Towards Strehl-Optimal Adaptive Optics Control Donald Gavel, Donald Wiberg, Center for Adaptive Optics, U.C. Santa Cruz Marcos Van Dam, Lawrence Livermore National Laboaratory The goal of adaptive optics is to Maximize Strehl • Piston-removed atmospheric phase: x x x WA x dx • Phase correction by DM: n a a x ai ri x i 1 WA xdx 1 vector of actuator commands ai ai s vector of wavefront sensor readings actuator response functions • Max Strehl minimize residual wavefront variance (Marechal’s aproximation) x x a x 2 2 x WA x dx Strehl e aperture averaged residual 2 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 2 Strehl-optimizing adaptive optics Define the cost function, J = mean square wavefront residual: J 2 x WA x dx 2 Wavefront estimation and control problems are separable (proven on subsequent pages): J J E JC where • JE is the estimation part: 2 J E x ˆx WA x dx • JC is the control part: JC and ˆx 2 ˆ x a x WA x dx x | s is the conditional mean of the wavefront IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 3 The Conditional Mean The conditional probability distribution is defined via Bayes theorem: P|S | s P , S , s PS s The conditional mean is the expected value over the conditional distribution: P , S , s ˆ x x | s P|S | s d d PS s IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 4 Properties of the conditional mean 1. The conditional mean is unbiased: ~ ˆ P , S , s dds PS s P , S | s dds PS s ˆP|S | s dds PS s ˆds PS s ˆ P|S | s dds PS s ˆds PS s ˆds 0 2. The error in the conditional mean is uncorrelated to the data it is conditioned on: P , s ~ x s ˆ s sP , S , s dds s , S d PS s ds 0 PS s 3. The error in the conditional mean is uncorrelated to the conditional mean: x ˆx ˆ ˆ P|S | s d P , S , s dds PS s P|S | s d P|S | s d ds ~ P , S , s d P|S | s d ds PS s P|S | s d P|S | s d ds PS s P , S | s d P|S | s d ds PS s P|S | s d P|S | s d ds 0 4. The error in the conditional mean is uncorrelated to the actuator commands: na na ~ x a x ri x ai s ri x ai s ˆ P|S | s dds ~ i 1 i 1 na ri x ai s P|S | s d P|S | s dds i 1 na ri x ai s P|S | s d P|S | s d ds 0 i 1 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 5 Proof that J = JE+JC (the estimation and control problems are separable) J x ˆ x x ˆ x W x dx x ˆ x 2 x ˆ x x ˆ x x ˆ x 2 a A 2 2 a a W x dx A ~ ~ J E 2 x a x x ˆ x WA x dx J C J E JC 0 0 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 6 1) The conditional mean wavefront is the optimal estimate (minimizes JE) Proof: We show that any other wavefront estimate results in larger JE Let E s ˆs s E s 2 ~ 2 ~ 2 ~ 2 ~ 2 2 2 ˆ J E E Therefore, E 2 ~ s ˆs ~ 2 2 ~ 2 , s s P( , s )dds 2 ~ 2 s PS s , s P( | s )d ds 2 0 2 for any 0 minimizes JE IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 7 Calculating the conditional mean wavefront given wavefront sensor measurements wavefront sensor operator: (average-gradient operator in the Hartmann slope sensor case) The measurement equation Measurement noise si x Wi x dx vi S For Gaussian distributed and n, it is straightforward to show (see next page) that the conditional mean of must be a linear function of s: ˆs ns k xs i 1 i i Ks Cross-correlate both sides with s and solve for K ˆsT K ssT K ˆsT ssT so where (known as the “normal” equation) 1 s T T ss 1 since ~ s 0 ˆ x p T x S 1s pi x x si x x Wi s xdx Sij si s j Wi s x W js x x x dxdx vi v j IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 8 Aside: Proof that the conditional mean is a linear function of measurements if the wavefront and measurement noise are Gaussian si x Wi S x dx vi s H s P s | P d S | PS s 1 1 T exp s H v1 s H T 1 d 2 2 1 1 T arg max s H v1 s H T 1 2 2 H T v1H 1 Ks 1 Hv1s Measurement equation Measurement is a linear function of wavefront Bayesian conditional mean Gaussian distribution = maximum log-Likelihood of a-posteriori distribution = a linear (least squares) solution IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 9 2) The best-fit of the DM response functions to the conditional