Adjustment Theory
Nico Sneeuw
Geodatisches Institut
Universitat Stuttgart
November 4, 2007
Contents
1 Introduction
1.1 Adjustment theory a first look . . . . . . . . . . . . . . . . . . . . . . .
1.2 Historical development . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
5
8
. . . . . .
. . . . . .
. . . . . .
equations
4.1.2 Ubertragung
auf Ausgleichungsrechnung
4.1.3 Higher dimensions . . . . . . . . . . . .
4.2 Weighted condition equations . . . . . . . . . .
4.3 Stochastics . . . . . . . . . . . . . . . . . . . .
4.4 Best Linear Unbiased Estimation (blue) . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22
22
25
28
33
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
41
42
43
43
46
47
.
.
.
.
.
.
50
50
52
53
54
54
60
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
Contents
6.1
6.2
6.3
6.4
Expectation of sum
Basics . . . . . . .
Hypotheses . . . .
Distributions . . .
of squared residuals
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
7 Statistical Testing
7.1 Global model test: a first approach
7.2 Testing procedure . . . . . . . . . .
7.3 DIA-Testprinciple . . . . . . . . . .
7.4 Internal reliability . . . . . . . . .
7.5 External reliability . . . . . . . . .
7.6 Reliability: a synthesis . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
67
68
71
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
73
73
75
84
85
88
89
.
.
.
.
.
.
92
92
92
92
93
93
94
8 Recursive estimation
8.1 Partitioned model . . . . . . . . . . . . . . . . . .
8.1.1 Batch / offline / Stapel / standard . . . . .
8.1.2 Rekursiv / sequentiell / real-time . . . . . .
8.1.3 Umformen . . . . . . . . . . . . . . . . . . .
8.1.4 Formulierung nach Bedingungsgleichungen .
8.2 Allgemeiner . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A Partitioning
96
A.1 Inverse Partitioning Method . . . . . . . . . . . . . . . . . . . . . . . . . . 96
A.2 Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
B Buchempfehlungen
98
B.1 Wissenschaftliche B
ucher . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
B.2 Popularwissenschaftliche B
ucher, Literatur . . . . . . . . . . . . . . . . . . 99
1 Introduction
Adjustment theory deals with the optimal combination of redundant measurements together with the estimation of unknown parameters.
(Teunissen, 2000)
2 equations in 2 unknowns:
5 = 3x1 + 4x2
6 = 5x1 + 2x2
5
34
x1
as matrix-vector system:
=
6
52
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
2 4
5
1
x1
x = A1 y
=
=
x2
6
0.5
14 5 3
So each apple costs 1e 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.
Ausgleichungsrechnung
1 Introduction
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.
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 = 5e
5
34
x1
=
6a + 8p = 10e
10
68
x2
The matrix A has linearly dependent columns (and rows), i.e. it is singular. Correspondingly det A = 0 and the inverse A1 does not exist. The observations (5e and 10e) 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 8e 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.
x1
34
5
6 = 5 2 x2
12
3
6
mn n1
(1.1)
m1
grober Fehler
systematischer F.
Zufallsfehler
1 Introduction
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
bulls eye.
Zufallsvariable
Whatever the type, 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.
y = Ax
L
y = L
nm m1
invert:
A x
nm mn n1
x = (LA)1 Ly
= By
nn
n(mn)
The trouble with this approach, obviously, is the arbitrariness of the choice of n observations. There are m
n 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 ships longitude. Tobias Mayers1 approach was to determine longitude, or rather
time, through the motion 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
271
273
x
31
Mayer called them equations of conditions, which is, from todays view point, an unfortunate designation.
Mayers 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 0
L = 0 0 0 0 1 1 1 0 0 0 0
327
0 0 0 0 0 0 0 0 1 1 1
1
Tobias Mayer (17231762) 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
G
ottingen, 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.
1 Introduction
Mayer actually believed each aggregate of 9 observations to be9 times more precise
than a single observation. Today we know that this should be 9 = 3.
Eulers 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 15821745; 6 unknowns Given up!
Euler was mathematician Error bounds
Laplaces attempt ca. 1787
Laplace3
Background: Saturn, too
Reformulated: 4 unknowns
Best Data: 24 Observations
Approach: like Mayer, but other combinations:
= A x
241
244 41
= L
424 241
A x
424 244 41
x = (LA)1 Ly
Euler (17071783) 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 (17491827) 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 Mecanique Celeste (Celestial Mechanics) (17991825). 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 Laplaces 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
L =
424
1
1
1
0
1
1
0
1
1
1
1
0
1 1 1 1
1 1 1 1
1 0 0 1
0 1 1 0
1
1
0
1
1
1
0
1
1
1
1
0
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 0 0 1 0 0 1 1 0 1 0 0 1 1
0 1 1 0 1 1 0 0 1 0 1 1 0 0
M () =
a(1 e2 )
(1 e2 sin2 ) 2
3
= a(1 e2 )(1 + e2 sin2 + ...)
2
M (0) = a(1 e2 ) < a
M ( ) = a 1e2 = a > a
3
2
1e2
(1e2 ) 2
= x1 + sin2 x2
5!
=
First attempt All 52 = 2!(52)!
10 systems of equations (2 2)
10 solutions
Comparison of results.
54321
21321
His result: gross variations of the ellipticity Reject the ellipsoidal hypothesis.
Second attempt
ei = 0 ,
i=1
|ei | = min .
Rudjer Josip Boskovic aka. Roger Boscovich (17111787) 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.
11
1 Introduction
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
Methode der kleinsten
Quadrate
In 1805 Legendre5 published his method of least squares (in French: moindres carres).
The name least squares refers to the fact the sum of square residuals is minimized.
Legendre developed the method 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:
E =
5
X
i=1
L = AT
T
x
= (A
A})1 AT y
{z
|
nm m1
n1
nn
After Legendres 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 Gausss claim of priority is very likely valid but that he refrained from publication
because he found his results still premature.
12
13
vermittelnde
Ausgleichung
direkte
Beobachtungen
The column vector a spans up a line y = ax in R2 . This line is the 1D model space
or range space of A: R(A). Inconsistency of the observation vector means that y does
not lie on this line. Instead, 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.
y=ax
e?
e
e
ax
e
ax=y
ax
ax
ax
ax?
Figure 2.1
Adjustment of observations is about finding the optimal e and x. An intuitive choice
for optimality is to make the vector e as short asP
possible. The shortest possible e is
Te
indicated
by
a
hat:
e
.
The
squared
length
e
=
2i is the smallest of all possible
ie
P
eT e = i e2i , which explains the name least squares. If e is determined, we will at the
same time know the optimal x
.
How do we get the shortest e? The right panel of fig. 2.1 show that the shortest e is
perpendicular to a:
e a
14
aT e = 0
(2.3a)
aT (y a
x) = 0
(2.3b)
aT a
x = aT y
(2.3c)
LS estimate of x
x
= (aT a)1 aT y
LS estimate of y
LS estimate of e
sum square residuals
(2.3d)
1 T
y = a
x = a(a a)
a y
T
(2.3e)
1 T
e = y y = [I a(a a)
eT e = y T [I a(aT a)1 aT ]y
a ]y
(2.3f)
(2.3g)
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) in the last line indeed follows from
the line above. Two things should be shown: that P is symmetric, and that P P = 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) = eT e ,
2
(2.4)
which is half of the sum of square residuals. Its graph would be a parabola. The factor
1
2 shouldnt worry us. If we find the minimum La , then any scaled version of it is also
minimized. The task is now to find the x
that minimizes the Lagrangian. With e = yax
we get the minimization problem:
1
min La (x) = min (y ax)T (y ax)
x
x
2
1 T
1 T 2
T
= min
y y xa y + a ax
.
x
2
2
15
with
y = a(aT a)1 aT y
(2.5a)
y = Pa y
(2.5b)
Pa = a(aT a)1 aT .
(2.5c)
(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:
e = y Pa y = (I Pa ) y = Pa y ,
which is also a projection. In order to give e 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 .
16
eT e = 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 .
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 = Pa y + Pa y = y + e ,
which is visualized in fig. 2.1.
Numerical example
With a = (1 1)T we will follow the steps from (2.3a):
(aT a)
x = aT y
x
= (aT a)1 aT y
y = a(aT a)1 aT y
e = y y
eT e
2
x = y1 + y2
x
= 12 (y1 + y2 )
(average)
y1
y1 + y2
1
= 2
y2
y1 + y2
e1
y1 y2
= 12
(error distribution)
e2
y1 + y2
21 (y1 y2 )2
(least squares)
Pa =
and Pa = I Pa =
2 11
2 1 1
and check the equations y = Pa y and e = Pa y with the numerical results above.
