UNIT 2 APPLIED PROBABILITY II

Structure
2.0 " Objectives
2.1 Introduction
2.2 Working with Mean Deviation
2.2.1 Deviation below the Mean
2.2.2 Variance of a Binomial Distribution
2.3 Applications of Chebysliev's Theorem
2.3.1 Deviation of Repeated Ti+ials
2.3.2 Proof of the Weak Law
2.4 Random Walks and Gamblers' Ruin
2.4.1 The Probability Space
2.4.2 The Probability of Winning
2.4.3 A Recurrence for the Probability of Winning
2.4.4 Intuition
2.4.5 Duration of a Walk
2.4.6 Right Time to Quit
2.5 Chernoff Bounds
2.6 Strong Law of Large Numbers
2.7 Let Us Sum Up
2.8 Keywords
2.9 Some Useful Books
2.10 Answer or Hints to Check Your Progress
2.1 1 Exercises

After going through this unit, you will be able to:

r understand the degree of accuracy offered by an expected value of the
mean value:
compute the lower and upper bounds in which the expected value
varies; and
analyse problems with one-dimension random walk.

2.1 INTRODUCTION
In this unit we initiate preparatory wark for understanding thr5: basic tools used
in stochastic process while applying to finance and insurance. Basically themes
covered in the theory of probability have been extended tic, a state where
information a;ailability varies over time. Our main concern is to highlight the
results obtained in simple random walks, The following study material is
Quantitative Techniques adopted from the lecture notes of Prof. Albert Mayer and Dr. Radhika Nagpal
for Risk Analysis (2002) borrowing freely the theorems, definitions and examples without
mentioning the source. Our reliance on the above mentioned work is heavy,
indeed. Credit goes to the authors for presenting the theme for a non-technical
learner. We have included the select themes of the notes keeping in view the
overall requirement of the course. Perhaps it will be useful to go back to the
lecture notes for better appreciation of the themes dealt with.

2.2 WQRMNG WITH MEAN DEVIATION

We use' to take expectation of a random variable because it indicates the
"average value" the random variable will take.
It means, the average value from repeated experiments is likely to be close to
the expected value of one experiment. In addition it is more likely to be closer
as the number of experiments increases. This result was first formulated and
proved by Bernoulli.
Let us perform an experiment n times. We have G,, ...,G, independent random
variables with the same distribution and the same expectationp. If A, be the
average of the results, that is,

then by letting n grow large enough, we will be as certain as we want that tlie
average will be close to p . In other words, given any positive tolerance, E ,
A,, we will be within the tolerance of p as n grows. Effectively, we are
looking for the limit

According to Bernoulli, this limit equals one. This result is called the Weak
Law of Large Numbers. ~h'corern2.1 (Weak Law of Large Nupbers). Let
G,,...,G,,... be independent variables with the same distribution and the some
expectation, p. For any E > 0,

See, however, that the law, does not say anything about the rate at which the
, limit i s approached. That is, how big must n be to be within a given tolerance
of the expected value with a specific desired probability?
. In the following we will develop three basic results concerning the above
problem:

i) Markov's Theorem: It gives upper bound on the probability that the

value of the random variable is mare than a certain multiple of its mean.
Markov's result holds if we know nothing more than the value of the
mean of a random'variable. As such, it is very general, but also is much
 weaker than results which take more information about the random Applied Probn bility I I
variable into account.
ii) Chebyshev's Theorem: When we not only know the mean, but also
another numerical quantity called l:lie variance oi'ttie random variable, we
consider this theorem as it combines Markov's Theorem and information
about the variance to give more ret?ned bounds.
iii) Painvise Independent Sampling Theorern: 11 provides the adclitional
information about rate of convergence we need to calculate numerical
answers to questions siich as those above. Thc Sampling Theorem
follows fr40rn CIiebyshev'sTlieore~nand properLies 01' the variance of a
suln 01' independe11.t variables. Finally, the Weali I.aw of I,arge Numbers
will be seeti as a corollary of [lie Pairwise Independent Sampling
'Theorem.

Marltov's Tlieorern
When we want to consider 1.11~probleni ol' bounding ilic probahilitj because
the value of a random variable is far ilway l'rorn tlie mean. Mnrltov',s thcore~n
, gives a vcry rough estimate, bascd otily on the valilc of tlie mean.
Theorem 2.2 (Markov's Theorem). If'j'li is lr I I ~ I ~ I I L ~ S ( I ~ ~ r~r17(Iot11
I)LJ ~~~ri~lhle,
thcr7.fi,r. crll x > 0

Markov's '['heorcni is ofte~iexpressed in at1 nltcn~ativeI'orm; we state it as a

corollary. F

Corollary 2.1. If' R is GI non~c~crtive

rclngiom vat-icrhlc,the11,fi)t.(111 c' = 0

Examples of Markov's Tt~eorem

Mayer and Nagpal (2002) give the followillg two examples,
Suppose that N llieri fittending a dinner party deposit tlicir hats in tlic counter.
At the end of tlie night, eac'h is given back his own Iiilt with probability i / N .
What is the probability that x or more men get the right hat'?
Filst, we derive an upper bound with Marltov's 'Theorem. 1,et the randorn
variable, R , be the number 01' !lien tliat get tlie right list. By linearity of
expectation d [ ~= 1.] Using Msrkov's Theorem, the probability ttiat r or
more men get the right hat is:

