Ug
out by applying the weighted least squares method after introducing
appropriate fictitious weighted observations for the unknowns/s. Now
still assume model (1) to be true. And suppose that zero observa- .
tions with weight matrix Pg are introduced for the additional
parameter vector S and then the weighted least squares method is
conducted. Thus, the following error equations can be established:
vq A H 1
= X + S -
Y 0 T 0
with the weight matrix of ond being IT oss
0 o 2.
After adjustment the estimators of the unknowns is obtained as
follows:
€ - (AAA I-GEGS £jEQ
§ = (ag + RS
in which Q, = I- AAA) a, Qas = (H'Q,,H) ;
are respectively the cofactor matrice of v and 3 obtained from the
adjustment treating s as free parameters. X and S are obviously
biased. Their biases are respectively
EX) u (ala A ut SD RSS) (14)
and = 4
BS) - 8 = ({I + QsP,) - I) s ; (12)
Assume that Ps = diag(Ps, » Psy» +++» Psmp) is a diagonal matrix
and s has been transformed so that Qaa = diag(as, » Agrees dsm5)*
Then the ith component of Eq.(12) is
-- Is; Ps.
E(S;) - 8; = anor Tio Si (13)
1 + Qs Psy
From section 2 we know that to assess the reliability is to evaluate
the biases as in Eq.(11),(12) or (13). In the application of
robust methods, the least squares estimation with all the unknowns
being free is usually also carried out as the preliminary
adjustment. So its result can be used to evaluate the biases. By
using 81, we acquired the maximum of the absolute values of the
bias for S, which are possible in the sense of probability 8, the
posteriori internal reliability for S; . It is expressed as
follows:
^
VSi
imu
Sj6{S;: 8 probable given 8]
ds, Ps;
= [Si] + 18/2 S IO. (14)
1 As; Ps;
- 371 -