17
zerlegen
Widerspruch
y1
y2
bT y = 0 .
12 21
(2.7)
11
In reality, though, both observations do contain errors, i.e. they are not equal: y1 y2 6= 0
or bT y 6= 0. Instead of 0 one would obtain a misclosure w. If we recast the observation
equation into y e = ax, it is clear that it is (y e) that has to obey the above condition:
bT (y e) = 0
Bedingungsgleichung
=0
w := bT y = bT e .
(2.8)
In this condition equation the vector e is unknown. The task of adjustment according
to the model of condition equations is to find the smallest possible e that fulfills the
condition (2.8). At this stage, 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 e is perpendicular to the line
bT e = w. So e lies on a line through b.
Sch
atzer
(2.9a)
1 T
y = y e = y b(b b)
b y
= [I b(bT b)1 bT ]y = Pb y
eT e = y T Pb y = y T b(bT b)1 bT y
(2.9b)
(2.9c)
Exercise 2.3 Confirm that the orthogonal projector Pb is idempotent and verify that
the equation for eT e is correct.
18
b e=b y
bTe=bTy
y
y
ee
e?
b
Figure 2.2
Numerical example
With bT = 1 1 we get
Pb = b(bT b)1 bT =
=
e = Pb y =
Pb = I Pb =
=
1
(bT b)1 =
2
1
1
1 1
1
1 1
=
2 1
2 1 1
1
y1 y2
2 y1 + y2
1
1 11
10
1 1
=
01
2 1 1
2 11
1 y1 + y2
2 y1 + y2
bT b = 2
y = Pb y =
These results for y and e are the same as those for the adjustment with observation
equations. The estimator y describes the mean of the two observations, whereas the
estimator e distributes the inconsistencies equally. Also note that Pb = Pa and vice
versa.
19
min Lb (e, ) =
e,
1 T
e e + (bT y bT e) .
2
(2.10a)
(2.10b)
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
. 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 .
that minimize the extended
The minimization problem now exists in finding the e and
Lb . Thus we need to derive the partial derivatives of Lb towards e and . Next, we
impose the conditions that these partial derivatives are zero when evaluated in e and .
L
=0
(
e, ) = 0 =
e b
e
L
(
e, ) = 0 = bT y bT e = 0
(2.11)
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:
=0
e b
.
e = b
= (bT b)1 bT y .
20
2.3 Synthesis
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 1
1
T
b a=
= 0,
1
(2.12)
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
bT a=0
y = ax + e bT y = bT ax + bT e bT y = bT e .
T
b e=b y
y
e
y=ax
y=Pa y=Pb y
e=Pb y=Pa y
21
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 + a0 + e
and B T (y e) = b0 .
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. Mathematically we have rank A = 0 so that the normal matrix has det AT A = 0
and is not invertible.
A = [ a1 , a2 , a3 , . . . , an ]
mn
22
m1
m1
m1
m1
y =
m1
n
X
i=1
ai xi + e ,
m1 11
m1
y3
y R(A)
a1
y
e1
e2
a2
y2
Ax=y
a2
Ax?
Ax?
a1
y1
Figure 3.1
e = PA y = [I A(AT A)1 AT ]y
y = PA y = A(AT A)1 AT y = A
x
x
= (AT A)1 AT y
(AT A)1 exists iff rank A = n = rank(AT A)
23
3 Generalizations
Calculus
1 T
e e
2
1
=
(y Ax)T (y Ax)
2
1 T
1
1
1
=
y y y T Ax xT AT y + xT AT Ax
2
2
2
2
LA (x) =
min
L
(
x) = 0 = e = y y = [I A(AT A)1 AT ]y = PA y
x
PA idempotent?
PA PA = [I A(AT A)1 AT ][I A(AT A)1 AT ]
= I A(A A)
= PA
y = PA y = A(AT A)1 AT y
km
0,9
0,8
1k
5k
PB
PA
P2
P3
0,5km
Figure 3.2
24
h1B = HB H1 + e1B
h13 = H3 H1 + e13
h12 = H2 H1 + e12
h32 = H2 H3 + e32
h1A = HA H1 + e1A
hT := [h1B , h13 , h12 , h32 , h1A ] vector of levelled height differences
H1 , H2 , H3 unknown heights of points P1 , P2 , P3
HA , HB given bench marks
In matrix notation:
h1B
1
h13 1
h12 = 1
h32 0
h1A
1
0 0
H1
HB
e1B
0 e13
0 1
H2
+ e12
1 0 H3
0
+
0 e32
1 1
0 0
HA
e1A
1
h1B HB
h
13
1
= 1
h12
h32
h1A HA
1
e1B
H1
0 0
e13
0 1
H2
e
1 0 H3
+
12
e32
1 1
e1A
0 0
(or y = Ax + e)
y =A x + e
51
53 31
51
25
3 Generalizations
1 0 0 0 1
0 1 1 1 0
h1B
1 0 0 0 1
HB
h13
0 1 1 1 0
0
h12 =
0 .
0
h32
h1A
HA
1 0 0 0 1
0 1 1 1 0
h1B e1B
1 0 0 0 1
HB
h13 e13
0 1 1 1 0
0
h12 e12 =
0 .
h32 e32
0
h1A e1A
HA
or
B
25
h e
51
51
= BTc .
21
= B T h B T e = B T c
B T e = B T h B T c
B T e = B T y =: w
rm m1
26
r1
1 T
e
e + T ( B T e w ) min
e,
2 1m m1 1r rm m1
r1
|
{z
}
11
LB
=
(
e, )
e
= 0
e + B
m1
mr r1
m1
LB
= B T e w = 0
(
e, )
(w = B T y)
rm m1
r1
r1
e
I
B
0
mr
= m1
= mm
BT 0
w
rm
rr
= B T B
=w
e = B
= (B T B)1 w
=
m1
rank(B T B) = b
= e = +B(B T B)1 w
= B(B T B)1 B T y
= PB y
y = y e
= [I B(B T B)1 B T ]y
= PB y
27
3 Generalizations
B T y = B T Ax + B T e B T A = 0
1 0 0 0 1
1 0 0
1 0 1
0 1 1 1 0
0
0
0
1 1 0 =
000
0 1 1
23
1 0 0
25
53
h12 = H2 H1
h12
1 1 0
H1
h13 = H3 H1 = h13 = 1 0 1 H2
h32 = H2 H3
h32
0 1 1
H3
= y = A x
31
33 31
wobei m = 3, n = 3, rank A = 2
0 1
1 0
= 1 + (1) = 0
det A = 1
(1)
1 1
1 1
= A and (AT A) are not invertible
dim N (A) > 0
Ax = 0 has a nontrivial solution = homogeneous solution xhom 6= 0
Konsequenz
math.: A(x + xhom ) = Ax + Axhom = Ax = y
| {z }
=0
28
math.: xhom
= 1
1
Rangdefekt in A = Datumproblem
- Beobachtungen nicht beeintraechtigt
- Unbekannte nicht loesbar
moegliche Loesung
- lege dim N (A) Parameter fest (Datumfestlegung oder Datumdefinition)
- Wie? Eine Spalte eliminieren und Beobachtungen anpassen
z.B. lege H1 fest
h12 + H1
1 0
H2
= h13 + H1 = 0 1 H3
h32
1 1
Approach 2: augment problem
augment solution space
1 1 1
h12
h13 = 0
h32
Redundanz = 1 = 1 Bedingung
Datumproblem?
B T y = B T Ax + B T e = 0 kein Datumproblem.
Material Krumm
im Streckennetz u
ber (Zusatz-) Bedingungsgleichungen
mn n1
m1
29
3 Generalizations
Bedingungsgleichungen zur Datumdefinition oder auch zur Erzwingung anderer Zwangsbedingungen, wobei r = n rank A (im Allgemeinen r Bedingungsgleichungen)
BT x = c
rn n1
r1
1 0 0 0 0 0 0 0 ... 0
BT = 0 0 0 0 1 0 0 0 . . . 0 ,
0 0 0 0 0 0 0 1 ... 0
0
c = 0
0
Im Allgemeinen sind auch beliebige andersartige Bedingungen zulassig (z. B. Summe der
Koordinatenzuschlage = vorgegebener Wert)
B T = 1 1 1 1 1 1 1 1 ... 1 ,
c = ...