, Consider the Chinese ,Appetizer problem where N people are eating Chinese
appetizers arranged on a circular, rotating tray. Someone then spins the tray so
that each person receives a randorn appetizer. We need to find the probability
that everyone gets the satile appetizer as before.
Quantitative Techniques
See that are N equally likely orientations for the tray after it stops spinning.
for Risk Analysis Therefore everyone is jikely getting the right appetizer in just one of these
a N
orientations. Therefore, probability is l / N .
Let us use the Markov theorem to get the probability. Let the random
variable, R , be the number of people who get the right appetizer. We know
that E [ R ] = 1. Applying Markov's Theorem, we find:

Markov's Theorem also gives the same 1/N bound for the probability
everyone gets their hat in the hatcheck problem. But in reality, the probability
of this event isl/N!. So Markov's Theorem i n this case gives probability
bounds that are way off.
Note that the proof of Markov's Theorem requires that the randoln variable, R , .
be nonnegative.
2.2.1 Deviation below the Mean
According Markov's Theorem, a random variable may not greatly exceed the
mean. There is another theorem which points out that a random variable is not
likely to be much slnaller than its mean.
. Theorem 2.2.1. Let L he a real number and let R be a random variable such
that R IL . For all x < L , we have:

Proof. The event that R 5 x is the same as the event that L - R 2 L - x.

Therefore,

(by Markov's Theorem)

Markov's Theorem can be applied since L - R is 'a nonnegative randoln

variable and L - x > 0,
Chebyshev's Theorem
We have seen above the Markov's Theorem for deviations above and below
the mean. Suppose we want bounds that apply in both directions, that is,,
bounds on the probability that J R- E [R]I
is large.
It is difficult to use Markov's inequality directly to bound the probability that
IR - E(1 11 L x , as we have to compute E [IR
- E [R]I] .
Since IRI and hence IRJ* are nonnegative variables For any R , Markov's
inequality also applies to the event [ I R2~x ~k ] .

Since such an event is equivalent to the event [IRI 2 x], we have

Corollary 2.2.1. For any random variable R , any positive integer k , and any Applied Probability 11
x>o,

Take k = 2 to get a special case of the above corollary. Then we can apply it to
bound the random variable, R - E [ R ], that measures R' s deviation from its
mean. Thus,

where the inequality (2) follows from Corollary 2.2..1 applied to the random
variable, R - E [ R ] . That is to say, we can bound the probability that the
random variable X deviates from its mean by Inore than x by an expression
decreasing as 1/x2 multiplied by the constant

Such a constant is variance of R .

Definition 2.2.1. The variance. Var [ R ], qf a random variable, R , is:

Accordingly we will rewrite (2) as Chebyshev theorem.

Theorem 2.2.2. (Chebyshev) Let R be a randon? variable, and let x be a
positive real number. Then

Example: Gambling Games

When we compare the following two gambling games, the usefulness of
variance can be easily judged.
Game A : We win $2 with probability 213 and lose $1 with probability 1/3.
GameB : We win $1002 with probability 213 and lose $2001 with probability
1/3.
Both the games have the same probability, 2 / 3 , of winning. However. &e will
see these to be very different, once we consider variance. Let random variables
A and B be the payoffs for the ,two games. For example, A is 2 with
probability 2/3 and -1 with probability l/3. The expected payoff for each
game is,
 Thus, the expected payoff is the same for both games. Look at the variance
Quantitative Techniques
for Risk Analysis however.
Let
1 with
A-E[A]
6
-2 with probability*

1 with probability3
( A- E [ A I ) ' 4 with probabilityf

var [ A ] = 2.

Similarly, for F'ar[~]:

B-E[B]

( B - E[B])'
= r 1001 with probability?
-2002 with probability+

1,002,001 with probability 5

4,008,004 with probability f

We may note that the variance of Game A is 2 whereas the variance of Game
B is more than two million. This means, the payoff in Game A remains closer
to the expected value of $1. On the contrary the payoff in Game B can deviate
very far from this expected value.
In economics high variance is often associated with high risk. For example, in
ten rounds of Game A, we expect to make $1 0, but could conceivably lose $10
instead. On the other hand, in ten rounds of game B , we also expect to make
$10, but could actually lose more than $20,000.
Theorem 2.2.3 (Pairwise Independent Additivity of Variance).
If R,, R,, ...,R,, are pairwise independent random variables, then
Vtzr[8,+R, +...+R,,]=v a r [ ~ , ] + v a r [+...+
~ , ] VW[R,,]

=
1sr+,i;n.
E [ R , ]E [R,,] +z
11

/=I
E [R,,] (painvise independence) ( 3 )
In 53), we use the fact that the expectation of the product o f two independent Applied Probability 111
variables is the product of their expectations.
Moreover,

Therefore we get

2.2.2 Variance of a Binomial Distribution

Take a random variable R . It has a binomial distribution, if

where n and p are parameters such that' n 2 1 and 0 < p < I .

If we take R as the sum of n independent Bernoulli variables, then we can
regard R as the number of heads that comes up when we toss n independent
coins, (where each head co~nesup with probabilityp). Thus, we can write
R = R , + & + ...+ Rn where

R, =( 1 with probabilityp,
0 with probability
. . 1 - p.
I

Quantitative Techniques
Then variance of the binomially distributed variable R is given by
for Risk Annlysis

'From the above result we see that tlie binomial distribution has variance
p(l - p ) n and standard deviation d m . It we consider a special case,
an unbiased binomial distribution with ( p = 1/2), then tlie variance is n/4 and
the standard deviation is &/2 .

