Chapter 7 — Advanced Discrete (Frequency) Distributions

Lecture Notes for Loss Models (Klugman, Panjer & Willmot)

Written for a student with no prior math/stats background

Table of Contents
How to Use These Notes......................................................................................................................................... 2
0. Before We Start: A Refresher of the Tools You’ll Need........................................................................2
0.1 What is a “counting random variable”?...............................................................................................2
0.2 What is a probability generating function (pgf), and why do we bother?...........................2
0.3 The a , b , 0 and a , b , 1 classes (from Chapter 6)...............................................................................3
0.4 What is a convolution?................................................................................................................................ 3
0.5 Moments in plain English.......................................................................................................................... 3
7.1 Compound Frequency Distributions......................................................................................................... 4
7.1.1 The Big Idea................................................................................................................................................. 4
7.1.2 The Compounding Formula.................................................................................................................. 4
7.1.3 How to Actually Compute the Probabilities: Panjer’s Recursion.........................................5
7.1.4 Section 7.1 Exercises — Worked........................................................................................................8
7.2 Further Properties of the Compound Poisson Class...........................................................................9
7.2.1 One Formula, Many Named Distributions......................................................................................9
7.2.2 Moments and Skewness — Choosing Between Models with the Same
Mean/Variance.................................................................................................................................................... 10
7.2.3 Compound Poisson Is Closed Under Convolution (Theorem 7.5).....................................11
7.2.4 Section 7.2 Exercises — Worked.....................................................................................................13
7.3 Mixed-Frequency Distributions................................................................................................................14
7.3.1 The General Idea: “Randomizing the Parameter Itself”.........................................................14
7.3.2 Mixed Poisson Distributions — The Most Important Special Case..................................15
7.3.3 Section 7.3 Exercises — Worked.....................................................................................................18
7.4 The Effect of Exposure on Frequency.....................................................................................................20
7.5 An Inventory of Discrete Distributions..................................................................................................21
7.5.1 Section 7.5 Exercises — Worked.....................................................................................................22
Chapter 7 — Master Summary Table.............................................................................................................24
How to Use These Notes
This chapter is genuinely hard the first time you see it, because it stacks three ideas on top
of each other: compounding, mixing, and recursion. Every single idea in this chapter is
really just answering one plain-English question:
“The number of claims a policy produces isn’t just one random number — it depends
on other random things too. How do we build a formula for it, and how do we
compute it fast?”
Read the notes in order. Each new idea is introduced with (1) a plain-English story, (2) the
formal math, (3) a fully worked numerical example, and (4) the matching textbook
example/exercise so you can cross-check against the book.

0. Before We Start: A Refresher of the Tools You’ll Need

Chapter 7 assumes you already know four things from Chapters 5–6. If any of these feel
shaky, read this section slowly — everything after depends on it.

0.1 What is a “counting random variable”?

In insurance, we’re constantly asking: “How many claims will this policy (or this whole
portfolio) produce in a year?” The answer is a whole number: 0, 1, 2, 3, …. A random
variable that only takes values 0, 1, 2, 3, … is called a counting random variable, and its
distribution is called a frequency distribution. We usually call it N (for “number of
claims”).
The probability function (pf) is just the list of probabilities:
pk =Pr ( N=k ) , k=0 , 1 , 2, …

0.2 What is a probability generating function (pgf), and why do we bother?

A pgf is nothing more than a clever bookkeeping trick: instead of writing down the whole
infinite list p0 , p 1 , p2 , …, we glue them together into one algebraic expression using a
dummy variable z :
∞
P N ( z ) =E ( z N )=∑ p k z k = p0 + p1 z+ p 2 z 2 +⋯
k=0

Think of z as a placeholder with no meaning of its own — it’s just there so that the
coefficient sitting next to z k tells you pk . The reason pgfs are worth the trouble:

• Sums of independent random variables become products of pgfs. If S= X 1 + X 2

with X 1 , X 2 independent, then PS ( z )=P X ( z ) P X ( z ). Adding random variables is hard
1 2
 directly (you’d need convolutions — see 0.4) but multiplying their pgfs is easy
algebra.
• A “random number of random variables” (a random sum) becomes function
composition of pgfs — this is literally the entire subject of Section 7.1.
• Derivatives of P ( z ) at z=1 give you moments (mean, variance, etc.) without
integrating or summing directly.
Rule of thumb: whenever you see P ( z ), mentally read it as “the whole distribution,
packaged as one formula.”

0.3 The ( a , b , 0 ) and ( a , b , 1 ) classes (from Chapter 6)

Many standard frequency distributions (Poisson, binomial, negative binomial, geometric)
share one beautiful recursive property. If
pk b
=a+ , k=1 , 2, 3 , …
p k−1 k

for constants a and b that don’t depend on k , the distribution is said to belong to the ( a , b , 0 )
class. You only need p0 and the constants a , b to generate every other probability — hugely
convenient for computing.
The ( a , b , 1 ) class is the same recursion but only required to hold from k =2 onward, which
lets p1 be adjusted freely relative to p0. This is how zero-truncated (no zero claims
allowed, pT0 =0) and zero-modified (the probability at zero is arbitrarily reset, e.g. to match
real data with “too many” or “too few” zeros) distributions are built. Superscripts T and M
denote zero-truncated and zero-modified versions throughout this chapter.

0.4 What is a convolution?

If X and Y are independent counting random variables, the pf of their sum X +Y at value k
is
k
Pr ( X +Y =k )=∑ Pr ( X= j ) Pr ( Y =k− j ) .
j=0

This “slide-and-multiply-and-add” operation is called a convolution. The notation f ¿n k

means: “take the distribution with pf f k , add n independent copies of it together, and read off
the probability of getting total k .” Convolutions are correct but computationally painful for
large n — this is precisely the pain that Chapter 7’s recursive formulas are designed to
avoid.

0.5 Moments in plain English

• Mean μ=E ( N ): the long-run average number of claims.
• Variance σ 2=Var ( N ): how spread out / unpredictable the claim count is around the
mean.
 • Third central moment μ3: measures skewness — whether the distribution has a
long right tail (occasional very large claim counts). Insurance claim counts are
almost always right-skewed: mostly 0 or 1 claim, but occasionally many more.
With these five tools (counting r.v.s, pgfs, ( a , b , 0 ) / ( a , b ,1 ) , convolution, moments) you have
everything you need for Chapter 7.

7.1 Compound Frequency Distributions

7.1.1 The Big Idea
Imagine a car insurance portfolio. Over a year:
• Some number of accidents happen. Call this N .
• Each accident can produce more than one claim (e.g., a multi-car pileup, or a single
accident that triggers both a collision claim and a liability claim). Call the number of
claims from the j -th accident M j .
The total number of claims is
S= M 1 + M 2 +⋯+ M N .

Notice something strange and important: the number of terms in the sum is itself random ( N
is random). This is called a random sum, and S is said to have a compound distribution.
N is the primary distribution and each M j is drawn from the secondary distribution.
We assume:
1. The M j ’s are i.i.d. (independent and identically distributed) — they all come from
the same secondary distribution and don’t influence each other.
2. The M j ’s don’t depend on N — how many claims one accident produces doesn’t tell
you anything about how many accidents there will be.
3. If N=0 (no accidents), then S=0 automatically.

7.1.2 The Compounding Formula

The pgf of S is remarkably simple — it’s just the primary pgf with the secondary pgf
plugged in for z :
PS ( z )=P N [ P M ( z ) ] .

