Netherlands Geodetic Commission: Publications On Geodesy New Series
Netherlands Geodetic Commission: Publications On Geodesy New Series
GEODETIC
COMMISSION
PUBLICATIONS O N GEODESY
NEW SERIES
NUMBER 2
VOLUME l
1963
PREFACE
CONTENTS
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The adjustment
2.1
Method of adjustment
2.2
3.1
3.2
3.3
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
10
12
16
18
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. .
. .
. .
Computational aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 38
FinalRemarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1 INTRODUCTION
2 T H E ADJUSTMENT
2.1
Method of adjustment
I t is well known that the two main forms in which an adjustment problem may be
expressed are :
a. condition equations,
b. observation equations.
Any adjustment problem may be split up into different steps, and each step may
be adjusted according to a. or b., so that even the most general adjustment problem
can be solved with the rules of computation of a. and b., provided they are used
in their general form, i.e. for correlated observations. A problem in the form of a.
was called by TIENSTRA
Standard Problem I, b. was called Standard Problem 11,
names which we shall also use here.
Adjustment by steps does not as a rule result in less computational work, but
splitting up the total adjustment affords greater possibilities for the analysis of
results, in particular when use is made of test methods of mathematical statistics.
The different steps can usually be interpreted directly, so that a clear outline of the
whole computation can be kept in mind. Smaller matrix inversions are involved in
the computation of the precision of the final results. Alterations in certain parts of
the adjustment or the addition of new parts can as a rule be effectuated easily.
The way adjustment by steps was used in the adjustment of U.E.L.N. can be
briefly described by saying thai the total net was split up in four partial nets, which
were separately adjusted (first step) and later joined together (second step).
First step. Let Fig. 1 represent a partial net with the antennae that connect it to
adjacent partial nets. The partial net is adjusted according to the method of observation equations. Let 01 be a datum point in the partial net, and let the differences
in geo-potential numbers (g.p.n.) between the nodal points and 01 be denoted by ra.
The observed differences in g.p.n. between the nodal
5
'4
points are fi their least-squares corrections U;. (The
underscore-denotes a stochastic quantity.) The observation equations are then :
etc.
or, in matrix notation:
Figure 1.
Ax = f + y
The weights of the observations are given; let their matrix be G . Since it is assumed
that there is no correlation between observations, G is a diagonal matrix. The normal
equations, which can be easily established without writing down the observation
equations, are :
A ~ G A =
~ ATG~
Let
(ATGA)-1 = Q
Then
QATGf
and Q is the matrix of weight coefficients of the variates g".
?4
fss
p =fr,+fqa
If the weight coefficients are denoted by p,frsetc.,
we have :
fYfT8
=fTQ,fT,+fqa,
f,s+2fTq,
f4s
l
Figure 2.
The weight g,,, ,, of p, which we can more simply denote by g,, in this case is
obtained by reciprocation :
I t is necessary to find the weight coefficients of the g.p.n. of points like g after the
adjustment. We define :
PUBLICATIONS ON
I t is seen that
a,"+a: = l
The g.p.n. of q will be denoted by -yq; it can be found in two ways
+p
yq = gr +frq
yq = gs-f
-
qs-_uqs
Multiplying the first equation by a: and the second by a: and adding the results,
we obtain
J
a:gT
+a:gs +a:f'"
a:fqs
b p + ...
in which bt is some coefficient.
For the non-diagonal weight coefficient yq, xt we get:
gt
yq, xt
a: xT,xt
According to (2) :
Similarly :
Hence
Y,X = DQ
If g' is another intermediate point, we have to distinguish two cases
a. q and g' are points on the same levelling line,
b. q and g' are situated on different lines.
I n the first case we can compute yq,yql as follows:
=frq,frq
a;frs, frs
frq, fq's
fqs,frqf
= fqq',
=
Figure 3.
0
p
fqs, f q t k
fqq'
,frq',frqf -frq,frq =
(ay1-a:) frs,frs
fq's,
fq's
a:'frs,
frs
S'
we obtain by a similar
I t is easily seen that, using the matrix D defined before, we have in general:
Y , i = DQD~
For points situated on the same line, an addition according to (2.1-4) has to be made.
We finally have to compute the weight coefficients of the g.p.n.'s of the junctionpoints with other partial nets, e.g. 14 in Figure 1. Denoting these g.p.n.'s by z, we
see :
etc.
The matrix of weight coefficients of all the variates g, y and g is finally
in which
Q has been defined before
214,
,/
10
11
Figure 4.
x4
All information concerning the adjusted partial net can now be expressed in the
new variates S, y and g and their matrix of weight coefficients B.
T h e situation is represented in Fig. 4.
1, NR. 2
Remark. If the partial net had been adjusted according to Standard Problem I, we
would have arrived at exactly the same result by writing the variates 3,y and g as
functions of adjusted observations, e.g. :
35 =
36
-(f5+u5)
+ (f6+u6)
-(f8+uS)
etc.
and then computing the weight coefficients of these functions by the appropriate
formulae of least squares theory. The numerical establishment of the functions is an
easy but rather tedious work, in which great care must be taken to avoid blunders.
Standard Problem I1 was chosen for the first step mainly to avoid this work. The
resulting matrix of normal equations had a higher order than the one that would
have resulted from a treatment according to Standard Problem I, but since it
could still be inverted, without partitioning, by the machine used, this offered no
inconvenience.
Second step. The unification of the partial nets was done by applying Standard
Problem I for correlated observations. I n Fig. 5, three partial nets have been drawn.
There is no more reason to distinguish the different types of observations g , and
~ g
from each other, and we shall denote by - the vector of all observations g, -y and g
of all partial nets together.
The conditions are easily established, e.g.:
etc.
or, in matrix notation:
U(_p+.)
Hence :
U.
-=t
= 0-Up
in which B1, B2 and BQ,denok the matrices of weight coefficients of the "observations" p in the first, second and third net respectively.
here is no correlation between variates belonging to different partial nets, so the
non-diagonal" sub-matrices in C are zero.
The normal equations are
Y5
UCUTk= t
k = (UCUT)-lt
The corrections are :
5 = CUTk
The matrix of weight coefficients of the correlates is
k, k = (UCUT)-1
and consequently that of the corrections
E, E
= CUTk, k UC
The matrix of weight coefficients of the adjusted observations is (see e.g. [13],
p. 112):
(P+), (P+) = p, p
- E, E =
C-CUT(UCUT)-1UC
The ultimate aim of the adjustment is to find the adjusted differences in g.p.n. c
between each point and a certain datum point. Let this datum point, which serves
as a reference point for the whole net, be 01.
Then it is seen in Fig. 5 that, e.g.:
- --p 3 +c3
c14 = 14
p +cl4
c3
c23
(plc'+ $0)
- (p21+g21)
(p23+g23)
etc.
In matrix notation:
E = Q+E)
The matrix of weight coefficients
c,is then
Figure 6.
m fromB1
Figure 7.
fromB2
Figure 8.
frornB3
PUBLICATIONS ON
GEODESY,
NEW SERIES,
VOL. 1, NR. 2
Most of the variates p will have a coefficient zero in all the condition equations.
Because they are correlated with variates whose coefficient is not zero in all the
condition equations, they will receive corrections from the second step of the adjustment. U p to and including the computation of the correlates the second step can
be treated as if it pertained only to the variates whose coefficient is not zero in all
condition equations (in Fig. 3 those indicated by a heavy line). This is easily seen
if we re-arrange the order of the variates, so that p', . . . , p m are the ones that have
non-zero coefficients, and pm+', . . . , p n the ones with only zero coefficients. The
matrix U is then built as in the diagram in Fig. 6. The matrix of weight coefficients
is as in Fig. 7.
I t is easily seen that UCUT = UICIIU:.
I n the computation one will of course as
much as possible take advantage of the many zero elements in the different matrices.
The matrix C deserves special attention in this respect. If the above-mentioned
re-arrangement of variates is made, C is built up as in Fig. 8 (considering the
example of three partial nets).
The formulae for the total adjustment are recapitulated on page 9 in matrix
notation, for the case of four partial nets.
Remark. The second step could also have been performed by the method of
Standard Problem I1 or rather by what TIENSTRA
has called Standard Problem I V :
condition equations containing unknowns.
If in Fig. 5 unknowns X 2 and g 3 are introduced for the differences in g.p.n.
between 0 2 and 0 1 and between 0 3 and 0 1 respectively, we have:
etc.
or, in general
=
vp-
1,1=vp, pvT
By inversion we obtain the matrix of weights, after which the whole problem can
be treated as Standard Problem I1 for correlated observations. From the corrections to 1, the corrections to p
- can be derived, the final differences with respect to
01 are, e.g.:
4.5 = p 5 + g 5
4.23
= x3+p23+g23
-
etc.
An application of the law of propagation of weight coefficients results in the
weight coefficients of the variates 4.. Although this application is simplified by the
well-known properties that any X is not correlated to any g and that
p
.
(pk+&k)= Pi, pk - E ~ ,
the number of matrices involved and their computation make this solution somewhat complicated from an organizational point of view. Therefore the method of
condition equations was used in the second step. Test computations indicated that
the total amount of purely computational work for both methods is about equal,
a t least in small nets.
RECAPITULATION OF FORMULAE
l st step Observations
Weights
f
G (diagonal)
Observation equations Ax = f + y
Normal equations
ATGAx = ATGf
Inversion
(ATGA)-1 = Q
Solution
"Sum of squares"
Intermediate points
-=
pi
G - D Q
Weight coefficients
.=
(E)
y,y
P
z, Y
= DQD~
g]
y, z
P
z, z
10
Weight coefficients
1, NR. 2
B10 0 0
0 B20
O O O B q
Conditions
U(P+E) = 0
Ug
-U2 = t
Normal equations
UCUTk = t
Correlates
k = (UCUT)-1
Corrections
g = CUTk
Weight coefficients
k, k = (UCUT)-1
P
E, E =
"Sum of squares"
CUTk, k U C
(P+), (P+) = P, P - E, E
(p+E), (p+E) = C -CUT(UCUT) -WC
Functions
E = A(P+E)
Weight coefficients
c,= A (p+),(p+)AT
11
cients 1 / P given there. These data have been established by using the different
values of the variance t 2 of 1 km levelling given by the various countries, expressed
in mm2/km (see [12], page 25). But the adjustment is executed with observations
expressed in geopotential units, so that the use of variances expressed in mm2/km
is theoretically not correct. However, the difference has little practical significance.
The adjusted differences after the first step in g.p.n. between the points of each
partial net and the respective datum points are given in Table I *), column 2. The
corrections to these differences resulting from the second step are listed in Table I,
column 3.
The final geo-potential numbers of the points with respect to the Amsterdam
datum are listed in Table I, column 4.
The corrections from the second step indicate the influence which the partial nets
exert on each other. A clear example is the partial net E, which is tilted by the
connection to the other nets: the western part is depressed, the depression diminishes
from 0.042 g.p.u. in 416 to 0.005 g.p.u. in JM-42. The eastern part is lifted by an
amount diminishing from 0.048 g.p.u. in 420 to 0.002 g.p.u. in 405. A "neutral
axis" may be imagined to lie between 417 and 418, between 421 and 422, through
429, between 427 and 428, ending between JM-44 and 406.
The corrections from the second step are biggest along the edges of the partial
nets. I t should be noted that the dividing lines often lie across polygons with large
misclosures, in the Pyrenees and the Alps. The decrease of the corrections for points
away from the edges is illustrative (corrections in 10-3 g.p.u.) :
I n partial net F:
Point
Correction from 2nd step
I n partial net C:
Point
Correction from 2nd step
I n partial net E:
Point
Correction from 2nd step
3.1
The four partial nets can be assumed to be fairly homogeneous as far as their
accuracy*) is concerned. One may now ask the following questions:
1. Are there such differences between the results of the partial nets that a conclusion can be drawn on the occurrence of model- or systematic errors in levelling?
2. Have the different countries given a good estimation of t2 (variance per km
levelling line) ?
An objective guide to the answers may be obtained by statistical tests.
A basic assumption is that our original observations are normally distributed and
mutually independent in the probability sense. Let oi be the standard deviation of
an observational quantity Pi.
- Weights gii are determined by introducing the constant variance factor 02, according to:
02
(0i)2 = gii
The weight formula for the original observations is
200
If the observations are adjusted by the method of least squares to fulfil1 b conditions, corrections _vi are found. The quantity
-E = [gii_vi_vi]
can then be computed; in the diagram on page 9 it has been called "sum of
squares" and the well-known formulae for its computation have been added. Of
course the notation [gii_viyi]is not appropriate in the case of correlated observations.
* ) The terms precision and accuracy are used in accordance with the definitions given by CHURCHILL
EISENHART
in: The reliability of measured values, Photogrammetric Engineering, Vol. 18, page 545.
See also: M. G. KENDALL
and W. R. BUCKLAND,
A dictionary of statistical terms, 2nd ed., Edinburgh, 1960.
U.E.L.N.
is an unbiased estimator of 0 2 .
The stochastic quantity
>l
14
1, NR. 2
Number of
conditions b
D-E
D-C
D-F
D-I1
E-C
E-F
E-I1
C-F
11-C
11-F
1-11
Starting with an analysis of Table A we see that the nets D and E, the total first
step, the second step and the adjustment as a whole lead to a rejection of the hypothesis of non-occurrence of model-errors, or to the assumption that several of the
estimates of t 2 given by the different nations, are too small. I t is remarked that all
F-values in Table A are greater than one, which fact also indicates that the conclusion drawn above is based on good evidence.
The test for the partial net F has been given for completeness only; it has no real
meaning because according to [12], page 39, t 2 for this net has been computed from
the polygons contributed to U.E.L.N. themselves. The resulting dependency invalidates the test.
If we now analyse Table B, it appears that the occurrence of model errors is not
15
very likely and that the results of Table A are more probably due to the value given
for t2 being too small. The fact that all F-values are greater than one need not cause
surprise; this is a consequence of the method of testing in which the greatest value
of G 2 is always put in the numerator. Nor does Table B give rise to the conclusion
that the used formula for the weights, inversely proportional to the length of the
levelling line, is wrong.
I n Table B the different partial nets are compared among themselves and with
the second step, and the whole first step with the second. The described test is not
suited for comparing the partial nets with the first step as a whole or with the whole
adjustment, since the required stochastic independence is not then present. T h e
result of these tests is that there are no significant differences.
In conclusion, it is evident that the value u2 = 200 cannot be used for evaluating the precision
of the results of the adjustment, since the estimate $2 computed from the adjustment is significantly
higher. This value $2 = 370 will therefore be used in the sequel to compute standard deviations
of adjusted geo-potential numbers.
Apart from the described tests, which follow from the method of adjustment, one
can easily test each polygon separately. Let t q e the misclosure of polygon nr. Q.
The weight coefficient gee is the sum of the weight coefficients of the sections of the
1
polygon; the weight of te is - = g,,. From the adjustment of the polygon follows
the estimator
g@
while
Nine out of 68 polygons turn out to have a significantly too large misclosure, namely
te
Nr.
Nr.
Nr.
Nr.
Nr.
Nr.
Nr.
Nr.
Nr.
2001
2010
3002
4003
4012
9003
9009
9010
9013
(nodal points
( ,,
,,
( ,,
,,
( ,,
,,
(
(
(
(
(
,,
.,
,.
,,
,,
,,
,,
,,
,,
,,
39.31.10-3 gpu
27.85
-204.34
123.91
69.36
- 65.11
- 152.92
- 179.76
75.74
-
+
+
+
1 Critical value
26.41
23.43
154.60
122.34
36.95
44.64
126.73
129.81
69.37
16
3.2
SERIES,
VOL. 1, NR. 2
Precision of results
The weight coefficients of the adjusted geo-potential numbers after the total adjustment were computed according to the method explained in Section 2.1. Not all the
weight coefficients are printed in this report, but only the most important ones,
those pertaining to the mareograph stations. These stations are denoted by RM,
and their geo-potential numbers are obtained by adding the observed g.p.n. differences of the antennae J M - R M to the g.p.n. of the J M points as found from the
adjustment. The weight coefficients of the g.p.n. of the R M points are found in the
same way as described in Section 2.1 for antennae-points. They are given in Table I1
(see folding page at the end of this publication) ; the unit is 10-6 (g.p.u.)2. The weight
coefficients pertaining to the mareograph stations of the Northern Block have been
computed by using the weights furnished by the Commission.
By multiplying the diagonal weight coefficients by the variance factor 370, one
obtains the squares of the standard deviations of the geo-potential numbers of the
mareograph stations with respect to the Amsterdam datum. The square of the
standard deviation of the g.p.n.-difference between two mareographs, c"-cB is,
according to the law of propagation of weight coefficients:
in which g"" and gBPdenote the diagonal weight coefficients pertaining to ca and cB,
and g"B their non-diagonal weight coefficient.
When interpreting the thus obtained standard deviation it should be borne in
mind that the value 370 is an estimate. We may obtain a 95% confidence interval
for the mean 3 2 of 62 by means of the following critical values obtained from [7],
Table V I I :
1.75
and
F 0 . 9 5 ~ 6 8 . 2 3W
1.85
R E P O R T ON T H E ADJUSTMENT O F U.E.L.N.
L
o
o
A* O+
Sm : o
xk 9:
mm
m d
$r 2o
8v
*t
"
O
d
o"
- . ..
0
oo
6)
.*
**
Figure 10.
,000 em r
$00
aaa
W:
:m
:ii
:NW
4f fit5
18
PUBLICATIONS ON
GEODESY,
NEW
SERIES, VOL.
1, NR. 2
the difference between the variance factors is not quite significant, so that the abovementioned difference of 30% is rather acceptable in view of the uncertainties of the
estimates. The resulting standard deviations are likely to be on the high side as far
as the Northern Block is concerned, but in view of the isolated position of the
constituent parts of this Block, this can be considered as a welcome safety-margin
which prevents over-estimation of the accuracy.
I n Fig. 10 the increase of the standard deviations in different directions, before
and after the adjustment, is pictured. The values before the adjustment have been
obtained by considering the observations along the most direct route between the
points considered, multiplying the sum of their weight coefficients (as furnished by
the Commission) by 370 and taking the square root.
Fig. 11 (sre folding page at the end of this publication) pictures the standard deviations in the whole Central Block by means of contour lines of a surface whose
height at a certain nodal point is equal to the standard deviation of the g.p.n. of
that point. Its visualises roughly the precision of the net; outside the mentioned
discrete points the surface has of course no significance.
3.3 Power of the tests used
I n Section 3.1, F-tests were used to investigate whether model errors made themselves felt in the adjustment. The level of significance adopted was 5%, which means
that there is a probability of 5% to make a "Type I error", i.e. to reject a true null
hypothesis. We will now deal with the following question: if our null hypothesis is
wrong, what probability do we have of nevertheless accepting it and make a "Type I1
error"?? I n other words, what size can a model error attain before it leads (with
a certain specified probability) to a significantly tbo high F-value? This question,
which forms an essential part of the application of statistical tests, has been treated
from a geodetic point of view by W. BAARDA
in [4], page 21 ff. By "model error"
we mean now a systematic error that manifests itself in the misclosures, and it is
convenient to think of a gross error or blunder, because we will consider cases where
one observation is "falsified". If we think of one polygon, a very large error in one
of its sections will almost certainly lead to rejection of the null hypothesis that no
model-errors are present, because it will result in a very large misclosure. A small
error, that has the size of the standard deviation of the observation in question will
very seldom lead to rejection: the probability P of detecting an error is dependent
on its size. This probability is called the power of the test with respect to the alternative hypothesis that an error of the given size occurs (this hypothesis is an alternative
to the null hypothesis).
We may now fix p, e.g. p = 0.8 and derive the size of the corresponding error,
which consequently is the size an error has to attain to be detected with 80%
probability ("be detected" means here: lead to rejection in the test. The test cannot
in general indicate the particular observation affected by the error.)
We will use the notation which is customary in tensor analysis, see [13]. Let us
consider observations Pt(i = 1, . . ., n), subject to b conditions. The weights are gtk,
the variance factor is
a?
up(p"+.a)
= U;
(i, k J l = l , . .,m)
Normal equations
gikufu;ia = te
Put
giku:~; = gw
and
gwgu7 = 8:
Then
i a = g0z.t'
is the solution of the normal equations.
The estimator of the variance factor is:
One of the basic assumptions of least squares theory is that the means of the corrections gi are zero. If the observations are affected by systematic errors or blunders,
V{, the means of the
will not be zero. The means of the misclosures will be
Ve = -uPV" the variate
We define
The distribution of
$2
-
02
Similar considerations hold for the type ofF-test according to (3.1-2), where we have
20
1,
NR.
in case the quadratic form in the numerator is non-central. For illustrative purposes
we will only discuss the case that b11 = m .
The probability that F f b m, , A is greater than the critical value F I - , ; ~m , of the test
is by definition the power of the test:
P(&,
m, A
> FI-,; b , m)
whereas g,,
-,
VtQz -VQ
Fixing a
0.05 and
The indices 1 and m of A are explained by the fact that 82, computed from one
supernumerous observation, is tested against the variance factor 0 2 (= 200). With
the aid of the nomogram [l01 we find
The result of this computation is fbund in Table 111, column 2, giving the result
per section. Most sections are part of two polygons; in that case the smallest of the
two V's is given.
We now consider an adjustment on h conditions, and we suppose that the adjustment is executed according to Standard Problem I1 (observation equations). For
convenience we recall the formulas in the notation used in this section: pi are the
observations, ci the corrections, _hr the unknowns (in this case geo-potentia! numbers
with respect to the datum).
Observation equations :
(i = l , . . . , m )
( r = 1 , . . .,m-b)
Weights :
Normal equations :
gtkafa%@
= gtka;pk
Put
gikafa,k = grs
grsgst = 6:
and
gixa:pk = _F,
Then
= grsfs
Weight coefficients
pi,
;Ib,m = -gtkVeiVeX
(S2
so that
giXO~iVek= d.10,m
Now the model error Vpi is the error in the observation between h 7 and
and
therefore we indicate it by Vrs. We suppose there is a model error in one observation
pi so that
-
gikvpivpk
= gii(vrs)2
Furthermore we have
VF, = gikafVp"
+ 1 or - 1, so that
(3.3-2)
22
1, NR. 2
and consequently
_hr
and
_hs,
gii is the weight ofp' before the adjustment, and gr-~,r-Scan be computed easily if
the matrix of weight coefficients of the unknowns is available.
and
Formula (3.3-4) shows clearly the relation existing between VrS, gig,
1
gr-ssr-S. Before the adjustment gr-szr-s = - (assuming non-correlated original
observations). Hence
u 2 ' I b ,m
CO
This confirms the self-evident truth that we cannot detect a model error if there
is no adjustment. Further we see from (3.3-4) that the smaller the relative weight
coefficient between the end points of a section, the smaller the model error that can
be detected. We can also compute as follows a lower bound for Vrs:
200
we get
Lt2
-,
The best test is obviously the one in which the smallest model error leads to rejection. I t is evident from (3.3-4) that to obtain this test we must not take the highest
possible number of conditions (polygons), for by adding more and more conditions,
23
gr-Ssr-swill finally hardly decrease, whereas l b . ,will increase with the number of
conditions.
I n order to verify this, the tests on
per partial net, as described in Section 3.1, have been investigated. The value of
V ~ was
S computed for each levelling line (section). The result is found in column 4
of Table 111. Of course, the lines between the partial nets had to be left out.
The result of this computation is remarkable. For each levelling line, the minimum
model error leading to rejection has increased, in most cases considerably. Evidently
the increase of l b , , has much more influence than the decrease of gr-ssr-8. Consequently it may be concluded that it is better to test per polygon than to test per
partial net, if one wants to detect model errors of minimum size. But we may also
use tests involving two, three or more polygons. Can we thus find a test that is still
better than the test per single polygon? T o answer this question we must try to find
a test, in which the decrease of gr-83'-8 has a greater influence in (3.3-4) than the
increase of l b , If we take a levelling line that occurs in two polygons, we see that
the adjustment of these two polygons has a great influence on gr-s.r-8. The obvious
thing to do is therefore to examine the test in which 6 2 is computed from two polygons
having a line in common. Actually, an adjustment was necessary for each line; this
adjustment has been done according to Standard Problem 11. One of the end points
of the line in question was taken as a reference point, the g.p.n. of the other end
point being the only unknown in the adjustment. We simply get:
.,
in which Z g t t is the sum of the weights of the three levelling lines occurring in
i
the problem.
The results of this computation are given in column 3 of Table 111. I n general
we find indeed that the minimum model error that leads to rejection in 80% of the
cases is smaller in the test on two polygons than it is in the test on a single polygon.
I n general, the improvement is not great; in several cases we even find that the
minimum model error is larger in the two polygon case. This is mainly the case
when the two polygons concerned are very different in precision. I n this case the
relative precision of the two points is only very slightly improved by the addition
of the least precise polygon.
The computation has been done only for lines which are situated between two
polygons, because it cannot be expected that the relative precision of the two end
points of a line is much improved by adding a polygon which does not contain
the line.
Since the relative weight coefficient gr-ssr-8 of geo-potential numbers resulting
from the adjustment of the total net are available, it is very easy to compute for
every line V's for the test using 6 2 of the total adjustment. The result can be found
24
PUBLICATIONS O N
GEODESY,
NEW SERIES,
VOL. 1, NR. 2
in Table 111, column 5. The model error corresponding to a power of 0.8 is considerably larger except for the lines forming the edges of the partial nets. This is caused
by the fact that the relative precision is increased most at those edges.
