Beruflich Dokumente
Kultur Dokumente
Abstract. Motivated by the existing theory of the nature of the problem has been clari®ed in two
geometric characteristics of linear generalized inverses important papers by Grafarend and Scharin (1974,
of linear mappings, an attempt is made to establish a 1976) while the relation among various solutions, as well
corresponding mathematical theory for nonlinear gen- as Meissl's ``inner'' solution have been established with
eralized inverses of nonlinear mappings in ®nite- the introduction of the famous Baarda S-transformation
dimensional spaces. The theory relies on the concept (Baarda 1973; see also Mierlo 1980; Koch 1982). It
of ®berings consisting of disjoint manifolds (®bers) in might therefore seem that every aspect of this problem
which the domain and range spaces of the mappings are has been fully investigated and well understood for quite
partitioned. Fiberings replace the quotient spaces gen- some time now, though some recent work (Xu 1995)
erated by some characteristic subspaces in the linear might point to the contrary. However, with the excep-
case. In addition to the simple generalized inverse, the tion of some simple cases, such as that of levelling in a
minimum-distance and the x0 -nearest generalized inverse small area, the mathematical treatment of the problem is
are introduced and characterized, in analogy with the con®ned to its linearized version, although it is well
least-squares and the minimum-norm generalized in- understood as a nonlinear problem. Nonlinear adjust-
verses of the linear case. The theory is specialized to the ment has been extensively studied in the geodetic
geodetic mapping from network coordinates to observ- literature, especially from the viewpoint of nonlinear
ables and the nonlinear transformations (Baarda's S- least squares, see, e.g., Teunissen (1985, 1989a, 1989b,
transformations) between dierent solutions are de®ned 1990), Grafarend and Scharin (1991) and Lohse
with the help of transformation parameters obtained (1994), where further references can be found. A
from the solution of nonlinear equations. In particular, completely dierent approach is that of Dermanis and
the transformations from any solution to an x0 -nearest Sanso (1995), where optimal nonlinear estimators have
solution (corresponding to Meissl's inner solution) are been investigated from a strictly probabilistic point of
given for two- and three-dimensional networks for both view. In both approaches the datum problem is solved in
the similarity and the rigid transformation case. Finally an implicit way. In the nonlinear least-squares adjust-
the nonlinear theory is specialized to the linear case with ment solved by an iteration scheme, the choice of datum
the help of the singular-value decomposition and depends on the choice of the initial parameter values
algebraic expressions with speci®c geometric meaning used for starting the iteration process, as well as on the
are given for all possible types of generalized inverses. principle used for ``improving'' these parameters. In the
nonlinear estimation case, the need to adopt a Bayesian
point of view in order to obtain a meaningful estimation
Key words. Generalised inverse á Nonlinearity á independent of the unknown ``true'' parameter values
Pseudoinverse á Inverse problems á Datum problem also solves the datum problem in an implicit way, where
the datum choice is hidden in the choice of the prior
probability distribution of the parameters.
The solution to the nonlinear datum problem pre-
1 Introduction sented here is based on the concept of the S-transfor-
mation and has the form of such a similarity (or rigid)
The datum problem or zero-order design problem transformation with parameters which come from the
(Grafarend 1974) arising in the adjustment of observa- solution of a system of nonlinear equations.
tions related to geodetic networks has received consid- Another important aspect of the geodetic datum
erable attention since the pioneering work of Meissl problem in its linear form is its relation to generalized
(1965, 1969) and its popularization by Blaha (1971). The inverses of matrices (linear operators) which led Bjer-
72
hammar (1951) to an independent introduction of the out full rank in the linear case), the determination of
Moore-Penrose generalized inverse, later than Moore parameter values is not trivial anymore, because there is
(1920) but before Penrose (1955). The question arises an in®nite set of parameter values which f maps on the
whether the nonlinear version of the geodetic datum same manifold point corresponding to the adjusted ob-
problem bears a similar relation to some type of gen- servations. The datum problem is exactly the problem of
eralized inverses of nonlinear operators. An attempt will choosing one out of all possible parameter sets.
be made to look mainly into the geometric aspects of Of course rank de®ciency in an observational model
such nonlinear generalized inverses, although the is not exclusively related to the datum problem, as
building of a concrete mathematical theory requires a demonstrated, e.g., in Dermanis and Grafarend (1981).
more rigorous treatment which is beyond our present Additional rank de®ciency may result from the inability
scope. We shall base our investigation on the represen- of the available observations to recover the shape (or
tation theory of various types of linear generalized in- shape and size) of a geodetic network. Such cases, al-
verses which has been introduced by Takos (1976) from though partly covered by the treatment in Sect. 3, need
an algebraic point of view and especially by Teunissen individual treatment and cannot be part of the general
(1985) from a geometric point of view. approach taken here, where our main subject is the
The datum problem is always a part of the problem of common datum defect resulting only from the use of
the adjustment of redundant observations which are re- coordinates as parameters.
lated to a set of parameters (coordinates in the geodetic We must point out that the above point of view is not
case) which in fact cannot be determined from the coordinate-free, since it depends on the choice of a
available observations. The reason is that the informa- speci®c set of m unknown parameters. This is perhaps of
tion contained in the angle and distance (angle) obser- little concern to geodesy where the choice of coordinates
vations relates only to the shape and size (shape) of the as parameters imposes itself as a matter of convenience.
network, while coordinates relate in addition to its posi- It is possible, however, to establish a theory where both
tion (position and size) with respect to a certain reference the adjustment and the datum problem are treated in a
frame. The problem of placing the network in relation to coordinate-free way in the spirit of modern dierential
a given reference frame can be also seen from the inverse geometry.
point of view of placing a reference frame in relation to a A more in-depth introduction to the (nonlinear) da-
given (i.e., physically existing) network. This choice of tum problem, especially in relation to the modelling
reference frame (or datum in geodetic terminology) poses problem, can be found in Dermanis (1991).
a problem, the datum problem, which must be solved in an
arbitrary but consistent way based in the introduction of
additional information not contained in the observations. 2 Geometric characteristics of generalized
The usual approach to the description of the adjust- inverses of linear operators
ment and choice of datum problem is to consider the n
observed parameters (n > r) as the coordinates of a A natural point of departure for the study of the
linear space Y , called the observation space, in which the nonlinear datum problem is our knowledge of the
r-dimensional manifold M modeling the physical sys- simpler linear case. For this reason we shall review some
tem, called the model manifold, is lying. (We con®ne of the geometric characteristics of the theory of
ourselves to the case of discrete and ®nite observations.) generalized inverses of linear mappings and point out
The model manifold is described as the range of a those which have proven to be more appropriate for
nonlinear operator f from an m-dimensional parameter generalization to the nonlinear case. Our exposition will
space X (m > r) into the observation space Y . The be more casual than in the rest of this work, since a
mapping f is established by the mathematical equations rigorous and thorough study exists already in the
which relate all observables to the unknown parameters. geodetic literature (Teunissen 1985).
In a ``normal'' situation, which is almost never the case A linear mapping f : X ! Y is characterized geo-
in geodesy, the number of parameters and the dimension metrically by its range
of the model manifold are equal (m r), in which case
the restriction fjM of f to M has an inverse which may R
f fy 2 Y jy f
x for some x 2 X g
serve as a coordinate mapping from M to X Rm . In
other words the chosen parameters can serve as a par- which is a linear subspace of Y and its null space
ticular system of coordinates on the model manifold and N
f fx 2 X jf
x 0g
the only problem to be solved is the adjustment prob-
lem. In the linear case the rank of f is r m and the which is a linear subspace of X . To any element y f
x
corresponding model is called a full-rank model. of R
f corresponds an ane subspace x N
f of X
The unavoidable observational errors are added to with elements which f maps to the same element
the true values of the observables (which correspond to y f
x. The parallel translates of N
f are thus the
a point on the model manifold) and give observations ``solution spaces'' of f which correspond one-to-one to
corresponding to a known point outside the manifold. the elements of R
f .
The adjustment problem can be simply de®ned as the If X is m-dimensional, Y is n-dimensional and rank
problem of ®nding an optimal way to ``return to the
f dim R
f r then d m ÿ r is the injectivity de-
model manifold''. In the case where m > r (model with- fect of f , while f n ÿ r is its surjectivity defect (usually
73
closest element to y from R
f . The projector p is the f
x1 f
x2 . To every element y 2 R
f in the range
orthogonal projector on R
f . R
f of f corresponds a ®ber Fy 2 F de®ned by
A minimum-norm generalized inverse is one for
which S N
f ? , in which case g
y is the element of Fy fx 2 X jf
x yg
1
the solution ane subspace corresponding to y^ p
y The function f gives rise to a bijection
which has the smallest norm. The projector q is the or-
thogonal projector on N
f ? . f: R
f ! F : y ! Fy
From all the preceding geometric characteristics of a
generalized inverse g : Y ! X of a linear mapping Obviously for every x 2 X ; x 2 Ff
x .
f : X ! Y , we summarize those which will be helpful in
dealing with nonlinear mappings. De®nition. The mapping p : X ! F : x ! Ff
x is called
The space Y is ``sliced'' by p f g into ane sub- the projection mapping from X to the ®bering F. It holds
spaces (p-slices) which are parallel to a speci®c (linear) that p f f .
subspace C. Each slice corresponds to a particular ele- When more than one ®berings are involved we denote
ment y^ of R
f Y , since p f g maps all the ele- p by pF . Thus pF
x is the unique ®ber from the ®bering
ments of the slice onto that y^. F which passes through a given point x.
