Do quantiles of quantiles converge to quantiles?

45 Views Asked by At

Is there any truth to the statement "the median of medians converges to the true median?".

While it's false that the median of medians is the median, is there a way to make this true asymptotically? I'm happy to take on regularity conditions on the underlying distribution (like continuous density function, etc).

More precisely, this is my setup. I have iid random samples $x_{i,j}$ of some random variable $X$.

For each $i$, I can take the median over $j$.

$$ m_i(n) := \text{median}(x_{i,j} | j \leq n) $$

And then take the median over those

$$ m(k,n) := \text{median}(m_i(n)| i \leq k) $$

Is it true that

$$ \lim_{k,n \to \infty} m(k,n) = \text{the actual median of } X $$

under some conditions on $X$?

And same question where median is replaced by other quantiles of $X$.

(though a counterexample with a non-super-weird $X$ (say continuous with finite mean and variance) would make me even happier)


I found this reference from another question, https://www.math.ucla.edu/~tom/papers/unpublished/meanmed.pdf

where it claims that the distribution of sample quantiles is asymptotically normal. So I guess the correct statement is that "average of quantile converges to quantile". IIUC, a counterexample to "quantile of quantile converges to quantile" would be greatly appreciated then!