Choosing a Method for a Sum Distribution
To find the distribution of \(S=X+Y\), first inspect the structure of the problem rather than immediately expanding a sum or integral. The useful questions are: Are \(X\) and \(Y\) independent? What values can they take? Does conditioning simplify their relationship? And do their distributions have a recognizable identity, such as a product of moment generating functions that matches a familiar distribution?
These questions often point to a more direct calculation. Convolution is a natural choice when the variables are independent and their supports give manageable bounds. Conditioning is useful when a variable or latent state makes the conditional distributions simpler. Moment generating functions (MGFs) can be efficient when the variables are independent and the product of their MGFs has a known form. Each method has conditions; a shorter calculation is not useful if those conditions are not satisfied.
A Practical Decision Rule
Use the following order of checks. Independence is important: ordinary convolution and multiplication of MGFs both require it. If independence is absent, ask whether conditioning on a shared state makes the summands independent, or otherwise gives a tractable conditional distribution. Then inspect the supports. Discrete supports may reduce a sum to a finite calculation, while continuous supports determine the integration bounds. Finally, check whether a familiar distribution identity avoids a longer convolution.
Establish whether the summands are independent. If only conditional independence is given, do not use an unconditional convolution or product of MGFs without conditioning.
List the possible values of each summand. For a target \(s\), only values or integration points that can satisfy \(X+Y=s\) contribute.
If independent summands have MGFs whose product is a recognized MGF, identify the sum directly. Otherwise, convolution or conditioning may be clearer.
Confirm that the result respects the support, has total probability one when it is a full distribution, and uses the correct independence assumptions.
For independent variables, the distribution function can be written as \(F_S(s)=P(X+Y\le s)=\int F_Y(s-x)\,dF_X(x)\). If \(X\) is discrete with probabilities \(p_x\), this becomes \(F_S(s)=\sum_x p_x F_Y(s-x)\). These expressions are useful when one summand is discrete or when the supports make the contributing terms easy to identify.
Worked Examples: Match the Method to the Structure
Worked Example: A Discrete and Continuous Sum
A hypothetical model has independent \(X\) and \(Y\). The discrete variable \(X\) has \(P(X=0)=0.2\), \(P(X=1)=0.5\), and \(P(X=2)=0.3\). The continuous variable \(Y\) is uniform on \((0,1)\). Find \(P(0.4<X+Y\le1.5)\).
Method: convolution through the distribution function. Independence is stated, and \(X\) has only three possible values, so the finite-sum form of the convolution is convenient. For each value \(x\), the contribution to \(F_S(s)\) is \(P(X=x)F_Y(s-x)\).
First calculate \(F_S(1.5)\). If \(X=0\), then \(Y\le1.5\) always, so the contribution is \(0.2(1)=0.2\). If \(X=1\), then \(Y\le0.5\), which has probability \(0.5\), so the contribution is \(0.5(0.5)=0.25\). If \(X=2\), then \(Y\le-0.5\), which has probability zero. Therefore,
Next calculate \(F_S(0.4)\). Only \(X=0\) can contribute: it requires \(Y\le0.4\), with probability \(0.4\). The cases \(X=1\) and \(X=2\) would require \(Y\le-0.6\) and \(Y\le-1.6\), respectively, both impossible. Thus \(F_S(0.4)=0.2(0.4)=0.08\). Since \(S\) has no point mass at either endpoint,
Check: The result lies between 0 and 1, and the support of \(S\) is contained in \((0,3)\), consistent with the interval used. The common mistake avoided is adding the probabilities of \(X\) values without accounting for the required range of \(Y\) for each one.
Worked Example: Conditioning on a Shared Latent State
A hypothetical count model has a latent state \(H\), with \(P(H=0)=0.6\) and \(P(H=1)=0.4\). Given \(H=0\), \(X\) and \(Y\) are independent Poisson variables, each with mean 1. Given \(H=1\), they are independent Poisson variables, each with mean 3. Find \(P(X+Y=2)\).
Method: condition on \(H\). The summands are independent given \(H\), not necessarily unconditionally. For independent Poisson variables, the sum is Poisson with mean equal to the sum of the means. Therefore, given \(H=0\), \(X+Y\) is Poisson with mean \(2\); given \(H=1\), it is Poisson with mean \(6\).
The law of total probability gives \(P(X+Y=2)=\sum_{h=0}^1 P(H=h)P(X+Y=2\mid H=h)\). For a Poisson variable with mean \(\lambda\), the probability of 2 is \(e^{-\lambda}\lambda^2/2!\). Substituting the two conditional means gives
Check: Both conditional probabilities are between 0 and 1, and their weighted average is also between 0 and 1. The common mistake avoided is replacing the two conditional models by one Poisson model using an averaged rate. Conditional independence does not imply unconditional independence: here \(E[X]=E[Y]=0.6(1)+0.4(3)=1.8\), while \(E[XY]=0.6(1)(1)+0.4(3)(3)=4.2\). Thus \(\operatorname{Cov}(X,Y)=4.2-(1.8)(1.8)=0.96\), which is not zero.
