I'm asking this question based off of Example 10.2.3 in Casella & Berger. We have a sample $X_1, ..., X_n$ from a population with pdf $f$ and (differentiable) cdf $F$. Let $\mu$ be the population median and $M_n$ be the sample median. We wish to compute

$$ \lim_{n \to \infty} P(\sqrt{n}(M_n - \mu) \leq a) $$

for some $a$. We then define $Y_i$ by

$$ Y_i = \begin{cases} 1 & \mbox{if } X_i \leq \mu + a/\sqrt{n} \\ 0 & o.w. \end{cases}$$

Clearly the $Y_i$ are Bernoulli with success probability $p_n = F(\mu + a/\sqrt{n})$. If we assume that $n$ is odd, then $\{M_n \leq \mu + a/\sqrt{n}\}$ is equivalent to $\{\sum Y_i \geq (n+1)/2\}$. Now, the book states "A straightforward limit calculation will also show that"

$$ \frac{(n+1)/2 - np_n}{\sqrt{np_n(1-p_n)}} \to -2aF'(\mu) = -2af(\mu) $$.

I follow completely up until this limit calculation at which point I'm completely lost. I tried substituting in the definition of $p_n$, separating, etc. I'm probably just overlooking something simple but this "straightforward" calculation has been eluding me for a couple hours now.


1 Answer 1


Welp, I've had an epiphany soon after posting. Here's what I have,

\begin{align*} \lim_{n \to \infty} \frac{(n+1)/2 - np_n}{\sqrt{np_n(1-p_n)}} &= \lim_{n \to \infty} \frac{(n+1)F(\mu) - nF(\mu + a/\sqrt{n})}{\sqrt{np_n(1-p_n)}} \\ &= \lim_{n \to \infty} \frac{n(F(\mu) - F(\mu + a/\sqrt{n}))}{\sqrt{n}\sqrt{p_n(1-p_n)}} + \lim_{n \to \infty}\frac{F(\mu)}{\sqrt{np_n(1-p_n)}} \\ &= \lim_{n \to \infty} \frac{F(\mu) - F(\mu + a/\sqrt{n})}{1/\sqrt{n}}\frac{1}{\sqrt{p_n(1-p_n)}} \\ &= \lim_{n \to \infty} a\frac{F(\mu) - F(\mu + a/\sqrt{n})}{a/\sqrt{n}}\frac{1}{\sqrt{p_n(1-p_n)}} \\ &= -aF'(\mu)\frac{1}{\sqrt{.25}} \\ &= -2af(\mu) \end{align*}

That took an embarrassingly long time to figure out but it's nice to understand it.

  • $\begingroup$ Nice answer. Thanks for taking the time to answer your own question. $\endgroup$
    – Glen_b
    Apr 18, 2017 at 0:48
  • $\begingroup$ Posting does that to a person. (+1). $\endgroup$ May 28, 2017 at 13:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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