In building probability and likelihood models, I often encounter the issue of transforming from either to their logarithms and vice versa. Code would be much simpler and less error-prone if the appropriate domain was defined, the natural arithmetic operators were used, and the conversions occurred (or could be prevented to catch errors) as necessary, but without necessarily requiring explicit bookkeeping to distinguish these quantities and their logarithms.
This suggests a (small) set of types with appropriate conversions, much as in the units library. However, probabilities are of course unitless.
Should this idea be developed as a separate library or does it make any sense to fold it into the existing units framework?
We've tried to make mcs::units as flexible and extensible as possible in order to accommodate a wide range of value types; as you point out, since probabilities and log probabilities don't have associated units (by definition, since taking the log of a unit doesn't make any sense from a dimensional analysis standpoint), so I don't think this is something that should be directly supported in the library itself. That being said, mcs::units can be used for tagging and conversion in this way if you're willing to specialize the conversion_helper class template for your probability/log_probability classes. If you want to try this, I can walk you through the steps... Matthias