Advanced Signal Processing Minimum Variance Unbiased Estimation (MVU)

Transcription

1 Advanced Signal Processing Minimum Variance Unbiased Estimation (MVU) Danilo Mandic room 813, ext: Department of Electrical and Electronic Engineering Imperial College London, UK URL: mandic c D. P. Mandic Advanced Signal Processing 1

2 Objectives Learn the concept of minumum variance unbiased (MVU) estimation Investigate how the accuracy of an estimator depends upon the relationship between the unknown parameter(s) and the PDF of noise Study the requirements for the design of an efficient estimator Analyse the Cramer Rao Lower Bound (CRLB) for the scalar case Extension to the Cramer Rao Lower Bound (CRLB) for the vector case Sinusoid parameter estimation, linear models Dependence on data length (motivation for sufficient statistics ) Examples: DC level in WGN (frequency estimation in power, bioengineering) Parameters of a sinusoid (scalar case, vector case) A new view of Fourier analysis System identification c D. P. Mandic Advanced Signal Processing 2

3 Example 1: Consider a single observation x[0] = A + w[0], where w[0] N(0,σ 2 ) The simplest estimator of the DC level A in white noise w[0] N(0,σ 2 ) is  = x[0]  is unbiased, with the variance σ2 To show that the estimator accuracy improves as σ 2 decreases: Consider [ ] p i (x[0];a) = 1 exp 1 (x[0] A) 2 2πσ 2 2σ 2 i i for x[0] = 3 and i = 1,2 with σ }{{} 1 = 1 3 and σ 2 = 1 f undamental step, we are f ixing the data value p (x[0]=3;a) 1 p (x[0]=3;a) 2 Case I σ 1=1/3 Case II σ 2=1 3 A 3 Clearly, as σ 1 < σ 2, the DC level A is estimated more accurately with p 1 (x[0];a) Likely candidates for values of A 3 ± 3σ therefore [2, 4] for σ 1 and [0, 6] for σ 2. A c D. P. Mandic Advanced Signal Processing 3

4 Likelihood function When the PDF is viewed as a function of an unknown parameter (with the dataset {x} = x[0],x[1],... fixed) it is termed the likelihood function. The sharpness of the likelihood function determines the accuracy at which the unknown parameter may be estimated. Sharpness is measured by the curvature a negative of the second derivative of the logarithm of the likelihood function at its peak. Example 2: Estimation based on one sample of a DC level in WGN lnp(x[0];a) = ln 2πσ 2 1 2σ 2 ( x[0] A ) 2 then ln p(x[0]; A) A = 1 σ 2 ( x[0] A ) and the curvature 2 lnp(x[0];a) A 2 = 1 σ 2 Therefore, as expected, the curvature increases as σ 2 decreases. c D. P. Mandic Advanced Signal Processing 4

5 Likelihood function: Curvature Since we know that the variance of the estimator equals σ 2, then var(â) = 1 2 lnp(x[0];a) A 2 and the variance decreases as the curvature increases. Generally, the second derivative does depend upon x[0], and hence a more appropriate measure of curvature is the statistical measure [ 2 ] lnp(x[0];a) E A 2 which measures the average curvature of the log-likelihood function Recall: The likelihood function is a random variable due to x[0] Recall: The Mean Square Error MSE = Bias 2 + variance Ê it makes perfect sense to look for a minimum variance unbiased (MVU) solution c D. P. Mandic Advanced Signal Processing 5

6 Link with human perception In the 50s, phsychologist Fred Attneave recorded eye dwellings on objects Example 3: The drawing of a bean (top) andthehistogramofeyedwellings(bottom) Example 4: Read the words below... now read letter by letter... are you still sure? Example 3: Is the drawing on the left still a penguin? So, what is the sufficient information to estimate an object? c D. P. Mandic Advanced Signal Processing 6

7 THE KEY: Cramer-Rao Lower Bound (CRLB) for a scalar parameter (performance of the theoretically best estimator) The Cramer Rao Lower Bound (CRLB) Theorem: [CRLB] Assumption: The PDF p(x; θ) satisfies the regularity condition [ ] lnp(x;θ) E = 0, θ θ where the expectation is taken with respect to p(x; θ). Then, the variance of any unbiased estimator, ˆθ, must satisfy var(ˆθ) E 1 { } 2 lnp(x;θ) θ 2 where the derivative is evaluated at the true value of θ. c D. P. Mandic Advanced Signal Processing 7

