Why Mixtures Need Not Stay in the Same Panjer Class

Why Mixtures Require a Fresh Class Check

A frequency model may describe a population made up of several groups, each with a different count distribution. The population count then has a mixture distribution: first a group is selected, then a count is generated from that group’s distribution. Even if every group has a Panjer frequency distribution, it does not follow that the mixture does. The key issue is whether the mixture probabilities obey one common recurrence at every required index.

A ratio calculation is useful when its denominator is positive, but a ratio-only definition can miss a failed recurrence at a zero-probability index. For class membership, check the recurrence itself, including indices where the preceding probability is zero and the terminating zero for finite support.

Definition: Let $p_k=P(N=k)$ for a count $N$ on the nonnegative integers. The distribution is in the $(a,b,0)$ Panjer class if $$ p_k=\left(a+\frac{b}{k}\right)p_{k-1},\qquad k\geq 1, $$ at every required index. The distribution is in the $(a,b,1)$ Panjer class if this recurrence holds for every $k\geq 2$; $p_0$ and $p_1$ are separately specified probabilities. In either definition, the recurrence must hold even when $p_{k-1}=0$. In particular, for finite support, it must correctly generate the zero probability at and beyond the support endpoint.

When $p_{k-1}>0$, define the probability ratio $r_k=p_k/p_{k-1}$. The corresponding recurrence requires $r_k=a+b/k$. But this ratio statement alone is not a complete membership test: if $p_{k-1}=0$, the ratio is undefined, while the recurrence still imposes a condition on $p_k$. A positive $p_k$ after a zero $p_{k-1}$ violates the recurrence.

What the Mixture Ratio Shows

Suppose a component $J$ is selected from $m$ distributions with probabilities $w_j$, where $w_j\geq 0$ and $\sum_{j=1}^m w_j=1$. Let $p_{k,j}$ be the probability of count $k$ in component $j$. The mixture probability is

p_k=\sum_{j=1}^{m}w_jp_{k,j}.

At an index where every positive-weight component with $p_{k-1,j}=0$ also has $p_{k,j}=0$, the ratio can be written as a weighted average of component ratios:

\frac{p_k}{p_{k-1}} = \sum_{j:\,p_{k-1,j}>0} \frac{w_jp_{k-1,j}}{\sum_{\ell=1}^{m}w_\ell p_{k-1,\ell}} \frac{p_{k,j}}{p_{k-1,j}}, \qquad p_{k-1}>0.

The weights in this expression sum to one. They depend on $k$, however, because the composition of the mixture among groups with count $k-1$ can change with $k$. Thus, even if each component ratio has the form $a_j+b_j/k$, their weighted average need not have the form $a+b/k$ with fixed $a,b$.

The stated condition matters. If a positive-weight component has $p_{k-1,j}=0$ but $p_{k,j}>0$, its contribution appears in the numerator of the mixture ratio but its component ratio is undefined. The weighted-average identity then does not apply. A safer route is to form the mixture probabilities and check the recurrence directly.

Worked Example: A Mixture of Two Poisson Counts

A hypothetical portfolio has two equally likely segments. Conditional on the segment, the claim count is Poisson with mean 1 or Poisson with mean 3. Determine whether the unconditional count belongs to an $(a,b,0)$ Panjer class. Use probability ratios and show the failed check.

Method. For a Poisson count with mean $\lambda$, $p_{k}=\lambda p_{k-1}/k$. Therefore each component is in the $(a,b,0)$ class with $a=0$ and $b=\lambda$. For the mixture, first calculate the probabilities needed for the ratios; then test whether the ratios follow one common $a+b/k$.

Let $t=e^{-2}$, so $t\approx 0.1353353$. A common factor $e^{-1}/2$ cancels from all ratios. The mixture probabilities are

p_0=\frac{e^{-1}}{2}(1+t),\quad p_1=\frac{e^{-1}}{2}(1+3t),\quad p_2=\frac{e^{-1}}{2}\frac{1+9t}{2},\quad p_3=\frac{e^{-1}}{2}\frac{1+27t}{6}.

For example, $p_2$ averages the two component probabilities $e^{-1}/2$ and $9e^{-3}/2$, each with weight $1/2$. The other displayed probabilities follow by the Poisson formula in the same way. The ratios are therefore

r_1=\frac{1+3t}{1+t}\approx1.2384,\qquad r_2=\frac{1+9t}{2(1+3t)}\approx0.7889,\qquad r_3=\frac{1+27t}{3(1+9t)}\approx0.6994.

If the mixture were in the $(a,b,0)$ class, $r_1=a+b$ and $r_2=a+b/2$. Solving these two equations gives

b=2(r_1-r_2)\approx0.8990,\qquad a=2r_2-r_1\approx0.3394.