Behandlung u
ber Lagrangefunktion (Qy = Py = I)
L(x, ) =
=
1 T
e
e + T ( B T x c ) min
x,
2 1m m1
1r
rn n1
r1
1
(y Ax)T (y Ax) + T (B T x c)
2
L
= AT y + AT A
c =! 0
(
x, )
x + B
x
L
!
= BTx
(
x, )
=0
T
T
x
A A B
A y
nr n1
n1
= nn
=
M z = v
BT 0
c
rn
rr
(n+r)(n+r)
r1
(n+r)1
r1
(n+r)1
x
= z = = M 1 v . . .
DT x = c
rn n1
e. g.
30
H1
1 0 0 H2 = H
H3
r1
1 T
e e + (DT x c)
2
1 T
1
=
y y y T Ax + xT AT Ax + (DT x c)
2
2
LD (x, ) =
LD
= AT y + AT Ax + D := 0
x
LD
= DT x c := 0
T
T
x
A AD
A y
=
= M z = v
=
DT 0
c
(n+r)(n+r)
(n+r)1
E. g.
1
A = 1
0
1 0
1
0 1 = AT A = 1
1 1
0
1 0
1
0 1 1
1 1
0
2 1 1
1 2 1
M =
1 1 2
1 0 0
1 0
2 1 1
0 1 = 1 2 1
1 1
1 1 2
1
0
0
0
1 2 1
2
1
det M = 1 det 1 1 2 = 1 1 det
= 3
1 2
1 0 0
= M regular
1
x
= AT A + DDT
AT y
L=
B T = 1 1 1 1 1 1 1 1 ... 1
1 T
e e + T (Ax + e y)
2
31
3 Generalizations
L
= e+
e
L
= Ax + e y
L
= AT
x
1
LD = (y Ax)T (y Ax) + T (DT x c)
2
T
T
x
A AD
A y
=
DT 0
c
= AT y
AT A
x + D
DDT x
= Dc
= (AT A + DDT )
x = AT y + Dc
Material Sharifi
y = A x + e
m1
mn n1
m1
DT x = 0
dn n1
d1
1 T
e e = min
2
1
= Lf (x, e) = (y Ax)T (y Ax) + T (DT x)
2
1 T
= (y xT AT )(y Ax) + T DT x
2
1
= [y T y y T Ax xT AT y + xT AT Ax] + T DT x
2
1
= [y T y 2y T Ax + xT AT Ax] + T DT x
2
Lf
x
Lf
32
T=0
= y T A + x
T AT A + D
T
= (DT x
)
=0
AT y + AT A
x + D
T
D x
=0
AT A D
DT 0
T
x
A y
=
mn nd
AH T = 0
AT AH T = 0
md
H(AT A) = 0
p
?
(xB xA )2 + (yB yA )2 y = Ax
y = f (x),
x0
1 d2 f
df
= f (x0 ) +
(x x0 ) +
(x x0 )2
dx x0
2 dx2 x0
|
{z
}
df
y y0 =
(x x0 ) + ...
dx x0
neglectable if x x0 small
df
y =
(x x0 ) +
O(x2 )
,
| {z }
dx 0
|
{z
} terms of higher order
linear model
x := x x0
= model errors
yi = fi (xj ),
i = 1, ..., m; j = 1, ..., n
33
3 Generalizations
f1
y1 =
x1 +
x1 0
f2
y2 =
x1 +
x1 0
..
.
ym
f1
x2 + +
x2 0
f2
x2 + +
x2 0
f1
xn
xn 0
f2
xn
xn 0
fm
fm
fm
=
x1 +
x2 + +
xn .
x1 0
x2 0
xn 0
y1
y2
..
.
ym
f1
x1
..
= .
x1
x1
.. x2 y = A(x )x
..
.
0
.
.
..
fm
fm
x2 xn
{z
}0 xn
f1
x2
f1
xn
Jacobian matrix A
Example:
Linearization of planar distance observation equation (given Taylor point of expansion
0 , x0 , y 0 = approximate values of unknown point coordinates); explicit differenis x0A , yA
B B
tiation
measured sAB =
q
p
2
(xB xA )2 + (yB yA )2 = x2AB + yAB
xA = x0A + xA ,
sAB
0
yA = yA
+ yA ,
0
xB = x0B + xB , yB = yB
+ yB
q
2
2
0 + y y 0 + y
=
x0B + xB x0A + xA
+ yB
B
A
A
q
2
0 y0 2
=
x0B x0A + yB
A
|
{z
}
s0AB
(distance from
approximate coordinates)
sAB
sAB
sAB
sAB
x
+
x
+
y
+
yB
A
B
A
xA 0
xB 0
yA 0
yB 0
sAB
sAB xAB
1
1
xB xA
=
= q
2xAB (1) =
xA
xAB xA
2 x2 + y 2
sAB
AB
AB
34
= sAB :=
s s0
|AB {z AB}
sAB
yB yA
,
=
yA
sAB
reduced observation
x0B x0A
s0AB
sAB
yB yA
=+
yB
sAB
0 y 0 x0 x0 y 0 y 0
yB
A
B
A
B
A
s0AB
s0AB
s0AB
y = A(x0 )x
xA
y
A
xB
yB
0 y0
yB
x0B x0A
A
(x
x
)
+
(yB yA )
B
A
s0AB
s0AB
X
f (n) (x0 )
n=0
n!
(x x0 )n
Linearization:
f (x) = f (x0 ) + f (x0 )(x x0 ) + O,
35
3 Generalizations
This results in linear model:
df
x + e = A(x0 )x + e
y =
dx x0
The datum problem again
Matrix A rank deficient (rank A < n)
mn
36
Answer b = r.
x0
linear model
x0 := x
^
PDxP2 < e
Stop criteria
x^ = x0 + Dx
No !
estimated unknown
parameters
Yes!
y^ = A x^
Error in iteration
process !
^
e
= y - y^
No !
A e^ = 0
Yes!
No !
^
y^ - f(x)
=0
Yes!
Error in iteration
process or erroneous
linearization !
Figure 3.3: Iterative scheme
37
3 Generalizations
726000
A
B
C
725000
724000
D
I
723000
F
E
722000
186000
185000
184000
183000
x
(a) distance network
A
A
B
or
D
I
Figure 3.4
38
1
w1 (y1 ax)2 + w2 (y2 ax)2
2
1
w1 0
= (y ax)T
(y ax)
0 w2
2
Eaw =
1
(y ax)T W (y ax)
2
1
1
= y T W y y T W ax + xT aT W ax
2
2
1
= eT W e
2
=
notwendige Bedingung:
dEa
(
x)
dx
dE
=
= y T W a + aT W ax = aT W y + aT W a
x=0
dx
x
: min Ea (x) =
x
= aT W a
x = aT W y Normalgleichung
hinreichende Bedingung:
d2 E
= aT W a > 0,
dx
erf
ullt, da W positiv definit
39
aT W (y a
x) = 0 = aT W e = 0
= e W a
aT W a
x = aT W y
Normalgleichung
WKQ-Schatzung von x
(gewichtete, kleinste Quadrate)
x
= (aT W a)1 aT W y
WKQ-Schatzung von y
y = a
x = a(aT W a)1 aT W y
h
i
e = y y = I a(aT W a)1 aT W y
WKQ-Schatzung von e
a=
w1 0
1
aT W = 1 1
= w1 w2
w1 0
0 w2
W =
0 w2
T
a W a = w1 w2
gewichtetes Mittel:
1
= w1 + w2
1
1
(w1 y1 + w2 y2 )
w1 + w2
w1
w2
=
y1 +
y2
w1 + w2
w1 + w2
x
=
e1
e2
y1
y2
1
=
w1 + w2
y1
y2
(w1 + w2 )y1 w1 y1 w2 y2
(w1 + w2 )y2 w1 y1 w2 y2
1
w2 (y1 y2 )
=
w1 + w2 w1 (y2 y1 )
w1 > w2 : mehr y1 als y2 in x
, |
e1 | < |
e2 |
40
4.1.1 Geometry
z2
z2
z2
T
z Wz = c
zTWz = c
z1
zTWz = c
z1
w11 = w22
w12 = 0
z1
w11 = w22
w12 = 0
(a) Kreis
w11 = w22
w12 = 0
(c) Allgemeine
Ellipse
Figure 4.1
F (z) = z T W z = c
w1 z12 + w2 z22 = c
w1 2 w2 2
z +
z =1
c 1
c 2
z12
z22
w1 + w2 = 1
c
z12 z22
+
=1
a
b
Ellipsengleichung
Eine Familie (weil c variieren kann) von Ellipsen, die im Allgemeinen nicht in Hauptachsenlage sind.
41
zTW z = c
u
bliche Ellipsenform aber:
z12 z22
+ 2 =1
a2
b
also
z12 w11 + 2w12 z1 z2 = 0, weil Hauptachsenlage + w22 z22 = c
| {z }
z2
z2
= q 1 2 + q 2 2 = 1
c
w11
c
w22
Allgemeine Ellipse
grad F (z0 ) = 2W z0
z z0 W z 0
oder
z0 T W (z z0 ) = 0
4.1.2 Ubertragung
auf Ausgleichungsrechnung
Gesucht: Vektor, der von der Linie ax startet, bei y endet und parallel zu z a oder
orthogonal zu aT W ist = e.
Aussagen:
y = a
x ist die Projektion von y
auf a
in Richtung orthogonal zu W a (entlang (W a) )
= y = Pa,(W a) y mit Pa,(W a) = a(aT W a)1 aT W
e ist die Projektion von y
auf (W a)
42
= e = P(W a) ,a y
P(W a) ,a = Pa,(W
a)
mit
= I a(aT W a)1 aT W
= [I a(aT W a)1 aT W ]y
Wegen e 6 a (oder aT e 6= 0) handelt es sich um schiefe Projektionen bzw. um
bez
uglich W orthogonale Projektionen, da ja e W a (oder aT W e = 0)
x + e
21 11
21
wird zu
y = A
m1
x + e
mn n1
m1
Ersetze a durch A!
P(Wa ) ,a = I A (AT W A)1 AT W
nm mm
mn
mm
{z
}
|
nn
43
bTe=bTy
z2
ax
e
ax
zTWz=aTWa
Wa
a
z1
b
T
a Wz=a Wa
W -1b
unter Bedingung bT e = bT y
Lagrange:
1
Lb (e, ) = eT W e + (bT e bT y)
2
Gesucht: e und , die Lb minimieren
=
44
etc.
Lb
= W e + b
=0
T
T
=b eb y =0
e, )
e (
Lb
(
e, )
b e=b y
e2
ax
e
e1
e TWe
=c1
e TWe
=c2 W -1b
e TWe
=c3
W b
bT 0
e
0
= bT y
Eliminierung 1. Zeile:
= 0 = e = W 1 b
W e + b
einsetzen:
= bT y
bT e = bT y = bT W 1 b
losen:
= (bT W 1 b)1 bT y
einsetzen:
e = W 1 b(bT W 1 b)1 bT y
h
i
y = y e = I W 1 b(bT W 1 b)1 bT y
Higher dimensions
Nur b durch B ersetzen.
b = m n Bedingungsgleichungen, Lagrange-Multiplikatoren
BTy = BTe
BTA = 0
y = Ax + e
= B T y = B T Ax + B T e = B T e
45
mm
BT
bm
m1 =
mb
bb
b1
0
BTy
e = W 1 B(B T W 1 B)1 B T y
h
i
y = I W 1 B(B T W 1 B)1 B T y
Constant term (RHS)
Ideal case without errors:
BTy = c
In reality:
B T (y e) = c = B T e = B T y c := w
e = W 1 B(B T W 1 B)1 [B T y c]
=
y = y e = etc.
4.3 Stochastics
Probabilistic formulation
Version 1:
y = Ax + e,
Version 2:
E {e} = 0,
D {e} = Qy
E y = Ax
D y = Qy
Kovarianzmatrix
46
Qz = M Qy M T
x
= (AT W A)1 AT W y
= Schatzer (LS)
= My
E {
x} = (AT W A)1 AT W E y
= (AT W A)1 AT W Ax
= x (erwartungstreuer/unverzerrter) Schatzer
Qx = M Qy M T
y = A
x
= PA y
E y = A E {
x} = Ax = E y
Qy = PA Qy PA T
e = y a
x = (I PA )y
E {
e} = E y Ax = 0
Qe = Qy PA Qy Qy PA T + PA Qy PA T
Questions:
Is x
the best estimator?
Or: When is Qx smallest?
47
E y = ax,
D y = Qy ,
L-property:
x
= lT y
U-property:
E {
x} = lT E y = lT ax = x = lT a = 1
B-property:
x
= lT y = x2 = lT Qy l
Wir suchen das l, das lT Qy l minimiert und das auf der Geraden lT a = 1 liegt.
= min lT Qy l unter lT a = 1
l
Solution?
Comparison WKQ, B-Model
min
unter
Schatzer
eT W e
= bT y = w
e = W 1 b(bT W 1 b)1 W
bT e
lT Qy l
= aT l = 1
l = Q1 a(aT Q1 a)1 /cdot1
y
y
lT a
1 T
= x
= lT y = (aT Q1
y a) a Qy y
Higher dimensions
a A,
48
Q1
y = Py
= BLUE, wenn W = Py = Q1
y
Variance-covariance propagation
x
= (AT Py A)1 AT Py y
= Qx = (AT Py A)1 AT Py Qy Py A(AT Py A)1 = (AT Py A)1
y = A
x = PA y
= Qy = A(AT Py A)1 AT Py Qy = PA Qy = PA Qy PA T = Qy PA
e = (I PA )y = PA y = y y
= Qe = Qy PA Qy Qy Py T + PA Qy PA T = PA Qy = Qy Qy
Besides:
I = PA + PA = Qy = PA Qy + PA = Qy + Qe
49
x
0
1
y
0
0
3
2
1
2
y
3
S13
T13
T12
S23
S12
50
x0
0
1
1
2
3
y0
0
0
3
2
1
2
Beobachtung
S12
S23
S13
y [Einheit]
:= 180
0.01[m]
0.03[m]
0.02[m]
0[rad]
0.3 [rad]
0.2 [rad]
Naherungswerte
0
S12
1m
0
S13
1m
0
S23
1m
0
60
0
60
0
60
Messwerte
S12
1.01 m
S13
1.02 m
S23
0.97 m
60
59.7
60.2
Koeffizientenmatrix, Designmatrix A
dx1 dy1
dx2 dy2
dx3 dy3
1
0
1
0 0
0
3
3
1
1
0
0
2 2
2
2
3
1
1
23 0
0
2
2
2
3
3
1
1
0 1
2
2
2
2
3
1
0
1
23 12
2
2
3
2
1
2
3
2
1
2
0.01 m
0.02 m
0.01 m
1
1
1
Einheiten
[]
[m1 ]
Unbekannte
[m]
dx1
dy1
dx2
dy2
dx3
dy3
Nichtlineare Beobachtungsgleichungen
Strecken:
Sij =
Winkel:
q
(xi xj )2 + (yi yj )2
= T12 T13
x2 x1
x3 x1
= arctan
arctan
y2 y1
y3 y1
= T23 T21
x3 x2
x1 x2
arctan
= arctan
y3 y2
y1 y2
= T31 T32
x1 x3
x2 x3
= arctan
arctan
y1 y3
y2 y3
Richtungswinkel:
Tij = arctan
xj xi
yj yi
Linearisierte Beobachtungsgleichungen
(Taylorpunkt = Nullstelle = Menge der Naherungskoordinaten)
51
x0i x0j
0
Sij
xi +
yi0 yj0
0
Sij
yi
x0i x0j
0
Sij
xj +
yi0 yj0
0
Sij
yj
0 . . . aus N
wobei: Sij . . . beobachtete Strecke (Mewert), Sij
aherungskoordinaten berech0
nete Strecke (Naherungsstrecke) und xi := xi xi , yi := yi yi0 etc. . . . unbekannte, im Ausgleichungsproze zu schatzende Parameter (= ausgeglichene Koordinaten
x
i = x0i +
xi etc.)
Tij = Tij0 +
1+
=
Tij0
= Tij0
x0j x0i
yj0 yi0
(yj0 yi0 )2
0 )2
(Sij
yj0 yi0
2
x0j x0i
x0j x0i
1
1
0
xi + 0
yi + 0
xj 0
yj
yj yi0
(yj yi0 )2
yj yi0
(yj yi0 )2
[. . .]
xi +
x0j x0i
yi +
yj0 yi0
xj +
x0j x0i
0 )2
0 )2
0 )2
0 )2 yj
(Sij
(Sij
(Sij
(Sij
0
0
y2 y10 y30 y10
x2 x01 x03 x01
0
0
= = T12 T13 + 0 2 +
x1 +
+
y1
(s12 )
(s013 )2
(s012 )2
(s013 )2
x02 x01
y30 y10
x03 x01
y20 y10
x
x
+
y3
2
2
3
(s012 )2
(s012 )2
(s013 )2
(s013 )2
= 0 + . . .