8 CRLB for a scalar parameter, continued Moreover, an unbiased estimator may be found that attains the bound for all θ, if and only if for some functions g and I lnp(x;θ) = I(θ) ( g(x) θ ) θ This estimator is the minimum variance unbiased (MVU) estimator, for which ˆθ = g(x) and its minimum variance 1 I(θ) end of CRLB theorem 1 Remark: Since the variance var(ˆθ) { E 2 lnp(x;θ) θ 2 the curvature term gives [ 2 ] lnp(x;θ) 2 lnp(x;θ) E θ 2 = θ 2 p(x;θ)dx }, the evaluation of Obviously, in general the bound depends on the parameter θ and the data length c D. P. Mandic Advanced Signal Processing 8

9 Example 5: Estimation of a DC level in WGN Consider the estimation of a DC level in WGN, assume N observations x[n] = A }{{} unknown DC level + w[n] }{{} noise with known pdf n = 0,1,2,...,N 1 where w[n] N(0,σ 2 ). Determine the CRLB for the unknown DC level A, starting from (θ = A) p(x;θ) = p(x;a) = = N 1 n=0 1 (2πσ 2 ) N/2exp [ 1 exp 1 ] ( ) 2 x[n] A 2πσ 2 2σ 2 [ 1 2σ 2 N 1 n=0 ( x[n] A ) 2 ] Estimation of a DC level is very useful, e.g. in the time-frequency plane a sinusoid of frequency f is represented by a straight line Ê c D. P. Mandic Advanced Signal Processing 9

10 Example 5: DC level in WGN continued Taking the first derivative, we have [ lnp(x;a) = ln [ 2πσ 2] N/2 1 A A 2σ 2 = 1 N 1 σ 2 n=0 where x denotes the sample mean. Connection with CRLB: x = 1 N Upon differentiating again N 1 n=0 ( x[n] A ) = N σ 2 ( x A ) N 1 n=0 2 lnp(x;a) A 2 = N σ 2 ( x[n] A ) 2 ] x(n) = g(x), I(A) = N/σ2 Therefore var(â) σ2 N is the CRLB, which implies that the sample mean estimator attains the Cramer-Rao lower bound and must, therefore, be an MVU estimator in WGN. c D. P. Mandic Advanced Signal Processing 10

11 Efficient estimator concept An estimator which is unbiased and attains the CRLB is said to be efficient in that it is an Minimum Variance Unbiased (MVU) etimator and that it efficiently uses the data. var( ^ θ ) var( θ^ ) CRLB θ^ 2 θ^ 3 θ^ 1 CRLB θ^ 2 ^ θ 3 ^ θ 1 ˆθ 1 is efficient and MVU, ˆθ 2,ˆθ 3 are not θ ˆθ1 may be MVU but is not efficient θ c D. P. Mandic Advanced Signal Processing 11

12 Fisher information The denominator in the CRLB I(θ) = E is referred to as the Fisher information. [ 2 ] lnp(x;θ) Intuitively: the more information available the lower the bound less variance θ 2 Essential properties of an information measure: Ê Non negative Additive for independent observations General CRLB for arbitrary signals in WGN (see the next slide) Ê var(ˆθ) N 1 n=0 σ 2 ( ) 2 s[n;θ] θ Accurate estimators: signals change rapidly with the parameter changes. c D. P. Mandic Advanced Signal Processing 12

13 General case: Arbitrary signal in noise Consider a deterministic signal s[n;θ] observed in WGN, w N(0,σ 2 ) x[n] = s[n;θ]+w[n], n = 0,1,...,N 1 Then the PDF for x parametrised by θ has the form 1 1 N 1 p(x;θ) = (2πσ 2 ) N/2e 2σ 2 n=0 so that ln p(x;θ) θ = 1 N 1 σ 2 2 ln p(x;θ) θ 2 = 1 σ 2 and the Fisher information n=0 N 1 n=0 ( x[n] s[n;θ] ( x[n] s[n;θ] ) s[n;θ] θ [ ( ) 2 s[n;θ] x[n] s[n;θ] }{{} θ 2 E{x[n]}=s[n;θ]} I(θ) = E [ 2 ln p(x;θ) θ 2 ] = 1 σ 2 N 1 n=0 ) 2 ( s[n;θ] θ ( s[n;θ] ) ] 2 θ ) 2 c D. P. Mandic Advanced Signal Processing 13