mean wavefront minimizes JC 2 J C ˆx a x WA x dx 2 JC 0 ˆ x a r x i i WA x dx i J C a j ˆ x a r x i i rj x WA x dx i ˆx rj x WA x dx ai ri x rj x WA x dx i r aT R a R 1r where r rj x ˆx WA x dx and R ri ( x)rj ( x)WA ( x)dx IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 10 Comparing to Wallner’s1 solution Combining the optimal estimator (1) and optimal controller (2) solutions gives Wallner’s “optimal correction” result: a R 1 AS 1s where A ri x p j x WA x dx • The two methods give the same result, a set of Strehl-optimizing actuator commands • The conditional mean approach separates the problem into two independent problems: 1) statistically optimal estimation of the wavefront given noisy data 2) deterministic optimal control of the wavefront to its optimal estimate given the deformable mirror’s actuator influence functions • We exploit the separation principle to derive a Strehl-optimizing closed-loop controller 1E. P. Wallner, Optimal wave-front correction using slope measurements, JOSA, 73, 1983. IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 11 The covariance statistics of (x) (piston-removed phase over an aperture A) x x WA d x x x x x xWA xdx x xWA xdx x xWA xWA xdxdx x x D x x g x g x a 1 2 where 2 2 2 D x x x x x x 2 x x g x 1 D x xWA xdx 2 1 a WA x D x xg xWA xdxdx 2 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 12 The g(x) function and a are “generic” under Kolmogorov statistics x 2 g x 6.88D r0 53 6.88D r0 53 • D(x) = 6.88(|x|/r0)5/3 • Circular aperture, diameter D • Factor out parameters 6.88(D/r0)5/3 and integrals are computable numerically x D x D a 0.149831 6.88 D r0 53 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 13 Towards a Strehl-optimizing control law for adaptive optics Remember our goal is to maximize Strehl = minimize wavefront variance in an adaptive optics system • But adaptive optic systems measure and control the wavefront in closed loop at sample times that are short compared to the wavefront correlation time. • So the optimum controller uses the conditional mean, conditioned on all the previous data: x x s t , s t 1 , s t 2 , IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 14 We need to progress the conditional mean through time (the Kalman filter2 concept) t1 x t 1 x s t 1 , s t 2 , t x t x s t 1 , s t 2 , t x t x s t , s t 1 , s t 2 , t x s t , t x 1. Take a conditional mean at time t-1 and progress it forward to time t 2. Take data at time t 3. Instantaneously update the conditional mean, incorporating the new data 4. Progress forward to time step t+1 5. etc. 2Kalman, R.E., A New Approach to Linear Filtering and Prediction Problems, J. Basic Eng., Trans. ASME, 82,1, 1960. IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 15 Kalman filtering new data new data st t1 x Time progress t x st 1 t x Update Time progress t1 x t1 x Update ... ... IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 16 Problems with calculating and progressing the conditional mean of an atmospheric wavefront through time • The wavefront is defined on a Hilbert Space (continuous domain) at an infinite number of points, x A (A = the aperture). • The progression of wavefronts with time is not a well-defined process (Taylor’s frozen flow hypothesis, etc.) • In addition to the estimate, the estimate’s error covariance must be updated at each time step. In the Hilbert Space, these are covariance bi-functions: ct (x,x’)=<t(x),t (x’)>, x A, x’ A. IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 17 Justifying the extra effort of the optimal estimator/optimal controller • If is interesting to compare “best possible” solutions to what we are getting now, with “non-optimal” controllers • Determine if there is room for much improvement. • Gain insights into the sensitivity of optimal solutions to modeling assumptions (e.g. knowledge of the wind, Cn2 profile, etc.) • Preliminary analysis of tomographic (MCAO) reconstructors suggest that Weiner (statistically optimal) filtering may be necessary to keep the noise propagation manageable IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 18 Updating a conditional mean given new data Say we are given a conditional mean wavefront given previous wavefront measurements ˆ s t 1 s t H vt And a measurement at time t The residual ~ et st H st 1 H st 1 vt H vt T is uncorrelated to previous measurements, e t s t 1 0 Applying the normal equation on the two pieces of data et and st-1: ˆ s t , s t 1 et , s t 1 sT where t 1 eT t s t 1sT t 1 T et s t 1 ~ et K t et eTt et eTt 0 1 1 s t 1e t s t 1 ~ s et 0 t 1 T e et e t t T et Summarizing: ˆ ˆ K st Hˆ IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 19 …written in Wallner’s notation • Estimate-update, given new data st: 1 t x t x p T t x S t s t s t s t s t s t 1 , s t 2 , W s x t x dx ~ ~ pt x x et t x t x W s xdx et st sˆ t Hartmann sensor applied to the wavefront estimate Correlation of wavefront to measurement T ~ ~ St et etT W s x W s x t x t x dxdx vv T Correlation of measurement to itself • Covariance-update: ~ ~ ~ ~ 1 t x t x t x t x p T x S t t p t x where the estimate error is defined: ~ t x t x t x IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 20 How it works in closed loop t x + at x x t W x x dx s Wavefront sensor Best fit to DM 1 t pT t x S t s t s Estimator t x + t1 x Predictor t x IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 21 Closed-loop measurements need a correction term ^ as s - s, the …since what the wavefront sensor sees is not exactly the same wavefront measurement prediction error s s W x x dx W x x a x dx W s x x x a x x dx s s W s x a x x dx Measurement prediction error Measurement prediction error = DM Fitting error Hartmann sensor residual (measured data) + DM Fitting error (can be computed from the wavefront estimate and knowledge of the DM) IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 22 Time-progressing the conditional mean t1 x Given how do we determine t x ? Example 1: On a finite aperture, the phase screen is unchanging and frozen in place t x t1 x ~ ~ ~ ~ t x t x t1 x t1 x Consequences: • Estimates corrections accrue (the integrator “has a pole at zero”) • If the noise covariance <vvT> is non-zero, then the updates cause the estimate error covariance to decrease monotonically with t. IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 23 Time-progressing the conditional mean Example 2: The aperture A is infinite, and the phase screen is frozen flow, with wind velocity w t x t1 x w ~ ~ ~ ~ t x t x t1 x w t1 x w Consequence: • An infinite plane of phase estimates must be updated at each measurement IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 24 Time-progressing the conditional mean Example 3: The aperture A is finite, and the phase screen is frozen flow, with wind velocity w w ˆt x A t x A s t 1 , s t , t x A ˆt1 x A A more on this approximation later ˆ xdx F x , x t 1 A A’ as we might expect F x, x x x w for x in the overlap region, AA’ The problem is to determine the progression operator, F(x,x’), for x in the newly blown in region, A A A’ ) IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 25 “Near Markov” approximation t x f t 1 x wx The property where w is random noise uncorrelated to t-1(x), is known as a Markov property. We see that if obeyed a Markov property t x st 1 , st 2 , t x t 1 x st 1 , st 2 , t x ˆt1 x that is, the conditional mean on a finite sized aperture retains all of the relevant statistical information from the growing history of prior measurements. Phase over the aperture however is not Markov, since some information in the “tail” portion, A’’ - (A’’ A’ ), which correlated to st-1, is dropped off and ignored. The fractal nature of Kolmogorov statistics does not allow us to write a Markov difference equation governing on a finite aperture. w We will nevertheless proceed assuming the Markov property since the effect of neglecting in A’’ - (A’’ A’ ) to estimates of in A - (A A’ ) is very small A A’ A’’ IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 