Sparse tensors suit data containing many zeros: they store only nonzero values, potentially reducing memory use. jax.experimental.sparse provides experimental support for sparse tensors and sparse matrix operations, but its interface may change. section 3.3.2
Use sparse representations when:
- A tensor contains many zero-valued elements and relatively few nonzero ones. section 3.3.2
- Specialized sparse linear algebra routines can exploit that structure efficiently. section 3.3.2
