What is bsparse?

bsparse implements sparse data containers, that always have an equivalent dense 2D array representation. As such, bsparse data containers can store …

• … 2D dense arrays of arbitrary shape and dtype.

• … scipy.sparse matrices of arbitrary shape and dtype.

• … bsparse matrices of arbitrary shape and dtype.

• • … a mixture of all of the above.

An important caveat is that matrix rows and columns must not be ragged, i. e. each data container row and column always has a uniform size.

All bsparse data containers implement…

• … basic arithmetic operations, i. e. addition, subtraction, multiplication, and division.

• … dedicated matrix multiplication.

• … (conjugate) transposition.

• … symmetry flags for square symmetric and Hermitian matrices to enable a higher compression.

• … NumPy-like indexing, data modification, as well as changes to the sparsity structure.

• … conversion routines between different storage formats.

• … a number of instantiation and conversion routines.

• … loading and saving data structures to disk.

It is the ideal framework if you want to access and modify block-wise sparse data, and implement algorithms exploiting this structure.

What bsparse is not

As it stands, bsparse is not particularly fast. It merely takes advantage of NumPy’s low-level accelerations.

Unlike scipy.sparse, it also currently does not give you access to high-performance linear algebra routines (although we would like it to do so in the future).

Why bsparse?

In our work we often encounter large, sparse, often diagonally dominant, block matrices. What we couldn’t find was a suitable data structure in Python, hence bsparse.