A deterministic approach is proposed for proving the convergence of st
ochastic algorithms of the most general form under necessary condition
s on the input noise and reasonable conditions on the (nonnecessarily
continuous) mean field, Emphasis is placed on the case where more than
one stationary point exists, We also use this approach to prove the c
onvergence of a stochastic algorithm with Markovian dynamics.