Lecture Notes

Adjustment Theory

Nico Sneeuw, Friedhelm Krumm, Matthias Roth

Geodätisches Institut
Universität Stuttgart

❤tt♣s✿✴✴✇✇✇✳❣✐s✳✉♥✐✲st✉tt❣❛rt✳❞❡

Rev. 4.48

August 12, 2022

© Nico Sneeuw, Friedhelm Krumm, Matthias Roth, Mohammad Tourian, 2006-2022
These are lecture notes in progress.
Please contact us (④s♥❡❡✉✇⑥❅❣✐s✳✉♥✐✲st✉tt❣❛rt✳❞❡) for remarks, errors, suggestions, etc.
Contents

1 Introduction 6
1.1 Adjustment theory – a first look . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Historical development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Least squares adjustment 13

2.1 Adjustment with observation equations . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Adjustment with condition equations . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Generalizations 21
3.1 Higher dimensions: the 𝐴-model (observation equations) . . . . . . . . . . . . . . . 21
3.2 The datum problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Linearization of non-linear observation equations . . . . . . . . . . . . . . . . . . . 26
3.4 Higher dimensions: the 𝐵-model (Condition equations) . . . . . . . . . . . . . . . . 32
3.5 Linearization of non-linear condition equations . . . . . . . . . . . . . . . . . . . . 36
3.6 Higher dimensions: the mixed model . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Weighted least squares 38

4.1 Weighted observation equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.1.1 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.1.2 Application to adjustment problems . . . . . . . . . . . . . . . . . . . . . . 41
4.1.3 Higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Weighted condition equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Stochastics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Best Linear Unbiased Estimation (blue) . . . . . . . . . . . . . . . . . . . . . . . . . 45

5 Geomatics examples 47
5.1 A-Model: Adjustment of observation equations . . . . . . . . . . . . . . . . . . . . . 47
5.1.1 Planar triangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.2 Distance Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.3 Distance and Direction Network (1) . . . . . . . . . . . . . . . . . . . . . . . 55
5.1.4 Distance and Direction Network (2a) . . . . . . . . . . . . . . . . . . . . . . 58
5.1.5 Free Adjustment: Distance and Direction Network (2b) . . . . . . . . . . . . 62
5.1.6 Overconstrained adjustment: Distance, direction and angle network . . . . 65
5.1.7 Polynomial fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 B-Model: Adjustment of condition equations . . . . . . . . . . . . . . . . . . . . . . 78
5.2.1 Planar triangle 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2.2 Planar triangle 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3
Contents

5.3 Mixed model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

5.3.1 Straight line fit using A-model with pseudo observation equations . . . . . 81
5.3.2 Straight line fit using extended B-Model . . . . . . . . . . . . . . . . . . . . 83
5.3.3 2D Similarity Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3.4 2D Affine Transformation Model I . . . . . . . . . . . . . . . . . . . . . . . 90
5.3.5 2D Affine Transformation Model II . . . . . . . . . . . . . . . . . . . . . . . 94
5.3.6 Ellipse fit under various restrictions . . . . . . . . . . . . . . . . . . . . . . 97

6 Statistics 106
6.1 Expectation of sum of squared residuals . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.3 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.4 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

7 Statistical Testing 113

7.1 Global model test: a first approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 Testing procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.3 DIA-Testprinciple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.4 Internal reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.5 External reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.6 Reliability: a synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

8 Recursive estimation 128

8.1 Partitioned model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.1.1 Batch / offline / Staple / standard . . . . . . . . . . . . . . . . . . . . . . . . 128
8.1.2 Recursive / sequential / real-time . . . . . . . . . . . . . . . . . . . . . . . . 128
8.1.3 Recursive formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.1.4 Formulation using condition equations . . . . . . . . . . . . . . . . . . . . . 129
8.2 More general . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

A Partitioning 131
A.1 Inverse Partitioning Method (IPM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
A.2 Inverse Partitioning Method: special case 1 . . . . . . . . . . . . . . . . . . . . . . . 131
A.3 Inverse Partitioning Method: special case 2 . . . . . . . . . . . . . . . . . . . . . . . 132

B Statistical Tables 134

B.1 Standard Normal Distribution z . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
B.2 Central 𝜒 2 -Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
B.3 Non-central 𝜒 2 -Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
B.4 Central t-Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
B.5 Central F-Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
B.6 Relation between F-Distribution and other distrubutions . . . . . . . . . . . . . . . 150

C Book recommendations and other material 151

C.1 Scientific books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
C.2 Popular science books, literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

4
 Contents

C.3 Other material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

5
 1 Introduction
Ausgleichungs-
rechnung
Adjustment theory deals with the optimal combination of redundant measurements together with
the estimation of unknown parameters.
Teunissen, 2000

1.1 Adjustment theory – a first look

To understand the purpose of adjustment theory consider the following simple highschool example
that is supposed to demonstrate how to solve for unknown quantities. In case 0 the price of apples
and pears is determined after doing groceries twice. After that we will discuss more interesting
shopping scenarios.

Case 0)

3 apples + 4 pears = 5.00 e
5 apples + 2 pears = 6.00 e

5 = 3𝑥 1 + 4𝑥 2
2 equations in 2 unknowns:
6 = 5𝑥 1 + 2𝑥 2
    
5 3 4 𝑥1
as matrix-vector system: =
6 5 2 𝑥2
linear algebra: 𝑦 = 𝐴𝑥

The determinant of matrix 𝐴 reads det 𝐴 = 3 · 2 − 5 · 4 = −14. Thus the above linear system can be
inverted:       
−1 𝑥1 1 2 −4 5 1
𝑥 = 𝐴 𝑦 ⇐⇒ = =
𝑥2 −14 −5 3 6 0.5
So each apple costs 1 e and each pear 50 cents. The price can be determined because there are as
many unknowns (the price of apples and the price of pears) as there are observations (shopping
twice). The square and regular matrix 𝐴 is invertible.

Remark 1.1 (terminology) The left-hand vector 𝑦 contains the observations. The vector 𝑥 contains
the unknown parameters. The two vectors are linked through the design matrix 𝐴. The linear model
𝑦 = 𝐴𝑥 is known as the model of observation equations.

The following cases demonstrate that the idea of determining unknowns from observations is not
as straightforward as may seem from the above example.

6
 1.1 Adjustment theory – a first look

Case 1a)
If one buys twice as much apples and pears the second time, and if one has to pay twice as much as
well, no new information is added to the system of linear equations
     
3a + 4p = 5 e 5 3 4 𝑥1
⇐⇒ =
6a + 8p = 10 e 10 6 8 𝑥2
The matrix 𝐴 has linearly dependent columns (and rows), i.e. it is singular. Correspondingly det 𝐴 =
0 and the inverse 𝐴−1 does not exist. The observations (5 e and 10 e) are consistent, but the vector
𝑥 of unknowns (price per apple or pear) cannot be determined. This situation will return later with
so-called datum problems. Seemingly trivial, case 1a) is of fundamental importance.

Case 1b)
Suppose the same shopping scenario as above, but now one needs to pay 8 e the second time.
 
5
𝑦=
8
In this alternative scenario, the matrix is still singular and 𝑥 cannot be determined. But worse still,
the observations 𝑦 are inconsistent with the linear model. Mathematically, they do not fulfil the
compatibility conditions. In data analysis inconsistency is not necessarily a weakness. In fact, it
may add information to the linear system. It might indicate observation errors (in 𝑦), for instance a
miscalculation of the total grocery bill. Or it might indicate an error in the linear model: the prices
may have changed in between, which leads to a different 𝐴.

Case 2)
We go back to the consistent and invertible case 0. Suppose a third combination of apples and pears
 
gives an inconsistent result.
© ª ©
5 3 4 ª 𝑥1
­ 6 ® = ­ 5 2 ® 𝑥2
­ ® ­ ®
« ¬ «
3 1 2 ¬
The third row is inconsistent with 𝑥 1 = 1, 𝑥 2 = 12 from case 0. But one can equally maintain that the
first row is inconsistent with the second and third. In short, we have redundant and inconsistent
information: the number of observations (𝑚 = 3) is larger than the number of unknowns (𝑛 = 2).
Consequently, matrix 𝐴 is not a square matrix.
Although a standard inversion is not possible anymore, redundancy is a positive characteristic in
engineering disciplines. In data analysis redundancy provides information on the quality of the
observations, it strengthens the estimation of the unknowns and allows us to perform statistical
tests. Thus, redundancy provides a handle to quality control.
But obviously the inconsistencies have to be eliminated. This is done by spreading them out in
an optimal way. This is the task of adjustment: to combine redundant and inconsistent data in an
optimal way. Two main questions will be addressed in the first part of this course:
• How to combine inconsistent data optimally?
• Which criterion defines what optimal is?

7
 1 Introduction

Errors
The inconsistencies may be caused by model errors. If the green grocer changed his prices between
two rounds of shopping we need to introduce new parameters. In surveying, however, the observa-
tion models are usually well-defined, e.g. the sum of angles in a plane triangle equals 𝜋. So usually
the inconsistencies arise from observation errors. To make the linear system 𝑦 = 𝐴𝑥 consistent
again, we need to introduce an error vector 𝑒 with the same dimension as the observation vector.

𝑦 =𝐴 𝑥 + 𝑒 . (1.1)
𝑚×1 𝑚×𝑛 𝑛×1 𝑚×1

Errors go under several names: inconsistencies, residuals, improvements, deviations, discrepancies,

and so on.

Remark 1.2 (sign convention) In many textbooks the error vector is put at the same side of the
equation as the observations: 𝑦 + 𝑒 = 𝐴𝑥. Where to put the 𝑒-vector is rather a philosophical question.
Practically, though, one should be aware of the definitions used, how the sign of 𝑒 is defined.

Three different types of errors are usually identified:

grober Fehler
i) Gross error, also known as blunder or outlier.
systematischer
Fehler ii) Systematic error, or bias.

Zufallsfehler iii) Random error.

These types are visualized in fig. 1.1. In this figure, one can think of the marks left behind by the
arrow points in a game of darts, in which one attempts to aim at the bull’s eye. Whatever the type,

(a) gross error (b) systematic error (c) random error

Figure 1.1: Different types of errors.

Zufallsvariable errors are stochastic quantities. Thus, the vector 𝑒 is a (𝑚-dimensional) stochastic variable. The
vector of observations is consequently also a stochastic variable. Such quantities will be underlined,
if necessary:
𝑦 = 𝐴𝑥 + 𝑒 .

Nevertheless, it will be assumed in the sequel that 𝑒 is drawn from a distribution of random errors.

8
 1.2 Historical development

1.2 Historical development

The question how to combine redundant and inconsistent data has been treated in many different
ways in the past. To compare the different approaches, the following mathematical framework is
used:

observation model: 𝑦 = 𝐴𝑥
combination: 𝐿 𝑦 = 𝐿 𝐴 𝑥
𝑛×𝑚 𝑚×1 𝑛×𝑚 𝑚×𝑛 𝑛×1

invert: 𝑥 = (𝐿𝐴) −1 𝐿𝑦
= 𝐵𝑦

From a modern viewpoint matrix 𝐵 is a left-inverse of 𝐴 because 𝐵𝐴 = 𝐼 . Note that such a left-inverse
is not unique, as it depends on the choice of the combination matrix 𝐿.

Method of selected points – before 1750

A simple way out of the overdetermined problem is to select only so many observations (“points”)
as there are unknowns. The remaining unused observations may be used to validate the estimated
result. This is the so-called method of selected points. Suppose one uses only the first 𝑛 observations.
Then:
𝐿 =[ 𝐼 0 ]
𝑛×𝑚 𝑛×𝑛 𝑛×(𝑚−𝑛)

The trouble with this approach, obviously, is the arbitrariness of the choice of 𝑛 observations. There
are 𝑚𝑛 choices.

From a modern perspective the method of selected points resembles the principle of cross-validation.
The idea of this principle is to deliberately leave out a limited number of observations during the
estimation and to use the estimated parameters to predict values for those observations that were left
out. A comparison between actual and predicted observations provides information on the quality
of the estimated parameters.

Method of averages – ca. 1750

In 1714 the British government offered the Longitude Prize for the precise determination of a ship’s
longitude. Tobias Mayer’s1 approach was to determine longitude, or rather time, through the mo-
tion of the moon. In the course of his investigations he needed to determine the libration of the
moon through measurement to lunar surface (craters). This led him to overdetermined systems of
observation equations:
𝑦 =𝐴 𝑥
27×1 27×3 3×1

Mayer called them equations of conditions, which is, from today’s view point, an unfortunate desig-
nation.
1 Tobias Mayer (1723–1762) made the breakthrough that enabled the lunar distance method to become a practicable way

of finding longitude at sea. As a young man, he displayed an interest in cartography and mathematics. In 1750, he
was appointed professor in the Georg-August Academy in Göttingen, where he was able to devote more time to his
interests in lunar theory and the longitude problem. From 1751 to 1755, he had an extensive correspondence with
Leonhard Euler, whose work on differential equations enabled Mayer to calculate lunar distance tables.

9
1 Introduction

Mayer’s adjustment strategy:

• distribute the observations into three groups
• sum up the equations within each group
• solve the 3 × 3-system.

©1 1 ··· 1 0 0 0 0 0 0 0 0ª
𝐿 = ­­ 0 0 0 0 1 1 · · · 1 0 0 0 0 ®®
«0 0 0 0 0 0 0 0 1 1 ··· 1¬
3×27

Mayer actually believed each aggregate of 9 observations

√ to be “9 times more precise” than a single
observation. Today we know that this should be 9 = 3.

Euler’s attempt – 1749

Leonhard Euler2
Background:
• Orbital motion of the Saturn under influence of Jupiter
• Stability of the solar system
• Prize (1748) of the Academy of Sciences, Paris
75 observations from the years 1582–1745; 6 unknowns =⇒ Given up!
Euler was mathematician −→ “Error bounds”

Laplace’s attempt – ca. 1787

Laplace3
Background: Saturn, too
Reformulated: 4 unknowns
Best data: 24 observations
Approach: like Mayer, but other combinations:

𝑦 =𝐴 𝑥
24×1 24×4 4×1

𝐿 𝑦 =𝐿 𝐴 𝑥
4×24 24×1 4×24 24×4 4×1

𝑥 = (𝐿𝐴) −1 𝐿𝑦
2 Euler (1707–1783) was a Swiss mathematician and physicist. He is considered to be one of the greatest mathematicians
who ever lived. Euler was the first to use the term function (defined by Leibniz in 1694) to describe an expression
involving various arguments; i.e. 𝑦 = 𝐹 (𝑥). He is credited with being one of the first to apply calculus to physics.
3 Pierre-Simon, Marquis de Laplace (1749–1827) was a French mathematician and astronomer who put the final capstone

on mathematical astronomy by summarizing and extending the work of his predecessors in his five volume Mécanique
Céleste (Celestial Mechanics) (1799–1825). This masterpiece translated the geometrical study of mechanics used by
Newton to one based on calculus, known as physical mechanics. He is also the discoverer of Laplace’s equation and the
Laplace transform, which appear in all branches of mathematical physics – a field he took a leading role in forming.
He became count of the Empire in 1806 and was named a marquis in 1817 after the restoration of the Bourbons.
Pierre-Simon Laplace was among the most influential scientists in history.

10
 1.2 Historical development

 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 −1 −1 −1 −1 −1 −1 −1 −1 −1 −1 −1 −1 
1 1 1 1 1 1 1 1 1
𝐿 =  −1 1 0 0 −1 0 0 1 1 0 −1 0 0 1 1 
1 1 1 1 1 1 1 1 1

 0 0 −1 −1 0 1 1 0 0 −1 0 1 1 0 0 
0 1 1 0 0 −1 0 0 1

4×24
1 0 0 −1 −1 0 1 1 0

Method of least absolute deviation – 1760

Roger Boscovich4
Ellipticity of the Earth
5 observations (Quito, Cape Town, Rome, Paris, Lapland)
2 unknowns
𝑎(1 − 𝑒 2 )
𝑀 (𝜑) = 3
(1 − 𝑒 2 sin2 𝜑) 2
3
= 𝑎(1 − 𝑒 2 ) (1 + 𝑒 2 sin2 𝜑 + ...)
2
𝑀 (0) = 𝑎(1 − 𝑒 2 ) < 𝑎
2
𝑀 ( 𝜋2 ) = 𝑎 1−𝑒 3 = √ 𝑎 >𝑎
(1−𝑒 2 ) 2 1−𝑒 2

= 𝑥 1 + sin2 𝜑𝑥 2

First attempt: All 52 = 2!(5−2)!
5!
= 5·4·3·2·1
2·1·3·2·1 = 10 combinations with 2 observations each.
=⇒ 10 systems of equations (2 × 2)
=⇒ 10 solutions
Comparison of results.
His result: gross variations of the ellipticity =⇒ reject the ellipsoidal hypothesis.

Second attempt: The mean deviation (or sum of deviations) should be zero:
Õ
5
𝑒𝑖 = 0 ,
𝑖=1

and the sum of absolute deviations should be minimum:

Õ
5
|𝑒𝑖 | = min .
𝑖=1

This is an objective adjustment criterion, although its implementation is mathematically difficult.

This is the approach of 𝐿1 -norm minimization.

Method of least squares – 1805

In 1805 Legendre5 published his method of least squares (in French: moindres carrés). The name least Methode der
squares refers to the fact the sum of square residuals is minimized. Legendre developed the method kleinsten
Quadrate
4 Rudjer Josip Bošković aka. Roger Boscovich (1711–1787) was a Croatian Jesuit, a mathematician and an innovative
physicist, he was active also in astronomy, nature philosophy and poetry as well as technician and geodesist.
5 Adrien-Marie Legendre (1752–1833) was a French mathematician. He made important contributions to statistics, num-

ber theory, abstract algebra and mathematical analysis.

11
1 Introduction

for the determination of orbits of comets and to derive the Earth ellipticity. As will be derived in the
next chapter, the matrix 𝐿 will be the transposed of the design matrix 𝐴:

Õ
5
L= 𝑒𝑖2 = 𝑒 T𝑒 = (𝑦 − 𝐴𝑥) T (𝑦 − 𝐴𝑥) = min
𝑥ˆ
𝑖=1

⇐⇒ 𝐿 = 𝐴T
⇐⇒ 𝑥ˆ = ( 𝐴T𝐴 ) −1 𝐴T 𝑦
𝑛×1
|{z} 𝑛×𝑚 𝑚×1
𝑛×𝑛

After Legendre’s publication Gauss states that he already developed and used the method of least
squares in 1794. He published his own theory only several years later. A bitter argument over the
scientific priority broke out. Nowadays it is acknowledged that Gauss’s claim of priority is very
likely valid but that he refrained from publication because he found his results still premature.

12
2 Least squares adjustment

Legendre’s method of least squares is actually not a method. Rather, it provides the criterion for the
optimal combination of inconsistent data: combine the observations such that the sum of squared
residuals is minimal. It was seen already that this criterion defines the combination matrix 𝐿:

𝐿𝑦 = 𝐿𝐴𝑥 =⇒ 𝑥 = (𝐿𝐴) −1 𝐿𝑦 .

But what is so special about 𝐿 = 𝐴T ? In this chapter we will derive the equations of least squares
adjustment from several mathematical viewpoints:
• geometry: smallest distance (Pythagoras)
• linear algebra: orthogonality between the optimal 𝑒 and the columns of 𝐴: 𝐴T𝑒 = 0
• calculus: minimizing target function → differentiation
• probability theory: BLUE (Best Linear Unbiased Estimate)
These viewpoints are elucidated by a simple but fundamental example in which a distance is mea-
sured twice.

2.1 Adjustment with observation equations

We will start with the model of the introduction 𝑦 = 𝐴𝑥. This is the model of observation equations, vermittelnde
in which observations are linearly related to unknowns. Ausgleichung

Suppose that, in order to determine a certain distance, it is measured twice. Let the unknown dis-
tance be 𝑥 and the observations 𝑦1 and 𝑦2 :
    
𝑦1 = 𝑥 𝑦1 1 direkte
=⇒ = 𝑥 =⇒ 𝑦 = 𝑎𝑥 (2.1)
𝑦2 = 𝑥 𝑦2 1 Beobachtungen

If 𝑦1 = 𝑦2 the equations are consistent and the parameter 𝑥 clearly solvable: 𝑥 = 𝑦1 = 𝑦2 . If, on
the other hand, 𝑦1 ≠ 𝑦2 the equations are inconsistent and 𝑥 not solvable directly. Given a limited
measurement precision the latter scenario will be more likely. Let’s therefore take into account
measurement errors 𝑒.      
𝑦1 1 𝑒
= 𝑥+ 1 =⇒ 𝑦 = 𝑎𝑥 + 𝑒 (2.2)
𝑦2 1 𝑒2

A geometric view
The column vector 𝑎 spans up a line 𝑦 = 𝑎𝑥 in R2 . This line is the 1D model space or range space Spaltenraum
of 𝐴: R (𝐴). Inconsistency of the observation vector means that 𝑦 does not lie on this line. Instead,

13
2 Least squares adjustment

there is some vector of discrepancies 𝑒 that connects the observations to the line. Both this vector
𝑒 and the point on the line, defined by the unknown parameter 𝑥, must be found, see the left panel
of fig. 2.1. Adjustment of observations is about finding the optimal 𝑒 and 𝑥. An intuitive choice for

(a) Inconsistent data: the observation vec- (b) Least squares adjustment means or-
tor 𝑦 is not in the model space, i. e. not thogonal projection of 𝑦 onto the line
on the line spanned by 𝑎. 𝑎𝑥. This guarantees the shortest 𝑒.

Figure 2.1

Í Í
“optimality” is to make the vector 𝑒 as short as possible. The shortest possible 𝑒 is indicated by a
ˆ The squared length 𝑒ˆT𝑒ˆ = 𝑖 𝑒ˆ𝑖2 is the smallest of all possible 𝑒 T𝑒 = 𝑖 𝑒𝑖2 , which explains the
hat: 𝑒.
name least squares. If 𝑒ˆ is determined, we will at the same time know the optimal 𝑥. ˆ
How do we get the shortest 𝑒? The right panel of fig. 2.1 show that the shortest 𝑒 is perpendicular
to 𝑎:
𝑒ˆ ⊥ 𝑎

Subtracting 𝑒ˆ from the vector of observations 𝑦 leads to the point 𝑦ˆ = 𝑎𝑥ˆ that is on the line and
closest to 𝑦. This is the vector of adjusted observations. Being on the line means that 𝑦ˆ is consistent.
If we now substitute 𝑒ˆ = 𝑦 −𝑎𝑥,
ˆ the least squares criterion leads us subsequently to optimal estimates
of 𝑥, 𝑦 and 𝑒:

orthogonality 𝑒ˆ ⊥ 𝑎 𝑎 T𝑒ˆ = 0 (2.3a)

𝑎 T (𝑦 − 𝑎𝑥)
ˆ =0 (2.3b)
normal equations 𝑎 T𝑎𝑥ˆ = 𝑎 T𝑦 (2.3c)
LS estimate of 𝑥 𝑥ˆ = (𝑎 T𝑎) −1𝑎 T𝑦 (2.3d)
LS estimate of 𝑦 𝑦ˆ = 𝑎𝑥ˆ = 𝑎(𝑎 T𝑎) −1𝑎 T𝑦 (2.3e)
LS estimate of 𝑒 𝑒ˆ = 𝑦 − 𝑦ˆ = [𝐼 − 𝑎(𝑎 T𝑎) −1𝑎 T ]𝑦 (2.3f)
sum square residuals 𝑒ˆT𝑒ˆ = 𝑦 T [𝐼 − 𝑎(𝑎 T𝑎) −1𝑎 T ]𝑦 (2.3g)

14
 2.1 Adjustment with observation equations

Exercise 2.1 Call the matrix in square brackets 𝑃 and convince yourself that the sum of squares of the
residuals (the squared length of 𝑒)
ˆ in the last line indeed follows from the line above. Two things should
be shown: that 𝑃 is symmetric, and that 𝑃𝑃 = 𝑃.

The least squares criterion leads us to the above algorithm. Indeed, the combination matrix reads
𝐿 = 𝐴T .

A calculus view
Let us define the Lagrangian or cost function:

1
L𝑎 (𝑥) = 𝑒 T𝑒 , (2.4)
2

which is half of the sum of square residuals. Its graph would be a parabola. The factor 21 shouldn’t
worry us. If we find the minimum L𝑎 , then any scaled version of it is also minimized. The task is
now to find the 𝑥ˆ that minimizes the Lagrangian. With 𝑒 = 𝑦 − 𝑎𝑥 we get the minimization problem:

1
min L𝑎 (𝑥) = min (𝑦 − 𝑎𝑥) T (𝑦 − 𝑎𝑥)
𝑥ˆ 2
 
𝑥ˆ

1 T T 1 T 2
= min 𝑦 𝑦 − 𝑥𝑎 𝑦 + 𝑎 𝑎𝑥 .
𝑥ˆ 2 2

The term 21 𝑦 T𝑦 is just a constant that doesn’t play a role in the minimization. The minimum occurs
at the location where the derivative of L𝑎 is zero (necessary condition):

dL𝑎
ˆ = −𝑎 T𝑦 + 𝑎 T𝑎𝑥ˆ = 0 .
(𝑥)
d𝑥
The solution of this equation, which happens to be the normal equation (2.3c), is the 𝑥ˆ we are looking
for:
𝑥ˆ = (𝑎 T𝑎) −1𝑎 T𝑦 .
To make sure that the derivative does not give us a maximum, we must check that the second
derivative of L𝑎 is positive at 𝑥ˆ (sufficiency condition):

d2 L𝑎
ˆ = 𝑎 T𝑎 > 0 ,
(𝑥)
d𝑥 2
which is a positive constant for all 𝑥 indeed.

Projectors
Figure 2.1 shows that the optimal, consistent 𝑦ˆ is obtained by an orthogonal projection of the original
𝑦 onto the line 𝑎𝑥. Mathematically this was translated by (2.3e) as:

𝑦ˆ = 𝑎(𝑎 T𝑎) −1𝑎 T𝑦 (2.5a)

⇐⇒ 𝑦ˆ = 𝑃𝑎𝑦 (2.5b)
T −1 T
with 𝑃𝑎 = 𝑎(𝑎 𝑎) 𝑎 . (2.5c)

15
 2 Least squares adjustment

The matrix 𝑃𝑎 is an orthogonal projector. It is an idempotent matrix, meaning:

𝑃𝑎 𝑃𝑎 = 𝑎(𝑎 T𝑎) −1𝑎 T𝑎(𝑎 T𝑎) −1𝑎 T = 𝑃𝑎 . (2.6)

It projects onto the line 𝑎𝑥 along a direction orthogonal to 𝑎. With this projection in mind, the
property 𝑃𝑎 𝑃𝑎 = 𝑃𝑎 becomes clear: if a vector has been projected already, the second projection has
no effect anymore.
Also (2.3f) can be abbreviated:

𝑒ˆ = 𝑦 − 𝑃𝑎𝑦 = (𝐼 − 𝑃𝑎 ) 𝑦 = 𝑃𝑎⊥𝑦 ,

which is also a projection. In order to give 𝑒ˆ the vector 𝑦 is projected onto a line perpendicular to
𝑎𝑥 along the direction 𝑎. And, of course, 𝑃𝑎⊥ is idempotent as well:

𝑃𝑎⊥ 𝑃𝑎⊥ = (𝐼 − 𝑃𝑎 ) (𝐼 − 𝑃𝑎 ) = 𝐼 − 2𝑃𝑎 + 𝑃𝑎 𝑃𝑎 = 𝐼 − 𝑃𝑎 = 𝑃𝑎⊥ .

Moreover, the definition (2.5c) makes clear that 𝑃𝑎 and 𝑃𝑎⊥ are symmetric. Therefore the square sum
of residuals (2.3g) could be simplified to:
T
𝑒ˆT𝑒ˆ = 𝑦 T 𝑃𝑎⊥ 𝑃𝑎⊥𝑦 = 𝑦 T 𝑃𝑎⊥ 𝑃𝑎⊥𝑦 = 𝑦 T 𝑃𝑎⊥𝑦 .

At a more fundamental level the definition of the orthogonal projector 𝑃𝑎⊥ = 𝐼 − 𝑃𝑎 can be recast
into the equation:
𝐼 = 𝑃𝑎 + 𝑃𝑎⊥ .
zerlegen Thus, we can decompose every vector, say 𝑧, into two components: one in component in a subspace
defined by 𝑃𝑎 , the other mapped onto a subspace by 𝑃𝑎⊥ :


𝑧 = 𝐼𝑧 = 𝑃𝑎 + 𝑃𝑎⊥ 𝑧 = 𝑃𝑎 𝑧 + 𝑃𝑎⊥𝑧 .

In the case of ls adjustment, the subspaces are defined by the range space R (𝑎) and its orthogonal
complement R (𝑎) ⊥ :
𝑦 = 𝑃𝑎𝑦 + 𝑃𝑎⊥𝑦 = 𝑦ˆ + 𝑒ˆ ,
which is visualized in fig. 2.1.

Numerical example
With 𝑎 = (1 1) T we will follow the steps from (2.3):

(𝑎 T𝑎)𝑥ˆ = 𝑎 T𝑦 ←→ 2𝑥ˆ = 𝑦1 + 𝑦2
1
𝑥ˆ = (𝑎 T𝑎) −1𝑎 T𝑦 ←→ 𝑥ˆ = (𝑦1 + 𝑦2 ) (average)
2
   
𝑦ˆ1 1 𝑦1 + 𝑦2
𝑦ˆ = 𝑎(𝑎 T𝑎) −1𝑎 T𝑦 ←→ =
𝑦ˆ2 2 𝑦1 + 𝑦2
   
𝑒ˆ1 1 𝑦1 − 𝑦2
𝑒ˆ = 𝑦 − 𝑦ˆ ←→ = (error distribution)
𝑒ˆ2 2 −𝑦1 + 𝑦2
1
𝑒ˆT𝑒ˆ ←→ (𝑦1 − 𝑦2 ) 2 (least squares)
2

16
 2.2 Adjustment with condition equations

Exercise 2.2 Verify that the projectors are

   
1 11 1 1 −1
𝑃𝑎 = and 𝑃𝑎⊥ = 𝐼 − 𝑃𝑎 =
2 11 2 −1 1

and check the equations 𝑦ˆ = 𝑃𝑎𝑦 and 𝑒ˆ = 𝑃𝑎⊥𝑦 with the numerical results above.

2.2 Adjustment with condition equations

In the ideal case, in which the measurements 𝑦1 and 𝑦2 are without error, both observations would
be equal: 𝑦1 = 𝑦2 or 𝑦1 − 𝑦2 = 0. In matrix notation:
  𝑦 
1
1 −1 =0 =⇒ 𝑏T 𝑦 = 0 . (2.7)
𝑦2 1×2 2×1 1×1

In reality, though, both observations do contain errors, i.e. they are not equal: 𝑦1 −𝑦2 ≠ 0 or 𝑏 T𝑦 ≠ 0.
Instead of 0 one would obtain a misclosure 𝑤. If we recast the observation equation into 𝑦 − 𝑒 = 𝑎𝑥, Widerspruch
it is clear that it is (𝑦 − 𝑒) that has to obey the above condition:

𝑏 T (𝑦 − 𝑒) = 0 =⇒ 𝑤 := 𝑏 T𝑦 = 𝑏 T𝑒 . (2.8)

In this condition equation the vector 𝑒 is unknown. The task of adjustment according to the model Bedingungs-
of condition equations is to find the smallest possible 𝑒 that fulfills the condition (2.8). At this stage, gleichung
the model of condition equations does not involve any kind of parameters 𝑥.

A geometric view
The condition (2.8) describes a line with normal vector 𝑏 that goes through the point 𝑦. This line is
the set of all possible vectors 𝑒. We are looking for the shortest 𝑒, i.e. the point closest to the origin.
Figure 2.2 makes it clear that 𝑒ˆ is perpendicular to the line 𝑏 T𝑒 = 𝑤. So 𝑒ˆ lies on a line through 𝑏.
Geometrically, 𝑒ˆ is achieved by projecting 𝑦 onto a line through 𝑏. Knowing the definition of the
projectors from the previous section, we here define the following estimates by using the projector Schätzungen
𝑃𝑏 :

𝑒ˆ = 𝑃𝑏 𝑦 = 𝑏 (𝑏 T𝑏) −1𝑏 T𝑦 = 𝑏 (𝑏 T𝑏) −1𝑤 (2.9a)

𝑦ˆ = 𝑦 − 𝑒ˆ = 𝑦 − 𝑏 (𝑏 T𝑏) −1𝑏 T𝑦
= [𝐼 − 𝑏 (𝑏 T𝑏) −1𝑏 T ]𝑦 = 𝑃𝑏⊥𝑦 (2.9b)
𝑒ˆT𝑒ˆ = 𝑦 T 𝑃𝑏 𝑦 = 𝑦 T𝑏 (𝑏 T𝑏) −1𝑏 T𝑦 (2.9c)

Exercise 2.3 Confirm that the orthogonal projector 𝑃𝑏 is idempotent and verify that the equation for
𝑒ˆT𝑒ˆ is correct.

17
2 Least squares adjustment

(a) The condition equation describes a line in R2 , (b) Least squares adjustment with condition
perpendicular to 𝑏 and going through 𝑦. We equations means orthogonal projection of 𝑦
are looking for a point 𝑒 on this line. onto the line through 𝑏. This guarantees the
shortest 𝑒.

Figure 2.2

Numerical example
 
With 𝑏 T = 1 −1 we get

1
=⇒ (𝑏 T𝑏) −1 =
𝑏 T𝑏 = 2
2
    
T −1 T 1 1 1 1 −1
𝑃𝑏 = 𝑏 (𝑏 𝑏) 𝑏 = 1 −1 =
2 −1 2 −1 1

 
1 𝑦1 − 𝑦2
=⇒ 𝑒ˆ = 𝑃𝑏 𝑦 =
2 −𝑦1 + 𝑦2
     
⊥ 10 1 1 −1 1 11
𝑃𝑏 = 𝐼 − 𝑃𝑏 = − =
01 2 −1 1 2 11
 
1 𝑦1 + 𝑦2
=⇒ 𝑦ˆ = 𝑃𝑏⊥𝑦 =
2 𝑦1 + 𝑦2

These results for 𝑦ˆ and 𝑒ˆ are the same as those for the adjustment with observation equations. The
estimator 𝑦ˆ describes the mean of the two observations, whereas the estimator 𝑒ˆ distributes the
inconsistencies equally. Also note that 𝑃𝑏 = 𝑃𝑎⊥ and vice versa.

18
 2.2 Adjustment with condition equations

A calculus view
Alternatively we can again determine the optimal 𝑒 by minimizing the target function L𝑏 (𝑒) = 𝑒 T𝑒,
but now under the condition 𝑏 T (𝑦 − 𝑒) = 0:

min L𝑏 (𝑒) = 𝑒 T𝑒 under 𝑏 T (𝑦 − 𝑒) = 0 , (2.10a)

𝑒ˆ
1
min L𝑏 (𝑒, 𝜆) = 𝑒 T𝑒 + 𝜆 T (𝑏 T𝑦 − 𝑏 T𝑒) . (2.10b)
ˆ 𝜆ˆ
𝑒, 2

The main trick here – due to Lagrange – is to not consider the condition as a constraint or limitation
of the minimization problem. Instead, the minimization problem is extended. To be precise, the
condition is added to the original cost function, multiplied by a factor 𝜆. Such factors are called
Lagrangian multipliers. In case of more than one condition, each gets its own multiplier. The target
function L𝑏 is now a function of 𝑒 and 𝜆.
The minimization problem now exists in finding the 𝑒ˆ and 𝜆ˆ that minimize the extended L𝑏 . Thus
we need to derive the partial derivatives of L𝑏 towards 𝑒 and 𝜆. Next, we impose the conditions that
ˆ
these partial derivatives are zero when evaluated in 𝑒ˆ and 𝜆.

𝜕L ˆ =0
(𝑒,
ˆ 𝜆) =⇒ 𝑒ˆ − 𝑏 𝜆ˆ = 0
𝜕𝑒
𝜕L ˆ =0
(𝑒,
ˆ 𝜆) =⇒ 𝑏 T𝑦 − 𝑏 T𝑒ˆ = 0
𝜕𝜆
In matrix terms, the minimization problem leads to:
    
𝐼 −𝑏 𝑒ˆ 0
= . (2.11)
−𝑏 T 0 𝜆ˆ −𝑏 T𝑦

Because of the extension of the original minimization problem, this system is square. It might be
inverted in a straightforward manner, see also A.1. Instead, we will solve it stepwise. First, rewrite
the first line:
𝑒ˆ − 𝑏 𝜆ˆ = 0 =⇒ 𝑒ˆ = 𝑏 𝜆ˆ .

This result is then used to eliminate 𝑒ˆ in the second line:

𝑏 T𝑦 − 𝑏 T𝑏 𝜆ˆ = 0 ,

which is solved by:

𝜆ˆ = (𝑏 T𝑏) −1𝑏 T𝑦 .

With this result we go back to the first line:

𝑒ˆ − 𝑏 (𝑏 T𝑏) −1𝑏 T𝑦 = 0 ,

which is finally solved by:

𝑒ˆ = 𝑏 (𝑏 T𝑏) −1𝑏 T𝑦 = 𝑃𝑏 𝑦 .

This is the same estimator 𝑒ˆ as (2.9a).

19
2 Least squares adjustment

2.3 Synthesis

Both the calculus and geometric approach provide the same ls estimators. This is due to

𝑃𝑎 = 𝑃𝑏⊥ and 𝑃𝑏 = 𝑃𝑎⊥ ,

as can be seen in fig. 2.3. The deeper reason is that 𝑎 is perpendicular to 𝑏:

 1
T
𝑏 𝑎 = 1 −1 = 0, (2.12)
1

which fundamentally connects the model with observation equations to the model with condition
equations. Starting with the observation equation, and applying the orthogonality, one ends up with
the condition equation:

𝑏T 𝑏 T𝑎=0
𝑦 = 𝑎𝑥 + 𝑒 −→ 𝑏 T𝑦 = 𝑏 T𝑎𝑥 + 𝑏 T𝑒 −→ 𝑏 T𝑦 = 𝑏 T𝑒 .

Figure 2.3: Least squares adjustment with

observation equations and
with condition equations in
terms of the projectors 𝑃𝑎 and
𝑃𝑏 .

20
3 Generalizations

In this chapter we will apply several generalizations. First we will take the ls adjustment problems
to higher dimensions. What we will basically do is replace the vector 𝑎 by an (𝑚 × 𝑛) matrix 𝐴
and replace the vector 𝑏 by an (𝑚 × (𝑚 − 𝑛)) matrix 𝐵. The basic structure of the projectors and
estimators will remain the same.
Moreover, we need to be able to formulate the 2 ls problems with constant terms:

𝑦 = 𝐴𝑥 + 𝑎 0 + 𝑒 and 𝐵 T (𝑦 − 𝑒) = 𝑏 0 .

Next, we will deal with nonlinear observation equations and nonlinear condition equations. This
will involve linearization, the use of approximate values, and iteration.
We will also touch upon the datum problem, which arises if 𝐴 contains dependent columns. Math-
ematically we have rank 𝐴 < 𝑛 so that the normal matrix has det 𝐴T𝐴 = 0 and is not invertible.
At the end we will merge both models in order to establish the so-called general model of adjustment
theory.

3.1 Higher dimensions: the 𝑨-model (observation equations)

The vector of observations 𝑦, the vector of inconsistencies 𝑒 and their respective ls-estimators will
be (𝑚 × 1) vectors. The vector 𝑥 will contain 𝑛 unknown parameters. Thus the redundancy, that is
the number of redundant observations, is:

redundancy: 𝑟 = 𝑚 − 𝑛 .

Geometry
𝑦 = 𝐴𝑥 +𝑒 is the multidimensional extension of 𝑦 = 𝑎𝑥 +𝑒 with given (reduced) vector of observations Absolutglied-
𝑦. vektor
We split 𝐴 in its 𝑛 column vectors 𝑎𝑖 , 𝑖 = 1, . . . , 𝑛
𝑚×1

𝐴 = [ 𝑎 1 , 𝑎 2 , 𝑎 3 , . . . , 𝑎𝑛 ]
𝑚×𝑛 𝑚×1 𝑚×1 𝑚×1 𝑚×1

Õ
𝑛
𝑦 = 𝑎𝑖 𝑥 𝑖 + 𝑒 ,
𝑚×1 𝑚×1 1×1 𝑚×1
𝑖=1

which span an 𝑛-dimensional vector space as a subspace of E𝑚 .

21
 3 Generalizations

Example: 𝑚 = 3, 𝑛 = 2 ( 𝑦 spans an E3 )
𝑚×1

(a) The vectors 𝑦, 𝑎 1 and 𝑎 2 all lie in R3 . (b) To see that 𝑦 is inconsistent, the space
spanned by 𝑎 1 and 𝑎 2 , is shown as the base
plane. It is clear that the observation vector
is not in this plane 𝑦 ∉ R (𝐴), i.e. 𝑦 cannot be
written as a linear combination of 𝑎 1 and 𝑎 2 .

Figure 3.1

𝑒ˆ = 𝑃𝐴⊥𝑦 = [𝐼 − 𝐴(𝐴T𝐴) −1𝐴T ]𝑦

𝑦ˆ = 𝑃𝐴𝑦 = 𝐴(𝐴T𝐴) −1𝐴T𝑦 = 𝐴𝑥ˆ
𝑥ˆ = (𝐴T𝐴) −1𝐴T𝑦

genau dann, (𝐴T𝐴) −1 exists iff rank 𝐴 = 𝑛 = rank(𝐴T𝐴).

wenn
Calculus

1
L𝐴 (𝑥) = 𝑒 T𝑒
2
1
= (𝑦 − 𝐴𝑥) T (𝑦 − 𝐴𝑥)
2
1 1 1 1 𝑥
= 𝑦 T𝑦 − 𝑦 T𝐴𝑥 − 𝑥 T𝐴T𝑦 + 𝑥 T𝐴T𝐴𝑥 −→ min
2 2 2 2

𝜕L
(𝑥)
ˆ =0 =⇒ 𝑒ˆ = 𝑦 − 𝑦ˆ = [𝐼 − 𝐴(𝐴T𝐴) −1𝐴T ]𝑦 = 𝑃𝐴⊥𝑦
𝜕𝑥

22
 3.1 Higher dimensions: the 𝐴-model (observation equations)

𝑃𝐴⊥ idempotent?
𝑃𝐴⊥ 𝑃𝐴⊥ = [𝐼 − 𝐴(𝐴T𝐴) −1𝐴T ] [𝐼 − 𝐴(𝐴T𝐴) −1𝐴T ]
= 𝐼 − 2𝐴(𝐴T𝐴) −1𝐴T + 𝐴(𝐴T𝐴) −1 𝐴T𝐴(𝐴T𝐴) −1 𝐴T
| {z }
=𝐼
T −1 T
= 𝐼 − 𝐴(𝐴 𝐴) 𝐴
= 𝑃𝐴⊥
𝑦ˆ = 𝑃𝐴𝑦 = 𝐴(𝐴T𝐴) −1𝐴T𝑦

Example: height network

Figure 3.2: Height network with distances between points.

ℎ 1B = 𝐻 B − 𝐻 1 + 𝑒 1B
ℎ 13 = 𝐻 3 − 𝐻 1 + 𝑒 13
ℎ 12 = 𝐻 2 − 𝐻 1 + 𝑒 12
ℎ 32 = 𝐻 2 − 𝐻 3 + 𝑒 32
ℎ 1A = 𝐻 A − 𝐻 1 + 𝑒 1A
Δℎ T := [ℎ 1B , ℎ 13 , ℎ 12 , ℎ 32 , ℎ 1A ] vector of levelled height differences
𝐻 1 , 𝐻 2 , 𝐻 3 unknown heights of points 𝑃 1, 𝑃2, 𝑃3
𝐻 A , 𝐻 B given bench marks
In matrix notation:
0 0 © 𝐻1 ª
© 1B ª © ª © B ª © 1B ª
ℎ −1 𝐻 𝑒
­ ℎ 13 ® ­ −1
­ ® ­ 0 1 ®® ­­ 𝐻 2 ®® ­­ 0 ®® ­­ 𝑒 13 ®®
­ ℎ 12 ® = ­ −1 1 0 ®® « 𝐻 3 ¬ + ­­ 0 ®® + ­­ 𝑒 12 ®®
­ ® ­
­ℎ ® ­ 0 1 −1 ®® ­ 0 ® ­𝑒 ®
­ 32 ® ­ ­ ® ­ 32 ®
« ℎ 1A ¬ « −1 0 0¬ « A ¬ « 𝑒 1A ¬
𝐻
0 0 © 𝐻1 ª
© 1B ª © ª © 1B ª
ℎ − 𝐻𝐵 −1 𝑒
­ ℎ 13 ® ­ −1 ­ ®
0 1 ®® ­ 𝐻 2 ® ­­ 𝑒 13 ®®
­ ® ­
­ ℎ 12 ® = ­ −1 1 0 ®® « 𝐻 3 ¬ + ­­ 𝑒 12 ®®
­ ® ­ ∼ 𝑦 =𝐴 𝑥 + 𝑒
­ ℎ ® ­ 0 1 −1 ®® ­𝑒 ®
­ ® ­ ­ 32 ®
5×1 5×3 3×1 5×1
32

« ℎ 1A − 𝐻 A ¬ « −1 0 0¬ « 𝑒 1A ¬

23
3 Generalizations

3.2 The datum problem

So far we have disregarded the fact that the matrix 𝐴T𝐴 might not be invertible because it is rank defi-
cient. From matrix algebra it is known that the rank of the normal equation matrix 𝑁 := 𝐴T𝐴, rank 𝑁 ,
equals the the rank of 𝐴, rank 𝐴. If it should happen now that – for some reason – matrix 𝐴 is rank
deficient, then the normal equation matrix 𝑁 =𝐴T𝐴 cannot be inverted. The following statements
are equivalent:
• Matrix 𝐴 rank deficient (rank 𝐴 < 𝑛),
𝑚×𝑛

• 𝐴 has linear dependent columns,

• 𝐴𝑥 = 0 has non-trivial solution 𝑥 hom ≠ 0, i.e. the null space N (𝐴) of 𝐴 is not empty,
• det(𝐴T𝐴) = 0,
• 𝐴T𝐴 has zero eigenvalues.
Let us investigate this problem of rank deficiency of 𝐴 and 𝑁 using levelling observations between
points 𝑃 1, 𝑃2 and 𝑃 3 of the height network shown in fig. 3.2.

ℎ 12 = 𝐻 2 − 𝐻 1 


 © ℎ 12 ª © −1 1 0 ª © 𝐻 1 ª
­ ℎ 13 ® = ­ −1 0 1 ® ­ 𝐻 2 ®
ℎ 13 = 𝐻 3 − 𝐻 1

=⇒ ­ ® ­ ®­ ®
ℎ 32 = 𝐻 2 − 𝐻 3 
 « 32 ¬ «
ℎ 0 1 −1 ¬ « 𝐻3 ¬
=⇒ 𝑦 =𝐴 𝑥
3×1 3×3 3×1

• 𝑚 = 3, 𝑛 = 3, rank 𝐴 = 2 =⇒ 𝑑 = 𝑛 − rank 𝐴 = 1 =⇒ 𝑟 = 𝑚 − (𝑛 − 𝑑) = 1,
0 1 1 0
• det 𝐴 = −1 , −(−1) = 1 + (−1) = 0,
1 −1 1 −1
• =⇒ 𝐴 and 𝑁 = 𝐴T𝐴 are not invertible,
• 𝑑 := dim N (𝐴) > 0,
• 𝐴𝑥 = 0 has a nontrivial solution =⇒ homogeneous solution 𝑥 hom ≠ 0.
=⇒ 𝑥 + 𝜆𝑥 hom is a solution of 𝑦 = 𝐴𝑥 because

|{z}
𝐴 (𝑥 + 𝜆𝑥 hom ) = 𝐴𝑥 + 𝜆 𝐴𝑥 hom = 𝐴𝑥 = 𝑦
=0

is fullfilled.

Interpretation:
• Unknown heights can be changed by an arbitrary constant height shift without affecting the
observations.
• Observed height differences are not sensitive to the null space N (𝐴).

24
 3.2 The datum problem

Solution approach 1: reduce solution space

• Fix 𝑑 = dim N (𝐴) unknowns and eliminate corresponding columns in 𝐴 so that the rank of
A, rank 𝐴 = 𝑛 − 𝑑, is full.
• Move fixed unknowns to the observation vector, e.g. fix 𝐻 1 :

 
© ℎ 12 + 𝐻 1 ª © 1 0 ª 𝐻 2
­ ℎ 13 + 𝐻 1 ® = ­ 0 1 ® 𝐻 3
=⇒ ­ ® ­ ®
« ℎ 32 ¬ « 1 −1 ¬

Solution approach 2: augment solution space

Augment solution space by adding 𝑑 = dim N (𝐴) constraints, e.g.

  © 𝐻1 ª
𝐻1 = 0 =⇒ 1 0 0 ­­ 𝐻 2 ®® = 0 ∼ 𝐷T 𝑥 = 𝑐
« 𝐻3 ¬
𝑑×𝑛 𝑛×1 𝑑×1

In order to remove the rank deficiency of 𝐴, matrix 𝐷 T must be chosen in such a way that

 
T
rank [ 𝐴 | 𝐷 ] =𝑛.
𝑛×𝑚 𝑛×𝑑

𝐴𝐷 = 0, however is not required. As an example, 𝐷 T = [1, −1, 0] is not permitted. The approach of
augmenting the solution space is far more flexible as compared to approach 1: no changes of original
quantities 𝑦, 𝐴 are necessary. Even curious constraints are allowed as long as datum deficiency is
resolved. However, we are faced with the constrained Lagrangian

1
L𝐷 (𝑥, 𝜆) = 𝑒 T𝑒 + 𝜆(𝐷 T𝑥 − 𝑐)
2
1 1
= 𝑦 T𝑦 − 𝑦 T𝐴𝑥 + 𝑥 T𝐴T𝐴𝑥 + 𝜆(𝐷 T𝑥 − 𝑐)
2 2
𝜕L𝐷
= −𝐴T𝑦 + 𝐴T𝐴𝑥 + 𝐷𝜆 = 0
𝜕𝑥
𝜕L𝐷
= 𝐷 T𝑥 − 𝑐 = 0
𝜕𝜆

     T 
𝐴T𝐴 𝐷 𝑥ˆ 𝐴𝑦
=⇒ ˆ = =⇒ 𝑀 𝑧ˆ = 𝑣
𝐷T 0 𝜆 𝑐
(𝑛+𝑑 )×(𝑛+𝑑 ) (𝑛+𝑑 )×1

25
3 Generalizations

E.g.

© −1 1 0ª © −1 1 0 ª © −1 1 0 ª © 2 1 −1 ª
𝐴 = ­­ −1 0 1 ®® =⇒ 𝐴T𝐴 = ­­ −1 0 1 ®® ­­ −1 0 1 ®® = ­­ 1 2 −1 ®®
« 0 1 −1 ¬ « 0 1 −1 ¬ « 0 1 −1 ¬ « −1 −1 2 ¬

© 2 1 −1 1ª
­ 1 2 −1 0 ®®
𝑀 = ­­
­ −1 −1 2 0 ®®
« 1 0 0 0¬
 
© 1 2 −1 ª
­ ®
det 𝑀 = −1 · det ­ −1 −1 2 ® = −1 · 1 · det
2 −1
= −3
−1 2
« 1 0 0¬
=⇒ 𝑀 regular =⇒ 𝑧ˆ = 𝑀 −1𝑣
  
𝑥ˆ = 𝑁 −1 𝐴T𝑦 + 𝐷𝑐 − 𝐷 (𝐷 T 𝑁 −1 𝐷) −1 𝐷 T 𝑁 −1𝐴T𝑦 + (𝐷 T 𝑁 −1 𝐷 − 𝐼 )𝑐
𝑁 := 𝐴T𝐴 + 𝐷𝐷 T

3.3 Linearization of non-linear observation equations

General 1-D-formulation
The functional model
𝑦 = 𝑓 (𝑥) ,

expressed by Taylor’s theorem, becomes

Õ∞
𝑓 (𝑛) (𝑥 0 )
𝑓 (𝑥) = (𝑥 − 𝑥 0 )𝑛
𝑛=0
𝑛!
d𝑓 1 d2 𝑓
= 𝑓 (𝑥 0 ) + (𝑥 − 𝑥 0 ) + (𝑥 − 𝑥 0 ) 2 + . . .
d𝑥 2 d𝑥 2
| {z }
𝑥0 𝑥0

negligible if 𝑥 − 𝑥 0 small

Substracting 𝑓 (𝑥 0 ) yields

d𝑓
𝑓 (𝑥) − 𝑓 (𝑥 0 ) = 𝑦 − 𝑦0 = (𝑥 − 𝑥 0 ) + . . .
d𝑥 𝑥0

d𝑓
(Δ𝑥) + O (Δ𝑥 2 )
| {z }
Δ𝑦 =
d𝑥 0
| {z } terms of higher order
linear model = model errors

with Δ𝑥 := 𝑥 − 𝑥 0 and Δ𝑦 := 𝑦 − 𝑦0 .

26
 3.3 Linearization of non-linear observation equations

General multi-D formulation

𝑦𝑖 = 𝑓𝑖 (𝑥 𝑗 ), 𝑖 = 1, . . . , 𝑚; 𝑗 = 1, . . . , 𝑛
𝑥 𝑗,0 −→ 𝑦𝑖,0 = 𝑓𝑖 (𝑥 𝑗,0 )

𝜕𝑓1 𝜕𝑓1 𝜕𝑓1

Δ𝑦1 = Δ𝑥 1 + Δ𝑥 2 + · · · + Δ𝑥𝑛
𝜕𝑥 1 0 𝜕𝑥 2 0 𝜕𝑥𝑛 0
𝜕𝑓2 𝜕𝑓2 𝜕𝑓2
Δ𝑦2 = Δ𝑥 1 + Δ𝑥 2 + · · · + Δ𝑥𝑛
𝜕𝑥 1 0 𝜕𝑥 2 0 𝜕𝑥𝑛 0
..
.
𝜕𝑓𝑚 𝜕𝑓𝑚 𝜕𝑓𝑚
Δ𝑦𝑚 = Δ𝑥 1 + Δ𝑥 2 + · · · + Δ𝑥𝑛 .
𝜕𝑥 1 0 𝜕𝑥 2 0 𝜕𝑥𝑛 0

Terms of second order and higher have been neglected.

© ª © 𝜕𝑥 1 ª ©­ Δ𝑥 2 ª®
Δ𝑦1 𝜕𝑓1 𝜕𝑓1 𝜕𝑓1 Δ𝑥 1
···
­ Δ𝑦2 ® ­ 𝜕𝑥 2
.. ® ­
𝜕𝑥𝑛
=⇒ ­­ . ®® = ­ ... ..
. ®® ­ .. ®®
­ .. ® ­ 𝜕𝑓𝑚
. ∼ Δ𝑦 = 𝐴(𝑥 0 ) Δ𝑥
­ . ®
« 𝜕𝑥 1
𝜕𝑓𝑚 𝜕𝑓𝑚
𝜕𝑥 2 · · · 𝜕𝑥𝑛 ¬ 0 Δ𝑥𝑛
« Δ𝑦𝑚 ¬ | {z }« ¬
Jacobian matrix 𝐴

Planar distance observation:

q
?
𝑠𝑖 𝑗 = (𝑥 𝑗 − 𝑥𝑖 ) 2 + (𝑦 𝑗 − 𝑦𝑖 ) 2 −→ 𝑦 = 𝐴𝑥

answer: linearize, Taylor series expansion

Linearization of planar distance observation equation (given Taylor point of expansion is 𝑥𝑖0 , 𝑦𝑖0 , 𝑥 0𝑗 ,
𝑦 0𝑗 = approximate values of unknown point coordinates); explicit differentiation
q q
𝑠𝑖 𝑗 = (𝑥 𝑗 − 𝑥𝑖 ) 2 + (𝑦 𝑗 − 𝑦𝑖 ) 2 = 𝑥𝑖2𝑗 + 𝑦𝑖2𝑗
“measured”

𝑥𝑖 = 𝑥𝑖0 + Δ𝑥𝑖 , 𝑦𝑖 = 𝑦𝑖0 + Δ𝑦𝑖 ,

𝑥 𝑗 = 𝑥 0𝑗 + Δ𝑥 𝑗 , 𝑦 𝑗 = 𝑦 0𝑗 + Δ𝑦 𝑗
r

2 
2
𝑠𝑖 𝑗 = 𝑥 0𝑗 + Δ𝑥 𝑗 − 𝑥𝑖0 + Δ𝑥𝑖 + 𝑦 0𝑗 + Δ𝑦 𝑗 − 𝑦𝑖0 + Δ𝑦𝑖
r
 2   2 𝜕𝑠 𝜕𝑠𝑖 𝑗 𝜕𝑠𝑖 𝑗 𝜕𝑠𝑖 𝑗
𝑖𝑗
= 𝑥 0𝑗 − 𝑥𝑖0 + 𝑦 0𝑗 − 𝑦𝑖0 + Δ𝑥𝑖 + Δ𝑥 𝑗 + Δ𝑦𝑖 + Δ𝑦 𝑗
| {z }
𝜕𝑥𝑖 0 𝜕𝑥 𝑗 0 𝜕𝑦𝑖 0 𝜕𝑦 𝑗 0

= 𝑠𝑖0𝑗 (distance from

approximate coordinates)

27
3 Generalizations

𝜕𝑠𝑖 𝑗 𝜕𝑠𝑖 𝑗 𝜕𝑥𝑖 𝑗 1 𝑥 𝑗 − 𝑥𝑖

= q
1
= 2𝑥𝑖 𝑗 (−1) = −
𝜕𝑥𝑖 𝜕𝑥𝑖 𝑗 𝜕𝑥𝑖 2 𝑥 2 + 𝑦2 𝑠𝑖 𝑗
𝑖𝑗 𝑖𝑗

𝜕𝑠𝑖 𝑗 𝑥 𝑗 − 𝑥𝑖 𝜕𝑠𝑖 𝑗 𝑦 𝑗 − 𝑦𝑖 𝜕𝑠𝑖 𝑗 𝑦 𝑗 − 𝑦𝑖

=+ , =− , =+
𝜕𝑥 𝑗 𝑠𝑖 𝑗 𝜕𝑦𝑖 𝑠𝑖 𝑗 𝜕𝑦 𝑗 𝑠𝑖 𝑗

 © 𝑖 ª
Δ𝑥
­ Δ𝑦𝑖 ®
𝑥 0𝑗 −𝑥𝑖0 𝑦 0𝑗 −𝑦𝑖0 𝑥 0𝑗 −𝑥𝑖0 𝑦 0𝑗 −𝑦𝑖0
­ ®
=⇒ Δ𝑠𝑖 𝑗 := 𝑠𝑖 𝑗 − 𝑠𝑖0𝑗 = − − ­ Δ𝑥 𝑗 ®
| {z } ­ ®
𝑠𝑖0𝑗 𝑠𝑖0𝑗 𝑠𝑖0𝑗 𝑠𝑖0𝑗

“reduced observation” « Δ𝑦 𝑗 ¬
Δ𝑦 = 𝐴(𝑥 0 ) Δ𝑥

Sometimes it is more convenient to use implicit differentiation within the linearization of observa-
tion equations.

2
2
Depart from 𝑠𝑖2𝑗 = 𝑥 𝑗 − 𝑥𝑖 + 𝑦 𝑗 − 𝑦𝑖 instead from 𝑠𝑖 𝑗 and calculate the total differential:





2/𝑠𝑖 𝑗 d𝑠𝑖 𝑗 = 2/ 𝑥 𝑗 − 𝑥𝑖 d𝑥 𝑗 − d𝑥𝑖 + 2/ 𝑦 𝑗 − 𝑦𝑖 d𝑦 𝑗 − d𝑦𝑖

Solve for d𝑠𝑖 𝑗 , introduce approximate value and switch from d −→ Δ:

𝑥 0𝑗 − 𝑥𝑖0
𝑦 0𝑗 − 𝑦𝑖0

Δ𝑠𝑖 𝑗 := 𝑠𝑖 𝑗 − 𝑠𝑖0𝑗 = Δ𝑥 𝑗 − Δ𝑥𝑖 + Δ𝑦 𝑗 − Δ𝑦𝑖
𝑠𝑖0𝑗 𝑠𝑖0𝑗

Grid bearings:
𝑥 𝑗 − 𝑥𝑖
𝑇𝑖 𝑗 = arctan
𝑦 𝑗 − 𝑦𝑖

=⇒ Linearized grid bearing observation equation:

!
1 1 𝑥 0𝑗 − 𝑥𝑖0 1 𝑥 0𝑗 − 𝑥𝑖0
𝑇𝑖 𝑗 = 𝑇𝑖0𝑗 +  2 − 0 Δ𝑥𝑖 + 0 Δ𝑦𝑖 + 0 Δ𝑥 𝑗 − 0 Δ𝑦 𝑗
𝑥 0𝑗 −𝑥𝑖0 𝑦 𝑗 − 𝑦𝑖0 (𝑦 𝑗 − 𝑦𝑖0 ) 2 𝑦 𝑗 − 𝑦𝑖0 (𝑦 𝑗 − 𝑦𝑖0 ) 2
1+ 𝑦 0𝑗 −𝑦𝑖0
!
(𝑦 0𝑗 − 𝑦𝑖0 ) 2 1 𝑥 0𝑗 − 𝑥𝑖0 1 𝑥 0𝑗 − 𝑥𝑖0
= 𝑇𝑖0𝑗 + − 0 Δ𝑥 𝑖 + Δ𝑦𝑖 + Δ𝑥 𝑗 − Δ𝑦 𝑗
(𝑠𝑖0𝑗 ) 2 𝑦 𝑗 − 𝑦𝑖0 (𝑦 0𝑗 − 𝑦𝑖0 ) 2 𝑦 0𝑗 − 𝑦𝑖0 (𝑦 0𝑗 − 𝑦𝑖0 ) 2
𝑦 0𝑗 − 𝑦𝑖0 𝑥 0𝑗 − 𝑥𝑖0 𝑦 0𝑗 − 𝑦𝑖0 𝑥 0𝑗 − 𝑥𝑖0
= 𝑇𝑖0𝑗 − Δ𝑥𝑖 + Δ𝑦𝑖 + Δ𝑥 𝑗 − Δ𝑦 𝑗
(𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2

Directions:

𝑟𝑖 𝑗 = 𝑇𝑖 𝑗 − 𝜔𝑖 (𝜔𝑖 additional unknown)

=⇒ linearization of bearing observation equation (see also Fig. 3.3)

28
 3.3 Linearization of non-linear observation equations

Figure 3.3: Linearization of bearing observation equation, bearing 𝑟𝑖 𝑗 , orientation unknown 𝜔𝑖 .

𝑟𝑖 𝑗 = 𝑇𝑖 𝑗 − 𝜔𝑖
𝑥 𝑗 − 𝑥𝑖
= arctan − 𝜔𝑖
𝑦 𝑗 − 𝑦𝑖
𝑦 0𝑗 − 𝑦𝑖0 𝑥 0𝑗 − 𝑥𝑖0 𝑦 0𝑗 − 𝑦𝑖0 𝑥 0𝑗 − 𝑥𝑖0
= 𝑟𝑖0𝑗 − Δ𝑥𝑖 + Δ𝑦𝑖 + Δ𝑥 𝑗 − Δ𝑦 𝑗 − 𝜔𝑖
(𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2

Angles:

𝛼𝑖 𝑗𝑘 = 𝑇𝑖𝑘 − 𝑇𝑖 𝑗
𝑥𝑘 − 𝑥𝑖 𝑥 𝑗 − 𝑥𝑖
= arctan − arctan
𝑦𝑘 − 𝑦𝑖 𝑦 𝑗 − 𝑦𝑖

=⇒ Linearized angle observation equation:

! !
𝑦𝑘0 − 𝑦𝑖0 𝑦 0𝑗 − 𝑦𝑖0 𝑥𝑘0 − 𝑥𝑖0 𝑥 0𝑗 − 𝑥𝑖0
𝛼𝑖 𝑗𝑘 = 𝑇𝑖𝑘0 − 𝑇𝑖0𝑗 + − 0 )2
+ Δ𝑥𝑖 + − Δ𝑦𝑖
(𝑠𝑖𝑘 (𝑠𝑖0𝑗 ) 2 0 )2
(𝑠𝑖𝑘 (𝑠𝑖0𝑗 ) 2
𝑦𝑘0 − 𝑦𝑖0 𝑥𝑘0 − 𝑥𝑖0 𝑦 0𝑗 − 𝑦𝑖0 𝑥 0𝑗 − 𝑥𝑖0
+ 0 )2
Δ𝑥𝑘 − 0 )2
Δ𝑦𝑘 − Δ𝑥 𝑗 + Δ𝑦 𝑗
(𝑠𝑖𝑘 (𝑠𝑖𝑘 (𝑠𝑖0𝑗 ) 2 (𝑠𝑖0𝑗 ) 2
= 𝛼𝑖0𝑗𝑘 + . . .

3D intersection with additional vertical angles

3D distances:
q
𝑑𝑖 𝑗 = (𝑥 𝑗 − 𝑥𝑖 ) 2 + (𝑦 𝑗 − 𝑦𝑖 ) 2 + (𝑧 𝑗 − 𝑧𝑖 ) 2 (𝑖 = 1, . . . , 4; 𝑗 ≡ 𝑃)

. . . linearization as usual.

29
3 Generalizations

Vertical angles:
p
(𝑥 𝑗 − 𝑥𝑖 ) 2 + (𝑦 𝑗 − 𝑦𝑖 ) 2
𝛽𝑖 𝑗 = arccot other trigonometric relations applicable
𝑧 𝑗 − 𝑧𝑖
𝑠𝑖 𝑗
= arccot
𝑧 𝑗 − 𝑧𝑖
1
= 𝛽𝑖0𝑗 −  2 · . . . Δ𝑥𝑖 + . . . Δ𝑦𝑖 + ... + . . . Δ𝑧 𝑗
𝑠𝑖 𝑗
1+ 𝑧 𝑗 −𝑧𝑖

Attention: physical units!

d4 d1
P4 4 1 P1
d3

3
d2
2
P3 P2

Figure 3.4: 3D intersection and vertical angles.

Iteration (see fig. 3.5)

Linearization (see 3.3) of the functional model 𝑦 = 𝑓 (𝑥) yields the linear model:

d𝑓
Δ𝑦 = Δ𝑥 + 𝑒 = 𝐴(𝑥 0 ) Δ𝑥 + 𝑒 .
d𝑥 𝑥0

The datum problem again

• Matrix 𝐴 is rank deficient (rank 𝐴 < 𝑛),
𝑚×𝑛

• 𝐴 has linear dependent columns,

• 𝐴𝑥 = 0 has non-trivial solution 𝑥 hom ≠ 0, i.e. the null space N (𝐴) of 𝐴 is not empty,
• det(𝐴T𝐴) = 0,
• 𝐴T𝐴 has zero eigenvalues.

30
 3.3 Linearization of non-linear observation equations

Nonlinear observation equations y = f(x)

Initial approximate values

x0

Dy(x0) = A(x0) Dx + e, Dy(x0) := y - f(x0) linear model

Redefined Additional constraints, e.g.

approximate x0 := x^ datum constraints
values
^ T -1 T estimated unknown
Dx = [ A (x0) A(x0) ] A (x0) Dy(x0) parameters

^ updated approximate values

x^ = x0 + Dx adjusted original parameters

No ! ^
PDxP2 < e Stop criteria

Yes!
^
y^ = f(x0)+ADx adjusted (estimated) observations

Error in iteration
process ! ^
e = y - y^ estimated residuals (inconsistencies)

No !
ATe^ = 0 orthogonality check satisfied ?

Yes!
No ! main check: nonlinear observation
y^ - f(x)
^
=0 equations satisfied by adjusted
observations ?

Yes!
Error in iteration
process or erroneous
linearization !

Figure 3.5: Iterative scheme

31
3 Generalizations

Example: planar distance network (fig. 3.6)

Rank defect:
• Translation −→ 2 parameters (𝑥-, 𝑦-direction),
• Rotation −→ 1 parameter,
=⇒ total of 𝑑 = 3 parameters,
=⇒ rank 𝐴 = 𝑛 − 𝑑 = 𝑛 − 3,
9 points −→ 𝑛 − 𝑑 = 18 − 3 = 15, 𝑚 = 19, thus 𝑟 = 4.
Conditional adjustment: How many conditions? Answer: 𝑟 condition equations.

3.4 Higher dimensions: the 𝑩-model (Condition equations)

In the ideal case we had

ℎ 1B − ℎ 1A = (𝐻 B − 𝐻 1 ) − (𝐻 A − 𝐻 1 ) = 𝐻 B − 𝐻 A
ℎ 13 + ℎ 32 − ℎ 12 = (𝐻 3 − 𝐻 1 ) + (𝐻 2 − 𝐻 3 ) − (𝐻 2 − 𝐻 1 ) = 0

or    
© ª © ª
1 0 0 0 −1 ℎ 1B 1 0 0 0 −1 𝐻 B
0 1 −1 1 0 ­­ ℎ 13 ®® 0 1 −1 1 0 ­­ 0 ®®
­ ℎ 12 ® = ­ 0 ®.
­ ® ­ ®
­ℎ ® ­ 0 ®
­ 32 ® ­ ®
« 1A ¬
ℎ « A¬
𝐻
Due to erroneous observations, a vector 𝑒 of unknown inconsistencies must be introduced in order
to make our linear model consistent.
   
© ª © ª
1 0 0 0 −1 ℎ 1B − 𝑒 1B 1 0 0 0 −1 𝐻 B
0 1 −1 1 0 ­­ ℎ 13 − 𝑒 13 ®® 0 1 −1 1 0 ­­ 0 ®®
­ ℎ 12 − 𝑒 12 ® = ­ 0 ®
­ ® ­ ®
­ ℎ −𝑒 ® ­ 0 ®
­ 32 32 ® ­ ®
« ℎ 1A − 𝑒 1A ¬ « 𝐻 A ¬.
or  
T
𝐵 Δℎ − 𝑒 = 𝐵 T𝑐 .
2×5 5×1 5×1 2×1

Connected with this example are the questions

Q 1: How to handle constants like the vector 𝑐?
Q 2: How many conditions must be set up?
Q 3: Is the solution of the 𝐵-model identical to the one of the 𝐴-model?

A 1: Starting from
𝐵 T (Δℎ − 𝑒) = 𝐵 T𝑐 ,

32
 3.4 Higher dimensions: the 𝐵-model (Condition equations)

(a) distance network

(b) Four lines may be deleted without destabilizing the net.

Figure 3.6

33
3 Generalizations

where solely 𝑒 is unknown, we collect all unknown parts on the left and all known quantities
on the right hand side

=⇒ 𝐵 T Δℎ − 𝐵 T𝑒 = 𝐵 T𝑐
𝐵 T𝑒 = 𝐵 T Δℎ − 𝐵 T𝑐
𝐵 T 𝑒 = 𝐵 T𝑦 =: 𝑤
𝑟 ×𝑚 𝑚×1 𝑟 ×1

𝑤 : vector of misclosures 𝑤 := 𝐵 T𝑦
𝑦 : reduced vector of observations
𝑟 : number of conditions

A 2: The number of conditions equals the redundancy

𝑟 =𝑚 −𝑛

Sometimes the number of conditions can hardly be determined without knowledge on the
number 𝑛 of unknowns in the 𝐴-model. This will be treated later in more detail together with
the so-called datum problem.
A 3:

1 T
L𝐵 (𝑒, 𝜆) = 𝑒 𝑒 + 𝜆 T ( 𝐵 T 𝑦 − 𝐵 T 𝑒 ) −→ min
2 1×𝑚 𝑚×1 1×𝑟 𝑟 ×𝑚 𝑚×1 𝑟 ×𝑚 𝑚×1
| {z }
𝑒,𝜆

1×1
𝜕L𝐵 ˆ = 𝑒ˆ − 𝐵 𝜆ˆ = 0
(𝑒,
ˆ 𝜆)
𝜕𝑒 𝑚×1 𝑚×𝑟 𝑟 ×1 𝑚×1

𝜕L𝐵 ˆ = − 𝐵 T 𝑒ˆ + 𝐵 T 𝑦 = 0
(𝑒,
ˆ 𝜆) (𝑤 = 𝐵 T𝑦)
𝜕𝜆
 
𝑟 ×𝑚 𝑚×1 𝑟 ×𝑚 𝑚×1 𝑟 ×1

𝑒ˆ
© ª © ª
𝐼 −𝐵 0
=⇒ ­ 𝑚×𝑚T 𝑚×𝑟 ® 𝜆ˆ = ­ 𝑚×1 ®
−𝐵 0 −𝑤
« 𝑟 ×𝑚 𝑟 ×𝑟 ¬ « 𝑚×1 ¬
𝑒ˆ = 𝐵 𝜆ˆ =⇒ 𝐵 T 𝐵 𝜆ˆ = 𝑤
=⇒ 𝜆ˆ = (𝐵 T 𝐵) −1𝑤 rank(𝐵 T 𝐵) = 𝑟
=⇒ 𝑒ˆ = 𝐵(𝐵 T 𝐵) −1𝑤
= 𝐵(𝐵 T 𝐵) −1 𝐵 T𝑦
= 𝑃𝐵𝑦
𝑦ˆ = 𝑦 − 𝑒ˆ
 
= 𝐼 − 𝐵(𝐵 T 𝐵) −1 𝐵 T 𝑦
= 𝑃𝐵⊥𝑦

34
 3.4 Higher dimensions: the 𝐵-model (Condition equations)

For the transition

parametric model ←→ model of condition equations

𝑦 = 𝐴𝑥 + 𝑒 ←→ 𝐵 T𝑒 = 𝐵 T𝑦 ,

left multiply 𝑦 = 𝐴𝑥 + 𝑒 by 𝐵 T :

𝐵 T𝑦 = 𝐵 T𝐴𝑥 + 𝐵 T𝑒 ⇐⇒ 𝐵 T𝐴 = 0 .

E.g.:
© ª
−1 0 0
  ­ −1 0 1 ®®  
1 0 0 0 −1 ­­ ® 000
0 1 −1 1 0 ­­
−1 1 0®= .
® 000
2×5 ­ 0 1 −1 ® 2×3

« −1 0 0¬
5×3

35
3 Generalizations

3.5 Linearization of non-linear condition equations

Figure 3.7: Linearization of condition equations

Ideal situation: error free observations

𝑎 𝑏
= ∼ 𝑎 sin 𝛽 − 𝑏 sin 𝛼 = 0
sin 𝛼 sin 𝛽

Real situation with “errors” 𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽

(𝑎 − 𝑒𝑎 ) sin(𝛽 − 𝑒 𝛽 ) − (𝑏 − 𝑒𝑏 ) sin(𝛼 − 𝑒𝛼 ) = 0

𝜕𝑓 𝜕𝑓
𝑓 (𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽 ) = 𝑓 (𝑒𝑎0 , 𝑒𝑏0 , 𝑒𝛼0 , 𝑒 𝛽0 ) + (𝑒𝑎 − 𝑒𝑎0 ) + . . . + (𝑒 𝛽 − 𝑒 𝛽0 )
𝜕𝑒𝑎 0 𝜕𝑒 𝛽 0
𝜕𝑓 𝜕𝑓 𝜕𝑓 𝜕𝑓
= 𝑓 (𝑒𝑎0 , 𝑒𝑏0 , 𝑒𝛼0 , 𝑒 𝛽0 ) + 𝑒𝑎 + . . . + 𝑒𝛽 − 𝑒𝑎0 − . . . − 𝑒0
𝜕𝑒𝑎 0 𝜕𝑒 𝛽 0 𝜕𝑒𝑎 0 𝜕𝑒 𝛽 0 𝛽

Model adjustment condition equations

𝑤 − 𝐵 T𝑒 = 0

36
 3.6 Higher dimensions: the mixed model

3.6 Higher dimensions: the mixed model (Gauss-Helmert model)

In the 𝐴-model, every observation is – in general – a linear or non-linear function of all unknown
quantities, i.e.

𝑦𝑖 = 𝑓𝑖 (𝑥 1, 𝑥 2, . . . , 𝑥𝑛 ) = 𝑓𝑖 (𝑥 𝑗 ) = 𝑓𝑖 (𝑥), 𝑖 = 1, . . . , 𝑚; 𝑗 = 1, . . . , 𝑛

and every observation equation 𝑦𝑖 contains just one single inconsistency 𝑒𝑖 . In contrast, in the 𝐵-
model no unknown parameter 𝑥 exist and we have linear or non-linear relationships between the
observations only,
𝑓 𝑗 (𝑦𝑖 ) = 𝑓 𝑗 (𝑦) = 0, 𝑖 = 1, . . . , 𝑚; 𝑗 = 1, . . . , 𝑟 .
However, in many applications, functional relationships exist between both, parameters 𝑥 and ob-
servations 𝑦, which can be formulated only as an implicit function

𝑓 (𝑥 𝑗 , 𝑦𝑖 ) = 0, 𝑖 = 1, . . . , 𝑚; 𝑗 = 1, . . . , 𝑛 .

This will lead to a combination of both, 𝐴- and 𝐵-model, which is known as the general model of
adjustment, mixed model or Gauss-Helmert model, in honor of Friedrich Robert Helmert1 .

Example: Best fitting circle with unknown radius 𝑟 , and unknown centre coordinates 𝑢 M , 𝑣 M ;
observations 𝑢𝑖 and 𝑣𝑖 inconsistent.

2
2
𝑓 (𝑟, 𝑢 M, 𝑦M, 𝑢𝑖 − 𝑒𝑢𝑖 , 𝑣𝑖 − 𝑒 𝑣𝑖 ) = 𝑢𝑖 − 𝑒𝑢𝑖 − 𝑢 M + 𝑣𝑖 − 𝑒 𝑣𝑖 − 𝑣 M − 𝑟2 = 0
| {z } | {z }
unknown observations “𝑦” −
parameters inconsistencies “𝑒”
“𝑥”

1 Friedrich Robert Helmert (1843–1917) was a famous German geodesist and mathematician who introduced this model

1872 in his book “Die Ausgleichungsrechnung nach der Methode der kleinsten Quadrate” (Adjustment theory using
the method of least squares). He is also known as the father of many mathematical and physical theories of modern
geodesy.

37
4 Weighted least squares

Observations have different weights ⇐⇒ different quality.

4.1 Weighted observation equations

Analytical interpretation
Target function:

𝑦1 −→ 𝑤 1 1 
L𝑎𝑤 = 𝑤 1 (𝑦1 − 𝑎𝑥) 2 + 𝑤 2 (𝑦2 − 𝑎𝑥) 2
𝑦2 −→ 𝑤 2 2
 
1 T 𝑤1 0
= (𝑦 − 𝑎𝑥) (𝑦 − 𝑎𝑥)
2 0 𝑤2
1
= (𝑦 − 𝑎𝑥) T𝑊 (𝑦 − 𝑎𝑥)
2
1 1
= 𝑦 T𝑊 𝑦 − 𝑦 T𝑊 𝑎𝑥 + 𝑥 T𝑎 T𝑊 𝑎𝑥
2 2
1 T
= 𝑒 𝑊𝑒
2

Necessary condition:

𝜕L𝑎
𝑥ˆ : min L𝑎 (𝑥) =⇒ (𝑥)
ˆ
𝑥 𝜕𝑥
𝜕L
=⇒ = −𝑦 T𝑊 𝑎 + 𝑎 T𝑊 𝑎𝑥 = −𝑎 T𝑊 𝑦 + 𝑎 T𝑊 𝑎𝑥ˆ = 0
𝜕𝑥
=⇒ 𝑎 T𝑊 𝑎𝑥ˆ = 𝑎 T𝑊 𝑦 normal equation

Sufficient condition:
𝜕2 L
= 𝑎 T𝑊 𝑎 > 0, since 𝑊 is positive definite
𝜕𝑥 2

Normal equation =⇒

𝑎 T𝑊 (𝑦 − 𝑎𝑥)
ˆ =0 =⇒ 𝑎 T𝑊 𝑒ˆ = 0
= 𝑒ˆ ⊥ 𝑊 𝑎

38
 4.1 Weighted observation equations

normal equations 𝑎 T𝑊 𝑎𝑥ˆ = 𝑎 T𝑊 𝑦

WLS estimate of 𝑥
𝑥ˆ = (𝑎 T𝑊 𝑎) −1𝑎 T𝑊 𝑦
(weighted least squares)
WLS estimate of 𝑦 𝑦ˆ = 𝑎𝑥ˆ = 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 𝑦
 
WLS estimate of 𝑒 𝑒ˆ = 𝑦 − 𝑦ˆ = 𝐼 − 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 𝑦

Example
 



  𝑤 0  
1
𝑎= 

  
1 ! T 1

𝑎𝑊 = 11 = 𝑤1 𝑤2
𝑤1 0 

0 𝑤2
𝑊 = 

0 𝑤2 

 1

T
𝑎 𝑊 𝑎 = 𝑤1 𝑤2 = 𝑤1 + 𝑤2
1

=⇒

1
𝑥ˆ = (𝑤 1𝑦1 + 𝑤 2𝑦2 ) (weighted mean)
𝑤1 + 𝑤2
𝑤1 𝑤2
= 𝑦1 + 𝑦2
𝑤1 + 𝑤2 𝑤1 + 𝑤2

     
𝑒ˆ1 𝑦 𝑦ˆ
= 1 − 1
𝑒ˆ2 𝑦2 𝑦ˆ2
 
1 (𝑤 1 + 𝑤 2 )𝑦1 − 𝑤 1𝑦1 − 𝑤 2𝑦2
=
𝑤 1 + 𝑤 2 (𝑤 1 + 𝑤 2 )𝑦2 − 𝑤 1𝑦1 − 𝑤 2𝑦2
 
1 𝑤 2 (𝑦1 − 𝑦2 )
=
𝑤 1 + 𝑤 2 𝑤 1 (𝑦2 − 𝑦1 )

𝑤1 > 𝑤2 : “𝑦1 is more important than 𝑦2 ” =⇒ |𝑒ˆ1 | < |𝑒ˆ2 |

Projectors

𝑃𝑎 = 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 := 𝑃𝑎,(𝑊 𝑎) ⊥

𝑃𝑎 𝑃𝑎 = 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊
= 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 = 𝑃𝑎

𝑃𝑎 idempotent matrix, oblique projector.

39
4 Weighted least squares

(a) Circle (b) Ellipse aligned with coordi- (c) General ellipse, not aligned
nate axes. with coordinate axes.

Figure 4.1

4.1.1 Geometry

𝐹 (𝑧) = 𝑧 T𝑊 𝑧 = 𝑐
𝑤 1𝑧 12 + 𝑤 2𝑧 22 = 𝑐
𝑤1 2 𝑤2 2
𝑧 + 𝑧 =1
𝑐 1 𝑐 2
𝑧 12 𝑧 22
𝑐 + 𝑐 =1
𝑤1 𝑤2

𝑧 12 𝑧 22
+ =1 ellipse equation
𝑎 𝑏

A family (𝑐 may vary!) of ellipses, the principal exes of which are not aligned with coordinate axes,
in general.

Principal axes not aligned with coordinate axes

𝑧 T𝑊 𝑧 = 𝑐 ∼ 𝑧 12𝑤 11 + 2𝑤 12𝑧 1𝑧 2 + 𝑤 22𝑧 22 = 𝑐

General ellipse

grad 𝐹 (𝑧 0 ) = 2𝑊 𝑧 0 vector in 𝑧 0 , orthogonal to the tangent of the ellipse in 𝑧 0

𝑧 − 𝑧0 ⊥ 𝑊 𝑧0 or 𝑧 0 T𝑊 (𝑧 − 𝑧 0 ) = 0

40
 4.2 Weighted condition equations

4.1.2 Application to adjustment problems

Find a vector starting on line 𝑎𝑥, ending in 𝑦 being parallel to 𝑧 − 𝑎 or orthogonal to 𝑎 T𝑊 : 𝑒ˆ

• 𝑦ˆ = 𝑎𝑥ˆ is the projection of 𝑦

– onto 𝑎
– in the direction orthogonal to 𝑊 𝑎 (along (𝑊 𝑎) ⊥ )

=⇒ 𝑦ˆ = 𝑃𝑎,(𝑊 𝑎) ⊥ 𝑦 with 𝑃𝑎,(𝑊 𝑎) ⊥ = 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊

• 𝑒ˆ is the projection of 𝑦
– onto (𝑊 𝑎) ⊥
– in direction of 𝑎

⊥
=⇒ 𝑒ˆ = 𝑃 (𝑊 𝑎) ⊥,𝑎𝑦 with 𝑃 (𝑊 𝑎) ⊥,𝑎 = 𝑃𝑎,(𝑊 𝑎) ⊥

= 𝐼 − 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊
 
= 𝐼 − 𝑎(𝑎 T𝑊 𝑎) −1𝑎 T𝑊 𝑦

• Because of 𝑒ˆ 6⊥ 𝑎 (or 𝑎 T𝑒ˆ ≠ 0) projections are oblique projections (or orthogonal projections
with respect to the metric 𝑊 ; 𝑒ˆ ⊥ 𝑊 𝑎 or 𝑎 T𝑊 𝑒ˆ = 0)

4.1.3 Higher dimensions

From one unknown to many unknowns.

𝑚=2
𝑦 =𝑎 𝑥 + 𝑒
2×1 2×1 1×1 2×1

becomes
𝑦 =𝐴 𝑥 + 𝑒
𝑚×1 𝑚×𝑛 𝑛×1 𝑚×1

Replace 𝑎 by 𝐴!

𝑃 (𝑊𝑎 ) ⊥,𝑎 = 𝐼 − 𝐴 (𝐴T 𝑊 𝐴) −1 𝐴T 𝑊

| {z }
𝑚×𝑛 𝑚×𝑚 𝑛×𝑚 𝑚×𝑚

𝑛×𝑛

4.2 Weighted condition equations

Geometry
Starting point again: 𝑏 T𝑎 = 0 (𝑎 ⊥ 𝑏):

41
4 Weighted least squares

Direction of (𝑊 𝑎) ⊥ :

𝑏 T𝑎 = 0 =⇒ 𝑏 T𝑊 −1𝑊 𝑎 = 0 =⇒ 𝑊 𝑎 ⊥ 𝑊 −1𝑏 =⇒ 𝑊 −1𝑏 = (𝑊 𝑎) ⊥

Figure 4.2: weighted condition

Target function to be minimized: 𝑒 T𝑊 𝑒 under 𝑏 T𝑒 = 𝑏 T𝑦 or 𝑏 T (𝑦 − 𝑒) = 0.

From all possible 𝑒’s find that 𝑒 = 𝑒ˆ which ends on the line 𝑏 T𝑒 = 𝑏 T𝑦 and generates the smallest
𝑒 T𝑊 𝑒 = 𝑐! =⇒ line 𝑏 T𝑦 = 𝑏 T𝑒 is tangent to 𝑒 T𝑊 𝑒 = 𝑒ˆT𝑊 𝑒ˆ = 𝑐 min .
Point of Tangency: normal of the ellipse = normal of the line 𝑏 T𝑦 = 𝑏 T𝑒 = direction of 𝑏 ⇐⇒ 𝑒ˆ
is parallel to 𝑊 −1𝑏 =⇒ 𝑒ˆ = 𝑊 −1𝑏𝛼, 𝛼 an unknown scalar.
Determine 𝛼: 𝑒ˆ lies on 𝑏 T𝑒 = 𝑏 T𝑦

=⇒ 𝑏 T𝑒ˆ = 𝑏 T𝑊 −1𝑏𝛼 = 𝑏 T𝑦
=⇒ 𝛼 = (𝑏 T𝑊 −1𝑏) −1𝑏 T𝑦
=⇒ 𝑒ˆ = 𝑊 −1𝑏 (𝑏 T𝑊 −1𝑏) −1𝑏 T𝑦
 
=⇒ 𝑦ˆ = 𝑦 − 𝑒ˆ = 𝐼 − 𝑊 −1𝑏 (𝑏 T𝑊 −1𝑏) −1𝑏 T 𝑦

Remark: 𝑒ˆ is not the smallest 𝑒, orthogonal to 𝑏 T𝑒 = 𝑏 T𝑦!

Calculus

1
L𝑏 (𝑒, 𝑦) = 𝑒 T𝑊 𝑒 + 𝜆 T (𝑏 T𝑦 − 𝑏 T𝑒) etc.
2
𝑒ˆ : min 𝑒 T𝑊 𝑒 under constraint 𝑏 T𝑒 = 𝑏 T𝑦
𝑒

42
 4.2 Weighted condition equations

Figure 4.3: possible ellipses

Lagrange:
1
L𝑏 (𝑒, 𝜆) = 𝑒 T𝑊 𝑒 + 𝜆 T (𝑏 T𝑦 − 𝑏 T𝑒)
2
Find 𝑒 and 𝜆 which minimize 𝐿𝑏 .
( 𝜕L
𝑏 ˆ = 𝑊 𝑒ˆ − 𝑏 𝜆ˆ = 0
𝜕𝑒 (𝑒,
ˆ 𝜆)
=⇒ 𝜕 L𝑏 ˆ = −𝑏 T𝑒 + 𝑏 T𝑦 = 0
𝜕𝜆 (𝑒,
ˆ 𝜆)
    
𝑊 −𝑏 𝑒ˆ 0
⇐⇒ =
−𝑏 T 0 𝜆ˆ −𝑏 T𝑦

1. row
𝑊 𝑒ˆ − 𝑏 𝜆ˆ = 0 =⇒ 𝑒ˆ = 𝑊 −1𝑏 𝜆ˆ
2. row
𝑏 T𝑒ˆ = 𝑏 T𝑦 =⇒ 𝑏 T𝑊 −1𝑏 𝜆ˆ = 𝑏 T𝑦
solve for 𝜆ˆ
𝜆ˆ = (𝑏 T𝑊 −1𝑏) −1𝑏 T𝑦
substitute in 1. row

𝑒ˆ = 𝑊 −1𝑏 (𝑏 T𝑊 −1𝑏) −1𝑏 T𝑦

 
𝑦ˆ = 𝑦 − 𝑒ˆ = 𝐼 − 𝑊 −1𝑏 (𝑏 T𝑊 −1𝑏) −1𝑏 T 𝑦

Higher dimensions
Replace 𝑏 with 𝐵.

43
4 Weighted least squares

𝑟 = 𝑚 − 𝑛 condition equations, Lagrange multipliers

𝐵 T𝑦 = 𝐵 T𝑒
)
𝑦 = 𝐴𝑥 + 𝑒
=⇒ 𝐵 T𝑦 = 𝐵 T𝐴𝑥 + 𝐵 T𝑒 = 𝐵 T𝑒
𝐵 T𝐴 = 0
!
© 𝑚×𝑚 𝑚×𝑟 ª © 𝑚×1 ª
𝑊 𝐵 𝑒ˆ 0
­ T ®­ ˆ ® =
𝐵 0 𝜆 𝐵 T𝑦
« 𝑟 ×𝑚 𝑟 ×𝑟 ¬ « 𝑟 ×1 ¬

𝑒ˆ = 𝑊 −1 𝐵(𝐵 T𝑊 −1 𝐵) −1 𝐵 T𝑦
 
𝑦ˆ = 𝐼 − 𝑊 −1 𝐵(𝐵 T𝑊 −1 𝐵) −1 𝐵 T 𝑦

Constant term (RHS)

Ideal case without errors:
𝐵 T𝑦 = 𝑐
In reality:

𝐵 T (𝑦 − 𝑒) = 𝑐 =⇒ 𝐵 T𝑒 = 𝐵 T𝑦 − 𝑐 =: 𝑤





𝑒ˆ = 𝑊 −1 𝐵(𝐵 T𝑊 −1 𝐵) −1 [𝐵 T𝑦 − 𝑐]
| {z }

=⇒

 𝑦ˆ = 𝑦 − 𝑒ˆ = . . .
𝑤


4.3 Stochastics

Probabilistic formulation
(stochastic quantities are underlined)

 
Version 1: 𝑦 = 𝐴𝑥 + 𝑒, E 𝑒 = 0 D 𝑒 = 𝑄𝑦
n o 
Version 2: E 𝑦 = 𝐴𝑥 D 𝑒 = 𝑄𝑦

| {z | } {z }
Stochastic model:
Functional model
variance-covariance matrix
| {z }
Mathematical model

Linear Variance-covariance propagation

In general:
𝑧 = 𝐿𝑦, 𝑄𝑧 = 𝐿𝑄 𝑦 𝐿 T

44
 4.4 Best Linear Unbiased Estimation (blue)

𝑥ˆ = (𝐴T𝑊 𝐴) −1𝐴T𝑊 𝑦

= 𝐿𝑦
 n o
−→ E 𝑥ˆ = (𝐴T𝑊 𝐴) −1𝐴T𝑊 E 𝑦

= (𝐴T𝑊 𝐴) −1𝐴T𝑊 𝐴𝑥
=𝑥 (unbiased estimate)
−→ 𝑄 𝑥ˆ = 𝐿𝑄 𝑦 𝐿 T
= (𝐴T𝑊 𝐴) −1𝐴T𝑊 𝑄 𝑦𝑊 𝐴(𝐴T𝑊 𝐴) −1
𝑦ˆ = 𝐴𝑥ˆ

= 𝑃𝐴𝑦
n o  n o
−→ E 𝑦ˆ = 𝐴 E 𝑥ˆ = 𝐴𝑥 = E 𝑦

−→ 𝑄 𝑦ˆ = 𝑃𝐴𝑄 𝑦 𝑃𝐴 T
𝑒ˆ = 𝑦 − 𝐴𝑥ˆ = (𝐼 − 𝑃𝐴 )𝑦
 n o
−→ E 𝑒ˆ = E 𝑦 −𝐴𝑥 = 0

−→ 𝑄𝑒ˆ = 𝑄 𝑦 − 𝑃𝐴𝑄 𝑦 − 𝑄 𝑦 𝑃𝐴 T + 𝑃𝐴𝑄 𝑦 𝑃𝐴 T

Questions:
• Is 𝑥ˆ the best estimator?
• Or: When is 𝑄 𝑥ˆ smallest?

4.4 Best Linear Unbiased Estimation (BLUE)

Best 𝑄 𝑥ˆ minimal (in LU-Class)

Linear 𝑥ˆ = 𝐿𝑦

Unbiased E 𝑥ˆ = 𝑥
Estimate

2D-example (old)

n o  
1
E 𝑦 = 𝑎𝑥, 𝑎=
1
n o  2 
𝜎1 𝜎12
D 𝑦 = 𝑄 𝑦, 𝑄𝑦 =
𝜎12 𝜎22

45
4 Weighted least squares

L-property:
𝑥ˆ = 𝑙 T𝑦
U-property:
 n o
E 𝑥ˆ = 𝑙 T E 𝑦 = 𝑙 T𝑎𝑥 = 𝑥 =⇒ 𝑙 T𝑎 = 1
B-property:
𝑥ˆ = 𝑙 T𝑦 =⇒ 𝜎𝑥2ˆ = 𝑙 T𝑄 𝑦 𝑙

Find that 𝑙 which minimizes 𝑙 T𝑄 𝑦 𝑙 and satisfies 𝑙 T𝑎 = 1!

=⇒ min 𝑙 T𝑄 𝑦 𝑙 under 𝑙 T𝑎 = 1
𝑙

Solution?

Comparison LS, B-Model

min 𝑒 T𝑊 𝑒 𝑙 T𝑄 𝑦 𝑙
under 𝑏 T𝑒 = 𝑏 T𝑦 = 𝑤 𝑙 T𝑎 = 𝑎 T𝑙 = 1
estimate 𝑒ˆ = 𝑊 −1𝑏 (𝑏 T𝑊 −1𝑏) −1𝑤 𝑙ˆ = 𝑄 𝑦−1𝑎(𝑎 T𝑄 𝑦−1𝑎) −1

=⇒ 𝑥ˆ = 𝑙ˆT𝑦 = (𝑎 T𝑄 𝑦−1𝑎) −1𝑎 T𝑄 𝑦−1𝑦

Higher dimensions
𝑎 −→ 𝐴, 𝑄 𝑦−1 = 𝑃 𝑦
Gauss coined the variable 𝑃 from the Latin pondus, which means weight.

)
BLUE: 𝑥ˆ = (𝐴T 𝑃 𝑦 𝐴) −1𝐴T 𝑃 𝑦𝑦
=⇒ BLUE, if 𝑊 = 𝑃 𝑦 = 𝑄 𝑦−1
Det.: 𝑥ˆ = (𝐴T𝑊 𝐴) −1𝐴T𝑊 𝑦

Linear Variance-covariance propagation

𝑥ˆ = (𝐴T 𝑃 𝑦 𝐴) −1𝐴T 𝑃 𝑦𝑦

=⇒ 𝑄 𝑥ˆ = (𝐴T 𝑃 𝑦 𝐴) −1𝐴T 𝑃 𝑦 𝑄 𝑦 𝑃 𝑦 𝐴(𝐴T 𝑃 𝑦 𝐴) −1 = (𝐴T 𝑃 𝑦 𝐴) −1

𝑦ˆ = 𝐴𝑥ˆ = 𝑃𝐴𝑦

=⇒ 𝑄 𝑦ˆ = 𝐴(𝐴T 𝑃 𝑦 𝐴) −1𝐴T 𝑃 𝑦 𝑄 𝑦 = 𝑃𝐴𝑄 𝑦 = 𝑃𝐴𝑄 𝑦 𝑃𝐴 T = 𝑄 𝑦 𝑃𝐴

𝑒ˆ = (𝐼 − 𝑃𝐴 )𝑦 = 𝑃𝐴⊥𝑦 = 𝑦 − 𝑦ˆ

=⇒ 𝑄𝑒ˆ = 𝑄 𝑦 − 𝑃𝐴𝑄 𝑦 − 𝑄 𝑦 𝑃𝐴 T + 𝑃𝐴𝑄 𝑦 𝑃𝐴 T = 𝑃𝐴⊥𝑄 𝑦 = 𝑄 𝑦 − 𝑄 𝑦ˆ

Besides:
𝐼 = 𝑃𝐴 + 𝑃𝐴⊥ =⇒ 𝑄 𝑦 = 𝑃𝐴𝑄 𝑦 + 𝑃𝐴⊥𝑄 𝑦 = 𝑄 𝑦ˆ + 𝑄𝑒ˆ
Note: 𝑃𝐴 is a projector, but 𝑃 𝑦 is a weight matrix.

46
5 Geomatics examples

Further (simple and more advanced) examples including data files can be found in "Geodetic Net-
work Adjustment Examples" (http://www.gis.uni-stuttgart.de/lehre/campus-docs/adjustment_examples.pdf).

5.1 A-Model: Adjustment of observation equations

5.1.1 Planar triangle

Observations: Angles 𝛼, 𝛽, 𝛾 and distances 𝑠 12 , 𝑠 13 , 𝑠 23

Auxiliary quantities: Bearings 𝑇12 , 𝑇13
Bearings:

𝑥 𝑗 − 𝑥𝑖
𝑇𝑖 𝑗 = arctan
𝑦 𝑗 − 𝑦𝑖

Angles:

𝑥2 − 𝑥1 𝑥3 − 𝑥1
𝛼 = 𝑇12 − 𝑇13 = arctan − arctan
𝑦2 − 𝑦1 𝑦3 − 𝑦1
𝑥3 − 𝑥2 𝑥1 − 𝑥2
𝛽 = 𝑇23 − 𝑇21 = arctan − arctan
𝑦3 − 𝑦2 𝑦1 − 𝑦2
𝑥1 − 𝑥3 𝑥2 − 𝑥3
𝛾 = 𝑇31 − 𝑇32 = arctan − arctan
𝑦1 − 𝑦3 𝑦2 − 𝑦3

Approximate coordinates: point 𝑥/𝑚 𝑦/𝑚

1 0 0
2 1 0
√
1 3
3 2 2

47
5 Geomatics examples

s13 s23

s12

Figure 5.1: Sketch Planar triangle

“Observations”
Approx. coordinates
from approx. Observations 𝜎
in m
coordinates
0
𝑠 12 1m 𝑠 12 1.01 m ±0.01 m
0
𝑠 13 1m 𝑠 13 1.02 m ±0.02 m
𝑥0 𝑦0 0
𝑠 23 1m 𝑠 23 0.97 m ±0.01 m
1 0 0 𝛼0 60◦ 𝛼 60◦ ±1 ′′
2 1 0 𝛽0 60◦ 𝛽 59.7◦ ±1 ′
√
1 3
3 2 2 𝛾0 60◦ 𝛾 60.2◦ ±1 ′

Observation 𝑦 Designmatrix 𝐴 Units Unknowns

180◦
𝜌 := 𝜋 d𝑥 1 d𝑦1 d𝑥 2 d𝑦2 d𝑥 3 d𝑦3 in m
𝑠 12 0.01 m −1 0 1 0 0 0 d𝑥 1
√ √
1 3 3
𝑠 23 −0.03 m 0 0 2 − 2 − 12 2 [−] d𝑦1
√ √
3 3
𝑠 13 0.02 m − 12 − 2 0 0 1
2 2 d𝑥 2
√ √
3 1 3 1
𝛼 0 rad 2 2 0 −1 − 2 2 d𝑦2
√ √
◦
𝛽 −0.3 𝜌 rad 0 −1 − 23 1
2 2
3 1
2 [m−1 ] d𝑥 3
√ √
◦ 3 1 3 1
𝛾 0.2 𝜌 rad − 2 2 2 2 0 −1 d𝑦3

=⇒ Linearized distance observation equation (Taylor point = point of expansion = set of approximate
coordinates).

48
 5.1 A-Model: Adjustment of observation equations

5.1.2 Distance Network

In this example, measured distances between the points of the network in figure 3.6 are adjusted.
The standard deviation of observations is 𝜎𝑠 = ±1 cm, the a priori standard deviation 𝜎0 = ±1 cm.

𝜎02
=⇒ 𝑃𝑠 = =1
𝜎𝑠2

Table 5.1 contains measured distances (observations 𝑦) between respective network points, while
table 5.2 contains approximate coordinates of the points. Points A and B are datum points with the
minimum number of datum parameters 𝑋 A , 𝑌A , 𝑋 B fixed.

leg length in m leg length in m Point ID 𝑋 0 /m 𝑌0 /m

A–B 1309.155 D–H 2179.147 A 184 270.031 725 830.033
A–C 1188.464 D– I 1461.074 B 185 549.974 725 400.000
A–G 1267.520 E– F 1031.232 C 183 200.000 725 450.000
B–G 1447.552 E– I 1353.146 D 183 800.000 723 550.000
B–H 1077.634 F –H 1991.004 E 184 300.000 722 050.000
C–D 1715.405 F– I 997.285 F 185 200.000 722 450.000
C–G 1504.039 G–H 1149.345 G 184 500.000 724 400.000
C– I 2688.088 G– I 1310.957 H 185 700.000 724 650.000
D–E 1780.446 H– I 1241.810 I 184 800.000 723 400.000
D–G 1260.133

Table 5.1: Observed distances 𝑦. Table 5.2: Approximate coordinates.

Table 5.3 contains the reduced vector Δ𝑦 and table 5.4 the estimated parameters at first iteration.

leg length in m leg length in m

A–B −41.098 D–H −16.303 c
Δ𝜉/m c
Δ𝜉/m
A–C 52.950 D– I 449.887 𝑋A 0.000 00 𝑌E 70.034 91
A–G −180.886 E– F 46.346 𝑌A 0.000 00 𝑋F 220.074 41
B–G −2.429 E– I −86.472 𝑋B 0.000 00 𝑌F 12.078 83
B–H 312.776 F –H −265.099 𝑌B 124.267 84 𝑋G −44.702 62
C–D −277.081 F– I −33.491 𝑋C −29.046 95 𝑌G 176.988 00
C–G −167.038 G–H −76.420 𝑌C −80.516 22 𝑋H −52.354 95
C– I 87.607 G– I 266.926 𝑋D −241.435 89 𝑌H −206.799 48
D–E 199.307 H– I −298.482 𝑌D 143.805 90 𝑋I 190.824 30
D–G 158.997 𝑋E 149.155 11 𝑌I −29.001 63

Table 5.3: Reduced observations Δ𝑦. Table 5.4: Estimated parameters.

Tables 5.5 and 5.6 contain the adjusted coordinates and observations (𝑦)
ˆ respectively, of the network
points after 6 iterations. Table 5.7 shows estimated inconsistencies in measured distances.

=⇒ 𝑒ˆT 𝑃 𝑒ˆ = 0.035 cm2, c k < 10−10 )

(6 iterations, stop criteria k Δ𝜉

49
5 Geomatics examples

leg ˆ
𝑦/m leg ˆ
𝑦/m
point ID 𝑋ˆ /m 𝑌ˆ /m A–B 1309.155 D–H 2179.1462
A 184 270.031 725 830.033 A–C 1188.464 D– I 1461.0749
B 185 549.974 725 555.019 A–G 1267.520 E– F 1031.2318
C 183 185.048 725 344.999 B–G 1447.552 E– I 1353.1463
D 183 598.001 723 680.041 B–H 1077.634 F –H 1991.0036
E 184 499.996 722 144.987 C–D 1715.4053 F– I 997.2854
F 185 469.997 722 495.040 C–G 1504.0395 G–H 1149.3454
G 184 480.021 724 580.029 C– I 2688.0873 G– I 1310.9572
H 185 625.005 724 480.000 D–E 1780.4458 H– I 1241.8109
I 185 030.002 723 390.016 D–G 1260.133

Table 5.5: Adjusted coordinates. Table 5.6: Adjusted observations.

leg ˆ
𝑒/mm leg ˆ
𝑒/mm
A–B −0.01 D–H 0.78
A–C −0.01 D– I −0.88
A–G 0.00 E– F 0.22
B–G 0.01 E– I −0.27
B–H −0.01 F –H 0.43
C–D −0.27 F– I −0.39
C–G −0.46 G–H −0.35
C– I 0.67 G– I −0.18
D–E 0.20 H– I −0.86
D–G −0.04

Table 5.7: Estimated inconsistencies.

50
 5.1 A-Model: Adjustment of observation equations

While figure 5.2 indicates the convergence of estimated corrections to approximate coordinates,
figure 5.3 depicts the overall convergence in adjustment iteration. Figure 5.4 represents approximate
points, adjusted and datum points. Finally in figure 5.5 adjusted and datum points are shown with
error ellipses. Table 5.8 displays the 𝐴-matrix after the first iteration.
Convergence of estimated corrections to approximate coordinates
250
1) ∆XA=0

200 2) ∆YA=0
3) ∆XB=0

150 4) ∆YB
5) ∆XC

100 6) ∆YC
7) ∆XD
Parameter value in m

50 8) ∆YD
9) ∆XE

0 10) ∆YE
11) ∆X F

−50 12) ∆YF

13) ∆XG

−100 14) ∆YG

15) ∆XH

−150 16) ∆YH

17) ∆XI
18) ∆YI
−200

−250
1 2 3 4 5 6
Iteration count

Figure 5.2: Convergence of estimated corrections.

Convergence in adjustment iteration

4
10

2
10

0
10
RMS of all estimated corrections in m

−2
10

−4
10

−6
10

−8
10

−10
10

−12
10
1 2 3 4 5 6
Iteration count

Figure 5.3: Convergence in adjustment iteration.

51
5 Geomatics examples

Approximate Points (blue), Adjusted Points (red), Datum Points (Triangles): ∆XA= ∆YA= ∆XB=0
726000
(A)
A

B
725500 (C)
(B)
C

725000

(H)
G
724500 H
(G)

724000
Y/m

(D)
723500
(I) I

723000

722500 (F)
F

E
(E)
722000
183000 183500 184000 184500 185000 185500 186000 186500
X/m

Figure 5.4: Approximate, adjusted and datum points.

52
 5.1 A-Model: Adjustment of observation equations

Adjusted Points (Circles), Datum Points (Triangles): ∆XA= ∆YA= ∆XB=0

726000
A

B
725500
C

725000

G
724500 H

724000
Y/m

723500
I

723000

1 cm
722500 Error Ellipses
F

722000
183000 183500 184000 184500 185000 185500 186000 186500
X/m

Figure 5.5: Adjusted and datum points with error ellipses.

53
 leg Δ𝑌B Δ𝑋 C Δ𝑌C Δ𝑋 D Δ𝑌D Δ𝑋 E Δ𝑌E Δ𝑋 F Δ𝑌F Δ𝑋 G Δ𝑌G Δ𝑋 H Δ𝑌H Δ𝑋 I Δ𝑌I
A-B −0.318 48 0 0 0 0 0 0 0 0 0 0 0 0 0 0
A-C 0 −0.942 33 −0.334 68 0 0 0 0 0 0 0 0 0 0 0 0
A-G 0 0 0 0 0 0 0 0 0 0.158 77 −0.987 31 0 0 0 0
B-G 0.689 66 0 0 0 0 0 0 0 0 −0.724 13 −0.689 66 0 0 0 0
B-H 0.980 57 0 0 0 0 0 0 0 0 0 0 0.196 15 −0.980 57 0 0
C-D 0 −0.301 13 0.953 58 0.301 13 −0.953 58 0 0 0 0 0 0 0 0 0 0
C-G 0 −0.777 94 0.628 34 0 0 0 0 0 0 0.777 94 −0.628 34 0 0 0 0
C-I 0 −0.615 27 0.788 32 0 0 0 0 0 0 0 0 0 0 0.615 27 −0.788 32
D-E 0 0 0 −0.316 23 0.948 68 0.316 23 −0.948 68 0 0 0 0 0 0 0 0
D-G 0 0 0 −0.635 71 −0.771 93 0 0 0 0 0.635 71 0.771 93 0 0 0 0
D-H 0 0 0 −0.865 43 −0.501 04 0 0 0 0 0 0 0.865 43 0.501 04 0 0
D-I 0 0 0 −0.988 94 0.148 34 0 0 0 0 0 0 0 0 0.988 94 −0.148 34
E-F 0 0 0 0 0 −0.913 81 −0.406 14 0.913 81 0.406 14 0 0 0 0 0 0
E-I 0 0 0 0 0 −0.347 31 −0.937 75 0 0 0 0 0 0 0.347 31 0.937 75
F-H 0 0 0 0 0 0 0 −0.221 62 −0.975 13 0 0 0.221 62 0.975 13 0 0
F-I 0 0 0 0 0 0 0 0.388 06 −0.921 64 0 0 0 0 −0.388 06 0.921 64
G-H 0 0 0 0 0 0 0 0 0 −0.978 98 −0.203 95 0.978 98 0.203 95 0 0
G-I 0 0 0 0 0 0 0 0 0 −0.287 35 0.957 83 0 0 0.287 35 −0.957 83
H-I 0 0 0 0 0 0 0 0 0 0 0 0.584 30 0.811 53 −0.584 30 −0.811 53
Table 5.8: 𝐴-Matrix (1st iteration).
5 Geomatics examples

54
 5.1 A-Model: Adjustment of observation equations

5.1.3 Distance and Direction Network (1)

A monitoring situation where directions and distances to 4 points A, B, C, and D are measured from
point N, see table 5.10. Coordinates (Table 5.9) of points 1 to 4 including N0 and the orientation
𝜔 N0 = 63.5610 gon are approximately given (see Jäger, 2005, pg. 241–242).

2000
Datum point
New Point
B

1500

1000 N
Y/m

500
C
A

D
0
0 500 1000 1500 2000 2500
X/m

Figure 5.6: Distance and direction network (1).

The standard deviation of observations are assumed to be 𝜎𝑠 = ±1 cm for distances and 𝜎𝑟 =

±0.5 mgon for directions. Thus, for the choice of 𝜎0 = ±1 m as a priori standard deviation the
elements of the weight matrix 𝑃 are obtained:

𝜎02 𝜎02 m2
𝑃𝑠 = = 10000, 𝑃𝑟 = = 1.6211 · 1010
𝜎𝑠2 𝜎𝑟2 rad2

point ID X/m Y/m measured: distance in m direction in gon

A 410.780 380.130 from to 𝜎𝑠 = ±1 cm 𝜎𝑟 = ±0.5 mgon
B 1183.460 1762.670 N A 982.690 193.1749
C 2077.030 433.380 B 765.000 337.1304
D 1207.570 124.630 C 1063.890 72.0344
N0 1175.150 997.720 D — 134.0758

Table 5.9: Coordinates. Table 5.10: Distance and direction observations.

55
5 Geomatics examples

leg Δ𝑦/m
leg Δ𝑋 N Δ𝑌N Δ𝜔 0 phys. unit N–A 0.000 37
N–A 0.777 83 0.628 47 0 N–B 0.004 86
m N–C
distance N–B −0.010 86 −0.999 94 0 −, rad
−0.002 46
N–C −0.847 72 0.530 45 0 leg Δ𝑦/rad
N–A 0.000 64 −0.000 79 −1 N–A −7.4935 · 10−6
N–B −0.001 31 0.000 01 −1 rad N–B −2.5420 · 10−6
direction m , −
N–C 0.000 50 0.000 80 −1 N–C 1.8678 · 10−6
N–D 0.001 14 0.000 04 −1 N–D −5.6276 · 10−6

Table 5.11: Designmatrix 𝐴. Table 5.12: Δ𝑦.

Table 5.11 shows the Designmatrix 𝐴, table 5.12 the reduced observation vector.
Table 5.13 contains the estimated parameters updates after the 1st iteration and table 5.14 contains
the adjusted coordinates for point N, including the adjusted orientation.

c
Δ𝜉/m c
Δ𝜉/rad
Δ𝑋ˆ N0 −0.000 49 Δ𝜔ˆ N0 3.3544 · 10−6 𝑋ˆ N 1175.150 m 𝜔ˆ N 63.5612 gon
Δ𝑌ˆN0 0.001 60 𝑌ˆN 997.722 m

Table 5.13: Estimated parameters. Table 5.14: Adjusted Coordinates and orientation.

Table 5.15 gives the adjusted observations including the distance 𝑠 ND , which is approximately given
by the adjusted coordinates of point N. Table 5.16 shows the inconsistencies 𝑒ˆ of the observations.

leg ˆ
𝑠/m leg 𝑟ˆ/rad leg ˆ
𝑒/m leg ˆ
𝑒/gon
N–A 982.690 N–A 193.1751 N–A −0.0003 N–A −0.000 16
N–B 764.994 N–B 337.1304 N–B 0.0065 N–B 9.6 · 10−6
N–C 1063.894 N–C 72.0341 N–C −0.0037 N–C 0.000 27
N–D 873.693 N–D 134.0759 N–D — N–C −0.000 11

Table 5.15: Adjusted observations. Table 5.16: Inconsistencies.

=⇒ 𝑒ˆT 𝑃 𝑒ˆ = 0.999 321 8 m2, c k < 10−10 ).

(3 iterations, stop criterion k Δ𝜉

Finally, tables 5.17 and 5.18 show the standard deviations for coordinates for adjusted point N, as
well as for adjusted orientation and observations.
Figure 5.7 shows the situation in detail, including the error ellipse for the adjusted coordinates of
point N.

56
 5.1 A-Model: Adjustment of observation equations

𝜎ˆ𝑋ˆ N /cm 𝜎ˆ𝑌ˆN /cm 𝜎ˆ𝜔ˆ N /mgon

±0.19 ±0.26 ±0.13

Table 5.17: Standard deviations of coordinates and orientation.

𝜎ˆ𝑟ˆNA 𝜎ˆ𝑟ˆNB 𝜎ˆ𝑟ˆNC 𝜎ˆ𝑟ˆND 𝜎ˆ𝑠ˆNA 𝜎ˆ𝑠ˆNB 𝜎ˆ𝑠ˆNC 𝜎ˆ𝑠ˆND

in mgon in cm
±0.19 ±0.23 ±0.18 ±0.17 ±0.22 ±0.26 ±0.21 ±0.26

Table 5.18: Standard deviations of directions and distances.

Approximate point
Adjusted Point
997.724

997.722

997.72 N
Y/m

997.718

1cm
Error Ellipses
997.716

1175.146 1175.148 1175.15 1175.152 1175.154 1175.156

X/m

Figure 5.7: Detailed view of point N.

57
5 Geomatics examples

5.1.4 Distance and Direction Network (2a)

1 Benchmark 2
1000 New Point

800

600
Y/m

400

200

0
3 4

0 200 400 600 800 1000

X/m

Figure 5.8: Network of Points.

This example (Benning, 2011, pg. 258–261) treats a network (Figure 5.8) of measured distances and
directions between two given points and two new points. The standard deviation of observations:
𝜎𝑠 = ±1 cm for distances, 𝜎𝑟 = ±1 mgon for directions and 𝜎0 = ±1 cm as a priori standard deviation.
This give the elements for the weight-matrix 𝑃:

𝜎02 𝜎02 m2
=⇒ 𝑃𝑠 = = 1, 𝑃𝑟 = = 4.0528 · 105
𝜎𝑠2 𝜎𝑟2 rad2

Table 5.19 contains coordinates for benchmarks 1 and 2, table 5.20 approximate coordinates for
points 3 and 4.

Point ID 𝑋 /m 𝑌 /m Point ID 𝑋 0 /m 𝑌0 /m
1 0.00 1000.00 3 0.00 0.00
2 1000.00 1000.00 4 1000.00 0.00

Table 5.19: Benchmarks. Table 5.20: Approximate coordinates.

Table 5.21 contains measured distances (𝑠) between individual points, approximate distances (𝑠 0 ) and
reduced observations (Δ𝑠).

58
 5.1 A-Model: Adjustment of observation equations

leg 𝑠/m 𝑠 0 /m Δ𝑠/m

1–3 1000.02 1000.0000 0.0200
1–4 1414.20 1414.2136 −0.0136
2–3 1414.24 1414.2136 0.0264
2–4 999.98 1000.0000 −0.0200
3–4 1000.00 1000.0000 0.0000

Table 5.21: Distance observations.

Table 5.22 displays direction observations (𝑟 ), approximate grid bearings (𝑇0 ), approximate orienta-
tion unknows (𝜔 0 ), approximate directions (𝑟 0 ) and reduced direction observations (Δ𝑟 0 ).

leg 𝑟 /gon 𝑇0 /gon 𝜔 0 /gon 𝑟 0 = 𝑇0 − 𝜔 0 in gon Δ𝑟 0 = 𝑟 − 𝑟 0 in gon

1–3 50.001 200.000 150.000 50.000 0.001
1–4 0.000 150.000 0.000 0
2–3 49.998 250.000 200.000 50.000 −0.002
2–4 0.000 200.000 0.000 0
3–1 0.000 0.000 0.000 0.000 0
3–2 49.999 50.000 50.000 −0.001
3–4 99.997 100.000 100.000 −0.003

Table 5.22: Direction observations.

Table 5.23 contains the designmatrix 𝐴 and table 5.24 the reduced observation vector Δ𝑦 after 1st
iteration.
leg Δ𝑦 /m
leg Δ𝑋 3 Δ𝑌3 Δ𝑋 4 Δ𝑌4 Δ𝜔 1 Δ𝜔 2 Δ𝜔 3 phys. unit 1–3 0.02
1–3 0 −1 0 0 0 0 0 1–4 −0.0136
1–4 0 0 0.7071 −0.7071 0 0 0 2–3 0.0264
distance 2–3 −0.7071 −0.7071 0 0 0 0 0 m
−, rad 2–4 −0.02
2–4 0 0 0 −1 0 0 0 3–4 0
3–4 −1 0 1 0 0 0 0 leg Δ𝑦 /rad
1–3 −0.001 0 0 0 −1 0 0 1–3 1.5708 · 10−5
1–4 0 0 −0.0005 −0.0005 −1 0 0 1–4 0
2–3 −0.0005 0.0005 0 0 0 −1 0 2–3 −3.1416 · 10−5
direction 2–4 0 0 −0.001 0 0 −1 0 rad , − 2–4 0
m
3–1 −0.001 0 0 0 0 0 −1 3–1 0
3–2 −0.0005 0.0005 0 0 0 0 −1 3–2 −1.5708 · 10−5
3–4 0 0.001 0 −0.001 0 0 −1 3–4 −4.7124 · 10−5

Table 5.23: Designmatrix 𝐴. Table 5.24: Δ𝑦.

Table 5.25 shows the estimated parameter updates after the 1st iteration and table 5.26 contains the
adjusted coordinates for point 3 and 4 and adjusted orientations.
Table 5.27 contains the adjusted observations for distances and directions, and 5.28 the estimated
inconsistencies 𝑒ˆ of the observations.

=⇒ 𝑒ˆT 𝑃 𝑒ˆ = 1.0463 cm2 c k < 10−12 )

(3 iterations with stop criterion k Δ𝜉

Finally, the tables 5.29, 5.30 and 5.31 contain standard deviations for adjusted coordinates and ori-
entations as well as for adjusted distance and direction observations.

59
5 Geomatics examples

c
Δ𝜉/m c
Δ𝜉/gon
Δ𝑋ˆ 3 −0.0101 Δ𝜔ˆ 1 −0.0003 𝑋ˆ 3 −0.010 m 𝜔ˆ 1 149.9997 gon
Δ𝑌ˆ3 −0.0231 Δ𝜔ˆ 2 0.0011 𝑌ˆ3 −0.023 m 𝜔ˆ 2 200.0011 gon
Δ𝑋ˆ 4 −0.0096 Δ𝜔ˆ 3 0.0006 𝑋ˆ 4 999.990 m 𝜔ˆ 3 0.0006 gon
Δ𝑌ˆ4 0.0163 𝑌ˆ4 0.016 m

Table 5.25: Estimated parameters. Table 5.26: Adjusted coordinates and orientations.

leg ˆ
𝑠/m leg 𝑟ˆ/gon leg ˆ
𝑒/m leg ˆ
𝑒/gon
1–3 1000.023 1–3 50.0009 1–3 −0.0031 1–3 0.0001
1–4 1414.195 1–4 0.0001 1–4 0.0048 1–4 −0.0001
2–3 1414.237 2–3 49.9985 2–3 0.0029 2–3 −0.0005
2–4 999.984 2–4 399.9995 2–4 −0.0037 2–4 0.0005
3–4 1000.001 3–1 0.0001 3–4 −0.0005 3–1 −0.0001
3–2 49.9990 3–2 0.0000
3–4 99.9969 3–4 0.0001

Table 5.27: Adjusted observations. Table 5.28: Inconsistencies.

𝜎ˆ𝑋ˆ 3 𝜎ˆ𝑌ˆ3 𝜎ˆ𝑋ˆ 4 𝜎ˆ𝑌ˆ4 𝜎ˆ𝜔ˆ 1 𝜎ˆ𝜔ˆ 2 𝜎ˆ𝜔ˆ 3

±0.56 cm ±0.41 cm ±0.57 cm ±0.40 cm ±0.44 mgon ±0.44 mgon ±0.41 mgon

Table 5.29: Standard deviations of adjusted coordinates and orientations.

𝜎ˆ𝑠ˆ13 𝜎ˆ𝑠ˆ14 𝜎ˆ𝑠ˆ23 𝜎ˆ𝑠ˆ24 𝜎ˆ𝑠ˆ34

±0.41 cm ±0.36 cm ±0.35 cm ±0.40 cm ±0.38 cm

Table 5.30: Standard deviations of distance observations.

𝜎ˆ𝑟ˆ13 = 𝜎ˆ𝑟ˆ14 𝜎ˆ𝑟ˆ23 = 𝜎ˆ𝑟ˆ24 𝜎ˆ𝑟ˆ31 𝜎ˆ𝑟ˆ32 𝜎ˆ𝑟ˆ34

±0.34 mgon ±0.35 mgon ±0.30 mgon ±0.28 mgon ±0.32 mgon

Table 5.31: Standard deviations of direction observations.

Figure 5.9 shows network of points including error ellipses, figure 5.10 and figure 5.11 give a detailed
view for points 3 and 4.

60
 5.1 A-Model: Adjustment of observation equations

1 Benchmark 2
1000 New Point

800

600
Y/m

400

200

1cm
Error Ellipses

0
3 4

0 200 400 600 800 1000

X/m

Figure 5.9: Network of points with error ellipses.

0.01 0.03
Approximate Point
Approximate Point Adjusted Point
Adjusted Point
0.005 0.025

0 0.02
3

−0.005 0.015 4

−0.01
Y/m

0.01
Y/m

1cm 1cm

Error Ellipses Error Ellipses

−0.015 0.005

−0.02 0
4
3

−0.025 −0.005

−0.03 −0.01
−0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 999.98 999.985 999.99 999.995 1000 1000.005 1000.01 1000.015 1000.02
X/m X/m

Figure 5.10: Detailed view point 3. Figure 5.11: Detailed view point 4.

61
5 Geomatics examples

5.1.5 Free Adjustment: Distance and Direction Network (2b)

This example (Benning, 2011, pg. 273–281) processes the same network as before (Figure 5.8). There-
fore, observations, weights and reduced observations Δ𝑦 (Table 5.24) do not change. However, since
we deal with a free adjustment here, designmatrix 𝐴 (Table 5.32) is augmented by four additional
columns comprising partial derivatives of those observations which also involve points 1 and 2.

leg Δ𝑋 1 Δ𝑌1 Δ𝑋 2 Δ𝑌2 Δ𝑋 3 Δ𝑌3 Δ𝑋 4 Δ𝑌4 Δ𝜔 1 Δ𝜔 2 Δ𝜔 3 phys. unit

1–3 0 1 0 0 0 −1 0 0 0 0 0
1–4 −0.707 11 0.707 11 0 0 0 0 0.707 11 −0.707 11 0 0 0
m
−, rad
distance 2–3 0 0 0.707 11 0.707 11 −0.707 11 −0.707 11 0 0 0 0 0
2–4 0 0 0 1 0 0 0 −1 0 0 0
3–4 0 0 0 0 −1 0 1 0 0 0 0
1–3 0.001 0 0 0 −0.001 0 0 0 −1 0 0
1–4 0.0005 0.0005 0 0 0 0 −0.0005 −0.0005 −1 0 0
2–3 0 0 0.0005 −0.0005 −0.0005 0.0005 0 0 0 −1 0
direction 2–4 0 0 0.001 0 0 0 −0.001 0 0 −1 0 rad , −
m
3–1 0.001 0 0 0 −0.001 0 0 0 0 0 −1
3–2 0 0 0.0005 −0.0005 −0.0005 0.0005 0 0 0 0 −1
3–4 0 0 0 0 0 0.001 0 −0.001 0 0 −1

Table 5.32: Designmatrix A for distances and directions.

The free adjustment process makes use of the pseudoinverse 𝑁 + shown in table 5.33 (1st iteration).

Δ𝑋 1 Δ𝑌1 Δ𝑋 2 Δ𝑌2 Δ𝑋 3 Δ𝑌3 Δ𝑋 4 Δ𝑌4 Δ𝜔 1 Δ𝜔 2 Δ𝜔 3

Δ𝑋 1 0.796 57 0.084 53 −0.762 01 0.144 78 −0.052 62 −0.119 09 0.018 06 −0.110 22 0.000 67 −0.000 63 0.000 12
Δ𝑌1 0.084 53 0.290 79 −0.209 99 −0.080 26 0.074 64 −0.165 33 0.050 82 −0.045 20 0.000 10 −0.000 22 −0.000 10
Δ𝑋 2 −0.762 01 −0.209 99 0.928 06 −0.036 65 −0.032 15 0.043 95 −0.133 90 0.202 69 −0.000 63 0.000 79 −0.000 12
Δ𝑌2 0.144 78 −0.080 26 −0.036 65 0.263 51 −0.041 16 −0.027 87 −0.066 96 −0.155 39 0.000 16 −0.000 06 0.000 06
Δ𝑋 3 −0.052 62 0.074 64 −0.032 15 −0.041 16 0.205 97 0.010 13 −0.121 21 −0.043 60 −0.000 08 0.000 00 −0.000 10
Δ𝑌3 −0.119 09 −0.165 33 0.043 95 −0.027 87 0.010 13 0.240 47 0.065 01 −0.047 27 −0.000 14 0.000 07 0.000 10
Δ𝑋 4 0.018 06 0.050 82 −0.133 90 −0.066 96 −0.121 21 0.065 01 0.237 04 −0.048 87 0.000 04 −0.000 16 0.000 10
Δ𝑌4 −0.110 22 −0.045 20 0.202 69 −0.155 39 −0.043 60 −0.047 27 −0.048 87 0.247 86 −0.000 12 0.000 21 −0.000 06
Δ𝜔 1 0.000 67 0.000 10 −0.000 63 0.000 16 −0.000 08 −0.000 14 0.000 04 −0.000 12 0.000 00 0.000 00 0.000 00
Δ𝜔 2 −0.000 63 −0.000 22 0.000 79 −0.000 06 0.000 00 0.000 07 −0.000 16 0.000 21 0.000 00 0.000 00 0.000 00
Δ𝜔 3 0.000 12 −0.000 10 −0.000 12 0.000 06 −0.000 10 0.000 10 0.000 10 −0.000 06 0.000 00 0.000 00 0.000 00

Table 5.33: Pseudoinverse 𝑁 + .

Table 5.34 shows estimated parameters (1st iteration) and table 5.35 contains the adjusted coordinates
for point 1–4 and also adjusted orientations.

c
Δ𝜉/m c
Δ𝜉/gon in m in gon
Δ𝑋ˆ 1 0.0018 Δ𝜔ˆ 1 −0.0003 𝑋ˆ 1 0.002 𝜔ˆ 1 149.9997
Δ𝑌ˆ1 0.0031 Δ𝜔ˆ 2 0.0017 𝑌ˆ1 1000.003 𝜔ˆ 2 200.0017
Δ𝑋ˆ 2 0.0135 Δ𝜔ˆ 3 0.0008 𝑋ˆ 2 1000.013 𝜔ˆ 3 0.0008
Δ𝑌ˆ2 −0.0014 𝑌ˆ2 999.999
Δ𝑋ˆ 3 −0.0076 𝑋ˆ 3 −0.008
Δ𝑌ˆ3 −0.0184 𝑌ˆ3 −0.018
Δ𝑋ˆ 4 −0.0077 𝑋ˆ 4 999.992
Δ𝑌ˆ4 0.0167 𝑌ˆ4 0.017

Table 5.34: Estimated parameters. Table 5.35: Adjusted coordinates and orientations.

62
 5.1 A-Model: Adjustment of observation equations

Table 5.36 contains the adjusted observations for distances and directions, while table 5.37 shows
the estimated inconsistencies 𝑒ˆ of the observations.

leg ˆ
𝑠/m leg 𝑟ˆ/gon leg ˆ
𝑒/gon leg ˆ
𝑒/gon
1–3 1000.021 1–3 50.0009 1–3 −0.0015 1–3 0.0001
1–4 1414.197 1–4 0.0001 1–4 0.0027 1–4 −0.0001
2–3 1414.240 2–3 49.9984 2–3 −0.0004 2–3 −0.0004
2–4 999.982 2–4 399.9996 2–4 −0.0019 2–4 0.0004
3–4 1000.000 3–1 399.9998 3–4 0.0001 3–1 0.0002
3–2 49.9993 3–2 −0.0003
3–4 99.9969 3–4 0.0001

Table 5.36: Adjusted observations. Table 5.37: Inconsistencies.

=⇒ 𝑒ˆT 𝑃 𝑒ˆ = 0.628 cm2 c k < 10−12 )

(3 iterations with stop criterion k Δ𝜉

Finally again, the tables 5.38, 5.39, 5.40 and 5.41 contain standard deviations for adjusted coordinates
and orientations as well as for adjusted distance and direction observations.

𝜎ˆ𝑋ˆ 1 𝜎ˆ𝑌ˆ1 𝜎ˆ𝑋ˆ 2 𝜎ˆ𝑌ˆ2 𝜎ˆ𝑋ˆ 3 𝜎ˆ𝑌ˆ3 𝜎ˆ𝑋ˆ 4 𝜎ˆ𝑌ˆ4

±0.35 cm ±0.21 cm ±0.38 cm ±0.20 cm ±0.18 cm ±0.19 cm ±0.19 cm ±0.20 cm

Table 5.38: Standard deviations of adjusted coordinates.

𝜎ˆ𝜔ˆ 1 𝜎ˆ𝜔ˆ 2 𝜎ˆ𝜔ˆ 3

±0.34 mgon ±0.35 mgon ±0.25 mgon

Table 5.39: Standard deviations of adjusted orientations.

𝜎ˆ𝑠ˆ13 𝜎ˆ𝑠ˆ14 𝜎ˆ𝑠ˆ23 𝜎ˆ𝑠ˆ24 𝜎ˆ𝑠ˆ34

±0.37 cm ±0.34 cm ±0.37 cm ±0.36 cm ±0.33 cm

Table 5.40: Standard deviations of adjusted distance observations.

𝜎ˆ𝑟ˆ13 = 𝜎ˆ𝑟ˆ14 𝜎ˆ𝑟ˆ23 = 𝜎ˆ𝑟ˆ24 𝜎ˆ𝑟ˆ31 𝜎ˆ𝑟ˆ32 𝜎ˆ𝑟ˆ34

±0.30 mgon ±0.31 mgon ±0.32 mgon ±0.30 mgon ±0.28 mgon

Table 5.41: Standard deviations of adjusted direction observations.

Figure 5.12 shows network of points including error ellipses, figure 5.13, 5.14, 5.15 and 5.16 give a
detailed view for points 1, 2, 3 and 4.

63
5 Geomatics examples

1 Approximate Point 2
1000 Adjusted Point

800

600

Y/m
400

200

1cm
Error Ellipses

0
3 4

0 200 400 600 800 1000

X/m

Figure 5.12: Network of points with error ellipses.

1000.02 1000.02
Approximate Point Approximate Point
Adjusted Point Adjusted Point
1000.015 1000.015
1cm 1cm
Error Ellipses Error Ellipses
1000.01 1000.01

1000.005 1000.005
1

1000 1000
Y/m

Y/m

1 2
2

999.995 999.995

999.99 999.99

999.985 999.985

999.98 999.98
−0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 999.985 999.99 999.995 1000 1000.005 1000.01 1000.015 1000.02 1000.025
X/m X/m

Figure 5.13: Detailed view point 1. Figure 5.14: Detailed view point 2

0.015 0.03
Approximate Point Approximate Point
Adjusted Point Adjusted Point
0.01 0.025

0.005 0.02

4
0 0.015
3

−0.005 0.01
Y/m
Y/m

1cm
Error Ellipses
−0.01 0.005
1cm
Error Ellipses
−0.015 0
4
3

−0.02 −0.005

−0.025 −0.01
−0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 999.98 999.985 999.99 999.995 1000 1000.005 1000.01 1000.015 1000.02
X/m X/m

Figure 5.15: Detailed view point 3 Figure 5.16: Detailed view point 4

64
 5.1 A-Model: Adjustment of observation equations

5.1.6 Overconstrained adjustment: Distance, direction and angle network

This example, taken from Wolf (1979, pg. 66–78), consists of a 10-point network observed by direc-
tions, one distance and one angle, see (5.17). The network is overconstrained because its datum is
defined by 6 benchmarks A–F.

727000

A B
726500

726000

725500
C G H

725000

724500
Y/m

724000

723500
I
D
723000

722500

722000
Benchmark F
New Point E
721500
183000 184000 185000 186000 187000
X/m

Figure 5.17: Network design

The observations are collected in table 5.43 (2nd column), and the following standard deviations have
been assumed:

𝜎𝑟 =±2.5 mgon (directions),

𝜎𝛼 =±3.5 mgon (angle),
𝜎𝑠 = ±3 cm (distance).

If the a priori standard deviation is taken as 𝜎0 = ±2.5 mgon then the weight matrix elements turn
out to be
𝜎2 𝜎2 𝜎2 rad2
𝑃𝑟 = 02 = 1, 𝑃𝛼 = 02 = 0.5102 and 𝑃𝑠 = 02 = 1.7135 · 10−6 2 .
𝜎𝑟 𝜎𝛼 𝜎𝑠 m

Table 5.42 contains coordinates for benchmarks A to F and approximate coordinates G, H and I.

65
5 Geomatics examples

benchmarks A–F
point ID 𝑋 /m 𝑌 /m
A 184 423.28 726 419.33
B 186 444.18 726 476.66
C 183 257.84 725 490.35
D 184 292.00 723 313.00
E 185 487.00 721 829.00
F 186 708.72 722 104.58
approximate coordinates
new points G–I
point ID 𝑋 0 /m 𝑌0 /m
G 184 868.20 725 139.70
H 186 579.30 725 336.60
I 185 963.07 723 322.02

Table 5.42: Benchmarks and approximate coordinates.

Additionally to observations, table 5.43 includes approximate grid bearings (𝑇0 ), angle (𝛼 0 ) and dis-
tance (𝑠 0 ), approximate orientation unknowns (𝜔 0 ), approximate directions (𝑟 0 ) and reduced obser-
vations (Δ𝑟 0 ). Orientation unknowns (𝜔 0 ) are mean values calculated from Δ𝑟 0 .
Table 5.44 contains the designmatrix 𝐴 after 1st iteration.
Table 5.45 shows the estimated parameters after 1st iteration and table 5.46 the adjusted coordinates
and orientations.
Table 5.47 contains the estimated inconsistencies 𝑒ˆ of the observations leading to a weighted square
sum of residuals

𝑒ˆT 𝑃 𝑒ˆ = 0.00225 gon2 c k < 10−10 )

(4 iterations, stop criteria k Δ𝜉

Finally, the tables 5.48 and 5.49 contain the standard deviations of coordinates, orientations and
observations.
Figures 5.18 and 5.19 show the adjusted network with corresponding absolute and relative error
ellipses at/between the new points.
Table 5.50 contains the elements (semi major axis 𝑎, semi minor axis 𝑏 and bearing 𝜙 of 𝑎) for absolute
error ellipses for new points and also for relative error ellipses between new points.
Figures 5.20, 5.21 and 5.22 give a detailed view, for the new points G, H and I.

66
 5.1 A-Model: Adjustment of observation equations

Distance observation in m
leg 𝑠 𝑠0 Δ𝑠 0
G– I 2121.90 2121.96 −0.06 — — —
Direction observations in gon
leg 𝑟 𝑇0 Δ𝑟 0 = 𝑇0 − 𝑟 𝜔0 𝑟 𝜔0 = 𝑇0 − 𝜔 0 Δ𝑟 𝜔0 = 𝑟 − 𝑟 𝜔0
A–B 0.0000 98.1945 98.1945 98.1960 −0.0008 0.0008
A–G 80.5000 178.6975 98.1975 80.5022 −0.0022
A–C 158.9610 257.1571 98.1961 158.9618 −0.0008
B–H 0.0000 192.4898 192.4861 192.4861 0.0073 −0.0073
B–G 62.7260 255.2121 192.4861 62.7296 −0.0036
B–A 105.7120 298.1945 192.4824 105.7120 0.0000
C–A 0.0000 57.1571 57.1571 57.1640 −0.0095 0.0095
C–G 56.4960 113.6491 57.1531 56.4825 0.0135
C– I 85.8450 143.0148 57.1698 85.8482 −0.0032
C–D 114.5950 171.7711 57.1761 114.6045 −0.0095
D–G 0.0000 19.4522 19.4522 19.4476 0.0015 −0.0015
D–H 34.4500 53.8894 19.4394 34.4387 0.0113
D– I 80.2110 99.6564 19.4454 80.2057 0.0053
D– E 137.4020 156.8412 19.4391 137.3905 0.0115
D–C 352.3090 371.7711 19.4621 352.3205 −0.0115
E– I 0.0000 19.6507 19.6507 19.6357 0.0224 −0.0224
E– F 66.2450 85.8763 19.6313 66.2481 −0.0031
E–D 337.2160 356.8412 19.6251 337.2129 0.0031
F–E 0.0000 285.8763 285.8763 285.8642 0.0000 0.0000
F– I 79.1690 365.0151 285.8461 79.1388 0.0302
F –H 111.5820 397.4521 285.8701 111.5758 0.0062
G–B 0.0000 55.2121 55.2121 55.2147 −0.0027 0.0027
G–H 37.4980 92.7064 55.2083 37.4916 0.0064
G– I 110.2580 165.4862 55.2282 110.2715 −0.0135
G–D 164.2320 219.4522 55.2201 164.2374 −0.0054
G–C 258.4410 313.6491 55.2081 258.4344 0.0066
G–A 323.4860 378.6975 55.2115 323.4827 0.0033
H– F 0.0000 197.4521 197.4521 197.4508 0.0013 −0.0013
H– I 21.4450 218.8979 197.4529 21.4471 −0.0021
H–D 56.4420 253.8889 197.4474 56.4386 0.0034
I –H 0.0000 18.8979 18.8979 18.9025 −0.0046 0.0046
I –F 146.1430 165.0151 18.8721 146.1126 0.0304
I –E 200.7330 219.6507 18.9177 200.7482 −0.0152
I –D 280.7560 299.6564 18.9004 280.7539 0.0021
I –C 324.1050 343.0148 18.9098 324.1123 −0.0073
I –G 346.5690 365.4862 18.9172 346.5837 −0.0147
Angle observation in gon
leg 𝛼 𝛼0 Δ𝛼 0
𝛼 HGB 99.7810 99.7834 −0.0024 — — —

Table 5.43: Observations.

