How to Identify Latent Causes in a Multivariate Setting

You mentioned disentangle representation and variational autoencoders. The idea is that disembedding spaces is somehow capturing latent causes. And so there's some very important question about, okay, statistically is it even possible to identify latent causes in a multivariate setting? You can show them mathematically that you get a lot of problems if you try doing that.