Conditional Variance for Random Counts and Sums

Why Condition on the Count?

For a random sum, $S=\sum_{j=1}^{N}X_j$, variation can arise in two distinct ways: the number of terms $N$ can vary, and the terms themselves can vary. Conditioning on $N$ separates these sources. If the summands also share a latent influence, or their dependence changes with the count, conditioning on the count alone may not be enough; the conditioning variable must capture the relevant structure.

The central tool is the law of total variance. It gives a general decomposition first; familiar formulas for random sums follow only after the conditional mean, conditional variance, and required moment conditions have been established. This approach is aligned with the SOA Exam P (Probability) syllabus.

Definition: Let $S$ be a random variable with finite second moment, and let $C$ be a conditioning variable. The conditional variance decomposition is $$ \operatorname{Var}(S)=E[\operatorname{Var}(S\mid C)]+\operatorname{Var}(E[S\mid C]). $$ The first term is the expected variation remaining after $C$ is known. The second is the variation in the conditional mean as $C$ changes.

Deriving the Decomposition

Write $m(C)=E[S\mid C]$. Add and subtract this conditional mean:

$$ S-E[S]=(S-m(C))+(m(C)-E[S]). $$

Squaring and taking expectations produces two squared terms and a cross term. The cross term is zero: conditional on $C$, the first factor has mean $E[S-m(C)\mid C]=0$, while $m(C)-E[S]$ is determined by $C$. Therefore,

\begin{aligned} \operatorname{Var}(S) &=E[(S-m(C))^2]+E[(m(C)-E[S])^2]\\ &\quad+2E[(S-m(C))(m(C)-E[S])]\\ &=E[\operatorname{Var}(S\mid C)]+\operatorname{Var}(E[S\mid C]). \end{aligned}

The finite-second-moment condition ensures these variances are defined. In applications, it is important to verify that condition rather than mechanically substituting into a variance formula.

Independent Summands and a Random Count

Suppose $N$ is a nonnegative integer-valued random variable, and $X_1,X_2,\ldots$ are independent and identically distributed summands, independent of $N$. Let $E[X_j]=\mu$ and $\operatorname{Var}(X_j)=\sigma^2<\infty$, and define $S=0$ when $N=0$. A sufficient condition for the usual random-sum variance formula is $E[N^2]<\infty$. Then $S$ has a finite second moment, and conditioning on $N=n$ gives

E[S\mid N=n]=n\mu,\qquad \operatorname{Var}(S\mid N=n)=n\sigma^2.

The conditional mean follows by adding $n$ means. The conditional variance is the sum of $n$ variances because, given the count, the summands are independent. Applying the decomposition with $C=N$ yields

\begin{aligned} \operatorname{Var}(S) &=E[N\sigma^2]+\operatorname{Var}(N\mu)\\ &=\sigma^2 E[N]+\mu^2\operatorname{Var}(N). \end{aligned}

The condition $E[N^2]<\infty$ is not a technical detail to ignore. Finite mean and variance for each summand do not by themselves guarantee that $S$ has a finite variance when the count can be very large. For example, a count with infinite mean and bounded Bernoulli summands can produce a sum with infinite mean, so the ordinary variance decomposition cannot be used as if all terms were finite.

Worked Example: An Independent Random Sum

A hypothetical device can produce $N\sim\operatorname{Binomial}(3,0.5)$ signals. Independently, each signal contributes $1$ or $3$ units with equal probability, independently of the other contributions. Find the mean and variance of the total $S$.

Method and assumptions. The count is independent of an iid sequence of summands. The count is bounded, so $E[N^2]<\infty$; the summands have finite variance. Thus the independent random-sum formula applies.

For one contribution,

\mu=\frac{1+3}{2}=2,\qquad \sigma^2=\frac{(1-2)^2+(3-2)^2}{2}=1.

For the count,

E[N]=3(0.5)=1.5,\qquad \operatorname{Var}(N)=3(0.5)(0.5)=0.75.

The expected total is $E[S]=\mu E[N]=2(1.5)=3$. Its variance is

\operatorname{Var}(S) =(1)(1.5)+(2^2)(0.75) =1.5+3 =4.5.

