Why Transform Calculations Need Checks
Generating functions and transforms turn distributional questions into algebra, but the algebra is valid only when its assumptions hold. A power series can diverge at the argument where it is evaluated; an MGF can be finite at zero but fail to exist on any open interval around zero; and a product formula for a sum can fail when the summands are dependent. Similarly, a characteristic function exists for every random variable, but differentiating its expectation requires moment conditions.
Throughout, let \(X\) and \(Y\) be real-valued random variables on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). When \(X\) is nonnegative and integer-valued, its probability generating function (PGF) is \(G_X(s)=\mathbb{E}[s^X]\) for arguments where this expectation is defined. The moment generating function (MGF) is \(M_X(t)=\mathbb{E}[e^{tX}]\) when finite, and the characteristic function (CF) is \(\phi_X(t)=\mathbb{E}[e^{itX}]\) for real \(t\). Since \(|e^{itX}|=1\), the CF is always defined. These definitions do not make every formal manipulation valid: each calculation must respect its domain and hypotheses.
PGF Arguments Must Stay in the Convergence Domain
For any nonnegative integer-valued \(X\), the PGF series converges absolutely when \(|s|\le 1\), because \(\sum_k\mathbb{P}(X=k)|s|^k\le\sum_k\mathbb{P}(X=k)=1\). For larger arguments, convergence depends on the distribution. A formula obtained by summing a power series is valid only where that series converges. Substituting into an algebraic expression outside that domain may give a number, but it does not thereby give the expectation.
Worked Example: A Geometric PGF Beyond Its Domain
Let \(X\) have probabilities \(\mathbb{P}(X=k)=2^{-(k+1)}\), \(k=0,1,2,\ldots\). These probabilities sum to \(\sum_{k=0}^{\infty}2^{-(k+1)}=1\). Find \(G_X(s)\) where the series converges, and assess the proposed value at \(s=3\).
For real \(s\ge0\), the defining series is \(\sum_{k=0}^{\infty}2^{-(k+1)}s^k=\tfrac12\sum_{k=0}^{\infty}(s/2)^k\). The geometric-series formula applies when \(s/2<1\), so for \(0\le s<2\), \(G_X(s)=\frac{1/2}{1-s/2}=\frac{1}{2-s}\). At \(s=1\), this gives \(G_X(1)=1\), as required.
At \(s=\tfrac32\), which is inside the convergence domain, the formula gives \(G_X(\tfrac32)=1/(2-\tfrac32)=2\). At \(s=3\), the expectation is the series \(\tfrac12\sum_{k=0}^{\infty}(3/2)^k\). Its \(k\)-th term is \(\tfrac12(3/2)^k\), which does not tend to zero, so the series diverges and \(G_X(3)=\infty\). Substitution into the expression \(1/(2-s)\) instead gives \(-1\), but that expression was derived only for \(s<2\). It is an algebraic continuation, not the expectation at \(s=3\).
Check: at \(s=\tfrac32\), the series ratio is \(\tfrac34<1\) and the formula gives \(2\). At \(s=3\), the ratio is \(\tfrac32>1\), so the terms fail to approach zero. The two checks agree with the domain restriction.
A common error is to treat a closed-form expression for a PGF as if it were valid at every real argument. The series defines the expectation; the simplified expression represents it only where the derivation is valid. Derivative formulas at a boundary also need care: for example, a finite value \(G_X(1)=1\) does not imply that \(G_X'(1)\) is finite.
An MGF at Zero Does Not Guarantee a Neighborhood
For every real-valued \(X\), \(M_X(0)=\mathbb{E}[1]=1\). This single finite value does not imply that the MGF exists for positive and negative \(t\) sufficiently close to zero. In particular, existence on a one-sided interval is not the same as existence near zero. This matters when applying results that require an MGF on an open neighborhood, rather than merely evaluating it at zero.
Worked Example: Finite at Zero but Not Near Zero
Let \(X\) have density \(f(x)=x^{-2}\) for \(x\ge1\), and \(f(x)=0\) otherwise. First check that this is a probability density: \(\int_1^\infty x^{-2}\,dx=[-x^{-1}]_1^\infty=1\). Determine where its MGF is finite.
