Recognizing Mixed Independent Sums by Their MGFs

Why Mixed Sums Require an MGF Check

When independent random variables come from different distribution families, their sum does not automatically belong to a familiar family. The moment generating function (MGF) gives a systematic way to investigate: multiply the component MGFs, simplify the product, and compare it with known MGF forms. Sometimes the product reveals a standard distribution. In other cases, it still identifies the sum’s distribution uniquely even though it does not match a familiar named form.

The method depends on independence and on the existence of the relevant MGFs near zero. The key task is not to force every product into a standard family, but to state exactly what the MGF establishes.

Definition: The MGF of a random variable \(X\) is \(M_X(t)=E[e^{tX}]\) for values of \(t\) for which this expectation is finite. If \(X\) and \(Y\) are independent and their MGFs exist on an open interval containing zero, then \(M_{X+Y}(t)=M_X(t)M_Y(t)\) wherever both factors are finite. If this product matches the MGF of a known distribution on an open interval containing zero, the uniqueness theorem for MGFs identifies the distribution of \(X+Y\).

A Practical Recognition Procedure

For a sum \(S=X+Y\), first record the distribution and parameterization of each component. For instance, an exponential distribution is often specified by its rate \(\lambda\), for which the mean is \(1/\lambda\); this differs from specifying its scale. Then write the component MGFs, state their domains, and multiply them. Finally, compare the simplified expression with standard MGF forms. The product must agree on an interval around zero, not merely at a single value of \(t\).

State the assumptions.
Name the component distributions and parameters, and confirm that the variables are independent.

Write each MGF and its domain.
The product is usable only where both component MGFs are finite.

Multiply and simplify.
Look for a standard MGF form, keeping its parameters and domain visible.

Conclude carefully.
If the product matches a known MGF, identify the distribution. If not, report the product as the identified MGF rather than assigning an unsupported name.

Useful forms include \(M_X(t)=(1-p+pe^t)^n\) for \(X\sim\operatorname{Binomial}(n,p)\), \(M_X(t)=\exp\{\lambda(e^t-1)\}\) for \(X\sim\operatorname{Poisson}(\lambda)\), \(M_X(t)=\lambda/(\lambda-t)\) for an exponential variable with rate \(\lambda\), and \(M_X(t)=(\lambda/(\lambda-t))^\alpha\) for a gamma variable with shape \(\alpha\) and rate \(\lambda\). The exponential form is the gamma form with shape one.

Worked Example: A Uniform Loss Component Plus an Exponential Component

Suppose a hypothetical model has independent variables \(C\sim\operatorname{Uniform}(1,3)\) and \(D\sim\operatorname{Exponential}(\text{rate }2)\). Find the MGF of \(S=C+D\), determine its mean and variance, and decide whether the product identifies a familiar standard family.

Method and assumptions. The variables are independent, so their MGFs multiply. For the uniform variable, direct integration gives

M_C(t)=\frac{1}{2}\int_1^3 e^{tx}\,dx =\frac{e^{3t}-e^t}{2t},\qquad t\ne0,

with \(M_C(0)=1\) by continuity. For the rate-2 exponential variable,

M_D(t)=\frac{2}{2-t},\qquad t<2.

Thus the product, finite for \(t<2\), is

M_S(t)=\frac{e^{3t}-e^t}{2t}\frac{2}{2-t},\qquad t\ne0,\ t<2, \quad\text{and}\quad M_S(0)=1.

This product does not have the standard form of a single uniform, exponential, gamma, or normal MGF. It nevertheless identifies the distribution of \(S\), because it is an MGF on an open interval containing zero and the MGF uniqueness theorem applies.

Mean and variance. The uniform mean is \((1+3)/2=2\), and its variance is \((3-1)^2/12=1/3\). The rate-2 exponential mean is \(1/2\), and its variance is \(1/2^2=1/4\). Independence gives

E[S]=2+\frac12=\frac52=2.5,\qquad \operatorname{Var}(S)=\frac13+\frac14=\frac{4+3}{12}=\frac7{12}.

