Date: Sat, 28 Apr 2007 12:17:57 -0700 From: Henry Baker <hbaker1@pipeline.com>
I wasn't aware of the ill-conditioned nature of this problem. References? Perhaps I should try a semilog transformation?
Strangely, and somewhat embarassingly, there's an appendix in my old thesis on approximately this problem. It was on the numerical stability of fitting gaussians instead of exponentials, but the root of the problem is the same: fitting log data means any error from the fit eventually get exponentiated. Exponentially growing error is the signature of this particular numerical disaster. I ended up supplying an initial guess for the parameters by other means, expanding in a power series around that guess, and fitting the first correction term of that series. Seemed to work ok. Even more embarassingly, MIT has put a lot of old theses online. If can stand to watch huffing and puffing by a near-insanely frustrated and pedantic grad student, look at: http://dspace.mit.edu/handle/1721.1/26852 http://hdl.handle.net/1721.1/26852 (dunno why there are 2) The gaussian fitting stuff is in Appendix C, pp. 255-261. -- Steve Rowley <sgr@alum.mit.edu> http://alum.mit.edu/www/sgr/ Skype: sgr000