In the VAE framework where x is an input data (a vector) and z is a vector of continuous latent variables,

the posterior p(z|x) is intractable because p(x) is intractable, so we approximate it using an approximate posterior q(z|x).

But, what happens if we know p(x)?

  1. e.g., if p(x) = N(mu, covariance mtx) is known, does this make the posterior tractable?

  2. If yes, does this also apply even if the covariance mtx has off-diagonal terms (i.e., elements in the vector x are correlated) - or does it not matter?


1 Answer 1


Bayes theorem is

$$ p(z|x) = \frac{p(x|z)\, p(z)}{p(x)} $$

so if you know all the components, it's just a multiplication and division. If you don't, you usually need to take the integral

$$ p(x) = \int \, p(x|z)\, p(z) \,dz $$

and the integral is often intractable. Correlated elements have nothing to do with that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.