feat: add new projection operators

[nperraud]

New features

• Adding proj_spsd : projection of a matrix on the symmetric positive semi definite cone
• Adding proj_lineq: linear equality projection operator

[coveralls]

Coverage increased (+0.06%) to 98.689% when pulling f809a01 on new-projections into b0eb893 on master.

[nperraud]

A small PR with minimum modification and clean commits. The one we like the most.

I moved the paramter epsilon and method in proj_b2 instead of proj because proj_b2 is the only projection using them.

[nperraud]

Here I need to add the pinvA matrix because the proj_lineq cannot compute it A is given as an operator. The automatic computation only works if A is a numpy array.

[nperraud]

@mdeff PR ready to be reviewed.

[mdeff]

Thanks! Looks good to me apart from the small comments above. (I simplified the proj_lineq test.)

[nperraud]

@mdeff Can I merge?

[mdeff]

proj_lineq is a special case of proj_b2 for epsilon=0 right? When is it preferable to use one or the other? That would be nice to put in the docs.

[nperraud]

Yes, it can be seen as a special case. If epsilon=0, you should always use proj_lineq as the implementation of proj_b2 is not really made of the case epsilon=0.

[nperraud]

Where would you put the information? In the proj_b2 doc?

[mdeff]

Great! Yep, in the proj_b2 doc.

[mdeff]

Actually, a better way is to have

        See Also
        --------
        proj_lineq : use instead of ``epsilon=0``

in proj_b2 and

        See Also
        --------
        proj_b2 : general case

in proj_lineq (between Parameters and Notes). That way, they will be linked in the HTML docs. Feel free to tweak the text.

[mdeff]

All good. Thanks!

[nperraud]

Thank you for the good and efficient review process