Actuarial Mathematics

Lecture Notes
on
Actuarial Mathematics
Jerry Alan Veeh
May 9, 2003
Copyright  2003 Jerry Alan Veeh. All rights reserved.

§0. Introduction
The objective of these notes is to present the basic aspects of the theory of
insurance, concentrating on the part of this theory related to life insurance. An
understanding of the basic principles underlying this part of the subject will form a
solid foundation for further study of the theory in a more general setting.
Throughout these notes are various exercises and problems. The reader should
attempt to work all of these.
A calculator, such as the one allowed on the Society of Actuaries examinations,

will be useful in solving some of the problems here. The problems contained here
are not all amenable to solution using only this simple calculator. A computer
equipped with spreadsheet software will sometimes be useful, especially for the
laboratory exercises.

§1. Overview
The central theme of these notes is embodied in the question, “What is the value
today of a random sum of money which will be paid at a random time in the future?”
Such a random payment is called a contingent payment.
The theory of insurance can be viewed as the theory of contingent payments.

The insurance company makes payments to its insureds contingent upon the oc-
currence of some event, such as the death of the insured, an auto accident by an
insured, and so on. The insured makes premium payments to the insurance company
contingent upon being alive, having sufficient funds, and so on. A natural way to
model these contingencies mathematically is to use probability theory. Probabilistic
considerations will, therefore, play an important role in the discussion that follows.
The other central consideration in the theory of insurance is the time value of
money. Both claims and premium payments occur at various, possibly random,
points of time in the future. Since the value of a sum of money depends on the
point in time at which the funds are available, a method of comparing the value
of sums of money which become available at different points of time is needed.
This methodology is provided by the theory of interest. The theory of interest will
be studied first in a non-random setting in which all payments are assumed to be
sure to be made. Then the theory will be developed in a random environment, and
will be seen to provide a complete framework for the understanding of contingent
payments.

§2. Elements of the Theory of Interest
A typical part of most insurance contracts is that the insured pays the insurer
a fixed premium on a periodic (usually annual or semi–annual) basis. Money has
time value, that is, $1 in hand today is more valuable than $1 to be received one year
hence. A careful analysis of insurance problems must take this effect into account.
The purpose of this section is to examine the basic aspects of the theory of interest.
A thorough understanding of the concepts discussed here is essential.
To begin, remember the way in which compound interest works. Suppose an

amount A is invested at interest rate i per year and this interest is compounded
annually. After n years the amount will be A(1 + i)n . The factor (1 + i)n is sometimes
called the accumulation factor. If interest is compounded daily after the same n
i 365n
years the amount will be A(1 + 365 ) . In this discussion the interest rate i is called
the nominal annual rate of interest. The effective rate of interest in the example
) − 1. This is the rate of interest
i 365
in which interest is compounded daily is (1 + 365
which compounded annually would provide the same return.
Exercise 2–1. What is the effective rate of interest corresponding to an interest rate
of 5% compounded quarterly?
It is possible that two different investment schemes with two different nominal
annual rates of interest may in fact be equivalent, that is, may have equal dollar
value at any fixed date in the future. This possibility is illustrated by means of an
example.
Example 2–1. Suppose I have the opportunity to invest $1 in Bank A which pays
5% interest compounded monthly. What interest rate does Bank B have to pay,
compounded daily, to provide an equivalent investment? At any time t in years the
12t 365t
.05
amount in the two banks is given by 1 + 012 i
and 1 + 365 respectively. It
is now an easy exercise to find the nominal interest rate i which makes these two
functions equal.
Exercise 2–2. Find the interest rate i. What is the effective rate of interest?
Situations in which interest is compounded more often than annually will arise
frequently. Some notation will be needed to discuss these situations conveniently.
Denote by i(m) the nominal annual interest rate compounded m times per year which
is equivalent to the interest rate i compounded annually. This means that
m
i(m)
1+ = 1 + i.
m
Exercise 2–3. Compute 0.05(12) .

§2: Elements of the Theory of Interest 5
An important abstraction of the idea of compound interest is the idea of con-
tinuous compounding.
If interest
nt
is compounded n times per year the amount after
t years is given by 1 + n . Letting n → ∞ in this expression produces eit , and
i
this corresponds to the notion of instantaneous compounding of interest. In this

context denote by δ the rate of instantaneous compounding which is equivalent
to interest rate i. Here δ is called the force of interest. The force of interest is
extremely important from a theoretical standpoint and also provides some useful
quick approximations.
Exercise 2–4. Show that δ = log(1 + i).
Exercise 2–5. Find the force of interest which is equivalent to 5% compounded

daily.
The converse of the problem of finding the amount after n years at compound
interest is as follows. Suppose the objective is to have an amount A n years hence.
If money can be invested at interest rate i, how much should be deposited today
in order to achieve this objective? It is readily seen that the amount required is
A(1 + i)−n . This quantity is called the present value of A. The factor (1 + i)−1 is
often called the discount factor and is denoted by v.
Example 2–2. Suppose the annual interest rate is 5%. What is the present value
of a payment of $2000 payable in the year 2014? The present value (in 2004) is
$2000(1 + 0.05)−10 = $1227.83.
The notion of present value is used to move payments of money through time in
order to simplify the analysis of a complex sequence of payments. In the simple case
of the last example the important idea is this. Suppose you were given the following
choice. You may either receive $1227.83 today or you may receive $2000 in the
year 2014. If you can earn 5% on your money (compounded annually) you should
be indifferent between these two choices. Under the assumption of an interest rate
of 5%, the payment of $2000 in 2014 can be replaced by a payment of $1227.83
today. Thus the payment of $2000 can be moved through time using the idea of
present value. A visual aid that is often used is that of a time diagram which shows
the time and amounts that are paid. Under the assumption of an interest rate of 5%,
the following two diagrams are equivalent.
Two Equivalent Cash Flows
$2000 $1227.83
....................................................................................................................................................... .......................................................................................................................................................
2004 2014 2004 2014
The advantage of moving amounts of money through time is that once all
amounts are paid at the same point in time, the most favorable option is readily
apparent.
Exercise 2–6. What happens in comparing these cash flows if the interest rate is
6% rather than 5%?
In an interest payment setting, the payment of interest of i at the end of the period
is equivalent to the payment of d at the beginning of the period. Such a payment
at the beginning of a period is called a discount. What relationship between i and
d must hold for a discount payment to be equivalent to the interest payment? The
time diagram is as follows.
Equivalence of Interest and Discount
i d
....................................................................................................................................................... .......................................................................................................................................................
0 1 0 1
The relationship is d = iv follows by moving the interest payment back in time

to the equivalent payment of iv at time 0.
Exercise 2–7. Denote by d (m) the rate of discount payable m times per year that is
equivalent to a nominal annual rate of interest i. What is the relationship between
d (m) and i? Between d (m) and i(m) ? Hint: Draw the time diagram illustrating the two
payments made at time 0 and 1/ m.
Exercise 2–8. Treasury bills (United States debt obligations) pay discount rather
than interest. At a recent sale the discount rate for a 3 month bill was 5%. What is
the equivalent rate of interest?
The notation and the relationships thus far are summarized in the string of
equalities m −m
i(m) d (m)
1+i= 1+ = 1− = v−1 = eδ .
m m
Problems
Problem 2–1. Show that if i > 0 then
d < d (2) < d (3) < ⋅ ⋅ ⋅ < δ < ⋅ ⋅ ⋅ < i(3) < i(2) < i.
Problem 2–2. Show that limm→∞ d (m) = limm→∞ i(m) = δ .
Problem 2–3. Calculate the nominal rate of interest convertible once every 4 years
that is equivalent to a nominal rate of discount convertible quarterly.
Problem 2–4. Interest rates are not always the same throughout time. In theoretical
studies such scenarios are usually modelled by allowing the force of interest to
depend on time. Consider the situation in which $1 is invested at time 0 in an
account which pays interest at a constant force of interest δ . What is the amount
A(t) in the account at time t? What is the relationship between A′ (t) and A(t)? More
generally, suppose the force of interest at time t is δ (t). Argue that A′ (t) = δ (t)A(t),
and solve this equation to find an explicit formula for A(t) in terms of δ (t) alone.
Problem 2–5. Show that d = 1 − v. Is there a similar equation involving d(m) ?
Problem 2–6. Show that d = iv. Is there a similar equation involving d(m) and i(m) ?
Solutions to Problems
Problem 2–1. An analytic argument is possible directly from the formulas.
For example, (1 + i(m) / m)m = 1 + i = eδ so i(m) = m(eδ / m − 1). Consider m as a
continuous variable and show that the right hand side is a decreasing function
of m for fixed i. Can you give a purely verbal argument? Hint: How does
an investment with nominal rate i(2) compounded annually compare with an
investment at nominal rate i(2) compounded twice a year?
Problem 2–2. Since i(m) = m((1 + i)1/ m − 1) the limit can be evaluated directly
using L’Hopitals rule, Maclaurin expansions, or the definition of derivative.
1/ 4 −4
Problem 2–3. The relevant equation is 1 + 4i(1/ 4) = 1 − d (4) / 4 .
Problem 2–4. In the constant force setting A(t) = eδ t and A′ (t) = δ A(t). The
equation A′ (t) = δ (t)A(t) can be solved by separation of variables.
Problem 2–5. 1 − d (m) / m = v1/ m .
Problem 2–6. d (m) / m = v1/ m i(m) / m.

Solutions to Exercises
Exercise 2–1. The equation to be solved is (1 + 0.05/ 4)4 = 1 + i, where i is the
effective rate of interest.
Exercise 2–2. Taking tth roots of both sides of the equation shows that t plays
no role in determining i and leads to the equation i = 365((1+0.05/ 12)12/ 365 −1) =
0.04989.
Exercise 2–3. 0.05(12) = 12((1 + 0.05)1/ 12 − 1) = 0.04888.
Exercise 2–4. The requirement for equivalence is that eδ = 1 + i.
Exercise 2–5. Here eδ = (1 + 0.05/ 365)365, so that δ = 0.4999. So as a rough

approximation when compounding daily the force of interest is the same as the
nominal interest rate.
Exercise 2–6. The present value in this case is $2000(1 + 0.06)−10 = $1116.79.
Exercise 2–7. A payment of d (m) / m made at time 0 is required to be equivalent

to a payment of i(m) / m made at time 1/ m. Hence d (m) / m = v1/ m i(m) / m. Since
v−1/ m = (1 +i)1/ m = 1 +i(m) / m this gives d (m) / m = 1 − v1/ m or 1 +i = (1 − d (m) / m)−m .
Another relation is that d (m) / m − i(m) / m = (d (m) / m)(i(m) / m).
Exercise 2–8. The given information is d (4) = 0.05, from which i can be
obtained using the formula of the previous exercise as i = (1 − 0.05/ 4)−4 − 1 =
0.0410.
§3. Annuities Certain
Many different types of financial transactions involve the payment of a fixed

amount of money at regularly spaced intervals of time for a predetermined period.
Such a sequence of payments is called an annuity certain or, more simply, an
annuity. A common example is loan payments. It is easy to use the idea of present
value to evaluate the worth of such a cash stream at any point in time. Here is an
example.
Example 3–1. Suppose you have the opportunity to buy an annuity, that is, for a
certain amount paid by you today you will receive monthly payments of $400, say,
for the next 20 years. How much is this annuity worth to you? Suppose that the
payments are to begin one month from today. Such an annuity is called an annuity
immediate (a truly unfortunate choice of terminology). It is useful to visualize the
cash stream represented by the annuity on a time diagram.
An Annuity Immediate
$400 $400 ... ... $400 $400

↓ ↓ ↓ ↓
0 1 2 239 240
Clearly you would be willing to pay today no more than the present value of the
total payments made by the annuity. Assume that you are able to earn 5% interest
(nominal annual rate) compounded monthly. The present value of the payments is

240
.05 −j
(1 + ) 400.
j=1 12
This sum is simply the sum of the present value of each of the payments using the
indicated interest rate. It is easy to find this sum since it involves a very simple
geometric series.
Exercise 3–1. Evaluate the sum.
Since expressions of this sort occur rather often, actuaries have developed some
special notation for this sum. Write an for the present value of an annuity which
pays $1 at the end of each period for n periods. Then

n
j 1 − vn
an = v =
j=1 i
where the last equality follows from the summation formula for a geometric series.
The interest rate per period is usually not included in this notation, but when such
§3: Annuities Certain 11
information is necessary the notation is an i . The present value of the annuity in the
previous example may thus be expressed as 400a240 .05/12 .
A slightly different annuity is the annuity due which is an annuity in which the
payments are made starting immediately. The notation än denotes the present value
of an annuity which pays $1 at the beginning of each period for n periods. Clearly

n−1
1 − vn
än = vj =
j=0 d
where again the last equality follows by summing the geometric series. Note that n
still refers to the number of payments. If the present time is denoted by time 0, then
for an annuity immediate the last payment is made at time n, while for an annuity
due the last payment is made at time n − 1, that is, the beginning of the nth period.
It is quite evident that an = v än , and there are many other similar relationships.
Exercise 3–2. Show that an = v än .
The connection between an annuity due and an annuity immediate can be viewed
in the following way. In an annuity due the payment for the period is made at the
beginning of the period, whereas for an annuity immediate the payment for the
period is made at the end of the period. Clearly a payment of 1 at the end of the
period is equivalent to the payment of v = 1/ (1 + i) at the beginning of the period.
This gives an intuitive description of the equality of the previous exercise.
Example 3–2. Suppose that the annuity is paid continuously, that is, that the annu-
itant receives money at a constant rate of σ dollars per unit time. What value of σ
makes this continuous 20 year annuity equivalent to the discrete annuity described
above? Two annuities are said to be equivalent if they have the same present value.
For the continuous annuity the present value of the σ dt dollars received in the time
interval (t, t + dt) is e−δ t σ dt. The present value of this annuity is therefore
20
σ e−δ t dt.
0
It is now a simple matter to find σ .
Exercise 3–3. Find σ . Note that you must first find δ which is equivalent to an
interest rate of 5% compounded monthly.
A continuous annuity of the type above which is payable at rate σ = 1 for n

periods has present value which is denoted by an .
1 − vn
Exercise 3–4. Show that an = .
δ
Thus far the value of an annuity has been computed at time 0. Another common
time point at which the value of an annuity consisting of n payments of 1 is
computed is time n. Denote by sn the value of an annuity immediate at time n, that
is, immediately after the nth payment. Then sn = (1 + i)n an from the time diagram.
The value sn is called the accumulated value of the annuity immediate. Similarly
s̈n is the accumulated value of an annuity due at time n and s̈n = (1 + i)n än .
Here are two examples which further develop skill in the use of these ideas.
Example 3–3. You are going to buy a house for which the purchase price is $100,000
and the downpayment is $20,000. You will finance the $80,000 by borrowing this
amount from a bank at 10% interest with a 30 year term. What is your monthly
payment? Typically such a loan is amortized, that is, you will make equal monthly
payments for the life of the loan and each payment consists partially of interest and
partially of principal. From the banks point of view this transaction represents the
purchase by the bank of an annuity immediate. The monthly payment, p, is thus the
solution of the equation 80000 = pa360 0.10/12 . In this setting the quoted interest rate
on the loan is assumed to be compounded at the same frequency as the payment
period unless stated otherwise.
Exercise 3–5. Find the monthly payment. What is the total amount of the payments
made?
Example 3–4. Long ago I bought a new car from a local dealer. Let us say the total
cost to me was $15,000. The dealer seemed very anxious that I finance the purchase
through him, and he presented several arguments as to why I should do so. I could
borrow the entire purchase price at 11.95% ammortized over 60 months. His first
sales pitch was as follows. If I financed the car I would pay about $5000 in interest.
If I paid cash I would lose about $8000 in interest that I could earn by investing my
money in a savings account at 8.5% interest. Thus I would gain almost $3000 by
financing the car. Is this argument correct? If not, what’s wrong with it?
Exercise 3–6. Find the monthly payment on the car if it is financed through the
dealer. What is the total interest paid? Is this a relevant fact?
The dealers second argument ran as follows. Suppose that at the end of 5
years the car is worth only half its present value, namely, $7500. Let’s analyze my
available assets under the two alternatives. If I pay cash for the car, at the end of
5 years I will have clear title to a $7500 asset. If I finance the car, at the end of 5
years I will have clear title to the car ($7500) plus the cash I did not pay originally
($15000) plus the interest on this cash ($8000) for a total of $30500. Obviously
only a fool would pass up this type of opportunity!
Exercise 3–7. What are the flaws, if any, in this second argument?
Problems
Problem 3–1. Show that an < an < än . Hint: This should be obvious from the
picture.
Problem 3–2. John borrows $1,000 from Jane at an annual effective rate of interest
i. He agrees to pay back $1,000 after six years and $1,366.87 after another 6 years.
Three years after his first payment, John repays the outstanding balance. What is
the amount of John’s second payment?
Problem 3–3. Suppose a loan of amount A is amortized by a series of n payments.

Denote by bk the loan balance immediately after the kth payment and write b0 = A.
Find a relationship that expresses bk+1 in terms of bk , the interest rate i, and P = A/ an .
Problem 3–4. There are two common ways of analyzing loans which are amortized.
In the prospective method the loan balance at any point in time is seen to be the
present value of the remaining loan payments. Use the previous problem to show
that this statement is correct.
Problem 3–5. In the retrospective method the loan balance at any point in time is
seen to be the accumulated original loan amount less the accumulated value of the
past loan payments. Show that this formula for the loan balance is correct.
Problem 3–6. An annuity immediate pays an initial benefit of one per year, in-
creasing by 10.25% every four years. The annuity is payable for 40 years. If the
effective interest rate is 5% find an expression for the present value of this annuity.
Problem 3–7. You are given an annuity immediate paying $10 for 10 years, then
decreasing by $1 per year for nine years and paying $1 per year thereafter, forever.
If the annual effective rate of interest is 5%, find the present value of this annuity.
Problem 3–8. Humphrey purchases a home with a $100,000 mortgage. Mortgage

payments are to be made monthly for 30 years, with the first payment to be made one
month from now. The rate of interest is 10%. After 10 years, Humphrey increases
the amount of each monthly payment by $325 in order to repay the mortgage more
quickly. What amount of interest is paid over the life of the loan?
Problem 3–9. On January 1, an insurance company has $100,000 which is due to

Linden as a life insurance death benefit. He chooses to receive the benefit annually
over a period of 15 years, with the first payment made immediately. The benefit
he receives is based on an effective interest rate of 4% per annum. The insurance
company earns interest at an effective rate of 5% per annum. Every July 1 the
company pays $100 in expenses and taxes to maintain the policy. How much
money does the company have remaining after 9 years?
Problem 3–1. Isn’t the chain of inequalities simply expressing the fact that
getting a given amount of money sooner makes it worth more? An analytic
proof should be easy to give too.
Problem 3–2. From Jane’s point of view the equation 1000 = 1000(1 + i)−6 +
1366.87(1 + i)−12 must hold. The outstanding balance at the indicated time is
1366.87(1 + i)−3 , which is the amount of the second payment.
Problem 3–3. bk+1 = bk + ibk − P.
Problem 3–4. The assertion of the prospective method is that bk = Pan−k .

Plug in and show that this choice satisfies the initial condition b0 = A and the
recursion of the previous problem.
Problem 3–5. Here the assertion is that bk = A(1 + i)k − Psk . Show that this
choice satisfies the required recursion and initial condition.
Problem 3–6. Each 4 year chunk is a simple annuity immediate. Taking the
present value of these chunks forms an annuity due with payments every 4 years
that are increasing.
Problem 3–7. What is the present value of an annuity immediate paying $1 per
year forever? What is the present value of such an annuity that begins payments
k years from now? The annuity desribed here is the difference of a few of these.
Problem 3–8. The initial monthly payment P is the solution of 100, 000 =
Pa360 . The balance after 10 years is Pa240 so the interest paid in the first 10
years is 120P − (100, 000 − Pa240 ). To determine the number of new monthly
payments required to repay the loan the equation Pa240 = (P + 325)ax should be
solved for x. Since after x payments the loan balance is 0 the amount of interest
paid in the second stage can then be easily determined.
Problem 3–9. Since the effective rate of interest for the insurance company is
5%, the rate (0.05)(2) should be used to move the insurance company’s expenses
from July 1 to January 1.
Exercise
240 3–1. The sum is the sum of the terms of a geometric series. So
.05 −j −1 −241
j=1 (1+ 12 ) 400 = 400((1+0.05/ 12) −(1+0.05/ 12) )/ (1−(1+0.05/ 12)−1) =
60, 610.12.
Exercise 3–2. This follows from the formulas for the present value of the two
annuities and the fact that d = iv.
Exercise 3–3. The force of interest that is equivalent to an interest rate of 5%

compounded monthly is δ = ln(1 +0.05/ 12)12. Integration then gives the present
value of the continuous annuity as σ (1 − e−20δ )/ δ . So σ = 400δ / i = 4790.03
will make the continuous annuity equivalent to the original discrete annuity.
Exercise 3–4. This formula follows by straightforward integration.
Exercise 3–5. Using the earlier formula gives a360 0.10/ 12 = 113.95 from which
p = 702.06 and the total amount of the payments is 360p = 252740.60.
Exercise 3–6. The monthly payment is 15, 000/ a60 .1195/ 12 = 333.29 so that the
total of the payments is 19, 997.27 of which 4, 997.27 is interest. The total of
the interest payments is irrelevant, since the time at which the interest payment
is made is not taken into account.
Exercise 3–7. Where did all those payments go?

§4. Laboratory 1
1. A loan of 10,000 carries an interest rate of 9% compounded quarterly. Equal

loan payments are to be made monthly for 36 months. What is the size of each
payment?
2. An amortization table is a table which lists the principal and interest portions
of each payment for a loan which is being amortized. Construct an amortization
table for the loan of the previous problem. The table should have four columns:
the payment number, the principal part of that payment, the interest part of that
payment, and the loan balance immediately after that payment is made.
3. A loan of 10,000 is to be repaid with equal monthly payments of p. The

interest rate for the first year is 1.9%, while the interest rate for the remaining 2
years is 10.9%. What is p? What is the balance after the 6th payment? After the
15th payment? What are the principal and interest components of the 6th payment?
Of the 15th payment?
4. A loan of 10,000 is to be repaid as follows. Payments of p are to be made at

the end of each month for 36 months and a balloon payment of 2500 is to be made
at the end of the 36th month as well. If the interest rate is 5%, what is p? What is
the loan balance at the end of the 12th month? What part of the 15th payment is
interest? Principal?
5. The symbol a(m)

n denotes the present value of an annuity immediate that pays
1/ m at the end of each mth part of a year for n years. For example, if m = 12
payments of 1/12 are made at the end of each month. Find a formula for a(m)
n .
6. The symbol ä(m)

n denotes the present value of an annuity due that pays 1/ m at
the beginning of each mth part of a year for n years. Find a formula for ä(m)
n .
7. An increasing annuity immediate with a term of n periods pays 1 at the end

of the first period, 2 at the end of the second period, 3 at the end of the third period,
. . . , n at the end of the nth period. Find (Ia)n , the present value of such an annuity.
8. A decreasing annuity immediate with a term of n periods pays n at the end

of the first period, n − 1 at the end of the second period, n − 2 at the end of the third
period, . . . , 1 at the end of the nth period. Find (Da)n , the present value of such an
annuity.

§5. Brief Review of Probability Theory
Another aspect of insurance is that money is paid by the company only if some
event, which may be considered random, occurs within a specific time frame. For
example, an automobile insurance policy will experience a claim only if there is an
accident involving the insured auto. In this section a brief outline of the essential
material from the theory of probability is given. Almost all of the material presented
here should be familiar to the reader. The concepts presented here will play a crucial
role in the rest of these notes.
The underlying object in probability theory is a sample space S, which is simply

a set. This set is sometimes thought of as the collection of all possible outcomes
of a random experiment. Certain subsets of the sample space, called events, are
assigned probabilities by the probability measure (or probability set function)
which is usually denoted by P. This function has a few defining properties.
(1) For any event E ⊂ S, 0 ≤ P[E] ≤ 1.
(2) P[∅] = 0 and P[S] = 1.

(3) If E1 , E2 , . . . are events and Ei ∩ Ej = ∅ for i ≠ j then P[ ∞i=1 Ei ] = ∞i=1 P[Ei ].
From these basic facts one can deduce all manner of useful computational formulas.
Exercise 5–1. Show that if A ⊂ B are events, then P[A] ≤ P[B].
Another of the basic concepts is that of a random variable. A random variable

is a function whose domain is the sample space of a random experiment and whose
range is the real numbers.
In practice, the sample space of the experiment fades into the background and
one simply identifies the random variables of interest. Once a random variable has
been identified, one may ask about its values and their associated probabilities. All
of the interesting probability information is bound up in the distribution function of
the random variable. The distribution function of the random variable X, denoted
FX (t), is defined by the formula FX (t) = P[X ≤ t].
Two types of random variables are quite common. A random variable X with
distribution function FX is discrete if FX is constant except at at most countably
many jumps. A random t
variable X with distribution function FX is absolutely
d
continuous if FX (t) = FX (s) ds holds for all real numbers t.
−∞ ds
If X is a discrete random variable, the density of X, denoted fX (t) is defined by

the formula fX (t) = P[X = t]. There are only countably many values of t for which
the density of a discrete random variable is not 0. If X is an absolutely continuous
d
random variable, the density of X is defined by the formula fX (t) = FX (t).
dt
§5: Brief Review of Probability Theory 18
Example 5–1. A Bernoulli random variable is a random variable which takes on
exactly two values, 0 and 1. Such random variables commonly arise to indicate the
success or failure of some operation. A Bernoulli random variable is discrete.
Exercise 5–2. Sketch the distribution function of a Bernoulli random variable with
P[X = 1] = 1/ 3.
Example 5–2. An exponentially distributed random variable Y with parameter

λ > 0 is a non-negative random variable for which P[Y ≥ t] = e−λ t for t ≥ 0. Such a
random variable is often used to model the waiting time until a certain event occurs.
An exponential random variable is absolutely continuous.
Exercise 5–3. Sketch the distribution function of an exponential random variable

with parameter λ = 1. Sketch its density function also.
Exercise 5–4. A random variable X is uniformly distributed on an interval (a, b) if

X represents the result of selecting a number at random from (a, b). Find the density
and distribution function of a random variable which is uniformly distributed on the
interval (0, 1).
Exercise 5–5. Draw a picture of a distribution function of a random variable which

is neither discrete nor absolutely continuous.
Another useful tool is the indicator function. Suppose A is a set. The indicator
function of the set A, denoted 1A (t), is defined by the equation

1 if t ∈ A
1A (t) =
0 if t ∉ A.
Exercise 5–6. Graph the function 1[0,1) (t).
Exercise 5–7. Verify that the density of a random variable which is exponential
with parameter λ may be written λ e−λ x 1(0,∞) (x).
Example 5–3. Random variables which are neither of the discrete nor absolutely
continuous type will arise frequently. As an example, suppose that person has a fire
insurance policy on a house. The amount of insurance is $50,000 and there is a $250
deductible. Suppose that if there is a fire the amount of damage may be represented
by a random variable D which has the uniform distribution on the interval (0, 70000).
(This assumption means that the person is underinsured.) Suppose further that in the
time period under consideration there is a probability p = 0.001 that a fire will occur.
Let F denote the random variable which is 1 if a fire occurs and is 0 otherwise. It is
easy to see that the size X of the claim to the insurer in this setting is given by

X = F (D − 250)1[250,50250] (D) + 500001(50250,∞) (D) .
This random variable X is neither discrete nor absolutely continuous.
Exercise 5–8. Verify the correctness of the formula for X. Find the distribution
function of the random variable X.
Often only the average value of a random variable and the spread of the
values around this average are all that are needed. The
expectation (or mean)
of a discrete random variable X is defined by E[X] = t fX (t), while the ex-
t
pectation of an absolutely continuous random variable X is defined by E[X] =
∞
t fX (t) dt. Notice that in both cases the sum (or integral) involves terms of the
−∞
form (possible value of X) × (probability X takes on that value). When X is neither
discrete nor absolutely continuous, the expectation is defined by the Riemann–
∞
Stieltjes integral E[X] = −∞ t dFX (t), which again has the same form.
Exercise 5–9. Find the mean of a Bernoulli random variable Z with P[Z = 1] = 1/ 3.
Exercise 5–10. Find the mean of an exponential random variable with parameter
λ = 3.
Exercise 5–11. Find the mean and variance of the random variable in the fire
insurance example given above. (The variance of a random variable X is defined
by Var(X) = E[(X − E[X])2 ] and is often computed using the alternate formula
Var(X) = E[X 2 ] − (E[X])2 .)
Computation of conditional probabilities will play an important role. If A and

B are events, the conditional probability of A given B, denoted P[A| B], is defined
by P[A| B] = P[A ∩ B]/ P[B] as long as P[B] ≠ 0.
The events A and B are independent if P[A| B] = P[A]. The intuition underlying
the notion of independent events is that the occurance of one of the events does not
alter the probability that the other event occurs.
Similar definitions can be given in the case of random variables. Intuitively,

the random variables X and Y are independent if knowledge of the value of one of
them does not effect the probabilities of events involving the other. Independence
of random variables is usually assumed based on this intuition. One important fact
is that if X and Y are independent random variables then E[XY] = E[X] E[Y].
Problems
Problem 5–1. Suppose X has the uniform distribution on the interval (0, a) where
a > 0 is given. What is the mean and variance of X?
Problem 5–2. The moment generating function of a random variable X, denoted

MX (t), is defined by the formula MX (t) = E[etX ]. What is the relationship between
MX′ (0) and E[X]? Find a formula for Var(X) in terms of the moment generating
function of X and its derivatives at t = 0.
Problem 5–3. Express the Maclaurin expansion of MX (t) in terms of the moments
E[X], E[X 2 ], E[X 3 ],. . . of X. Hint: What is the Maclaurin expansion of ex ?
Problem 5–4. Find the moment generating function of a Bernoulli random variable
Y for which P[Y = 1] = 1/ 4.

d d2

Problem 5–5. Show that ln MX (t) = E[X] and 2 ln MX (t) = Var(X). This
dt t=0 dt
t=0
is useful when the moment generating function has a certain form.
Problem 5–6. Find the moment generating function of a random variable Z which
has the exponential distribution with parameter λ . Use the moment generating
function to find the mean and variance of Z.
Problem 5–7. A double indemnity life insurance policy has been issued to a person
aged 30. This policy pays $100,000 in the event of non-accidental death and
$200,000 in the event of accidental death. The probability of death during the next
year is 0.002, and if death occurs there is a 70% chance that it was due to an accident.
Write a random variable X which represents the size of the claim filed in the next
year. Find the distribution function, mean, and variance of X.
Problem 5–8. In the preceding problem suppose that if death occurs the day of the
year on which it occurs is uniformly distributed. Assume also that the claim will be
paid immediately at death and the interest rate is 5%. What is the expected present
value of the size of the claim during the next year?
Problem 5–9. Show that for a random

∞
variable Y which is discrete and takes non–

negative integer values, E[Y] = P[Y ≥ i]. Find a similar alternate expression for
i=1
E[Y 2 ].
Problem 5–10. Suppose Y is a non-negative, absolutely continuous random vari-

∞
able. Show that E[Y] = P[Y > t] dt.
0
Problem 5–1. E[X] = a/ 2 and Var(X) = a2 / 12.
Problem 5–2. MX′ (0) = E[X].
∞

Problem 5–3. Since ex = xk / k!,
k=0
MX (t) = E[etX ]
∞
= E[ (tX)k / k!]
k=0
∞

= E[X k ]tk / k!.
k=0
Thus the coefficient of tk in the Maclaurin expansion of MX (t) is E[X k ]/ k!.
Problem 5–4. MY (t) = 3/ 4 + et / 4.
Problem 5–6. MZ (t) = λ / (λ − t) for 0 ≤ t < λ .
Problem 5–7. Let D be a random variable which is 1 if the insured dies in the
next year and 0 otherwise. Let A be a random variable which is 2 if death is due
to an accident and 1 otherwise. Then X = 100000AD.
Problem 5–8. If U is uniformly distributed on the integers from 1 to 365 then

E[100000ADvU ] is the desired expectation. Here v = 1/ (1 + 0.05(365) / 365).
Problem
5–9. Hint: In the usual formula for the expectation of Y write
i = ij=1 1 and then interchange the order of summation.
Problem 5–10. Use a trick like that of the previous problem. Double integrals
anyone?
Exercise 5–1. Using the fact that B = A ∪ (B \ A) and property (3) gives
P[B] = P[A] + P[B \ A]. By property (1), P[B \ A] ≥ 0, so the inequality
P[B] ≥ P[A] follows.
Exercise 5–2. The distribution function F(t) takes the value 0 for t < 0, the
value 2/ 3 for 0 ≤ t < 1 and the value 1 for t ≥ 1.
Exercise 5–3. The distribution function F(t) is 0 if t < 0 and 1 − e−t for t ≥ 0.
The density function is 0 for t < 0 and e−t for t ≥ 0.
Exercise 5–4. The distribution function F(t) takes the value 0 for t < 0, the
value t for 0 ≤ t ≤ 1 and the value 1 for t > 1. The density function takes the
value 1 for 0 < t < 1 and 0 otherwise.
Exercise 5–5. The picture should have a jump and also a smoothly increasing
portion.
Exercise 5–6. This function takes the value 1 for 0 ≤ t < 1 and the value 0
otherwise.
Exercise 5–8. The distribution function F(t) takes the value 0 if t < 0, the value
(1 − 0.001) + 0.001 × (250/ 70000) for 0 ≤ t < 250 (because X = 0 if either there is
no fire or the loss caused by a fire is less than 250),the value0.999+0.001×t/ 70000
for 250 ≤ t < 50250 and the value 1 for t ≥ 50250.
Exercise 5–9. E[Z] = 0 × (2/ 3) + 1 × (1/ 3) = 1/ 3.

