The reduce operation computes many times the exponential and the logarithm. It induces numerical instabilities. In addition, the logsum used to replace it is quicker, as it makes less calls to transcendal functions (exp, log).