+
rij = Tij wi
xj xi
= arctan
wi
yj yi
0
= rij
52
yj0 yi0
(s0ij )2
xi +
x0j x0i
(s0ij )2
yi +
yj0 yi0
(s0ij )2
xj
x0j x0i
(s0ij )2
yi wi
zusatzliche Unbekannte
y
0 Teilkreis
i
Pj
Tij
rij
Pi
a sin b sin = 0
f (ea , eb , e , e ) =
f (e0a , e0b , e0 , e0 )
f
f
0
+
(ea ea ) + . . . +
(e e0 )
ea 0
e 0
w = a sin b sin
Widerspruch
(i = 1, . . . , 4; j P )
p
(xi xj )2 + (yi yj )2
= arccot
zi zj
= arccot
0
= ij
dij
zi zj
dij
zi zj
2 . . . xi + . . . yi + . . . + . . . zj
54
z
P
P4
S4
4
S3 3
S1
1
S2
P1
P2
P3
x
Figure 5.4: Hohenwinkel
gesucht: Parameter des ausgleichenden Polynoms
f (x) = y =
nX
max
an xn
n=0
nX
max
an xn + ei
n=0
55
cubic (n=3)
quadratic (n=2)
linear (n=1)
x1 x2 x3
x4 x5
x7 x8
x6
x9
x10
y1
y2
..
.
1
1
= ..
.
ym
| {z }
y
a0
x1 xn1 max
e1
x2 xn2 max
a1 e2
.. + ..
..
. .
.
n
1 xm xmmax
anmax
em
|
{z
} | {z } | {z }
1. Ausgleichungsprinzip eT e min x
56
Parabel
f (x) = a1 + 2a2 x
Tangente in xT : g(x) = a0 + a1 xT + a2 x2T + (a1 + 2a2 xT )(x xT )
Tangente in xT , die durch xP , yP geht
yP = a0 + a1 xT + a2 x2T + (a1 + 2a2 xT )(xP xT )
= B T x = yP
BT =
1 xP xT (2xP xT )
Einbinden u
ber Lagrange oder eine Unbekannte zugunsten der anderen eliminieren
Im Allgemeinfall des Polynoms von Grad n gilt
B T = 1 xP xT (2xP xT ) . . . xn1
T [nxP (n 1)xT ]
c = yP
57
yT yP
xT xP
yT yP
xT xP
=: aT
xT =: bT =
yP xT
yT x P
xT xP
= yi = aT xi + bT
Problem: ausgleichender Kreis als Spezialfall einer Ellipse gemischtes Modell
gemessen: x- und y-Koordinaten von Punkten, die nahezu auf einer Ellipse liegen
gesucht: ausgeglichene Positionen der Messpunkte, Mittelpunktkoordinaten, Halbachsen
der Ellipse
Bedingungsgleichung:
(xi + exi xM )2 (yi + eyi yM )2
+
1=0
a2
b2
Ellipsengleichung
0 , a , b , e0 , e0
Einf
uhrung von Naherungspunkten x0M , yM
0 0 x i yi
xi x0M
a0
2
0
yi yM
b0
2
0 )
2(xi x0M )
2(yi yM
xM
yM
2
2
a0
b0
0 )
2(yi yM
2(xi x0M )
+
e
eyi
x
i
a20
b20
0 )2
2(xi x0M )2
2(yi yM
a
b = 0
a30
b30
+
58
0 )
2(xi x0M ) 2(yi yM
2
2
a0
b0
2(xi x0M )
a20
e
0 )
2(yi yM
2
b0
xi
eyi
2(xi x0M )2
a30
0 )2
2(yi yM
3
b0
xM
y
M
a
b
0 )2
(xi x0M )2 (yi yM
+
1=0
a20
b20
T
B = 2
m2m
0
x1 x0M y1 yM
a20
b20
0
..
.
A = 2
m4
x1 x0M
a20
x2 x0M
a20
..
.
0
x2 x0M y2 yM
a20
b20
0
y1 yM
2
b0
0
y2 yM
2
b0
...
...
...
(x1 x0M )2
a30
(x2 x0M )2
a30
0
xm x0M ym yM
a20
b20
0 )2
(y1 yM
3
b0
0 )2
(y2 yM
3
b0
0
0 )2
(xm x0M )2 (ym yM
xm x0M ym yM
a20
b20
a30
b30
2m1
x = xM yM a b
41
w =
m1
(x1 x0M )2
a20
(x2 x0M )2
a20
(xm x0M )2
a20
0 )2
(y1 yM
b20
0 )2
(y2 yM
b20
T
1
T
..
0 )2
(ym yM
+
1
2
b
+
1
L(e, x, ) = eT e + T (B T e + Ax + w) min
e,x,
2
L
= e + B
=0
(
e, x
, )
e
L
= AT
=0
(
e, x
, )
x
L
= B T e + A
(
e, x
, )
x = w
59
2m2m
0
42m
BT
2m4
2mm
AT
44
4m
m2m
m4
mm
3m+43m+4
e
0
x
=
0
3m+41
3m+41
= B T B
+ A
e = B
x = w
B T B A
w
=
=
0
AT 0
x
= (B T B)1 (A
=
x + w)
= AT (B T B)1 A
x + AT (B T B)1 w = 0
= x
= (AT (B T B)1 A)1 AT (B T B)1 w
= e = B(B T B)1 (A
x + w)
f (x0 , y) +
g(x0 ) +
f
x x0 ,y
g
x x
f (x, y) = 0
=
g(x) = 0
x +
f
y x ,y
0
x + HOT = 0
Ax B T + w = 0
A x + wf = 0
1 fT
2 e W e min
f (x0 + x, y e) = 0
g(x0 + x) = 0
(e) + HOT = 0
=
w + Ax B T e = 0
wf + Af x = 0
1
Lf (x, e, ) = eT W e+T (AxB T e+w)+f T (Af x+wf )
2
Lf
T A + f T Af = 0 = AT
+ Af T
f = 0
= T A + f T Af =
x
Lf
T B T = eT W
T B T = 0 = W e B
=0
= eT W
e
60
Lf
T
x B T e + wf = 0
= (Af x B T e + wf ) = Af
f
W
B T
0
0
B 0
0
0
0
0
AT Af T
e
0
0
A w
f = wf
Af
0
0
B T W 1 row 1 + row 2 :
W
B
0
0 B T W 1 B 0
0
0
0
T
0
A
Af T
e
0
0
w
A
f = wf
Af
0
0
B T W 1 B 0 A
w
f = wf
0
0 Af
T
T
0
A
Af 0
x
AT (B T W 1 B)1 row 1 + row 3 :
B T W 1 B 0
A
w
f =
0
0
Af
wf
=
T
T
T
1
1
T
T
1
1
A (B W B) w
0
Af A (B W B) A
x
second reduction
f
0
Af
wf
=
=
AT (B T W 1 B)1 w
Af T AT (B T W 1 B)1 A
x
or
=
AT (B T W 1 B)1 A Af T
Af
0
AT (B T W 1 B)1 w
wf
61
N11 N12
N21 N22
Q11 Q12
Q21 Q22
I 0
0I
Q22
Q12
=
Q
21
Q11
1
= (N22 N21 N11
N12 )1
1
= N11 N12 Q22
1
= Q22 N21 N11
1
1
1
= N11 + N11 N12 Q22 N21 N11
N11 = AT (B T W 1 B)1 A = AT M 1 A
N12 = Af T
N21 = N12 T = Af
N22 = 0
Q22 = (0 Af (AT M 1 A)1 Af T )1 = (Af (AT M 1 A)1 Af T )1
Q12 = (AT M 1 A)1 Af T (Af (AT M 1 A)1 Af T )1
Q21 = Q12 T
Q11 = (AT M 1 A)1 (I Af T (Af (AT M 1 A)1 Af T )1 Af )
x = Q11 AT M 1 w Q12 wf
x =
xwithout +
x
T
1
f = Q21 A M w Q22 wf
+ A
w = M
x
= M 1 (A
=
x + w)
= W 1 BM 1 (A
e = W 1 B
xw)
Case 2: AT M 1 A is a rank deficient matrix
rank(AT M 1 A) = rank A = n d
62
N Af T
Af 0
1
R ST
S Q
1
N R + Af T S = I
(5.1)
N S T + Af T Q = 0
(5.2)
Af R = 0
(5.3)
Af S T = I
(5.4)
HN = 0
H : d n therefore
N is symmetric
H (5.1)
= H |{z}
N R +HAf T S = H = S = (HAf T )1 H
0
H (5.2)
= H N
S T} +HAf T Q = 0 = HAf T Q = 0
| {z
0
(5.1) = N R + Af T (HAf T )1 H = I
(5.3) = Af R = 0 = Af T Af R = 0
(+)
(N + Af T Af )R = I Af T (HAf T )1 H
= R = (N + Af T Af )1 (I Af T (HAf T )1 H)
x = RAT N 1 w + S T wf
= (N + Af T Af )1 AT M 1 w
+ (N + Af T Af )1 Af T (HAf T )1 HAT M 1 w S T wf
|
{z
}
=0
= (N + Af Af )
A M 1 w H T (Af H T )1 wf
T
63
x = (N + Af T Af )1 AT M 1 w
= M 1 (A
x
x + w)
= M 1 ((N + Af T Af )1 AT M 1 w
= M 1 ((N + Af T Af )1 AT M 1 I)w
e = W 1 B
= W 1 BM 1 ((N + Af T Af )1 AT M 1 I)w
64
6 Statistics
6.1 Expectation of sum of squared residuals
n
o
T 1
E e Qy e
NB.: eT Q1
oe, die minimiert wird
y e ist die Gr
T
Q1
y
e =
1m mm m1
m X
m
X
(Py )ij ei ej
i=1 j=1
m X
m
o
n
X
T 1
(Py )ij E ei ej
= E e Qy e =
i=1 j=1
m X
m
X
i=1 j=1
m
m X
X
i=1 j=1
[Py Qe]ii
= trace(Py Qe)
= trace(Py (Qy Qy))
= trace(Im Py Qy)
= m trace Py Qy
trace Py Qy = trace QyPy
= trace AQx AT Py
= trace A(AT Py A)1 AT Py
= trace PA
Projektor
65
6 Statistics
Aus Linearer Algebra:
trace X = Summe der Eigenwerte von X
Frage: Eigenwerte des Projektors?