26 Validity of approximating wind-blown Kolmogorov turbulence as near-Markov using the estimate at this point To predict this point what is the effect of neglecting this point? A A , A A A A eA eA A A A var A A contribution of point in A’ var A eA A A’ A’’ contribution of neglected point in A’’ wind Information contained in points neglected by the nearMarkov approximation is negligible IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 27 The progression operator from A’ to A We write the conditional mean of the wavefront in A, conditioned on knowing it in A’ x x w A’ x A, x A A G(x,x’’) solves A G x , x x dx x x G x , x x x dx (a normal equation) A We can then say that x G x , x x dx q x A Note: q(x) = 0 and G(x,x’) = (x-x’-w) for x in the overlap A A’ where q(x) is the error in the conditional mean (x) - <(x)|(x’)>. q(x) is uncorrelated to the “data” ((x’)) q x x 0 Also true in the overlap since q(x) = 0 there q x s t 1 0 and, consequently since the measurement at t-1 depends only on (x’) and random measurement noise. Then t x G x , x t1 x dx i.e. F x, x Gx, x A IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 28 In summary: The time-progression of the conditional mean is w ˆt x F x, xˆt1 xdx A A A’ where F(x,x’) solves x x F x, x x x dx A x A, x A • If we assume the wavefront phase covariance function is constant or slowly varying with time, then the Green’s function F(x,x’) need only be computed infrequently (e.g. in slowly varying seeing conditions) • To solve this equation, we now need the cross-covariance statistics of the phase, piston-removed on two different apertures. IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 29 Cross-covariance of Kolmogorov phase, piston-removed on two different apertures x A x A A A’ 1 D x x g x c g x c a c c 2 x A, x A Where c and c’ are the centers of the respective apertures, and and g x 1 D x xWA xdx 2 ax 1 g x x W A x dx 2 as before also a “generic” function IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 30 The error covariance must also progress, since it is used in the update formulas using t x t x t x the error in the conditional mean is ~ t x F x, xt 1 xdx qt x ~ ~ A and the error covariance is t x t x ~ ~ ~ ~ F x , x x t 1 t 1 x F x , x dx dx Q x, x A A x, x A and x, x A where Q x, x qx q x x x F x, x x x F x, xdxdx A A Q is defined simply to preserve the Kolmogorov turbulence strength on the subsequent aperture IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 31 Simulations • Nominal parameters – D = 3m, d = 43cm (D/d = 7) – r0(l=0.5m) = 10cm ( r0(l=2m) d ) – w = 11m/s 1 ms (w = D/300) – Noise = 0.1 arcsec rms • Simulations – Wallner’s equations strictly applied, even though the wind is blowing – Strehl-optimal controller – Optimal controller with update matrix, K, set at converged value (allows precomputing error covariances) – Sensitivity to assumed r0 – Sensitivity to assumed wind speed – Sensitivity to assumed wind direction IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 32 Noise performance after convergence Single-step (Wallner) Strehl-optimal IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 33 Convergence time history K matrix fixed at converged value K matrix optimal at each time step IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 34 Sensitivity to r0 IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 35 Sensitivity to wind speed and direction IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 36 Conclusions • Kalman filtering techniques can be applied to better optimize the closed-loop Strehl of adaptive optics wavefront controllers • A-priori knowledge of r0 and wind velocity is required • Simulations show – Considerable improvement in performance over a single step optimized control law (Wallner) – Insensitivity to the exact knowledge of the seeing parameters over reasonably practical variations in these parameters IPAM Workshop on Estimation and Control Problems in Adaptive Optics, Jan., 2004 37