14 Example 6: Sinusoidal frequency estimation (the CRLB depends both on the unknown parameter f 0 and the data length N) Consider a general sinewave in noise: x[n] = Acos ( 2πf 0 n+φ ) +w[n] If only the freqeuncy f 0 is unknown, then (for normalised frequency) and s[n;f 0 ] = var(ˆf 0 ) A }{{} known A 2 N 1 n=0 cos(2πf 0 n+ Φ }{{} known), 0 < f 0 < 1 2 σ 2 [ 2πn sin(2πf0 n+φ) ] 2 Cramer Rao lower bound 5 x Frequency Note the preferred frequencies, e.g. f 0.03, and that for f 0 {0,1/2} the CRLB Paramet.: N = 10, Φ = 0, SNR=A 2 /σ 2 = 1 c D. P. Mandic Advanced Signal Processing 14

15 Extension to a vector parameter we now have the Fisher Information Matrix I, s.t. [I(θ)] ij = E [ 2 ] ln p(x;θ) θ i θ j Formulation: Estimate a vector parameter θ = [θ 1,θ 2,...,θ p ] T Recall that an unbiased estimator ˆθ is efficient (and therefore an MVU estimator) when it satisfies the conditions of the CRLB It is assumed that the PDF p(x; θ) satisfies the regularity conditions [ ] lnp(x;θ) E = 0, θ θ Then the covariance matrix of any unbiased estimator ˆθ satisfies Cˆθ I 1 (θ) 0 (symbol 0 means that Cˆθ is positive semidefinite) [ ] The Fisher Information Matrix is given by [I(θ)] ij = E 2 lnp(x;θ) θ i θ j An unbiased estimator ˆθ = g(x) exists that satisfies the bound Cˆθ = I 1 (θ) if and only if lnp(x;θ) = I(θ) ( g(x) θ ) θ Ê c D. P. Mandic Advanced Signal Processing 15

16 Extension to a vector parameter: Fisher information Some observations: Elements of the Information Matrix I(θ) are given by [I(θ)] ij = E [ 2 ] lnp(x;θ) θ i θ j where the derivatives are evaluated at the true values of the parameter vector. The CRLB theorem provides a powerful tool for finding MVU estimators for a vector parameter. MVU estimators for linear models are found with the Cramer Rao Lower Bound (CRLB) theorem. Ê c D. P. Mandic Advanced Signal Processing 16

17 Example 7: Sinusoid parameter estimation vector case Consider again a general sinewave s[n] = Acos ( 2πf 0 n+φ ) where A, f 0 and Φ are all unknown. Then, the data model becomes x[n] = A cos(2πf 0 n+φ)+w[n] n = 0,1,...,N 1 where A > 0, 0 < f 0 < 1/2, and w[n] N(0,σ 2 ). Task: Determine CRLB for the parameter vector θ = [A f 0 Φ] T. Solution: The elements of the Fisher Information Matrix become N/2 0 0 I(θ) = 1 0 2A σ 2 2 π 2 N 1 n=0 n2 πa N 1 n=0 n 0 πa N 1 n=0 n NA 2 2 c D. P. Mandic Advanced Signal Processing 17

18 Example 7: Sinusoid parameter estimation continued since Cˆθ = I 1 (θ) (slide 15) make an inverse of the FIM After inversion, the diagonal components yield (where η = A2 2σ 2 is SNR): var(â) 2σ2 N var(ˆf 0 ) 12 (2π) 2 ηn(n 2 1) var(ˆφ) CRLB for Sinusoidal Parameter Estimates at SNR = 3dB (Dashed Lines) and 10dB (Solid Lines) 2(2N 1) ηn(n +1) 0 Amplitude Frequency Phase CRLB [db] Number of data samples, N the variance of the estimated parameters of a sinusoid behaves 1/η and 1/N 3, thus exhibiting strong sensitivity to data length Ê c D. P. Mandic Advanced Signal Processing 18