The mapping f ``slices'' the space X into ane sub-
spaces (f -slices) which are parallel to the subspace N
f . De®nition. A section of a ®bering F of X is the range
Each slice corresponds to a particular element y^ of R
f S R
s of a mapping s : F ! X such that p s idF ,
since f maps all the elements of the slice on that y^, we where idF is the identity mapping in F.
may therefore call this slice ``solution space of y^''.
The generalized inverse g maps R
f to a subset S of A section S of a ®bering F intersects any one of its
R
g which has the property that it intersects each so- ®bers in a single element: for any x 2 S; S \ pF
x fxg.
lution space (f -slice) in a single element.
The mapping q g f slices X into q-slices, each De®nition. Let F and K be two ®berings of the same set
slice corresponding to a particular element x^ of S, since q A. K is called a re®nement of F if every ®ber Kt of K is
maps all the elements of the slice on that x^. The q-slices a subset of some ®ber Fs of F : 8 Kt 2 K : 9 Fs 2 F such
are identical to the f -slices. that Kt Fs .
Finally g ``slices'' Y into ane subspaces which are If K is a re®nement of F then pK
z pF
z for
parallel to the subspace N
g. These g-slices have the any z.
property that each is contained in one of the p-slices.
Thus each p-slice is itself further sliced in a more ``re- De®nition. Two ®berings F and H of the same space X
®ned'' set of g-slices. Only in the case of a re¯exive are called complementary if every ®ber of F is a section
generalized inverse are the p-slices and g-slices of H, and vice versa: pF
x \ pH
x fxg; 8 x 2 X .
identical.
These slices, all of which are ane subspaces, De®nition. A ®bering F of X induces a ®bering
can be described by a single ``generating'' linear sub- F0 FjM on a manifold M X with ®bers pF0
x
space (f-slices by N
f , g-slices, and q-slices by N
g, p- M \ pF
x for every x 2 M. The ®bering F0 FjM is
slices by C). Of course, this possibility cannot survive in called the restriction of the ®bering F on M.
the nonlinear case.
Example 1. We shall use the simplest possible geodetic
3 Generalized inverses of nonlinear operators network in order to illustrate some of the results.
Following Grafarend and Scharin (1974) we consider a
Let X ; Y be ®nite-dimensional spaces and f : X ! Y a three-point horizontal network P1 P2 P3 with coordinates
(nonlinear) mapping from X to Y . The question arises
whether it is possible to de®ne, under some appropriate x x1 y1 x2 y2 x3 y3 T
conditions, a class of mappings g : Y ! X which can where all sides a P2 P3 ; b P1 P3 ; c P1 P2 , and all
serve as generalized inverses of the mapping f . The angles A; B; C at points P1 ; P2 ; P3 , respectively, have been
similar theory developed for linear mappings may serve observed. The observation vector
as a guide to a certain extend, although the nonlinearity
of f poses problems which do not allow for a direct y a b c A B CT
extension of the existing theory.
is a function y f
x of the coordinates, explicitly given
De®nition. A partition of a set A into disjoint sets At , by
A [t At it is called a ®bering of A, and its elements At q
are called ®bers. Such a partition can arise from an a
x3 ÿ x2 2
y3 ÿ y2 2
equivalence relation on A, each ®ber consisting of q
equivalent elements of A (Loomis and Sternberg 1980.) b
x3 ÿ x1 2
y3 ÿ y1 2
q
A mapping f de®ned on X gives rise to a ®bering F
of X through the equivalence relation x1 x2 if c
x2 ÿ x1 2
y2 ÿ y1 2
76
mapping /d
/
d; . The inverse mapping /ÿ1 is in r rg min
n; m
5
fact a coordinate mapping on X where the coordinates
d; p /ÿ1
x are ``adapted'' to the ®bering: each ®ber Once S is given, g is de®ned on R
f Y : The question
corresponds to a ®xed value of d, while the ``free'' is how this particular g is extended outside R
f . A step
parameters p may serve as (intrinsic) coordinates on the in this direction is to note that as a consequence of (G1)
®ber. both q g f and p f g are idempotent mappings
We assume for the moment that f is de®ned on the
q2 q; p2 p
whole of X , it is continuous, and has a dierential f g f f )
g f
g f g f and
mapping dfx at each x which is a linear mapping from
the tangent vector space Tx at x to the tangent vector
f g
f g f g
6
space Tf
x at f
x (Choquet-Bruhat et al. 1977, p. 121).
which can be considered as ``nonlinear projections''
The mapping dfx is de®ned by
q g f : X ! R
g f R
g X
dfx : Tx ! Tf
x : v ! u
2 p f g : Y ! R
f g R
f Y
7
u
f / v
/ for every f : X ! R
We have set R
g f R
g, which is obvious, and
(where the tangent vectors v and u are visualized as R
f g R
f according to the previous lemma.
``directional derivatives'' acting on functions). Elements belonging to the range of idempotent
We further assume that dfx has constant rank over X , mappings are invariant under the mapping. Idempotent
rank
dfx r min
n; m; where m dim X and mappings induce also a ®bering of the space they act on.
n dim Y . Consequently R
f is an r-dimensional sub- We denote by P, Q the ®berings induced by p and q,
manifold of Y . respectively, with corresponding elements Py ;y 2 R
p
We now come to the possibility of de®ning a gener- R
f and Qx ;x 2 R
q R
g.
alized inverse g of the mapping f by a similar way as in The mapping g gives rise to a ®bering G of Y with
the linear case. elements Gx fy 2 Y jg
y xg. Obviously for every
element y of Y it holds that y 2 Gg
y pG
y. If
De®nition. A mapping g : Y ! X is called a generalized
y 2 R
f then y 2 Gx pG
f
x, where fxg Fy \ S:
inverse of a given mapping f : X ! Y when
f gf f
3
G1 Lemma. For the ®berings of the space Y it holds that the
®bering G induced by g is a re®nement of the ®bering P
[We denote Eq. (3) by (G1).] This means that for every induced by p f g.
x 2 X ;
f g f
x f
x, so that for every y f
x
2 R
f it holds that
f g
y f
g
y y. As a Proof. For any z 2 Y ; pG
z 2 G and pP
z 2 P. For
consequence g
y 2 Fy , i.e., g must map every element every y 2 pG
z; g
y g
z ) p
y f
g
y f
g
z
of R
f into an element of its corresponding ®ber. In p
z and y 2 pP
z: Consequently pG
z pP
z for any
other words g maps R
f onto a section S of the ®bering z 2 Y and thus G is a re®nement of P. (
F where S R
g. The restriction of g on R
f is a
bijection between R
f and S which are both r-dimen-
sional manifolds.
The relation (G1) satis®ed by the generalized inverse
g implies a corresponding relation between the dier-
entials dfx and dgf
x of f and g respectively:
dfx dgf
x dfx dfx
4
which follows by implementing the implicit function
theorem (Choquet-Bruhat et al. 1977, p. 91).
Proof. y 2 R
f ) 9 x 2 X:
y f
x
f g f
x
f gf
x
f g
y 2 R
f g ) R
f R
f g
which combined with the obvious relation
R
f g R
f implies that in fact R
f g R
f : (
Assuming that g has a constant rank rg rank
dgy Fig. 3. The geometry of the generalized inverse g of a nonlinear
for every y, application of the property of the rank of a mapping f . All elements of pG
y are mapped into the same element
linear map composition r
AB minr
A; r
B to g
y. The elements of pP
y are mapped into elements of pF
x and
Eq. (3), gives r min
r; rg rg so that thus projected by p on the same y^
78
Example 1 (continued). We claim that the mapping g ®xed y 2 R
f must simultaneously satisfy the condi-
de®ned by x g
y, explicitly tions for x 2 Fy
f
g
z z or f
g
f
x
f g f
x f
x and Since y0 2 R
f is ®xed, the image g
pP0
y0 consists
since x has been arbitrarily chosen f g f f and g of elements x having x1 ; x2 ; x3 constant, while y1 ; y2 ; y3
is indeed a generalized inverse of f . ( depend on a single parameter a. It has therefore di-
mension 1. Since x0 g
y0 is ®xed, x g
y may be also
Example 1 (continued). We now consider the choice of a expressed in terms of x0 :
section M of G. We have seen that y 2 pG
y0 when
x1 x01 ; y1 y01 aÿa
3
0
up one element from each one of its ®bers, and this can or
be done by assigning values to both B and C. If in a ÿ a0
addition we want R
f M, we must assign the known x x0 0 0 0 1 1 1T
3
values
a2 c20 ÿ b20 a2 b20 ÿ c20 We can also express g
pP0
y0 by ®ve conditions.
