Some sparse matrix formats store explicit zero-valued elements (see e.g.

I know that in some applications of sparse matrices in which implicit zeros have different semantics (e.g. unknown value), explicit zeros are exploited, such as for data imputation.

It seems explicit zeros are maintained in matrix conversion to CSR, CSC, BSR, COO and DOK, but not to LIL.

To my knowledge the properties of the sparse matrix implementations with respect to explicit zeros are undocumented and untested. I presume they should be.

For an example of inconsistency:

>>> A = sp.coo_matrix([[0]])
>>> A.nnz
0
>>> B = A.tocsr()
>>> B[0, 0] = 0
>>> B.nnz
1
>>> B = A.todok()
>>> B[0, 0] = 0
>>> B.nnz
0
>>> B = A.tolil()
>>> B[0, 0] = 0
>>> B.nnz
0

What is correct behaviour?