Minimisation Techniques 1 Assimilation Algorithms: Minimisation Techniques Yannick Trémolet ECMWF Data Assimilation Training Course March 2006
Minimisation Techniques 2 4D Variational Data Assimilation
Minimisation Techniques 3 Incremental 4D-Var
Minimisation Techniques 4 Incremental 4D-Var
Minimisation Techniques 5 Incremental 4D-Var
Minimisation Techniques 6 The Outer Iterations
Minimisation Techniques 7 The Inner Iterations
Minimisation Techniques 8 Minimisation: Newton method
Minimisation Techniques 9 Minimisation: Newton method
Minimisation Techniques 10 Minimisation: Quasi-Newton method
Minimisation Techniques 11 Minimisation: Quasi-Newton method
Minimisation Techniques 12 Limited Memory Quasi-Newton
Minimisation Techniques 13 Minimisation: Steepest Descent
Minimisation Techniques 14 Minimisation: Steepest Descent The first step fully minimizes the function in the descent direction, but this is undone by subsequent steps. We want to avoid this.
Minimisation Techniques 15 Minimisation: Conjugate Gradient
Minimisation Techniques 16 Conjugate Gradient Convergence
Minimisation Techniques 17 Preconditioning
Minimisation Techniques 18 4D-Var Preconditioning
Minimisation Techniques 19 A case of poor convergence
Minimisation Techniques 20 Theoretical example
Minimisation Techniques 21 Theoretical example
Minimisation Techniques 22 A case of poor convergence
Minimisation Techniques 23 Hessian Preconditioning
Minimisation Techniques 24 Hessian Eigenvectors Preconditioning
Minimisation Techniques 25 4D-Var Eigenvalues Eigenvalue N =1 1 = = Preconditioning reduces the condition number k= 1 / N from to
Minimisation Techniques 26 Conjugate Gradient and Lanczos Algorithm
Minimisation Techniques 27 Lanczos Algorithm
Minimisation Techniques 28 Lanczos Algorithm
Minimisation Techniques 29 Conjugate Gradient and Lanczos Algorithm
Minimisation Techniques 30 Superlinear Convergence
Minimisation Techniques 31 Rounding Error
Minimisation Techniques 32 CG Cost function reduction Quasi-Newton with inexact line searches Quasi-Newton with exact line searches Conjugate Gradient without orthogonalisation Conjugate Gradient with orthogonalisation
Minimisation Techniques 33 CG Gradient norm reduction Quasi-Newton with inexact line searches Quasi-Newton with exact line searches Conjugate Gradient without orthogonalisation Conjugate Gradient with orthogonalisation
Minimisation Techniques 34 4D-Var Cost function reduction Variational Quality Control
Minimisation Techniques 35 4D-Var gradient norm reduction Convergence is roughly twice as fast with Hessian preconditioning.
Minimisation Techniques 36 CG reduction of norm of gradient 0.05
Minimisation Techniques 37 Spectrum of preconditioned Hessian
Minimisation Techniques 38 RMS of T analysis increments
Minimisation Techniques 39 4D-Var Convergence
Minimisation Techniques 40 4D-Var Convergence
Minimisation Techniques 41 Summary
Minimisation Techniques 42