Gradient adaptation under unit-norm constraints

Abstract
We study gradient-based adaptive algorithms within parameter spaces specified by /spl par/w/spl par/=1, where /spl par//spl middot//spl par/ is any vector norm. Several approximate algorithms for this task have already been developed when /spl par/w/spl par/ is the L/sub 2/ norm. We derive general algorithm forms for arbitrary vector norms and relate them to true gradient procedures via their geometric structures. We also give algorithms that mitigate an inherent numerical instability for tangent-vector L/sub 2/-norm methods. Simulations showing the performance of the techniques for minor component analysis are provided.

This publication has 14 references indexed in Scilit: