Robust Aitchison changes with the newest update 2.7-1

[m3j]

Hi

I have encountered quite a big change with the newest update of vegan with regards to vegdist with the robust.aitchison method, probably relating to the change noted in the release notes:

• Robust Aitchinson distance with matrix completion for zero entries was added to vegdist and corresponding standardization in decostand.

Before the change I had a nice screeplot using ape::pcoa for calculating the eigenvalues

But after the update I only have three valid PCs as the rest have very small values below e-11 and I get the same with stats::cmdscale and eigen():

How can this be?
From #570 I understand that before the zeros was just ignored doing clr on nonzero data not considering the zeros and then taking the euclidian distances discarding the original zeros. But now instead using the method from marino et al 2019 where the centralization is done after the log transform of nonzero data thereby including the zeros when centering. Or as stated in the changelog "uses matrix completion to estimate the missing values." - so does that mean changing the -inf from the log(0) to something else than 0?
Either way I am not sure why it affect my data so much. Only 4% of my counts are zero.

Anyone have any idea?

[jarioksa]

No answer yet, so here my go: the "completed" matrix (imputed matrix) defaults to be of rank 3 which also means that it will have three positive eigenvalues. The matrix completion also changes non-zero values, and even if you have no zeros the imputed matrix will be of the given rank. The decostand help page tells that the rank can be specified with argument ropt. See also the help of optspace which is the function used for matrix completion.