67
5 Geomatics examples

leg Δ𝑋 G Δ𝑌G Δ𝑋𝐻 Δ𝑌𝐻 Δ𝑋𝐼 Δ𝑌𝐼 Δ𝜔𝐴 Δ𝜔 𝐵 Δ𝜔𝐶 Δ𝜔 𝐷 Δ𝜔 𝐸 Δ𝜔 𝐹 Δ𝜔𝐺 Δ𝜔 𝐻 Δ𝜔 𝐼 phys. unit
Distance observation
G– I −0.5159 0.8566 0 0 0.5159 −0.8566 0 0 0 0 0 0 0 0 0 m
−, rad
Direction observations
A–B 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0
A –G −0.0007 −0.0002 0 0 0 0 −1 0 0 0 0 0 0 0 0
A –C 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0
B –H 0 0 −0.0009 −0.0001 0 0 0 −1 0 0 0 0 0 0 0
B –G −0.0003 0.0004 0 0 0 0 0 −1 0 0 0 0 0 0 0
B –A 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0
C –A 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0
C –G −0.0001 −0.0006 0 0 0 0 0 0 −1 0 0 0 0 0 0
C–I 0 0 0 0 −0.0002 −0.0002 0 0 −1 0 0 0 0 0 0
C –D 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0
D –G 0.0005 −0.0002 0 0 0 0 0 0 0 −1 0 0 0 0 0
D –H 0 0 0.0002 −0.0002 0 0 0 0 0 −1 0 0 0 0 0
D– I 0 0 0 0 3.2 · 10−6 −0.0006 0 0 0 −1 0 0 0 0 0
D–E 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0
D –C 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0
E–I 0 0 0 0 0.0006 −0.0002 0 0 0 0 −1 0 0 0 0
E–F 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0
E –D 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 rad , −
F –E 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 m
F–I 0 0 0 0 0.0006 0.0004 0 0 0 0 0 −1 0 0 0
F –H 0 0 0.0003 1.2 · 10−5 0 0 0 0 0 0 0 −1 0 0 0
G–B −0.0003 0.0004 0 0 0 0 0 0 0 0 0 0 −1 0 0
G –H −6.6 · 10−5 0.0006 6.6 · 10−5 −0.0006 0 0 0 0 0 0 0 0 −1 0 0
G– I 0.0004 0.0002 0 0 −0.0004 −0.0002 0 0 0 0 0 0 −1 0 0
G –D 0.0005 −0.0001 0 0 0 0 0 0 0 0 0 0 −1 0 0
G –C −0.0001 −0.0006 0 0 0 0 0 0 0 0 0 0 −1 0 0
G –A −0.0007 −0.0002 0 0 0 0 0 0 0 0 0 0 −1 0 0
H–F 0 0 0.0003 1.2 · 10−5 0 0 0 0 0 0 0 0 0 −1 0
H– I 0 0 0.0005 −0.0001 −0.0005 0.0001 0 0 0 0 0 0 0 −1 0
H –D 0 0 0.0002 −0.0002 0 0 0 0 0 0 0 0 0 −1 0
I –H 0 0 0.0005 −0.0001 −0.0005 0.0001 0 0 0 0 0 0 0 0 −1
I –F 0 0 0 0 0.0006 0.0004 0 0 0 0 0 0 0 0 −1
I –E 0 0 0 0 0.0006 −0.0002 0 0 0 0 0 0 0 0 −1
I –D 0 0 0 0 3.2 · 10−6 −0.0006 0 0 0 0 0 0 0 0 −1
I –C 0 0 0 0 −0.0002 −0.0002 0 0 0 0 0 0 0 0 −1
I –G 0.0004 0.0002 0 0 −0.0004 −0.0002 0 0 0 0 0 0 0 0 −1
Angle observation
𝛼 HGB 6.6 · 10−5 −0.0006 −0.0009 0.0004 0 0 0 0 0 0 0 0 0 0 0 rad , −
m

Table 5.44: Designmatrix 𝐴 (after 1st iteration).

c
Δ𝜉/m in m
Δ𝑋ˆ𝐺 −0.162 𝑋ˆ𝐺 184 868.038
Δ𝑌ˆ𝐺 −0.043 𝑌ˆ𝐺 725 139.657
Δ𝑋ˆ𝐻 0.037 𝑋ˆ𝐻 186 579.337
Δ𝑌ˆ𝐻 −0.186 𝑌ˆ𝐻 725 336.414
Δ𝑋ˆ𝐼 0.145 𝑋ˆ𝐼 185 963.215
Δ𝑌ˆ𝐼 0.283 𝑌ˆ𝐼 723 322.303
c
Δ𝜉/gon in gon
Δ𝜔ˆ 𝐴 0.0026 𝜔ˆ 𝐴 98.1987
Δ𝜔ˆ 𝐵 0.0005 𝜔ˆ 𝐵 192.4866
Δ𝜔ˆ 𝐶 −0.0007 𝜔ˆ 𝐶 57.1634
Δ𝜔ˆ 𝐷 −0.0024 𝜔ˆ 𝐷 19.4452
Δ𝜔ˆ 𝐸 0.0007 𝜔ˆ 𝐸 19.6364
Δ𝜔ˆ 𝐹 0.0042 𝜔ˆ 𝐹 285.8684
Δ𝜔ˆ 𝐺 0.0002 𝜔ˆ 𝐺 55.2150
Δ𝜔ˆ 𝐻 0.0017 𝜔ˆ 𝐻 197.4525
Δ𝜔ˆ 𝐼 −0.0024 𝜔ˆ 𝐼 18.9001

Table 5.45: Estimated parameters after Table 5.46: Adjusted coordinates and
1st iteration. orientations.

68
 5.1 A-Model: Adjustment of observation equations

leg ˆ
𝑠/m ˆ
𝑒/m
G– I 2121.836 0.0638
leg 𝑟ˆ/gon ˆ
𝑒/gon leg 𝑟ˆ/gon ˆ
𝑒/gon leg 𝑟ˆ/gon ˆ
𝑒/gon
A–B −0.0042 0.0042 D– I 80.2004 0.0106 G–D 164.2325 −0.0005
A–G 80.5067 −0.0067 D–E 137.3959 0.0061 G–C 258.4371 0.0039
A–C 158.9585 0.0025 D–C 352.3259 −0.0169 G–A 323.4904 −0.0044
B–H 0.0024 −0.0024 E– I 0.0164 −0.0164 H– F 0.0003 −0.0003
B–G 62.7277 −0.0017 E– F 66.2399 0.0051 H– I 21.4464 −0.0014
B–A 105.7079 0.0041 E–D 337.2047 0.0113 H–D 56.4403 0.0017
C–A −0.0062 0.0062 F –E 0.0079 −0.0079 I –H −0.0012 0.0012
C–G 56.4887 0.0072 F– I 79.1588 0.0102 I –F 146.1271 0.0159
C– I 85.8457 −0.0007 F –H 111.5843 −0.0023 I –E 200.7527 −0.0197
C–D 114.6078 −0.0128 G–B −0.0007 0.0007 I –D 280.7455 0.0105
D–G 0.0022 −0.0022 G–H 37.4975 0.0005 I –C 324.1089 −0.0039
D–H 34.4476 0.0024 G– I 110.2583 −0.0003 I –G 346.5731 −0.0041
leg ˆ
𝛼/gon ˆ
𝑒/gon
𝛼 HGB 99.7765 0.0045

Table 5.47: Estimated inconsistencies.

in cm in mgon
𝜎ˆ𝑋ˆ G ±11.866 𝜎ˆ𝜔ˆ A ±6.0023
𝜎ˆ𝑌ˆG ±13.078 𝜎ˆ𝜔ˆ B ±6.7376
𝜎ˆ𝑋ˆ H ±15.816 𝜎ˆ𝜔ˆ C ±5.1859
𝜎ˆ𝑌ˆH ±26.380 𝜎ˆ𝜔ˆ D ±4.8772
𝜎ˆ𝑋ˆ I ±11.470 𝜎ˆ𝜔ˆ E ±5.9353
𝜎ˆ𝑌ˆI ±13.537 𝜎ˆ𝜔ˆ F ±6.1002
𝜎ˆ𝜔ˆ G ±4.3863
𝜎ˆ𝜔ˆ H ±6.5588
𝜎ˆ𝜔ˆ I ±4.3554

Table 5.48: Estimated standard deviations of coordinates and orientations.

in cm
𝜎ˆ𝑠ˆ𝐺𝐼 ±10.2660
in mgon in mgon in mgon in mgon
𝜎ˆ𝑟ˆAB ±6.0023 𝜎ˆ𝑟ˆCD ±5.1859 𝜎ˆ𝑟ˆFE ±6.1002 𝜎ˆ𝑟ˆHF ±6.1800
𝜎ˆ𝑟ˆAG ±6.8033 𝜎ˆ𝑟ˆDG ±5.3020 𝜎ˆ𝑟ˆFI ±6.6864 𝜎ˆ𝑟ˆHI ±6.1533
𝜎ˆ𝑟ˆAC ±6.0023 𝜎ˆ𝑟ˆDH ±5.3830 𝜎ˆ𝑟ˆFH ±6.2494 𝜎ˆ𝑟ˆHD ±6.3495
𝜎ˆ𝑟ˆBH ±8.3169 𝜎ˆ𝑟ˆDI ±5.7282 𝜎ˆ𝑟ˆGB ±6.0317 𝜎ˆ𝑟ˆIH ±6.1128
𝜎ˆ𝑟ˆBG ±6.7802 𝜎ˆ𝑟ˆDE ±4.8772 𝜎ˆ𝑟ˆGH ±8.1933 𝜎ˆ𝑟ˆIF ±6.9589
𝜎ˆ𝑟ˆBA ±6.7376 𝜎ˆ𝑟ˆDC ±4.8772 𝜎ˆ𝑟ˆGI ±6.0859 𝜎ˆ𝑟ˆIE ±5.7188
𝜎ˆ𝑟ˆCA ±5.1859 𝜎ˆ𝑟ˆEI ±6.5637 𝜎ˆ𝑟ˆGD ±6.1072 𝜎ˆ𝑟ˆID ±5.6981
𝜎ˆ𝑟ˆCG ±6.0882 𝜎ˆ𝑟ˆEF ±5.9353 𝜎ˆ𝑟ˆGC ±6.1493 𝜎ˆ𝑟ˆIC ±4.5646
𝜎ˆ𝑟ˆCI ±5.2157 𝜎ˆ𝑟ˆED ±5.9353 𝜎ˆ𝑟ˆGA ±6.5839 𝜎ˆ𝑟ˆIG ±5.7435
in mgon
𝜎ˆ𝛼ˆHGB ±9.4045

Table 5.49: Estimated standard deviations of the observations.

69
5 Geomatics examples

727000
20 cm
Error Ellipses A B
726500
726000

726000 New Point

725500 725500
H
C G
H
725000 G

725000
724500
Y/m

724000 724500

Y/m
723500
I
D 724000
723000

722500
723500

722000 20 cm
Benchmark F Error Ellipses I
New Point E
721500 723000
183000 184000 185000 186000 187000 184500 185000 185500 186000 186500 187000
X/m X/m

Figure 5.18: Network with error ellipses. Figure 5.19: Detailed view: absolute and relative
error ellipses.

Element’s absolute error ellipse

point 𝜙/gon 𝑎/cm 𝑏/cm
G 185.2077 13.147 11.790
H 12.3417 26.717 15.240
I 186.9145 13.623 11.367
Element’s relative error ellipse
leg 𝜙/gon 𝑎/cm 𝑏/cm
G–H 26.3811 24.956 16.044
H–I 19.5521 26.328 15.502
G–I 60.6365 14.447 10.237

Table 5.50: Error ellipse elements.

70
 5.1 A-Model: Adjustment of observation equations

Approximate Point
Approximate Point Adjusted Point
Adjusted Point 725336.7

725139.9
H

725336.5

H
G

Y/m
725139.7
Y/m

725336.3
G
10cm
Error Ellipses
10cm
Error Ellipses

725139.5
725336.1

184868.1 184868.3 184868.5 186579.1 186579.3 186579.5

X/m X/m

Figure 5.20: Detailed view of point G. Figure 5.21: Detailed view of point H.

723322.5 Approximate Point

Adjusted Point

I
723322.3
Y/m

723322.1

10cm
I Error Ellipses

723321.9

185962.9 185963.1 185963.3 185963.5

X/m

Figure 5.22: Detailed view of point I.

71
5 Geomatics examples

5.1.7 Polynomial fit

Observations: 𝑦𝑖 , 𝑖 = 1, . . . , 𝑚.
Given: fixed x-coordinates 𝑥𝑖 , 𝑖 = 1, . . . , 𝑚.
Find parameters 𝑎𝑛 , 𝑛 = 0, . . . , 𝑛 max of fitting polynomial
𝑛Õ
max

𝑓 (𝑥) = 𝑦 = 𝑎𝑛 𝑥 𝑛 .
𝑛=0

Possible additional restrictions:

(a) tangent in (𝑥 T, 𝑦T ) should pass through (𝑥 P, 𝑦P ) or
(b) fitting polynomial should pass through (𝑥 Q, 𝑦Q ) or
(c) unknown coefficient 𝑎𝑘 shall get the numerical value 𝑎˜𝑘 .

Figure 5.23: Fitting polynomials of different degrees.

Observation equation
𝑛Õ
max

𝑦𝑖 = 𝑎𝑛 𝑥 𝑛 + 𝑒𝑖 ,
𝑛=0

𝑦1 = 𝑎 0𝑥 10 + 𝑎 1𝑥 11 + 𝑎 2𝑥 12 + . . . + 𝑒 1,
..
.
0 1 2
𝑦𝑚 = 𝑎 0𝑥𝑚 + 𝑎 1𝑥𝑚 + 𝑎 2𝑥𝑚 + . . . + 𝑒𝑚 .

Vandermonde matrix 𝐴

𝑥 1 · · · 𝑥 1𝑛max
© ª © ª© ª © ª
𝑦1 1 𝑎0 𝑒1
­ 𝑦2 ® ­1 𝑥 2 · · · 𝑥 2𝑛max ® ­ 𝑎 1 ® ­ 𝑒 2 ®
­ . ®=­. ®­ . ®+­ . ® .
­ . ® ­. .. ®­ . ® ­ . ®
­ . ® ­. . ®­ . ® ­ . ®
« 𝑦𝑚 ¬ « 1 𝑥𝑚 · · · 𝑥𝑚 ¬ « 𝑎𝑛max ¬ « 𝑒𝑚 ¬
𝑛 max
|{z} | {z } | {z } |{z}
𝑦 𝐴 𝜉 𝑒

72
 5.1 A-Model: Adjustment of observation equations

1. Adjustment principle 𝑒 T𝑒 −→ min 𝜉.

2. 𝑥-coordinates are error free: inconsistences only for 𝑦𝑖 .
3. 𝑦𝑖 may have matrix 𝑄 𝑦 .
4. The smaller 𝑒ˆT𝑒ˆ for varying 𝑛 max , the better the fit is. However, the larger 𝑛 max , the more the
polynomial oscillates. Using a sufficiently large value for 𝑛 max , even 𝑒ˆT𝑒ˆ = 0 can be achieved.
=⇒ Only low degree polynomials are used.
5. Possible additional restrictions
a) Tangent in (𝑥 T, 𝑦T ), 𝑥 T ∈ 𝑥, shall pass through the point (𝑥 P, 𝑦P ).
Tangent equation:

𝑔(𝑥) = 𝑓 (𝑥 T ) + 𝑓 ′ (𝑥 T ) (𝑥 − 𝑥 T ) =⇒ 𝑦P = 𝑔(𝑥 P ) = 𝑓 (𝑥 T ) + 𝑓 ′ (𝑥 T ) (𝑥 P − 𝑥 T ) .

Example for 𝑛 max = 2

𝑓 (𝑥) = 𝑎 0 + 𝑎 1𝑥 + 𝑎 2𝑥 2 parabola
′
𝑓 (𝑥) = 𝑎 1 + 2𝑎 2𝑥

Tangent in 𝑥 T : 𝑔(𝑥) = 𝑎 0 + 𝑎 1𝑥 T + 𝑎 2𝑥 T2 + (𝑎 1 + 2𝑎 2𝑥 T ) (𝑥 − 𝑥 T )
Tangent in 𝑥 T , passing through 𝑥 P, 𝑦P

𝑦P = 𝑎 0 + 𝑎 1𝑥 T + 𝑎 2𝑥 T2 + (𝑎 1 + 2𝑎 2𝑥𝑇 ) (𝑥 P − 𝑥 T )
= 𝑎 0 + 𝑎 1𝑥 T + 𝑎 2𝑥 T2 + 𝑎 1 (𝑥 P − 𝑥 T ) + 2𝑎 2 (𝑥 P − 𝑥 T )𝑥 T
= 𝑎 0 + 𝑥 P𝑎 1 + 𝑥 T (2𝑥 P − 𝑥 T )𝑎 2
 
=⇒ 𝐵 T 𝜉 = 𝑦P, with 𝜉 = [𝑎 0, 𝑎 1, 𝑎 2 ] T, 𝐵 T = 1 𝑥 P 𝑥 T (2𝑥 P − 𝑥 T ) .

Include restriction using techniques of Lagrange multipliers or eliminate one unknown

coefficient, e. g. 𝑎 0 , in favor of the other unknown coefficients:
𝑎 0 = 𝑦P − 𝑥 P𝑎 1 − 𝑥 T (2𝑥 P − 𝑥 T )𝑎 2 .
General case for a polynomial of degree 𝑛 max
 
𝐵 T = 1 𝑥 P 𝑥 T (2𝑥 P − 𝑥 T ) . . . 𝑥 T𝑛max −1 (𝑛 max𝑥 P − (𝑛 max − 1)𝑥 T )

 T
=⇒ Tangent equation with adjusted parameters 𝜉ˆ = 𝑎ˆ0, . . . , 𝑎ˆ𝑛max

𝑦ˆT − 𝑦P
𝑦 = 𝑎 T𝑥 + 𝑏 T, 𝑎 T := “tangent slope”,
𝑥T − 𝑥P
𝑦ˆT − 𝑦P
𝑏 T := 𝑦ˆT − 𝑥T “axis intercept”,
𝑥T − 𝑥P
𝑦ˆT = 𝑎ˆ0 + 𝑎ˆ1𝑥 T + . . . + 𝑎ˆ𝑛max 𝑥 T𝑛max “estimated ordinate”.

73
5 Geomatics examples

b) adjusted polynomial shall pass through the point (𝑥 Q, 𝑦Q )

𝑛Õ
max  
𝑛
𝑦Q = 𝑎𝑛 𝑥 Q𝑛 =⇒ 𝐵 T 𝜉 = 𝑦Q, 𝐵 T = 1 𝑥 Q 𝑥 Q2 . . . 𝑥 Qmax .
𝑛=0

c) The unknown coefficient 𝑎𝑘 should have the fixed numerical value 𝑎˜𝑘 .
" #
|{z}
T T
0 ... 1 ...
𝐵 𝜉 = 𝑎˜𝑘 , 𝐵 =
position 𝑘 + 1

or eliminate unknown 𝑎𝑘 from 𝜉 by setting it to 𝑎˜𝑘 from the very beginning.

Examples
 T
𝑥𝑖 = −1, 0, 1, 2, 3, 4, 5 ,
 T
𝑦𝑖 = 1.3, 0.8, 0.9, 1.2, 2.0, 3.5, 4.1 .

1) No restrictions: see Fig. 5.24.

2) With tangent restriction: see Fig. 5.25.
3) With point restriction: see Fig. 5.26.
4) With coefficient restriction: see Fig. 5.27.

1st order polynomial: 𝑒ˆT𝑒ˆ = 2.5 2nd order polynomial: 𝑒ˆT𝑒ˆ = 3.4 · 10−1
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

3rd order polynomial: 𝑒ˆT𝑒ˆ = 1.7 · 10−1 4th order polynomial: 𝑒ˆT𝑒ˆ = 7.5 · 10−2
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

5th order polynomial: 𝑒ˆT𝑒ˆ = 8.8 · 10−4 6th order polynomial: 𝑒ˆT𝑒ˆ = 2.4 · 10−29
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

Figure 5.24: Polynomial fit without restrictions.

74
 5.1 A-Model: Adjustment of observation equations

1st order polynomial: 𝑒ˆT𝑒ˆ = 6.3 2nd order polynomial: 𝑒ˆT𝑒ˆ = 4.7 · 10−1
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

3rd order polynomial: 𝑒ˆT𝑒ˆ = 2.1 · 10−1 4th order polynomial: 𝑒ˆT𝑒ˆ = 2.1 · 10−1
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

5th order polynomial: 𝑒ˆT𝑒ˆ = 5.1 · 10−2 6th order polynomial: 𝑒ˆT𝑒ˆ = 1.3 · 10−2
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

Figure 5.25: Polynomial fit with tangent restriction: tangent in 𝑥 T = 1, 𝑦ˆT (𝑥 T ) shall pass through the
point 𝑥 P = 4, 𝑦P = 2.

1st order polynomial: 𝑒ˆT𝑒ˆ = 3.1 2nd order polynomial: 𝑒ˆT𝑒ˆ = 2.9
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

3rd order polynomial: 𝑒ˆT𝑒ˆ = 2.8 4th order polynomial: 𝑒ˆT𝑒ˆ = 2.5
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

5th order polynomial: 𝑒ˆT𝑒ˆ = 1.4 6th order polynomial: 𝑒ˆT𝑒ˆ = 1.3
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

Figure 5.26: Polynomial fit with point restriction: adjusted polynomial shall pass through the point
𝑥 Q = 1.5, 𝑦Q = 2.

75
5 Geomatics examples

1st order polynomial: 𝑒ˆT𝑒ˆ = 1.0 · 101 2nd order polynomial: 𝑒ˆT𝑒ˆ = 3.9 · 10−1
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

3rd order polynomial: 𝑒ˆT𝑒ˆ = 3.5 · 10−1 4th order polynomial: 𝑒ˆT𝑒ˆ = 3.4 · 10−1
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

5th order polynomial: 𝑒ˆT𝑒ˆ = 1.2 · 10−3 6th order polynomial: 𝑒ˆT𝑒ˆ = 3.7 · 10−4
4 4
2 2
0 0
−1 0 1 2 3 4 5 −1 0 1 2 3 4 5

Figure 5.27: Polynomial fit with coefficient restriction: coefficient 𝑎ˆ1 shall vanish, i. e. 𝑎ˆ1 = 0.

More examples: Various straight line fits. For the numerics, the values on page 74 were reused.
1) Straight line fit using A-Model, with inconsistencies 𝑒 𝑦𝑖 in observations 𝑦𝑖 (𝑄 𝑦−1 = 𝐼 ). Observation
equation: 𝑦𝑖 = 𝑎 0 + 𝑎 1𝑥𝑖 .
Results (see also figure 5.28):

𝑎ˆ0 = 0.907, 𝑎ˆ1 = 0.532, 𝑒ˆT 𝑃 𝑒ˆ = 2.505,

 T
𝑦ˆ = 0.375, 0.907, 1.439, 1.971, 2.504, 3.036, 3.568 ,
 T
𝑒ˆ𝑦 = 0.925, −0.107, −0.539, −0.771, −0.504, 0.464, 0.532 .

4
3 data points
adjusted data
𝑦

2
residuals
1
0
−1 0 1 2 3 4 5
𝑥

Figure 5.28: A-model with inconsistencies in 𝑦𝑖 , uniform weights.

76
 5.1 A-Model: Adjustment of observation equations

2) Straight line fit using A-Model, with inconsistencies 𝑒𝑥𝑖 in observations 𝑥𝑖 (𝑄 𝑥−1 = 𝐼 ). Observation
equation: 𝑥𝑖 = 𝑎 0 + 𝑎 1𝑦𝑖 .
Results (see also figure 5.29):

𝑎ˆ0 = −0.815, 𝑎ˆ1 = 1.428, 𝑒ˆT 𝑃 𝑒ˆ = 6.723,

 T
𝑥ˆ = 1.041, 0.327, 0.470, 0.898, 2.041, 4.183, 5.040 ,
 T
𝑒ˆ𝑥 = −2.041, −0.327, 0.530, 1.102, 0.959, −0.183, −0.040 .

4
3 data points
adjusted data
𝑦

2
residuals
1
0
−1 0 1 2 3 4 5
𝑥

Figure 5.29: A-model with inconsistencies in 𝑥𝑖 , uniform weights.

77
5 Geomatics examples

5.2 B-Model: Adjustment of condition equations

5.2.1 Planar triangle 1

Figure 5.30: Triangle observed by angles

Observations: angles 𝛼, 𝛽, 𝛾
Unknowns: inconsistencies 𝑒𝛼 , 𝑒 𝛽 , 𝑒𝛾 =⇒ linear function

𝑓 (𝑒𝛼 , 𝑒 𝛽 , 𝑒𝛾 ) = (𝛼 − 𝑒𝛼 ) + (𝛽 − 𝑒 𝛽 ) + (𝛾 − 𝑒𝛾 ) − 180° = 0 .

Model adjustment condition equations

𝐵 T (𝑦 − 𝑒) − 180° = 𝐵 T𝑦 − 180° − 𝐵 T𝑒 = 𝑤 − 𝐵 T𝑒 = 0
 T  T
with 𝑒 = 𝑒𝛼 , 𝑒 𝛽 , 𝑒𝛾 , 𝑦 = 𝛼, 𝛽, 𝛾 and 𝑤 = 𝐵 T𝑦 − 180° (“misclosure”).

78
 5.2 B-Model: Adjustment of condition equations

5.2.2 Planar triangle 2

Figure 5.31: Triangle observed by angles and distances.

Observations: angles 𝛼, 𝛽, distances 𝑎, 𝑏

Unknowns: inconsistencies 𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽 =⇒ non-linear function 𝑓

𝑓 (𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽 ) = (𝑎 − 𝑒𝑎 ) sin(𝛽 − 𝑒 𝛽 ) − (𝑏 − 𝑒𝑏 ) sin(𝛼 − 𝑒𝛼 ) = 0,

linearized with respect to the “Taylor point” (𝑒𝑎0 , 𝑒𝑏0 , 𝑒𝛼0 , 𝑒 𝛽0 ) =: | 0

  𝜕𝑓
𝜕𝑓

𝑓 𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽 = 𝑓 𝑒𝑎0 , 𝑒𝑏0 , 𝑒𝛼0 , 𝑒 𝛽0 + 𝑒𝑎 − 𝑒𝑎0 + 𝑒𝑏 − 𝑒𝑏0
𝜕𝑒𝑎
0 𝜕𝑒𝑏 0
𝜕𝑓
𝜕𝑓  
!
+ 𝑒𝛼 − 𝑒𝛼0 + 𝑒 𝛽 − 𝑒 𝛽0 = 0,
𝜕𝑒𝛼 0 𝜕𝑒 𝛽 0
 
𝑓 𝑒𝑎0 , 𝑒𝑏0 , 𝑒𝛼0 , 𝑒 𝛽0 = (𝑎 − 𝑒𝑎0 ) sin(𝛽 − 𝑒 𝛽0 ) − (𝑏 − 𝑒𝑏0 ) sin(𝛼 − 𝑒𝛼0 )

= 𝑎 sin(𝛽 − 𝑒 𝛽0 ) − sin(𝛽 − 𝑒 𝛽0 )𝑒𝑎0 − 𝑏 sin(𝛼 − 𝑒𝛼0 ) + sin(𝛼 − 𝑒𝛼0 )𝑒𝑏0 ,

𝜕𝑓


𝑒𝑎 − 𝑒𝑎0 = − sin(𝛽 − 𝑒 𝛽0 ) 𝑒𝑎 − 𝑒𝑎0
𝜕𝑒𝑎 0

= − sin(𝛽 − 𝑒 𝛽0 )𝑒𝑎 + sin(𝛽 − 𝑒 𝛽0 )𝑒𝑎0 ,

𝜕𝑓


𝑒𝑏 − 𝑒𝑏0 = sin(𝛼 − 𝑒𝛼0 ) 𝑒𝑏 − 𝑒𝑏0
𝜕𝑒𝑏 0

= sin(𝛼 − 𝑒𝛼0 )𝑒𝑏 − sin(𝛼 − 𝑒𝛼0 )𝑒𝑏0 ,

𝜕𝑓



𝑒𝛼 − 𝑒𝛼0 = 𝑏 − 𝑒𝑏0 cos(𝛼 − 𝑒𝛼0 ) 𝑒𝛼 − 𝑒𝛼0
𝜕𝑒𝛼



0

= 𝑏 − 𝑒𝑏0 cos(𝛼 − 𝑒𝛼0 )𝑒𝛼 − 𝑏 − 𝑒𝑏0 cos(𝛼 − 𝑒𝛼0 )𝑒𝛼0 ,

𝜕𝑓  
 
𝑒 𝛽 − 𝑒 𝛽0 = − 𝑎 − 𝑒𝑎0 cos(𝛽 − 𝑒 𝛽0 ) 𝑒 𝛽 − 𝑒 𝛽0
𝜕𝑒 𝛽



0

= − 𝑎 − 𝑒𝑎0 cos(𝛽 − 𝑒 𝛽0 )𝑒 𝛽 + 𝑎 − 𝑒𝑎0 cos(𝛽 − 𝑒 𝛽0 )𝑒 𝛽0 .

Model adjustment condition equations

 T
𝑤 − 𝐵 T𝑒 = 0 with 𝑒 = 𝑒𝑎 , 𝑒𝑏 , 𝑒𝛼 , 𝑒 𝛽 .

79
5 Geomatics examples

Collect the coefficients of all terms with 𝑒 in −𝐵 T , all remaining terms go into the vector 𝑤 of mis-
closures.
 
=⇒ 𝐵 T = sin(𝛽 − 𝑒 𝛽0 ), − sin(𝛼 − 𝑒𝛼0 ), −(𝑏 − 𝑒𝑏0 ) cos(𝛼 − 𝑒𝛼0 ), (𝑎 − 𝑒𝑎0 ) cos(𝛽 − 𝑒 𝛽0 ) ,
 


𝑤 = 𝑎 sin 𝛽 − 𝑒 𝛽0 − 𝑏 sin 𝛼 − 𝑒𝛼0 − 𝑏 − 𝑒𝑏0 cos(𝛼 − 𝑒𝛼0 )𝑒𝛼0


+ 𝑎 − 𝑒𝑎0 cos(𝛽 − 𝑒 𝛽0 )𝑒 𝛽0 (“misclosure”)

Example: observations
𝑎 = 10, 𝑏 = 5, 𝛼 = 60°, 𝛽 = 23.7°
with associated weights

𝑃𝑎 = 1, 𝑃𝑏 = 0.1, 𝑃𝛼 = 1°, 𝑃 𝛽 = 0.2° .

Initial approximate values for unknown inconsistencies

𝑒𝑎0 = 𝑒𝑏0 = 0, 𝑒𝛼0 = 𝑒 𝛽0 = 0° .

c k < 10−12 )
Results: parameters (after 6 iterations, k Δ𝑒

𝑒ˆ𝑎 = −5.63 · 10−6, 𝑒ˆ𝑏 = 1.13 · 10−4, 𝑒ˆ𝛼 = 6′26.241′′, 𝑒ˆ𝛽 = −1°55′42.492′′,
𝑒ˆT 𝑃 𝑒ˆ = 4.017 · 10−6 .

80
 5.3 Mixed model

5.3 Mixed model

5.3.1 Straight line fit using A-model with pseudo observation equations

Example: Straight line fit using A-Model, with inconsistencies 𝑒𝑥𝑖 and 𝑒 𝑦𝑖 in both observations 𝑥𝑖
and 𝑦𝑖 (𝑄 𝑦−1 = 𝑄 𝑥−1 = 𝐼 , 𝑃 = diag(𝑄 𝑦−1, 𝑄 𝑥−1 )). For the numerics, the values on page 74 have been used.
Unknown parameters 𝑎 0 , 𝑎 1 , 𝑥¯𝑖 , 𝑖 = 1, . . . , 𝑚

𝑦𝑖 − 𝑒 𝑦𝑖 = 𝑎 0 + 𝑎 1 (𝑥𝑖 − 𝑒𝑥𝑖 ) = 𝑎 0 + 𝑎 1𝑥¯𝑖 (5.1)

𝑥𝑖 − 𝑒𝑥𝑖 = 𝑥¯𝑖 (5.2)

Approximate values: 𝑎 0 = 𝑎 00 + Δ𝑎 0 , 𝑎 1 = 𝑎 01 + Δ𝑎 1 , 𝑥¯𝑖 = 𝑥¯𝑖 0 + Δ𝑥¯𝑖 .

Linearized equations 5.1 and 5.2:

𝑦𝑖 − (𝑎 00 + 𝑎 01𝑥¯𝑖0 ) −𝑒 𝑦𝑖 = Δ𝑎 0 + 𝑎 01 Δ𝑥¯𝑖 + 𝑥¯𝑖0 Δ𝑎 1

| {z }
Δ𝑦𝑖

𝑥𝑖 − 𝑥¯𝑖0 − 𝑒𝑥𝑖 = Δ𝑥¯𝑖 .

This leads to
Δ𝑦𝑖 − 𝑒 𝑦𝑖 = Δ𝑎 0 + 𝑎 01 Δ𝑥¯𝑖 + 𝑥¯𝑖0 Δ𝑎 1

and
Δ𝑥𝑖 − 𝑒𝑥𝑖 = Δ𝑥¯𝑖 .

In matrix notation:
    Δ𝑎 0  
Δ𝑦 − 𝑒 𝑦 𝐴1 ©­ ª
® 𝐴1
𝐴2 ­
= Δ𝑎 1 ® = Δ𝜉
Δ𝑥 − 𝑒𝑥 𝐴2
2𝑚×(𝑚+2) «
2𝑚×1
Δ𝑥¯ ¬
(𝑚+2)×1

where
𝑥¯10 𝑎 01
© ª © ª
1 0... 0 0 0 1 0 ... 0
­1 𝑥¯20 𝑎 01 ® ­0 0®
𝐴1 = ­­ . ®; 𝐴2 = ­­ . .. ®® .
0 ... 0 0 0... 1
.. .. .. ..® .. .. ..
­ .. ® ­ .. .®
.. ..
𝑚×(𝑚+2) . . .. . 𝑚×(𝑚+2) . . . .
«1 𝑎1 ¬ «0 1¬
0
𝑥¯𝑚 0 0 ... 0 0 0 0 ...

Results: Initial approximate values for unknown parameters:

𝑎 00 = 0.800, 𝑎 01 = 0.550, 𝑥¯𝑖0 = 𝑥𝑖 .

c k < 10−12 ):
Parameters (after 20 iterations, k Δ𝜉

𝑎ˆ0 = 0.829, 𝑎ˆ1 = 0.571, 𝑒ˆT 𝑃 𝑒ˆ = 1.921

81
5 Geomatics examples

 T
𝑦ˆ = 0.514, 0.822, 1.277, 1.782, 2.409, 3.209, 3.787
 T
𝑒ˆ𝑦 = 0.786, −0.022, −0.377, −0.582, −0.409, 0.291, 0.313
 T
𝑥ˆ = −0.551, −0.012, 0.785, 1.668, 2.766, 4.166, 5.179
 T
𝑒ˆ𝑥 = −0.449, 0.012, 0.215, 0.332, 0.234, −0.166, −0.179

See figure 5.32.

4
3 data points
adjusted data
𝑦

2
residuals
1
0
−1 0 1 2 3 4 5
𝑥

Figure 5.32: A-model with inconsistencies in 𝑥𝑖 and 𝑦𝑖 , uniform weights.

82
 5.3 Mixed model

5.3.2 Straight line fit using extended B-Model

Example: Straight line fit using extended B-Model with 𝑄 𝑦−1 = 𝑄 𝑥−1 = 𝐼 , 𝑃 = diag(𝑄 𝑦−1, 𝑄 𝑥−1 ).
Non linear condition equation with unknowns 𝑎 0 , 𝑎 1 :


𝑦𝑖 − 𝑒 𝑦𝑖 − 𝑎 0 + 𝑎 1 (𝑥𝑖 − 𝑒𝑥𝑖 ) = 0

Initial approximate values 𝑒𝑥0𝑖 = 0, 𝑒 𝑦0 𝑖 = 0, 𝑎 00 , 𝑎 01 so that

𝑒𝑥𝑖 = 𝑒𝑥0𝑖 + Δ𝑒𝑥𝑖 , 𝑒 𝑦𝑖 = 𝑒 𝑦0 𝑖 + Δ𝑒 𝑦𝑖 , 𝑎 0 = 𝑎 00 + Δ𝑎 0, 𝑎 1 = 𝑎 01 + Δ𝑎 1 .


𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑎 00 + 𝑎 01 (𝑥𝑖 − 𝑒𝑥0𝑖 ) − Δ𝑒 𝑦𝑖 − Δ𝑎 0 − Δ𝑎 1 (𝑥𝑖 − 𝑒𝑥0𝑖 ) + 𝑎 01 Δ𝑒𝑥𝑖 = 0
 
0 0 0 0
0
  Δ𝑎 0
𝑦𝑖 − 𝑒 𝑦𝑖 − 𝑎 0 + 𝑎 1 (𝑥𝑖 − 𝑒𝑥𝑖 ) − 𝑒 𝑦𝑖 + 𝑒 𝑦𝑖 − 1 (𝑥𝑖 − 𝑒𝑥𝑖 ) 0 + 𝑎 01𝑒𝑥𝑖 − 𝑎 01𝑒𝑥0𝑖 = 0
Δ𝑎 1
   
0 0
 0
 Δ𝑎 0  0  𝑒𝑥𝑖
𝑦𝑖 − (𝑎 0 + 𝑎 1𝑥𝑖 ) − 1 (𝑥𝑖 − 𝑒𝑥𝑖 ) + 𝑎 1 −1 =0
| {z }| {z } Δ𝑎 1 | {z } 𝑒 𝑦𝑖
𝑤𝑖 𝐴
| {z } T
|{z}
𝑖 𝐵𝑖
Δ𝜉 𝑒𝑖

 1 𝑥 1 − 𝑒𝑥0 
 1 
   
 1 𝑥2 − 𝑒 0   T
 𝑥2 
𝐴 =−. ;
Δ𝑎 0 𝑒
𝑒 = 𝑒𝑥 1 · · · 𝑒𝑥𝑚 𝑒 𝑦1 · · · 𝑒 𝑦𝑚 = 𝑥 ;
. 
. Δ𝜉 = ;
 
. .
. Δ𝑎 1 𝑒𝑦
 1 𝑥𝑚 − 𝑒 0 
𝑚×2 2×1 2𝑚×1

 𝑥𝑚 
𝑎 0 
 1 −1   
 
𝐵T =  . . . ..  = 𝑎 01 𝐼𝑚 , −𝐼𝑚 ; 𝑤 = 𝑦 − (𝑎 00 + 𝑎 01𝑥) .
 . 
 −1

𝑚×2𝑚 𝑚×1
𝑎 01
Lagrangian:
1
L (Δ𝜉, 𝑒, 𝜆) = 𝑒 T 𝑃𝑒 + 𝜆 T (𝑤 + 𝐴Δ𝜉 + 𝐵 T𝑒) −→ min
2 Δ𝜉,𝑒,𝜆

𝜕L
(𝑒,
ˆ 𝜆, c = 𝑃 𝑒ˆ
ˆ Δ𝜉) + 𝐵 𝜆ˆ = 0
𝜕𝑒 2𝑚×2𝑚 2𝑚×1 2𝑚×𝑚 𝑚×1 2𝑚×1

𝜕L
(𝑒,
ˆ 𝜆, c =
ˆ Δ𝜉) 𝐴T 𝜆ˆ = 0
𝜕Δ𝜉 2×𝑚 𝑚×1 2×1

𝜕L
(𝑒, c = 𝐵 T 𝑒ˆ + 𝐴 Δ𝜉
ˆ Δ𝜉)
ˆ 𝜆, c = −𝑤
𝜕𝜆 𝑚×2𝑚 2𝑚×1 𝑚×2 2×1 𝑚×1

Results (see also figure 5.33):

𝑎ˆ0 = 0.829, 𝑎ˆ1 = 0.571, 𝑒ˆT 𝑃 𝑒ˆ = 1.921
 T
𝑦ˆ = 0.514, 0.822, 1.277, 1.782, 2.409, 3.209, 3.787
 T
𝑒ˆ𝑦 = 0.786, −0.022, −0.377, −0.582, −0.409, 0.291, 0.313
 T
𝑥ˆ = −0.551, −0.012, 0.785, 1.668, 2.766, 4.166, 5.179
 T
𝑒ˆ𝑥 = −0.449, 0.012, 0.215, 0.332, 0.234, −0.166, −0.179

83
5 Geomatics examples

4
3 data points
adjusted data

𝑦
2
residuals
1
0
−1 0 1 2 3 4 5
𝑥

Figure 5.33: Extended B-model with inconsistencies in 𝑥𝑖 and 𝑦𝑖 , uniform weights.

Example: The following results and figure show the cases for the previous two examples, ob-
servations having weights (𝑃𝑥 = 𝑄 𝑥−1 ≠ 𝐼 , 𝑃 𝑦 = 𝑄 𝑦−1 ≠ 𝐼 , 𝑃 = diag (𝑃𝑥 , 𝑃 𝑦 )). We introduce the
weights
 T
diag 𝑃𝑥 = 3, 9, 8, 4, 5, 7, 10
 T
diag 𝑃 𝑦 = 2, 8, 7, 5, 10, 8, 6

Both, the A-model with inconsistencies 𝑒𝑥𝑖 and 𝑒 𝑦𝑖 and the extended B-model, give identical results.
Due to 𝑃 ≠ 𝐼 residuals are not orthogonal to the adjusted line. See figure 5.34.

𝑎ˆ0 = 0.5512, 𝑎ˆ1 = 0.6580, 𝑒ˆT 𝑃 𝑒ˆ = 7.6931

 T
𝑦ˆ = 0.208, 0.620, 1.124, 1.633, 2.281, 3.288, 3.895
 T
𝑒ˆ𝑦 = 1.092, 0.180, −0.224, −0.433, −0.281, 0.212, 0.205
 T
𝑥ˆ = −0.521, 0.105, 0.871, 1.644, 2.630, 4.159, 5.081
 T
𝑒ˆ𝑥 = −0.479, −0.105, 0.129, 0.356, 0.370, −0.159, −0.081

4
3 data points
adjusted data
𝑦

2
residuals
1
0
−1 0 1 2 3 4 5
𝑥

Figure 5.34: Extended B-model with inconsistencies in 𝑥𝑖 and 𝑦𝑖 , non-uniform weights.

84
 5.3 Mixed model

5.3.3 2D Similarity Transformation

The following two tables (see Niemeier, 2008, pg. 374–375) give coordinates with respect to the
source (𝑢, 𝑣)-system and the target (𝑥, 𝑦)-system. Points 1–4 are identical to both systems (con-
trol points). We assume inconsistencies in both source and target system coordinates and they are
uncorrelated having equal unit variances, i. e.

𝑃𝑥 = 𝑄 𝑥−1 = 𝐼, 𝑃 𝑦 = 𝑄 𝑦−1 = 𝐼, 𝑃𝑢 = 𝑄𝑢−1 = 𝐼, 𝑃 𝑣 = 𝑄 𝑣−1 = 𝐼, 𝑃 = diag(𝑃𝑥 , 𝑃 𝑦 , 𝑃𝑢 , 𝑃 𝑣 ) .

Table 5.51: Source coordinates. Table 5.52: Target coordinates.

Point 𝑢/m 𝑣/m Point 𝑥/m 𝑦/m
1 14 029.640 12 786.840 1 19 405.518 23 159.823
2 14 914.630 12 535.560 2 20 291.232 22 909.817
3 14 771.830 11 404.660 3 20 150.035 21 778.202
4 13 221.620 11 840.320 4 18 598.550 22 211.755
13 14 735.090 12 127.380
14 14 253.840 11 923.950
15 13 603.740 11 836.700
16 14 291.760 12 495.310
17 13 931.500 12 307.610

A 2D similarity transformation (Helmert transformation) to transform the set of coordinates from

the source system to the target system will be performed.
       
𝑥𝑖 cos 𝛼 sin 𝛼 𝑢𝑖 𝑡
=𝜆 + 𝑥
𝑦𝑖 − sin 𝛼 cos 𝛼 𝑣𝑖 𝑡𝑦
|{z} |{z}
target source

Usual adjustment: only target coordinates 𝑥𝑖 and 𝑦𝑖 have inconsistencies

𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆𝑢𝑖 cos 𝛼 + 𝜆𝑣𝑖 sin 𝛼 + 𝑡𝑥 ,

𝑦𝑖 − 𝑒 𝑦𝑖 = −𝜆𝑢𝑖 sin 𝛼 + 𝜆𝑣𝑖 cos 𝛼 + 𝑡 𝑦 .

Mixed model approach I: A-model with inconsistencies in both [𝑥𝑖 , 𝑦𝑖 ] and [𝑢𝑖 , 𝑣𝑖 ] coordinates
(𝑖 = 1, . . . , 𝑝, with 𝑝 number of control points).

𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆𝑢¯𝑖 cos 𝛼 + 𝜆𝑣¯𝑖 sin 𝛼 + 𝑡𝑥 ,

𝑦𝑖 − 𝑒 𝑦𝑖 = −𝜆𝑢¯𝑖 sin 𝛼 + 𝜆𝑣¯𝑖 cos 𝛼 + 𝑡 𝑦 ,
𝑢𝑖 − 𝑒𝑢𝑖 = 𝑢¯𝑖 ,
𝑣𝑖 − 𝑒 𝑣𝑖 = 𝑣¯𝑖 .

85
5 Geomatics examples

Approximate values:

𝜆 = 𝜆 0 + Δ𝜆, 𝛼 = 𝛼 0 + Δ𝛼, 𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 ,

𝑢¯𝑖 = 𝑢¯𝑖0 + Δ𝑢¯𝑖 , 𝑣¯𝑖 = 𝑣¯𝑖0 + Δ¯𝑣𝑖 .

Linearization process:

𝑥𝑖 − 𝑒𝑥𝑖 = (𝜆 0 cos 𝛼 0𝑢¯𝑖0 + 𝜆 0 sin 𝛼 0𝑣¯𝑖0 + 𝑡𝑥0 ) +Δ𝑡𝑥 + (−𝜆 0 sin 𝛼 0𝑢¯𝑖0 + 𝜆 0 cos 𝛼 0𝑣¯𝑖0 ) Δ𝛼
| {z } | {z }
𝑥𝑖0 𝑎𝑖

+ (cos 𝛼 0𝑢¯𝑖0 + sin 𝛼 0𝑣¯𝑖0 ) Δ𝜆 + 𝜆 0 cos 𝛼 0 Δ𝑢¯𝑖 + 𝜆 0 sin 𝛼 0 Δ¯𝑣𝑖

| {z }
𝑏𝑖

𝑦𝑖 − 𝑒 𝑦𝑖 = (−𝜆 0 sin 𝛼 0𝑢¯𝑖0 + 𝜆 0 cos 𝛼 𝑣¯𝑖0 + 𝑡 𝑦0 ) +Δ𝑡 𝑦 −(𝜆 0 cos 𝛼 0𝑢¯𝑖0 + 𝜆 0 sin 𝛼 0𝑣¯𝑖0 ) Δ𝛼
| {z } | {z }
𝑦𝑖0 𝑐𝑖

+ (− sin 𝛼 0𝑢¯𝑖0 + cos 𝛼 0𝑣¯𝑖0 ) Δ𝜆 − 𝜆 0 sin 𝛼 0 Δ𝑢¯𝑖 + 𝜆 0 cos 𝛼 0 Δ¯𝑣𝑖

| {z }
𝑑𝑖

𝑢𝑖 − 𝑒𝑢𝑖 = 𝑢¯𝑖0 + Δ𝑢¯𝑖0

𝑣𝑖 − 𝑒 𝑣𝑖 = 𝑣¯𝑖0 + Δ¯𝑣𝑖0

In matrix form:
𝑥1 − 𝑥 0   𝑒𝑥 1  1 
 1    0 𝑎1 𝑏 1 𝜆 0 cos 𝛼 0 . . . 0 𝜆 0 sin 𝛼 0 . . . 0 
 .   ..   .. 
 ..   .   .
.. .. .. .. .. .. .. .. ..

    
. .

𝑥 − 𝑥 0 
. . . . . . .
 𝑝 𝑝
𝑒 𝑥 𝑝  1 𝜆 sin 𝛼 
   . . . 𝜆 0 cos 𝛼 0 0 0
 
0 𝑎𝑝 𝑏 𝑝 0 0 ...
       Δ𝑡𝑥 
       
𝑦1 − 𝑦1  𝑒 𝑦  0   Δ𝑡 𝑦 
 1    
0
 . 
1 𝑐1 𝑑1 −𝜆 0 sin 𝛼 0 ... 0 𝜆 0 cos 𝛼 0 ... 0
 .   .   .   Δ𝛼 
 .   ..   .. .. .. .. .. .. .. .. .. ..   
     . . . . . . . . .   Δ𝜆 
𝑦𝑝 − 𝑦𝑝0  𝑒  0 𝜆 cos 𝛼   
   𝑦𝑝   . . . −𝜆 0 sin 𝛼 0 0 0
 
 −   =    Δ𝑢¯1 
1 𝑐 𝑝 𝑑𝑝 0 0 ...
    
       ... 
𝑢 1 − 𝑢¯0   𝑒𝑢1  0   
 1  .   .
0 0 0 1 ... 0 0 ... 0
 Δ𝑢¯𝑝 
 .   .   .   
 ..   .   .
.... .. .. .. .. .. .. ..
  Δ¯𝑣 1 
    
. . . . . . . . .
  
𝑢 − 𝑢¯0  𝑒𝑢𝑝  0   ... 
 𝑝     0 0 0 0 ... 1 0 ... 0   
       Δ¯𝑣 𝑝 
𝑝
       
   𝑒𝑣  0  |{z}
 𝑣 1 − 𝑣¯10   1  
 . 
0 0 0 0 ... 0 1 ... 0
 .   .   . 
 .   ..   .. .... .. .. .. .. .. 
.. .. Δ𝜉
     . . . . . . . . . 
𝑒  0 
(2𝑝+4)×1
 𝑣 𝑝 − 𝑣¯𝑝   𝑣𝑝   
 
0 0 0 0 0 ... 0 0 ... 1
| {z } |{z} | {z }
𝑙 𝑒 𝐴
4𝑝×1 4𝑝×1 4𝑝×(2𝑝+4)

Results: by using the initial approximate values for unknown parameters

   
𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝛼 0 = 1.5′′, 𝜆 0 = 1, 𝑢¯0 = 𝑢 1, . . . , 𝑢 4 T, 𝑣¯0 = 𝑣 1, . . . , 𝑣 4 T,

86
 5.3 Mixed model

c k < 10−11 ):
we obtain the parameters (after 5 Iterations, k Δ𝜉

𝑡ˆ𝑥 = 5389.091 m, 𝑡ˆ𝑦 = 10 347.006 m, 𝛼ˆ = −5′5.557′′, 𝜆ˆ = 1.000 409 017,

𝑒ˆT 𝑃 𝑒ˆ = 0.001 284 79 m2 .

Coordinates of data points in the target system are listed in Tab. 5.53.

Table 5.53: Coordinates of data points in the target system.

Point 𝑥/m 𝑦/m
13 20 112.219 22 501.170
14 19 631.075 22 296.944
15 18 980.839 22 208.695
16 19 668.163 22 868.593
17 19 308.035 22 680.283

source system target system

13 000 23 500
1
12 500 23 000
1
16 2
16 2
17
13 17
12 000 22 500 13
𝑦/m
𝑣/m

4 14
15 14
4 15
11 500 22 000
3
3
11 000 21 500
control points control points
source data source data
13 000 13 500 14 000 14 500 15 000 18 500 19 000 19 500 20 000 20 500
𝑢/m 𝑥/m

Figure 5.35: 2D similarity transformation: Gauss Markov model; inconsistencies in both source and
target system.

Mixed model approach II: extended B-model with inconsistencies in both [𝑥𝑖 , 𝑦𝑖 ] and [𝑢𝑖 , 𝑣𝑖 ]
coordinates (𝑖 = 1, . . . , 𝑝, with 𝑝 number of control points).

𝑓𝑥𝑖 := 𝑥𝑖 − 𝑒𝑥𝑖 − 𝜆 cos 𝛼 (𝑢𝑖 − 𝑒𝑢𝑖 ) + 𝜆 sin 𝛼 (𝑣𝑖 − 𝑒 𝑣𝑖 ) + 𝑡𝑥 = 0,


𝑓𝑦𝑖 := 𝑦𝑖 − 𝑒 𝑦𝑖 − −𝜆 sin 𝛼 (𝑢𝑖 − 𝑒𝑢𝑖 ) + 𝜆 cos 𝛼 (𝑣𝑖 − 𝑒 𝑣𝑖 ) + 𝑡 𝑦 = 0 .

87
5 Geomatics examples

Linearization using Taylor point 𝑒𝑥0𝑖 , 𝑒 𝑦0 𝑖 , 𝑒𝑢0𝑖 , 𝑒 𝑣0𝑖 , 𝑡𝑥0 , 𝑡 𝑦0 , 𝛼 0 , 𝜆 0 so that

𝑒𝑥𝑖 = 𝑒𝑥0𝑖 + Δ𝑒𝑥𝑖 , 𝑒 𝑦𝑖 = 𝑒 𝑦0 𝑖 + Δ𝑒 𝑦𝑖 , 𝑒𝑢𝑖 = 𝑒𝑢0𝑖 + Δ𝑒𝑢𝑖 , 𝑒 𝑣𝑖 = 𝑒 𝑣0𝑖 + Δ𝑒 𝑣𝑖 ,

𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 , 𝛼 = 𝛼 0 + Δ𝛼, 𝜆 = 𝜆 0 + Δ𝜆 .

" #    Δ𝑒𝑥𝑖   Δ𝑡𝑥 

       
 Δ𝑒 𝑦𝑖     Δ𝑡 𝑦 
𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖
𝑓𝑥0𝑖    +    
+  
𝜕𝑒𝑥𝑖 𝜕𝑒 𝑦𝑖 𝜕𝑒𝑢𝑖 𝜕𝑒 𝑣𝑖 𝜕𝑡𝑥 𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆
   Δ𝑒𝑢     Δ𝛼  = 0
𝑓𝑦0𝑖 
𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
    𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
  
   Δ𝑒     Δ𝜆 
𝑖

 𝑣𝑖   
𝜕𝑒𝑥𝑖 𝜕𝑒 𝑦𝑖 𝜕𝑒𝑢𝑖 𝜕𝑒 𝑣𝑖 𝜕𝑡𝑥 𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆 0
0

where

𝑓𝑥0𝑖 = 𝑥𝑖 − 𝑒𝑥0𝑖 − 𝜆 0 cos 𝛼 0 (𝑢𝑖 − 𝑒𝑢0𝑖 ) + 𝜆 0 sin 𝛼 0 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) + 𝑡𝑥0
 
𝑓𝑦0𝑖 = 𝑦𝑖 − 𝑒 𝑦0 𝑖 − −𝜆 0 sin 𝛼 0 (𝑢𝑖 − 𝑒𝑢0𝑖 ) + 𝜆 0 cos 𝛼 0 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) + 𝑡 𝑦0 .

First, we replace Δ𝑒𝑥𝑖 = 𝑒𝑥𝑖 − 𝑒𝑥0𝑖 , Δ𝑒 𝑦𝑖 = 𝑒 𝑦𝑖 − 𝑒 𝑦0 𝑖 etc. and get

" #  𝜕𝑓𝑥𝑖   𝑒𝑥𝑖   Δ𝑡𝑥 

     
− 𝑒𝑥0𝑖
 𝜕𝑒𝑥   𝑒 𝑦𝑖     Δ𝑡 𝑦 
𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖
𝑓𝑥0𝑖  𝑖   − 𝑒 𝑦0 𝑖 +    
+  
𝜕𝑒 𝑦𝑖 𝜕𝑒𝑢𝑖 𝜕𝑒 𝑣𝑖 𝜕𝑡𝑥 𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆
 𝜕𝑓𝑦𝑖   𝑒𝑢     Δ𝛼  = 0 .
𝑓𝑦0𝑖  𝜕𝑒𝑥𝑖
𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
  𝑖 − 𝑒𝑢0𝑖   𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
  
  0  𝑒 𝑣𝑖  

  Δ𝜆 
 
𝜕𝑒 𝑦𝑖 𝜕𝑒𝑢𝑖 𝜕𝑒 𝑣𝑖 𝜕𝑡𝑥 𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆
| {z } − 𝑒 𝑣0𝑖 0

𝐵𝑖 T

Then, we rearrange the equation to

" #  𝑒𝑥0𝑖   𝑒𝑥𝑖   Δ𝑡𝑥 

 0     𝜕𝑓𝑥𝑖   
 𝑒𝑦   𝑒 𝑦𝑖   𝜕𝑡𝑥   Δ𝑡 𝑦 
𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖 𝜕𝑓𝑥𝑖
𝑓𝑥0𝑖
 𝑖  + 𝐵𝑖 T      
𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆
− 𝐵𝑖 T  𝑒0   𝑒𝑢  +  𝜕𝑓𝑦𝑖   Δ𝛼  = 0 .
𝑓𝑦0𝑖  𝑢𝑖   𝑖  𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
  
|{z}  𝑒0  𝑒   𝜕𝑡𝑥  0  Δ𝜆 
2×4 2×4

 𝑣𝑖   𝑣𝑖  | {z }  
𝜕𝑡 𝑦 𝜕𝛼 𝜕𝜆

𝑓0 |{z} |{z} | {z }
𝐴𝑖
2×1 𝑒𝑖0 𝑒𝑖 2×4 Δ𝜉

| {z }
4×1 4×1
4×1

𝑤𝑖
2×1

88
 5.3 Mixed model

In matrix notation:
 𝑒𝑥 1 
 
 .. 
 . 
 
 𝑒𝑥 
 𝑥1 − 𝑥 0   𝑝  −1 0 𝜆 0𝑎 0 −𝑏 10 
 1  𝑒  
   𝑦1   . . .. 
1
 ..   .   .. .. ..
 .   ..   . . 
 Δ𝑡𝑥 
 𝑥 − 𝑥0   −𝐼𝑝 0𝑝 𝜆 0 cos 𝛼 0 𝐼𝑝 𝜆 0 sin 𝛼 0 𝐼𝑝     −1 0 𝜆 0𝑎 0 −𝑏 𝑝0   
 𝑝 𝑝        Δ𝑡 𝑦 
   𝑒 𝑦𝑝   
 +    
𝑝
    +    Δ𝛼  = 0,
     𝑒𝑢1     
 𝑦1 − 𝑦1 
0
 0𝑝 −𝐼𝑝 −𝜆 sin 𝛼 𝐼𝑝 𝜆 cos 𝛼 𝐼𝑝 
0 0 0 0  .   0 −1 𝜆 0𝑏 10 𝑎 01   Δ𝜆 
  | {z }  .   . . ..   
   .   . . | {z }
     . . . 
.. ..
   𝑒𝑢𝑝   
. .
 𝑦𝑝 − 𝑦𝑝0     0 −1 𝜆 0𝑏 𝑝0 𝑎𝑝0 
𝐵T
   𝑒 𝑣1  
2𝑝×4𝑝 Δ𝜉
| {z }   | {z }
 .. 
4×1

 . 
 
𝑤 𝐴
2𝑝×1
 𝑒𝑣  2𝑝×4
 𝑝
| {z }
𝑒
4𝑝×1

where 𝐼𝑝 is the unit matrix of size 𝑝 × 𝑝 and 0𝑝 the zero matrix of the same size. Additionally, we
define 𝑢¯𝑖0 = 𝑢𝑖 − 𝑒𝑢0𝑖 and 𝑣¯𝑖0 = 𝑣𝑖 − 𝑒 𝑣0𝑖 to get the abbreviations

𝑥𝑖0 = 𝜆 0 (cos 𝛼 0𝑢𝑖 + sin 𝛼 0𝑣𝑖 ) + 𝑡𝑥0 , 𝑦𝑖0 = 𝜆 0 (− sin 𝛼 0𝑢𝑖 + cos 𝛼 0𝑣𝑖 ) + 𝑡 𝑦0 ,
𝑎𝑖0 = sin 𝛼 0𝑢¯𝑖0 − cos 𝛼 0𝑣¯𝑖0, 𝑏𝑖0 = cos 𝛼 0𝑢¯𝑖0 + sin 𝛼 0𝑣¯𝑖0 .

Results: by using the following initial approximate values for unknown parameters

𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝜆 0 = 1, 𝛼 0 = 1.5′′, 𝑒𝑢0𝑖 = 𝑒 𝑣0𝑖 = 0 ∀𝑖,

c k < 10−11 ):
we get the parameters (after 7 iterations, k Δ𝜉

𝑡ˆ𝑥 = 5389.091 m, 𝑡ˆ𝑦 = 10 347.006 m, 𝛼ˆ = −5′5.557′′, 𝜆ˆ = 1.000 409 017,

𝑒ˆT 𝑃 𝑒ˆ = 0.001 284 79 m2 .

89
5 Geomatics examples

5.3.4 2D Affine Transformation Model I

Example: 6-parameter affine transformation – model I

The numerical data on page 85 (from Niemeier, 2008, pg. 374–375) are transformed using the 6-
parameter affine transformtion
          
𝑥𝑖 𝜆1 0 1𝑘 cos 𝛼 sin 𝛼 𝑢𝑖 𝑡
= + 𝑥 .
𝑦𝑖 0 𝜆2 0 1 − sin 𝛼 cos 𝛼 𝑣𝑖 𝑡𝑦
|{z} | {z } |{z} |{z}
target scale factors shear source

Mixed model approach I: A-model with inconsistencies in both (𝑥𝑖 , 𝑦𝑖 ) and (𝑢𝑖 , 𝑣𝑖 ) coordinates.

𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆1𝑢¯𝑖 (cos 𝛼 − 𝑘 sin 𝛼) + 𝜆1𝑣¯𝑖 (sin 𝛼 + 𝑘 cos 𝛼) + 𝑡𝑥 ,

𝑦𝑖 − 𝑒 𝑦𝑖 = −𝜆2𝑢¯𝑖 sin 𝛼 + 𝜆2𝑣¯𝑖 cos 𝛼 + 𝑡 𝑦 ,
𝑢𝑖 − 𝑒𝑢𝑖 = 𝑢¯𝑖 ,
𝑣𝑖 − 𝑒 𝑣𝑖 = 𝑣¯𝑖 .

Approximate values:

𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝛼 = 𝛼 0 + Δ𝛼, 𝑢¯𝑖 = 𝑢¯𝑖0 + Δ𝑢¯𝑖 , 𝜆1 = 𝜆10 + Δ𝜆1,

𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 , 𝑘 = 𝑘 0 + Δ𝑘, 𝑣¯𝑖 = 𝑣¯𝑖0 + Δ¯𝑣𝑖 𝜆2 = 𝜆20 + Δ𝜆2, .

Linearization:


𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆10𝑢¯𝑖0 (cos 𝛼 0 − 𝑘 0 sin 𝛼 0 ) + 𝜆10𝑣¯𝑖0 (sin 𝛼 0 + 𝑘 0 cos 𝛼 0 ) + 𝑡𝑥0 +Δ𝑡𝑥
| {z }
𝑥𝑖0


+ −𝜆10𝑢¯𝑖0 (sin 𝛼 0 + 𝑘 0 cos 𝛼 0 ) + 𝜆10𝑣¯𝑖0 (cos 𝛼 0 − 𝑘 0 sin 𝛼 0 ) Δ𝛼
| {z }
𝑎𝑖



+ 𝑢¯𝑖0 (cos 𝛼 0 − 𝑘 sin 𝛼 ) + 𝑣¯𝑖0 (sin 𝛼 0 + 𝑘 0 cos 𝛼 0 ) Δ𝜆1 + −𝜆10𝑢¯𝑖0 sin 𝛼 0 + 𝜆10𝑣¯𝑖0 cos 𝛼 0 Δ𝑘
0 0
| {z } | {z }
𝑏𝑖 𝑓𝑖

+ 𝜆10 (cos 𝛼 0 − 𝑘 0 sin 𝛼 0 ) Δ𝑢¯𝑖 + 𝜆10 (sin 𝛼 0 + 𝑘 0 cos 𝛼 0 ) Δ¯𝑣𝑖 ,

| {z } | {z }
𝑔 ℎ
 

𝑦𝑖 − 𝑒 𝑦𝑖 = −𝜆20𝑢¯𝑖0 sin 𝛼 0 + 𝜆20𝑣¯𝑖0 cos 𝛼 0 + 𝑡 𝑦0 +Δ𝑡 𝑦 + −𝜆20 (𝑢¯𝑖0 cos 𝛼 0 + 𝑣¯10 sin 𝛼 0 ) Δ𝛼
| {z } | {z }
𝑐𝑖
𝑦𝑖0

+ (−𝑢¯𝑖0 sin 𝛼 0 + 𝑣¯𝑖0 cos 𝛼 0 ) Δ𝜆2 −𝜆20 sin 𝛼 0 Δ𝑢¯𝑖 + 𝜆20 cos 𝛼 0 Δ¯𝑣𝑖 .
| {z } | {z } | {z }
𝑑𝑖 𝑞 𝑟

90
 5.3 Mixed model

In matrix notation:
 𝑥1 − 𝑥 0   𝑒𝑥 1  1 
 1     0 𝑎1 𝑏1 0 𝑓1
𝑔 ... 0 ℎ ... 0 
   ..   .. 
 ..     .
.. .. .. .. ..
.. . . .... . . ..

 .     . . . .   Δ𝑡𝑥 
 𝑥 − 𝑥0 
. . . . . . . .
 𝑝 𝑝 
 𝑒𝑥𝑝  1 ℎ   
    0 𝑎𝑝 𝑏 𝑝 0 𝑓𝑝 0 . . . 𝑔 0 ...  Δ𝑡 𝑦 
       
       Δ𝛼 
 𝑦1 − 𝑦1   𝑒 𝑦1  0 0   
   Δ𝜆 
0
 
1 𝑐1 0 𝑑1 0 𝑞 ... 0 𝑟 ...
   .   . ..   1

..
  ..   .. .. .. .. .. .... . . .. .. . .  Δ𝜆 

.
    . . . . . . . . . . .   2
 𝑦𝑝 − 𝑦𝑝0  𝑒  0 𝑟   
      Δ𝑘 
 −      
1 𝑐𝑝 0 𝑑𝑝 0 0 ... 𝑞 0 ...

𝑦
=    Δ𝑢¯1  .
𝑝

       . 
 𝑢 1 − 𝑢¯0  
 1  1  0 0  . 
   . ..   . 
𝑒 0 0 0 0 0 1 ... 0 0 ...
 
𝑢
   ..   .  
 .   . .   Δ𝑢¯𝑝 
.. .... .. .. .. .. . . .. .. . .
 .    
. . . . . . . . . .
  
 𝑢 − 𝑢¯0   𝑒𝑢𝑝  0 0   Δ¯𝑣 1 
 𝑝 𝑝    0 0 0 0 0 0 ... 1 0 ...  
       .. 
       . 
  
 𝑣 1 − 𝑣¯1 
0
 𝑒 𝑣1  0
 0   
 Δ¯𝑣 𝑝 
 
0 0 0 0 0 0 ... 0 1 ...
   .   . ..   

..
  ..   .. .... .. .. .. .. . . .. .. . .
| {z }

.
    . . . . . . . . . . . 
 𝑣 𝑝 − 𝑣¯𝑝  𝑒  0 1 
    
0 0 0 0 0 0 0 ... 0 0 ... 𝜉
| {z } | {z } | {z }
𝑣 𝑝
(2𝑝+6)×1

𝑙 𝑒 𝐴
4𝑝×1 4𝑝×1 4𝑝×(2𝑝+6)

Results: by using these initial approximate values for unknown parameters

𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝜆10 = 1, 𝜆20 = 1, 𝛼 0 = 1.5′′, 𝑘 0 = 0,

   
𝑢¯0 = 𝑢 1, . . . , 𝑢 4 T, 𝑣¯0 = 𝑣 1, . . . , 𝑣 4 T

c k < 10−11 ):
we obtain the parameters (after 5 iterations, k Δ𝜉

𝑡ˆ𝑥 = 5388.876 m, 𝛼ˆ = −5′7.89′′, 𝜆ˆ1 = 1.000 409 692,

𝑡ˆ𝑦 = 10 346.871 m, 𝑘ˆ = 2.8233 · 10−5, 𝜆ˆ2 = 1.000 406 924, 𝑒ˆT 𝑃 𝑒ˆ = 0.000 993 2 m2 .

Coordinates of data points in the target system are listed in Tab. 5.54.

Table 5.54: Coordinates of data points in the target system.

Point 𝑥/m 𝑦/m
13 20 112.220 22 501.176
14 19 631.071 22 296.945
15 18 980.833 22 208.689
16 19 668.169 22 868.593
17 19 308.037 22 680.279

91
5 Geomatics examples

source system target system

13 000 23 500
1
12 500 23 000
1
16 2
16 2
17
13 17
12 000 22 500 13

𝑦/m
𝑣/m

4 14
15 14
4 15
11 500 22 000
3
3
11 000 21 500
control points control points
source data source data
13 000 13 500 14 000 14 500 15 000 18 500 19 000 19 500 20 000 20 500
𝑢/m 𝑥/m

Figure 5.36: 6-parameter affine transformation: Gauss Markov model; inconsistencies in both source
and target systems.

Mixed model approach II: Extended B-model with inconsistencies in both (𝑥𝑖 , 𝑦𝑖 ) and (𝑢𝑖 , 𝑣𝑖 )
coordinates (𝑖 = 1, . . . , 𝑝 with 𝑝 number of control points).

𝑓𝑥𝑖 := 𝑥𝑖 − 𝑒𝑥𝑖 − 𝜆1 (𝑢𝑖 − 𝑒𝑢𝑖 ) (cos 𝛼 − 𝑘 sin 𝛼) + 𝜆1 (𝑣𝑖 − 𝑒 𝑣𝑖 ) (sin 𝛼 + 𝑘 cos 𝛼) + 𝑡𝑥 = 0


𝑓𝑦𝑖 := 𝑦𝑖 − 𝑒 𝑦𝑖 − −𝜆2 (𝑢𝑖 − 𝑒𝑢𝑖 ) sin 𝛼 + 𝜆2 (𝑣𝑖 − 𝑒 𝑣𝑖 ) cos 𝛼 + 𝑡 𝑦 = 0

Linearization using Taylor point 𝑒𝑥0𝑖 , 𝑒 𝑦0 𝑖 , 𝑒𝑢0𝑖 , 𝑒 𝑣0𝑖 , 𝑡𝑥0 , 𝑡 𝑦0 , 𝛼 0 , 𝜆10 , 𝜆20 , 𝑘 0 so that

𝑒𝑥𝑖 = 𝑒𝑥0𝑖 + Δ𝑒𝑥𝑖 , 𝑒𝑢𝑖 = 𝑒𝑢0𝑖 + Δ𝑒𝑢𝑖 , 𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝛼 = 𝛼 0 + Δ𝛼, 𝜆2 = 𝜆20 + Δ𝜆2,
𝑒 𝑦𝑖 = 𝑒 𝑦0 𝑖 + Δ𝑒 𝑦𝑖 , 𝑒 𝑣𝑖 = 𝑒 𝑣0𝑖 + Δ𝑒 𝑣𝑖 , 𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 , 𝑘 = 𝑘 0 + Δ𝑘, 𝜆1 = 𝜆10 + Δ𝜆1 .

 Δ𝑡𝑥 
 
 Δ𝑒𝑥𝑖   Δ𝑡 
" #        𝑦

𝜕𝑓𝑥𝑖
  Δ𝑒 𝑦𝑖  
𝜕𝑓𝑥𝑖
  
𝑓𝑥0𝑖  Δ𝛼 
+    +    = 0,
𝜕 (𝑒𝑥𝑖 ,𝑒 𝑦𝑖 ,𝑒𝑢𝑖 ,𝑒 𝑣𝑖 ) 0 𝜕 (𝑡𝑥 ,𝑡 𝑦 ,𝛼,𝜆1 ,𝜆2 ,𝑘) 0
  Δ𝑒𝑢     Δ𝜆1 
       
𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
𝑓𝑦0𝑖
   Δ𝑒     Δ𝜆2 
𝑖

 𝑣𝑖 
𝜕 (𝑒𝑥𝑖 ,𝑒 𝑦𝑖 ,𝑒𝑢𝑖 ,𝑒 𝑣𝑖 ) 0 𝜕 (𝑡𝑥 ,𝑡 𝑦 ,𝛼,𝜆1 ,𝜆2 ,𝑘) 0
 
 Δ𝑘 
 
where

𝑓𝑥0𝑖 = 𝑥𝑖 − 𝑒𝑥0𝑖 − 𝜆10 (𝑢𝑖 − 𝑒𝑢0𝑖 ) (cos 𝛼 0 − 𝑘 0 sin 𝛼 0 ) + 𝜆10 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) (sin 𝛼 0 + 𝑘 0 cos 𝛼 0 ) + 𝑡𝑥0
 
and 𝑓𝑦0𝑖 = 𝑦𝑖 − 𝑒 𝑦0 𝑖 − −𝜆20 (𝑢𝑖 − 𝑒𝑢0𝑖 ) sin 𝛼 0 + 𝜆20 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) cos 𝛼 0 + 𝑡 𝑦0 .

92
 5.3 Mixed model

In matrix notation:
 𝑒𝑥 1 
 
 .. 
 . 
 
 𝑒𝑥 
 𝑥1 − 𝑥 0   𝑝  −1 0 𝑑𝑎,1 𝑑𝑏,1 0 𝑑𝑐,1 
 1  𝑒   
   𝑦1   .. .. ..   Δ𝑡𝑥 
 ..   .   . .
.. .. ..
.   
 .   ..   . . .  Δ𝑡 𝑦 
 𝑥 − 𝑥0   −𝐼𝑝 0𝑝 𝜆10𝑏 0 𝐼𝑝     −1 0 𝑑𝑎,𝑝 𝑑𝑏,𝑝 0 𝑑𝑐,𝑝   
 𝑝 𝑝   𝜆10𝑎 0 𝐼𝑝     Δ𝛼 
     𝑒 𝑦𝑝 
  +     +  

 
 Δ𝜆1  = 0,
     𝑒𝑢1   0 −1
 𝑦1 − 𝑦1 
0
 0𝑝 −𝐼𝑝 −𝜆2 sin 𝛼 𝐼𝑝 𝜆2 cos 𝛼 𝐼𝑝 
0 0 0 0  .   𝑑𝑑,1 0 𝑑𝑒,1 0  
 Δ𝜆2 

  | {z }  .   . . ..   
   .   .. ..  Δ𝑘 
   
.. .. .. ..

.
  𝑒𝑢𝑝   . . . .   
 𝑦𝑝 − 𝑦𝑝0 
𝐵T
   0 −1 𝑑𝑑,𝑝 0 𝑑𝑒,𝑝 0  | {z }
 
2𝑝×4𝑝
 𝑒 𝑣1  
| {z }   | {z }
 .. 
Δ𝜉
 . 
 
𝑤 𝐴 6×1

 𝑒𝑣 
2𝑝×1 2𝑝×6
 𝑝
| {z }
𝑒
4𝑝×1

where 𝐼𝑝 is the unit matrix of size 𝑝 × 𝑝 and 0𝑝 the zero matrix of the same size. Additionally, we
put the abbreviations 𝑎 0 = cos 𝛼 0 − 𝑘 0 sin 𝛼 0 , 𝑏 0 = sin 𝛼 0 + 𝑘 0 cos 𝛼 0 , 𝑢¯𝑖0 = 𝑢𝑖 − 𝑒𝑢0𝑖 and 𝑣¯𝑖0 = 𝑣𝑖 − 𝑒 𝑣0𝑖 to
get



𝑥𝑖0 = 𝜆10 𝑎 0𝑢𝑖 + 𝑏 0𝑣𝑖 + 𝑡𝑥0 , 𝑦𝑖0 = 𝜆20 − sin 𝛼 0𝑢𝑖 + cos 𝛼 0𝑣𝑖 + 𝑡 𝑦0




𝑑𝑎,𝑖 = 𝜆10 𝑏 0𝑢¯𝑖0 − 𝑎 0𝑣¯𝑖0 , 𝑑𝑏,𝑖 = − 𝑎 0𝑢¯𝑖0 + 𝑏 0𝑣¯𝑖0 , 𝑑𝑐,𝑖 = 𝜆10 sin 𝛼 0𝑢¯𝑖0 − cos 𝛼 0𝑣¯𝑖0 ,


𝑑𝑑,𝑖 = 𝜆20 cos 𝛼 0𝑢¯𝑖0 + sin 𝛼 0𝑣¯𝑖0 , 𝑑𝑒,𝑖 = sin 𝛼 0𝑢¯𝑖0 − cos 𝛼 0𝑣¯𝑖0 .

Results: with the following initial approximate values for unknown parameters and inconsis-
tencies

𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝜆10 = 1, 𝜆20 = 1, 𝛼 0 = 1.5′′, 𝑘 0 = 0, 𝑒𝑢0𝑖 = 𝑒 𝑣0𝑖 = 0 ∀𝑖

c k < 10−11 ):
we get the parameters (after 4 iterations, k Δ𝜉

𝑡ˆ𝑥 = 5388.876 m, 𝛼ˆ = −5′7.89′′, 𝜆ˆ1 = 1.000 409 692,

𝑡ˆ𝑦 = 10 346.871 m, 𝑘ˆ = 2.8233 · 10−5, 𝜆ˆ2 = 1.000 406 924, 𝑒ˆT 𝑃 𝑒ˆ = 0.000 993 2 m2 .

93
5 Geomatics examples

5.3.5 2D Affine Transformation Model II

Example: 6-parameter affine transformation – model II

The same data sets (85) will be analysed using a second model for the 6-parameter affine transfor-
mation.
         
𝑥𝑖 cos 𝜀 − sin 𝛿 𝜆1 0 𝑢 𝑡
= + 𝑥
𝑦𝑖 sin 𝜀 cos 𝛿 0 𝜆2 𝑣 𝑡𝑦
|{z} | {z } | {z } |{z}
target rotation angles scale factors source

Mixed model approach I: A-model with inconsistencies in both [𝑥𝑖 , 𝑦𝑖 ] and [𝑢𝑖 , 𝑣𝑖 ] coordinates.

𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆1𝑢¯𝑖 cos 𝜀 − 𝜆2𝑣¯𝑖 sin 𝛿 + 𝑡𝑥

𝑦𝑖 − 𝑒 𝑦𝑖 = 𝜆1𝑢¯𝑖 sin 𝜀 + 𝜆2𝑣¯𝑖 cos 𝛿 + 𝑡 𝑦
𝑢𝑖 − 𝑒𝑢𝑖 = 𝑢¯𝑖
𝑣𝑖 − 𝑒 𝑣𝑖 = 𝑣¯𝑖

Approximate values:

𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 , 𝜀 = 𝜀 0 + Δ𝜀, 𝛿 = 𝛿 0 + Δ𝛿, 𝜆1 = 𝜆10 + Δ𝜆, 𝜆2 = 𝜆20 + Δ𝜆2 .

Linearization process:

𝑥𝑖 − 𝑒𝑥𝑖 = 𝜆10𝑢¯𝑖0 cos 𝜀 0 − 𝜆20𝑣¯𝑖0 sin 𝛿 0 + 𝑡𝑥0 +Δ𝑡𝑥 − 𝜆10𝑢¯𝑖0 sin 𝜀 0 Δ𝜀 − 𝜆20𝑣¯𝑖0 cos 𝛿 0 Δ𝛿
| {z }
𝑥𝑖0

+ 𝑢¯𝑖0 cos 𝜀 0 Δ𝜆1 − 𝑣¯𝑖0 sin 𝛿 0 Δ𝜆2 + 𝜆10 cos 𝜀 0 Δ𝑢¯𝑖 − 𝜆20 sin 𝛿 0 Δ¯𝑣𝑖
 
𝑦𝑖 − 𝑒 𝑦𝑖 = 𝜆10𝑢¯𝑖0 sin 𝜀 0 + 𝜆20𝑣¯𝑖0 cos 𝛿 0 + 𝑡 𝑦0 +Δ𝑡 𝑦 + 𝜆10𝑢¯𝑖0 cos 𝜀 0 Δ𝜀 − 𝜆20𝑣¯𝑖0 sin 𝛿 0 Δ𝛿
| {z }
𝑦𝑖0

+ 𝑢¯𝑖0 sin 𝜀 0 Δ𝜆1 + 𝑣¯𝑖0 cos 𝛿 0 Δ𝜆2 + 𝜆10 sin 𝜀 0 Δ𝑢¯𝑖 + 𝜆20 cos 𝛿 0 Δ¯𝑣𝑖
𝑢𝑖 − 𝑒𝑢𝑖 = 𝑢¯𝑖0 + Δ𝑢¯𝑖
𝑣𝑖 − 𝑒 𝑣𝑖 = 𝑣¯𝑖0 + Δ¯𝑣𝑖

94
 5.3 Mixed model

In matrix notation:
 𝑥1 − 𝑥 0   𝑒𝑥 1  1 0 −𝑏𝑠 . . . 0 
 1    
0 −𝑎𝑠 𝑢¯10 −𝑏𝑐 𝑣¯10 𝑐𝑐,1 −𝑑𝑠,1 𝑎𝑐 . . .

   ..   .. .. .. . . .. 
 .     .
.. .. .. .. .. .. . .
.  
 
..
 .   Δ𝑡 
. . . . . . . . . .
 𝑥 − 𝑥0    1 . . . 𝑎𝑐 0 . . . −𝑏𝑠   𝑥 
 𝑝 𝑝   𝑒𝑥 𝑝   0 −𝑎𝑠 𝑢¯𝑝0 −𝑏𝑐 𝑣¯𝑝0 𝑐𝑐,𝑝 −𝑑𝑠,𝑝
 
0
       Δ𝑡 𝑦 
       Δ𝜀 
 𝑦1 − 𝑦1   𝑒 𝑦1  0 . . . 0 𝑏𝑐 . . . 0   
  
0
 
1 𝑎𝑐 𝑢¯10 −𝑏𝑠 𝑣¯10 𝑐𝑠,1 𝑑𝑐,1 𝑎𝑠
   .   .   Δ𝛿 

..
  ..   .. .. .. .. .. .. .. . . .. .. . . ..  
. .   Δ𝜆1 

.
    . . . . . . . . .
 𝑦𝑝 − 𝑦𝑝  𝑒  0 . . . 𝑎𝑠 0 . . . 𝑏𝑐   Δ𝜆2 
   𝑝  
0
1 𝑎𝑐 𝑢¯𝑝0 −𝑏𝑠 𝑣¯𝑝0 𝑐𝑠,𝑝 𝑑𝑐,𝑝 0
  −      
𝑦

  =    Δ𝑢¯1  ,
 𝑢1 − 𝑢       . 
 1 
0  𝑢1  0 ... 0 0 ... 0   . 
     . 
𝑒 0 0 0 0 0 1
   ..   .. . . .. .. . . ..   
 ..     .
.. .. .. .. .. ..
. .   Δ𝑢¯𝑝 
 .  
.
  . . . . . . . . .
  
 𝑢 − 𝑢0    0 . . . 1 0 . . . 0   Δ¯𝑣 1 
 𝑝   𝑒 𝑝  
 
0 0 0 0 0 0
     .. 
𝑝 𝑢
       . 
   𝑒𝑣  0
 𝑣 1 − 𝑣 10   1   . . . 0 1 . . . 0   

 
0 0 0 0 0 0
   .   . . . .. .. . . ..  |  Δ¯𝑣 𝑝 

..
  ..   .. .. .. .. .. .. .. {z }

.
    . . . . . . . . . . . 
 𝑣𝑝 − 𝑣𝑝  𝑒  0 . . . 0 0 . . . 1 
   𝑣𝑝  
0 0 0 0 0 0 0 Δ𝜉
| {z } | {z } | {z } (2𝑝+6)×1

𝑙 𝑒 𝐴
4𝑝×1 4𝑝×1 4𝑝×(2𝑝+6)

where
𝑎𝑐 = 𝜆10 cos 𝜀 0, 𝑏𝑐 = 𝜆20 cos 𝛿 0, 𝑐𝑐,𝑖 = 𝑢¯𝑖0 cos 𝜀 0, 𝑑𝑐,𝑖 = 𝑣¯𝑖0 cos 𝛿 0,
𝑎𝑠 = 𝜆10 sin 𝜀 0, 𝑏𝑠 = 𝜆20 sin 𝛿 0, 𝑐𝑠,𝑖 = 𝑢¯𝑖0 sin 𝜀 0, 𝑑𝑠,𝑖 = 𝑣¯𝑖0 sin 𝛿 0 .

Results: Initial approximate values for unknown parameters

𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝜀 0 = 1.5′′, 𝛿 0 = 3.5′′, 𝜆10 = 1, 𝜆20 = 1,
   
𝑢¯0 = 𝑢 1, . . . , 𝑢 4 T, 𝑣¯0 = 𝑣 1, . . . , 𝑣 4 T .

c k < 10−11 ):
Parameters (after 7 iterations, k Δ𝜉
𝑡ˆ𝑥 = 5388.876 m, 𝜀ˆ = 5′7.89′′, 𝜆ˆ1 = 1.000 409 734,
𝑡ˆ𝑦 = 10 346.871 m, 𝛿ˆ = 5′2.06′′, 𝜆ˆ2 = 1.000 406 883, 𝑒ˆT 𝑃 𝑒ˆ = 0.000 993 2 m2 .

Mixed model approach II: Extended B-model with inconsistences in both [𝑥𝑖 , 𝑦𝑖 ] and [𝑢𝑖 , 𝑣𝑖 ]
coordinates.




𝑓𝑥𝑖 := 𝑥𝑖 − 𝑒𝑥𝑖 − 𝜆1 𝑢𝑖 − 𝑒𝑢𝑖 cos 𝜀 − 𝜆2 𝑣𝑖 − 𝑒 𝑣𝑖 sin 𝛿 + 𝑡𝑥 = 0




𝑓𝑦𝑖 := 𝑦𝑖 − 𝑒 𝑦𝑖 − 𝜆1 𝑢𝑖 − 𝑒𝑢𝑖 sin 𝜀 + 𝜆2 𝑣𝑖 − 𝑒 𝑣𝑖 cos 𝛿 + 𝑡 𝑦 = 0

Initial approximate values:

𝑒𝑥𝑖 = 𝑒𝑥0𝑖 + Δ𝑒𝑥𝑖 , 𝑒 𝑦𝑖 = 𝑒 𝑦0 𝑖 + Δ𝑒 𝑦𝑖 , 𝑒𝑢𝑖 = 𝑒𝑢0𝑖 + Δ𝑒𝑢𝑖 , 𝑒 𝑣𝑖 = 𝑒 𝑣0𝑖 + Δ𝑒 𝑣𝑖 ,
𝑡𝑥 = 𝑡𝑥0 + Δ𝑡𝑥 , 𝑡 𝑦 = 𝑡 𝑦0 + Δ𝑡 𝑦 , 𝜀 = 𝜀 0 + Δ𝜀, 𝛿 = 𝛿 0 + Δ𝛿, 𝜆1 = 𝜆10 + Δ𝜆1, 𝜆2 = 𝜆20 + Δ𝜆2 .

95
5 Geomatics examples

Linearization:
 Δ𝑡𝑥 
 
 Δ𝑒𝑥𝑖   Δ𝑡 
" #        𝑦

𝜕𝑓𝑥𝑖
  Δ𝑒 𝑦𝑖  
𝜕𝑓𝑥𝑖
  
𝑓𝑥0𝑖  Δ𝜀 
+    +    = 0,
𝜕 (𝑒𝑥𝑖 ,𝑒 𝑦𝑖 ,𝑒𝑢𝑖 ,𝑒 𝑣𝑖 ) 0 𝜕 (𝑡𝑥 ,𝑡 𝑦 ,𝜀,𝛿,𝜆1 ,𝜆2 ) 0
  Δ𝑒𝑢     Δ𝛿 
       
𝜕𝑓𝑦𝑖 𝜕𝑓𝑦𝑖
𝑓𝑦0𝑖
   Δ𝑒     Δ𝜆1 
𝑖

 𝑣𝑖 
𝜕 (𝑒𝑥𝑖 ,𝑒 𝑦𝑖 ,𝑒𝑢𝑖 ,𝑒 𝑣𝑖 ) 0 𝜕 (𝑡𝑥 ,𝑡 𝑦 ,𝜀,𝛿,𝜆1 ,𝜆2 ) 0
 
 Δ𝜆2 
 
where


𝑓𝑥0𝑖 = 𝑥𝑖 − 𝑒𝑥0𝑖 − 𝜆10 (𝑢𝑖 − 𝑒𝑢0𝑖 ) cos 𝜀 0 − 𝜆20 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) sin 𝛿 0 + 𝑡𝑥0
 
and 𝑓𝑦0𝑖 = 𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝜆10 (𝑢𝑖 − 𝑒𝑢0𝑖 ) sin 𝜀 0 + 𝜆20 (𝑣𝑖 − 𝑒 𝑣0𝑖 ) cos 𝛿 0 + 𝑡 𝑦0 .

In matrix notation:
 𝑒𝑥 1 
 
 .. 
 . 
 
 𝑒𝑥 
 𝑥1 − 𝑥 0   𝑝  −1 
 1  𝑒  
0 𝑎𝑠 𝑢¯10 𝑏𝑐 𝑣¯10 −𝑐𝑐,1 𝑑𝑠,1

   𝑦1   ..   Δ𝑡𝑥 
 ..   .   .
.. .. .. .. ..
  
 .   ..   . . . . .   Δ𝑡 𝑦 
 𝑥 − 𝑥0   −𝐼𝑝 0𝑝 𝑎𝑐 𝐼𝑝 −𝑏𝑠 𝐼𝑝     −1   
 𝑝 𝑝       𝑎𝑠 𝑢¯𝑝0 𝑏𝑐 𝑣¯𝑝0 −𝑐𝑐,𝑝 𝑑𝑠,𝑝   Δ𝜀 
   𝑒 𝑦𝑝 
0
  +     +    
     𝑒𝑢1     Δ𝛿  = 0,
 𝑦1 − 𝑦1   0𝑝 −𝐼𝑝 𝑎𝑠 𝐼𝑝 𝑏𝑐 𝐼𝑝   .   0 −𝑑𝑐,1   
    Δ𝜆1 
0
   . 
−1 −𝑎𝑐 𝑢¯1 𝑏𝑠 𝑣¯1 −𝑐𝑠,1
  | {z }  .   . ..   

..
    .. .. .. .. ..  Δ𝜆 

.
  𝑒𝑢𝑝   . . . . .   2
 𝑦𝑝 − 𝑦𝑝 
𝐵T
   0 −𝑑𝑐,𝑝  | {z }
   𝑒 𝑣1  
0 2𝑝×4𝑝 −1 −𝑎𝑐 𝑢¯𝑝 𝑏𝑠 𝑣¯𝑝 −𝑐𝑠,𝑝
| {z }   | {z }
 .. 
Δ𝜉
 . 
 
𝑤 𝐴 6×1

 𝑒𝑣 
2𝑝×1 2𝑝×6
 𝑝
| {z }
𝑒
4𝑝×1

where 𝐼𝑝 is the unit matrix of size 𝑝 × 𝑝 and 0𝑝 the zero matrix of the same size. Additionally, we
use the following abbreviations
𝑎𝑐 = 𝜆10 cos 𝜀 0, 𝑏𝑐 = 𝜆20 cos 𝛿 0, 𝑐𝑐,𝑖 = 𝑢¯𝑖0 cos 𝜀 0, 𝑑𝑐,𝑖 = 𝑣¯𝑖0 cos 𝛿 0, 𝑢¯𝑖0 = 𝑢𝑖 − 𝑒𝑢0𝑖 ,
𝑎𝑠 = 𝜆10 sin 𝜀 0, 𝑏𝑠 = 𝜆20 sin 𝛿 0, 𝑐𝑠,𝑖 = 𝑢¯𝑖0 sin 𝜀 0, 𝑑𝑠,𝑖 = 𝑣¯𝑖0 sin 𝛿 0, 𝑣¯𝑖0 = 𝑣𝑖 − 𝑒𝑢0𝑖 .

Results: with the following initial approximate values for unknown parameters
𝑡𝑥0 = 5500 m, 𝑡 𝑦0 = 10 200 m, 𝜀 0 = 1.5′′, 𝛿 0 = 3.5′′, 𝜆10 = 1, 𝜆20 = 1,
𝑒𝑥0𝑖 = 𝑒 𝑦0 𝑖 = 𝑒𝑢0𝑖 = 𝑒 𝑣0𝑖 = 0 ∀𝑖 .
c k < 10−11 ):
we get the parameters (after 20 iterations, k Δ𝜉
𝑡ˆ𝑥 = 5388.876 m, 𝜀ˆ = 5′7.89′′, 𝜆ˆ1 = 1.000 409 734,
𝑡ˆ𝑦 = 10 346.871 m, 𝛿ˆ = 5′2.06′′, 𝜆ˆ2 = 1.000 406 882, 𝑒ˆT 𝑃 𝑒ˆ = 0.000 993 2 m2 .

96
 5.3 Mixed model

5.3.6 Ellipse fit under various restrictions

Example: Best fitting ellipse (here: principal axes aligned with coordinate axes) with unknown
semi major axis 𝑎, semi minor axis 𝑏 and centre coordinates (𝑥 M, 𝑦M ); observations 𝑥𝑖 and 𝑦𝑖 are
inconsistent.

