Michael Baudin

- [ ] 3.1 Defining the objective function based on the squared residuals on probabilities might be easier than the squared residuals on quantiles. This may weight the objective giving...

OK. Let's design this. ```python ncomponents = 2 algo = PrincipalComponentAnalysisAlgorithm(ncomponents, sample) algo.run() result = algo.getResult() transform = result.getTransform() # Function inverse = result.getInverseTransform() # Function ev = result.getEigenvalues() #...

There are several papers on the rank-1 update of the SVD. - Gandhi, R., & Rajgor, A. (2017). Updating singular value decomposition for rank one matrix perturbation. arXiv preprint arXiv:1707.08369....

> The proposed class could be implemented using the `KarhunenLoeveSVDAlgorithm` instead of low-level algebra classes, in order to benefit from the compressed matrices or the stochastic algorithms to compute SVD....

> If z and z' are defined on RN, the pair (z,z') is defined on R2N You are right, because the vector is in a cartesian product of two vector...

@josephmure : May I suggest to finalize this review?

@jschueller : I think this is ready for merging.

@ariste-sarl, @josephmure : This PR might be interesting for your graphical works.

- This PR should be paused so that @ariste-sarl can push this into a C++ code. - It is not necessary to convert to PolygonArray: a Polygon does the job.

This is useless after #2462. This is why I close this PR.