∞ ∞
Exercise 5–10. The expectation is 0 tf (t) dt = 0 t3e−3t dt = 1/ 3 using
integration by parts.
Exercise 5–11. Notice that the loss random variable X is neither discrete nor
absolutely continuous. The distribution function of X has two jumps: one at
t = 0 of size 0.999 + 250/ 70000 and another at 50250of size 0.001 − 0.001 ×
50250/ 70000. So E[X] = 0 × (0.999 + 250/ 70000) + 250 t0.001/ 70000 dt +
50250
50250 × (0.001 − 0.001 × 50250/ 70000). The quantity E[X 2 ] can be computed
similarly.
§6. Laboratory 2
One iteration of an experiment is conducted as follows. A single six sided die

is rolled and the number D of spots up is noted. Then D coins are tossed and the
number H of heads observed is noted.
1. What are the possible values of the random variable D? What are the possible
values of the random variable H?
2. For each relevant value of h and d compute P[H = h, D = d]. This probability
is called the joint density of H and D and is denoted fH,D (h, d). Put your computed
values in the form of a rectangular table with rows indexed by h and columns
indexed by d.
3. Find the density and expectation of D. Find the density and expectation of
H.
4. For each value of d find the conditional density of H given D = d. Notation-

ally, fH | D (h| d) = P[H = h| D = d].
5. For each value of d find the conditional expectation of H given D = d.

Notationally, E[H | D = d] = h fH | D (h| d).
h
6. The conditional expectation of H given D, denoted E[H | D], is the random

variable whose value on the event D = d is E[H | D = d]. Find the density of the
random variable E[H | D].
7. Compute E[E[H | D]]. The Theorem of Total Expectation states that for
any two random variables X and Y, E[E[Y | X]] = E[Y]. Do your computations of
E[H] in this problem and problem 3 reflect this fact?
8. The computations above show that it is easier to compute E[H] by first

finding E[H | D]. Sometimes probabilistic intuition can be used to find a conditional
expectation without computation. The intuition behind the conditional expectation
E[H | D] is that this should be the expected value of H computed after taking the
value of D into account. Give a probabilistic argument (without computation) that
E[H | D] = D/ 2.

§7. Survival Distributions
An insurance policy can embody two different types of risk. For some types
of insurance (such as life insurance) the variability in the claim is only the time at
which the claim is made, since the amount of the claim is specified by the policy.
In other types of insurance (such as auto or casualty) there is variability in both the
time and amount of the claim. The problems associated with life insurance will be
studied first, since this is both an important type of insurance and also relatively
simple in some of its aspects.
The central difficulty in issuing life insurance is that of determining the length
of the future life of the insured. Denote by X the random variable which represents
the future lifetime of a newborn. For mathematical simplicity, assume that the
distribution function of X is absolutely continuous. The survival function of X,
denoted by s(x) is defined by the formula
s(x) = P[X > x] = P[X ≥ x]
where the last equality follows from the continuity assumption. The assumption
that s(0) = 1 will always be made.
Example 7–1. In the past there has been some interest in modelling survival func-
tions in an analytic way. The simplest model is that due to Abraham DeMoivre. He
x
assumed that s(x) = 1 − for 0 < x < ω where ω is the limiting age by which
ω
all have died. The DeMoivre law is simply the assertion that X has the uniform
distribution on the interval (0, ω ).
Life insurance is usually issued on a person who has already attained a certain
age x. For notational convenience denote such a life aged x by (x), and denote the
future lifetime of a life aged x by T(x). What is the survival function for (x)? From
the discussion above, the survival function for (x) is P[T(x) > t]. Some standard
notation is now introduced. Set
t px = P[T(x) > t]
and
t qx = P[T(x) ≤ t].
When t = 1 the prefix is ommitted and one just writes px and qx respectively.
Generally speaking, having observed (x) some additional information about the
survival of (x) can be inferred. For example, (x) may have just passed a physical
exam given as a requirement for obtaining life insurance. For now this type of
possibility is disregarded. Operating under this assumption
s(x + t)
t px = P[T(x) > t] = P[X > x + t| X > x] = .
s(x)
§7: Survival Distributions 25
Exercise 7–1. Write a similar expression for t qx .
Exercise 7–2. Show that for t ≥ s, t px = t−s px+s s px .
There is one more special symbol. Set
t| u qx = P[t < T(x) ≤ t + u]
which represents the probability that (x) survives at least t and no more than t + u
years. Again, if u = 1 one writes t| qx . The relations t| u qx = t+u qx − t qx = t px − t+u px
follow immediately from the definition.
Exercise 7–3. Prove these two equalitites. Show that t| u qx = t px u qx+t .
Exercise 7–4. Compute t px for the DeMoivre law of mortality. Conclude that under
the DeMoivre law T(x) has the uniform distribution on the interval (0, ω − x).
Under the assumption that X is absolutely continuous the random variable T(x)
will be absolutely continuous as well. Indeed
s(x + t)
P[T(x) ≤ t] = P[x ≤ X ≤ x + t| X > x] = 1 −
s(x)
so the density of T(x) is given by
−s′ (x + t) fX (x + t)
fT(x) (t) = = .
s(x) 1 − FX (x)
Intuitively this density represents the rate of death of (x) at time t.
∞
Exercise 7–5. Use integration by parts to show that E[T(x)] = t px dt. This
0
expectation is called the
∞complete expectation of life and is denoted by e̊x . Show
2
also that E[T(x) ] = 2 t t px dt.
0
Exercise 7–6. If X follows DeMoivre’s law, what is e̊x ?
It is often useful to consider then the quantity
=− ′
fX (x) s (x)
µx =
1 − FX (x) s(x)
which is called the force of mortality. Intuitively the force of mortality is the
instantaneous rate of death of (x). (In component reliability theory this function is
often referred to as the hazard rate.) Integrating both sides of this equality gives the
useful relation x
s(x) = exp − µt dt .
0
Exercise 7–7. Derive this last expression.
x+t
µs ds
Exercise 7–8. Show that t px = e− x .
Exercise 7–9. Show that the density of T(x) can be written fT(x) (t) = t px µx+t .
If the force of mortality is constant the life random variable X has an expo-
nential distribution. This is directly in accord with the “memoryless” property of
exponential random variables. This memoryless property also has the interpretation
that a used article is as good as a new one. For human lives (and most manufactured
components) this is a fairly poor assumption, at least over the long term. The force
of mortality usually is increasing, although this is not always so.
Exercise 7–10. Find the force of mortality for DeMoivre’s law.
The curtate future lifetime of (x), denoted by K(x), is defined by the relation
K(x) = [T(x)]. Here [t] is the greatest integer function. Note that K(x) is a discrete
random variable with density P[K(x) = k] = P[k ≤ T(x) < k + 1]. The curtate
lifetime, K(x), represents the number of complete future years lived by (x).
Exercise 7–11. Show that P[K(x) = k] = k px qx+k .
Exercise 7–12. Show that the curtate expectation of life ex = E[K(x)] is given by

the formula ex = ∞i=0 i+1 px . Hint: E[Y] = ∞i=1 P[Y ≥ i].
Problems
Problem 7–1. Suppose µx+t = t for t ≥ 0. Calculate t px µx+t and e̊x .
∂ d
Problem 7–2. Calculate t px and e̊x .
∂x dx
Problem 7–3. A life aged (40) is subject to an extra risk for the next year only.
Suppose the normal probability of death is 0.004, and that the extra risk may be
expressed by adding the function 0.03(1 − t) to the normal force of mortality for this
year. What is the probability of survival to age 41?
Problem 7–4. Suppose qx is computed using force of mortality µx , and that q′x is
computed using force of mortality 2µx . What is the relationship between qx and q′x ?
Problem 7–5. Show that the conditional distribution of K(x) given that K(x) ≥ k is
the same as the unconditional distribution of K(x + k) + k.
Problem 7–6. Show that the conditional distribution of T(x) given that T(x) ≥ t is
the same as the unconditional distribution of T(x + t) + t.
Problem 7–7. The Gompertz law of mortality is defined by the requirement that
µt = Act for some constants A and c. What restrictions are there on A and c for this
to be a force of mortality? Write an expression for t px and e̊x under Gompertz’ law.
Problem 7–8. Makeham’s law of mortality is defined by the requirement that

µt = A + Bct for some constants A, B, and c. What restrictions are there on A, B
and c for this to be a force of mortality? Write an expression for t px and e̊x under
Makeham’s law.
t ∞
− µx+s ds
= e−t /2
= √2π / 2.
2
Problem 7–1. Here t px = e 0 and e̊x = t px dt
0
∞
∂ d ∂
Problem 7–2. t px = t px (µx − µx+t ) and e̊x = t px dt = µx e̊x − 1.
∂x dx 0 ∂x
1
− µ40+s +0.03(1−s) ds
Problem 7–3. If µt is the usual force of mortality then p40 = e 0 .
Problem 7–4. The relation p′x = (px )2 holds, which gives a relation for the
death probability.
Problem 7–5. P[K(x) ≤ k + l| K(x) ≥ k] = P[k ≤ K(x) ≤ k + l]/ P[K(x) ≥ k] =

= P[K(x + k) + k ≤ l + k].
l qx+k
Problem 7–6. Proceed as in the previous problem.

Exercise 7–1. t qx = P[T(x) ≤ t] = P[X ≤ x + t| X > x] = P[x < X ≤ x + t]/ P[X >
x] = (s(x) − s(x + t))/ s(x).
Exercise 7–2. t px = s(x + t)/ s(x) = s(x + s + (t − s))/ s(x) = (s(x + s + (t − s))/ s(x +
s))(s(x + s)/ s(x)) = t−s px+ss px . What does this mean in words?
Exercise 7–3. For the first one, t| u qx = P[t < T(x) ≤ t + u] = P[x + t < X ≤
t+u+x| X > x] = (s(x+t)−s(t+u+x))/ s(x) = (s(x+t)−s(x)+s(x)−s(t+u+x))/ s(x) =
t+u qx − t qx . The second identity follows from the fourth term by simplifying
(s(x + t) − s(t + u + x))/ s(x) = t px − t+u px . For the last one, t| u qx = P[t < T(x) ≤
t + u] = P[x + t < X ≤ t + u + x| X > x] = (s(x + t) − s(t + u + x))/ s(x) =
(s(x + t)/ s(x))(s(t + x) − s(t + u + x))/ s(x + t) = t px u qx+t .
Exercise 7–4. Under the DeMoivre law, s(x) = (ω − x)/ ω so that t px =

(ω − x − t)/ (ω − x) for 0 < t < ω − x. Thus the distribution function of T(x) is
t/ (ω − x) for 0 < t < ω − x, which is the distribution function of a uniformly
distributed random variable.
∞ ∞ ∞
Exercise 7–5. ∞ E[T(x)] = 0 tfT(x) (t) dt = − 0 ts′ (x + t)/ s(x) dt = 0 s(x +
t)/ s(x) dt = 0 t px dt. The fact the limt→∞ s(x + t) = 0 is assumed, since everyone
eventually dies. The other expectation is computed similarly.
Exercise 7–6. Under DeMoivre’s law, e̊x = (ω − x)/ 2, since T(x) is uniform on
the interval (0, ω − x).
x x
Exercise 7–7. 0 µt dt = 0 −s′ (t)/ s(t) dt = − ln(s(x)) + ln(s(0)) = − ln(s(x)),
since s(0) = 1. Exponentiating both sides to solve for s(x) gives the result.
x+t
− µ ds
Exercise 7–8. From the previous exercise, s(x + t) = e 0 s . Using this
fact, the previous exercise, and the fact that t px = s(x + t)/ s(x) gives the formula.
Exercise 7–9. Since fT(x) (t) = −s′ (x + t)/ s(x) and s′ (x + t) = −s(x + t)µx+t by the
previous exercise, the result follows.
Exercise 7–10. From the earlier expression for the survival function under
DeMoivre’s law s(x) = (ω − x)/ ω , so that µx = −s′ (x)/ s(x) = 1/ (ω − x), for
0 < x < ω.
Exercise 7–11. P[K(x) = k] = P[k ≤ T(x) < k+1] = (s(x+k)−s(x+k+1))/ s(x) =

((s(x + k) − s(x + k + 1))/ s(x + k))(s(x + k)/ s(x)) = k px qx+k .
∞ ∞ ∞
Exercise
∞ 7–12. E[K(x)] = i=1 P[K(x) ≥ i] = i=1 P[T(x) ≥ i] = i=1 i px =
j=0 j+1 x .
p
§8. Life Tables
In practice the survival distribution is estimated by compiling mortality data in

the form of a life table. An example of a life table appears at the end of these notes.
Here is the conceptual model behind the entries in the table. Imagine that at time
0 there are l0 newborns. Here l0 is called the radix of the life table and is usually
taken to be some large number such as 100,000. Denote by lx the number of these
original newborns who are still alive at age x. Similarly n dx denotes the number of
persons alive at age x who die before reaching age x + n. As usual, when n = 1 it is
supressed in the notation.
Exercise 8–1. Show that n dx = lx − lx+n .

lx
The ratio is an estimate of s(x) based on the collected data. Assume that in
l0
lx
fact s(x) = for non-negative integer values of x. Since earlier the assumption
l0
was made that the life random variable X is absolutely continuous, the question
arises as to how the values of the survival function will be computed at non-integer
values of x. There are three commonly used methods of doing this, and these
methods produce slightly different numerical results. For the remainder of this
discussion suppose x is fixed (and an integer) and that 0 ≤ t ≤ 1.
Under the assumption of the uniform distribution of deaths in the year of

death, denoted UDD, the survival function is computed by the formula
s(x + t) = (1 − t)s(x) + ts(x + 1).

The UDD assumption is the one most commonly made.
The assumption of a constant force of mortality in each year of age leads to

the formula
s(x + t) = s(x) e−µ t
where µ = − log px .
The Balducci assumption is expressed in the formula

1 1−t t
= + .
s(x + t) s(x) s(x + 1)
Under each of the assumptions an explicit expression for all of the survivor
functions can be found.
Exercise 8–2. Find expressions for t qx and µx+t , 0 ≤ t ≤ 1, under each of the above
3 assumptions.
§8: Life Tables 31
Having observed (x) may mean more than simply having seen a person aged x.
It may well mean that (x) has just passed a physical exam in preparation for buying
a life insurance policy. One would expect that the survival distribution of such a
person could be different from s(x). If this is believed to be the case the survival
function is actually dependent on two variables: the age at selection (application
for insurance) and the amount of time passed after the time of selection. A life
table which takes this effect into account is called a select table. A family of
survival functions indexed by both the age at selection and time are then required
and notation such as q[x]+i denotes the probability that a person dies between years
x + i and x + i + 1 given that selection ocurred at age x. As one might expect it
is reasonable to suppose that after a certain period of time the effect of selection
on mortality is negligable. The length of time until the selection effect becomes
negligable is called the select period. The Society of Actuaries (based in Illinois)
uses a 15 year select period in its mortality tables. The Institute of Actuaries in
Britain uses a 2 year select period. The implication of the select period of 15 years
in computations is that for each j ≥ 0, l[x]+15+j = lx+15+j .
A life table in which the survival functions are tabulated for attained ages only
is called an aggregrate table. Generally, a select life table contains a final column
which constitutes an aggregate table. The whole table is then referred to as a select
and ultimate table and the last column is usually called an ultimate table. With
these observations in mind it is easy to utilize select life tables in computations.
Exercise 8–3. You are given the following extract from a 3 year select and ultimate
mortality table.
x l[x] l[x]+1 l[x]+2 lx+3 x+3
70 7600 73
71 7984 74
72 8016 7592 75
Assume that the ultimate table follows DeMoivre’s law and that d[x] = d[x]+1 =
d[x]+2 for all x. Find 1000(2| 2 q[71] ).
§8: Life Tables 32
Problems
Problem 8–1. Graph µx+t , 0 ≤ t ≤ 1, under each of the 3 assumptions for fractional
years.
Problem 8–2. For each of the 3 assumptions for fractional years find a formula for
t px , 0 ≤ t ≤ 1 in terms of t and px . For each of 20 equally space values of px between
0 and 1, make a plot of t px for 0 ≤ t ≤ 1 under each of the 3 assumptions. For what
value(s) of px are the assumptions numerically indistinguishable?
Problem 8–3. Use the life table to compute 1/2 p20 under each of the 3 assumptions
for fractional years.
Problem 8–4. Show that under the assumption of uniform distribution of deaths in
the year of death that K(x) and T(x) − K(x) are independent and that T(x) − K(x) has
the uniform distribution on the interval (0, 1).
Problem 8–5. Show that under UDD e̊x = ex + 12 .

§8: Life Tables 33
Problem 8–3. Under UDD, t px = (1 − t) + tpx so 1/ 2 p20 = 1/ 2 + 1/ 2p20 .
Under constant force, t px = et log px so 1/ 2 p20 = e(1/ 2) log p20 . Under Balducci,
1/ t px = 1 − t + t/ px so 1/ 2 p20 = 1/ (1/ 2 + 1/ 2p20 ).
Problem 8–4. For 0 ≤ t < 1, P[K(x) = k, T(x) − K(x) ≤ t] = P[k ≤ T(x) ≤

k + t] = k px t qx+k = k px (t − tpx+k ) = tP[K(x) = k].
Problem 8–5. Use the previous problem.

§8: Life Tables 34
Exercise 8–1. Since n dx is the number alive at age x who die by age x + n, this
is simply the number alive at age x, which is lx , minus the number alive at age
x + n, which is lx+n .
Exercise 8–2. Under UDD, t qx = (s(x) − s(x + t))/ s(x) = (ts(x) − ts(x + 1))/ s(x) =
tqx and µx+t = −s′ (x+t)/ s(x+t) = (s(x)−s(x+1))/ ((1−t)s(x)+ts(x+1)) = (1−px)/ (1−
t + tpx ). Under constant force, t qx = 1 − (px )t while µx+t = µx . Under Balducci,
s(x + t) = s(x)s(x + 1)/ ((1 − t)s(x + 1) + ts(x)) so that s(x + t)/ s(x) = px / (px + tqx ).
Thus t qx = tqx / (px + tqx ). Also µx+t = −s′ (x + t)/ s(x + t) = (1 − qx )/ (1 + tqx ).
Exercise 8–3. The objective is to compute 10002| 2 q[71] = 1000(2p[71] − 4 p[71] ) =

1000(l[71]+2 − l[71]+4 )/ l[71] = 1000(l[71]+2 − l75 )/ l[71] , where the effect of the selec-
tion period has been used. To find the required entries in the table proceed as
follows. Since 8016 − 7592 = 424 and using the assumption about the number of
deaths, l[72]+1 = 8016−212 = 7804 and l72+3 = 7592−212 = 7380. Since the ulti-
mate table follows DeMoivre’s Law, l71+3 = (7600 + 7380)/ 2 = 7490. Again us-
ing the assumption about the number of deaths, l[71]+2 = (7984 + 7490)/ 2 = 7737
and l[71] = 7984 + 247 = 8231. So 10002| 2 q[71] = 1000(7737 − 7380)/ 8231 =
43.37.
§9. Laboratory 3
1. Below is a table which gives the values of qx for ages 1 through 105. Define
l1 = 1, 000, 000 and use the values of qx to compute lx and dx for 1 ≤ x ≤ 106.
2. Make a plot of lx for 1 ≤ x ≤ 106.
x qx x qx x qx
1 0.000637 36 0.000841 71 0.026627
2 0.000430 37 0.000904 72 0.029565
3 0.000357 38 0.000964 73 0.032931
4 0.000278 39 0.001021 74 0.036738
5 0.000255 40 0.001079 75 0.041002
6 0.000244 41 0.001142 76 0.045699
7 0.000234 42 0.001215 77 0.050833
8 0.000216 43 0.001299 78 0.056487
9 0.000209 44 0.001397 79 0.062777
10 0.000212 45 0.001508 80 0.069757
11 0.000219 46 0.001629 81 0.077444
12 0.000228 47 0.001762 82 0.085828
13 0.000240 48 0.001905 83 0.094904
14 0.000254 49 0.002060 84 0.104700
15 0.000269 50 0.002225 85 0.115289
16 0.000284 51 0.002401 86 0.126798
17 0.000301 52 0.002589 87 0.139353
18 0.000316 53 0.002795 88 0.153021
19 0.000331 54 0.003023 89 0.167757
20 0.000345 55 0.003283 90 0.183408
21 0.000357 56 0.003583 91 0.199769
22 0.000366 57 0.003932 92 0.216605
23 0.000373 58 0.004332 93 0.233662
24 0.000376 59 0.004784 94 0.250693
25 0.000376 60 0.005286 95 0.267491
26 0.000378 61 0.005833 96 0.283905
27 0.000382 62 0.006414 97 0.299852
28 0.000393 63 0.007014 98 0.315296
29 0.000412 64 0.007616 99 0.330207
30 0.000444 65 0.008207 100 0.344556
31 0.000499 66 0.008777 101 0.358628
32 0.000562 67 0.009318 102 0.371685
33 0.000631 68 0.009828 103 0.383040
34 0.000702 69 0.010306 104 0.392003
35 0.000773 70 0.010753 105 1.000000

§10. Status
A life insurance policy is sometimes issued which pays a benefit at a time

which depends on the survival characteristics of two or more people. A status is an
artificially constructed life form for which the notion of life and death can be well
defined.
Example 10–1. A common artificial life form is the status which is denoted n . This
is the life form which survives for exactly n time units and then dies.
Example 10–2. Another common status is the joint life status which is constructed
as follows. Given two life forms (x) and (y) the joint life status, denoted x : y, dies
exactly at the time of death of the first to die of (x) and (y).
Exercise 10–1. If (x) and (y) are independent lives, what is the survival function of
the status x : y?
Exercise 10–2. What is survival function of x : n ?
Occasionally, even the order in which death occurs is important. The status
1
x : n is a status which dies at the time of death of (x) if the death of (x) occurs before
time n. Otherwise, this status never dies.
1
Exercise 10–3. Under what circumstances does x : n die?

§10: Status 37
Problems
1
Problem 10–1. Find a formula for the survival function of x : n in terms of the
survival function of (x).
1
Problem 10–2. If the UDD assumption is valid for (x), does UDD hold for x : n ?
1
Problem 10–3. Find a formula for the survival function of x : n .
1
Problem 10–6. If the UDD assumption is valid for each of (x) and (y) and if (x)
and (y) are independent lives, does UDD hold for x : y?
§10: Status 38
1 1
Problem 10–1. P[T(x : n ) ≥ t] = t px for 0 ≤ t < n and P[T(x : n ) ≥ t] = n px
for t ≥ n.
1 1
Problem 10–2. The UDD assuption holds for x : n if and only if P[T(x : n ) ≥
1 1
k + t] = (1 − t)P[T(x : n ) ≥ k] + tP[T(x : n ) ≥ k + 1] for all integers k and all
0 ≤ t ≤ 1. Now use the formula for the survival function found in the previous
problem.
§10: Status 39
Exercise 10–1. The joint life status survives t time units if and only if both (x)
and (y) survive t time units. Using the independence gives s(t) = t px t py .
Exercise 10–2. Since a constant random variable is independent of any other

random variable, s(t) = t px t pn = t px if t ≤ n and 0 if t > n, by using the previous
exercise.
1
Exercise 10–3. The status x : n dies at time n if (x) is still alive at time n,
otherwise this status never dies.
§11. Valuing Contingent Payments
Earlier, the central theme of these notes was asserted to be embodied in the
question, “What is the value today of a random sum of money which will be paid
at a random time in the future?” This question can now be answered. Suppose the
random amount of money is denoted by A and the random time at which it will be paid
is denoted by T. The value of this payment today is computed in two steps. First, an
expression for the present value of the payment is written. This expression will be a
random variable. Here the present value is AvT . Then the expectation of this random
variable is computed. This expectation, E[AvT ], is the value today of the random
future payment. The interpretation of this amount, E[AvT ], is as the average present
value of the actual payment. Averages are reasonable in the insurance context since,
from the company’s point of view, there are many probabilistically similar policies
for which the company is obliged to pay benefits. The average cost (and income) per
policy is therefore a reasonable starting point from which to determine the premium.
The expected present value is usually referred to as the actuarial present value
in the insurance context. In the next few sections the actuarial present value of
certain standard parts of insurance contracts are computed.

§12. Life Insurance
In the case of life insurance the determination of the value of the insurance
depends on the random time of death of the insured. The amount that is paid at the
time of death is usually fixed by the policy and is non-random. Assume that the
force of interest is constant and known to be equal to δ . Also simply write T = T(x)
whenever clarity does not demand the full notation. The actuarial present value of
an insurance which pays 1 at the time of death is then
E[vT ]
by the priciple above. Intuitively, the actuarial present value of the benefit is the
single premium payment that an insurance company with no operating expenses
and no desire for profit would charge today in order to provide the benefit payment.
For this reason the actuarial present value of a benefit is also called the net single
premium. The net single premium would be the idealized amount an insured would
pay as a lump sum (single premium) at the time that the policy is issued. The case
of periodic premium payments will be discussed later.
A catalog of the various standard types of life insurance policies and the standard
notation for the associated net single premium follows. In most cases the benefit
amount is assumed to be $1, and in all cases the benefit is assumed to be paid at the
time of death. Keep in mind that a fixed constant force of interest is also assumed
and that v = 1/ (1 + i) = e−δ .
Insurances Payable at the Time of Death
Type Net Single Premium
n-year pure endowment A 1 = n Ex = E[vn 1(n,∞) (T)]
x:n
n-year term A1x:n = E[vT 1[0,n] (T)]
whole life Ax = E[vT ]
n-year endowment Ax:n = E[vT∧n ]
T
m-year deferred n-year term m| n Ax = E[v 1(m,n+m] (T)]
whole life increasing mthly (IA)(m) T

x = E[v [Tm + 1]/ m]
n-year term increasing annually (IA)1x:n = E[vT [T + 1]1[0,n) (T)]
n-year term decreasing annually (DA)1x:n = E[vT (n − [T])1[0,n) (T)]
Using this table it is a simple matter to compute the net single premium in the
various cases. The bar is indicative of an insurance paid at the time of death, while
the subscripts denote the status whose death causes the insurance to be paid. These
insurances are now reviewed on a case-by-case basis.
§12: Life Insurance 42
The first type of insurance is n-year pure endowment insurance which pays
the full benefit amount at the end of the nth year if the insured survives at least n
years. The notation for the net single premium for a benefit amount of 1 is A 1 (or
x:n
occasionally in this context n Ex ). The net single premium for a pure endowment is
just the actuarial present value of a lump sum payment made at a future date. This
differs from the ordinary present value simply because it also takes into account the
mortality characteristics of the recipient.
Exercise 12–1. Show that n Ex = vn n px .
The second type of insurance is n-year term insurance. The net single premium
with a benefit of 1 payable at the time of death for an insured (x) is denoted A1x:n .
This type insurance provides for a benefit payment only if the insured dies within n
years of policy inception.
The third type of insurance is whole life in which the full benefit is paid no
matter when the insured dies in the future. The whole life benefit can be obtained
by taking the limit as n → ∞ in the n-year term insurance setting. The notation for
the net single premium for a benefit of 1 is Ax .
Exercise 12–2. Suppose that T(x) has an exponential distribution with mean 50. If
the force of interest is 5%, find the net single premium for a whole life policy for
(x), if the benefit of $1000 is payable at the moment of death.
Exercise 12–3. Show that Ax = A1x:n + vn n px Ax+n by conditioning on the event

T(x) ≥ n and also by direct reasoning from a time diagram by looking at the
difference of two policies.
The fourth type of insurance, n-year endowment insurance, provides for the
payment of the full benefit at the time of death of the insured if this occurs before
time n and for the payment of the full benefit at time n otherwise. The net single
premium for a benefit of 1 is denoted Ax:n .
Exercise 12–4. Show that Ax:n = A1x:n + A 1 .

x:n
Exercise 12–5. Use the life table to find the net single premium for a 5 year pure
endowment policy for (30) assuming an interest rate of 5%.
The m-year deferred n-year term insurance policy provides provides the same
benefits as n year term insurance between times m and m + n provided the insured
lives m years.
All of the insurances discussed thus far have a fixed constant benefit. Increasing
whole life insurance provides a benefit which increase linearly in time. Similarly,
increasing and decreasing n-year term insurance provides for linearly increasing
(decreasing) benefit over the term of the insurance.
Corresponding to the insurances payable at the time of death are the same type
of policies available with the benefit being paid at the end of the year of death. The
only difference between these insurances and those already described is that these
insurances depend on the distribution of the curtate life variable K = K(x) instead
of T. The following table introduces the notation.
Insurances Payable the End of the Year of Death
Type Net Single Premium
n-year term A1x:n = E[vK+1 1[0,n) (K)]
whole life Ax = E[vK+1 ]
n-year endowment Ax:n = E[v(K+1)∧n ]
m-year deferred n-year term = E[vK+1 1[m,n+m) (K)]
m| n Ax
whole life increasing annually (IA)x = E[vK+1 (K + 1)]

n-year term increasing annually (IA)1x:n = E[vK+1 (K + 1)1[0,n) (K)]
n-year term decreasing annually (DA)1x:n = E[vK+1 (n − K)1[0,n) (K)]
These policies have net single premiums which can be easily computed from
the information in the life table. The primary use for these types of policies is the
computational connection between them and the ‘continuous’ policies described
above. To illustrate the ease of computation when using a life table observe that
from the definition
∞
∞
dx+k
Ax = k+1
v k px qx+k = vk+1 .
k=0 k=0 lx
In practice, of course, the sum is finite. Similar computational formulas are readily
obtained in the other cases.
Exercise 12–6. Show that A 1 = A 1 and interpret the result verbally. How would
x:n x:n
you compute A 1 using the life table?
x:n
Under the UDD assumption it is fairly easy to find formulas which relate the
insurances payable at the time of death to the corresponding insurance payable at
the end of the year of death. For example, in the case of a whole life policy
Ax = E[e−δ T(x) ]
= E[e−δ (T(x)−K(x)+K(x)) ]
= E[e−δ (T(x)−K(x)) ] E[e−δ K(x) ]
1
= (1 − e−δ )eδ E[e−δ (K(x)+1) ]
δ
i
= Ax
δ
where the third equality springs from the independence of K(x) and T(x) − K(x)
under UDD, and the fourth equality comes from the fact that under UDD the
random variable T(x) − K(x) has the uniform distribution on the interval (0,1).
Exercise 12–7. Can similar relationships be established for term and endowment
policies?
Exercise 12–8. Use the life table to find the net single premium for a 5 year
endowment policy for (30) assuming an interest rate of 5%.
Exercise 12–9. An insurance which pays a benefit amount of 1 at the end of the
mth part of the year in which death occurs has net single premium denoted by A(m)
x .
Show that under UDD i Ax = δ Ax .
(m) (m)
One consequence of the exercise above is that only the net single premiums for
insurances payable at the end of the year of death need to be tabulated, if the UDD
assumption is made. This leads to a certain amount of computational simplicity.
Problems
Problem 12–1. Write expressions for all of the net single premiums in terms of
either integrals or sums. Hint: Recall the form of the density of T(x) and K(x).
Problem 12–2. Show that δ A1x:n = iA1x:n , but that δ Ax:n ≠ iAx:n , in general.
Problem 12–3. Use the life table and UDD assumption (if necessary) to compute
A21 , A21:5 , and A1 .
21:5
Problem 12–4. Show that

dAx
= −v(IA)x .
di
Problem 12–5. Assume that DeMoivre’s law holds with ω = 100 and i = 0.10.
Find A30 and A30 . Which is larger? Why?
Problem 12–6. Suppose µx+t = µ and i = 0.10. Compute Ax and A1x:n . Do your
answers depend on x? Why?
Problem 12–7. Suppose Ax = 0.25, Ax+20 = 0.40, and Ax:20 = 0.55. Compute A 1
x:20
and A1x:20 .
(IA)x = vqx + v[Ax+1 + (IA)x+1 ]px .
What assumptions (if any) did you make?
Problem 12–9. What change in Ax results if for some fixed n the quantity qx+n is
replaced with qx+n + c?
Problem 12–10. What is the change in Ax if δ is replaced by δ + c? If µ is replaced

by µ + c?
Problem 12–1. The densities required are fT(x) (t) = t px µx+t and fK(x) (k) =
k px qx+k respectively.
Problem 12–2. δ A1 = δ (Ax − e−δ n n px Ax+n ) = iAx − ivn n px Ax+n = iA1 .

x:n x:n
Problem 12–3. Use δ A21 = iA21 , A21:5 = A1 + 5 E21 and the previous problem.
21:5
Problem 12–4. Just differentiate under the expectation in the definition of Ax .
Problem 12–5. Clearly A30 > A30 since the insurance is paid sooner in the
continuous case. Under DeMoivre’s law the UDD assumption is automatic and
1
70 −δ t
A30 = 70 0 e dt.
Problem 12–6. The answers do not depend on x since the lifetime is exponential
and therefore ageless.
Problem 12–7. The two relations Ax = A1 + v20 n px Ax+20 and Ax:20 = A1 +

x:20 x:20
vn n px along with the fact that A 1 = vn n px give two equations in the two sought
x:20
after unknowns.
Problem 12–8. Either the person dies in the first year, or doesn’t. If she doesn’t
buy an increasing annually policy for (x + 1) and a whole life policy to make up
for the increasing part the original policy would provide.
Problem 12–9. The new benefit is the old benefit plus a pure endowment
benefit of cv at time n.
Exercise 12–1. Since if the benefit is paid, the benefit payment occurs at time
n, n Ex = E[vn 1[n,∞) (T(x))] = vn P[T(x) ≥ n] = vn n px .
Exercise 12–2. Under the assumptions given the net single premium is
∞
E[1000vT(x) ] = 0 1000e−0.05t (1/ 50)e−t/ 50 dt = 285.71.
Exercise 12–3. For the conditioning argument, break the expectation into two
pieces by writing Ax = E[vT ] = E[vT 1[0,n] (T)] + E[vT 1(n,∞) (T)]. The first expecta-
tion is exactly A1 . For the second expectation, using the Theorem of Total Ex-
x:n
pectation gives E[vT 1(n,∞) (T)] = E[E[vT 1(n,∞) (T)| T ≥ n]]. Now the conditional
distribution of T given that T ≥ n is the same as the unconditional distribution of
T(x+n)+n. Using this fact gives the conditional expectation as E[vT 1(n,∞) (T)| T ≥
n] = E[vT(x+n)+n 1(n,∞) (T(x + n) + n)]1(n,∞) (T) = vn Ax+n 1(n,∞) (T). Taking expecta-
tions gives the result. To use the time diagram, imagine that instead of buying
a whole life policy now, the insured pledges to buy an n year term policy now,
and if alive after n years, to buy a whole life policy at time n (at age x + n). This
will produce the same result. The premium for the term policy paid now is A1
x:n
and the premium for the whole life policy at time n is Ax+n . This latter premium
is only paid if the insured survives, so the present value of this premium is the
second term in the solution.
Exercise 12–4. Using the definition and properties of expectation gives Ax:n =
E[vT 1[0,n] (T) + vn 1(n,∞) (T)] = E[vT 1[0,n] (T)] + E[vn 1(n,∞) (T)] = A1 + A 1 .
x:n x:n
Exercise 12–5. The net single premium for the pure endowment policy is
v5 5 p30 = (1.05)−5l35 / l30 = (1.05)−595808/ 96477 = 0.778.
Exercise 12–6. A 1 = E[vn 1[n,∞) (T)] = E[vn 1[n,∞) (K)] = A 1 = vn n p x =

x:n x:n
vn lx+n / lx .
Exercise 12–7. Since term policies can be expressed as a difference of premi-

ums for whole life policies, the answer is yes.
Exercise 12–8. The net single premium for a pure endowment policy is
v5 5 p30 = (1.05)−5l35 / l30 = (1.05)−595808/ 96477 = 0.778. For the endowment
policy, the net single premium for a 5 year term policy must be added to this
amount. From the relation given earlier, A1 = A30 −v5 5 p30 A35 . The relationship
30:5
between insurances payable at the time of death and insurances payable at the
end of the year of death is used to complete the calculation.
Exercise 12–9. Notice that [mT(x)] is the number of full mths of a year
that (x) lives before dying. (Here [a] is the greatest integer function.) So the
number of mths of a year that pass until the benefit for the insurance is paid
is [mT(x)] + 1, that is, the benefit is paid at time ([mT(x)] + 1)/ m. From here
the derivation proceeds as above. A(m) x = E[v([mT]+1)/ m ] = E[v([m(T−K+K)]+1)/ m ] =
K
E[v ]E[v ([m(T−K)]+1)/ m
]. Now T − K has the uniform distribution on the interval
(0, 1) under UDD, so [m(T − K)] has the uniform distribution over the integers
0,. . . , m − 1. So E[v([m(T−K)]/ m ] = m−1j=0 v j/ m
× (1/ m) = (1/ m)(1 − v)/ (1 − v1/ m )
from the geometric series formula. Substituting this in the earlier expression
−1 1/ m
gives A(m)
x = Ax v v (1/ m)(1 − v)/ (1 − v1/ m) = Ax δ / i(m) since i(m) = m(v−1/ m − 1).
§13. Laboratory 4
1. Show that Ax = vqx + vpx Ax+1 . Derive a similar formula for Ax .
2. The one step recursion formulas derived in problem 1 are especially useful
for computational purposes. The formulas are used to work backwards from large
attained ages to smaller ones, since at large attained ages everyone is dead and
the net premium for the insurance must be zero. Use the values of qx given in
Laboratory 3 and i = 5% to compute the values of Ax and Ax for x = 1 to x = 106.
Place the result of your computations into a nice table.

