I think they completely misstated in the original paper what they did. It was a tensor decomposition of complex of 4x4 matrices up to the factor 0.5. Which is a nice result, but it is not really anything practical for a computer program doing 4x4 complex matrix multiplication.