# How can we decompose generalization gap as done in the paper “Generalization in Deep Learning”?

I am reading a recent paper "Generalization in Deep Learning" and I am unable to understand a step. In this step, they first take neural network as a direct acyclic graph(DAG) and described output of neural network as follows,

and then they decomposed generalization gap as follows,

without any explanation. Here z is,

Can someone explain this or point me towards the right direction.