A Spectral Approach to Linear Bayesian Updating

Similar documents
A Stochastic Collocation based. for Data Assimilation

NON-LINEAR APPROXIMATION OF BAYESIAN UPDATE

Parametric Problems, Stochastics, and Identification

Addressing the nonlinear problem of low order clustering in deterministic filters by using mean-preserving non-symmetric solutions of the ETKF

Stochastic Collocation Methods for Polynomial Chaos: Analysis and Applications

A Note on the Particle Filter with Posterior Gaussian Resampling

Ensemble square-root filters

Adaptive ensemble Kalman filtering of nonlinear systems

Smoothers: Types and Benchmarks

Organization. I MCMC discussion. I project talks. I Lecture.

Ensemble Kalman Filter

Benjamin L. Pence 1, Hosam K. Fathy 2, and Jeffrey L. Stein 3

A new Hierarchical Bayes approach to ensemble-variational data assimilation

State and Parameter Estimation in Stochastic Dynamical Models

Fundamentals of Data Assimila1on

Convergence of Square Root Ensemble Kalman Filters in the Large Ensemble Limit

New Fast Kalman filter method

Sampling and low-rank tensor approximation of the response surface

Inverse Problems in a Bayesian Setting

Gaussian Process Approximations of Stochastic Differential Equations

Lagrangian Data Assimilation and Manifold Detection for a Point-Vortex Model. David Darmon, AMSC Kayo Ide, AOSC, IPST, CSCAMM, ESSIC

Handling nonlinearity in Ensemble Kalman Filter: Experiments with the three-variable Lorenz model

Ensemble forecasting and flow-dependent estimates of initial uncertainty. Martin Leutbecher

Kalman Filter and Ensemble Kalman Filter

Gaussian Filtering Strategies for Nonlinear Systems

Dynamic System Identification using HDMR-Bayesian Technique

Maximum Likelihood Ensemble Filter Applied to Multisensor Systems

Ensemble Data Assimilation and Uncertainty Quantification

An Efficient Ensemble Data Assimilation Approach To Deal With Range Limited Observation

ECONOMETRIC METHODS II: TIME SERIES LECTURE NOTES ON THE KALMAN FILTER. The Kalman Filter. We will be concerned with state space systems of the form

Data Assimilation Research Testbed Tutorial

Estimating functional uncertainty using polynomial chaos and adjoint equations

Relative Merits of 4D-Var and Ensemble Kalman Filter

4 Derivations of the Discrete-Time Kalman Filter

Convergence of the Ensemble Kalman Filter in Hilbert Space

An introduction to data assimilation. Eric Blayo University of Grenoble and INRIA

The Ensemble Kalman Filter:

Optimisation under Uncertainty with Stochastic PDEs for the History Matching Problem in Reservoir Engineering

A data-driven method for improving the correlation estimation in serial ensemble Kalman filter

Local Ensemble Transform Kalman Filter: An Efficient Scheme for Assimilating Atmospheric Data

Localization in the ensemble Kalman Filter

Bred vectors: theory and applications in operational forecasting. Eugenia Kalnay Lecture 3 Alghero, May 2008

Application of the Ensemble Kalman Filter to History Matching

The Ensemble Kalman Filter:

Kalman Filter. Predict: Update: x k k 1 = F k x k 1 k 1 + B k u k P k k 1 = F k P k 1 k 1 F T k + Q

Stability of Ensemble Kalman Filters

The Laplace driven moving average a non-gaussian stationary process

Short tutorial on data assimilation

Nonlinear error dynamics for cycled data assimilation methods

L06. LINEAR KALMAN FILTERS. NA568 Mobile Robotics: Methods & Algorithms

Practical Aspects of Ensemble-based Kalman Filters

Efficient Solvers for Stochastic Finite Element Saddle Point Problems

(Extended) Kalman Filter

Data assimilation in high dimensions

Fast Numerical Methods for Stochastic Computations

2D Image Processing. Bayes filter implementation: Kalman filter

EnKF-based particle filters

A new unscented Kalman filter with higher order moment-matching

Fundamentals of Data Assimila1on

Hierarchical Bayes Ensemble Kalman Filter

DATA ASSIMILATION FOR FLOOD FORECASTING

Local Ensemble Transform Kalman Filter

Stochastic Spectral Approaches to Bayesian Inference

2D Image Processing. Bayes filter implementation: Kalman filter

Aspects of the practical application of ensemble-based Kalman filters

Bayesian Inverse problem, Data assimilation and Localization

Methods of Data Assimilation and Comparisons for Lagrangian Data

Mini-Course 07 Kalman Particle Filters. Henrique Massard da Fonseca Cesar Cunha Pacheco Wellington Bettencurte Julio Dutra