2
2
𝑥 𝑖 − 𝑒𝑥𝑖 − 𝑥 M 𝑦𝑖 − 𝑒 𝑦𝑖 − 𝑦M
𝑓 (𝑎, 𝑏, 𝑥 M, 𝑦M, 𝑥𝑖 − 𝑒𝑥𝑖 , 𝑦𝑖 − 𝑒 𝑦𝑖 ) = + −1=0
| {z } | {z } 𝑎2 𝑏2
unknown observations 𝑦
parameters 𝜉 − inconsistencies 𝑒

Possible restriction: Best fitting ellipse shall pass through the point (𝑥 P, 𝑦P )

(𝑥 P − 𝑥 M ) 2 (𝑦P − 𝑦M ) 2
𝑔(𝑎, 𝑏, 𝑥 M, 𝑦M ) = + −1=0
| {z } 𝑎2 𝑏2
𝜉

Linearization (with 𝑒𝑥0𝑖 = 𝑒 𝑦0 𝑖 = 0 in the first iteration and given 𝜉 0 )

𝜉 = 𝜉 0 + Δ𝜉, 𝑒𝑥𝑖 = 𝑒𝑥0𝑖 + Δ𝑒𝑥𝑖 , 𝑒 𝑦𝑖 = 𝑒 𝑦0 𝑖 + Δ𝑒 𝑦𝑖

𝜕𝑓 𝜕𝑓
𝑓 (𝜉, 𝑒) = 𝑓 (𝜉 0, 𝑒 0 ) + Δ𝜉 + 𝑒+O =0
𝜕𝜉 𝑥 0 ,𝑒 0 𝜕𝑒 𝜉 0 ,𝑒 0

𝑤 + 𝐴 Δ𝜉 + 𝐵 T𝑒 =0
𝜕𝑔
𝑔(𝜉) = 𝑔(𝜉 0 ) + Δ𝜉 +O =0
𝜕𝜉 𝜉0

 𝑤𝑅 + 𝑅 Δ𝜉 =0

with O terms of higher order.

Linear model and adjustment principle
)
𝐴Δ𝜉 + 𝐵 T𝑒 = −𝑤 1 T
𝑒 𝑊 𝑒 −→ min
𝑅Δ𝜉 = −𝑤 𝑅 2

Constrained Lagrangian (𝑚 observation equations, 𝑛 unknown parameters, 𝑝 inconsistencies, 𝑟 re-

strictions)

1 T
L𝑅 (Δ𝜉, 𝑒, 𝜆, 𝜆𝑅 ) = 𝑒 𝑊 𝑒 + 𝜆 T ( 𝐴 Δ𝜉 + 𝐵 T 𝑒 + 𝑤 )
2 1×𝑝 𝑝×𝑝 𝑝×1 1×𝑚 𝑚×𝑛 𝑛×1 𝑚×𝑝 𝑝×1 𝑚×1
+ 𝜆𝑅 T ( 𝑅 Δ𝜉 + 𝑤 𝑅 ) −→ min
1×𝑟 𝑟 ×𝑛 𝑛×1 𝑟 ×1 Δ𝜉,𝑒,𝜆,𝜆𝑅

97
5 Geomatics examples

Necessary condition

𝜕L𝑅 c ˆ ˆ !
( Δ𝜉, 𝑒,
ˆ 𝜆, 𝜆𝑅 ) = 0 =⇒ 𝐴T 𝜆ˆ + 𝑅 T 𝜆ˆ𝑅 = 0
𝜕Δ𝜉
𝜕L𝑅 c ˆ ˆ !
( Δ𝜉, 𝑒,
ˆ 𝜆, 𝜆𝑅 ) = 0 =⇒ 𝑊 𝑒ˆ + 𝐵 𝜆ˆ =0
𝜕𝑒
𝜕L𝑅 c ˆ ˆ ! c + 𝐵 T𝑒ˆ
( Δ𝜉, 𝑒,
ˆ 𝜆, 𝜆𝑅 ) = 0 =⇒ 𝐴 Δ𝜉 = −𝑤
𝜕𝜆
𝜕L𝑅 c ˆ ˆ ! c
( Δ𝜉, 𝑒,
ˆ 𝜆, 𝜆𝑅 ) = 0 =⇒ 𝑅 Δ𝜉 = −𝑤 𝑅
𝜕𝜆𝑅

𝑊 𝐵 0   0 
 𝑝×𝑝 𝑝×𝑚 0
 𝑝×1 
   𝑒ˆ   
 𝐵T 0 𝐴 0     −𝑤 
𝑝×𝑛 𝑝×𝑟

  𝜆ˆ   
 𝑚×𝑝 𝑚×𝑚 𝑚×𝑛 𝑚×𝑟   
   c  =  𝑚×1 

 0 𝐴T 0 𝑅T   Δ𝜉   0 
 𝑛×𝑝 𝑛×𝑚 𝑛×𝑛 𝑛×𝑟     𝑛×1 
   𝜆ˆ𝑅   −𝑤 
 0 0 𝑅 0     
 𝑟 ×𝑝 𝑟 ×𝑚  𝑟 ×1 
 𝑟 ×𝑛 𝑟 ×𝑟 
(𝑝+𝑚+𝑛+𝑟 )×1
𝑅

(𝑝+𝑚+𝑛+𝑟 )×(𝑝+𝑚+𝑛+𝑟 )

1st row multiplied with −𝐵 T𝑊 −1 (from left) is added to 2nd row

𝑊
 𝐵 0 0   𝑒ˆ   0 
 0  
 −𝐵 𝑊 −1 𝐵
T 𝐴 0   𝜆ˆ   −𝑤 
 0 c = 
 0 𝑅 T   Δ𝜉  0 
  
𝐴T
 0 
𝑅 0   𝜆ˆ𝑅   −𝑤 𝑅 
 0

−1
2nd row multiplied with 𝐴T 𝐵 T𝑊 −1 𝐵 (from left) is added to 3rd row

𝑊 0   𝑒ˆ   
 𝐵 0 0 
 0     
 −𝐵 T𝑊 −1 𝐵


0  𝜆  ˆ −𝑤


    
c   −𝐴T 𝐵 T𝑊 −1 𝐵 𝑤 
𝐴
 0 𝐴T 𝐵 T𝑊 −1 𝐵 𝐴 𝑅 T   Δ𝜉
−1 = −1
    
0
 0 0   𝜆ˆ𝑅   
 0 𝑅 −𝑤 𝑅 
 T T −1 −1 " #  
𝐴 (𝐵 𝑊 𝐵) 𝐴 𝑅 T c
Δ𝜉 T T −1
−𝐴 (𝐵 𝑊 𝐵) 𝑤 −1
=⇒ ˆ =
𝑅 0 𝜆𝑅 −𝑤 𝑅

−1
Case 1: 𝐴T 𝐵 T𝑊 −1 𝐵 𝐴 = 𝐴T 𝑀 −1𝐴 is a full-rank matrix. =⇒ Use partitioning formula:





𝑄 22 −1 𝑁 ) −1
= (𝑁 22 − 𝑁 21 𝑁 11
     
12

 𝑄 12
 −1 𝑁 𝑄
= −𝑁 11
𝑁 11 𝑁 12 𝑄 11 𝑄 12 𝐼 0 12 22

= =⇒


𝑁 21 𝑁 22 𝑄 21 𝑄 22 0𝐼 −1
= −𝑄 22 𝑁 21 𝑁 11

𝑄 21

 𝑄 11

−1 + 𝑁 −1 𝑁 𝑄 𝑁 𝑁 −1
= 𝑁 11 11 12 22 21 11

98
 5.3 Mixed model

𝑁 11 = 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝐴 = 𝐴T 𝑀 −1𝐴
𝑁 12 = 𝑅 T
𝑁 21 = 𝑁 12 T = 𝑅
𝑁 22 = 0
𝑄 22 = [0 − 𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T ] −1 = −[𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T ] −1
𝑄 12 = (𝐴T 𝑀 −1𝐴) −1𝑅 T [𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T ] −1 = −𝑁 11
−1
𝑁 12𝑄 22
𝑄 21 = 𝑄 12 T
𝑄 11 = (𝐴T 𝑀 −1𝐴) −1 {𝐼 − 𝑅 T [𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T ] −1𝑅(𝐴T 𝑀 −1𝐴) −1 }
−1 −1
= 𝑁 11 − 𝑄 12 𝑁 12 T 𝑁 11

c = −𝑄 11𝐴T 𝑀 −1𝑤 − 𝑄 12𝑤 𝑅

Δ𝜉
= −(𝐴T 𝑀 −1𝐴) −1𝐴T 𝑀 −1𝑤
  −1
+ (𝐴T 𝑀 −1𝐴) −1𝑅 T 𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T 𝑅(𝐴T 𝑀 −1𝐴) −1𝐴T 𝑀 −1𝑤
  −1
− (𝐴T 𝑀 −1𝐴) −1𝑅 T 𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T 𝑤 𝑅

c
= −(𝐴T 𝑀 −1𝐴) −1𝐴T 𝑀 −1𝑤 + 𝛿 Δ𝜉
c
= Δ𝜉 (without restrictions 𝑔(𝜉) = 0) c
+ 𝛿 Δ𝜉

𝜆ˆ𝑅 = −𝑄 21𝐴T 𝑀 −1𝑤 − 𝑄 22𝑤 𝑅

 
= 𝑄 22 𝑅 T (𝐴T 𝑀 −1𝐴) −1𝐴T 𝑀 −1𝑤 − 𝑤 𝑅
  
= 𝑅(𝐴T 𝑀 −1𝐴) −1𝑅 T 𝑤 𝑅 − 𝑅 T (𝐴T 𝑀 −1𝐴) −1𝐴T 𝑀 −1𝑤

c
−𝑤 = −𝑀 𝜆ˆ + 𝐴Δ𝜉

=⇒ c + 𝑤)
𝜆ˆ = 𝑀 −1 (𝐴Δ𝜉
c + 𝑤)
𝑒ˆ = 𝑊 −1 𝐵 𝜆ˆ = 𝑊 −1 𝐵𝑀 −1 (𝐴Δ𝜉

Case 2: 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝐴 = 𝐴T 𝑀 −1𝐴 is a rank deficient matrix

rank(𝐴T 𝑀 −1𝐴) = rank 𝐴 = 𝑛 − 𝑑

  −1   −1
𝑁 𝑅T 𝑅 𝑆T
=
𝑅 0 𝑆 𝑄

99
5 Geomatics examples

𝑁 𝑅 + 𝑅 T𝑆 = 𝐼 (5.3)
𝑁 𝑆 T + 𝑅 T𝑄 = 0 (5.4)
𝑅𝑅 = 0 (5.5)
𝑅𝑆 T = 𝐼 (5.6)

Since 𝐴 is rank deficient 𝐴𝐻 T = 0 where 𝐻 = null(𝐴) 𝐻 : 𝑑 × 𝑛 therefore



 𝐴 𝑀 𝐴𝐻 = 0

T −1 T

𝑁𝐻T

=0
 𝐻𝑁
 =0

𝑁 is symmetric.

𝐻 𝑁 𝑅 +𝐻𝑅 T𝑆 = 𝐻 𝑆 = (𝐻𝑅 T ) −1𝐻

|{z}
𝐻 · (5.3) =⇒ =⇒
0

𝐻 𝑁 𝑆 T +𝐻𝑅 T𝑄 = 0 𝐻𝑅 T𝑄 = 0
|{z}
𝐻 · (5.4) =⇒ =⇒
0

𝐻𝑅 T full rank =⇒ 𝑄 = 0.

(5.3) =⇒ 𝑁 𝑅 + 𝑅 T (𝐻𝑅 T ) −1𝐻 = 𝐼 




(𝑁 + 𝑅 T𝑅)𝑅 = 𝐼 − 𝑅 T (𝐻𝑅 T ) −1𝐻
=0 
⊕ =⇒


(5.5) =⇒ 𝑅𝑅 = 0 =⇒ 𝑅 T𝑅𝑅
 
=⇒ 𝑅 = (𝑁 + 𝑅 T𝑅) −1 𝐼 − 𝑅 T (𝐻𝑅 T ) −1𝐻

c = −𝑅𝐴T 𝑁 −1𝑤 + 𝑆 T𝑤 𝑅
Δ𝜉
= −(𝑁 + 𝑅 T𝑅) −1𝐴T 𝑀 −1𝑤
+ (𝑁 + 𝑅 T𝑅) −1𝑅 T (𝐻𝑅 T ) −1𝐻𝐴T 𝑀 −1𝑤 −𝑆 T𝑤 𝑅
| {z }
=0

= −(𝑁 + 𝑅 𝑅) 𝐴 𝑀 −1𝑤 − 𝐻 T (𝑅𝐻 T ) −1𝑤 𝑅

T −1 T

if 𝑤 𝑅 = 0:
c = −(𝑁 + 𝑅 T𝑅) −1𝐴T 𝑀 −1𝑤
Δ𝜉

c −→ 𝜆ˆ = 𝑀 −1 (𝐴Δ𝜉
Δ𝜉 c + 𝑤)

= 𝑀 −1 (−(𝑁 + 𝑅 T𝑅) −1𝐴T 𝑀 −1𝑤

 
𝜆ˆ = −𝑀 −1 (𝑁 + 𝑅 T𝑅) −1𝐴T 𝑀 −1 − 𝐼 𝑤

𝑒ˆ = 𝑊 −1 𝐵 𝜆ˆ
 
= −𝑊 −1 𝐵𝑀 −1 (𝑁 + 𝑅 T𝑅) −1𝐴T 𝑀 −1 − 𝐼 𝑤

100
 5.3 Mixed model

Examples for case 1: 𝐴T (𝐵 T 𝐵) −1𝐴 is a full-rank matrix.

Best fitting ellipse with unknown semi major axes 𝑎 and 𝑏, unknown centre coordinates 𝑥 M , 𝑦M ,
inconsistent observations 𝑥𝑖 and 𝑦𝑖 , 𝑖 = 1, . . . , 𝑚, no restrictions 𝑔(𝑥), ellipse aligned with coordinate
axes!
Ellipse equation

!2 !2
𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M
0 𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0
𝑓 (𝑎, 𝑏, 𝑥 M, 𝑦M, 𝑥𝑖 − 𝑒𝑥𝑖 , 𝑦𝑖 − 𝑒 𝑦𝑖 ) = + −1
𝑎0 𝑏0

2(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M
0) 2(𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0 )
+ 𝑒𝑥0𝑖 + 𝑒 𝑦0 𝑖
𝑎 20 𝑏 02
2(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M
0) 0 )
2(𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
− Δ𝑥 M − Δ𝑦M
𝑎 20 𝑏 02
2(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M
0) 2(𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0 )
− 𝑒𝑥𝑖 − 𝑒 𝑦𝑖
𝑎 20 𝑏 02
2(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M
0 )2 2(𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0 )2
− Δ𝑎 − Δ𝑏 = 0
𝑎 30 𝑏 03

h 0 ) 2(𝑦 −𝑒 0 −𝑦 0 )
2(𝑥𝑖 −𝑒𝑥0 𝑖 −𝑥 M
i 𝑒 
𝑖 𝑦𝑖 M 𝑥𝑖
=⇒ 𝑎 20 𝑏 02 𝑒 𝑦𝑖
 Δ𝑥 M 
h i  Δ𝑦 
2(𝑥𝑖 −𝑒𝑥0 𝑖 −𝑥 M
 M
0) 2(𝑦𝑖 −𝑒 𝑦0 𝑖 −𝑦M
0 ) 2(𝑥𝑖 −𝑒𝑥0 𝑖 −𝑥 M
0 )2 2(𝑦𝑖 −𝑒 𝑦0 𝑖 −𝑦M
0 )2
+ − − − −  Δ𝑎 
 
𝑎02 𝑏0 2 𝑎03 𝑏0 3

 Δ𝑏 
 
0 )2
(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M (𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0 )2 0)
2(𝑥𝑖 − 𝑒𝑥0𝑖 − 𝑥 M 2(𝑦𝑖 − 𝑒 𝑦0 𝑖 − 𝑦M
0 )
+ + −1+ 𝑒𝑥0𝑖 + 𝑒 𝑦0 𝑖 = 0
𝑎 20 𝑏 02 𝑎 20 𝑏 02

=⇒ 𝑝 = 2𝑚, 𝑊 = 𝐼𝑝 (𝑝 × 𝑝 identity matrix), 𝑛=4

101
5 Geomatics examples

=⇒ 𝐵 T𝑒 + 𝐴Δ𝜉 + 𝑤 = 0

 𝑥 1 −𝑒𝑥0 1 −𝑥 M0 𝑦1 −𝑒 𝑦0 1 −𝑦M
0

 
 
𝑎 20 𝑏 02
0 0 ... 0 0
 
 
0 𝑦 −𝑒 0 −𝑦 0
𝑥 2 −𝑒𝑥0 2 −𝑥 M 2 𝑦2

𝐵 T = −2  
M
0 0 ... 0 0

𝑎 20 𝑏 02
 .. .. .. .. .. .. .. 
 . 
𝑚×𝑝
. . . . . .
 
 
0 𝑦 −𝑒 0 −𝑦 0
𝑥𝑚 −𝑒𝑥0𝑚 −𝑥 M
 
𝑚 𝑦𝑚 M
0 0 0 0 ... 𝑎02 𝑏 02

 𝑥 1 −𝑒𝑥0 1 −𝑥 M0 𝑦1 −𝑒 𝑦0 1 −𝑦M0 (𝑥 1 −𝑒𝑥0 1 −𝑥 M0 ) 2 0 )2

(𝑦1 −𝑒 𝑦0 1 −𝑦M 
 
 𝑎 20 𝑏 02 𝑎 30 𝑏 03 
 𝑥 2 −𝑒𝑥0 2 −𝑥 M0 𝑦2 −𝑒 𝑦0 2 −𝑦M0 (𝑥 2 −𝑒𝑥0 2 −𝑥 M0 ) 2 
 
(𝑦2 −𝑒 𝑦0 2 −𝑦M
0 )2

𝐴 = −2  𝑎 20 𝑏 02 𝑎 30 𝑏 03 

 . . .. .. 
 
𝑚×4 .. .. . .
 𝑥𝑚 −𝑒𝑥0𝑚 −𝑥 M0 𝑦𝑚 −𝑒 𝑦0 𝑚 −𝑦M0 (𝑥𝑚 −𝑒𝑥0𝑚 −𝑥 M0 ) 2 
 
0 )2
(𝑦𝑚 −𝑒 𝑦0 𝑚 −𝑦M
 𝑎 20 𝑏 02 𝑎 30 𝑏 03 
 T
𝑒 = 𝑒𝑥 1 𝑒 𝑦1 𝑒𝑥 2 𝑒 𝑦2 . . . 𝑒𝑥𝑚 𝑒 𝑦𝑚
𝑝×1

 T
Δ𝜉 = Δ𝑥 M Δ𝑦M Δ𝑎 Δ𝑏
4×1

 (𝑥 1 −𝑒𝑥0 1 −𝑥 M
0 )2 (𝑦1 −𝑒 𝑦0 1 −𝑦M
0 )2 𝑥 1 −𝑒𝑥0 1 −𝑥 M
0

𝑦1 −𝑒 𝑦0 1 −𝑦M
0
 + +2 𝑒𝑥01 +2  𝑒 𝑦0 1 −1
 𝑎 20 𝑏 02 𝑎 20 𝑏 02 
 
 
(𝑥 2 −𝑒𝑥0 2 −𝑥 M
0 )2 (𝑦2 −𝑒 𝑦0 2 −𝑦M
0 )2 𝑥 2 −𝑒𝑥0 2 −𝑥 M
0 𝑦2 −𝑒 𝑦0 2 −𝑦M
0
𝑒𝑥02 𝑒 𝑦0 2
𝑤 =  
+ + 2 + 2 −1

𝑎 20 𝑏 02 𝑎
0
2
0 𝑏 2

 .. 
 
𝑚×1
.
 
 𝑒 𝑦𝑚 − 1 
0 )2
(𝑥𝑚 −𝑒𝑥0𝑚 −𝑥 M 0 )2
(𝑦𝑚 −𝑒 𝑦0 𝑚 −𝑦M 0 0
𝑥𝑚 −𝑒𝑥𝑚 −𝑥 M 0 0 0
𝑦𝑚 −𝑒 𝑦𝑚 −𝑦M 0
 2
𝑎0
+ 𝑏02 +2 𝑎 20
𝑒𝑥𝑚 + 2 𝑏 02

 𝑊 0   𝑒ˆ   0 
 2𝑚×2𝑚 2𝑚×𝑚
𝐵  2𝑚×1 
 2𝑚×𝑛   2𝑚×1   
 𝐵T 𝐴   𝜆ˆ  =  −𝑤 
 0  
 𝑚×2𝑚 𝑚×𝑚 𝑚×𝑛   𝑚×1   𝑚×1 
=⇒
 0T 𝐴T 0   Δ𝜉   0 
   c   
 𝑛×2𝑚 𝑛×𝑚 𝑛×𝑛   𝑛×1   𝑛×1 
(3𝑚+4)×(3𝑚+4) (3𝑚+4)×1 (3𝑚+4)×1

)
𝑊 𝑒ˆ + 𝐵 𝜆ˆ = 0 =⇒ 𝑒ˆ = −𝑊 −1 𝐵𝜆ˆ
c = −𝑤
=⇒ −𝐵 T𝑊 −1 𝐵 𝜆ˆ + 𝐴Δ𝜉
c = −𝑤
𝐵 T𝑒ˆ + 𝐴Δ𝜉

=⇒ c + 𝑤)
𝜆ˆ = (𝐵 T𝑊 −1 𝐵) −1 (𝐴Δ𝜉

=⇒ c + 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝑤 = 0
𝐴T (𝐵 T𝑊 −1 𝐵) −1𝐴Δ𝜉
  −1
=⇒ c = − 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝐴 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝑤
Δ𝜉

=⇒ c + 𝑤)
𝑒ˆ = −𝑊 −1 𝐵(𝐵 T𝑊 −1 𝐵) −1 (𝐴Δ𝜉
   −1 
= 𝑊 −1 𝐵(𝐵 T𝑊 −1 𝐵) −1 𝐴 𝐴T (𝐵 T𝑊 −1 𝐵) −1𝐴 𝐴T (𝐵 T𝑊 −1 𝐵) −1 − 𝐼 𝑤

102
 5.3 Mixed model

Numerics
 T
𝑥= 0, 50, 90, 120, 130, −130, −100, −50, 0
 T
𝑦= 120, 110, 80, 0, −50, −50, 60, 100, −110

Approximated values:
0 0
𝑥M = 𝑦M = 0, 𝑎 0 = 𝑏 0 = 120 .
c k < 10−12 , k Δ𝑒
Parameters (after 10 iterations: k Δ𝜉 c k < 10−12 ):

𝑥ˆM = −0.598, 𝑦ˆM = −1.942, 𝑎ˆ = 131.087, 𝑏ˆ = 115.131, 𝑒ˆT𝑊 𝑒ˆ = 523.208

 T
𝑒ˆ𝑥 = 0.026, 1.793, −0.627, −10.466, 9.322, −8.324, 6.771, 1.534, −0.030
 T
𝑒ˆ𝑦 = 6.813, 5.089, −0.736, −0.224, −4.355, −3.933, −5.583, −4.142, 7.072

See figure 5.37.

150
Observed Points
Estimated Centre
Adjusted Points

100

50

0
y/m

-50

-100

-150
-150 -100 -50 0 50 100 150
x/m

Figure 5.37: Ellipse fit (mixed model), no restriction.

103
5 Geomatics examples

Example 2: as example 1, but with additional (linear) restriction 𝑔(𝜉) = 0 so that 𝑎ˆ = 𝑏ˆ (best fitting
circle).

𝑔(𝜉) = 𝑔(𝑥 M, 𝑦M, 𝑎, 𝑏) = 𝑎 − 𝑏 = 0

=⇒ 𝑅 = [0, 0, 1, −1], 𝑤𝑅 = 0

c k < 10−12 , k Δ𝑒
Parameters (after 10 iterations: k Δ𝜉 c k < 10−12 ):

𝑥ˆM = 1.119, 𝑦ˆM = −3.921, 𝑎ˆ = 𝑏ˆ = 122.939, 𝑒ˆT𝑊 𝑒ˆ = 815.668

 T
𝑒ˆ𝑥 = −0.009, 0.404, −0.509, −3.992, 13.118, −15.134, 2.798, 3.145, 0.178
 T
𝑒ˆ𝑦 = 0.987, 0.943, −0.480, −0.132, −4.690, −5.318, −1.769, −6.394, 16.854

See figure 5.38.

150
Observed Points
Estimated Centre
Adjusted Points

100

50

0
y/m

-50

-100

-150
-150 -100 -50 0 50 100 150
x/m

Figure 5.38: Ellipse fit (mixed model), circle restriction.

104
 5.3 Mixed model

Example 3: as example 1, but with additional (non-linear) constraint 𝑔(𝜉) so that best fitting el-
lipse passes through the point 𝑥 P = 100, 𝑦P = −100.

(𝑥 P − 𝑥 M ) 2 (𝑦P − 𝑦M ) 2
𝑔(𝜉) = 𝑔(𝑥𝑚 , 𝑦𝑚 , 𝑎, 𝑏) = + −1=0
𝑎2 𝑏2
h𝑥 0 0 0 )2 0 )2 i (𝑥 P − 𝑥 M0 )2 (𝑦P − 𝑦M 0 )2
P −𝑥 M 𝑦P −𝑦M (𝑥 P −𝑥 M (𝑦P −𝑦M
=⇒ 𝑅 = −2 2 2 3 3 , 𝑤 𝑅 = + −1
𝑎0 𝑏0 𝑎0 𝑏0 𝑎 20 𝑏 02

c k < 10−12 , k Δ𝑒
Parameters (after 10 iterations: k Δ𝜉 c k < 10−12 ):

𝑥ˆM = 5.402, 𝑦ˆM = −11.769, 𝑎ˆ = 134.124, 𝑏ˆ = 124.460, 𝑒ˆ′𝑊 𝑒ˆ = 1197.412

 T
𝑒ˆ𝑥 = −0.263, 1.262, −2.361, −18.668, −2.701, −6.996, 2.583, 0.569, 1.191
 T
𝑒ˆ𝑦 = 7.401, 3.985, −2.988, −2.287, 0.966, −2.275, −2.051, −1.336, 26.079

See figure 5.39.

150
Observed Points
Estimated Centre
Adjusted Points
Constrained Point
100

50

0
y/m

-50

-100

-150
-150 -100 -50 0 50 100 150
x/m

Figure 5.39: Ellipse fit (mixed model), point restriction.

105
6 Statistics

6.1 Expectation of sum of squared residuals

n o
E 𝑒ˆT𝑄 𝑦−1𝑒ˆ
Note: 𝑒 T𝑄 𝑦−1𝑒 is the quantity to be minimized.
𝑚 Õ
Õ 𝑚
𝑒ˆT 𝑄 𝑦−1 𝑒ˆ = (𝑃 𝑦 )𝑖 𝑗 𝑒ˆ𝑖 𝑒ˆ 𝑗
1×𝑚 𝑚×𝑚 𝑚×1 𝑖=1 𝑗=1

n o Õ𝑚 Õ
𝑚 n o
=⇒ E 𝑒ˆT𝑄 𝑦−1𝑒ˆ = (𝑃 𝑦 )𝑖 𝑗 E 𝑒ˆ𝑖 𝑒ˆ 𝑗
𝑖=1 𝑗=1
𝑚 Õ
Õ 𝑚
= (𝑃 𝑦 )𝑖 𝑗 (𝑄𝑒ˆ )𝑖 𝑗
𝑖=1 𝑗=1
𝑚 Õ
Õ 𝑚 Õ 
= (𝑃 𝑦 )𝑖 𝑗 (𝑄𝑒ˆ ) 𝑗𝑖 = 𝑃 𝑦 𝑄𝑒ˆ 𝑖𝑖
𝑖=1 𝑗=1 𝑖

= trace(𝑃 𝑦 𝑄𝑒ˆ )
= trace(𝑃 𝑦 (𝑄 𝑦 − 𝑄 𝑦ˆ ))
= trace(𝐼𝑚 − 𝑃 𝑦 𝑄 𝑦ˆ )
= 𝑚 − trace 𝑃 𝑦 𝑄 𝑦ˆ
trace 𝑃 𝑦 𝑄 𝑦ˆ = trace 𝑄 𝑦ˆ 𝑃 𝑦
= trace 𝐴𝑄 𝑥ˆ 𝐴T 𝑃 𝑦
= trace 𝐴(𝐴T 𝑃 𝑦 𝐴) −1𝐴T 𝑃 𝑦
= trace 𝑃𝐴
Linear algebra:
trace 𝑋 = sum of eigenvalues of 𝑋

Q: Eigenvalues of a projector?
𝑃𝐴𝑧 = 𝜆𝑧 (special) eigenvalue problem
 
𝑃𝐴 𝑃𝐴𝑧 = 𝑃𝐴𝑧 = 𝜆𝑧 0
𝜆 2𝑧 = 𝜆𝑧 =⇒ 𝜆(𝜆 − 1)𝑧 = 0 =⇒ 𝜆=
𝑃𝐴 𝑃𝐴𝑧 = 𝜆𝑃𝐴𝑧 = 𝜆 2𝑧 1
=⇒ trace 𝑃𝐴 = number of eigenvalues 1

106
 6.2 Basics

Q: How many eigenvalues 𝜆 = 1?

A:

dim R (𝐴) = 𝑛
 T
E 𝑒ˆ 𝑃 𝑦 𝑒ˆ = 𝑚 − 𝑛 (= 𝑟 redundancy)

6.2 Basics

Random variable: 𝑥
Realization: 𝑥

Wahrschein-
Probability density function (pdf) lichkeitsdichte

(a) Probability density function. (b) Probability calculations by inte- (c) Intervall.
grating over the pdf.

Figure 6.1

∫∞
𝑓 (𝑥) d𝑥 = 1
−∞
Note: not necessarily normal distribution.
∫∞

E 𝑥 =: 𝜇𝑥 = 𝑥 𝑓 (𝑥) d𝑥
−∞
∫∞
 
D 𝑥 =: 𝜎𝑥2 = (𝑥 − 𝜇𝑥 ) 2 𝑓 (𝑥) d𝑥 = E (𝑥 − 𝜇𝑥 ) 2
−∞

Probability calculations by integrating over the pdf.

∫𝑥 0
𝑃 (𝑥 < 𝑥 0 ) = 𝑓 (𝑥) d𝑥
−∞

107
 6 Statistics

Verteilungs-
funktion Cumulative distribution or density function (cdf)

Figure 6.2: Cumulative distribution or density function.

∫𝑥
𝐹 (𝑥) = 𝑓 (𝑦) d𝑦 = 𝑃 (𝑥 < 𝑥)
−∞
e. g.
∫𝑏 ∫𝑏 ∫𝑎
𝑃 (𝑎 ≤ 𝑥 ≤ 𝑏) = 𝑓 (𝑥) d𝑥 = 𝑓 (𝑥) d𝑥 − 𝑓 (𝑥) d𝑥
𝑎 −∞ −∞

= 𝐹 (𝑏) − 𝐹 (𝑎)

6.3 Hypotheses

Assumption or statement which can be statistically tested.

𝐻 : 𝑥 ∼ 𝑓 (𝑥)

Sicherheitswahr-
Assumption: 𝑥 is distributed with given 𝑓 (𝑥).
scheinlichkeit
𝑃 (𝑎 ≤ 𝑥 ≤ 𝑏) = 1 − 𝛼 = confidence level
Irrtumswahr-
scheinlichkeit 𝑃 (𝑥 ∉ [𝑎; 𝑏]) = 𝛼 = significance level
Konfidenzbe-
[𝑎; 𝑏] = confidence region
reich (Annahme-
bereich) [−∞; 𝑎] ∪ [𝑏; ∞] = critical region
kritischer Be-
reich (Ableh- Now: given a realization 𝑥 of 𝑥. If 𝑎 ≤ 𝑥 ≤ 𝑏, there is no reason to reject the hypothesis, otherwise:
nungs-, Verwer- reject hypothesis. E. g.
fungsbereich)
𝑒ˆ = 𝑃𝑎⊥𝑦

𝑄𝑒ˆ = 𝑃𝑎⊥𝑄 𝑦 = 𝑄 𝑦 − 𝑄 𝑦ˆ

108
 6.3 Hypotheses

Figure 6.3: Confidence and significance level.

Example Normal distribution

define 𝑎, 𝑏: determine 𝛼

𝑃 (𝜇 − 𝜎 ≤ 𝑥 ≤ 𝜇 + 𝜎) = 68.3% =⇒ 𝛼 = 0.317
𝑃 (𝜇 − 2𝜎 ≤ 𝑥 ≤ 𝜇 + 2𝜎) = 95.5% =⇒ 𝛼 = 0.045
𝑃 (𝜇 − 3𝜎 ≤ 𝑥 ≤ 𝜇 + 3𝜎) = 99.7% =⇒ 𝛼 = 0.003

Matlab: ♥♦r♠♣❞❢
1 − 𝛼 = 𝐹 (𝑏) − 𝐹 (𝛼) = 𝐹 (𝜇 + 𝑘𝜎) − 𝐹 (𝜇 − 𝑘𝜎)
𝑘 = critical value kritischer Wert

define 𝛼: determine 𝑎, 𝑏

𝑃 (𝜇 − 1.96𝜎 ≤ 𝑥 ≤ 𝜇 + 1.69𝜎) = 95% ⇐= 𝛼 = 0.05 (≈ 2𝜎)

𝑃 (𝜇 − 2.58𝜎 ≤ 𝑥 ≤ 𝜇 + 2.58𝜎) = 99% =⇒ 𝛼 = 0.01
𝑃 (𝜇 − 3.29𝜎 ≤ 𝑥 ≤ 𝜇 + 3.29𝜎) = 99.9% =⇒ 𝛼 = 0.001

Matlab: ♥♦r♠✐♥✈

Rejection of hypothesis
=⇒ an alternative hypothesis must hold

𝐻0 : 𝑥 ∼ 𝑓0 (𝑥) null-hypothesis
𝐻𝑎 : 𝑥 ∼ 𝑓𝑎 (𝑥) alternative hypothesis

109
 6 Statistics

Figure 6.4: Accept or reject hypothesis?

𝐻 0 true 𝐻 0 false
wrong =⇒ type I error
𝑥 ∈𝐾
(false alarm) OK
=⇒ reject 𝐻 0
𝑃 (𝑥 ∈ 𝐾 |𝐻 0 ) = 𝛼
wrong =⇒ type II error
𝑥 ∉𝐾
OK (failed alarm)
=⇒ accept 𝐻 0
𝑃 (𝑥 ∉ 𝐾 |𝐻𝑎 ) = 𝛽

𝛼 = level of significance of test = size of test

Testgüte 𝛾 = 1 − 𝛽 = power of test

6.4 Distributions

Standard normal distribution (univariate)

1 1 2
𝑥 ∼ 𝑁 (0, 1), 𝑓 (𝑥) = √ 𝑒 − 2 𝑥 ,
2𝜋

E 𝑥 = 0,
  
D 𝑥 = E 𝑥2 = 1 ←− 𝑥 2 ∼ 𝜒 2 (1, 0), E 𝑥2 = 1 .

110
 6.4 Distributions

Standard normal (multivariate) → 𝝌 2 -distribution

 
1 T 1
𝑥 ∼ 𝑁 ( 0 , 1), 𝑓 (𝑥) = 𝑘
exp − 𝑥 𝑥 ,
(2𝜋) 2 2
 
𝑘-vector 𝑘-vector

E 𝑥 = 0 ,
  
𝑘-vector 𝑘-vector

D 𝑥 = E 𝑥2 = 1 ←− 𝑥 2 ∼ 𝜒 2 (1, 0), E 𝑥𝑥 T = 𝐼,
𝑥 T𝑥 = 𝑥 21 + 𝑥 22 + . . . + 𝑥 𝑘2 ∼ 𝜒 2 (𝑘, 0),
  
E 𝑥 T𝑥 = E 𝑥 21 + . . . + E 𝑥 𝑘2 = 𝑘 .

Non-standard normal → central 𝝌 2 -distribution

𝜎12
© ª
­ ®
0
𝜎22
𝑥 ∼ 𝑁 (0, 𝑄 𝑥 ), 𝑄 𝑥 = ­­ ®,
®
­ ®
..
.
0
« 𝜎𝑘2 ¬
 
2 1 1 𝑥 𝑖2
𝑥 𝑖 ∼ 𝑁 (0, 𝜎𝑖 ), 𝑓 (𝑥 𝑖 ) = √ exp − 2 ,
2𝜋𝜎𝑖 2 𝜎𝑖
𝑥
𝑦 = 𝑖 ∼ 𝑁 (0, 1),
𝑖 𝜎𝑖
𝑥 12 𝑥 22 𝑥𝑘2 
𝑥 T𝑄 𝑥−1𝑥 = + +...+ ∼ 𝜒 2 (𝑘, 0) =⇒ E 𝑥 T𝑄 𝑥−1𝑥 = 𝑘,
𝜎12 𝜎22 𝜎𝑘2
1 1
𝑓 (𝑥) = 𝑘
exp (− 𝑥 T𝑄 𝑥−1𝑥) .
1
(2𝜋) (det 𝑄 𝑥 )
2 2 2

The same is true when

𝑥 ∼ 𝑁 (0, 𝑄 𝑥 ) with 𝑄 𝑥 full matrix .

Non-standard normal → non-central 𝝌 2 -distribution

𝑥 ∼ 𝑁 (𝜇, 𝐼 ) =⇒ 𝑥 T𝑥 ∼ 𝜒 2 (𝑘, 𝜆),


E 𝑥 =𝜇

E 𝑥 T𝑥 = 𝑘 + 𝜆; 𝜇 T 𝜇 = non-centrality parameter
= 𝜇 12 + 𝜇 22 + . . . + 𝜇𝑘2

111
6 Statistics

c2(k,0)
N(0,1) N(m,1)

c2(k,l)
1

0 m k k+l

Figure 6.5: Central/Non-central normal and 𝜒 2 -distribution

General case

𝑥 ∼ 𝑁 (𝜇, 𝑄 𝑥 ),
 
1 1 −1
𝑓 (𝑥) = 𝑘 1
exp − (𝑥 − 𝜇)𝑄 𝑥 (𝑥 − 𝜇) ,
(2𝜋) 2 (det 𝑄 𝑥 ) 2 2
 
E 𝑥 = 𝜇, D 𝑥 = 𝑄𝑥 ,
 T −1
E 𝑥 𝑄𝑥 𝑥 = 𝑘 + 𝜆 = 𝜇 T𝑄 𝑥−1 𝜇 .

112
7 Statistical Testing

7.1 Global model test: a first approach

Statistics of estimated residuals

𝑒ˆ = 𝑦 − 𝑦ˆ
= 𝑃𝐴⊥𝑦

E 𝑒ˆ = 0, 𝑒ˆ ∼ 𝑁 (0, 𝑄𝑒ˆ )

D 𝑒ˆ = 𝑄𝑒ˆ
= 𝑄 𝑦 − 𝑄 𝑦ˆ
T
= 𝑃𝐴⊥𝑄 𝑦 (𝑃𝐴⊥ )

 T −1
Question: 𝑒ˆT𝑄𝑒−1 ˆ 2
ˆ 𝑒 ∼ 𝜒 (𝑚, 0) and thus E 𝑒 𝑄 𝑒ˆ 𝑒 = 𝑚?
ˆ ˆ
No, because 𝑄𝑒ˆ is singular and therefore not invertible. However, in 6.1:
n o 
E 𝑒ˆT𝑄 𝑦−1𝑒ˆ = trace(𝑄 𝑦−1 E 𝑒ˆ𝑒ˆT ) = trace(𝑄 𝑦−1 (𝑄 𝑦 − 𝑄 𝑦ˆ )) = 𝑚 − 𝑛
| {z }
𝑄𝑒ˆ

Test statistic
As residuals tell us something about the mismatch between data and model, they will be the basis
for our testing. In particular the sum of squared estimated residuals will be used as our test statistic
𝑇:

𝑇 = 𝑒ˆT𝑄 𝑦−1𝑒ˆ ∼ 𝜒 2 (𝑚 − 𝑛, 0)

E 𝑇 = 𝑚 −𝑛

Thus, we have a test statistic and we know its distribution. This is the starting point for global model
testing.

𝑻 > 𝒌 𝜶 : reject 𝑯 0
In case𝑇 – the realization of𝑇 – is larger than a chosen critical value (based on 𝛼), the null hypothesis
𝐻 0 should be rejected. At this point, we haven’t formulated an alternative hypothesis 𝐻 a yet. The
rejection may be due to:

113
7 Statistical Testing

Figure 7.1: Distribution of the test statistic 𝑇 under the null and alternative hypotheses. (Non-
centrality parameter 𝜆 to be explained later)

• error in the (deterministic) observation model 𝐴,


• measurement error: E 𝑒 ≠ 0,

• wrong assumptions in the stochastic model: D 𝑒 ≁ 𝑄 𝑦 .

Variance of unit weight

A possible error in the stochastic model would be a wrong scale factor. Let us write 𝑄 𝑦 = 𝜎 2𝑄 and
see how an unknown variance factor 𝜎 2 propagates through the various estimates:

𝑥ˆ = (𝐴T𝑄 𝑦−1𝐴) −1𝐴T𝑄 𝑦−1𝑦

𝑄 𝑥ˆ = (𝐴T𝑄 𝑦−1𝐴) −1
 𝑄 𝑦 = 𝜎 2𝑄

y 𝑃 𝑦 = 𝑄 −1 = 𝜎 −2𝑄 −1

𝑦

𝑥ˆ = (𝐴T𝜎 −2𝑄 −1𝐴) −1𝐴T𝜎 −2𝑄 −1𝑦

= 𝜎 2 (𝐴T𝑄 −1𝐴) −1𝐴T𝜎 −2𝑄 −1𝑦

= (𝐴T𝑄 −1𝐴) −1𝐴T𝑄 −1𝑦 =⇒ independent on 𝜎 2

𝑄 𝑥ˆ = 𝜎 2 (𝐴T𝑄 −1𝐴) −1 =⇒ depending on 𝜎 2

Thus, the estimate 𝑥ˆ is independent of the variance factor and therefore insensitive to stochastic
model errors. However, the covariance matrix 𝑄 𝑥ˆ is scaled by the variance factor. This is also

114
 7.2 Testing procedure

true for functions 𝑓ˆ = 𝐹 (𝑥):

ˆ while 𝑓ˆ is not influenced by 𝜎 2 , its covariance-matrix 𝑄 𝑓ˆ is changed
accordingly. How about the test statistic 𝑇 ?
n o 
E 𝑒ˆT𝑄 𝑦−1𝑒ˆ = E 𝜎 −2𝑒ˆT𝑄 −1𝑒ˆ = 𝑚 − 𝑛

=⇒ E 𝑒ˆT𝑄 −1𝑒ˆ = 𝜎 2 (𝑚 − 𝑛)

Alternative test statistic

This leads to a new test statistic:
𝑒ˆT𝑄 −1𝑒ˆ 
𝜎ˆ 2 = =⇒ E 𝜎ˆ 2 = 𝜎 2 ,
𝑚 −𝑛
which shows that 𝜎ˆ 2 is an unbiased estimate of 𝜎 2 . unverzerrte
Schätzung
If we consider 𝑄 as the a priori variance-covariance matrix, then 𝑄ˆ 𝑦 = 𝜎ˆ 2 is the a posteriori one.
Now consider the ratio between a posteriori and a priori variance as an alternative test statistic:
T
𝜎ˆ 2 𝑒ˆT𝜎 −2𝑄 −1𝑒ˆ 𝑒ˆ 𝑄 𝑦−1𝑒ˆ 𝜒 2 (𝑚 − 𝑛, 0)
= = ∼ = 𝐹 (𝑚 − 𝑛, ∞, 0)
𝜎2 𝑚 −𝑛 𝑚 −𝑛 𝑚 −𝑛
( )
The ratio has a so-called Fisher distribution.
𝜎ˆ 2
E 2 =1
𝜎

7.2 Testing procedure

Null hypothesis and alternative hypothesis

n o n o
If the null hypothesis is described by E 𝑦 = 𝐴𝑥, D 𝑦 = 𝑄 𝑦 , and if we assume that our stochastic
model is correct, then we formulate an alternative hypothesis by augmenting the model. We will
add 𝑞 new parameters ∇ (which is not an operator here). Consequently we will need a design matrix
𝐶 for ∇.
𝑚×𝑞

𝐻0 𝐻a
n o n o n o n o
E 𝑦 = 𝐴𝑥; D 𝑦 = 𝑄 𝑦 E 𝑦 = 𝐴𝑥 + 𝐶∇; D 𝑦 = 𝑄 𝑦
  𝑥 
= 𝐴𝐶
∇
↓ ↓
𝑥ˆ 0 ˆ
𝑥ˆ , ∇ a
↓ ↓
𝑦ˆ = 𝐴𝑥ˆ 0 𝑦ˆ = 𝐴𝑥ˆ a + 𝐶 ∇ˆ
0 a
↓ ↓
𝑒ˆ 0 𝑒ˆ a
↓ ↓
𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0 ∼ 𝜒 2 (𝑚 − 𝑛) 𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a ∼ 𝜒 2 (𝑚 − 𝑛 − 𝑞)

115
7 Statistical Testing

𝐻 a more parameters =⇒ sum of squared residuals smaller

𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a < 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0

=⇒ difference, which is a measure of improvement as test statistic:

𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0 − 𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a First version

How is it distributed?
𝐻 0 : 𝑇 ∼ 𝜒 2 (𝑞, 0) and 𝐻 a : 𝑇 ∼ 𝜒 2 (𝑞, 𝜆)

116
 7.2 Testing procedure

Geometry of 𝑯 0 und 𝑯 a

(a) (b)

(c)

Figure 7.2: Hypotheses 𝐻 0 and 𝐻𝑎

ˆ
𝑦ˆ = 𝐴𝑥ˆ a + 𝐶 ∇
a
ˆ +𝑃 ⊥𝐶 ∇
= 𝐴𝑥ˆ a + 𝑃𝐴𝐶 ∇ ˆ
| {z }
𝐴

𝑦ˆ
0

ˆ
=⇒ 𝑦ˆ − 𝑦ˆ = 𝑃𝐴⊥𝐶 ∇
a 0

ˆ T𝑄 𝑦−1 𝑃 ⊥𝐶 ∇
=⇒ 𝑇 = 𝑃𝐴⊥𝐶 ∇ ˆ T𝐶 T
ˆ =∇ T
𝑃𝐴⊥ 𝑄 𝑦−1 𝑃𝐴⊥ ˆ
𝐶∇
| {z }
𝐴

⊥ =𝑄 −1𝑄 𝑄 −1
=𝑄 𝑦−1 𝑃𝐴 𝑦 𝑒ˆ0 𝑦

ˆ T𝐶 T𝑄 𝑦−1𝑄𝑒ˆ 𝑄 𝑦−1𝐶 ∇
=∇ ˆ Second version
0

𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0 − 𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a
= (𝑦ˆ − 𝑦ˆ ) T𝑄 𝑦−1 (𝑦ˆ − 𝑦ˆ ) Third version
0 a 0 a

ˆ T𝐶 T𝑄 𝑦−1𝑄𝑒ˆ 𝑄 𝑦−1𝐶 ∇
=∇ ˆ
0

117
7 Statistical Testing

All versions of 𝑇 require adjustment under 𝐻 a

ˆ
(ˆ𝑒 a, 𝑦ˆ , ∇)
a

Now: Version only with 𝑒ˆ 0 and 𝐶

Normal equations under 𝑯 0, 𝑯 a

𝐻0 : 𝐴T𝑄 𝑦−1𝐴𝑥ˆ 0 = 𝐴T𝑄 𝑦−1𝑦

 T    𝑥ˆ   𝐴T 
𝐴 −1 −1
𝐻a : 𝑄𝑦 𝐴 𝐶 ˆ = 𝐶T 𝑄𝑦 𝑦
𝐶T ∇
𝐴T𝑄 −1𝐴 𝐴T𝑄 𝑦−1𝐶   !
© 𝑛×𝑛𝑦 ª 𝑥ˆ
⇐⇒ ­­ T −1 ® a =
𝐴 T𝑄 −1𝑦

® ∇
𝑛×𝑞 𝑦
𝐶 𝑄 𝑦 𝐴 𝐶 𝑄 𝑦−1𝐶
T ˆ 𝐶 T𝑄 𝑦−1𝑦
« 𝑞×𝑛 𝑞×𝑞 ¬

1. row: solve for 𝑥ˆ a

ˆ = 𝐴T𝑄 𝑦−1𝐴𝑥ˆ
𝐴T𝑄 𝑦−1𝐴𝑥ˆ a + 𝐴T𝑄 𝑦−1𝐶 ∇ 0
ˆ
=⇒ 𝑥ˆ a = 𝑥ˆ 0 − (𝐴T𝑄 𝑦−1𝐴) −1𝐴T𝑄 𝑦−1𝐶 ∇
ˆ
=⇒ 𝐴𝑥ˆ a = 𝐴𝑥ˆ 0 − 𝑃𝐴𝐶 ∇
ˆ = 𝐴𝑥ˆ + (𝐼 − 𝑃𝐴 )𝐶 ∇
=⇒ 𝐴𝑥ˆ a + 𝐶 ∇ ˆ
0
ˆ
=⇒ 𝑦ˆ = 𝑦ˆ + 𝑃𝐴⊥𝐶 ∇
a 0

ˆ −→ laborious derivation!
2. row: substitute 𝑥ˆ a and solve for ∇
Result:
ˆ = (𝐶 T𝑄 𝑦−1𝑄𝑒ˆ 𝑄 𝑦−1𝐶) −1𝐶 T𝑄 𝑦−1𝑒ˆ
∇ 0 0

ˆ in second version of 𝑇 −→ Fourth version

Substitute ∇

𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝐶 (...) −1 (...) (...) −1𝐶 T𝑄 𝑦−1𝑒ˆ 0

= 𝑒ˆ T0 𝑄 𝑦−1𝐶 (𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶) −1𝐶 T𝑄 𝑦−1𝑒ˆ 0

Distribution of 𝑇
Transformation of variables

𝑧 = 𝐶 T 𝑄 𝑦−1 𝑒ˆ 0
𝑞×1 𝑞×𝑚 𝑚×𝑚 𝑚×1

𝑄𝑧 = 𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶
ˆ = 𝑄𝑧−1𝑧 =⇒ 𝑧 = 𝑄𝑧 ∇
∇ ˆ

𝑇 = 𝑧 T𝑄𝑧−1𝑧 ∼ 𝜒𝑞2

118
 7.2 Testing procedure

𝐻0 𝐻a
𝑧 ∼ 𝑁 (0, 𝑄𝑧 ) ˆ 𝑄𝑧 )
𝑧 ∼ 𝑁 (𝑄𝑧 ∇,
𝑇 ∼ 𝜒 2 (𝑞, 0) 2
𝑇 ∼ 𝜒 (𝑞, 𝜆)
𝜆 = ∇T𝑄𝑧 𝑄𝑧−1𝑄𝑧 ∇ = ∇T𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶∇

Summary
Test quantity 𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0 − 𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a exhibits that 𝐻 0 has to be rejected in favor of 𝐻 a : Model
E {𝑦} = 𝐴𝑥 is not suitable. In case 𝐻 0 is true 𝑇 is (central) 𝜒 2 -distributed with 𝑞 degrees of freedom,
2 , otherwise 𝑇 ∼ 𝜒 2 with 𝜆 being the non-centrality parameter 𝜆 = ∇T𝐶 T𝑄 −1𝑄 𝑄 −1𝐶∇.
𝑇 ∼ 𝜒𝑞,0 𝑞,𝜆 𝑦 𝑒ˆ0 𝑦

Five alternative versions for 𝑇

(1) 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0 − 𝑒ˆ Ta 𝑄 𝑦−1𝑒ˆ a

(2) (𝑦ˆ − 𝑦ˆ ) T𝑄 𝑦−1 (𝑦ˆ − 𝑦ˆ )
0 a 0 a

ˆT T
(3) ∇ 𝐶 𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶 ∇ ˆ
  −1
(4) 𝑒ˆ T0 𝑄 𝑦−1𝐶 𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶 𝐶 T𝑄 𝑦−1𝑒ˆ 0

(5) 𝑧 T𝑄𝑧−1𝑧; 𝑧 := 𝐶 T𝑄 𝑦−1𝑒ˆ 0 ; 𝑄𝑧 = 𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶

Versions (1)–(3) explicitely involve the computation of 𝐻 a while cases (4) and (5) require only 𝑒ˆ 0 and
some 𝐶.
For the reason that

𝑧 ∼ 𝑁 (0, 𝑄𝑧 ) under 𝐻 0
𝑧 ∼ 𝑁 (𝑄𝑧 ∇, 𝑄𝑧 ) under 𝐻 a

test quantity 𝑇 is distributed

𝑇 ∼ 𝜒 2 (𝑞, 0) under 𝐻 0
𝑇 ∼ 𝜒 2 (𝑞, 𝜆), 𝜆 = ∇ T𝑄 𝑧 ∇ under 𝐻 a

Question: How is the minimal/maximal number of additional parameters ∇?