19 Linear models Generally, it is difficult to determine the MVU estimator. In signal processing, however, a linear data model can often be employed straightforward to determine the MVU estimator. Example 8: Linear model of a straight line in noise x[n] = A+Bn+w[n] n = 0,1,...,N 1 x[n] noisy line where w[n] N(0,σ 2 ), B - slope and A - intercept. A 0 ideal noiseless line n c D. P. Mandic Advanced Signal Processing 19

20 Linear models: Compact notation (Example 8 contd.) This data model can be written more compactly in the matrix notation as where x = and x[0] x[1]. x[n 1] x = Hθ +w or x = Hθ +w = [ x[0],x[1],...,x[n 1] ] T θ = [A B] T w = [w[0],w[1],...,w[n 1]] T w N(0,σ 2 I) I = H = = diag(1,1,...,1) N 1 c D. P. Mandic Advanced Signal Processing 20

21 Linear models: Fisher information matrix NB: p(x;θ) = 1 (2πσ 2 ) N/2e (x Hθ) The CRLB theorem can be used to obtain the MVU estimator for θ The MVU estimator, ˆθ = g(x), will satisfy T (x Hθ) 2σ 2 lnp(x;θ) θ = I(θ) ( g(x) θ ) where I(θ) is the Fisher information matrix, whose elements are [ 2 ln p(x;θ) ] [I] ij = E θ i θ j Applying the Linear Model lnp(x;θ) = [ N θ θ 2 ln(2πσ2 ) 1 ] ( ) T ( ) x Hθ x Hθ 2σ 2 = 1 2σ 2 θ [ Nσ 2 ln(2πσ 2 )+x T x 2x T Hθ +θ T H T Hθ Note that only the quadratic term in θ involves the matrix H ] c D. P. Mandic Advanced Signal Processing 21

22 Linear models: Some useful matrix/vector derivatives the derivation is given in the Lecture Supplement Use the identities (remember that both b T θ and θ T Aθ are scalars) b T θ θ θ T Aθ θ = b xt Hθ θ = 2Aθ θt H T Hθ θ = (x T H) T = H T x = 2H T Hθ (which you should prove for yourself), that is, follow the rules of vector/matrix differentiation. As a rule of thumb, watch for the position of the ( ) T operator Then, for A symmetric: lnp ( x;θ ) θ = 1 σ 2 [ H T x H T Hθ ] c D. P. Mandic Advanced Signal Processing 22

23 Linear models: Cramer-Rao lower bound lnp(x;θ) Find the MVU estimator: θ = I(θ) ( g(x) θ ) Similarly (recall that (H T H) T = H T H), I(θ) = T θ [ ] lnp(x;θ) = 1 H θ σ 2HT Therefore lnp(x;θ) θ = 1 σ 2HT H }{{} I(θ) [ (H T H ) ] 1 H T x θ }{{} g(x) By inspection, the linear estimator ˆθ = g(x) = ( H T H ) 1 H T x provided ( H T H ) 1 is invertible (it is, as H is full rank, with orthogonal rows and columns). The covariance matrix of ˆθ now becomes Cˆθ = I 1 (θ) = σ 2( H T H ) 1 c D. P. Mandic Advanced Signal Processing 23

24 CRLB for linear models continued The MVU estimator for the linear model is efficient it attains the CRLB The columns of H must be linearly independent for ( H T H ) to be invertible Theorem: (Minimum Variance Unbiased Estimator for the Linear Model) If the observed data can be modelled as x = Hθ +w where x is an N 1 vector of observed data H is an N p observation (measurement) matrix of rank p θ is a p 1 unknown parameter vector w is an N 1 additive noise vector N ( 0,σ 2 I ) c D. P. Mandic Advanced Signal Processing 24

25 CRLB for linear models: Theorem Then, the MVU estimator is then given by ˆθ = ( H T H ) 1 H T x for which the covariance matrix Cˆθ = σ2( H T H ) 1 Note that the statistical performance of ˆθ is completely described because ˆθ is a linear transformation of a Gaussian vector x, i.e. ˆθ N (θ,σ 2( H T H ) 1 ) c D. P. Mandic Advanced Signal Processing 25

