symbolically equal to the wellknown normal equation matrix of least squares.
The limiting Variance of the estimators ü for u follows then to:
lim Vartû) = T1”!
n ^
Although theoretically interesting, this limit theorem is of limited practical
importance due to the impatience of most surveyors to make infinitely many
observalions.
5. NUMERICAL SOLUTION
In order to establish the link with the conventional least squares method
and robust estimation, we rewrite problem (5), (6) into:
I p(rji)rj;^ —' min ; r, = Z, - E(2;)
1/21 for (ri) < a (9)
with p(r;) =
0 :sfor {r,) > a (10)
and obtain a formulation similar to Hampels method of robust estimation,
(Hampel, 1973). As the problem (9) cannot be readily solved, an iterative
solution is usually used, approximating the stepfunction (10) for p by a
continuous function, and iteratively approaching its final shape (either from
the inside or the outside of p(r;)). Possible choices are:
plr;).- 1/42 for Ir! <a (11)
i 7 :
La? exp — (r-a)* for lr,l > a
increasing throughout iteration
or
1/2 for lpi: <a
per PER a (12)
o? .ı7ı8 for ril 2. 8
r!
These weight choices relate this method to the wellknown Danish Method
(Kubik, 1982) and L8 - Method (Barrodale, 1966). The numerical method will
converge to a unique solution, if sufficiently small changes are made in the
approximating function for (10) from iteration to iteration.
In the past, we experienced problems with convergence, mainly due to
improper initial values for û and r, which were computed by the classical
least squares method. Other numerical solution methods should, therefore,
be investigated.
- 407 -