Answer: Total number of all parameters 𝑥 and ∇ is 𝑛 + 𝑞 which must not exceed number of obser-
vations 𝑚 =⇒

𝑛 + 𝑞 ≤ 𝑚 =⇒ 0 < 𝑞 ≤ 𝑚 − 𝑛

119
7 Statistical Testing

Case (i) 𝒒 = 𝒎 − 𝒏: global model test

!
rank(𝐴|𝐶) = 𝑛 + 𝑞 = 𝑛 + (𝑚 − 𝑛) = 𝑚
=⇒ 𝑜 (𝐴|𝐶) = 𝑚 × 𝑛 + 𝑞 = 𝑚 × 𝑚 “quadratic”
=⇒ redundancy = 𝑚 − 𝑛 − 𝑞 = 0
=⇒ 𝑒ˆ a = 0
=⇒ 𝑦ˆ = 𝑦
a

=⇒ 𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝑒ˆ 0

Figure 7.3: Test quantity 𝑇

n o n o
𝐻 0 : E 𝑦 = 𝐴𝑥 versus 𝐻 a : E 𝑦 ∈ R𝑚

2
𝑇 ∼ 𝜒𝑚−𝑛,0
2
𝑇 ∼ 𝜒𝑚−𝑛,𝜆 , 𝜆 = ∇ T𝑄 𝑧 ∇

For the reason that 𝑒ˆ a = 0, it is obviously not necessary to specify any matrix 𝐶. The test can always
be carried out, that is why it is called overal model test or global test.

120
 7.3 DIA-Testprinciple

Case (ii) 𝑞 = 1: data snooping

=⇒ 𝐶 is an 𝑚 × 1-vector, ∇ a scalar
  −1
𝑇 = 𝑒ˆ T0 𝑄 𝑦−1𝐶 𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶 𝐶 T𝑄 𝑦−1𝑒ˆ 0
 2
𝑒ˆ T0 𝑄 𝑦−1𝐶
= T −1
𝐶 𝑄 𝑦 𝑄𝑒ˆ0 𝑄 𝑦−1𝐶
ˆ2
∇
=
−1
𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶

∇ˆ2
= 2
𝜎ˆ
∇
𝐶 T𝑄 𝑦−1𝑒ˆ
ˆ2 =
∇
𝐶 T𝑄 𝑦−1𝑄𝑒ˆ0 𝑄 𝑦−1𝐶

Important application:
Detection of a gross error (outlier, blunder) in the observations, which leads to a wrong model
specification.

n o n o
𝐻 0 : E 𝑦 = 𝐴𝑥 𝐻 a : E 𝑦 = 𝐴𝑥 + 𝐶∇

 T
 
 
𝐶 = 0, 0, . . . , 1 , 0, . . . , 0
 |{z} 
 
 position i 

ˆ2
∇ p ∇ˆ √ p ∇ˆ √
Reject 𝐻 0 if 𝑇 = > 𝑘𝛼 or if 𝑇 = < − 𝑘𝛼 and 𝑇 = ˆ can be positive or negative)!
> 𝑘 𝛼 (∇
𝜎 2ˆ 𝜎∇ˆ 𝜎∇ˆ
∇

Should 𝐻 0 be rejected, observation 𝑦𝑖 must be checked and corrected, discarded or even be remea-
sured. The test is performed for every 𝑖 = 1, . . . , 𝑚 if necessary in an iterative manner. The test is
called data snooping. For a diagonal matrix 𝑄 𝑦 we get
p 𝑒ˆ
𝑇 = 𝑖
𝜎𝑒ˆ𝑖

p p p
𝐻 0 : 𝑇 ∼ 𝑁 (0, 1) 𝐻 a : 𝑇 ∼ 𝑁 (∇ 𝑇 , 1)
p q
with ∇ 𝑇 = 𝐶 T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝐶∇

7.3 DIA-Testprinciple

DIA ⇐⇒ Detection, Identification, Adaptation

121
 7 Statistical Testing

Detection: Check the overall validity of 𝐻 0 , perform the overall model test, answer the question
whether or not we have generally to expect any model error, e. g. an outlier in the data, search
for a possible model misspecification.
Identification: Perform data snooping in order to locate a possible gross error. Identify it in the
collection of observations. Screen each individual observation for the presence of a blunder.
Adaptation: React to the outcomes of detection and identification step. Perform a corrective action
in order to get the null hypothesis accepted. Repair, replace or discard the corrupted obser-
vation. Remeasure part of the observations or change the model in order to account for the
identified model errors.

Question: How to ensure consistent testing parameters? How can we avoid the situation of a conflict
between the overall model test in the detection step and individual test of the identification step?
Answer: Consistency is guaranteed if the probability of detecting an outlier under the alternative
hypothesis with 𝑞 = 1 (data snooping) is the same for the global test. Thus, both tests must use the
same 𝛾 = 1 − 𝛽, which is called 𝛾 0 here.

𝜆0 = 𝜆 (𝛼, 𝑞 = 𝑚 − 𝑛, 𝛾 = 𝛾 0 ) = 𝜆 (𝛼 1, 𝑞 = 1, 𝛾 = 𝛾 0 )

𝑞 = 1: 
𝛾0 = 1 − 𝛽0
=⇒ 𝜆(−→ 𝜇) = 𝜆0
𝛼1
𝑞 = 𝑚 − 𝑛: 
𝜆0
=⇒ 𝛼 = 𝛼𝑚−𝑛
𝛾0 = 1 − 𝛽0

e. g.: 𝛼 1 = 1% (usually small), 𝛽 0 = 20% =⇒ 𝛼𝑚−𝑛 ≈ 30%

7.4 Internal reliability

Which model error 𝐶∇ results in the power of test 𝛾 0 ? Or the other way around: Which model error
innere 𝐶∇ can be just detected with probability 𝛾 0 ? This question is discussed in the framework of internal
Zuverlässigkeit reliability.

Analysis 𝝀

𝜆 = ∇T𝐶 T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝐶∇

𝑄𝑒ˆ = 𝑄 𝑦 − 𝑄 𝑦ˆ
  −1
= 𝑄 𝑦 − 𝐴 𝐴T𝑄 𝑦−1𝐴 𝐴T
   −1 
T T −1 −1 T −1 T −1
|{z}
=⇒ 𝜆 = ∇ 𝐶 𝑄 𝑦 − 𝑄 𝑦 𝐴 𝐴 𝑄 𝑦 𝐴 𝐴 𝑄 𝑦 𝐶∇
(𝐶 ∇) T

122
 7.4 Internal reliability

Question: For given fixed 𝜆 = 𝜆0 , how can 𝐶∇ be manipulated?

Using 𝑸 𝒚 :

“better” observations =⇒ 𝑄 𝑦 smaller

=⇒ 𝑄 𝑦−1 larger
=⇒ 𝐶∇ smaller (in order to keep 𝜆 = 𝜆0 constant)

=⇒ the more precise the observations are, the smaller the model error 𝐶∇ may be. It will be detected
with probability 𝛾 0 .

Using 𝑨:
• more observations =⇒ larger redundancy
=⇒ for a constant 𝐶∇: 𝜆 increases or the other way around for a constant 𝜆, 𝐶∇ gets smaller
• better network design, better configuration, improved distribution of observations, avoid bad
geometries in resection problems
=⇒ 𝐶∇ can be decreased

Minimum Detectable Bias (MDB)

kleinster
n o n o aufdeckbarer
𝛿𝑦 := 𝐶 ∇ = E 𝑦|𝐻 a − E 𝑦|𝐻 0 Fehler
𝑚×1 𝑚×𝑞 𝑞×1

𝛿𝑦 describes the internal reliability; it measures the smallest possible error which can be detected
with probability 𝛾.
Question: How can ∇ be determined from 𝜆0 = (𝐶∇) T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝐶∇?

Case 𝒒 = 1 (datasnooping):
∇ is a scalar, 𝐶 = 𝑐𝑖 , 𝛿𝑦𝑖 = 𝑐𝑖 ∇

𝜆0 = 𝑐𝑖 T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝑐𝑖 ∇2
s
𝜆0
=⇒ |∇𝑖 | =
𝑐𝑖 𝑄 𝑦 𝑄𝑒ˆ𝑄 𝑦−1𝑐𝑖
T −1

|∇𝑖 | = minimal detectable bias

123
7 Statistical Testing

Assumption: 𝑄 𝑦 is diagonal
h i
𝑐𝑖 T𝑄 𝑦−1 = 0, 0, . . . , 𝜎 𝑦−2𝑖 , 0, . . .
1×𝑚

=⇒ 𝑐𝑖 T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝑐𝑖 = 𝜎 𝑦−2𝑖 [𝑄𝑒ˆ ] 𝑖𝑖 𝜎 𝑦−2𝑖 = 𝜎 𝑦−4𝑖 𝜎𝑒ˆ−2

 
𝑖

= 𝜎 𝑦−4𝑖 𝜎 𝑦2𝑖 − 𝜎 𝑦2ˆ𝑖

!
𝜎 𝑦2ˆ𝑖
−2
= 𝜎 𝑦𝑖 1 − 2
𝜎 𝑦𝑖
√
=⇒ |∇𝑖 | = r
𝜆0
 
𝜎 𝑦−4𝑖 𝜎 𝑦2𝑖 − 𝜎 𝑦2ˆ𝑖
√
= 𝜎 𝑦𝑖 r
𝜆0
𝜎 𝑦2ˆ
𝑖
1− 𝜎 𝑦2 𝑖
√
𝜆0
= 𝜎 𝑦𝑖 √
𝑟𝑖

a) If no improvement through adjustment

𝜎 𝑦ˆ𝑖 = 𝜎 𝑦𝑖 =⇒ |∇𝑖 | = ∞

√
b) If 𝜎 𝑦ˆ𝑖 ≪ 𝜎 𝑦𝑖 : |∇𝑖 | = 𝜎 𝑦𝑖 𝜆0 is detectable

Local redundancy
𝜎 𝑦2ˆ𝑖
𝑟𝑖 = 1 − = local redundancy number
𝜎 𝑦2𝑖

𝑦𝑖 poorly controlled −→ 0 ≤ 𝑟𝑖 ≤ 1 ←− well controlled

Õ
𝑚
𝑟𝑖 = 𝑚 − 𝑛
𝑖=1

 
𝑟𝑖 = 𝑐𝑖 T
𝐼 − 𝑄 𝑦ˆ 𝑄 𝑦−1 𝑐𝑖

= 𝑐𝑖 T (𝐼 − 𝑃𝐴 ) 𝑐𝑖
= 𝑐𝑖 T 𝑃𝐴⊥𝑐𝑖
Õ
=⇒ 𝑟𝑖 = trace𝑃𝐴⊥ = 𝑚 − 𝑛
𝑖
n o
NB.: E 𝑒ˆ T𝑄 𝑦−1𝑒ˆ T = 𝑚 − 𝑛

124
 7.5 External reliability

Mean local redundancy number

Í
𝑚
s
𝑟𝑖
𝑖=1 𝑚 −𝑛 𝜆
𝑟¯ = = =⇒ | ∇¯𝑖 | = 𝜎 𝑦𝑖 𝑚−𝑛
𝑚 𝑚 𝑚

Redundancy

𝑒ˆ = 𝑃𝐴⊥𝑦
   −1 
T −1 T −1
= 𝐼 − 𝐴 𝐴 𝑄𝑦 𝐴 𝐴 𝑄𝑦 𝑦

= 𝑅𝑦 𝑅 = redundancy matrix
𝑒ˆ𝑖 = 𝑅𝑖 𝑗 𝑦
𝑗
= 𝑟𝑖 𝑦 + ...
𝑖
=⇒ 𝛿 𝑒ˆ = 𝑟𝑖 𝛿𝑦
𝑖

=⇒ Local redundancy is a quantity how redundancy is distributed among the single observations
or how a model error 𝛿𝑦 = 𝐶∇ is projected onto the residuals.

7.5 External reliability

How does an undetected error corrupt the adjustment results?

𝛿𝑦 := 𝐶∇ −→ 𝛿 𝑥?
ˆ

𝑥ˆ = (𝐴T𝑄 𝑦−1𝐴) −1𝐴T𝑄 𝑦−1𝑦

ˆ = (𝐴T𝑄 𝑦−1𝐴) −1𝐴T𝑄 𝑦−1 (𝑦 + 𝛿𝑦)

(𝑥ˆ + 𝛿 𝑥)

𝛿 𝑥ˆ = (𝐴T𝑄 𝑦−1𝐴) −1𝐴T𝑄 𝑦−1𝛿𝑦

Problems:
• 𝛿 𝑥ˆ is a vector-valued quantity
• 𝛿 𝑥ˆ depends on possibly inhomogenous quantities with different physical units.
Remedy: Normalize 𝛿 𝑥ˆ using 𝑄 𝑥−1
ˆ =⇒ squared bias-to-noise-ratio

𝜆𝑥ˆ = 𝛿 𝑥ˆ T𝑄 𝑥−1
ˆ 𝛿 𝑥ˆ

= 𝛿 𝑥ˆ T𝐴T𝑄 𝑦−1𝐴𝛿 𝑥ˆ
= (𝑃𝐴𝛿𝑦) T𝑄 𝑦−1 (𝑃𝐴𝛿𝑦)
2
= k𝑃𝐴𝛿𝑦 k𝑄 −1
𝑦

125
7 Statistical Testing

𝜆𝑥ˆ : large =⇒ large influence of a model error 𝛿𝑦

𝜆𝑥ˆ : small =⇒ insignificant influence of a model error 𝛿𝑦

7.6 Reliability: a synthesis

𝛿𝑦 = 𝐼𝛿𝑦 = (𝑃𝐴 + 𝑃𝐴⊥ )𝛿𝑦 = 𝑃𝐴𝛿𝑦 + 𝑃𝐴⊥𝛿𝑦

⇓
2 2 ⊥ 2
k𝛿𝑦 k𝑄 −1 = k𝑃𝐴 𝛿𝑦 k 𝑄 −1 + k𝑃𝐴 𝛿𝑦 k 𝑄 −1
𝑦 𝑦 𝑦

T T T T T
or 𝛿𝑦 𝑄 𝑦−1𝛿𝑦 = 𝛿 𝑥ˆ 𝐴 𝑄 𝑦−1𝐴𝛿 𝑥ˆ + 𝛿𝑦 (𝑃𝐴⊥ ) 𝑄 𝑦−1 𝑃𝐴⊥𝛿𝑦

or 𝜆 𝑦 = 𝜆𝑥ˆ + 𝜆 0

=⇒ 𝜆𝑥ˆ = 𝜆 𝑦 − 𝜆 0

Question: Why is k𝑃𝐴⊥𝛿𝑦 k𝑄

2
−1 = 𝜆 0 ?
𝑦

Answer:
𝜆 0 = 𝛿𝑦 T𝑄 𝑦−1𝑄𝑒ˆ𝑄 𝑦−1𝛿𝑦 = 𝛿𝑦 T𝑄 𝑦−1 𝑃𝐴⊥𝛿𝑦
T
= 𝛿𝑦 T (𝑃𝐴⊥ ) 𝑄 𝑦−1 𝑃𝐴⊥𝛿𝑦
T
= (𝑃𝐴⊥𝛿𝑦) 𝑄 𝑦−1 𝑃𝐴⊥𝛿𝑦
= k𝑃𝐴⊥𝛿𝑦 k𝑄
2
−1
𝑦

special case
𝑞 = 1, 𝑐𝑖 , 𝑄 𝑦 = diagonal

=⇒ 𝜆𝑥ˆ = 𝜆 𝑦𝑖 − 𝜆 0
1
= 𝜆 − 𝜆0
𝑟𝑖 0
1 − 𝑟𝑖
= 𝜆
𝑟𝑖 0
𝜎 𝑦2ˆ𝑖 𝜎 𝑦2𝑖
= 𝜆0
𝜎 𝑦2𝑖 (𝜎 𝑦2𝑖 − 𝜎 𝑦2ˆ𝑖 )
𝜎 𝑦2ˆ𝑖
= 𝜆0
𝜎 𝑦2𝑖 − 𝜎 𝑦2ˆ𝑖
1
= 𝜆0
𝜎 𝑦2 𝑖
𝜎 𝑦2ˆ
−1
𝑖

126
 7.6 Reliability: a synthesis

Figure 7.4: Decomposition of 𝜆𝑦

127
 8 Recursive estimation

8.1 Partitioned model

( !)   ( !)  
n o 𝑦 𝐴1 𝑦 𝑄1 0
E 𝑦 =E 1 = 𝑥; D 1 =
𝑦 𝐴2 𝑦 0 𝑄2
2 2
(𝑚 1 +𝑚 2 )×1 (𝑚 1 +𝑚 2 )×𝑛 (𝑚 1 +𝑚 2 )×(𝑚 1 +𝑚 2 )

8.1.1 Batch / offline / Staple / standard

    −1
𝑥ˆ (1) = 𝐴1 T𝑄 1−1𝐴1 𝐴1 T𝑄 1−1𝑦 , 𝑄 𝑥ˆ (1) = 𝐴1 T𝑄 1−1𝐴1
  −1  
1

𝑥ˆ (2) = 𝐴1 T𝑄 1−1𝐴1 + 𝐴2 T𝑄 2−1𝐴2 𝐴1 T𝑄 1−1𝑦 + 𝐴2 T𝑄 2−1𝑦 ,

  −1
1 2

𝑄 𝑥ˆ (2) = 𝐴1 T𝑄 1−1𝐴1 + 𝐴T2 𝑄 2−1𝐴2

8.1.2 Recursive / sequential / real-time

n o n o
E 𝑦 = 𝐴1𝑥 D 𝑦 = 𝑄1
1 1

(
−1
𝑥ˆ (1) = 𝐴1 T𝑄 1−1𝐴1 𝐴1 T𝑄 1−1𝑦
=⇒
−1 1
𝑄 𝑥ˆ (1) = 𝐴1 T𝑄 1−1𝐴1

−1  T −1 



T −1 T −1 T −1
ˆ
𝑥 𝐴 𝑄 𝐴 + 𝐴 𝑄 𝐴 𝐴 𝑄 𝑦 + 𝐴 𝑄 𝑦

=
  −1  
1 1 2 2 1


(2) 1 2 1 1 2 2 2
T −1 T −1
= 𝑄 𝑥−1 𝑄 𝑥−1
 ˆ (1) + 𝐴2 𝑄 2 𝐴2 ˆ (1) 𝑥 (1) + 𝐴2 𝑄 2 𝑦 2
ˆ

  

 𝑄 𝑥ˆ (2) = 𝑄 −1 + 𝐴2 T𝑄 2−1𝐴2
−1
 𝑥ˆ (1)


Aufdatierungs-
measurement update
=⇒
gleichungen covariance update

This is also the solution of the problem

( !) 
 ( !)  
𝑥ˆ (1) 𝐼 𝑥ˆ (1) 𝑄 𝑥ˆ (1) 0
E = 𝑥; D =
𝑦 𝐴2 𝑦 0 𝑄2
2 2

128
 8.1 Partitioned model

8.1.3 Recursive formulation

Intuitively, it would be nice to have something like 𝑥ˆ (2) = 𝑥ˆ (1) + . . .

=⇒ Solve
𝑄 𝑥−1 −1 T −1
ˆ (2) = 𝑄 𝑥ˆ (1) + 𝐴2 𝑄 2 𝐴2

for 𝑄 𝑥−1
ˆ
(1)

=⇒ 𝑄 𝑥−1 −1 T −1
ˆ (1) = 𝑄 𝑥ˆ (2) − 𝐴2 𝑄 2 𝐴2

  −1  
T −1 T −1
Substitute the result in 𝑥ˆ (2) = 𝑄 𝑥−1
ˆ + 𝐴 2 𝑄 2 𝐴 2 𝑄 −1 𝑥ˆ
𝑥ˆ (1) + 𝐴 2 𝑄 2 𝑦
(1) (1) 2
 
=⇒ 𝑥ˆ (2) = 𝑄 𝑥ˆ (2) 𝑄 𝑥−1 ˆ
𝑥
ˆ (2) (1) − 𝐴 2
T −1
𝑄 2 𝐴 ˆ
𝑥
2 (1) + 𝐴 2
T −1
𝑄 𝑦
2 2
 
= 𝑥ˆ (1) + 𝑄 𝑥ˆ (2) 𝐴2 T𝑄 2−1 𝑦 − 𝐴2𝑥ˆ (1)
2
= 𝑥ˆ (1) + 𝐾𝑣 2
𝑣 2 = 𝑦 − 𝐴2𝑥ˆ (1)
2

𝐾 = 𝑄 𝑥ˆ (2) 𝐴2 T𝑄 2−1

𝐴2𝑥ˆ (1) . . . predicted observation

𝑣 2 . . . predicted residual (attention!)
𝐾 . . . gain matrix Verstärkungs-
matrix
disadvantage: too many matrix inversions
  −1
𝑄 𝑥−1
ˆ (1) , 𝑄 𝑥−1
ˆ (1) + 𝐴 2
T −1
𝑄 2 𝐴 2 , 𝑄 2−1
𝑚 2 ×𝑚 2
𝑛×𝑛 𝑛×𝑛

8.1.4 Formulation using condition equations

𝐵 T𝐴 = 0
  𝐼 
=⇒ −𝐴2 𝐼 =0
𝐴2
( !) ( !)  
  𝑥ˆ (1) 𝑥ˆ (1) 𝑄 𝑥ˆ (1) 0
−𝐴2 𝐼 E = 0; D =
𝑦 𝑦 0 𝑄2
2 2
n o
   −1 
𝐵T E 𝑦 = 0  
 
n o T
𝐵T 𝑦
D 𝑦 = 𝑄𝑦 
=⇒ 𝑦ˆ = 𝐼 − 𝑄 𝑦 𝐵 𝐵 𝑄 𝑦 𝐵


!     !
𝑥ˆ (2) 𝐼 0 −𝑄 𝑥ˆ (1) 𝐴2 T   −1    𝑥ˆ
T (1)
=⇒ = − 𝑄 2 + 𝐴2𝑄 𝑥ˆ (1) 𝐴2 −𝐴2 𝐼
𝑦ˆ 0𝐼 𝑄2 𝑦
2 2

129
8 Recursive estimation

=⇒ Measurement update
  −1  
𝑥ˆ (2) = 𝑥ˆ (1) + 𝑄 𝑥ˆ (1) 𝐴2 T 𝑄 2 + 𝐴2𝑄 𝑥ˆ (1) 𝐴2 T 𝑦 − 𝐴2𝑥ˆ (1)
2
= 𝑥ˆ (1) + 𝐾𝑣 2
  −1
𝐾 = 𝑄 𝑥ˆ (1) 𝐴2 T 𝑄 2 + 𝐴2𝑄 𝑥ˆ (1) 𝐴2 T
𝑚 2 ×𝑚 2

= 𝑄 𝑥ˆ (1) 𝐴2 T𝑄 𝑣−12

=⇒ Covariance update
  −1
𝑄 𝑥ˆ (2) = 𝑄 𝑥ˆ (1) − 𝑄 𝑥ˆ (1) 𝐴2 T 𝑄 2 + 𝐴2𝑄 𝑥ˆ (1) 𝐴2 T 𝐴2𝑄 𝑥ˆ (1)
= 𝑄 𝑥ˆ (1) − 𝐾𝐴2𝑄 𝑥ˆ (1)
= (𝐼 − 𝐾𝐴2 ) 𝑄 𝑥ˆ (1)

Remark: Variance decreases as more observations are included.

8.2 More general


 𝑦   
 𝑦  

 © 1 ª  
 © 1 ª 
  © ª   © ª
𝐴1 𝑄1
 ­
­ ®
 ® 
 ­ 𝐴2 ®  ­
­ ®
 ® 
 ­
0
®
E ­ .2 ® = ­­ . ®® 𝑥; D ­ .2 ® = ­­ ®
𝑦 𝑦 𝑄2
 ­ .. ®
 ­ .. ®   ­
 ­ .. ®  ®
  ®
..

 ­ ®  
 ­ ® 
.
 𝑦 
  𝑦 

0
  «
« 𝑘 ¬ « 𝑘 ¬
𝑄𝑘 ¬
« 𝑘 ¬
𝐴

Batch: ! −1 !
Õ
𝑘 Õ
𝑘
𝑥ˆ = 𝐴𝑖 T𝑄𝑖−1𝐴𝑖 𝐴𝑖 T𝑄𝑖−1𝑦
𝑖
𝑖=1 𝑖=1

Recursive:

𝑥ˆ (𝑘) = 𝑥ˆ (𝑘−1) + 𝐾𝑘 𝑣 𝑘
𝑣 𝑘 = 𝑦 − 𝐴𝑘 𝑥ˆ (𝑘−1)
  −1
𝑘

𝐾𝑘 = 𝑄 𝑥ˆ (𝑘−1) 𝐴𝑘 T 𝑄𝑘 + 𝐴𝑘 𝑄 𝑥ˆ (𝑘−1) 𝐴𝑘 T

= 𝑄 𝑥ˆ (𝑘−1) 𝐴𝑘 T𝑄 𝑣−1
𝑘

𝑄 𝑥ˆ (𝑘 ) = (𝐼 − 𝐾𝑘 𝐴𝑘 ) 𝑄 𝑥ˆ (𝑘−1)

130
A Partitioning

A.1 Inverse Partitioning Method (IPM)

𝑊 𝑋   𝐴 𝐵   
 𝑛×𝑛 𝑛×𝑘   𝑛×𝑛 𝑛×𝑘 
   = 𝐼𝑛 0
 𝑌 𝑍  𝐶 𝐷  A, B, C, D are unknown
   0 𝐼𝑘
 𝑘×𝑛 𝑘×𝑘   𝑘×𝑛 𝑘×𝑘 
| {z }
Inverse

(1) 𝑊 𝐴 + 𝑋𝐶 = 𝐼𝑛 , rank𝑊 = 𝑛
(2) 𝑊 𝐵 + 𝑋 𝐷 = 0
(3) 𝑌 𝐴 + 𝑍𝐶 = 0
(4) 𝑌 𝐵 + 𝑍 𝐷 = 𝐼𝑘
(5) 𝑊 −1 · (1) : 𝐴 + 𝑊 −1𝑋𝐶 = 𝑊 −1 =⇒ 𝐴 = 𝑊 −1 − 𝑊 −1𝑋𝐶
(6) Insert (5) into (3) : 𝑌𝑊 −1 − 𝑌𝑊 −1𝑋𝐶 + 𝑍𝐶 = 0 =⇒ 𝐶 = −(𝑍 − 𝑌𝑊 −1𝑋 ) −1𝑌𝑊 −1 (provided
𝐺 = 𝑍 − 𝑌𝑊 −1𝑋 is non-singular)
(7) 𝐷 = 𝐺 −1 = (𝑍 − 𝑌𝑊 −1𝑋 ) −1
(8) 𝐵 = −𝑊 −1𝑋𝐺 −1 = −𝑊 −1𝑋 (𝑍 − 𝑌𝑊 −1𝑋 ) −1

A.2 Inverse Partitioning Method: special case 1

    
𝐼 −𝑏 𝐴 𝐵 𝐼 0
= A, B, C, D are unknown
−𝑏 T 0 𝐶 𝐷 0𝐼
| {z }
Inverse

(1) 𝐴 − 𝑏𝐶 = 𝐼
(2) 𝐵 − 𝑏𝐷 = 0
(3) −𝑏 T𝐴 = 0
(4) −𝑏 T 𝐵 = 𝐼
(5) −𝑏 T𝐴 −𝑏 T𝑏𝐶 = −𝑏 T =⇒ 𝐶 = −(𝑏 T𝑏) −1𝑏 T
|{z}
=0

(6) −𝑏 T 𝐵 + 𝑏 T𝑏𝐷 = 0 =⇒ 𝐼 + 𝑏 T𝑏𝐷 = 0 =⇒ 𝐷 = −(𝑏 T𝑏) −1

131
A Partitioning

(7) 𝐴 + 𝑏 (𝑏 T𝑏) −1𝑏 T = 𝐼 =⇒ 𝐴 = 𝐼 − 𝑏 (𝑏 T𝑏) −1𝑏 T

(8) 𝐵 + 𝑏 (𝑏 T𝑏) −1 = 0 =⇒ 𝐵 = −𝑏 (𝑏 T𝑏) −1

  −1  
𝐼 −𝑏 𝐼 − 𝑏 (𝑏 T𝑏) −1𝑏 T −𝑏 (𝑏 T𝑏) −1
=
−𝑏 T 0 −(𝑏 T𝑏) −1𝑏 T −(𝑏 T𝑏) −1

𝑒ˆ = 𝑏 (𝑏 T𝑏) −1𝑏 T𝑦
𝜆ˆ = (𝑏 T𝑏) −1𝑏 T𝑦

A.3 Inverse Partitioning Method: special case 2

(𝐴 rank deficient, constraint 𝐷 T𝑥 = 𝑐).

The normal matrix of the linear system is symmetric, therefore
 T    
𝐴 𝐴𝐷 𝑅 𝑆T 𝐼 0
=
𝐷T 0 𝑆 𝑄 0𝐼
| {z }
Inverse

then

(𝐴T𝐴)𝑅 + 𝐷𝑆 = 𝐼 (A.1)
(𝐴T𝐴)𝑆 T + 𝐷𝑄 = 0 (A.2)
𝐷 T𝑅 = 0 (A.3)
𝐷 T𝑆 T = 𝐼 (A.4)

and with 𝐻 = null(𝐴), 𝐴𝐻 T = 0

𝐻 · (A.1) =⇒ 𝐻 (𝐴T 𝐴)𝑅 + 𝐻 𝐷𝑆 = 𝐻 =⇒ 𝑆 = (𝐻 𝐷) −1𝐻

|{z}
0

𝐻 · (A.2) =⇒ 𝐻 (𝐴T 𝐴)𝑆 T + 𝐻 𝐷𝑄 = 0 =⇒ 𝐻 𝐷𝑄 = 0

|{z}
0

since 𝐻 𝐷 is a 𝑑 × 𝑑 full-rank matrix

𝐻 𝐷𝑄 = 0 =⇒ 𝑄 = 0

(A.1) + 𝐷 · (A.3) =⇒ (𝐴T𝐴)𝑅 + 𝐷 (𝐻 𝐷) −1𝐻 + 𝐷𝐷 T𝑅 = 𝐼 (A.5)

=⇒ (𝐴T𝐴 + 𝐷𝐷 T )𝑅 = 𝐼 − 𝐷 (𝐻 𝐷) −1𝐻
=⇒ 𝑅 = (𝐴T𝐴 + 𝐷𝐷 T ) −1 (𝐼 − 𝐷 (𝐻 𝐷) −1𝐻 )

132
 A.3 Inverse Partitioning Method: special case 2

Inserting 𝑅 and 𝑆 into the normal equations





𝑥ˆ = 𝑅𝐴T𝑦 = (𝐴T𝐴 + 𝐷𝐷 T ) −1𝐴T𝑦 − (𝐴T𝐴 + 𝐷𝐷 T ) −1 𝐷 (𝐻 𝐷) −1 𝐻𝐴T 𝑦
|{z}



0

 𝜆ˆ = 𝑆𝐴T𝑦 = (𝐻 𝐷) −1 𝐻𝐴T 𝑦 = 0

 |{z}

 0

=⇒ 𝜆ˆ = 0
=⇒ 𝑥ˆ = (𝐴T𝐴 + 𝐷𝐷 T ) −1𝐴T𝑦 + 𝐻 (𝐻 T 𝐷) −1𝑐
𝑒ˆ = 𝑦 − 𝐴𝑥ˆ = (𝐼 − 𝐴(𝐴T𝐴 + 𝐷𝐷 T ) −1𝐴T )𝑦

133
B Statistical Tables

B.1 Standard Normal Distribution z

∫𝑘𝛼
Computation of one-sided level of significance 𝛼 = 1 − 𝑓 (𝑥) d𝑥 and
−∞
∫ 𝛼/2
𝑘 1−
two-sided level of significance 𝛼 = 2 𝑓 (𝑥) d𝑥.
−∞

f(x) f(x)

k1- -a = -k -a
2 2

a a/2 a/2
- ∞ 0 ka x - ∞ k1- -a 0 k -a x
2 2

𝑘𝛼 0 1 2 3 4 5 6 7 8 9
0.0 0.5000 0.4960 0.4920 0.4880 0.4840 0.4801 0.4761 0.4721 0.4681 0.4641
0.1 0.4602 0.4562 0.4522 0.4483 0.4443 0.4404 0.4364 0.4325 0.4286 0.4247
0.2 0.4207 0.4168 0.4129 0.4090 0.4052 0.4013 0.3974 0.3936 0.3897 0.3859
0.3 0.3821 0.3783 0.3745 0.3707 0.3669 0.3632 0.3594 0.3557 0.3520 0.3483
0.4 0.3446 0.3409 0.3372 0.3336 0.3300 0.3264 0.3228 0.3192 0.3156 0.3121
0.5 0.3085 0.3050 0.3015 0.2981 0.2946 0.2912 0.2877 0.2843 0.2810 0.2776
0.6 0.2743 0.2709 0.2676 0.2643 0.2611 0.2578 0.2546 0.2514 0.2483 0.2451
0.7 0.2420 0.2389 0.2358 0.2327 0.2296 0.2266 0.2236 0.2206 0.2177 0.2148
0.8 0.2119 0.2090 0.2061 0.2033 0.2005 0.1977 0.1949 0.1922 0.1894 0.1867
0.9 0.1841 0.1814 0.1788 0.1762 0.1736 0.1711 0.1685 0.1660 0.1635 0.1611
1.0 0.1587 0.1562 0.1539 0.1515 0.1492 0.1469 0.1446 0.1423 0.1401 0.1379
1.1 0.1357 0.1335 0.1314 0.1292 0.1271 0.1251 0.1230 0.1210 0.1190 0.1170
1.2 0.1151 0.1131 0.1112 0.1093 0.1075 0.1056 0.1038 0.1020 0.1003 0.0985
1.3 0.0968 0.0951 0.0934 0.0918 0.0901 0.0885 0.0869 0.0853 0.0838 0.0823
1.4 0.0808 0.0793 0.0778 0.0764 0.0749 0.0735 0.0721 0.0708 0.0694 0.0681

134
 B.1 Standard Normal Distribution z

∫𝑘𝛼
Computation of one-sided level of significance 𝛼 = 1 − 𝑓 (𝑥) d𝑥 and
−∞
∫ 𝛼/2
𝑘 1−
two-sided level of significance 𝛼 = 2 𝑓 (𝑥) d𝑥 (continued).
−∞

𝑘𝛼 0 1 2 3 4 5 6 7 8 9
1.5 0.0668 0.0655 0.0643 0.0630 0.0618 0.0606 0.0594 0.0582 0.0571 0.0559
1.6 0.0548 0.0537 0.0526 0.0516 0.0505 0.0495 0.0485 0.0475 0.0465 0.0455
1.7 0.0446 0.0436 0.0427 0.0418 0.0409 0.0401 0.0392 0.0384 0.0375 0.0367
1.8 0.0359 0.0351 0.0344 0.0336 0.0329 0.0322 0.0314 0.0307 0.0301 0.0294
1.9 0.0287 0.0281 0.0274 0.0268 0.0262 0.0256 0.0250 0.0244 0.0239 0.0233
2.0 0.0228 0.0222 0.0217 0.0212 0.0207 0.0202 0.0197 0.0192 0.0188 0.0183
2.1 0.0179 0.0174 0.0170 0.0166 0.0162 0.0158 0.0154 0.0150 0.0146 0.0143
2.2 0.0139 0.0136 0.0132 0.0129 0.0125 0.0122 0.0119 0.0116 0.0113 0.0110
2.3 0.0107 0.0104 0.0102 0.0099 0.0096 0.0094 0.0091 0.0089 0.0087 0.0084
2.4 0.0082 0.0080 0.0078 0.0075 0.0073 0.0071 0.0069 0.0068 0.0066 0.0064
2.5 0.0062 0.0060 0.0059 0.0057 0.0055 0.0054 0.0052 0.0051 0.0049 0.0048
2.6 0.0047 0.0045 0.0044 0.0043 0.0041 0.0040 0.0039 0.0038 0.0037 0.0036
2.7 0.0035 0.0034 0.0033 0.0032 0.0031 0.0030 0.0029 0.0028 0.0027 0.0026
2.8 0.0026 0.0025 0.0024 0.0023 0.0023 0.0022 0.0021 0.0021 0.0020 0.0019
2.9 0.0019 0.0018 0.0018 0.0017 0.0016 0.0016 0.0015 0.0015 0.0014 0.0014
3.0 0.0013 0.0013 0.0013 0.0012 0.0012 0.0011 0.0011 0.0011 0.0010 0.0010
3.1 0.0010 0.0009 0.0009 0.0009 0.0008 0.0008 0.0008 0.0008 0.0007 0.0007
3.2 0.0007 0.0007 0.0006 0.0006 0.0006 0.0006 0.0006 0.0005 0.0005 0.0005
3.3 0.0005 0.0005 0.0005 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0003
3.4 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0002

Calculation in Matlab:

𝛼 = 1 − ♥♦r♠❝❞❢ (𝑘𝛼 ) 𝑘𝛼 = ♥♦r♠✐♥✈ (1 − 𝛼)

Example (𝑘𝛼 = 0.87, one-sided):

𝛼 = 0.1922 = 1 − ♥♦r♠❝❞❢ (0.87) 𝑘𝛼 = 0.87 = ♥♦r♠✐♥✈ (1 − 0.1922)

135
B Statistical Tables

B.2 Central 𝜒 2 -Distribution

2
Computation of critical value 𝑘𝛼 = 𝜒1−𝛼;𝑟,𝜆=0 (𝑟 degrees of freedom).

f(x) f(x)

2 2
c c
r,l=0 r,l=0

a a/2 a/2
0 ka x 0 k1- -a2 k -a
2
x

𝑟 \𝛼 0.995 0.990 0.975 0.950 0.900 0.500

1 0.000 0.000 0.001 0.004 0.016 0.455
2 0.010 0.020 0.051 0.103 0.211 1.386
3 0.072 0.115 0.216 0.352 0.584 2.366
4 0.207 0.297 0.484 0.711 1.064 3.357
5 0.412 0.554 0.831 1.145 1.610 4.351
6 0.676 0.872 1.237 1.635 2.204 5.348
7 0.989 1.239 1.690 2.167 2.833 6.346
8 1.344 1.646 2.180 2.733 3.490 7.344
9 1.735 2.088 2.700 3.325 4.168 8.343
10 2.156 2.558 3.247 3.940 4.865 9.342
11 2.603 3.053 3.816 4.575 5.578 10.34
12 3.074 3.571 4.404 5.226 6.304 11.34
13 3.565 4.107 5.009 5.892 7.042 12.34
14 4.075 4.660 5.629 6.571 7.790 13.34
15 4.601 5.229 6.262 7.261 8.547 14.34
16 5.142 5.812 6.908 7.962 9.312 15.34
17 5.697 6.408 7.564 8.672 10.09 16.34
18 6.265 7.015 8.231 9.390 10.86 17.34
19 6.844 7.633 8.907 10.12 11.65 18.34
20 7.434 8.260 9.591 10.85 12.44 19.34
21 8.034 8.897 10.28 11.59 13.24 20.34
22 8.643 9.542 10.98 12.34 14.04 21.34
23 9.260 10.20 11.69 13.09 14.85 22.34
24 9.886 10.86 12.40 13.85 15.66 23.34
25 10.52 11.52 13.12 14.61 16.47 24.34
26 11.16 12.20 13.84 15.38 17.29 25.34
27 11.81 12.88 14.57 16.15 18.11 26.34
28 12.46 13.56 15.31 16.93 18.94 27.34
29 13.12 14.26 16.05 17.71 19.77 28.34
30 13.79 14.95 16.79 18.49 20.60 29.34

136
 B.2 Central 𝜒 2 -Distribution

2
Computation of critical value 𝑘𝛼 = 𝜒1−𝛼;𝑟,𝜆=0 (continued).

𝑟 \𝛼 0.995 0.990 0.975 0.950 0.900 0.500

35 17.19 18.51 20.57 22.47 24.80 34.34
40 20.71 22.16 24.43 26.51 29.05 39.34
45 24.31 25.90 28.37 30.61 33.35 44.34
50 27.99 29.71 32.36 34.76 37.69 49.33
60 35.53 37.48 40.48 43.19 46.46 59.33
70 43.28 45.44 48.76 51.74 55.33 69.33
80 51.17 53.54 57.15 60.39 64.28 79.33
90 59.20 61.75 65.65 69.13 73.29 89.33
100 67.33 70.06 74.22 77.93 82.36 99.33

𝑟 \𝛼 0.100 0.050 0.025 0.010 0.005 0.001

1 2.706 3.841 5.024 6.635 7.879 10.83
2 4.605 5.991 7.378 9.210 10.60 13.82
3 6.251 7.815 9.348 11.34 12.84 16.27
4 7.779 9.488 11.14 13.28 14.86 18.47
5 9.236 11.07 12.83 15.09 16.75 20.52
6 10.64 12.59 14.45 16.81 18.55 22.46
7 12.02 14.07 16.01 18.48 20.28 24.32
8 13.36 15.51 17.53 20.09 21.95 26.12
9 14.68 16.92 19.02 21.67 23.59 27.88
10 15.99 18.31 20.48 23.21 25.19 29.59
11 17.28 19.68 21.92 24.72 26.76 31.26
12 18.55 21.03 23.34 26.22 28.30 32.91
13 19.81 22.36 24.74 27.69 29.82 34.53
14 21.06 23.68 26.12 29.14 31.32 36.12
15 22.31 25.00 27.49 30.58 32.80 37.70
16 23.54 26.30 28.85 32.00 34.27 39.25
17 24.77 27.59 30.19 33.41 35.72 40.79
18 25.99 28.87 31.53 34.81 37.16 42.31
19 27.20 30.14 32.85 36.19 38.58 43.82
20 28.41 31.41 34.17 37.57 40.00 45.31
21 29.62 32.67 35.48 38.93 41.40 46.80
22 30.81 33.92 36.78 40.29 42.80 48.27
23 32.01 35.17 38.08 41.64 44.18 49.73
24 33.20 36.42 39.36 42.98 45.56 51.18
25 34.38 37.65 40.65 44.31 46.93 52.62
26 35.56 38.89 41.92 45.64 48.29 54.05
27 36.74 40.11 43.19 46.96 49.64 55.48
28 37.92 41.34 44.46 48.28 50.99 56.89
29 39.09 42.56 45.72 49.59 52.34 58.30
30 40.26 43.77 46.98 50.89 53.67 59.70

137
B Statistical Tables

2
Computation of critical value 𝑘𝛼 = 𝜒1−𝛼;𝑟,𝜆=0 (continued).

𝑟 \𝛼 0.100 0.050 0.025 0.010 0.005 0.001

35 46.06 49.80 53.20 57.34 60.27 66.62
40 51.81 55.76 59.34 63.69 66.77 73.40
45 57.51 61.66 65.41 69.96 73.17 80.08
50 63.17 67.50 71.42 76.15 79.49 86.66
60 74.40 79.08 83.30 88.38 91.95 99.61
70 85.53 90.53 95.02 100.4 104.2 112.3
80 96.58 101.9 106.6 112.3 116.3 124.8
90 107.6 113.1 118.1 124.1 128.3 137.2
100 118.5 124.3 129.6 135.8 140.2 149.4

Calculation in Matlab:

𝑘𝛼 = ❝❤✐✷✐♥✈ (1 − 𝛼, 𝑟 ) 𝛼 = 1 − ❝❤✐✷❝❞❢ (𝑘𝛼 , 𝑟 )

Example (𝛼 = 0.95, one-sided):

𝑘𝛼 = 11.59 = ❝❤✐✷✐♥✈ (1 − 0.95, 21) 𝛼 = 0.95 = 1 − ❝❤✐✷❝❞❢ (11.59, 21)

Example (𝛼 = 0.01, two-sided):

𝑘 1−𝛼/2 = 41.4 = ❝❤✐✷✐♥✈ (1 − 0.01/2, 21) 𝑘𝛼/2 = 8.034 = ❝❤✐✷✐♥✈ ( 0.01/2, 21)

138
 B.3 Non-central 𝜒 2 -Distribution

B.3 Non-central 𝜒 2 -Distribution

Computation of power of test 𝛾 = 1 − 𝛽

(Non-centrality parameter 𝜆, 𝑟 = 1 degrees of freedom).

