Pn(t)+ Pn + l ̂ >1+ 1 (0 T tl > 1
I - / / * 1 - (0.5)’*
352
353
UNIVERSITY OF IB
DAN LIBRARY
and By mathematical induction, one can prove that this formula is correct
"h Pn n ^n -l
d ^n+1 — P n - ' ^ 71-1Pn+1 Pn+1
d t P° ^ ~ ~*nPo(t) +
7 1 -1
Since Pn (t) is independent o f time, the steady-state solution =1 nipi-1=0 +i
^ P n (t) = 0 and the differential-difference equation reduce to
o = - (^n "h Pti ) + P ti+1 n+l + 1* H > 1 Making use of the boundary condition, we obtain PQPn P ^ n - lP n -
and 0. andpn = p f o m > 1
P2P1 °
then
A _= h ± P 2 n A1
.3
^2^ 1^0
Po = l - p
P3P2 Pi Thus
pn = pn( 1 - p), fo r n > 0
So that in general
'71 — ' 0 (same as model 1)n _ ^n-l^n -2 —^0 n
. PnPn- i ~Pi II. When Xnn = —n+l f’o r n > 0 ,
andpn = p fo rn > 1
Then
-1
= n,"=ro #-*7J+-1 7̂0 . » 2 1 V A" Po = 1 + Z-,1n ! p nn
355
354
UNIVERSITY OF IBADAN LIBRARY
It is given that the service increases with increase in the number of persons.
-1 Thus, pn = np. where there are n persons.
1 + p + | p 2 + j f />3 + - ]
= e~P X OP
Thus £(7.) = ^n =n0 p „ = n^=0n . ( - p " ) e -<’
Pn = { ^ . pn) e ~P f o r n ^ °
Here we can see that pn follows the Poisson distribution where p = - . But, p > 1 or e~p.p .e p = p
p < 1 most be finite. = 5/g persons
III. When An = Afo rn > 0 , andpn = np_ fo rn > 1 The average solving time is inversely proportional to the number of people solvingon
Then
1 the problem is given by day problem.i-iZ Xn Expected time for a person entering the line jVPo = + n!
71 = 1 ^ E(?T) = -l day or 8 hours.
= e~p
Thus Practice Questions
1. Derive, using both methods, the probability that a gambler will be ruined
p* m G s pn) e ~p f o r n ~ 0 given that his initial capital is Z.
Here, service rate increase with increase in queue length. Hence it is known as the 2. Show that gambler's expected gain is given asN (l — qz ) - N.
queuing problem with infinite number of channels= (M\M\co): (oo|F I F O ) 3. Under what condition can the expected gain be zero?
4. Company A enters into a project deal with another company B. 4 's initial
Example 18.4
deposit is /V577t, while f?’s initial deposit is NAm. For every success, A gains
Problems arrive at a computing center in a Poisson fashion at an average rate of five more naira from 6 , otherwise it loses same to B. If the probability of success
per day. The rules of the computing center are that any man waiting to get his is 0.7, what is the probability of losing the entire deal?
problem solved must aid the man whose problem is being solved. Tf the time to solve 5. A gambler's initial fortune is t. On each play of the game the gambler wins 1
a problem with one man has an exponential distribution with mean time of 'A day,
with probability p, or loses 1 with probability 1 — p. He or she continues
and if the average solving time is inversely proportional to the number people
playing until he/she is n ahead (that is, the fortune is t + ?t). or losing by m.
working on the problem, approximate the expected time in the center for a person
Here 0 < i - m and i + n < N. What is the probability that the gambler-quits
entering the line.
as a winner?
Solution
A = 5 problem per day, p = 3 problems per day
356
UNIVERSITY OF IBADAN LIBRARY
6 . Given an initial capital, Z, show that expected duration of the game is be the number of customer in the system (serviced and waiting) immediately
M after an event. Suppose that an event is equally likely to be an arrival or a N 1 - completed service.
q - p q - p Ml (a) State the transition graph and transition matrix and find the stationary distribution.
Describe the model 1 of the M|Af llqueue discipline, and show that (b) If a customer arrivers. what is the probability that he finds the system empty?
(a) the average number of customers in the system is given as Full?
