Sat Jul 12 14:21:05 CEST 2008

Inner loop over angle/tangent

( I'm not sure if this is useful: only works when the angle/tangent
loop is the inner one. )

Since we're not using any resolution-reducing histogram, and
floats/doubles have a large dynamic range, can the parameter space be
turned into something more convenient so the kernel method for
evaluatiing p(R,t) becomes simpler?

Goal: make a kernel method that evaluates p(R,t) in a different
parameter space. 

Let's try to replace it by (o,m) : offset and tangent of angle, and
use the computation of p(R,m) to compute p(R,t).

The bundle equation for (x,m) becomes:

      (o,m) | o = y - m x

Voting for e^{i o/O(m)} = e^{ i (y - m x) / O(m) } becomes fairly
trivial: the exponential can be accumulated since it's linear in both
m and (x,y). The polar coord version is only linear in (x,y), so works
only for a fixed t.