§14. Life Annuities
The basic study of life insurance concludes by developing techniques for un-
derstanding what happens when premiums are paid monthly or annually instead of
just when the insurance is issued. In the non–random setting a sequence of equal
payments made at equal intervals in time was referred to as an annuity. Here
interest centers on annuities in which the payments are made (or received) only as
long as the insured survives.
An annuity in which the payments are made for a non–random period of time
is called an annuity certain. From the earlier discussion, the present value of an
annuity immediate (payments begin one period in the future) with a payment of 1
in each period is
n
1 − vn
an = vj =
j=1 i
while the present value of an annuity due (payments begin immediately) with a
payment of 1 in each period is

n−1
1 − vn 1 − vn
än = j
v = = .
j=0 1−v d
These formulas will now be adapted to the case of contingent annuities in which
payments are made for a random time interval.
Suppose that (x) wishes to buy a life insurance policy. Then (x) will pay a
premium at the beginning of each year until (x) dies. Thus the premium payments
represent a life annuity due for (x). Consider the case in which the payment amount
is 1. Since the premiums are only paid annually the term of this life annuity depends
only on the curtate life of (x). There will be a total of K(x) + 1 payments, so the
actuarial present value of the payments is äx = E[äK(x)+1 ] where the left member is
a notational convention. This formula gives
1 − vK(x)+1 1 − Ax
äx = E[äK(x)+1 ] = E[ ]=
d d
as the relationship between this life annuity due and the net single premium for a
whole life policy. A similar analysis holds for life annuities immediate.
Exercise 14–1. Compute the actuarial present value of a life annuity immediate.
What is the connection with a whole life policy?
Exercise 14–2. A life annuity due in which payments are made m times per year
and each payment is 1/ m has actuarial present value denoted by ä(m)
x . Show that
(m) (m) (m)
Ax + d äx = 1.
§14: Life Annuities 51
Example 14–1. The Mathematical Association of America offers the following
alternative to members aged 60. You can pay the annual dues and subscription rate
of $90, or you can become a life member for a single fee of $675. Life members
are entitled to all the benefits of ordinary members, including subscriptions. Should
one become a life member? To answer this question, assume that the interest rate is
5% so that the Life Table at the end of the notes can be used. The actuarial present
value of a life annuity due of $90 per year is
1 − A60 1 − 0.412195
90 = 90 = 1110.95.
1−v 1 − 1/ 1.05
Thus one should definitely consider becoming a life member.
Exercise 14–3. What is the probability that you will get at least your money’s worth
if you become a life member? What assumptions have you made?
Pension benefits often take the form of a life annuity immediate. Sometimes
one has the option of receiving a higher benefit, but only for a fixed number of years
or until death occurs, whichever comes first. Such an annuity is called a temporary
life annuity.
Example 14–2. Suppose a life annuity immediate pays a benefit of 1 each year
for n years or until (x) dies, whichever comes first. The symbol for the actuarial
present value of such a policy is ax:n . How does one compute the actuarial present
value of such a policy? Remember that for a life annuity immediate, payments are
made at the end of each year, provided the annuitant is alive. So there will be a

total of K(x) ∧ n payments, and ax:n = E[ K(x)∧n j=1 vj ]. A similar argument applies
in the case of an n year temporary life annuity due. In this case, payments are
made at the beginning of each of n years, provided the annuitant is alive. In this
(K(x)+1)∧n
case äx:n = E[ K(x)∧(n−1)
j=0 vj ] = E[ 1−v d ] where the left member of this equality
introduces the notation.
Exercise 14–4. Show that Ax:n = 1 − d äx:n . Find a similar relationship for ax:n .
Especially in the case of pension benefits it is more realistic to assume that the
payments are made monthly. Suppose payments are made m times per year. In this
case each payment is 1/ m. One could begin from first principles (this makes a good
exercise), but instead the previously established facts for insurances together with
the relationships between insurances and annuities given above will be used. Using
the obvious notation gives
1 − A(m) x
ä(m)
x = (m)
d
1 − i(m)i
Ax
= (m)
d
1 − i(m)i
(1 − d äx )
=
d (m)
id i(m) − i
= ä x +
i(m) d (m) i(m) d (m)
where at the second equality the UDD assumption was used.
Exercise 14–5. Find a similar relationship for an annuity immediate which pays
1/ m m times per year.
A useful idealization of annuities payable at discrete times is an annuity payable

continuously. Such an annuity does not exist in the ‘real world’, but it provides a
useful connecting bridge between certain types of discrete annuities. Suppose that
the rate at which the benefit is paid is constant and is 1 per unit time. Then during
the time interval (t, t + dt) the amount paid is dt and the present value of this amount
is e−δ t dt. Thus the present value of such a continuously paid annuity over a period
of n years is
n
1 − e−δ n
an = e−δ t dt = .
0 δ
A life annuity which is payable continuously will thus have actuarial present value
1 − e−δ T(x)
ax = E[aT(x) ] = E[ ].
δ
Exercise 14–6. Show that Ax = 1 − δ ax . Find a similar relationship for ax:n .
This point of view makes it easy to understand certain modifications of discrete

annuities. When annuity benefit payments are made at discrete intervals it is
customary to provide a final adjustment which takes into account the death of the
annuitant between payment periods. One such modified annuity is called a complete
annuity immediate whose actuarial present value is denoted by åx . The payment
scheme for such an annuity is $1 at the end of each of the years 1 through K(x) − 1
plus a final adjustment payment made at the time of death in order to make this
scheme actuarially equivalent to a continuous annuity. To determine the size of the
adjustment payment, first determine the rate at which a continuous annuity must
be paid in order to be equivalent to a discrete annuity immediate. The following
picture illustrates the cash flows.
Equivalence of an Annuity Immediate and a Continuous Annuity
1 σ
....................................................................................................................................................... .......................................................................................................................................................
0 1 0 1
1
Equating the present value of the two cash streams gives v = σ e−δ t dt from
0
which σ = δ / i in order for the streams to be equivalent. It follows that the amount
of the final payment, made at time T(x), for the complete annuity immediate must
be T(x) T(x)−K(x)
δ T(x) δ −δ t δ δt
e e dt = e dt.
K(x) i 0 i
Also T(x)
δ −δ t δ
åx = E[ e dt] = ax .
0 i i
Exercise 14–7. When payments are made on an mthly basis (each payment being
1/ m) the actuarial present value of a complete annuity immediate is denoted by å(m)
x .
Find a formula for the adjustment payment and the actuarial present value in this
case.
Premium payments to an insurance company often take the form of an appor-

tioned annuity due. Here the payment scheme consists of $1 immediately and at
the beginning of each year through and including time K(x) less a refund payment
(from the company) at the time of death. The refund payment is made because
(in the premium context) the insured payed for a full year of coverage in advance.
Using techniques similar to those above shows that the refund payment, made at
time T(x), should be in the amount of
K(x)+1 K(x)+1−T(x)
δ −δ t δ −δ t
eδ T(x) e dt = e dt
T(x) d 0 d
and that the actuarial present value of an apportioned annuity due is
T(x)
δ −δ t δ
ä{1}
x = E[ e dt] = ax .
0 d d
Here, of course, the left most member of the equalities is the notational convention.
Exercise 14–8. In the case of mthly payments (each of size 1/ m) find a formula for
ä{m}
x as well as for the size of the refund payment.
In order to compare apportioned and complete annuities, let us see how pre-
miums paid by a complete annuity immediate would operate. In such a scheme,
premiums would be paid at the end of each year, except that in the year of death
a reduced premium would be paid at the time of death. When viewed from the
insurer’s viewpoint in this way it is obvious that ä{1}
x > åx .
There is one other idea of importance. In the annuity certain setting one may be
interested in the accumulated value of the annuity at a certain time. For an annuity
due for a period of n years the accumulated value of the annuity at time n, denoted
by s̈n , is given by s̈n = (1 + i)n än = (1+i)d −1 . The present value of s̈n is the same as
n
the present value of the annuity. Thus the cash stream represented by the annuity is
equivalent to the single payment of the amount s̈n at time n. This last notion has an
analog in the case of life annuities. In the life annuity context
n Ex s̈x:n = äx:n
where n Ex = vn n px is the actuarial present value (net single premium) of a pure

endowment of $1 payable at time n. Thus s̈x:n represents the amount of pure
endowment payable at time n which is actuarially equivalent to the annuity.
Problems
Problem 14–1. Show that under UDD
x < ax < ⋅ ⋅ ⋅ < ax < ⋅ ⋅ ⋅ < äx < äx < äx .
ax < a(2) (3) (3) (2)
Give an example to show that without the UDD assumption the inequalities may
fail.

{3} {2} {1}
x < åx < ⋅ ⋅ ⋅ < ax < ⋅ ⋅ ⋅ < äx < äx < äx .
åx < å(2) (3)
Problem 14–3. Show that for any m we have ä{m}

x < ä(m)
x and that ax < åx .
(m) (m)
1
Problem 14–4. True or false: A1x:n = 1 − d ä1x:n . Hint: When does x : n die?
Problem 14–5. True or false: s̈x:n ≤ s̈n .
Problem 14–6. Use the life table to calculate the actuarial present value of $1000
due in 30 years if (40) survives.
Problem 14–7. Use the life table to compute a21 and ä{4}
21 .
Problem 14–8. Find a general formula for m| n äx and use it together with the life
table to compute 5| 10 ä20 .
Problem 14–9. Prove ax:n = 1 Ex äx+1:n .
Problem 14–10. Show that under UDD
x = α (m)äx − β (m).
ä(m)
id i − i(m)
Here α (m) = (m) (m) and β (m) = (m) (m) . The functions α (m) and β (m) defined
i d i d
here are standard actuarial functions.
Problem 14–11. Show that δ (Ia)T + TvT = aT .
Problem 14–12. Use the previous problem to show that δ (Ia)x + (IA)x = ax . Here
(Ia)x is the actuarial present value of an annuity in which payments are made at rate
t at time t. Is there a similar formula in discrete time?
Problem 14–13. Show that äx = 1 + ax and that 1

m
+ a(m) (m) 1 n
x:n = äx:n + m v n px .
Problem 14–14. Show that äx:n = äx − vn n px äx+n and use this to compute ä21:5 .
Problem 14–1. As the type of annuity varies from left to right, the annuitant
receives funds sooner and thus the present value is higher.
Problem 14–2. Since å(m)x = δ / i ax the result follows for the earlier relation-
(m)
ship between the rates of interest. A similar argument resolves the other half of
the inequalities.
Problem 14–3. The difference betwee the two sides of the inequalities is the
amount of the refund (or extra) payment.
Problem 14–4. The status dies only if (x) dies before time n. The result is true.
Problem 14–5. False.
Problem 14–8. m| n äx = m Ex äx+m − m+n Ex äx+m+n .
Problem 14–10. See the text above.
Problem 14–11. Use integration by parts starting with the formula δ (Ia)T =
T
δ t e−δ t dt.
0
K(x) 1−v−1 Ax
Exercise 14–1. In this case, E[aK(x) ] = E[ 1−vi ]= i .
Exercise 14–2. Here there are [mT] +j/ m1 payments, so using the geometric series
formula gives ä(m) j=0 (1/ m)v ] = E[(1/ m)(1 − v
= E[ [mT] )/ (1 − v1/ m )].
([mT]+1)/ m
x
Now m(1 − v ) = d , which gives the result.
1/ m (m)
Exercise 14–3. To get your money’s worth, you must live long enough so that
the present value of the annual dues you would pay if you were not a life member
will exceed $675. This gives a condition that K(60) must satisfy if you are to
get your moneys worth.
(K+1)∧n (K+1)∧n
Exercise 14–4. For the first one äx:n = E[ 1−v d ] = E[ 1−v d 1[0,n−1] (K)] +
d 1[n,∞) (K)] = (1/ d)(1 − Ax:n ).
(K+1)∧n (K+1) n
E[ 1−v d 1[n,∞) (K)] = E[ 1−vd 1[0,n−1] (K)] + E[ 1−v
A similar argument shows that ax:n = (1/ i)(A1 + n px vn+1 ).
x:n
Exercise 14–5. The argument proceeds in a similar way, beginning with the
1−v−1/ m A(m)
relation a(m)
x = i(m)
x
.
Exercise 14–6. The first relationship follows directly from the given equation
and the fact that Ax = E[e−δ T(x) ]. Since T(x : n ) = T(x) ∧ n a similar argument
gives ax:n = (1/ δ )(1 − Ax:n ).
Exercise 14–7. In this case the payment rate for the corresponding continuous
δ
annuity is δ / i(m) , which gives å(m) = i(m) ax and the adjustment payment as

δ T(x) T(x) (m) −δ t
T(x)−[mT(x)]/
x
[mT(x)]/ m (δ / i )e (δ / i(m) )eδ t dt.

m
e dt = 0
Exercise 14–8. Here the rate is δ / d (m) for the corresponding continuous annuity
[mT]/ m+1/ m δ −δ t
so that ä{m} = (δ / d (m) )ax and the refund payment is eδ T T d (m) e dt =
[mT]/ m+1/ m−T) δ −δ t
x
0 d (m) e dt.
§15. Laboratory 5
{m}
1. Show that äx = 1 + vpx äx+1 . Find formulas expressing ä(m)
x and äx in terms
of äx .
2. The one step recursion formulas for annuities can be used just like the one
step recursions for insurances themselves. Use the qx values from Laboratory 3 and
{12}
i = 5% and compute äx , ä(12)
x , and äx for x = 1 to x = 106. Place the result of your
computations into a nice table.

§16. Net Premiums
The common types of insurance policies can now be realistically analyzed from
an insurers point of view.
To develop the ideas consider the case of an insurer who wishes to sell a fully
discrete whole life policy which will be paid for by equal annual premium payments
during the life of the insured. The terminology fully discrete refers to the fact that
the benefit is to be paid at the end of the year of death and the premiums are to
paid on a discrete basis as well. How should the insurer set the premium? A first
approximation is given by the net premium. The net premium is found by using the
equivalence principle: the premium should be set so that actuarial present value
of the benefits paid is equal to the actuarial present value of the premiums received.
Using the equivalence principle the net premium P should satisfy
E[vK(x)+1 ] = PE[äK(x)+1 ]
or
Ax − Päx = 0.
From here it is easy to determine the net premium, which in this case is denoted Px .
Exercise 16–1. Use the life table to find the net premium, P30 , for (30) if i = 0.05.
The notation for other net premiums for fully discrete insurances parallel the
notation for the insurance policies themselves. For example, the net annual premium
for an n year term policy with premiums payable monthly is denoted P(12)1 .
x:n
Exercise 16–2. Use the life table to find P1 . What is P(12)

1 ? What is P{1}
1 ?
30:10 30:10 30:10
Exercise 16–3. An h payment whole life policy is one in which the premiums are
paid for h years, beginnning immediately. Find a formula for h Px , the net annual
premium for an h payment whole life policy.
Example 16–1. As a more complicated example consider a recent insurance ad-

vertisement which I received. For a fixed monthly premium payment (which is
constant over time) one may receive a death benefit defined as follows:
100000 1[0,65) (K(x)) + 75000 1[65,75)(K(x))
+ 50000 1[75,80)(K(x)) + 25000 1[80,85) (K(x)).
What is the net premium for such a policy? Assume that the interest rate is 5% so
that the life table can be used for computations. Using the equivalence principle,
the net annual premium P is the solution of the equation
Pä(12)
x:85−x
= 100000 A1x:65−x + 75000 65−x| 10 Ax + 50000 75−x| 5 Ax + 25000 80−x| 5 Ax

§16: Net Premiums 60
in terms of certain term and deferred term insurances.
Exercise 16–4. Compute the actual net monthly premium for (21).
The methodology for finding the net premium for other types of insurance is
exactly the same. The notation in the other cases is now briefly discussed. The most
common type of insurance policy is one issued on a semi-continuous basis. Here
the benefit is paid at the time of death, but the premiums are paid on a discrete basis.
The notation for the net annual premium in the case of a whole life policy is P(Ax ).
The net annual premium for a semi-continuous term policy with premiums payable
mthly is P(m) (A1x:n ). The notation for other semi-continuous policies is similar.
Exercise 16–5. What type of policy has net annual premium P{m} (Ax:n )?
Policies issued on a fully continuous basis pay the benefit amount at the time
of death and collect premiums in the form of a continuous annuity. Obviously, such
policies are of theoretical interest only. The notation here is similar to that of the
semi-continuous case, with a bar placed over the P. Thus P(Ax ) is the premium rate
for a fully continuous whole life policy.
Problems
Problem 16–1. Show that if δ = 0 then P(Ax ) = 1/ e̊x .
Problem 16–2. Arrange in increasing order of magnitude: P(2) (A40:25 ), P(A40:25 ),

P{4} (A40:25 ), P(A40:25 ), P{12} (A40:25 ).
Problem 16–3. If P(Ax ) = 0.03 and if interest is at the effective rate of 5%, find
P{2}
x .
Problem 16–4. If 15 P45 = 0.038, P45:15 = 0.056 and A60 = 0.625 find P1 .
45:15
Problem 16–5. Recall that apportionable annuities differ from annuities due only
in the fact that the apportionable annuity offers the additional benefit of a ‘pre-
mium refund’. Let AxPR denote the net single premium for this refund benefit for a
continuous whole life policy with apportioned premiums payable annually. Show
that
P(Ax )
AxPR = (Ax − Ax ).
δ
Problem 16–6. Refer to the previous problem and show that
P(Ax )
P(AxPR ) = (Ax − Ax ).
δ äx
Problem 16–7. Use the equivalence principle to find the net annual premium for
a fully discrete 10 year term policy with benefit equal to $10,000 plus the return,
with interest, of the premiums paid. Assume that the interest rate earned on the
premiums is the same as the interest rate used in determining the premium. Use the
life table to compute the premium for this policy for (21). How does this premium
compare with 10000P1 ?
21:10
Problem 16–8. A level premium whole life insurance of 1, payable at the end
of the year of death, is issued to (x). A premium of G is due at the beginning
of each year provided (x) survives. Suppose L denotes the insurer’s loss when
G = Px , L∗ denotes the insurer’s loss when G is chosen so that E[L∗ ] = −0.20, and
Var(L) = 0.30. Compute Var(L∗ ).
Problem 16–9. A policy issued to (x) has the following features.

(1) Apportionable premiums are payable annually.
(2) The first premium is twice the renewal premium.
(3) Term insurance coverage for $100,000 plus the difference between the first
and second premium is provided for 10 years.
(4) An endowment equal to the first year premium is paid at the end of 10 years.
(5) Death claims are paid at the moment of death.
Use the equivalence principle to find an expression for the renewal net annual
premium.
Problem 16–10. A $1000 whole life policy is issued to (50). The premiums are
payable twice a year, and are calculated on an apportionable basis. The benefit is
payable at the moment of death. Calculate the semi-annual net premium given that
A50 = 0.3, δ = 0.07, and e−0.035 = 0.9656.
Problem 16–11. Polly, aged 25, wishes to provide cash for her son Tad, currently
aged 5, to go to college. Polly buys a policy which will provide a benefit in the
form of a temporary life annuity due (contingent on Tad’s survival) in the amount of
$25,000 per year for 4 years commencing on Tad’s 18th birthday. Polly will make 10
equal annual premium payments beginning today. The 10 premium payments take
the form of a temporary life annuity due (contingent on Polly’s survival). According
to the equivalence principle, what is the amount of each premium payment? Use
the life table and UDD assumption (if necessary).
Problem 16–12. Snow White, presently aged 21, wishes to provide for the welfare
of the 7 dwarfs in the event of her premature demise. She buys a whole life policy
which will pay $7,000,000 at the moment of her death. The premium payments for
the first 5 years will be $5,000 per year. According to the equivalence principle,
what should her net level annual premium payment be thereafter? Use the life table
and UDD assumption (if necessary).
Problem 16–13. The Ponce de Leon Insurance Company computes premiums for
its policies under the assumptions that i = 0.05 and µx = 0.01 for all x > 0. What
is the net annual premium for a whole life policy for (21) which pays a benefit
of $100,000 at the moment of death and has level apportioned premiums payable
annually?
Problem 16–2. This is really a question about the present value of annuities.
Problem 16–3. Since iAx = δ Ax and d (2) ä{2} {2}

x = δ ax , Px = (d / i)P(Ax ).
(2)
Problem 16–4. Use the two equations P45:15 = P1 + 15 E45 / ä45:15 and
45:15
A45 − 15 E45 A60
P1 = = 15 P45 − 15 E45 A60 / ä45:15 with the given information.
45:15 ä45:15
Problem 16–7. The present value of the benefit is 10000vK+11[0,10) (K) +

pvK+1 s̈K+1 1[0,10) (K) where p is the premium.
Problem 16–8. The loss random variable is (1 + G/ d)vK+1 − G/ d from which

the mean and variance in the two cases can be computed and compared.
Problem 16–10. The annual premium p satisfies pä{2}

50 = 1000A50 .
Problem 16–11. The premium p satisfies pä25:10 = 2500013| 4ä5 .
Problem 16–12. The premium p satisfies 7000000A21 = 5000ä21:5 + p5 E21 ä26 .

Exercise 16–1. From the table, P30 = A30 / ä30 = 0.133866/ 18.189 = 0.0735.
Exercise 16–2. Now P1 = A1 / ä30:10 . Also A1 = A30 − v10 10 p30 A40 =

30:10 30:10 30:10
0.133866 − (1.05)−10(94926/ 96477)(0.201506) = 0.01214. Similarly, ä30:10 =
ä30 − v10 10 p30 ä40 = 8.060, giving the premium as 0.001506. The other two
premiums differ only in the denominator, since P(12) 1 = A1 / ä(12)
30:10
and
30:10 30:10
P{1}
1 = A1 / ä{1}
30:10
. Now ä(12)
30:10 30 − v 10 p30 ä40 .
= ä(12) 10 (12)
Since ä(12)
30 =
30:10 30:10
(id/ i(12) d(12) )ä30 + (i(12) − i)/ i(12) d(12) and a similar expression holds for ä(12)
40 ,
the value of the annuity can be computed from the life table. A similar argument
and the fact that ä{1} x = (i/ δ )äx + (δ − i)/ δ d allows the computation of the value
of the apportioned annuity. Note that the UDD assumption has been used here.
Exercise 16–3. h Px = Ax / äx:h .
Exercise 16–4. The net monthly premium is P/ 12 where P = (100000A1 +

21:44
75000v4444 p21 A1 + 50000v5454 p21 A1 + 25000v5959 p21 A1 )/ ä(12)
21:64
. These
65:10 75:5 80:5
values can be computed from the life table using the techniques of an earlier
exercise.
Exercise 16–5. This is the premium for a continuous endowment policy with
mthly apportioned premiums.
§17. Laboratory 6
1. A 40 year term insurance policy with semi-annual premiums is issued to

(30). Assume that i = 5% and mortality follows the table given in Laboratory 3.
Compute the semi-annual net premium if the benefit amount is 100,000 and the
benefit is paid at the moment of death.
2. Rework the computation in problem 1 for an interest rate of 4% and 6%.

How much does the change in interest rate change the semi-annual premium?
3. Suppose the benefit amount of the policy in problem 1 is 100,000 for death
in the first 30 years (until age 60), and then decreases by 5,000 per year for each of
the remaining years. What is the semi-annual premium?

§18. Insurance Models Including Expenses
A more realistic view of the insurance business includes provisions for expenses.
The profit for the company can also be included here as an expense.
The common method used for the determination of the expense loaded pre-
mium (or the gross premium) is a modification of the equivalence principle. Ac-
cording to the modified equivalence principle the gross premium G is set so that
on the policy issue date the actuarial present value of the benefit plus expenses is
equal to the actuarial present value of the premium income. The premium is usually
assumed to be constant. Under these assumptions it is fairly easy to write a formula
to determine G. Assume that the expenses in policy year k are ek−1 and are paid at
time k − 1, that is, at the beginning of the year. The actuarial present value of the
expenses is then given by

K(x) ∞

E[ k
v ek ] = vk ek k px .
k=0 k=0
Typically expenses are dependent on the premium. Also the sales commission is
usually dependent on the policy size.
Example 18–1. Suppose that the first year expenses for a $100,000 semi-continuous
whole life policy are 20% of premiums plus a sales commission equal to 0.5% of
the policy amount, and that the expenses for subsequent years are 10% of premium
plus $5. The gross premium G for such a policy satisfies
100000Ax + (0.20G + 500) + (0.10G + 5)ax = Gäx .
An important, and realistic, feature of the above example is the large amount of
first year expense. Expenses are now examined in greater detail.
Example 18–2. Let’s look at the previous example in the case of a policy for a
person aged 21. Assume that the interest rate is 5% and that the life table applies.
Then
100, 000A21 + 495 + 5ä21
G= = $604.24.
0.9ä21 − 0.1
From this gross premium the company must pay $500 in fixed expenses plus 20%
of the gross premium in expenses ($120.85), plus provide term insurance coverage
for the first year, for which the net single premium is 100, 000A1 = $123.97. Thus
21:1
there is a severe expected cash flow strain in the first policy year! The interested
reader may wish to examine the article “Surplus Loophole” in Forbes, September
4, 1989, pages 44-48.
§18: Insurance Models Including Expenses 67
Expenses typically consist of two parts. The first part of the expenses can be
expressed as a fraction of gross premium. These are expenses which depend on
policy amount, such as sales commission, taxes, licenses, and fees. The other part
of expenses consist of those items which are independent of policy amount such
as data processing fees, printing of actual policy documents, clerical salaries, and
mailing expenses.
Studying the gross premium as a function of the benefit provided can be useful.
Denote by G(b) the gross premium for a policy with benefit amount b. The value
G(0) represents the overhead involved in providing the policy and is called the
policy fee. Typically the policy fee is not zero. The ratio G(b)/ b is called the
premium rate for a policy of benefit b and reflects (approximately) the premium
change per dollar of benefit change when the benefit amount is b.
Exercise 18–1. In the example above find R(b), the premium rate for a policy of
benefit b.
Problems
Problem 18–1. The expense loaded annual premium for an 35 year endowment
policy of $10,000 issued to (30) is computed under the assumptions that
(1) sales commission is 40% of the gross premium in the first year
(2) renewal commissions are 5% of the gross premium in year 2 through 10
(3) taxes are 2% of the gross premium each year
(4) per policy expenses are $12.50 per 1000 in the first year and $2.50 per 1000
thereafter
(5) i = 0.05
Find the gross premium using the life table.
Problem 18–2. A semi-continuous whole life policy issued to (21) has the following
expense structure. The first year expense is 0.4% of the policy amount plus $50. The
expenses in years 2 through 10 are 0.2% of the policy amount plus $25. Expenses in
the remaining years are $25, and at the time of death there is an additional expense
of $100. Find a formula for G(b). Compute G(1) and compare it to A21 .
Problem 18–3. Your company sells supplemental retirement annuity plans. The
benefit under such a plan takes the form of an annuity immediate, payable monthly,
beginning on the annuitant’s 65th birthday. Let the amount of the monthly benefit
payment be b. The premiums for this annuity are collected via payroll deduction
at the end of each month during the annuitant’s working life. Set up expenses for
such a plan are $100. Subsequent expenses are $5 each month during the premium
collection period, $100 at the time of the first annuity payment, and $5 per month
thereafter. Find G(b) for a person buying the plan at age x. What is R(b)?
Problem 18–4. A single premium life insurance policy with benefits payable at the
end of the year of death is issued to (x). Suppose that
(1) Ax = 0.25
(2) d = 0.05
(3) Sales commission is 18% of gross premium
(4) Taxes are 2% of gross premium
(5) per policy expenses are $40 the first year and $5 per year thereafter
Calculate the policy fee that should be charged.

Problem 18–1. 10000A30:35 + 0.35G + 0.05Gä30:10 + (0.02G + 25)ä30:35 + 100 =
Gä30:35 .
Problem 18–2. bA21 + 0.002b + 25 + 0.002bä21:10 + 25ä21 + 100A21 = G(b)ä21.
(12)
Problem 18–3. 12b65−x| a(12) (12)
x +5×12äx +95+10065+ 1 −x Ex = G(b)×12ax:65−x .
12
Problem 18–4. G(b) = bAx + 0.18G(b) + 0.02G(b) + 35 + 5äx .