2.3 A P PLIC A T I O N S OF CHEBYSHEV'S

THEOREM
Let us reformulate Chebyshev's Theorerii in ternis of standard deviation.
Corollary 5.1. Let R be a rand on^ varicrhle, crud let c he u positive real
number.

That is, the "likely" values of R are clustered in an ~(a,,)

-sized region
around E [ R ] . This result. therefore, confirms that the standard deviation
measures spread out the distribution of R is around its mean.
A One-Sided Bound
Chebyshev's Theorem gives a "two-sided bound". That is we gct the bounds
with tlie probability that a random variable deviates above or below i.he mean
by some amount. We develop one-sided bound for it below. For example, let
us find the probability or a random variable that deviates above the mean.
2.3.1 Deviation of Repeated Trials
Using Chebyshev's Theorem we will show how the average of many trials
approaches the mean in a large number of repeated trials.
Estimation from Repeated Trials
Let's take p as proportion of success of variable, a Bernoulli variable. G , with
G = 1 for success and G = 0 otherwise. In this case, (7 = G" so

and
To estimate p , we take a large number of trials, n , and record the fraction of Applied Probability II
success. Thus, we are taking Bernoulli variables G,, G,, ...,G, which are
independent each with the same expectation as G . Computing their sum

S,, = 1GI. l

, our estimate of p .
and then using the average, S , , / I I as
Speaking more generally, we may take any set of random variables
G,,G,,..., GI,,with the s a n ~ emean, p, and use the average, S,,/n, to estimate
p. A critical property of S,/n is that it has the same expectation as the ~ ~ ' s ,
namely,

Note that the random variables G, need not be either Bernoulli or independent
for (6) to hold.
To derive its variance, suppose that the G,'Salso have the same deviation, cr .
In that case, we get the second critical property of S,,/n, that is,

Thus, the variance af S, is the sum of the variances of the G,'s. For example,
by Theorem 2.3.1, the variances can be summed if the G,'S are pairwise
independent. Then we derive,

With this result, we apply Chebyshev 's Bound and conclude as follows:
Theorem 2.3.1 [Pairwise Independent Sampling]. Let

where G,,...,Gn are pairwise independent variables with the same mean, p,
and deviation, 0 .Then
 Note that Theorem 2.3.1 provides a precise statement about the average of
j Quantitative Techniques
for Risk Analysis * indebendent samples of a randoin variable approaching the mean. We .can
generalise it to many cases when S, is the sum of independent variables whose
mean and dev.iation are not necessarily all the same.
2.3.2 Proof of the Weak Law
We can state the conclusion of the Weak Law of Large Numbers, Theorem 2.1,
is that the probability average differing from the expectation by more than any
given tolerance approaches zero.
Theorem 2.3.2 [Weak Law of Large Numbers]. Let

where G,, .... Gn,... are pairwise independent variables with the same
expectation, ,LL and standard deviation, a . For any c;. > 0 ,

Proof. Choose x in Theorem 2.3.1 to be 6 .Then, given any 6 > 0 , choose n

large enough to make ( o / x 2 ) / n< 6. By Theorern 2.3.1,

So at the limit probability must equal zero.

Notice that this version of the Weak Law is slightly different from the version
we first stated in Theorem 2.1. Whereas Theorem 2.3.2 requires that the G,'S
lhave finite variance, Theorem 2.1 requires finite expectation only. Again, the
original version of Theorem 2.1 requires mutual independence, while Theorem
2.3.2 requires only pairwise independence.
Note that both the Weak Law and Pairwise Independence Sampling Theorem
2.3.1 do not provide any information about the way the average value of the
observations may be expected to oscillate in the course of repeated
experiments. We will see such a feature in Strong Law of Large Numbers
which deals with the oscillations. Such oscillations are critical in gambling
situations, where large oscillations can bankr~rpta player, even though the
player's average winnings are assured in the long run.

Ch~eckYour Progress 1
1) What is the result of Weak Law of Large Number?
2) Write the meaning of Applied Probability 11
(i) Markov's theorem and (ii) Cliebyshev's tlieorem.

....... ..........
A
I
r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3) How do you find two-sided bound given by Chybyshev's theorem?

4) Consider n independent Bernoillli trials with Probability of success ps.

Find the probability of no failures.

2.4 FUNDOM WALKS AND GAMBLER'SORUIN

Random walks model processes taking steps in a randomly chosen sequence of
directions, For example, in Computer Science, the Google search engine uses
random walks through the graph of world-wide web links and determines the
relative importance of websites, In case of finance, there are models which
consider one-dimensional random walks in the day-to-day fluctuations of
market prices. In the following we consider l-dimensional random walks, i.e., '
walks along a straight line.
:I\

t
./)
,. ,,
. ,
I .
Quantitative Techniques
h r Risk Analysis
, T=Y1+In , - - - - - - - - - - - - - - - - - - - - - ^ - - - - - - - - - - - - - - - - - -

, , , ,

. , ! ..

bet olttcomes:
WLLWLWWLLL

Source: Meyer and Nagpal (2002)

