(Extended) Kalman Filter

Transcription

1 (Extended) Kalman Filter Brian Hunt 7 June 2013

2 Goals of Data Assimilation (DA) Estimate the state of a system based on both current and all past observations of the system, using a model for the system dynamics. Perform the estimation iteratively: compute the current estimate in terms of a recent past estimate. Ideally, quantify the uncertainty in the state estimate.

3 Terminology and Notation Forecast model: a known function M on a vector space of model states. Truth: an unknown sequence {x n } of model states to be estimated. Model error: δ n = x n+1 M(x n ). Observations: a sequence {y n } of vectors in observation space (may depend on n). Forward operator: a known function H n from model space to observation space. Observation error: ε n = y n H n (x n ).

4 When DA is not Necessary If the forward operator H n is invertible and ε n = 0 then x n = H 1 n (y n ). If H n is invertible and the statistics of ε n are known, then we can compute the pdf of x n (but data assimilation can improve the estimate). Note: the pdf (probability density function) of x gives the relative likelihood of the possible values of x. The maximizer ( mode ) of the pdf is the most likely value.

5 More Terminology Background ( first guess ): estimate x b n of the current model state x n given past observations y 1,..., y n 1. Analysis: estimate x a n of x n given current and past observations y 1,..., y n. A data assimilation cycle consists of: Analysis step: Determine analysis x a n from background x b n and observations y n. Forecast step: Typically x b n+1 = M(x a n ).

6 Remarks on the Analysis Step If the observation error ε n is zero, we should seek x a n close to x b n such that H n (x a n ) = y n. Otherwise, we should just make H n (x a n ) closer to y n than H n (x b n ) is. How much closer depends on the relative uncertainties of the background estimate x n b and the observation y n. The better we understand the uncertainties, the clearer it is how to do the analysis step.

7 Bayes Rule Definition of conditional probability: Then P(V W ) = P(V W )/P(W ) P(V given W ) = P(V and W )/P(W ) P(V W )P(W ) = P(V W ) = P(V )P(W V ). Corollary: P(V W ) = P(V )P(W V )/P(W ) posterior = prior likelihood/normalization

8 Bayesian Data Assimilation Assume that the statistics of the model error δ n and observation error ε n are known. Theoretically, given an analysis pdf p(x n 1 y 1,..., y n 1 ), we can use the forecast model to determine a background ( prior ) pdf p(x n y 1,..., y n 1 ). The forward operator tells us p(y n x n ). Bayes rule tells us that the analysis ( posterior ) pdf p(x n y 1,..., y n ) is proportional to p(x n y 1,..., y n 1 )p(y n x n ).

9 Advantages and Disadvantages Advantage: the analysis step is simple just multiply two functions. Disadvantage: the forecast step is generally unfeasible in practice. If x is high-dimensional, we can t numerically keep track of an arbitrary pdf for x too much information! We need to make some simplifying assumptions.

10 Linearity and Gaussianity Assume that M and H n are linear. Assume model and observation errors are Gaussian with known covariances and no time correlations: δ n N(0, Q n ) and ε n N(0, R n ). Then in the analysis step, a Gaussian background pdf leads to a Gaussian analysis pdf. Gaussian input yields Gaussian output in the forecast step too. Let the background pdf have mean x b n and covariance P b n.

11 Bayesian DA with Gaussians The (unnormalized) background pdf is: exp[ (x n x b n ) T (P b n) 1 (x n x b n )/2] The pdf of y n given x n is exp[ (H n x n y n ) T R 1 n (H n x n y n )/2] The analysis pdf is the exp( J n /2) where: J n =(x n x b n ) T (P b n) 1 (x n x b n ) + (H n x n y n ) T R 1 n (H n x n y n ) To find the mean and covariance of the analysis pdf, we want to write: J n = (x n x a n ) T (P a n) 1 (x n x a n ) + c

12 The Kalman Filter [Kalman 1960] After some linear algebra, the analysis mean x a n and covariance P a n are x a n = x b n + K n (y n H n x b n ) Pn a = [(Pn) b 1 + Hn T Rn 1 H n ] 1 = [I + PnH b n T Rn 1 H n ] 1 Pn b where K n = PnH a n T Rn 1 matrix. is the Kalman gain The forecast step is x b n+1 = Mx a n and P b n+1 = MPb nm T + Q n

13 Observation Space Formulation After some further linear algebra, the Kalman filter analysis equations can be written K n =P b nh T n [H n P b nh T n + R n ] 1 x a n =x b n + K n (y n H n x b n ) P a n =(I K n H n )P b n The size of the matrix that must be inverted is determined by the number of (current) observations, not by the number of model state variables.

14 Example Assume that M = H n = I, that x is a scalar, and that Q n = 0 and R n = r > 0. We are making independent measurements y 1, y 2,... of a constant-in-time quantity x. The analysis equations are: x a n =x b n + P a nr 1 (y n x b n ) (P a n) 1 =(P b n) 1 + r 1 Start with a uniform prior pdf: (P b 1 ) 1 = 0 and x b 1 arbitrary. Then by induction, P b n+1 = Pa n = r/n and x b n+1 = x a n = (y y n )/n.

