Geoscience Reference
In-Depth Information
Fig. 10.6
P
p
)of
the
posterior
error to the
prior
projected in data-space.
The overfit (
black
) method
generates a higher error
estimate compared to the
constrained (
grey
) method
across all choices of
background and observational
errors as well as over varying
numbers of observations
Ratio (
P
a
=
3 Obs
12 Obs
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
P
Fig. 10.7
Comparison of
prior
R
and
posterior
R
a
observational error for overfit
(
black
) and constrained (
grey
)
methods
3
2.5
2
1.5
1
0.5
0
0
0.5
1
1.5
2
2.5
3
R
the observations. The constrained method slightly underestimates the observational
error, but is consistent with the
prior
. This exposes the overfit method for placing
too much emphasis on the observations.
As a final comparison, the theoretical cost-function minimum is compared to the
minimized cost-function at the end of each outer-loop. Because the overfit method
is unconstrained by
P
after the first outer-loop, it was shown that it will minimize
J
o
at the expense of
J
b
.
Bennett
(
2002
) showed that for correctly specified
P
and
R
, the minimum value
of the cost-function is
. This measure has been used as a useful
diagnostic by
Weaver et al.
(
2003
)and
Powell et al.
(
2008
). Unfortunately, this
measure does not quantify the contribution of each component of the cost function.
Moore et al.
(
2011a
) provides a concise review of the work of
Talag r an d
(
1999
),
Chapnik et al.
(
2006
), and
Desroziers et al.
(
2009
) along with the derivation for
determining the minimum theoretical values of each cost-function component. The
J
min
D
N
obs
=2
Search WWH ::
Custom Search