We at SIAI have been working on building a mathematical model of this since summer 2008. See Michael Anissimov's blog post at http://www.acceleratingfuture.com/michael/blog/2009/02/the-uncertain-future-simple-ai-self-improvement-models/. You (or anyone else reading this) can contact us at uncertainfuture@intelligence.org if you're interested in helping us test the model.