26 Example 9: Fourier analysis The data model is given by (n = 0,1,...,N 1, w[n] N ( 0,σ 2) ) x[n] = M ( ) 2πkn a k cos + N k=1 M ( ) 2πkn b k sin +w[n] N where the Fourier coefficients, a k and b k, that is the amplitudes of the cosine and sine terms, are to be estimated. k=1 Frequencies are harmonically related, i.e. f 1 = 1 N, and f k = k N. Then, the parameter vector is θ = [a 1,a 2,...,a M,b 1,b 2,...,b M ] T and the observation matrix H ( N }{{} 2M dimensional ) takes the form H = cos 2π N... cos 2πM N sin 2π N... sin 2πM N cos 2πM(N 1) N sin 2π(N 1) N... sin 2πM(N 1) N cos 2π(N 1) N p N 2M c D. P. Mandic Advanced Signal Processing 26

27 Example 9: Fourier analysis continued For H not to be under-determined, it has to satisfy N > p M < N 2, and the columns of H have to be orthogonal. This is obvious if we rewrite the measurement matrix in the form H = [h 1,h 2,...,h 2M ] where h i = h i is the i-th column of H. Then (products of sines and cosines of different frequencies) h T i h j = 0 for i j That is, h i h j, and the columns of matrix H are orthogonal c D. P. Mandic Advanced Signal Processing 27

28 Example 9: Fourier analysis contd. contd. The orthogonality of the columns (for large N) follows from the discrete Fourier transform (DFT) relationships N 1 ( ) ( ) 2πin 2πjn cos cos = N N N 2 δ ij n=0 N 1 n=0 N 1 n=0 where ( ) ( ) 2πin 2πjn sin sin N N ( ) ( ) 2πin 2πjn cos sin N N δ ij = = N 2 δ ij = 0 i,j, s.t. i,j = 1,2,...,M < N 2 { 1, i = j 0, i j In other words: (i) cosiα sinjα, i,j, (ii) cosiα cosjα, i j, (iii) siniα sinjα, i j c D. P. Mandic Advanced Signal Processing 28

29 Example 9: Fourier analysis observation matrix Therefore H T H = ht 1. h T 2M [ h 1,...,h 2M ] = N N N = N 2 I The MVU estimator of the amplitudes of the cosine and sine terms (Fourier coefficients) is therefore given by ˆθ = ( H T H ) 1 H T x ˆθ = 2 N HT x = 2 N h T 1. h T 2M x = that is 2 N ht 1x. 2 N ht 2M x and finally, the estimators of Fourier coefficients become â k = 2 N N 1 n=0 x[n]cos( ) 2πkn N ˆbk = 2 N 1 N n=0 x[n]sin( ) 2πkn N c D. P. Mandic Advanced Signal Processing 29

30 Example 9: Finally Fourier coefficients Therefore, the Fourier analysis is a linear estimator, given by â k = 2 N 1 N n=0 x[n]cos( ) 2πkn N ˆbk = 2 N 1 N n=0 x[n]sin( ) 2πkn N where the a k and b k are discrete Fourier transform coefficients. From CRLB for Linear Model, the covariance matrix of this estimator is Cˆθ = σ2( H T H ) 1 = 2σ 2 N I i) Note that, as ˆθ is a Gaussian random variable and the covariance matrix is diagonal, the amplitude estimates are statistically independent; ii) The orthogonality of the columns of H is fundamental in the computation of the MVU estimator. c D. P. Mandic Advanced Signal Processing 30

31 Example 10: System Identification (SYS ID) Aim: To identify the model of a system (filter coefficients {h}) from input/output data. Assume an FIR filter system model given below u[ n ] 1 Z 1 Z Z 1 h (0) x h (1) x h ( p 1) x Σ Σ p 1 Σ h(k) k =0 u[n k] The input u[n] probes the system, then the output of the FIR filter is given by the convolution x[n] = p 1 k=0 h(k)x[n k] We wish to estimate the filter coefficients [h(0),...,h(p 1)] T In practice, the output is corrupted by additive WGN c D. P. Mandic Advanced Signal Processing 31

