I wondering if anyone has seen this "averaging" technique before, and what its advantages might be? The values being averages are probabilities and as such have a value of 1 or less.
$\frac{1}{log(n)} \sum_{i=1}^n \log(a_i)$ where $n > 1$
This is being used in a language model claiming to be using Kneser-Ney Smoothing.