Another side to this argument concerns the notion of a "rhythm space" -- a metric space for rhythms. One would imagine the construction of this space to be facilitated by the inclusion of quantized tatum-grids in the representation. Specifically, in that case the space would be essentially discrete, aside from "feel factors" which would provide small continuous variations of a fixed underlying rhythmic pattern. But this would not be general enough. Consider the possible placements of five clave hits over the span of 2 seconds -- a subspace of rhythm space. The range of all such rhythms is continuous, clearly, and the representation should reflect that. Thus the relationships among, say, various Afro-Cuban claves and various West African bell patterns are not lost because of quantization issues.
So imagine a rhythm space, for a monophonic voice, as follows: A particular point in the space (i.e., a particular rhythm) is characterized by
t_on[i], i = 1 to N, t_on[i] < t_on[i+1] < tfor all
t_off[i], i = 1 to N, t_on[i] < t_off[i] < t_on[i+1] < tfor all
We could imagine a psychoacoustics experiment similar to the timbre-space multidimensional scaling, that would help us find other relevant dimensions.