Let $X_1,\dots, X_n$ denote i.i.d. random variables, with a smooth density $f$ having unique median $m$.

It is known that the sample median $M_n$ of $X_1,\dots, X_n$, $M_n = \operatorname{med}(X_1, \dots, X_n)$, is asymptotically normal with mean $m$ and variance $\frac 1 {4 n f(m)^2}$. (Or phrased more precisely, $Z_n = (M_n-m)\sqrt{4 n f(m)^2}$ approaches a standard normal distribution; e.g. Grimmet & Welsh, Probability, an introduction, problem 8.6.16).

Now, for $k \leq n$, let $M_k$ denote the sample median of the first $k$ variables, $M_k := \operatorname{med}(X_1, \dots, X_k)$.

What is known about the asymptotic joint distribution of $(M_k, M_n)$ as $k \rightarrow \infty$ and $n \rightarrow \infty$ (perhaps for $k = \lfloor \alpha n \rfloor $ for some $0 < \alpha < 1$)?

I am able to obtain a rather involved non-asymptotic expression for the joint distribution of $(M_k, M_n)$ but before trying to derive an asymptotic expression, I wonder what is known about this.

Update It is natural to expect that $M_k$ and $M_n$ are correlated Gaussian random variables with correlation coefficient $\sqrt{\alpha} \approx \sqrt{k/n}$. Is it true?


If each $X$ is Bernoulli with probability $p$, then the correlation between $M_k$ and $M_n$ has the following graph for $k=21$, $n=41$:

enter image description here

So also for smooth distributions approximating the Bernoulli distributions, the correlations of medians would depend on $p$, and not just on $k$ and $n$. In general, the correlations of medians would depend on the parent distribution and not just on the sample and subsample sizes.


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.