Geoscience Reference
In-Depth Information
and
x.k
C
1/
D
x
f
.k/
C
GŒ
.k/
h.x
f
.k//
z
Define an innovation
.k/
h
x
f
.k/
2
R
m
d.k/
D
z
(2.56)
and define
d.1
W
N/
D
d
T
.1/;d
T
.2/;:::;d
T
.N/
T
2
R
Nm
(2.57)
Define
x.0/
x
b
.0/
T
B
1
x.0/
x
b
.0/
J
b
.x.0//
D
1
2
(2.58)
J
n
.x.0/;G/
D
1
2
d
T
.1
W
N/G
T
.P
f
/
1
Gd.1
W
N/
(2.59)
x
b
.0/
Clearly,
J
b
.x.0//
measures the weighted squared distance between
x.0/
and
and
is called the nudging term that measures the weighted square of the
model error term in (
2.55
), and
J
n
.x.0/;G/
Rm
is the model error covariance matrix
computed using the standard method used in the Kalman filter literature (
Lewis
et al.
(
2006
)).
Vidard et al.
(
2003
) then pose the estimation problem as one of minimizing
P
f
2
R
Nm
Q
3
.x.0/;G/
D
J
2
.G/
C
J
b
.x.0//
C
J
n
.x.0/;G/
(2.60)
where
is defined in (
2.22
) and the nudged dynamics in (
2.56
)isusedasa
strong constraint. This minimization is again solved by invoking the standard adjoint
method [
Lewis et al.
(
2006
)].
Following the arguments at the end of Sect.
2.3.1
, it can be readily verified that
the error vector e(1:N) which is a part of
J
2
.G/
J
2
.G/
in (
2.22
) is temporally correlated.
Hence, the correct formulation
J
2
.G/
in (
2.60
) must be replaced by
J
3
.G/
in (
2.33
).
Similarly, it can be verified that the innovation vector
d.1
W
N/
is also temporally
correlated and the
in (
2.60
) by a similar correct form of the functional
that takes the serial correlation of
J
n
.x.0/;G/
W
2
R
Nm
Nm
be the
d.1
W
N/
into account. Let
serial correlation of
d.1
W
N/
. Then define
J
3
.x.0/;G/
D
1
2
d
T
.1
W
N/G
T
W
1
Gd.1
W
N/
(2.61)
Hence the correct formulation is to minimize
Q
3
.x.0/;G/
D
J
3
.G/
C
J
b
.x.0//
C
J
3
.x.0/;G/
(2.62)
W
We leave the computation of the elements of
as an exercise to the reader.
Search WWH ::
Custom Search