Lee and Shin's Quaternion filters paper

Lee and Shin give a method to apply time-domain filters to rotational
data, and show some example filters. These are, importantly,
coordinate-invariant, time-invariant and symmetric. They give a 
discussion of why generalizing standard signal processing methods to 
non-linear spaces (like rotations). And they show how they overcome this 
issue by mapping rotational data to a vector space, applying filters, 
and mapping back.

Key ideas: Non-linear spaces don't allow for easy filtering, also have
non-intuitive continuity problems. Renormalization of translated
quaternions can have side effects such as singularities and unexpected
distortion.

Contributions: A way around these side effects, and a method of signal
processing rotations, using the exponential map.