# Data Normalized Filters

date post

02-Jun-2018Category

## Documents

view

220download

0

Embed Size (px)

### Transcript of Data Normalized Filters

8/11/2019 Data Normalized Filters

1/9

w e xtend the transient analysis of the earlier chaptersto mo re general filter recursions,starting with the normalizedLMS algorithm.

25.1 NLMSFILTER

We thus consider the6-NLMSrecursion

for which the data normalization in(22.2) is given by

S[Ui] + l lUiIl2

(25.1)

(25.2)

In this case, relations(22.26)-(22.27)and (22.29)become

and we see that we need to evaluate the moments

Unfortunately, closed form expressions for these moments are not available in general,even for Gauss ian regressors. Still, we w ill be able to show that the filter is convergent inthe mean and is also mean-square stable fo r step-sizes satisfyingp < 2, and regardless

of the input distribution (Gaussian or otherwise)- ee App. 25.B. We therefore treat thegeneral case directly. Since the argum ents are similar to those in Chapter24 for LMS, weshall be brief.

Thus, introduce theM 2 x 1 vectors

A Au = vec(C), r = vec(R,)

371

Adaptive Filters, by Ali H. ayedCopyright @ 2008 John Wiley Sons, Inc

8/11/2019 Data Normalized Filters

2/9

as well as theM 2 x M 2 matrices72

wi

CHAPTER25DATA-NORMALIZEDFILTERS

- E llwillzE I IwiIILE I I ~ i I I $ z u (25.1 1)

E I I ~ i l l F ( M Z - l ) u2

= ( ;12T @ E :,12)]and theM x M matrix

25.5)

25.6)

The matrixA is positive-definite, whileB is nonnegative-definite ee Prob.V.6. pply-ing the vec notation to both side s of the ex pression forC' in (25.3) we find that it reducesto

d F r 25.7)

where F is M 2 x M 2 and given by

AF = I - p A + p 2 B 25.8)

Moreover, the recursion forEw an be w ritten as

The same argum ents that were used in Sec. 24.1 will show that the mean-square behaviorof E - N L M S is characterized by the following M2-dim ensional state-space model:

where s the companion matrix

=

with

0 1

0 0 10 0 0 1

0 0 0 1-Po -p1 -p2 . . . -PMz-l

M 2 x M 2 )

Ma-1

p(z) = det(z1- F ) = xM2+ pkxk O

8/11/2019 Data Normalized Filters

3/9

and the k-th entry ofY is given by 373ECTION25.2

NLMSFILTERk = 0 ' 1 , . . ,M 2 - 1 (25.12)

The definitions of{Wi, Y } re in terms of any of interest, e.g., most commonly,= qor u = r . It is shown in App. 25.B that anyp < 2 is a sufficient condition for the stabilityof (25.10).

Theorem 25.1 (Stability of E-NLMS)Consider the E-NLMS recursion (25.1)and assume the data { d ( i ) , u i } atisfy model (22.1) and the independenceassumption (22.23). Th e regressors need not be Gaussian. Then the filt eris convergent in the mean and is also mean-square stable for any p < 2.Moreover, the transient behavior of the filter is characterized by the state-space recursion (25.10)-(25.12), and th e mean-square dev iati on and the ex-cess mean-square error are given by

where r = vec(R,) and q = vec(1).

The expressions for theMSD and EMSE in the statem ent of the theo rem are derived ina manner sim ilar to (23.51 ) and (23.56). They can be rewritten as

~ MSD = p20:Tr(SCm,d) and EMSE = p 2 0 : T r ( S E e m q~

where

)A E ( u f u

6 + llu 112)2and the weigh ting matrices {Cmsd,C correspond to the vectorsUmsd = (I - F ) - l qand g e m s = (I F ) - r . That is,

Cmsd = vec-l(um,d) and c,,,, = vec-'(oemSe)

Learning CurveObserve that sinceE lea i)I2 = E l l t & - . l ~ l ~ u ,e again find that the time evolutionofE iea i)12 s described by the top entryof the state vectorWi in (25.10)-(25.12) withochosen asu = r . The lea rning curve of the filter will beE le i)i2 = 0, E lea i)12.

Small Step-Size ApproximationsSeveral approxim ations for theEMSE and MSD expression s that appe ar in the above theo-rem are derived in Prob s. V.18-V.20. The ultima te conclusion from these problems is thatfor small enough,Y and E , we get

11EMSE = &E (m) r(R,) and MSD = B E m)2 P 2 P

I

(25.13)The exp ression for theEMSE is the same we derived earlier in Lemm a 17 .1.

8/11/2019 Data Normalized Filters

4/9

74 Gaussian RegressorsCHAPTER25NORMALIZEDFILTERS

If the regressors happen to be Gaussian, thenit can be shown that the M2-dimensionalstate-space m odel(25.10)-(25.12) reduces to an M-dimensional model- his assertionis proved in Probs.V.16 and V.17.