Why does this work? (Plain-English derivation) Condition on the value of N . If N=n
n
exactly, then S is the sum of n independent copies of M , so its pgf is [ P M ( z ) ] (remember: pgf
of an independent sum = product of pgfs). Now average over all possible values of n,
weighting by how likely each value of N is:
 ∞
PS ( z )=∑ Pr ( N=n ) [ P M ( z ) ] =PN [ P M ( z ) ] ,
n

n=0

because the last sum is exactly the definition of P N evaluated at the point P M ( z ) instead of at
z . That’s the whole trick: plug one pgf into another.
You do not need the accident-and-injuries story to justify using a compound distribution —
if a compound model fits your claims data well, that alone is good enough reason to use it.

Worked Example 7.1 — Every zero-modified distribution is secretly a compound distribution

Claim: Take any secondary distribution M . Let the primary distribution N be Bernoulli (i.e.,
N is 1 with probability q and 0 with probability 1−q; think of it as “did anything happen at
all?”). Its pgf is P N ( z ) =1−q +qz .
Plug in:
PS ( z )=P N [ P M ( z ) ]=1−q+ q P M ( z ) .

Compare this to the formula for a zero-modified (ZM) distribution (from Chapter 6): a ZM
distribution takes an ordinary distribution and resets the probability at zero to some
chosen value p0M , rescaling everything else to keep total probability 1. Matching up the
algebra, the q above equals
M
1− p0
q= ,
1− p0

where p0 is the original probability of zero claims from M . In plain words: “zero-
modifying” a distribution is the same thing as compounding a Bernoulli primary
distribution with that distribution as the secondary. This tells you ZM distributions aren’t a
separate, weird construction — they’re just the simplest possible compound distribution.

Worked Example 7.2 — Poisson primary + Poisson secondary = Neyman Type A

λ 1 ( z −1 ) λ2 ( z−1 )
Let both N and M be Poisson: P N ( z ) =e and P M ( z )=e . Then
λ1 [ e −1]
λ2 ( z −1 )

PS ( z )=e .

This is called the Neyman Type A (or Poisson–Poisson) distribution. Story: N = number
of insect egg clusters (or number of accidents); M = number of larvae per cluster (or claims
per accident). When λ 2 ≫ λ1 (secondary mean much bigger than primary mean), the
resulting distribution can even have two humps (bimodal) — e.g. a big group of zeros
from “no cluster occurred” and a second hump around λ 1 λ 2 from “a cluster occurred and
produced its typical count.”
7.1.3 How to Actually Compute the Probabilities: Panjer’s Recursion
Formula (7.1) is elegant but if you tried to get actual numbers gk =Pr ( S=k ) by literally
expanding the composed pgf, you’d be doing infinite convolutions — extremely slow. The
genius of this section is a recursive shortcut that works whenever the primary
distribution belongs to the ( a , b , 0 ) or ( a , b , 1 ) class.

Write gk =Pr ( S=k ), pk =Pr ( N=k ), f k =Pr ( M =k ) .

Theorem 7.1 (primary distribution in the ( a , b , 0 ) class):

( )
k
1 bj
gk = ∑
1−a f 0 j =1
a+ f g , k =1 ,2 , 3 , …
k j k− j

with starting value g0=P N ( f 0 ) (Theorem 7.3 below).

Theorem 7.2 (primary distribution in the ( a , b , 1 ) class):

k

[ p1− ( a+b ) p0 ] f k + ∑ ( a+bj /k ) f j gk− j

j =1
gk = , k=1, 2 , 3 ,…
1−a f 0

Where does this come from? (Intuition, not the full proof) The ( a , b , 0 ) / ( a , b ,1 )
recursion pn= ( a+b /n ) p n−1 is a statement purely about the primary distribution. If you
n−1
multiply both sides by [ P M ( z ) ] P M ' ( z ) and sum over n , the left side becomes (after some
calculus) the derivative of the composed pgf PS ( z ), and matching up coefficients of z k−1 on
both sides of the resulting differential equation gives (7.5)/(7.6). The full algebraic proof is
in the textbook — what matters for you is that the primary distribution’s recursive
structure transfers, term by term, into a recursion for the compound distribution’s
probabilities, weighted by the secondary probabilities f j .
Theorem 7.3 (the starting value, always true, for any compound distribution):
g0=Pr ( S=0 )=P N ( f 0 ) ,

where f 0=Pr ( M =0 ). In words: S=0 can only happen if either no “events” occur ( N=0), or
events occur but every single one of them produces zero secondary claims. Plugging z=0
into P M ( z ) literally gives you f 0, and plugging that into P N gives the probability that a
weighted-random number of “all-zero” secondary draws still nets to zero.

Worked Example 7.3 — Poisson primary (the single most useful case)
If N ∼ Poisson ( λ ), then a=0 , b=λ in the ( a , b , 0 ) notation, and (7.5) simplifies beautifully to
k
λ
gk = ∑
k j=1
j f j g k− j , g0=e ( ) .
−λ 1−f 0
Any compound distribution with a Poisson primary distribution is called compound
Poisson, and specifically “Poisson–X” where X names the secondary distribution (e.g.,
Poisson–binomial, Poisson–ETNB, etc.). This is the workhorse formula of the whole chapter
— memorize it.

Fully Worked Numerical Example (= textbook Example 7.4)

Setup: Poisson primary with λ=3. Secondary is ETNB (extended truncated negative
binomial) with r =−0.5 , β=1, giving secondary probabilities (computed in Chapter 6):
f 0=0 , f 1=0.853553 , f 2=0.106694 , f 3 =0.026674 .

Step 1 — starting value:

−3 ( 1−0) −3
g0=e =e =0.049787 .

Step 2 — recursion, using (7.7) with λ=3:

k
3j
gk =∑ f g .
j =1 k j k− j

3 ( 1)
g1 = ( 0.853553 ) ( 0.049787 )=0.127488
1
3 ( 1) 3 ( 2)
g2 = ( 0.853553 ) ( 0.127488 ) + ( 0.106694 ) ( 0.049787 )=0.179163
2 2
3 ( 1) 3 ( 2) 3 (3 )
g3 = ( 0.853553 )( 0.179163 )+ ( 0.106694 ) ( 0.127488 ) + ( 0.026674 )( 0.049787 )=0.184114
3 3 3
Reading the pattern: at each step k , you loop j from 1 to k , multiply “how many secondary
events contribute” (3 j/k ) times the probability of a secondary jump of size j ( f j ) times the
already computed probability of hitting k − j with everything before ( gk− j), and add these
up. This is exactly how you’d program it in a spreadsheet or in R/Python: build g0, then
compute g1 , g 2 , g3 , … in order, each one only needing previously computed values.

Theorem 7.4 — Adding/removing the secondary distribution’s zero doesn’t create anything
new
This is a subtle but very practical result. Split any secondary pgf into “chance of secondary
zero” plus “chance of a genuine positive value”:
T
P M ( z )=f 0 + ( 1−f 0 ) P M ( z ) ,

where PTM ( z ) is the zero-truncated version of M (i.e., M conditioned on being positive).

Theorem 7.4 says: if the primary pgf has the special form P N ( z ; θ ) =B [ θ ( z−1 ) ] — meaning
the parameter θ only ever appears multiplied against ( z−1 ) — then compounding with M
gives exactly the same distribution as compounding with the zero-truncated version M T , but
with the primary parameter rescaled to θ ( 1−f 0 ) .

Why this matters in practice: it means that whether your secondary distribution allows
zeros or not (Poisson secondary vs. zero-truncated Poisson secondary vs. zero-modified
Poisson secondary) makes no real difference — you always land on the same family
(e.g. Neyman Type A), just with a different value of the primary parameter. So don’t waste
time trying dozens of secondary zero-inflation variants hoping for a genuinely new shape
— you’ll just be re-parameterizing the same distribution.