Figure 1: This is a graph of the gambler's capital versus time for one possible sequence of
bet outcornes. At each time step, the graph goes up with pl.obability p and down
with probability 1 - p . The gambler continues betting until the graph reaches
either 0 or T = 17 + ??I.
S~ipposethat a gambler starts with n dollars - his initial money (capital). He
makes a sequence of $1 bets. If he wins an individual bet, he gets his money
back plus another $ I . If he loses, he loses the $ 1 . In each bet, he wins with
probability p > 0 and loses with probability q = 1 - p > 0 . The ganlbler plays
until either he is bankrupt qr increases his capital to a goal amount of. T
dollars, If he reaches his goal, then he is called an overall winner, and his profit
will be m = T - 17 dollars. If his capital reaches zero dollars before reaching his
goal, then we say that he is "ruined".
Figure 1 sliows the ganibler's position as he proceeds with his $ 1 bets. The
randoni walk has boundaries at 0 and T . tf.the random walk ever reaches
either of these. boundary values, then it terminates. With these constraints, our
ob.jective is to determine the probability, v v , that the walk terminates at
boundary T . That is, the probability that the gambler is a winner.
It is easy to see that p = q = 112 in a fair gaine. In that case random walk is
called zmbiased. The ganibler is more Iiltely to will if p > 1 I 2 and lcss likely to
win if p < 112; these case's give the random walks which are bicrsetl.
Example I . Suppose that the gambler is flipping a coin, winning $ 1 on Heads
and losing $1 on Tails, Also, the gambler's starting capital is n = 500 dollars,
arid he wants to tilake m = 1 00 dollars. That is, lie plays until he goes broke or
reaches a goal of I' = n -t m = $600. What is the probability that he is a winner?
We will show that in this case the probability w = 516 . So his chances of
winning are really very good, namely, 5 chances out o f 6,
Let us suppose that the gambler chooses to play roulette in an American
casino, always betting $1 on red. A roulette wheel has 18 black nu1nbers,l8 red
nu~nbers,and 2 green numbers. In this game, the probability o r winning a
single bet is p = 18/38 cs 0.47 . There are two green numbers, which introduce
bias to the bets and give the casino an edge. However, the bets are allnost fair,
and we might expect that the gambler has a reasonable chance of reaching his
 -
goal - the 516 probability of winning ill the unbiased game surely gets Applied Probability 11
reduced, but perhaps not too drastically.
But our gness is not correct. His odds o r winning against the "sliglitly" ~ ~ n f a i r
roulette wlieel a r e less than I in 37,000. No riiatter how much money the
gambler has to start, e.g., $5000, $50,000, 01. $5.1 0", his odds are still less
than 1 in 37,000 of winning a mere 100 dollars. Perhaps it will be better to
suggest for not playing the game,
The theory of random walks draws a lot of sucli counler-intuitive conclusions.
2.4.1 The Probability Space
A game like tlie one described above follows in Figure I . It starts at the poi111
(!?,(I).
A winning path does not touch the .Y axis: I n fact it ends when it first
touches the line y = T .
In a similar logic, a losing path will not toi~chthe line y = T and will end
when it first toilches the x axis.
We can take the length k path, wliicli is chnracterised by tlie history of wins
and losses on individual $ 1 bets. So we use a length k string of W ' s and L' s
to nod el a path, assigning probability f q k - ' ' to a string that contains r. W's.
The outcomes in s ~ ~ ac sample
h space are the string that shows tlie winning or
losing wallts.
I11 an infinite wallcs in which the gali~blesplays forever, neither reaches his
goal nor goes bankrupt. In such a game the probability of playing forever is
I
zero, so we don 't need to include any sucli outcomes in our sarilple space.
2.4.2 The Probability of Winning
The UIIbiased Game
Cotisider the case of a fair coin, Sa the probability of winnilig p = 112. Let us
determine the probability, YV, that tlie gamblel. wins. 1.f the random variable G
is equal to the gambler's dollar gain, the G = T - 11 if the gambles wins, and
G' = -n if lie loses, Thus, tlie expected value of win is
E [ G ]= I ~ J ( T - n ) - ( I- w ) n = i,vT--n
Since the gambler wins or loses, the probability of losing is 1 - iu.
Let G, be the amount tlie gsmblcr gains on thc it11 flip: So, G,= 1 if tlie
gambler wins the flip, G,= -1 i f t l ? e garnbler loses thc flip. and (i,= O if'the
game has ended before the ith flip. As the coin is fair, E [ c ( ]= 0 .

Theorem 2.42. In the unbiased Gambler's Ruin game with probability

p = 1 / 2 of winning each individual bet, with initial capital, n , and goal, T ,
 n
Quantitative Techniques ~r {the gambler is a winner} = - (9)
for Risk Analysis T
'
Example 2. Suppose we have $1 00 and we start flipping a fair coin, betting $1
with the aim of winning $100. Then the probability of reaching the $200 goal
is 100 1200 =I 12 - the same as the probability of going bankrupt. In general,
if T = 2 n , then the probability of doubling your money or losing all your
mcney is the same.
Example 3. Suppose we have $500 and we start flipping a fair coin, betting $1
I with the aim of winning. $1 00.