Worked Example: Recognizing a Sum Through MGFs
Let \(X\sim\operatorname{Binomial}(2,0.25)\) and \(Y\sim\operatorname{Bernoulli}(0.25)\), independently. Find \(P(X+Y\ge2)\).
Method: multiply MGFs and identify the distribution. A Bernoulli variable with success probability \(p\) has MGF \(1-p+pe^t\). A binomial variable with parameters \(n,p\) has MGF \((1-p+pe^t)^n\). Independence permits multiplication, so
This is the MGF of a \(\operatorname{Binomial}(3,0.25)\) variable. Hence
Check: The identified binomial distribution has support \(\{0,1,2,3\}\), as a sum of a variable supported on \(\{0,1,2\}\) and one supported on \(\{0,1\}\) must. The common mistake avoided is multiplying the MGFs without verifying independence.
Why the Methods Work—and When They Do Not
The convolution formula follows by partitioning the event \(X+Y\le s\) according to the value of \(X\). If \(X=x\), the event becomes \(Y\le s-x\), whose probability is \(F_Y(s-x)\). Averaging this conditional probability over the distribution of \(X\) gives
When \(X\) is discrete, the integral is a sum over its possible values. Independence is essential here because it allows the distribution of \(Y\) in each part of the partition to remain \(F_Y\). If the conditional distribution of \(Y\) changes with \(X\), use \(P(Y\le s-X\mid X)\) instead, or choose another conditioning variable that makes the calculation manageable.
The MGF product rule has a similarly direct justification. Since \(e^{t(X+Y)}=e^{tX}e^{tY}\), independence gives
If the MGFs exist in a neighborhood of zero, they determine the distributions uniquely. Thus, if the product matches a known MGF, it identifies the sum’s distribution. If the relevant MGF does not exist, or the product does not simplify to a recognizable form, this approach may not help; convolution or conditioning can still be available.
Conditioning is justified by the law of total probability: for a discrete state \(H\), \(P(S\in A)=\sum_h P(H=h)P(S\in A\mid H=h)\). For a continuous state, the corresponding sum is an integral. This makes conditioning particularly useful for mixtures. It does not mean the unconditional distribution has the same form as any one conditional distribution.
Common Traps and a Selection Checklist
The method should follow the assumptions and structure, not just the names of the distributions. In particular, an identity for independent summands cannot be used merely because each marginal distribution has a familiar form. A shared latent variable can create dependence even when the variables are independent after conditioning on that state.
- Choose convolution when independence is established and the supports lead to a manageable sum or integral.
- Choose conditioning when a state, event, or one summand makes the conditional model simpler.
- Choose MGFs when the summands are independent, the MGFs exist near zero, and their product has a recognizable form.
- Regardless of method, verify support, endpoint treatment, and whether the final probabilities or density are valid.
This method-selection skill is aligned with the SOA VEE Mathematical Statistics syllabus. In a timed problem, a brief assumption check can prevent a longer calculation based on an invalid independence claim or an incorrect support range.
Check Your Understanding
For each problem, choose a suitable method and work through the key distribution calculation before checking the final answer.
- Independent \(X\) takes values 0 and 2 with probabilities 0.7 and 0.3. Independently, \(Y\) is uniform on \((0,2)\). Find \(P(X+Y\le1)\). (Answer: \(0.35\))
- Independent \(X\sim\operatorname{Poisson}(2)\) and \(Y\sim\operatorname{Poisson}(5)\). Find \(P(X+Y=4)\). (Answer: \(\frac{2401}{24}e^{-7}\approx0.0912\))
- A state \(H\) has \(P(H=0)=0.25\) and \(P(H=1)=0.75\). Given \(H=0\), \(X,Y\) are independent Poisson variables with means 0.5 each; given \(H=1\), they are independent Poisson variables with means 2 each. Find \(P(X+Y=1)\). (Answer: \(0.25e^{-1}+3e^{-4}\approx0.1469\))
- Five independent Bernoulli trials each have success probability 0.4. Use an MGF or distribution identity to find the probability of exactly two successes. (Answer: \(0.3456\))
- Independent gamma variables have shapes 2 and 3, respectively, and common rate 2. Use their MGFs to identify the sum distribution, then find \(P(X+Y\le2)\). (Answer: \(1-\frac{103}{3}e^{-4}\approx0.3712\))