Mateusz Baran

Here is a (relatively) simple reproducer: ```julia using Manopt using Manifolds using LineSearches using LinearAlgebra norm_inf(M::AbstractManifold, p, X) = norm(X, Inf) function f_rosenbrock_manopt(::AbstractManifold, x) result = 0.0 for i in...

One of the top causes of losing normalization in optimization on a sphere in Manopt is accumulation of numerical errors. The default exponential map isn't particularly stable and this is...

This is not an issue with the JuMP interface and Manopt works nearly completely fine. You just didn't anticipate the influence of compounding numerical errors. This doesn't expose any fundamental...

We could change the default retraction on sphere so that it does `exp` and then projects the point for numerical stability. It would still properly reflect the geometry of the...