2 Preliminaries Weconsidertheproblemoftrainingofatwo-layerReLUneuralnetworkwith mscalarinputsanda singlescalaroutputusingtheleast-squaresloss: min