One trifling question- the bottom of the first col. of page 13 says 300 examples, whereas the middle of the second col. says 3000? I assume this is a typo but does anyone know which is right? Just curious....

As for the choice-of-kernel debate, I'm pretty sure that since the idea is to use a kernel of the form
k(x,y) = (\Phi(x) \dot \Phi(y))
where \Phi maps from the input space to some useful space where things are nicely seperable and such, the kernel has to make a difference. If you used a kernel that related to a \Phi that mapped the input space into some lame space where the data wasn't usefully distinguishable, I don't see how it would work. To give a sneek preview of my paper, it involves finding Gaussian-weghted distances between objects (which can be seen as a kernel operation), which works better than a bunch of other measures you could pick. Therefore, I cast my vote for the kernel-has-to-matter side.

I was going to say something about the notation being unfamiliar and getting somewhat lost in the math, but since we're not delving into that in great detail I'll save that comment for later ;-)