Debugging a Gauss–Newton Fit When Translating MATLAB Code to R

# What is wrong in my non-linear estimation sample code?


# What is wrong in my non-linear estimation sample code?












I am trying to reproduce the code and plot you see here on pages 8,9 and 10 which was coded in MATLAB, but I'd like to convert it to R code.

I believe I converted the MATLAB code below to R syntax and if you run it you will see a plot but it does not match the plot you see in the pdf. The R estimation is much worse.

BACKGROUND - The matlab code makes a non-linear fit to population data. It is well described on page 8. Basically it uses a gauss newton method to fit a non linear model. I am just trying to get it to work in R.

Any ideas why?

```
df =  function(p,q,a1,a2,index) #calculate partial derivatives
{

  if  (index == 1){
    value = exp(a2*p);
  }
  if  (index == 2){
    value = p*a1*exp(a2*p);
  }
  return(value)
}

      tol = 1e-8  #set a value for the accuracy
      maxstep = 30 #set maximum number of steps to run for
      p =   c(1,2,3,4,5,6,7,8,9,10,11,12,13)                #for convenience p is set as 1-13
      #set q as population of NYC from 1810 to 1930
      q =   c(119734,152056,242278,391114,696115,1174779,1478103,1911698,2507414,3437202,4766883,5620048,6930446)
      a =   c(110000, 0.5) #set initial guess for P0 and r
      m =   length(p); #determine number of functions
      n =   length(a); #determine number of unkowns
      J = matrix(0,m,n)
      JT = matrix(0,n,m)
      r = numeric(13)
      aold = a;
      for (k in 1:maxstep){ #iterate through process
        S = 0;
        #k=1
        #i = 1
        #j = 2
        for (i in 1:m){
          for (j in 1:n){
              J[i,j] = df(p[i],q[i],a[1],a[2],j)  #calculate Jacobian
              JT[j,i] = J[i,j] #and its trnaspose
              J
              JT
            }
        }

      Jz = -JT %*% J #multiply Jacobian and  negative transpose
      for (i in 1:m){
         r[i] = q[i] - a[1]*exp(a[2]*p[i]); #calculate r
          S = S + r[i]^2; #calculate the sum of the squares of the residuals
      }

      g = solve(Jz)  %*% JT  #mulitply Jz inverse by J transpose
      a = aold-g*r  #calculate new approximation
      unknowns = a  #set w equal to most recent approximations of the unkowns
      #abs(a(1)-aold(1)) #calculate error
            if (abs(a[1]-aold[1]) <= tol){
              break #if less than tolerance break
            } 
      aold = a  #set aold to a
      }
        steps = k
        f = unknowns[1]*exp( unknowns[2] * p  ) #determine the malthusian  approximation using P0 and r determined by Gauss-Newton method
        plot(p,q) #plot the measured population
        lines(p,f, type = 'l', col = "red", main = "Population of NYC 1810 to 1930", xlab = "year", ylab = "population") #plot the approximation
        title('Population of NYC 1810 to 1930') #set axis lables, title and legend
```

Here is the R plot vs matlab plot in the pdf

## Answer by SRKX (score 1)

https://quant.stackexchange.com/a/21320

You should try to run both codes step by step and see what the differences are.

The only thing that would justify different outputs would be if you're using some built-in functions by either software, which you don't seem to be doing.

If you don't have MATLAB available, you can try using Octave (which is kind of an open-source version of it and there is an online version here).