TRUNCATED ESTIMATION
Kurt Kubik
Anton Schenk r
Department of Geodetic Science and Surveying
Ohio State University
1958 Neil Avenue
Columbus, Ohio 43220
1. INTRODUCTION
In this paper, we formulate the method of truncated estimation and relate it
to other well-known estimation methods. We show that truncated estimation
may be regarded as robust Maximum Likelihood Method, also yielding a
theoretical foundation for the Danish Method of estimation.
2. CLASSICAL TRUNCATED ESTIMATION
Classical truncated estimation is only applied to mean estimation: Consider
a sample of (uncorrelated and equally accurate) measurements Z;. In order
to estimate their mean, the «X largest and smallest measurements are
rejected and the arithmetic mean is computed from the remaining (l-«)%
measurements. For the case of Z, =: N(E(Z), e), the estimator has =a
variance, which is a% larger than the minimal attainable variance. In the
presence of outlyer in the measurements, the estimator is by far superior
to the least squares estimator (Andrews et al, 1972). In the following
sections, we generalize this idea to the least squares adjustment problem.
3. THEORETICAL FOUNDATION OF THE METHOD
Consider the model:
EZ) - B . u
B coefficient matrix’
u vector of unknown coefficients
E(Z) expectation of Z
and a set of uncorrelated measurements Z, for E(Z), possibly containing
outliers. We assume a truncated normal distribution:
^
m
c. exp ( - gl, (X-E(x))T (X-E(x))} for (X - EGO)
Na(x) =
(1)
0
with P (IX -EB(x)) « 2) 21-4«-
else
- 405 -