Check and common mistake avoided. The distribution of \(C\) is continuous over the interval from 1 to 3, so its mean is the midpoint, 2, not the upper endpoint, 3. The sum’s mean is therefore \(2.5\), not \(3.5\) or \(3\). The variance check is \(1/3+1/4=7/12\), as above.

When the Product Does Not Match a Standard Form

A product of valid MGFs is still useful even when it does not simplify to a familiar family. In that case, report the product, its domain, and any requested moments. Do not infer a named distribution just because the support or first two moments look familiar: many distributions can share a mean and variance.

For independent variables with finite variances, additivity gives \(E[X+Y]=E[X]+E[Y]\) and \(\operatorname{Var}(X+Y)=\operatorname{Var}(X)+\operatorname{Var}(Y)\). These component moments are often quicker to use than differentiating a complicated product. The MGF still provides the distributional identification when it exists near zero.

Worked Example: A Binomial Component Plus a Poisson Component

Let \(B\sim\operatorname{Binomial}(3,0.25)\) and \(N\sim\operatorname{Poisson}(2)\) be independent. Find the MGF of \(T=B+N\), calculate its mean and variance, and state what distributional conclusion is justified.

Method and assumptions. Use the binomial and Poisson MGFs and multiply them because \(B\) and \(N\) are independent:

M_T(t)=M_B(t)M_N(t) =(0.75+0.25e^t)^3\exp\{2(e^t-1)\}.

Both factors are finite for every real \(t\), so the product MGF exists for every real \(t\). It does not simplify to the MGF of a single binomial or Poisson variable: it retains both a finite-power factor and an exponential factor. The product identifies the distribution of \(T\), but this calculation does not justify labeling it as one of those standard families.

Moments. The binomial component has mean \(3(0.25)=0.75\) and variance \(3(0.25)(0.75)=0.5625=9/16\). The Poisson component has mean and variance both equal to 2. Therefore

E[T]=0.75+2=2.75=\frac{11}{4},\qquad \operatorname{Var}(T)=\frac9{16}+2=\frac{9+32}{16}=\frac{41}{16}.

Check and common mistake avoided. The sum takes nonnegative integer values, consistent with both components. But matching that support does not make the sum Poisson. In particular, a Poisson variable has equal mean and variance, whereas here \(11/4\ne41/16\). The unequal moments rule out a Poisson identification.

When Different Parameterized Forms Combine

Sometimes components described using different names nevertheless have MGFs that combine into a standard form. For example, an exponential variable is a gamma variable with shape one. The rate parameters matter: equal rates allow the powers in the gamma MGFs to combine, while different rates generally leave distinct factors.

Worked Example: Gamma Plus Exponential With a Shared Rate

Let \(G\sim\operatorname{Gamma}(\text{shape }3,\text{ rate }2)\) and \(E\sim\operatorname{Exponential}(\text{rate }2)\) be independent. Identify the distribution of \(R=G+E\), and verify the result using its moments.

Method and assumptions. For a gamma variable with shape \(\alpha\) and rate \(\lambda\), the MGF is \((\lambda/(\lambda-t))^\alpha\) for \(t<\lambda\). An exponential variable with rate 2 has MGF \(2/(2-t)\) on the same domain. Hence

M_R(t)=\left(\frac{2}{2-t}\right)^3\left(\frac{2}{2-t}\right) =\left(\frac{2}{2-t}\right)^4,\qquad t<2.

This is the MGF of a gamma variable with shape 4 and rate 2. Since the MGFs agree on an interval containing zero, \(R\sim\operatorname{Gamma}(\text{shape }4,\text{ rate }2)\).

Moment check. The component means are \(3/2\) and \(1/2\), so \(E[R]=2\). Their variances are \(3/2^2=3/4\) and \(1/2^2=1/4\), so \(\operatorname{Var}(R)=1\). The identified gamma distribution gives mean \(4/2=2\) and variance \(4/2^2=1\), agreeing with both checks.

Common mistake avoided. The common rate is essential to combining these powers. If the exponential rate differed, the product would have factors with different denominators and would not equal this gamma MGF.

