Nicolas Boumal

Hello Yuval, You could create a new problem structure for each cost function? That is: you create a problem structure for your first cost function and run a solver on...

I cannot speak for the PyManopt lead team, but I imagine GPU support is not part of near future plans. We have partial GPU support in the [Matlab version](https://github.com/NicolasBoumal/manopt) of...

Thanks for the detailed description of your attempts so far. The following example shows one way of defining egrad directly: [https://github.com/pymanopt/pymanopt/blob/master/examples/dominant_invariant_subspace.py](https://github.com/pymanopt/pymanopt/blob/master/examples/dominant_invariant_subspace.py) If that does help, could you post a minimal...

Indeed, that is something to look into. Thanks for reporting it!

> It also would not be difficult to make the stochastic axis a parameter with a default value when creating the manifold Yes, that would be ideal. Under the hood,...

Hello @bamdevm, Did you have a chance to take a look at this PR? Best, Nicolas

> Adding information about the computational backend to manifolds would be a prime example of a leaky abstraction. These classes shouldn't have a notion of the particular execution context at...

Agreed with your follow up @antoinecollas, and then zooming in on: > [...] the user specifies a cost function (and thus a backend) and a manifold. The manifold uses the...

Also, as a proxy for now, you can work with the sphere in R^2n instead of the sphere in C^n, and use the map g(x) = x(1:n) + i *...

Hello, * The math behind the second-order test is described in Section 6.8 of [my book](https://www.nicolasboumal.net/book/): mentioning this here, just in case you want to try to write your own...