Example 4. Suppose John starts with $100, and Sita starts with $10. They flip
a fair coin, and every time a head appears, John wins $I from Sita, and vice
versa for tails. They play this game until one person goes bankrupt. What is the
probability of John winning?
Frame this problem as is done in case of Gambler's Ruin problem with n=100
and T=100+10 =I 10. The probability of John winning is 10011 10=10111,
namely, the ratio of his wealth to the combined wealth. Sita's chances of
winning are 1 / I 1.
It is easy to see that although John will win most of the time, the game is still
fair. When John wins, he wins $10 only; when he loses, he loses big: $,loo.
The expected win of both players is zero.
Another .result of the analysis is: the larger the gambler's initial,stake, the
larger the probability that he will win a fixed amount.
Example 5. If the gambler started with one million dollars instead of 500, but
aimed to win the same 100 dollars as in the Example 3, the probability of
winning would increase to 1 M / ( IM + 100) > 0.9999,
2.4.3 - A Recurrence for the Probability of Winning
Let p and T be fixed, and let w,, be the gambler's probability of winning
when his initial capital is n dollars. For example, w, is the probability that the
gambler will win given that he starts off broke; clearly, w, = 0 . Similarly, for
d7.= 1 .
Suppose that, the gambler starts with n dollars, where 0 c n < T . Consider the
outcome of his first bet. The gambler wins the ficst bet with probability p . In
that case, he is Ieb with n t 1 dollars and becomes a winner with probability
w,,,, . On the other hand, he loses the first bet with probability 1 - p . Now he is
left with n - 1 dollars and becomes a winner with probability wn-,. Thus, he
wins with probability wn = pw,,,, +qw,-, . Solving for w,,,, we have
Homogeneous Linear Recurrence Applied Probability 11

The definition of a quantity wn+, in terms of a linear combination of values w,

for k < n + l is called a homogeneous linear recurrence. These can be solved
with the help of a general method. Suppose that the form of the solution is
w,, = c" for some c > 0 . It can be shown it is value of a constant c

From (1 0) we have

For the correct solution, this is equivalent to

Factoring out c"-' we get

From the above equation we get two roots, (1 - p ) / p and 1. If

wn = ((1 - P)/P)'' = ( q l ~ )5 " a

then (1 I), and hence (10) will be satisfied. Taking w,, = In condition (I 1) can
be satisfied, Since the left-hand side of ( I I ) is zero using either definition, it
follows that any definition of the form

will also satisfy (11). For the values of w, and w T ,the boundary co~iditions
we solve for A and B to gel

I
Thus, A t B=-A, (12)
( 4 / ~- )1 '~
and

The solution (1 3) only applies to biased walks, since we require p r q so the

denominator is not zero.
Thearcrn 2.4.3. In the biased Gambler's Ruin game with probability, p 7t 112
of winning each bet, with initial capital, n , and goal, T ,
Quantitative Techniques
~r { the gambler is winner) =
(Y 1 1 ) ~-
~1
for Risk Analysis
((P)'
I -1
To appreciate the idea of the probability that the gambler wins in the biased
game, suppose that y < 112; that is, the game is biased against the gambler.
Then both the numerator and denominator in the quotient in (14) are positive
and the quotient is less than one. So adding 1 to both the numerator and
denominator increases the quotient, and the bound (14) simplifies to

Corollary 10.7. In the Gambler's Ruin game biased against tlie gambler. that
is, with probability y < 112 of winning each bet, with initial capital, n , and
goal, T ,

~ r ( t l i gambler

where m = T - n ,
e is a winner} <
(9 n'
,

the gambler's intended profif. Tlius, the gambler gains his intended profit, m ,
before going broke with probability at most ( P / q ) " J . In this result, the upper
bo~lnd does not depend on the gambler's starting capital. To see the
consequences of thsis, we give tlie following examples:
Example 6. Suppose that the gambler starts with $500 aiming to profit $100,
this time by making $1 bets on red in roulette. By (15), the probability, 1vJl,
that he is a winner is less than

This is in sharp contrast to the unbiased game, where we saw (in Example 3)
that his probability of winning was 516.
Example 7. Consider Example 3 and the capital to be $1,000,000 to start in the
unbiased game. The ganible was alniost certain to win $100. But betting
against the "slightly" unfair roulette wheel, even starting with $1,000,000, his
chance of winning $100 remains less than 1 in 37,648. He will alnlost surely
lose all his $1,000,000. In fact, because the bound (15) depends only on his
intended profit, his chance of going up a mere $100 is less than I in 37,648 no
mutjer how much money he starts with.
9
The bound (1 5) is exponential in m . So, for example, doubling his intended
profit will square his probability of winning.
Example 8. Suppose that ganibler's stake goes up ,200 dollars in the above
example. What is probability that he wins before going broke in the roulette
game?

wliich is about 1 in '70 billion.

Example 9. Apply the formula (14) and find that the probability of winning Applied Probability II
$1 0 before losing $1 0

See that this is somewhat worse than the 1 in 2 chance in the fair game.
What is lesson of fair and unfair game's formulation?
In case of a fair game, it helps if we to have a large capital, whereas in the
unfair case, it doesn't.
2.4.4 Intuition
It will be useful to see the intuition behind the gambler's failitre to makc
money when the game is slightly biased against him. Seen in terms of roulette
game, the gambler wins a dollar with probability 9/19 and loses a dollar with
probability10119.
Therefore, his expected return on each bet is
9119-10/19=-l/19x-0.053 dollars.
hat is, on each bet: his capital is expected to drift downward by n little over 5
cents.
If the gambler starts with a trillion ciollar~then he will play for a very long
time. So at sorile point there would be a lucky outcome. Perhaps an upward
swing will put him $100 ahead. However, there is a problem in this
formulation. His capital drifts downward steadily. IFtlie gambler does not have
luck for an upward swing early, then lie is doomed. The logic is simple: After
his capital drifts downward a few hundred dollars, lie weds a huge upward
swing to save himself', Perl~aps,silch a huge swing is extremely unlikely.
To quantify these drifts and swings, consider the position after k rounds, Let
the nuriiber of wills by tlic player be a binomial distribution with parameters
.
p < 1 12 and k Then his expected win on any single bet is p - q = 2 p - 1
do1 lars. Thus. his expected capital is n - k ( 1 - 2~2).
The actual number of wins
must exceed the expected number by nl-I-k(1-217) to be a winner. Since the
binomial distribution has a standard deviation or only J G j , for the
gambler to win, he needs his number of wins to deviate by

