Dan Schult - Sparse arrays in scipy.sparse | SciPy 2024

SciPy is transitioning from sparse matrices to sparse arrays to better align with NumPy’s array API and modern practices

Key differences between sparse matrices and arrays:

• Matrices are always 2D while arrays can be 1D or 2D
• Arrays use @ operator for matrix multiplication (matrices use *)
• Element-wise operations behave differently
• Indexing returns different dimensional results

Migration path is gradual:

• New array classes exist alongside matrix classes
• Construction functions have new array-based versions
• Migration guide and developer support coming soon
• Full deprecation planned after at least two release cycles

Current status:

• 1D and 2D sparse arrays are fully functional
• Indexing support coming in version 1.15
• ND (>2D) support planned for future

Storage formats:

• CSR (Compressed Sparse Row)
• CSC (Compressed Sparse Column)
• COO (Coordinate format)
• DIA (Diagonal format)
• Different formats optimize for different sparsity patterns

Migration recommendations:

• Update existing matrix code first
• Convert constructors from matrix to array versions
• Change multiplication operators (* to @)
• Update helper functions and shape handling
• Ensure good test coverage

Backward compatibility maintained:

• Old matrix code continues to work
• Side-by-side implementation reduces migration friction
• Matrix features being maintained during transition