Robustness Analysis

3. 5

Final Report

External Reliability
Another reliability measure that Baarda developed in 1968 is called the external reliability.

External reliability tells us about the effect of

voei on the positions obtained through the least-

squares adjustment. External reliability, however, is not used in our method; the effect of V ~ on
the result is handled in a different way. To make our discussion more complete and to make a
comparison between our method and Baarda' s method in assessing the effect of V

oei , a brief

discussion on the external reliability is still given below.

In the least-squares adjustment, the unknown parameters are estimated using equation (1.2). If
we pre-multiply equation (1.2) by N-1, we get:
(3.21a)
or
(3.21b)
Supposing that the observation vector tis burdened by blunders Vt, i.e., t =

e + ve and

substituting t into equation (3.21b), we get

~ = (At Pt A)-1 At Pt t ,
where ~ are the shifted unknown parameters affected by blunders

(3.22)

ve. Substituting fort into

equation (3.22), we get

~ =(At Pt A)-1 At Pt (t + Vt)

=(At Pt A)-1 At Pt t +(At Pt A)-1 At Pt Vt
and denoting by v~ the shift of ~ due to

(3.23)

ve yields

V~ =(At Pt A)-1 At Pt Vt

(3.24)

The effect of the maximum undetectable blunder Voei on the estimated parameters can be
determined by substituting Voei for

ve :
(3.25)

Here, VoQi is dependent on the coordinate definition, i.e., it is datum dependent. Baarda [1976;
1979] proposed another kind of external reliability measure:

3. Reliability Analysis

21

Robustness Analysis

Final Report

(3.26)
known as the relative external reliability measure.
As we shall see in the next chapter, the effect of blunders on the network is better handled as a
virtual deformation and thus depicted by a more appropriate technique than the 'external
reliability.'

3. Reliability Analysis

22

Robust Analysis

Final Report

4. GEOMETRICAL STRENGTH ANALYSIS

In this chapter, the use of strain for the strength analysis of geodetic networks is described.
The basic concepts of strain are given followed by its application to geodetic networks, and
specifically its use as a tool for analysing the geometrical strength of networks. We also show that
changes in the network datum have only a second-order effect on strength.

4.1

Concept of Strain

Strain is a purely geometric approach to the analysis of the deformation of a physical body. It
is based on the theory of elasticity in mechanics where it is applied to the description of the
relative deformation of a body with respect to some initial state. Here deformation is taken to
mean the change in shape or configuration of the body.
Deformation can be classified as either homogeneous or nonhomogeneous. If the deformation
is homogeneous, straight or parallel lines will remain as straight or parallel lines after
deformation. If, on the other hand, the deformation is nonhomogeneous, initially straight or
parallel lines become curved or nonparallel after deformation. These deformations are illustrated
in Figure 4.1.

U ndeformed state

Homogeneous strain

Nonhomogeneous strain

Figure 4.1 Examples of homogeneous and nonhomogeneous deformations.

4. Geometrical Strength Analysis

23

Robust Analysis

Final Report

Strain induced by homogeneous deformation is called homogeneous strain and is constant

over all parts of the region of deformation. Nonhomogeneous deformation, on the other hand,
produces a more complicated nonhomogeneous strain field.
Deformation and strain can also be classified as fmite or infinitesimal. Finite strain usually
describes an instantaneous deformation of a continually deforming body with respect to its
original undeformed state, i.e., cumulative strain. On the other hand, infinitesimal or incremental
strain describes the instantaneous deformation of the current deformed state with respect to some
earlier, not necessarily undeformed, instantaneous state.
Only nonhomogeneous and infinitesimal deformation is needed in strength analysis due to the
following considerations:
(a) the deformation of a geodetic network is generally nonhomogeneous, and
(b) in the strength analysis of a geodetic network, the deformation is much smaller compared to
the size of the network and we can thus use infinitesimal strain theory.
The latter allows us to take advantage of the fact that infmitesimal deformation is differentially
small in order to simplify the mathematical description of strain.
Mathematically, infinitesimal strain is defined as the rate of change (i.e., gradient or slope) of
an object's displacement field with respect to position. Given a three-dimensional (3D)
displacement field u(x,y,z)=(u,v,w)T, as a function of position x=(x,y,z)T, the strain matrix E
consists of 9 linear displacement gradients given by

E =

grad(u)

= au(x,y,z)
ax

au
ax

au
ay

au
az

av

av

av

ax

ay

aw aw aw
ax
ay

az

dz

euy euz ]
evy evz
,
ewx ewy ewz

[ eux

evx

(4.1)

where the derivatives are evaluated at the point of concern. These linear strains e correspond to
the rate of change of displacement in each of the three coordinate components along the three
coordinate axes. For example, euy is the rate of change or gradient of displacement in the xdirection with respect to position in the x-direction.

4. Geometrical Strength Analysis

24

Robust Analysis

Final Report

Note that the mechanical properties of the material are not involved in strain. Strain is
applicable, whatever the mechanical behaviour of the material. Note also that strains describe
only the relative displacement of points so that rigid body translations do not affect strain. This
will be discussed in more detail later in this chapter.

4. 2

Deformation Primitives

The strain matrix contains all of the strain information about the displacement field. It is not
easily interpreted, however. Various scalar parameters can be derived from the strain matrix in
order to make the interpretation of strain more convenient and illustrative. We call these
parameters deformation primitives.
The strain matrix E can be decomposed into its symmetric S and anti-symmetric A parts; i.e.,

E
where

.. ,

Euy
Evx Evy Evz
Ewx Ewy Ewz

[ Eux

~ [ ro,
0

-COy

-COz

roy

- ;x
0
COx

]
]

= S+A
eux

1
euy+evx)
1
euz+ewx)

1
euy-evx)
1
-2(euz-ewx)

(4.2)

'
1

euy+evx)
evy

2<evz+ewy)

1
-euy-evx)

euz+ewx)

evz+ewy)

ewz

1
euz-ewx)
1
- 2(evz--ewy)

2< evz--ewy)

(4.3)

(4.4)

The symmetric part is often referred to as the symmetric strain tensor.

The symmetric strain tensor S describes the expansion and contraction as well as the shearing
deformation at a point. The strain tensor is usually parameterized in terms of the so-called strain
ellipse or ellipsoid in the same manner that error ellipses and ellipsoids are computed from
covariance matrices, except that no square roots of the semi-axis lengths are taken. The principal
strains (At. A.2, A.3) are the eigenvalues of the strain tensor and the
4. Geometrical Strength Analysis

25

Robust Analysis

Final Report

eigenvectors are the directions of the principal axes. Negative principal strains indicate
contraction and positive principal strains expansion.
The anti-symmetric strain matrix A describes the twisting deformation at a point. The
quantities ro are called average differential rotations and describe the twisting about each of the
three coordinate axes at a point. Note that in the two-dimensional case there is only a twist COz
about the local z-axis (i.e., in the x-y horizontal plane).
More convenient scalar deformation primitives for strength analysis can also be derived from
the strain matrix (see Schneider [1982]). Dilation cr describes the average extension or
contraction at a point and is defmed as the average of the principal strains; e.g., for 3D
(4.5)

Note that the sum of principal strains is equivalent to the trace of the symmetric strain tensor
which is equal to the trace of the strain matrix. Total strain Ais a similar quantity, defined as the
geometric mean of the principal strains [Dare, 1983]; i.e.,

(4.6)
Shear strain can be classified as either pure shear or simple shear. Pure shear 't deforms a
square into a rectangle so that separation between lines changes. It is defined by [Schneider,
1982]
1
txy = -'tyx

= 2 (eux- evy)

1
'txz = -'tzx =

(au av)

(au aw)

= 2 ax- ay

2 (eux- ewz) = 2
1

'tyz

ax - dZ

(av aw)

= -'tzy = 2 (evz- ewy) = 2 az

'
'

- ay .

Simple shear u deforms a rectangle into a rhombus so that angles between lines change. defined as
[Schneider, 1982]
(4.7)

(4.8)

(4.9)

It is

4. Geometrical Strength Analysis

Robust Analysis

26

Final Report

'Uxy

= -'Uyx = 2 (euy + evx)

'Uxz

= -'Uzx =

1
2 (euz + ewx)

= -'Uzy =

1
2 (evy +ewz)

'Uzz

(au a v)

(4.10)

(au a w)

(4.11)

(av a w)

(4.12)

= 2 ay + dx '
1

= 2 az + dx '
1

= 2 dy + dz .

Neither type of shear produces any rotation. These two types of shear are illustrated in Figure 4.2.
Another type of shear, total shear y, is the geometric mean of the components of pure and simple
shear; i.e.,
'Yxy
'Yxz

= ""''txy2 + 'Ux/
= ""''txz

2
2

+ 'Uxz

(4.13)

(4.14)

(4.15)
'Yyz = ""''tyz + 'Uyz
The principal axes of the strain tensor define the directions in which no shear takes place. The
directions of maximum shear are at 45 to the principal axes of the strain ellipse/ellipsoid. The
magnitude of shear can also be determined indirectly from the difference of the principal strains
(lengths of the principal axes of the strain ellipse/ellipsoid) [Schneider, 1982].

X
~

r---------1

I
I
I
I
I
I

I
I
-

I
.....

--

~-

--1

I
I

_,J

.I

Pure shear

Simple shear

Figure 4.2 Pure and simple shear. Solid lines are the undeformed state and
broken lines are the deformed state.

4. Geometrical Strength Analysis

27

Robust Analysis

Final Report

Note that the symmetric strain tensorS can be represented in terms of dilation, pure shear, and
simple shear. Using the above definitions, we find that

CJ+'txy+'txz
'U yx
'U zx

'Uxy

'Uxz

CJ+'txy+'txz
'U yz
'U zy
CJ+'txy+'txz

(4.16)

Although the expressions for the various deformation primitives have been developed in 3D,
previous investigations by Craymer et al. [1987] have found that only 2D primitives have any
practical meaning in the context of geodetic networks. The problem is that geodetic networks are
inherently only 2D in nature since they lie on the surface of the Earth whose variations in height
are much smaller than those in the horizontal dimension. When two points have very nearly the
same height (a common occurrence), the displacement field gradients with respect to height can
become extremely large or even discontinuous, resulting in artificially large and misleading
results.
The deformation can be displayed in a variety of ways (see Thapa [1980], Schneider [1982],
Dare [1983], and Craymer [1987]). In network strength applications, the only scalar primitives
needed are differential rotation, dilation, and total shear. These scalar deformation primitives are
most conveniently displayed using either 3D surfaces or contour plots. Only the latter is currently
supported in the NETAN software which implements this analysis.

4. 3

Virtual Deformation of Geodetic Networks

The concept of strain can be readily applied to the analysis of geodetic networks by
considering the network to be a structure in itself. That is, stations are held together by the
interconnecting observations as a building is held together by its beams. In this analogy, stations
are considered to be the joints and observations are the beams and brackets. Distance observations
can be thought of as beams of rigid length whose orientation in space is not fixed. Angles can be
considered as brackets which fix the relative orientation (angles) between beams of arbitrary
length. Azimuths can be thought of as brackets that fix the orientation of a beam of

4. Geometrical Strength Analysis

28

Robust Analysis

Final Report

arbitrary length with respect to the foundation (which acts as the datum definition). We have
found that using such an analogy helps in the interpretation of the strain parameters.
In practice, the displacement field over a structure is never known as a continuous function of
the position of points on the body. The displacements are known only for a discrete set of points
describing the structure. Only a discrete displacement field can therefore be obtained which
approximates the actual continuous displacement field.
For the strain analysis of geodetic networks, the displacements are of a virtual nature. They
represent changes to the coordinates of the points in the network that may result from a variety of
changes (perturbations) of the network. Some of these are:

changes in observation values,

changes in observation weights,

deletion or addition of observations,

deletion or addition of points,

changes in network constraints.

The virtual displacement field is the set of coordinate changes for all points in the network. Only
virtual displacements due to changes of observation values are needed in strength analysis.

Given a local displacement field l) around a point, the strain can be easily determined from
the displacement gradient evaluated at the point. For geodetic networks, we can define the "local
displacement field" at a point to consist of displacements of either all interconnected points (i.e.,
all points connected by observations to the point of interest) or all stations within a specified
radius of the point of interest. The virtual displacements (i.e., changes in coordinates) of all points
within the local displacement field can then be approximated by a simple surface such as a plane
or low-degree algebraic surface (surface described by a low-degree algebraic polynomial). In our
experience, we have found a plane to be the most robust approximation of the local displacement
field at each point. Higher-order algebraic polynomials are not suitable for such applications
since they tend to produce spurious gradients
4. Geometrical Strength Analysis

29

Robust Analysis

Final Report

when the points are not regularly distributed in space (i.e., they tend to 'fall through' areas without
stations).
The gradients of the local displacement field are evaluated separately for each of the
coordinate components. A separate local displacement field is detennined for each coordinate
component and the gradients along each of the coordinate axes are evaluated to give the
components of the strain matrix. Fitting a plane surface to each displacement field results in a
very simple detennination of the strain; the strain components are just the slopes of the planes
along each of the coordinate axes.
For the 2D case, the local displacement field components u and v are approximated by
u
v

= ao + a1 x + a2 y ,
= bo + b1 x + b2 y ,

(4.17)
(4.18)

where x and y are the coordinate components of the points in the local displacement field, and the
a's and b's are the coefficients defining the planes. For numerical stability, these coordinates are
expressed relative to the point of interest. Solving for the coefficients in both sets of equations
results in

(4.19)

(4.20)

where N=A T A and A=[l x y] with 1 being a column of ones.

The strain elements are then

(4.21)

(4.22)

Letting N denote the reduced normal equation matrix with ao or bo eliminated, the elements of
the strain matrix can be expressed together in vector form as

4. Geometrical Strength Analysis

Robust Analysis

30

Final Report

_ eux]
e - [ evx
euy

evy

_ "'IT
- [ N- 1 A Tv] N- A u

_ Q l)

'
(4.23)

where the local displacement field vector is ordered as l;T=(u T, vT).

4. 4

Strength Analysis Using Strain

The use of strain to analyse the strength of a geodetic network was first proposed by Vamcek
et al. [1981] and later developed by Dare [1983]. Rather than describing the ability of a network
to resist the propagation and accumulation of random errors, the strain approach is based on the
ability to resist the propagation and accumulation of systematic errors or blunders
(i.e., changes of a non-random nature).
The basic approach is to perform a series of separate strain analyses by individually changing
the observation values. Each such perturbation produces a new displacement field and thus strain
at each point. The most realistic results were obtained when changing each observation by its
standard deviation [Dare, 1983]. A measure of strength is obtained by assuming the network is
only as strong as its weakest link. The weakest link corresponds to the largest strain parameter at
each station from the entire series of strain solutions for all
observation perturbations.
With this technique, a virtual displacement field must be generated for every observation in
the network. Although this may seem like a daunting task, sequential estimation methods can be
used to advantage here (see Craymer et al. [1989]). The displacement field

l)

in response to a

change of an observation can then be given directly in terms of the perturbed observation vector
~1,

which contains only one non-zero element equal to the standard deviation of the

observation; i.e.,

5 = -N-1 AT P M = T 6.1 '

where A is the design matrix, P is the weight matrix of the observations and N

(4.24)

= AT P A is the

normal equation matrix. Because only one observation is changed at a time, only one

4. Geometrical Strength Analysis

31

Robust Analysis

Final Report

column of the normal equation matrix is needed to evaluate the displacement field if the
observations are independently weighted. Since the strain elements are linear functions of the
displacements (see eqn. (4.23)), we can write the following system of linear equations for each
change to an observation:

e = Ql)

= QT~I = R~I.

(4.25)

We are now interested only in the largest deformations at each point as measured by the
deformation primitives: dilation 0', total shear y, and differential rotation ro (cf. eqns. (4.5), (4.13)
to (4.15) and (4.4)). New deformation primitives are computed one at a time for a change in each
observation by its standard deviation. Only the largest primitives (in absolute value) at each point
are retained as a measure of the weakest link. These maximum values (denoted by O'max. "fmax.
and COmax) at each point in the network describe the network strength and are referred to as
strength in scale, strength in shear, and strength in rotation (twist), respectively. They can be
displayed as contour plots or 3D surface plots. Only the former is currently supported in the
NETAN program [Craymer et al., 1988; 1989].

4. 5

Datum Independence of Strength

The effect of the coordinate system or datum definition on the computed strain parameters is

an important issue in the strength analysis of networks. Datum is taken here to mean the
defmition of the origin and orientation of the coordinate system as well as the scale. The origin is
usually defined by specifying fixed or heavily weighted coordinates for one or several points in
the network. The orientation is often defined using weighted observations such as azimuths or
observed position differences between points. The scale is generally defined using weighted
distances or, again, position difference observations. Two or more weighted position observations
can also be used to defme datum orientation and scale.
Ideally the strength of a network should not depend on the choice of a datum so that different
people analysing the same network, but using different datums, will get the same strength
parameters. It is shown here that rotations and scale changes have only a very small
4. Geometrical Strength Analysis

32

Final Report

Robust Analysis

and negligible effect on the strength parameters and that translations of the datum origin have no
effect at all. The effects of translations, rotations, and scale changes on strength parameters will
be evaluated in terms of strain parameters only, strength being just the largest strain parameter at
each station resulting from a series of virtual displacement fields.
It is important to bear in mind that only one datum definition is used in a single strength .
analysis. The virtual displacement fields generated for the strength analysis are due only to changes
in the network observations ('blunders') and not to changes in the datum. The question is whether a
displacement field generated by such a 'blunder' gives the same strain as another displacement field
also generated by the same 'blunder' but using a different datum. In practice, the differences in
datums are likely to be very small; say, less than a degree in the orientation of the coordinate axes,
and a few hundred parts per million in scale.

Translations of Datum Origin

Differences in the datum origin between different strain (or strength) solutions completely
cancel in the determination of the displacement field. That is, the displacement fields for both
solutions are identical even though they may be based on datums with different origins.
This can be proven very easily by considering one strain solution where x1 are the original
coordinates of points in the local displacement field and x2 are the coordinates after the network
has been perturbed by a single blunder. The local displacement field l) is then
(4.26)
Consider now a second strain solution using a different datum origin which is offset from that for
the first solution by a translation .6.x. For this solution, the original coordinates XI* and those X2

* after perturbation by the same blunder can be expressed in terms of the coordinates for the frrst
solution as
XI*

= XI + dX ,

(4.27)

X2 *

= X2 + dX .

(4.28)

The displacement field l)* for this second solution is then

4. Geometrical Strength Analysis
Robust Analysis

33
Final Report

~*

= X2 -XI
= X2 -XI = u ,
(4.29)
u
which is identical to that for the first solution. Any translation of the datum origin therefore cancels
in the virtual displacement field. Since the virtual displacement fields are identical in both strain
solutions, the strain and strength parameters also will be identical. Strain and strength parameters
are therefore invariant to translations of the datum origin.
It is important to point out here that strain is also invariant to displacements resulting from the
translation of all points in a network.. The [Link] fields in this case will have constant
values for both the u and v components at all stations (i.e., they will be horizontal planes). The
gradients (strain) of these displacement fields will be zero since a horizontal plane will have zero
slope. This is the reason that strain is preferred in studies of crustal motion where it is not known
whether the point fixed in a previous adjustment has moved.

Rotations of Datum Coordinate Axes

A change in the orientation of the coordinate axes defining the datum in a strain (or strength)
analysis results in only a very small and negligible effect on the resulting strain and strength
parameters. Given the same displacement field 5 for the first strain solution as above (generated by
a blunder), the strain matrix E is defmed by
(4.30)
Consider a second solution where the coordinate system has been rotated by an arbitrary rotation
matrix R. The new coordinates before and after perturbation by the same blunder are, respectively,
XI*
X2*

= R X}'
= R X2.

(4.31)
(4.32)

The displacement field 5 * for this second solution is then

li*

= X2 * -XI* = R x2 -

R xl

= R li .

(4.33)

Note that for a small angle of rotation, the term R li is only a second-order effect. The
corresponding strain matrix E* is
4. Geometrical Strength Analysis

34

Robust Analysis

Final Report

E* = grad(ll*) = R grad(ll) = R E .

(4.34)

If the rotation angle &a. is small, the rotation matrix for a single rotation about, e.g., the zaxis can be simplified to
R-

[-~a

&a.
1

0] = I+ [ 0 &a. g]=l+AR.

-&a.

0 1

(4.35)

The strain matrix in this new datum is then

E* = R E "" E + &R E = E + &E .

(4.36)

The effect of a change in datum orientation is therefore &E=&R E. For strength analyses of
geodetic networks, the changes to this will result in only a very small effect which will be
negligible in all practical cases.
A worst case effect can be estimated by considering a solution with very large strains of about
e=1xlo-4 (100 ppm). If the datum is changed by a rotation of the coordinate system by a large amount,
say &a=1x1Q-2 radians (over half a degree), the change in the strain matrix from the first solution is
only &ae=1x1o-6 (1 ppm). In a strength analysis, the strain elements are unlikely to exceed 50 ppm in
which case a rotation of over 1 will be required to produce a
1 ppm effect on strain. In practice, the orientation of the network datum will generally be known to
much better than 1 degree accuracy. These estimates have been verified using numerical tests.

Changes of Datum Scale

The effect on strain (and strength) parameters due to changes in datum scale can be
determined in a similar manner. In this case, the strain solutions before and after a change in scale
of &s results in the following coordinates
XI*

= (1 + &s) XI ,

(4.37)

X2 *

= (1 + &s) X2

(4.38)

The displacement field ll* for this second solution is then

ll* = x2* -xi* = (1 + &s) ll.
4. Geometrical Strength Analysis

(4.39)
35

Robust Analysis

Final Report

and the corresponding strain matrix E* is

E*

= grad(fi*) = (1 +~s) grad(fi) = (1 +&) E = E- ~s E .

(4.40)

Note that ~s l;, and thus ~s E, is again only a second-order effect

A worst case effect can also be estimated by again considering a solution with very large
strains of about e=1x1o-4 (100 ppm). If the datum is changed in scale by an extremely large
amount, say ~s=1x1o-2 (10 000 ppm), the change in the strain matrix from the frrst solution is
only ~s-e=lx1o-6 (1 ppm). In a strength analysis, the strain elements are unlikely to exceed

50 ppm in which case a scale change of over 20 000 ppm will be required to produce a 1 ppm effect
on strain. In practice, the orientation of the network datum will generally be known to much better
than 10 ppm accuracy which would result in scale effects of only 0.001 ppm for this example. These
estimates have been verified using numerical tests.

4. Geometrical Strength Analysis

36

Robustness Analysis

Final Report

5. ROBUSTNESS ANALYSIS
5. 1

Merging Reliability and Geometrical Strength Analysis

In the 20 plus years since Baarda [1968] proposed the concept of reliability analysis, the
technique has found, albeit quite slowly, many proponents and followers. Based on a rigorous
statistical foundation, the technique offers an alternative tool for analysing networks of
various kinds, e.g., geodetic, photogrammetric , and those for engineering surveys. The only
problem with reliability analysis is that the interpretation of its results, particularly those
pertaining to positions as opposed to observations, is rather difficult. We can quantify the
maximum expected observation residual that can escape purging as an outlier (blunder) by the
standard statistical test for outliers. What we cannot learn from the analysis is just how much
damage (distortion) such an undetected error (possible blunder) can cause to the network. On
the other hand, there is a global indicator of 'external reliability' (equation (3.26)) provided in
Baarda's technique, but this is far too coarse a measure to be of much real use. What is really
needed is a much finer measure, commensurate in its fineness with the distinguishing power
of the internal reliability, that would pinpoint areas, or even better, points, where one can
expect the damage to be significant and other points where the damage should be expected to
be insignificant. The individual indicators of 'external reliability' (equation (3.25)) associated
with individual points provide a fine enough measure but they are datum dependent and there
are far too many of them to be practical.
Some 10 years ago, work on one such measure started at UNB and culminated in 1983 with
Dare's [1983] formulation of 'geometrical strength analysis' (GSA). This technique approaches
the problem of network deformability or lack of it, i.e., strength, from a purely geometrical point
of view. The question GSA answers is: How could a geodetic network deform in the worst case if
the observations were burdened with some undetected non-random errors? The answer includes a
somewhat unexpected complication - there does not exist one

5. Robustness Analysis

37

Robustness Analysis

Final Report

single scalar measure of such a deformation, but three independent measures. In Chapter 4,
these three measures are called the 'deformation primitives.' They are the pure strain (scale),
the total shear (shape), and the differential rotation (twist). Every one of these primitives
shows one independent aspect of network deformation.
In Dare's formulation of GSA and the later application program NETAN [Craymer et al.,
1988], little attention was paid to possible values of observation errors that could cause the
virtual deformation analysed by GSA. Values equal to the standard deviations of individual
observations were used to generate the studied deformation. GSA thus starts where Baarda's
reliability analysis ends. This became obvious to us at the outset of this contract, and we
decided to put these two techniques together to obtain a full image of the strength of a
geodetic network. All that is required to 'marry' the two techniques is to take the maximum
errors (blunder) undetectable by the standard statistical tests for outliers as estimated by the
reliability analysis and use them as values that can cause the virtual deformation of the
network in GSA. The result is that equation (4.25), which generates the vector of
displacement gradients (strains), changes to
(5.1)

e = ..[Link]* ,
where
a.

*
1

O"j

(5.2)

= :-r- .

-vri

In other words, this equation is created by substituting

vei from equation (3.20b) for O"i in

equation (4.25). The subsequent treatment of e remains the same as in GSA.

5. 2

Properties

of Robustness

Analysis

The resulting analysis, which combines the statistical aspects of Baarda's theory with the
geometrical aspects of GSA, gives an answer to the proper question that one must ask if analysing
the strength of a network; namely, What would be the worst possible deformation of a network
whose observations had been tested for outliers (and detected outliers purged) on a

5. Robustness Analysis

38

Robustness Analysis

Final Report

specific confidence level l-ao? The analysis- we call it 'network robustness analysis,' to reflect
contemporary statistical terminology where robustness means insensitivity to blunders - gives
a picture of the network's potential worst deformation in terms of the three independent
deformation primitives. Clearly, a high value of one of the primitives associated with a point,
or a region, shows a weakness (in resistance to deformation) of the network at that point, or
region, in the sense of that particular primitive (aspect). Low values, on the other hand, are
indicative of good resistance to deformation, i.e., indicative of strength.
For the purpose of designing a network that would meet specific strength criteria, it would
be necessary to formulate meaningful tolerance limits for admissible weaknesses in the three
independent primitives: scale, shear, and twist. In other words, for specifications dealing with
a design of desirably strong networks, it will be necessary to come up with a specific value of

A.o, which scales the three indicators of strength. (As we see from equation (5.1), --./'Ao is a
common scale factor to all the results obtained from the robustness analysis.) In Chapter 3, it was
shown that AO is a function of ao, the significance level selected for testing adjusted
observations for outliers, and

Po. the probability of Type II error in the testing. While a.o is

selected beforehand, prior to the outlier testing,

Po is free but should be specified for the purpose

of choosing tolerance limits in the robustness specifications.

Two types of singularities can appear in a network subjected to robustness analysis: a
geometrical singularity, and a singularity due to no redundancy. The first singularity is caused by
specific geometrical configurations when the point to be analysed is either connected only to one
other point, or when it is colinear with all the connected points. This type of singularity does not
have anything to do with the strength of the network at that point; strength simply cannot be
(reliably or at all) determined at that point. In the enhanced NETAN (see Chapter 6), strength
indicators at these singular points are simply not plotted at all.

The other type of singularity occurs at points attached to the network by observation(s)
whose redundancy number equals zero, i.e., at points whose position is derived from the
minimal number of observations with no redundancy (e.g., two intersecting directions).
5. Robustness Analysis

39

Robustness Analysis

Final Report

Because such observations are not checked at all by the test for outliers, there is no guarantee
that such observations are not burdened with huge blunders and the point in question
represents a point of infinite weakness - zero strength - in the network.

The strength

indicators at these points show very large values.

Finally, we wish to note here that robustness analysis is datum independent The proof for
the independence was shown in the previous chapter for the geometrical strength analysis and
it fully applies here as well. The consequence of this property is that any choice of a
minimally constrained adjustment model will yield the same results as far as network strength
is concerned. It must be emphasized, however, that a network adjusted with some weighted
constraints, e.g., a network for which positions of some points, together with their errors, are
known a priori, cannot be viewed in the same light. Weighted constraints become an
indivisible part of the network itself as much as the observations are, and the strength of such
a network is as much affected by the constraints (and their weights) as it is by the
observations (and their weights). The fact that the weighted constrains may also supply a
datum for the adjustment is only incidental.

5. 3

Comparison of Robustness and Geometrical Strength Analyses

To demonstrate the difference between the GSA and the robustness analysis, results of
both analyses applied to the HOACS 3D synthetic network, are shown here. For a full
discussion of the HOACS 3D network and its robustness, the reader is referred to Chapter 7.
Figures 5.1 and 5.2 show strength in scale as estimated by both techniques. The results
are somewhat similar insofar as the extreme values are concerned: the maxima (in absolute
value) are located at the same points, the northeast and the southwest corners, i.e., the
weaknesses of the network have been pinpointed by both techniques the same way. On the
other hand, the details are quite different and so are the magnitudes, as one should expect.

5. Robustness Analysis

40

Robustness Analysis
Final Report

5. Robustness Analysis

41