times its standard deviation, which is ~ ~ ~ ~ l idire

k e ltoy tail position of binomial
distribution.
In contrast, a fair game does not witness drift; swings are the only effect, Dire
to this, our earlier result holds good. Therefore, if the gambler starts with a
trillion dollars, then alrnost certainly there will be a lucky swing that puts him
$100 ahead.
 I Quantitative Techniques
for Risk A~ralysis The discussion above pointed out that the gambler would be a winner in both
fair and unfair games. Let us enquire, how many bets he needs on average to
either win or go broke.
Biased Walk Case
Let Q be the number of bets the gambler makes until the game ends. Since the
gambler's expected win on any bet is 2 p - 1, Wald's Theorem (seb :\eppendix
given at the end) should tell us that his game winnings, G , will have
expectation E [Q]( 2 p- 1) . Thus,

E[G] = ( 2 p - l ) ~ [ Q ] (16)
In an unbiased game (16) is a trivial result as both 2y - 1 and the expected
overall winnings, E [GI,
are zero. In the unfair case, however, 2 p - 1 # 0 .

I Moreover,

So assuming (1 6), we say:

Theorem 2.4.5. In the biased Gambler's Ruin game with initial capital, n ,
goal, T , and probability, p $ 1 / 2 oFwinning each bet,
~r {gambler is a winner] T - n
E [nober of bets till game ends] = (17)
2p-I

Since G = xR !=I
G, is not a sum of nonnegative variables we do not have a
special case of Wald's theorem and if the gambler loses the ith bet, then
random variable GI equals 1. To confirm that (16) applies in general cases,
take the random variable GI+ 1 is nonnegative, and *

E [ G , + ~ ~ Q > ~ ] = El[~Gk, i ] + 1 = 2 ~ ,
so by Wald's Theorem

L 1-1 _I

= E [ G ] +E[Q] (1 9)
Combining (18) and (19) we get the result implied by (16).
 Example 10. If the gambler aims to profit $100 playing roulette with n dollars Probability 11
~pplitd
to start, he can expect to male ((n + 100) 137,648- n)1(2(18/38)- 1) 1 19n
bets before the game ends. So he can enjoy playing for a-good while before
almost surely going broke.
Duration of an Unbiased Walk I
Consider the expected number of bets as a function of the gambler's initial
capital. That is, for fixed p and T , let e,, be the expected fiumber of bets until
the game ends when the gambler's initial capital is n dollars. Since the game
is over in no steps if n = 0 or T , the corresponding boundary conditions are
eo = e,. = 0.

Alternatively, let the gambler start with n dollars, where 0 ir; < T. By the
conditional expectation, the expected number of steps can be broken down. So
that,
e,, = p ~ I gambler
[ ~wins first bet] + q~ [ Q I gambler loses first bet] .
When the gambler wins the first bet, his capital is n + 1, so he can expect to
I
make another e,,,, beti. That is,
E [ Q I gambler wins first bet] = I + e,,,, .
and similarly,
I
E [ Q I gambler loses first bet] = I + e,,-,
So we have
. II
which yields the linear recurrence

For p = q = 112, this equation sinlplifies to

e,,,, = 2e, -2e,, -en-, - 2 . (20)
There is a general theory for solving linear recurrences like (20) in which the
value at n + l is a linear combinati alues at some arguments k < n + 1
plus another simple term - in this lus the constant - 2. This theory
implies that
e, =(T -n)n. (21)
Equation (21) can be verified easily from the boundary condition along with
(20). Using strong inductic~rion n , we have,
Theorem 2.4,5.1. In the unbiased Gambler's Ruin game with initial capital, n ,
and goal, T , and probability,
E [number of bets till game ends] = n ( -~n) (22)
or, n number of bets till game ends] = initi&capital intended profit. (23)
Quantitative Techniques
To see the result, consider that if the gambler starts with $1 0 dollars and plays
for Risk Analysis until he is broke or ahead $10, then 10 10 =I00 bets are required on average.
lf he starts with $500 and plays until he is broke or ahead $100, then the
expected number of bets until the game is over is 500 x 100 =50,000.
2.4.6 Right Time to Quit
Let the gambler start with n > 0 dollars, without a target T , and let him play
until going broke. Then we find that if the game is not favourable, i.e.,
p 5 1/2, the gambler is sure to go broke. In particular, he is even sure to go
broke in a "fair" game with p = 112.
Lemma 2.4.6.1. If the gambler starts with one or more dollars and plays a fair
game until he is broke, then he will go broke with probability 1.
However, if the game is fair, he can "expect" to play for a very long time
before going broke; in fact, he can expect to play forever.
Lemma 2.4.6.2. If the gambler starts with one or more dollars and plays a fair
game until he goes broke, then his expected number of plays is infinite.
A game without bounds will have a larger expected number of bets compared
to the bourded one because, in addition to the possibility that the gambler goes
broke, in the bounded game there is also the possibility that the game will end
when the gambler reachei his goal, T . That is,

