only used for analysis and discussions about gross error
location rather than to calculation of value of Qvv.p
matrix.
Equation(7) and (8) satisfy the condition tr(6)=r, where
r is the redundant number because the main diagonal
element of equation(8,b) is
E m
AB SIEH au Sich Dir Eu (Sho. (exi)
m -—
x m = .
and tr (AG) 7, (gi 7D gu Si E aa gy "Si +. (Kat)
Matrix Qvv.P is a singular idempotent matrix from which
we get
S 22 AKXi)
gii -E sa "Ski ‚Bi "Bi 7yzBik Exi o ^
m -
i @i- Dei -§ =-8 X a $us (Kéj3
= 3
so tr( G) = 0 and tr(G) =r
The expanded expressions of matrix Qvv.P are very useful
tool for the discussion of localizing gross errors. The
present author gave the results of capability of
localizing gross errors about iterated weight least
squares method ( Wang Renxiang, 1988b ) and gave the
conclusion about the power value of negative power
function that taken 2.5--4.0 is better then 1.0--2.0
in ( Wang Renxiang, 1989c ) by using the first order
expression of matrix Qvv.P. In this paper, we will
further extend the applications about the expanded
Qvv.P matrix.
2. CALCULATING Qvv.P MATRIX BY USING
FAST RECURSIVE ALGORITHM
As iterated weight leaset squares is performed, Qvv.P
matrix will be changed with changing weight matrix P.
It is time consuming for caculating Qvv.P matrix
according to the egation(2), when the normal equation
with large dimension. If matrix Qvv.P is computed in
first iteration, then the sequential iterated Qvv.P
matrix can be calculated by using equation(6) and get
o” = firt -œ (9)
where
m,k = up foot-not, denoting the number of repeated
computation
i = the number of row and column of changing
diagonal element of weight matrix P
G® =1-A (APA) 1 A-P , where P is the initial
weight matrix P
G = computed Qvv.P according to the fast recursive
algorithm with changed weight matrix
^ Aere Sig
Jon 0
rfi 4t ee (10)
0. 0-Sigki 0-1
x
If all the elements of matrix P have got their own
imcrement, calculate with the fast recursive algorithm
according to k=i, otherwisse, if some elements of matrix
P nothing changed,then the calculating will have to jump
over the order.
Qvv.P matrix calculated from the normal equation with
large dimension by equation(9) is more fast then by
equation(2), because after the first iteration, there
is no inverb matrix in the computation.
3. THE PROPERTIES OF WEIGHT ZERO RESIDUAL
For localizing gross errors, in some robust estimate or
iterated least squares method, at least the last
iteration is always weighted zero value ( or near zero )
to the observation of which residual is rather large
value in absolute. In this paper,we defined the residual
computed with weighted zero to the observation as
so-called ' weighted zero residual ' and symbolized Vi
( Stefanovic, 1985 called ' swep residual 3. Tt is
necessary to investigate the properties of weighted zero
residual for further discussion about gross errors
localization.
Asumming that P = I. Firstly, we assign zero weight to
observation i, ie. spi=-1 and Si--l/ gj. According
to equation(9) we have
1
G! - Ti -G
Using equatio(9) and (10), we get
$= g. =1 j=1,m
Qq,
From above, we know that all the elements of i th
column are enlarged by a factor of 1/g; If only one
observation i is assigned zero weight then its weighted
zero residual y; can be calculated from vi directly.
ie.
Vi=V;/% (11)
On the otherhand, the elements of k th column will be
BH Spada faut k=1, m, Ki
Using
Oik = dik/ y didkk 7 du / / qi Gre
1 2
then Bix = kk - Pj, k* Aki =(1-P; kx
Let P =0.7 9 = 940,5
In this cse,
we get 20,0. 1 g. f. 25
if gross error V, is included in
observation k then to compare the magnitude of
residuals, the observation i will be larger then the
observation k. So, when correlation coefficient of
residuals is big. The gross errors localizing is not
reliable.. If the absolute value of residual is used as
statistical equantity.
Secondly, we further assign zero weight to observation k
, le. 8p,--1, Sx »-1/8k , then we get
G6 -G!- T .Ti-G? - T]. Gg
According to equation(9) and (10), we have
Bü =(Qij + Quk — Qk; * Aix)/(Qii | di - dii * dik)
x; =(Qk; + Di - ij: yfgig "Ok - gk da)
and gi-1 ,B«-1 > Bk 20 , 8 =0
Now there is nolonger correlation between weighted zero
residules Vi and y, In the same way, we obtain
: zx PRR 2
Bij z (Qj * di - 9k; :Pi.k/di.dkk y (17 83*x) Bi *-Qkk (12)
&ij = (4; > di = 9, -Obik/du- di, )/ (1- Pi*x) Bii * dy
In the following,
we take two conditions for further
discussions.
3.1 Let Mi,k #0
If Pisx=0 , then
dij = 4i;/ di, 8kj = àj/ dk