Except for linear fits, all curve fits in Logger Pro and Graphical Analysis are done using a non-linear n-dimensional search. As such there is an element of randomness that can prevent the fit from converging absolutely every time a fit is performed. Simply repeating the fit (or clicking the Try Fit button a second or third time) can give you convergence. The difference between a properly converged fit and a fit that ends up in a numerical cul de sac is dramatic. The bad fit will be obviously poor. You are unlikely to find a poorly-converged fit that looks sort of close.

Most data sets will converge quickly to the same parameters each time, so a user gets the impression that the curve fit is a determinate process. However, a few badly-behaved data sets will be unstable, and will show different fits on successive attempts. The root mean square error is usually quite large for the poorly converged fits, so that you have a quick way (in addition to the visual presentation) of discarding some fits. Since the fit is determined by minimizing the RMSE, if the RMSE surface has multiple minima, the fit can converge to multiple sets of parameters. The more the parameters, the more likely this is to happen.

Logger and Graphical recalculate curve fits when the file is opened; the fit results are not actually saved. As a result it can sometimes happen that you save a file with a bad fit, and then open it to find a good fit–or vice versa.

In later versions of Logger Pro we tweaked the search to favor common parameters, so convergence is somewhat better than using, say, Graphical Analysis 3.0.

If you have a particularly ill-behaved data set, you can aid the fit by choosing the manual fit option, entering very rough parameters as a starting point (within a factor of 5 is plenty close) and then clicking Try Fit. The program will start with those parameters, among others, in the search. This won’t guarantee a convergence, but it makes it more likely.