Awange J.,Grafarend E., Palancz B., Zaletnyik P. Algebraic Geodesy and Geoinformatics
Подождите немного. Документ загружается.
9-3 Partial procrustes solution 117
obtain the sum of squares kA − BTk
2
which is then minimized. One proceeds via
Frobenius norm as follows:
min
kX − YTk :=
q
tr(X
0
− T
0
Y
0
)(X − YT)
T
0
T = I.
(9.2)
In-order to obtain T in (9.2), the following prop erties of a matrix in Table (9.1)
are essential.
Table 9.1. Matrix properties for procrustes analysis
(a) tr A =
n
P
i=1
a
ii
Definition of trace function
(b) tr A = tr A
0
Invariant under transpose
(c) tr ABC = tr C AB = tr BCA Invariant under ‘cyclic’ permutation
(d) tr A
0
B = tr (A
0
B)
0
= tr B
0
A = tr AB
0
Combining prop erties (b) and (c)
(e) tr (A + B) = tr A + trB Summation rule
Using (9.2) and the properties of Table 9.1, one writes
kA − BTk
2
:= tr(A
0
− T
0
B
0
)(A − BT)
T
0
T = I
= tr(A
0
A − 2A
0
BT + T
0
B
0
BT)
= trA
0
A − 2trA
0
BT + trB
0
B
trT
0
B
0
BT = trTT
0
B
0
B = trB
0
B.
(9.3)
The simplification trT
0
B
0
BT = trB
0
B in (9.3) is obtained by using the property
of invariance of the trace function under cyclic permutation (i.e., property (c) in
Table 9.1). Since tr(A
0
A) and tr(B
0
B) are not dependent on T, we note from
(9.3) that
kA − BTk
2
= min ⇔ tr(A
0
BT) = max
T
0
T = TT
0
= I
k
.
(9.4)
If UΣV
0
is the singular value decomposition of A
0
B and C = A
0
B, then we have
A
0
B = UΣV
0
if C = UΣV
0
, U, V
0
∈ SO(3)
Σ = Diag(σ
1
, . . . , σ
k
)then
tr(CT) ≤
k
P
i=1
σ
k
,
with
k = 3.
(9.5)
The proof for (9.5) is given by [294, p. 34] as follows: Substituting for C from its
singular value decomposition and with the property (c) in Table 9.1, one writes
118 9 Procrustes solution
tr(CT) = tr(UΣV
0
T) = tr(ΣV
0
TU)
taking
R = (ij)1 ≤ i, j ≤ k = V
0
TU orthogonal and |r
ii
| ≤ 1
then
tr(ΣV
0
TU) =
k
P
i=1
σ
i
r
ii
≤
k
P
i=1
σ
i
.
(9.6)
From (9.5) and (9.6), one notes that
tr(A
0
BT) = max ⇔ tr(A
0
BT) ≤
k
X
i=1
γ
i
, (9.7)
subject to the singular value decomposition
A
0
B = UΣV
0
, U, V ∈ SO(3) and orthogonal.
(9.8)
Finally, the maximum value is obtained as
max(trA
0
BT) =
k
X
i=1
γ
i
⇔ T = VU
0
. (9.9)
Thus the solution of the rotation matrix by Procrustes method is
T = VU
0
. (9.10)
9-32 Partial derivative formulation
This approach is attributed to P. H. Schonemann [355] as well as [356]. Proceeding
from the Frobenius norm in (9.2) and using (9.1) leads to
d
1
= tr A
0
A − 2tr A
0
BT, +T
0
B
0
BT, (9.11)
while the condition that T
0
T = I leads to
d
2
= Λ(T
0
T − I). (9.12)
where Λ is the m x m unknown matrix of Lagrange multipliers. Equations (9.11)
and (9.12) are added to give
d = d
1
+ d
2
. (9.13)
The derivative of (9.13) are obtained with respect to T as
∂d
∂T
=
∂d
1
∂T
+
∂d
2
∂T
=
∂
tr A
0
A −2tr A
0
BT + tr T
0
B
0
BT
∂T
+
∂
Λ T
0
T − Λ I
∂T
= −2 B
0
A + B
0
BT + B
0
BT + TΛ + TΛ
0
=
B
0
B + B
0
B
T −2 B
0
A + T
Λ + Λ
0
.
