I don't understand why thanks to the reconstruction of a given probability distribution -given a number of observations- it is possible to create new observations.
I know that obtaining an unknown pdf, can give us information on other samples than the observed( with the related frequencies) and in this sense, we are able to have a new observation but in what direction we can apply this approach on data sets like the MNIST or faces that are used in generative models in ML? I suppose that we have to consider P(X) in which X is a vector in multidimensional space. Am I right? Are there books or reosurces to fix this concept?
Many thanks