Information Technology Reference
In-Depth Information
x
n
of a real Hilbert space
H
, we know that their linear
combinations with real coefficients provide a closed subspace
K
(a subset of
H
closed under the vector operations) of
H
. Therefore the vector
v
Given
n
vectors
x
1
,
x
2
,...,
∈
H
closest to
K
has to minimize the norm
||
−
−
−...−
||.
v
c
1
x
1
c
2
x
2
c
n
x
n
According to the projection theorem, the unique minimizing vector
x
0
∈
K
has to be
orthogonal to
K
. Therefore, for
i
=
1
,
2
,...,
n
:
((
v
−
c
1
x
1
−
c
2
x
2
−...−
c
n
x
n
)
|
x
i
)=
0
.
This means:
c
1
(
x
1
|
x
1
)+
c
2
(
x
2
|
x
1
)+
...
+
c
n
(
x
n
|
x
1
)=(
v
|
x
1
)
(7.16)
c
1
(
x
2
)
.......................................... ... ......
c
1
(
x
1
|
x
2
)+
c
2
(
x
2
|
x
2
)+
...
+
c
n
(
x
n
|
x
2
)=(
v
|
x
1
|
x
n
)+
c
2
(
x
2
|
x
n
)+
...
+
c
n
(
x
n
|
x
n
)=(
v
|
x
n
)
The matrix
G
which is transpose to the matrix of system 7.16 is called Gram matrix
of the
n
vectors generating the vector space
K
.
Given a set of
m
equations with
n
unknowns, which expresses a linear combina-
tion of
n
linearly independent vectors of
m
for some (unknown) coefficients, there
exists a unique
n
-vector of these coefficients providing the best approximation to a
given
m
-vector
v
. The unicity of this
n
-vector, and its algebraic form, is determined
by the theorem of Table 7.5, as a natural consequence of the projection theorem
for Hilbert spaces, and the solution is expressed in terms of the Gram matrix of the
vectors.
R
Ta b l e 7 . 5
The Least Square Evaluation of matrix
W
Least-Square-Estimate
.Let
W
a
m
×
n
matrix (
m
>
n
) with linearly independent col-
umn vectors and let
Wz
be the matrix product of
W
by
z
∈
R
n
. Then, for any
m
-vector
m
, there exists a unique vector
z
0
∈
R
n
v
∈
R
minimizing
||
v
−
Wz
||
(the Euclidean
m
-
n
. Moreover, if
W
T
dimensional norm) over all
z
∈
R
denotes the transpose of
W
,this
vector
z
0
is given by
z
0
=(
W
T
W
)
−
1
W
T
v
.
The existence and unicity of vector
z
0
, as the Least-Square-Estimate claims (see
7.5), follows directly from the projection theorem. Moreover, the Gram matrix cor-
responding to the column vectors of matrix
W
is easily seen to be
W
T
W
(see left
sides of Eqs. (7.16)), and
W
T
v
corresponds to the right hand vector of equation