boundlab.polysp.softmax.l1_optimal_step#

boundlab.polysp.softmax.l1_optimal_step(direction, residual)[source]#

Per-column step t minimizing |residual + t * direction|.sum(0).

The minimizer of this piecewise-linear objective is the weighted median of the ratios -residual / direction under the weights |direction|. Zero is always a candidate, so stepping never increases the L1 norm.