Comparing robustness of RNN with and without activation function using POPQORN

[talrub]

Hi,

If i want to compare the robustness of the Vanilla RNN to that of the Vanilla RNN without the activation function in the recurrent, should i use the closed-form global bounds from Theorem A.1 (page 7) in the appendix for both architectures?
Or should i use Theorem A.2 (page 7) from the appendix for the RNN without the activation function? ( but if i will set 'v' to 'm' in Theorem A.2 in order to get closed-form global bounds for the RNN without the activation function i will get Theorem A.1).

I would appreciate some advice on this topic.

Thanks!