Towards Evading the Limits of Randomized Smoothing: A Theoretical Analysis