Suppose I have $(\mathbf{X}_1, \cdots, \mathbf{X}_n)$ from a multivariate distribution $f$. The multivariate KDE is \begin{align*} \widehat{f}_\mathbf{H}(\mathbf{x}) = n^{-1}\sum_{i=1}^{n}K_\mathbf{H}(\mathbf{x} - \mathbf{X}_i) \end{align*} with bandwidth matrix H. The bandwidth matrix can be computed via a plethora of methods; I'm leaning towards either plug-in selectors or smoothed cross-validation. In my specific example, each $\mathbf{X}_i \in \mathbb{R}^2$ and my $n$ is tremendous ($n \approx 130,000$). I'm thinking about reducing computing time by performing on a subset of observations $(\mathbf{X}^*_1, \cdots, \mathbf{X}^*_m) \subset (\mathbf{X}_1, \cdots, \mathbf{X}_n)$ with $m \ll n$ to obtain some estimate $\widetilde{H}$. I know that subsetted bandwidth estimate $\widetilde{H}$ is not optimal for the entire dataset $(\mathbf{X}_1, \cdots, \mathbf{X}_n)$ because people have shown that $\widetilde{H}$ decreases with respect to $n$. However, is it possible to simply scale $\widehat{H} = c \widetilde{H}$ such that $\widehat{H}$ is approximately optimal for the entire data? If it makes it any easier, I plan to also restrict $\mathbf{H}$ to be diagonal.
Thanks!