This is an aside, but if have a general transform, is doing Jacobi SVD the
fastest? Did anyone investigate using the newton type methods? eg.
https://www.hindawi.com/journals/ddns/2015/649423/
That's a good question. What's sure is that Jacobi SVD is not the fastest among all possible approaches but (1) it's already in Eigen and (2) it's safe.
On my side, I've also written a polar decomposition code based on the direct 3x3 symmetric eigenvalue solver we have in Eigen. IIRC, it's about 5x faster, but it's also numerically less accurate and robust in corner cases.
It would be very interesting to implement the method described in this paper to see how it compares.
gael.
-Edward