Consider GF for a real-analytic objective function $L(\theta)$.


By Theorem 2.2 in 1, either $\theta_t$ goes unbounded, or it converges to a critical point. (In many cases, boundedness can be obtained via conservation laws.)



Categories: Learning