At \(t=0\), \(M_X(0)=\int_1^\infty x^{-2}\,dx=1\). If \(t<0\), then \(0<e^{tx}\le1\) for \(x\ge1\), so \(0\le M_X(t)=\int_1^\infty e^{tx}x^{-2}\,dx\le\int_1^\infty x^{-2}\,dx=1\); it is finite.
If \(t>0\), the integrand \(e^{tx}/x^2\) does not tend to zero as \(x\to\infty\), and its integral diverges. For example, for sufficiently large \(x\), \(tx-2\log x\ge0\), so \(e^{tx}/x^2\ge1\) on an infinite interval. Thus \(M_X(t)=\infty\) for every \(t>0\). Although the MGF is finite at zero and for all \(t<0\), every open interval containing zero includes positive \(t\), where it is infinite. This MGF does not exist near zero.
Check: the density integrates to \(1\), \(M_X(0)=1\), and for \(t>0\) the integrand is eventually at least \(1\), forcing divergence. These checks establish both the distribution and the claimed domain behavior.
The key diagnostic is to ask what an invoked result actually assumes. If it requires an MGF on a neighborhood of zero, a value at \(t=0\) or a finite value for \(t<0\) does not meet that condition. A formal derivative of an expression at zero cannot supply the missing existence assumption.
Products of Transforms Require Independence
If \(X\) and \(Y\) are independent and the relevant expectations are finite, the MGF of their sum factors: \(M_{X+Y}(t)=M_X(t)M_Y(t)\). The same product rule applies to PGFs of independent nonnegative integer-valued variables at arguments where the expectations are finite. Independence is essential: the expectation of a product does not in general equal the product of expectations. The following result states and proves the MGF version used in the example.
Proof. The identity \(e^{t(X+Y)}=e^{tX}e^{tY}\) holds pointwise. Independence of \(X\) and \(Y\) implies independence of the measurable functions \(e^{tX}\) and \(e^{tY}\). Since both are nonnegative and have finite expectations by hypothesis, the expectation of their product factors, giving \(\mathbb{E}[e^{t(X+Y)}]=\mathbb{E}[e^{tX}]\mathbb{E}[e^{tY}]\). The left side is \(M_{X+Y}(t)\), and the right side is \(M_X(t)M_Y(t)\), as claimed. \(\square\)
Worked Example: A Product Formula for Dependent Variables
Let \(X=Y\), where \(X\) is Bernoulli with \(\mathbb{P}(X=0)=\mathbb{P}(X=1)=\tfrac12\). The variables have the same marginal distribution but are not independent. Compare the actual MGF of \(X+Y\) with the product of their marginal MGFs at \(t=\log 2\).
Since \(X+Y=2X\), its MGF is \(M_{X+Y}(\log 2)=\tfrac12e^{0}+\tfrac12e^{2\log 2}=\tfrac12+\tfrac12\cdot4=\tfrac52\). Each marginal MGF is \(M_X(\log 2)=\tfrac12e^0+\tfrac12e^{\log 2}=\tfrac12+\tfrac12\cdot2=\tfrac32\). Multiplying the marginal MGFs gives \(\tfrac32\cdot\tfrac32=\tfrac94\), which is not \(\tfrac52\).
Check: \(X+Y\) equals \(0\) with probability \(\tfrac12\) and \(2\) with probability \(\tfrac12\); its MGF at \(\log 2\) is therefore \(\tfrac12+\tfrac12\cdot4=\tfrac52\). This verifies the direct calculation and confirms that the product formula cannot be used without independence.
Characteristic-Function Derivatives Need Moments
A CF exists for every real-valued random variable, but this fact alone does not authorize differentiation inside its expectation. A convenient sufficient condition is a finite absolute moment. The next proposition gives the first- and second-derivative formulas and proves the precise conditions used below.