B0 arcos 0 and C0 arcos 0 One possible choice is
2a0 c0 2a0 b0
to B and C, respectively, whenever y0 2 R
f . x1 x2 ; x1 x2 x3 c0 ; y2 ÿ y1 c0
Thus M may be determined from the two conditions x3 ÿ x2 b0 sin A0 ; y3 ÿ y1 b0 cos A0
a2 c 2 ÿ b2 a 2 b 2 ÿ c2 Comparison with the identical conditions for x 2 Fy00
cos B ; cos C
2ac 2ab shows that in fact g
pP0
y0 Fy00 Fy0 \ R
g is an
and it has dimension 6±2=4. M contains the three- element of the ®bering F0 induced by F on R
g.
dimensional R
f , which satis®es these conditions and in Returning once again to the theory, we would like the
addition the similar one for cos A. The independent pa- generalized inverse g of a continuous mapping f to be
rameters a; b; c; A may serve as curvilinear coordinates itself continuous, and this means that the mappings cy
for M which is then described by y y
q, explicitly cannot be independent. Loosely speaking, they must
a a; b b; c c; A A map neighboring elements of Y belonging to neighbor-
ing ®bers of P into neighboring elements of X . Thus cy
a2 c2 ÿ b2 a 2 b 2 ÿ c2 may be considered as the restriction of a continuous
B arcos ; C arcos
2ac 2ab mapping c : R
f Y ! X , de®ned by c
y; z cy
z,
which is continuous in both y and z.
We now consider the intersection M \ pP
y0 pP0
y0 . We can overcome the need to use a family of map-
If we combine the requirements for y 2 M pings cy , one for each y 2 R
f with the introduction of
a any; b any; c any; A any additional ®berings for Y and X and a single mapping
between them. Let M be a section of the ®bering G. In
a2 c2 ÿ b2 a 2 b 2 ÿ c2 order to de®ne g on pP
y for every y 2 R
f it is su-
B arcos ; C arcos
2ac 2ab cient to de®ne it on M \ pP
y since the ®bering G, being
with those for y 2 pP
y0 a re®nement of P, naturally extends g on the whole of
pP
y: for any z 2 pP
y 2= M there exists a unique ®ber
a any; b b0 ; c c0 ; A A0 ; B any; C any pG
z pP
z from G such that pG
z \ M fz0 g
M \ pP
z and g
z g
z0 . Although any section M
we conclude that y 2 M \ pP
y0 pP0
y0 when
of G is appropriate, it is more convenient to choose M so
a any; b b0 ; c c0 ; A A0 that it contains R
f . This is possible in view of the next
lemma.
a2 c20 ÿ b20 a2 b20 ÿ c20
B arcos ; C arcos Lemma. For every y 2 R
f , it holds that pG
y
2ac0 2ab0
\R
f fyg.
This means that pP0
y0 M \ pP
y0 is of dimension 1,
where the parameter a may serve as the single curvili- Proof. Since G is a re®nement of P; pG
y pP
y.
near coordinate. When, in particular, y0 2 R
f then Since R
f is a section of P it holds pG
y \ R
f fyg
b2 c2 ÿa2
A A0 arcos 0 2b00c0 0 . which combined with pG
y pP
y yields the desired
From the de®nition of g we see that when y 2 pP0
y0 result. (
where y0 2 R
f then x g
y cy0
y is given by The ®bering P induces a ®bering P0 PjM on M
c0 ÿ b0 sin A0 a ÿ c0 ÿ b0 cos A0 with ®bers pPjM
z M \ pP
z. R
f is a section of P
x1 y1 and since R
f M by choice, it is also a section of P0 .
3 3
It is possible to consider a ®bering R of M containing
c0 ÿ b0 sin A0 a 2c0 ÿ b0 cos A0
x2 y2 R
f which is complementary to P0 . Thus every z 2 M is
3 3 the unique element pR
z \ pP0
z fzg. The general-
c0 2b0 sin A0 a ÿ c0 2b0 cos A0 ized inverse g must be a mapping whose restriction to M
x3 y3
3 3 is a one-to-one mapping gjM from M to R
g, since
82
Let y 2 Rk \ pP0
y0 : The y must satisfy simultaneously c0 ÿ b0 sin A0 c0 2b0 sin A0
the conditions for y 2 Rk x1 x 2 ; x3
3 3
a0 k ÿ c0 ÿ b0 cos A0
a ÿ k2 b2 c2 ÿ 2 b c cos A y1 ;
3
a0 k 2c0 ÿ b0 cos A0
a2 c 2 ÿ b2 a2 b2 ÿ c2 y2 ;
cos B ; cos C 3
2ac 2ab a0 k ÿ c0 2b0 cos A0
y3
and the conditions for y 2 pP0
y0 3
b b0 ; c c0 ; A A0 But the above components of x determined by
fxg Sk \ Fy00 are in fact identical to those of the image
a2 c20 ÿ b20 a2 b20 ÿ c20 x g
y where y is determined by fyg Rk \ pP 0
y0 .
cos B ; cos C Therefore g is uniquely determined from the mapping
2 a c0 2 a b0
C : Rk ! Sk , which to each ®ber Rk 2 R assigns a ®ber
The combination gives the six conditions Sk 2 S, the correspondence having been established
through the common parameter k. The mapping
a a0 k; b b0 ; c c0 ; A A0 C : R ! S, maps
84
This means that pP
y 0 \ R
f fyg and R
f is indeed
Rk yj
a ÿ k2 b2 c2 ÿ 2bc cos A; a section of P, as required. We can therefore describe
the ®bers of P using y0 2 R
f , in which case y 2 pP
y0
a2 c2 ÿ b2 a 2 b 2 ÿ c2 when
cos B ; cos C
2ac 2ab
a any; b any; c c0 ; A A0 ; B B0 ; C any
to
From the preceding conditions it follows that for an
Sk xjx1 x2 ; x1 x2 x3 y2 ÿ y1 ; arbitrary y its projection y0 p
y f
g
y is given by
2 2 2 sin A sin B
x3 ÿ x2
y3 ÿ y2
y1 y2 y3 ÿ k a0 c ; b0 c ; c0 c
sin
A B sin
A B
The other required correspondence between pP0
y0 A0 A; B0 B; C0 p ÿ
A B
and Fy00 is naturally introduced through the common
element y0 used in the de®nitions of the ®bers The next step is to introduce a re®nement G of P
pP0
y0 2 P0 and the corresponding ones Fy00 2 F0 , such that each ®ber of G intersects R
f at most one
where F0 is thus a ®bering of R
g complementary to element. Our choice is y 2 pG
y 0 when
S. For given y0 2 R
f the ®ber Fy00 depends only on the
a a0 ; b b0 ; c c0 ; A A0 ; B B0 ; C any
mapping f and not on the generalized inverse map-
ping g. Obviously y 2 pG
y 0 ) y 2 pP
y 0 and therefore
pG
y 0 pP
y 0 , so that G is indeed a re®nement of P.
Example 2. We shall now give an example where the If y 0 2= R
f then obviously for any y 2 pG
y 0 also
generalized inverse is not de®ned directly but follows y2= R
f , i.e. pG
y 0 \ R
f [: If y 0 y0 2 R
f then
from the choice of ®berings and subspaces as well
as a ®bering mapping C, along the lines of Propo- sin A0 sin B0
a 0 c0 ; b0 c0 ;
sition 2. sin
A0 B0 sin
A0 B0
We consider the same mapping f as in Example 1, C0 p ÿ
A0 B0
but we will choose a dierent description of the range
R
f and the solution spaces Fy , using this time c; A; B as and y 2 pG
y0 when
the free parameters. In this case y a b c A B CT
sin A0 sin B0
2 R
f when a c0 ; b c0 ; c c0
sin
A0 B0 sin
A0 B0
sin A sin B
ac ; bc ; c any A A0 ; B B0 ; C any
sin
A B sin
A B
A any; B any; C p ÿ
A B in which case pG
y0 \ R
f fy0 g as required.
In order to choose a section M of G we must pick up
The conditions in this above description are in fact the a single element y from each ®ber pG
y 0 . Since the ele-
well-known angle and sine conditions ments a; b; c; A; B are already determined from the equal
a b c ones of y 0 , we are left with the possibility of choosing
A B C p; one value of C for every ®ber of G. Our choice is
sin A sin B sin C
C p ÿ
A B so that y 2 M when
We will introduce a ®bering P having R
f as a section
by y 2 pP
y 0 when a any; b any; c any
0 0 0
a any; b any; c c ; A A ; B B ; C any A any; B any; C p ÿ
A B
0
If y 2 pP
y \ R
f then y must satisfy simultaneously M is obviously a section of G since y 2 M \ pG
y 0 is
the conditions for y 2 pP
y 0 : uniquely determined from
c c0 ; A A0 ; B B0 a a0 ; b b0 ; c c0 ; A A0 ; B B0 ; C p ÿ
A0 B0
and the conditions for y 2 R
f :
The ®bering P induces a ®bering P0 on M with ®bers
sin A sin B pP0
y0 pP
y0 \ M. Combining the conditions for
ac ; bc ; C p ÿ
A B y 2 M with those for y 2 pP
y0 we conclude that
sin
A B sin
A B
y 2 pP0
y0 when
which, combined, show that y is uniquely determined
from a any; b any; c c0 ; A A0 ; B B0 ;
C p ÿ
A0 B0
sin A0 sin B0
a c0 ; b c0 ; c c0
sin
A0 B0 sin
A0 B0 For the ®bering F of the solution spaces Fy
y 2 R
f
0 0 0 0
we need a representation which uses c; A; B as free
AA; BB; C p ÿ
A B parameters: x 2 Fy whenever f
x y; i.e.,
85
x2 ÿ x1 2
y2 ÿ y1 2 c2 x1 a; x2 a c0 ; y1 b; y2 b
x3 a b0 cos A0 ; y3 b b0 sin A0
x3 ÿ x1
x2 ÿ x1
y3 ÿ y1
y2 ÿ y1
qq cos A Since Sa;b \ Fy00 fxg, the ®bering S is complementary
x3 ÿ x1 2
y3 ÿ y1 2
x2 ÿ x1 2
y2 ÿ y1 2 to F0 within R
g:
In order to obtain ®nally an explicit description of
x3 ÿ x2
x2 ÿ x1
y3 ÿ y2
y2 ÿ y1
qq cos B the generalized inverse mapping g : y ! x g
y, let
x3 ÿ x2 2
y3 ÿ y2 2
x2 ÿ x1 2
y2 ÿ y1 2 y a b c A B CT be an arbitrary element of Y . Then
y 2 Ra;b \ pP0
y0 when y0 p
y is given by
For R
g g
Y we must choose a subset of X which
sin A sin B
has the same dimension 5 as M, since R
g g
M also. a0 c ; b0 c ; c0 c
Our choice is to set x 2 R
g when sin
A B sin
A B
y2 y1 A0 A; B0 B; C0 p ÿ
A B
The ®bering F0 induced by F on R
g has ®bers and x g
y is uniquely determined from x 2 Sa;b \ Fy00 .