Worked Example 7.6 — Same numbers, different “secondary zero” story

Poisson primary λ=7.5; secondary is zero-modified ETNB with p0M =0.6 , r=−0.5 , β=1,
giving f 0=0.6 , f 1=0.341421 , f 2=0.042678 , f 3=0.010670 . Running through exactly the same
recursion gives g0=0.049787 , g 1=0.127487 , g2=0.179161 , g3=0.184112 — the same
numbers (up to rounding) as Example 7.4! This is Theorem 7.4 in action:
λ × ( 1−f 0 )=7.5 × 0.4=3, recovering exactly the λ=3 case from before.

7.1.4 Section 7.1 Exercises — Worked

Exercise 7.1. Do all members of the ( a , b , 0 ) class satisfy Theorem 7.4’s condition
P N ( z ; θ ) =B [ θ ( z−1 ) ] ? For those that do, identify θ .

• Poisson( λ ): P ( z )=e λ ( z−1). Here θ=λ and B ( x ) =e x . Satisfies it.

−1
• Geometric( β ) (special case of negative binomial with r =1): P ( z )=[ 1−β ( z−1 ) ] .
Here θ=β , B ( x ) =( 1−x )−1.
Negative binomial( r , β ): P ( z )=[ 1−β ( z−1 ) ] . Again write it as B [ β ( z−1 ) ] with
−r
•
−r
B ( x ) =( 1−x ) — here θ=β (the extra parameter r just rides along inside B, which
the theorem explicitly allows).
m
• Binomial( m , q ) : P ( z )=[ 1+q ( z−1 ) ] . You can write this as θ=q , B ( x ) =( 1+ x )m. It also
satisfies the condition algebraically; but note the binomial is the one ( a , b , 0 )
member singled out elsewhere in the chapter as not infinitely divisible (Section
7.3.2) — that is a different property (about raising the whole pgf to a fractional
power), not about Theorem 7.4. Theorem 7.4 only needs the parameter to enter as
θ ( z−1 ) , which the binomial’s q does.

Punchline: every member of the ( a , b , 0 ) class has one parameter that enters exactly as
θ ( z−1 ) (it’s the parameter that controls “average rate/size”), so Theorem 7.4 always
applies — modifying the secondary distribution’s zero probability is always absorbable
into that one primary parameter.
Exercise 7.2. Show that (1) geometric–geometric, (2) Bernoulli–geometric, and (3) zero-
modified geometric are all the same distribution.
Solution sketch: Geometric is negative binomial with r =1, so its pgf is
−1
P ( z )=[ 1−β ( z−1 ) ] =B [ β ( z−1 ) ] with B ( x ) =( 1−x ) — exactly the Theorem-7.4 form with
−1

θ=β . By Theorem 7.4, compounding a geometric primary with a geometric secondary

(compound (1)) gives the same distribution as compounding a geometric primary with a
zero-truncated geometric secondary but with rescaled θ . Meanwhile, by Example 7.1,
compounding any secondary with a Bernoulli primary (compound (2)) produces the zero-
modified version of that secondary (compound (3)) directly, by definition of “zero-
modified.” The algebra (matching coefficients) confirms that (1), a geometric–geometric
compound, reduces exactly to a zero-modified geometric with a specific relationship
between the geometric–geometric parameters and the ZM parameters — which is exactly
(2), a Bernoulli primary compounded with a (plain) geometric secondary. All three
descriptions land on the identical two-parameter family (a zero-modified geometric
distribution); they are just three different “stories” for building the same object.
Exercise 7.3. Show binomial–geometric and negative-binomial–geometric (with integer r )
are identical.
Solution sketch: A negative binomial with integer r can itself be built as the sum of r
independent geometric random variables (a standard fact — it’s literally “r independent
geometrics added together,” the discrete analogue of how a Gamma with integer shape is a
sum of Exponentials). So “negative-binomial–geometric” (NB primary, geometric
secondary) is the same as compounding r independent copies of “geometric–geometric,”
each of which, by Exercise 7.2, is a zero-modified geometric. On the other side, using
Theorem 7.5 (Section 7.2, closure of compound Poisson under convolution) you can show
the sum collapses to exactly the same binomial–geometric pgf, because a binomial’s pgf
raised to compounding with a geometric secondary produces the identical algebraic form
as the r -fold NB compound when the parameters are matched up one-to-one. The key
takeaway (rather than grinding every algebraic step) is structural: both routes describe “a
sum of a random, geometrically-ish number of geometric jumps,” and there is essentially
only one distribution family that arises from combining these two building blocks.

7.2 Further Properties of the Compound Poisson Class

The compound Poisson family (Poisson primary + any secondary distribution) is the single
most important sub-family in this chapter, because it shows up everywhere: as a physically
motivated claim model, as a way to unify several named distributions, and as a
computational shortcut for adding independent risks together.

7.2.1 One Formula, Many Named Distributions

Write the compound Poisson pgf as
P ( z )=exp { λ [ Q ( z )−1 ] },

where Q ( z ) is the secondary distribution’s pgf.

Worked Example 7.7 — Poisson–ETNB unifies negative binomial, Polya–Aeppli, and Neyman
Type A
Take the secondary distribution to be ETNB (extended truncated negative binomial) with
parameters r , β . After algebra (moving the “−1” inside and simplifying), the log-pgf of the
Poisson–ETNB distribution comes out to
−r λ
ln P ( z )=μ {[ 1−β ( z−1 ) ] −1 }, μ= −r
.
1−( 1+ β )
This is again compound Poisson form, but now with mean parameter μ and secondary pgf
−r
[ 1−β ( z−1 ) ] — which is just a negative binomial pgf! So: Poisson–ETNB,
reparametrized, is “Poisson mixed with a negative-binomial-shaped secondary.” As
you vary the single parameter r over (−1 , ∞ ) ¿ 0 }¿, you sweep through an entire family of
named distributions:

As r → Secondary distribution becomes Named result

r →0 Logarithmic Negative binomial
r =1 Geometric Polya–Aeppli
r →∞ Poisson Neyman Type A

This is a wonderful “one-parameter dial” for controlling the shape/tail of your frequency
model without switching to a completely different formula each time.

7.2.2 Moments and Skewness — Choosing Between Models with the Same
Mean/Variance
If a secondary distribution has raw moments m1 ' ,m2 ' , m3 ' , the compound Poisson
distribution’s own moments are simply the Poisson mean λ times the corresponding
secondary raw moment:
2
μ= λ m1 ' , σ =λ m2 ' , μ3= λ m3 ' .

In plain English: if you know the mean, variance, and skewness of a single secondary
“jump,” multiplying each by λ (the expected number of jumps) gives you the same
summary statistic for the total. This is a direct consequence of how moments of random
sums combine (it comes from differentiating (7.9) at z=1; see Exercise 7.5).
Why this section matters for model selection: Two frequency distributions can have the
exact same mean and variance and still look completely different in the tail, because they
have different skewness. The book derives μ3 in terms of μ , σ 2 for five named distributions,
all sharing the pattern
2
2 ( σ 2−μ )
μ3=3 σ −2 μ +c ⋅ ,
μ
where the constant c differs by distribution:
Distribution c
Poisson c=0 (no third term)
(baseline)
Poisson–binomial m−2
0≤ <1
m−1
Negative binomial c=2
Polya–Aeppli c=3 /2
Neyman Type A c=1
Poisson–ETNB r +2
c= ∈ (1 , ∞)
r +1

How to read this table: given a fixed mean and variance, the only thing that changes the
skewness is this multiplier c in front of the “excess variance squared over mean” term. So:
if you fit two distributions to the same mean/variance and get very different answers
for higher moments, this table tells you exactly which named family is flexible
enough to match the actual skewness in your data — and in particular, the Poisson–
ETNB family, because c ranges continuously from just above 1 to infinity, can match
essentially any skewness greater than Poisson’s own.

