DNA shuffling would be more efficient if you already have a big repertoire of variants (which have a lot of point differences among them; see [Cohen, 2001][1]). When you do not have that, then introducing point mutations using error prone PCR would be better.

A detailed mathematical framework to compare these two methods, is not available. However there are many mutation models and there is at least one model for DNA shuffling too ([Sun F., 1999][2]). 

Basically, for a point mutation experiment the number of mutants will depend on the error rate of polymerase. In a DNA shuffling experiment you are just creating random combination from existing variants. So, which one is more efficient depends on these parameters.


  [1]: http://dx.doi.org/10.1126/science.293.5528.237
  [2]: http://dx.doi.org/doi:10.1089/cmb.1999.6.77