HTML automatically generated with rman
Table of Contents

Name

tabnllsqfit - general purpose non-linear least squares fitting program

Synopsis

tabnllsqfit in=table-file [parameter=value]

Description

tabnllsqfit takes a non-linear least squares fit (minimized ChiSquared) of a set of datapoints (x(i),y(i)) where a functional form y=f(x;p0,p1,...,pN) is assumed. The function is not required to be non-linear in it parameters. Although a number of common pre-defined functional forms can be selected (see fit=), you can also write a small C routine, and dynamically load the fitted function during runtime (see load=).

The inputfile is an ascii file in a tabular format, where on every line there must be an assigned column for the X- and Y-coordinate(s). An optional column can be assigned to the errors in Y (the DY-coordinates, the inverse square of these being used as a weight). If no DY-column is assigned, it is assumed that the errors in Y are all the same and equal to 1. Lines starting with the ’#’ symbol are comment lines. Certain models (see fit= below) allow multiple X or Y columns.

Each fit can also produce an output file with an additional column computing the difference between the data and the model (in that order). In the special case where the user has supplied no free parameters, the residuals are still computed correctly.

This routine can also be used to plot the function, although only in 1D cases, see x=.

Parameters

The following parameters are recognized in any order if the keyword is also given:
in=in-file
(ascii) input file, a table of values from which data is taken. No default.
xcol=col,...
The column(s) from which the (independant) X-values are taken. The first column is numbered 1 [default: 1].
ycol=col,...
The column(s) from which the (dependant) Y-values are taken [default: 2].
dycol=int
Optional column from which the weights will be derived. By default this column is meant to refer to an error bar, and normally the inverse square of the errors are uses as weights for those fits where these are used (see dypow= next). [default: none assigned].
dypow=power
By changing this value, you can adjust the weights column (dycol). Instead of the normal weight = 1/sigma^2, you will be able to use them as weight^dypow. The most common use is when your dycol does not refer to an error, but to the weight directly (e.g. an intensity), in which case dypow=-0.5 is appropriate. [default: 1].
xrange=xmin:xmax,xmin:xmax,...
Range in X-values for which the fit is done. A set of min:max, separated by comma’s, can be given. This keyword cannot be used if more than one X-column is used. [default: no entry, i.e. all].
fit=model
Model name what to fit. Currently implemented are line, plane, poly, gauss1d, loren, and exp. [default: line].
order=order
Highest order of the polynomial fit (fit=poly) or number of dimension of the hyper-plane fit (fit=plane). 0 would fit a constant. [Default: 0].
out=filename
Optional output filename where the data and residuals are stored. The first few columns contain the X and Y columns, the last column contains the residual Y-F(X). See an example below how to plot the data with a model fit overlayed. [default: no output file created].
nsigma=sigma_factor
A positive number will delete points more than sigma_factor from the fit, and fit again. Multiple values can be given here, in which case multiple iterations are done. [Default: -1].
par=
Initial estimates for the parameters. For non-linear fits, or linear fits where certain parameters are fixed (see free= below) they need to be given here. Some fits (e.g. gauss1d) do not require an initial estimate and attempt to come up with a reasonable one. Otherwise all initial parameters are 0.
free=
A list of 1’s and 0’s to set which parameters are to be kept free (1) and which are fixed (0). By default all parameters are free, i.e. fitted. If any parameter is fixed, initial estimates for all need to be given (see par=). Default: all parameters free.
load=
If given, this should contain the full path to a compiled and shared object (.so) file) that contains the function to be fitted in terms of the nllsqfit parameters f and df. They must be named func_loadobj and derv_loadobj (maybe modified with fit=). See LOAD FUNCTIONS below.
x=
If given, these are the X-values for which the load function can be tested. It is useful for testing and debugging your function. It will report the X, Y and the dY/dPAR vector values. The program will exit after this, and no fitting is done. If the first two values are the same, a numerical derivate is computed (see below). Default: none.
nmax=max_lines
Maximum number of lines allocated if the input file was being read from a pipe. If not, the routine file_lines(3NEMO) is used to allocate space for the table. [Default: 10000].
tol=
Tolerance for convergence of nllsqfit [Default: 0.0].
lab=
Mixing parameter for nllsqfit [Default: 0.0 for linear, 0.1 for non-linear fits].
itmax=
Maximum number of allowed nllsqfit iterations [Default: 50]
format=
Output format. Default %g
bootstrap=
Number of bootstrap samples to take to estimate the error in the parameters. Output off all parameters and their errors are on one line containing the word bootstrap. Parameters and their errors are output in pairs. Default:0
seed=
Initial seed. See xrandom(3) for more details. Default: 0
method=g|n|m
Fitting method. Gipsy uses their nllsqfit(3NEMO) , NumRec uses their mrqfit and MINPACK uses mpfit(3NEMO) . Only first character is used, case insensitive. [Default: g]

