TV distance/KL divergence between $f(\theta_1, X)$ and $f(\theta_2, X)$?

68 Views Asked by At

Given a function $f(\theta, X)$ and a random variable X (such as normal distribution). Assume that $f$ is L-Lipschitz with respect to $\theta$, how can I bound the divergence (TV/KL) between $f(\theta_1, X)$ and $f(\theta_2, X)$? $f$ can be seen as a $K$-layers neural network with parameters $\theta$, and $X$ is the input.