Check. The within-count contribution is $1.5$, and the variation in the conditional mean $2N$ contributes $4\operatorname{Var}(N)=3$. These add to $4.5$. The common mistake to avoid is omitting the second term and reporting only the expected conditional variance, $1.5$.

When the Summands Depend on One Another

If the first $n$ terms have conditional mean $\mu$, conditional variance $\sigma^2$, and common pairwise conditional covariance $c$ given $N=n$, then

E[S\mid N=n]=n\mu,\qquad \operatorname{Var}(S\mid N=n)=n\sigma^2+n(n-1)c.

There are $n$ individual variance terms and $n(n-1)$ ordered cross-covariance terms in the variance of a sum. If these conditional moments are the same for every $n$, and $E[N^2]<\infty$, the decomposition gives

\operatorname{Var}(S) =\sigma^2E[N]+cE[N(N-1)]+\mu^2\operatorname{Var}(N).

This formula requires the stated conditional means, variances, and covariances; it is not a general formula for arbitrary dependent summands. If these quantities vary with $n$, use $E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])$ directly.

Worked Example: A Shared Latent Probability

Let $N\sim\operatorname{Binomial}(2,0.5)$. Independently of $N$, a latent state $Z$ equals $0$ or $1$ with probability $0.5$ each. Given $Z$, the summands are conditionally iid Bernoulli, with success probability $q_0=0.25$ if $Z=0$ and $q_1=0.75$ if $Z=1$. Find $\operatorname{Var}(S)$.

Method and assumptions. Condition first on both $N$ and $Z$. Given these, the included Bernoulli terms are independent. The count is bounded, so all required second moments are finite.

The average success probability and the average conditional Bernoulli variance are

E[q_Z]=\frac{0.25+0.75}{2}=0.5,\qquad E[q_Z(1-q_Z)]=\frac{0.25(0.75)+0.75(0.25)}{2}=0.1875=\frac{3}{16}.

Also,

\operatorname{Var}(q_Z) =\frac{(0.25-0.5)^2+(0.75-0.5)^2}{2} =0.0625=\frac{1}{16}.

Given $N=n$, conditioning on $Z$ gives conditional mean $nq_Z$ and conditional variance $nq_Z(1-q_Z)$. Applying the decomposition over $Z$,

E[S\mid N=n]=\frac{n}{2},\qquad \operatorname{Var}(S\mid N=n) =\frac{3n}{16}+\frac{n^2}{16}.

For the count, $E[N]=1$, $\operatorname{Var}(N)=0.5$, and $E[N^2]=1.5$. Therefore,

\begin{aligned} E[\operatorname{Var}(S\mid N)] &=\frac{3E[N]}{16}+\frac{E[N^2]}{16} =\frac{3}{16}+\frac{1.5}{16} =\frac{9}{32},\\ \operatorname{Var}(E[S\mid N]) &=\operatorname{Var}(N/2) =\frac{0.5}{4} =\frac{1}{8}=\frac{4}{32},\\ \operatorname{Var}(S)&=\frac{9}{32}+\frac{4}{32} =\frac{13}{32}=0.40625. \end{aligned}

Check and trap. A single summand has marginal mean $0.5$ and variance $0.25$. Treating the summands as independent would give $0.25(1)+0.25(0.5)=0.375$, which is too small. The shared $Z$ makes two summands positively dependent: their covariance is $\operatorname{Var}(q_Z)=1/16$. The expected pair contribution is $E[N(N-1)]/16=0.5/16=1/32$, exactly the missing amount. The correct variance is $0.375+0.03125=0.40625$.

When the Count and Summand Distribution Are Related

A random count need not be independent of the mechanism generating the summands. A useful model has a latent variable $Z$ such that, given $Z$ and $N$, the included terms are independent with conditional mean $\mu_Z$ and conditional variance $\sigma_Z^2$. Provided the relevant second moments are finite,

E[S\mid N,Z]=N\mu_Z,\qquad \operatorname{Var}(S\mid N,Z)=N\sigma_Z^2,

and hence

\operatorname{Var}(S)=E[N\sigma_Z^2]+\operatorname{Var}(N\mu_Z).

