downsampling a kde / combining kde and histogram

I'm calculating a KDE of one parameter (y, particle density) in bins of another parameter (x, distance from the origin). At small x I have poorly sampled data (10s to 1000s of points per x-bin) while at large x values it is very well sampled (millions of points per x-bin). At small radii, using a KDE seems very important/effective, while at large radii the result is effectively identical to a histogram, but it extremely slow to compute (at least using scipy in python) *. Ultimately I don't need the KDE per se, I just need the smoothed/sampled PDF it produces (i.e. on a regular grid).

It seems like a hybrid approach would be possible in which the KDE is used when the sampling is sparse, but I revert to simple binning when it is very well sampled. Is there a standard procedure for hybridizing these approaches? Or are there techniques for adaptively choosing the bandwidth such that I can use a kernel with finite support that shrinks as the sampling becomes more dense?

* Ultimately, I assume this is because I'm using Gaussian kernels with infinite support, and thus N*M evaluations (for N particles in the given KDE, which is resampled onto a grid of M points).

• A simple procedure would be to compute the KDE of a nonlinear transform of the parameter and then convert it back. Assuming anything close to a slowly-varying density near the origin, a good choice would be the $n^\text{th}$ root of the distance in $n$ dimensions. – whuber Feb 1 at 22:41
• @whuber can you expand a bit? I don't see how a transform helps the computational costs. Or is the idea that you reshape the parameter space, to mimic the effect of changing the kernel (/ bandwidth)? Do you know any references/tutorials for this, offhand? – DilithiumMatrix Feb 2 at 15:09
• An appropriate transformation will place approximately the same numbers of data in each one of a regularly-spaced set of bins. Some examples appear at stats.stackexchange.com/questions/65866 – whuber Feb 2 at 16:21