Fit Functions, Log scale and Xscale wierdness

Share icon

I have data that increases monotonically in a sigmoid fashion from 0 to 100, which I fit with a user defined Hill function. If I graph the data (x and y waves), then change the x axis to log in the ModifyAxis/Axis/Mode window, and then fit the data with my function, everything looks perfectly fine, and the returned fit parameters are correct, but there is one little thing:

The xscaling of the fit_wavename wave returned by the fit function is incorrect (although it initially graphs as it should). Trying to use the findlevel function returns erroneous results, while the binarysearch method works just fine. Regraphing fit_wavename vs calculated in a new window reveals the incorrect xscaling visually. This behavior is resolved if the original data is fit before the axis log transformation.

Is this working as intended, a bug, or am I making a mistake somehow when I define my fit functions?
I just brought this up because it seemed glitchy, I would think that a simple display transform that doesnt actually change any data values, shouldn't have any effect whatsoever on the output of a fit. It was also pretty much invisible to me until I tried extracting y values, which if it is reproduceable for other people as well, could be a nasty surprise.

Graph y vs x waves.
Change display scaling on x axis to logarithmic.
Fit the data.
Output fit wave has erroneous x scaling but displays correctly in the orginal data graph. Findlevel will return erroneous results that clearly do not match the data.

versus

Graph y vs x waves.
Fit the data.
Change display scaling on x axis to logarithmic.
Output fit wave displays correctly and has the correct x scaling.

January 11, 2013 at 02:02 am - Permalink

When you do a fit with /D (autodestination) and the data are in the top graph window, Igor adds a fit curve to the graph. If the horizontal axis is a log axis, the fit curve is an XY pair with the X values exponentially spaced so that the data points are evenly spaced on the log axis. In this case, the X scaling the Y wave of the pair has no significance.

I added this feature a few years ago because I got tired of people reporting a "bug"- the left end of a fit curve looked odd because the data points of the fit curve were very far apart at the left end of a log axis. This feature has reduced the incidence of bug reports, but it is still confusing. I guess not so many people change the log axis back to linear, or I would see this question more often. I'm not sure there is a good solution to the problem.

John Weeks
WaveMetrics, Inc.
support@wavemetrics.com

January 11, 2013 at 01:47 pm - Permalink

Are you sure it has no significance? I only found it because I was calling findlevel and getting an erroneous result, the x value returned was well at the end of the wave, but the value i was looking for was clearly (both data and visual inspection) in the middle of the wave. This is while in logarithmic view, which is when I did the fit as well. Doing a binarysearch (i think its called) gave me the correct point value located in the middle, so I realized it was something to do with the xscaling. Thats the only reason I tried to regraph the fit curve by itself, and in this case can see that it graphs way up the x axis where it doesnt belong.

I will note again, the fit parameters returned by the fit function were correct despite the (invisible) xscaling problem, but findlevel returns a bad answer.

January 11, 2013 at 02:33 pm - Permalink

If the trace is an XY pair of waves, the X values come from the X wave of the pair, and the X scaling is not used. I think in this case CurveFit and FuncFit omit the step of setting the X scaling because it is actually wrong for the XY pair. There is no correct way to set the X scaling so that X values from scaling are meaningful for and XY pair.

John Weeks
WaveMetrics, Inc.
support@wavemetrics.com

January 14, 2013 at 05:03 pm - Permalink