Exercise 18–1. Since the premium is bR(b) when the benefit is b, the modified
equivalence principle gives bAx + (0.20bR(b) + 0.005b) + (0.20bR(b) + 5)äx =
bR(b)äx from which R(b) = (bAx + 0.005b + 5äx )/ b(0.9äx − 0.20).
§19. Multiple Lives
The study of the basic aspects of life insurance is now complete. Two different
but similar directions will now be followed in the ensuing sections. On the one
hand, types of insurance in which the benefit is paid contingent on the death or
survival of more than one life will be examined. On the other hand, the effects of
competing risks on the cost of insurance will be studied.
The first area of study will be insurance in which the time of the benefit payment
depends on more than one life. Recall that a status is an artificially constructed
life form for which there is a definition of survival and death (or decrement). The
simplest type of status is the single life status. The single life status (x) dies exactly
when (x) does. Another simple status is the certain status n . This status dies at
the end of n years. The joint life status for the n lives (x1 ), . . . (xn ) is the status
which survives until the first member of the group dies. This status is denoted by
(x1 x2 . . . xn ). The last survivor status, denoted by (x1 x2 . . . xn ) is the status which
survives until the last member of the group dies.
When discussing a given status the question naturally arises as to how one
would issue insurance to such a status. If the constituents of the status are assumed
to die independently this problem can be easily solved in terms of what is already
known.
Example 19–1. Consider a fully discrete whole life policy issued to the joint status
(xy). The net annual premium to be paid for such a policy is computed as follows.
Using the obvious notation, the premium, P, must satisfy
Axy = Päxy .
Using the definition of the joint life status gives
Axy = E[vK(x)∧K(y)+1 ]
and
1 − Axy
äxy =
d
which are obtained as previously.
Exercise 19–1. Obtain an expression for Axy in terms that can be computed from
the life table.
Exercise 19–2. Is Axy + Axy = Ax + Ay ?
A useful technique for writing computational formulas directly is to ask the

question “Under what conditions is a payment made at time t?” The answer will
usually provide a computational formula.
§19: Multiple Lives 72
Example 19–2. What is äxy ? This annuity makes a payment of 1 at time k if and
only if both (x) and (y) are alive at time k, and the probability of this is k pxy = k px k py .
∞

Thus äxy = vk k px k py .
k=0
If one is willing to assume an analytical law of mortality computations involving

joint lives can be simplified. Recall that two of the common analytical laws of
mortality are the Gompertz and Makeham laws. The Gompertz law has force of
mortality µx = Bcx . It is easily seen that the joint survival of two independent lives
(x) and (y) is probabilistically identical with the survival of a single life (w) if and
only if
µ(xy)+s = µx+s + µy+s = µw+s .
When (x) and (y) have mortality which follows Gompertz’ Law this relation holds if
w satisfies cx + cy = cw . A similar observation applies to Makeham’s law for which
the force of mortality is µx = A + Bcx . In this case, however, it is necessary to mimic
the joint life (xy) by using a joint life (ww) at equal ages. Here w is the solution of
2cw = cx + cy .
Exercise 19–3. Verify these assertions.
A status can also be determined by the order in which death occurs. The idea
1
here is similar to that used for term insurance earlier in which the status x : n fails
at the time of death of (x) provided (x) dies first. As a more complicated example
2
the status (x : y) dies at the time of death of (y) provided (y) is the second to die.
Hence this status lives forever if (y) dies before (x). An insurance for such a status
is a simple case of what is known as a contingent insurance. Again, if the lives are
assumed to fail independently it is a simple matter to reduce computations involving
contingent insurance to the cases already considered.
Problems
Problem 19–1. Show
t pxy = t pxy + t px (1 − t py ) + t py (1 − t px ).
Problem 19–2. Suppose µx = 1/ (110 − x) for 0 ≤ x < 110. Find 10 p20:30 , 10 p20:30 ,
and e̊20:30 .
Problem 19–3. Find an expression for the actuarial present value of a deferred
annuity of $1 payable at the end of any year as long as either (20) or (25) is living
after age 50.
Problem 19–4. Find the actuarial present value of a 20 year annuity due which
provides annual payments of $50,000 while both (x) and (y) survive, reducing by
25,000 on the death of (x) and by 12,500 on the death of (y).
Problem 19–5. Show that n q1xy = n qxy2 + n qx n py .
Problem 19–6. Show that A1xy − Axy2 = Axy − Ay .
Problem 19–7. If µx = 1/ (100 − x) for 0 ≤ x < 100, calculate 25 q 2 .

25:50
Problem 19–8. In a mortality table which follows Makeham’s Law you are given
A = 0.003 and c10 = 3. Calculate ∞ q1 if e̊40:50 = 17.
40:50
Problem 19–9. If the probability that (35) will survive for 10 years is a and the
probability that (35) will die before (45) is b, what is the probability that (35) will
die within 10 years after the death of (45)? Assume the lives are independent.
Problem 19–10. Plot the survival function for Gompertz law of mortality with
A = 0.001 and c = 1.06.
Problem 19–11. Suppose (20) and (30) are independent lives that follow the Gom-
pertz law of mortality given in the previous problem. Plot the survival for the joint
life status 20 : 30. Is there a single age (x) whose survival function is the same as
the survival function of 20 : 30?
Problem 19–12. Plot the survival function for Makeham’s law of mortality with
A = 0.003, B = 0.001, and c = 1.06.
Problem 19–13. Suppose (20) and (30) are independent lives that follow the Make-
ham law of mortality given in the previous problem. Plot the survival for the joint
life status 20 : 30. Is there a single age (x) whose survival function is the same as
the survival function of 20 : 30?
Problem 19–1. t pxy = P[[T(x) ≥ t] ∪ [T(y) ≥ t]] = P[T(x) ≥ t, T(y) ≥
t] + P[T(x) ≥ t, T(y) ≤ t] + P[T(x) ≤ t, T(y) ≥ t].
Problem 19–2. From the form of the force of mortality, DeMoivre’s Law
holds.
Problem 19–3. 30| a20 + 25| a25 − 30| a20:25 .
Problem 19–4. The annuity pays 12,500 for 20 years no matter what so the
actuarial present value consists of 3 layers.
Problem 19–7. DeMoivre’s Law holds.

∞
Problem 19–8. Here ∞ e̊40:50 = 0 t p40:50 dt = 17.
∞ Also, by conditioning,
P[T(40) < T(50)] = 0 P[T(50) > t]fT(40)(t) dt = 0 t p40:50 µ40+t dt while by a
∞
symmetric argument P[T(40) > T(50)] = 0 t p40:50 µ50+t dt. Using the form of
the force of mortality under Makeham’s Law, the fact that these two probabilities
sum to 1, and the given information completes the argument.
∞
Problem
∞ 19–9. P[T(35) > T(45) + 10] = 0 P[T(35) > t +10]t p45 µ45+t dt =
∞
0 P[T(35) > t+10| T(35) ≥ 10]P[T(35) ≥ 10]t p45 µ45+t dt = a 0 P[T(45)+10 >
∞ ∞
t + 10]t p45 µ45+t dt = a 0 (t p45 )2 µ45+t dt = 0 −t p45 dtd t p45 dt = a/ 2. Thus the
desired probability is 1 − a/ 2 − b.
Exercise 19–1. Using the independence gives t pxy = t pxt py , so that Axy =
E[vK(xy)+1 ] = ∞k=0 vk+1 (k pxy − k+1 pxy ) = ∞k=0 vk+1 (k pxk py − k+1 pxk+1 py ).
Exercise 19–2. Intuitively, either (x) dies first or (y) dies first, so the equation
is true. This can be verified by writing the expectations in terms of indicators.
Exercise 19–3. Under Makeham the requirement is that (A+Bcx+s)+(A+Bcy+s ) =

(A + Bcw+s ) + (A + Bcw+s) for all s, and this holds if cx + cy = 2cw .
§20. Laboratory 7
1. Find an expression for the net single premium for a whole life policy issued
1
to (xy), a status which fails when (x) dies if T(x) < T(y). Use this expression and
the life table data of Laboratory 3 to compute the premium for a $1,000,000 policy
1
issued to (30 : 40). Use 6% as the interest rate.
2. Find an expression for the net single premium for a whole life policy issued to
2
(xy), where the benefit is paid on the death of (y) if T(x) < T(y). Use this expression
and the life table data of Laboratory 3 to compute the premium for a $1,000,000
2
policy issued to (30 : 40). Use 6% as the interest rate.
3. Show that if X and Y are independent random variables and one of them is
absolutely continuous then P[X = Y] = 0. Hence under the standard assumptions of
this section no two people can die simultaneously.
4. One model for joint lives which allows for simultaneous death is the common
shock model. The intuition is that the two lives behave almost independently except
for the possibility of death by a common cause. The model is as follows. Let T ∗ (x),
T ∗ (y), and Z be independent random variables. Assume that T ∗ (x) and T ∗ (y) have
the distribution of the remaining lifetimes of (x) and (y) as given by the life table.
The random variable Z represents the time of occurrence of the common catastrophe
which will kill any survivors. The common shock model is that the true remaining
lifetimes of (x) and (y) are given as T(x) = min{T ∗ (x), Z} and T(y) = min{T ∗ (y), Z}
respectively. What is the probability that (x) and (y) die simultaneously in this
model? What is the survival function for the joint life status (xy) in this model?
Answer these two questions in general, and then in the special case in which
T ∗ (x), T ∗ (y), and Z have exponential distributions with parameters µx , µy , and µz
respectively.

§21. Multiple Decrement Models
In contrast to the case in which a status is defined in terms of multiple lives, the
way in which a single life fails can be studied. This point of view is particularly
important in the context of the analysis of pension plans. In such a setting a person
may withdraw from the workforce (a ‘death’) due to accident, death, or retirement.
Different benefits may be payable in each of these cases. Another common type
of insurance in which a multiple decrement model is appropriate is in the double
indemnity life policy. Here the benefit is twice the face amount of the policy if
the death is accidental. In actuarial parlance the termination of a status is called a
decrement and multiple decrement models will now be developed. These models
also go by the name of competing risk models in other contexts.
To analyze the new situation, introduce the random variable J = J(x) which is a
discrete random variable denoting the cause of decrement of the status (x). Assume
that J(x) has as possible values the integers 1, . . . , m. It is clear that all of the
information of interest is contained in the joint distribution of the random variables
T(x) and J(x). Note that this joint distribution is of mixed type since (as always) T(x)
is assumed to be absolutely continuous while J(x) is discrete. The earlier notation is
modified in a fairly obvious way to take into account the new model. For example,
(j)
t qx = P[0 < T(x) ≤ t, J(x) = j].
and
(j)
t px = P[T(x) > t, J(x) = j].
Here ∞ q(j)
x gives the marginal density of J(x). To discuss the probability of death
due to all causes the superscript (τ ) is used. For example,

m
(τ )
t qx = P[T(x) ≤ t] = (j)
t qx
j=1
and similar expressions for the survival probability and the force of mortality can
be obtained. Although t qx(τ ) + t px(τ ) = 1 a similar equation for the individual causes
of death fails unless m = 1. For the force of mortality
(τ ) fT(x) (t)
µx+t =
P[T(x) > t]
and
fT(x) J(x) (t, j)
µx+t
(j)
= .
P[T(x) > t]
One must be careful in the use of these formulas. In particular, while
t
(τ ) (τ )
t px = exp{− µx+s ds}
0

§21: Multiple Decrement Models 78
it is also the case that t
t px ≠ exp{−
(j)
µx+s
(j)
ds}
0
as one might not at first expect. This latter integral does have an important use
which is explored below.
An important practical problem is that of constructing a multiple decrement life

table. To see how such a problem arises consider the case of a double indemnity
whole life policy. Assume that the policy will pay an amount $1 at the end of the
year of death if death occurs due to non-accidental causes and an amount of $2 if
the death is accidental. Denote the type of decrement as 1 and 2 respectively. The
present value of the benefit is then
vK(x)+1 1{1} (J(x)) + 2vK(x)+1 1{2} (J(x)) = J(x)vK(x)+1 .
To compute the net premium it remains to compute the expectation of this quantity.
This computation can only be completed if p(j)
x is known. How are these probabilities
calculated?
There are two basic methodologies used. If a large group of people for which
extensive records are maintained is available the actual survival data with the deaths
in each year of age broken down by cause would also be known. It is then very easy
to construct the multiple decrement table. This is seldom the case.
Example 21–1. An insurance company has a thriving business issuing life insurance
to coal miners. There are three causes of decrement (death): mining accidents, lung
disease, and other causes. From the company’s vast experience with coal miners a
decrement (life) table for these three causes of decrement is available. The company
now wants to enter the life insurance business for salt miners. Here the two causes of
decrement (death) are mining accidents and other. How can the information about
mining accidents for coal miners be used to get useful information about mining
accidents for salt miners?
A simple-minded answer to the question raised in the example would be to

simply lift the appropriate column from the coal miners life table and use it for the
salt miners. Such an approach fails, because it does not take into account the fact
that there are competing risks, that is, the accident rate for coal miners is affected by
the fact that some miners die from lung disease and thus are denied the opportunity
to die from an accident. The death rate for each cause in the absence of competing
risk is needed.
To see how to proceed the multiple decrement process is examined in a bit more
detail. Some auxillary quantities are introduced. Define
t
′
t px
(j)
= exp{−
0
µx+s
(j)
ds}
and
t qx′(j) = 1 − t p′x(j).
The “probability” t q′x (j) is called the net probability of decrement (or absolute
rate of decrement). It is these “probabilities” that represent the death rates in the
absence of competing risks. To see why this interpretation is reasonable, note that
fT(x) J(x) (t, j)

µx+t
(j)
=
P[T(x) > t]
fX J(x) (x + t, j)/ s(x)
=
s(x + t)/ s(x)
fX J(x) (x + t, j)
= .
P[X > x + t]
This shows that µx+t

(j)
represents the rate of death due to cause j among those surviving
up to time x + t.
These “probabilities” may be used to obtain the desired entries in a multiple

decrement table as follows. First

m
(τ )
t px = ′(j) .
t px
j=1
This shows how one can pass from the absolute rate of decrement to total survival
probabilities. Note that this relationship implies that the rates are generally larger
than the total survival probability. Then, under the assumption of constant force of
mortality for each decrement over each year of age in the multiple decrement table,
1
(τ )
q(j)
x = s px µx+s
(j)
ds
0
1
(τ )
= s px µx(j) ds
0
µx(j) 1 (τ ) (τ )
= s px µx ds
µx(τ ) 0
µ (j)
= (xτ ) qx(τ )
µx
log p′x (j) (τ )
= qx .
log px(τ )
This solves the problem of computing the entries in a multiple decrement table
under the stated assumption about the structure of the causes of decrement in that
table.
Exercise 21–1. What happens if px(τ ) = 1?

Exercise 21–2. Show that the same formula results if one assumes instead that the
time of decrement due to each cause of decrement in the multiple decrement table
has the uniform distribution over the year of decrement. (This assumption means
x = t ⋅ qx .)
that t q(j) (j)
Exercise 21–3. Assume that two thirds of all deaths at any age are due to accident.
What is the net single premium for (30) for a double indemnity whole life policy?
How does this premium compare with that of a conventional whole life policy?
The previous computations were based on assumptions about the causes of

decrement within the multiple decrement table. In some contexts it is more sensible
to make assumptions about the structure of the individual causes of decrement as
if each were acting independently, that is, to make assumptions about the absolute
rate of decrement in the single decrement tables.
Example 21–2. Suppose we are designing a pension plan and that there are two
causes of decrement: death and retirement. In many contexts (such as teaching) it
is reasonable to assume that retirements all occur at the end of a year, while deaths
can occur at any time. How could we construct a multiple decrement table which
reflects this assumption?
One common assumption about a single decrement is the assumption of uniform

distribution of deaths in the year of death. In the multiple decrement context this
translates in the statement that for 0 ≤ t ≤ 1
t qx ′(j) = t q′x(j) .
Exercise 21–4. Show that under this assumption we have t p′x (j) µx+t
(j)
= q′x (j) for 0 ≤
d
t ≤ 1. Hint: Compute t p′x (j) in two different ways.
dt
If this uniformity assumption is made for all causes of decrement it is then easy
to construct the multiple decrement table. The computations are illustrated for the
case of 2 causes of decrement. In this setting
1
(τ ) (1)
q(1)
x = s px µx+s ds
0
1
=
0
s px′(1)s px′(2)µx+s
(1)
ds
1
= qx′ (1)
′(2) ds
s px
0
1
= qx′(1) (1 − sqx′(2) ) ds
0
1
= qx′(1) (1 − qx′(2) )
2
with a similar formula for q(2)
x . It is easy to see how this procedure could be modified
for different assumptions about the decrement in each single decrement table.
Exercise 21–5. Construct a multiple decrement table in which the first cause of
decrement is uniformly distributed and the second cause has all decrements occur
at the end of the year. The pension plan described in the example above illustrates
the utility of this technique.
Another approximation which is used to connect single and multiple decrement

tables makes use of the functions
1
Lx = lx+t dt
0
lx − lx+1
mx = .
Lx
The function mx is called the central death rate at age x. The central rate of death
is used in a special technique, called the central rate bridge, in the context of
multiple decrement tables. This technique is now briefly described. Define
1
(τ ) (τ )
t px µx+t dt
mx(τ ) = 0
1
(τ )
t px dt
0
and 1
(τ ) (j)
t px µx+t dt
m(j)
x = 0
1
(τ )
t px dt
0
and 1
t px′(j) µx+t
(j)
dt
m′x =
(j) 0
1 .
t p′x dt
(j)
0
The central rate bridge is based on the following approximation. First, under the
UDD assumption in each single decrement table
q′x (j)
m′x (j) = .
1 − 12 q′x (j)
Second, under the UDD assumption in the multiple decrement table
q(j)
m(j)
x =
x
.
1 − 12 qx(τ )
Thirdly, under the constant force assumption in the multiple decrement table
x = µx = m′x .
m(j) (j) (j)
Now assume that all of these equalities are good approximations in any case. This
assumption provides a way of connecting the single and multiple decrement tables.
There is no guarantee of the internal consistency of the quantities computed in this
way, since, in general, the three assumptions made are not consistent with each
other. The advantage of this method is that the computations are usually simpler
than for any of the ‘exact’ methods.
Exercise 21–6. Show that each of the above equalities hold under the stated as-
sumptions.
Problems
Problem 21–1. Assume that each decrement has a uniform distribution over each
year of age in the multiple decrement table to construct a multiple decrement table
from the following data.
Age qx′(1) qx′(2) qx′(3)

62 0.020 0.030 0.200
63 0.022 0.034 0.100
64 0.028 0.040 0.120
Problem 21–2. Rework the preceding exercise using the central rate bridge. How
different is the multiple decrement table?
Problem 21–3. In a double decrement table where cause 1 is death and cause 2 is
withdrawal it is assumed that deaths are uniformly distributed over each year of age
while withdrawals between ages h and h + 1 occur immediately after attainment of
(τ )
50 = 0.24, and d50 = 0.06d50 . What
= 1000, q(2) (1) (2)
age h. In this table one sees that l50
′ ? How does your answer change if all withdrawals occur at midyear? At the
(1)
is q50
end of the year?
Problem 21–4. How would you construct a multiple decrement table if you were
given qx′(1) , qx′(2) , and q(3)
x ? What assumptions would you make, and what formulas
would you use? What if you were given qx′(1) , q(2) (3)
x , and qx ?
Problem 21–1. First, p(62τ ) = (.98)(.97)(.80) and q(62τ ) = 1 − p(62τ ) . Also p′62(j) =
log p′62(j) (τ )
1 − q′62(j) . From the relation q(j) 62 = q62 the first row of the multiple
log p(62τ )
decrement table can be found.
(2) (1)
Problem 21–3. From the information d50 = 240 and d50 = 14. Since with-
drawals occur at the beginning of the year there are 1000 − 240 = 760 people
under observation of whom 14 die. So q′50(1) = 14/ 760. If withdrawals occur at
year end all 1000 had a chance to die so q′50(1) = 14/ 1000.
Problem 21–4. The central rate bridge could be used. Is there an exact method
available?
Exercise 21–1. What would this mean for µx(τ ) and the derivation?
(τ )
Exercise 21–2. The assumption is that t q(j) (j)
x = tqx for all j. Hence t px =
(τ ) (τ ) (τ )
1 − tqx and µx+s = ds s qx / s px = qx / s px . Substitution and integration gives
(j) d (j) (j)
1 (j)
p′x (j) = e 0 x+s = (1 − q(xτ ) )qx / qx . Since p(xτ ) = 1 − q(xτ ) , the result follows by
− µ ds (j) (τ )
substitution.
Exercise 21–3. The actuarial present value of the benefit is (1/ 3) × 1 × Ax +

(2/ 3) × 2 × Ax = (5/ 3)Ax, from which the premium is easily calculated.
Exercise 21–4. From the definition, d

′
dt t px
(j)
= −t p′x (j) µx+t
(j)
, while from the
relation t p′x (j) = 1 − tq′x (j) , dtd t p′x (j) = −q′x (j) .

x = q′x 0 s p′x
(1) 1
Exercise 21–5. Since cause 1 obeys UDD, q(1) (2)
ds as in the
derivation above. For cause 2, s p′x = 1 for s < 1, so qx = q′x . For cause 2
(2) (1) (1)
1 (1) d (2)
x = − 0 s p′x ds s p′x
proceed as in the derivation above to get q(2) ds. Now s p′x (2)
is constant except for a jump of size −q′x at s = 1. Hence qx = q′x (2) p′x (1) =
(2) (2)
q′x (2) (1 − q′x (1) ).
Exercise 21–6. Under UDD in the single decrement table t p′x (j) = 1 − tq′x (j) and
1 1
t p′x µx+t = q′x so m′x = 0 q′x (j) dt/ 0 (1 − tq′x (j) ) dt = q′x (j) / (1 − 12 q′x (j) ). Under
(j) (j) (j) (j)
(τ )
UDD in the multiple decrement table µx+s (j)
x / s px so that substitution gives
= q(j)
the result. Under the constant force assumption in the multiple decrement table,
µx+s x = µx = m′x by substitution.
= µx(j) for all j and m(i)
(j) (j) (j)
§22. Laboratory 8
The service table at the end of these notes contains information about a group
of workers. There are 4 causes of decrement for this population. The first cause
is death (d), the second cause is withdrawal (w) (termination of employment), the
third cause is incapacity (i), and the fourth cause is retirement (r).
1. Suppose a concerted effort by the company reduces the rate of on the job
injury (incapacity) by 1/ 3 at all ages. Recompute the entries in the service table.
2. Suppose that in addition to the incapacity improvement an enhanced salary

and benefits plan reduces the withdrawal rate at all ages by 1/ 4. Recompute the
entries in the service table.

§23. Insurance Company Operations
The discussion thus far has been about individual policies. In the next few
sections the operations of the company as a whole are examined. This examination
begins with an overview of the accounting practices of an insurance company. This
is followed by a study of the behavior of the loss characteristics of groups of similar
policies. This last study leads to another method of setting premiums for a policy.

§24. Net Premium Reserves
A realistic model for both insurance policies and the method and amount of
premium payment is now in hand. The next question is how accounting principles
are applied to the financial operations of insurance companies.
A basic review of accounting principles is given first. There are three broad
categories of items for accounting purposes: assets, liabilities, and equity. Assets
include everything which is owned by the business. Liabilities include everything
which is owed by the business. Equity consists of the difference in the value of the
assets and liabilities. Equity could be negative. In the insurance context liabilities
are referred to as reserve and equity as surplus. When an insurance policy is
issued the insurance company is accepting certain financial obligations in return for
the premium income. The basic question is how this information is reflected in the
accounting statements of the company. Some of the different accounting procedures
available will now be described. Keep in mind that this discussion only concerns
how the insurance company prepares accounting statements reflecting transactions
which have occurred. The method by which gross (or net) premiums are calculated
is not being changed!
Example 24–1. Suppose the following data for an insurance company is given.
Income for Year Ending December 31, 1990

Premiums 341,000
Investment Income 108,000
Expenses 112,000
Claims and Maturities 93,000
Increases in Reserves —
Net Income —
Balance Sheet
December 31, 1989 December 31, 1990
Assets 1,725,000 —
Reserves — 1,433,000
Surplus 500,000 —
The missing entries in the tables can be filled in as follows (amounts in thou-
sands). Total income is 341 + 108 = 449 while total expenses are 112 + 93 = 205,
so net income (before reserve contributions) is 449 − 205 = 244. Now the reserves
at the end of 1989 are 1, 725 − 500 = 1, 225, so the increase in reserves must be
1, 433 − 1, 225 = 208. The net income is 244 − 208 = 36. Hence the 1990 surplus
is 536 and the 1990 assets are 1,969.
§24: Net Premium Reserves 89
The central question in insurance accounting is “How are liabilities measured?”
The answer to this question has some very important consequences for the operation
of the company, as well as for the financial soundness of the company. The general
equation is
Reserve at time t = Actuarial Present Value at time t of future benefits
− Actuarial Present Value at time t of future premiums.
The only accounting assumption required is one regarding the premium to be used
in this formula. Is it the net premium, gross premium, or ???
The only point of view adopted here is that liabilities are measured as the
net level premium reserves. This is the reserve computed under the accounting
assumption that the premium charged for the policy is the net level premium. To see
that this might be a reasonable approach, recall that the equivalence principle sets
the premium so that the actuarial present value of the benefit is equal to the actuarial
present value of the premiums collected. However, it is clear that after the policy
is issued the present value of the benefits and of the un-collected premiums will no
longer be equal, but will diverge in time. This is because the present value of the
unpaid benefits will be increasing in time and the present value of the uncollected
premiums will decrease in time. The discrepency between these two amounts at any
time represents an unrealized liability to the company. To avoid a negative surplus
(technical bankruptcy), this liability must be offset in the accounting statments of
the company by a corresponding asset. Assume (for simplicity) that this asset takes
the form of cash on hand of the insurance company at that time. How does one
compute the amount of the reserve at any time t under this accounting assumption?
This computation is illustrated in the context of an example.
Example 24–2. Consider a fully discrete whole life policy issued to (x) in which
the premium is payable annually and is equal to the net premium. What is the
reserve at time k, where k is an integer? To compute the reserve simply note that
if (x) has survived until time k then the (curtate) remaining life of x has the same
distribution as K(x + k). The outstanding benefit has present value vK(x+k)+1 while
the present value of the remaining premium income is äK(x+k)+1 times the annual
premium payment. Denote by k L the random variable which denotes the size of the
future loss at time k. Then
kL = vK(x+k)+1 − Px äK(x+k)+1 .
The reserve, denoted in this case by k Vx , is the expectation of this loss variable.
Hence
k Vx = E[k L] = Ax+k − Px äx+k .
This is called the prospective reserve formula, since it is based on a look at the
future performance of the insurance portfolio.
A word about notation. In the example above the reserve has been computed for
a discrete whole life policy. The notation for the reserves for other types of policies
parallel the notation for the premiums for the policy. Thus k V1x:n is the reserve at
time k for an n year term policy. When discussing general principals the notation
k V is used to denote the reserve at time k for a general policy.
Exercise 24–1. What types of policies have reserves t V(A1x:n ), k V(A1x:n ), and k V(äx )?
Certain timing assumptions regarding disbursements and receipts have been

made in the previous computation. Such assumptions are always necessary, so they
are now made explicit. Assume that a premium payment which is due at time t is
paid at time t+; an endowment benefit due at time t is paid at time t+; a death benefit
payment due at time t is assumed to be paid at time t−, that is, just before time t.
Interest earned for the period is received at time t−. Thus t Vx includes any interest
earned and also the effects of any non-endowment benefit payments but excludes
any premium income payable at time t and any endowment payments to be made at
time t. Also assume that the premium charged is the net level premium. Therefore
the full technical description of what has been computed is the net level premium
terminal reserve. One can also compute the net level premium initial reserve
which is the reserve computed right at time t. This initial reserve differs from the
terminal reserve by the amount of premium received at time t and the amount of the
endowment benefit paid at time t. Ordinarily one is interested only in the terminal
reserve.
In the remainder of this section methods of computing the net level premium
terminal reserve are discussed. For succintness, the term ‘reserve’ is always taken
to mean the net level premium terminal reserve unless there is an explicit statement
to the contrary.
äx+k
Exercise 24–2. Show that k Vx = 1 − . From this lim k Vx = 1. Why is this
äx k→∞
reasonable?
Exercise 24–3. Use the Life Table to compute the reserve for the first five years
after policy issue for a fully discrete whole life policy to (20). Assume the policy
amount is equal to $100,000 and the premium is the net premium.
The reserve can be viewed in a different way. Typically an insurance company

has many identical policies in force. One may benefit by studying the cash flow
associated with this group of policies on the average. Here is an example.
Example 24–3. Let us examine the expected cash flow associated with a whole life
policy issued to (x). Assume the premium is the net level premium and that the
policy is fully discrete. In policy year k + 1 (that is in the time interval [k, k + 1))
there are the following expected cash flows.
Time Income Cash on Hand
k− (benefits just paid, interest just received) k Vx
k Px k Vx + Px
k + 1− −qx+k k Vx + Px − qx+k
k + 1− i(k Vx + Px ) (1 + i)(k Vx + Px ) − qx+k
This final cash on hand at time k+1− must be equal to the reserve for the policies
of the survivors. Thus
px+k k+1 Vx = (1 + i)(k Vx + Px ) − qx+k .
This provides an important formula connecting successive reserves.
Exercise 24–4. Show that 1 Ex+k k+1 Vx = k Vx + Px − vqx+k .
The analysis of the previous example illustrates a general argument connecting

the reserves at successive time points.
kV = Actuarial Present Value at time k of benefits payable in [k, k + 1)

− Actuarial Present Value at time k of premiums payable in [k, k + 1)
+ vpx+kk+1 V.
Such recursive formulas for reserves are especially useful for computational pur-
poses.
There are other ways to compute the reserve. First the reserve may be viewed as
maintaining the balance between income and expenses. Since at time 0 the reserve is
0 (because of the equivalence principle) the reserve can also be viewed as balancing
past income and expenses. This leads to the retrospective reserve formula
k Ex k Vx = Px äx:k − A1x:k .
This formula is derived as follows. Recall that
Ax = A1x:k + vk k px Ax+k
and
äx = äx:k + vk k px äx+k .
Since the reserve at time 0 is zero,

0 = Ax − Px äx = A1x:k + vk k px Ax+k − Px äx:k + vk k px äx+k
where k is an arbitrary positive integer. Rearranging terms and using the prospective
formula for the reserve given above produces the retrospective reserve formula.
Exercise 24–5. Sometimes the retrospective reserve formula is written as
h Vx = Px äx:h / hEx − h kx = Px s̈x:h − h kx
where h kx is called the accumulated cost of insurance. Find an expression for h kx .

How would t kx be computed?
It is relatively straightforward to write expressions for the reserve for any of

the many possible types of insurance policy. Doing this is left as an exercise for
the reader. One should keep in mind that the important point here is to be able to
(ultimately) write a formula for the reserve which one can compute with the data
available in the life table. Hence continuous and/or mthly payment schemes need
to be reduced to their equivalent annual forms. Recursive formulas are also often
used.
Problems
äx+k:n−k
Problem 24–1. True or False: For 0 ≤ k < n, k Vx:n = 1 − . What happens at
äx:n
k = n?
Problem 24–2. Find a formula for the reserve at the end of 5 years for a 10 year
term policy with benefit $1 issued to (30) on a net single premium basis.
Problem 24–3. Show that for 0 ≤ t ≤ n

t V(Ax:n ) = P(Ax+t:n−t ) − P(Ax:n ) ax+t:n−t .
This is called the premium difference formula for reserves. Find similar formulas
for the other types of insurance.
Problem 24–4. Show that for 0 ≤ t ≤ n

P(Ax:n )
t V(Ax:n ) = 1 − Ax+t:n−t .
P(Ax+t:n−t )
This is called the paid up insurance formula for reserves. Find similar formulas
for the other types of insurance.
Problem 24–5. Find P1x:n if n Vx = 0.080, Px = 0.024 and P 1 = 0.2.

x:n
Problem 24–6. Given that 10 V35 = 0.150 and that 20 V35 = 0.354 find 10 V45 .
Problem 24–7. Write prospective and retrospective formulas for 40

20 V(A20 ), the re-
serve at time 20 for a semi-continuous 40 payment whole life policy issued to
(20).
Problem 24–8. For a general fully discrete insurance let us suppose that the benefit
payable if death occurs in the time interval (h − 1, h] is bh and that this benefit is
paid at time h, that is, at the end of the year of death. Suppose also that the premium
paid for this policy at time h is π h . Show that for 0 ≤ t ≤ 1
t px+k k+t V + v1−t t qx+k bk+1 = (1 + i)t (k V + π k ).
This gives a correct way to interpolate reserves at fractional years.
Problem 24–9. In the notation of the preceding problem show that for 0 ≤ t ≤ 1
k+t V = v1−t (1−t qx+k+t bk+1 + 1−t px+k+t k+1 V) .
Problem 24–10. Show that under UDD, hk V(Ax:n ) = (i/ δ )hk V1x:n + hk V 1 .
x:n
Problem 24–11. Show that under UDD, hk Vx:n(m)

= hk Vx:n + β (m)h P(m)
x:n k V1
x:n
. This gives
the reserves for a policy with mthly premium payments in terms of the reserves for
a policy with annual premium payments.
Problem 24–12. Show that hk V (m) (Ax:n ) = hk V(Ax:n ) + β (m)h P(m) (Ax:n )k V1x:n under
UDD.
Problem 24–13. The amount at risk in year k for a discrete insurance is the
difference between the benefit payment payable at the end of year k and k V. Find
the mean and variance of the amount at risk in year 3 of a 5 year term policy issued
to (30) which pays a benefit of 1 at the end of the year of death and has net level
premiums.
Problem 24–14. Suppose that 1000 t V(Ax ) = 100, 1000P(Ax ) = 10.50, and δ =
0.03. Find ax+t .
Problem 24–15. Calculate 20 V45 given that P45 = 0.014, P 1 = 0.022, and
45:20
P45:20 = 0.030.
Problem 24–16. A fully discrete life insurance issued to (35) has a death benefit of
$2500 in year 10. Reserves are calculated at i = 0.10 and the net annual premium
P. Calculate q44 given that 9 V + P = 10 V = 500.
Problem 24–1. Use the prospective formula and Ax:n + däx:n = 1 to see the
formula is true. When k = n the reserve is 1 by the timing assumptions.
Problem 24–2. Prospectively the reserve is A1 .

35:5
Problem 24–3. Use the prospective formula and the premium definitions.
Problem 24–5. From the retrospective formula n Exn Vx = Px äx:n − A1 . Now

x:n
divide by äx:n .
Problem 24–6. Use the prospective formula and the relation Ax + däx = 1 to
obtain k Vx = 1 − äx+k / äx .
40
Problem 24–7. The prospective and retrospective formulas are 20 V(A20 ) =
A40 − Pä40:20 and 40
20 V(A20 ) = (Pä20:20 − A1 )/ 20 E20 .
20:20
Problem 24–8. The value of the reserve, given survival, plus the present value
of the benefit, given death, must equal the accumulated value of the prior reserve
and premium.
Problem 24–13. The amount at risk random variable is 1{2} (K(30)) − 3 V1 .

30:5
Problem 24–14. Use the prospective reserve formula and the relationship
Ax + δ ax = 1.
Problem 24–15. Use the retrospective formula.
Problem 24–16. By the general recursion formula 1 E4410 V = 9 V +P−2500vq44.

