Optimized localization and hybridization to filter ensemble-based covariances Benjamin Ménétrier and Tom Auligné NCAR - Boulder - Colorado NASA GSFC - 07/21/2015
Acknowledgement: AFWA
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context:
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances.
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts.
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size.
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size. • Systematic error arises because of ensemble misspecifications.
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size. • Systematic error arises because of ensemble misspecifications. • Question: how to tackle both errors?
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size. • Systematic error arises because of ensemble misspecifications. • Question: how to tackle both errors?
Usual methods:
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size. • Systematic error arises because of ensemble misspecifications. • Question: how to tackle both errors?
Usual methods: • Covariance localization
→ tapering with a localization matrix
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Context: • DA often relies on forecast error covariances. • This matrix can be sampled from an ensemble of forecasts. • Sampling noise arises because of the limited ensemble size. • Systematic error arises because of ensemble misspecifications. • Question: how to tackle both errors?
Usual methods: • Covariance localization
→ tapering with a localization matrix • Covariance hybridization
→ linear combination with a static covariance matrix
1
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions:
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method:
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method: •
using data from the ensemble only,
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method: •
using data from the ensemble only,
•
affordable for high-dimensional systems.
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method: •
using data from the ensemble only,
•
affordable for high-dimensional systems.
2. Can localization and hybridization be considered together, and optimized simultaneously?
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method: •
using data from the ensemble only,
•
affordable for high-dimensional systems.
2. Can localization and hybridization be considered together, and optimized simultaneously? 3. Is hybridization always improving the accuracy of forecast error covariances?
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Introduction Questions: 1. Can we compute an optimized localization with a method: •
using data from the ensemble only,
•
affordable for high-dimensional systems.
2. Can localization and hybridization be considered together, and optimized simultaneously? 3. Is hybridization always improving the accuracy of forecast error covariances? 4. Can ensemble systematic error be taken into account?
2
Introduction
Localization
Hybridization
Systematic error
Conclusions
Outline
Introduction Objectively optimized localization Jointly optimized localization and hybridization Accounting for systematic error Conclusions
3
Introduction
Localization
Hybridization
Systematic error
Conclusions
Outline
Introduction Objectively optimized localization Jointly optimized localization and hybridization Accounting for systematic error Conclusions
4
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance sampling
5
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance sampling e An ensemble of N forecasts {e xbp } is used to sample B: e= B
N T 1 δe xb δe xb ∑ N − 1 p=1
where: δe xbp = e xbp − he xb i and he xb i =
1 N
N
∑ exbp
p=1
5
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance sampling e An ensemble of N forecasts {e xbp } is used to sample B: e= B
N T 1 δe xb δe xb ∑ N − 1 p=1
where: δe xbp = e xbp − he xb i and he xb i =
1 N
N
∑ exbp
p=1
e →B e? Asymptotic behavior: if N → ∞ , then B
5
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance sampling e An ensemble of N forecasts {e xbp } is used to sample B: e= B
N T 1 δe xb δe xb ∑ N − 1 p=1
where: δe xbp = e xbp − he xb i and he xb i =
1 N
N
∑ exbp
p=1
e →B e? Asymptotic behavior: if N → ∞ , then B ee = B e −B e? In practice, N < ∞ ⇒ sampling noise B
5
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: homogeneous covariances
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: heterogeneous variances
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: heterogeneous variances
Sampling noise amplitude related to the asymptotic variance
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: homogeneous covariances
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: heterogeneous length-scales
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Sampling noise properties Simple 1D example: heterogeneous length-scales