Worked Example 7.8 — Selecting a model using real accident data (Hossack et al.)
Number of
claims Observed frequency
0 565,664
1 68,714
2 5,177
3 365
4 24
5 6
6+ 0

From this data the sample mean, variance, and third central moment are μ=0.1254614,
σ =0.1299599, μ3=0.1401737. Compute:
2

2
μ 3−3 σ +2 μ
2
=7.543865 .
( σ 2−μ ) /μ
Checking the table: this value is way above 2 (rules out negative binomial), way above 1.5
(rules out Polya–Aeppli), way above 1 (rules out Neyman Type A), and it isn’t in ¿ (rules out
Poisson–binomial). Only the Poisson–ETNB’s range ( 1 , ∞ ) accommodates it. Solving
r +2
7.543865= ⟹ r=−0.8471851 ,
r +1
which gives a specific, negative, non-integer r — a value only reachable by the flexible
Poisson–ETNB family. Lesson: matching just three simple sample statistics (mean,
variance, third moment) against this one table is often enough to immediately rule out
most of the “usual suspect” distributions before you even do a formal statistical fit.

7.2.3 Compound Poisson Is Closed Under Convolution (Theorem 7.5)

Theorem 7.5: If S1 , … , S k are independent compound Poisson random variables (with
(i )
parameters λ i and secondary distributions q n ), then their sum S=S 1+⋯+ S k is also
compound Poisson, with

λ 1 q (n1 )+⋯+ λ k q(nk )

λ=λ 1+⋯+ λk , qn= .
λ
Plain English: adding several independent compound-Poisson risks together doesn’t
create some exotic new type of distribution — it just produces another compound Poisson
distribution, whose secondary distribution is a λ i-weighted average (a mixture) of the
individual secondary distributions.
Why this is practically huge: without this theorem, if you wanted the distribution of the
sum of, say, 5 different lines of business (each compound Poisson), you would need to
convolve 5 separate recursions together. With this theorem, you build one combined
secondary distribution (a weighted blend) and run the compound-Poisson recursion
(7.7) only once.

Worked Example 7.9 — Combining two compound Poisson risks numerically

S1: λ 1=2, secondary q (11 )=0.2 ,q (21 )=0.7 , q3(1)=0.1. S2: λ 2=3, secondary
(2 ) ( 2) (2 )
q 2 =0.25 , q3 =0.6 , q 4 =0.15.

Combined λ=2+3=5. Weights are λ 1 / λ=0.4 and λ 2 / λ=0.6 :

q 1=0.4 ( 0.2 )+ 0.6 ( 0 )=0.08

q 2=0.4 ( 0.7 ) +0.6 ( 0.25 ) =0.43

q 3=0.4 ( 0.1 )+ 0.6 ( 0.6 ) =0.40

q 4=0.4 ( 0 ) +0.6 ( 0.15 )=0.09

Now run the ordinary compound-Poisson recursion (7.7) once, with λ=5 and this blended
q n, starting from g0=e−5:
x
5
Pr ( S=x )= ∑ n q n Pr ( S=x−n ) , x=1 , 2, …
x n=1
Worked Example 7.10 — Convolving several negative binomials
If N 1 , … , N k are independent negative binomials with a common β but possibly different r i ,
their sum is again negative binomial with r =r 1+ ⋯+r k and the same β — no need for
Theorem 7.5 at all, just add the r ’s (this is the discrete-distributions version of “sums of
independent Gammas with the same scale add their shape parameters”).
If the β i’s differ, it’s trickier — but Example 7.5 already showed every negative binomial is
compound Poisson with a logarithmic secondary distribution and λ i=r i ln ( 1+ β i ). Now
Theorem 7.5 applies directly: the sum is compound Poisson with

∑ r i [ β i / ( 1+ β i ) ] /n
n

λ=∑ r i ln ( 1+ β i ) , qn=
i
,n=1 ,2 , …
i ∑ r i ln ( 1+ βi )
i

and you compute the whole sum’s distribution with a single recursive pass (7.7), rather
than convolving k separate negative binomials pairwise — an enormous computational
saving when k is large (e.g., many lines of business or many territories).

7.2.4 Section 7.2 Exercises — Worked

Exercise 7.4. Si are independent compound Poisson with parameters λ i, all sharing the same
secondary pgf P2 ( z ) . Find the distribution of S=S 1+⋯+ S n.
Solution: This is Theorem 7.5 in the special case where every secondary distribution is
identical, so the weighted average of identical things is just that thing again: S is compound
Poisson with λ=λ 1+⋯+ λn and the same secondary pgf P2 ( z ) unchanged. (Sanity check: this
matches the ordinary Poisson-additivity rule you already know — sums of independent
Poissons add their means — just lifted up one level to compound Poissons.)

Exercise 7.5. Show P( k ) (1 )=E [ N ( N −1 ) ⋯ ( N −k +1 ) ] for any pgf, and use it to confirm (7.10).

Solution sketch: Differentiate P ( z )=∑ pn z term by term k times: each differentiation

n

n
brings down one factor from the exponent, so

P ( z ) =∑ n ( n−1 ) ⋯ ( n−k +1 ) p n z .
( k) n−k

n ≥k

Setting z=1 collapses z n−k → 1, leaving exactly

∑ n ( n−1 ) ⋯ ( n−k + 1 ) pn =E [ N ( N −1 ) ⋯ ( N −k +1 ) ] — this quantity is called a factorial
n
moment. Applying this to the compound Poisson pgf P ( z )=exp { λ [ Q ( z )−1 ] } for k =1 ,2 , 3
and converting the resulting factorial moments into ordinary raw/central moments (using
the standard identities relating factorial moments to raw moments, e.g.
E ( N 2 ) =E [ N ( N −1 ) ] + E ( N )) reproduces exactly the three lines of (7.10): each moment of S
equals λ times the matching raw moment of the secondary distribution.
Exercise 7.6. Verify the three moments in (7.11) for the Poisson–binomial distribution.
Solution sketch: the secondary distribution here is binomial( m , q ) , whose raw moments are
standard: m1 ' =mq , and m2 ' , m3 ' follow from the binomial’s known variance mq ( 1−q ) and
skewness formulas. Plug these into (7.10): μ= λ ( mq ); for the variance, using
2 2 2
m2 ' =Var ( M ) + ( m1 ' ) =mq ( 1−q ) +m q and simplifying σ 2=λ m2 ' leads, after factoring out
μ= λmq, to σ 2=μ [ 1+ ( m−1 ) q ] exactly as given. The third-moment line follows the same
substitute-and-simplify pattern using the binomial’s third raw moment; the algebra is
mechanical once you have m1 ' ,m2 ' , m3 ' for the binomial secondary distribution, so the
“hard part” is just correctly recalling/deriving those three binomial moments, not the
compound-Poisson step itself.

