Re: [Help-gsl] fitting 3-peak gauss

[Patrick Alken]

Hello,

  I've corrected your code a bit (attached). Also included is a plot of 
the data and fitted model. The output I get from running the program is 
below. Note that I rescaled the x axis by subtracting the mean of all x 
values - this makes the problem better conditioned and easier to 
converge. You can add back the mean later to get back to the original 
scale of the problem. To solve the problem with the original X values 
you'll probably need either a very accurate initial guess vector, or an 
analytic Jacobian function.

====
$ ./nlfit > dat
summary from method 'lmsder'
number of iterations: 7
function evaluations: 80
Jacobian evaluations: 0
reason for stopping: small step size
initial |f(x)| = 6393.58
final   |f(x)| = 2100.08
chisq/dof = 19092.3
Amp      = 946.59642 +/- 65.37942
mu       = -2.41481 +/- 0.01003
sigma    = 0.12576 +/- 0.01003
Amp      = 5620.88963 +/- 67.58526
mu       = 0.13784 +/- 0.00163
sigma    = 0.11744 +/- 0.00163
Amp      = 2608.71157 +/- 65.71170
mu       = 1.27723 +/- 0.00361
sigma    = 0.12409 +/- 0.00361
status = success
====

On 03/21/2016 09:26 AM, Hinxx wrote:
> Hi all,
>
> I'm new to GSL and numerical methods so please bear with me. I'm using 
> GSL-2.1.
>
> I'm trying to fit a 3-peak gaussian curve onto experiment data that I 
> have (800W5-95kv_ROI.txt).
> I started from expfit.c and nlfit.c and modified the code to handle 
> gauss curve.
> For my initial tests I also used code/data from test_gaussian.c. I've 
> played around with the test set and managed to get proper amplitude, 
> sigma and peak location(s), also when extending test set to hold three 
> peaks (copies of first) and code to handle three peaks.
>
> When using the real data of some 240 points, I'm seeing weird 
> behavior. Solver bails out after 1st iteration and I get my initial 
> guessed parameters back, unchanged.
>
> If I fiddle with the data set a bit to make all the points between the 
> peaks near zero I get better results, bat still good enough.
>
> Maybe it is worth noting that I also have some sample Matlab code that 
> manages to find the fit on the presented data set. It is using code 
> from http://www2.imm.dtu.dk/projects/immoptibox/.
>
> I'm attaching the modified GSL code and real data set if someone might 
> be able to help out.
>
> Thanks in advance,
> Hinko

expfit.c
Description: Text Data

nlfit.c
Description: Text Data

plot.png
Description: PNG image