Exercise 24–1. t V(A1 ) is the reserve at time t for a fully continuous n year
x:n
term insurance policy, k V(A1 ) is the reserve at time k for a semi-continuous n
x:n
year term policy, and k V(äx ) is the reserve at time k for a life annuity.
Exercise 24–2. Since Ax + däx = 1, k Vx = Ax+k − (Ax / äx )äx+k = 1 − däx+k − (1 −

däx )äx+k / äx = 1 − äx+k / äx .
Exercise 24–3. The reserve amounts are easily computed using the previous
exercise as 1000001V20 = 100000(1 − 19.014/ 19.087) = 382.46, 1000002V20 =
775.40, 1000003V20 = 1184.06, 1000004V20 = 1613.67, and 1000005V20 =
2064.24.
Exercise 24–4. Just multiply the previous equation by v = (1 + i)−1 .
Exercise 24–5. h kx = A1 / k Ex , which can be easily computed from the life

x:k
table using previous identities.
§25. Laboratory 9
1. Return to Laboratory 6 and find the reserve at the end of each policy year
for the term policy of problem 1 at an interest rate of 5%. Begin by deriving a
recurrence relation between the reserves for successive years.
2. Find the reserve at the end of each policy year for the term policy with de-
clining benefits given in problem 3 of Laboratory 6. Begin by deriving a recurrence
relation between the reserves for successive years.
The problems below all concern the following situation. Your company has
just sold 3 year term insurance policies to a group of 20 persons, each aged 30. The
benefit amount for each policy is $1,000,000 and is payable at the moment of death.
Each policy has level premiums payable at the beginning of each policy year. The
mortality characteristics are q30 = 0.000444, q31 = 0.000499, q32 = 0.000562, and
q33 = 0.000631. The interest rate is 4%.
3. Make a table showing the reserves for this group of policies at the end of
each policy year.
4. Your company charges a premium which is twice the net premium. Make a
table showing the expected free cash flow generated by this group of policies at the
end of each year. Assume that there are no expenses.
5. Estimate the probability that this group of policies will require a cash
infusion from the other business of the company at some point during the life of
these policies. Also estimate the amount of cash required (if any). Carefully explain
your methodology and state your assumptions.

§26. The Individual Risk Model
An insurance company has a large number of policies in force at any given

time. This creates a financial risk for the company. There are two aspects to the
problem of analyzing this risk. First, one must be able to estimate the amount of
risk. Secondly, one must be able to model the times at which claims will occur in
order to avoid cash flow difficulties. The problem of modeling the amount of risk
will be studied first.
In the individual risk model the insurers total risk S is assumed to be expressable
in the form S = X1 + . . . + Xn where X1 , . . . , Xn are independent random variables
with Xi representing the loss to the insurer on insured unit i. Here Xi may be quite
different than the actual damages suffered by insured unit i. In the closed model
the number of insured units n is assumed to be known and fixed. A model in which
migration in and out of the insurance system is allowed is called an open model.
The individual risk model is appropriate when the analysis does not require the
effect of time to be taken into account.
The first difficulty is to find at least a reasonable approximation to the proba-

bilistic properties of the loss random variables Xi . This can often be done using data
from the past experience of the company.
Example 26–1. For short term disability insurance the amount paid by the insurance
company can often be modeled as X = cY where c is a constant representing the
daily rate of disability payments and Y is the number of days a person is disabled.
One then is simply interested in modelling the random variable Y. Historical data
can be used to estimate P[Y > y]. In this context P[Y > y], which was previously
called the survival function, is referred to as the continuance function. The same
notion can be used for the daily costs of a hospitalization policy.
The second difficulty is to uncover the probabilistic properties of the random

variable S. In theoretical discussions the idea of conditioning can be used to find
an explicit formula for the distribution function of a sum of independent random
varibles.
Example 26–2. Suppose X and Y are independent random variables each having
the exponential distribution with parameter 1. By conditioning
∞
P[X + Y ≤ t] = P[X + Y ≤ t| Y = y] fY (y) dy
−∞
t
= P[X ≤ t − y] fY (y) dy
0
t
= 1 − e−(t−y) fY (y) dy
0
= 1 − e−t − te−t
§26: The Individual Risk Model 99
for t ≥ 0.
This argument has actually shown that if X and Y are independent and absolutely
continuous then ∞
FX+Y (t) = FX (t − y) fY (y) dy.
−∞
This last integral is called the convolution of the two distribution functions and is
often denoted FX ∗ FY .
Exercise 26–1. If X and Y are absolutely continuous random variables show that
X + Y is also absolutely continuous and find a formula for the density function of
X + Y.
Exercise 26–2. Find a similar formula if X and Y are both discrete. Use this formula
to find the density of X + Y if X and Y are independent Bernoulli random variables
with the same success probability.
An approach which requires less detailed computation is to appeal to the Central

Limit Theorem.
Central Limit Theorem. If X1 , . . . , Xn are independent random variables then the

n
n
distribution of Xi is approximately the normal distribution with mean E[Xi ]
i=1 i=1

n
and variance Var(Xi ).
i=1
The importance of this theorem lies in the fact that the approximating normal
distribution does not depend on the detailed nature of the original distribution but
only on the first two moments. The accuracy of this approximation will be explored
in the exercises and laboratory.
Example 26–3. You are a claims adjuster for the Good Driver Insurance Company
of Auburn. Based on past experience the chance of one of your 1000 insureds being
involved in an accident on any given day is 0.001. Your typical claim is $500. What
is the probability that there are no claims made today? If you have $1000 cash on
hand with which to pay claims, what is the probability you will be able to pay all
of todays claims? How much cash should you have on hand in order to have a
99% chance of being able to pay all of todays claims? What assumptions have you
made? How reasonable are they? What does this say about the solvency of your
company?
It is easily seen, by using the Central Limit Theorem, that if an insurance

company sold insurance at the pure premium not only would the company only
break even (in the long run) but due to random fluctuations of the amount of
claims the company would likely go bankrupt. Thus insurance companies charge
an amount greater than the pure premium. A common methodology is for the
company to charge (1 + θ ) times the pure premium. When this scheme is followed θ
is called the relative security loading and the amount θ × (pure premium) is called
the security loading. This is a reasonable procedure since the insureds with larger
expected claims pay a proportionate share of the loading. The relative loading θ is
usually adjusted to achieve a certain measure of protection for the company.
Example 26–4. Suppose that a company is going to issue 1,000 fire insurance
policies each having a $250 deductible, and a policy amount of $50,000. Denote
by Fi the Bernoulli random variable which is 1 if the ith insured suffers a loss, and
by Di the amount of damage to the ith insureds property. Suppose Fi has success
probability 0.001 and that the actual damage Di is uniformly distributed on the
interval (0,70000)). What is the relative loading so that the premium income will
be 95% certain to cover the claims made? Using the obvious notation, the total
amount of claims made is given by the formula

1000

S= Fi (Di − 250)1[250,50250] (Di ) + 500001(50250,∞) (Di )
i=1
where the F’s and the D’s are independent (why?) and for each i the conditional
distribution of Di given Fi = 1 is uniform on the interval (0,70000). The relative
security loading is determined so that
P[S ≤ (1 + θ ) E[S]] = 0.95.
This is easily accomplished by using the Central Limit Theorem.
Exercise 26–3. Compute E[S] and Var(S) and then use the Central Limit Theorem
to find θ . What is the probability of bankruptcy when θ = 0?
Another illustration is in connection with reinsurance. It is generally not good

practice for an insurance company to have all of its policy holders homogeneous,
such as all located in one geographical area, or all of the same physical type. A
moments reflection on the effect of a hurricane on an insurance company with all
of its property insurance business located in one geographic area makes this point
clear. An insurance company may diversify its portfolio of policies (or just protect
itself from such a concentration of business) by buying or selling reinsurance. The
company seeking reinsurance (the ceding company) buys an insurance policy from
the reinsurer which will reimburse the company for claims above the retention
limit. For stop loss reinsurance, the retention limit applies on a policy-by-policy
basis to those policies covered by the reinsurance. The retention limit plays the
same role here as a deductible limit in a stop loss policy. Usually there is one
reinsurance policy which covers an entire package of original policies. For excess
of loss reinsurance, the retention limit is applied to the total amount of claims for the
package of policies covered by the insurance, not the claims of individual policies.
Example 26–5. You operate a life insurance company which has insured 2,000 30
year olds. These policies are issued in varying amounts: 1,000 people with $100,000
policies, 500 people with $500,000 policies, and 500 people with $1,000,000 poli-
cies. The probability that any one of the policy holders will die in the next year is
0.001. Stop loss reinsurance may be purchased at the rate of 0.0015 per dollar of
coverage. How should the retention limit be set in order to minimize the probabil-
ity that the total expenses (claims plus reinsurance expense) exceed $1,000,000 is
minimized? Let X, Y, and Z denote the number of policy holders in the 3 catagories
dying in the next year. Then X has the binomial distribution based on 1000 trials
each with success probability 0.001, Y has the binomial distribution based on 500
trials each with success probability 0.001, and Z has the binomial distribution based
on 500 trials each with success probability 0.001. If the retention limit is set at r
then the cost C of claims and reinsurance is given by
C = (100000 ∧ r)X + (500000 ∧ r)Y + (1000000 ∧ r)Z

+ 0.0015 1000(100000 − r)+ + 500(500000 − r)+ + 500(1000000 − r)+ .
It is then a relatively straightforward, though tedious, task to use the central limit
theorem to estimate P[C ≥ 1, 000, 000].
Exercise 26–4. Verify the validity of the above formula. Use the central limit
theorem to estimate P[C ≥ 1, 000, 000] as a function of r. Find the value(s) of r
which minimize this probability.
Problems
Problem 26–1. The probability of an automobile accident in a given time period is

0.001. If an accident occurs the amount of damage is uniformly distibuted on the
interval (0,15000). Find the expectation and variance of the amount of damage.
Problem 26–2. Find the distribution and density for the sum of three independent
random variables each uniformly distributed on the interval (0,1). Compare the
exact value of the distribution function at a few selected points (say 0.25, 1, 2.25)
with the approximation obtained from the central limit theorem.
Problem 26–3. Repeat the previous problem for 3 independent exponential random
variables each having mean 1. It may help to recall the gamma distribution here.
Problem 26–4. A company insures 1000 essentially identical cars. The probability
that any one car is in an accident in any given year is 0.001. The damage to a car
that is involved in an accident is uniformly distributed on the interval (0,15000).
What relative security loading θ should be used if the company wishes to be 99%
sure that it does not lose money?
Problem 26–1. The amount of damage is BU where B is a Bernoulli variable
with success probability 0.001 and U has the uniform distribution.

1000
Problem 26–4. The loss random variable is of the form B i Ui .
i=1
Exercise 26–1. Differentiation
∞ of the general distribution function formula
above gives fX+Y (t) = −∞ fX (t − y)fY (y) dy.
Exercise
26–2. In the discrete case the same line of reasoning n fX+Yt−y(t) =
mgives
y fX (t−y)f
Y (y). Applying this in the Bernoulli case, fX+Y (t) = t−y p (1−
n m n+m t y=0
p)n−t+y my py (1 − p)m−y = pt (1 − p)n+m−t m y=0 t−y y = t p (1 − p) m+n−t
.
Exercise 26–3. The loading θ is chosen so that θ E[S]/ √Var(S) = 1.645, from
the normal table. When θ = 0 the bankruptcy probability is about 1/ 2.
Exercise 26–4. Direct computation using properties of the binomial distribu-

E[C] = (100000 ∧ r) × 1 + (500000 ∧ r) × (1/ 2) + (1000000 ∧ r)
tion gives × (1/ 2) +
0.0015 1000(100000 − r)+ + 500(500000 − r)+ + 500(1000000 − r)+ and also
Var(C) = (100000∧r)2 ×0.999+(500000∧r)2×0.999/ 2+(1000000∧r)2×0.999/ 2.
The probability can now be investigated numerically using the Central Limit
Theorem approximation.
§27. Laboratory 10
Another method of uncovering the probabilistic properties of the random vari-

able S is to use simulation. Suppose that a company insures 10,000 essentially
identical cars. The probability that any one car is in an accident in any given year is
0.001. The damage to a car that is involved in an accident is uniformly distributed
on the interval (0,20000).
1. Write the loss random variable S in this case. Find E[S] and Var(S).
2. Use the Central Limit Theorem to find the relative security loading θ that
should be used if the company wishes to be 99% sure that it does not lose money.
3. Consider one of the loss random variables X that occur in the expression for
S. Explain how a random number generator could be used to simulate X.
4. Use the method of problem 3 to simulate 5,000 observations on S. Make

a histogram of these observations. Based on this histogram, what relative security
loading should be used if the company wishes to be 99% sure that it does not lose
money? Compare the result here with that of problem 2.
5. Show that if Y is a random variable with increasing distribution function

FY (t) and if U is uniformly distributed on the interval (0, 1), then FY−1 (U) has the
same distribution as Y. This gives a general method to use for simulating random
variables.

§28. The Collective Risk Model and Ruin Probabilities
Some of the consequences of the collective risk model will now be examined.
In the collective risk model the time at which claims are made is taken to account.

Here the aggregate claims up to time t is assumed to be given by N(t) k=1 Xk where
X1 , X2 , . . . are independent identically distributed random variables representing the
sizes of the respective claims, N(t) is a stochastic process representing the number
of claims up to time t, and N and the X’s are independent. The object of interest is
the insurer’s surplus at time t, denoted by U(t), which is assumed to be of the form

N(t)
U(t) = u + ct − Xk
k=1
where u is the surplus at time t = 0, and c represents the rate of premium income.
Special attention will be given to the problem of estimating the probability that the
insurance company has negative surplus at some time t since this would mean that
the company is ruined.
To gain familiarity with some of the ideas involved, the simpler classical gam-
bler’s ruin problem will be studied first.

§29. Stopping Times and Martingales
A discrete time version of the collective risk model will be studied and some
important new concepts will be introduced.
Suppose that a gambler enters a casino with z dollars and plays a game of chance
in which the gambler wins $1 with probability p and loses $1 with probability
q = 1 − p. Suppose also that the gambler will quit playing if his fortune ever reaches
a > z and will be forced to quit by being ruined if his fortune reaches 0. The main
interest is in finding the probability that the gambler is ultimately ruined and also
the expected number of the plays in the game.
In order to keep details to a minimum, the case in which p = q = 1/ 2 will

be examined first. Denote by Xj the amount won or lost on the jth play of the
game. These random variables are all independent and have the same underlying
distribution function. Absent any restrictions about having to quit the game, the
fortune of the gambler after k plays of the game is

k
z+ Xj .
j=1
Now in the actual game being played the gambler either reaches his goal or is ruined.
Introduce a random variable, T, which marks the play of the game on which this
occurs. Technically

T = inf{k : z + kj=1 Xj = 0 or a}.
Such a random variable is called a random time. Observe that for this specific
random variable the event [T ≤ k] depends only on the random variables X1 , . . . , Xk .
That is, in order to decide at time k whether or not the game has ended it is not
necessary to look into the future. Such special random times are called stopping
times. The precise definition is as follows. If X1 , X2, . . . are random variables and
T is a nonnegative integer valued random variable with the property that for each
integer k the event [T ≤ k] depends only on X1 , . . . , Xk then T is said to be a stopping
time (relative to the sequence X1 , X2 , . . .).

The random variable z + Tj=1 Xj is the gambler’s fortune when he leaves the
casino, which is either a or 0. Denote by ψ (z) the probability that the gambler

leaves the casino with 0. Then by direct computation E[z + Tj=1 Xj ] = a(1 − ψ (z)).
A formula for the ruin probability ψ (z) will be obtained by computing this same
expectation in a second way.
Each of the random variables Xj takes values 1 and −1 with equal probability,

so E[Xj ] = 0. Hence for any integer k, E[ kj=1 Xj ] = 0 too. So it is at least plausible

that E[ Tj=1 Xj ] = 0 as well. Using this fact, E[z + Tj=1 Xj ] = z, and equating this

§29: Stopping Times and Martingales 108
with the expression above gives z = a(1 − ψ (z)). Thus ψ (z) = 1 − z/ a for 0 ≤ z ≤ a
are the ruin probabilities.
There are two important technical ingredients behind this computation. The
first is the fact that T is a stopping time. The second is the fact that the gambling
game with p = q = 1/ 2 is a fair game. The notion of a fair game motivates the
definition of a martingale. Suppose M0 , M1 , M2 , . . . are random variables. The
sequence is said to be a martingale if E[Mk | Mk−1 , . . . , M0 ] = Mk−1 for all k ≥ 1. In
the gambling context, if Mk is the gambler’s fortune after k plays of a fair game then
given Mk−1 the expected fortune after one more play is still Mk−1 .
k
Exercise 29–1. Show that Mk = z + j=1 Xj (with M0 = z) is a martingale.
2
Example 29–1. The sequence M0 = z2 and Mk = z + kj=1 Xj − k for k ≥ 1 is also
a martingale. This follows from the fact that knowing M0 , . . . , Mk−1 is the same as
knowing X1 , . . . , Xk−1 and the fact that the X’s are independent.
Exercise 29–2. Fill in the details behind this example.
The important computational fact is the Optional Stopping Theorem which

states that if {Mk } is a martingale and T is a stopping time then E[MT ] = E[M0 ]. In
the gambling context this says that no gambling strategy T can make a fair game
biased.
2
Example 29–2. Using the martingale M0 = z2 and Mk = z + kj=1 Xj − k for
k ≥ 1 along with the same stopping time T as before can provide information
about the duration of the gambler’s stay in the casino. The random variable MT =
2
z + Tj=1 Xj −T has an expectation which is easily computed directly to be E[MT ] =
a2 (1 − ψ (z)) − E[T]. By the optional stopping theorem, E[MT ] = E[M0 ] = z2 .
Comparing these two expressions gives E[T] = a2 (1 − ψ (z)) − z2 = az − z2 as the
expected duration of the game.
The preceding example illustrates the general method. To analyze a particular

problem identify a martingale Mk and stopping time T. Then compute E[MT ] in two
ways, directly from the definition and by using the optional stopping theorem. The
resulting equation will often reveal useful information.
Uncovering the appropriate martingale is often the most difficult part of the
process. One standard method is the following. If X1 , X2 , . . . are independent and
identically distributed random variables define
k
et j=1
Xj
Wk = k .
t Xj
E[e j=1 ]
Notice that the denominator is nothing more than the moment generating function
of the sum evaluated at t. For each fixed t the sequence Wk is a martingale (here
W0 = 1). This follows easily from the fact that if X and Y are independent then
E[et(X+Y) ] = E[etX ] E[etY ]. This martingale is called Wald’s martingale (or the
exponential martingale) for the X sequence.
Exercise 29–3. Show that {Wk : k ≥ 0} is a martingale no matter what the fixed
value of t is.
In many important cases a non-zero value of t can be found so that the denomi-
nator part of the Wald martingale is 1. Using this particular value of t then makes
application of the optional stopping theorem neat and easy.
To illustrate the technique consider the following situation which is closer

to that of the collective risk model. Suppose the insurer has initial reserve z
and that premium income is collected at the rate of c per unit time. Also, Xk
denotes the claims that are payable at time k, and the X’s are independent and
identically distributed random variables. The insurers reserve at time k is then

z + ck − kj=1 Xj = z + kj=1 (c − Xj ). Denote by T the time of ruin, so that

k
T = min{k : z + ck − Xj ≤ 0}.
j=1
The objective is to study the probability ψ (z) that ruin occurs in this setting.
As a first step, notice that if E[c − Xj ] ≤ 0, ruin is guaranteed since premium

income in each period is not adequate to balance the average amount of claims in
the period. So to continue, assume that E[c − Xj ] > 0.
Under this assumption, suppose there is a number τ so that E[eτ (c−Xj ) ] = 1.

This choice of

τ in Wald’s martingale makes the denominator 1, and shows that
k
Mk = eτ (z+ck− j=1 Xj ) is a martingale. Computing the expectation of MT using the
Optional Stopping Theorem gives E[MT ] = E[M0 ] = eτ z . Computing directly gives
T
E[MT ] = E[eτ (z+cT− j=1 Xj ) | T < ∞] ψ (z). Hence
T
ψ (z) = eτ z / E[eτ (z+cT− j=1
Xj )
| T < ∞].
A problem below will show that τ < 0, so the denominator of this fraction is larger
than 1. Hence ψ (z) ≤ eτ z . The ruin probability decays exponentially as the initial
reserve increases.
The usual terminology defines the adjustment coefficient R = −τ . Thus

ψ (z) ≤ e−Rz . So a large adjustment coefficient implies that the ruin probability
declines rapidly as the initial reserve increases.
Problems
Problem 29–1. By conditioning on the outcome of the first play of the game show
that in the gambler’s ruin problem ψ (z) = pψ (z + 1) + qψ (z − 1). Show that if p = q
there is a solution of this equation of the form ψ (z) = C1 + C2 z and find C1 and C2
by using the natural definitions ψ (0) = 1 and ψ (a) = 0. Show that if p ≠ q there is a
solution of the form ψ (z) = C1 + C2 (q/ p)z and find the two constants. This provides
a solution to the gambler’s ruin problem by using difference equations instead of
probabilistic reasoning.
Problem 29–2. In the gambler’s ruin problem, show that if p ≠ q the choice t =
ln(q/ p) makes the denominator of Wald’s martingale 1. Use this choice of t and the
optional stopping theorem to find the ruin probability in this case.
Problem 29–3. Suppose p ≠ q in the gambler’s ruin problem. Define M0 = z and

Mk = z + kj=1 Xj − k(p − q) for k ≥ 1. Show that the sequence Mk is a martingale and
use it to compute E[T] in this case.
Problem 29–4. Suppose that c > 0 is a number and X is a random variable which
takes on only non-negative values. Suppose also that E[c − X] > 0. Show that if
c − X takes on positive and negative values then there is a number τ < 0 so that
E[eτ (c−X) ] = 1.
(q/ p)a −(q/ p)z
Problem 29–2. ψ (z) = (q/ p)a −1 .
a 1−(q/ p)z
Problem 29–3. E[T] = z
q−p − q−p 1−(q/ p)a .
Problem 29–4. Define a function f (v) = E[ev(c−X) ]. Then f ′ (v) = E[(c −

X)ev(c−X)] and f ′′ (v) = E[(c − X)2 ev(c−X) ] > 0. Thus f is a convex function and
the graph of f is concave up. Now f (0) = 1 and f ′ (0) = E[(c − X)] > 0. Thus
the graph of f is above 1 to the right of 0, and below 1 (initially) to the left of
0. Since c − X takes on negative values, limv→−∞ f (v) = ∞, so there is a negative
value of v at which f (v) = 1, by continuity.
Exercise 29–1. Knowing M0 , . . . , Mk−1 is the same as knowing X1 , . . . , Xk−1 . So
k−1
E[Mk | M0 , . . . , Mk−1 ] = E[Mk | X0 , . . . , Xk−1 ] = z + j=1 Xj + E[Xk | X0 , . . . , Xk−1 ] =
Mk−1 since the last expectation is 0 by independence.
2 2
Exercise 29–2. First write Mk = z + j=1 k−1
Xj + Xk − k = z + k−1 j=1 X j +
k−1
2Xk (z + j=1 Xj ) + Xk − k. Take conditional expectations using the fact that Xk
2
is independent of the other X’s and E[Xk ] = 0 and E[Xk2 ] = 1 to obtain the result.
k k−1
t X t X
Exercise 29–3. Independence gives E[e j=1 j ] = E[e j=1 j ]×E[etXk ]. Direct
computation of the conditional expectation gives the result.
§30. The Collective Risk Model Revisited
The ideas developed in connection with the gambler’s ruin problem will now
be used to compute the ruin probability in the collective risk model. Since the
processes are now operating in continuous time the details are more complicated
and not every step of the arguments will be fully justified.

In this setting the claims process is N(t)k=1 Xk where X1 , X2 , . . . are independent
identically distributed random variables representing the sizes of the respective
claims, N(t) is a stochastic process representing the number of claims up to time t,
and N and the X’s are assumed to be independent. The insurer’s surplus is given by

k=1 Xk , where u > 0 is the surplus at time t = 0 and c > 0 is the rate
U(t) = u + ct − N(t)
at which premium income arrives per unit time. The probability of ruin with initial
surplus u will be denoted by ψ (u).
As in the discrete time setting, the Wald martingale will be used together with the
Optional Stopping Theorem in order to obtain information about the ruin probability.
Here the denominator of the Wald martingale is E[eν U(t) ], and the first step is to find
N(t)
a ν ≠ 0 so that E[eν (ct− k=1 Xk ) ] = 1 no matter the value of t.

The new element in this analysis is the random sum N(t) k=1 Xk . Now for each
fixed t, N(t) is a random variable which is independent of the X’s. The moment
generating function of this sum can be easily computed by conditioning on the value
of the discrete random variable N(t).
N(t) N(t)
E[eν k=1
Xk
] = E[E[eν k=1
Xk
| N(t)]]
∞
N(t)
= E[eν k=1
Xk
| N(t) = j] P[N(t) = j]
j=0
∞ j
= E[eν k=1
Xk
] P[N(t) = j]
j=0
∞ j
= E[eν X ] P[N(t) = j]
j=0
∞
ν X ])
= ej ln(E[e P[N(t) = j]
j=0
= MN(t) (ln(MX (ν ))).

N(t)
Hence there is a ν ≠0 so that E[eν (ct− k=1 Xk ) ] = 1 if and only if eν ct MN(t) (ln(MX (−ν ))) =
1 for all t. Suppose for now that there is a number R > 0 so that
e−Rct MN(t) (ln(MX (R))) = 1
for all t. This number R is called the adjustment coefficient. The existence of an
adjustment coefficient will be investigated a bit later. Using −R as the value of ν in
§30: The Collective Risk Model Revisited 114
Wald’s martingale shows that
N(t)
Wt = e−R(u+ct− k=1
Xk )
is a martingale.

Define a stopping time Ta by Ta = inf{s : u + cs − N(s)k=1 Xk ≤ 0 or ≥ a} where a
is an arbitrary but fixed positive number. It is intuitively clear that Ta is a stopping
time in an appropriate sense in the new continuous time setting. Now by the Optional
Stopping Theorem, E[WTa ] = e−Ru . Direct computation gives
N(Ta ) a )
N(T a )
N(T
E[WTa ] = E[e−R(u+cTa − k=1
Xk )
| u + cTa − Xk ≤ 0] P[u + cTa − Xk ≤ 0]
k=1 k=1
N(t) a )
N(T a )
N(T
+ E[e−R(u+ct− k=1
Xk )
| u + cTa − Xk ≥ a] P[u + cTa − Xk ≥ a].
k=1 k=1
Since this equation is valid for any fixed positive a, and since R > 0, it is possible
a)
to take limits as a → ∞. Since lima→∞ P[u + cTa − N(T k=1 Xk ) ≤ 0] = ψ (u) and
−Ra
lima→∞ e = 0 the following result is obtained.
Theorem. Suppose that in the collective risk model the adjustment coefficient R > 0

satisfies e−Rct MN(t) (ln(MX (R))) = 1 for all t. Let T = inf{s : u + cs − N(s)
k=1 Xk ≤ 0}
be the random time at which ruin occurs. Then
e−Ru
ψ (u) = N(T) ≤ e−Ru .
−R(u+cT−
E[e k=1
Xk )
| T < ∞]
Exercise 30–1. Why is the last inequality true?
As in the discrete time model, the existence of an adjustment coefficient guar-

antees that the ruin probability decreases exponentially as the initial surplus u
increases.
In general there is no guarantee that an adjustment coefficient will exist. For

certain particular types of models the adjustment coefficient can explicitly be found.
Moreover, a more detailed analysis of the claims process can be made in these
special cases.
The more restrictive discussion begins by examining the nature of the process
N(t), the total number of claims up to time t. A common assumption is that this
process is a Poisson process with constant intensity λ > 0. What this assumption
means is the following. Suppose W1 , W2 , . . . are independent identically distributed
exponential random variables with mean 1/ λ and common density λ e−λ x 1(0,∞) (x).
The W’s are the waiting times between claims. The Poisson process can then
be viewed as the number of claims that arrive up to time t. This means that

j=1 Wj > t}. It can be shown that for any fixed t the random variable
N(t) = inf{k : k+1
N(t) has the Poisson distribution with parameter λ t and that the stochastic process
{N(t) : t ≥ 0} has independent increments, that is, whenever t1 < t2 < . . . < tn are
fixed real numbers then the random variables N(t2 ) − N(t1 ), . . . , N(tn ) − N(tn−1 ) are
independent. Using this, direct computation gives
∞

E[eν N(t) ] = νj
e P[N(t) = j]
j=0
∞
= eν je−λ t (λ t)j / j!
j=0
∞

= e−λ t (eν λ t)j / j!
j=0
λ t(eν −1)
=e .
This simple formula for the moment generating function of N(t) leads to a simple
formula for the adjustment coefficient in this case. The general equation for the
adjustment coefficient was earlier found to be e−Rct MN(t) (ln(MX (R))) = 1. Taking
logarithms and using the form of the moment generating function of N(t) shows that
the adjustment coefficient is the positive solution of the equation
λ + cR = λ MX (R).
An argument similar to that given in the discrete time case can be used to show that
there is a unique adjustment coefficient in this setting.
Exercise 30–2. Verify that the adjustment coefficient, if it exists, must satisfy this
equation.
Example 30–1. Suppose all claims are for a unit amount. Then MX (ν ) = eν so the
adjustment coefficient is the positive solution of λ + cR = λ eR . Note that there is
no solution if c ≤ λ . But in this case the ruin probability is clearly 1.
Exercise 30–3. Show that if c ≤ λ E[X] the ruin probability is 1. Show that if
c > λ E[X] the adjustment coefficient always exists and hence the ruin probability
is less than 1.
The previous exercises suggest that only the case in which c > λ E[X] is of
interest. Henceforth write c = (1 + θ )λ E[X] for some θ > 0. Here θ is the relative
security loading.
Even more detailed information can be obtained when N(t) is a Poisson process.
To do this define a stopping time Tu = inf{s : U(s) < u} to be the first time that
the surplus falls below its initial level and denote by L1 = u − U(Tu ) the amount by
which the surplus falls below its initial level. Then
∞
1
P[Tu < ∞, L1 ≥ y] = (1 − FX (x)) dx.
(1 + θ )E[X] y
The proof of this fact is rather technical.

proof : Let h > 0 be small. Then P[N(h) = 0] = e−λ h ≈ 1, P[N(h) = 1] = λ he−λ h ≈ λ h and
P[N(h) ≥ 2] ≈ 0. Denote by R(u, y) the probability that with an initial surplus of u the first
time the surplus drops below 0, the surplus actually drops below −y. Conditioning on the
value of N(h) gives
u ∞
R(u, y) ≈ (1 − λ h)R(u + ch, y) + λ h R(u − x, y)fX (x) dx + fX (x) dx .
0 u+ch+y
Re-arranging gives

R(u, y) − R(u + ch, y) λ λ u
λ
= − R(u + ch, y) + R(u − x, y)fX (x) dx + fX (x) dx.
ch c c 0 c u+ch+y
Now take limits as h → 0 to obtain

λ λ u λ
−R′ (u, y) = − R(u, y) + R(u − x, y)fX (x) dx + fX (x) dx.
c c 0 c u+y
Since R(u, y) ≤ ψ (u) ≤ e−Ru , both sides can be integrated with respect to u from 0 to ∞.
Doing this gives
∞ ∞ ∞ ∞
λ λ u
λ
R(0, y) = − R(u, y) du + R(u − x, y)fX (x) dx du + fX (x) dx du.
c 0 c 0 0 c 0 u+y
Interchanging the order

∞ of integration in the double integrals shows that the first double
integral is equal to R(u, y) du, while the second double integral is equal to
0
∞ x−y ∞
fX (x) du dx = (x − y)fX (x) dx
y 0 y
∞
= xfX (x) dx − yP[X ≥ y]
y
∞
= (1 − FX (x)) dx
y
after integration by parts. Substitution now completes the proof after using c = (1+θ )λ E[X].
This formula has two useful consequences. First, by taking y = 0, the probability
that the surplus ever drops below its initial level is 1/ (1 + θ ). Second, an explicit
formula for the size of the drop below the initial level is obtained as
y
1
P[L1 ≤ y| Tu < ∞] = (1 − FX (x)) dx.
E[X] 0
This expression can be evaluated in certain cases.
Exercise 30–4. Derive this expression for P[L1 ≤ y| Tu < ∞].
Exercise 30–5. What is the conditional distribution of L1 given Tu < ∞ if the claim
size has an exponential distribution with mean 1/ δ ?
Exercise 30–6. Show that the conditional moment generating function of L1 given
Tu < ∞ is (MX (t) − 1)/ (tE[X]).
This information can also be used to study the random variable L which repre-

sents the maximum aggregrate loss and is defined by L = maxt≥0 { N(t) k=1 Xk − ct}.
Note that P[L ≤ u] = 1 − ψ (u) from which it is immediately seen that the distribution
of L has a discontinuity at the origin of size 1 − ψ (0) = θ / (1 + θ ), and is continuous
otherwise. In fact a reasonably explicit formula for the moment generating function
of L can be obtained.
N(t)
Theorem. If N(t) is a Poisson process and L = maxt≥0 { k=1 Xk − ct} then
θ E[X]ν
ML (ν ) = .
1 + (1 + θ )E[X]ν − MX (ν )
proof : Note from above that the size of each new deficit does not depend on the initial starting
point of the surplus process. Thus

