“…The latter is a measure of the distance between the two distributions, as we will further discuss in Sec. V-D (see [59], [60]). The analytical advantages of the ELBO L(q, θ) over the original log-likelihood are that: (i) it entails an expectation of the logarithm of the model p(x|z, θ), which, as mentioned, is typically a tractable function; and (ii) the average is over a fixed distribution q(z), which does not depend on the model parameter θ.…”