A Derivation of the Score Function Estimator Given K samples, the objective being maximized is L K (x): = E h log ˆ Z i ˆ Z: = 1 K

K!1 and get the asymptotic variance: Var[ g ] = Var " X D.2 Control V ariate for Small ESS In the case ESS 1 we can write log ˆ Z as a sum of two terms: log ˆ Z = log w We will leave out a derivation for non-leading terms for brevity. D.3 Unified Interpolation We unify the two ESS limits under a unifying expression OVIS In this paper, gradient ascent is considered (i.e. However, the term (a) may dominate the term (b) . The model is trained for 5.000 epochs using the Adam optimizer with a base learning