7.3 Mixed-Frequency Distributions

7.3.1 The General Idea: “Randomizing the Parameter Itself”
Compounding (Section 7.1) builds a new random variable as a random sum. Mixing is a
completely different mental model: it states that the parameter of an otherwise ordinary
distribution is itself unknown/random, varying across the population.
Story: In auto insurance, not all drivers are equally risky. Suppose that, given a particular
driver’s personal “risk level” θ , their annual claim count follows a Poisson( θ ) distribution.
But θ itself varies from driver to driver according to some distribution U ( θ ) (some drivers
are just riskier than others, even after you’ve accounted for all the rating variables you
know about — age, experience, violation history, etc.). Because you cannot observe θ
directly for any individual driver, when you pool everyone’s data together you get a
mixture:
P ( z )=∫ P ( z ∣θ ) u ( θ ) dθ or P ( z )=∑ P ( z ∣ θ j ) u ( θ j )
j

depending on whether the mixing variable Θ is continuous or discrete, with matching

formulas for the probabilities themselves:
pk = ∫ p k ( θ ) u ( θ ) dθ or pk =∑ pk ( θ j ) u ( θ j ) .
j

Crucial practical point: the mixing distribution U ( θ ) is unobservable — your data only
ever shows you the mixed (pooled) distribution, never which θ any individual policyholder
actually had.

Worked Example 7.11 — Zero-modification is also a two-point mixture

P ( z )= p ( 1 )+ ( 1− p ) P 2 ( z ) .

Here the mixing is between “degenerate at zero, always” (probability p) and “the
distribution P2” (probability 1− p ). This is literally the same object as compounding a
Bernoulli primary with P2 as secondary (Example 7.1) — a nice illustration that the same
final distribution can be derived in more than one conceptual way (compounding
vs. mixing); which story you tell doesn’t change the math, but the two stories offer different
intuitions and different practical mixing-parameter ideas.

Worked Example 7.12 — Binomial mixed over a Beta distribution (“binomial–beta”)

If q (the binomial success probability) has a Beta( a , b ) prior instead of being fixed,
integrating it out gives

( )( )( )
pk = −a −b / −a−b ,k =0 , 1 , … ,m ,
k m−k m

known as the negative hypergeometric or Polya–Eggenberger distribution. (You don’t

need to memorize the Beta-integral algebra — the point is the mechanism: replace the fixed
binomial parameter with a random one, integrate over its distribution, and out pops a new
named distribution.)

Worked Example 7.13 — Negative binomial mixed over Beta (“generalized Waring”)

Mixing the negative binomial’s parameter p= (1+ β )−1 over a Beta( a , b ) distribution gives the
generalized Waring distribution. Special cases have their own names: b=1 gives the
Waring distribution; r =b=1 gives the Yule distribution. These are historically
important in ecology/linguistics (word-frequency and species-abundance models) as well
as insurance.

7.3.2 Mixed Poisson Distributions — The Most Important Special Case

Take pk ( θ ) in (7.13) to be Poisson( θ ) .

Worked Example 7.14 — A simple, real, two-point mixture (good/bad drivers)

Suppose a fraction p of drivers are “good” (Poisson( λ 1 )) and the rest are “bad” (Poisson( λ 2 )):
− λ1 k −λ2 k
eλ1 e λ2
pk = p + ( 1− p ) .
k! k!
Fitting this to real Swedish/Swiss auto data (Tröbliger), the estimated parameters were
^p=0.94, ^λ 1=0.11, ^λ 2=0.70 — i.e., about 94% of drivers are “good” (0.11 expected
claims/year) and 6% are “bad” (0.70 expected claims/year). Important caveat: even after
fitting this model, you can never go back and say which specific drivers were the “bad” ones
— the labels are a modeling convenience, not something recoverable from individual data.
Two general lessons from this example:
1. Reality probably has a whole continuum of risk levels, not just two discrete groups
— a two-point mixture is a simplification.
2. Finite mixtures cost you parameters fast: a 2-group mixture already has 3
parameters ( p , λ1 , λ2); going to r groups costs r −1 mixing weights plus all the
 component parameters. This is why, in practice, continuous mixing distributions
are usually preferred — they typically need only 1–2 extra parameters no matter
how “many groups” of risk they effectively represent.

The Key Structural Facts About Mixed Poisson Distributions

Introduce a scale parameter λ so that, given Θ=θ , the Poisson mean is λθ . Then

u ( θ ) dθ=M Θ [ λ ( z−1 ) ] ,
λθ ( z−1)
P ( z )=∫ e

where M Θ ( ⋅ ) is the moment generating function (mgf) of the mixing distribution. In plain
words: a mixed Poisson pgf is always just the mixing distribution’s mgf, evaluated at the
point λ ( z−1 ) instead of at its usual argument. This is an extremely convenient shortcut —
you don’t have to do the integral in (7.12) yourself if you already know Θ ’s mgf from a
probability table.
Consequence 1 — mean: differentiate once and set z=1: E ( N )= λ E ( Θ ).
Consequence 2 — variance is always bigger than the mean (overdispersion):
differentiating twice gives, after simplification,

Var ( N )=λ 2 Var ( Θ ) + E ( N ) > E ( N ) .

This is one of the most important facts in the whole chapter. A plain Poisson always
has variance exactly equal to its mean. Real insurance claim-count data is almost always
more spread out than a Poisson would predict (this is called overdispersion), and mixed
Poisson models are the natural, mathematically clean way to build that extra spread into
your model — the “extra” variance λ 2 Var ( Θ ) is literally just the amount of hidden
heterogeneity in risk across your population.
Consequence 3 — a convenient normalization: because most continuous mixing
distributions used here have a scale parameter, changing λ in (7.14) is the same thing as
changing the mixing distribution’s own scale parameter. So without loss of generality you
can just set λ=1 whenever the mixing distribution already has its own scale parameter —
one fewer thing to keep track of.
Consequence 4 — uniqueness (Douglas’ theorem): two different mixing distributions
can never produce the exact same mixed Poisson distribution. This means that, in principle,
if you know the mixed (pooled) distribution is Poisson-mixed-over-something, you can
work backward and uniquely identify what that “something” (the mixing distribution) must
have been.

Infinite Divisibility — the Bridge Between Mixing and Compounding

Definition 7.6 (infinitely divisible): A distribution’s characteristic function φ ( z ) is
n
infinitely divisible if, for every n=1 ,2 , 3 , …, φ ( z )=[ φn ( z ) ] for some legitimate characteristic
function φ n. In plain words: you can always “split” the distribution into n independent
identical pieces that add up to it, no matter how large n is. (Contrast: you cannot split a
single coin flip into two independent “half coin flips” that sum to it — Bernoulli/binomial-
type variables with a hard upper cap are generally not infinitely divisible.)

Definition 7.7 (characteristic function): φ X ( z )=E ( e izX ) — like a pgf or mgf, just
guaranteed to exist for literally every distribution (even ones with very heavy tails where
the mgf might blow up), which is why the formal definition uses it. For the distributions in
this book, if the pgf exists, Theorem 7.8 shows P X ( z )=φ (−iln z ) and φ X ( z )=P ( eiz ) , so you can
freely translate between the two — you don’t need to actually work with characteristic
functions in the exercises, just know they exist as the “universal” version of the
transform.
Known infinitely divisible distributions: normal, gamma, Poisson, negative binomial.
Not infinitely divisible: binomial (its exponent m must be a whole number, and m/n
generally isn’t) — more generally, no distribution with a hard finite upper limit can be
infinitely divisible.
Theorem 7.9 (the big unifying result): if a mixed Poisson distribution’s mixing
distribution is infinitely divisible, then that same mixed Poisson distribution can also be
written as a compound Poisson distribution:
λ [ P2 ( z ) −1]
P ( z )=e

for some (unique, if you require P2 ( 0 )=0 ) secondary pgf P2 ( z ) .