The Kalman Filter ImPr Talk

Chapter 6: Ensemble Forecasting and Atmospheric Predictability. Introduction

Stochastic structural dynamic analysis with random damping parameters

Bayesian Statistics and Data Assimilation. Jonathan Stroud. Department of Statistics The George Washington University

Stochastic Processes. M. Sami Fadali Professor of Electrical Engineering University of Nevada, Reno

Parameter Estimation for Mechanical Systems Using an Extended Kalman Filter

Hierarchical Parallel Solution of Stochastic Systems

Multilevel stochastic collocations with dimensionality reduction

Winter 2019 Math 106 Topics in Applied Mathematics. Lecture 1: Introduction

ECE521 week 3: 23/26 January 2017

Ergodicity in data assimilation methods

Model error and parameter estimation

NON-LINEAR NOISE ADAPTIVE KALMAN FILTERING VIA VARIATIONAL BAYES

Data assimilation with and without a model

Revision of TR-09-25: A Hybrid Variational/Ensemble Filter Approach to Data Assimilation

DART_LAB Tutorial Section 2: How should observations impact an unobserved state variable? Multivariate assimilation.

Lecture 16: State Space Model and Kalman Filter Bus 41910, Time Series Analysis, Mr. R. Tsay

Accepted in Tellus A 2 October, *Correspondence

Asynchronous data assimilation

Robust Ensemble Filtering With Improved Storm Surge Forecasting

What do we know about EnKF?

Covariance Matrix Simplification For Efficient Uncertainty Management

Robotics 2 Target Tracking. Kai Arras, Cyrill Stachniss, Maren Bennewitz, Wolfram Burgard

Data assimilation with and without a model

Miscellaneous. Regarding reading materials. Again, ask questions (if you have) and ask them earlier

Data Assimilation with the Ensemble Kalman Filter and the SEIK Filter applied to a Finite Element Model of the North Atlantic

Constrained State Estimation Using the Unscented Kalman Filter

Systematic strategies for real time filtering of turbulent signals in complex systems

ROBOTICS 01PEEQW. Basilio Bona DAUIN Politecnico di Torino

Lagrangian Data Assimilation and Its Application to Geophysical Fluid Flows

Estimation of State Noise for the Ensemble Kalman filter algorithm for 2D shallow water equations.

Analysis Scheme in the Ensemble Kalman Filter

Transcription:

A Spectral Approach to Linear Bayesian Updating Oliver Pajonk 1,2, Bojana V. Rosic 1, Alexander Litvinenko 1, and Hermann G. Matthies 1 1 Institute of Scientific Computing, TU Braunschweig, Germany 2 SPT Group GmbH, Hamburg, Germany 1

Outline Motivation / Problem Statement Proposed Solution Examples Non-trivial scalar update Gauss-linear model State estimation of Lorenz-63 model Conclusions / Outlook 2

Motivation Application: stochastic inverse & control problems on dynamical systems Uncertain parameter & state estimation from noisy evidence, Subsequent optimal control under uncertainty (or closed-loop; not considered further in this talk) Most existing (linear) Bayesian methods use sampling Spectral representations of RVs possess nice convergence properties, are deterministic make use of that! Updating and spectral representation should be tightly integrated ( uncertainty quantification) Approaches exist, but have specific problems 3

Linear Bayesian Updating of Polynomial Chaos Coefficients (1/3) First ingredient: Linear Update Formula for RVs 1 : Second ingredient: Polynomial Chaos Expansion 2 : x ω = x ω + K z ω y ω with K = C xy C z +C y 1 r ω = r α α J H α θ 1 ω,, θ k ω, y(ω) = h x ω z ω = h x + ε(ω) x(ω) prior, x(ω) posterior, x unknown truth to identify, ε ω measurement error, z ω evidence + error model, h measurement operator 1 Kalman (1960) [Gaussian RVs]; Luenberger (1969) [L 2 RVs] 2 Possibly other spectral expansions. Projection on PCE gives: α J: x α = x α + K z α y α Truncation & limitation to finite amount of basis RVs θ i ω gives: α J Z : x α = x α + K z α y α This LPCU is a formula which actually can be implemented: 4

