I'm working with quite large rotation matrices, which have the inherent property to have a large number of zeros. In order to reduce memory use and possibly reduce computation cost when multiplying these rotation matrices with other matrices/vectors, I would like to use a sparse matrix data structure. I have found documentation on the SciPy sparse matrices, but I don't quite understand how these work and what the differences are. (SciPy docs)
What would be the best sparse data structure to use for rotation matrices in Python?
If you already have the rotation matrix as a dense array you can simply do
m = csr_matrix(dense_rot_matrix)
One of the two types csr_matrix or csc_matrix should be used.
A better option would be to populate already the sparse matrix which can be easily accomplished using the coo_matrix type, which has efficient methods to convert to csr_matrix or csc_matrix. I've been using Cython to create sparse matrices in this way very efficiently.