So I'm going through Casella's Statistical Inference, and in Definition 2.3.6 he defines the moment generating function of a random variable $X$ with cdf $F_X$, denoted by $M_X(t)$, as $$M_X(t) = Ee^{tx},$$ provided that the expectation exists for $t$ in some neighborhood of $0$.

In Theorem 2.3.7, he then states that if $X$ has a mgf $M_X(t)$, then the nth moment of $X$ is equal to the nth derivative of $M_X(t)$ evaluated at $0$.

But why do all of these derivatives have to existed? Is it implicitly assumed without it being stated, or does it somehow just follow from the definition above? If so, how?


These derivatives exist as soon as the moment generating function is finite on some interval containing $0$ (i.e. $M_x(t)$ exists for all $t \in [0,\varepsilon)$ where $\varepsilon > 0$). Often such details are glossed over in textbooks because they can sometimes require some more advanced machinery to prove.

In this case, a straightforward proof is possible. For $t_0 \in [0, \varepsilon)$, we have $$\frac1t \mathbb{E}[e^{(t_0 + t)X} - e^{t_0X}] = \mathbb{E}[e^{t_0X}X] + o(t)\mathbb{E}[e^{t_oX}]$$ by Taylor's theorem. It follows that $\frac{d}{dt}M_X \big|_{t_0} = \mathbb{E}[e^{t_0X} X]$. Repeating this argument will give $$\frac{d^n}{dt^n}M_X \big|_{t_0} = \mathbb{E}[e^{t_0X} X^n]$$ since the presence of $X$ really doesn't make a difference. Evaluating at $t_0 = 0$ gives the result.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.