This form does not assume that $N$ and $Z$ are independent. If they are dependent, the expectation and variance must retain their joint relationship. Conditioning on $Z$ can be especially useful when it explains both the count and the summand distribution.

Worked Example: A Count Linked to the Environment

A hypothetical process has a latent state $Z$, with $P(Z=0)=P(Z=1)=0.5$. If $Z=0$, then $N=1$ and each included term is Bernoulli with success probability $0.25$. If $Z=1$, then $N=3$ and each included term is Bernoulli with success probability $0.75$. Given the state, the terms are independent. Find the mean and variance of $S$.

Method and assumptions. Condition on $Z$, which determines both the count and the conditional summand distribution. The count is at most three, so the sum has a finite second moment.

When $Z=0$, the conditional mean and variance are $1(0.25)=0.25$ and $1(0.25)(0.75)=3/16$. When $Z=1$, they are $3(0.75)=2.25$ and $3(0.75)(0.25)=9/16$. Thus,

E[S]=\frac{0.25+2.25}{2}=1.25=\frac{5}{4}, \qquad E[\operatorname{Var}(S\mid Z)] =\frac{3/16+9/16}{2} =\frac{3}{8}.

The two conditional means are $0.25$ and $2.25$. Their average is $1.25$, and each is one unit from that average, so

\operatorname{Var}(E[S\mid Z]) =\frac{(0.25-1.25)^2+(2.25-1.25)^2}{2}=1.

Therefore,

\operatorname{Var}(S)=\frac{3}{8}+1=\frac{11}{8}=1.375.

Check. The conditional second moments are $3/16+(1/4)^2=1/4$ in state $0$, and $9/16+(9/4)^2=45/8$ in state $1$. Their average is $47/16$; subtracting $(E[S])^2=(5/4)^2=25/16$ gives $22/16=11/8$. The trap is to use an unconditional summand mean and variance as though the count were independent of the summand mechanism; here the state links them.

Choosing and Checking the Conditioning Variable

For a random sum, conditioning on $N$ is effective when it makes the conditional mean and variance manageable. If summands share an environmental effect, condition on that effect as well, or explicitly include the resulting conditional covariance. A reliable calculation follows these steps:

State the model.
Define the count, summands, and any latent variable. Specify which quantities are independent or independent only after conditioning.

Verify moments.
Check that the sum has a finite second moment. For an independent random count with iid finite-variance summands, $E[N^2]<\infty$ is a sufficient count condition.

Find conditional moments.
Compute $E[S\mid C]$ and $\operatorname{Var}(S\mid C)$, including any cross-covariance terms that remain after conditioning.

Decompose and check.
Add the expected conditional variance and the variance of the conditional mean. Check whether a shared source of variation has been omitted.

The main exam trap is confusing “identically distributed” with “independent.” Identical marginal distributions do not eliminate covariance. Another is using a correct-looking formula without the assumptions that justify it. The safest practice is to write the conditional mean and variance before substituting into the total-variance identity.

Check Your Understanding

For each problem, identify the relevant conditioning structure before calculating.

$N\sim\operatorname{Binomial}(4,0.25)$ is independent of iid Bernoulli$(0.5)$ summands. Find the mean and variance of their random sum. (Answer: mean $0.5$, variance $7/16=0.4375$)
Given $N=4$, four summands each have conditional mean $3$, conditional variance $2$, and every pair has conditional covariance $0.5$. Find the conditional variance of their sum. (Answer: $14$)
A fixed count of $2$ terms shares a latent state: $Z=0$ or $1$ with equal probabilities, and given $Z$ the terms are iid Bernoulli with success probability $0.2$ or $0.8$, respectively. Find the variance of the sum. (Answer: $0.68$)
$N\sim\operatorname{Poisson}(2)$ is independent of iid summands with mean $1$ and variance $4$. Find the variance of the random sum. (Answer: $10$)
$Z=0$ or $1$ with equal probabilities. If $Z=0$, $N=1$ and the summand success probability is $0.2$; if $Z=1$, $N=2$ and the success probability is $0.6$. Given $Z$, the summands are independent Bernoulli variables. Find the variance of the random sum. (Answer: $0.57$)

What You'll Learn