fixedrankfactory_3factors egrad2rgrad issue

[Florent Bouchard]

Hello,

I'm new to manopt and to riemannian optimisation in general so maybe I'm completely wrong but I think that there is a problem in the function converting the euclidian gradient in fixedrankfactory_3factors.
Indeed, rgrad obtained after a call to the function is not in the horizontal space of the manifold (it can easily be checked using check gradient where the residual is far from 0).
If I'm not wrong, adding the line: rgrad = M.proj(X,rgrad); at the end of the function egrad2rgrad solve the problem.

Can someone confirm that ?

Thank you.

Florent

[Nicolas Boumal]

Hello Florent,

Thank you for your message, and sorry for the inconvenience!

We'll look into this rapidly. (I don't know that geometry very much myself, I'll contact Bamdev.)

Cheers,

Nicolas

[BM]

Hello Florian, 

Thanks for the interest in the toolbox.

It could be a numerical issue. Mathematically, one can show that the egrad2rgrad operation in the fixedrankfactory_3factors factory is correct. Could you share us the code that used? 

I have also attached a sample file that showed correct result last time when I ran. Could you verify this?

Regards,

Bamdev

P.S. I am on a trip and I have limited Matlab functionality. 

[Florent Bouchard]

Hello,

Thank you a lot for your answer and the time you accorded me. I tried your code and it worked fine. It also enabled me to see that I  did a silly mistake because I used the manifold as if X = X.L*X.S*X.R instead of X = X.L*X.S*X.R'. And I did not realize it as I work with full column rank matrices.

Anyway now that I have changed that it works just fine. Sorry for the inconvenient and thank you again for your time.

Cheers,

Florent

[Nicolas Boumal]

Cool!

Thanks a lot, Bamdev! Enjoy your trip!

[BM]

@Nicolas: Thanks.

@Florent: Nice. Do let us know how the toolbox is useful to you. Also, let us know if you have any other query. 

Regards,

Bamdev