15 A Least Squares Formulation In terms of all the observations y 1,..., y n, what problem did we solve to estimate x n? Assume no model error (δ n = 0). The likelihood of a model trajectory x 1,..., x n is exp( J n /2) where: J n = n i=1 (H i x i y i ) T R 1 i (H i x i y i ) Problem: minimize the cost function J n (x 1,..., x n ) subject to the constraints x i+1 = Mx i.

16 Kalman Filter Revisited The Kalman filter expresses the minimizer x a n of J n in terms of the minimizer x a n 1 of J n 1 as follows. It expresses J n 1 as a function of x n 1 only. It keeps track of an auxiliary matrix P a n 1 that is the 2nd derivative (Hessian) of J n 1. Assuming it has done so correctly at time n 1, the next slide explains why it does so at time n.

17 Kalman Filter Revisited If xn 1 a minimizes J n 1 and Pn 1 a Hessian, then is its J n 1 = (x n 1 x a n 1) T (P a n 1) 1 (x n 1 x a n 1)+c n 1 Then substituting x n = Mx n 1, x b n = Mx a n 1, and P b n = MP a n 1 MT yields: J n 1 = (x n x b n ) T (P b n) 1 (x n x b n ) + c n 1 We get the same cost function as before: J n = J n 1 + (Hx n y n ) T R 1 n (Hx n y n ) The KF completes the square as before.

18 Nonlinear Least Squares Now let s eliminate the assumption that M and H i are linear. As before, assume no model error and Gaussian observation errors. The maximum likelihood estimate for the true trajectory is the minimizer of: J n = n i=1 (H i (x i ) y i ) T R 1 i (H i (x i ) y i ) subject to the constraints x i+1 = M(x i ).

19 Approximate Solution Methods Use an approximate solution at time n 1 to find an approximate solution at time n. If we track covariances associated with our estimates, we can write: J n 1 (x n 1 x a n 1) T (P a n 1) 1 (x n 1 x a n 1)+c As a further approximation, we can write: J n 1 (x n x b n ) T (P b n) 1 (x n x b n ) + c It seems clear that x b n should be M(x a n 1 ), but what choice of P b n is best?

20 Extended Kalman Filter Matching the second derivatives of the two approximate cost functions yields P b n = (DM)P a n 1 (DM)T where DM is the derivative of M at x a n 1. The remaining equations are like the Kalman filter (linearizing H n near x b n ). Advantage: The approximation error may be smaller than for other methods. Disadvantage: For a high-dimensional model, the covariance forecast is computationally expensive.

21 Extended KF (Square Root Form) If M is computed by solving a system of differential equations, then DM is computed by solving the associated tangent linear model (TLM). If P a n 1 = X a n 1 (X a n 1 )T, then compute Xn b = (DM)Xn 1 a, followed by Pb n = Xn b (Xn b ) T. This is easier if Xn 1 a has (many) fewer columns than rows; the resulting covariance has reduced rank. The Kalman covariance update becomes X a n = X b n (H n X b n ) T [H n X b n (H n X b n ) T + R n ] 1/2.

22 Tangent Linear Model Suppose x n = x(n) where dx/dt = F (x). Then for all solutions, M(x(0)) = x(1). Consider a family of solutions with x γ (0) = x 0 + γv; then DM(x 0 )v = ( / γ)x γ (1) γ=0. Let v(t) = ( / γ)x γ (t) γ=0. Substituting x γ (t) into the ODE and differentiating w.r.t. γ yields dv/dt = DF(x 0 (t))v Compute v(1) with v(0) = v to get DM(x 0 )v.

23 Ensemble Kalman Filter Use an ensemble of model states whose mean and covariance are transformed according to the Kalman filter equations. Forecast each ensemble member separately. Advantage: Relatively easy to implement and the analysis step is computationally efficient. Disadvantage: Only represents uncertainty in a space whose dimension is bounded by the ensemble size (inherently reduced rank).

24 3D-Var Replace P b n with a time-independent background covariance matrix B, determined empirically. Numerically minimize the resulting cost function (allowing nonlinear H n ). Advantage: The covariance B and associated matrices (B 1/2 is used in the analysis) only need to be computed once. Disadvantage: Ignores time dependence of background uncertainty, which can vary considerably.

25 (Strong Constraint) 4D-Var [le Dimet & Talagrand 1985] Numerically minimize the cost function J n =(x n p xn p) b T B 1 (x n p xn p) b n + (H i (x i ) y i ) T R 1 i (H i (x i ) y i ) i=n p subject to the constraints x i+1 = M(x i ). Advantage: Accuracy, especially as p increases. Disadvantage: Difficult to implement and computationally expensive.