The Ensemble Kalman Filter:

Theoretical formulation and practical

implementation
Geir Evensen

Norsk Hydro Research Centre, Bergen, Norway

Based on Evensen 2003, Ocean Dynamics, Vol 53, No 3

– p.1
The Ensemble Kalman Filter (EnKF)
Represents error statistics using an ensemble of model states
Evolves error statistics by ensemble integrations
“Variance minimizing” analysis scheme operating on the
ensemble

Monte Carlo, low rank, error subspace method

Converges to the Kalman Filter with increasing ensemble size
Fully nonlinear error evolution, contrary to EKF
Assumption of Gaussian statistics in analysis scheme

– p.2
 The error covariance matrix
Define ensemble covariances around the ensemble mean























The ensemble mean is the best-guess.


The ensemble spread defines the error variance.
The covariance is determined by the smoothness of
the ensemble members.
A covariance matrix can be represented by an
ensemble of model states (not unique).

– p.3
Dynamical evolution of error statistics
Each ensemble member evolve according to the model
dynamics which is expressed by a stochastic
differential equation











 


The probability density then evolve according to
Kolmogorov’s equation

!







 





"













 


This is the fundamental equation for evolution of error
statistics and can be solved using Monte Carlo
methods.
– p.4
 Analysis scheme (1)
Given an ensemble of model forcasts .





Create an ensemble of observations

%$




"
with
the first guess observations,
#

a vector of observation noise.

$

The measurement error covariance matrix is


&

' &
(


$
$

– p.5
 Analysis scheme (2)
Update each ensemble member according to




*
)







'


where

0


./
)

*
+

-


'

'
'

'
This is equivalent to updating the mean




 #

*
)



'


The posterior error covariance becomes

1






– p.6
Analysis of the Analysis scheme (1)
Define the ensemble matrix

3
7

6
5
2



4


3
0
"

"


"


The ensemble mean is (defining )

3
7
5

:;

8

9
3
2

2
8


 3
The ensemble perturbations becomes
<

3
2

2
1

8






The ensemble covariance matrix becomes

7
5
6

6

4


<

= <

2

2







– p.7
 Analysis equation (2)
Given a vector of measurements , define

>5
#

4
#

;

?
%$






"

"


"

"
stored in

3
7
5>


3
#

4


0
"

"


"

"
The ensemble perturbations are stored in

3
7


5>
@

4
$

" $

" $


"



"
thus, the measurement error covariance matrix
becomes

@

@
&





;


– p.8
 Analysis equation (3)
The analysis equation can now be written

./

2
2

&









Defining the innovations and using

<

2



previous definitions:

0
.



<

<

<

<

<

2

@






i.e., analysis expressed entirely in terms of the
ensemble

– p.9
Practical computation of analysis (1)
Use pseudo inverse


<

<
2

B
@

A


.C 


0
0
<

<


B
@

A
.



Computational cost is:
to form ,


<

<
2

2
;
D

for eigenvalue decomposition.



Unafordable for large !

D

– p.10
Practical computation of analysis (2)
Note that , thus


<
2

F
9



<

<

<

<


@


@
2

2
@

@






Compute SVD, , giving


<
2

GH

I



<

<
2

GH

IH

GH
H

G



Computational cost is for SVD.

D ;

– p.11
Practical computation of analysis (3)
The analysis equation can now be written


0

<

<

<



2

B
G

G
.




The computation goes as follows with

;
K
J


3
0

>7
5
B

G
.

D
" J

0

3
<

7
5

D ;
4

J


"
3
7
5>
G

D ;
4

J

L

"
3

3

<

7


5
2

D ;

;
4

M

"
3
<

6
5
2

N ;

;


4


"
All computations reduced to .
;

D ;
D

– p.12
Practical computation of analysis (4)
The final update can be written as

2
2

2



M


3
2

2
1

8



M

M
2
1




2


O
"
thus, the analysis is a “weakly nonlinear combination”
of the forecast ensemble.
Note also



;


2




L


2
9

– p.13
Remarks on the analysis equation (1)
Covariances only needed between observed variables
at measurement locations.
Covariances never computed but indirectly used to
determine .




Analysis may be interpreted as:
combination of ensemble members, or,
forecast pluss combination of covariance functions.
Covariances only needed to compute .

O
Accuracy of analysis is determined by
the accuracy of
O

the properties of the ensemble error space

– p.14
Remarks on the analysis equation (2)
For a linear model, any choice of will result in an

O
analysis which is also a solution of the model.
Filtering of covariance functions introduces
nondynamical modes in the analysis.

– p.15
 Local analysis
A local analysis is computed grid point by grid point
using only nearby measurements.
Introduces nondynamical modes in the analysis.
Different for each grid point.
O
Allows us to reach a larger class of solutions.

– p.16
 Nonlinear measurements
Measurement equation

P


$



Define ensemble of model prediction of the
measurements
2 Q

3
7


5>
P


P


4


3
0
"


"

"


The analysis then becomes

.
<

<

<
2 Q

2 Q

2 Q

2 Q

<



2

@





"
Analysis based on covariances between and .

P



– p.17
Ensemble Kalman Smoother (EnKS)
Starts with EnKF solution.
Computes updates backward in time;
sequentially for each measurement time,
using covariances in time,
no backward integrations.
The analysis becomes for :

S <
T
K

K


U 
V



.

U
ZYX
W 


<


<



2

2


ZYW

[]

O
\
[

^ 
_

– p.18
Some recent applications of the EnKF
Haugan and Evensen (2002), Ocean Dynamics.
Mitchell et al. (2002), MWR.
Brusdal et al (2003), JMS.
Natvik and Evensen (2003a,b), JMS.
Keppenne and Rienecker (2003), JMS.
DIADEM project
TOPAZ project (topaz.nersc.no)
MERSEA project

– p.19
 Time correlated model noise
Scalar model

g
 
`

Ua
. `

a
f
U

U
0

0
.




dc b


h


Ue
`
U

0
.

– p.20
 Results ( )

i
4

-1

-2

-3

-4
0 1 2 3 4 5 6 7 8 9 10
Time
– p.21
 Results ( )

j
4

-1

-2

-3

-4
0 1 2 3 4 5 6 7 8 9 10
Time
– p.22
 Estimate of model noise, EnKF
2
4

2
1

-1

-2
-1

-3

-4
-2
0 1 2 3 4 5 6 7 8 9 10
Time
– p.23
 Estimate of model noise, EnKS
2

-1

-2
0 1 2 3 4 5 6 7 8 9 10
Time
– p.24
Model forced by estimated model error

dc b


` k


Ue

U

0
.
k
l 
m

lm


– p.25
 o o
p

r t r s rq

u u
p

v
opo

x r t
y

{
}~|

{
r s x r s
y w
x rq
y
 €
z
z

{
€ ‚

ƒ
„
rq

u u
p
Introduces poorly known parameter

U
{ n
o o
p

† … ‚
‡

Š Š w
‰~ˆ
in model

x r
y
Parameter and bias estimation

‹p
u u
– p.26
Estimate and model error, EnKF
4
5

3
4

2
3

1
2

1
-1

0
-2

-1
-3

-4
-2
0 10 20 30 40 50
Time
– p.27
Estimate and model error, EnKS
5

-1

-2
0 10 20 30 40 50
Time
– p.28
Estimate, model error and bias, EnKF
4
5

3
4

2
3

1
2

1
-1

0
-2

-1
-3

-4
-2
0 10 20 30 40 50
Time
– p.29
Estimate, model error and bias, EnKS
5

-1

-2
0 10 20 30 40 50
Time
– p.30
 Estimated bias and std dev
2

-1

-2
0 10 20 30 40 50
Time
– p.31