RLSELE Adaptive Signal Processing 1 Recursive Least-Squares (RLS) Adaptive Filters
ELE Adaptive Signal Processing2 RLS Definition With the arrival of new data samples estimates are updated recursively. Introduce a weighting factor to the sum-of-error-squares definition Weighting factor Forgetting factor : real, positive, <1, →1 =1 → ordinary LS 1/(1- ): memory of the algorithm (ordinary LS has infinite memory) w(n) is kept fixed during the observation interval 1≤i ≤n for which the cost function (n) is defined. two time-indices n: outer, i: inner
ELE Adaptive Signal Processing3 RLS Definition
ELE Adaptive Signal Processing4 RLS Regularisation LS cost function can be ill-posed There is insufficient information in the input data to reconstruct the input-output mapping uniquely Uncertainty in the mapping due to measurement noise. To overcome the problem, take ‘prior information’ into account Prewindowing is assumed! (not the covariance method) Regularisation term Smooths and stabilises the solution : regularisation parameter
ELE Adaptive Signal Processing5 RLS Normal Equations From method of least-squares we know that then the time-average autocorrelation matrix of the input u(n) becomes Similarly, the time-average cross-correlation vector between the tap inputs and the desired response is (unaffected from regularisation) Hence, the optimum (in the LS sense) filter coefficients should satisfy autocorrelation matrix is always non-singular due to this term. ( -1 always exists!)
ELE Adaptive Signal Processing6 RLS Recursive Computation Isolate the last term for i=n: Similarly We need to calculate -1 to find w → direct calculation can be costly! Use Matrix Inversion Lemma (MIL)
ELE Adaptive Signal Processing7 RLS Recursive Least-Squares Algorithm Let Then, using MIL Now, letting We obtain inverse correlation matrix gain vector Riccati equation
ELE Adaptive Signal Processing8 RLS Recursive Least-Squares Algorithm Rearranging How can w be calculated recursively? Let After substituting the recursion for P(n) into the first term we obtain But P(n)u(n)=k(n), hence
ELE Adaptive Signal Processing9 RLS Recursive Least-Squares Algorithm The term is called the a priori estimation error, Whereas the term is called the a posteriori estimation error. (Why?) Summary; the update eqn. -1 is calculated recursively and with scalar division Initialisation: (n=0) If no a priori information exists gain vector a priori error regularisation parameter
ELE Adaptive Signal Processing10 RLS Recursive Least-Squares Algorithm
ELE Adaptive Signal Processing11 RLS Recursive Least-Squares Algorithm
ELE Adaptive Signal Processing12 RLS Recursion for the Sum-of-Weighted-Error-Squares From LS, we know that where Then Hence
ELE Adaptive Signal Processing13 RLS Convergence Analysis Assume stationary environment and =1 To avoid transitions, consider times n>M Assumption I: The desired response d(n) and the tap-input vector u(n) are related by the linear regression model where w o is the regression parameter vector and e o (n) is the measurement noise. The noise e o (n) is white with zero mean and variance o 2 which makes it independent of the regressor u(n).
ELE Adaptive Signal Processing14 RLS Convergence Analysis Assumption II: The input vector u(n) is drawn from a stochastic process, which is ergodic in the autocorrelation function. R: ensemble average, : time average autocorrelation matrices Assumption III: The fluctuations in the weight-error vector (n) are slow compared with those of the input signal vector u(n). Justification: (n) is an accumulation of the a priori error → hence, the input →Smoothing (low-pass filtering) effect. Consequence:
ELE Adaptive Signal Processing15 RLS Convergence in Mean Value Then, Substituting into w(n) and taking the expectation, we get Applying Assumptions I and II, above expression simplifies to biased estimate due to the initialization, but bias →0 as n→∞. =1
ELE Adaptive Signal Processing16 RLS Mean-Square Deviation Weight-error correlation matrix and invoking Assumption I and simplifying we obtain Then But, mean-square-deviation is
ELE Adaptive Signal Processing17 RLS Mean-Square Deviation Observations: Mean-Square Deviation D (n) is proportional to the sum of reciprocal of eigenvalues of R The sensitivity of the RLS algorithm to eigenvalue spread is determined by the reciprocal of the smallest eigenvalue. ill-conditioned LS problems may lead to poor convergence behaviour. decays almost linearly with the number of iterations w(n) converges to the Wiener solution w o as n grows. ^
ELE Adaptive Signal Processing18 RLS Ensemble-Average Learning Curve There are two error terms A priori error, A posteriori error, Learning curve considering (n) yields the same general shape as that for the LMS algorithm. Both RLS and LMS learning curves can be compared with this choice. The learning curve for RLS (a posteriori error) is We know that
ELE Adaptive Signal Processing19 RLS Ensemble-Average Learning Curve Substitution yields 1 st term (Assumption I) 2 nd term (Assumption III) 3 & 4 th terms (Assumption I)
ELE Adaptive Signal Processing20 RLS Ensemble-Average Learning Curve Combining all terms Observations The ensemble-average learning curve of the RLS algorithm converges in about 2M iterations Typically an order of magnitude faster than LMS As the number of iterations n→∞ the MSE J’(n) approaches the final value σ o 2 which is the variance of the measur. error e o (n). in theory RLS produces zero excess MSE!. Convergence of the RLS algorithm in the mean square is independent of the eigenvalues of the ensemble-average correlation matrix R of the input vector u(n).
ELE Adaptive Signal Processing21 RLS Ensemble-Average Learning Curve