It is far from elegant!

Consider allowing any number of vector arguments such that the entropy is calculated for their sum but divided by the sum of entropies of each vector. This could serve as a normalization different to the one currently used, but possibly more relevant to the purpose at hand.

It is also possible that entropy differences could be used to calculate (replace) cosine similarity to determine the similarity of any two spectra.

Also simplify the coding, and ensure everything works nicely with sigma = 0 (see #4 (comment)), and normalization.

Given that there are at least three different types of input -- low resolution indicator vectors (no smoothing), smoothed high resolution tensors, and pitch/time class sets, it would be useful for the expected input type of the function's arguments to be made explicit with their names.

Alternatively, the input type could be flexible and the function checks the input and acts accordingly.

Options?:
x_ind for binary indicator vector (pitch/time-domain)
x_? for non-binary weight vector (pitch/time-domain)
x_p and x_w for pitch/time vector and associated weight vector (event-domain)
x_pc, x_w, and period for pitch/time vector with associated weight vector and period size (event-domain)