Inverse Approximation Theory for Nonlinear Recurrent Neural Networks