Skip to content

Add missing regularization term in the quality measurement for step acceptance

Klaus Böhnlein requested to merge FixRiemannianPN into master

When checking for the acceptance of a new iterate in the Riemannian proximal Newton solver the regularization-term which is part of the quadratic model is missing in the quality measurement of the step. The model decrease should be: $ m(x) - m(x+s) = -<g,s> - 0.5 <s, (H + regularization * I) s> $ However $ m(x) - m(x+s) = -<g,s> - 0.5 <s, H s> $ is currently used.

Merge request reports