[dongecko]

The Frank-Wolfe algorithm for optimization. Christian Bauckhage has a couple of papers, where he uses it to connect optimization problems to the fix states of a certain type of recurrent neutral networks (echo state networks).

The Marchenko-Pastur distribution derives from random matrix theory a nice theoretical border to estimate if a principal component is more noise then data.

Also I am a huge fan of all sorts of embedding/projection/matrix factorization algorithms and I use them quite regularly.