Linear Bayesian Updating of Polynomial Chaos Coefficients (2/3) Write truncated PCE as matrices of coefficient column vectors: X =, x α, X = X + K Z Y Kalman gain computation: Introduce diagonal Gram matrix: αβ = E H α H β = diag(α!) Some matrix shorthands: X = X α>0, X = X α=0 = E X, therefore: X = [X X] Then the involved (cross-) covariance matrices are easy to approximate, e.g.: C xy X Y T, and the computation of K is complete. 5

Linear Bayesian Updating of Polynomial Chaos Coefficients (3/3) Now all ingredients of the formula are ready... aren t they? not quite. Important: ε ω and y ω are assumed as uncorrelated, therefore Z Y (as defined above) is technically wrong! The problem: C yε = C yz = 0 is assumed for the formula. The variance of z ω and y ω must add. Possible solutions: 1. ignore it, 2. fix it, 3. circumvent it. 6

The Implementation Issue and Possible Treatments Ignore it: Simply re-use the same basis RVs for evidence error and forward model. In other words: directly compute Z Y, thereby introducing that ε ω and y ω are correlated. See this as necessary approximation (denoted co-linear in the following). However, this approach over-estimates the posterior variance. Fix it: Introduce new basis RVs with each update. Not really applicable for sequential updating in dynamical systems. Circumvent it: Come up with some consistent way to avoid the additional RVs; e.g. Zeng (2010) and Blanchard (2010) argue that the PCE coefficients are zero due to independence. However, this approach under-estimates the posterior variance (cf. Burgers (1998) for EnKF case). 7

Circumvent It, Differently: A Square Root Implementation (1/3) Idea of square root approaches: Update mean x(ω) and varying part x ω of RV x(ω) independently (cf. Potter (1963). First: Update for the mean remains as-is: X = X + K Z Y. Second: Update for the varying part: Realise that: C x X X T = X X T = SS T ( is diagonal) S is a (very specific) square root of the prior covariance matrix. Idea: Transform S into S, a square root of the posterior covariance matrix? Ansatz: Find matrix A with S = SAT. (T some orthonormal matrix) 8

Circumvent It, Differently: A Square Root Implementation (2/3) Start with update for covariance (e.g. Kalman, 1960): (req.: linear measurement H) C x = (I KH)C x = C x C x H T HC x H T + C z 1 HC x = SS T SS T H T HSS T H T + C z 1 HSS T = SMS T Substitute: C x SS T with M = I S T H T HSS T H T + C z 1 HS. A matrix A with AA T = M would be a solution to the ansatz. Therefore: Compute eigenvalue decomposition HSS T H T + C z = BΛB T. Then: M = I S T H T BΛB T 1 HS = I S T H T BΛ 1 B T HS = I Λ 1 2B T HS = I W T W (cf. Evensen, 2004: EnSRF) T Λ 1 2B T HS M = I (UΣV T ) T UΣV T = I VΣ T ΣV T = V I Σ T Σ V T = V I Σ T Σ V I Σ T Σ T giving the desired square root of M. 9

Circumvent It, Differently: A Square Root Implementation (3/3) With this result, the ansatz becomes: S = SAT = S V I Σ T Σ T. It remains to choose T. See V as mapping between normalized PCE space and covariance structure space. We need to map back; therefore, choose T = V T, giving the final update equation for the varying part: S = S V I Σ T Σ V T. To obtain posterior PCE, set X = S Δ and join with updated mean: X = [XX]. This will be denoted as SRPCU in the following. Approach is not uniqe: Use a pre-multiplication ansatz Use a different derivation for the square root Choice of T may be conclusive but it is still a choice! others exist. 10

Numerical Example: Non-trivial Scalar Update 11

Numerical Example: Non-trivial Scalar Update (Zoom) 12

Numerical Example: Lorenz-63 State Estimation System model: 3d, non-linear, chaotic Task: State estimation from evidence Noise model: N(μ = 0, σ = 3) for all three variables Initial conditions & parameters: standard choices Governing equations: dx dt dy dt = s y x = rx y xz dz = xy bz dt 13