Proof. For fixed real \(t\) and nonzero \(h\), the difference quotient inside the expectation is \(e^{itX}(e^{ihX}-1)/h\). The inequality \(|e^{iu}-1|\le|u|\) for real \(u\), obtained by integrating the derivative \(ie^{iv}\) from \(0\) to \(u\), bounds its absolute value by \(|X|\). As \(h\to0\), it converges pointwise to \(iXe^{itX}\). Since \(\mathbb{E}|X|<\infty\), dominated convergence permits taking the limit inside the expectation and yields the first formula. If \(\mathbb{E}[X^2]<\infty\), apply the same argument to the difference quotient of \(iXe^{itX}\): its absolute value is bounded by \(X^2\), and its pointwise limit is \(-X^2e^{itX}\). Dominated convergence gives the second formula. \(\square\)
Worked Example: A First Derivative Exists but a Second-Moment Formula Fails
Let \(X\) have density \(f(x)=\tfrac32 x^{-5/2}\) for \(x\ge1\), and zero otherwise. The density integrates to \(\tfrac32\int_1^\infty x^{-5/2}\,dx=\tfrac32\cdot\tfrac23=1\). Calculate its first and second moments, then identify which CF derivative formula is justified.
The first absolute moment is \(\mathbb{E}|X|=\tfrac32\int_1^\infty x^{-3/2}\,dx=\tfrac32\cdot2=3\), so it is finite. The proposition gives \(\phi_X'(0)=i\mathbb{E}[X]=3i\). In contrast, \(\mathbb{E}[X^2]=\tfrac32\int_1^\infty x^{-1/2}\,dx=\infty\). The second-derivative formula requiring a finite second moment therefore does not apply; writing \(\phi_X''(0)=-\mathbb{E}[X^2]\) as a finite derivative calculation is unjustified.
One can see the obstruction directly. For real \(t\ne0\), \(\frac{2(1-\operatorname{Re}\phi_X(t))}{t^2}=\mathbb{E}\left[\frac{2(1-\cos(tX))}{t^2}\right]\). As \(t\to0\), the integrand tends pointwise to \(X^2\). Fatou's lemma therefore gives a lower limit at least \(\mathbb{E}[X^2]=\infty\). Thus this quotient is unbounded near zero, consistent with the failure of a finite second-order expansion.
Check: the first-moment integral evaluates to \(3\), while the second-moment integral is \(\tfrac32\int_1^\infty x^{-1/2}\,dx=\infty\). The finite first moment licenses the first derivative; the infinite second moment removes the hypothesis needed for the second-derivative formula.
A Practical Error-Checking Routine
Before accepting a transform calculation, check the definition and the conditions at the point where each operation is used. This short routine catches the errors illustrated above:
- For a PGF: identify its defining series and verify convergence at the proposed argument before using a closed form.
- For an MGF: distinguish finiteness at one point from finiteness on the interval required by the result.
- For a sum: check independence before replacing the transform of a sum by a product of marginal transforms.
- For a CF derivative: check the corresponding absolute moment before differentiating under the expectation.
- For a final expression: compare it with the defining expectation or a direct computation whenever possible.
Common mistake. Starting from the geometric PGF above, one might substitute \(s=3\) into \(1/(2-s)\) and report \(-1\). The correction is to return to the defining series: its terms are \(\tfrac12(3/2)^k\), which do not tend to zero, so the expectation diverges. A compact algebraic expression is not a substitute for checking the domain in which it represents the transform.
The same discipline applies even when the final expression looks familiar. A correct formula used outside its hypotheses can produce a plausible but invalid answer; checking convergence, integrability, and independence is part of the calculation, not an optional afterthought.
Check Your Understanding
For each question, state the assumption or calculation that determines whether the proposed transform step is valid.
- For the geometric distribution in the first example, is the defining PGF series finite at \(s=\tfrac32\)? What is its value?
- Why does \(M_X(0)=1\) not imply that an MGF exists near zero? Use the density \(x^{-2}\) on \(x\ge1\) to explain.
- If \(X=Y\) and \(X\) is a non-degenerate Bernoulli random variable, which condition needed for \(M_{X+Y}(t)=M_X(t)M_Y(t)\) fails?
- Which moment condition justifies the formula \(\phi_X'(t)=i\mathbb{E}[Xe^{itX}]\), and which additional condition justifies the second-derivative formula?
- For the density \(\tfrac32x^{-5/2}\) on \(x\ge1\), which of the first and second moments is finite? Show the corresponding integral test.