^2 (c) If the system is empty, the time until it is empty again is called a “busy
(b) the average queue length is given as —.
period". During a busy period, what is the expected number of times that the
8. In the M |M |1 system of a queuing process, show that the system is full?
(a) steady state probability of model 1 is Pn = pn( l - p ) , where p < 1 and n > (d) Show that a limit distribution is a stationary distribution.
0.
(b) the waiting distribution is given as
( 1 - p . t = o
m =
[ l - p e - r t ' - P * , t > 0
9. SAO Super market has one cashier at its counter. The service discipline of the
cashier is FIFO. It is observed that the supermarket has 18 arrivals on average
of every 10 minutes while the cashier can serve 12 customers in 6 minutes. If
the distributions of arrivals and service rates arc poisson and exponential
respectively. Calculate.
(a) The traffic intensity and interpret the figure obtained
(b) The average number of customers in the system
(c) The average queue length
(d) The average time a customer spends in the system
(e) The average time a customer waits before being served
10. Customers arrive at an ATM where there is room for three customers to wait
in line. Customers arrive alone with probability and in pairs with
probability ^ (but only one can be served at a time). If both cannot join, they
both leaver call a completed services or an arrival an “event” and let the slate
358
UNIVERSITY OF IBADAN LIBRARY
R E F E R E N C E S
Amah in. u . N.(2()07). STA 211- Probability II. Ibadan Distance Learning Centre
Rosen. K. H., Michaels. J. G.. Gross. J. L.. Grossman. J. W. and Shier, D. R. (2000).
Series. Distance Learning Centre. University o f Ibadan, Handbook o f Discrete and Combinatorial Mathematics. CRC Press.
Alawode. (). A. andShittu 0 . 1. (2011). Probability III. An Unpublished lecture note Odeyinka. J. A. and Oscni, B. A. (2008). Basic Tools in Statistical Theory. Highland
developed lor the Ibadan Distance Learning Programme. Publishers.
Bhat, B. R. ( 1985). Modern Probability Theory: An Introductory Textbook. 2"'1 Olubusoye, O. F.. (2000). UnpublishedLec/u/v notes on STA 311.
Edition. Wiley Eastern Ltd.. Bombay. Roussas, G. G. (1973). A First Course in Mathematical Statistics. Addison-Wesley
Bhat U. N. (1971). Elements of Applied Stochastic Processes. John Wiley & Sons Publishing Company.
Ltd.. London. Shittu. O. I. (2011 ).UnpublishcdAG/c.v on Probability 11.
Attenborough. M. (2003). Mathematics for Electrical Engineering anti Computing. ii'ii'ir. wikinedia.ore
Nevvnes. Elsevier. Linacre House, Jordan Hill, Oxford. Encyclopaedia of Physics (2nd Edition), R.G. Lcmcr. G.L. Trigg. VHC publishers.
Brualdi, R. A. (1999). Introductory Combinatorics. Pearson Education Asia Limited 1991, ISBN (Verlagsgesellschaft) 3-527-26954-1. ISBN (VHC Inc.) 0-89573-752-3
and China Machine Press.
Bunneheka BMSG. Ekanayake CiEMUPD (2009) "A new point estimator for the Feller W. (1970). An Introduction to Probability theory and its applications. 3rJ
median of gamma distribution". Viyoihtyu J Science. 14:95-103 Edition. John Wiley and Sons Inc. New York.
Fisz M. (1963): Probability Theory and Mathematical Statistics. John Wiley and Sons,
Barry C. Arnold (1983). Pareto Distributions. International Co-operative Publishing New York.
I louse. ISBN 0-N9974-Q12-X.
Fisher Snedccor (19.38). The true characteristic function of the F distribution.
Dan Musa (2010): http://probabilityandstats.wordpress.com/2010/02/18/the- Biometrika. 69. 261-264.
malching-problem/ Ilori S. A. and Ajayi O. O. (2000). Algebra. University Mathematics Scries. Y-Books
Grimaldi, R. P. (1999). Discrete and Combinatorial Mathematics. Pearson Addison- (a
Wesley.
Gupta. B. D. (2001). Mathematical Physics. Vikas Publishing House PVT Ltd. Johnson NL, Kolz S. Balakrislinan N (1994) Continuous univariate distributions Vol
Hogg. R. V. and Craig. A. T. (1970). Introduction to Mathematical Statistics. New 1. Wiley Series in Probability and Statistics.
York: Macmillan Publishing Co.