Fy0 Fy \ R
g. Thus taking y2 y1 into account, the This means that x g
y if we replace a a; b b and
preceding conditions for x 2 Fy we obtain, after some the values of the elements of y0 in the expression giving
algebraic manipulation, that x 2 Fy0 when x 2 Sa;b \ Fy00 . This results in
y3 ÿ y1 y3 ÿ y2 x1 a; x2 a c; y1 b; y2 b
y2 y1 ; x2 ÿ x1 c; tan A; tan B
x3 ÿ x1 x3 ÿ x2
sin B sin B
Since for every y the ®ber pP0
y will be mapped onto the x3 a c cos A; y3 b c sin A
0 sin
A B sin
A B
®ber Fp
y , it remains to introduce a ®bering R of M
which is complementary to P0 . We correspond to each thus bringing to an end Example 2.
pair of positive numbers a; b a ®ber Ra;b 2 R which is A special case of interest is the case of a minimal rank
de®ned by y 2 Ra;b when generalized inverse g where rg r. In this case
a a; b b; c any; A any; B any; rg dim
R
g r dim
R
f
C p ÿ
A B which combined with S R
g implies that S R
g:
In order to show that R is complementary to P0 (within
M), let y 2 Ra;b \ pP0
y0 . Combining the conditions for Lemma. When rg r the ®bering G induced by g
y 2 Ra;b with the ones for y 2 pP0
y0 we conclude that y coincides with the ®bering P induced by p f g:
is uniquely determined from
Proof. Since S R
g, the restriction gjR
f of g to the
a a; b b; c c0 ; A A0 ; B B0 ; range of f is in this case a bijection between R
f and
C p ÿ
A0 B0 R
g. For every z 2= R
f ; g
z 2 R
g S and since S is
a section of F, fg
zg S \ Fy for a unique ®ber Fy
Therefore Ra;b \ pP0
y0 fyg and R is complementary corresponding to a unique y 2 R
f . Therefore
to P0 . y f
g
z
f g
z p
z and z 2 pP
y
One more step remains to complete the de®nition of pP
f g
z. Consequently the ®bering G induced by
the generalized inverse g of f , that is to introduce a g coincides with the ®bering P induced by p f g. (
mapping
For the particular choice rg r, the last two re-
C : R ! S : Ra;b ! Sa;b quirements in Proposition 1 can be relaxed. The same
simpli®ed situation can arise in a dierent way by re-
where S is a ®bering of R
g complementary to F0 . quiring that in addition to g being a generalized inverse
We choose Sa;b C
Ra;b by requiring that x 2 Sa;b of f , at the same time f is a generalized inverse of g.
when Repeating property (G1) by interchanging the roles of f
x1 a; y1 b; y2 b and g leads to the following de®nition.
Obviously y1 y2 and thus Sa;b R
g. If x 2 Sa;b \ Fy00 De®nition. A generalized inverse g of a given mapping f
then combining these conditions for x 2 Sa;b with those is called a re¯exive generalized inverse if in addition to
for x 2 Fy00 we obtain (G1) it satis®es the following condition, henceforth
x1 a; y1 b; y2 b denoted (G2):
p ÿ
A B C p ÿ
A B C
A0 A ; B0 B ;
2 2
C0 C
These equations describe also the projection y0 p
y
for any given y.
For the corresponding ®ber Fy0 of F we need the
appropriate description x 2 Fy0 when
Lemma. For a re¯exive generalized inverse g, the
f g- The solution is easily veri®ed to be
induced ®bers are identical to the g-induced ®bers, i.e.,
P G: x1 23 c0 sin B0 ; y1 ÿ 13 a0 23 c0 cos B0
cos AÿBC
2
De®nition. The unique re¯exive generalized inverse g of
x1 23 a cos C ; a given mapping f which is also a minimum distance
cos BCÿA
2
and x0 -nearest generalized inverse is called the pseudo-
sin AÿBC
2 inverse of f .
y1 ÿ13a 23a cos C
cos BCÿA
2 Various other nonunique generalized inverses can be
de®ned by combination of (G1) with the some of the
cos AÿBC
2
other properties (G2), (G3), and (G4). The inverses
x2 ÿ13a cos C ; within the same class, i.e., the ones satisfying the same
cos BCÿA
2 set of properties, may dier in the following aspects:
sin AÿBC
2
y2 ÿ13a ÿ 13a cos C (1) In the section S when (G4) is not satis®ed.
cos BCÿA
2 (2) In the ®bering P induced by p f g when (G3) is
not satis®ed.
cos AÿBC
2
(3) In the re®nement G of P and the mapping
x3 ÿ13a cos C ; c
y; z cy
z extending g outside R
f when (G2) is
cos BCÿA
2 not satis®ed.
sin AÿBC
2
y3 23a ÿ 13a cos C ( In the geodetic case property (G3) is a ``must'', since
cos BCÿA
2 it solves the adjustment problem. Re¯exivity (G2) is not
necessary but convenient. The section S is speci®ed ei-
For re¯exive generalized inverses we have a possibi- ther directly by a set of minimal constraints or by in-
lity of choices, corresponding to the choices of the sec- troducing (G4). (G4) is introduced, either directly
tion S of F and of the ®bering P of Y over the elements (pseudo-inverse solution) or indirectly by a set of inner
of R
f . In analogy to the uniquely de®ned pseudo-in- constraints which describe the particular section S cor-
verse of a linear operator we seek two conditions, ad- responding to (G4).
ditional to (G1) and (G2), such that (G3) speci®es the
®bering P and (G4) speci®es the section S of F. De®nition. A set of (nonlinear) equations v
x; d 0, is
We introduce the minimum-distance generalized in- called a set of minimal constraints (with respect to f ) if
verse. Let qY be the distance function of the metric space the corresponding mapping v : X Rmÿr ! Rmÿr gives
Y . For any ®xed element y^ 2 R
f de®ne rise to a ®bering H of X , such that its member
Hd fx 2 Hd jv
x; d 0g, corresponding to a ®xed d,
Ry^ fz 2 Y ; qY
z; y^ min qY
z; yg
9 is a section of the ®bering F induced by f .
y2R
f
Eq. (9) we name (G3). In the particular case that the De®nition. A set of minimal constraints h
x d are
subsets Ry^ are ®bers of Y , as y^ varies over R
f , then we called inner constraints (with respect to a given point x0
can employ the following de®nition: g is a minimum of X ) when the ®ber Hd coincides with the section of F
distance generalized inverse of f , if P R, i.e., when the speci®ed by the property (G4), i.e., when for every
®bering P induced by p f g is identical with the Fy 2 F the unique element x such that Fy \ Hd fxg is
®bering R with ®bers Ry^. the one closest to x0 among all elements of Fy .
If Y is an inner-product vector space with metric
induced by a quadratic norm, we may use the term Remark. In the special case that f and its generalized
``least squares'' in place of ``minimum distance''. inverse are linear or ane mappings, R
f , the ®bers
We now de®ne the x0 -nearest generalized inverse. Fy ; Py ; Gx ; Qx , and the section S are all ane subspaces
Let qX be the distance function of X ; x0 2 X be ®xed and of the corresponding spaces Y and X . The ®bers
let F; P; G; Q consist of parallel ane subspaces, i.e., they
xFy min qX
x0 ; z
10 are equivalent to quotient spaces of Y and X . Each
z2Fy quotient space is uniquely determined by its member
passing through zero, i.e., by their ``modulo'' linear
This equation we denote (G4). In the particular case subspace in X or Y (Halmos 1974, Sect. 21) . The ®ber
that there exists only one element on every Fy which F, e.g., is equivalent to the quotient space of ane
satis®es (G4), then the set of all xFy constitutes a section subspaces parallel to the null space of f . Thus the linear
S of F, and we can employ the following de®nition: g is of ane generalized inverses are determined by a set of
an x0 -nearest generalized inverse of f if S is taken as the linear subspaces, instead of the ®berings (see Teunissen
image of R
f under g, i.e., S g
R
f : 1985). By the way, there is no reason to restrict the class
of generalized inverses of a linear operator to be itself
Remark. If X is an inner-product vector space with linear, as is usually done within the linear theory. A
metric induced by the norm, while x0 0, we may use linear operator may well have a nonlinear generalized
the term ``minimum norm'' in place of ``0-nearest''. inverse. A geodetic example is the ane generalized
However in the geodetic case 0 is a ``prohibited'' point, inverse obtained implicitly by imposing a set of inho-
since it corresponds to a network with all its points mogeneous linear constrains Hx d; d 6 0; on the
coinciding. solution of the least-squares adjustment problem.
88
The approach followed here can be characterized as a contradiction: let y be a point on the geodesic joining
naive, and falls short of a proper mathematical theory of y0 with y^2 such that y 2 BR
y0 , in which case
generalized inverses of nonlinear mappings, even in the q
y0 ; y^2 q
y0 ; y q
y; y^2 and q
y; y^1 q
y; y^2
relatively simple case of ®nite-dimensional spaces. The ) q
y0 ; y^1 q
y0 ; y q
y; y^1 , i.e., y belongs also to
reason is that the assumptions made are too strong for the geodesic joining y0 and y^1 , which is obviously im-
the purpose of building which to build a general theory, possible.
a fact which will become obvious even in the simple For the de®nition based on (G4) we can ascertain
geodetic case. Some of the assumptions that do not that the set of problematic points forms a subset of X
generally hold are the following: with measure zero, provided that the ®bers Ff
x do not
coincide locally with the boundaries of balls having x0 as
(1) f may not be de®ned at all points of the space X but center.
only on an open subset U X . In any case our de®nitions hold almost everywhere in
(2) g may not be possible to de®ne at all points of the X and Y , i.e., except on subsets of measure zero. From a
space Y but only on an open subset V Y : physical point of view we can always work in a neigh-
(3) The minimum distance property (G3) may not suf- borhood of X and the corresponding neighborhood of Y
®ce for the determination of a ®bering of Y or even which do not contain critical (problematic) points.
V . This has to do with the dierential geometric
properties of R
f as a submanifold of Y and in
particular with its curvature tensor. There might be 4 The geodetic nonlinear mapping
points in V such that their minimum distance from
R
f does not correspond to a unique point of R
f The nonlinear mapping f in the geodetic case arises
but either on a set of discrete points or even a non from n individual mappings fk ; k 1; . . . n, which map
discrete subset of R
f . (Think, e.g., of the case coordinates of points into various types of observables.
where R
f is a ball in Y and consider its center We consider a geodetic network of N points with
which belongs to any ®ber induced by the minimum Cartesian coordinates xi ; i 1; . . . ; N , where xi is a
distance principle.) vector of dimension either 2 (plane networks) or 3
(4) The x0 -nearest property (G4) may not suce for the (spatial networks). These vectors constitute a coordinate
determination of a ®bering of X or even U . This has vector
to do with the dierential geometric properties of 2 3
the ®bers Fy as submanifolds of X . There might be x1
one or more ®bers Fy on which there is more than 6 .. 7
x4 . 5
11
one point attaining the minimum distance from x0 .
xN
(5) The rank of f (=rank of dfx ) may not be the same at
all points x of U . The consequence is that R
f is no of dimension 2N or 3N , which is considered an element
longer a smooth submanifold of Y and its behavior of the m-dimensional Euclidean space (m 2N or
at singular points (points with rank (dfx < r) may m 3N )
pose additional problems.
ÿ N ÿ N
X E2N E2 or X E3N E3
We shall keep these problems in mind when study-
ing the particular case of the mappings f arising in equipped with the simple inner product
xa ; xb xTa xb .
geodetic problems. The geodetic observables we will examine here can be
The last three problems have to do with points in Y distinguished as ``angular'' observables yijk and ``dis-
or X which do not behave ``nicely'' with respect to the tance'' observables dik , which dier with respect to their
de®nitions employing (G3), (G4), and the de®nition of invariance characteristics under coordinate transforma-
the rank r of f , r rank
dfx ). Starting from the latter, tions. Angular observables are invariant under similarity
we note that the set of points y f
x 2 R
f , for which coordinates transformations x0 S
x while distance
rank
dfx < r, forms within R( f ) a set of measure zero. observables are invariant under rigid coordinate trans-
This is in fact a consequence of Sard's theorem (Stern- formations x0 R
x, which are de®ned pointwise by
berg 1964, ch. II.3, Theorem 3.1), provided we impose
some mild assumptions on the dierentiability of the x0i S
xi k R xi t; x0i R
xi R xi t
12
manifolds X and Y . In our case, where dim Y > dim X , it
suces to require that they are C 1 manifolds and f is t is a displacement vector, k > 0 a scale parameter and R
also a mapping of class C 1 . a proper orthogonal matrix
RT R R; RT I and
jRj 1) depending on a single rotational parameter h
A similar fact holds for the ``problematic'' points of in the two-dimensional case or on three rotational
Y which happen to belong to more than two sets Ry^ parameters h1 ; h2 ; h3 in the three-dimensional case. The
which thus intersect and fail to be ®bers of Y . If such observables are functions of the coordinates
points form a subset A Ry^1 \ Ry^2 Y which is not of
measure zero, then it will be possible to ®nd a y0 2 A and yijk y
xi ; xj ; xk yijk
x
a positive constant R such that BR
y0 A, where BR
y0 dik d
xi ; xk dik
x
13
is the ball with center y0 and radius R. We shall show
that the assumption that A has nonzero measure leads to which have the invariance properties
89
y
xi ; xj ; xk y
S
xi ; S
xj ; S
xk Since all Dik are closed, the same holds for D V C and
14 V is an open subset of X .
d
xi ; xk d
R
xi ; R
xk
Turning to the range space Y we note that distances
or are mappings from X to R , since they obtain only
yijk
x yijk
S
x; dik
x dik
R
x
15 positive values, while angles are mappings from X to the
unit circle S. Thus, strictly speaking, f is a mapping
The mapping f which maps the network coordinates x n2
to the observables y f
x 2 R
f Y consists in f : V !
Sn1
R
20
general of n1 angular mappings and n2 distance map- However it is standard practice to replace S with an
pings
n1 n2 n and we can distinguish two cases interval of R, our choice being I
0; 2p, in which case
with respect to its invariance properties. When n2 0 f becomes a mapping
(only angular observations) then
n
f
x f
S
x
16 f : V ! U
In1
R 2 Y En1 n2 En
21
while when n2 6 0 (observations of distances or both Since I does not contain its limit point 0 and the same
distances and angles) then holds true for R , U is an open subset of Y . We have
assumed that Y En is Rn equipped with the euclidean
f
x f
R
x inner product, though a more general inner product
ya ; yb yTa Pyb may be used, in which case Y EPn .
Remark. We must emphasize here that f is a nonlinear The restriction of f from a mapping f : X ! Y to a
mapping and S or R are general transformations not mapping f : V ! U , poses some serious problems as far
necessarily close to the identity (represented by as the properties (G3) and (G4) of generalized inverses
k 1; R I; t 0. It has been pointed out by Grafa- are concerned. Now the range R
f is restricted from
rend and Kampmann (1996) that there are other groups f
X to U \ f
X and further to U \ f
V . The ad-
of transformations, such as the ten-parameter conformal justment problem may have a solution y^ 2 f
X closest
group G C10
3, in three dimensions; although it does to a given y 2 U , such that y^ 2 = U \ f
V . This means
not satisfy Eq. (16), i.e., f
x 6 f
G
x, when it is close that adjusted observations y^ may include negative or
to the identity (G Id), it satis®es the corresponding zero distances, or angles outside the prescribed interval.
relation dfx0
x dfx0
G
x for the dierential map- Turning to the domain V we must distinguish be-
ping dfx0 of f at any Taylor point x0 . In other words tween the similarity transformation and the rigid
when G is close to the identity (``small'' transformation) transformation case. When distances are observed and
it leaves linearized angular observations invariant, a y^ 2 f
X , the ®bers Fy^ may be initially de®ned with the
property shared of course with S. Therefore the C10
3 help of any element x 2 X , such that ^ y f
x0 by means
group, although of essentially no relevance to the strictly of
nonlinear approach followed here, should be taken into
consideration in numerical applications, where we have Fy^ fx 2 X jx R
x0 g
22
to resort, explicitly or implicitly (iterative solutions), to
linearization and solutions x; x0 close to x0 , in which This means that if x0 2 V the same holds for any
case the transformation from any such x to any such x0 x R
x0 and therefore Fy^ V . Thus Fy^ is either
is necessarily close to the identity. included in V or lies completely outside V . This means
The presence of angular observations has as a con- that the problem of ®nding an element x^ of a ®ber closest
sequence that f is not de®ned at every point of X . For to a given element x0 2 V has a solution in V provided
an angular observable yijk y
xi ; xj ; xk ) to be de®ned it that Fy^ V , i.e., that y^ 2 f
V . When distance mea-
is necessary that the relevant network points are distinct, surements are not included the similarity transformation
i.e., we must exclude from X points x with xi xj , or S gives rise to ®bers
xj xk , or xk xi . Although the same problem does Fy^ fx 2 X jx S
x0 g
23
not appear in the case of distance observations, we will
stick to the restriction that those network points which which are no longer closed subsets of X , because they
are joined by observations must be distinct. Let J be the ``converge'' at the zero element 0 2 X . Although the
index set of those pairs
i; k of point indices corre- scale is restricted to positive values k > 0, it is possible to
sponding to points joined by observations and let Dik be consider a sequence ki ! 0 , giving rise to a sequence of
the diagonal subspace of X de®ned by similarity transformations Si such that
xi Si
x0 2 Fy^ ! 0 2
= Fy^ (convergence in norm). As a
Dik fx 2 X jxi xk g
18 consequence, the problem of ®nding an element x^ of a
We must exclude all inappropriate diagonal subspaces, ®ber closest to a given element, x0 2 V may have no
which leaves as the domain of de®nition of f the open solution (consider the case where 0 is the closest to x0
subset V X de®ned by its complement V C with respect element from the closure F y^ of the ®ber Fy^).
to X These are problems of great concern to the mathe-
matician, but of little or no concern at all to the geo-
VC [ Dik D
19 desist. The reason is that in geodetic applications we are
i;k 2 J con®ned to a small neighborhood Ny U \ f
V of the
90
true observables y 2 U \ f
V , such that the observa- the ®xed coordinates and azimuth, we specify the value
tions y and the adjusted observations y^ also belong to of the parameter set d. The ®rst step of this procedure
Ny . In the same way it is possible to choose an element corresponds to the choice of a ®bering S complemen-
x0 2 V such that Fy^ crosses a small neighborhood tary to F, where each ®ber corresponds to a particular
Nx0 V of x0 and furthermore the element x^ of Fy^ set of values for d. In the second step the assignment of
closest to x0 also lies in the same neighborhood Nx0 . speci®c values for d picks up a particular element S from
Although we con®ne ourselves here to the deter- S which will serve as the section of F into which g will
ministic aspects of generalized inverses, we are tempted map R
f .
to say that such a restriction to neighborhoods can be A particular choice for S is as the set of points x such
also justi®ed when a probabilistic criterion is used for that if S \ Fy fxg then x is the closest point to a given
the ``optimal'' choice of a generalized inverse, which x0 2 V among all elements of Fy . The description of this
provides an estimator of x based on the observations y particular section by the corresponding set of minimal
(outcomes of random variables). In this case a Bayesian constraints (inner constraints) is not as easy as in the
point of view can be followed with a noninformative linear case. Instead we shall follow an approach similar
prior distribution for x, such that it is homogeneous to that of Baarda's S-transformation in the linear case,
within the desired neighborhood in X , and zero outside. where a known reference ``minimal'' solution is trans-
For a glimpse into the possible connection between formed into the desired ``inner'' solution, i.e., the x0 -
the deterministic approach taken here and probabilistic nearest solution.
aspects, consider the case where y is the outcome of a In order to proceed with the solution we need an
random variable Y with likelihood function L
y; x analytical description of the solution ®bers as subman-
pY
yjx. If it so happens that L has the form L
x; y ifolds of X , e.g., by appropriate curvilinear coordinates.
L
q
y; f
x where L
q has a maximum when q is Let G denote the group of applicable transformations
minimized, while q quali®es as a metric, then maximum (i.e., either similarity or rigid) with elements g : X ! X .
likelihood estimate can be identi®ed with the solution If z belongs to a speci®c ®ber z 2 Fy then the same holds
following from the use of a minimum-distance general- true for x g
z 2 Fy for every g 2 G.
ized inverse. This establishes for every ®ber Fy of X a natural
With these considerations in mind we turn to the correspondence h : Fy Fy ! G :
nonlinear datum problem. We assume that the adjust-
ment problem has already been solved and its solution y^ g h
z; x , x g
z
24
de®nes a speci®c ®ber Fy^ V X , described by either This means that by ®xing one of the arguments of h
Eq. (22) or Eq. (23). we can obtain a mapping hz : x ! hz
x h
z; x :
Fy ! G which is invertible. If further fG : G ! Rd is a
coordinate system on G (i.e., a parametrization of G),
5 Solution to the nonlinear geodetic datum problem the composite mapping
The datum problem can be viewed as the problem of f fG hz : x ! p
x : Fy ! Rd
25
completing the projection mapping p : Y ! R
f which
solved the adjustment problem into a (minimum-dis- establishes a coordinate system on Fy . We have already
tance) generalized inverse g c p of f , with the choice introduced the coordinate systems f and the corre-
of a complementary mapping c : R
f ! S, where S is a sponding coordinates p in the descriptions, Eq. (12), of
section of the ®bering F induced by f in V . In fact it is the similarity and rigid transformations. The dimension
sucient to determine only S, in which case c (and thus d of Fy depends on the applicable transformation group
g) is automatically de®ned as the mapping of any and the dimension of the network. For two-dimensional
y 2 R
f into c
y x, where x is the unique element of networks d 4 for the similarity case (coordinates
S \ Fy . One way of describing the d-dimensional man- h; t1 ; t2 ; k) and d 3 for the rigid case (coordinates
ifold S is by means of a set of d m ÿ r minimal h; t1 ; t2 ). For three-dimensional networks d 7 for the
constraints v
x; d 0, where d is a ®xed vector of d similarity case (coordinates h1 ; h2 ; h3 ; t1 ; t2 ; t3 ; k) and
parameters, which usually appear in the explicit form d 6 for the rigid case (coordinates h1 ; h2 ; h3 ; t1 ; t2 ; t3 ).
h
x d. The alternative possible description of S by the In fact we have introduced not one but a family of
m constraints x /
d; q, where d is again a ®xed vector coordinate systems depending on the ``reference point'' z
of d parameters and q a vector of r free parameters, is chosen for any ®ber Fy . This choice can be made by
hardly used in geodesy. introducing a section Z of the ®bering F of X , e.g., by
Strictly speaking, a set of minimal constraints con- the use of minimal constraints which are easy to handle
sists of two things. First a family of mappings v
x; , or in computations, e.g., of the form xik 0 (trivial
/
; q, which corresponds to the choice of the type of constraints), where k is the network point index and i
constraints, and second a set of speci®c values for the the coordinate index (i 1; 2 for plane and i 1; 2; 3 for
parameters d. For example, in a horizontal network we three-dimensional networks).
may decide to ®x the two coordinates of a particular Once a reference solution z is established, all other
point and the azimuth of a particular side and thus solutions in the same ®ber x x
z; p depend on the
choose the function v
x; (or h
x, or /
; q, analo- coordinates p which are identical to the parameters of
gously). When we decide on the values to be assigned to the transformation g
p : z ! x. To obtain the solution
91
speci®ed by a given set of d minimal constraints three-dimensional networks and similarity or rigid
v
x; d 0 (usually of the form h
x d we have to transformations. The results are:
solve a system of d nonlinear equations with d un- Three-dimensional networks ± similarity transforma-
knowns tion:
P
v
p v
x
z; p; d 0
26
x0i ÿ x0 T R
zi ÿ z
or in the usual form in geodesy xi x 0 i P R
h
zi ÿ z
30
zi ÿ zT
zi ÿ z
i
h
p h
x
z; p d
27
where
The solution p of the preceding system speci®es the
transformation gp g
p which maps the reference 1X 1X
solution z into the minimal constraint solution z zi ; x0 x0i
31
N i N i
x gp
z. This transformation is called the nonlinear
Baarda S-transformation in the similarity, and the and h is the solution of the three nonlinear equations
nonlinear Baarda R-transformation in the rigid case.
The solution x closest to a given element x0 2 X is 1X
x0i ÿ x0 R
h
zi ÿ z 0
32
obtained either by minimizing N i
F
p x
z; p ÿ x0 T x
z; p ÿ x0
28 Three-dimensional networks ± rigid transformation:
or by imposing the condition that the vector x0 ÿ x is xi x0 R
h
zi ÿ z
33
orthogonal to the tangent space to the ®ber Fy^ at the
point x. Following the ®rst approach we are led to a where h is the solution of the three nonlinear equations
system of d nonlinear equations with d unknowns 1X
T
x0i ÿ x0 R
h
zi ÿ z 0
34
oF ox N i
p 0 )
z; px
z; p ÿ x0 0
29
op op
Two-dimensional networks ± similarity transformation:
The solution p^ of this system speci®es the transfor- P
mation gp^ g
^ p which maps the reference solution z
x0i ÿ x0 T R
zi ÿ z
i
into the solution x gp^
z closest to x0 . gp^ is the non- xi x0 P R
#
zi ÿ z
35
linear Baarda S-transformation (or R-transformation) to
zi ÿ zT
zi ÿ z
i
the x0 -nearest solution corresponding to the inner so-
lution of Meissl in the linear case. where # is the solution of the single nonlinear equation
What about the famous inner constraints of Meissl
which appear in the linear case? Well, oF 1X oR
op 0 can be
x0i ÿ x0 T
zi ÿ z 0
36
solved in principle to give p p
x0 ; z, which substituted N i o#
in x g
p
z x
z; p gives an expression of the form cos # sin #
x x
x0 ; z x
z, since x0 is ®xed a priori. Recalling When R
#
that z is determined by the intersection of a solution ÿ sin# cos #
1 b a
®ber Fy 2 F with a section S0 of F determined by a set xi x0 2
zi ÿ z
37
of minimal constraints v
z; d 0 or z /
d; q, it is sz ÿa b
possible in principle to have a description of the r-di- where
mensional manifold S0 in the form z z
q, where q is a
vector of r free parameters serving in fact as coordinates 1X 0 1
for S0 . If this representation is replaced in the inner a
x0i ÿ x0 T
zi ÿ z
38
N i ÿ1 0
solution x x
z we obtain x x
z
q w
q. The
resulting relation x w
q is the set of inner constraints 1X
which to every q assigns the element of the ®ber b
x0i ÿ x0 T
zi ÿ z
39
N i
pF
z
q of F passing through z z
q its element
closest to x0 . Unlike the linear case, the inner constraints 1X
are not independent of the reference minimal constraints s2z
zi ÿ zT
zi ÿ z
40
because they depend on their parametrization in terms N i
of q. The above description of the inner constraints is a
Two-dimensional networks ± rigid transformation:
conceptual rather than a practical one since the analyt-
ical solution of nonlinear equations or the derivation of xi x0 R
#
zi ÿ z
41
desired alternative descriptions of sections are not pos-
sible as a rule. where # is the solution of the single nonlinear equation
The detailed derivation of the speci®c form of the 1 Xh oR i
nonlinear system given by Eq. (29) is given in Appendix
x0i ÿ x0 T
zi ÿ z 0
42
A for the four special cases corresponding to two- or N i o#
92
cos # sin #
When R
# F0 FT0 vi l2i vi ) F0 FT0 VMVT
50
ÿ sin # cos#
1 b a V v1 v2 . . . vn ; Mik dik l2i
xi x0 p
zi ÿ z
43
2
a b 2 ÿa b
The eigenvectors fui g; fvi g form orthonormal systems
where a and b are de®ned as in the previous case. and they can be seen as the representations with respect
to the initial bases of the new bases, in which case U and
V become the matrices of change of bases, already
6 Application to the linear case introduced with the same notation. Furthermore the
number of non-zero eigenvalues is r rank
f for both
We have outlined an approach to the representation of k2i and l2i , which are also identical, i.e., l2i k2i ;
generalized inverses of nonlinear operators, and the i 1; 2; . . . r. In this case, if K is the r r diagonal
question arises how this approach agrees with the matrix with elements
existing representation theory for (linear) generalized
inverses of linear operators (Takos 1976; Teunissen Kik dik ki
51
1985). Since linear operators are simply special cases of the representations with respect to the new bases are
nonlinear operators we will proceed to show that the
existing theory can be derived as a special case of our K 0
nonlinear case. FT F UT
FT0 F0 U L
52
0 0
If f : X ! Y is a linear operator, we seek a linear
operator g : Y ! X which is a generalized inverse of f , T T T K 0
i.e., it satis®es property (G1). In terms of given ``initial'' FF V
F0 F0 V M
53
0 0
orthonormal bases feX Y0
i g for X and fei g for Y , f is
0
Proof. It follows directly from the property (G2): Remark. Upon looking into the explicit form of the
GFG G. ( matrices
In the linear case where X and Y are inner-product I KH Ir 0
FG r and GF , it is easy to
linear spaces with distances determined by the norm, the 0 0 JK 0
minimum-distance generalized inverse is called the least- see that H 0 ,
FGT FG, while J 0 ,
squares generalized inverse and choosing x0 0, the
GFT GF. Thus we obtain the more familiar form of
0-nearest generalized inverse is called minimum- norm the properties (G3) and (G4) for least-squares and
generalized inverse. minimum-norm generalized inverses of matrices.
Lemma. For a least-squares generalized inverse it holds Corollary. A re¯exive least-squares inverse is represent-
that H 0, i.e., ed by
ÿ1 ÿ1
K 0 K 0
G
63 G
65
J K J 0
a re¯exive minimum-norm inverse by
Proof. If G represents a least-squares inverse g then FG ÿ1
must represent an orthogonal projection p on R
f . In K H
G
66
other words, for any y 2 Y ; y^ p
y represented by 0 0
^
y FGy is the closest element to y from R
f if a least squares-minimum norm inverse by
y^ ÿ y ? R
f , i.e., y^ ÿ y 2 R
f ? , where R
f ? is the ÿ1
orthogonal complement of R
f in X . Since y^ ÿ y 2 C K 0
and C is a subspace complementary to R
f it must hold G
67
0 K
that C R
f ? . ÿKH
Since C is represented by R it must hold and ®nally a re¯exive least squares-minimum norm
If
that inverse (pseudo inverse) by
ÿ1
T K 0
ÿKH G
68
Fx 0 for every x 0 0
If
T T (
ÿKH K 0 x1
, 0 A generalized inverse G of the matrix F is speci®ed if the
If 0 0 x2 subspaces C Y and S X are known by means of
, ÿHT Kx1 0 for every x1 , H 0 ( speci®c representations and in addition the action of g
on C is speci®ed. Alternatively, G is uniquely determined
once (in addition to C and S) we know the images under
Lemma. For a minimum-norm generalized inverse it g of the elements of a basis of C, which images span and
holds that J 0, i.e., thus determine both the action of g on C and the
ÿ1 subspace D g
C N
f X .
K H Let C; S and D be represented in the original bases by
G
64
0 K R
C0 ; R
S0 and R
D0 with D0 G0 C0 .
In the special bases introduced by the singular value
decomposition
C is represented by the range R
C
Proof. If G represents a minimum-norm inverse g then q C1
represented by GF must project every x0 2 Fy to an R with jC2 j 6 0 (since dimC f ; C is n f ,
C2
element x^ 2 Fy which has the minimum norm among all
elements of the same ®ber Fy . Since Fy x0 N
f this so that C1 is r f and C2 is f f , which means that
happens only when x^ ? N
f . We have seen that q maps the f columns of C represent the elements of a basis for
elements of X on the section S, and in view of the same C.
dimension of these subspaces In the same special bases S is represented by the
ÿ1it must hold that
K S1
S N
f ? . The columns of span S and must range R
S R with jS1 j 6 0 (since dimS r; S
J S2
95
ÿ U2 UT2 S0
UT1 S0 ÿ1 Kÿ1 VT1 C0
VT2 C0 ÿ1 VT2 G0 S0
U1 KT S0 ÿ1 VT1 D0 VT2 C0 ÿ1 VT2
84
T
T U Minimum-norm generalized inverse:
which in view of I UU U1 U2 1T
U2
U1 UT1 U2 UT2 and the condition UT1 D0 0 becomes G0 U1
V1 Kÿ1 ÿ S0
U1 KT S0 ÿ1 VT1 C0
VT2 C0 ÿ1 VT2
D0 VT2 C0 ÿ1 VT2
85
G0 S0
UT1 S0 ÿ1 Kÿ1 VT1
ÿ S0
UT1 S0 ÿ1 Kÿ1 VT1 C0
VT2 C0 ÿ1 VT2 Re¯exive least-squares inverse:
I ÿ U1 UT1 D0 VT2 C0 ÿ1 VT2 G0 S0 U1 KT S0 ÿ1 VT1 86
ÿ S0
UT1 S0 ÿ1 Kÿ1 VT1 C0
VT2 C0 ÿ1 VT2 G0 S0
U1 KT S0 ÿ1 VT1 ÿ VT1 C0
VT2 C0 ÿ1 VT2
D0
VT2 C0 ÿ1 VT2 D0 VT2 C0 ÿ1 VT2
87
S0 KUT1 S0 ÿ1 VT1 ÿ VT1 C0
VT2 C0 ÿ1 VT2 Least squares-minimum norm inverse:
D0 VT2 C0 ÿ1 VT2 ( G0 U1
V1 Kÿ1 I ÿ C0
VT2 C0 ÿ1 VT2
88
97
Pseudo-inverse: X oxi
xi ÿ x0i T
oh
G0 U1 Kÿ1 VT1
U1 KKÿ1
V1 Kÿ1
89 i
X
A7
kRzi t ÿ x0i T
ÿkRzi RT X 0
All these representations involve the submatrices C0 , S0 , i
V1 , V2 , (V1 K) and (U1 K) which have a clear geometrical X
meaning. oxi X
xi ÿ x0i T
kRzi t ÿ x0i T I 0
A8
i
ot i
Appendix A X oxi X
xi ÿ x0i T
kRzi t ÿ x0i T
Rzi 0
Derivation of the nonlinear Baarda transformations i
ok i
to the inner solution
A9
We shall derive here the solution to the nonlinear which assuming that k 6 0 and jXj 6 0, can be rewritten
Baarda transformations, which are coordinate transfor- as
mations x S
z from any minimum constraints solu- X X X
tion z to the x0 -nearest element x of the f -induced ®ber k zi zi zi RT t zi RT x0i
A10
i i i
Ff
z . Here X X X
kR zi t x0i
A11
xT xT1 xT2 . . . xTN ; xT0 xT01 xT02 . . . xT0N
A1 i i i
X X X
and in the most general case the three-dimensional k zTi zi zTi RT t zTi RT x0i
A12
transformation sought can be expressed pointwise by i i i
the similarity transformation
Taking into account that zi zi 0 and setting
xi kR
hzi t xi
zi ; p; pT hT tT k
A2 1X 1X
z zi ; x0 x0i
A13
where the transformation parameters p consist of three N i N i
rotational parameters h h1 h2 h3 T de®ning the orthog- Eqs. (A10)±(A12) take the form
onal matrix R R
h, three parallel displacement X
parameters t t1 t2 t3 T , and a scale parameter k. N zRT t zi RT x0i
A14
The problem is to ®nd the optimal values of h; t; k i
which minimize f f
p
x ÿ x0 T
x ÿ x0 . We ®rst kRz t x0
A15
note that for X X
k zTi zi N zT RT t zTi RT x0i
A16
o
R_ R : i i
oa Solving Eq. (A15) for t
T _ T RR_ T 0
RR I ) RR t x0 ÿ kRz
A17
) RR _ T T x ) R_ xR
_ T ÿ
RR and replacing in Eq. (A16) gives
P T T
This can be applied for the derivatives with respect to hk zi R
x0i ÿ x0
by setting oho k R xk R. i
k P T
A18
zi zi ÿ N zT z
oxi o i
k Rzi kxk Rzi ÿk
Rzi xk
ohk ohk
A3 which can be also written in the form
ÿkRzi RT xk P
x0i ÿ x0 T R
zi ÿ z
oxi oxi oxi oxi i
k P
A19
A4
zi ÿ zT
zi ÿ z
oh oh1 oh2 oh3
i
ÿkRzi RT x1 x2 x3 ÿkRzi RT X
To determine R (in fact h) we replace t in Eq. (A14)
where X
N zRT x0 ÿ kN zz zi RT x0i
A20
oxi oxi i
I; Rzi
A5
ot ok which in view of zz 0 can be written as
To minimize f we simply set 1X
zi RT x0i ÿ zRT x0
of of ox T ox ox ox
N i
2
x ÿ x0 0
A6
A21
op ox op oh ot ok 1X
x0i Rzi ÿ x0 Rz 0
N i
or explicitly
98
or in the equivalent form With this particular choice Eq. (A28) takes the form
1X a cos # b sin #
A32
x0i ÿ x0 R
zi ÿ z 0
A22
N i
where
Equation (A22) is a nonlinear equation which can be 1X
solved to obtain the values of the parameters h in any a
x0i ÿ x0 T W
zi ÿ z
N i
particular parametrization of the rotation matrix R
h.
A33
The obtained values should be next substituted in 1X
b
x0i ÿ x0 T
zi ÿ z
Eq. (A19) in order to determine the value of the scale N i
parameter k. The similarity transformation can be re-
alized using these values, once t and k are replaced from There are two solutions: cos # p 6
; sin #
Eq. (A17) and (A19), respectively, into Eq. (A2) to ob- a2 b2
tain p
6
2
a b 2
(both positive or both negative). Replacing in
P Eq. (A29) we obtain
x0i ÿ x0 T R
zi ÿ z
i p
xi x0 P R
h
zi ÿ z
A23 b cos # a sin # a2 b2
zi ÿ zT
zi ÿ z k
A34
i s2z s2z
In the case of the rigid transformation xi R
hzi t we where
obtain only Eds. (A7), (A8) with k 1 and following the 1X
same procedure as before we arrive at the solution s2z
zi ÿ zT
zi ÿ z
A35
N i
1X
x0i ÿ x0 R
zi ÿ z 0
A24 With the obtained values of sin #; cos # and k, the planar
N i
similarity transformation of Eq. (A30) becomes
xi x0 R
h
zi ÿ z
A25
1 b a
xi x0 2
zi ÿ z
A36
The planar (two-dimensional) case is somewhat dif- sz ÿa b
ferent since the 2 2 rotation matrix R
# depends only
on a single parameter #. Setting In the case of the planar rigid transformation, the
solution is given by
o _ T
W RRT RR
A26 1X
o#
x0i ÿ xT WR
zi ÿ z 0
A37
N i
where W is a 2 2 antisymmetric matrix, the only
dierence from the three-dimensional similarity trans-
formation case (apart from the fact that the vectors xi x0 R
#
zi ÿ z
A38
xi ; x0i ; zi ; x0 ; z involved are now two-dimensional) is that
Eq. (A7) is replaced by and for the above speci®c parametrization of R
# we
obtain two solutions
X oxi X
xi ÿ x0i T
kRzi t ÿ x0i T
kWRzi 0 1 b a
i
o# i xi x0 p
zi ÿ z
A39
a2 b2 ÿa b
A27
which correspond to two rotation angles # and # p. A
and the ®nal solution for the planar similarity transfor- further investigation is needed in each particular appli-
mation takes the form cation, in order to determine which of the two solutions
1X minimizes in fact the target function f
x0i ÿ x0 T WR
zi ÿ z 0
A28
x ÿ x0 T
x ÿ x0 :
N i
P
x0i ÿ x0 T R
zi ÿ z
i Appendix B
k P
A29
zi ÿ zT
zi ÿ z
i Proof of the equivalence of the representations
of Eqs. (81) and (82)
xi x0 kR
#
zi ÿ z
A30
Our starting point is Eq. (81)
We can proceed further with the solution by choosing
the usual parametrization
G0 S0 KUT1 S0 ÿ1 VT1 ÿ VT1 C0
VT2 C0 ÿ1 VT2
cos # sin # 0 1 (B1)
R
# )W
A31
ÿ sin # cos # ÿ1 0 D0 VT2 C0 ÿ1 VT2
99
We need an interpretation of this formula which will be With the obvious simplest choice B Id we obtain
independent of U1 ; U2 ; V1 ; V2 and depends only on
S0 ; C0 ; D0 , and related matrices. ? 0
FS : (
Let us consider in addition to the matrix C with Id
column span R
C a matrix C? with column span
R
C? R
C? . This means that C? must be such that Returning to the original bases we have
International Association of Geodesy, Vienna, August 11±24, Meissl P (1969) Zusammengfassung und Ausbau der inneren
1991 Fehlertheorie eines Punkthaufens. Deutsche Geod Komm A
Dermanis A, Grafarend E (1981) Estimability analysis of geodetic, 61: 8±21
astrometric and geodynamical quantities in Very Long Baseline Mierlo, J van (1980) Free network adjustment and S-transforma-
Interferometry. Geophys J R Astronom Soc 64: 31±56 tions. Deutsche Geod Komm B 252: 41±54
Dermanis A, Sanso F (1995) Nonlinear estimation problems for Moore EH (1920) On the reciprocal of the general algebraic matrix
nonlinear models. Manuscr Geod 20: 110±122 (abstract). Bull Amer Math Soc 26: 394±395
Grafarend E (1974) Optimization of geodetic networks. Boll Geod Penrose R (1955): A generalized inverse for matrices. Proc Cam-
Sci Ani 33: 351±406 bridge Philos Soc 52: 17±19
Grafarend E, Kampmann G (1996) C10
3: the ten-parameter Rao CR, Mitra SK (1971) Generalized inverse of matrices and its
conformal group as a datum transformation in three-dimen- applications. Wiley, New York
sional Euclidean space. Z Vermessungswesen 121: 68±77 Scharin B, Heidenreich E, Grafarend E (1977) A representation
Grafarend E, Scharin B (1974) Unbiased free net adjustment. of the standard gravity ®eld. Manuscr Geod 2: 135±174
Surv Rev XXII, 171: 200±218 Sternberg S (1964) Lectures on dierential geometry. Prentice-Hall,
Grafarend E, Scharin B (1976) Equivalence of estimable quanti- Englewood Clis, N.J.
ties and invariants in geodetic networks. Z Vermessungswesen Takos I (1976) Generalized inverse of matrices and its applications
101: 485±491 (in Greek). Bull Hellenic Army Geogr Service 1976: 77±133
Grafarend E, Scharin B (1991) The planar trisection problem and Teunissen PJG (1985) The geometry of geodetic inverse linear
the impact of curvature on non-linear least squares estimation. mapping and non-linear adjustment. Netherlands Geodetic
Comput Stat Data Anal 12: 187±199 Commission, Publications on Geodesy, New Series, Vol. 8,
Halmos PR (1974) Finite-dimensional vector spaces. Springer, No. 1
Berlin Heidelberg New York Teunissen PJG (1989a) Nonlinear inversion of geodetic and geo-
Koch K-R (1982) S-transformations and projections for obtaining physical data: diagnosing nonlinearity. In: Brunner FK, Rizos
estimable parameters. In: Blotwijk, M.J., F.J.J. Brouwer, D.T. C (eds) Developments in four-dimensional geodesy. Lecture
van Daalen, H.M. de Heus, J.T. Gravesteijn, H.C. van der Notes in Earth Sciences, No. 29. Springer, Berlin Heidelberg
Hoek, J.J. Kok, P.J.G. Teunissen (eds) 40 Years of Thought, New York pp 241±264
Anniversary volume for Prof. Baarda's 65th Birthday Vol. 1, Teunissen PJG (1989b) First and second moments of non-linear
Technische Hogeschool Delft, Delft, pp 136±144 least-squares estimators. Bull Geod 63: 253±262
Lohse P (1994) Ausgleichungsrechnung in nichtlinearen modellen. Teunissen PJG (1990) Non-linear least squares. Manuscr Geod 15:
Deutsche Geod Komm C 429 137±150
Loomis LH, Sternberg S (1980) Advanced calculus. Addison- Xu P (1995) Testing the hypotheses of non-estimable functions in
Wesley, Reading Mass free net adjustment models. Manuscr Geod 20: 73±81
Meissl P (1965) UÈber die Innere Genauigkeit dreidimensionaler
Punkthaufen. Z Vermessungswesen 90: 109±118