We might now want to investigate the power of the test when 3 2 is computed
from e.g. three polygons. I t is not impossible that we find a V ~ that
S is still somewhat smaller than in the case for two -polygons.
But
we
meet
some
difficulties. I n
. the first place, one can add a third polygon to two other ones in many different
ways, and it is not self-evident which combination results in the highest power. I n
the second place, this investigation involves much extra computational work, therefore it has not been pursued.
One might investigate the power of the tests on &,/B$ concerning estimates of
from different partial nets. I n that case formula (3.3-4) is still valid, when Ab,
is replaced by Abl,b2. I n our case, Abl,b2 M 2Ab, m, SO that the model error leading to
rejection in 80% of the cases will become approximately 4 2 times as large. The
power of this test is therefore much smaller than that of other tests. Nevertheless,
this type of test has the advantage that the assumed value of 0 2 has no influence.
This investigation has not resulted in a statement saying which test has the
greatest power: this question has not been answered fully. But the tests having
greatest power are definitely not the tests involving two estimates, nor the tests on
estimates obtained from a large number of supernumerous observations. It is well
to test the estimate resulting from two contiguous polygons. I n the line they have
in common one can then detect a model error that is not likely to be much greater
than the model error that can be detected by any other F-test. If the line to be
investigated is only contained in one polygon, it is probably best to test the estimate
resulting from that polygon. If the line is part of two polygons of very different
quality, the best one can do is test the estimate from the most precise polygon.
Tests on $2 obtained from all observations will have minimum detectable errors that
are nearly twice as large as the tests recommended above.
We give a small survey where for convenience the model errors are expressed
in cm:
Number of conditions
from which $2 is computed
1
2
11- 15 (partial net)
68
(total net)
These values illustrate that even the hypotheses that were not rejected may be
quite unreliable; the fact that a hypothesis is not rejected in a test does not imply
that it is a true hypothesis.
25
4 T H E G E O - P O T E N T I A L O F M E A N SEA L E V E L
4.1
27
(i,j, k
l,.
. ., n)
The actually observed values c$ do not fulfil1 these conditions; they have to be
adjusted. The corrections ~ h e s u l t i n gfrom the adjustment are immediately known,
namely
The number n of mareographs considered may vary from one to the total number
of mareographs attached to the net. I t may be noted that the number of condition
equations is equal to the number of observations, and in this case it is more evident
than usual that the condition equations, forming the condition model, are the
expression of certain hypotheses.
To test the hypothesis expressed by the conditions
we use the estimate of the variance factor resulting from the adjustment. Denoting
any estimate of the variance factor from mareograph observations by an index M,
we can compute for any combination of mareographs
The values ci are given, and the matrix of weights l lgvl l is obtained by inverting
the n, n (partial) matrix of weight coefficients I lg"1.
The estimate obtained can be tested with respect to the estimate found from the
adjustment of the net:
$L= gwcgc"
The test is based on the identity
28
PUBLICATIONS ON
GEODESY,
NEW SERIES,
VOL. 1, NR. 2
The level of significance we use is 5%. For the practical computation we compare
ci with the critical value yi which follows from
in which VcGs the model error which on the average in 8 out of 10 tests will lead
to rejection of the null hypothesis. The difficulty is, however, that to the authors'
knowledge no tables are available for the power function of this more complicated
non-central F-distribution. I n the following we will therefore take 0 2 = 370 in the
above formula and consider this as the true variance factor. Theoretically, the
results will be more or less incorrect but it is hoped they give a sufficient practical
indication.
We find
From the geo-potential difference between MSL at any two mareographs numbered
i and j one can estimate 0 2 by exactly the same formulae as used in the previous
section. Instead of c b n e must introduce ci-cj and instead ofgii their relative weight
coefficient, which may be denoted by g(i-j).
The results are given in Table V. Of course, not all combinations have been
investigated, but it was tried to make a representative choice. Of 132 combinations
considered, 59 show a significant difference.
A closer investigation shows that strong significance can be ascribed to some combination Gulf of Bothnia - North Sea (1-23, 12-23, 15-1 7, 15-24, 15-34), to several
29
A
MA
MZ
4886
10351
11869
13.2
28.0
32.1
16
6
4
'
Result
1.79
2.23
2.50
reject
reject
reject
~ are
z found
~
in Table VI.
We see that all of these tests lead to rejection, and even with considerable significance. The null hypothesis was: the g.p.n. of all mareographs in a group is zero.