D
L= Aj
j=1
where A1 , A2 , . . . are independent identically distributed random variables each having the
same distribution as the conditional distribution of L1 given Tu < ∞, and D is a random
variable independent of the A’s which counts the number of times a new deficit level is
reached. From here it is a simple matter to compute the moment generating function of
L.
Exercise 30–7. Complete the details of the proof.
This formula for the moment generating function of L can sometimes be used
to find an explicit formula for the distribution function of L, and hence ψ (u) =
1 − P[L ≤ u].
There are some other interesting consequences of the assumption that the claim
number process N(t) is a Poisson process.
d
First, a bit of notation. If A and B are random variables, write A = B to denote
that A and B have the same distribution.
A random variable S is said to have the compound Poisson distribution
with Poisson parameter λ and mixing distribution F(x), denoted S = CP(λ , F),
d

if S = Nj=1 Xj where X1 , X2 , . . . are independent identically distributed random vari-
d
ables with common distribution function F and N is a random variable which is

independent of the X’s and has a Poisson distribution with parameter λ .
Example 30–2. For each fixed t, the aggregate claims process CP(λ t, FX ).
Example 30–3. If S = CP(λ , F) then the moment generating function of S is

d
∞
MS (ν ) = exp{λ (euν − 1) dF(u)}.
−∞
This follows from the earlier general derivation of the moment generating function
of a random sum.
Exercise 30–8. Suppose that S = CP(λ , F) and T = CP(δ, G) and that S and T are
d d
independent. Show that S + T = CP(λ + δ, λλ+δ F + λ δ+δ G).

d
This last property is very useful in the insurance context. Because of this
property the results of the analysis of different policy types can be easily combined
into one grand analysis of the company’s prospects as a whole. A compound Poisson
distribution can also be decomposed.
Example 30–4. Suppose each claim is either for $1 or $2, each event having
probability 0.5. If the number of claims is Poisson with parameter λ then the
amount of total claims, S, is compound Poisson distributed with moment generating
function
MS (ν ) = exp{0.5λ (eν − 1) + 0.5λ (e2ν − 1)}.
d
Hence S = Y1 + 2Y2 where Y1 and Y2 are independent Poisson random variables with
mean λ / 2. Thus the number of claims of each size are independent!
Example 30–5. The collective risk model can be used as an approximation to

the individual risk model. In the individual risk model the claim amount is often
represented by a product Bj Xj in which B is a Bernoulli random variable which
represents whether a claim is paid or not and X is the amount of the claim. Then

B
N
Xj′ ≈ Xj′
d
BX =
j=1 j=1
where N has a Poisson distribution with parameter P[B = 1] and X1′ , X2′ , . . . are
independent random variables each having the same distribution as X. Thus the
distribution of BX may be approximated by the CP(P[B = 1], FX ) distribution.
Problems
Problem 30–1. If N has a Poisson distribution with parameter λ express P[N = k]

in terms of P[N = k − 1]. This gives a recursive method of computing Poisson
probabilities.
Problem 30–2. Show that if X takes positive integer values and S = CP(λ , FX )
d

then x P[S = x] = ∞k=1 λ kP[X = k]P[S = x − k] for x > 0. This is called Panjer’s
recursion formula. Hint: First show, using symmetry, that E[Xj | S = x, N = n] =
x/ n for 1 ≤ j ≤ n and then write out what this means.
Problem 30–3. Suppose in the previous problem that λ = 3 and that X takes on the
values 1, 2, 3, and 4 with probabilities 0.3, 0.2, 0.1, and 0.4 respectively. Calculate
P[S = k] for 0 ≤ k ≤ 40.
Problem 30–4. Suppose S1 has a compound Poisson distribution with λ = 2 and

that the compounded variable takes on the values 1, 2, or 3 with probabilities 0.2,
0.6, and 0.2 respectively. Suppose S2 has a compound Poisson distribution with
parameter λ = 6 and the compounded variable takes on the values 3 or 4 with
probabilities 1/2 each. If S1 and S2 are independent, what is the distribution of
S1 + S2 ?
Problem 30–5. The compound Poisson distribution is not symmetric about its
mean, as the normal distribution is. One might therefore consider approximation of
the compound Poisson distribution by some other skewed distribution. A random
variable G is said to have the Gamma distribution with parameters α and β if G has
density function
β α α −1 −β x
fG (x) = x e 1(0,∞) (x).
Γ(α )
It is useful to recall the definition and basic properties of the Gamma function in this
connection. One easily computes the moments of such a random variable. In fact
the moment generating function is MG (ν ) = (β / β − ν )α . The case in which β = 1/ 2
and 2α is a positive integer corresponds to the chi–square distribution with 2α
degrees of freedom. Also the distribution of the sum of n independent exponential
random variables with mean 1/ β is a gamma distribution with parameters n and β .
For approximation purposes the shifted gamma distribution is used to approximate
the compound Poisson distribution. This means that an α , β , and x is found so that
x + G has approximately the same distribution as the compound Poisson variate.
The quantities x, α , and β are found by using the method of moments. The first
three central moments of both random variables are equated, and the equations are
then solved. Show that when approximating the distribution of a compound Poisson
random variable S the method of moments leads to
3
4[Var(S)] 3 4λ E[X 2 ]
α= = 2
[E[(S − E[S])3 ]]2 E[X 3 ]
2Var(S) 2E[X 2 ]
β= =
E[(S − E[S])3 ] E[X 3 ]
2
2[Var(S)]2 2λ E[X 2 ]
x = E[S] − = λ E[X] − .
E[(S − E[S])3 ] E[X 3 ]
Problem 30–6. A random variable X is a mixture of exponential random variables

if the value of X is determined in the following way. Fix a number 0 < p < 1.
Perform a two stage experiment. In the first stage, select a number U at random in
the interval (0, 1). For the second stage, proceed as follows. If U < p select the
value of X to be the value of an exponential random variable with parameter λ1 . If
U > p select the value of X to be the value of an exponential random variable with
parameter λ2 . Show that the density of X is fX (x) = pλ1 e−λ1 x + (1 − p)λ2 e−λ2 x for
x ≥ 0. Show that the moment generating function of X is E[etX ] = λ1p−t + λ1−p
2 −t
.
Problem 30–7. What is the density of a random variable X with moment generating
function E[etX ] = (30 − 9t)/ 2(5 − t)(3 − t) for 0 < t < 3?
Problem 30–8. In the continuous time model, if the individual claims X have
density fX (x) = (3e−3x + 7e−7x )/ 2 for x > 0 and θ = 1, find the adjustment coefficient
and ψ (u).
Problem 30–9. In the continuous time model, if the individual claims X are discrete
with possible values 1 or 2 with probabilities 1/ 4 and 3/ 4 respectively, and if the
adjustment coefficient is ln(2), find the relative security loading.
Problem 30–10. Use integration by parts to show that the adjustment coefficient in
∞
the continuous time model is the solution of the equation erx (1 − FX (x)) dx = c/ λ .
0
Problem 30–11. In the continuous time model, use integration by parts to find
ML1 (t). Find expressions for E[L1 ], E[L21 ] and Var(L1 ). Here L1 is the random
variable which is the amount by which the surplus first falls below its initial level,
given that this occurs.
Problem 30–12. Find the moment generating function of the maximum aggregate
loss random variable in the case in which all claims are of size 5. What is E[L]?
Hint: Use the Maclaurin expansion of MX (t) to find the Maclaurin expansion of
ML (t).
Problem 30–13. If ψ (u) = 0.3e−2u + 0.2e−4u + 0.1e−7u , what is the relative security
loading?
Problem 30–14. If L is the maximum aggregate loss random variable, find expres-
sions for E[L], E[L2 ], and Var(L) in terms of moments of X.
Problem 30–15. In the compound Poisson continous time model suppose that
λ = 3, c = 1, and X has density fX (x) = (e−3x + 16e−6x )/ 3 for x > 0. Find the relative
security loading, the adjustment coefficient, and an explicit formula for the ruin
probability.
Problem 30–16. In the compound Poisson continous time model suppose that
9x
λ = 3, c = 1, and X has density fX (x) = e−3x/5 for x > 0. Find the relative security
25
loading, the adjustment coefficient, and an explicit formula for the ruin probability.
What happens if c = 20?
Problem 30–17. The claim number random variable is sometimes assumed to have
the negative binomial distribution. A random variable N is said to have the negative
¯
binomial distribution with parameters p and r if N counts the number of failures
before the rth success in a sequence of independent Bernoulli trials, each having
success probability p. Find the density and moment generating function of a random
variable N with the negative binomial distribution. Define the compound negative
binomial distribution and find the moment generating function, mean, and variance
of a random variable with the compound negative binomial distribution.
Problem 30–18. In the case of fire insurance the amount of damage may be quite
large. Three common assumptions are made about the nature of the loss variables in
d
this case. One is that X has a lognormal distribution. This means that X = eZ where
Z = N(µ, σ 2 ). A second possible assumption is that X has a Pareto distribution.
d
This means that X has a density of the form α x0 / xα +1 1[x0 ,∞) (x) for some α > 0. Note
that a Pareto distribution has very heavy tails, and the mean and/or variance may
not exist. A final assumption which is sometimes made is that the density of X is a
mixture of exponentials, that is,
fX (t) = (0.7)λ1 e−λ1 t + (0.3)λ2 e−λ2 t
for example. After an assumption is made about the nature of the underlying
distribution one may use actual data to estimate the unknown parameters. For each
of the three models find the maximum likelihood estimators and the method of
moments estimators of the unknown parameters.
Problem 30–19. For automobile physical damage a gamma distribution is often

postulated. Find the maximum likelihood and method of moments estimators of the
unknown parameters in this case.
Problem 30–20. One may also examine the benefits, in terms of risk reduction,
of using reinsurance. Begin by noting the possible types of reinsurance available.
First there is proportional reinsurance. Here the reinsurer agrees to pay a fraction
α , 0 ≤ α ≤ 1, of each individual claim amount. Secondly, there is stop–loss
reinsurance, in which the reinsurer pays the amount of the individual claim in
excess of the deductible amount. Finally, there is excess of loss reinsurance in
which the reinsurer pays the amount by which the claims of a portfolio of policies
exceeds the deductible amount. As an example, the effect of stop–loss reinsurance
with deductible d on an insurer’s risk will be analyzed. The amount of insurer’s
risk will be measured by the ruin probability. In fact, since the ruin probability is
so difficult to compute, the effect of reinsurance on the adjustment coefficient will
be measured. Recall that the larger the adjustment coefficient, the smaller the ruin
probability. Initially (before the purchase of reinsurance) the insurer’s surplus at
time t is

N(t)
U(t) = u + ct − Xj
j=1
where c = (1 + θ )λ E[X] and N(t) is a Poisson process with intensity λ . The

adjustment coefficient before the purchase of reinsurance is the positive solution of
λ + cr = λ MX (r).
After the purchase of stop loss reinsurance with deductible d the insurer’s surplus is

N(t)
U ′ (t) = u + c′ t − (Xj ∧ d)
j=1
where c′ = c − reinsurance premium. Note that this process has the same structure
as the original one. The new adjustment coefficient is therefore the solution of
λ + c′ r = λ MX∧d (r).
By examining the reinsurance procedure from the reinsurer’s standpoint it is clear

that the reinsurer’s premium is given by
(1 + θ ′ )λ E[(X − d)1[d,∞) (X)]
where θ ′ is the reinsurer’s relative security loading. With this information the new
adjustment coefficient can be computed. Carry out these computations when λ = 2,
θ = 0.50, θ ′ = 0.25, d = 750, and X has a exponential distribution with mean 500.
Problem 30–21. Repeat the previous problem for the case of proportional reinsur-
ance.
Problem 30–22. The case of excess of loss reinsurance leads to a discrete time
model, since the reinsurance is applied to a portfolio of policies and the reinsurance
is paid annually (say). The details here a similar to those in the discussion of the
discrete time gamblers ruin problem. Analyze the situation described in the previous
problem if the deductible for an excess of loss policy is 1500 and the rest of the
assumptions are the same. Which type of reinsurance is better?
Problem 30–23. In the discrete time model, suppose the X’s have the N(10, 4)
distribution and the relative security loading is 25%. A reinsurer will reinsure a
fraction f of the total portfolio on a proportional basis for a premium which is 140%
of the expected claim amount. Find the insurers adjustment coefficient as a function
of f . What value of f maximizes the security of the ceding company?
Problem 30–2. By symmetry, E[X1 | S = x, N = n] = x/ n while direct com-
∞
putation gives E[X1 | S = x, N = n] = k P[X1 = k| S = x, N = n]. Now
k=1
P[X1 = k, S = x, N = n] = P[X1 = k, nj=1 Xj = x, N = n] = P[X1 =

k] P[ nj=2 Xj = x − k] P[N = n] = P[X1 = k] P[ nj=2 Xj = x − k] P[N = n −
1]λ / n = P[X1 = k] P[S
= x − k, N = n − 1]λ / n. Making this substitution gives
xP[S = x, N = n] = ∞k=1 λ kP[X1 = k] P[S = x − k, N = n − 1]. Summing both
sides on n from 0 to ∞ gives the result.
Problem 30–4. The sum has a compound Poisson distribution with λ = 8.
Problem 30–6. Compute the distribution function of X by conditioning on

U to obtain FX (x) = P[X1 ≤ x] p + P[X2 ≤ x] (1 − p) for x ≥ 0 where X1 and
X2 are exponentially distributed random variables with parameters λ1 and λ2
respectively.
Problem 30–7. Use partial fractions and the previous problem to see that X
is a mixture of two exponentially distributed random variables with parameters
λ1 = 3 and λ2 = 5 and p = 1/ 4.
Problem 30–8. Here MX (t) = (5t − 21)/ (t − 3)(t − 7) and E[X] = 5/ 21. This
leads to R = 1.69. Also ML (t) = 1/ 2 − 0.769/ (t − 1.69) − 0.280/ (t − 6.20) using
partial fractions. Hence the density of L is fL (t) = 0.769e−1.69t + 0.280e−6.20t
together with a jump of size 1/ 2 at t = 0. (Recall that L has both a discrete and
absolutely continuous part.) Thus ψ (u) = 0.454e−1.69u + 0.045e−6.20u.
Problem 30–9. Here θ = 10/ 7 ln(2) − 1 = 1.0609.
Problem 30–11. The density of L1 is fL1 (t) = (1 − FX (t))/ E[X] for t ≥ 0.

Integration by parts then gives ML1 (t) = (MX (t) − 1)/ tE[X]. Using the Maclaurin
E[X 2]
expansion of MX (t) = 1 +tE[X] +t2 E[X 2 ]/ 2 + . . . then gives ML1 (t) = 1 + t+
2E[X]
E[X 3 ] 2
t + . . ., from which the first two moments of L1 can be read off.
6E[X]
Problem 30–12. ML (t) = 5θ t/ (1 + 5(1 + θ )t − e5t ) = 1 + t 2E[X

θ E[X] + . . ..
2
]
Problem 30–13. Here θ = 2/ 3 since ψ (0) = 1/ (1 + θ ).
Problem 30–14. Substitute the Maclaurin expansion of MX (t) into the expres-
sion for moment generating function of L in order to get the Macluarin expansion
of ML (t).
Problem 30–15. Here θ = 4/ 5 and R = 2. Also ML (t) = 4/ 9+(8/ 9) 2−t

1 1
+(4/ 9) 4−t
−2u −4u
so that ψ (u) = (4/ 9)e + (1/ 9)e .
Problem 30–16. Here MX (t) = 25 9

(3/ 5 − t)−2 so that E[X] = 10/ 3 and θ = −9/ 10
when c = 1. Since θ < 0, there is no adjustment coefficient and the ruin
1
probability is 1. When c = 20, θ = 1 and R = 0.215. Also ML (t) = − 0.119/ (t −
2
0.215) + 0.044/ (t − 0.834) by partial fractions. The density of the absolutely
continuous part of L is fL (t) = 0.119e−.0.215t − 0.044e−0.834t, and the distribution
of L has a jump of size 1/ 2 at the origin. So ψ (u) = 0.553e−0.215u − 0.053e−0.834u.
Problem 30–17. The compound negative binomial distribution is the distribu-

N
tion of the random sum Xi where N and the X’s are independent, N has the
i=1
negative binomial
distribution, and the X’s all have the same distribution. Now
P[N = k] = k+r−1
r−1 p r
(1 − p) k
for k ≥ 0 and MN (t) = pr (1 − (1 − p)et )−r . Now use
the general result about the moment generating function of a random sum.
Problem 30–20. Here MX (t) = (1 − 500t)−1 for t < 1/ 500. The adjustment
coefficient before reinsurance is then R = 1/ 1500. The reinsurance premium is
(1 + 0.25)2E[(X − 750)1(750,∞) (X)] = 278.91 and the insurer’s new adjustment
coefficient is the solution of 2 + (1500 − 278.91)R = MX∧750 (R) which gives
R = 0.00143.
Problem 30–21. As in the preceding problem, R = 1/ 1500 before reinsurance.

Suppose the insurer retains 100(1−α )% of the liability. The reinsurance premium
is then (1 + θ ′ )λ α E[X] = 1250α . The adjustment coefficient after reinsurance
is then the solution of 2 + (1500 − 1250α )R = 2M(1−α )X (R) = 2MX ((1 − α )R). So
R = (2− α )/ 500(5α 2 −11α +6) which is always at least 1/ 1500. Notice that since
θ ′ < θ here, the insurer should pass off all of the risk to the reinsurer. By using
α = 1 the insurer collects the difference between the original and reinsurance
premiums, and has no risk of paying a claim.
Problem 30–22. The computational details here are quite N complicated. In

a time interval of unit length the total claims are C = j=1 Xj where N is a
Poisson random variable with parameter λ . Now recall that in the discrete time
setting the adjustment coefficient is the solution of the equation E[eR(c−C) ] = 1.
As before c = 1500. Also MC (t) = eλ (MX (t)−1) . So the adjustment coefficient
before reinsurance is 1/ 1500. The reinsurance premium with deductible 1500
is λ (1 + θ ′ )E[(C − 1500)1(1500,∞)(C)] = 568.12. This is obtained numerically
by conditioning on the value of N and using the fact that conditional on N = k,
C has a gamma distribution with parameters α = k and β = 1/ 500. The new
adjustment coefficient solves E[eR(1500−568.12−C∧1500) ] = 1.
Problem 30–23. Here MX (t) = e10t+2t , the premium income is 12.5 for each
2
time period, and the adjustment coefficient is the solution of e−12.5t MX (t) = 1
which gives R = 1.25. The reinsurance premium is 14f so that after reinsurance
the adjustment coefficient satisfies e−(12.5−14f )t MX ((1 − f )t) = 1, which gives
R = (5 − 8f )/ 4(1 − 2f + f 2 ). The value f = 1/ 4 produces the maximum value of
R, namely 4/ 3.
N(T)
Exercise 30–1. Since R > 0 and u + cT − k=1 Xk ≤ 0 when T < ∞ the
denominator expectation is at least 1.
Exercise 30–2. MU(t)−u (−R) = 1 holds if and only if −ctR − λ t(MX (R) − 1) = 0,
which translates into the given condition.
Exercise 30–3. If c ≤ λ E[X] premium income is less than or equal to the

average rate of the claim process. So eventually the company will be ruined
by a run of above average size claims. By Maclaurin expansion, MX (R) =
1 + E[X]R + E[X 2]R2 / 2 + . . . and all of the coefficients are positive since X is a
positive random variable. So λ + cR − λ MX (R) = (c − λ E[X])R − E[X 2]R2 / 2 − . . .
is a function which is positive for R near 0 and negative for large values of R.
Thus there is some positive value of R for which this function is zero.
Exercise 30–4. From the definition of conditional probability, P[L1 ≤ y| Tu <

∞] = P[L1 ≤ y, Tu < ∞]/ P[Tu < ∞] and the result follows from the previous
formula and the fact that P[Tu < ∞] = 1/ (1 + θ ).
Exercise 30–5. Since in this case FX (t) = 1 − e−δ t for t > 0, direct substitution
gives P[L1 ≤ y| Tu < ∞] = 1 − e−δ y for y > 0.
Exercise 30–6. Given Tu < ∞ the density of L1 is (1−FX (y))/ E[X] for y > 0. Us-
ing integration
∞ by parts then gives the conditional moment
∞ generating function of
∞
L1 as 0 ety (1−FX (y))/ E[X] dy = ety (1 − FX (y))/ tE[X]0 + 0 ety fX (y)/ tE[X] dy =
(MX (t) − 1)/ tE[X]. Notice that the unconditional distribution of L1 has a jump of
size θ / (1 + θ ) at the origin. The unconditional moment generating function of
L1 is θ / (1 + θ ) + (MX (t) − 1)/ (1 + θ )tE[X].
Exercise 30–7. Since P[D = k] = (θ / (1 + θ ))(1/ (1 + θ ))k for k = 0, 1, 2, . . .,

t
D
A ∞
conditioning gives ML (t) = E[e j=1 j ] = E[MA (t)D ] = k=0 MA (t) (θ / (1 +
k
θ ))(1/ (1 + θ )) = (θ / (1 + θ ))/ (1 − MA (t)/ (1 + θ )) = θ / (1 + θ − MA (t)) and this

k
simplifies to the desired result using the formula of the previous exercise.
Exercise 30–8. Using the independence, MS+T (ν ) = MS (ν )MT (ν ) and the result
follows by substituion and algebraic rearrangement.
§31. Related Probability Models
In the next few sections some probability models are discussed which can be
used as models for transactions other than life insurance.
Discrete and continuous time Markov chains are often used as models for
a sequence of random variables which are dependent. One application of such
stochastic processes is as a model for the length of stay of a patient in a nursing
home.
The Brownian motion process is often used as a building block for a model of
stock prices. The definition and simple properties of Brownian motion are developed
here.

§32. Discrete Time Markov Chains
In many situations the random variables which serve naturally as a model are
not independent. The simplest kind of dependence allows future behavior to depend
on the present situation.
Example 32–1. Patients in a nursing home fall into 3 categories, and each category
of patient has a differing expense level. Patients who can care for themselves with
minimal assistance are in the lowest expense category. Other patients require some
skilled nursing assistance on a regular basis and are in the next higher expense
category. Finally, some patients require continuous skilled nursing assistance and
are in the highest expense category. One way of modeling the level of care a
particular patient requires on a given day is as follows. Denote by Xi the level of
care this patient requires on day i. Here the value of Xi would be either 1, 2, or
3 depending on which of the 3 expense categories is appropriate for day i. It is
intuitively clear that the random variables {Xi } are not independent.
Possibly the simplest type of dependence structure for a sequence of random

variables is that in which the future probabilistic behavior of the sequence depends
only on the present value of the sequence and not on the entire history of the
sequence. A sequence of random variables {Xn : n = 0, 1, . . .} is said to be a
Markov chain if
(1) P[Xn ∈ {0, 1, 2, 3, . . .}] = 1 for all n and
(2) for any real numbers a < b and any finite sequence of non-negative integers
t1 < t2 < ⋅ ⋅ ⋅ < tn < tn+1 ,
P[a < Xtn+1 ≤ b| Xt1 , . . . , Xtn ] = P[a < Xtn+1 ≤ b| Xtn ].
The second requirement is referred to as the Markov property.
The possible values of the chain are called states.
Exercise 32–1. Show that any sequence of independent discrete random variables
is a Markov chain.
Because of the simple dependence structure a vital role is played by the tran-
sition probabilities P[Xn+1 = j| Xn = i]. In principle, this probability depends not
only on the two states i and j, but also on n. A Markov chain is said to have
stationary transition probabilities if the transition probabilities P[Xn+1 = j| Xn = i]
do not depend on n. In the discussion here, the transition probabilities will always
be assumed to be stationary, and the notation Pi,j = P[Xn+1 = j| Xn = i] will be used.
The transition probabilities together with the distribution of X0 determine com-

pletely the probabilistic behavior of the Markov chain. This is seen in the following
§32: Discrete Time Markov Chains 129
computation.
P[Xn = in , . . . , X0 = i0 ]
= P[Xn = in | Xn−1 = in−1 , . . . , X0 = i0 ] × P[Xn−1 = in−1 , . . . , X0 = i0 ]
= P[Xn = in | Xn−1 = in−1 ] × P[Xn−1 = in−1 , . . . , X0 = i0 ]
= Pin−1 ,in P[Xn−1 = in−1 , . . . , X0 = i0 ]
= ...
= Pin−1 ,in Pin−2 ,in−1 . . . Pi0 ,i1 P[X0 = i0 ].
Exercise 32–2. Justify each of the steps here completely. Where was the Markov
property used?
Exercise 32–3. Show that for a Markov chain with stationary transition probabilities
P[X4 = 3, X3 ≠ 3| X2 = 3, X1 ≠ 3, X0 = 3] = P[X2 = 3, X1 ≠ 3| X0 = 3]. Generalize.
For a Markov chain with stationary transition probabilities it is useful to collect

the transition probabilities into the transition matrix P = [Pi,j ] of the chain. This
matrix may be infinite in extent.

Exercise 32–4. Show that if P is a transition matrix then j Pi,j = 1 for each i.
Example 32–2.  In the previous nursing home example, suppose the transition
0.9 0.05 0.05
 
matrix is P =  0.1 0.8 0.1 . Then using conditioning it is easy to compute
0 0.05 0.95
P[X3 = 2| X0 = 1].
Example 32–3. The gambler’s ruin problem illustrates many of the features of a
Markov chain. A gambler enters a casino with $z available for wagering and sits
down at her favorite game. On each play of the game, the gambler wins $1 with
probability p and loses $1 with probability q = 1 − p. She will happily leave the
casino if her fortune reaches $a > 0, and will definitely leave, rather unhappily,
if her fortune reaches $0. Denote by Xn the gambler’s fortune after the nth play.
Clearly {Xn } is a Markov chain with P[X0 = z] = 1. The natural state space here is
{0, 1, . . . , a}.
Exercise 32–5. Find the (a + 1) × (a + 1) transition matrix.
Even with the simplifying assumption of stationary transition probabilities the

formula for the joint distribution of the values of the chain is unwieldy, especially
since in most cases it is the long term behavior of the chain that is of interest.
Fortunately, it is possible to find relatively simple answers to the following central
questions.
(1) If {Xn } is a Markov chain with stationary transition probabilities, what is the
limiting distribution of Xn ?
(2) If s is a state of a Markov chain with stationary transition probabilities how
often is the process in state s?
As a warm up exercise for studying these questions the n step transition

probabilities defined by Pni,j = P[Xn+m = j| Xm = i] and the corresponding n step
transition probability matrix P(n) will now be computed.
Exercise 32–6. Show that P[Xn+m = j| Xm = i] does not depend on m.
Theorem. The n step transition probability matrix is given by P(n) = Pn where P is

the transition probability matrix.
proof : The case n = 1 being clear, the induction step is supplied.
Pni,j = P[Xn+m = j| Xm = i]
∞

= P[[Xn+m = j] ∩ [Xn+m−1 = k] ]| Xm = i]
k=0
∞

= P[[Xn+m = j, Xn+m−1 = k]| Xm = i]
k=0
∞
= P[Xn+m = j| Xn+m−1 = k, Xm = i] P[Xn+m−1 = k| Xm = i]
k=0
∞
= P[Xn+m = j| Xn+m−1 = k] P[Xn+m−1 = k| Xm = i]
k=0
∞
= i,k .
Pk,j P(n−1)
k=0
The induction hypothesis together with the formula for the multiplication of matrices
conclude the proof.
Using this lemma gives the following formula for the density of Xn in terms of
the density of X0 .
( P[Xn = 0] P[Xn = 1] . . . ) = ( P[X0 = 0] P[X0 = 1] . . . )Pn .
Exercise 32–7. Verify that this formula is correct.
Consequently, if Xn converges in distribution to Y as n → ∞ then

( P[Y = 0] P[Y = 1] . . . ) = n→∞
lim ( P[Xn = 0] P[Xn = 1] . . . )
= lim ( P[X0 = 0] P[X0 = 1] . . . )Pn
n→∞
= n→∞
lim ( P[X0 = 0] P[X0 = 1] . . . )Pn+1
= ( P[Y = 0] P[Y = 1] . . . )P
which gives a necessary condition for Y to be a distributional limit for the chain,
namely, the density of Y must be a left eigenvector of P corresponding to the
eigenvalue 1.
Example 32–4. For the nursing home chain given earlier there is a unique left
eigenvector of P corresponding to the eigenvalue 1, after normalizing so that the
sum of the coordinates is 1. That eigenvector is (0.1202, 0.1202, 0.7595). Thus a
patient will, in the long run, spend about 12% of the time in each of categories 1
and 2 and about 76% of the time in category 3.
Exercise 32–8. Find the left eigenvectors corresponding to the eigenvalue 1 of the
transition matrix for the gambler’s ruin chain.

0 1
Example 32–5. Consider the Markov chain with transition matrix P = .
1 0
This chain will be called the oscillating chain. The left eigenvector of P corre-
sponding to the eigenvalue 1 is ( 1/ 2 1/ 2 ). If the chain starts in one of the states,
there is clearly no limiting distribution.
Exercise 32–9. Show that this last chain does not have a limiting distribution.
The oscillating chain example shows that a Markov chain need not have a
limiting distribution. Even so, this chain does spend half the time in each state, so
the entries in the left eigenvector do have an intuitive interpretation as properties of
the chain. To explore this possiblilty further, some terminology is introduced. Let
P be the transition probability matrix of a Markov chain X with stationary transition
probabilities. A vector π = ( π 0 , π 1 , . . . ) is said to be a stationary distribution for
the chain X if
(1) π i ≥ 0 for all i,

(2) ∞i=0 π i = 1, and
(3) π P = π .
If a limiting distribution for the chain X exists then that limiting distribution will
be a stationary distribution. In the example above, ( 1/ 2 1/ 2 ) is a stationary
distribution even in though the chain has no limiting distribution.
From a relative frequency viewpoint a stationary distribution should arise as the

1 L
limit of the occupation times: π i = limL→∞ L+1 n=0 1{i} (Xn ). Indeed, this is the case
even in the rather poorly behaved example above. It will be easier to differentiate
the possible cases that can arise by studying the states of the chain more closely.

For each state i define the random variable Ni = ∞n=0 1{i} (Xn ). Clearly Ni counts
the total number of visits of the Markov chain X to the state i. It is possible that
Ni = ∞. A state i for which P[Ni = ∞| X0 = i] = 1 is a state which is sure to be
revisited infinitely many times. Such a state is said to be recurrent. A non-recurrent
state, that is, a state i for which P[Ni = ∞| X0 = i] < 1 is said to be transient. It is a
rather amazing fact that for a transient state i, P[Ni = ∞| X0 = i] = 0. Thus for each
state i the random variable Ni is either always infinite or never infinite.
Exercise 32–10. Show that if i is a transient state then Ni is a geometric random

variable.
Checking each state to see whether that state is transient or recurrent is clearly
a difficult task with only the tools available now. Some other useful notions can
greatly simplify the job.
One key notion is that of accessibility. The state j is said to be accessible from
the state i if there is a positive probability that the chain can start in state i and reach
state j. Technically, the requirement for j to be accessible from i is that Pni,j > 0 for
some n ≥ 0. Two states i and j are said to communicate, denoted i↔j, if each is
accessible from the other.
Example 32–6. Consider the Markov chain X in which Xn denotes the outcome of
the nth toss of a fair coin in which 1 corresponds to a head and 0 to a tail. Clearly
0↔1.
Example 32–7. In the gambler’s ruin problem it is intuitively clear that the states
0 and a are accessible from any other state but do not communicate with any state
except themselves. Such states are absorbing. The other states all communicate
with each other.
Exercise 32–11. Prove that the intuition of the preceding example is correct.
Example 32–8. In the nursing home example, all states communicate with each
other.
It is simple to check that the communication relation between states is an equiv-

alence relation, and therefore partitions the state space of the chain into equivalence
classes. If the chain has but a single equivalence class the chain is said to be
irreducible.
It is an important fact that if i↔j then i is recurrent if and only if j is recurrent.