Why you should care: this theorem hands you a free upgrade in computational power.
Mixed Poisson distributions, as originally defined via the integral in (7.12), can be nasty to
compute (you’d need numerical integration). But if the mixing distribution happens to be
infinitely divisible, Theorem 7.9 tells you the very same distribution can also be computed
via the simple, fast Panjer recursion (7.7) from Section 7.1 — you just need to identify λ
and P2 ( z ) first, using
ln P ( 0 )−ln P ( z )
λ=−ln P ( 0 ) , P2 ( z )= .
ln P ( 0 )
A philosophical warning worth remembering: because the same distribution can arise
from two totally different stories (mixing heterogeneous risk vs. random-sum
compounding of accidents-and-claims), a good statistical fit does not prove which story
is “really” happening in your data. Negative binomial claim counts fitting well doesn’t
prove individuals are secretly Poisson with gamma-distributed risk — it’s equally
consistent with an accident-and-multiple-claims compounding story.

Worked Example 7.15 — Negative binomial as both mixed and compound Poisson
1 /n
The gamma distribution’s mgf is M Θ ( t ) =( 1−βt ) , which is infinitely divisible ([ M Θ ( t ) ]
−r
is
again a gamma mgf, just with shape r /n ). Plug into (7.14) with λ=1:
−r
P ( z )=M Θ ( z −1 )=[ 1−β ( z−1 ) ] ,
— the familiar negative binomial pgf, now derived as a Poisson mixed over a Gamma.
Since gamma is infinitely divisible, Theorem 7.9 guarantees the same distribution is also
compound Poisson; using (7.15)-(7.16) directly gives λ=r ln ( 1+ β ) and P2 ( z ) works out to
the logarithmic pgf — matching Example 7.5’s earlier, purely algebraic derivation, but
arrived at here with more insight into why it must work (because gamma is infinitely
divisible).

Worked Example 7.16 — Poisson–inverse Gaussian

Mixing Poisson over an inverse Gaussian distribution is shown (via the same mgf-
substitution technique) to be infinitely divisible, and by Theorem 7.9 is therefore also
compound Poisson — specifically, Poisson primary with an ETNB(r =−0.5) secondary
distribution. This distribution’s official name in the book, Poisson–inverse Gaussian
(PIG), is a genuinely useful, flexible two-parameter frequency distribution beyond the
( a , b , 0 ) / ( a , b ,1 ) classes.

Worked Example 7.17 — Neyman Type A via mixing

If the mixing distribution is itself Poisson( μ ) , (7.14) directly gives

P ( z )=exp {μ [ e λ ( z−1)−1 ] },
i.e., exactly the Neyman Type A / Poisson–Poisson pgf from Example 7.2 — confirming that
Neyman Type A can be built either as a compound distribution (Section 7.1’s story) or as a
mixed Poisson distribution (this section’s story). Table 7.2 below summarizes these three
double-identity distributions.
Table 7.2 — Distributions with two identities (compound and mixed):

Compound-Poisson
Name secondary Mixing distribution
Negative binomial Logarithmic Gamma
Neyman Type A Poisson Poisson
Poisson–inverse ETNB (r =−0.5) Inverse Gaussian
Gaussian

One more shape fact worth knowing: Holgate’s result — if the mixing distribution is
continuous and unimodal (one hump), the resulting mixed Poisson distribution is also
unimodal. Multiple modes (humps) can only appear when the mixing distribution is itself
discrete (as in the two-point good/bad-driver example, or in Neyman Type A for extreme
parameter combinations).

7.3.3 Section 7.3 Exercises — Worked

Exercise 7.7. Show negative-binomial–Poisson compounding equals a mixed Poisson with
negative binomial mixing.
Solution sketch: “negative-binomial–Poisson” means the primary distribution is negative
binomial and the secondary is Poisson. Using Example 7.5’s identity (negative binomial ¿
compound Poisson with logarithmic secondary, i.e. NB ( z ) =e λ [Q ( z)−1 ]) is a different
log

compounding than what’s asked here — instead, directly use (7.9)-style algebra: writing
out P NB [ PPoisson ( z ) ] and comparing against (7.14) with Θ ∼ negative binomial shows the two
expressions are algebraically identical term by term, because both reduce to the same
exponential-of-power-function form once you substitute the negative binomial pgf into
itself. The practical takeaway (rather than the line-by-line algebra) is: just as Poisson-
mixed-over-Gamma gives negative binomial (Example 7.15), negative-binomial-
mixed-over-Poisson-primary via compounding gives back a mixed Poisson with
negative-binomial mixing — these dual constructions keep reappearing throughout the
chapter.
Exercise 7.8. N i (i=1 , … , n) are mixed Poisson with a common scale λ but each its own
mixing pgf Pi ( z ) (for the mixing random variable). Find the mixing pgf for N=N 1+⋯+ N n.

Solution: By (7.14), P N ( z )=M Θ [ λ ( z−1 ) ] . Since the N i are independent,

i i

P N ( z ) =∏ P N ( z )=∏ M Θ [ λ ( z−1 ) ] . Because a product of mgfs evaluated at the same

i i
i i
argument is the mgf of the sum of the underlying (independent) random variables, this
equals M Θ +⋯ +Θ [ λ ( z−1 ) ]. So N is again mixed Poisson with the same scale λ , and its mixing
1 n

distribution is simply Θ=Θ1 +⋯+Θn (the sum of the individual mixing random variables) —
in pgf terms (for discrete Θ ) the mixing pgf of N is ∏ P Θ ( z ) .
i i
i

Exercise 7.9. Poisson given Θ=θ has mean λθ ; Θ ’s distribution has a scale parameter. Show
the mixed distribution doesn’t depend on λ .
Solution: If Θ has a scale parameter c (so that cΘ has the same distributional shape, just
rescaled), then multiplying the Poisson-conditional mean by λ is indistinguishable from
rescaling Θ ’s own scale parameter by λ — i.e., λΘ and Θ (with a rescaled parameter)
describe the exact same family of mixing distributions. Formally, from (7.14),
P ( z )=M Θ [ λ ( z−1 ) ]; if Θ has scale parameter c , then M Θ ( λt )=M Θ ' ( t ) where Θ ' has scale
parameter cλ — still just a member of the same mixing family, only relabeled. So the mixed
distribution’s shape doesn’t change with λ ; only the mixing distribution’s scale parameter
silently absorbs it. This is exactly Consequence 3 above, generalized.

Exercise 7.10. Given a specific pdf u ( θ )=α 2 ( α + 1 )−1 ( θ+1 ) e−αθ , find the mixed Poisson pf, and
show it is also a compound distribution.
Solution approach: Plug u ( θ ) into (7.13)’s integral form and integrate term by term (this pdf
is a simple mixture of two Gamma-type pieces — a Gamma (1 , α ) term and a Gamma ( 2, α )
term, weighted — so the integral splits into two standard Gamma-Poisson mixture
integrals, each of which is already known to produce geometric/negative-binomial-shaped
terms). The resulting pn turns out to be expressible as a mixture of two negative binomial
pfs — and because a finite mixture of distributions can always be written as a compound
distribution with a two-point (or more generally discrete) primary distribution (exactly the
logic of Example 7.11 applied here with more than one non-trivial component), this
confirms the mixed distribution is simultaneously a compound distribution.
Exercise 7.11. Show the mixed Poisson with a positive stable mixing distribution is
compound Poisson with a Sibuya secondary distribution.

