What is the impact of the size of the hidden layer in Autoencoders

Question

I am revisiting autoencoders recently and I found most of the discussions are on its relation to PCA. But there is little discussion on the size of the hidden layers.

Are