Sampling noise length-scale related to the asymptotic length-scale
6
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory of sampling noise Covariance of sampled covariances: ? 1 ? ? e e ?, B e? + 1 E Ξ e e ,B e = Cov B e B − E B Cov B ij kl ij kl ijkl ij kl N N ? ? 1 e e B e? + E B e? B + E B il jk ik jl N(N − 1)
7
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory of sampling noise Covariance of sampled covariances: ? 1 ? ? e e ?, B e? + 1 E Ξ e e ,B e = Cov B e B − E B Cov B ij kl ij kl ijkl ij kl N N ? ? 1 e e B e? + E B e? B + E B il jk ik jl N(N − 1) involving: • the ensemble size N, e ?, • the asymptotic covariance B e ?. • the asymptotic fourth-order centered moment Ξ
7
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory of sampling noise Covariance of sampled covariances: ? 1 ? ? e e ?, B e? + 1 E Ξ e e ,B e = Cov B e B − E B Cov B ij kl ij kl ijkl ij kl N N ? ? 1 e e B e? + E B e? B + E B il jk ik jl N(N − 1) involving: • the ensemble size N, e ?, • the asymptotic covariance B e ?. • the asymptotic fourth-order centered moment Ξ e : Expectation of the sample fourth-order centered moment Ξ ijkl (N − 1)(N 2 − 3N + 3) ? e e E Ξ = E Ξ ijkl ijkl N3 ? ? ? ? (N − 1)(2N − 3) e ?B e? + E B e B e +E B e B e + E B ij kl ik jl il jk N3
7
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix:
b = L◦B e B
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Covariance localization Localization = Schur product with a localization matrix L • Covariance matrix: • Increment:
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
8
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
Light assumptions:
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
Light assumptions: ee = B e −B e ? is not correlated • The unbiased sampling noise B e ?. with the asymptotic sample covariance matrix B
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
Light assumptions: ee = B e −B e ? is not correlated • The unbiased sampling noise B e ?. with the asymptotic sample covariance matrix B
e ? and • The two random processes generating the asymptotic B the sample distribution are independent.
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
Light assumptions: ee = B e −B e ? is not correlated • The unbiased sampling noise B e ?. with the asymptotic sample covariance matrix B
e ? and • The two random processes generating the asymptotic B the sample distribution are independent. Use of the covariance sampling theory... and a lot of calculus!
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization An "optimal" localization should minimize the expected quadratic error: i h 2 e? e − B k e=E k L ◦ B |{z} | {z } e Localized B
e Asymptotic B
Light assumptions: ee = B e −B e ? is not correlated • The unbiased sampling noise B e ?. with the asymptotic sample covariance matrix B
e ? and • The two random processes generating the asymptotic B the sample distribution are independent. Use of the covariance sampling theory... and a lot of calculus! e e B e E Ξ E B (N − 1)2 N N −1 ijij ii jj Lij = − 2 + 2 e e N(N − 3) (N − 2)(N − 3) E B N(N − 2)(N − 3) E B ij
ij
9
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization This formula of optimal localization L involves: • the ensemble size N,
e • the sample covariance B, e • the sample fourth-order centered moment Ξ.
10
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization This formula of optimal localization L involves: • the ensemble size N,
e • the sample covariance B, e • the sample fourth-order centered moment Ξ. An ergodicity assumption is required to estimate the statistical expectations E in practice:
10
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization This formula of optimal localization L involves: • the ensemble size N,
e • the sample covariance B, e • the sample fourth-order centered moment Ξ. An ergodicity assumption is required to estimate the statistical expectations E in practice: • whole domain average, • local average, • scale dependent average,
10
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization This formula of optimal localization L involves: • the ensemble size N,
e • the sample covariance B, e • the sample fourth-order centered moment Ξ. An ergodicity assumption is required to estimate the statistical expectations E in practice: • whole domain average, • local average, • scale dependent average,
→ This assumption is independent from the basic theory.
10
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization optimization This formula of optimal localization L involves: • the ensemble size N,
e • the sample covariance B, e • the sample fourth-order centered moment Ξ. An ergodicity assumption is required to estimate the statistical expectations E in practice: • whole domain average, • local average, • scale dependent average,
→ This assumption is independent from the basic theory. Theory and results published recently in Ménétrier et al. 2015 (Monthly Weather Review).