From (21) we have

u,, 2 n ( -~n)
If n 2 ( T -n), and T t n then lower bound on u,, can be arbitrarily large,
which implies u, must be infinity.
Take Lemma 2.4.6.1, along with a probability.1. Then the unbounded game
ends when the gambler goes broke. Therefore, the expected time for the
uhbounded game to end is the same as the expected time for the gambler to go
broke and the expected time to go broke is infinite.
Probability of at least one event
Theorem 2.4.6. Let A,, A,,,.. A,, be independent events, and let T be the
number of these events that occur. The probability that none of the events
occur is at most e-EITI.
, , T = ) ; + T , + ..AT,, (26)
where I: is the indicator variable for the event A,. We also use the fact that '

for all x , which follows from the Taylor expansion

Two useful special cases of Theorem 2.4.6 are given in the following:
Corollary 15.2. Suppose an event has probqbility l/m. Then the probability Apfilied ~robabilit~"11
that the event will occur at least once in m independent trials is at least
approximately 1. l/e = 63% . There is at least 50% chance the evekwill occur
--
in n = m log 2 = 0.69m trials.

Check Your Progress 2

I ) What is gambler's ruin problem?

2) What is the probability of a gambler winning in a fair coin toss game?

..........................................................................................
.........................................................................................
.........................................................................................
.........................................................................................
.........................................................................................
.........................................................................................
3) Write the meaning of recurrence for the probability of winning.

2.5 CHERNOFF BOUNDS

Probability of at least k Events
We generalise the above theorem from probability that one event occurs, to
probability that k events occur, assuming mutual independence of the events
A,. In other words, we find the probability Pr { T 2 k), given that the events A,
are mutually independent.
Consider two examples, which describe different Scenario.
Example A . Suppose we want to know the probability that at least k heads
come up in N tosses of a coin. Here A, is the event that the coin is heads on
the i th toss, T is the total number of heads,.tind P~(T
2 k} is the probability
that at least k heads come up.
 I
I

< I
Quantitative Techniques
Example B . Suppose that we want the probability of a student answering at
I
for Risk Analysis least k questions .correctly on an exam with N questions. In this case, A, is
the event that the student answers the i th question correctly, T is the total
\ number of questions answered correctly, and ~r{ T 2 k} is the probability that
. k questions correctly.
the student answers at least
The difference between these two examples: In the first, all events A, have
equal probability, i.e., the coin is as likely to come up heads on one flip as on
another. So T has a binomial distribution whose tail bounds we have already
characterised above.
On the other hand we have a situation where, some examination questions
might be more difficult than others. If Question 1 is easier than Question 2,
then the probability of event A, is greater than the probability of event A,.
In following we develop a method to handle this moie general situation in
which the events A, may have different probabilities.
Taking Chernoff bound we will show that the number of events that occur is
almost never much greater. than the expectation. For example, if we toss N
coins, the expected number of heads is N/2 heads. The Chernoff Bound
implies that for sufficiently large N , the number of heads is almost always not
much greater than N/2.
Statement of the.Bound
Chernoff s Theorem can be stated in terms of Bernoulli variables instead of
events. However, we oan take I; as an indicator for the event A,.
Theorem 2.5. (Chernoff Bound). Let I;, T,, ...,T, be mutually independent
Bernoulli variables, and let
T = I ; + T , + ...+T,.
Then for all c 2 1, we have

The formula for the exponent in the bound'is a little awkward. The situation is
simpler when c = e = 2.718... 111 this case,

so we have as an immediate corollary of Theorem 2.5:

~orolltky15.2. Let I;,T,,...,qbe mutually independent Bernoulli variables,
and let T=T,+T,+ ...+T,. Then

2.6 STRONG LAW OF LARGE NUMBERS

We'described the Weak Law of Large Numbers above and took up the question
of what was tfie strong law of large numbers we might prove. Roughly
speaking, the strong law says that with probability 1, the bound of the weak
law will hold for all but a finite number of the S,, simultaneously-there will Applied Probability I1
only be finitely many exceptions to it.
'Theorem 2.6. [The Strong Law of Large Numbers]
x"
Let S, = I = I X , where X , ,....Xi ....are mutually indepenbent, identically
distributed random variables with finite expectation, p .'Then .- y

Although Theorem 2.6 can be proven without this assumption, we will assume
for simplicity that the random variables Xi have a finite fourth moment. That
is, we will suppose that
B [ x > ]= K < m. (29)

Check Your Progress 3

1) , What do you mean by symmetric random walk?

2) If a gambler is playing with the aim of winning a fixed amount, what are
the chances of his success?

3) Suppose an evelit has probability of -.m1 Find the probability that the event
will occur at least once in independent trials.
Quantitative Techniques 4) Take T,,T,, ...q,to be mutually independent Bernou Ili variables and let
for Risk Analysis T=T,+T,+ ...+ T,. So that
~r{(T 2 CE[ T ] ) )5 exp(-(c log c - c + I) E [TI)

5) Write the meaning of Strong Law of Large Numbers.

", .........................................................................................
.........................................................................................
.........................................................................................
.........................................................................................