These values predict $r_3=a+b/3\approx0.6391$, not the actual $r_3\approx0.6994$. The mixture is therefore not in a single $(a,b,0)$ class. The calculation uses unrounded values in the comparison; displayed ratios are rounded to four decimal places.

Common mistake avoided. It would be incorrect to average the component $b$ parameters and conclude that the mixture has $a=0$ and $b=2$. The component weights relevant to a ratio change with the count, and the actual ratios fail the required common recurrence.

Why a Shared Recurrence Is Preserved

There is an important exception to the general lack of closure: if all mixture components obey the same recurrence coefficients at every required index, their mixture obeys that recurrence too. This follows from linearity of the probabilities, not from a claim that all mixtures preserve class membership.

Result: If each positive-weight component satisfies $p_{k,j}=(a+b/k)p_{k-1,j}$ at every index required for the chosen Panjer class, with the same $a,b$ for every component, then the mixture satisfies $p_k=(a+b/k)p_{k-1}$ at those indices. The result applies to $(a,b,0)$ for $k\geq1$, or to $(a,b,1)$ for $k\geq2$.

Write the mixture probabilities.
At each index, $p_k=\sum_j w_jp_{k,j}$, using the same component weights for every $k$.

Apply the shared component recurrence.
If every component obeys it, then $p_{k,j}=(a+b/k)p_{k-1,j}$ at the index being checked.

Factor out the common coefficient.
Summing gives $p_k=(a+b/k)\sum_jw_jp_{k-1,j}=(a+b/k)p_{k-1}$. This calculation remains valid when some or all preceding component probabilities are zero.

The zero case is handled directly by this proof. If $p_{k-1,j}=0$, the component recurrence with shared finite coefficients forces $p_{k,j}=0$ at that index. No undefined component ratio is needed. In contrast, if component parameters differ, the coefficient cannot generally be factored out of the sum.

Worked Example: Mixing Two Distributions With a Common Recurrence

Two segments have count distributions with probabilities $p_0=0.80,\ p_1=0.10$ and $p_0=0.60,\ p_1=0.20$, respectively. In each segment, the probabilities satisfy $p_k=0.5p_{k-1}$ for $k\geq2$. The segments are selected with probabilities $0.25$ and $0.75$. Find the first four mixture probabilities and verify the applicable recurrence.

Method. These are $(a,b,1)$ distributions with shared coefficients $a=0.5,\ b=0$. Compute the component probabilities from the recurrence, take the weighted averages, and check the recurrence for the mixture from $k=2$ onward.

In the first segment, $p_2=0.5(0.10)=0.05$ and $p_3=0.5(0.05)=0.025$. In the second, $p_2=0.5(0.20)=0.10$ and $p_3=0.5(0.10)=0.05$. Each component is a probability distribution: its positive-count probabilities sum to $p_1(1+0.5+0.5^2+\cdots)=2p_1$, so the totals are $0.80+0.20=1$ and $0.60+0.40=1$.

The mixture probabilities are

\begin{aligned} p_0&=0.25(0.80)+0.75(0.60)=0.65,\\ p_1&=0.25(0.10)+0.75(0.20)=0.175,\\ p_2&=0.25(0.05)+0.75(0.10)=0.0875,\\ p_3&=0.25(0.025)+0.75(0.05)=0.04375. \end{aligned}

The total probability is $0.65+0.175+0.0875+0.04375+\cdots=0.65+0.35=1$, since the remaining geometric tail sums to $0.04375/(1-0.5)=0.0875$. Also $p_2/p_1=0.5$ and $p_3/p_2=0.5$; the same shared recurrence continues at every $k\geq2$ by the linearity argument. Thus the mixture remains in the $(0.5,0,1)$ class.

Common mistake avoided. Do not conclude that all mixtures fail to remain in a Panjer class. Shared recurrence coefficients are sufficient for preservation, even when the components have different $p_0$ and $p_1$.

Support Gaps and the Limits of Ratio Checks

A gap in the support requires particular care. For example, suppose a count has $p_0=1/2,\ p_1=0,\ p_2=1/2$, with all other probabilities zero. At $k=1$, the $(a,b,0)$ recurrence requires $p_1=(a+b)p_0$, so $a+b=0$. At $k=2$, it then requires

p_2=\left(a+\frac{b}{2}\right)p_1=0,

contradicting $p_2=1/2$. A test that considers only ratios where the denominator is positive could miss this failure by leaving $k=2$ unchecked. In a mixture, a component can create exactly this problem when it contributes positive probability at $k$ despite having zero probability at $k-1$.

