One research group at our university is particularly interested in the statistical properties of language. One professor, Michael Ramscar, is teaching us some classes this semester on related topics. And basically the idea is that this kind of logarithmic/exponential distribution is considered optimal in an information theory point of view, since it ensures that the entropy of the whole probability distribution would be the lowest, ensuring the most effective communication in general. Some of these ideas can also be found at his blog, for example this article about the distribution of names.
P.S.: Actually he also made the argument that the "Zipfian power law" view is not entirely accurate while an "logarithmic/exponential" distribution might be a better description. He said that the shape of the curve under the Zipfian power law would be largely impacted by the sample size, while the shape of the curve under simple logarithmic/exponential distribution wouldn't, because it would only have one parameter, while the Zipfian power law distribution also gets a scaling factor besides the simple exponential factor. I'm not sure I completely got this point to be honest, but the point about entropy should still stand regardless, since I think Zipfian distribution is basically logarithmic/exponential as well.