PA z = z
PA PA z = PA z = z
PA PA z = PA z = 2 z
Eigenwertproblem
z = z = ( 1)z = 0 = =
66
0
1
6.2 Basics
6.2 Basics
Zufallsvariable: x
Realisierung: x
Probability density function
probability density function (pdf)
f (x)
Wahrscheinlichkeitsdichte
f (x)
f (x)
x0
ab
(c) intervall
Figure 6.1
Z
f (x) dx = 1
E {x} = x =
xf (x) dx
D {x} = x2 =
(x x )2 f (x) dx = E (x x )2
P (x < x0 ) =
Zx0
f (x) dx
67
6 Statistics
Cumulative density function
Verteilungsfunktion
F(x)
F (x) =
Zx
f (y) dy = P (x < x)
z.B.:
P (a x b) =
Zb
Zb
f (x) dx
f (x) dx
Za
f (x) dx
= F (b) F (a)
6.3 Hypotheses
Annahme/Aussage, die statistisch getestet werden kann.
H : x f (x)
Annahme: x sei verteilt nach vorgegebem f (x).
P (a x b) = 1
= Sicherheitswahrscheinlichkeit = Konfidenzniveau
68
6.3 Hypotheses
1
/2
/2
a
K
reject
x
A
accept
b
K
reject
P ( x + ) = 68.3% = = 0.317
P ( 2 x + 2) = 95.5% = = 0.045
P ( 3 x + 3) = 99.7% = = 0.003
Matlab: normpdf
1 = F (b) F () = F ( + k) F ( k)
k = kritischer Wert (critical value)
69
6 Statistics
festlegen a, b bestimmen
H0 : x f0 (x)
: Null-Hypothese
Ha : x fa (x)
: alternative Hypothese
Ha
H0
0
A
a
k
accept
K
reject
xK
= H0 verwerfen
x 6 K
= H0 annehmen
H0 wahr
Fehlschlu 1. Art
(falscher Alarm)
P (x K|H0 ) =
OK
70
H0 falsch
OK
Fehlschlu 2. Art
(unterlassener Alarm)
P (x 6 K|Ha ) =
6.4 Distributions
6.4 Distributions
Standard normal distribution (univariate)
x N (0, 1)
f (x) =
1 2
1
e 2 x
sqrt2
E {x} = 0
D {x} = E x2 = 1 x2 2 (1, 0)
E x2 = 1
N(
, 1)
f (x) =
1 T
exp
x
x
k
2
(2) 2
kV ector
x
kV ektor
kV ektor
n
o
E xxT = I
D {x} = E x2 = 1 x2 2 (1, 0)
x N (0, Qx ),
Qx =
12
22
0
..
.
k2
1
1 x2i
xi
f (xi ) =
exp 2
2 i
2i
x
y i = i N (0, 1)
i
n
o
x2k
x21
x22
2
T 1
xT Q1
x x = 2 + 2 + . . . + 2 (k, 0) = E x Qx x = k
1
2
k
N (0, i2 )
71
6 Statistics
f (x) =
1
1
(det Qx ) 2 exp ( xT Q1
x x)
2
(2)
k
2
T = Nichtzentralitatsparameter
= 21 + 22 + . . . + 2k
General case
x N (, Qx )
1
1
1
f (x) =
exp
(x
)Q
(x
)
x
k
1
2
(2) 2 (det Qx ) 2
E {x} =
D {x} = Qx
n
o
E xT Q1
x x = k + ,
72
= T Q1
x
7 Statistical Testing
7.1 Global model test: a first approach
Statistics of estimated residuals
e = y y
= PA y
E {
e} = 0,
e N (0, Qe)
D {
e} = Qe
= Qy Qy
= PA Qy (PA )
2 (m, 0) and thus E eT Q1
= m?
Question: eT Q1
e e
e e
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 = eT Q1
2 (m n, 0)
y e
E {T } = m n
Thus, we have a test statistic and we know its distribution. This is the starting point
for global model testing.
73
7 Statistical Testing
Figure 7.1: Distribution of the test statistic T under the null and alternative hypotheses.
(Non-centrality parameter to be explained later)
T > k: reject H0
In case T the realization of T is larger than a chosen critical value (based on ),
the null hypothesis H0 should be rejected. At this point, we havent formulated an
alternative hypothesis Ha yet. The rejection may be due to:
error in the (deterministic) observation model A,
measurement error: E {e} 6= 0,
wrong assumptions in the stochastic model: D {e} 6 Qy .
Variance of unit weigth
A possible error in the stochastic model would be a wrong scale factor. Let us write
Qy = 2 Q and see how an unknown variance factor 2 propagates through the various
estimators:
1 T 1
x
= (AT Q1
y A) A Qy y
1
Qx = (AT Q1
y A)
Qy = 2 Q
y Py = Q1 = 2 Q1
y
x
= (AT 2 Q1 A)1 AT 2 Q1 y
= 2 (AT Q1 A)1 AT 2 Q1 y
74
= unabhangig von 2
=
E
Q
e
=mn
E eT Q1
y
n
o
= E eT Q1 e = 2 (m n)
Alternative test statistic
This leads to a new test statistic:
2 =
2
eT Q1 e
= E
= 2 ,
mn
unverzerrter Sch
atzer
y =
If we consider Q as the a priori variance-covariance matrix, then Q
2 is the a
posteriori one.
Now consider the ratio between a posteriori and a priori variance as an alternative test
statistic:
T
2 (m n, 0)
2
eT 2 Q1 e e Q1
y e
=
=
= F (m n, , 0)
2
mn
mn
mn
E
=1
2
75
7 Statistical Testing
H0
E y = Ax;
Ha
D y = Qy
x
0
y0 = A
x0
e0
T
1
e0 Qy e0 2 (m n)
E y = Ax + C; D y = Qy
x
= AC
x
a ,
ya = A
xa + C
ea
T
1
ea Qy ea 2 (m n q)
How is it distributed?
