Implicit gradients and Dirichlet uncertainty

Goal is to compute

\[\nabla_\theta \mathcal{L} = \nabla_\theta \mathbb{E}_{q_\theta(z)}[f_\theta(z)]\]

The transform of \(z \sim q_\theta(z)\) by the cdf \(F_\theta(z)\) is uniformly distributed.

Let \(u \sim U(0,1)\) be a randomly sampled uniform value. Therefore, for some \(z \in \mathbb{R}\), we have that

\[u = F_\theta(x) = \int_{-\infty}^z q_\theta(z') dz'\]

Taking the derivative of both sides, we get that:

\[0 = \dfrac{\partial z}{\partial \theta} q_\theta(z) + \int_{-\infty}^z \dfrac{\partial}{\partial \theta } q_\theta(z')dz'\]

Hence,

\[\dfrac{\partial z}{\partial \theta} = \frac{-\dfrac{\partial}{\partial \theta} F_\theta(z)}{q_\theta(z)}\]

??? (I don’t understand how to fill in the steps here)

Profit: \(\nabla_\theta \mathcal{L}=\mathbb{E}_{q_\theta(z)}\left[\frac{d f_\theta(z)}{d z} \frac{d z}{d \theta}+\frac{\partial f_\theta(z)}{\partial \theta}\right]\)