This paper is about sparse multi-dimensional arrays in Python. We discuss their applications, layouts, and current implementations in the SciPy ecosystem along with strengths and weaknesses. We then introduce a new package for sparse arrays that builds on the legacy of the scipy.sparse implementation, but supports more modern interfaces, dimensions greater than two, and improved integration with newer array packages, like XArray and Dask. We end with performance benchmarks and notes on future work. Additionally, this work provides a concrete implementation of the recent NumPy array protocols to build generic array interfaces for improved interoperability, and so may be useful for broader community discussion.

Keywords:sparsesparse arrayssparse matricesscipy.sparsendarrayndarray interface