32 Example 10: SYS ID data model in noise w N(0,σ 2 ) Data model x[n] = p 1 k=0 Equivalently, in the matrix vector form x[0] x[1]. = x[n 1] } {{ } obs. vec. x that is h(k)u[n k]+w[n] n = 0,1,...,N 1 u[0] u[1] u[0] u[n 1] u[n 2]... u[n p] } {{ } measurement matrix H h(0) h(1). h(p 1) } {{ } coeff. vec. θ x = Hθ +w where w N(0,σ 2 I) The MVU estimator ˆθ = ( H T H ) 1 H T x with Cˆθ = σ 2( H T H ) 1 w[0] + w[1]. w[n 1] } {{ } noise vec. w This representation also lends itslef to state-space modelling c D. P. Mandic Advanced Signal Processing 32

33 Example 10: SYS ID more about H Now, H T H becomes a symmetric Toeplitz autocorrelation matrix, given by r uu (0) r uu (1)... r uu (p 1) H T H = N r uu (1) r uu (0)... r uu (p 2) r uu (p 1) r uu (p 2)... r uu (0) where r uu (k) = 1 N N 1 k n=0 u[n]u[n+k] For H T H to be diagonal, we must have r uu (k) = 0 for k 0, which holds for a pseudorandom (PRN) input sequence. Finally, when H T H = N r uu (0)I then var ( ĥ(i) ) = σ 2 N r uu (0), i = 0,1,...,p 1 c D. P. Mandic Advanced Signal Processing 33

34 Example 10: SYS ID MVU estimator For a PRN sequence, the MVU estimator becomes ˆθ = ( H T H ) 1 H T x Then ĥ(k) = and r ux (k) r uu (0) = 1 N r uu (0) 1 N N 1 n=0 u[n k]x[n] N 1 k n=0 u[n]x[n + k] r uu (0) k = 0,1,...,p 1 Thus, the MVU estimator is the ratio of the input-output cross-correlation to the input autocorrelation. ÊCompare with the Wiener filter in Lecture 7. c D. P. Mandic Advanced Signal Processing 34

35 Theorem: The MVU Estimator for a General Linear Model (GLM) i) Data model x = Hθ + }{{} s +w known signal w N ( 0,C ) ii) Then, the MVU estimator ˆθ = ( H T C 1 H) 1 H T C 1( x s ) iii) with covariance matrix Cˆθ = ( H T CH) 1 c D. P. Mandic Advanced Signal Processing 35

36 What to remember about MVU estimators An estimator is a random variable and as such its performance can only be described statistically by its PDF The use of computer simulations for assessing the performance of an estimator is rarely conclusive Unbiased estimators tend to have symmetric PDFs centred about the true value of θ The minimum mean square error (MMSE) criterion is natural to search for optimal estimators, but it most often leads to unrealisable estimators (those that cannot be written solely as a function of data) Since MSE= Bias 2 + variance, any criterion that depends on bias should be abandoned we need to consider alternative approaches Remedy: Constrain the bias to zero and find an estimator which minimises the variance the minimum variance unbiased (MVU) estim. Minimising the variance of an unbiased estimator also has the effect of concentrating the PDF of the estimation error, ˆθ θ, about zero estimation error less likely to be large c D. P. Mandic Advanced Signal Processing 36

37 A few things about CRLB to remember Even if the MVU estimator exists, there is no turn of the crank procedure to find it. The CRLB sets a lower bound on the variance of any unbiased estimator! This can be extremely useful in several ways: If we find an estimator that achieves the CRLB we known we have found an MVU estimator The CRLB can provide a benchmark against which we can compare the performance of any unbiased estimator The CRLB enables us to rule out impossible estimators. It is physically impossible to find an unbiased estimator that beats the CRLB We may require the estimator to be linear, which is not necessarily a severe restriction, as shown in the example on the estimation of Fourier coefficients c D. P. Mandic Advanced Signal Processing 37

38 Notes: c D. P. Mandic Advanced Signal Processing 38

39 Notes: c D. P. Mandic Advanced Signal Processing 39

40 Notes: c D. P. Mandic Advanced Signal Processing 40