H0 : T 2 (q, 0)
76
and Ha : T 2 (q, )
ea
R(C)
e0
R(A|C)
ya
ea
e0
y0 ya
y0
y0
ya
y0 ya
R(A)
(a)
(b)
R(C)
C
R(A
R(A|C)
PAC
ya
PA C
y0-Axa
Axa
(c)
ya = A
xa + C
+PA C
= A
x + P C
| a {z A }
y0
= ya y0 = PA C
Q1
T T
= T = PA C
y PA C = C
PA Q1
y PA
|
{z
}
1
1
=Q1
y PA =Qy Qe
0 Qy
1
T C T Q1
=
y Qe0 Qy C
= (
y0 ya )T Q1
y0 ya )
y (
77
7 Statistical Testing
1
T C T Q1
=
y Qe0 Qy C
(
ea , ya , )
Jetzt: Version nur mit e0 und C
Normal equations under H0 , Ha
AT Q1
x0 = AT Q1
y A
y y
T
T
x
A
A
1
Ha :
Qy A C
=
Q1
y y
T
C
CT
T 1
AT Q1
T 1
y A A Qy C
a
A Qy y
x
nn
nq
T 1
=
T Q1 y
C Qy A C T Q1
C
C
y
y
H0
qn
1. Zeile:
T 1
AT Q1
xa + AT Q1
x0
y A
y C = A Qy A
1 T 1
= x
a = x
0 (AT Q1
y A) A Qy C
= A
xa = A
x0 PA C
= A
= A
xa + C
x0 + (I PA )C
= ya = y0 + PA C
0
y Qe0 Qy C) C Qy e
T...
4. Form
Einsetzen in
1
1
1 T 1
T = eT
0
0 Qy C() ()() C Qy e
1
T 1
1
1 T 1
= eT
0
0 Qy C(C Qy Qe0 Qy C) C Qy e
78
q1
qm mm m1
1
Qz = C T Q1
y Qe0 Qy C
= Q1
z z = z = Qz
2
T = z T Q1
z z q
H0
Ha
z N (0, Qz )
T 2 (q, 0)
Qz )
z N (Qz ,
2
T (q, )
T
1
1
= Qz Qz Qz = T C T Q1
y Qe0 Qy C
Summary
1 e
1 zeigt m
Testgroe T = eT
T
oglicherweise an, da H0 zugunsten Ha zu
0
a
a Qy e
0 Qy e
verwerfen ist, d. h. das Modell E {y} = Ax ist moglicherweise unpassend.
(2) (
y0 ya )T Q1
y0 ya )
y (
1
(3) T C T Q1
y Qe0 Qy C
1
1
T 1
1
(4) eT
Q
C
C
Q
Q
Q
C
C T Q1
0
e0 y
0 y
y
y e
(5) z T Q1
z z;
z := C T Q1
0 ;
y e
Qz = . . .
wobei bei (1), (2) und (3) die Berechnung der Alternativhypothese involviert ist,
wahrend (4) und (5) nur e0 und C benotigt
Wegen
z N (0, Qz ) unter H0
z N (Qz , Qz ) unter Ha
79
7 Statistical Testing
folgt die Aussage
T 2 (q, 0) unter H0 , T 2 (q, ), = T Qz unter Ha
Wieviel Zusatzparameter konnen minimal/maximal gewahlt werden?
Gesamtzahl aller Parameter x und ist n + q
Anzahl der Beobachtungen m
Losbarkeit gegeben, wenn n + q m = 0 < q m n
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 quadratisch
Redundanz = m n q = 0
ea = 0
ya = y
1
T = eT
0
0 Qy e
80
2 =
mn
T
mn
1
eT
0
0 Qy e
mn
verwendet und
H0 :
2 F (m n, , 0)
Ha :
2 F (m n, , 0), wie zuvor
Meaning of
2
Im Fall, da die VK-Matrix D y = Qy zu 2 Q gegeben ist 2 ist ein unbekannter
Skalierungsfaktor folgt f
ur die quadratische Form eT Q1
y e
n
o
E eT Q1
e
= 2 (m n)
y
2
=E
(m n)
2
= E
= 2
D {y} = Qy = 2 Q = W 1
2 . . . unbekannter Varianzfaktor
Q1 = W . . . gegebene Gewichtsmatrix
1 T 1
x
= (AT Q1
y A) A Qy y
= (AT 2 Q1 A)1 AT 2 Q1 y
= 2 (AT Q1 A)1 AT 2 Q1 y
= (AT Q1 A)1 AT Q1 y
von 2 unabh
angig (egal ob bekannt oder unbekannt)
81
7 Statistical Testing
T 1
Qy = (AQ1
x A )
= 2 A(AT Q1 A)1 AT
! abhangig
e = y y
unabhangig
! abhangig
allgemein
Qe = Qy Qy
h
i
= 2 Q A(AT Q1 A)1 AT
f = F x
Qf = F Qx F T
= 2 F (AT Q1 A)1 F T
Die VK-Matrizen sind von 2 abhangig, d. h. nicht berechenbar, wenn 2 unbekannt
ist. Aber wir haben mit
2 ja eine erwartungstreue Schatzung f
ur 2 , so da anstatt
Qx , Qy, Qe, . . . , Qf die VK-Matrizen
x =
=
Q
2 Qx , . . . , Q
2 Qf
f
82
x widerspiegelt.
in Q
was sich u
ber 2 = e mn
x = 0wird,
b) Parabelanpassung: . . . Das Modell pat perfekt zum Datensatz 2, so da Q
T
1
denn e Q e = 0 wegen e = 0!
Bedeutung des Tests
Wegen ea = 0 mu offensichtlich keine C-Matrix spezifiziert werden, um den Test
durchf
uhren zu konnen. Im Fall q < m n ist das anders ,und es ist nicht immer trivial, ein f
ur alle Situationen geeignetes C zu finden. Der Test kann immer durchgef
uhrt
werden und deshalb als allgemeiner Modelltest bezeichnet werden (overall model test).
Case (ii) q = 1: data snooping
= C ist ein m 1-Vektor, ein Skalar
1
1
T 1
1
T = w2 = eT
Q
C
C
Q
Q
Q
C
C T Q1
0
e0 y
0 y
y
y e
=
1
eT
0 Qy C
2
1
C T Q1
y Qe0 Qy C
83
7 Statistical Testing
2
1
C T Q1
y Qe0 Qy C
2 =
mit
=
1
C T Q1
y e
1
C T Q1
y Qe0 Qy C
Wichtigste Anwendung:
Detektion eines groben Fehlers in den Beobachtungen, der zu einer falschen Modellformulierung f
uhrt.
H0 : E {y} = Ax
Verwerfe H0 , falls T =
> k oder
H : E {y} = Ax + C
a
C = 0, 0, . . . ,
T =
1
, 0, . . . , 0
|{z}
i-te Pos.
1
< k2 bzw.
T =
kann
> k2 (
ei
w= T =
ei
H0 :
T N (0, 1)
Ha : Tq N ( T , 1)
1
mit T = C T Q1
y QeQy C
7.3 DIA-Testprinciple
Detektion: Globaltest Modellfehler (
uberhaupt?)
Identifikation: Data snooping Ausreier (wo?) max(wi )
Anpassung Iteration, Neumodellierung, Neubeobachtung
Question: how to ensure consistent testing parameters? We must make sure that we
would reject the null hypothesis but cannot identify the error during the data snooping.
84
0
0 = 1 0
= ( ) = 0
= = mn
z. B.: 1 = 1% (
ublicherweise klein), 0 = 20% = mn 30%
1
= T C T Q1
y QeQy C
Qe = Qy Qy
1
= Qy A AT Q1
A
AT
y
1
T T
1
1
T 1
T 1
= =
C
Q
Q
A
Q
A
A
Q
C
y
y
y
| {z } y
(C)T
85
interne
keit
Zuverl
assig-
7 Statistical Testing
= Q1y groer
= C kleiner
= Bei genaueren Messungen ist der Modellfehler C,
der (gerade noch) bestimmt werden kann, kleiner
u
ber A:
mehr Beobachtungen = groere Redundanz
Bei gleichbleibendem C = groer
oder: bei gleichem = C kleiner
besserer Netzentwurf, Messanordnung (= Selbstkontrolle)
Minimum Detectable Bias (MDB)
y = C = E {y|Ha } E {y|H0 }
m1
mq q1
Case q = 1 (datasnooping):
ist skalar / yi = ci
1
2
0 = ci T Q1
y QeQy ci
s
0
= |i | =
1
T
ci Qy QeQ1
y ci
Qe ci
|{z}
=Qy Qy
= = y4
y2i y2i
i
86
= |i | = s
y2i
= yi r
|i | = yi r
y2i y2
y2
i
y2
y2i
y2
0
= yi
ri
y2i
b) Wenn yi yi : |i | = yi 0
ri = 1
y2i
y2i
= lokale Redundnaz
ri = m n
Red. weil: ri = ci T I QyQ1
ci
y
= ci T (I PA ) ci
= ci T PA ci
=
X
i
ri = tracePA = m n
n
o
T 1 T
NB.: E e Qy e
= mn
mn
i | = y
r = i=1 =
= |
i
m
m
mn
m
87
7 Statistical Testing
Ausgleichung: Redundanz verteilen abhangig von A, Qy
Redundancy
e = RA
y
1
T 1
T 1
= I A A Qy A
A Qy y
= Ry
R = Redundanzmatrix
ei = Rij yj
= ri yi + ...