DATA-

25.2 DATA-NORMALIZED FILTERS

The arguments that were employed in the last two sections forLMS and 6-NLMS aregeneral enough an d can be applied to ada ptive filters with generic data non linearities of theform (22.2)-(22.3). To see this, consider again the variance and mean relations(22.26)-(22.27) and (22.29),which are reproduced below:

(25.14)If we now introduce theM2 x M 2 matrices

A (E [ IT@ 1) ,,, (25.15)

B E ( [ Z l T @ [ ] ) (25.16)

and theM x M matrix

(25.17)

thenEt3 i = (I - P ) E t 3 i - l

and the expression for C' ca n be written in terms of the linear vector relation

6' = F o (25.18)

where F is MZ x M2 and given by

AF = I - p A + p 2 B (25.19)

Let

H = -:I2 ] 2 M 2 x 2 M 2 ) (25.20)Then the same arguments that were used in Chapter24 will lead to the statement ofThm. 25.2 listed further ahead. The expressions for theMSD and EMSE in the state-ment of the theore m are derived in a manner similar to(23.51) and (23.56). They can berewritten as

I MSD = p2a:Tr(SC,,d) and EMSE = p2a:Tr(SCe,,e) I

8/11/2019 Data Normalized Filters

5/9

whereS = E ( ~ b ~ i / g ~ [ ~ i ] )

and the weightingmatrices {Cmsd,C,,,,} correspond to the vectorsOmsd = (I - F ) - l qand gemse (I - F ) - ~ T . hat is,

Cmsd= VeC-l(Om,d) and C, = VeC-l(Oemse)

Theorem 25.2 (Stability of data-normalized filters) Consider data normalizedadaptive filters o f the form (22.2)-(22.3), and assume th e data { d ( i ) , i} at-isfy model (22.1) and the independence assumption (22.23). Then th e filteris convergent in the mean and is mean-square stable for step-sizes satisfying

0 < p < min{2/Xm,(P), l/Xmax(A-'B), l /max{X(H) E R }}

where the matrices {A, B , P, H} are defined by (25.15)-(25.17) and (25.20)and B is assumed finite. Moreover, the transient behavior of the filter ischaracterized by the M2-dimensional state-space recursion Wi = W i - 1 +p 2 g : y , where F s the companion mat rix

=

with

0 10 0 10 0 0 1

M 2 x M 2 )

MZ-1A

p ( z ) = d e t ( s 1 - F ) = s M Z p k z kk O

denoting the characteristic polynomial of F in (25.19). Also,

Aw

for any 0 of interest, e.g., g = q or = T . In addition, the mean-squaredeviation and the excess mean-square error are given by

375SECTION25.2

DATA-NORMALIZED

FILTERS

where T = vec(R,) and q = vec(1).

8/11/2019 Data Normalized Filters

6/9

76 Learning CurveCHAPTER25NORMALIZEDFILTERS

As before, sinceE lea i)i2 E lIGi 1 Ilkuwe find that the time evo lution ofE lea z)l2 sdescribed by the top entry of the sta te vectorWi in with hosen aso = T . The learningcurve of the filter will beE le i)I2 = c: + E l ea i ) I2 .

DATA-

Small Step-Size ApproximationsIn Rob. V.39 it is shown that under a boundedness requirement on the m atrixB of fourthmoments, data-normalized adaptive filters can be guaranteed to be mean-square stable for

sufficiently small step-sizes. Th atis, there always exists a small-en ough step-size that lieswithin the stability range described inThm. 25.2.

Now observe that the performance results ofThm. 25.2 are in terms of the momentmatrices { A ,B , P } . These mom ents are generally hard to evaluate for arbitrary inputdistributions and data nonlinearitiesg[. . However, some simplifications occur when thestep-sizeis sufficiently small.This is because, inthis case, we may ignore the quadraticterm in p that appears in the expression forC in (25.14), and thereby approximate thevariance and m ean relations by

where P is as in (25.17).Using the weigh ting vector notation, we can write

(25.21)

(25.22)

F = I - p A (25.23)

where nowA = ( P ~ C N ) + ( I B P )

The variance relation(25.22) would then lead to the following approximate expressionsfor the filterEMSEand MSD:

EMSE = p2aiTr(SCemse) and MSD = p20~Tr(sCmsd)

where

and the weigh ting matrices {Cemse,Cmsd} correspon d to the vectorsoemse A-'vec( R,)/pand cmsd = A-lvec(I)/p. That is, {Cemse,Cmsd} are the un ique solution s of the Lya-punov equations

S = E ~ t u i / g ~ [ ~ i ] )

p p z m s d +PCrnsdP = I and p p z e m s e -I- p x e m s e p = u

It is easy to Verify thatCmsd = p-'P- ' /2 so that the performance expressions can berewrittenas

I EMSE = p20:Tr(SCemse), MSD = pa:Tr(SP-')/2

Remark 25.1 (Filters with error nonlinearities) There is more to say about the transient per-formance of adaptive filters, especiallyfor filters witherrornonlinearities in their update equations.This is a more challenging class of filters to study and their performance is examined in App. 9.C

8/11/2019 Data Normalized Filters

7/9

of Sayed(2003)by using the same energy-conservation arguments of this part. The derivation usedin that appendix to study adaptive filters with error nonlinearities can also be used to provide analternative simplified transient analysis for data-no rmalized filters. The derivation is based on a longfilter assumption in order to justify a Gaussian condition on the distribution of thea priori errorsignal. Among other results, it is shown in App.9.C of Sayed (2003)that the transient behavior ofdata-normalized filters can be approximated by an M-dimensional linear time-invariant state-spacemodel even for non-Gaussian regressors. Appendix9.E of the same reference further

*View more*