Minimizing the Cauchy function

A good trick (I discovered accidentally) is to use the weight  

$$\bold W \eq \ {\bf diag} \left( {1\over{\sqrt{1+r_i^2/\bar r^2}}} \right)$$ (16)

Sergey Fomel points out that this weight arises from minimizing the Cauchy function:  

$$E(\bold r) \eq \sum_i \ \log (1+r_i^2/\bar r^2)$$ (17)

A plot of this function is found in Figure 3.

 

cauchy
Figure 3 The coordinate is m. Top is Cauchy measures of m-1. Bottom is the same measures of the data set (1,1,2,3,5). Left, center, and right are for $\bar r = (2, 1, .2)$.

Because the second derivative is not positive everywhere, the Cauchy function introduces the possibility of multiple solutions, but because of the good results we see in Figure 4, you might like to try it anyway. Perhaps the reason it seems to work so well is that it uses mostly residuals of ``average size,'' not the big ones or the small ones. This happens because $\Delta \bold m$ is made from $\bold F'$ and the components of $\bold W^2\bold r$ which are a function $r_i/(1+r_i^2/\bar r^2)$that is maximum for those residuals near $\bar r$.

Module irls supplies two useful weighting functions that can be interchanged as arguments to the reweighted scheme .  

module irls {
  use quantile_mod
contains
  function l1 (res, weight) result (stat) {
    integer              :: stat
    real, dimension (:)  :: res, weight
    real                 :: rbar
    rbar = quantile (int(0.5*size(res)), abs (res)) # median
    weight = 1. / sqrt (sqrt (1. + (res/rbar)**2))
    stat = 0
  }
  function cauchy (res, weight) result (stat) {
    integer              :: stat
    real, dimension (:)  :: res, weight
    real                 :: rbar
    rbar = quantile (int(0.5*size(res)), abs (res)) # median
    weight = 1. / sqrt (1. + (res/rbar)**2)
    stat = 0
  }
}