=
e = ri yi
= Lokale Redundanz ist ein Ma daf
ur wie Redundanz auf einzelne Messungen umverteilt
wird bzw. wie Modellfehler y auf Residuen projiziert werden.
nicht behandelt: q > 1, z. B. m n
y( = C) x
1 T 1
x
= (AT Q1
y A) A Qy y
1 T 1
(
x + x
) = (AT Q1
y A) A Qy (y + y)
1 T 1
x
= (AT Q1
y A) A Qy y
88
x = x
T Q1
x
x
= x
T AT Q1
y A x
= (PA y)T Q1
y (PA y)
= |PA y|2
y = PA y + PA y
oder = PA x
+ PA y
2
|y|2 = |PA y|2 + PA y
T AT Q1
+ y T (PA ) Q1
y
oder y T Q1
y y = x
y A x
y PA |{z}
| {z }
|
{z
} |
{z
}
y
x + 0
y =
1
1
Q1
y Qe
Qy =Qy PA
{z
0
special case
q = 1, ci , Qy = diag
yi = 0 ,
ri = 1
y2i
y2
=
x = yi 0
1
= 0 0
ri
1 ri
=
0
ri
89
7 Statistical Testing
=
=
=
90
y2i y2i
y2i (y2i y2i )
y2i
y2i y2i
1
y2i
y2
i
91
8 Recursive estimation
8.1 Partitioned model
E
y1
y2
(m1 +m2 )1
A1
A2
x;
(m1 +m2 )n
y1
y2
Q1 0
0 Q2
1
T 1
T 1
T 1
A
y
+
A
Q
y
= x
= A1 T Q1
A
+
A
Q
A
Q
1
2
2
1
1
2
2
1
2
1
2
x
(2)
1
T 1
Qx = A1 T Q1
A
+
A
Q
A
1
2
2
1
2
n o
n o
E y 1 = A1 x
D y 1 = Q1
(
1 T 1
x
(1) = A1 T Q1
A1 Q1 y
1 A1
1
=
T 1
Qx(1) = A1 Q1 A1
1
1
T 1
T 1
x
(2) = Q1
+
A
Q
A
Q
x
+
A
Q
y
2
2
2
2
2
x
x
(1) (1)
2
(1)
1
Q
1
T 1
x
(2) = Qx
(1) + A2 Q2 A2
92
measurement update
covariance update
= Aufdatierungsgleichungen
x
(1)
y2
I
A2
x;
x
(1)
y2
Qx(1) 0
0 Q2
8.1.3 Umformen
mit
1
T 1
Q1
x
(2) = Qx
(1) + A2 Q2 A2
1
T 1
= Q1
x
(1) = Qx
(2) A2 Q2 A2
Einsetzen:
T 1
T 1
x
(2) = Qx(2) Q1
x
A
Q
A
x
+
A
Q
y
2
2 (1)
2
2
2
2
x
(2) (1)
y2 A2 x
(1)
=x
(1) + Qx2 A2 T Q1
2
=x
(1) + KV2
V2 = y2 A2 x
2
K = Qx2 A2 T Q1
2
A2 x
(1) . . . predicted observation
V2 . . . prdicted residual
K . . . gain matrix
Problem: viele Matrixinversionen:
Qx(1) ,
Qx1 + A2 T Q1
A
2 ,
2
nn
nn
Q2
m2 m2
93
8 Recursive estimation
= A2 I
I
A2
=0
Q
0
x
(1)
x
(1)
x
(1)
A2 I E
= 0;
D
= 0
y2
y2
Q2
1
B T E y = 0
T
T
y = I Qy B B QyB
B y
D y = Qy
1
Qx(1) A2 T
x
(2)
I 0
x
1
T
A2 I
=
=
Q2 + A2 Qx(1) A2
0I
y2
y2
Q2
1
= x
(2) = x
(1) + Qx(1) A2 T Q2 + A2 Qx(1) A2 T
(y2 A2 x
1 )
=x
(1) + KV2
K = Qx(1) A2 T
Q2 + A2 Qx(1) A2 T
m2 m2
1
A2 Qx(1)
= (I KA2 ) Qx(1)
Verstarkungsmatrix numerisch identisch. Nur 1 Inversion m2 m2. Zum Beispiel nach
jeder Neubeobachtung m2 = 1.
8.2 Allgemeiner
Batch:
y1
A1
y2 A2
E . = . x;
.
.
.
.
yk
Ak
x
=
k
X
i=1
94
y1
Q1
y2
Q2
D . =
..
yk
Ai T Q1i Ai
!1
k
X
i=1
Ai T Q1i yi
0
..
.
Qk
8.2 Allgemeiner
Rekursiv:
x
k = x
k1 + Kk zk
zk = yk Ak x
k1
1
T
T
Kk = Qx(k1) Ak Qk + Ak Qk Ak
Qx(k) = [I Kk Ak ] Qxk1
95
A Partitioning
A.1 Inverse Partitioning Method
I b
bT 0
1. A bC = I
A B
I 0
=
C D
0I
| {z }
inverse
A, B, C, D are unknown
2. B bD = 0
3. bT A = 0
4. bT B = I
T
T
T
T 1 T
1. b
| {z A} b bC = b = C = (b b) b
=0
2. bT B + bT bD = 0 = I + bT bD = 0 = D = (bT b)1
1. A + b(bT b)1 bT = I = A = I b(bT b)1 bT
2. B + b(bT b)1 = 0 = B = b(bT b)1
I b
bT 0
1
e = b(bT b)1 bT y
= (bT b)1 bT y
96
A.2 Partitioning
A.2 Partitioning
The normal matrix of the linear system is symmetric, therefore
T
A AD
I 0
R ST
=
S Q
0I
DT 0
then
(AT A)R + DS = I
T
(A.1)
(A A)S + DQ = 0
(A.2)
DT R = 0
(A.3)
DT S T = I
(A.4)
HDQ = 0 = Q = 0
x
= RAT y = (AT A + DDT )1 AT y (AT A + DDT )1 D(HD)1 HT
| {z } y
= SAT y = (HD)1 HT T y = 0
| {z }
=0
=
= x
= (AT A + DDT )1 AT y
e = y A
x = (I A(AT A + DDT )1 AT )y
97
B Buchempfehlungen
B.1 Wissenschaftliche B
ucher
Teunissen, P. J. G.
Adjustment theory an introduction
Delft University Press, 2003
ISBN 90-407-1974-8
Teunissen, P. J. G.
Testing theory an introduction
Delft University Press, 20002006
ISBN 90-407-1975-6
Teunissen, P. J. G.
Dynamic data processing recursive least-squares
Delft University Press, 2001
ISBN 90-407-1976-4
Niemeier, Wolfgang
Ausgleichungsrechnung
de Gruyter, 2002
ISBN 3-11-014080-2
Grafarend, Erik W.
Linear and Nonlinear Models Fixed Effects, Random Effects, and
Mixed Models
de Gruyter, 2006
ISBN 978-3-11-016216-5
Koch, Karl-Rudolf
Parametersch
atzung und Hypothesentests in linearen Modellen
D
ummlers
ISBN 3-427-78923-3
98
B.2 Popularwissenschaftliche B
ucher, Literatur
B.2 Popul
arwissenschaftliche B
ucher, Literatur
Sobel, Dava
Longitude: The True Story of a Lone Genius Who Solved the Greatest
Scientific Problem of His Time
Fourth Estate, 1996
ISBN 1-85702-502-4
Deutsche Ubersetzung:
L
angengrad. Die wahre Geschichte eines einsamen Genies, welches das
gr
ote wissenschaftliche Problem seiner Zeit l
oste
Berliner Taschenbuch Verlag, 2003
ISBN 3-8333-0271-2
Kehlmann, Daniel
Die Vermessung der Welt
Rowohlt, Reinbek, 2005
ISBN 3-498-03528-2
99
Index
Legendre, 13
minimization problem, 15
optimal estimates, 15
unknowns, 13
100