Solution sketch: The positive stable distribution’s Laplace transform is given as exp (−s α ) ,
0< α <1. Laplace transforms of positive random variables are mgfs evaluated at t=−s, so
plugging s= λ ( 1−z ) into the given Laplace transform reproduces exactly (7.14)’s form:
P ( z )=exp [−λ α ( 1−z ) α ] . The positive stable distribution is infinitely divisible essentially by
construction (raising its Laplace transform to the 1/n power just rescales s inside the
exponential, giving another positive stable transform) so Theorem 7.9 applies, and
α
matching (7.15)-(7.16) to this expression identifies the secondary pgf as P2 ( z ) =1−( 1−z )
— which is precisely the Sibuya distribution’s pgf (with parameter r =−α ; see Exercise
7.17(d) below for its probabilities).
Exercise 7.12 & 7.13. These ask you to substitute the given Poisson-inverse-Gaussian pgf /
reciprocal-inverse-Gaussian mgf into the general formulas and simplify — mechanical
exercises in “plug into (7.14)/(7.15)/(7.16) and simplify the resulting square roots and
factorials,” directly extending Example 7.16’s approach. The result of Exercise 7.13 in
particular identifies the secondary distribution as an ETNB with r =−0.5, the same family
that appears throughout this section whenever an inverse-Gaussian-type mixing
distribution is involved.
¿
Exercise 7.14. (a) If N∨ Λ=λ ∼ Poisson( λ ) and Λ = Λ+ μ (a constant shift), then,
conditioning as usual,

P N ( z )=E ( e Λ ( z −1 ) )=e μ ( z−1 ) E ( e Λ ( z −1 ) )=e μ ( z−1 ) P N ( z ) ,

¿

which is exactly the pgf of N compounded (added independently) with an extra

¿
Poisson( μ ) component — i.e., N has the same distribution as N +Poisson ( μ ), independent
of N . In words: shifting the mixing variable by a fixed constant μ is the same as just adding
an independent Poisson( μ ) noise term on top of the original mixed distribution. (b) If Λ has
an inverse Gaussian distribution, then N ’s pgf P N [ Q ( z ) ] =e μ ( Q ( z)−1 ) P N [ Q ( z ) ] is a product of
¿
¿

two compound Poisson-type pgfs (Example 7.16’s Poisson–ETNB piece times a new
Poisson–Q piece); by Theorem 7.5 (closure under convolution) you can combine them into
a single compound Poisson recursion by blending their two secondary distributions exactly
as in Example 7.9 — compute the combined λ and blended secondary probabilities first,
then run (7.7) once.

7.4 The Effect of Exposure on Frequency

The business question: if your book of business currently has n policyholders (“entities”)
and you know the current frequency distribution N , what does the frequency distribution
¿
look like if the book grows to n policyholders — assuming the new policyholders are
statistically identical to the current ones?
If N j is the (i.i.d.) claim count for the j -th entity, then the whole book’s count is
N=N 1+⋯+ N n, so P N ( z ) =[ PN ( z ) ] . If we scale up to n¿ entities,
n
1

¿ ¿
n n /n
P N ( z )=[ P N ( z ) ] =[ P N ( z ) ]
¿
1
.

In plain English: to rescale a frequency distribution for exposure changes, you literally
raise its pgf to the power (new size / old size). If this operation still produces a legitimate
pgf — i.e., if N is infinitely divisible (the same concept from Section 7.3.2!) — then the
¿
rescaled distribution N is the same family, just with adjusted parameters. This is precisely
why infinite divisibility, introduced earlier purely to connect mixing and compounding,
turns out to also be the right tool for handling exposure/portfolio-size adjustments.

Worked Example 7.18

A group of 300 workers has negative binomial claims with β=0.3 , r =10. Find the frequency
distribution for an expanded group of 500 workers.
−10 500/300 −16.67
P N ( z )={ [ 1−0.3 ( z−1 ) ]
¿ } =[ 1−0.3 ( z−1 ) ] ,

still negative binomial, same β=0.3 , but with r rescaled from 10 to 16.67 (simply
10 ×500 /300 ). Notice how simple this is: because negative binomial is infinitely divisible,
scaling exposure only ever touches the r parameter — you never need to refit the whole
model from scratch.

Which Distributions Behave This Way?

• ( a , b , 0 ) class: every member except the binomial is infinitely divisible and rescales
cleanly this way. (The binomial fails because its m parameter must stay a whole
¿
number, and m ×n /n generally isn’t.)
• ( a , b , 1 ) class: none of its members (zero-truncated/zero-modified versions) are
infinitely divisible — the “reset zero probability” operation breaks the clean
rescaling property.
• Compound distributions: rescale cleanly whenever the primary distribution is
infinitely divisible — so compound Poisson and compound negative binomial
(including geometric) distributions are always safe choices if you expect your
book of business to grow or shrink and want your frequency model to remain
internally consistent under that growth.
Practical modeling advice from the book: if you anticipate future exposure changes, it
may be worth choosing a compound-Poisson-type model even if a zero-modified ( a , b , 1 )-
class model technically fits your current data slightly better — because the zero-modified
model has no clean, principled way to rescale for a different portfolio size, while the
compound model does. As a bonus, compound models are also good at producing a large
spike of probability exactly at zero, similar to what zero-modification achieves, so you often
don’t lose much fit quality by making this choice.
7.5 An Inventory of Discrete Distributions
By this point in the chapter you have assembled: the ( a , b , 0 ) class, its extension to ( a , b , 1 )
(adding zero-truncation/zero-modification), and then two entirely different construction
methods — compounding (Section 7.1) and mixing (Section 7.3) — each of which can
produce either brand-new distribution families or turn out to secretly equal a distribution
you already had.
Table 7.3 in the book organizes all of this into a single reference: which distributions are
special cases of others (you get them by fixing a parameter at a boundary value), and which
are limiting cases of others (you get them by sending a parameter to 0 or ∞ ).

Distribution Special case of Limiting case of

Poisson ZM Poisson Negative binomial, Poisson–binomial,
Poisson–inverse Gaussian, Polya–Aeppli,
Neyman Type A
ZT Poisson ZM Poisson ZT negative binomial
ZM Poisson — ZM negative binomial
Geometric Negative binomial, —
Geometric–Poisson, ZM
geometric
ZT geometric ZT negative binomial —
ZM geometric ZM negative binomial —
Logarithmic ZM logarithmic ZT negative binomial
Binomial ZM binomial —
Negative ZM negative binomial, —
binomial Poisson–ETNB
Poisson–inverse — Poisson–ETNB
Gaussian
Polya–Aeppli — Poisson–ETNB
Neyman Type A — Poisson–ETNB

How to read a row: e.g. “Poisson is a special case of ZM Poisson” means: setting the ZM
Poisson’s arbitrary zero-probability parameter back to its “natural” value recovers plain
Poisson. “Poisson is a limiting case of negative binomial” means: sending the negative
binomial’s r → ∞ (with mean held fixed) converges to Poisson — this is the same limiting
relationship you saw back in Chapter 6.
Why the book stops here: compound distributions where the primary distribution is itself
a two-parameter distribution (like negative binomial or Poisson-inverse Gaussian as the
primary, rather than as the secondary) are deliberately excluded from the inventory,
because — as Sections 7.2–7.3 repeatedly demonstrated — these tend to just be
generalizations of compound Poisson distributions already on the list (remember: negative
binomial itself is compound Poisson via Example 7.5/7.15). Including them wouldn’t add
genuinely new shapes, just duplicate bookkeeping.

7.5.1 Section 7.5 Exercises — Worked

Exercise 7.15. Calculate Pr ( N =0 ) , Pr ( N =1 ) , Pr ( N =2 ) for a list of named distributions with
given parameters. This exercise is pure “plug the given parameters into the formula you
already know from Chapters 5–6 (or this chapter for the compound ones) and evaluate.”
Below are the key formulas and worked answers for the most instructive parts; the
remaining parts follow the identical mechanical pattern.

• (a) Poisson( λ=4 ): pk =e−λ λ k /k !. p0=e−4=0.018316, p1=4 e−4 =0.073263,

−4
p2=8 e =0.146525 .
k k +1
• (b) Geometric( β=4): pk = β / ( 1+ β ) . p0=1 /5=0.2, p1=4 /25=0.16,
p2=16 /125=0.128 .

