Choosing Between Convolution and Distributional Closure

Why the Method Matters

For independent random variables, there is more than one route to the distribution of their sum. Direct convolution builds the distribution from the component distributions. Moment generating functions (MGFs) turn independence into multiplication and can make a distribution recognizable. Standard-family closure can identify the sum immediately when the variables and parameters meet the required conditions.

The efficient method depends on both the distributions and the question. A short sum of discrete variables may be easiest by convolution, especially when only one or two probabilities are needed. A recognized closure result avoids unnecessary summation. An MGF is useful when its product has a recognizable form, but it does not automatically make an unfamiliar sum a standard distribution. These choices are aligned with the SOA Exam P (Probability) syllabus.

Definition: If \(X\) and \(Y\) are independent, the distribution of \(S=X+Y\) can be obtained by convolution. For discrete variables, \(P(S=s)=\sum_x P(X=x)P(Y=s-x)\). For independent continuous variables with densities \(f_X\) and \(f_Y\), \(f_S(s)=\int_{-\infty}^{\infty} f_X(x)f_Y(s-x)\,dx\). In either formula, independence is essential, and terms outside the variables' supports contribute zero.

A Decision Process

Before calculating, state what is known about the components and what is being asked about the sum. Independence permits the product rule for MGFs and the convolution formulas, but it does not by itself imply that the sum belongs to a familiar family. Check any proposed closure result against its exact conditions, including the parameters.

Check for direct closure.
If the components belong to a family with a known sum rule, verify its conditions and use that rule. This is usually the shortest route.

Consider the MGF product.
When the MGFs exist near zero, multiply them. If the product matches a recognized MGF, identify the distribution by its parameters.

Use convolution when needed.
If no useful closure is available, or a particular probability is simpler to obtain directly, set up the sum over the valid support or integrate over the valid region.

Check the result.
Confirm the support, total probability where relevant, parameterization, and whether the numerical answer is within its possible range.

Why the MGF Product Rule Works

The MGF of \(X\), when finite for \(t\) in an interval containing zero, is \(M_X(t)=E[e^{tX}]\). If \(X\) and \(Y\) are independent and the required expectations are finite, then the MGF of their sum factors:

M_{X+Y}(t) =E[e^{t(X+Y)}] =E[e^{tX}e^{tY}] =E[e^{tX}]E[e^{tY}] =M_X(t)M_Y(t).

The factorization follows because \(e^{tX}\) and \(e^{tY}\) are functions of independent variables. If the product matches the MGF of a known distribution on an interval around zero, the sum has that distribution. The MGF method is not available merely because a formula can be written: the MGFs must exist around zero, and a match must be made with a properly parameterized known distribution.

Worked Example: Convolving Exponentials With Different Rates

Let \(X\) and \(Y\) be independent exponential random variables with rates 1 and 2, respectively. Find \(P(X+Y\leq 1)\). Use convolution, and explain why equal-rate exponential closure does not apply.

Method and supports. The rates differ, so the equal-rate exponential-sum result does not apply. Direct convolution is straightforward. For \(s\geq 0\), the density of \(S=X+Y\) is

f_S(s)=\int_0^s e^{-x}\,2e^{-2(s-x)}\,dx.

The integration limits are \(0\) and \(s\) because both exponential variables are nonnegative and \(x+(s-x)=s\). Simplifying the integrand gives

f_S(s)=2e^{-2s}\int_0^s e^x\,dx =2e^{-2s}(e^s-1) =2(e^{-s}-e^{-2s}).

Integrate this density from 0 to 1:

\begin{aligned} P(S\leq 1) &=\int_0^1 2(e^{-s}-e^{-2s})\,ds\\ &=2(1-e^{-1})-(1-e^{-2})\\ &=1-2e^{-1}+e^{-2}\\ &\approx 1-2(0.3678794)+0.1353353\\ &\approx 0.3996. \end{aligned}

The probability is between 0 and 1, as it must be. A direct tail check gives \(P(S>1)=2e^{-1}-e^{-2}\approx 0.6004\); the two probabilities add to 1. This check also shows why a verbal comparison with the mean alone is not a valid way to decide whether a probability is above or below one-half.

Common mistake avoided: applying the equal-rate result to unequal rates. The rates here are 1 and 2, so the shape and rate parameters of an equal-rate gamma sum cannot be assigned to this sum by that rule.

Worked Example: Using MGFs to Identify a Gamma Sum

Let \(X\) and \(Y\) be independent, with \(X\sim\operatorname{Gamma}(2,\text{ rate }3)\) and \(Y\sim\operatorname{Gamma}(3,\text{ rate }3)\). Find \(P(X+Y>1)\). Use the MGF product to identify the sum, then evaluate the probability.

Parameter convention. Here a gamma distribution with shape \(\alpha\) and rate \(\lambda\) has MGF \((\lambda/(\lambda-t))^\alpha\) for \(t<\lambda\). Thus

M_X(t)=\left(\frac{3}{3-t}\right)^2,\qquad M_Y(t)=\left(\frac{3}{3-t}\right)^3,\qquad t<3.

By independence, the sum \(S=X+Y\) has MGF

M_S(t)=M_X(t)M_Y(t) =\left(\frac{3}{3-t}\right)^5,\qquad t<3.

