Any writing on this topic without a huge section on convergence is a BS. In practice, what you need to know is that algorithm almost never converges just like a

Any writing on this topic without a huge section on convergence is a BS. In practice, what you need to know is that algorithm almost never converges just like any complex randomized algorithm. If you want to train a reasonable model, you need to look for something convex instead.