( )
k
• (c) Negative binomial(r =2 , β=2): pk = ( k )
k + r−1 (1+ β )−r β
1+ β
2
. p0= (1 /3 ) =0.111111
2 2 2
, p1=2 ( 1/3 ) ( 2/3 )=0.148148 , p2=3 ( 1/3 ) ( 2/3 ) =0.148148 .

•
8 8
()
(d) Binomial(m=8 ,q=0.5 ): pk = 0.5 . p0=0.5 =0.003906 , p1=8 ( 0.58 )=0.031250,
k
8

p2=28 ( 0.5 8 )=0.109375 .

k
1 β
• (e) Logarithmic( β=4): pk = for k ≥ 1, p0=0. ln ( 5 ) =1.609438;
ln ( 1+ β ) k
4 1 1 1 β
p1 = ⋅ ⋅ — more precisely p1= ⋅ =4 /1.609438=2.48562… (note:
1.609438 1 1 ln ( 1+ β ) 1
p
since 1 must be , remember the correct normalized formula is
≤ 1
k

pk =
1 [ β / (1+ β ) ]
; using β / ( 1+ β )=0.8): p0=0, p1=0.8 /1.609438=0.497087 ,
ln ( 1+ β ) k
p2=0.64 / ( 2 ×1.609438 ) =0.198835.

• (f) ETNB(r =−0.5 , β=4 ): built from the truncated negative binomial recursion in
T rβ / ( 1+ β )
Chapter 6, starting from p1 = style formulas and recursing — same
(1+ β )r −1
mechanics as the worked ETNB numbers already used throughout this chapter’s
Examples 7.4/7.6.

• (g)–(k) (Poisson–inverse Gaussian, ZM geometric, Neyman Type A, Poisson–ETNB,

Poisson–ZM-geometric): each of these is computed using exactly the compound-
Poisson machinery of Section 7.1 — get f 0 , f 1 , f 2 for the named secondary
distribution from Chapter 6’s formulas, get g0=e ( ) from (7.7), then apply the
−λ 1−f 0
 k
λ
recursion gk = ∑ j f j g k− j for k =1 ,2 — identical arithmetic pattern to Examples
k j=1
7.4 and 7.6 worked in full above, just with different secondary-distribution inputs.
This is deliberate: the whole point of Exercise 7.15 is to make you practice the same
two-step recipe (get secondary f j ’s, then Panjer-recurse) across every named
distribution in the chapter until it’s automatic.

Exercise 7.16. Verify the zeta distribution is not a member of the ( a , b , 1 ) class.
T − ( ρ+1)
Solution: The zeta distribution has pk =k /ζ ( ρ+1 ) . Compute the ratio
T

( )
ρ +1
pk k −1
T
= .
p k−1 k

For this to be of the required ( a , b , 1 )-class form a+ b/k , this ratio would need to be

( )
ρ+1
k −1 ρ+1
expressible as a constant plus a constant divided by . But k =( 1−1 /k ) is a
k
genuinely nonlinear function of 1/k (unless ρ+1 happens to be exactly 1, which isn’t
allowed generically) — expanding it in powers of 1/k produces infinitely many terms (a
binomial-series expansion), not just a constant-plus-1/k combination. Since no choice of
constants a , b can match ( 1−1 /k )ρ +1 for every k simultaneously, the zeta distribution’s ratio
is not of the required linear form, so it is not a member of the ( a , b , 1 ) class.

Exercise 7.17. Tail probabilities a n=Pr ( N > n ).

∞
(a) E ( N )=∑ an : this is the classic “tail-sum” formula for the expectation of a nonnegative
n=0
∞ ∞ k−1
integer random variable. Proof: ∑ an=∑ ∑ p k =∑ p k ∑ 1=∑ k pk =E ( N ) — i.e., swap the
n=0 n =0 k >n k n=0 k
order of summation and count, for each k , how many n ’s satisfy n< k (there are exactly k of
them), which directly reproduces the definition of the mean.

(b) A ( z )=[ 1−P ( z ) ] / ( 1−z ): from part (a)’s swapped-sum logic generalized to a generating
k−1
1−z k 1−P ( z )
function, A ( z )=∑ an z =∑ p k ∑ z =∑ pk
n n
= , using the finite geometric series
n k n=0 k 1−z 1−z
k −1
formula for ∑ z . As z → 1: both numerator and denominator go to zero, so by L’Hôpital’s
n

n=0
rule (or just recognizing this as the definition of a derivative),
1−P ( z )
A ( 1 )=lim ¿ z →1 =P ' ( 1 )=E ( N ) ¿ — consistent with part (a).
1−z
(c) For negative binomial with integer r : writing a n as a sum over k =1 , … ,r of “sub-
negative-binomial” terms is proved by using the fact (mentioned in Exercise 7.3’s
discussion) that a negative binomial with integer r is a sum of r i.i.d. geometrics — the
survival function of this sum decomposes via a standard combinatorial identity into exactly
the given sum of r terms, each itself a negative-binomial-shaped probability with
parameter k =1 , … ,r .

(d) Sibuya distribution: using P ( z )=1−( 1−z )−r with −1<r < 0, its probabilities pn come from
a direct binomial-series (Newton’s generalized binomial theorem) expansion of ( 1−z )−r,
giving the stated Gamma-function formula; then applying part (b)’s formula
A ( z )=[ 1−P ( z ) ] / ( 1−z )=( 1−z )
−r−1
and expanding this again via the binomial series directly
produces a n= ( ) n+r
n
, a strikingly clean closed form (this is the reason the Sibuya
distribution reappears as the “compound Poisson secondary distribution” behind the
positive-stable mixing result of Exercise 7.11).
(e) Mixed Poisson case: substituting the mixed-Poisson pf formula directly into the tail-sum
definition of a n and swapping the order of integration and summation (same technique as
part (a), just with an extra integral over θ ) gives exactly the stated integral formula, where
[ 1−U ( θ ) ] appears because you are now summing the conditional Poisson tail probabilities
weighted by the mixing density, and the Poisson conditional-tail-sum-formula (itself a mini
version of part (a) applied to a plain Poisson, scaled by λ ) produces the λ multiplying
everything.

Chapter 7 — Master Summary Table

Concept One-line meaning Key formula
Compound Random sum: primary count PS ( z )=P N [ P M ( z ) ]
distribution of secondary-sized jumps
Panjer-type Fast way to compute (7.5) / (7.6), starting g0=P N ( f 0 )
recursion compound probabilities
Compound Poisson primary + any P ( z )=e λ [ Q ( z )−1], recursion (7.7)
Poisson secondary
Closure under Sum of independent Theorem 7.5, blend secondary
convolution compound Poissons is distributions by λ i weights
compound Poisson
Mixed Parameter itself is random P ( z )=∫ P ( z ∣θ ) u ( θ ) dθ
distribution across the population
Mixed Poisson Poisson with random mean ⇒ P ( z )=M Θ [ λ ( z−1 ) ], Var > E
overdispersion
Infinite Distribution splittable into n Bridges mixed ⇔ compound Poisson
divisibility i.i.d. pieces for any n (Thm 7.9); also governs exposure
scaling (Sec 7.4)
¿
Exposure Rescale a book of business P N ( z )=[ P N ( z ) ]
¿
n /n

scaling from n to n¿ entities

The one-sentence takeaway for the whole chapter: almost every “new” discrete
distribution you meet is really just the Poisson (or another familiar building block) wearing
a disguise — either randomly summed (compounding) or randomly reparametrized
(mixing) — and infinite divisibility is the mathematical fact that tells you exactly when
those two disguises are secretly the same distribution.