Posterior Covariance vs. Analysis Error Covariance in Data Assimilation

Transcription

1 Posterior Covariance vs. Analysis Error Covariance in Data Assimilation F.-X. Le Dimet(1), I. Gejadze(2), V. Shutyaev(3) (1) Université de Grenole (2)University of Strathclyde, Glasgow, UK (3) Institute of Numerical Mathematics, RAS, Moscow, Russia Septemer 27, 2013 F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

2 Overview Introduction Analysis Error Covariance via Hessian Posterior Covariance : A Bayesian approach Effective Covariance Estimates Implementation : some remarks Asymptotic Properties Numerical Example Conclusion F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

3 Introduction 1 There are two asic approaches for Data Assimilation: Variational Methods Kalman Filter Both lead to minimize a cost function : J(u) = (V (u u ), u u ) X (Vo (Cϕ y), Cϕ y) Yo, (1) 2 where u X is a prior initial-value function (ackground state), y Y o is a prescried function (oservational data), Y o is an oservation space, C : Y Y o is a linear ounded operator. We get the same optimal solution ū F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

4 Introduction 2 ū is the solution of the Optimality System : { ϕ t { ϕ = F (ϕ) + f, t (0, T ) t ϕ t=0 = u, + (F (ϕ)) ϕ = C V 1 o (Cϕ y), t (0, T ) ϕ t=t = 0, (2) (3) V 1 (u u ) ϕ t=0 = 0 (4) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

5 Introduction 3 In an analysis there are two inputs: The ackground u the oservation y Both have error and the question is what is the impact of these errors on the analysis? The ackground can e considered from two different viewpoints: Variational viewpoint: the ackground is a regularization term in the Tykhonov s sense to make the prolem well posed Bayesian view point: the ackground is an a priori information on the analysis F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

6 Introduction 4 In the linear case we get the same covariance error for the analysis : the inverse of the Hessian of the cost function. In the non linear case we get two different items: Variational approach : Analysis Error Covariance Bayesian Approach : Posterior Covariance Questions: How to compute, approximate, these elements? What are the differences? F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

7 Analysis Error Covariance 1: True solution and Errors We assume the existence of a true solution u t and an associated true state ϕ t verifying: ϕ t t = F (ϕ t ) + f, t (0, T ) ϕ t t=0 = u t. (5) Then the errors are defined y : u = u t + ξ, y = Cϕ t + ξ o with covariances V andv o F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

8 Analysis Error Covariance 2: Discrepancy Evolution Let δϕ = ϕ ϕ t, δu = u u t. Then for regular F there exists ϕ = ϕ t + τ(ϕ ϕ t ), τ [0, 1], such that { δϕ t F ( ϕ)δϕ = 0, t (0, T ), δϕ t=0 = δu, (6) { ϕ t + (F (ϕ)) ϕ = C Vo 1 (Cδϕ ξ o ), ϕ t=t = 0, (7) V 1 (δu ξ ) ϕ t=0 = 0. (8) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

9 Analysis Error Covariance 3: Exact Equation for Analysis Error Let us introduce the operator R(ϕ) : X Y as follows: R(ϕ)v = ψ, v X, (9) where ψ is the solution of the tangent linear prolem ψ t F (ϕ)ψ = 0, ψ t=0 = v. (10) Then, the system for errors can e represented as a single operator equation for δu: H(ϕ, ϕ)δu = V 1 ξ + R (ϕ)c Vo 1 ξ o, (11) where H(ϕ, ϕ) = V 1 + R (ϕ)c V 1 o CR( ϕ). (12) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

10 Analysis Error Covariance 4 : H Operator The operator H(ϕ, ϕ) : X X can e defined y ψ t ψ t F ( ϕ)ψ = 0, ψ t=0 = v, (13) (F (ϕ)) ψ = C V 1 o Cψ, ψ t=t = 0, (14) H(ϕ, ϕ)v = V 1 v ψ t=0. (15) The operator H(ϕ, ϕ) is neither symmetric, nor positive definite. ϕ = ϕ = θ, it ecomes the Hessian H(θ) of the cost function J 1 in the following auxiliary DA prolem: find δu and δϕ such that J 1 (δu) = inf v J 1(v), where J 1 (δu) = 1 1 (V 2 (δu ξ ), δu ξ ) X (Vo (Cδϕ ξ o ), Cδϕ ξ o ) Yo, 2 (16) and δϕ satisfies the prolem δϕ F (θ)δϕ = 0, δϕ t t=0 = δu. (17) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

11 Analysis Error Covariance 5 : Analysis Error Covariance via Hessian The optimal solution (analysis) error δu is assumed to e uniased, i.e. E[δu] = 0, and V δu = E[(, δu) X δu] = E[(, u u t ) X (u u t )]. (18) The est value of ϕ and ϕ independent of ξ o, ξ is apparently ϕ t and using the error equation reduces to H(ϕ t )δu = V 1 R( ϕ) R(ϕ t ), R (ϕ) R (ϕ t ), (19) ξ + R (ϕ t )C V 1 ξ o, H( ) = V 1 We express δu from equation (??) δu = H 1 (ϕ t )(V 1 ξ + R (ϕ t )C Vo 1 ξ o ) and otain for the analysis error covariance V δu = H 1 (ϕ t )(V 1 o + R ( )C Vo 1 CR( ). (20) + R (ϕ t )C V 1 o CR(ϕ t ))H 1 (ϕ t ) = H 1 (ϕ t ). (21) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

12 Analysis Error Covariance 6 : Approximations In practice the true field ϕ t is not known, thus we have to use an approximation ϕ associated to a certain optimal solution ū defined y the real data (ū, ȳ), i.e. we use V δu = H 1 ( ϕ). (22) In (Raier and Courtier,1992), the error equation is derived in the form (V 1 + R (ϕ)c V 1 o CR(ϕ))δu = V 1 ξ + R (ϕ)c V 1 o ξ o. (23) The error due to transitions R( ϕ) R(ϕ t ) and R (ϕ) R (ϕ t ); we call it the linearization error. The use of ϕ instead of ϕ t in the Hessian computations leads to another error, which shall e called the origin error. F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

13 Posterior Covariance : Bayesian Approach 1 Given u N (ū, V ), y N (ȳ, V o ), the following expression for the posterior distriution of u is derived from the Bayes theorem: p(u ȳ) = C exp ( 1 1 (V 2 (u ū ), u ū ) X ) exp( 1 1 (Vo (Cϕ ȳ), Cϕ ȳ) Yo ). 2 (24) The solution to the variational DA prolem with the data y = ȳ and u = ū is equal to the mode of p(u, ȳ) (see e.g. Lorenc, 1986; Tarantola, 1987). Accordingly, the Bayesian posterior covariance is defined y : with u p(u ȳ). V δu = E[(, u E[u]) X (u E[u]) (25) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

14 Posterior Covariance : Bayesian Approach 2 In order to compute V δu y the Monte Carlo method, one must generate a sample of pseudo-random realizations u i from p(u ȳ). We will consider u i to e the solutions to the DA prolem with the pertured data u = ū + ξ, and y = ȳ + ξ o, where ξ N (0, V ), ξ o N (0, V o ). Further we assume that E[u] = ū, where ū is the solution to the unpertured prolem in which case V δu can e approximated as follows V δu = E[(, u ū) X (u ū)] = E[(, δu) X δu]. (26) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

15 Posterior Covariance : O.S. for Errors Unpertured O.S. with u = ū, y = ȳ: ϕ t ϕ t = F ( ϕ) + f, ϕ t=0 = ū, (27) + (F ( ϕ)) ϕ = C V 1 o (C ϕ ȳ), ϕ t=t = 0, (28) V 1 (ū ū ) ϕ t=0 = 0 (29) With perturations : u = ū + ξ, y = ȳ + ξ o, where ξ X, ξ o Y o. δu = u ū, δϕ = ϕ ϕ, δϕ = ϕ ϕ. δϕ t δϕ t = F (ϕ) F ( ϕ), δϕ t=0 = δu, (30) +(F (ϕ)) δϕ = [((F ( ϕ)) F (ϕ)) ] ϕ +C V 1 o (Cδϕ ξ o ), (31) V 1 (δu ξ ) δϕ t=0 = 0. (32) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

16 Posterior Covariance : Exact Errors Equations Introducing ϕ 1 = ϕ + τ 1 δϕ, ϕ 2 = ϕ + τ 2 δϕ, τ 1, τ 2 [0, 1], we derive the exact system for errors: δϕ t δϕ t = F ( ϕ 1 )δϕ, δϕ t=0 = δu, (33) + (F (ϕ)) δϕ = [(F ( ϕ 2 )) ϕ ] δϕ + C V 1 o (Cδϕ ξ o ), (34) V 1 (δu ξ ) δϕ t=0 = 0, (35) equivalent to a single operator equation for δu: where H(ϕ, ϕ 1, ϕ 2 ) = V 1 H(ϕ, ϕ 1, ϕ 2 )δu = V 1 ξ + R (ϕ)c Vo 1 ξ o, (36) + R (ϕ)(c V 1 o C [(F ( ϕ 2 )) ϕ ] )R( ϕ 1 ). (37) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

17 Posterior Covariance : Operator H H(ϕ, ϕ 1, ϕ 2 ) : X X is defined y solving: ψ t ψ t = F ( ϕ 1 )ψ, ψ t=0 = v, (38) (F (ϕ)) ψ = [(F ( ϕ 2 )) ϕ ] ψ C V 1 o Cψ, (39) H(ϕ, ϕ 1, ϕ 2 )v = V 1 v ψ t=0. (40) If we had ϕ = ϕ 1 = ϕ 2, H(ϕ) ecomes the Hessian of the cost function in the original DA prolem, it is symmetric and positive definite if u is a minimum of J(u). The equation is often referred as the second order adjoint (Le Dimet et al., 2002). As aove, we assume that E(δu) 0, and we consider the following approximations R( ϕ 1 ) R( ϕ), R (ϕ) R ( ϕ), [(F ( ϕ 2 )) ϕ ] [(F ( ϕ)) ϕ ]. (41) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

18 Posterior Covariance via Hessian The exact error equation (??) is approximated as follows where H( ) = V 1 Now, we express δu : H( ϕ)δu = V 1 ξ + R( ϕ) C Vo 1 ξ o, (42) + R ( )(C V 1 o C [(F ( )) ϕ ] )R( ). (43) δu = H 1 ( ϕ)(v 1 ξ + R( ϕ) C Vo 1 ξ o ), and otain an approximate expression for the posterior error covariance V 1 = H 1 ( ϕ)(v 1 + R ( ϕ)v o 1 R( ϕ))h 1 ( ϕ) = H 1 ( ϕ)h( ϕ)h 1 ( ϕ), (44) where H( ϕ) is the Hessian of the cost function J 1 computed at θ = ϕ. Other approximations of the posterior covariance: V 2 = H 1 ( ϕ), V 3 = H 1 ( ϕ). (45) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

19 Posterior Covariance : Effective Estimates To supress the linearization errors, the effective inverse Hessian may e used for estimating the analysis error covariance (see Gejadze et al., 2011): V δu = E [ H 1 (ϕ) ]. (46) The same is true for the posterior error covariance V δu : V δu = E [ H 1 (ϕ, ϕ 1, ϕ 2 ) H(ϕ) H 1 (ϕ, ϕ 1, ϕ 2 ) ]. First, we sustitute a possily asymmetric and indefinite operator H(ϕ, ϕ 1, ϕ 2 ) y H(ϕ): V δu V e 1 = E [ H 1 (ϕ) H(ϕ) H 1 (ϕ) ]. (47) Next, y assuming H(ϕ)H 1 (ϕ) I we get V δu V e 2 = E [ H 1 (ϕ) ]. (48) Finally, y assuming H 1 (ϕ) H 1 (ϕ) we otain yet another approximation V δu V e 3 = E [ H 1 (ϕ) ]. (49) F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

20 Numerical Example ϕ(x, t) is governed y the 1D Burgers equation with a nonlinear viscous term : ϕ t + 1 (ϕ 2 ) = ( ν(ϕ) ϕ ), (50) 2 x x x ϕ = ϕ(x, t), t (0, T ), x (0, 1), with the Neumann oundary conditions ϕ = ϕ = 0 (51) x x x=0 x=1 and the viscosity coefficient ( ) ϕ 2 ν(ϕ) = ν 0 + ν 1, ν 0, ν 1 = const > 0. (52) x Two initial conditions u t = ϕ t (x, 0) are considered (case A and case B), F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

21 Numerical Example : Cas A ϕ F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

22 Numerical Example : Cas B ϕ F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

23 Numerical Example : Squared Riemann distance ( M ) 1/2. µ(a, B) = i=1 ln2 γ i Case µ 2 (V 3, ˆV) µ 2 (V 2, ˆV) µ 2 (V 1, ˆV) µ 2 (V3 e, ˆV) µ 2 (V2 e, ˆV) µ 2 (V1 e, ˆV) A A A A A A A B B B B B B F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

24 The reference mean deviation ˆσ(x) (related to ˆV). Cases A2, A case A2 case A8 σ x The mean deviation vector σ is defined as follows: σ(i) = V 1/2 (i, i). F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

25 The reference mean deviation ˆσ(x) (related to ˆV). Cases B6, B case B6 case B9 σ x F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

26 Asolute errors in the correlation matrix: 3, e3, e1 (case A2) a) ) c) case A2 F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1

27 Asolute errors in the correlation matrix: 3, e3, e1 (case A8) a) ) c) F.-X. Le Dimet (INRIA) case A8 Posterior covariance Septemer 27, / 1

28 Asolute errors in the correlation matrix: 3, e3, e1 (case B6) a) ) c) F.-X. Le Dimet (INRIA) case B6 Posterior covariance Septemer 27, / 1

29 Conclusion Variational and Bayesian approaches of DA lead to the minimization of the same function For Error Estimation we otain two different concepts Analysis Error Covariance for the Variational Approach Posterior Covariance for the Bayesian Approach In the linear case the covariances coincide In the nonlinear case strong discrepancies can occur. Algorithms for the estimation and the approximation of these covariances are proposed. Gejadze, I.,Shutyaev, V., Le Dimet, F.-X. Analysis error covariance versus posterior covariance in variational data assimilation prolems. Q.J.R.Meteolol. Soc. (2013). F.-X. Le Dimet (INRIA) Posterior covariance Septemer 27, / 1