boundlab.polysp.softmax.softmax_opt

boundlab.polysp.softmax.softmax_opt(polysp, axis=-1)

Add the optimal multiple of the zero-valued constraint \(\sum_i \sigma_i - 1 = 0\) to each element.

Sound because the added expression is identically zero on every concrete input; effective because its error rows are anti-correlated with the output’s and cancel them.