Almost sure convergence of the difference of two quantiles Let $X_1, ..., X_n$ be i.i.d. random variables with the same distribution as the random variable $X$. Let $F$ be the distribution function of $X$. We assume that $F$ and its quantile function $F^{-1}$ are $C^1$. Let $\alpha \in [0 ,1]$ be a probability and let $x_\alpha = F^{-1}(\alpha)$ be the corresponding quantile. Let $\widehat{F}_n(x)$ be the empirical distribution of $\{X_1, ..., X_n\}$. Let
$$
\widehat{X}_{\alpha} = X_{(\lceil \alpha n \rceil)}
$$
be the sample quantile based on order statistics. Can we prove that :
$$
F^{-1}\left(\widehat{F}_n \left(x_{\alpha}\right)\right)
- \widehat{X}_{\alpha} \xrightarrow[]{a.s.} 0 \qquad ?
$$
 A: The proof uses asymptotic expansions of $\widehat{X}_{\alpha}$ and $\widehat{F}_n \left(x_{\alpha}\right)$ and a Taylor expansion of the quantile function $F^{-1}$ at point $\alpha$. We know that the sample quantile converges almost surely to the quantile, i.e.
$$
\widehat{X}_{\alpha} \xrightarrow[]{a.s.} x_\alpha.
$$
This implies :
$$
\widehat{X}_{\alpha} = x_\alpha + \epsilon_1
$$
where $\epsilon_1$ is a random variable such that $\epsilon_1 \xrightarrow[]{a.s.} 0$. On the other hand, the empirical distribution converges almost surely to the distribution function :
$$
\widehat{F}_n(x) \xrightarrow[]{a.s.} F(x)
$$
for any $x \in \mathbb{R}$. This must be true for the quantile $x_\alpha$ :
$$
\widehat{F}_n(x_\alpha) \xrightarrow[]{a.s.} F(x_\alpha) = \alpha.
$$
Hence,
$$
\widehat{F}_n(x_\alpha) = \alpha + \epsilon_2
$$
where $\epsilon_2$ is a random variable such that $\epsilon_2 \xrightarrow[]{a.s.} 0$. The Taylor expansion of the quantile function $F^{-1}$ at point $\alpha$ :
$$
\begin{aligned}
F^{-1}\left(\widehat{F}_n \left(x_{\alpha}\right)\right)
& = F^{-1}\left(\alpha + \epsilon_2\right) \\
& = F^{-1}(\alpha) + \left(F^{-1}\right)'(\xi) \epsilon_3 \\
& = x_\alpha + \left(F^{-1}\right)'(\xi) \epsilon_3
\end{aligned}
$$
for some $\xi$ between $\alpha$ and $\alpha + \epsilon_2$. Therefore :
$$
\begin{aligned}
F^{-1}\left(\widehat{F}_n \left(x_{\alpha}\right)\right) - \widehat{X}_{\alpha}
& = x_\alpha + \left(F^{-1}\right)'(\xi) \epsilon_3
- (x_\alpha + \epsilon_1) \\
& = \left(F^{-1}\right)'(\xi) \epsilon_3 - \epsilon_1
\end{aligned}
$$
because the quantiles $x_\alpha$ cancel. Since $\left(F^{-1}\right)'(\xi) \epsilon_3 - \epsilon_1 \xrightarrow[]{a.s.} 0$, this concludes the proof.
