I work in machine learning, and often need to do probability computations in the log scale to avoid underflows. A nice way to do this is to use a class where the add, divide etc. operators are overloaded by their log scale equivalents (e.g. log(a*b) = log(a) + log(b)). In other words a class implementing the log semiring <http://en.wikipedia.org/wiki/Semiring>. Also useful in this context is the tropical semiring, where adding is replaced by taking the maximum (or minimum) of the arguments. I looked for these in the boost math library, but couldn't find anything. So I wrote my own. Did I miss something? And if not, is anyone interested in adding my implementation to boost math?

No you didn't miss anything, I think this would need reviewing as a new "sub-library" of Boost.Math, but I would encourage you to make your code available and seek a review if folks are interested. Cheers, John.