Fit Parameters

Parameters are referred to as p0,p1,p2,p3,.....
line         y=p0+p1*x                        (same as poly,order=1)
plane           z=p0+p1*x+p2*y                   (example order=2)
polynomial      y=p0+p1*x+p2*x^2+p3*x^3...       (example order=3)
gauss1d           y=p0+p1*exp(-(x-p2)^2/(2*p3^2))
gauss2d           y=p0+p1*exp(-[(x-p2)^2+(x-p3)^2]/(2*p4^2))
exp        y=p0+p1*exp(-(x-p2)/p3)
grow        y=p0*(exp(x/p1)-1)
arm        y=p0+p1*cos(x)+p2*sin(x)
arm3        y=p0+p1*cos(x)+p2*sin(x)+p3*cos(3*x)+p4*sin(3*x) 
loren        y=p1/PI( (x-p0)^2 + p1^2 )
psf        y=p1*x^p2*sin(y)^p3+p4

Example

Here is an example of a linear fit: a straight line, with some added noise and random weights between 1 and 2:
% nemoinp 1:100 |\
    tabmath - - "%1+rang(0,10),ranu(1,2)" seed=123 |\
    tabnllsqfit - 1 2 3
nrt=0
Fitting a+bx:  
a= 1.50492 2.25695 
b= 1.00159 0.0380961
Here is an example of a 2D plane in 3D: (1+2x+3y)
% ccdmath "" - ’1+2*%x+3*%y+rang(0,0.1)’ 5,5 seed=123 |\
    ccdprint - x= y= label=x,y newline=t |\
    tabnllsqfit - 1,2 3 fit=plane order=2
nrt=0
Fitting p0+p1*x1+p2*x2+.....pN*xN: (N=2)
p0= 1.0688 0.0523819
p1= 2.01497 0.0165646
p2= 2.97436 0.0165646
And a fit to a gaussian:
% nemoinp 1:100 |\
    tabmath - - ’4+exp(-(%1-50)**2/(200))+ranu(0,1)’ seed=123 |\
    tabnllsqfit - fit=gauss1d par=4,1,50,10
nrt=13
Fitting a+b*exp(-(x-c)^2/(2*d^2)):  
a= 4.46714 0.0416026 
b= 1.13036 0.0994723 
c= 50.2263 0.845469
d= 8.70728  0.959347
rms2/chi2= 8.92068
rms/chi = 1
Here is a contrived example of plotting the function to be plotted, by fixing all parameters and computing a residual table from 0s:
% nemoinp 0:10:0.1 | tabmath - tab0 0
% tabnllsqfit tab0 1 2 fit=gauss par=1,2,5,1 free=0,0,0,0 out=tab0.d
% tabmath tab0.d - %1,-%3 | tabplot -

Here is an example of removing outlier points and fitting again:


% nemoinp 1:10 |\
   tabmath - - ’2*%1+1+rang(0,0.1)’ seed=123 |\
   tabnllsqfit - fit=line nsigma=1.5::3
nrt=0
Fitting a+bx:  
a= 1.09548 0.0775617 
b= 1.99937 0.0125002
2/10 points outside 1.5*sigma (0.152328)
nrt=0
Fitting a+bx:  
a= 1.02651 0.0452119 
b= 2.01358 0.00753531
0/8 points outside 1.5*sigma (0.080422)
Although 3 iterations were requested, after the first iteration no more points were removed, and the iterations were stopped.

Here is an example of estimating the errors via a bootstrap (resampling of errors) method. Fitting a polynomial of order 2 and taking 100 bootstrap samples:

% tabnllsqfit tab11 fit=poly order=2 bootstrap=100
nrt=0
Fitting p0+p1*x+p2*x^2+.....pN*x^N: (N=2)
p0= 3.11325 0.192787
p1= 1.97474 0.0896959
p2= 0.00157311 0.008639
bootstrap= 3.11603 0.164922 1.96464 0.086429 0.00293632 0.00820007 
             ^^^^    ^^^^^   ^^^^^    ^^^^^    ^^^^^^      ^^^^^^
              P0      dP0      P1      dP2       P3         dP3

Load Functions

With the load= keyword dynamic object files can be loaded using the loadobj(3NEMO) mechanism. The convention is that two functions must be externally visible, and named func_method and derv_method (where method is the same as the fit= keyword.