Worked Example: A Component Ratio That Is Undefined

Choose with equal probability between two hypothetical count components: one is concentrated at 1 and the other is concentrated at 2. Calculate the mixture ratio at $k=2$, and explain why it cannot be obtained as a weighted average of both component ratios.

Method. Form the mixture probabilities directly. This avoids assigning a ratio to a component with zero denominator.

The first component has $p_{1,1}=1,\ p_{2,1}=0$; the second has $p_{1,2}=0,\ p_{2,2}=1$. With weights $1/2$ each,

p_1=\frac12(1)+\frac12(0)=\frac12,\qquad p_2=\frac12(0)+\frac12(1)=\frac12, \qquad \frac{p_2}{p_1}=1.

The component ratio for the count-2 component at this index would be $p_{2,2}/p_{1,2}=1/0$, which is undefined. Its weight in a proposed ratio average based on $p_{1,2}$ would be zero, even though it contributes $1/2$ to the mixture numerator. Therefore the weighted-average identity does not apply. The mixture calculation itself is valid; it is the component-ratio argument that fails.

Common mistake avoided. A zero weight in the ratio-average formula does not mean a component contributes nothing to the mixture numerator. First check whether a component can move from zero probability at $k-1$ to positive probability at $k$.

A Reliable Test for a Mixture

For exam problems, separate probability construction from class classification. The following checks prevent the two main errors: assuming a mixture inherits its components’ class, and overlooking an index with a zero denominator.

Write the mixture probabilities $p_k=\sum_jw_jp_{k,j}$, with nonnegative weights summing to one.
Identify whether the target is $(a,b,0)$, requiring all indices $k\geq1$, or $(a,b,1)$, requiring all indices $k\geq2$.
At positive denominators, compare the mixture ratios with one pair of fixed coefficients $a,b$, not a new pair at each index.
At zero denominators, check the recurrence directly. A positive $p_k$ following $p_{k-1}=0$ cannot satisfy a finite-coefficient recurrence.
For finite support, check the first zero at the endpoint as well as all later indices required by the recurrence.

This distinction matters when deciding whether a frequency model is eligible for a Panjer recursion in a compound-loss calculation. A component-by-component classification is not enough: the actual unconditional count distribution must satisfy the required recurrence. Always verify the mixture sequence itself, or justify it using common coefficients and the full recurrence conditions.

Check Your Understanding

For each problem, apply the recurrence at every required index, not just where a convenient ratio is available.

A count is an equal mixture of Poisson distributions with means 2 and 4. Write the mixture probabilities $p_0,p_1,p_2$ in terms of $e^{-2}$ and $e^{-4}$, and state the ratio $r_1$. (Answer: $p_0=(e^{-2}+e^{-4})/2,\ p_1=(2e^{-2}+4e^{-4})/2,\ p_2=(2e^{-2}+8e^{-4})/2;\ r_1=(2e^{-2}+4e^{-4})/(e^{-2}+e^{-4})$.)
A distribution has $p_0=0.4,\ p_1=0,\ p_2=0.6$. Can it satisfy an $(a,b,0)$ recurrence? Identify the index that decides the answer. (Answer: No; at $k=1$, the recurrence forces $a+b=0$, and at $k=2$ it then forces $p_2=0$, contradicting $p_2=0.6$.)
Two components have $p_0=0.7,\ p_1=0.2$ and $p_0=0.5,\ p_1=0.4$, and both satisfy $p_k=0.25p_{k-1}$ for $k\geq2$. They are mixed with weights $0.6$ and $0.4$. Find the mixture $p_0,p_1,p_2$, and identify its class recurrence. (Answer: $p_0=0.62,\ p_1=0.28,\ p_2=0.07$; it satisfies the shared $(0.25,0,1)$ recurrence.)
An equal mixture consists of a $\operatorname{Binomial}(1,0.2)$ count and a $\operatorname{Binomial}(2,0.2)$ count. Its probabilities are $p_0=0.72,\ p_1=0.26,\ p_2=0.02,\ p_3=0$. Fit $a,b$ from $k=1,2$ and use $k=3$ to test $(a,b,0)$ membership. (Answer: $a+b=0.26/0.72=13/36$ and $a+b/2=0.02/0.26=1/13$, giving $a=-75/364,\ b=167/364$. Then $a+b/3=-58/1092\ne0$, so the recurrence predicts a negative $p_3$, not zero; the mixture is not in $(a,b,0)$.)
At a particular required index $k$, a positive-weight component has $p_{k-1,j}=0$ and $p_{k,j}>0$. Can the mixture ratio at $k$ be written as the weighted average of component ratios with weights proportional to $w_jp_{k-1,j}$? (Answer: No; that component ratio is undefined, its proposed ratio weight is zero, but it contributes positively to the mixture numerator.)

What You'll Learn