Levenberg Marquardt curve fitting pt.2

[altenbach]

---

DSPGuy wrote:
Darn! Lost the link to that screenshot.  Here it is:

---

Yes, for A it probably won't be a good solution, but for a case like B, where the useful parameter range is a very small fraction of the absolute value, it could help.

An example is the simulation of EPR spectra. The elements of the g tensor are very close to 2 and typically vary only in the fourth decimal digit (e.g. gx=2.0088, gy=2.0058. gz=2.0021), but a change by 0.0001 in either one of them would already significantly alter the spectral shape.

[DSPGuy]

Yes, true.  There are lots of good reasons to scale parameters before fitting.  An application can make educated guesses about scaling as it proceeds, but it is hard to outperform knowlege of the problem at hand.