Why the Product Rule Holds

The product rule follows directly from the definition of an MGF and the meaning of independence. For values of \(t\) where the expectations are finite, independence makes the joint distribution factor into its marginal distributions. Since \(e^{t(X+Y)}=e^{tX}e^{tY}\), the expectation of this product factors.

\begin{aligned} 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). \end{aligned}

The factorization on the third line uses independence; it is not a general rule for dependent variables. The product must also be considered only on the intersection of the component MGFs’ domains. For instance, a bounded uniform variable has an MGF for every real \(t\), whereas a rate-\(\lambda\) exponential variable has an MGF only for \(t<\lambda\). Their sum’s product is therefore restricted to \(t<\lambda\).

Common Traps and a Final Check

One tempting but incorrect calculation is to add MGFs when adding independent variables. For the uniform-plus-exponential example, that error would give \(M_C(t)+M_D(t)\). At \(t=0\), it equals \(1+1=2\), but every MGF must satisfy \(M(0)=E[1]=1\). The correction is to multiply: \(M_{C+D}(t)=M_C(t)M_D(t)\).

Do not omit independence. The product rule applies to independent variables.
Keep rate and scale distinct. For an exponential rate \(\lambda\), the mean is \(1/\lambda\); its MGF is \(\lambda/(\lambda-t)\).
Check the domain. A product is valid only where every factor is finite, and identification requires agreement near zero.
Do not overstate a match. If no standard MGF form results, give the product as the identified MGF and calculate moments separately if needed.
Use moments as a check, not as proof of a distribution. Agreement of means and variances alone does not establish that two distributions are identical.

Key takeaway: For an independent mixed sum, multiply the component MGFs, retain the valid domain, and compare the product with known forms. A match identifies a familiar distribution; no match still leaves a precise, unique MGF description when the MGF exists near zero.

Check Your Understanding

For each independent sum, write or identify its MGF and use component moments where requested. Answers are provided for checking.

Let \(X\sim\operatorname{Bernoulli}(0.4)\) and \(Y\sim\operatorname{Exponential}(\text{rate }4)\) be independent. Give the MGF of \(X+Y\), its domain, mean, and variance. (Answer: \(M(t)=(0.6+0.4e^t)\frac{4}{4-t}\) for \(t<4\); \(E[X+Y]=13/20=0.65\); \(\operatorname{Var}(X+Y)=121/400=0.3025\).)
Let \(B\sim\operatorname{Binomial}(4,0.5)\) and \(N\sim\operatorname{Poisson}(1)\) be independent. Give the MGF, mean, and variance, and say whether the product is the MGF of a Poisson variable. (Answer: \(M(t)=(0.5+0.5e^t)^4\exp(e^t-1)\); mean \(3\), variance \(2\); it is not Poisson, since a Poisson variable has equal mean and variance.)
Let \(U\sim\operatorname{Uniform}(0,2)\) and \(E\sim\operatorname{Exponential}(\text{rate }1)\) be independent. Give the MGF of \(U+E\), its domain, mean, and variance. (Answer: \(M(t)=\frac{e^{2t}-1}{2t}\frac{1}{1-t}\) for \(t\ne0,\ t<1\), with \(M(0)=1\); mean \(2\); variance \(4/3\).)
Let \(G\sim\operatorname{Gamma}(\text{shape }5,\text{ rate }3)\) and \(E\sim\operatorname{Exponential}(\text{rate }2)\) be independent. Give the MGF and calculate the mean and variance. Does the product have the standard gamma MGF form? (Answer: \(M(t)=(\frac{3}{3-t})^5\frac{2}{2-t}\) for \(t<2\); mean \(13/6\); variance \(29/36\); no, the different rate denominators prevent it from having the standard gamma MGF form.)
Let \(G\sim\operatorname{Gamma}(\text{shape }2,\text{ rate }5)\) and \(E\sim\operatorname{Exponential}(\text{rate }5)\) be independent. Identify the distribution of \(G+E\), and verify its mean and variance. (Answer: \(\operatorname{Gamma}(\text{shape }3,\text{ rate }5)\); mean \(3/5\); variance \(3/25\).)

What You'll Learn