2.7 LET US SUM UP

In thii h i t w e have discussed the probabilistic idea that the average of a large
number of trials is likely to be within an acceptable level of the expectation.
For that purpose we have examined the results of Markov's theorem,
Chebychev's theorem and the pairwise independent sampling theorem.
Building on such findings we extend the coverage to include random walk and
discussed the gambler's ruin problem with the different scenarios of initial
capital as well as conditions of variance.

' 2.8 KEY WORDS - - - - - --

Chernoff Bound: A lower bound for the success of majority agreement for n
independent, equally likely events. The Chernoff bound is a special case of
Chernoff s inequality.
Cliernoff's inequality: It states the fol,lowing:
Let ...X,
X,X,,
. be independent random variables, such that
< .

E [ x , ] = o and Ix,I51forall i .
Gambler's Ruin: A 'gamblerts loss of the last of her unit of gambling money
and coukquent inability to continue gambling. It is is also sometimes used to
refer to, a final large losin bet placed in the hopes of winning back all the
U' gambler has lost during a gambling session. More generally, the phrase refers
to the ever decreasing expected value of a gambler's initial capital as she Applied
continues to gamble with her winnings.
Markov Property: A property of a stochastic process in which the past and
future are independent.
Random Walk: Taking successive steps, each in a random direction.
Recurrence: A relation in an equation which defines a sequence recursively,
i.e., each term of the sequence is defined as a function of the preceding terms.
Stopping Time: With respect to a sequence of random variables XI,Xz,... a
random variable z with the property that for each t, the occurrence or non-
occurrence of the event z = t depends only on the values of XI, Xz,..., 4.'Thus,
at any particular time t, you can look at the sequence so far and tell if it is time
to stop.
Strong Law of Large Numbers: As the sample size grows larger, the
probabiIity that the sample mean and the population mean will be exactly equal
and approaches 1.
Weak Law of Large Numbers: As the sample size grows larger, the
difference between the sample mean and the population mean will approach
zero.

2.9 SOME USEFUL BOOKS

Meyer, A, m d Radhika Nagpal (2002), Mathematics for Con~puterScience
(course Notes 13-14), MIT, Cambridge, MA (see internet)
Ross, Sheldon, (1996), Stochastic Processes, Second Edition, John Wiley &
Sons, Inc, New York

2.10 ANSWER OR HINTS TO CHECK YOUR,

PROGRESS
Check Your Progress 1
1) See Section 2.3 and answer.
2) See Section 2.3 and answer.
3) See Section 2.3 and answer.
4) Hint: Take Pr ( F= 0) = p,:' and proceed to find
P ~ { F51) = p : +n(l-p,v)p.:.'-'.Get

Check Your Progress 2

1) See Section 2.4 cu~danswer ,

2) See Sub-section 2.4.2 and answer

3) See Sub-section 2.4,3 and answer
Quantitative Techniques Check Your Progress 3
For Risk Analysis
1) Read Section 2.4 and answer
2) Hint: Always succeeds in the'long run
1 = 63%. So there is a 50% chance the event will occur in
Hint: Get 1- -
3)
e
n = log 2m 0.69m trials
4) Hint: Find Chernoff .Bound
5) See Section 2.6 and answer

2.1 1 EXERCISES
1) Suppose you roll n dice that are 6-sided, fair and mutually independent.
What is the expected value of the largest number that comes up?
1" + 2" + 3" + 4".c5"
Ans. 6--
6" I

v,

2) In casino following game is p1ayed:;The gambler bets a fixed wager and

the owner of casino flips a fair coin until it comes of heads. The gambler
receives $1 if the coin turns up heads the first.time, $2 if it turns up the
first head at second toss. Let the owner tosses the coin :k times, and
gambler gets $2"' to get the first head.
i) What is the expected amount of money that the gambler will win
with a fixed wager of $1 O?
Ans. oo amount
ii) What is the probability that the gambler does not lose money in a
game when fixed wagei is $1 0,000?
1
Ans. -
16,384
3) Suppose that all dice are fair and'six-sided and that dice rolls are mutually
independent. Find the expected value for each of the following random
variables.
21
i) Sum of the rolls of three dice. (Ans. -)
2
343
ii) Product of the rolls of three dice (Ans. -)
8
4) Two gambling games have the following patterns of bets .

Game 1: You win $2 with probability 213 and lose $1 with probability
113.
Game 2: You win $1002 with probability 213 and lose $2001 with
probability 113. \
i) What is the ekpected win in each case? (Ans. 1)
ii) What is the variance in each case (Ans. 2)
5) Suppose you have learned tat the average MEC students total number of Applied Probabiilty II
marks 200.
a) Use ~ a r k o v ' sinequality to find a best possible. upper
-
bound for the
200
fraction of MCE students with at least 235 marks. (Ans. -)
235
b) Suppos,e you are tald that no student can pass with less than 170
marks. How does this help to improve your previous bound? Show
30
that this is the best possible bound. (Ans. -)
65
C) Suppose that you further learn that the standard deviation of the total
marks per studdnt is 7. Give a best possible bound on the fractian of
I
students who can graduate with at least 235 marks. (Ans. -)
25
6) Let X, Y be independent Binomial random variables with parameters
( n ,J J ) and ( m , F ) , respectively. What is ~r { x + Y = k) ?
 -
Quantitrrive Techniques Appendix to Unit 2
for Risk Analysis Wald Theorem
Let C,,C,, ..., be a sequence of nonnegative random variables, and let Q be a
positive interger-valued random variable, all with finite expectations, suppose
that

for some ,u E R and for all i 2 1. Then