(9.14)
9-4 The General Procrustes solution 119
From (9.14), let
B
0
B = B
∗
, B
0
A = C and
Λ + Λ
0
= 2Λ
∗
. (9.15)
For an extremum value of d, we set
∂d
∂T
= 0 such that
2C = 2B
∗
T + 2TΛ
∗
C = B
∗
T + TΛ
∗
,
(9.16)
leading to both B
∗
and Λ
∗
being symmetric. Hence
Λ
∗
= T
0
C − T
0
B
0
T. (9.17)
But B
∗
→ symmetric and thus T
0
B
∗
T is also symmetric. T
0
C is therefore sym-
metric or
T
0
C = C
0
T
from the side condition
T
0
T = TT
0
= I
3
we have that
C = TC
0
T.
(9.18)
From (9.8), we had C = A
0
B = UΣV
0
by SVD. In the present case we note that
C = B
0
A, thus C = B
0
A = VΣU
0
. From (9.18) we have
with U
0
U = UU
0
= V
0
V = V V
0
= I
3
C = TC
0
T
V Σ U
0
= TUΣV
0
T
V = T U
or T = VU
0
,
(9.19)
which is identical to (9.10).
9-4 The General Procrustes solution
In Sect. 9-3, we presented the partial Procrustes algorithm and referred to it
as “partial” because it was applied to solve only the rotation component of the
datum transformation problem. The analysis of the parameterized conformal group
C
7
(3) (7-parameter similarity transformation) as presented in Sect. 17-1, however,
requires the estimation of the dilatation unknown (scale factor), unknown vector of
translation and the unknown matrix of rotation. These unknowns are determined
from a given matrix data set of Cartesian coordinates as pseudo-observations.
In addition to the unknown rotation matrix which was determined in Sect. 9-
3, therefore, one has to determine the scale and translation elements. The partial
120 9 Procrustes solution
Procrustes algorithm gives way to the general Procrustes algorithm. The transpose
which was indicated by {
0
} in Sect. 9-3 will be denoted by {∗} in this section. In
Sect. 17-1, the 7-parameter datum transformation problem will be formulated such
that the solution of (17.1) lead to the desired seven parameters.
The unknown parameters for the 7-parameter transformation problem are
a scalar-valued scale factor x
1
∈ R, a vector-valued translation parameters
x
2
∈ R
3×1
(column vector) and a matrix valued rotation parameters X
3
∈
O
+
(3) := {X
3
∈ R
3×3
| X
∗
3
X
3
= I
3
, | X
3
|= +1}, which in total constitute
the 7-dimensional parameter space. x
1
represents the dilatation unknown (scale
factor), x
2
the translation vector unknown (3 parameters) and X
3
the unknown
orthonormal matrix (rotation matrix) which is an element of the special orthog-
onal group in three dimension. In other words, the O
+
(3) differentiable manifold
can be coordinated by three parameters. In (10.23) on p. 148, relative position
vectors are used to form the two matrices A and B in the same dimensional space.
If the actual c oordinates are used instead, the matrix-valued pseudo-observations
{Y
1
, Y
2
} become
x
1
x
2
. . . x
n
y
1
y
2
. . . y
n
z
1
z
2
. . . z
n
∗
=: Y
1
, Y
2
:=
X
1
X
2
. . . X
n
Y
1
Y
2
. . . Y
n
Z
1
Z
2
. . . Z
n
∗
, (9.20)
with {Y
1
and Y
2
} replacing {A and B}. The coordinate matrices of the n points
(n−dimensional simplex) of a left three-dimensional Weitzenb¨ock space as well as
a right three-dimensional Weitzenb¨ock space, namely Y
1
∈ R
n×3
and Y
2
∈ R
n×3
constitute the 6n dimensional observation space. Left and right matrices {Y
1
, Y
2
}
are related by means of the passive 7-parameter conformal group C
7
(3) in three
dimensions (similarity transformation, orthogonal Procrustes transformation) by
(cf., 17.1 on p. 304)
Y
1
.
