Why does the constrained optimization perspective explain why L1 produces sparse solutions while L2 does not?