Representing smooth functions as compositions of near-identity functions with implications for deep network optimization