Unnormalized probability distributions are guaranteed to be nonnegative, but not guaranteed to sum or integrate to 1.
We need to consider a partition function to obtain a valid probability distribution.
But how do you calculate that..? Since it is most often intractable, we need to find a way to confront that by resorting to approximations.
Reference:
Deep Learning By Ian Goodfellow, Yoshua Bengio and Aaron Courville
