`CoulombMatrix(permutation="sorted_l2")` is not symmetric about dscribe HOT 5 CLOSED

Hi @MomoLangenstein,

Thanks for the report. I need to check what is going on. If this is the case, we definitely need to fix it and introduce a better test that also checks the symmetricity.

from dscribe.

Hi @MomoLangenstein,

Indeed there was an issue where the rows and columns of the matrix were not being permuted as expected, causing non-symmetric matrices. There is now a fix in the branch v1.2.2, which will be made into a release in the near future.

It would be great if you can verify this fix as well. I have added a new symmetricity test into our test suite and your examples now also pass after changing the sorting on the python side slightly like this:

sorted_row_indices = np.argsort(-row_norms)

The sorting in your manual implementation needs to be exactly the same as in our C++ implementation due to ambiguity in sorting by row norm. In several cases the norms of the matrix rows are identical and in order to get the exact same results, the sorting logic needs to be identical in your example and in our C++ implementation.

from dscribe.

Hi @lauri-codes,

Thank you for working out the fix for this issue! I can confirm that my small test now works on the v1.2.2 branch. I just wanted to clarify your proposed change to the manual Python implementation. If I do use

sorted_row_indices = np.argsort(-row_norms)

instead of

sorted_row_indices = np.argsort(row_norms)[::-1]

I get exactly the same results with my manual implementation as with dscribe. Are both methods semantically correct, and only differ in their handling of how to order rows/columns with equivalent norms?

from dscribe.

Hi,

Yes, the sorting simply affects the way that rows with equal norm are handled. If there are no equal norms, then the output is completely independent of the sorting algorithm (see e.g. the kind attribute in the numpy docs) and whether you do np.argsort(-row_norms) or np.argsort(row_norms)[::-1].

The small differences only start kicking in when there are equal norms. I think np.argsort(-row_norms) and np.argsort(row_norms)[::-1] should not be equal in these cases, e.g.:

import numpy as np

a = [
    [2, 1],
    [1, 2],
]

row_norms = np.linalg.norm(a, axis=1)
indices1 = np.argsort(row_norms)[::-1]
print(indices1) # [1, 0]
indices2 = np.argsort(-row_norms)
print(indices2) # [0, 1]

In the end it does not really matter which way you sort the rows with equal norms, just as long as you are consistent throughout your data.

from dscribe.

Now fixed in v1.2.2.

from dscribe.