Well, log does that. But so does square root also. Lots of functions have diminishing marginal returns.
I can think of two good reasons to model diminishing returns with the natural log.
Logs produce nice units in the regression coefficients. A log-lin function (that is -- log'd dependent, linear independent) says that a percent increase in X results in a unit increase in Y. Similar statements are true for lin-log and log-log, the latter of which produces elasticities.
y=ln(x) and y=sqrt(x) will both fit data in a similar manner, so it makes sense to go with the one that makes for easy interpretation.
Additionally, the natural log frequently shows up in financ...
A monthly thread for posting rationality-related quotes you've seen recently (or had stored in your quotesfile for ages).