This is the MGF of a gamma distribution with shape 5 and rate 3. Therefore \(S\sim\operatorname{Gamma}(5,\text{ rate }3)\). For a positive integer shape \(k\), a gamma variable with shape \(k\) and rate \(\lambda\) has survival probability \(P(S>x)=e^{-\lambda x}\sum_{j=0}^{k-1}(\lambda x)^j/j!\). At \(x=1\), \(k=5\), and \(\lambda=3\):

\begin{aligned} P(S>1) &=e^{-3}\left(1+3+\frac{3^2}{2!}+\frac{3^3}{3!}+\frac{3^4}{4!}\right)\\ &=e^{-3}(1+3+4.5+4.5+3.375)\\ &=16.375e^{-3}\\ &\approx 0.8153. \end{aligned}

The result is a probability in \([0,1]\). The shared rate of 3 is essential: the product of the MGFs has a single denominator \(3-t\), with the exponents adding to 5. Common mistake avoided: adding gamma rates or overlooking a rate mismatch. Here the shapes add; the common rate stays 3.

Worked Example: Applying Poisson Closure

A hypothetical monitoring system records two independent types of alert during one shift. Let their counts be \(N_1\sim\operatorname{Poisson}(1.2)\) and \(N_2\sim\operatorname{Poisson}(0.8)\). Find the probability that the total count is at most 1.

Method. Independent Poisson variables have a Poisson sum, with parameter equal to the sum of their parameters. Therefore \(N_1+N_2\sim\operatorname{Poisson}(1.2+0.8)=\operatorname{Poisson}(2)\). This closure is more efficient than summing the component probabilities for every way the total could be 0 or 1.

\begin{aligned} P(N_1+N_2\leq 1) &=P(N_1+N_2=0)+P(N_1+N_2=1)\\ &=e^{-2}\frac{2^0}{0!}+e^{-2}\frac{2^1}{1!}\\ &=3e^{-2}\\ &\approx 0.4060. \end{aligned}

The parameter is a mean count over the same shift for each component; the sum is not obtained by adding event probabilities. Common mistake avoided: assuming the total is Poisson without checking that the component counts are independent Poisson variables. Both conditions are given here.

When Convolution Is the Better Choice

Not every independent sum has a convenient standard distribution. In that case, convolution is not a fallback to avoid; it is the defining calculation. It can also be preferable when the sum has a small discrete support or the question asks for a specific event that requires only a few terms.

For example, if \(N\sim\operatorname{Poisson}(1)\) and \(B\sim\operatorname{Bernoulli}(0.4)\) are independent, \(N+B\) is not generally Poisson. Its MGF is \(e^{e^t-1}(0.6+0.4e^t)\), which does not have the Poisson form \(e^{\lambda(e^t-1)}\). For the event \(N+B\leq 1\), convolution needs only the cases \((N,B)=(0,0),(1,0),(0,1)\):

\begin{aligned} P(N+B\leq 1) &=P(N=0)P(B=0)+P(N=1)P(B=0)+P(N=0)P(B=1)\\ &=e^{-1}(0.6)+e^{-1}(0.6)+e^{-1}(0.4)\\ &=1.6e^{-1}\approx 0.5886. \end{aligned}

The terms are multiplied because of independence, and the listed cases are exactly those for which the sum is at most 1. An MGF can still describe this sum, but convolution is the direct route to this particular probability.

Choosing Carefully Under Exam Conditions

A useful comparison is not “which formula do I remember?” but “which method answers this question with the fewest justified steps?” Check the variable supports before writing convolution limits. Check whether rates or probabilities really match a closure condition. If using an MGF, state its parameter domain and identify the distribution by matching the complete functional form, not just one feature such as the mean.

Use closure when the family and parameter conditions match exactly. State the resulting distribution and its parameters before calculating a probability.
Use an MGF product when independence makes the product manageable and the result can be matched to a known MGF. It can also expose why a proposed closure does not hold.
Use convolution when no suitable standard form is established or only a few support combinations contribute to the requested event.

A common exam trap is to treat these methods as interchangeable without checking assumptions. Convolution and the MGF product rule require independence in the forms used here. A named closure rule has its own parameter conditions. If those conditions fail, either derive the sum's distribution by a valid method or leave it in a nonstandard form; do not force it into a familiar family.

Key takeaway: Identify the sum's distribution only after checking independence, support, and the exact closure conditions. Use the shortest valid method, and verify the resulting probability or distribution before moving on.

Check Your Understanding

For each problem, choose a valid method and check the assumptions before calculating.

Independent counts \(N_1\sim\operatorname{Poisson}(0.7)\) and \(N_2\sim\operatorname{Poisson}(1.3)\) occur during a fixed interval. Find \(P(N_1+N_2=2)\). (Answer: \(2e^{-2}\approx 0.2707\))
Independent variables \(X,Y\) are exponential with common rate 2. Find \(P(X+Y>1)\). (Answer: \(3e^{-2}\approx 0.4060\))
Three independent Bernoulli variables have success probabilities 0.2, 0.5, and 0.7. Find the probability their sum equals 1. (Answer: \(0.4300\))
Let \(N\sim\operatorname{Poisson}(1)\) and \(B\sim\operatorname{Bernoulli}(0.25)\) be independent. Find \(P(N+B\geq 2)\). (Answer: \(1-\frac{7}{4e}\approx 0.3562\))
Independent \(X\sim N(4,1)\) and \(Y\sim N(1,4)\), where the second parameter is the variance. Find \(P(X+Y>8)\). (Answer: \(1-\Phi(3/\sqrt{5})\approx 0.0899\))

What You'll Learn