page 1 of 26 a pv control variable ross bannister* mike cullen *data assimilation research centre,...
TRANSCRIPT
Page 1 of 26
A ‘PV’ control variable
Ross Bannister*
Mike Cullen†
*Data Assimilation Research Centre, Univ. Reading, UK
†Met Office, Exeter, UK
Page 2 of 26
Definition of the problem
A variational assimilation system needs a background state, and a PDF that described its uncertainty. Errors in are assumed to be:
In MetO, ECMWF, NCEP, etc, is implied by a model:
Bx
Bx
B
Unbiased Normally distributed Correlated
Define a change of variables Impose that elements of are mutually uncorrelated and have unit variance
vxx B
)()(expPDF 121
BT
B xxxx(x) B
vv(v) T21expPDF
v
Page 3 of 26
Questions
What is the best transformation we can conceive for a climatological ? Are errors in the components of uncorrelated? Is the transformation practical? Is it invertible? Will it give realistic implied covariances . Will it give an appropriately ‘balanced’ analysis?
vB
TUUB
'vxx B U
)(' Bxxv T
Page 4 of 26
What is required from the PV-control variable project?
1) A control variable ‘parameter transform’ (for use in the inner loop).
2) The transpose of (for use in the inner loop gradient calculation).
3) The inverse of , (for the calculation of statistics).B
'
'
v
vxx B
hvp UUU
U
pU
pU pT
Doesn’t the current parameter transform already do this?
Page 5 of 26
Current parameter transform
• LBE = Linear Balance Eq (rotational wind to ‘balanced’ pressure).• F = vertical regression operator (for vertical consistency).
streamfunction
unbalanced pressure
velocity potential
Potential temperature, density, specific humidity and vertical velocity increments follow diagnostically.
'
'
'
0)LBE(
/0/
/0/
'
'
'
pyx
xy
p
v
uA
IF
Page 6 of 26
How can we improve on this?
It is assumed that B is univariate in (ignoring moisture for now)
','', pA
A better choice of parameters for use in the assimilation:
the slow, ‘balanced’ part of the flow (1 parameter),the unbalanced part (2 parameters)
Page 7 of 26
Shallow water result
represents the balanced part at small horizontal scales only We require a variable that is ‘balanced’ at all scales
Page 8 of 26
The ‘PV’ formulation
A. Define alternative parameters.B. Formulate the U-transform.C. Formulate the T-transform.D. Other technical information.E. Tests.F. Achievements and problems.G. References.
Page 9 of 26
A. Three new parameters
1 Describes the ‘balanced’ component of the rotational flow
2 Describes the ‘unbalanced’ component of the rotational flow
3 Describes the divergent component of the flow
's
'PV
'pu
'PV
'
'u
Page 10 of 26
B. Formulation of U-trans … 1
Definition of variables Model perturbations Control parameters
Associated parameters Transforms
'
'
'
'
'
'
'
'
'
'
'
'
'
p
v
u
p
v
u
p
v
u
p
v
u
x
up
up
up
s
s
s
'
'
'
'
p
s
v U
'
'
'
'
u
y PV
PV
''
''
''
xv
xy
vx
T
A
U
Page 11 of 26
B. Formulation of U-trans ... 2
'
'
')(
'
'
'
'
'
'
333231
232221
131211
p
s
p
s
p
v
uU
upsU
UUU
UUU
UUU
UUU
What are the column vectors ?UUU ,, ups
U-transform
A-transform
'
'
'
'
'
'
'
'
'
*
*
*
333231
232221
131211
p
v
u
p
v
u
u u
PV
PV
A
A
A
AAA
AAA
AAA
PV
PV
A is a known linear operator (later)
Page 12 of 26
B. Formulation of U-trans … 3
Design strategy & definition of anti-PV1. The ‘balanced’ transform Choose the ‘balanced’ set of increments to satisfy LBE=0.
)'(
/'
/'
'
'
'
'
02 sf
xs
ys
s
p
v
u
s
s
s
s
U
0')'(
0LBE from increments '2
0
s
s
psf
p
0'PV
')'(' Define 20 pf PV
Page 13 of 26
B. Formulation of U-trans … 4
2. The unbalanced component of the vortical flow Choose the ‘unbalanced’ set of increments to satisfy PV’=0.
'
'
'
'
'
'
'
p
p
p
p
p
v
u
U
U
U
Uup
up
up
up
S
R
UR and S are complicated operators giving winds that have zero linearized PV’
0)',','('
'''')',','('
2
2
upupup pvuz
pzp
ppvu
PV
PV 1. Calculate from2. Convert to3. Compute
up'upup '' 2 'pU
upup vu ','
Page 14 of 26
B. Formulation of U-trans … 5
3. The divergent component of the flow The divergent component automatically has no PV or anti-PV.
0
/
/
'
'
'
y
x
p
v
u
U
Page 15 of 26
B. Summary of U-transform
'
'
')(
'
'
'
p
s
p
v
uU
ups UUUU-transform
A-transform
'
'
'
'
'
'
*
*
*
p
v
u
u u
PV
PV
A
A
A
PV
PV
• Zero anti-PV• Zero Divergence
• Zero PV• Zero Divergence
• Zero PV• Zero anti-PV
0'
0'
*
*
s
s
su
sPV
UA
UA
0'
0'
*
*
p
p
Uupu
UupPV
UA
UA
0'
0'
*
*
UA
UA
PV
PV
Page 16 of 26
B. ‘Footnote’ to the U - transformpU
Recall the linearized PV formula:
Problem: This cannot be computed at the top and bottom boundaries.Solution: Avoid computing at top and bottom Compute PV’ of first two vertical modes instead.
2
2 '''')',','('
z
p
z
pppvu
PV
''dd''ddd'
'd'd '
0 0'0 0'
'
0
'2
0
'
01
top
z
z
z
top
z
z
zzp
top
zzp
top
zzp
top
z
zzpzzfzf
pzfz
PV
PV like PV of external mode
like PV of 1st internal mode
Page 17 of 26
C. Formulation of T-transform
''
''
''
''
vx
xv
xy
vx
AUA
T
A
U
Recall
Until now, is given, what is ? 'v 'x
'
'
'
'
'
'
***
***
***
p
s
u
U
uupusu
PVupPVsPV
PVupPVsPV
UAUAUA
UAUAUA
UAUAUA
PV
PV
For calibration of B, ask: is given what is ?'v'x
Page 18 of 26
D. Other technical information
Grid positions
The reference state
's
,'pU
'PV
' ,' uPV
Zonal mean reference state
Page 19 of 26
E. Tests
1. What do PV’, anti-PV’ and divergence’ look like?
2. Linearity test for PV - is the linear approximation reasonable?
3. Vertical mode test 1 – are the two vertical modes independent?
4. Vertical mode test 2 – are they PV-like?
5. Adjoint test for U-transform – is the adjoint code correct?
6. ‘Cog’ test of U-transform – is information carried through the
assimilation system with the new transform?
7. Inverse test – is the inverse transform valid?
Page 20 of 26
E.1 PV’, anti-PV’, divergence
PV’ anti-PV’
divergence’
All level 17 (~5km)
Page 21 of 26
E.2 Linearity of PV’
)(12 1)(' LSZMLSLSPV )()( 22 LSPVLSPV
level 17
Page 22 of 26
E.3 Are the ‘extra PV’ modes independent?
PV1 PV2
''dd''ddd'
'd'd '
0 0'0 0'
'
0
'2
0
'
01
top
z
z
z
top
z
z
zzp
top
zzp
top
zzp
top
z
zzpzzfzf
pzfz
PV
PV''d
'd'
''d
0'
0'
1
0'
topz
z
z
z
z
zz
z
z
PV
Page 23 of 26
E.4 Are the modes PV-like?
nnsnn pcPV F][ 2 PV of vertical mode n (spectral space)
PV1 PV2
Small scales only
External mode 1st internal mode
Page 24 of 26
Test with conversion to and from ‘adjoint’ variables for in
Test by bypassing in transforms
2 TU
2
E.5 Adjoint test
vvvv T ||?
UUUU
27
27
10...163153.89RHS
10...611309.02LHS
41757611969.97543RHS
71757611969.97543LHS
Page 25 of 26
F. Summary, achievements, problems, what next?
The current choice of control parameters is
A better choice of control parameters is expected to beStarted to implement the new PV-based scheme
Current problems
Next stage
',',' pA
• Expected to be strong correlations between their errors at large horizontal length scales
',',' ps U
• Handing of inverse Laplacian in adjoint
• ‘Cog’ test in preparation• Tp-transform
• Forms of Up transforms (+complications)• Adjoint code• Strategy for Tp-transform
Page 26 of 26
G. References
This talk and other documents at “file:///home/mm0200/frxb/public_html/PVcv/PVcv.html” on intranet.
Cullen M.J.P., 4d Var: A new formulation based on a PV representation, QJRMS 129, pp. 2777-2796 (2003).