Chain products

[amasotti]

Matlab and Python seem to handle matrix multiplication quite differently.
The weights in the C-space, for example, are computed according to the following formula:

TP^T x W x TP

implemented in MATLAB as:

TP'*W*TP

and in Python as:

# with numpy 
TP.T @ W @ TP

# with pytorch
TP.T.matmul(W).matmul(TP) 

With identical TP, TP^T and W matrices these products give slightly differently results. Although the sum of values per row/col is the same, the single values in the matrices are differently distributed.

I wonder if this is an issue for the GSC implemented here or whether this is just a by-product of different algorithms used in Matlab and Python but doesn't affect the final result.