Unexpected linalg.matrix_rank behavior

[amueller]

I had a couple of unexpected results using matrix_rank. I'm on 1.6.1.
One was that removing a row from a matrix increased the rank, another was that adding a row to a matrix increased the rank by 2.
Also, these are not really reproducible and seem to be related to slicing.
I get something like:

>>>np.linalg.matrix_rank(mat[:58])
55
>>>np.linalg.matrix_rank(mat[:59])
57
>>>np.linalg.matrix_rank(mat[:58].copy())
56

I'm not sure where to go from here.
The matrix in this case has shape (80, 92) and three non-zero entries per row.

Is / was that a known issue that has been fixed in later version? Should I try master?
Any help would be appreciated.
Thanks,
Andy

[pv]

It counts singular values larger than cutoff eps*max(svals). This may be sensitive to rounding error, which can depend on data alignment, hence the dependence on .copy(). You can maybe check what SVD gives you and see if that helps to explain things.

OTOH, the behavior was slightly tweaked in 78b7693, but I don't think this matters here.

[amueller]

I thought it might be a precision issue. I started playing with the SVD, maybe it'll help. I thought it may just be a numerical issue but the dependence on copying and slicing was a bit irritating.

Thanks :)

[seberg]

As noted, a precision issue of matrix_rank and not really a bug. I don't think there is any action possible here, so closing.