What operations on sparse tensor keep it coalesced?

I cannot find in the documentation which operations on sparse tensors keep them coalesced.
In particular, I would like to know if torch.sparse.sum keep them coalesced.

If the sum is implemented by “simply” removing one index, then it doesn’t. But I think that, using the fact the indices are sorted (when coalesced), the sum could be implemented efficiently while keeping the tensor coalesced.