Hi there,
I am relatively new to numpyro, and I am exploring a bit with different features. In one scenario, I am using Gaussian copulas to model some variables, one of which has a discrete marginal distribution (say, Bernoulli).
In my pipeline, I would generally start from some latent normal distributions with a dependent structure, apply PIT to transform to uniforms, then call icdf from the marginal distribution to get back the target marginal distributions. I understand that ICDF (and the entirety of this approach) might be a bit trickier for discrete rather than continuous distributions, but I am a bit puzzled by the fact that ICDF is completely absent from any discrete distributions I can find in the documentation (except DiscreteUniform).
Is there a reason for this? Are there any standard workarounds, rather than direct computation? In the github there is an approved merge request (#1652) for this, limited to Bernoulli, from 2023. Has there been a change in philosophy since then?