DEPARTMENT OF INFORMATICS March 2003 David Gesbert Signal and Image Processing Group (DSB) http://www.ifi.uio.no/~gesbert D. Gesbert: IN357 Statistical Signal Processing1 of 17 Course book: Chap. 7.2 Statistical Digital Signal Processing and modeling, M. Hayes 1996. IN357: WIENER FILTERING UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing2 of 17 • Application 3: Multi-antennas diversity combining • Application 2: Noise canceling • Application 1: Linear prediction • Finding the solution: The Wiener-Hopf equations • Geometrical interpretation • Optimum linear filtering • Example 1: temporal Wiener filtering • Purpose of Wiener filtering Outline UNIVERSITY OF OSLO .. DEPARTMENT OF INFORMATICS p−1 x (n) x 0(n) x2 (n) desired signal d(n) W filter error signal e(n) estimated signal D. Gesbert: IN357 Statistical Signal Processing3 of 17 ^ d(n) observed random process desired random process (unobserved) observed random process observed random process p observations {xp−1(n)} {d(n)} {x0(n)} {x2(n)} Goal: “To estimate an unknown random process (RP) from a set of observed RPs, with which it is correlated.” Purpose of Wiener filtering UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing4 of 17 • (2) one uses knowledge of correlation between {xi(n)} and d(n). • (1) one has a training signal for d(n) and one adjusts W to minimize the power of e(n). Two solutions to find the operator (filter) W : Purpose of Wiener filtering (II) UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing5 of 17 xi(n) = x(n − i), i = 0.., p − 1 W (z) = wo + w1z −1 + w1z −2 + .. + wp−1z −p+1 .. ˆ d(n) = w0x(n) + w1x(n − 1) + w2x(n − 2) + ... + wp−1x(n − p + 1) ⇒ {xi(n)}, i = 0, ..p − 1 are the samples of a wide sense stationary (WSS) random signal. ⇒ W is a linear FIR filter. Example 1: Linear temporal filtering UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS where T is the transpose operator. ˆ d(n) = W T X(n) D. Gesbert: IN357 Statistical Signal Processing6 of 17 W = [w0, w1, .., wp−1]T X(n) = [x0(n), x1(n), .., xp−1(n)]T Vector Formulation UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing7 of 17 Find Wo such that J(Wo) is minimum. Wo is the optimum linear filter in the Wiener sense. where E() is the expectation. ˆ e(n) = d(n) − d(n) J(W ) = E|e(n)|2 Problem: One wishes to find the linear filter W that minimizes the error ˆ between d(n) and d(n). Optimum linear filtering UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS x0 (n) 1 x (n) • Orthogonality = decorrelation d(n) ^ d(n) e(n) D. Gesbert: IN357 Statistical Signal Processing8 of 17 p=2 • The scalar product between two RPs {u(n)} and {v(n)} is defined by the correlation: < u, v >= E(u(n)v(n)∗). • WSS random processes ({e(n)}, {d(n)}, {x0(n)}, .., {xp−1(n)} can be viewed as elements in a vector space. Vector space representation Geometrical Interpretation UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS (2) (3) (1) D. Gesbert: IN357 Statistical Signal Processing9 of 17 Rx = E(X(n)∗X(n)T ) rdx = E(d(n)X(n)∗) RxWo = rdx where The solution Wo is given by the Wiener-Hopf equations. Finding the solution UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS o ∂J ∗ ∂wp−1 |W =W .. ∂J ∂w0∗ |W =Wo ∂J ∂w1∗ |W =Wo (4) D. Gesbert: IN357 Statistical Signal Processing10 of 17 0 0 =. . 0 Direct method using optimization theory: The extrema of the cost function J(W ) are found from (see book by Hayes p.49): Derivation of the Wiener-Hopf equation UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS x0 (n) 1 x (n) d(n) ^ d(n) e(n) (5) D. Gesbert: IN357 Statistical Signal Processing11 of 17 p=2 < e(n), x0(n) >= E(e(n)x0(n)∗) = 0 < e(n), x1(n) >= E(e(n)x1(n)∗) = 0 .. < e(n), xp−1(n) >= E(e(n)xp−1(n)∗) = 0 • {e(n)} is orthogonal to the space spanned by {x0(n)}, .., {xp−1(n)}. ˆ • {d(n)} is the projection of {d(n)} onto the space spanned by {x0(n)}, .., {xp−1(n)}, or equivalently: ˆ {d(n)} is the optimal estimate iff Derivation using geometry UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing12 of 17 H Wo J(Wo) = E|d(n)|2 + WoH RxWo − WoH rdx − rdx H J(Wo) = E|d(n)|2 − rdx Wo Wo The minimum error, obtained with Wo, is J(Wo): Expression for the minimum error UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS Hayes’ book, p. 342 D. Gesbert: IN357 Statistical Signal Processing13 of 17 Application 1: Linear prediction UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS Hayes’ book, p. 349 D. Gesbert: IN357 Statistical Signal Processing14 of 17 Application 2: Noise canceling UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS s 2 1 h h 2 2 w =h* 1 w =h*1 y D. Gesbert: IN357 Statistical Signal Processing15 of 17 Weights phases estimation Application 3: Multi-antenna diversity combining UNIVERSITY OF OSLO X(n) = Hs(n) + V (n) DEPARTMENT OF INFORMATICS (6) D. Gesbert: IN357 Statistical Signal Processing16 of 17 E|WoT X(n) − s(n)|2 minimum Problem: Find Wo such that where X(N ) collects the signals received on p antennas. s(n) is the transmitted symbol sequence (e.g. s(n) = +/ − 1). His the channel vector of size p. V (n) is the white noise vector. Received signal model: Antenna diversity combining (II) UNIVERSITY OF OSLO DEPARTMENT OF INFORMATICS D. Gesbert: IN357 Statistical Signal Processing17 of 17 NOTE: This receiver is equivalent to the minimum mean square error (MMSE) receiver. where σv2 is the variance of noise samples. Wo = (H ∗H T + σv2I)−1H ∗ RxWo = rdx (H ∗H T + σv2I)Wo = H ∗ Answer (Wiener-Hopf ): Antenna diversity combining (III) UNIVERSITY OF OSLO