The conclusion that the test indicates that not all mareographs in a group have a
zero g.p.n. However, this does not all mean that the group as a whole lies higher
or lower than NAP. We can illustrate this by considering the case that two mareographs a and b are tested with respect to NAP. Then we have
30
1, NR. 2
~a = - gab c b
gas
This extreme value is a minimum, for
If a and b are not too far apart the computations show that
gab W -gas
Consequently, 8 ~ reaches
2
a minimum if ca W cb. If, e.g., ca = -cb, 8 ~ is2 much
greater; the probability of rejecting the null hypothesis is also greater, meanwhile
we can certainly not say that the combination of the two mareographs lies higher
or lower than NAP.
We can also in this case compute VC" i.e. the model error in the mareograph i,
which with 80% probability leads to rejection of the null hypothesis tested. The
computation is done according to
(Vc"2
02An,68
=-
gti
For 02 we use again the value 5 2 = 370 obtained from the net adjustment, which,
as indicated before, is theoretically wrong.
The computation is found in Table VI. O n comparing the results with those in
Table I V one notes that the power in the former case is in general smaller, but in
some cases a little higher. An exception is formed by the mareographs 22 and 23,
for which the "detectable" model error has become much smaller in Table VI.
I t is very likely that this can be explained by the fact that these mareographs are
strongly dependent on each other.
An entirely different result is obtained if we do not assume a model error in a
single mareograph, but in a group of mareographs, e.g.:
VC"
for i
We then get
v2
An,6802
=-
gtj
i j
We find:
al,
. . ., an (n mareographs)
31
If a model error occurs in the assumed way, we can detect with 80% probability
in the Atlantic group a model error of about 4 cm, in the Mediterranean one of
about 1 dm.
4.2.4
I n all tests so far executed (except the testing of pairs of mareographs in Section
4.2.2) geo-potential numbers were used whose zero reference surface was defined
by the NAP datum. However, it is a well-known fact that NAP differs considerably
from MSL. This is of course a model error which contributed to the high significance
with which many estimates 62 exceeded their critical value in the tests described.
We will have to use another model to compare the level of the different seas. We
do not put the adjusted g.p.n.'s of MSL at the mareographs equal to zero but to a
value CM which is the same per group of mareographs. We then get three separate
adjustments for the determination of CA, c M A and cMZ. Each adjustment has the following form (we use Standard Problem 11):
~ t + c t = 4.M
(i = al,
. . ., a,;
At4.M = ~
with At
n mareographs)
or
t + ~ t
1 for i
al,
. . ., a,
gix
and
FM =
i=a, k = a ,
02
gtkck
i=n,
is:
- _FM CM
- gtk~t4.~
n-l
The estimator is denoted by
3 ~The
~ computation
.
of CM and MM is found in Table VII. The computation of
~
32
A
MA
MZ
30818
51797
38937
78175
62106
47475
47357
10309
8538
15
5
3157
2062
2846
8.53
5.57
7.69
1, NR. 2
1.81
2.35
2.74
reject
reject
reject
I t is seen that the significance is much lower than in Table C in Section 4.2.3,
2
be
but still it is very high. This means that the result of the tests on 2 ~ cannot
explained completely by the difference in level between the seas compared, for the
null hypothesis tested is now that all the mareographs of a group indicate the
same MSL.
Another test can be designed by using the estimator
The result is
TABLE
E
A
MA
MZ
30694
51717
38927
83.0
139.8
105.2
3.98
3.98
3.98
Result
reject
reject
reject
The result is striking. We can draw the conclusion that the "mean" level of the
Atlantic as well as that of the Mediterranean differs significantly from NAP. From
the computations in Table V11 it follows that these levels are -75.10-3 gpu and
-297. 10p3gpu with estimated standard deviations of about
2062
gpu and
u 60 10-3 gpu respectively (see also Table D).
m 24.10-3
We can examine the power of these tests too. The following relation is valid:
We find
VCA
24.10-3 gpu
O n comparing these results with (4.2.3-2) we see that the power is increased considerably. This is also evident from a comparison of the formulae: if the V from
(4.2.3-1) is denoted by VI and the V from (4.2.4-1) by VII it is evident that:
Consequently
Then
cM2
is found by
The matrix Jlck, cl11 is known; the coefficients Bk are easily computed. The computation is carried out in Table VIII. We find
= gMIMl
fgM2M2-
from:
~ 2
2gMlM2
1