(Said briefly, recurrence is a class property.)
Exercise 32–12. For the coin tossing chain, is the state 1 recurrent?
Exercise 32–13. What are the recurrent states for the gambler’s ruin chain?
Because of the possible existence of transient states, the discussion of the

limiting behavior of the chain is a bit complicated. Denote by fi,j the probability that
the chain ever enters state j given that the chain is currently in state i. Denote by µi,i
the expected number of time steps between visits to state i. (For a transient state,
µi,i = ∞ and it is possible for µi,i = ∞ even for a recurrent state.) With this notation,
the central result in the theory of Markov chains can be stated. For any two states i
and j, given that the chain begins at time zero in state i,
1 L
lim 1{j} (Xn ) = fi,j / µj,j .
L→∞ L + 1
n=0
This is the result that was anticipated based on previous examples.
This result can be used to provide an interpretation of the entries in a stationary

distribution. To do this, take expectations of both sides of the above limit to obtain
1 L
lim Pni,j = fi,j / µj,j . This can be expressed in matrix terms by defining S to be
L→∞ L + 1
n=0
1 L
the matrix with entries fi,j / µj,j . The foregoing statement is then lim Pn = S.
L→∞ L + 1
n=0
If π is a stationary distribution, π P = π , and by multiplication of the limiting
statement by π , π S = π too. So every stationary distribution is in the row space of
the matrix S.
Now consider the case in all states i and j are recurrent and communicate. Then
fi,j = 1 for all i and j, and all of the rows of S are the same. Thus there is only one
stationary distribution and the value π j = 1/ µj,j , which is the expected fraction of the
time the chain spends in state j. Since there is only one stationary distribution in this
case, there is also at most one limiting distribution in this case. The oscillating chain
example shows that the stationary distribution need not be a limiting distribution.
If some of the states are transient, the rows of S may not all be the same. This
is in accord with the behavior of the gambler’s ruin chain which shows that the
limiting behavior may depend on the initial state.
Exercise 32–14. Show that if a Markov chain is irreducible and every state is
transient then there is no stationary distribution for the chain.
Even for an irreducible and recurrent chain a limiting distribution need not exist
for another reason. If µi,i = ∞ there is no stationary distribution. Such a chain is
null recurrent. If µi,i < ∞, then as above there is one stationary distribution. In
this case the chain is strongly ergodic (or positive recurrent).
To identify the cases in which the chain has a limiting distribution one additional
idea is needed. The period of the state i, denoted by d(i), is the greatest common
divisor of the set {n ≥ 1 : Pni,i > 0}. If this set is empty, the period is defined to be
0. If d(i) = 1 for all states of the chain then the chain is said to be aperiodic. If a
Markov chain is aperiodic, a limiting distribution exists (given that the chain starts
lim P[Xn = j| X0 = i] = fi,j / µj,j . If all states communicate
in a particular state) and n→∞
and are recurrent, this limiting distribution will not depend on the initial state.
Exercise 32–15. What is the period of each state in the nursing home chain? The
gambler’s ruin chain? The oscillating chain?
Example 32–9. For the gambler’s ruin chain ( c, 0, . . . , 0, 1 − c ) is a stationary

distribution for any 0 ≤ c ≤ 1. One easily sees that only the two recurrent classes
contribute non-zero probabilities here. A formula for fi,0 was found earlier.
Problems
Problem 32–1. Show that Pni,i = 0 if d(i) does not divide n. Show that if i↔j then
d(i) = d(j). (Said briefly, period is a class property.)
Problem 32–2. Suppose the chain has only finitely many states all of which com-
municate with each other. Are any of the states transient?
Problem 32–3. Suppose Ni,j is the total number of visits of the chain to state j given
∞

that the chain begins in state i. Show that for i ≠ j, E[Ni,j ] = E[Nk,j ] Pi,k . What
k=0
happens if i = j?
Problem 32–4. Suppose the chain has both transient and recurrent states. Relabel
the states so that
the transient
states are listed first. Partition the transition matrix in
PT Q
to blocks P = . Explain why the lower left block is a zero matrix. Show
0 PR
that the T × T matrix of expectations E[Ni,j ] as i and j range over the transient states
is (I − PT )−1 .
Problem 32–5. Show that if i ≠ j and both are transient, fi,j = E[Ni,i ]/ E[Nj,j ]. What
happens if i = j?
Problem 32–6. In addition to the 3 categories of expenses in the nursing home ex-
ample, consider also the possibilities of withdrawal from the home and death. Sup-

0.8 0.05 0.01 0.09 0.05 
 0.5 0.45 0.04 0.0 0.01 
 
 
pose the corresponding transition matrix is P =  0.05 0.15 0.70 0.0 0.10 
 
 0 .0 0 .0 0 .0 1 .0 0 .0 
0 .0 0 .0 0 .0 0 .0 1 .0
where the states are the 3 expense categories in order followed by withdrawal and
death. Find the stationary distribution(s) of the chain. Which states communicate,
which states are transient, and what are the absorption probabilities given the initial
state?
Problem 32–7. An auto insurance company classifies insureds in 2 classes: (1)

preferred, and (2) standard. Preferred customers have an expected loss of $400 in
any one year, while standard customers have an expected loss of $900 in any one
year. A driver who is classified as preferred this year has an 85% chance of being
classified as preferred next year; a driver classified as standard this year has a 40%
chance of being classified as standard next year. Losses are paid at the end of each
year and i = 5%. What is the net single premium for a 3 year term policy for an
entering standard driver?
Problem 32–1. Since i↔j there are integers a and b so that Pai,j > 0 and
Pbj,i > 0. As shown earlier, Pja+n+b
,j ≥ Pbj,i Pni,i Pai,j . If Pni,i > 0 this inequality shows
n 2
that Pja+b+n
,j > 0 too and therefore d(j) divides a + b + n. But since P2n i,i ≥ Pi,i a
similar inequality shows that d(j) divides a + b + 2n as well. Hence d(j) divides
a + b + 2n − (a + b + n) = n, and so d(j) ≤ d(i). Interchanging the roles of i and j
shows that d(i) ≤ d(j).
Problem 32–2. No. Since all states communicate, either all are transient or
all are recurrent. Since there are only finitely many states they can not all be
transient. Hence all states are recurrent.
Problem 32–3. The formula follows by conditioning on the first step leaving
∞

state i. When i = j the formula is E[Ni,i ] = 1 + E[Nk,i ] Pi,k , by the same
k=0
argument.
Problem 32–4. It is not possible to go from a recurrent state to a transient state.

Express the equations of the previous problem in matrix form and solve.
Problem 32–5. fi,i = (E[Ni,i ] − 1)/ E[Ni,i ].
Problem 32–6. The 3 expense category states communicate with each other and
are transient. The other 2 states are recurrent and absorbing. The probabilities
fi,j satisfy f0,4 = 0.8f0,4 + 0.05f1,4 + 0.05f2,4 + 0.09 and 2 other similar equations,
from which f0,4 = 0.382, f1,4 = 0.409, and f2,4 = 0.601.
Problem
32–7.
From the given information the transition matrix is P =
.85 .15
. The two year transition probabilities for an entering standard
.6 .4
driver are found from the second row of P2 to be ( .75 .25 ). The premium is
900v + (.6 × 400 + .4 × 900)v2 + (.75 × 400 + .25 × 900)v3 = 1854.90.
Exercise 32–1. Because of the independence both of the conditional probabil-
ities in the definition are equal to the unconditional probability P[a < Xtn+1 < b].
Exercise 32–2. This is just the definition of conditional probability together

with the use of the Markov property to simplify each of the conditional proba-
bilities.

Exercise 32–3. Write P[X4 = 3, X3 ≠ 3| X2 = 3, X1 ≠ 3, X0 = 3] = j≠3 P[X4 =
3, X3 = j| X2 = 3, X1 ≠ 3, X0 = 3]
= j≠3 P[X4 = 3, X3 = j, X2 = 3, X1 ≠ 3, X0 =
3]/ P[X2 = 3, X1 ≠ 3, X0 = 3] = j≠3 k≠3 P[X4 = 3, X3 = j, X2 = 3, X1 = k, X0 =
3]/ P[X2 = 3, X1 ≠ 3, X0 = 3] = j≠3 k≠3 P[X4 = 3, X3 = j| X 2 = 3,
X1 = k , X0 =
3]P[X2 = 3, X1 = k, X0 = 3]/ P[X2 = 3, X1 ≠ 3, X0 = 3] = j≠3 k≠3 P[X4 =
, X3 = j| X2 = 3]P[X2 = 3, X1 =
3 k, X0 = 3]/ P[X2 = 3, X1 ≠ 3, X0 = 3] =
j≠3 P[X4 = 3, X3 = j| X2 = 3] = j≠3 P[X2 = 3, X1 = j| X0 = 3] = P[X2 =
3, X1 ≠ 3| X0 = 3], where the Markov property and stationarity have been used.

Exercise 32–4. j Pi,j = j P[X1 = j| X0 = i] = P[X1 ∈ R | X0 = i] = 1.
 
1 0 0 0 0 ... 0
q 0 p 0 0 ... 0
 
0 q 0 p 0 ... 0
Exercise 32–5. 
0 0 q 0 p ...

0 .
. .. .. .. .. .. .. 
. 
. . . . . . .
0 0 0 0 0 ... 1
Exercise 32–6. Use induction on n. The case n = 1 is true from the definition
of stationarity. For the induction
step assume the result holds when n = k. Then
P[Xk+1+m = j| Xm = i] = b P[Xk+1+m = j, X k+m = b| X m = i] = b P[Xk+1+m =
j| Xk+m = b]P[Xk+m = b| Xm = i] = b P[Xk+1 = j| Xk = b]P[Xk = b| X0 = i] =
P[Xk+1 = j| X0 = i], as desired.

Exercise 32–7. P[Xn = k] = i P[Xn = k| X0 = i]P[X0 = i] = i Pni,k P[X0 = i]
which agrees with the matrix multiplication.
Exercise 32–8. Matrix multiplication shows that the left eigenvector condition
implies that the left eigenvector x = (x0 , . . . , xa ) has coordinates that satisfy
x0 + qx1 = x0 , qx2 = x1 , pxk−1 + qxk+1 = xk for 2 ≤ k ≤ a − 2, pxa−2 = xa−1 and
pxa−1 + xa = xa . From these equations it follows that only x0 and xa can be
non-zero, and that these two values can be arbitrary. Hence all left eigenvectors
corresponding to the eigenvalue 1 are of the form (c, 0, 0, . . . , 0, 1 − c) for some
0 ≤ c ≤ 1.
Exercise 32–9. P[Xn = 1] is 0 or 1 depending on whether n is odd or even, so

this probability has no limit.
Exercise 32–10. Let p = P[Ni < ∞| X0 = i]. Then P[Ni = k| X0 = i] = pk (1 − p)

for k = 0, 1, . . ..
Exercise 32–11. If the current fortune is i, and i is not 0 or a, then it is possible

to obtain fortune j is | j − i| plays of the game by having | j − i| wins (or losses)
in a row.
Exercise 32–12. Yes, 1 is recurrent since state 1 is sure to be visited infinitely

often.
Exercise 32–13. The only recurrent states are 0 and a.
Exercise 32–14. Under these assumptions, fi,j = 0 so the limit above is always
zero. Hence there is no stationary distribution.
Exercise 32–15. For the nursing home chain and the gambler’s ruin chain
each state has period 1. For the oscillating chain each state has period 2. This
explains why the oscillating chain does not have a limiting distribution.
§33. Continuous Time Markov Chains
The next step in the study of Markov chains retains a discrete state space but
allows time to vary continuously.
A continuous time Markov chain is a stochastic process {Xt : t ≥ 0} with state

space {0, 1, 2, . . .} which satisfies the Markov property: whenever n > 0 and 0 ≤
t1 < t2 < . . . < tn < t then P[Xt = xt | Xtn = xtn , . . . , Xt1 = xt1 ] = P[Xt = xt | Xtn = xtn ].
As in the discrete time case the discussion here assumes that the Markov chain has
stationary transition probabilities, that is, P[Xt+s = i| Xt = j] does not depend on t.
A discrete time chain always spent one time unit in each state before the next
transition. For a continuous time chain the time spent in each state before a transition
will be random. Intuitively, this is the only difference between discrete time and
continuous time Markov chains. The bulk of the work consists of verifying that this
intuition is indeed correct and in identifying the probabilistic properties of the times
between transitions for the chain.
As for discrete time chains the transition probabilities Pi,j (t) = P[Xt = j| X0 = i]
and the associated transition probability matrix P(t) = [Pi,j (t)] play an important
role. By convention, P(0) = I, the identity matrix.
The first result parallels a result for discrete time chains.

Transition Semi-group Property. The transition matrices {P(t) : t ≥ 0} form a
semi-group under matrix multiplication, that is, P(s + t) = P(s)P(t) for all s, t ≥ 0.
Exercise 33–1. Prove the theorem.
For a discrete time chain, the one step transition probability matrix determined
all of the interesting properties of the chain. For continuous time chains, a similar
role is played by the matrix P′ (0), which is called the infinitesimal generator of
the Markov chain (or of the transition semi-group {P(t) : t ≥ 0}). Since typically
the infinitesimal generator is specified during model building the central question is
how the infinitesimal generator determines properties of the chain.
Example 33–1. The Poisson process provides a typical example of the way in which
the transition probabilities are specified for continuous time Markov chains. As a
specific example let Xt denote the number of calls that have arrived at a telephone
switchboard by time t. In a very short time interval essentially either 0 or 1 call can
arrive. Denote by λ > 0 the average rate at which calls arrive. A model for the
behavior of the chain Xt can be expressed by saying that for small h > 0
(1) P[Xt+h − Xt = 1| Xt = i] ≈ λ h,
(2) P[Xt+h − Xt = 0| Xt = i] ≈ 1 − λ h, and
§33: Continuous Time Markov Chains 140
(3) P[Xt+h − Xt ≥ 2| Xt = i] ≈ 0
or equivalently in terms of the transition probabilities

(1) Pi,i (h) ≈ 1 − λ h,
(2) Pi,i+1 (h) ≈ λ h, and
(3) Pi,j (h) ≈ 0 if j ≠ i, i + 1.
Since P(0) = I these assumptions show that the derivative of P(t) exists at t = 0 and
is given by
 
−λ λ 0 0 ...
 0 −λ λ 0 . . . 
P′ (0) = 


.
 0 0 −λ λ . . . 
.. .. .. .. ..
. . . . .
In the case of the Poisson process, the row sums of the generator are zero. An
infinitesimal generator for which the row sums are all zero is conservative. Also,
the off-diagonal entries of the generator are non-negative, while the diagonal entries
of the generator are non-positive.
Exercise 33–2. Show that if the chain has only finitely many states then the in-
finitesimal generator must be conservative.
Exercise 33–3. How could a generator fail to be conservative?
Exercise 33–4. Are the off-diagonal entries of an infinitesimal generator always

non-negative? Are the diagonal entries of an infinitesimal generator always non-
positive?
Suppose P(t) is a Markov semi-group which is continuous at 0. Then using

results from real analysis, the infinitesimal generator A = P′ (0) exists. Using the
semi-group property gives
P(t + h) − P(t)
P′ (t) = lim
h→0 h
P(h)P(t) − P(t)
= lim
h→0

h
P(h) − I
= lim P(t)
h→0 h
= A P(t).
This equation is known as Kolmogorov’s backward equation. The same argument

leads also to Komogorov’s forward equation P′ (t) = P(t) A.
Exercise 33–5. Why aren’t these derivations rigorous?

These two equations, especially the forward equation, are very useful in appli-
cations. Unfortunately there are no simple general conditions which guarantee that
the forward equation holds.
In the case of chains with only a finite number of states the theory is very simple:
the transition semi-group is uniquely determined by the infinitesimal generator
(which must be conservative) and both the forward and backward equations hold.
Also, the semi-group can be obtained from the generator by the formula P(t) = etP′ (0) .
In this formula the exponential of a matrix is computed from the Maclaurin series
for the exponential function.
The case in which the chain has infinitely many states is more complex. The
main result in this case is just stated here.
Theorem. Suppose the infinitesimal generator is conservative. If P(t) is the unique

solution of either the forward or the backward equation and if ∞j=0 Pi,j (t) = 1 for all
i then P(t) is the unique solution to both the forward and the backward equation and
is the unique transition semi-group with the given generator.
Example 33–2. Assume for the moment that the forward equation holds for the
Poisson process. (Later this will be shown to be true.) Recall that the infinitesimal
generator in this case is
 
−λ λ 0 0 ...
 0 −λ λ 0 ...
 
A = P′ (0) =  .
 0 0 −λ λ ...
.. .. .. .. ..
. . . . .
Translating the matrix form of the forward equation P′ (t) = P(t)A into statements
about the individual transition probabilities gives
P′i,j (t) = −λ Pi,j (t) + λ Pi,j−1 (t).
To compute M(t) = E[Xt | X0 = i] use the definition of expectation and this equation
to obtain ∞
M ′ (t) = jP′i,j (t)
j=0
∞

= j −λ Pi,j (t) + λ Pi,j−1 (t)
j=0
∞
∞

= −λ jPi,j (t) + λ (j − 1)Pi,j−1 (t) + λ
j=0 j=0
= λ.
So M(t) = λ t + i. Thus the forward equation can be used to find the conditional
expectation without first finding the transition matrices.
Exercise 33–6. Try the same computation using the backward equations.
A method will now be developed to directly interpret the entries of the generator
in terms of the probabilistic behavior of the chain.
Define T0 = 0 and inductively set Tn = inf{t ≥ Tn−1 : Xt ≠ XTn−1 }. The T’s are
the times of changes of state for the chain.
To study the behavior of the chain let t > 0 and i ≠ j be states. Then
P[T1 > t, XT1 = j| X0 = i]

∞

= lim P[Xk/2n = i, 1 ≤ k ≤ [2n t] + l, X([2n t]+l+1)/2n = j| X0 = i]
n→∞
l=0
∞
[2n t+l]
= lim Pi,i (1/ 2n) Pi,j (1/ 2n)
n→∞
l=0
∞
Pi,j (1/ 2n ) 2n [2n t]/2n 1
= lim Pi,i (1/ 2n ) Pi,i (1/ 2n)l
n→∞ 1/ 2n l=0 2 n
n
1/ 2
= ai,j eai,i t lim
n→∞ 1 − Pi,i (1/ 2n )
−ai,j ai,i t
= e .
ai,i
−ai,j
(Recall that ai,i ≤ 0.) Hence P[XT1 = j| X0 = i] = ai,i
and P[T1 > t| X0 = i] = eai,i t for
t > 0.
Exercise 33–7. Fill in the details for all of these calculations.
This computation means that the time spent in state i until a transition occurs has
an exponential distribution with mean −1/ ai,i and that the probability upon leaving
state i of entering state j is −ai,j / ai,i .
Exercise 33–8. Extend the argument above to show that

n−1
−ail−1 ,il
P[Tl > tl , XTl = il , 0 ≤ l ≤ n| X0 = i0 ] = eail−1 ,il−1 tl .
l=0 ail−1 ,il−1
The exercise justifies the following view of continuous time Markov chains.
Begin at time 0 in state i0 . Stay in state i0 for a random time T1 , where T1 has the
exponential distribution with mean −1/ ai0 ,i0 . Then move to state i1 with probability
−ai0 ,i1 / ai0 ,i0 . Remain in state i1 an exponentially distributed time, etc.
If interest is only in the states that are visited by the chain and not in the time
spent in each state one may as well study the embedded discrete time chain with
transition matrix P = [−ai,j / ai,i (1 − δi,j )]. Note that if ai,i = 0 the corresponding
transition probability in the embedded chain is Pi,i = 1 since state i is obviously an
absorbing state for the continuous time chain.
Example 33–3. For the Poisson process the embedded chain has transition matrix
 
0 1 0 0 0 ...
0 0 1 0 0 ...
 
P= .
0 0 0 1 0 ...
.. .. .. .. .. ..
. . . . . .
Hence all states are transient and there is no stationary distribution.
The relationship between the infinitesimal generator and the chain itself has been
made rather precise. The question of the behavior of the chain is now considered.
Suppose I = ( P[X0 = 0] P[X0 = 1] . . . ) is the initial distribution of the chain.

If a limiting distribution π for the chain exists then
π = t→∞
lim I P(t) = lim I P(t + s) = lim I P(t)P(s) = π P(s)
t→∞ t→∞
for all s ≥ 0.
Exercise 33–9. Verify that this computation is correct.
Once again the notion of a stationary distribution will play an important role.
In the continuous time context π = ( π 0 π 1 . . . ) is said to be a stationary
distribution if
(1) π i ≥ 0 for all i,

(2) ∞i=0 π i = 1, and
(3) π P(s) = π for all s ≥ 0.
Because of the close relationship between the infinitesimal generator and the
transition semi-group it should be possible to find the stationary distribution of the
chain using only the generator.
Theorem. Suppose the infinitesimal generator A of the chain is conservative and

that π i ≥ 0 for all i and ∞i=0 π i = 1. Then π is a stationary distribution if and only
if π A = 0.
There is no notion of period in the continuous time setting. Thus the distinction
between a stationary distribution and a limiting distribution depends only on whether
there are transient states, and the mean recurrence time for the recurrent states. This
makes the continuous time case somewhat simpler than the discrete time case.
Some additional examples of continuous time chains will now be given. One of
the main features is the way in which the behavior of the chain is determined from
the infinitesimal generator.
Example 33–4. Look once again at the Poisson process. Here the forward equation
is P′i,j (t) = −λ Pi,j (t) + λ Pi,j−1 (t). This can be rewritten using an integrating factor

as dtd eλ t Pi,j (t) = λ eλ t Pi,j−1 (t). Thus P0,0 (t) = e−λ t and by induction P0,n (t) =
(λ t)n e−λ t / n!. Similarly, Pi,0 (t) = 0 if i > 0 and by induction Pi,j (t) = (λ t)j−i e−λ t / (j −

i)!1[0,∞) (j − i). This solution is unique and ∞j=0 Pi,j (t) = 1 for all i. From the general
theory Pi,j (t) is therefore the unique solution to the backward equation as well and
is also the unique transition semi-group with this infinitesimal generator.
Exercise 33–10. Fill in the details of the induction arguments above.
Example 33–5. The next example is a pure birth process. This is a variant of the
Poisson process in which the probability of additional calls depends on the number
of calls received. Specifically assume that
(1) P[Xt+h − Xt = 1| Xt = k] ≈ λk h,
(2) P[Xt+h − Xt = 0| Xt = k] ≈ 1 − λk h, and
(3) P[Xt+h − Xt ≥ 2| Xt = k] ≈ 0.
As before this leads to the generator


−λ0 λ0 0 0 ...
 0 −λ1 λ1 0 ...
 
A=
−λ2 λ2 ... .
 0 0 
.. .. .. .. ..
. . . . .
Using the forward equation and the integrating factor eλj t yields
t
Pi,j (t) = δi,j e−λj t + λj−1 e−λj t eλj t Pi,j−1 (s) ds.
0
This shows inductively that Pi,j (t) ≥ 0 and that the solution is unique. The remaining

question is whether or not ∞j=0 Pi,j (t) = 1 for all i. Fix i and define Sn (t) = nj=0 Pi,j (t).

Using the forward equation gives Sn′ (t) = −λn Pi,n (t) and so 1 − Sn (t) = 0t −Sn′ (s) ds =
t ∞
λn 0 Pi,n (s) ds. Now from the definition of Sn (t), 1 − j=0 Pi,j (t) ≤ 1 − Sn (t) ≤ 1 so
 
∞ t
1  1
1− Pi,j (t) ≤ Pi,n (s) ds ≤ .
λn j=0 0 λn
Summing on n gives
 
∞
∞ t∞ ∞
1  1
1− Pi,j (t) ≤ Pi,n (s) ds ≤ .
n=0 λn j=0 0 n=0 n=0 λn

The right hand part of this inequality shows that if ∞n=0 λ1n < ∞ then ∞j=0 Pi,j (t) can

not be 1 for all t, while the left hand part of this inequality shows that if ∞n=0 λ1n = ∞

then ∞j=0 Pi,j (t) = 1 for all t ≥ 0. Thus the pure birth process exists if and only if
∞ 1
n=0 λn = ∞. Note that this condition is obviously satisfied for the Poisson process.
Exercise 33–11. Explain intuitively what goes wrong with the pure birth process

when ∞n=0 λ1n < ∞.
Example 33–6. The Yule process is a pure birth process for which λn = nβ , that
is, the birth rate is proportional to the number present. This process clearly meets
the criteria for existence which was established in the previous example. If Mi (t) =
E[Xt | X0 = i] then the forward equation can be used to show that Mi′ (t) = β Mi (t).
Hence Mi (t) = ieβ t .
Exercise 33–12. Fill in the details of how the forward equation is used in this
computation. What is the conditional variance of Xt ?
Example 33–7. The final example is the birth and death process. Suppose Xt
is the size of a population at time t. Then in a short time period there will be
(essentially) a single birth or a single death or neither. Formally the model is
(1) P[Xt+h − Xt = 1| Xt = k] ≈ λk h,
(2) P[Xt+h − Xt = 0| Xt = k] ≈ 1 − (λk + µk )h,
(3) P[Xt+h − Xt = −1| Xt = k] ≈ µk h, and
(4) P[| Xt+h − Xt | ≥ 2| Xt = k] ≈ 0.
Here λk is birth rate and µk is the death rate when the population size is k.
As before this leads to the generator

 
−λ0 λ0 0 0 ...

 µ1 −µ1 − λ1 λ1 0 ...
A=
 0 µ2 −µ2 − λ2 λ2 ...
.
 
.. .. .. .. ..
. . . . .
∞
As in the case of the pure birth process the condition 1
n=0 λn = ∞ will guarantee
that the process is well defined.
Example 33–8. For the birth and death process the embedded chain has transition
matrix  
0 1 0 0 ...
 µ1
P= 0 µ1λ+1λ1 0 . . . 

 µ1 +λ1 
.. .. .. .. ..
. . . . .
if λ0 > 0 so this chain behaves in a manner similar to the gambler’s ruin chain,
except that when the gambler reaches fortune 0, she begins again with 1 at the
next play. If λ0 = 0 and 0 < λi for i > 0 then finding the absorption probability
for the birth and death chain is exactly the same problem as finding the absorption
probability at 0 for the gambler’s ruin chain.
Exercise 33–13. How does the first row of P differ from that shown when λ0 = 0?
For a continuous time Markov chain with conservative generator A the stationary
distributions π are the solutions of π A = 0. Also −1/ ai,i is the mean of the
exponentially distributed time that the chain spends in state i prior to a transition
and −ai,j / ai,i is the probability that when a transition occurs from state i the chain
will move to state j ≠ i. The probability of a transition from state i to itself is 0 unless
ai,i = 0 in which case i is an absorbing state.
One consequence of the Markov property is that the time spent in each state is
exponential. In many models this is not realistic because the exponential distribution
is memoryless.
Problems
Problem 33–1. Consider a continuous time version of the nursing home chain in
which the states are (1) minimal care, (2) some skilled assistance required, (3)
continuous skilled assistance required, and (4) dead. Suppose the infinitesimal
generator A of the chain has entries a1,2 = 0.12, a1,3 = 0.05, a1,4 = 0.08, a2,1 = 0.05,
a2,3 = 0.07, a2,4 = 0.12, a3,4 = 0.20 and all other non-diagonal entries are zero.
What is the matrix A? What are the stationary distributions of the chain? What is
the expected time until absorption as a function of the initial state?
Problem 33–2. For the birth-death process show that a stationary distribution must
λj 1
satisfy π j+1 = π j for j ≥ 0 and that π 0 = ∞ λ0 ⋅⋅⋅λn−1 . Thus the chain has a
µj+1 1 + n=1 µ1 ⋅⋅⋅µn
∞
λ0 ⋅ ⋅ ⋅ λn−1
limiting distribution as long as < ∞. What happens if λ0 = 0?
n=1 µ1 ⋅ ⋅ ⋅ µn
Problem 33–3. Find the conditional moment generating function E[evXt | X0 = i] for
a Poisson process.
Problem 33–4. A continuing care retirement community is modeled by a contin-

uous time Markov chain with five classifications for the status of a patient: (1)
Individual Living Unit, (2) Skilled Nursing Facility (Temporary), (3) Skilled Nurs-
ing Facility (Permanent), (4) Dead, and (5) Withdrawn from Care. The non-zero
off-diagonal entries of the generator are

0 .3 0 .1 0 .1 0 .2 
 0 .5 0 .3 0 .3 
 
 


0 .4 .

 
If a patient is currently in the Individual Living Unit, what is the probability of more
than one visit to the Skilled Nursing Facility? What is the expected duration of a
visit to the Skilled Nursing Facility?
Problem 33–5. An HMO has 3 classes of patients: (1) healthy, (2) moderately
imparied, and (3) severely impaired. The status of a patient is modeled by a
continuous
 time Markov  chain with generator whose non-zero off-diagonal entries
1/ 2 1/ 2
 
are  1/ 3 1/ 3 . What is the stationary distribution of this chain? If a
1/ 4 1/ 4
healthy person arrives today, about how long will it be until this person becomes
severely impaired?
 
−0.25 0.12 0.05 0.08
 0.05 −0.24 0.07 0.12 
Problem 33–1. A =  , from which the sta-
0 0 −0.20 0.20
0 0 0 0
tionary distribution is easily found to be ( 0 0 0 1 ). If ei is the expected
time until absorption starting from state i then e1 = 4 + (.12/ .25)e2 + (.05/ .25)e3
and two other similar equations also hold. Solving gives e1 = 8.55, e2 = 7.40,
and e3 = 5.
Problem 33–2. The relation π A = 0 gives λ0 π 0 = µ1 π 1 and λi−1 π i−1 − (µi +

λi )π i + µi+1 π i+1 = 0 for i ≥ 1. Sum this last equality on i from 1 to j and use
telescoping and the first equation to simplify. The rest follows by normalization.
Problem 33–3. If g(t) = E[evXt | X0 = i] the forward equation shows that

g′ (t) = −λ g(t) + λ ev g(t). Now solve this differential equation and use the initial
condition g(0) = evi .
Problem 33–4. The probability of no visits to the Skilled Nursing Facility is

3/ 7. The probability of exactly one visit is 3/ 7(6/ 11 +(5/ 11)(3/ 7)) +1/ 7. Similar
reasoning applies when computing the expected duration, keeping in mind that
the probabilities are conditional on visiting the Skilled Nursing Facility.
Problem 33–5. The stationary distribution is π 1 = 2/ 9, π 2 = 3/ 9, and π 3 = 4/ 9.

If e1 and e2 are the expected waiting times until the first entry into state 3 given
the starting state, then e1 = 1/ 2 + 1/ 2(1 + e2 ) and e2 = 1/ 2(3/ 2) + 1/ 2(3/ 2 + e1 ).
Hence e1 = 7/ 3 (and e2 = 8/ 3).
Exercise 33–1. Hint: Look at the proof of the parallel result for discrete time
chains.
Exercise 33–2. Since there are only finitely many states, the derivative of
the sum is the sum of the derivatives, so the equation j Pi,j (t) = 1 can be
differentiated to show that the generator is conservative.
Exercise 33–3. If there are infinitely many states, the derivative and the sum
in the previous exercise might not be interchangeable, and this could make the
generator not be conservative.
Exercise 33–4. Yes in both cases. Look at the sign of the difference quotients
used to compute the entries of the generator.
Exercise 33–5. If there are infinitely many states there could be problems since
these ‘derivations’ involve passing a limit through the infinite sum occurring in
the matrix multiplication.
Exercise 33–6. The backward equation gives P′i,j (t) = λ Pi−1,j (t)− λ Pi,j (t) which
does not produce a differential equation for M(t).
Exercise 33–10. The inductive step is dtd (eλ t Pi,j (t)) = λ j tj−1 / (j − 1)! from which
integration gives eλ t Pi,j (t) = (λ t)j / j!, as desired.
Exercise 33–11. If the process is currently in state k the expected waiting time
until a birth is 1/ λk . If the sum is finite then there is a finite expected waiting
time for infinitely many births to occur, that is, the population size will become
infinite in a finite amount of time.
Exercise 33–12. The forward equation gives P′i,j (t) = β jPi,j (t) + β (j − 1)Pi,j−1(t).
Now multiply both sides of this equation by j and sum on j = j − 1 + 1 to obtain
the equation.
Exercise 33–13. When λ0 = 0 the first row is ( 1 0 0 0 . . . ).

§34. Introduction to Brownian Motion
The Brownian motion model originated as a description of the movement of

microscopic particles in liquid. Applications to modeling stock prices and other
situations quickly followed. Brownian motion is stochastic process for which both
time and state space are continuous. To make this transition from earlier studies
a simpler discrete time and state space model will be constructed which mimicks
some of the important features of Brownian motion. A suitable passage to the limit
will enable this simpler model to become a Brownian motion.
The discrete time and state space model will be constructed to model the motion
of a particle in one dimension. Let Sn denote the position of the particle at time n and
assume S0 = 0. The change in position of the particle at time i will be denoted Di .
The random variables D1 , D2 , . . . will be assumed to be independent and identically
distributed random variables each taking the values 1 and −1 with equal probability.

Hence Sn = ni=0 Di and {Sn : n ≥ 0} is nothing more than a random walk model.
A convenient visual and conceptual aid is the path generated by D1 , . . . , Dn

which consists of the vertices {(j, Sj ) : 0 ≤ j ≤ n} together with the line segments
joining the vertices. Many interesting probabilistic problems can be turned into path
counting problems by using this device.

Proposition. The number of paths joining the origin to a point (n, x) is n
n+x if
2
(n + x)/ 2 is a non-negative integer.
proof : Denote by Pn the number of displacements D1 , . . . , Dn which are positive and let Nn = n−Pn
denote the number of negative displacements. The D’s generate a path from the origin to
(n, x) if and only if n = Pn + Nn and x = Pn − Nn . This shows that there are Pnn such paths.
From the two equations Pn = (n + x)/ 2 and the result follows.
The reflection principle is another path inspired device. The notation corre-
sponds to that of the picture below, which also provides the proof.
Reflection Principle. The number of paths from A to B which touch or cross the
level T is equal to the number of all paths from A′ to B.
Sn
.•.......
A •......... .....•........... ....•...........
•. .•...............
.... ....
....•
... B
T •
....
. .
. •
.....
...
•..... .•.. . .•.......

.... ... ...
A′ •..... ...•.. ..... .•.... ..

•
.....
n
§34: Introduction to Brownian Motion 151
As an application consider an election in which candidate A wins with a votes
compared to the b < a votes for candidate B. If the ballots are counted one at a time,
what is the probability that candidate A is always ahead in the count?
Ballot Theorem. Suppose n and x are positive integers. The number of paths from
the origin to (n, x) which do not touch the n axis is n n+x .
x n
2
proof : The number of paths meeting the requirement is the same as the number of paths from
(1, 1) to (n, x) which do not touch or cross the n axis. Now the total number of paths from
,n−1
(1 1) to (n, x) is the same as the total number of paths from (0, 0) to (n − 1, x − 1), which is
n−1+x−1 by the proposition. The number of paths from (1, 1) to (n, x) which touch or cross
2
the n axis can be computed using the reflection principle. By the reflection principle this
is the same as the number of paths from (1, −1) to (n, x) which in turn is the same as the
of paths from (0, 0) to (n − 1, x + 1). Again using the proposition gives this number
number
as n−1 n+x . Subtraction gives the desired result.
2
Exercise 34–1. Use the Ballot Theorem to show that the probability that candidate
A is always ahead in the tally is (a − b)/ (a + b). Thus if A wins 70 votes to 30 the
probability that A was always ahead is 0.40.
Some of the basic properties of Brownian motion will now be developed by

passing to the limit, in an appropriate way, in the discrete time and state space
model.
One of the first applications of the Brownian motion model was to describe the
displacement of a small particle suspended in fluid. As a first approximation to the
behavior of the particle assume that the particle moves only because of collisions
with other particles. Assume also that collisions are as likely to come from the
left as from the right. To simplify even further suppose that each collision causes
a displacement of magnitude ∆ and that the time between collisions is τ . If the
particle begins at the origin at time 0 the position Xt of the particle at time t is
then ∆S[t/ τ ] in the notation above. The idea is to see what happens as both ∆
and τ go to 0. This passage to the limit makes the approximate model become
more realistic. Note that E[Xt ] = 0 and Var(Xt ) ≈ ∆2 t/ τ . To avoid disaster some
relationship between ∆2 and τ must be maintained in the passage to the limit. The
assumption made here is that ∆2 / τ = σ 2 for some σ > 0. After passing to the limit
under this assumption the Central Limit Theorem shows that Xt = N(0, σ 2 t). The
d
process Xt obtained in this way is also a process with independent increments.

