Nonlinear Regression and 'nlinfit'

In MATLAB, all three fucntions 'lsqnonlin', 'lsqcurvefit' and 'nlinfit' are used to perform non-linear curve fitting.

To better understand the differences and similarities in these functions, consider the model function:

$y= \beta_1+\beta_2  \text{exp}(-\beta_3x)$

We wish to estimate the $\beta=\{\beta_1,\beta_2,\beta_3\}$ for the set of independents {$x_i$} and observed values {$y_i$} such that the model fits the data.

Both 'nlinfit' and 'lsqcurvefit' are very similar as we can pass the regression function to compute the parameters. 'lsqnonlin' on the other hand, solves optimization problems of the type $min_{\beta} \sum_k f_k(\beta)^2$, so we cannot directly specify the regression function and instead, an error function has to be provided.  This is shown in the code below:


modelfun = @(b,x)(b(1)+b(2)*exp(-b(3)*x));
b = [1;3;2]; %actual
x = exprnd(2,100,1); %independents
y = modelfun(b,x) + normrnd(0,0.1,100,1); %noisy observation
beta0 = [2;2;2]; %guess
beta = nlinfit(x,y,modelfun,beta0)
beta = lsqcurvefit(modelfun,beta0,x,y)
beta = lsqnonlin(@(b)err_fun(b,x,y),beta0) %err_fun = modelfun-y

All three functions generate:

beta =

    1.0071
    3.0805
    2.1418

Observations:

• lsqcurvefit is more superior in the sense that we can define the bounds for the design variable (unlike nlinfit) while inputting the observed values separately (unlike lsqnonlin). 
• Nlinfit provides extra statistics such as covariance matrix of the fitted coefficients and information about error model.
• As an alternative to defining weights for the observed values in 'nlinfit', 'RobustWgtFtn' option can choose from different pre-defined weight functions for robust regression (with robust regression, fitting criterion is not as vulnerable to unusual data as least squares weighting function.)

References:

http://uk.mathworks.com/help/stats/nlinfit.html

http://uk.mathworks.com/help/optim/ug/lsqcurvefit.html

http://uk.mathworks.com/help/optim/ug/lsqnonlin.html