Forward and Inverse Approximation Theory for Linear Temporal Convolutional Networks