5
$\begingroup$

The following question is straightforward to show with Markov's Inequality:

Let $X_n$ be uniformly distributed on the set {1, 2, ..., n}, and let $Y_n = 1/X_n$. Prove that $Y_n \overset{p}{\rightarrow} 0$.

For practice, I also wanted to use MGFs (on the justification $Z_n\overset{p}{\rightarrow} c \iff Z_n \overset{d}{\rightarrow} c$ where $c$ is a constant). It follows that $M_{Y_n}(t) = \frac{1}{n} \sum_{x=1}^n e^{t/x}$, but when I do the limit by hand I get 0. I also get 0 using Mathematica:

Limit[ Sum[  (1/n) Exp[t/x]   , {x, 1, n}], n -> Infinity]
0

Clearly these are both "wrong" because the MGF of a degenerate random variable equal to 0 with probability 1 is "1", and so that must be the "proper" limit.

The MGF has to only exist in a neighborhood of 0 for $t$, and clearly at $t = 0$ the limit is 1 (but not for any specific but arbitrarily small $t$).

My only thoughts are that I have the MGF wrong itself or that there's something subtle happening with the limit of the arithmetic mean of infinite terms close to 1. Any thoughts would be appreciated!

Edit: Adding a note that my Mathematica version is 11.3, primarily for reference if it helps someone in the future.

$\endgroup$
  • 1
    $\begingroup$ Mathematica 9.0.1.0 does not make the same error with the limit. $\endgroup$ – whuber Jul 28 '18 at 20:46
5
$\begingroup$

It looks like the mgf is correct, because by definition

$$M_{Y_n}(t) = E[\exp(t Y_n)] = \sum_{x=1}^n \exp(t / x) \Pr(X_n=x) = \frac{1}{n}\sum_{x=1}^n \exp(t/x).$$

Since $t$ is real, all these terms are positive. The arithmetic-geometric mean inequality implies

$$M_{Y_n}(t) \ge \left(\prod_{x=1}^n \exp(t/x)\right)^{1/n} = \exp\left(t H_n/n\right)$$

where I have written $H_n=1+1/2+\cdots + 1/n$ for the $n^\text{th}$ Harmonic number. As is well known (and easy to show), $H_n/n\to 0$ as $n$ grows large. The continuity of $\exp$ then implies the right hand side of $(*)$ approaches $\exp(t0)=1$ in the limit. That is,

$$\lim_{n\to\infty}M_{Y_n}(t) \ge 1.\tag{*}$$

On the other hand, consider any real $t$ and select an arbitrary $\varepsilon \gt 0.$ For all integers $n \gt |t| / \log(1+\varepsilon),$ $\exp(t/n) \lt 1 + \varepsilon.$ Let $N \ge 1$ be an integer and break the sum for $M_{Y_{N+n}}(t)$ into two parts,

$$\eqalign{ M_{Y_{N+n}}(t) &= \frac{1}{N+n} \sum_{x=1}^n \exp(t/x) + \frac{1}{N+n} \sum_{x=n+1}^{N+n} \exp(t/x) \\ &= \frac{n}{N+n} M_{Y_n}(t) + \frac{1}{N+n} \sum_{x=n+1}^{N+n} \exp(t/x) \\ &\lt \frac{n}{N+n} M_{Y_n}(t) + \frac{1}{N+n} \sum_{x=n+1}^{N+n}(1+\varepsilon) \\ &= \frac{n}{N+n} M_{Y_n}(t) + \frac{N}{N+n}(1 + \varepsilon). }$$

As $N$ grows large (still keeping $t,$ $\varepsilon,$ and $n$ fixed), the first term converges to $0$ while the second converges to $1 + \varepsilon,$ establishing

$$\lim_{n\to\infty} M_{Y_n}(t) \le 1 + \varepsilon.\tag{**}$$

Since $\varepsilon$ is arbitrarily small, results $(*)$ and $(**)$ together show the limit of the mgf can only be the unit constant function.

$\endgroup$

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.