Bounding the Width of Neural Networks via Coupled Initialization -- A Worst Case Analysis