= F (x
1
, x
2
, X
3
| Y
2
) = Y
2
X
∗
3
x
1
+ 1x
∗
2
, 1 ∈ R
n×1
. (9.21)
The nonlinear matrix-valued equation F (x
1
, x
2
, X
3
| Y
2
)
.
= Y
1
is inconsistent
since the image <(F) ⊂
6=
D(Y
1
) of F (range space <(F )) is constrained in the
domain D(Y
1
) of Y
1
∈ R
n×3
(domain space D(Y
1
)). First, as a mapping, F is
“not onto, but into” or “not surjective”. Second, by means of the error matrix
E ∈ R
n×3
which accounts for errors in the pseudo-observation matrices Y
1
as
well as Y
2
, respectively, we are able to make the nonlinear matrix-valued equation
F (x
1
, x
2
, X
3
| Y
2
)
.
= Y
1
as identity. In this case,
Y
1
= F (x
1
, x
2
, X
3
| Y
2
) + E = Y
2
X
∗
3
x
1
+ 1x
∗
2
+ E. (9.22)
Furthermore, excluding configuration defect which can be detected a priori we shall
assume ℵ(F ) = {0}, the kernel of F (null space ℵ(F )) to contain only the zero
element (empty null space ℵ(F)). A simplex of minimal dimension which allows
the computation of the seven parameters of the space X is constituted by n = 4
points, namely a tetrahedron which is presented in the next examples.
9-4 The General Procrustes solution 121
Example 9.1 (Simplex of minimal dimension, n = 4 points, tetrahedron).
Y
1
=
x
1
x
2
x
3
x
4
y
1
y
2
y
3
y
4
z
1
z
2
z
3
z
4
∗
∈ R
n×3
, Y
2
=
X
1
X
2
X
3
X
4
Y
1
Y
2
Y
3
Y
4
Z
1
Z
2
Z
3
Z
4
∗
∈ R
n×3
,
x
1
y
1
z
1
x
2
y
2
z
2
x
3
y
3
z
3
x
4
y
4
z
4
=
X
1
Y
1
Z
1
X
2
Y
2
Z
2
X
3
Y
3
Z
3
X
4
Y
4
Z
4
X
∗
3
x
1
+ 1x
∗
2
+
e
11
e
12
e
13
e
21
e
22
e
23
e
31
e
32
e
33
e
41
e
42
e
43
.
We will now introduce the weight component and solve it the problem using Gen-
eral Procrustes solution.
Example 9.2 (Weighted LEast Squares’ Solution W-LESS). We depart from the
set up of the pseudo-observation equations given in Example 9.1 (simplex of
minimal dimension, n = 4 points, tetrahedron). For a diagonal weight W =
Diag (w
1
, . . . , w
4
) ∈ R
4×4
we c ompute the Frobenius error matrix W-semi-norm
k E k
2
W
:= tr(E
∗
WE) =
tr
e
11
e
21
e
31
e
41
e
12
e
22
e
32
e
42
e
13
e
23
e
33
e
43
w
1
0 0 0
0 w
2
0 0
0 0 w
3
0
0 0 0 w
4
e
11
e
12
e
13
e
21
e
22
e
23
e
31
e
32
e
33
e
41
e
42
e
43
= tr
e
11
w
1
e
21
w
2
e
31
w
3
e
41
w
4
e
12
w
1
e
22
w
2
e
32
w
3
e
42
w
4
e
13
w
1
e
23
w
2
e
33
w
3
e
43
w
4
e
11
e
12
e
13
e
21
e
22
e
23
e
31
e
32
e
33
e
41
e
42
e
43
=
w
1
e
2
11
+ w
2
e
2
21
+ w
3
e
2
31
+ w
4
e
2
41
+w
1
e
2
12
+ w
2
e
2
22
+ w
3
e
2
32
+ w
4
e
2
42
+w
1
e
2
13
+ w
2
e
2
23
+ w
3
e
2
33
+ w
4
e
2
43
.
Obviously the coordinate errors (e
11
, e
12
, e
13
) have the same weight w
1
, (e
21
, e
22
, e
23
) →
w
2
, (e
31
, e
32
, e
33
) → w
3
and finally (e
41
, e
42
, e
43
) → w
4
. We may also say that
the error weight is pointwise isotropic, namely weight e
11
=weight e
12
=weight
e
13
=weight w
1
etc. But the error weight is not homogeneous since w
1
=weight
e
11
6=weight e
21
= w
2
. Of course, an ideal homogeneous and isotropic weight dis-
tribution is guaranteed by the criterion w
1
= w
2
= w
3
= w
4
= w.
By means of Solution 9.1 we have summarized the parameter space (x
1
, x
2
, X
3
) ∈
R × R
3
× R
3×3
. In contrast, Solution 9.2 reviews the pseudo-observation space
(Y
1
, Y
2
) ∈ R
n×3
× R
n×3
equipp e d with the Frobenius matrix W–semi-norm.
Solution 9.1 (The parameter space X).
x
1
∈ R dilatation parameter (scale factor)
x
2
∈ R
3×1
column vector of translation parameters
122 9 Procrustes solution
X
3
∈ O
+
(3) := {X
3
∈ R
3×3
| X
∗
3
X
3
= I
3
, | X
3
|= +1}
orthonormal matrix,
rotation matrix of three
parameters
Solution 9.2 (The observation space Y).
Y
1
=
x
1
x
2
. . . x
n
y
1
y
2
. . . y
n
z
1
z
2
. . . z
n
∗
∈ R
n×3
, Y
2
=
X
1
X
2
. . . X
n
Y
1
Y
2
. . . Y
n
Z
1
Z
2
. . . Z
n
∗
∈ R
n×3
left three-dimensional coordinate right three-dimensional coordinate
matrix of an n− dimensional matrix of an n− dimensional
simplex simplex
The immediate problem that one is faced with is how to solve the inconsistent
matrix-valued nonlinear equation (9.22). Esse ntially, this is the same problem that
we introduced in (17.1) on p. 304. The difference between (17.1) and (9.22) is the
incorporation of the error matrix E in the latter. This takes into consideration
the stochasticity of the systems Y
1
and Y
2
. In what follows W-LESS (i.e., the
Weighted LEast Squares’ Solution) is defined and materialized by the Procrustes
algorithm presented by means of:
• Corollary 9.1 (partial W-LESS for the unknown vector x
2l
).
• Corollary 9.2 (partial W-LESS for the unknown scalar x
1l
).
• Corollary 9.3 (partial W-LESS for the unknown matrix X
3l
).
The partial optimization results are collected in Theorem 9.1 (W-LESS of Y
1
=
Y
2
X
∗
3
x
1
+ 1x
∗
2
+ E) and Corollary 9.4 (I-LESS of Y
1
= Y
2
X
∗
3
x
1
+ 1x
∗
2
+ E).
Solution 9.3 summarizes the general Procrustes algorithm.
Definition 9.1 (W-LESS). The parameter array {x
1l
, x
2l
, X
3l
} is called W-
LESS (least squares solution with respect to Frobe nius matrix W–semi-norm) of
the inconsistent nonlinear matrix-valued system of equations
Y
2
X
∗
3
x
1
+ 1x
∗
2
+ E = Y
1
, (9.23)
subject to
X
∗
3
X
3
= I
3
, | X
3
|= +1, (9.24)
if for the parameter array in comparison to all other parameter arrays {x
1l
, x
2l
, X
3l
},
the inequality
k Y
1
− Y
2
X
∗
3l
x
1l
− 1x
∗
2l
k
2
W
:= tr((Y
1
− Y
2
X
∗
3l
x
1l
− 1x
∗
2l
)
∗
W(Y
1
− Y
2
X
∗
3l
x
1l
− 1x
∗
2l
)
≤ tr((Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
)
∗
W(Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
)
=:k Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
k
2
W
(9.25)
9-4 The General Procrustes solution 123
holds, in particular if E
l
:= Y
1
− Y
2
X
∗
3l
x
1l
− 1x
∗
2l
has the minimal Frobenius
matrix W–semi-norm such that W ∈ R
n×n
is positive semi-definite.
Note that k E k
2
W
:= tr(E
∗
WE) characterizes the method of least squares tuned to
an error matrix E ∈ R
n×3
and a positive semi-definite weight matrix W. Indeed
a positive semi-definite we ight matrix W of weights is chosen in-order to have
the option to exclude by means of zero weight a particular pseudo-observation,
say a particular coordinate row vector [x
i
, y
i
, z
i
], i ∈ N arbitrary, but fixed by
w
ii
= w
i
= 0, which may be an outlier. Example 9.2 illustrates details of Definition
9.1.
In order to construct W-LESS of the inconsistent nonlinear matrix-valued sys-
tem of equations (9.23) subject to (9.24) we introduce the Procrustes algorithm.
The first algorithmic step is constituted by the forward computation of the trans-
formation parameters x
2l
from the unconstraint Lagrangean L(x
1
, x
2
, X
3
) which is
twice the value of the Frobenius error matrix W–semi-norm. As soon as the trans-
lation parameters x
2l
are backward substituted we gain a Lagrangean L(x
1
, X
3
)
which is centralized with resp e ct to the observation matrix Y
1
− Y
2
X
∗
3
x
1
. In the
second algorithmic step the scale parameter x
1l
is forward computed from the cen-
tralized Lagrangean L(x
1
, X
3
). Its backward substitution leads to the Lagrangean
L(X
3
) which is only dependent on the rotation matrix X
3
. Finally the optimiza-
tion problem L(X
3
) = min subject to X
∗
3
X
3
= I
3
, | X
3
|= +1 generates the third
algorithmic step. This computational step is similar to that of partial Procrustes
algorithm of Sect. 9-3. By means of singular value decomposition SVD the rotation
matrix X
3l
is forward computed and backward substituted to gain (x
1
, x
2
, X
3
) at
the end. The results are collected in Corollaries 9.1 to 9.3.
Corollary 9.1 (Partial W-LESS for the translation vector x
2l
). A 3 × 1
vector x
2l
is partial W-LESS of (9.23) subject to (9.24) if and only if x
2l
fulfills
the system of normal equations
1
∗
W1x
2l
= (Y
1
− Y
2
X
∗
3
x
1
)
∗
W1. (9.26)
The translation vector x
2l
always exist and is represented by
x
2l
= (1
∗
W1)
−1
(Y
1
− Y
2
X
∗
3
x
1
)
∗
W1. (9.27)
For the special case W = I
n
, i.e., the weight matrix is unit, the translational
parameter vector x
2l
is given by
x
2l
=
1
n
(Y
1
− Y
2
X
∗
3
x
1
)
∗
1.
proof
W-LESS is constructed by unconstraint Lagrangean
124 9 Procrustes solution
L(x
1
, x
2
, X
3
) :=
1
2
k E k
2
W
=k Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
k
2
W
=
1
2
tr(Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
)
∗
W(Y
1
− Y
2
X
∗
3
x
1
− 1x
∗
2
) = min,
subject to {x
1
≥ 0, x
2
∈ R
3×1
, X
∗
3
X
3
= I
3
}
∂L
∂x
∗
2
(x
2l
) = (1
∗
W1)x
2
− (Y
1
− Y
2
X
∗
3
x
1
)
∗
W1 = 0,
(9.28)
constitutes the first necessary condition. Basics of vector-valued differentials are
as given in Table 9.1, p. 117. For more details on matrix properties and manipula-
tions, we refer to [180, pp. 439–451]. As soon as we back-substitute the translation
parameter x
2l
, we are led to the centralized Lagrangean
L(x
1
, X
3
) =
1
2
tr{
Y
1
− (Y
2
X
∗
3
x
1
+ (1
∗
W1)
−1
11
∗
W(Y
1
− Y
2
X
∗
3
x
1
))
∗
W
∗ ∗
Y
1
− (Y
2
X
∗
3
x
1
+ (1
∗
W1)
−1
11
∗
W(Y
1
− Y
2
X
∗
3
x
1
))
}
(9.29)
L(x
1
, X
3
) =
1
2
tr{
(I − (1
∗
W1)
−1
11
∗
)W(Y
1
− Y
2
X
∗
3
x
1
)
∗
W
∗ ∗
(I − (1
∗
W1)
−1
11
∗
)W(Y
1
− Y
2
X
∗
3
x
1
)
}
(9.30)
C := I
n
− (1
∗
W1)
−1
11
∗
W (9.31)
is a definition of the centering matrix , namely for W = I
n
C := I
n
−
1
n
11,
∗
(9.32)
being symmetric, in general. Substituting the centering matrix into the reduced
Lagrangean L(x
1
, X
3
), we gain the centralized Lagrangean
L(x
1
, X
3
) =
1
2
tr{[Y
1
− Y
2
X
∗
3
x
1
]
∗
C
∗
WC [Y
1
− Y
2
X
∗
3
x
1
]}
(9.33)
♣
Corollary 9.2 (Partial W-LESS for the scale factor x
1l
). A scalar x
1l
is
partial W-LESS of (9.23) subject to (9.24) if and only if
x
1l
=
trY
∗
1
C
∗
WCY
2
X
∗
3
trY
∗
2
C
∗
WCY
2
(9.34)
holds. For special case W = I
n
the scale parameter vector x
1l
is given by
x
1l
=
trY
∗
1
C
∗
CY
2
X
∗
3
trY
∗
2
C
∗
CY
2
(9.35)
9-4 The General Procrustes solution 125
proof
Partial W-LESS is constructed by the unconstraint centralized Lagrangean
L(x
1
, X
3
) =
=
1
2
tr{[(Y
1
− Y
2
X
∗
3
x
1
)]
∗
C
∗
WC [Y
1
− Y
2
X
∗
3
x
1
]} = min
x
1
,X
3
,
subject to {x
1
≥ 0, X
∗
3
X
3
= I
3
}.
(9.36)
∂L
∂x
1
(x
1l
) = x
1l
trX
3
Y
∗
2
C
∗
WCY
2
X
∗
3
− trY
∗
1
C
∗
WCY
2
X
∗
3
= 0
(9.37)
constitutes the second necessary condition. Due to (e.g., cyclic property in Table
9.1, p. 117)
trX
3
Y
∗
2
C
∗
WCY
2
X
∗
3
= trY
∗
2
C
∗
WCY
2
X
∗
3
X
3
= Y
∗
2
C
∗
WCY
2
,
(9.37) leads to (9.34). While the forward computation of
∂L
∂x
1
(x
1l
) = 0 enjoyed a
representation of the optimal scale parameter x
1l
, its backward substitution into
the Lagrangean L(x
1
, X
3
) amounts to
L(X
3
) =
tr{
Y
1
− Y
2
X
∗
3
trY
∗
1
C
∗
WCY
2
X
∗
3
trY
∗
2
C
∗
WCY
2
C
∗
WC
∗ ∗
Y
1
− Y
2
X
∗
3
trY
∗
1
C
∗
WCY
2
X
∗
3
trY
∗
2
C
∗
WCY
2
}
(9.38)
L(X
3
) =
1
2
tr{(Y
∗
1
C
∗
WCY
1
) −tr(Y
∗
1
C
∗
WCY
2
X
∗
3
)
trY
∗
1
C
∗
WCY
2
X
∗
3
trY
∗
2
C
∗
WCY
2
−tr(X
3
Y
∗
2
C
∗
WCY
1
)
trY
∗
1
C
∗
WCY
2
X
∗
3
trY
∗
2
C
∗
WCY
2
+tr(X
3
Y
∗
2
C
∗
WCY
2
X
∗
3
)
[trY
∗
1
C
∗
WCY
2
X
∗
3
]
2
[trY
∗
2
C
∗
WCY
2
]
2
}
(9.39)
L(X
3
) =
=
1
2
tr(Y
∗
1
C
∗
WCY
1
) −
[trY
∗
1
C
∗
WCY
2
X
∗
3
]
2
[trY
∗
2
C
∗
WCY
2
]
+
1
2
[trY
∗
1
C
∗
WCY
2
X
∗
3
]
2
[trY
∗
2
C
∗
WCY
2
]
(9.40)
126 9 Procrustes solution
L(X
3
) =
=
1
2
tr(Y
∗
1
C
∗
WCY
1
) −
1
2
[trY
∗
1
C
∗
WCY
2
X
∗
3
]
2
[trY
∗
2
C
∗
WCY
2
]
= min,
subject to {X
∗
3
X
3
= I
3
}
(9.41)
♣
Corollary 9.3 (Partial W-LESS for the rotation matrix X
3l
). A 3 × 3 or-
thonormal matrix X
3
is partial W-LESS of (9.41) if and only if
X
3l
= UV
∗
(9.42)
holds, where A := Y
∗
1
C
∗
WCY
2
= UΣ
s
V
∗
is a singular value decomposition
with respect to a left orthonormal matrix U, U
∗
U = I
3
, a right orthonormal
matrix V, VV
∗
= I
3
, and Σ
s
= Diag(σ
1
, σ
2
, σ
3
) a diagonal matrix of sin-
gular values (σ
1
, σ
2
, σ
3
). The singular values are the canonical coordinates of
the right eigenspace (A
∗
A − Σ
2
s
I)V = 0. The left eigenspace is based upon
U = AVΣ
−1
s
.
proof
In (9.41) L(X
3
) subject to X
∗
3
X
3
= I
3
is minimal if
tr(Y
∗
1
C
∗
WCY
2
X
∗
3
) = min, subject to {x
1
≥ 0, X
∗
3
X
3
= I
3
}. (9.43)
Let A := Y
∗
1
C
∗
WCY
2
= UΣ
s
V
∗
be a singular value decomposition with respect
to a left orthonormal matrix U, U
∗
U = I
3
, a right orthonormal matrix V, VV
∗
=
I
3
, and Σ
s
= Diag(σ
1
, σ
2
, σ
3
) a diagonal matrix of singular values (σ
1
, σ
2
, σ
3
).
Then
tr(AX
∗
3
) = tr(UΣ
s
V
∗
X
∗
3
)
= tr(Σ
s
V
∗
X
∗
3
U) =
P
3
i=1
σ
i
r
ii
≤
P
3
i=1
σ
i
,
(9.44)
holds since R = V
∗
X
∗
3
U = [r
ij
] ∈ R
3×3
is orthonormal with | r
ii
|≤ 1. The
identity tr(AX
∗
3
) =
3
P
i=1
σ
i
applies if V
∗
X
∗
3
U = I
3
, that is X
∗
3
= VU
∗
, X
3
= UV
∗
,
namely if tr(AX
∗
3
) is maximal:
"
tr(AX
∗
3
) = max
{X
∗
3
X
3
=I
3
}
⇔
⇔ R = V
∗
X
∗
3
U = I
3
.
trAX
∗
3
=
3
X
i=1
σ.
i
(9.45)
♣
An alternative proof of Corollary 9.3 based on formal differentiation of traces
and determinants has been given in Sect. 9-32 of Sect. 9-3. Finally we collect our
sequential results in Theorem 9.1 identifying the stationary point of W-LESS of
(9.23) specialized for W = I, i.e., matrix of unit weight in Corollary 9.4. The
highlight is the General Procrustes algorithm we have developed in Solution 9.3.