An Uncertainty Principle for Linear Recurrent Neural Networks