Uniform consistency of an empirical moment-generating function (source code)

= Uniform consistency of an empirical moment-generating function
{title2=$\sup_{0\le s\le t}|\widehat m_n(s)-m(s)|\to0$}

Let $X_1,X_2,\ldots$ be <independent and identically distributed> nonnegative <random variables>. Fix $t>0$ with $\mathbb E e^{tX_1}<\infty$, and define
$$
\widehat m_n(s)=\frac1n\sum_{i=1}^ne^{sX_i},\qquad
m(s)=\mathbb E e^{sX_1},\qquad 0\leq s\leq t.
$$
Then the empirical <moment-generating function> is <uniform law of large numbers>[uniformly strongly consistent] on this <interval>:
$$
\sup_{0\leq s\leq t}|\widehat m_n(s)-m(s)|\longrightarrow0
\quad\text{almost surely}.
$$
The <dominated convergence theorem>, with envelope $e^{tX_1}$, makes $m$ <continuous function>[continuous] on $[0,t]$. Both $m$ and every $\widehat m_n$ are <monotone function>[nondecreasing] because the observations are nonnegative. Given $\varepsilon>0$, choose a finite grid $0=s_0<\cdots<s_k=t$ whose successive $m$ increments are at most $\varepsilon$. The <strong law of large numbers> holds simultaneously at its finitely many endpoints. Squeezing between adjacent endpoints gives
$$
\sup_s|\widehat m_n(s)-m(s)|
\leq\max_j|\widehat m_n(s_j)-m(s_j)|+\varepsilon.
$$
Use a countable sequence $\varepsilon\downarrow0$ to conclude. <continuity> of the sample and population <functions> also makes the supremum <measurable function>[measurable] by reducing it to <rational number>[rational] parameter values. This is the finite-bracketing mechanism behind the <uniform strong law from finite L1 bracketing>. The <independent and identically distributed>[iid] sampling assumption cannot simply be omitted: repeatedly observing the same nonconstant <random variable> would not create averaging.