𝜆\𝛼 0.100 0.010 0.001 𝜆\𝛼 0.100 0.010 0.001

1.000 0.264 0.058 0.011 11.000 0.953 0.771 0.510
1.250 0.302 0.073 0.015 11.250 0.956 0.782 0.525
1.500 0.339 0.088 0.019 11.500 0.960 0.793 0.540
1.750 0.375 0.105 0.025 11.750 0.963 0.803 0.555
2.000 0.410 0.123 0.030 12.000 0.966 0.813 0.569
2.250 0.443 0.141 0.037 12.250 0.968 0.822 0.583
2.500 0.475 0.160 0.044 12.500 0.971 0.831 0.597
2.750 0.506 0.179 0.051 12.750 0.973 0.840 0.610
3.000 0.535 0.199 0.060 13.000 0.975 0.848 0.624
3.250 0.563 0.220 0.068 13.250 0.977 0.856 0.637
3.500 0.590 0.240 0.078 13.500 0.979 0.864 0.649
3.750 0.615 0.261 0.088 13.750 0.980 0.871 0.662
4.000 0.639 0.282 0.098 14.000 0.982 0.878 0.674
4.250 0.662 0.304 0.110 14.250 0.983 0.885 0.686
4.500 0.683 0.325 0.121 14.500 0.985 0.891 0.698
4.750 0.704 0.346 0.133 14.750 0.986 0.897 0.709
5.000 0.723 0.367 0.146 15.000 0.987 0.903 0.720
5.250 0.741 0.388 0.159 15.250 0.988 0.908 0.731
5.500 0.758 0.409 0.172 15.500 0.989 0.913 0.741
5.750 0.774 0.429 0.186 15.750 0.990 0.918 0.751
6.000 0.790 0.450 0.200 16.000 0.991 0.923 0.761
6.250 0.804 0.470 0.215 16.250 0.991 0.927 0.771
6.500 0.817 0.490 0.229 16.500 0.992 0.931 0.780
6.750 0.830 0.509 0.244 16.750 0.993 0.935 0.789
7.000 0.842 0.528 0.260 17.000 0.993 0.939 0.797
7.250 0.853 0.546 0.275 17.250 0.994 0.943 0.806
7.500 0.863 0.565 0.291 17.500 0.994 0.946 0.814
7.750 0.873 0.582 0.306 17.750 0.995 0.949 0.822
8.000 0.882 0.600 0.322 18.000 0.995 0.952 0.829
8.250 0.890 0.617 0.338 18.250 0.996 0.955 0.837
8.500 0.898 0.633 0.354 18.500 0.996 0.958 0.844
8.750 0.905 0.649 0.370 18.750 0.996 0.960 0.851
9.000 0.912 0.664 0.386 19.000 0.997 0.963 0.857
9.250 0.919 0.679 0.402 19.250 0.997 0.965 0.864
9.500 0.925 0.694 0.417 19.500 0.997 0.967 0.870
9.750 0.930 0.708 0.433 19.750 0.997 0.969 0.876
10.000 0.935 0.721 0.449 20.000 0.998 0.971 0.881
10.250 0.940 0.734 0.465 20.250 0.998 0.973 0.887
10.500 0.945 0.747 0.480 20.500 0.998 0.975 0.892
10.750 0.949 0.759 0.495 20.750 0.998 0.976 0.897

Calculation in Matlab:

𝑘𝛼 = ❝❤✐✷✐♥✈ (1 − 𝛼, 𝑟 ) 𝛾 = 1 − ♥❝①✷❝❞❢ (𝑘𝛼 , 𝑟, 𝜆)

139
B Statistical Tables

B.4 Central t-Distribution

Computation of critical value 𝑘𝛼 = 𝑡 1−𝛼;𝑟 (𝑟 degrees of freedom).

f(x) f(x)

k1- -a = -k -a
2 2

a a/2 a/2
- ∞ 0 ka x - ∞ k1- -a 0 k -a x
2 2

𝑟 \𝛼 0.100 0.050 0.025 0.010 0.005 0.001

1 3.078 6.314 12.71 31.82 63.66 318.3
2 1.886 2.920 4.303 6.965 9.925 22.33
3 1.638 2.353 3.182 4.541 5.841 10.21
4 1.533 2.132 2.776 3.747 4.604 7.173
5 1.476 2.015 2.571 3.365 4.032 5.893
6 1.440 1.943 2.447 3.143 3.707 5.208
7 1.415 1.895 2.365 2.998 3.499 4.785
8 1.397 1.860 2.306 2.896 3.355 4.501
9 1.383 1.833 2.262 2.821 3.250 4.297
10 1.372 1.812 2.228 2.764 3.169 4.144
11 1.363 1.796 2.201 2.718 3.106 4.025
12 1.356 1.782 2.179 2.681 3.055 3.930
13 1.350 1.771 2.160 2.650 3.012 3.852
14 1.345 1.761 2.145 2.624 2.977 3.787
15 1.341 1.753 2.131 2.602 2.947 3.733
16 1.337 1.746 2.120 2.583 2.921 3.686
17 1.333 1.740 2.110 2.567 2.898 3.646
18 1.330 1.734 2.101 2.552 2.878 3.610
19 1.328 1.729 2.093 2.539 2.861 3.579
20 1.325 1.725 2.086 2.528 2.845 3.552

140
 B.4 Central t-Distribution

Computation of critical value 𝑘𝛼 = 𝑡 1−𝛼;𝑟 (continued).

𝑟 \𝛼 0.100 0.050 0.025 0.010 0.005 0.001

21 1.323 1.721 2.080 2.518 2.831 3.527
22 1.321 1.717 2.074 2.508 2.819 3.505
23 1.319 1.714 2.069 2.500 2.807 3.485
24 1.318 1.711 2.064 2.492 2.797 3.467
25 1.316 1.708 2.060 2.485 2.787 3.450
26 1.315 1.706 2.056 2.479 2.779 3.435
27 1.314 1.703 2.052 2.473 2.771 3.421
28 1.313 1.701 2.048 2.467 2.763 3.408
29 1.311 1.699 2.045 2.462 2.756 3.396
30 1.310 1.697 2.042 2.457 2.750 3.385
35 1.306 1.690 2.030 2.438 2.724 3.340
40 1.303 1.684 2.021 2.423 2.704 3.307
45 1.301 1.679 2.014 2.412 2.690 3.281
50 1.299 1.676 2.009 2.403 2.678 3.261
60 1.296 1.671 2.000 2.390 2.660 3.232
70 1.294 1.667 1.994 2.381 2.648 3.211
80 1.292 1.664 1.990 2.374 2.639 3.195
90 1.291 1.662 1.987 2.368 2.632 3.183
100 1.290 1.660 1.984 2.364 2.626 3.174
200 1.286 1.653 1.972 2.345 2.601 3.131
500 1.283 1.648 1.965 2.334 2.586 3.107
∞ 1.282 1.645 1.960 2.327 2.576 3.091

Calculation in Matlab:

𝑘𝛼 = t✐♥✈ (1 − 𝛼, 𝑟 ) 𝛼 = 1 − t❝❞❢ (𝑘𝛼 , 𝑟 )

Example (𝛼 = 0.005, one-sided):

𝑘𝛼 = 2.898 = t✐♥✈ (1 − 0.005, 17) 𝛼 = 0.005 = 1 − t❝❞❢ (2.898, 17)

141
B Statistical Tables

B.5 Central F-Distribution

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 (𝑟 1 , 𝑟 2 degrees of freedom).

f(x) f(x)

1
F r1,r2 k1-a/2 =Fa/2;r2,r1 =
F1-a/2;r ,r
1 2

a a/2 a/2
0 ka x 0 k1-a/2 ka/2 x

𝛼 = 0.10, 1 − 𝛼 = 0.90
𝑟 2 \𝑟 1 1 2 3 4 5 6 7 8 9 10 12
1 39.86 49.50 53.59 55.83 57.24 58.20 58.91 59.44 59.86 60.19 60.71
2 8.526 9.000 9.162 9.243 9.293 9.326 9.349 9.367 9.381 9.392 9.408
3 5.538 5.462 5.391 5.343 5.309 5.285 5.266 5.252 5.240 5.230 5.216
4 4.545 4.325 4.191 4.107 4.051 4.010 3.979 3.955 3.936 3.920 3.896
5 4.060 3.780 3.619 3.520 3.453 3.405 3.368 3.339 3.316 3.297 3.268
6 3.776 3.463 3.289 3.181 3.108 3.055 3.014 2.983 2.958 2.937 2.905
7 3.589 3.257 3.074 2.961 2.883 2.827 2.785 2.752 2.725 2.703 2.668
8 3.458 3.113 2.924 2.806 2.726 2.668 2.624 2.589 2.561 2.538 2.502
9 3.360 3.006 2.813 2.693 2.611 2.551 2.505 2.469 2.440 2.416 2.379
10 3.285 2.924 2.728 2.605 2.522 2.461 2.414 2.377 2.347 2.323 2.284
11 3.225 2.860 2.660 2.536 2.451 2.389 2.342 2.304 2.274 2.248 2.209
12 3.177 2.807 2.606 2.480 2.394 2.331 2.283 2.245 2.214 2.188 2.147
13 3.136 2.763 2.560 2.434 2.347 2.283 2.234 2.195 2.164 2.138 2.097
14 3.102 2.726 2.522 2.395 2.307 2.243 2.193 2.154 2.122 2.095 2.054
15 3.073 2.695 2.490 2.361 2.273 2.208 2.158 2.119 2.086 2.059 2.017
16 3.048 2.668 2.462 2.333 2.244 2.178 2.128 2.088 2.055 2.028 1.985
17 3.026 2.645 2.437 2.308 2.218 2.152 2.102 2.061 2.028 2.001 1.958
18 3.007 2.624 2.416 2.286 2.196 2.130 2.079 2.038 2.005 1.977 1.933
19 2.990 2.606 2.397 2.266 2.176 2.109 2.058 2.017 1.984 1.956 1.912
20 2.975 2.589 2.380 2.249 2.158 2.091 2.040 1.999 1.965 1.937 1.892
22 2.949 2.561 2.351 2.219 2.128 2.060 2.008 1.967 1.933 1.904 1.859
24 2.927 2.538 2.327 2.195 2.103 2.035 1.983 1.941 1.906 1.877 1.832
26 2.909 2.519 2.307 2.174 2.082 2.014 1.961 1.919 1.884 1.855 1.809
28 2.894 2.503 2.291 2.157 2.064 1.996 1.943 1.900 1.865 1.836 1.790
30 2.881 2.489 2.276 2.142 2.049 1.980 1.927 1.884 1.849 1.819 1.773
40 2.835 2.440 2.226 2.091 1.997 1.927 1.873 1.829 1.793 1.763 1.715
50 2.809 2.412 2.197 2.061 1.966 1.895 1.840 1.796 1.760 1.729 1.680
60 2.791 2.393 2.177 2.041 1.946 1.875 1.819 1.775 1.738 1.707 1.657
80 2.769 2.370 2.154 2.016 1.921 1.849 1.793 1.748 1.711 1.680 1.629
100 2.756 2.356 2.139 2.002 1.906 1.834 1.778 1.732 1.695 1.663 1.612
200 2.731 2.329 2.111 1.973 1.876 1.804 1.747 1.701 1.663 1.631 1.579
500 2.716 2.313 2.095 1.956 1.859 1.786 1.729 1.683 1.644 1.612 1.559
∞ 2.706 2.303 2.084 1.945 1.847 1.774 1.717 1.670 1.632 1.599 1.546

142
 B.5 Central F-Distribution

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 (continued).

𝛼 = 0.10, 1 − 𝛼 = 0.90
𝑟 2 \𝑟 1 14 16 18 20 30 40 50 100 200 500 ∞
1 61.07 61.35 61.57 61.74 62.26 62.53 62.69 63.01 63.17 63.26 63.33
2 9.420 9.429 9.436 9.441 9.458 9.466 9.471 9.481 9.486 9.489 9.491
3 5.205 5.196 5.190 5.184 5.168 5.160 5.155 5.144 5.139 5.136 5.134
4 3.878 3.864 3.853 3.844 3.817 3.804 3.795 3.778 3.769 3.764 3.761
5 3.247 3.230 3.217 3.207 3.174 3.157 3.147 3.126 3.116 3.109 3.105
6 2.881 2.863 2.848 2.836 2.800 2.781 2.770 2.746 2.734 2.727 2.722
7 2.643 2.623 2.607 2.595 2.555 2.535 2.523 2.497 2.484 2.476 2.471
8 2.475 2.455 2.438 2.425 2.383 2.361 2.348 2.321 2.307 2.298 2.293
9 2.351 2.329 2.312 2.298 2.255 2.232 2.218 2.189 2.174 2.165 2.159
10 2.255 2.233 2.215 2.201 2.155 2.132 2.117 2.087 2.071 2.062 2.055
11 2.179 2.156 2.138 2.123 2.076 2.052 2.036 2.005 1.989 1.979 1.972
12 2.117 2.094 2.075 2.060 2.011 1.986 1.970 1.938 1.921 1.911 1.904
13 2.066 2.042 2.023 2.007 1.958 1.931 1.915 1.882 1.864 1.853 1.846
14 2.022 1.998 1.978 1.962 1.912 1.885 1.869 1.834 1.816 1.805 1.797
15 1.985 1.961 1.941 1.924 1.873 1.845 1.828 1.793 1.774 1.763 1.755
16 1.953 1.928 1.908 1.891 1.839 1.811 1.793 1.757 1.738 1.726 1.718
17 1.925 1.900 1.879 1.862 1.809 1.781 1.763 1.726 1.706 1.694 1.686
18 1.900 1.875 1.854 1.837 1.783 1.754 1.736 1.698 1.678 1.665 1.657
19 1.878 1.852 1.831 1.814 1.759 1.730 1.711 1.673 1.652 1.639 1.631
20 1.859 1.833 1.811 1.794 1.738 1.708 1.690 1.650 1.629 1.616 1.607
22 1.825 1.798 1.777 1.759 1.702 1.671 1.652 1.611 1.590 1.576 1.567
24 1.797 1.770 1.748 1.730 1.672 1.641 1.621 1.579 1.556 1.542 1.533
26 1.774 1.747 1.724 1.706 1.647 1.615 1.594 1.551 1.528 1.514 1.504
28 1.754 1.726 1.704 1.685 1.625 1.592 1.572 1.528 1.504 1.489 1.478
30 1.737 1.709 1.686 1.667 1.606 1.573 1.552 1.507 1.482 1.467 1.456
40 1.678 1.649 1.625 1.605 1.541 1.506 1.483 1.434 1.406 1.389 1.377
50 1.643 1.613 1.588 1.568 1.502 1.465 1.441 1.388 1.359 1.340 1.327
60 1.619 1.589 1.564 1.543 1.476 1.437 1.413 1.358 1.326 1.306 1.292
80 1.590 1.559 1.534 1.513 1.443 1.403 1.377 1.318 1.284 1.261 1.245
100 1.573 1.542 1.516 1.494 1.423 1.382 1.355 1.293 1.257 1.232 1.214
200 1.539 1.507 1.480 1.458 1.383 1.339 1.310 1.242 1.199 1.168 1.144
500 1.518 1.485 1.458 1.435 1.358 1.313 1.282 1.209 1.160 1.122 1.087
∞ 1.505 1.471 1.444 1.421 1.342 1.295 1.263 1.185 1.130 1.082 1.008

143
B Statistical Tables

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 .

𝛼 = 0.05, 1 − 𝛼 = 0.95
𝑟 2 \𝑟 1 1 2 3 4 5 6 7 8 9 10 12
1 161.4 199.5 215.7 224.6 230.2 234.0 236.8 238.9 240.5 241.9 243.9
2 18.51 19.00 19.16 19.25 19.30 19.33 19.35 19.37 19.38 19.40 19.41
3 10.13 9.552 9.277 9.117 9.013 8.941 8.887 8.845 8.812 8.786 8.745
4 7.709 6.944 6.591 6.388 6.256 6.163 6.094 6.041 5.999 5.964 5.912
5 6.608 5.786 5.409 5.192 5.050 4.950 4.876 4.818 4.772 4.735 4.678
6 5.987 5.143 4.757 4.534 4.387 4.284 4.207 4.147 4.099 4.060 4.000
7 5.591 4.737 4.347 4.120 3.972 3.866 3.787 3.726 3.677 3.637 3.575
8 5.318 4.459 4.066 3.838 3.687 3.581 3.500 3.438 3.388 3.347 3.284
9 5.117 4.256 3.863 3.633 3.482 3.374 3.293 3.230 3.179 3.137 3.073
10 4.965 4.103 3.708 3.478 3.326 3.217 3.135 3.072 3.020 2.978 2.913
11 4.844 3.982 3.587 3.357 3.204 3.095 3.012 2.948 2.896 2.854 2.788
12 4.747 3.885 3.490 3.259 3.106 2.996 2.913 2.849 2.796 2.753 2.687
13 4.667 3.806 3.411 3.179 3.025 2.915 2.832 2.767 2.714 2.671 2.604
14 4.600 3.739 3.344 3.112 2.958 2.848 2.764 2.699 2.646 2.602 2.534
15 4.543 3.682 3.287 3.056 2.901 2.790 2.707 2.641 2.588 2.544 2.475
16 4.494 3.634 3.239 3.007 2.852 2.741 2.657 2.591 2.538 2.494 2.425
17 4.451 3.592 3.197 2.965 2.810 2.699 2.614 2.548 2.494 2.450 2.381
18 4.414 3.555 3.160 2.928 2.773 2.661 2.577 2.510 2.456 2.412 2.342
19 4.381 3.522 3.127 2.895 2.740 2.628 2.544 2.477 2.423 2.378 2.308
20 4.351 3.493 3.098 2.866 2.711 2.599 2.514 2.447 2.393 2.348 2.278
22 4.301 3.443 3.049 2.817 2.661 2.549 2.464 2.397 2.342 2.297 2.226
24 4.260 3.403 3.009 2.776 2.621 2.508 2.423 2.355 2.300 2.255 2.183
26 4.225 3.369 2.975 2.743 2.587 2.474 2.388 2.321 2.265 2.220 2.148
28 4.196 3.340 2.947 2.714 2.558 2.445 2.359 2.291 2.236 2.190 2.118
30 4.171 3.316 2.922 2.690 2.534 2.421 2.334 2.266 2.211 2.165 2.092
40 4.085 3.232 2.839 2.606 2.449 2.336 2.249 2.180 2.124 2.077 2.003
50 4.034 3.183 2.790 2.557 2.400 2.286 2.199 2.130 2.073 2.026 1.952
60 4.001 3.150 2.758 2.525 2.368 2.254 2.167 2.097 2.040 1.993 1.917
80 3.960 3.111 2.719 2.486 2.329 2.214 2.126 2.056 1.999 1.951 1.875
100 3.936 3.087 2.696 2.463 2.305 2.191 2.103 2.032 1.975 1.927 1.850
200 3.888 3.041 2.650 2.417 2.259 2.144 2.056 1.985 1.927 1.878 1.801
500 3.860 3.014 2.623 2.390 2.232 2.117 2.028 1.957 1.899 1.850 1.772
∞ 3.842 2.996 2.605 2.372 2.214 2.099 2.010 1.939 1.880 1.831 1.752

144
 B.5 Central F-Distribution

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 (continued).

𝛼 = 0.05, 1 − 𝛼 = 0.95
𝑟 2 \𝑟 1 14 16 18 20 30 40 50 100 200 500 ∞
1 245.4 246.5 247.3 248.0 250.1 251.1 251.8 253.0 253.7 254.1 254.3
2 19.42 19.43 19.44 19.45 19.46 19.47 19.48 19.49 19.49 19.49 19.50
3 8.715 8.692 8.675 8.660 8.617 8.594 8.581 8.554 8.540 8.532 8.526
4 5.873 5.844 5.821 5.803 5.746 5.717 5.699 5.664 5.646 5.635 5.628
5 4.636 4.604 4.579 4.558 4.496 4.464 4.444 4.405 4.385 4.373 4.365
6 3.956 3.922 3.896 3.874 3.808 3.774 3.754 3.712 3.690 3.678 3.669
7 3.529 3.494 3.467 3.445 3.376 3.340 3.319 3.275 3.252 3.239 3.230
8 3.237 3.202 3.173 3.150 3.079 3.043 3.020 2.975 2.951 2.937 2.928
9 3.025 2.989 2.960 2.936 2.864 2.826 2.803 2.756 2.731 2.717 2.707
10 2.865 2.828 2.798 2.774 2.700 2.661 2.637 2.588 2.563 2.548 2.538
11 2.739 2.701 2.671 2.646 2.570 2.531 2.507 2.457 2.431 2.415 2.405
12 2.637 2.599 2.568 2.544 2.466 2.426 2.401 2.350 2.323 2.307 2.296
13 2.554 2.515 2.484 2.459 2.380 2.339 2.314 2.261 2.234 2.218 2.206
14 2.484 2.445 2.413 2.388 2.308 2.266 2.241 2.187 2.159 2.142 2.131
15 2.424 2.385 2.353 2.328 2.247 2.204 2.178 2.123 2.095 2.078 2.066
16 2.373 2.333 2.302 2.276 2.194 2.151 2.124 2.068 2.039 2.022 2.010
17 2.329 2.289 2.257 2.230 2.148 2.104 2.077 2.020 1.991 1.973 1.960
18 2.290 2.250 2.217 2.191 2.107 2.063 2.035 1.978 1.948 1.929 1.917
19 2.256 2.215 2.182 2.155 2.071 2.026 1.999 1.940 1.910 1.891 1.878
20 2.225 2.184 2.151 2.124 2.039 1.994 1.966 1.907 1.875 1.856 1.843
22 2.173 2.131 2.098 2.071 1.984 1.938 1.909 1.849 1.817 1.797 1.783
24 2.130 2.088 2.054 2.027 1.939 1.892 1.863 1.800 1.768 1.747 1.733
26 2.094 2.052 2.018 1.990 1.901 1.853 1.823 1.760 1.726 1.705 1.691
28 2.064 2.021 1.987 1.959 1.869 1.820 1.790 1.725 1.691 1.669 1.654
30 2.037 1.995 1.960 1.932 1.841 1.792 1.761 1.695 1.660 1.637 1.622
40 1.948 1.904 1.868 1.839 1.744 1.693 1.660 1.589 1.551 1.526 1.509
50 1.895 1.850 1.814 1.784 1.687 1.634 1.599 1.525 1.484 1.457 1.438
60 1.860 1.815 1.778 1.748 1.649 1.594 1.559 1.481 1.438 1.409 1.389
80 1.817 1.772 1.734 1.703 1.602 1.545 1.508 1.426 1.379 1.347 1.325
100 1.792 1.746 1.708 1.676 1.573 1.515 1.477 1.392 1.342 1.308 1.283
200 1.742 1.694 1.656 1.623 1.516 1.455 1.415 1.321 1.263 1.221 1.189
500 1.712 1.664 1.625 1.592 1.482 1.419 1.376 1.275 1.210 1.159 1.113
∞ 1.692 1.644 1.604 1.571 1.459 1.394 1.350 1.244 1.170 1.107 1.010

145
B Statistical Tables

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 .

𝛼 = 0.025, 1 − 𝛼 = 0.975
𝑟 2 \𝑟 1 1 2 3 4 5 6 7 8 9 10 12
1 647.8 799.5 864.2 899.6 921.8 937.1 948.2 956.7 963.3 968.6 976.7
2 38.51 39.00 39.17 39.25 39.30 39.33 39.36 39.37 39.39 39.40 39.41
3 17.44 16.04 15.44 15.10 14.88 14.73 14.62 14.54 14.47 14.42 14.34
4 12.22 10.65 9.979 9.605 9.364 9.197 9.074 8.980 8.905 8.844 8.751
5 10.01 8.434 7.764 7.388 7.146 6.978 6.853 6.757 6.681 6.619 6.525
6 8.813 7.260 6.599 6.227 5.988 5.820 5.695 5.600 5.523 5.461 5.366
7 8.073 6.542 5.890 5.523 5.285 5.119 4.995 4.899 4.823 4.761 4.666
8 7.571 6.059 5.416 5.053 4.817 4.652 4.529 4.433 4.357 4.295 4.200
9 7.209 5.715 5.078 4.718 4.484 4.320 4.197 4.102 4.026 3.964 3.868
10 6.937 5.456 4.826 4.468 4.236 4.072 3.950 3.855 3.779 3.717 3.621
11 6.724 5.256 4.630 4.275 4.044 3.881 3.759 3.664 3.588 3.526 3.430
12 6.554 5.096 4.474 4.121 3.891 3.728 3.607 3.512 3.436 3.374 3.277
13 6.414 4.965 4.347 3.996 3.767 3.604 3.483 3.388 3.312 3.250 3.153
14 6.298 4.857 4.242 3.892 3.663 3.501 3.380 3.285 3.209 3.147 3.050
15 6.200 4.765 4.153 3.804 3.576 3.415 3.293 3.199 3.123 3.060 2.963
16 6.115 4.687 4.077 3.729 3.502 3.341 3.219 3.125 3.049 2.986 2.889
17 6.042 4.619 4.011 3.665 3.438 3.277 3.156 3.061 2.985 2.922 2.825
18 5.978 4.560 3.954 3.608 3.382 3.221 3.100 3.005 2.929 2.866 2.769
19 5.922 4.508 3.903 3.559 3.333 3.172 3.051 2.956 2.880 2.817 2.720
20 5.871 4.461 3.859 3.515 3.289 3.128 3.007 2.913 2.837 2.774 2.676
22 5.786 4.383 3.783 3.440 3.215 3.055 2.934 2.839 2.763 2.700 2.602
24 5.717 4.319 3.721 3.379 3.155 2.995 2.874 2.779 2.703 2.640 2.541
26 5.659 4.265 3.670 3.329 3.105 2.945 2.824 2.729 2.653 2.590 2.491
28 5.610 4.221 3.626 3.286 3.063 2.903 2.782 2.687 2.611 2.547 2.448
30 5.568 4.182 3.589 3.250 3.026 2.867 2.746 2.651 2.575 2.511 2.412
40 5.424 4.051 3.463 3.126 2.904 2.744 2.624 2.529 2.452 2.388 2.288
50 5.340 3.975 3.390 3.054 2.833 2.674 2.553 2.458 2.381 2.317 2.216
60 5.286 3.925 3.343 3.008 2.786 2.627 2.507 2.412 2.334 2.270 2.169
80 5.218 3.864 3.284 2.950 2.730 2.571 2.450 2.355 2.277 2.213 2.111
100 5.179 3.828 3.250 2.917 2.696 2.537 2.417 2.321 2.244 2.179 2.077
200 5.100 3.758 3.182 2.850 2.630 2.472 2.351 2.256 2.178 2.113 2.010
500 5.054 3.716 3.142 2.811 2.592 2.434 2.313 2.217 2.139 2.074 1.971
∞ 5.024 3.689 3.116 2.786 2.567 2.408 2.288 2.192 2.114 2.048 1.945

146
 B.5 Central F-Distribution

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 (continued).

𝛼 = 0.025, 1 − 𝛼 = 0.975
𝑟 2 \𝑟 1 14 16 18 20 30 40 50 100 200 500 ∞
1 982.5 986.9 990.3 993.1 1001. 1006. 1008. 1013. 1016. 1017. 1018.
2 39.43 39.44 39.44 39.45 39.46 39.47 39.48 39.49 39.49 39.50 39.50
3 14.28 14.23 14.20 14.17 14.08 14.04 14.01 13.96 13.93 13.91 13.90
4 8.684 8.633 8.592 8.560 8.461 8.411 8.381 8.319 8.289 8.270 8.257
5 6.456 6.403 6.362 6.329 6.227 6.175 6.144 6.080 6.048 6.028 6.015
6 5.297 5.244 5.202 5.168 5.065 5.012 4.980 4.915 4.882 4.862 4.849
7 4.596 4.543 4.501 4.467 4.362 4.309 4.276 4.210 4.176 4.156 4.142
8 4.130 4.076 4.034 3.999 3.894 3.840 3.807 3.739 3.705 3.684 3.670
9 3.798 3.744 3.701 3.667 3.560 3.505 3.472 3.403 3.368 3.347 3.333
10 3.550 3.496 3.453 3.419 3.311 3.255 3.221 3.152 3.116 3.094 3.080
11 3.359 3.304 3.261 3.226 3.118 3.061 3.027 2.956 2.920 2.898 2.883
12 3.206 3.152 3.108 3.073 2.963 2.906 2.871 2.800 2.763 2.740 2.725
13 3.082 3.027 2.983 2.948 2.837 2.780 2.744 2.671 2.634 2.611 2.596
14 2.979 2.923 2.879 2.844 2.732 2.674 2.638 2.565 2.526 2.503 2.487
15 2.891 2.836 2.792 2.756 2.644 2.585 2.549 2.474 2.435 2.411 2.395
16 2.817 2.761 2.717 2.681 2.568 2.509 2.472 2.396 2.357 2.333 2.316
17 2.753 2.697 2.652 2.616 2.502 2.442 2.405 2.329 2.289 2.264 2.248
18 2.696 2.640 2.596 2.559 2.445 2.384 2.347 2.269 2.229 2.204 2.187
19 2.647 2.591 2.546 2.509 2.394 2.333 2.295 2.217 2.176 2.150 2.133
20 2.603 2.547 2.501 2.464 2.349 2.287 2.249 2.170 2.128 2.103 2.085
22 2.528 2.472 2.426 2.389 2.272 2.210 2.171 2.090 2.047 2.021 2.003
24 2.468 2.411 2.365 2.327 2.209 2.146 2.107 2.024 1.981 1.954 1.935
26 2.417 2.360 2.314 2.276 2.157 2.093 2.053 1.969 1.925 1.897 1.878
28 2.374 2.317 2.270 2.232 2.112 2.048 2.007 1.922 1.877 1.848 1.829
30 2.338 2.280 2.233 2.195 2.074 2.009 1.968 1.882 1.835 1.806 1.787
40 2.213 2.154 2.107 2.068 1.943 1.875 1.832 1.741 1.691 1.659 1.637
50 2.140 2.081 2.033 1.993 1.866 1.796 1.752 1.656 1.603 1.569 1.545
60 2.093 2.033 1.985 1.944 1.815 1.744 1.699 1.599 1.543 1.507 1.482
80 2.035 1.974 1.925 1.884 1.752 1.679 1.632 1.527 1.467 1.428 1.400
100 2.000 1.939 1.890 1.849 1.715 1.640 1.592 1.483 1.420 1.378 1.347
200 1.932 1.870 1.820 1.778 1.640 1.562 1.511 1.393 1.320 1.269 1.229
500 1.892 1.830 1.779 1.736 1.596 1.515 1.462 1.336 1.254 1.192 1.137
∞ 1.866 1.803 1.752 1.709 1.566 1.484 1.429 1.296 1.206 1.128 1.012

147
B Statistical Tables

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 .

𝛼 = 0.01, 1 − 𝛼 = 0.99
𝑟 2 \𝑟 1 1 2 3 4 5 6 7 8 9 10 12
1 4052. 4999. 5403. 5625. 5764. 5859. 5928. 5981. 6022. 6056. 6106.
2 98.50 99.00 99.17 99.25 99.30 99.33 99.36 99.37 99.39 99.40 99.42
3 34.12 30.82 29.46 28.71 28.24 27.91 27.67 27.49 27.35 27.23 27.05
4 21.20 18.00 16.69 15.98 15.52 15.21 14.98 14.80 14.66 14.55 14.37
5 16.26 13.27 12.06 11.39 10.97 10.67 10.46 10.29 10.16 10.05 9.888
6 13.75 10.92 9.780 9.148 8.746 8.466 8.260 8.102 7.976 7.874 7.718
7 12.25 9.547 8.451 7.847 7.460 7.191 6.993 6.840 6.719 6.620 6.469
8 11.26 8.649 7.591 7.006 6.632 6.371 6.178 6.029 5.911 5.814 5.667
9 10.56 8.022 6.992 6.422 6.057 5.802 5.613 5.467 5.351 5.257 5.111
10 10.04 7.559 6.552 5.994 5.636 5.386 5.200 5.057 4.942 4.849 4.706
11 9.646 7.206 6.217 5.668 5.316 5.069 4.886 4.744 4.632 4.539 4.397
12 9.330 6.927 5.953 5.412 5.064 4.821 4.640 4.499 4.388 4.296 4.155
13 9.074 6.701 5.739 5.205 4.862 4.620 4.441 4.302 4.191 4.100 3.960
14 8.862 6.515 5.564 5.035 4.695 4.456 4.278 4.140 4.030 3.939 3.800
15 8.683 6.359 5.417 4.893 4.556 4.318 4.142 4.004 3.895 3.805 3.666
16 8.531 6.226 5.292 4.773 4.437 4.202 4.026 3.890 3.780 3.691 3.553
17 8.400 6.112 5.185 4.669 4.336 4.102 3.927 3.791 3.682 3.593 3.455
18 8.285 6.013 5.092 4.579 4.248 4.015 3.841 3.705 3.597 3.508 3.371
19 8.185 5.926 5.010 4.500 4.171 3.939 3.765 3.631 3.523 3.434 3.297
20 8.096 5.849 4.938 4.431 4.103 3.871 3.699 3.564 3.457 3.368 3.231
22 7.945 5.719 4.817 4.313 3.988 3.758 3.587 3.453 3.346 3.258 3.121
24 7.823 5.614 4.718 4.218 3.895 3.667 3.496 3.363 3.256 3.168 3.032
26 7.721 5.526 4.637 4.140 3.818 3.591 3.421 3.288 3.182 3.094 2.958
28 7.636 5.453 4.568 4.074 3.754 3.528 3.358 3.226 3.120 3.032 2.896
30 7.562 5.390 4.510 4.018 3.699 3.473 3.304 3.173 3.067 2.979 2.843
40 7.314 5.179 4.313 3.828 3.514 3.291 3.124 2.993 2.888 2.801 2.665
50 7.171 5.057 4.199 3.720 3.408 3.186 3.020 2.890 2.785 2.698 2.562
60 7.077 4.977 4.126 3.649 3.339 3.119 2.953 2.823 2.718 2.632 2.496
80 6.963 4.881 4.036 3.563 3.255 3.036 2.871 2.742 2.637 2.551 2.415
100 6.895 4.824 3.984 3.513 3.206 2.988 2.823 2.694 2.590 2.503 2.368
200 6.763 4.713 3.881 3.414 3.110 2.893 2.730 2.601 2.497 2.411 2.275
500 6.686 4.648 3.821 3.357 3.054 2.838 2.675 2.547 2.443 2.356 2.220
∞ 6.635 4.605 3.782 3.319 3.017 2.802 2.640 2.511 2.408 2.321 2.185

148
 B.5 Central F-Distribution

Computation of critical value 𝑘𝛼 = 𝐹 1−𝛼;𝑟 1,𝑟 2,𝜆=0 (continued).

𝛼 = 0.01, 1 − 𝛼 = 0.99
𝑟 2 \𝑟 1 14 16 18 20 30 40 50 100 200 500 ∞
1 6143. 6170. 6192. 6209. 6261. 6287. 6303. 6334. 6350. 6360. 6366.
2 99.43 99.44 99.44 99.45 99.47 99.47 99.48 99.49 99.49 99.50 99.50
3 26.92 26.83 26.75 26.69 26.50 26.41 26.35 26.24 26.18 26.15 26.13
4 14.25 14.15 14.08 14.02 13.84 13.75 13.69 13.58 13.52 13.49 13.46
5 9.770 9.680 9.610 9.553 9.379 9.291 9.238 9.130 9.075 9.042 9.021
6 7.605 7.519 7.451 7.396 7.229 7.143 7.091 6.987 6.934 6.902 6.880
7 6.359 6.275 6.209 6.155 5.992 5.908 5.858 5.755 5.702 5.671 5.650
8 5.559 5.477 5.412 5.359 5.198 5.116 5.065 4.963 4.911 4.880 4.859
9 5.005 4.924 4.860 4.808 4.649 4.567 4.517 4.415 4.363 4.332 4.311
10 4.601 4.520 4.457 4.405 4.247 4.165 4.115 4.014 3.962 3.930 3.909
11 4.293 4.213 4.150 4.099 3.941 3.860 3.810 3.708 3.656 3.624 3.603
12 4.052 3.972 3.909 3.858 3.701 3.619 3.569 3.467 3.414 3.382 3.361
13 3.857 3.778 3.716 3.665 3.507 3.425 3.375 3.272 3.219 3.187 3.166
14 3.698 3.619 3.556 3.505 3.348 3.266 3.215 3.112 3.059 3.026 3.004
15 3.564 3.485 3.423 3.372 3.214 3.132 3.081 2.977 2.923 2.891 2.869
16 3.451 3.372 3.310 3.259 3.101 3.018 2.967 2.863 2.808 2.775 2.753
17 3.353 3.275 3.212 3.162 3.003 2.920 2.869 2.764 2.709 2.676 2.653
18 3.269 3.190 3.128 3.077 2.919 2.835 2.784 2.678 2.623 2.589 2.566
19 3.195 3.116 3.054 3.003 2.844 2.761 2.709 2.602 2.547 2.512 2.489
20 3.130 3.051 2.989 2.938 2.778 2.695 2.643 2.535 2.479 2.445 2.421
22 3.019 2.941 2.879 2.827 2.667 2.583 2.531 2.422 2.365 2.329 2.306
24 2.930 2.852 2.789 2.738 2.577 2.492 2.440 2.329 2.271 2.235 2.211
26 2.857 2.778 2.715 2.664 2.503 2.417 2.364 2.252 2.193 2.156 2.132
28 2.795 2.716 2.653 2.602 2.440 2.354 2.300 2.187 2.127 2.090 2.064
30 2.742 2.663 2.600 2.549 2.386 2.299 2.245 2.131 2.070 2.032 2.006
40 2.563 2.484 2.421 2.369 2.203 2.114 2.058 1.938 1.874 1.833 1.805
50 2.461 2.382 2.318 2.265 2.098 2.007 1.949 1.825 1.757 1.713 1.683
60 2.394 2.315 2.251 2.198 2.028 1.936 1.877 1.749 1.678 1.633 1.601
80 2.313 2.233 2.169 2.115 1.944 1.849 1.788 1.655 1.579 1.530 1.494
100 2.265 2.185 2.120 2.067 1.893 1.797 1.735 1.598 1.518 1.466 1.427
200 2.172 2.091 2.026 1.971 1.794 1.694 1.629 1.481 1.391 1.328 1.279
500 2.117 2.036 1.970 1.915 1.735 1.633 1.566 1.408 1.308 1.232 1.165
∞ 2.082 2.000 1.934 1.878 1.697 1.592 1.523 1.358 1.248 1.153 1.015

Calculation in Matlab:
𝑘𝛼 = ❢✐♥✈ (1 − 𝛼, 𝑟 1, 𝑟 2 )

Example (𝛼 = 0.05, one-sided):

𝑘𝛼 = 2.774 = ❢✐♥✈ (1 − 0.05, 20, 10)

1
𝑘 1−𝛼 = 0.360 = = ❢✐♥✈ (0.05, 10, 20)
❢✐♥✈ (1 − 0.05, 20, 10)

149
B Statistical Tables

B.6 Relation between F-Distribution and other distrubutions

2
𝜒 2 -Distribution 𝜒 1−𝛼;𝑟 = 𝑟 𝐹 1−𝛼;𝑟,∞
p
Standard Normal Distribution 𝑧 1−𝛼/2 = 𝐹 1−𝛼;1,∞
p
t-Distribution 𝑡 1−𝛼/2;𝑟 = 𝐹 1−𝛼;1,𝑟

𝜏-Distribution 𝜏 r
𝑟 𝐹 1−𝛼 ;𝑞,𝑟 −𝑞,𝜆
1−𝛼;𝑞,𝑟 −𝑞,𝜆= 𝑟 −𝑞+𝑞𝐹
1−𝛼 ;𝑞,𝑟 −𝑞,𝜆

150
C Book recommendations and other material

C.1 Scientific books

• Anderson, James M. and Edward M. Mikhail

Surveying. Theory and Practice
7th edition
McGraw-Hill, 1998
ISBN 0-07-015914-9
• Benning, Wilhelm
Statistik in Geodäsie, Geoinformation und Bauwesen
4., überarbeitete und erweiterte Auflage
Wichmann, 2011
ISBN 978-3-87907-512-6
• Beucher, Ottmar
Wahrscheinlichkeitsrechnung und Statistik mit Matlab
Springer, 2007
ISBN 978-3-540-72155-0
• Caspary, Wilhelm and Klaus Wichmann
Auswertung von Messdaten. Statistische Methoden für Geo- und Ingenieurwis-
senschaften
Oldenbourg, 2007
ISBN 978-3-486-58351-9
• Chatterjee, Samprit and Ali S. Hadi
Regresson Analysis by Example
Fourth Edition
John Wiley & Sons, Inc., 2006
ISBN 13 978-0-471-74696-6
• Ghilani Charles D.
Adjustment Computations. Spatial Data Analysis
Fifth Edition
John Wiley & Sons, Inc., 2010
ISBN 978-0-470-46491-5

151
C Book recommendations and other material

• Grafarend, Erik W.
Linear and Nonlinear Models – Fixed Effects, Random Effects, and Mixed Models
de Gruyter, 2006
ISBN 978-3-11-016216-5
• Jäger, Reiner et al.
Klassische und robuste Ausgleichungsverfahren. Ein Leitfaden für Ausbildung und
Praxis von Geodäten und Geoinformatikern
Wichmann, 2005
ISBN 3-87907-370-8
• Koch, Karl-Rudolf
Parameter Estimation and Hypothesis Testing in Linear Models
2nd updated and enlarged edition
Springer, 1999
ISBN 978-3-540-65257-1
• Koch, Karl-Rudolf
Parameterschätzung und Hypothesentests in linearen Modellen
Dritte, bearbeitete Auflage
Dümmlers, 1997
ISBN 3-427-78923-3
• Lay, David C.
Linear Algebra and its Applications
3rd edition
Addison-Wesley Publishing Company, 2003
ISBN 0-201-70970-8
• Magnus, Jan R. and Heinz Neudecker
Matrix Differential Calculus with Applications in Statistics and Ecnonometrics
John Wiley & Sons Ltd., 1988
ISBN 0-471-91516-5
• Mikhail, Edward M. and Fritz Ackermann
Observations and Least Squares
IEP-A Dun-Donnelley Publisher, 1976
ISBN 0-7002-2481-5
• Niemeier, Wolfgang
Ausgleichungsrechnung, statistische Auswertemethoden
2., überarbeitete und erweiterte Auflage
de Gruyter, 2008
ISBN 978-3-11-019055-7
• Strang, Gilbert
Linear Algebra and its Applications
4th edition

152
 C.2 Popular science books, literature

Harcourt Brace & Company, 2005

ISBN 0030105676
• Strang, Gilbert
Introduction to Linear Algebra
4th edition
Wellesley-Cambridge Press, 2009
ISBN 078-0-9802327-1-4
• Strang, Gilbert and Kai Borre
Linear Algebra, Geodesy, and GPS
Wellesley-Cambridge Press, 1997
ISBN 0-9614088-6-3
• Teunissen, P. J. G.
Adjustment theory – an introduction
Delft University Press, 2003
ISBN 13 978-90-407-1974-5
• Teunissen, P. J. G.
Testing theory – an introduction
Delft University Press, 2000–2006
ISBN 13 978-90-407-1975-2
• Teunissen, P. J. G.
Dynamic data processing – recursive least-squares
Delft University Press, 2001
ISBN 13 978-90-407-1976-9
• Wolf, H.
Ausgleichungsrechnung II, Aufgaben und Beispiele zur praktischen Anwendung
Ferd. Dümmlers Verlag, 1979
ISBN 3-427-78361-8

C.2 Popular science books, literature

• Sobel, Dava
Longitude: The True Story of a Lone Genius Who Solved the Greatest Scientific Prob-
lem of His Time
Fourth Estate, 1996
ISBN 1-85702-502-4
Deutsche Übersetzung:
Längengrad. Die wahre Geschichte eines einsamen Genies, welches das größte wis-
senschaftliche Problem seiner Zeit löste
Berliner Taschenbuch Verlag, 2003
ISBN 3-8333-0271-2

153
C Book recommendations and other material

• Kehlmann, Daniel
Die Vermessung der Welt
Rowohlt, Reinbek, 2005
ISBN 3-498-03528-2

C.3 Other material

• Krumm, Friedhelm
Geodetic Network Adjustment Examples
❤tt♣✿✴✴✇✇✇✳❣✐s✳✉♥✐✲st✉tt❣❛rt✳❞❡✴❧❡❤r❡✴❝❛♠♣✉s✲❞♦❝s✴❛❞❥✉st♠❡♥t❴❡①❛♠♣❧❡s✳♣❞❢
• Strang, Gilbert
Linear Algebra, MIT Course 18.06
❤tt♣✿✴✴♦❝✇✳♠✐t✳❡❞✉✴❝♦✉rs❡s✴♠❛t❤❡♠❛t✐❝s✴✶✽✲✵✻✲❧✐♥❡❛r✲❛❧❣❡❜r❛✲s♣r✐♥❣✲✷✵✶✵✴

154