Following Plots: Some Functionals and their Reliability Functionals f: root mean square error (RMSE = 1 N variance, skewness, and kurtosis 1000 repetitions of each experiment Plots contain M f and a reliability measure of functional f: V( ), the unbiased sample variance, and M( ), the sample mean, V(f) M f 2 are computed over the 1000 repetitions. N E x i x 2 i=1 i ), probabilistic components are randomized (evidence noise, initial ensemble noise, simulated data noise) 14

Lorenz-63: RMSE (SRPCU, EnSRF) 15

Lorenz-63: Variance (SRPCU, EnSRF) 16

Lorenz-63: Skewness (SRPCU, EnSRF) 17

Lorenz-63: Kurtosis (SRPCU, EnSRF) 18

Lorenz-63: PDF Estimates @ t = 80 (SRPCU, EnSRF) EnSRF: tends to produce outliers, clusters (known effect) SRPCU: smooth, non- Gaussian 19

Lorenz-63: PDF Estimates @ t = 80 (SRPCU, EnKF) EnKF: smooth, but strong tendency to Gaussian estimates SRPCU: smooth, non- Gaussian 20

Conclusions Fully deterministic method (as opposed to EnKF, EnSRF) Applications where this is mandatory Very efficient (update takes practically no time, evolution of the model is expensive part) Exact for Gauss-linear problems (as theory would predict) Higher moments are transferred from prior to posterior, mean and variance are corrected Avoids the outlier problem of EnSRF Avoids the growing PCE basis problem of correct, but non-square-root schemes Drawback: only evidence and assumed covariance enter the update not distributional form (e.g. non-gaussian) 21

Outlook Combination with adaptive subspace selection schemes Other spectral expansions Collocation methods (cf. Zeng (2010)) Different V to change variance re-distribution Iterative variants for improved non-linear identification Pre-multiplication schemes (as opposed to the post-multiplication used here) Non-linear h( )? Regularization techniques like covariance localization (cf. EnKF) 22

References Pajonk, O.; Rosić, B. V.; Litvinenko, A. & Matthies, H. G., A Deterministic Filter for Non-Gaussian Bayesian Estimation, Physica D: Nonlinear Phenomena, 2012, 241, 775-788, DOI:10.1016/j.physd.2012.01.001 Rosić, B. V.; Litvinenko, A.; Pajonk, O. & Matthies, H. G., Direct Bayesian Update of Polynomial Chaos Representations, Journal of Computational Physics, 2011, Submitted for publication Related Methods: Blanchard, E. D.; Sandu, A. & Sandu, C., A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems, Journal of Dynamic Systems, Measurement, and Control, ASME, 2010, 132, 061404 Zeng, L. & Zhang, D., A stochastic collocation based Kalman filter for data assimilation, Computational Geosciences, Springer Netherlands, 2010, 14, 721-744 Saad, G. A., Stochastic Data Assimilation with Application to Multi-Phase Flow and Health Monitoring Problems, Faculty of the Graduate School, University of Southern California, 2007 Bibliography: Kálmán, R. E., A New Approach to Linear Filtering and Prediction Problems, Transactions of the ASME - Journal of Basic Engineering, 1960, 82, 35-45 Potter, J. E. & Stern, R. G., Statistical filtering of space navigation measurements, Proceedings of the AIAA Guidance and Control Conference, Massachusetts Institute of Technology, August, 1963 Evensen, G., Sampling strategies and square root analysis schemes for the EnKF, Ocean Dynamics, 2004, 54, 539-560 23

24

Numerical Example: Gauss-Linear Model (1/3) Case with σ = 0.01 Mean good for all methods (watch scale!) Dirac evidence underestimates variance RMSE is smaller Colinear evidence overestimates variance, but order of EnKF noise (not visible) 25

Numerical Example: Gauss-Linear Model (2/3) Case with σ = 0.1 Mean bad for colinear and Dirac evidence Variance estimates are constantly wrong for both, but not completely off RMSE therefore worse than KF solution 26

Numerical Example: Gauss-Linear Model (3/3) Case with σ = 1.0 Mean is partly completely off, variance too LPCU with square root, EnKF, and EnSRF reproduce the KF result in all cases Other LPCU variants over/underestimate variance 27