Geoscience Reference
In-Depth Information
2 Variational Methods
Successful and efficient minimization of the cost function ((
19.1
) from main text)
requires the so-called Hessian preconditioning (e.g.,
Axelsson and Barker 1984
;
Zupanski 1993
,
1995
). The square-root forecast error covariance is commonly used
for this purpose in variational methods, introduced as a change of variable
D
P
1=2
f
x
x
f
w
(19.26)
Where
w
is the preconditioned control variable. The iterative minimization solution
w
a
is obtained as a limit of sequence
f
w
k
D
w
k
1
C
˛
k
1
d
k
1
I
k
D
1;2;:::
g
where index
is descent direction.
After substituting the minimization solution
w
a
in (
19.26
) one obtains the analysis
solution in terms of the physical state variable
k
is the iteration index,
˛
is step-size, and
d
x
a
x
f
D
P
1=2
f
w
a
:
(19.27)
After substituting (
19.21
)in(
19.27
), and denoting
i
D
i
v
i
w
a
(19.28)
The variational method solution (
19.27
) becomes
D
X
i
x
a
x
f
i
u
i
:
(19.29)
Therefore, the variational solution can also be represented as a linear combination
of forecast error covariance singular vectors.
Since majority of currently used data assimilation algorithms are based on
KF and/or variational methods, one can see from (
19.25
)to(
19.29
) that analysis
correction
x
a
x
f
lies in the space defined by the forecast error covariance singular
vectors.
Appendix 2
Entropy and Mutual Information
We follow
Cover and Thomas
(
2006
) to quantify the information content of
observations, based on Shannon information theory (
Shannon and Weaver 1949
)
and relative entropy (
Kullback and Leibler 1951
). Entropy of a random variable
X
is defined as a non-negative measure of uncertainty
Search WWH ::
Custom Search