Comparison of an Apocalypse-Free and an Apocalypse-Prone First-Order Low-Rank Optimization Algorithm

02/18/2022

∙

by   Guillaume Olikier, et al.

∙

0

∙

share

We compare two first-order low-rank optimization algorithms, namely P^2GD (Schneider and Uschmajew, 2015), which has been proven to be apocalypse-prone (Levin et al., 2021), and its apocalypse-free version P^2GDR obtained by equipping P^2GD with a suitable rank reduction mechanism (Olikier et al., 2022). Here an apocalypse refers to the situation where the stationarity measure goes to zero along a convergent sequence whereas it is nonzero at the limit. The comparison is conducted on two simple examples of apocalypses, the original one (Levin et al., 2021) and a new one. We also present a potential side effect of the rank reduction mechanism of P^2GDR and discuss the choice of the rank reduction parameter.