This means that if t1 < t2 < . . . < tn then the random variables Xtn − Xtn−1 , . . . , Xt2 − Xt1
are independent. The increments are also stationary which means that for t > s
d
Xt − Xs = Xt−s . Intuition further suggests that the sample paths of the process Xt
should be continuous functions. This means that for each point ω in the underlying
probability space the function of t defined by t → Xt (ω ) should be continuous. This
intuitive motivation leads to the following more formal definition. A stochastic
process {Xt : t ≥ 0} is said to be a Brownian motion process (or Wiener process)
with parameter σ 2 if
(1) Xt is a process with stationary independent increments,
(2) Xt = N(0, σ 2 t) for each t > 0, and
d
(3) X0 = 0.
The term standard Brownian motion refers to the case in which σ 2 = 1.
Exercise 34–2. Use the independent increments property to show Cov(Xt , Xs ) =

σ 2 (t ∧ s).
The important but difficult to prove fact about Brownian motion is that the
sample paths of the process can be assumed to be continous. This result can be
intuitively deduced from the approximation argument given earlier. The sample
paths can also be shown to be nowhere differentiable. Again this fact is intuitively
apparent from the earlier construction since the sample paths of the approximating
sums have a lot of corners.
For some purposes the notion of a Brownian motion starting from x will be
useful. Such a process is a Brownian motion as above except for the fact that X0 = x
is assumed.
Some of the basic properties of Brownian motion are derived here in an intuitive
way. The continuous time analog of the reflection principle plays an important role.
Example 34–1. Suppose a > 0 and denote by Ta the first time at which the standard
Brownian motion Xt takes the value a. Then Ta = inf{s > 0 : Xs = a}. The
random variable Ta is well defined because of the continuity of the sample paths
of the process X. Let t > 0 be fixed. Now by the Theorem of Total Probability
P[Xt ≥ a] = P[Xt ≥ a| Ta ≤ t] P[Ta ≤ t] + P[Xt ≥ a| Ta > t] P[Ta > t]. Clearly
P[Xt ≥ a| Ta > t] = 0. If Ta ≤ t the independent increment property of X suggests
that after time Ta the Brownian motion restarts as though from scratch except that
the intial value is a rather than 0. Hence P[Xt ≥ a| Ta ≤ t] = P[Xt ≤ a| Ta ≤ t] = 1/ 2.

So P[Ta ≤ t] = 2 P[Xt ≥ a] = √22π a/∞√t e−x /2 dx. From this formula P[Ta < ∞] = 1
2
but E[Ta ] = ∞.
Exercise 34–3. Show that P[Ta < ∞] = 1 and E[Ta ] = ∞.
Exercise 34–4. What happens if a < 0?
Example 34–2. Again let a > 0 and fix t > 0. Then P[max0≤x≤t Xs ≥ a] = P[Ta ≤ t].
Exercise 34–5. Provide the details of this argument.

Example 34–3. Suppose 0 < t1 < t2 . What is the probability that Xt is 0 at least
once in the interval (t1 , t2 )? Let Z denote the event that Xt is zero for at least one
value of t in (t1 , t2).
P[Z] = E[P[Z | Xt1 ]]

∞
= P[T | x| ≤ t2 − t1 ] dFXt1 (x)
−∞
∞
1 −x2 /2t1
= P[T | x| ≤ t2 − t1 ] e dx.
−∞ √2π t1
Exercise 34–6. Simplify this expression to obtain the Arcsine Law.
Martingales can also be exploited to obtain some results about Brownian motion.
The verification that {Xt : t ≥ 0}, {Xt2 − t : t ≥ 0} and {exp{λ Xt − λ 2 t/ 2} : t ≥ 0}
are martingales is quite straightforward.
Exercise 34–7. Verify that these are martingales.
Example 34–4. Suppose a < 0 < b and let T = inf{s : Xs = a or Xs = b}. Applying
the Optional Stopping Theorem gives 0 = E[X0 ] = E[XT ] = a P[XT = a] + b P[XT =
b]. Solving gives P[XT = a] = b/ (b − a) and P[XT = b] = −a/ (b − a).
Exercise 34–8. Verify this computation. Why does the Optional Stopping Theorem
apply?
Exercise 34–9. Use the martingale {Xt2 − t : t ≥ 0} to compute E[T].
Example 34–5. A Brownian motion with drift is a stochastic process Bt of the

form Bt = Xt + µ t where X is a standard Brownian motion and µ is a nonzero
constant. Again a simple computation shows that {exp{λ Bt − µλ t − λ 2 t/ 2} : t ≥ 0}
is a martingale for each fixed value of λ . Choose for simplicity λ = −2µ so that
e−2µ Bt is a martingale. Let T = inf{s : Bs = a or Bs = b} where a < 0 < b are fixed.
The Optional Stopping Theorem gives P[BT = b] = (1 − e−2µ a )/ (e−2µ b − e−2µ a ). This
is comparable to the result for the gambler’s ruin process when the game is biased.
Exercise 34–10. Fill in the details of the computation.
Example 34–6. Brownian motion in several dimensions is defined in way which

imitates the definition in dimension 1. A d dimensional Brownian motion process
{Xt : t ≥ 0} is a process with stationary independent increments for which X0 = 0
and for each t > 0 the random vector Xt has the d variate normal distribution with
mean vector 0 and covariance matrix tΣ for some positive definite matrix Σ. The
standard d dimensional Brownian motion has Σ as the identity matrix. This implies
that the coordinates of a standard d dimensional Brownian motion are independent.
Example 34–7. In the multidimensional setting the distance of a standard Brownian
motion process Xt from the origin is often of interest. The process Yt = || Xt || is
called a radial Brownian motion or a Bessel process.
Exercise 34–11. Find the expected time required for a standard multidimensional
Brownian motion to leave the ball of radius a > 0 centered at the origin. Hint:
Show that || Xt || 2 − dt is a martingale.
Exercise 34–1. The number of paths from the origin to (a + b, a − b) which
don’t touch the axis is the number of counting paths in which A is always ahead.
The Ballot Theorem gives the number of these paths. The total number of paths
from the origin to (a + b, a − b) is the total number of ways the ballots could be
counted. This number is given by the Proposition.
Exercise 34–2. If t > s write Xt = Xt −Xs +Xs and use the independent increments
property to compute as follows: Cov(Xt , Xs ) = E[Xt Xs ] = E[(Xt −Xs )Xs ]+E[Xs2] =
0 + Var(Xs ) = σ 2 s. When s > t a similar argument shows that the covariance is
σ 2 t.
∞ 2
Exercise 34–3. Let t → ∞ to obtain P[Ta < ∞] = (2/ √2π ) 0 e−x / 2 dx = 1.
Differentiation gives the density of Ta as e−a / 2t / √2π t3 for t > 0, from which it
2
is easy to see that E[Ta ] = ∞.
d
Exercise 34–4. The Reflection Principle gives Ta = T | a| .
Exercise 34–5. If the maximum exceeds a then the time required to reach level
a must have been smaller than t.
Exercise 34–6. Substitute from the earlier formula for P[T | x| ≤ t2 − t1 ] and then
integrate by parts.
Exercise 34–7. Just write Xt = Xt − Xs + Xs and use the independent increments

property.
Exercise 34–9. 0 = E[XT2 − T] = a2 P[XT = a] + b2 P[XT = b] − E[T] so that

E[T] = −ab after substitution for the probabilities from the example above.
Exercise 34–11. Here || Xt || 2 − dt is a martingale so if Ta is the time at which

the Brownian motion hits the surface of the ball of radius a, E[Ta ] = a2 / d by the
Optional Stopping Theorem.
§35. Laboratory 11
One theory about the behavior of stock price Pt over time is that Pt should
behave like a stochastic process with independent ratios, that is, if t1 < . . . < tn
then Ptn / Ptn−1 , . . . , Pt2 / Pt1 should be independent. A simple model for this sort of
behaviour is given by a geometric Brownian motion process Pt = eBt where Bt is
a Brownian motion with drift. One use of this model of stock prices is to construct
a theoretical pricing model for stock options. A call option is a contract which
gives the owner of the contract the right to buy one share of the underlying stock at
a particular price K called the strike price of the option. The call option contract
expires after a fixed time T. Thus a call that is not used by time T becomes worthless.
1. If the drift and variance parameters of the underlying Brownian motion with
drift are 0.10 and 0.005 respectively, what is the probability that a stock with a price
of $20 today has a price exceeding $25 one year from now? What is the probability
that the stock price exceeds $25 at some time during the next year?
2. The economist Robert Merton has presented an economic argument that call
options will not be exercised before the expiration time T. This implies that the
value of a call option today depends only on the price of the underlying stock at
time T. Since the value of the call at time T is max{PT − K, 0}, the value of the call
today (at time 0) is just the actuarial present value of this amount. Thus the value
of a call today is E[vT max{PT − K, 0}]. This is the Black–Scholes option pricing
formula. Simplify this expression to obtain a form suitable for computation.
3. Explain how stock price data could be used in conjunction with the formula
of the previous question in order to value a call option on a particular stock. Choose
a stock for which call options are traded on the exchange and estimate the parameters
governing the stock price process. Compare the option price with that given by the
Black–Scholes formula. Most traders believe that the interest rate on 3 month U.S.
Treasury bills is the interest rate to use when computing the present value.

§36. Utility Functions
In this concluding section a fundamental question will be examined. Why would

an insurer assume a risk which the insured is unwilling to assume? To answer this
question, a basic understanding of how value is attached to quantities of money will
be required.
The concept of utility was invented to provide a mathematical methodology

for at least formally measuring the value of money (that is the utility of money).
Naively, one may think that a dollar is a dollar and that’s it. A moments reflection
shows that your desire (or your measure of worth) of a payment of, say, $10,000
might differ considerably from that of the person next to you. Consider also the
importance of the same $10,000 to Bill Gates.
These ideas will now be formalized. In this discussion quantities of money

will always be measured in dollars. Denote by u(w) the utility of w dollars. It
is convenient to make a strictly mathematical assumption. Practically speaking, a
utility function can only be defined on a certain discrete set (the multiples of $.01).
However for mathematical tractibility it is assumed that utility functions are defined
on the entire real line or an interval.
As already noted above, different people may well have different utility func-
tions. What features should utility functions have in common? One typical as-
sumption is that for any individual the utility function should be non–decreasing.
This property is the mathematical expression of the fact that having more money
is ‘better.’ Furthermore one would also expect that as one’s wealth increases the
utility of an additional dollar should decrease. This expresses the notion that to
someone having only $10 the prospect of gaining an additional dollar is greater than
the prospect of gaining an additional dollar if one already has $1,000,000.
Exercise 36–1. Argue that if the utility function is sufficiently smooth then u′ (w) ≥
0 and u′′ (w) ≤ 0. Thus a utility function will be concave (down).
A real valued function u(w) is said to be a risk averse utility function if u(w) is
non–decreasing and concave down.
The reason for the terminology risk averse will be explained shortly.
Example 36–1. It is straightforward to check that each of the functions −e−α w ,

log w, and α w + β is a risk averse utility function as long as α ≥ 0. A utility function
of the form −e−α w is called an exponential utility function.
Example 36–2. Suppose a person with a risk averse utility function has current
wealth w and is given a choice between two investment schemes. In the first
§36: Utility Functions 158
scheme the person will receive an amount b outright. In the second scheme the
person will receive a random amount W where E[W] = b. How does the person
decide between these two alternatives?
To answer this question, decision makers will be assumed to act according to the
Expected Utility Principle: a decision maker always chooses the available option
with the largest expected utility.
Example 36–3. In the previous example the expected utility of the first investment
scheme is E[u(w+b)] = u(w+b), while the expected utility of the second investment
scheme is E[u(w + W)].
Generally it is not easy to perform the actual computations required by the

expected utility principal. The following theorem can provide some helpful infor-
mation.
Jensen’s Inequality. If X is a random variable and f is a function which is concave
then
E[f (X)] ≤ f (E[X]).
proof : This proof is valid under the additional assumption that the function f is twice continuously
differentiable. For any fixed x and a, the Fundamental Theorem of Calculus and an
integration by parts gives
x
f (x) = f (a) + f ′ (t) dt
a
x
= f (a) + f ′ (a) (x − a) + (x − t) f ′′ (t) dt
a
≤ f (a) + f ′ (a) (x − a)
since f is concave. Substituting X for x and E[X] for a yields
f (X) ≤ f (E[X]) + f ′ (E[X]) (X − E[X]).
Taking expectations of both sides of this last inequality proves the theorem.
Exercise 36–2. Under what conditions does equality hold in Jensen’s inequality?
Continuing the example, from Jensen’s inequality and the fact that a risk averse
utility function is concave
E[u(w + W)] ≤ u(E[w + W]) = u(w + b)
so that such a person will always select the sure payoff b! Would such a person buy
a lottery ticket?
Having discussed the methodology by which a measure of value is assigned to

a given quantity of money, the question of why insurance exists can be addressed.
Consider a situation in which a person with utility function u and current fortune
w faces the prospect of a potential financial loss of amount A. For concreteness
suppose that the person is insuring a house against the possibility of a loss due to
fire. It is natural to view the amount of loss A as a random variable since the amount
of loss depends on the severity of the fire, the speed with which the fire department
responds, and other such factors which are unpredictable in nature. Suppose that
for a fixed non–random premium amount P an insurer will indemnify the insured
against such a loss. What premium would the person be willing to pay? If the
offer of insurance is accepted, the expected utility is u(w − P). If the person does
not buy insurance then E[u(w − A)] is the persons expected utility. Because of the
expected utility principle, the person would be willing to buy insurance whenever
the premium satisfies
u(w − P) ≥ E[u(w − A)].
This analysis applies to the person seeking insurance. Now analyze the position
of the insurer. Assume that the insurer has utility function ui and current wealth
wi . Reasoning as before shows that the insurance company will offer insurance at a
premium P if
ui (wi ) ≤ E[ui (wi + P − A)].
If there is a set of values P which satisfy simultaneously both of these inequalities

then there is the possibility of insurance being issued.
Exercise 36–3. What is the situation if both the person and the insurer have linear
utility functions?
Exercise 36–4. What happens if the insurer and the person have the same expo-
nential utility function −e−w/1000 , the persons wealth is $5,000, the insurers wealth is
$5,000,000 and the loss variable A has an exponential distribution with mean $500?
Exercise 36–5. What happens in the previous exercise if the utility function is
log w?
Jensen’s inequality can provide some interesting information about the condi-
tions under which a person will purchase insurance. Consider the largest premium
that an individual would pay for insurance. This premium P must satisfy
u(w − P) = E[u(w − A)].
Using Jensen’s inequality on the right member gives
u(w − P) ≤ u(w − E[A])
and since the utility function is non–decreasing, P ≥ E[A]. Thus a risk averse
decision maker would be willing to pay a premium greater than the pure premium
(expected loss) in order to obtain insurance. Similarly the premium charged by the
insurance company must also exceed the pure premium. These two facts reinforce
intuition and lend a certain credibility to the analysis and assumptions.
As a final note, consider a type of insurance policy that is typically issued.

A standard type of insurance policy is the stop loss policy in which the insurer
pays only the amount of the claim which exceeds a certain pre–arranged deductible
amount d. In such a policy the insurers payment when the loss is x is given by
(x − d) 1[d,∞) (x). There are many other conceivable types of payment policies, and
the following discussion shows why such policies do not generally occur. Suppose
there is another type of payout procedure which for a claim of amount x would pay
R(x). Naturally, 0 ≤ R(x) ≤ x, but even more is true.
Theorem. Suppose a person with risk averse utility function u is to be insured
against a loss of amount A. If there is a stop loss policy with expected payout
E[R(A)] which is offered for the same premium as the policy with payout R(A) then
the stop loss policy has greater expected utility than the policy with payout R(A).
proof : Let d be the deductible for the stop loss policy which exists by hypothesis, let the current
wealth of the person be w, and let the premium be P. Arguing as in the proof of Jensen’s
inequality gives
u(w+R(A) − A − P) − u(w + (A − d) 1[d,∞) (A) − A − P)

≤ u′ (w + (A − d) 1[d,∞) (A) − A − P) R(A) − (A − d) 1[d,∞) (A)

≤ u′ (w − d − P) R(A) − (A − d) 1[d,∞) (A)
where the fact that u′ is decreasing has been used. Taking expectations gives the result,
since the expectation of the last term is zero (why?).
Problems
Problem 36–1. Consider a game of chance in which a fair coin is tossed until a head
apprears. Let N denote the toss on which this occurs. Find the density, expectation,
and variance of N. Suppose a prize of X = 2N is paid when a head first appears.
What is E[X]? If your utility function is u(w) = log(w) what is the expected utility
of the prize?
Problem 36–2. In a typical state lotto game a person who buys a $1 ticket wins
$1,000,000 with probability 10−8 . Find a possible utility function for a person who
plays the lotto. Can this person have a risk averse utility function?
Problem 36–3. Suppose a person has a utility function which is increasing and
concave up. Give an example to explain why such a person might be called a risk
lover.
Problem 36–4. Show that if the amount of random loss X has a uniform distribution
on the interval (0, 1) and if the insured has utility function u(w) = log(w) then the
maximum amount the insured will pay as a premium is
ww
w− .
e(w − 1)w−1
Problem 36–5. Repeat the previous problem if the utility function is u(w) = −e−α w
and the loss random variable X has a chi square distribution with n degrees of
freedom.
Problem 36–6. Use the Taylor expansion of u(w − x) about x = µ to show that the
maximum premium an insured will pay is approximately
1 u′′ (w − µ ) 2
µ− σ .
2 u′ (w − µ )
Here µ and σ 2 are the mean and variance of the loss random variable.
Problem 36–7. A group medical insurance policy pays $D each time a member of
the group is hospitalized. The group consists of g distinct subgroups, which differ in
the rate of hospitalization. Suppose that the annual number of hospital admissions
for subgroup i has a Poisson distribution with parameter λi . Find an expression
for the expected claims payments in one year, and also find the distribution of the
number of admissions in one year.
Problem 36–8. Suppose the loss random variable X has an exponential distribution
with mean 10. Suppose a premium of $5 will be paid. Show that the propor-
tional insurance policy with benefit X/ 2 and the stop loss policy with benefit
(X − 10 log 2)1[10 log 2,∞) (X) are both feasible insurance policies. Which policy would
the insured choose, and why?
Problem 36–9. True or False: A person with an exponential utility function, −e−α w ,
considers wealth irrelevent when deciding the maximum premium to pay for com-
plete protection against a random loss.
Problem 36–10. An insurer with wealth w insures a loss X which has the following
probability distribution:
1
P[X = 0] = P[X = 16] = .
2
The insurer’s utility function is u(x) = log x. The insurer is willing to pay a maximum
of 6 to a reinsurer who accepts 50 percent of the loss. Find w.
Problem 36–11. An insurer and an individual have exponential utility functions

with parameters α and 2α respectively. The individual faces a random loss that is
normally distributed with mean 100. The insurer assumes that the variance is σ 2 .
The individual assumes that the variance is 25. Determine the largest value of σ 2
for which the insurer can charge a premium acceptable to the individual.
Problem 36–12. A decision maker has utility function u(x) = −e−3x and initial
wealth w. The decision maker faces two random losses. The loss X has a normal
distribution with mean α and variance 4. The loss Y has a normal distribution with
mean 10 and variance 8. Determine the maximum value of α for which the decision
maker prefers X to Y.
Problem 36–13. Three insurers have identical utility functions u(w) = log w, w > 0,
and a wealth of 36, 25, and 16 respectively. All three companies insure the same
risk. In the event of a loss each insurer will pay 11. The probability of loss is
0.5. Another company offers to reinsure each company’s complete risk at the same
premium π . Each company is willing to accept the reinsurance at the premium π
if its expected utility is maximized. Determine π such that the reinsurer maximizes
its total expected profit.
Problem 36–1. Here P[N = n] = 2−n for n = 1, 2, . . .. Hence E[X] = +∞, and
∞

E[log(X)] = n log(2)2−n ≈ 1.386.
n=1
Problem 36–3. See the previous problem.
Problem 36–7. Denote by Ni the number of hospitalizations for group i. Then

g

g
the amount of claims payments in one year is D Ni . Also Ni has a Poisson
i=1 i=1

g
distribution with parameter λi .
i=1
Problem 36–9. True. The variable w cancels from both sides of the equation.
Exercise 36–1. Since more wealth is better, u(w) is increasing. Thus u′ (w) ≥ 0.
Since the utility of an additional dollar decreases as wealth increases, u(w + 1) −
u(w) ≈ u′ (w) is decreasing. Thus u′′ (w) = (u′ (w))′ ≤ 0.
Exercise 36–2. Equality holds if f ′′ (x) = 0 for all x, which means f (x) is a
linear function of x.
Exercise 36–3. The person requires P ≤ E[A] while the insurer requires
P ≥ E[A] so that the only possible premium is P = E[A].
Exercise 36–4. In this case the wealth of each party is irrelevant and the only
common premium is P = 1000 ln E[eA/ 1000 ] = 2000.
Exercise 36–5. The person requires the premium to satisfy ln(5000 − P) ≥

E[ln(5000 − A)] while the insurer requires ln(5000000) ≤ E[ln(5000000 +P− A)].
To compute the expectations the utility function u(w) is taken to have the value
0 when w ≤ 0. Then the requirement for the person is P ≤ 536.46 while for
the company, P ≥ 500.03 by numerical computation using a computer algebra
package.
§37. Life Table at 5% Interest
Age lx dx äx 1000 Ax 1000 2Ax 1000 Axx
0 100000 1260 19.922 51.329 17.924 82.836
1 98740 92 20.122 41.823 7.253 63.528
2 98648 64 20.097 43.022 7.071 64.962
3 98584 49 20.064 44.553 7.152 67.000
4 98535 40 20.028 46.307 7.392 69.426
5 98495 36 19.987 48.236 7.747 72.144
6 98459 33 19.944 50.301 8.178 75.075
7 98426 30 19.898 52.498 8.684 78.211
8 98396 26 19.848 54.835 9.272 81.562
9 98370 23 19.796 57.327 9.961 85.156
10 98347 19 19.741 59.974 10.751 88.988
11 98328 19 19.681 62.792 11.662 93.087
12 98309 24 19.619 65.751 12.666 97.393
13 98285 37 19.555 68.811 13.724 101.824
14 98248 52 19.490 71.902 14.759 106.242
15 98196 67 19.425 75.008 15.751 110.613
16 98129 82 19.359 78.129 16.695 114.936
17 98047 94 19.293 81.268 17.585 119.212
18 97953 102 19.226 84.453 18.447 123.493
19 97851 110 19.158 87.726 19.316 127.852
20 97741 118 19.087 91.091 20.195 132.295
21 97623 124 19.014 94.552 21.083 136.826
22 97499 129 18.939 98.134 22.002 141.488
23 97370 130 18.861 101.852 22.964 146.305
24 97240 130 18.779 105.751 24.015 151.356
25 97110 128 18.693 109.849 25.173 156.670
26 96982 126 18.602 114.174 26.470 162.297
27 96856 126 18.507 118.737 27.921 168.252
28 96730 126 18.406 123.534 29.520 174.518
29 96604 127 18.300 128.576 31.284 181.112
30 96477 127 18.189 133.866 33.220 188.034
31 96350 130 18.072 139.426 35.355 195.319
32 96220 132 17.950 145.244 37.680 202.935
33 96088 137 17.822 151.342 40.226 210.918
34 95951 143 17.688 157.708 42.984 219.240
35 95808 153 17.549 164.348 45.968 227.902
36 95655 163 17.404 171.242 49.162 236.861
37 95492 175 17.254 178.404 52.586 246.138
38 95317 188 17.098 185.832 56.247 255.719
39 95129 203 16.936 193.533 60.159 265.611
40 94926 220 16.768 201.506 64.328 275.804

§37: Life Table at 5% Interest 166
41 94706 241 16.595 209.750 68.764 286.290
42 94465 264 16.417 218.248 73.454 297.031
43 94201 288 16.233 227.000 78.408 308.020
44 93913 314 16.044 236.014 83.643 319.265
45 93599 343 15.849 245.292 89.171 330.761
46 93256 374 15.649 254.826 94.994 342.489
47 92882 410 15.443 264.618 101.127 354.445
48 92472 451 15.232 274.647 107.553 366.588
49 92021 495 15.017 284.891 114.257 378.874
50 91526 540 14.798 295.346 121.241 391.286
51 90986 584 14.574 306.018 128.527 403.836
52 90402 631 14.344 316.935 136.156 416.563
53 89771 684 14.110 328.092 144.138 429.454
54 89087 739 13.871 339.464 152.455 442.463
55 88348 797 13.628 351.054 161.123 455.591
56 87551 856 13.380 362.858 170.152 468.830
57 86695 919 13.127 374.890 179.571 482.196
58 85776 987 12.870 387.137 189.384 495.670
59 84789 1063 12.609 399.586 199.586 509.225
60 83726 1145 12.344 412.195 210.141 522.795
61 82581 1233 12.076 424.941 221.027 536.340
62 81348 1324 11.806 437.794 232.219 549.814
63 80024 1415 11.534 450.744 243.712 563.202
64 78609 1502 11.260 463.800 255.529 576.519
65 77107 1587 10.983 476.997 267.729 589.820
66 75520 1674 10.702 490.357 280.360 603.143
67 73846 1764 10.419 503.878 293.435 616.486
68 72082 1864 10.132 517.547 306.957 629.837
69 70218 1970 9.843 531.305 320.858 643.109
70 68248 2083 9.553 545.108 335.091 656.247
71 66165 2193 9.263 558.900 349.587 669.173
72 63972 2299 8.974 572.682 364.352 681.895
73 61673 2394 8.684 586.454 379.395 694.420
74 59279 2480 8.395 600.260 394.790 706.822
75 56799 2560 8.103 614.130 410.598 719.156
76 54239 2640 7.810 628.073 426.851 731.452
77 51599 2721 7.517 642.054 443.518 743.680
78 48878 2807 7.224 656.017 460.530 755.786
79 46071 2891 6.933 669.858 477.742 767.655
80 43180 2972 6.647 683.490 495.023 779.196
§37: Life Table at 5% Interest 167
81 40208 3036 6.367 696.796 512.188 790.281
82 37172 3077 6.096 709.717 529.133 800.856
83 34095 3083 5.834 722.208 545.770 810.886
84 31012 3052 5.580 734.292 562.116 820.423
85 27960 2999 5.334 746.010 578.225 829.545
86 24961 2923 5.097 757.276 593.938 838.167
87 22038 2803 4.873 767.969 609.034 846.154
88 19235 2637 4.659 778.150 623.584 853.585
89 16598 2444 4.452 787.993 637.853 860.684
90 14154 2246 4.251 797.588 651.990 867.598
91 11908 2045 4.057 806.812 665.785 874.229
92 9863 1831 3.875 815.461 678.882 880.384
93 8032 1608 3.707 823.460 691.127 886.009
94 6424 1381 3.554 830.749 702.385 891.054
95 5043 1159 3.416 837.313 712.595 895.509
96 3884 945 3.294 843.126 721.669 899.329
97 2939 754 3.184 848.396 729.931 902.714
98 2185 587 3.084 853.138 737.372 905.648
99 1598 448 2.992 857.517 744.245 908.260
100 1150 335 2.907 861.589 750.614 910.548
101 815 245 2.825 865.483 756.669 912.543
102 570 177 2.740 869.539 762.975 914.484
103 393 126 2.649 873.840 769.651 916.294
104 267 88 2.549 878.615 777.064 917.908
105 179 60 2.426 884.468 786.271 919.465
106 119 41 2.253 892.737 799.737 921.800
107 78 27 2.006 904.455 819.532 925.261
108 51 18 1.616 923.036 852.465 933.393
109 33 33 1.000 952.381 907.029 952.381
§38. Service Table
x l(xτ ) dx(d) dx(w) dx(i) dx(r)
30 100000 100 19990
31 79910 80 14376
32 65454 72 9858
33 55524 61 5702
34 49761 60 3971
35 45730 64 2693 46
36 42927 64 1927 43
37 40893 65 1431 45
38 39352 71 1181 47
39 38053 72 989 49
40 36943 78 813 52
41 36000 83 720 54
42 35143 91 633 56
43 34363 96 550 58
44 33659 104 505 61
45 32989 112 462 66
46 32349 123 421 71
47 31734 133 413 79
48 31109 143 373 87
49 30506 156 336 95
50 29919 168 299 102
51 29350 182 293 112
52 28763 198 259 121
53 28185 209 251 132
54 27593 226 218 143
55 27006 240 213 157
56 26396 259 182 169
57 25786 276 178 183
58 25149 297 148 199
59 24505 316 120 213
60 23856 313 3552
61 19991 298 1587
62 18106 284 2692
63 15130 271 1350
64 13509 257 2006
65 11246 204 4448
66 6594 147 1302
67 5145 119 1522
68 3504 83 1381
69 2040 49 1004
70 987 17 970

Actuarial Mathematics

Enviado por

Dados do documento

Descrição original:

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Actuarial Mathematics

Enviado por

Direitos autorais:

Formatos disponíveis

Lecture Notes

Jerry Alan Veeh

Copyright  2003 Jerry Alan Veeh. All rights reserved.

A calculator, such as the one allowed on the Society of Actuaries examinations,

Copyright  2003 Jerry Alan Veeh. All rights reserved.

The theory of insurance can be viewed as the theory of contingent payments.

Copyright  2003 Jerry Alan Veeh. All rights reserved.

To begin, remember the way in which compound interest works. Suppose an

Exercise 2–3. Compute 0.05(12) .

this corresponds to the notion of instantaneous compounding of interest. In this

Exercise 2–4. Show that δ = log(1 + i).

Exercise 2–5. Find the force of interest which is equivalent to 5% compounded

Two Equivalent Cash Flows

2004 2014 2004 2014

Equivalence of Interest and Discount

The relationship is d = iv follows by moving the interest payment back in time

Problem 2–1. Show that if i > 0 then

Problem 2–2. Show that limm→∞ d (m) = limm→∞ i(m) = δ .

Problem 2–5. Show that d = 1 − v. Is there a similar equation involving d(m) ?

Problem 2–5. 1 − d (m) / m = v1/ m .

Problem 2–6. d (m) / m = v1/ m i(m) / m.

Exercise 2–3. 0.05(12) = 12((1 + 0.05)1/ 12 − 1) = 0.04888.

Exercise 2–4. The requirement for equivalence is that eδ = 1 + i.

Exercise 2–5. Here eδ = (1 + 0.05/ 365)365, so that δ = 0.4999. So as a rough

Exercise 2–7. A payment of d (m) / m made at time 0 is required to be equivalent

Many different types of financial transactions involve the payment of a fixed

$400 $400 ... ... $400 $400

Exercise 3–1. Evaluate the sum.

Exercise 3–2. Show that an = v än .

It is now a simple matter to find σ .

A continuous annuity of the type above which is payable at rate σ = 1 for n

Problem 3–3. Suppose a loan of amount A is amortized by a series of n payments.

Problem 3–8. Humphrey purchases a home with a $100,000 mortgage. Mortgage

Problem 3–9. On January 1, an insurance company has $100,000 which is due to

Problem 3–3. bk+1 = bk + ibk − P.

Problem 3–4. The assertion of the prospective method is that bk = Pan−k .

Exercise 3–3. The force of interest that is equivalent to an interest rate of 5%

Exercise 3–4. This formula follows by straightforward integration.

Exercise 3–7. Where did all those payments go?

1. A loan of 10,000 carries an interest rate of 9% compounded quarterly. Equal

3. A loan of 10,000 is to be repaid with equal monthly payments of p. The

4. A loan of 10,000 is to be repaid as follows. Payments of p are to be made at

5. The symbol a(m)

6. The symbol ä(m)

7. An increasing annuity immediate with a term of n periods pays 1 at the end

8. A decreasing annuity immediate with a term of n periods pays n at the end

Copyright  2003 Jerry Alan Veeh. All rights reserved.

The underlying object in probability theory is a sample space S, which is simply

Exercise 5–1. Show that if A ⊂ B are events, then P[A] ≤ P[B].

Another of the basic concepts is that of a random variable. A random variable

If X is a discrete random variable, the density of X, denoted fX (t) is defined by

Example 5–2. An exponentially distributed random variable Y with parameter

Exercise 5–3. Sketch the distribution function of an exponential random variable

Exercise 5–4. A random variable X is uniformly distributed on an interval (a, b) if

Exercise 5–5. Draw a picture of a distribution function of a random variable which

Exercise 5–6. Graph the function 1[0,1) (t).

Computation of conditional probabilities will play an important role. If A and

Similar definitions can be given in the case of random variables. Intuitively,

Problem 5–2. The moment generating function of a random variable X, denoted

Problem 5–9. Show that for a random

Problem 5–10. Suppose Y is a non-negative, absolutely continuous random vari-

Problem 5–2. MX′ (0) = E[X].

Thus the coefficient of tk in the Maclaurin expansion of MX (t) is E[X k ]/ k!.

Problem 5–10. Suppose Y is a non-negative, absolutely continuous random vari-

5. For each value of d find the conditional expectation of H given D = d.