10
Introduction
Localization
Hybridization
Systematic error
Conclusions
Illustration of localization Data: • Ensemble: mature WRF-ARW EnKF over the CONUS domain
11
Introduction
Localization
Hybridization
Systematic error
Conclusions
Illustration of localization Data: • Ensemble: mature WRF-ARW EnKF over the CONUS domain • Field: temperature at level 10 (∼ 1 km height)
11
Introduction
Localization
Hybridization
Systematic error
Conclusions
Illustration of localization Data: • Ensemble: mature WRF-ARW EnKF over the CONUS domain • Field: temperature at level 10 (∼ 1 km height)
11
Introduction
Localization
Hybridization
Systematic error
Conclusions
Illustration of localization Data: • Ensemble: mature WRF-ARW EnKF over the CONUS domain • Field: temperature at level 10 (∼ 1 km height)
11
Introduction
Localization
Hybridization
Systematic error
Conclusions
Outline
Introduction Objectively optimized localization Jointly optimized localization and hybridization Accounting for systematic error Conclusions
12
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix b = L◦B e B
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix
Increment
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix
Increment
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
Localization by L + hybridization with B Increment e
δ x = β δ xe + β c B1/2 vc
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix
Increment
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
Localization by L + hybridization with B Covariance matrix bh
e 2
c 2
e + (β ) B B = (β ) L ◦ B
Increment e
δ x = β δ xe + β c B1/2 vc
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix
Increment
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
Localization by L + hybridization with B Covariance matrix bh
e 2
c 2
Gain Lh
Offset
e + (β ) B B = (β ) L ◦ B | {z } | {z }
Increment e
δ x = β δ xe + β c B1/2 vc
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
From localization to hybridization Localization by L (Schur product) Covariance matrix
Increment
b = L◦B e B
N 1 δ xe = √ δe xbp ◦ L1/2 vpα ∑ N − 1 p=1
Localization by L + hybridization with B Covariance matrix bh
e 2
c 2
Gain Lh
Offset
e + (β ) B B = (β ) L ◦ B | {z } | {z }
Increment e
δ x = β δ xe + β c B1/2 vc
e Localization + hybridization = linear filtering of B Lh and β c have to be optimized together
13
Introduction
Localization
Hybridization
Systematic error
Conclusions
Joint optimization of localization and hybridization An "optimal" hybridization should minimize the expected quadratic error: e + (β c )2 B − B e? e h = E k Lh ◦ B k2 | {z } |{z} e Localized / hybridized B
e Asymptotic B
14
Introduction
Localization
Hybridization
Systematic error
Conclusions
Joint optimization of localization and hybridization An "optimal" hybridization should minimize the expected quadratic error: e + (β c )2 B − B e? e h = E k Lh ◦ B k2 | {z } |{z} e Localized / hybridized B
e Asymptotic B
Same assumptions as before.
14
Introduction
Localization
Hybridization
Systematic error
Conclusions
Joint optimization of localization and hybridization An "optimal" hybridization should minimize the expected quadratic error: e + (β c )2 B − B e? e h = E k Lh ◦ B k2 | {z } |{z} e Localized / hybridized B
e Asymptotic B
Same assumptions as before. Result of the minimization: a linear system in Lh and (β c )2 e E B ij h Lij = Lij − 2 B ij (β c )2 e E Bij h e B 1 − L E Bij ∑ ij ij ij (β c )2 = 2 ∑ij B ij where Lij is the localization optimized alone.
14
Introduction
Localization
Hybridization
Systematic error
Conclusions
Hybridization benefits Comparison of: b = L ◦ B, e with an optimal L minimizing e • B b h = Lh ◦ B e + (β c )2 B, with optimal Lh and β c minimizing e h • B
15
Introduction
Localization
Hybridization
Systematic error
Conclusions
Hybridization benefits Comparison of: b = L ◦ B, e with an optimal L minimizing e • B b h = Lh ◦ B e + (β c )2 B, with optimal Lh and β c minimizing e h • B We can show that: e B 2ij Var B ij e h − e = −(β c )2 ∑ 2 e E B ij ij | {z } ≤0
15
Introduction
Localization
Hybridization
Systematic error
Conclusions
Hybridization benefits Comparison of: b = L ◦ B, e with an optimal L minimizing e • B b h = Lh ◦ B e + (β c )2 B, with optimal Lh and β c minimizing e h • B We can show that: e B 2ij Var B ij e h − e = −(β c )2 ∑ 2 e E B ij ij | {z } ≤0
With optimal parameters, whatever the static B: Localization + hybridization is better than localization alone
15
Introduction
Localization
Hybridization
Systematic error
Conclusions
Experimental setup • WRF-ARW model, large domain, 25 km-resolution, 40 levels
16
Introduction
Localization
Hybridization
Systematic error
Conclusions
Experimental setup • WRF-ARW model, large domain, 25 km-resolution, 40 levels • Initial conditions randomized from a homogeneous static B
16
Introduction
Localization
Hybridization
Systematic error
Conclusions
Experimental setup • WRF-ARW model, large domain, 25 km-resolution, 40 levels • Initial conditions randomized from a homogeneous static B • Reference and test ensembles (1000 / 100 members)
16
Introduction
Localization
Hybridization
Systematic error
Conclusions
Experimental setup • WRF-ARW model, large domain, 25 km-resolution, 40 levels • Initial conditions randomized from a homogeneous static B • Reference and test ensembles (1000 / 100 members) • Forecast ranges: 12, 24, 36 and 48 h
16
Introduction
Localization
Hybridization
Systematic error
Conclusions
Experimental setup • WRF-ARW model, large domain, 25 km-resolution, 40 levels • Initial conditions randomized from a homogeneous static B • Reference and test ensembles (1000 / 100 members) • Forecast ranges: 12, 24, 36 and 48 h Temperature at level 7 (∼ 1 km above ground), 48 h-range forecasts
Standard-deviation (K)
Correlations functions
16
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level.
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Localization length-scale:
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Hybridization coefficients for zonal wind:
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Impact of the hybridization:
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Impact of the hybridization: e ? is estimated with the reference ensemble • B
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Impact of the hybridization: e ? is estimated with the reference ensemble • B
• Expected quadratic errors e and e h are computed
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Impact of the hybridization: e ? is estimated with the reference ensemble • B
• Expected quadratic errors e and e h are computed
Error reduction from e to e h for 25 members Zonal wind 4.5 %
Meridian wind 4.2 %
Temperature 3.9 %
Specific humidity 1.7 %
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Localization and hybridization • Optimization of the horizontal localization Lh and of the hor
hybridization coefficient β c at each vertical level. e • Static B = horizontal average of B • Impact of the hybridization: e ? is estimated with the reference ensemble • B
• Expected quadratic errors e and e h are computed
Error reduction from e to e h for 25 members Zonal wind 4.5 %
Meridian wind 4.2 %
Temperature 3.9 %
Specific humidity 1.7 %
→ Hybridization with B improves the accuracy of the forecast error covariance matrix
17
Introduction
Localization
Hybridization
Systematic error
Conclusions
Outline
Introduction Objectively optimized localization Jointly optimized localization and hybridization Accounting for systematic error Conclusions
18
Introduction
Localization
Hybridization
Systematic error
Conclusions
An imperfect ensemble Sample covariance matrix decomposition: e= e? e B e? B B + B |{z} | − {z } e Asymptotic B
ee Sampling error B
19
Introduction
Localization
Hybridization
Systematic error
Conclusions
An imperfect ensemble Sample covariance matrix decomposition: e= e? e B e? B B + B |{z} | − {z } e Asymptotic B
= |{z} Bt + "Truth"
ee Sampling error B
e ? − Bt B | {z }
+ sys
e Systematic error B
e B e? B | − {z }
e
e Sampling error B
19
Introduction
Localization
Hybridization
Systematic error
Conclusions
An imperfect ensemble Sample covariance matrix decomposition: e= e? e B e? B B + B |{z} | − {z } e Asymptotic B
= |{z} Bt + "Truth"
ee Sampling error B
e ? − Bt B | {z }
+ sys
e Systematic error B
e B e? B | − {z }
e
e Sampling error B
Systematic error is coming from ensemble misspecifications: • oversimplified observation and model error models, • missrepresentation of some uncertainty sources (ex: SST), • inconsistencies of the ensemble update scheme, • ...
19
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory extension Expected quadratic error to minimize: h i bh − B e ? k2 e h = E kB
20
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory extension Expected quadratic error to minimize: h i h i bh − B e ? k2 b h − Bt k 2 e h = E kB → e t = E kB
20
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory extension Expected quadratic error to minimize: h i h i bh − B e ? k2 b h − Bt k 2 e h = E kB → e t = E kB Linear system to solve: e E B ij = Lij − 2 B ij (β c )2 e E Bij e B ij 1 − Lhij E B ∑ ij ij c 2 (β ) = 2 ∑ij B ij Lhij
20
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory extension Expected quadratic error to minimize: h i h i bh − B e ? k2 b h − Bt k 2 e h = E kB → e t = E kB Linear system to solve: e sys e e B E B E B ij ij ij c 2 = Lij − 2 B ij (β ) − 2 e e E B E B ij ij sys h e e B ij 1 − Lij E Bij B ij E B ∑ ∑ ij ij ij c 2 (β ) = − 2 2 ∑ij B ij ∑ij B ij Lhij
20
Introduction
Localization
Hybridization
Systematic error
Conclusions
Theory extension Expected quadratic error to minimize: h i h i bh − B e ? k2 b h − Bt k 2 e h = E kB → e t = E kB Linear system to solve: e sys e e B E B E B ij ij ij c 2 = Lij − 2 B ij (β ) − 2 e e E B E B ij ij sys h e e B ij 1 − Lij E Bij B ij E B ∑ ∑ ij ij ij c 2 (β ) = − 2 2 ∑ij B ij ∑ij B ij Lhij
sys e B e sys and E B e ... We just have to precompute E B ij ij ij
20
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem:
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem: • Sampling noise had know statistical properties...
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem: • Sampling noise had know statistical properties... • But the truth Bt is unknown, and so is the systematic error!
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem: • Sampling noise had know statistical properties... • But the truth Bt is unknown, and so is the systematic error! • However, a "target" Bt can be modeled.
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem: • Sampling noise had know statistical properties... • But the truth Bt is unknown, and so is the systematic error! • However, a "target" Bt can be modeled.
Consider multiplicative inflation, dealing with ensemble spread issue only: δe xbk ← α ◦ δe xbk where α is a vectorial inflation factor.
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Systematic error modeling There is a problem: • Sampling noise had know statistical properties... • But the truth Bt is unknown, and so is the systematic error! • However, a "target" Bt can be modeled.
Consider multiplicative inflation, dealing with ensemble spread issue only: δe xbk ← α ◦ δe xbk where α is a vectorial inflation factor. This model relies on the implicit assumption that: e? Bt ' α α B ij
i
j
ij
21
Introduction
Localization
Hybridization
Systematic error
Conclusions
Model of systematic error In this case, we can obtain: sys e e B = (1 − α α )E E B ij ij i j 2 sys e e e E Bij Bij = (1 − αi αj )Lij E B ij
22
Introduction
Localization
Hybridization
Systematic error
Conclusions
Model of systematic error In this case, we can obtain: sys e e B = (1 − α α )E E B ij ij i j 2 sys e e e E Bij Bij = (1 − αi αj )Lij E B ij Linear system to solve: e E B ij h Lij = Lij − 2 B ij (β c )2 e E Bij h e B 1 − L E Bij ∑ ij ij ij (β c )2 = 2 ∑ij B ij
22
Introduction
Localization
Hybridization
Systematic error
Conclusions
Model of systematic error In this case, we can obtain: sys e e B = (1 − α α )E E B ij ij i j 2 sys e e e E Bij Bij = (1 − αi αj )Lij E B ij Linear system to solve: e E B ij h Lij = αi αj Lij − 2 B ij (β c )2 e E Bij h e B α α − L E Bij ∑ ij ij i j ij (β c )2 = 2 ∑ij B ij
22
Introduction
Localization
Hybridization
Systematic error
Conclusions
Model of systematic error In this case, we can obtain: sys e e B = (1 − α α )E E B ij ij i j 2 sys e e e E Bij Bij = (1 − αi αj )Lij E B ij Linear system to solve: e E B ij h Lij = αi αj Lij − 2 B ij (β c )2 e E Bij h e B α α − L E Bij ∑ ij ij i j ij (β c )2 = 2 ∑ij B ij b = Lh B e + (β c )2 B is The resulting localized-hybridized covariance B ij ij different from an inflation of the original hybrid covariance by αi αj
22
Introduction
Localization
Hybridization
Systematic error
Conclusions
Outline
Introduction Objectively optimized localization Jointly optimized localization and hybridization Accounting for systematic error Conclusions
23
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method:
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
•
affordable for high-dimensional systems,
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
•
affordable for high-dimensional systems,
•
tackling the sampling noise issue only.
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
•
affordable for high-dimensional systems,
•
tackling the sampling noise issue only.
2. Localization and hybridization are two joint aspects of the linear filtering of sample covariances and can be optimized simultaneously.
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
•
affordable for high-dimensional systems,
•
tackling the sampling noise issue only.
2. Localization and hybridization are two joint aspects of the linear filtering of sample covariances and can be optimized simultaneously. 3. If done optimally, hybridization always improves the accuracy of forecast error covariances.
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Conclusions 1. Localization can be optimized with a method: •
based on properties of the ensemble only,
•
affordable for high-dimensional systems,
•
tackling the sampling noise issue only.
2. Localization and hybridization are two joint aspects of the linear filtering of sample covariances and can be optimized simultaneously. 3. If done optimally, hybridization always improves the accuracy of forecast error covariances. Ménétrier, B. and T. Auligné: Optimized Localization and Hybridization to Filter Ensemble-Based Covariances Monthly Weather Review, 2015, accepted
24
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper:
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization.
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization. • Heterogeneous optimization: local averages over subdomains
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization. • Heterogeneous optimization: local averages over subdomains • 3D optimization: joint computation of horizontal and vertical
localizations, and hybridization coefficients
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization. • Heterogeneous optimization: local averages over subdomains • 3D optimization: joint computation of horizontal and vertical
localizations, and hybridization coefficients New preliminary results: e? • Extension of the theory to account for systematic errors in B
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization. • Heterogeneous optimization: local averages over subdomains • 3D optimization: joint computation of horizontal and vertical
localizations, and hybridization coefficients New preliminary results: e? • Extension of the theory to account for systematic errors in B To be done: tests in a cycled quasi-operational configuration
25
Introduction
Localization
Hybridization
Systematic error
Conclusions
Perspectives Thank you
for your attention! Any question? Already done in the paper: • Extension to vectorial hybridization weights:
δ x = β e ◦ δ xe + β c ◦ δ xc → Requires the solution of a nonlinear system A(Lh , β c ) = 0, performed by a bound-constrained minimization. • Heterogeneous optimization: local averages over subdomains • 3D optimization: joint computation of horizontal and vertical
localizations, and hybridization coefficients New preliminary results: e? • Extension of the theory to account for systematic errors in B To be done: tests in a cycled quasi-operational configuration
25