Let $X_n \sim Beta(k,n+1-k)$, with $k = [\nu n]$ for some $\nu \in (0,1)$.

I want to show that

$$X_n \to \nu $$

almost surely as $n \to \infty$.

My attempt:

I have successfully shown convergence in probability by writing out the probability $P(|X_n - k/n| > \epsilon)$ and applying Markov's inequality. However here I am stuck, and haven't been able to progress further.

I have heard that a beta random variable can be expressed as a ratio of sums of standard exponential distributions:

$$X = \frac{\sum_{i=1}^k Z_i}{\sum_{i=1}^{n+1} Z_i}$$

where $Z_i \sim \exp (1)$.

If this is true (and I think it is, using the relation between Gamma and exponential, and gamma and beta), the Strong Law of Large Numbers could be used to show that $X_n = Y_n/Z_n$ where $Y_n \to k$ almost surely and $Z_n$ converges to $n$ almost surely. But, again, I'm stuck. Since Slutsky's thereom can only be used to show convergence in probability here, not almost sure.


According to Wikipedia, if $X_n, Y_n$ converge almost surely to $X,Y$ respectively, then $X_nY_n$ converges to $XY$ almost surely. Hence, since we can show, under the exponential representation, we have $X_n = Y_n / Z_n$ where $Y_n \to k$ and $Z_n \to n$, then applying the continuous mapping theorem to $Z_n$ I believe we can write

$$X_n = Y_n Z^*_n$$

Where $Y_n \to k$ and $Z_n^* \to n^{-1}$. It then follows by the result I found on Wikipedia that $X_n \to k/n$ almost surely. Can someone confirm this proof is correct?

  • $\begingroup$ How does the limiting random variable still depend on $n$ after taking $n$ to infinity? $\endgroup$ – Shashi Oct 14 '18 at 8:50
  • $\begingroup$ Sorry, I should have specified, $k$ also increases with $n$. It is defined as $k = [n \nu ]$ for some $\nu \in (0,1)$. So $k/n \to \nu$. $\endgroup$ – Xiaomi Oct 14 '18 at 8:54
  • $\begingroup$ I would suggest you write $X_n\to \nu$ in that case... BTW what are these square brackets? Floor function? $\endgroup$ – Shashi Oct 14 '18 at 9:02
  • $\begingroup$ Yes, floor function/integer part. $\endgroup$ – Xiaomi Oct 14 '18 at 9:05

You are really close. I denote $k$ as $k_n:=[n\nu]$ where $[\cdot] $ as you said is the integer part. We can indeed write $$X_n=\frac{\sum_{i=1}^{k_n}Z_i} {\sum_{i=1}^{n+1}Z_i}, $$ where $Z_1,Z_2,...$ are i.i.d. $\exp(1)$ distributed random variables. We know from the SLLN that $$\frac{1}{k_n}{\sum_{i=1}^{k_n}Z_i}\to 1 \ \ \ \text{ a.s. for } n\to\infty, $$ and similarly $$\frac{1}{n+1}{\sum_{i=1}^{n+1}Z_i}\to 1 \ \ \ \text{ a.s. for } n\to\infty. $$ But then we have $$X_n=\frac{k_n}{n+1}\frac{\frac{1}{k_n}\sum_{i=1}^{k_n}Z_i} {\frac{1}{n+1}\sum_{i=1}^{n+1}Z_i}\to \nu \ \ \ \text{a.s., }$$ where we have used the fact that $$\lim_{n\to\infty} \frac{k_n}{n+1}=\nu$$

  • $\begingroup$ Thank you, especially for setting me straight on notation. I realise now it's non-sensical to write $lim X_n = k/n$. Will make sure to keep that in mind in future $\endgroup$ – Xiaomi Oct 14 '18 at 9:32
  • $\begingroup$ @Xiaomi Good to hear that I could help! $\endgroup$ – Shashi Oct 14 '18 at 9:33

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.