I'm looking to re-read a paper I read 2-3 years back, but my search-fu is failing me now at finding it again.
The paper was about a new (at the time) approach for kernel smoothing by implicitly defining the kernel in a data-dependent way by looking in the Fourier domain. The basic idea of the paper was to look at the spectrum of the empirical distribution, take the region inside a level curve around the origin, then transform just that region back to the spatial domain (More or less. Like I said, it's been a few years). So, the kernel itself is only ever implicitly defined by being the inverse Fourier transform of a 0-1 mask.
The paper wasn't new even when I read it (maybe 10 years old then, so ~15 years now?).
The example given by the authors was a series of ridge distributions forming letters spelling out the name of the lab (whose name I don't remember) over a central gradient. The big selling point was that it did a good job on distributions that are locally non-isotropic in many different directions and have both high and low frequency components.
This may be a standard approach these days for all I know, so I would be interested in any follow-on work as well.
