Rather than using a new random initialization for factorizing the input Y at each rank, use the factorization of smaller ranks with padded random numbers to warm start the next largest rank.
Possible complications:
- constraints may not be immediately satisfied and need to be reapplied when padded
- the initialization function needs to be able to generate single row/slices rather than just new entries, otherwise we should generate a full sized factorization and in-place fill the first entries with the smaller factorization output
Rather than using a new random initialization for factorizing the input Y at each rank, use the factorization of smaller ranks with padded random numbers to warm start the next largest rank.
Possible complications: