1. Computing a posteriori covariance in variational DA I.Gejadze, F.-X. Le Dimet, V.Shutyaev

2. What is a posteriori covariance and why one needs to compute it ? - 1 Model of evolution process Objective function (for the initial value control) Uncertainty in initial, boundary conditions, distributed coefficients, model error …. Uncertainty in the event measure - target state Measure of event Data assimilation is used to reduce uncertainty in controls / parameters, and eventually in the event measure can be defined in 2D-3D, can be a vector-function

3. What is a posteriori covariance and why one needs to compute it ? – 2(idea of adjoint sensitivities) Trivial relationship between (small) uncertainties in controls / parameters and in the event measure How to compute the gradient? 1. Direct method ( ) : for solve forward model with compute 2. Adjoint method: form the Lagrangian zero take variation integrate by parts If = zero and For initial value problemGenerally Gateaux derivative

4. What is a posteriori covariance and why one needs to compute it ? - 3 scalar - n-vector sensitivity vector Objective function - the covariance matrix of the background error or a priori covariance (measure of uncertainty in before DA) Discussion is mainly focused on the question: does the linearised error evolution model (adjoint sensitivity) approximate the non-linear model well? However, another issue is equally (if not more) important: do and how well we really now ? -must be a posteriori covariance of the estimation error (measure of uncertainty in after DA)

5. How to compute a-posteriori covariance?

6. Hessian Definitions: Gradient, where Hessian Function For large-scale problems, keeping Hessian or its inverse in the matrix form is not always feasible. If the state vector dimension is, the Hessian contains elements.

7. Hessian – operator-function product form 1 Control problem: Optimality system: Void optimality system for exact solution

8. Hessian – operator-function product form - 2 Non-linear optimality system for errors: Definition of errors: There exists a unique representation In the form of non-linear operator equation, we call - the Hessian operator. Hessian operator-function product definition: All operators are defined similarly.

9. Hessian – operator-function product form - 3 We do not know (and do not want to know), only. Hessian operator – is the only one to be inverted in error analysis for any inverse problem, therefore plays the key role in defining ! Since we do not store full matrices, we must find a way to define. Main Theorem Assume errors are normally distributed, unbiased and mutually uncorrelated, and, ; a) if b) if (tangent linear hypothesis is a ‘local’ sufficient condition, not necessary!

10. Optimal solution error covariance reminder Holds for any control or combination of controls!

11. Optimal solution error covariance Case I: 1) Errors are normally distributed, 2) is moderately non-linear or are small, i.e. Case II: 1) errors have arbitrary pdf, 2) ‘weak’ non-linear conditions hold a) pre-compute (define). b) produce single errors implementation using certain pdf generators. c) compute (no inversion involved)! d) generate ensemble of Case III: 1) Errors have arbitrary pdf, 2) is strongly non-linear (chaotic?) or/and are very big All as above, but compute by iterative procedure using as a pre-conditioner e) consider pdf of, which is not normal! f) use it, if you can.

12. Methods for computing the inverse Hessian Direct methods : 1. Fisher information matrix: Requires full matrix storage and algebraic inversion. 2. Sensitivity matrix method (A. Bennett), not suitable for large space-temporal data sets, i.e. good for 3DVar; 3. Finite difference method. Not suitable when constraints are partial differential equations due to large truncation errors; requires full matrix storage and algebraic matrix Inversion. Iterative methods: 4. Lanczos type methods Require only the product. The convergence could be erratic and slow if eigenvalues are no well separated. 5. Quasi-Newton BFGS/LBFGS (generate inverse Hessian as a collateral result) Require only the product. The convergence seems uniform. Required storage can be controlled. Exact method (J. Greenstadt, J.Bard)

13. Preconditioned LBFGS Solve by LBFGS the following auxiliary control problem: BFGS algorithm: - pairs of vectors to be kept in memory. LBFGS keeps only M latest pairs and the diagonal. for chosen. Retrieve pairs which define projected inverse Hessian. Then, How to get ? For example one can use Cholesky factors.

14. Stupid, yet 100% non-linear ensemble method 1. Consider function as the exact solution to the problem 2. Assume that and. 3. The solution of the control problem is 4. Start ensemble loop k=1,…,m. Generate, which correspond to the given pdf. Put Solve the non-linear optimization problem defined in 3). Compute. 5. End ensemble loop. 6. Compute statistic

15. NUMERICAL EXAMPLES: case 1 Model (non-linear convection-diffusion): - driving bc

16. NUMERICAL EXAMPLES: case 1 upstream boundary downstream boundary

18. NUMERICAL EXAMPLES: case 2

21. NUMERICAL EXAMPLES: case 3

24. SUMMARY 1.Hessian plays the crucial role in the analysis of the inverse problem solution errors as the only invertible operator; 2.If errors are normally distributed and constraints are linear the inverse Hessian is itself the covariance operator (matrix) of the optimal solution error; 3.If the problem is moderately non-linear, the inverse Hessian could be a good approximation of the optimal solution error covariance far beyond the validity of the tangent linear hypothesis. Higher order terms could be considered in problem canonical decomposition. 4. Inverse Hessian can be well approximated by a sequence of quasi-Newton updates (LBFGS) using the operator-vector product only. This sequence seems to converge uniformly; 5. Preconditioning dramatically accelerates the computing; 6. The computational cost of computing inverse Hessian should not exceed the cost of data assimilation procedure itself. 7. Inverse Hessian is useful for uncertainty analysis, experiment design, adaptive measuring techniques, etc.

25. PUBLICATIONS 1.I.Yu. Gejadze, V.P. Shutyaev, An optimal control problem of initial data restoration, Comput. Math. & Math. Physics, 39/9 (1999), pp.1416-1425 2.F-X. Le-Dimet, V.P. Shutyaev, On deterministic error analysis in variational data assimilation, Non-linear Processes in Geophysics 14(2005), pp1-10 3.F.-X. Le-Dimet, V.P. Shutyaev, I.Yu. Gejadze, On optimal solution error in variational data assimilation: theoretical aspects. Russ. J. Numer. Analysis and Math. Modelling (2006), v21/2, pp.139-152 4. I. Gejadze, F-X. Le-Dimet and V. Shutyaev. On analysis error covariances in variational data assimilation, SIAM J. Sci. Comp. (2008), v.30, no.4, 1847-74. 5. I. Gejadze, F-X. Le-Dimet and V. Shutyaev. Optimal solution error covariances in variational data assimilation problems, SIAM J. Sci. Comp. (2008), to be published. Covariance matrix of the initial control problem for the diffusion equation. 3-sensor configuration