〈 Berger | Dillon 〉

What if I told you that I had a fast Singular Value Decomposition (SVD) function? What would you use it for?

Depends on _how_ fast. Realistically, probably what @eramirem said, amd for estimating covariance matrices. But if were somehow faster for solving Ax=b for large, pos. semi-def A than Cholesky, I have a quadratic programming solver I’d use it for.

So we are talking on CPUs only at the moment. I still don't fully understand when the new algorithm is good and when it isn't but it's beating Intel's MKL by a few times when the input matrix is rank deficient.

Is it a randomized algo? Does it perform the same way for dense and sparse inputs?

Yes to the first question. Don't yet know for the second.

Would be very nice if it’s similarly fast on sparse matrices—if it is, I may have a computer vision application it could be useful for.

Particularly useful for masking moving objects and static frames in video.