Here is an example of the file myline.c that can be used with fit=line load=myline.so and compiled with

    bake myline.so


/* File:  myline.c  */
#include <stdinc.h>
real func_line(real *x, real *p, int np) 
{
  return p[0] + p[1]*x[0];
}
void derv_line(real *x, real *p, real *e, int np) 
{
  e[0] = 1.0;
  e[1] = x[0];
}

One word of caution: if you find the program having a hard time finding a solution in complex cases, it is quite possible that this is not due to the fact that the function is complex, but due to noise or bad initial conditions.

Derivatives Check

A common debug problem is to check the analytical derivatives in your external loaded function. The x= keyword can be used to check the function values (useful to plot up the function for a given set of parameters in par=), in which also all the partial derivatives are listed:
   % cd $NEMO/src/kernel/tab/fit ; bake gaussn.so
   % echo 1 1 | tabnllsqfit - load=gaussn.so fit=gaussn par=1,2,1,0.2 x=1.1:1.4:0.1
   1.1 2.76499   1 0.882497 4.41248 2.20624
   1.2 2.21306   1 0.606531 6.06531 6.06531
   1.3 1.6493   1 0.324652 4.86979 7.30468
   1.4 1.27067   1 0.135335 2.70671 5.41341
But to check if the analytical derivates in derv_gaussn() are correct, repeating the same value of x a number of times, each parameter is incremented by a decreasing factor of 0.1


   % echo 1 1 | tabnllsqfit - load=gaussn.so fit=gaussn par=1,2,1,0.2 x=1.4::4
   # par iter Y dP  (Y-Y0)/dP   [(Y-Y0)/dP  - dY/dP]
   0 0 1.37067 0.1 1 [8.88178e-16]
   0 1 1.32067 0.05 1 [5.32907e-15]
   ...
   0 9 1.27087 0.000195313 1 [1.36424e-12]
   1 0 1.2842 0.1 0.135335 [2.16493e-15]
   ...
   2 0 1.6493 0.1 3.78634 [1.07964]
   2 1 1.43253 0.05 3.2372 [0.53049]
   ...
   2 8 1.27173 0.000390625 2.71067 [0.00396662]
   2 9 1.2712 0.000195313 2.70869 [0.00198288]
   3 0 1.82222 0.1 5.51554 [0.102129]
   3 1 1.55607 0.05 5.70808 [0.294669]    
   ...
   3 8 1.27279 0.000390625 5.41867 [0.00525903]
   3 9 1.27173 0.000195313 5.41605 [0.00263639]
You can see that the offset (par 0) and amplitude (par 1) of the gauss are well behaved (as they are linear), but the centroid (par 2) and width (par 3) of the gauss only slowly converge linearly with the step. The last number in square brackets is the error between numerical and analytical derivative.

Caveats

It will not recognize linear fits if the non-linear parameters are kept fixed, e.g. the offset 0 in fit=gauss1d.

See Also

tablsqfit(1NEMO) , tabhist(1NEMO) , tabmath(1NEMO) , gaussfit(1NEMO) , linreg(1NEMO) , nllsqfit(3NEMO) , fit.dc1(GIPSY)

Numerical Recipies in C, Ch.14

NLREG: http://www.nlreg.com

NIST non-linear: http://www.itl.nist.gov/div898/strd/lls/lls.shtml

NIST linear: http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml

fityk: http://fityk.nieto.pl/

A new scheme for calculating weights and describing correlations in nonlinear least squares fits. (Hessler, Curent & Ogren, C.I.P. 10, 186, 1996): http://dx.doi.org/10.1063/1.168569

Author

Peter Teuben

Files


~/src/kernel/tab    tabnllsqfit.c
~/src/kernel/tab/fit    example fitting functions

Update History


12-jul-02    V1.0 cloned off tablsqfit    PJT
17-jul-02    V1.1 added load=, x=, numrec=        PJT
11-sep-02    V1.1e  changes error/warning to accomodate residual writen    PJT
21-nov-02    V1.4 nsigma= can be an array of iterations    PJT
14-feb-03    V1.6 arm,arm3 for Rahul        PJT
21-mar-03    V1.7 added bootstrap=, seed=    PJT
4-apr-03    V1.8 fixed error in using dycol=, and introduced dypow=        PJT
15-mar-04    V1.8b added fit=loren and corrected lab= setting for functions    PJT/RS
21-nov-05    V2.0 added fit=gauss1d,gauss2d        PJT
24-apr-08    V2.1 added psf    PJT
7-may-10    V2.2 added grow        PJT
24-dec-11    V2.3b estimate gauss1d if no initial par given    PJT
9-dec-12    V3.0 new style xrange= with multiple segments    PJT
9-oct-13    V4.0 numrec= is now method=  for mpfit trials, add function deriv
checker    PJT


Table of Contents