Krishnamoorlhy, K. (2006). Handbook o f Statistical Distributions with Applications. Krishnamoorlhy K. (2006): Handbook of Statistical Distributions with Applications.
Chapman^ llall/C'RC. Chapman A-1 la 11/CRC
Merris. R. (2003). Combinatorics. Wiley-lntcrseienee.
Mood. A. M.. ( iraybill. F. A. and Boos, D. C. (1974). Introduction to theory o/ Lord, Nick. (July 2010). "Binomial averages when the mean is an integer", IJie
Statistics. McGrow-1 lill Inc. Mathematical Gazelle 94. 331-332. division of Associated Book Makers Nig. Ltd.)
3f> I
UNIVERSITY OF IBADAN LIBRARY
. Muiu.ll I (2(M)S).SVt///.v//a// Physics (2nd Edition), I . Mandl, Manchester Physics, John
Wiley & Sons. ISBN 97X0471915331 Ross S. (1988). A First Course in Probability. Macmillan Publishing Coy. New York.
Ross. Sheldon M. (2009). Introduction to probability and statistics fo r engineers and
Maxwell. J.C. (I860) Illustrations of the dynamical tneory ot gases. Philosophical scientists (4th ed.). Associated Press, p. 267. ISBN 97S-0-12-370483-2.
Magazine 19. 19-32 and Philosophical Magazine 20. 21-37.
Shangodoyin. D. K... Olubusoye. O. E., Shittu, O.l. and Adepoju. A. A. (2002).
Mood, A. M., Graybill, F. A. and Boes. D. C. (1963). Introduction to theory o f Statistical Theory and Methods. Joytal Printing Press, ISBN: 978-2906-23-9
Statistics. McGraw-Hill Books Coy. Swarup K... Gupta P.K.. and Mohan M. (1978): Operation Research. Sultcm Chant/and
Moners P. and Peres Y. (2008). Brownian Motion. An Unpublished lecture note on Sons. New Delhi. Second Ed. Reprinted.
the web.
Nadarajah, S. and Kotz, S (2006).The beta-exponential distribution. Reliability Udofia G. (1997): Lecture Series on Stochastic Processes. University' ofUyo, Uvo.
Engineering and System Safety, 91(1 ):689-697. Nigeria. (Unpublished)
Neumann. P. (1966). "Uber den Median der Binomial- and Poissonverteilung".
Hissenscluifiliche Zeitschrift tier Technischen Universitat Dresden (in German) 19: Udomboso. C. G. and Shittu O. 1. (2011). Stochastic Processes. A lecture note
29-33. developed for the Ibadan Distance Learning Programme.
Ugbebor. O. O. and Bassey U. K. (2003). Mathematics for Users. University
Odeyiiika J.A., and Oseni B.A. (2008): Basic Tools in Statistical Theory. Highland Mathematics Series. Y-Books (a division of Associated Book Makers Nig. Ltd.)
Publishers Ugbebor, O. O. (2010). Unpublished Lecture notes on Probability in Mathematics and
Statistics Departments, University of Ibadan.
Ollolsson P. (2005): Probability, Statistics and Stochastic Processes. John Wiley and Wadsworth, G. P. (1960). Introduction to probability and random variables. USA:
Sons, U.S.A. McGraw-Hill New York. p. 52.
Rogers L. C. G. and Williams D. (1994): Diffusions. Markov processes and
Martingles, Vol. I . second ed„ Wiley Series in Probability and Mathematical Young. G.A. and Smith R. L. (2005). Essentials of Statistical Inference. Cambridge
Statistics: Probability and Mathematical Statistics, John Wiley & Sons Ltd.. University Press
Chichester, 1994. Foundations.
Young, H.D and Freedman R.A. (2008): University Physics - With Modern Physics
Rogers L C. G. and Williams D. (1994): Diffusions, Markov processes and (12th Edition), (Original edition), Addison-Wesley (Pearson International). 1st
Martingles. Vol. 2, Cambridge Mathematical Library, Cambridge University Press. Edition: 1949, 12th Edition:, ISBN (10-) 0-321-50130-6, ISBN (13-) 978-0-321-
Cambridge. 2000, Ito calculus. Reprint of the second (1994) EDITION. 50130-1
Roussas G.G. ( 1973): A First Course in Mathematical Statistics. Addison-Wiley
Publishing Company
362 363
UNIVERSITY OF IBADAN LIBRARY